Skip to contentSkip to conversation

drm/i915: prevent gt fifo count underflow

PATCH New Mika Kuoppala 3 participants 5 messages May 14, 2014

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: prevent gt fifo count underflow

If we get the final value of zero as a count of free entries available, we will underflow our own fifo_count and then it will take a long time before we check things again. Admittedly we are in trouble already if we get into this situation, but prevent the underflow by checking against zero before decreasing fifo_count.

---
 drivers/gpu/drm/i915/intel_uncore.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_uncore.c +1 -1
index 76dc185..159f8ab 100644
@@ -174,7 +174,7 @@ static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 		}
 		if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
 			++ret;
-		dev_priv->uncore.fifo_count = fifo;
+		dev_priv->uncore.fifo_count = fifo ?: 1;
 	}
 	dev_priv->uncore.fifo_count--;
 
Signature · 1 line
1.7.9.5
#1 lore raw

PATCH 1/1 drm/i915: prevent gt fifo count underflow

If we get the final value of zero as a count of free entries available, we will underflow our own fifo_count and then it will take a long time before we check things again. Admittedly we are in trouble already if we get into this situation, but prevent the underflow by returning early.

v2: Less convoluted control flow (Daniel Vetter)

---
 drivers/gpu/drm/i915/intel_uncore.c |   20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)
drivers/gpu/drm/i915/intel_uncore.c +9 -11
index 76dc185..bf1b661 100644
@@ -154,10 +154,8 @@ static void __gen7_gt_force_wake_mt_put(struct drm_i915_private *dev_priv,
 		gen6_gt_check_fifodbg(dev_priv);
 }
 
-static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
+static bool __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 {
-	int ret = 0;
-
 	/* On VLV, FIFO will be shared by both SW and HW.
 	 * So, we need to read the FREE_ENTRIES everytime */
 	if (IS_VALLEYVIEW(dev_priv->dev))
@@ -173,12 +171,12 @@ static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 			fifo = __raw_i915_read32(dev_priv, GTFIFOCTL) & GT_FIFO_FREE_ENTRIES_MASK;
 		}
 		if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
-			++ret;
+			return true;
 		dev_priv->uncore.fifo_count = fifo;
 	}
 	dev_priv->uncore.fifo_count--;
 
-	return ret;
+	return false;
 }
 
 static void vlv_force_wake_reset(struct drm_i915_private *dev_priv)
@@ -642,13 +640,13 @@ gen5_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __gen6_write(x) \
 static void \
 gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	REG_WRITE_FOOTER; \
@@ -657,14 +655,14 @@ gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __hsw_write(x) \
 static void \
 hsw_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	hsw_unclaimed_reg_clear(dev_priv, reg); \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	hsw_unclaimed_reg_check(dev_priv, reg); \
Signature · 1 line
1.7.9.5
#2 lore raw
Quoted 30 lines · Mika Kuoppala

If we get the final value of zero as a count of free entries available, we will underflow our own fifo_count and then it will take a long time before we check things again. Admittedly we are in trouble already if we get into this situation, but prevent the underflow by returning early.

v2: Less convoluted control flow (Daniel Vetter)

---
 drivers/gpu/drm/i915/intel_uncore.c |   20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)
drivers/gpu/drm/i915/intel_uncore.c +1 -3
index 76dc185..bf1b661 100644
@@ -154,10 +154,8 @@ static void __gen7_gt_force_wake_mt_put(struct drm_i915_private *dev_priv,
 		gen6_gt_check_fifodbg(dev_priv);
 }
 
-static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
+static bool __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 {
-	int ret = 0;
-
 	/* On VLV, FIFO will be shared by both SW and HW.
 	 * So, we need to read the FREE_ENTRIES everytime */
 	if (IS_VALLEYVIEW(dev_priv->dev))

drivers/gpu/drm/i915/intel_uncore.c static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv) · line 173

Quoted 3 lines
drivers/gpu/drm/i915/intel_uncore.c +0 -0
@@ -173,12 +171,12 @@ static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 			fifo = __raw_i915_read32(dev_priv, GTFIFOCTL) & GT_FIFO_FREE_ENTRIES_MASK;
 		}
 		if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))

Maybe kill the 'loop<0' check while at it. It's redundant and IMO just makes things less obvious.

Also I don't get why we first check for
'fifo_count < GT_FIFO_NUM_RESERVED_ENTRIES', but then the while loop checks for 'fifo <= GT_FIFO_NUM_RESERVED_ENTRIES'.

drivers/gpu/drm/i915/intel_uncore.c static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv) · line 175

Quoted 3 lines
drivers/gpu/drm/i915/intel_uncore.c +1 -1
-			++ret;
+			return true;
 		dev_priv->uncore.fifo_count = fifo;

We no longer update fifo_count on failure. Not really a problem, but since we've already done all the work maybe we should still update it.

Quoted 50 lines
drivers/gpu/drm/i915/intel_uncore.c +7 -7
 	}
 	dev_priv->uncore.fifo_count--;
 
-	return ret;
+	return false;
 }
 
 static void vlv_force_wake_reset(struct drm_i915_private *dev_priv)
@@ -642,13 +640,13 @@ gen5_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __gen6_write(x) \
 static void \
 gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	REG_WRITE_FOOTER; \
@@ -657,14 +655,14 @@ gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __hsw_write(x) \
 static void \
 hsw_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	hsw_unclaimed_reg_clear(dev_priv, reg); \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	hsw_unclaimed_reg_check(dev_priv, reg); \
Signature · 6 lines
1.7.9.5

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
#3 lore raw
Quoted 38 lines · Ville Syrjälä
Quoted 30 lines · Mika Kuoppala

If we get the final value of zero as a count of free entries available, we will underflow our own fifo_count and then it will take a long time before we check things again. Admittedly we are in trouble already if we get into this situation, but prevent the underflow by returning early.

v2: Less convoluted control flow (Daniel Vetter)

---
 drivers/gpu/drm/i915/intel_uncore.c |   20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)
drivers/gpu/drm/i915/intel_uncore.c +1 -3
index 76dc185..bf1b661 100644
@@ -154,10 +154,8 @@ static void __gen7_gt_force_wake_mt_put(struct drm_i915_private *dev_priv,
 		gen6_gt_check_fifodbg(dev_priv);
 }
 
-static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
+static bool __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 {
-	int ret = 0;
-
 	/* On VLV, FIFO will be shared by both SW and HW.
 	 * So, we need to read the FREE_ENTRIES everytime */
 	if (IS_VALLEYVIEW(dev_priv->dev))

drivers/gpu/drm/i915/intel_uncore.c static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv) · line 173

Quoted 3 lines
drivers/gpu/drm/i915/intel_uncore.c +0 -0
@@ -173,12 +171,12 @@ static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 			fifo = __raw_i915_read32(dev_priv, GTFIFOCTL) & GT_FIFO_FREE_ENTRIES_MASK;
 		}
 		if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))

Maybe kill the 'loop<0' check while at it. It's redundant and IMO just makes things less obvious.

Quoted 3 lines

Also I don't get why we first check for
'fifo_count < GT_FIFO_NUM_RESERVED_ENTRIES', but then the while loop checks for 'fifo <= GT_FIFO_NUM_RESERVED_ENTRIES'.

No good reason, I thought a little bit of hysteresis would be useful.

Quoted 6 lines

drivers/gpu/drm/i915/intel_uncore.c static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv) · line 175

Quoted 3 lines
drivers/gpu/drm/i915/intel_uncore.c +1 -1
-			++ret;
+			return true;
 		dev_priv->uncore.fifo_count = fifo;

We no longer update fifo_count on failure. Not really a problem, but since we've already done all the work maybe we should still update it.

It doesn't matter, the value will still trigger the loop again, so by ignoring it we can keep the code neater.
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#4 lore raw

PATCH v3 1/1 drm/i915: prevent gt fifo count underflow

In reply to Ville Syrjälä

If we get the final value of zero as a count of free entries available, we will underflow our own fifo_count and then it will take a long time before we check things again. Admittedly we are in trouble already if we get into this situation, but prevent the underflow by returning early.

v2: Less convoluted control flow (Daniel Vetter)

v3: Kill redundant loop<0 check (Ville Syrjälä)

---
 drivers/gpu/drm/i915/intel_uncore.c |   23 +++++++++++------------
 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/gpu/drm/i915/intel_uncore.c +11 -12
index 76dc185..72f74d0 100644
@@ -154,10 +154,8 @@ static void __gen7_gt_force_wake_mt_put(struct drm_i915_private *dev_priv,
 		gen6_gt_check_fifodbg(dev_priv);
 }
 
-static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
+static bool __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 {
-	int ret = 0;
-
 	/* On VLV, FIFO will be shared by both SW and HW.
 	 * So, we need to read the FREE_ENTRIES everytime */
 	if (IS_VALLEYVIEW(dev_priv->dev))
@@ -172,13 +170,14 @@ static int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
 			udelay(10);
 			fifo = __raw_i915_read32(dev_priv, GTFIFOCTL) & GT_FIFO_FREE_ENTRIES_MASK;
 		}
-		if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
-			++ret;
+		if (WARN_ON(fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
+			return true;
+
 		dev_priv->uncore.fifo_count = fifo;
 	}
 	dev_priv->uncore.fifo_count--;
 
-	return ret;
+	return false;
 }
 
 static void vlv_force_wake_reset(struct drm_i915_private *dev_priv)
@@ -642,13 +641,13 @@ gen5_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __gen6_write(x) \
 static void \
 gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	REG_WRITE_FOOTER; \
@@ -657,14 +656,14 @@ gen6_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace
 #define __hsw_write(x) \
 static void \
 hsw_write##x(struct drm_i915_private *dev_priv, off_t reg, u##x val, bool trace) { \
-	u32 __fifo_ret = 0; \
+	bool __fifo_failed = false; \
 	REG_WRITE_HEADER; \
 	if (NEEDS_FORCE_WAKE((dev_priv), (reg))) { \
-		__fifo_ret = __gen6_gt_wait_for_fifo(dev_priv); \
+		__fifo_failed = __gen6_gt_wait_for_fifo(dev_priv); \
 	} \
 	hsw_unclaimed_reg_clear(dev_priv, reg); \
 	__raw_i915_write##x(dev_priv, reg, val); \
-	if (unlikely(__fifo_ret)) { \
+	if (unlikely(__fifo_failed)) { \
 		gen6_gt_check_fifodbg(dev_priv); \
 	} \
 	hsw_unclaimed_reg_check(dev_priv, reg); \
Signature · 6 lines
1.7.9.5

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx