Skip to contentSkip to conversation

drm/i915: Force wake restore for TDR

RFC In review Lister, Ian 2 participants 2 messages Dec 16, 2013

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

RFC 4/13 drm/i915: Force wake restore for TDR

Quoted 1 line

From 698fdaf8a85fb7039dacdff3068be2018a486816 Mon Sep 17 00:00:00 2001

Message-Id: <698fdaf8a85fb7039dacdff3068be2018a486816.1387201899.git.ian.lister@intel.com>
In-Reply-To: <[email protected]>
References: <[email protected]>
From: ian-lister <[email protected]>
Date: Mon, 9 Dec 2013 09:47:20 +0000
Subject: [RFC 04/13] drm/i915: Force wake restore for TDR

Abstracted force wake restore code from gen6_do_reset into its own function "gen6_gt_force_wake_restore" as it will also be needed for per-engine reset. It has been expanded to cope with the Valleyview specific forcewake restore.

---
 drivers/gpu/drm/i915/i915_drv.h         |   4 +
 drivers/gpu/drm/i915/i915_reg.h         |   1 +
 drivers/gpu/drm/i915/intel_ringbuffer.h |   1 +
 drivers/gpu/drm/i915/intel_uncore.c     | 197 ++++++++++++++++++++++++++++++--
 4 files changed, 192 insertions(+), 11 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +4 -0
index f261ab5..b9c4876 100644
@@ -459,6 +459,8 @@ struct intel_uncore {
 	unsigned fw_rendercount;
 	unsigned fw_mediacount;
 
+	unsigned ecobus;
+
 	struct delayed_work force_wake_work;
 };
 
@@ -1927,6 +1929,8 @@ extern int i915_emit_box(struct drm_device *dev,
 			 struct drm_clip_rect *box,
 			 int DR1, int DR4);
 extern int intel_gpu_reset(struct drm_device *dev);
+extern int intel_gpu_engine_reset(struct drm_device *dev,
+				enum intel_ring_id engine);
 extern int i915_reset(struct drm_device *dev);
 extern unsigned long i915_chipset_val(struct drm_i915_private *dev_priv);
 extern unsigned long i915_mch_val(struct drm_i915_private *dev_priv);
drivers/gpu/drm/i915/i915_reg.h +1 -0
index bda7562..e896bca 100644
@@ -107,6 +107,7 @@
 #define  GEN6_GRDOM_RENDER		(1 << 1)
 #define  GEN6_GRDOM_MEDIA		(1 << 2)
 #define  GEN6_GRDOM_BLT			(1 << 3)
+#define  GEN6_GRDOM_VEBOX               (1 << 4)
 
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
drivers/gpu/drm/i915/intel_ringbuffer.h +1 -0
index c15b97b..6417de1 100644
@@ -65,6 +65,7 @@ struct intel_ring_hangcheck {
 	u32 hd;
 	int score;
 	enum intel_ring_hangcheck_action action;
+	u32 resets; /* Total resets applied to this ring/engine*/
 };
 
 struct  intel_ring_buffer {
drivers/gpu/drm/i915/intel_uncore.c +186 -11

239 lines not on this page · raw message

Signature · 1 line
1.8.5.1
#1 lore raw
Quoted 8 lines · Ian

From 698fdaf8a85fb7039dacdff3068be2018a486816 Mon Sep 17 00:00:00 2001
Message-Id: <698fdaf8a85fb7039dacdff3068be2018a486816.1387201899.git.ian.lister@intel.com>
In-Reply-To: <[email protected]>
References: <[email protected]>
From: ian-lister <[email protected]>
Date: Mon, 9 Dec 2013 09:47:20 +0000
Subject: [RFC 04/13] drm/i915: Force wake restore for TDR

Quoted 5 lines

Abstracted force wake restore code from gen6_do_reset into its own function "gen6_gt_force_wake_restore" as it will also be needed for per-engine reset. It has been expanded to cope with the Valleyview specific forcewake restore.

Some comments inline.
-Daniel

Quoted 12 lines
---
 drivers/gpu/drm/i915/i915_drv.h         |   4 +
 drivers/gpu/drm/i915/i915_reg.h         |   1 +
 drivers/gpu/drm/i915/intel_ringbuffer.h |   1 +
 drivers/gpu/drm/i915/intel_uncore.c     | 197 ++++++++++++++++++++++++++++++--
 4 files changed, 192 insertions(+), 11 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +0 -0
index f261ab5..b9c4876 100644

drivers/gpu/drm/i915/i915_drv.h struct intel_uncore { · line 462

Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -459,6 +459,8 @@ struct intel_uncore {
 	unsigned fw_rendercount;
 	unsigned fw_mediacount;
 
+	unsigned ecobus;

This shouldn't be needed since we should assign proper vfunc entries at init time. If I'm wrong and we indeed need this pls give it a descriptive name like has_mt_forcewake or similar. Plus one-line comment here.

But on a quick read through of the patch it looks unneeded. If a later patch needs it the commit message should mention this.

Quoted 72 lines
drivers/gpu/drm/i915/i915_drv.h +3 -0
+
 	struct delayed_work force_wake_work;
 };
 
@@ -1927,6 +1929,8 @@ extern int i915_emit_box(struct drm_device *dev,
 			 struct drm_clip_rect *box,
 			 int DR1, int DR4);
 extern int intel_gpu_reset(struct drm_device *dev);
+extern int intel_gpu_engine_reset(struct drm_device *dev,
+				enum intel_ring_id engine);
 extern int i915_reset(struct drm_device *dev);
 extern unsigned long i915_chipset_val(struct drm_i915_private *dev_priv);
 extern unsigned long i915_mch_val(struct drm_i915_private *dev_priv);
drivers/gpu/drm/i915/i915_reg.h +1 -0
index bda7562..e896bca 100644
@@ -107,6 +107,7 @@
 #define  GEN6_GRDOM_RENDER		(1 << 1)
 #define  GEN6_GRDOM_MEDIA		(1 << 2)
 #define  GEN6_GRDOM_BLT			(1 << 3)
+#define  GEN6_GRDOM_VEBOX               (1 << 4)
 
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
drivers/gpu/drm/i915/intel_ringbuffer.h +1 -0
index c15b97b..6417de1 100644
@@ -65,6 +65,7 @@ struct intel_ring_hangcheck {
 	u32 hd;
 	int score;
 	enum intel_ring_hangcheck_action action;
+	u32 resets; /* Total resets applied to this ring/engine*/
 };
 
 struct  intel_ring_buffer {
drivers/gpu/drm/i915/intel_uncore.c +27 -0
index 848bdaf..764e2af 100644
@@ -283,6 +283,44 @@ void vlv_force_wake_put(struct drm_i915_private *dev_priv,
 	spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
 }
 
+void vlv_force_wake_restore(struct drm_i915_private *dev_priv,
+				int fw_engine)
+{
+	/* Restore the current force wake state with the hardware
+	 * WARNING: Caller *MUST* hold uncore.lock whilst calling this function
+	 */
+
+	if (FORCEWAKE_RENDER & fw_engine) {
+		if (dev_priv->uncore.fw_rendercount
+		&& (__raw_i915_read32(dev_priv, FORCEWAKE_ACK_VLV)
+			& FORCEWAKE_KERNEL) == 0)
+			dev_priv->uncore.funcs.force_wake_get(dev_priv,
+				FORCEWAKE_RENDER);
+		else if (!dev_priv->uncore.fw_rendercount
+		&& (__raw_i915_read32(dev_priv, FORCEWAKE_ACK_VLV)
+			& FORCEWAKE_KERNEL) == 1)
+			dev_priv->uncore.funcs.force_wake_put(dev_priv,
+				FORCEWAKE_RENDER);
+	}
+
+	if (FORCEWAKE_MEDIA & fw_engine) {
+		if (dev_priv->uncore.fw_mediacount
+		&& (__raw_i915_read32(dev_priv, FORCEWAKE_ACK_MEDIA_VLV)
+			& FORCEWAKE_KERNEL) == 0)
+			dev_priv->uncore.funcs.force_wake_get(dev_priv,
+				FORCEWAKE_MEDIA);
+		else if (!dev_priv->uncore.fw_mediacount

drivers/gpu/drm/i915/intel_uncore.c void vlv_force_wake_put(struct drm_i915_private *dev_priv, · line 322

Quoted 10 lines
drivers/gpu/drm/i915/intel_uncore.c +10 -0
+		&& (__raw_i915_read32(dev_priv, FORCEWAKE_ACK_MEDIA_VLV)
+                        & FORCEWAKE_KERNEL) == 1)
+			dev_priv->uncore.funcs.force_wake_put(dev_priv,
+				FORCEWAKE_MEDIA);
+	}
+
+	dev_priv->uncore.fifo_count =
+		__raw_i915_read32(dev_priv, GTFIFOCTL)
+			& GT_FIFO_FREE_ENTRIES_MASK;
+}

Can't we have a generic forcewake restore function which does the right thing depending upon domains? I.e. create a new function intel_uncore_forcewake_restore which dtrt ... By the looks of it the current code in gen6_do_reset is already broken on vlv.

Bonus points if this function is shared with the normal reset code (i.e. the global reset) in gen6_do_reset.

Quoted 163 lines
drivers/gpu/drm/i915/intel_uncore.c +119 -11
+
 static void gen6_force_wake_work(struct work_struct *work)
 {
 	struct drm_i915_private *dev_priv =
@@ -397,6 +435,50 @@ void gen6_gt_force_wake_put(struct drm_i915_private *dev_priv, int fw_engine)
 	spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
 }
 
+void gen6_gt_force_wake_restore(struct drm_i915_private *dev_priv)
+{
+	/* Restore the current expected force wake state with the
+	 * hardware. This may be required following a reset.
+	 *
+	 * WARNING: Caller *MUST* hold uncore.lock
+	 *
+	 * uncore.lock isn't taken in this function to allow the caller the
+	 * flexibility to do other work immediately before/after
+	 * whilst holding the lock.
+	 */
+	struct drm_device *dev = dev_priv->dev;
+	u32 forcewake_ack;
+
+	if (IS_VALLEYVIEW(dev))
+		return vlv_force_wake_restore(dev_priv, FORCEWAKE_ALL);
+
+	if (IS_HASWELL(dev) || IS_GEN8(dev))
+		forcewake_ack = FORCEWAKE_ACK_HSW;
+	else if (IS_IVYBRIDGE(dev)
+		&& (dev_priv->uncore.ecobus & FORCEWAKE_MT_ENABLE))
+		forcewake_ack = FORCEWAKE_MT_ACK;
+	else
+		forcewake_ack = 0;
+
+	if (dev_priv->uncore.forcewake_count
+	&& (!forcewake_ack 
+		|| ((__raw_i915_read32(dev_priv, forcewake_ack)
+			& FORCEWAKE_KERNEL) == 0))) {
+		/* It was enabled, so re-enable it */
+		dev_priv->uncore.funcs.force_wake_get(dev_priv, FORCEWAKE_ALL);
+	} else if ((dev_priv->uncore.forcewake_count == 0)
+	&& (!forcewake_ack
+		|| ((__raw_i915_read32(dev_priv, forcewake_ack)
+			& FORCEWAKE_KERNEL) == 1))) {
+		/* It was disabled, so disable it */
+		dev_priv->uncore.funcs.force_wake_put(dev_priv, FORCEWAKE_ALL);
+	}
+
+	dev_priv->uncore.fifo_count =
+		__raw_i915_read32(dev_priv, GTFIFOCTL)
+			& GT_FIFO_FREE_ENTRIES_MASK;
+}
+
 /* We give fast paths for the really cool registers */
 #define NEEDS_FORCE_WAKE(dev_priv, reg) \
 	 ((reg) < 0x40000 && (reg) != FORCEWAKE)
@@ -661,8 +743,6 @@ void intel_uncore_init(struct drm_device *dev)
 		dev_priv->uncore.funcs.force_wake_get = __gen6_gt_force_wake_mt_get;
 		dev_priv->uncore.funcs.force_wake_put = __gen6_gt_force_wake_mt_put;
 	} else if (IS_IVYBRIDGE(dev)) {
-		u32 ecobus;
-
 		/* IVB configs may use multi-threaded forcewake */
 
 		/* A small trick here - if the bios hasn't configured
@@ -674,11 +754,11 @@ void intel_uncore_init(struct drm_device *dev)
 		 */
 		mutex_lock(&dev->struct_mutex);
 		__gen6_gt_force_wake_mt_get(dev_priv, FORCEWAKE_ALL);
-		ecobus = __raw_i915_read32(dev_priv, ECOBUS);
+		dev_priv->uncore.ecobus = __raw_i915_read32(dev_priv, ECOBUS);
 		__gen6_gt_force_wake_mt_put(dev_priv, FORCEWAKE_ALL);
 		mutex_unlock(&dev->struct_mutex);
 
-		if (ecobus & FORCEWAKE_MT_ENABLE) {
+		if (dev_priv->uncore.ecobus & FORCEWAKE_MT_ENABLE) {
 			dev_priv->uncore.funcs.force_wake_get =
 				__gen6_gt_force_wake_mt_get;
 			dev_priv->uncore.funcs.force_wake_put =
@@ -897,16 +977,88 @@ static int gen6_do_reset(struct drm_device *dev)
 
 	intel_uncore_forcewake_reset(dev);
 
-	/* If reset with a user forcewake, try to restore, otherwise turn it off */
-	if (dev_priv->uncore.forcewake_count)
-		dev_priv->uncore.funcs.force_wake_get(dev_priv, FORCEWAKE_ALL);
-	else
-		dev_priv->uncore.funcs.force_wake_put(dev_priv, FORCEWAKE_ALL);
+	spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
+	return ret;
+}
+
+static int gen6_do_engine_reset(struct drm_device *dev,
+                               enum intel_ring_id engine)
+{
+	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_ring_buffer *ring = &dev_priv->ring[engine];
+	int     ret = -ENODEV;
+	unsigned long irqflags;
+	char *reset_event[2];
+	reset_event[1] = NULL;
 
-	/* Restore fifo count */
-	dev_priv->uncore.fifo_count = __raw_i915_read32(dev_priv, GTFIFOCTL) & GT_FIFO_FREE_ENTRIES_MASK;
+	DRM_DEBUG_TDR("Engine reset %d\n", engine);
+
+	/* Hold uncore.lock across reset to prevent any register access
+	 * with forcewake not set correctly
+	 */
+	spin_lock_irqsave(&dev_priv->uncore.lock, irqflags);
+
+	ring->hangcheck.resets++;
+
+	/* Reset the engine.
+	 * GEN6_GDRST is not in the gt power well so no need to check
+	 * for fifo space for the write or forcewake the chip for
+	 * the read.
+	 */
+	switch (engine) {
+	case RCS:
+		__raw_i915_write32(dev_priv, GEN6_GDRST, GEN6_GRDOM_RENDER);
+		/* Spin waiting for the device to ack the reset request */
+		ret = wait_for_atomic((__raw_i915_read32(dev_priv,
+					GEN6_GDRST)
+					& GEN6_GRDOM_RENDER) == 0, 500);
+		DRM_DEBUG_TDR("RCS Reset\n");
+		break;
+
+	case BCS:
+		__raw_i915_write32(dev_priv, GEN6_GDRST, GEN6_GRDOM_BLT);
+
+		/* Spin waiting for the device to ack the reset request */
+		ret = wait_for_atomic((__raw_i915_read32(dev_priv,
+					GEN6_GDRST)
+					& GEN6_GRDOM_BLT) == 0, 500);
+		DRM_DEBUG_TDR("BCS Reset\n");
+		break;
+
+	case VCS:
+		__raw_i915_write32(dev_priv, GEN6_GDRST, GEN6_GRDOM_MEDIA);
+
+		/* Spin waiting for the device to ack the reset request */
+		ret = wait_for_atomic((__raw_i915_read32(dev_priv,
+					GEN6_GDRST)
+					& GEN6_GRDOM_MEDIA) == 0, 500);
+		DRM_DEBUG_TDR("VCS Reset\n");
+	break;
+
+	case VECS:
+		I915_WRITE_NOTRACE(GEN6_GDRST, GEN6_GRDOM_VEBOX);
+
+		/* Spin waiting for the device to ack the reset request */
+		ret = wait_for_atomic((I915_READ_NOTRACE(GEN6_GDRST)
+			& GEN6_GRDOM_VEBOX) == 0, 500);
+		DRM_DEBUG_TDR("VECS Reset\n");
+		break;
+
+	default:
+		DRM_ERROR("Unexpected engine\n");
+		break;
+	}
+
+	gen6_gt_force_wake_restore(dev_priv);
 

drivers/gpu/drm/i915/intel_uncore.c static int gen6_do_reset(struct drm_device *dev) · line 1063

Quoted 10 lines
drivers/gpu/drm/i915/intel_uncore.c +7 -0
 	spin_unlock_irqrestore(&dev_priv->uncore.lock, irqflags);
+
+	/* Do uevent outside of spinlock as uevent can sleep */
+	reset_event[0] = kasprintf(GFP_KERNEL, "RESET RING=%d", engine);
+	kobject_uevent_env(&dev->primary->kdev->kobj,
+		KOBJ_CHANGE, reset_event);
+	kfree(reset_event[0]);
+
 	return ret;
 }

If your patch says "forcewake restore support" it shouldn't contain other stuff like the per-engine reset code here. Or you need to adjust the commit headline, but I think splitting out the forcewake dance makes sense.

drivers/gpu/drm/i915/intel_uncore.c int intel_gpu_reset(struct drm_device *dev) · line 1083

Quoted 12 lines
drivers/gpu/drm/i915/intel_uncore.c +7 -0
@@ -922,6 +1074,29 @@ int intel_gpu_reset(struct drm_device *dev)
 	}
 }
 
+int intel_gpu_engine_reset(struct drm_device *dev, enum intel_ring_id engine)
+{
+	/* Reset an individual engine */
+	int ret = -ENODEV;
+
+	if (!dev)
+		return -EINVAL;

We don't cather to driver bugs, so a BUG_ON here is the right choice (since we'll blow up in a NULL deref later on anyway).

The exception to that rule if it's hard to prove by review that the invariatn holds (e.g. when the invariant depends upon long-term driver state), then a WARN_ON+early exit is appropriate. But checking all callers of this function to pass a proper dev pointer is trivial, so not such a case.

Quoted 27 lines
drivers/gpu/drm/i915/intel_uncore.c +16 -0
+
+	switch (INTEL_INFO(dev)->gen) {
+	case 7:
+	case 6:
+		ret = gen6_do_engine_reset(dev, engine);
+		break;
+	default:
+		DRM_ERROR("Engine reset not supported\n");
+		ret = -ENODEV;
+		break;
+	}
+
+	return ret;
+}
+
+
 void intel_uncore_clear_errors(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
Signature · 7 lines
1.8.5.1


_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch