[PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[Chris Wilson]

Since the hardware sometimes mysteriously totally flummoxes the 64bit
read of a 64bit register when read using a single instruction, split the
read into two instructions. Since the read here is of automatically
incrementing timestamp counters, we also have to be very careful in
order to make sure that it does not increment between the two
instructions.

However, since userspace tried to workaround this issue and so enshrined
this ABI for a broken hardware read and in the process neglected that
the read only fails in some environments, we have to introduce a new
uABI flag for userspace to request the 2x32 bit accurate read of the
timestamp.

v2: Fix alignment check and include details of the workaround for
userspace.

Reported-by: Karol Herbst <freedesktop@karolherbst.de>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91317
Testcase: igt/gem_reg_read
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: stable@vger.kernel.org
---
 drivers/gpu/drm/i915/intel_uncore.c | 26 +++++++++++++++++++-------
 include/uapi/drm/i915_drm.h         |  8 ++++++++
 2 files changed, 27 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/i915/intel_uncore.c b/drivers/gpu/drm/i915/intel_uncore.c
index 2c477663d378..eb244b57b3fd 100644
--- a/drivers/gpu/drm/i915/intel_uncore.c
+++ b/drivers/gpu/drm/i915/intel_uncore.c
@@ -1310,10 +1310,12 @@ int i915_reg_read_ioctl(struct drm_device *dev,
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_reg_read *reg = data;
 	struct register_whitelist const *entry = whitelist;
+	unsigned size;
+	u64 offset;
 	int i, ret = 0;
 
 	for (i = 0; i < ARRAY_SIZE(whitelist); i++, entry++) {
-		if (entry->offset == reg->offset &&
+		if (entry->offset == (reg->offset & -entry->size) &&
 		    (1 << INTEL_INFO(dev)->gen & entry->gen_bitmask))
 			break;
 	}
@@ -1321,23 +1323,33 @@ int i915_reg_read_ioctl(struct drm_device *dev,
 	if (i == ARRAY_SIZE(whitelist))
 		return -EINVAL;
 
+	/* We use the low bits to encode extra flags as the register should
+	 * be naturally aligned (and those that are not so aligned merely
+	 * limit the available flags for that register).
+	 */
+	offset = entry->offset;
+	size = entry->size;
+	size |= reg->offset ^ offset;
+
 	intel_runtime_pm_get(dev_priv);
 
-	switch (entry->size) {
+	switch (size) {
+	case 8 | 1:
+		reg->val = I915_READ64_2x32(offset, offset+4);
+		break;
 	case 8:
-		reg->val = I915_READ64(reg->offset);
+		reg->val = I915_READ64(offset);
 		break;
 	case 4:
-		reg->val = I915_READ(reg->offset);
+		reg->val = I915_READ(offset);
 		break;
 	case 2:
-		reg->val = I915_READ16(reg->offset);
+		reg->val = I915_READ16(offset);
 		break;
 	case 1:
-		reg->val = I915_READ8(reg->offset);
+		reg->val = I915_READ8(offset);
 		break;
 	default:
-		MISSING_CASE(entry->size);
 		ret = -EINVAL;
 		goto out;
 	}
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index b0f82ddab987..83f60f01dca2 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -1087,6 +1087,14 @@ struct drm_i915_reg_read {
 	__u64 offset;
 	__u64 val; /* Return value */
 };
+/* Known registers:
+ *
+ * Render engine timestamp - 0x2358 + 64bit - gen7+
+ * - Note this register returns an invalid value if using the default
+ *   single instruction 8byte read, in order to workaround that use
+ *   offset (0x2538 | 1) instead.
+ *
+ */
 
 struct drm_i915_reset_stats {
 	__u32 ctx_id;
-- 
2.1.4

Re: [PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[Michał Winiarski]

On Thu, Jul 16, 2015 at 12:37:56PM +0100, Chris Wilson wrote:
> Since the hardware sometimes mysteriously totally flummoxes the 64bit
> read of a 64bit register when read using a single instruction, split the
> read into two instructions. Since the read here is of automatically
> incrementing timestamp counters, we also have to be very careful in
> order to make sure that it does not increment between the two
> instructions.
> 
> However, since userspace tried to workaround this issue and so enshrined
> this ABI for a broken hardware read and in the process neglected that
> the read only fails in some environments, we have to introduce a new
> uABI flag for userspace to request the 2x32 bit accurate read of the
> timestamp.
> 
> v2: Fix alignment check and include details of the workaround for
> userspace.
> 
> Reported-by: Karol Herbst <freedesktop@karolherbst.de>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91317
> Testcase: igt/gem_reg_read
Tested-by: Michał Winiarski <michal.winiarski@intel.com>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Michał Winiarski <michal.winiarski@intel.com>
> Cc: stable@vger.kernel.org
> ---
>  drivers/gpu/drm/i915/intel_uncore.c | 26 +++++++++++++++++++-------
>  include/uapi/drm/i915_drm.h         |  8 ++++++++
>  2 files changed, 27 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_uncore.c b/drivers/gpu/drm/i915/intel_uncore.c
> index 2c477663d378..eb244b57b3fd 100644
> --- a/drivers/gpu/drm/i915/intel_uncore.c
> +++ b/drivers/gpu/drm/i915/intel_uncore.c
> @@ -1310,10 +1310,12 @@ int i915_reg_read_ioctl(struct drm_device *dev,
>  	struct drm_i915_private *dev_priv = dev->dev_private;
>  	struct drm_i915_reg_read *reg = data;
>  	struct register_whitelist const *entry = whitelist;
> +	unsigned size;
> +	u64 offset;
>  	int i, ret = 0;
>  
>  	for (i = 0; i < ARRAY_SIZE(whitelist); i++, entry++) {
> -		if (entry->offset == reg->offset &&
> +		if (entry->offset == (reg->offset & -entry->size) &&
>  		    (1 << INTEL_INFO(dev)->gen & entry->gen_bitmask))
>  			break;
>  	}
> @@ -1321,23 +1323,33 @@ int i915_reg_read_ioctl(struct drm_device *dev,
>  	if (i == ARRAY_SIZE(whitelist))
>  		return -EINVAL;
>  
> +	/* We use the low bits to encode extra flags as the register should
> +	 * be naturally aligned (and those that are not so aligned merely
> +	 * limit the available flags for that register).
> +	 */
> +	offset = entry->offset;
> +	size = entry->size;
> +	size |= reg->offset ^ offset;
> +
>  	intel_runtime_pm_get(dev_priv);
>  
> -	switch (entry->size) {
> +	switch (size) {
> +	case 8 | 1:
> +		reg->val = I915_READ64_2x32(offset, offset+4);
> +		break;
>  	case 8:
> -		reg->val = I915_READ64(reg->offset);
> +		reg->val = I915_READ64(offset);
>  		break;
>  	case 4:
> -		reg->val = I915_READ(reg->offset);
> +		reg->val = I915_READ(offset);
>  		break;
>  	case 2:
> -		reg->val = I915_READ16(reg->offset);
> +		reg->val = I915_READ16(offset);
>  		break;
>  	case 1:
> -		reg->val = I915_READ8(reg->offset);
> +		reg->val = I915_READ8(offset);
>  		break;
>  	default:
> -		MISSING_CASE(entry->size);
>  		ret = -EINVAL;
>  		goto out;
>  	}
> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> index b0f82ddab987..83f60f01dca2 100644
> --- a/include/uapi/drm/i915_drm.h
> +++ b/include/uapi/drm/i915_drm.h
> @@ -1087,6 +1087,14 @@ struct drm_i915_reg_read {
>  	__u64 offset;
>  	__u64 val; /* Return value */
>  };
> +/* Known registers:
> + *
> + * Render engine timestamp - 0x2358 + 64bit - gen7+
> + * - Note this register returns an invalid value if using the default
> + *   single instruction 8byte read, in order to workaround that use
> + *   offset (0x2538 | 1) instead.
> + *
> + */
>  
>  struct drm_i915_reset_stats {
>  	__u32 ctx_id;
> -- 
> 2.1.4
> 

Re: [Intel-gfx] [PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[Daniel Vetter]

On Fri, Jul 17, 2015 at 05:10:25PM +0200, Michał Winiarski wrote:
> On Thu, Jul 16, 2015 at 12:37:56PM +0100, Chris Wilson wrote:
> > Since the hardware sometimes mysteriously totally flummoxes the 64bit
> > read of a 64bit register when read using a single instruction, split the
> > read into two instructions. Since the read here is of automatically
> > incrementing timestamp counters, we also have to be very careful in
> > order to make sure that it does not increment between the two
> > instructions.
> > 
> > However, since userspace tried to workaround this issue and so enshrined
> > this ABI for a broken hardware read and in the process neglected that
> > the read only fails in some environments, we have to introduce a new
> > uABI flag for userspace to request the 2x32 bit accurate read of the
> > timestamp.
> > 
> > v2: Fix alignment check and include details of the workaround for
> > userspace.
> > 
> > Reported-by: Karol Herbst <freedesktop@karolherbst.de>
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91317
> > Testcase: igt/gem_reg_read
> Tested-by: Michał Winiarski <michal.winiarski@intel.com>

Where are the mesa/beignet/libva patches for this?
-Daniel

> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Michał Winiarski <michal.winiarski@intel.com>
> > Cc: stable@vger.kernel.org
> > ---
> >  drivers/gpu/drm/i915/intel_uncore.c | 26 +++++++++++++++++++-------
> >  include/uapi/drm/i915_drm.h         |  8 ++++++++
> >  2 files changed, 27 insertions(+), 7 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_uncore.c b/drivers/gpu/drm/i915/intel_uncore.c
> > index 2c477663d378..eb244b57b3fd 100644
> > --- a/drivers/gpu/drm/i915/intel_uncore.c
> > +++ b/drivers/gpu/drm/i915/intel_uncore.c
> > @@ -1310,10 +1310,12 @@ int i915_reg_read_ioctl(struct drm_device *dev,
> >  	struct drm_i915_private *dev_priv = dev->dev_private;
> >  	struct drm_i915_reg_read *reg = data;
> >  	struct register_whitelist const *entry = whitelist;
> > +	unsigned size;
> > +	u64 offset;
> >  	int i, ret = 0;
> >  
> >  	for (i = 0; i < ARRAY_SIZE(whitelist); i++, entry++) {
> > -		if (entry->offset == reg->offset &&
> > +		if (entry->offset == (reg->offset & -entry->size) &&
> >  		    (1 << INTEL_INFO(dev)->gen & entry->gen_bitmask))
> >  			break;
> >  	}
> > @@ -1321,23 +1323,33 @@ int i915_reg_read_ioctl(struct drm_device *dev,
> >  	if (i == ARRAY_SIZE(whitelist))
> >  		return -EINVAL;
> >  
> > +	/* We use the low bits to encode extra flags as the register should
> > +	 * be naturally aligned (and those that are not so aligned merely
> > +	 * limit the available flags for that register).
> > +	 */
> > +	offset = entry->offset;
> > +	size = entry->size;
> > +	size |= reg->offset ^ offset;
> > +
> >  	intel_runtime_pm_get(dev_priv);
> >  
> > -	switch (entry->size) {
> > +	switch (size) {
> > +	case 8 | 1:
> > +		reg->val = I915_READ64_2x32(offset, offset+4);
> > +		break;
> >  	case 8:
> > -		reg->val = I915_READ64(reg->offset);
> > +		reg->val = I915_READ64(offset);
> >  		break;
> >  	case 4:
> > -		reg->val = I915_READ(reg->offset);
> > +		reg->val = I915_READ(offset);
> >  		break;
> >  	case 2:
> > -		reg->val = I915_READ16(reg->offset);
> > +		reg->val = I915_READ16(offset);
> >  		break;
> >  	case 1:
> > -		reg->val = I915_READ8(reg->offset);
> > +		reg->val = I915_READ8(offset);
> >  		break;
> >  	default:
> > -		MISSING_CASE(entry->size);
> >  		ret = -EINVAL;
> >  		goto out;
> >  	}
> > diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> > index b0f82ddab987..83f60f01dca2 100644
> > --- a/include/uapi/drm/i915_drm.h
> > +++ b/include/uapi/drm/i915_drm.h
> > @@ -1087,6 +1087,14 @@ struct drm_i915_reg_read {
> >  	__u64 offset;
> >  	__u64 val; /* Return value */
> >  };
> > +/* Known registers:
> > + *
> > + * Render engine timestamp - 0x2358 + 64bit - gen7+
> > + * - Note this register returns an invalid value if using the default
> > + *   single instruction 8byte read, in order to workaround that use
> > + *   offset (0x2538 | 1) instead.
> > + *
> > + */
> >  
> >  struct drm_i915_reset_stats {
> >  	__u32 ctx_id;
> > -- 
> > 2.1.4
> > 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Re: [Intel-gfx] [PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[Chris Wilson]

On Tue, Jul 21, 2015 at 08:49:31AM +0200, Daniel Vetter wrote:
> On Fri, Jul 17, 2015 at 05:10:25PM +0200, Michał Winiarski wrote:
> > On Thu, Jul 16, 2015 at 12:37:56PM +0100, Chris Wilson wrote:
> > > Since the hardware sometimes mysteriously totally flummoxes the 64bit
> > > read of a 64bit register when read using a single instruction, split the
> > > read into two instructions. Since the read here is of automatically
> > > incrementing timestamp counters, we also have to be very careful in
> > > order to make sure that it does not increment between the two
> > > instructions.
> > > 
> > > However, since userspace tried to workaround this issue and so enshrined
> > > this ABI for a broken hardware read and in the process neglected that
> > > the read only fails in some environments, we have to introduce a new
> > > uABI flag for userspace to request the 2x32 bit accurate read of the
> > > timestamp.
> > > 
> > > v2: Fix alignment check and include details of the workaround for
> > > userspace.
> > > 
> > > Reported-by: Karol Herbst <freedesktop@karolherbst.de>
> > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91317
> > > Testcase: igt/gem_reg_read
> > Tested-by: Michał Winiarski <michal.winiarski@intel.com>
> 
> Where are the mesa/beignet/libva patches for this?

Trivial. Absolutely trivial. Just waiting for the kernel.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre

Re: [Intel-gfx] [PATCH v3] drm/i915: Use two 32bit reads for select 64bit REG_READ ioctls

[Daniel Vetter]

On Tue, Jul 21, 2015 at 10:45:45AM +0100, Chris Wilson wrote:
> On Tue, Jul 21, 2015 at 08:49:31AM +0200, Daniel Vetter wrote:
> > On Fri, Jul 17, 2015 at 05:10:25PM +0200, Michał Winiarski wrote:
> > > On Thu, Jul 16, 2015 at 12:37:56PM +0100, Chris Wilson wrote:
> > > > Since the hardware sometimes mysteriously totally flummoxes the 64bit
> > > > read of a 64bit register when read using a single instruction, split the
> > > > read into two instructions. Since the read here is of automatically
> > > > incrementing timestamp counters, we also have to be very careful in
> > > > order to make sure that it does not increment between the two
> > > > instructions.
> > > > 
> > > > However, since userspace tried to workaround this issue and so enshrined
> > > > this ABI for a broken hardware read and in the process neglected that
> > > > the read only fails in some environments, we have to introduce a new
> > > > uABI flag for userspace to request the 2x32 bit accurate read of the
> > > > timestamp.
> > > > 
> > > > v2: Fix alignment check and include details of the workaround for
> > > > userspace.
> > > > 
> > > > Reported-by: Karol Herbst <freedesktop@karolherbst.de>
> > > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91317
> > > > Testcase: igt/gem_reg_read
> > > Tested-by: Michał Winiarski <michal.winiarski@intel.com>
> > 
> > Where are the mesa/beignet/libva patches for this?
> 
> Trivial. Absolutely trivial. Just waiting for the kernel.

Well still not how it should be done, so I guess you owe me them all ;-)

Anyway, applied to -fixes.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch