[Intel-gfx] [PATCH 5/8] drm/i915: Add second slice l3 remapping

Ben Widawsky ben at bwidawsk.net
Tue Sep 17 20:45:30 CEST 2013


On Fri, Sep 13, 2013 at 12:38:01PM +0300, Ville Syrjälä wrote:
> On Thu, Sep 12, 2013 at 10:28:31PM -0700, Ben Widawsky wrote:
> > Certain HSW SKUs have a second bank of L3. This L3 remapping has a
> > separate register set, and interrupt from the first "slice". A slice is
> > simply a term to define some subset of the GPU's l3 cache. This patch
> > implements both the interrupt handler, and ability to communicate with
> > userspace about this second slice.
> > 
> > Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> > ---
> >  drivers/gpu/drm/i915/i915_drv.h         |  9 ++--
> >  drivers/gpu/drm/i915/i915_gem.c         | 26 ++++++----
> >  drivers/gpu/drm/i915/i915_irq.c         | 84 +++++++++++++++++++++------------
> >  drivers/gpu/drm/i915/i915_reg.h         |  6 +++
> >  drivers/gpu/drm/i915/i915_sysfs.c       | 34 ++++++++++---
> >  drivers/gpu/drm/i915/intel_ringbuffer.c |  6 +--
> >  include/uapi/drm/i915_drm.h             |  8 ++--
> >  7 files changed, 115 insertions(+), 58 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index 81ba5bb..eb90461 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -918,9 +918,11 @@ struct i915_ums_state {
> >  	int mm_suspended;
> >  };
> >  
> > +#define MAX_L3_SLICES 2
> >  struct intel_l3_parity {
> > -	u32 *remap_info;
> > +	u32 *remap_info[MAX_L3_SLICES];
> >  	struct work_struct error_work;
> > +	int which_slice;
> >  };
> >  
> >  struct i915_gem_mm {
> > @@ -1686,7 +1688,8 @@ struct drm_i915_file_private {
> >  
> >  #define HAS_FORCE_WAKE(dev) (INTEL_INFO(dev)->has_force_wake)
> >  
> > -#define HAS_L3_GPU_CACHE(dev) (IS_IVYBRIDGE(dev) || IS_HASWELL(dev))
> > +#define HAS_L3_GPU_CACHE(dev) (INTEL_INFO(dev)->gen >= 7)
> > +#define NUM_L3_SLICES(dev) (IS_HSW_GT3(dev) ? 2 : HAS_L3_GPU_CACHE(dev))
> >  
> >  #define GT_FREQUENCY_MULTIPLIER 50
> >  
> > @@ -1947,7 +1950,7 @@ bool i915_gem_clflush_object(struct drm_i915_gem_object *obj, bool force);
> >  int __must_check i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj);
> >  int __must_check i915_gem_init(struct drm_device *dev);
> >  int __must_check i915_gem_init_hw(struct drm_device *dev);
> > -void i915_gem_l3_remap(struct drm_device *dev);
> > +void i915_gem_l3_remap(struct drm_device *dev, int slice);
> >  void i915_gem_init_swizzling(struct drm_device *dev);
> >  void i915_gem_cleanup_ringbuffer(struct drm_device *dev);
> >  int __must_check i915_gpu_idle(struct drm_device *dev);
> > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > index 5b510a3..b11f7d6c 100644
> > --- a/drivers/gpu/drm/i915/i915_gem.c
> > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > @@ -4256,16 +4256,21 @@ i915_gem_idle(struct drm_device *dev)
> >  	return 0;
> >  }
> >  
> > -void i915_gem_l3_remap(struct drm_device *dev)
> > +void i915_gem_l3_remap(struct drm_device *dev, int slice)
> >  {
> >  	drm_i915_private_t *dev_priv = dev->dev_private;
> > +	u32 reg_base = GEN7_L3LOG_BASE + (slice * 0x200);
> > +	u32 *remap_info = dev_priv->l3_parity.remap_info[slice];
> >  	u32 misccpctl;
> >  	int i;
> >  
> >  	if (!HAS_L3_GPU_CACHE(dev))
> >  		return;
> >  
> > -	if (!dev_priv->l3_parity.remap_info)
> > +	if (NUM_L3_SLICES(dev) < 2 && slice)
> > +		return;
> 
> This check is redundant as we should never populate
> l3_parity.remap_info[1] when there's no second slice.
> 

Got it. Smashed the early exit check together while at it.

> > +
> > +	if (!remap_info)
> >  		return;
> >  
> >  	misccpctl = I915_READ(GEN7_MISCCPCTL);
> > @@ -4273,17 +4278,17 @@ void i915_gem_l3_remap(struct drm_device *dev)
> >  	POSTING_READ(GEN7_MISCCPCTL);
> >  
> >  	for (i = 0; i < GEN7_L3LOG_SIZE; i += 4) {
> > -		u32 remap = I915_READ(GEN7_L3LOG_BASE + i);
> > -		if (remap && remap != dev_priv->l3_parity.remap_info[i/4])
> > +		u32 remap = I915_READ(reg_base + i);
> > +		if (remap && remap != remap_info[i/4])
> >  			DRM_DEBUG("0x%x was already programmed to %x\n",
> > -				  GEN7_L3LOG_BASE + i, remap);
> > -		if (remap && !dev_priv->l3_parity.remap_info[i/4])
> > +				  reg_base + i, remap);
> > +		if (remap && !remap_info[i/4])
> >  			DRM_DEBUG_DRIVER("Clearing remapped register\n");
> > -		I915_WRITE(GEN7_L3LOG_BASE + i, dev_priv->l3_parity.remap_info[i/4]);
> > +		I915_WRITE(reg_base + i, remap_info[i/4]);
> >  	}
> >  
> >  	/* Make sure all the writes land before disabling dop clock gating */
> > -	POSTING_READ(GEN7_L3LOG_BASE);
> > +	POSTING_READ(reg_base);
> >  
> >  	I915_WRITE(GEN7_MISCCPCTL, misccpctl);
> >  }
> > @@ -4377,7 +4382,7 @@ int
> >  i915_gem_init_hw(struct drm_device *dev)
> >  {
> >  	drm_i915_private_t *dev_priv = dev->dev_private;
> > -	int ret;
> > +	int ret, i;
> >  
> >  	if (INTEL_INFO(dev)->gen < 6 && !intel_enable_gtt())
> >  		return -EIO;
> > @@ -4396,7 +4401,8 @@ i915_gem_init_hw(struct drm_device *dev)
> >  		I915_WRITE(GEN7_MSG_CTL, temp);
> >  	}
> >  
> > -	i915_gem_l3_remap(dev);
> > +	for (i = 0; i < NUM_L3_SLICES(dev); i++)
> > +		i915_gem_l3_remap(dev, i);
> >  
> >  	i915_gem_init_swizzling(dev);
> >  
> > diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> > index 13d26cf..62cdf05 100644
> > --- a/drivers/gpu/drm/i915/i915_irq.c
> > +++ b/drivers/gpu/drm/i915/i915_irq.c
> > @@ -882,9 +882,10 @@ static void ivybridge_parity_work(struct work_struct *work)
> >  	drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
> >  						    l3_parity.error_work);
> >  	u32 error_status, row, bank, subbank;
> > -	char *parity_event[5];
> > +	char *parity_event[6];
> >  	uint32_t misccpctl;
> >  	unsigned long flags;
> > +	uint8_t slice = 0;
> >  
> >  	/* We must turn off DOP level clock gating to access the L3 registers.
> >  	 * In order to prevent a get/put style interface, acquire struct mutex
> > @@ -892,45 +893,63 @@ static void ivybridge_parity_work(struct work_struct *work)
> >  	 */
> >  	mutex_lock(&dev_priv->dev->struct_mutex);
> >  
> > +	/* If we've screwed up tracking, just let the interrupt fire again */
> > +	if (WARN_ON(!dev_priv->l3_parity.which_slice))
> > +		goto out;
> > +
> >  	misccpctl = I915_READ(GEN7_MISCCPCTL);
> >  	I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
> >  	POSTING_READ(GEN7_MISCCPCTL);
> >  
> > -	error_status = I915_READ(GEN7_L3CDERRST1);
> > -	row = GEN7_PARITY_ERROR_ROW(error_status);
> > -	bank = GEN7_PARITY_ERROR_BANK(error_status);
> > -	subbank = GEN7_PARITY_ERROR_SUBBANK(error_status);
> > +	while ((slice = ffs(dev_priv->l3_parity.which_slice)) != 0) {
> > +		u32 reg;
> >  
> > -	I915_WRITE(GEN7_L3CDERRST1, GEN7_PARITY_ERROR_VALID |
> > -				    GEN7_L3CDERRST1_ENABLE);
> > -	POSTING_READ(GEN7_L3CDERRST1);
> > +		if (WARN_ON(slice >= MAX_L3_SLICES))
> > +			break;
> 
> Could be >= NUM_L3_SLICES(dev) for a bit of extra paranoia. Also we
> would fail to clear invalid bits from which_slice in this case, and
> thus we'd get the WARN every time the work runs. But I guess this
> should never happen in any case so probably not worth worrying about
> this too much.

Not worth worrying, but I didn't mean to be so noisy. I've fixed this
with WARN_ON_ONCE.

> 
> >  
> > -	I915_WRITE(GEN7_MISCCPCTL, misccpctl);
> > +		dev_priv->l3_parity.which_slice &= ~(1<<slice);
> >  
> > -	spin_lock_irqsave(&dev_priv->irq_lock, flags);
> > -	ilk_enable_gt_irq(dev_priv, GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
> > -	spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
> > +		reg = GEN7_L3CDERRST1 + (slice * 0x200);
> >  
> > -	mutex_unlock(&dev_priv->dev->struct_mutex);
> > +		error_status = I915_READ(reg);
> > +		row = GEN7_PARITY_ERROR_ROW(error_status);
> > +		bank = GEN7_PARITY_ERROR_BANK(error_status);
> > +		subbank = GEN7_PARITY_ERROR_SUBBANK(error_status);
> > +
> > +		I915_WRITE(reg, GEN7_PARITY_ERROR_VALID | GEN7_L3CDERRST1_ENABLE);
> > +		POSTING_READ(reg);
> > +
> > +		parity_event[0] = I915_L3_PARITY_UEVENT "=1";
> > +		parity_event[1] = kasprintf(GFP_KERNEL, "ROW=%d", row);
> > +		parity_event[2] = kasprintf(GFP_KERNEL, "BANK=%d", bank);
> > +		parity_event[3] = kasprintf(GFP_KERNEL, "SUBBANK=%d", subbank);
> > +		parity_event[4] = kasprintf(GFP_KERNEL, "SLICE=%d", slice);
> > +		parity_event[5] = NULL;
> >  
> > -	parity_event[0] = I915_L3_PARITY_UEVENT "=1";
> > -	parity_event[1] = kasprintf(GFP_KERNEL, "ROW=%d", row);
> > -	parity_event[2] = kasprintf(GFP_KERNEL, "BANK=%d", bank);
> > -	parity_event[3] = kasprintf(GFP_KERNEL, "SUBBANK=%d", subbank);
> > -	parity_event[4] = NULL;
> > +		kobject_uevent_env(&dev_priv->dev->primary->kdev.kobj,
> > +				   KOBJ_CHANGE, parity_event);
> >  
> > -	kobject_uevent_env(&dev_priv->dev->primary->kdev.kobj,
> > -			   KOBJ_CHANGE, parity_event);
> > +		DRM_DEBUG("Parity error: Slice = %d, Row = %d, Bank = %d, Sub bank = %d.\n",
> > +			  slice, row, bank, subbank);
> >  
> > -	DRM_DEBUG("Parity error: Row = %d, Bank = %d, Sub bank = %d.\n",
> > -		  row, bank, subbank);
> > +		kfree(parity_event[4]);
> > +		kfree(parity_event[3]);
> > +		kfree(parity_event[2]);
> > +		kfree(parity_event[1]);
> > +	}
> > +
> > +	I915_WRITE(GEN7_MISCCPCTL, misccpctl);
> > +
> > +out:
> > +	WARN_ON(dev_priv->l3_parity.which_slice);
> 
> First I figured the irq could rearm this behind our back, but we disable
> the irq until the work is done. So yeah, this is fine.
> 
> > +	spin_lock_irqsave(&dev_priv->irq_lock, flags);
> > +	ilk_enable_gt_irq(dev_priv, GT_PARITY_ERROR);
> 
> Is it actually safe to enable the second slice irq when there's no second
> slice? This docs say it's just "reserved", but no mention whether it RO or
> could there be side effects.

Tests on my machine appear to work. But I don't know for certain. Bryan,
could you answer this?

> 
> > +	spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
> >  
> > -	kfree(parity_event[3]);
> > -	kfree(parity_event[2]);
> > -	kfree(parity_event[1]);
> > +	mutex_unlock(&dev_priv->dev->struct_mutex);
> >  }
> >  
> > -static void ivybridge_parity_error_irq_handler(struct drm_device *dev)
> > +static void ivybridge_parity_error_irq_handler(struct drm_device *dev, u32 iir)
> >  {
> >  	drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
> >  
> > @@ -938,9 +957,12 @@ static void ivybridge_parity_error_irq_handler(struct drm_device *dev)
> >  		return;
> >  
> >  	spin_lock(&dev_priv->irq_lock);
> > -	ilk_disable_gt_irq(dev_priv, GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
> > +	ilk_disable_gt_irq(dev_priv, GT_PARITY_ERROR);
> >  	spin_unlock(&dev_priv->irq_lock);
> >  
> > +	iir &= GT_PARITY_ERROR;
> > +	dev_priv->l3_parity.which_slice =
> > +		1 << (iir & GT_RENDER_L3_PARITY_ERROR_INTERRUPT_S1 ? 1 : 0);
> 
> What if both slices report an error at the same time?

I was thinking that such an event can not occur, but on rethinking it
you are right that it's possible. I really hope this never happens, but
it's fixed. Anyway, it should have been |=, not =


[snip]

I'll resend the patch after Bryan answers the question about both
interrupts.

-- 
Ben Widawsky, Intel Open Source Technology Center



More information about the Intel-gfx mailing list