In our platform, we hit the the stolen region initialization failure case, such as below log: [drm:i915_stolen_to_physical] *ERROR* conflict detected with stolen region: [0x7b000000]
And it causes the dev_priv->mm.stolen_base is NULL, in this case, we should avoid accessing it any more.
Here is possible call trace: intel_enable_gt_powersave -- > valleyview_enable_rps -- > valleyview_setup_pctx
Cc: Li Fei fei.li@intel.com Signed-off-by: Liu, Chuansheng chuansheng.liu@intel.com --- drivers/gpu/drm/i915/intel_pm.c | 3 +++ 1 file changed, 3 insertions(+)
diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c index 26c2ea3..1069b24 100644 --- a/drivers/gpu/drm/i915/intel_pm.c +++ b/drivers/gpu/drm/i915/intel_pm.c @@ -3735,6 +3735,9 @@ static void valleyview_setup_pctx(struct drm_device *dev) u32 pcbr; int pctx_size = 24*1024;
+ if (!dev_priv->mm.stolen_base) + return; + pcbr = I915_READ(VLV_PCBR); if (pcbr) { /* BIOS set it up already, grab the pre-alloc'd space */
On Fri, Oct 25, 2013 at 12:33:47AM +0800, Chuansheng Liu wrote:
In our platform, we hit the the stolen region initialization failure case, such as below log: [drm:i915_stolen_to_physical] *ERROR* conflict detected with stolen region: [0x7b000000]
And it causes the dev_priv->mm.stolen_base is NULL, in this case, we should avoid accessing it any more.
Here is possible call trace: intel_enable_gt_powersave -- > valleyview_enable_rps -- > valleyview_setup_pctx
The two create_stolen routines are no-ops in that case so all that happens instead is that we read VLV_PCBR. However, really if i915_gem_object_create_stolen_for_preallocated() fails we should abort loading the driver as it means we have a hardware conflict and undefined behaviour. -Chris
On Thu, Oct 24, 2013 at 01:17:06PM +0100, Chris Wilson wrote:
On Fri, Oct 25, 2013 at 12:33:47AM +0800, Chuansheng Liu wrote:
In our platform, we hit the the stolen region initialization failure case, such as below log: [drm:i915_stolen_to_physical] *ERROR* conflict detected with stolen region: [0x7b000000]
And it causes the dev_priv->mm.stolen_base is NULL, in this case, we should avoid accessing it any more.
Here is possible call trace: intel_enable_gt_powersave -- > valleyview_enable_rps -- > valleyview_setup_pctx
The two create_stolen routines are no-ops in that case so all that happens instead is that we read VLV_PCBR. However, really if i915_gem_object_create_stolen_for_preallocated() fails we should abort loading the driver as it means we have a hardware conflict and undefined behaviour. -Chris
-- Chris Wilson, Intel Open Source Technology Centre
I agree. We should start treating these things as errors since no RPS/RC6 is essentially not what anyone wants.
For another immediate solution it seems you can demote the DRM_ERROR to DRM_DEBUG_DRIVER, and add a check in valleyview_enable_rps for the pctx value.
Hello Chris and Ben,
-----Original Message----- From: Ben Widawsky [mailto:ben@bwidawsk.net] Sent: Friday, October 25, 2013 4:57 AM To: Chris Wilson; Liu, Chuansheng; daniel.vetter@ffwll.ch; airlied@linux.ie; intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; linux-kernel@vger.kernel.org; Li, Fei Subject: Re: [Intel-gfx] drm/i915: Avoid accessing the stolen address when it is unavailable
On Thu, Oct 24, 2013 at 01:17:06PM +0100, Chris Wilson wrote:
On Fri, Oct 25, 2013 at 12:33:47AM +0800, Chuansheng Liu wrote:
In our platform, we hit the the stolen region initialization failure case, such as below log: [drm:i915_stolen_to_physical] *ERROR* conflict detected with stolen
region: [0x7b000000]
And it causes the dev_priv->mm.stolen_base is NULL, in this case, we
should
avoid accessing it any more.
Here is possible call trace: intel_enable_gt_powersave -- > valleyview_enable_rps -- > valleyview_setup_pctx
The two create_stolen routines are no-ops in that case so all that happens instead is that we read VLV_PCBR. However, really if i915_gem_object_create_stolen_for_preallocated() fails we should abort loading the driver as it means we have a hardware conflict and undefined behaviour.
In case of dev_priv->mm.stolen_base == NULL, and the valleyview_setup_pctx() is called at the first time, it will call i915_gem_object_create_stolen_for_preallocated(), which should should return NULL always due to (!drm_mm_initialized(&dev_priv->mm.stolen)).
After that, every time specially when doing pm operation, the above scenario will be called again and again.
Here this patch is to save some time for PM operation, we do not need to read VLV_PCBR and pcbr_offset calculation in case of stolen_base == NULL.
Is it making sense? Thanks.
I agree. We should start treating these things as errors since no RPS/RC6 is essentially not what anyone wants.
DRM_DEBUG_DRIVER, and add a check in valleyview_enable_rps for the pctx value.
The pctx is already checked in valleyview_disable_rps(). Do we need more checking in case of pctx == NULL?
On Fri, Oct 25, 2013 at 12:27:42AM +0000, Liu, Chuansheng wrote:
Hello Chris and Ben,
-----Original Message----- From: Ben Widawsky [mailto:ben@bwidawsk.net] Sent: Friday, October 25, 2013 4:57 AM To: Chris Wilson; Liu, Chuansheng; daniel.vetter@ffwll.ch; airlied@linux.ie; intel-gfx@lists.freedesktop.org; dri-devel@lists.freedesktop.org; linux-kernel@vger.kernel.org; Li, Fei Subject: Re: [Intel-gfx] drm/i915: Avoid accessing the stolen address when it is unavailable
On Thu, Oct 24, 2013 at 01:17:06PM +0100, Chris Wilson wrote:
On Fri, Oct 25, 2013 at 12:33:47AM +0800, Chuansheng Liu wrote:
In our platform, we hit the the stolen region initialization failure case, such as below log: [drm:i915_stolen_to_physical] *ERROR* conflict detected with stolen
region: [0x7b000000]
And it causes the dev_priv->mm.stolen_base is NULL, in this case, we
should
avoid accessing it any more.
Here is possible call trace: intel_enable_gt_powersave -- > valleyview_enable_rps -- > valleyview_setup_pctx
The two create_stolen routines are no-ops in that case so all that happens instead is that we read VLV_PCBR. However, really if i915_gem_object_create_stolen_for_preallocated() fails we should abort loading the driver as it means we have a hardware conflict and undefined behaviour.
In case of dev_priv->mm.stolen_base == NULL, and the valleyview_setup_pctx() is called at the first time, it will call i915_gem_object_create_stolen_for_preallocated(), which should should return NULL always due to (!drm_mm_initialized(&dev_priv->mm.stolen)).
After that, every time specially when doing pm operation, the above scenario will be called again and again.
Here this patch is to save some time for PM operation, we do not need to read VLV_PCBR and pcbr_offset calculation in case of stolen_base == NULL.
Is it making sense? Thanks.
I see. No, it is a pointless optimisation that leaks knowledge about internals of another subsystem to paper over a kernel bug. -Chris
dri-devel@lists.freedesktop.org
-
Ben Widawsky -
Chris Wilson -
Chuansheng Liu -
Liu, Chuansheng