Attach a iommu [DMAR] fault handler for our device and try and reset the
GPU upon a fault. At worst this will allow us to more quickly recover
from a fault, rather than wait 10s for the hangcheck to determine a
stuctk GPU. At best, it will immediately restart the GPU and paper over
the bad iommu.
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/i915_drv.c | 21 +++++++++++++++++++++
1 file changed, 21 insertions(+)
diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index f2389ba49c69..f881de6e4583 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -501,6 +501,22 @@ static int i915_set_dma_info(struct drm_i915_private *i915)
return ret;
}
+static int fault_handler(struct iommu_fault *f, void *arg)
+{
+ struct drm_i915_private *i915 = arg;
+
+ intel_gt_handle_error(&i915->gt, ALL_ENGINES, 0, "DMAR fault");
+
+ /*
+ * If we successfully handle the fault, eg mapping a new page,
+ * we should call iommu_page_response().
+ *
+ * We make no attempt to resolve the cause of the fault, as it
+ * should only be from misconfiguration of the iommu device itself.
+ */
+ return 0;
+}
+
/**
* i915_driver_hw_probe - setup state requiring device access
* @dev_priv: device private
@@ -621,6 +637,9 @@ static int i915_driver_hw_probe(struct drm_i915_private *dev_priv)
intel_bw_init_hw(dev_priv);
+ iommu_register_device_fault_handler(dev_priv->drm.dev,
+ fault_handler, dev_priv);
+
return 0;
err_msi:
@@ -644,6 +663,8 @@ static void i915_driver_hw_remove(struct drm_i915_private *dev_priv)
{
struct pci_dev *pdev = dev_priv->drm.pdev;
+ iommu_unregister_device_fault_handler(dev_priv->drm.dev);
+
i915_perf_fini(dev_priv);
if (pdev->msi_enabled)
--
2.20.1
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
