Save the initial HW state snapshot in i915 so that the rest code of GVT-g
can be moved into a dedicated module while it can still get a clean
initial HW state saved at the correct time during the initialization of
i915. The futhrer vGPU created by GVT-g will use this HW state as the
initial HW state.
v6:
- Remove the reference of intel_gvt_device_info.(Christoph)
- Refine the save_mmio() function. (Christoph)
Cc: Christoph Hellwig <[email protected]>
Cc: Jason Gunthorpe <[email protected]>
Cc: Jani Nikula <[email protected]>
Cc: Joonas Lahtinen <[email protected]>
Cc: Vivi Rodrigo <[email protected]>
Cc: Zhenyu Wang <[email protected]>
Cc: Zhi Wang <[email protected]>
Signed-off-by: Zhi Wang <[email protected]>
---
drivers/gpu/drm/i915/i915_drv.h | 2 +
drivers/gpu/drm/i915/intel_gvt.c | 92 +++++++++++++++++++++++++++++++-
2 files changed, 92 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 44c1f98144b4..2a230840cdfa 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -606,6 +606,8 @@ struct i915_virtual_gpu {
struct mutex lock; /* serialises sending of g2v_notify command pkts */
bool active;
u32 caps;
+ u32 *initial_mmio;
+ u8 *initial_cfg_space;
};
struct i915_selftest_stash {
diff --git a/drivers/gpu/drm/i915/intel_gvt.c b/drivers/gpu/drm/i915/intel_gvt.c
index 32aa6f111d6b..c5019173ac44 100644
--- a/drivers/gpu/drm/i915/intel_gvt.c
+++ b/drivers/gpu/drm/i915/intel_gvt.c
@@ -68,6 +68,85 @@ void intel_gvt_sanitize_options(struct drm_i915_private *dev_priv)
dev_priv->params.enable_gvt = 0;
}
+static void free_initial_hw_state(struct drm_i915_private *dev_priv)
+{
+ struct i915_virtual_gpu *vgpu = &dev_priv->vgpu;
+
+ vfree(vgpu->initial_mmio);
+ vgpu->initial_mmio = NULL;
+
+ kfree(vgpu->initial_cfg_space);
+ vgpu->initial_cfg_space = NULL;
+}
+
+static void save_mmio(struct intel_gvt_mmio_table_iter *iter, u32 offset,
+ u32 size)
+{
+ struct drm_i915_private *dev_priv = iter->i915;
+ u32 *mmio, i;
+
+ for (i = offset; i < offset + size; i += 4) {
+ mmio = iter->data + i;
+ *mmio = intel_uncore_read_notrace(to_gt(dev_priv)->uncore,
+ _MMIO(i));
+ }
+}
+
+static int handle_mmio(struct intel_gvt_mmio_table_iter *iter,
+ u32 offset, u32 device, u32 size)
+{
+ if (WARN_ON(!IS_ALIGNED(offset, 4)))
+ return -EINVAL;
+
+ save_mmio(iter, offset, size);
+ return 0;
+}
+
+static int save_initial_hw_state(struct drm_i915_private *dev_priv)
+{
+ struct pci_dev *pdev = to_pci_dev(dev_priv->drm.dev);
+ struct i915_virtual_gpu *vgpu = &dev_priv->vgpu;
+ struct intel_gvt_mmio_table_iter iter;
+ void *mem;
+ int i, ret;
+
+ mem = kzalloc(PCI_CFG_SPACE_EXP_SIZE, GFP_KERNEL);
+ if (!mem)
+ return -ENOMEM;
+
+ vgpu->initial_cfg_space = mem;
+
+ for (i = 0; i < PCI_CFG_SPACE_EXP_SIZE; i += 4)
+ pci_read_config_dword(pdev, i, mem + i);
+
+ mem = vzalloc(2 * SZ_1M);
+ if (!mem) {
+ ret = -ENOMEM;
+ goto err_mmio;
+ }
+
+ vgpu->initial_mmio = mem;
+
+ iter.i915 = dev_priv;
+ iter.data = vgpu->initial_mmio;
+ iter.handle_mmio_cb = handle_mmio;
+
+ ret = intel_gvt_iterate_mmio_table(&iter);
+ if (ret)
+ goto err_iterate;
+
+ return 0;
+
+err_iterate:
+ vfree(vgpu->initial_mmio);
+ vgpu->initial_mmio = NULL;
+err_mmio:
+ kfree(vgpu->initial_cfg_space);
+ vgpu->initial_cfg_space = NULL;
+
+ return ret;
+}
+
/**
* intel_gvt_init - initialize GVT components
* @dev_priv: drm i915 private data
@@ -97,15 +176,23 @@ int intel_gvt_init(struct drm_i915_private *dev_priv)
return -EIO;
}
+ ret = save_initial_hw_state(dev_priv);
+ if (ret) {
+ drm_dbg(&dev_priv->drm, "Fail to save initial HW state\n");
+ goto err_save_hw_state;
+ }
+
ret = intel_gvt_init_device(dev_priv);
if (ret) {
drm_dbg(&dev_priv->drm, "Fail to init GVT device\n");
- goto bail;
+ goto err_init_device;
}
return 0;
-bail:
+err_init_device:
+ free_initial_hw_state(dev_priv);
+err_save_hw_state:
dev_priv->params.enable_gvt = 0;
return 0;
}
@@ -129,6 +216,7 @@ void intel_gvt_driver_remove(struct drm_i915_private *dev_priv)
return;
intel_gvt_clean_device(dev_priv);
+ free_initial_hw_state(dev_priv);
}
/**
--
2.25.1
On Tue, Feb 08, 2022 at 06:11:50AM -0500, Zhi Wang wrote:
> + struct drm_i915_private *dev_priv = iter->i915;
> + u32 *mmio, i;
> +
> + for (i = offset; i < offset + size; i += 4) {
> + mmio = iter->data + i;
> + *mmio = intel_uncore_read_notrace(to_gt(dev_priv)->uncore,
> + _MMIO(i));
This reads much stranger than:
u32 *mmio = iter->data;
for (i = offset; i < offset + size; i += 4) {
mmio[i] = intel_uncore_read_notrace(to_gt(dev_priv)->uncore,
_MMIO(i));
}
> +static int handle_mmio(struct intel_gvt_mmio_table_iter *iter,
> + u32 offset, u32 device, u32 size)
> +{
> + if (WARN_ON(!IS_ALIGNED(offset, 4)))
> + return -EINVAL;
Shouldn't this be in the caller of the method?
> + save_mmio(iter, offset, size);
> + return 0;
Now that the block callback is gone save_mmio and handle_mmio
can be merged.
> + mem = vzalloc(2 * SZ_1M);
Don't we want a driver-wide constant for this instead of a magic number?
On 2/9/22 7:32 AM, Christoph Hellwig wrote:
> On Tue, Feb 08, 2022 at 06:11:50AM -0500, Zhi Wang wrote:
>> + struct drm_i915_private *dev_priv = iter->i915;
>> + u32 *mmio, i;
>> +
>> + for (i = offset; i < offset + size; i += 4) {
>> + mmio = iter->data + i;
>> + *mmio = intel_uncore_read_notrace(to_gt(dev_priv)->uncore,
>> + _MMIO(i));
>
> This reads much stranger than:
>
> u32 *mmio = iter->data;
>
> for (i = offset; i < offset + size; i += 4) {
> mmio[i] = intel_uncore_read_notrace(to_gt(dev_priv)->uncore,
> _MMIO(i));
> }
>
I am not sure the suggestion is correct. That's the reason why I didn't take the comments in the previous review.
if mmio is u32 *, the step of mmio[0] -> mmio[1] will be 4, and i will be increased by 4 in each loop.
I guess the correct one would be mmio[i/4] = xxxxx; would that looks better? if yes, I will do that in the next version.
>> +static int handle_mmio(struct intel_gvt_mmio_table_iter *iter,
>> + u32 offset, u32 device, u32 size)
>> +{
>> + if (WARN_ON(!IS_ALIGNED(offset, 4)))
>> + return -EINVAL;
>
> Shouldn't this be in the caller of the method?
>
>> + save_mmio(iter, offset, size);
>> + return 0;
>
Yes. You are right. It's because I get rid of the mmio_block in intel_gvt.c
> Now that the block callback is gone save_mmio and handle_mmio
> can be merged.
>
>> + mem = vzalloc(2 * SZ_1M);
>
> Don't we want a driver-wide constant for this instead of a magic number?
>
We actually have one in i915, but it's not exported. Should we export that one?
Thanks,
Zhi.