[PATCH v1 1/1] vfio/nvgrace-gpu: Add Blackwell-Next GPU readiness check via CXL DVSEC

[PATCH v1 1/1] vfio/nvgrace-gpu: Add Blackwell-Next GPU readiness check via CXL DVSEC

[ankita]

From: Ankit Agrawal <ankita@nvidia.com>

Blackwell-Next GPUs report device readiness via the CXL DVSEC Range 1 Low
register (offset 0x1C) instead of the BAR0 HBM training register used by
GB200. The GPU memory readiness is checked by polling for the Memory_Active
bit (bit 1) for the Memory_Active_Timeout (bits 15:13).

Add runtime detection by checking the presence of the DVSEC register.
Wire a wait_device_ready ops pointer on nvgrace_gpu_pci_core_device,
which is set at probe to either the Blackwell-Next (CXL DVSEC) or
legacy variant.

Signed-off-by: Ankit Agrawal <ankita@nvidia.com>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 80 ++++++++++++++++++++++++++---
 1 file changed, 72 insertions(+), 8 deletions(-)

diff --git a/drivers/vfio/pci/nvgrace-gpu/main.c b/drivers/vfio/pci/nvgrace-gpu/main.c
index fa056b69f899..8b6b3577a8ea 100644
--- a/drivers/vfio/pci/nvgrace-gpu/main.c
+++ b/drivers/vfio/pci/nvgrace-gpu/main.c
@@ -34,6 +34,12 @@
 #define HBM_TRAINING_BAR0_OFFSET 0x200BC
 #define STATUS_READY 0xFF
 
+#define CXL_DEVICE_DVSEC_ID		0
+#define CXL_DVSEC_RANGE_1_LOW		0x1C
+#define CXL_DVSEC_MEMORY_VALID		BIT(0)
+#define CXL_DVSEC_MEMORY_ACTIVE		BIT(1)
+#define CXL_DVSEC_MEMORY_ACTIVE_TIMEOUT	GENMASK(15, 13)
+
 #define POLL_QUANTUM_MS 1000
 #define POLL_TIMEOUT_MS (30 * 1000)
 
@@ -64,6 +70,10 @@ struct nvgrace_gpu_pci_core_device {
 	bool has_mig_hw_bug;
 	/* GPU has just been reset */
 	bool reset_done;
+	/* CXL Device DVSEC offset; 0 if not present (legacy GB path) */
+	int cxl_dvsec;
+	int (*wait_device_ready)(struct nvgrace_gpu_pci_core_device *nvdev,
+				 void __iomem *io);
 };
 
 static void nvgrace_gpu_init_fake_bar_emu_regs(struct vfio_device *core_vdev)
@@ -242,7 +252,8 @@ static void nvgrace_gpu_close_device(struct vfio_device *core_vdev)
 	vfio_pci_core_close_device(core_vdev);
 }
 
-static int nvgrace_gpu_wait_device_ready(void __iomem *io)
+static int nvgrace_gpu_wait_device_ready_legacy(struct nvgrace_gpu_pci_core_device *nvdev,
+						void __iomem *io)
 {
 	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
 
@@ -256,6 +267,52 @@ static int nvgrace_gpu_wait_device_ready(void __iomem *io)
 	return -ETIME;
 }
 
+/*
+ * Decode the 3-bit Memory_Active_Timeout field from CXL DVSEC Range 1 Low
+ * (bits 15:13) into milliseconds. Encoding per CXL spec r4.0 sec 8.1.3.8.2:
+ * 000b = 1s, 001b = 4s, 010b = 16s, 011b = 64s, 100b = 256s,
+ * 101b-111b = reserved (clamped to 256s).
+ */
+static inline unsigned long nvgrace_gpu_cxl_mem_active_timeout_ms(u8 timeout)
+{
+	return 1000UL << (2 * min_t(u8, timeout, 4));
+}
+
+static int nvgrace_gpu_wait_device_ready_bw_next(struct nvgrace_gpu_pci_core_device *nvdev,
+						 void __iomem *io)
+{
+	struct pci_dev *pdev = nvdev->core_device.pdev;
+	int pcie_dvsec = nvdev->cxl_dvsec;
+	unsigned long timeout;
+	u32 dvsec_memory_status;
+	u8 mem_active_timeout;
+
+	pci_read_config_dword(pdev, pcie_dvsec + CXL_DVSEC_RANGE_1_LOW,
+			      &dvsec_memory_status);
+
+	if (!(dvsec_memory_status & CXL_DVSEC_MEMORY_VALID))
+		return -ENODEV;
+
+	mem_active_timeout = FIELD_GET(CXL_DVSEC_MEMORY_ACTIVE_TIMEOUT,
+				       dvsec_memory_status);
+
+	timeout = jiffies +
+		  msecs_to_jiffies(nvgrace_gpu_cxl_mem_active_timeout_ms(mem_active_timeout));
+
+	do {
+		pci_read_config_dword(pdev,
+				      pcie_dvsec + CXL_DVSEC_RANGE_1_LOW,
+				      &dvsec_memory_status);
+
+		if (dvsec_memory_status & CXL_DVSEC_MEMORY_ACTIVE)
+			return 0;
+
+		msleep(POLL_QUANTUM_MS);
+	} while (!time_after(jiffies, timeout));
+
+	return -ETIME;
+}
+
 /*
  * If the GPU memory is accessed by the CPU while the GPU is not ready
  * after reset, it can cause harmless corrected RAS events to be logged.
@@ -275,7 +332,7 @@ nvgrace_gpu_check_device_ready(struct nvgrace_gpu_pci_core_device *nvdev)
 	if (!__vfio_pci_memory_enabled(vdev))
 		return -EIO;
 
-	ret = nvgrace_gpu_wait_device_ready(vdev->barmap[0]);
+	ret = nvdev->wait_device_ready(nvdev, vdev->barmap[0]);
 	if (ret)
 		return ret;
 
@@ -1146,8 +1203,9 @@ static bool nvgrace_gpu_has_mig_hw_bug(struct pci_dev *pdev)
  * Ensure that the BAR0 region is enabled before accessing the
  * registers.
  */
-static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
+static int nvgrace_gpu_probe_check_device_ready(struct nvgrace_gpu_pci_core_device *nvdev)
 {
+	struct pci_dev *pdev = nvdev->core_device.pdev;
 	void __iomem *io;
 	int ret;
 
@@ -1165,7 +1223,7 @@ static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
 		goto iomap_exit;
 	}
 
-	ret = nvgrace_gpu_wait_device_ready(io);
+	ret = nvdev->wait_device_ready(nvdev, io);
 
 	pci_iounmap(pdev, io);
 iomap_exit:
@@ -1183,10 +1241,6 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
 	u64 memphys, memlength;
 	int ret;
 
-	ret = nvgrace_gpu_probe_check_device_ready(pdev);
-	if (ret)
-		return ret;
-
 	ret = nvgrace_gpu_fetch_memory_property(pdev, &memphys, &memlength);
 	if (!ret)
 		ops = &nvgrace_gpu_pci_ops;
@@ -1198,6 +1252,16 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
 
 	dev_set_drvdata(&pdev->dev, &nvdev->core_device);
 
+	nvdev->cxl_dvsec = pci_find_dvsec_capability(pdev, PCI_VENDOR_ID_CXL,
+						     CXL_DEVICE_DVSEC_ID);
+	nvdev->wait_device_ready = nvdev->cxl_dvsec ?
+		nvgrace_gpu_wait_device_ready_bw_next :
+		nvgrace_gpu_wait_device_ready_legacy;
+
+	ret = nvgrace_gpu_probe_check_device_ready(nvdev);
+	if (ret)
+		goto out_put_vdev;
+
 	if (ops == &nvgrace_gpu_pci_ops) {
 		nvdev->has_mig_hw_bug = nvgrace_gpu_has_mig_hw_bug(pdev);
 
-- 
2.34.1

Re: [PATCH v1 1/1] vfio/nvgrace-gpu: Add Blackwell-Next GPU readiness check via CXL DVSEC

[Alex Williamson]

On Mon, 30 Mar 2026 05:42:20 +0000
<ankita@nvidia.com> wrote:

> From: Ankit Agrawal <ankita@nvidia.com>
> 
> Blackwell-Next GPUs report device readiness via the CXL DVSEC Range 1 Low
> register (offset 0x1C) instead of the BAR0 HBM training register used by
> GB200. The GPU memory readiness is checked by polling for the Memory_Active
> bit (bit 1) for the Memory_Active_Timeout (bits 15:13).
> 
> Add runtime detection by checking the presence of the DVSEC register.
> Wire a wait_device_ready ops pointer on nvgrace_gpu_pci_core_device,
> which is set at probe to either the Blackwell-Next (CXL DVSEC) or
> legacy variant.
> 
> Signed-off-by: Ankit Agrawal <ankita@nvidia.com>
> ---
>  drivers/vfio/pci/nvgrace-gpu/main.c | 80 ++++++++++++++++++++++++++---
>  1 file changed, 72 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/vfio/pci/nvgrace-gpu/main.c b/drivers/vfio/pci/nvgrace-gpu/main.c
> index fa056b69f899..8b6b3577a8ea 100644
> --- a/drivers/vfio/pci/nvgrace-gpu/main.c
> +++ b/drivers/vfio/pci/nvgrace-gpu/main.c
> @@ -34,6 +34,12 @@
>  #define HBM_TRAINING_BAR0_OFFSET 0x200BC
>  #define STATUS_READY 0xFF
>  
> +#define CXL_DEVICE_DVSEC_ID		0
> +#define CXL_DVSEC_RANGE_1_LOW		0x1C
> +#define CXL_DVSEC_MEMORY_VALID		BIT(0)
> +#define CXL_DVSEC_MEMORY_ACTIVE		BIT(1)

Looks like pci_regs.h has all these:

#define PCI_DVSEC_CXL_DEVICE                            0
#define  PCI_DVSEC_CXL_RANGE_SIZE_LOW(i)                (0x1C + (i * 0x10))
#define   PCI_DVSEC_CXL_MEM_INFO_VALID                  _BITUL(0)
#define   PCI_DVSEC_CXL_MEM_ACTIVE                      _BITUL(1)

We should search and replace to use the common macros and drop our
local versions.

> +#define CXL_DVSEC_MEMORY_ACTIVE_TIMEOUT	GENMASK(15, 13)
> +
>  #define POLL_QUANTUM_MS 1000
>  #define POLL_TIMEOUT_MS (30 * 1000)
>  
> @@ -64,6 +70,10 @@ struct nvgrace_gpu_pci_core_device {
>  	bool has_mig_hw_bug;
>  	/* GPU has just been reset */
>  	bool reset_done;
> +	/* CXL Device DVSEC offset; 0 if not present (legacy GB path) */
> +	int cxl_dvsec;
> +	int (*wait_device_ready)(struct nvgrace_gpu_pci_core_device *nvdev,
> +				 void __iomem *io);
>  };
>  
>  static void nvgrace_gpu_init_fake_bar_emu_regs(struct vfio_device *core_vdev)
> @@ -242,7 +252,8 @@ static void nvgrace_gpu_close_device(struct vfio_device *core_vdev)
>  	vfio_pci_core_close_device(core_vdev);
>  }
>  
> -static int nvgrace_gpu_wait_device_ready(void __iomem *io)
> +static int nvgrace_gpu_wait_device_ready_legacy(struct nvgrace_gpu_pci_core_device *nvdev,
> +						void __iomem *io)
>  {
>  	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
>  
> @@ -256,6 +267,52 @@ static int nvgrace_gpu_wait_device_ready(void __iomem *io)
>  	return -ETIME;
>  }
>  
> +/*
> + * Decode the 3-bit Memory_Active_Timeout field from CXL DVSEC Range 1 Low

Is this range 1 or range 0 bits?  The pci_regs.h macro suggests 0.
Same for the commit log.

> + * (bits 15:13) into milliseconds. Encoding per CXL spec r4.0 sec 8.1.3.8.2:
> + * 000b = 1s, 001b = 4s, 010b = 16s, 011b = 64s, 100b = 256s,
> + * 101b-111b = reserved (clamped to 256s).
> + */
> +static inline unsigned long nvgrace_gpu_cxl_mem_active_timeout_ms(u8 timeout)
> +{
> +	return 1000UL << (2 * min_t(u8, timeout, 4));
> +}
> +
> +static int nvgrace_gpu_wait_device_ready_bw_next(struct nvgrace_gpu_pci_core_device *nvdev,
> +						 void __iomem *io)
> +{
> +	struct pci_dev *pdev = nvdev->core_device.pdev;
> +	int pcie_dvsec = nvdev->cxl_dvsec;
> +	unsigned long timeout;
> +	u32 dvsec_memory_status;
> +	u8 mem_active_timeout;
> +
> +	pci_read_config_dword(pdev, pcie_dvsec + CXL_DVSEC_RANGE_1_LOW,
> +			      &dvsec_memory_status);
> +
> +	if (!(dvsec_memory_status & CXL_DVSEC_MEMORY_VALID))
> +		return -ENODEV;
> +
> +	mem_active_timeout = FIELD_GET(CXL_DVSEC_MEMORY_ACTIVE_TIMEOUT,
> +				       dvsec_memory_status);
> +
> +	timeout = jiffies +
> +		  msecs_to_jiffies(nvgrace_gpu_cxl_mem_active_timeout_ms(mem_active_timeout));
> +
> +	do {
> +		pci_read_config_dword(pdev,
> +				      pcie_dvsec + CXL_DVSEC_RANGE_1_LOW,
> +				      &dvsec_memory_status);
> +
> +		if (dvsec_memory_status & CXL_DVSEC_MEMORY_ACTIVE)
> +			return 0;
> +
> +		msleep(POLL_QUANTUM_MS);
> +	} while (!time_after(jiffies, timeout));
> +
> +	return -ETIME;
> +}
> +
>  /*
>   * If the GPU memory is accessed by the CPU while the GPU is not ready
>   * after reset, it can cause harmless corrected RAS events to be logged.
> @@ -275,7 +332,7 @@ nvgrace_gpu_check_device_ready(struct nvgrace_gpu_pci_core_device *nvdev)
>  	if (!__vfio_pci_memory_enabled(vdev))
>  		return -EIO;
>  
> -	ret = nvgrace_gpu_wait_device_ready(vdev->barmap[0]);
> +	ret = nvdev->wait_device_ready(nvdev, vdev->barmap[0]);

A bit of a nit, but it's a little redundant to have both cxl_dvsec and
.wait_device_ready on the object.  It also presents the problem that
each variant has the same prototype, but requires a different
parameter.  Wouldn't a wrapper be a bit cleaner instead?

static inline int nvgrace_gpu_wait_device_ready(struct nvgrace_gpu_pci_core_device *nvdev,
                                                void __iomem *io)
{
	return nvdev->cxl_dvsec ?
		nvgrace_gpu_wait_device_ready_bw_next(nvdev) :
		nvgrace_gpu_wait_device_ready_legacy(io);
}


	ret = nvgrace_gpu_wait_device_ready(nvdev, vdev->barmap[0]);

Thanks,
Alex

>  	if (ret)
>  		return ret;
>  
> @@ -1146,8 +1203,9 @@ static bool nvgrace_gpu_has_mig_hw_bug(struct pci_dev *pdev)
>   * Ensure that the BAR0 region is enabled before accessing the
>   * registers.
>   */
> -static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
> +static int nvgrace_gpu_probe_check_device_ready(struct nvgrace_gpu_pci_core_device *nvdev)
>  {
> +	struct pci_dev *pdev = nvdev->core_device.pdev;
>  	void __iomem *io;
>  	int ret;
>  
> @@ -1165,7 +1223,7 @@ static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
>  		goto iomap_exit;
>  	}
>  
> -	ret = nvgrace_gpu_wait_device_ready(io);
> +	ret = nvdev->wait_device_ready(nvdev, io);
>  
>  	pci_iounmap(pdev, io);
>  iomap_exit:
> @@ -1183,10 +1241,6 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
>  	u64 memphys, memlength;
>  	int ret;
>  
> -	ret = nvgrace_gpu_probe_check_device_ready(pdev);
> -	if (ret)
> -		return ret;
> -
>  	ret = nvgrace_gpu_fetch_memory_property(pdev, &memphys, &memlength);
>  	if (!ret)
>  		ops = &nvgrace_gpu_pci_ops;
> @@ -1198,6 +1252,16 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
>  
>  	dev_set_drvdata(&pdev->dev, &nvdev->core_device);
>  
> +	nvdev->cxl_dvsec = pci_find_dvsec_capability(pdev, PCI_VENDOR_ID_CXL,
> +						     CXL_DEVICE_DVSEC_ID);
> +	nvdev->wait_device_ready = nvdev->cxl_dvsec ?
> +		nvgrace_gpu_wait_device_ready_bw_next :
> +		nvgrace_gpu_wait_device_ready_legacy;
> +
> +	ret = nvgrace_gpu_probe_check_device_ready(nvdev);
> +	if (ret)
> +		goto out_put_vdev;
> +
>  	if (ops == &nvgrace_gpu_pci_ops) {
>  		nvdev->has_mig_hw_bug = nvgrace_gpu_has_mig_hw_bug(pdev);
>  

Re: [PATCH v1 1/1] vfio/nvgrace-gpu: Add Blackwell-Next GPU readiness check via CXL DVSEC

[Ankit Agrawal]

Thanks for the suggestion Alex, comments inline.

>>  
>> +#define CXL_DEVICE_DVSEC_ID          0
>> +#define CXL_DVSEC_RANGE_1_LOW                0x1C
>> +#define CXL_DVSEC_MEMORY_VALID               BIT(0)
>> +#define CXL_DVSEC_MEMORY_ACTIVE              BIT(1)
>
> Looks like pci_regs.h has all these:
>
> #define PCI_DVSEC_CXL_DEVICE                            0
> #define  PCI_DVSEC_CXL_RANGE_SIZE_LOW(i)                (0x1C + (i * 0x10))
> #define   PCI_DVSEC_CXL_MEM_INFO_VALID                  _BITUL(0)
> #define   PCI_DVSEC_CXL_MEM_ACTIVE                      _BITUL(1)
>
> We should search and replace to use the common macros and drop our
> local versions.

Ack, will fix that. Also add the missing MEMORY_ACTIVE_TIMEOUT in the
pci_regs.h from Table 8-13. DVSEC CXL Range 1 Size Low (Offset 1Ch)
This will replace CXL_DVSEC_MEMORY_ACTIVE_TIMEOUT in the patch.

>>  
>> +/*
>> + * Decode the 3-bit Memory_Active_Timeout field from CXL DVSEC Range 1 Low
>
> Is this range 1 or range 0 bits?  The pci_regs.h macro suggests 0.
> Same for the commit log.

This has been a cause of confusion for me. The CXL spec 4.0 has the following
8.1.3.8.2 DVSEC CXL Range 1 Size Low (Offset 1Ch). So that's why I used range 1.

The pci_regs.h does not have any clarification. It does point to the same section.
/* CXL r4.0, 8.1.3: PCIe DVSEC for CXL Device */
#define PCI_DVSEC_CXL_DEVICE                            0
#define  PCI_DVSEC_CXL_CAP                              0xA
#define   PCI_DVSEC_CXL_MEM_CAPABLE                     _BITUL(2)
#define   PCI_DVSEC_CXL_HDM_COUNT                       __GENMASK(5, 4)
#define  PCI_DVSEC_CXL_CTRL                             0xC
#define   PCI_DVSEC_CXL_MEM_ENABLE                      _BITUL(2)
#define  PCI_DVSEC_CXL_RANGE_SIZE_HIGH(i)               (0x18 + (i * 0x10))
#define  PCI_DVSEC_CXL_RANGE_SIZE_LOW(i)                (0x1C + (i * 0x10))

>>        if (!__vfio_pci_memory_enabled(vdev))
>>                return -EIO;
>>  
>> -     ret = nvgrace_gpu_wait_device_ready(vdev->barmap[0]);
>> +     ret = nvdev->wait_device_ready(nvdev, vdev->barmap[0]);
>
> A bit of a nit, but it's a little redundant to have both cxl_dvsec and
> .wait_device_ready on the object.  It also presents the problem that
> each variant has the same prototype, but requires a different
> parameter.  Wouldn't a wrapper be a bit cleaner instead?
>
> static inline int nvgrace_gpu_wait_device_ready(struct nvgrace_gpu_pci_core_device *nvdev,
>                                                 void __iomem *io)
> {
>         return nvdev->cxl_dvsec ?
>                 nvgrace_gpu_wait_device_ready_bw_next(nvdev) :
>                 nvgrace_gpu_wait_device_ready_legacy(io);
> }

Sounds better, will make the change.

>
> Thanks,
> Alex