Re: [PATCH v7 6/6] vfio/nvgrace-gpu: wait for the GPU mem to be ready

[Alex Williamson <alex@shazbot.org>]

On Wed, 26 Nov 2025 05:26:27 +0000
<ankita@nvidia.com> wrote:

> From: Ankit Agrawal <ankita@nvidia.com>
> 
> Speculative prefetches from CPU to GPU memory until the GPU is
> ready after reset can cause harmless corrected RAS events to
> be logged on Grace systems. It is thus preferred that the
> mapping not be re-established until the GPU is ready post reset.
> 
> The GPU readiness can be checked through BAR0 registers similar
> to the checking at the time of device probe.
> 
> It can take several seconds for the GPU to be ready. So it is
> desirable that the time overlaps as much of the VM startup as
> possible to reduce impact on the VM bootup time. The GPU
> readiness state is thus checked on the first fault/huge_fault
> request or read/write access which amortizes the GPU readiness
> time.
> 
> The first fault and read/write checks the GPU state when the
> reset_done flag - which denotes whether the GPU has just been
> reset. The memory_lock is taken across map/access to avoid
> races with GPU reset.
> 
> Cc: Shameer Kolothum <skolothumtho@nvidia.com>
> Cc: Alex Williamson <alex@shazbot.org>
> Cc: Jason Gunthorpe <jgg@ziepe.ca>
> Cc: Vikram Sethi <vsethi@nvidia.com>
> Suggested-by: Alex Williamson <alex@shazbot.org>
> Signed-off-by: Ankit Agrawal <ankita@nvidia.com>
> ---
>  drivers/vfio/pci/nvgrace-gpu/main.c | 81 +++++++++++++++++++++++++----
>  1 file changed, 72 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/vfio/pci/nvgrace-gpu/main.c b/drivers/vfio/pci/nvgrace-gpu/main.c
> index b46984e76be7..3064f8aca858 100644
> --- a/drivers/vfio/pci/nvgrace-gpu/main.c
> +++ b/drivers/vfio/pci/nvgrace-gpu/main.c
> @@ -104,6 +104,17 @@ static int nvgrace_gpu_open_device(struct vfio_device *core_vdev)
>  		mutex_init(&nvdev->remap_lock);
>  	}
>  
> +	/*
> +	 * GPU readiness is checked by reading the BAR0 registers.
> +	 *
> +	 * ioremap BAR0 to ensure that the BAR0 mapping is present before
> +	 * register reads on first fault before establishing any GPU
> +	 * memory mapping.
> +	 */
> +	ret = vfio_pci_core_setup_barmap(vdev, 0);
> +	if (ret)
> +		return ret;
> +
>  	vfio_pci_core_finish_enable(vdev);
>  
>  	return 0;
> @@ -146,6 +157,31 @@ static int nvgrace_gpu_wait_device_ready(void __iomem *io)
>  	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.
> + * Make sure the GPU is ready before establishing the mappings.
> + */
> +static int
> +nvgrace_gpu_check_device_ready(struct nvgrace_gpu_pci_core_device *nvdev)
> +{
> +	struct vfio_pci_core_device *vdev = &nvdev->core_device;
> +	int ret;
> +
> +	lockdep_assert_held_read(&vdev->memory_lock);
> +
> +	if (!nvdev->reset_done)
> +		return 0;
> +
> +	ret = nvgrace_gpu_wait_device_ready(vdev->barmap[0]);
> +	if (ret)
> +		return ret;
> +
> +	nvdev->reset_done = false;
> +
> +	return 0;
> +}
> +
>  static unsigned long addr_to_pgoff(struct vm_area_struct *vma,
>  				   unsigned long addr)
>  {
> @@ -163,13 +199,13 @@ static vm_fault_t nvgrace_gpu_vfio_pci_huge_fault(struct vm_fault *vmf,
>  	struct vfio_pci_core_device *vdev = &nvdev->core_device;
>  	unsigned int index =
>  		vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
> -	vm_fault_t ret = VM_FAULT_SIGBUS;
> +	vm_fault_t ret;
>  	struct mem_region *memregion;
>  	unsigned long pfn, addr;
>  
>  	memregion = nvgrace_gpu_memregion(index, nvdev);
>  	if (!memregion)
> -		return ret;
> +		return VM_FAULT_SIGBUS;
>  
>  	addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
>  	pfn = PHYS_PFN(memregion->memphys) + addr_to_pgoff(vma, addr);
> @@ -179,8 +215,14 @@ static vm_fault_t nvgrace_gpu_vfio_pci_huge_fault(struct vm_fault *vmf,
>  		goto out;
>  	}
>  
> -	scoped_guard(rwsem_read, &vdev->memory_lock)
> +	scoped_guard(rwsem_read, &vdev->memory_lock) {
> +		if (nvgrace_gpu_check_device_ready(nvdev)) {
> +			ret = VM_FAULT_SIGBUS;
> +			goto out;
> +		}
> +
>  		ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
> +	}
>  
>  out:
>  	dev_dbg_ratelimited(&vdev->pdev->dev,
> @@ -593,9 +635,15 @@ nvgrace_gpu_read_mem(struct nvgrace_gpu_pci_core_device *nvdev,
>  	else
>  		mem_count = min(count, memregion->memlength - (size_t)offset);
>  
> -	ret = nvgrace_gpu_map_and_read(nvdev, buf, mem_count, ppos);
> -	if (ret)
> -		return ret;
> +	scoped_guard(rwsem_read, &nvdev->core_device.memory_lock) {
> +		ret = nvgrace_gpu_check_device_ready(nvdev);
> +		if (ret)
> +			return ret;
> +
> +		ret = nvgrace_gpu_map_and_read(nvdev, buf, mem_count, ppos);
> +		if (ret)
> +			return ret;
> +	}
>  
>  	/*
>  	 * Only the device memory present on the hardware is mapped, which may
> @@ -713,9 +761,15 @@ nvgrace_gpu_write_mem(struct nvgrace_gpu_pci_core_device *nvdev,
>  	 */
>  	mem_count = min(count, memregion->memlength - (size_t)offset);
>  
> -	ret = nvgrace_gpu_map_and_write(nvdev, buf, mem_count, ppos);
> -	if (ret)
> -		return ret;
> +	scoped_guard(rwsem_read, &nvdev->core_device.memory_lock) {
> +		ret = nvgrace_gpu_check_device_ready(nvdev);
> +		if (ret)
> +			return ret;
> +
> +		ret = nvgrace_gpu_map_and_write(nvdev, buf, mem_count, ppos);
> +		if (ret)
> +			return ret;
> +	}
>  
>  exitfn:
>  	*ppos += count;
> @@ -1056,6 +1110,15 @@ MODULE_DEVICE_TABLE(pci, nvgrace_gpu_vfio_pci_table);
>   *
>   * The reset_done implementation is triggered on every reset and is used
>   * set the reset_done variable that assists in achieving the serialization.
> + *
> + * The writer memory_lock is held during reset through ioctl and FLR, while
> + * the reader is held in the fault and access code.
> + *
> + * The lock is however not taken during reset called through
> + * vfio_pci_core_enable during open. Whilst a serialization is not
> + * required at that early stage, it still prevents from putting
> + * lockdep_assert_held_write in this function.
> + *

I think the key point is that the readiness test is done holding the
memory_lock read lock and we expect all vfio-pci initiated resets to
hold the memory_lock write lock to avoid races.  However, .reset_done
extends beyond the scope of vfio-pci initiated resets therefore we
cannot assert this behavior.  Thanks,

Alex