Skip to contentSkip to conversation

vfio/nvgrace-gpu: Support huge PFNMAP and wait for GPU ready post reset

PATCH Superseded [email protected] 3 participants 17 messages Nov 26, 2025

Subsystems (4)

v7 of 9 Superseded Latest is v9 →

lore
Share & tools
lore

[email protected]

author maintainer
#0 lore raw

v7 cover vfio/nvgrace-gpu: Support huge PFNMAP and wait for GPU ready post reset

From: Ankit Agrawal <[email protected]>

NVIDIA's Grace based system have large GPU device memory. The device memory is mapped as VM_PFNMAP in the VMM VMA. The nvgrace-gpu module could make use of the huge PFNMAP support added in mm [1].

To achieve this, nvgrace-gpu module is updated to implement huge_fault ops. The implementation establishes mapping according to the order request. Note that if the PFN or the VMA address is unaligned to the order, the mapping fallbacks to the PTE level.

Secondly, it is expected that the mapping not be re-established until the GPU is ready post reset. Presence of the mappings during that time could potentially leads to harmless corrected RAS events to be logged if the CPU attempts to do speculative reads on the GPU memory on the Grace systems.

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 which amortizes the GPU readiness time. The GPU readiness is checked through BAR0 registers as is done at the device probe.

Patch 1 Refactor vfio_pci_mmap_huge_fault and export the code to map at the various levels.

Patch 2 implements support for huge pfnmap.

Patch 3 vfio_pci_core_mmap cleanup.

Patch 4 split the code to check the device readiness.

Patch 5 reset_done handler implementation

Patch 6 Ensures that the GPU is ready before re-establishing the mapping after reset.

Applied over 6.18-rc6.

Changelog:
[v7]
- Code cleanup to use ALIGN functions and move the mapping sanity check to the vfio header file. (Thanks Alex Williamson, Zhi Wang)
- Added comments for reset_done implementation. (Thanks Alex Williamson)
- Collected Reviewed-by. (Thanks Zhi Wang)
- Miscellaneous cleanup (Thanks Alex Williamson, Zhi Wang)

- Updated the vfio_pci_vmf_insert_pfn function to add more common code. (Thanks Shameer Kolothum)
- Added missing userspace offset for pgoff calculation for huge pfnmap. (Thanks Shameer Kolothum)
- Removed alignment for GPU memory. (Thanks Jason Gunthorpe)
- Collected Reviewed-by. (Thanks Shameer Kolothum)
- Miscellaneous cleanup, log messages fixup
(Thanks Alex Williamson, Jason Gunthorpe, Shameer Kolothum)

- Updated gpu_mem_mapped with reset_done flag for clearer semantics. (6/7) (Thanks Alex Williamson)
- Renamed vfio_pci_map_pfn to vfio_pci_vmf_insert_pfn. (2/7) (Thanks Alex Williamson)
- Updated to hold memory_lock across the vmf_insert_pfn and the read/write access of the device. (7/7) (Thanks Alex Williamson)
- Used scoped_guard to simplify critical region. (1/7, 7/7) [v4]
- Implemented reset_done handler to set gpu_mem_mapped flag. Cleaned up FLR detection path (Thanks Alex Williamson)
- Moved the premap check of the device readiness to a new function. Added locking to avoid races. (Thanks Alex Williamson)
- vfio_pci_core_mmap cleanup.
- Added ioremap to BAR0 during open.

- Moved the code for BAR mapping to a separate function.
- Added BAR0 mapping during open. Ensures BAR0 is mapped when registers are checked. (Thanks Alex Williamson, Jason Gunthorpe for suggestion)
- Added check for GPU readiness on nvgrace_gpu_map_device_mem. (Thanks Alex Williamson for the suggestion.

- Fixed build kernel warning
- subject text changes
- Rebased to 6.18-rc6.

Ankit Agrawal (6):
vfio: refactor vfio_pci_mmap_huge_fault function vfio/nvgrace-gpu: Add support for huge pfnmap
vfio: use vfio_pci_core_setup_barmap to map bar in mmap vfio/nvgrace-gpu: split the code to wait for GPU ready vfio/nvgrace-gpu: Inform devmem unmapped after reset vfio/nvgrace-gpu: wait for the GPU mem to be ready

 drivers/vfio/pci/nvgrace-gpu/main.c | 215 ++++++++++++++++++++++------
 drivers/vfio/pci/vfio_pci_core.c    |  69 +++++----
 include/linux/vfio_pci_core.h       |  16 +++
 3 files changed, 222 insertions(+), 78 deletions(-)
Signature · 1 line
2.34.1
author maintainer
#1 lore raw

PATCH v7 1/6 vfio: refactor vfio_pci_mmap_huge_fault function

From: Ankit Agrawal <[email protected]>

Refactor vfio_pci_mmap_huge_fault to take out the implementation to map the VMA to the PTE/PMD/PUD as a separate function.

Export the new function to be used by nvgrace-gpu module.

No functional change is intended.

Cc Shameer Kolothum <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
---
 drivers/vfio/pci/vfio_pci_core.c | 54 +++++++++++++++++---------------
 include/linux/vfio_pci_core.h    | 16 ++++++++++
 2 files changed, 45 insertions(+), 25 deletions(-)
drivers/vfio/pci/vfio_pci_core.c +29 -25
index 7dcf5439dedc..52e3a10d776b 100644
@@ -1640,48 +1640,52 @@ static unsigned long vma_to_pfn(struct vm_area_struct *vma)
 	return (pci_resource_start(vdev->pdev, index) >> PAGE_SHIFT) + pgoff;
 }
 
-static vm_fault_t vfio_pci_mmap_huge_fault(struct vm_fault *vmf,
-					   unsigned int order)
+vm_fault_t vfio_pci_vmf_insert_pfn(struct vfio_pci_core_device *vdev,
+				   struct vm_fault *vmf,
+				   unsigned long pfn,
+				   unsigned int order)
 {
-	struct vm_area_struct *vma = vmf->vma;
-	struct vfio_pci_core_device *vdev = vma->vm_private_data;
-	unsigned long addr = vmf->address & ~((PAGE_SIZE << order) - 1);
-	unsigned long pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
-	unsigned long pfn = vma_to_pfn(vma) + pgoff;
-	vm_fault_t ret = VM_FAULT_SIGBUS;
-
-	if (order && (addr < vma->vm_start ||
-		      addr + (PAGE_SIZE << order) > vma->vm_end ||
-		      pfn & ((1 << order) - 1))) {
-		ret = VM_FAULT_FALLBACK;
-		goto out;
-	}
-
-	down_read(&vdev->memory_lock);
+	lockdep_assert_held_read(&vdev->memory_lock);
 
 	if (vdev->pm_runtime_engaged || !__vfio_pci_memory_enabled(vdev))
-		goto out_unlock;
+		return VM_FAULT_SIGBUS;
 
 	switch (order) {
 	case 0:
-		ret = vmf_insert_pfn(vma, vmf->address, pfn);
-		break;
+		return vmf_insert_pfn(vmf->vma, vmf->address, pfn);
 #ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
 	case PMD_ORDER:
-		ret = vmf_insert_pfn_pmd(vmf, pfn, false);
-		break;
+		return vmf_insert_pfn_pmd(vmf, pfn, false);
 #endif
 #ifdef CONFIG_ARCH_SUPPORTS_PUD_PFNMAP
 	case PUD_ORDER:
-		ret = vmf_insert_pfn_pud(vmf, pfn, false);
+		return vmf_insert_pfn_pud(vmf, pfn, false);
 		break;
 #endif
 	default:
+		return VM_FAULT_FALLBACK;
+	}
+}
+EXPORT_SYMBOL_GPL(vfio_pci_vmf_insert_pfn);
+
+static vm_fault_t vfio_pci_mmap_huge_fault(struct vm_fault *vmf,
+					   unsigned int order)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct vfio_pci_core_device *vdev = vma->vm_private_data;
+	unsigned long addr = vmf->address & ~((PAGE_SIZE << order) - 1);
+	unsigned long pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
+	unsigned long pfn = vma_to_pfn(vma) + pgoff;
+	vm_fault_t ret;
+
+	if (unmappable_for_order(vma, addr, pfn, order)) {
 		ret = VM_FAULT_FALLBACK;
+		goto out;
 	}
 
-out_unlock:
-	up_read(&vdev->memory_lock);
+	scoped_guard(rwsem_read, &vdev->memory_lock)
+		ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
+
 out:
 	dev_dbg_ratelimited(&vdev->pdev->dev,
 			   "%s(,order = %d) BAR %ld page offset 0x%lx: 0x%x\n",
include/linux/vfio_pci_core.h +16 -0
index f541044e42a2..1d457216ce4d 100644
@@ -119,6 +119,9 @@ ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
 		size_t count, loff_t *ppos);
 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
 		size_t count, loff_t *ppos);
+vm_fault_t vfio_pci_vmf_insert_pfn(struct vfio_pci_core_device *vdev,
+				   struct vm_fault *vmf, unsigned long pfn,
+				   unsigned int order);
 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma);
 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count);
 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf);
@@ -161,4 +164,17 @@ VFIO_IOREAD_DECLARATION(32)
 VFIO_IOREAD_DECLARATION(64)
 #endif
 
+static inline bool unmappable_for_order(struct vm_area_struct *vma,
+					unsigned long addr,
+					unsigned long pfn,
+					unsigned int order)
+{
+	if (order && (addr < vma->vm_start ||
+		      addr + (PAGE_SIZE << order) > vma->vm_end ||
+		      !IS_ALIGNED(pfn, 1 << order)))
+		return true;
+
+	return false;
+}
+
 #endif /* VFIO_PCI_CORE_H */
Signature · 1 line
2.34.1
maintainer
#2 lore raw

Re: [PATCH v7 1/6] vfio: refactor vfio_pci_mmap_huge_fault function

Quoted 89 lines · On Wed, 26 Nov 2025 05:26:22 +0000 <[email protected]> wrote:

From: Ankit Agrawal <[email protected]>

Refactor vfio_pci_mmap_huge_fault to take out the implementation to map the VMA to the PTE/PMD/PUD as a separate function.

Export the new function to be used by nvgrace-gpu module.

No functional change is intended.

Cc Shameer Kolothum <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
---
 drivers/vfio/pci/vfio_pci_core.c | 54 +++++++++++++++++---------------
 include/linux/vfio_pci_core.h    | 16 ++++++++++
 2 files changed, 45 insertions(+), 25 deletions(-)
drivers/vfio/pci/vfio_pci_core.c +25 -23
index 7dcf5439dedc..52e3a10d776b 100644
@@ -1640,48 +1640,52 @@ static unsigned long vma_to_pfn(struct vm_area_struct *vma)
 	return (pci_resource_start(vdev->pdev, index) >> PAGE_SHIFT) + pgoff;
 }
 
-static vm_fault_t vfio_pci_mmap_huge_fault(struct vm_fault *vmf,
-					   unsigned int order)
+vm_fault_t vfio_pci_vmf_insert_pfn(struct vfio_pci_core_device *vdev,
+				   struct vm_fault *vmf,
+				   unsigned long pfn,
+				   unsigned int order)
 {
-	struct vm_area_struct *vma = vmf->vma;
-	struct vfio_pci_core_device *vdev = vma->vm_private_data;
-	unsigned long addr = vmf->address & ~((PAGE_SIZE << order) - 1);
-	unsigned long pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
-	unsigned long pfn = vma_to_pfn(vma) + pgoff;
-	vm_fault_t ret = VM_FAULT_SIGBUS;
-
-	if (order && (addr < vma->vm_start ||
-		      addr + (PAGE_SIZE << order) > vma->vm_end ||
-		      pfn & ((1 << order) - 1))) {
-		ret = VM_FAULT_FALLBACK;
-		goto out;
-	}
-
-	down_read(&vdev->memory_lock);
+	lockdep_assert_held_read(&vdev->memory_lock);
 
 	if (vdev->pm_runtime_engaged || !__vfio_pci_memory_enabled(vdev))
-		goto out_unlock;
+		return VM_FAULT_SIGBUS;
 
 	switch (order) {
 	case 0:
-		ret = vmf_insert_pfn(vma, vmf->address, pfn);
-		break;
+		return vmf_insert_pfn(vmf->vma, vmf->address, pfn);
 #ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
 	case PMD_ORDER:
-		ret = vmf_insert_pfn_pmd(vmf, pfn, false);
-		break;
+		return vmf_insert_pfn_pmd(vmf, pfn, false);
 #endif
 #ifdef CONFIG_ARCH_SUPPORTS_PUD_PFNMAP
 	case PUD_ORDER:
-		ret = vmf_insert_pfn_pud(vmf, pfn, false);
+		return vmf_insert_pfn_pud(vmf, pfn, false);
 		break;
 #endif
 	default:
+		return VM_FAULT_FALLBACK;
+	}
+}
+EXPORT_SYMBOL_GPL(vfio_pci_vmf_insert_pfn);
+
+static vm_fault_t vfio_pci_mmap_huge_fault(struct vm_fault *vmf,
+					   unsigned int order)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct vfio_pci_core_device *vdev = vma->vm_private_data;
+	unsigned long addr = vmf->address & ~((PAGE_SIZE << order) - 1);
+	unsigned long pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
+	unsigned long pfn = vma_to_pfn(vma) + pgoff;
+	vm_fault_t ret;
+
+	if (unmappable_for_order(vma, addr, pfn, order)) {

drivers/vfio/pci/vfio_pci_core.c static unsigned long vma_to_pfn(struct vm_area_struct *vma) · line 1689

Quoted 10 lines
drivers/vfio/pci/vfio_pci_core.c +4 -2
 		ret = VM_FAULT_FALLBACK;
+		goto out;
 	}
 
-out_unlock:
-	up_read(&vdev->memory_lock);
+	scoped_guard(rwsem_read, &vdev->memory_lock)
+		ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
+
 out:

We really don't need a goto to jump over this tiny section of code. With the naming/polarity change below this can just be:

	vm_fault_t ret = VM_FAULT_FALLBACK;

	if (is_aligned_for_order(vma, addr, pfn, order)) {
		scoped_guard(rwsem_read, &vdev->memory_lock)
			ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
	}
Quoted 22 lines
drivers/vfio/pci/vfio_pci_core.c +0 -0
 	dev_dbg_ratelimited(&vdev->pdev->dev,
 			   "%s(,order = %d) BAR %ld page offset 0x%lx: 0x%x\n",
include/linux/vfio_pci_core.h +5 -0
index f541044e42a2..1d457216ce4d 100644
@@ -119,6 +119,9 @@ ssize_t vfio_pci_core_read(struct vfio_device *core_vdev, char __user *buf,
 		size_t count, loff_t *ppos);
 ssize_t vfio_pci_core_write(struct vfio_device *core_vdev, const char __user *buf,
 		size_t count, loff_t *ppos);
+vm_fault_t vfio_pci_vmf_insert_pfn(struct vfio_pci_core_device *vdev,
+				   struct vm_fault *vmf, unsigned long pfn,
+				   unsigned int order);
 int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma);
 void vfio_pci_core_request(struct vfio_device *core_vdev, unsigned int count);
 int vfio_pci_core_match(struct vfio_device *core_vdev, char *buf);
@@ -161,4 +164,17 @@ VFIO_IOREAD_DECLARATION(32)
 VFIO_IOREAD_DECLARATION(64)
 #endif
 
+static inline bool unmappable_for_order(struct vm_area_struct *vma,
+					unsigned long addr,

include/linux/vfio_pci_core.h VFIO_IOREAD_DECLARATION(32) · line 178

Quoted 10 lines
include/linux/vfio_pci_core.h +10 -0
+					unsigned long pfn,
+					unsigned int order)
+{
+	if (order && (addr < vma->vm_start ||
+		      addr + (PAGE_SIZE << order) > vma->vm_end ||
+		      !IS_ALIGNED(pfn, 1 << order)))
+		return true;
+
+	return false;
+}

Change polarity and rename to is_aligned_for_order()? No need for branched return.

	return !(order && (addr < vma->vm_start ||
			   addr + (PAGE_SIZE << order) > vma->vm_end ||
			   !IS_ALIGNED(pfn, 1 << order)));

Describe this change in the commit log. Thanks,

Alex

author maintainer
#3 lore raw

PATCH v7 2/6 vfio/nvgrace-gpu: Add support for huge pfnmap

In reply to [email protected]

From: Ankit Agrawal <[email protected]>

NVIDIA's Grace based systems have large device memory. The device memory is mapped as VM_PFNMAP in the VMM VMA. The nvgrace-gpu module could make use of the huge PFNMAP support added in mm [1].

To make use of the huge pfnmap support, fault/huge_fault ops based mapping mechanism needs to be implemented. Currently nvgrace-gpu module relies on remap_pfn_range to do the mapping during VM bootup. Replace it to instead rely on fault and use vfio_pci_vmf_insert_pfn to setup the mapping.

Moreover to enable huge pfnmap, nvgrace-gpu module is updated by adding huge_fault ops implementation. The implementation establishes mapping according to the order request. Note that if the PFN or the VMA address is unaligned to the order, the mapping fallbacks to the PTE level.

Cc Shameer Kolothum <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 85 +++++++++++++++++++++--------
 1 file changed, 63 insertions(+), 22 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +63 -22
index e346392b72f6..ac9551b9e4b6 100644
@@ -130,6 +130,63 @@ static void nvgrace_gpu_close_device(struct vfio_device *core_vdev)
 	vfio_pci_core_close_device(core_vdev);
 }
 
+static unsigned long addr_to_pgoff(struct vm_area_struct *vma,
+				   unsigned long addr)
+{
+	u64 pgoff = vma->vm_pgoff &
+		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
+
+	return ((addr - vma->vm_start) >> PAGE_SHIFT) + pgoff;
+}
+
+static vm_fault_t nvgrace_gpu_vfio_pci_huge_fault(struct vm_fault *vmf,
+						  unsigned int order)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct nvgrace_gpu_pci_core_device *nvdev = vma->vm_private_data;
+	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;
+	struct mem_region *memregion;
+	unsigned long pfn, addr;
+
+	memregion = nvgrace_gpu_memregion(index, nvdev);
+	if (!memregion)
+		return ret;
+
+	addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
+	pfn = PHYS_PFN(memregion->memphys) + addr_to_pgoff(vma, addr);
+
+	if (unmappable_for_order(vma, addr, pfn, order)) {
+		ret = VM_FAULT_FALLBACK;
+		goto out;
+	}
+
+	scoped_guard(rwsem_read, &vdev->memory_lock)
+		ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
+
+out:
+	dev_dbg_ratelimited(&vdev->pdev->dev,
+			    "%s order = %d pfn 0x%lx: 0x%x\n",
+			    __func__, order, pfn,
+			    (unsigned int)ret);
+
+	return ret;
+}
+
+static vm_fault_t nvgrace_gpu_vfio_pci_fault(struct vm_fault *vmf)
+{
+	return nvgrace_gpu_vfio_pci_huge_fault(vmf, 0);
+}
+
+static const struct vm_operations_struct nvgrace_gpu_vfio_pci_mmap_ops = {
+	.fault = nvgrace_gpu_vfio_pci_fault,
+#ifdef CONFIG_ARCH_SUPPORTS_HUGE_PFNMAP
+	.huge_fault = nvgrace_gpu_vfio_pci_huge_fault,
+#endif
+};
+
 static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 			    struct vm_area_struct *vma)
 {
@@ -137,10 +194,8 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		container_of(core_vdev, struct nvgrace_gpu_pci_core_device,
 			     core_device.vdev);
 	struct mem_region *memregion;
-	unsigned long start_pfn;
 	u64 req_len, pgoff, end;
 	unsigned int index;
-	int ret = 0;
 
 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
 
@@ -157,17 +212,18 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
 
 	if (check_sub_overflow(vma->vm_end, vma->vm_start, &req_len) ||
-	    check_add_overflow(PHYS_PFN(memregion->memphys), pgoff, &start_pfn) ||
 	    check_add_overflow(PFN_PHYS(pgoff), req_len, &end))
 		return -EOVERFLOW;
 
 	/*
-	 * Check that the mapping request does not go beyond available device
-	 * memory size
+	 * Check that the mapping request does not go beyond the exposed
+	 * device memory size.
 	 */
 	if (end > memregion->memlength)
 		return -EINVAL;
 
+	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
+
 	/*
 	 * The carved out region of the device memory needs the NORMAL_NC
 	 * property. Communicate as such to the hypervisor.
@@ -184,23 +240,8 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 	}
 
-	/*
-	 * Perform a PFN map to the memory and back the device BAR by the
-	 * GPU memory.
-	 *
-	 * The available GPU memory size may not be power-of-2 aligned. The
-	 * remainder is only backed by vfio_device_ops read/write handlers.
-	 *
-	 * During device reset, the GPU is safely disconnected to the CPU
-	 * and access to the BAR will be immediately returned preventing
-	 * machine check.
-	 */
-	ret = remap_pfn_range(vma, vma->vm_start, start_pfn,
-			      req_len, vma->vm_page_prot);
-	if (ret)
-		return ret;
-
-	vma->vm_pgoff = start_pfn;
+	vma->vm_ops = &nvgrace_gpu_vfio_pci_mmap_ops;
+	vma->vm_private_data = nvdev;
 
 	return 0;
 }
Signature · 1 line
2.34.1
reviewer Reviewed-by on v7 2/6
#4 lore raw

RE: [PATCH v7 2/6] vfio/nvgrace-gpu: Add support for huge pfnmap

Quoted 39 lines

-----Original Message-----
From: Ankit Agrawal <[email protected]>
Sent: 26 November 2025 05:26
To: Ankit Agrawal <[email protected]>; [email protected]; Yishai Hadas <[email protected]>; Shameer Kolothum <[email protected]>; [email protected]; [email protected]; Aniket Agashe <[email protected]>; Vikram Sethi <[email protected]>; Matt Ochs <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; Avihai Horon <[email protected]>; [email protected]; [email protected]; [email protected]; Alistair Popple <[email protected]>; [email protected]; linux- [email protected]; Neo Jia <[email protected]>; Kirti Wankhede <[email protected]>; Tarun Gupta (SW-GPU) <[email protected]>; Zhi Wang <[email protected]>; Dan Williams <[email protected]>; Dheeraj Nigam <[email protected]>; Krishnakant Jaju <[email protected]>
Subject: [PATCH v7 2/6] vfio/nvgrace-gpu: Add support for huge pfnmap

From: Ankit Agrawal <[email protected]>

NVIDIA's Grace based systems have large device memory. The device memory is mapped as VM_PFNMAP in the VMM VMA. The nvgrace-gpu module could make use of the huge PFNMAP support added in mm [1].

To make use of the huge pfnmap support, fault/huge_fault ops based mapping mechanism needs to be implemented. Currently nvgrace-gpu module relies on remap_pfn_range to do the mapping during VM bootup. Replace it to instead rely on fault and use vfio_pci_vmf_insert_pfn to setup the mapping.

Moreover to enable huge pfnmap, nvgrace-gpu module is updated by adding huge_fault ops implementation. The implementation establishes mapping according to the order request. Note that if the PFN or the VMA address is unaligned to the order, the mapping fallbacks to the PTE level.

Quoted 8 lines

[email protected]/ [1]

Cc Shameer Kolothum <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
maintainer
#5 lore raw

Re: [PATCH v7 2/6] vfio/nvgrace-gpu: Add support for huge pfnmap

In reply to [email protected]
Quoted 65 lines · On Wed, 26 Nov 2025 05:26:23 +0000 <[email protected]> wrote:

From: Ankit Agrawal <[email protected]>

NVIDIA's Grace based systems have large device memory. The device memory is mapped as VM_PFNMAP in the VMM VMA. The nvgrace-gpu module could make use of the huge PFNMAP support added in mm [1].

To make use of the huge pfnmap support, fault/huge_fault ops based mapping mechanism needs to be implemented. Currently nvgrace-gpu module relies on remap_pfn_range to do the mapping during VM bootup. Replace it to instead rely on fault and use vfio_pci_vmf_insert_pfn to setup the mapping.

Moreover to enable huge pfnmap, nvgrace-gpu module is updated by adding huge_fault ops implementation. The implementation establishes mapping according to the order request. Note that if the PFN or the VMA address is unaligned to the order, the mapping fallbacks to the PTE level.

Cc Shameer Kolothum <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 85 +++++++++++++++++++++--------
 1 file changed, 63 insertions(+), 22 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +27 -0
index e346392b72f6..ac9551b9e4b6 100644
@@ -130,6 +130,63 @@ static void nvgrace_gpu_close_device(struct vfio_device *core_vdev)
 	vfio_pci_core_close_device(core_vdev);
 }
 
+static unsigned long addr_to_pgoff(struct vm_area_struct *vma,
+				   unsigned long addr)
+{
+	u64 pgoff = vma->vm_pgoff &
+		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
+
+	return ((addr - vma->vm_start) >> PAGE_SHIFT) + pgoff;
+}
+
+static vm_fault_t nvgrace_gpu_vfio_pci_huge_fault(struct vm_fault *vmf,
+						  unsigned int order)
+{
+	struct vm_area_struct *vma = vmf->vma;
+	struct nvgrace_gpu_pci_core_device *nvdev = vma->vm_private_data;
+	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;
+	struct mem_region *memregion;
+	unsigned long pfn, addr;
+
+	memregion = nvgrace_gpu_memregion(index, nvdev);
+	if (!memregion)
+		return ret;
+
+	addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
+	pfn = PHYS_PFN(memregion->memphys) + addr_to_pgoff(vma, addr);

drivers/vfio/pci/nvgrace-gpu/main.c static void nvgrace_gpu_close_device(struct vfio_device *core_vdev) · line 169

Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c +10 -0
+
+	if (unmappable_for_order(vma, addr, pfn, order)) {
+		ret = VM_FAULT_FALLBACK;
+		goto out;
+	}
+
+	scoped_guard(rwsem_read, &vdev->memory_lock)
+		ret = vfio_pci_vmf_insert_pfn(vdev, vmf, pfn, order);
+
+out:

Adopt same code flow as suggested in 1/, return SIGBUS directly for !memregion, initialize ret to FALLBACK. Thanks,

Alex

Quoted 82 lines
drivers/vfio/pci/nvgrace-gpu/main.c +26 -22
+	dev_dbg_ratelimited(&vdev->pdev->dev,
+			    "%s order = %d pfn 0x%lx: 0x%x\n",
+			    __func__, order, pfn,
+			    (unsigned int)ret);
+
+	return ret;
+}
+
+static vm_fault_t nvgrace_gpu_vfio_pci_fault(struct vm_fault *vmf)
+{
+	return nvgrace_gpu_vfio_pci_huge_fault(vmf, 0);
+}
+
+static const struct vm_operations_struct nvgrace_gpu_vfio_pci_mmap_ops = {
+	.fault = nvgrace_gpu_vfio_pci_fault,
+#ifdef CONFIG_ARCH_SUPPORTS_HUGE_PFNMAP
+	.huge_fault = nvgrace_gpu_vfio_pci_huge_fault,
+#endif
+};
+
 static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 			    struct vm_area_struct *vma)
 {
@@ -137,10 +194,8 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		container_of(core_vdev, struct nvgrace_gpu_pci_core_device,
 			     core_device.vdev);
 	struct mem_region *memregion;
-	unsigned long start_pfn;
 	u64 req_len, pgoff, end;
 	unsigned int index;
-	int ret = 0;
 
 	index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT);
 
@@ -157,17 +212,18 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		((1U << (VFIO_PCI_OFFSET_SHIFT - PAGE_SHIFT)) - 1);
 
 	if (check_sub_overflow(vma->vm_end, vma->vm_start, &req_len) ||
-	    check_add_overflow(PHYS_PFN(memregion->memphys), pgoff, &start_pfn) ||
 	    check_add_overflow(PFN_PHYS(pgoff), req_len, &end))
 		return -EOVERFLOW;
 
 	/*
-	 * Check that the mapping request does not go beyond available device
-	 * memory size
+	 * Check that the mapping request does not go beyond the exposed
+	 * device memory size.
 	 */
 	if (end > memregion->memlength)
 		return -EINVAL;
 
+	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
+
 	/*
 	 * The carved out region of the device memory needs the NORMAL_NC
 	 * property. Communicate as such to the hypervisor.
@@ -184,23 +240,8 @@ static int nvgrace_gpu_mmap(struct vfio_device *core_vdev,
 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 	}
 
-	/*
-	 * Perform a PFN map to the memory and back the device BAR by the
-	 * GPU memory.
-	 *
-	 * The available GPU memory size may not be power-of-2 aligned. The
-	 * remainder is only backed by vfio_device_ops read/write handlers.
-	 *
-	 * During device reset, the GPU is safely disconnected to the CPU
-	 * and access to the BAR will be immediately returned preventing
-	 * machine check.
-	 */
-	ret = remap_pfn_range(vma, vma->vm_start, start_pfn,
-			      req_len, vma->vm_page_prot);
-	if (ret)
-		return ret;
-
-	vma->vm_pgoff = start_pfn;
+	vma->vm_ops = &nvgrace_gpu_vfio_pci_mmap_ops;
+	vma->vm_private_data = nvdev;
 
 	return 0;
 }
author maintainer
#6 lore raw

PATCH v7 3/6 vfio: use vfio_pci_core_setup_barmap to map bar in mmap

In reply to [email protected]

From: Ankit Agrawal <[email protected]>

Remove code duplication in vfio_pci_core_mmap by calling vfio_pci_core_setup_barmap to perform the bar mapping.

No functional change is intended.

---
 drivers/vfio/pci/vfio_pci_core.c | 15 +++------------
 1 file changed, 3 insertions(+), 12 deletions(-)
drivers/vfio/pci/vfio_pci_core.c +3 -12
index 52e3a10d776b..020bf3aa22e9 100644
@@ -1753,18 +1753,9 @@ int vfio_pci_core_mmap(struct vfio_device *core_vdev, struct vm_area_struct *vma
 	 * Even though we don't make use of the barmap for the mmap,
 	 * we need to request the region and the barmap tracks that.
 	 */
-	if (!vdev->barmap[index]) {
-		ret = pci_request_selected_regions(pdev,
-						   1 << index, "vfio-pci");
-		if (ret)
-			return ret;
-
-		vdev->barmap[index] = pci_iomap(pdev, index, 0);
-		if (!vdev->barmap[index]) {
-			pci_release_selected_regions(pdev, 1 << index);
-			return -ENOMEM;
-		}
-	}
+	ret = vfio_pci_core_setup_barmap(vdev, index);
+	if (ret)
+		return ret;
 
 	vma->vm_private_data = vdev;
 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
Signature · 1 line
2.34.1
author maintainer
#7 lore raw

PATCH v7 4/6 vfio/nvgrace-gpu: split the code to wait for GPU ready

In reply to [email protected]

From: Ankit Agrawal <[email protected]>

Split the function that check for the GPU device being ready on the probe.

Move the code to wait for the GPU to be ready through BAR0 register reads to a separate function. This would help reuse the code.

This also fixes a bug where the return status in case of timeout gets overridden by return from pci_enable_device. With the fix, a timeout generate an error as initially intended.

Fixes d85f69d520e6 ("vfio/nvgrace-gpu: Check the HBM training and C2C link status")
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 29 +++++++++++++++++------------
 1 file changed, 17 insertions(+), 12 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +17 -12
index ac9551b9e4b6..f691deb8e43c 100644
@@ -130,6 +130,20 @@ 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)
+{
+	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
+
+	do {
+		if ((ioread32(io + C2C_LINK_BAR0_OFFSET) == STATUS_READY) &&
+		    (ioread32(io + HBM_TRAINING_BAR0_OFFSET) == STATUS_READY))
+			return 0;
+		msleep(POLL_QUANTUM_MS);
+	} while (!time_after(jiffies, timeout));
+
+	return -ETIME;
+}
+
 static unsigned long addr_to_pgoff(struct vm_area_struct *vma,
 				   unsigned long addr)
 {
@@ -934,9 +948,8 @@ 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_wait_device_ready(struct pci_dev *pdev)
+static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
 {
-	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
 	void __iomem *io;
 	int ret = -ETIME;
 
@@ -954,16 +967,8 @@ static int nvgrace_gpu_wait_device_ready(struct pci_dev *pdev)
 		goto iomap_exit;
 	}
 
-	do {
-		if ((ioread32(io + C2C_LINK_BAR0_OFFSET) == STATUS_READY) &&
-		    (ioread32(io + HBM_TRAINING_BAR0_OFFSET) == STATUS_READY)) {
-			ret = 0;
-			goto reg_check_exit;
-		}
-		msleep(POLL_QUANTUM_MS);
-	} while (!time_after(jiffies, timeout));
+	ret = nvgrace_gpu_wait_device_ready(io);
 
-reg_check_exit:
 	pci_iounmap(pdev, io);
 iomap_exit:
 	pci_release_selected_regions(pdev, 1 << 0);
@@ -980,7 +985,7 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
 	u64 memphys, memlength;
 	int ret;
 
-	ret = nvgrace_gpu_wait_device_ready(pdev);
+	ret = nvgrace_gpu_probe_check_device_ready(pdev);
 	if (ret)
 		return ret;
 
Signature · 1 line
2.34.1
maintainer
#8 lore raw

Re: [PATCH v7 4/6] vfio/nvgrace-gpu: split the code to wait for GPU ready

Quoted 47 lines · On Wed, 26 Nov 2025 05:26:25 +0000 <[email protected]> wrote:

From: Ankit Agrawal <[email protected]>

Split the function that check for the GPU device being ready on the probe.

Move the code to wait for the GPU to be ready through BAR0 register reads to a separate function. This would help reuse the code.

This also fixes a bug where the return status in case of timeout gets overridden by return from pci_enable_device. With the fix, a timeout generate an error as initially intended.

Fixes d85f69d520e6 ("vfio/nvgrace-gpu: Check the HBM training and C2C link status")
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 29 +++++++++++++++++------------
 1 file changed, 17 insertions(+), 12 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +14 -0
index ac9551b9e4b6..f691deb8e43c 100644
@@ -130,6 +130,20 @@ 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)
+{
+	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
+
+	do {
+		if ((ioread32(io + C2C_LINK_BAR0_OFFSET) == STATUS_READY) &&
+		    (ioread32(io + HBM_TRAINING_BAR0_OFFSET) == STATUS_READY))
+			return 0;
+		msleep(POLL_QUANTUM_MS);
+	} while (!time_after(jiffies, timeout));
+
+	return -ETIME;
+}
+
 static unsigned long addr_to_pgoff(struct vm_area_struct *vma,
 				   unsigned long addr)
 {

drivers/vfio/pci/nvgrace-gpu/main.c static bool nvgrace_gpu_has_mig_hw_bug(struct pci_dev *pdev) · line 954

Quoted 9 lines
drivers/vfio/pci/nvgrace-gpu/main.c +1 -2
@@ -934,9 +948,8 @@ 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_wait_device_ready(struct pci_dev *pdev)
+static int nvgrace_gpu_probe_check_device_ready(struct pci_dev *pdev)
 {
-	unsigned long timeout = jiffies + msecs_to_jiffies(POLL_TIMEOUT_MS);
 	void __iomem *io;
 	int ret = -ETIME;

ret initialization is not needed, it's immediately clobbered. Thanks,

Alex

Quoted 28 lines
drivers/vfio/pci/nvgrace-gpu/main.c +2 -10
@@ -954,16 +967,8 @@ static int nvgrace_gpu_wait_device_ready(struct pci_dev *pdev)
 		goto iomap_exit;
 	}
 
-	do {
-		if ((ioread32(io + C2C_LINK_BAR0_OFFSET) == STATUS_READY) &&
-		    (ioread32(io + HBM_TRAINING_BAR0_OFFSET) == STATUS_READY)) {
-			ret = 0;
-			goto reg_check_exit;
-		}
-		msleep(POLL_QUANTUM_MS);
-	} while (!time_after(jiffies, timeout));
+	ret = nvgrace_gpu_wait_device_ready(io);
 
-reg_check_exit:
 	pci_iounmap(pdev, io);
 iomap_exit:
 	pci_release_selected_regions(pdev, 1 << 0);
@@ -980,7 +985,7 @@ static int nvgrace_gpu_probe(struct pci_dev *pdev,
 	u64 memphys, memlength;
 	int ret;
 
-	ret = nvgrace_gpu_wait_device_ready(pdev);
+	ret = nvgrace_gpu_probe_check_device_ready(pdev);
 	if (ret)
 		return ret;
 
author maintainer
#9 lore raw

PATCH v7 5/6 vfio/nvgrace-gpu: Inform devmem unmapped after reset

In reply to [email protected]

From: Ankit Agrawal <[email protected]>

Introduce a new flag reset_done to notify that the GPU has just been reset and the mapping to the GPU memory is zapped.

Implement the reset_done handler to set this new variable. It will be used later in the patches to wait for the GPU memory to be ready before doing any mapping or access.

---
 drivers/vfio/pci/nvgrace-gpu/main.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
drivers/vfio/pci/nvgrace-gpu/main.c +25 -1
index f691deb8e43c..b46984e76be7 100644
@@ -58,6 +58,8 @@ struct nvgrace_gpu_pci_core_device {
 	/* Lock to control device memory kernel mapping */
 	struct mutex remap_lock;
 	bool has_mig_hw_bug;
+	/* GPU has just been reset */
+	bool reset_done;
 };
 
 static void nvgrace_gpu_init_fake_bar_emu_regs(struct vfio_device *core_vdev)
@@ -1048,12 +1050,34 @@ static const struct pci_device_id nvgrace_gpu_vfio_pci_table[] = {
 
 MODULE_DEVICE_TABLE(pci, nvgrace_gpu_vfio_pci_table);
 
+/*
+ * The GPU reset is required to be serialized against the *first* mapping
+ * faults and read/writes accesses to prevent potential RAS events logging.
+ *
+ * The reset_done implementation is triggered on every reset and is used
+ * set the reset_done variable that assists in achieving the serialization.
+ */
+static void nvgrace_gpu_vfio_pci_reset_done(struct pci_dev *pdev)
+{
+	struct vfio_pci_core_device *core_device = dev_get_drvdata(&pdev->dev);
+	struct nvgrace_gpu_pci_core_device *nvdev =
+		container_of(core_device, struct nvgrace_gpu_pci_core_device,
+			     core_device);
+
+	nvdev->reset_done = true;
+}
+
+static const struct pci_error_handlers nvgrace_gpu_vfio_pci_err_handlers = {
+	.reset_done = nvgrace_gpu_vfio_pci_reset_done,
+	.error_detected = vfio_pci_core_aer_err_detected,
+};
+
 static struct pci_driver nvgrace_gpu_vfio_pci_driver = {
 	.name = KBUILD_MODNAME,
 	.id_table = nvgrace_gpu_vfio_pci_table,
 	.probe = nvgrace_gpu_probe,
 	.remove = nvgrace_gpu_remove,
-	.err_handler = &vfio_pci_core_err_handlers,
+	.err_handler = &nvgrace_gpu_vfio_pci_err_handlers,
 	.driver_managed_dma = true,
 };
 
Signature · 1 line
2.34.1
reviewer Reviewed-by on v7 5/6
#10 lore raw

RE: [PATCH v7 5/6] vfio/nvgrace-gpu: Inform devmem unmapped after reset

Quoted 26 lines

-----Original Message-----
From: Ankit Agrawal <[email protected]>
Sent: 26 November 2025 05:26
To: Ankit Agrawal <[email protected]>; [email protected]; Yishai Hadas <[email protected]>; Shameer Kolothum <[email protected]>; [email protected]; [email protected]; Aniket Agashe <[email protected]>; Vikram Sethi <[email protected]>; Matt Ochs <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; Avihai Horon <[email protected]>; [email protected]; [email protected]; [email protected]; Alistair Popple <[email protected]>; [email protected]; linux- [email protected]; Neo Jia <[email protected]>; Kirti Wankhede <[email protected]>; Tarun Gupta (SW-GPU) <[email protected]>; Zhi Wang <[email protected]>; Dan Williams <[email protected]>; Dheeraj Nigam <[email protected]>; Krishnakant Jaju <[email protected]>
Subject: [PATCH v7 5/6] vfio/nvgrace-gpu: Inform devmem unmapped after reset

From: Ankit Agrawal <[email protected]>

Introduce a new flag reset_done to notify that the GPU has just been reset and the mapping to the GPU memory is zapped.

Quoted 7 lines

Implement the reset_done handler to set this new variable. It will be used later in the patches to wait for the GPU memory to be ready before doing any mapping or access.

maintainer
#11 lore raw

Re: [PATCH v7 5/6] vfio/nvgrace-gpu: Inform devmem unmapped after reset

In reply to [email protected]
Quoted 30 lines · On Wed, 26 Nov 2025 05:26:26 +0000 <[email protected]> wrote:

From: Ankit Agrawal <[email protected]>

Introduce a new flag reset_done to notify that the GPU has just been reset and the mapping to the GPU memory is zapped.

Implement the reset_done handler to set this new variable. It will be used later in the patches to wait for the GPU memory to be ready before doing any mapping or access.

---
 drivers/vfio/pci/nvgrace-gpu/main.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)
drivers/vfio/pci/nvgrace-gpu/main.c +2 -0
index f691deb8e43c..b46984e76be7 100644
@@ -58,6 +58,8 @@ struct nvgrace_gpu_pci_core_device {
 	/* Lock to control device memory kernel mapping */
 	struct mutex remap_lock;
 	bool has_mig_hw_bug;
+	/* GPU has just been reset */
+	bool reset_done;
 };
 
 static void nvgrace_gpu_init_fake_bar_emu_regs(struct vfio_device *core_vdev)

drivers/vfio/pci/nvgrace-gpu/main.c static const struct pci_device_id nvgrace_gpu_vfio_pci_table[] = { · line 1058

Quoted 9 lines
drivers/vfio/pci/nvgrace-gpu/main.c +6 -0
@@ -1048,12 +1050,34 @@ static const struct pci_device_id nvgrace_gpu_vfio_pci_table[] = {
 
 MODULE_DEVICE_TABLE(pci, nvgrace_gpu_vfio_pci_table);
 
+/*
+ * The GPU reset is required to be serialized against the *first* mapping
+ * faults and read/writes accesses to prevent potential RAS events logging.
+ *
+ * The reset_done implementation is triggered on every reset and is used
+ * set the reset_done variable that assists in achieving the serialization.

I can't parse this last sentence, "used _to_ set"?. Maybe something as simple as:

/*
* First fault or access after a reset needs to poll device readiness,
* flag that a reset has occurred.
*/

Thanks,
Alex

Quoted 26 lines
drivers/vfio/pci/nvgrace-gpu/main.c +17 -1
+ */
+static void nvgrace_gpu_vfio_pci_reset_done(struct pci_dev *pdev)
+{
+	struct vfio_pci_core_device *core_device = dev_get_drvdata(&pdev->dev);
+	struct nvgrace_gpu_pci_core_device *nvdev =
+		container_of(core_device, struct nvgrace_gpu_pci_core_device,
+			     core_device);
+
+	nvdev->reset_done = true;
+}
+
+static const struct pci_error_handlers nvgrace_gpu_vfio_pci_err_handlers = {
+	.reset_done = nvgrace_gpu_vfio_pci_reset_done,
+	.error_detected = vfio_pci_core_aer_err_detected,
+};
+
 static struct pci_driver nvgrace_gpu_vfio_pci_driver = {
 	.name = KBUILD_MODNAME,
 	.id_table = nvgrace_gpu_vfio_pci_table,
 	.probe = nvgrace_gpu_probe,
 	.remove = nvgrace_gpu_remove,
-	.err_handler = &vfio_pci_core_err_handlers,
+	.err_handler = &nvgrace_gpu_vfio_pci_err_handlers,
 	.driver_managed_dma = true,
 };
 
author maintainer
#12 lore raw

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

In reply to [email protected]

From: Ankit Agrawal <[email protected]>

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 <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 81 +++++++++++++++++++++++++----
 1 file changed, 72 insertions(+), 9 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +72 -9
index b46984e76be7..3064f8aca858 100644
@@ -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.
+ *
  */
 static void nvgrace_gpu_vfio_pci_reset_done(struct pci_dev *pdev)
 {
Signature · 1 line
2.34.1
reviewer Reviewed-by on v7 6/6
#13 lore raw

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

Quoted 63 lines

-----Original Message-----
From: Ankit Agrawal <[email protected]>
Sent: 26 November 2025 05:26
To: Ankit Agrawal <[email protected]>; [email protected]; Yishai Hadas <[email protected]>; Shameer Kolothum <[email protected]>; [email protected]; [email protected]; Aniket Agashe <[email protected]>; Vikram Sethi <[email protected]>; Matt Ochs <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; Avihai Horon <[email protected]>; [email protected]; [email protected]; [email protected]; Alistair Popple <[email protected]>; [email protected]; linux- [email protected]; Neo Jia <[email protected]>; Kirti Wankhede <[email protected]>; Tarun Gupta (SW-GPU) <[email protected]>; Zhi Wang <[email protected]>; Dan Williams <[email protected]>; Dheeraj Nigam <[email protected]>; Krishnakant Jaju <[email protected]>
Subject: [PATCH v7 6/6] vfio/nvgrace-gpu: wait for the GPU mem to be ready

From: Ankit Agrawal <[email protected]>

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 <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 81 +++++++++++++++++++++++++---
(quoted diff) +0 -1
-
 1 file changed, 72 insertions(+), 9 deletions(-)
drivers/vfio/pci/nvgrace- renamed from drivers/vfio/pci/nvgrace-gpu/main.c +0 -0

gpu/main.c
index b46984e76be7..3064f8aca858 100644

drivers/vfio/pci/nvgrace-gpu/main.c +0 -0
@@ -104,6 +104,17 @@ static int nvgrace_gpu_open_device(struct
 vfio_device *core_vdev)
 		mutex_init(&nvdev->remap_lock);
 	}
 

drivers/vfio/pci/nvgrace-gpu/main.c static int nvgrace_gpu_open_device(struct · line 117

Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c +10 -0
+	/*
+	 * 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;

Should make sure vfio_pci_core_disable() is called on err path above.

With that,

author maintainer
#14 lore raw

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

Quoted 5 lines
Quoted 3 lines

+     ret = vfio_pci_core_setup_barmap(vdev, 0);
+     if (ret)
+             return ret;

Should make sure vfio_pci_core_disable() is called on err path above.

Thanks Shameer for the review! Yeah, we should not have that if (ret) check and let the flow fall to the vfio_pci_core_disable().

Quoted 2 lines

With that,

author maintainer
#15 lore raw

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

Quoted 5 lines
Quoted 3 lines

+     ret = vfio_pci_core_setup_barmap(vdev, 0);
+     if (ret)
+             return ret;

Should make sure vfio_pci_core_disable() is called on err path above.

Yes, will have to do the following.

@@ -112,8 +112,10 @@ static int nvgrace_gpu_open_device(struct vfio_device *core_vdev)

         * memory mapping.
         */
        ret = vfio_pci_core_setup_barmap(vdev, 0);
-       if (ret)
+       if (ret) {
+               vfio_pci_core_disable(vdev);
                return ret;

+ }

        vfio_pci_core_finish_enable(vdev);
Quoted 2 lines

With that,

maintainer
#16 lore raw

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

In reply to [email protected]
Quoted 159 lines · On Wed, 26 Nov 2025 05:26:27 +0000 <[email protected]> wrote:

From: Ankit Agrawal <[email protected]>

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 <[email protected]>
Cc Alex Williamson <[email protected]>
Cc Jason Gunthorpe <[email protected]>
Cc Vikram Sethi <[email protected]>
---
 drivers/vfio/pci/nvgrace-gpu/main.c | 81 +++++++++++++++++++++++++----
 1 file changed, 72 insertions(+), 9 deletions(-)
drivers/vfio/pci/nvgrace-gpu/main.c +63 -9
index b46984e76be7..3064f8aca858 100644
@@ -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

drivers/vfio/pci/nvgrace-gpu/main.c MODULE_DEVICE_TABLE(pci, nvgrace_gpu_vfio_pci_table); · line 1121

Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c +9 -0
  * 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