vfio/nvgpu: Add vfio pci variant module for grace hopper
[email protected]
maintainerPATCH v7 1/1 vfio/nvgpu: Add vfio pci variant module for grace hopper
From: Ankit Agrawal <[email protected]>
NVIDIA's upcoming Grace Hopper Superchip provides a PCI-like device for the on-chip GPU that is the logical OS representation of the internal proprietary cache coherent interconnect.
This representation has a number of limitations compared to a real PCI device, in particular, it does not model the coherent GPU memory aperture as a PCI config space BAR, and PCI doesn't know anything about cacheable memory types.
Provide a VFIO PCI variant driver that adapts the unique PCI representation into a more standard PCI representation facing userspace. The GPU memory aperture is obtained from ACPI using device_property_read_u64(), according to the FW specification, and exported to userspace as a separate VFIO_REGION. Since the device implements only one 64-bit BAR (BAR0), the GPU memory aperture is mapped to the next available PCI BAR (BAR2). Qemu will then naturally generate a PCI device in the VM with two 64-bit BARs (where the cacheable aperture reported in BAR2).
Since this memory region is actually cache coherent with the CPU, the VFIO variant driver will mmap it into VMA using a cacheable mapping. The mapping is done using remap_pfn_range().
PCI BAR are aligned to the power-of-2, but the actual memory on the device may not. A read or write access to the physical address from the last device PFN up to the next power-of-2 aligned physical address results in reading ~1 and dropped writes.
This goes along with a qemu series to provides the necessary implementation of the Grace Hopper Superchip firmware specification so that the guest operating system can see the correct ACPI modeling for the coherent GPU device. Verified with the CUDA workload in the VM.
https://www.mail-archive.com/[email protected]/msg967557.html
This patch is split from a patch series being pursued separately:
https://lore.kernel.org/lkml/[email protected]/
Applied and tested over next-20230821.
---
Link for v6: https://lore.kernel.org/all/[email protected]/
v6 -> v7
- Handled out-of-bound and overflow conditions at various places to validate input offset and length.
- Added code to return EINVAL for offset beyond region size.
- Memremap the device memory region and cached in the nvdev object until the device is closed
v5 -> v6
- Added the code to handle BAR2 read/write using memremap to the device memory.
v4 -> v5
- Changed the module name from nvgpu-vfio-pci to nvgrace-gpu-vfio-pci.
- Fixed memory leak and added suggested boundary checks on device memory mapping.
- Added code to read all Fs and ignored write on region outside of the physical memory.
- Other miscellaneous cleanup suggestions.
v3 -> v4
- Mapping the available device memory using sparse mmap. The region outside the device memory is handled by read/write ops.
- Removed the fault handler added in v3.
v2 -> v3
- Added fault handler to map the region outside the physical GPU memory up to the next power-of-2 to a dummy PFN.
- Changed to select instead of "depends on" VFIO_PCI_CORE for all the vfio-pci variant driver.
- Code cleanup based on feedback comments.
- Code implemented and tested against v6.4-rc4.
v1 -> v2
- Updated the wording of reference to BAR offset and replaced with index.
- The GPU memory is exposed at the fixed BAR2_REGION_INDEX.
- Code cleanup based on feedback comments.
MAINTAINERS | 6 + drivers/vfio/pci/Kconfig | 2 + drivers/vfio/pci/Makefile | 2 + drivers/vfio/pci/nvgrace-gpu/Kconfig | 10 + drivers/vfio/pci/nvgrace-gpu/Makefile | 3 + drivers/vfio/pci/nvgrace-gpu/main.c | 444 ++++++++++++++++++++++++++ 6 files changed, 467 insertions(+) create mode 100644 drivers/vfio/pci/nvgrace-gpu/Kconfig create mode 100644 drivers/vfio/pci/nvgrace-gpu/Makefile create mode 100644 drivers/vfio/pci/nvgrace-gpu/main.c
MAINTAINERS +6 -0
@@ -22347,6 +22347,12 @@ L: [email protected] S: Maintained F: drivers/vfio/platform/ +VFIO NVIDIA GRACE GPU DRIVER +M: Ankit Agrawal <[email protected]> +L: [email protected] +S: Maintained +F: drivers/vfio/pci/nvgrace-gpu/ + VGA_SWITCHEROO R: Lukas Wunner <[email protected]> S: Maintained
drivers/vfio/pci/Kconfig +2 -0
@@ -63,4 +63,6 @@ source "drivers/vfio/pci/mlx5/Kconfig" source "drivers/vfio/pci/hisilicon/Kconfig" +source "drivers/vfio/pci/nvgrace-gpu/Kconfig" + endmenu
drivers/vfio/pci/Makefile +2 -0
@@ -11,3 +11,5 @@ obj-$(CONFIG_VFIO_PCI) += vfio-pci.o obj-$(CONFIG_MLX5_VFIO_PCI) += mlx5/ obj-$(CONFIG_HISI_ACC_VFIO_PCI) += hisilicon/ + +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/
drivers/vfio/pci/nvgrace-gpu/Kconfig new file +10 -0
@@ -0,0 +1,10 @@ +# SPDX-License-Identifier: GPL-2.0-only +config NVGRACE_GPU_VFIO_PCI + tristate "VFIO support for the GPU in the NVIDIA Grace Hopper Superchip" + depends on ARM64 || (COMPILE_TEST && 64BIT) + select VFIO_PCI_CORE + help + VFIO support for the GPU in the NVIDIA Grace Hopper Superchip is + required to assign the GPU device to a VM using KVM/qemu/etc. + + If you don't know what to do here, say N.
drivers/vfio/pci/nvgrace-gpu/Makefile new file +3 -0
@@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0-only +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu-vfio-pci.o +nvgrace-gpu-vfio-pci-y := main.o
drivers/vfio/pci/nvgrace-gpu/main.c new file +444 -0
444 lines not on this page · raw message
Signature · 1 line
2.17.1
Quoted 128 lines · On Tue, 22 Aug 2023 13:23:03 -0700 <[email protected]> wrote:
From: Ankit Agrawal <[email protected]>
NVIDIA's upcoming Grace Hopper Superchip provides a PCI-like device for the on-chip GPU that is the logical OS representation of the internal proprietary cache coherent interconnect.
This representation has a number of limitations compared to a real PCI device, in particular, it does not model the coherent GPU memory aperture as a PCI config space BAR, and PCI doesn't know anything about cacheable memory types.
Provide a VFIO PCI variant driver that adapts the unique PCI representation into a more standard PCI representation facing userspace. The GPU memory aperture is obtained from ACPI using
device_property_read_u64(), according to the FW specification, and exported to userspace as a separateVFIO_REGION. Since the device implements only one 64-bit BAR (BAR0), the GPU memory aperture is mapped to the next available PCI BAR (BAR2). Qemu will then naturally generate a PCI device in the VM with two 64-bit BARs (where the cacheable aperture reported in BAR2).Since this memory region is actually cache coherent with the CPU, the VFIO variant driver will mmap it into VMA using a cacheable mapping. The mapping is done using
remap_pfn_range().PCI BAR are aligned to the power-of-2, but the actual memory on the device may not. A read or write access to the physical address from the last device PFN up to the next power-of-2 aligned physical address results in reading ~1 and dropped writes.
This goes along with a qemu series to provides the necessary implementation of the Grace Hopper Superchip firmware specification so that the guest operating system can see the correct ACPI modeling for the coherent GPU device. Verified with the CUDA workload in the VM.
https://www.mail-archive.com/[email protected]/msg967557.htmlThis patch is split from a patch series being pursued separately:
https://lore.kernel.org/lkml/[email protected]/Applied and tested over next-20230821.
Signed-off-by Ankit Agrawal <[email protected]>---
Link for v6: https://lore.kernel.org/all/[email protected]/
v6 -> v7
- Handled out-of-bound and overflow conditions at various places to validate input offset and length.
- Added code to return EINVAL for offset beyond region size.
- Memremap the device memory region and cached in the nvdev object until the device is closedv5 -> v6
- Added the code to handle BAR2 read/write using memremap to the device memory.v4 -> v5
- Changed the module name from nvgpu-vfio-pci to nvgrace-gpu-vfio-pci.
- Fixed memory leak and added suggested boundary checks on device memory mapping.
- Added code to read all Fs and ignored write on region outside of the physical memory.
- Other miscellaneous cleanup suggestions.v3 -> v4
- Mapping the available device memory using sparse mmap. The region outside the device memory is handled by read/write ops.
- Removed the fault handler added in v3.v2 -> v3
- Added fault handler to map the region outside the physical GPU memory up to the next power-of-2 to a dummy PFN.
- Changed to select instead of "depends on"VFIO_PCI_COREfor all the vfio-pci variant driver.
- Code cleanup based on feedback comments.
- Code implemented and tested against v6.4-rc4.v1 -> v2
- Updated the wording of reference to BAR offset and replaced with index.
- The GPU memory is exposed at the fixedBAR2_REGION_INDEX.
- Code cleanup based on feedback comments.MAINTAINERS | 6 + drivers/vfio/pci/Kconfig | 2 + drivers/vfio/pci/Makefile | 2 + drivers/vfio/pci/nvgrace-gpu/Kconfig | 10 + drivers/vfio/pci/nvgrace-gpu/Makefile | 3 + drivers/vfio/pci/nvgrace-gpu/main.c | 444 ++++++++++++++++++++++++++ 6 files changed, 467 insertions(+) create mode 100644 drivers/vfio/pci/nvgrace-gpu/Kconfig create mode 100644 drivers/vfio/pci/nvgrace-gpu/Makefile create mode 100644 drivers/vfio/pci/nvgrace-gpu/main.cMAINTAINERS +6
-0@@ -22347,6 +22347,12 @@ L: [email protected] S: Maintained F: drivers/vfio/platform/ +VFIO NVIDIA GRACE GPU DRIVER +M: Ankit Agrawal <[email protected]> +L: [email protected] +S: Maintained +F: drivers/vfio/pci/nvgrace-gpu/ + VGA_SWITCHEROO R: Lukas Wunner <[email protected]> S: Maintaineddrivers/vfio/pci/Kconfig +2
-0@@ -63,4 +63,6 @@ source "drivers/vfio/pci/mlx5/Kconfig" source "drivers/vfio/pci/hisilicon/Kconfig" +source "drivers/vfio/pci/nvgrace-gpu/Kconfig" + endmenudrivers/vfio/pci/Makefile +0
-0
drivers/vfio/pci/Makefile obj-$(CONFIG_VFIO_PCI) += vfio-pci.o · line 15
Quoted 5 lines
drivers/vfio/pci/Makefile +2
-0@@ -11,3 +11,5 @@ obj-$(CONFIG_VFIO_PCI) += vfio-pci.o obj-$(CONFIG_MLX5_VFIO_PCI) += mlx5/ obj-$(CONFIG_HISI_ACC_VFIO_PCI) += hisilicon/ + +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/
The pds driver should have been here using the linux-next base reported above.
Quoted 35 lines
drivers/vfio/pci/nvgrace-gpu/Kconfig new file +10
-0@@ -0,0 +1,10 @@ +# SPDX-License-Identifier: GPL-2.0-only +config NVGRACE_GPU_VFIO_PCI + tristate "VFIO support for the GPU in the NVIDIA Grace Hopper Superchip" + depends on ARM64 || (COMPILE_TEST && 64BIT) + select VFIO_PCI_CORE + help + VFIO support for the GPU in the NVIDIA Grace Hopper Superchip is + required to assign the GPU device to a VM using KVM/qemu/etc. + + If you don't know what to do here, say N.drivers/vfio/pci/nvgrace-gpu/Makefile new file +3
-0@@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0-only +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu-vfio-pci.o +nvgrace-gpu-vfio-pci-y := main.odrivers/vfio/pci/nvgrace-gpu/main.c new file +4
-0@@ -0,0 +1,444 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* +* Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved +*/
drivers/vfio/pci/nvgrace-gpu/main.c line 14
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+ +#include <linux/pci.h> +#include <linux/vfio_pci_core.h> +#include <linux/vfio.h> + +struct nvgrace_gpu_vfio_pci_core_device { +struct vfio_pci_core_device core_device; +u64 hpa; +u64 mem_length; +void *opregion;
Maybe mem_hpa, mem_lenth, mem_map?
Quoted 122 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +122
-0+}; + +static int nvgrace_gpu_vfio_pci_open_device (struct vfio_device *core_vdev) +{ +struct vfio_pci_core_device *vdev = + container_of(core_vdev, struct vfio_pci_core_device, vdev); +int ret; + + ret =vfio_pci_core_enable (vdev); +if (ret) +return ret; + +vfio_pci_core_finish_enable (vdev); + +return 0 ; +} + +static void nvgrace_gpu_vfio_pci_close_device (struct vfio_device *core_vdev) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (nvdev->opregion) { +memunmap (nvdev->opregion); + nvdev->opregion =NULL ; + } + +vfio_pci_core_close_device (core_vdev); +} + +static int nvgrace_gpu_vfio_pci_mmap (struct vfio_device *core_vdev, +struct vm_area_struct *vma) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +unsigned long start_pfn; +unsigned int index; +u64 req_len, pgoff, end; +int ret =0 ; + + index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT -PAGE_SHIFT ); +if (index !=VFIO_PCI_BAR2_REGION_INDEX ) +return vfio_pci_core_mmap (core_vdev, vma); + +/* +* Request to mmap the BAR. Map to the CPU accessible memory on the +* GPU using the memory information gathered from the system ACPI +* tables. +*/ + pgoff = vma->vm_pgoff & + ((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 (nvdev->hpa), pgoff, &start_pfn) || +check_add_overflow (PFN_PHYS (pgoff), req_len, &end)) +return -EOVERFLOW ; + +if (end > nvdev->mem_length) +return -EINVAL ; + +/* +* Perform a PFN map to the memory. The device BAR is backed by the +* GPU memory now. Check that the mapping does not overflow out of +* the GPU memory size. +* +* The available GPU memory size may not be power-of-2 aligned. Given +* that the memory is exposed as a BAR, the mapping request is of the +* power-of-2 aligned size. Map only up to the size of the GPU memory. +* If the memory access is beyond the actual GPU memory size, it will +* be handled by the vfio_device_ops read/write. +* +* 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; + +return 0 ; +} + +static long nvgrace_gpu_vfio_pci_ioctl (struct vfio_device *core_vdev, +unsigned int cmd,unsigned long arg) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + + unsigned long minsz = offsetofend(struct vfio_region_info, offset); +struct vfio_region_info info; + +if (cmd ==VFIO_DEVICE_GET_REGION_INFO ) { +if (copy_from_user (&info, (void __user *)arg, minsz)) +return -EFAULT ; + +if (info.argsz < minsz) +return -EINVAL ; + +if (info.index ==VFIO_PCI_BAR2_REGION_INDEX ) { +/* +* Request to determine the BAR region information. Send the +* GPU memory information. +*/ +uint32_t size; +struct vfio_region_info_cap_sparse_mmap *sparse; +struct vfio_info_cap caps = { .buf =NULL , .size =0 }; + + size =struct_size (sparse, areas,1 ); + +/* +* Setup for sparse mapping for the device memory. Only the +* available device memory on the hardware is shown as a +* mappable region. +*/ + sparse =kzalloc (size,GFP_KERNEL ); +if (!sparse) +return -ENOMEM ; +
drivers/vfio/pci/nvgrace-gpu/main.c line 146
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+ sparse->nr_areas =1 ; + sparse->areas[0 ].offset =0 ; + sparse->areas[0 ].size = nvdev->mem_length; + sparse->header.id =VFIO_REGION_INFO_CAP_SPARSE_MMAP ; + sparse->header.version =1 ; + +if (vfio_info_add_capability (&caps, &sparse->header, size)) { +kfree (sparse); +return -EINVAL ; + }
Ideally this would be:
ret = vfio_info_add_capability(...
if (ret) {
kfree(sparse);
return ret;
}Which avoids inventing return values, but also I think we're done with @sparse here, so it would simplify later error handling to have:
ret = vfio_info_add_capability(... kfree(sparse); if (ret) return ret;
Quoted 37 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +37
-0+ + info.offset =VFIO_PCI_INDEX_TO_OFFSET (info.index); +/* +* The available GPU memory size may not be power-of-2 aligned. +* Given that the memory is exposed as a BAR and may not be +* aligned, roundup to the next power-of-2. +*/ + info.size =roundup_pow_of_two (nvdev->mem_length); + info.flags =VFIO_REGION_INFO_FLAG_READ | +VFIO_REGION_INFO_FLAG_WRITE | +VFIO_REGION_INFO_FLAG_MMAP ; + +if (caps.size) { + info.flags |=VFIO_REGION_INFO_FLAG_CAPS ; +if (info.argsz <sizeof (info) + caps.size) { + info.argsz =sizeof (info) + caps.size; + info.cap_offset =0 ; + }else { +vfio_info_cap_shift (&caps,sizeof (info)); +if (copy_to_user ((void __user *)arg + +sizeof (info), caps.buf, + caps.size)) { +kfree (caps.buf); +kfree (sparse); +return -EFAULT ; + } + info.cap_offset =sizeof (info); + } +kfree (caps.buf); + } + +kfree (sparse); +return copy_to_user ((void __user *)arg, &info, minsz) ? + -EFAULT :0 ; + } + } +
drivers/vfio/pci/nvgrace-gpu/main.c line 193
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+return vfio_pci_core_ioctl (core_vdev, cmd, arg); +} + +/* +* Read count bytes from the device memory at an offset. The actual device +* memory size (available) may not be a power-of-2. So the driver fakes +* the size to a power-of-2 (reported) when exposing to a user space driver. +* +* Read request beyond the actual device size is filled with ~1, while +* those beyond the actual reported size is skipped.
I'm not sure why the commit log and description here describe this as "~1", which is the bit-wise NOT of 1, ie. 0x...fe, when we're using -1, ie. ~0.
drivers/vfio/pci/nvgrace-gpu/main.c line 199
Quoted 6 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +6
-0+* +* A read from a reported+ offset is considered error conditions and +* returned with an -EINVAL. Overflow conditions are also handled. +*/ +ssize_t nvgrace_gpu_read_mem (void __user *buf,size_t count,loff_t *ppos, +const void *addr,size_t available,size_t reported)
The translation to a new set of variable names is distracting here, we could simply pass nvdev itself or pull out "opregion" and "mem_length" (or the above suggested names) and let this function calculate "bar_size" (or "region_size") itself.
Quoted 5 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +5
-0+{ +u64 offset = *ppos &VFIO_PCI_OFFSET_MASK ; +u64 end; +size_t read_count, i; +u8 val =0xFF ;
drivers/vfio/pci/nvgrace-gpu/main.c line 214
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+ +if (offset >= reported) +return -EINVAL ; + +if (check_add_overflow (offset, count, &end)) +return -EOVERFLOW ; + +/* Clip short the read request beyond reported BAR size */ +if (end >= reported) + count = reported - offset;
This would typically be something like:
count = min(count, reported - offset);
drivers/vfio/pci/nvgrace-gpu/main.c line 223
Quoted 9 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +9
-0+ +/* +* Determine how many bytes to be actually read from the device memory. +* Do not read from the offset beyond available size. +*/ +if (offset >= available) + read_count =0 ; +else + read_count =min (count, available - (size_t )offset);
phys_count? mem_count?
drivers/vfio/pci/nvgrace-gpu/main.c line 230
Quoted 7 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +7
-0+ +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ;
Just to verify, does this memory allow access of arbitrary alignment and size?
drivers/vfio/pci/nvgrace-gpu/main.c line 238
Quoted 8 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +8
-0+ +/* +* Only the device memory present on the hardware is mapped, which may +* not be power-of-2 aligned. A read to the BAR2 region implies an +* access outside the available device memory on the hardware. Fill +* such read request with ~1. +*/ +for (i =0 ; i < count - read_count; i++)
Why not:
for (i = read_count; i < count; i++)
Quoted 9 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +9
-0+if (copy_to_user (buf + read_count + i, &val,1 )) +return -EFAULT ; + +return count; +} + +static ssize_t nvgrace_gpu_vfio_pci_read (struct vfio_device *core_vdev, +char __user *buf,size_t count,loff_t *ppos) +{
drivers/vfio/pci/nvgrace-gpu/main.c line 257
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+unsigned int index =VFIO_PCI_OFFSET_TO_INDEX (*ppos); +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }
Seems like this would be susceptible to concurrent accesses causing duplicate mappings.
The write() path should adopt similar changes to the read() path above. Thanks,
Alex
Quoted 187 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +187
-0+ +return nvgrace_gpu_read_mem (buf, count, ppos, nvdev->opregion, + nvdev->mem_length,roundup_pow_of_two (nvdev->mem_length)); + } + +return vfio_pci_core_read (core_vdev, buf, count, ppos); +} + +/* +* Write count bytes to the device memory at a given offset. The actual device +* memory size (available) may not be a power-of-2. So the driver fakes the +* size to a power-of-2 (reported) when exposing to a user space driver. +* +* Write request beyond the actual device size are dropped, while those +* beyond the actual reported size are skipped entirely. +* +* A write to a reported+ offset is considered error conditions and +* returned with an -EINVAL. Overflow conditions are also handled. +*/ +ssize_t nvgrace_gpu_write_mem (const void *addr,size_t count,loff_t *ppos, +const void __user *buf,size_t available,size_t reported) +{ +u64 offset = *ppos &VFIO_PCI_OFFSET_MASK ; +u64 end; +size_t write_count; + +if (offset >= reported) +return -EINVAL ; + +if (check_add_overflow (offset, count, &end)) +return -EOVERFLOW ; + +/* Clip short the read request beyond reported BAR size */ +if (end >= reported) + count = reported - offset; + +/* +* Determine how many bytes to be actually written to the device memory. +* Do not write to the offset beyond available size. +*/ +if (offset >= available) + write_count =0 ; +else + write_count =min (count, available - (size_t )offset); + +/* +* Only the device memory present on the hardware is mapped, which may +* not be power-of-2 aligned. A write to the BAR2 region implies an +* access outside the available device memory on the hardware. Drop +* those write requests. +*/ +if (copy_from_user ((u8 *)addr + offset, buf, write_count)) +return -EFAULT ; + +return count; +} + +static ssize_t nvgrace_gpu_vfio_pci_write (struct vfio_device *core_vdev, +const char __user *buf,size_t count,loff_t *ppos) +{ +unsigned int index =VFIO_PCI_OFFSET_TO_INDEX (*ppos); +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + } + +return nvgrace_gpu_write_mem (nvdev->opregion, count, ppos, buf, + nvdev->mem_length,roundup_pow_of_two (nvdev->mem_length)); + } + +return vfio_pci_core_write (core_vdev, buf, count, ppos); +} + +static const struct vfio_device_ops nvgrace_gpu_vfio_pci_ops = { + .name ="nvgrace-gpu-vfio-pci" , + .init = vfio_pci_core_init_dev, + .release = vfio_pci_core_release_dev, + .open_device = nvgrace_gpu_vfio_pci_open_device, + .close_device = nvgrace_gpu_vfio_pci_close_device, + .ioctl = nvgrace_gpu_vfio_pci_ioctl, + .read = nvgrace_gpu_vfio_pci_read, + .write = nvgrace_gpu_vfio_pci_write, + .mmap = nvgrace_gpu_vfio_pci_mmap, + .request = vfio_pci_core_request, + .match = vfio_pci_core_match, + .bind_iommufd = vfio_iommufd_physical_bind, + .unbind_iommufd = vfio_iommufd_physical_unbind, + .attach_ioas = vfio_iommufd_physical_attach_ioas, +}; + +static struct +nvgrace_gpu_vfio_pci_core_device *nvgrace_gpu_drvdata (struct pci_dev *pdev) +{ +struct vfio_pci_core_device *core_device =dev_get_drvdata (&pdev->dev); + + return container_of(core_device, struct nvgrace_gpu_vfio_pci_core_device, + core_device); +} + +static int +nvgrace_gpu_vfio_pci_fetch_memory_property (struct pci_dev *pdev, +struct nvgrace_gpu_vfio_pci_core_device *nvdev) +{ +int ret; + +/* +* The memory information is present in the system ACPI tables as DSD +* properties nvidia,gpu-mem-base-pa and nvidia,gpu-mem-size. +*/ + ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-base-pa" , + &(nvdev->hpa)); +if (ret) +return ret; + + ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-size" , + &(nvdev->mem_length)); +return ret; +} + +static int nvgrace_gpu_vfio_pci_probe (struct pci_dev *pdev, +const struct pci_device_id *id) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev; +int ret; + + nvdev =vfio_alloc_device (nvgrace_gpu_vfio_pci_core_device, core_device.vdev, + &pdev->dev, &nvgrace_gpu_vfio_pci_ops); +if (IS_ERR (nvdev)) +return PTR_ERR (nvdev); + +dev_set_drvdata (&pdev->dev, nvdev); + + ret =nvgrace_gpu_vfio_pci_fetch_memory_property (pdev, nvdev); +if (ret) + goto out_put_vdev; + + ret =vfio_pci_core_register_device (&nvdev->core_device); +if (ret) + goto out_put_vdev; + +return ret; + +out_put_vdev: +vfio_put_device (&nvdev->core_device.vdev); +return ret; +} + +static void nvgrace_gpu_vfio_pci_remove (struct pci_dev *pdev) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =nvgrace_gpu_drvdata (pdev); +struct vfio_pci_core_device *vdev = &nvdev->core_device; + +vfio_pci_core_unregister_device (vdev); +vfio_put_device (&vdev->vdev); +} + +static const struct pci_device_id nvgrace_gpu_vfio_pci_table[] = { +/* GH200 120GB */ + {PCI_DRIVER_OVERRIDE_DEVICE_VFIO (PCI_VENDOR_ID_NVIDIA ,0x2342 ) }, +/* GH200 480GB */ + {PCI_DRIVER_OVERRIDE_DEVICE_VFIO (PCI_VENDOR_ID_NVIDIA ,0x2345 ) }, + {} +}; + +MODULE_DEVICE_TABLE (pci, nvgrace_gpu_vfio_pci_table); + +static struct pci_driver nvgrace_gpu_vfio_pci_driver = { + .name =KBUILD_MODNAME , + .id_table = nvgrace_gpu_vfio_pci_table, + .probe = nvgrace_gpu_vfio_pci_probe, + .remove = nvgrace_gpu_vfio_pci_remove, + .err_handler = &vfio_pci_core_err_handlers, + .driver_managed_dma = true, +}; + +module_pci_driver (nvgrace_gpu_vfio_pci_driver); + +MODULE_LICENSE ("GPL v2" ); +MODULE_AUTHOR ("Ankit Agrawal <[email protected]>" ); +MODULE_AUTHOR ("Aniket Agashe <[email protected]>" ); +MODULE_DESCRIPTION ( +"VFIO NVGRACE GPU PF - User Level driver for NVIDIA devices with CPU coherently accessible device memory" );
Jason Gunthorpe
reviewerQuoted 6 lines · On Tue, Aug 22, 2023 at 01:23:03PM -0700, [email protected] wrote:
drivers/vfio/pci/nvgrace-gpu/main.c new file +5
-0@@ -0,0 +1,444 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* +* Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved +*/ +
drivers/vfio/pci/nvgrace-gpu/main.c line 15
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+#include <linux/pci.h> +#include <linux/vfio_pci_core.h> +#include <linux/vfio.h> + +struct nvgrace_gpu_vfio_pci_core_device { +struct vfio_pci_core_device core_device; +u64 hpa; +u64 mem_length; +void *opregion; +};
opregion is some word that has meaning for the intel drivers, use something else please.
'hpa' has no business being in a VFIO driver.
phys_addr_t memphyssize_t memlength
void __iomem *memmap;
drivers/vfio/pci/nvgrace-gpu/main.c line 110
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+static long nvgrace_gpu_vfio_pci_ioctl (struct vfio_device *core_vdev, +unsigned int cmd,unsigned long arg) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + + unsigned long minsz = offsetofend(struct vfio_region_info, offset); +struct vfio_region_info info; + +if (cmd ==VFIO_DEVICE_GET_REGION_INFO ) {
This block is long enough to put in its own function and remove a level of indenting
Quoted 44 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +44
-0+/* +* Read count bytes from the device memory at an offset. The actual device +* memory size (available) may not be a power-of-2. So the driver fakes +* the size to a power-of-2 (reported) when exposing to a user space driver. +* +* Read request beyond the actual device size is filled with ~1, while +* those beyond the actual reported size is skipped. +* +* A read from a reported+ offset is considered error conditions and +* returned with an -EINVAL. Overflow conditions are also handled. +*/ +ssize_t nvgrace_gpu_read_mem (void __user *buf,size_t count,loff_t *ppos, +const void *addr,size_t available,size_t reported) +{ +u64 offset = *ppos &VFIO_PCI_OFFSET_MASK ; +u64 end; +size_t read_count, i; +u8 val =0xFF ; + +if (offset >= reported) +return -EINVAL ; + +if (check_add_overflow (offset, count, &end)) +return -EOVERFLOW ; + +/* Clip short the read request beyond reported BAR size */ +if (end >= reported) + count = reported - offset; + +/* +* Determine how many bytes to be actually read from the device memory. +* Do not read from the offset beyond available size. +*/ +if (offset >= available) + read_count =0 ; +else + read_count =min (count, available - (size_t )offset); + +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ;
drivers/vfio/pci/nvgrace-gpu/main.c line 240
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+ +/* +* Only the device memory present on the hardware is mapped, which may +* not be power-of-2 aligned. A read to the BAR2 region implies an +* access outside the available device memory on the hardware. Fill +* such read request with ~1. +*/ +for (i =0 ; i < count - read_count; i++) +if (copy_to_user (buf + read_count + i, &val,1 )) +return -EFAULT ;
Use put_user()
drivers/vfio/pci/nvgrace-gpu/main.c line 327
Quoted 13 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +13
-0+static ssize_t nvgrace_gpu_vfio_pci_write (struct vfio_device *core_vdev, +const char __user *buf,size_t count,loff_t *ppos) +{ +unsigned int index =VFIO_PCI_OFFSET_TO_INDEX (*ppos); +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }
Needs some kind of locking on opregion
Quoted 4 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +4
-0+static int +nvgrace_gpu_vfio_pci_fetch_memory_property (struct pci_dev *pdev, +struct nvgrace_gpu_vfio_pci_core_device *nvdev) +{
drivers/vfio/pci/nvgrace-gpu/main.c line 375
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +10
-0+int ret; + +/* +* The memory information is present in the system ACPI tables as DSD +* properties nvidia,gpu-mem-base-pa and nvidia,gpu-mem-size. +*/ + ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-base-pa" , + &(nvdev->hpa)); +if (ret) +return ret;
Technically you need to check that the read_u64 doesn't overflow phys_addr_t
drivers/vfio/pci/nvgrace-gpu/main.c line 379
Quoted 3 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +3
-0+ ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-size" , + &(nvdev->mem_length)); +return ret;
And that mem_length doesn't overflow size_t
Jason
Ankit Agrawal
maintainerQuoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c line 230
Quoted 7 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +7
-0+ +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ;Just to verify, does this memory allow access of arbitrary alignment and size?
Please correct me if I'm wrong, but based on following gdb dump data on the corresponding MemoryRegion->ops, unaligned access isn't supported, and a read of size upto 8 may be done.
(gdb) p/x *(mr->ops)
$7 = {read = 0xaaab5e0b1c50, write = 0xaaab5e0b1a50, read_with_attrs = 0x0, write_with_attrs = 0x0,
endianness = 0x2, valid = {min_access_size = 0x1, max_access_size = 0x8, unaligned = 0x0, accepts = 0x0}, impl = {min_access_size = 0x1, max_access_size = 0x8, unaligned = 0x0}}
Thanks
Ankit Agrawal
Jason Gunthorpe
reviewerQuoted 8 lines · Ankit Agrawal
Quoted 8 lines
Quoted 7 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +7
-0+ +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ;
Quoted 6 lines
Quoted 2 lines
Just to verify, does this memory allow access of arbitrary alignment and size?
Please correct me if I'm wrong, but based on following gdb dump data on the corresponding
MemoryRegion->ops, unaligned access isn't supported, and a read of size upto 8 may be done.
Regardless, you used MEMREMAP_WB which is equivalent to normal system memory. It supports all accesses, including atomics, and doesn't require __iomem accessors.
Jason
Quoted 11 lines · On Wed, 23 Aug 2023 14:18:52 +0000 Ankit Agrawal <[email protected]> wrote:
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c line 230
Quoted 7 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +7
-0+ +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ;Just to verify, does this memory allow access of arbitrary alignment and size?
Quoted 8 lines
Please correct me if I'm wrong, but based on following gdb dump data on the corresponding
MemoryRegion->ops, unaligned access isn't supported, and a read of size upto 8 may be done.(gdb) p/x *(mr->ops)
$7 = {read =0xaaab5e0b1c50, write =0xaaab5e0b1a50,read_with_attrs=0x0,write_with_attrs=0x0,
endianness =0x2, valid = {min_access_size=0x1,max_access_size=0x8, unaligned =0x0, accepts =0x0}, impl = {min_access_size=0x1,max_access_size=0x8, unaligned =0x0}}
This is QEMU policy relative to this region, the kernel interface is not exclusive to QEMU usage. Thanks,
Alex
Ankit Agrawal
maintainerQuoted 11 lines
drivers/vfio/pci/nvgrace-gpu/main.c line 257
Quoted 6 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +6
-0+if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }[AW] Seems like this would be susceptible to concurrent accesses causing duplicate mappings.
[JG] Needs some kind of locking on opregion
Right, will add a new lock item in nvdev to control the access to opregion/memmap. Please let me know if it is preferable to do memremap in open_device instead of read/write.
Quoted 7 lines · On Wed, 23 Aug 2023 14:50:31 +0000 Ankit Agrawal <[email protected]> wrote:
Quoted 7 lines
Quoted 6 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +6
-0+if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }
Quoted 8 lines
Quoted 4 lines
[AW] Seems like this would be susceptible to concurrent accesses causing duplicate mappings.
[JG] Needs some kind of locking on opregion
Right, will add a new lock item in nvdev to control the access to opregion/memmap. Please let me know if it is preferable to do memremap in
open_deviceinstead of read/write.
That's a valid option also, certainly avoids the locking and serialization per access. Thanks,
Alex
Jason Gunthorpe
reviewerQuoted 13 lines · Alex Williamson
Quoted 10 lines · On Wed, 23 Aug 2023 14:50:31 +0000 Ankit Agrawal <[email protected]> wrote:
Quoted 10 lines
drivers/vfio/pci/nvgrace-gpu/main.c line 257
Quoted 6 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +6
-0+if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }[AW] Seems like this would be susceptible to concurrent accesses causing duplicate mappings.
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
[JG] Needs some kind of locking on opregion
Right, will add a new lock item in nvdev to control the access to opregion/memmap. Please let me know if it is preferable to do memremap in
open_deviceinstead of read/write.That's a valid option also, certainly avoids the locking and serialization per access. Thanks,
open_device is no good, that would waste large amounts of kernel memory for page tables to support something we don't expect to be used.
Jason
Quoted 20 lines · On Wed, 23 Aug 2023 12:16:23 -0300 Jason Gunthorpe <[email protected]> wrote:
Quoted 19 lines · Alex Williamson
Quoted 15 lines · On Wed, 23 Aug 2023 14:50:31 +0000 Ankit Agrawal <[email protected]> wrote:
Quoted 11 lines
drivers/vfio/pci/nvgrace-gpu/main.c line 257
Quoted 6 lines
drivers/vfio/pci/nvgrace-gpu/main.c new file +6
-0+if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + }[AW] Seems like this would be susceptible to concurrent accesses causing duplicate mappings.
[JG] Needs some kind of locking on opregion
Right, will add a new lock item in nvdev to control the access to opregion/memmap. Please let me know if it is preferable to do memremap in
open_deviceinstead of read/write.
Quoted 6 lines
Quoted 2 lines
That's a valid option also, certainly avoids the locking and serialization per access. Thanks,
open_deviceis no good, that would waste large amounts of kernel memory for page tables to support something we don't expect to be used.
A lock it is then, I guess this is a very large range of memory. The contention/serialization is also not a priority, we expect access through the mmap for anything other than debug. Thanks,
Alex
I know I'm chiming in a bit late, but what ultimate user space is going to use this? We should not add anything to the kernel that can't be used without fully open user space.
vfio has traditionally been a bit special as it "just" passes devices through, so any user space could just be a user space driver for a random device on $FOO bus, including an actual Linux driver in a VM, but this driver has very specific semantics for a very specific piece of hardware, so it really needs to be treated like a generic GPU driver or accelerator driver.
Quoted 611 lines · On Tue, Aug 22, 2023 at 01:23:03PM -0700, [email protected] wrote:
From: Ankit Agrawal <[email protected]>
NVIDIA's upcoming Grace Hopper Superchip provides a PCI-like device for the on-chip GPU that is the logical OS representation of the internal proprietary cache coherent interconnect.
This representation has a number of limitations compared to a real PCI device, in particular, it does not model the coherent GPU memory aperture as a PCI config space BAR, and PCI doesn't know anything about cacheable memory types.
Provide a VFIO PCI variant driver that adapts the unique PCI representation into a more standard PCI representation facing userspace. The GPU memory aperture is obtained from ACPI using
device_property_read_u64(), according to the FW specification, and exported to userspace as a separateVFIO_REGION. Since the device implements only one 64-bit BAR (BAR0), the GPU memory aperture is mapped to the next available PCI BAR (BAR2). Qemu will then naturally generate a PCI device in the VM with two 64-bit BARs (where the cacheable aperture reported in BAR2).Since this memory region is actually cache coherent with the CPU, the VFIO variant driver will mmap it into VMA using a cacheable mapping. The mapping is done using
remap_pfn_range().PCI BAR are aligned to the power-of-2, but the actual memory on the device may not. A read or write access to the physical address from the last device PFN up to the next power-of-2 aligned physical address results in reading ~1 and dropped writes.
This goes along with a qemu series to provides the necessary implementation of the Grace Hopper Superchip firmware specification so that the guest operating system can see the correct ACPI modeling for the coherent GPU device. Verified with the CUDA workload in the VM.
https://www.mail-archive.com/[email protected]/msg967557.htmlThis patch is split from a patch series being pursued separately:
https://lore.kernel.org/lkml/[email protected]/Applied and tested over next-20230821.
Signed-off-by Ankit Agrawal <[email protected]>---
Link for v6: https://lore.kernel.org/all/[email protected]/
v6 -> v7
- Handled out-of-bound and overflow conditions at various places to validate input offset and length.
- Added code to return EINVAL for offset beyond region size.
- Memremap the device memory region and cached in the nvdev object until the device is closedv5 -> v6
- Added the code to handle BAR2 read/write using memremap to the device memory.v4 -> v5
- Changed the module name from nvgpu-vfio-pci to nvgrace-gpu-vfio-pci.
- Fixed memory leak and added suggested boundary checks on device memory mapping.
- Added code to read all Fs and ignored write on region outside of the physical memory.
- Other miscellaneous cleanup suggestions.v3 -> v4
- Mapping the available device memory using sparse mmap. The region outside the device memory is handled by read/write ops.
- Removed the fault handler added in v3.v2 -> v3
- Added fault handler to map the region outside the physical GPU memory up to the next power-of-2 to a dummy PFN.
- Changed to select instead of "depends on"VFIO_PCI_COREfor all the vfio-pci variant driver.
- Code cleanup based on feedback comments.
- Code implemented and tested against v6.4-rc4.v1 -> v2
- Updated the wording of reference to BAR offset and replaced with index.
- The GPU memory is exposed at the fixedBAR2_REGION_INDEX.
- Code cleanup based on feedback comments.MAINTAINERS | 6 + drivers/vfio/pci/Kconfig | 2 + drivers/vfio/pci/Makefile | 2 + drivers/vfio/pci/nvgrace-gpu/Kconfig | 10 + drivers/vfio/pci/nvgrace-gpu/Makefile | 3 + drivers/vfio/pci/nvgrace-gpu/main.c | 444 ++++++++++++++++++++++++++ 6 files changed, 467 insertions(+) create mode 100644 drivers/vfio/pci/nvgrace-gpu/Kconfig create mode 100644 drivers/vfio/pci/nvgrace-gpu/Makefile create mode 100644 drivers/vfio/pci/nvgrace-gpu/main.cMAINTAINERS +6
-0@@ -22347,6 +22347,12 @@ L: [email protected] S: Maintained F: drivers/vfio/platform/ +VFIO NVIDIA GRACE GPU DRIVER +M: Ankit Agrawal <[email protected]> +L: [email protected] +S: Maintained +F: drivers/vfio/pci/nvgrace-gpu/ + VGA_SWITCHEROO R: Lukas Wunner <[email protected]> S: Maintaineddrivers/vfio/pci/Kconfig +2
-0@@ -63,4 +63,6 @@ source "drivers/vfio/pci/mlx5/Kconfig" source "drivers/vfio/pci/hisilicon/Kconfig" +source "drivers/vfio/pci/nvgrace-gpu/Kconfig" + endmenudrivers/vfio/pci/Makefile +2
-0@@ -11,3 +11,5 @@ obj-$(CONFIG_VFIO_PCI) += vfio-pci.o obj-$(CONFIG_MLX5_VFIO_PCI) += mlx5/ obj-$(CONFIG_HISI_ACC_VFIO_PCI) += hisilicon/ + +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/drivers/vfio/pci/nvgrace-gpu/Kconfig new file +10
-0@@ -0,0 +1,10 @@ +# SPDX-License-Identifier: GPL-2.0-only +config NVGRACE_GPU_VFIO_PCI + tristate "VFIO support for the GPU in the NVIDIA Grace Hopper Superchip" + depends on ARM64 || (COMPILE_TEST && 64BIT) + select VFIO_PCI_CORE + help + VFIO support for the GPU in the NVIDIA Grace Hopper Superchip is + required to assign the GPU device to a VM using KVM/qemu/etc. + + If you don't know what to do here, say N.drivers/vfio/pci/nvgrace-gpu/Makefile new file +3
-0@@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0-only +obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu-vfio-pci.o +nvgrace-gpu-vfio-pci-y := main.odrivers/vfio/pci/nvgrace-gpu/main.c new file +444
-0@@ -0,0 +1,444 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* +* Copyright (c) 2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved +*/ + +#include <linux/pci.h> +#include <linux/vfio_pci_core.h> +#include <linux/vfio.h> + +struct nvgrace_gpu_vfio_pci_core_device { +struct vfio_pci_core_device core_device; +u64 hpa; +u64 mem_length; +void *opregion; +}; + +static int nvgrace_gpu_vfio_pci_open_device (struct vfio_device *core_vdev) +{ +struct vfio_pci_core_device *vdev = + container_of(core_vdev, struct vfio_pci_core_device, vdev); +int ret; + + ret =vfio_pci_core_enable (vdev); +if (ret) +return ret; + +vfio_pci_core_finish_enable (vdev); + +return 0 ; +} + +static void nvgrace_gpu_vfio_pci_close_device (struct vfio_device *core_vdev) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (nvdev->opregion) { +memunmap (nvdev->opregion); + nvdev->opregion =NULL ; + } + +vfio_pci_core_close_device (core_vdev); +} + +static int nvgrace_gpu_vfio_pci_mmap (struct vfio_device *core_vdev, +struct vm_area_struct *vma) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +unsigned long start_pfn; +unsigned int index; +u64 req_len, pgoff, end; +int ret =0 ; + + index = vma->vm_pgoff >> (VFIO_PCI_OFFSET_SHIFT -PAGE_SHIFT ); +if (index !=VFIO_PCI_BAR2_REGION_INDEX ) +return vfio_pci_core_mmap (core_vdev, vma); + +/* +* Request to mmap the BAR. Map to the CPU accessible memory on the +* GPU using the memory information gathered from the system ACPI +* tables. +*/ + pgoff = vma->vm_pgoff & + ((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 (nvdev->hpa), pgoff, &start_pfn) || +check_add_overflow (PFN_PHYS (pgoff), req_len, &end)) +return -EOVERFLOW ; + +if (end > nvdev->mem_length) +return -EINVAL ; + +/* +* Perform a PFN map to the memory. The device BAR is backed by the +* GPU memory now. Check that the mapping does not overflow out of +* the GPU memory size. +* +* The available GPU memory size may not be power-of-2 aligned. Given +* that the memory is exposed as a BAR, the mapping request is of the +* power-of-2 aligned size. Map only up to the size of the GPU memory. +* If the memory access is beyond the actual GPU memory size, it will +* be handled by the vfio_device_ops read/write. +* +* 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; + +return 0 ; +} + +static long nvgrace_gpu_vfio_pci_ioctl (struct vfio_device *core_vdev, +unsigned int cmd,unsigned long arg) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + + unsigned long minsz = offsetofend(struct vfio_region_info, offset); +struct vfio_region_info info; + +if (cmd ==VFIO_DEVICE_GET_REGION_INFO ) { +if (copy_from_user (&info, (void __user *)arg, minsz)) +return -EFAULT ; + +if (info.argsz < minsz) +return -EINVAL ; + +if (info.index ==VFIO_PCI_BAR2_REGION_INDEX ) { +/* +* Request to determine the BAR region information. Send the +* GPU memory information. +*/ +uint32_t size; +struct vfio_region_info_cap_sparse_mmap *sparse; +struct vfio_info_cap caps = { .buf =NULL , .size =0 }; + + size =struct_size (sparse, areas,1 ); + +/* +* Setup for sparse mapping for the device memory. Only the +* available device memory on the hardware is shown as a +* mappable region. +*/ + sparse =kzalloc (size,GFP_KERNEL ); +if (!sparse) +return -ENOMEM ; + + sparse->nr_areas =1 ; + sparse->areas[0 ].offset =0 ; + sparse->areas[0 ].size = nvdev->mem_length; + sparse->header.id =VFIO_REGION_INFO_CAP_SPARSE_MMAP ; + sparse->header.version =1 ; + +if (vfio_info_add_capability (&caps, &sparse->header, size)) { +kfree (sparse); +return -EINVAL ; + } + + info.offset =VFIO_PCI_INDEX_TO_OFFSET (info.index); +/* +* The available GPU memory size may not be power-of-2 aligned. +* Given that the memory is exposed as a BAR and may not be +* aligned, roundup to the next power-of-2. +*/ + info.size =roundup_pow_of_two (nvdev->mem_length); + info.flags =VFIO_REGION_INFO_FLAG_READ | +VFIO_REGION_INFO_FLAG_WRITE | +VFIO_REGION_INFO_FLAG_MMAP ; + +if (caps.size) { + info.flags |=VFIO_REGION_INFO_FLAG_CAPS ; +if (info.argsz <sizeof (info) + caps.size) { + info.argsz =sizeof (info) + caps.size; + info.cap_offset =0 ; + }else { +vfio_info_cap_shift (&caps,sizeof (info)); +if (copy_to_user ((void __user *)arg + +sizeof (info), caps.buf, + caps.size)) { +kfree (caps.buf); +kfree (sparse); +return -EFAULT ; + } + info.cap_offset =sizeof (info); + } +kfree (caps.buf); + } + +kfree (sparse); +return copy_to_user ((void __user *)arg, &info, minsz) ? + -EFAULT :0 ; + } + } + +return vfio_pci_core_ioctl (core_vdev, cmd, arg); +} + +/* +* Read count bytes from the device memory at an offset. The actual device +* memory size (available) may not be a power-of-2. So the driver fakes +* the size to a power-of-2 (reported) when exposing to a user space driver. +* +* Read request beyond the actual device size is filled with ~1, while +* those beyond the actual reported size is skipped. +* +* A read from a reported+ offset is considered error conditions and +* returned with an -EINVAL. Overflow conditions are also handled. +*/ +ssize_t nvgrace_gpu_read_mem (void __user *buf,size_t count,loff_t *ppos, +const void *addr,size_t available,size_t reported) +{ +u64 offset = *ppos &VFIO_PCI_OFFSET_MASK ; +u64 end; +size_t read_count, i; +u8 val =0xFF ; + +if (offset >= reported) +return -EINVAL ; + +if (check_add_overflow (offset, count, &end)) +return -EOVERFLOW ; + +/* Clip short the read request beyond reported BAR size */ +if (end >= reported) + count = reported - offset; + +/* +* Determine how many bytes to be actually read from the device memory. +* Do not read from the offset beyond available size. +*/ +if (offset >= available) + read_count =0 ; +else + read_count =min (count, available - (size_t )offset); + +/* +* Handle read on the BAR2 region. Map to the target device memory +* physical address and copy to the request read buffer. +*/ +if (copy_to_user (buf, (u8 *)addr + offset, read_count)) +return -EFAULT ; + +/* +* Only the device memory present on the hardware is mapped, which may +* not be power-of-2 aligned. A read to the BAR2 region implies an +* access outside the available device memory on the hardware. Fill +* such read request with ~1. +*/ +for (i =0 ; i < count - read_count; i++) +if (copy_to_user (buf + read_count + i, &val,1 )) +return -EFAULT ; + +return count; +} + +static ssize_t nvgrace_gpu_vfio_pci_read (struct vfio_device *core_vdev, +char __user *buf,size_t count,loff_t *ppos) +{ +unsigned int index =VFIO_PCI_OFFSET_TO_INDEX (*ppos); +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + } + +return nvgrace_gpu_read_mem (buf, count, ppos, nvdev->opregion, + nvdev->mem_length,roundup_pow_of_two (nvdev->mem_length)); + } + +return vfio_pci_core_read (core_vdev, buf, count, ppos); +} + +/* +* Write count bytes to the device memory at a given offset. The actual device +* memory size (available) may not be a power-of-2. So the driver fakes the +* size to a power-of-2 (reported) when exposing to a user space driver. +* +* Write request beyond the actual device size are dropped, while those +* beyond the actual reported size are skipped entirely. +* +* A write to a reported+ offset is considered error conditions and +* returned with an -EINVAL. Overflow conditions are also handled. +*/ +ssize_t nvgrace_gpu_write_mem (const void *addr,size_t count,loff_t *ppos, +const void __user *buf,size_t available,size_t reported) +{ +u64 offset = *ppos &VFIO_PCI_OFFSET_MASK ; +u64 end; +size_t write_count; + +if (offset >= reported) +return -EINVAL ; + +if (check_add_overflow (offset, count, &end)) +return -EOVERFLOW ; + +/* Clip short the read request beyond reported BAR size */ +if (end >= reported) + count = reported - offset; + +/* +* Determine how many bytes to be actually written to the device memory. +* Do not write to the offset beyond available size. +*/ +if (offset >= available) + write_count =0 ; +else + write_count =min (count, available - (size_t )offset); + +/* +* Only the device memory present on the hardware is mapped, which may +* not be power-of-2 aligned. A write to the BAR2 region implies an +* access outside the available device memory on the hardware. Drop +* those write requests. +*/ +if (copy_from_user ((u8 *)addr + offset, buf, write_count)) +return -EFAULT ; + +return count; +} + +static ssize_t nvgrace_gpu_vfio_pci_write (struct vfio_device *core_vdev, +const char __user *buf,size_t count,loff_t *ppos) +{ +unsigned int index =VFIO_PCI_OFFSET_TO_INDEX (*ppos); +struct nvgrace_gpu_vfio_pci_core_device *nvdev =container_of ( + core_vdev, struct nvgrace_gpu_vfio_pci_core_device, core_device.vdev); + +if (index ==VFIO_PCI_BAR2_REGION_INDEX ) { +if (!nvdev->opregion) { + nvdev->opregion =memremap (nvdev->hpa, nvdev->mem_length,MEMREMAP_WB ); +if (!nvdev->opregion) +return -ENOMEM ; + } + +return nvgrace_gpu_write_mem (nvdev->opregion, count, ppos, buf, + nvdev->mem_length,roundup_pow_of_two (nvdev->mem_length)); + } + +return vfio_pci_core_write (core_vdev, buf, count, ppos); +} + +static const struct vfio_device_ops nvgrace_gpu_vfio_pci_ops = { + .name ="nvgrace-gpu-vfio-pci" , + .init = vfio_pci_core_init_dev, + .release = vfio_pci_core_release_dev, + .open_device = nvgrace_gpu_vfio_pci_open_device, + .close_device = nvgrace_gpu_vfio_pci_close_device, + .ioctl = nvgrace_gpu_vfio_pci_ioctl, + .read = nvgrace_gpu_vfio_pci_read, + .write = nvgrace_gpu_vfio_pci_write, + .mmap = nvgrace_gpu_vfio_pci_mmap, + .request = vfio_pci_core_request, + .match = vfio_pci_core_match, + .bind_iommufd = vfio_iommufd_physical_bind, + .unbind_iommufd = vfio_iommufd_physical_unbind, + .attach_ioas = vfio_iommufd_physical_attach_ioas, +}; + +static struct +nvgrace_gpu_vfio_pci_core_device *nvgrace_gpu_drvdata (struct pci_dev *pdev) +{ +struct vfio_pci_core_device *core_device =dev_get_drvdata (&pdev->dev); + + return container_of(core_device, struct nvgrace_gpu_vfio_pci_core_device, + core_device); +} + +static int +nvgrace_gpu_vfio_pci_fetch_memory_property (struct pci_dev *pdev, +struct nvgrace_gpu_vfio_pci_core_device *nvdev) +{ +int ret; + +/* +* The memory information is present in the system ACPI tables as DSD +* properties nvidia,gpu-mem-base-pa and nvidia,gpu-mem-size. +*/ + ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-base-pa" , + &(nvdev->hpa)); +if (ret) +return ret; + + ret =device_property_read_u64 (&pdev->dev,"nvidia,gpu-mem-size" , + &(nvdev->mem_length)); +return ret; +} + +static int nvgrace_gpu_vfio_pci_probe (struct pci_dev *pdev, +const struct pci_device_id *id) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev; +int ret; + + nvdev =vfio_alloc_device (nvgrace_gpu_vfio_pci_core_device, core_device.vdev, + &pdev->dev, &nvgrace_gpu_vfio_pci_ops); +if (IS_ERR (nvdev)) +return PTR_ERR (nvdev); + +dev_set_drvdata (&pdev->dev, nvdev); + + ret =nvgrace_gpu_vfio_pci_fetch_memory_property (pdev, nvdev); +if (ret) + goto out_put_vdev; + + ret =vfio_pci_core_register_device (&nvdev->core_device); +if (ret) + goto out_put_vdev; + +return ret; + +out_put_vdev: +vfio_put_device (&nvdev->core_device.vdev); +return ret; +} + +static void nvgrace_gpu_vfio_pci_remove (struct pci_dev *pdev) +{ +struct nvgrace_gpu_vfio_pci_core_device *nvdev =nvgrace_gpu_drvdata (pdev); +struct vfio_pci_core_device *vdev = &nvdev->core_device; + +vfio_pci_core_unregister_device (vdev); +vfio_put_device (&vdev->vdev); +} + +static const struct pci_device_id nvgrace_gpu_vfio_pci_table[] = { +/* GH200 120GB */ + {PCI_DRIVER_OVERRIDE_DEVICE_VFIO (PCI_VENDOR_ID_NVIDIA ,0x2342 ) }, +/* GH200 480GB */ + {PCI_DRIVER_OVERRIDE_DEVICE_VFIO (PCI_VENDOR_ID_NVIDIA ,0x2345 ) }, + {} +}; + +MODULE_DEVICE_TABLE (pci, nvgrace_gpu_vfio_pci_table); + +static struct pci_driver nvgrace_gpu_vfio_pci_driver = { + .name =KBUILD_MODNAME , + .id_table = nvgrace_gpu_vfio_pci_table, + .probe = nvgrace_gpu_vfio_pci_probe, + .remove = nvgrace_gpu_vfio_pci_remove, + .err_handler = &vfio_pci_core_err_handlers, + .driver_managed_dma = true, +}; + +module_pci_driver (nvgrace_gpu_vfio_pci_driver); + +MODULE_LICENSE ("GPL v2" ); +MODULE_AUTHOR ("Ankit Agrawal <[email protected]>" ); +MODULE_AUTHOR ("Aniket Agashe <[email protected]>" ); +MODULE_DESCRIPTION ( +"VFIO NVGRACE GPU PF - User Level driver for NVIDIA devices with CPU coherently accessible device memory" );Signature · 1 line
2.17.1
---end quoted text---
Jason Gunthorpe
reviewerQuoted 3 lines · Christoph Hellwig
I know I'm chiming in a bit late, but what ultimate user space is going to use this? We should not add anything to the kernel that can't be used without fully open user space.
qemu will get the matching VFIO userspace patches, I think they were posted someplace already.
Quoted 6 lines
vfio has traditionally been a bit special as it "just" passes devices through, so any user space could just be a user space driver for a random device on $FOO bus, including an actual Linux driver in a VM, but this driver has very specific semantics for a very specific piece of hardware, so it really needs to be treated like a generic GPU driver or accelerator driver.
This is basically a pre-CXL driver. It takes a PCI device and some non-standard CXL-ish metadata and adapts it to VFIO.
In a post-CXL world this same functionality of managing the 'cache coherent BAR' for VFIO would be done generically by some generic vfio-cxl driver.
Jason
Quoted 7 lines · Jason Gunthorpe
Quoted 3 lines · Christoph Hellwig
I know I'm chiming in a bit late, but what ultimate user space is going to use this? We should not add anything to the kernel that can't be used without fully open user space.
qemu will get the matching VFIO userspace patches, I think they were posted someplace already.
Well, that's not what I mean with full userspace. Whats the actual consumer running in a qemu VM here?
Ankit Agrawal
maintainerHi Christoph,
Quoted 1 line
Whats the actual consumer running in a qemu VM here?
The primary use case in the VM is to run the open source Nvidia driver (https://github.com/NVIDIA/open-gpu-kernel-modules)
and workloads.
Quoted 6 lines · Ankit Agrawal
Hi Christoph,
Quoted 1 line
Whats the actual consumer running in a qemu VM here?
The primary use case in the VM is to run the open source Nvidia driver (https://github.com/NVIDIA/open-gpu-kernel-modules)
and workloads.
So this infrastructure to run things in a VM that we don't even support in mainline? I think we need nouveau support for this hardware in the drm driver first, before adding magic vfio support.
Quoted 11 lines · On Thu, 31 Aug 2023 07:04:10 -0700 Christoph Hellwig <[email protected]> wrote:
Quoted 6 lines · Ankit Agrawal
Hi Christoph,
Quoted 1 line
Whats the actual consumer running in a qemu VM here?
The primary use case in the VM is to run the open source Nvidia driver (https://github.com/NVIDIA/open-gpu-kernel-modules)
and workloads.So this infrastructure to run things in a VM that we don't even support in mainline? I think we need nouveau support for this hardware in the drm driver first, before adding magic vfio support.
There's really never a guarantee that the thing we're exposing via the vfio uAPI has mainline drivers, for example we don't consult the nouveau device table before we expose an NVIDIA GPU to a Windows guest running proprietary device drivers.
We've also never previously made a requirement that any new code in vfio must directly contribute to supporting a mainline driver, in fact I think you'll find examples where we do have such code.
This driver is proposing to expose a coherent memory region associated with the device, composed as a PCI BAR, largely to bring it into the vfio device model. Access to that memory region is still pass-through. This is essentially behavior that we also enable though mdev drivers like kvmgt (modulo the coherent aspect).
I assume the above driver understands how to access and make use of this coherent memory whether running bare-metal or virtualized, so potentially we have some understanding of how it's used by the driver, which can't be said for all devices used with vfio. I'm therefore not sure how we can suddenly decide to impose a mainline driver requirement for exposing a device to userspace. Thanks,
Alex
Jason Gunthorpe
reviewerQuoted 6 lines · Alex Williamson
I assume the above driver understands how to access and make use of this coherent memory whether running bare-metal or virtualized, so potentially we have some understanding of how it's used by the driver, which can't be said for all devices used with vfio. I'm therefore not sure how we can suddenly decide to impose a mainline driver requirement for exposing a device to userspace. Thanks,
Yeah, I was comfortable with removing the old powernv VFIO stuff based on the combined logic that the platform was dead, powernv has weird arch entanglements and there was no open source driver anyhow so maintaining the mess past the vendor lifetime was looking bad. This has none of those issues.
I think the threshold here should be the maintainability of the kernel and its associated open ecosystem. An open source qemu, and a open source VM kernel driver is a pretty good situation to sustain this driver. In particular if Alex&co think the qemu side should not advance then this should not be merged.
For comparison, I'm much more unhappy about VFIO_UPDATE_VADDR from a maintainability perspective than I am about this. There was a half hearted effort to get the userspace side in qemu and now we are now stuck with an ugly kernel side mess with no open source userspace so we can't even test. :\ Considering the vigorous objections when I tried to remove it I assume a cloud operator is running it with a proprietary userspace.
Certainly I would strongly support removing kernel side parts if the qemu side doesn't get merged within a year or something like that.
Jason
Quoted 2 lines · Jason Gunthorpe
Certainly I would strongly support removing kernel side parts if the qemu side doesn't get merged within a year or something like that.
That should go away as well, yes.
Quoted 4 lines · Alex Williamson
There's really never a guarantee that the thing we're exposing via the vfio uAPI has mainline drivers, for example we don't consult the nouveau device table before we expose an NVIDIA GPU to a Windows guest running proprietary device drivers.
As I said in my initial mail there's a huge different between a simple passthrough like vfio-pci that doesn't care about the actual device, and a driver like this one that is completely device specific.