On 24/07/2025 15:09, Aneesh Kumar K.V wrote:
Will Deacon <will@xxxxxxxxxx> writes:
On Sun, May 25, 2025 at 01:19:15PM +0530, Aneesh Kumar K.V (Arm) wrote:
This also allocates a stage1 bypass and stage2 translate table.
Please write your commit messages as per Linux kernel guidelines.
Signed-off-by: Aneesh Kumar K.V (Arm) <aneesh.kumar@xxxxxxxxxx>
---
builtin-run.c | 2 +
include/kvm/kvm-config.h | 1 +
vfio/core.c | 4 +-
vfio/iommufd.c | 115 ++++++++++++++++++++++++++++++++++++++-
[...]
4 files changed, 119 insertions(+), 3 deletions(-)
diff --git a/vfio/iommufd.c b/vfio/iommufd.c
index 742550705746..39870320e4ac 100644
--- a/vfio/iommufd.c
+++ b/vfio/iommufd.c
@@ -108,6 +108,116 @@ err_out:
return ret;
}
+static int iommufd_alloc_s1bypass_hwpt(struct vfio_device *vdev)
+{
+ int ret;
+ unsigned long dev_num;
+ unsigned long guest_bdf;
+ struct vfio_device_bind_iommufd bind;
+ struct vfio_device_attach_iommufd_pt attach_data;
+ struct iommu_hwpt_alloc alloc_hwpt;
+ struct iommu_viommu_alloc alloc_viommu;
+ struct iommu_hwpt_arm_smmuv3 bypass_ste;
+ struct iommu_vdevice_alloc alloc_vdev;
+
+ bind.argsz = sizeof(bind);
+ bind.flags = 0;
+ bind.iommufd = iommu_fd;
+
+ /* now bind the iommufd */
+ if (ioctl(vdev->fd, VFIO_DEVICE_BIND_IOMMUFD, &bind)) {
+ ret = -errno;
+ vfio_dev_err(vdev, "failed to get info");
+ goto err_out;
+ }
+
+ alloc_hwpt.size = sizeof(struct iommu_hwpt_alloc);
+ alloc_hwpt.flags = IOMMU_HWPT_ALLOC_NEST_PARENT;
+ alloc_hwpt.dev_id = bind.out_devid;
+ alloc_hwpt.pt_id = ioas_id;
+ alloc_hwpt.data_type = IOMMU_HWPT_DATA_NONE;
+ alloc_hwpt.data_len = 0;
+ alloc_hwpt.data_uptr = 0;
+
+ if (ioctl(iommu_fd, IOMMU_HWPT_ALLOC, &alloc_hwpt)) {
+ ret = -errno;
+ pr_err("Failed to allocate HWPT");
+ goto err_out;
+ }
+
+ attach_data.argsz = sizeof(attach_data);
+ attach_data.flags = 0;
+ attach_data.pt_id = alloc_hwpt.out_hwpt_id;
+
+ if (ioctl(vdev->fd, VFIO_DEVICE_ATTACH_IOMMUFD_PT, &attach_data)) {
+ ret = -errno;
+ vfio_dev_err(vdev, "failed to attach to IOAS ");
+ goto err_out;
+ }
+
+ alloc_viommu.size = sizeof(alloc_viommu);
+ alloc_viommu.flags = 0;
+ alloc_viommu.type = IOMMU_VIOMMU_TYPE_ARM_SMMUV3;
+ alloc_viommu.dev_id = bind.out_devid;
+ alloc_viommu.hwpt_id = alloc_hwpt.out_hwpt_id;
+
+ if (ioctl(iommu_fd, IOMMU_VIOMMU_ALLOC, &alloc_viommu)) {
+ ret = -errno;
+ vfio_dev_err(vdev, "failed to allocate VIOMMU %d", ret);
+ goto err_out;
+ }
+#define STRTAB_STE_0_V (1UL << 0)
+#define STRTAB_STE_0_CFG_S2_TRANS 6
+#define STRTAB_STE_0_CFG_S1_TRANS 5
+#define STRTAB_STE_0_CFG_BYPASS 4
+
+ /* set up virtual ste as bypass ste */
+ bypass_ste.ste[0] = STRTAB_STE_0_V | (STRTAB_STE_0_CFG_BYPASS << 1);
+ bypass_ste.ste[1] = 0x0UL;
+
+ alloc_hwpt.size = sizeof(struct iommu_hwpt_alloc);
+ alloc_hwpt.flags = 0;
+ alloc_hwpt.dev_id = bind.out_devid;
+ alloc_hwpt.pt_id = alloc_viommu.out_viommu_id;
+ alloc_hwpt.data_type = IOMMU_HWPT_DATA_ARM_SMMUV3;
+ alloc_hwpt.data_len = sizeof(bypass_ste);
+ alloc_hwpt.data_uptr = (unsigned long)&bypass_ste;
+
+ if (ioctl(iommu_fd, IOMMU_HWPT_ALLOC, &alloc_hwpt)) {
+ ret = -errno;
+ pr_err("Failed to allocate S1 bypass HWPT %d", ret);
+ goto err_out;
+ }
+
+ alloc_vdev.size = sizeof(alloc_vdev),
+ alloc_vdev.viommu_id = alloc_viommu.out_viommu_id;
+ alloc_vdev.dev_id = bind.out_devid;
+
+ dev_num = vdev->dev_hdr.dev_num;
+ /* kvmtool only do 0 domain, 0 bus and 0 function devices. */
+ guest_bdf = (0ULL << 32) | (0 << 16) | dev_num << 11 | (0 << 8);
I don't understand this. Shouldn't the BDF correspond to the virtual
configuration space? That's not allocated until later, but just going
with 0 isn't going to work.
What am I missing?
As I understand it, kvmtool supports only bus 0 and does not allow
multifunction devices. Based on that, I derived the guest BDF as follows
(correcting what was wrong in the original patch):
guest_bdf = (0ULL << 16) | (0 << 8) | dev_num << 3 | (0 << 0);
Are you suggesting that this approach is incorrect, and that we can use
a bus number other than 0?
To put this other way, the emulation of the configuration space is based
on the "dev_num". i.e., CFG address is converted to the offset and
mapped to the "dev_num". So I think what we have here is correct.
Suzuki
