Xu Yilun <yilun.xu@xxxxxxxxxxxxxxx> writes: > On Fri, Feb 21, 2025 at 01:43:28PM +0530, Aneesh Kumar K.V wrote: >> Alexey Kardashevskiy <aik@xxxxxxx> writes: >> >> .... >> >> > >> > I am trying to wrap my head around your tsm. here is what I got in my tree: >> > https://github.com/aik/linux/blob/tsm/include/linux/tsm.h >> > >> > Shortly: >> > >> > drivers/virt/coco/tsm.ko does sysfs (including "connect" and "bind" to >> > control and "certs"/"report" to attest) and implements tsm_dev/tsm_tdi, >> > it does not know pci_dev; >> > >> > drivers/pci/tsm-pci.ko creates/destroys tsm_dev/tsm_dev using tsm.ko; >> > >> > drivers/crypto/ccp/ccp.ko (the PSP guy) registers: >> > - tsm_subsys in tsm.ko (which does "connect" and "bind" and >> > - tsm_bus_subsys in tsm-pci.ko (which does "spdm_forward") >> > ccp.ko knows about pci_dev and whatever else comes in the future, and >> > ccp.ko's "connect" implementation calls the IDE library (I am adopting >> > yours now, with some tweaks). >> > >> > tsm-dev and tsm-tdi embed struct dev each and are added as children to >> > PCI devices: no hide/show attrs, no additional TSM pointer in struct >> > device or pci_dev, looks like: >> > >> > aik@sc ~> ls /sys/bus/pci/devices/0000:e1:04.0/tsm-tdi/tdi:0000:e1:04.0/ >> > device power subsystem tsm_report tsm_report_user tsm_tdi_bind >> > tsm_tdi_status tsm_tdi_status_user uevent >> > >> > aik@sc ~> ls /sys/bus/pci/devices/0000:e1:04.0/tsm_dev/ >> > device power subsystem tsm_certs tsm_cert_slot tsm_certs_user >> > tsm_dev_connect tsm_dev_status tsm_meas tsm_meas_user uevent >> > >> > aik@sc ~> ls /sys/class/tsm/tsm0/ >> > device power stream0:0000:e1:00.0 subsystem uevent >> > >> > aik@sc ~> ls /sys/class/tsm-dev/ >> > tdev:0000:c0:01.1 tdev:0000:e0:01.1 tdev:0000:e1:00.0 >> > >> > aik@sc ~> ls /sys/class/tsm-tdi/ >> > tdi:0000:c0:01.1 tdi:0000:e0:01.1 tdi:0000:e1:00.0 tdi:0000:e1:04.0 >> > tdi:0000:e1:04.1 tdi:0000:e1:04.2 tdi:0000:e1:04.3 >> > >> > >> > SPDM forwarding seems a bus-agnostic concept, "connect" is a PCI thing >> > but pci_dev is only needed for DOE/IDE. >> > >> > Or is separating struct pci_dev from struct device not worth it and most >> > of it should go to tsm-pci.ko? Then what is left for tsm.ko? Thanks, >> > >> >> For the Arm CCA DA, I have structured the flow as follows. I am >> currently refining my changes to prepare them for posting. I am using >> tsm-core in both the host and guest. There is no bind interface at the >> sysfs level; instead, it is managed via the KVM ioctl >> >> Host: >> step 1. >> echo ${DEVICE} > /sys/bus/pci/devices/${DEVICE}/driver/unbind >> echo vfio-pci > /sys/bus/pci/devices/${DEVICE}/driver_override >> echo ${DEVICE} > /sys/bus/pci/drivers_probe >> >> step 2. >> echo 1 > /sys/bus/pci/devices/$DEVICE/tsm/connect >> >> step 3. >> using VMM to make the new KVM_SET_DEVICE_ATTR ioctl >> >> + dev_num = vfio_devices[i].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); >> + >> + struct kvm_vfio_tsm_bind param = { >> + .guest_rid = guest_bdf, >> + .devfd = vfio_devices[i].fd, >> + }; >> + struct kvm_device_attr attr = { >> + .group = KVM_DEV_VFIO_DEVICE, >> + .attr = KVM_DEV_VFIO_DEVICE_TDI_BIND, >> + .addr = (__u64)¶m, >> + }; >> + >> + if (ioctl(kvm_vfio_device, KVM_SET_DEVICE_ATTR, &attr)) { >> + pr_err("Failed KVM_SET_DEVICE_ATTR for KVM_DEV_VFIO_DEVICE"); >> + return -ENODEV; >> + } >> + > > I think bind (which brings device to a LOCKED state, no MMIO, no DMA) > cannot be a driver agnostic behavior. So I think it should be a VFIO > ioctl. > For the current CCA implementation bind is equivalent to VDEV_CREATE which doesn't mark the device LOCKED. Marking the device LOCKED is driven by the guest as shown in the steps below. >> >> Now in the guest we follow the below steps >> >> step 1: >> echo ${DEVICE} > /sys/bus/pci/devices/${DEVICE}/driver/unbind >> >> step 2: Move the device to TDISP LOCK state >> echo 1 > /sys/bus/pci/devices/0000:00:00.0/tsm/connect >> echo 3 > /sys/bus/pci/devices/0000:00:00.0/tsm/connect > > Reuse the 'connect' interface? I think it conceptually brings chaos. Is > it better we create a new interface? > I was looking at converting these numbers to strings. "1" -> connect "2" -> lock "3" -> run > >> >> step 3: Moves the device to TDISP RUN state >> echo 4 > /sys/bus/pci/devices/0000:00:00.0/tsm/connect > > Could you elaborate what '1'/'3'/'4' stand for? > As mentioned above, them move the device to different TDISP state. I will reply to this patch with my early RFC chnages for tsm framework. I am not yet ready to share the CCA backend changes. But I assume having the tsm framework changes alone can be useful? -aneesh
