From: Nathan Chen <nathanc@nvidia.com>

Open VFIO FDs from libvirt backend without exposing
these FDs to XML users, i.e. one per iommufd hostdev
for /dev/vfio/devices/vfioX, and pass the FD to qemu
command line.

Suggested-by: Ján Tomko <jtomko@redhat.com>
Signed-off-by: Nathan Chen <nathanc@nvidia.com>
---
 src/libvirt_private.syms |  1 +
 src/qemu/qemu_command.c  | 21 +++++++++++
 src/qemu/qemu_process.c  | 79 ++++++++++++++++++++++++++++++++++++++++
 src/util/virpci.c        | 42 +++++++++++++++++++++
 src/util/virpci.h        |  2 +
 5 files changed, 145 insertions(+)

diff --git a/src/libvirt_private.syms b/src/libvirt_private.syms
index a8eadbfb8a..0904265459 100644
--- a/src/libvirt_private.syms
+++ b/src/libvirt_private.syms
@@ -3163,6 +3163,7 @@ virPCIDeviceGetStubDriverName;
 virPCIDeviceGetStubDriverType;
 virPCIDeviceGetUnbindFromStub;
 virPCIDeviceGetUsedBy;
+virPCIDeviceGetVfioPath;
 virPCIDeviceGetVPD;
 virPCIDeviceHasPCIExpressLink;
 virPCIDeviceIsAssignable;
diff --git a/src/qemu/qemu_command.c b/src/qemu/qemu_command.c
index 98e4469c25..2a16f9df63 100644
--- a/src/qemu/qemu_command.c
+++ b/src/qemu/qemu_command.c
@@ -4809,6 +4809,18 @@ qemuBuildPCIHostdevDevProps(const virDomainDef *def,
                               NULL) < 0)
         return NULL;
 
+    if (pcisrc->driver.name == VIR_DEVICE_HOSTDEV_PCI_DRIVER_NAME_VFIO &&
+        pcisrc->driver.iommufd == VIR_TRISTATE_BOOL_YES) {
+        qemuDomainHostdevPrivate *hostdevPriv = QEMU_DOMAIN_HOSTDEV_PRIVATE(dev);
+
+        if (hostdevPriv->vfioDeviceFd != -1) {
+            g_autofree char *fdstr = g_strdup_printf("%d", hostdevPriv->vfioDeviceFd);
+            if (virJSONValueObjectAdd(&props, "S:fd", fdstr, NULL) < 0)
+                return NULL;
+            hostdevPriv->vfioDeviceFd = -1;
+        }
+    }
+
     if (qemuBuildDeviceAddressProps(props, def, dev->info) < 0)
         return NULL;
 
@@ -5253,6 +5265,15 @@ qemuBuildHostdevCommandLine(virCommand *cmd,
             if (qemuCommandAddExtDevice(cmd, hostdev->info, def, qemuCaps) < 0)
                 return -1;
 
+            if (subsys->u.pci.driver.iommufd == VIR_TRISTATE_BOOL_YES) {
+                qemuDomainHostdevPrivate *hostdevPriv = QEMU_DOMAIN_HOSTDEV_PRIVATE(hostdev);
+
+                if (hostdevPriv->vfioDeviceFd != -1) {
+                    virCommandPassFD(cmd, hostdevPriv->vfioDeviceFd,
+                                     VIR_COMMAND_PASS_FD_CLOSE_PARENT);
+                }
+            }
+
             if (!(devprops = qemuBuildPCIHostdevDevProps(def, hostdev)))
                 return -1;
 
diff --git a/src/qemu/qemu_process.c b/src/qemu/qemu_process.c
index 0e50cd1ccc..ab88a6bf62 100644
--- a/src/qemu/qemu_process.c
+++ b/src/qemu/qemu_process.c
@@ -103,6 +103,7 @@
 #include "storage_source.h"
 #include "backup_conf.h"
 #include "storage_file_probe.h"
+#include "virpci.h"
 
 #include "logging/log_manager.h"
 #include "logging/log_protocol.h"
@@ -8181,6 +8182,9 @@ qemuProcessLaunch(virConnectPtr conn,
     if (qemuExtDevicesStart(driver, vm, incomingMigrationExtDevices) < 0)
         goto cleanup;
 
+    if (qemuProcessOpenVfioFds(vm) < 0)
+        goto cleanup;
+
     if (!(cmd = qemuBuildCommandLine(vm,
                                      incoming ? "defer" : NULL,
                                      vmop,
@@ -10360,3 +10364,78 @@ qemuProcessHandleNbdkitExit(qemuNbdkitProcess *nbdkit,
     qemuProcessEventSubmit(vm, QEMU_PROCESS_EVENT_NBDKIT_EXITED, 0, 0, nbdkit);
     virObjectUnlock(vm);
 }
+
+/**
+ * qemuProcessOpenVfioDeviceFd:
+ * @hostdev: host device definition
+ * @vfioFd: returned file descriptor
+ *
+ * Opens the VFIO device file descriptor for a hostdev.
+ *
+ * Returns: FD on success, -1 on failure
+ */
+static int
+qemuProcessOpenVfioDeviceFd(virDomainHostdevDef *hostdev)
+{
+    g_autofree char *vfioPath = NULL;
+    int fd = -1;
+
+    if (hostdev->mode != VIR_DOMAIN_HOSTDEV_MODE_SUBSYS ||
+        hostdev->source.subsys.type != VIR_DOMAIN_HOSTDEV_SUBSYS_TYPE_PCI) {
+        virReportError(VIR_ERR_INTERNAL_ERROR, "%s",
+                       _("VFIO FD only supported for PCI hostdevs"));
+        return -1;
+    }
+
+    if (virPCIDeviceGetVfioPath(&hostdev->source.subsys.u.pci.addr, &vfioPath) < 0)
+        return -1;
+
+    VIR_DEBUG("Opening VFIO device %s", vfioPath);
+
+    if ((fd = open(vfioPath, O_RDWR | O_CLOEXEC)) < 0) {
+        if (errno == ENOENT) {
+            virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
+                           _("VFIO device %1$s not found - ensure device is bound to vfio-pci driver"),
+                           vfioPath);
+        } else {
+            virReportSystemError(errno,
+                                 _("cannot open VFIO device %1$s"), vfioPath);
+        }
+        return -1;
+    }
+
+    VIR_DEBUG("Opened VFIO device FD %d for %s", fd, vfioPath);
+    return fd;
+}
+
+/**
+ * qemuProcessOpenVfioFds:
+ * @vm: domain object
+ *
+ * Opens all necessary VFIO file descriptors for the domain.
+ *
+ * Returns: 0 on success, -1 on failure
+ */
+int
+qemuProcessOpenVfioFds(virDomainObj *vm)
+{
+    size_t i;
+
+    /* Check if we have any hostdevs that need VFIO FDs */
+    for (i = 0; i < vm->def->nhostdevs; i++) {
+        virDomainHostdevDef *hostdev = vm->def->hostdevs[i];
+        qemuDomainHostdevPrivate *hostdevPriv = QEMU_DOMAIN_HOSTDEV_PRIVATE(hostdev);
+
+        if (hostdev->mode == VIR_DOMAIN_HOSTDEV_MODE_SUBSYS &&
+            hostdev->source.subsys.type == VIR_DOMAIN_HOSTDEV_SUBSYS_TYPE_PCI &&
+            hostdev->source.subsys.u.pci.driver.name == VIR_DEVICE_HOSTDEV_PCI_DRIVER_NAME_VFIO &&
+            hostdev->source.subsys.u.pci.driver.iommufd == VIR_TRISTATE_BOOL_YES) {
+            /* Open VFIO device FD */
+            hostdevPriv->vfioDeviceFd = qemuProcessOpenVfioDeviceFd(hostdev);
+            if (hostdevPriv->vfioDeviceFd == -1)
+                 return -1;
+        }
+    }
+
+    return 0;
+}
diff --git a/src/util/virpci.c b/src/util/virpci.c
index 90617e69c6..886e2c55e6 100644
--- a/src/util/virpci.c
+++ b/src/util/virpci.c
@@ -3320,3 +3320,45 @@ virPCIDeviceAddressFree(virPCIDeviceAddress *address)
 {
     g_free(address);
 }
+
+/**
+ * virPCIDeviceGetVfioPath:
+ * @addr: host device PCI address
+ * @vfioPath: returned VFIO device path
+ *
+ * Constructs the VFIO device path for a PCI hostdev.
+ *
+ * Returns: 0 on success, -1 on failure
+ */
+int
+virPCIDeviceGetVfioPath(virPCIDeviceAddress *addr,
+                        char **vfioPath)
+{
+    g_autofree char *addrStr = NULL;
+
+    *vfioPath = NULL;
+    addrStr = virPCIDeviceAddressAsString(addr);
+
+    /* First try: Direct lookup in device's vfio-dev subdirectory */
+    {
+        g_autofree char *sysfsPath = NULL;
+        g_autoptr(DIR) dir = NULL;
+        struct dirent *entry = NULL;
+
+        sysfsPath = g_strdup_printf("/sys/bus/pci/devices/%s/vfio-dev/", addrStr);
+
+        if (virDirOpen(&dir, sysfsPath) == 1) {
+            while (virDirRead(dir, &entry, sysfsPath) > 0) {
+                if (STRPREFIX(entry->d_name, "vfio")) {
+                    *vfioPath = g_strdup_printf("/dev/vfio/devices/%s", entry->d_name);
+                    return 0;
+                }
+            }
+        }
+    }
+
+    virReportError(VIR_ERR_INTERNAL_ERROR,
+                  _("cannot find VFIO device for PCI device %1$s"),
+                  addrStr);
+    return -1;
+}
diff --git a/src/util/virpci.h b/src/util/virpci.h
index fc538566e1..24ede10755 100644
--- a/src/util/virpci.h
+++ b/src/util/virpci.h
@@ -296,6 +296,8 @@ void virPCIEDeviceInfoFree(virPCIEDeviceInfo *dev);
 
 void virPCIDeviceAddressFree(virPCIDeviceAddress *address);
 
+int virPCIDeviceGetVfioPath(virPCIDeviceAddress *addr, char **vfioPath);
+
 G_DEFINE_AUTOPTR_CLEANUP_FUNC(virPCIDevice, virPCIDeviceFree);
 G_DEFINE_AUTOPTR_CLEANUP_FUNC(virPCIDeviceAddress, virPCIDeviceAddressFree);
 G_DEFINE_AUTOPTR_CLEANUP_FUNC(virPCIEDeviceInfo, virPCIEDeviceInfoFree);
-- 
2.43.0