On 7/25/19 8:09 PM, Daniel Henrique Barboza wrote:
From: Shivaprasad G Bhat <sbhat(a)linux.vnet.ibm.com>
The iommu group, /dev/vfio/<iommuGroupNumber> behaviors
of the host are mocked. This patch implements support for
multifunction/multiple devices per iommu groups and emulates
the /dev/vfio/<iommuGroupNumber> file correctly.
This code helps adding necessary testcases for pci-hotplug
code.
Signed-off-by: Shivaprasad G Bhat <sbhat(a)linux.vnet.ibm.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413(a)gmail.com>
---
tests/virpcimock.c | 177 +++++++++++++++++++++++++++++++++++++++++----
1 file changed, 162 insertions(+), 15 deletions(-)
Looks good. We will have some merge conflicts because we fix the same issue.
diff --git a/tests/virpcimock.c b/tests/virpcimock.c
index 5ec9abe05f..1b44ed0a44 100644
--- a/tests/virpcimock.c
+++ b/tests/virpcimock.c
@@ -44,6 +44,8 @@ char *fakerootdir;
char *fakesysfspcidir;
# define SYSFS_PCI_PREFIX "/sys/bus/pci/"
+# define SYSFS_KERNEL_PREFIX "/sys/kernel/"
+
# define STDERR(...) \
fprintf(stderr, "%s %zu: ", __FUNCTION__, (size_t) __LINE__); \
@@ -123,6 +125,11 @@ struct pciDevice {
struct pciDriver *driver; /* Driver attached. NULL if attached to no driver */
};
+struct pciIommuGroup {
+ int iommu;
+ size_t nDevicesBoundToVFIO; /* Indicates the devices in the group */
+};
+
struct fdCallback {
int fd;
char *path;
@@ -134,6 +141,9 @@ size_t nPCIDevices = 0;
struct pciDriver **pciDrivers = NULL;
size_t nPCIDrivers = 0;
+struct pciIommuGroup **pciIommuGroups = NULL;
+size_t npciIommuGroups = 0;
+
struct fdCallback *callbacks = NULL;
size_t nCallbacks = 0;
@@ -247,6 +257,13 @@ getrealpath(char **newpath,
errno = ENOMEM;
return -1;
}
+ } else if (STRPREFIX(path, SYSFS_KERNEL_PREFIX)) {
+ if (virAsprintfQuiet(newpath, "%s/%s",
+ fakerootdir,
+ path) < 0) {
+ errno = ENOMEM;
+ return -1;
+ }
} else {
if (VIR_STRDUP_QUIET(*newpath, path) < 0)
return -1;
@@ -460,6 +477,101 @@ pci_device_autobind(struct pciDevice *dev)
return pci_driver_bind(driver, dev);
}
+static void
+pci_iommu_new(int num)
+{
+ char *iommupath, *kerneldir;
+ struct pciIommuGroup *iommuGroup;
+
+ if (VIR_ALLOC_QUIET(iommuGroup) < 0)
+ ABORT_OOM();
+
+ iommuGroup->iommu = num;
+ iommuGroup->nDevicesBoundToVFIO = 0; /* No device bound to vfio by default */
+
+ if (virAsprintfQuiet(&kerneldir, "%s%s",
+ fakerootdir, SYSFS_KERNEL_PREFIX) < 0)
+ ABORT_OOM();
+
+ if (virAsprintfQuiet(&iommupath, "%s/iommu_groups/%d/devices",
kerneldir, num) < 0)
+ ABORT_OOM();
+ VIR_FREE(kerneldir);
+
+ if (virFileMakePath(iommupath) < 0)
+ ABORT("Unable to create: %s", iommupath);
+ VIR_FREE(iommupath);
+
+ if (VIR_APPEND_ELEMENT_QUIET(pciIommuGroups, npciIommuGroups, iommuGroup) < 0)
+ ABORT_OOM();
+}
+
+static int
+pci_vfio_release_iommu(struct pciDevice *device)
+{
+ char *vfiopath = NULL;
+ int ret = -1;
+ size_t i = 0;
+
+ for (i = 0; i < npciIommuGroups; i++) {
+ if (pciIommuGroups[i]->iommu == device->iommuGroup)
+ break;
+ }
+
+ if (i != npciIommuGroups) {
+ if (pciIommuGroups[i]->nDevicesBoundToVFIO == 0) {
+ ret = 0;
+ goto cleanup;
+ }
+ pciIommuGroups[i]->nDevicesBoundToVFIO--;
+ if (!pciIommuGroups[i]->nDevicesBoundToVFIO) {
+ if (virAsprintfQuiet(&vfiopath, "%s/dev/vfio/%d",
+ fakesysfspcidir, device->iommuGroup) < 0) {
+ errno = ENOMEM;
+ goto cleanup;
+ }
+ if (unlink(vfiopath) < 0)
+ goto cleanup;
+ }
+ }
+
+ ret = 0;
+ cleanup:
+ VIR_FREE(vfiopath);
+ return ret;
+}
+
+static int
+pci_vfio_lock_iommu(struct pciDevice *device)
+{
+ char *vfiopath = NULL;
+ int ret = -1;
+ size_t i = 0;
+ int fd = -1;
+
+ for (i = 0; i < npciIommuGroups; i++) {
+ if (pciIommuGroups[i]->iommu == device->iommuGroup)
+ break;
+ }
+
+ if (i != npciIommuGroups) {
+ if (!pciIommuGroups[i]->nDevicesBoundToVFIO) {
+ if (virAsprintfQuiet(&vfiopath, "%s/dev/vfio/%d",
+ fakesysfspcidir, device->iommuGroup) < 0) {
+ errno = ENOMEM;
+ goto cleanup;
+ }
+ if ((fd = real_open(vfiopath, O_CREAT)) < 0)
+ goto cleanup;
+ }
+ pciIommuGroups[i]->nDevicesBoundToVFIO++;
+ }
+
+ ret = 0;
+ cleanup:
+ real_close(fd);
+ VIR_FREE(vfiopath);
+ return ret;
+}
/*
* PCI Driver functions
@@ -583,6 +695,10 @@ pci_driver_bind(struct pciDriver *driver,
if (symlink(devpath, driverpath) < 0)
goto cleanup;
+ if (STREQ(driver->name, "vfio-pci"))
+ if (pci_vfio_lock_iommu(dev) < 0)
+ goto cleanup;
+
dev->driver = driver;
ret = 0;
cleanup:
@@ -617,6 +733,10 @@ pci_driver_unbind(struct pciDriver *driver,
unlink(driverpath) < 0)
goto cleanup;
+ if (STREQ(driver->name, "vfio-pci"))
+ if (pci_vfio_release_iommu(dev) < 0)
+ goto cleanup;
+
dev->driver = NULL;
ret = 0;
cleanup:
@@ -813,6 +933,8 @@ init_syms(void)
static void
init_env(void)
{
+ char *devVFIO;
+
if (fakerootdir && fakesysfspcidir)
return;
@@ -828,6 +950,29 @@ init_env(void)
make_file(fakesysfspcidir, "drivers_probe", NULL, -1);
+ if (virAsprintfQuiet(&devVFIO, "%s/dev/vfio", fakesysfspcidir) <
0)
+ ABORT_OOM();
+
+ if (virFileMakePath(devVFIO) < 0)
+ ABORT("Unable to create: %s", devVFIO);
+
+ /* Create /dev/vfio/vfio file */
+ make_file(devVFIO, "vfio", NULL, -1);
+ VIR_FREE(devVFIO);
+
+ pci_iommu_new(0);
+ pci_iommu_new(1);
+ pci_iommu_new(2);
+ pci_iommu_new(3);
+ pci_iommu_new(4);
+ pci_iommu_new(5);
+ pci_iommu_new(6);
+ pci_iommu_new(7);
+ pci_iommu_new(8);
+ pci_iommu_new(9);
+ pci_iommu_new(10);
+ pci_iommu_new(11);
I wonder if we can make these somehow dyamic. E.g. called from
pci_device_new_from_stub() and require iommuGroup in MAKE_PCI_DEVICE()
because every device belongs to an IOMMU group.
Michal