On Wed, Jul 14 2021, Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > Currently the driver ops have an open/release pair that is called once > each time a device FD is opened or closed. Add an additional set of > open/close_device() ops which are called when the device FD is opened for > the first time and closed for the last time. > > An analysis shows that all of the drivers require this semantic. Some are > open coding it as part of their reflck implementation, and some are just > buggy and miss it completely. > > To retain the current semantics PCI and FSL depend on, introduce the idea > of a "device set" which is a grouping of vfio_device's that share the same > lock around opening. > > The device set is established by providing a 'set_id' pointer. All > vfio_device's that provide the same pointer will be joined to the same > singleton memory and lock across the whole set. This effectively replaces > the oddly named reflck. > > After conversion the set_id will be sourced from: > - A struct device from a fsl_mc_device (fsl) > - A struct pci_slot (pci) > - A struct pci_bus (pci) > - The struct vfio_device (everything) > > The design ensures that the above pointers are live as long as the > vfio_device is registered, so they form reliable unique keys to group > vfio_devices into sets. > > This implementation uses xarray instead of searching through the driver > core structures, which simplifies the somewhat tricky locking in this > area. > > Following patches convert all the drivers. > > Signed-off-by: Yishai Hadas <yishaih@xxxxxxxxxx> > Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxx> > --- > drivers/vfio/mdev/vfio_mdev.c | 22 ++++++ > drivers/vfio/vfio.c | 144 ++++++++++++++++++++++++++++------ > include/linux/mdev.h | 2 + > include/linux/vfio.h | 19 +++++ > 4 files changed, 165 insertions(+), 22 deletions(-) > (...) > @@ -760,6 +829,13 @@ int vfio_register_group_dev(struct vfio_device *device) > struct iommu_group *iommu_group; > struct vfio_group *group; > > + /* > + * If the driver doesn't specify a set then the device is added to a > + * signleton set just for itself. s/signleton/singleton/ > + */ > + if (!device->dev_set) > + vfio_assign_device_set(device, device); > + > iommu_group = iommu_group_get(device->dev); > if (!iommu_group) > return -EINVAL; > @@ -1361,7 +1437,8 @@ static int vfio_group_get_device_fd(struct vfio_group *group, char *buf) > { > struct vfio_device *device; > struct file *filep; > - int ret; > + int fdno; > + int ret = 0; > > if (0 == atomic_read(&group->container_users) || > !group->container->iommu_driver || !vfio_group_viable(group)) > @@ -1375,38 +1452,38 @@ static int vfio_group_get_device_fd(struct vfio_group *group, char *buf) > return PTR_ERR(device); > > if (!try_module_get(device->dev->driver->owner)) { > - vfio_device_put(device); > - return -ENODEV; > + ret = -ENODEV; > + goto err_device_put; > } > > - ret = device->ops->open(device); > - if (ret) { > - module_put(device->dev->driver->owner); > - vfio_device_put(device); > - return ret; > + mutex_lock(&device->dev_set->lock); > + device->open_count++; > + if (device->open_count == 1 && device->ops->open_device) { > + ret = device->ops->open_device(device); > + if (ret) > + goto err_undo_count; Won't that fail for mdev devices, until the patches later in this series have been applied? (i.e. bad for bisect) > + } > + mutex_unlock(&device->dev_set->lock); > + > + if (device->ops->open) { > + ret = device->ops->open(device); > + if (ret) > + goto err_close_device; > }