> From: Robin Murphy <robin.murphy@xxxxxxx> > Sent: Thursday, July 7, 2022 5:59 PM > > On 2022-07-07 07:31, Tian, Kevin wrote: > >> From: Robin Murphy <robin.murphy@xxxxxxx> > >> Sent: Wednesday, July 6, 2022 1:08 AM > >> > >> The number of bus types that the IOMMU subsystem deals with is small > and > >> manageable, so pull that list into core code as a first step towards > >> cleaning up all the boilerplate bus-awareness from drivers. Calling > >> iommu_probe_device() before bus->iommu_ops is set will simply return > >> -ENODEV and not break the notifier call chain, so there should be no > >> harm in proactively registering all our bus notifiers at init time. > >> > > > > Suppose we miss a check on iommu ops in iommu_release_device(): > > > > if (!dev->iommu) <<<<<<< > > return; > > > > iommu_device_unlink(dev->iommu->iommu_dev, dev); > > > > ops = dev_iommu_ops(dev); > > ops->release_device(dev); > > > > following the rationale in patch01 a device could be removed when > > it's associated with a known but not registered instance. > > No, because at that point the instance is only known internally to the > driver. As long as it isn't erroneously returned from > ->probe_device(dev), dev->iommu will remain NULL and the rest of the > core code works as expected. > You are correct. I overlooked dev->iommu as device_to_iommu() in patch01. As long as the device hasn't been probed or ->probe_device doesn't do bad thing then dev->iommu should be NULL in this case.
