On Wed, Feb 5, 2025 at 5:44 PM Renyong Wan <wa...@yunsilicon.com> wrote: > >> That's how we designed it. > >> We designed a low-level device operations framework named xsc_dev_ops to > >> support both VFIO drivers and kernel drivers. Each xsc_dev_ops is > >> registered before the main function runs. During the PCI device probe > >> phase, the XSC PMD selects the corresponding xsc_dev_ops based on > >> rte_pci_device->driver, therefore, there is no need for devargs. > > I don't understand. > > If both VFIO and the kernel driver are loaded, > > we'll scan the device twice? > > Will it be probed 2 times? > > > > > No, it won't probe twice. > I suppose that each PCI device will only be bound to either the VFIO > driver or a kernel driver. The drv_flags of the xsc_pmd PCI driver will > not preset with RTE_PCI_DRV_NEED_MAPPING. Therefore, in the > rte_pci_probe_one_driver function, rte_pci_map_device() will not be > called. After entering the probe phase of the xsc PMD PCI driver, > rte_pci_map_device() will be called in xsc_dev_ops->init() based on > whether it is a VFIO driver.
(side note, this means that this driver should probably call rte_pci_unmap_device() in its release path, though I see none) -- David Marchand
