On Fri, Jun 23, 2023 at 05:59:55AM +0200, Felix Kuehling wrote: > On 2023-06-22 23:42, Lukas Wunner wrote: > > On Thu, Jun 22, 2023 at 02:02:12PM -0700, Smita Koralahalli wrote: > > > Would it be fair to just reuse pci_enable_atomic_ops_to_root() for > > > Atomic_Ops configuration? > > > > Hm, that's a good question. I'm not an expert on that corner of > > the PCI core. > > > > But indeed what you could try is amend that function to not only > > *set* PCI_EXP_DEVCTL2_ATOMIC_REQ if it's supported, but to also > > *clear* it if it's not supported. > > > > And you'd have to call pci_enable_atomic_ops_to_root() on enumeration, > > e.g. from pci_init_capabilities(). > > > > That should obviate the need to call pci_enable_atomic_ops_to_root() > > from drivers, so you could probably remove the call from all the > > drivers which currently call it (amdgpu, infiniband, mellanox), > > in one separate patch per driver. > > > > An then you could drop the EXPORT clause for pci_enable_atomic_ops_to_root() > > and make it private to the PCI core. > > Then our driver would need an alternative way to determine whether atomic > capabilities are enabled for a device. We currently use the return value > from pci_enable_atomic_ops_to_root to determine this. Just read PCI_EXP_DEVCTL2 and check whether PCI_EXP_DEVCTL2_ATOMIC_REQ is set. (I.e. has been set by the PCI core on device enumeration.) Problem solved, I guess? Thanks, Lukas
