On Wed, Dec 13, 2023 at 07:05:24PM +0200, Margolin, Michael wrote:
> > On Mon, Dec 11, 2023 at 05:47:15PM +0000, Michael Margolin wrote:
> >> diff --git a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
> >> index 9c65bd27bae0..597f7ca6f31d 100644
> >> --- a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
> >> +++ b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
> >> @@ -415,6 +415,32 @@ struct efa_admin_reg_mr_resp {
> >> * memory region
> >> */
> >> u32 r_key;
> >> +
> >> + /*
> >> + * Mask indicating which fields have valid values
> >> + * 0 : recv_pci_bus_id
> >> + * 1 : rdma_read_pci_bus_id
> >> + * 2 : rdma_recv_pci_bus_id
> >> + */
> >> + u8 validity;
> >> +
> >> + /*
> >> + * Physical PCIe bus used by the device to reach the MR for receive
> >> + * operation
> >> + */
> >> + u8 recv_pci_bus_id;
> >> +
> >> + /*
> >> + * Physical PCIe bus used by the device to reach the MR for RDMA read
> >> + * operation
> >> + */
> >> + u8 rdma_read_pci_bus_id;
> >> +
> >> + /*
> >> + * Physical PCIe bus used by the device to reach the MR for RDMA write
> >> + * receive
> >> + */
> >> + u8 rdma_recv_pci_bus_id;
> > What driver is bound to this other PCIe bus and how did the iommu get
> > setup for it?
>
> It's internal bus that is not directly exposed to the host. Addresses
> mapping is acquired from accelerator's driver as for any MR residing in
> accelerator memory, and the translation is owned by devices on that bus.
So if it isn't visible to the host, or connectable to anything Linux
would call a PCI RID, why are you giving it such specific names? Just
call it 'interconnect path id' or something and make it opaque?
Jason
