23/10/2024 21:14, Akhil Goyal: > > 22/10/2024 14:06, Akhil Goyal: > > > > On Tue, Oct 22, 2024 at 8:06 AM Akhil Goyal <gak...@marvell.com> wrote: > > > > > > The rational to NOT pull "Hardware abstraction library using the BAR > > > > > > address" to DPDK are > > > > > > -Yet another 200K of driver C++ code which does not make sense to > > > > > > keep > > > > > > in dpdk.org > > > > > > -It can not implemenent any of the current subsystems > > > > > > > > > > > > In this context, let me know what you think? > > > > > > > > This hardware abstraction library will have to call some driver > > > > specific API (like the added raw/cnxk_rvu API). > > > > Can this library directly use the PCI driver API, and call the cnxk > > > > common driver? > > > > If so, there is no need to add another driver API (that breaks > > > > layers), and the rawdev driver is only about adding the mailbox > > > > features. > > > > > > cnxk common has a lot of internal APIs which > > > we do not want to expose directly to application. > > > > Actually the target is an out-of-tree driver, not an application. > > > > > It seems the only contentious API is to get the BAR addresses. > > > All other APIs seems to be fine. Right? > > > > Interrupt management is also something we would like to see reserved to > > drivers. > > > > > If so, we can remove that API and get the bar addresses as you suggested > > below. > > > rte_rawdev_info_get() -> get rte_device -> RTE_DEV_TO_PCI -> get bar addr > > > via bus_pci_driver.h header, exported via enable_driver_sdk option > > > > Yes please, it is better to avoid confusion between driver and application. > > Any driver type layer should be exposed only when the Meson option > > enable_driver_sdk is enabled. > > > > You probably should export the interrupt functions in a separate file > > through driver_sdk_headers, so it is clear it is not an application level. > > Does it apply to the mailbox functions as well? > > > Ok, if we are having a separate file as driver_sdk_header, > We can also keep the BAR address API in that. Right?
I think yes but it is redundant with the path rte_rawdev_info_get() -> get rte_device -> RTE_DEV_TO_PCI -> get bar addr Is it performance sensitive? In this case you could save/cache it.
