Re: [dpdk-dev] [PATCH v6 1/8] eal: pci: add api to rd/wr pci bar region

[Yuanhan Liu <yuanhan.liu@linux.intel.com>]

On Tue, Feb 02, 2016 at 09:48:44PM +0530, Santosh Shukla wrote:
> On Tue, Feb 2, 2016 at 9:21 PM, Santosh Shukla <sshukla@mvista.com> wrote:
> > On Tue, Feb 2, 2016 at 2:19 PM, Yuanhan Liu <yuanhan.liu@linux.intel.com> wrote:
> >> On Tue, Feb 02, 2016 at 06:50:18AM +0100, David Marchand wrote:
> >>> On Tue, Feb 2, 2016 at 6:43 AM, Yuanhan Liu <yuanhan.liu@linux.intel.com> wrote:
> >>> > On Tue, Feb 02, 2016 at 09:44:14AM +0530, Santosh Shukla wrote:
> >>> >> Current use-case is virtio: It is used as io_bar which is first
> >>> >> bar[1]. But implementation is generic, can be used to do rd/wr for
> >>> >> other bar index too. Also vfio facilitate user to do rd/wr to pci_bars
> >>> >> w/o mapping that bar, So apis will be useful for such cases in future.
> >>> >>
> >>> >> AFAIU: uio has read/write_config api only and Yes if bar region mapped
> >>> >> then no need to do rd/wr, user can directly access the pci_memory. But
> >>> >> use-case of this api entirely different: unmapped memory by
> >>> >> application context i.e.. vfio_rd/wr-way {pread/pwrite-way}.
> >>> >>
> >>> >> Is above explanation convincing? Pl. let me know.
> >>> >
> >>> > TBH, not really. So, as you stated, it should be generic APIs to
> >>> > read/write bar space, but limiting it to VFIO only and claiming
> >>> > that read/write bar space is not support by other drivers (such
> >>> > as UIO) while in fact it can (in some ways) doesn't seem right
> >>> > to me.
> >>> >
> >>> > Anyway, it's just some thoughts from me. David, comments?
> >>>
> >>> >From the very start, same opinion.
> >>> We should have a unique api to access those, and eal should hide
> >>> details like kernel drivers (uio, vfio, whatever) to the pmd.
> >>>
> >>> Now the thing is, how to do this in an elegant and efficient way.
> >>
> >> I was thinking that we may just make it be IO port specific read/
> >> write functions:
> >>
> >
> > Ok,
> >
> >>         rte_eal_pci_ioport_read(dev, bar, buf, size)
> >>         {
> >>
> >>                 return if not an IO bar;
> >>
> >>                 if (has io)
> >>                         return inb/w/l();
> >>
> >
> > In that case, It may be r / if (has io) / if (drv->kdrv == UIO)

Nope, I meant platform supports inb/w/l() command.

> >
> >>                 if (vfio)
> >>                         return vfio_ioport_read();
> >>
> >>                 else, claim aloud that io port read is not allowed
> >>         }
> >>
> >> Let us not handle memory bar resource here: in such case, you should
> >> go with rte_eal_pci_map_device() and do it with memory mapped io.
> >>
> >> Does that make any sense?
> >>
> > I am not entirely sure.
> > Are you considering IGB_UIO, UIO_GENERIC and NIC_UIO: all the cases ?
> >
> 
> Just came-up something below what Yuanhan has proposed, Does this look okay?
> 
> int rte_eal_pci_ioport_read(const struct rte_pci_device *device,
>                                               void *buf, size_t len,
> off_t offset,
>                                               int bar_idx)

Your implementation doesn't look right to me. But anyway, that's not
important so far; the feasibility is.

David, would you please comment on this?

	--yliu