Re: [PATCH] NTB: ntb_perf: fix cast to restricted __le32

From: Jon Mason
Date: Tue Jan 23 2018 - 23:37:17 EST


On Fri, Jan 19, 2018 at 10:26:37PM +0100, Arnd Bergmann wrote:
> On Fri, Jan 19, 2018 at 10:03 PM, Serge Semin <fancer.lancer@xxxxxxxxx> wrote:
> >
> > Actually the provided patch is the best solution I could come up with.
> > The thing is, that the methods can't be changed. Those functions are
> > the part of the NTB API methods used by many drivers. So basically they
> > are like pci_{read,write}_config_{byte,word,dword}() methods. We can't
> > change their prototypes only because it's suit some driver. The methods
> > give an access to the NTB device dummy u32-sized registers, nothing
> > else. So endianness is the transmitted data settings in this case.
> >
> > NTB is the technology to interconnect some two systems with possibly
> > different endianness (unlike PCI, which interconnect CPU with LE devices).
> > In this case I'd need to set some agreement up between two systems about
> > the endianness of the exchanged data like host and network types in
> > Linux networking. I've chosen the network data to be little-endian,
> > that's why I needed first to convert them from CPU to le32, then on
> > remote side convert them back from le32 to CPU.
> >
> > If you have any better suggestion how the warning can be fixed, I'd
> > be glad to stick to it.
>
> I don't think your description matches what you actually do: The
> underlying ntb hardware drivers (amd, idt, intel, mscc) all treat the
> incoming data as CPU-endian and convert it to little-endian on
> the register side, so the framework already assumes that whatever
> you do here uses a little-endian wire-level protocol.
>
> On a little-endian kernel/CPU, nothing is ever swapped here, neither
> in the ntb_perf front-end nor in the back-ends. On a big-endian
> kernel/CPU, they both swap, so you end up with CPU-endian
> data on the wire, so it should be impossible for a big-endian
> system to talk to a little-endian one. Have you actually tried that
> combination with the current code?

I do not believe anyone has every tried NTB on a big endien system,
let alone tried it with one side LE and the other BE. To my
knowledge, this has only ever been used on x86 to x86.

> If my interpretation is correct, then the best solution would be to
> completely remove the cpu_to_le32/le32_to_cpu conversions
> from ntb_perf, and just define that it works like any other PCI
> device, exchanging little-endian data.

Yes, this would be the best solution. Thank you for the insight.

Serge, when you get a chance, please make this change and resumbit.

Thanks,
Jon


>
> There are two interesting cases to consider though:
>
> - if someone wants to implement an NTB based protocol
> using big-endian data on the wire, you probably want to add
> a ntb_peer_spad_read_be()/ntb_peer_msg_write_be()
> set of interfaces, to go along with ioread32_be()/iowrite32_be()
> the same way that ntb_peer_spad_read()/ntb_peer_msg_write()
> ends up doing ioread32()/iowrite32() with the implied little-endian
> behavior.
>
> - memcpy_toio()/memcpy_fromio() and ioread32_rep()/iowrite32_rep
> importantly do not do any byteswap, they are meant to
> transfer byte streams.
>
> Arnd