On Friday, November 18, 2016 7:12:35 PM CET zhichang.yuan wrote:
> Hi, Arnd,
>
>
> On 2016/11/18 17:20, Arnd Bergmann wrote:
> > On Friday, November 11, 2016 6:07:07 PM CET zhichang.yuan wrote:
> >>
> >> I have similar idea as your PPC MMIO.
> >>
> >> We notice the prototype of {in/out()} is something like that:
> >>
> >> static inline u8 inb(unsigned long addr)
> >> static inline void outb(u8 value, unsigned long addr)
> >>
> >> The type of parameter 'addr' is unsigned long. For I/O space, it is big enough.
> >> So, could you divide this 'addr' into several bit segments? The top 8 bits is
> >> defined as bus index. For normal direct IO, the bus index is 0. For those bus
> >> device which need indirectIO or some special I/O accessors, when these devices
> >> are initializing, can request to allocate an unique ID to them, and register
> >> their own accessors to the entry which is corresponding to the ID.
> >
> > Ah, have you looked at the IA64 code? It does exactly this.
> > For ARM64 we decided to use the same basic approach as powerpc with
> > a single range of virtual memory for mapping it as that somewhat
> > simplified all cases we knew about at the time.
>
> Yes. I spent some time to trace how to work on PPC. But the code is a bit long,
> I am not clear on how the indirectIO there was supported.
>
> I noticed there are CONFIG_PPC_INDIRECT_PIO and CONFIG_PPC_INDIRECT_MMIO on PPC.
> It seems that only CONFIG_PPC_INDIRECT_MMIO applied some MSB to store the bus
> tokens which are used to get iowa_busses[] for specific operation helpers.
> I can not find how CONFIG_PPC_INDIRECT_PIO support multiple ISA domains. It
> seems only Opal-lpc.c adopt this INDIRECT_PIO method.
>
> Although CONFIG_PPC_INDIRECT_MMIO is for MMIO, seems not suitable for ISA/LPC
> I/O. But this idea is helpful.
>
> what else did I miss??
I mentioned two different things here: ia64 IIRC uses some bits of the
port number to look up the domain, while powerpc traditionally had no
support for any such lookup, it did the same thing as ARM64 with
virtual remapping of MMIO ranges into an address range starting at
a fixed virtual address.
CONFIG_PPC_INDIRECT_PIO is a fairly recent addition, I was not thinking
of that.
> >> In this way, we can support multiple domains, I think.
> >> But I am not sure whether it is feasible, for example, are there some
> >> architectures/platforms had populated the top 8 bits? Do we need to request IO
> >> region from ioport_resource for those devices? etc...
> >
> > On a 64-bit architecture, the top 32 bits of the port number are
> > definitely free to use for this, and 8 bits are probably sufficient.
> >
> > Even on 32 bit architectures, I can't see why we'd ever need more than
> > 16 bits worth of addressing within a domain, so using 8 bit domain
> > and 16 bit address leaves 8 or 40 unused bits.
>
> Yes. 8 bits are enough.
> But the maximal PIO on some architectures are defined as ~0 or -1. There is no
> any bare space left. Probably we can not ensure the upper 8 bits available.
Right, we clearly can't use it across all architectures. The trick with
architectures using ULONG_MAX as the limit for port numbers is that they
treat it as a 1:1 mapping between port numbers and virtual addresses,
which is yet another way to handle the MMIO-based devices, but that has
a number of downsides we don't need to get into now.
What I think the code should do is a generic workaround handling that
architectures can opt-in to. We'd start doing this on ARM64 only,
and can then decide whether to change ARM or PowerPC over to use
that as well.
Arnd
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
