Re: [Qemu-devel] [PATCH 2/2] linux-user: SOCK_PACKET uses network endian to encode protocol in socket()

[Laurent Vivier]

Le mardi 01 janvier 2013 à 15:03 +0000, Peter Maydell a écrit :
> On 31 December 2012 22:19, Laurent Vivier <address@hidden> wrote:
> > Le lundi 31 décembre 2012 à 21:32 +0000, Peter Maydell a écrit :
> >> Also it seems rather involved since we swap things three times and
> >> have an entirely new abi_* function. Either I'm completely confused
> >> or it should be enough to just have
> >>
> >> if (type == SOCK_PACKET) {
> >>       protocol = tswap16(protocol);
> >> }
> 
> Looking more carefully at packet(7) this is actually the wrong
> guard anyway. You need to check for
>  (domain == AF_PACKET) || (type == SOCK_PACKET)

I agree.

> since SOCK_PACKET is the obsolete Linux 2.0 way of doing packet sockets.

But dhclient is always using this...

> > works... sometime. In fact, work if target endianess is network endianess.
> >
> > Correct me if I'm wrong.
> >
> > target          host
> > little endian / big endian
> >
> > memory   00 00 00 03
> 
> Syscall arguments aren't generally passed in memory, they're
> in registers (and if they were pased in memory for some architecture
> then that arch would do a load-and-swap-from-memory in main.c).
> So the value you see in do_socket() is always "the integer passed
> as a syscall parameter, as a host-order integer".

Yes, I missed that.

> So in this case, with a simple guest program:
> #include <sys/socket.h>
> #include <netpacket/packet.h>
> #include <net/ethernet.h>
> #include <arpa/inet.h>
> 
> int main(void) {
>    return socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
> }
> 
> you will find that do_socket() in QEMU is passed either 0x3 [if the
> guest is bigendian and the guest htons() is a no-op] or 0x0300
> [if the guest is littleendian]. Since what we want to pass to the
> host socket() call is 0x3 if the host is bigendian and 0x0300 if
> the host is little endian, this amounts to needing to do a 16 bit
> byteswap if the host and guest are different endianness, which
> is exactly what tswap16() does. I checked with i386-to-i386
> that do_socket() gets passed 0x300 and we correctly send it
> through to the host socket().

Yes, I agree. I correct the patch.

Thank you,
Laurent

-- 
"Just play. Have fun. Enjoy the game."
- Michael Jordan