[musl] socketcall syscall on x32 ABI producing incorrect results

[musl] socketcall syscall on x32 ABI producing incorrect results

[ca0]

I am running busybox with musl-1.2.3 on x32 linux with gcc-12.2.0.

The syscall returns bad address in the case of the bind call. I have traced this down to the pointer being 0xf...... in 32bits being sign extended to 32bits in the syscall.

Here is the code:
0xf7f687ec	65		else r = __syscall(sys, a, b, c, d, e, f);
0xf7f687e9 <__alt_socketcall+178>:	48 63 d0           	movslq %eax,%rdx

I can see the address being sign extended as intended by the movslq instruction.

If I change the bind implementation to cast this address I do not see this issue, but other issues arise with the soecket call

ie. in bind 
int bind(int fd, const struct sockaddr *addr, socklen_t len)
{
    long b = ((unsigned long)(addr)  & 0xffffffff);
    freturn socketcall(bind, fd, b, len, 0, 0, 0);
}

Is this a known issue as x32 is marked as experimental.

Sent with Proton Mail secure email.

Re: [musl] socketcall syscall on x32 ABI producing incorrect results

[Rich Felker]

On Sun, Mar 26, 2023 at 09:12:41PM +0000, ca0 wrote:
> I am running busybox with musl-1.2.3 on x32 linux with gcc-12.2.0.
> 
> The syscall returns bad address in the case of the bind call. I have traced this down to the pointer being 0xf...... in 32bits being sign extended to 32bits in the syscall.
> 
> Here is the code:
> 0xf7f687ec	65		else r = __syscall(sys, a, b, c, d, e, f);
> 0xf7f687e9 <__alt_socketcall+178>:	48 63 d0           	movslq %eax,%rdx
> 
> I can see the address being sign extended as intended by the movslq instruction.
> 
> If I change the bind implementation to cast this address I do not see this issue, but other issues arise with the soecket call
> 
> ie. in bind 
> int bind(int fd, const struct sockaddr *addr, socklen_t len)
> {
>     long b = ((unsigned long)(addr)  & 0xffffffff);
>     freturn socketcall(bind, fd, b, len, 0, 0, 0);
> }
> 
> Is this a known issue as x32 is marked as experimental.
> 
> Sent with Proton Mail secure email.

Thanks! I think this is fixed upstream already (not in a release) in
commit 2404d9d643763e6eceafa9a1918925d80a84ad44. Let us know if you
think that doesn't solve it.

Rich