Re: linuxulator problems in linux_socket.c (and friends)

[Andrew Atrens <atrens@xxxxxxxxxxxxxxxxxx>]

Oops, with all the crashing and rebooting my realtime clock got a bit 
skewed :(


This change seems to have been introduced in -

http://gomerbud.com/daver/patches/dragonfly/linux-no43-3.diff

by David Reese ...


Looks like

-/* Return 0 if IP_HDRINCL is set for the given socket. */
-static int
-linux_check_hdrincl(int s)
-{


and

-/*
- * Updated sendto() when IP_HDRINCL is set:
- * tweak endian-dependent fields in the IP packet.
- */
-static int
-linux_sendto_hdrincl(struct sendto_args *bsd_args)


were removed.


Indeed, looks like

linux_check_hdrincl() did something fancy when calling getsockopt()

-	sg = stackgap_init();
-	val = stackgap_alloc(&sg, sizeof(int));
-	valsize = stackgap_alloc(&sg, sizeof(int));
-
-	if ((error = copyout(&size_val, valsize, sizeof(size_val))))
-		return (error);
-
-	bsd_args.s = s;
-	bsd_args.level = IPPROTO_IP;
-	bsd_args.name = IP_HDRINCL;
-	bsd_args.val = val;
-	bsd_args.avalsize = (int *)valsize;
-	bsd_args.sysmsg_result = 0;
-	if ((error = getsockopt(&bsd_args)))
-		return (error);
-	/* return value not used */
-
-	if ((error = copyin(val, &optval, sizeof(optval))))
-		return (error);
-

this was replaced by the below ref'd code that now calls 
kern_getsockopt() without doing the stackgap_alloc/copyin/copyout ...


Not sure why all that was in there, but if I have some time on Thursday 
I'll take a run at it :P

Cheers,

Andrew.



Andrew Atrens wrote:
> Hi All,
>
> I've had clearcase 5.0 running on FreeBSD-stable for quite some time 
> (probably 3+ years) and I thought I'd give it a try on DragonFly ...
>
> Well, the short version is that it doesn't work... and I've not found a 
> workaround -
>
> in linux_socket.c -
>
> kern_getsockopt() (quoted below) is failing with error 42...
>
> #define ENOPROTOOPT     42              /* Protocol not available */
>
>
> If I kludge around the error by forcing both error and sopt to zero, I 
> can get simple commands to work, but this is clearly not the right thing
> to do.
>
> linux_socket.c is very different in df from what's in 4.9. Looks like 
> the df version has been substantially refactored.
>
> In 4.9 IP_HDRINCL sockets are treated very specially (with their own 
> special identification test func. and special sendto func as well). 
> Don't know if this is required in df, but what's there now doesn't work 
> for me :(
>
> I'd certainly like to get this working ... let me know how you'd like me 
>   to proceed.
>
> Cheers,
>
> Andrew.
>
>
>
> static int
> linux_sendto(struct linux_sendto_args *args, int *res)
> {
>     struct linux_sendto_args linux_args;
>     struct thread *td = curthread;
>     struct uio auio;
>     struct iovec aiov;
>     struct sockopt sopt;
>     struct sockaddr *sa = NULL;
>     caddr_t msg = NULL;
>     int error, optval;
>
>     error = copyin(args, &linux_args, sizeof(linux_args));
>     if (error)
>         return (error);
>
>     if (linux_args.to) {
>         error = linux_getsockaddr(&sa, linux_args.to,
>             linux_args.tolen);
>         if (error)
>             return (error);
>     }
>
>     /*
>      * Check to see if the IP_HDRINCL option is set.
>      */
>     sopt.sopt_dir = SOPT_GET;
>     sopt.sopt_level = IPPROTO_IP;
>     sopt.sopt_name = IP_HDRINCL;
>     sopt.sopt_val = &optval;
>     sopt.sopt_valsize = sizeof(optval);
>     sopt.sopt_td = NULL;
>
>     error = kern_getsockopt(linux_args.s, &sopt);    <--- returns 42
>
>     error = sopt = 0 ; /* BROKEN: assume the option is not set */
>
>     if (error)
>         return (error);