Re: tcpcb (was Re: sockbuf (was Re: BGL-free net stack))

["Sepherosa Ziehau" <sepherosa@xxxxxxxxx>]

On 6/10/08, Aggelos Economopoulos <aoiko@cc.ece.ntua.gr> wrote:
> On Friday 06 June 2008, Aggelos Economopoulos wrote:
>  [...]
>
> > OK, same thing, but now it's the pcbs. TCP is "easy". The inpcb is inserted on
>  > a per-cpu hash table so that the corresponding protocol thread runs on the
>  > same cpu. Some tcpcb fields however, are accessed directly from user-thread
>  > context. The interesting fields are:
>  >
>  > t_flags:
>  >       need to do early copyin / delayed copyout in so_pr_ctloutput
>
>
> So, I was thinking something like the following:
>
>  diff --git a/sys/kern/uipc_msg.c b/sys/kern/uipc_msg.c
>  index fde4c93..e786e02 100644
>  --- a/sys/kern/uipc_msg.c
>  +++ b/sys/kern/uipc_msg.c
>  @@ -35,6 +35,7 @@
>
>   #include <sys/param.h>
>   #include <sys/systm.h>
>  +#include <sys/kernel.h>
>   #include <sys/msgport.h>
>   #include <sys/protosw.h>
>   #include <sys/socket.h>
>  @@ -379,22 +380,38 @@ so_pru_sopoll(struct socket *so, int events, struct ucred *cred)
>         return (error);
>   }
>
>  +MALLOC_DEFINE(M_SOPT, "sopt", "sopt temp storage");
>  +
>   int
>   so_pr_ctloutput(struct socket *so, struct sockopt *sopt)
>   {
>  -       return ((*so->so_proto->pr_ctloutput)(so, sopt));
>   #ifdef gag     /* does copyin and copyout deep inside stack XXX JH */
>  +       return ((*so->so_proto->pr_ctloutput)(so, sopt));
>  +#else
>         struct netmsg_pr_ctloutput msg;
>         lwkt_port_t port;
>         int error;
>  +       void *uval;
>  +
>  +       uval = sopt->sopt_val;
>
>  -       port = so->so_proto->pr_mport(so, NULL);
>  +       /* we keep duplicate copies, but for option {s,g}etting who cares? */
>  +       sopt->sopt_val = kmalloc(sopt->sopt_valsize, M_SOPT, M_WAITOK);
>  +       error = copyin(uval, sopt->sopt_val, sopt->sopt_valsize);
>  +       if (error)
>  +               goto out;
>  +       port = so->so_proto->pr_mport(so, NULL, NULL, XXX);
>         netmsg_init(&msg.nm_netmsg, &curthread->td_msgport, 0,
>                     netmsg_pru_ctloutput);
>         msg.nm_prfn = so->so_proto->pr_ctloutput;
>         msg.nm_so = so;
>         msg.nm_sopt = sopt;
>         error = lwkt_domsg(port, &msg.nm_netmsg.nm_lmsg, 0);
>  +       if (error)
>  +               goto out;
>  +       error = copyout(sopt->sopt_val, uval, sopt->sopt_valsize);
>  +out:
>  +       kfree(sopt->sopt_val, M_SOPT);
>         return (error);
>   #endif
>   }
>
>  But before I update all callees to remove copy{in,out}(), does anybody have
>  any objections? Also, what should be the value of XXX? Perhaps ctloutput
>  belongs to pr_usrreqs...

I don't think above change is enough.  You need to change
sooptcopy{in,out}() or at least set sopt->sopt_td to NULL.  Please
take a look at ip_ctloutput() for what the above patch may break.

Best Regards,
sephe

-- 
Live Free or Die