illumos gate

We encountered an instance where rpcbind wasn't properly reporting an open file via pfiles. After a bit of digging to find this, the following is the final analysis:

This is exactly the purpose of the userland function dotli() in pfiles. However, with the change for dfc0fed81813380fb526f1a003d3be17ab26ce3b (illumos#3686), this code can obviously never work:

        strcmd.sc_cmd = TI_GETMYNAME;
        if (pr_ioctl(Pr, fd, _I_CMD, &strcmd, sizeof (strcmd)) == 0)
                show_sockaddr("sockname", (void *)&strcmd.sc_buf, 0);

        strcmd.sc_cmd = TI_GETPEERNAME;
        if (pr_ioctl(Pr, fd, _I_CMD, &strcmd, sizeof (strcmd)) == 0)
                show_sockaddr("peername", (void *)&strcmd.sc_buf, 0);

Note that we always pass a 0 length to show_sockaddr(); however, this change introduced a shortcut check that bails out when length is 0. So since then, TLI socket names have been broken. This should instead pass the proper length obtained from the ioctl. With this change, we now see:

   5: S_IFCHR mode:0000 dev:545,0 ino:57437 uid:0 gid:0 rdev:67,116
      O_RDWR
        sockname: AF_INET6 ::  port: 62129
      /devices/pseudo/udp6@0:udp6
      offset:0

We're not done yet, though. In the GZ, these paths are links to the /devices paths, which end in :udp (for example) and are correctly identified by pfiles as TLI endpoints. In zones, however, the /dev paths are the only paths there are, and libproc correctly presents them to pfiles as /dev/udp (for example). These are not sent through dotli() and thus are not shown, either. So we also need to try the last component of the filename if there's no : in it when checking for known TLI endpoints. With this, I now also get:

   5: S_IFCHR mode:0000 dev:546,3 ino:22746 uid:0 gid:0 rdev:67,24
      O_RDWR
        sockname: AF_INET6 ::  port: 50612
      /dev/udp6
      offset:0