Not only because of its the ugliest mbuf structure and the design of the network packet processing. I did not work with *bsd kernels more than 4 years already and very happy this horror was finished not even seriously started.

But even its userspace part is ugly. Let's see how Solaris and Linux implemented <code>sendfile()</code> call (although Solaris put it into special <code>libsendfile.so</code> library) compared to FreeBSD one:#ifdef HAVE_SENDFILE4_SUPPORT
#include <sys/sendfile.h>
int dnet_sendfile(struct dnet_net_state *st, int fd, off_t *offset, size_t size)
{
       int err;
   
   err = sendfile(st->s, fd, offset, size);
        if (err < 0)
            return -errno;

        return err;
}
#elif HAVE_SENDFILE7_SUPPORT
#include <sys/uio.h>
int dnet_sendfile(struct dnet_net_state *st, int fd, off_t *offset, size_t size)
{
       int err;
   err = sendfile(fd, st->s, *offset, size, NULL, &size, 0);
   if (err && errno != EAGAIN)
                return -errno;

        if (size) {
                *offset += size;
           return size;
       }

     return -EAGAIN;
}
#else
#error "Your platform does not support sendfile. Sorry."
#endif

As you might expect the latter part of the compat layer is for FreeBSD - 7 arguments for the syscall (forget about registers), and weird return values: it may be zero, positive number of bytes and negative error. And <code>EAGAIN</code> mess. What prevented it from following Solaris way of having simple and vectorized sendfiles remains a riddle.

And that's my 'simple' m4 autoconf script to detect the proper sendfile usage.
AC_DEFUN([AC_CHECK_SENDFILE],[
AC_MSG_CHECKING([whether sendfile() is supported and what prototype it has])

AC_CACHE_VAL([ac_sendfile4_supported],[
      AC_TRY_LINK([#include <sys/sendfile.h>
                       #include <stdio.h>],
         [sendfile(1, 1, NULL, 0);],
                [
                  ac_sendfile4_supported=yes
                 ac_sendfile_supported=yes
                  AC_MSG_RESULT([usual Linux/Solaris sendfile()])
            ],
         [
                  dnl Checking wether we need libsendfile
                    dnl Presumably on Solaris
                  AC_CHECK_LIB(sendfile, sendfile,
                           [
                                  ac_sendfile4_supported=yes
                                 ac_sendfile_supported=yes
                                  SENDFILE_LIBS="-lsendfile"
                                       AC_SUBST(SENDFILE_LIBS)
                            ],
                         [ac_sendfile_supported=no])
                ])
])
if test x$ac_sendfile_supported = xno; then
  AC_CACHE_VAL([ac_sendfile7_supported],[
            AC_TRY_LINK([#include <sys/socket.h>
                         #include <stdio.h>],
                 [sendfile(1, 1, 0, 0, NULL, NULL, 0);],
                    [
                          ac_sendfile7_supported=yes
                         ac_sendfile_supported=yes
                          AC_MSG_RESULT([ugly FreeBSD sendfile()])
                   ],
                 [ac_sendfile_supported=no])
        ])
fi
if test x$ac_sendfile_supported = xno; then
  AC_MSG_ERROR([no sendfile support])
fi
if test x$ac_sendfile4_supported = xyes; then
       AC_DEFINE(HAVE_SENDFILE4_SUPPORT, 1, [Define this if Linux/Solaris sendfile() is supported])
fi
if test x$ac_sendfile7_supported = xyes; then
      AC_DEFINE(HAVE_SENDFILE7_SUPPORT, 1, [Define this if FreeBSD sendfile() is supported])
fi
])

I'm slowly starting to like autotools - although my project is about 5 thousands of lines and only generated <code>configure</code> script is about 5 times bigger, it is still simple and somewhat clear (although limited sometimes).

<a href="/projects/elliptics">Elliptics</a> network got full FreeBSD support and I even used it slightly to catch some bugs (mostly because of its <code>sendfile()</code> support). Project also got log events mask, compile-time ID size changes (<code>--with-id-size=$bytes</code> switch) and lots of other smaller bits (especially in the automatic configuration).

Tomorrow (if things will be completed before running out climbing) I will implement the testing tool described and start the process. Do you expect a failure? :)