Andrew Stitcher escribió:
On Tue, 2008-05-27 at 09:13 +0200, Manuel Teira wrote:
...
Yes, my plan was to make something like this:
AC_CHECK_HEADERS([sys/epoll.h],[with_epoll=yes], [with_epoll=no])
if test x$with_epoll = xno; then
AC_CHECK_HEADERS([port.h],[with_solecf=yes], [with_solecf=no])
if test x$with_solecf = xno; then
AC_MSG_ERROR([No supported polling infrastructure found])
fi
fi
AM_CONDITIONAL([EPOLL], [test x$with_epoll = xyes])
AM_CONDITIONAL([SOLECF], [test x$with_solecf = xyes])
in configure.ac. Or perhaps something more $(UNAME) oriented would be
preferred? I mean, just using the Solaris Event Completion Framework
when in solaris, instead of that. Anyway, we have to take into account
that ECF exists since Solaris 10. For previous versions of the OS, they
had a poll interface.
Then, selecting the appropiate implementation in the proper Makefile.am:
if EPOLL
posix_plat_src += qpid/sys/epoll/EpollPoller.cpp
else
posix_plat_src += qpid/sys/sol-ecf/SolECFPoller.cpp
endif
Do you agree with this type of approach?
This looks about right. The only comment I'd make is that it's not
really a binary epoll/solecf thing, so an approach that more easily
allows us to add in say kqueue (for FreeBSD) and poll (for everyone else
posix) approaches would be favoured (by mo at least).
In practice it will be a combined uname/version/header/library check I'd
guess.
I think that the only user visible switch to configure should be to
select posix as against platform specific. So having
--with-epoll/--with-kqueue/--with-ecf is not an option I like. I'd go
for --with-posix-only for the case that we want poll even though we've
detected that the platform supports something more efficient.
As an aside, unless you know you've got Solaris 10 or better you will
need to implement something poll() or /dev/poll based as well. In which
case can I encourage you to do a poll() port as it would be generally
useful to anyone with any Unix like system.
I totally agree with your approach. Actually, the posix poll interface
is the most generic, and then, we would have specific mechanisms in
different OSs. So, perhaps the configure.ac stuff would automatically
try to guess the better option based in the $host variable (instead of
the uname I was proposing, and so, providing some hope for cross
compilations), something like (pseudocode) :
POLLING_IMPL=None
case *sun-solaris2.10)
- Try to locate port.h && POLLING_IMPL=sol-ecf
- Else POLLING_IMPL=posix-poll
case *sun-solaris*)
- POLLING_IMPL=posix-poll
case *linux*)
- Try to locate sys/epoll.h && POLLING_IMPL=linux-epoll
- Else POLLING_IMPL=posix-poll
case *bsd*)
- Try to locate kqueue.h or whatever && POLLING_IMPL=bsd-kqueue
- Else POLLING_IMPL=posix-poll
case *windows*)
- Whatever mechanism is used in windows && POLLING_IMPL=windows-whatever
After this, a comparison with the set of existing polling
implementations would be nice, to decide if we can follow on or we just
have to throw an error:
case $POLLING_IMPL in
None)
AC_MSG_ERROR([No polling mechanism detected for $host])
;;
posix-poll)
bsd-kqueue)
windows-whatever)
AC_MSG_ERROR([No $POLLING_IMPL mechanism implemented yet. Please
contrib one])
;;
linux-epoll)
AM_CONDITIONAL([LINUX_EPOLL], [1])
;;
solaris-ecf)
AM_CONDITIONAL([SOLARIS_ECF], [1])
esac
Then, in the sources Makefile.am, we would add the right sources for the
selected epoll mechanism.
Regards.
Andrew
.
