[libvirt] mingw: virsh event loop failure in current git
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored. virEventRegisterDefaultImpl calls virEventPollInit that calls virSetNonBlock that calls ioctl that is replaced by gnulib and calls ioctlsocket. ioctlsocket fails because the given FD is a pipe but ioctlsocket expects a socket. A version that works on pipes on Windows looks like this. Although the pipe is actually not a named pipe the call to SetNamedPipeHandleState doesn't fail at least. int virSetPipeNonBlock(int fd) { DWORD mode = PIPE_NOWAIT; HANDLE handle = _get_osfhandle(fd) BOOL result = SetNamedPipeHandleState(handle, &mode, NULL, NULL); return result ? 0 : -1; } So, is the event loop stuff supposed to work on Windows at all and we should get it fixed? Or do we just put an #ifndef WIN32 around virEventRegisterDefaultImpl in virsh, because the only event loop user in virsh is the console command that is disabled on Windows anyway? Matthias
On Tue, Mar 15, 2011 at 09:29:14AM +0100, Matthias Bolte wrote:
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored.
virEventRegisterDefaultImpl calls virEventPollInit that calls virSetNonBlock that calls ioctl that is replaced by gnulib and calls ioctlsocket. ioctlsocket fails because the given FD is a pipe but ioctlsocket expects a socket.
Hmm, isn't this actually a bug in gnulib's ioctl replacement then ? The goal of gnulib is that you never use a SOCKET in any of the POSIX APIs it wraps. It hides the WinSock horribleness so that you can just use normal FDs everywhere.
A version that works on pipes on Windows looks like this. Although the pipe is actually not a named pipe the call to SetNamedPipeHandleState doesn't fail at least.
int virSetPipeNonBlock(int fd) { DWORD mode = PIPE_NOWAIT; HANDLE handle = _get_osfhandle(fd) BOOL result = SetNamedPipeHandleState(handle, &mode, NULL, NULL);
return result ? 0 : -1; }
So, is the event loop stuff supposed to work on Windows at all and we should get it fixed? Or do we just put an #ifndef WIN32 around virEventRegisterDefaultImpl in virsh, because the only event loop user in virsh is the console command that is disabled on Windows anyway?
There were several reasons the console stuff used to be disabled - It relied on opening /dev/pts/XXX which obviously only makes sense for local connection, not remote ones. - It uses the terminal attribute APIs to put the console into raw mode. The event loop is intended to be portable, and certainly, now it is a public API we should make it work. The opening of /dev/pts/XXX is no longer a blocker for virsh console on Win32, since we access the console via the streams APIs. So if we can figure out what todo with the terminal attribute APIs, virsh console should work on Win32 (or any other platform). This is one of the nice benefits of the streams API Regards, Daniel -- |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :| |: http://libvirt.org -o- http://virt-manager.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
On 03/15/2011 05:02 AM, Daniel P. Berrange wrote:
On Tue, Mar 15, 2011 at 09:29:14AM +0100, Matthias Bolte wrote:
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored.
virEventRegisterDefaultImpl calls virEventPollInit that calls virSetNonBlock that calls ioctl that is replaced by gnulib and calls ioctlsocket. ioctlsocket fails because the given FD is a pipe but ioctlsocket expects a socket.
Hmm, isn't this actually a bug in gnulib's ioctl replacement then ?
Not necessarily. FIONBIO is not a standardized ioctl() call (for that matter, ioctl() is only standardized for STREAMS, which is an obsolete Solaris concept not really supported in Linux). Rather, the gnulib ioctl() wrapper exists so you can call undo the rest of gnulib's munging of mingw socket fds, since the real mingw ioctl() has to operate on the socket rather than the (nicer) fd.
The goal of gnulib is that you never use a SOCKET in any of the POSIX APIs it wraps. It hides the WinSock horribleness so that you can just use normal FDs everywhere.
But use of ioctl() is already non-portable; so a better solution would be one that reduces the amount of direct ioctl in gnulib in the first place (well, some ioctls are necessary, like lxc or kvm startup, but that's in Linux-specific code and not expected to be portable elsewhere, when contrasted with this bug report about the use of ioctl() in the main generic polling loop). -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
[adding bug-gnulib and Paolo as author of rpl_ioctl] On 03/15/2011 02:29 AM, Matthias Bolte wrote:
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored.
virEventRegisterDefaultImpl calls virEventPollInit that calls virSetNonBlock that calls ioctl that is replaced by gnulib and calls ioctlsocket. ioctlsocket fails because the given FD is a pipe but ioctlsocket expects a socket.
A version that works on pipes on Windows looks like this. Although the pipe is actually not a named pipe the call to SetNamedPipeHandleState doesn't fail at least.
int virSetPipeNonBlock(int fd) { DWORD mode = PIPE_NOWAIT; HANDLE handle = _get_osfhandle(fd) BOOL result = SetNamedPipeHandleState(handle, &mode, NULL, NULL);
return result ? 0 : -1; }
So, is the event loop stuff supposed to work on Windows at all and we should get it fixed? Or do we just put an #ifndef WIN32 around virEventRegisterDefaultImpl in virsh, because the only event loop user in virsh is the console command that is disabled on Windows anyway?
Right now, gnulib's rpl_ioctl doesn't do any inspection of the various requests handed it, nor does it check whether the request makes sense on a non-socket. Sniffing the FIONBIO request and handling it for pipes as well as sockets might make sense. And while gnulib has rpl_fcntl for mingw, it does not yet support F_SETFL or F_GETFL to inspect/set O_NONBLOCK on non-file fds. I'm not sure how much effort that would be to add; I suppose that if F_GETFL doesn't support all possible flags, but does support enough that we could detect whether a read-modify-write F_GETFL/F_SETFL sequence is trying to set or clear just O_NONBLOCK, that it would be sufficient for the task at hand. But rather than exposing fcntl, maybe the better thing to do is have gnulib provide a simple wrapper API module that controls whether an fd is blocking (similar to how the gnulib cloexec module hides fcntl(F_GETFD/F_SETFD) or a mingw counterpart). Paolo, any thoughts on the best approach to take? (I know which way I'm leaning, but want some feedback before I give away my bias). -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
On 03/29/2011 01:27 AM, Eric Blake wrote:
Paolo, any thoughts on the best approach to take? (I know which way I'm leaning, but want some feedback before I give away my bias).
Without guessing what your bias is, I also :) prefer to implement {g,s}et_nonblock_flag functions. It would use either SetNamedPipeHandleState or ioctlsocket (using the socket detection trick in sockets.c to detect sockets, and then GetFileType to detect pipes if it fails). This module can then be used to build a fcntl F_GETFL/F_SETFL implementation, but it is not very important to do so. Paolo
On Tue, Mar 29, 2011 at 8:48 AM, Paolo Bonzini <bonzini@gnu.org> wrote:
On 03/29/2011 01:27 AM, Eric Blake wrote:
Paolo, any thoughts on the best approach to take? (I know which way I'm leaning, but want some feedback before I give away my bias).
Without guessing what your bias is, I also :) prefer to implement {g,s}et_nonblock_flag functions. It would use either SetNamedPipeHandleState or ioctlsocket (using the socket detection trick in sockets.c to detect sockets, and then GetFileType to detect pipes if it fails).
This module can then be used to build a fcntl F_GETFL/F_SETFL implementation, but it is not very important to do so.
Could you also implement SOCK_NONBLOCK ? Bastien
Paolo
On 03/15/2011 02:29 AM, Matthias Bolte wrote:
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored.
Does this patch fix things? If so, we should definitely get it into 0.9.0: https://www.redhat.com/archives/libvir-list/2011-March/msg01553.html (but be sure to pick up the logic correction before testing): https://www.redhat.com/archives/libvir-list/2011-March/msg01554.html -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
2011/4/1 Eric Blake <eblake@redhat.com>:
On 03/15/2011 02:29 AM, Matthias Bolte wrote:
Commit 2ed6cc7bec41dd344d41ea1531f6760c93099128 "Expose event loop implementation as a public API" turned a failure to initialize the default event loop into a fatal error in virsh on Windows. Before that commit such a failure was ignored.
Does this patch fix things? If so, we should definitely get it into 0.9.0: https://www.redhat.com/archives/libvir-list/2011-March/msg01553.html
(but be sure to pick up the logic correction before testing): https://www.redhat.com/archives/libvir-list/2011-March/msg01554.html
Yes, at least virSetNonBlock doesn't fail anymore. But now virSetCloseExec is the next thing that fails, because it calls virSetInherit and virSetInherit always returns -1 on Windows. Matthias
participants (5)
-
Bastien ROUCARIES -
Daniel P. Berrange -
Eric Blake -
Matthias Bolte -
Paolo Bonzini