On Wed, May 07, 2008 at 08:25:58AM -0400, Daniel Veillard wrote:
On Tue, May 06, 2008 at 12:51:08AM -0700, Dave Leskovec wrote:
This patch changes the lxc tty forwarding process to use epoll instead of poll. This is done to avoid a cpu consuming loop when a user disconnects from the container console.
During some testing, we found that when the slave end of a tty is closed, calls to poll() on the master end will return immediately with POLLHUP until the slave end is opened again. The result of this is that if you connect to the container console using virsh and then ctrl-] out of it, the container tty process will chew up the processor handling POLLHUP. This event can't be disabled regardless of what is set in the events field.
This can be avoided by using epoll. When used in edge triggered mode, you see the initial EPOLLHUP but will not see another one until someone connects and then disconnects from the console again. This also drove some changes into how the regular input data is handled. Once an EPOLLIN is seen from an fd, another one will not be surfaced until all data has been read from the file (EAGAIN is returned by read).
Sounds fine in principle but i have a couple of questions with the patch
+#include <stdbool.h>
err ... what is that ? looks like a linux specific header, do we really need this ? epoll is linux specific I think but #include <sys/epoll.h> should be sufficient no ?
It is kinda academic whether epool is linux specific - the entire LXC driver is Linux specific, so you're not compiling it on Solaris/Windows anyway. So any Linux-isms in LXC driver code is fine Dan. -- |: Red Hat, Engineering, Boston -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|