[libvirt] Memory leak in libvirt / gnutls
I'm trying to track down: https://bugzilla.redhat.com/show_bug.cgi?id=810613 I *think* this is a libvirt bug, but I can only reproduce it on one machine, not on another machine that has a virtually identical setup. Anyway, something is confusing me about libvirt. In this commit: ---------------------------------------------------------------------- commit 74c75671331d284e1f777f9692b72e9737520bf0 Author: Michal Privoznik <mprivozn@redhat.com> Date: Thu Aug 18 10:44:08 2011 +0200 daemon: initialize GnuTLS When spice_tls is set but listen_tls is not, we don't initialize GnuTLS library. So any later gnutls call (e.g. during migration, where we initialize a certificate) will access uninitialized GnuTLS internal structs and throws an error. Although, we might now initialize GnuTLS twice, it is safe according to the documentation: This function can be called many times, but will only do something the first time. This patch creates 2 functions: virNetTLSInit and virNetTLSDeinit with respect to written above. ---------------------------------------------------------------------- ... a pair of functions called virNetTLSInit & virNetTLSDeinit are introduced, which would appear to do the right thing, calling first gnutls_global_init, then gnutls_global_deinit. This looks correct to me. However the above commit is later amended by this commit: ---------------------------------------------------------------------- commit eaddec976ef06457fee4a4ce86b8c7ee906183b7 Author: Michal Privoznik <mprivozn@redhat.com> Date: Wed Aug 24 16:16:45 2011 +0200 daemon: Move TLS initialization to virInitialize My previous patch 74c75671331d284e1f777f9692b72e9737520bf0 introduced a regression by removing TLS initialization from client. ---------------------------------------------------------------------- which removes virNetTLSDeinit. This appears to be a mistake, or at least I can't see the logical reason for it, and according to the gnutls docs, it would introduce a memory leak looking exactly like the one I am chasing down. The current code mentions virNetTLSDeinit only in a comment, and never calls gnutls_global_deinit. Any ideas? Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones virt-top is 'top' for virtual machines. Tiny program with many powerful monitoring features, net stats, disk stats, logging, etc. http://et.redhat.com/~rjones/virt-top
On 04/07/2012 03:33 AM, Richard W.M. Jones wrote:
However the above commit is later amended by this commit:
---------------------------------------------------------------------- commit eaddec976ef06457fee4a4ce86b8c7ee906183b7 Author: Michal Privoznik <mprivozn@redhat.com> Date: Wed Aug 24 16:16:45 2011 +0200
daemon: Move TLS initialization to virInitialize
My previous patch 74c75671331d284e1f777f9692b72e9737520bf0 introduced a regression by removing TLS initialization from client. ----------------------------------------------------------------------
which removes virNetTLSDeinit. This appears to be a mistake, or at least I can't see the logical reason for it, and according to the gnutls docs, it would introduce a memory leak looking exactly like the one I am chasing down.
I remember asking at the time, and seem to remember this answer: gnutls_global_init is not thread-safe, and therefore must not be called in the context of a library that might be in use by a multi-threaded parent application. Same goes for gnutls_global_deinit. We'd rather leak the memory (which isn't really a leak unless you dynamically load and unload libvirt multiple times in your app) than risk thread unsafety, all while waiting for gnutls to fix their thread-safety issue. -- Eric Blake eblake@redhat.com +1-919-301-3266 Libvirt virtualization library http://libvirt.org
On Mon, Apr 09, 2012 at 10:19:38AM -0600, Eric Blake wrote:
On 04/07/2012 03:33 AM, Richard W.M. Jones wrote:
However the above commit is later amended by this commit:
---------------------------------------------------------------------- commit eaddec976ef06457fee4a4ce86b8c7ee906183b7 Author: Michal Privoznik <mprivozn@redhat.com> Date: Wed Aug 24 16:16:45 2011 +0200
daemon: Move TLS initialization to virInitialize
My previous patch 74c75671331d284e1f777f9692b72e9737520bf0 introduced a regression by removing TLS initialization from client. ----------------------------------------------------------------------
which removes virNetTLSDeinit. This appears to be a mistake, or at least I can't see the logical reason for it, and according to the gnutls docs, it would introduce a memory leak looking exactly like the one I am chasing down.
I remember asking at the time, and seem to remember this answer:
gnutls_global_init is not thread-safe, and therefore must not be called in the context of a library that might be in use by a multi-threaded parent application.
OK, although this function we *do* call :-)
Same goes for gnutls_global_deinit.
There's still a confusing comment in libvirt. Take a look at the very bottom of 'src/rpc/virnettlscontext.c'.
We'd rather leak the memory (which isn't really a leak unless you dynamically load and unload libvirt multiple times in your app) than risk thread unsafety, all while waiting for gnutls to fix their thread-safety issue.
I had a look at the initialization code in gnutls, and it is pretty complex ... Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones virt-df lists disk usage of guests without needing to install any software inside the virtual machine. Supports Linux and Windows. http://et.redhat.com/~rjones/virt-df/
On Tue, Apr 10, 2012 at 12:08:14PM +0100, Richard W.M. Jones wrote:
On Mon, Apr 09, 2012 at 10:19:38AM -0600, Eric Blake wrote:
On 04/07/2012 03:33 AM, Richard W.M. Jones wrote:
However the above commit is later amended by this commit:
---------------------------------------------------------------------- commit eaddec976ef06457fee4a4ce86b8c7ee906183b7 Author: Michal Privoznik <mprivozn@redhat.com> Date: Wed Aug 24 16:16:45 2011 +0200
daemon: Move TLS initialization to virInitialize
My previous patch 74c75671331d284e1f777f9692b72e9737520bf0 introduced a regression by removing TLS initialization from client. ----------------------------------------------------------------------
which removes virNetTLSDeinit. This appears to be a mistake, or at least I can't see the logical reason for it, and according to the gnutls docs, it would introduce a memory leak looking exactly like the one I am chasing down.
I remember asking at the time, and seem to remember this answer:
gnutls_global_init is not thread-safe, and therefore must not be called in the context of a library that might be in use by a multi-threaded parent application.
OK, although this function we *do* call :-)
Same goes for gnutls_global_deinit.
There's still a confusing comment in libvirt. Take a look at the very bottom of 'src/rpc/virnettlscontext.c'.
Yep that comment is bogus 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 :|
participants (3)
-
Daniel P. Berrange -
Eric Blake -
Richard W.M. Jones