When a server decides to close a client, the
virNetServerClientCloseLocked() is called. In here various
cleanup steps are taken, but the most important part (from this
commit's POV at least) is the way that the socket is closed.
Firstly, removal of the socket associated with the client from
the event loop is signalized and then the socket is unrefed. The
socket is not closed just yet though, because the event loop
holds a reference to it. This reference will be freed as soon as
the event loop wakes up and starts issuing callbacks (in this
case virNetSocketEventFree()).
So far, this is how things usually work. But if the daemon
reaches the number of opened files limit, things start to work
differently.
If the RLIMIT_NOFILE limit is reached and there's a client that
wants to connect then the event loop wakes up, sees POLLIN on the
socket and calls virNetServerServiceAccept() which in turn calls
virNetSocketAccept(). But because of the limit, accept() fails
with EMFILE leaving the POLLIN event unhandled. The dispatch then
continues to next FDs with events on them. BUT, it will NOT call
the socket removal callback (virNetSocketEventFree()) because it
has low priority (G_PRIORITY_DEFAULT_IDLE). Per glib's
documentation:
* Each event source is assigned a priority. The default priority,
* %G_PRIORITY_DEFAULT, is 0. Values less than 0 denote higher priorities.
* Values greater than 0 denote lower priorities. Events from high priority
* sources are always processed before events from lower priority sources.
and per g_idle_add() documentation:
* Adds a function to be called whenever there are no higher priority
* events pending to the default main loop. The function is given the
* default idle priority, %G_PRIORITY_DEFAULT_IDLE.
Now, because we did not accept() the client we are constantly
seeing POLLIN on the main socket and thus the removal of the
client socket won't ever happen.
The fix is to set at least the same priority as other sources,
but since we want to just close an FD, let's give it the highest
priority and call it before handling other events.
This issue can be easily reproduced, for instance:
# ulimit -S -n 40 (tweak this number if needed)
# ./src/libvirtd
from another terminal:
# for ((i=0; i<100; i++)); do virsh list & done; virsh list
The last `virsh list` must not get stuck.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=2007168
Signed-off-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
The g_idle_add function adds a callback to the primary GMainContext.
To workaround the GSource unref bugs, we need to add our callbacks
to the GMainContext that is associated with the GSource being
unref'd. Thus code using the per-VM virEventThread must use its
private GMainContext.
Reviewed-by: Martin Kletzander <mkletzan@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
This way it can be used from other places as well.
Signed-off-by: Martin Kletzander <mkletzan@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
I previously did a workaround for a glib event loop race
that causes crashes:
commit 0db4743645
Author: Daniel P. Berrangé <berrange@redhat.com>
Date: Tue Jul 28 16:52:47 2020 +0100
util: avoid crash due to race in glib event loop code
it turns out that the workaround has a significant performance
penalty on I/O intensive workloads. We thus need to avoid the
workaround if we know we have a new enough glib to avoid the
race condition.
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Tested-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
There is a fairly long standing race condition bug in glib which can hit
if you call g_source_destroy or g_source_unref from a non-main thread:
https://gitlab.gnome.org/GNOME/glib/-/merge_requests/1358
Unfortunately it is really common for libvirt to call g_source_destroy
from a non-main thread. This glib bug is the cause of non-determinstic
crashes in eventtest, and probably in libvirtd too.
To work around the problem we need to ensure that we never release
the last reference on a GSource from a non-main thread. The previous
patch replaced our use of g_source_destroy with a pair of
g_source_remove and g_source_unref. We can now delay the g_source_unref
call by using a idle callback to invoke it from the main thread which
avoids the race condition.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
The source ID number is an alternative way to identify a source that has
been added to a GMainContext. Internally when a source ID is given, glib
will lookup the corresponding GSource and use that. The use of a source
ID is racy in some cases though, because it is invalid to continue to
use an ID number after the GSource has been removed. It is thus safer
to use the GSource object directly and have full control over the ref
counting and thus cleanup.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
The libvirt-glib project has provided a GMainContext based
event loop impl for applications. This imports it and sets
it up for use by libvirt as the primary event loop. This
remains a private impl detail of libvirt.
IOW, applications must *NOT* assume that a call to
"virEventRegisterDefaultImpl" results in a GLib based
event loop. They should continue to use the libvirt-glib
API gvir_event_register() if they explicitly want to
guarantee a GLib event loop.
This follows the general principle that the libvirt public
API should not expose the fact that GLib is being used
internally.
Reviewed-by: Pavel Hrdina <phrdina@redhat.com>
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
