[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[libvirt] RFC: libxl race fixes

I've been investigating some races in the libxl driver and would like to
get comments on some potential solutions.

The first race is in the fd/timeout event handling code, which maps
libxl's osevent interface to libvirt's event loop interface.  This
mapping opens the possibility for libvirt's event loop to invoke
event callbacks after libxl has already deregistered the event,
potentially accessing an event object that has already been freed.

One solution to this race I've found successful is reference counting the
objects associated with the events.  When libxl registers an event, an
object encapsulating the event is created and it's reference count is set
to 1.  When the event is injected into libvirt's event loop, another
reference is taken on the object.  When libxl deregisters the event, it's
reference count is decremented.  Once the event is removed from libvirt's
event loop, the final reference is decremented and the object is disposed.
This approach ensures the object is not disposed until it is removed
from libvirt's event loop *and* libxl had explicitly deregistered the event.
The notion of an event being 'disabled' found in libvirt's event loop impl
also had to be added to the libxl event object, to ensure the driver doesn't
call into libxl for a previously deregistered event.

The second race is between destroying a vm (i.e., calling
privateDataFreeFunc, which frees the libxl ctx) and deregistration/cleanup
of all events that have been registered by libxl.

One solution for this race is to convert libxlDomainObjPrivate to a
virObject and increment its reference for each fd and timeout registration.
Only when all fds and timeouts are deregistered and destroyed will the
libxlDomainObjPrivate be destroyed.  One downside to this approach is that
an API to "cleanup" the libxl ctx is needed.  Without such an API, some
fds are not deregistered until calling libxl_ctx_free.  But since the fd
events have references on the libxlDomainObjPrivate, libxl_ctx_free is
never called.  (BTW I have a patch adding libxl_ctx_quiesce() to libxl,
which upstream xen folks seem receptive to, including backporting to
Xen 4.2 branch, but I don't think this is ideal.)

An alternate solution that can be used to address both of these races
is to maintain a list of the fd/timeout registrations in the
libxlDomainObjPrivate object, and take a more brute force approach to
managing the registrations

Other solutions that I might be overlooking are certainly appreciated :).


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]