Re: [Xen-devel] [PATCH 2/6] libxl: events: Deregister xenstore watch fd when not needed

[Ian Jackson <Ian.Jackson@xxxxxxxxxxxxx>]

Ian Campbell writes ("Re: [PATCH 2/6] libxl: events: Deregister xenstore watch 
fd when not needed"):
> On Thu, 2014-11-27 at 18:27 +0000, Ian Jackson wrote:
> > * On libxl teardown, the watch fd should therefore be unregistered.
> >   assert that this is the case.
> 
> A bunch of the patches in this series basically assume that the ctx is
> idle when it is freed, i.e. it requires everything to be explicitly
> cancelled rather than implicitly doing so on free.

libxl_ctx_free explicitly disables all the application-requested event
generators.  (free_disable_deaths and libxl__evdisable_disk_eject.)

Destroying the ctx during the execution of an asynchronous operation
is forbidden by this text in libxl.h (near line 813):
  * *ao_how does not need to remain valid after the initiating function
  * returns. All other parameters must remain valid for the lifetime of
  * the asynchronous operation, unless otherwise specified.
That implies that the ctx must remain valid during the ao, so it may
not be destroyed beforehand.

Those are the two ways that, even without any threads inside libxl, a
ctx can be other than idle.

It should be obvious to the application programmer that destroying the
ctx when there are other threads inside libxl is not going to work.
Indeed a programmer who tries to destroy the ctx when they have
threads which might be inside libxl cannot ensure that the ctx is
valid even on entry to libxl.

> I think this is a fine restriction, but it probably ought to be written
> down.

Does the above demonstrate that the existing restrictions are
documented ?  I'd rather avoid writing the restrictions twice if it
can be avoided - these docs are long enough as they are.

Ian.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel