[libvirt] [PATCH] virt-admin: Call virInitialize to fix startup crash

[Daniel P. Berrange]

On Tue, Jun 28, 2016 at 01:01:17PM +0200, Martin Kletzander wrote:
> On Tue, Jun 28, 2016 at 09:57:20AM +0100, Daniel P. Berrange wrote:
> > On Tue, Jun 28, 2016 at 10:49:33AM +0200, Martin Kletzander wrote:
> > > On Tue, Jun 28, 2016 at 10:12:01AM +0200, Martin Kletzander wrote:
> > > > On Mon, Jun 27, 2016 at 02:28:39PM -0400, Cole Robinson wrote:
> > > > > Similar to what virsh and virt-login-shell do
> > > > >
> > > > > https://bugzilla.redhat.com/show_bug.cgi?id=1350315
> > > > > ---
> > > > > I can't actually reproduce the bug, the backtrace is similar to
> > > > > 97973ebb7 which added the same fix for virt-login-shell, and
> > > > > that commit also mentions the randomness of reproducing...
> > > > >
> > > >
> > > > I think we should try finding out what the cause for that is.  It might
> > > > be as simple as adding similar fix to yours and asking the user what
> > > > error does he see with that fix in.  The idea behind that is that
> > > > scripts shouldn't need to call that initialization as that should be
> > > > part of any first call they make to the library.
> > > >
> > > > > tools/virt-admin.c | 5 +++++
> > > > > 1 file changed, 5 insertions(+)
> > > > >
> > > > > diff --git a/tools/virt-admin.c b/tools/virt-admin.c
> > > > > index 4acac65..7bff5c3 100644
> > > > > --- a/tools/virt-admin.c
> > > > > +++ b/tools/virt-admin.c
> > > > > @@ -1371,6 +1371,11 @@ main(int argc, char **argv)
> > > > >         return EXIT_FAILURE;
> > > > >     }
> > > > >
> > > > > +    if (virInitialize() < 0) {
> > > >
> > > > If this is the right thing to do, it should be virAdmInitialize().  By
> > > > using virInitialize() we're giving bad example to others as it only
> > > > works in virt-admin thanks to internal.h being included, I guess.
> > > >
> > > 
> > > I tried this instead, what do you think about it?
> > > 
> > > diff --git i/src/util/virerror.c w/src/util/virerror.c
> > > index 1177570ef0d5..300b0b90252f 100644
> > > --- i/src/util/virerror.c
> > > +++ w/src/util/virerror.c
> > > @@ -37,7 +37,7 @@
> > > 
> > > VIR_LOG_INIT("util.error");
> > > 
> > > -virThreadLocal virLastErr;
> > > +static virThreadLocal virLastErr;
> > > 
> > > virErrorFunc virErrorHandler = NULL;     /* global error handler */
> > > void *virUserData = NULL;        /* associated data */
> > 
> > Has no functional effect.
> > 
> 
> I was under the impression that only static globals were guaranteed to
> be initialized to zeros.
> 
> > > diff --git i/src/util/virevent.c w/src/util/virevent.c
> > > index e0fd35e41644..8f9d15e46454 100644
> > > --- i/src/util/virevent.c
> > > +++ w/src/util/virevent.c
> > > @@ -266,6 +266,9 @@ int virEventRegisterDefaultImpl(void)
> > > {
> > >     VIR_DEBUG("registering default event implementation");
> > > 
> > > +    if (virErrorInitialize() < 0)
> > > +        return -1;
> > > +
> > >     virResetLastError();
> > > 
> > >     if (virEventPollInit() < 0) {
> > 
> > This is just hacking around fact that we didn't call
> > virInitialize/virAdmInitialize().
> > 
> 
> Which is what we have in other function so that users don't have to call
> these specifically.  Each virAdmConnect and virConnect etc. are calling
> respective initializers.

This is the currently documented semantics for the virInitialize
call. It is mandatory except when calling virConnect as the first
API call. It specifically notes you must call virInitialize() if
any other libvirt API call is to be used first.

The difference is that your call to virErrorInitialize is not
thread safe and when virInitialize() is later invoked, you'll get
a second call to virErrorInitialize, which you're then hacking
around by adding the 'bool initialized' attribute to the thread
local.

If we want to make virInitialize() optional, then we should
make sure that virInitialize() itself is invoked at all the
right places - not call virErrorInitialize() directly.

Amuzingly virThreadInitialize() is a no-op since we deleted
the Win32 custom threads impl, so we can just kill that one
off entirely.

> > > diff --git i/src/util/virthread.c w/src/util/virthread.c
> > > index 6c495158f566..4b69451dd355 100644
> > > --- i/src/util/virthread.c
> > > +++ w/src/util/virthread.c
> > > @@ -308,7 +308,8 @@ int virThreadLocalInit(virThreadLocalPtr l,
> > >                        virThreadLocalCleanup c)
> > > {
> > >     int ret;
> > > -    if ((ret = pthread_key_create(&l->key, c)) != 0) {
> > > +    if (!l->initialized &&
> > > +        (ret = pthread_key_create(&l->key, c)) != 0) {
> > >         errno = ret;
> > >         return -1;
> > >     }
> > 
> > Has no functional effect, since !l->initialized will
> > always be true.
> > 
> 
> Yeah there should've been 'l->initialized = true;' here of course.
> 
> > > diff --git i/src/util/virthread.h w/src/util/virthread.h
> > > index e466d9bf0184..1ba8a0fe46bb 100644
> > > --- i/src/util/virthread.h
> > > +++ w/src/util/virthread.h
> > > @@ -54,6 +54,7 @@ typedef virThreadLocal *virThreadLocalPtr;
> > > 
> > > struct virThreadLocal {
> > >     pthread_key_t key;
> > > +    bool initialized;
> > > };
> > > 
> > > typedef struct virThread virThread;


Regards,
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 :|