[libvirt] RFC: Use __attribute__ ((cleanup) in libvirt ?

[Daniel P. Berrange]

On Tue, Jan 10, 2017 at 10:48:47AM +0100, Martin Kletzander wrote:
> On Tue, Jan 10, 2017 at 02:58:21AM -0500, Laine Stump wrote:
> > On 01/09/2017 08:09 PM, Richard W.M. Jones wrote:
> > > On Mon, Jan 09, 2017 at 04:58:49PM +0000, Daniel P. Berrange wrote:
> > > > For those who don't already know, GCC and CLang both implement a C language
> > > > extension that enables automatic free'ing of resources when variables go
> > > > out of scope. This is done by annotating the variable with the "cleanup"
> > > > attribute, pointing to a function the compiler will wire up a call to when
> > > > unwinding the stack. Since the annotation points to an arbitrary user
> > > > defined function, you're not limited to simple free() like semantics. The
> > > > cleanup function could unlock a mutex, or decrement a reference count, etc
> > > > 
> > > > This annotation is used extensively by systemd, and libguestfs, amongst
> > > > other projects. This obviously doesn't bring full garbage collection to
> > > > C, but it does enable the code to be simplified. By removing the need to
> > > > put in many free() (or equiv) calls to cleanup state, the "interesting"
> > > > logic in the code stands out more, not being obscured by cleanup calls
> > > > and goto jumps.
> > > > 
> > > > I'm wondering what people think of making use of this in libvirt ?
> > > > 
> > > > To my mind the only real reason to *not* use it, would be to maintain
> > > > code portability to non-GCC/non-CLang compilers. OS-X, *BSD and *Linux
> > > > all use GCC or CLang or both, so its a non-issue there. So the only place
> > > > this could cause pain is people building libvirt on Win32, who are using
> > > > the Microsoft compilers instead og GCC.
> > > > 
> 
> Only reason I see for not using it is the "temporary" mess it will
> cause.  Yes, we can change to that incrementally, but it will take some
> time and effort and it will never be all of the code that uses it.
> Don't get me wrong, I would love using more builtin compiler features
> and shortening the code here and there.  I'm just worried this
> particular one might be more disrupting than useful.  Most of us are
> pretty used to the code flow we already have and there's nothing you
> can't achieve without the cleanup attribute.
> 
> And yes, I used quotation marks around the word temporary intentionally.

Yes, that's why I thought of it as something that would make for a GSoc
project - have someone do a full conversion of particular areas of code.
eg convert all of util/ or convert the domain XML parser, etc. Basically
if we did it, I think we'd want to have entire files converted at once.
Only converting individual methods ad-hoc would be quite messy.

> > > > IMHO, it is perfectly valid for us to declare that MSVC is unsupported
> > > > with Libvirt and users must use GCC to build on Windows, either natively
> > > > via cygwin, or cross-build from Linux hosts.
> 
> I would love to know if anyone actually tried doing that lately.  Given
> how often we are broken with mingw and we only foind out thanks to our
> test suite (and sometomes the fixing takes more than a release cycle), I
> think nobody does that and from what I know, it might not even work.

We have mingw in the CI system for a while now and its generally fixed
as quickly as native arch builds are fixed these days.


> > > (2) You must not write code like:
> > > 
> > >    fn ()
> > >    {
> > >      CLEANUP_FREE char *v; // uninitialized
> > > 
> > >      if (some error condition) {
> > >        return -1;
> > >      }
> > >      ...
> > >    }
> > > 
> > > because that will call free (v) on the uninitialized variable.
> > > Sometimes GCC can spot this.  In libguestfs we tend to initialize
> > > every CLEANUP_* variable to either an explicit value or NULL.  GCC
> > > optimizes away calls to free (NULL).
> > 
> 
> I'm trying to initialize all variables, always, so I don't see this as a
> problem, but there are some of us that (I have the feeling) are trying
> to initialize as few as possible, so this (although it's a different
> story) might still be a problem for someone.

We pretty much have the same problem already with 'goto cleanup' - you
have to make sure everything is initialized sanely before the first
"goto cleanup". So I think we're safe in this respect already and
the cleanup attributes wouldn't make it any more complex.

> > You've covered one of the worries that I had about it (accidentally
> > marking for CLEANUP a pointer whose value gets returned, and the fact
> > that you can't use it for the cleanup of objects that would have
> > normally been returned, in the case that the function encounters an
> > error and has to dump everything). And since the nice cleanup isn't
> > happening for *everything*, people will have to be paying attention to
> > which objects are auto-cleaned up and which aren't, which will
> > inevitably lead to incorrect classification and/or accidentally adding
> > manual cleanup for something that's auto-cleaned or vice versa. (and
> > merging this into the code bit by bit is going to exacerbate this
> > problem). Also, there is something to be said for having all the code
> > that's executed sitting out there in the open in an easy to follow
> > format rather than obscured behind a macro and a compiler directive that
> > points you up to somewhere else.
> > 
> 
> I don't really like our macros around __attribute__ although I
> understand we need to have some of them to be dynamically defined to
> nothing in some cases.  However with __attribute__((cleanup)), we will
> need to have that all the time.  What's even better, you immediatelly
> see what function will be called on the cleanup and you can jump to the
> tag definition more easily.

If we mandate use of gcc / clang, then we wouldn't need to hide it
behind a macro - we'd be able to use it inline. That said, using a
macro makes it smaller and gives a bit of standardization. eg with
libguestfs style:

  #define CLEANUP_FREE __attribute__((cleanup(free)))
  #define CLEANUP_OBJECT_UNREF __attribute__((cleanup(virObjectUnref)))

  CLEANUP_FREE char *str;
  CLEANUP_OBJECT_UNREF virDomainPtr dom;

vs full inline style:

  __attribute__((cleanup(free))) char *str;
  __attribute__((cleanup(virObjectUnref))) virDomainPtr dom;

That said I see systemd took a halfway house

  #define _cleanup_(x) __attribute__((cleanup(x)))

  _cleanup(free) char *str;
  _cleanup(virObjectUnref) virDomainPtr dom;


I think the systemd style is quite reasonable, as its shorter
and the function called is still clear.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://entangle-photo.org       -o-    http://search.cpan.org/~danberr/ :|