Re: [PATCH v2 1/5] qemu/qarray.h: introduce QArray

["Daniel P. Berrangé" <berrange@redhat.com>]

On Thu, Sep 30, 2021 at 03:55:36PM +0200, Christian Schoenebeck wrote:
> On Donnerstag, 30. September 2021 15:31:10 CEST Daniel P. Berrangé wrote:
> > On Thu, Sep 30, 2021 at 03:20:19PM +0200, Christian Schoenebeck wrote:
> > > On Mittwoch, 29. September 2021 19:48:38 CEST Daniel P. Berrangé wrote:
> > > > On Wed, Sep 29, 2021 at 07:32:39PM +0200, Christian Schoenebeck wrote:
> > > > > On Dienstag, 28. September 2021 18:41:17 CEST Daniel P. Berrangé 
> wrote:
> > > > > > On Tue, Sep 28, 2021 at 06:23:23PM +0200, Christian Schoenebeck 
> wrote:
> > > > > > > On Dienstag, 28. September 2021 15:04:36 CEST Daniel P. Berrangé
> > > 
> > > wrote:
> > > > > > > > On Sun, Aug 22, 2021 at 03:16:46PM +0200, Christian Schoenebeck
> > > 
> > > wrote:
> > > > > [...]
> > > > > 
> > > > > > The GLib automatic memory support is explicitly designed to be
> > > > > > extendd
> > > > > > with support for application specific types. We already do exactly
> > > > > > that
> > > > > > all over QEMU with many calls to G_DEFINE_AUTOPTR_CLEANUP_FUNC(..)
> > > > > > to
> > > > > > register functions for free'ing specific types, such that you can
> > > > > > use 'g_autoptr' with them.
> > > > > 
> > > > > Ok, just to make sure that I am not missing something here, because
> > > > > really
> > > > > if there is already something that does the job that I simply haven't
> > > > > seen, then I happily drop this QArray code.
> > > > 
> > > > I don't believe there is anything that currently addresses this well.
> > > > 
> > > > > But AFAICS this G_DEFINE_AUTOPTR_CLEANUP_FUNC() & g_autoptr concept
> > > > > does
> > > > > not have any notion of "size" or "amount", right?
> > > > 
> > > > Correct, all it knows is that there's a data type and an associated
> > > > free function.
> > > 
> > > Ok, thanks for the clarification.
> > > 
> > > > > So let's say you already have the following type and cleanup function
> > > > > in
> > > > > your existing code:
> > > > > 
> > > > > typedef struct MyScalar {
> > > > > 
> > > > >     int a;
> > > > >     char *b;
> > > > > 
> > > > > } MyScalar;
> > > > > 
> > > > > void myscalar_free(MayScalar *s) {
> > > > > 
> > > > >     g_free(s->b);
> > > > > 
> > > > > }
> > > > > 
> > > > > Then if you want to use G_DEFINE_AUTOPTR_CLEANUP_FUNC() for an array
> > > > > on
> > > > > that scalar type, then you still would need to *manually* write
> > > > > additionally a separate type and cleanup function like:
> > > > > 
> > > > > typedef struct MyArray {
> > > > > 
> > > > >     MyScalar *s;
> > > > >     int n;
> > > > > 
> > > > > };
> > > > > 
> > > > > void myarray_free(MyArray *a) {
> > > > > 
> > > > >     for (int i = 0; i < a->n; ++i) {
> > > > >     
> > > > >         myscalar_free(a->s[i]);
> > > > >     
> > > > >     }
> > > > >     g_free(a);
> > > > > 
> > > > > }
> > > > > 
> > > > > Plus you have to manually populate that field 'n' after allocation.
> > > > > 
> > > > > Am I wrong?
> > > > 
> > > > Yes and no.  You can of course manually write all these stuff
> > > > as you describe, but since we expect the array wrappers to be
> > > > needed for more than one type it makes more sense to have
> > > > that all done via macros.
> > > > 
> > > > Your patch contains a DECLARE_QARRAY_TYPE and DEFINE_QARRAY_TYPE
> > > > that provide all this reqiured boilerplate code.  The essential
> > > > difference that I'm suggesting is that the array struct type emitted
> > > > by the macro is explicitly visible as a concept to calling code such
> > > > that it is used directly used with g_autoptr.
> > > 
> > > I got that, but your preferred user pattern was this:
> > >     DECLARE_QARRAY_TYPE(Foo);
> > > 	 
> > > 	 ...
> > > 	 
> > >     g_autoptr(FooArray) foos = foo_array_new(n);
> > > 
> > > I don't see a portable way to do upper-case to lower-case conversion with
> > > the> 
> > > C preprocessor. So you would end up like this instead:
> > >     g_autoptr(FooArray) foos = Foo_array_new(n);
> > > 
> > > Which does not really fit into common QEMU naming conventions either, does
> > > it?
> > Right, it would need to be a two arg macro:
> > 
> >   DECLARE_QARRAY_TYPE(Foo, foo);
> > 
> > similar to what we do with macros for declaring QOM types becuase of
> > the same case conversion needs.
> > 
> > > And I can help it, I don't see what's wrong in exposing a regular C-array
> > > to user code. I mean in the Linux kernel for instance it is absolutely
> > > normal to convert from a compound structure to its parent structure. I
> > > don't find anything magical about that and it is simply less code and
> > > better readable.
> > QEMU code is not Linux code. We're following the GLib practices for
> > automatic memory deallocation, and QOM is also modelled on GLib. The
> > proposal looks magical from the POV of QEMU's code patterns, as it is
> > not making use of GLib's g_auto* code.
> 
> Hmm, I start to think whether I should just make it some 9p local utility code 
> for now instead, e.g. "P9Array" or something.

IMHO even if it was private to a subsystem it should still be using the
standard g_auto functionality for automatically deallocating memory,
because this is a QEMU wide standard.


Regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|