[libvirt] [RFC: PATCH 0/4] Update array allocation macros

[Daniel P. Berrange]

On Sat, Aug 14, 2010 at 01:08:33PM -0600, Eric Blake wrote:
> On 08/14/2010 10:23 AM, Eric Blake wrote:
> > So, after some IRC discussions, two ideas emerged in addition to the
> > obvious documentation update.  One is the simpler VIR_EXPAND_N, where
> > the array size always matches the array allocation, along with a
> > VIR_SHRINK_N counterpart which can trim an array back down (and
> > failure to trim need not be fatal).  But this can scale quadratically
> > if you end up with a large array, as the existing contents must be
> > copied on every realloc.  See patches 1 and 2 for an implementation
> > and use.
> > 
> > The other idea is VIR_RESIZE_N, which tracks allocation separately
> > from usage, and has better scaling by doing geometric growth only when
> > needed rather than on every array usage increase.  But it takes more
> > parameters (making it slightly harder to remember how to use
> > correctly), along with possible modifications of the struct tracking
> > an array (so it might not be possible to modify all uses of
> > VIR_REALLOC_N to use this, if it would end up altering a public struct
> > layout).  See patches 3 and 4 for an implementation and use.
> 
> Again on IRC, Matthias had yet another idea, by borrowing after c++'s
> new[] and delete[] operators.  Basically, VIR_ALLOC_N overallocates,
> reserves the first 8 bytes for itself (or, more likely,
> __alignof__(union{size_t;long double;}) bytes, to be precise to platform
> alignment needs), and stores the length of the array in that chunk while
> returning an offset pointer to the user.  That way, the user does not
> have to track allocations; all other VIR_*_N macros would be able to
> look before the user's pointer to find that hidden length field.  On the
> other hand, it means we would need a VIR_FREE_N for arrays, since the
> user's pointer is not the same as the pointer returned by malloc; and
> mixing VIR_ALLOC/VIR_FREE_N or VIR_ALLOC_N/VIR_FREE is a disaster
> waiting to happen.
> 
> It sounds like it might have some appeal for reducing some of the user's
> book-keeping, but would require a careful audit of code to safely match
> VIR_ALLOC_N exactly with VIR_FREE_N.  Thoughts on this approach?

The #1 goal of the memory allocation APIs is to make it hard to make
programming mistakes. Having a VIR_FREE and VIR_FREE_N somewhat
compromises that goal, for only a small convenience, so I don't think
we need to go down that route.

Regards,
Daniel
-- 
|: Red Hat, Engineering, London    -o-   http://people.redhat.com/berrange/ :|
|: http://libvirt.org -o- http://virt-manager.org -o- http://deltacloud.org :|
|: http://autobuild.org        -o-         http://search.cpan.org/~danberr/ :|
|: GnuPG: 7D3B9505  -o-   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|