Chris Lalancette <clalance(a)redhat.com> wrote:
For 0.4.3, danpb's new memory management scheme went into
libvirt. This is
fine, except that is subtly alters the semantics of malloc(), calloc(), and
realloc(). In particular, if you say:
foo = malloc(0);
glibc will happily return a non-NULL pointer to you. However, with the new
memory management stuff, if you say:
foo = VIR_ALLOC(0);
you will actually get a NULL pointer back. Personally, I think this is a
dangerous deviation from malloc() semantics that everyone is used to, and is
indeed causing problems with the remote driver. The short of it is that the
remote driver allocates memory on behalf of the remote side using VIR_ALLOC_N,
and this call is returning NULL so that the NULL checks elsewhere in the code
fire and return failure.
The attached patch fixes this situation by removing the 0 checks from the memory
allocation paths, and just lets them fall through to the normal malloc(),
calloc(), or realloc() routines, restoring old semantics.
Wow. That's a nasty one, violating such assumptions.
Good catch.
In case anyone reading along wonders about the portability of libvirt's
assumption that those functions return non-NULL for N=0, it's fine
because gnulib wrappers protect us from libc functions with the
undesirable semantics.
ACK.