Re: [Qemu-devel] [PATCH v2] qpci_free_pc: sdhci-test and vhost-user-test could free() NULL pointers.

[Paolo Bonzini <pbonzini@redhat.com>]

On 02/07/2018 16:05, Emanuele Giuseppe Esposito wrote:
> @@ -152,6 +152,8 @@ QPCIBus *qpci_init_pc(QTestState *qts, QGuestAllocator *alloc)
>  
>  void qpci_free_pc(QPCIBus *bus)
>  {
> +    g_assert(bus);
> +
>      QPCIBusPC *s = container_of(bus, QPCIBusPC, bus);
>  
>      g_free(s);
> diff --git a/tests/sdhci-test.c b/tests/sdhci-test.c
> index 1d825eb010..9b486b93bf 100644
> --- a/tests/sdhci-test.c
> +++ b/tests/sdhci-test.c
> @@ -209,7 +209,9 @@ static QSDHCI *machine_start(const struct sdhci_t *test)
>  
>  static void machine_stop(QSDHCI *s)
>  {
> -    qpci_free_pc(s->pci.bus);
> +    if (s->pci.bus) {
> +        qpci_free_pc(s->pci.bus);
> +    }

Sorry for chiming in just now.

In general, freeing a NULL pointer is a fine thing to do in C.  In your
code you do

    QPCIBusPC *ret = g_new0(QPCIBusPC, 1);
    qpci_set_pc(ret, qts, alloc);
    return &ret->bus;

But now &ret->bus can be inside the pointer.  qpci_free_pc must
therefore check for NULL before doing the container_of.

It is debatable whether this change should go in QEMU before your code,
or together with it.  There are good arguments for both sides:

- the container_of is assuming that QPCIBus is the first field of the
struct, but that's a strange assumption: container_of usually is used to
go from an interior pointer to an outside struct, and passing NULL to it
is usually wrong

- but, the struct _does_ have QPCIBus as the first field, so we can
assume that if bus == NULL, s will be NULL too.  And g_free(NULL) is okay.

I suggest that you add the "if (!bus) { return; }" in your code, in the
same patch that adds the field before QPCIBusPC.

Thanks,

Paolo