Hi,
On Tue, May 14, 2019 at 10:51:33AM +0200, Phil Sutter wrote:
> Conditional cache flush logic was inverted.
>
> Fixes: 862818ac3a0de ("xtables: add and use nft_build_cache")
> Signed-off-by: Phil Sutter <phil@xxxxxx>
> ---
> iptables/nft.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/iptables/nft.c b/iptables/nft.c
> index 6354b7e8e72fe..83e0d9a69b37c 100644
> --- a/iptables/nft.c
> +++ b/iptables/nft.c
> @@ -1541,7 +1541,7 @@ void nft_build_cache(struct nft_handle *h)
>
> void nft_rebuild_cache(struct nft_handle *h)
> {
> - if (!h->have_cache)
> + if (h->have_cache)
> flush_chain_cache(h, NULL);
>
> __nft_build_cache(h);
So with this change I broke your transaction reload logic. The problem
is that data in h->obj_list potentially sits in cache, too. At least
rules have to be there so insert with index works correctly. If the
cache is flushed before regenerating the batch, use-after-free occurs
which crashes the program.
We need to either keep the old cache around or keep locally generated
entries when flushing (which might require more intelligent cache
update, too).
I also have a fix for your testcase, will submit that in a minute as
well.
Cheers, Phil
