On Fri, Dec 21, 2018 at 12:41:31PM +0100, Pablo Neira Ayuso wrote: > On Wed, Dec 12, 2018 at 08:04:12PM +0100, Phil Sutter wrote: > > Kernel prefers to identify chain by handle if it was given which causes > > manual traversal of the chain list. In contrast, chain lookup by name in > > kernel makes use of a hash table so is considerably faster. Force this > > code path by removing the cached chain's handle when removing it. > > > > Signed-off-by: Phil Sutter <[email protected]> > > --- > > iptables/nft.c | 1 + > > 1 file changed, 1 insertion(+) > > > > diff --git a/iptables/nft.c b/iptables/nft.c > > index 5ef3a75efcde5..8ff21e09f0344 100644 > > --- a/iptables/nft.c > > +++ b/iptables/nft.c > > @@ -1643,6 +1643,7 @@ static int __nft_chain_user_del(struct nftnl_chain > > *c, void *data) > > fprintf(stdout, "Deleting chain `%s'\n", > > nftnl_chain_get_str(c, NFTNL_CHAIN_NAME)); > > > > + nftnl_chain_unset(c, NFTNL_CHAIN_HANDLE); > > LGTM. > > We can probably add a hashtable for chain handle lookups in the > kernel too so we can re-enable this in the future.
Florian suggested to use an rbtree and some seqcount trickery to avoid concurrency issues. I already started with that but decided I need to learn more about kernel locking mechanisms before I submit something stupid. :) Cheers, Phil
