On 12/2/2023 20:53, Florian Westphal wrote:
Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote:
One way would be to return 0 in that case (in
nf_conntrack_hash_check_insert()). What do you think?
This is misleading to the user that adds an entry via ctnetlink?
ETIMEDOUT also looks a bit confusing to report to userspace.
Rewinding: if the intention is to deal with stale conntrack extension,
for example, helper module has been removed while this entry was
added. Then, probably call EAGAIN so nfnetlink has a chance to retry
transparently?
Seems we first need to add a "bool *inserted" so we know when the ct
entry went public.
I don't think so.
nf_conntrack_hash_check_insert(struct nf_conn *ct)
{
...
/* The caller holds a reference to this object */
refcount_set(&ct->ct_general.use, 2); // [1]
__nf_conntrack_hash_insert(ct, hash, reply_hash);
nf_conntrack_double_unlock(hash, reply_hash);
NF_CT_STAT_INC(net, insert);
local_bh_enable();
if (!nf_ct_ext_valid_post(ct->ext)) {
nf_ct_kill(ct); // [2]
NF_CT_STAT_INC_ATOMIC(net, drop);
return -ETIMEDOUT;
}
...
}
We set ct->ct_general.use to 2 in nf_conntrack_hash_check_insert()([1]).
nf_ct_kill willn't put the last refcount. So ct->master will not be
freed in this way. But this means the situation not only causes
ct->master's refcount leak but also releases ct whose refcount is still
1 in nf_conntrack_free() (in ctnetlink_create_conntrack() err1).
I think it may be a good idea to set ct->ct_general.use to 0 after
nf_ct_kill() ([2]) to put the caller's reference. What do you think?
Thanks,
Hangyu
I'll also have a look at switching to a refcount based model for
all extensions that reference external objects, this would avoid
the entire problem, but thats likely more intrusive.