nfp: flower-ct: remove callback delete deadlock

The current location of the callback delete can lead to a race
condition where deleting the callback requires a write_lock on
the nf_table, but at the same time another thread from netfilter
could have taken a read lock on the table before trying to offload.
Since the driver is taking a rtnl_lock this can lead into a deadlock
situation, where the netfilter offload will wait for the cls_flower
rtnl_lock to be released, but this cannot happen since this is
waiting for the nf_table read_lock to be released before it can
delete the callback.

Solve this by completely removing the nf_flow_table_offload_del_cb
call, as this will now be cleaned up by act_ct itself when cleaning
up the specific nf_table.

Fixes: 62268e7814 ("nfp: flower-ct: add nft callback stubs")
Signed-off-by: Louis Peens <louis.peens@corigine.com>
Signed-off-by: Yinjun Zhang <yinjun.zhang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
Louis Peens 2021-07-02 11:21:39 +02:00 committed by David S. Miller
parent 77ac5e40c4
commit 7cc93d888d

View file

@ -1141,20 +1141,7 @@ int nfp_fl_ct_del_flow(struct nfp_fl_ct_map_entry *ct_map_ent)
nfp_fl_ct_clean_flow_entry(ct_entry);
kfree(ct_map_ent);
/* If this is the last pre_ct_rule it means that it is
* very likely that the nft table will be cleaned up next,
* as this happens on the removal of the last act_ct flow.
* However we cannot deregister the callback on the removal
* of the last nft flow as this runs into a deadlock situation.
* So deregister the callback on removal of the last pre_ct flow
* and remove any remaining nft flow entries. We also cannot
* save this state and delete the callback later since the
* nft table would already have been freed at that time.
*/
if (!zt->pre_ct_count) {
nf_flow_table_offload_del_cb(zt->nft,
nfp_fl_ct_handle_nft_flow,
zt);
zt->nft = NULL;
nfp_fl_ct_clean_nft_entries(zt);
}