netfilter: nf_tables: fix addition/deletion of elements from commit/abort
Subsystems (2)
Pablo Neira Ayuso
maintainerPATCH 1/1 netfilter: nf_tables: fix addition/deletion of elements from commit/abort
We have several problems in this path:
1) There is a use-after-free when removing individual elements from the commit path.
2) We have to uninit() the data part of the element from the abort path to avoid a chain refcount leak.
3) We have to check for set->flags to see if there's a mapping, instead of the element flags.
4) We have to check for !(flags & NFT_SET_ELEM_INTERVAL_END) to skip elements that are part of the interval that have no data part, so they don't need to be uninit().
---
This fix replaces the following two patches:
http://patchwork.ozlabs.org/patch/440781/
http://patchwork.ozlabs.org/patch/440782/
Thanks to Patrick for helping on discussing this fix!
net/netfilter/nf_tables_api.c | 21 ++++++++++++--------- 1 file changed, 12 insertions(+), 9 deletions(-)
net/netfilter/nf_tables_api.c +12 -9
@@ -3612,12 +3612,11 @@ static int nf_tables_commit(struct sk_buff *skb) &te->elem, NFT_MSG_DELSETELEM, 0); te->set->ops->get (te->set, &te->elem); - te->set->ops->remove (te->set, &te->elem);nft_data_uninit (&te->elem.key,NFT_DATA_VALUE ); -if (te->elem.flags &NFT_SET_MAP ) { -nft_data_uninit (&te->elem.data, - te->set->dtype); - } +if (te->set->flags &NFT_SET_MAP && + !(te->elem.flags &NFT_SET_ELEM_INTERVAL_END )) +nft_data_uninit (&te->elem.data, te->set->dtype); + te->set->ops->remove (te->set, &te->elem);nft_trans_destroy (trans);break ; }
@@ -3658,7 +3657,7 @@ static int nf_tables_abort(struct sk_buff *skb) {struct net *net =sock_net (skb->sk);struct nft_trans *trans, *next; -struct nft_set *set; +struct nft_trans_elem *te;list_for_each_entry_safe (trans, next, &net->nft.commit_list, list) {switch (trans->msg_type) {
@@ -3719,9 +3718,13 @@ static int nf_tables_abort(struct sk_buff *skb)break ;case NFT_MSG_NEWSETELEM :nft_trans_elem_set (trans)->nelems--; - set =nft_trans_elem_set (trans); - set->ops->get (set, &nft_trans_elem (trans)); - set->ops->remove (set, &nft_trans_elem (trans)); + te = (struct nft_trans_elem *)trans->data; + te->set->ops->get (te->set, &te->elem); +nft_data_uninit (&te->elem.key,NFT_DATA_VALUE ); +if (te->set->flags &NFT_SET_MAP && + !(te->elem.flags &NFT_SET_ELEM_INTERVAL_END )) +nft_data_uninit (&te->elem.data, te->set->dtype); + te->set->ops->remove (te->set, &te->elem);nft_trans_destroy (trans);break ;case NFT_MSG_DELSETELEM :
Signature · 1 line
1.7.10.4
Quoted 8 lines · Pablo Neira Ayuso
We have several problems in this path:
1) There is a use-after-free when removing individual elements from the commit path.
2) We have to
uninit()the data part of the element from the abort path to avoid a chain refcount leak.
Quoted 6 lines
3) We have to check for
set->flagsto see if there's a mapping, instead of the element flags.4) We have to check for !(flags &
NFT_SET_ELEM_INTERVAL_END) to skip elements that are part of the interval that have no data part, so they don't need to beuninit().
Just wondering, in the delete case, don't we need to set the flags in the sets' ->get() function for this to work?
Pablo Neira Ayuso
maintainerQuoted 12 lines · Patrick McHardy
Quoted 11 lines · Pablo Neira Ayuso
We have several problems in this path:
1) There is a use-after-free when removing individual elements from the commit path.
2) We have to
uninit()the data part of the element from the abort path to avoid a chain refcount leak.3) We have to check for
set->flagsto see if there's a mapping, instead of the element flags.
Quoted 6 lines
Quoted 3 lines
4) We have to check for !(flags &
NFT_SET_ELEM_INTERVAL_END) to skip elements that are part of the interval that have no data part, so they don't need to beuninit().Just wondering, in the delete case, don't we need to set the flags in the sets' ->get() function for this to work?
They are already set from hash and rbtree, so we only need to add the check for NFT_SET_ELEM_INTERVAL_END from the commit path in nf_tables_api.c
Unless you have any further concern, I'll pass up this soon.
Quoted 17 lines · Pablo Neira Ayuso
Quoted 16 lines · Patrick McHardy
Quoted 14 lines · Pablo Neira Ayuso
We have several problems in this path:
1) There is a use-after-free when removing individual elements from the commit path.
2) We have to
uninit()the data part of the element from the abort path to avoid a chain refcount leak.3) We have to check for
set->flagsto see if there's a mapping, instead of the element flags.4) We have to check for !(flags &
NFT_SET_ELEM_INTERVAL_END) to skip elements that are part of the interval that have no data part, so they don't need to beuninit().
Quoted 7 lines
Quoted 2 lines
Just wondering, in the delete case, don't we need to set the flags in the sets' ->get() function for this to work?
They are already set from hash and rbtree, so we only need to add the check for
NFT_SET_ELEM_INTERVAL_ENDfrom the commit path innf_tables_api.cUnless you have any further concern, I'll pass up this soon.
Right, in the hash case it's 0 anyways and rbtree does set them.
Looks good to me.
Pablo Neira Ayuso
maintainerQuoted 21 lines · Patrick McHardy
Quoted 20 lines · Pablo Neira Ayuso
Quoted 18 lines · Patrick McHardy
Quoted 8 lines · Pablo Neira Ayuso
We have several problems in this path:
1) There is a use-after-free when removing individual elements from the commit path.
2) We have to
uninit()the data part of the element from the abort path to avoid a chain refcount leak.Quoted 6 lines
3) We have to check for
set->flagsto see if there's a mapping, instead of the element flags.4) We have to check for !(flags &
NFT_SET_ELEM_INTERVAL_END) to skip elements that are part of the interval that have no data part, so they don't need to beuninit().Just wondering, in the delete case, don't we need to set the flags in the sets' ->get() function for this to work?
Quoted 8 lines
Quoted 4 lines
They are already set from hash and rbtree, so we only need to add the check for
NFT_SET_ELEM_INTERVAL_ENDfrom the commit path innf_tables_api.cUnless you have any further concern, I'll pass up this soon.
Right, in the hash case it's 0 anyways and rbtree does set them.
Looks good to me.
Ok, I'll pass this to David, thanks.