netfilter: nft_set_rbtree: bogus lookup/get on consecutive elements in named sets
Subsystems (2)
Pablo Neira Ayuso
maintainerPATCH 1/1 netfilter: nft_set_rbtree: bogus lookup/get on consecutive elements in named sets
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)
This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
--- net/netfilter/nft_set_rbtree.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-)
net/netfilter/nft_set_rbtree.c +15 -4
@@ -74,8 +74,13 @@ static bool __nft_rbtree_lookup(const struct net *net, const struct nft_set *set parent =rcu_dereference_raw (parent->rb_left);continue ; } -if (nft_rbtree_interval_end (rbe)) - goto out; +if (nft_rbtree_interval_end (rbe)) { +if (set->flags &NFT_SET_ANONYMOUS ) +return false; + parent =rcu_dereference_raw (parent->rb_left); + interval =NULL ; +continue ; + } *ext = &rbe->ext;return true;
@@ -88,7 +93,7 @@ static bool __nft_rbtree_lookup(const struct net *net, const struct nft_set *set *ext = &interval->ext;return true; } -out: +return false; }
@@ -141,6 +146,8 @@ static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, } else {if (!nft_set_elem_active (&rbe->ext, genmask)) parent =rcu_dereference_raw (parent->rb_left); +continue ; + }if (!nft_set_ext_exists (&rbe->ext,NFT_SET_EXT_FLAGS ) || (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==
@@ -148,7 +155,11 @@ static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, *elem = rbe;return true; } -return false; + +if (nft_rbtree_interval_end (rbe)) + interval =NULL ; + + parent =rcu_dereference_raw (parent->rb_left); } }
Signature · 1 line
2.11.0
Phil Sutter
reviewerHi Pablo,
Quoted 11 lines · Pablo Neira Ayuso
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
I didn't get what the actual problem is?
[...]
Quoted 4 lines
net/netfilter/nft_set_rbtree.c +0
-0
[...]
net/netfilter/nft_set_rbtree.c static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, · line 153
Quoted 9 lines
net/netfilter/nft_set_rbtree.c +2
-0@@ -141,6 +146,8 @@ static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, } else {if (!nft_set_elem_active (&rbe->ext, genmask)) parent =rcu_dereference_raw (parent->rb_left); +continue ; + }if (!nft_set_ext_exists (&rbe->ext,NFT_SET_EXT_FLAGS ) || (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==
Are you sure about that chunk? It adds a closing brace without a matching opening one. Either this patch ignores whitespace change or there's something fishy. :)
Cheers, Phil
Pablo Neira Ayuso
maintainerQuoted 8 lines · Phil Sutter
Hi Pablo,
Quoted 5 lines · Pablo Neira Ayuso
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)
Quoted 8 lines
Quoted 6 lines
This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
I didn't get what the actual problem is?
The lookup/get results false, while there is an element in the rbtree. Moreover, the get operation returns true as if a+2 would be in the tree. This happens with named sets after several set updates, I could reproduce the issue with several elements mixed with insertion and deletions in one batch.
I managed to trigger the problem using the test file coming in this
patch: https://patchwork.ozlabs.org/patch/1204779/
I can extend the patch description if you like.
Quoted 13 lines
[...]
Quoted 4 lines
net/netfilter/nft_set_rbtree.c +0
-0[...]
Quoted 7 lines
net/netfilter/nft_set_rbtree.c +2
-0@@ -141,6 +146,8 @@ static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, }else {if (!nft_set_elem_active (&rbe->ext, genmask)) parent =rcu_dereference_raw (parent->rb_left); +continue ; + }
Quoted 6 lines
net/netfilter/nft_set_rbtree.c static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, · line 153
Quoted 2 lines
net/netfilter/nft_set_rbtree.c +0
-0if (!nft_set_ext_exists (&rbe->ext,NFT_SET_EXT_FLAGS ) || (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==Are you sure about that chunk? It adds a closing brace without a matching opening one. Either this patch ignores whitespace change or there's something fishy. :)
Yes, I botched it when squashing my patches before submission. I just sent a v2.
Thanks.
Pablo Neira Ayuso
maintainerQuoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Phil Sutter
Hi Pablo,
Quoted 11 lines · Pablo Neira Ayuso
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
Quoted 7 lines
Quoted 1 line
I didn't get what the actual problem is?
The lookup/get results false, while there is an element in the rbtree. Moreover, the get operation returns true as if a+2 would be in the tree. This happens with named sets after several set updates, I could reproduce the issue with several elements mixed with insertion and deletions in one batch.
To extend the problem description: The issue is that the existing lookup optimization (that only works for the anonymous sets) might not reach the opening [ a+1, ... element if the closing ... , a+1) is found in first place when walking over the rbtree. Hence, walking the full tree in that case is needed.
Quoted 31 lines
I managed to trigger the problem using the test file coming in this
patch: https://patchwork.ozlabs.org/patch/1204779/I can extend the patch description if you like.
Quoted 13 lines
[...]
Quoted 4 lines
net/netfilter/nft_set_rbtree.c +0
-0[...]
Quoted 7 lines
net/netfilter/nft_set_rbtree.c +2
-0@@ -141,6 +146,8 @@ static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, }else {if (!nft_set_elem_active (&rbe->ext, genmask)) parent =rcu_dereference_raw (parent->rb_left); +continue ; + }Quoted 6 lines
net/netfilter/nft_set_rbtree.c static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set, · line 153
Quoted 2 lines
net/netfilter/nft_set_rbtree.c +0
-0if (!nft_set_ext_exists (&rbe->ext,NFT_SET_EXT_FLAGS ) || (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==Are you sure about that chunk? It adds a closing brace without a matching opening one. Either this patch ignores whitespace change or there's something fishy. :)
Yes, I botched it when squashing my patches before submission. I just sent a v2.
Thanks.
Phil Sutter
reviewerHi Pablo,
Quoted 25 lines · Pablo Neira Ayuso
Quoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Phil Sutter
Hi Pablo,
Quoted 11 lines · Pablo Neira Ayuso
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
Quoted 7 lines
Quoted 1 line
I didn't get what the actual problem is?
The lookup/get results false, while there is an element in the rbtree. Moreover, the get operation returns true as if a+2 would be in the tree. This happens with named sets after several set updates, I could reproduce the issue with several elements mixed with insertion and deletions in one batch.
Quoted 5 lines
To extend the problem description: The issue is that the existing lookup optimization (that only works for the anonymous sets) might not reach the opening [ a+1, ... element if the closing ... , a+1) is found in first place when walking over the rbtree. Hence, walking the full tree in that case is needed.
Ah! Thanks a lot for your elaborations. It was really hard to grasp what all this is about from the initial commit message. :)
Sometimes I wonder if we should do more set optimizations under the hood when adding elements. Right now, we don't touch existing ones although it would make sense. And we could be more intelligent for example if a set contains 20-30 and a user adds 25-35.
Cheers, Phil
Pablo Neira Ayuso
maintainerQuoted 28 lines · Phil Sutter
Hi Pablo,
Quoted 25 lines · Pablo Neira Ayuso
Quoted 23 lines · Pablo Neira Ayuso
Quoted 8 lines · Phil Sutter
Hi Pablo,
Quoted 5 lines · Pablo Neira Ayuso
The existing rbtree implementation might store consecutive elements where the closing element and the opening element might overlap, eg.
[ a, a+1) [ a+1, a+2)Quoted 8 lines
Quoted 6 lines
This patch removes the optimization for non-anonymous sets in the exact matching case, where it is assumed to stop searching in case that the closing element is found. Instead, invalidate candidate interval and keep looking further in the tree.
This patch fixes the lookup and get operations.
I didn't get what the actual problem is?
The lookup/get results false, while there is an element in the rbtree. Moreover, the get operation returns true as if a+2 would be in the tree. This happens with named sets after several set updates, I could reproduce the issue with several elements mixed with insertion and deletions in one batch.
Quoted 8 lines
Quoted 5 lines
To extend the problem description: The issue is that the existing lookup optimization (that only works for the anonymous sets) might not reach the opening [ a+1, ... element if the closing ... , a+1) is found in first place when walking over the rbtree. Hence, walking the full tree in that case is needed.
Ah! Thanks a lot for your elaborations. It was really hard to grasp what all this is about from the initial commit message. :)
I'll append this to the description before applying, no problem, thanks for asking.
Quoted 4 lines
Sometimes I wonder if we should do more set optimizations under the hood when adding elements. Right now, we don't touch existing ones although it would make sense. And we could be more intelligent for example if a set contains 20-30 and a user adds 25-35.
Yes, if 'automerge' is set on, then nft should start doing some more intelligent stuff. Basically the idea is: check if this interval overlaps with an existing one, if so transaction looks like this:
[ delete 25-35 | add 20-35 ]
With the existing transaction infrastructure in place, I think it should not be too hard. There's already a routine to check for overlaps. If this is a data mapping, this needs to be careful to not merge intervals that have different data on the right hand side.
There's another problem, I think incremental userspace cache is still incomplete. I can probably scratch time and look into this, this might not happen before 2020 though.
Thanks.