netlink_delinearize: incorrect meta protocol dependency kill
Pablo Neira Ayuso
maintainerPATCH 1/1 netlink_delinearize: incorrect meta protocol dependency kill
meta protocol is meaningful in bridge, netdev and inet familiiess, do not remove this.
--- src/netlink_delinearize.c | 22 ++++++++++++-- tests/py/bridge/meta.t | 3 ++ tests/py/bridge/meta.t.json | 54 ++++++++++++++++++++++++++++++++++ tests/py/bridge/meta.t.payload | 18 ++++++++++++ tests/py/inet/meta.t | 4 +++ tests/py/inet/meta.t.json | 54 ++++++++++++++++++++++++++++++++++ tests/py/inet/meta.t.payload | 18 ++++++++++++ tests/py/ip/meta.t | 2 ++ tests/py/ip/meta.t.json | 16 ++++++++++ tests/py/ip/meta.t.payload | 9 ++++++ tests/py/ip6/meta.t | 3 ++ tests/py/ip6/meta.t.json | 54 ++++++++++++++++++++++++++++++++++ tests/py/ip6/meta.t.payload | 18 ++++++++++++ 13 files changed, 272 insertions(+), 3 deletions(-)
src/netlink_delinearize.c +19 -3
@@ -1993,7 +1993,7 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx, const struct expr *expr) {struct expr *dep = ctx->pdep->expr; -uint16_t l3proto; +uint16_t l3proto, protocol;uint8_t l4proto;if (ctx->pbase !=PROTO_BASE_NETWORK_HDR )
@@ -2005,7 +2005,22 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,case NFPROTO_BRIDGE :break ;default : -return true; +if (dep->left->etype !=EXPR_META || + dep->right->etype !=EXPR_VALUE ) +return false; + +if (dep->left->meta.key ==NFT_META_PROTOCOL ) { + protocol =mpz_get_uint16 (dep->right->value); + +if (family ==NFPROTO_IPV4 && + protocol ==ETH_P_IP ) +return true; +else if (family ==NFPROTO_IPV6 && + protocol ==ETH_P_IPV6 ) +return true; + } + +return false; }if (expr->left->meta.key !=NFT_META_L4PROTO )
@@ -2015,7 +2030,8 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,switch (dep->left->etype) {case EXPR_META : -if (dep->left->meta.key !=NFT_META_NFPROTO ) +if (dep->left->meta.key !=NFT_META_NFPROTO && + dep->left->meta.key !=NFT_META_PROTOCOL )return true;break ;case EXPR_PAYLOAD :
tests/py/bridge/meta.t +3 -0
@@ -6,3 +6,6 @@ meta obrname "br0";ok meta ibrname "br0";ok meta ibrvproto vlan;ok;meta ibrvproto 8021q meta ibrpvid 100;ok + +meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok
tests/py/bridge/meta.t.json +54 -0
57 lines not on this page · raw message
tests/py/bridge/meta.t.payload +18 -0
@@ -17,3 +17,21 @@ bridge test-bridge input bridge test-bridge input [ meta load bri_iifpvid => reg 1 ] [ cmp eq reg 1 0x00000064 ] + +# meta protocol ip udp dport 67 +bridge + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x00000008 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ] + +# meta protocol ip6 udp dport 67 +bridge + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x0000dd86 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ]
tests/py/inet/meta.t +4 -0
@@ -12,6 +12,10 @@ meta nfproto ipv4 tcp dport 22;ok meta nfproto ipv4 ip saddr 1.2.3.4;ok;ip saddr 1.2.3.4 meta nfproto ipv6 meta l4proto tcp;ok;meta nfproto ipv6 meta l4proto 6 meta nfproto ipv4 counter ip saddr 1.2.3.4;ok + +meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok + meta ipsec exists;ok meta secpath missing;ok;meta ipsec missing meta ibrname "br0";fail
tests/py/inet/meta.t.json +54 -0
57 lines not on this page · raw message
tests/py/inet/meta.t.payload +18 -0
@@ -79,3 +79,21 @@ inet test-inet input [ ct load mark => reg 1 ] [ bitwise reg 1 = ( reg 1 >> 0x00000008 ) ] [ meta set mark with reg 1 ] + +# meta protocol ip udp dport 67 +inet + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x00000008 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ] + +# meta protocol ip6 udp dport 67 +inet + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x0000dd86 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ]
tests/py/ip/meta.t +2 -0
@@ -8,6 +8,8 @@ meta l4proto ipv6-icmp icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-adv meta l4proto 58 icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-advert icmpv6 type nd-router-advert;ok +meta protocol ip udp dport 67;ok + meta ibrname "br0";fail meta obrname "br0";fail
tests/py/ip/meta.t.json +16 -0
@@ -140,3 +140,19 @@ "accept": null } ] + +# meta protocol ip udp dport 67 +[ + { + "match": { + "left": { + "payload": { + "field": "dport", + "protocol": "udp" + } + }, + "op": "==", + "right": 67 + } + } +]
tests/py/ip/meta.t.payload +9 -0
@@ -44,3 +44,12 @@ ip6 test-ip4 input [ meta load sdifname => reg 1 ] [ cmp neq reg 1 0x31667276 0x00000000 0x00000000 0x00000000 ] [ immediate reg 0 accept ] + +# meta protocol ip udp dport 67 +ip + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x00000008 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ]
tests/py/ip6/meta.t +3 -0
@@ -9,5 +9,8 @@ meta l4proto icmp icmp type echo-request;ok;icmp type echo-request meta l4proto 1 icmp type echo-request;ok;icmp type echo-request icmp type echo-request;ok +meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok + meta sdif "lo" accept;ok meta sdifname != "vrf1" accept;ok
tests/py/ip6/meta.t.json +54 -0
57 lines not on this page · raw message
tests/py/ip6/meta.t.payload +18 -0
@@ -44,3 +44,21 @@ ip6 test-ip6 input [ meta load sdifname => reg 1 ] [ cmp neq reg 1 0x31667276 0x00000000 0x00000000 0x00000000 ] [ immediate reg 0 accept ] + +# meta protocol ip udp dport 67 +ip6 + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x00000008 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ] + +# meta protocol ip6 udp dport 67 +ip6 + [ meta load protocol => reg 1 ] + [ cmp eq reg 1 0x0000dd86 ] + [ meta load l4proto => reg 1 ] + [ cmp eq reg 1 0x00000011 ] + [ payload load 2b @ transport header + 2 => reg 1 ] + [ cmp eq reg 1 0x00004300 ]
Signature · 1 line
2.20.1
Florian Westphal
maintainerQuoted 4 lines · Pablo Neira Ayuso
meta protocol is meaningful in bridge, netdev and inet familiiess, do not remove this.
Fixes a1bcf8a34975 ("payload: add payload_may_dependency_kill()")
Looks like its a regression from "netlink_delinearize: Refactor meta_may_dependency_kill()", 056aaa3e6dc65aced5e552233ac3e7f89fb81f86.
Phil Sutter
reviewerHi,
Quoted 2 lines · Pablo Neira Ayuso
meta protocol is meaningful in bridge, netdev and inet familiiess, do not remove this.
So you want to avoid dependency killing in above families, but:
[...]
Quoted 4 lines
src/netlink_delinearize.c +0
-0
[...]
Quoted 14 lines
src/netlink_delinearize.c +9
-1@@ -2005,7 +2005,22 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,case NFPROTO_BRIDGE :break ;default : -return true; +if (dep->left->etype !=EXPR_META || + dep->right->etype !=EXPR_VALUE ) +return false; + +if (dep->left->meta.key ==NFT_META_PROTOCOL ) { + protocol =mpz_get_uint16 (dep->right->value); + +if (family ==NFPROTO_IPV4 && + protocol ==ETH_P_IP )
src/netlink_delinearize.c static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx, · line 2026
Quoted 10 lines
src/netlink_delinearize.c +7
-0+return true; + else if (family == NFPROTO_IPV6 && + protocol == ETH_P_IPV6) + return true; + } + +return false; }if (expr->left->meta.key !=NFT_META_L4PROTO )
The above code only applies to families other than inet, netdev or bridge?! Am I missing something?
AFAIU, dependency killing defaults to true and examines the cases where it should not happen. So the old behaviour was for unexpected dependency expressions to just drop them instead of keeping them. The code you add above does the opposite. That's not wrong per se, but I assume there was a reason why dependency elimination was implemented "greedy", therefore I kept it that way when refactoring the code.
src/netlink_delinearize.c static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx, · line 2037
Quoted 10 lines
src/netlink_delinearize.c +2
-1@@ -2015,7 +2030,8 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,switch (dep->left->etype) {case EXPR_META : -if (dep->left->meta.key !=NFT_META_NFPROTO ) +if (dep->left->meta.key !=NFT_META_NFPROTO && + dep->left->meta.key !=NFT_META_PROTOCOL )return true;break ;case EXPR_PAYLOAD :
If we continue evaluation for 'meta protocol' payload dependencies, RHS values need adjustment: with 'meta nfproto' we'll see NFPROTO_IPV4 or NFPROTO_IPV6 while with 'meta protocol' we'll see ETH_P_IP or ETH_P_IPV6. The EXPR_PAYLOAD case has such conversion code.
[...]
tests/py/ip/meta.t meta l4proto ipv6-icmp icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-adv · line 12
Quoted 10 lines
tests/py/ip/meta.t +2
-0@@ -8,6 +8,8 @@ meta l4proto ipv6-icmp icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-adv meta l4proto 58 icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-advert icmpv6 type nd-router-advert;ok +meta protocol ip udp dport 67;ok +
Hmm. Shouldn't this drop the dependency, i.e. list as 'udp dport 67' as the family is defined by table family already?
There should also be a case for 'meta protocol ip6 udp dport 67' which makes sure the dependency is not dropped, right?
Quoted 17 lines
tests/py/ip/meta.t +0
-0meta ibrname "br0";fail meta obrname "br0";failtests/py/ip/meta.t.json +6
-0@@ -140,3 +140,19 @@ "accept": null } ] + +# meta protocol ip udp dport 67 +[ + { + "match": { + "left": {
tests/py/ip/meta.t.json line 158
Quoted 10 lines
tests/py/ip/meta.t.json +10
-0+ "payload": { + "field": "dport", + "protocol": "udp" + } + }, + "op": "==", + "right": 67 + } + } +]
This translation is not correct, the 'meta protocol' match is missing in JSON. Even though it is not present when listing ruleset, the idea is that JSON is identical to standard syntax input to avoid any "short cuts" (I guess it's just a typo, other spots are fine).
[...]
tests/py/ip6/meta.t meta l4proto icmp icmp type echo-request;ok;icmp type echo-request · line 14
Quoted 11 lines
tests/py/ip6/meta.t +3
-0@@ -9,5 +9,8 @@ meta l4proto icmp icmp type echo-request;ok;icmp type echo-request meta l4proto 1 icmp type echo-request;ok;icmp type echo-request icmp type echo-request;ok +meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok +
Here, 'meta protocol ip6' should be dropped.
Pablo, do you want to have another look at your patch or should I try to fix things up?
Cheers, Phil
Pablo Neira Ayuso
maintainerHi Phil,
Quoted 37 lines · Phil Sutter
Hi,
Quoted 2 lines · Pablo Neira Ayuso
meta protocol is meaningful in bridge, netdev and inet familiiess, do not remove this.
So you want to avoid dependency killing in above families, but:
[...]
Quoted 4 lines
src/netlink_delinearize.c +0
-0[...]
Quoted 23 lines
src/netlink_delinearize.c +16
-1@@ -2005,7 +2005,22 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,case NFPROTO_BRIDGE :break ;default : -return true; +if (dep->left->etype !=EXPR_META || + dep->right->etype !=EXPR_VALUE ) +return false; + +if (dep->left->meta.key ==NFT_META_PROTOCOL ) { + protocol =mpz_get_uint16 (dep->right->value); + +if (family ==NFPROTO_IPV4 && + protocol ==ETH_P_IP ) +return true; +else if (family ==NFPROTO_IPV6 && + protocol ==ETH_P_IPV6 ) +return true; + } + +return false; }
Quoted 4 lines
src/netlink_delinearize.c static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx, · line 2026
Quoted 1 line
src/netlink_delinearize.c +0
-0if (expr->left->meta.key !=NFT_META_L4PROTO )The above code only applies to families other than inet, netdev or bridge?! Am I missing something?
Yes, this code only applies to ip and ip6, this removes "meta protocol" in this cases:
rule ip x y meta protocol ip udp dport 67 => udp dport 67 rule ip6 x y meta protocol ip6 udp dport 67 => udp dport 67
I should have keep it in a separated patch actually since it is not related to bridge/inet/netdev. Sorry for the noise.
Quoted 9 lines
AFAIU, dependency killing defaults to true and examines the cases where it should not happen. So the old behaviour was for unexpected dependency expressions to just drop them instead of keeping them. The code you add above does the opposite. That's not wrong per se, but I assume there was a reason why dependency elimination was implemented "greedy", therefore I kept it that way when refactoring the code.
Quoted 2 lines
src/netlink_delinearize.c +0
-0@@ -2015,7 +2030,8 @@ static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx,
Quoted 13 lines
src/netlink_delinearize.c static bool meta_may_dependency_kill(struct payload_dep_ctx *ctx, · line 2037
Quoted 8 lines
src/netlink_delinearize.c +2
-1switch (dep->left->etype) {case EXPR_META : -if (dep->left->meta.key !=NFT_META_NFPROTO ) +if (dep->left->meta.key !=NFT_META_NFPROTO && + dep->left->meta.key !=NFT_META_PROTOCOL )return true;break ;case EXPR_PAYLOAD :If we continue evaluation for 'meta protocol' payload dependencies, RHS values need adjustment: with 'meta nfproto' we'll see
NFPROTO_IPV4orNFPROTO_IPV6while with 'meta protocol' we'll seeETH_P_IPorETH_P_IPV6. TheEXPR_PAYLOADcase has such conversion code.
The chunk above the actual fix for the reported issue.
Yes, checking for ETH_P_IP and ETH_P_IPV6 would be good to narrow down the dependency killing more precisely.
Quoted 9 lines
[...]
Quoted 8 lines
tests/py/ip/meta.t +0
-0@@ -8,6 +8,8 @@ meta l4proto ipv6-icmp icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-adv meta l4proto 58 icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-advert icmpv6 type nd-router-advert;ok
Quoted 5 lines
tests/py/ip/meta.t meta l4proto ipv6-icmp icmpv6 type nd-router-advert;ok;icmpv6 type nd-router-adv · line 12
Quoted 2 lines
tests/py/ip/meta.t +2
-0+meta protocol ip udp dport 67;ok +Hmm. Shouldn't this drop the dependency, i.e. list as 'udp dport 67' as the family is defined by table family already?
It does actually, this spews a warning if you run tests/py.
Quoted 2 lines
There should also be a case for 'meta protocol ip6 udp dport 67' which makes sure the dependency is not dropped, right?
Yes, I forgot to add it to ip/meta.t
I added it to ip6/meta.t
tests/py/ip6/meta.t +2 -0
@@ -9,5 +9,8 @@ meta l4proto icmp icmp type echo-request;ok;icmp type echo-request meta l4proto 1 icmp type echo-request;ok;icmp type echo-request icmp type echo-request;ok +meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok
Quoted 24 lines
Quoted 24 lines
tests/py/ip/meta.t +0
-0meta ibrname "br0";fail meta obrname "br0";failtests/py/ip/meta.t.json +13
-0@@ -140,3 +140,19 @@ "accept": null } ] + +# meta protocol ip udp dport 67 +[ + { + "match": { + "left": { + "payload": { + "field": "dport", + "protocol": "udp" + } + }, + "op": "==", + "right": 67
Quoted 8 lines
tests/py/ip/meta.t.json line 158
Quoted 3 lines
tests/py/ip/meta.t.json +3
-0+ } + } +]This translation is not correct, the 'meta protocol' match is missing in JSON. Even though it is not present when listing ruleset, the idea is that JSON is identical to standard syntax input to avoid any "short cuts" (I guess it's just a typo, other spots are fine).
If I understand correctly the json tests, this actually checking for the listing output, which has already removed the 'meta protocol' match.
IIRC, if I add 'meta protocol' here, the json tests report a warning.
Quoted 9 lines
[...]
Quoted 8 lines
tests/py/ip6/meta.t +0
-0@@ -9,5 +9,8 @@ meta l4proto icmp icmp type echo-request;ok;icmp type echo-request meta l4proto 1 icmp type echo-request;ok;icmp type echo-request icmp type echo-request;ok
Quoted 5 lines
tests/py/ip6/meta.t meta l4proto icmp icmp type echo-request;ok;icmp type echo-request · line 14
Quoted 3 lines
tests/py/ip6/meta.t +3
-0+meta protocol ip udp dport 67;ok +meta protocol ip6 udp dport 67;ok +Here, 'meta protocol ip6' should be dropped.
No, it is checking that this is removed. Note that this spews a warning in tests/py when running tests.
Look: If we want to fix this right(tm), we should update payload_try_merge() to actually remove this redundant dependency at rule load time, instead of doing this lazy dependency removal at listing time.
Quoted 2 lines
Pablo, do you want to have another look at your patch or should I try to fix things up?
Sure, you can follow up to refine. You're welcome.
Thanks for reviewing.
Pablo Neira Ayuso
maintainerOn Mon, Aug 30, 2021 at 07:21:14PM +0200, Pablo Neira Ayuso wrote:
[...]
Quoted 4 lines
Look: If we want to fix this right(tm), we should update
payload_try_merge()to actually remove this redundant dependency at rule load time, instead of doing this lazy dependency removal at listing time.
Or actually, this match statement could be just canceled from the evaluation phase for these cases:
ip family meta protocol ip
ip6 family meta protocol ip6