fix crash bug during rule restore
Florian Westphal
maintainercover fix crash bug during rule restore
https://bugzilla.netfilter.org/show_bug.cgi?id=1351
nft -f <<EOF
flush ruleset
table inet filter {
}
table inet filter {
chain test {
counter
}
}EOF
segfaults during error reporting.
First patch makes error handling more robust.
Second patch passes the right handle -- with above ruleset this highlights "chain test" in the resulting error message.
Last patch skips rule cache updates for invalid op to restore the 0.9.0 behaviour.
Florian Westphal
maintainerPATCH 1/3 src: erec: fall back to internal location if its null
This should never happen (we should pass valid locations to the error reporting functions), but in case we screw up we will segfault during error reporting.
cat crash
table inet filter {
}
table inet filter {
chain test {
counter
}
}"nft -f crash" Now reports:
internal:0:0-0: Error: No such file or directory
... which is both bogus and useless, but better than crashing.
--- src/erec.c | 3 +++ 1 file changed, 3 insertions(+)
src/erec.c +3 -0
@@ -92,6 +92,9 @@ void erec_print(struct output_ctx *octx, const struct error_record *erec,FILE *f;int l; +if (!indesc) + indesc = &internal_indesc; +switch (indesc->type) {case INDESC_BUFFER :case INDESC_CLI :
Signature · 1 line
2.21.0
Pablo Neira Ayuso
maintainerRe: [PATCH nft 1/3] src: erec: fall back to internal location if its null
Quoted 12 lines · Florian Westphal
This should never happen (we should pass valid locations to the error reporting functions), but in case we screw up we will segfault during error reporting.
cat crash
table inet filter {
}table inet filter { chain test { counter } }
Quoted 4 lines
"nft -f crash" Now reports:
internal:0:0-0: Error: No such file or directory... which is both bogus and useless, but better than crashing.
This should not ever happen, right?
Florian Westphal
maintainerRe: [PATCH nft 1/3] src: erec: fall back to internal location if its null
Quoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Florian Westphal
This should never happen (we should pass valid locations to the error reporting functions), but in case we screw up we will segfault during error reporting.
cat crash
table inet filter {
}table inet filter { chain test { counter } }"nft -f crash" Now reports:
internal:0:0-0: Error: No such file or directory
Quoted 3 lines
Quoted 1 line
... which is both bogus and useless, but better than crashing.
This should not ever happen, right?
It happens with current master plus above file.
Pablo Neira Ayuso
maintainerRe: [PATCH nft 1/3] src: erec: fall back to internal location if its null
In reply to Florian WestphalQuoted 12 lines · Florian Westphal
This should never happen (we should pass valid locations to the error reporting functions), but in case we screw up we will segfault during error reporting.
cat crash
table inet filter {
}table inet filter { chain test { counter } }
Quoted 4 lines
"nft -f crash" Now reports:
internal:0:0-0: Error: No such file or directory... which is both bogus and useless, but better than crashing.
I'd suggest we add BUG() here, so we catch missing location information via indesc == NULL. So we can fix the lack of it, otherwise users will rely on internal, which is very limited.
Thanks.
Florian Westphal
maintainerPATCH 2/3 src: evaluate: don't rely on global chain ctx for error reporting
In reply to Florian Westphaltable inet filter {
}
table inet filter {
chain test {
counter
}
}will now report:
crash:7:13-16: Error: No such file or directory
chain test {
^^^^... which is still bogus, but we won't fallback to the 'internal' location anymore and at least somehwat hint that there is a problem with 'test' chain.
The error occurs because we're looking up the chain in the first 'table inet filter' instance, not the second.
--- src/evaluate.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-)
src/evaluate.c +7 -7
@@ -182,17 +182,17 @@ static int table_not_found(struct eval_ctx *ctx) family2str(table->handle.family)); } -static int chain_not_found (struct eval_ctx *ctx) +static int chain_not_found (struct eval_ctx *ctx,struct handle *h) {const struct table *table;struct chain *chain; - chain =chain_lookup_fuzzy (&ctx->cmd->handle, &ctx->nft->cache, &table); + chain =chain_lookup_fuzzy (h, &ctx->nft->cache, &table);if (chain ==NULL ) -return cmd_error (ctx, &ctx->cmd->handle.chain.location, +return cmd_error (ctx, &h->chain.location,"%s" ,strerror (ENOENT )); - return cmd_error(ctx, &ctx->cmd->handle.chain.location, + return cmd_error(ctx, &h->chain.location, "%s; did you mean chain ‘%s’ in table %s ‘%s’?", strerror(ENOENT), chain->handle.chain.name, family2str(chain->handle.family),
@@ -3264,7 +3264,7 @@ static int rule_cache_update(struct eval_ctx *ctx, enum cmd_ops op) chain =chain_lookup (table, &rule->handle);if (!chain) -return chain_not_found (ctx); +return chain_not_found (ctx, &rule->handle);if (rule->handle.index.id) { ref =rule_lookup_by_index (chain, rule->handle.index.id);
@@ -3710,7 +3710,7 @@ static int cmd_evaluate_list(struct eval_ctx *ctx, struct cmd *cmd)return table_not_found (ctx);if (chain_lookup (table, &cmd->handle) ==NULL ) -return chain_not_found (ctx); +return chain_not_found (ctx, &cmd->handle);return 0 ;case CMD_OBJ_QUOTA :
@@ -3843,7 +3843,7 @@ static int cmd_evaluate_rename(struct eval_ctx *ctx, struct cmd *cmd)return table_not_found (ctx);if (chain_lookup (table, &ctx->cmd->handle) ==NULL ) -return chain_not_found (ctx); +return chain_not_found (ctx, &ctx->cmd->handle);break ;default :
Signature · 1 line
2.21.0
Florian Westphal
maintainerPATCH 3/3 src: evaluate: return immediately if no op was requested
In reply to Florian WestphalThis makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}
table inet filter {
chain test {
counter
}
}loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.
--- src/evaluate.c | 3 +++ tests/shell/testcases/cache/0003_cache_update_0 | 12 ++++++++++++ 2 files changed, 15 insertions(+)
src/evaluate.c +3 -0
@@ -3258,6 +3258,9 @@ static int rule_cache_update(struct eval_ctx *ctx, enum cmd_ops op)struct table *table;struct chain *chain; +if (op ==CMD_INVALID ) +return 0 ; + table =table_lookup (&rule->handle, &ctx->nft->cache);if (!table)return table_not_found (ctx);
tests/shell/testcases/cache/0003_cache_update_0 +12 -0
@@ -48,3 +48,15 @@ $NFT -f - >/dev/null <<EOF add rule ip t4 c meta l4proto igmp accept add rule ip t4 c index 2 drop EOF + +# Trigger a crash or rule restore error with nft 0.9.1 +$NFT -f - >/dev/null <<EOF +flush ruleset +table inet testfilter { +} +table inet testfilter { + chain test { + counter + } +} +EOF
Signature · 1 line
2.21.0
Pablo Neira Ayuso
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 11 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }
Quoted 6 lines
loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.Fixes e5382c0d08e3c ("src: Support intra-transaction rule references")
This one is causing the cache corruption, right?
Quoted 43 lines
Signed-off-by Florian Westphal <[email protected]>--- src/evaluate.c | 3 +++ tests/shell/testcases/cache/0003_cache_update_0 | 12 ++++++++++++ 2 files changed, 15 insertions(+)src/evaluate.c +3
-0@@ -3258,6 +3258,9 @@ static int rule_cache_update(struct eval_ctx *ctx, enum cmd_ops op)struct table *table;struct chain *chain; +if (op ==CMD_INVALID ) +return 0 ; + table =table_lookup (&rule->handle, &ctx->nft->cache);if (!table)return table_not_found (ctx);tests/shell/testcases/cache/0003_cache_update_0 +12
-0@@ -48,3 +48,15 @@ $NFT -f - >/dev/null <<EOF add rule ip t4 c meta l4proto igmp accept add rule ip t4 c index 2 drop EOF + +# Trigger a crash or rule restore error with nft 0.9.1 +$NFT -f - >/dev/null <<EOF +flush ruleset +table inet testfilter { +} +table inet testfilter { + chain test { + counter + } +} +EOFSignature · 1 line
2.21.0
Florian Westphal
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 12 lines · Pablo Neira Ayuso
Quoted 11 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }
Quoted 8 lines
Quoted 6 lines
loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.Fixes e5382c0d08e3c ("src: Support intra-transaction rule references")This one is causing the cache corruption, right?
There is no cache corruption. This patch makes us enter a code path that we did not take before.
Pablo Neira Ayuso
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 17 lines · Florian Westphal
Quoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Fixes e5382c0d08e3c ("src: Support intra-transaction rule references")This one is causing the cache corruption, right?
There is no cache corruption. This patch makes us enter a code path that we did not take before.
Sorry, I mean, this is a cache bug :-)
cache_flush() is cheating, it sets flags to CACHE_FULL, hence this enters this codepath we dit not take before. This propagates from the previous logic, as a workaround.
I made this patch, to skip the cache in case "flush ruleset" is requested.
This breaks testcases/transactions/0024rule_0, which is a recent test from Phil to check for intra-transaction references, I don't know yet what makes this code unhappy with my changes.
Phil, would you help me have a look at what assumption breaks? Thanks.
include/cache.h +1 -0
@@ -30,6 +30,7 @@ enum cache_level_flags {NFT_CACHE_CHAIN_BIT |NFT_CACHE_RULE_BIT ,NFT_CACHE_FULL = __NFT_CACHE_MAX_BIT -1 , + NFT_CACHE_FLUSHED = (1 << 31), };#endif /* _NFT_CACHE_H_ */
include/rule.h +0 -1
@@ -679,7 +679,6 @@ extern int do_command(struct netlink_ctx *ctx, struct cmd *cmd);extern unsigned int cache_evaluate (struct nft_ctx *nft,struct list_head *cmds);extern int cache_update (struct nft_ctx *ctx,enum cmd_ops cmd,struct list_head *msgs); -extern void cache_flush (struct nft_ctx *ctx,struct list_head *msgs);extern void cache_release (struct nft_cache *cache);extern bool cache_is_complete (struct nft_cache *cache,enum cmd_ops cmd);
src/cache.c +2 -0
@@ -72,6 +72,8 @@ static unsigned int evaluate_cache_flush(struct cmd *cmd, unsigned int flags) flags |=NFT_CACHE_SET ;break ;case CMD_OBJ_RULESET : + flags |=NFT_CACHE_FLUSHED ; +break ;default : flags =NFT_CACHE_EMPTY ;break ;
src/evaluate.c +0 -1
@@ -3776,7 +3776,6 @@ static int cmd_evaluate_flush(struct eval_ctx *ctx, struct cmd *cmd)switch (cmd->obj) {case CMD_OBJ_RULESET : -cache_flush (ctx->nft, ctx->msgs);break ;case CMD_OBJ_TABLE :/* Flushing a table does not empty the sets in the table nor remove
src/rule.c +9 -15
@@ -234,6 +234,11 @@ static bool cache_is_updated(struct nft_cache *cache, uint16_t genid)return genid && genid == cache->genid; } +static bool cache_is_flushed (struct nft_cache *cache) +{ +return cache->flags &NFT_CACHE_FLUSHED ; +} +int cache_update (struct nft_ctx *nft,unsigned int flags,struct list_head *msgs) {struct netlink_ctx ctx = {
@@ -255,6 +260,9 @@ replay:if (cache->genid)cache_release (cache); +if (cache_is_flushed (cache)) + goto skip; + ret =cache_init (&ctx, flags);if (ret <0 ) {cache_release (cache);
@@ -270,7 +278,7 @@ replay:cache_release (cache); goto replay; } - +skip: cache->genid = genid; cache->flags = flags;return 0 ;
@@ -286,20 +294,6 @@ static void __cache_flush(struct list_head *table_list) } } -void cache_flush (struct nft_ctx *nft,struct list_head *msgs) -{ -struct netlink_ctx ctx = { - .list =LIST_HEAD_INIT (ctx.list), - .nft = nft, - .msgs = msgs, - }; -struct nft_cache *cache = &nft->cache; - -__cache_flush (&cache->list); - cache->genid =mnl_genid_get (&ctx); - cache->flags =NFT_CACHE_FULL ; -} -void cache_release (struct nft_cache *cache) {__cache_flush (&cache->list);
Phil Sutter
reviewerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 32 lines · Pablo Neira Ayuso
Quoted 17 lines · Florian Westphal
Quoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Fixes e5382c0d08e3c ("src: Support intra-transaction rule references")This one is causing the cache corruption, right?
There is no cache corruption. This patch makes us enter a code path that we did not take before.
Sorry, I mean, this is a cache bug :-)
cache_flush()is cheating, it sets flags toCACHE_FULL, hence this enters this codepath we dit not take before. This propagates from the previous logic, as a workaround.
Quoted 8 lines
I made this patch, to skip the cache in case "flush ruleset" is requested.
This breaks testcases/transactions/0024rule_0, which is a recent test from Phil to check for intra-transaction references, I don't know yet what makes this code unhappy with my changes.
Phil, would you help me have a look at what assumption breaks? Thanks.
Sorry, I don't get it. What is happening in the first place? Florian writes, a lookup happens in the wrong table and it seems chain_evaluate() doesn't add the chain to cache. Yet I don't understand why given patch fixes the problem.
Cheers, Phil
Pablo Neira Ayuso
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 40 lines · Phil Sutter
Quoted 39 lines · Pablo Neira Ayuso
Quoted 17 lines · Florian Westphal
Quoted 16 lines · Pablo Neira Ayuso
Quoted 15 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
Fixes e5382c0d08e3c ("src: Support intra-transaction rule references")This one is causing the cache corruption, right?
There is no cache corruption. This patch makes us enter a code path that we did not take before.
Sorry, I mean, this is a cache bug :-)
cache_flush()is cheating, it sets flags toCACHE_FULL, hence this enters this codepath we dit not take before. This propagates from the previous logic, as a workaround.I made this patch, to skip the cache in case "flush ruleset" is requested.
This breaks testcases/transactions/0024rule_0, which is a recent test from Phil to check for intra-transaction references, I don't know yet what makes this code unhappy with my changes.
Quoted 6 lines
Quoted 1 line
Phil, would you help me have a look at what assumption breaks? Thanks.
Sorry, I don't get it. What is happening in the first place? Florian writes, a lookup happens in the wrong table and it seems
chain_evaluate()doesn't add the chain to cache. Yet I don't understand why given patch fixes the problem.
Just sent a patch for this.
Pablo Neira Ayuso
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
In reply to Florian WestphalQuoted 11 lines · Florian Westphal
This makes nft behave like 0.9.0 -- the ruleset
flush ruleset
table inet filter {
}table inet filter { chain test { counter } }
Quoted 5 lines
loads again without generating an error message.
I've added a test case for this, without this it will create an error, and with a checkout of the 'fixes' tag we get crash.
This one should fix this bugzilla:
http://git.netfilter.org/nftables/commit/?id=3ab02db5f836ae0cf9fe7fba616d7eb52139d537
more comments below.
[...]
Quoted 10 lines
tests/shell/testcases/cache/0003_cache_update_0 +2
-0@@ -48,3 +48,15 @@ $NFT -f - >/dev/null <<EOF add rule ip t4 c meta l4proto igmp accept add rule ip t4 c index 2 drop EOF + +# Trigger a crash or rule restore error with nft 0.9.1
tests/shell/testcases/cache/0003_cache_update_0 $NFT -f - >/dev/null <<EOF · line 62
Quoted 10 lines
tests/shell/testcases/cache/0003_cache_update_0 +10
-0+$NFT -f - >/dev/null <<EOF +flush ruleset +table inet testfilter { +} +table inet testfilter { + chain test { + counter + } +} +EOF
I have applied this test as an separated patch, thanks Florian.
Florian Westphal
maintainerRe: [PATCH nft 3/3] src: evaluate: return immediately if no op was requested
Quoted 3 lines · Pablo Neira Ayuso
This one should fix this bugzilla:
http://git.netfilter.org/nftables/commit/?id=3ab02db5f836ae0cf9fe7fba616d7eb52139d537
much better fix, thanks Pablo!