Skip to contentSkip to conversation

act_ife: load meta modules before tcf_idr_check_alloc()

PATCH In review Cong Wang 3 participants 4 messages Sep 4, 2020 – Sep 7, 2020

Subsystems (3)
lore
Share & tools
lore

Cong Wang

author
#0 lore raw

PATCH 1/1 act_ife: load meta modules before tcf_idr_check_alloc()

The following deadlock scenario is triggered by syzbot:

Thread A: Thread B:
tcf_idr_check_alloc()
...
populate_metalist()
rtnl_unlock()

					rtnl_lock()
					...

request_module() tcf_idr_check_alloc()
rtnl_lock()

At this point, thread A is waiting for thread B to release RTNL lock, while thread B is waiting for thread A to commit the IDR change with tcf_idr_insert() later.

Break this deadlock situation by preloading ife modules earlier, before tcf_idr_check_alloc(), this is fine because we only need to load modules we need potentially.

Reported-and-tested-by [email protected]
Fixes 0190c1d452a9 ("net: sched: atomically check-allocate action")
Cc Jamal Hadi Salim <[email protected]>
Cc Vlad Buslov <[email protected]>
Cc Jiri Pirko <[email protected]>
---
 net/sched/act_ife.c | 44 ++++++++++++++++++++++++++++++++++----------
 1 file changed, 34 insertions(+), 10 deletions(-)
net/sched/act_ife.c +34 -10
index c1fcd85719d6..5c568757643b 100644
@@ -436,6 +436,25 @@ static void tcf_ife_cleanup(struct tc_action *a)
 		kfree_rcu(p, rcu);
 }
 
+static int load_metalist(struct nlattr **tb, bool rtnl_held)
+{
+	int i;
+
+	for (i = 1; i < max_metacnt; i++) {
+		if (tb[i]) {
+			void *val = nla_data(tb[i]);
+			int len = nla_len(tb[i]);
+			int rc;
+
+			rc = load_metaops_and_vet(i, val, len, rtnl_held);
+			if (rc != 0)
+				return rc;
+		}
+	}
+
+	return 0;
+}
+
 static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 			     bool exists, bool rtnl_held)
 {
@@ -449,10 +468,6 @@ static int populate_metalist(struct tcf_ife_info *ife, struct nlattr **tb,
 			val = nla_data(tb[i]);
 			len = nla_len(tb[i]);
 
-			rc = load_metaops_and_vet(i, val, len, rtnl_held);
-			if (rc != 0)
-				return rc;
-
 			rc = add_metainfo(ife, i, val, len, exists);
 			if (rc)
 				return rc;
@@ -509,6 +524,21 @@ static int tcf_ife_init(struct net *net, struct nlattr *nla,
 	if (!p)
 		return -ENOMEM;
 
+	if (tb[TCA_IFE_METALST]) {
+		err = nla_parse_nested_deprecated(tb2, IFE_META_MAX,
+						  tb[TCA_IFE_METALST], NULL,
+						  NULL);
+		if (err) {
+			kfree(p);
+			return err;
+		}
+		err = load_metalist(tb2, rtnl_held);
+		if (err) {
+			kfree(p);
+			return err;
+		}
+	}
+
 	index = parm->index;
 	err = tcf_idr_check_alloc(tn, &index, a, bind);
 	if (err < 0) {
@@ -570,15 +600,9 @@ static int tcf_ife_init(struct net *net, struct nlattr *nla,
 	}
 
 	if (tb[TCA_IFE_METALST]) {
-		err = nla_parse_nested_deprecated(tb2, IFE_META_MAX,
-						  tb[TCA_IFE_METALST], NULL,
-						  NULL);
-		if (err)
-			goto metadata_parse_err;
 		err = populate_metalist(ife, tb2, exists, rtnl_held);
 		if (err)
 			goto metadata_parse_err;
-
 	} else {
 		/* if no passed metadata allow list or passed allow-all
 		 * then here we process by adding as many supported metadatum
Signature · 1 line
2.28.0
maintainer
#1 lore raw
Quoted 16 lines · On Thu, 3 Sep 2020 19:10:11 -0700 Cong Wang wrote:

The following deadlock scenario is triggered by syzbot:

Thread A: Thread B:
tcf_idr_check_alloc()
...
populate_metalist()
rtnl_unlock()

					rtnl_lock()
					...

request_module() tcf_idr_check_alloc()
rtnl_lock()

At this point, thread A is waiting for thread B to release RTNL lock, while thread B is waiting for thread A to commit the IDR change with tcf_idr_insert() later.

Quoted 8 lines

Break this deadlock situation by preloading ife modules earlier, before tcf_idr_check_alloc(), this is fine because we only need to load modules we need potentially.

Reported-and-tested-by [email protected]
Fixes 0190c1d452a9 ("net: sched: atomically check-allocate action")
Cc Jamal Hadi Salim <[email protected]>
Cc Vlad Buslov <[email protected]>

Vlad, it'd have been nice to see your review tag here.

Quoted 2 lines

LGTM, applied and queued for stable, thank you Cong!

#2 lore raw
Quoted 21 lines · Jakub Kicinski
Quoted 20 lines · On Thu, 3 Sep 2020 19:10:11 -0700 Cong Wang wrote:

The following deadlock scenario is triggered by syzbot:

Thread A: Thread B:
tcf_idr_check_alloc()
...
populate_metalist()
rtnl_unlock()

					rtnl_lock()
					...

request_module() tcf_idr_check_alloc()
rtnl_lock()

At this point, thread A is waiting for thread B to release RTNL lock, while thread B is waiting for thread A to commit the IDR change with tcf_idr_insert() later.

Break this deadlock situation by preloading ife modules earlier, before tcf_idr_check_alloc(), this is fine because we only need to load modules we need potentially.

Quoted 6 lines
Quoted 4 lines
Reported-and-tested-by [email protected]
Fixes 0190c1d452a9 ("net: sched: atomically check-allocate action")
Cc Jamal Hadi Salim <[email protected]>
Cc Vlad Buslov <[email protected]>

Vlad, it'd have been nice to see your review tag here.

Reviewed. Sorry for the delay.

Quoted 5 lines
Quoted 3 lines

LGTM, applied and queued for stable, thank you Cong!

Reviewed-by Tested-by on v1 1/1
#3 lore raw
In reply to Cong Wang
Quoted 20 lines · Cong Wang

The following deadlock scenario is triggered by syzbot:

Thread A: Thread B:
tcf_idr_check_alloc()
...
populate_metalist()
rtnl_unlock()

					rtnl_lock()
					...

request_module() tcf_idr_check_alloc()
rtnl_lock()

At this point, thread A is waiting for thread B to release RTNL lock, while thread B is waiting for thread A to commit the IDR change with tcf_idr_insert() later.

Break this deadlock situation by preloading ife modules earlier, before tcf_idr_check_alloc(), this is fine because we only need to load modules we need potentially.

Quoted 7 lines
Reported-and-tested-by [email protected]
Fixes 0190c1d452a9 ("net: sched: atomically check-allocate action")
Cc Jamal Hadi Salim <[email protected]>
Cc Vlad Buslov <[email protected]>
Cc Jiri Pirko <[email protected]>

---

Thanks for fixing this, Cong! I've verified that all tdc ife tests pass with this patch.