Skip to contentSkip to conversation

net/sched: keep tcf block alive during offload unbind

PATCH In review Chengfeng Ye 2 participants 2 messages Jul 31

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net/sched: keep tcf block alive during offload unbind

An RTNL-unlocked filter request can race with qdisc teardown on a shared tcf block. The filter request takes a temporary block reference before teardown drops the qdisc owner's reference. Teardown then enters tcf_block_offload_unbind() without retaining a reference of its own.

The following interleaving can occur:

filter request qdisc teardown
tcf_block_refcnt_get(): 1 -> 2

                                   __tcf_block_put(): 2 -> 1
                                   tcf_block_offload_unbind()
  tcf_block_refcnt_put(): 1 -> 0
  tcf_block_destroy()
                                   access block->nooffloaddevcnt

The unbind path may sleep in flow_indr_dev_setup_offload(), allowing the RCU grace period to expire and the block to be reclaimed. It then continues using block->cb_lock and block->nooffloaddevcnt.

KASAN reported:

  BUG: KASAN: slab-use-after-free in tcf_block_offload_unbind
  Read of size 4 at addr ffff88811ae590a4
  Call Trace:
   tcf_block_offload_unbind+0x158/0x190
   clsact_destroy+0xd6/0x690
   __qdisc_destroy+0xdf/0x290
   qdisc_graft+0x5d7/0x1560
   tc_get_qdisc+0x415/0xd50
   rtnetlink_rcv_msg+0x367/0xa40
  Allocated by task 85:
   tcf_block_get_ext+0x803/0x1570
   clsact_init+0x1ff/0xb50
   qdisc_create+0x392/0xde0
  Freed by task 88:
   __kasan_slab_free+0x43/0x70
   __rcu_free_sheaf_prepare+0x5c/0x220
   rcu_free_sheaf+0x1a/0x100

Keep the qdisc owner's reference until offload unbind completes by moving the unbind before the refcount decrement. A concurrent temporary reference release can no longer destroy the block while unbind uses it.

Fixes 92149190067d ("net: sched: flower: set unlocked flag for flower proto ops")
---
 net/sched/cls_api.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
net/sched/cls_api.c +3 -5
index fee4524adc98..d3b2506d87cf 100644
@@ -1312,6 +1312,9 @@ static struct tcf_block *__tcf_block_find(struct net *net, struct Qdisc *q,
 static void __tcf_block_put(struct tcf_block *block, struct Qdisc *q,
 			    struct tcf_block_ext_info *ei, bool rtnl_held)
 {
+	if (q)
+		tcf_block_offload_unbind(block, q, ei);
+
 	if (refcount_dec_and_mutex_lock(&block->refcnt, &block->lock)) {
 		/* Flushing/putting all chains will cause the block to be
 		 * deallocated when last chain is freed. However, if chain_list
@@ -1325,15 +1328,10 @@ static void __tcf_block_put(struct tcf_block *block, struct Qdisc *q,
 		if (tcf_block_shared(block))
 			tcf_block_remove(block, block->net);
 
-		if (q)
-			tcf_block_offload_unbind(block, q, ei);
-
 		if (free_block)
 			tcf_block_destroy(block);
 		else
 			tcf_block_flush_all_chains(block, rtnl_held);
-	} else if (q) {
-		tcf_block_offload_unbind(block, q, ei);
 	}
 }
 
Signature · 1 line
2.43.0
maintainer
#1 lore raw
Quoted 40 lines · Chengfeng Ye

An RTNL-unlocked filter request can race with qdisc teardown on a shared tcf block. The filter request takes a temporary block reference before teardown drops the qdisc owner's reference. Teardown then enters tcf_block_offload_unbind() without retaining a reference of its own.

The following interleaving can occur:

filter request qdisc teardown
tcf_block_refcnt_get(): 1 -> 2

                                   __tcf_block_put(): 2 -> 1
                                   tcf_block_offload_unbind()
  tcf_block_refcnt_put(): 1 -> 0
  tcf_block_destroy()
                                   access block->nooffloaddevcnt

The unbind path may sleep in flow_indr_dev_setup_offload(), allowing the RCU grace period to expire and the block to be reclaimed. It then continues using block->cb_lock and block->nooffloaddevcnt.

KASAN reported:

  BUG: KASAN: slab-use-after-free in tcf_block_offload_unbind
  Read of size 4 at addr ffff88811ae590a4
  Call Trace:
   tcf_block_offload_unbind+0x158/0x190
   clsact_destroy+0xd6/0x690
   __qdisc_destroy+0xdf/0x290
   qdisc_graft+0x5d7/0x1560
   tc_get_qdisc+0x415/0xd50
   rtnetlink_rcv_msg+0x367/0xa40
  Allocated by task 85:
   tcf_block_get_ext+0x803/0x1570
   clsact_init+0x1ff/0xb50
   qdisc_create+0x392/0xde0
  Freed by task 88:
   __kasan_slab_free+0x43/0x70
   __rcu_free_sheaf_prepare+0x5c/0x220
   rcu_free_sheaf+0x1a/0x100
Quoted 4 lines

Keep the qdisc owner's reference until offload unbind completes by moving the unbind before the refcount decrement. A concurrent temporary reference release can no longer destroy the block while unbind uses it.

Please always send a reproducer - either as a tdc test case or if it is sensitive send it privately to me and Cc the other maintainers. I will not look at it without a repro.

cheers,
jamal

Quoted 40 lines
Fixes 92149190067d ("net: sched: flower: set unlocked flag for flower proto ops")
---
 net/sched/cls_api.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
net/sched/cls_api.c +3 -6
index fee4524adc98..d3b2506d87cf 100644
@@ -1312,6 +1312,9 @@ static struct tcf_block *__tcf_block_find(struct net *net, struct Qdisc *q,
 static void __tcf_block_put(struct tcf_block *block, struct Qdisc *q,
                            struct tcf_block_ext_info *ei, bool rtnl_held)
 {
+       if (q)
+               tcf_block_offload_unbind(block, q, ei);
+
        if (refcount_dec_and_mutex_lock(&block->refcnt, &block->lock)) {
                /* Flushing/putting all chains will cause the block to be
                 * deallocated when last chain is freed. However, if chain_list
@@ -1325,15 +1328,10 @@ static void __tcf_block_put(struct tcf_block *block, struct Qdisc *q,
                if (tcf_block_shared(block))
                        tcf_block_remove(block, block->net);
 
-               if (q)
-                       tcf_block_offload_unbind(block, q, ei);
-
                if (free_block)
                        tcf_block_destroy(block);
                else
                        tcf_block_flush_all_chains(block, rtnl_held);
-       } else if (q) {
-               tcf_block_offload_unbind(block, q, ei);
        }
 }
 
--

2.43.0