[Question]: A non NULL req->sk in tcp_rtx_synack. Not a fastopen connection.

[Question]: A non NULL req->sk in tcp_rtx_synack. Not a fastopen connection.

[Martin KaFai Lau]

Hi,

We are seeing a use-after-free from a bpf prog attached to 
trace_tcp_retransmit_synack. The program passes the req->sk to the 
bpf_sk_storage_get_tracing kernel helper which does check for null before using it.

fastopen is not used.

We got a kfence report on use-after-free (pasted at the end). It is running with 
an older 6.4 kernel and we hardly hit this in production.

 From the upstream code, del_timer_sync() should have been done by 
inet_csk_reqsk_queue_drop() before "req->sk = child;" is assigned in 
inet_csk_reqsk_queue_add(). My understanding is the req->rsk_timer should have 
been stopped before the "req->sk = child;" assignment.

or there are cases that req->sk is not NULL in the reqsk_timer_handler()?

BUG: KFENCE: use-after-free read in bpf_sk_storage_get_tracing+0x2e/0x1b0

Use-after-free read at 0x00000000a891fb3a (in kfence-#1):
bpf_sk_storage_get_tracing+0x2e/0x1b0
bpf_prog_5ea3e95db6da0438_tcp_retransmit_synack+0x1d20/0x1dda
bpf_trace_run2+0x4c/0xc0
tcp_rtx_synack+0xf9/0x100
reqsk_timer_handler+0xda/0x3d0
run_timer_softirq+0x292/0x8a0
irq_exit_rcu+0xf5/0x320
sysvec_apic_timer_interrupt+0x6d/0x80
asm_sysvec_apic_timer_interrupt+0x16/0x20
intel_idle_irq+0x5a/0xa0
cpuidle_enter_state+0x94/0x273
cpu_startup_entry+0x15e/0x260
start_secondary+0x8a/0x90
secondary_startup_64_no_verify+0xfa/0xfb

kfence-#1: 0x00000000a72cc7b6-0x00000000d97616d9, size=2376, cache=TCPv6

allocated by task 0 on cpu 9 at 260507.901592s:
sk_prot_alloc+0x35/0x140
sk_clone_lock+0x1f/0x3f0
inet_csk_clone_lock+0x15/0x160
tcp_create_openreq_child+0x1f/0x410
tcp_v6_syn_recv_sock+0x1da/0x700
tcp_check_req+0x1fb/0x510
tcp_v6_rcv+0x98b/0x1420
ipv6_list_rcv+0x2258/0x26e0
napi_complete_done+0x5b1/0x2990
mlx5e_napi_poll+0x2ae/0x8d0
net_rx_action+0x13e/0x590
irq_exit_rcu+0xf5/0x320
common_interrupt+0x80/0x90
asm_common_interrupt+0x22/0x40
cpuidle_enter_state+0xfb/0x273
cpu_startup_entry+0x15e/0x260
start_secondary+0x8a/0x90
secondary_startup_64_no_verify+0xfa/0xfb

freed by task 0 on cpu 9 at 260507.927527s:
rcu_core_si+0x4ff/0xf10
irq_exit_rcu+0xf5/0x320
sysvec_apic_timer_interrupt+0x6d/0x80
asm_sysvec_apic_timer_interrupt+0x16/0x20
cpuidle_enter_state+0xfb/0x273
cpu_startup_entry+0x15e/0x260
start_secondary+0x8a/0x90
secondary_startup_64_no_verify+0xfa/0xfb

Thanks,
Martin

Re: [Question]: A non NULL req->sk in tcp_rtx_synack. Not a fastopen connection.

[Kuniyuki Iwashima]

From: Martin KaFai Lau <martin.lau@linux.dev>
Date: Thu, 3 Oct 2024 18:14:09 -0700
> Hi,
> 
> We are seeing a use-after-free from a bpf prog attached to 
> trace_tcp_retransmit_synack. The program passes the req->sk to the 
> bpf_sk_storage_get_tracing kernel helper which does check for null before using it.
> 
> fastopen is not used.
> 
> We got a kfence report on use-after-free (pasted at the end). It is running with 
> an older 6.4 kernel and we hardly hit this in production.
> 
>  From the upstream code, del_timer_sync() should have been done by 
> inet_csk_reqsk_queue_drop() before "req->sk = child;" is assigned in 
> inet_csk_reqsk_queue_add(). My understanding is the req->rsk_timer should have 
> been stopped before the "req->sk = child;" assignment.

There seems to be a small race window in reqsk_queue_unlink().

expire_timers() first calls detach_timer(, true), which marks the timer
as not pending, and then calls reqsk_timer_handler().

If reqsk_queue_unlink() calls timer_pending() just before expire_timers()
calls reqsk_timer_handler(), reqsk_queue_unlink() could miss
del_timer_sync() ?

---8<---
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 2c5632d4fddb..4ba47ee6c9da 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -1045,7 +1045,7 @@ static bool reqsk_queue_unlink(struct request_sock *req)
 		found = __sk_nulls_del_node_init_rcu(sk);
 		spin_unlock(lock);
 	}
-	if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
+	if (del_timer_sync(&req->rsk_timer))
 		reqsk_put(req);
 	return found;
 }
---8<---


> 
> or there are cases that req->sk is not NULL in the reqsk_timer_handler()?
> 
> BUG: KFENCE: use-after-free read in bpf_sk_storage_get_tracing+0x2e/0x1b0
> 
> Use-after-free read at 0x00000000a891fb3a (in kfence-#1):
> bpf_sk_storage_get_tracing+0x2e/0x1b0
> bpf_prog_5ea3e95db6da0438_tcp_retransmit_synack+0x1d20/0x1dda
> bpf_trace_run2+0x4c/0xc0
> tcp_rtx_synack+0xf9/0x100
> reqsk_timer_handler+0xda/0x3d0
> run_timer_softirq+0x292/0x8a0
> irq_exit_rcu+0xf5/0x320
> sysvec_apic_timer_interrupt+0x6d/0x80
> asm_sysvec_apic_timer_interrupt+0x16/0x20
> intel_idle_irq+0x5a/0xa0
> cpuidle_enter_state+0x94/0x273
> cpu_startup_entry+0x15e/0x260
> start_secondary+0x8a/0x90
> secondary_startup_64_no_verify+0xfa/0xfb
> 
> kfence-#1: 0x00000000a72cc7b6-0x00000000d97616d9, size=2376, cache=TCPv6
> 
> allocated by task 0 on cpu 9 at 260507.901592s:
> sk_prot_alloc+0x35/0x140
> sk_clone_lock+0x1f/0x3f0
> inet_csk_clone_lock+0x15/0x160
> tcp_create_openreq_child+0x1f/0x410
> tcp_v6_syn_recv_sock+0x1da/0x700
> tcp_check_req+0x1fb/0x510
> tcp_v6_rcv+0x98b/0x1420
> ipv6_list_rcv+0x2258/0x26e0
> napi_complete_done+0x5b1/0x2990
> mlx5e_napi_poll+0x2ae/0x8d0
> net_rx_action+0x13e/0x590
> irq_exit_rcu+0xf5/0x320
> common_interrupt+0x80/0x90
> asm_common_interrupt+0x22/0x40
> cpuidle_enter_state+0xfb/0x273
> cpu_startup_entry+0x15e/0x260
> start_secondary+0x8a/0x90
> secondary_startup_64_no_verify+0xfa/0xfb
> 
> freed by task 0 on cpu 9 at 260507.927527s:
> rcu_core_si+0x4ff/0xf10
> irq_exit_rcu+0xf5/0x320
> sysvec_apic_timer_interrupt+0x6d/0x80
> asm_sysvec_apic_timer_interrupt+0x16/0x20
> cpuidle_enter_state+0xfb/0x273
> cpu_startup_entry+0x15e/0x260
> start_secondary+0x8a/0x90
> secondary_startup_64_no_verify+0xfa/0xfb
> 
> Thanks,
> Martin

Re: [Question]: A non NULL req->sk in tcp_rtx_synack. Not a fastopen connection.

[Martin KaFai Lau]

On 10/3/24 7:02 PM, Kuniyuki Iwashima wrote:
> From: Martin KaFai Lau <martin.lau@linux.dev>
> Date: Thu, 3 Oct 2024 18:14:09 -0700
>> Hi,
>>
>> We are seeing a use-after-free from a bpf prog attached to
>> trace_tcp_retransmit_synack. The program passes the req->sk to the
>> bpf_sk_storage_get_tracing kernel helper which does check for null before using it.
>>
>> fastopen is not used.
>>
>> We got a kfence report on use-after-free (pasted at the end). It is running with
>> an older 6.4 kernel and we hardly hit this in production.
>>
>>   From the upstream code, del_timer_sync() should have been done by
>> inet_csk_reqsk_queue_drop() before "req->sk = child;" is assigned in
>> inet_csk_reqsk_queue_add(). My understanding is the req->rsk_timer should have
>> been stopped before the "req->sk = child;" assignment.
> 
> There seems to be a small race window in reqsk_queue_unlink().
> 
> expire_timers() first calls detach_timer(, true), which marks the timer
> as not pending, and then calls reqsk_timer_handler().
> 
> If reqsk_queue_unlink() calls timer_pending() just before expire_timers()
> calls reqsk_timer_handler(), reqsk_queue_unlink() could miss
> del_timer_sync() ?

This seems to explain it. :)

Does it mean there is a chance that the reqsk_timer_handler() may rearm the 
timer again and I guess only a few more synack will be sent in this case and 
should be no harm?

> 
> ---8<---
> diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> index 2c5632d4fddb..4ba47ee6c9da 100644
> --- a/net/ipv4/inet_connection_sock.c
> +++ b/net/ipv4/inet_connection_sock.c
> @@ -1045,7 +1045,7 @@ static bool reqsk_queue_unlink(struct request_sock *req)
>   		found = __sk_nulls_del_node_init_rcu(sk);
>   		spin_unlock(lock);
>   	}
> -	if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
> +	if (del_timer_sync(&req->rsk_timer))

It seems the reqsk_timer_handler() will also call reqsk_queue_unlink() through 
inet_csk_reqsk_queue_drop_and_put(). Not sure if the reqsk_timer_handler() can 
del_timer_sync() itself.

>   		reqsk_put(req);
>   	return found;
>   }
> ---8<---
> 
> 
>>
>> or there are cases that req->sk is not NULL in the reqsk_timer_handler()?
>>
>> BUG: KFENCE: use-after-free read in bpf_sk_storage_get_tracing+0x2e/0x1b0
>>
>> Use-after-free read at 0x00000000a891fb3a (in kfence-#1):
>> bpf_sk_storage_get_tracing+0x2e/0x1b0
>> bpf_prog_5ea3e95db6da0438_tcp_retransmit_synack+0x1d20/0x1dda
>> bpf_trace_run2+0x4c/0xc0
>> tcp_rtx_synack+0xf9/0x100
>> reqsk_timer_handler+0xda/0x3d0
>> run_timer_softirq+0x292/0x8a0
>> irq_exit_rcu+0xf5/0x320
>> sysvec_apic_timer_interrupt+0x6d/0x80
>> asm_sysvec_apic_timer_interrupt+0x16/0x20
>> intel_idle_irq+0x5a/0xa0
>> cpuidle_enter_state+0x94/0x273
>> cpu_startup_entry+0x15e/0x260
>> start_secondary+0x8a/0x90
>> secondary_startup_64_no_verify+0xfa/0xfb
>>
>> kfence-#1: 0x00000000a72cc7b6-0x00000000d97616d9, size=2376, cache=TCPv6
>>
>> allocated by task 0 on cpu 9 at 260507.901592s:
>> sk_prot_alloc+0x35/0x140
>> sk_clone_lock+0x1f/0x3f0
>> inet_csk_clone_lock+0x15/0x160
>> tcp_create_openreq_child+0x1f/0x410
>> tcp_v6_syn_recv_sock+0x1da/0x700
>> tcp_check_req+0x1fb/0x510
>> tcp_v6_rcv+0x98b/0x1420
>> ipv6_list_rcv+0x2258/0x26e0
>> napi_complete_done+0x5b1/0x2990
>> mlx5e_napi_poll+0x2ae/0x8d0
>> net_rx_action+0x13e/0x590
>> irq_exit_rcu+0xf5/0x320
>> common_interrupt+0x80/0x90
>> asm_common_interrupt+0x22/0x40
>> cpuidle_enter_state+0xfb/0x273
>> cpu_startup_entry+0x15e/0x260
>> start_secondary+0x8a/0x90
>> secondary_startup_64_no_verify+0xfa/0xfb
>>
>> freed by task 0 on cpu 9 at 260507.927527s:
>> rcu_core_si+0x4ff/0xf10
>> irq_exit_rcu+0xf5/0x320
>> sysvec_apic_timer_interrupt+0x6d/0x80
>> asm_sysvec_apic_timer_interrupt+0x16/0x20
>> cpuidle_enter_state+0xfb/0x273
>> cpu_startup_entry+0x15e/0x260
>> start_secondary+0x8a/0x90
>> secondary_startup_64_no_verify+0xfa/0xfb
>>
>> Thanks,
>> Martin

Re: [Question]: A non NULL req->sk in tcp_rtx_synack. Not a fastopen connection.

[Kuniyuki Iwashima]

From: Martin KaFai Lau <martin.lau@linux.dev>
Date: Thu, 3 Oct 2024 21:00:20 -0700
> On 10/3/24 7:02 PM, Kuniyuki Iwashima wrote:
> > From: Martin KaFai Lau <martin.lau@linux.dev>
> > Date: Thu, 3 Oct 2024 18:14:09 -0700
> >> Hi,
> >>
> >> We are seeing a use-after-free from a bpf prog attached to
> >> trace_tcp_retransmit_synack. The program passes the req->sk to the
> >> bpf_sk_storage_get_tracing kernel helper which does check for null before using it.
> >>
> >> fastopen is not used.
> >>
> >> We got a kfence report on use-after-free (pasted at the end). It is running with
> >> an older 6.4 kernel and we hardly hit this in production.
> >>
> >>   From the upstream code, del_timer_sync() should have been done by
> >> inet_csk_reqsk_queue_drop() before "req->sk = child;" is assigned in
> >> inet_csk_reqsk_queue_add(). My understanding is the req->rsk_timer should have
> >> been stopped before the "req->sk = child;" assignment.
> > 
> > There seems to be a small race window in reqsk_queue_unlink().
> > 
> > expire_timers() first calls detach_timer(, true), which marks the timer
> > as not pending, and then calls reqsk_timer_handler().
> > 
> > If reqsk_queue_unlink() calls timer_pending() just before expire_timers()
> > calls reqsk_timer_handler(), reqsk_queue_unlink() could miss
> > del_timer_sync() ?
> 
> This seems to explain it. :)
> 
> Does it mean there is a chance that the reqsk_timer_handler() may rearm the 
> timer again and I guess only a few more synack will be sent in this case and 
> should be no harm?

Ah, it seems possible.  I was wondering how the timer can be delayed
until sk is freed.  In such a case, the timer will just let the peer
generate some challenge ACKs.


> 
> > 
> > ---8<---
> > diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> > index 2c5632d4fddb..4ba47ee6c9da 100644
> > --- a/net/ipv4/inet_connection_sock.c
> > +++ b/net/ipv4/inet_connection_sock.c
> > @@ -1045,7 +1045,7 @@ static bool reqsk_queue_unlink(struct request_sock *req)
> >   		found = __sk_nulls_del_node_init_rcu(sk);
> >   		spin_unlock(lock);
> >   	}
> > -	if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
> > +	if (del_timer_sync(&req->rsk_timer))
> 
> It seems the reqsk_timer_handler() will also call reqsk_queue_unlink() through 
> inet_csk_reqsk_queue_drop_and_put(). Not sure if the reqsk_timer_handler() can 
> del_timer_sync() itself.

Exactly, it seems illegal to call it from the timer.
Then, we need a variant of inet_csk_reqsk_queue_drop() to see if
the caller is tiemr or not. (compile-test only)

---8<---
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 2c5632d4fddb..2623964d8817 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -1045,21 +1045,31 @@ static bool reqsk_queue_unlink(struct request_sock *req)
 		found = __sk_nulls_del_node_init_rcu(sk);
 		spin_unlock(lock);
 	}
-	if (timer_pending(&req->rsk_timer) && del_timer_sync(&req->rsk_timer))
-		reqsk_put(req);
+
 	return found;
 }
 
-bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
+static bool __inet_csk_reqsk_queue_drop(struct sock *sk,
+					struct request_sock *req,
+					bool from_timer)
 {
 	bool unlinked = reqsk_queue_unlink(req);
 
+	if (!from_timer && del_timer_sync(&req->rsk_timer))
+		reqsk_put(req);
+
 	if (unlinked) {
 		reqsk_queue_removed(&inet_csk(sk)->icsk_accept_queue, req);
 		reqsk_put(req);
 	}
+
 	return unlinked;
 }
+
+bool inet_csk_reqsk_queue_drop(struct sock *sk, struct request_sock *req)
+{
+	return __inet_csk_reqsk_queue_drop(sk, req, false);
+}
 EXPORT_SYMBOL(inet_csk_reqsk_queue_drop);
 
 void inet_csk_reqsk_queue_drop_and_put(struct sock *sk, struct request_sock *req)
@@ -1152,7 +1162,7 @@ static void reqsk_timer_handler(struct timer_list *t)
 
 		if (!inet_ehash_insert(req_to_sk(nreq), req_to_sk(oreq), NULL)) {
 			/* delete timer */
-			inet_csk_reqsk_queue_drop(sk_listener, nreq);
+			__inet_csk_reqsk_queue_drop(sk_listener, nreq, true);
 			goto no_ownership;
 		}
 
@@ -1178,7 +1188,8 @@ static void reqsk_timer_handler(struct timer_list *t)
 	}
 
 drop:
-	inet_csk_reqsk_queue_drop_and_put(oreq->rsk_listener, oreq);
+	__inet_csk_reqsk_queue_drop(sk_listener, nreq, true);
+	reqsk_put(req);
 }
 
 static bool reqsk_queue_hash_req(struct request_sock *req,
---8<---