[PATCH net-next] tcp: Revert "tcp: tcp_probe: use spin_lock_bh()"

[PATCH net-next] tcp: Revert "tcp: tcp_probe: use spin_lock_bh()"

[Eric Dumazet]

This reverts commit e70ac171658679ecf6bea4bbd9e9325cd6079d2b.

jtcp_rcv_established() is in fact called with hard irq being disabled.

Initial bug report from Ricardo Nabinger Sanchez [1] still needs
to be investigated, but does not look like a TCP bug.

[1] https://www.spinics.net/lists/netdev/msg420960.html

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: kernel test robot <xiaolong.ye@intel.com>
Cc: Ricardo Nabinger Sanchez <rnsanchez@gmail.com>
---
 net/ipv4/tcp_probe.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/tcp_probe.c b/net/ipv4/tcp_probe.c
index 3d063eb3784828b142874c92fd2db026bea0f3b3..f6c50af24a64737672f7ede2ff41158bfed5f1b4 100644
--- a/net/ipv4/tcp_probe.c
+++ b/net/ipv4/tcp_probe.c
@@ -117,7 +117,7 @@ static void jtcp_rcv_established(struct sock *sk, struct sk_buff *skb,
 	     (fwmark > 0 && skb->mark == fwmark)) &&
 	    (full || tp->snd_cwnd != tcp_probe.lastcwnd)) {
 
-		spin_lock_bh(&tcp_probe.lock);
+		spin_lock(&tcp_probe.lock);
 		/* If log fills, just silently drop */
 		if (tcp_probe_avail() > 1) {
 			struct tcp_log *p = tcp_probe.log + tcp_probe.head;
@@ -157,7 +157,7 @@ static void jtcp_rcv_established(struct sock *sk, struct sk_buff *skb,
 			tcp_probe.head = (tcp_probe.head + 1) & (bufsize - 1);
 		}
 		tcp_probe.lastcwnd = tp->snd_cwnd;
-		spin_unlock_bh(&tcp_probe.lock);
+		spin_unlock(&tcp_probe.lock);
 
 		wake_up(&tcp_probe.wait);
 	}
-- 
2.11.0.483.g087da7b7c-goog

Re: [PATCH net-next] tcp: Revert "tcp: tcp_probe: use spin_lock_bh()"

[David Miller]

From: Eric Dumazet <edumazet@google.com>
Date: Tue, 21 Feb 2017 06:21:47 -0800

> This reverts commit e70ac171658679ecf6bea4bbd9e9325cd6079d2b.
> 
> jtcp_rcv_established() is in fact called with hard irq being disabled.
> 
> Initial bug report from Ricardo Nabinger Sanchez [1] still needs
> to be investigated, but does not look like a TCP bug.
> 
> [1] https://www.spinics.net/lists/netdev/msg420960.html
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Reported-by: kernel test robot <xiaolong.ye@intel.com>
> Cc: Ricardo Nabinger Sanchez <rnsanchez@gmail.com>

Applied.

But your analysis was quite sound, we can now invoke TCP input path
from user context, therefore this jprobe needs to use BH locking.  What
is the problem.

Re: [PATCH net-next] tcp: Revert "tcp: tcp_probe: use spin_lock_bh()"

[Eric Dumazet]

On Tue, Feb 21, 2017 at 10:26 AM, David Miller <davem@davemloft.net> wrote:
> From: Eric Dumazet <edumazet@google.com>
> Date: Tue, 21 Feb 2017 06:21:47 -0800
>
>> This reverts commit e70ac171658679ecf6bea4bbd9e9325cd6079d2b.
>>
>> jtcp_rcv_established() is in fact called with hard irq being disabled.
>>
>> Initial bug report from Ricardo Nabinger Sanchez [1] still needs
>> to be investigated, but does not look like a TCP bug.
>>
>> [1] https://www.spinics.net/lists/netdev/msg420960.html
>>
>> Signed-off-by: Eric Dumazet <edumazet@google.com>
>> Reported-by: kernel test robot <xiaolong.ye@intel.com>
>> Cc: Ricardo Nabinger Sanchez <rnsanchez@gmail.com>
>
> Applied.
>
> But your analysis was quite sound, we can now invoke TCP input path
> from user context, therefore this jprobe needs to use BH locking.  What
> is the problem.


If hard irq are masked, then using spin_lock() should work without a
lockdep splat.

The problem here is some false positive, as if one
trace_hardirqs_off() was missing.

Peter originally fixed the issue back in this commit :

commit 58dfe883d3bc3b4c08c53a7f39e2ca3ec84f089e
Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date:   Thu Oct 11 22:25:25 2007 +0200

So really I have no idea what is happening, because the
trace_hardirqs_off() is still in x86 setjmp_pre_handler()

Maybe some FTRACE interaction ?