Re: kernel BUG at include/net/tcp.h:739

["Ilpo Järvinen" <ilpo.jarvinen@helsinki.fi>]

On Wed, 2 May 2007, Michal Piotrowski wrote:

> Please take a look at this bug
> 
> [15236.638092] kernel BUG at /mnt/md0/devel/linux-git/include/net/tcp.h:739!
> [15236.644860] invalid opcode: 0000 [#1]
> [15236.648514] PREEMPT SMP 
> [15236.651075] Modules linked in: ipt_MASQUERADE iptable_nat nf_nat autofs4 af_packet nf_conntrack_netbios_ns ipt_REJECT nf_conntrack_ipv4 xt_state nf_conntrack nfnetlink iptable_filter ip_tables ip6t_REJECT xt_tcpudp ip6table_filter ip6_tables x_tables ipv6 binfmt_misc thermal processor fan container nvram snd_intel8x0 snd_ac97_codec ac97_bus snd_seq_dummy snd_seq_oss snd_seq_midi_event snd_seq snd_seq_device snd_pcm_oss snd_mixer_oss snd_pcm evdev snd_timer snd intel_agp i2c_i801 soundcore agpgart snd_page_alloc ide_cd cdrom rtc unix
> [15236.698898] CPU:    0
> [15236.698899] EIP:    0060:[<c02f798b>]    Not tainted VLI
> [15236.698900] EFLAGS: 00010206   (2.6.21-gdc87c398 #169)
> [15236.711580] EIP is at tcp_ack+0xc54/0x16a0
> [15236.715664] eax: 00000017   ebx: 00000000   ecx: 00000003   edx: 0000010e
> [15236.722433] esi: d5bc1254   edi: 0000010e   ebp: c0462e18   esp: c0462da8
> [15236.729202] ds: 007b   es: 007b   fs: 00d8  gs: 0000  ss: 0068
> [15236.735019] Process swapper (pid: 0, ti=c0462000 task=c03f14e0 task.ti=c0427000)
> [15236.742219] Stack: 00000000 00000000 00000000 00000198 00000100 00000000 00000000 00000018 
> [15236.750698]        3b4d5775 0ffef1c0 3b4d79ad 00000001 00000018 00000006 3b4d79ad 003f14e0 
> [15236.759178]        00000006 0000082a d4b9cde0 00e4059a 0000000c 00000000 00000000 0130204a 
> [15236.767649] Call Trace:
> [15236.770286]  [<c0105039>] show_trace_log_lvl+0x1a/0x2f
> [15236.775438]  [<c01050eb>] show_stack_log_lvl+0x9d/0xa5
> [15236.780590]  [<c01052e0>] show_registers+0x1ed/0x32c
> [15236.785569]  [<c0105537>] die+0x118/0x22f
> [15236.789588]  [<c01056c7>] do_trap+0x79/0x91
> [15236.793781]  [<c0105f57>] do_invalid_op+0x97/0xa1
> [15236.798501]  [<c031f38c>] error_code+0x7c/0x84
> [15236.802960]  [<c02fafb7>] tcp_rcv_established+0x568/0x645
> [15236.808354]  [<c03000bd>] tcp_v4_do_rcv+0x2b/0x32c
> [15236.813144]  [<c030243a>] tcp_v4_rcv+0x7f9/0x86b
> [15236.817777]  [<c02e9dd3>] ip_local_deliver+0x170/0x235
> [15236.822928]  [<c02e9c2a>] ip_rcv+0x4f3/0x52c
> [15236.827199]  [<c02d4909>] netif_receive_skb+0x1b9/0x252
> [15236.832437]  [<c02635c2>] skge_poll+0x47a/0x545
> [15236.836967]  [<c02d68df>] net_rx_action+0x9f/0x192
> [15236.841772]  [<c0126408>] __do_softirq+0x6d/0xea
> [15236.846407]  [<c01069b5>] do_softirq+0x64/0xd1
> [15236.850867]  =======================
> [15236.854437] Code: 69 42 c0 f7 d0 64 8b 15 04 00 00 00 8b 04 90 ff 80 a4 00 00 00 8b 9e 70 05 00 00 89 d8 03 86 74 05 00 00 3b 86 8c 04 00 00 76 04 <0f> 0b eb fe 89 86 90 04 00 00 8a 86 88 03 00 00 84 c0 75 3c 83 
> [15236.874343] EIP: [<c02f798b>] tcp_ack+0xc54/0x16a0 SS:ESP 0068:c0462da8
> 
> l *0xc02f798b
> 0xc02f798b is in tcp_ack (/mnt/md0/devel/linux-git/include/net/tcp.h:739).
> 734                                 (tp->snd_cwnd >> 2)));
> 735     }
> 736
> 737     static inline void tcp_sync_left_out(struct tcp_sock *tp)
> 738     {
> 739             BUG_ON(tp->sacked_out + tp->lost_out > tp->packets_out);
> 740             tp->left_out = tp->sacked_out + tp->lost_out;
> 741     }
> 742
> 743     extern void tcp_enter_cwr(struct sock *sk, const int set_ssthresh);
> 
> Caused by commit 34588b4c046c34773e5a1a962da7b78b05c4d1bd 

I think I found the reason:

tcp_clean_rtx_queue without SACK it does not decrement sacked_out but
tcp_reset_reno_sack/remove_reno_sack is being called later in
tcp_fastretrans_alert. Before that, tcp_sync_left_out is being called, 
at least by tcp_fastretrans_alert itself, which sees sacked_out that 
includes also segments that are no longer in window (this also explains 
why the original code did not do the reduction in the non-SACK case). Also 
tcp_process_frto calls tcp_sync_left_out, so it would also lead to the 
same problem.

Here is a change that ignores this trap without SACK. However, it would be 
useful to trap this without SACK too as S+L skb causes potentially a 
negative packets in flight (= large one) disturbing cwnd compares. 


[PATCH] [TCP]: Use S+L catcher only with SACK for now

TCP has a transitional state when SACK is not in use during
which this invariant is temporarily broken. Without SACK,
tcp_clean_rtx_queue does not decrement sacked_out. Therefore
calls to tcp_sync_left_out before sacked_out is again
corrected by tcp_fastretrans_alert can trigger this trap as
sacked_out still has couple of segments that are already out
of window.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
---
 include/net/tcp.h |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

diff --git a/include/net/tcp.h b/include/net/tcp.h
index ef8f9d4..e22b4f0 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -736,7 +736,8 @@ static inline __u32 tcp_current_ssthresh
 
 static inline void tcp_sync_left_out(struct tcp_sock *tp)
 {
-	BUG_ON(tp->sacked_out + tp->lost_out > tp->packets_out);
+	BUG_ON(tp->rx_opt.sack_ok &&
+	       (tp->sacked_out + tp->lost_out > tp->packets_out));
 	tp->left_out = tp->sacked_out + tp->lost_out;
 }
 
-- 
1.4.2