Hello!

On Thu, Oct 16, 2014 at 10:17:15AM +0200, Jiri Horky wrote:

> Hi list,
>
> we are seeing sporadic nginx errors "upstream prematurely closed
> connection while reading response header from upstream" with nginx/1.6.2
> which seems to be some kind of race condition.
> For debugging purposes we only setup 1 upstream server on a public IP
> address of the same server as nginx, there is no keepalive configured
> between nginx and the upstream server. The upstream HTTP server is
> written in a way that it forcibly closes the connection when the
> response status code is 303. This may be part of the problem as well.

[...]

> Now, we tracked down, that this only happens when FIN packet from
> upstream server reaches nginx sooner than it's finished with parsing the
> response (headers) and thus sooner than nginx closes the connection
> itself. For example this packet order will trigger the problem:
> No. Time Source SrcPrt Destination Protocol
> Length Info
> 25571 10.297569 1.1.1.1 35481 1.1.1.1 TCP 76 35481 > 8888 [SYN] Seq=0 Win=3072 Len=0 MSS=16396 SACK_PERM=1 TSval=1902164528 TSecr=0 WS=8192
> 25572 10.297580 1.1.1.1 8888 1.1.1.1 TCP 76 8888 > 35481 [SYN, ACK] Seq=0 Ack=1 Win=3072 Len=0 MSS=16396 SACK_PERM=1 TSval=1902164528 TSecr=1902164528 WS=8192
> 25573 10.297589 1.1.1.1 35481 1.1.1.1 TCP 68 35481 > 8888 [ACK] Seq=1 Ack=1 Win=8192 Len=0 TSval=1902164528 TSecr=1902164528
> 25574 10.297609 1.1.1.1 35481 1.1.1.1 HTTP 1533 GET / HTTP/1.0
> 25575 10.297617 1.1.1.1 8888 1.1.1.1 TCP 68 8888 > 35481 [ACK] Seq=1 Ack=1466 Win=8192 Len=0 TSval=1902164528 TSecr=1902164528
> 25596 10.323092 1.1.1.1 8888 1.1.1.1 HTTP 480 HTTP/1.1 303 See Other
> 25597 10.323106 1.1.1.1 35481 1.1.1.1 TCP 68 35481 > 8888 [ACK] Seq=1466 Ack=413 Win=8192 Len=0 TSval=1902164554 TSecr=1902164554
> 25598 10.323161 1.1.1.1 8888 1.1.1.1 TCP 68 8888 > 35481 [FIN, ACK] Seq=413 Ack=1466 Win=8192 Len=0 TSval=1902164554 TSecr=1902164554
> 25599 10.323167 1.1.1.1 35481 1.1.1.1 TCP 68 35481 > 8888 [FIN, ACK] Seq=1466 Ack=413 Win=8192 Len=0 TSval=1902164554 TSecr=1902164554
> 25600 10.323180 1.1.1.1 8888 1.1.1.1 TCP 68 8888 > 35481 [ACK] Seq=414 Ack=1467 Win=8192 Len=0 TSval=1902164554 TSecr=1902164554
> 25601 10.323189 1.1.1.1 35481 1.1.1.1 TCP 68 35481 > 8888 [ACK] Seq=1467 Ack=414 Win=8192 Len=0 TSval=1902164554 TSecr=1902164554
>
> Note that the upstream HTTP (port 8888) sends the FIN packet sooner than
> nginx (port 35481 in this case).

Looking into the packet trace I suspect this commit may be
relevant to your case:

http://hg.nginx.org/nginx/rev/9d3a9c45fc43

Please test with nginx 1.7.3+ to see if it helps.

--
Maxim Dounin
http://nginx.org/

_______________________________________________
nginx mailing list
nginx@nginx.org
http://mailman.nginx.org/mailman/listinfo/nginx