nginx -s reload terminates connections

[rlx01]

Hi,

Using the official 1.12.0 package on debian 9 and the 1.12.0 ports package
on FreeBSD 11-RC1, calling nginx -s reload drops connections.

It's fairly easy to reproduce, I just install the fresh package from
wherever, and serving the default index.html, I run wrk, and in the middle
of the run I call 'nginx -s reload' without actually changing anything:

$ wrk -d 30 -c 100 -t 20 http://nginx-test01.lan/
Running 30s test @ http://nginx-test01.lan/
  20 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    10.06ms    4.10ms  53.02ms   79.07%
    Req/Sec   503.04    161.36   757.00     53.20%
  300801 requests in 30.05s, 243.82MB read
  Socket errors: connect 0, read 100, write 0, timeout 0
Requests/sec:  10008.67
Transfer/sec:      8.11MB

You can see 100 socket read errors in this example.

I tried this with the h2o (2.2.2) server just to see what happens and there
is no error:

$ wrk -d 30 -c 100 -t 20 http://h2o-test01.lan/
Running 30s test @ http://nginx-test01.lan/
  20 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     4.06ms    1.83ms  56.08ms   76.57%
    Req/Sec     1.24k   724.32     6.61k    97.67%
  741569 requests in 30.02s, 599.01MB read
Requests/sec:  24699.67
Transfer/sec:     19.95MB

--

The same issue occurs when I test HTTP/2 with 'h2load'. The h2load test just
stops as soon as I call 'nginx -s reload'.

Is this expected behaviour? If so, is there some configuration directive
that would fix it?

In my real world usage, I have a nginx server running on the edge and it
proxies (HTTP/1.1) to a different nginx server. If I call 'nginx -s reload'
on the internal nginx the external nginx shows "upstream prematurely closed
connection while reading response header from upstream" if it's under heavy
load.

Thanks!

Posted at Nginx Forum: https://forum.nginx.org/read.php?2,275328,275328#msg-275328