I've also seen the issue when running plain nginx 1.9.11 sort like builting a new nginx, only the issue is that upstream closes cnxs, not nginx.
Since I've discovered that our TCs had the Connector' keepAliveTimeout way low (10 msec), mistakenly thought the units were sec and not as actually msec.
Since increasing this a 1000 fold everything seems much better now :)

<Connector port="8081" protocol="HTTP/1.1"
maxConnections="8192" maxThreads="200"
enableLookups="false" tcpNoDelay="true"
connectionTimeout="10000" keepAliveTimeout="900000"
socket.soKeepAlive="true" maxKeepAliveRequests="-1" />


Thanks, will look into your suggestions next time I need to debug TC, right now they are running at high throttle ;)

BTW I'm using Apache Tomcat/7.0.27

tcpdump snippet between nginx(10.45.69.14) and TC(10.45.69.25), where TC RESETs cnx instead of replying:

01:15:49.744283 IP 10.45.69.14.58702 > 10.45.69.25.8081: Flags [P.], seq 31688:39753, ack 191, win 219, options [nop,nop,TS val 38708511 ecr 32578863], length 8065
/PUT /<redacted URL> HTTP/1.1^M
Host: mosgenericbackend^M
Content-Length: 7811^M
Content-Type: application/x-www-form-urlencoded^M
User-Agent: Java/1.8.0_66^M
Accept: text/html, image/gif, image/jpeg, *; q=.2, */*; q=.2^M
^M
<redacted args>
01:15:49.744311 IP 10.45.69.25.8081 > 10.45.69.14.58702: Flags [R], seq 1376851868, win 0, length 0
01:15:49.744467 IP 10.45.69.14.58702 > 10.45.69.25.8081: Flags [.], ack 192, win 219, options [nop,nop,TS val 38708512 ecr 32578874], length 0
01:15:49.744480 IP 10.45.69.25.8081 > 10.45.69.14.58702: Flags [R], seq 1376851869, win 0, length 0