php7.0 doesn't report accurate socket state in PHP_RINIT

On the 7.0 node:

```
# apt-get install php7.0-dev php-hypernode
root@2a1186-php7-magweb-vgr /vagrant/php-hypernode/php-7.0/hypernode-20160309.1 
# phpize

Configuring for:
PHP Api Version:         20151012
Zend Module Api No:      20151012
Zend Extension Api No:   320151012
```

On the 5.5 node:

```
# apt-get install php5-dev php5-hypernode

root@cd03d0-php5-magweb-vgr /vagrant/php-hypernode/php-5/hypernode-20160309.1 # phpize
Configuring for:
PHP Api Version:         20121113
Zend Module Api No:      20121212
Zend Extension Api No:   220121212
```

On both nodes:

In hypernode.c, above 

```
        if (tcp_info.tcpi_state == TCP_CLOSE_WAIT) {
```

add

```
        char msg[100];
        sprintf(msg, "tcp_info.tcp_state is %d, TCP_CLOSE_WAIT_STATE is %d\n", tcp_info.tcpi_state, TCP_CLOSE_WAIT);
        php_syslog(LOG_NOTICE, msg);
```

For reference, the TCP state mapping looks like:

```
enum {
  TCPF_ESTABLISHED = (1 << 1),
  TCPF_SYN_SENT  = (1 << 2),
  TCPF_SYN_RECV  = (1 << 3),
  TCPF_FIN_WAIT1 = (1 << 4),
  TCPF_FIN_WAIT2 = (1 << 5),
  TCPF_TIME_WAIT = (1 << 6),
  TCPF_CLOSE     = (1 << 7),
  TCPF_CLOSE_WAIT = (1 << 8),
  TCPF_LAST_ACK  = (1 << 9),
  TCPF_LISTEN    = (1 << 10),
  TCPF_CLOSING   = (1 << 11)
};
```

https://github.com/mirror/reactos/blob/master/reactos/drivers/network/tcpip/include/tcpdef.h#L77

Install build tools

```
# apt-get install build-essential autoconf
```

Run autoconf

```
# autoconf
```

Build the things

```
# ./configure && make && make install
```

On the 7.0 node:

```
in /etc/php/7.0/fpm/pool.d/www.conf
 pm.max_children = 3

# cp modules/hypernode.so /usr/lib/php/20151012/hypernode.so 
# service php7.0-fpm restart
```

On the 5.5 node:

```
in /etc/php5/fpm/pool.d/www.conf
 pm.max_children = 3
# cp modules/hypernode.so /usr/lib/php5/20121212/hypernode.so
# service php5-fpm restart
```

On both nodes:

in /data/web/public/index.php:

```
<?php
    sleep(2000000);
```

disable varnish

```
# echo -e 'vcl 4.0;\nbackend default {\n  .host = "127.0.0.1";\n  .port = "8080";\n}\nsub vcl_recv {\n  return(pass);\n}' > /data/web/disabled_caching.vcl && varnishadm vcl.load nocache /data/web/disabled_caching.vcl && varnishadm vcl.use nocache
```

tail syslog

```
# tail -f /var/log/syslog 
```

on the host spam some requests:

```
for i in 1 2 3 4 5 6 7 8 9; do curl -k --max-time 5 https://php5.hypernode.local & done
for i in 1 2 3 4 5 6 7 8 9; do curl -k --max-time 5 https://php7.hypernode.local & done
```
## php7:

```
20614 BUSY 3.2169 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
20615 BUSY 3.2192 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
21687 BUSY 2.3646 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
```
## important

```
## tcp_state 10 is TCP_LISTEN
Apr  4 12:24:36 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 10, TCP_CLOSE_WAIT_STATE is 8
```
## php5:

```
21746 BUSY 42.2295 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
21747 BUSY 42.2294 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
21855 BUSY 41.7218 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
```
## important

```
## tcp_state 1 is TCP_ESTABLISHED
Apr  4 12:24:04 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 1, TCP_CLOSE_WAIT_STATE is 8
```

Stop nginx so the client to FPM is gone:

```
# service nginx stop
```

```
20614 GONE 99.0227 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
20615 GONE 99.0255 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
21687 GONE 98.1713 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
```

```
21746 GONE 127.7270 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
21747 GONE 127.7275 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
21855 GONE 127.2204 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
```

Now, if we kill one of the FPM children we expect the tcp_info.tcp_state of the newly spawned child to be WAIT_CLOSE because the requests from nginx left queued in the kernel will already be disconnected.
## important

On the PHP5 node:

```
# kill 21746
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 8, TCP_CLOSE_WAIT_STATE is 8
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: Terminated request GET php5.hypernode.local/index.php because client at 10.0.3.1 is already gone
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 8, TCP_CLOSE_WAIT_STATE is 8
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: Terminated request GET php5.hypernode.local/index.php because client at 10.0.3.1 is already gone
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 8, TCP_CLOSE_WAIT_STATE is 8
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: Terminated request GET php5.hypernode.local/index.php because client at 10.0.3.1 is already gone
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 8, TCP_CLOSE_WAIT_STATE is 8
Apr  4 12:27:34 xxxxx-dummytag-vagrant ool www: Terminated request GET php5.hypernode.local/index.php because client at 10.0.3.1 is already gone
etc.. until the queue is empty

21747 GONE 226.7354 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
21855 GONE 226.2278 10.0.3.1        GET  php5.hypernode.local/   (curl/7.47.0)
27027 IDLE                                ()
```
## important

On the PHP7 node:

```
# kill 21687
## tcp_state 10 is TCP_LISTEN
Apr  4 12:28:34 xxxxx-dummytag-vagrant ool www: tcp_info.tcp_state is 10, TCP_CLOSE_WAIT_STATE is 8

20614 GONE 251.6493 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
20615 GONE 251.6518 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0)
27620 GONE 0.4032 10.0.3.1        GET  php7.hypernode.local/   (curl/7.47.0) <- this one should have been dropped immediately, instead the state is TCP_LISTEN in the RINIT context.
```

Now if immediately if we run netsat on the proc we do get a CLOSE_WAIT:

```
# netstat -nap | grep 27620
tcp        9      0 127.0.0.1:9000          127.0.0.1:57038         CLOSE_WAIT  28844/php-fpm: pool
```

_So it looks like the inner workings of fpm has changed and it deals with the request->fd differently_

I tried compiling fpm with the socket code from 5.5 but that made no difference

```
cd ~/code/projects/php; git checkout hypernode-7.0; git checkout -b get-listening-socket-again
cd sapi/fpm/
wget https://raw.githubusercontent.com/ByteInternet/php/master-5.5/sapi/fpm/fpm/fpm_sockets.c
wget https://raw.githubusercontent.com/ByteInternet/php/master-5.5/sapi/fpm/fpm/fpm_sockets.h
dpkg-source --commit
gbp buildpackage --git-pbuilder --git-dist=precise --git-arch=amd64 --git-debian-branch=get-listening-socket-again -us -sa
```

still:

```
Apr  4 13:17:09 2a1186-php7-magweb-vgr ool www: tcp_info.tcp_state is 10, TCP_CLOSE_WAIT_STATE is 8
```

I also tried using a [UNIX socket instead of a network socket](https://easyengine.io/tutorials/php/fpm-sysctl-tweaking/), but then even in php55 the state was always TCP_LISTEN


Provide feedback

Saved searches

Use saved searches to filter your results more quickly

php7.0 doesn't report accurate socket state in PHP_RINIT #2

php7:

important

php5:

important

important

important

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

php7.0 doesn't report accurate socket state in PHP_RINIT #2

Description

php7:

important

php5:

important

important

important

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions