So I had a FreeBSD jail today running a geriatric version of MySQL that was long past time to update. But sometimes when you perform big updates you discover problems along the way, and it takes much longer than usual. This particular jail can’t tolerate a lot of downtime so I thought it would be nice to perform a “practice” upgrade to flush out all of the problems and have solutions ready when I’m ready to perform the live upgrade. My idea was to clone the jail and perform a practice upgrade on that. It was easy.

First, clone the file system. That’s easy because I keep all of my jails on their own ZFS file systems and maintain daily snapshots:

zfs clone zroot/jails/vhost2@daily.0 zroot/jails/vhost2upgrade

And of course that happens pretty much instantly.

Next I found a free IP address on the server, created a new jail and started it up:

ezjail-admin create -x vhost2upgrade 10.0.0.42
ezjail-admin start vhost2upgrade
ezjail-admin console vhost2upgrade

Now I’m logged into a clone of my server jail and I can perform my MySQL upgrade and figure out all the little details before I perform it on the live server jail.

All done? Clean up after yourself. Log out of the jail and then from the host system:

ezjail-admin stop vhost2upgrade
ezjail-admin delete vhost2upgrade
zfs destroy zroot/jails/vhost2upgrade

An even better setup would be if I had set up my jail with server and user data on separate partitions. Then, once I had the server data working the way I wanted, I could zfs promote the clone’s server partition. To do.

I’ve finally got a working config for Apache with PHP-FPM, per-vhost pools, UIDs and chroots.  There seem to be a lot of tutorials around the net to help set up FPM with nginx, but very little with Apache.  The following instructions are for FreeBSD, but they would be easy to adapt to most any OS.  This document is still evolving, but I wanted to get it out to people in FreeNode #php-fpm who have been asking for help.

Why am I setting up PHP like this??

I have been using Apache with mod_php for years, and it works, but it has a number of problems, especially in a virtual hosting situation. All PHP scripts will run with the webserver’s UID, which is crummy for security. Users’ scripts can see the whole file system. When Apache services a non-PHP requests, such as an image or style sheet, it still has to load the whole PHP interpreter, using a bunch of memory.

This setup addresses each of these issues, hopefully making PHP sites more secure, and less memory hungry. Instead of including the mod_php interpreter in Apache uses the “FastCGI” protocol to parcel requests to a long-running “PHP-FPM” server. Each website I’m hosting is has its own configuration. Each runs under its own UID. Each is chroot-ed in the owner’s home directory. Only PHP requests are handled my PHP-FPM. Everything else stays in Apache.

Let’s get to the details

We’re going to install and configure a bunch of stuff software, and then set up a chroot environment.

Install Apache 2.2

cd /usr/ports/www/apache22
make install clean

Be sure to enable suexec in the Apache options dialog.

Enable Apache

Add apache22_enable=”YES” to /etc/rc.conf and start it up

service apache22 start

Install PHP-FPM

cd /usr/ports/lang/php5
make install clean
  • Do NOT build the Apache module.
  • DO build the FPM version
  • Building the CGI and CLI versions is fine as well
  • I add the mailhead patch too

Install the PHP extensions

This is a bit of a FreeBSD-ism, that you won’t have to do on most other OSs.  FreeBSD strips the PHP port down to a bare minimum, and moves all the plugins – including the default ones – into their own ports.  The php5-extensions meta-port collects them all into one place.

cd /usr/ports/lang/php5-extensions
make install clean

Add php_fpm_enable=”YES” to /etc/rc.conf and start it up

service php-fpm start

Install fastcgi

cd /usr/ports/www/mod_fastcgi/
make install clean

Edit httpd.conf, inserting:

LoadModule fastcgi_module     libexec/apache22/mod_fastcgi.so
LoadModule suexec_module        libexec/apache22/mod_suexec.so

and setting:

ServerAdmin webaster@example.com
ServerName server_ip_address_or_working_hostname

And un-comment the “Include” directives that make sense for me.
Now append:

NameVirtualHost *:80
Include etc/apache22/Includes/*.conf

and comment out this block:

#<Directory />
#    AllowOverride None
#    Order deny,allow
#    Deny from all
#</Directory>

Yes, that super-sucks.  Does anyone know of a workaround?

I like to keep each vhosts configuration in its own file, in a “vhosts/” directory, so I append:

Include etc/apache22/vhosts/*.conf

and

mkdir vhosts disabled-vhosts

You can guess what the second directory if for.  Now restart and see if that works.

service apache22 restart

You may get a warning like “NameVirtualHost *:80 has no VirtualHosts” because we haven’t added any yet.  Nothing to worry about.

Next create a Includes/php-fpm.conf for global fpm configs that will apply to every site.  Mine looks like:

FastCgiIpcDir /usr/local/etc/php-fpm/
FastCgiConfig -autoUpdate -singleThreshold 100 -killInterval 300 -idle-timeout 240 -maxClassProcesses 1 -pass-header HTTP_AUTHORIZATION
FastCgiWrapper /usr/local/sbin/suexec

<FilesMatch \.php$>
SetHandler php5-fcgi
</FilesMatch>

Action php5-fcgi /fcgi-bin

<Directory /usr/local/sbin>
Options ExecCGI FollowSymLinks
SetHandler fastcgi-script
Order allow,deny
Allow from all
</Directory>

See if Apache likes that:

service apache22 restart

Configure FPM

Now FPM needs some configuration.  Create a directory to store per-vhost fpm configs:

mkdir /usr/local/etc/fpm.d

Then edit the global php-fpm.conf, un-commenting:

include=etc/fpm.d/*.conf

switching the listen statement from a tcp port to:

listen = /tmp/php-fpm.sock

and changing the pm to:

pm = ondemand

There are a couple different types of process manager (pm).  On demand will prefork zero (0) processes.  They will only forked when needed.  I chose this for lots of small sites.  You may want a model that suits your setup better.

Now lets create a vhost.  Given a site named “example.com” owned by user “luser”, here’s my template:

<VirtualHost *:80>
ServerName        www.example.com
DocumentRoot    /home/luser/example.com/htdocs
SuexecUserGroup    luser luser
ServerAlias        example.com
ErrorLog        /home/luser/example.com/logs/example.com.error_log
CustomLog        /home/luser/example.com/logs/example.com.access_log combined

<Directory /home/luser/example.com/htdocs">
    Order allow,deny
    Allow from all
    Options +Indexes +FollowSymLinks +ExecCGI +Includes +MultiViews
    AllowOverride All
</Directory>

FastCgiExternalServer /tmp/fpm-example.com -socket /tmp/php-fpm-example.com.sock -user luser -group luser
Alias /fcgi-bin /tmp/fpm-example.com
<Location /fcgi-bin>
    Options +ExecCGI
    Order allow,deny
    Allow from all
</Location>

<LocationMatch "/(ping|fpm-status)">
    SetHandler php5-fcgi-virt
    Action php5-fcgi-virt /fcgi-bin virtual
</LocationMatch>
</VirtualHost>

And create a complimentary the FPM pool config:

[example.com]
user = luser
group = luser
listen = /tmp/php-fpm-example.com.sock
chroot = /home/luser
pm = ondemand
pm.max_children = 50
pm.status_path = /fpm-status
php_admin_value[doc_root] = /example.com/htdocs
php_admin_value[cgi.fix_pathinfo] = 0
php_admin_value[sendmail_path] = /bin/mini_sendmail -t -fwebmaster@internal.org

Living in a chroot
So PHP’s mail() function invokes your system’s sendmail binary, usually /usr/sbin/sendmail.  From within a chroot, that won’t be available.  However, there is the further problem that even if you copied sendmail and any libraries it needs into the chroot, it will want to write files to /var/spool, and again, that won’t be available.  We need a work around.  Install mini_sendmail.  It is a sendmail workalike that you can easily copy into a chroot, and instead of writing  to /var/spool, it will make an SMTP connection to localhost.  Be sure to set the -f envelope-sender in your fpm pool config, or mini_sendmail will use your username out of the environment when PHP or mini_sendmail was compiled, at the machine name.  PHP scripts can still override it using the mail() functions additional_parameters argument.

cd /usr/ports/mail/mini_sendmail
make install clean

Create a chroot environment for the vhost:

mkdir ~luser/tmp ~luser/bin
ln /tmp/mysql.sock ~luser/tmp/
cp /rescue/sh ~luser/bin/sh
ln /usr/local/bin/mini_sendmail ~luser/bin/mini_sendmail

PHP will need a /tmp directory.  If you are using MySQL, you will need to hardlink your mysql.sock into there or use TCP connections.  If you link the socket, you need to redo that EVERY time you restart MySQL.  (I should include my rc script here).  Hard link mini_sendmail into the chroot.  And finally, PHP needs a shell to invoke sendmail.  Yes this sucks.  You can copy /bin/sh in, but chances are, it needs libraries that aren’t in the chroot.  I could copy those too, but I just copied the crunched binary from FreeBSD’s /rescue dir.  Yes, this sucks even more because it includes stuff I don’t want or need, and I need a better solution.  TODO: crunch my own sh with a couple other useful items.  Maybe use busybox for this?

Set the tmp dir in php.ini to

upload_tmp_dir = /tmp

Update #1

I had a problem with a number of server variables not getting properly translated for use within the chroot, so I added a php prepend directive to the php-fpm conf files like:

php_admin_value[auto_prepend_file] = /bin/phpfix

And then linked this file into each chroot’s ~/bin/ directory:

$_SERVER['DOCUMENT_ROOT'] = ini_get('doc_root');
$_SERVER['PATH_TRANSLATED'] = str_replace($_SERVER['HOME'], '', $_SERVER['PATH_TRANSLATED']);
$_SERVER['SCRIPT_FILENAME'] = str_replace($_SERVER['HOME'], '', $_SERVER['SCRIPT_FILENAME']);

Update #2

PHP’s streams tools (like file_get_contents()) rely on openssl for HTTPS URLs, and many other plugins (like SOAP) in turn rely on those streams. Curl seems to function just fine in a chroot, but PHP’s openssl streams require certain device nodes to function. You will have to mount /dev inside your chroot in order to use them. More on this when I get a good system in place.