Edited the php version, since it affects avery version.

I want to patch the sources, but can't find a way to get from the function 
init_request_info info fpm_main.c the chroot path.
Since the documentation about it lacks, can anyone put me in the right 
direction? 
:P

I think I need to get the configuration (or add a variable to the worker) about 
the current chroot location. A more safe way to handle it would be to get, from 
the father process, the child's chroot location and filter it before passing the 
variables to the child.

Anyway, in both cases, there is lack of documentation and I've still been unable 
to patch anything, so suggestions would be really appreciated! :)

The symbolic trick does work (as long as the sym link is relative (ie ../../../chrootpath if you try to link directly from the root, then the symlink goes outside the jail and it will still fail).

Anyway thats a very messy hack, and I agree with you, php-fpm creates the jails, so it should fix the PHP variables.

Without the symlink, I just get file not found for all calls for scripts. 

If you change fix.cgi_pathinfo = 0 it will run the scripts, but the $_SERVER variables will still be wrong... Im sure most people would rather it was fixed so that the fix.cgi_pathinfo = 1 works, AND the variables are set correctly... But either way would be better than trying to hack in fixes with sym links, and/or preloading a fixup script.

I did take a look at the fpm_main.c but also been unable to create my own patch for it at this time. Surely it cant be that hard to read the chroot path from the configuration file, and remove that from all appropriate server variables, or recreate the broken variables some other way.

Symlink is good enough for a test environment, to see what other php functions break with the chroot, and to start building a system, but I wouldn't use that hack on a production web server.

In the meantime, I've already found a way to fix this behaviour: a proxy between 
PHP-FPM and Apache.

I needed a caching proxy to be on top of Apache to cache php scripts (I don't care 
about static files because what really causes CPU and memory problems is having 
the php processes running), so I've thought to implement a fcgi proxy to be used 
among PHP-FPM and Apache.

It does work and does a great job: I didn't need to implement anything about the 
HTTP protocol since there's already Apache caring about it; then, caching the 
results from PHP, I can avoid asking FPM to run the already cached pages, giving 
them away to Apache in an elegant manner. And, of course, the proxy does fix the 
wrong variables, removing the chrooted part and making the scripts able to work.

I'm still trying to fix the sources, but since they appears complicated (read: 
messy), and there's no documentation about it or about debugging them, I don't 
think I will fix them before I have more time to spend or having a clue about 
reading the configuration values from fpm_main.c.

Anyway, for anyone interested in it, the proxy has been made using python and flup 
server and client implementations.

From my point of view, I'd rather prefer to configure this on the web server 
side (if it's not possible with apache then switch to something else -- aka 
nginx for exemple. If apache does not suit well with fastcgi customizations, 
then ask apache guys :p)

more seriously, I can understand the need. I'll try to look at this to 
dynamicaly adapt fastcgi variables from the chroot env. The following algorithm 
should work:

for variable in  all_fastcgi_variables; do
  if ($_SERVER[$var] =~ /^$chroot(.*)$/); then
    $_SERVER[$var] = $1;
  fi
done

or somehting more restrictive:

for var in "DOCUMENT_ROOT" "SCRIPT_FILENAME" "PATH_TRANSLATED"]; do
  if ($_SERVER[$var] =~ /^$chroot(.*)$/); then
    $_SERVER[$var] = $1;
  fi
end

Thank your for your support but I'm sorry, I think I didn't understand your fix: 
does php-fpm chroot works starting it from bash, or how could it be implemented 
the way you said (from inside fpm I mean)?

I used pseudo bash syntax but it'll be implemanted inside fpm (with C code)

Moving to nginx would be lovely, but on many low cost virtual hosting scenarios where a dedicated server would be totally overkill, many end users are still heavily using .htaccess files (Yes I know the arguments against.. performance, excessive stats,  too many reads etc etc...)

Take just about any off the shelf, wordpress, drupal, magento etc, and it comes ready to roll on Apache, and fully loaded with .htaccess files.

The argument against changing it from the Apache side of things, is that Apache has simply followed the standards for FastCGI, and that since its PHP-FPM that has created the jail, Apache doesnt even need to know that its happened.

Your proposed solution to the problem should sort it out though, I assume it would also avoid the "not found" problem when fix.cgi_pathinfo is set to 1.

If you have a patch, Im quite happy to try it on one of my systems to check it corrects the problem in various "off the shelf" configurations, and if any problems still occur, im happy to share the config file.

PHP-FPM works brilliantly with Apache / FastCGI with sockets.. but would be nice to get the chroot working without resorting to symlink workarounds etc.

More 1 year (15 month) after the last comment, i have the same problem and i don't think this bug is solved.

Everybody have migrate to Nginx instead Apache ? I can't do it because my hosted need to use .htaccess for Apache.

Thanks so much for your reply ...

This was partially fixed in #69090

#69090 has nothing to do with this bug. This bug still exists in 7.1.