Background info fast-cgi-client v2.6.0

Background info about a faulty change in v2.5.0 and why that caused a minor BC break in v2.6.0 of hollodotme/fast-cgi-client.

Let's begin...

...with the history of the v2.5.0 release.

In February 2019 Arne Blankerts reached out to me via twitter DM and we had a discussion about the evergreen problem of the fast-cgi-client library that any relative script path or any script path containing path traversal characters lead to a (silent) fail of the FastCGI request.

Some example paths that do not work when requesting a script on php-fpm, which is the focus FastCGI server that the library is developed against:

Requesting such paths will lead to a File Not Found response from php-fpm, even though they might exist and would resolve to absolute paths in the file system. However, php-fpm does not do any automatic path resolution and expects an absolute path to the script that shall be executed. Full stop.

At that time v2.4.3 was the most recent version of the library and there already was a hint on the trouble shooting section of the README.md regarding this kind of issue. Arne's primary question was:

From user perspective:
Shouldn't the library be responsible to handle this problem in any better way
than silently returning a supposedly successful response?

Well, that would be great, right?! So I agreed.

I then had a look into the code because I was wondering why the response was not the original error message from php-fpm which is "Primary script unknown" in this case. After some testing and step-debugging I realized that there was a general bug in the response handling of the underlying Socket class. The FastCGI protocol separates STDOUT and STDERR byte streams when sending response packets, but the handling of the client at that time simply concatenated both streams to compose the complete server response. This resulted in responses that looked like this:

Primary script unknownStatus: 404 Not Found
X-Powered-By: PHP/7.3.1

File not found

The response object that the client provides for each request than made the error messages disappear in the first response header as it assumes all lines before the blank line to be a header that was sent from the server. So the header array of the parsed response looked like this:

[
    'Primary script unknownStatus' => '404 Not Found',
    'X-Powered-By' => 'PHP/7.3.1'
]

I slaped myself and called me names! -.-

So I decided to save the STDERR stream to a separate variable and throw a newly introduced exception whenever this variable is not empty. That solved two problems:

  1. The server response that gets parsed by the response class does not contain (and therefor hide) any prepended error messages.
  2. In case the server sent an error an exception is thrown that could be gracefully handled by the library user.

And that was basically the change that made it into the v2.5.0 release.

A short time later ...

... Andy Buckingham opened an issue on GitHub and stated that fetching custom output from e.g. error_log() is not accessible on the client side anymore, because all output from the STDERR stream is converted into an exception.

To summarize the whole discussion shortly: I tried to find a reliable way of identifying an error produced by the FastCGI server, so the client could be able to tell the user "Hey something went wrong, please have a look!".

php-fpm, which is only one FastCGI server implementation, tosses exactly 4 errors at the user in different circumstances as you can see here. But that might be different for other FastCGI server implementations.

Arne suggested to provide the status code from the equally named response header, but unfortunately those headers are not part of the FastCGI protocol (but of HTTP) and php-fpm does not always set such a status header as explained here.

So in the end there was the following conclusion for the next v2.6.0 release:

First and foremost, introducing the ProcessManagerException was a bad idea in the first place.

From a strict point of view the last point is a BC break, as users may have implemented an appropriate exception handling after they upgraded to v2.5.0, but after considering all the arguments from Andy, Arne and me, it was simply a wrong move to introduce this exception as it leads to uninteded behaviour for another common use case (retrieval of STDERR output).

You can find the new v2.6.0 release here.

So how to handle server-sent errors now?

Let's stick to the initial problem of using a non-absolute script path for a request to php-fpm:

<php declare(strict_types=1);

namespace YourVendor\YourProject;

use hollodotme\FastCGI\Client;
use hollodotme\FastCGI\Requests\GetRequest;
use hollodotme\FastCGI\SocketConnections\NetworkSocket;

$connection = new NetworkSocket('127.0.0.1', 9000);
$client     = new Client($connection);
$request    = new GetRequest('../relative/path/to/script.php', '');

$response = $client->sendRequest($request);

The response can be checked as follows for the error.

1. Check the error output for specific error messages

if (preg_match("#^Primary script unknown\n?$#", $response->getError()))
{
    throw new LogicException('The script cannot be found on FastCGI server, please check if path is absolute.');
}

Please note the optional \n? in the regex. In PHP versions prior to 7.3 error messages from php-fpm have a trailing line feed, since 7.3 they have not. (Had to find that out the hard way. -.-)

2. Check for the status header

if ('404 Not Found' === $response->getHeader('Status'))
{
    throw new LogicException('The script cannot be found on FastCGI server, please check if path is absolute.');
}

Please note, that such a header is not always sent if an error occurrs. But in this error case you can rely on it with php-fpm as server.

And please be also aware that all server error scenarios can be user-generated on server side, like this:

<?php declare(strict_types=1);

error_log('Primary script unknown');
header('Status: 404 Not Found');
echo 'File not found.';

I hope I could clarify the library changes from v2.4.3 to v2.6.0 and you still enjoy using the lib.

Please report any issues via GitHub at hollodotme/fast-cgi-client.

Thank you!

04/02/2019