...(PHP version < 5.5)

This little addition allows clear text error descriptions also for PHP 5.4 and earlier. I don't check for the existence of curl_error as it exists along curl_errno since PHP v4.0.3.

When the decode_contentoption is enabled and the server return an encoded response, the returned content-length and content-encoding headers must be fixed.

This problem was already reported and fixed in Guzzle 6+. See https://github.com/guzzle/guzzle/issues/1075

Hey!

I'm currently implementing SSL verification in egeloen/ivory-http-adapter#62 which test all adapters of guzzle http in order to be sure everything is working fine. When doing it, I discover an inconsitence between Curl (Single/Multi) and the Stream one. Basically, in my test suite I have generated a self signed certificate which allow me to test the feature. For Curl and MultiCurl, if I enable ssl verification, I got an exception because it does not allow self signed cetificate whereas the stream one does not use the same strategy and allow it. This PR rationalizes this behavior.

Feedback welcome!

When calling then() on a FutureArray or FutureValue, an instance of FutureInterface is returned. It would be nice if an instance of FutureArray or FutureValue would be returned instead.

Typical use case

I have a service that communicates with ElasticSearch via their PHP library (that for async calls returns a FutureArray) and the service somehow processes the result (the ES response contains lots of stuff I don't need) - that's where the then() method comes to play.

I want to use the service output in both a cli command (for one-time testing/debug/...) and a worker that runs completely asynchronously in an event loop. That means in the first case, I would get the result by calling wait() and in the second one, I would simply chain another then() call to the service output.

Problem

Calling then() on the FutureArray returns a PromiseInterface (the result of calling then() on the wrapped promise) and I am unable to block until a result is available using the wait() method because it simply does not exist in React promises.

Solution

The then() method can return a new FutureArray that simply wraps the result originally returned.

Considering the FutureArray already implements PromiseInterface, this should be no BC break.

Instead of new self(...) the method could use late static binding and call new static(...) but the only difference would be if someone subclassed FutureArray. And in such case, it is quite possible that they created their own constructor with a different signature.

The same applies to FutureValue.

I have included two unit tests that test that the then() method returns the apprioriate class instance and that the returned value was processed by the callback.

If you call Guzzle client using debug=false (having curl disabled) you get printed debug information, since actually right now its just validating that the key exists but your are not checking the value.

When received multiple Set-Cookie headers in same response, after parseHeaderOutput the cookies order is reversed. Can break in situations where order of cookies is important because later in guzzlehttp/guzzle setCookie is changing the content of the same-name cookie. Therefore instead of eventually ending in cookieJar with the newest cookie value, we are left with the oldest. This for example has the effect of breaking secure logins in many servers.

When curl has the error status "65" "rewind failed". Which can happen when putting a big file via a resource. Which gets a redirect as response. Then this flow will happen:

We have a curl error https://github.com/PBWebMedia/RingPHP/blob/master/src/Client/CurlFactory.php#L90

return !empty($response['curl']['errno']) || !isset($response['status'])
            ? self::createErrorResponse($handler, $request, $response)
            : $response;

We have error number 65 https://github.com/PBWebMedia/RingPHP/blob/master/src/Client/CurlFactory.php#L112

// Retry when nothing is present or when curl failed to rewind.
        if (!isset($response['err_message'])
            && (empty($response['curl']['errno'])
                || $response['curl']['errno'] == 65)
        ) {
            return self::retryFailedRewind($handler, $request, $response);
        }

Nothing of the above happens so we run the handle (gues he follows the redirect) https://github.com/PBWebMedia/RingPHP/blob/master/src/Client/CurlFactory.php#L544

return $handler($request);

But the problem here is that the handle does not return an array but an CompletedFutureArray

Here is the handle set: https://github.com/guzzle/RingPHP/blob/master/src/Client/CurlHandler.php#L78

return new CompletedFutureArray(
            CurlFactory::createResponse($this, $request, $response, $hd, $bd)
        );

As we can see in the row above that will return a CompletedFutureArray

Which it will put in CompletedFutureArray::__construct() But the constructor accepts an array and not a CompletedFutureArray.

I thought it would be weird to modify CompletedFutureArray so it can also accept itself. It seemed more clean to make sure that CurlFactory::createResponse always return an array instead having 1 case that returns an object.

Currently, the adapter may process any new requests added by listeners one curl_multi_select cycle late. For example, a complete listener sends a new request (on the line $this->processMessages();), which results in curl_multi_add_handle being called, but after that the execute loop will call curl_multi_select waiting for a previous request to be completed before adding the new request to the multi pool via curl_multi_exec.

Hope I managed to described the problem well. I know it's only a copy-paste of previous few lines, but I don't have a better idea at the moment. Here at ManageWP we use our own fork of the CurlMultiAdapter for Guzzle 4, but it seems it's much easier for Guzzle 5 to implement this functionality.

when headers coming from curl are like this Set-cookie: a=b Set-Cookie: c=d then there is a bug, because of associative array of headers keys of array of headers are case-sensitive

Make the test server aware that its http-auth module is not available, when responding to an authentified request. Now, in this situation (authentified query AND http-auth not here) the server will not crash, but will reply with an HTTP 501.

Hey!

This PR extracts the response protocol version in order to propagate it in the response. Without this extraction/parsing, the response always get an 1.1 protocol version (default value) even when the response is an 1.0 one.