111 lines
3.3 KiB
PHP
111 lines
3.3 KiB
PHP
|
<?php
|
||
|
|
||
|
namespace Guzzle\Parser\Message;
|
||
|
|
||
|
/**
|
||
|
* Default request and response parser used by Guzzle. Optimized for speed.
|
||
|
*/
|
||
|
class MessageParser extends AbstractMessageParser
|
||
|
{
|
||
|
public function parseRequest($message)
|
||
|
{
|
||
|
if (!$message) {
|
||
|
return false;
|
||
|
}
|
||
|
|
||
|
$parts = $this->parseMessage($message);
|
||
|
|
||
|
// Parse the protocol and protocol version
|
||
|
if (isset($parts['start_line'][2])) {
|
||
|
$startParts = explode('/', $parts['start_line'][2]);
|
||
|
$protocol = strtoupper($startParts[0]);
|
||
|
$version = isset($startParts[1]) ? $startParts[1] : '1.1';
|
||
|
} else {
|
||
|
$protocol = 'HTTP';
|
||
|
$version = '1.1';
|
||
|
}
|
||
|
|
||
|
$parsed = array(
|
||
|
'method' => strtoupper($parts['start_line'][0]),
|
||
|
'protocol' => $protocol,
|
||
|
'version' => $version,
|
||
|
'headers' => $parts['headers'],
|
||
|
'body' => $parts['body']
|
||
|
);
|
||
|
|
||
|
$parsed['request_url'] = $this->getUrlPartsFromMessage($parts['start_line'][1], $parsed);
|
||
|
|
||
|
return $parsed;
|
||
|
}
|
||
|
|
||
|
public function parseResponse($message)
|
||
|
{
|
||
|
if (!$message) {
|
||
|
return false;
|
||
|
}
|
||
|
|
||
|
$parts = $this->parseMessage($message);
|
||
|
list($protocol, $version) = explode('/', trim($parts['start_line'][0]));
|
||
|
|
||
|
return array(
|
||
|
'protocol' => $protocol,
|
||
|
'version' => $version,
|
||
|
'code' => $parts['start_line'][1],
|
||
|
'reason_phrase' => isset($parts['start_line'][2]) ? $parts['start_line'][2] : '',
|
||
|
'headers' => $parts['headers'],
|
||
|
'body' => $parts['body']
|
||
|
);
|
||
|
}
|
||
|
|
||
|
/**
|
||
|
* Parse a message into parts
|
||
|
*
|
||
|
* @param string $message Message to parse
|
||
|
*
|
||
|
* @return array
|
||
|
*/
|
||
|
protected function parseMessage($message)
|
||
|
{
|
||
|
$startLine = null;
|
||
|
$headers = array();
|
||
|
$body = '';
|
||
|
|
||
|
// Iterate over each line in the message, accounting for line endings
|
||
|
$lines = preg_split('/(\\r?\\n)/', $message, -1, PREG_SPLIT_DELIM_CAPTURE);
|
||
|
for ($i = 0, $totalLines = count($lines); $i < $totalLines; $i += 2) {
|
||
|
|
||
|
$line = $lines[$i];
|
||
|
|
||
|
// If two line breaks were encountered, then this is the end of body
|
||
|
if (empty($line)) {
|
||
|
if ($i < $totalLines - 1) {
|
||
|
$body = implode('', array_slice($lines, $i + 2));
|
||
|
}
|
||
|
break;
|
||
|
}
|
||
|
|
||
|
// Parse message headers
|
||
|
if (!$startLine) {
|
||
|
$startLine = explode(' ', $line, 3);
|
||
|
} elseif (strpos($line, ':')) {
|
||
|
$parts = explode(':', $line, 2);
|
||
|
$key = trim($parts[0]);
|
||
|
$value = isset($parts[1]) ? trim($parts[1]) : '';
|
||
|
if (!isset($headers[$key])) {
|
||
|
$headers[$key] = $value;
|
||
|
} elseif (!is_array($headers[$key])) {
|
||
|
$headers[$key] = array($headers[$key], $value);
|
||
|
} else {
|
||
|
$headers[$key][] = $value;
|
||
|
}
|
||
|
}
|
||
|
}
|
||
|
|
||
|
return array(
|
||
|
'start_line' => $startLine,
|
||
|
'headers' => $headers,
|
||
|
'body' => $body
|
||
|
);
|
||
|
}
|
||
|
}
|