Network Stack

Network Stack

目录

1. 1 Overview
2. 2 Code Layout
3. 3 Anatomy of a Network Request (focused on HTTP)
   1. 3.1 URLRequest
   2. 3.2 URLRequestHttpJob
   3. 3.3 HttpNetworkTransaction
   4. 3.4 HttpStreamFactory
      1. 3.4.1 Proxy Resolution
      2. 3.4.2 Connection Management
      3. 3.4.3 Host Resolution
      4. 3.4.4 SSL/TLS

Overview

The network stack is a mostly single-threaded cross-platform library primarily for resource fetching.  Its main interfaces are URLRequest and URLRequestContext.  URLRequest, as indicated by its name, represents the request for a URL.  URLRequestContext contains all the associated context necessary to fulfill the URL request, such as cookies, host resolver, proxy resolver, cache, etc. Many URLRequest objects may share the same URLRequestContext.  Most net objects are not threadsafe, although the disk cache can use a dedicated thread, and several components (host resolution, certificate verification, etc.) may use unjoined worker threads.  Since it primarily runs on a single network thread, no operation on the network thread is allowed to block.  Therefore we use non-blocking operations with asynchronous callbacks (typically CompletionCallback).  The network stack code also logs most operations to NetLog, which allows the consumer to record said operations in memory and render it in a user-friendly format for debugging purposes.

 

Chromium developers wrote the network stack in order to:

• Allow coding to cross-platform abstractions
• Provide greater control than would be available with higher-level system networking libraries (e.g. WinHTTP or WinINET)
• Avoid bugs that may exist in system libraries
• Enable greater opportunity for performance optimizations

Code Layout

• net/base - Grab bag of net utilities, such as host resolution, cookies, network change detection, SSL.
• net/disk_cache - Cache for web resources.
• net/ftp - FTP implementation.  Code is primarily based on the old HTTP implementation.
• net/http - HTTP implementation.
• net/ocsp - OCSP implementation when not using the system libraries or if the system does not provide an OCSP implementation. Currently only contains an NSS based implementation.
• net/proxy - Proxy (SOCKS and HTTP) configuration, resolution, script fetching, etc.
• net/quic - QUIC implementation.
• net/socket - Cross-platform implementations of TCP sockets, "SSL sockets", and socket pools.
• net/socket_stream - socket streams for WebSockets.
• net/spdy - HTTP2 (and its predecessor) SPDY implementation.
• net/url_request - URLRequest, URLRequestContext, and URLRequestJob implementations.
• net/websockets - WebSockets implementation.

Anatomy of a Network Request (focused on HTTP)

URLRequest

 

class URLRequest {

 public:

  // Construct a URLRequest for |url|, notifying events to |delegate|.

  URLRequest(const GURL& url, Delegate* delegate);

  

  // Specify the shared state

  void set_context(URLRequestContext* context);

 

  // Start the request. Notifications will be sent to |delegate|.

  void Start();

 

  // Read data from the request.

  bool Read(IOBuffer* buf, int max_bytes, int* bytes_read);

};

 

class URLRequest::Delegate {

 public:

  // Called after the response has started coming in or an error occurred.

  virtual void OnResponseStarted(...) = 0;

 

  // Called when Read() calls complete.

  virtual void OnReadCompleted(...) = 0;

};

 

When a URLRequest is started, the first thing it does is decide what type of URLRequestJob to create. The main job type is the URLRequestHttpJob which is used to fulfill http:// requests.  There are a variety of other jobs, such as URLRequestFileJob (file://), URLRequestFtpJob (ftp://), URLRequestDataJob (data://), and so on. The network stack will determine the appropriate job to fulfill the request, but it provides two ways for clients to customize the job creation: URLRequest::Interceptor and URLRequest::ProtocolFactory. These are fairly redundant, except that URLRequest::Interceptor's interface is more extensive. As the job progresses, it will notify the URLRequest which will notify the URLRequest::Delegate as needed.

URLRequestHttpJob

URLRequestHttpJob will first identify the cookies to set for the HTTP request, which requires querying the CookieMonster in the request context. This can be asynchronous since the CookieMonster may be backed by an sqlitedatabase. After doing so, it will ask the request context's HttpTransactionFactory to create a HttpTransaction. Typically, the HttpCache will be specified as the HttpTransactionFactory. The HttpCache will create a HttpCache::Transaction to handle the HTTP request. The HttpCache::Transaction will first check the HttpCache (which checks the disk cache) to see if the cache entry already exists. If so, that means that the response was already cached, or a network transaction already exists for this cache entry, so just read from that entry. If the cache entry does not exist, then we create it and ask the HttpCache's HttpNetworkLayer to create a HttpNetworkTransaction to service the request. The HttpNetworkTransaction is given a HttpNetworkSession which contains the contextual state for performing HTTP requests. Some of this state comes from the URLRequestContext.

HttpNetworkTransaction

 

class HttpNetworkSession {

 ...

 

 private:

  // Shim so we can mock out ClientSockets.

  ClientSocketFactory* const socket_factory_;

  // Pointer to URLRequestContext's HostResolver.

  HostResolver* const host_resolver_;

  // Reference to URLRequestContext's ProxyService

  scoped_refptr<ProxyService> proxy_service_;

  // Contains all the socket pools.

  ClientSocketPoolManager socket_pool_manager_;

  // Contains the active SpdySessions.

  scoped_ptr<SpdySessionPool> spdy_session_pool_;

  // Handles HttpStream creation.

  HttpStreamFactory http_stream_factory_;

};

 

HttpNetworkTransaction asks the HttpStreamFactory to create a HttpStream.  The HttpStreamFactory returns a HttpStreamRequest that is supposed to handle all the logic of figuring out how to establish the connection, and once the connection is established, wraps it with a HttpStream subclass that mediates talking directly to the network.

 

 

class HttpStream {

 public:

  virtual int SendRequest(...) = 0;

  virtual int ReadResponseHeaders(...) = 0;

  virtual int ReadResponseBody(...) = 0;

  ...

};

 

Currently, there are only two main HttpStream subclasses: HttpBasicStream and SpdyHttpStream, although we're planning on creating subclasses for HTTP pipelining.  HttpBasicStream assumes it is reading/writing directly to a socket.  SpdyHttpStream reads and writes to a SpdyStream. The network transaction will call methods on the stream, and on completion, will invoke callbacks back to the HttpCache::Transaction which will notify the URLRequestHttpJob and URLRequest as necessary. For the HTTP pathway, the generation and parsing of http requests and responses will be handled by the HttpStreamParser. For the SPDY pathway, request and response parsing are handled by SpdyStream and SpdySession. Based on the HTTP response, the HttpNetworkTransaction may need to perform HTTP authentication. This may involve restarting the network transaction.

HttpStreamFactory

HttpStreamFactory first does proxy resolution to determine whether or not a proxy is needed. The endpoint is set to the URL host or the proxy server. HttpStreamFactory then checks the SpdySessionPool to see if we have an available SpdySession for this endpoint. If not, then the stream factory requests a "socket" (TCP/proxy/SSL/etc) from the appropriate pool. If the socket is an SSL socket, then it checks to see if NPN indicated a protocol (which may be SPDY), and if so, uses the specified protocol. For SPDY, we'll check to see if a SpdySession already exists and use that if so, otherwise we'll create a new SpdySession from this SSL socket, and create a SpdyStreamfrom the SpdySession, which we wrap a SpdyHttpStream around. For HTTP, we'll simply take the socket and wrap it in a HttpBasicStream.

Proxy Resolution

HttpStreamFactory queries the ProxyService to return the ProxyInfo for the GURL. The proxy service first needs to check if it has an up-to-date proxy configuration. If not, it uses the ProxyConfigService to query the system for the current proxy settings. If the proxy settings are set to no proxy or a specific proxy, then proxy resolution is simple (we return no proxy or the specific proxy). Otherwise, we need to run a PAC script to determine the appropriate proxy (or lack thereof). If we don't already have the PAC script, then the proxy settings will indicate we're supposed to use WPAD auto-detection, or a custom PAC url will be specified, and we'll fetch the PAC script with the ProxyScriptFetcher. Once we have the PAC script, we'll execute it via the ProxyResolver. Note that we use a shim MultiThreadedProxyResolver object to dispatch the PAC script execution to threads, which run a ProxyResolverV8 instance. This is because PAC script execution may block on host resolution. Therefore, in order to prevent one stalled PAC script execution from blocking other proxy resolutions, we allow for executing multiple PAC scripts concurrently (caveat: V8 is not threadsafe, so we acquire locks for the javascript bindings, so while one V8 instance is blocked on host resolution, it releases the lock so another V8 instance can execute the PAC script to resolve the proxy for a different URL).

Connection Management

After the HttpStreamRequest has determined the appropriate endpoint (URL endpoint or proxy endpoint), it needs to establish a connection. It does so by identifying the appropriate "socket" pool and requesting a socket from it. Note that "socket" here basically means something that we can read and write to, to send data over the network. An SSL socket is built on top of a transport (TCP) socket, and encrypts/decrypts the raw TCP data for the user. Different socket types also handle different connection setups, for HTTP/SOCKS proxies, SSL handshakes, etc. Socket pools are designed to be layered, so the various connection setups can be layered on top of other sockets. HttpStream can be agnostic of the actual underlying socket type, since it just needs to read and write to the socket. The socket pools perform a variety of functions-- They implement our connections per proxy, per host, and per process limits. Currently these are set to 32 sockets per proxy, 6 sockets per destination host, and 256 sockets per process (not implemented exactly correctly, but good enough). Socket pools also abstract the socket request from the fulfillment, thereby giving us "late binding" of sockets. A socket request can be fulfilled by a newly connected socket or an idle socket (reused from a previous http transaction).

Host Resolution

Note that the connection setup for transport sockets not only requires the transport (TCP) handshake, but probably already requires host resolution. HostResolverImpl uses getaddrinfo() to perform host resolutions, which is a blocking call, so the resolver invokes these calls on unjoined worker threads.  Typically host resolution usually involves DNS resolution, but may involve non-DNS namespaces such as NetBIOS/WINS. Note that, as of time of writing, we cap the number of concurrent host resolutions to 8, but are looking to optimize this value. HostResolverImpl also contains a HostCache which caches up to 1000 hostnames.

SSL/TLS

SSL sockets require performing SSL connection setup as well as certificate verification. Currently, on all platforms, we use NSS's libssl to handle the SSL connection logic. However, we use platform specific APIs for certificate verification. We are moving towards using a certificate verification cache as well, which will consolidate multiple requests for certificate verification of the same certificate into a single certificate verification job and cache the results for a period of time.

 

SSLClientSocketNSS roughly follows this sequence of events (ignoring advanced features like Snap Start or DNSSEC based certificate verification):

• Connect() is called. We set up NSS's SSL options based on SSLConfig specified configuration or preprocessor macros. Then we kickoff the handshake.
• Handshake completes. Assuming we didn't hit any errors, we proceed to verify the server's certificate using CertVerifier. Certificate verification may take some amount of time, so CertVerifier uses the WorkerPool to actually call X509Certificate::Verify(), which is implemented using platform specific APIs.

Note that Chromium has its own NSS patches which support some advanced features which aren't necessarily in the system's NSS installation, such as support for NPN, False Start, Snap Start , OCSP stapling, etc.

 

TODO: talk about network change notifications