Squid Architecture

Broad Overview

Squid is operating at layers 4-7 on the OSI data model. So unlike most networking applications there is no relationship between packets (a layer 3 concept) and the traffic received by Squid. Instead of packets HTTP operates on a message basis (called segments in the OSI model definitions), where an HTTP request and response can each be loosely considered equivelent to one "packet" in a transport architecture. Just like IP packets HTTP messages are stateless and the delivery is entirely optional for process. See the RFC 7230 texts for a better description on HTTP specifics and how it operates.

At the broad level Squid consists of five generic processing areas;

  • client facing (was "client-side") which implements HTTP, HTTPS, PROXY, FTP, ICP, HTCP, and SNMP protocols to communicate with clients, and
  • server facing (was "server-side") which implements HTTP, HTTPS, ICY, FTP, Gopher, WAIS, URN-N2H, and blind TCP tunnels to communicate with web servers (with or without an upstream proxy/relay) , and
  • between them is the cache storage. Which in broad terms provides the buffering mechanisms for data transit, and provides switching logic to determine data source between disk, memory, and server-side.
  • on each side of storage after client facing logic, and before server-facing there are optional diversions to pass the traffic through content adaptation services (URI rewrite, redirection, ICAP, eCAP, or HTTP request/reply manglers).
  • there is also a set of components performing extra support tasks. Such as; security (authentication and access control), (m)DNS client, IDENT client, WHOIS client, logging, QoS, WCCP, NAT client, cache/peer integration. Some are obvious, some not - protocols implemented which are not documented as directly server or client facing usually fall into this group.
    • BroadOverview.png

/!\ TODO: image is outdated.

  • Client Side now includes SNMP, PROXY, FTP, (with many bugs) SFTP protocols and "other" protocols.
  • Storage removed COSS and added CLP caches
  • Extra Processing Logics includes URN, ICAP, eCAP
  • Server side missing URN N2H translator.

General Overview

  • {i} The general design level is where Squid-2 and Squid-3 differ. With Squid-2 being composed purely of event callback chains, Squid-3 adds the model of task encapsulation within Jobs.

/!\ TODO: images with overview of data flow.

/!\ TODO: pull in existing descriptions of I/O event model, AsyncJob model from source code.

/!\ TODO: data processing diagram with color-coded for display of AsyncJob vs Event callback coverage.

Transaction Processing

A connection (class Comm::Connection) begins with class TcpAcceptor accept() when the TCP socket is accepted, or for UDP when a packet is received. TCP connections end when the class Comm::Connection::close() method is called, which must be done explicitly. UDP connections close when the class Comm::Connection objects last reference is dropped, NOTE that UDP sockets are shared so Comm::Connection::close() must not be used by UDP transactions, it may close all active UDP traffic.

The following apply to TCP based protocols only:

A master transaction (class MasterXaction) manages information shared among the HTTP or FTP request and all related protocol messages (such as the corresponding HTTP/FTP/Gopher/etc. protocol response and control messages) as well as ICAP, eCAP, and helper transactions caused by protocol messages. Much of the code necessary to collect and share this information has yet to be developed.

A stream transaction is an HTTP or FTP request with the corresponding final protocol reply. It begins with a class !Parser successful parse() call when the protocol request arrives on the connection. The details from its MasterXaction are copied into a AccessLogEntry which accumulate the details about the stream and eventually winds up in access.log.

An ICAP transaction (class Adaptation::Icap::Xaction) is an ICAP request with the corresponding final ICAP response. Many ICAP transactions may occur for a single Master Transaction and, IIRC, ICAP OPTIONS revalidation transactions occur without a Master Transaction.

A helper transaction (class Helper::Xaction) may occur for each plugin helper which squid.conf settings may cause to be used by the stream transaction.

/!\ TODO: alter the master transaction definition to cope with UDP based protocols involving streams and content adapted. eg SNMPv3, HTTP/3, QUICK, CoAP, CoAPS, DNS, WebSockets3

HTTP Request

The following sequence of checks and adjustments is applied to most HTTP requests. This sequence starts after Squid parses the request header and ends before Squid starts satisfying the request from the cache or origin server. The checks are listed here in the order of their execution:

  1. Host header forgery checks
  2. http_access directive

  3. ICAP/eCAP adaptation

  4. redirector

  5. adapted_http_access directive

  6. store_id directive

  7. clientInterpretRequestHeaders()
  8. cache directive

  9. ToS marking
  10. nf marking
  11. ssl_bump directive

  12. callout sequence error handling

A failed check may prevent subsequent checks from running.

A typical HTTP transaction (i.e., a pair of HTTP request and response messages) goes through the above sequence once. However, multiple transactions may participate in processing of a single "web page download", confusing Squid admins. While all experienced Squid admins know that a single web page may contain dozens and sometimes hundreds of resources, each triggering an HTTP transaction, those multiple transactions may happen even when requesting a single resource and even when using simple command-line tools like curl or wget.

Internal Squid requests may cause even more confusion. For example, when SslBump is in use, Squid may create several fake CONNECT transactions for a given TLS connection, and each CONNECT may go through the above motions. If you use SslBump for intercepted port 443 traffic, then shortly after a new connection is accepted by Squid, SslBump creates a fake CONNECT request with TCP level information, and that CONNECT request goes through the above sequence (matching step SslBump1 ACL if any). If an "ssl_bump peek" or "ssl_bump stare" rule matches during that first SslBump step, then SslBump code gets SNI and creates a second fake CONNECT request that goes through the same sequence again.

Similarly (S)FTP native services have each message in a Stream Transaction translated into various HTTP messages which should go through the above above motions.

Your Squid directives and helpers must be prepared to deal with multiple [CONNECT] requests per connection.

/!\ TODO: document forwarding destination selection

/!\ TODO: HTTP response callback processing sequence

/!\ TODO: non-TCP transaction processing sequence?

/!\ TODO: non-HTTP stream transactions?


Discuss this page using the "Discussion" link in the main menu

ProgrammingGuide/Architecture (last edited 2022-01-17 16:38:58 by AmosJeffries)