Built in subroutines¶

Various built-in subroutines are called during processing of client- and backend requests as well as upon vcl.load and vcl.discard.

See Varnish Processing States for a detailed graphical overview of the states and how they relate to core code functions and VCL subroutines.

Subroutines always terminate with a return() on a keyword, which determines how processing continues in the request processing state machine.

The behaviour for return() keywords is identical or at least similar across subroutines, so differences are only documented where relevant.

common return keywords¶

fail
Transition to vcl_synth on the client side as for return(synth(503, "VCL Failed")), but with any request state changes undone as if std.rollback() was called and forcing a connection close.

Intended for fatal errors, for which only minimal error handling is possible.

synth(status code, reason)
Transition to vcl_synth with resp.status and resp.reason being preset to the arguments of synth().

pass
Switch to pass mode. Control will eventually pass to vcl_pass.

pipe
Switch to pipe mode. Control will eventually pass to vcl_pipe.

restart
Restart the transaction. Increases the req.restarts counter.

If the number of restarts is higher than the max_restarts parameter, control is passed to vcl_synth as for return(synth(503, "Too many restarts"))

For a restart, all modifications to req attributes are preserved except for req.restarts and req.xid, which need to change by design.

client side¶

vcl_recv¶

Called at the beginning of a request, after the complete request has been received and parsed, after a restart or as the result of an ESI include.

Its purpose is to decide whether or not to serve the request, possibly modify it and decide on how to process it further. A backend hint may be set as a default for the backend processing side.

The vcl_recv subroutine may terminate with calling return() on one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

pass
see pass

pipe
see pipe

hash
Continue processing the object as a potential candidate for caching. Passes the control over to vcl_hash.

purge
Purge the object and it’s variants. Control passes through vcl_hash to vcl_purge.

vcl(label)
Switch to vcl labelled label. This will continue vcl processing in this vcl’s vcl_recv as if it was the active vcl.

See the varnishadm vcl.label command.

vcl_pipe¶

Called upon entering pipe mode. In this mode, the request is passed on to the backend, and any further data from both the client and backend is passed on unaltered until either end closes the connection. Basically, Varnish will degrade into a simple TCP proxy, shuffling bytes back and forth. For a connection in pipe mode, no other VCL subroutine will ever get called after vcl_pipe.

The vcl_pipe subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

pipe
Proceed with pipe mode.

vcl_pass¶

Called upon entering pass mode. In this mode, the request is passed on to the backend, and the backend’s response is passed on to the client, but is not entered into the cache. Subsequent requests submitted over the same client connection are handled normally.

The vcl_pass subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

fetch
Proceed with pass mode - initiate a backend request.

vcl_hash¶

Called after vcl_recv to create a hash value for the request. This is used as a key to look up the object in Varnish.

The vcl_hash subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

lookup
Look up the object in cache.

Control passes to vcl_purge when coming from a purge return in vcl_recv.

Otherwise control passes to the next subroutine depending on the result of the cache lookup:

a hit: pass to vcl_hit

a miss or a hit on a hit-for-miss object (an object with obj.uncacheable == true): pass to vcl_miss

a hit on a hit-for-pass object (for which pass(DURATION) had been previously returned from vcl_backend_response): pass to vcl_pass

vcl_purge¶

Called after the purge has been executed and all its variants have been evicted.

The vcl_purge subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

vcl_miss¶

Called after a cache lookup if the requested document was not found in the cache or if vcl_hit returned fetch.

Its purpose is to decide whether or not to attempt to retrieve the document from the backend. A backend hint may be set as a default for the backend processing side.

The vcl_miss subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

pass
see pass

fetch
Retrieve the requested object from the backend. Control will eventually pass to vcl_backend_fetch.

vcl_hit¶

Called when a cache lookup is successful. The object being hit may be stale: It can have a zero or negative ttl with only grace or keep time left.

The vcl_hit subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

pass
see pass

miss
Synchronously refresh the object from the backend despite the cache hit. Control will eventually pass to vcl_miss.

deliver
Deliver the object. If it is stale, a background fetch to refresh it is triggered.

vcl_deliver¶

Called before any object except a vcl_synth result is delivered to the client.

The vcl_deliver subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

synth(status code, reason)
see synth

restart
see restart

deliver
Deliver the object to the client.

vcl_synth¶

Called to deliver a synthetic object. A synthetic object is generated in VCL, not fetched from the backend. Its body may be constructed using the synthetic() function.

A vcl_synth defined object never enters the cache, contrary to a vcl_backend_error defined object, which may end up in cache.

The subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

restart
see restart

deliver
Directly deliver the object defined by vcl_synth to the client without calling vcl_deliver.

Backend Side¶

vcl_backend_fetch¶

Called before sending the backend request. In this subroutine you typically alter the request before it gets to the backend.

The vcl_backend_fetch subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

fetch
Fetch the object from the backend.

abandon
Abandon the backend request. Unless the backend request was a background fetch, control is passed to vcl_synth on the client side with resp.status preset to 503.

error(status code, reason)
Transition to vcl_backend_error with beresp.status and beresp.reason being preset to the arguments of error() if arguments are provided.

Before calling vcl_backend_fetch, Varnish core prepares the bereq backend request as follows:

Unless the request is a pass,
- set bereq.method to GET and bereq.proto to HTTP/1.1 and
- set bereq.http.Accept_Encoding to gzip if http_gzip_support is enabled.
If there is an existing cache object to be revalidated, set bereq.http.If-Modified-Since from its Last-Modified header and/or set bereq.http.If-None-Match from its Etag header
Set bereq.http.X-Varnish to the current transaction id (vxid)

These changes can be undone or modified in vcl_backend_fetch before the backend request is issued.

In particular, to cache non-GET requests, req.method needs to be saved to a header or variable in vcl_recv and restored to bereq.method. Notice that caching non-GET requests typically also requires changing the cache key in vcl_hash e.g. by also hashing the request method and/or request body.

HEAD request can be satisfied from cached GET responses.

vcl_backend_response¶

Called after the response headers have been successfully retrieved from the backend.

The vcl_backend_response subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

deliver
For a 304 response, create an updated cache object. Otherwise, fetch the object body from the backend and initiate delivery to any waiting client requests, possibly in parallel (streaming).

retry
Retry the backend transaction. Increases the retries counter. If the number of retries is higher than max_retries, control will be passed to vcl_backend_error.

abandon
Abandon the backend request. Unless the backend request was a background fetch, control is passed to vcl_synth on the client side with resp.status preset to 503.

pass(duration)
Mark the object as a hit-for-pass for the given duration. Subsequent lookups hitting this object will be turned into passed transactions, as if vcl_recv had returned pass.

error(status code, reason)
Transition to vcl_backend_error with beresp.status and beresp.reason being preset to the arguments of error() if arguments are provided.

304 handling¶

For a 304 response, varnish core code amends beresp before calling vcl_backend_response:

If the gzip status changed, Content-Encoding is unset and any Etag is weakened
Any headers not present in the 304 response are copied from the existing cache object. Content-Length is copied if present in the existing cache object and discarded otherwise.
The status gets set to 200.

beresp.was_304 marks that this conditional response processing has happened.

Note: Backend conditional requests are independent of client conditional requests, so clients may receive 304 responses no matter if a backend request was conditional.

beresp.ttl / beresp.grace / beresp.keep¶

Before calling vcl_backend_response, core code sets beresp.ttl based on the response status and the response headers Age, Cache-Control or Expires and Date as follows:

If present and valid, the value of the Age header is effectively deduced from all ttl calculations.
For status codes 200, 203, 204, 300, 301, 304, 404, 410 and 414:
- If Cache-Control contains an s-maxage or max-age field (in that order of preference), the ttl is set to the respective non-negative value or 0 if negative.
- Otherwise, if no Expires header exists, the default ttl is used.
- Otherwise, if Expires contains a time stamp before Date, the ttl is set to 0.
- Otherwise, if no Date header is present or the Date header timestamp differs from the local clock by no more than the clock_skew parameter, the ttl is set to
  - 0 if Expires denotes a past timestamp or
  - the difference between the local clock and the Expires header otherwise.
- Otherwise, the ttl is set to the difference between Expires and Date
For status codes 302 and 307, the calculation is identical except that the default ttl is not used and -1 is returned if neither Cache-Control nor Expires exists.
For all other status codes, ttl -1 is returned.

beresp.grace defaults to the default_grace parameter.

For a non-negative ttl, if Cache-Control contains a stale-while-revalidate field value, beresp.grace is set to that value if non-negative or 0 otherwise.

beresp.keep defaults to the default_keep parameter.

vcl_backend_error¶

This subroutine is called if we fail the backend fetch or if max_retries has been exceeded.

A synthetic object is generated in VCL, whose body may be constructed using the synthetic() function.

The vcl_backend_error subroutine may terminate with calling return() with one of the following keywords:

fail
see fail

deliver
Deliver and possibly cache the object defined in vcl_backend_error as if it was fetched from the backend, also referred to as a “backend synth”.

retry
Retry the backend transaction. Increases the retries counter. If the number of retries is higher than max_retries, vcl_synth on the client side is called with resp.status preset to 503.

vcl.load / vcl.discard¶

vcl_init¶

Called when VCL is loaded, before any requests pass through it. Typically used to initialize VMODs.

The vcl_init subroutine may terminate with calling return() with one of the following keywords:

ok
Normal return, VCL continues loading.

fail
Abort loading of this VCL.

vcl_fini¶

Called when VCL is discarded only after all requests have exited the VCL. Typically used to clean up VMODs.

The vcl_fini subroutine may terminate with calling return() with one of the following keywords:

ok
Normal return, VCL will be discarded.