DOKK / manpages / debian 12 / varnish / vsl.7.en
VSL(7) VSL(7)

VSL - Varnish Shared Memory Logging

This document describes the format and content of all the Varnish shared memory logging tags. These tags are used by the varnishlog(1), varnishtop(1), etc. logging tools supplied with Varnish.

Logged when a connection is selected for handling a backend request.

The format is:

%d %s %s
|  |  |
|  |  +- Backend display name
|  +---- VCL name
+------- Connection file descriptor


NOTE: This tag is currently not in use in the Varnish log. It is mentioned here to document legacy versions of the log, or reserved for possible use in future versions.

Logged when a backend connection is closed.

The format is:

%d %s %s [ %s ]
|  |  |    |
|  |  |    +- Optional reason
|  |  +------ "close" or "recycle"
|  +--------- Backend display name
+------------ Connection file descriptor


Logged when a new backend connection is opened.

The format is:

%d %s %s %s %s %s %s
|  |  |  |  |  |  |
|  |  |  |  |  |  +- "connect" or "reuse"
|  |  |  |  |  +---- Local port
|  |  |  |  +------- Local address
|  |  |  +---------- Remote port
|  |  +------------- Remote address
|  +---------------- Backend display name
+------------------- Connection file descriptor


Logged when a backend connection is put up for reuse by a later connection.

The format is:

%d %s
|  |
|  +- Backend display name
+---- Connection file descriptor


NOTE: This tag is currently not in use in the Varnish log. It is mentioned here to document legacy versions of the log, or reserved for possible use in future versions.

Start of backend processing. Logs the backend IP address and port number.

The format is:

%s %s
|  |
|  +- Backend Port number
+---- Backend IP4/6 address


NOTE: This tag is currently not in use in the Varnish log. It is mentioned here to document legacy versions of the log, or reserved for possible use in future versions.

The result of a backend health probe.

The format is:

%s %s %s %s %u %u %u %f %f %s
|  |  |  |  |  |  |  |  |  |
|  |  |  |  |  |  |  |  |  +- Probe HTTP response / error information
|  |  |  |  |  |  |  |  +---- Average response time
|  |  |  |  |  |  |  +------- Response time
|  |  |  |  |  |  +---------- Probe window size
|  |  |  |  |  +------------- Probe threshold level
|  |  |  |  +---------------- Number of good probes in window
|  |  |  +------------------- Probe window bits
|  |  +---------------------- "healthy" or "sick"
|  +------------------------- "Back", "Still" or "Went"
+---------------------------- Backend name


Probe window bits are:

'-': Could not connect
'4': Good IPv4
'6': Good IPv6
'U': Good UNIX
'x': Error Xmit
'X': Good Xmit
'r': Error Recv
'R': Good Recv
'H': Happy


When the backend is just created, the window bits for health check slots that haven't run yet appear as '-' like failures to connect.

The first record of a VXID transaction.

The format is:

%s %d %s
|  |  |
|  |  +- Reason
|  +---- Parent vxid
+------- Type ("sess", "req" or "bereq")


Contains byte counters from backend request processing.

The format is:

%d %d %d %d %d %d
|  |  |  |  |  |
|  |  |  |  |  +- Total bytes received
|  |  |  |  +---- Body bytes received
|  |  |  +------- Header bytes received
|  |  +---------- Total bytes transmitted
|  +------------- Body bytes transmitted
+---------------- Header bytes transmitted


HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

The HTTP request method used.
The HTTP protocol version information.
The HTTP request URL.
HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

The HTTP protocol version information.
The HTTP response reason string.
The HTTP response status code.
HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

Contains the first 20 characters of received HTTP headers we could not make sense of. Applies to both req.http and beresp.http.
CLI communication between varnishd master and child process.
Debug messages can normally be ignored, but are sometimes helpful during trouble-shooting. Most debug messages must be explicitly enabled with parameters.

Debug messages may be added, changed or removed without prior notice and shouldn't be considered stable.

NB: This log record is masked by default.

An error or warning was generated during parsing of an ESI object. The log record describes the problem encountered.
The last record of a VXID transaction.
Error - Error messages
Error messages are stuff you probably want to know.
Logs the VXID when an object is banned.
Logs events related to object expiry. The events are:
Logged when the expiry time of an object changes.
Logged when the expiry thread picks an object from the inbox for processing.
Logged when the expiry thread kills an object from the inbox.
Logged when the expiry thread moves an object on the binheap.
Logged when the expiry thread expires an object.
Logged when an object is evaluated for LRU force expiry.
Logged when an object is force expired due to LRU.
Logged when no suitable candidate object is found for LRU force expiry.

The format is:

EXP_Rearm p=%p E=%f e=%f f=0x%x
EXP_Inbox p=%p e=%f f=0x%x
EXP_Kill p=%p e=%f f=0x%x
EXP_When p=%p e=%f f=0x%x
EXP_Expired x=%u t=%f
LRU_Cand p=%p f=0x%x r=%d
LRU x=%u
LRU_Fail
Legend:
p=%p         Objcore pointer
t=%f         Remaining TTL (s)
e=%f         Expiry time (unix epoch)
E=%f         Old expiry time (unix epoch)
f=0x%x       Objcore flags
r=%d         Objcore refcount
x=%u         Object VXID


Logs the error message of a failed fetch operation.

Error messages should be self-explanatory, yet the http connection (HTC) class of errors is reported with these symbols:

  • junk (-5): Received unexpected data
  • close (-4): Connection closed
  • timeout (-3): Timed out
  • overflow (-2): Buffer/workspace too small
  • eof (-1): Unexpected end of input
  • empty (0): Empty response
  • more (1): More data required
  • complete (2): Data complete (no error)
  • idle (3): Connection was closed while idle



Notice that some HTC errors are never emitted.

Ready to fetch body from backend.

The format is:

%d %s %s
|  |  |
|  |  +---- 'stream' or '-'
|  +------- Text description of body fetch mode
+---------- Body fetch mode


List of filters applied to the body
A Gzip record is emitted for each instance of gzip or gunzip work performed. Worst case, an ESI transaction stored in gzip'ed objects but delivered gunziped, will run into many of these.

The format is:

%c %c %c %d %d %d %d %d
|  |  |  |  |  |  |  |
|  |  |  |  |  |  |  +- Bit length of compressed data
|  |  |  |  |  |  +---- Bit location of 'last' bit
|  |  |  |  |  +------- Bit location of first deflate block
|  |  |  |  +---------- Bytes output
|  |  |  +------------- Bytes input
|  |  +---------------- 'E': ESI, '-': Plain object
|  +------------------- 'F': Fetch, 'D': Deliver
+---------------------- 'G': Gzip, 'U': Gunzip, 'u': Gunzip-test


Examples:

U F E 182 159 80 80 1392
G F E 159 173 80 1304 1314


Binary data
Binary data
Binary data
Binary data
This value was added to the object lookup hash.

NB: This log record is masked by default.

Object looked up in cache.

The format is:

%u %f %f %f [%u [%u]]
|  |  |  |   |   |
|  |  |  |   |   +- Content length
|  |  |  |   +----- Fetched so far
|  |  |  +--------- Keep period
|  |  +------------ Grace period
|  +--------------- Remaining TTL
+------------------ VXID of the object


Hit-for-miss object looked up in cache.

The format is:

%u %f
|  |
|  +- Remaining TTL
+---- VXID of the object


Hit-for-pass object looked up in cache.

The format is:

%u %f
|  |
|  +- Remaining TTL
+---- VXID of the object


Logs the content of unparseable HTTP requests.
Logs the size of a fetch object body.
Links this VXID to any child VXID it initiates.

The format is:

%s %d %s
|  |  |
|  |  +- Reason
|  +---- Child vxid
+------- Child type ("req" or "bereq")


Logs the header name of a failed HTTP header operation due to resource exhaustion or configured limits.
Informational log messages on events occurred during request handling.

The format is:

%s: %s
|   |
|   +- Short description of the notice message
+----- Manual page containing the detailed description


See the NOTICE MESSAGES section below or the individual VMOD manual pages for detailed information of notice messages.

HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

The HTTP protocol version information.
The HTTP response reason string.
The HTTP response status code.
HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

Contains byte counters for pipe sessions.

The format is:

%d %d %d %d
|  |  |  |
|  |  |  +------- Piped bytes to client
|  |  +---------- Piped bytes from client
|  +------------- Backend request headers
+---------------- Client request headers


PROXY protocol information.

The format is:

%d %s %d %s %d
|  |  |  |  |
|  |  |  |  +- server port
|  |  |  +---- server ip
|  |  +------- client port
|  +---------- client ip
+------------- PROXY protocol version
All fields are "local" for PROXY local connections (command 0x0)


A PROXY protocol header was unparseable.
Contains byte counts for the request handling. The body bytes count includes transmission overhead (ie: chunked encoding). ESI sub-requests show the body bytes this ESI fragment including any subfragments contributed to the top level request. The format is:

%d %d %d %d %d %d
|  |  |  |  |  |
|  |  |  |  |  +- Total bytes transmitted
|  |  |  |  +---- Body bytes transmitted
|  |  |  +------- Header bytes transmitted
|  |  +---------- Total bytes received
|  +------------- Body bytes received
+---------------- Header bytes received


HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

The HTTP request method used.
The HTTP protocol version information.
Start of request processing. Logs the client address, port number and listener endpoint name (from the -a command-line argument).

The format is:

%s %s %s
|  |  |
|  |  +-- Listener name (from -a)
|  +----- Client Port number (0 for Unix domain sockets)
+-------- Client IP4/6 address (0.0.0.0 for UDS)


The HTTP request URL.
HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

The HTTP protocol version information.
The HTTP response reason string.
The HTTP response status code.
HTTP header contents.

The format is:

%s: %s
|   |
|   +- Header value
+----- Header name


NOTE: HTTP header fields are free form records and not strictly made of 2 fields. Accessing a specific header with the prefix notation helps treating the header value as a single string.

SessClose is the last record for any client connection.

The format is:

%s %f
|  |
|  +- How long the session was open
+---- Why the connection closed


Accepting a client connection has failed.

The format is:

%s %s %s %d %d %s
|  |  |  |  |  |
|  |  |  |  |  +- Detailed error message
|  |  |  |  +---- Error Number (errno) from accept(2)
|  |  |  +------- File descriptor number
|  |  +---------- Local TCP port / 0 for UDS
|  +------------- Local IPv4/6 address / 0.0.0.0 for UDS
+---------------- Socket name (from -a argument)


The first record for a client connection, with the socket-endpoints of the connection.

The format is:

%s %d %s %s %s %f %d
|  |  |  |  |  |  |
|  |  |  |  |  |  +- File descriptor number
|  |  |  |  |  +---- Session start time (unix epoch)
|  |  |  |  +------- Local TCP port / 0 for UDS
|  |  |  +---------- Local IPv4/6 address / 0.0.0.0 for UDS
|  |  +------------- Socket name (from -a argument)
|  +---------------- Remote TCP port / 0 for UDS
+------------------- Remote IPv4/6 address / 0.0.0.0 for UDS


Type and name of the storage backend the object is stored in.

The format is:

%s %s
|  |
|  +- Name of storage backend
+---- Type ("malloc", "file", "persistent" etc.)


A TTL record is emitted whenever the ttl, grace or keep values for an object is set as well as whether the object is cacheable or not.

The format is:

%s %d %d %d %d [ %d %d %u %u ] %s
|  |  |  |  |    |  |  |  |    |
|  |  |  |  |    |  |  |  |    +- "cacheable" or "uncacheable"
|  |  |  |  |    |  |  |  +------ Max-Age from Cache-Control header
|  |  |  |  |    |  |  +--------- Expires header
|  |  |  |  |    |  +------------ Date header
|  |  |  |  |    +--------------- Age (incl Age: header value)
|  |  |  |  +-------------------- Reference time for TTL
|  |  |  +----------------------- Keep
|  |  +-------------------------- Grace
|  +----------------------------- TTL
+-------------------------------- "RFC", "VCL" or "HFP"


The four optional fields are only present in "RFC" headers.

Examples:

RFC 60 10 -1 1312966109 1312966109 1312966109 0 60 cacheable
VCL 120 10 0 1312966111 uncacheable
HFP 2 0 0 1312966113 uncacheable


Contains timing information for the Varnish worker threads.

Time stamps are issued by Varnish on certain events, and show the absolute time of the event, the time spent since the start of the work unit, and the time spent since the last timestamp was logged. See the TIMESTAMPS section below for information about the individual time stamps.

The format is:

%s: %f %f %f
|   |  |  |
|   |  |  +- Time since last timestamp
|   |  +---- Time since start of work unit
|   +------- Absolute time of event
+----------- Event label


Logs error messages generated during VCL execution.
User generated log messages insert from VCL through std.log()
ACLs with the +log flag emits this record with the result.

The format is:

%s %s [%s [fixed: %s]]
|  |   |          |
|  |   |          +- Fixed entry (see vcc_acl_pedantic parameter)
|  |   +------------ Matching entry (only for MATCH)
|  +---------------- Name of the ACL
+-------------------- MATCH or NO_MATCH


MATCH denotes an ACL match NO_MATCH denotes that a checked ACL has not matched

Logs the VCL method name when a VCL method is called.
Logs the VCL method terminating statement.
Logs VCL execution trace data.

The format is:

%s %u %u.%u.%u
|  |  |  |  |
|  |  |  |  +- VCL program line position
|  |  |  +---- VCL program line number
|  |  +------- VCL program source index
|  +---------- VCL trace point index
+------------- VCL configname


NB: This log record is masked by default.

Records the name of the VCL being used.

The format is:

%s [ %s %s ]
|    |  |
|    |  +- Name of label used to find it
|    +---- "via"
+--------- Name of VCL put in use


Warnings and error messages generated by the VSL API while reading the shared memory log.
Contains name of VDP and statistics.

The format is:

%s %d %d
|  |  |
|  |  +- Total bytes produced
|  +---- Number of calls made
+------- Name of filter


NB: This log record is masked by default.

Contains name of VFP and statistics.

The format is:

%s %d %d
|  |  |
|  |  +- Total bytes produced
|  +---- Number of calls made
+------- Name of filter


NB: This log record is masked by default.

Diagnostic recording of locking order.
Logs worker thread creation and termination events.

The format is:

%p %s
|  |
|  +- [start|end]
+---- Worker struct pointer


NB: This log record is masked by default.


Timestamps are inserted in the log on completing certain events during the worker thread's task handling. The timestamps has a label showing which event was completed. The reported fields show the absolute time of the event, the time spent since the start of the task and the time spent since the last timestamp was logged.

The timestamps logged automatically by Varnish are inserted after completing events that are expected to have delays (e.g. network IO or spending time on a waitinglist). Timestamps can also be inserted from VCL using the std.timestamp() function. If one is doing time consuming tasks in the VCL configuration, it's a good idea to log a timestamp after completing that task. This keeps the timing information in subsequent timestamps from including the time spent on the VCL event.

The start of request processing (first byte received or restart).
Complete client request received.
Client request body processed (discarded, cached or passed to the backend).
Came off waitinglist.
Fetch processing finished (completely fetched or ready for streaming).
Processing finished, ready to deliver the client response.
Delivery of response to the client finished.
Client request is being restarted.

The following timestamps are client timestamps specific to pipe transactions:

Opened a pipe to the backend and forwarded the request.
The pipe session has finished.

The following timestamps change meaning in a pipe transaction:

Processing finished, ready to begin the pipe delivery.

Start of the backend fetch processing.
Came off vcl_backend_fetch ready to send the backend request.
Successfully established a backend connection, or selected a recycled connection from the pool.
Backend request sent.
Backend response headers received.
Processing finished, ready to fetch the response body.
Backend response body received.
Backend request is being retried.
Backend request failed to vcl_backend_error.

Notice messages contain informational messages about the handling of a request. These can be exceptional circumstances encountered that causes deviation from the normal handling. The messages are prefixed with vsl: for core Varnish generated messages, and VMOD authors are encouraged to use vmod_<name>: for their own notice messages. This matches the name of the manual page where detailed descriptions of notice messages are expected.

The core messages are described below.

The backend answered 304 Not Modified on a conditional fetch using an object that has not yet been fully fetched as the stale template object. This can only happen when the TTL of the object is less than the time it takes to fetch it. The fetch is halted until the stale object is fully fetched, upon which the new object is created as normal. While waiting, any grace time on the stale object will be in effect.

Objects are primarily looked up from an index using the hash key computed from the hash_data() VCL function. When variants are involved, that is to say when a backend response was stored with a Vary header, a secondary lookup is performed but it is not indexed. As the number of variants for a given key increases, this can slow cache lookups down, and since this happens under a lock, this can greatly increase lock contention, even more so for frequently requested objects. Variants should therefore be used sparingly on cacheable responses, but since they can originate from either VCL or origin servers, this notice should help identify problematic resources.

This document was initially written by Poul-Henning Kamp, and later updated by Martin Blix Grydeland.


This document is licensed under the same licence as Varnish itself. See LICENCE for details.

  • Copyright (c) 2006 Verdens Gang AS
  • Copyright (c) 2006-2015 Varnish Software AS