Skip to content

Latest commit

 

History

History
1023 lines (740 loc) · 32.1 KB

File metadata and controls

1023 lines (740 loc) · 32.1 KB

vmod_dynamic

Varnish dynamic backends module

Manual section:3

SYNOPSIS

import dynamic [as name] [from "path"]

new xdirector = dynamic.director(STRING port, STRING host_header, ENUM share, PROBE probe, ACL whitelist, DURATION ttl, DURATION connect_timeout, DURATION first_byte_timeout, DURATION between_bytes_timeout, DURATION domain_usage_timeout, DURATION first_lookup_timeout, INT max_connections, INT proxy_header, BLOB resolver, ENUM ttl_from, DURATION retry_after, BACKEND via, INT keep, STRING authority, DURATION wait_timeout, INT wait_limit)

    BACKEND xdirector.backend(STRING host, STRING port, STRING authority)

    BACKEND xdirector.service(STRING service)

    VOID xdirector.debug(BOOL)

new xresolver = dynamic.resolver(BOOL set_from_os, INT parallel)

    BLOB xresolver.use()

    BOOL xresolver.set_resolution_type(ENUM)

    BOOL xresolver.clear_namespaces()

    BOOL xresolver.add_namespace(ENUM)

    BOOL xresolver.set_namespaces()

    BOOL xresolver.clear_transports()

    BOOL xresolver.add_transport(ENUM)

    BOOL xresolver.set_transports()

    BOOL xresolver.set_idle_timeout(DURATION)

    BOOL xresolver.set_limit_outstanding_queries(INT)

    BOOL xresolver.set_timeout(DURATION)

    BOOL xresolver.set_follow_redirects(ENUM)

DESCRIPTION

This module provides a varnish director for dynamic creation of backends based on calls to

  • the system's network address resolution service which, in turn, typically use information from the /etc/hosts file and the Domain Name Service (DNS), but can be configured to use other sources like LDAP (see nsswitch.conf(5)).
  • or more advanced DNS resolution where getdns is available.

While standard varnish backends defined in VCL may also be defined in terms of host names, changes of the name service information will only be picked up with a VCL reload.

In contrast, for dynamic backends provided by this module,

  • name resolution information will be refreshed by background threads after a configurable time to live (ttl) or after the ttl from DNS with a getdns dynamic.resolver().
  • resolution to multiple network addresses is supported

In addition, with a getdns dynamic.resolver(), service discovery by DNS SRV records is possible, in which case this module also allows to configure host names (targets), their ports, priority and weight though DNS. See https://en.wikipedia.org/wiki/SRV_record for a good basic explanation and xdirector.service() for details.

BACKEND SHARING

The share parameter specifies if backends are shared per director or per hostname.

  • share = "DIRECTOR"

    Default if via is not specified.

    Backends are shared per director: Only one backend per director instance and ip address (and optionally port) is created.

  • share = "HOST"

    Default if via is specified.

    Sharing of backends is limited to the same hostname (host argument to the .backend() method).

    This is the default for via for the common use case of TLS: For TLS connections, the TLS onloader usually should verify the server's certificate based on a hostname. For this purpose, the dynamic backend's name is put into the authority TLV, so, consequently, the backend needs to be defined per hostname.

PROBING

A probe to be used with dynamically created backends can be specified.

With share = HOST, the Host: header for probes defaults to the backend's hostname (host argument to the .backend() method).

With share = DIRECTOR, the probe is not specific to any particular host. If the probe has the .request attribute set, it will be used as the probe request. Otherwise, if a host_header argument was given to dynamic.director(), it will be used for the probe's Host header, otherwise Varnish-Cache core code will send the backend's IP Address as the Host header.

Consider setting the initial attribute of probes at least as high as the threshold attribute. Otherwise transactions that trigger the first lookup of a domain will see a sick backend and fail.

Irrespective of the initial attribute, transactions may still fail for backends which are actually sick. This can be mitigated using the retry transition in VCL.

TTLs from DNS

With the default system resolver, TTLs from DNS are not supported optimally. While a good combination of a ttl parameter in combination with a system name service caching service like nscd(8) can achieve good results, to use TTLs from DNS, we recommend to compile this module with getdns support, configure a dynamic.resolver() object and set the ttl_from parameter to either dns, min or max as in this example:

sub vcl_init {
        new r = dynamic.resolver();
        new d = dynamic.director(
                resolver = r.use(),
                ttl_from = dns
                );
}

See ttl_from for details.

NAMES

Directors and backends created by this vmod follow this naming scheme, which will be referred to as <name> in the following documentation

  • <directorname>

    The name of the VCL object created by dynamic.director()

  • <directorname>(<hostname>:<port>)

    A director (internally called domain), created as a result of a xdirector.backend() or, indirectly, for targets resulting from a xdirector.service() call.

  • <directorname>(service)

    A director created as a result of a xdirector.backend() or xdirector.service() call.

  • <directorname>(address:port)

    A dynamic backend for a <hostname>:<port> with DIRECTOR scope sharing

  • <directorname>(hostname.address:port)

    A dynamic backend for a <hostname>:<port> with HOST scope sharing

port may be represented symbolically (http by default)

If an authority is set, it is appended to address:portseparated by /.

STATISTICS

Dynamic backends are created and deleted on demand and can be monitored just like VCL-defined backends. Their statistics will appear in VSM-tools like varnishstat as:

VBE.<configname>.<name>.*

LOGGING

This module may log VCL_Log, Error, and Debug records following a common pattern:

vmod-dynamic %s %s %s %s [ %s ]
             |  |  |  |    |
             |  |  |  |    +- Additional information
             |  |  |  +------ Event
             |  |  +--------- <name> without the <director> part
             |  +------------ Director name
             +--------------- VCL name

Lookup timestamps are also logged to help troubleshooting, using regular Timestamp records with the following pattern for event labels:

vmod-dynamic <vcl>.<name> <Lookup|Results|Update>

When a lookup thread is terminated, either because the VCL is cooling down or the domain_usage_timeout triggered, Timestamp records are logged with the event:

vmod-dynamic <vcl>.<name> Done

Not all logs belong to HTTP transactions, especially since DNS lookups happen in the background. In order to capture all logs from this module the simplest way with varnishlog is the following:

varnishlog -g raw -q '* ~ vmod-dynamic'

It displays any individual record that contains the string vmod-dynamic whether it belongs to a transaction or not.

When a lookup fails, the backends are left untouched and the error will be logged with the following event:

getaddrinfo <errno> (<reason>)

new xdirector = dynamic.director(STRING port, STRING host_header, ENUM share, PROBE probe, ACL whitelist, DURATION ttl, DURATION connect_timeout, DURATION first_byte_timeout, DURATION between_bytes_timeout, DURATION domain_usage_timeout, DURATION first_lookup_timeout, INT max_connections, INT proxy_header, BLOB resolver, ENUM ttl_from, DURATION retry_after, BACKEND via, INT keep, STRING authority, DURATION wait_timeout, INT wait_limit)

new xdirector = dynamic.director(
   STRING port="http",
   STRING host_header=0,
   ENUM {DEFAULT, DIRECTOR, HOST} share=DEFAULT,
   PROBE probe=0,
   ACL whitelist=0,
   DURATION ttl=3600,
   DURATION connect_timeout=-1,
   DURATION first_byte_timeout=-1,
   DURATION between_bytes_timeout=-1,
   DURATION domain_usage_timeout=7200,
   DURATION first_lookup_timeout=10,
   INT max_connections=0,
   INT proxy_header=0,
   BLOB resolver=NULL,
   ENUM {cfg, dns, min, max} ttl_from=cfg,
   DURATION retry_after=30,
   BACKEND via=NULL,
   INT keep=3,
   STRING authority=NULL,
   DURATION wait_timeout=-1,
   INT wait_limit=0
)
Description

Create a DNS director.

The director creates backends with DNS lookups and chooses them in a round robin fashion. It accepts the following

Parameters:
  • port (defaults to "http")

    The port to conncet to

  • share (defaults to "DIRECTOR")

    Backend sharing scope, see BACKEND SHARING

  • host_header (defaults to none)

    host_header attribute for dynamically created backends. See also PROBING

  • probe (defaults to none)

    Probe to use. See also PROBING

  • whitelist - an acl (defaults to none)

    Only name resolution results matching the acl will be used.

  • ttl - interval between lookups (defaults to one hour)

    Minimum configured backend lifetime before address resultion is re-checked, see also ttl_from

    Related parameter: keep

  • domain_usage_timeout

    Delay until an unused domain and its backends are removed (defaults to two hours)

  • first_lookup_timeout

    Delay until the director fails lookup when a domain is requested for the first time and there is no response from the name service (defaults to ten seconds)

  • resolver

    Use a particular resolver instance created with the dynamic.resolver() constructor, if available.

    The argument to the resolver parameter must be the return value of the xresolver.use() method.

  • ttl_from

    How to determine the minimum backend lifetime before address resultion is re-checked when a resolver is also used:

    • cfg: Always use the ttl argument value (default)
    • dns: Always use the ttl from the DNS response(s), falling back to the ttl argument value
    • min: Use the minimum of the DNS response and the ttl argument value
    • max: Use the maximum of the DNS response and the ttl argument value

    If there is more than one DNS response, the minimum if taken as the DNS ttl.

    For no resolver, only "cfg" is valid.

  • retry_after

    Delay to retry lookups after DNS failures (defaults to 30 seconds).

    Notice that, for all practical purposes, this only concerns lookup errors (NXDOMAIN). After successful resolution, dynamic directors preserve existing address information if DNS resources are temporarily unavailable.

  • keep (default: 3)

    For how many subsequent updates to keep backends configured which are no longer returned by name resolution.

    Some DNS services do not return a stable set of IP addresses for consecutive queries, but rather a varying subset from a larger pool. The keep parameter is intended to keep backends configured but unused for such scenarios, in case they re-appear later.

    keep = 0 releases backends immediately as soon as they are no longer contained in a resolution result.

    Any other number specifies for how many updates (in ttl intervals) to keep backends configured while they are not contained in the resolution result.

    Until Varnish-Cache issue 3949 is addressed, it is recommended to use a keep value of at least 1.

    The keep value must not be negative and it is capped to a high positive value (UINT_MAX, usually 4 294 967 295, see limits.h(7POSIX)).

Parameters to set attributes of backends

See varnish documentation for details

  • connect_timeout (defaults to global connect_timeout)
  • first_byte_timeout (defaults to global first_byte_timeout)
  • between_bytes_timeout (defaults to global between_bytes_timeout)
  • max_connections (defaults to zero, unlimited)
  • proxy_header - version of the PROXY protocol to use, or zero to disable it (defaults to zero, valid versions are one or two)
  • via backend to make the connection through with an additional PROXY protocol header containing the actual address to connect to. Primary use case is SSL/TLS onloading via haproxy.
  • authority default for the xdirector.backend() argument. The main use case for this argument to dynamic.director() is to disable SNI by default by setting it to the empty string as authority = "".
  • wait_timeout (defaults to global backend_wait_timeout)
  • wait_limit (defaults to global backend_wait_limit)
Example
probe www_probe {
     .window = 8;
     .initial = 7;
     .threshold = 6;
     .interval = 5s;
}

acl www_acl {
     "192.168"/24;
}

sub vcl_init {
     new www_dir = dynamic.director(
             port = "80",
             probe = www_probe,
             whitelist = www_acl,
             ttl = 5m);
}

sub vcl_recv {
     set req.backend_hint = www_dir.backend("production.acme.com");
}

BACKEND xdirector.backend(STRING host, STRING port, STRING authority)

BACKEND xdirector.backend(
      STRING host="",
      STRING port="",
      STRING authority=NULL
)
Description

Return a backend from the director for a given host name and, optionally, authority. If the host name resolves to multiple addresses, they are used in a round-robin fashion.

If possible, a healhy backend is returned. If no healthy backends are known, any backend is returned. Note that it may or may not be actually responsive.

If the host is not specified, it is picked from either bereq or req.

If the port is not speficied, it is taken from the director.

If authority is set, it's value is used as the authority TLV in the PROXY header sent to the via host, which, in turn, will probably use it as the SNI for an outgoing TLS connection.

If authority is not set, it will fall back to these options in order:

  • authority from the director
  • host_header from the director
  • host argument

If authority is set to the empty string, or if via is not set, then the TLV is not sent.

BACKEND xdirector.service(STRING service)

Description

Return a backend from the director for a service name (DNS SRV record).

This Method is only supported when a dynamic.resolver() object has been passed to the dynamic.director() constructor.

SRV records contain host (called target) and port information. Dynamic backends are automatically added and maintained based on this information as if xdirector.backend() had been called.

SRV records also contain priority and weight information. The xdirector.service() method returns a target (backend) from the lowest priority found healthy. If there are multiple healthy targets for the lowest priority, one is chosen randomly based on the probabilities defined by the weight attributes.

If there are no healthy targets at all, the last target from the lowest priority is returned.

Note that it highly recommended to use a probe argument to the dynamic.director() to avoid returning unresponsive backends.

VOID xdirector.debug(BOOL)

Description
Enable or disable debugging for a dynamic director, logging background operations related to backends management.

new xresolver = dynamic.resolver(BOOL set_from_os, INT parallel)

new xresolver = dynamic.resolver(
   BOOL set_from_os=1,
   INT parallel=16
)

Create parallel getdns contexts to be used with the dynamic.director() constructor - see xresolver.use()

Parameters:

Additional configuration of the resolver contexts is possible through the methods documented below. Attempts to call these methods from outside vcl_init{} will trigger a VCL failure.

BLOB xresolver.use()

return a reference to the dynamic.resolver() object for use as a parameter to the dynamic.director() constructor.

BOOL xresolver.set_resolution_type(ENUM {RECURSING, STUB})

Specifies whether DNS queries are performed with nonrecurive lookups or as a stub resolver.

May only be called from vcl_init{}

See https://getdnsapi.net/documentation/spec/#83-contexts-for-basic-resolution

BOOL xresolver.clear_namespaces()

Clear the list of namespaces to be configured, see below.

May only be called from vcl_init{}

BOOL xresolver.add_namespace(ENUM)

BOOL xresolver.add_namespace(
      ENUM {DNS, LOCALNAMES, NETBIOS, MDNS, NIS}
)

Add a namespace to the list of namespaces to be queried.

This method only adds the namespace to an internal list, The actual configuration is only done once xresolver.set_namespaces() is called.

May only be called from vcl_init{}

See https://getdnsapi.net/documentation/spec/#83-contexts-for-basic-resolution

Notice that not all namespaces are available on all platforms. VCL load will fail with error 312 (The library did not have the requested API feature implemented.) in this case when xresolver.set_namespaces() is called.

BOOL xresolver.set_namespaces()

Apply namespace configuration, see above.

May only be called from vcl_init{}

BOOL xresolver.clear_transports()

Clear the list of transports to be configured, see below.

May only be called from vcl_init{}

BOOL xresolver.add_transport(ENUM {UDP, TCP, TLS})

Add a transport to the list of transports to be tried.

This method only adds the transport to an internal list, The actual configuration is only done once xresolver.set_transports() is called.

May only be called from vcl_init{}

See https://getdnsapi.net/documentation/spec/#83-contexts-for-basic-resolution

BOOL xresolver.set_transports()

Apply transport configuration, see above.

May only be called from vcl_init{}

BOOL xresolver.set_idle_timeout(DURATION)

Specifies the duration the API will leave an idle TCP or TLS connection open for (idle means no outstanding responses and no pending queries).

May only be called from vcl_init{}

BOOL xresolver.set_limit_outstanding_queries(INT)

May only be called from vcl_init{}

BOOL xresolver.set_timeout(DURATION)

May only be called from vcl_init{}

BOOL xresolver.set_follow_redirects(ENUM)

BOOL xresolver.set_follow_redirects(
      ENUM {REDIRECTS_FOLLOW, REDIRECTS_DO_NOT_FOLLOW}
)

May only be called from vcl_init{}

FULL EXAMPLE: BEHAVE LIKE SQUID

For illustrative purposes, here is an example to turn Varnish into a caching forward proxy for any host. This example is for http only, for https support, via support from the proxy_via_6 branch is required.

While the same functionality could be achieved without it, this example uses vmod_re for clarity:

vcl 4.1;

import dynamic;
import re;

backend proforma None;

acl ipv4_only { "0.0.0.0"/0; }

sub vcl_init {
        new http = dynamic.director(
            whitelist = ipv4_only     # remove if IPv6 is ok
            );
        new proxyurl = re.regex("^http://([^:/]+(?::(\d+))?)(/.*)");
        new hostport = re.regex("^(?:[^:]+)(?::(\d+))?");
}

sub vcl_recv {
        if (req.method == "CONNECT") {
                return (synth(400, "CONNECT is not supported"));
        }
        if (proxyurl.match(req.url)) {
                set req.url = proxyurl.backref(3, "");
                set req.http.Host = proxyurl.backref(1, "");
                set req.backend_hint =
                    http.backend(port=proxyurl.backref(2, "80"));
        } else if (hostport.match(req.http.Host)) {
                set req.backend_hint =
                    http.backend(port=hostport.backref(1, "80"));
        } else {
                return (synth(400, "URL/Host format unknown"));
        }
}

STATUS DETAILS

Status about dynamic backends can be queried using the backend.list varnish-cli(7) command. There are four variants of the output, explained as examples with a vcl named vcl and a dynamic director named dyn with a probe definition, having resolved www.****.de (some actual domain hosted by Akamai). All examples are abbreviated.

  • backend.list

    Shows terse, textual information about the director for the dynamic domain with the number of backends in the Probe column and all configured backends (which may be more than the actively used ones, see keep):

    Backend name                 Admin  Probe  Health   Last change
    vcl.dyn(www.****.de:(null))  probe  9/9    healthy  Tue, 04 Jul 2023 15:08:55 GMT
    vcl.dyn(88.221.123.83:http)  probe  8/8    healthy  Tue, 04 Jul 2023 15:08:55 GMT
    vcl.dyn(88.221.123.65:http)  probe  8/8    healthy  Tue, 04 Jul 2023 15:08:55 GMT
    vcl.dyn(88.221.123.43:http)  probe  8/8    healthy  Tue, 04 Jul 2023 15:08:55 GMT
    vcl.dyn(88.221.123.73:http)  probe  8/8    healthy  Tue, 04 Jul 2023 15:08:55 GMT
    ...
  • backend.list -p

    Shows detailed, textual information about the director for the dynamic domain with details about all active backends, followed by configured backends (which may be more than the actively used ones, see keep):

    Backend name                 Admin                     Probe    Health   Last change
    vcl.dyn(www.****.de:(null))  probe                     9/9      healthy  Tue, 04 Jul 2023 15:08:55 GMT
    
                                 Backend                   Health
                                 dyn(88.221.123.106:http)  healthy
                                 dyn(88.221.123.120:http)  healthy
                                 dyn(88.221.123.98:http)   healthy
                                 dyn(88.221.123.83:http)   healthy
                                 dyn(88.221.123.105:http)  healthy
                                 dyn(88.221.123.112:http)  healthy
                                 dyn(88.221.123.122:http)  healthy
                                 dyn(88.221.123.91:http)   healthy
                                 dyn(88.221.123.88:http)   healthy
    
    vcl.dyn(88.221.123.83:http)  probe                     8/8      healthy  Tue, 04 Jul 2023 15:08:55 GMT
     Current states  good:  8 threshold:  3 window:  8
      Average response time of good probes: 0.114404
      Oldest ================================================== Newest
      --------------------------------44444444444444444444444444444444 Good IPv4
      --------------------------------XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX Good Xmit
      --------------------------------RRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRR Good Recv
      ------------------------------HHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH Happy
    ...
  • backend.list -j

    Shows terse information in JSON format. Information about the dynamic domain does not exceed that of an ordinary backend, the probe message gives the number of healthy backends, total number of backends and the overall health status:

    [ 3, ["backend.list", "-j"], 1688484208.766,
      {
        "vcl.dyn(www.****.de:(null))": {
          "type": "dynamic",
          "admin_health": "probe",
          "probe_message": [9, 9, "healthy"],
          "last_change": 1688484005.035
        },
        "vcl.dyn(88.221.123.19:http)": {
          "type": "backend",
          "admin_health": "probe",
          "probe_message": [8, 8, "healthy"],
          "last_change": 1688484005.037
        },
        ...
  • backend.list -pj

    Shows detailed information in JSON format. The dynamic domain object contains as probe_details most properties of the dynamic director VCL object. Note that, due to Varnish-Cache API limitations, for probe and whitelist, only a boolean value can be returned. The list of backends represents the active backends:

    [ 3, ["backend.list", "-jp"], 1688484368.378,
      {
        "vcl.dyn(www.****.de:(null))": {
          "type": "dynamic",
          "admin_health": "probe",
          "probe_message": [8, 9, "healthy"],
          "probe_details": {
            "port": "http",
            "hosthdr": "",
            "share": "DIRECTOR",
            "probe": true,
            "whitelist": false,
            "connect_timeout": -1.00,
            "first_byte_timeout": -1.00,
            "between_bytes_timeout": -1.00,
            "domain_usage_timeout": 7200.00,
            "first_lookup_timeout": 10.00,
            "max_connections": 0,
            "proxy_header": 0,
            "resolver": "getaddrinfo",
            "retry_after": 30.00,
            "via": "",
            "ttl_from": "cfg",
            "ttl": 1.00,
            "keep": 600,
            "backends": {
              "dyn(88.221.123.120:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.106:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.90:http)": {
                "health": "sick"
              },
              "dyn(88.221.123.112:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.24:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.16:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.82:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.25:http)": {
                "health": "healthy"
              },
              "dyn(88.221.123.18:http)": {
                "health": "healthy"
              }
            }
          },
          "last_change": 1688484005.035
        },
        "vcl.dyn(88.221.123.19:http)": {
          "type": "backend",
          "admin_health": "probe",
          "probe_message": [8, 8, "healthy"],
          "probe_details": {
            "bits_4": 18446744073709551615,
            "bits_6": 0,
            "bits_U": 0,
            "bits_x": 0,
            "bits_X": 18446744073709551615,
            "bits_r": 0,
            "bits_R": 18446744073709551615,
            "bits_H": 18446744073709551615,
            "good": 8,
            "threshold": 3,
            "window": 8
          },
          "last_change": 1688484005.037
        },
        ...

PITFALLS

There is no support for lookups limited to IPv4 or IPv6 only. However it can be achieved by the means of a white list:

acl ipv4_only { "0.0.0.0"/0; }
acl ipv6_only { "::0"/0; }

With that you can restrict backends to the desired IP network, and monitor error logs with the whitelist mismatch event. Knowing which addresses were rejected, you can fix your domains registration (DNS records, hosts file etc).

SUPPORT

To report bugs, use github.com issues.

For enquiries about professional service and support, please contact info@uplex.de.

SEE ALSO

  • vcl(7)
  • vsl(7)
  • vsl-query(7)
  • varnish-cli(7)
  • varnish-counters(7)
  • varnishstat(1)
  • getaddrinfo(3)
  • nscd(8)
  • nsswitch.conf(5)

If you want to learn more about DNS, you can start with RFC 1034 and other RFCs that updated it over time. You may also have DNS already in place, or may be interested in setting up a name server in your infrastructure. Below is a non-exhaustive list of tools and services, but for free software name servers you can have a look at debianadmin.

DNS in the cloud (in alphabetic order):

DNS and containers (in alphabetic order):

COPYRIGHT

Copyright (c) 2015-2016 Dridi Boukelmoune
Copyright 2017-2023 UPLEX - Nils Goroll Systemoptimierung

Authors: Dridi Boukelmoune <dridi.boukelmoune@gmail.com>
         Nils Goroll <nils.goroll@uplex.de>

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
1. Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.