Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

258 lines (203 sloc) 8.448 kb
<erl>
out(A) ->
{ssi, "TAB.inc", "%%",[{"internals", "choosen"}]}.
</erl>
<div id="entry">
<h1>Internals</h1>
<h2>Introduction</h2>
<p>I'll try to describe some of the internal workings of Yaws in this page.
The page is thus mostly interesting for people interested in either hacking Yaws
or simply wanting to get a better understanding.
</p>
<p>I'll describe how Yaws pages get compiled, the process structure
and other things which can make it easier to understand the code. This page
is ment to be read by programmers that wish to either work on Yaws or
just get a better understanding.
</p>
<h2> JIT Compiling a .yaws page</h2>
<p>
When the client GETs a a page that has a .yaws suffix. The Yaws server
will read that page from the hard disk and divide it in parts
that consist of HTML code and Erlang code. Each chunk of Erlang code
will be compiled into a module. The chunk of Erlang code must contain
a function <tt>out/1</tt> If it doesn't the Yaws server will insert a
proper error message into the generated HTML output.
</p>
<p>When the Yaws server ships a .yaws page it will process it chunk by chunk
through the .yaws file. If it is HTML code, the server will ship that
as is, whereas if it is Erlang code, the Yaws server will invoke the
<tt>out/1</tt> function in that code and insert the output of that <tt>out/1</tt>
function into the stream
of HTML that is being shipped to the client.
</p>
<p>Yaws will cache the result of the compilation
and the next time a client requests the same .yaws page Yaws will
be able to invoke the already compiled modules directly.
</p>
<p>This is best explained by an example:</p>
<p>Say that a file consists of 400 bytes, we have "foo.yaws"
and it looks like:</p>
<p>
<img src="compile_layout.png" />
</p>
<p>When a client request the file "foo.yaws", the webserver will
look in its cache for the file, (more on that later). For the sake of
argument, we assume the file is not in the cache.
</p>
<p>The file will be processes by the code in <tt>yaws_compile.erl</tt>
and the result will be a structure that looks like:</p>
<div class="box">
<verbatim>
[CodeSpec]
CodeSpec = Data | Code | Error
Data = {data, NumChars}
Code = {mod, LineNo, YawsFile, NumSkipChars, Mod, Func}
Err = {error, NumSkipChars, E}
</verbatim>
</div>
<p>In the particular case of our "foo.yaws" file above, the JIT
compiler will return:
</p>
<div class="box">
<verbatim>
[{mod, 1, "/foo.yaws", 100, m1, out},
{data, 200},
{mod, 30, "/foo.yaws", 100, m2, out}
]
</verbatim>
</div>
<p>
This structure gets stored in the cache and will continue
to be associated to the file "foo.yaws".
</p>
<p>When the server "ships" a .yaws page, it needs the <tt>CodeSpec</tt>
structure to do it. If the structure is not in the cache, the page
gets JIT compiled and inserted into the cache.
</p>
<p>To ship the above <tt>CodeSpec</tt> structure, the server
performs the following steps:</p>
<ol>
<li>Create the Arg structure which is a #arg{} record, this
structure is wellknown to all yaws programmers since it's the
main mechanism to pass data from the server to the .yaws
page.</li>
<li>Item (1) Invoke <tt>m1:out(Arg)</tt></li>
<li>Look at the return value from <tt>m1:out(Arg)</tt> and
perform whatever is requested. This typically involves generating
some dynamic ehtml code, generate headers or whatever.
</li>
<li>Finally jump ahead 100 bytes in the file as a result of
processing the first <tt>CodeSpec</tt> item.</li>
<li>Item (2) Next <tt>CodeSpec</tt> is just plain data from the file,
thus we read 200 bytes from the file (or rather from the cache
since the data will be there) and ship to the client.</li>
<li>Item (3) Yet another {mod structure which is handled
the same way as Item (1) above except that the erlang module
is <tt>m2</tt> instead of <tt>m1</tt></li>
</ol>
<p>Another thing that is worth mentioning is that yaws will
not ship (write on the socket) data until all content is generated.
This is questionable
and different from what i.e. PHP does. This makes it possible to
generate headers after content has been generated.
</p>
<h2>Process structure</h2>
<p>Before describing the process structure, I need to describe
the two most important datastructures in Yaws. The <tt>#gconf{}</tt>
and the <tt>#sconf{}</tt> records.
</p>
<h3>The <tt>#gconf{}</tt> record</h3>
<p>This record is used to hold all global state, i.e. state and configuration
data which is valid for all Virtual servers.
The record looks like:
</p>
<div class="box">
<verbatim>
%%global conf
record(gconf,{
yaws_dir, %% topdir of Yaws installation
trace, %% false | {true,http}|{true,traffic}
flags = ?GC_DEF, %% boolean flags
logdir,
ebin_dir = [],
runmods = [], %% runmods for entire server
keepalive_timeout = 15000,
max_num_cached_files = 400,
max_num_cached_bytes = 1000000, %% 1 MEG
max_size_cached_file = 8000,
large_file_chunk_size = 10240,
mnesia_dir = [],
log_wrap_size = 1000000, % wrap logs after 1M
cache_refresh_secs = 30, % seconds (auto zero when debug)
include_dir = [], %% list of inc dirs for .yaws files
phpexe = "php", %% cgi capable php executable
yaws, %% server string
username, %% maybe run as a different user than root
uid, %% unix uid of user that started yaws
id = "default" %% string identifying this instance of yaws
}).
</verbatim>
</div>
<p>The structure is derived from the /etc/yaws/yaws.conf file and is passed
around all through the functions in the server.
</p>
<h3> The <tt>#sconf{}</tt> record</h3>
<p>The next important datastructure is the <tt>#sconf{}</tt> record. It
is used to describe a single virtual server.
<p>Each:
</p>
<p>
<verbatim>
<server>
.....
</server>
</verbatim>
</p>
<p>In the /etc/yaws/yaws.conf file corresponds to one <tt>#sconf{}</tt>
record. We have: </p>
<div class="box">
<verbatim>
%% server conf
-record(sconf,
{port = 8000, %% which port is this server listening to
flags = ?SC_DEF,
rhost, %% forced redirect host (+ optional port)
rmethod, %% forced redirect method
docroot, %% path to the docs
listen = {127,0,0,1}, %% bind to this IP, {0,0,0,0} is possible
servername = "localhost", %% servername is what Host: header is
ets, %% local store for this server
ssl,
authdirs = [],
partial_post_size = nolimit,
appmods = [], %% list of modules for this app
errormod_404 = yaws_404, %% the default 404 error module
errormod_crash = yaws_404, %% use the same module for crashes
arg_rewrite_mod = yaws,
opaque = [], %% useful in embedded mode
start_mod, %% user provided module to be started
allowed_scripts = [yaws],
revproxy = []
}).
</verbatim>
</div>
<p>Both of these two structures are defined in "yaws.hrl"</p>
<p>Now we're ready to describe the process structure. We have:</p>
<img src="process_tree.png" />
<p>Thus, all the different "servers" defined in the configuration
file are clumped together in groups. For HTTP (i.e. not HTTPS) servers
there can be multiple virtual servers per IP address. Each group is
defined by the pair <tt>{IpAddr, Port}</tt> and they all need to
have different server names.</p>
<p>The client will send the server name in the "Host:" header and that
header is used to pick a <tt>#sconf{}</tt> record out of the list
of virtual servers for a specific <tt>{Ip,Port}</tt> pair.
</p>
<p>SSL servers are different, we cannot read the headers before we
decide which virtual server to choose because the certificate is connected
to a server name. Thus, there can only be one HTTPS server per
<tt>{Ip,Port}</tt> pair.
</div>
<erl>
out(A) -> {ssi, "END2",[],[]}.
</erl>
Jump to Line
Something went wrong with that request. Please try again.