<?xml version="1.0" encoding="iso-8859-1"?>
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1" />
<title>Courier-IMAP: IMAP keywords implementation</title>
<meta name="generator" content="amaya 8.6, see http://www.w3.org/Amaya/" />
<style type="text/css">
.filename { font-family: courier new,courier,fixed; font-weight: bold; }
.imapflag { font-family: courier new,courier,fixed; }</style>
</head>
<body>
<h1>Courier-IMAP: IMAP keywords implementation</h1>
<p>This white paper describes how Courier-IMAP implements IMAP keywords. This
document is provided for informational purposes only.</p>
<h2>Background</h2>
<p>Courier-IMAP is a maildir-based IMAP server. The reader is presumed to
know how maildirs work.</p>
<p>There are five pre-defined flags that may be set on each message in an
IMAP folder: <span class="imapflag">\Seen</span>, <span
class="imapflag">\Answered</span>, <span class="imapflag">\Draft</span>,
<span class="imapflag">\Deleted</span>, and <span
class="imapflag">\Flagged</span>. An IMAP server may also optionally offer
the ability to set arbitrary client-defined flags for any message.</p>
<h2>Implementation Requirements</h2>
<ul>
<li>Maintain the high-performance, lock-free nature of maildir-based mail
stores.</li>
<li>The current version of Courier-IMAP offers an option to use light,
dot-lock based locking to minimize undesirable side-effects brought by
concurrent folder updates by multiple IMAP clients. Keyword usage should
not rely on locking being enabled.</li>
<li>Reading and saving keywords should be reasonably fast, even with large
folders.</li>
<li>Obtaining a list of all keywords set for a given message should be a
fast operation.</li>
<li>Obtaining a list of all messages with a given keyword set should also
be a fast operation.</li>
<li>Updating keywords should be a reasonably fast operation.</li>
<li>Should not have any noticable overhead unless keywords are actually
used.</li>
</ul>
<h2>Implementation Details</h2>
<p>The rest of this document describes the technical keyword implementation.
This is a short summary of the implement issues that should be understood
when using IMAP keywords with Courier-IMAP.</p>
<ul>
<li>On systems that impose a fixed upper limit on the maximum number of
files in a directory, the number of messages whose keywords may be
adjusted within a 15-20 minute window may not exceed 1/3rd of the upper
limit. For example, Linux ext2 filesystem directories can hold about
30,000 files, maximum. On Linux systems, no more than 10,000 messages (in
the same folder, of course) may have their keywords changed within any
15-20 minute window.</li>
<li>The atomicity is on a per-message basis. All keywords set for a
particular message are saved as an atomic unit. A client adjusts the
keywords that are set for a particular message by reading the existing
set of keywords, and then replacing them with a new set of keywords. This
means that when multiple clients update the keyword set of the same
message, the last update wins. Changes made by the losing client are
lost. Moral of the story: do not allow multiple clients to mess with the
same message, at the same time.</li>
</ul>
<h2>Data storage</h2>
<p>A new subdirectory, <span class="filename">courierimapkeywords</span>, is
created in the maildir. It stores keyword-related data.</p>
<p>The file <span class="filename">courierimapkeywords/:list</span> contains
a "stable, known list" of all keywords sets for all messages. It is,
essentially, a list of the base filenames of each message in the <span
class="filename">cur</span> directory that has keywords, without the ":2,"
suffix, and any message flags. Messages without any set keywords are not
listed in this file.</p>
<p>Additional files may also exist in this subdirectory, named either <span
class="filename">.N.file</span>, or <span class="filename">file</span>. <span
class="filename">file</span> is the base filename of a message, while "N" is
a numeric value.</p>
<p>The list of keywords set for all messages is obtained by reading the
contents of <span class="filename">courierimapkeywords</span> according to
the process described below.</p>
<h2>Updating keywords</h2>
<p>A keywords set for a message may be updated as follows:</p>
<ul>
<li>Create a file in tmp, containing the new keywords that are set for the
message. To remove all existing keywords, the file should be empty.</li>
<li>Rename the file as <span
class="filename">courierimapkeywords/file</span>, with <span
class="filename">file</span> matching the message's base filename.</li>
</ul>
<h2>Reading keywords</h2>
<p>First, a list of all messages present in <span class="filename">new</span>
and <span class="filename">cur</span> is obtained. Then:</p>
<ul>
<li>Read <span class="filename">courierimapkeywords/:list</span>. Ignore
non-existent base filenames read from the <span
class="filename">:list</span> file.</li>
<li>Divide the current time, in seconds, by 300. Call the result T.</li>
<li>Read the contents of the <span
class="filename">courierimapkeywords</span> directory. Ignore <span
class="filename">:list</span>, the remaining files in the directory will
be named either "<span class="filename">file</span>", or ".N.file" where
N is a number. When encountering a file that cannot be found in the
current list of messages present in <span class="filename">new</span> and
<span class="filename">cur</span>, stat the file, and remove it if its
ctime is at least fifteen minutes old (prevents removal of keywords for a
message that's just been added to the folder, and the scan for messages
in new and cur just missed it).</li>
<li>When encountering "<span class="filename">.N.file</span>" after another
"<span class="filename">.N.file</span>" was encountered earlier, remove
the file with the lesser N, unless the larger of the two Ns is greater
than or equals to T. Keep track of the largest N seen that's less than T,
and the largest N that's seen that's greater than or equals to T. When
encountering a "<span class="filename">file</span>", add it a list of all
"<span class="filename">file</span>"s that were encountered, and process
this entry as if it were <span class="filename">.X.file</span>, where
X=T+1.</li>
<li>After reading the entire directory, apply the following changes to the
keywords read from the <span class="filename">:list</span> file: the
contents of every <span class="filename">file</span> seen; if <span
class="filename">file</span> was not seen but a <span
class="filename">.N.file</span> was seen, then the contents of the file
with the largest N. If an attempt to open an update file failed with
ENOENT, restart everything from step 1.</li>
<li>Write the new set of keywords to a temporary file in <span
class="filename">tmp</span>, then rename it as <span
class="filename">courierimapkeywords/:list</span>. Afterwords go through
the list of all "<span class="filename">file</span>"s that were
encountered two steps ago, and rename each "<span
class="filename">file</span>" to "<span class="filename">.X.file</span>".
This step should be omitted unless at least one nonexistent file was
skipped in the old <span class="filename">:list</span> file, or the
contents of at least one <span class="filename">.N.file</span> was
updated to <span class="filename">:list</span>.</li>
<li>If exactly one <span class="filename">.N.file</span> was seen, and
N<T, remove the lone <span class="filename">.N.file</span>.</li>
</ul>
</body>
</html>
