Cyrus IMAP Server: Mailbox File Formats

Intro

This documentation refers to the "version 12" cyrus index format and associated mailbox files.

No external tools should make use of this information. The only supported method of access to the mail store is through the standard interfaces: IMAP, POP, NNTP, LMTP, etc.

A cyrus mailbox is a directory in the filesystem. It contains the following files:

  • zero or more message files

  • the cyrus.header metadata file

  • the cyrus.index metadata file

  • the cyrus.cache metadata file

  • zero or one cyrus.squat search indexes

  • zero or more subdirectories

With "split metadata" configuration, the mailbox may actually be split between multiple disks, with the files being in the same relative directory on the meta disk. See the imapd.conf option metapartition_files for more information

Message Files

The message files are named by their UID, followed by a ".", so UID 423 would be named "423.". They are stored in wire-format: lines are terminated by CRLF and binary data is not allowed.

cyrus.header

This file contains mailbox-wide information that does not change often. Its format:

<Mailbox Header Magic String>
<Quota Root>\t<Mailbox Unique ID String>\n
<Space-separated list of user flags>\n
<Mailbox ACL>\n

The Mailbox Unique ID String is used for non-owner per-user \Seen flags so they remain with the mailbox during renames, and also by the replication subsystem to detect mailbox renames.

The ACL is a copy of the value stored in mailboxes.db, and isn't actually used.

Locking Considerations

The cyrus.index file must be locked in exclusive mode while making changes to the cyrus.header file to ensure consistency. All changes are made by rewriting the entire file and renaming the new version into place.

cyrus.cache

The cyrus.cache file is a pure cache of information that's also present in the message files. It exists to make ENVELOPE and specific header fetches more efficient, as well as to assist with searches and sorts.

If a cyrus.cache file is missing or corrupted, it can be re-generated by running a reconstruct on the mailbox.

The format is 10 individual records each prefixed with a 32 bit length value in network byte order. The offset of each message's cache record is stored in the cyrus.index file (documented below). The records in a cyrus.cache file are of variable length, depending on the contents of the associated message.

The first 4 bytes of the cyrus.cache file are a "generation number" which must match the first 4 bytes of the associated cyrus.index file. In the past this was used to track consistency between the files, but the name locking scheme and per-record CRC check in cyrus 2.4 and above means this is just a backup consistency check rather than an essential format feature.

+------------------------------------------------------------------------+
|Gen # (32bits)|Size 1 (32bits)|Data 1                                   |
+------------------------------------------------------------------------+
|           |Size 2 (32bits)|Data 2            |Size 3 (32bits)| Data 3  |
+------------------------------------------------------------------------+
| .....                                                                  |
+------------------------------------------------------------------------+

While there are occasional changes to the cache format, this information is NOT stored in the cyrus.cache file. Instead, there is a "cache_version" field in the cyrus.index record, so multiple different versions of cache data may exist in the same cache file.

The order of fields per record in the cache file is as follows: (keep in mind that they are all preceded by a 4 byte network byte order size).

Envelope Response

Raw IMAP response for a request for the envelope.

Bodystructure Response

Raw IMAP response for a request for the bodystructure.

Body Response

Raw IMAP response for an (old style) request for the body.

Binary Bodystructure

Offsets into the message file to pull out various body parts. Because of the nature of MIME parts, this is somewhat recursive.

This looks like the following (starting the octet following the cache field size). All of the fields are bit32s.

[
 [Number of message parts+1 for the rfc822 header if present]
 [
  [Offset in the message file of the header of this part]
  [Size (octets) of the header of this part]
  [Offset in the message file of the content of this part]
  [Size (octets) of the content of this part]
  [Encoding Type of this part]
 ]
    (repeat for each part as well as once for the headers)
 [zero *or* number of sub-parts in the case of a multipart.
  if nonzero, this is a recursion into the top structure]
    (repeat for each part)
]

Note if this is not a message/rfc822, than the values for the sizes of the part 0 are -1 (to indicate that it doesn't exist). Sub-parts are not possible for a part 0, so they aren't included when finding recursive entries.

The offset and size info for both the mime header and content part are useful in order to do fast indexing on the appropriate parts of the message file when a client does a FETCH request for BODY[HEADER], or BODY[2.MIME].

Note that the top level RFC822 headers are a treated as a separate part from their body text ("0" or "HEADER").

In the case of a multipart/alternative, the content size & offset refers to the size of the entire mime part.

A very simple message (with a single text/plain part) would therefore look like:

[[2][rfc822 header][text/plain body part info][0]]

A simple multipart/alternative message might look like:

[[3][rfc822 header][text/plain message part info]
    [second message part info][0][0]]

A message with an attachment that has two subparts:

[[3][rfc822 header info][rfc822 first body part info][attachment info][0][
      [3][NIL header info][sub part 1 info][sub part 2 info][0][0]]]

A message with an attached message/rfc822 message with the following total structure:

message/rfc822
  0 headers; content-type: multipart/mixed
  1 text/plain
  2 message/rfc822
    0 headers; content-type: multipart/alternative
    1 text/plain
    2 text/html
[[3][rfc822 header part 0][text/plain part 1][overall attachment info][0][
     [3][rfc822 header part 2.0][text/plain part 2.1][text/html part 2.2]
        [0][0]]]
Cache Header

Any cached header fields. The exact set of fields here depends on the cache record version - there is a function in imap/mailbox.c to determine if a named header would be cached based on the version. These are in the same format they would appear in the message file:

HeaderName: headerdata\r\n

Examples include: References, In-Reply-To, etc.

From

The from header.

To

The to header.

Cc

The CC header.

Bcc

The BCC header.

Subject

The Subject header.

Locking Considerations

The cyrus.index file must be locked in exclusive mode while making changes to the cyrus.cache file to ensure consistency. All new cache records are created by reading the current end-of-file offset, appending the new cache record, and storing that start offset into the associated cyrus.index record.

cyrus.index

The cyrus.index file is NOT just a cache - it stores information not present in the message file!

The cyrus.index file consists of a fixed width header, followed by fixed width records. In the past, it would be rewritten on every expunge, but since Cyrus 2.4 the expunged records remain in the cyrus.index file for a configurable time to support QRESYNC and more efficient delayed expunge.

The cyrus.index file is the "heart" of the mailbox format - containing checksums (CRC32) of everything else, and the most frequently updated fields. All fields are stored in network byte order and aligned on 4 byte boundaries. Due to some 64 bit values being stored, the header and individual records are aligned on 8 byte boundaries.

The overall format looks sort of like this:

cyrus.index:
+----------------+
| Mailbox Header |
+----------------+
| Msg: Num 1     |
+----------------+
| Msg: Num 2     |
+----------------+
|     ...        |
+----------------+

The basic idea being that there is one header, and then all the message records are evenly spaced throughout the file. All of the message records are at well-known offsets, making any part of the file accessable at roughly equal speed.

Locking Considerations

cyrus.index files can not be repacked (i.e. records can not change UID for a particular offset, and the file can't be rewritten or deleted) unless there's an exclusive namelock held for the mailbox name. This is to avoid race conditions and simplify the use of mailboxes. Whenever a mailbox is opened, the caller holds a shared namelock on the mailbox name for the duration of the "mailbox object"'s existence.

All reads of a cyrus.index file must be done with a lock held, and all writes must be done with an exclusive lock held. This ensures CRC32 checksums of individual headers and records are always consistent. There are no direct "offset" reads done any more, instead the mailbox API provides a way to read an entire cyrus.index header or cyrus.index record into a struct, performing consistency checks. Writes are also done with a complete record struct.

Detail of cyrus.index header

The index header contains the following information, in order:

Generation Number (4 bytes)

A number that is basically the "revision number" of the mailbox. It must match between the cache and index files. This is to ensure that if we fail to sync both the cache and index files and a crash happens (so that only one is synced), we do not provide bad data to the user. This is also backed by having individual cache checksums on each record.

Format (4 bytes)

Basically obsolete (indicates netnews or regular).

Minor Version (4 bytes)

Indicates the version number of the index file. This can be used for on-the-fly upgrades of the index and cache files.

Start Offset (4 bytes)

Size of index header.

Record Size (4 bytes)

Size of an index record.

Num Records (4 bytes)

How many records are in this index (including records for expunged records. See below for "Exists" which has moved from pre-version 12 files.

Last Appenddate (4 bytes)

(time_t) of the last time a message was appended

Last UID (4 bytes)

Highest UID of all messages in the mailbox (UIDNEXT - 1).

Quota Mailbox Used (8 bytes)

Total amount of storage used by all of the messages in the mailbox. Platforms that don't support 64-bit integers only use the last 4 bytes.

POP3 Last Login (4 bytes)

(time_t) of the last pop3 login to this INBOX, used to enforce the "poptimeout" imapd.conf option.

UIDvalidity (4 bytes)

The UID validitiy of this mailbox. Cyrus currently uses the time() when this mailbox was created.

Deleted, Answered, and Flagged (4 bytes each)

Counts of how many messages have each flag.

Mailbox Options (4 bytes)

Bitmask of mailbox options, consisting of any combination of the following:

POP3_NEW_UIDL

Flag signalling that we're using "uidvalidity.uid" instead of just "uid" for the output of the POP3 UIDL command.

IMAP_SHAREDSEEN

Flag signalling that we're supporting a shared \Seen flag on the mailbox.

IMAP_DUPDELIVER

Flag signalling that we're allowing duplicate delivery of messages to the mailbox, overriding system-wide duplicate suppression.

MAILBOX_NEEDS_REPACK

Flag signalling that the mailbox is due to be repacked. During mailbox_close() every process will attempt to take an exclusive namelock on the mailbox and repack.

MAILBOX_DELETED

Flag signalling that the mailbox is deleted. This can be set with a shared namelock, and indicates to all other users of the mailbox that they need to close it and attempt cleanup. The last process to close the mailbox will perform the final cleanup under an exclusive namelock, giving the other processes a chance to finish their current operation first without files disappearing from under them!

Leaked Cache (4 bytes)

Number of leaked records in the cache file.

Highest ModSeq (8 bytes)

Highest Modification Sequence of all the messages in the mailbox (CONDSTORE).

Deleted ModSeq (8 bytes)

Lowest Modification Sequence before which expunged message data may have been purged from the mailbox and forgotten (CONDSTORE/QRESYNC support).

Exists (4 bytes)

See NumRecords above. This is the count of non-expunged records in the mailbox and corresponds to the IMAP status item "EXISTS".

First Expunged (4 bytes)

lowest modified time of an expunged message in this mailbox (or zero if there are no expunged messages) - used to determine if the mailbox needs repacking.

Last Repack Time (4 bytes)

a timestamp for the last repack, to ensure repacks aren't done too close together if expunges were closely spaced

Header File CRC (4 bytes)

CRC32 value of the bytes in the cyrus.header file for this mailbox. Must be rewritten whenever the cyrus.header file is changed (see locking considerations above - this is why the cyrus.index must be exclusively locked!)

Sync CRC (4 bytes)

An XOR of the CRC32 of a specially generated value for each of the non-expunged records in this mailbox. This is a cached value which allows the replication subsystem to quickly determine that all non-expunged records in a mailbox are in sync and detect possible "split brain" scenarios with low bandwidth use.

Recent UID (4 bytes)

The highest UID last time an IMAP client logged in as the mailbox owner (or anybody if SHAREDSEEN is enabled) selected this mailbox. Used to generate the \Recent flags in IMAP

Recent Time (4 bytes)

Used for consistency with the seen_db code, but probably not actually necessary. Oh well

Header CRC (4 bytes)

Must always be the LAST record of the header. This is the CRC32 of the actual bytes on disk (network order format) for the rest of the cyrus.index. By keeping it last, it can be easily calculated with the following snippet of code: crc = crc32_map(buf, OFFSET_HEADER_CRC); - i.e. crc32 from the start of the buffer to just before this field.

There are also spare fields in the index header, to allow for future expansion without forcing an upgrade of the file, and to round up to be divisible by 8 bytes.

Detail of cyrus.index records

These records start immediately following the cyrus.index header, and are all fixed size. They are in-order by uid of the message.

UID (4 bytes)

UID of the message

INTERNALDATE (4 bytes)

INTERNALDATE of the message (where possible, this matches the creation and modification times of the file on disk to help reconstruct in the event of data loss)

SENTDATE (4 bytes)

Contents of the Date: header chomped to day resolution with timezone stripped.

SIZE (4 bytes)

Size of the whole message (in octets)

HEADER SIZE (4 bytes)

Size of the message header (in octets)

GMTIME (4 bytes)

Contents of the Date: header at 1 second resolution and converted to GMT (for sort)

CACHE_OFFSET (4 bytes)

Offset into the cyrus.cache file for the beginning of this message's cache entry.

LAST UPDATED (4 bytes)

(time_t) of the last time this record was changed

SYSTEM FLAGS (4 bytes)

Bitmask showing which system flags are set/unset

USER FLAGS (MAX_USER_FLAGS / 32 bytes)

Bitmask showing which user flags are set/unset (bits correspond to positions in the cyrus.header flag list, i.e. (1<<0) == the flag name

CONTENT_LINES (4 bytes)

Number of text lines contained in the message content (body).

CACHE_VERSION (4 bytes)

Indicates the version number of the cache record for the message (determines which headers are cached, see list in mailbox.c).

GUID (MESSAGE_GUID_SIZE bytes)

Globally Unique IDentifier of the message (used by replication engine). This is the sha1 value of the bytes as stored on disk.

MODSEQ (8 bytes)

Modification Sequence of the message (CONDSTORE).

CACHE_CRC (4 bytes)

This is the CRC32 of all the bytes of the cache record (all 10 fields) as stored on disk. Again, calculated over the exact bytes stored in the cyrus.cache file.

RECORD_CRC (4 bytes)

Like the header CRC - this is the CRC32 of all the bytes in on-disk order that exist in this record. Records are always rewritten as the entire record, including the updated CRC, so it's always consistent if you have a lock on the cyrus.index file, because writers will wait until they get an exclusive lock to make modifications.

Notes

  • Expunge is super quick now - it's just a flag update!

  • Append is relatively fast (it only adds to the end of both the cache and index files and modifies the index header)

  • Message unlinks always happen during the "close" phase - which may be noticed when you select another mailbox, but otherwise are delayed from the actual action. With delayed expunge, the unlinks are pushed off to cyr_expire which is a background task, and will never be noticed by the user.

  • Message delivery is something like this:

    1. write/sync message file

    2. write/sync new cyrus.cache record

    3. write/sync new cyrus.index record

    4. calculate, write, sync new cyrus.index header

    5. acknowledge message delivery

    The message isn't delivered until the new index header is written. In case of a crash before the new index header is written, any previous writes will be overwritten on the next delivery (and will not be noticed by the readers).

    Note that certain power failure situations (power failure in the middle of a disk sector write) could cause a mailbox to need reconstruction (possibly even losing some flag state). These failure modes are not possible in the "Hardware RAID disk model" (which we will describe somewhere else when we get around to it).

Future considerations

  • Cache all header fields? (or all up to Xk?) This could greatly improve speeds of clients that just ask for everything, but also increases the expense of rewriting the cache file (as well as the size it takes on disk).

  • Reformat cache file to use a (size)(size)(size)(size)(data)(data)(data) format. This makes accesses anywhere in the cache file equally fast, as opposed to having to iterate through all the entires for a given message to get to the last one. Note that either way is still O(1) so maybe it doesn't matter much.

  • It would be useful to store a uniqueid -> mailbox name index, so that we could fix arbitron again.