kmail

KMail architectural overview

design principles

This file is intended to guide the reader's way through the KMail codebase. It should esp. be handy for people not hacking full-time on KMail as well as people that want to trace bugs in parts of KMail which they don't know well.

Contents:

TODO: reader, composer, messages, accounts, ...

KERNEL

Files: kmkernel.h, kmkernel.cpp

Contact Zack Rusin <zack@kde.org> with questions...

The first thing you'll notice about KMail is the extensive use of kmkernel->xxx() constructs. The "kmkernel" is a define in kmkernel.h declared as : define kmkernel KMKernel::self() KMKernel is the central object in KMail. It's always created before any other class, therefore you are _guaranteed_ that KMKernel::self() (and therefore "kmkernel" construct) won't return 0 (null).

KMKernel implements the KMailIface (our DCOP interface) and gives access to all the core KMail functionality.

IDENTITY

FIXME this has moved to libkpimidentities, right?

Files: identity*, kmidentity.{h,cpp}, configuredialog.cpp, signatureconfigurator.{h,cpp}

Contact Marc Mutz <mutz@kde.org> on questions...

Identities consists of various fields represented by QStrings. Currently, those fields are hardcoded, but feel free to implement KPIM::Identity as a map from strings to QVariants or somesuch.

One part of identities are signatures. They can represent four modes (Signature::Type) of operation (disabled, text from file or command and inline text), which correspond to the combo box in the identitydialog.

Identities are designed to be used through the KPIM::IdentityManager: const KPIM::Identity & ident = kmkernel->identityManager()->identityForUoidOrDefault(...) Make sure you assign to a _const_ reference, since the identityForFoo methods are overloaded with non-const methods that access a different list of identities in the manager that is used while configuring. That is known source of errors when you use identityForFoo() as a parameter to a method taking const KPIM::Identity &.

WARNING: Don't hold the reference longer than the current functions scope or next return to the event loop. That's b/c the config dialog is not modal and the user may hit apply/ok anytime between calls to function that want to use the identity reference. Store the UOID instead if you need to keep track of the identity. You may also want to connect to one of the KPIM::IdentityManager::changed() or deleted() signals, if you want to do special processing in case the identity changes.

Thus, in the ConfigureDialog, you will see non-const KPIM::Identity references being used, while everywhere else (KMMessage, IdentityCombo) const references are used.

The KPIM::IdentityCombo is what you see in the composer. It's a self-updating combo box of KPIM::Identity's. Use this if you want the user to choose an identity, e.g. in the folder dialog.

Ihe IdentityListView is what you see in the config dialog's identity management page. It's not meant to be used elsewhere, but is DnD enabled (well, at the time of this writing, only drag-enabled). This is going to be used to dnd identities around between KNode and KMail, e.g.

The SignatureConfigurator is the third tab in the identity dialog. It's separate since it is used by the identity manager to request a new file/command if the current value somehow fails.

FILTER

Contact Marc Mutz <mutz@kde.org> on questions...

Filters consist of a search pattern and a list of actions plus a few flags to indicate when they are to be applied (kmfilter.h). They are managed in a QPtrList<KMFilter>, called KMFilterMgr. This filter magnager is responsible for loading and storing filters (read/writeConfig) and for executing them (process). The unique instance of the filter manager is held by the kernel (KMKernel::filterMgr()).

The search pattern is a QPtrList of search rules (kmsearchpattern.h) and a boolean operator that defines their relation (and/or).

A search rule consists of a field-QString, a "function"-enum and a "contents" or "value" QString. The first gives the header (or pseudoheader) to match against, the second says how to match (equals, consists, is less than,...) and the third holds the pattern to match against. Currently, there are two types of search rules, which are mixed together into a single class: String-valued and int-valued. The latter is a hack to enable

<size>
and
<age in days>
pseudo-header matching. KMSearchRules should better be organized like KMFilterActions are.

A filter action (kmfilteraction.h) inherits from KMFilterAction or one of it's convenience sub-classes. They have three sub-interfaces: (1) argument handling, (2) processing and (3) parameter widget handling. Interface (1) consists of args{From,As}String(), name() and isEmpty() and is used to read and write the arguments (if any) from/to the config. Interface (2) is used by the filter manager to execute the action (process() / ReturnCode). Interface (3) is used by the filter dialog to allow editing of actions and consists of name(), label() and the ParamWidget*(). Complex parameter widgets are collected in kmfawidget.{h,cpp}.

A typical call for applying filters is

KMKernel::filterMgr() foreach message { KMFilterMgr::process(): }

CONFIGURE DIALOG

Files: configuredialog*.{h,cpp} ( identitylistview.{h,cpp} )

Contact Marc Mutz <mutz@kde.org> on questions...

The configuredialog is made up of pages that in turn may consist of a number of tabs. The genral rule of thumb is that each page and tab is responsible for reading and writing the config options presented on it, although in the future, that may be reduced to interacting with the corresponding config manager instead. But that won't change the basic principle.

Thus, there is an abstract base class ConfigurePage (defined in configuredialog_p.h), which derives from QWidget. It has four methods of which you have to reimplement at least the first two:

  • void setup() Re-read the config (from the config file or the manager) and update the widgets correspondingly. That is, you should only create the widgets in the ctor, not set the options yet. The reason for that is that the config dialog, once created, is simply hidden and shown subsequently, so we need a reset-like method anyway.

  • void apply() Read the config from the widgets and write it into the config file or the corresponding config manager.

  • void installProfile() This is called when the user selected a profile and hit apply. A profile is just another KConfig object. Therefore, this method should be the same as setup(), except that you should only alter widgets for configs that really exist in the profile.

For tabbed config pages, there exists a convenience class called TabbedConfigurationPage, which (as of this writing only offers the addTab() convenience method. It is planned to also provide reimplementations of setup, dismiss, apply and installProfile that just call the same functions for each tab.

MDNs

Files: libkdenetwork/kmime_mdn.{h,cpp} and kmmessage.{h,cpp}, mostly

Contact Marc Mutz <mutz@kde.org> on questions...

MDNs (Message Disposition Notifications; RFC 2298) are a way to send back information regarding received messages back to their sender. Examples include "message read/deleted/forwarded/processed".

The code in kmime_mdn.{h,cpp} is responsible for creating the message/disposition-notification body part (2nd child of multipart/report that makes the MDN) and for providing the template for human-readable text that goes into the text/plain part (1st child of the multipart/report).

The code in KMMessage::createMDN() actually constructs a message containing a MDN for this message, using the kmime_mdn helper functions. It starts by checking the index for an already sent MDN, since the RFC demands that MDNs be sent only once for every message. If that test succeeds, it goes on to check various other constraints as per RFC and if all goes well the message containing the multipart/report is created.

If you need to use this functionality, see KMReaderWin::touchMsg() and KMFilterAction::sendMDN() for examples. The touchMsg() code is invoked on display of a message and sends a "displayed" MDN back (if so configured), whereas the KMFilterAction method is a convenience helper for the various filter actions that can provoke a MDN (move to trash, redirect, forward, ...).

Folders

Files: kmfolder*.{h,cpp}, folderstorage.{h,cpp} and *job.{h,cpp}

Contact Zack Rusin <zack@kde.org> with questions...

The collaboration among KMail folder classes looks as follows :

KMFolderNode / \ / \ KMFolderDir \ KMFolder v FolderStorage | | KMFolderIndex | | ---< actual folder types: KMFolderImap, KMFolderMbox... >--

At the base KMail's folder design starts with KMFolderNode which inherits QObject. KMFolderNode is the base class encapsulating common folder properties such as the name and a boolean signifying whether the folder is a folder holding mail directly or a KMFolderDir. KMFolderNode's often do not have an on-disk representation, they are entities existing only within KMail's design.

KMFolder acts as the runtime representation of a folder with the physical storage part being represented by a member of type FolderStorage. KMFolder and FolderStorage have many functions with the same names and signatures, but there is no inheritance. KMFolderIndex contains some common indexing functionality for physical folders. Subclasses of KMFolderIndex finally interact directly with physical storage or with storage providers over the network.

KMFolderDir is a directory abstraction which holds KMFolderNode's. It inherits KMFolderNode and KMFolderNodeList which is a QPtrList<KMFolderNode>. A special case of a KMFolderDir is KMFolderRootDir; it represents the toplevel KMFolderDir in KMail's folder hierarchy.

KMFolderDir's contents are managed by KMFolderMgr's. KMail contains three main KMFolderMgr's. They can be accessed via KMKernel ( the "kmkernel" construct ). Those methods are : 1) KMFolderMgr *folderMgr() - which returns the folder manager for the folders stored locally. 2) KMFolderMgr *imapFolderMgr() - which returns the folder manager for all imap folders. They're handled a little differently because for all imap messages only headers are cached locally while the main contents of all messages is kept on the server. 3) KMFolderMgr *dimapFolderMgr() - which returns disconnected IMAP (dimap) folder manager. In dimap, both the headers and a copy of the full message are cached locally. 4) KMFolderMgr *searchFolderMgr() - which returns the folder manager for search folders (folders created by using the "find messages" tool). Other email clients call this type of folder "virtual folders".

FolderJob classes - These classes allow asynchronous operations on KMFolder's. You create a Job on the heap, connect to one of its signals and wait for the job to finish. Folders serve as FolderJob factories. For example, to retrieve the full message from a folder you do :

FolderJob *job = folderParent->createJob( aMsg, tGetMessage ); connect( job, SIGNAL(messageRetrieved(KMMessage*)), SLOT(msgWasRetrieved(KMMessage*)) ); job->start();

KMail architectural overview

Files: kmfolderindex.{h,cpp} and kmmsg{base,info}.{h,cpp}

Contact Marc Mutz <mutz@kde.org> or Till Adam <adam@kde.org> or Don Sanders <sanders@kde.org> with questions...

index := header *entry

header := magic LF NUL header-length byte-order-marker sizeof-long

magic := "# KMail-Index V" 1*DIGITS

header-length := Q_UINT32

byte-order-marker := Q_UINT32( 0x12345678 )

sizeof-long := Q_UINT32( 4 / 8 )

entry := tag length value

tag := Q_UINT32 ; little endian (native?)

length := Q_UINT16 ; little endian (native?)

value := unicode-string-value / ulong-value

unicode-string-value := 0*256QChar ; network-byte-order

ulong-value := unsigned_long ; little endian

Currently defined tag values are:

Msg*Part num. val type obtained by:

No 0 u - From 1 u fromStrip().stripWhitespace() Subject 2 u subject().stripWhitespace() To 3 u toStrip().stripWhiteSpace() ReplyToIdMD5 4 u replyToIdMD5().stripWhiteSpace() IdMD5 5 u msgIdMD5().stripWhiteSpace() XMark 6 u xmark().stripWhiteSpace() Offset 7 l folderOffset() (not only mbox!) LegacyStatus 8 l mLegacyStatus Size 9 l msgSize() Date 10 l date() File 11 u fileName() (not only maildir!) CryptoState 12 l (signatureState() << 16) | encryptionState()) MDNSent 13 l mdnSentState() ReplyToAuxIdMD5 14 u replyToAuxIdMD5() StrippedSubject 15 u strippedSubjectMD5().stripWhiteSpace() Status 16 l status()

u: unicode-string-value; l: ulong-value

Proposed new (KDE 3.2 / KMail 1.6) entry format:

index := header *entry

entry := sync 1*( tag type content ) crc-tag crc-value

sync := Q_UINT16 (32?) ; resync mark, some magic bit pattern ; (together with preceding crc-tag provides ; 24(40)bits to resync on in case of index ; corruption)

tag := Q_UINT8

type := Q_UINT8

content := variable-length-content / fixed-length-content

crc-tag := tag type ; tag=CRC16, type=CRC16

crc-value := Q_UINT16 ; the CRC16 sum is calculated over all of ; 1*( tag type content )

variable-length-content := length *512byte padding

padding := *3NUL ; make the string a multiple of 4 octets in length

fixed-length-content := 1*byte

length := Q_UINT16 (Q_UINT8?)

byte := Q_UINT8

The type field is pseudo-structured:

bit: 7 6 5 4 3 2 1 0 +-----+-----+-----+-----+-----+-----+-----+-----+ | uniq | chunk | len | +-----+-----+-----+-----+-----+-----+-----+-----+

uniq: 3 bits = max. 8 different types with same chunk size:

for chunk = (0)00 (LSB(base)=0: octets): 00(0) Utf8String 01(0) BitField 10(0) reserved 11(0) Extend

for chunk = (1)00 (LSB(base)=1: 16-octet blocks): 00(1) MD5(Hash/List) 01(1) reserved 10(1) reserved 11(1) Extend

for chunk = 01 (shorts): 000 Utf16String 001-110 reserved 111 Extend

for chunk = 10 (int32): 000 Utf32String (4; not to be used yet) 001 Size32 010 Offset32 011 SerNum/UOID 100 DateTime 101 Color (QRgb: (a,r,g,b)) 110 reserved 111 Extend

for chunk = 11 (int64): 000 reserved 001 Size64 010 Offset64 011-110 reserved 111 Extend

len: length in chunks 000 (variable width -> Q_UINT16 with the real width follows) 001..111: fixed-width data of length 2^len (2^1=2..2^6=128)

You find all defined values for the type field in indexentrybase.cpp

Currently defined tags are:

tag type content

DateSent DateTime Date: DateReceived DateTime last Received:'s date-time FromDisplayName String decoded display-name of first From: addr ToDisplayName String dto. for To: CcDisplayName String dto. for Cc: FromAddrSpecs String possibly IMAA-encoded, comma-separated addr-spec's of From: ToAddrSpecs String dto. for To: CcAddrSpecs String dto. for Cc: Subject String decoded Subject: BaseSubjectMD5 String md5(utf8(stripOffPrefixes(subject()))) BodyPeek String body preview MaildirFile String Filename of maildir file for this messagea MBoxOffset Offset(64) Offset in mbox file (pointing to From_) MBoxLength Size(64) Length of message in mbox file (incl. From_) Size Size(64) rfc2822-size of message (in mbox: excl. From_) Status BitField (see below) MessageIdMD5 MD5Hash MD5Hash of _normalized_ Message-Id: MDNLink SerialNumber SerNum of MDN received for this message DNSLink SerialNumber SerNUm of DSN received for this message ThreadHeads SerialNumberList MD5Hash's of all (so far discovered) _top-level thread parents_ ThreadParents SerialNumberList MD5Hash's of all (so far discovered) thread parents

"String" is either Utf8String or (Utf16String or Latin1String), depending on content

Currently allocated bits for the Status BitField are:

Bit Value: on(/off) (\imapflag)

# "std stati": 0 New (\Recent) 1 Read (\Seen) 2 Answered (\Answered) 3 Deleted (\Deleted) 4 Flagged (\Flagged) 5 Draft (\Draft) 6 Forwarded 7 reserved

# message properties: 8 HasAttachments 9 MDNSent ($MDNSent) 10..11 00: unspecified, 01: Low, 10: Normal, 11: High Priority 12..15 0001: Queued, 0010: Sent, 0011: DeliveryDelayed, 0100: Delivered, 0101: DeliveryFailed, 0110: DisplayedToReceiver, 0111: DeletedByReceiver, 1001: ProcessedByReceiver, 1010: ForwardedByReceiver, 1011-1110: reserved 1111: AnswerReceived

# signature / encryption state: 16..19 0001: FullyEncrypted, 0010: PartiallyEncrypted, 1xxx: Problem 20..23 0001: FullySigned, 0010: PartiallySigned, 1xxx: Problem

# "workflow stati": 24..25 01: Important, 10: ToDo, 11: Later (from Moz/Evo) 26..27 01: Work, 10: Personal, 11: reserved (dto.) 28..29 01: ThreadWatched, 10: ThreadIgnored, 11: reserved 30..31 reserved

All bits and combinations marked as reserved MUST NOT be altered if set and MUST be set to zero (0) when creating the bitfield.

Headers (Threading and Sorting)

Contact Till Adam <adam@kde.org> or Don Sanders <sanders@kde.org> with questions...

Threading and sorting is implemented in kmheaders.[cpp|h] and headeritem.[cpp|h] and involves a handfull of players, namely:

class KMHeaders: this is the listview that contains the subject, date etc of each mail. It's a singleton, which means there is only one, per mainwidget headers list. It is actually a member of KMMainwidget and accessible there.

class HeaderItem: these are the [Q|K]ListViewItem descendend items the KMHeaders listview consists of. There's one for each message.

class SortCacheItem: these are what the threading and sorting as well as the caching of those operate on. Each is paired with a HeaderItem, such that each holds a pointer to one, and vice versa, each header item holds a pointer to it's associated sort cache item.

.sorted file: The order of the sorted and threaded (if threading is turned on for this folder) messages is cached on disk in a file named .$FOLDER.index.sorted so if, for example the name of a folder is foo, the associated sorting cache file would be called ".foo.index.sorted". For each message, its serial number, that of its parent, the length of its sorting key, and the key itself are written to this file. At the start of the file several per folder counts and flags are cached additionally, immediately after a short file headers. The entries from the start of the file are:

  • "## KMail Sort V%04d\n\t" (magic header and version string)
  • byteOrder flag containing 0x12345678 for byte order testing
  • column, the sort column used for the folder
  • ascending, the sort direction
  • threaded, is the view threaded or is it not?
  • appended, have there been items appended to the file (obsolete?)
  • discovered_count, number of new mail discovered since the last sort file generation
  • sorted_count, number of sorted messages in the header list

What is used for figuring out threading?

  • messages can have an In-Reply-To header that contains the message id of another message. This is called a perfect parent.
  • additionally there is the References header which, if present, holds a list of message ids that the current message is a follow up to. We currently use the second to last entry in that list only. See further down for the reasoning behind that.
  • If the above two fail and the message is prefixed (Re: foo, Fwd: foo etc.) an attempt is made to find a parent by looking for messages with the same subject. How that is done is explained below as well.

For all these comparisons of header contents, the md5 hashes of the headers are used for speed reasons. They are stored in the index entry for each message. All data structures described below use md5 hash strings unless stated otherwise.

Strategy: When a folder is opened, updateMessageList is called, which in turn calls readSortOrder where all the fun happens. If there is a .sorted file at the expected location, it is openend and parsed. The header flags are read in order to figure out the state in which this .sorted file was written. This means the sort direction, whether the folder is threaded or not etc. FIXME: is there currently sanity checking going on? Now the file is parsed and for each message in the file a SortCacheItem is created, which holds the offset in the .sorted file for this message as well as it's sort key as read from the file. That sort cache item is entered into the global mSortCache structure (member of the KMHeaders instance), which is a QMemArray<SortCacheItem *> of the size mFolder->count(). Note that this is the size reported by the folder, not as read from the .sorted file. The mSortCache (along with some other structures) is updated when messages are added or removed. More on that further down. As soon as the serial number is read from the file, that number is looked up in the message dict, to ensure it is still in the current folder. If not, it has been moved away in the meantime (possibly by outside forces such as other clients) and a deleted counter is incremented and all further processing stopped for this message. The messages parent serial number, as read from the sorted file is then used to look up the parent and reset it to -1 should it not be in the current folder anymore. -1 and -2 are special values that can show up as parent serial numbers and are used to encode the following: -1 means this message has no perfect parent, a parent for it needs to be found from among the other messages, if there is a suitable one -2 means this message is top level and will forever stay so, no need to even try to find a parent. This is also used for the non-threaded case. These are messages that have neither an In-Reply-To header nor a References header and have a subject that is not prefixed. In case there is a perfect parent, the current sort cache item is appended to the parents list of unsorted children, or to that of root, if there is not. A sort cache item is created in the mSortCache for the parent, if it is not already there. Messages with a parent of -1 are appended to the "unparented" list, which is later traversed and its elements threaded. Messages with -2 as the parent are children of root as well, as noted above, and will remain so.

Once the end of the file is reached, we should have a nicely filled mSortCache, containing a sort cache item for each message that was in the sorted file. Messages with perfect parents know about them, top level messages know about that as well, all others are on a list and will be threaded later.

Now, what happens when messages have been added to the store since the last update of the .sorted file? Right, they don't have a sort cache item yet, and would be overlooked. Consequently all message ids in the folder from 0 to mFolder->count() are looked at and a SortCacheItem is created for the ones that do not have one yet. This is where all sort cache items are created if there was no sorted file. The items created here are by definition un- sorted as well as unparented. On creation their sort key is figured out as well.

The next step is finding parents for those messages that are either new, or had a parent of -1 in the .sorted file. To that end, a dict of all sort cache items indexed by the md5 hash of their messsage id headers is created, that will be used for looking up sort cache items by message id. The list of yet unparented messages is then traversed and findParent() called for each element wihch checks In-Reply-To and References headers and looks up the sort cache item of those parents in the above mentioned dict. Should none be found, the item is added to a second list the items of which will be subject threaded.

How does findParent() work, well, it tries to find the message pointed to by the In-Reply-To header and if that fails looks for the one pointed to by the second to last entry of the References header list. Why the second to last? Imagine the following situation in Bob's kmail:

  • Bob get's mail from Alice
  • Bob replies to Alice's mail, but his mail is stored in the Outbox, not the Inbox.
  • Alice replies again, so Bob now has two mails from Alice which are part of the same thread. The mail in the middle is somewhere else. We want that to thread as follows: Bob <- In-Reply-To: Alice1 ============ Alice1 |_Alice <- In-Reply-To: Bob (not here), References: Alice, Bob

  • since the above is a common special case, it is worth trying. I think. ;)

If the parent found is perfect (In-Reply-To), the sort cache items is marked as such. mIsImperfectlyThreaded is used for that purposer, we will soon see why that flag is needed.

On this first pass, no subject threading is attempted yet. Once it is done, the messages that are now top-level, the current thread heads, so to speak, are collected into a second dict ( QDict< QPtrList< SortCacheItem > > ) that contains for each different subject an entry holding a list of (so far top level) messages with that subject, that are potential parents for threading by subjects. These lists are sorted by date, so the parent closest by date can be chosen. Sorting of these lists happens via insertion sort while they are built because not only are they expected to be short (apart from hard corner cases such as cvs commit lists, for which subject threading makes little sense anyhow and where it should be turned off), but also since the messages should be roughly pre sorted by date in the store already. Some cursory benchmarking supports that assumption. If now a parent is needed for a message with a certain subject, the list of top level messages with that subject is traversed and the first one that is older than our message is chosen as it's parent. Parents more than six weeks older than the message are not accepted. The reasoning being that if a new message with the same subject turns up after such a long time, the chances that it is still part of the same thread are slim. The value of six weeks is chosen as a result of a poll conducted on kde-devel, so it's probably bogus. :) All of this happens in the aptly named: findParentBySubject().

Everthing that still has no parent after this ends up at toplevel, no further attemp is made at finding one. If you are reading this because you want to implement threading by body search or somesuch, please go away, I don't like you. :)

Ok, so far we have only operated on sort cache items, nothing ui wise has happened yet. Now that we have established the parent/child relations of all messages, it's time to create HeaderItems for them for use in the header list. But wait, you say, what about sorting? Wouldn't it make sense to do that first? Exactly, you're a clever bugger, ey? Here, have a cookie. ;) Both creation of header items and sorting of the as of yet unsorted sort cache items happen at the same time.

As previously mentioned (or not) each sort cache item holds a list of its sorted and one of its unsorted children. Starting with the root node the unsorted list is first qsorted, and then merged with the list of already sorted children. To achieve that, the heads of both lists are compared and the one with the "better" key is added to the list view next by creating a KMHeaderListItem for it. That header item receives both its sort key as well as its id from the sort cache item. Should the current sort cache item have children, it is added to the end of a queue of nodes to repeat these steps on after the current level is sorted. This way, a breadth first merge sort is performed on the sort cache items and header items are created at each node.

What follows is another iteration over all message ids in the folder, to make sure that there are HeaderItems for each now. Should that not be the case, top level emergency items are created. This loop is also used to add all header items that are imperfectly threaded to a list so they can be reevalutated when a new message arrives. Here the reverse mapping from header items to sort cache items are made as well. Those are also necessary so the sort cache item based data structures can be updated when a message is removed.

The rest of readSortOrder should make sense on itself, I guess, if not, drop me an email.

What happens when a message arrives in the folder? Among other things, the msgAdded slot is called, which creates the necessary sort cache item and header item for the new message and makes sure the data structures described above are updated accordingly. If threading is enabled, the new message is threaded using the same findParent and findParentBySubject methods used on folder open. If the message ends up in a watched or ignored thread, those status bits are inherited from the parent. The message is also added to the dict of header items, the index of messages by message id and, if applicable and if the message is threaded at top level, to the list of potential parents for subject threading.

After those house keeping tasks are performed, the list of as of yet imper- fectly threaded messages is traversed and our newly arrived message is considered as a new parent for each item on it. This is especially important to ensure that parents arriving out of order after their children still end up as parents. If necessary, the entries in the .sorted file of rethreaded messages are updated. An entry for the new message itself is appended to the .sorted file as well.

Note that as an optimization newly arriving mail as part of a mailcheck in an imap folder is not added via msgAdded, but rather via complete reload of the folder via readSortOrder(). That's because only the headers are gotten for each mail on imap and that is so fast, that adding them individually to the list view is very slow by comparison. Thus we need to make sure that writeSortOrder() is called whenever something related to sorting changes, otherwise we read stale info from the .sorted file. The reload is triggered by the folderComplete() signal of imap folders.

What happens when a message is removed from the folder? In this case the msgRemoved slot kicks in and updates the headers list. First the sort cache item and header item representing our message are removed from the data structures and the ids of all items after it in the store decre- mented. Then a list of children of the message is assembled containing those children that have to be reparented now that our message has gone away. If one of those children has been marked as toBeDeleted, it is simply added to root at top level, because there is no need to find a parent for it if it is to go away as well. This is an optimization to avoid rethreading all messages in a thread when deleting several messages in a thread, or even the whole thread. The KMMoveCommand marks all messages that will be moved out of the folder as such so that can be detected here and the useless reparenting can be avoided. Note that that does not work when moving messages via filter action.

That list of children is then traversed and a new parent found for each one using, again, findParent and findParentBySubject. When a message becomes imperfectly threaded in the process, it is added to the corresponding list.

The message itself is removed from the list of imperfectly threaded messages and its header item is finally deleted. The HeaderItem destructor destroys the SortCacheItem as well, which is hopefully no longer referenced anywhere at this point.

DISPLAY (reader window - new)

Contact Marc Mutz <mutz@kde.org> with questions...

What happens when a message is to displayed in the reader window? First, since KMMessagePart and KMMessage don't work with nested body parts, a hierarchical tree of MIME body parts is constructed with partNode's (being wrappers around KMMessagePart and DwBodyPart). This is done in KMReaderWin::parseMsg(KMMessage*). After some legacy handling, an ObjectTreeParser is instantiated and it's parseObjectTree() method called on the root partNode. ObjectTreeParser is the result of an ongoing refactoring to enhance the design of the message display code. It's an implementation of the Method Object pattern, used to break down a huuuge method into many smaller ones without the need to pass around a whole lot of paramters (those are made members instead). parseObjectTree() recurses into the partNode tree and generates an HTML representation of each node in the tree (although some can be hidden by setting them to processed beforehand or - the new way - by using AttachmentStrategy::Hidden). The HTML generation is delegated to BodyPartFormatters, while the HTML is written to HTMLWriters, which abstract away HTML sinks. One of those is KHTMLPartHTMLWriter, which writes to the KHTMLPart in the readerwindow. This is the current state of the code. The goal of the ongoing refactoring is to make the HTML generation blissfully ignorant of the readerwindow and to allow display plugins that can operate against stable interfaces even though the internal KMail classes change dramatically.

To this end, we designed a new set of interfaces that allows plugins to be written that need or want to asynchronously update their HTML representation. This set of interfaces consists of the following:

  • BodyPartFormatterPlugin the plugin base class. Allows the BodyPartFormatterFactory to query it for an arbitray number of BodyPartFormatters and their associated metadata and url handlers.

  • BodyPartFormatter the formatter interface. Contains a single method format() that takes a BodyPart to process and an HTMLWriter to write the generated HTML to and returns a result code with which it can request more information or delegate the formatting back to some default processor. BodyPartFormatters are meant to be Flyweights, implying that the format() method is not allowed to maintain state between calls, but see Mememto below.

  • BodyPart body part interface. Contains methods to retrieve the content of a message part and some of the more interesting header information. This interface decouples the bodypart formatters from the implementation used in KMail (or KNode, for that matter). Also contains methods to set and retrieve a Memento, which is used by BodyPartFormatters to maintain state between calls of format().

  • BodyPartMemento interface for opaque state storage and event handling. Contains only a virtual destructor to enable the backend processing code to destroy the object without the help of the bodypart formatter.

During the design phase we identified a need for BodyPartFormatters to request their being called on some form of events, e.g. a dcop signal. Thus, the Memento interface also includes the IObserver and ISubject interfaces. If a BodyPartFormatter needs to react to a signal (Qt or DCOP), it implements the Memento interface using a QObject, connects the signal to a slot on the Memento and (as an ISubject) notifies it's IObservers when the slot is called. If a Memento is created, the reader window registers itself as an observer of the Memento and will magically invoke the corresponding BodyPartFormatter, passing along the Memento to be retrieved from the BodyPart interface.

An example should make this clearer. Suppose, we want to update our display after 10 seconds. Initially, we just write out an icon, and after 10 seconds, we want to replace the icon by a "Hello world!" line. The following code does exactly this:

class DelayedHelloWorldBodyPartFormatter
      : public KMail::BodyPartFormatter {
public:
  Result format( KMail::BodyPart * bodyPart,
         KMail::HTMLWriter * htmlWriter ) {
    if ( !bodyPart->memento() ) {
      bodyPart->registerMemento( new DelayedHelloWorldBodyPartMemento() );
      return AsIcon;
    } else {
      htmlWriter->write( "Hello, world!" );
      return Ok;
    }
  }
};

class DelayedHelloWorldBodyPartMemento
      : public QObject, public KMail::BodyPartMemento {
public:
  DelayedHelloWorldBodyPartMemento()
    : QObject( 0, "DelayedHelloWorldBodyPartMemento" ),
      KMail::BodyPartMemento()
  {
    QTimer::singleShot( 10*1000, this, SLOT(slotTimeout()) );
  }

private slots:
  void slotTimeout() { notify(): }

private:
  // need to reimplement this abstract method...
  bool update() { return true; }
};