~bzr-pqm/bzr/bzr.dev

Bazaar Windows Shell Extension Options
======================================

.. contents:: :local:

Introduction
------------

This document details the implementation strategy chosen for the
Bazaar Windows Shell Extensions, otherwise known as TortoiseBzr, or TBZR.
As justification for the strategy, it also describes the general architecture
of Windows Shell Extensions, then looks at the C++ implemented TortoiseSvn
and the Python implemented TortoiseBzr, and discusses alternative
implementation strategies, and the reasons they were not chosen.

The following points summarize the  strategy:

* Main shell extension code will be implemented in C++, and be as thin as
  possible.  It will not directly do any VCS work, but instead will perform
  all operations via either external applications or an RPC server.

* Most VCS operations will be performed by external applications.  For
  example, committing changes or viewing history will spawn a child
  process that provides its own UI.

* For operations where spawning a child process is not practical, an
  external RPC server will be implemented in Python and will directly use
  the VCS library. In the short term, there will be no attempt to create a
  general purpose RPC mechanism, but instead will be focused on keeping the
  C++ RPC client as thin, fast and dumb as possible.

Background Information
----------------------

The facts about shell extensions
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Well - the facts as I understand them :)

Shell Extensions are COM objects. They are implemented as DLLs which are
loaded by the Windows shell. There is no facility for shell extensions to
exist in a separate process - DLLs are the only option, and they are loaded
into other processes which take advantage of the Windows shell (although
obviously this DLL is free to do whatever it likes).
 
For the sake of this discussion, there are 2 categories of shell extensions:
 
* Ones that create a new "namespace". The file-system itself is an example of
  such a namespace, as is the "Recycle Bin". For a user-created example,
  picture a new tree under "My Computer" which allows you to browse a remote
  server - it creates a new, stand-alone tree that doesn't really interact
  with the existing namespaces.
 
* Ones that enhance existing namespaces, including the filesystem. An example
  would be an extension which uses Icon Overlays to modify how existing files
  on disk are displayed or add items to their context menu, for example.

The latter category is the kind of shell extension relevant for TortoiseBzr,
and it has an important implication - it will be pulled into any process
which uses the shell to display a list of files. While this is somewhat
obvious for Windows Explorer (which many people consider the shell), every
other process that shows a FileOpen/FileSave dialog will have these shell
extensions loaded into its process space. This may surprise many people - the
simple fact of allowing the user to select a filename will result in an
unknown number of DLLs being loaded into your process. For a concrete
example, when notepad.exe first starts with an empty file it is using around
3.5MB of RAM. As soon as the FileOpen dialog is loaded, TortoiseSvn loads
well over 20 additional DLLs, including the MSVC8 runtime, into the Notepad
process causing its memory usage (as reported by task manager) to more than
double - all without doing anything tortoise specific at all. (In fairness,
this illustration is contrived - the code from these DLLs are already in
memory and there is no reason to suggest TSVN adds any other unreasonable
burden - but the general point remains valid.)
 
This has wide-ranging implications. It means that such shell extensions
should be developed using a tool which can never cause conflict with
arbitrary processes. For this very reason, MS recommend against using .NET
to write shell extensions[1], as there is a significant risk of being loaded
into a process that uses a different version of the .NET runtime, and this
will kill the process. Similarly, Python implemented shell extension may well
conflict badly with other Python implemented applications (and will certainly
kill them in some situations). A similar issue exists with GUI toolkits used
- using (say) PyGTK directly in the shell extension would need to be avoided
(which it currently is best I can tell). It should also be obvious that the
shell extension will be in many processes simultaneously, meaning use of a
simple log-file (for example) is problematic.

In practice, there is only 1 truly safe option - a low-level language (such
as C/C++) which makes use of only the win32 API, and a static version of the
C runtime library if necessary. Obviously, this sucks from our POV. :)

[1]: http://blogs.msdn.com/oldnewthing/archive/2006/12/18/1317290.aspx

Analysis of TortoiseSVN code
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

TortoiseSVN is implemented in C++. It relies on an external process to
perform most UI (such as diff, log, commit etc.) commands, but it appears to
directly embed the SVN C libraries for the purposes of obtaining status for
icon overlays, context menu, drag&drop, etc.

The use of an external process to perform commands is fairly simplistic in
terms of parent and modal windows. For example, when selecting "Commit", a
new process starts and *usually* ends up as the foreground window, but it may
occasionally be lost underneath the window which created it, and the user may
accidently start many processes when they only need 1. Best I can tell, this
isn't necessarily a limitation of the approach, just the implementation.

Advantages of using the external process is that it keeps all the UI code
outside Windows explorer - only the minimum needed to perform operations
directly needed by the shell are part of the "shell extension" and the rest
of TortoiseSvn is "just" a fairly large GUI application implementing many
commands. The command-line to the app has even been documented for people who
wish to automate tasks using that GUI. This GUI is also implemented in C++
using Windows resource files.
 
TortoiseSvn has an option (enabled by default) which enabled a cache using a
separate process, aptly named TSVNCache.exe. It uses a named pipe to accept
connections from other processes for various operations. When enabled, TSVN
fetches most (all?) status information from this process, but it also has the
option to talk directly to the VCS, along with options to disable functionality
in various cases.

There doesn't seem to be a good story for logging or debugging - which is
what you expect from C++ based apps. :( Most of the heavy lifting is done by
the external application, which might offer better facilities.

Analysis of existing TortoiseBzr code
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The existing code is actually quite cool given its history (SoC student,
etc), so this should not be taken as criticism of the implementer nor of the
implementation. Indeed, many criticisms are also true of the TortoiseSvn
implementation - see above. However, I have attempted to list the bad things
rather than the good things so a clear future strategy can be agreed, with
all limitations understood.

The existing TortoiseBzr code has been ported into Python from other tortoise
implementations (probably svn). This means it is very nice to implement and
develop, but suffers the problems described above - it is likely to conflict
with other Python based processes, and it means the entire CPython runtime
and libraries are pulled into many arbitrary processes.

The existing TortoiseBzr code pulls in the bzrlib library to determine the
path of the bzr library, and also to determine the status of files, but uses
an external process for most GUI commands - ie, very similar to TortoiseSvn
as described above - and as such, all comments above apply equally here - but
note that the bzr library *is* pulled into the shell, and therefore every
application using the shell. The GUI in the external application is written
in PyGTK, which may not offer the best Windows "look and feel", but that
discussion is beyond the scope of this document.
 
It has a better story for logging and debugging for the developer - but not
for diagnosing issues in the field - although again, much of the heavy
lifting remains done by the external application.
 
It uses a rudimentary in-memory cache for the status of files and
directories, the implementation of which isn't really suitable (ie, no
theoretical upper bound on cache size), and also means that there is no
sharing of cached information between processes, which is unfortunate (eg,
imagine a user using Windows explorer, then switching back to their editor)
and also error prone (it's possible the editor will check the file in,
meaning Windows explorer will be showing stale data). This may be possible to
address via file-system notifications, but a shared cache would be preferred
(although clearly more difficult to implement).

One tortoise port recently announced a technique for all tortoise ports to
share the same icon overlays to help work around a limitation in Windows on
the total number of overlays (it's limited to 15, due to the number of bits
reserved in a 32bit int for overlays). TBZR needs to take advantage of that
(but to be fair, this overlay sharing technique was probably done after the
TBZR implementation).

The current code appears to recursively walk a tree to check if *any* file in
the tree has changed, so it can reflect this in the parent directory status.
This is almost certainly an evil thing to do (Shell Extensions are optimized
so that a folder doesn't even need to look in its direct children for another
folder, let alone recurse for any reason at all. It may be a network mounted
drive that doesn't perform at all.)

Although somewhat dependent on bzr itself, we need a strategy for binary
releases (ie, it assumes python.exe, etc) and integration into an existing
"blessed" installer.

Trivially, the code is not PEP8 compliant and was written by someone fairly
inexperienced with the language.

Detailed Implementation Strategy
--------------------------------

We will create a hybrid Python and C++ implementation.  In this model, we
would still use something like TSVNCache.exe (this external
process doesn't have the same restrictions as the shell extension itself) but
go one step further - use this remote process for *all* interactions with
bzr, including status and other "must be fast" operations. This would allow
the shell extension itself to be implemented in C++, but still take advantage
of Python for much of the logic.

A pragmatic implementation strategy will be used to work towards the above
infrastructure - we will keep the shell extension implemented in Python - but
without using bzrlib. This allows us to focus on this
shared-cache/remote-process infrastructure without immediately
re-implementing a shell extension in C++. Longer term, once the
infrastructure is in place and as optimized as possible, we can move to C++
code in the shell calling our remote Python process. This port should try and
share as much code as possible from TortoiseSvn, including overlay handlers.

External Command Processor
~~~~~~~~~~~~~~~~~~~~~~~~~~

The external command application (ie, the app invoked by the shell extension
to perform commands) can remain as-is, and remain a "shell" for other
external commands. The implementation of this application is not particularly
relevant to the shell extension, just the interface to the application (ie,
its command-line) is. In the short term this will remain PyGTK and will only
change if there is compelling reason - cross-platform GUI tools are a better
for bazaar than Windows specific ones, although native look-and-feel is
important. Either way, this can change independently from the shell
extension.

Performance considerations
~~~~~~~~~~~~~~~~~~~~~~~~~~

As discussed above, the model used by Tortoise is that most "interesting"
things are done by external applications. Most Tortoise implementations
show read-only columns in the "detail" view, and shows a few read only
properties in the "Properties" dialog - but most of these properties are
"state" related (eg, revision number), or editing of others is done by
launching an external application. This means that the shell extension itself
really has 2 basic requirements WRT RPC: 1) get the local state of a file and
2) get some named state-related "properties" for a file. Everything else can
be built on that.

There are 2 aspects of the shell integration which are performance critical -
the "icon overlays" and "column providers".

The short-story with Icon Overlays is that we need to register 12 global
"overlay providers" - one for each state we show. Each provider is called for
every icon ever shown in Windows explorer or in any application's FileOpen
dialog. While most versions of Windows update icons in the background, we
still need to perform well. On the positive side, this just needs the simple
"local state" of a file - information that can probably be carried in a
single byte. On the negative side, it is the shell which makes a synchronous
call to us with a single filename as an arg, which makes it difficult to
"batch" multiple status requests into a single RPC call.

The story with columns is messier - these have changed significantly for
Vista and the new system may not work with the VCS model (see below).
However, if we implement this, it will be fairly critical to have
high-performance name/value pairs implemented, as described above.

Note that the nature of the shell implementation means we will have a large
number of "unrelated" handlers, each called somewhat independently by the
shell, often for information about the same file (eg, imagine each of our
overlay providers all called in turn with the same filename, followed by our
column providers called in turn with the same filename. However, that isn't
exactly what happens!). This means we will need a kind of cache, geared
towards reducing the number of status or property requests we make to the RPC
server.

We will also allow all of the above to be disabled via user preferences.
Thus, Icon Overlays could be disabled if it did cause a problem for some
people, for example.
 
RPC options
~~~~~~~~~~~

Due to the high number of calls for icon overlays, the RPC overhead must be
kept as low as possible. Due to the client side being implemented in C++,
reducing complexity is also a goal. Our requirements are quite simple and no
existing RPC options exist we can leverage. It does not seen prudent to build
an XMLRPC solution for tbzr - which is not to preclude the use of such a
server in the future, but tbzr need not become the "pilot" project for an
XMLRPC server given these constraints.

I propose that a custom RPC mechanism, built initially using windows-specific
named-pipes, be used. A binary format, designed with an eye towards
implementation speed and C++ simplicity, will be used. If we succeed here, we
can build on that infrastructure, and even replace it should other more
general frameworks materialize.

FWIW, with a Python process at each end, my P4 2.4G machine can achieve
around 25000 "calls" per-second across an open named pipe. C++ at one end
should increase this a little, but obviously any real work done by the Python
side of the process will be the bottle-neck. However, this throughput would
appear sufficient to implement a prototype.

Vista versus XP
~~~~~~~~~~~~~~~

Let's try and avoid an OS advocacy debate :) But it is probably true that
TBZR will, over its life, be used by more Vista computers than XP ones. In
short, Vista has changed a number of shell related interfaces, and while TSVN
is slowly catching up (http://tortoisesvn.net/vistaproblems) they are a pain.

XP has IColumnProvider (as implemented by Tortoise), but Vista changes this
model. The new model is based around "file types" (eg, .jpg files) and it
appears each file type can only have 1 provider! TSVN also seems to think the
Vista model isn't going to work (see previous link). It's not clear how much
effort we should expend on a column system that has already been abandoned by
MS. I would argue we spend effort on other parts of the system (ie, the
external GUI apps themselves, etc) and see if a path forward does emerge for
Vista. We can re-evaluate this based on user feedback and more information
about features of the Vista property system.

Reuse of TSVNCache?
~~~~~~~~~~~~~~~~~~~

The RPC mechanism and the tasks performed by the RPC server (rpc, file system
crawling and watching, device notifications, caching) are very similar to
those already implemented for TSVN and analysis of that code shows that
it is not particularly tied to any VCS model.  As a result, consideration
should be given to making the best use of this existing debugged and
optimized technology.

Discussions with the TSVN developers have indicated that they would prefer us
to fork their code rather than introduce complexity and instability into
their code by attempting to share it. See the follow-ups to
http://thread.gmane.org/gmane.comp.version-control.subversion.tortoisesvn.devel/32635/focus=32651
for details.

For background, the TSVNCache process is fairly sophisticated - but mainly in
areas not related to source control. It has had various performance tweaks
and is smart in terms of minimizing its use of resources when possible. The
'cloc' utility counts ~5000 lines of C++ code and weighs in just under 200KB
on disk (not including headers), so this is not a trivial application.
However, the code that is of most interest (the crawlers, watchers and cache)
are roughly ~2500 lines of C++. Most of the source files only depend lightly
on SVN specifics, so it would not be a huge job to make the existing code
talk to Bazaar. The code is thread-safe, but not particularly thread-friendly
(ie, fairly coarse-grained locks are taken in most cases).

In practice, this give us 2 options - "fork" or "port":

* Fork the existing C++ code, replacing the existing source-control code with
  code that talks to Bazaar. This would involve introducing a Python layer,
  but only at the layers where we need to talk to bzrlib. The bulk of the
  code would remain in C++.

  This would have the following benefits:

  - May offer significant performance advantages in some cases (eg, a
    cache-hit would never enter Python at all.)

  - Quickest time to a prototype working - the existing working code can be
    used quicker.

  And the following drawbacks:

  - More complex to develop. People wishing to hack on it must be on Windows,
    know C++ and own the most recent MSVC8.

  - More complex to build and package: people making binaries must be on
    Windows and have the most recent MSVC8.

  - Is tied to Windows - it would be impractical for this to be
    cross-platform, even just for test purposes (although parts of it
    obviously could).

* Port the existing C++ code to Python. We would do this almost
  "line-for-line", and attempt to keep many optimizations in place (or at
  least document what the optimizations were for ones we consider dubious).
  For the windows versions, pywin32 and ctypes would be leaned on - there
  would be no C++ at all.

  This would have the following benefits:

  - Only need Python and Python skills to hack on it.

  - No C++ compiler needed means easier to cut releases

  - Python makes it easier to understand and maintain - it should appear much
    less complex than the C++ version.

  And the following drawbacks:

  - Will be slower in some cases - eg, a cache-hit will involve executing
    Python code.

  - Will take longer to get a minimal system working. In practice this
    probably means the initial versions will not be as sophisticated.

Given the above, there are two issues which prevent Python being the clear
winner: (1) will it perform OK? (2) How much longer to a prototype?

My gut feeling on (1) is that it will perform fine, given a suitable Python
implementation. For example, Python code that simply looked up a dictionary
would be fast enough - so it all depends on how fast we can make our cache.
Re (2), it should be possible to have a "stub" process (did almost nothing in
terms of caching or crawling, but could be connected to by the shell) in a 8
hours, and some crawling and caching in 40. Note that this is separate from
the work included for the shell extension itself (the implementation of which
is largely independent of the TBZRCache implementation). So given the lack of
a deadline for any particular feature and the better long-term fit of using
Python, the conclusion is that we should "port" TSVN for bazaar.

Reuse of this code by Mercurial or other Python based VCS systems?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Incidentally, the hope is that this work can be picked up by the Mercurial
project (or anyone else who thinks it is of use). However, we will limit
ourselves to attempting to find a clean abstraction for the parts that talk
to the VCS (as good design would dictate regardless) and then try and assist
other projects in providing patches which work for both of us. In other
words, supporting multiple VCS systems is not an explicit goal at this stage,
but we would hope it is possible in the future.

Implementation plan
-------------------

The following is a high-level set of milestones for the implementation:

* Design the RPC mechanism used for icon overlays (ie, binary format used for
  communication).

* Create Python prototype of the C++ "shim": modify the existing TBZR Python
  code so that all references to "bzrlib" are removed. Implement the client
  side of the RPC mechanism and implement icon overlays using this RPC
  mechanism.

* Create initial implementation of RPC server in Python. This will use
  bzrlib, but will also maintain a local cache to achieve the required
  performance. File crawling and watching will not be implemented at this
  stage, but caching will (although cache persistence might be skipped).

* Analyze performance of prototype. Verify that technique is feasible and
  will offer reasonable performance and user experience.

* Implement file watching, crawling etc by "porting" TSVNCache code to
  Python, as described above.

* Implement C++ shim: replace the Python prototype with a light-weight C++
  version. We will fork the current TSVN sources, including its new
  support for sharing icon overlays (although advice on how to setup this
  fork is needed!)
 
* Implement property pages and context menus in C++. Expand RPC server as
  necessary.

* Create binary for alpha releases, then go round-and-round until its baked.

Alternative Implementation Strategies
-------------------------------------

Only one credible alternative strategy was identified, as discussed below. No
languages other than Python and C++ were considered; Python as the bzr
library and existing extensions are written in Python and otherwise only C++
for reasons outlined in the background on shell extensions above.

Implement Completely in Python
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This would keep the basic structure of the existing TBZR code, with the
shell extension continuing to pull in Python and all libraries used by Bzr
into various processes.

Although implementation simplicity is a key benefit to this option, it was
not chosen for various reasons, e.g. the use of Python means that there is a
larger chance of conflicting with existing applications, or even existing
Python implemented shell extensions. It will also increase the memory usage
of all applications which use the shell. While this may create problems for a
small number of users, it may create a wider perception of instability or
resource hogging.