1
0
Fork 0
mutter-performance-source/HACKING

300 lines
14 KiB
Text
Raw Normal View History

Intro...
Window managers have a few ways in which they are significantly different
from other applications. This file, combined with the code overview in
doc/code-overview.txt, should hopefully provide a series of relatively
quick pointers (hopefully only a few minutes each) to some of the places
one can look to orient themselves and get started. Some of this will be
general to window managers on X, much will be specific to Metacity, and
there's probably some information that's common to programs in general but
is nonetheless useful.
Overview
Administrative issues
Minimal Building/Testing Environment
Relevant standards and X properties
Debugging and testing
Debugging logs
Adding information to the log
Valgrind
Testing Utilities
Technical gotchas to keep in mind
Other important reading
Extra reading
Ideas for tasks to work on
Administrative issues
Don't commit substantive code in here without asking hp@redhat.com.
Adding translations, no-brainer typo fixes, etc. is fine.
The code could use cleanup in a lot of places, feel free to do so.
See http://developer.gnome.org/dotplan/for_maintainers.html for
information on how to make a release. The only difference from those
instructions is that the minor version number of a Metacity release
should always be a number from the Fibonacci sequence.
Minimal Building/Testing Environment
You do not need to _install_ a development version of Metacity to
build, run and test it; you can run it from some temporary
directory. Also, you do not need to build all of Gnome in order to
build a development version of Metacity -- odds are, you may be able
to build metacity from CVS without building any other modules.
As long as you have gtk+ >= 2.10 with your distro (gtk+ >= 2.6 if
you manually revert the change from bug 348633), you should be able
to install your distro's development packages (e.g. gtk2-devel,
GConf2-devel, startup-notification-devel on Fedora; also, remember
to install the gnome-common package which is needed for building cvs
versions of Gnome modules like Metacity) as well as the standard
development tools (gcc, autoconf, automake, pkg-config, intltool,
and libtool) and be ready to build and test Metacity. Steps to do
so:
$ cvs -q -z3 -d :pserver:anonymous@anoncvs.gnome.org:/cvs/gnome \
checkout metacity
$ cd metacity
$ ./autogen.sh --prefix /usr
$ make
$ ./src/metacity --replace
Again, note that you do not need to run 'make install'.
Relevant standards and X properties
There are two documents that describe some basics about how window
managers should behave: the ICCCM (Inter-Client Communication Conventions
Manual) and EWMH (Extended Window Manager Hints). You can find these at
the following locations:
ICCCM - http://tronche.com/gui/x/icccm/
EWMH - :pserver:anoncvs@pdx.freedesktop.org:/cvs
The ICCCM is usually available in RPM or DEB format as well. There is
actually an online version of the EWMH, but it is almost always woefully
out of date. Just get it from cvs with these commands (the backslash
means include the stuff from the next line):
cvs -d :pserver:anoncvs@cvs.freedesktop.org:/cvs/icccm-extensions login
cvs -d :pserver:anoncvs@cvs.freedesktop.org:/cvs/icccm-extensions \
checkout wm-spec
DO NOT GO AND READ THOSE THINGS. THEY ARE REALLY, REALLY BORING.
If you do, you'll probably end up catching up on your sleep instead of
hacking on Metacity. ;-) Instead, just look at the table of contents and
glance at a page or two to get an idea of what's in there. Then only
refer to it if you see something weird in the code and you don't know
what it is but has some funny looking name like you see in one of those
two documents.
You can refer to the COMPLIANCE file for additional information on these
specifications and Metacity's compliance therewith.
One of the major things those documents cover that are useful to learn
about immediately are X properties. The right way to learn about those,
though, is through hand on experimentation with the xprop command (and
then look up things you find from xprop in those two manuals if you're
curious enough). First, try running
xprop
in a terminal and click on one of the windows on your screen. That gives
you the x properties for that window. Look through them and get a basic
idea of what's there for kicks. Note that you can get rid of some of the
verboseness by grepping out the _NET_WM_ICON stuff, i.e.
xprop | grep -v _NET_WM_ICON
Next, try running
xprop -root
in a terminal. There's all the properties of the root window (which you
can think of as the "main" Xserver window). You can also manually
specify individual windows that you want the properties of with
xprop -id <id>
if you know the id of the window in question. You can get the id of a
given window by either running xwininfo, e.g.
xwininfo | grep "Window id" | cut -f 4 -d ' '
or by looking at the _NET_CLIENT_STACKING property of the root
window. Finally, it can also be useful to add "-spy" (without the
quotes) to the xprop command to get it to continually monitor that
window and report any changes to you.
Debugging information
Trying to run a window manager under a typical debugger, such as gdb,
unfortunately just doesn't work very well. So, we have to resort to
other methods.
Debugging logs
First, note that you can start a new version of metacity to replace the
existing one by running
metacity --replace
(which also comes in handy in the form "./src/metacity --replace" when
trying to quickly test a small change while hacking on metacity without
doing a full "make install", though I'm going off topic...) This will
allow you to see any warnings printed at the terminal. Sometimes it's
useful to have these directed to a logfile instead, which you can do by
running
METACITY_USE_LOGFILE=1 metacity --replace
The logfile it uses will be printed in the terminal. Sometimes, it's
useful to get more information than just warnings. You can set
METACITY_VERBOSE to do that, like so:
METACITY_VERBOSE=1 METACITY_USE_LOGFILE=1 metacity --replace
(note that METACITY_VERBOSE=1 can be problematic without
METACITY_USE_LOGFILE=1; avoid it unless running in from something that
won't be managed by the new Metacity--see bug 305091 for more details).
There are also other flags, such as METACITY_DEBUG, most of which I
haven't tried and don't know what they do. Go to the source code
directory and run
grep "METACITY_" * | grep getenv
to find out what the other ones are.
Adding information to the log
Since we can't single step with a debugger, we often have to fall back to
the primitive method of getting information we want to know: adding
"print" statements. Metacity has a fairly structured way to do this,
using the functions meta_warning, meta_topic, and meta_verbose. All
three have the same basic format as printf, except that meta_topic also
takes a leading enumeration parameter to specify the type of message
being shown (makes it easier for grepping in a verbose log). You'll find
tons of examples in the source code if you need them; just do a quick
grep or look in most any file. Note that meta_topic and meta_verbose
messages only appear if verbosity is turned on. I tend to frequently add
temporary meta_warning statements (or switch meta_topic or meta_verbose
ones to meta_warning ones) and then undo the changes once I've learned
the info that I needed.
There is also a meta_print_backtrace (which again is only active if
verbosity is turned on) that can also be useful if you want to learn how
a particular line of code gets called. And, of course, there's always
g_assert if you want to make sure some section isn't executed (or isn't
executed under certain conditions).
Valgrind
Valgrind is awesome for finding memory leaks or corruption and
uninitialized variables. But I also tend to use it in a non-traditional
way as a partial substitute for a normal debugger: it can provide me with
a stack trace of where metacity is crashing if I made a change that
caused it to do so, which is one of the major uses of debuggers. (And,
what makes it cooler than a debugger is that there will also often be
warnings pinpointing the cause of the crash from either some kind of
simple memory corruption or an uninitialized variable). Sometimes, when
I merely want to know what is calling a particular function I'll just
throw in an "int i; printf("%d\n", i);" just because valgrind will give
me a full stacktrace whenever it sees that uninitialized variable being
used (yes, I could use meta_print_backtrace, but that means I have to
turn verbosity on).
To run metacity under valgrind, use options typical for any Gnome
program, such as
valgrind --logfile=metacity.log --tool=memcheck --num-callers=48 \
--leak-check=yes --leak-resolution=high --show-reachable=yes \
./src/metacity --replace
where, again, the backslashes mean to join all the stuff on the following
line with the previous one.
However, there is a downside. Things run a little bit slowly, and it
appears that you'll need about 1.5GB of ram, which unfortunately prevents
most people from trying this.
Testing Utilities
src/run-metacity.sh
The script src/run-metacity.sh is useful to hack on the window manager.
It runs metacity in an Xnest. e.g.:
CLIENTS=3 ./run-metacity.sh
or
DEBUG=memprof ./run-metacity.sh
or
DEBUG_TEST=1 ./run-metacity-sh
or whatever.
metacity-message
The tool metacity-message can be used as follows:
metacity-message reload-theme
metacity-message restart
metacity-message enable-keybindings
metacity-message disable-keybindings
The first of these is useful for testing themes, the second is just
another way (besides the --restart flag to metacity itself) of
restarting metacity, and the third is useful for testing Metacity when
running it under an Xnest (typically, the Metacity under the Xnest
wouldn't get keybinding notifications--making keyboard navigation not
work--but if you disable the keybindings for the global Metacity then
the Metacity under the Xnest can then get those keybinding notifications).
metacity-window-demo
metacity-window-demo is good for trying behavior of various kinds
of window without launching a full desktop.
Technical gotchas to keep in mind
Files that include gdk.h or gtk.h are not supposed to include
display.h or window.h or other core files. Files in the core
(display.[hc], window.[hc]) are not supposed to include gdk.h or
gtk.h. Reasons:
"Basically you don't want GDK most of the time. It adds
abstractions that cause problems, because they aren't designed to
be used in a WM where we do weird stuff (display grabs, and just
being the WM). At best GDK adds inefficiency, at worst it breaks
things in weird ways where you have to be a GDK guru to figure
them out. Owen also told me that they didn't want to start adding
a lot of hacks to GDK to let a WM use it; we both agreed back in
the mists of time that metacity would only use it for the "UI"
bits as it does.
Having the split in the source code contains and makes very clear
the interface between the WM and GDK/GTK. This keeps people from
introducing extra GDK/GTK usage when it isn't needed or
appropriate. Also, it speeds up the compilation a bit, though this
was perhaps more relevant 5 years ago than it is now.
There was also a very old worry that the GDK stuff might have to
be in a separate process to work right; that turned out to be
untrue. Though who knows what issues the CM will introduce."
Remember that strings stored in X properties are not in UTF-8, and they
have to end up in UTF-8 before we try putting them through Pango.
If you make any X request involving a client window, you have to
meta_error_trap_push() around the call; this is not necessary for X
requests on the frame windows.
Remember that not all windows have frames, and window->frame can be NULL.
Other important reading & where to get started
Extra reading
There are some other important things to read to get oriented as well.
These are:
http://pobox.com/~hp/features.html
rationales.txt
doc/code-overview.txt
It pays to read http://pobox.com/~hp/features.html in order
to understand the philosophy of Metacity.
The rationales.txt file has two things: (1) a list of design choices with
links in the form of bugzilla bugs that discuss the issue, and (2) a list
outstanding bug categories, each of which is tracked by a particular
tracker bug in bugzilla from which you can find several closely related
bug reports.
doc/code-overview.txt provides a fairly good overview of the code,
including coverage of the function of the various files, the main
structures and their relationships, and places to start looking in the
code tailored to general categories of tasks.
Ideas for tasks to work on
There are a variety of things you could work on in the code. You may
have ideas of your own, but in case you don't, let me provide a list of
ideas you could choose from:
If you're ambitious, there's a list of things Havoc made that he'd really
like to see tackled, which you can find at
http://log.ometer.com/2004-05.html. Be sure to double check with someone
to make sure the item is still relevant if you're interested in one of
these. Another place to look for ideas, of course, is bugzilla. One can
just do queries and look for things that look fixable.
However, perhaps the best way of getting ideas of related tasks to work
on, is to look at the second half of the rationales.txt file, which tries
to group bugs by type.