As part of the .eap to .desktop conversion, raster wants me to redo the
relevant caching to suit. He has made many "eap caching needs
fixing" noises in the recent past, so maybe a fresh start is needed
at this point.
This is a discussion and design document. Major parts of it where sent
to the devel list but got no response. Raster said to put it in here,
and keep it up to date with whatever information and design is relevant.
The problem -
It's really the E_App structure as defined in e_apps.h that may need
caching, no matter what their source. I say "no matter what their
source" because at the moment the sources are .eap files and .desktop
files, and at some unspecified date in the future .eap files might go
away. I say "may need caching" because this is a fresh start, do they
really need caching?
To answer that last question we need to look at all the places that use
E_Apps and check to see if those uses are performance critical. Then
we need to see how much time it takes to load an E_App from it's source.
Grepping through the source code shows that E_App is used in the ibar,
engage, and mbar modules, plus in E itself. Devilhorns tells me that
mbar will be replaced, so I will ignore that for now. HandyAndE tells
me that the stand alone version of engage is deprecated, so I can
ignore that. Both ibar and engage essentially do the same things with
E_App, so I'll speak about those in generic terms. I'll also try to
avoid spending time fully groking all the code that uses E_Apps, as I
have some time pressure here. That's why I'm thinking out loud, if I
miss something, other devs will point it out.
Interesting that neither entangle, nor the stand alone eapp_edit came up
in the grep results. There may be other things that deal with
application icons in their own way. I'll ignore them for now. I know
raster will say "engage entangle and such ar not part of corr e, ignre
them". Entangle and the stand alone eapp_edit are now obsolete, as they
where not included in the .desktop update.
Ibar and engage both display a handful of application icons, allow the
user to click on those icons to start the application. Ibar runs a
"starting application" animation, and engage keeps track of open
windows that belong to all applications, but tries to keep the open
window icons next to the application icon if it has one for that
application. I don't know if ibar uses the freedesktop.org startup
notification spec, some other method, or just runs the animation for an
arbitrary time. Either way, there are three operations in these
modules that use E_Apps - show the icon, start the application, and
match newly created windows with the E_App.
E also needs to do those three operations, and more. Menus, winlist,
pager, exebuf, and borders all need to show the icon. Menus and exebuf
need to start the application. When a window is opened, it needs to be
matched to it's E_App to show it's icon and other things. Exebuf needs
to search through all the E_Apps looking for matches to the partially
typed in command. EFM needs to show an icon for any .eap or .desktop it
is currently displaying, maybe needs to run the application. The
built in eap editor needs to read, display, and save all the
application icon information.
These operations fall into two distinct categories - dealing with one
E_App at a time, or searching through all E_Apps with arbitrary search
criteria. One at a time is often done in user time (the user only
operates one icon at a time) so caching is not so important there.
Searching through all E_Apps is an expensive operation that should be
optimised as much as possible. Note that I count the showing of menus
as the later category, as the user may want to quickly access many sub
menus full of E_Apps.
Since the source of the image data used for the icon can be something
that needs time to decode, at the times when lots of icons are being
used at once (exebuf and menus mostly) we probably want that image data
cached for speed as well. The searching through all E_Apps can occur
on data that is part of the original file that is the source, so an in
memory search will be a lot faster than any file searching.
So the problem is - as fast as possible search through all the E_Apps,
no matter what their source, and give access to all the information
they have ,including image data. Since this is useful in the "one at a
time" category, might as well use the same method there to.
The issues with possible solutions -
An in memory cache of all E_App information is needed. Since the
information that could be used to search this cache can be quite
arbitrary, hashes and other such structures may not be useful. On the
other hand, some sort of index mechanism may be useful for the most
common searches, and those that need to be quickest.
Lets try to come up with some numbers. The eap torture test involves
over two thousand .eaps, and that's from a full install of SuSE 9.3
Pro. Certain distros have just as many applications, or more, and they
are growing all the time. 120x120 pixels is probably the largest
commonly used icon size, 32 bits per pixel to include the alpha
channel. I have seen .desktop files up to 25 KB in size, but that
includes all the translations for name, comment, and keywords,
something that is not currently stored in E_App. A much more common
size is 2 - 4 KB, but that still mostly is for the name translations.
Note, .desktop files only include the name of the icon, not the image
data itself. I use the .desktop files for this part of the discussion
because they include all the non image data that is typically in an
E_App. 1 KB is probably maximum storage requirement for non image,
single translation E_App data. That could be a maximum of 56 KB per
E_App, and possibly 100 MB for the entire cache including image data.
Ouch, no wonder raster suggested mmaping the cache file.
Note, this is a likely maximum for those people that install two
thousand applications and like really big icons. I have over two
thousand applications installed, and I like the icons in engage to zoom
out really big. B-) Typical users will require a much smaller amount
The original caching strategy uses multiple caches, on a per directory
basis, but still has a single cache for the all directory, the one with
all the .eaps in it. raster has mentioned that a single cache might be
a better solution. With .desktop files scattered all over the place, a
single cache is more desirable for searching purposes.
This does raise some interesting issues. While .eap files have in the
past been per user, the majority of .desktop files are likely to be
system files, mostly unchanging, but with new ones added whenever new
software is installed. The user can create their own, and override
system ones. There is also the translations to deal with. Do we
include only one translation in the cache? Do we have a system cache
and per user caches that allow overrides? Do we really need image data
in the cache? Although the freedesktop.org spec says that the user
chooses a single icon size, in practise the various things that render
icons render them at different sizes.
There is the cache update strategy to consider. The current one causes
problems. Raster has thought about changing to no auto update, but with
a manually triggered update. I would prefer automatic updates that
work smoothly and without great pauses in the operation of the wm, as
that gives the users less to complain about. Maybe we can make use of
the thumb nailing stuff?
Answers to these questions will help choose a caching mechanism.
All of the ecore_desktop code.
At startup. They hardly ever change.
The use of GNOME and KDE apps to figure out some of the paths.
They are curently commented out, and guesses are made.
Go with the guesses, run the apps later to correct the guesses.
Cache the results to disk.
"Applications" menu and "Applications" config dialog.
The menus are parsed at first startup, then manually by the dialog.
When new applications are installed, will need to update.
Parse them during idle time, maybe as a seperate process. Monitor
relevant directories. ~/.e/e/applications/menu/all is the results
cached to disk.
The icon theme config dialog.
When the icon theme config dialog is started.
The actual icon theme currently in use is a small amount of info that
can always be in ram, and is quick to get at startup or theme change.
It's the complete list of themes that are currently installed that can
take ages to find. The list itself should also be small.
Find them during idle time, and finish the job async when the dialog
is started. Monitor the icon paths for the installation of new
themes. Keep them all in a hash.
Whenever an E_App is displayed and others.
Combined .edj and FDO searching. FDO searching can take a long time.
Things that display lots of icons wont want to wait for all those
complex searches. Searching for icons that don't exist is what
takes the longest, as the FDO algo means we have to look through all
the directories, in this theme, and the next theme.
e_app_icon_add() should be used everywhere, and it should register a
rectangle for use by the icon. The caller then shows an empty icon.
A thumbnailing style process then does all the searching, and does
the fm2 bouncy icon thing when the icon is found. raster forbids
The results of the search should be cached somewhere on disk for
future reference. That needs to be nuked when the user changes
icon theme. Changing the icon theme currently does not update all
displayed icons, it just does a restart like wm theme changing does.
The X-Enlightenment-IconPath field of the .desktop file could be
used for the on disk cache. X-Enlightenment-Icon-Theme and
X-Enlightenment-Icon-TimeStamp can be used to determine if the icon
path stored on disk is out of date, or in a different theme. An
idle task can go through all the .desktops in
~/.e/e/applications/all and check if the icons need updating. The
spec allows caching of dirs, and we can monitor just the top level
icon theme directory.
E_Apps and others.
Whenever the E_App needs to be read from disk.
Currently a fairly simple in memory hash of them is kept. Since
each Ecore_Desktop holds a hash of the fields, this can be
expensive in memory. Also, we would like a pre parsed binary on
disk cache for the sheer speed. ~/.e/e/applications/all is a
directory with all of them that are for E_Apps, some are just links
to system files.
Remove after implementing the E_App cache then see if these still
needs caching. Note that the menu generation probably reads the
.desktop files twice each.
Menus, bars, startup, restart.
When displaying the menu or bar. At startup and restart. When
Menus, bars, borders, exebuf, lots of other places.
Currently all read in at startup time. Should change this to an
idle time task. Still need to read in them all near startup time
though. We need all of them for searching.
a_actions searches for filename, name, generic, and exe to find an
app to execute. Updates the most recently used list.
e_border searches for border and pid to find the border icon. The
border search involves searching all E_Apps for win_* and exe that
match the app in the border. The pid search involves searching all
instances of all E_Apps, looking for a matching pid. Both searches
update the most recently used list. None of this is needed for
e_exebuf searches/lists all apps and most recently used apps based
on a partially typed in string. It globs name, exe, generic, and
comment. When actually listing the matches, it searches for exe.
When showing an icon for the current match, it searches exe, name,
then generic. Updates the most recently used list.
e_zone searches for exe to find a matching E_App when it is
executing things. Updates the most recently used list.
e_int_config_apps shows a sorted list of all apps.
Everything else loads E_Apps by file name or directory name. Some
have .order files.
The current eap caching code is not up to scratch, and the .desktop
conversion didn't help. It does hide some other bugs though, so that
needs to be cleaned up first.
How this all hangs together is complex.
Only contains ~/.e/e/applications/all for now.
E_App *_e_apps_all = ~/.e/e/applications/all
Currently has all the E_Apps as subbapps, but only used as if it was a single E_App.
On the other hand, this is scanned at startup time, so is used to preload the caches.
_e_apps_list = Most recently used list / all apps list.
E_App members -
E_App *parent; /* the parent e_app node */
This is in a directory, the parent is the E_App for that directory.
Evas_List *subapps; /* if this a directory, a list of more E_App's */
e_apps.c internal except for _e_startup and _e_int_menus_apps_scan
This will be a directory with a .order file, subapps points to the content E_Apps.
Evas_List *references; /* If this app is in a main repository, this would
be a list to other eapp pointing to this */
E_Apps in a repository are copied before use, this points to those copies.
E_App *orig; /* if this is a copy, point to the original */
e_apps.c internal except for _e_border_menu_cb_icon_edit
E_Apps in a repository are copied before use, this points to the original.
Evas_List *instances; /* a list of all the exe handles for executions */
e_apps.c internal except for e_zone_exec
Tracks E_Apps that are currently running.
Ecore_File_Monitor *monitor; /* Check for changes and files */
Only E_Apps that are directories are monitored.
Start by putting all filenames in ~/.e/e/applications/all in a hash
with "empty" E_Apps. Fill the "empty" E_Apps during idle time, also
as needed. Everytime a new .desktop file is created, add it to the
hash and put it in the directory.
The most recently used list can be built from whatever gets used, it
could actually speed things up if it does not contain everything, as
it is usually a linear search. On the other hand, we could reuse
the subapps of _e_apps_all as this list, since it already exists,
and it doesn't care what order its subapps are in.
Executing an app can probably afford to use a slow search. Border
icons could make use of bouncy icon thumbnailing. e_int_config_apps
can just show ~/.e/e/applications/all, but e_fm will need to know to
sort based on visible name. e_exebuf can also use bouncy icon
thumbnailing, but may need more speedups for its globbing.
Later we shall see if an on disk copy is needed.
comments on cache doc:
1. ignore caching of image data - evas already does that. just provide
file paths and "keys" if inside .edj/eap/eet files. don't duplicate
image data or do anything "smart" as it will invariably end up less
smart that the existing evas cache :) so don't worry about icons and
their pixels - just the file paths/keys
2. instead of "in memory" let's try a new idea. lets put it in a flat
mmap'ed file (not eet). let's not make it architecture dependant
(write any integers in a fixed byte order) and make it all offset
based. fill the file with strings and indexes and fast search indexes
etc. 1 big flat file for the whole cache.. the eap structs themselves
are "in memory" and allocated but all strings and "content" are pointed
to this mmaped file - map and unmap this file whenever needed (need to
add calls that implicitly map it when you want to access eap struct
contents and then unmap when done - refcount) and on map check the
header of the file for some "modification count) and keep the last mod
count value - if they don't match re-evaluate all internal tabels in
the map file. re-evaluate all pointers in eap structs. always try and
map the file to the same mem address (remember it and try map TO that
addr so you don't need to do address fixups). if you can't map to there
map to unspecified address (let mmap choose) and now keep trying to
map to this addr from now on. if the addr changes or the mmap file
changes all eap files need a fixup of their str pointers done.
so now you have a e_app_use() and e_app_unuse() thatg basically makes
sure the apps map file is mappeed after a use and unmapped after an
unuse (when refcount goes to 0 for the maps file - maybe with a
timeout to avoid map thrashing).
the advantage of all this is most of the time e is not accessing eapp
data - it just has references TO them so during this time no memory
is consumed by the data and no kernel mappings. we should bring mem
use down to a tiny fraction.
this is a quick braindump... so it may need more discussion. nb - this
is just a bug flat dump for all ".desktop files" - hash it for fast
lookup - have fast search hash indexes for lookup by exe, name, class
etc. other cache files for menu structures etc. would implicitly
reference data in this big fat cache.
3. on any change of .desktop files on the system or by users - rewrite the
whole cache and re-fix pointers etc. - this shouldn't happen often.
4. same mmap setup for lookups of icon name in .desktop -> fully resolved
path based on icon themes and inheritance. if the themes change then
re-evaluate this cache.
5. .order files and menu structs can just live in the same cache.
6. all eaps can be quickyl created at startup from an all apps cache that
referenced the desktop content cache file as above. putting it into
an idel taks means on addition of new eaps all window icons need
re-evaluation as to all eap instances in ibar etc. etc. - makes life
more complex (for now). maybe simpler on that front and better cache
will work well here.
7. eap instances are created as you lanich apps - no need to cache or
worry about these - the search by exe, win name/class etc. are what
we need to worry about :)
8. the above should mean no need to fill in eaps in idlers - map the
"big fat desktop cache" that contains everything - icon paths, labels
exe strings etc. when needed - fox up all pointers in eapp structs
if the contents or base address ever change - unmap when refcount
goes to 0 (with a timeout to avoid thrashing).