I have problem with Eclipse - it slows down after some time, it take seconds to refresh window/open menu/... and later stops to respond and must be killed. I found that xrestop shows still growing values in Pxms and Misc columns and top shows Xorg eating lots of cpu time. When I use another default style then Oxygen or gtk style other then oxygen-gtk, it is ok. Reproducible: Always Steps to Reproduce: 1. Use Oxygen/oxygen-gtk style 2. Have Eclipse running long time or do lots of GUI operation. Current config: openSUSE 12.1 KDE 4.8.97 nvidia drivers 302.17 gtk2 oxygen engine 1.3 eclipse juno but i have this problem longer time so it is not caused by latest versions of SW...
I"ll have a look. In priciple, pixmaps allocation via xrestop should significantly increase when you start the application, and ultimately saturate to some maximum, because we are using caches (to optimize rendering). Now, maybe there is a leak somewhere. Question: do you experience similar kind of issues with other gtk applications (e.g. gimp), notably concerning the xrestop behavior ? Cheers, Hugo
I am trying it now on another pc (OSS 12.1, kde 4.8.97, nvidia 302.17, oxygen-gtk2 1.2.50git.1342458469-41.1). I have tried pidgin - values grow while maximizing/minimizing window, opening/closing menus etc. In a few minutes it grows to thousands, more than any other app. GIMP and opening menus still around - the same, just faster (i suppose it has bigger menus:). GIMP with industrial gtk style after same procedure use just 19 pixmaps and 69 misc.
My experience with gimp: growing number of allocated pixmaps at startup and during first minutes of operations, then saturating to something like 1200 pixmaps, and 1.4M pixmap memory. -> these are small pixmaps (one must compare the *memory* -not number of pixmaps- with other apps) After that, number of pixmaps and memory might still grow by one or two every now and then but definitly not at the same pace as at the beginning. this is the *expected* behavior, and corresponds to caching. It is *normal* that you don't see it with other themes since they are not caching their (usually simpler) rendering. => comparison with "industrial" is irrelevant. Now: do you see the same behavior (that is: the number of allocated pixmaps reaching a maximum after some point), or do you see an ever-increasing amount of pixmaps being allocated, at constant pace ?
Also, please provide some actual numbers rather than "grows to thousands". What I see is 1 thousand pixmaps allocated. Which is expected. Not thousand*s*
I can confirm I have this problem for a very long time. Various KDE versions and using openSUSE and Kubuntu. Currently, Kubuntu 12.10, KDE 4.10. Hugo, how could I help fixing this ?
For example, current xrestop for eclipse, launched for a few hours, very slow now : res-base Wins GCs Fnts Pxms Misc Pxm mem Other Total PID Identifier 1000000 48 88 1 20859 21830 77901K 515K 78417K 6697 Java EE - Eclipse
And a freshly launched eclipse : 3e00000 16 47 1 464 569 2725K 15K 2741K 24176 Java EE - Eclipse And after a window and a context menu opened, the total is more than 32k.
@Gael thanks for the numbers. 20 thousands pixmaps is definitely too much. Now I'm not sure they are allocated by us. Will check the caches again ... Also: which version of oxygen-gtk do you have (it is largely uncorrelated with the version of KDE and is not shipped with KDE). Concerning your question "how could I help to fix this". Ideally: use valgrind with eclipse. But 1/ I have not succeeded to do so myself 2/ I'm not sure it is even possible due to its plugin nature and the use of Java ... I've googled a bit already but with no success ... (my concern, to be honest, is that as far as I know the same behavior is _not_ achieved with other gtk applications, running with oxygen-gtk. A lot of pixmaps are created, but the number "saturates" to around 1000, which corresponds to the caches being full. So that there must be something upstream (eclipse) that triggers this behavior together with oxygen-gtk ...
OK, I can reproduce. @Hugo, if you need a way to reproduce, here it is: 1. Make some C++ project 2. Open some tree items in Project Explorer so that they give a screenful of rows (this will make it easier to reproduce) 3. Now start moving mouse up-down over that treeview, so that hover rect moves many times over the list. At the same time watch xrestop, Eclipse will linearly increase its memory consumption. If I change "SimpleCache( size_t size=100..." to size of 1 (and the same for Cache()), nothing changes. Likewise, if I disable inner shadows hack, nothing is changed either. Now if I add a return; at the beginning of "if( d.isCell() )", I don't reproduce the problem.
OK, traced it down to StyleHelper::selection(). Looks like _selectionCache.insert() is called for each prelight, with different surface, even if it has the same key as just before.
Created attachment 77692 [details] Proposed patch @Hugo Apparently, TileSet's h1 and h3 appear both 0: h1 is passed as 0 constant, h3 looks like gets subtracted from surface height which appears equal it. And TileSet::isValid() checks for height()>0, where height = h1+h3. So, should this check be changed somehow? Maybe "(width()>0||height()>0)&&_surfaces.size() == 9"? instead of both width()>0 && height()>0? The change is attached as a proposed patch. I'm not sure it's the way to go, but at least I no longer reproduce this bug with the patch.
@Ruslan thank's for tracking this down (and for the aurorae patch). Will investigate your patch tomorrow or monday. Then there is the ecclipse "expander" bug. Then I guess we'll make a release :)
Now great news: this bug is that same bug for pidgin becoming unresponsive with time. I can reproduce it with pidgin's buddy list just the same way I did for Eclipse. In fact, you can reproduce it even with twf! So, I'm now happy we know what's going on. Thanks to reporters :)
Git commit 118aba00061ee239b07cff00093a6a7de1293a60 by Ruslan Kabatsayev. Committed on 03/03/2013 at 11:12. Pushed by kabatsayev into branch '1.3'. Add some debug warnings M +6 -0 src/oxygentileset.cpp http://commits.kde.org/oxygen-gtk/118aba00061ee239b07cff00093a6a7de1293a60
Ah well, thank you for looking into this! This was really annoying. Thank you thank you thank you! :)
@Ruslan I don't think the patch is correct, sadly (might create other issues) A tileSet is valid only if *both* (not either) its dimensions are right. I believe there must be a way to fix this "upstream" ... investigating.
... interestingly, the same issue might be present in oxygen@kde ...
ok. I 'think' that the only test that is needed for TileSet validity is in fact _surfaces.size()==9 Which means that proper constructor was called. Now, I also need to add some checks in the rendering method that the surfaces being rendered are valid. ... and then that leaves the fact that deleted tilesets are not properly clearing the allocated resources
Git commit 37d090ece6b81aa4c30a66ac637c571c210af84b by Hugo Pereira Da Costa. Committed on 05/03/2013 at 16:18. Pushed by hpereiradacosta into branch '1.3'. only check pixmap list size in ::isValid M +1 -1 src/oxygentileset.h http://commits.kde.org/oxygen-gtk/37d090ece6b81aa4c30a66ac637c571c210af84b
Commit above prevents the bug to occur, but does not fix the deep-down issue that resources are apparently leaked at tileset deletion. (it just prevents the cache to be filled too quickly). ... still investigating the upstream issue.
Git commit 5149fe325591aa5da370aa374cd2c1ec680a400e by Hugo Pereira Da Costa. Committed on 05/03/2013 at 16:18. Pushed by hpereiradacosta into branch 'gtk3'. only check pixmap list size in ::isValid M +1 -1 src/oxygentileset.h http://commits.kde.org/oxygen-gtk/5149fe325591aa5da370aa374cd2c1ec680a400e
closing down this one since both the not using cached value bug, and the underlying resource leak are fixed.