Tuesday, 5 May 2020

linux-5.6-ck2, MuQSS version 0.2 for linux-5.6

Announcing a new -ck release, 5.6-ck2  with the latest version of the Multiple Queue Skiplist Scheduler, version 0.2. These are patches designed to improve system responsiveness and interactivity with specific emphasis on the desktop, but configurable for any workload. This is a maintenance release to address a build failure on -ck1 when built with full dynticks, and fix some cosmetic CPU load accounting issues. Upgrading is not required unless you are affected by the aforementioned issues or are rebuilding for a new stable release. It's worth pointing out that changing the reported load might have repercussions with how CPU frequency scaling behaves.



linux-5.6-ck2:
-ck2 patch:

Git tree:

MuQSS only:
Download:
Git tree:
Web: http://kernel.kolivas.org


As an aside, it has been brought to my attention that the MESA code uses SCHED_IDLEPRIO for what it considers low priority threads. In the mainline kernel this only makes them lower priority than regular tasks, but on MuQSS which has true idle scheduling, it can potentially lead to stalls under conditions of load. Once a thread has stalled for an extended time, it is possible that it will not progress normally depending on how the code expects to run. This could lead to GUI stalls in applications that use MESA, of which there are quite a few now, such as firefox. I've been considering submitting a change to the MESA code in the hope they approach this differently, but I am a pragmatist and expect the turnaround time and acceptability of the changes may be drawn out and unsatisfactory. So I am considering alternatively softening the idle scheduling and making it configurable to behave more like mainline's by default and optionally be set to be true idle scheduling. In the meantime, I've prepared some hacked mesa packages for those on ubuntu 20.04 variants that disable this behaviour, but this is a cludge only for the time being:

mesa-ubuntu20.04 packages

Here is a much better patch for Mesa that converts threads to nice 19 SCHED_BATCH instead:

0001-Linux-Change-minimum-priority-threads-from-SCHED_IDL.patch


 Enjoy!
お楽しみ下さい
-ck

52 comments:

  1. Could you link to the mesa email thread or bug you've created, please?

    ReplyDelete
    Replies
    1. I haven't made a thread or bug report (yet?)

      Delete
  2. Thank you very much.

    ReplyDelete
  3. Could you clarify what was changed for the custom mesa packages, so that could possibly be implemented for other distributions as well? I tried digging trough the package files and comparing to the original ubuntu package files, but couldn't figure this out.

    ReplyDelete
    Replies
    1. Just this https://aargh.no-ip.org/u_queue-no-sched-idle.patch

      Delete
    2. Thank you.

      From a quick glance at the code, it looks like currently Mesa only uses SCHED_ISO when creating shader caches, so I'd imagine there wouldn't be any real world difference between the SCHED_ISO implementations in -ck and mainline kernels in this case. In theory -ck kernel's SCHED_ISO may even result in better performance, since the cache creation would leave more resources available for more important things. (Note that I'm not any kind of expert in Mesa or GPU stuff in general, so I could be mistaken on this).

      In any case, this would mainly have any effect with some games, and only on the first run when the caches are being created (which often happens when loading the game, rather than during the actual gameplay).

      Of course, there's a chance that Mesa, or other applications would use SCHED_ISO somewhere where the different implementations actually matter, so it probably wouldn't hurt to have an upstream-compatible idle scheduling solution available.

      Nevertheless, I think recompiling Mesa over this is probably not worth the trouble at the moment.

      Delete
    3. As far as I can see, it uses SCHED_IDLE which is the issue entirely. Not sure what you're looking at. There is no use of SCHED_ISO.

      Delete
    4. See:
      https://cgit.freedesktop.org/mesa/mesa/tree/src/util/u_queue.c#n334

      Delete
    5. Sorry, I meant SCHED_IDLE but my tired brain messed up.

      You can replace SCHED_ISO with SCHED_IDLE on every instance on the above post.

      Delete
    6. This has been confirmed to create GUI stalls which go away when the scheduling policy is removed, so whilst it sounds good in principle, in reality it is problematic. SCHED_IDLE in mainline is very different to that in MuQSS.

      Delete
    7. Interesting. Thanks for clarification.

      Is there some specific test case for the GUI stalls? Or other information/discussions about this somewhere?

      Delete
    8. Here is a more comprehensive solution to the problem which still makes the most of Linux's ability to set policy and priority by thread only by switching to nice 19 SCHED_BATCH instead:
      http://ck.kolivas.org/mesa-ubuntu20.04/0001-Linux-Change-minimum-priority-threads-from-SCHED_IDL.patch

      If I can ever figure out the mesa submission process I will try to submit a merge request.

      Delete
    9. And here's the merge request. Fingers crossed:

      https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4912

      Delete
    10. Thanks! https://www.phoronix.com/scan.php?page=news_item&px=Mesa-GUI-Stalls-Fixed-Kolivas

      Delete
  4. may be correct?
    --- pid_t tid = gettid();
    +++ pid_t tid = syscall(SYS_gettid);

    ReplyDelete
    Replies
    1. I see your point. gettid() was added to glibc 2.30. I suppose it's possible someone will be building a new mesa on an old distribution which would require the syscall wrapper.

      Delete
    2. Thanks, I've added it to the git merge request.

      Delete
    3. The mesa devs were very accommodating and helped me negotiate their convoluted merge process and the code is now in the master mesa git which is great news.

      Delete
    4. cool.
      when to expect release with these changes?
      will these changes be backported to older versions of the mesa?

      Delete
    5. No idea of their process of merging to stable releases. I seriously doubt it will be backported unless there's a strong push by someone saying it's a bugfix. I suspect we won't be seeing this change in stable packages in distros till their next release cycle unless you're on a rolling distro.

      Delete
  5. So, does this mean it's "ok" to use full dynticks config? (tickless), or is it just to fix compile problems for those that insist on using tickless? :)

    I seem to remember some posts a while back about that kinda concluded that it was not really recommended with ck/muqss?

    ReplyDelete
    Replies
    1. Yes it's fine now with -ck2, it was always a problem with forced context tracking, but is still unlikely to be advantageous.

      Delete
  6. Somehow, I'm never able to boot both ck1 and ck2 kernel on my new ThinkPad X390 (i7 Gen10). I tried with/without irqthread and both idle tickless and full dynticks, all combinations with no success.

    ReplyDelete
    Replies
    1. Worst case scenario, try disabling all runqueue sharing which is the riskiest part of the code. Add the following kernel command line to test it without needing to rebuild:
      rqshare=none

      Delete
    2. I press the power button a few times for a short while when booting stalls, it will boot up on my machine eventually.

      Delete
    3. I tried rqshare=none and also other option (e.g. smt, all) nothing work. I also tried them with nothreadirqs still unable to boot.

      Delete
    4. I remember having similar issue with some combination of CONFIG_HZ, threadirqs, rqshare, rr_interval & iso_cpu settings.
      You can also try with different HZ (100 or 1000), periodic ticks (nohz=off command line parameter). Or maybe some combination of nohz, 1000HZ and muqss-only (without other patches).
      There also might be some distro or config-related problem.

      Delete
    5. Sounds strange cwt.
      Are you sure you use the same kernel (source/patches) as the one that you normally use?

      Just asking cos if you are using a distro kernel (eg. Ubuntu kernel), and do not use the same source you could run into some patched kernel driver or whatnot that is not added to your "custom" kernel with MuQSS.

      Delete
    6. I use vanilla kernel with minimal patches, one from https://github.com/graysky2/kernel_gcc_patch/blob/master/enable_additional_cpu_optimizations_for_gcc_v9.1%2B_kernel_v5.5%2B.patch to enable native optimization, another one is https://github.com/dolohow/uksm/blob/master/v5.x/uksm-5.6.patch.

      with these two patches and config from distro (fedora 32), I just make some changes like "Preemptible Kernel" and remove some drivers for hardware that don't exist on my laptop, the kernel is working perfectly. However after added ck patch or even just MuQSS patch, the kernel won't boot.

      Delete
    7. Can You remove "quiet" from the kernel commandline, try to boot and share what's on the screen when it crashes?

      Delete
    8. Yeah, "it won't boot" mean it won't boot at all, not boot for a while and crash. However, here the screen, just a cursor on the top left. http://imgur.com/gallery/kn7HlhA

      Delete
    9. The only time recently i had that happen, was when i tried to compile the kernel with gcc-10. Just black screen with the "cursor" on the left corner.

      Did not bother to debug it more, cos gcc-10 i use is not really "production" ready i guess (Ubuntu testing PPA). I am only mentioning this, since i think Fedora 32 got gcc-10 now?

      Something about CONFIG_STACKPROTECTOR_STRONG or some shit, but it should then happen anyways, and not only when using MuQSS i guess?

      Delete
    10. Yes, Fedora 32 came with gcc10, but with the same compiler I can build and run vanilla kernel. I will try again without native optimization or with gcc9.

      Delete
    11. Some information about gcc-10 vs. kernel
      https://github.com/Frogging-Family/linux-tkg/issues/7
      https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/commit?id=f670269a42bfdd2c83a1118cc3d1b475547eac22

      Mentioning the first one, cos this is a custom kernel with different schedulers - like muqss/bmq/pds and the last one actually mentioning that 5.6 kernel is not 100% ready for gcc-10.

      https://lore.kernel.org/lkml/20200423161126.GD26021@zn.tnic/

      I have not tried it, but if you find the kernel boots fine when compiled with gcc-9 (like it does for me), you could try skim through this and test the patch with gcc-10?

      Delete
    12. I also cannot boot with CK kernel 5.6 on Fedora 32. I always build a custom kernel, and vanilla (5.6/5.7) / BMQ (5.6) kernels both work fine. With CK kernel using GCC or LLVM I cannot boot at all, just gets a black screen, removing quiet option does nothing the kernel never outputs anything.

      I also tried the GCC patch and it does not help, but given it didn't work with LLVM I guess that makes sense.

      Delete
    13. You could try the 5.6-muqss tree. There are some fixes in there.

      Delete
    14. Fedora 32 cannot install and run gcc9 on it easily, so I tried something by built the kernel with clang10 instead, and the result was still the same that ck kernel cannot boot. However, vanilla kernel that built with clang10 can boot and run perfectly. I also tried BMQ patch with gcc10 and the kernel is working too. Is it possible that there is something in CK or MuQSS patch that cause (or trigger) a problem with gcc10?

      Delete
    15. Using 5.6-muqss tree makes no difference, still unable to boot.

      Delete
    16. Arch Linux just updated to gcc-10.1, and I was able to successfully build and boot 5.6.13-ck1 kernel (built with -O3 -march=native, and no other changes) with it.

      Delete
    17. This comment has been removed by the author.

      Delete
  7. This comment has been removed by the author.

    ReplyDelete
  8. This comment has been removed by the author.

    ReplyDelete
  9. Hi.
    I have panic when kernel is building with option CONFIG_PSI=y
    https://pastebin.com/vnV1GkdD

    ReplyDelete
    Replies
    1. Did you try to boot with psi=0? dmesg don't report about psi with this option
      In config
      CONFIG_PSI_DEFAULT_DISABLED=y

      ---

      Currently schedutil also doesn't work correctly with alternative schedulers (muqss, pds, bmq)

      Delete
  10. Hi,

    I get an error with the eth0 interface in my USB-C adapter when using the linux-ck kernel in Archlinux. The problem is not present with stock kernels, so I guess it may be due to the ck patchset. The issue is described here:

    https://bbs.archlinux.org/viewtopic.php?pid=1907858#p1907858

    I can provide additional info if needed. Thanks.

    ReplyDelete
    Replies
    1. If you've built with irqthreads force enabled, try disabling them. Add the nothreadirqs parameter to your kernel command line. Otherwise, try the extra bugfixes committed to the muqss git tree.

      Delete
    2. Thanks Con, I tried with the nothreadirqs flag to no avail. I never built the patched kernel so I don't know how to try the muqss git. If there's nothing else that I can try then I'll stick with the stock kernel and see if the new patchset for Linux 5.7 will fix the problem.

      Delete
  11. It turned out that the problem was not related to CK patchset, and was present also in vanilla kernel. The cuplrit was the powersave for the ethernet adapter of the USB-C dock, disabling it by means of a udev rule fixed the problem. I'm sure I had no problem at the end of February, but even reverting to kernel/firmware versions at that date didn't fix the issue (just gave different, mitigated phenomenology). Nevertheless, the patchset was ok, so sorry for the noise and thanks for the support.

    ReplyDelete