Subject: PROBLEM: CPU accounting/scheduling regression in v4.6 CPU scheduling patchset?

Hi,

Since updating my PC to Linux 4.6, I noticed the following problems:

1. CPU-bound tasks which use all CPU cores have a severe impact on responsiveness. For example, the following bash command (which simply starts one busyloop per core) is enough to make the machine almost completely unresponsive:

   for N in $(seq $(nproc)) ; do while true ; do ; done & ; done

2. Nearly all tasks in the process listing are shown with 0% CPU usage, even when they're CPU-bound. The only exceptions are the kernel migration and kthreadd tasks, and occasionally the init process.

I have bisected the problem to commit 1cf4f629d9d246519a1e76c021806f2a51ddba4d ("cpu/hotplug: Move online calls to hotplugged cpu"), which is part of Thomas Gleixner's CPU hotplug refactoring patchset [1]. It introduces both problems described above.

My system is a GIGABYTE X79S-UP5-WIFI motherboard (F5f BIOS) with an i7-4960X CPU, running Arch Linux. I've reproduced with both the distro's kernel config [2], as well as a minimal config for my system. I can reproduce the problems on the latest rc at the moment, v4.7-rc5.

Comparing dmesg output before and after 1cf4f629, I see no notable differences.

I noticed an existing thread "S3 resume regression" [3] referencing this commit, however it describes a different problem. I also found a Bugzilla issue for the zero CPU usage problem [4], however it has no replies.

[1]: https://lkml.org/lkml/2016/2/26/806
[2]: https://aur.archlinux.org/cgit/aur.git/tree/config.x86_64?h=linux-git
[3]: https://lkml.org/lkml/2016/5/11/238
[4]: https://bugzilla.kernel.org/show_bug.cgi?id=120151

Stuff REPORTING-BUGS told me to include:

ver_linux output:
https://dump.v.panteleev.md/616390d43a4c6a3d085acc5eaa390c82/16%3A58%3A08-stdin.txt

/proc/cpuinfo:
https://dump.v.panteleev.md/5dfeba5d7c64028de51d50559b566088/16%3A58%3A49-stdin.txt

/proc/modules:
https://dump.v.panteleev.md/868c0f2b23651be8164975fa5d7e7aab/16%3A59%3A18-stdin.txt

/proc/ioports:
https://dump.v.panteleev.md/5e44aa12cc403dbd783b0273bd3edab4/17%3A01%3A33-stdin.txt

/proc/iomem:
https://dump.v.panteleev.md/110a8fdd0f647fd8d729c54f4f01a3d0/17%3A01%3A49-stdin.txt

"lspci -vvv" output:
https://dump.v.panteleev.md/0c2448fa8a872e34c4555d876b656013/17%3A02%3A18-stdin.txt

/proc/scsi/scsi:
https://dump.v.panteleev.md/6efa007ce74f0bf4ce10ae56690c63de/17%3A02%3A54-stdin.txt

dmesg output:
https://dump.v.panteleev.md/b8a3ba608a914a3d70667dad697dddfb/1467563818.log

I doubt that, but that commit has been a bisect victim before ...

That one says:
* After an hour or less (I have no idea), the top/ps start working
* I do have exactly the same problem with the LTS branch 4.4.14
* With 4.5.4 I cannot reproduce the problem just after booting

I tried to reproduce the issue on a couple of machines, but no luck.

No idea at the moment, but Cc'ed scheduler folks.

Thanks,

tglx

[    0.059350] smpboot: CPU0: Intel(R) Core(TM) i7-4960X CPU @ 3.60GHz (family: 0x6, model: 0x3e, stepping: 0x4)

...

[    0.069372] x86: Booting SMP configuration:
[    0.069375] .... node  #0, CPUs:        #1
[    0.132420] TSC synchronization [CPU#0 -> CPU#1]:
[    0.132427] Measured 170505122558937 cycles TSC warp between CPUs, turning off TSC clock.
[    0.132435] tsc: Marking TSC unstable due to check_tsc_sync_source failed
[    0.136636]   #2  #3  #4  #5  #6  #7  #8  #9 #10 #11
[    0.773136] x86: Booted up 1 node, 12 CPUs

How can TSC be borken on single socket IvyBridge !?