[PATCH 0/6] powerpc/qspinlock: Fix yield latency bug and other

Shrikanth Hegde sshegde at linux.vnet.ibm.com
Wed Oct 18 18:41:42 AEDT 2023



On 10/16/23 6:12 PM, Nicholas Piggin wrote:
> This fixes a long-standing latency bug in the powerpc qspinlock
> implementation that quite a few people have reported and helped
> out with debugging.
> 
> The first patch is a minimal fix that avoids the problem. The
> other patches are streamlining and improvements after the fix.
> 

Hi Nick, Thanks for the fix. This issue has been happening in various
scenarios when there was vCPU contention.

Tested this on Power10 Shared processor LPAR(SPLPAR) based on powerVM.
System has two SPLPARs. on LPAR1 trying various scenarios and
LPAR2 is running constant stress-ng threads consuming 100% its CPU.
LPAR1: 96VP, 64EC and LPAR2 is 32VP, 32EC.

lscpu of LPAR1:
Architecture:            ppc64le
  Byte Order:            Little Endian
CPU(s):                  768


  On-line CPU(s) list:   0-767


Model name:              POWER10 (architected), altivec supported
  Model:                 2.0 (pvr 0080 0200)
  Thread(s) per core:    8

Scenarios tried on LPAR1:
1. run ppc64_cpu --smt=1 and ppc64_cpu --smt=8 to switch between SMT=1
   and SMT=8
2. create a cgroup, assign 5% quota to it and run same number of
   stress-ng as number of CPUs within that cgroup.
3. Run a suite of microbenchmarks such as unixbench, schbench, hackbench
   stress-ng with perf enabled.

baseline was tip/master at 84ab57184ff4 (origin/master, origin/HEAD)
Merge branch into tip/master: 'x86/tdx'

Hard lockup was SEEN in each of the above scenario with baseline.
With this patch series applied hard lockup was NOT SEEN in each of
the above scenario.

So,
Tested-by: Shrikanth Hegde <sshegde at linux.vnet.ibm.com>

> Thanks,
> Nick
> 
> Nicholas Piggin (6):
>   powerpc/qspinlock: Fix stale propagated yield_cpu
>   powerpc/qspinlock: stop queued waiters trying to set lock sleepy
>   powerpc/qspinlock: propagate owner preemptedness rather than CPU
>     number
>   powerpc/qspinlock: don't propagate the not-sleepy state
>   powerpc/qspinlock: Propagate sleepy if previous waiter is preempted
>   powerpc/qspinlock: Rename yield_propagate_owner tunable
> 
>  arch/powerpc/lib/qspinlock.c | 119 +++++++++++++++--------------------
>  1 file changed, 52 insertions(+), 67 deletions(-)
> 


More information about the Linuxppc-dev mailing list