Re: [linus:master] [x86] 4817f70c25: stress-ng.mmapaddr.ops_per_sec 63.0% regression

[Peter Zijlstra]

On Tue, Jan 28, 2025 at 11:05:17AM +0100, David Hildenbrand wrote:
> On 28.01.25 10:57, kernel test robot wrote:
> > 
> > hi, Qi Zheng,
> > 
> > this is more a FYI report than a regression report.
> > 
> > by 4817f70c25, parent/4817f70c25 configs have below diff,
> > 
> > --- /pkg/linux/x86_64-rhel-9.4/gcc-12/718b13861d2256ac95d65b892953282a63faf240/.config  2025-01-27 16:20:43.419181382 +0800
> > +++ /pkg/linux/x86_64-rhel-9.4/gcc-12/4817f70c25b63ee5e6fd42d376700c058ae16a96/.config  2025-01-26 09:27:16.848625105 +0800
> > @@ -1236,6 +1236,8 @@ CONFIG_IOMMU_MM_DATA=y
> >   CONFIG_EXECMEM=y
> >   CONFIG_NUMA_MEMBLKS=y
> >   CONFIG_NUMA_EMU=y
> > +CONFIG_ARCH_SUPPORTS_PT_RECLAIM=y
> > +CONFIG_PT_RECLAIM=y
> > 
> >   #
> >   # Data Access Monitoring
> > 
> > 
> > this report seems show the impact of PT_CLAIM feature for this stress-ng case.
> > 
> > To us, this is not a code logic regression, but is kind of 'regression' from a
> > new feature. anyway, below full report just FYI.
> 
> mmapaddr test case seems to mostly do mmap+munmap. No obvious sign of
> MADV_DONTNEED, unless buried somewhere :)
> 
> So either
> 
> (1) The series is reclaiming page tables outside of MADV_DONTNEED, which
>     it shouldn't -- in particular not during munmap() where that happens
>     already using the "ordinary" page table removal code for removed
>     VMAs.
> 
> (2) This is the effect of MMU_GATHER_RCU_TABLE_FREE that gets selected?
> 
> 
> I recall a recent series to select MMU_GATHER_RCU_TABLE_FREE on x86
> unconditionally (@Peter, @Rik).

Those changes should not have made it to Linus yet.

/me updates git and checks...

nope, nothing changed there ... yet