Re: [x86/modules] f2c65fb322: will-it-scale.per_process_ops -2.9% regression

[Dave Hansen]

On 7/18/19 10:59 AM, Nadav Amit wrote:
> I donât understand how this patch has any impact on this workload.
> 
> I ran it and set a function tracer for any function that is impacted by this
> patch:
> 
>   # cd /sys/kernel/debug/tracing
>   # echo text_poke_early > set_ftrace_filter
>   # echo module_alloc >> set_ftrace_filter
>   # echo bpf_int_jit_compile >> set_ftrace_filter
>   # tail -f trace
> 
> Nothing came up. Can you please check if you see any of them invoked on your
> setup? Perhaps you have some bpf filters being installed, although even then
> this is a one-time (small) overhead for each process invocation.

I think the direct map's structure changed.  I noticed the following the
0day data:

> 7298e24f904224fa f2c65fb3221adc6b73b0549fc7b 
> ---------------- --------------------------- 
>        fail:runs  %reproduction    fail:runs
>            |             |             |    
>            :4           50%           2:4     dmesg.WARNING:at#for_ip_interrupt_entry/0x
>          %stddev     %change         %stddev
>              \          |                \  
>   10541056 Â  7%     +12.1%   11820032        meminfo.DirectMap2>  2.909e+08           -12.0%   2.56e+08 Â  2%  perf-stat.i.iTLB-load-misses
>       1876           +10.4%       2071 Â  2%  perf-stat.i.instructions-per-iTLB-miss
>       1872           +10.4%       2068 Â  2%  perf-stat.overall.instructions-per-iTLB-miss
>  2.899e+08           -12.0%  2.551e+08 Â  2%  perf-stat.ps.iTLB-load-misses

So something with the direct map changed as did iTLB miss behavior.