5.5-rc1 regression with BNXT firmware driver

[Valentin Schneider <valentin.schneider@arm.com>]

Hi,

I've been hunting down some hackbench regression between 5.4-rc8 and 5.5-rc1
on my Juno r0, one of the offenders seems to be:

  246880958ac9 ("firmware: broadcom: add OP-TEE based BNXT f/w manager")

This is tested on a kernel built with defconfig (TEE_BNXT_FW gets selected)
and with:

  echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
  echo performance > /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor
  ./perf stat --null --sync --repeat 200 ./hackbench

The regression is easily reproducible on my end, this is 3 runs of the above
comparing the patch and its parent:

  -PATCH:
  0.71062 +- 0.00150 seconds time elapsed  ( +-  0.21% )                                                                                                                                                      
  0.71121 +- 0.00181 seconds time elapsed  ( +-  0.25% )                                                                                                                                                      
  0.71277 +- 0.00181 seconds time elapsed  ( +-  0.25% )  

  +PATCH:
  0.72556 +- 0.00174 seconds time elapsed  ( +-  0.24% )                                                                                                                                                      
  0.72695 +- 0.00192 seconds time elapsed  ( +-  0.26% )                                                                                                                                                      
  0.72559 +- 0.00178 seconds time elapsed  ( +-  0.25% ) 


AIUI Vincent found something different while hunting down a similar
regression:

  df323337e507 ("apparmor: Use a memory pool instead per-CPU caches")

but it seems this one is another cause. Seeing as this involves security
stuff the overhead may be acceptable, nevertheless now that I have some
reproducer I figured I'd send this out.

Cheers,
Valentin