On Sun, Nov 01, 2015 at 06:14:43PM -0800, Eric Dumazet wrote:

On Mon, 2015-11-02 at 00:24 +0000, Al Viro wrote:

quoted

This ought to be a bit cleaner.  Eric, could you test the variant below on your
setup?

Sure !

5 runs of :
lpaa24:~# taskset ff0ff ./opensock -t 16 -n 10000000 -l 10

total = 4386311
total = 4560402
total = 4437309
total = 4516227
total = 4478778

Umm...  With Linus' variant it was what, around 4000000?  +10% or so, then...

With 48 threads :

./opensock -t 48 -n 10000000 -l 10
total = 4940245
total = 4848513
total = 4813153
total = 4813946
total = 5127804

And that - +40%?  Interesting...  And it looks like at 48 threads you are
still seeing arseloads of contention, but apparently less than with Linus'
variant...  What if you throw the __clear_close_on_exec() patch on
top of that?

Looks like it's spending less time under ->files_lock...  Could you get
information on fs/file.o hotspots?