On 8/17/22 09:07, Ilya Leoshkevich wrote:
Oh my. Well, we could
(1) revert that patch because the premise is wrong,
(2) go with your per-tb clearing,
(3) clear the whole thing with cpu_tb_jmp_cache_clear
Ideally we'd have some benchmark numbers to inform the choice...
FWIW 6f1653180f570 still looks useful.
Reverting it caused 620.omnetpp_s to take ~4% more time.
I ran the benchmark with reduced values in omnetpp.ini so as not to
wait forever, therefore the real figures might be closer to what the
commit message says. In any case this still shows that the patch has
measurable impact.
Thanks for the testing.
I think option (3) will be best for user-only, because mprotect/munmap of existing code
pages is rare -- usually only at process startup.
r~