On 02/15/2018 09:47 AM, Linus Torvalds wrote: > On Thu, Feb 15, 2018 at 5:20 AM, Dave Hansen > <[email protected]> wrote: >> During the switch over to PTI, we seem to have lost our ability to have >> GLOBAL mappings. ... > Did you perhaps re-run any benchmark numbers just to verify? Because > it's always good to back up patches that should improve performance > with actual numbers..
Same test as last time except I'm using all 4 cores on a Skylake desktop instead of just 1. The test is this: > https://github.com/antonblanchard/will-it-scale/blob/master/tests/lseek1.c With PCIDs, lseek()s/second go up around 2% to 3% with the these patches enabling the global bit (it's noisy). I measured it at 3% before, so definitely the same ballpark. That was also before all of Andy's trampoline stuff and the syscall fast path removal.
