On 1/2/19 11:07 AM, Linus Torvalds wrote: > On Wed, Jan 2, 2019 at 9:55 AM Jens Axboe <[email protected]> wrote: >> >> A previous commit deemed it safe to use __set_current_state() for IRQ >> driven O_DIRECT, but that's isn't necessarily the case. Be safer and >> only apply that optimization to polled IO, where we know the the task is >> going to find the completions itself. > > No. > > We're just reverting 849a370016a5. > > There is absolutely no point in trying to avoid a memory barrier that > is maybe 15 CPU cycles. > > Stop this craziness. The optimization is garbage. If you want to save > 15 cycles, get rid of code, don't add new code in an area where the > block maintainers have already shown that they can't get it right.
Fine, we'll just revert the damn thing then. -- Jens Axboe
