On March 18, 2019 11:10:22 AM PDT, Linus Torvalds <torva...@linux-foundation.org> wrote: >On Mon, Mar 18, 2019 at 10:51 AM Peter Zijlstra <pet...@infradead.org> >wrote: >> >> How about I do a patch that schedules EFLAGS for both 32bit and >64bit, >> mark this for backporting to infinity. >> >> And then at the end, after the objtool-ac bits land, I do a patch >> removing the EFLAGS scheduling for x86_64. > >Sounds sane to me. > >And we can make it AC-conditional if it's actually shown to be visible >from a performance standpoint. > >But iirc pushf/popf isn't really that expensive - in fact I think it's >pretty cheap when system flags don't change. Which would be the common >case unless you'd also make the popf do the irq restore part and >simply make finish_lock_switch() re-enable irq's by doing an >irqrestore? > >I think popf is like 20 cycles or something (and pushf is just single >cycles). Probably not worth worrying about in the task switch context. > > Linus
Yes, the problem isn't scheduling per se but the risk of hiding problems. -- Sent from my Android device with K-9 Mail. Please excuse my brevity.
