> On Jul 19, 2020, at 3:17 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > > Andy Lutomirski <luto@xxxxxxxxxx> writes: >>> On Sat, Jul 18, 2020 at 7:16 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: >>> Andy Lutomirski <luto@xxxxxxxxxx> writes: >>>> FWIW, TIF_USER_RETURN_NOTIFY is a bit of an odd duck: it's an >>>> entry/exit word *and* a context switch word. The latter is because >>>> it's logically a per-cpu flag, not a per-task flag, and the context >>>> switch code moves it around so it's always set on the running task. >>> >>> Gah, I missed the context switch thing of that. That stuff is hideous. >> >> It's also delightful because anything that screws up that dance (such >> as failure to do the exit-to-usermode path exactly right) likely >> results in an insta-root-hole. If we fail to run user return >> notifiers, we can run user code with incorrect syscall MSRs, etc. > > Looking at it deeper, having that thing in the loop is a pointless > exercise. This really wants to be done _after_ the loop. > As long as we’re confident that nothing after the loop can set the flag again. > Thanks, > > tglx
