On Tue, Jan 05, 2021 at 01:03:48PM -0500, Andrea Arcangeli wrote: > On Tue, Jan 05, 2021 at 04:37:27PM +0100, Peter Zijlstra wrote: > > (your other email clarified this point; the COW needs to copy while > > holding the PTL and we need TLBI under PTL if we're to change this) > > The COW doesn't need to hold the PT lock, the TLBI broadcast doesn't > need to be delivered under PT lock either. > > Simply there need to be a TLBI broadcast before the copy. The patch I > sent here https://lkml.kernel.org/r/X+QLr1WmGXMs33Ld@xxxxxxxxxx that > needs to be cleaned up with some abstraction and better commentary > also misses a smp_mb() in the case flush_tlb_page is not called, but > that's a small detail. That's horrific crap. All of that tlb-pending stuff is batshit, and this makes it worse. > > And I'm thinking the speculative page fault series steps right into all > > this, it fundamentally avoids mmap_sem and entirely relies on the PTL. > > I thought about that but that only applies to some kind of "anon" page > fault. That must be something new; it used to handle all faults. I specifically spend quite a bit of time getting the file crud right (which Linus initially fingered for being horrible broken). SPF fundamentally elides the mmap_sem, which Linus said must serialize faults. > Here the problem isn't just the page fault, the problem is not to > regress clear_refs to block on page fault I/O, and all IIRC we do the actual reads without any locks held, just like VM_FAULT_RETRY does today. You take the fault, find you need IO, drop locks, do IO, retake fault. > MAP_PRIVATE/MAP_SHARED filebacked faults bitting the disk to read > /usr/ will still prevent clear_refs from running (and the other way > around) if it has to take the mmap_sem for writing. > > I don't look at the speculative page fault for a while but last I > checked there was nothing there that can tame the above major > regression from CPU speed to disk I/O speed that would be inflicted on > both clear_refs on huge mm and on uffd-wp. All of the clear_refs nonsense is immaterial to SPF. Also, who again cares about clear_refs? Why is it important?
