On Thu, Jul 16, 2020 at 1:55 PM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Thu, Jul 16, 2020 at 6:16 AM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > > > > It can also lead to performance regression: for small mremap() if only one > > side of the range got aligned and there's no PMD_SIZE range to move, > > kernel will still iterate over PTEs, but it would need to handle more > > pte_none()s than without the patch. > > Ack, I've dropped the patch from my queue of experiments, because it > doesn't work for the case I wanted to do, and the other cases could > regress, as you say. > > Plus considering how many problems that patch had, I decided it wasn't > as simple as I initially thought it would be anyway ;) > > Joel - while it's gone from my mind, if you're still interested in > this, maybe you can do something _similar_ that patch, except perhaps > also start out checking that the initial size is large enough for this > to make sense even when one of the sides doesn't align, for example. > > (It might be as simple as checking that the initial 'len' is at least > PMD_SIZE - then you're guaranteed that whichever side gets aligned, > it's not doing extra work because the other side didn't). Hi Linus, Yes I'm quite interested in doing a similar patch and also adding some test cases to make sure it always works properly. Probably I can add some kselftest cases doing mremap on various ranges. I think it will be a great optimization to trigger the PMD move more often. I'll work on it and get back hopefully soon on this, Thanks! - Joel
