On Fri, Apr 15, 2016 at 02:42:33PM +0100, Dr. David Alan Gilbert wrote: > * Kirill A. Shutemov (kirill@xxxxxxxxxxxxx) wrote: > > On Thu, Apr 14, 2016 at 12:22:30PM -0400, Andrea Arcangeli wrote: > > > Adding linux-mm too, > > > > > > On Thu, Apr 14, 2016 at 01:34:41PM +0100, Dr. David Alan Gilbert wrote: > > > > * Andrea Arcangeli (aarcange@xxxxxxxxxx) wrote: > > > > > > > > > The next suspect is the massive THP refcounting change that went > > > > > upstream recently: > > > > > > > > > As further debug hint, can you try to disable THP and see if that > > > > > makes the problem go away? > > > > > > > > Yep, this seems to be the problem (cc'ing in Kirill). > > > > > > > > 122afea9626ab3f717b250a8dd3d5ebf57cdb56c - works (just before Kirill disables THP) > > > > 61f5d698cc97600e813ca5cf8e449b1ea1c11492 - breaks (when THP is reenabled) > > > > > > > > It's pretty reliable; as you say disabling THP makes it work again > > > > and putting it back to THP/madvise mode makes it break. And you need > > > > to test on a machine with some free ram to make sure THP has a chance > > > > to have happened. > > > > > > > > I'm not sure of all of the rework that happened in that series, > > > > but my reading of it is that splitting of THP pages gets deferred; > > > > so I wonder if when I do the madvise to turn THP off, if it's actually > > > > still got THP pages and thus we end up with a whole THP mapped > > > > when I'm expecting to be userfaulting those pages. > > > > > > Good thing at least I didn't make UFFDIO_COPY THP aware yet so there's > > > less variables (as no user was interested to handle userfaults at THP > > > granularity yet, and from userland such an improvement would be > > > completely invisible in terms of API, so if an user starts doing that > > > we can just optimize the kernel for it, criu restore could do that as > > > the faults will come from disk-I/O, when network is involved THP > > > userfaults wouldn't have a great tradeoff with regard to the increased > > > fault latency). > > > > > > I suspect there is an handle_userfault missing somewhere in connection > > > with trans_huge_pmd splits (not anymore THP splits) that you're doing > > > with MADV_DONTNEED to zap those pages in the destination that got > > > redirtied in source during the last precopy stage. Or more simply > > > MADV_DONTNEED isn't zapping all the right ptes after the trans huge > > > pmd got splitted. > > > > > > The fact the page isn't splitted shouldn't matter too much, all we care > > > about is the pte triggers handle_userfault after MADV_DONTNEED. > > > > > > The userfaultfd testcase in the kernel isn't exercising this case > > > unfortunately, that should probably be improved too, so there is a > > > simpler way to reproduce than running precopy before postcopy in qemu. > > > > I've tested current Linus' tree and v4.5 using qemu postcopy test case for > > both x86-64 and i386 and it never failed for me: > > > > /x86_64/postcopy: first_byte = 7e last_byte = 7d hit_edge = 1 OK > > OK > > /i386/postcopy: first_byte = f6 last_byte = f5 hit_edge = 1 OK > > OK > > > > I've run it directly, setting relevant QTEST_QEMU_BINARY. > > Interesting; it's failing reliably for me - but only with a reasonably > freshly booted machine (so that the pages get THPd). The same here. Freshly booted machine with 64GiB ram. I've checked /proc/vmstat: huge pages were allocated -- Kirill A. Shutemov -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>