> collapse_swapin_single_pte() and collapse_max_ptes_swap() swap a range out > and then require smaps to report exactly the count they asked for. Two > things keep that count from arriving. > > MADV_PAGEOUT is best effort, so the count often turns up a moment late. > > And wait_for_scan() leaves MADV_HUGEPAGE behind, so khugepaged is still
"... leaves range as eligible for collapsing, so khugepaged is still working on it." > working on the range. Collapsing a range with up to max_ptes_swap pages > swapped out means reading them back in, so the daemon empties the swap as > fast as the case fills it. On arm64 with 64K pages max_ptes_swap is 1024 > pages, which is 64M a step, and the case loses: ".. the case looses the race:" > > # Swapout 1024 of 8192 pages... Fail > not ok 10 collapse_max_ptes_swap > > Ask again for up to two seconds, with the range held out of the daemon's > reach while asking. The collapse each case runs next puts MADV_HUGEPAGE > back, so only the setup is affected. "Retry several times for up to two seconds meanwhile holding the range out of the khugepaged reach. The collapse each case runs next restores MADV_NOHUGEPAGE, so only setup is affected" > > If the pages still will not go, skip. A machine with no swap, or swap too "If the pages still won't swap out, skip". > small, full, capped by a memcg or busy with writeback, is not the kernel > under test refusing. An error from madvise() itself still ends the run. > > Assisted-by: Claude-Code:claude-opus-5 > Reviewed-by: Muhammad Usama Anjum <[email protected]> > Tested-by: Muhammad Usama Anjum <[email protected]> > Signed-off-by: Kiryl Shutsemau (Meta) <[email protected]> > > diff --git a/tools/testing/selftests/mm/khugepaged.c > b/tools/testing/selftests/mm/khugepaged.c > index b436b401e2ea..a3d14cc1a2bf 100644 > --- a/tools/testing/selftests/mm/khugepaged.c > +++ b/tools/testing/selftests/mm/khugepaged.c > @@ -241,6 +241,41 @@ static bool check_swap(void *addr, unsigned long size) > return swap; > } > > +/* > + * Page the range out and wait for the swap count to say so. > + * > + * Two things get in the way. MADV_PAGEOUT is best effort: > + * shrink_folio_list() leaves a folio alone when it cannot reclaim it right > + * away, and one still under writeback from an earlier pageout is the common > + * case, so the count the caller asks for arrives a moment later. And a > range > + * an earlier collapse left MADV_HUGEPAGE is one khugepaged is still working > + * on: collapsing a range with up to max_ptes_swap pages swapped out means > + * reading those pages back in, so the daemon undoes the pageout as fast as > it > + * is asked for. Keep the range out of its reach; the collapse the caller > runs > + * next puts MADV_HUGEPAGE back. > + * > + * Failing to get the pages out is the machine's answer, not the kernel's -- > + * swap too small, swap full, a memcg cap, a folio still under writeback -- > so > + * callers skip rather than fail. An error from madvise() is different, and > + * ends the run here. I'd replace this cumbersome comment with > + */ > +static bool swapout_range(void *p, unsigned long size) > +{ > + int i; > + /* prevent khugepaged from swapping into this range */ > + if (madvise(p, size, MADV_NOHUGEPAGE)) > + ksft_exit_fail_perror("madvise(MADV_NOHUGEPAGE)"); > + /* * retry several times because MADV_PAGEOUT is best effort. * sleep between the retries to give a chance for outstanding * writebacks for pages being swapped out to finish. */ -- Sincerely yours, Mike.
