On Fri, Apr 22, 2011 at 1:00 AM, Mel Gorman <mgorman@xxxxxxx> wrote: > On Thu, Apr 21, 2011 at 11:26:36PM +0900, Minchan Kim wrote: >> On Thu, Apr 21, 2011 at 12:08:41PM +0100, Mel Gorman wrote: >> > On Thu, Apr 21, 2011 at 03:59:47PM +0900, Minchan Kim wrote: >> > > Hi Mel, >> > > >> > > On Fri, Apr 15, 2011 at 7:12 PM, Mel Gorman <mgorman@xxxxxxx> wrote: >> > > > With transparent hugepage support, handle_mm_fault() has to be careful >> > > > that a normal PMD has been established before handling a PTE fault. To >> > > > achieve this, it used __pte_alloc() directly instead of pte_alloc_map >> > > > as pte_alloc_map is unsafe to run against a huge PMD. pte_offset_map() >> > > > is called once it is known the PMD is safe. >> > > > >> > > > pte_alloc_map() is smart enough to check if a PTE is already present >> > > > before calling __pte_alloc but this check was lost. As a consequence, >> > > > PTEs may be allocated unnecessarily and the page table lock taken. >> > > > Thi useless PTE does get cleaned up but it's a performance hit which >> > > > is visible in page_test from aim9. >> > > > >> > > > This patch simply re-adds the check normally done by pte_alloc_map to >> > > > check if the PTE needs to be allocated before taking the page table >> > > > lock. The effect is noticable in page_test from aim9. >> > > > >> > > > AIM9 >> > > >        Â2.6.38-vanilla 2.6.38-checkptenone >> > > > creat-clo   Â446.10 ( 0.00%)  424.47 (-5.10%) >> > > > page_test    38.10 ( 0.00%)  Â42.04 ( 9.37%) >> > > > brk_test    Â52.45 ( 0.00%)  Â51.57 (-1.71%) >> > > > exec_test   Â382.00 ( 0.00%)  456.90 (16.39%) >> > > > fork_test    60.11 ( 0.00%)  Â67.79 (11.34%) >> > > > MMTests Statistics: duration >> > > > Total Elapsed Time (seconds)        Â611.90  Â612.22 >> > > > >> > > > (While this affects 2.6.38, it is a performance rather than a >> > > > functional bug and normally outside the rules -stable. While the big >> > > > performance differences are to a microbench, the difference in fork >> > > > and exec performance may be significant enough that -stable wants to >> > > > consider the patch) >> > > > >> > > > Reported-by: Raz Ben Yehuda <raziebe@xxxxxxxxx> >> > > > Signed-off-by: Mel Gorman <mgorman@xxxxxxx> >> > > > -- >> > > > Âmm/memory.c |  Â2 +- >> > > > Â1 file changed, 1 insertion(+), 1 deletion(-) >> > > > >> > > > diff --git a/mm/memory.c b/mm/memory.c >> > > > index 5823698..1659574 100644 >> > > > --- a/mm/memory.c >> > > > +++ b/mm/memory.c >> > > > @@ -3322,7 +3322,7 @@ int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, >> > > >     * run pte_offset_map on the pmd, if an huge pmd could >> > > >     * materialize from under us from a different thread. >> > > >     */ >> > > > -    if (unlikely(__pte_alloc(mm, vma, pmd, address))) >> > > > +    if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address)) >> > > >        Âreturn VM_FAULT_OOM; >> > > >    Â/* if an huge pmd materialized from under us just retry later */ >> > > >    Âif (unlikely(pmd_trans_huge(*pmd))) >> > > > >> > > >> > > Reviewed-by: Minchan Kim <minchan.kim@xxxxxxxxx> >> > > >> > > Sorry for jumping in too late. I have a just nitpick. >> > > >> > >> > Better late than never :) >> > >> > > We have another place, do_huge_pmd_anonymous_page. >> > > Although it isn't workload of page_test, is it valuable to expand your >> > > patch to cover it? >> > > If there is workload there are many thread and share one shared anon >> > > vma in ALWAYS THP mode, same problem would happen. >> > >> > We already checked pmd_none() in handle_mm_fault() before calling >> > into do_huge_pmd_anonymous_page(). We could race for the fault while >> > attempting to allocate a huge page but it wouldn't be as severe a >> > problem particularly as it is encountered after failing a 2M allocation. >> >> Right you are. Fail ot 2M allocation would affect as throttle. >> Thanks. >> >> As I failed let you add the check, I have to reveal my mind. :) >> Actually, what I want is consistency of the code. > > This is a stronger arguement than as a performance fix. I was concerned > that if such a check was added that it would confuse someone in a years > time trying to figure out why the pmd_none check was really necessary. > >> The code have been same in two places but you find the problem in page_test of aim9, >> you changed one of them slightly. I think in future someone will >> have a question about that and he will start grep git log but it will take >> a long time as the log is buried other code piled up. >> > > Fair point. > >> I hope adding the comment in this case. >> >>     /* >>     Â* PTEs may be allocated unnecessarily and the page table lock taken. >>     Â* The useless PTE does get cleaned up but it's a performance hit in >>     Â* some micro-benchmark. Let's check pmd_none before __pte_alloc to >>     Â* reduce the overhead. >>     Â*/ >> -    if (unlikely(__pte_alloc(mm, vma, pmd, address))) >> +    if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address)) >> > > I think a better justification is > >    Â/* >     * Even though handle_mm_fault has already checked pmd_none, we >     * have failed a huge allocation at this point during which a >     * valid PTE could have been inserted. Double check a PTE alloc >     * is still necessary to avoid additional overhead >     */ > Hmm. If we disable thp, the comment about failing a huge allocation. was not true. So I prefer mine :) But Andrea suggested defining new pte_alloc which checks pmd_none internally for code consistency POV. In such case, I have no concern about comment. Is it okay? >> If you mind it as someone who have a question can find the log at last >> although he need some time, I wouldn't care of the nitpick any more. :) >> It's up to you. >> > > If you want to create a new patch with either your comment or mine > (whichever you prefer) I'll add my ack. I'm about to drop offline > for a few days but if it's still there Tuesday, I'll put together an > appropriate patch and submit. I'd keep it separate from the other patch > because it's a performance fix (which I'd like to see in -stable) where > as this is more of a cleanup IMO. Okay. Thanks, Mel. -- Kind regards, Minchan Kim -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href