lkml.org 
[lkml]   [2009]   [Jul]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    Date
    From
    SubjectRe: [PATCH] mm: Warn once when a page is freed with PG_mlocked set V2
    On Thu, Jul 23, 2009 at 10:23:16AM -0700, Andrew Morton wrote:
    > On Thu, 23 Jul 2009 11:29:39 +0100 Mel Gorman <mel@csn.ul.ie> wrote:
    >
    > > On Wed, Jul 22, 2009 at 04:06:49PM -0700, Andrew Morton wrote:
    > > > On Wed, 15 Jul 2009 10:31:54 -0400 (EDT)
    > > > Christoph Lameter <cl@linux-foundation.org> wrote:
    > > >
    > > > > On Wed, 15 Jul 2009, Mel Gorman wrote:
    > > > >
    > > > > > -static inline int free_pages_check(struct page *page)
    > > > > > +static inline int free_pages_check(struct page *page, int wasMlocked)
    > > > > > {
    > > > > > + WARN_ONCE(wasMlocked, KERN_WARNING
    > > > > > + "Page flag mlocked set for process %s at pfn:%05lx\n"
    > > > > > + "page:%p flags:0x%lX\n",
    > > > > > + current->comm, page_to_pfn(page),
    > > > > > + page, page->flags|__PG_MLOCKED);
    > > > > > +
    > > > > > if (unlikely(page_mapcount(page) |
    > > > >
    > > > > There is already a free_page_mlocked() that is only called if the mlock
    > > > > bit is set. Move it into there to avoid having to run two checks in the
    > > > > hot codee path?
    > > >
    > > > Agreed.
    > > >
    > > > This patch gratuitously adds hotpath overhead. Moving the change to be
    > > > inside those preexisting wasMlocked tests will reduce its overhead a lot.
    > > >
    > >
    > > It adds code duplication then, one of which is in a fast path.
    > >
    > > > As it stands, I'm really doubting that the patch's utility is worth its
    > > > cost.
    > > >
    > >
    > > I'm happy to let this one drop. It seemed like it would be nice for debugging
    > > while there are still corner cases where mlocked pages are getting freed
    > > instead of torn down but we already account for that situation occuring. While
    > > I think it'll be tricky to spot, it's probably preferable to having warnings
    > > spew out onto dmesg.
    >
    > If we do in it the way which Christoph recommends, the additional
    > overhead is miniscule?
    >

    Yep, it should be. I misinterpreted what you said with doubting the patch's
    utility. The cost as it was was too high rather than the warning itself was
    useless. When moved to free_page_mlock(), patch looks like;

    ==== CUT HERE ====
    mm: Warn once when a page is freed with PG_mlocked set V3

    Changelog since V2
    o Move warning to free_page_mlock()
    o Use %#lx instead of 0x%lX in printf format string

    Changelog since V1
    o Remove unnecessary branch

    When a page is freed with the PG_mlocked set, it is considered an unexpected
    but recoverable situation. A counter records how often this event happens
    but it is easy to miss that this event has occured at all. This patch warns
    once when PG_mlocked is set to prompt debuggers to check the counter to
    see how often it is happening.

    Signed-off-by: Mel Gorman <mel@csn.ul.ie>
    ---
    mm/page_alloc.c | 5 +++++
    1 file changed, 5 insertions(+)

    diff --git a/mm/page_alloc.c b/mm/page_alloc.c
    index b8283e8..d3d0707 100644
    --- a/mm/page_alloc.c
    +++ b/mm/page_alloc.c
    @@ -488,6 +488,11 @@ static inline void __free_one_page(struct page *page,
    */
    static inline void free_page_mlock(struct page *page)
    {
    + WARN_ONCE(1, KERN_WARNING
    + "Page flag mlocked set for process %s at pfn:%05lx\n"
    + "page:%p flags:%#lx\n",
    + current->comm, page_to_pfn(page),
    + page, page->flags|__PG_MLOCKED);
    __dec_zone_page_state(page, NR_MLOCK);
    __count_vm_event(UNEVICTABLE_MLOCKFREED);
    }

    \
     
     \ /
      Last update: 2009-07-24 12:39    [W:0.024 / U:153.328 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site