[lkml]   [2011]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCH] ksm: use FAULT_FLAG_ALLOW_RETRY in breaking COW
    On Mon, Nov 21, 2011 at 12:16 PM, Hugh Dickins <> wrote:
    > On Sat, 19 Nov 2011, Hillf Danton wrote:
    >> The flag, FAULT_FLAG_ALLOW_RETRY, was introduced by the patch,
    >>       mm: retry page fault when blocking on disk transfer
    >>       commit: d065bd810b6deb67d4897a14bfe21f8eb526ba99
    >> for reducing mmap_sem hold times that are caused by waiting for disk
    >> transfers when accessing file mapped VMAs.
    >> To break COW, handle_mm_fault() is repeated with mmap_sem held, where
    >> the introduced flag could be used again.
    >> The straight way is to add changes in break_ksm(), but the function could be
    >> under write-mode mmap_sem, so it has to be dupilcated.
    >> Signed-off-by: Hillf Danton <>
    > Thank you for making the patch; but unless I'm mistaken - please
    > correct me if so - I think it's better to keep break_cow() simple
    > than add special FAULT_FLAG_ALLOW_RETRY handling there.  Do you
    > have any evidence that its down_read of mmap_sem is a problem in
    > some workload?  I sense that you're using it "because it's there".
    > I'm sceptical on several grounds.
    > One, break_cow() is itself only called on an "error" path: not
    > really an error, but when KSM's bet that it can merge pages turns
    > out to be wrong before it can complete the merge; not a rare case,
    > but not on the hot path.
    > Two, break_ksm()'s loop is required for correctness, but it
    > is a rare case that it actually needs to go round a second time.
    > The typical case it's needed (am I forgetting a more common one?)
    > is when userspace access flips a pte bit in between handle_pte_fault()
    > noting faulting pte, and the chosen fault handler checking pte_same()
    > before committing to its action.  With the page marked PageKsm, yet
    > not in the stable tree, even page reclaim is unable to interfere.
    > Three, FAULT_FLAG_ALLOW_RETRY is acted upon only in lock_page_or_retry(),
    > which is called only from filemap_fault() (not the case here since we
    > don't consider file pages for conversion to PageKsm) or do_swap_page();
    > yet the fault we're provoking would be handled by do_wp_page().
    > Four, lock_page_or_retry() is called in those places when there's a
    > possibility that the page is being read in from disk, to drop the
    > mmap_sem across the slow I/O.  There is no precedent for dropping
    > mmap_sem here while allocating a new page, nor when pte_same() fails:
    > in the former case it could only be a win when the system is already
    > slowed by memory pressure, in the latter case there's little point,
    > since mmap_sem would be reacquired in a moment.
    > I think that amounts to a genial Nack!
    Hello Hugh,

    After reading your reply and the comments in break_ksm(), if the patch does
    not mess up
    "The important thing is to not let VM_MERGEABLE be cleared while any
    such pages might remain in the area",
    "because handle_mm_fault() may back out if there's
    any difficulty e.g. if pte accessed bit gets updated concurrently",

    then if the path in which lock_page_or_retry() is called is not involved,
    mmap_sem is not upped, so the patch has nearly same behavior with break_ksm.

    And the overhead of the patch, I think, could match break_ksm.

    With dozen cases of writers of mmap_sem in the mm directory, the patch looks
    more flexible in rare and rare corners.

    Best regards
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to
    More majordomo info at
    Please read the FAQ at

     \ /
      Last update: 2011-11-21 14:49    [W:0.025 / U:0.892 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site