On Mon, Nov 21, 2011 at 12:16 PM, Hugh Dickins <hughd@xxxxxxxxxx> 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 <dhillf@xxxxxxxxx> > > 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", and "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 Hillf -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href