On Fri, Jan 04, 2013 at 01:28:09PM -0800, Hugh Dickins wrote: > I've added Alexander, Hillf and Alex to the Cc. > > On Fri, 4 Jan 2013, Mel Gorman wrote: > > Zhouping, please test this patch. > > > > Andrea and Hugh, any comments on whether this could be improved? > > Your patch itself looks just right to me, no improvement required; > and it's easy to understand how the bug crept in, from a blanket > rwsem replacement of anon_vma mutex meeting the harmless-looking > anon_vma_interval_tree_foreach in __split_huge_page, which looked > as if it needed only the readlock provided by the usual method. > Indeed. Thanks Hugh for taking a look over it. > But I'd fight shy myself of trying to describe all the THP locking > conventions in the commit message: I haven't really tried to work > out just how right you've got all those details. > I thought it was risky myself but it was the best way of getting Andrea to object if I missed some subtlety! If I had infinite time I would follow up with a patch to Documentation/vm/transhuge.txt explaining how the anon_vma lock is used by THP. > The actual race in question here was just two processes (one or both > forked) doing split_huge_page() on the same THPage at the same time, > wasn't it? (Though of course we only see the backtrace from one of > them.) Which would be very confusing, and no surprise that the > pmd_trans_splitting test ends up skipping pmds already updated by > the racing process, so the mapcount doesn't match what's expected. > Of course we need exclusive lock against that, which you give it. > Ok, thanks. Will resend to Andrew with some changelog edits. -- Mel Gorman SUSE Labs -- 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/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>