On 01/09, Andrea Arcangeli wrote: > > On Thu, Jan 09, 2014 at 03:04:47PM +0100, Oleg Nesterov wrote: > > OK. Even if I am right, we can probably make another fix. > > I think the confusion here was to think this was related to the futex > code, it isn't. This was just a generic theoretical problem found > doing the futex cleanups but it's not related to the futex code.
Yes, yes, sure. I mentioned get_futex_key() just for example. > > put_compound_page() and __get_page_tail() can do yet another PageTail() > > check _before_ compound_lock(). > > The above alternate fix looks good to me too. > > Only thing to sort out is in the common code (not just x86) then we > may need a smp_mb() between PageTail check and the bit_spin_lock... We > just can't risk writing the bit_spin_lock before reading PageTail. I do not think we need mb() in between... other callers of compound_lock() look fine, get/put(page_tail) can't have the false positive after successful get_page_unless_zero(), and recently it was documented that the kernel can rely on the control dependency to serialize LOAD + STORE. But we probably need barrier() in between, we can't use ACCESS_ONCE(). > And regardless of gup_fast, like Linus said, for increased NUMA > fairness we could move the compound lock from page->flags to an hashed > array of proper spinlocks sized in function of ram. The contention on > these locks is so low that I doubt we can run into lock starvation, > but because the contention is so low, the array would be fine as well, > and it would be more theoretically correct for NUMA usages than the > bit spinlock. So this problem also goes away if we convert the > bit_spin_lock to an hashed array of spin_lock. Yes. But in this case I really think we should cleanup get/put first and add the helper, like the patch I mentioned does. > I personally prefer to keep the complexity in one place so adding to > get/put_page OK. I'll send v3. > > Although personally I'd prefer this patch. And if we change get/put > > I think it would be better to do this on top of > > > > "[PATCH -mm 6/7] mm: thp: introduce get_lock_thp_head()" > > http://marc.info/?l=linux-kernel&m=138739438800899 > > Not against the cleanups of course, but about the order, it gets > harder to backport it for distros if applied after the cleanups. Oh, I don't think this highly theoreitical fix should be backported but I agree, lets fix the bug first. Oleg. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
