On Thu, May 07, 2020 at 06:14:19PM +0100, Catalin Marinas wrote: > On Wed, May 06, 2020 at 10:40:19AM -0700, Paul E. McKenney wrote: > > On Wed, May 06, 2020 at 12:22:37PM -0400, Qian Cai wrote: > > > == call_rcu() leaks == > > > Another issue that might be relevant is that it seems sometimes, > > > kmemleak will give a lot of false positives (hundreds) because the > > > memory was supposed to be freed by call_rcu() (for example, in > > > dst_release()) but for some reasons, it takes a long time probably > > > waiting for grace periods or some kind of RCU self-stall, but the > > > memory had already became an orphan. I am not sure how we are going > > > to resolve this properly until we have to figure out why call_rcu() > > > is taking so long to finish? > > > > I know nothing about kmemleak, but I won't let that stop me from making > > random suggestions... > > > > One approach is to do an rcu_barrier() inside kmemleak just before > > printing leaked blocks, and check to see if any are still leaked after > > the rcu_barrier(). > > The main issue is that kmemleak doesn't stop the world when scanning > (which can take over a minute, depending on your hardware), so we get > lots of transient pointer misses. There are some heuristics but > obviously they don't always work. > > With RCU, objects are queued for RCU freeing later and chained via > rcu_head.next (IIUC). Under load, this list can be pretty volatile and > if this happen during kmemleak scanning, it's sufficient to lose track > of a next pointer and the rest of the list would be reported as a leak. > > I think rcu_barrier() just before the starting the kmemleak scanning may > help if it reduces the number of objects queued. It might, especially if the call_rcu() rate is lower after the rcu_barrier() than it was beforehand. Which might well be the case when a large cleanup activity ended just before rcu_barrier() was invoked. > Now, I wonder whether kmemleak itself can break this RCU chain. The > kmemleak metadata is allocated on a slab alloc callback. The freeing, > however, is done using call_rcu() because originally calling back into > the slab freeing from kmemleak_free() didn't go well. Since the > kmemleak_object structure is not tracked by kmemleak, I wonder whether > its rcu_head would break this directed pointer reference graph. It is true that kmemleak could decide that being passed to call_rcu() as being freed. However, it would need to know the rcu_head offset. And there are (or were) a few places that pass linked structures to call_rcu(), and kmemleak would presumably need to mark them all free at that point. Or maybe accept the much lower false-positive rate from not marking them. > Let's try the rcu_barrier() first and I'll think about the metadata case > over the weekend. Looking forward to hearing how it goes! Thanx, Paul
