On Tue, May 05, 2020 at 09:00:44AM -0700, Eric Dumazet wrote:
Not exactly the 10,000,000, as it is only the possible highest number, but I was able to observe clear exponential increase of the number of the objects using slabtop. Before the start of the problematic workload, the number of objects of 'kmalloc-64' was 5760, but I was able to observe the number increase to 1,136,576.
OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAMEbefore: 5760 5088 88% 0.06K 90 64 360K kmalloc-64 after: 1136576 1136576 100% 0.06K 17759 64 71036K kmalloc-64
Great, thanks.
How recent is the kernel you are running for your experiment ?
Let's make sure the bug is not in RCU.
After Al changes, RCU got slightly better under stress.
The thing that worries me here is that this is far from being the only source of RCU-delayed freeing of objects. If we really see bogus OOM kills due to that (IRL, not in an artificial microbenchmark), we'd better do something that would help with all those sources, not just paper over the contributions from one of those. Because there's no chance in hell to get rid of RCU-delayed freeing in general...
Does the problem extend to kfree_rcu()? And there's a lot of RCU callbacks that boil down to kmem_cache_free(); those really look like they should have exact same issue - sock_free_inode() is one of those, after all.