quoted

I take that back. I think it's leaking memory - my tests never used
to run out of memory, but now they eventually do.

quoted

I'll try to figure out more.

Interesting.  The kernel test robot found a bug in the insertion
slowpath where we end up inserting without taking the inner spinlock
in case of a nested table.  Not sure whether that's the same issue
as you're seeing but I'll do a v2 posting.

Increasing the memory for the VMs from 384MB to 512MB didn't avoid the
issue, so there's a definite leak somewhere, although this time fewer
VMs crashed :)

Reverting the patches and running with 384MB then works, so it's not
something else added in the meantime (I ran last a few days ago, but
don't think I merged anything interesting in the meantime).

I'll test your new patches in a minute.

johannes