On Fri, Jul 17, 2015 at 01:13:30AM +0200, Richard Braun wrote: > Those patches look good but since we don't have a good reproducer for > this particular problem, the only way to know is to get it running on > busy buildds.
Actually, we do have a good reproducer. I inserted calls to force a context switch each time a radix tree node is allocated, and I got the assertion right away. I then applied Justus' patches along with the error injection code and could run the system fine. I then proceeded to build some packages and run a benchmark tool (sysbench) to stress the file system, which is the biggest ipc space users, and had no trouble. -- Richard Braun
