Re: use-after-free in __perf_install_in_context

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Fri, Dec 04, 2015 at 09:04:35PM +0100, Dmitry Vyukov wrote:
> Hello,
> 
> While running syzkaller fuzzer I am seeing lots of the following
> use-after-free reports. Unfortunately all my numerous attempts to
> reproduce them in a controlled environment failed. They pop up during
> fuzzing periodically (once in several hours in a single VM), but
> whenever I try to stress-replay what happened in the VM before the
> report, the use-after-free does not reproduce. Can somebody
> knowledgeable in perf subsystem look at the report? Maybe it is
> possible to figure out what happened based purely on the report. I can
> pretty reliably test any proposed fixes.
> All reports look like this one. Then it is usually followed by other
> reports and eventually kernel hangs or dies. What happens in the
> fuzzer is essentially random syscalls with random arguments, tasks
> born and die concurrently and so on. I was able to reproduce it by
> restricting syscalls only to perf_event_open, perf ioctls and bpf
> syscall.

For the sake of trying to narrow it down:
does the error disappear when you stop using bpf syscall in your fuzzing?
If yes, then I could have missed some interaction between perf_event_free,
kprobe free and bpf_prog_free.
There was a race there before.
May be there is still something else.