[Bug 271490] Deadlock between _rtld_atfork_pre and _thr_attr_init

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=271490

--- Comment #7 from KJ Tsanaktsidis <kj@kjtsanaktsidis.id.au> ---
I did make a bit more progress debugging the issue though - I don't think the
problem has anything to do with jemalloc at all - it's _just_ an RTLD issue.

I captured a different backtrace of the problem and attached it. In this one
thread #2 is just sitting there chilling on an unrelated mutex; it is _not_
doing any jemalloc things. So, thread #1 manages to get deadlocked during
`_rtld_atfork_pre` all by itself.

In particular, it's blocked acquiring the `rtld_bind_lock` in write mode, and
the lock appears to already be held in write mode (by who??):

```
(gdb) print *((struct rtld_lock *)lock)

$21 = {lock = {rw_state = -1610612736, rw_flags = 2, rw_blocked_readers = 1,
rw_blocked_writers = 0, rw_spare = {0, 0, 0, 0}},
  _pad = '\000' <repeats 31 times>}
# bitmask against URWLOCK constants, from
https://github.com/freebsd/freebsd-src/blob/1377eb268a5a8a5573bda33a651963259c43c3ed/sys/sys/umtx.h#L62
(gdb) print /x ((struct rtld_lock *)lock)->lock.rw_state & 0x80000000U

$22 = 0x80000000
(gdb) print /x ((struct rtld_lock *)lock)->lock.rw_state & 0x40000000U

$23 = 0x0
(gdb) print /x ((struct rtld_lock *)lock)->lock.rw_state & 0x1FFFFFFFU

$24 = 0x0
(gdb)
```

Also also, I'm not 100% sure yet that this wasn't just luck so far, but it
_seems_ I can avoid the deadlock by setting `LD_BIND_NOW`...

Anyway I'll keep hunting for an isolated reproduction. If you have a stab in
the dark though I'm happy to try a patch as well.

-- 
You are receiving this mail because:
You are the assignee for the bug.