locking/osq_lock: Annotate a data race in osq_lock
authorQian Cai <cai@lca.pw>
Tue, 11 Feb 2020 13:54:15 +0000 (08:54 -0500)
committerPaul E. McKenney <paulmck@kernel.org>
Mon, 29 Jun 2020 19:04:48 +0000 (12:04 -0700)
commit33190b675ce2eacbeb4e75168c05b41110b506ec
treec1290359ba769363b2d0bdbaebd169b86d716b85
parent1fe84fd4a4027a17d511a832f89ab14107650ba4
locking/osq_lock: Annotate a data race in osq_lock

The prev->next pointer can be accessed concurrently as noticed by KCSAN:

 write (marked) to 0xffff9d3370dbbe40 of 8 bytes by task 3294 on cpu 107:
  osq_lock+0x25f/0x350
  osq_wait_next at kernel/locking/osq_lock.c:79
  (inlined by) osq_lock at kernel/locking/osq_lock.c:185
  rwsem_optimistic_spin
  <snip>

 read to 0xffff9d3370dbbe40 of 8 bytes by task 3398 on cpu 100:
  osq_lock+0x196/0x350
  osq_lock at kernel/locking/osq_lock.c:157
  rwsem_optimistic_spin
  <snip>

Since the write only stores NULL to prev->next and the read tests if
prev->next equals to this_cpu_ptr(&osq_node). Even if the value is
shattered, the code is still working correctly. Thus, mark it as an
intentional data race using the data_race() macro.

Signed-off-by: Qian Cai <cai@lca.pw>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/locking/osq_lock.c