GitLab

The more intense grace-period processing resulting from the 50x RCU
Tasks Trace grace-period speedups exposed the following race condition:

o	Task A running on CPU 0 executes rcu_read_lock_trace(),
	entering a read-side critical section.

o	When Task A eventually invokes rcu_read_unlock_trace()
	to exit its read-side critical section, this function
	notes that the ->trc_reader_special.s flag is zero and
	and therefore invoke wil set ->trc_reader_nesting to zero
	using WRITE_ONCE().  But before that happens...

o	The RCU Tasks Trace grace-period kthread running on some other
	CPU interrogates Task A, but this fails because this task is
	currently running.  This kthread therefore sends an IPI to CPU 0.

o	CPU 0 receives the IPI, and thus invokes trc_read_check_handler().
	Because Task A has not yet cleared its ->trc_reader_nesting
	counter, this function sees that Task A is still within its
	read-side critical section.  This function therefore sets the
	->trc_reader_nesting.b.need_qs flag, AKA the .need_qs flag.

	Except that Task A has already checked the .need_qs flag, which
	is part of the ->trc_reader_special.s flag.  The .need_qs flag
	therefore remains set until Task A's next rcu_read_unlock_trace().

o	Task A now invokes synchronize_rcu_tasks_trace(), which cannot
	start a new grace period until the current grace period completes.
	And thus cannot return until after that time.

	But Task A's .need_qs flag is still set, which prevents the current
	grace period from completing.  And because Task A is blocked, it
	will never execute rcu_read_unlock_trace() until its call to
	synchronize_rcu_tasks_trace() returns.

	We are therefore deadlocked.

This race is improbable, but 80 hours of rcutorture made it happen twice.
The race was possible before the grace-period speedup, but roughly 50x
less probable.  Several thousand hours of rcutorture would have been
necessary to have a reasonable chance of making this happen before this
50x speedup.

This commit therefore eliminates this deadlock by setting
->trc_reader_nesting to a large negative number before checking the
.need_qs and zeroing (or decrementing with respect to its initial
value) ->trc_reader_nesting.  For its part, the IPI handler's
trc_read_check_handler() function adds a check for negative values,
deferring evaluation of the task in this case.  Taken together, these
changes avoid this deadlock scenario.

Fixes: 276c4104

 ("rcu-tasks: Split ->trc_reader_need_end")
Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: <bpf@vger.kernel.org>
Cc: <stable@vger.kernel.org> # 5.7.x
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>