futex: Prevent requeue_pi() lock nesting issue on RT

The requeue_pi() operation on RT kernels creates a problem versus the
task::pi_blocked_on state when a waiter is woken early (signal, timeout)
and that early wake up interleaves with the requeue_pi() operation.

When the requeue manages to block the waiter on the rtmutex which is
associated to the second futex, then a concurrent early wakeup of that
waiter faces the problem that it has to acquire the hash bucket spinlock,
which is not an issue on non-RT kernels, but on RT kernels spinlocks are
substituted by 'sleeping' spinlocks based on rtmutex. If the hash bucket
lock is contended then blocking on that spinlock would result in a
impossible situation: blocking on two locks at the same time (the hash
bucket lock and the rtmutex representing the PI futex).

It was considered to make the hash bucket locks raw_spinlocks, but
especially requeue operations with a large amount of waiters can introduce
significant latencies, so that's not an option for RT.

The RT tree carried a solution which (ab)used task::pi_blocked_on to store
the information about an ongoing requeue and an early wakeup which worked,
but required to add checks for these special states all over the place.

The distangling of an early wakeup of a waiter for a requeue_pi() operation
is already looking at quite some different states and the task::pi_blocked_on
magic just expanded that to a hard to understand 'state machine'.

This can be avoided by keeping track of the waiter/requeue state in the
futex_q object itself.

Add a requeue_state field to struct futex_q with the following possible
states:

	Q_REQUEUE_PI_NONE
	Q_REQUEUE_PI_IGNORE
	Q_REQUEUE_PI_IN_PROGRESS
	Q_REQUEUE_PI_WAIT
	Q_REQUEUE_PI_DONE
	Q_REQUEUE_PI_LOCKED

The waiter starts with state = NONE and the following state transitions are
valid:

On the waiter side:
  Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_IGNORE
  Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_WAIT

On the requeue side:
  Q_REQUEUE_PI_NONE		-> Q_REQUEUE_PI_INPROGRESS
  Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_DONE/LOCKED
  Q_REQUEUE_PI_IN_PROGRESS	-> Q_REQUEUE_PI_NONE (requeue failed)
  Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_DONE/LOCKED
  Q_REQUEUE_PI_WAIT		-> Q_REQUEUE_PI_IGNORE (requeue failed)

The requeue side ignores a waiter with state Q_REQUEUE_PI_IGNORE as this
signals that the waiter is already on the way out. It also means that
the waiter is still on the 'wait' futex, i.e. uaddr1.

The waiter side signals early wakeup to the requeue side either through
setting state to Q_REQUEUE_PI_IGNORE or to Q_REQUEUE_PI_WAIT depending
on the current state. In case of Q_REQUEUE_PI_IGNORE it can immediately
proceed to take the hash bucket lock of uaddr1. If it set state to WAIT,
which means the wakeup is interleaving with a requeue in progress it has
to wait for the requeue side to change the state. Either to DONE/LOCKED
or to IGNORE. DONE/LOCKED means the waiter q is now on the uaddr2 futex
and either blocked (DONE) or has acquired it (LOCKED). IGNORE is set by
the requeue side when the requeue attempt failed via deadlock detection
and therefore the waiter's futex_q is still on the uaddr1 futex.

While this is not strictly required on !RT making this unconditional has
the benefit of common code and it also allows the waiter to avoid taking
the hash bucket lock on the way out in certain cases, which reduces
contention.

Add the required helpers required for the state transitions, invoke them at
the right places and restructure the futex_wait_requeue_pi() code to handle
the return from wait (early or not) based on the state machine values.

On !RT enabled kernels the waiter spin waits for the state going from
Q_REQUEUE_PI_WAIT to some other state, on RT enabled kernels this is
handled by rcuwait_wait_event() and the corresponding wake up on the
requeue side.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20210815211305.693317658@linutronix.de

0 files changed, 0 insertions, 0 deletions