From: Thomas Gleixner Date: Wed, 8 Mar 2017 14:23:35 +0100 Subject: [PATCH] futex: workaround migrate_disable/enable in different context Origin: https://www.kernel.org/pub/linux/kernel/projects/rt/4.19/older/patches-4.19.1-rt3.tar.xz migrate_disable()/migrate_enable() takes a different path in atomic() vs !atomic() context. These little hacks ensure that we don't underflow / overflow the migrate code counts properly while we lock the hb lockwith interrupts enabled and unlock it with interrupts disabled. Signed-off-by: Thomas Gleixner Signed-off-by: Sebastian Andrzej Siewior --- kernel/futex.c | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) --- a/kernel/futex.c +++ b/kernel/futex.c @@ -2798,9 +2798,18 @@ static int futex_lock_pi(u32 __user *uad * lock handoff sequence. */ raw_spin_lock_irq(&q.pi_state->pi_mutex.wait_lock); + /* + * the migrate_disable() here disables migration in the in_atomic() fast + * path which is enabled again in the following spin_unlock(). We have + * one migrate_disable() pending in the slow-path which is reversed + * after the raw_spin_unlock_irq() where we leave the atomic context. + */ + migrate_disable(); + spin_unlock(q.lock_ptr); ret = __rt_mutex_start_proxy_lock(&q.pi_state->pi_mutex, &rt_waiter, current); raw_spin_unlock_irq(&q.pi_state->pi_mutex.wait_lock); + migrate_enable(); if (ret) { if (ret == 1) @@ -2947,11 +2956,21 @@ static int futex_unlock_pi(u32 __user *u * observed. */ raw_spin_lock_irq(&pi_state->pi_mutex.wait_lock); + /* + * Magic trickery for now to make the RT migrate disable + * logic happy. The following spin_unlock() happens with + * interrupts disabled so the internal migrate_enable() + * won't undo the migrate_disable() which was issued when + * locking hb->lock. + */ + migrate_disable(); spin_unlock(&hb->lock); /* drops pi_state->pi_mutex.wait_lock */ ret = wake_futex_pi(uaddr, uval, pi_state); + migrate_enable(); + put_pi_state(pi_state); /*