MFC r202889, r202940:

- Fix a race in sched_switch() of sched_4bsd.
  Block the td_lock when acquiring explicitly sched_lock in order to prevent
  races with other td_lock contenders.
- Merge the ULE's internal function thread_block_switch() into the global
  thread_lock_block() and make the former semantic as the default for
  thread_lock_block().
- Split out an invariant in order to have better checks.

Tested by:	Giovanni Trematerra
		<giovanni dot trematerra at gmail dot com>
Approved by:	re (kib)

3 files changed, 15 insertions, 27 deletions

diff --git a/sys/kern/kern_mutex.c b/sys/kern/kern_mutex.c
index 9f7f1c87bee0..94c8c2313ce0 100644
--- a/sys/kern/kern_mutex.c
+++ b/sys/kern/kern_mutex.c
@@ -557,7 +557,6 @@ thread_lock_block(struct thread *td)
 {
 	struct mtx *lock;
 
-	spinlock_enter();
 	THREAD_LOCK_ASSERT(td, MA_OWNED);
 	lock = td->td_lock;
 	td->td_lock = &blocked_lock;
@@ -572,7 +571,6 @@ thread_lock_unblock(struct thread *td, struct mtx *new)
 	mtx_assert(new, MA_OWNED);
 	MPASS(td->td_lock == &blocked_lock);
 	atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new);
-	spinlock_exit();
 }
 
 void
diff --git a/sys/kern/sched_4bsd.c b/sys/kern/sched_4bsd.c
index 56f69bd34af8..c0989062cb06 100644
--- a/sys/kern/sched_4bsd.c
+++ b/sys/kern/sched_4bsd.c
@@ -824,9 +824,11 @@ sched_sleep(struct thread *td)
 void
 sched_switch(struct thread *td, struct thread *newtd, int flags)
 {
+	struct mtx *tmtx;
 	struct td_sched *ts;
 	struct proc *p;
 
+	tmtx = NULL;
 	ts = td->td_sched;
 	p = td->td_proc;
 
@@ -835,17 +837,20 @@ sched_switch(struct thread *td, struct thread *newtd, int flags)
 	/*
 	 * Switch to the sched lock to fix things up and pick
 	 * a new thread.
+	 * Block the td_lock in order to avoid breaking the critical path.
 	 */
 	if (td->td_lock != &sched_lock) {
 		mtx_lock_spin(&sched_lock);
-		thread_unlock(td);
+		tmtx = thread_lock_block(td);
 	}
 
 	if ((p->p_flag & P_NOLOAD) == 0)
 		sched_load_rem();
 
-	if (newtd)
+	if (newtd) {
+		MPASS(newtd->td_lock == &sched_lock);
 		newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED);
+	}
 
 	td->td_lastcpu = td->td_oncpu;
 	td->td_flags &= ~TDF_NEEDRESCHED;
@@ -888,8 +893,8 @@ sched_switch(struct thread *td, struct thread *newtd, int flags)
 			sched_load_add();
 	} else {
 		newtd = choosethread();
+		MPASS(newtd->td_lock == &sched_lock);
 	}
-	MPASS(newtd->td_lock == &sched_lock);
 
 	if (td != newtd) {
 #ifdef	HWPMC_HOOKS
@@ -907,7 +912,7 @@ sched_switch(struct thread *td, struct thread *newtd, int flags)
 			(*dtrace_vtime_switch_func)(newtd);
 #endif
                 /* I feel sleepy */
-		cpu_switch(td, newtd, td->td_lock);
+		cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock);
 		/*
 		 * Where am I?  What year is it?
 		 * We are in the same thread that went to sleep above,
diff --git a/sys/kern/sched_ule.c b/sys/kern/sched_ule.c
index d1d4bbeb5130..97c7bcd52b2f 100644
--- a/sys/kern/sched_ule.c
+++ b/sys/kern/sched_ule.c
@@ -318,7 +318,6 @@ static void sched_balance_groups(void);
 static void sched_balance_group(struct tdq_group *);
 static void sched_balance_pair(struct tdq *, struct tdq *);
 static inline struct tdq *sched_setcpu(struct td_sched *, int, int);
-static inline struct mtx *thread_block_switch(struct thread *);
 static inline void thread_unblock_switch(struct thread *, struct mtx *);
 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
 #endif
@@ -989,9 +988,11 @@ sched_setcpu(struct td_sched *ts, int cpu, int flags)
 	 * The hard case, migration, we need to block the thread first to
 	 * prevent order reversals with other cpus locks.
 	 */
+	spinlock_enter();
 	thread_lock_block(td);
 	TDQ_LOCK(tdq);
 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
+	spinlock_exit();
 	return (tdq);
 }
 
@@ -1789,23 +1790,6 @@ sched_switchin(struct tdq *tdq, struct thread *td)
 }
 
 /*
- * Block a thread for switching.  Similar to thread_block() but does not
- * bump the spin count.
- */
-static inline struct mtx *
-thread_block_switch(struct thread *td)
-{
-	struct mtx *lock;
-
-	THREAD_LOCK_ASSERT(td, MA_OWNED);
-	lock = td->td_lock;
-	td->td_lock = &blocked_lock;
-	mtx_unlock_spin(lock);
-
-	return (lock);
-}
-
-/*
  * Handle migration from sched_switch().  This happens only for
  * cpu binding.
  */
@@ -1822,7 +1806,7 @@ sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
 	 * not holding either run-queue lock.
 	 */
 	spinlock_enter();
-	thread_block_switch(td);	/* This releases the lock on tdq. */
+	thread_lock_block(td);	/* This releases the lock on tdq. */
 
 	/*
 	 * Acquire both run-queue locks before placing the thread on the new
@@ -1848,7 +1832,8 @@ sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
 }
 
 /*
- * Release a thread that was blocked with thread_block_switch().
+ * Variadic version of thread_lock_unblock() that does not assume td_lock
+ * is blocked.
  */
 static inline void
 thread_unblock_switch(struct thread *td, struct mtx *mtx)
@@ -1907,7 +1892,7 @@ sched_switch(struct thread *td, struct thread *newtd, int flags)
 	} else {
 		/* This thread must be going to sleep. */
 		TDQ_LOCK(tdq);
-		mtx = thread_block_switch(td);
+		mtx = thread_lock_block(td);
 		tdq_load_rem(tdq, ts);
 	}
 	/*