SCHED_4BSD: maybe_resched() cannot schedule ast() for curthread

maybe_resched() needs to schedule TDA_SCHED for curthread, but this
requires taking curthread lock while owning some other thread lock.
To avoid introducing the order:
- Use a scheduler-private TDP flag.
- Register an unconditional TDA_SCHED_PRIV for 4BSD.
When an AST needs to be scheduled, i.e. the current thread must do
context switch in the return to userspace path, set the flag.
Then the ast handler calls ast_scheduler(), which gives the same
effect as scheduling TDA_AST.

The overhead is a single function call on each userspace return, for
4BSD case.

Reported and tested by:	pho (previous version)
Reviewed by:	markj
Sponsored by:	The FreeBSD Foundation
MFC after:	1 week
Differential revision:	https://reviews.freebsd.org/D54945

sys/kern/kern_synch.c

@@ -600,7 +600,7 @@ loadav(void *arg)
 	    loadav, NULL, C_DIRECT_EXEC | C_PREL(32));
 }
 
-static void
+void
 ast_scheduler(struct thread *td, int tda __unused)
 {
 #ifdef KTRACE

sys/kern/sched_4bsd.c

@@ -108,6 +108,8 @@ struct td_sched {
 #define TDF_BOUND	TDF_SCHED1	/* Bound to one CPU. */
 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
 
+#define	TDP_RESCHED	TDP_SCHED1	/* Reschedule due to maybe_resched(). */
+
 /* flags kept in ts_flags */
 #define	TSF_AFFINITY	0x0001		/* Has a non-"full" CPU set. */
 
@@ -274,6 +276,17 @@ sched_load_rem(void)
 	KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt);
 	SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt);
 }
+
+static void
+maybe_resched_ast(struct thread *td, int tda)
+{
+	MPASS(td == curthread);		/* We are AST */
+	if ((td->td_pflags & TDP_RESCHED) != 0) {
+		td->td_pflags &= ~TDP_RESCHED;
+		ast_scheduler(td, tda);
+	}
+}
+
 /*
  * Arrange to reschedule if necessary, taking the priorities and
  * schedulers into account.
@@ -281,10 +294,12 @@ sched_load_rem(void)
 static void
 maybe_resched(struct thread *td)
 {
+	struct thread *ctd;
 
+	ctd = curthread;
 	THREAD_LOCK_ASSERT(td, MA_OWNED);
-	if (td->td_priority < curthread->td_priority)
-		ast_sched_locked(curthread, TDA_SCHED);
+	if (td->td_priority < ctd->td_priority)
+		ctd->td_pflags |= TDP_RESCHED;
 }
 
 /*
@@ -621,6 +636,8 @@ sched_4bsd_setup(void)
 
 	/* Account for thread0. */
 	sched_load_add();
+
+	ast_register(TDA_SCHED_PRIV, ASTR_UNCOND, 0, maybe_resched_ast);
 }
 
 /*

sys/sys/proc.h

@@ -503,6 +503,7 @@ enum {
 	TDA_SIGSUSPEND,
 	TDA_MOD3,		/* .. and after */
 	TDA_MOD4,
+	TDA_SCHED_PRIV,
 	TDA_MAX,
 };
 #define	TDAI(tda)		(1U << (tda))

sys/sys/sched.h

@@ -188,6 +188,8 @@ sched_unpin(void)
 	curthread->td_pinned--;
 }
 
+void ast_scheduler(struct thread *td, int tda);
+
 /* sched_add arguments (formerly setrunqueue) */
 #define	SRQ_BORING	0x0000		/* No special circumstances. */
 #define	SRQ_YIELDING	0x0001		/* We are yielding (from mi_switch). */