From: Tim Leslie <tleslie@protonmail.com>
Subject: kern_resorce: move to lazy updates after donice
To: "tech@openbsd.org" <tech@openbsd.org>
Date: Sat, 14 Feb 2026 13:18:31 +0000


@tech,

Diff below to remove the immediate setpriority update loop in kern_resource.

Currently, donice grabs ps_mtx and SCHED_LOCK to call setpriority on all threads. However, setpriority only updates p_usrpri; it does not move SRUN threads to their new runqueue buckets. That re-bucketing only happens in schedcpu (once per second).

Since the scheduler doesn't act on the new priority until schedcpu anyway, there is no latency benefit to updating p_usrpri immediately for runnable threads.

The other states are also covered:
SSLEEP: Priority is recalculated in setrunnable upon wakeup.
SONPROC: Priority is updated in schedclock (hardclock).

As such, doing less is an improvement. WRITE_ONCE added to avoid any read tears.

Tim


diff --git a/sys/kern/kern_resource.c b/sys/kern/kern_resource.c
--- a/sys/kern/kern_resource.c
+++ b/sys/kern/kern_resource.c
@@ -47,6 +47,7 @@
 #include <sys/ktrace.h>
 #include <sys/sched.h>
 #include <sys/signalvar.h>
+#include <sys/atomic.h>
 
 #include <sys/mount.h>
 #include <sys/syscallargs.h>
@@ -196,7 +197,6 @@ int
 donice(struct proc *curp, struct process *chgpr, int n)
 {
 	struct ucred *ucred = curp->p_ucred;
-	struct proc *p;
 
 	if (ucred->cr_uid != 0 && ucred->cr_ruid != 0 &&
 	    ucred->cr_uid != chgpr->ps_ucred->cr_uid &&
@@ -209,14 +209,7 @@ donice(struct proc *curp, struct process *chgpr, int n)
 	n += NZERO;
 	if (n < chgpr->ps_nice && suser(curp))
 		return (EACCES);
-	chgpr->ps_nice = n;
-	mtx_enter(&chgpr->ps_mtx);
-	SCHED_LOCK();
-	TAILQ_FOREACH(p, &chgpr->ps_threads, p_thr_link) {
-		setpriority(p, p->p_estcpu, n);
-	}
-	SCHED_UNLOCK();
-	mtx_leave(&chgpr->ps_mtx);
+	WRITE_ONCE(chgpr->ps_nice, n);
 	return (0);
 }