[PATCH] sched: Preserve reset-on-fork across concurrent sched_setparam()

SCHED_RESET_ON_FORK prevents scheduling attributes from propagating to a
new child. The request is stored in p->sched_reset_on_fork; at fork
time, the scheduler uses it to reset inherited DL/RT policy, negative
nice values, custom slices and utilization clamps, then clears the flag
in the child.

sched_setparam() uses SETPARAM_POLICY to update scheduler parameters
while preserving both the current policy and the reset-on-fork setting.

However, __sched_setscheduler() samples p->sched_reset_on_fork before
acquiring the task's rq lock. A concurrent sched_setscheduler() can
change the flag after that snapshot and the lagging sched_setparam()
call can overwrite it with the stale value.

Remember when the caller requested to keep the policy and refresh
reset_on_fork after acquiring the rq lock. Both the early-exit and full
update paths then preserve the most recent serialized setting.

Fixes: ca94c442535a ("sched: Introduce SCHED_RESET_ON_FORK scheduling policy flag")
Reported-by: K Prateek Nayak <[email protected]>
Link: https://lore.kernel.org/[email protected]
Signed-off-by: Andrea Righi <[email protected]>
---
 kernel/sched/syscalls.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/kernel/sched/syscalls.c b/kernel/sched/syscalls.c
index 8fb8474d0a0ec..b149aeb2fd490 100644
--- a/kernel/sched/syscalls.c
+++ b/kernel/sched/syscalls.c
@@ -503,6 +503,7 @@ int __sched_setscheduler(struct task_struct *p,
 	int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
 	struct rq *rq;
 	bool cpuset_locked = false;
+	bool keep_policy = policy < 0;
 
 	/* The pi code expects interrupts enabled */
 	BUG_ON(pi && in_interrupt());
@@ -571,6 +572,10 @@ int __sched_setscheduler(struct task_struct *p,
 	rq = task_rq_lock(p, &rf);
 	update_rq_clock(rq);
 
+	/* Preserve reset_on_fork changes made while the rq lock was not held. */
+	if (keep_policy)
+		reset_on_fork = p->sched_reset_on_fork;
+
 	/*
 	 * Changing the policy of the stop threads its a very bad idea:
 	 */
-- 
2.55.0