[PATCH] sched/rt: Fix rt_runtime leaks with cpu hotplug and RT_RUNTIME_SHARE

[Zhao Wenhui]

When using cgroup rt_bandwidth with RT_RUNTIME_SHARE, if there are cpu
hotplug and cpu.rt_runtime_us changing concurrently, the warning in
__disable_runtime may occur:
[  991.697692] WARNING: CPU: 0 PID: 49573 at kernel/sched/rt.c:802
rq_offline_rt+0x24d/0x260
[  991.697795] CPU: 0 PID: 49573 Comm: kworker/1:0 Kdump: loaded Not
tainted 6.9.0-rc1+ #4
[  991.697800] Workqueue: events cpuset_hotplug_workfn
[  991.697803] RIP: 0010:rq_offline_rt+0x24d/0x260
[  991.697825] Call Trace:
[  991.697827]  <TASK>
[  991.697858]  set_rq_offline.part.125+0x2d/0x70
[  991.697864]  rq_attach_root+0xda/0x110
[  991.697867]  cpu_attach_domain+0x433/0x860
[  991.697880]  partition_sched_domains_locked+0x2a8/0x3a0
[  991.697885]  rebuild_sched_domains_locked+0x608/0x800
[  991.697895]  rebuild_sched_domains+0x1b/0x30
[  991.697897]  cpuset_hotplug_workfn+0x4b6/0x1160
[  991.697909]  process_scheduled_works+0xad/0x430
[  991.697917]  worker_thread+0x105/0x270
[  991.697922]  kthread+0xde/0x110
[  991.697928]  ret_from_fork+0x2d/0x50
[  991.697935]  ret_from_fork_asm+0x11/0x20
[  991.697940]  </TASK>
[  991.697941] ---[ end trace 0000000000000000 ]---

That's how it happens:
CPU0                                   CPU1
-----                                  -----

set_rq_offline(rq)
    __disable_runtime(rq) (1)
                                      tg_set_rt_bandwidth (2)
                                      do_balance_runtime  (3)
set_rq_online(rq)
    __enable_runtime(rq)  (4)

In step(1) rt_rq->rt_runtime is set to RUNTIME_INF, and this rtrq's
runtime is not supposed to change until its rq gets online. However,
in step(2) tg_set_rt_bandwidth can set rt_rq->rt_runtime to
rt_bandwidth.rt_runtime. Then, in step(3) rtrq's runtime is not
RUNTIME_INF, so others can borrow rt_runtime from it. Finally, in
step(4) the rq gets online, so its rtrq's runtime is set to
rt_bandwidth.rt_runtime again, and Since then the total rt_runtime in
the domain is increased by this way. After these steps, when offline cpu,
rebuilding the sched_domain will offline all rq, and the last rq will
find the rt_runtime is increased but nowhere to return.

To fix this, we can add a state RUNTIME_DISABLED, which means the runtime
is disabled and should not be used. When rq get offline, we can set its
rtrq's rt_runtime to RUNTIME_DISABLED, and when rq get online, reset it.
And in tg_set_rt_bandwidth and do_balance_runtime, never change a
disabled rt_runtime.

Fixes: 7def2be1dc67 ("sched: fix hotplug cpus on ia64")
Closes: https://lore.kernel.org/all/47b4a790-9a27-2fc5-f2aa-f9981c6da015@xxxxxxxxxx/
Co-developed-by: Hui Tang <tanghui20@xxxxxxxxxx>
Signed-off-by: Hui Tang <tanghui20@xxxxxxxxxx>
Signed-off-by: Zhao Wenhui <zhaowenhui8@xxxxxxxxxx>
---
 kernel/sched/rt.c    | 15 +++++++++------
 kernel/sched/sched.h |  5 +++++
 2 files changed, 14 insertions(+), 6 deletions(-)

diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index aa4c1c874fa4..44b8cc5a2f5f 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c
@@ -704,7 +704,8 @@ static void do_balance_runtime(struct rt_rq *rt_rq)
 		 * or __disable_runtime() below sets a specific rq to inf to
 		 * indicate its been disabled and disallow stealing.
 		 */
-		if (iter->rt_runtime == RUNTIME_INF)
+		if (iter->rt_runtime == RUNTIME_INF ||
+				iter->rt_runtime == RUNTIME_DISABLED)
 			goto next;
 
 		/*
@@ -775,7 +776,9 @@ static void __disable_runtime(struct rq *rq)
 			/*
 			 * Can't reclaim from ourselves or disabled runqueues.
 			 */
-			if (iter == rt_rq || iter->rt_runtime == RUNTIME_INF)
+			if (iter == rt_rq ||
+					iter->rt_runtime == RUNTIME_INF ||
+					iter->rt_runtime == RUNTIME_DISABLED)
 				continue;
 
 			raw_spin_lock(&iter->rt_runtime_lock);
@@ -801,10 +804,9 @@ static void __disable_runtime(struct rq *rq)
 		WARN_ON_ONCE(want);
 balanced:
 		/*
-		 * Disable all the borrow logic by pretending we have inf
-		 * runtime - in which case borrowing doesn't make sense.
+		 * Disable all the borrow logic by marking runtime disabled.
 		 */
-		rt_rq->rt_runtime = RUNTIME_INF;
+		rt_rq->rt_runtime = RUNTIME_DISABLED;
 		rt_rq->rt_throttled = 0;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 		raw_spin_unlock(&rt_b->rt_runtime_lock);
@@ -2827,7 +2829,8 @@ static int tg_set_rt_bandwidth(struct task_group *tg,
 		struct rt_rq *rt_rq = tg->rt_rq[i];
 
 		raw_spin_lock(&rt_rq->rt_runtime_lock);
-		rt_rq->rt_runtime = rt_runtime;
+		if (rt_rq->rt_runtime != RUNTIME_DISABLED)
+			rt_rq->rt_runtime = rt_runtime;
 		raw_spin_unlock(&rt_rq->rt_runtime_lock);
 	}
 	raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index a831af102070..c2ad9102b8fa 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -183,6 +183,11 @@ extern struct list_head asym_cap_list;
  */
 #define RUNTIME_INF		((u64)~0ULL)
 
+/*
+ * Single value that denotes runtime is disabled, and it should not be used.
+ */
+#define RUNTIME_DISABLED	(-2ULL)
+
 static inline int idle_policy(int policy)
 {
 	return policy == SCHED_IDLE;
-- 
2.34.1