[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240524034227.1871565-1-zhaowenhui8@huawei.com>
Date: Fri, 24 May 2024 03:42:27 +0000
From: Zhao Wenhui <zhaowenhui8@...wei.com>
To: <mingo@...hat.com>, <peterz@...radead.org>, <juri.lelli@...hat.com>,
<vincent.guittot@...aro.org>, <dietmar.eggemann@....com>,
<rostedt@...dmis.org>, <bsegall@...gle.com>, <mgorman@...e.de>,
<bristot@...hat.com>, <vschneid@...hat.com>, <linux-kernel@...r.kernel.org>
CC: <zhaowenhui8@...wei.com>, <tanghui20@...wei.com>
Subject: [PATCH] sched/rt: Fix rt_runtime leaks with cpu hotplug and RT_RUNTIME_SHARE
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@huawei.com/
Co-developed-by: Hui Tang <tanghui20@...wei.com>
Signed-off-by: Hui Tang <tanghui20@...wei.com>
Signed-off-by: Zhao Wenhui <zhaowenhui8@...wei.com>
---
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
Powered by blists - more mailing lists