| 1 |
From 55628334c7c73c8b948d7d97be0e79b61b28cd11 Mon Sep 17 00:00:00 2001 |
| 2 |
Message-ID: <55628334c7c73c8b948d7d97be0e79b61b28cd11.1699049719.git.osandov@fb.com> |
| 3 |
From: Omar Sandoval <[email protected]> |
| 4 |
Date: Fri, 3 Nov 2023 14:50:46 -0700 |
| 5 |
Subject: [PATCH] sched: work around mystery QEMU hang |
| 6 |
|
| 7 |
Before commit 2558aacff858 ("sched/hotplug: Ensure only per-cpu kthreads |
| 8 |
run during hotplug") (in v5.11), GCC inlined the call to |
| 9 |
raw_spin_unlock_irq() after __balance_callbacks() in __schedule(). That |
| 10 |
change incidentally turned it into an uninlined function call. For some |
| 11 |
reason, this causes the kernel to hang while calibrating loops_per_jiffy |
| 12 |
for secondary CPUs, but only under QEMU full system emulation, not KVM. |
| 13 |
|
| 14 |
Commit 5cb9eaa3d274 ("sched: Wrap rq::lock access") (in v5.14) added |
| 15 |
some wrappers that essentially turned the call to raw_spin_unlock_irq() |
| 16 |
into a (still not inlined by GCC) call to raw_spin_unlock() followed by |
| 17 |
local_irq_enable(). This mysteriously gets rid of the hang. |
| 18 |
|
| 19 |
This feels like a QEMU bug, but since it only affects EOL kernels, let's |
| 20 |
just work around it. |
| 21 |
|
| 22 |
See osandov/drgn#365. |
| 23 |
|
| 24 |
Signed-off-by: Omar Sandoval <[email protected]> |
| 25 |
--- |
| 26 |
kernel/sched/core.c | 3 ++- |
| 27 |
1 file changed, 2 insertions(+), 1 deletion(-) |
| 28 |
|
| 29 |
diff --git a/kernel/sched/core.c b/kernel/sched/core.c |
| 30 |
index 4ca80df205ce..4901f47e84d8 100644 |
| 31 |
--- a/kernel/sched/core.c |
| 32 |
+++ b/kernel/sched/core.c |
| 33 |
@@ -5150,7 +5150,8 @@ static void __sched notrace __schedule(bool preempt) |
| 34 |
|
| 35 |
rq_unpin_lock(rq, &rf); |
| 36 |
__balance_callbacks(rq); |
| 37 |
- raw_spin_unlock_irq(&rq->lock); |
| 38 |
+ raw_spin_unlock(&rq->lock); |
| 39 |
+ local_irq_enable(); |
| 40 |
} |
| 41 |
} |
| 42 |
|
| 43 |
-- |
| 44 |
2.41.0 |
| 45 |
|