Re: [RFC] cpus: avoid get stuck in pause_all_vcpus

[Paolo Bonzini <pbonzini@redhat.com>]

On 10/03/20 10:14, Longpeng(Mike) wrote:
> From: Longpeng <longpeng2@huawei.com>
> 
> We find an issue when repeat reboot in guest during migration, it cause the
> migration thread never be waken up again.
> 
> <main loop>                        |<migration_thread>
>                                    |
> LOCK BQL                           |
> ...                                |
> main_loop_should_exit              |
>  pause_all_vcpus                   |
>   1. set all cpus ->stop=true      |
>      and then kick                 |
>   2. return if all cpus is paused  |
>      (by '->stopped == true'), else|
>   3. qemu_cond_wait [BQL UNLOCK]   |
>                                    |LOCK BQL
>                                    |...
>                                    |do_vm_stop
>                                    | pause_all_vcpus
>                                    |  (A)set all cpus ->stop=true
>                                    |     and then kick
>                                    |  (B)return if all cpus is paused
>                                    |     (by '->stopped == true'), else
>                                    |  (C)qemu_cond_wait [BQL UNLOCK]
>   4. be waken up and LOCK BQL      |  (D)be waken up BUT wait for  BQL
>   5. goto 2.                       |
>  (BQL is still LOCKed)             |
>  resume_all_vcpus                  |
>   1. set all cpus ->stop=false     |
>      and ->stopped=false           |
> ...                                |
> BQL UNLOCK                         |  (E)LOCK BQL
>                                    |  (F)goto B. [but stopped is false now!]
>                                    |Finally, sleep at step 3 forever.
> 
> 
> Note: This patch is just for discuss this issue, I'm looking forward to
>       your suggestions, thanks!

Thanks Mike,

the above sketch is really helpful.

I think the problem is not that pause_all_vcpus() is not pausing hard
enough; the problem is rather than resume_all_vcpus(), when used outside
vm_start(), should know about the race and do nothing if it happens.

Fortunately resume_all_vcpus does not release the BQL so it should be
enough to test once; translated to code, this would be the patch to fix it:

diff --git a/cpus.c b/cpus.c
index b4f8b84b61..1eb7533a91 100644
--- a/cpus.c
+++ b/cpus.c
@@ -1899,6 +1899,10 @@ void resume_all_vcpus(void)
 {
     CPUState *cpu;

+    if (!runstate_is_running()) {
+        return;
+    }
+
     qemu_clock_enable(QEMU_CLOCK_VIRTUAL, true);
     CPU_FOREACH(cpu) {
         cpu_resume(cpu);


Thanks,

Paolo