diff mbox

[RFC,6/8] linux-user: Support CPU work queue

Message ID 57763236.8000203@gmail.com
State New
Headers show

Commit Message

Sergey Fedorov July 1, 2016, 9:04 a.m. UTC 
On 01/07/16 11:56, Sergey Fedorov wrote:
> On 30/06/16 13:35, Sergey Fedorov wrote:
>> On 30/06/16 13:32, Alex Bennée wrote:
>>> Sergey Fedorov <serge.fdrv@gmail.com> writes:
>>>
>>>> On 29/06/16 19:17, Alex Bennée wrote:
>>>>> So I think there is a deadlock we can get with the async work:
>>>>>
>>>>> (gdb) thread apply all bt
>>>>>
>>>>> Thread 11 (Thread 0x7ffefeca7700 (LWP 2912)):
>>>>> #0  pthread_cond_wait@@GLIBC_2.3.2 () at ../sysdeps/unix/sysv/linux/x86_64/pthread_cond_wait.S:185
>>>>> #1  0x00005555555cb777 in wait_cpu_work () at /home/alex/lsrc/qemu/qemu.git/linux-user/main.c:155
>>>>> #2  0x00005555555a0cee in wait_safe_cpu_work () at /home/alex/lsrc/qemu/qemu.git/cpu-exec-common.c:87
>>>>> #3  0x00005555555cb8fe in cpu_exec_end (cpu=0x555555bb67e0) at /home/alex/lsrc/qemu/qemu.git/linux-user/main.c:222
>>>>> #4  0x00005555555cc7a7 in cpu_loop (env=0x555555bbea58) at /home/alex/lsrc/qemu/qemu.git/linux-user/main.c:749
>>>>> #5  0x00005555555db0b2 in clone_func (arg=0x7fffffffc9c0) at /home/alex/lsrc/qemu/qemu.git/linux-user/syscall.c:5424
>>>>> #6  0x00007ffff6bed6fa in start_thread (arg=0x7ffefeca7700) at pthread_create.c:333
>>>>> #7  0x00007ffff6923b5d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:109
>>>>>
>>>>> <a bunch of other threads doing the same and then...>
>>>>>
>>>>> Thread 3 (Thread 0x7ffff7f38700 (LWP 2904)):
>>>>> #0  0x00005555555faf5d in safe_syscall_base ()
>>>>> #1  0x00005555555cfeaf in safe_futex (uaddr=0x7ffff528a0a4, op=128, val=1, timeout=0x0, uaddr2=0x0, val3=-162668384)
>>>>>     at /home/alex/lsrc/qemu/qemu.git/linux-user/syscall.c:706
>>>>> #2  0x00005555555dd7cc in do_futex (uaddr=4132298916, op=128, val=1, timeout=0, uaddr2=0, val3=-162668384)
>>>>>     at /home/alex/lsrc/qemu/qemu.git/linux-user/syscall.c:6246
>>>>> #3  0x00005555555e8cdb in do_syscall (cpu_env=0x555555a81118, num=240, arg1=-162668380, arg2=128, arg3=1, arg4=0, arg5=0, arg6=-162668384,
>>>>>     arg7=0, arg8=0) at /home/alex/lsrc/qemu/qemu.git/linux-user/syscall.c:10642
>>>>> #4  0x00005555555cd20e in cpu_loop (env=0x555555a81118) at /home/alex/lsrc/qemu/qemu.git/linux-user/main.c:883
>>>>> #5  0x00005555555db0b2 in clone_func (arg=0x7fffffffc9c0) at /home/alex/lsrc/qemu/qemu.git/linux-user/syscall.c:5424
>>>>> #6  0x00007ffff6bed6fa in start_thread (arg=0x7ffff7f38700) at pthread_create.c:333
>>>>> #7  0x00007ffff6923b5d in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:109
>>>>>
>>>>> So everything is stalled awaiting this thread waking up and draining
>>>>> its queue. So for linux-user I think we need some mechanism to kick
>>>>> these syscalls which I assume means throwing a signal at it.
>>>> Nice catch! How did you get it?
>>> Running pigz (armhf, debian) to compress stuff.
>>>
>>>> We always go through cpu_exec_end()
>>>> before serving a guest syscall and always go through cpu_exec_start()
>>>> before entering the guest code execution loop. If we always schedule
>>>> safe work on the current thread's queue then I think there's a way to
>>>> make it safe and avoid kicking syscalls.
>>> Not let the signals complete until safe work is done?
>> I'm thinking of waiting for completion of safe works in cpu_exec_start()
>> as well as in cpu_exec_end().
> I found a mistake in my code which causes deadlocks in my run of pigz.
> Could you also try running it after applying the following patch?
>
> diff --git a/linux-user/main.c b/linux-user/main.c
> index 6da3bb32186b..1dca55145c56 100644
> --- a/linux-user/main.c
> +++ b/linux-user/main.c
> @@ -214,7 +214,7 @@ static inline void cpu_exec_end(CPUState *cpu)
>      cpu->running = false;
>      tcg_pending_cpus--;
>      if (!tcg_pending_cpus) {
> -        pthread_cond_broadcast(&exclusive_cond);
> +        signal_cpu_work();
>      }
>      exclusive_idle();
>      flush_queued_work(cpu);
>
>

Or better this one:

 /* Make sure everything is in a consistent state for calling fork().  */
@@ -140,7 +139,6 @@ void fork_end(int child)
         pthread_mutex_init(&cpu_list_mutex, NULL);
         pthread_cond_init(&exclusive_cond, NULL);
         pthread_cond_init(&exclusive_resume, NULL);
-        pthread_cond_init(&work_cond, NULL);
         qemu_mutex_init(&tcg_ctx.tb_ctx.tb_lock);
         gdbserver_fork(thread_cpu);
     } else {
@@ -151,12 +149,12 @@ void fork_end(int child)
 
 void wait_cpu_work(void)
 {
-    pthread_cond_wait(&work_cond, &exclusive_lock);
+    pthread_cond_wait(&exclusive_cond, &exclusive_lock);
 }
 
 void signal_cpu_work(void)
 {
-    pthread_cond_broadcast(&work_cond);
+    pthread_cond_broadcast(&exclusive_cond);
 }
 
 /* Wait for pending exclusive operations to complete.  The exclusive lock
@@ -214,7 +212,7 @@ static inline void cpu_exec_end(CPUState *cpu)
     cpu->running = false;
     tcg_pending_cpus--;
     if (!tcg_pending_cpus) {
-        pthread_cond_broadcast(&exclusive_cond);
+        signal_cpu_work();
     }
     exclusive_idle();
     flush_queued_work(cpu);

Thanks,
Sergey
diff mbox

Patch

diff --git a/linux-user/main.c b/linux-user/main.c
index 6da3bb32186b..aba550b69611 100644
--- a/linux-user/main.c
+++ b/linux-user/main.c
@@ -111,7 +111,6 @@  static pthread_mutex_t cpu_list_mutex =
PTHREAD_MUTEX_INITIALIZER;
 static pthread_mutex_t exclusive_lock = PTHREAD_MUTEX_INITIALIZER;
 static pthread_cond_t exclusive_cond = PTHREAD_COND_INITIALIZER;
 static pthread_cond_t exclusive_resume = PTHREAD_COND_INITIALIZER;
-static pthread_cond_t work_cond = PTHREAD_COND_INITIALIZER;
 static bool exclusive_pending;