Re: [Qemu-devel] [PATCH v4 06/10] iohandler: switch to GPollFD

[Laszlo Ersek]

On 02/22/13 18:33, Jan Kiszka wrote:
> On 2013-02-20 11:28, Stefan Hajnoczi wrote:
>> Convert iohandler_select_fill() and iohandler_select_poll() to use
>> GPollFD instead of rfds/wfds/xfds.
> 
> Since this commmit, I'm getting QEMU lock-ups, apparently slirp is
> involved (the Linux guest tries to start its network at this point):
> 
> (gdb) thread apply all bt
> 
> Thread 3 (Thread 0x7fffed0e3700 (LWP 26788)):
> #0  __lll_lock_wait () at 
> ../nptl/sysdeps/unix/sysv/linux/x86_64/lowlevellock.S:136
> #1  0x00007ffff44c7294 in _L_lock_999 () from /lib64/libpthread.so.0
> #2  0x00007ffff44c70aa in __pthread_mutex_lock (mutex=0x5555560afcc0) at 
> pthread_mutex_lock.c:61
> #3  0x00005555558945e9 in qemu_mutex_lock (mutex=<value optimized out>) at 
> /data/qemu/util/qemu-thread-posix.c:57
> #4  0x00005555557d9da5 in kvm_cpu_exec (env=0x55555689c9f0) at 
> /data/qemu/kvm-all.c:1564
> #5  0x0000555555780091 in qemu_kvm_cpu_thread_fn (arg=0x55555689c9f0) at 
> /data/qemu/cpus.c:759
> #6  0x00007ffff44c4a3f in start_thread (arg=0x7fffed0e3700) at 
> pthread_create.c:297
> #7  0x00007ffff2fb871d in clone () at 
> ../sysdeps/unix/sysv/linux/x86_64/clone.S:112
> #8  0x0000000000000000 in ?? ()
> 
> Thread 2 (Thread 0x7fffed8e4700 (LWP 26787)):
> #0  sem_timedwait () at 
> ../nptl/sysdeps/unix/sysv/linux/x86_64/sem_timedwait.S:103
> #1  0x0000555555894a93 in qemu_sem_timedwait (sem=0x555556090020, ms=<value 
> optimized out>) at /data/qemu/util/qemu-thread-posix.c:237
> #2  0x000055555575116e in worker_thread (unused=<value optimized out>) at 
> /data/qemu/thread-pool.c:88
> #3  0x00007ffff44c4a3f in start_thread (arg=0x7fffed8e4700) at 
> pthread_create.c:297
> #4  0x00007ffff2fb871d in clone () at 
> ../sysdeps/unix/sysv/linux/x86_64/clone.S:112
> #5  0x0000000000000000 in ?? ()
> 
> Thread 1 (Thread 0x7ffff7fab760 (LWP 26784)):
> #0  0x00007ffff44cc763 in recvfrom () at ../sysdeps/unix/syscall-template.S:82
> #1  0x000055555574b67d in recvfrom (so=0x555556bd2f50) at 
> /usr/include/bits/socket2.h:77
> #2  sorecvfrom (so=0x555556bd2f50) at /data/qemu/slirp/socket.c:498
> #3  0x000055555574a160 in slirp_pollfds_poll (pollfds=0x555556511240, 
> select_error=0) at /data/qemu/slirp/slirp.c:619
> #4  0x000055555570ec99 in main_loop_wait (nonblocking=<value optimized out>) 
> at /data/qemu/main-loop.c:514
> #5  0x000055555577821d in main_loop (argc=<value optimized out>, argv=<value 
> optimized out>, envp=<value optimized out>) at /data/qemu/vl.c:2002
> #6  main (argc=<value optimized out>, argv=<value optimized out>, envp=<value 
> optimized out>) at /data/qemu/vl.c:4334
> 
> Thread 1 blocks in recvfrom, not returning, not releasing the global lock.
> 
> Any idea?

Well I guess recvfrom() shouldn't block in slirp / thread 1, so maybe
slirp_pollfds_poll() finds readiness where it shouldn't -- we possibly
shouldn't even call sorecvfrom().

sorecvfrom() belongs to the UDP branch in slirp_pollfds_poll().

Could this be related to the change we discussed in
<http://thread.gmane.org/gmane.comp.emulators.qemu/192801/focus=193181>?

I guess trace calls would be handy...

FWIW I find it interesting that slirp doesn't hang after patch 05/10
(which is the slirp conversion) but here. This patch (06/10) converts
qemu iohandler. It looks as if qemu_iohandler_poll(), called just before
slirp_pollfds_poll() in main_loop_wait(), "stole" data from slirp. Mixup
between file descriptors?

Laszlo