Hello,
one of openSUSE Leap 42.2 users encountered (repeatedly) a warning
[ 4057.170653] WARNING: CPU: 1 PID: 3471 at ../kernel/sched/core.c:7913
__might_sleep+0x76/0x80()
[ 4057.170661] do not call blocking ops when !TASK_RUNNING; state=1 set at
[<ffffffff810c25ab>] prepare_to_wait+0x2b/0x80
with stack
[ 4057.170786] [<ffffffff81019e69>] dump_trace+0x59/0x320
[ 4057.170789] [<ffffffff8101a22a>] show_stack_log_lvl+0xfa/0x180
[ 4057.170792] [<ffffffff8101afd1>] show_stack+0x21/0x40
[ 4057.170798] [<ffffffff81327657>] dump_stack+0x5c/0x85
[ 4057.170803] [<ffffffff8107e821>] warn_slowpath_common+0x81/0xb0
[ 4057.170806] [<ffffffff8107e89c>] warn_slowpath_fmt+0x4c/0x50
[ 4057.170809] [<ffffffff810a3106>] __might_sleep+0x76/0x80
[ 4057.170814] [<ffffffff816071ac>] mutex_lock+0x1c/0x38
[ 4057.170822] [<ffffffffa0cfb477>] vmci_qpair_produce_free_space+0x97/0xd0
[vmw_vmci]
[ 4057.170848] [<ffffffffa0d10d36>] vsock_stream_sendmsg+0x1f6/0x320 [vsock]
[ 4057.170855] [<ffffffff814f6fb0>] sock_sendmsg+0x30/0x40
[ 4057.170859] [<ffffffff814f7039>] sock_write_iter+0x79/0xd0
[ 4057.170864] [<ffffffff81204d49>] __vfs_write+0xa9/0xf0
[ 4057.170867] [<ffffffff8120534d>] vfs_write+0x9d/0x190
[ 4057.170870] [<ffffffff81206012>] SyS_write+0x42/0xa0
[ 4057.170873] [<ffffffff816093f2>] entry_SYSCALL_64_fastpath+0x16/0x71
The kernel is 4.4.27 but it already has commit f7f9b5e7f8ec ("AF_VSOCK:
Shrink the area influenced by prepare_to_wait") applied. The issue comes
from this part of vsock_stream_sendmsg():
prepare_to_wait(sk_sleep(sk), &wait, TASK_INTERRUPTIBLE);
while (vsock_stream_has_space(vsk) == 0 &&
sk->sk_err == 0 &&
!(sk->sk_shutdown & SEND_SHUTDOWN) &&
!(vsk->peer_shutdown & RCV_SHUTDOWN)) {
where vsock_stream_has_space() can sleep:
vsock_stream_has_space
vmci_transport_stream_has_space
vmci_qpair_produce_free_space
qp_lock
qp_acquire_queue_mutex
mutex_lock
but this is not allowed between prepare_to_wait() and either the actual
waiting or finish_wait().
I tried to think about a solution but there doesn't seem to be an easy
way to fix this in vmw_stream_sendmsg() as moving prepare_to_wait()
inside the loop would result in missed wake-ups (that was the problem
with the original fix); IMHO the right way to resolve the issue would be
rewriting the vmci queue pair code to allow performing the has_space()
check without taking a mutex.
Michal Kubecek
