On 3/19/21 6:25 PM, Stefan Metzmacher wrote:
> Am 20.03.21 um 00:46 schrieb Jens Axboe:
>> On Mar 19, 2021, at 5:27 PM, Stefan Metzmacher <metze@xxxxxxxxx> wrote:
>>>
>>> Hi Jens,
>>>
>>> as said before I found some problems related to
>>> the new io_threads together with signals.
>>>
>>> I applied the diff (at the end) to examples/io_uring-cp.c
>>> in order to run endless in order to give me time to
>>> look at /proc/...
>>>
>>> Trying to attach gdb --pid to the pid of the main process (thread group)
>>> it goes into an endless loop because it can't attach to the io_threads.
>>>
>>> Sending kill -STOP to the main pid causes the io_threads to spin cpu
>>> at 100%.
>>>
>>> Can you try to reproduce and fix it? Maybe same_thread_group() should not match?
>>
>> Definitely, I’ll go over this shortly and make sure we handle (and ignore) signals correctly.
>
> Thanks! Also a kill -9 to a io_thread kills the application.
OK, this I believe should take care of it - ignore STOP specifically for
PF_IO_WORKER, and disallow any signal sent to a worker.
diff --git a/kernel/signal.c b/kernel/signal.c
index ba4d1ef39a9e..b113bf647fb4 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -833,6 +833,9 @@ static int check_kill_permission(int sig, struct kernel_siginfo *info,
if (!valid_signal(sig))
return -EINVAL;
+ /* PF_IO_WORKER threads don't take any signals */
+ if (t->flags & PF_IO_WORKER)
+ return -EPERM;
if (!si_fromuser(info))
return 0;
@@ -2346,6 +2349,10 @@ static bool do_signal_stop(int signr)
t = current;
while_each_thread(current, t) {
+ /* don't STOP PF_IO_WORKER threads */
+ if (t->flags & PF_IO_WORKER)
+ continue;
+
/*
* Setting state to TASK_STOPPED for a group
* stop is always done with the siglock held,
--
Jens Axboe
