Re: [PATCH 1/1] fork, vhost: Use CLONE_THREAD to fix freezer/ps regression

["Eric W. Biederman" <ebiederm@xmission.com>]

Oleg Nesterov <oleg@redhat.com> writes:

> Hi Mike,
>
> sorry, but somehow I can't understand this patch...
>
> I'll try to read it with a fresh head on Weekend, but for example,
>
> On 06/01, Mike Christie wrote:
>>
>>  static int vhost_task_fn(void *data)
>>  {
>>  	struct vhost_task *vtsk = data;
>> -	int ret;
>> +	bool dead = false;
>> +
>> +	for (;;) {
>> +		bool did_work;
>> +
>> +		/* mb paired w/ vhost_task_stop */
>> +		if (test_bit(VHOST_TASK_FLAGS_STOP, &vtsk->flags))
>> +			break;
>> +
>> +		if (!dead && signal_pending(current)) {
>> +			struct ksignal ksig;
>> +			/*
>> +			 * Calling get_signal will block in SIGSTOP,
>> +			 * or clear fatal_signal_pending, but remember
>> +			 * what was set.
>> +			 *
>> +			 * This thread won't actually exit until all
>> +			 * of the file descriptors are closed, and
>> +			 * the release function is called.
>> +			 */
>> +			dead = get_signal(&ksig);
>> +			if (dead)
>> +				clear_thread_flag(TIF_SIGPENDING);
>
> this can't be right or I am totally confused.
>
> Another signal_wake_up() can come right after clear(SIGPENDING).

Technically yes.

However please not that prepare_signal does:
	if (signal->flags & SIGNAL_GROUP_EXIT)
		return false;

In general it is wrong to receive or attempt to process a signal
after task death has been decided.

Strictly speaking that doesn't cover de_thread, and coredumping
but still receiving any kind of signal at that point is rare
and generally wrong behavior.

Beyond that clearing TIF_SIGPENDING is just an optimization so
the thread can sleep in schedule and not spin.

> Again, I'll try to re-read this patch, but let me ask anyway...
>
> Do we have a plan B? I mean... iirc you have mentioned that you can
> change these code paths to do something like
>
> 	if (killed)
> 		tell_the_drivers_that_all_callbacks_will_fail();
>
>
> so that vhost_worker() can exit after get_signal() returns SIGKILL.
>
> Probably I misunderstood you, but it would be nice to avoid the changes
> in coredump/etc code just to add a temporary (iiuc!) fix.

One saving grace with the the vhost code is that you need to open
device nodes that normally have root-only permissions.

If we are willing to allow races in process shutdown to cause leaks I
think we can do something better, and put the burden of work on vhost
layer.

I will follow up with a patch doing that.

Eric