Re: [PATCH 2/8] kernel: unmask SIGSTOP for IO threads

[Stefan Metzmacher <metze@samba.org>]

Am 26.03.21 um 16:29 schrieb Jens Axboe:
> On 3/26/21 9:23 AM, Stefan Metzmacher wrote:
>> Am 26.03.21 um 16:01 schrieb Jens Axboe:
>>> On 3/26/21 7:48 AM, Oleg Nesterov wrote:
>>>> Jens, sorry, I got lost :/
>>>
>>> Let's bring you back in :-)
>>>
>>>> On 03/25, Jens Axboe wrote:
>>>>>
>>>>> With IO threads accepting signals, including SIGSTOP,
>>>>
>>>> where can I find this change? Looks like I wasn't cc'ed...
>>>
>>> It's this very series.
>>>
>>>>> unmask the
>>>>> SIGSTOP signal from the default blocked mask.
>>>>>
>>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>>>> ---
>>>>>  kernel/fork.c | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/kernel/fork.c b/kernel/fork.c
>>>>> index d3171e8e88e5..d5a40552910f 100644
>>>>> --- a/kernel/fork.c
>>>>> +++ b/kernel/fork.c
>>>>> @@ -2435,7 +2435,7 @@ struct task_struct *create_io_thread(int (*fn)(void *), void *arg, int node)
>>>>>  	tsk = copy_process(NULL, 0, node, &args);
>>>>>  	if (!IS_ERR(tsk)) {
>>>>>  		sigfillset(&tsk->blocked);
>>>>> -		sigdelsetmask(&tsk->blocked, sigmask(SIGKILL));
>>>>> +		sigdelsetmask(&tsk->blocked, sigmask(SIGKILL)|sigmask(SIGSTOP));
>>>>
>>>> siginitsetinv(blocked, sigmask(SIGKILL)|sigmask(SIGSTOP)) but this is minor.
>>>
>>> Ah thanks.
>>>
>>>> To remind, either way this is racy and can't really help.
>>>>
>>>> And if "IO threads accepting signals" then I don't understand why. Sorry,
>>>> I must have missed something.
>>>
>>> I do think the above is a no-op at this point, and we can probably just
>>> kill it. Let me double check, hopefully we can just remove this blocked
>>> part.
>>
>> Is this really correct to drop in your "kernel: stop masking signals in create_io_thread()"
>> commit?
>>
>> I don't assume signals wanted by userspace should potentially handled in an io_thread...
>> e.g. things set with fcntl(fd, F_SETSIG,) used together with F_SETLEASE?
> 
> I guess we do actually need it, if we're not fiddling with
> wants_signal() for them. To quell Oleg's concerns, we can just move it
> to post dup_task_struct(), that should eliminate any race concerns
> there.

If that one is racy, don' we better also want this one?
https://lore.kernel.org/io-uring/438b738c1e4827a7fdfe43087da88bbe17eedc72.1616197787.git.metze@samba.org/T/#u

And clear tsk->pf_io_worker ?

metze