Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH 2/7] io_uring: handle signals for IO threads like a normal
 thread
To:     "Eric W. Biederman" <ebiederm@xmission.com>
Cc:     io-uring@vger.kernel.org, torvalds@linux-foundation.org,
        metze@samba.org, oleg@redhat.com, linux-kernel@vger.kernel.org
References: <20210326155128.1057078-1-axboe@kernel.dk>
 <20210326155128.1057078-3-axboe@kernel.dk> <m1wntty7yn.fsf@fess.ebiederm.org>
 <106a38d3-5a5f-17fd-41f7-890f5e9a3602@kernel.dk>
 <m1k0ptv9kj.fsf@fess.ebiederm.org>
 <01058178-dd66-1bff-4d74-5ff610817ed6@kernel.dk>
 <m18s69v8zb.fsf@fess.ebiederm.org>
From:   Jens Axboe <axboe@kernel.dk>
Message-ID: <7a71da2f-ca39-6bbf-28c1-bcc2eec43943@kernel.dk>
Date:   Fri, 26 Mar 2021 16:38:16 -0600
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.10.0
MIME-Version: 1.0
In-Reply-To: <m18s69v8zb.fsf@fess.ebiederm.org>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 3/26/21 4:35 PM, Eric W. Biederman wrote:
> Jens Axboe <axboe@kernel.dk> writes:
> 
>> On 3/26/21 4:23 PM, Eric W. Biederman wrote:
>>> Jens Axboe <axboe@kernel.dk> writes:
>>>
>>>> On 3/26/21 2:29 PM, Eric W. Biederman wrote:
>>>>> Jens Axboe <axboe@kernel.dk> writes:
>>>>>
>>>>>> We go through various hoops to disallow signals for the IO threads, but
>>>>>> there's really no reason why we cannot just allow them. The IO threads
>>>>>> never return to userspace like a normal thread, and hence don't go through
>>>>>> normal signal processing. Instead, just check for a pending signal as part
>>>>>> of the work loop, and call get_signal() to handle it for us if anything
>>>>>> is pending.
>>>>>>
>>>>>> With that, we can support receiving signals, including special ones like
>>>>>> SIGSTOP.
>>>>>>
>>>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>>>>> ---
>>>>>>  fs/io-wq.c    | 24 +++++++++++++++++-------
>>>>>>  fs/io_uring.c | 12 ++++++++----
>>>>>>  2 files changed, 25 insertions(+), 11 deletions(-)
>>>>>>
>>>>>> diff --git a/fs/io-wq.c b/fs/io-wq.c
>>>>>> index b7c1fa932cb3..3e2f059a1737 100644
>>>>>> --- a/fs/io-wq.c
>>>>>> +++ b/fs/io-wq.c
>>>>>> @@ -16,7 +16,6 @@
>>>>>>  #include <linux/rculist_nulls.h>
>>>>>>  #include <linux/cpu.h>
>>>>>>  #include <linux/tracehook.h>
>>>>>> -#include <linux/freezer.h>
>>>>>>  
>>>>>>  #include "../kernel/sched/sched.h"
>>>>>>  #include "io-wq.h"
>>>>>> @@ -503,10 +502,16 @@ static int io_wqe_worker(void *data)
>>>>>>  		if (io_flush_signals())
>>>>>>  			continue;
>>>>>>  		ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
>>>>>> -		if (try_to_freeze() || ret)
>>>>>> +		if (signal_pending(current)) {
>>>>>> +			struct ksignal ksig;
>>>>>> +
>>>>>> +			if (fatal_signal_pending(current))
>>>>>> +				break;
>>>>>> +			if (get_signal(&ksig))
>>>>>> +				continue;
>>>>>                         ^^^^^^^^^^^^^^^^^^^^^^
>>>>>
>>>>> That is wrong.  You are promising to deliver a signal to signal
>>>>> handler and them simply discarding it.  Perhaps:
>>>>>
>>>>> 			if (!get_signal(&ksig))
>>>>>                         	continue;
>>>>> 			WARN_ON(!sig_kernel_stop(ksig->sig));
>>>>>                         break;
>>>>
>>>> Thanks, updated.
>>>
>>> Gah.  Kill the WARN_ON.
>>>
>>> I was thinking "WARN_ON(!sig_kernel_fatal(ksig->sig));"
>>> The function sig_kernel_fatal does not exist.
>>>
>>> Fatal is the state that is left when a signal is neither
>>> ignored nor a stop signal, and does not have a handler.
>>>
>>> The rest of the logic still works.
>>
>> I've just come to the same conclusion myself after testing it.
>> Of the 3 cases, most of them can do the continue, but doesn't
>> really matter with the way the loop is structured. Anyway, looks
>> like this now:
> 
> This idiom in the code:
>> +		if (signal_pending(current)) {
>> +			struct ksignal ksig;
>> +
>> +			if (fatal_signal_pending(current))
>> +				break;
>> +			if (!get_signal(&ksig))
>> +				continue;
>>  }
> 
> Needs to be:
>> +		if (signal_pending(current)) {
>> +			struct ksignal ksig;
>> +
>> +			if (!get_signal(&ksig))
>> +				continue;
>> +			break;
>>  }
> 
> Because any signal returned from get_signal is fatal in this case.
> It might make sense to "WARN_ON(ksig->ka.sa.sa_handler != SIG_DFL)".
> As the io workers don't handle that case.
> 
> It won't happen because you have everything blocked.
>
> The extra fatal_signal_pending(current) logic is just confusing in this
> case.

OK good point, and follows the same logic even if it won't make a
difference in my case. I'll make the change.

-- 
Jens Axboe