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 3/3] task_work: use TIF_TASKWORK if available
To:     Thomas Gleixner <tglx@linutronix.de>,
        Oleg Nesterov <oleg@redhat.com>
Cc:     linux-kernel@vger.kernel.org, io-uring@vger.kernel.org,
        peterz@infradead.org
References: <20201001194208.1153522-1-axboe@kernel.dk>
 <20201001194208.1153522-4-axboe@kernel.dk>
 <20201002151415.GA29066@redhat.com> <871rigejb8.fsf@nanos.tec.linutronix.de>
From:   Jens Axboe <axboe@kernel.dk>
Message-ID: <4c9dbcc4-cae7-c7ad-8066-31d49239750a@kernel.dk>
Date:   Fri, 2 Oct 2020 09:52:24 -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: <871rigejb8.fsf@nanos.tec.linutronix.de>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 10/2/20 9:31 AM, Thomas Gleixner wrote:
> On Fri, Oct 02 2020 at 17:14, Oleg Nesterov wrote:
>> Heh. To be honest I don't really like 1-2 ;)
> 
> I do not like any of this :)
> 
>> So I think that if we are going to add TIF_TASKWORK we should generalize
>> this logic and turn it into TIF_NOTIFY_SIGNAL. Similar to TIF_NOTIFY_RESUME
>> but implies signal_pending().
>>
>> IOW, something like
>>
>> 	void set_notify_signal(task)
>> 	{
>> 		if (!test_and_set_tsk_thread_flag(task, TIF_NOTIFY_SIGNAL)) {
>> 			if (!wake_up_state(task, TASK_INTERRUPTIBLE))
>> 				kick_process(t);
>> 		}
>> 	}
>>
>> 	// called by exit_to_user_mode_loop() if ti_work & _TIF_NOTIFY_SIGNAL
>> 	void tracehook_notify_signal(regs)
>> 	{
>> 		clear_thread_flag(TIF_NOTIFY_SIGNAL);
>> 		smp_mb__after_atomic();
>> 		if (unlikely(current->task_works))
>> 			task_work_run();
>> 	}
>>
>> This way task_work_run() doesn't need to clear TIF_NOTIFY_SIGNAL and it can
>> have more users.
> 
> I think it's fundamentaly wrong that we have several places and several
> flags which handle task_work_run() instead of having exactly one place
> and one flag.

I don't disagree with that. I know it's not happening in this series, but
if we to the TIF_NOTIFY_SIGNAL route and get all archs supporting that,
then we can kill the signal and notify resume part of running task_work.
And that leaves us with exactly one place that runs it.

So we can potentially improve the current situation in that regard.

-- 
Jens Axboe