Re: [RFC][PATCH v2 5/5] sched: User Mode Concurency Groups

From: Peter Zijlstra <peterz@infradead.org>
To: mingo@redhat.com, tglx@linutronix.de, juri.lelli@redhat.com,
	vincent.guittot@linaro.org, dietmar.eggemann@arm.com,
	rostedt@goodmis.org, bsegall@google.com, mgorman@suse.de,
	bristot@redhat.com
Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	linux-api@vger.kernel.org, x86@kernel.org, pjt@google.com,
	posk@google.com, avagin@google.com, jannh@google.com,
	tdelisle@uwaterloo.ca, mark.rutland@arm.com, posk@posk.io
Subject: Re: [RFC][PATCH v2 5/5] sched: User Mode Concurency Groups
Date: Fri, 21 Jan 2022 12:47:58 +0100	[thread overview]
Message-ID: <20220121114758.GF20638@worktop.programming.kicks-ass.net> (raw)
In-Reply-To: <20220120160822.914418096@infradead.org>

On Thu, Jan 20, 2022 at 04:55:22PM +0100, Peter Zijlstra wrote:

> +SYSCALL_DEFINE2(umcg_wait, u32, flags, u64, timo)
> +{
> +	struct task_struct *tsk = current;
> +	struct umcg_task __user *self = READ_ONCE(tsk->umcg_task);
> +	bool worker = tsk->flags & PF_UMCG_WORKER;
> +	int ret;
> +
> +	if (!self || flags)
> +		return -EINVAL;
> +
> +	if (worker) {
> +		tsk->flags &= ~PF_UMCG_WORKER;
> +		if (timo)
> +			return -ERANGE;
> +	}
> +
> +	/* see umcg_sys_{enter,exit}() syscall exceptions */
> +	ret = umcg_pin_pages();
> +	if (ret)
> +		goto unblock;
> +
> +	/*
> +	 * Clear UMCG_TF_COND_WAIT *and* check state == RUNNABLE.
> +	 */
> +	ret = umcg_update_state(tsk, self, UMCG_TASK_RUNNABLE, UMCG_TASK_RUNNABLE);
> +	if (ret)
> +		goto unpin;
> +
> +	ret = umcg_wake_next(tsk, self);
> +	if (ret)
> +		goto unpin;
> +
> +	if (worker) {
> +		/*
> +		 * If this fails it is possible ::next_tid is already running
> +		 * while this task is not going to block. This violates our
> +		 * constraints.
> +		 *
> +		 * That said, pretty much the only way to make this fail is by
> +		 * force munmap()'ing things. In which case one is most welcome
> +		 * to the pieces.
> +		 */
> +		ret = umcg_enqueue_and_wake(tsk);
> +		if (ret)
> +			goto unpin;
> +	}
> +
> +	umcg_unpin_pages();
> +
> +	ret = umcg_wait(timo);
> +	switch (ret) {
> +	case 0:		/* all done */
> +	case -EINTR:	/* umcg_notify_resume() will continue the wait */

So I was playing with the whole worker timeout thing last night and
realized this is broken. If we get a signal while we have a timeout, the
timeout gets lost.

I think the easiest solution is to have umcg_notify_resume() also resume
the timeout, but the first pass of that was yuck, so I need to try
again.

Related, by moving the whole enqueue-and-wake thing into the timeout, we
get more 'fun' failure cases :-(

Oh well..

> +		ret = 0;
> +		break;
> +
> +	default:
> +		goto unblock;
> +	}
> +out:
> +	if (worker)
> +		tsk->flags |= PF_UMCG_WORKER;
> +	return ret;
> +
> +unpin:
> +	umcg_unpin_pages();
> +unblock:
> +	umcg_update_state(tsk, self, UMCG_TASK_RUNNABLE, UMCG_TASK_RUNNING);
> +	goto out;
> +}