Date:   Tue, 14 Dec 2021 00:08:05 -0800
From:   Christoph Hellwig <hch@infradead.org>
To:     Davidlohr Bueso <dave@stgolabs.net>
Cc:     Christoph Hellwig <hch@infradead.org>, axboe@kernel.dk,
        bigeasy@linutronix.de, tglx@linutronix.de,
        linux-block@vger.kernel.org, linux-kernel@vger.kernel.org,
        Davidlohr Bueso <dbueso@suse.de>
Subject: Re: [PATCH] blk-mq: make synchronous hw_queue runs RT friendly
Message-ID: <YbhQ5b2qZN0Jpk5i@infradead.org>
References: <20211213054425.28121-1-dave@stgolabs.net>
 <YbdFeHVnQbT0E5kR@infradead.org>
 <20211213190529.j54qikq7mk5zuc3o@offworld>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20211213190529.j54qikq7mk5zuc3o@offworld>
Precedence: bulk

On Mon, Dec 13, 2021 at 11:05:29AM -0800, Davidlohr Bueso wrote:
> On Mon, 13 Dec 2021, Christoph Hellwig wrote:
> > But more importantly:  why isn't migrate_disable/enable doing the right
> > thing for !PREEMPT_RT to avoid this mess?
> 
> Please see Peter's description of the situation in af449901b84.

That explains why migrate_disable is a bad idea in PREEMPT_RT, not why it
can't do something sensible for !PREEMPT_RT…

> 
> While I'm not at all a fan of sprinkling migrate_disabling around code,
> I didn't want to add any overhead for the common case. If this, however,
> were not an issue (if most cases are async runs, for example) the ideal
> solution I think would be to just pin current to the hctx->cpumask.

sync running is the performance case.