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;
Date:   Tue, 21 Jul 2020 17:00:24 +0200
From:   Michal Hocko <mhocko@kernel.org>
To:     Chris Down <chris@chrisdown.name>
Cc:     linux-mm@kvack.org, LKML <linux-kernel@vger.kernel.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Tim Chen <tim.c.chen@linux.intel.com>
Subject: Re: [RFC PATCH] mm: silence soft lockups from unlock_page
Message-ID: <20200721150024.GM4061@dhcp22.suse.cz>
References: <20200721063258.17140-1-mhocko@kernel.org>
 <20200721141749.GA742741@chrisdown.name>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200721141749.GA742741@chrisdown.name>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue 21-07-20 15:17:49, Chris Down wrote:
> I understand the pragmatic considerations here, but I'm quite concerned
> about the maintainability and long-term ability to reason about a patch like
> this.  For example, how do we know when this patch is safe to remove? Also,
> what other precedent does this set for us covering for poor userspace
> behaviour?
> 
> Speaking as a systemd maintainer, if udev could be doing something better on
> these machines, we'd be more than receptive to help fix it. In general I am
> against explicit watchdog tweaking here because a.) there's potential to
> mask other problems, and b.) it seems like the kind of one-off trivia nobody
> is going to remember exists when doing complex debugging in future.
> 
> Is there anything preventing this being remedied in udev, instead of the
> kernel?

Yes, I believe that there is a configuration to cap the maximum number
of workers. This is not my area but my understanding is that the maximum
is tuned based on available memory and/or cpus. We have been hit byt
this quite heavily on SLES. Maybe newer version of systemd have a better
tuning.

But, it seems that udev is just a messenger here. There is nothing
really fundamentally udev specific in the underlying problem unless I
miss something. It is quite possible that this could be triggered by
other userspace which happens to fire many workers at the same time and
condending on a shared page.

Not that I like this workaround in the first place but it seems that the
existing code allows very long wait chains and !PREEMPT kernels simply
do not have any scheduling point for a long time potentially. I believe
we should focus on that even if the systemd as the current trigger can
be tuned better. I do not insist on this patch, hence RFC, but I am
simply not seeing a much better, yet not convoluted, solution.

-- 
Michal Hocko
SUSE Labs