Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 29 Jan 2019 13:50:13 -0500
From:   Josef Bacik <josef@toxicpanda.com>
To:     Andrea Righi <righi.andrea@gmail.com>
Cc:     Vivek Goyal <vgoyal@redhat.com>,
        Josef Bacik <josef@toxicpanda.com>, Tejun Heo <tj@kernel.org>,
        Li Zefan <lizefan@huawei.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Jens Axboe <axboe@kernel.dk>, Dennis Zhou <dennis@kernel.org>,
        cgroups@vger.kernel.org, linux-block@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH 0/3] cgroup: fsio throttle controller
Message-ID: <20190129185012.jieed26ddcbz7jmb@MacBook-Pro-91.local>
References: <20190118103127.325-1-righi.andrea@gmail.com>
 <20190118163530.w5wpzpjkcnkektsp@macbook-pro-91.dhcp.thefacebook.com>
 <20190118184403.GB1535@xps-13>
 <20190118194652.gg5j2yz3h2llecpj@macbook-pro-91.dhcp.thefacebook.com>
 <20190119100827.GA1630@xps-13>
 <20190121214715.GA27713@redhat.com>
 <20190128174129.GB8272@xps-13>
 <20190128192620.GB10240@redhat.com>
 <20190129183938.GA2960@xps-13>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20190129183938.GA2960@xps-13>
User-Agent: NeoMutt/20180716
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Tue, Jan 29, 2019 at 07:39:38PM +0100, Andrea Righi wrote:
> On Mon, Jan 28, 2019 at 02:26:20PM -0500, Vivek Goyal wrote:
> > On Mon, Jan 28, 2019 at 06:41:29PM +0100, Andrea Righi wrote:
> > > Hi Vivek,
> > > 
> > > sorry for the late reply.
> > > 
> > > On Mon, Jan 21, 2019 at 04:47:15PM -0500, Vivek Goyal wrote:
> > > > On Sat, Jan 19, 2019 at 11:08:27AM +0100, Andrea Righi wrote:
> > > > 
> > > > [..]
> > > > > Alright, let's skip the root cgroup for now. I think the point here is
> > > > > if we want to provide sync() isolation among cgroups or not.
> > > > > 
> > > > > According to the manpage:
> > > > > 
> > > > >        sync()  causes  all  pending  modifications  to filesystem metadata and cached file data to be
> > > > >        written to the underlying filesystems.
> > > > > 
> > > > > And:
> > > > >        According to the standard specification (e.g., POSIX.1-2001), sync() schedules the writes, but
> > > > >        may  return  before  the actual writing is done.  However Linux waits for I/O completions, and
> > > > >        thus sync() or syncfs() provide the same guarantees as fsync called on every file in the  sys‐
> > > > >        tem or filesystem respectively.
> > > > > 
> > > > > Excluding the root cgroup, do you think a sync() issued inside a
> > > > > specific cgroup should wait for I/O completions only for the writes that
> > > > > have been generated by that cgroup?
> > > > 
> > > > Can we account I/O towards the cgroup which issued "sync" only if write
> > > > rate of sync cgroup is higher than cgroup to which page belongs to. Will
> > > > that solve problem, assuming its doable?
> > > 
> > > Maybe this would mitigate the problem, in part, but it doesn't solve it.
> > > 
> > > The thing is, if a dirty page belongs to a slow cgroup and a fast cgroup
> > > issues "sync", the fast cgroup needs to wait a lot, because writeback is
> > > happening at the speed of the slow cgroup.
> > 
> > Hi Andrea,
> > 
> > But that's true only for I/O which has already been submitted to block
> > layer, right? Any new I/O yet to be submitted could still be attributed
> > to faster cgroup requesting sync.
> 
> Right. If we could bump up the new I/O yet to be submitted I think we
> could effectively prevent the priority inversion problem (the ongoing
> writeback I/O should be negligible).
> 
> > 
> > Until and unless cgroups limits are absurdly low, it should not take very
> > long for already submitted I/O to finish. If yes, then in practice, it
> > might not be a big problem?
> 
> I was actually doing my tests with a very low limit (1MB/s both for rbps
> and wbps), but this shows the problem very well I think.
> 
> Here's what I'm doing:
> 
>  [ slow cgroup (1Mbps read/write) ]
> 
>    $ cat /sys/fs/cgroup/unified/cg1/io.max
>    259:0 rbps=1048576 wbps=1048576 riops=max wiops=max
>    $ cat /proc/self/cgroup
>    0::/cg1
> 
>    $ fio --rw=write --bs=1M --size=32M --numjobs=16 --name=writer --time_based --runtime=30
> 
>  [ fast cgroup (root cgroup, no limitation) ]
> 
>    # cat /proc/self/cgroup
>    0::/
> 
>    # time sync
>    real	9m32,618s
>    user	0m0,000s
>    sys	0m0,018s
> 
> With this simple test I can easily trigger hung task timeout warnings
> and make the whole system totally sluggish (even the processes running
> in the root cgroup).
> 
> When fio ends, writeback is still taking forever to complete, as you can
> see by the insane amount that sync takes to complete.
> 

Yeah sync() needs to be treated differently, but its kind of special too.  We
don't want slow to run sync() and backup fast doing sync() because we make all
of the io go based on the submitting cgroup.  The problem here is we don't know
who's more important until we get to the blk cgroup layer, and even then
sometimes we can't tell (different hierarchies would make this tricky with
io.weight or io.latency).

We could treat it like REQ_META and just let everything go through and back
charge.  This feels like a way for the slow group to cheat though, unless we
just throttle the shit out of him before returning to user space.  I'll have to
think about this some more.  Thanks,

Josef