Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0
From:   "NeilBrown" <neilb@suse.de>
To:     "Dave Chinner" <david@fromorbit.com>
Cc:     "Andrew Morton" <akpm@linux-foundation.org>,
        "Mel Gorman" <mgorman@suse.de>,
        "Philipp Reisner" <philipp.reisner@linbit.com>,
        "Lars Ellenberg" <lars.ellenberg@linbit.com>,
        "Jan Kara" <jack@suse.com>,
        "Ryusuke Konishi" <konishi.ryusuke@gmail.com>,
        "Darrick J. Wong" <djwong@kernel.org>, linux-xfs@vger.kernel.org,
        linux-ext4@vger.kernel.org, linux-nilfs@vger.kernel.org,
        linux-mm@kvack.org, linux-fsdevel@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH 2/2] Remove bdi_congested() and wb_congested() and related
 functions
In-reply-to: <20211213050736.GS449541@dread.disaster.area>
References: <163936868317.23860.5037433897004720387.stgit@noble.brown>,
 <163936886727.23860.5245364396572576756.stgit@noble.brown>,
 <20211213050736.GS449541@dread.disaster.area>
Date:   Mon, 13 Dec 2021 18:04:10 +1100
Message-id: <163937905049.22433.10716750573737410875@noble.neil.brown.name>
Precedence: bulk

On Mon, 13 Dec 2021, Dave Chinner wrote:
> On Mon, Dec 13, 2021 at 03:14:27PM +1100, NeilBrown wrote:
> > These functions are no longer useful as the only bdis that report
> > congestion are in ceph, fuse, and nfs.  None of those bdis can be the
> > target of the calls in drbd, ext2, nilfs2, or xfs.
> > 
> > Removing the test on bdi_write_contested() in current_may_throttle()
> > could cause a small change in behaviour, but only when PF_LOCAL_THROTTLE
> > is set.
> > 
> > So replace the calls by 'false' and simplify the code - and remove the
> > functions.
> > 
> > Signed-off-by: NeilBrown <neilb@suse.de>
> ....
> > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> > index 631c5a61d89b..22f73b3e888e 100644
> > --- a/fs/xfs/xfs_buf.c
> > +++ b/fs/xfs/xfs_buf.c
> > @@ -843,9 +843,6 @@ xfs_buf_readahead_map(
> >  {
> >  	struct xfs_buf		*bp;
> >  
> > -	if (bdi_read_congested(target->bt_bdev->bd_disk->bdi))
> > -		return;
> 
> Ok, but this isn't a "throttle writeback" test here - it's trying to
> avoid having speculative readahead blocking on a full request queue
> instead of just skipping the readahead IO. i.e. prevent readahead
> thrashing and/or adding unnecessary read load when we already have a
> full read queue...
> 
> So what is the replacement for that? We want to skip the entire
> buffer lookup/setup/read overhead if we're likely to block on IO
> submission - is there anything we can use to do this these days?

I don't think there is a concept of a "full read queue" any more.
There are things that can block an IO submission though.
There is allocation of the bio from a mempool, and there is
rq_qos_throttle, and there are probably other places where submission
can block.  I don't think you can tell in advance if a submission is
likely to block.

I think the idea is that the top level of the submission stack should
rate-limit based on the estimated throughput of the stack.  I think
write-back does this.  I don't know about read-ahead.

NeilBrown