Date: Wed, 23 Mar 2016 09:04:31 +1100
From: Dave Chinner <david@fromorbit.com>
To: Jens Axboe <axboe@fb.com>
Cc: linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        linux-block@vger.kernel.org
Subject: Re: [PATCH 2/6] writeback: wb_start_writeback() should use
 WB_SYNC_ALL for WB_REASON_SYNC
Message-ID: <20160322220431.GT11812@dastard>
References: <1458669320-6819-1-git-send-email-axboe@fb.com>
 <1458669320-6819-3-git-send-email-axboe@fb.com>
 <20160322213427.GR11812@dastard>
 <56F1BBCC.8050004@fb.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <56F1BBCC.8050004@fb.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2276
Lines: 56

On Tue, Mar 22, 2016 at 03:40:28PM -0600, Jens Axboe wrote:
> On 03/22/2016 03:34 PM, Dave Chinner wrote:
> >On Tue, Mar 22, 2016 at 11:55:16AM -0600, Jens Axboe wrote:
> >>If you call sync, the initial call to wakeup_flusher_threads() ends up
> >>calling wb_start_writeback() with reason=WB_REASON_SYNC, but
> >>wb_start_writeback() always uses WB_SYNC_NONE as the writeback mode.
> >>Ensure that we use WB_SYNC_ALL for a sync operation.
> >
> >This seems wrong to me. We want background write to happen as
> >quickly as possible and /not block/ when we first kick sync.
> 
> It's not going to block. wakeup_flusher_threads() async queues
> writeback work through wb_start_writeback().

The flusher threads block, not the initial wakeup. e.g. they will
now block waiting for data writeback to complete before writing the
inode. i.e. this code in __writeback_single_inode() is now triggered
by the background flusher:

        /*
         * Make sure to wait on the data before writing out the metadata.
         * This is important for filesystems that modify metadata on data
         * I/O completion. We don't do it for sync(2) writeback because it has a
         * separate, external IO completion path and ->sync_fs for guaranteeing
         * inode metadata is written back correctly.
         */
        if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
                int err = filemap_fdatawait(mapping);
                if (ret == 0)
                        ret = err;
        }

It also changes the writeback chunk size in write_cache_pages(), so
instead of doing a bit of writeback from all dirty inodes, it tries
to write everything from each inode in turn (see
writeback_chunk_size()) which will further exacerbate the wait
above.

> >The latter blocking passes of sync use WB_SYNC_ALL to ensure that we
> >block waiting for all remaining IO to be issued and waited on, but
> >the background writeback doesn't need to do this.
> 
> That's fine, they can get to wait on the previously issued IO, which
> was properly submitted with WB_SYNC_ALL.
> 
> Maybe I'm missing your point?

Making the background flusher block and wait for data makes it
completely ineffective in speeding up sync() processing...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com