Re: [RFC PATCH] btrfs: don't call btrfs_sync_file from iomap context

From: Goldwyn Rodrigues <rgoldwyn@suse.com>
To: Johannes Thumshirn <Johannes.Thumshirn@wdc.com>
Cc: "fdmanana@gmail.com" <fdmanana@gmail.com>,
	David Sterba <dsterba@suse.com>,
	"linux-btrfs @ vger . kernel . org" <linux-btrfs@vger.kernel.org>,
	Josef Bacik <josef@toxicpanda.com>
Subject: Re: [RFC PATCH] btrfs: don't call btrfs_sync_file from iomap context
Date: Tue, 1 Sep 2020 13:40:40 -0500	[thread overview]
Message-ID: <20200901184040.gxa3bh3nvrddi7ai@fiona> (raw)
In-Reply-To: <SN4PR0401MB3598F35FDA34EF5DE377C2CE9B2E0@SN4PR0401MB3598.namprd04.prod.outlook.com>

On 14:44 01/09, Johannes Thumshirn wrote:
> On 01/09/2020 16:37, Filipe Manana wrote:
> >> +               iocb->ki_flags |= IOCB_DSYNC;
> > 
> > This should be set before calling generic_write_sync() above,
> > otherwise we don't do the persistence required by dsync.
> > 
> > I have to be honest, I don't like this approach that much, it's a bit
> > fragile - what if some other code, invoked in between clearing and
> > setting back the flag, needs that flag to operate correctly?
> 
> Yes I don't like it either. 
> 
> I've compared btrfs to ext4 and xfs and both don't hold on the inode 
> lock for (nearly) the whole xxx_file_write_iter() time, so I think 
> the correct fix would be to relax this time. But I haven't found any 
> documentation yet on which functions need to be under the inode_lock 
> and which not.

I think you have got this reversed. ext4 and xfs hold the inode lock for
the entire xxx_file_write_iter(). OTOH, they don't take the inode lock
during fsync operations. The comments in btrfs_sync_file() mention that
it needs to make sure inode lock is acquired to make sure no other
process dirty the pages while sync is running.

> 
> Any input is appreciated here.
> 

-- 
Goldwyn