On Tue 22-10-19 09:38:19, Dave Chinner wrote:
> On Mon, Oct 21, 2019 at 09:43:30PM +0200, Jan Kara wrote:
> > On Mon 21-10-19 09:31:12, Theodore Y. Ts'o wrote:
> > > Hi Matthew, thanks for your work on this patch series!
> > >
> > > I applied it against 4c3, and ran a quick test run on it, and found
> > > the following locking problem. To reproduce:
> > >
> > > kvm-xfstests -c nojournal generic/113
> > >
> > > generic/113 [09:27:19][ 5.841937] run fstests generic/113 at 2019-10-21 09:27:19
> > > [ 7.959477]
> > > [ 7.959798] ============================================
> > > [ 7.960518] WARNING: possible recursive locking detected
> > > [ 7.961225] 5.4.0-rc3-xfstests-00012-g7fe6ea084e48 #1238 Not tainted
> > > [ 7.961991] --------------------------------------------
> > > [ 7.962569] aio-stress/1516 is trying to acquire lock:
> > > [ 7.963129] ffff9fd4791148c8 (&sb->s_type->i_mutex_key#12){++++}, at: __generic_file_fsync+0x3e/0xb0
> > > [ 7.964109]
> > > [ 7.964109] but task is already holding lock:
> > > [ 7.964740] ffff9fd4791148c8 (&sb->s_type->i_mutex_key#12){++++}, at: ext4_dio_write_iter+0x15b/0x430
> >
> > This is going to be a tricky one. With iomap, the inode locking is handled
> > by the filesystem while calling generic_write_sync() is done by
> > iomap_dio_rw(). I would really prefer to avoid tweaking iomap_dio_rw() not
> > to call generic_write_sync().
>
> You can't remove it from there, because that will break O_DSYNC
> AIO+DIO. i.e. generic_write_sync() needs to be called before
> iocb->ki_complete() is called in the AIO completion path, and that
> means filesystems using iomap_dio_rw need to be be able to run
> generic_write_sync() without taking the inode_lock().
>
> > So we need to remove inode_lock from
> > __generic_file_fsync() (used from ext4_sync_file()). This locking is mostly
> > for legacy purposes and we don't need this in ext4 AFAICT - but removing
> > the lock from __generic_file_fsync() would mean auditing all legacy
> > filesystems that use this to make sure flushing inode & its metadata buffer
> > list while it is possibly changing cannot result in something unexpected. I
> > don't want to clutter this series with it so we are left with
> > reimplementing __generic_file_fsync() inside ext4 without inode_lock. Not
> > too bad but not great either. Thoughts?
>
> XFS has implemented it's own ->fsync operation pretty much forever
> without issues. It's basically:
>
> 1. flush dirty cached data w/ WB_SYNC_ALL
> 2. flush dirty cached metadata (i.e. journal force)
> 3. flush device caches if journal force didn't, keeping in
> mind the requirements of data and journal being placed on
> different devices.
>
> The ext4 variant shouldn't need to be any more complex than that...
Yeah, that's what we do for the common case as well. But when the
filesystem is created without a journal (i.e., ext2 compatibility mode) we
currently use the old fsync implementation including
__generic_file_fsync(). But as I wrote above, duplicating those ~5 lines
out of __generic_file_fsync() we really care about is not a big deal.
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
