Re: [PATCH v5 2/2] buffer: record blockdev write errors in super_block that it backs

[Jeff Layton]

On Thu, 2020-04-16 at 11:35 +0200, Jan Kara wrote:
> On Wed 15-04-20 12:22:27, Jeff Layton wrote:
> > On Wed, 2020-04-15 at 16:06 +0200, Jan Kara wrote:
> > > On Wed 15-04-20 08:13:00, Jeff Layton wrote:
> > > > From: Jeff Layton <jlayton@xxxxxxxxxx>
> > > > 
> > > > When syncing out a block device (a'la __sync_blockdev), any error
> > > > encountered will only be recorded in the bd_inode's mapping. When the
> > > > blockdev contains a filesystem however, we'd like to also record the
> > > > error in the super_block that's stored there.
> > > > 
> > > > Make mark_buffer_write_io_error also record the error in the
> > > > corresponding super_block when a writeback error occurs and the block
> > > > device contains a mounted superblock.
> > > > 
> > > > Since superblocks are RCU freed, hold the rcu_read_lock to ensure
> > > > that the superblock doesn't go away while we're marking it.
> > > > 
> > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > > ---
> > > >  fs/buffer.c | 7 +++++++
> > > >  1 file changed, 7 insertions(+)
> > > > 
> > > > diff --git a/fs/buffer.c b/fs/buffer.c
> > > > index f73276d746bb..2a4a5cc20418 100644
> > > > --- a/fs/buffer.c
> > > > +++ b/fs/buffer.c
> > > > @@ -1154,12 +1154,19 @@ EXPORT_SYMBOL(mark_buffer_dirty);
> > > >  
> > > >  void mark_buffer_write_io_error(struct buffer_head *bh)
> > > >  {
> > > > +	struct super_block *sb;
> > > > +
> > > >  	set_buffer_write_io_error(bh);
> > > >  	/* FIXME: do we need to set this in both places? */
> > > >  	if (bh->b_page && bh->b_page->mapping)
> > > >  		mapping_set_error(bh->b_page->mapping, -EIO);
> > > >  	if (bh->b_assoc_map)
> > > >  		mapping_set_error(bh->b_assoc_map, -EIO);
> > > > +	rcu_read_lock();
> > > > +	sb = bh->b_bdev->bd_super;
> > > 
> > > You still need READ_ONCE() here. Otherwise the dereference below can still
> > > result in refetch and NULL ptr deref.
> > > 
> > > 								Honza
> > > 
> > 
> > Huh? That seems like a really suspicious thing for the compiler/arch to
> > do. We are checking that sb isn't NULL before we dereference it. Doesn't
> > that imply a data dependency? How could the value of "sb" change after
> > that?
> 
> Because the compiler is free to optimize the local variable away and
> actually compile the dereference below as bh->b_bdev->bd_super->s_wb_err
> (from C11 standard POV such code is equivalent since in C11 memory model
> it is assumed there are no concurrent accesses). And READ_ONCE() is a way
> to forbid compiler from doing such optimization - through 'volatile'
> keyword it tells the compiler there may be concurrent accesses happening
> and makes sure the value is really fetched into the local variable and used
> from there. There are good articles about this on LWN - I'd give you a link
> but LWN seems to be down today. But the latest article is about KCSAN and
> from there are links to older articles about compiler optimizations.
> 
> > I'm also not sure I understand how using READ_ONCE really helps there if
> > we can't count on the value of a local variable not changing.
> 
> I hope I've explained this above.
> 

Got it. Thanks for the explanation. Now I'll have nightmares about all
of the race conditions I've created in the past by making this
assumption!

I'll send a v6 set in a few mins.

> > > > +	if (sb)
> > > > +		errseq_set(&sb->s_wb_err, -EIO);
> > > > +	rcu_read_unlock();
> > > >  }
> > > >  EXPORT_SYMBOL(mark_buffer_write_io_error);
> > > >  
> > > > -- 
> > > > 2.25.2
> > > > 
> > 
> > -- 
> > Jeff Layton <jlayton@xxxxxxxxxx>
> > 

-- 
Jeff Layton <jlayton@xxxxxxxxxx>