Re: [PATCH v5] add barriers to buffer functions

[Matthew Wilcox]

On Mon, Aug 08, 2022 at 10:57:45AM -0400, Mikulas Patocka wrote:
> On Mon, 8 Aug 2022, Matthew Wilcox wrote:
> 
> > On Mon, Aug 08, 2022 at 10:26:10AM -0400, Mikulas Patocka wrote:
> > > On Sun, 7 Aug 2022, Matthew Wilcox wrote:
> > > > > +static __always_inline void set_buffer_locked(struct buffer_head *bh)
> > > > > +{
> > > > > +	set_bit(BH_Lock, &bh->b_state);
> > > > > +}
> > > > > +
> > > > > +static __always_inline int buffer_locked(const struct buffer_head *bh)
> > > > > +{
> > > > > +	bool ret = test_bit(BH_Lock, &bh->b_state);
> > > > > +	/*
> > > > > +	 * pairs with smp_mb__after_atomic in unlock_buffer
> > > > > +	 */
> > > > > +	if (!ret)
> > > > > +		smp_acquire__after_ctrl_dep();
> > > > > +	return ret;
> > > > > +}
> > > > 
> > > > Are there places that think that lock/unlock buffer implies a memory
> > > > barrier?
> > > 
> > > There's this in fs/reiserfs:
> > > 
> > > if (!buffer_dirty(bh) && !buffer_locked(bh)) {
> > > 	reiserfs_free_jh(bh); <--- this could be moved before buffer_locked
> > 
> > It might be better to think of buffer_locked() as
> > buffer_someone_has_exclusive_access().  I can't see the problem with
> > moving the reads in reiserfs_free_jh() before the read of buffer_locked.
> > 
> > > if (buffer_locked((journal->j_header_bh))) {
> > > 	...
> > > }
> > > journal->j_last_flush_trans_id = trans_id;
> > > journal->j_first_unflushed_offset = offset;
> > > jh = (struct reiserfs_journal_header *)(journal->j_header_bh->b_data); <--- this could be moved before buffer_locked
> > 
> > I don't think b_data is going to be changed while someone else holds
> > the buffer locked.  That's initialised by set_bh_page(), which is an
> > initialisation-time thing, before the BH is visible to any other thread.
> 
> So, do you think that we don't need a barrier in buffer_locked()?

That's my feeling.  Of course, you might not be the only one confused,
and if fs authors in general have made the mistake of thinking that
buffer_locked is serialising, then it might be better to live up to
that expectation.

> There is also this (where the BUG_ON(!buffer_uptodate(bh)) saves it).
>                 if (buffer_locked(bh)) {
>                         int depth;
>                         PROC_INFO_INC(sb, scan_bitmap.wait);
>                         depth = reiserfs_write_unlock_nested(sb);
>                         __wait_on_buffer(bh);
>                         reiserfs_write_lock_nested(sb, depth);
>                 }
>                 BUG_ON(!buffer_uptodate(bh));
>                 BUG_ON(atomic_read(&bh->b_count) == 0);
> 
>                 if (info->free_count == UINT_MAX)
>                         reiserfs_cache_bitmap_metadata(sb, bh, info); <--- this could be moved before buffer_locked if there were no BUG_ONs

It could be moved before buffer_locked(), but I don't see the harm in
that.  Look at how reiserfs_read_bitmap_block() gets the bh:

        bh = sb_bread(sb, block);

__bread_gfp() has either already read the buffer (and it's uptodate),
in which case it returns it.  Or it calls __bread_slow() which will do
the read and check uptodate before returning it.  I wouldn't be surprised
to find that this buffer_locked() check is actually dead code, but I have
no desire to dive into reiserfs far enough to find out whether it's dead
code or not.