In message <20080501170819.bdcb9035.akpm@linux-foundation.org>, Andrew Morton writes:
quoted text
> On Thu, 1 May 2008 19:44:18 -0400
> Erez Zadok <ezk@cs.sunysb.edu> wrote:
[...]
quoted text
> i_size is much more important because glitches in there can result in
> incorrect data being returned from read() and things like that.  i_blocks
> is just a beancounting curiosity.
> 
> > 
> > 2. I've rewritten your suggested code a bit to reduce stack use.  Modulo
> >    having 32-bit spin_lock/unlock variants, do you see any problem with this
> >    code below?  My testing of it so far on 32/64-bit SMP/UMP have all
> >    passed.
> > 
> > void fsstack_copy_inode_size(struct inode *dst, struct inode *src)
> > {
> > #if BITS_PER_LONG == 32
> > 	blkcnt_t i_blocks;
> > 
> > 	spin_lock(&src->i_lock);
> > 	i_blocks = src->i_blocks;
> > 	spin_unlock(&src->i_lock);
> > 	spin_lock(&dst->i_lock);
> > 	dst->i_blocks = i_blocks;
> > 	spin_unlock(&dst->i_lock);
> > #else
> > 	dst->i_blocks = src->i_blocks;
> > #endif
> > 	i_size_write(dst, i_size_read(src));
> > }
> 
> That looks sane, as long as we don't care about i_size-vs-i_blocks
> coherency.

quoted text
> However I expect that approximately zero of the sites which modify i_blocks
> are taking i_lock to do so.

If i_blocks is indeed less important than i_size, then we can live with some
incoherency b/t i_size and i_blocks, for now.  Nevertheless, I propose
adding this to linux/fs.h:

static inline blkcnt_t i_blocks_read(const struct inode *inode)
{
#if BITS_PER_LONG == 32
	blkcnt_t i_blocks;
	spin_lock(&src->i_lock);
	i_blocks = src->i_blocks;
	spin_unlock(&src->i_lock);
	return i_blocks;
#else
	return src->i_blocks;
#endif
}

and a matching i_blocks_write function.  We can then gradually convert those
"unsafe" users of i_blocks to use the new i_blocks_read/write helpers.

The nice thing about these two helpers is fsstack_copy_inode_size becomes a
lot cleaner and more elegant:

void fsstack_copy_inode_size(struct inode *dst, struct inode *src)
{
	i_blocks_write(dst, i_blocks_read(src));
	i_size_write(dst, i_size_read(src));
}

And, if we ever wanted to ensure coherency b/t i_blocks and i_size, we'll
need to create helpers that merge the functionality of i_size_read/write and
i_blocks_read/write.

What do you think?

Erez.
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/