Re: memory barrier question

[Jamie Lokier]

Miklos Szeredi wrote:
> On Thu, 16 Sep 2010, Paul E. McKenney wrote:
> > On Thu, Sep 16, 2010 at 03:30:56PM +0100, David Howells wrote:
> > > Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
> > > 
> > > > Is the rmb() really needed?
> > > > 
> > > > Take this code from fs/namei.c for example:
> > > > 
> > > > 		inode = next.dentry->d_inode;
> > > > 		if (!inode)
> > > > 			goto out_dput;
> > > > 
> > > > 		if (inode->i_op->follow_link) {
> > > > 
> > > > It happily dereferences dentry->d_inode without a barrier after
> > > > checking it for non-null, while that d_inode might have just been
> > > > initialized on another CPU with a freshly created inode.  There's
> > > > absolutely no synchornization with that on this side.
> > > 
> > > Perhaps it's not necessary; once set, how likely is i_op to be changed once
> > > I_NEW is cleared?
> > 
> > Are the path_get()s protecting this?
> 
> No, when creating a file the dentry will go from negative to positive
> independently from lookup.  The dentry can get instantiated with an
> inode between the path_get() and dereferencing ->d_inode.
> 
> > 
> > If there is no protection, then something like rcu_dereference() is
> > needed for the assignment from next.dentry->d_inode.
> 
> Do I understand correctly that the problem is that a CPU may have a
> stale cache associated with *inode, one that was loaded before the
> write barrier took effect?
> 
> Funny that such a bug could stay unnoticed in so often excercised
> code.  Yeah I know it's alpha only.

When I first saw read_barrier_depends(), I thought it must be Alpha's
speculative execution, fetching memory out of order and confirming
it's valid later.  I was really surprised to find out it's not that -
it's a quirk of the Alpha's cache/forwarding protocol.  Others
presumably don't have it because they were designed with awareness of
this coding pattern.

But...

I wonder if it can happen on IA64 with it's funky memory-alias
compiler optimisations.

I wonder if it can happen on x86 and others, if the compiler decides
this is a valid transformation (it is with a single CPU):

Original code:

    foo = global_ptr_to_foo;
    foo_x = foo->x;

    bar = global_ptr_to_bar;
    bar_y = bar->y;
    // use bar_y;

Transformed by compiler:

    foo = global_ptr_to_foo;
    foo_x = foo->x;

    bar = global_ptr_to_bar;
    bar_y = (__typeof__(bar->y))foo_x;
    if ((void *)bar != (void *)foo)
        bar_y = bar->y;

    // use bar_y;

In other words, without a barrier, the compiler doesn't have to order
the executed bar->y dereference *instruction* after the bar =
global_ptr_to_bar instruction.  Thus making it a compiler property,
not a CPU one.

There is no danger of dereferencing NULL in that example, but
dereferencing the values from the wrong object is just as wrong.

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