Re: [PATCH 7/5] Optimise d_find_alias()

[Ingo Molnar <mingo@elte.hu>]

* David Howells <dhowells@redhat.com> wrote:

> Andrew Morton <akpm@osdl.org> wrote:
> 
> > How can we get away without a barrier?
> 
> Is this what you're thinking of:
> 
> 	struct dentry * d_find_alias(struct inode *inode)
> 	{
> 		struct dentry *de = NULL;
> 
> 		smp_rmb();
> 		if (!list_empty(&inode->i_dentry)) {
> 			spin_lock(&dcache_lock);

no barrier is needed i think. This code just picks an alias, with no 
guarantee of any ordering wrt. add/removal ops from the alias list. In 
other words, d_find_alias() depends on external serialization - when 
used standalone it is fundamentally nondeterministic anyway.

it is true that the list_empty() might observe an intermediate state of 
the list [e.g. backlink not modified yet but forward link modified - or 
the other way around], but that's not a problem: it either returns true 
or false, both of which are correct results when the modification is 
underway in parallel on another CPU. When d_find_alias() returns 
non-NULL it will only do it by taking dcache_lock, so the correct use of 
the list is always guaranteed.

[ a barrier would have no useful effect here anyway - it only guarantees
  instruction ordering locally, and has no deterministic effect on the
  order of memory values being written from another CPU. A barrier makes
  sense when ordering of writes to two different memory values is
  deterministic - e.g. a flag and a resource pointer. But in the case of
  list_empty(), the order of modifications to the two pointers in the
  list is not guaranteed. ]

	Ingo