Re: d_splice_alias() problem.

[Greg Banks <gnb@melbourne.sgi.com>]

G'day Dipankar,

I don't know if you've been following this thread on lkml, but Al Viro
wants an RCU expert's opinion on the following dcache patch.

Greg Banks wrote:
> 
> > > *   Dentry_stat.nr_unused can be be spuriously decremented when dput()
> > >     races with __dget_unlocked().  Eventual result is nr_unused<0
> > >     and kswapd loops.  This is the problem I mentioned earlier.  Note
> > >     that this is not an NFS-specific problem.  Fix is:
> > > [...first attempt elided...]
> Ok, how about this...it's portable, and not racy, but may perturb the
> logic slightly by also taking dentries off the unused list in the case
> where they already had d_count>=1.  I'm not sure how significant that is.
> In any case this also passes my tests.
> 
> --- linux.orig/fs/dcache.c      Mon May  3 21:46:30 2004
> +++ linux/fs/dcache.c   Tue May  4 14:34:44 2004
> @@ -256,7 +256,7 @@
>  static inline struct dentry * __dget_locked(struct dentry *dentry)
>  {
>         atomic_inc(&dentry->d_count);
> -       if (atomic_read(&dentry->d_count) == 1) {
> +       if (!list_empty(&dentry->d_lru)) {
>                 dentry_stat.nr_unused--;
>                 list_del_init(&dentry->d_lru);
>         }
> @@ -663,6 +663,7 @@
>                 if (gfp_mask & __GFP_FS)
>                         prune_dcache(nr);
>         }
> +       BUG_ON(dentry_stat.nr_unused < 0);
>         return dentry_stat.nr_unused;
>  }
> 

viro@parcelfarce.linux.theplanet.co.uk wrote:
> 
> a) ask RCU folks to review - the current logics in dcache.c is extremely
> brittle as it is.

Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.