Re: [PATCH] Fix shrink_dcache_parent() against shrink_dcache_memory() race (3rd updated patch)]

[Balbir Singh <balbir@in.ibm.com>]

On Fri, Mar 10, 2006 at 01:31:53PM +0100, Jan Blunck wrote:
> On Fri, Mar 10, Neil Brown wrote:
> 
> > -static void prune_dcache(int count)
> > +static void prune_dcache(int count, struct super_block *sb)
> >  {
> >  	spin_lock(&dcache_lock);
> >  	for (; count ; count--) {
> > @@ -417,8 +425,10 @@ static void prune_dcache(int count)
> >   			spin_unlock(&dentry->d_lock);
> >  			continue;
> >  		}
> > -		/* If the dentry was recently referenced, don't free it. */
> > -		if (dentry->d_flags & DCACHE_REFERENCED) {
> > +		/* If the dentry was recently referenced, or is for
> > +		 * a unmounting filesystem, don't free it. */
> > +		if ((dentry->d_flags & DCACHE_REFERENCED) ||
> > +		    (dentry->d_sb != sb && dentry->d_sb->s_root == NULL)) {
> >  			dentry->d_flags &= ~DCACHE_REFERENCED;
> >   			list_add(&dentry->d_lru, &dentry_unused);
> >   			dentry_stat.nr_unused++;
> 
> You have to down_read the rw-semaphore sb->s_umount since sb->s_root is
> protected by it :(

<snip>

Please do not beat me up for suggesting this. I was wondering if it
makes sense to add a PF_SHRINKER flag and set it in shrink_slab().
Use my solution, instead of PF_MEMALLOC check against PF_SHRINKER.

I think PF_SHRINKER might be a good idea, it might help us detect
races between the shrinker and other subsystems too - not only dcache.

It might be well worth adding in. If there is sufficient interest I can
send create and send out a patch.

Balbir