On 22/07/14 04:36PM, Jan Kara wrote:
> On Thu 14-07-22 17:17:02, Ritesh Harjani wrote:
> > On 22/07/12 12:54PM, Jan Kara wrote:
> > > Do not reclaim entries that are currently used by somebody from a
> > > shrinker. Firstly, these entries are likely useful. Secondly, we will
> > > need to keep such entries to protect pending increment of xattr block
> > > refcount.
> > >
> > > CC: stable@xxxxxxxxxxxxxxx
> > > Fixes: 82939d7999df ("ext4: convert to mbcache2")
> > > Signed-off-by: Jan Kara <jack@xxxxxxx>
> > > ---
> > > fs/mbcache.c | 10 +++++++++-
> > > 1 file changed, 9 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/mbcache.c b/fs/mbcache.c
> > > index 97c54d3a2227..cfc28129fb6f 100644
> > > --- a/fs/mbcache.c
> > > +++ b/fs/mbcache.c
> > > @@ -288,7 +288,7 @@ static unsigned long mb_cache_shrink(struct mb_cache *cache,
> > > while (nr_to_scan-- && !list_empty(&cache->c_list)) {
> > > entry = list_first_entry(&cache->c_list,
> > > struct mb_cache_entry, e_list);
> > > - if (entry->e_referenced) {
> > > + if (entry->e_referenced || atomic_read(&entry->e_refcnt) > 2) {
> > > entry->e_referenced = 0;
> > > list_move_tail(&entry->e_list, &cache->c_list);
> > > continue;
> > > @@ -302,6 +302,14 @@ static unsigned long mb_cache_shrink(struct mb_cache *cache,
> > > spin_unlock(&cache->c_list_lock);
> > > head = mb_cache_entry_head(cache, entry->e_key);
> > > hlist_bl_lock(head);
> > > + /* Now a reliable check if the entry didn't get used... */
> > > + if (atomic_read(&entry->e_refcnt) > 2) {
> >
> > On taking a look at this patchset again. I think if we move this "if" condition
> > of checking refcnt to above i.e. before we delete the entry from c_list.
> > Then we can avoid =>
> > removing of the entry -> checking it's refcnt under lock -> adding it back
> > if the refcnt is elevated.
> >
> > Thoughts?
>
> Well, but synchronization would get more complicated because we don't want
> to acquire hlist_bl_lock() under c_list_lock (technically we could at this
Ok, yes. I tried implementing it and it becomes lock()/unlock() mess.
> point in the series but it would make life harder for the last patch in the
> series). And we need c_list_lock to remove entry from the LRU list. It
> could be all done but I don't think what you suggest is really that simpler
> and this code will go away later in the patchset anyway...
I agree. Thanks for re-checking it.
-ritesh
