Messages in this thread | | | Date | Sun, 17 Oct 2010 04:29:24 +1100 | From | Nick Piggin <> | Subject | Re: [PATCH 04/19] fs: Implement lazy LRU updates for inodes. |
| |
On Sat, Oct 16, 2010 at 12:59:30PM -0400, Christoph Hellwig wrote: > On Sat, Oct 16, 2010 at 08:29:16PM +1100, Nick Piggin wrote: > > On Sat, Oct 16, 2010 at 07:13:58PM +1100, Dave Chinner wrote: > > > @@ -502,11 +527,15 @@ static void prune_icache(int nr_to_scan) > > > iput(inode); > > > spin_lock(&inode_lock); > > > > > > - if (inode != list_entry(inode_unused.next, > > > - struct inode, i_list)) > > > - continue; /* wrong inode or list_empty */ > > > - if (!can_unuse(inode)) > > > + /* > > > + * if we can't reclaim this inode immediately, give it > > > + * another pass through the free list so we don't spin > > > + * on it. > > > + */ > > > + if (!can_unuse(inode)) { > > > + list_move(&inode->i_list, &inode_unused); > > > continue; > > > + } > > > } > > > list_move(&inode->i_list, &freeable); > > > WARN_ON(inode->i_state & I_NEW); > > > > This is a bug, actually 2 bugs, which is why I omitted it in the version > > you picked up. I agree we want the optimisation though, so I've added it > > back in my tree. > > > > After you iput() and then re take the inode lock, you can't reference > > the inode because you don't know what happened to it. You need to keep > > that pointer check to verify it is still there. > > I don't think the pointer check will work either. By the time we retake > the lru lock the inode might already have been reaped through a call > to invalidate_inodes. There's no way we can do anything with it after
I don't think you're right. If we re take inode_lock, ensure it is on the LRU, and call the can_unuse checks, there is no more problem than the regular loop taking items from the LRU, AFAIKS.
> iput. What we could do is using variant of can_unuse to decide to move > the inode to the front of the lru before doing the iput. That way > we'll get to it next after retaking the lru lock if it's still there.
This might actually be the better approach anyway (even for upstream) -- it means we don't have to worry about the "check head element" heuristic of the LRU check which could get false negatives if there is a lot of concurrency on the LRU.
| |