On 06/25/2012 07:15 PM, Christoph Hellwig wrote: > I've not been fully through this series, but two comments: > > - we absolutely need the split out struct node_id as a preparatory > patch before adding the new sockfd cache. Letting the casting mess > get in first and clean up later will be way more work. > This will also have various advantages like cleaning up the existing > qsort/bsearch callbacks. Point taken. I'll send the patch that add the struct node_id first. > - I've previously been non-exited about more liburcu use, but for > a cache like this which gets hit by all threads I don't think > using the slow sleeping pthread locks is a good idea. We need > a lockless reader algorithm for this cache as soon as possible. I have been thinking of it too, but I plan to add RCU in later after basic idea of sockfd cache is evaluated and merged. Current pthread-based implementation speed up the prototyping and easy to prove correctness. Thanks, Yuan |