Re: [Lse-tech] lockmeter results comparing 2.4.17, 2.5.3, and 2.5.5

[Andrew Morton]

"Martin J. Bligh" wrote:
> 
> ...
> looks a little distressing - the hold times on inode_lock by prune_icache
> look bad in terms of latency (contention is still low, but people are still
> waiting on it for a very long time). Is this a transient thing, or do people
> think this is going to be a problem?

inode_lock hold times are a problem for other reasons.  Leaving this
unfixed makes the preepmtible kernel rather pointless....  An ideal
fix would be to release inodes based on VM pressure against their backing
page.  But I don't think anyone's started looking at inode_lock yet.

The big one is lru_list_lock, of course.  I'll be releasing code in
the next couple of days which should take that off the map.  Testing
would be appreciated.

I have a concern about the lockmeter results.  Lockmeter appears
to be measuring lock frequency and hold times and contention.  But
is it measuring the cost of the cacheline transfers?   For example,
I expect that with delayed allocation and radix-tree pagecache, one
of the major remaining bottlenecks will be ownership of the superblock
semaphore's cacheline.   Is this measurable?  (Actually, we may
then be at the point where copy_from_user costs dominate).

-