Re: inetpeer with create==0

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 02 mars 2011 à 20:45 -0800, David Miller a écrit :
> Eric, I was profiling the non-routing-cache case and something that stuck
> out is the case of calling inet_getpeer() with create==0.
> 
> If an entry is not found, we have to redo the lookup under a spinlock
> to make certain that a concurrent writer rebalancing the tree does
> not "hide" an existing entry from us.
> 
> This makes the case of a create==0 lookup for a not-present entry
> really expensive.  It is on the order of 600 cpu cycles on my
> Niagara2.
> 

Well, your test assumes all data is already on cpu caches ?

I'll take a look, but my reasoning was that the real cost in DDOS
situation is to bring data into caches. With a 20 depth, cache misses
costs are the problem.

The second lookup was basically free.

> I added a hack to not do the relookup under the lock when create==0
> and it now costs less than 300 cycles.
> 

Hmm... I'm curious you send this hack to me ;)

> This is now a pretty common operation with the way we handle COW'd
> metrics, so I think it's definitely worth optimizing.
> 
> I looked at the generic radix tree implementation, and it supports
> full RCU lookups in parallel with insert/delete.  It handles the race
> case without the relookup under lock because it creates fixed paths
> to "slots" where nodes live using shifts and masks.  So if a path
> to a slot ever existed, it will always exist.
> 
> Take a look at lib/radix-tree.c and include/linux/radix-tree.h if
> you are curious.
> 
> I think we should do something similar for inetpeer.  Currently we
> cannot just use the existing generic radix-tree code because it only
> supports indexes as large as "unsigned long" and we need to handle
> 128-bit ipv6 addresses.