On Tue, 2011-03-15 at 23:42 +0100, Eric Dumazet wrote: > Le mardi 15 mars 2011 Ã 20:48 +0100, Peter Zijlstra a Ãcrit : > > On Tue, 2011-03-15 at 20:22 +0100, Thomas Gleixner wrote: > > > I am not sure if its a good idea to walk the tree > > > > as and when the tree is changing either because of a insertion or > > > > deletion of a probe. > > > > > > I know that you cannot walk the tree lockless except you would use > > > some rcu based container for your probes. > > > > You can in fact combine a seqlock, rb-trees and RCU to do lockless > > walks. > > > > https://lkml.org/lkml/2010/10/20/160 > > > > and > > > > https://lkml.org/lkml/2010/10/20/437 > > > > But doing that would be an optimization best done once we get all this > > working nicely. > > > > We have such schem in net/ipv4/inetpeer.c function inet_getpeer() (using > a seqlock on latest net-next-2.6 tree), but we added a counter to make > sure a reader could not enter an infinite loop while traversing tree Right, Linus suggested a single lockless iteration, but a limited count works too. > (AVL tree in inetpeer case). Ooh, there's an AVL implementation in the kernel? I have to ask, why not use the RB-tree? (I know AVL has a slightly stronger balancing condition which reduces the max depth from 2*log(n) to 1+log(n)). Also, if it does make sense to have both and AVL and RB implementation, does it then also make sense to lift the AVL thing to generic code into lib/ ? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>