The attached patch contains a new implementation of liballoc. It is a
bit more streamlined, uses less memory for overhead, and is faster.
Blocks are now strung together via a small header (instead of a separate
linked lists of structs that point at blocks). This implementation of
liballoc has been in production on SGI systems for over a year now
without any difficulties.
I plan on checking this patch in by the end of this week. Please provide
feedback regarding any problems you might see.