Manuel points out that we could be exhausting kernel virtual address space. That's plausible on i386 and macppc, and not quite as much on amd64. So I would be tempted to understand the memory allocators and make failure to allocate kernel virtual space more obvious. Someone at BBN is looking into that (we have a lockup that seems similar, with some private code that allocates a ridiculous amount of kernel memory). It's also possible that there is one issue on amd64 and one issue on i386/macppc, or something else. Have you set kern.maxvnodes different from default? Is anyone else seeing this problem?
Description: PGP signature