Re: Patch(?): linux-2.4.3-pre6/mm/vmalloc.c could return with

[Linus Torvalds]

On Fri, 23 Mar 2001, Marcelo Tosatti wrote:
>
> There is no need to hold mm->page_table_lock for vmalloced memory.

But there is. You do need _some_ protection to protect the kernel from
inserting two different pmd/pgd entries for two different areas in the
same slot. And that's exactly what page_table_lock does for us.

> I guess a better solution is to make the vmalloc codepath use
> "pte_alloc_vmalloc" (or something like that) which would be a
> spinlock-free version of pte_alloc (like the old one).

The old one avoided the race by using the big kernel lock. Which is
totally non-sensical, but works. It's much better to use the spinlock that
is meant for exactly this thing.

		Linus