Re: [patch] mmap bug with drivers that adjust vm_start

[Benjamin LaHaise <bcrl@redhat.com>]

On Tue, Mar 26, 2002 at 05:42:36PM +0100, Andrea Arcangeli wrote:
> However if the patch is needed it means the ->mmap also must do the
> do_munmap stuff by hand internally, which is very ugly given we also did
> our own do_munmap in a completly different region (the one requested by
> the user).

At least my own code checks for that and fails if there is a mapping 
already placed at the fixed address it needs to use.  If we're paranoid, 
we could BUG() on getting a vma back from the new find_vma_prepare call.

> Our do_munmap should not happen if we place the mapping
> elsewhere. If possible I would prefer to change those drivers to
> advertise their enforced vm_start with a proper callback, the current
> way is halfway broken still. BTW, which are those drivers, and why they
> needs to enforce a certain vm_start (also despite MAP_FIXED that they
> cannot check within the ->mmap callback)?

Video drivers, others that require specific alignment (4MB pages for 
example).  Historically, the mmap call has been the hook for doing this, 
hence the comment in do_mmap from davem.  Unless there's a really good 
reason for changing the hook, I don't see doing so as providing much 
benefit other than making source compatibility hard.

		-ben
-- 
"A man with a bass just walked in,
 and he's putting it down
 on the floor."