Re: [RFC PATCH 1/5] mm: intorduce __GFP_UNMAPPED and unmapped_alloc()

[Mike Rapoport]

On Fri, May 19, 2023 at 11:47:42AM -0400, Kent Overstreet wrote:
> On Fri, May 19, 2023 at 11:29:45AM +0300, Mike Rapoport wrote:
> > Your allocator implicitly relies on vmalloc because of module_alloc ;-)
> > 
> > What I was thinking is that we can replace module_alloc() calls in your
> > allocator with something based on my unmapped_alloc(). If we make the part
> > that refills the cache also take care of creating the mapping in the
> > module address space, that should cover everything.
> 
> Yeah, that's exactly what I was thinking :)
> 
> Liam was also just mentioning on IRC vmalloc lock contention came up
> again at LSF, and that's historically always been an isuse - going with
> your patchset for the backend nicely avoids that.

Unfortunately not because we still need to map the pages in the modules
area which is essentially a subset of vmalloc address space.
 
> If I have time (hah! big if :) I'll see if I can cook up a patchset that
> combines our two approaches over the weekend.

Now there is also an interest about unmapped allocations from KVM folks, so
I might continue pursuing unmapped allocator, probably just without a new
GFP flag and hooks into page allocator.

-- 
Sincerely yours,
Mike.