Re: kvmalloc: always use vmalloc if CONFIG_DEBUG_VM

[Mike Snitzer <snitzer@redhat.com>]

On Tue, May 01 2018 at  8:36pm -0400,
Andrew Morton <akpm@linux-foundation.org> wrote:

> On Tue, 24 Apr 2018 12:33:01 -0400 (EDT) Mikulas Patocka <mpatocka@redhat.com> wrote:
> 
> > 
> > 
> > On Tue, 24 Apr 2018, Michal Hocko wrote:
> > 
> > > On Tue 24-04-18 11:30:40, Mikulas Patocka wrote:
> > > > 
> > > > 
> > > > On Tue, 24 Apr 2018, Michal Hocko wrote:
> > > > 
> > > > > On Mon 23-04-18 20:25:15, Mikulas Patocka wrote:
> > > > > 
> > > > > > Fixing __vmalloc code 
> > > > > > is easy and it doesn't require cooperation with maintainers.
> > > > > 
> > > > > But it is a hack against the intention of the scope api.
> > > > 
> > > > It is not!
> > > 
> > > This discussion simply doesn't make much sense it seems. The scope API
> > > is to document the scope of the reclaim recursion critical section. That
> > > certainly is not a utility function like vmalloc.
> > 
> > That 15-line __vmalloc bugfix doesn't prevent you (or any other kernel 
> > developer) from converting the code to the scope API. You make nonsensical 
> > excuses.
> > 
> 
> Fun thread!
> 
> Winding back to the original problem, I'd state it as
> 
> - Caller uses kvmalloc() but passes the address into vmalloc-naive
>   DMA API and
> 
> - Caller uses kvmalloc() but passes the address into kfree()
> 
> Yes?

I think so.

> If so, then...
> 
> Is there a way in which, in the kvmalloc-called-kmalloc path, we can
> tag the slab-allocated memory with a "this memory was allocated with
> kvmalloc()" flag?  I *think* there's extra per-object storage available
> with suitable slab/slub debugging options?  Perhaps we could steal one
> bit from the redzone, dunno.
> 
> If so then we can
> 
> a) set that flag in kvmalloc() if the kmalloc() call succeeded
> 
> b) check for that flag in the DMA code, WARN if it is set.
> 
> c) in kvfree(), clear that flag before calling kfree()
> 
> d) in kfree(), check for that flag and go WARN() if set.
> 
> So both potential bugs are detected all the time, dependent upon
> CONFIG_SLUB_DEBUG (and perhaps other slub config options).

Thanks Andrew, definitely the most sane proposal I've seen to resolve
this.