Re: mm/vmalloc.c:3689 vread_iter() error: we previously assumed 'vm' could be null (see line 3667)

[Baoquan He]

On 10/19/23 at 08:40am, Dan Carpenter wrote:
> On Thu, Oct 19, 2023 at 10:28:21AM +0800, Baoquan He wrote:
> > On 10/18/23 at 08:52am, Andrew Morton wrote:
> > > On Wed, 18 Oct 2023 23:15:31 +0800 Baoquan He <bhe@xxxxxxxxxx> wrote:
> > > 
> > > > From: Baoquan He <bhe@xxxxxxxxxx>
> > > > Date: Wed, 18 Oct 2023 22:50:14 +0800
> > > > Subject: [PATCH] mm/vmalloc: fix the unchecked dereference warning in vread_iter()
> > > > Content-type: text/plain
> > > > 
> > > > LKP reported smatch warning as below:
> > > > 
> > > > ===================
> > > > smatch warnings:
> > > > mm/vmalloc.c:3689 vread_iter() error: we previously assumed 'vm' could be null (see line 3667)
> > > > ......
> > > > 06c8994626d1b7  @3667 size = vm ? get_vm_area_size(vm) : va_size(va);
> > > > ......
> > > > 06c8994626d1b7  @3689 else if (!(vm->flags & VM_IOREMAP))
> > > >                                  ^^^^^^^^^
> > > > Unchecked dereference
> > > > =====================
> > > > 
> > > > So add checking on whether 'vm' is not null when dereferencing it in
> > > > vread_iter(). This mutes smatch complaint.
> > > > 
> > > > ...
> > > >
> > > > --- a/mm/vmalloc.c
> > > > +++ b/mm/vmalloc.c
> > > > @@ -3813,7 +3813,7 @@ long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
> > > >  
> > > >  		if (flags & VMAP_RAM)
> > > >  			copied = vmap_ram_vread_iter(iter, addr, n, flags);
> > > > -		else if (!(vm->flags & VM_IOREMAP))
> > > > +		else if (!(vm && (vm->flags & VM_IOREMAP)))
> > > >  			copied = aligned_vread_iter(iter, addr, n);
> > > >  		else /* IOREMAP area is treated as memory hole */
> > > >  			copied = zero_iter(iter, n);
> > > 
> > > So is this not a real runtime bug?  We're only doing this to suppress a
> > > smatch warning?
> > > 
> > > If so, can we please include a description of *why* this wasn't a bug? 
> > > What conditions ensure that vm!=NULL at this point?
> > 
> > I think this is not a real runtime bug. The only chance it can hapen is
> > when (flags == VMAP_BLOCK) is true. That has been warned and could never
> > happen. I updated patch log and paste v2 here. 
> > 
> >                 /*
> >                  * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
> >                  * be set together with VMAP_RAM.
> >                  */
> >                 WARN_ON(flags == VMAP_BLOCK);
> >  
> >                 if (!vm && !flags)
> >                         continue;
> > 
> > 
> 
> Thanks.  If you want you could just ignore the warning.  It's a one time
> warning so we won't send the mail again and if people have questions
> about it, they can just look it up on lore.
> 
> The truth is when I was reviewing this code the first time I got mixed
> up between flags and vm->flags so that's part of why I reported it.
> 
> Smatch ignores the WARN_ON().  Historically WARN_ON() has been useless
> for indicating whether something can happen or not.  These days,
> WARN_ON() is treated as a syzkaller bug so we prefer pr_warn() if
> something can actually happen.  We still see a lot of WARN_ON()s
> happening in real life so I'm not eager to make Smatch treat them like a
> BUG_ON().
> 
> Also, sadly, even if we changed the WARN_ON() to a BUG_ON() it still
> wouldn't silence the warning because Smatch is not quite clever enough
> to parse that.

Thanks for your sharing, Dan.

My understanding was that our current code alwasys have va->flags with
VMAP_RAM when it's set. So the case va->flags == VMAP_BLOCK won't
happen. I now understand your worry that it possibly happens. People
could change that in the futuer with buggy code or intented action while
not noticing that. I may not get your suggestion clearly, wonder if you
are suggesting this could be a realtime bug and need be stated in patch
log clearly.

Thanks
Baoquan