[PATCH] drm/nouveau/mmu: fix Use after Free bug in nvkm_vmm_node_split

[Takashi Iwai]

On Sat, 28 Jan 2023 03:17:15 +0100,
Danilo Krummrich wrote:
> 
> On Fri, Jan 27, 2023 at 01:10:46PM +0100, Takashi Iwai wrote:
> > On Tue, 03 Jan 2023 15:07:55 +0100,
> > Takashi Iwai wrote:
> > > 
> > > On Fri, 30 Dec 2022 08:27:58 +0100,
> > > Zheng Wang wrote:
> > > > 
> > > > Here is a function call chain.
> > > > nvkm_vmm_pfn_map->nvkm_vmm_pfn_split_merge->nvkm_vmm_node_split
> > > > If nvkm_vma_tail return NULL in nvkm_vmm_node_split, it will
> > > > finally invoke nvkm_vmm_node_merge->nvkm_vmm_node_delete, which
> > > > will free the vma. However, nvkm_vmm_pfn_map didn't notice that.
> > > > It goes into next label and UAF happens.
> > > > 
> > > > Fix it by returning the return-value of nvkm_vmm_node_merge
> > > > instead of NULL.
> > > > 
> > > > Signed-off-by: Zheng Wang <zyytlz.wz at 163.com>
> > > 
> > > FWIW, CVE-2023-0030 has been assigned to this bug.
> > > It's a question whether it really deserves as a security issue, but a
> > > bug is a bug...
> > > 
> > > Ben, could you review this please?
> > 
> > A gentle ping as reminder.  The bug is still present.
> 
> This was also reported in [1]. I had a closer look and FWICT this code is fine
> and there isn't a bug.
> 
> Zheng Wang, the reporter of the BZ, also confirmed this to be a false positive
> from CodeQL.
> 
> Anyway, here's the explaination I also posted in the BZ:
> 
> "In nvkm_vmm_node_merge() nvkm_vmm_node_delete() is only called when prev is
> set. However, prev is NULL unless we enter the "if (vma->addr != addr)" path in
> nvkm_vmm_node_split(). In such a case the vma pointer, which is also passed to
> nvkm_vmm_node_merge(), is set to a freshly allocated struct nvkm_vma with
> nvkm_vma_tail() right before prev is set to the old vma pointer.
> 
> Hence, the only thing happening there when nvkm_vma_tail() fails in the
> "if (vma->size != size)" path is that either nvkm_vmm_node_merge() does nothing
> in case prev wasn't set or it merges and frees the new vma created in the
> "if (vma->addr != addr)" path. Or in other words the proper cleanup for the
> error condition is done.
> 
> I can't see any case where the original vma pointer given by nvkm_vmm_pfn_map()
> is actually freed."
> 
> [1] https://bugzilla.redhat.com/show_bug.cgi?id=2157041

Thanks for the information!  Then we should try to dispute the CVE.
I'll ask our security team.


Takashi

> 
> - Danilo
> 
> > 
> > 
> > thanks,
> > 
> > Takashi
> > 
> > > 
> > > 
> > > thanks,
> > > 
> > > Takashi
> > > 
> > > > ---
> > > >  drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c | 4 ++--
> > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
> > > > index ae793f400ba1..84d6fc87b2e8 100644
> > > > --- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
> > > > +++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
> > > > @@ -937,8 +937,8 @@ nvkm_vmm_node_split(struct nvkm_vmm *vmm,
> > > >  	if (vma->size != size) {
> > > >  		struct nvkm_vma *tmp;
> > > >  		if (!(tmp = nvkm_vma_tail(vma, vma->size - size))) {
> > > > -			nvkm_vmm_node_merge(vmm, prev, vma, NULL, vma->size);
> > > > -			return NULL;
> > > > +			tmp = nvkm_vmm_node_merge(vmm, prev, vma, NULL, vma->size);
> > > > +			return tmp;
> > > >  		}
> > > >  		tmp->part = true;
> > > >  		nvkm_vmm_node_insert(vmm, tmp);
> > > > -- 
> > > > 2.25.1
> > > > 
> > 
>