[Nouveau] [PATCH] nouveau/gem: fix user-after-free in nouveau_gem_new

[Jeremy Cline]

On Mon, May 17, 2021 at 11:19:02AM +0200, Thierry Reding wrote:
> On Mon, May 17, 2021 at 10:56:29AM +0200, Thierry Reding wrote:
> > On Tue, May 11, 2021 at 06:35:53PM +0200, Karol Herbst wrote:
> > > If ttm_bo_init fails it will already call ttm_bo_put, so we don't have to
> > > do it through nouveau_bo_ref.
> > > 
> > > ==================================================================
> > > BUG: KFENCE: use-after-free write in ttm_bo_put+0x11/0x40 [ttm]
> > > 
> > > Use-after-free write at 0x000000004dc4663c (in kfence-#44):
> > >  ttm_bo_put+0x11/0x40 [ttm]
> > >  nouveau_gem_new+0xc1/0xf0 [nouveau]
> > >  nouveau_gem_ioctl_new+0x53/0xf0 [nouveau]
> > >  drm_ioctl_kernel+0xb2/0x100 [drm]
> > >  drm_ioctl+0x215/0x390 [drm]
> > >  nouveau_drm_ioctl+0x55/0xa0 [nouveau]
> > >  __x64_sys_ioctl+0x83/0xb0
> > >  do_syscall_64+0x33/0x40
> > >  entry_SYSCALL_64_after_hwframe+0x44/0xae
> > > 
> > > kfence-#44 [0x00000000c0593b31-0x000000002e74122b, size=792, cache=kmalloc-1k] allocated by task 2657:
> > >  nouveau_bo_alloc+0x63/0x4c0 [nouveau]
> > >  nouveau_gem_new+0x38/0xf0 [nouveau]
> > >  nouveau_gem_ioctl_new+0x53/0xf0 [nouveau]
> > >  drm_ioctl_kernel+0xb2/0x100 [drm]
> > >  drm_ioctl+0x215/0x390 [drm]
> > >  nouveau_drm_ioctl+0x55/0xa0 [nouveau]
> > >  __x64_sys_ioctl+0x83/0xb0
> > >  do_syscall_64+0x33/0x40
> > >  entry_SYSCALL_64_after_hwframe+0x44/0xae
> > > 
> > > freed by task 2657:
> > >  ttm_bo_release+0x1cc/0x300 [ttm]
> > >  ttm_bo_init_reserved+0x2ec/0x300 [ttm]
> > >  ttm_bo_init+0x5e/0xd0 [ttm]
> > >  nouveau_bo_init+0xaf/0xc0 [nouveau]
> > >  nouveau_gem_new+0x7f/0xf0 [nouveau]
> > >  nouveau_gem_ioctl_new+0x53/0xf0 [nouveau]
> > >  drm_ioctl_kernel+0xb2/0x100 [drm]
> > >  drm_ioctl+0x215/0x390 [drm]
> > >  nouveau_drm_ioctl+0x55/0xa0 [nouveau]
> > >  __x64_sys_ioctl+0x83/0xb0
> > >  do_syscall_64+0x33/0x40
> > >  entry_SYSCALL_64_after_hwframe+0x44/0xae
> > > 
> > > Fixes: 019cbd4a4feb3 "drm/nouveau: Initialize GEM object before TTM object"
> > > Cc: Thierry Reding <treding at nvidia.com>
> > > Signed-off-by: Karol Herbst <kherbst at redhat.com>
> > > ---
> > >  drivers/gpu/drm/nouveau/nouveau_gem.c | 1 -
> > >  1 file changed, 1 deletion(-)
> > > 
> > > diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
> > > index c88cbb85f101..1165ff990fb5 100644
> > > --- a/drivers/gpu/drm/nouveau/nouveau_gem.c
> > > +++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
> > > @@ -212,7 +212,6 @@ nouveau_gem_new(struct nouveau_cli *cli, u64 size, int align, uint32_t domain,
> > >  
> > >  	ret = nouveau_bo_init(nvbo, size, align, domain, NULL, NULL);
> > >  	if (ret) {
> > > -		nouveau_bo_ref(NULL, &nvbo);
> > >  		return ret;
> > >  	}
> > 
> > Looking at the surrounding code, I wonder if I just managed to jumble
> > the cleanup paths for drm_gem_object_init() and nouveau_bo_init(). If
> > drm_gem_object_init() fails, I don't think it's necessary (though it
> > also doesn't look harmful) to call drm_gem_object_release().
> > 
> > However, if nouveau_bo_init() fails, then I think we'd still need to
> > call drm_gem_object_release(), to make sure to undo the effects of
> > drm_gem_object_init().
> > 
> > So I wonder if we need something like this instead:
> > 
> > --- >8 ---
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
> > index c88cbb85f101..9b6055116f30 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_gem.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
> > @@ -205,14 +205,13 @@ nouveau_gem_new(struct nouveau_cli *cli, u64 size, int align, uint32_t domain,
> >  	 * to the caller, instead of a normal nouveau_bo ttm reference. */
> >  	ret = drm_gem_object_init(drm->dev, &nvbo->bo.base, size);
> >  	if (ret) {
> > -		drm_gem_object_release(&nvbo->bo.base);
> >  		kfree(nvbo);
> >  		return ret;
> >  	}
> >  
> >  	ret = nouveau_bo_init(nvbo, size, align, domain, NULL, NULL);
> >  	if (ret) {
> > -		nouveau_bo_ref(NULL, &nvbo);
> > +		drm_gem_object_release(&nvbo->bo.base);
> >  		return ret;
> >  	}
> >  
> > --- >8 ---
> > 
> > Thierry
> 
> Adding Jeremy for visibility.
> 

Admittedly I only skimmed the code so I'm not extremely confident in my
analysis, but isn't that handled by the nouveau_bo_del_ttm() callback
which should get called after the last reference is dropped with
nouveau_bo_ref?

- Jeremy