Re: [PATCH v2 2/2] usercopy: Convert test_user_copy to KUnit test

[Kees Cook <kees@kernel.org>]

On Wed, Jun 12, 2024 at 09:21:52PM +0200, Geert Uytterhoeven wrote:
> Hi Kees,
> 
> On Wed, Jun 12, 2024 at 6:51 PM Kees Cook <kees@kernel.org> wrote:
> > On Wed, Jun 12, 2024 at 05:13:39PM +0800, David Gow wrote:
> > > On Tue, 11 Jun 2024 at 05:33, Kees Cook <kees@kernel.org> wrote:
> > > > Convert the runtime tests of hardened usercopy to standard KUnit tests.
> > > >
> > > > Co-developed-by: Vitor Massaru Iha <vitor@massaru.org>
> > > > Signed-off-by: Vitor Massaru Iha <vitor@massaru.org>
> > > > Link: https://lore.kernel.org/r/20200721174654.72132-1-vitor@massaru.org
> > > > Tested-by: Ivan Orlov <ivan.orlov0322@gmail.com>
> > > > Signed-off-by: Kees Cook <kees@kernel.org>
> > > > ---
> > >
> > > This looks good, particularly with the x86 fix applied.
> > >
> > > It's still hanging on m68k -- I think at the 'illegal reversed
> > > copy_to_user passed' test -- but I'll admit to not having tried to
> > > debug it further.
> > >
> > > One other (set of) notes below about using KUNIT_EXPECT_MEMEQ_MSG(),
> > > otherwise (assuming the m68k stuff isn't actually a regression, which
> > > I haven't tested but I imagine is unlikely),
> >
> > I'm trying to debug a hang on m68k in the usercopy behavioral testing
> > routines. It's testing for the pathological case of having inverted
> > arguments to copy_to_user():
> >
> >         user_addr = kunit_vm_mmap(test, NULL, 0, priv->size,
> >                             PROT_READ | PROT_WRITE | PROT_EXEC,
> >                             MAP_ANONYMOUS | MAP_PRIVATE, 0);
> >         ...
> >         bad_usermem = (char *)user_addr;
> >         ...
> >         KUNIT_EXPECT_NE_MSG(test, copy_to_user((char __user *)kmem, bad_usermem,
> >                                                PAGE_SIZE), 0,
> >                 "illegal reversed copy_to_user passed");
> >
> > On other architectures, this immediate fails because the access_ok()
> > check rejects it. On m68k with CONFIG_ALTERNATE_USER_ADDRESS_SPACE,
> > access_ok() short-circuits to "true". I've tried reading
> > arch/m68k/include/asm/uaccess.h but I'm not sure what's happening under
> > CONFIG_CPU_HAS_ADDRESS_SPACES.
> 
> On m68k CPUs that support CPU_HAS_ADDRESS_SPACES (i.e. all traditional
> 680x0 that can run real Linux), the CPU has separate address spaces
> for kernel and user addresses.  Accessing userspace addresses is done
> using the special "moves" instruction, so we can just use the MMU to
> catch invalid accesses.

Okay, that's what I suspected. I think I'll need to just not test this
particular case for archs with separate address spaces, since it would
be meaningless.

> 
> > For now I've excluded that test for m68k, but I'm not sure what's
> > expected to happen here on m68k for this set of bad arguments. Can you
> > advise?
> 
> Perhaps the kernel address is actually a valid user address, or
> vice versa?

Right -- I think that's what's happened.

> 
> Does the test work on systems that use 4G/4G for kernel/userspace
> instead of the usual 1G/3G split?
> 
> /me runs the old test_user_copy.ko on ARAnyM
> Seems to take a while? Or it hangs, too?

Sounds like the same behavior.

> Related reading material
> https://lore.kernel.org/all/CAMuHMdUzHwm5_TL7TNAOF+uqheJnKgsqF+_vzqGRzB_3eufKug@mail.gmail.com/
> https://lore.kernel.org/all/CAMuHMdVQ93ihgcxUbjptTaHdPjxXLyVAsAr-m3tWBJV0krS2vw@mail.gmail.com/

Thanks!

-Kees

-- 
Kees Cook