Re: [PATCH -next] bpf, test_run: fix alignment problem in bpf_prog_test_run_skb()

[Eric Dumazet]

On Tue, Nov 1, 2022 at 9:27 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Tue, Nov 01, 2022 at 09:05:42PM -0700, Jakub Kicinski wrote:
> > On Wed, 2 Nov 2022 10:59:44 +0800 zhongbaisong wrote:
> > > On 2022/11/2 0:45, Daniel Borkmann wrote:
> > > > [ +kfence folks ]
> > >
> > > + cc: Alexander Potapenko, Marco Elver, Dmitry Vyukov
> > >
> > > Do you have any suggestions about this problem?
> >
> > + Kees who has been sending similar patches for drivers
> >
> > > > On 11/1/22 5:04 AM, Baisong Zhong wrote:
> > > >> Recently, we got a syzkaller problem because of aarch64
> > > >> alignment fault if KFENCE enabled.
> > > >>
> > > >> When the size from user bpf program is an odd number, like
> > > >> 399, 407, etc, it will cause skb shard info's alignment access,
> > > >> as seen below:
> > > >>
> > > >> BUG: KFENCE: use-after-free read in __skb_clone+0x23c/0x2a0
> > > >> net/core/skbuff.c:1032
> > > >>
> > > >> Use-after-free read at 0xffff6254fffac077 (in kfence-#213):
> > > >>   __lse_atomic_add arch/arm64/include/asm/atomic_lse.h:26 [inline]
> > > >>   arch_atomic_add arch/arm64/include/asm/atomic.h:28 [inline]
> > > >>   arch_atomic_inc include/linux/atomic-arch-fallback.h:270 [inline]
> > > >>   atomic_inc include/asm-generic/atomic-instrumented.h:241 [inline]
> > > >>   __skb_clone+0x23c/0x2a0 net/core/skbuff.c:1032
> > > >>   skb_clone+0xf4/0x214 net/core/skbuff.c:1481
> > > >>   ____bpf_clone_redirect net/core/filter.c:2433 [inline]
> > > >>   bpf_clone_redirect+0x78/0x1c0 net/core/filter.c:2420
> > > >>   bpf_prog_d3839dd9068ceb51+0x80/0x330
> > > >>   bpf_dispatcher_nop_func include/linux/bpf.h:728 [inline]
> > > >>   bpf_test_run+0x3c0/0x6c0 net/bpf/test_run.c:53
> > > >>   bpf_prog_test_run_skb+0x638/0xa7c net/bpf/test_run.c:594
> > > >>   bpf_prog_test_run kernel/bpf/syscall.c:3148 [inline]
> > > >>   __do_sys_bpf kernel/bpf/syscall.c:4441 [inline]
> > > >>   __se_sys_bpf+0xad0/0x1634 kernel/bpf/syscall.c:4381
> > > >>
> > > >> kfence-#213: 0xffff6254fffac000-0xffff6254fffac196, size=407,
> > > >> cache=kmalloc-512
> > > >>
> > > >> allocated by task 15074 on cpu 0 at 1342.585390s:
> > > >>   kmalloc include/linux/slab.h:568 [inline]
> > > >>   kzalloc include/linux/slab.h:675 [inline]
> > > >>   bpf_test_init.isra.0+0xac/0x290 net/bpf/test_run.c:191
> > > >>   bpf_prog_test_run_skb+0x11c/0xa7c net/bpf/test_run.c:512
> > > >>   bpf_prog_test_run kernel/bpf/syscall.c:3148 [inline]
> > > >>   __do_sys_bpf kernel/bpf/syscall.c:4441 [inline]
> > > >>   __se_sys_bpf+0xad0/0x1634 kernel/bpf/syscall.c:4381
> > > >>   __arm64_sys_bpf+0x50/0x60 kernel/bpf/syscall.c:4381
> > > >>
> > > >> To fix the problem, we round up allocations with kmalloc_size_roundup()
> > > >> so that build_skb()'s use of kize() is always alignment and no special
> > > >> handling of the memory is needed by KFENCE.
> > > >>
> > > >> Fixes: 1cf1cae963c2 ("bpf: introduce BPF_PROG_TEST_RUN command")
> > > >> Signed-off-by: Baisong Zhong <zhongbaisong@xxxxxxxxxx>
> > > >> ---
> > > >>   net/bpf/test_run.c | 1 +
> > > >>   1 file changed, 1 insertion(+)
> > > >>
> > > >> diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
> > > >> index 13d578ce2a09..058b67108873 100644
> > > >> --- a/net/bpf/test_run.c
> > > >> +++ b/net/bpf/test_run.c
> > > >> @@ -774,6 +774,7 @@ static void *bpf_test_init(const union bpf_attr
> > > >> *kattr, u32 user_size,
> > > >>       if (user_size > size)
> > > >>           return ERR_PTR(-EMSGSIZE);
> > > >> +    size = kmalloc_size_roundup(size);
> > > >>       data = kzalloc(size + headroom + tailroom, GFP_USER);
> > > >
> > > > The fact that you need to do this roundup on call sites feels broken, no?
> > > > Was there some discussion / consensus that now all k*alloc() call sites
> > > > would need to be fixed up? Couldn't this be done transparently in k*alloc()
> > > > when KFENCE is enabled? I presume there may be lots of other such occasions
> > > > in the kernel where similar issue triggers, fixing up all call-sites feels
> > > > like ton of churn compared to api-internal, generic fix.
>
> I hope I answer this in more detail here:
> https://lore.kernel.org/lkml/202211010937.4631CB1B0E@keescook/
>
> The problem is that ksize() should never have existed in the first
> place. :P Every runtime bounds checker has tripped over it, and with
> the addition of the __alloc_size attribute, I had to start ripping
> ksize() out: it can't be used to pretend an allocation grew in size.
> Things need to either preallocate more or go through *realloc() like
> everything else. Luckily, ksize() is rare.
>
> FWIW, the above fix doesn't look correct to me -- I would expect this to
> be:
>
>         size_t alloc_size;
>         ...
>         alloc_size = kmalloc_size_roundup(size + headroom + tailroom);
>         data = kzalloc(alloc_size, GFP_USER);

Making sure the struct skb_shared_info is aligned to a cache line does
not need kmalloc_size_roundup().

What is needed is to adjust @size so that (@size + @headroom) is a
multiple of SMP_CACHE_BYTES