Received-SPF: pass (google.com: domain of linux-kernel+bounces-112461-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Precedence: bulk
MIME-Version: 1.0
References: <20240321124640.8870-1-puranjay12@gmail.com> <9f2b63b5-569c-1e00-a635-93d9cd695517@iogearbox.net>
 <mb61p4jcyxq5m.fsf@gmail.com> <15ba79e3-14b2-d92e-3f94-e4f5f963e15d@iogearbox.net>
In-Reply-To: <15ba79e3-14b2-d92e-3f94-e4f5f963e15d@iogearbox.net>
From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Date: Sat, 23 Mar 2024 12:13:06 -0700
Message-ID: <CAADnVQJhxQV0xBJDWVkTGXJ+XYpsq5kFmMsuMEHBEaqUmkh0iw@mail.gmail.com>
Subject: Re: [PATCH bpf v4] bpf: verifier: prevent userspace memory access
To: Daniel Borkmann <daniel@iogearbox.net>
Cc: Puranjay Mohan <puranjay12@gmail.com>, "David S. Miller" <davem@davemloft.net>, 
	David Ahern <dsahern@kernel.org>, Alexei Starovoitov <ast@kernel.org>, Andrii Nakryiko <andrii@kernel.org>, 
	Martin KaFai Lau <martin.lau@linux.dev>, Eduard Zingerman <eddyz87@gmail.com>, Song Liu <song@kernel.org>, 
	Yonghong Song <yonghong.song@linux.dev>, John Fastabend <john.fastabend@gmail.com>, 
	KP Singh <kpsingh@kernel.org>, Stanislav Fomichev <sdf@google.com>, Hao Luo <haoluo@google.com>, 
	Jiri Olsa <jolsa@kernel.org>, Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@redhat.com>, 
	Borislav Petkov <bp@alien8.de>, Dave Hansen <dave.hansen@linux.intel.com>, X86 ML <x86@kernel.org>, 
	"H. Peter Anvin" <hpa@zytor.com>, Jean-Philippe Brucker <jean-philippe@linaro.org>, 
	Network Development <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>, 
	LKML <linux-kernel@vger.kernel.org>, Ilya Leoshkevich <iii@linux.ibm.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Fri, Mar 22, 2024 at 9:28=E2=80=AFAM Daniel Borkmann <daniel@iogearbox.n=
et> wrote:
>
> On 3/22/24 4:05 PM, Puranjay Mohan wrote:
> [...]
> >>> +           /* Make it impossible to de-reference a userspace address=
 */
> >>> +           if (BPF_CLASS(insn->code) =3D=3D BPF_LDX &&
> >>> +               (BPF_MODE(insn->code) =3D=3D BPF_PROBE_MEM ||
> >>> +                BPF_MODE(insn->code) =3D=3D BPF_PROBE_MEMSX)) {
> >>> +                   struct bpf_insn *patch =3D &insn_buf[0];
> >>> +                   u64 uaddress_limit =3D bpf_arch_uaddress_limit();
> >>> +
> >>> +                   if (!uaddress_limit)
> >>> +                           goto next_insn;
> >>> +
> >>> +                   *patch++ =3D BPF_MOV64_REG(BPF_REG_AX, insn->src_=
reg);
> >>> +                   if (insn->off)
> >>> +                           *patch++ =3D BPF_ALU64_IMM(BPF_ADD, BPF_R=
EG_AX, insn->off);
> >>> +                   *patch++ =3D BPF_ALU64_IMM(BPF_RSH, BPF_REG_AX, 3=
2);
> >>> +                   *patch++ =3D BPF_JMP_IMM(BPF_JLE, BPF_REG_AX, uad=
dress_limit >> 32, 2);
> >>> +                   *patch++ =3D *insn;
> >>> +                   *patch++ =3D BPF_JMP_IMM(BPF_JA, 0, 0, 1);
> >>> +                   *patch++ =3D BPF_MOV64_IMM(insn->dst_reg, 0);
> >>
> >> But how does this address other cases where we could fault e.g. non-ca=
nonical,
> >> vsyscall page, etc? Technically, we would have to call to copy_from_ke=
rnel_nofault_allowed()
> >> to really address all the cases aside from the overflow (good catch bt=
w!) where kernel
> >> turns into user address.
> >
> > So, we are trying to ~simulate a call to
> > copy_from_kernel_nofault_allowed() here. If the address under
> > consideration is below TASK_SIZE (TASK_SIZE + 4GB to be precise) then w=
e
> > skip that load because that address could be mapped by the user.
> >
> > If the address is above TASK_SIZE + 4GB, we allow the load and it could
> > cause a fault if the address is invalid, non-canonical etc. Taking the
> > fault is fine because JIT will add an exception table entry for
> > for that load with BPF_PBOBE_MEM.
>
> Are you sure? I don't think the kernel handles non-canonical fixup.

I believe it handles it fine otherwise our selftest bpf_testmod_return_ptr:
   case 4: return (void *)(1ull << 60);    /* non-canonical and invalid */
would have been crashing for the last 3 years,
since we've been running it.

> > The vsyscall page is special, this approach skips all loads from this
> > page. I am not sure if that is acceptable.
>
> The bpf_probe_read_kernel() does handle it fine via copy_from_kernel_nofa=
ult().
>
> So there is tail risk that BPF_PROBE_* could trigger a crash.

For this patch let's do
return max(TASK_SIZE_MAX + PAGE_SIZE, VSYSCALL_ADDR)
to cover both with one check?

> Other archs might
> have other quirks, e.g. in case of loongarch it says highest bit set mean=
s kernel
> space.

let's tackle loongarch with whatever quirks it has separately.