For PTR_TO_FLOW_KEYS, check_flow_keys_access() only uses fixed off
for validation. However, variable offset ptr alu is not prohibited
for this ptr kind. So the variable offset is not checked.
The following prog is accepted:
func#0 @0
0: R1=ctx() R10=fp0
0: (bf) r6 = r1 ; R1=ctx() R6_w=ctx()
1: (79) r7 = *(u64 *)(r6 +144) ; R6_w=ctx() R7_w=flow_keys()
2: (b7) r8 = 1024 ; R8_w=1024
3: (37) r8 /= 1 ; R8_w=scalar()
4: (57) r8 &= 1024 ; R8_w=scalar(smin=smin32=0,
smax=umax=smax32=umax32=1024,var_off=(0x0; 0x400))
5: (0f) r7 += r8
mark_precise: frame0: last_idx 5 first_idx 0 subseq_idx -1
mark_precise: frame0: regs=r8 stack= before 4: (57) r8 &= 1024
mark_precise: frame0: regs=r8 stack= before 3: (37) r8 /= 1
mark_precise: frame0: regs=r8 stack= before 2: (b7) r8 = 1024
6: R7_w=flow_keys(smin=smin32=0,smax=umax=smax32=umax32=1024,var_off
=(0x0; 0x400)) R8_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=1024,
var_off=(0x0; 0x400))
6: (79) r0 = *(u64 *)(r7 +0) ; R0_w=scalar()
7: (95) exit
This prog loads flow_keys to r7, and adds the variable offset r8
to r7, and finally causes out-of-bounds access:
BUG: unable to handle page fault for address: ffffc90014c80038
..
Call Trace:
<TASK>
bpf_dispatcher_nop_func include/linux/bpf.h:1231 [inline]
__bpf_prog_run include/linux/filter.h:651 [inline]
bpf_prog_run include/linux/filter.h:658 [inline]
bpf_prog_run_pin_on_cpu include/linux/filter.h:675 [inline]
bpf_flow_dissect+0x15f/0x350 net/core/flow_dissector.c:991
bpf_prog_test_run_flow_dissector+0x39d/0x620 net/bpf/test_run.c:1359
bpf_prog_test_run kernel/bpf/syscall.c:4107 [inline]
__sys_bpf+0xf8f/0x4560 kernel/bpf/syscall.c:5475
__do_sys_bpf kernel/bpf/syscall.c:5561 [inline]
__se_sys_bpf kernel/bpf/syscall.c:5559 [inline]
__x64_sys_bpf+0x73/0xb0 kernel/bpf/syscall.c:5559
do_syscall_x64 arch/x86/entry/common.c:52 [inline]
do_syscall_64+0x3f/0x110 arch/x86/entry/common.c:83
entry_SYSCALL_64_after_hwframe+0x63/0x6b
Fix this by rejecting ptr alu with variable offset on flow_keys.
Applying the patch makes the program rejected with "R7 pointer
arithmetic on flow_keys prohibited"
Fixes: d58e468b1112 ("flow_dissector: implements flow dissector BPF hook")
Signed-off-by: Hao Sun <[email protected]>
---
kernel/bpf/verifier.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index adbf330d364b..65f598694d55 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c
@@ -12826,6 +12826,10 @@ static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
}
switch (base_type(ptr_reg->type)) {
+ case PTR_TO_FLOW_KEYS:
+ if (known)
+ break;
+ fallthrough;
case CONST_PTR_TO_MAP:
/* smin_val represents the known value */
if (known && smin_val == 0 && opcode == BPF_ADD)
--
2.34.1
On Tue, 2024-01-09 at 16:36 +0100, Hao Sun wrote:
> For PTR_TO_FLOW_KEYS, check_flow_keys_access() only uses fixed off
> for validation. However, variable offset ptr alu is not prohibited
> for this ptr kind. So the variable offset is not checked.
>
[...]
>
> Fixes: d58e468b1112 ("flow_dissector: implements flow dissector BPF hook")
> Signed-off-by: Hao Sun <[email protected]>
> ---
> kernel/bpf/verifier.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index adbf330d364b..65f598694d55 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -12826,6 +12826,10 @@ static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
> }
>
> switch (base_type(ptr_reg->type)) {
> + case PTR_TO_FLOW_KEYS:
> + if (known)
> + break;
> + fallthrough;
> case CONST_PTR_TO_MAP:
> /* smin_val represents the known value */
> if (known && smin_val == 0 && opcode == BPF_ADD)
This change makes sense, could you please add a testcase?
Also, this switch is written to explicitly disallow and implicitly allow
pointer arithmetics, which might be a bit unsafe when new ptr types are added.
Would it make more sense to instead rewrite it to explicitly allow?
E.g. here is what it currently allows / disallows:
| Pointer type | Arithmetics allowed |
|---------------------+---------------------|
| PTR_TO_CTX | yes |
| CONST_PTR_TO_MAP | conditionally |
| PTR_TO_MAP_VALUE | yes |
| PTR_TO_MAP_KEY | yes |
| PTR_TO_STACK | yes |
| PTR_TO_PACKET_META | yes |
| PTR_TO_PACKET | yes |
| PTR_TO_PACKET_END | no |
| PTR_TO_FLOW_KEYS | yes |
| PTR_TO_SOCKET | no |
| PTR_TO_SOCK_COMMON | no |
| PTR_TO_TCP_SOCK | no |
| PTR_TO_TP_BUFFER | yes |
| PTR_TO_XDP_SOCK | no |
| PTR_TO_BTF_ID | yes |
| PTR_TO_MEM | yes |
| PTR_TO_BUF | yes |
| PTR_TO_FUNC | yes |
| CONST_PTR_TO_DYNPTR | yes |
Of these PTR_TO_FUNC and CONST_PTR_TO_DYNPTR (?) should not be allowed
as well, probably (not sure if that could be exploited).
On Tue, Jan 9, 2024 at 5:21 PM Eduard Zingerman <[email protected]> wrote:
>
> On Tue, 2024-01-09 at 16:36 +0100, Hao Sun wrote:
> > For PTR_TO_FLOW_KEYS, check_flow_keys_access() only uses fixed off
> > for validation. However, variable offset ptr alu is not prohibited
> > for this ptr kind. So the variable offset is not checked.
> >
> [...]
> >
> > Fixes: d58e468b1112 ("flow_dissector: implements flow dissector BPF hook")
> > Signed-off-by: Hao Sun <[email protected]>
> > ---
> > kernel/bpf/verifier.c | 4 ++++
> > 1 file changed, 4 insertions(+)
> >
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index adbf330d364b..65f598694d55 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> > @@ -12826,6 +12826,10 @@ static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
> > }
> >
> > switch (base_type(ptr_reg->type)) {
> > + case PTR_TO_FLOW_KEYS:
> > + if (known)
> > + break;
> > + fallthrough;
> > case CONST_PTR_TO_MAP:
> > /* smin_val represents the known value */
> > if (known && smin_val == 0 && opcode == BPF_ADD)
>
> This change makes sense, could you please add a testcase?
>
OK, will do it in the next version tomorrow.
> Also, this switch is written to explicitly disallow and implicitly allow
> pointer arithmetics, which might be a bit unsafe when new ptr types are added.
> Would it make more sense to instead rewrite it to explicitly allow?
Yes, this sounds more safe and clear to me, should be done in another patch.
> E.g. here is what it currently allows / disallows:
>
> | Pointer type | Arithmetics allowed |
> |---------------------+---------------------|
> | PTR_TO_CTX | yes |
> | CONST_PTR_TO_MAP | conditionally |
> | PTR_TO_MAP_VALUE | yes |
> | PTR_TO_MAP_KEY | yes |
> | PTR_TO_STACK | yes |
> | PTR_TO_PACKET_META | yes |
> | PTR_TO_PACKET | yes |
> | PTR_TO_PACKET_END | no |
> | PTR_TO_FLOW_KEYS | yes |
This one should be `conditionally`, variable offset disallowed, fixed allowed.
> | PTR_TO_SOCKET | no |
> | PTR_TO_SOCK_COMMON | no |
> | PTR_TO_TCP_SOCK | no |
> | PTR_TO_TP_BUFFER | yes |
> | PTR_TO_XDP_SOCK | no |
> | PTR_TO_BTF_ID | yes |
> | PTR_TO_MEM | yes |
> | PTR_TO_BUF | yes |
> | PTR_TO_FUNC | yes |
> | CONST_PTR_TO_DYNPTR | yes |
>
> Of these PTR_TO_FUNC and CONST_PTR_TO_DYNPTR (?) should not be allowed
> as well, probably (not sure if that could be exploited).
I think both should be disallowed.
If alu sanitation is triggered, alu op on func and dynptr would be
rejected by retrieve_ptr_limit();
otherwise, it could be dangerous.