Re: [PATCH bpf-next v3 3/9] bpf/verifier: refine retval R0 state for bpf_get_stack helper

[Yonghong Song <yhs@fb.com>]

On 4/22/18 4:55 PM, Alexei Starovoitov wrote:
> On Fri, Apr 20, 2018 at 03:18:36PM -0700, Yonghong Song wrote:
>> The special property of return values for helpers bpf_get_stack
>> and bpf_probe_read_str are captured in verifier.
>> Both helpers return a negative error code or
>> a length, which is equal to or smaller than the buffer
>> size argument. This additional information in the
>> verifier can avoid the condition such as "retval > bufsize"
>> in the bpf program. For example, for the code blow,
>>      usize = bpf_get_stack(ctx, raw_data, max_len, BPF_F_USER_STACK);
>>      if (usize < 0 || usize > max_len)
>>          return 0;
>> The verifier may have the following errors:
>>      52: (85) call bpf_get_stack#65
>>       R0=map_value(id=0,off=0,ks=4,vs=1600,imm=0) R1_w=ctx(id=0,off=0,imm=0)
>>       R2_w=map_value(id=0,off=0,ks=4,vs=1600,imm=0) R3_w=inv800 R4_w=inv256
>>       R6=ctx(id=0,off=0,imm=0) R7=map_value(id=0,off=0,ks=4,vs=1600,imm=0)
>>       R9_w=inv800 R10=fp0,call_-1
>>      53: (bf) r8 = r0
>>      54: (bf) r1 = r8
>>      55: (67) r1 <<= 32
>>      56: (bf) r2 = r1
>>      57: (77) r2 >>= 32
>>      58: (25) if r2 > 0x31f goto pc+33
>>       R0=inv(id=0) R1=inv(id=0,smax_value=9223372032559808512,
>>                           umax_value=18446744069414584320,
>>                           var_off=(0x0; 0xffffffff00000000))
>>       R2=inv(id=0,umax_value=799,var_off=(0x0; 0x3ff))
>>       R6=ctx(id=0,off=0,imm=0) R7=map_value(id=0,off=0,ks=4,vs=1600,imm=0)
>>       R8=inv(id=0) R9=inv800 R10=fp0,call_-1
>>      59: (1f) r9 -= r8
>>      60: (c7) r1 s>>= 32
>>      61: (bf) r2 = r7
>>      62: (0f) r2 += r1
>>      math between map_value pointer and register with unbounded
>>      min value is not allowed
>> The failure is due to llvm compiler optimization where register "r2",
>> which is a copy of "r1", is tested for condition while later on "r1"
>> is used for map_ptr operation. The verifier is not able to track such
>> inst sequence effectively.
>>
>> Without the "usize > max_len" condition, there is no llvm optimization
>> and the below generated code passed verifier:
>>      52: (85) call bpf_get_stack#65
>>       R0=map_value(id=0,off=0,ks=4,vs=1600,imm=0) R1_w=ctx(id=0,off=0,imm=0)
>>       R2_w=map_value(id=0,off=0,ks=4,vs=1600,imm=0) R3_w=inv800 R4_w=inv256
>>       R6=ctx(id=0,off=0,imm=0) R7=map_value(id=0,off=0,ks=4,vs=1600,imm=0)
>>       R9_w=inv800 R10=fp0,call_-1
>>      53: (b7) r1 = 0
>>      54: (bf) r8 = r0
>>      55: (67) r8 <<= 32
>>      56: (c7) r8 s>>= 32
>>      57: (6d) if r1 s> r8 goto pc+24
>>       R0=inv(id=0,umax_value=800) R1=inv0 R6=ctx(id=0,off=0,imm=0)
>>       R7=map_value(id=0,off=0,ks=4,vs=1600,imm=0)
>>       R8=inv(id=0,umax_value=800,var_off=(0x0; 0x3ff)) R9=inv800
>>       R10=fp0,call_-1
>>      58: (bf) r2 = r7
>>      59: (0f) r2 += r8
>>      60: (1f) r9 -= r8
>>      61: (bf) r1 = r6
>>
>> Signed-off-by: Yonghong Song <yhs@fb.com>
>> ---
>>   kernel/bpf/verifier.c | 27 +++++++++++++++++++++++++++
>>   1 file changed, 27 insertions(+)
>>
>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>> index aba9425..3c8bb92 100644
>> --- a/kernel/bpf/verifier.c
>> +++ b/kernel/bpf/verifier.c
>> @@ -164,6 +164,8 @@ struct bpf_call_arg_meta {
>>   	bool pkt_access;
>>   	int regno;
>>   	int access_size;
>> +	s64 msize_smax_value;
>> +	u64 msize_umax_value;
>>   };
>>   
>>   static DEFINE_MUTEX(bpf_verifier_lock);
>> @@ -2027,6 +2029,14 @@ static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
>>   		err = check_helper_mem_access(env, regno - 1,
>>   					      reg->umax_value,
>>   					      zero_size_allowed, meta);
>> +
>> +		if (!err && !!meta) {
> 
> Please drop !! in the above.
> 
> Also what happens when
> if (!tnum_is_const(reg->var_off))
>    meta = NULL;
> ?
> it seems two new fields of meta will stay zero initialized
> that later do_refine_retval_range() will set R0->umax_value = 0
> which seems incorrect.

Thanks for catching this. In do_refine_retval_range(), if meta is NULL,
the function should just return. Otherwise, a page fault will happen.

> 
>> +			/* remember the mem_size which may be used later
>> +			 * to refine return values.
>> +			 */
>> +			meta->msize_smax_value = reg->smax_value;
>> +			meta->msize_umax_value = reg->umax_value;
>> +		}
>>   	}
>>   
>>   	return err;
>> @@ -2333,6 +2343,21 @@ static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
>>   	return 0;
>>   }
>>   
>> +static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
>> +				   int func_id,
>> +				   struct bpf_call_arg_meta *meta)
>> +{
>> +	struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
>> +
>> +	if (ret_type != RET_INTEGER ||
>> +	    (func_id != BPF_FUNC_get_stack &&
>> +	     func_id != BPF_FUNC_probe_read_str))
>> +		return;
>> +
>> +	ret_reg->smax_value = meta->msize_smax_value;
>> +	ret_reg->umax_value = meta->msize_umax_value;
>> +}
>> +
>>   static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
>>   {
>>   	const struct bpf_func_proto *fn = NULL;
>> @@ -2456,6 +2481,8 @@ static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn
>>   		return -EINVAL;
>>   	}
>>   
>> +	do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
>> +
>>   	err = check_map_func_compatibility(env, meta.map_ptr, func_id);
>>   	if (err)
>>   		return err;
>> -- 
>> 2.9.5
>>