Re: Fwd: [PATCH bpf] bpf: Do more tight ALU bounds tracking

[Daniel Borkmann <daniel@iogearbox.net>]

On 8/8/22 5:14 PM, Kuee k1r0a wrote:
> ---------- Forwarded message ---------
> From: Kuee k1r0a <liulin063@gmail.com>
> Date: Mon, Aug 8, 2022 at 11:11 PM
> Subject: Re: [PATCH bpf] bpf: Do more tight ALU bounds tracking
> To: Daniel Borkmann <daniel@iogearbox.net>
> 
> 
> On Mon, Aug 8, 2022 at 9:25 PM Daniel Borkmann <daniel@iogearbox.net> wrote:
>>
>> On 7/30/22 12:48 AM, Hao Luo wrote:
>>> On Fri, Jul 29, 2022 at 3:43 PM Youlin Li <liulin063@gmail.com> wrote:
>>>>
>>>> In adjust_scalar_min_max_vals(), let 32bit bounds learn from 64bit bounds
>>>> to get more tight bounds tracking. Similar operation can be found in
>>>> reg_set_min_max().
>>>>
>>>> Also, we can now fold reg_bounds_sync() into zext_32_to_64().
>>>>
>>>> Before:
>>>>
>>>>       func#0 @0
>>>>       0: R1=ctx(off=0,imm=0) R10=fp0
>>>>       0: (b7) r0 = 0                        ; R0_w=0
>>>>       1: (b7) r1 = 0                        ; R1_w=0
>>>>       2: (87) r1 = -r1                      ; R1_w=scalar()
>>>>       3: (87) r1 = -r1                      ; R1_w=scalar()
>>>>       4: (c7) r1 s>>= 63                    ; R1_w=scalar(smin=-1,smax=0)
>>>>       5: (07) r1 += 2                       ; R1_w=scalar(umin=1,umax=2,var_off=(0x0; 0xffffffff))  <--- [*]
>>>>       6: (95) exit
>>>>
>>>> It can be seen that even if the 64bit bounds is clear here, the 32bit
>>>> bounds is still in the state of 'UNKNOWN'.
>>>>
>>>> After:
>>>>
>>>>       func#0 @0
>>>>       0: R1=ctx(off=0,imm=0) R10=fp0
>>>>       0: (b7) r0 = 0                        ; R0_w=0
>>>>       1: (b7) r1 = 0                        ; R1_w=0
>>>>       2: (87) r1 = -r1                      ; R1_w=scalar()
>>>>       3: (87) r1 = -r1                      ; R1_w=scalar()
>>>>       4: (c7) r1 s>>= 63                    ; R1_w=scalar(smin=-1,smax=0)
>>>>       5: (07) r1 += 2                       ; R1_w=scalar(umin=1,umax=2,var_off=(0x0; 0x3))  <--- [*]
>>>>       6: (95) exit
>>>>
>>>> Signed-off-by: Youlin Li <liulin063@gmail.com>
>>>
>>> Looks good to me. Thanks Youlin.
>>>
>>> Acked-by: Hao Luo <haoluo@google.com>
>>
>> Thanks Youlin! Looks like the patch breaks CI [0] e.g.:
>>
>>     #142/p bounds check after truncation of non-boundary-crossing range FAIL
>>     Failed to load prog 'Permission denied'!
>>     invalid access to map value, value_size=8 off=16777215 size=1
>>     R0 max value is outside of the allowed memory range
>>     verification time 296 usec
>>     stack depth 8
>>     processed 15 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
>>
>> Please take a look. Also it would be great to add a test_verifier selftest to
>> assert above case from commit log against future changes.
>>
>> Thanks,
>> Daniel
>>
>>     [0] https://github.com/kernel-patches/bpf/runs/7696324041?check_suite_focus=true
> 
> This test case fails because the 32bit boundary information is lost
> after the 11th instruction is executed:
> Before:
>      11: (07) r1 += 2147483647             ;
> R1_w=scalar(umin=70866960383,umax=70866960638,var_off=(0x1000000000;
> 0xffffffff),u32_min=2147483647,u32_max=-2147483394)
> After:
>      11: (07) r1 += 2147483647             ;
> R1_w=scalar(umin=70866960383,umax=70866960638,var_off=(0x1000000000;
> 0xffffffff))
> 
> This may be because, in previous versions of the code, when
> __reg_combine_64_into_32() was called, the 32bit boundary was
> completely deduced from the 64bit boundary, so there was a call to
> __mark_reg32_unbounded() in __reg_combine_64_into_32().
> 
> But now, before adjust_scalar_min_max_vals() calls
> __reg_combine_64_into_32() , the 32bit bounds are already calculated
> to some extent, and __mark_reg32_unbounded() will eliminate these
> information.
> 
> Simply copying a code without __mark_reg32_unbounded() should work,
> perhaps it would be more elegant to introduce a flag into
> __reg_combine_64_into_32()?
> 
> Sorry for not completing the tests because I did not 'make selftests'
> successfully, and uploaded the code that caused the error.

Under tools/testing/selftests/bpf/, you can run test_progs and test_verifier
through the vmtest script, e.g. `./vmtest.sh -- ./test_progs` should ease
running it. The whole `make selftests` is not necessary given here we care
about BPF, CI is running these where 2 failed and need investigation:

           test_progs: PASS
  test_progs-no_alu32: FAIL (returned 1)
            test_maps: PASS
        test_verifier: FAIL (returned 1)

Fwiw, for the test_verifier failure case at least, we should then adapt it
in a separate commit with an analysis explaining why it is okay to alter the
test; plus a 3rd commit adding new test cases as mentioned earlier.

Thanks a lot, Kuee!
Daniel