Date:   Mon, 29 Aug 2016 22:17:10 +0200
From:   Daniel Borkmann <daniel@...earbox.net>
To:     Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Jakub Kicinski <jakub.kicinski@...ronome.com>
CC:     netdev@...r.kernel.org, ast@...nel.org,
        dinan.gunawardena@...ronome.com, jiri@...nulli.us,
        john.fastabend@...il.com, kubakici@...pl
Subject: Re: [RFCv2 07/16] bpf: enable non-core use of the verfier

On 08/29/2016 10:13 PM, Daniel Borkmann wrote:
> On 08/27/2016 07:32 PM, Alexei Starovoitov wrote:
>> On Sat, Aug 27, 2016 at 12:40:04PM +0100, Jakub Kicinski wrote:
>>> On Fri, 26 Aug 2016 16:29:05 -0700, Alexei Starovoitov wrote:
>>>> On Fri, Aug 26, 2016 at 07:06:06PM +0100, Jakub Kicinski wrote:
>>>>> Advanced JIT compilers and translators may want to use
>>>>> eBPF verifier as a base for parsers or to perform custom
>>>>> checks and validations.
>>>>>
>>>>> Add ability for external users to invoke the verifier
>>>>> and provide callbacks to be invoked for every intruction
>>>>> checked.  For now only add most basic callback for
>>>>> per-instruction pre-interpretation checks is added.  More
>>>>> advanced users may also like to have per-instruction post
>>>>> callback and state comparison callback.
>>>>>
>>>>> Signed-off-by: Jakub Kicinski <jakub.kicinski@...ronome.com>
>>>>
>>>> I like the apporach. Making verifier into 'bytecode parser'
>>>> that JITs can reuse is a good design choice.
>
> +1
>
>>>> The only thing I would suggest is to tweak the verifier to
>>>> avoid in-place state recording. Then I think patch 8 for
>>>> clone/unclone of the program won't be needed, since verifier
>>>> will be read-only from bytecode point of view and patch 9
>>>> also will be slightly cleaner.
>>>> I think there are very few places in verifier that do this
>>>> state keeping inside insn. It was bugging me for some time.
>>>> Good time to clean that up.
>>>> Unless I misunderstand the patches 7,8,9...
>>>
>>> Agreed, I think the verifier only modifies the program to
>>> store pointer types in imm field.  I will try to come up
>>> a way around this, any suggestions?  Perhaps state_equal()
>>
>> probably array_of_insn_aux_data[num_insns] should do it.
>> Unlike reg_state that is forked on branches, this array
>> is only one.
>
> This would be for struct nfp_insn_meta, right? So, struct
> bpf_ext_parser_ops could become:
>
> static const struct bpf_ext_parser_ops nfp_bpf_pops = {
>      .insn_hook = nfp_verify_insn,
>      .insn_size = sizeof(struct nfp_insn_meta),
> };
>
> ... where bpf_parse() would prealloc that f.e. in env->insn_meta[].

(Well, actually everything can live in env->private_data.)

>>> logic could be modified to downgrade pointers to UNKONWNs
>>> when it detects other state had incompatible pointer type.
>>>
>>>> There is also small concern for patches 5 and 6 that add
>>>> more register state information. Potentially that extra
>>>> state can prevent states_equal() to recognize equivalent
>>>> states. Only patch 9 uses that info, right?
>>>
>>> 5 and 6 recognize more constant loads, those can only
>>> upgrade some UNKNOWN_VALUEs to CONST_IMMs.  So yes, if the
>>> verifier hits the CONST first and then tries with UNKNOWN
>>> it will have to reverify the path.
>
> Agree, was also my concern when I read patch 5 and 6. It would
> not only be related to types, but also different imm values,
> where the memcmp() could fail on. Potentially the latter can be
> avoided by only checking types which should be sufficient. Hmm,
> maybe only bpf_parse() should go through this stricter mode since
> only relevant for drivers (otoh downside would be that bugs
> would end up less likely to be found).
>
>>>> Another question is do you need all state walking that
>>>> verifier does or single linear pass through insns
>>>> would have worked?
>>>> Looks like you're only using CONST_IMM and PTR_TO_CTX
>>>> state, right?
>>>
>>> I think I need all the parsing.  Right now I mostly need
>>> the verification to check that exit codes are specific
>>> CONST_IMMs.  Clang quite happily does this:
>>>
>>> r0 <- 0
>>> if (...)
>>>     r0 <- 1
>>> exit
>>
>> I see. Indeed then you'd need the verifier to walk all paths
>> to make sure constant return values.
>
> I think this would still not cover the cases where you'd fetch
> a return value/verdict from a map, but this should be ignored/
> rejected for now, also since majority of programs are not written
> in such a way.
>
>> If you only need yes/no check then such info can probably be
>> collected unconditionally during initial program load.
>> Like prog->cb_access flag.
>
> One other comment wrt the header, when you move these things
> there, would be good to prefix with bpf_* so that this doesn't
> clash in future with other header files.

Powered by blists - more mailing lists