From: Yonghong Song <yhs@meta.com>
To: Andrii Nakryiko <andrii.nakryiko@gmail.com>, Yonghong Song <yhs@fb.com>
Cc: bpf@vger.kernel.org, Alexei Starovoitov <ast@kernel.org>,
Andrii Nakryiko <andrii@kernel.org>,
Daniel Borkmann <daniel@iogearbox.net>,
kernel-team@fb.com, Martin KaFai Lau <martin.lau@kernel.org>
Subject: Re: [PATCH bpf-next 0/7] bpf: Improve verifier for cond_op and spilled loop index variables
Date: Thu, 6 Apr 2023 09:49:19 -0700 [thread overview]
Message-ID: <b73075cb-67c4-2144-64f9-fc564eb00833@meta.com> (raw)
In-Reply-To: <CAEf4Bzayt_FUG6JyMzU060swqP_w=W9TFJOKD15ux6GNDm3qSg@mail.gmail.com>
On 4/4/23 2:46 PM, Andrii Nakryiko wrote:
> On Wed, Mar 29, 2023 at 10:56 PM Yonghong Song <yhs@fb.com> wrote:
>>
>> LLVM commit [1] introduced hoistMinMax optimization like
>> (i < VIRTIO_MAX_SGS) && (i < out_sgs)
>> to
>> upper = MIN(VIRTIO_MAX_SGS, out_sgs)
>> ... i < upper ...
>> and caused the verification failure. Commit [2] workarounded the issue by
>> adding some bpf assembly code to prohibit the above optimization.
>> This patch improved verifier such that verification can succeed without
>> the above workaround.
>>
>> Without [2], the current verifier will hit the following failures:
>> ...
>> 119: (15) if r1 == 0x0 goto pc+1
>> The sequence of 8193 jumps is too complex.
>> verification time 525829 usec
>> stack depth 64
>> processed 156616 insns (limit 1000000) max_states_per_insn 8 total_states 1754 peak_states 1712 mark_read 12
>> -- END PROG LOAD LOG --
>> libbpf: prog 'trace_virtqueue_add_sgs': failed to load: -14
>> libbpf: failed to load object 'loop6.bpf.o'
>> ...
>> The failure is due to verifier inadequately handling '<const> <cond_op> <non_const>' which will
>> go through both pathes and generate the following verificaiton states:
>> ...
>> 89: (07) r2 += 1 ; R2_w=5
>> 90: (79) r8 = *(u64 *)(r10 -48) ; R8_w=scalar() R10=fp0
>> 91: (79) r1 = *(u64 *)(r10 -56) ; R1_w=scalar(umax=5,var_off=(0x0; 0x7)) R10=fp0
>> 92: (ad) if r2 < r1 goto pc+41 ; R0_w=scalar() R1_w=scalar(umin=6,umax=5,var_off=(0x4; 0x3))
>
> offtopic, but if this is a real output, then something is wrong with
> scratching register logic for conditional, it should have emitted
> states of R1 and R2, maybe you can take a look while working on this
> patch set?
Yes, this is the real output. Yes, the above R1_w should be an
impossible state. This is what this patch tries to fix.
I am not what verifier should do if this state indeed happens,
return an -EFAULT or something?
>
>> R2_w=5 R6_w=scalar(id=385) R7_w=0 R8_w=scalar() R9_w=scalar(umax=21474836475,var_off=(0x0; 0x7ffffffff))
>> R10=fp0 fp-8=mmmmmmmm fp-16=mmmmmmmm fp-24=mmmm???? fp-32= fp-40_w=4 fp-48=mmmmmmmm fp-56= fp-64=mmmmmmmm
>> ...
>> 89: (07) r2 += 1 ; R2_w=6
>> 90: (79) r8 = *(u64 *)(r10 -48) ; R8_w=scalar() R10=fp0
>> 91: (79) r1 = *(u64 *)(r10 -56) ; R1_w=scalar(umax=5,var_off=(0x0; 0x7)) R10=fp0
>> 92: (ad) if r2 < r1 goto pc+41 ; R0_w=scalar() R1_w=scalar(umin=7,umax=5,var_off=(0x4; 0x3))
>> R2_w=6 R6=scalar(id=388) R7=0 R8_w=scalar() R9_w=scalar(umax=25769803770,var_off=(0x0; 0x7ffffffff))
>> R10=fp0 fp-8=mmmmmmmm fp-16=mmmmmmmm fp-24=mmmm???? fp-32= fp-40=5 fp-48=mmmmmmmm fp-56= fp-64=mmmmmmmm
>> ...
>> 89: (07) r2 += 1 ; R2_w=4088
>> 90: (79) r8 = *(u64 *)(r10 -48) ; R8_w=scalar() R10=fp0
>> 91: (79) r1 = *(u64 *)(r10 -56) ; R1_w=scalar(umax=5,var_off=(0x0; 0x7)) R10=fp0
>> 92: (ad) if r2 < r1 goto pc+41 ; R0=scalar() R1=scalar(umin=4089,umax=5,var_off=(0x0; 0x7))
>> R2=4088 R6=scalar(id=12634) R7=0 R8=scalar() R9=scalar(umax=17557826301960,var_off=(0x0; 0xfffffffffff))
>> R10=fp0 fp-8=mmmmmmmm fp-16=mmmmmmmm fp-24=mmmm???? fp-32= fp-40=4087 fp-48=mmmmmmmm fp-56= fp-64=mmmmmmmm
>>
>> Patch 3 fixed the above issue by handling '<const> <cond_op> <non_const>' properly.
>> During developing selftests for Patch 3, I found some issues with bound deduction with
>> BPF_EQ/BPF_NE and fixed the issue in Patch 1.
[...]
next prev parent reply other threads:[~2023-04-06 16:49 UTC|newest]
Thread overview: 26+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-03-30 5:56 [PATCH bpf-next 0/7] bpf: Improve verifier for cond_op and spilled loop index variables Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 1/7] bpf: Improve verifier JEQ/JNE insn branch taken checking Yonghong Song
2023-03-30 21:14 ` Dave Marchevsky
2023-03-31 6:40 ` Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 2/7] selftests/bpf: Add tests for non-constant cond_op NE/EQ bound deduction Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 3/7] bpf: Improve handling of pattern '<const> <cond_op> <non_const>' in verifier Yonghong Song
2023-03-30 22:54 ` Dave Marchevsky
2023-03-31 15:26 ` Yonghong Song
2023-04-04 22:04 ` Andrii Nakryiko
2023-04-06 16:51 ` Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 4/7] selftests/bpf: Add verifier tests for code pattern '<const> <cond_op> <non_const>' Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 5/7] bpf: Mark potential spilled loop index variable as precise Yonghong Song
2023-03-31 21:54 ` Eduard Zingerman
2023-03-31 23:39 ` Yonghong Song
2023-04-03 1:48 ` Alexei Starovoitov
2023-04-03 4:04 ` Yonghong Song
2023-04-04 22:09 ` Andrii Nakryiko
2023-04-06 16:55 ` Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 6/7] selftests/bpf: Remove previous workaround for test verif_scale_loop6 Yonghong Song
2023-03-30 5:56 ` [PATCH bpf-next 7/7] selftests/bpf: Add a new test based on loop6.c Yonghong Song
2023-04-03 1:39 ` Alexei Starovoitov
2023-04-03 3:59 ` Yonghong Song
2023-04-04 21:46 ` [PATCH bpf-next 0/7] bpf: Improve verifier for cond_op and spilled loop index variables Andrii Nakryiko
2023-04-06 16:49 ` Yonghong Song [this message]
2023-04-06 18:38 ` Andrii Nakryiko
2023-04-06 19:54 ` Alexei Starovoitov
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=b73075cb-67c4-2144-64f9-fc564eb00833@meta.com \
--to=yhs@meta.com \
--cc=andrii.nakryiko@gmail.com \
--cc=andrii@kernel.org \
--cc=ast@kernel.org \
--cc=bpf@vger.kernel.org \
--cc=daniel@iogearbox.net \
--cc=kernel-team@fb.com \
--cc=martin.lau@kernel.org \
--cc=yhs@fb.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).