regressions.lists.linux.dev archive mirror
 help / color / mirror / Atom feed
From: Lorenz Bauer <lmb@cloudflare.com>
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Alexei Starovoitov <ast@kernel.org>,
	kernel-team <kernel-team@cloudflare.com>,
	 bpf <bpf@vger.kernel.org>,
	regressions@lists.linux.dev,  Andrii Nakryiko <andrii@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>
Subject: Re: Verifier rejects previously accepted program
Date: Mon, 8 Nov 2021 13:21:12 +0000	[thread overview]
Message-ID: <CACAyw99KGdTAz+G3aU8G3eqC926YYpgD57q-A+NFNVqqiJPY3g@mail.gmail.com> (raw)
In-Reply-To: <20211105194952.xve6u6lgh2oy46dy@ast-mbp.dhcp.thefacebook.com>

On Fri, 5 Nov 2021 at 19:49, Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Fri, Nov 05, 2021 at 10:41:40AM +0000, Lorenz Bauer wrote:
> >
> > bpf-next with f30d4968e9ae on top:
> >
> >     works!
>
> Awesome.
>
> > commit 3e8ce29850f1 ("bpf: Prevent pointer mismatch in
> > bpf_timer_init.") (found via bisection):
> >
> >     BPF program is too large. Processed 1000001 insn
> >
> > commit 3e8ce29850f1^ ("bpf: Add map side support for bpf timers."):
> >
> >    works!
>
> So with just 3e8ce29850f1 it's "too large" and with parent commit it works ?
> I've analyzed offending commit again and don't see how it can be causing
> state pruning to be more conservative for your asm.
> reg->map_uid should only be non-zero for lookups from inner maps,
> but your asm doesn't have lookups at all in that loop.

I misattributed the problem to the loop, since it was really prominent
in the verifier output. We use nested maps extensively, most likely
those are what's causing the problem.

> Maybe in some case map_uid doesn't get cleared, but I couldn't find
> such code path with manual code analysis.
> I think it's worth investigating further.
> Please craft a reproducer.

I've started with some verifier log analysis to narrow the problem down.

* Same test case as before
* Dump verifier output with log_level=2 for both 3e8ce29850f1 and 3e8ce29850f1^
* Use diff to find the first non-matching line

3e8ce29850f1 makes the verifier do a lot more work on our code. Some
later commit then drops the complexity below what the verifier will
accept, probably the more precise scalar spill tracking.

3e8ce29850f1^:                  295498 insns
3e8ce29850f1:                > 1000000 insns
be2f2d1680df + bd479d103883:    450161 insns

Trace from 3e8ce29850f1^ (working):

1033: R0=map_value(id=0,off=0,ks=4,vs=36,imm=0) R1_w=invP0
R3_w=map_value(id=0,off=0,ks=4,vs=36,imm=0) R6=ctx(id=0,off=0,imm=0)
R7=inv(id=0) R8=pkt(id=0,off=18,r=38,imm=0) R9=inv0 R10=fp0
fp-24=mmmmmmmm fp-32=mmmmmmmm fp-40=mmmm00m0 fp-48=mmmm0000
fp-56=00000000 fp-64=00000000 fp-72=0000mmmm fp-80=mmmmmmmm
fp-88=map_value fp-96=pkt_end fp-104=map_value fp-112=pkt fp-120=fp
fp-128=map_value
1033: (16) if w1 == 0x0 goto pc+43
1077: safe
1178: R0=inv0 R1=map_ptr(id=0,off=0,ks=4,vs=4,imm=0) R2_w=inv0
R3=inv2388976653695081527 R4=inv-8645972361240307355 R5=inv(id=6898)
R6=ctx(id=0,off=0,imm=0) R7=inv(id=0) R8=pkt(id=0,off=18,r=38,imm=0)
R9=inv0 R10=fp0 fp-24=mmmmmmmm fp-32=mmmmmmmm fp-40=mmmm00m0
fp-48=mmmm0000 fp-56=00000000 fp-64=00000000 fp-72=0000mmmm
fp-80=mmmmmmmm fp-88=map_value fp-96=pkt_end fp-104=map_value
fp-112=pkt fp-120=fp fp-128=map_value
1178: (63) *(u32 *)(r10 -32) = r7
<...>
processed 295498 insns (limit 1000000) max_states_per_insn 29
total_states 14527 peak_states 1322 mark_read 53

Trace from 3e8ce29850f1 (broken):

1033: R0=map_value(id=0,off=0,ks=4,vs=36,imm=0) R1_w=invP0
R3_w=map_value(id=0,off=0,ks=4,vs=36,imm=0) R6=ctx(id=0,off=0,imm=0)
R7=inv(id=0) R8=pkt(id=0,off=18,r=38,imm=0) R9=inv0 R10=fp0
fp-24=mmmmmmmm fp-32=mmmmmmmm fp-40=mmmm00m0 fp-48=mmmm0000
fp-56=00000000 fp-64=00000000 fp-72=0000mmmm fp-80=mmmmmmmm
fp-88=map_value fp-96=pkt_end fp-104=map_value fp-112=pkt fp-120=fp
fp-128=map_value
1033: (16) if w1 == 0x0 goto pc+43
1077: R0=map_value(id=0,off=0,ks=4,vs=36,imm=0) R1_w=invP0
R3_w=map_value(id=0,off=0,ks=4,vs=36,imm=0) R6=ctx(id=0,off=0,imm=0)
R7=inv(id=0) R8=pkt(id=0,off=18,r=38,imm=0) R9=inv0 R10=fp0
fp-24=mmmmmmmm fp-32=mmmmmmmm fp-40=mmmm00m0 fp-48=mmmm0000
fp-56=00000000 fp-64=00000000 fp-72=0000mmmm fp-80=mmmmmmmm
fp-88=map_value fp-96=pkt_end fp-104=map_value fp-112=pkt fp-120=fp
fp-128=map_value
1077: (79) r2 = *(u64 *)(r10 -128)
1078: R0=map_value(id=0,off=0,ks=4,vs=36,imm=0) R1_w=invP0
R2_w=map_value(id=0,off=0,ks=4,vs=32,imm=0)
R3_w=map_value(id=0,off=0,ks=4,vs=36,imm=0) R6=ctx(id=0,off=0,imm=0)
R7=inv(id=0) R8=pkt(id=0,off=18,r=38,imm=0) R9=inv0 R10=fp0
fp-24=mmmmmmmm fp-32=mmmmmmmm fp-40=mmmm00m0 fp-48=mmmm0000
fp-56=00000000 fp-64=00000000 fp-72=0000mmmm fp-80=mmmmmmmm
fp-88=map_value fp-96=pkt_end fp-104=map_value fp-112=pkt fp-120=fp
fp-128=map_value
1078: (79) r1 = *(u64 *)(r2 +0)
<...>
(truncated)

Trace from be2f2d1680df ("libbpf: Deprecate bpf_program__load() API")
with bd479d103883 ("bpf: Do not reject when the stack read size is
different from the tracked scalar size") cherry picked:

processed 450161 insns (limit 1000000) max_states_per_insn 19
total_states 19452 peak_states 1319 mark_read 53

r2 is the result of a lookup from a per-CPU array, ts_metrics in the
snippet below:

struct bpf_map_def traffic_set_metrics_map __section("maps") = {
    .type        = BPF_MAP_TYPE_PERCPU_ARRAY,
    .key_size    = sizeof(traffic_set_id_t),
    .value_size  = sizeof(traffic_set_metrics_t),
    .max_entries = SET_BY_USERSPACE,
};

    traffic_set_metrics_t *ts_metrics =
bpf_map_lookup_elem(&traffic_set_metrics_map, &meta->ts_id);
    if (ts_metrics == NULL) {
        return XDP_ABORTED;
    }

   <...>

   if (meta->from_plurimog) {
        ts_metrics->packets_total_plurimog_ingress++;
    } else {
        ts_metrics->packets_total_main++; // insn 1078
    }

Lorenz

-- 
Lorenz Bauer  |  Systems Engineer
6th Floor, County Hall/The Riverside Building, SE1 7PB, UK

www.cloudflare.com

  reply	other threads:[~2021-11-08 13:21 UTC|newest]

Thread overview: 17+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-11-03 11:55 Verifier rejects previously accepted program Lorenz Bauer
2021-11-04 16:50 ` Alexei Starovoitov
2021-11-04 23:30   ` sdf
2021-11-05  1:20     ` Alexei Starovoitov
2021-11-05  4:13       ` Stanislav Fomichev
2021-11-05 10:41   ` Lorenz Bauer
2021-11-05 19:49     ` Alexei Starovoitov
2021-11-08 13:21       ` Lorenz Bauer [this message]
2021-11-10  4:25         ` Alexei Starovoitov
2021-11-10 11:41           ` Lorenz Bauer
2021-11-10 16:50             ` Alexei Starovoitov
2021-11-10 17:05               ` Lorenz Bauer
2021-11-10 18:01               ` Thorsten Leemhuis
2021-11-10 19:16                 ` Alexei Starovoitov
2021-11-10 19:49                   ` Thorsten Leemhuis
2021-11-16  9:26 ` Lorenz Bauer
2021-11-16 10:59   ` Thorsten Leemhuis

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=CACAyw99KGdTAz+G3aU8G3eqC926YYpgD57q-A+NFNVqqiJPY3g@mail.gmail.com \
    --to=lmb@cloudflare.com \
    --cc=alexei.starovoitov@gmail.com \
    --cc=andrii@kernel.org \
    --cc=ast@kernel.org \
    --cc=bpf@vger.kernel.org \
    --cc=daniel@iogearbox.net \
    --cc=kernel-team@cloudflare.com \
    --cc=regressions@lists.linux.dev \
    /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).