On Sun, Aug 1, 2021 at 1:38 AM Johan Almbladh
[off-list ref] wrote:

On Fri, Jul 30, 2021 at 12:48 AM Andrii Nakryiko
[off-list ref] wrote:

quoted

On Thu, Jul 29, 2021 at 3:29 PM Andrii Nakryiko
[off-list ref] wrote:

quoted

On Thu, Jul 29, 2021 at 2:38 PM Johan Almbladh
[off-list ref] wrote:

quoted

On Wed, Jul 28, 2021 at 9:13 PM Yonghong Song [off-list ref] wrote:

quoted

I also checked arm/arm64 jit. I saw the following comments:

         /* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
          *      goto out;
          * tail_call_cnt++;
          */

Maybe we have this MAX_TAIL_CALL_CNT + 1 issue
for arm/arm64 jit?

That wouldn't be unreasonable. I don't have an arm or arm64 setup
available right now, but I can try to test it in qemu.

On a brief check, there seems to be quite a mess in terms of the code
and comments.

E.g., in arch/x86/net/bpf_jit_comp32.c:

        /*
         * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
         *     goto out;
         */

                            ^^^^ here comment is wrong

        [...]

        /* cmp edx,hi */
        EMIT3(0x83, add_1reg(0xF8, IA32_EBX), hi);
        EMIT2(IA32_JNE, 3);
        /* cmp ecx,lo */
        EMIT3(0x83, add_1reg(0xF8, IA32_ECX), lo);

        /* ja out */
        EMIT2(IA32_JAE, jmp_label(jmp_label1, 2));

        ^^^ JAE is >=, right? But the comment says JA.


As for arch/x86/net/bpf_jit_comp.c, both comment and the code seem to
do > MAX_TAIL_CALL_CNT, but you are saying JIT is correct. What am I
missing?

Can you please check all the places where MAX_TAIL_CALL_CNT is used
throughout the code? Let's clean this up in one go.

Also, given it's so easy to do this off-by-one error, can you please
add a negative test validating that 33 tail calls are not allowed? I
assume we have a positive test that allows exactly MAX_TAIL_CALL_CNT,
but please double-check that as well.

Ok, I see that you've added this in your bpf tests patch set. Please
consider, additionally, implementing a similar test as part of
selftests/bpf (specifically in test_progs). We run test_progs
continuously in CI for every incoming patch/patchset, so it has much
higher chances of capturing any regressions.

I'm also thinking that this MAX_TAIL_CALL_CNT change should probably
go into the bpf-next tree. First, this off-by-one behavior was around
for a while and it doesn't cause serious issues, even if abused. But
on the other hand, it will make your tail call tests fail, when
applied into bpf-next without your change. So I think we should apply
both into bpf-next.

I can confirm that the off-by-one behaviour is present on arm. Below
is the test output running on qemu. Test #4 calls itself recursively
and increments a counter each time, so the correct result should be 1
+ MAX_TAIL_CALL_CNT.

test_bpf: #0 Tail call leaf jited:1 71 PASS
test_bpf: #1 Tail call 2 jited:1 134 PASS
test_bpf: #2 Tail call 3 jited:1 164 PASS
test_bpf: #3 Tail call 4 jited:1 257 PASS
test_bpf: #4 Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
test_bpf: #5 Tail call error path, NULL target jited:1 114 PASS
test_bpf: #6 Tail call error path, index out of range jited:1 112 PASS
test_bpf: test_tail_calls: Summary: 6 PASSED, 1 FAILED, [7/7 JIT'ed]

The MAX_TAIL_CALL_CNT constant is referenced in the following JITs.

arch/arm64/net/bpf_jit_comp.c
arch/arm/net/bpf_jit_32.c
arch/mips/net/ebpf_jit.c
arch/powerpc/net/bpf_jit_comp32.c
arch/powerpc/net/bpf_jit_comp64.c
arch/riscv/net/bpf_jit_comp32.c
arch/riscv/net/bpf_jit_comp64.c
arch/s390/net/bpf_jit_comp.c
arch/sparc/net/bpf_jit_comp_64.c
arch/x86/net/bpf_jit_comp32.c
arch/x86/net/bpf_jit_comp.c

The x86 JITs all pass the test, even though the comments are wrong.
The comments can easily be fixed of course. For JITs that have the
off-by-one behaviour, an easy fix would be to change all occurrences
of MAX_TAIL_CALL_CNT to MAX_TAIL_CALL_CNT - 1. We must first know
which JITs affected though.

If you are going to fix ARM, please send a fix to comments for x86 as well.

The fix is easy but setting up the test is hard. It took me quite some
time to get the qemu/arm setup up and running. If the same has to be
done for arm64, mips64, powerpc, powerpc64, riscv32, risc64, sparc and
s390, I will need some help with this. If someone already has a
working setup for any of the systems, the test can be performed on
that.

Unfortunately, I myself have only x86-64 setup. libbpf
CI/kernel-patches CI we use to run all tests are running selftests
against x86-64 only as well. There was temporarily halted effort to
add s390x support as well, but it's not done yet. No one yet
volunteered to set up any other platforms and I don't know if that's
possible and how hard it would be to do within Github Actions platform
we are currently using.

So in short, I understand the challenges of testing all those
platforms and I don't really expect any single person to do all that
work. I've applied your fix, please follow up with ARM and comment
fixes.

Or perhaps there is a better way to do this? If I implement a similar
test in selftest/bpf, that would trigger the CI when the patch is
submitted and we will see which JITs we need to fix.

The other nice benefit of implementing this in selftest/bpf, besides
continuous testing, is that you write it in C, which allows you to
express much more complicated logic more easily.

quoted

On a related topic, please don't forget to include the target kernel
tree for your patches: [PATCH bpf] or [PATCH bpf-next].

I'll add that! All patches I sent related to this are for the bpf-next tree.

Johan