On Mon, Aug 2, 2021 at 10:28 PM Andrii Nakryiko
[off-list ref] wrote:

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

quoted

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.

quoted

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.

Thanks! I will fix the ARM JIT and the comments, then submit an
updated patch set for the test suite with changes after Yonghong's
review.

My current test setup can easily cross-compile the kernel with busybox
as userspace. However, getting it to run on QEMU has required some
amount of detective work. Every platforms seems to be different in
terms of what to boot (vmlinux, zImage, bzImage), how to boot it (dtb,
bios, uBoot requirements) and QEMU vs Kconfig settings. Currently I
can run i386, x86_64, MIPS, MIPS64 and ARM under QEMU. I can verify
and if needed fix the JIT on some of the other platforms as well, if I
can get it to run on QEMU with a reasonable effort. However, I cannot
build for RISC-V since I don't have a toolchain for that. I build my
toolchains with crosstool-ng using libmusl, and the latter does not
currently support RISC-V.

As a side note, I think having a QEMU-compatible defconfig for each
platform would make it easier to test arch-specific code. It could
also be a first step towards fully automated arch-specific CI.

Sorry for being a bit slow to respond. I am currently travelling with
only sporadic access to e-mail.

quoted

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

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