On Tue, Nov 2, 2021 at 6:48 AM Michael Ellerman [off-list ref] wrote:

"Naveen N. Rao" [off-list ref] writes:

quoted

Michael Ellerman wrote:

quoted

Daniel Borkmann [off-list ref] writes:

quoted

On 10/25/21 8:15 AM, Naveen N. Rao wrote:

quoted

Hari Bathini wrote:

quoted

Running program with bpf-to-bpf function calls results in data access
exception (0x300) with the below call trace:

    [c000000000113f28] bpf_int_jit_compile+0x238/0x750 (unreliable)
    [c00000000037d2f8] bpf_check+0x2008/0x2710
    [c000000000360050] bpf_prog_load+0xb00/0x13a0
    [c000000000361d94] __sys_bpf+0x6f4/0x27c0
    [c000000000363f0c] sys_bpf+0x2c/0x40
    [c000000000032434] system_call_exception+0x164/0x330
    [c00000000000c1e8] system_call_vectored_common+0xe8/0x278

as bpf_int_jit_compile() tries writing to write protected JIT code
location during the extra pass.

Fix it by holding off write protection of JIT code until the extra
pass, where branch target addresses fixup happens.

Cc: stable@vger.kernel.org
Fixes: 62e3d4210ac9 ("powerpc/bpf: Write protect JIT code")
Signed-off-by: Hari Bathini <hbathini@linux.ibm.com>
---
 arch/powerpc/net/bpf_jit_comp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Thanks for the fix!

Reviewed-by: Naveen N. Rao <redacted>

LGTM, I presume this fix will be routed via Michael.

Thanks for reviewing, I've picked it up.

quoted

BPF selftests have plenty of BPF-to-BPF calls in there, too bad this was
caught so late. :/

Yeah :/

STRICT_KERNEL_RWX is not on by default in all our defconfigs, so that's
probably why no one caught it.

Yeah, sorry - we should have caught this sooner.

quoted

I used to run the BPF selftests but they stopped building for me a while
back, I'll see if I can get them going again.

Ravi had started looking into getting the selftests working well before
he left. I will take a look at this.

Thanks.

I got them building with something like:

 - turning on DEBUG_INFO and DEBUG_INFO_BTF and rebuilding vmlinux
 - grabbing clang 13 from:
   https://github.com/llvm/llvm-project/releases/download/llvmorg-13.0.0/clang+llvm-13.0.0-powerpc64le-linux-ubuntu-18.04.tar.xz
 - PATH=$HOME/clang+llvm-13.0.0-powerpc64le-linux-ubuntu-18.04/bin/:$PATH
 - apt install:
   - libelf-dev
   - dwarves
   - python-docutils
   - libcap-dev


The DEBUG_INFO requirement is a bit of a pain for me. I generally don't

We do need DWARF to be present during BTF generation. We don't really
need to preserve DWARF after BTF is generated, though. But no one
added that config option and corresponding optimization. If you can
figure out how to do that, I'm sure a bunch of folks will appreciate
being able to specify CONFIG_DEBUG_INFO_BTF without CONFIG_DEBUG_INFO
dependency.

build with that enabled, because the resulting kernels are stupidly
large. I'm not sure if that's a hard requirement, or if the vmlinux has
to match the running kernel exactly?

There is logic in tools/testing/bpf/Makefile to use VMLINUX_H instead of
extracting the BTF from the vmlinux (line 247), but AFAICS that's
unreachable since 1a3449c19407 ("selftests/bpf: Clarify build error if
no vmlinux"), which makes it a hard error to not have a VMLINUX_BTF.

Yeah, you can pass pre-generated vmlinux.h through VMLINUX_H, which we
do for libbpf CI (see [0]) when running latest selftests against old
kernels (we test 4.9 and 5.5 currently). Latest vmlinux image (which
you can override with VMLINUX_BTF) is required for custom kernel
module which we use during selftests. But if you don't provide the
matching kernel, everything should still build fine, the test module
won't load properly and we'll skip a few tests. You still should get a
good coverage.

So in short, given we are able to build selftests and run it against
4.9 and 5.5, you should be able to as well.

  [0] https://github.com/libbpf/libbpf/blob/master/travis-ci/vmtest/build_selftests.sh#L29-L30

cheers