Re: [PATCH bpf-next v4 0/3] Annotate kfuncs in .BTF_ids section

From: Viktor Malik
Date: Mon Feb 05 2024 - 13:43:23 EST


On 2/3/24 19:45, Manu Bretelle wrote:
> On Sat, Feb 03, 2024 at 03:40:24PM +0100, Jiri Olsa wrote:
>> On Fri, Feb 02, 2024 at 03:09:05PM -0800, Manu Bretelle wrote:
>>> On Sun, Jan 28, 2024 at 06:24:05PM -0700, Daniel Xu wrote:
>>>> === Description ===
>>>>
>>>> This is a bpf-treewide change that annotates all kfuncs as such inside
>>>> .BTF_ids. This annotation eventually allows us to automatically generate
>>>> kfunc prototypes from bpftool.
>>>>
>>>> We store this metadata inside a yet-unused flags field inside struct
>>>> btf_id_set8 (thanks Kumar!). pahole will be taught where to look.
>>>>
>>>> More details about the full chain of events are available in commit 3's
>>>> description.
>>>>
>>>> The accompanying pahole and bpftool changes can be viewed
>>>> here on these "frozen" branches [0][1].
>>>>
>>>> [0]: https://github.com/danobi/pahole/tree/kfunc_btf-v3-mailed
>>>> [1]: https://github.com/danobi/linux/tree/kfunc_bpftool-mailed
>>>
>>>
>>> I hit a similar issue to [0] on master
>>> 943b043aeecc ("selftests/bpf: Fix bench runner SIGSEGV")
>>> when cross-compiling on x86_64 (LE) to s390x (BE).
>>> I do have CONFIG_DEBUG_INFO_BTF enable and the issue would not trigger if
>>> I disabled CONFIG_DEBUG_INFO_BTF (and with the fix mentioned in [0]).
>>>
>>> What seems to happen is that `tools/resolve_btfids` is ran in the context of the
>>> host endianess and if I printk before the WARN_ON:
>>> diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
>>> index ef380e546952..a9ed7a1a4936 100644
>>> --- a/kernel/bpf/btf.c
>>> +++ b/kernel/bpf/btf.c
>>> @@ -8128,6 +8128,7 @@ int register_btf_kfunc_id_set(enum bpf_prog_type prog_type,
>>> * WARN() for initcall registrations that do not check errors.
>>> */
>>> if (!(kset->set->flags & BTF_SET8_KFUNCS)) {
>>> + printk("Flag 0x%08X, expected 0x%08X\n", kset->set->flags, BTF_SET8_KFUNCS);
>>> WARN_ON(!kset->owner);
>>> return -EINVAL;
>>> }
>>>
>>> the boot logs would show:
>>> Flag 0x01000000, expected 0x00000001
>>>
>>> The issue did not happen prior to
>>> 6f3189f38a3e ("bpf: treewide: Annotate BPF kfuncs in BTF")
>>> has only 0 was written before.
>>>
>>> It seems [1] will be addressing cross-compilation, but it did not fix it as is
>>> by just applying on top of master, so probably some of the changes will also need
>>> to be ported to `tools/include/linux/btf_ids.h`?
>>
>> the fix in [1] is fixing flags in set8's pairs, but not the global flags
>>
>> it looks like Viktor's fix should now also swap that as well? like in the
>> change below on top of Viktor's changes (untested)
>>
>> jirka
>>
>>
>> ---
>> diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c
>> index d01603ef6283..c44d57fec390 100644
>> --- a/tools/bpf/resolve_btfids/main.c
>> +++ b/tools/bpf/resolve_btfids/main.c
>> @@ -706,6 +706,8 @@ static int sets_patch(struct object *obj)
>> * correctly translate everything.
>> */
>> if (need_bswap) {
>> + set8->flags = bswap_32(set8->flags);
>> +
>> for (i = 0; i < cnt; i++) {
>> set8->pairs[i].flags =
>> bswap_32(set8->pairs[i].flags);
>>
>
> That should work. Here are a few tests I ran:
>
> $ md5sum /tmp/kbuild-s390x/vmlinux.*
> eb658e51e089f3c5b2c8909a29dc9997 /tmp/kbuild-s390x/vmlinux.a
> # plain vmlinux before running resolv_btfids (all 0s)
> ea907cd46a1a73b8276b5f2a82af00ca /tmp/kbuild-s390x/vmlinux.before_resolv
> # x86_64 resolv_btfids on master without Viktor's patch
> 980a40c3a3ff563d1c2d1ebdd5071a23 /tmp/kbuild-s390x/vmlinux.resolv_native
> # x86_64 resolv_btfids on master with Viktor's patch
> b986d19e242719ebea41c578235da662 /tmp/kbuild-s390x/vmlinux.resolv_native_patch_viktor
> # x86_64 resolv_btfids on master with Viktor's patch and your suggested patch
> 4edd8752ff01129945bd442689b1927b /tmp/kbuild-s390x/vmlinux.resolv_native_patch_viktor_patched
> # s390x resolv_btfids run with qemu-s390x-static
> 4edd8752ff01129945bd442689b1927b /tmp/kbuild-s390x/vmlinux.resolv_s390x
>
>
> and some hexdiff of those binaries:
>
>
> # difference between master's native build and s390x build.... has byte swapping for set8 and others
> diff -ruN <(xxd /tmp/kbuild-s390x/vmlinux.resolv_s390x) <(xxd /tmp/kbuild-s390x/vmlinux.resolv_native) > diff_s390x_native.diff
> https://gist.github.com/chantra/c3d58637a08a6f7340953dc155bb18cc
>
> # difference betwee Viktor's version and s390x build.... squinting my eyes I only see the global set8 is missing
> diff -ruN <(xxd /tmp/kbuild-s390x/vmlinux.resolv_s390x) <(xxd /tmp/kbuild-s390x/vmlinux.resolv_native_patch_viktor) > diff_s390x_native_viktor.diff
> https://gist.github.com/chantra/61cfff02b456ae72d3c0161ce1897097

Thanks for the testing Manu!

Jiri's suggested fix is now a part of [1].

Viktor

[1] https://lore.kernel.org/bpf/cover.1707157553.git.vmalik@xxxxxxxxxx/

>
> Have a good weekend all!
>
> Manu
>