Re: [PATCH] Revert "bpf: Take return from set_memory_rox() into account with bpf_jit_binary_lock_ro()" for linux-6.6.37

[LEROY Christophe]

Le 08/07/2024 à 14:36, Greg KH a écrit :
> On Sun, Jul 07, 2024 at 03:34:15PM +0800, WangYuli wrote:
>>
>> On 2024/7/6 17:30, Greg KH wrote:
>>> This makes it sound like you are reverting this because of a build
>>> error, which is not the case here, right?  Isn't this because of the
>>> powerpc issue reported here:
>>>     https://lore.kernel.org/r/20240705203413.wbv2nw3747vjeibk@xxxxxxxxxxxx
>>> ?
>>
>> No, it only occurs on ARM64 architecture. The reason is that before being
>> modified, the function
>>
>> bpf_jit_binary_lock_ro() in arch/arm64/net/bpf_jit_comp.c +1651
>>
>> was introduced with __must_check, which is defined as
>> __attribute__((__warn_unused_result__)).
>>
>>
>> However, at this point, calling bpf_jit_binary_lock_ro(header)
>> coincidentally results in an unused-result
>>
>> warning.
>
> Ok, thanks, but why is no one else seeing this in their testing?

Probably the configs used by robots do not activate BPF JIT ?

>
>>> If not, why not just backport the single missing arm64 commit,
>>
>> Upstream commit 1dad391daef1 ("bpf, arm64: use bpf_prog_pack for memory
>> management") is part of
>>
>> a larger change that involves multiple commits. It's not an isolated commit.
>>
>>
>> We could certainly backport all of them to solve this problem, buthas it's not
>> the simplest solution.
>
> reverting the change feels wrong in that you will still have the bug
> present that it was trying to solve, right?  If so, can you then provide
> a working version?

Indeed, by reverting the change you "punish" all architectures because
arm64 hasn't properly been backported, is it fair ?

In fact, when I implemented commit e60adf513275 ("bpf: Take return from
set_memory_rox() into account with bpf_jit_binary_lock_ro()"), we had
the following users for function bpf_jit_binary_lock_ro() :

$ git grep bpf_jit_binary_lock_ro e60adf513275~
e60adf513275~:arch/arm/net/bpf_jit_32.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:arch/loongarch/net/bpf_jit.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:arch/mips/net/bpf_jit_comp.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:arch/parisc/net/bpf_jit_core.c:
bpf_jit_binary_lock_ro(jit_data->header);
e60adf513275~:arch/s390/net/bpf_jit_comp.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:arch/sparc/net/bpf_jit_comp_64.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:arch/x86/net/bpf_jit_comp32.c:
bpf_jit_binary_lock_ro(header);
e60adf513275~:include/linux/filter.h:static inline void
bpf_jit_binary_lock_ro(struct bpf_binary_header *hdr)

But when commit 08f6c05feb1d ("bpf: Take return from set_memory_rox()
into account with bpf_jit_binary_lock_ro()") was applied, we had one
more user which is arm64:

$ git grep bpf_jit_binary_lock_ro 08f6c05feb1d~
08f6c05feb1d~:arch/arm/net/bpf_jit_32.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/arm64/net/bpf_jit_comp.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/loongarch/net/bpf_jit.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/mips/net/bpf_jit_comp.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/parisc/net/bpf_jit_core.c:
bpf_jit_binary_lock_ro(jit_data->header);
08f6c05feb1d~:arch/s390/net/bpf_jit_comp.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/sparc/net/bpf_jit_comp_64.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:arch/x86/net/bpf_jit_comp32.c:
bpf_jit_binary_lock_ro(header);
08f6c05feb1d~:include/linux/filter.h:static inline void
bpf_jit_binary_lock_ro(struct bpf_binary_header *hdr)

Therefore, commit 08f6c05feb1d should have included a backport for arm64.

So yes, I agree with Greg, the correct fix should be to backport to
ARM64 the changes done on other architectures in order to properly
handle return of set_memory_rox() in bpf_jit_binary_lock_ro().

Christophe