Re: [PATCH bpf v3 2/2] bpf_trace: bail out from bpf_kprobe_multi_link_attach when in compat

[Yonghong Song <yhs@fb.com>]

On 5/18/22 1:03 PM, Eugene Syromiatnikov wrote:
> On Wed, May 18, 2022 at 09:55:05AM -0700, Yonghong Song wrote:
>>
>>
>> On 5/18/22 5:22 AM, Eugene Syromiatnikov wrote:
>>> Since bpf_kprobe_multi_link_attach doesn't support 32-bit kernels
>>> for whatever reason, having it enabled for compat processes on 64-bit
>>> kernels makes even less sense due to discrepances in the type sizes
>>> that it does not handle.
>>
>> If I understand correctly, the reason is due to
>> in libbpf we have
>> struct bpf_link_create_opts {
>>          size_t sz; /* size of this struct for forward/backward compatibility
>> */
>>          __u32 flags;
>>          union bpf_iter_link_info *iter_info;
>>          __u32 iter_info_len;
>>          __u32 target_btf_id;
>>          union {
>>                  struct {
>>                          __u64 bpf_cookie;
>>                  } perf_event;
>>                  struct {
>>                          __u32 flags;
>>                          __u32 cnt;
>>                          const char **syms;
>>                          const unsigned long *addrs;
>>                          const __u64 *cookies;
>>                  } kprobe_multi;
>>          };
>>          size_t :0;
>> };
>>
>> Note that we have `const unsigned long *addrs;`
>>
>> If we have 32-bit user space application and 64bit kernel,
>> and we will have userspace 32-bit pointers and kernel as
>> 64bit pointers and current kernel doesn't handle 32-bit
>> user pointer properly.
>>
>> Consider this may involve libbpf uapi change, maybe
>> we should change "const unsigned long *addrs;" to
>> "const __u64 *addrs;" considering we haven't freeze
>> libbpf UAPI yet.
>>
>> Otherwise, we stick to current code with this patch,
>> it will make it difficult to support 32-bit app with
>> 64-bit kernel for kprobe_multi in the future due to
>> uapi issues.
>>
>> WDYT?
> 
> As 32 bit arches are "unsupported" currently, the change would be more
> a semantic one rather then practical;  I don't mind having it here (basically,
> the tools/* part of [1]), though (assuming it is still possible to get it
> in 5.18).
> 
> [1] https://lore.kernel.org/lkml/6ef675aeeea442fa8fc168cd1cb4e4e474f65a3f.1652772731.git.esyr@redhat.com/

I think for patch [1], we only need libbpf and selftest change, no
kernel change is needed since we
explicitly does not support 32bit kernel in the
beginning of function bpf_kprobe_multi_link_attach():

         /* no support for 32bit archs yet */
         if (sizeof(u64) != sizeof(void *))
                 return -EOPNOTSUPP;

and in kernel, address (pointer) size will be considered
long (64bit) which is exactly the libbpf change did that.

> 
>>>
>>> Fixes: 0dcac272540613d4 ("bpf: Add multi kprobe link")
>>> Signed-off-by: Eugene Syromiatnikov <esyr@redhat.com>
>>> ---
>>>   kernel/trace/bpf_trace.c | 2 +-
>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index 212faa4..2f83489 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -2412,7 +2412,7 @@ int bpf_kprobe_multi_link_attach(const union bpf_attr *attr, struct bpf_prog *pr
>>>   	int err;
>>>   	/* no support for 32bit archs yet */
>>> -	if (sizeof(u64) != sizeof(void *))
>>> +	if (sizeof(u64) != sizeof(void *) || in_compat_syscall())
>>>   		return -EOPNOTSUPP;
>>>   	if (prog->expected_attach_type != BPF_TRACE_KPROBE_MULTI)
>>
>