Re: [PATCH bpf-next] libbpf: Add LoongArch support to bpf_tracing.h

[Hengqi Chen <hengqi.chen@gmail.com>]

On 2022/12/14 06:09, Andrii Nakryiko wrote:
> On Mon, Dec 12, 2022 at 4:17 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
>>
>> On Mon, Dec 12, 2022 at 1:11 AM Hengqi Chen <hengqi.chen@gmail.com> wrote:
>>>
>>> Add PT_REGS macros for LoongArch64.
>>>
>>> Signed-off-by: Hengqi Chen <hengqi.chen@gmail.com>
>>> ---
>>>  tools/lib/bpf/bpf_tracing.h | 21 +++++++++++++++++++++
>>>  1 file changed, 21 insertions(+)
>>>
>>> diff --git a/tools/lib/bpf/bpf_tracing.h b/tools/lib/bpf/bpf_tracing.h
>>> index 2972dc25ff72..2d7da1caa961 100644
>>> --- a/tools/lib/bpf/bpf_tracing.h
>>> +++ b/tools/lib/bpf/bpf_tracing.h
>>> @@ -32,6 +32,9 @@
>>>  #elif defined(__TARGET_ARCH_arc)
>>>         #define bpf_target_arc
>>>         #define bpf_target_defined
>>> +#elif defined(__TARGET_ARCH_loongarch)
>>> +       #define bpf_target_loongarch
>>> +       #define bpf_target_defined
>>>  #else
>>>
>>>  /* Fall back to what the compiler says */
>>> @@ -62,6 +65,9 @@
>>>  #elif defined(__arc__)
>>>         #define bpf_target_arc
>>>         #define bpf_target_defined
>>> +#elif defined(__loongarch__) && __loongarch_grlen == 64
>>> +       #define bpf_target_loongarch
>>> +       #define bpf_target_defined
>>>  #endif /* no compiler target */
>>>
>>>  #endif
>>> @@ -258,6 +264,21 @@ struct pt_regs___arm64 {
>>>  /* arc does not select ARCH_HAS_SYSCALL_WRAPPER. */
>>>  #define PT_REGS_SYSCALL_REGS(ctx) ctx
>>>
>>> +#elif defined(bpf_target_loongarch)
>>> +
>>> +#define __PT_PARM1_REG regs[5]
>>> +#define __PT_PARM2_REG regs[6]
>>> +#define __PT_PARM3_REG regs[7]
>>> +#define __PT_PARM4_REG regs[8]
>>> +#define __PT_PARM5_REG regs[9]
>>> +#define __PT_RET_REG regs[1]
>>> +#define __PT_FP_REG regs[22]
>>> +#define __PT_RC_REG regs[4]
>>> +#define __PT_SP_REG regs[3]
>>> +#define __PT_IP_REG csr_era
>>> +/* loongarch does not select ARCH_HAS_SYSCALL_WRAPPER. */
>>> +#define PT_REGS_SYSCALL_REGS(ctx) ctx
>>
>> Is there some online documentation explaining this architecture's
>> calling conventions? It would be useful to include that as a comment
>> to be able to refer back to it. On a related note, are there any
>> syscall specific calling convention differences, similar to
>> PT_REGS_PARM1_SYSCALL for arm64 or PT_REGS_PARM4_SYSCALL for x86-64?
>>
> 
> Ok, I think [0] would be a good resource, please add a link to it in
> the comment. But also it seems like PARM1-5 should map to regs[6]
> through regs[10] (not regs[5] - regs[9] that you have here). And BTW,
> seems like architecture supports passing more than five, PARM6 would
> be regs[11]. I've been wanting to add 6th+ argument to libbpf macros'
> for a while (it came up in x86-64 world for uprobes as well), so if
> you have cycles, please consider helping with that as well.
> 

I've seen this on GitHub. Let me have a try.

> Also I see orig_a0 in struct pt_regs, which seems suspiciously similar
> to arm64's PT_REGS_PARM1_SYSCALL's use of orig_x0, please check about
> that as well. As I said, syscalls usually have some additional quirks.
> 
> 
>   [0] https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
> 
> 
>>> +
>>>  #endif
>>>
>>>  #if defined(bpf_target_defined)
>>> --
>>> 2.31.1

Thanks, Andrii.

After some investigation, I do find some quirks on syscalls. Will update this patch.