Le 24/11/2022 à 14:49, Naveen N. Rao a écrit :
> Christophe Leroy wrote:
>>
>>
>> Le 24/11/2022 à 11:13, Naveen N. Rao a écrit :
>>> Christophe Leroy wrote:
>>>> ldimm64 is not only used for loading function addresses, and
>>>
>>> That's probably true today, but I worry that that can change upstream
>>> and we may not notice at all.
>>
>> Not sure what you mean.
>>
>> Today POWERPC considers that ldimm64 is _always_ loading a function
>> address whereas upstream BPF considers that ldimm64 is a function only
>> when it is flagged BPF_PSEUDO_FUNC.
>
> Not sure why you think we consider ldimm64 to always be loading a
> function address. Perhaps it is due to the poorly chosen variable name
> func_addr in bpf_jit_fixup_addresses(), or due to the fact that we
> always update the JIT code for ldimm64. In any case, we simply overwrite
> imm64 load instructions to ensure we are using the updated address.
Well that's the padding which make me think that. When ldimm64 is used
with immediate value, it won't change from one pass to the other. We
have the need for the padding only because it may contain addresses that
will change from one pass to another.
>
>>
>> In what direction could that change in the future ?
>>
>> For me if they change that it becomes an API change.
>
> More of an extension, which is exactly what we had when BPF_PSEUDO_FUNC
> was introduced. Took us nearly a year before we noticed.
>
> Because we do not do a full JIT during the extra pass today like other
> architectures, we are the exception - there is always the risk of bpf
> core changes breaking our JIT. So, I still think it is better if we do a
> full JIT during extra pass.
>
I like the idea of a full JIT during extra passes and will start looking
at it.
Will it also allow us to revert your commit fab07611fb2e
("powerpc32/bpf: Fix codegen for bpf-to-bpf calls") ?
Thanks
Christophe
