Re: [PATCH bpf-next v2] arm64, bpf: Add 12-argument support for bpf trampoline

From: Puranjay Mohan
Date: Mon Jul 08 2024 - 05:01:24 EST


Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> writes:

> On 7/5/2024 8:53 PM, Puranjay Mohan wrote:
>> The arm64 bpf JIT currently supports attaching the trampoline to
>> functions with <= 8 arguments. This is because up to 8 arguments can be
>> passed in registers r0-r7. If there are more than 8 arguments then the
>> 9th and later arguments are passed on the stack, with SP pointing to the
>> first stacked argument. See aapcs64[1] for more details.
>>
>> If the 8th argument is a structure of size > 8B, then it is passed fully
>> on stack and r7 is not used for passing any argument. If there is a 9th
>> argument, it will be passed on the stack, even though r7 is available.
>>
>> Add the support of storing and restoring arguments passed on the stack
>> to the arm64 bpf trampoline. This will allow attaching the trampoline to
>> functions that take up to 12 arguments.
>>
>> [1] https://github.com/ARM-software/abi-aa/blob/main/aapcs64/aapcs64.rst#parameter-passing
>>
>> Signed-off-by: Puranjay Mohan <puranjay@xxxxxxxxxx>
>> ---
>> Changes in V1 -> V2:
>> V1: https://lore.kernel.org/all/20240704173227.130491-1-puranjay@xxxxxxxxxx/
>> - Fixed the argument handling for composite types (structs)
>> ---
>> arch/arm64/net/bpf_jit_comp.c | 139 ++++++++++++++-----
>> tools/testing/selftests/bpf/DENYLIST.aarch64 | 3 -
>> 2 files changed, 107 insertions(+), 35 deletions(-)
>>

[SNIP]

>> fill_link_info/kprobe_multi_invalid_ubuff # bpf_program__attach_kprobe_multi_opts unexpected error: -95
>
> It looks like this patch, similar to [1], also does not handle
> parameter alignment properly [2].
>
> For example:
>
> int func(__int128 a, __int128 b, __int128 c, int64_t d, __int128 e, int64_t f, __int128 g)
> {
> }
>
> parameter a~d are passed through register x0~x6, while parameter
> e~g are passed through stack. Since __int128 is 16-byte aligned,
> parameter e, f, and g should be placed at sp + 0, +16, and +32
> respectively, with 8 bytes **padding** between f and g.
>
>
> And the compiler's alignment or packed attribute may make things
> worse, causing parameters to be placed on the stack at positions
> that are not naturally aligned.

Hi Xu,
Thanks for explaining this. I was not aware that you have already sent a
patch[1] to add this support to arm64.

So, I see that it will be non-trivial to calculate padding for each
argument passed on the stack. If you are not planning to work on this
then I can try to implement it.

Alsoi, do we currently have a selftest that checks for this edge case? if
not I can try to add that too.

Thanks,
Puranjay

> [1] https://lore.kernel.org/bpf/20230917150752.69612-1-xukuohai@xxxxxxxxxxxxxxx/
> [2] https://lore.kernel.org/bpf/CABRcYmLtk8aQEzoUFw+j5Rdd-MXf-q+i7RHXZtu-skjRz11ZDw@xxxxxxxxxxxxxx/

Attachment: signature.asc
Description: PGP signature