[PATCH 2/6] powerpc64/bpf: Tailcall handling with trampolines

[bot+bpf-ci at kernel.org]

> diff --git a/arch/powerpc/net/bpf_jit_comp.c b/arch/powerpc/net/bpf_jit_comp.c
> index 069a8822c30d..4aaa0a287a45 100644
> --- a/arch/powerpc/net/bpf_jit_comp.c
> +++ b/arch/powerpc/net/bpf_jit_comp.c

[ ... ]

> -static void bpf_trampoline_setup_tail_call_cnt(u32 *image, struct codegen_context *ctx,
> -					       int func_frame_offset, int r4_off)
> +static void bpf_trampoline_setup_tail_call_info(u32 *image, struct codegen_context *ctx,
> +					       int func_frame_offset,
> +					       int bpf_dummy_frame_size, int r4_off)
>  {
>  	if (IS_ENABLED(CONFIG_PPC64)) {
>  		/* See bpf_jit_stack_tailcallinfo_offset() */
> -		int tailcallcnt_offset = 7 * 8;
> +		int tailcallinfo_offset = BPF_PPC_STACK_SAVE + SZL;

[ ... ]

> +		EMIT(PPC_RAW_STL(_R3, _R1, func_frame_offset
> +				- bpf_dummy_frame_size - tailcallinfo_offset));
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

The new setup function stores tail_call_info at offset
(func_frame_offset - bpf_dummy_frame_size - tailcallinfo_offset) from R1.
Since func_frame_offset = bpf_dummy_frame_size + bpf_frame_size, this
resolves to (bpf_frame_size - 56), which is a positive offset from R1.

However, the corresponding bpf_trampoline_restore_tail_call_cnt() function
was not updated. It still reads from (R1 - 56):

    static void bpf_trampoline_restore_tail_call_cnt(...)
    {
        if (IS_ENABLED(CONFIG_PPC64)) {
            int tailcallcnt_offset = 7 * 8;
            EMIT(PPC_RAW_LL(_R3, _R1, -tailcallcnt_offset));
                                      ^^^^^^^^^^^^^^^^^^
            ...
        }
    }

Is there an offset mismatch here? The setup function stores at
R1 + (bpf_frame_size - 56) while restore reads from R1 - 56. These point
to different memory locations since bpf_frame_size is a positive value
(at minimum ~200 bytes given the frame contents).

When BPF_TRAMP_F_TAIL_CALL_CTX is set, the restore function would read
incorrect data after the traced function returns, potentially causing
wrong tail call context to be propagated.

> @@ -801,6 +830,12 @@ static int __arch_prepare_bpf_trampoline(...)
>  		}
>  	}
>
> +	/* Room for 64-bit tail_call_cnt */
> +	bpf_frame_size += SZL;
> +
> +	/* Room for nvr save area */
> +	bpf_frame_size += BPF_PPC_STACK_SAVE;
> +
>  	/* Padding to align stack frame, if any */
>  	bpf_frame_size = round_up(bpf_frame_size, SZL * 2);

The frame size is increased by 56 bytes (SZL + BPF_PPC_STACK_SAVE = 8 + 48)
for the tail_call_info and nvr save area. This affects all calculations
using bpf_frame_size, including the store offset in the setup function,
but the restore function continues to use a hardcoded negative offset.


---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md

CI run summary: https://github.com/kernel-patches/bpf/actions/runs/20713322833