Re: [PATCH bpf 2/7] bpf/x86: prevent trampoline attachment when args location on stack is uncertain

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Jun 13, 2025 at 10:26:37AM +0200, Alexis Lothoré wrote:
> Hi Peter,
> 
> On Fri Jun 13, 2025 at 10:11 AM CEST, Peter Zijlstra wrote:
> > On Fri, Jun 13, 2025 at 09:37:11AM +0200, Alexis Lothoré (eBPF Foundation) wrote:
> >> When the target function receives more arguments than available
> >> registers, the additional arguments are passed on stack, and so the
> >> generated trampoline needs to read those to prepare the bpf context,
> >> but also to prepare the target function stack when it is in charge of
> >> calling it. This works well for scalar types, but if the value is a
> >> struct, we can not know for sure the exact struct location, as it may
> >> have been packed or manually aligned to a greater value.
> >
> > https://refspecs.linuxbase.org/elf/x86_64-abi-0.99.pdf
> >
> > Has fairly clear rules on how arguments are encoded. Broadly speaking
> > for the kernel, if the structure exceeds 2 registers in size, it is
> > passed as a reference, otherwise it is passed as two registers.
> 
> Maybe my commit wording is not precise enough, but indeed, there's not
> doubt about whether the struct value is passed on the stack or through a
> register/a pair of registers. The doubt is rather about the struct location
> when it is passed _by value_ and _on the stack_: the ABI indeed clearly
> states that "Structures and unions assume the alignment of their most
> strictly aligned component" (p.13), but this rule is "silently broken" when
> a struct has an __attribute__((packed)) or and __attribute__((aligned(X))),
> and AFAICT this case can not be detected at runtime with current BTF info.

Ah, okay. So it is a failure of BTF. That was indeed not clear.