Re: [PATCH RFC bpf-next 1/4] bpf: add struct largest member size in func model

[Xu Kuohai]

On 4/17/2025 3:14 PM, Alexis Lothoré wrote:
> Hi Andrii,
>
> On Wed Apr 16, 2025 at 11:24 PM CEST, Andrii Nakryiko wrote:
> > On Fri, Apr 11, 2025 at 1:32 PM Alexis Lothoré (eBPF Foundation)
> > <alexis.lothore@xxxxxxxxxxx> wrote:
> > > In order to properly JIT the trampolines needed to attach BPF programs
> > > to functions, some architectures like ARM64 need to know about the
> > > alignment needed for the function arguments. Such alignment can
> > > generally be deduced from the argument size, but that's not completely
> > > true for composite types. In the specific case of ARM64, the AAPCS64 ABI
> > > defines that a composite type which needs to be passed through stack
> > > must be aligned on the maximum between 8 and the largest alignment
> > > constraint of its first-level members. So the JIT compiler needs more
> > > information about the arguments to make sure to generate code that
> > > respects those alignment constraints.
> > >
> > > For struct arguments, add information about the size of the largest
> > > first-level member in the struct btf_func_model to allow the JIT
> > > compiler to guess the needed alignment. The information is quite
> >
> > I might be missing something, but how can the *size* of the field be
> > used to calculate that argument's *alignment*? i.e., I don't
> > understand why arg_largest_member_size needs to be calculated instead
> > of arg_largest_member_alignment...
>
> Indeed I initially checked whether I could return directly some alignment
> info from btf, but it then involves the alignment computation in the btf
> module. Since there could be minor differences between architectures about
> alignment requirements, I though it would be better to in fact keep alignment
> computation out of the btf module. For example, I see that 128 bits values
> are aligned on 16 bytes on ARM64, while being aligned on 8 bytes on S390.
>
> And since for ARM64, all needed alignments are somehow derived from size
> (it is either directly size for fundamental types, or alignment of the
> largest member for structs, which is then size of largest member),
> returning the size seems to be enough to allow the JIT side to compute
> alignments.

Not exactly. The compiler's "packed" and "alignment" attributes cause a
structure to be aligned differently from its natural alignment.

For example, with the following three structures:

struct s0 {
    __int128 x;
};

struct s1 {
    __int128 x;
} __attribute__((packed));

struct s2 {
    __int128 x;
} __attribute__((aligned(64)));

Even though the largest member size is the same, s0 will be aligned to 16
bytes, s1 and s2 are not aligned the same way. s1 has no alignment due to
the "packed" attribute, while s2 will be aligned to 64 bytes.

When these three structures are passed as function arguments, they will be
located on different positions on the stack.

For the following three functions:

int f0(__int128 a, __int128 b, __int128 c, int64_t d, __int128 e, int64_t f, struct s0 g);
int f1(__int128 a, __int128 b, __int128 c, int64_t d, __int128 e, int64_t f, struct s1 g);
int f2(__int128 a, __int128 b, __int128 c, int64_t d, __int128 e, int64_t f, struct s2 g);

g will be located at sp+32 in f0, sp + 24 in f1, and some 64-byte aligned
stack address in f2.

> > > specific, but it allows to keep arch-specific concerns (ie: guessing the
> > > final needed alignment for an argument) isolated in each JIT compiler.
> >
> > couldn't all this information be calculated in the JIT compiler (if
> > JIT needs that) from BTF?
>
> > From what I understand, the JIT compiler does not have access to BTF info,
> only a substract from it, arranged in a struct btf_func_model ? This
> struct btf_func_model already has size info for standard types, but for
> structs we need some additional info about the members, hence this
> arg_largest_member_alignment addition in btf_func_model.
>
> Thanks,
>
> Alexis