[PATCH v2 2/2] x86/cfi,bpf: Fix BPF JIT call

[Peter Zijlstra]

On Wed, Dec 06, 2023 at 01:39:43PM -0800, Alexei Starovoitov wrote:


> All is ok until kCFI comes into picture.
> Here we probably need to teach arch_prepare_bpf_trampoline() to emit
> different __kcfi_typeid depending on kernel function proto,
> so that caller hash checking logic won't be tripped.
> I suspect that requires to reverse engineer an algorithm of computing kcfi from clang.
> other ideas?

I was going to try and extend bpf_struct_ops with a pointer, this
pointer will point to a struct of the right type with all ops filled out
as stubs.

Then I was going to have bpf_struct_ops_map_update_elem() pass a pointer
to the stub op (using moff) into bpf_struct_ops_prepare_trampoline() and
eventually arch_prepare_bpf_trampoline().

Additionally I was going to add BPF_TRAMP_F_INDIRECT.

Then when F_INDIRECT is set, have it generate the CFI preamble based on
the stub passed -- which will have the correct preamble for that method.

At least, that's what I'm thinking now, I've yet to try and implement
it.

> > > The other case:

> In the case of bpf_for_each_map_elem() the 'bloom_callback' is a subprog
> of bpf_callback_t type.
> So the kernel is doing:
>                 ret = callback_fn((u64)(long)map, (u64)(long)&key,
>                                   (u64)(long)val, (u64)(long)callback_ctx, 0);
> and that works on all archs including 32-bit.
> The kernel is doing conversion from native calling convention to bpf calling convention
> and for lucky archs like x86-64 the conversion is a true nop.
> It's a plain indirect call to JITed bpf prog.
> Note there is no interpreter support here. This works on archs with JITs only.
> No ftrace and no trampoline.
> 
> This case is easier to make work with kCFI.
> The JIT will use:
> cfi_bpf_hash:
>       .long   __kcfi_typeid___bpf_prog_runX  
> like your patch already does.
> And will use
> extern u64 __bpf_callback_fn(u64, u64, u64, u64, u64);
> cfi_bpf_subprog_hash:
>       .long   __kcfi_typeid___bpf_callback_fn
> to JIT all subprogs. See bpf_is_subprog().

Aaah!, yes it should be trivial to use another hash value when
is_subprog in emit_prologue().

> btw there are two patchsets in progress that will touch core bits of JITs.
> This one:
> https://patchwork.kernel.org/project/netdevbpf/cover/20231201190654.1233153-1-song@kernel.org/
> and this one:
> https://patchwork.kernel.org/project/netdevbpf/cover/20231011152725.95895-1-hffilwlqm@gmail.com/
> 
> so do you mind resending your current set with get_cfi_offset() change and
> I can land it into bpf-next, so we can fix one bug at a time,
> build on top, and avoid conflicts?

I can do.

> The more we dig the more it looks like that the follow up you planned to do
> on top of this set isn't going to happen soon.
> So should be ok going through bpf-next and then you can follow up with x86 things
> after merge window?

Yes, we can do that. Plans have changed on my side too -- I'm taking a 6
week break soon, so I'll do whatever I can before I'm out, and then
continue from whatever state I find when I get back.


Thanks for the details!