On Mon, Sep 28, 2020 at 7:39 AM Daniel Borkmann [off-list ref] wrote:

Port of tail_call_static() helper function from Cilium's BPF code base [0]
to libbpf, so others can easily consume it as well. We've been using this
in production code for some time now. The main idea is that we guarantee
that the kernel's BPF infrastructure and JIT (here: x86_64) can patch the
JITed BPF insns with direct jumps instead of having to fall back to using
expensive retpolines. By using inline asm, we guarantee that the compiler
won't merge the call from different paths with potentially different
content of r2/r3.

We're also using Cilium's __throw_build_bug() macro (here as: __bpf_unreachable())
in different places as a neat trick to trigger compilation errors when
compiler does not remove code at compilation time. This works for the BPF
back end as it does not implement the __builtin_trap().

  [0] https://github.com/cilium/cilium/commit/f5537c26020d5297b70936c6b7d03a1e412a1035

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Andrii Nakryiko <redacted>
---

few optional nits below, but looks good to me:

Acked-by: Andrii Nakryiko <redacted>

 tools/lib/bpf/bpf_helpers.h | 46 +++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

[...]

+/*
+ * Helper function to perform a tail call with a constant/immediate map slot.
+ */
+static __always_inline void
+bpf_tail_call_static(void *ctx, const void *map, const __u32 slot)

nit: const void *ctx would work here, right? would avoid users having
to do unnecessary casts in some cases

+{
+       if (!__builtin_constant_p(slot))
+               __bpf_unreachable();
+
+       /*
+        * Provide a hard guarantee that LLVM won't optimize setting r2 (map
+        * pointer) and r3 (constant map index) from _different paths_ ending
+        * up at the _same_ call insn as otherwise we won't be able to use the
+        * jmpq/nopl retpoline-free patching by the x86-64 JIT in the kernel
+        * given they mismatch. See also d2e4c1e6c294 ("bpf: Constant map key
+        * tracking for prog array pokes") for details on verifier tracking.
+        *
+        * Note on clobber list: we need to stay in-line with BPF calling
+        * convention, so even if we don't end up using r0, r4, r5, we need
+        * to mark them as clobber so that LLVM doesn't end up using them
+        * before / after the call.
+        */
+       asm volatile("r1 = %[ctx]\n\t"
+                    "r2 = %[map]\n\t"
+                    "r3 = %[slot]\n\t"
+                    "call 12\n\t"

nit: it's weird to have tabs at the end of each string literal,
especially that r1 doesn't start with a tab...

+                    :: [ctx]"r"(ctx), [map]"r"(map), [slot]"i"(slot)
+                    : "r0", "r1", "r2", "r3", "r4", "r5");
+}
+
 /*
  * Helper structure used by eBPF C program
  * to describe BPF map attributes to libbpf loader
--
2.21.0