feat: Detect eBPF features without relying on kernel version checks

Kernel version checks are unreliable:

* Features differ across architectures.
* Linux distros tend to backport features to their stable/LTS kernels.

Instead of relying on version, use small probes (written in eBPF
assembly) to figure out what features are available.

crates/bpf-builder/include/output.bpf.h

@@ -5,7 +5,6 @@
 
 #include "bpf/bpf_helpers.h"
 #include "buffer.bpf.h"
-#include "compatibility.bpf.h"
 #include "interest_tracking.bpf.h"
 
 // eBPF programs could interrupt each other, see "Are BPF programs preemptible?"
@@ -151,27 +150,27 @@ static __always_inline int is_initialized() {
 
 // Get a value and increment it by one
 static __always_inline u64 sync_increment(u64 *value) {
-  if (HAVE_FEATURE_ATOMICS) {
-    return __sync_fetch_and_add(value, 1);
-  } else {
-    // If we miss atomic operations, it still shouldn't cause problems:
-    // - the nesting levels are kept on a PERCPU array
-    // - even if an eBPF program interrupts this, the nesting level
-    //   will be left in a consistent state as all eBPF programs will
-    //   reset the counter to its previous value before exiting.
-    u64 old_value = *value;
-    *value += 1;
-    return old_value;
-  }
+#ifdef FEATURE_ATOMICS
+  return __sync_fetch_and_add(value, 1);
+#else
+  // If we miss atomic operations, it still shouldn't cause problems:
+  // - the nesting levels are kept on a PERCPU array
+  // - even if an eBPF program interrupts this, the nesting level
+  //   will be left in a consistent state as all eBPF programs will
+  //   reset the counter to its previous value before exiting.
+  u64 old_value = *value;
+  *value += 1;
+  return old_value;
+#endif
 }
 
-// Decremnet value by one
+// Decrement value by one
 static __always_inline u64 sync_decrement(u64 *value) {
-  if (HAVE_FEATURE_ATOMICS) {
-    return __sync_fetch_and_sub(value, 1);
-  } else {
-    u64 old_value = *value;
-    *value -= 1;
-    return old_value;
-  }
+#ifdef FEATURE_ATOMICS
+  return __sync_fetch_and_sub(value, 1);
+#else
+  u64 old_value = *value;
+  *value -= 1;
+  return old_value;
+#endif
 }

crates/bpf-builder/include/task.bpf.h

@@ -0,0 +1,13 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+
+#pragma once
+
+#include "common.bpf.h"
+
+static __always_inline struct task_struct *get_current_task() {
+#ifdef FEATURE_TASK_BTF
+  return bpf_get_current_task_btf();
+#else
+  return (struct task_struct*)bpf_get_current_task();
+#endif
+}

crates/bpf-builder/src/lib.rs

@@ -1,4 +1,9 @@
-use std::{env, path::PathBuf, process::Command, string::String};
+use std::{
+    env,
+    path::{Path, PathBuf},
+    process::Command,
+    string::String,
+};
 
 use anyhow::{bail, Context};
 
@@ -7,10 +12,8 @@ static LLVM_STRIP: &str = "llvm-strip";
 static INCLUDE_PATH: &str = concat!(env!("CARGO_MANIFEST_DIR"), "/include");
 
 // Given a probe name and the eBPF program source code path, compile it to OUT_DIR.
-// We'll build multiple versions:
-// - `${OUT_DIR}/{name}.5_13.bpf.o`: will contain the full version
-// - `${OUT_DIR}/{name}.5_5.bpf.o`: will contain a version with the FEATURE_5_5 constant
-//   defined. This version should be loaded on kernel < 5.13, see ../include/compatibility.bpf.h
+// We'll build multiple versions with all combinations of eBPF features we rely
+// on. Lack of certain feature will result in using legacy replacements.
 pub fn build(name: &str, source: &str) -> Result<(), Box<dyn std::error::Error>> {
     println!("cargo:rerun-if-changed={source}");
     println!("cargo:rerun-if-changed={INCLUDE_PATH}/common.bpf.h");
@@ -19,23 +22,42 @@ pub fn build(name: &str, source: &str) -> Result<(), Box<dyn std::error::Error>>
     println!("cargo:rerun-if-changed={INCLUDE_PATH}/interest_tracking.bpf.h");
     println!("cargo:rerun-if-changed={INCLUDE_PATH}/loop.bpf.h");
     println!("cargo:rerun-if-changed={INCLUDE_PATH}/get_path.bpf.h");
-    println!("cargo:rerun-if-changed={INCLUDE_PATH}/compatibility.bpf.h");
+    println!("cargo:rerun-if-changed={INCLUDE_PATH}/task.bpf.h");
 
-    let out_file = PathBuf::from(env::var("OUT_DIR")?).join(name);
+    let features = [
+        ("FEATURE_ATOMICS", "a"),
+        ("FEATURE_CGROUP_TASK_BTF", "c"),
+        ("FEATURE_FN_POINTERS", "f"),
+    ];
 
-    compile(
-        source,
-        out_file.with_extension("5_13.bpf.o"),
-        &["-DVERSION_5_13"],
-    )
-    .context("Error compiling 5.13 version")?;
-    compile(source, out_file.with_extension("5_5.bpf.o"), &[])
-        .context("Error compiling 5.5 version")?;
+    let out_dir = env::var("OUT_DIR").context("OUT_DIR not set")?;
+    let out_path = Path::new(&out_dir).join(name);
+
+    for bits in 0..(1 << features.len()) {
+        let mut feature_args = Vec::new();
+        let mut suffix = String::new();
+
+        for (i, (feature, code)) in features.iter().enumerate() {
+            if bits & (1 << i) != 0 {
+                feature_args.push(format!("-D{}", feature));
+                suffix.push_str(code);
+            }
+        }
+
+        if suffix.is_empty() {
+            suffix.push_str("none");
+        }
+
+        let filename = format!("{}.{}.bpf.o", name, suffix);
+        let full_path = out_path.with_file_name(filename);
+        compile(source, full_path, &feature_args)
+            .context("Error compiling programs with features: {feature_args:?}")?;
+    }
 
     Ok(())
 }
 
-fn compile(probe: &str, out_object: PathBuf, extra_args: &[&str]) -> anyhow::Result<()> {
+fn compile(probe: &str, out_object: PathBuf, extra_args: &[String]) -> anyhow::Result<()> {
     let clang = env::var("CLANG").unwrap_or_else(|_| String::from(CLANG_DEFAULT));
     let arch = env::var("CARGO_CFG_TARGET_ARCH").unwrap();
     let include_path = PathBuf::from(INCLUDE_PATH);

crates/bpf-common/Cargo.toml

@@ -13,6 +13,7 @@ sqlite3-vendored = ["libsqlite3-sys/bundled"]
 
 [dependencies]
 aya = { workspace = true, features = ["async_tokio"] }
+aya-ebpf-bindings = { workspace = true }
 aya-obj = { workspace = true }
 bytes = { workspace = true }
 thiserror = { workspace = true }

crates/bpf-common/src/feature_autodetect/atomic.rs

@@ -0,0 +1,49 @@
+use aya_ebpf_bindings::bindings::{BPF_DW, BPF_JEQ, BPF_REG_0, BPF_REG_1, BPF_REG_10, BPF_XCHG};
+use aya_obj::generated::bpf_insn;
+use log::warn;
+
+use super::{load_program, BpfProgType};
+use crate::insn;
+
+/// eBPF program bytecode with simple atomic operations.
+fn bpf_prog_atomic() -> Vec<bpf_insn> {
+    vec![
+        // val = 3;
+        insn::bpf_st_mem(BPF_DW as u8, BPF_REG_10 as u8, -8, 3),
+        // old = atomic_xchg(&val, 4);
+        insn::bpf_mov64_imm(BPF_REG_1 as u8, 4),
+        insn::bpf_atomic_op(
+            BPF_DW as u8,
+            BPF_XCHG,
+            BPF_REG_10 as u8,
+            BPF_REG_1 as u8,
+            -8,
+        ),
+        // if (old != 3) exit(1);
+        insn::bpf_jmp_imm(BPF_JEQ as u8, BPF_REG_1 as u8, 3, 2),
+        insn::bpf_mov64_imm(BPF_REG_0 as u8, 1),
+        insn::bpf_exit_insn(),
+        // if (val != 4) exit(2);
+        insn::bpf_ldx_mem(BPF_DW as u8, BPF_REG_0 as u8, BPF_REG_10 as u8, -8),
+        insn::bpf_jmp_imm(BPF_JEQ as u8, BPF_REG_0 as u8, 4, 2),
+        insn::bpf_mov64_imm(BPF_REG_0 as u8, 2),
+        insn::bpf_exit_insn(),
+        // exit(0);
+        insn::bpf_mov64_imm(BPF_REG_0 as u8, 0),
+        insn::bpf_exit_insn(),
+    ]
+}
+
+/// Checks whether the current kernel supports atomic operations in eBPF.
+pub fn atomics_supported() -> bool {
+    let insns = bpf_prog_atomic();
+    // Program type doesn't matter, kprobe is just the most basic one.
+    let res = load_program(BpfProgType::BPF_PROG_TYPE_KPROBE, insns);
+    match res {
+        Ok(_) => true,
+        Err(e) => {
+            warn!("Atomic operations in eBPF are not supported by the kernel: {e}");
+            false
+        }
+    }
+}

crates/bpf-common/src/feature_autodetect/func.rs

@@ -0,0 +1,35 @@
+use aya_ebpf_bindings::bindings::BPF_REG_0;
+use aya_obj::generated::bpf_insn;
+use log::warn;
+
+use super::{load_program, BpfProgType};
+use crate::insn;
+
+/// eBPF program bytecode with a simple function call (with `bpf_emit_call`
+/// instruction) to the function with the given ID.
+fn bpf_prog_func_id(func_id: u32) -> Vec<bpf_insn> {
+    vec![
+        insn::bpf_emit_call(func_id),
+        insn::bpf_mov64_imm(BPF_REG_0 as u8, 0),
+        insn::bpf_exit_insn(),
+    ]
+}
+
+/// Checks whether the provided `func_id` for the given `prog_type` is
+/// supported by the current kernel, by loading a minimal program trying to use
+/// it.
+///
+/// Similar checks are performed by [`bpftool`].
+///
+/// [`bpftool`]: https://github.com/torvalds/linux/blob/v6.8/tools/bpf/bpftool/feature.c#L534-L544
+pub fn func_id_supported(func_id: u32, prog_type: BpfProgType) -> bool {
+    let insns = bpf_prog_func_id(func_id);
+    let res = load_program(prog_type, insns);
+    match res {
+        Ok(_) => true,
+        Err(err) => {
+            warn!("Function {func_id} in program type {prog_type:?} not supported: {err}");
+            false
+        }
+    }
+}

crates/bpf-common/src/feature_autodetect/lsm.rs

@@ -28,7 +28,7 @@ pub fn lsm_supported() -> bool {
 
 const PATH: &str = "/sys/kernel/security/lsm";
 static TEST_LSM_PROBE: &[u8] =
-    include_bytes_aligned!(concat!(env!("OUT_DIR"), "/test_lsm.5_13.bpf.o"));
+    include_bytes_aligned!(concat!(env!("OUT_DIR"), "/test_lsm.none.bpf.o"));
 
 fn try_load() -> Result<()> {
     // Check if LSM enabled