[sairex53]

Summary

Add libxdp_cache_kernel_features() so applications can probe and cache XDP frags and device-bound support during single-threaded startup. Dispatcher attachments then reuse those results instead of repeating unsupported-feature probes and their libbpf diagnostics.

Document the API and its startup requirements, export it in the LIBXDP_1.7.0 ABI, and add a regression test for repeated calls.

Fixes #478.

Testing

• make -s on Fedora 44 ARM64
• test_kernel_feature_cache passed when run directly
• Shared and static API link checks passed
• git diff --check passed

The full privileged test runner was not run because this SSH account lacks sudo and the required namespace privileges.

[tohojo]

I am OK with adding the API, but I don't think we should add a whole new code path for it. Rather, we can transparently introduce caching to the existing probe method, something like:

diff --git i/lib/libxdp/libxdp.c w/lib/libxdp/libxdp.c
index fa57b00329e1..7877e1da1b17 100644
--- i/lib/libxdp/libxdp.c
+++ w/lib/libxdp/libxdp.c
@@ -1694,7 +1694,7 @@ static const char *get_bpf_flag_name(__u32 flag)
        }
 }
 
-static bool kernel_has_bpf_flag(__u32 flag)
+static bool kernel_probe_bpf_flag(__u32 flag)
 {
        struct xdp_program *test_prog;
        bool ret = false;
@@ -1723,6 +1723,38 @@ static bool kernel_has_bpf_flag(__u32 flag)
        return ret;
 }
 
+/* We cache the two flags defined as bits 5 and 6 in the kernel headers */
+#define FEATURE_CACHE_OFFSET  5
+#define FEATURE_CACHE_SIZE   2
+#define FEATURE_PROBED       (1U << 0)
+#define FEATURE_AVAILABLE    (1U << 1)
+static unsigned int feature_cache[FEATURE_CACHE_SIZE];
+
+static bool kernel_has_bpf_flag(__u32 flag) {
+       unsigned int value, slot = __builtin_ctz(flag) - FEATURE_CACHE_OFFSET;
+
+       if (slot >= FEATURE_CACHE_SIZE) {
+               pr_warn("Invalid probe flag %u - please report this as a bug!\n", flag);
+               return false;
+       }
+
+       value = __atomic_load_n(&feature_cache[slot], __ATOMIC_ACQUIRE);
+       if (value & FEATURE_PROBED)
+               goto out;
+
+       value = FEATURE_PROBED;
+       if (kernel_probe_bpf_flag(flag))
+               value |= FEATURE_AVAILABLE;
+
+       /* We assume probing always returns the same result, so we don't bother
+        * with a cmpxchg() when setting this.
+        */
+       __atomic_store_n(&feature_cache[slot], value, __ATOMIC_RELEASE);
+
+out:
+       return !!(value & FEATURE_AVAILABLE);
+}
+
 static bool kernel_has_frags_support(void)
 {
        return kernel_has_bpf_flag(BPF_F_XDP_HAS_FRAGS);

The new function then becomes simply:

void libxdp_cache_kernel_features(void)
{
	(void)kernel_has_frags_support();
	(void)kernel_has_dev_bound();
}