Skip to contentSkip to conversation

selftests/bpf: skip verif_scale_pyperf600 on -E2BIG jump complexity limit

PATCH In review Sun Jian 3 participants 6 messages Mar 2 – Mar 6

Subsystems (3)
lore
Share & tools
lore

Sun Jian

author
#0 lore raw

PATCH 1/1 selftests/bpf: skip verif_scale_pyperf600 on -E2BIG jump complexity limit

verif_scale_pyperf600 loads pyperf600.bpf.o, a large scale case produced by partial LLVM unrolling.

On some kernel/toolchain combinations, the verifier rejects it with -E2BIG:

The sequence of 8193 jumps is too complex.

This hits the verifier jump sequence complexity limit (8192).

Treat -E2BIG for this test as SKIP rather than FAIL. Other pyperf600 variants are unaffected.

---
 .../bpf/prog_tests/bpf_verif_scale.c          | 22 ++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)
tools/testing/selftests/bpf/prog_tests/bpf_verif_scale.c +21 -1
index 73f669014b69..181acdddea1c 100644
@@ -112,13 +112,33 @@ void test_verif_scale_pyperf180()
 
 void test_verif_scale_pyperf600()
 {
+	libbpf_print_fn_t old_print_fn = NULL;
+	int err;
+
 	/* partial unroll. llvm will unroll loop ~150 times.
 	 * C loop count -> 600.
 	 * Asm loop count -> 4.
 	 * 16k insns in loop body.
 	 * Total of 5 such loops. Total program size ~82k insns.
 	 */
-	scale_test("pyperf600.bpf.o", BPF_PROG_TYPE_RAW_TRACEPOINT, false);
+
+	if (env.verifier_stats) {
+		test__force_log();
+		old_print_fn = libbpf_set_print(libbpf_debug_print);
+	}
+
+	err = check_load("pyperf600.bpf.o", BPF_PROG_TYPE_RAW_TRACEPOINT);
+
+	if (env.verifier_stats)
+		libbpf_set_print(old_print_fn);
+
+	if (err == -E2BIG) {
+		test__skip();
+		return;
+	}
+
+	ASSERT_OK(err, "expect_success");
+
 }
 
 void test_verif_scale_pyperf600_bpf_loop(void)

base-commit: 11439c4635edd669ae435eec308f4ab8a0804808

Signature · 1 line
2.43.0
#1 lore raw
Quoted 11 lines · Sun Jian

verif_scale_pyperf600 loads pyperf600.bpf.o, a large scale case produced by partial LLVM unrolling.

On some kernel/toolchain combinations, the verifier rejects it with -E2BIG:

The sequence of 8193 jumps is too complex.

This hits the verifier jump sequence complexity limit (8192).

Treat -E2BIG for this test as SKIP rather than FAIL. Other pyperf600 variants are unaffected.

This is a known issue with LLVM 19+ [1] and this test is already skipped in CI [2, 3]. I don't think ignoring the error is the proper fix here.

1 - https://lore.kernel.org/all/[email protected]/
2 - https://github.com/kernel-patches/vmtest/commit/da142df34750386cac2ba10283d15216d5c8d033
3 - https://github.com/libbpf/libbpf/commit/02b3ec9ffc1b7000afa2a9ccc3dfd37de3bad8cb

Paul

Quoted 51 lines
---
 .../bpf/prog_tests/bpf_verif_scale.c          | 22 ++++++++++++++++++-
 1 file changed, 21 insertions(+), 1 deletion(-)
tools/testing/selftests/bpf/prog_tests/bpf_verif_scale.c +21 -1
index 73f669014b69..181acdddea1c 100644
@@ -112,13 +112,33 @@ void test_verif_scale_pyperf180()
 
 void test_verif_scale_pyperf600()
 {
+	libbpf_print_fn_t old_print_fn = NULL;
+	int err;
+
 	/* partial unroll. llvm will unroll loop ~150 times.
 	 * C loop count -> 600.
 	 * Asm loop count -> 4.
 	 * 16k insns in loop body.
 	 * Total of 5 such loops. Total program size ~82k insns.
 	 */
-	scale_test("pyperf600.bpf.o", BPF_PROG_TYPE_RAW_TRACEPOINT, false);
+
+	if (env.verifier_stats) {
+		test__force_log();
+		old_print_fn = libbpf_set_print(libbpf_debug_print);
+	}
+
+	err = check_load("pyperf600.bpf.o", BPF_PROG_TYPE_RAW_TRACEPOINT);
+
+	if (env.verifier_stats)
+		libbpf_set_print(old_print_fn);
+
+	if (err == -E2BIG) {
+		test__skip();
+		return;
+	}
+
+	ASSERT_OK(err, "expect_success");
+
 }
 
 void test_verif_scale_pyperf600_bpf_loop(void)

base-commit: 11439c4635edd669ae435eec308f4ab8a0804808

Signature · 1 line
2.43.0
author
#2 lore raw
Quoted 9 lines · Paul Chaignon
Quoted 2 lines · Sun Jian

verif_scale_pyperf600 loads pyperf600.bpf.o, a large scale case produced by partial LLVM unrolling.

This is a known issue with LLVM 19+ [1] and this test is already skipped in CI [2, 3]. I don't think ignoring the error is the proper fix here.

Thanks for the pointers. Agreed — since this is a known LLVM issue, I'll drop
this patch.

#3 lore raw
Quoted 7 lines · sun jian
Quoted 5 lines · Paul Chaignon
Quoted 2 lines · Sun Jian

verif_scale_pyperf600 loads pyperf600.bpf.o, a large scale case produced by partial LLVM unrolling.

Quoted 6 lines
Quoted 4 lines

This is a known issue with LLVM 19+ [1] and this test is already skipped in CI [2, 3]. I don't think ignoring the error is the proper fix here.

Thanks for the pointers. Agreed — since this is a known LLVM issue, I'll drop
this patch.

It's been failing for too long.
Pls do a minimal change to its source to workaround this 8k limit.

pw-bot: cr

author
#4 lore raw
Quoted 2 lines · Alexei Starovoitov

It's been failing for too long.
Pls do a minimal change to its source to workaround this 8k limit.

Will do, sending v2.