Skip to contentSkip to conversation

selftests/seccomp: Expand benchmark to per-filter measurements

PATCH New Kees Cook 1 participant 1 message Jun 1, 2020

Subsystems (2)
lore
Share & tools
lore

Kees Cook

author
#0 lore raw

PATCH 1/1 selftests/seccomp: Expand benchmark to per-filter measurements

It's useful to see how much (at a minimum) each filter adds to the syscall overhead. Add additional calculations.

---
As part of the performance discussions, this is what I'm adding to the seccomp selftest in for-next/seccomp to get more details.
https://lore.kernel.org/linux-security-module/202006011116.3F7109A@keescook/T/#md8a6fd608cfd1f3c70aaf9ccc4f09fcc33b5fc1b
This does not include the BPF-bypass mode, which I don't see a way to do without creating major problems with seccomp. ;)

---
 .../selftests/seccomp/seccomp_benchmark.c     | 36 +++++++++++++++----
 tools/testing/selftests/seccomp/seccomp_bpf.c |  2 --
 2 files changed, 29 insertions(+), 9 deletions(-)
tools/testing/selftests/seccomp/seccomp_benchmark.c +29 -7
index 5838c8697ec3..eca13fe1fba9 100644
@@ -68,32 +68,54 @@ int main(int argc, char *argv[])
 	};
 	long ret;
 	unsigned long long samples;
-	unsigned long long native, filtered;
+	unsigned long long native, filter1, filter2;
 
 	if (argc > 1)
 		samples = strtoull(argv[1], NULL, 0);
 	else
 		samples = calibrate();
 
+	printf("Current BPF sysctl settings:\n");
+	system("sysctl net.core.bpf_jit_enable");
+	system("sysctl net.core.bpf_jit_harden");
 	printf("Benchmarking %llu samples...\n", samples);
 
+	/* Native call */
 	native = timing(CLOCK_PROCESS_CPUTIME_ID, samples) / samples;
 	printf("getpid native: %llu ns\n", native);
 
 	ret = prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
 	assert(ret == 0);
 
+	/* One filter */
 	ret = prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog);
 	assert(ret == 0);
 
-	filtered = timing(CLOCK_PROCESS_CPUTIME_ID, samples) / samples;
-	printf("getpid RET_ALLOW: %llu ns\n", filtered);
+	filter1 = timing(CLOCK_PROCESS_CPUTIME_ID, samples) / samples;
+	printf("getpid RET_ALLOW 1 filter: %llu ns\n", filter1);
 
-	printf("Estimated seccomp overhead per syscall: %llu ns\n",
-		filtered - native);
+	if (filter1 == native)
+		printf("No overhead measured!? Try running again with more samples.\n");
 
-	if (filtered == native)
-		printf("Trying running again with more samples.\n");
+	/* Two filters */
+	ret = prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &prog);
+	assert(ret == 0);
+
+	filter2 = timing(CLOCK_PROCESS_CPUTIME_ID, samples) / samples;
+	printf("getpid RET_ALLOW 2 filters: %llu ns\n", filter2);
+
+	/* Calculations */
+	printf("Estimated total seccomp overhead for 1 filter: %llu ns\n",
+		filter1 - native);
+
+	printf("Estimated total seccomp overhead for 2 filters: %llu ns\n",
+		filter2 - native);
+
+	printf("Estimated seccomp per-filter overhead: %llu ns\n",
+		filter2 - filter1);
+
+	printf("Estimated seccomp entry overhead: %llu ns\n",
+		filter1 - native - (filter2 - filter1));
 
 	return 0;
 }
tools/testing/selftests/seccomp/seccomp_bpf.c +0 -2
index 4dae278cf77e..402ccb3a4e52 100644
@@ -3824,7 +3824,6 @@ TEST(user_notification_filter_empty_threaded)
 
 /*
  * TODO:
- * - add microbenchmarks
  * - expand NNP testing
  * - better arch-specific TRACE and TRAP handlers.
  * - endianness checking when appropriate
@@ -3832,7 +3831,6 @@ TEST(user_notification_filter_empty_threaded)
  * - arch value testing (x86 modes especially)
  * - verify that FILTER_FLAG_LOG filters generate log messages
  * - verify that RET_LOG generates log messages
- * - ...
  */
 
 TEST_HARNESS_MAIN
Signature · 5 lines
2.25.1


-- 
Kees Cook