Add stack benchmarks (#10)

* add stack benchmarks Signed-off-by: Matthias J. Kannwischer <[email protected]> * format Signed-off-by: Matthias J. Kannwischer <[email protected]> * change ldscript generation to make stack benchmarks work Signed-off-by: Matthias J. Kannwischer <[email protected]> * update CI to check correct number of OKs Signed-off-by: Matthias J. Kannwischer <[email protected]> --------- Signed-off-by: Matthias J. Kannwischer <[email protected]>
pq-code-package · Mar 29, 2024 · 8bb4b0a · 8bb4b0a
1 parent 9f49561
commit 8bb4b0a
Show file tree

Hide file tree

Showing 6 changed files with 111 additions and 3 deletions.
diff --git a/.github/workflows/build.yaml b/.github/workflows/build.yaml
@@ -67,7 +67,9 @@ jobs:
             # check if num of ok is as expected
             if [[ "$file" == *"test"* ]]; then
               expect=$((tests*3))
-            elif [[ "$file" == *"speed"* ]] then
+            elif [[ "$file" == *"stack"* ]]; then
+              expect=1
+            else
               expect=$tests
             fi
 

diff --git a/Makefile b/Makefile
@@ -3,7 +3,7 @@
 
 .PHONY: all
 
-all: test speed
+all: test speed stack
 
 
 include mk/config.mk
@@ -12,6 +12,7 @@ include mk/rules.mk
 
 test: $(foreach scheme,$(KEM_SCHEMES),$(scheme)-test)
 speed: $(foreach scheme,$(KEM_SCHEMES),$(scheme)-speed)
+stack: $(foreach scheme,$(KEM_SCHEMES),$(scheme)-stack)
 
 .PHONY: emulate clean libclean
 

diff --git a/hal/devices.data b/hal/devices.data
@@ -0,0 +1,3 @@
+# SPDX-License-Identifier: Apache-2.0
+stm32f407vg stm32f4 ROM=1024K RAM=128K
+stm32f4 END ROM_OFF=0x08000000 RAM_OFF=0x20000000 CPU=cortex-m4 FPU=hard-fpv4-sp-d16
diff --git a/mk/opencm3.mk b/mk/opencm3.mk
@@ -16,7 +16,7 @@ ifeq ($(DEVICE),)
 $(warning no DEVICE specified for linker script generator)
 endif
 
-DEVICES_DATA ?= $(OPENCM3_DIR)/ld/devices.data
+DEVICES_DATA ?= hal/devices.data
 
 genlink_family		:=$(shell $(OPENCM3_DIR)/scripts/genlink.py $(DEVICES_DATA) $(DEVICE) FAMILY)
 genlink_subfamily	:=$(shell $(OPENCM3_DIR)/scripts/genlink.py $(DEVICES_DATA) $(DEVICE) SUBFAMILY)

diff --git a/mk/schemes.mk b/mk/schemes.mk
@@ -10,11 +10,15 @@ $(1)-test: bin/$(1)-test.hex
 
 $(1)-speed: CPPFLAGS += -DKYBER_K=$(2) -DMUPQ_NAMESPACE=$$(MUPQ_NAMESPACE) -DNTESTS=$$(NTESTS)
 $(1)-speed: bin/$(1)-speed.hex
+
+$(1)-stack: CPPFLAGS += -DKYBER_K=$(2) -DMUPQ_NAMESPACE=$$(MUPQ_NAMESPACE)
+$(1)-stack: bin/$(1)-stack.hex
 endef
 
 define scheme-elf
 elf/$(1)-test.elf: obj/test/$(1)-test.o $(call OBJS,$(1)) $(LINKDEPS)
 elf/$(1)-speed.elf: obj/test/$(1)-speed.o $(call OBJS,$(1)) $(LINKDEPS)
+elf/$(1)-stack.elf: obj/test/$(1)-stack.o $(call OBJS,$(1)) $(LINKDEPS)
 endef
 
 define compile-obj
@@ -34,6 +38,9 @@ obj/test/$(1)-test.o: test/test.c
 
 obj/test/$(1)-speed.o: test/speed.c
 	$$(compile-obj)
+
+obj/test/$(1)-stack.o: test/stack.c
+	$$(compile-obj)
 endef
 
 # iterate over schemes and params to generate targets for functional test and speed benchmark

diff --git a/test/stack.c b/test/stack.c
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: Apache-2.0 or CC0-1.0
+#include "kem.h"
+#include "randombytes.h"
+#include "hal.h"
+#include "sendfn.h"
+
+#include <string.h>
+
+#ifndef MAX_STACK_SIZE
+#define MAX_STACK_SIZE hal_get_stack_size()
+#endif
+
+#ifndef STACK_SIZE_INCR
+#define STACK_SIZE_INCR 0x1000
+#endif
+
+// https://stackoverflow.com/a/1489985/1711232
+#define PASTER(x, y) x##y
+#define EVALUATOR(x, y) PASTER(x, y)
+#define NAMESPACE(fun) EVALUATOR(MUPQ_NAMESPACE, fun)
+
+// use different names so we can have empty namespaces
+#define MUPQ_CRYPTO_BYTES           NAMESPACE(CRYPTO_BYTES)
+#define MUPQ_CRYPTO_PUBLICKEYBYTES  NAMESPACE(CRYPTO_PUBLICKEYBYTES)
+#define MUPQ_CRYPTO_SECRETKEYBYTES  NAMESPACE(CRYPTO_SECRETKEYBYTES)
+#define MUPQ_CRYPTO_CIPHERTEXTBYTES NAMESPACE(CRYPTO_CIPHERTEXTBYTES)
+#define MUPQ_CRYPTO_ALGNAME NAMESPACE(CRYPTO_ALGNAME)
+
+#define MUPQ_crypto_kem_keypair NAMESPACE(crypto_kem_keypair)
+#define MUPQ_crypto_kem_enc NAMESPACE(crypto_kem_enc)
+#define MUPQ_crypto_kem_dec NAMESPACE(crypto_kem_dec)
+
+#define send_stack_usage(S, U) send_unsigned((S), (U))
+
+unsigned int canary_size;
+volatile unsigned char *p;
+unsigned int c;
+uint8_t canary = 0x42;
+
+unsigned char key_a[MUPQ_CRYPTO_BYTES], key_b[MUPQ_CRYPTO_BYTES];
+unsigned char pk[MUPQ_CRYPTO_PUBLICKEYBYTES];
+unsigned char sendb[MUPQ_CRYPTO_CIPHERTEXTBYTES];
+unsigned char sk_a[MUPQ_CRYPTO_SECRETKEYBYTES];
+unsigned int stack_key_gen, stack_encaps, stack_decaps;
+
+#define FILL_STACK()                                                           \
+    p = &a;                                                                      \
+    while (p > &a - canary_size)                                                 \
+        *(p--) = canary;
+#define CHECK_STACK()                                                          \
+    c = canary_size;                                                             \
+    p = &a - canary_size + 1;                                                    \
+    while (*p == canary && p < &a) {                                             \
+        p++;                                                                       \
+        c--;                                                                       \
+    }
+
+static int test_keys(void) {
+    // Alice generates a public key
+    hal_spraystack();
+    MUPQ_crypto_kem_keypair(pk, sk_a);
+    stack_key_gen = hal_checkstack();
+
+    // Bob derives a secret key and creates a response
+    hal_spraystack();
+    MUPQ_crypto_kem_enc(sendb, key_b, pk);
+    stack_encaps = hal_checkstack();
+
+    // Alice uses Bobs response to get her secret key
+    hal_spraystack();
+    MUPQ_crypto_kem_dec(key_a, sendb, sk_a);
+    stack_decaps = hal_checkstack();
+
+    if (memcmp(key_a, key_b, MUPQ_CRYPTO_BYTES)) {
+        return -1;
+    } else {
+        send_stack_usage("keypair stack usage:", stack_key_gen);
+        send_stack_usage("encaps stack usage:", stack_encaps);
+        send_stack_usage("decaps stack usage:", stack_decaps);
+        hal_send_str("OK KEYS\n");
+        return 0;
+    }
+}
+
+int main(void) {
+    hal_setup(CLOCK_FAST);
+
+    // marker for automated benchmarks
+    hal_send_str("==========================");
+    test_keys();
+    // marker for automated benchmarks
+    hal_send_str("#");
+
+    return 0;
+}