Merge pull request #1797 from SparkiDev/stack_size

Small stack usage fixes

configure.ac

@@ -1645,6 +1645,19 @@ then
 fi
 
 
+# STACK usage logging
+AC_ARG_ENABLE([stacklog],
+    [AS_HELP_STRING([--enable-stacklog],[Enable stack logging (default: disabled)])],
+    [ ENABLED_STACKLOG=$enableval ],
+    [ ENABLED_STACKLOG=no ]
+    )
+
+if test "$ENABLED_STACKLOG" = "yes"
+then
+    AM_CFLAGS="$AM_CFLAGS -DWOLFSSL_STACK_LOG -finstrument-functions"
+fi
+
+
 
 # RSA
 AC_ARG_ENABLE([rsa],

src/internal.c

@@ -8679,6 +8679,8 @@ int ProcessPeerCerts(WOLFSSL* ssl, byte* input, word32* inOutIdx,
     (void)sizeof(args_test);
 #elif defined(WOLFSSL_NONBLOCK_OCSP)
     ProcPeerCertArgs* args = ssl->nonblockarg;
+#elif defined(WOLFSSL_SMALL_STACK)
+    ProcPeerCertArgs* args = NULL;
 #else
     ProcPeerCertArgs  args[1];
 #endif
@@ -8707,6 +8709,12 @@ int ProcessPeerCerts(WOLFSSL* ssl, byte* input, word32* inOutIdx,
         }
     }
     if (ssl->nonblockarg == NULL) /* new args */
+#elif defined(WOLFSSL_SMALL_STACK)
+    args = (ProcPeerCertArgs*)XMALLOC(
+        sizeof(ProcPeerCertArgs), ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
+    if (args == NULL) {
+        ERROR_OUT(MEMORY_E, exit_ppc);
+    }
 #endif
     {
         /* Reset state */
@@ -9837,9 +9845,12 @@ exit_ppc:
 
     FreeProcPeerCertArgs(ssl, args);
 
-#if !defined(WOLFSSL_ASYNC_CRYPT) && defined(WOLFSSL_NONBLOCK_OCSP)
+#if defined(WOLFSSL_ASYNC_CRYPT)
+#elif defined(WOLFSSL_NONBLOCK_OCSP)
     XFREE(args, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
     ssl->nonblockarg = NULL;
+#elif defined(WOLFSSL_SMALL_STACK)
+    XFREE(args, ssl->heap, DYNAMIC_TYPE_TMP_BUFFER);
 #endif
 
     FreeKeyExchange(ssl);

wolfcrypt/src/memory.c

@@ -1074,3 +1074,22 @@ void xfree(void *p, void* heap, int type, const char* func, const char* file,
 }
 #endif /* WOLFSSL_MEMORY_LOG */
 
+#ifdef WOLFSSL_STACK_LOG
+/* Note: this code only works with GCC using -finstrument-functions. */
+void __attribute__((no_instrument_function))
+     __cyg_profile_func_enter(void *func,  void *caller)
+{
+    register void* sp asm("sp");
+    fprintf(stderr, "ENTER: %016lx %p\n", (size_t)func, sp);
+    (void)caller;
+}
+
+void __attribute__((no_instrument_function))
+     __cyg_profile_func_exit(void *func, void *caller)
+{
+    register void* sp asm("sp");
+    fprintf(stderr, "EXIT: %016lx %p\n", (size_t)func, sp);
+    (void)caller;
+}
+#endif
+

wolfcrypt/src/sp_arm32.c

@@ -19185,7 +19185,6 @@ static void sp_256_div2_8(sp_digit* r, sp_digit* a, sp_digit* m)
 static void sp_256_proj_point_dbl_8(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* x;
@@ -19195,7 +19194,8 @@ static void sp_256_proj_point_dbl_8(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -19372,7 +19372,6 @@ static void sp_256_proj_point_add_8(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* t3 = t + 4*8;
@@ -19399,8 +19398,8 @@ static void sp_256_proj_point_add_8(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -19596,7 +19595,6 @@ static void sp_256_proj_point_dbl_n_8(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*8;
     sp_digit* b = t + 4*8;
@@ -19608,7 +19606,8 @@ static void sp_256_proj_point_dbl_n_8(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -19672,7 +19671,6 @@ static void sp_256_proj_point_add_qz1_8(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*8;
     sp_digit* t3 = t + 4*8;
@@ -19692,8 +19690,8 @@ static void sp_256_proj_point_add_qz1_8(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;

wolfcrypt/src/sp_arm64.c

@@ -13408,7 +13408,6 @@ static void sp_256_div2_4(sp_digit* r, sp_digit* a, sp_digit* m)
 static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -13418,7 +13417,8 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -13483,7 +13483,6 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -13495,7 +13494,8 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -13568,7 +13568,6 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -13595,8 +13594,8 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -13998,7 +13997,6 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -14018,8 +14016,8 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;

wolfcrypt/src/sp_c32.c

@@ -8504,7 +8504,6 @@ static void sp_256_div2_10(sp_digit* r, sp_digit* a, sp_digit* m)
 static void sp_256_proj_point_dbl_10(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* x;
@@ -8514,7 +8513,8 @@ static void sp_256_proj_point_dbl_10(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -8594,7 +8594,6 @@ static void sp_256_proj_point_add_10(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* t3 = t + 4*10;
@@ -8621,8 +8620,8 @@ static void sp_256_proj_point_add_10(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -9018,7 +9017,6 @@ static void sp_256_proj_point_dbl_n_10(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*10;
     sp_digit* b = t + 4*10;
@@ -9030,7 +9028,8 @@ static void sp_256_proj_point_dbl_n_10(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -9094,7 +9093,6 @@ static void sp_256_proj_point_add_qz1_10(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*10;
     sp_digit* t3 = t + 4*10;
@@ -9114,8 +9112,8 @@ static void sp_256_proj_point_add_qz1_10(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;

wolfcrypt/src/sp_c64.c

@@ -7727,7 +7727,6 @@ static void sp_256_div2_5(sp_digit* r, sp_digit* a, sp_digit* m)
 static void sp_256_proj_point_dbl_5(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* x;
@@ -7737,7 +7736,8 @@ static void sp_256_proj_point_dbl_5(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -7816,7 +7816,6 @@ static void sp_256_proj_point_add_5(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* t3 = t + 4*5;
@@ -7843,8 +7842,8 @@ static void sp_256_proj_point_add_5(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -8240,7 +8239,6 @@ static void sp_256_proj_point_dbl_n_5(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*5;
     sp_digit* b = t + 4*5;
@@ -8252,7 +8250,8 @@ static void sp_256_proj_point_dbl_n_5(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -8316,7 +8315,6 @@ static void sp_256_proj_point_add_qz1_5(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*5;
     sp_digit* t3 = t + 4*5;
@@ -8336,8 +8334,8 @@ static void sp_256_proj_point_add_qz1_5(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;

wolfcrypt/src/sp_x86_64.c

@@ -28686,7 +28686,6 @@ SP_NOINLINE static void sp_256_div2_4(sp_digit* r, sp_digit* a, sp_digit* m)
 static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -28696,7 +28695,8 @@ static void sp_256_proj_point_dbl_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -28761,7 +28761,6 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -28773,7 +28772,8 @@ static void sp_256_proj_point_dbl_n_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -28846,7 +28846,6 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -28873,8 +28872,8 @@ static void sp_256_proj_point_add_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -29708,7 +29707,6 @@ static void sp_256_map_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 static void sp_256_proj_point_dbl_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* x;
@@ -29718,7 +29716,8 @@ static void sp_256_proj_point_dbl_avx2_4(sp_point* r, sp_point* p, sp_digit* t)
 
     /* When infinity don't double point passed in - constant time. */
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -29783,7 +29782,6 @@ static void sp_256_proj_point_dbl_n_avx2_4(sp_point* r, sp_point* p, int n,
         sp_digit* t)
 {
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* w = t;
     sp_digit* a = t + 2*4;
     sp_digit* b = t + 4*4;
@@ -29795,7 +29793,8 @@ static void sp_256_proj_point_dbl_n_avx2_4(sp_point* r, sp_point* p, int n,
     int i;
 
     rp[0] = r;
-    rp[1] = &tp;
+    rp[1] = (sp_point*)t;
+    XMEMSET(rp[1], 0, sizeof(sp_point));
     x = rp[p->infinity]->x;
     y = rp[p->infinity]->y;
     z = rp[p->infinity]->z;
@@ -29856,7 +29855,6 @@ static void sp_256_proj_point_add_avx2_4(sp_point* r, sp_point* p, sp_point* q,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -29883,8 +29881,8 @@ static void sp_256_proj_point_add_avx2_4(sp_point* r, sp_point* p, sp_point* q,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -30217,7 +30215,6 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -30237,8 +30234,8 @@ static void sp_256_proj_point_add_qz1_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;
@@ -30610,7 +30607,6 @@ static void sp_256_proj_point_add_qz1_avx2_4(sp_point* r, sp_point* p,
 {
     sp_point* ap[2];
     sp_point* rp[2];
-    sp_point tp;
     sp_digit* t1 = t;
     sp_digit* t2 = t + 2*4;
     sp_digit* t3 = t + 4*4;
@@ -30630,8 +30626,8 @@ static void sp_256_proj_point_add_qz1_avx2_4(sp_point* r, sp_point* p,
     }
     else {
         rp[0] = r;
-        rp[1] = &tp;
+        rp[1] = (sp_point*)t;
-        XMEMSET(&tp, 0, sizeof(tp));
+        XMEMSET(rp[1], 0, sizeof(sp_point));
         x = rp[p->infinity | q->infinity]->x;
         y = rp[p->infinity | q->infinity]->y;
         z = rp[p->infinity | q->infinity]->z;