diff --git a/hal/stm32wb.c b/hal/stm32wb.c
index a3e94ce2..2e6e5934 100644
--- a/hal/stm32wb.c
+++ b/hal/stm32wb.c
@@ -97,7 +97,8 @@ PKA_HandleTypeDef hpka = { };
 #define FLASH_ACR_LATENCY_MASK                (0x07)
 
 #ifndef WOLFSSL_STM32_PKA
-#define FLASH_SR_BSY                         (1 << 16)
+#define FLASH_SR_BSY                          (1 << 16)
+#define FLASH_SR_CFGBSY                       (1 << 18)
 #define FLASH_SR_SIZERR                       (1 << 6)
 #define FLASH_SR_PGAERR                       (1 << 5)
 #define FLASH_SR_WRPERR                       (1 << 4)
@@ -109,11 +110,12 @@ PKA_HandleTypeDef hpka = { };
 
 #define FLASH_CR_PER                          (1 << 1)
 #define FLASH_CR_PG                           (1 << 0)
+#define FLASH_CR_FSTPG                        (1 << 18)
 
 #endif /* !WOLFSSL_STM32_PKA */
 
 #define FLASH_CR_PNB_SHIFT                     3
-#define FLASH_CR_PNB_MASK                      0x3f
+#define FLASH_CR_PNB_MASK                      0xFF
 
 #define FLASH_KEY1                            (0x45670123)
 #define FLASH_KEY2                            (0xCDEF89AB)
@@ -128,7 +130,7 @@ static void RAMFUNCTION flash_set_waitstates(unsigned int waitstates)
 
 static RAMFUNCTION void flash_wait_complete(void)
 {
-    while ((FLASH_SR & FLASH_SR_BSY) == FLASH_SR_BSY)
+    while ((FLASH_SR & (FLASH_SR_BSY | FLASH_SR_CFGBSY)) != 0)
         ;
 }
 
@@ -137,21 +139,50 @@ static void RAMFUNCTION flash_clear_errors(void)
     FLASH_SR |= ( FLASH_SR_SIZERR | FLASH_SR_PGAERR | FLASH_SR_WRPERR |  FLASH_SR_PROGERR);
 }
 
+
+
+void RAMFUNCTION hal_flash_unlock(void)
+{
+    flash_wait_complete();
+    if ((FLASH_CR & FLASH_CR_LOCK) != 0) {
+        FLASH_KEY = FLASH_KEY1;
+        DMB();
+        FLASH_KEY = FLASH_KEY2;
+        DMB();
+        while ((FLASH_CR & FLASH_CR_LOCK) != 0)
+            ;
+    }
+}
+
+void RAMFUNCTION hal_flash_lock(void)
+{
+    flash_wait_complete();
+    if ((FLASH_CR & FLASH_CR_LOCK) == 0)
+        FLASH_CR |= FLASH_CR_LOCK;
+}
+
 int RAMFUNCTION hal_flash_write(uint32_t address, const uint8_t *data, int len)
 {
     int i = 0;
     uint32_t *src, *dst;
+    uint32_t pdword[2] __attribute__((aligned(16)));
+    uint32_t reg;
+
     flash_clear_errors();
-    FLASH_CR |= FLASH_CR_PG;
+    reg = FLASH_CR & (~FLASH_CR_FSTPG);
+    FLASH_CR = reg | FLASH_CR_PG;
 
     while (i < len) {
         flash_clear_errors();
         if ((len - i > 3) && ((((address + i) & 0x07) == 0)  && ((((uint32_t)data) + i) & 0x07) == 0)) {
+            uint32_t idx = i >> 2;
             src = (uint32_t *)data;
-            dst = (uint32_t *)(address + FLASHMEM_ADDRESS_SPACE);
+            dst = (uint32_t *)(address);
+            pdword[0] = src[idx];
+            pdword[1] = src[idx + 1];
             flash_wait_complete();
-            dst[i >> 2] = src[i >> 2];
-            dst[(i >> 2) + 1] = src[(i >> 2) + 1];
+            dst[idx] = pdword[0];
+            dst[idx + 1] = pdword[1];
             flash_wait_complete();
             i+=8;
         } else {
@@ -176,42 +207,26 @@ int RAMFUNCTION hal_flash_write(uint32_t address, const uint8_t *data, int len)
     return 0;
 }
 
-void RAMFUNCTION hal_flash_unlock(void)
-{
-    flash_wait_complete();
-    if ((FLASH_CR & FLASH_CR_LOCK) != 0) {
-        FLASH_KEY = FLASH_KEY1;
-        DMB();
-        FLASH_KEY = FLASH_KEY2;
-        DMB();
-        while ((FLASH_CR & FLASH_CR_LOCK) != 0)
-            ;
-    }
-}
-
-void RAMFUNCTION hal_flash_lock(void)
-{
-    flash_wait_complete();
-    if ((FLASH_CR & FLASH_CR_LOCK) == 0)
-        FLASH_CR |= FLASH_CR_LOCK;
-}
-
 
 int RAMFUNCTION hal_flash_erase(uint32_t address, int len)
 {
-    int start = -1, end = -1;
     uint32_t end_address;
     uint32_t p;
     if (len == 0)
         return -1;
+    address -= FLASHMEM_ADDRESS_SPACE;
     end_address = address + len - 1;
+    flash_wait_complete();
     for (p = address; p < end_address; p += FLASH_PAGE_SIZE) {
-        uint32_t reg = FLASH_CR & (~(FLASH_CR_PNB_MASK << FLASH_CR_PNB_SHIFT));
-        FLASH_CR = reg | ((p >> 12) << FLASH_CR_PNB_SHIFT) | FLASH_CR_PER | FLASH_CR_PG;
+        uint32_t reg;
+        flash_clear_errors();
+        reg = FLASH_CR & ~((FLASH_CR_PNB_MASK << FLASH_CR_PNB_SHIFT) | FLASH_CR_FSTPG | FLASH_CR_PG);
+        FLASH_CR = reg | ((p >> 12) << FLASH_CR_PNB_SHIFT) | FLASH_CR_PER;
         DMB();
         FLASH_CR |= FLASH_CR_STRT;
+        DMB();
         flash_wait_complete();
-        FLASH_CR &= ~(FLASH_CR_PER | FLASH_CR_PG);
+        FLASH_CR &= ~(FLASH_CR_PER);
     }
     return 0;
 }
@@ -310,7 +325,6 @@ void hal_prepare_boot(void)
 #ifdef SPI_FLASH
     spi_release();
 #endif
-    hal_flash_lock();
     clock_pll_off();
 }
 
diff --git a/lib/wolfssl b/lib/wolfssl
index a104caef..02327aee 160000
--- a/lib/wolfssl
+++ b/lib/wolfssl
@@ -1 +1 @@
-Subproject commit a104caef130ec58027f51d99b8254af83f5c744f
+Subproject commit 02327aee290596777f48f2db982650ee835f3da6