chore(core): use cortex-m33 stack protection

core/SConscript.bootloader_emu

@@ -127,6 +127,7 @@ SOURCE_BOOTLOADER = [
 SOURCE_TREZORHAL = [
     'embed/trezorhal/unix/boot_args.c',
     'embed/trezorhal/unix/display-unix.c',
+    'embed/trezorhal/unix/fault_handlers.c',
     'embed/trezorhal/unix/flash.c',
     'embed/trezorhal/unix/common.c',
     'embed/trezorhal/unix/touch/touch.c',

core/embed/boardloader/main.c

@@ -24,6 +24,7 @@
 #include "common.h"
 #include "compiler_traits.h"
 #include "display.h"
+#include "fault_handlers.h"
 #include "flash.h"
 #include "image.h"
 #include "model.h"
@@ -256,6 +257,8 @@ int main(void) {
 
   mpu_config_boardloader();
 
+  fault_handlers_init();
+
 #ifdef USE_SDRAM
   sdram_init();
 #endif

core/embed/boardloader/memory_stm32u5a.ld

@@ -14,7 +14,7 @@ MEMORY {
 }
 
 main_stack_base = ORIGIN(SRAM2) + SIZEOF(.stack); /* 8-byte aligned full descending stack */
-_sstack = ORIGIN(SRAM2) + 0x100;
+_sstack = ORIGIN(SRAM2);
 _estack = main_stack_base;
 
 /* used by the startup code to populate variables used by the C code */
@@ -80,7 +80,7 @@ SECTIONS {
   } >SRAM1
 
   .stack : ALIGN(8) {
-    . = 16K + 0x100; /* Exactly 16K allocated for stack. Overflow causes MemManage fault (when using MPU). */
+    . = 16K; /* Overflow causes UsageFault */
   } >SRAM2
 
   .sensitive : ALIGN(8) {

core/embed/boardloader/startup_stm32u5.s

@@ -5,6 +5,11 @@
   .global reset_handler
   .type reset_handler, STT_FUNC
 reset_handler:
+  // set the stack protection
+  ldr r0, =_sstack
+  add r0, r0, #16        // padding
+  msr MSPLIM, r0
+
   bl SystemInit
 
   // read the first rng data and save it

core/embed/bootloader/main.c

@@ -23,6 +23,7 @@
 #include "boot_args.h"
 #include "common.h"
 #include "display.h"
+#include "fault_handlers.h"
 #include "flash.h"
 #include "image.h"
 #include "lowlevel.h"
@@ -423,6 +424,8 @@ int bootloader_main(void) {
 
   mpu_config_bootloader();
 
+  fault_handlers_init();
+
 #ifdef TREZOR_EMULATOR
   // wait a bit so that the empty lock icon is visible
   // (on a real device, we are waiting for touch init which takes longer)

core/embed/bootloader/memory_stm32u5a.ld

@@ -14,7 +14,7 @@ MEMORY {
 }
 
 main_stack_base = ORIGIN(SRAM2) + SIZEOF(.stack); /* 8-byte aligned full descending stack */
-_sstack = ORIGIN(SRAM2) + 0x100;
+_sstack = ORIGIN(SRAM2);
 _estack = main_stack_base;
 
 /* used by the startup code to populate variables used by the C code */
@@ -81,7 +81,7 @@ SECTIONS {
   } >SRAM1
 
   .stack : ALIGN(8) {
-    . = 16K + 0x100; /* Exactly 16K allocated for stack. Overflow causes MemManage fault (when using MPU). */
+    . = 16K; /* Overflow causes UsageFault */
   } >SRAM2
 
   .sensitive : ALIGN(512) {

core/embed/bootloader/startup_stm32u5.s

@@ -5,6 +5,11 @@
   .global reset_handler
   .type reset_handler, STT_FUNC
 reset_handler:
+  // set the stack protection
+  ldr r0, =_sstack
+  add r0, r0, #16        // padding
+  msr MSPLIM, r0
+
   // setup environment for subsequent stage of code
   ldr r2, =0             // r2 - the word-sized value to be written
 

core/embed/firmware/main.c

@@ -41,6 +41,7 @@
 #include "common.h"
 #include "compiler_traits.h"
 #include "display.h"
+#include "fault_handlers.h"
 #include "flash.h"
 #include "image.h"
 #include "memzero.h"
@@ -95,6 +96,19 @@
 // from util.s
 extern void shutdown_privileged(void);
 
+volatile int aa = 1;
+
+int test_stk(int i) {
+  int a[4096];
+  a[40] = i;
+  a[50] = i;
+  memset(a, 0, sizeof(a));
+  for (i = 0; i < 4096; i++) {
+    aa = a[i];
+  }
+  return 0;
+}
+
 int main(void) {
   random_delays_init();
 
@@ -143,10 +157,7 @@ int main(void) {
   dma2d_init();
 #endif
 
-#if !PRODUCTION
-  // enable BUS fault and USAGE fault handlers
-  SCB->SHCSR |= (SCB_SHCSR_USGFAULTENA_Msk | SCB_SHCSR_BUSFAULTENA_Msk);
-#endif
+  fault_handlers_init();
 
 #if defined TREZOR_MODEL_T
   set_core_clock(CLOCK_180_MHZ);
@@ -193,6 +204,8 @@ int main(void) {
   drop_privileges();
 #endif
 
+  // aa = test_stk(4096);
+
 #ifdef USE_SECP256K1_ZKP
   ensure(sectrue * (zkp_context_init() == 0), NULL);
 #endif

core/embed/firmware/memory_DISC2.ld

@@ -14,7 +14,7 @@ MEMORY {
 }
 
 main_stack_base = ORIGIN(SRAM2) + SIZEOF(.stack); /* 8-byte aligned full descending stack */
-_sstack = ORIGIN(SRAM2) + 0x100;
+_sstack = ORIGIN(SRAM2);
 _estack = main_stack_base;
 
 /* used by the startup code to populate variables used by the C code */
@@ -97,7 +97,7 @@ SECTIONS {
   } >SRAM1
 
   .stack : ALIGN(8) {
-    . = 16K + 0x100; /* Exactly 16K allocated for stack. Overflow causes MemManage fault (when using MPU). */
+    . = 16K; /* Overflow causes UsageFault */
   } >SRAM2
 
   .sensitive : ALIGN(512) {

core/embed/firmware/startup_stm32u5.S

@@ -5,6 +5,11 @@
   .global reset_handler
   .type reset_handler, STT_FUNC
 reset_handler:
+  // set the stack protection
+  ldr r0, =_sstack
+  add r0, r0, #16        // padding
+  msr MSPLIM, r0
+
   // setup environment for subsequent stage of code
   ldr r2, =0             // r2 - the word-sized value to be written
 

core/embed/prodtest/main.c

@@ -26,6 +26,7 @@
 #include "button.h"
 #include "common.h"
 #include "display.h"
+#include "fault_handlers.h"
 #include "flash.h"
 #include "i2c.h"
 #include "model.h"
@@ -569,6 +570,8 @@ int main(void) {
 #endif
 
   mpu_config_prodtest();
+  fault_handlers_init();
+
   drop_privileges();
 
   display_clear();

core/embed/prodtest/memory_stm32u5a.ld

@@ -14,7 +14,7 @@ MEMORY {
 }
 
 main_stack_base = ORIGIN(SRAM2) + SIZEOF(.stack); /* 8-byte aligned full descending stack */
-_sstack = ORIGIN(SRAM2) + 0x100;
+_sstack = ORIGIN(SRAM2);
 _estack = main_stack_base;
 
 /* used by the startup code to populate variables used by the C code */
@@ -96,7 +96,7 @@ SECTIONS {
   } >SRAM1
 
   .stack : ALIGN(8) {
-    . = 16K + 0x100; /* Exactly 16K allocated for stack. Overflow causes MemManage fault (when using MPU). */
+    . = 16K; /* Overflow causes UsageFault */
   } >SRAM2
 
   .sensitive : ALIGN(512) {

core/embed/prodtest/startup_stm32u5.s

@@ -5,6 +5,11 @@
   .global reset_handler
   .type reset_handler, STT_FUNC
 reset_handler:
+  // set the stack protection
+  ldr r0, =_sstack
+  add r0, r0, #16        // padding
+  msr MSPLIM, r0
+
   // setup environment for subsequent stage of code
   ldr r2, =0             // r2 - the word-sized value to be written
 

core/embed/reflash/memory_stm32u5a.ld

@@ -14,7 +14,7 @@ MEMORY {
 }
 
 main_stack_base = ORIGIN(SRAM2) + SIZEOF(.stack); /* 8-byte aligned full descending stack */
-_sstack = ORIGIN(SRAM2) + 0x100;
+_sstack = ORIGIN(SRAM2);
 _estack = main_stack_base;
 
 /* used by the startup code to populate variables used by the C code */
@@ -97,7 +97,7 @@ SECTIONS {
   } >SRAM1
 
   .stack : ALIGN(8) {
-    . = 16K + 0x100; /* Exactly 16K allocated for stack. Overflow causes MemManage fault (when using MPU). */
+    . = 16K + 0x100; /* Overflow causes UsageFault */
   } >SRAM2
 
   .sensitive : ALIGN(512) {

core/embed/reflash/startup_stm32u5.s

@@ -5,6 +5,11 @@
   .global reset_handler
   .type reset_handler, STT_FUNC
 reset_handler:
+  // set the stack protection
+  ldr r0, =_sstack
+  add r0, r0, #16        // padding
+  msr MSPLIM, r0
+
   // setup environment for subsequent stage of code
   ldr r2, =0             // r2 - the word-sized value to be written
 

core/embed/trezorhal/fault_handlers.h

@@ -0,0 +1,7 @@
+#ifndef TREZORHAL_FAULT_HANDLERS_H
+#define TREZORHAL_FAULT_HANDLERS_H
+
+// Initializes and enables fault handlers
+void fault_handlers_init(void);
+
+#endif  // TREZORHAL_FAULT_HANDLERS_H

core/embed/trezorhal/stm32f4/fault_handlers.c

@@ -1,5 +1,10 @@
 #include "common.h"
 
+void fault_handlers_init(void) {
+  // Enable BUS fault and USAGE fault handlers
+  SCB->SHCSR |= (SCB_SHCSR_USGFAULTENA_Msk | SCB_SHCSR_BUSFAULTENA_Msk);
+}
+
 void HardFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(HF)"); }
 
 void MemManage_Handler_MM(void) { error_shutdown("INTERNAL ERROR", "(MM)"); }

core/embed/trezorhal/stm32u5/fault_handlers.c

@@ -1,14 +1,28 @@
 #include "common.h"
 
-void HardFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(HF)"); }
+void fault_handlers_init(void) {
+  // Enable BUS fault and USAGE fault handlers
+  SCB->SHCSR |= (SCB_SHCSR_USGFAULTENA_Msk | SCB_SHCSR_BUSFAULTENA_Msk);
+}
 
-void MemManage_Handler_MM(void) { error_shutdown("INTERNAL ERROR", "(MM)"); }
+void HardFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(HF)"); }
 
-void MemManage_Handler_SO(void) { error_shutdown("INTERNAL ERROR", "(SO)"); }
+void MemManage_Handler(void) { error_shutdown("INTERNAL ERROR", "(MM)"); }
 
 void BusFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(BF)"); }
 
-void UsageFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(UF)"); }
+void UsageFault_Handler(void) {
+  if (SCB->CFSR & SCB_CFSR_STKOF_Msk) {
+    // Stack overflow
+    extern uint8_t _estack;  // linker script symbol
+    // Fix stack pointer
+    __set_MSP((uint32_t)&_estack);
+    error_shutdown("INTERNAL ERROR", "(SO)");
+  } else {
+    // Other error
+    error_shutdown("INTERNAL ERROR", "(UF)");
+  }
+}
 
 void SecureFault_Handler(void) { error_shutdown("INTERNAL ERROR", "(SF)"); }
 

core/embed/trezorhal/stm32u5/limited_util.s

@@ -130,16 +130,4 @@ shutdown_privileged:
   ldr r0, =0
   b . // loop forever
 
-  .global MemManage_Handler
-  .type MemManage_Handler, STT_FUNC
-MemManage_Handler:
-  ldr r2, =_sstack
-  mrs r1, msp
-  ldr r0, =_estack
-  msr msp, r0
-  cmp r1, r2
-  IT lt
-  bllt MemManage_Handler_SO
-  bl MemManage_Handler_MM
-
   .end

core/embed/trezorhal/stm32u5/mpu.c

@@ -114,8 +114,7 @@ static void mpu_set_attributes() {
 #define SIZE_128K (128 * 1024)
 #define SIZE_192K (192 * 1024)
 #define SIZE_320K (320 * 1024)
-#define SIZE_768K (768 * 1024)
-#define SIZE_1728K ((832 * 2 + 64) * 1024)
+#define SIZE_2496K (2496 * 1024)
 #define SIZE_3776K ((4096 - 320) * 1024)
 #define SIZE_3904K ((4096 - 192) * 1024)
 #define SIZE_4032K ((4096 - 64) * 1024)
@@ -131,10 +130,10 @@ void mpu_config_boardloader() {
   SET_REGION( 0, FLASH_BASE_S,             SIZE_16K,           FLASH_DATA,  YES,    NO ); // Secret
   SET_REGION( 1, FLASH_BASE_S + SIZE_16K,  SIZE_48K,           FLASH_CODE,   NO,    NO ); // Boardloader code
   SET_REGION( 2, FLASH_BASE_S + SIZE_64K,  SIZE_4032K,         FLASH_DATA,  YES,    NO ); // Bootloader + Storage + Firmware
-  SET_REGION( 3, SRAM1_BASE_S,             SIZE_768K,          SRAM,        YES,    NO ); // SRAM1
-  SET_REGION( 4, SRAM2_BASE_S + 0x100,     SIZE_1728K - 0x100, SRAM,        YES,    NO ); // SRAM2/3/5 + stack guard
-  SET_REGION( 5, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,    NO ); // Frame buffer
-  SET_REGION( 6, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,    NO ); // Peripherals
+  SET_REGION( 3, SRAM1_BASE_S,             SIZE_2496K,         SRAM,        YES,    NO ); // SRAM1/2/3/5
+  SET_REGION( 4, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,    NO ); // Frame buffer
+  SET_REGION( 5, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,    NO ); // Peripherals
+  DIS_REGION( 6 );
   DIS_REGION( 7 );
   // clang-format on
   HAL_MPU_Enable(LL_MPU_CTRL_HARDFAULT_NMI);
@@ -148,11 +147,11 @@ void mpu_config_bootloader() {
   SET_REGION( 0, FLASH_BASE_S,             SIZE_64K,           FLASH_DATA,  YES,    NO ); // Secret + Boardloader
   SET_REGION( 1, FLASH_BASE_S + SIZE_64K,  SIZE_128K,          FLASH_CODE,  NO,     NO ); // Bootloader code
   SET_REGION( 2, FLASH_BASE_S + SIZE_192K, SIZE_3904K,         FLASH_DATA,  YES,    NO ); // Storage + Firmware
-  SET_REGION( 3, SRAM1_BASE_S,             SIZE_768K,          SRAM,        YES,    NO ); // SRAM1
-  SET_REGION( 4, SRAM2_BASE_S + 0x100,     SIZE_1728K - 0x100, SRAM,        YES,    NO ); // SRAM2/3/5 + stack guard
-  SET_REGION( 5, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,    NO ); // Frame buffer
-  SET_REGION( 6, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,    NO ); // Peripherals
-  SET_REGION( 7, FLASH_OTP_BASE,           FLASH_OTP_SIZE,     FLASH_DATA,  YES,    NO ); // OTP
+  SET_REGION( 3, SRAM1_BASE_S,             SIZE_2496K,         SRAM,        YES,    NO ); // SRAM1/2/3/5
+  SET_REGION( 4, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,    NO ); // Frame buffer
+  SET_REGION( 5, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,    NO ); // Peripherals
+  SET_REGION( 6, FLASH_OTP_BASE,           FLASH_OTP_SIZE,     FLASH_DATA,  YES,    NO ); // OTP
+  DIS_REGION( 7 );
   // clang-format on
   HAL_MPU_Enable(LL_MPU_CTRL_HARDFAULT_NMI);
 }
@@ -164,11 +163,11 @@ void mpu_config_firmware() {
   //   REGION    ADDRESS                   SIZE                TYPE       WRITE   UNPRIV
   SET_REGION( 0, FLASH_BASE_S + SIZE_192K, SIZE_128K,          FLASH_DATA,  YES,   YES ); // Storage
   SET_REGION( 1, FLASH_BASE_S + SIZE_320K, SIZE_3776K,         FLASH_CODE,   NO,   YES ); // Firmware
-  SET_REGION( 2, SRAM1_BASE_S,             SIZE_768K,          SRAM,        YES,   YES ); // SRAM1
-  SET_REGION( 3, SRAM2_BASE_S + 0x100,     SIZE_1728K - 0x100, SRAM,        YES,   YES ); // SRAM2/3/5 + stack guard
-  SET_REGION( 4, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,   YES ); // Frame buffer
-  SET_REGION( 5, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,   YES ); // Peripherals
-  SET_REGION( 6, FLASH_OTP_BASE,           FLASH_OTP_SIZE,     FLASH_DATA,  YES,   YES ); // OTP
+  SET_REGION( 2, SRAM1_BASE_S,             SIZE_2496K,         SRAM,        YES,   YES ); // SRAM1/2/3/5
+  SET_REGION( 3, GFXMMU_BUFFERS_S,         SIZE_16M,           SRAM,        YES,   YES ); // Frame buffer
+  SET_REGION( 4, PERIPH_BASE_S,            SIZE_256M,          PERIPHERAL,  YES,   YES ); // Peripherals
+  SET_REGION( 5, FLASH_OTP_BASE,           FLASH_OTP_SIZE,     FLASH_DATA,  YES,   YES ); // OTP
+  DIS_REGION( 6 );
   DIS_REGION( 7 );
   // clang-format on
   HAL_MPU_Enable(LL_MPU_CTRL_HARDFAULT_NMI);