fix(core): improve framebuffer switching

[no changelog]

core/embed/trezorhal/bg_copy.h

@@ -7,9 +7,10 @@
 #include <stdint.h>
 
 /**
- * Wait for the data transfer completion
+ * Callback function invoked from the IRQ context
+ * when the transfer is complete
  */
-void bg_copy_wait(void);
+typedef void (*bg_copy_callback_t)(void);
 
 /**
  * Performs data copy from src to dst in the background. The destination is
@@ -19,8 +20,21 @@ void bg_copy_wait(void);
  * @param src source data address
  * @param dst destination data address
  * @param size size of data to be transferred in bytes
+ * @param callback optional callback to be called when the transfer is complete
+ */
+void bg_copy_start_const_out_8(const uint8_t *src, uint8_t *dst, size_t size,
+                               bg_copy_callback_t callback);
+
+/**
+ * Waits for the data transfer completion
+ */
+void bg_copy_wait(void);
+
+/**
+ * Immediately aborts the data transfer
+ *
+ * @note The callback will not be called
  */
-void bg_copy_start_const_out_8(const uint8_t *src, uint8_t *dst, size_t size);
 
 void bg_copy_abort(void);
 

core/embed/trezorhal/stm32f4/displays/st7789v.c

@@ -567,12 +567,12 @@ void DISPLAY_TE_INTERRUPT_HANDLER(void) {
   if (act_frame_buffer == 1) {
     bg_copy_start_const_out_8((uint8_t *)PhysFrameBuffer1,
                               (uint8_t *)DISPLAY_DATA_ADDRESS,
-                              DISPLAY_RESX * DISPLAY_RESY * 2);
+                              DISPLAY_RESX * DISPLAY_RESY * 2, NULL);
 
   } else {
     bg_copy_start_const_out_8((uint8_t *)PhysFrameBuffer0,
                               (uint8_t *)DISPLAY_DATA_ADDRESS,
-                              DISPLAY_RESX * DISPLAY_RESY * 2);
+                              DISPLAY_RESX * DISPLAY_RESY * 2, NULL);
   }
 
   pending_fb_switch = false;

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_driver.c

@@ -22,6 +22,7 @@
 #include <xdisplay.h>
 
 #include "display_fb.h"
+#include "display_internal.h"
 #include "display_io.h"
 #include "display_panel.h"
 
@@ -39,14 +40,8 @@
 #error "Incompatible display resolution"
 #endif
 
-// Display driver context.
-typedef struct {
-  // Current display orientation (0, 90, 180, 270)
-  int orientation_angle;
-} display_driver_t;
-
 // Display driver instance
-static display_driver_t g_display_driver;
+display_driver_t g_display_driver;
 
 void display_init(void) {
   display_driver_t* drv = &g_display_driver;
@@ -84,7 +79,8 @@ void display_reinit(void) {
 void display_finish_actions(void) {
 #ifdef XFRAMEBUFFER
 #ifndef BOARDLOADER
-  wait_for_fb_switch();
+  display_ensure_refreshed();
+  svc_disableIRQ(DISPLAY_TE_INTERRUPT_NUM);
 #endif
 #endif
 }
@@ -135,18 +131,4 @@ int display_get_orientation(void) {
   return drv->orientation_angle;
 }
 
-void display_wait_for_sync(void) {
-#ifdef DISPLAY_TE_PIN
-  uint32_t id = display_panel_identify();
-  if (id && (id != DISPLAY_ID_GC9307)) {
-    // synchronize with the panel synchronization signal
-    // in order to avoid visual tearing effects
-    while (GPIO_PIN_SET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN))
-      ;
-    while (GPIO_PIN_RESET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN))
-      ;
-  }
-#endif
-}
-
 void display_set_compatible_settings(void) { display_panel_set_big_endian(); }

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_fb.c

@@ -25,6 +25,7 @@
 #include STM32_HAL_H
 
 #include "display_fb.h"
+#include "display_internal.h"
 #include "display_io.h"
 #include "display_panel.h"
 #include "xdisplay.h"
@@ -41,173 +42,220 @@
 #error Framebuffer only supported on STM32U5 for now
 #endif
 
+// The following code supports only 1 or 2 frame buffers
+_Static_assert(FRAME_BUFFER_COUNT == 1 || FRAME_BUFFER_COUNT == 2);
+
 // Size of the physical frame buffer in bytes
 #define PHYSICAL_FRAME_BUFFER_SIZE (DISPLAY_RESX * DISPLAY_RESY * 2)
 
 // Physical frame buffers in internal SRAM memory.
 // Both frame buffers layes in the fixed addresses that
 // are shared between bootloaders and the firmware.
-__attribute__((section(".fb1")))
+static __attribute__((section(".fb1")))
 ALIGN_32BYTES(uint8_t physical_frame_buffer_0[PHYSICAL_FRAME_BUFFER_SIZE]);
-__attribute__((section(".fb2")))
-ALIGN_32BYTES(uint8_t physical_frame_buffer_1[PHYSICAL_FRAME_BUFFER_SIZE]);
-
-// The current frame buffer selector at fixed memory address
-// It's shared between bootloaders and the firmware
-__attribute__((section(".framebuffer_select"))) uint32_t current_frame_buffer =
-    0;
 
-void display_physical_fb_clear(void) {
-  memset(physical_frame_buffer_0, 0, sizeof(physical_frame_buffer_0));
-  memset(physical_frame_buffer_1, 0, sizeof(physical_frame_buffer_1));
-}
-
-#ifndef BOARDLOADER
-static volatile uint16_t pending_fb_switch = 0;
-static volatile uint32_t last_fb_update_time = 0;
+#if (FRAME_BUFFER_COUNT > 1)
+static __attribute__((section(".fb2")))
+ALIGN_32BYTES(uint8_t physical_frame_buffer_1[PHYSICAL_FRAME_BUFFER_SIZE]);
 #endif
 
-#ifndef BOARDLOADER
-void DISPLAY_TE_INTERRUPT_HANDLER(void) {
-  if (pending_fb_switch == 1) {
-    if (current_frame_buffer == 1) {
-      bg_copy_start_const_out_8((uint8_t *)physical_frame_buffer_1,
-                                (uint8_t *)DISPLAY_DATA_ADDRESS,
-                                DISPLAY_RESX * DISPLAY_RESY * 2);
-
-    } else {
-      bg_copy_start_const_out_8((uint8_t *)physical_frame_buffer_0,
-                                (uint8_t *)DISPLAY_DATA_ADDRESS,
-                                DISPLAY_RESX * DISPLAY_RESY * 2);
-    }
-    last_fb_update_time = HAL_GetTick();
-    pending_fb_switch = 2;
-  } else if (pending_fb_switch == 2) {
-    HAL_NVIC_DisableIRQ(DISPLAY_TE_INTERRUPT_NUM);
-    pending_fb_switch = 0;
+// Returns the pointer to the physical frame buffer (0.. FRAME_BUFFER_COUNT-1)
+// Returns NULL if the framebuffer index is out of range.
+static uint8_t *get_fb_ptr(uint32_t index) {
+  if (index == 0) {
+    return physical_frame_buffer_0;
+#if (FRAME_BUFFER_COUNT > 1)
+  } else if (index == 1) {
+    return physical_frame_buffer_1;
+#endif
   } else {
-    HAL_NVIC_DisableIRQ(DISPLAY_TE_INTERRUPT_NUM);
-    pending_fb_switch = 0;
+    return NULL;
   }
-  __HAL_GPIO_EXTI_CLEAR_FLAG(DISPLAY_TE_PIN);
 }
 
-static void copy_fb_to_display(const uint16_t *fb) {
-  for (int i = 0; i < DISPLAY_RESX * DISPLAY_RESY; i++) {
-    // 2 bytes per pixel because we're using RGB 5-6-5 format
-    ISSUE_PIXEL_DATA(fb[i]);
+void display_physical_fb_clear(void) {
+  for (int i = 0; i < FRAME_BUFFER_COUNT; i++) {
+    memset(get_fb_ptr(i), 0, PHYSICAL_FRAME_BUFFER_SIZE);
   }
 }
 
-void wait_for_fb_switch(void) {
-  if (is_mode_handler()) {
-    if (pending_fb_switch != 0) {
-      if (current_frame_buffer == 0) {
-        copy_fb_to_display((uint16_t *)physical_frame_buffer_1);
-      } else {
-        copy_fb_to_display((uint16_t *)physical_frame_buffer_0);
-      }
-      pending_fb_switch = 0;
-    }
-  } else {
-    while (pending_fb_switch != 0) {
-      __WFI();
-    }
-    bg_copy_wait();
+#ifndef BOARDLOADER
+
+// Callback called when the background copying is done
+// It's called from the IRQ context
+static void bg_copy_callback(void) {
+  display_driver_t *drv = &g_display_driver;
+
+  if (drv->queue.rix >= FRAME_BUFFER_COUNT) {
+    // This is an invalid state and we should never get here
+    return;
   }
+
+  drv->queue.entry[drv->queue.rix] = FB_STATE_EMPTY;
+  drv->queue.rix = (drv->queue.rix + 1) % FRAME_BUFFER_COUNT;
 }
-#endif
 
-static void switch_fb_manually(void) {
-  // sync with the panel refresh
-  while (GPIO_PIN_SET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN)) {
-  }
-  while (GPIO_PIN_RESET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN)) {
-  }
+// Interrupt routing handling TE signal
+void DISPLAY_TE_INTERRUPT_HANDLER(void) {
+  display_driver_t *drv = &g_display_driver;
 
-  if (current_frame_buffer == 0) {
-    current_frame_buffer = 1;
-    copy_fb_to_display((uint16_t *)physical_frame_buffer_1);
-    memcpy(physical_frame_buffer_0, physical_frame_buffer_1,
-           sizeof(physical_frame_buffer_0));
+  __HAL_GPIO_EXTI_CLEAR_FLAG(DISPLAY_TE_PIN);
 
-  } else {
-    current_frame_buffer = 0;
-    copy_fb_to_display((uint16_t *)physical_frame_buffer_0);
-    memcpy(physical_frame_buffer_1, physical_frame_buffer_0,
-           sizeof(physical_frame_buffer_1));
+  if (drv->queue.rix >= FRAME_BUFFER_COUNT) {
+    // This is an invalid state and we should never get here
+    return;
   }
-}
-
-#ifndef BOARDLOADER
-static void switch_fb_in_background(void) {
-  if (current_frame_buffer == 0) {
-    current_frame_buffer = 1;
 
-    memcpy(physical_frame_buffer_0, physical_frame_buffer_1,
-           sizeof(physical_frame_buffer_0));
+  switch (drv->queue.entry[drv->queue.rix]) {
+    case FB_STATE_EMPTY:
+    case FB_STATE_PREPARING:
+      // No new frame queued
+      break;
+
+    case FB_STATE_COPYING:
+      // Currently we are copying a data to the display.
+      // We need to wait for the next TE interrupt.
+      break;
+
+    case FB_STATE_READY:
+      // Now it's proper time to copy the data to the display
+      drv->queue.entry[drv->queue.rix] = FB_STATE_COPYING;
+      display_panel_set_window(0, 0, DISPLAY_RESX - 1, DISPLAY_RESY - 1);
+      bg_copy_start_const_out_8(get_fb_ptr(drv->queue.rix),
+                                (uint8_t *)DISPLAY_DATA_ADDRESS,
+                                PHYSICAL_FRAME_BUFFER_SIZE, bg_copy_callback);
 
-    pending_fb_switch = 1;
-    __HAL_GPIO_EXTI_CLEAR_FLAG(DISPLAY_TE_PIN);
-    svc_enableIRQ(DISPLAY_TE_INTERRUPT_NUM);
-  } else {
-    current_frame_buffer = 0;
-    memcpy(physical_frame_buffer_1, physical_frame_buffer_0,
-           sizeof(physical_frame_buffer_1));
+      // NOTE: when copying is done, this queue slot is marked empty
+      // (see bg_copy_callback())
+      break;
 
-    pending_fb_switch = 1;
-    __HAL_GPIO_EXTI_CLEAR_FLAG(DISPLAY_TE_PIN);
-    svc_enableIRQ(DISPLAY_TE_INTERRUPT_NUM);
+    default:
+      // This is an invalid state and we should never get here
+      break;
   }
 }
 #endif
 
 display_fb_info_t display_get_frame_buffer(void) {
-  void *addr;
+  display_driver_t *drv = &g_display_driver;
+
+  frame_buffer_state_t state;
+
+  // We have to wait if the buffer was passed for copying
+  // to the interrupt handler
+  do {
+    state = drv->queue.entry[drv->queue.wix];
+  } while (state == FB_STATE_READY || state == FB_STATE_COPYING);
+
+  if (state == FB_STATE_EMPTY) {
+    // First use of this buffer, copy the previous buffer into it
+#if (FRAME_BUFFER_COUNT > 1)
+    uint8_t *src = get_fb_ptr((FRAME_BUFFER_COUNT + drv->queue.wix - 1) %
+                              FRAME_BUFFER_COUNT);
+    uint8_t *dst = get_fb_ptr(drv->queue.wix);
+    memcpy(dst, src, PHYSICAL_FRAME_BUFFER_SIZE);
+#endif
+  };
 
-  if (current_frame_buffer == 0) {
-    addr = (void *)physical_frame_buffer_1;
-  } else {
-    addr = (void *)physical_frame_buffer_0;
-  }
+  drv->queue.entry[drv->queue.wix] = FB_STATE_PREPARING;
 
   display_fb_info_t fb = {
-      .ptr = addr,
+      .ptr = get_fb_ptr(drv->queue.wix),
       .stride = DISPLAY_RESX * sizeof(uint16_t),
   };
 
   return fb;
 }
 
+// Copies the frame buffer with the given index to the display
+static void copy_fb_to_display(uint8_t index) {
+  uint16_t *fb = (uint16_t *)get_fb_ptr(index);
+
+  if (fb != NULL) {
+    display_panel_set_window(0, 0, DISPLAY_RESX - 1, DISPLAY_RESY - 1);
+    for (int i = 0; i < DISPLAY_RESX * DISPLAY_RESY; i++) {
+      // 2 bytes per pixel because we're using RGB 5-6-5 format
+      ISSUE_PIXEL_DATA(fb[i]);
+    }
+  }
+}
+
+static void wait_for_te_signal(void) {
+  // sync with the panel refresh
+  while (GPIO_PIN_SET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN)) {
+  }
+  while (GPIO_PIN_RESET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN)) {
+  }
+}
+
 void display_refresh(void) {
-#ifndef BOARDLOADER
+  display_driver_t *drv = &g_display_driver;
+
+  if (drv->queue.entry[drv->queue.wix] != FB_STATE_PREPARING) {
+    // No refresh needed as the frame buffer is not in
+    // the state to be copied to the display
+    return;
+  }
 
+#ifndef BOARDLOADER
   if (is_mode_handler()) {
-    if (pending_fb_switch != 0) {
-      pending_fb_switch = 0;
-      bg_copy_abort();
+    // Disable scheduling of any new background copying
+    HAL_NVIC_DisableIRQ(DISPLAY_TE_INTERRUPT_NUM);
+    // Wait for next TE signal. During this time the
+    // display might be updated in the background
+    wait_for_te_signal();
+    // Stop any background copying even if it is not finished yet
+    bg_copy_abort();
+    // Copy the frame buffer to the display manually
+    copy_fb_to_display(drv->queue.wix);
+    // Reset the buffer queue so we can eventually continue
+    // safely in thread mode
+    drv->queue.wix = 0;
+    drv->queue.rix = 0;
+    for (int i = 0; i < FRAME_BUFFER_COUNT; i++) {
+      drv->queue.entry[i] = FB_STATE_EMPTY;
     }
-    display_panel_set_window(0, 0, DISPLAY_RESX - 1, DISPLAY_RESY - 1);
-    switch_fb_manually();
+    // Enable normal processing again
+    HAL_NVIC_EnableIRQ(DISPLAY_TE_INTERRUPT_NUM);
   } else {
-    wait_for_fb_switch();
-    display_panel_set_window(0, 0, DISPLAY_RESX - 1, DISPLAY_RESY - 1);
-    switch_fb_in_background();
+    // Mark the buffer ready to switch to
+    drv->queue.entry[drv->queue.wix] = FB_STATE_READY;
+    drv->queue.wix = (drv->queue.wix + 1) % FRAME_BUFFER_COUNT;
   }
-#else
-  display_panel_set_window(0, 0, DISPLAY_RESX - 1, DISPLAY_RESY - 1);
-  switch_fb_manually();
+
+#else  // BOARDLOADER
+  wait_for_te_signal();
+  copy_fb_to_display(drv->queue.wix);
+  drv->queue.entry[drv->queue.wix] = FB_STATE_EMPTY;
 #endif
 }
 
 void display_ensure_refreshed(void) {
 #ifndef BOARDLOADER
+  display_driver_t *drv = &g_display_driver;
+
   if (!is_mode_handler()) {
-    wait_for_fb_switch();
-    // the update time is collected after starting the BG copy, then we need to
-    // wait: for the bg copy to finish and for at least one full refresh cycle
-    // before we can consider the display fully redrawn
-    while (HAL_GetTick() - last_fb_update_time < 40) {
+    bool copy_pending;
+
+    // Wait until all frame buffers are written to the display
+    //  so we can be sure there's not scheduled or pending
+    // background copying
+    do {
+      copy_pending = false;
+      for (int i = 0; i < FRAME_BUFFER_COUNT; i++) {
+        frame_buffer_state_t state = drv->queue.entry[i];
+        if (state == FB_STATE_READY || state == FB_STATE_COPYING) {
+          copy_pending = true;
+          break;
+        }
+      }
+      __WFI();
+    } while (copy_pending);
+
+    // Wait until the display is fully refreshed
+    // (TE signal is low when the display is updating)
+    while (GPIO_PIN_RESET ==
+           HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN)) {
       __WFI();
     }
   }

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_fb.h

@@ -29,8 +29,6 @@ void display_physical_fb_clear(void);
 
 void display_ensure_refreshed(void);
 
-void wait_for_fb_switch(void);
-
 #endif  // XFRAMEBUFFER
 
 #endif  // TREZORHAL_DISPLAY_FB_H

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_internal.h

@@ -0,0 +1,59 @@
+#ifndef TREZORHAL_DISPLAY_INTERNAL_H
+#define TREZORHAL_DISPLAY_INTERNAL_H
+
+#include <stdint.h>
+
+#ifdef XFRAMEBUFFER
+
+// Number of frame buffers used (1 or 2)
+// If 1 buffer is selected, some animations may not
+// be so smooth but the memory usage is lower.
+#define FRAME_BUFFER_COUNT 2
+
+// Each frame buffer can be in one of the following states:
+typedef enum {
+  // The frame buffer is empty and can be written to
+  FB_STATE_EMPTY = 0,
+  // The frame buffer pass passed to application
+  FB_STATE_PREPARING = 1,
+  // The frame buffer was written to and is ready
+  // to be copied to the display
+  FB_STATE_READY = 2,
+  // The frame buffer is currently being copied to
+  // the display
+  FB_STATE_COPYING = 3,
+
+} frame_buffer_state_t;
+
+typedef struct {
+  // Queue entries
+  volatile frame_buffer_state_t entry[FRAME_BUFFER_COUNT];
+  // Read index
+  // (accessed & updated in the context of the interrupt handlers
+  uint8_t rix;
+  // Write index
+  // (accessed & updated in context of the main thread)
+  uint8_t wix;
+
+} frame_buffer_queue_t;
+
+#endif  // XFRAMEBUFFER
+
+// Display driver state
+typedef struct {
+#ifdef XFRAMEBUFFER
+  // Framebuffer queue
+  // (accessed & updated in the context of the main thread
+  // and the interrupt context)
+  volatile frame_buffer_queue_t queue;
+#endif
+
+  // Current display orientation (0, 90, 180, 270)
+  int orientation_angle;
+
+} display_driver_t;
+
+// Display driver instance
+extern display_driver_t g_display_driver;
+
+#endif  // TREZORHAL_DISPLAY_INTERNAL_H

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_io.c

@@ -22,6 +22,7 @@
 
 #include "display_io.h"
 #include "irq.h"
+#include "supervise.h"
 
 __IO DISP_MEM_TYPE *const DISPLAY_CMD_ADDRESS =
     (__IO DISP_MEM_TYPE *const)((uint32_t)DISPLAY_MEMORY_BASE);
@@ -35,7 +36,6 @@ void display_io_init_gpio(void) {
   __HAL_RCC_GPIOA_CLK_ENABLE();
   __HAL_RCC_GPIOC_CLK_ENABLE();
   __HAL_RCC_GPIOD_CLK_ENABLE();
-  __HAL_RCC_FMC_CLK_ENABLE();
 
   GPIO_InitTypeDef GPIO_InitStructure;
 
@@ -87,6 +87,8 @@ void display_io_init_gpio(void) {
 }
 
 void display_io_init_fmc(void) {
+  __HAL_RCC_FMC_CLK_ENABLE();
+
   // Reference UM1725 "Description of STM32F4 HAL and LL drivers",
   // section 64.2.1 "How to use this driver"
   SRAM_HandleTypeDef external_display_data_sram = {0};
@@ -141,5 +143,6 @@ void display_io_init_te_interrupt(void) {
 
   // setup interrupt for tearing effect pin
   HAL_NVIC_SetPriority(DISPLAY_TE_INTERRUPT_NUM, IRQ_PRI_DMA, 0);
+  svc_enableIRQ(DISPLAY_TE_INTERRUPT_NUM);
 }
 #endif

core/embed/trezorhal/stm32f4/xdisplay/st-7789/display_nofb.c

@@ -25,7 +25,23 @@
 #include "display_panel.h"
 
 void display_refresh(void) {
-  // if the framebuffer is not used the implementation is empty
+  // If the framebuffer is not used the, we do not need
+  // to refresh the display explicitly as we write the data
+  // directly to the display internal RAM.
+}
+
+void display_wait_for_sync(void) {
+#ifdef DISPLAY_TE_PIN
+  uint32_t id = display_panel_identify();
+  if (id && (id != DISPLAY_ID_GC9307)) {
+    // synchronize with the panel synchronization signal
+    // in order to avoid visual tearing effects
+    while (GPIO_PIN_SET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN))
+      ;
+    while (GPIO_PIN_RESET == HAL_GPIO_ReadPin(DISPLAY_TE_PORT, DISPLAY_TE_PIN))
+      ;
+  }
+#endif
 }
 
 static inline void set_window(const gfx_bitblt_t* bb) {

core/embed/trezorhal/stm32u5/bg_copy.c

@@ -9,6 +9,7 @@ static volatile uint32_t dma_transfer_remaining = 0;
 static volatile uint32_t dma_data_transferred = 0;
 static void *data_src = NULL;
 static void *data_dst = NULL;
+static bg_copy_callback_t bg_copy_callback = NULL;
 static DMA_HandleTypeDef DMA_Handle = {0};
 
 void HAL_DMA_XferCpltCallback(DMA_HandleTypeDef *hdma) {
@@ -46,6 +47,10 @@ void GPDMA1_Channel0_IRQHandler(void) {
     HAL_NVIC_DisableIRQ(GPDMA1_Channel0_IRQn);
     data_src = NULL;
     data_dst = NULL;
+
+    if (bg_copy_callback != NULL) {
+      bg_copy_callback();
+    }
   }
 }
 
@@ -57,12 +62,14 @@ void bg_copy_wait(void) {
   }
 }
 
-void bg_copy_start_const_out_8(const uint8_t *src, uint8_t *dst, size_t size) {
+void bg_copy_start_const_out_8(const uint8_t *src, uint8_t *dst, size_t size,
+                               bg_copy_callback_t callback) {
   uint32_t data_to_send = size > MAX_DATA_SIZE ? MAX_DATA_SIZE : size;
   dma_transfer_remaining = size;
   dma_data_transferred = 0;
   data_src = (void *)src;
   data_dst = (void *)dst;
+  bg_copy_callback = callback;
 
   // setup DMA for data copy to constant output address
 

core/embed/trezorhal/xdisplay_legacy.c

@@ -40,4 +40,8 @@ int display_backlight(int level) {
   }
 }
 
-void display_sync(void) {}
+void display_sync(void) {
+#ifndef XFRAMEBUFFER
+  display_wait_for_sync();
+#endif
+}