feature(dcd_dwc2): Added cache synchronization

Author: roma-jam

src/portable/synopsys/dwc2/dcd_dwc2.c

@@ -67,6 +67,16 @@ static bool _sof_en;
 //--------------------------------------------------------------------
 // DMA
 //--------------------------------------------------------------------
+// When DMA requires cache synchronization for memory
+// Data synchronization: cache to memory
+#ifndef dsync_c2m
+#define dsync_c2m(_addr, _size)
+#endif // dsync_c2m
+
+// Data synchronization: memory to cache
+#ifndef dsync_m2c
+#define dsync_m2c(_addr, _size)
+#endif // dsync_m2c
 
 TU_ATTR_ALWAYS_INLINE static inline bool dma_device_enabled(const dwc2_regs_t* dwc2) {
   (void) dwc2;
@@ -359,7 +369,7 @@ static void edpt_schedule_packets(uint8_t rhport, uint8_t const epnum, uint8_t c
 
     if(dma_device_enabled(dwc2)) {
       dep->diepdma = (uintptr_t)xfer->buffer;
-
+      dsync_c2m(xfer->buffer, total_bytes);
       // For ISO endpoint set correct odd/even bit for next frame.
       if ((dep->diepctl & DIEPCTL_EPTYP) == DIEPCTL_EPTYP_0 && (XFER_CTL_BASE(epnum, dir))->interval == 1) {
         // Take odd/even bit from frame counter.
@@ -397,6 +407,7 @@ static void edpt_schedule_packets(uint8_t rhport, uint8_t const epnum, uint8_t c
 
     if(dma_device_enabled(dwc2)) {
       dep->doepdma = (uintptr_t)xfer->buffer;
+      dsync_c2m(xfer->buffer, total_bytes);
     }
 
     dep->doepctl |= DOEPCTL_EPENA | DOEPCTL_CNAK;
@@ -775,6 +786,7 @@ static void handle_epout_irq(uint8_t rhport) {
 
         if(dma_device_enabled(dwc2)) {
           dma_setup_prepare(rhport);
+          dsync_m2c((uint8_t*) _setup_packet, sizeof(_setup_packet));
         }
 
         dcd_event_setup_received(rhport, (uint8_t*) _setup_packet, true);
@@ -801,7 +813,7 @@ static void handle_epout_irq(uint8_t rhport) {
               if(epnum == 0 && xfer->total_len == 0) {
                 dma_setup_prepare(rhport);
               }
-
+              dsync_m2c(xfer->buffer, xfer->total_len);
               dcd_event_xfer_complete(rhport, epnum, xfer->total_len, XFER_RESULT_SUCCESS, true);
             }
           } else {

src/portable/synopsys/dwc2/dwc2_esp32.h

@@ -38,6 +38,17 @@
 #include "esp_intr_alloc.h"
 #include "soc/periph_defs.h"
 #include "soc/usb_wrap_struct.h"
+#include "esp_log.h"
+
+#if TU_CHECK_MCU(OPT_MCU_ESP32P4)
+#if (CFG_TUD_DWC2_DMA && SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE)
+#include "sdkconfig.h"
+#include "hal/cache_hal.h"
+#include "esp_cache.h"
+#define DWC2_ENABLE_MEM_CACHE         1
+#endif // SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE
+#endif // OPT_MCU_ESP32P4
+
 
 #if TU_CHECK_MCU(OPT_MCU_ESP32S2, OPT_MCU_ESP32S3)
 #define DWC2_FS_REG_BASE   0x60080000UL
@@ -111,6 +122,42 @@ TU_ATTR_ALWAYS_INLINE static inline void dwc2_phy_update(dwc2_regs_t* dwc2, uint
   // maybe usb_utmi_hal_disable()
 }
 
+// MCU specific cache synchronization call
+TU_ATTR_ALWAYS_INLINE static inline void dwc2_dcd_sync_cache_to_memory(void *addr, size_t size) {
+#if DWC2_ENABLE_MEM_CACHE
+  ESP_EARLY_LOGV("dwc2_esp32", "cache to mem sync, addr 0x%"PRIx32", size %d", (uintptr_t)addr, size);
+  int flags = ESP_CACHE_MSYNC_FLAG_DIR_C2M | ESP_CACHE_MSYNC_FLAG_UNALIGNED;
+  if (addr != NULL && size) {
+    esp_err_t ret = esp_cache_msync(addr, size, flags);
+    assert(ret == ESP_OK);
+  }
+#else
+  (void) addr;
+  (void) size;
+  // nothing to do
+#endif // DWC2_ENABLE_MEM_CACHE
+}
+
+TU_ATTR_ALWAYS_INLINE static inline void dwc2_dcd_sync_memory_to_cache(void *addr, size_t size) {
+#if DWC2_ENABLE_MEM_CACHE
+  ESP_EARLY_LOGV("dwc2", "mem to cache sync, addr 0x%"PRIx32", size %d", (uintptr_t)addr, size);
+  int flags = ESP_CACHE_MSYNC_FLAG_DIR_M2C;
+  if (addr != NULL && size) {
+    // TODO: size should be multiply of CONFIG_CACHE_L1_CACHE_LINE_SIZE?
+    size = (size < CONFIG_CACHE_L1_CACHE_LINE_SIZE)? CONFIG_CACHE_L1_CACHE_LINE_SIZE : size;
+    esp_err_t ret = esp_cache_msync(addr, size, flags);
+    assert(ret == ESP_OK);
+  }
+#else
+  (void) addr;
+  (void) size;
+  // nothing to do
+#endif // DWC2_ENABLE_MEM_CACHE
+}
+
+#define dsync_c2m(_addr, _size)   dwc2_dcd_sync_cache_to_memory((_addr), (_size))
+#define dsync_m2c(_addr, _size)   dwc2_dcd_sync_memory_to_cache((_addr), (_size))
+
 #ifdef __cplusplus
 }
 #endif