From b8d31a59eec6b27b900a6d51d602a6d6f47ac4d4 Mon Sep 17 00:00:00 2001
From: Roman Leonov <roman.leonov@espressif.com>
Date: Tue, 19 Nov 2024 11:28:04 +0100
Subject: [PATCH] feature(dcd_dwc2): Added cache synchronization

---
 src/portable/synopsys/dwc2/dcd_dwc2.c   | 27 ++++++++++++++++++++++++-
 src/portable/synopsys/dwc2/dwc2_esp32.h | 27 +++++++++++++++++++++++++
 2 files changed, 53 insertions(+), 1 deletion(-)

diff --git a/src/portable/synopsys/dwc2/dcd_dwc2.c b/src/portable/synopsys/dwc2/dcd_dwc2.c
index 2b3ef096f..1addd28f9 100644
--- a/src/portable/synopsys/dwc2/dcd_dwc2.c
+++ b/src/portable/synopsys/dwc2/dcd_dwc2.c
@@ -47,7 +47,20 @@
 // MACRO TYPEDEF CONSTANT ENUM
 //--------------------------------------------------------------------+
 
+#ifdef DWC2_MEM_CACHE_LINE_SIZE
+CFG_TUD_MEM_SECTION struct {
+  union {
+    uint32_t data[2];
+    uint8_t buffer[DWC2_MEM_CACHE_LINE_SIZE];
+  };
+} _cache_aligned_setup_packet;
+
+#define _setup_packet                         _cache_aligned_setup_packet.data
+#define _sizeof_setup_packet()                DWC2_MEM_CACHE_LINE_SIZE
+#else
 static CFG_TUD_MEM_SECTION TU_ATTR_ALIGNED(4) uint32_t _setup_packet[2];
+#define _sizeof_setup_packet()                sizeof(_setup_packet)
+#endif // DWC2_MEM_CACHE_LINE_SIZE
 
 typedef struct {
   uint8_t* buffer;
@@ -348,6 +361,11 @@ static void edpt_schedule_packets(uint8_t rhport, const uint8_t epnum, const uin
 
   const bool is_dma = dma_device_enabled(dwc2);
   if(is_dma) {
+    if (dir == TUSB_DIR_IN && total_bytes != 0) {
+      // CACHE HINT
+      // The xfer->buffer has new data for Host, move it to memory for DMA to transfer it
+      dcd_dcache_clean(xfer->buffer, total_bytes);
+    }
     dep->diepdma = (uintptr_t) xfer->buffer;
   }
 
@@ -847,6 +865,11 @@ static void handle_epout_dma(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepi
 
   if (doepint_bm.setup_phase_done) {
     dma_setup_prepare(rhport);
+    // CACHE HINT
+    // When cache is enabled, _setup_packet must have cache line size alignment
+    // and there should be no valuable data in memory after.
+    // Thus, specific struct is used as a buffer for setup packet data
+    dcd_dcache_invalidate((uint8_t*) _setup_packet, _sizeof_setup_packet());
     dcd_event_setup_received(rhport, (uint8_t*) _setup_packet, true);
     return;
   }
@@ -872,7 +895,9 @@ static void handle_epout_dma(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepi
         if(epnum == 0 && xfer->total_len == 0) {
           dma_setup_prepare(rhport);
         }
-
+        // CACHE HINT
+        // Some data has been received by DMA, fetch the data from memory to cache
+        dcd_dcache_invalidate(xfer->buffer, xfer->total_len);
         dcd_event_xfer_complete(rhport, epnum, xfer->total_len, XFER_RESULT_SUCCESS, true);
       }
     }
diff --git a/src/portable/synopsys/dwc2/dwc2_esp32.h b/src/portable/synopsys/dwc2/dwc2_esp32.h
index 42ab4b80f..c7f1f78e9 100644
--- a/src/portable/synopsys/dwc2/dwc2_esp32.h
+++ b/src/portable/synopsys/dwc2/dwc2_esp32.h
@@ -39,6 +39,14 @@
 #include "soc/periph_defs.h"
 #include "soc/usb_wrap_struct.h"
 
+#if (CFG_TUD_DWC2_DMA_ENABLE && SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE)
+#include "sdkconfig.h"
+#include "esp_cache.h"
+#include "esp_log.h"
+
+#define DWC2_MEM_CACHE_LINE_SIZE    CONFIG_CACHE_L1_CACHE_LINE_SIZE
+#endif // SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE
+
 #if TU_CHECK_MCU(OPT_MCU_ESP32S2, OPT_MCU_ESP32S3)
 #define DWC2_FS_REG_BASE   0x60080000UL
 #define DWC2_EP_MAX        7
@@ -111,6 +119,25 @@ TU_ATTR_ALWAYS_INLINE static inline void dwc2_phy_update(dwc2_regs_t* dwc2, uint
   // maybe usb_utmi_hal_disable()
 }
 
+#if (CFG_TUD_DWC2_DMA_ENABLE && SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE)
+void dcd_dcache_clean(void const* addr, uint32_t data_size) {
+  int flags = ESP_CACHE_MSYNC_FLAG_DIR_C2M | ESP_CACHE_MSYNC_FLAG_UNALIGNED;
+  if (addr != NULL && data_size) {
+    esp_err_t ret = esp_cache_msync((void *) addr, data_size, flags);
+    assert(ret == ESP_OK);
+  }
+}
+
+void dcd_dcache_invalidate(void const* addr, uint32_t data_size) {
+  int flags = ESP_CACHE_MSYNC_FLAG_DIR_M2C;
+  if (addr != NULL && data_size) {
+    data_size = (data_size < DWC2_MEM_CACHE_LINE_SIZE)? DWC2_MEM_CACHE_LINE_SIZE : data_size;
+    esp_err_t ret = esp_cache_msync((void *) addr, data_size, flags);
+    assert(ret == ESP_OK);
+  }
+}
+#endif // CFG_TUD_DWC2_DMA_ENABLE && SOC_CACHE_INTERNAL_MEM_VIA_L1CACHE
+
 #ifdef __cplusplus
 }
 #endif