pbio/drv: Handle cache ops for EV3 drivers

This allows the main memory to be write-back cached.

Author: ArcaneNibble

lib/pbio/drv/block_device/block_device_ev3.c

@@ -34,6 +34,7 @@
 #include "block_device_ev3.h"
 
 #include <pbdrv/block_device.h>
+#include <pbdrv/cache.h>
 #include <pbdrv/clock.h>
 #include <pbdrv/compiler.h>
 #include <pbdrv/gpio.h>
@@ -103,6 +104,10 @@ static struct {
     uint8_t spi_cmd_buf_tx[SPI_CMD_BUF_SZ];
     // This is used to hold the replies to commands to the SPI peripheral.
     uint8_t spi_cmd_buf_rx[SPI_CMD_BUF_SZ];
+    // This stores the RX buffer address so that we clean the cache when DMA is complete
+    uint32_t rx_user_buf_addr;
+    // This stores the RX buffer size;
+    uint32_t rx_user_buf_sz;
 } spi_dev;
 
 static uint32_t last_spi_dma_complete_time;
@@ -114,6 +119,10 @@ static void spi_dma_complete(void) {
     }
     SPIIntDisable(SOC_SPI_0_REGS, SPI_DMA_REQUEST_ENA_INT);
     pbio_os_request_poll();
+    pbdrv_cache_prepare_after_dma(spi_dev.spi_cmd_buf_rx, SPI_CMD_BUF_SZ);
+    if (spi_dev.rx_user_buf_addr && spi_dev.rx_user_buf_sz) {
+        pbdrv_cache_prepare_after_dma((void *)spi_dev.rx_user_buf_addr, spi_dev.rx_user_buf_sz);
+    }
     last_spi_dma_complete_time = pbdrv_clock_get_ms();
 }
 
@@ -351,16 +360,21 @@ static pbio_error_t spi_begin_for_flash(
 
         spi_dev.status = SPI_STATUS_WAIT_RX;
 
+        // Prevent write to spi_dev.status from being reordered
+        pbdrv_compiler_memory_barrier();
+
         uint32_t tx = spi0_last_dat1_for_flash(cmd[0]);
         SPIIntEnable(SOC_SPI_0_REGS, SPI_RECV_INT);
         HWREG(SOC_SPI_0_REGS + SPI_SPIDAT1) = tx;
     } else {
-        memcpy(&spi_dev.spi_cmd_buf_tx, cmd, cmd_len);
+        memcpy(PBDRV_UNCACHED_ADDR(spi_dev.spi_cmd_buf_tx), cmd, cmd_len);
+        spi_dev.rx_user_buf_addr = (uint32_t)user_data_rx;
+        spi_dev.rx_user_buf_sz = user_data_len;
 
         if (user_data_len == 0) {
             // Only a command, no user data
 
-            spi_dev.tx_last_word = spi0_last_dat1_for_flash(cmd[cmd_len - 1]);
+            PBDRV_UNCACHED(spi_dev.tx_last_word) = spi0_last_dat1_for_flash(cmd[cmd_len - 1]);
 
             // TX everything except last byte
             ps.p.srcAddr = (unsigned int)(&spi_dev.spi_cmd_buf_tx);
@@ -418,7 +432,8 @@ static pbio_error_t spi_begin_for_flash(
             edma3_set_param(EDMA3_CHA_SPI0_TX, &ps);
 
             if (user_data_tx) {
-                spi_dev.tx_last_word = spi0_last_dat1_for_flash(user_data_tx[user_data_len - 1]);
+                pbdrv_cache_prepare_before_dma(user_data_tx, user_data_len);
+                PBDRV_UNCACHED(spi_dev.tx_last_word) = spi0_last_dat1_for_flash(user_data_tx[user_data_len - 1]);
 
                 // TX all but the last byte
                 ps.p.srcAddr = (unsigned int)(user_data_tx);
@@ -434,7 +449,7 @@ static pbio_error_t spi_begin_for_flash(
                 ps.p.opt = EDMA3CC_OPT_TCINTEN | (EDMA3_CHA_SPI0_TX << EDMA3CC_OPT_TCC_SHIFT);
                 edma3_set_param(127, &ps);
             } else {
-                spi_dev.tx_last_word = spi0_last_dat1_for_flash(0);
+                PBDRV_UNCACHED(spi_dev.tx_last_word) = spi0_last_dat1_for_flash(0);
 
                 // TX all but the last byte
                 ps.p.srcAddr = (unsigned int)(&spi_dev.tx_dummy_byte);
@@ -487,7 +502,7 @@ static pbio_error_t spi_begin_for_flash(
 
         spi_dev.status = SPI_STATUS_WAIT_TX | SPI_STATUS_WAIT_RX;
 
-        // TODO: eventually needs DMA cache management
+        // Prevent write to spi_dev.status from being reordered
         pbdrv_compiler_memory_barrier();
 
         EDMA3EnableTransfer(SOC_EDMA30CC_0_REGS, EDMA3_CHA_SPI0_TX, EDMA3_TRIG_MODE_EVENT);
@@ -804,9 +819,14 @@ static pbio_error_t pbdrv_block_device_ev3_spi_begin_for_adc(const uint32_t *cmd
     ps.p.opt = EDMA3CC_OPT_TCINTEN | (EDMA3_CHA_SPI0_RX << EDMA3CC_OPT_TCC_SHIFT);
     edma3_set_param(EDMA3_CHA_SPI0_RX, &ps);
 
+    // We play dangerously and don't flush the cache for commands (since they're const)
+    // but we do need to flush the cache for the data which is read.
+    spi_dev.rx_user_buf_addr = (uint32_t)data;
+    spi_dev.rx_user_buf_sz = sizeof(uint16_t) * len;
+
     spi_dev.status = SPI_STATUS_WAIT_TX | SPI_STATUS_WAIT_RX;
 
-    // TODO: eventually needs DMA cache management
+    // Prevent write to spi_dev.status from being reordered
     pbdrv_compiler_memory_barrier();
 
     EDMA3EnableTransfer(SOC_EDMA30CC_0_REGS, EDMA3_CHA_SPI0_TX, EDMA3_TRIG_MODE_EVENT);

lib/pbio/drv/display/display_ev3.c

@@ -13,6 +13,7 @@
 #include <stdio.h>
 #include <string.h>
 
+#include <pbdrv/cache.h>
 #include <pbdrv/display.h>
 #include <pbdrv/gpio.h>
 
@@ -363,6 +364,8 @@ void pbdrv_display_st7586s_write_data_begin(uint8_t *data, uint32_t size) {
     spi_status = SPI_STATUS_WAIT;
     pbdrv_gpio_out_low(&pin_lcd_cs);
 
+    pbdrv_cache_prepare_before_dma(data, size);
+
     // Parameter object must be volatile since it is copied byte-by-byte in the
     // TI API, causing it to be optimized out.
     volatile EDMA3CCPaRAMEntry paramSet = {

lib/pbio/drv/uart/uart_ev3.c

@@ -14,6 +14,7 @@
 #include <contiki.h>
 #include <contiki-lib.h>
 
+#include <pbdrv/cache.h>
 #include <pbdrv/uart.h>
 
 #include <pbio/busy_count.h>
@@ -220,6 +221,8 @@ pbio_error_t pbdrv_uart_write_hw(pbio_os_state_t *state, pbdrv_uart_dev_t *uart,
         .opt = EDMA3CC_OPT_DAM | ((pdata->sys_int_uart_tx_int_id << EDMA3CC_OPT_TCC_SHIFT) & EDMA3CC_OPT_TCC) | (1 << EDMA3CC_OPT_TCINTEN_SHIFT),
     };
 
+    pbdrv_cache_prepare_before_dma(uart->write_buf, length);
+
     // Save configuration and start transfer.
     EDMA3SetPaRAM(SOC_EDMA30CC_0_REGS, pdata->sys_int_uart_tx_int_id, (EDMA3CCPaRAMEntry *)&paramSet);
     EDMA3EnableTransfer(SOC_EDMA30CC_0_REGS, pdata->sys_int_uart_tx_int_id, EDMA3_TRIG_MODE_EVENT);

lib/pbio/drv/usb/usb_ev3.c

@@ -13,6 +13,7 @@
 #include <string.h>
 
 #include <pbdrv/bluetooth.h>
+#include <pbdrv/cache.h>
 #include <pbdrv/compiler.h>
 #include <pbdrv/usb.h>
 #include <pbio/os.h>
@@ -283,7 +284,7 @@ static uint32_t cppi_linking_ram[CPPI_DESC_COUNT];
 
 // Fill in the CPPI DMA descriptor to receive a packet
 static void usb_setup_rx_dma_desc(void) {
-    cppi_descriptors[CPPI_DESC_RX] = (usb_cppi_hpd_t) {
+    PBDRV_UNCACHED(cppi_descriptors[CPPI_DESC_RX]) = (usb_cppi_hpd_t) {
         .word0 = {
             .hostPktType = CPPI_HOST_PACKET_DESCRIPTOR_TYPE,
         },
@@ -307,7 +308,7 @@ static void usb_setup_rx_dma_desc(void) {
 
 // Fill in the CPPI DMA descriptor to send a packet
 static void usb_setup_tx_dma_desc(int tx_type, void *buf, uint32_t buf_len) {
-    cppi_descriptors[tx_type] = (usb_cppi_hpd_t) {
+    PBDRV_UNCACHED(cppi_descriptors[tx_type]) = (usb_cppi_hpd_t) {
         .word0 = {
             .hostPktType = CPPI_HOST_PACKET_DESCRIPTOR_TYPE,
             .pktLength = buf_len,
@@ -904,11 +905,13 @@ static pbio_error_t pbdrv_usb_ev3_process_thread(pbio_os_state_t *state, void *c
             // (which is technically allowed by the as-if rule).
             pbdrv_compiler_memory_barrier();
 
-            uint32_t usb_rx_sz = cppi_descriptors[CPPI_DESC_RX].word0.pktLength;
+            uint32_t usb_rx_sz = PBDRV_UNCACHED(cppi_descriptors[CPPI_DESC_RX].word0).pktLength;
             pbio_pybricks_error_t result;
             bool usb_send_response = false;
             // Skip empty commands, or commands sent when previous response is still busy
             if (usb_rx_sz && !usb_tx_response_is_not_ready) {
+                pbdrv_cache_prepare_after_dma(ep1_rx_buf, sizeof(ep1_rx_buf));
+
                 switch (ep1_rx_buf[0]) {
                     case PBIO_PYBRICKS_OUT_EP_MSG_SUBSCRIBE:
                         pbio_os_timer_set(&keepalive_timer, 1000);
@@ -928,6 +931,7 @@ static pbio_error_t pbdrv_usb_ev3_process_thread(pbio_os_state_t *state, void *c
 
             if (usb_send_response) {
                 usb_tx_response_is_not_ready = true;
+                pbdrv_cache_prepare_before_dma(ep1_tx_response_buf, sizeof(ep1_tx_response_buf));
                 usb_setup_tx_dma_desc(CPPI_DESC_TX_RESPONSE, ep1_tx_response_buf, PBIO_PYBRICKS_USB_MESSAGE_SIZE(sizeof(uint32_t)));
             }
 
@@ -944,6 +948,7 @@ static pbio_error_t pbdrv_usb_ev3_process_thread(pbio_os_state_t *state, void *c
             uint32_t usb_status_sz = PBIO_PYBRICKS_USB_MESSAGE_SIZE(pbsys_status_get_status_report(&ep1_tx_status_buf[1]));
 
             usb_tx_status_is_not_ready = true;
+            pbdrv_cache_prepare_before_dma(ep1_tx_status_buf, sizeof(ep1_tx_status_buf));
             usb_setup_tx_dma_desc(CPPI_DESC_TX_STATUS, ep1_tx_status_buf, usb_status_sz);
 
             prev_status_flags = new_status_flags;
@@ -1124,6 +1129,7 @@ pbio_error_t pbdrv_usb_stdout_tx(const uint8_t *data, uint32_t *size) {
     memcpy(ptr, data, *size);
 
     usb_tx_stdout_is_not_ready = true;
+    pbdrv_cache_prepare_before_dma(ep1_tx_stdout_buf, sizeof(ep1_tx_stdout_buf));
     usb_setup_tx_dma_desc(CPPI_DESC_TX_STDOUT, ep1_tx_stdout_buf, 2 + *size);
 
     return PBIO_SUCCESS;

lib/pbio/platform/ev3/platform.c

@@ -721,8 +721,8 @@ static void mmu_init(void) {
     // Off-chip main DDR RAM
     for (unsigned int i = 0; i < SYSTEM_RAM_SZ_MB; i++) {
         uint32_t addr = 0xC0000000 + i * MMU_SECTION_SZ;
-        // TODO: Enable caching once DMA code is upgraded to handle cache
-        l1_page_table[addr >> MMU_SECTION_SHIFT] = MMU_L1_SECTION(addr, 0, 1, 0, 0);
+        // Enable write-back caching
+        l1_page_table[addr >> MMU_SECTION_SHIFT] = MMU_L1_SECTION(addr, 0, 1, 1, 1);
     }
     // Off-chip main DDR RAM, uncacheable mirror @ 0xD0000000
     for (unsigned int i = 0; i < SYSTEM_RAM_SZ_MB; i++) {