zynq-7000学习笔记(十一)——Linux下VDMA的使用

PC平台：WINDOWS 10 64位 + 虚拟机Ubuntu 14.04

Xilinx设计开发套件：Xilinx_vivado_sdk_2015.4

开发板：Zed Board

USB摄像头：罗技 C270(720P)

Linux源码：2016_R1

Linaro文件系统：linaro-vivid-developer-20150618-705.tar.gz

      在zynq平台做视频应用，VDMA是一定会用到的，在linux下如何使用呢？我在网上找到了一个例子，稍微做了一下修改，代码没怎么整理，将就着先用

#include <stdio.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
#include <pthread.h>

/* Register offsets */
#define OFFSET_PARK_PTR_REG 0x28
#define OFFSET_VERSION 0x2c

#define OFFSET_VDMA_MM2S_CONTROL_REGISTER 0x00
#define OFFSET_VDMA_MM2S_STATUS_REGISTER 0x04
#define OFFSET_VDMA_MM2S_VSIZE 0x50
#define OFFSET_VDMA_MM2S_HSIZE 0x54
#define OFFSET_VDMA_MM2S_FRMDLY_STRIDE 0x58
#define OFFSET_VDMA_MM2S_FRAMEBUFFER1 0x5c
#define OFFSET_VDMA_MM2S_FRAMEBUFFER2 0x60
#define OFFSET_VDMA_MM2S_FRAMEBUFFER3 0x64
#define OFFSET_VDMA_MM2S_FRAMEBUFFER4 0x68

#define OFFSET_VDMA_S2MM_CONTROL_REGISTER 0x30
#define OFFSET_VDMA_S2MM_STATUS_REGISTER 0x34
#define OFFSET_VDMA_S2MM_IRQ_MASK 0x3c
#define OFFSET_VDMA_S2MM_REG_INDEX 0x44
#define OFFSET_VDMA_S2MM_VSIZE 0xa0
#define OFFSET_VDMA_S2MM_HSIZE 0xa4
#define OFFSET_VDMA_S2MM_FRMDLY_STRIDE 0xa8
#define OFFSET_VDMA_S2MM_FRAMEBUFFER1 0xac
#define OFFSET_VDMA_S2MM_FRAMEBUFFER2 0xb0
#define OFFSET_VDMA_S2MM_FRAMEBUFFER3 0xb4
#define OFFSET_VDMA_S2MM_FRAMEBUFFER4 0xb8

/* S2MM and MM2S control register flags */
#define VDMA_CONTROL_REGISTER_START 0x00000001
#define VDMA_CONTROL_REGISTER_CIRCULAR_PARK 0x00000002
#define VDMA_CONTROL_REGISTER_RESET 0x00000004
#define VDMA_CONTROL_REGISTER_GENLOCK_ENABLE 0x00000008
#define VDMA_CONTROL_REGISTER_FrameCntEn 0x00000010
#define VDMA_CONTROL_REGISTER_INTERNAL_GENLOCK 0x00000080
#define VDMA_CONTROL_REGISTER_WrPntr 0x00000f00
#define VDMA_CONTROL_REGISTER_FrmCtn_IrqEn 0x00001000
#define VDMA_CONTROL_REGISTER_DlyCnt_IrqEn 0x00002000
#define VDMA_CONTROL_REGISTER_ERR_IrqEn 0x00004000
#define VDMA_CONTROL_REGISTER_Repeat_En 0x00008000
#define VDMA_CONTROL_REGISTER_InterruptFrameCount 0x00ff0000
#define VDMA_CONTROL_REGISTER_IRQDelayCount 0xff000000

/* S2MM status register */
#define VDMA_STATUS_REGISTER_HALTED 0x00000001 // Read-only
#define VDMA_STATUS_REGISTER_VDMAInternalError 0x00000010 // Read or write-clear
#define VDMA_STATUS_REGISTER_VDMASlaveError 0x00000020 // Read-only
#define VDMA_STATUS_REGISTER_VDMADecodeError 0x00000040 // Read-only
#define VDMA_STATUS_REGISTER_StartOfFrameEarlyError 0x00000080 // Read-only
#define VDMA_STATUS_REGISTER_EndOfLineEarlyError 0x00000100 // Read-only
#define VDMA_STATUS_REGISTER_StartOfFrameLateError 0x00000800 // Read-only
#define VDMA_STATUS_REGISTER_FrameCountInterrupt 0x00001000 // Read-only
#define VDMA_STATUS_REGISTER_DelayCountInterrupt 0x00002000 // Read-only
#define VDMA_STATUS_REGISTER_ErrorInterrupt 0x00004000 // Read-only
#define VDMA_STATUS_REGISTER_EndOfLineLateError 0x00008000 // Read-only
#define VDMA_STATUS_REGISTER_FrameCount 0x00ff0000 // Read-only
#define VDMA_STATUS_REGISTER_DelayCount 0xff000000 // Read-only

typedef struct {
unsigned int baseAddr;
int vdmaHandler;
int width;
int height;
int pixelLength;
int fbLength;
unsigned int* vdmaVirtualAddress;
unsigned char* fb1VirtualAddress;
unsigned char* fb1PhysicalAddress;
unsigned char* fb2VirtualAddress;
unsigned char* fb2PhysicalAddress;
unsigned char* fb3VirtualAddress;
unsigned char* fb3PhysicalAddress;

pthread_mutex_t lock;
} vdma_handle;

int vdma_setup(vdma_handle *handle, unsigned int baseAddr, int width, int height, int pixelLength, unsigned int fb1Addr, unsigned int fb2Addr, unsigned int fb3Addr) {
handle->baseAddr=baseAddr;
handle->width=width;
handle->height=height;
handle->pixelLength=pixelLength;
handle->fbLength=pixelLength*width*height;
handle->vdmaHandler = open("/dev/mem", O_RDWR | O_SYNC);
handle->vdmaVirtualAddress = (unsigned int*)mmap(NULL, 65535, PROT_READ | PROT_WRITE, MAP_SHARED, handle->vdmaHandler, (off_t)handle->baseAddr);
if(handle->vdmaVirtualAddress == MAP_FAILED) {
perror("vdmaVirtualAddress mapping for absolute memory access failed.\n");
return -1;
}

handle->fb1PhysicalAddress = fb1Addr;
handle->fb1VirtualAddress = (unsigned char*)mmap(NULL, handle->fbLength, PROT_READ | PROT_WRITE, MAP_SHARED, handle->vdmaHandler, (off_t)fb1Addr);
if(handle->fb1VirtualAddress == MAP_FAILED) {
perror("fb1VirtualAddress mapping for absolute memory access failed.\n");
return -2;
}

handle->fb2PhysicalAddress = fb2Addr;
handle->fb2VirtualAddress = (unsigned char*)mmap(NULL, handle->fbLength, PROT_READ | PROT_WRITE, MAP_SHARED, handle->vdmaHandler, (off_t)fb2Addr);
if(handle->fb2VirtualAddress == MAP_FAILED) {
perror("fb2VirtualAddress mapping for absolute memory access failed.\n");
return -3;
}

handle->fb3PhysicalAddress = fb3Addr;
handle->fb3VirtualAddress = (unsigned char*)mmap(NULL, handle->fbLength, PROT_READ | PROT_WRITE, MAP_SHARED, handle->vdmaHandler, (off_t)fb3Addr);
if(handle->fb3VirtualAddress == MAP_FAILED)
{
perror("fb3VirtualAddress mapping for absolute memory access failed.\n");
return -3;
}

memset(handle->fb1VirtualAddress, 255, handle->width*handle->height*handle->pixelLength);
memset(handle->fb2VirtualAddress, 255, handle->width*handle->height*handle->pixelLength);
memset(handle->fb3VirtualAddress, 255, handle->width*handle->height*handle->pixelLength);
return 0;
}

void vdma_halt(vdma_handle *handle) {
vdma_set(handle, OFFSET_VDMA_S2MM_CONTROL_REGISTER, VDMA_CONTROL_REGISTER_RESET);
vdma_set(handle, OFFSET_VDMA_MM2S_CONTROL_REGISTER, VDMA_CONTROL_REGISTER_RESET);
munmap((void *)handle->vdmaVirtualAddress, 65535);
munmap((void *)handle->fb1VirtualAddress, handle->fbLength);
munmap((void *)handle->fb2VirtualAddress, handle->fbLength);
munmap((void *)handle->fb3VirtualAddress, handle->fbLength);
close(handle->vdmaHandler);
}

unsigned int vdma_get(vdma_handle *handle, int num) {
return handle->vdmaVirtualAddress[num>>2];
}

void vdma_set(vdma_handle *handle, int num, unsigned int val) {
handle->vdmaVirtualAddress[num>>2]=val;
}

void vdma_status_dump(int status) {
if (status & VDMA_STATUS_REGISTER_HALTED) printf(" halted"); else printf("running");
if (status & VDMA_STATUS_REGISTER_VDMAInternalError) printf(" vdma-internal-error");
if (status & VDMA_STATUS_REGISTER_VDMASlaveError) printf(" vdma-slave-error");
if (status & VDMA_STATUS_REGISTER_VDMADecodeError) printf(" vdma-decode-error");
if (status & VDMA_STATUS_REGISTER_StartOfFrameEarlyError) printf(" start-of-frame-early-error");
if (status & VDMA_STATUS_REGISTER_EndOfLineEarlyError) printf(" end-of-line-early-error");
if (status & VDMA_STATUS_REGISTER_StartOfFrameLateError) printf(" start-of-frame-late-error");
if (status & VDMA_STATUS_REGISTER_FrameCountInterrupt) printf(" frame-count-interrupt");
if (status & VDMA_STATUS_REGISTER_DelayCountInterrupt) printf(" delay-count-interrupt");
if (status & VDMA_STATUS_REGISTER_ErrorInterrupt) printf(" error-interrupt");
if (status & VDMA_STATUS_REGISTER_EndOfLineLateError) printf(" end-of-line-late-error");
printf(" frame-count:%d", (status & VDMA_STATUS_REGISTER_FrameCount) >> 16);
printf(" delay-count:%d", (status & VDMA_STATUS_REGISTER_DelayCount) >> 24);
printf("\n");
}

void vdma_s2mm_status_dump(vdma_handle *handle) {
int status = vdma_get(handle, OFFSET_VDMA_S2MM_STATUS_REGISTER);
printf("S2MM status register (%08x):", status);
vdma_status_dump(status);
}

void vdma_mm2s_status_dump(vdma_handle *handle) {
int status = vdma_get(handle, OFFSET_VDMA_MM2S_STATUS_REGISTER);
printf("MM2S status register (%08x):", status);
vdma_status_dump(status);
}

void vdma_start_triple_buffering(vdma_handle *handle) {
// Reset VDMA
vdma_set(handle, OFFSET_VDMA_S2MM_CONTROL_REGISTER, VDMA_CONTROL_REGISTER_RESET);
vdma_set(handle, OFFSET_VDMA_MM2S_CONTROL_REGISTER, VDMA_CONTROL_REGISTER_RESET);

// Wait for reset to finish
while((vdma_get(handle, OFFSET_VDMA_S2MM_CONTROL_REGISTER) & VDMA_CONTROL_REGISTER_RESET)==4);
while((vdma_get(handle, OFFSET_VDMA_MM2S_CONTROL_REGISTER) & VDMA_CONTROL_REGISTER_RESET)==4);

// Clear all error bits in status register
vdma_set(handle, OFFSET_VDMA_S2MM_STATUS_REGISTER, 0);
vdma_set(handle, OFFSET_VDMA_MM2S_STATUS_REGISTER, 0);

// Do not mask interrupts
vdma_set(handle, OFFSET_VDMA_S2MM_IRQ_MASK, 0xf);

int interrupt_frame_count = 3;

// Start both S2MM and MM2S in triple buffering mode
vdma_set(handle, OFFSET_VDMA_S2MM_CONTROL_REGISTER,
(interrupt_frame_count << 16) |
VDMA_CONTROL_REGISTER_START |
VDMA_CONTROL_REGISTER_GENLOCK_ENABLE |
VDMA_CONTROL_REGISTER_INTERNAL_GENLOCK |
VDMA_CONTROL_REGISTER_CIRCULAR_PARK);
vdma_set(handle, OFFSET_VDMA_MM2S_CONTROL_REGISTER,
(interrupt_frame_count << 16) |
VDMA_CONTROL_REGISTER_START |
VDMA_CONTROL_REGISTER_GENLOCK_ENABLE |
VDMA_CONTROL_REGISTER_INTERNAL_GENLOCK |
VDMA_CONTROL_REGISTER_CIRCULAR_PARK);

while((vdma_get(handle, 0x30)&1)==0 || (vdma_get(handle, 0x34)&1)==1) {
printf("Waiting for VDMA to start running...\n");
sleep(1);
}

// Extra register index, use first 16 frame pointer registers
vdma_set(handle, OFFSET_VDMA_S2MM_REG_INDEX, 0);

// Write physical addresses to control register
vdma_set(handle, OFFSET_VDMA_S2MM_FRAMEBUFFER1, handle->fb1PhysicalAddress);
vdma_set(handle, OFFSET_VDMA_S2MM_FRAMEBUFFER2, handle->fb1PhysicalAddress);
vdma_set(handle, OFFSET_VDMA_S2MM_FRAMEBUFFER3, handle->fb1PhysicalAddress);

vdma_set(handle, OFFSET_VDMA_MM2S_FRAMEBUFFER1, handle->fb2PhysicalAddress);
vdma_set(handle, OFFSET_VDMA_MM2S_FRAMEBUFFER2, handle->fb2PhysicalAddress);
vdma_set(handle, OFFSET_VDMA_MM2S_FRAMEBUFFER3, handle->fb2PhysicalAddress);

// Write Park pointer register
vdma_set(handle, OFFSET_PARK_PTR_REG, 0);

// Frame delay and stride (bytes)
vdma_set(handle, OFFSET_VDMA_S2MM_FRMDLY_STRIDE, handle->width*handle->pixelLength);
vdma_set(handle, OFFSET_VDMA_MM2S_FRMDLY_STRIDE, handle->width*handle->pixelLength);

// Write horizontal size (bytes)
vdma_set(handle, OFFSET_VDMA_S2MM_HSIZE, handle->width*handle->pixelLength);
vdma_set(handle, OFFSET_VDMA_MM2S_HSIZE, handle->width*handle->pixelLength);

// Write vertical size (lines), this actually starts the transfer
vdma_set(handle, OFFSET_VDMA_S2MM_VSIZE, handle->height);
vdma_set(handle, OFFSET_VDMA_MM2S_VSIZE, handle->height);
}

int vdma_running(vdma_handle *handle) {
// Check whether VDMA is running, that is ready to start transfers
return (vdma_get(handle, 0x34)&1)==1;
}

int vdma_idle(vdma_handle *handle) {
// Check whtether VDMA is transferring
return (vdma_get(handle, OFFSET_VDMA_S2MM_STATUS_REGISTER) & VDMA_STATUS_REGISTER_FrameCountInterrupt)!=0;
}

int main() {
int j, i;
vdma_handle handle;

// Setup VDMA handle and memory-mapped ranges
vdma_setup(&handle, 0x43010000, 640, 480, 4, 0x1f400000, 0x1f500000, 0x1f600000);

// Start triple buffering
vdma_start_triple_buffering(&handle);

// Run for 10 seconds, just monitor status registers
for(i=0; i<10; i++) {
vdma_s2mm_status_dump(&handle);
vdma_mm2s_status_dump(&handle);

memset(handle.fb2VirtualAddress, 0xCC+i, 640*480*4);

printf("FB1:\n");
for (j = 0; j < 256; j++)
printf(" %02x", handle.fb1VirtualAddress[j]); printf("\n");
sleep(1);
}

// Halt VDMA and unmap memory ranges
vdma_halt(&handle);
}

      main函数测试方法很简单，就是往MM2S写数据，然后把S2MM的数据打印出来看看是不是对的；不过先要把vivado工程里的VDMA stream端的MM2S和S2MM直接连起来，形成一个回环才能测试；还要把VDMA加到设备树里面去，下面是个例子，要注意地址是否与vivado分配的地址一致

axi_vdma_2: fast_corner_vdma@43020000 {
compatible = "xlnx,axi-vdma-1.00.a";
#dma-cells = <1>;
reg = <0x43020000 0x10000>;
xlnx,num-fstores = <0x8>;
xlnx,flush-fsync = <0x1>;
interrupt-parent = <&intc>;
dma-channel@43020000 {
compatible = "xlnx,axi-vdma-mm2s-channel";
interrupts = <0 52 0x4>;
xlnx,datawidth = <0x40>;
};
dma-channel@43020030 {
compatible = "xlnx,axi-vdma-s2mm-channel";
interrupts = <0 51 0x4>;
xlnx,datawidth = <0x40>;
};
};