一、概述
framebuffer是啥就不用详细说了吧，需要注意的是android emulator的framebuffer貌似用处不大，因为我之前用android emulator运行x86镜像时，可能是分辨率选得太大了，/dev/graphics/fb0文件都没有产生，系统却可以正常跑，因为系统用的是DRM方式。
另外，GOLDFISH-VIRTUAL-HARDWARE.TXT说了
IMPORTANT NOTE: When GPU emulation is enabled, the framebuffer will typically only be used during boot.
所以，本篇可以跳过不看。
二、驱动
老规矩，先看文档
V. Goldfish framebuffer:
========================
Relevant files:
$QEMU/hw/android/goldfish/fb.c
$KERNEL/drivers/video/goldfish_fb.c
Device properties:
Name: goldfish_fb
Id: 0 to N (only one used in practice).
IrqCount: 0
I/O Registers:
0x00 GET_WIDTH R: Read framebuffer width in pixels.
0x04 GET_HEIGHT R: Read framebuffer height in pixels.
0x08 INT_STATUS
0x0c INT_ENABLE
0x10 SET_BASE
0x14 SET_ROTATION
0x18 SET_BLANK W: Set 'blank' flag.
0x1c GET_PHYS_WIDTH R: Read framebuffer width in millimeters.
0x20 GET_PHYS_HEIGHT R: Read framebuffer height in millimeters.
0x24 GET_FORMAT R: Read framebuffer pixel format.
The framebuffer device is a bit peculiar, because it uses, in addition to the
typical I/O registers and IRQs, a large area of physical memory, allocated by
the kernel, but visible to the emulator, to store a large pixel buffer.
The emulator is responsible for displaying the framebuffer content in its UI
window, which can be rotated, as instructed by the kernel.
IMPORTANT NOTE: When GPU emulation is enabled, the framebuffer will typically
only be used during boot. Note that GPU emulation doesn't rely on a specific
virtual GPU device, however, it uses the "QEMU Pipe" device described below.
For more information, please read:
https://android.googlesource.com/platform/sdk/+/master/emulator/opengl/DESIGN
On boot, the kernel will read various properties of the framebuffer:
IO_READ(GET_WIDTH) and IO_READ(GET_HEIGHT) return the width and height of
the framebuffer in pixels. Note that a 'row' corresponds to consecutive bytes
in memory, but doesn't necessarily to an horizontal line on the final display,
due to possible rotation (see SET_ROTATION below).
IO_READ(GET_PHYS_WIDTH) and IO_READ(GET_PHYS_HEIGHT) return the emulated
physical width and height in millimeters, this is later used by the kernel
and the platform to determine the device's emulated density.
IO_READ(GET_FORMAT) returns a value matching the format of pixels in the
framebuffer. Note that these values are specified by the Android hardware
abstraction layer (HAL) and cannot change:
0x01 HAL_PIXEL_FORMAT_BRGA_8888
0x02 HAL_PIXEL_FORMAT_RGBX_8888
0x03 HAL_PIXEL_FORMAT_RGB_888
0x04 HAL_PIXEL_FORMAT_RGB_565
0x05 HAL_PIXEL_FORMAT_BGRA_8888
0x06 HAL_PIXEL_FORMAT_RGBA_5551
0x08 HAL_PIXEL_FORMAT_RGBA_4444
HOWEVER, the kernel driver only expects a value of HAL_PIXEL_FORMAT_RGB_565
at the moment. Until this is fixed, the virtual device should always return
the value 0x04 here. Rows are not padded, so the size in bytes of a single
framebuffer will always be exactly 'width * heigth * 2'.
Note that GPU emulation doesn't have this limitation and can use and display
32-bit surfaces properly, because it doesn't use the framebuffer.
The device has a 'blank' flag. When set to 1, the UI should only display an
empty/blank framebuffer, ignoring the content of the framebuffer memory.
It is set with IO_WRITE(SET_BLANK, <value>), where value can be 1 or 0. This is
used when simulating suspend/resume.
IMPORTANT: The framebuffer memory is allocated by the kernel, which will send
its physical address to the device by using IO_WRITE(SET_BASE, <address>).
The kernel really allocates a memory buffer large enough to hold *two*
framebuffers, in order to implement panning / double-buffering. This also means
that calls to IO_WRITE(SET_BASE, <address>) will be frequent.
The allocation happens with dma_alloc_writecombine() on ARM, which can only
allocate a maximum of 4 MB, this limits the size of each framebuffer to 2 MB,
which may not be enough to emulate high-density devices :-(
For other architectures, dma_alloc_coherent() is used instead, and has the same
upper limit / limitation.
TODO(digit): Explain how it's possible to raise this limit by modifyinf
CONSISTENT_DMA_SIZE and/or MAX_ORDER in the kernel configuration.
The device uses a single IRQ to notify the kernel of several events. When it
is raised, the kernel IRQ handler must IO_READ(INT_STATUS), which will return
a value containing the following bit flags:
bit 0: Set to 1 to indicate a VSYNC event.
bit 1: Set to 1 to indicate that the content of a previous SET_BASE has
been properly displayed.
Note that reading this register also lowers the device's IRQ level.
The second flag is essentially a way to notify the kernel that an
IO_WRITE(SET_BASE, <address>) operation has been succesfully processed by
the emulator, i.e. that the new content has been displayed to the user.
The kernel can control which flags should raise an IRQ by using
IO_WRITE(INT_ENABLE, <flags>), where <flags> has the same format as the
result of IO_READ(INT_STATUS). If the corresponding bit is 0, the an IRQ
for the corresponding event will never be generated,注意，framebuffer的驱动只支持HAL_PIXEL_FORMAT_RGB_565这一种格式，GPU的不受这个限制。另外，驱动程序中VSYNC的中断其实没有使用。
驱动程序为goldfish中的drivers/video/goldfishfb.c
初始化：
static struct platform_driver goldfish_fb_driver = {
.probe = goldfish_fb_probe,
.remove = goldfish_fb_remove,
.driver = {
.name = "goldfish_fb"
}
};
static int __init goldfish_fb_init(void)
{
return platform_driver_register(&goldfish_fb_driver);
}
static void __exit goldfish_fb_exit(void)
{
platform_driver_unregister(&goldfish_fb_driver);
}
module_init(goldfish_fb_init);
module_exit(goldfish_fb_exit);
goldfish_fb_probe，初始化了一下fb中的spin_lock和waitqueue，然后还是获得IO内存，ioremap，获得中断号，设置中断函数，老套路了。
设置fb的一些属性，比如像素，物理尺寸，图像格式等。
注意申请framebuffer使用的内存时，goldfish_fb_memblock_map总是失败的，因为goldfish platform bus上的设备， IORESOURCE_MEM只有一个，fbmem = platform_get_resource(pdev, IORESOURCE_MEM, 1)肯定会失败的，使用-show-kernel可以看到错误提示"no framebuffer memblock"。
实际使用的是goldfish_fb_dma_alloc，使用dma分配的物理内存，大小为两屏，使用FB_SET_BASE时将其中一屏的地址传递给虚拟设备(双缓存时，这种操作会执行很多次)，虚拟设备去把里面的东西画出来。
static int __devinit goldfish_fb_probe(struct platform_device *pdev)
{
int ret;
struct resource *r;
struct goldfish_fb *fb;
size_t framesize;
uint32_t width, height;
fb = kzalloc(sizeof(*fb), GFP_KERNEL);
if(fb == NULL) {
ret = -ENOMEM;
goto err_fb_alloc_failed;
}
spin_lock_init(&fb->lock);
init_waitqueue_head(&fb->wait);
platform_set_drvdata(pdev, fb);
r = platform_get_resource(pdev, IORESOURCE_MEM, 0);
if(r == NULL) {
ret = -ENODEV;
goto err_no_io_base;
}
#if defined(CONFIG_ARM)
fb->reg_base = (void __iomem *)IO_ADDRESS(r->start - IO_START);
#elif defined(CONFIG_X86) || defined(CONFIG_MIPS)
fb->reg_base = ioremap(r->start, PAGE_SIZE);
#else
#error NOT SUPPORTED
#endif
fb->irq = platform_get_irq(pdev, 0);
if(fb->irq < 0) {
ret = -ENODEV;
goto err_no_irq;
}
width = readl(fb->reg_base + FB_GET_WIDTH);
height = readl(fb->reg_base + FB_GET_HEIGHT);
fb->fb.fbops = &goldfish_fb_ops;
fb->fb.flags = FBINFO_FLAG_DEFAULT;
fb->fb.pseudo_palette = fb->cmap;
//strncpy(fb->fb.fix.id, clcd_name, sizeof(fb->fb.fix.id));
fb->fb.fix.type = FB_TYPE_PACKED_PIXELS;
fb->fb.fix.visual = FB_VISUAL_TRUECOLOR;
fb->fb.fix.line_length = width * 2;
fb->fb.fix.accel = FB_ACCEL_NONE;
fb->fb.fix.ypanstep = 1;
fb->fb.var.xres = width;
fb->fb.var.yres = height;
fb->fb.var.xres_virtual = width;
fb->fb.var.yres_virtual = height * 2;
fb->fb.var.bits_per_pixel = 16;
fb->fb.var.activate = FB_ACTIVATE_NOW;
fb->fb.var.height = readl(fb->reg_base + FB_GET_PHYS_HEIGHT);
fb->fb.var.width = readl(fb->reg_base + FB_GET_PHYS_WIDTH);
fb->fb.var.pixclock = 10000;
fb->fb.var.red.offset = 11;
fb->fb.var.red.length = 5;
fb->fb.var.green.offset = 5;
fb->fb.var.green.length = 6;
fb->fb.var.blue.offset = 0;
fb->fb.var.blue.length = 5;
framesize = width * height * 2 * 2;
ret = goldfish_fb_memblock_map(fb, pdev, framesize);
if (ret < 0)
ret = goldfish_fb_dma_alloc(fb, pdev, width, height, framesize);
if (ret < 0)
goto err_alloc_screen_base_failed;
ret = fb_set_var(&fb->fb, &fb->fb.var);
if(ret)
goto err_fb_set_var_failed;
ret = request_irq(fb->irq, goldfish_fb_interrupt, IRQF_SHARED, pdev->name, fb);
if(ret)
goto err_request_irq_failed;
writel(FB_INT_BASE_UPDATE_DONE, fb->reg_base + FB_INT_ENABLE);
goldfish_fb_pan_display(&fb->fb.var, &fb->fb); // updates base
ret = register_framebuffer(&fb->fb);
if(ret)
goto err_register_framebuffer_failed;
#ifdef CONFIG_ANDROID_POWER
fb->early_suspend.suspend = goldfish_fb_early_suspend;
fb->early_suspend.resume = goldfish_fb_late_resume;
android_register_early_suspend(&fb->early_suspend);
#endif
return 0;
err_register_framebuffer_failed:
free_irq(fb->irq, fb);
err_request_irq_failed:
err_fb_set_var_failed:
goldfish_fb_mem_free(fb, pdev);
err_alloc_screen_base_failed:
err_no_irq:
goldfish_fb_regs_free(fb);
err_no_io_base:
kfree(fb);
err_fb_alloc_failed:
return ret;
}
goldfish_fb_interrupt中断函数，如果中断标志位FB_INT_BASE_UPDATE_DONE有效，说明虚拟设备已得知base改变，准备重画(尚未开始画)，然后就去唤醒waitqueue上等待的线程。虚拟设备中对应的代码为goldfish_fb_update_display函数。
static irqreturn_t
goldfish_fb_interrupt(int irq, void *dev_id)
{
unsigned long irq_flags;
struct goldfish_fb *fb = dev_id;
uint32_t status;
spin_lock_irqsave(&fb->lock, irq_flags);
status = readl(fb->reg_base + FB_INT_STATUS);
if(status & FB_INT_BASE_UPDATE_DONE) {
fb->base_update_count++;
wake_up(&fb->wait);
}
spin_unlock_irqrestore(&fb->lock, irq_flags);
return status ? IRQ_HANDLED : IRQ_NONE;
}
static int goldfish_fb_pan_display(struct fb_var_screeninfo *var, struct fb_info *info)
{
unsigned long irq_flags;
int base_update_count;
struct goldfish_fb *fb = container_of(info, struct goldfish_fb, fb);
spin_lock_irqsave(&fb->lock, irq_flags);
base_update_count = fb->base_update_count;
writel(fb->fb.fix.smem_start + fb->fb.var.xres * 2 * var->yoffset, fb->reg_base + FB_SET_BASE);
spin_unlock_irqrestore(&fb->lock, irq_flags);
wait_event_timeout(fb->wait, fb->base_update_count != base_update_count, HZ / 15);
if(fb->base_update_count == base_update_count)
printk("goldfish_fb_pan_display: timeout wating for base update\n");
return 0;
}
goldfish_fb_set_par设置一些参数的，比如旋转
static int goldfish_fb_set_par(struct fb_info *info)
{
struct goldfish_fb *fb = container_of(info, struct goldfish_fb, fb);
if(fb->rotation != fb->fb.var.rotate) {
info->fix.line_length = info->var.xres * 2;
fb->rotation = fb->fb.var.rotate;
writel(fb->rotation, fb->reg_base + FB_SET_ROTATION);
}
return 0;
}
goldfish_fb_set_par和goldfish_fb_pan_display都是在fbmem.c的do_fb_ioctl中调用的。
goldfish_fb_early_suspend用于在suspend时，画一个纯黑的图
static void goldfish_fb_early_suspend(android_early_suspend_t *h)
{
struct goldfish_fb *fb = container_of(h, struct goldfish_fb, early_suspend);
writel(1, fb->reg_base + FB_SET_BLANK);
}
goldfish_fb_late_resume和goldfish_fb_early_suspend对应的
static void goldfish_fb_late_resume(android_early_suspend_t *h)
{
struct goldfish_fb *fb = container_of(h, struct goldfish_fb, early_suspend);
writel(0, fb->reg_base + FB_SET_BLANK);
}
三、虚拟设备
虚拟设备的代码为hw/android/goldfish/fb.c
goldfish_fb_init初始化，s->ds = graphic_console_init返回的格式是32bit的，和驱动中固定支持的565不同，不过驱动压根就没有用GET_FORMAT去看虚拟设备支持的格式，所以就不管它了。另外，dpi被写死了s->dpi = 165
void goldfish_fb_init(int id)
{
struct goldfish_fb_state *s;
s = (struct goldfish_fb_state *)g_malloc0(sizeof(*s));
s->dev.name = "goldfish_fb";
s->dev.id = id;
s->dev.size = 0x1000;
s->dev.irq_count = 1;
s->ds = graphic_console_init(goldfish_fb_update_display,
goldfish_fb_invalidate_display,
NULL,
NULL,
s);
s->dpi = 165; /* XXX: Find better way to get actual value ! */
/* IMPORTANT: DO NOT COMPUTE s->pixel_format and s->bytes_per_pixel
* here because the display surface is going to change later.
*/
s->bytes_per_pixel = 0;
s->pixel_format = -1;
goldfish_device_add(&s->dev, goldfish_fb_readfn, goldfish_fb_writefn, s);
register_savevm(NULL,
"goldfish_fb",
0,
GOLDFISH_FB_SAVE_VERSION,
goldfish_fb_save,
goldfish_fb_load,
s);
}
读写函数也比较简单，注意读函数中的FB_GET_FORMAT没有被驱动程序使用，写函数FB_SET_BASE后会设置need_update=1, need_int=1之类的状态
static uint32_t goldfish_fb_read(void *opaque, hwaddr offset)
{
uint32_t ret;
struct goldfish_fb_state *s = opaque;
switch(offset) {
case FB_GET_WIDTH:
ret = ds_get_width(s->ds);
//printf("FB_GET_WIDTH => %d\n", ret);
return ret;
case FB_GET_HEIGHT:
ret = ds_get_height(s->ds);
//printf( "FB_GET_HEIGHT = %d\n", ret);
return ret;
case FB_INT_STATUS:
ret = s->int_status & s->int_enable;
if(ret) {
s->int_status &= ~ret;
goldfish_device_set_irq(&s->dev, 0, 0);
}
return ret;
case FB_GET_PHYS_WIDTH:
ret = pixels_to_mm( ds_get_width(s->ds), s->dpi );
//printf( "FB_GET_PHYS_WIDTH => %d\n", ret );
return ret;
case FB_GET_PHYS_HEIGHT:
ret = pixels_to_mm( ds_get_height(s->ds), s->dpi );
//printf( "FB_GET_PHYS_HEIGHT => %d\n", ret );
return ret;
case FB_GET_FORMAT:
return goldfish_fb_get_pixel_format(s);
default:
cpu_abort(cpu_single_env,
"goldfish_fb_read: Bad offset %" HWADDR_PRIx "\n",
offset);
return 0;
}
}
static void goldfish_fb_write(void *opaque, hwaddr offset,
uint32_t val)
{
struct goldfish_fb_state *s = opaque;
switch(offset) {
case FB_INT_ENABLE:
s->int_enable = val;
goldfish_device_set_irq(&s->dev, 0, (s->int_status & s->int_enable));
break;
case FB_SET_BASE: {
int need_resize = !s->base_valid;
s->fb_base = val;
s->int_status &= ~FB_INT_BASE_UPDATE_DONE;
s->need_update = 1;
s->need_int = 1;
s->base_valid = 1;
if(s->set_rotation != s->rotation) {
//printf("FB_SET_BASE: rotation : %d => %d\n", s->rotation, s->set_rotation);
s->rotation = s->set_rotation;
need_resize = 1;
}
goldfish_device_set_irq(&s->dev, 0, (s->int_status & s->int_enable));
if (need_resize) {
//printf("FB_SET_BASE: need resize (rotation=%d)\n", s->rotation );
dpy_resize(s->ds);
}
} break;
case FB_SET_ROTATION:
//printf( "FB_SET_ROTATION %d\n", val);
s->set_rotation = val;
break;
case FB_SET_BLANK:
s->blank = val;
s->need_update = 1;
break;
default:
cpu_abort(cpu_single_env,
"goldfish_fb_write: Bad offset %" HWADDR_PRIx "\n",
offset);
}
}
The pixel buffer is itself a set of physical pages allocated by the kernel driver in the emulated system. These pages are contiguous in the emulated system, but not in the emulator's process space which places them randomly in the heap. Also, a function called goldfish_fb_update_display() is in charge of checking the dirty bits of the framebuffer physical pages, in order to compute the bounding rectangle of pixel updates since the last call, and send them to the UI through qframebuffer_update(). More on this later.
goldfish_fb_update_display会被循环调用，如果base变化了，会重画全屏(常用的双缓存方式)。否则调用compute_fb_update_rect_linear判断需要重画的区域并重画(直接打开/dev/fb0开画就是这样的)。
static void goldfish_fb_update_display(void *opaque)
{
struct goldfish_fb_state *s = (struct goldfish_fb_state *)opaque;
uint32_t base;
uint8_t* dst_line;
uint8_t* src_line;
int full_update = 0;
int width, height, pitch;
base = s->fb_base;
if(base == 0)
return;
if((s->int_enable & FB_INT_VSYNC) && !(s->int_status & FB_INT_VSYNC)) {
s->int_status |= FB_INT_VSYNC;
goldfish_device_set_irq(&s->dev, 0, 1);
}
if(s->need_update) {
full_update = 1;
if(s->need_int) {
s->int_status |= FB_INT_BASE_UPDATE_DONE;
if(s->int_enable & FB_INT_BASE_UPDATE_DONE)
goldfish_device_set_irq(&s->dev, 0, 1);
}
s->need_int = 0;
s->need_update = 0;
}
src_line = qemu_get_ram_ptr( base );
dst_line = s->ds->surface->data;
pitch = s->ds->surface->linesize;
width = s->ds->surface->width;
height = s->ds->surface->height;
FbUpdateState fbs;
FbUpdateRect rect;
fbs.width = width;
fbs.height = height;
fbs.dst_pixels = dst_line;
fbs.dst_pitch = pitch;
fbs.bytes_per_pixel = goldfish_fb_get_bytes_per_pixel(s);
fbs.src_pixels = src_line;
fbs.src_pitch = width*s->ds->surface->pf.bytes_per_pixel;
#if STATS
if (full_update)
stats_full_updates += 1;
if (++stats_counter == 120) {
stats_total += stats_counter;
stats_total_full_updates += stats_full_updates;
printf( "full update stats: peak %.2f %% total %.2f %%\n",
stats_full_updates*100.0/stats_counter,
stats_total_full_updates*100.0/stats_total );
stats_counter = 0;
stats_full_updates = 0;
}
#endif /* STATS */
if (s->blank)
{
memset( dst_line, 0, height*pitch );
rect.xmin = 0;
rect.ymin = 0;
rect.xmax = width-1;
rect.ymax = height-1;
}
else
{
if (full_update) { /* don't use dirty-bits optimization */
base = 0;
}
if (compute_fb_update_rect_linear(&fbs, base, &rect) == 0) {
return;
}
}
rect.xmax += 1;
rect.ymax += 1;
#if 0
printf("goldfish_fb_update_display (y:%d,h:%d,x=%d,w=%d)\n",
rect.ymin, rect.ymax-rect.ymin, rect.xmin, rect.xmax-rect.xmin);
#endif
dpy_update(s->ds, rect.xmin, rect.ymin, rect.xmax-rect.xmin, rect.ymax-rect.ymin);
}
goldfish_fb_get_pixel_format，由于驱动没有使用FB_GET_FORMAT寄存器，所以这个函数没啥用处
四、测试程序
来自http://www.linuxquestions.org/questions/programming-9/hello-world-or-display-pixels-on-framebuffer-fb-h-4175539854/上的一个例子
#include <stdlib.h>
#include <unistd.h>
#include <stdio.h>
#include <fcntl.h>
#include <linux/fb.h>
#include <sys/mman.h>
#include <sys/ioctl.h>
int main()
{
int fbfd = 0;
struct fb_var_screeninfo vinfo;
struct fb_fix_screeninfo finfo;
long int screensize = 0;
char *fbp = 0;
int x = 0, y = 0;
long int location = 0;
// Open the file for reading and writing
fbfd = open("/dev/fb0", O_RDWR);
if (fbfd == -1) {
perror("Error: cannot open framebuffer device");
exit(1);
}
printf("The framebuffer device was opened successfully.\n");
// Get fixed screen information
if (ioctl(fbfd, FBIOGET_FSCREENINFO, &finfo) == -1) {
perror("Error reading fixed information");
exit(2);
}
// Get variable screen information
if (ioctl(fbfd, FBIOGET_VSCREENINFO, &vinfo) == -1) {
perror("Error reading variable information");
exit(3);
}
printf("%dx%d, %dbpp\n", vinfo.xres, vinfo.yres, vinfo.bits_per_pixel);
// Figure out the size of the screen in bytes
screensize = vinfo.xres * vinfo.yres * vinfo.bits_per_pixel / 8;
// Map the device to memory
fbp = (char *)mmap(0, screensize, PROT_READ | PROT_WRITE, MAP_SHARED, fbfd, 0);
if ((int)fbp == -1) {
perror("Error: failed to map framebuffer device to memory");
exit(4);
}
printf("The framebuffer device was mapped to memory successfully.\n");
// Figure out where in memory to put the pixel
for (y = 100; y < 300; y++)
for (x = 100; x < 300; x++) {
location = (x+vinfo.xoffset) * (vinfo.bits_per_pixel/8) +
(y+vinfo.yoffset) * finfo.line_length;
if (vinfo.bits_per_pixel == 32) {
*(fbp + location) = 100; // Some blue
*(fbp + location + 1) = 15+(x-100)/2; // A little green
*(fbp + location + 2) = 200-(y-100)/5; // A lot of red
*(fbp + location + 3) = 0; // No transparency
} else { //assume 16bpp
int b = 10;
int g = (x-100)/6; // A little green
int r = 31-(y-100)/16; // A lot of red
unsigned short int t = r<<11 | g << 5 | b;
*((unsigned short int*)(fbp + location)) = t;
}
}
munmap(fbp, screensize);
close(fbfd);
return 0;
}
在android emulator上运行，没有反应。。。
找一个linux机器，带有framebuffer的，测试OK
难怪前面说android的framebuffer没有什么用处。。。
还有一个狂野点的测试程序dd
sudo dd if=/dev/fb0 of=fbdata.bin bs=1M count=2
随便移动窗口或者其他方式改变一下桌面
sudo dd if=fbdata.bin of=/dev/fb0 bs=1M count=2
看到桌面恢复了，然后鼠标滑过的地方会被更新
PS：生命诚可贵，不要用dd。以前搞路由器时，就这样把电脑的sda给dd没了的。
