target/linux/brcm2708/patches-4.14/950-0056-fbdev-add-FBIOCOPYAREA-ioctl.patch

   1 From d56622a9c76ab913da66bb22c3158b10f9fb0543 Mon Sep 17 00:00:00 2001
   2 From: Siarhei Siamashka <siarhei.siamashka@gmail.com>
   3 Date: Mon, 17 Jun 2013 13:32:11 +0300
   4 Subject: [PATCH 056/454] fbdev: add FBIOCOPYAREA ioctl
   5
   6 Based on the patch authored by Ali Gholami Rudi at
   7     https://lkml.org/lkml/2009/7/13/153
   8
   9 Provide an ioctl for userspace applications, but only if this operation
  10 is hardware accelerated (otherwide it does not make any sense).
  11
  12 Signed-off-by: Siarhei Siamashka <siarhei.siamashka@gmail.com>
  13
  14 bcm2708_fb: Add ioctl for reading gpu memory through dma
  15 ---
  16  drivers/video/fbdev/bcm2708_fb.c | 111 +++++++++++++++++++++++++++++++
  17  drivers/video/fbdev/core/fbmem.c |  36 ++++++++++
  18  include/uapi/linux/fb.h          |  12 ++++
  19  3 files changed, 159 insertions(+)
  20
  21 --- a/drivers/video/fbdev/bcm2708_fb.c
  22 +++ b/drivers/video/fbdev/bcm2708_fb.c
  23 @@ -31,8 +31,10 @@
  24  #include <linux/console.h>
  25  #include <linux/debugfs.h>
  26  #include <asm/sizes.h>
  27 +#include <linux/uaccess.h>
  28  #include <linux/io.h>
  29  #include <linux/dma-mapping.h>
  30 +#include <linux/cred.h>
  31  #include <soc/bcm2835/raspberrypi-firmware.h>
  32
  33  //#define BCM2708_FB_DEBUG
  34 @@ -94,6 +96,7 @@ struct bcm2708_fb {
  35         wait_queue_head_t dma_waitq;
  36         struct bcm2708_fb_stats stats;
  37         unsigned long fb_bus_address;
  38 +       struct { u32 base, length; } gpu;
  39  };
  40
  41  #define to_bcm2708(info)       container_of(info, struct bcm2708_fb, fb)
  42 @@ -426,6 +429,106 @@ static int bcm2708_fb_pan_display(struct
  43         return result;
  44  }
  45
  46 +static void dma_memcpy(struct bcm2708_fb *fb, dma_addr_t dst, dma_addr_t src, int size)
  47 +{
  48 +       int burst_size = (fb->dma_chan == 0) ? 8 : 2;
  49 +       struct bcm2708_dma_cb *cb = fb->cb_base;
  50 +
  51 +       cb->info = BCM2708_DMA_BURST(burst_size) | BCM2708_DMA_S_WIDTH |
  52 +                  BCM2708_DMA_S_INC | BCM2708_DMA_D_WIDTH |
  53 +                  BCM2708_DMA_D_INC;
  54 +       cb->dst = dst;
  55 +       cb->src = src;
  56 +       cb->length = size;
  57 +       cb->stride = 0;
  58 +       cb->pad[0] = 0;
  59 +       cb->pad[1] = 0;
  60 +       cb->next = 0;
  61 +
  62 +       if (size < dma_busy_wait_threshold) {
  63 +               bcm_dma_start(fb->dma_chan_base, fb->cb_handle);
  64 +               bcm_dma_wait_idle(fb->dma_chan_base);
  65 +       } else {
  66 +               void __iomem *dma_chan = fb->dma_chan_base;
  67 +               cb->info |= BCM2708_DMA_INT_EN;
  68 +               bcm_dma_start(fb->dma_chan_base, fb->cb_handle);
  69 +               while (bcm_dma_is_busy(dma_chan)) {
  70 +                       wait_event_interruptible(
  71 +                               fb->dma_waitq,
  72 +                               !bcm_dma_is_busy(dma_chan));
  73 +               }
  74 +               fb->stats.dma_irqs++;
  75 +       }
  76 +       fb->stats.dma_copies++;
  77 +}
  78 +
  79 +#define INTALIAS_NORMAL(x) ((x)&~0xc0000000) // address with no aliases
  80 +#define INTALIAS_L1L2_NONALLOCATING(x) (((x)&~0xc0000000)|0x80000000) // cache coherent but non-allocating in L1 and L2
  81 +
  82 +static long vc_mem_copy(struct bcm2708_fb *fb, unsigned long arg)
  83 +{
  84 +       struct fb_dmacopy ioparam;
  85 +       size_t size = PAGE_SIZE;
  86 +       u32 *buf = NULL;
  87 +       dma_addr_t bus_addr;
  88 +       long rc = 0;
  89 +       size_t offset;
  90 +
  91 +       /* restrict this to root user */
  92 +       if (!uid_eq(current_euid(), GLOBAL_ROOT_UID))
  93 +       {
  94 +               rc = -EFAULT;
  95 +               goto out;
  96 +       }
  97 +
  98 +       /* Get the parameter data.
  99 +        */
 100 +       if (copy_from_user
 101 +           (&ioparam, (void *)arg, sizeof(ioparam)) != 0) {
 102 +               pr_err("[%s]: failed to copy-from-user\n",
 103 +                               __func__);
 104 +               rc = -EFAULT;
 105 +               goto out;
 106 +       }
 107 +
 108 +       if (fb->gpu.base == 0 || fb->gpu.length == 0) {
 109 +               pr_err("[%s]: Unable to determine gpu memory (%x,%x)\n", __func__, fb->gpu.base, fb->gpu.length);
 110 +               return -EFAULT;
 111 +       }
 112 +
 113 +       if (INTALIAS_NORMAL(ioparam.src) < fb->gpu.base || INTALIAS_NORMAL(ioparam.src) >= fb->gpu.base + fb->gpu.length) {
 114 +               pr_err("[%s]: Invalid memory access %x (%x-%x)", __func__, INTALIAS_NORMAL(ioparam.src), fb->gpu.base, fb->gpu.base + fb->gpu.length);
 115 +               return -EFAULT;
 116 +       }
 117 +
 118 +       buf = dma_alloc_coherent(fb->fb.device, PAGE_ALIGN(size), &bus_addr,
 119 +                                GFP_ATOMIC);
 120 +       if (!buf) {
 121 +               pr_err("[%s]: failed to dma_alloc_coherent(%d)\n",
 122 +                               __func__, size);
 123 +               rc = -ENOMEM;
 124 +               goto out;
 125 +       }
 126 +
 127 +       for (offset = 0; offset < ioparam.length; offset += size) {
 128 +               size_t remaining = ioparam.length - offset;
 129 +               size_t s = min(size, remaining);
 130 +               unsigned char *p = (unsigned char *)ioparam.src + offset;
 131 +               unsigned char *q = (unsigned char *)ioparam.dst + offset;
 132 +               dma_memcpy(fb, bus_addr, INTALIAS_L1L2_NONALLOCATING((dma_addr_t)p), size);
 133 +               if (copy_to_user(q, buf, s) != 0) {
 134 +                       pr_err("[%s]: failed to copy-to-user\n",
 135 +                                       __func__);
 136 +                       rc = -EFAULT;
 137 +                       goto out;
 138 +               }
 139 +       }
 140 +out:
 141 +       if (buf)
 142 +               dma_free_coherent(fb->fb.device, PAGE_ALIGN(size), buf, bus_addr);
 143 +       return rc;
 144 +}
 145 +
 146  static int bcm2708_ioctl(struct fb_info *info, unsigned int cmd, unsigned long arg)
 147  {
 148         struct bcm2708_fb *fb = to_bcm2708(info);
 149 @@ -438,6 +541,9 @@ static int bcm2708_ioctl(struct fb_info
 150                                             RPI_FIRMWARE_FRAMEBUFFER_SET_VSYNC,
 151                                             &dummy, sizeof(dummy));
 152                 break;
 153 +       case FBIODMACOPY:
 154 +               ret = vc_mem_copy(fb, arg);
 155 +               break;
 156         default:
 157                 dev_dbg(info->device, "Unknown ioctl 0x%x\n", cmd);
 158                 return -ENOTTY;
 159 @@ -760,6 +866,11 @@ static int bcm2708_fb_probe(struct platf
 160         fb->dev = dev;
 161         fb->fb.device = &dev->dev;
 162
 163 +       // failure here isn't fatal, but we'll fail in vc_mem_copy if fb->gpu is not valid
 164 +       rpi_firmware_property(fb->fw,
 165 +                                   RPI_FIRMWARE_GET_VC_MEMORY,
 166 +                                   &fb->gpu, sizeof(fb->gpu));
 167 +
 168         ret = bcm2708_fb_register(fb);
 169         if (ret == 0) {
 170                 platform_set_drvdata(dev, fb);
 171 --- a/drivers/video/fbdev/core/fbmem.c
 172 +++ b/drivers/video/fbdev/core/fbmem.c
 173 @@ -1086,6 +1086,31 @@ fb_blank(struct fb_info *info, int blank
 174  }
 175  EXPORT_SYMBOL(fb_blank);
 176
 177 +static int fb_copyarea_user(struct fb_info *info,
 178 +                           struct fb_copyarea *copy)
 179 +{
 180 +       int ret = 0;
 181 +       if (!lock_fb_info(info))
 182 +               return -ENODEV;
 183 +       if (copy->dx >= info->var.xres ||
 184 +           copy->sx >= info->var.xres ||
 185 +           copy->width > info->var.xres ||
 186 +           copy->dy >= info->var.yres ||
 187 +           copy->sy >= info->var.yres ||
 188 +           copy->height > info->var.yres ||
 189 +           copy->dx + copy->width > info->var.xres ||
 190 +           copy->sx + copy->width > info->var.xres ||
 191 +           copy->dy + copy->height > info->var.yres ||
 192 +           copy->sy + copy->height > info->var.yres) {
 193 +               ret = -EINVAL;
 194 +               goto out;
 195 +       }
 196 +       info->fbops->fb_copyarea(info, copy);
 197 +out:
 198 +       unlock_fb_info(info);
 199 +       return ret;
 200 +}
 201 +
 202  static long do_fb_ioctl(struct fb_info *info, unsigned int cmd,
 203                         unsigned long arg)
 204  {
 205 @@ -1096,6 +1121,7 @@ static long do_fb_ioctl(struct fb_info *
 206         struct fb_cmap cmap_from;
 207         struct fb_cmap_user cmap;
 208         struct fb_event event;
 209 +       struct fb_copyarea copy;
 210         void __user *argp = (void __user *)arg;
 211         long ret = 0;
 212
 213 @@ -1213,6 +1239,15 @@ static long do_fb_ioctl(struct fb_info *
 214                 unlock_fb_info(info);
 215                 console_unlock();
 216                 break;
 217 +       case FBIOCOPYAREA:
 218 +               if (info->flags & FBINFO_HWACCEL_COPYAREA) {
 219 +                       /* only provide this ioctl if it is accelerated */
 220 +                       if (copy_from_user(&copy, argp, sizeof(copy)))
 221 +                               return -EFAULT;
 222 +                       ret = fb_copyarea_user(info, &copy);
 223 +                       break;
 224 +               }
 225 +               /* fall through */
 226         default:
 227                 if (!lock_fb_info(info))
 228                         return -ENODEV;
 229 @@ -1358,6 +1393,7 @@ static long fb_compat_ioctl(struct file
 230         case FBIOPAN_DISPLAY:
 231         case FBIOGET_CON2FBMAP:
 232         case FBIOPUT_CON2FBMAP:
 233 +       case FBIOCOPYAREA:
 234                 arg = (unsigned long) compat_ptr(arg);
 235         case FBIOBLANK:
 236                 ret = do_fb_ioctl(info, cmd, arg);
 237 --- a/include/uapi/linux/fb.h
 238 +++ b/include/uapi/linux/fb.h
 239 @@ -35,6 +35,12 @@
 240  #define FBIOPUT_MODEINFO        0x4617
 241  #define FBIOGET_DISPINFO        0x4618
 242  #define FBIO_WAITFORVSYNC      _IOW('F', 0x20, __u32)
 243 +/*
 244 + * HACK: use 'z' in order not to clash with any other ioctl numbers which might
 245 + * be concurrently added to the mainline kernel
 246 + */
 247 +#define FBIOCOPYAREA           _IOW('z', 0x21, struct fb_copyarea)
 248 +#define FBIODMACOPY            _IOW('z', 0x22, struct fb_dmacopy)
 249
 250  #define FB_TYPE_PACKED_PIXELS          0       /* Packed Pixels        */
 251  #define FB_TYPE_PLANES                 1       /* Non interleaved planes */
 252 @@ -347,6 +353,12 @@ struct fb_copyarea {
 253         __u32 sy;
 254  };
 255
 256 +struct fb_dmacopy {
 257 +       void *dst;
 258 +       __u32 src;
 259 +       __u32 length;
 260 +};
 261 +
 262  struct fb_fillrect {
 263         __u32 dx;       /* screen-relative */
 264         __u32 dy;