drivers/gpu/drm/i915/i915_gem_tiling.c

   1 /*
   2  * Copyright © 2008 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eric Anholt <eric@anholt.net>
  25  *
  26  */
  27
  28 #include "drmP.h"
  29 #include "drm.h"
  30 #include "i915_drm.h"
  31 #include "i915_drv.h"
  32
  33 /** @file i915_gem_tiling.c
  34  *
  35  * Support for managing tiling state of buffer objects.
  36  *
  37  * The idea behind tiling is to increase cache hit rates by rearranging
  38  * pixel data so that a group of pixel accesses are in the same cacheline.
  39  * Performance improvement from doing this on the back/depth buffer are on
  40  * the order of 30%.
  41  *
  42  * Intel architectures make this somewhat more complicated, though, by
  43  * adjustments made to addressing of data when the memory is in interleaved
  44  * mode (matched pairs of DIMMS) to improve memory bandwidth.
  45  * For interleaved memory, the CPU sends every sequential 64 bytes
  46  * to an alternate memory channel so it can get the bandwidth from both.
  47  *
  48  * The GPU also rearranges its accesses for increased bandwidth to interleaved
  49  * memory, and it matches what the CPU does for non-tiled.  However, when tiled
  50  * it does it a little differently, since one walks addresses not just in the
  51  * X direction but also Y.  So, along with alternating channels when bit
  52  * 6 of the address flips, it also alternates when other bits flip --  Bits 9
  53  * (every 512 bytes, an X tile scanline) and 10 (every two X tile scanlines)
  54  * are common to both the 915 and 965-class hardware.
  55  *
  56  * The CPU also sometimes XORs in higher bits as well, to improve
  57  * bandwidth doing strided access like we do so frequently in graphics.  This
  58  * is called "Channel XOR Randomization" in the MCH documentation.  The result
  59  * is that the CPU is XORing in either bit 11 or bit 17 to bit 6 of its address
  60  * decode.
  61  *
  62  * All of this bit 6 XORing has an effect on our memory management,
  63  * as we need to make sure that the 3d driver can correctly address object
  64  * contents.
  65  *
  66  * If we don't have interleaved memory, all tiling is safe and no swizzling is
  67  * required.
  68  *
  69  * When bit 17 is XORed in, we simply refuse to tile at all.  Bit
  70  * 17 is not just a page offset, so as we page an objet out and back in,
  71  * individual pages in it will have different bit 17 addresses, resulting in
  72  * each 64 bytes being swapped with its neighbor!
  73  *
  74  * Otherwise, if interleaved, we have to tell the 3d driver what the address
  75  * swizzling it needs to do is, since it's writing with the CPU to the pages
  76  * (bit 6 and potentially bit 11 XORed in), and the GPU is reading from the
  77  * pages (bit 6, 9, and 10 XORed in), resulting in a cumulative bit swizzling
  78  * required by the CPU of XORing in bit 6, 9, 10, and potentially 11, in order
  79  * to match what the GPU expects.
  80  */
  81
  82 /**
  83  * Detects bit 6 swizzling of address lookup between IGD access and CPU
  84  * access through main memory.
  85  */
  86 void
  87 i915_gem_detect_bit_6_swizzle(struct drm_device *dev)
  88 {
  89         drm_i915_private_t *dev_priv = dev->dev_private;
  90         uint32_t swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
  91         uint32_t swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
  92
  93         if (!IS_I9XX(dev)) {
  94                 /* As far as we know, the 865 doesn't have these bit 6
  95                  * swizzling issues.
  96                  */
  97                 swizzle_x = I915_BIT_6_SWIZZLE_NONE;
  98                 swizzle_y = I915_BIT_6_SWIZZLE_NONE;
  99         } else if ((!IS_I965G(dev) && !IS_G33(dev)) || IS_I965GM(dev) ||
 100                    IS_GM45(dev)) {
 101                 uint32_t dcc;
 102
 103                 /* On 915-945 and GM965, channel interleave by the CPU is
 104                  * determined by DCC.  The CPU will alternate based on bit 6
 105                  * in interleaved mode, and the GPU will then also alternate
 106                  * on bit 6, 9, and 10 for X, but the CPU may also optionally
 107                  * alternate based on bit 17 (XOR not disabled and XOR
 108                  * bit == 17).
 109                  */
 110                 dcc = I915_READ(DCC);
 111                 switch (dcc & DCC_ADDRESSING_MODE_MASK) {
 112                 case DCC_ADDRESSING_MODE_SINGLE_CHANNEL:
 113                 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_ASYMMETRIC:
 114                         swizzle_x = I915_BIT_6_SWIZZLE_NONE;
 115                         swizzle_y = I915_BIT_6_SWIZZLE_NONE;
 116                         break;
 117                 case DCC_ADDRESSING_MODE_DUAL_CHANNEL_INTERLEAVED:
 118                         if (IS_I915G(dev) || IS_I915GM(dev) ||
 119                             dcc & DCC_CHANNEL_XOR_DISABLE) {
 120                                 swizzle_x = I915_BIT_6_SWIZZLE_9_10;
 121                                 swizzle_y = I915_BIT_6_SWIZZLE_9;
 122                         } else if ((IS_I965GM(dev) || IS_GM45(dev)) &&
 123                                    (dcc & DCC_CHANNEL_XOR_BIT_17) == 0) {
 124                                 /* GM965/GM45 does either bit 11 or bit 17
 125                                  * swizzling.
 126                                  */
 127                                 swizzle_x = I915_BIT_6_SWIZZLE_9_10_11;
 128                                 swizzle_y = I915_BIT_6_SWIZZLE_9_11;
 129                         } else {
 130                                 /* Bit 17 or perhaps other swizzling */
 131                                 swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
 132                                 swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
 133                         }
 134                         break;
 135                 }
 136                 if (dcc == 0xffffffff) {
 137                         DRM_ERROR("Couldn't read from MCHBAR.  "
 138                                   "Disabling tiling.\n");
 139                         swizzle_x = I915_BIT_6_SWIZZLE_UNKNOWN;
 140                         swizzle_y = I915_BIT_6_SWIZZLE_UNKNOWN;
 141                 }
 142         } else {
 143                 /* The 965, G33, and newer, have a very flexible memory
 144                  * configuration.  It will enable dual-channel mode
 145                  * (interleaving) on as much memory as it can, and the GPU
 146                  * will additionally sometimes enable different bit 6
 147                  * swizzling for tiled objects from the CPU.
 148                  *
 149                  * Here's what I found on the G965:
 150                  *    slot fill         memory size  swizzling
 151                  * 0A   0B   1A   1B    1-ch   2-ch
 152                  * 512  0    0    0     512    0     O
 153                  * 512  0    512  0     16     1008  X
 154                  * 512  0    0    512   16     1008  X
 155                  * 0    512  0    512   16     1008  X
 156                  * 1024 1024 1024 0     2048   1024  O
 157                  *
 158                  * We could probably detect this based on either the DRB
 159                  * matching, which was the case for the swizzling required in
 160                  * the table above, or from the 1-ch value being less than
 161                  * the minimum size of a rank.
 162                  */
 163                 if (I915_READ16(C0DRB3) != I915_READ16(C1DRB3)) {
 164                         swizzle_x = I915_BIT_6_SWIZZLE_NONE;
 165                         swizzle_y = I915_BIT_6_SWIZZLE_NONE;
 166                 } else {
 167                         swizzle_x = I915_BIT_6_SWIZZLE_9_10;
 168                         swizzle_y = I915_BIT_6_SWIZZLE_9;
 169                 }
 170         }
 171
 172         dev_priv->mm.bit_6_swizzle_x = swizzle_x;
 173         dev_priv->mm.bit_6_swizzle_y = swizzle_y;
 174 }
 175
 176
 177 /**
 178  * Returns the size of the fence for a tiled object of the given size.
 179  */
 180 static int
 181 i915_get_fence_size(struct drm_device *dev, int size)
 182 {
 183         int i;
 184         int start;
 185
 186         if (IS_I965G(dev)) {
 187                 /* The 965 can have fences at any page boundary. */
 188                 return ALIGN(size, 4096);
 189         } else {
 190                 /* Align the size to a power of two greater than the smallest
 191                  * fence size.
 192                  */
 193                 if (IS_I9XX(dev))
 194                         start = 1024 * 1024;
 195                 else
 196                         start = 512 * 1024;
 197
 198                 for (i = start; i < size; i <<= 1)
 199                         ;
 200
 201                 return i;
 202         }
 203 }
 204
 205 /* Check pitch constriants for all chips & tiling formats */
 206 static bool
 207 i915_tiling_ok(struct drm_device *dev, int stride, int size, int tiling_mode)
 208 {
 209         int tile_width;
 210
 211         /* Linear is always fine */
 212         if (tiling_mode == I915_TILING_NONE)
 213                 return true;
 214
 215         if (tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
 216                 tile_width = 128;
 217         else
 218                 tile_width = 512;
 219
 220         /* 965+ just needs multiples of tile width */
 221         if (IS_I965G(dev)) {
 222                 if (stride & (tile_width - 1))
 223                         return false;
 224                 return true;
 225         }
 226
 227         /* Pre-965 needs power of two tile widths */
 228         if (stride < tile_width)
 229                 return false;
 230
 231         if (stride & (stride - 1))
 232                 return false;
 233
 234         /* We don't handle the aperture area covered by the fence being bigger
 235          * than the object size.
 236          */
 237         if (i915_get_fence_size(dev, size) != size)
 238                 return false;
 239
 240         return true;
 241 }
 242
 243 /**
 244  * Sets the tiling mode of an object, returning the required swizzling of
 245  * bit 6 of addresses in the object.
 246  */
 247 int
 248 i915_gem_set_tiling(struct drm_device *dev, void *data,
 249                    struct drm_file *file_priv)
 250 {
 251         struct drm_i915_gem_set_tiling *args = data;
 252         drm_i915_private_t *dev_priv = dev->dev_private;
 253         struct drm_gem_object *obj;
 254         struct drm_i915_gem_object *obj_priv;
 255
 256         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
 257         if (obj == NULL)
 258                 return -EINVAL;
 259         obj_priv = obj->driver_private;
 260
 261         if (!i915_tiling_ok(dev, args->stride, obj->size, args->tiling_mode)) {
 262                 drm_gem_object_unreference(obj);
 263                 return -EINVAL;
 264         }
 265
 266         mutex_lock(&dev->struct_mutex);
 267
 268         if (args->tiling_mode == I915_TILING_NONE) {
 269                 obj_priv->tiling_mode = I915_TILING_NONE;
 270                 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
 271         } else {
 272                 if (args->tiling_mode == I915_TILING_X)
 273                         args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
 274                 else
 275                         args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
 276                 /* If we can't handle the swizzling, make it untiled. */
 277                 if (args->swizzle_mode == I915_BIT_6_SWIZZLE_UNKNOWN) {
 278                         args->tiling_mode = I915_TILING_NONE;
 279                         args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
 280                 }
 281         }
 282         if (args->tiling_mode != obj_priv->tiling_mode) {
 283                 int ret;
 284
 285                 /* Unbind the object, as switching tiling means we're
 286                  * switching the cache organization due to fencing, probably.
 287                  */
 288                 ret = i915_gem_object_unbind(obj);
 289                 if (ret != 0) {
 290                         WARN(ret != -ERESTARTSYS,
 291                              "failed to unbind object for tiling switch");
 292                         args->tiling_mode = obj_priv->tiling_mode;
 293                         mutex_unlock(&dev->struct_mutex);
 294                         drm_gem_object_unreference(obj);
 295
 296                         return ret;
 297                 }
 298                 obj_priv->tiling_mode = args->tiling_mode;
 299         }
 300         obj_priv->stride = args->stride;
 301
 302         mutex_unlock(&dev->struct_mutex);
 303
 304         drm_gem_object_unreference(obj);
 305
 306         return 0;
 307 }
 308
 309 /**
 310  * Returns the current tiling mode and required bit 6 swizzling for the object.
 311  */
 312 int
 313 i915_gem_get_tiling(struct drm_device *dev, void *data,
 314                    struct drm_file *file_priv)
 315 {
 316         struct drm_i915_gem_get_tiling *args = data;
 317         drm_i915_private_t *dev_priv = dev->dev_private;
 318         struct drm_gem_object *obj;
 319         struct drm_i915_gem_object *obj_priv;
 320
 321         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
 322         if (obj == NULL)
 323                 return -EINVAL;
 324         obj_priv = obj->driver_private;
 325
 326         mutex_lock(&dev->struct_mutex);
 327
 328         args->tiling_mode = obj_priv->tiling_mode;
 329         switch (obj_priv->tiling_mode) {
 330         case I915_TILING_X:
 331                 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_x;
 332                 break;
 333         case I915_TILING_Y:
 334                 args->swizzle_mode = dev_priv->mm.bit_6_swizzle_y;
 335                 break;
 336         case I915_TILING_NONE:
 337                 args->swizzle_mode = I915_BIT_6_SWIZZLE_NONE;
 338                 break;
 339         default:
 340                 DRM_ERROR("unknown tiling mode\n");
 341         }
 342
 343         mutex_unlock(&dev->struct_mutex);
 344
 345         drm_gem_object_unreference(obj);
 346
 347         return 0;
 348 }