drivers/gpu/drm/i915/i915_gem_render_state.c

   1 /*
   2  * Copyright © 2014 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  *    Mika Kuoppala <mika.kuoppala@intel.com>
  25  *
  26  */
  27
  28 #include "i915_drv.h"
  29 #include "intel_renderstate.h"
  30
  31 struct render_state {
  32         const struct intel_renderstate_rodata *rodata;
  33         struct i915_vma *vma;
  34         u32 aux_batch_size;
  35         u32 aux_batch_offset;
  36 };
  37
  38 static const struct intel_renderstate_rodata *
  39 render_state_get_rodata(const struct drm_i915_gem_request *req)
  40 {
  41         switch (INTEL_GEN(req->i915)) {
  42         case 6:
  43                 return &gen6_null_state;
  44         case 7:
  45                 return &gen7_null_state;
  46         case 8:
  47                 return &gen8_null_state;
  48         case 9:
  49                 return &gen9_null_state;
  50         }
  51
  52         return NULL;
  53 }
  54
  55 /*
  56  * Macro to add commands to auxiliary batch.
  57  * This macro only checks for page overflow before inserting the commands,
  58  * this is sufficient as the null state generator makes the final batch
  59  * with two passes to build command and state separately. At this point
  60  * the size of both are known and it compacts them by relocating the state
  61  * right after the commands taking care of aligment so we should sufficient
  62  * space below them for adding new commands.
  63  */
  64 #define OUT_BATCH(batch, i, val)                                \
  65         do {                                                    \
  66                 if (WARN_ON((i) >= PAGE_SIZE / sizeof(u32))) {  \
  67                         ret = -ENOSPC;                          \
  68                         goto err_out;                           \
  69                 }                                               \
  70                 (batch)[(i)++] = (val);                         \
  71         } while(0)
  72
  73 static int render_state_setup(struct render_state *so)
  74 {
  75         struct drm_device *dev = so->vma->vm->dev;
  76         const struct intel_renderstate_rodata *rodata = so->rodata;
  77         const bool has_64bit_reloc = INTEL_GEN(dev) >= 8;
  78         unsigned int i = 0, reloc_index = 0;
  79         struct page *page;
  80         u32 *d;
  81         int ret;
  82
  83         ret = i915_gem_object_set_to_cpu_domain(so->vma->obj, true);
  84         if (ret)
  85                 return ret;
  86
  87         page = i915_gem_object_get_dirty_page(so->vma->obj, 0);
  88         d = kmap(page);
  89
  90         while (i < rodata->batch_items) {
  91                 u32 s = rodata->batch[i];
  92
  93                 if (i * 4  == rodata->reloc[reloc_index]) {
  94                         u64 r = s + so->vma->node.start;
  95                         s = lower_32_bits(r);
  96                         if (has_64bit_reloc) {
  97                                 if (i + 1 >= rodata->batch_items ||
  98                                     rodata->batch[i + 1] != 0) {
  99                                         ret = -EINVAL;
 100                                         goto err_out;
 101                                 }
 102
 103                                 d[i++] = s;
 104                                 s = upper_32_bits(r);
 105                         }
 106
 107                         reloc_index++;
 108                 }
 109
 110                 d[i++] = s;
 111         }
 112
 113         while (i % CACHELINE_DWORDS)
 114                 OUT_BATCH(d, i, MI_NOOP);
 115
 116         so->aux_batch_offset = i * sizeof(u32);
 117
 118         if (HAS_POOLED_EU(dev)) {
 119                 /*
 120                  * We always program 3x6 pool config but depending upon which
 121                  * subslice is disabled HW drops down to appropriate config
 122                  * shown below.
 123                  *
 124                  * In the below table 2x6 config always refers to
 125                  * fused-down version, native 2x6 is not available and can
 126                  * be ignored
 127                  *
 128                  * SNo  subslices config                eu pool configuration
 129                  * -----------------------------------------------------------
 130                  * 1    3 subslices enabled (3x6)  -    0x00777000  (9+9)
 131                  * 2    ss0 disabled (2x6)         -    0x00777000  (3+9)
 132                  * 3    ss1 disabled (2x6)         -    0x00770000  (6+6)
 133                  * 4    ss2 disabled (2x6)         -    0x00007000  (9+3)
 134                  */
 135                 u32 eu_pool_config = 0x00777000;
 136
 137                 OUT_BATCH(d, i, GEN9_MEDIA_POOL_STATE);
 138                 OUT_BATCH(d, i, GEN9_MEDIA_POOL_ENABLE);
 139                 OUT_BATCH(d, i, eu_pool_config);
 140                 OUT_BATCH(d, i, 0);
 141                 OUT_BATCH(d, i, 0);
 142                 OUT_BATCH(d, i, 0);
 143         }
 144
 145         OUT_BATCH(d, i, MI_BATCH_BUFFER_END);
 146         so->aux_batch_size = (i * sizeof(u32)) - so->aux_batch_offset;
 147
 148         /*
 149          * Since we are sending length, we need to strictly conform to
 150          * all requirements. For Gen2 this must be a multiple of 8.
 151          */
 152         so->aux_batch_size = ALIGN(so->aux_batch_size, 8);
 153
 154         kunmap(page);
 155
 156         ret = i915_gem_object_set_to_gtt_domain(so->vma->obj, false);
 157         if (ret)
 158                 return ret;
 159
 160         if (rodata->reloc[reloc_index] != -1) {
 161                 DRM_ERROR("only %d relocs resolved\n", reloc_index);
 162                 return -EINVAL;
 163         }
 164
 165         return 0;
 166
 167 err_out:
 168         kunmap(page);
 169         return ret;
 170 }
 171
 172 #undef OUT_BATCH
 173
 174 int i915_gem_render_state_init(struct drm_i915_gem_request *req)
 175 {
 176         struct render_state so;
 177         struct drm_i915_gem_object *obj;
 178         int ret;
 179
 180         if (WARN_ON(req->engine->id != RCS))
 181                 return -ENOENT;
 182
 183         so.rodata = render_state_get_rodata(req);
 184         if (!so.rodata)
 185                 return 0;
 186
 187         if (so.rodata->batch_items * 4 > 4096)
 188                 return -EINVAL;
 189
 190         obj = i915_gem_object_create(&req->i915->drm, 4096);
 191         if (IS_ERR(obj))
 192                 return PTR_ERR(obj);
 193
 194         so.vma = i915_vma_create(obj, &req->i915->ggtt.base, NULL);
 195         if (IS_ERR(so.vma)) {
 196                 ret = PTR_ERR(so.vma);
 197                 goto err_obj;
 198         }
 199
 200         ret = i915_vma_pin(so.vma, 0, 0, PIN_GLOBAL);
 201         if (ret)
 202                 goto err_obj;
 203
 204         ret = render_state_setup(&so);
 205         if (ret)
 206                 goto err_unpin;
 207
 208         ret = req->engine->emit_bb_start(req, so.vma->node.start,
 209                                          so.rodata->batch_items * 4,
 210                                          I915_DISPATCH_SECURE);
 211         if (ret)
 212                 goto err_unpin;
 213
 214         if (so.aux_batch_size > 8) {
 215                 ret = req->engine->emit_bb_start(req,
 216                                                  (so.vma->node.start +
 217                                                   so.aux_batch_offset),
 218                                                  so.aux_batch_size,
 219                                                  I915_DISPATCH_SECURE);
 220                 if (ret)
 221                         goto err_unpin;
 222         }
 223
 224         i915_vma_move_to_active(so.vma, req, 0);
 225 err_unpin:
 226         i915_vma_unpin(so.vma);
 227 err_obj:
 228         i915_gem_object_put(obj);
 229         return ret;
 230 }