drivers/gpu/drm/msm/adreno/a3xx_gpu.c

   1 /*
   2  * Copyright (C) 2013 Red Hat
   3  * Author: Rob Clark <robdclark@gmail.com>
   4  *
   5  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 as published by
   9  * the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program.  If not, see <http://www.gnu.org/licenses/>.
  18  */
  19
  20 #ifdef CONFIG_MSM_OCMEM
  21 #  include <mach/ocmem.h>
  22 #endif
  23
  24 #include "a3xx_gpu.h"
  25
  26 #define A3XX_INT0_MASK \
  27         (A3XX_INT0_RBBM_AHB_ERROR |        \
  28          A3XX_INT0_RBBM_ATB_BUS_OVERFLOW | \
  29          A3XX_INT0_CP_T0_PACKET_IN_IB |    \
  30          A3XX_INT0_CP_OPCODE_ERROR |       \
  31          A3XX_INT0_CP_RESERVED_BIT_ERROR | \
  32          A3XX_INT0_CP_HW_FAULT |           \
  33          A3XX_INT0_CP_IB1_INT |            \
  34          A3XX_INT0_CP_IB2_INT |            \
  35          A3XX_INT0_CP_RB_INT |             \
  36          A3XX_INT0_CP_REG_PROTECT_FAULT |  \
  37          A3XX_INT0_CP_AHB_ERROR_HALT |     \
  38          A3XX_INT0_CACHE_FLUSH_TS |        \
  39          A3XX_INT0_UCHE_OOB_ACCESS)
  40
  41 extern bool hang_debug;
  42
  43 static void a3xx_dump(struct msm_gpu *gpu);
  44 static bool a3xx_idle(struct msm_gpu *gpu);
  45
  46 static bool a3xx_me_init(struct msm_gpu *gpu)
  47 {
  48         struct msm_ringbuffer *ring = gpu->rb[0];
  49
  50         OUT_PKT3(ring, CP_ME_INIT, 17);
  51         OUT_RING(ring, 0x000003f7);
  52         OUT_RING(ring, 0x00000000);
  53         OUT_RING(ring, 0x00000000);
  54         OUT_RING(ring, 0x00000000);
  55         OUT_RING(ring, 0x00000080);
  56         OUT_RING(ring, 0x00000100);
  57         OUT_RING(ring, 0x00000180);
  58         OUT_RING(ring, 0x00006600);
  59         OUT_RING(ring, 0x00000150);
  60         OUT_RING(ring, 0x0000014e);
  61         OUT_RING(ring, 0x00000154);
  62         OUT_RING(ring, 0x00000001);
  63         OUT_RING(ring, 0x00000000);
  64         OUT_RING(ring, 0x00000000);
  65         OUT_RING(ring, 0x00000000);
  66         OUT_RING(ring, 0x00000000);
  67         OUT_RING(ring, 0x00000000);
  68
  69         gpu->funcs->flush(gpu, ring);
  70         return a3xx_idle(gpu);
  71 }
  72
  73 static int a3xx_hw_init(struct msm_gpu *gpu)
  74 {
  75         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
  76         struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
  77         uint32_t *ptr, len;
  78         int i, ret;
  79
  80         DBG("%s", gpu->name);
  81
  82         if (adreno_is_a305(adreno_gpu)) {
  83                 /* Set up 16 deep read/write request queues: */
  84                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
  85                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
  86                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
  87                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
  88                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
  89                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
  90                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
  91                 /* Enable WR-REQ: */
  92                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
  93                 /* Set up round robin arbitration between both AXI ports: */
  94                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
  95                 /* Set up AOOO: */
  96                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
  97                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
  98         } else if (adreno_is_a306(adreno_gpu)) {
  99                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
 100                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x0000000a);
 101                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x0000000a);
 102         } else if (adreno_is_a320(adreno_gpu)) {
 103                 /* Set up 16 deep read/write request queues: */
 104                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
 105                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
 106                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
 107                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
 108                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
 109                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
 110                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
 111                 /* Enable WR-REQ: */
 112                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
 113                 /* Set up round robin arbitration between both AXI ports: */
 114                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
 115                 /* Set up AOOO: */
 116                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
 117                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
 118                 /* Enable 1K sort: */
 119                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x000000ff);
 120                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
 121
 122         } else if (adreno_is_a330v2(adreno_gpu)) {
 123                 /*
 124                  * Most of the VBIF registers on 8974v2 have the correct
 125                  * values at power on, so we won't modify those if we don't
 126                  * need to
 127                  */
 128                 /* Enable 1k sort: */
 129                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
 130                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
 131                 /* Enable WR-REQ: */
 132                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
 133                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
 134                 /* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
 135                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
 136
 137         } else if (adreno_is_a330(adreno_gpu)) {
 138                 /* Set up 16 deep read/write request queues: */
 139                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x18181818);
 140                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x18181818);
 141                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x18181818);
 142                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x18181818);
 143                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
 144                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x18181818);
 145                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x18181818);
 146                 /* Enable WR-REQ: */
 147                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
 148                 /* Set up round robin arbitration between both AXI ports: */
 149                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
 150                 /* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
 151                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0001);
 152                 /* Set up AOOO: */
 153                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003f);
 154                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003f003f);
 155                 /* Enable 1K sort: */
 156                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
 157                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
 158                 /* Disable VBIF clock gating. This is to enable AXI running
 159                  * higher frequency than GPU:
 160                  */
 161                 gpu_write(gpu, REG_A3XX_VBIF_CLKON, 0x00000001);
 162
 163         } else {
 164                 BUG();
 165         }
 166
 167         /* Make all blocks contribute to the GPU BUSY perf counter: */
 168         gpu_write(gpu, REG_A3XX_RBBM_GPU_BUSY_MASKED, 0xffffffff);
 169
 170         /* Tune the hystersis counters for SP and CP idle detection: */
 171         gpu_write(gpu, REG_A3XX_RBBM_SP_HYST_CNT, 0x10);
 172         gpu_write(gpu, REG_A3XX_RBBM_WAIT_IDLE_CLOCKS_CTL, 0x10);
 173
 174         /* Enable the RBBM error reporting bits.  This lets us get
 175          * useful information on failure:
 176          */
 177         gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL0, 0x00000001);
 178
 179         /* Enable AHB error reporting: */
 180         gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL1, 0xa6ffffff);
 181
 182         /* Turn on the power counters: */
 183         gpu_write(gpu, REG_A3XX_RBBM_RBBM_CTL, 0x00030000);
 184
 185         /* Turn on hang detection - this spews a lot of useful information
 186          * into the RBBM registers on a hang:
 187          */
 188         gpu_write(gpu, REG_A3XX_RBBM_INTERFACE_HANG_INT_CTL, 0x00010fff);
 189
 190         /* Enable 64-byte cacheline size. HW Default is 32-byte (0x000000E0): */
 191         gpu_write(gpu, REG_A3XX_UCHE_CACHE_MODE_CONTROL_REG, 0x00000001);
 192
 193         /* Enable Clock gating: */
 194         if (adreno_is_a306(adreno_gpu))
 195                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
 196         else if (adreno_is_a320(adreno_gpu))
 197                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbfffffff);
 198         else if (adreno_is_a330v2(adreno_gpu))
 199                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
 200         else if (adreno_is_a330(adreno_gpu))
 201                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbffcffff);
 202
 203         if (adreno_is_a330v2(adreno_gpu))
 204                 gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x05515455);
 205         else if (adreno_is_a330(adreno_gpu))
 206                 gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x00000000);
 207
 208         /* Set the OCMEM base address for A330, etc */
 209         if (a3xx_gpu->ocmem_hdl) {
 210                 gpu_write(gpu, REG_A3XX_RB_GMEM_BASE_ADDR,
 211                         (unsigned int)(a3xx_gpu->ocmem_base >> 14));
 212         }
 213
 214         /* Turn on performance counters: */
 215         gpu_write(gpu, REG_A3XX_RBBM_PERFCTR_CTL, 0x01);
 216
 217         /* Enable the perfcntrs that we use.. */
 218         for (i = 0; i < gpu->num_perfcntrs; i++) {
 219                 const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i];
 220                 gpu_write(gpu, perfcntr->select_reg, perfcntr->select_val);
 221         }
 222
 223         gpu_write(gpu, REG_A3XX_RBBM_INT_0_MASK, A3XX_INT0_MASK);
 224
 225         ret = adreno_hw_init(gpu);
 226         if (ret)
 227                 return ret;
 228
 229         /* setup access protection: */
 230         gpu_write(gpu, REG_A3XX_CP_PROTECT_CTRL, 0x00000007);
 231
 232         /* RBBM registers */
 233         gpu_write(gpu, REG_A3XX_CP_PROTECT(0), 0x63000040);
 234         gpu_write(gpu, REG_A3XX_CP_PROTECT(1), 0x62000080);
 235         gpu_write(gpu, REG_A3XX_CP_PROTECT(2), 0x600000cc);
 236         gpu_write(gpu, REG_A3XX_CP_PROTECT(3), 0x60000108);
 237         gpu_write(gpu, REG_A3XX_CP_PROTECT(4), 0x64000140);
 238         gpu_write(gpu, REG_A3XX_CP_PROTECT(5), 0x66000400);
 239
 240         /* CP registers */
 241         gpu_write(gpu, REG_A3XX_CP_PROTECT(6), 0x65000700);
 242         gpu_write(gpu, REG_A3XX_CP_PROTECT(7), 0x610007d8);
 243         gpu_write(gpu, REG_A3XX_CP_PROTECT(8), 0x620007e0);
 244         gpu_write(gpu, REG_A3XX_CP_PROTECT(9), 0x61001178);
 245         gpu_write(gpu, REG_A3XX_CP_PROTECT(10), 0x64001180);
 246
 247         /* RB registers */
 248         gpu_write(gpu, REG_A3XX_CP_PROTECT(11), 0x60003300);
 249
 250         /* VBIF registers */
 251         gpu_write(gpu, REG_A3XX_CP_PROTECT(12), 0x6b00c000);
 252
 253         /* NOTE: PM4/micro-engine firmware registers look to be the same
 254          * for a2xx and a3xx.. we could possibly push that part down to
 255          * adreno_gpu base class.  Or push both PM4 and PFP but
 256          * parameterize the pfp ucode addr/data registers..
 257          */
 258
 259         /* Load PM4: */
 260         ptr = (uint32_t *)(adreno_gpu->fw[ADRENO_FW_PM4]->data);
 261         len = adreno_gpu->fw[ADRENO_FW_PM4]->size / 4;
 262         DBG("loading PM4 ucode version: %x", ptr[1]);
 263
 264         gpu_write(gpu, REG_AXXX_CP_DEBUG,
 265                         AXXX_CP_DEBUG_DYNAMIC_CLK_DISABLE |
 266                         AXXX_CP_DEBUG_MIU_128BIT_WRITE_ENABLE);
 267         gpu_write(gpu, REG_AXXX_CP_ME_RAM_WADDR, 0);
 268         for (i = 1; i < len; i++)
 269                 gpu_write(gpu, REG_AXXX_CP_ME_RAM_DATA, ptr[i]);
 270
 271         /* Load PFP: */
 272         ptr = (uint32_t *)(adreno_gpu->fw[ADRENO_FW_PFP]->data);
 273         len = adreno_gpu->fw[ADRENO_FW_PFP]->size / 4;
 274         DBG("loading PFP ucode version: %x", ptr[5]);
 275
 276         gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_ADDR, 0);
 277         for (i = 1; i < len; i++)
 278                 gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_DATA, ptr[i]);
 279
 280         /* CP ROQ queue sizes (bytes) - RB:16, ST:16, IB1:32, IB2:64 */
 281         if (adreno_is_a305(adreno_gpu) || adreno_is_a306(adreno_gpu) ||
 282                         adreno_is_a320(adreno_gpu)) {
 283                 gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS,
 284                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB1_START(2) |
 285                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB2_START(6) |
 286                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_ST_START(14));
 287         } else if (adreno_is_a330(adreno_gpu)) {
 288                 /* NOTE: this (value take from downstream android driver)
 289                  * includes some bits outside of the known bitfields.  But
 290                  * A330 has this "MERCIU queue" thing too, which might
 291                  * explain a new bitfield or reshuffling:
 292                  */
 293                 gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS, 0x003e2008);
 294         }
 295
 296         /* clear ME_HALT to start micro engine */
 297         gpu_write(gpu, REG_AXXX_CP_ME_CNTL, 0);
 298
 299         return a3xx_me_init(gpu) ? 0 : -EINVAL;
 300 }
 301
 302 static void a3xx_recover(struct msm_gpu *gpu)
 303 {
 304         int i;
 305
 306         adreno_dump_info(gpu);
 307
 308         for (i = 0; i < 8; i++) {
 309                 printk("CP_SCRATCH_REG%d: %u\n", i,
 310                         gpu_read(gpu, REG_AXXX_CP_SCRATCH_REG0 + i));
 311         }
 312
 313         /* dump registers before resetting gpu, if enabled: */
 314         if (hang_debug)
 315                 a3xx_dump(gpu);
 316
 317         gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 1);
 318         gpu_read(gpu, REG_A3XX_RBBM_SW_RESET_CMD);
 319         gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 0);
 320         adreno_recover(gpu);
 321 }
 322
 323 static void a3xx_destroy(struct msm_gpu *gpu)
 324 {
 325         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 326         struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
 327
 328         DBG("%s", gpu->name);
 329
 330         adreno_gpu_cleanup(adreno_gpu);
 331
 332 #ifdef CONFIG_MSM_OCMEM
 333         if (a3xx_gpu->ocmem_base)
 334                 ocmem_free(OCMEM_GRAPHICS, a3xx_gpu->ocmem_hdl);
 335 #endif
 336
 337         kfree(a3xx_gpu);
 338 }
 339
 340 static bool a3xx_idle(struct msm_gpu *gpu)
 341 {
 342         /* wait for ringbuffer to drain: */
 343         if (!adreno_idle(gpu, gpu->rb[0]))
 344                 return false;
 345
 346         /* then wait for GPU to finish: */
 347         if (spin_until(!(gpu_read(gpu, REG_A3XX_RBBM_STATUS) &
 348                         A3XX_RBBM_STATUS_GPU_BUSY))) {
 349                 DRM_ERROR("%s: timeout waiting for GPU to idle!\n", gpu->name);
 350
 351                 /* TODO maybe we need to reset GPU here to recover from hang? */
 352                 return false;
 353         }
 354
 355         return true;
 356 }
 357
 358 static irqreturn_t a3xx_irq(struct msm_gpu *gpu)
 359 {
 360         uint32_t status;
 361
 362         status = gpu_read(gpu, REG_A3XX_RBBM_INT_0_STATUS);
 363         DBG("%s: %08x", gpu->name, status);
 364
 365         // TODO
 366
 367         gpu_write(gpu, REG_A3XX_RBBM_INT_CLEAR_CMD, status);
 368
 369         msm_gpu_retire(gpu);
 370
 371         return IRQ_HANDLED;
 372 }
 373
 374 static const unsigned int a3xx_registers[] = {
 375         0x0000, 0x0002, 0x0010, 0x0012, 0x0018, 0x0018, 0x0020, 0x0027,
 376         0x0029, 0x002b, 0x002e, 0x0033, 0x0040, 0x0042, 0x0050, 0x005c,
 377         0x0060, 0x006c, 0x0080, 0x0082, 0x0084, 0x0088, 0x0090, 0x00e5,
 378         0x00ea, 0x00ed, 0x0100, 0x0100, 0x0110, 0x0123, 0x01c0, 0x01c1,
 379         0x01c3, 0x01c5, 0x01c7, 0x01c7, 0x01d5, 0x01d9, 0x01dc, 0x01dd,
 380         0x01ea, 0x01ea, 0x01ee, 0x01f1, 0x01f5, 0x01f5, 0x01fc, 0x01ff,
 381         0x0440, 0x0440, 0x0443, 0x0443, 0x0445, 0x0445, 0x044d, 0x044f,
 382         0x0452, 0x0452, 0x0454, 0x046f, 0x047c, 0x047c, 0x047f, 0x047f,
 383         0x0578, 0x057f, 0x0600, 0x0602, 0x0605, 0x0607, 0x060a, 0x060e,
 384         0x0612, 0x0614, 0x0c01, 0x0c02, 0x0c06, 0x0c1d, 0x0c3d, 0x0c3f,
 385         0x0c48, 0x0c4b, 0x0c80, 0x0c80, 0x0c88, 0x0c8b, 0x0ca0, 0x0cb7,
 386         0x0cc0, 0x0cc1, 0x0cc6, 0x0cc7, 0x0ce4, 0x0ce5, 0x0e00, 0x0e05,
 387         0x0e0c, 0x0e0c, 0x0e22, 0x0e23, 0x0e41, 0x0e45, 0x0e64, 0x0e65,
 388         0x0e80, 0x0e82, 0x0e84, 0x0e89, 0x0ea0, 0x0ea1, 0x0ea4, 0x0ea7,
 389         0x0ec4, 0x0ecb, 0x0ee0, 0x0ee0, 0x0f00, 0x0f01, 0x0f03, 0x0f09,
 390         0x2040, 0x2040, 0x2044, 0x2044, 0x2048, 0x204d, 0x2068, 0x2069,
 391         0x206c, 0x206d, 0x2070, 0x2070, 0x2072, 0x2072, 0x2074, 0x2075,
 392         0x2079, 0x207a, 0x20c0, 0x20d3, 0x20e4, 0x20ef, 0x2100, 0x2109,
 393         0x210c, 0x210c, 0x210e, 0x210e, 0x2110, 0x2111, 0x2114, 0x2115,
 394         0x21e4, 0x21e4, 0x21ea, 0x21ea, 0x21ec, 0x21ed, 0x21f0, 0x21f0,
 395         0x2200, 0x2212, 0x2214, 0x2217, 0x221a, 0x221a, 0x2240, 0x227e,
 396         0x2280, 0x228b, 0x22c0, 0x22c0, 0x22c4, 0x22ce, 0x22d0, 0x22d8,
 397         0x22df, 0x22e6, 0x22e8, 0x22e9, 0x22ec, 0x22ec, 0x22f0, 0x22f7,
 398         0x22ff, 0x22ff, 0x2340, 0x2343, 0x2440, 0x2440, 0x2444, 0x2444,
 399         0x2448, 0x244d, 0x2468, 0x2469, 0x246c, 0x246d, 0x2470, 0x2470,
 400         0x2472, 0x2472, 0x2474, 0x2475, 0x2479, 0x247a, 0x24c0, 0x24d3,
 401         0x24e4, 0x24ef, 0x2500, 0x2509, 0x250c, 0x250c, 0x250e, 0x250e,
 402         0x2510, 0x2511, 0x2514, 0x2515, 0x25e4, 0x25e4, 0x25ea, 0x25ea,
 403         0x25ec, 0x25ed, 0x25f0, 0x25f0, 0x2600, 0x2612, 0x2614, 0x2617,
 404         0x261a, 0x261a, 0x2640, 0x267e, 0x2680, 0x268b, 0x26c0, 0x26c0,
 405         0x26c4, 0x26ce, 0x26d0, 0x26d8, 0x26df, 0x26e6, 0x26e8, 0x26e9,
 406         0x26ec, 0x26ec, 0x26f0, 0x26f7, 0x26ff, 0x26ff, 0x2740, 0x2743,
 407         0x300c, 0x300e, 0x301c, 0x301d, 0x302a, 0x302a, 0x302c, 0x302d,
 408         0x3030, 0x3031, 0x3034, 0x3036, 0x303c, 0x303c, 0x305e, 0x305f,
 409         ~0   /* sentinel */
 410 };
 411
 412 /* would be nice to not have to duplicate the _show() stuff with printk(): */
 413 static void a3xx_dump(struct msm_gpu *gpu)
 414 {
 415         printk("status:   %08x\n",
 416                         gpu_read(gpu, REG_A3XX_RBBM_STATUS));
 417         adreno_dump(gpu);
 418 }
 419
 420 static struct msm_gpu_state *a3xx_gpu_state_get(struct msm_gpu *gpu)
 421 {
 422         struct msm_gpu_state *state = kzalloc(sizeof(*state), GFP_KERNEL);
 423
 424         if (!state)
 425                 return ERR_PTR(-ENOMEM);
 426
 427         adreno_gpu_state_get(gpu, state);
 428
 429         state->rbbm_status = gpu_read(gpu, REG_A3XX_RBBM_STATUS);
 430
 431         return state;
 432 }
 433
 434 /* Register offset defines for A3XX */
 435 static const unsigned int a3xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
 436         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_AXXX_CP_RB_BASE),
 437         REG_ADRENO_SKIP(REG_ADRENO_CP_RB_BASE_HI),
 438         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR, REG_AXXX_CP_RB_RPTR_ADDR),
 439         REG_ADRENO_SKIP(REG_ADRENO_CP_RB_RPTR_ADDR_HI),
 440         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_AXXX_CP_RB_RPTR),
 441         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_AXXX_CP_RB_WPTR),
 442         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_AXXX_CP_RB_CNTL),
 443 };
 444
 445 static const struct adreno_gpu_funcs funcs = {
 446         .base = {
 447                 .get_param = adreno_get_param,
 448                 .hw_init = a3xx_hw_init,
 449                 .pm_suspend = msm_gpu_pm_suspend,
 450                 .pm_resume = msm_gpu_pm_resume,
 451                 .recover = a3xx_recover,
 452                 .submit = adreno_submit,
 453                 .flush = adreno_flush,
 454                 .active_ring = adreno_active_ring,
 455                 .irq = a3xx_irq,
 456                 .destroy = a3xx_destroy,
 457 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
 458                 .show = adreno_show,
 459 #endif
 460                 .gpu_state_get = a3xx_gpu_state_get,
 461                 .gpu_state_put = adreno_gpu_state_put,
 462         },
 463 };
 464
 465 static const struct msm_gpu_perfcntr perfcntrs[] = {
 466         { REG_A3XX_SP_PERFCOUNTER6_SELECT, REG_A3XX_RBBM_PERFCTR_SP_6_LO,
 467                         SP_ALU_ACTIVE_CYCLES, "ALUACTIVE" },
 468         { REG_A3XX_SP_PERFCOUNTER7_SELECT, REG_A3XX_RBBM_PERFCTR_SP_7_LO,
 469                         SP_FS_FULL_ALU_INSTRUCTIONS, "ALUFULL" },
 470 };
 471
 472 struct msm_gpu *a3xx_gpu_init(struct drm_device *dev)
 473 {
 474         struct a3xx_gpu *a3xx_gpu = NULL;
 475         struct adreno_gpu *adreno_gpu;
 476         struct msm_gpu *gpu;
 477         struct msm_drm_private *priv = dev->dev_private;
 478         struct platform_device *pdev = priv->gpu_pdev;
 479         int ret;
 480
 481         if (!pdev) {
 482                 dev_err(dev->dev, "no a3xx device\n");
 483                 ret = -ENXIO;
 484                 goto fail;
 485         }
 486
 487         a3xx_gpu = kzalloc(sizeof(*a3xx_gpu), GFP_KERNEL);
 488         if (!a3xx_gpu) {
 489                 ret = -ENOMEM;
 490                 goto fail;
 491         }
 492
 493         adreno_gpu = &a3xx_gpu->base;
 494         gpu = &adreno_gpu->base;
 495
 496         gpu->perfcntrs = perfcntrs;
 497         gpu->num_perfcntrs = ARRAY_SIZE(perfcntrs);
 498
 499         adreno_gpu->registers = a3xx_registers;
 500         adreno_gpu->reg_offsets = a3xx_register_offsets;
 501
 502         ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
 503         if (ret)
 504                 goto fail;
 505
 506         /* if needed, allocate gmem: */
 507         if (adreno_is_a330(adreno_gpu)) {
 508 #ifdef CONFIG_MSM_OCMEM
 509                 /* TODO this is different/missing upstream: */
 510                 struct ocmem_buf *ocmem_hdl =
 511                                 ocmem_allocate(OCMEM_GRAPHICS, adreno_gpu->gmem);
 512
 513                 a3xx_gpu->ocmem_hdl = ocmem_hdl;
 514                 a3xx_gpu->ocmem_base = ocmem_hdl->addr;
 515                 adreno_gpu->gmem = ocmem_hdl->len;
 516                 DBG("using %dK of OCMEM at 0x%08x", adreno_gpu->gmem / 1024,
 517                                 a3xx_gpu->ocmem_base);
 518 #endif
 519         }
 520
 521         if (!gpu->aspace) {
 522                 /* TODO we think it is possible to configure the GPU to
 523                  * restrict access to VRAM carveout.  But the required
 524                  * registers are unknown.  For now just bail out and
 525                  * limp along with just modesetting.  If it turns out
 526                  * to not be possible to restrict access, then we must
 527                  * implement a cmdstream validator.
 528                  */
 529                 dev_err(dev->dev, "No memory protection without IOMMU\n");
 530                 ret = -ENXIO;
 531                 goto fail;
 532         }
 533
 534         return gpu;
 535
 536 fail:
 537         if (a3xx_gpu)
 538                 a3xx_destroy(&a3xx_gpu->base.base);
 539
 540         return ERR_PTR(ret);
 541 }