drivers/gpu/drm/radeon/r600.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/slab.h>
  29 #include <linux/seq_file.h>
  30 #include <linux/firmware.h>
  31 #include <linux/platform_device.h>
  32 #include <linux/module.h>
  33 #include <drm/drmP.h>
  34 #include <drm/radeon_drm.h>
  35 #include "radeon.h"
  36 #include "radeon_asic.h"
  37 #include "radeon_mode.h"
  38 #include "r600d.h"
  39 #include "atom.h"
  40 #include "avivod.h"
  41
  42 #define PFP_UCODE_SIZE 576
  43 #define PM4_UCODE_SIZE 1792
  44 #define RLC_UCODE_SIZE 768
  45 #define R700_PFP_UCODE_SIZE 848
  46 #define R700_PM4_UCODE_SIZE 1360
  47 #define R700_RLC_UCODE_SIZE 1024
  48 #define EVERGREEN_PFP_UCODE_SIZE 1120
  49 #define EVERGREEN_PM4_UCODE_SIZE 1376
  50 #define EVERGREEN_RLC_UCODE_SIZE 768
  51 #define CAYMAN_RLC_UCODE_SIZE 1024
  52 #define ARUBA_RLC_UCODE_SIZE 1536
  53
  54 /* Firmware Names */
  55 MODULE_FIRMWARE("radeon/R600_pfp.bin");
  56 MODULE_FIRMWARE("radeon/R600_me.bin");
  57 MODULE_FIRMWARE("radeon/RV610_pfp.bin");
  58 MODULE_FIRMWARE("radeon/RV610_me.bin");
  59 MODULE_FIRMWARE("radeon/RV630_pfp.bin");
  60 MODULE_FIRMWARE("radeon/RV630_me.bin");
  61 MODULE_FIRMWARE("radeon/RV620_pfp.bin");
  62 MODULE_FIRMWARE("radeon/RV620_me.bin");
  63 MODULE_FIRMWARE("radeon/RV635_pfp.bin");
  64 MODULE_FIRMWARE("radeon/RV635_me.bin");
  65 MODULE_FIRMWARE("radeon/RV670_pfp.bin");
  66 MODULE_FIRMWARE("radeon/RV670_me.bin");
  67 MODULE_FIRMWARE("radeon/RS780_pfp.bin");
  68 MODULE_FIRMWARE("radeon/RS780_me.bin");
  69 MODULE_FIRMWARE("radeon/RV770_pfp.bin");
  70 MODULE_FIRMWARE("radeon/RV770_me.bin");
  71 MODULE_FIRMWARE("radeon/RV730_pfp.bin");
  72 MODULE_FIRMWARE("radeon/RV730_me.bin");
  73 MODULE_FIRMWARE("radeon/RV710_pfp.bin");
  74 MODULE_FIRMWARE("radeon/RV710_me.bin");
  75 MODULE_FIRMWARE("radeon/R600_rlc.bin");
  76 MODULE_FIRMWARE("radeon/R700_rlc.bin");
  77 MODULE_FIRMWARE("radeon/CEDAR_pfp.bin");
  78 MODULE_FIRMWARE("radeon/CEDAR_me.bin");
  79 MODULE_FIRMWARE("radeon/CEDAR_rlc.bin");
  80 MODULE_FIRMWARE("radeon/REDWOOD_pfp.bin");
  81 MODULE_FIRMWARE("radeon/REDWOOD_me.bin");
  82 MODULE_FIRMWARE("radeon/REDWOOD_rlc.bin");
  83 MODULE_FIRMWARE("radeon/JUNIPER_pfp.bin");
  84 MODULE_FIRMWARE("radeon/JUNIPER_me.bin");
  85 MODULE_FIRMWARE("radeon/JUNIPER_rlc.bin");
  86 MODULE_FIRMWARE("radeon/CYPRESS_pfp.bin");
  87 MODULE_FIRMWARE("radeon/CYPRESS_me.bin");
  88 MODULE_FIRMWARE("radeon/CYPRESS_rlc.bin");
  89 MODULE_FIRMWARE("radeon/PALM_pfp.bin");
  90 MODULE_FIRMWARE("radeon/PALM_me.bin");
  91 MODULE_FIRMWARE("radeon/SUMO_rlc.bin");
  92 MODULE_FIRMWARE("radeon/SUMO_pfp.bin");
  93 MODULE_FIRMWARE("radeon/SUMO_me.bin");
  94 MODULE_FIRMWARE("radeon/SUMO2_pfp.bin");
  95 MODULE_FIRMWARE("radeon/SUMO2_me.bin");
  96
  97 int r600_debugfs_mc_info_init(struct radeon_device *rdev);
  98
  99 /* r600,rv610,rv630,rv620,rv635,rv670 */
 100 int r600_mc_wait_for_idle(struct radeon_device *rdev);
 101 static void r600_gpu_init(struct radeon_device *rdev);
 102 void r600_fini(struct radeon_device *rdev);
 103 void r600_irq_disable(struct radeon_device *rdev);
 104 static void r600_pcie_gen2_enable(struct radeon_device *rdev);
 105
 106 /* get temperature in millidegrees */
 107 int rv6xx_get_temp(struct radeon_device *rdev)
 108 {
 109         u32 temp = (RREG32(CG_THERMAL_STATUS) & ASIC_T_MASK) >>
 110                 ASIC_T_SHIFT;
 111         int actual_temp = temp & 0xff;
 112
 113         if (temp & 0x100)
 114                 actual_temp -= 256;
 115
 116         return actual_temp * 1000;
 117 }
 118
 119 void r600_pm_get_dynpm_state(struct radeon_device *rdev)
 120 {
 121         int i;
 122
 123         rdev->pm.dynpm_can_upclock = true;
 124         rdev->pm.dynpm_can_downclock = true;
 125
 126         /* power state array is low to high, default is first */
 127         if ((rdev->flags & RADEON_IS_IGP) || (rdev->family == CHIP_R600)) {
 128                 int min_power_state_index = 0;
 129
 130                 if (rdev->pm.num_power_states > 2)
 131                         min_power_state_index = 1;
 132
 133                 switch (rdev->pm.dynpm_planned_action) {
 134                 case DYNPM_ACTION_MINIMUM:
 135                         rdev->pm.requested_power_state_index = min_power_state_index;
 136                         rdev->pm.requested_clock_mode_index = 0;
 137                         rdev->pm.dynpm_can_downclock = false;
 138                         break;
 139                 case DYNPM_ACTION_DOWNCLOCK:
 140                         if (rdev->pm.current_power_state_index == min_power_state_index) {
 141                                 rdev->pm.requested_power_state_index = rdev->pm.current_power_state_index;
 142                                 rdev->pm.dynpm_can_downclock = false;
 143                         } else {
 144                                 if (rdev->pm.active_crtc_count > 1) {
 145                                         for (i = 0; i < rdev->pm.num_power_states; i++) {
 146                                                 if (rdev->pm.power_state[i].flags & RADEON_PM_STATE_SINGLE_DISPLAY_ONLY)
 147                                                         continue;
 148                                                 else if (i >= rdev->pm.current_power_state_index) {
 149                                                         rdev->pm.requested_power_state_index =
 150                                                                 rdev->pm.current_power_state_index;
 151                                                         break;
 152                                                 } else {
 153                                                         rdev->pm.requested_power_state_index = i;
 154                                                         break;
 155                                                 }
 156                                         }
 157                                 } else {
 158                                         if (rdev->pm.current_power_state_index == 0)
 159                                                 rdev->pm.requested_power_state_index =
 160                                                         rdev->pm.num_power_states - 1;
 161                                         else
 162                                                 rdev->pm.requested_power_state_index =
 163                                                         rdev->pm.current_power_state_index - 1;
 164                                 }
 165                         }
 166                         rdev->pm.requested_clock_mode_index = 0;
 167                         /* don't use the power state if crtcs are active and no display flag is set */
 168                         if ((rdev->pm.active_crtc_count > 0) &&
 169                             (rdev->pm.power_state[rdev->pm.requested_power_state_index].
 170                              clock_info[rdev->pm.requested_clock_mode_index].flags &
 171                              RADEON_PM_MODE_NO_DISPLAY)) {
 172                                 rdev->pm.requested_power_state_index++;
 173                         }
 174                         break;
 175                 case DYNPM_ACTION_UPCLOCK:
 176                         if (rdev->pm.current_power_state_index == (rdev->pm.num_power_states - 1)) {
 177                                 rdev->pm.requested_power_state_index = rdev->pm.current_power_state_index;
 178                                 rdev->pm.dynpm_can_upclock = false;
 179                         } else {
 180                                 if (rdev->pm.active_crtc_count > 1) {
 181                                         for (i = (rdev->pm.num_power_states - 1); i >= 0; i--) {
 182                                                 if (rdev->pm.power_state[i].flags & RADEON_PM_STATE_SINGLE_DISPLAY_ONLY)
 183                                                         continue;
 184                                                 else if (i <= rdev->pm.current_power_state_index) {
 185                                                         rdev->pm.requested_power_state_index =
 186                                                                 rdev->pm.current_power_state_index;
 187                                                         break;
 188                                                 } else {
 189                                                         rdev->pm.requested_power_state_index = i;
 190                                                         break;
 191                                                 }
 192                                         }
 193                                 } else
 194                                         rdev->pm.requested_power_state_index =
 195                                                 rdev->pm.current_power_state_index + 1;
 196                         }
 197                         rdev->pm.requested_clock_mode_index = 0;
 198                         break;
 199                 case DYNPM_ACTION_DEFAULT:
 200                         rdev->pm.requested_power_state_index = rdev->pm.default_power_state_index;
 201                         rdev->pm.requested_clock_mode_index = 0;
 202                         rdev->pm.dynpm_can_upclock = false;
 203                         break;
 204                 case DYNPM_ACTION_NONE:
 205                 default:
 206                         DRM_ERROR("Requested mode for not defined action\n");
 207                         return;
 208                 }
 209         } else {
 210                 /* XXX select a power state based on AC/DC, single/dualhead, etc. */
 211                 /* for now just select the first power state and switch between clock modes */
 212                 /* power state array is low to high, default is first (0) */
 213                 if (rdev->pm.active_crtc_count > 1) {
 214                         rdev->pm.requested_power_state_index = -1;
 215                         /* start at 1 as we don't want the default mode */
 216                         for (i = 1; i < rdev->pm.num_power_states; i++) {
 217                                 if (rdev->pm.power_state[i].flags & RADEON_PM_STATE_SINGLE_DISPLAY_ONLY)
 218                                         continue;
 219                                 else if ((rdev->pm.power_state[i].type == POWER_STATE_TYPE_PERFORMANCE) ||
 220                                          (rdev->pm.power_state[i].type == POWER_STATE_TYPE_BATTERY)) {
 221                                         rdev->pm.requested_power_state_index = i;
 222                                         break;
 223                                 }
 224                         }
 225                         /* if nothing selected, grab the default state. */
 226                         if (rdev->pm.requested_power_state_index == -1)
 227                                 rdev->pm.requested_power_state_index = 0;
 228                 } else
 229                         rdev->pm.requested_power_state_index = 1;
 230
 231                 switch (rdev->pm.dynpm_planned_action) {
 232                 case DYNPM_ACTION_MINIMUM:
 233                         rdev->pm.requested_clock_mode_index = 0;
 234                         rdev->pm.dynpm_can_downclock = false;
 235                         break;
 236                 case DYNPM_ACTION_DOWNCLOCK:
 237                         if (rdev->pm.requested_power_state_index == rdev->pm.current_power_state_index) {
 238                                 if (rdev->pm.current_clock_mode_index == 0) {
 239                                         rdev->pm.requested_clock_mode_index = 0;
 240                                         rdev->pm.dynpm_can_downclock = false;
 241                                 } else
 242                                         rdev->pm.requested_clock_mode_index =
 243                                                 rdev->pm.current_clock_mode_index - 1;
 244                         } else {
 245                                 rdev->pm.requested_clock_mode_index = 0;
 246                                 rdev->pm.dynpm_can_downclock = false;
 247                         }
 248                         /* don't use the power state if crtcs are active and no display flag is set */
 249                         if ((rdev->pm.active_crtc_count > 0) &&
 250                             (rdev->pm.power_state[rdev->pm.requested_power_state_index].
 251                              clock_info[rdev->pm.requested_clock_mode_index].flags &
 252                              RADEON_PM_MODE_NO_DISPLAY)) {
 253                                 rdev->pm.requested_clock_mode_index++;
 254                         }
 255                         break;
 256                 case DYNPM_ACTION_UPCLOCK:
 257                         if (rdev->pm.requested_power_state_index == rdev->pm.current_power_state_index) {
 258                                 if (rdev->pm.current_clock_mode_index ==
 259                                     (rdev->pm.power_state[rdev->pm.requested_power_state_index].num_clock_modes - 1)) {
 260                                         rdev->pm.requested_clock_mode_index = rdev->pm.current_clock_mode_index;
 261                                         rdev->pm.dynpm_can_upclock = false;
 262                                 } else
 263                                         rdev->pm.requested_clock_mode_index =
 264                                                 rdev->pm.current_clock_mode_index + 1;
 265                         } else {
 266                                 rdev->pm.requested_clock_mode_index =
 267                                         rdev->pm.power_state[rdev->pm.requested_power_state_index].num_clock_modes - 1;
 268                                 rdev->pm.dynpm_can_upclock = false;
 269                         }
 270                         break;
 271                 case DYNPM_ACTION_DEFAULT:
 272                         rdev->pm.requested_power_state_index = rdev->pm.default_power_state_index;
 273                         rdev->pm.requested_clock_mode_index = 0;
 274                         rdev->pm.dynpm_can_upclock = false;
 275                         break;
 276                 case DYNPM_ACTION_NONE:
 277                 default:
 278                         DRM_ERROR("Requested mode for not defined action\n");
 279                         return;
 280                 }
 281         }
 282
 283         DRM_DEBUG_DRIVER("Requested: e: %d m: %d p: %d\n",
 284                   rdev->pm.power_state[rdev->pm.requested_power_state_index].
 285                   clock_info[rdev->pm.requested_clock_mode_index].sclk,
 286                   rdev->pm.power_state[rdev->pm.requested_power_state_index].
 287                   clock_info[rdev->pm.requested_clock_mode_index].mclk,
 288                   rdev->pm.power_state[rdev->pm.requested_power_state_index].
 289                   pcie_lanes);
 290 }
 291
 292 void rs780_pm_init_profile(struct radeon_device *rdev)
 293 {
 294         if (rdev->pm.num_power_states == 2) {
 295                 /* default */
 296                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 297                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 298                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 299                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 0;
 300                 /* low sh */
 301                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = 0;
 302                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = 0;
 303                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 304                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 305                 /* mid sh */
 306                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = 0;
 307                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = 0;
 308                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 309                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 0;
 310                 /* high sh */
 311                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = 0;
 312                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = 1;
 313                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 314                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 0;
 315                 /* low mh */
 316                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = 0;
 317                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = 0;
 318                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 319                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 320                 /* mid mh */
 321                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = 0;
 322                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = 0;
 323                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 324                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 0;
 325                 /* high mh */
 326                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = 0;
 327                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = 1;
 328                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 329                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 0;
 330         } else if (rdev->pm.num_power_states == 3) {
 331                 /* default */
 332                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 333                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 334                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 335                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 0;
 336                 /* low sh */
 337                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = 1;
 338                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = 1;
 339                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 340                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 341                 /* mid sh */
 342                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = 1;
 343                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = 1;
 344                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 345                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 0;
 346                 /* high sh */
 347                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = 1;
 348                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = 2;
 349                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 350                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 0;
 351                 /* low mh */
 352                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = 1;
 353                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = 1;
 354                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 355                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 356                 /* mid mh */
 357                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = 1;
 358                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = 1;
 359                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 360                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 0;
 361                 /* high mh */
 362                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = 1;
 363                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = 2;
 364                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 365                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 0;
 366         } else {
 367                 /* default */
 368                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 369                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 370                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 371                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 0;
 372                 /* low sh */
 373                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = 2;
 374                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = 2;
 375                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 376                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 377                 /* mid sh */
 378                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = 2;
 379                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = 2;
 380                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 381                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 0;
 382                 /* high sh */
 383                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = 2;
 384                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = 3;
 385                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 386                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 0;
 387                 /* low mh */
 388                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = 2;
 389                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = 0;
 390                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 391                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 392                 /* mid mh */
 393                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = 2;
 394                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = 0;
 395                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 396                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 0;
 397                 /* high mh */
 398                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = 2;
 399                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = 3;
 400                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 401                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 0;
 402         }
 403 }
 404
 405 void r600_pm_init_profile(struct radeon_device *rdev)
 406 {
 407         int idx;
 408
 409         if (rdev->family == CHIP_R600) {
 410                 /* XXX */
 411                 /* default */
 412                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 413                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 414                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 415                 rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 0;
 416                 /* low sh */
 417                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 418                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 419                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 420                 rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 421                 /* mid sh */
 422                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 423                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 424                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 425                 rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 0;
 426                 /* high sh */
 427                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 428                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 429                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 430                 rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 0;
 431                 /* low mh */
 432                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 433                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 434                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 435                 rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 436                 /* mid mh */
 437                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 438                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 439                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 440                 rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 0;
 441                 /* high mh */
 442                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 443                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 444                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 445                 rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 0;
 446         } else {
 447                 if (rdev->pm.num_power_states < 4) {
 448                         /* default */
 449                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 450                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 451                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 452                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 2;
 453                         /* low sh */
 454                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = 1;
 455                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = 1;
 456                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 457                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 458                         /* mid sh */
 459                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = 1;
 460                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = 1;
 461                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 462                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 1;
 463                         /* high sh */
 464                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = 1;
 465                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = 1;
 466                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 467                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 2;
 468                         /* low mh */
 469                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = 2;
 470                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = 2;
 471                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 472                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 473                         /* low mh */
 474                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = 2;
 475                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = 2;
 476                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 477                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 1;
 478                         /* high mh */
 479                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = 2;
 480                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = 2;
 481                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 482                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 2;
 483                 } else {
 484                         /* default */
 485                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 486                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 487                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 488                         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 2;
 489                         /* low sh */
 490                         if (rdev->flags & RADEON_IS_MOBILITY)
 491                                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_BATTERY, 0);
 492                         else
 493                                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 0);
 494                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = idx;
 495                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = idx;
 496                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 497                         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 498                         /* mid sh */
 499                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = idx;
 500                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = idx;
 501                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 502                         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 1;
 503                         /* high sh */
 504                         idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 0);
 505                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = idx;
 506                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = idx;
 507                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 508                         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx = 2;
 509                         /* low mh */
 510                         if (rdev->flags & RADEON_IS_MOBILITY)
 511                                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_BATTERY, 1);
 512                         else
 513                                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 1);
 514                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = idx;
 515                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = idx;
 516                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 517                         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 518                         /* mid mh */
 519                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = idx;
 520                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = idx;
 521                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 522                         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 1;
 523                         /* high mh */
 524                         idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 1);
 525                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = idx;
 526                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = idx;
 527                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 528                         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx = 2;
 529                 }
 530         }
 531 }
 532
 533 void r600_pm_misc(struct radeon_device *rdev)
 534 {
 535         int req_ps_idx = rdev->pm.requested_power_state_index;
 536         int req_cm_idx = rdev->pm.requested_clock_mode_index;
 537         struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx];
 538         struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage;
 539
 540         if ((voltage->type == VOLTAGE_SW) && voltage->voltage) {
 541                 /* 0xff01 is a flag rather then an actual voltage */
 542                 if (voltage->voltage == 0xff01)
 543                         return;
 544                 if (voltage->voltage != rdev->pm.current_vddc) {
 545                         radeon_atom_set_voltage(rdev, voltage->voltage, SET_VOLTAGE_TYPE_ASIC_VDDC);
 546                         rdev->pm.current_vddc = voltage->voltage;
 547                         DRM_DEBUG_DRIVER("Setting: v: %d\n", voltage->voltage);
 548                 }
 549         }
 550 }
 551
 552 bool r600_gui_idle(struct radeon_device *rdev)
 553 {
 554         if (RREG32(GRBM_STATUS) & GUI_ACTIVE)
 555                 return false;
 556         else
 557                 return true;
 558 }
 559
 560 /* hpd for digital panel detect/disconnect */
 561 bool r600_hpd_sense(struct radeon_device *rdev, enum radeon_hpd_id hpd)
 562 {
 563         bool connected = false;
 564
 565         if (ASIC_IS_DCE3(rdev)) {
 566                 switch (hpd) {
 567                 case RADEON_HPD_1:
 568                         if (RREG32(DC_HPD1_INT_STATUS) & DC_HPDx_SENSE)
 569                                 connected = true;
 570                         break;
 571                 case RADEON_HPD_2:
 572                         if (RREG32(DC_HPD2_INT_STATUS) & DC_HPDx_SENSE)
 573                                 connected = true;
 574                         break;
 575                 case RADEON_HPD_3:
 576                         if (RREG32(DC_HPD3_INT_STATUS) & DC_HPDx_SENSE)
 577                                 connected = true;
 578                         break;
 579                 case RADEON_HPD_4:
 580                         if (RREG32(DC_HPD4_INT_STATUS) & DC_HPDx_SENSE)
 581                                 connected = true;
 582                         break;
 583                         /* DCE 3.2 */
 584                 case RADEON_HPD_5:
 585                         if (RREG32(DC_HPD5_INT_STATUS) & DC_HPDx_SENSE)
 586                                 connected = true;
 587                         break;
 588                 case RADEON_HPD_6:
 589                         if (RREG32(DC_HPD6_INT_STATUS) & DC_HPDx_SENSE)
 590                                 connected = true;
 591                         break;
 592                 default:
 593                         break;
 594                 }
 595         } else {
 596                 switch (hpd) {
 597                 case RADEON_HPD_1:
 598                         if (RREG32(DC_HOT_PLUG_DETECT1_INT_STATUS) & DC_HOT_PLUG_DETECTx_SENSE)
 599                                 connected = true;
 600                         break;
 601                 case RADEON_HPD_2:
 602                         if (RREG32(DC_HOT_PLUG_DETECT2_INT_STATUS) & DC_HOT_PLUG_DETECTx_SENSE)
 603                                 connected = true;
 604                         break;
 605                 case RADEON_HPD_3:
 606                         if (RREG32(DC_HOT_PLUG_DETECT3_INT_STATUS) & DC_HOT_PLUG_DETECTx_SENSE)
 607                                 connected = true;
 608                         break;
 609                 default:
 610                         break;
 611                 }
 612         }
 613         return connected;
 614 }
 615
 616 void r600_hpd_set_polarity(struct radeon_device *rdev,
 617                            enum radeon_hpd_id hpd)
 618 {
 619         u32 tmp;
 620         bool connected = r600_hpd_sense(rdev, hpd);
 621
 622         if (ASIC_IS_DCE3(rdev)) {
 623                 switch (hpd) {
 624                 case RADEON_HPD_1:
 625                         tmp = RREG32(DC_HPD1_INT_CONTROL);
 626                         if (connected)
 627                                 tmp &= ~DC_HPDx_INT_POLARITY;
 628                         else
 629                                 tmp |= DC_HPDx_INT_POLARITY;
 630                         WREG32(DC_HPD1_INT_CONTROL, tmp);
 631                         break;
 632                 case RADEON_HPD_2:
 633                         tmp = RREG32(DC_HPD2_INT_CONTROL);
 634                         if (connected)
 635                                 tmp &= ~DC_HPDx_INT_POLARITY;
 636                         else
 637                                 tmp |= DC_HPDx_INT_POLARITY;
 638                         WREG32(DC_HPD2_INT_CONTROL, tmp);
 639                         break;
 640                 case RADEON_HPD_3:
 641                         tmp = RREG32(DC_HPD3_INT_CONTROL);
 642                         if (connected)
 643                                 tmp &= ~DC_HPDx_INT_POLARITY;
 644                         else
 645                                 tmp |= DC_HPDx_INT_POLARITY;
 646                         WREG32(DC_HPD3_INT_CONTROL, tmp);
 647                         break;
 648                 case RADEON_HPD_4:
 649                         tmp = RREG32(DC_HPD4_INT_CONTROL);
 650                         if (connected)
 651                                 tmp &= ~DC_HPDx_INT_POLARITY;
 652                         else
 653                                 tmp |= DC_HPDx_INT_POLARITY;
 654                         WREG32(DC_HPD4_INT_CONTROL, tmp);
 655                         break;
 656                 case RADEON_HPD_5:
 657                         tmp = RREG32(DC_HPD5_INT_CONTROL);
 658                         if (connected)
 659                                 tmp &= ~DC_HPDx_INT_POLARITY;
 660                         else
 661                                 tmp |= DC_HPDx_INT_POLARITY;
 662                         WREG32(DC_HPD5_INT_CONTROL, tmp);
 663                         break;
 664                         /* DCE 3.2 */
 665                 case RADEON_HPD_6:
 666                         tmp = RREG32(DC_HPD6_INT_CONTROL);
 667                         if (connected)
 668                                 tmp &= ~DC_HPDx_INT_POLARITY;
 669                         else
 670                                 tmp |= DC_HPDx_INT_POLARITY;
 671                         WREG32(DC_HPD6_INT_CONTROL, tmp);
 672                         break;
 673                 default:
 674                         break;
 675                 }
 676         } else {
 677                 switch (hpd) {
 678                 case RADEON_HPD_1:
 679                         tmp = RREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL);
 680                         if (connected)
 681                                 tmp &= ~DC_HOT_PLUG_DETECTx_INT_POLARITY;
 682                         else
 683                                 tmp |= DC_HOT_PLUG_DETECTx_INT_POLARITY;
 684                         WREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL, tmp);
 685                         break;
 686                 case RADEON_HPD_2:
 687                         tmp = RREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL);
 688                         if (connected)
 689                                 tmp &= ~DC_HOT_PLUG_DETECTx_INT_POLARITY;
 690                         else
 691                                 tmp |= DC_HOT_PLUG_DETECTx_INT_POLARITY;
 692                         WREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL, tmp);
 693                         break;
 694                 case RADEON_HPD_3:
 695                         tmp = RREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL);
 696                         if (connected)
 697                                 tmp &= ~DC_HOT_PLUG_DETECTx_INT_POLARITY;
 698                         else
 699                                 tmp |= DC_HOT_PLUG_DETECTx_INT_POLARITY;
 700                         WREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL, tmp);
 701                         break;
 702                 default:
 703                         break;
 704                 }
 705         }
 706 }
 707
 708 void r600_hpd_init(struct radeon_device *rdev)
 709 {
 710         struct drm_device *dev = rdev->ddev;
 711         struct drm_connector *connector;
 712         unsigned enable = 0;
 713
 714         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
 715                 struct radeon_connector *radeon_connector = to_radeon_connector(connector);
 716
 717                 if (connector->connector_type == DRM_MODE_CONNECTOR_eDP ||
 718                     connector->connector_type == DRM_MODE_CONNECTOR_LVDS) {
 719                         /* don't try to enable hpd on eDP or LVDS avoid breaking the
 720                          * aux dp channel on imac and help (but not completely fix)
 721                          * https://bugzilla.redhat.com/show_bug.cgi?id=726143
 722                          */
 723                         continue;
 724                 }
 725                 if (ASIC_IS_DCE3(rdev)) {
 726                         u32 tmp = DC_HPDx_CONNECTION_TIMER(0x9c4) | DC_HPDx_RX_INT_TIMER(0xfa);
 727                         if (ASIC_IS_DCE32(rdev))
 728                                 tmp |= DC_HPDx_EN;
 729
 730                         switch (radeon_connector->hpd.hpd) {
 731                         case RADEON_HPD_1:
 732                                 WREG32(DC_HPD1_CONTROL, tmp);
 733                                 break;
 734                         case RADEON_HPD_2:
 735                                 WREG32(DC_HPD2_CONTROL, tmp);
 736                                 break;
 737                         case RADEON_HPD_3:
 738                                 WREG32(DC_HPD3_CONTROL, tmp);
 739                                 break;
 740                         case RADEON_HPD_4:
 741                                 WREG32(DC_HPD4_CONTROL, tmp);
 742                                 break;
 743                                 /* DCE 3.2 */
 744                         case RADEON_HPD_5:
 745                                 WREG32(DC_HPD5_CONTROL, tmp);
 746                                 break;
 747                         case RADEON_HPD_6:
 748                                 WREG32(DC_HPD6_CONTROL, tmp);
 749                                 break;
 750                         default:
 751                                 break;
 752                         }
 753                 } else {
 754                         switch (radeon_connector->hpd.hpd) {
 755                         case RADEON_HPD_1:
 756                                 WREG32(DC_HOT_PLUG_DETECT1_CONTROL, DC_HOT_PLUG_DETECTx_EN);
 757                                 break;
 758                         case RADEON_HPD_2:
 759                                 WREG32(DC_HOT_PLUG_DETECT2_CONTROL, DC_HOT_PLUG_DETECTx_EN);
 760                                 break;
 761                         case RADEON_HPD_3:
 762                                 WREG32(DC_HOT_PLUG_DETECT3_CONTROL, DC_HOT_PLUG_DETECTx_EN);
 763                                 break;
 764                         default:
 765                                 break;
 766                         }
 767                 }
 768                 enable |= 1 << radeon_connector->hpd.hpd;
 769                 radeon_hpd_set_polarity(rdev, radeon_connector->hpd.hpd);
 770         }
 771         radeon_irq_kms_enable_hpd(rdev, enable);
 772 }
 773
 774 void r600_hpd_fini(struct radeon_device *rdev)
 775 {
 776         struct drm_device *dev = rdev->ddev;
 777         struct drm_connector *connector;
 778         unsigned disable = 0;
 779
 780         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
 781                 struct radeon_connector *radeon_connector = to_radeon_connector(connector);
 782                 if (ASIC_IS_DCE3(rdev)) {
 783                         switch (radeon_connector->hpd.hpd) {
 784                         case RADEON_HPD_1:
 785                                 WREG32(DC_HPD1_CONTROL, 0);
 786                                 break;
 787                         case RADEON_HPD_2:
 788                                 WREG32(DC_HPD2_CONTROL, 0);
 789                                 break;
 790                         case RADEON_HPD_3:
 791                                 WREG32(DC_HPD3_CONTROL, 0);
 792                                 break;
 793                         case RADEON_HPD_4:
 794                                 WREG32(DC_HPD4_CONTROL, 0);
 795                                 break;
 796                                 /* DCE 3.2 */
 797                         case RADEON_HPD_5:
 798                                 WREG32(DC_HPD5_CONTROL, 0);
 799                                 break;
 800                         case RADEON_HPD_6:
 801                                 WREG32(DC_HPD6_CONTROL, 0);
 802                                 break;
 803                         default:
 804                                 break;
 805                         }
 806                 } else {
 807                         switch (radeon_connector->hpd.hpd) {
 808                         case RADEON_HPD_1:
 809                                 WREG32(DC_HOT_PLUG_DETECT1_CONTROL, 0);
 810                                 break;
 811                         case RADEON_HPD_2:
 812                                 WREG32(DC_HOT_PLUG_DETECT2_CONTROL, 0);
 813                                 break;
 814                         case RADEON_HPD_3:
 815                                 WREG32(DC_HOT_PLUG_DETECT3_CONTROL, 0);
 816                                 break;
 817                         default:
 818                                 break;
 819                         }
 820                 }
 821                 disable |= 1 << radeon_connector->hpd.hpd;
 822         }
 823         radeon_irq_kms_disable_hpd(rdev, disable);
 824 }
 825
 826 /*
 827  * R600 PCIE GART
 828  */
 829 void r600_pcie_gart_tlb_flush(struct radeon_device *rdev)
 830 {
 831         unsigned i;
 832         u32 tmp;
 833
 834         /* flush hdp cache so updates hit vram */
 835         if ((rdev->family >= CHIP_RV770) && (rdev->family <= CHIP_RV740) &&
 836             !(rdev->flags & RADEON_IS_AGP)) {
 837                 void __iomem *ptr = (void *)rdev->gart.ptr;
 838                 u32 tmp;
 839
 840                 /* r7xx hw bug.  write to HDP_DEBUG1 followed by fb read
 841                  * rather than write to HDP_REG_COHERENCY_FLUSH_CNTL
 842                  * This seems to cause problems on some AGP cards. Just use the old
 843                  * method for them.
 844                  */
 845                 WREG32(HDP_DEBUG1, 0);
 846                 tmp = readl((void __iomem *)ptr);
 847         } else
 848                 WREG32(R_005480_HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
 849
 850         WREG32(VM_CONTEXT0_INVALIDATION_LOW_ADDR, rdev->mc.gtt_start >> 12);
 851         WREG32(VM_CONTEXT0_INVALIDATION_HIGH_ADDR, (rdev->mc.gtt_end - 1) >> 12);
 852         WREG32(VM_CONTEXT0_REQUEST_RESPONSE, REQUEST_TYPE(1));
 853         for (i = 0; i < rdev->usec_timeout; i++) {
 854                 /* read MC_STATUS */
 855                 tmp = RREG32(VM_CONTEXT0_REQUEST_RESPONSE);
 856                 tmp = (tmp & RESPONSE_TYPE_MASK) >> RESPONSE_TYPE_SHIFT;
 857                 if (tmp == 2) {
 858                         printk(KERN_WARNING "[drm] r600 flush TLB failed\n");
 859                         return;
 860                 }
 861                 if (tmp) {
 862                         return;
 863                 }
 864                 udelay(1);
 865         }
 866 }
 867
 868 int r600_pcie_gart_init(struct radeon_device *rdev)
 869 {
 870         int r;
 871
 872         if (rdev->gart.robj) {
 873                 WARN(1, "R600 PCIE GART already initialized\n");
 874                 return 0;
 875         }
 876         /* Initialize common gart structure */
 877         r = radeon_gart_init(rdev);
 878         if (r)
 879                 return r;
 880         rdev->gart.table_size = rdev->gart.num_gpu_pages * 8;
 881         return radeon_gart_table_vram_alloc(rdev);
 882 }
 883
 884 static int r600_pcie_gart_enable(struct radeon_device *rdev)
 885 {
 886         u32 tmp;
 887         int r, i;
 888
 889         if (rdev->gart.robj == NULL) {
 890                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
 891                 return -EINVAL;
 892         }
 893         r = radeon_gart_table_vram_pin(rdev);
 894         if (r)
 895                 return r;
 896         radeon_gart_restore(rdev);
 897
 898         /* Setup L2 cache */
 899         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 900                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 901                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 902         WREG32(VM_L2_CNTL2, 0);
 903         WREG32(VM_L2_CNTL3, BANK_SELECT_0(0) | BANK_SELECT_1(1));
 904         /* Setup TLB control */
 905         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 906                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 907                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5) |
 908                 ENABLE_WAIT_L2_QUERY;
 909         WREG32(MC_VM_L1_TLB_MCB_RD_SYS_CNTL, tmp);
 910         WREG32(MC_VM_L1_TLB_MCB_WR_SYS_CNTL, tmp);
 911         WREG32(MC_VM_L1_TLB_MCB_RD_HDP_CNTL, tmp | ENABLE_L1_STRICT_ORDERING);
 912         WREG32(MC_VM_L1_TLB_MCB_WR_HDP_CNTL, tmp);
 913         WREG32(MC_VM_L1_TLB_MCD_RD_A_CNTL, tmp);
 914         WREG32(MC_VM_L1_TLB_MCD_WR_A_CNTL, tmp);
 915         WREG32(MC_VM_L1_TLB_MCD_RD_B_CNTL, tmp);
 916         WREG32(MC_VM_L1_TLB_MCD_WR_B_CNTL, tmp);
 917         WREG32(MC_VM_L1_TLB_MCB_RD_GFX_CNTL, tmp);
 918         WREG32(MC_VM_L1_TLB_MCB_WR_GFX_CNTL, tmp);
 919         WREG32(MC_VM_L1_TLB_MCB_RD_PDMA_CNTL, tmp);
 920         WREG32(MC_VM_L1_TLB_MCB_WR_PDMA_CNTL, tmp);
 921         WREG32(MC_VM_L1_TLB_MCB_RD_SEM_CNTL, tmp | ENABLE_SEMAPHORE_MODE);
 922         WREG32(MC_VM_L1_TLB_MCB_WR_SEM_CNTL, tmp | ENABLE_SEMAPHORE_MODE);
 923         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 924         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
 925         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
 926         WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
 927                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
 928         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 929                         (u32)(rdev->dummy_page.addr >> 12));
 930         for (i = 1; i < 7; i++)
 931                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 932
 933         r600_pcie_gart_tlb_flush(rdev);
 934         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 935                  (unsigned)(rdev->mc.gtt_size >> 20),
 936                  (unsigned long long)rdev->gart.table_addr);
 937         rdev->gart.ready = true;
 938         return 0;
 939 }
 940
 941 static void r600_pcie_gart_disable(struct radeon_device *rdev)
 942 {
 943         u32 tmp;
 944         int i;
 945
 946         /* Disable all tables */
 947         for (i = 0; i < 7; i++)
 948                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 949
 950         /* Disable L2 cache */
 951         WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
 952                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 953         WREG32(VM_L2_CNTL3, BANK_SELECT_0(0) | BANK_SELECT_1(1));
 954         /* Setup L1 TLB control */
 955         tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5) |
 956                 ENABLE_WAIT_L2_QUERY;
 957         WREG32(MC_VM_L1_TLB_MCD_RD_A_CNTL, tmp);
 958         WREG32(MC_VM_L1_TLB_MCD_WR_A_CNTL, tmp);
 959         WREG32(MC_VM_L1_TLB_MCD_RD_B_CNTL, tmp);
 960         WREG32(MC_VM_L1_TLB_MCD_WR_B_CNTL, tmp);
 961         WREG32(MC_VM_L1_TLB_MCB_RD_GFX_CNTL, tmp);
 962         WREG32(MC_VM_L1_TLB_MCB_WR_GFX_CNTL, tmp);
 963         WREG32(MC_VM_L1_TLB_MCB_RD_PDMA_CNTL, tmp);
 964         WREG32(MC_VM_L1_TLB_MCB_WR_PDMA_CNTL, tmp);
 965         WREG32(MC_VM_L1_TLB_MCB_RD_SEM_CNTL, tmp);
 966         WREG32(MC_VM_L1_TLB_MCB_WR_SEM_CNTL, tmp);
 967         WREG32(MC_VM_L1_TLB_MCB_RD_SYS_CNTL, tmp);
 968         WREG32(MC_VM_L1_TLB_MCB_WR_SYS_CNTL, tmp);
 969         WREG32(MC_VM_L1_TLB_MCB_RD_HDP_CNTL, tmp);
 970         WREG32(MC_VM_L1_TLB_MCB_WR_HDP_CNTL, tmp);
 971         radeon_gart_table_vram_unpin(rdev);
 972 }
 973
 974 static void r600_pcie_gart_fini(struct radeon_device *rdev)
 975 {
 976         radeon_gart_fini(rdev);
 977         r600_pcie_gart_disable(rdev);
 978         radeon_gart_table_vram_free(rdev);
 979 }
 980
 981 static void r600_agp_enable(struct radeon_device *rdev)
 982 {
 983         u32 tmp;
 984         int i;
 985
 986         /* Setup L2 cache */
 987         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 988                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 989                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 990         WREG32(VM_L2_CNTL2, 0);
 991         WREG32(VM_L2_CNTL3, BANK_SELECT_0(0) | BANK_SELECT_1(1));
 992         /* Setup TLB control */
 993         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 994                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 995                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5) |
 996                 ENABLE_WAIT_L2_QUERY;
 997         WREG32(MC_VM_L1_TLB_MCB_RD_SYS_CNTL, tmp);
 998         WREG32(MC_VM_L1_TLB_MCB_WR_SYS_CNTL, tmp);
 999         WREG32(MC_VM_L1_TLB_MCB_RD_HDP_CNTL, tmp | ENABLE_L1_STRICT_ORDERING);
1000         WREG32(MC_VM_L1_TLB_MCB_WR_HDP_CNTL, tmp);
1001         WREG32(MC_VM_L1_TLB_MCD_RD_A_CNTL, tmp);
1002         WREG32(MC_VM_L1_TLB_MCD_WR_A_CNTL, tmp);
1003         WREG32(MC_VM_L1_TLB_MCD_RD_B_CNTL, tmp);
1004         WREG32(MC_VM_L1_TLB_MCD_WR_B_CNTL, tmp);
1005         WREG32(MC_VM_L1_TLB_MCB_RD_GFX_CNTL, tmp);
1006         WREG32(MC_VM_L1_TLB_MCB_WR_GFX_CNTL, tmp);
1007         WREG32(MC_VM_L1_TLB_MCB_RD_PDMA_CNTL, tmp);
1008         WREG32(MC_VM_L1_TLB_MCB_WR_PDMA_CNTL, tmp);
1009         WREG32(MC_VM_L1_TLB_MCB_RD_SEM_CNTL, tmp | ENABLE_SEMAPHORE_MODE);
1010         WREG32(MC_VM_L1_TLB_MCB_WR_SEM_CNTL, tmp | ENABLE_SEMAPHORE_MODE);
1011         for (i = 0; i < 7; i++)
1012                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
1013 }
1014
1015 int r600_mc_wait_for_idle(struct radeon_device *rdev)
1016 {
1017         unsigned i;
1018         u32 tmp;
1019
1020         for (i = 0; i < rdev->usec_timeout; i++) {
1021                 /* read MC_STATUS */
1022                 tmp = RREG32(R_000E50_SRBM_STATUS) & 0x3F00;
1023                 if (!tmp)
1024                         return 0;
1025                 udelay(1);
1026         }
1027         return -1;
1028 }
1029
1030 static void r600_mc_program(struct radeon_device *rdev)
1031 {
1032         struct rv515_mc_save save;
1033         u32 tmp;
1034         int i, j;
1035
1036         /* Initialize HDP */
1037         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
1038                 WREG32((0x2c14 + j), 0x00000000);
1039                 WREG32((0x2c18 + j), 0x00000000);
1040                 WREG32((0x2c1c + j), 0x00000000);
1041                 WREG32((0x2c20 + j), 0x00000000);
1042                 WREG32((0x2c24 + j), 0x00000000);
1043         }
1044         WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
1045
1046         rv515_mc_stop(rdev, &save);
1047         if (r600_mc_wait_for_idle(rdev)) {
1048                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1049         }
1050         /* Lockout access through VGA aperture (doesn't exist before R600) */
1051         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
1052         /* Update configuration */
1053         if (rdev->flags & RADEON_IS_AGP) {
1054                 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
1055                         /* VRAM before AGP */
1056                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
1057                                 rdev->mc.vram_start >> 12);
1058                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
1059                                 rdev->mc.gtt_end >> 12);
1060                 } else {
1061                         /* VRAM after AGP */
1062                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
1063                                 rdev->mc.gtt_start >> 12);
1064                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
1065                                 rdev->mc.vram_end >> 12);
1066                 }
1067         } else {
1068                 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR, rdev->mc.vram_start >> 12);
1069                 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR, rdev->mc.vram_end >> 12);
1070         }
1071         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, rdev->vram_scratch.gpu_addr >> 12);
1072         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
1073         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
1074         WREG32(MC_VM_FB_LOCATION, tmp);
1075         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
1076         WREG32(HDP_NONSURFACE_INFO, (2 << 7));
1077         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
1078         if (rdev->flags & RADEON_IS_AGP) {
1079                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 22);
1080                 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 22);
1081                 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
1082         } else {
1083                 WREG32(MC_VM_AGP_BASE, 0);
1084                 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
1085                 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
1086         }
1087         if (r600_mc_wait_for_idle(rdev)) {
1088                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1089         }
1090         rv515_mc_resume(rdev, &save);
1091         /* we need to own VRAM, so turn off the VGA renderer here
1092          * to stop it overwriting our objects */
1093         rv515_vga_render_disable(rdev);
1094 }
1095
1096 /**
1097  * r600_vram_gtt_location - try to find VRAM & GTT location
1098  * @rdev: radeon device structure holding all necessary informations
1099  * @mc: memory controller structure holding memory informations
1100  *
1101  * Function will place try to place VRAM at same place as in CPU (PCI)
1102  * address space as some GPU seems to have issue when we reprogram at
1103  * different address space.
1104  *
1105  * If there is not enough space to fit the unvisible VRAM after the
1106  * aperture then we limit the VRAM size to the aperture.
1107  *
1108  * If we are using AGP then place VRAM adjacent to AGP aperture are we need
1109  * them to be in one from GPU point of view so that we can program GPU to
1110  * catch access outside them (weird GPU policy see ??).
1111  *
1112  * This function will never fails, worst case are limiting VRAM or GTT.
1113  *
1114  * Note: GTT start, end, size should be initialized before calling this
1115  * function on AGP platform.
1116  */
1117 static void r600_vram_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc)
1118 {
1119         u64 size_bf, size_af;
1120
1121         if (mc->mc_vram_size > 0xE0000000) {
1122                 /* leave room for at least 512M GTT */
1123                 dev_warn(rdev->dev, "limiting VRAM\n");
1124                 mc->real_vram_size = 0xE0000000;
1125                 mc->mc_vram_size = 0xE0000000;
1126         }
1127         if (rdev->flags & RADEON_IS_AGP) {
1128                 size_bf = mc->gtt_start;
1129                 size_af = 0xFFFFFFFF - mc->gtt_end;
1130                 if (size_bf > size_af) {
1131                         if (mc->mc_vram_size > size_bf) {
1132                                 dev_warn(rdev->dev, "limiting VRAM\n");
1133                                 mc->real_vram_size = size_bf;
1134                                 mc->mc_vram_size = size_bf;
1135                         }
1136                         mc->vram_start = mc->gtt_start - mc->mc_vram_size;
1137                 } else {
1138                         if (mc->mc_vram_size > size_af) {
1139                                 dev_warn(rdev->dev, "limiting VRAM\n");
1140                                 mc->real_vram_size = size_af;
1141                                 mc->mc_vram_size = size_af;
1142                         }
1143                         mc->vram_start = mc->gtt_end + 1;
1144                 }
1145                 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
1146                 dev_info(rdev->dev, "VRAM: %lluM 0x%08llX - 0x%08llX (%lluM used)\n",
1147                                 mc->mc_vram_size >> 20, mc->vram_start,
1148                                 mc->vram_end, mc->real_vram_size >> 20);
1149         } else {
1150                 u64 base = 0;
1151                 if (rdev->flags & RADEON_IS_IGP) {
1152                         base = RREG32(MC_VM_FB_LOCATION) & 0xFFFF;
1153                         base <<= 24;
1154                 }
1155                 radeon_vram_location(rdev, &rdev->mc, base);
1156                 rdev->mc.gtt_base_align = 0;
1157                 radeon_gtt_location(rdev, mc);
1158         }
1159 }
1160
1161 static int r600_mc_init(struct radeon_device *rdev)
1162 {
1163         u32 tmp;
1164         int chansize, numchan;
1165
1166         /* Get VRAM informations */
1167         rdev->mc.vram_is_ddr = true;
1168         tmp = RREG32(RAMCFG);
1169         if (tmp & CHANSIZE_OVERRIDE) {
1170                 chansize = 16;
1171         } else if (tmp & CHANSIZE_MASK) {
1172                 chansize = 64;
1173         } else {
1174                 chansize = 32;
1175         }
1176         tmp = RREG32(CHMAP);
1177         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
1178         case 0:
1179         default:
1180                 numchan = 1;
1181                 break;
1182         case 1:
1183                 numchan = 2;
1184                 break;
1185         case 2:
1186                 numchan = 4;
1187                 break;
1188         case 3:
1189                 numchan = 8;
1190                 break;
1191         }
1192         rdev->mc.vram_width = numchan * chansize;
1193         /* Could aper size report 0 ? */
1194         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
1195         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
1196         /* Setup GPU memory space */
1197         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
1198         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
1199         rdev->mc.visible_vram_size = rdev->mc.aper_size;
1200         r600_vram_gtt_location(rdev, &rdev->mc);
1201
1202         if (rdev->flags & RADEON_IS_IGP) {
1203                 rs690_pm_info(rdev);
1204                 rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
1205         }
1206         radeon_update_bandwidth_info(rdev);
1207         return 0;
1208 }
1209
1210 int r600_vram_scratch_init(struct radeon_device *rdev)
1211 {
1212         int r;
1213
1214         if (rdev->vram_scratch.robj == NULL) {
1215                 r = radeon_bo_create(rdev, RADEON_GPU_PAGE_SIZE,
1216                                      PAGE_SIZE, true, RADEON_GEM_DOMAIN_VRAM,
1217                                      NULL, &rdev->vram_scratch.robj);
1218                 if (r) {
1219                         return r;
1220                 }
1221         }
1222
1223         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
1224         if (unlikely(r != 0))
1225                 return r;
1226         r = radeon_bo_pin(rdev->vram_scratch.robj,
1227                           RADEON_GEM_DOMAIN_VRAM, &rdev->vram_scratch.gpu_addr);
1228         if (r) {
1229                 radeon_bo_unreserve(rdev->vram_scratch.robj);
1230                 return r;
1231         }
1232         r = radeon_bo_kmap(rdev->vram_scratch.robj,
1233                                 (void **)&rdev->vram_scratch.ptr);
1234         if (r)
1235                 radeon_bo_unpin(rdev->vram_scratch.robj);
1236         radeon_bo_unreserve(rdev->vram_scratch.robj);
1237
1238         return r;
1239 }
1240
1241 void r600_vram_scratch_fini(struct radeon_device *rdev)
1242 {
1243         int r;
1244
1245         if (rdev->vram_scratch.robj == NULL) {
1246                 return;
1247         }
1248         r = radeon_bo_reserve(rdev->vram_scratch.robj, false);
1249         if (likely(r == 0)) {
1250                 radeon_bo_kunmap(rdev->vram_scratch.robj);
1251                 radeon_bo_unpin(rdev->vram_scratch.robj);
1252                 radeon_bo_unreserve(rdev->vram_scratch.robj);
1253         }
1254         radeon_bo_unref(&rdev->vram_scratch.robj);
1255 }
1256
1257 /* We doesn't check that the GPU really needs a reset we simply do the
1258  * reset, it's up to the caller to determine if the GPU needs one. We
1259  * might add an helper function to check that.
1260  */
1261 static void r600_gpu_soft_reset_gfx(struct radeon_device *rdev)
1262 {
1263         u32 grbm_busy_mask = S_008010_VC_BUSY(1) | S_008010_VGT_BUSY_NO_DMA(1) |
1264                                 S_008010_VGT_BUSY(1) | S_008010_TA03_BUSY(1) |
1265                                 S_008010_TC_BUSY(1) | S_008010_SX_BUSY(1) |
1266                                 S_008010_SH_BUSY(1) | S_008010_SPI03_BUSY(1) |
1267                                 S_008010_SMX_BUSY(1) | S_008010_SC_BUSY(1) |
1268                                 S_008010_PA_BUSY(1) | S_008010_DB03_BUSY(1) |
1269                                 S_008010_CR_BUSY(1) | S_008010_CB03_BUSY(1) |
1270                                 S_008010_GUI_ACTIVE(1);
1271         u32 grbm2_busy_mask = S_008014_SPI0_BUSY(1) | S_008014_SPI1_BUSY(1) |
1272                         S_008014_SPI2_BUSY(1) | S_008014_SPI3_BUSY(1) |
1273                         S_008014_TA0_BUSY(1) | S_008014_TA1_BUSY(1) |
1274                         S_008014_TA2_BUSY(1) | S_008014_TA3_BUSY(1) |
1275                         S_008014_DB0_BUSY(1) | S_008014_DB1_BUSY(1) |
1276                         S_008014_DB2_BUSY(1) | S_008014_DB3_BUSY(1) |
1277                         S_008014_CB0_BUSY(1) | S_008014_CB1_BUSY(1) |
1278                         S_008014_CB2_BUSY(1) | S_008014_CB3_BUSY(1);
1279         u32 tmp;
1280
1281         if (!(RREG32(GRBM_STATUS) & GUI_ACTIVE))
1282                 return;
1283
1284         dev_info(rdev->dev, "  R_008010_GRBM_STATUS      = 0x%08X\n",
1285                 RREG32(R_008010_GRBM_STATUS));
1286         dev_info(rdev->dev, "  R_008014_GRBM_STATUS2     = 0x%08X\n",
1287                 RREG32(R_008014_GRBM_STATUS2));
1288         dev_info(rdev->dev, "  R_000E50_SRBM_STATUS      = 0x%08X\n",
1289                 RREG32(R_000E50_SRBM_STATUS));
1290         dev_info(rdev->dev, "  R_008674_CP_STALLED_STAT1 = 0x%08X\n",
1291                 RREG32(CP_STALLED_STAT1));
1292         dev_info(rdev->dev, "  R_008678_CP_STALLED_STAT2 = 0x%08X\n",
1293                 RREG32(CP_STALLED_STAT2));
1294         dev_info(rdev->dev, "  R_00867C_CP_BUSY_STAT     = 0x%08X\n",
1295                 RREG32(CP_BUSY_STAT));
1296         dev_info(rdev->dev, "  R_008680_CP_STAT          = 0x%08X\n",
1297                 RREG32(CP_STAT));
1298
1299         /* Disable CP parsing/prefetching */
1300         WREG32(R_0086D8_CP_ME_CNTL, S_0086D8_CP_ME_HALT(1));
1301
1302         /* Check if any of the rendering block is busy and reset it */
1303         if ((RREG32(R_008010_GRBM_STATUS) & grbm_busy_mask) ||
1304             (RREG32(R_008014_GRBM_STATUS2) & grbm2_busy_mask)) {
1305                 tmp = S_008020_SOFT_RESET_CR(1) |
1306                         S_008020_SOFT_RESET_DB(1) |
1307                         S_008020_SOFT_RESET_CB(1) |
1308                         S_008020_SOFT_RESET_PA(1) |
1309                         S_008020_SOFT_RESET_SC(1) |
1310                         S_008020_SOFT_RESET_SMX(1) |
1311                         S_008020_SOFT_RESET_SPI(1) |
1312                         S_008020_SOFT_RESET_SX(1) |
1313                         S_008020_SOFT_RESET_SH(1) |
1314                         S_008020_SOFT_RESET_TC(1) |
1315                         S_008020_SOFT_RESET_TA(1) |
1316                         S_008020_SOFT_RESET_VC(1) |
1317                         S_008020_SOFT_RESET_VGT(1);
1318                 dev_info(rdev->dev, "  R_008020_GRBM_SOFT_RESET=0x%08X\n", tmp);
1319                 WREG32(R_008020_GRBM_SOFT_RESET, tmp);
1320                 RREG32(R_008020_GRBM_SOFT_RESET);
1321                 mdelay(15);
1322                 WREG32(R_008020_GRBM_SOFT_RESET, 0);
1323         }
1324         /* Reset CP (we always reset CP) */
1325         tmp = S_008020_SOFT_RESET_CP(1);
1326         dev_info(rdev->dev, "R_008020_GRBM_SOFT_RESET=0x%08X\n", tmp);
1327         WREG32(R_008020_GRBM_SOFT_RESET, tmp);
1328         RREG32(R_008020_GRBM_SOFT_RESET);
1329         mdelay(15);
1330         WREG32(R_008020_GRBM_SOFT_RESET, 0);
1331
1332         dev_info(rdev->dev, "  R_008010_GRBM_STATUS      = 0x%08X\n",
1333                 RREG32(R_008010_GRBM_STATUS));
1334         dev_info(rdev->dev, "  R_008014_GRBM_STATUS2     = 0x%08X\n",
1335                 RREG32(R_008014_GRBM_STATUS2));
1336         dev_info(rdev->dev, "  R_000E50_SRBM_STATUS      = 0x%08X\n",
1337                 RREG32(R_000E50_SRBM_STATUS));
1338         dev_info(rdev->dev, "  R_008674_CP_STALLED_STAT1 = 0x%08X\n",
1339                 RREG32(CP_STALLED_STAT1));
1340         dev_info(rdev->dev, "  R_008678_CP_STALLED_STAT2 = 0x%08X\n",
1341                 RREG32(CP_STALLED_STAT2));
1342         dev_info(rdev->dev, "  R_00867C_CP_BUSY_STAT     = 0x%08X\n",
1343                 RREG32(CP_BUSY_STAT));
1344         dev_info(rdev->dev, "  R_008680_CP_STAT          = 0x%08X\n",
1345                 RREG32(CP_STAT));
1346
1347 }
1348
1349 static void r600_gpu_soft_reset_dma(struct radeon_device *rdev)
1350 {
1351         u32 tmp;
1352
1353         if (RREG32(DMA_STATUS_REG) & DMA_IDLE)
1354                 return;
1355
1356         dev_info(rdev->dev, "  R_00D034_DMA_STATUS_REG   = 0x%08X\n",
1357                 RREG32(DMA_STATUS_REG));
1358
1359         /* Disable DMA */
1360         tmp = RREG32(DMA_RB_CNTL);
1361         tmp &= ~DMA_RB_ENABLE;
1362         WREG32(DMA_RB_CNTL, tmp);
1363
1364         /* Reset dma */
1365         if (rdev->family >= CHIP_RV770)
1366                 WREG32(SRBM_SOFT_RESET, RV770_SOFT_RESET_DMA);
1367         else
1368                 WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA);
1369         RREG32(SRBM_SOFT_RESET);
1370         udelay(50);
1371         WREG32(SRBM_SOFT_RESET, 0);
1372
1373         dev_info(rdev->dev, "  R_00D034_DMA_STATUS_REG   = 0x%08X\n",
1374                 RREG32(DMA_STATUS_REG));
1375 }
1376
1377 static int r600_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
1378 {
1379         struct rv515_mc_save save;
1380
1381         if (!(RREG32(GRBM_STATUS) & GUI_ACTIVE))
1382                 reset_mask &= ~(RADEON_RESET_GFX | RADEON_RESET_COMPUTE);
1383
1384         if (RREG32(DMA_STATUS_REG) & DMA_IDLE)
1385                 reset_mask &= ~RADEON_RESET_DMA;
1386
1387         if (reset_mask == 0)
1388                 return 0;
1389
1390         dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
1391
1392         rv515_mc_stop(rdev, &save);
1393         if (r600_mc_wait_for_idle(rdev)) {
1394                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1395         }
1396
1397         if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE))
1398                 r600_gpu_soft_reset_gfx(rdev);
1399
1400         if (reset_mask & RADEON_RESET_DMA)
1401                 r600_gpu_soft_reset_dma(rdev);
1402
1403         /* Wait a little for things to settle down */
1404         mdelay(1);
1405
1406         rv515_mc_resume(rdev, &save);
1407         return 0;
1408 }
1409
1410 bool r600_gpu_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
1411 {
1412         u32 srbm_status;
1413         u32 grbm_status;
1414         u32 grbm_status2;
1415
1416         srbm_status = RREG32(R_000E50_SRBM_STATUS);
1417         grbm_status = RREG32(R_008010_GRBM_STATUS);
1418         grbm_status2 = RREG32(R_008014_GRBM_STATUS2);
1419         if (!G_008010_GUI_ACTIVE(grbm_status)) {
1420                 radeon_ring_lockup_update(ring);
1421                 return false;
1422         }
1423         /* force CP activities */
1424         radeon_ring_force_activity(rdev, ring);
1425         return radeon_ring_test_lockup(rdev, ring);
1426 }
1427
1428 /**
1429  * r600_dma_is_lockup - Check if the DMA engine is locked up
1430  *
1431  * @rdev: radeon_device pointer
1432  * @ring: radeon_ring structure holding ring information
1433  *
1434  * Check if the async DMA engine is locked up (r6xx-evergreen).
1435  * Returns true if the engine appears to be locked up, false if not.
1436  */
1437 bool r600_dma_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
1438 {
1439         u32 dma_status_reg;
1440
1441         dma_status_reg = RREG32(DMA_STATUS_REG);
1442         if (dma_status_reg & DMA_IDLE) {
1443                 radeon_ring_lockup_update(ring);
1444                 return false;
1445         }
1446         /* force ring activities */
1447         radeon_ring_force_activity(rdev, ring);
1448         return radeon_ring_test_lockup(rdev, ring);
1449 }
1450
1451 int r600_asic_reset(struct radeon_device *rdev)
1452 {
1453         return r600_gpu_soft_reset(rdev, (RADEON_RESET_GFX |
1454                                           RADEON_RESET_COMPUTE |
1455                                           RADEON_RESET_DMA));
1456 }
1457
1458 u32 r6xx_remap_render_backend(struct radeon_device *rdev,
1459                               u32 tiling_pipe_num,
1460                               u32 max_rb_num,
1461                               u32 total_max_rb_num,
1462                               u32 disabled_rb_mask)
1463 {
1464         u32 rendering_pipe_num, rb_num_width, req_rb_num;
1465         u32 pipe_rb_ratio, pipe_rb_remain;
1466         u32 data = 0, mask = 1 << (max_rb_num - 1);
1467         unsigned i, j;
1468
1469         /* mask out the RBs that don't exist on that asic */
1470         disabled_rb_mask |= (0xff << max_rb_num) & 0xff;
1471
1472         rendering_pipe_num = 1 << tiling_pipe_num;
1473         req_rb_num = total_max_rb_num - r600_count_pipe_bits(disabled_rb_mask);
1474         BUG_ON(rendering_pipe_num < req_rb_num);
1475
1476         pipe_rb_ratio = rendering_pipe_num / req_rb_num;
1477         pipe_rb_remain = rendering_pipe_num - pipe_rb_ratio * req_rb_num;
1478
1479         if (rdev->family <= CHIP_RV740) {
1480                 /* r6xx/r7xx */
1481                 rb_num_width = 2;
1482         } else {
1483                 /* eg+ */
1484                 rb_num_width = 4;
1485         }
1486
1487         for (i = 0; i < max_rb_num; i++) {
1488                 if (!(mask & disabled_rb_mask)) {
1489                         for (j = 0; j < pipe_rb_ratio; j++) {
1490                                 data <<= rb_num_width;
1491                                 data |= max_rb_num - i - 1;
1492                         }
1493                         if (pipe_rb_remain) {
1494                                 data <<= rb_num_width;
1495                                 data |= max_rb_num - i - 1;
1496                                 pipe_rb_remain--;
1497                         }
1498                 }
1499                 mask >>= 1;
1500         }
1501
1502         return data;
1503 }
1504
1505 int r600_count_pipe_bits(uint32_t val)
1506 {
1507         return hweight32(val);
1508 }
1509
1510 static void r600_gpu_init(struct radeon_device *rdev)
1511 {
1512         u32 tiling_config;
1513         u32 ramcfg;
1514         u32 cc_rb_backend_disable;
1515         u32 cc_gc_shader_pipe_config;
1516         u32 tmp;
1517         int i, j;
1518         u32 sq_config;
1519         u32 sq_gpr_resource_mgmt_1 = 0;
1520         u32 sq_gpr_resource_mgmt_2 = 0;
1521         u32 sq_thread_resource_mgmt = 0;
1522         u32 sq_stack_resource_mgmt_1 = 0;
1523         u32 sq_stack_resource_mgmt_2 = 0;
1524         u32 disabled_rb_mask;
1525
1526         rdev->config.r600.tiling_group_size = 256;
1527         switch (rdev->family) {
1528         case CHIP_R600:
1529                 rdev->config.r600.max_pipes = 4;
1530                 rdev->config.r600.max_tile_pipes = 8;
1531                 rdev->config.r600.max_simds = 4;
1532                 rdev->config.r600.max_backends = 4;
1533                 rdev->config.r600.max_gprs = 256;
1534                 rdev->config.r600.max_threads = 192;
1535                 rdev->config.r600.max_stack_entries = 256;
1536                 rdev->config.r600.max_hw_contexts = 8;
1537                 rdev->config.r600.max_gs_threads = 16;
1538                 rdev->config.r600.sx_max_export_size = 128;
1539                 rdev->config.r600.sx_max_export_pos_size = 16;
1540                 rdev->config.r600.sx_max_export_smx_size = 128;
1541                 rdev->config.r600.sq_num_cf_insts = 2;
1542                 break;
1543         case CHIP_RV630:
1544         case CHIP_RV635:
1545                 rdev->config.r600.max_pipes = 2;
1546                 rdev->config.r600.max_tile_pipes = 2;
1547                 rdev->config.r600.max_simds = 3;
1548                 rdev->config.r600.max_backends = 1;
1549                 rdev->config.r600.max_gprs = 128;
1550                 rdev->config.r600.max_threads = 192;
1551                 rdev->config.r600.max_stack_entries = 128;
1552                 rdev->config.r600.max_hw_contexts = 8;
1553                 rdev->config.r600.max_gs_threads = 4;
1554                 rdev->config.r600.sx_max_export_size = 128;
1555                 rdev->config.r600.sx_max_export_pos_size = 16;
1556                 rdev->config.r600.sx_max_export_smx_size = 128;
1557                 rdev->config.r600.sq_num_cf_insts = 2;
1558                 break;
1559         case CHIP_RV610:
1560         case CHIP_RV620:
1561         case CHIP_RS780:
1562         case CHIP_RS880:
1563                 rdev->config.r600.max_pipes = 1;
1564                 rdev->config.r600.max_tile_pipes = 1;
1565                 rdev->config.r600.max_simds = 2;
1566                 rdev->config.r600.max_backends = 1;
1567                 rdev->config.r600.max_gprs = 128;
1568                 rdev->config.r600.max_threads = 192;
1569                 rdev->config.r600.max_stack_entries = 128;
1570                 rdev->config.r600.max_hw_contexts = 4;
1571                 rdev->config.r600.max_gs_threads = 4;
1572                 rdev->config.r600.sx_max_export_size = 128;
1573                 rdev->config.r600.sx_max_export_pos_size = 16;
1574                 rdev->config.r600.sx_max_export_smx_size = 128;
1575                 rdev->config.r600.sq_num_cf_insts = 1;
1576                 break;
1577         case CHIP_RV670:
1578                 rdev->config.r600.max_pipes = 4;
1579                 rdev->config.r600.max_tile_pipes = 4;
1580                 rdev->config.r600.max_simds = 4;
1581                 rdev->config.r600.max_backends = 4;
1582                 rdev->config.r600.max_gprs = 192;
1583                 rdev->config.r600.max_threads = 192;
1584                 rdev->config.r600.max_stack_entries = 256;
1585                 rdev->config.r600.max_hw_contexts = 8;
1586                 rdev->config.r600.max_gs_threads = 16;
1587                 rdev->config.r600.sx_max_export_size = 128;
1588                 rdev->config.r600.sx_max_export_pos_size = 16;
1589                 rdev->config.r600.sx_max_export_smx_size = 128;
1590                 rdev->config.r600.sq_num_cf_insts = 2;
1591                 break;
1592         default:
1593                 break;
1594         }
1595
1596         /* Initialize HDP */
1597         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
1598                 WREG32((0x2c14 + j), 0x00000000);
1599                 WREG32((0x2c18 + j), 0x00000000);
1600                 WREG32((0x2c1c + j), 0x00000000);
1601                 WREG32((0x2c20 + j), 0x00000000);
1602                 WREG32((0x2c24 + j), 0x00000000);
1603         }
1604
1605         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
1606
1607         /* Setup tiling */
1608         tiling_config = 0;
1609         ramcfg = RREG32(RAMCFG);
1610         switch (rdev->config.r600.max_tile_pipes) {
1611         case 1:
1612                 tiling_config |= PIPE_TILING(0);
1613                 break;
1614         case 2:
1615                 tiling_config |= PIPE_TILING(1);
1616                 break;
1617         case 4:
1618                 tiling_config |= PIPE_TILING(2);
1619                 break;
1620         case 8:
1621                 tiling_config |= PIPE_TILING(3);
1622                 break;
1623         default:
1624                 break;
1625         }
1626         rdev->config.r600.tiling_npipes = rdev->config.r600.max_tile_pipes;
1627         rdev->config.r600.tiling_nbanks = 4 << ((ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
1628         tiling_config |= BANK_TILING((ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT);
1629         tiling_config |= GROUP_SIZE((ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT);
1630
1631         tmp = (ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT;
1632         if (tmp > 3) {
1633                 tiling_config |= ROW_TILING(3);
1634                 tiling_config |= SAMPLE_SPLIT(3);
1635         } else {
1636                 tiling_config |= ROW_TILING(tmp);
1637                 tiling_config |= SAMPLE_SPLIT(tmp);
1638         }
1639         tiling_config |= BANK_SWAPS(1);
1640
1641         cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
1642         tmp = R6XX_MAX_BACKENDS -
1643                 r600_count_pipe_bits((cc_rb_backend_disable >> 16) & R6XX_MAX_BACKENDS_MASK);
1644         if (tmp < rdev->config.r600.max_backends) {
1645                 rdev->config.r600.max_backends = tmp;
1646         }
1647
1648         cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0x00ffff00;
1649         tmp = R6XX_MAX_PIPES -
1650                 r600_count_pipe_bits((cc_gc_shader_pipe_config >> 8) & R6XX_MAX_PIPES_MASK);
1651         if (tmp < rdev->config.r600.max_pipes) {
1652                 rdev->config.r600.max_pipes = tmp;
1653         }
1654         tmp = R6XX_MAX_SIMDS -
1655                 r600_count_pipe_bits((cc_gc_shader_pipe_config >> 16) & R6XX_MAX_SIMDS_MASK);
1656         if (tmp < rdev->config.r600.max_simds) {
1657                 rdev->config.r600.max_simds = tmp;
1658         }
1659
1660         disabled_rb_mask = (RREG32(CC_RB_BACKEND_DISABLE) >> 16) & R6XX_MAX_BACKENDS_MASK;
1661         tmp = (tiling_config & PIPE_TILING__MASK) >> PIPE_TILING__SHIFT;
1662         tmp = r6xx_remap_render_backend(rdev, tmp, rdev->config.r600.max_backends,
1663                                         R6XX_MAX_BACKENDS, disabled_rb_mask);
1664         tiling_config |= tmp << 16;
1665         rdev->config.r600.backend_map = tmp;
1666
1667         rdev->config.r600.tile_config = tiling_config;
1668         WREG32(GB_TILING_CONFIG, tiling_config);
1669         WREG32(DCP_TILING_CONFIG, tiling_config & 0xffff);
1670         WREG32(HDP_TILING_CONFIG, tiling_config & 0xffff);
1671         WREG32(DMA_TILING_CONFIG, tiling_config & 0xffff);
1672
1673         tmp = R6XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
1674         WREG32(VGT_OUT_DEALLOC_CNTL, (tmp * 4) & DEALLOC_DIST_MASK);
1675         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((tmp * 4) - 2) & VTX_REUSE_DEPTH_MASK);
1676
1677         /* Setup some CP states */
1678         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) | ROQ_IB2_START(0x2b)));
1679         WREG32(CP_MEQ_THRESHOLDS, (MEQ_END(0x40) | ROQ_END(0x40)));
1680
1681         WREG32(TA_CNTL_AUX, (DISABLE_CUBE_ANISO | SYNC_GRADIENT |
1682                              SYNC_WALKER | SYNC_ALIGNER));
1683         /* Setup various GPU states */
1684         if (rdev->family == CHIP_RV670)
1685                 WREG32(ARB_GDEC_RD_CNTL, 0x00000021);
1686
1687         tmp = RREG32(SX_DEBUG_1);
1688         tmp |= SMX_EVENT_RELEASE;
1689         if ((rdev->family > CHIP_R600))
1690                 tmp |= ENABLE_NEW_SMX_ADDRESS;
1691         WREG32(SX_DEBUG_1, tmp);
1692
1693         if (((rdev->family) == CHIP_R600) ||
1694             ((rdev->family) == CHIP_RV630) ||
1695             ((rdev->family) == CHIP_RV610) ||
1696             ((rdev->family) == CHIP_RV620) ||
1697             ((rdev->family) == CHIP_RS780) ||
1698             ((rdev->family) == CHIP_RS880)) {
1699                 WREG32(DB_DEBUG, PREZ_MUST_WAIT_FOR_POSTZ_DONE);
1700         } else {
1701                 WREG32(DB_DEBUG, 0);
1702         }
1703         WREG32(DB_WATERMARKS, (DEPTH_FREE(4) | DEPTH_CACHELINE_FREE(16) |
1704                                DEPTH_FLUSH(16) | DEPTH_PENDING_FREE(4)));
1705
1706         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
1707         WREG32(VGT_NUM_INSTANCES, 0);
1708
1709         WREG32(SPI_CONFIG_CNTL, GPR_WRITE_PRIORITY(0));
1710         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(0));
1711
1712         tmp = RREG32(SQ_MS_FIFO_SIZES);
1713         if (((rdev->family) == CHIP_RV610) ||
1714             ((rdev->family) == CHIP_RV620) ||
1715             ((rdev->family) == CHIP_RS780) ||
1716             ((rdev->family) == CHIP_RS880)) {
1717                 tmp = (CACHE_FIFO_SIZE(0xa) |
1718                        FETCH_FIFO_HIWATER(0xa) |
1719                        DONE_FIFO_HIWATER(0xe0) |
1720                        ALU_UPDATE_FIFO_HIWATER(0x8));
1721         } else if (((rdev->family) == CHIP_R600) ||
1722                    ((rdev->family) == CHIP_RV630)) {
1723                 tmp &= ~DONE_FIFO_HIWATER(0xff);
1724                 tmp |= DONE_FIFO_HIWATER(0x4);
1725         }
1726         WREG32(SQ_MS_FIFO_SIZES, tmp);
1727
1728         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
1729          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
1730          */
1731         sq_config = RREG32(SQ_CONFIG);
1732         sq_config &= ~(PS_PRIO(3) |
1733                        VS_PRIO(3) |
1734                        GS_PRIO(3) |
1735                        ES_PRIO(3));
1736         sq_config |= (DX9_CONSTS |
1737                       VC_ENABLE |
1738                       PS_PRIO(0) |
1739                       VS_PRIO(1) |
1740                       GS_PRIO(2) |
1741                       ES_PRIO(3));
1742
1743         if ((rdev->family) == CHIP_R600) {
1744                 sq_gpr_resource_mgmt_1 = (NUM_PS_GPRS(124) |
1745                                           NUM_VS_GPRS(124) |
1746                                           NUM_CLAUSE_TEMP_GPRS(4));
1747                 sq_gpr_resource_mgmt_2 = (NUM_GS_GPRS(0) |
1748                                           NUM_ES_GPRS(0));
1749                 sq_thread_resource_mgmt = (NUM_PS_THREADS(136) |
1750                                            NUM_VS_THREADS(48) |
1751                                            NUM_GS_THREADS(4) |
1752                                            NUM_ES_THREADS(4));
1753                 sq_stack_resource_mgmt_1 = (NUM_PS_STACK_ENTRIES(128) |
1754                                             NUM_VS_STACK_ENTRIES(128));
1755                 sq_stack_resource_mgmt_2 = (NUM_GS_STACK_ENTRIES(0) |
1756                                             NUM_ES_STACK_ENTRIES(0));
1757         } else if (((rdev->family) == CHIP_RV610) ||
1758                    ((rdev->family) == CHIP_RV620) ||
1759                    ((rdev->family) == CHIP_RS780) ||
1760                    ((rdev->family) == CHIP_RS880)) {
1761                 /* no vertex cache */
1762                 sq_config &= ~VC_ENABLE;
1763
1764                 sq_gpr_resource_mgmt_1 = (NUM_PS_GPRS(44) |
1765                                           NUM_VS_GPRS(44) |
1766                                           NUM_CLAUSE_TEMP_GPRS(2));
1767                 sq_gpr_resource_mgmt_2 = (NUM_GS_GPRS(17) |
1768                                           NUM_ES_GPRS(17));
1769                 sq_thread_resource_mgmt = (NUM_PS_THREADS(79) |
1770                                            NUM_VS_THREADS(78) |
1771                                            NUM_GS_THREADS(4) |
1772                                            NUM_ES_THREADS(31));
1773                 sq_stack_resource_mgmt_1 = (NUM_PS_STACK_ENTRIES(40) |
1774                                             NUM_VS_STACK_ENTRIES(40));
1775                 sq_stack_resource_mgmt_2 = (NUM_GS_STACK_ENTRIES(32) |
1776                                             NUM_ES_STACK_ENTRIES(16));
1777         } else if (((rdev->family) == CHIP_RV630) ||
1778                    ((rdev->family) == CHIP_RV635)) {
1779                 sq_gpr_resource_mgmt_1 = (NUM_PS_GPRS(44) |
1780                                           NUM_VS_GPRS(44) |
1781                                           NUM_CLAUSE_TEMP_GPRS(2));
1782                 sq_gpr_resource_mgmt_2 = (NUM_GS_GPRS(18) |
1783                                           NUM_ES_GPRS(18));
1784                 sq_thread_resource_mgmt = (NUM_PS_THREADS(79) |
1785                                            NUM_VS_THREADS(78) |
1786                                            NUM_GS_THREADS(4) |
1787                                            NUM_ES_THREADS(31));
1788                 sq_stack_resource_mgmt_1 = (NUM_PS_STACK_ENTRIES(40) |
1789                                             NUM_VS_STACK_ENTRIES(40));
1790                 sq_stack_resource_mgmt_2 = (NUM_GS_STACK_ENTRIES(32) |
1791                                             NUM_ES_STACK_ENTRIES(16));
1792         } else if ((rdev->family) == CHIP_RV670) {
1793                 sq_gpr_resource_mgmt_1 = (NUM_PS_GPRS(44) |
1794                                           NUM_VS_GPRS(44) |
1795                                           NUM_CLAUSE_TEMP_GPRS(2));
1796                 sq_gpr_resource_mgmt_2 = (NUM_GS_GPRS(17) |
1797                                           NUM_ES_GPRS(17));
1798                 sq_thread_resource_mgmt = (NUM_PS_THREADS(79) |
1799                                            NUM_VS_THREADS(78) |
1800                                            NUM_GS_THREADS(4) |
1801                                            NUM_ES_THREADS(31));
1802                 sq_stack_resource_mgmt_1 = (NUM_PS_STACK_ENTRIES(64) |
1803                                             NUM_VS_STACK_ENTRIES(64));
1804                 sq_stack_resource_mgmt_2 = (NUM_GS_STACK_ENTRIES(64) |
1805                                             NUM_ES_STACK_ENTRIES(64));
1806         }
1807
1808         WREG32(SQ_CONFIG, sq_config);
1809         WREG32(SQ_GPR_RESOURCE_MGMT_1,  sq_gpr_resource_mgmt_1);
1810         WREG32(SQ_GPR_RESOURCE_MGMT_2,  sq_gpr_resource_mgmt_2);
1811         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1812         WREG32(SQ_STACK_RESOURCE_MGMT_1, sq_stack_resource_mgmt_1);
1813         WREG32(SQ_STACK_RESOURCE_MGMT_2, sq_stack_resource_mgmt_2);
1814
1815         if (((rdev->family) == CHIP_RV610) ||
1816             ((rdev->family) == CHIP_RV620) ||
1817             ((rdev->family) == CHIP_RS780) ||
1818             ((rdev->family) == CHIP_RS880)) {
1819                 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(TC_ONLY));
1820         } else {
1821                 WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC));
1822         }
1823
1824         /* More default values. 2D/3D driver should adjust as needed */
1825         WREG32(PA_SC_AA_SAMPLE_LOCS_2S, (S0_X(0xc) | S0_Y(0x4) |
1826                                          S1_X(0x4) | S1_Y(0xc)));
1827         WREG32(PA_SC_AA_SAMPLE_LOCS_4S, (S0_X(0xe) | S0_Y(0xe) |
1828                                          S1_X(0x2) | S1_Y(0x2) |
1829                                          S2_X(0xa) | S2_Y(0x6) |
1830                                          S3_X(0x6) | S3_Y(0xa)));
1831         WREG32(PA_SC_AA_SAMPLE_LOCS_8S_WD0, (S0_X(0xe) | S0_Y(0xb) |
1832                                              S1_X(0x4) | S1_Y(0xc) |
1833                                              S2_X(0x1) | S2_Y(0x6) |
1834                                              S3_X(0xa) | S3_Y(0xe)));
1835         WREG32(PA_SC_AA_SAMPLE_LOCS_8S_WD1, (S4_X(0x6) | S4_Y(0x1) |
1836                                              S5_X(0x0) | S5_Y(0x0) |
1837                                              S6_X(0xb) | S6_Y(0x4) |
1838                                              S7_X(0x7) | S7_Y(0x8)));
1839
1840         WREG32(VGT_STRMOUT_EN, 0);
1841         tmp = rdev->config.r600.max_pipes * 16;
1842         switch (rdev->family) {
1843         case CHIP_RV610:
1844         case CHIP_RV620:
1845         case CHIP_RS780:
1846         case CHIP_RS880:
1847                 tmp += 32;
1848                 break;
1849         case CHIP_RV670:
1850                 tmp += 128;
1851                 break;
1852         default:
1853                 break;
1854         }
1855         if (tmp > 256) {
1856                 tmp = 256;
1857         }
1858         WREG32(VGT_ES_PER_GS, 128);
1859         WREG32(VGT_GS_PER_ES, tmp);
1860         WREG32(VGT_GS_PER_VS, 2);
1861         WREG32(VGT_GS_VERTEX_REUSE, 16);
1862
1863         /* more default values. 2D/3D driver should adjust as needed */
1864         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
1865         WREG32(VGT_STRMOUT_EN, 0);
1866         WREG32(SX_MISC, 0);
1867         WREG32(PA_SC_MODE_CNTL, 0);
1868         WREG32(PA_SC_AA_CONFIG, 0);
1869         WREG32(PA_SC_LINE_STIPPLE, 0);
1870         WREG32(SPI_INPUT_Z, 0);
1871         WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
1872         WREG32(CB_COLOR7_FRAG, 0);
1873
1874         /* Clear render buffer base addresses */
1875         WREG32(CB_COLOR0_BASE, 0);
1876         WREG32(CB_COLOR1_BASE, 0);
1877         WREG32(CB_COLOR2_BASE, 0);
1878         WREG32(CB_COLOR3_BASE, 0);
1879         WREG32(CB_COLOR4_BASE, 0);
1880         WREG32(CB_COLOR5_BASE, 0);
1881         WREG32(CB_COLOR6_BASE, 0);
1882         WREG32(CB_COLOR7_BASE, 0);
1883         WREG32(CB_COLOR7_FRAG, 0);
1884
1885         switch (rdev->family) {
1886         case CHIP_RV610:
1887         case CHIP_RV620:
1888         case CHIP_RS780:
1889         case CHIP_RS880:
1890                 tmp = TC_L2_SIZE(8);
1891                 break;
1892         case CHIP_RV630:
1893         case CHIP_RV635:
1894                 tmp = TC_L2_SIZE(4);
1895                 break;
1896         case CHIP_R600:
1897                 tmp = TC_L2_SIZE(0) | L2_DISABLE_LATE_HIT;
1898                 break;
1899         default:
1900                 tmp = TC_L2_SIZE(0);
1901                 break;
1902         }
1903         WREG32(TC_CNTL, tmp);
1904
1905         tmp = RREG32(HDP_HOST_PATH_CNTL);
1906         WREG32(HDP_HOST_PATH_CNTL, tmp);
1907
1908         tmp = RREG32(ARB_POP);
1909         tmp |= ENABLE_TC128;
1910         WREG32(ARB_POP, tmp);
1911
1912         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
1913         WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
1914                                NUM_CLIP_SEQ(3)));
1915         WREG32(PA_SC_ENHANCE, FORCE_EOV_MAX_CLK_CNT(4095));
1916         WREG32(VC_ENHANCE, 0);
1917 }
1918
1919
1920 /*
1921  * Indirect registers accessor
1922  */
1923 u32 r600_pciep_rreg(struct radeon_device *rdev, u32 reg)
1924 {
1925         u32 r;
1926
1927         WREG32(PCIE_PORT_INDEX, ((reg) & 0xff));
1928         (void)RREG32(PCIE_PORT_INDEX);
1929         r = RREG32(PCIE_PORT_DATA);
1930         return r;
1931 }
1932
1933 void r600_pciep_wreg(struct radeon_device *rdev, u32 reg, u32 v)
1934 {
1935         WREG32(PCIE_PORT_INDEX, ((reg) & 0xff));
1936         (void)RREG32(PCIE_PORT_INDEX);
1937         WREG32(PCIE_PORT_DATA, (v));
1938         (void)RREG32(PCIE_PORT_DATA);
1939 }
1940
1941 /*
1942  * CP & Ring
1943  */
1944 void r600_cp_stop(struct radeon_device *rdev)
1945 {
1946         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
1947         WREG32(R_0086D8_CP_ME_CNTL, S_0086D8_CP_ME_HALT(1));
1948         WREG32(SCRATCH_UMSK, 0);
1949         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
1950 }
1951
1952 int r600_init_microcode(struct radeon_device *rdev)
1953 {
1954         struct platform_device *pdev;
1955         const char *chip_name;
1956         const char *rlc_chip_name;
1957         size_t pfp_req_size, me_req_size, rlc_req_size;
1958         char fw_name[30];
1959         int err;
1960
1961         DRM_DEBUG("\n");
1962
1963         pdev = platform_device_register_simple("radeon_cp", 0, NULL, 0);
1964         err = IS_ERR(pdev);
1965         if (err) {
1966                 printk(KERN_ERR "radeon_cp: Failed to register firmware\n");
1967                 return -EINVAL;
1968         }
1969
1970         switch (rdev->family) {
1971         case CHIP_R600:
1972                 chip_name = "R600";
1973                 rlc_chip_name = "R600";
1974                 break;
1975         case CHIP_RV610:
1976                 chip_name = "RV610";
1977                 rlc_chip_name = "R600";
1978                 break;
1979         case CHIP_RV630:
1980                 chip_name = "RV630";
1981                 rlc_chip_name = "R600";
1982                 break;
1983         case CHIP_RV620:
1984                 chip_name = "RV620";
1985                 rlc_chip_name = "R600";
1986                 break;
1987         case CHIP_RV635:
1988                 chip_name = "RV635";
1989                 rlc_chip_name = "R600";
1990                 break;
1991         case CHIP_RV670:
1992                 chip_name = "RV670";
1993                 rlc_chip_name = "R600";
1994                 break;
1995         case CHIP_RS780:
1996         case CHIP_RS880:
1997                 chip_name = "RS780";
1998                 rlc_chip_name = "R600";
1999                 break;
2000         case CHIP_RV770:
2001                 chip_name = "RV770";
2002                 rlc_chip_name = "R700";
2003                 break;
2004         case CHIP_RV730:
2005         case CHIP_RV740:
2006                 chip_name = "RV730";
2007                 rlc_chip_name = "R700";
2008                 break;
2009         case CHIP_RV710:
2010                 chip_name = "RV710";
2011                 rlc_chip_name = "R700";
2012                 break;
2013         case CHIP_CEDAR:
2014                 chip_name = "CEDAR";
2015                 rlc_chip_name = "CEDAR";
2016                 break;
2017         case CHIP_REDWOOD:
2018                 chip_name = "REDWOOD";
2019                 rlc_chip_name = "REDWOOD";
2020                 break;
2021         case CHIP_JUNIPER:
2022                 chip_name = "JUNIPER";
2023                 rlc_chip_name = "JUNIPER";
2024                 break;
2025         case CHIP_CYPRESS:
2026         case CHIP_HEMLOCK:
2027                 chip_name = "CYPRESS";
2028                 rlc_chip_name = "CYPRESS";
2029                 break;
2030         case CHIP_PALM:
2031                 chip_name = "PALM";
2032                 rlc_chip_name = "SUMO";
2033                 break;
2034         case CHIP_SUMO:
2035                 chip_name = "SUMO";
2036                 rlc_chip_name = "SUMO";
2037                 break;
2038         case CHIP_SUMO2:
2039                 chip_name = "SUMO2";
2040                 rlc_chip_name = "SUMO";
2041                 break;
2042         default: BUG();
2043         }
2044
2045         if (rdev->family >= CHIP_CEDAR) {
2046                 pfp_req_size = EVERGREEN_PFP_UCODE_SIZE * 4;
2047                 me_req_size = EVERGREEN_PM4_UCODE_SIZE * 4;
2048                 rlc_req_size = EVERGREEN_RLC_UCODE_SIZE * 4;
2049         } else if (rdev->family >= CHIP_RV770) {
2050                 pfp_req_size = R700_PFP_UCODE_SIZE * 4;
2051                 me_req_size = R700_PM4_UCODE_SIZE * 4;
2052                 rlc_req_size = R700_RLC_UCODE_SIZE * 4;
2053         } else {
2054                 pfp_req_size = PFP_UCODE_SIZE * 4;
2055                 me_req_size = PM4_UCODE_SIZE * 12;
2056                 rlc_req_size = RLC_UCODE_SIZE * 4;
2057         }
2058
2059         DRM_INFO("Loading %s Microcode\n", chip_name);
2060
2061         snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
2062         err = request_firmware(&rdev->pfp_fw, fw_name, &pdev->dev);
2063         if (err)
2064                 goto out;
2065         if (rdev->pfp_fw->size != pfp_req_size) {
2066                 printk(KERN_ERR
2067                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
2068                        rdev->pfp_fw->size, fw_name);
2069                 err = -EINVAL;
2070                 goto out;
2071         }
2072
2073         snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
2074         err = request_firmware(&rdev->me_fw, fw_name, &pdev->dev);
2075         if (err)
2076                 goto out;
2077         if (rdev->me_fw->size != me_req_size) {
2078                 printk(KERN_ERR
2079                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
2080                        rdev->me_fw->size, fw_name);
2081                 err = -EINVAL;
2082         }
2083
2084         snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", rlc_chip_name);
2085         err = request_firmware(&rdev->rlc_fw, fw_name, &pdev->dev);
2086         if (err)
2087                 goto out;
2088         if (rdev->rlc_fw->size != rlc_req_size) {
2089                 printk(KERN_ERR
2090                        "r600_rlc: Bogus length %zu in firmware \"%s\"\n",
2091                        rdev->rlc_fw->size, fw_name);
2092                 err = -EINVAL;
2093         }
2094
2095 out:
2096         platform_device_unregister(pdev);
2097
2098         if (err) {
2099                 if (err != -EINVAL)
2100                         printk(KERN_ERR
2101                                "r600_cp: Failed to load firmware \"%s\"\n",
2102                                fw_name);
2103                 release_firmware(rdev->pfp_fw);
2104                 rdev->pfp_fw = NULL;
2105                 release_firmware(rdev->me_fw);
2106                 rdev->me_fw = NULL;
2107                 release_firmware(rdev->rlc_fw);
2108                 rdev->rlc_fw = NULL;
2109         }
2110         return err;
2111 }
2112
2113 static int r600_cp_load_microcode(struct radeon_device *rdev)
2114 {
2115         const __be32 *fw_data;
2116         int i;
2117
2118         if (!rdev->me_fw || !rdev->pfp_fw)
2119                 return -EINVAL;
2120
2121         r600_cp_stop(rdev);
2122
2123         WREG32(CP_RB_CNTL,
2124 #ifdef __BIG_ENDIAN
2125                BUF_SWAP_32BIT |
2126 #endif
2127                RB_NO_UPDATE | RB_BLKSZ(15) | RB_BUFSZ(3));
2128
2129         /* Reset cp */
2130         WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
2131         RREG32(GRBM_SOFT_RESET);
2132         mdelay(15);
2133         WREG32(GRBM_SOFT_RESET, 0);
2134
2135         WREG32(CP_ME_RAM_WADDR, 0);
2136
2137         fw_data = (const __be32 *)rdev->me_fw->data;
2138         WREG32(CP_ME_RAM_WADDR, 0);
2139         for (i = 0; i < PM4_UCODE_SIZE * 3; i++)
2140                 WREG32(CP_ME_RAM_DATA,
2141                        be32_to_cpup(fw_data++));
2142
2143         fw_data = (const __be32 *)rdev->pfp_fw->data;
2144         WREG32(CP_PFP_UCODE_ADDR, 0);
2145         for (i = 0; i < PFP_UCODE_SIZE; i++)
2146                 WREG32(CP_PFP_UCODE_DATA,
2147                        be32_to_cpup(fw_data++));
2148
2149         WREG32(CP_PFP_UCODE_ADDR, 0);
2150         WREG32(CP_ME_RAM_WADDR, 0);
2151         WREG32(CP_ME_RAM_RADDR, 0);
2152         return 0;
2153 }
2154
2155 int r600_cp_start(struct radeon_device *rdev)
2156 {
2157         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
2158         int r;
2159         uint32_t cp_me;
2160
2161         r = radeon_ring_lock(rdev, ring, 7);
2162         if (r) {
2163                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
2164                 return r;
2165         }
2166         radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
2167         radeon_ring_write(ring, 0x1);
2168         if (rdev->family >= CHIP_RV770) {
2169                 radeon_ring_write(ring, 0x0);
2170                 radeon_ring_write(ring, rdev->config.rv770.max_hw_contexts - 1);
2171         } else {
2172                 radeon_ring_write(ring, 0x3);
2173                 radeon_ring_write(ring, rdev->config.r600.max_hw_contexts - 1);
2174         }
2175         radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
2176         radeon_ring_write(ring, 0);
2177         radeon_ring_write(ring, 0);
2178         radeon_ring_unlock_commit(rdev, ring);
2179
2180         cp_me = 0xff;
2181         WREG32(R_0086D8_CP_ME_CNTL, cp_me);
2182         return 0;
2183 }
2184
2185 int r600_cp_resume(struct radeon_device *rdev)
2186 {
2187         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
2188         u32 tmp;
2189         u32 rb_bufsz;
2190         int r;
2191
2192         /* Reset cp */
2193         WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
2194         RREG32(GRBM_SOFT_RESET);
2195         mdelay(15);
2196         WREG32(GRBM_SOFT_RESET, 0);
2197
2198         /* Set ring buffer size */
2199         rb_bufsz = drm_order(ring->ring_size / 8);
2200         tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2201 #ifdef __BIG_ENDIAN
2202         tmp |= BUF_SWAP_32BIT;
2203 #endif
2204         WREG32(CP_RB_CNTL, tmp);
2205         WREG32(CP_SEM_WAIT_TIMER, 0x0);
2206
2207         /* Set the write pointer delay */
2208         WREG32(CP_RB_WPTR_DELAY, 0);
2209
2210         /* Initialize the ring buffer's read and write pointers */
2211         WREG32(CP_RB_CNTL, tmp | RB_RPTR_WR_ENA);
2212         WREG32(CP_RB_RPTR_WR, 0);
2213         ring->wptr = 0;
2214         WREG32(CP_RB_WPTR, ring->wptr);
2215
2216         /* set the wb address whether it's enabled or not */
2217         WREG32(CP_RB_RPTR_ADDR,
2218                ((rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC));
2219         WREG32(CP_RB_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
2220         WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
2221
2222         if (rdev->wb.enabled)
2223                 WREG32(SCRATCH_UMSK, 0xff);
2224         else {
2225                 tmp |= RB_NO_UPDATE;
2226                 WREG32(SCRATCH_UMSK, 0);
2227         }
2228
2229         mdelay(1);
2230         WREG32(CP_RB_CNTL, tmp);
2231
2232         WREG32(CP_RB_BASE, ring->gpu_addr >> 8);
2233         WREG32(CP_DEBUG, (1 << 27) | (1 << 28));
2234
2235         ring->rptr = RREG32(CP_RB_RPTR);
2236
2237         r600_cp_start(rdev);
2238         ring->ready = true;
2239         r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, ring);
2240         if (r) {
2241                 ring->ready = false;
2242                 return r;
2243         }
2244         return 0;
2245 }
2246
2247 void r600_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size)
2248 {
2249         u32 rb_bufsz;
2250         int r;
2251
2252         /* Align ring size */
2253         rb_bufsz = drm_order(ring_size / 8);
2254         ring_size = (1 << (rb_bufsz + 1)) * 4;
2255         ring->ring_size = ring_size;
2256         ring->align_mask = 16 - 1;
2257
2258         if (radeon_ring_supports_scratch_reg(rdev, ring)) {
2259                 r = radeon_scratch_get(rdev, &ring->rptr_save_reg);
2260                 if (r) {
2261                         DRM_ERROR("failed to get scratch reg for rptr save (%d).\n", r);
2262                         ring->rptr_save_reg = 0;
2263                 }
2264         }
2265 }
2266
2267 void r600_cp_fini(struct radeon_device *rdev)
2268 {
2269         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
2270         r600_cp_stop(rdev);
2271         radeon_ring_fini(rdev, ring);
2272         radeon_scratch_free(rdev, ring->rptr_save_reg);
2273 }
2274
2275 /*
2276  * DMA
2277  * Starting with R600, the GPU has an asynchronous
2278  * DMA engine.  The programming model is very similar
2279  * to the 3D engine (ring buffer, IBs, etc.), but the
2280  * DMA controller has it's own packet format that is
2281  * different form the PM4 format used by the 3D engine.
2282  * It supports copying data, writing embedded data,
2283  * solid fills, and a number of other things.  It also
2284  * has support for tiling/detiling of buffers.
2285  */
2286 /**
2287  * r600_dma_stop - stop the async dma engine
2288  *
2289  * @rdev: radeon_device pointer
2290  *
2291  * Stop the async dma engine (r6xx-evergreen).
2292  */
2293 void r600_dma_stop(struct radeon_device *rdev)
2294 {
2295         u32 rb_cntl = RREG32(DMA_RB_CNTL);
2296
2297         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
2298
2299         rb_cntl &= ~DMA_RB_ENABLE;
2300         WREG32(DMA_RB_CNTL, rb_cntl);
2301
2302         rdev->ring[R600_RING_TYPE_DMA_INDEX].ready = false;
2303 }
2304
2305 /**
2306  * r600_dma_resume - setup and start the async dma engine
2307  *
2308  * @rdev: radeon_device pointer
2309  *
2310  * Set up the DMA ring buffer and enable it. (r6xx-evergreen).
2311  * Returns 0 for success, error for failure.
2312  */
2313 int r600_dma_resume(struct radeon_device *rdev)
2314 {
2315         struct radeon_ring *ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
2316         u32 rb_cntl, dma_cntl, ib_cntl;
2317         u32 rb_bufsz;
2318         int r;
2319
2320         /* Reset dma */
2321         if (rdev->family >= CHIP_RV770)
2322                 WREG32(SRBM_SOFT_RESET, RV770_SOFT_RESET_DMA);
2323         else
2324                 WREG32(SRBM_SOFT_RESET, SOFT_RESET_DMA);
2325         RREG32(SRBM_SOFT_RESET);
2326         udelay(50);
2327         WREG32(SRBM_SOFT_RESET, 0);
2328
2329         WREG32(DMA_SEM_INCOMPLETE_TIMER_CNTL, 0);
2330         WREG32(DMA_SEM_WAIT_FAIL_TIMER_CNTL, 0);
2331
2332         /* Set ring buffer size in dwords */
2333         rb_bufsz = drm_order(ring->ring_size / 4);
2334         rb_cntl = rb_bufsz << 1;
2335 #ifdef __BIG_ENDIAN
2336         rb_cntl |= DMA_RB_SWAP_ENABLE | DMA_RPTR_WRITEBACK_SWAP_ENABLE;
2337 #endif
2338         WREG32(DMA_RB_CNTL, rb_cntl);
2339
2340         /* Initialize the ring buffer's read and write pointers */
2341         WREG32(DMA_RB_RPTR, 0);
2342         WREG32(DMA_RB_WPTR, 0);
2343
2344         /* set the wb address whether it's enabled or not */
2345         WREG32(DMA_RB_RPTR_ADDR_HI,
2346                upper_32_bits(rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFF);
2347         WREG32(DMA_RB_RPTR_ADDR_LO,
2348                ((rdev->wb.gpu_addr + R600_WB_DMA_RPTR_OFFSET) & 0xFFFFFFFC));
2349
2350         if (rdev->wb.enabled)
2351                 rb_cntl |= DMA_RPTR_WRITEBACK_ENABLE;
2352
2353         WREG32(DMA_RB_BASE, ring->gpu_addr >> 8);
2354
2355         /* enable DMA IBs */
2356         ib_cntl = DMA_IB_ENABLE;
2357 #ifdef __BIG_ENDIAN
2358         ib_cntl |= DMA_IB_SWAP_ENABLE;
2359 #endif
2360         WREG32(DMA_IB_CNTL, ib_cntl);
2361
2362         dma_cntl = RREG32(DMA_CNTL);
2363         dma_cntl &= ~CTXEMPTY_INT_ENABLE;
2364         WREG32(DMA_CNTL, dma_cntl);
2365
2366         if (rdev->family >= CHIP_RV770)
2367                 WREG32(DMA_MODE, 1);
2368
2369         ring->wptr = 0;
2370         WREG32(DMA_RB_WPTR, ring->wptr << 2);
2371
2372         ring->rptr = RREG32(DMA_RB_RPTR) >> 2;
2373
2374         WREG32(DMA_RB_CNTL, rb_cntl | DMA_RB_ENABLE);
2375
2376         ring->ready = true;
2377
2378         r = radeon_ring_test(rdev, R600_RING_TYPE_DMA_INDEX, ring);
2379         if (r) {
2380                 ring->ready = false;
2381                 return r;
2382         }
2383
2384         radeon_ttm_set_active_vram_size(rdev, rdev->mc.real_vram_size);
2385
2386         return 0;
2387 }
2388
2389 /**
2390  * r600_dma_fini - tear down the async dma engine
2391  *
2392  * @rdev: radeon_device pointer
2393  *
2394  * Stop the async dma engine and free the ring (r6xx-evergreen).
2395  */
2396 void r600_dma_fini(struct radeon_device *rdev)
2397 {
2398         r600_dma_stop(rdev);
2399         radeon_ring_fini(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX]);
2400 }
2401
2402 /*
2403  * GPU scratch registers helpers function.
2404  */
2405 void r600_scratch_init(struct radeon_device *rdev)
2406 {
2407         int i;
2408
2409         rdev->scratch.num_reg = 7;
2410         rdev->scratch.reg_base = SCRATCH_REG0;
2411         for (i = 0; i < rdev->scratch.num_reg; i++) {
2412                 rdev->scratch.free[i] = true;
2413                 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
2414         }
2415 }
2416
2417 int r600_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
2418 {
2419         uint32_t scratch;
2420         uint32_t tmp = 0;
2421         unsigned i;
2422         int r;
2423
2424         r = radeon_scratch_get(rdev, &scratch);
2425         if (r) {
2426                 DRM_ERROR("radeon: cp failed to get scratch reg (%d).\n", r);
2427                 return r;
2428         }
2429         WREG32(scratch, 0xCAFEDEAD);
2430         r = radeon_ring_lock(rdev, ring, 3);
2431         if (r) {
2432                 DRM_ERROR("radeon: cp failed to lock ring %d (%d).\n", ring->idx, r);
2433                 radeon_scratch_free(rdev, scratch);
2434                 return r;
2435         }
2436         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
2437         radeon_ring_write(ring, ((scratch - PACKET3_SET_CONFIG_REG_OFFSET) >> 2));
2438         radeon_ring_write(ring, 0xDEADBEEF);
2439         radeon_ring_unlock_commit(rdev, ring);
2440         for (i = 0; i < rdev->usec_timeout; i++) {
2441                 tmp = RREG32(scratch);
2442                 if (tmp == 0xDEADBEEF)
2443                         break;
2444                 DRM_UDELAY(1);
2445         }
2446         if (i < rdev->usec_timeout) {
2447                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
2448         } else {
2449                 DRM_ERROR("radeon: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
2450                           ring->idx, scratch, tmp);
2451                 r = -EINVAL;
2452         }
2453         radeon_scratch_free(rdev, scratch);
2454         return r;
2455 }
2456
2457 /**
2458  * r600_dma_ring_test - simple async dma engine test
2459  *
2460  * @rdev: radeon_device pointer
2461  * @ring: radeon_ring structure holding ring information
2462  *
2463  * Test the DMA engine by writing using it to write an
2464  * value to memory. (r6xx-SI).
2465  * Returns 0 for success, error for failure.
2466  */
2467 int r600_dma_ring_test(struct radeon_device *rdev,
2468                        struct radeon_ring *ring)
2469 {
2470         unsigned i;
2471         int r;
2472         void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
2473         u32 tmp;
2474
2475         if (!ptr) {
2476                 DRM_ERROR("invalid vram scratch pointer\n");
2477                 return -EINVAL;
2478         }
2479
2480         tmp = 0xCAFEDEAD;
2481         writel(tmp, ptr);
2482
2483         r = radeon_ring_lock(rdev, ring, 4);
2484         if (r) {
2485                 DRM_ERROR("radeon: dma failed to lock ring %d (%d).\n", ring->idx, r);
2486                 return r;
2487         }
2488         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
2489         radeon_ring_write(ring, rdev->vram_scratch.gpu_addr & 0xfffffffc);
2490         radeon_ring_write(ring, upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff);
2491         radeon_ring_write(ring, 0xDEADBEEF);
2492         radeon_ring_unlock_commit(rdev, ring);
2493
2494         for (i = 0; i < rdev->usec_timeout; i++) {
2495                 tmp = readl(ptr);
2496                 if (tmp == 0xDEADBEEF)
2497                         break;
2498                 DRM_UDELAY(1);
2499         }
2500
2501         if (i < rdev->usec_timeout) {
2502                 DRM_INFO("ring test on %d succeeded in %d usecs\n", ring->idx, i);
2503         } else {
2504                 DRM_ERROR("radeon: ring %d test failed (0x%08X)\n",
2505                           ring->idx, tmp);
2506                 r = -EINVAL;
2507         }
2508         return r;
2509 }
2510
2511 /*
2512  * CP fences/semaphores
2513  */
2514
2515 void r600_fence_ring_emit(struct radeon_device *rdev,
2516                           struct radeon_fence *fence)
2517 {
2518         struct radeon_ring *ring = &rdev->ring[fence->ring];
2519
2520         if (rdev->wb.use_event) {
2521                 u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
2522                 /* flush read cache over gart */
2523                 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
2524                 radeon_ring_write(ring, PACKET3_TC_ACTION_ENA |
2525                                         PACKET3_VC_ACTION_ENA |
2526                                         PACKET3_SH_ACTION_ENA);
2527                 radeon_ring_write(ring, 0xFFFFFFFF);
2528                 radeon_ring_write(ring, 0);
2529                 radeon_ring_write(ring, 10); /* poll interval */
2530                 /* EVENT_WRITE_EOP - flush caches, send int */
2531                 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
2532                 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_EVENT_TS) | EVENT_INDEX(5));
2533                 radeon_ring_write(ring, addr & 0xffffffff);
2534                 radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
2535                 radeon_ring_write(ring, fence->seq);
2536                 radeon_ring_write(ring, 0);
2537         } else {
2538                 /* flush read cache over gart */
2539                 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
2540                 radeon_ring_write(ring, PACKET3_TC_ACTION_ENA |
2541                                         PACKET3_VC_ACTION_ENA |
2542                                         PACKET3_SH_ACTION_ENA);
2543                 radeon_ring_write(ring, 0xFFFFFFFF);
2544                 radeon_ring_write(ring, 0);
2545                 radeon_ring_write(ring, 10); /* poll interval */
2546                 radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE, 0));
2547                 radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0));
2548                 /* wait for 3D idle clean */
2549                 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
2550                 radeon_ring_write(ring, (WAIT_UNTIL - PACKET3_SET_CONFIG_REG_OFFSET) >> 2);
2551                 radeon_ring_write(ring, WAIT_3D_IDLE_bit | WAIT_3D_IDLECLEAN_bit);
2552                 /* Emit fence sequence & fire IRQ */
2553                 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
2554                 radeon_ring_write(ring, ((rdev->fence_drv[fence->ring].scratch_reg - PACKET3_SET_CONFIG_REG_OFFSET) >> 2));
2555                 radeon_ring_write(ring, fence->seq);
2556                 /* CP_INTERRUPT packet 3 no longer exists, use packet 0 */
2557                 radeon_ring_write(ring, PACKET0(CP_INT_STATUS, 0));
2558                 radeon_ring_write(ring, RB_INT_STAT);
2559         }
2560 }
2561
2562 void r600_semaphore_ring_emit(struct radeon_device *rdev,
2563                               struct radeon_ring *ring,
2564                               struct radeon_semaphore *semaphore,
2565                               bool emit_wait)
2566 {
2567         uint64_t addr = semaphore->gpu_addr;
2568         unsigned sel = emit_wait ? PACKET3_SEM_SEL_WAIT : PACKET3_SEM_SEL_SIGNAL;
2569
2570         if (rdev->family < CHIP_CAYMAN)
2571                 sel |= PACKET3_SEM_WAIT_ON_SIGNAL;
2572
2573         radeon_ring_write(ring, PACKET3(PACKET3_MEM_SEMAPHORE, 1));
2574         radeon_ring_write(ring, addr & 0xffffffff);
2575         radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | sel);
2576 }
2577
2578 /*
2579  * DMA fences/semaphores
2580  */
2581
2582 /**
2583  * r600_dma_fence_ring_emit - emit a fence on the DMA ring
2584  *
2585  * @rdev: radeon_device pointer
2586  * @fence: radeon fence object
2587  *
2588  * Add a DMA fence packet to the ring to write
2589  * the fence seq number and DMA trap packet to generate
2590  * an interrupt if needed (r6xx-r7xx).
2591  */
2592 void r600_dma_fence_ring_emit(struct radeon_device *rdev,
2593                               struct radeon_fence *fence)
2594 {
2595         struct radeon_ring *ring = &rdev->ring[fence->ring];
2596         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
2597
2598         /* write the fence */
2599         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_FENCE, 0, 0, 0));
2600         radeon_ring_write(ring, addr & 0xfffffffc);
2601         radeon_ring_write(ring, (upper_32_bits(addr) & 0xff));
2602         radeon_ring_write(ring, lower_32_bits(fence->seq));
2603         /* generate an interrupt */
2604         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_TRAP, 0, 0, 0));
2605 }
2606
2607 /**
2608  * r600_dma_semaphore_ring_emit - emit a semaphore on the dma ring
2609  *
2610  * @rdev: radeon_device pointer
2611  * @ring: radeon_ring structure holding ring information
2612  * @semaphore: radeon semaphore object
2613  * @emit_wait: wait or signal semaphore
2614  *
2615  * Add a DMA semaphore packet to the ring wait on or signal
2616  * other rings (r6xx-SI).
2617  */
2618 void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
2619                                   struct radeon_ring *ring,
2620                                   struct radeon_semaphore *semaphore,
2621                                   bool emit_wait)
2622 {
2623         u64 addr = semaphore->gpu_addr;
2624         u32 s = emit_wait ? 0 : 1;
2625
2626         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SEMAPHORE, 0, s, 0));
2627         radeon_ring_write(ring, addr & 0xfffffffc);
2628         radeon_ring_write(ring, upper_32_bits(addr) & 0xff);
2629 }
2630
2631 int r600_copy_blit(struct radeon_device *rdev,
2632                    uint64_t src_offset,
2633                    uint64_t dst_offset,
2634                    unsigned num_gpu_pages,
2635                    struct radeon_fence **fence)
2636 {
2637         struct radeon_semaphore *sem = NULL;
2638         struct radeon_sa_bo *vb = NULL;
2639         int r;
2640
2641         r = r600_blit_prepare_copy(rdev, num_gpu_pages, fence, &vb, &sem);
2642         if (r) {
2643                 return r;
2644         }
2645         r600_kms_blit_copy(rdev, src_offset, dst_offset, num_gpu_pages, vb);
2646         r600_blit_done_copy(rdev, fence, vb, sem);
2647         return 0;
2648 }
2649
2650 /**
2651  * r600_copy_dma - copy pages using the DMA engine
2652  *
2653  * @rdev: radeon_device pointer
2654  * @src_offset: src GPU address
2655  * @dst_offset: dst GPU address
2656  * @num_gpu_pages: number of GPU pages to xfer
2657  * @fence: radeon fence object
2658  *
2659  * Copy GPU paging using the DMA engine (r6xx).
2660  * Used by the radeon ttm implementation to move pages if
2661  * registered as the asic copy callback.
2662  */
2663 int r600_copy_dma(struct radeon_device *rdev,
2664                   uint64_t src_offset, uint64_t dst_offset,
2665                   unsigned num_gpu_pages,
2666                   struct radeon_fence **fence)
2667 {
2668         struct radeon_semaphore *sem = NULL;
2669         int ring_index = rdev->asic->copy.dma_ring_index;
2670         struct radeon_ring *ring = &rdev->ring[ring_index];
2671         u32 size_in_dw, cur_size_in_dw;
2672         int i, num_loops;
2673         int r = 0;
2674
2675         r = radeon_semaphore_create(rdev, &sem);
2676         if (r) {
2677                 DRM_ERROR("radeon: moving bo (%d).\n", r);
2678                 return r;
2679         }
2680
2681         size_in_dw = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT) / 4;
2682         num_loops = DIV_ROUND_UP(size_in_dw, 0xFFFE);
2683         r = radeon_ring_lock(rdev, ring, num_loops * 4 + 8);
2684         if (r) {
2685                 DRM_ERROR("radeon: moving bo (%d).\n", r);
2686                 radeon_semaphore_free(rdev, &sem, NULL);
2687                 return r;
2688         }
2689
2690         if (radeon_fence_need_sync(*fence, ring->idx)) {
2691                 radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
2692                                             ring->idx);
2693                 radeon_fence_note_sync(*fence, ring->idx);
2694         } else {
2695                 radeon_semaphore_free(rdev, &sem, NULL);
2696         }
2697
2698         for (i = 0; i < num_loops; i++) {
2699                 cur_size_in_dw = size_in_dw;
2700                 if (cur_size_in_dw > 0xFFFE)
2701                         cur_size_in_dw = 0xFFFE;
2702                 size_in_dw -= cur_size_in_dw;
2703                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 0, 0, cur_size_in_dw));
2704                 radeon_ring_write(ring, dst_offset & 0xfffffffc);
2705                 radeon_ring_write(ring, src_offset & 0xfffffffc);
2706                 radeon_ring_write(ring, (((upper_32_bits(dst_offset) & 0xff) << 16) |
2707                                          (upper_32_bits(src_offset) & 0xff)));
2708                 src_offset += cur_size_in_dw * 4;
2709                 dst_offset += cur_size_in_dw * 4;
2710         }
2711
2712         r = radeon_fence_emit(rdev, fence, ring->idx);
2713         if (r) {
2714                 radeon_ring_unlock_undo(rdev, ring);
2715                 return r;
2716         }
2717
2718         radeon_ring_unlock_commit(rdev, ring);
2719         radeon_semaphore_free(rdev, &sem, *fence);
2720
2721         return r;
2722 }
2723
2724 int r600_set_surface_reg(struct radeon_device *rdev, int reg,
2725                          uint32_t tiling_flags, uint32_t pitch,
2726                          uint32_t offset, uint32_t obj_size)
2727 {
2728         /* FIXME: implement */
2729         return 0;
2730 }
2731
2732 void r600_clear_surface_reg(struct radeon_device *rdev, int reg)
2733 {
2734         /* FIXME: implement */
2735 }
2736
2737 static int r600_startup(struct radeon_device *rdev)
2738 {
2739         struct radeon_ring *ring;
2740         int r;
2741
2742         /* enable pcie gen2 link */
2743         r600_pcie_gen2_enable(rdev);
2744
2745         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
2746                 r = r600_init_microcode(rdev);
2747                 if (r) {
2748                         DRM_ERROR("Failed to load firmware!\n");
2749                         return r;
2750                 }
2751         }
2752
2753         r = r600_vram_scratch_init(rdev);
2754         if (r)
2755                 return r;
2756
2757         r600_mc_program(rdev);
2758         if (rdev->flags & RADEON_IS_AGP) {
2759                 r600_agp_enable(rdev);
2760         } else {
2761                 r = r600_pcie_gart_enable(rdev);
2762                 if (r)
2763                         return r;
2764         }
2765         r600_gpu_init(rdev);
2766         r = r600_blit_init(rdev);
2767         if (r) {
2768                 r600_blit_fini(rdev);
2769                 rdev->asic->copy.copy = NULL;
2770                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
2771         }
2772
2773         /* allocate wb buffer */
2774         r = radeon_wb_init(rdev);
2775         if (r)
2776                 return r;
2777
2778         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
2779         if (r) {
2780                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
2781                 return r;
2782         }
2783
2784         r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
2785         if (r) {
2786                 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
2787                 return r;
2788         }
2789
2790         /* Enable IRQ */
2791         r = r600_irq_init(rdev);
2792         if (r) {
2793                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
2794                 radeon_irq_kms_fini(rdev);
2795                 return r;
2796         }
2797         r600_irq_set(rdev);
2798
2799         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
2800         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
2801                              R600_CP_RB_RPTR, R600_CP_RB_WPTR,
2802                              0, 0xfffff, RADEON_CP_PACKET2);
2803         if (r)
2804                 return r;
2805
2806         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
2807         r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
2808                              DMA_RB_RPTR, DMA_RB_WPTR,
2809                              2, 0x3fffc, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
2810         if (r)
2811                 return r;
2812
2813         r = r600_cp_load_microcode(rdev);
2814         if (r)
2815                 return r;
2816         r = r600_cp_resume(rdev);
2817         if (r)
2818                 return r;
2819
2820         r = r600_dma_resume(rdev);
2821         if (r)
2822                 return r;
2823
2824         r = radeon_ib_pool_init(rdev);
2825         if (r) {
2826                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
2827                 return r;
2828         }
2829
2830         r = r600_audio_init(rdev);
2831         if (r) {
2832                 DRM_ERROR("radeon: audio init failed\n");
2833                 return r;
2834         }
2835
2836         return 0;
2837 }
2838
2839 void r600_vga_set_state(struct radeon_device *rdev, bool state)
2840 {
2841         uint32_t temp;
2842
2843         temp = RREG32(CONFIG_CNTL);
2844         if (state == false) {
2845                 temp &= ~(1<<0);
2846                 temp |= (1<<1);
2847         } else {
2848                 temp &= ~(1<<1);
2849         }
2850         WREG32(CONFIG_CNTL, temp);
2851 }
2852
2853 int r600_resume(struct radeon_device *rdev)
2854 {
2855         int r;
2856
2857         /* Do not reset GPU before posting, on r600 hw unlike on r500 hw,
2858          * posting will perform necessary task to bring back GPU into good
2859          * shape.
2860          */
2861         /* post card */
2862         atom_asic_init(rdev->mode_info.atom_context);
2863
2864         rdev->accel_working = true;
2865         r = r600_startup(rdev);
2866         if (r) {
2867                 DRM_ERROR("r600 startup failed on resume\n");
2868                 rdev->accel_working = false;
2869                 return r;
2870         }
2871
2872         return r;
2873 }
2874
2875 int r600_suspend(struct radeon_device *rdev)
2876 {
2877         r600_audio_fini(rdev);
2878         r600_cp_stop(rdev);
2879         r600_dma_stop(rdev);
2880         r600_irq_suspend(rdev);
2881         radeon_wb_disable(rdev);
2882         r600_pcie_gart_disable(rdev);
2883
2884         return 0;
2885 }
2886
2887 /* Plan is to move initialization in that function and use
2888  * helper function so that radeon_device_init pretty much
2889  * do nothing more than calling asic specific function. This
2890  * should also allow to remove a bunch of callback function
2891  * like vram_info.
2892  */
2893 int r600_init(struct radeon_device *rdev)
2894 {
2895         int r;
2896
2897         if (r600_debugfs_mc_info_init(rdev)) {
2898                 DRM_ERROR("Failed to register debugfs file for mc !\n");
2899         }
2900         /* Read BIOS */
2901         if (!radeon_get_bios(rdev)) {
2902                 if (ASIC_IS_AVIVO(rdev))
2903                         return -EINVAL;
2904         }
2905         /* Must be an ATOMBIOS */
2906         if (!rdev->is_atom_bios) {
2907                 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
2908                 return -EINVAL;
2909         }
2910         r = radeon_atombios_init(rdev);
2911         if (r)
2912                 return r;
2913         /* Post card if necessary */
2914         if (!radeon_card_posted(rdev)) {
2915                 if (!rdev->bios) {
2916                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
2917                         return -EINVAL;
2918                 }
2919                 DRM_INFO("GPU not posted. posting now...\n");
2920                 atom_asic_init(rdev->mode_info.atom_context);
2921         }
2922         /* Initialize scratch registers */
2923         r600_scratch_init(rdev);
2924         /* Initialize surface registers */
2925         radeon_surface_init(rdev);
2926         /* Initialize clocks */
2927         radeon_get_clock_info(rdev->ddev);
2928         /* Fence driver */
2929         r = radeon_fence_driver_init(rdev);
2930         if (r)
2931                 return r;
2932         if (rdev->flags & RADEON_IS_AGP) {
2933                 r = radeon_agp_init(rdev);
2934                 if (r)
2935                         radeon_agp_disable(rdev);
2936         }
2937         r = r600_mc_init(rdev);
2938         if (r)
2939                 return r;
2940         /* Memory manager */
2941         r = radeon_bo_init(rdev);
2942         if (r)
2943                 return r;
2944
2945         r = radeon_irq_kms_init(rdev);
2946         if (r)
2947                 return r;
2948
2949         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ring_obj = NULL;
2950         r600_ring_init(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX], 1024 * 1024);
2951
2952         rdev->ring[R600_RING_TYPE_DMA_INDEX].ring_obj = NULL;
2953         r600_ring_init(rdev, &rdev->ring[R600_RING_TYPE_DMA_INDEX], 64 * 1024);
2954
2955         rdev->ih.ring_obj = NULL;
2956         r600_ih_ring_init(rdev, 64 * 1024);
2957
2958         r = r600_pcie_gart_init(rdev);
2959         if (r)
2960                 return r;
2961
2962         rdev->accel_working = true;
2963         r = r600_startup(rdev);
2964         if (r) {
2965                 dev_err(rdev->dev, "disabling GPU acceleration\n");
2966                 r600_cp_fini(rdev);
2967                 r600_dma_fini(rdev);
2968                 r600_irq_fini(rdev);
2969                 radeon_wb_fini(rdev);
2970                 radeon_ib_pool_fini(rdev);
2971                 radeon_irq_kms_fini(rdev);
2972                 r600_pcie_gart_fini(rdev);
2973                 rdev->accel_working = false;
2974         }
2975
2976         return 0;
2977 }
2978
2979 void r600_fini(struct radeon_device *rdev)
2980 {
2981         r600_audio_fini(rdev);
2982         r600_blit_fini(rdev);
2983         r600_cp_fini(rdev);
2984         r600_dma_fini(rdev);
2985         r600_irq_fini(rdev);
2986         radeon_wb_fini(rdev);
2987         radeon_ib_pool_fini(rdev);
2988         radeon_irq_kms_fini(rdev);
2989         r600_pcie_gart_fini(rdev);
2990         r600_vram_scratch_fini(rdev);
2991         radeon_agp_fini(rdev);
2992         radeon_gem_fini(rdev);
2993         radeon_fence_driver_fini(rdev);
2994         radeon_bo_fini(rdev);
2995         radeon_atombios_fini(rdev);
2996         kfree(rdev->bios);
2997         rdev->bios = NULL;
2998 }
2999
3000
3001 /*
3002  * CS stuff
3003  */
3004 void r600_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3005 {
3006         struct radeon_ring *ring = &rdev->ring[ib->ring];
3007         u32 next_rptr;
3008
3009         if (ring->rptr_save_reg) {
3010                 next_rptr = ring->wptr + 3 + 4;
3011                 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
3012                 radeon_ring_write(ring, ((ring->rptr_save_reg -
3013                                          PACKET3_SET_CONFIG_REG_OFFSET) >> 2));
3014                 radeon_ring_write(ring, next_rptr);
3015         } else if (rdev->wb.enabled) {
3016                 next_rptr = ring->wptr + 5 + 4;
3017                 radeon_ring_write(ring, PACKET3(PACKET3_MEM_WRITE, 3));
3018                 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
3019                 radeon_ring_write(ring, (upper_32_bits(ring->next_rptr_gpu_addr) & 0xff) | (1 << 18));
3020                 radeon_ring_write(ring, next_rptr);
3021                 radeon_ring_write(ring, 0);
3022         }
3023
3024         radeon_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
3025         radeon_ring_write(ring,
3026 #ifdef __BIG_ENDIAN
3027                           (2 << 0) |
3028 #endif
3029                           (ib->gpu_addr & 0xFFFFFFFC));
3030         radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFF);
3031         radeon_ring_write(ring, ib->length_dw);
3032 }
3033
3034 int r600_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
3035 {
3036         struct radeon_ib ib;
3037         uint32_t scratch;
3038         uint32_t tmp = 0;
3039         unsigned i;
3040         int r;
3041
3042         r = radeon_scratch_get(rdev, &scratch);
3043         if (r) {
3044                 DRM_ERROR("radeon: failed to get scratch reg (%d).\n", r);
3045                 return r;
3046         }
3047         WREG32(scratch, 0xCAFEDEAD);
3048         r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
3049         if (r) {
3050                 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
3051                 goto free_scratch;
3052         }
3053         ib.ptr[0] = PACKET3(PACKET3_SET_CONFIG_REG, 1);
3054         ib.ptr[1] = ((scratch - PACKET3_SET_CONFIG_REG_OFFSET) >> 2);
3055         ib.ptr[2] = 0xDEADBEEF;
3056         ib.length_dw = 3;
3057         r = radeon_ib_schedule(rdev, &ib, NULL);
3058         if (r) {
3059                 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
3060                 goto free_ib;
3061         }
3062         r = radeon_fence_wait(ib.fence, false);
3063         if (r) {
3064                 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
3065                 goto free_ib;
3066         }
3067         for (i = 0; i < rdev->usec_timeout; i++) {
3068                 tmp = RREG32(scratch);
3069                 if (tmp == 0xDEADBEEF)
3070                         break;
3071                 DRM_UDELAY(1);
3072         }
3073         if (i < rdev->usec_timeout) {
3074                 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
3075         } else {
3076                 DRM_ERROR("radeon: ib test failed (scratch(0x%04X)=0x%08X)\n",
3077                           scratch, tmp);
3078                 r = -EINVAL;
3079         }
3080 free_ib:
3081         radeon_ib_free(rdev, &ib);
3082 free_scratch:
3083         radeon_scratch_free(rdev, scratch);
3084         return r;
3085 }
3086
3087 /**
3088  * r600_dma_ib_test - test an IB on the DMA engine
3089  *
3090  * @rdev: radeon_device pointer
3091  * @ring: radeon_ring structure holding ring information
3092  *
3093  * Test a simple IB in the DMA ring (r6xx-SI).
3094  * Returns 0 on success, error on failure.
3095  */
3096 int r600_dma_ib_test(struct radeon_device *rdev, struct radeon_ring *ring)
3097 {
3098         struct radeon_ib ib;
3099         unsigned i;
3100         int r;
3101         void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
3102         u32 tmp = 0;
3103
3104         if (!ptr) {
3105                 DRM_ERROR("invalid vram scratch pointer\n");
3106                 return -EINVAL;
3107         }
3108
3109         tmp = 0xCAFEDEAD;
3110         writel(tmp, ptr);
3111
3112         r = radeon_ib_get(rdev, ring->idx, &ib, NULL, 256);
3113         if (r) {
3114                 DRM_ERROR("radeon: failed to get ib (%d).\n", r);
3115                 return r;
3116         }
3117
3118         ib.ptr[0] = DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1);
3119         ib.ptr[1] = rdev->vram_scratch.gpu_addr & 0xfffffffc;
3120         ib.ptr[2] = upper_32_bits(rdev->vram_scratch.gpu_addr) & 0xff;
3121         ib.ptr[3] = 0xDEADBEEF;
3122         ib.length_dw = 4;
3123
3124         r = radeon_ib_schedule(rdev, &ib, NULL);
3125         if (r) {
3126                 radeon_ib_free(rdev, &ib);
3127                 DRM_ERROR("radeon: failed to schedule ib (%d).\n", r);
3128                 return r;
3129         }
3130         r = radeon_fence_wait(ib.fence, false);
3131         if (r) {
3132                 DRM_ERROR("radeon: fence wait failed (%d).\n", r);
3133                 return r;
3134         }
3135         for (i = 0; i < rdev->usec_timeout; i++) {
3136                 tmp = readl(ptr);
3137                 if (tmp == 0xDEADBEEF)
3138                         break;
3139                 DRM_UDELAY(1);
3140         }
3141         if (i < rdev->usec_timeout) {
3142                 DRM_INFO("ib test on ring %d succeeded in %u usecs\n", ib.fence->ring, i);
3143         } else {
3144                 DRM_ERROR("radeon: ib test failed (0x%08X)\n", tmp);
3145                 r = -EINVAL;
3146         }
3147         radeon_ib_free(rdev, &ib);
3148         return r;
3149 }
3150
3151 /**
3152  * r600_dma_ring_ib_execute - Schedule an IB on the DMA engine
3153  *
3154  * @rdev: radeon_device pointer
3155  * @ib: IB object to schedule
3156  *
3157  * Schedule an IB in the DMA ring (r6xx-r7xx).
3158  */
3159 void r600_dma_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
3160 {
3161         struct radeon_ring *ring = &rdev->ring[ib->ring];
3162
3163         if (rdev->wb.enabled) {
3164                 u32 next_rptr = ring->wptr + 4;
3165                 while ((next_rptr & 7) != 5)
3166                         next_rptr++;
3167                 next_rptr += 3;
3168                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 1));
3169                 radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
3170                 radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xff);
3171                 radeon_ring_write(ring, next_rptr);
3172         }
3173
3174         /* The indirect buffer packet must end on an 8 DW boundary in the DMA ring.
3175          * Pad as necessary with NOPs.
3176          */
3177         while ((ring->wptr & 7) != 5)
3178                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0));
3179         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_INDIRECT_BUFFER, 0, 0, 0));
3180         radeon_ring_write(ring, (ib->gpu_addr & 0xFFFFFFE0));
3181         radeon_ring_write(ring, (ib->length_dw << 16) | (upper_32_bits(ib->gpu_addr) & 0xFF));
3182
3183 }
3184
3185 /*
3186  * Interrupts
3187  *
3188  * Interrupts use a ring buffer on r6xx/r7xx hardware.  It works pretty
3189  * the same as the CP ring buffer, but in reverse.  Rather than the CPU
3190  * writing to the ring and the GPU consuming, the GPU writes to the ring
3191  * and host consumes.  As the host irq handler processes interrupts, it
3192  * increments the rptr.  When the rptr catches up with the wptr, all the
3193  * current interrupts have been processed.
3194  */
3195
3196 void r600_ih_ring_init(struct radeon_device *rdev, unsigned ring_size)
3197 {
3198         u32 rb_bufsz;
3199
3200         /* Align ring size */
3201         rb_bufsz = drm_order(ring_size / 4);
3202         ring_size = (1 << rb_bufsz) * 4;
3203         rdev->ih.ring_size = ring_size;
3204         rdev->ih.ptr_mask = rdev->ih.ring_size - 1;
3205         rdev->ih.rptr = 0;
3206 }
3207
3208 int r600_ih_ring_alloc(struct radeon_device *rdev)
3209 {
3210         int r;
3211
3212         /* Allocate ring buffer */
3213         if (rdev->ih.ring_obj == NULL) {
3214                 r = radeon_bo_create(rdev, rdev->ih.ring_size,
3215                                      PAGE_SIZE, true,
3216                                      RADEON_GEM_DOMAIN_GTT,
3217                                      NULL, &rdev->ih.ring_obj);
3218                 if (r) {
3219                         DRM_ERROR("radeon: failed to create ih ring buffer (%d).\n", r);
3220                         return r;
3221                 }
3222                 r = radeon_bo_reserve(rdev->ih.ring_obj, false);
3223                 if (unlikely(r != 0))
3224                         return r;
3225                 r = radeon_bo_pin(rdev->ih.ring_obj,
3226                                   RADEON_GEM_DOMAIN_GTT,
3227                                   &rdev->ih.gpu_addr);
3228                 if (r) {
3229                         radeon_bo_unreserve(rdev->ih.ring_obj);
3230                         DRM_ERROR("radeon: failed to pin ih ring buffer (%d).\n", r);
3231                         return r;
3232                 }
3233                 r = radeon_bo_kmap(rdev->ih.ring_obj,
3234                                    (void **)&rdev->ih.ring);
3235                 radeon_bo_unreserve(rdev->ih.ring_obj);
3236                 if (r) {
3237                         DRM_ERROR("radeon: failed to map ih ring buffer (%d).\n", r);
3238                         return r;
3239                 }
3240         }
3241         return 0;
3242 }
3243
3244 void r600_ih_ring_fini(struct radeon_device *rdev)
3245 {
3246         int r;
3247         if (rdev->ih.ring_obj) {
3248                 r = radeon_bo_reserve(rdev->ih.ring_obj, false);
3249                 if (likely(r == 0)) {
3250                         radeon_bo_kunmap(rdev->ih.ring_obj);
3251                         radeon_bo_unpin(rdev->ih.ring_obj);
3252                         radeon_bo_unreserve(rdev->ih.ring_obj);
3253                 }
3254                 radeon_bo_unref(&rdev->ih.ring_obj);
3255                 rdev->ih.ring = NULL;
3256                 rdev->ih.ring_obj = NULL;
3257         }
3258 }
3259
3260 void r600_rlc_stop(struct radeon_device *rdev)
3261 {
3262
3263         if ((rdev->family >= CHIP_RV770) &&
3264             (rdev->family <= CHIP_RV740)) {
3265                 /* r7xx asics need to soft reset RLC before halting */
3266                 WREG32(SRBM_SOFT_RESET, SOFT_RESET_RLC);
3267                 RREG32(SRBM_SOFT_RESET);
3268                 mdelay(15);
3269                 WREG32(SRBM_SOFT_RESET, 0);
3270                 RREG32(SRBM_SOFT_RESET);
3271         }
3272
3273         WREG32(RLC_CNTL, 0);
3274 }
3275
3276 static void r600_rlc_start(struct radeon_device *rdev)
3277 {
3278         WREG32(RLC_CNTL, RLC_ENABLE);
3279 }
3280
3281 static int r600_rlc_init(struct radeon_device *rdev)
3282 {
3283         u32 i;
3284         const __be32 *fw_data;
3285
3286         if (!rdev->rlc_fw)
3287                 return -EINVAL;
3288
3289         r600_rlc_stop(rdev);
3290
3291         WREG32(RLC_HB_CNTL, 0);
3292
3293         if (rdev->family == CHIP_ARUBA) {
3294                 WREG32(TN_RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
3295                 WREG32(TN_RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
3296         }
3297         if (rdev->family <= CHIP_CAYMAN) {
3298                 WREG32(RLC_HB_BASE, 0);
3299                 WREG32(RLC_HB_RPTR, 0);
3300                 WREG32(RLC_HB_WPTR, 0);
3301         }
3302         if (rdev->family <= CHIP_CAICOS) {
3303                 WREG32(RLC_HB_WPTR_LSB_ADDR, 0);
3304                 WREG32(RLC_HB_WPTR_MSB_ADDR, 0);
3305         }
3306         WREG32(RLC_MC_CNTL, 0);
3307         WREG32(RLC_UCODE_CNTL, 0);
3308
3309         fw_data = (const __be32 *)rdev->rlc_fw->data;
3310         if (rdev->family >= CHIP_ARUBA) {
3311                 for (i = 0; i < ARUBA_RLC_UCODE_SIZE; i++) {
3312                         WREG32(RLC_UCODE_ADDR, i);
3313                         WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3314                 }
3315         } else if (rdev->family >= CHIP_CAYMAN) {
3316                 for (i = 0; i < CAYMAN_RLC_UCODE_SIZE; i++) {
3317                         WREG32(RLC_UCODE_ADDR, i);
3318                         WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3319                 }
3320         } else if (rdev->family >= CHIP_CEDAR) {
3321                 for (i = 0; i < EVERGREEN_RLC_UCODE_SIZE; i++) {
3322                         WREG32(RLC_UCODE_ADDR, i);
3323                         WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3324                 }
3325         } else if (rdev->family >= CHIP_RV770) {
3326                 for (i = 0; i < R700_RLC_UCODE_SIZE; i++) {
3327                         WREG32(RLC_UCODE_ADDR, i);
3328                         WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3329                 }
3330         } else {
3331                 for (i = 0; i < RLC_UCODE_SIZE; i++) {
3332                         WREG32(RLC_UCODE_ADDR, i);
3333                         WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3334                 }
3335         }
3336         WREG32(RLC_UCODE_ADDR, 0);
3337
3338         r600_rlc_start(rdev);
3339
3340         return 0;
3341 }
3342
3343 static void r600_enable_interrupts(struct radeon_device *rdev)
3344 {
3345         u32 ih_cntl = RREG32(IH_CNTL);
3346         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
3347
3348         ih_cntl |= ENABLE_INTR;
3349         ih_rb_cntl |= IH_RB_ENABLE;
3350         WREG32(IH_CNTL, ih_cntl);
3351         WREG32(IH_RB_CNTL, ih_rb_cntl);
3352         rdev->ih.enabled = true;
3353 }
3354
3355 void r600_disable_interrupts(struct radeon_device *rdev)
3356 {
3357         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
3358         u32 ih_cntl = RREG32(IH_CNTL);
3359
3360         ih_rb_cntl &= ~IH_RB_ENABLE;
3361         ih_cntl &= ~ENABLE_INTR;
3362         WREG32(IH_RB_CNTL, ih_rb_cntl);
3363         WREG32(IH_CNTL, ih_cntl);
3364         /* set rptr, wptr to 0 */
3365         WREG32(IH_RB_RPTR, 0);
3366         WREG32(IH_RB_WPTR, 0);
3367         rdev->ih.enabled = false;
3368         rdev->ih.rptr = 0;
3369 }
3370
3371 static void r600_disable_interrupt_state(struct radeon_device *rdev)
3372 {
3373         u32 tmp;
3374
3375         WREG32(CP_INT_CNTL, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
3376         tmp = RREG32(DMA_CNTL) & ~TRAP_ENABLE;
3377         WREG32(DMA_CNTL, tmp);
3378         WREG32(GRBM_INT_CNTL, 0);
3379         WREG32(DxMODE_INT_MASK, 0);
3380         WREG32(D1GRPH_INTERRUPT_CONTROL, 0);
3381         WREG32(D2GRPH_INTERRUPT_CONTROL, 0);
3382         if (ASIC_IS_DCE3(rdev)) {
3383                 WREG32(DCE3_DACA_AUTODETECT_INT_CONTROL, 0);
3384                 WREG32(DCE3_DACB_AUTODETECT_INT_CONTROL, 0);
3385                 tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3386                 WREG32(DC_HPD1_INT_CONTROL, tmp);
3387                 tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3388                 WREG32(DC_HPD2_INT_CONTROL, tmp);
3389                 tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3390                 WREG32(DC_HPD3_INT_CONTROL, tmp);
3391                 tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3392                 WREG32(DC_HPD4_INT_CONTROL, tmp);
3393                 if (ASIC_IS_DCE32(rdev)) {
3394                         tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3395                         WREG32(DC_HPD5_INT_CONTROL, tmp);
3396                         tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3397                         WREG32(DC_HPD6_INT_CONTROL, tmp);
3398                         tmp = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3399                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0, tmp);
3400                         tmp = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3401                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1, tmp);
3402                 } else {
3403                         tmp = RREG32(HDMI0_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3404                         WREG32(HDMI0_AUDIO_PACKET_CONTROL, tmp);
3405                         tmp = RREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3406                         WREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL, tmp);
3407                 }
3408         } else {
3409                 WREG32(DACA_AUTODETECT_INT_CONTROL, 0);
3410                 WREG32(DACB_AUTODETECT_INT_CONTROL, 0);
3411                 tmp = RREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL) & DC_HOT_PLUG_DETECTx_INT_POLARITY;
3412                 WREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL, tmp);
3413                 tmp = RREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL) & DC_HOT_PLUG_DETECTx_INT_POLARITY;
3414                 WREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL, tmp);
3415                 tmp = RREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL) & DC_HOT_PLUG_DETECTx_INT_POLARITY;
3416                 WREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL, tmp);
3417                 tmp = RREG32(HDMI0_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3418                 WREG32(HDMI0_AUDIO_PACKET_CONTROL, tmp);
3419                 tmp = RREG32(HDMI1_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3420                 WREG32(HDMI1_AUDIO_PACKET_CONTROL, tmp);
3421         }
3422 }
3423
3424 int r600_irq_init(struct radeon_device *rdev)
3425 {
3426         int ret = 0;
3427         int rb_bufsz;
3428         u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
3429
3430         /* allocate ring */
3431         ret = r600_ih_ring_alloc(rdev);
3432         if (ret)
3433                 return ret;
3434
3435         /* disable irqs */
3436         r600_disable_interrupts(rdev);
3437
3438         /* init rlc */
3439         ret = r600_rlc_init(rdev);
3440         if (ret) {
3441                 r600_ih_ring_fini(rdev);
3442                 return ret;
3443         }
3444
3445         /* setup interrupt control */
3446         /* set dummy read address to ring address */
3447         WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
3448         interrupt_cntl = RREG32(INTERRUPT_CNTL);
3449         /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
3450          * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
3451          */
3452         interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
3453         /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
3454         interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
3455         WREG32(INTERRUPT_CNTL, interrupt_cntl);
3456
3457         WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
3458         rb_bufsz = drm_order(rdev->ih.ring_size / 4);
3459
3460         ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
3461                       IH_WPTR_OVERFLOW_CLEAR |
3462                       (rb_bufsz << 1));
3463
3464         if (rdev->wb.enabled)
3465                 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
3466
3467         /* set the writeback address whether it's enabled or not */
3468         WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
3469         WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
3470
3471         WREG32(IH_RB_CNTL, ih_rb_cntl);
3472
3473         /* set rptr, wptr to 0 */
3474         WREG32(IH_RB_RPTR, 0);
3475         WREG32(IH_RB_WPTR, 0);
3476
3477         /* Default settings for IH_CNTL (disabled at first) */
3478         ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10);
3479         /* RPTR_REARM only works if msi's are enabled */
3480         if (rdev->msi_enabled)
3481                 ih_cntl |= RPTR_REARM;
3482         WREG32(IH_CNTL, ih_cntl);
3483
3484         /* force the active interrupt state to all disabled */
3485         if (rdev->family >= CHIP_CEDAR)
3486                 evergreen_disable_interrupt_state(rdev);
3487         else
3488                 r600_disable_interrupt_state(rdev);
3489
3490         /* at this point everything should be setup correctly to enable master */
3491         pci_set_master(rdev->pdev);
3492
3493         /* enable irqs */
3494         r600_enable_interrupts(rdev);
3495
3496         return ret;
3497 }
3498
3499 void r600_irq_suspend(struct radeon_device *rdev)
3500 {
3501         r600_irq_disable(rdev);
3502         r600_rlc_stop(rdev);
3503 }
3504
3505 void r600_irq_fini(struct radeon_device *rdev)
3506 {
3507         r600_irq_suspend(rdev);
3508         r600_ih_ring_fini(rdev);
3509 }
3510
3511 int r600_irq_set(struct radeon_device *rdev)
3512 {
3513         u32 cp_int_cntl = CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE;
3514         u32 mode_int = 0;
3515         u32 hpd1, hpd2, hpd3, hpd4 = 0, hpd5 = 0, hpd6 = 0;
3516         u32 grbm_int_cntl = 0;
3517         u32 hdmi0, hdmi1;
3518         u32 d1grph = 0, d2grph = 0;
3519         u32 dma_cntl;
3520
3521         if (!rdev->irq.installed) {
3522                 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
3523                 return -EINVAL;
3524         }
3525         /* don't enable anything if the ih is disabled */
3526         if (!rdev->ih.enabled) {
3527                 r600_disable_interrupts(rdev);
3528                 /* force the active interrupt state to all disabled */
3529                 r600_disable_interrupt_state(rdev);
3530                 return 0;
3531         }
3532
3533         if (ASIC_IS_DCE3(rdev)) {
3534                 hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
3535                 hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
3536                 hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
3537                 hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
3538                 if (ASIC_IS_DCE32(rdev)) {
3539                         hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
3540                         hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
3541                         hdmi0 = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0) & ~AFMT_AZ_FORMAT_WTRIG_MASK;
3542                         hdmi1 = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1) & ~AFMT_AZ_FORMAT_WTRIG_MASK;
3543                 } else {
3544                         hdmi0 = RREG32(HDMI0_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3545                         hdmi1 = RREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3546                 }
3547         } else {
3548                 hpd1 = RREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL) & ~DC_HPDx_INT_EN;
3549                 hpd2 = RREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL) & ~DC_HPDx_INT_EN;
3550                 hpd3 = RREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL) & ~DC_HPDx_INT_EN;
3551                 hdmi0 = RREG32(HDMI0_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3552                 hdmi1 = RREG32(HDMI1_AUDIO_PACKET_CONTROL) & ~HDMI0_AZ_FORMAT_WTRIG_MASK;
3553         }
3554         dma_cntl = RREG32(DMA_CNTL) & ~TRAP_ENABLE;
3555
3556         if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
3557                 DRM_DEBUG("r600_irq_set: sw int\n");
3558                 cp_int_cntl |= RB_INT_ENABLE;
3559                 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
3560         }
3561
3562         if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
3563                 DRM_DEBUG("r600_irq_set: sw int dma\n");
3564                 dma_cntl |= TRAP_ENABLE;
3565         }
3566
3567         if (rdev->irq.crtc_vblank_int[0] ||
3568             atomic_read(&rdev->irq.pflip[0])) {
3569                 DRM_DEBUG("r600_irq_set: vblank 0\n");
3570                 mode_int |= D1MODE_VBLANK_INT_MASK;
3571         }
3572         if (rdev->irq.crtc_vblank_int[1] ||
3573             atomic_read(&rdev->irq.pflip[1])) {
3574                 DRM_DEBUG("r600_irq_set: vblank 1\n");
3575                 mode_int |= D2MODE_VBLANK_INT_MASK;
3576         }
3577         if (rdev->irq.hpd[0]) {
3578                 DRM_DEBUG("r600_irq_set: hpd 1\n");
3579                 hpd1 |= DC_HPDx_INT_EN;
3580         }
3581         if (rdev->irq.hpd[1]) {
3582                 DRM_DEBUG("r600_irq_set: hpd 2\n");
3583                 hpd2 |= DC_HPDx_INT_EN;
3584         }
3585         if (rdev->irq.hpd[2]) {
3586                 DRM_DEBUG("r600_irq_set: hpd 3\n");
3587                 hpd3 |= DC_HPDx_INT_EN;
3588         }
3589         if (rdev->irq.hpd[3]) {
3590                 DRM_DEBUG("r600_irq_set: hpd 4\n");
3591                 hpd4 |= DC_HPDx_INT_EN;
3592         }
3593         if (rdev->irq.hpd[4]) {
3594                 DRM_DEBUG("r600_irq_set: hpd 5\n");
3595                 hpd5 |= DC_HPDx_INT_EN;
3596         }
3597         if (rdev->irq.hpd[5]) {
3598                 DRM_DEBUG("r600_irq_set: hpd 6\n");
3599                 hpd6 |= DC_HPDx_INT_EN;
3600         }
3601         if (rdev->irq.afmt[0]) {
3602                 DRM_DEBUG("r600_irq_set: hdmi 0\n");
3603                 hdmi0 |= HDMI0_AZ_FORMAT_WTRIG_MASK;
3604         }
3605         if (rdev->irq.afmt[1]) {
3606                 DRM_DEBUG("r600_irq_set: hdmi 0\n");
3607                 hdmi1 |= HDMI0_AZ_FORMAT_WTRIG_MASK;
3608         }
3609
3610         WREG32(CP_INT_CNTL, cp_int_cntl);
3611         WREG32(DMA_CNTL, dma_cntl);
3612         WREG32(DxMODE_INT_MASK, mode_int);
3613         WREG32(D1GRPH_INTERRUPT_CONTROL, d1grph);
3614         WREG32(D2GRPH_INTERRUPT_CONTROL, d2grph);
3615         WREG32(GRBM_INT_CNTL, grbm_int_cntl);
3616         if (ASIC_IS_DCE3(rdev)) {
3617                 WREG32(DC_HPD1_INT_CONTROL, hpd1);
3618                 WREG32(DC_HPD2_INT_CONTROL, hpd2);
3619                 WREG32(DC_HPD3_INT_CONTROL, hpd3);
3620                 WREG32(DC_HPD4_INT_CONTROL, hpd4);
3621                 if (ASIC_IS_DCE32(rdev)) {
3622                         WREG32(DC_HPD5_INT_CONTROL, hpd5);
3623                         WREG32(DC_HPD6_INT_CONTROL, hpd6);
3624                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0, hdmi0);
3625                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1, hdmi1);
3626                 } else {
3627                         WREG32(HDMI0_AUDIO_PACKET_CONTROL, hdmi0);
3628                         WREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL, hdmi1);
3629                 }
3630         } else {
3631                 WREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL, hpd1);
3632                 WREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL, hpd2);
3633                 WREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL, hpd3);
3634                 WREG32(HDMI0_AUDIO_PACKET_CONTROL, hdmi0);
3635                 WREG32(HDMI1_AUDIO_PACKET_CONTROL, hdmi1);
3636         }
3637
3638         return 0;
3639 }
3640
3641 static void r600_irq_ack(struct radeon_device *rdev)
3642 {
3643         u32 tmp;
3644
3645         if (ASIC_IS_DCE3(rdev)) {
3646                 rdev->irq.stat_regs.r600.disp_int = RREG32(DCE3_DISP_INTERRUPT_STATUS);
3647                 rdev->irq.stat_regs.r600.disp_int_cont = RREG32(DCE3_DISP_INTERRUPT_STATUS_CONTINUE);
3648                 rdev->irq.stat_regs.r600.disp_int_cont2 = RREG32(DCE3_DISP_INTERRUPT_STATUS_CONTINUE2);
3649                 if (ASIC_IS_DCE32(rdev)) {
3650                         rdev->irq.stat_regs.r600.hdmi0_status = RREG32(AFMT_STATUS + DCE3_HDMI_OFFSET0);
3651                         rdev->irq.stat_regs.r600.hdmi1_status = RREG32(AFMT_STATUS + DCE3_HDMI_OFFSET1);
3652                 } else {
3653                         rdev->irq.stat_regs.r600.hdmi0_status = RREG32(HDMI0_STATUS);
3654                         rdev->irq.stat_regs.r600.hdmi1_status = RREG32(DCE3_HDMI1_STATUS);
3655                 }
3656         } else {
3657                 rdev->irq.stat_regs.r600.disp_int = RREG32(DISP_INTERRUPT_STATUS);
3658                 rdev->irq.stat_regs.r600.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
3659                 rdev->irq.stat_regs.r600.disp_int_cont2 = 0;
3660                 rdev->irq.stat_regs.r600.hdmi0_status = RREG32(HDMI0_STATUS);
3661                 rdev->irq.stat_regs.r600.hdmi1_status = RREG32(HDMI1_STATUS);
3662         }
3663         rdev->irq.stat_regs.r600.d1grph_int = RREG32(D1GRPH_INTERRUPT_STATUS);
3664         rdev->irq.stat_regs.r600.d2grph_int = RREG32(D2GRPH_INTERRUPT_STATUS);
3665
3666         if (rdev->irq.stat_regs.r600.d1grph_int & DxGRPH_PFLIP_INT_OCCURRED)
3667                 WREG32(D1GRPH_INTERRUPT_STATUS, DxGRPH_PFLIP_INT_CLEAR);
3668         if (rdev->irq.stat_regs.r600.d2grph_int & DxGRPH_PFLIP_INT_OCCURRED)
3669                 WREG32(D2GRPH_INTERRUPT_STATUS, DxGRPH_PFLIP_INT_CLEAR);
3670         if (rdev->irq.stat_regs.r600.disp_int & LB_D1_VBLANK_INTERRUPT)
3671                 WREG32(D1MODE_VBLANK_STATUS, DxMODE_VBLANK_ACK);
3672         if (rdev->irq.stat_regs.r600.disp_int & LB_D1_VLINE_INTERRUPT)
3673                 WREG32(D1MODE_VLINE_STATUS, DxMODE_VLINE_ACK);
3674         if (rdev->irq.stat_regs.r600.disp_int & LB_D2_VBLANK_INTERRUPT)
3675                 WREG32(D2MODE_VBLANK_STATUS, DxMODE_VBLANK_ACK);
3676         if (rdev->irq.stat_regs.r600.disp_int & LB_D2_VLINE_INTERRUPT)
3677                 WREG32(D2MODE_VLINE_STATUS, DxMODE_VLINE_ACK);
3678         if (rdev->irq.stat_regs.r600.disp_int & DC_HPD1_INTERRUPT) {
3679                 if (ASIC_IS_DCE3(rdev)) {
3680                         tmp = RREG32(DC_HPD1_INT_CONTROL);
3681                         tmp |= DC_HPDx_INT_ACK;
3682                         WREG32(DC_HPD1_INT_CONTROL, tmp);
3683                 } else {
3684                         tmp = RREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL);
3685                         tmp |= DC_HPDx_INT_ACK;
3686                         WREG32(DC_HOT_PLUG_DETECT1_INT_CONTROL, tmp);
3687                 }
3688         }
3689         if (rdev->irq.stat_regs.r600.disp_int & DC_HPD2_INTERRUPT) {
3690                 if (ASIC_IS_DCE3(rdev)) {
3691                         tmp = RREG32(DC_HPD2_INT_CONTROL);
3692                         tmp |= DC_HPDx_INT_ACK;
3693                         WREG32(DC_HPD2_INT_CONTROL, tmp);
3694                 } else {
3695                         tmp = RREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL);
3696                         tmp |= DC_HPDx_INT_ACK;
3697                         WREG32(DC_HOT_PLUG_DETECT2_INT_CONTROL, tmp);
3698                 }
3699         }
3700         if (rdev->irq.stat_regs.r600.disp_int_cont & DC_HPD3_INTERRUPT) {
3701                 if (ASIC_IS_DCE3(rdev)) {
3702                         tmp = RREG32(DC_HPD3_INT_CONTROL);
3703                         tmp |= DC_HPDx_INT_ACK;
3704                         WREG32(DC_HPD3_INT_CONTROL, tmp);
3705                 } else {
3706                         tmp = RREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL);
3707                         tmp |= DC_HPDx_INT_ACK;
3708                         WREG32(DC_HOT_PLUG_DETECT3_INT_CONTROL, tmp);
3709                 }
3710         }
3711         if (rdev->irq.stat_regs.r600.disp_int_cont & DC_HPD4_INTERRUPT) {
3712                 tmp = RREG32(DC_HPD4_INT_CONTROL);
3713                 tmp |= DC_HPDx_INT_ACK;
3714                 WREG32(DC_HPD4_INT_CONTROL, tmp);
3715         }
3716         if (ASIC_IS_DCE32(rdev)) {
3717                 if (rdev->irq.stat_regs.r600.disp_int_cont2 & DC_HPD5_INTERRUPT) {
3718                         tmp = RREG32(DC_HPD5_INT_CONTROL);
3719                         tmp |= DC_HPDx_INT_ACK;
3720                         WREG32(DC_HPD5_INT_CONTROL, tmp);
3721                 }
3722                 if (rdev->irq.stat_regs.r600.disp_int_cont2 & DC_HPD6_INTERRUPT) {
3723                         tmp = RREG32(DC_HPD5_INT_CONTROL);
3724                         tmp |= DC_HPDx_INT_ACK;
3725                         WREG32(DC_HPD6_INT_CONTROL, tmp);
3726                 }
3727                 if (rdev->irq.stat_regs.r600.hdmi0_status & AFMT_AZ_FORMAT_WTRIG) {
3728                         tmp = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0);
3729                         tmp |= AFMT_AZ_FORMAT_WTRIG_ACK;
3730                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET0, tmp);
3731                 }
3732                 if (rdev->irq.stat_regs.r600.hdmi1_status & AFMT_AZ_FORMAT_WTRIG) {
3733                         tmp = RREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1);
3734                         tmp |= AFMT_AZ_FORMAT_WTRIG_ACK;
3735                         WREG32(AFMT_AUDIO_PACKET_CONTROL + DCE3_HDMI_OFFSET1, tmp);
3736                 }
3737         } else {
3738                 if (rdev->irq.stat_regs.r600.hdmi0_status & HDMI0_AZ_FORMAT_WTRIG) {
3739                         tmp = RREG32(HDMI0_AUDIO_PACKET_CONTROL);
3740                         tmp |= HDMI0_AZ_FORMAT_WTRIG_ACK;
3741                         WREG32(HDMI0_AUDIO_PACKET_CONTROL, tmp);
3742                 }
3743                 if (rdev->irq.stat_regs.r600.hdmi1_status & HDMI0_AZ_FORMAT_WTRIG) {
3744                         if (ASIC_IS_DCE3(rdev)) {
3745                                 tmp = RREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL);
3746                                 tmp |= HDMI0_AZ_FORMAT_WTRIG_ACK;
3747                                 WREG32(DCE3_HDMI1_AUDIO_PACKET_CONTROL, tmp);
3748                         } else {
3749                                 tmp = RREG32(HDMI1_AUDIO_PACKET_CONTROL);
3750                                 tmp |= HDMI0_AZ_FORMAT_WTRIG_ACK;
3751                                 WREG32(HDMI1_AUDIO_PACKET_CONTROL, tmp);
3752                         }
3753                 }
3754         }
3755 }
3756
3757 void r600_irq_disable(struct radeon_device *rdev)
3758 {
3759         r600_disable_interrupts(rdev);
3760         /* Wait and acknowledge irq */
3761         mdelay(1);
3762         r600_irq_ack(rdev);
3763         r600_disable_interrupt_state(rdev);
3764 }
3765
3766 static u32 r600_get_ih_wptr(struct radeon_device *rdev)
3767 {
3768         u32 wptr, tmp;
3769
3770         if (rdev->wb.enabled)
3771                 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
3772         else
3773                 wptr = RREG32(IH_RB_WPTR);
3774
3775         if (wptr & RB_OVERFLOW) {
3776                 /* When a ring buffer overflow happen start parsing interrupt
3777                  * from the last not overwritten vector (wptr + 16). Hopefully
3778                  * this should allow us to catchup.
3779                  */
3780                 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
3781                         wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
3782                 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
3783                 tmp = RREG32(IH_RB_CNTL);
3784                 tmp |= IH_WPTR_OVERFLOW_CLEAR;
3785                 WREG32(IH_RB_CNTL, tmp);
3786         }
3787         return (wptr & rdev->ih.ptr_mask);
3788 }
3789
3790 /*        r600 IV Ring
3791  * Each IV ring entry is 128 bits:
3792  * [7:0]    - interrupt source id
3793  * [31:8]   - reserved
3794  * [59:32]  - interrupt source data
3795  * [127:60]  - reserved
3796  *
3797  * The basic interrupt vector entries
3798  * are decoded as follows:
3799  * src_id  src_data  description
3800  *      1         0  D1 Vblank
3801  *      1         1  D1 Vline
3802  *      5         0  D2 Vblank
3803  *      5         1  D2 Vline
3804  *     19         0  FP Hot plug detection A
3805  *     19         1  FP Hot plug detection B
3806  *     19         2  DAC A auto-detection
3807  *     19         3  DAC B auto-detection
3808  *     21         4  HDMI block A
3809  *     21         5  HDMI block B
3810  *    176         -  CP_INT RB
3811  *    177         -  CP_INT IB1
3812  *    178         -  CP_INT IB2
3813  *    181         -  EOP Interrupt
3814  *    233         -  GUI Idle
3815  *
3816  * Note, these are based on r600 and may need to be
3817  * adjusted or added to on newer asics
3818  */
3819
3820 int r600_irq_process(struct radeon_device *rdev)
3821 {
3822         u32 wptr;
3823         u32 rptr;
3824         u32 src_id, src_data;
3825         u32 ring_index;
3826         bool queue_hotplug = false;
3827         bool queue_hdmi = false;
3828
3829         if (!rdev->ih.enabled || rdev->shutdown)
3830                 return IRQ_NONE;
3831
3832         /* No MSIs, need a dummy read to flush PCI DMAs */
3833         if (!rdev->msi_enabled)
3834                 RREG32(IH_RB_WPTR);
3835
3836         wptr = r600_get_ih_wptr(rdev);
3837
3838 restart_ih:
3839         /* is somebody else already processing irqs? */
3840         if (atomic_xchg(&rdev->ih.lock, 1))
3841                 return IRQ_NONE;
3842
3843         rptr = rdev->ih.rptr;
3844         DRM_DEBUG("r600_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
3845
3846         /* Order reading of wptr vs. reading of IH ring data */
3847         rmb();
3848
3849         /* display interrupts */
3850         r600_irq_ack(rdev);
3851
3852         while (rptr != wptr) {
3853                 /* wptr/rptr are in bytes! */
3854                 ring_index = rptr / 4;
3855                 src_id = le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
3856                 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
3857
3858                 switch (src_id) {
3859                 case 1: /* D1 vblank/vline */
3860                         switch (src_data) {
3861                         case 0: /* D1 vblank */
3862                                 if (rdev->irq.stat_regs.r600.disp_int & LB_D1_VBLANK_INTERRUPT) {
3863                                         if (rdev->irq.crtc_vblank_int[0]) {
3864                                                 drm_handle_vblank(rdev->ddev, 0);
3865                                                 rdev->pm.vblank_sync = true;
3866                                                 wake_up(&rdev->irq.vblank_queue);
3867                                         }
3868                                         if (atomic_read(&rdev->irq.pflip[0]))
3869                                                 radeon_crtc_handle_flip(rdev, 0);
3870                                         rdev->irq.stat_regs.r600.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
3871                                         DRM_DEBUG("IH: D1 vblank\n");
3872                                 }
3873                                 break;
3874                         case 1: /* D1 vline */
3875                                 if (rdev->irq.stat_regs.r600.disp_int & LB_D1_VLINE_INTERRUPT) {
3876                                         rdev->irq.stat_regs.r600.disp_int &= ~LB_D1_VLINE_INTERRUPT;
3877                                         DRM_DEBUG("IH: D1 vline\n");
3878                                 }
3879                                 break;
3880                         default:
3881                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3882                                 break;
3883                         }
3884                         break;
3885                 case 5: /* D2 vblank/vline */
3886                         switch (src_data) {
3887                         case 0: /* D2 vblank */
3888                                 if (rdev->irq.stat_regs.r600.disp_int & LB_D2_VBLANK_INTERRUPT) {
3889                                         if (rdev->irq.crtc_vblank_int[1]) {
3890                                                 drm_handle_vblank(rdev->ddev, 1);
3891                                                 rdev->pm.vblank_sync = true;
3892                                                 wake_up(&rdev->irq.vblank_queue);
3893                                         }
3894                                         if (atomic_read(&rdev->irq.pflip[1]))
3895                                                 radeon_crtc_handle_flip(rdev, 1);
3896                                         rdev->irq.stat_regs.r600.disp_int &= ~LB_D2_VBLANK_INTERRUPT;
3897                                         DRM_DEBUG("IH: D2 vblank\n");
3898                                 }
3899                                 break;
3900                         case 1: /* D1 vline */
3901                                 if (rdev->irq.stat_regs.r600.disp_int & LB_D2_VLINE_INTERRUPT) {
3902                                         rdev->irq.stat_regs.r600.disp_int &= ~LB_D2_VLINE_INTERRUPT;
3903                                         DRM_DEBUG("IH: D2 vline\n");
3904                                 }
3905                                 break;
3906                         default:
3907                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3908                                 break;
3909                         }
3910                         break;
3911                 case 19: /* HPD/DAC hotplug */
3912                         switch (src_data) {
3913                         case 0:
3914                                 if (rdev->irq.stat_regs.r600.disp_int & DC_HPD1_INTERRUPT) {
3915                                         rdev->irq.stat_regs.r600.disp_int &= ~DC_HPD1_INTERRUPT;
3916                                         queue_hotplug = true;
3917                                         DRM_DEBUG("IH: HPD1\n");
3918                                 }
3919                                 break;
3920                         case 1:
3921                                 if (rdev->irq.stat_regs.r600.disp_int & DC_HPD2_INTERRUPT) {
3922                                         rdev->irq.stat_regs.r600.disp_int &= ~DC_HPD2_INTERRUPT;
3923                                         queue_hotplug = true;
3924                                         DRM_DEBUG("IH: HPD2\n");
3925                                 }
3926                                 break;
3927                         case 4:
3928                                 if (rdev->irq.stat_regs.r600.disp_int_cont & DC_HPD3_INTERRUPT) {
3929                                         rdev->irq.stat_regs.r600.disp_int_cont &= ~DC_HPD3_INTERRUPT;
3930                                         queue_hotplug = true;
3931                                         DRM_DEBUG("IH: HPD3\n");
3932                                 }
3933                                 break;
3934                         case 5:
3935                                 if (rdev->irq.stat_regs.r600.disp_int_cont & DC_HPD4_INTERRUPT) {
3936                                         rdev->irq.stat_regs.r600.disp_int_cont &= ~DC_HPD4_INTERRUPT;
3937                                         queue_hotplug = true;
3938                                         DRM_DEBUG("IH: HPD4\n");
3939                                 }
3940                                 break;
3941                         case 10:
3942                                 if (rdev->irq.stat_regs.r600.disp_int_cont2 & DC_HPD5_INTERRUPT) {
3943                                         rdev->irq.stat_regs.r600.disp_int_cont2 &= ~DC_HPD5_INTERRUPT;
3944                                         queue_hotplug = true;
3945                                         DRM_DEBUG("IH: HPD5\n");
3946                                 }
3947                                 break;
3948                         case 12:
3949                                 if (rdev->irq.stat_regs.r600.disp_int_cont2 & DC_HPD6_INTERRUPT) {
3950                                         rdev->irq.stat_regs.r600.disp_int_cont2 &= ~DC_HPD6_INTERRUPT;
3951                                         queue_hotplug = true;
3952                                         DRM_DEBUG("IH: HPD6\n");
3953                                 }
3954                                 break;
3955                         default:
3956                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3957                                 break;
3958                         }
3959                         break;
3960                 case 21: /* hdmi */
3961                         switch (src_data) {
3962                         case 4:
3963                                 if (rdev->irq.stat_regs.r600.hdmi0_status & HDMI0_AZ_FORMAT_WTRIG) {
3964                                         rdev->irq.stat_regs.r600.hdmi0_status &= ~HDMI0_AZ_FORMAT_WTRIG;
3965                                         queue_hdmi = true;
3966                                         DRM_DEBUG("IH: HDMI0\n");
3967                                 }
3968                                 break;
3969                         case 5:
3970                                 if (rdev->irq.stat_regs.r600.hdmi1_status & HDMI0_AZ_FORMAT_WTRIG) {
3971                                         rdev->irq.stat_regs.r600.hdmi1_status &= ~HDMI0_AZ_FORMAT_WTRIG;
3972                                         queue_hdmi = true;
3973                                         DRM_DEBUG("IH: HDMI1\n");
3974                                 }
3975                                 break;
3976                         default:
3977                                 DRM_ERROR("Unhandled interrupt: %d %d\n", src_id, src_data);
3978                                 break;
3979                         }
3980                         break;
3981                 case 176: /* CP_INT in ring buffer */
3982                 case 177: /* CP_INT in IB1 */
3983                 case 178: /* CP_INT in IB2 */
3984                         DRM_DEBUG("IH: CP int: 0x%08x\n", src_data);
3985                         radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
3986                         break;
3987                 case 181: /* CP EOP event */
3988                         DRM_DEBUG("IH: CP EOP\n");
3989                         radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
3990                         break;
3991                 case 224: /* DMA trap event */
3992                         DRM_DEBUG("IH: DMA trap\n");
3993                         radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
3994                         break;
3995                 case 233: /* GUI IDLE */
3996                         DRM_DEBUG("IH: GUI idle\n");
3997                         break;
3998                 default:
3999                         DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
4000                         break;
4001                 }
4002
4003                 /* wptr/rptr are in bytes! */
4004                 rptr += 16;
4005                 rptr &= rdev->ih.ptr_mask;
4006         }
4007         if (queue_hotplug)
4008                 schedule_work(&rdev->hotplug_work);
4009         if (queue_hdmi)
4010                 schedule_work(&rdev->audio_work);
4011         rdev->ih.rptr = rptr;
4012         WREG32(IH_RB_RPTR, rdev->ih.rptr);
4013         atomic_set(&rdev->ih.lock, 0);
4014
4015         /* make sure wptr hasn't changed while processing */
4016         wptr = r600_get_ih_wptr(rdev);
4017         if (wptr != rptr)
4018                 goto restart_ih;
4019
4020         return IRQ_HANDLED;
4021 }
4022
4023 /*
4024  * Debugfs info
4025  */
4026 #if defined(CONFIG_DEBUG_FS)
4027
4028 static int r600_debugfs_mc_info(struct seq_file *m, void *data)
4029 {
4030         struct drm_info_node *node = (struct drm_info_node *) m->private;
4031         struct drm_device *dev = node->minor->dev;
4032         struct radeon_device *rdev = dev->dev_private;
4033
4034         DREG32_SYS(m, rdev, R_000E50_SRBM_STATUS);
4035         DREG32_SYS(m, rdev, VM_L2_STATUS);
4036         return 0;
4037 }
4038
4039 static struct drm_info_list r600_mc_info_list[] = {
4040         {"r600_mc_info", r600_debugfs_mc_info, 0, NULL},
4041 };
4042 #endif
4043
4044 int r600_debugfs_mc_info_init(struct radeon_device *rdev)
4045 {
4046 #if defined(CONFIG_DEBUG_FS)
4047         return radeon_debugfs_add_files(rdev, r600_mc_info_list, ARRAY_SIZE(r600_mc_info_list));
4048 #else
4049         return 0;
4050 #endif
4051 }
4052
4053 /**
4054  * r600_ioctl_wait_idle - flush host path cache on wait idle ioctl
4055  * rdev: radeon device structure
4056  * bo: buffer object struct which userspace is waiting for idle
4057  *
4058  * Some R6XX/R7XX doesn't seems to take into account HDP flush performed
4059  * through ring buffer, this leads to corruption in rendering, see
4060  * http://bugzilla.kernel.org/show_bug.cgi?id=15186 to avoid this we
4061  * directly perform HDP flush by writing register through MMIO.
4062  */
4063 void r600_ioctl_wait_idle(struct radeon_device *rdev, struct radeon_bo *bo)
4064 {
4065         /* r7xx hw bug.  write to HDP_DEBUG1 followed by fb read
4066          * rather than write to HDP_REG_COHERENCY_FLUSH_CNTL.
4067          * This seems to cause problems on some AGP cards. Just use the old
4068          * method for them.
4069          */
4070         if ((rdev->family >= CHIP_RV770) && (rdev->family <= CHIP_RV740) &&
4071             rdev->vram_scratch.ptr && !(rdev->flags & RADEON_IS_AGP)) {
4072                 void __iomem *ptr = (void *)rdev->vram_scratch.ptr;
4073                 u32 tmp;
4074
4075                 WREG32(HDP_DEBUG1, 0);
4076                 tmp = readl((void __iomem *)ptr);
4077         } else
4078                 WREG32(R_005480_HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
4079 }
4080
4081 void r600_set_pcie_lanes(struct radeon_device *rdev, int lanes)
4082 {
4083         u32 link_width_cntl, mask, target_reg;
4084
4085         if (rdev->flags & RADEON_IS_IGP)
4086                 return;
4087
4088         if (!(rdev->flags & RADEON_IS_PCIE))
4089                 return;
4090
4091         /* x2 cards have a special sequence */
4092         if (ASIC_IS_X2(rdev))
4093                 return;
4094
4095         /* FIXME wait for idle */
4096
4097         switch (lanes) {
4098         case 0:
4099                 mask = RADEON_PCIE_LC_LINK_WIDTH_X0;
4100                 break;
4101         case 1:
4102                 mask = RADEON_PCIE_LC_LINK_WIDTH_X1;
4103                 break;
4104         case 2:
4105                 mask = RADEON_PCIE_LC_LINK_WIDTH_X2;
4106                 break;
4107         case 4:
4108                 mask = RADEON_PCIE_LC_LINK_WIDTH_X4;
4109                 break;
4110         case 8:
4111                 mask = RADEON_PCIE_LC_LINK_WIDTH_X8;
4112                 break;
4113         case 12:
4114                 mask = RADEON_PCIE_LC_LINK_WIDTH_X12;
4115                 break;
4116         case 16:
4117         default:
4118                 mask = RADEON_PCIE_LC_LINK_WIDTH_X16;
4119                 break;
4120         }
4121
4122         link_width_cntl = RREG32_PCIE_P(RADEON_PCIE_LC_LINK_WIDTH_CNTL);
4123
4124         if ((link_width_cntl & RADEON_PCIE_LC_LINK_WIDTH_RD_MASK) ==
4125             (mask << RADEON_PCIE_LC_LINK_WIDTH_RD_SHIFT))
4126                 return;
4127
4128         if (link_width_cntl & R600_PCIE_LC_UPCONFIGURE_DIS)
4129                 return;
4130
4131         link_width_cntl &= ~(RADEON_PCIE_LC_LINK_WIDTH_MASK |
4132                              RADEON_PCIE_LC_RECONFIG_NOW |
4133                              R600_PCIE_LC_RENEGOTIATE_EN |
4134                              R600_PCIE_LC_RECONFIG_ARC_MISSING_ESCAPE);
4135         link_width_cntl |= mask;
4136
4137         WREG32_PCIE_P(RADEON_PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
4138
4139         /* some northbridges can renegotiate the link rather than requiring
4140          * a complete re-config.
4141          * e.g., AMD 780/790 northbridges (pci ids: 0x5956, 0x5957, 0x5958, etc.)
4142          */
4143         if (link_width_cntl & R600_PCIE_LC_RENEGOTIATION_SUPPORT)
4144                 link_width_cntl |= R600_PCIE_LC_RENEGOTIATE_EN | R600_PCIE_LC_UPCONFIGURE_SUPPORT;
4145         else
4146                 link_width_cntl |= R600_PCIE_LC_RECONFIG_ARC_MISSING_ESCAPE;
4147
4148         WREG32_PCIE_P(RADEON_PCIE_LC_LINK_WIDTH_CNTL, (link_width_cntl |
4149                                                        RADEON_PCIE_LC_RECONFIG_NOW));
4150
4151         if (rdev->family >= CHIP_RV770)
4152                 target_reg = R700_TARGET_AND_CURRENT_PROFILE_INDEX;
4153         else
4154                 target_reg = R600_TARGET_AND_CURRENT_PROFILE_INDEX;
4155
4156         /* wait for lane set to complete */
4157         link_width_cntl = RREG32(target_reg);
4158         while (link_width_cntl == 0xffffffff)
4159                 link_width_cntl = RREG32(target_reg);
4160
4161 }
4162
4163 int r600_get_pcie_lanes(struct radeon_device *rdev)
4164 {
4165         u32 link_width_cntl;
4166
4167         if (rdev->flags & RADEON_IS_IGP)
4168                 return 0;
4169
4170         if (!(rdev->flags & RADEON_IS_PCIE))
4171                 return 0;
4172
4173         /* x2 cards have a special sequence */
4174         if (ASIC_IS_X2(rdev))
4175                 return 0;
4176
4177         /* FIXME wait for idle */
4178
4179         link_width_cntl = RREG32_PCIE_P(RADEON_PCIE_LC_LINK_WIDTH_CNTL);
4180
4181         switch ((link_width_cntl & RADEON_PCIE_LC_LINK_WIDTH_RD_MASK) >> RADEON_PCIE_LC_LINK_WIDTH_RD_SHIFT) {
4182         case RADEON_PCIE_LC_LINK_WIDTH_X0:
4183                 return 0;
4184         case RADEON_PCIE_LC_LINK_WIDTH_X1:
4185                 return 1;
4186         case RADEON_PCIE_LC_LINK_WIDTH_X2:
4187                 return 2;
4188         case RADEON_PCIE_LC_LINK_WIDTH_X4:
4189                 return 4;
4190         case RADEON_PCIE_LC_LINK_WIDTH_X8:
4191                 return 8;
4192         case RADEON_PCIE_LC_LINK_WIDTH_X16:
4193         default:
4194                 return 16;
4195         }
4196 }
4197
4198 static void r600_pcie_gen2_enable(struct radeon_device *rdev)
4199 {
4200         u32 link_width_cntl, lanes, speed_cntl, training_cntl, tmp;
4201         u16 link_cntl2;
4202         u32 mask;
4203         int ret;
4204
4205         if (radeon_pcie_gen2 == 0)
4206                 return;
4207
4208         if (rdev->flags & RADEON_IS_IGP)
4209                 return;
4210
4211         if (!(rdev->flags & RADEON_IS_PCIE))
4212                 return;
4213
4214         /* x2 cards have a special sequence */
4215         if (ASIC_IS_X2(rdev))
4216                 return;
4217
4218         /* only RV6xx+ chips are supported */
4219         if (rdev->family <= CHIP_R600)
4220                 return;
4221
4222         ret = drm_pcie_get_speed_cap_mask(rdev->ddev, &mask);
4223         if (ret != 0)
4224                 return;
4225
4226         if (!(mask & DRM_PCIE_SPEED_50))
4227                 return;
4228
4229         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
4230         if (speed_cntl & LC_CURRENT_DATA_RATE) {
4231                 DRM_INFO("PCIE gen 2 link speeds already enabled\n");
4232                 return;
4233         }
4234
4235         DRM_INFO("enabling PCIE gen 2 link speeds, disable with radeon.pcie_gen2=0\n");
4236
4237         /* 55 nm r6xx asics */
4238         if ((rdev->family == CHIP_RV670) ||
4239             (rdev->family == CHIP_RV620) ||
4240             (rdev->family == CHIP_RV635)) {
4241                 /* advertise upconfig capability */
4242                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
4243                 link_width_cntl &= ~LC_UPCONFIGURE_DIS;
4244                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
4245                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
4246                 if (link_width_cntl & LC_RENEGOTIATION_SUPPORT) {
4247                         lanes = (link_width_cntl & LC_LINK_WIDTH_RD_MASK) >> LC_LINK_WIDTH_RD_SHIFT;
4248                         link_width_cntl &= ~(LC_LINK_WIDTH_MASK |
4249                                              LC_RECONFIG_ARC_MISSING_ESCAPE);
4250                         link_width_cntl |= lanes | LC_RECONFIG_NOW | LC_RENEGOTIATE_EN;
4251                         WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
4252                 } else {
4253                         link_width_cntl |= LC_UPCONFIGURE_DIS;
4254                         WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
4255                 }
4256         }
4257
4258         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
4259         if ((speed_cntl & LC_OTHER_SIDE_EVER_SENT_GEN2) &&
4260             (speed_cntl & LC_OTHER_SIDE_SUPPORTS_GEN2)) {
4261
4262                 /* 55 nm r6xx asics */
4263                 if ((rdev->family == CHIP_RV670) ||
4264                     (rdev->family == CHIP_RV620) ||
4265                     (rdev->family == CHIP_RV635)) {
4266                         WREG32(MM_CFGREGS_CNTL, 0x8);
4267                         link_cntl2 = RREG32(0x4088);
4268                         WREG32(MM_CFGREGS_CNTL, 0);
4269                         /* not supported yet */
4270                         if (link_cntl2 & SELECTABLE_DEEMPHASIS)
4271                                 return;
4272                 }
4273
4274                 speed_cntl &= ~LC_SPEED_CHANGE_ATTEMPTS_ALLOWED_MASK;
4275                 speed_cntl |= (0x3 << LC_SPEED_CHANGE_ATTEMPTS_ALLOWED_SHIFT);
4276                 speed_cntl &= ~LC_VOLTAGE_TIMER_SEL_MASK;
4277                 speed_cntl &= ~LC_FORCE_DIS_HW_SPEED_CHANGE;
4278                 speed_cntl |= LC_FORCE_EN_HW_SPEED_CHANGE;
4279                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
4280
4281                 tmp = RREG32(0x541c);
4282                 WREG32(0x541c, tmp | 0x8);
4283                 WREG32(MM_CFGREGS_CNTL, MM_WR_TO_CFG_EN);
4284                 link_cntl2 = RREG16(0x4088);
4285                 link_cntl2 &= ~TARGET_LINK_SPEED_MASK;
4286                 link_cntl2 |= 0x2;
4287                 WREG16(0x4088, link_cntl2);
4288                 WREG32(MM_CFGREGS_CNTL, 0);
4289
4290                 if ((rdev->family == CHIP_RV670) ||
4291                     (rdev->family == CHIP_RV620) ||
4292                     (rdev->family == CHIP_RV635)) {
4293                         training_cntl = RREG32_PCIE_P(PCIE_LC_TRAINING_CNTL);
4294                         training_cntl &= ~LC_POINT_7_PLUS_EN;
4295                         WREG32_PCIE_P(PCIE_LC_TRAINING_CNTL, training_cntl);
4296                 } else {
4297                         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
4298                         speed_cntl &= ~LC_TARGET_LINK_SPEED_OVERRIDE_EN;
4299                         WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
4300                 }
4301
4302                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
4303                 speed_cntl |= LC_GEN2_EN_STRAP;
4304                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
4305
4306         } else {
4307                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
4308                 /* XXX: only disable it if gen1 bridge vendor == 0x111d or 0x1106 */
4309                 if (1)
4310                         link_width_cntl |= LC_UPCONFIGURE_DIS;
4311                 else
4312                         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
4313                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
4314         }
4315 }
4316
4317 /**
4318  * r600_get_gpu_clock - return GPU clock counter snapshot
4319  *
4320  * @rdev: radeon_device pointer
4321  *
4322  * Fetches a GPU clock counter snapshot (R6xx-cayman).
4323  * Returns the 64 bit clock counter snapshot.
4324  */
4325 uint64_t r600_get_gpu_clock(struct radeon_device *rdev)
4326 {
4327         uint64_t clock;
4328
4329         mutex_lock(&rdev->gpu_clock_mutex);
4330         WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
4331         clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
4332                 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
4333         mutex_unlock(&rdev->gpu_clock_mutex);
4334         return clock;
4335 }