drivers/gpu/drm/radeon/evergreen.c

   1 /*
   2  * Copyright 2010 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Alex Deucher
  23  */
  24 #include <linux/firmware.h>
  25 #include <linux/platform_device.h>
  26 #include <linux/slab.h>
  27 #include "drmP.h"
  28 #include "radeon.h"
  29 #include "radeon_asic.h"
  30 #include "radeon_drm.h"
  31 #include "evergreend.h"
  32 #include "atom.h"
  33 #include "avivod.h"
  34 #include "evergreen_reg.h"
  35 #include "evergreen_blit_shaders.h"
  36
  37 #define EVERGREEN_PFP_UCODE_SIZE 1120
  38 #define EVERGREEN_PM4_UCODE_SIZE 1376
  39
  40 static void evergreen_gpu_init(struct radeon_device *rdev);
  41 void evergreen_fini(struct radeon_device *rdev);
  42 void evergreen_pcie_gen2_enable(struct radeon_device *rdev);
  43
  44 void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev)
  45 {
  46         u16 ctl, v;
  47         int cap, err;
  48
  49         cap = pci_pcie_cap(rdev->pdev);
  50         if (!cap)
  51                 return;
  52
  53         err = pci_read_config_word(rdev->pdev, cap + PCI_EXP_DEVCTL, &ctl);
  54         if (err)
  55                 return;
  56
  57         v = (ctl & PCI_EXP_DEVCTL_READRQ) >> 12;
  58
  59         /* if bios or OS sets MAX_READ_REQUEST_SIZE to an invalid value, fix it
  60          * to avoid hangs or perfomance issues
  61          */
  62         if ((v == 0) || (v == 6) || (v == 7)) {
  63                 ctl &= ~PCI_EXP_DEVCTL_READRQ;
  64                 ctl |= (2 << 12);
  65                 pci_write_config_word(rdev->pdev, cap + PCI_EXP_DEVCTL, ctl);
  66         }
  67 }
  68
  69 void evergreen_pre_page_flip(struct radeon_device *rdev, int crtc)
  70 {
  71         /* enable the pflip int */
  72         radeon_irq_kms_pflip_irq_get(rdev, crtc);
  73 }
  74
  75 void evergreen_post_page_flip(struct radeon_device *rdev, int crtc)
  76 {
  77         /* disable the pflip int */
  78         radeon_irq_kms_pflip_irq_put(rdev, crtc);
  79 }
  80
  81 u32 evergreen_page_flip(struct radeon_device *rdev, int crtc_id, u64 crtc_base)
  82 {
  83         struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[crtc_id];
  84         u32 tmp = RREG32(EVERGREEN_GRPH_UPDATE + radeon_crtc->crtc_offset);
  85
  86         /* Lock the graphics update lock */
  87         tmp |= EVERGREEN_GRPH_UPDATE_LOCK;
  88         WREG32(EVERGREEN_GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
  89
  90         /* update the scanout addresses */
  91         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + radeon_crtc->crtc_offset,
  92                upper_32_bits(crtc_base));
  93         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  94                (u32)crtc_base);
  95
  96         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + radeon_crtc->crtc_offset,
  97                upper_32_bits(crtc_base));
  98         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  99                (u32)crtc_base);
 100
 101         /* Wait for update_pending to go high. */
 102         while (!(RREG32(EVERGREEN_GRPH_UPDATE + radeon_crtc->crtc_offset) & EVERGREEN_GRPH_SURFACE_UPDATE_PENDING));
 103         DRM_DEBUG("Update pending now high. Unlocking vupdate_lock.\n");
 104
 105         /* Unlock the lock, so double-buffering can take place inside vblank */
 106         tmp &= ~EVERGREEN_GRPH_UPDATE_LOCK;
 107         WREG32(EVERGREEN_GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
 108
 109         /* Return current update_pending status: */
 110         return RREG32(EVERGREEN_GRPH_UPDATE + radeon_crtc->crtc_offset) & EVERGREEN_GRPH_SURFACE_UPDATE_PENDING;
 111 }
 112
 113 /* get temperature in millidegrees */
 114 int evergreen_get_temp(struct radeon_device *rdev)
 115 {
 116         u32 temp, toffset;
 117         int actual_temp = 0;
 118
 119         if (rdev->family == CHIP_JUNIPER) {
 120                 toffset = (RREG32(CG_THERMAL_CTRL) & TOFFSET_MASK) >>
 121                         TOFFSET_SHIFT;
 122                 temp = (RREG32(CG_TS0_STATUS) & TS0_ADC_DOUT_MASK) >>
 123                         TS0_ADC_DOUT_SHIFT;
 124
 125                 if (toffset & 0x100)
 126                         actual_temp = temp / 2 - (0x200 - toffset);
 127                 else
 128                         actual_temp = temp / 2 + toffset;
 129
 130                 actual_temp = actual_temp * 1000;
 131
 132         } else {
 133                 temp = (RREG32(CG_MULT_THERMAL_STATUS) & ASIC_T_MASK) >>
 134                         ASIC_T_SHIFT;
 135
 136                 if (temp & 0x400)
 137                         actual_temp = -256;
 138                 else if (temp & 0x200)
 139                         actual_temp = 255;
 140                 else if (temp & 0x100) {
 141                         actual_temp = temp & 0x1ff;
 142                         actual_temp |= ~0x1ff;
 143                 } else
 144                         actual_temp = temp & 0xff;
 145
 146                 actual_temp = (actual_temp * 1000) / 2;
 147         }
 148
 149         return actual_temp;
 150 }
 151
 152 int sumo_get_temp(struct radeon_device *rdev)
 153 {
 154         u32 temp = RREG32(CG_THERMAL_STATUS) & 0xff;
 155         int actual_temp = temp - 49;
 156
 157         return actual_temp * 1000;
 158 }
 159
 160 void sumo_pm_init_profile(struct radeon_device *rdev)
 161 {
 162         int idx;
 163
 164         /* default */
 165         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_ps_idx = rdev->pm.default_power_state_index;
 166         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_ps_idx = rdev->pm.default_power_state_index;
 167         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_off_cm_idx = 0;
 168         rdev->pm.profiles[PM_PROFILE_DEFAULT_IDX].dpms_on_cm_idx = 0;
 169
 170         /* low,mid sh/mh */
 171         if (rdev->flags & RADEON_IS_MOBILITY)
 172                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_BATTERY, 0);
 173         else
 174                 idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 0);
 175
 176         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_ps_idx = idx;
 177         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_ps_idx = idx;
 178         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_off_cm_idx = 0;
 179         rdev->pm.profiles[PM_PROFILE_LOW_SH_IDX].dpms_on_cm_idx = 0;
 180
 181         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_ps_idx = idx;
 182         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_ps_idx = idx;
 183         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_off_cm_idx = 0;
 184         rdev->pm.profiles[PM_PROFILE_LOW_MH_IDX].dpms_on_cm_idx = 0;
 185
 186         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_ps_idx = idx;
 187         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_ps_idx = idx;
 188         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_off_cm_idx = 0;
 189         rdev->pm.profiles[PM_PROFILE_MID_SH_IDX].dpms_on_cm_idx = 0;
 190
 191         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_ps_idx = idx;
 192         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_ps_idx = idx;
 193         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_off_cm_idx = 0;
 194         rdev->pm.profiles[PM_PROFILE_MID_MH_IDX].dpms_on_cm_idx = 0;
 195
 196         /* high sh/mh */
 197         idx = radeon_pm_get_type_index(rdev, POWER_STATE_TYPE_PERFORMANCE, 0);
 198         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_ps_idx = idx;
 199         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_ps_idx = idx;
 200         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_off_cm_idx = 0;
 201         rdev->pm.profiles[PM_PROFILE_HIGH_SH_IDX].dpms_on_cm_idx =
 202                 rdev->pm.power_state[idx].num_clock_modes - 1;
 203
 204         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_ps_idx = idx;
 205         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_ps_idx = idx;
 206         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_off_cm_idx = 0;
 207         rdev->pm.profiles[PM_PROFILE_HIGH_MH_IDX].dpms_on_cm_idx =
 208                 rdev->pm.power_state[idx].num_clock_modes - 1;
 209 }
 210
 211 void evergreen_pm_misc(struct radeon_device *rdev)
 212 {
 213         int req_ps_idx = rdev->pm.requested_power_state_index;
 214         int req_cm_idx = rdev->pm.requested_clock_mode_index;
 215         struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx];
 216         struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage;
 217
 218         if (voltage->type == VOLTAGE_SW) {
 219                 /* 0xff01 is a flag rather then an actual voltage */
 220                 if (voltage->voltage == 0xff01)
 221                         return;
 222                 if (voltage->voltage && (voltage->voltage != rdev->pm.current_vddc)) {
 223                         radeon_atom_set_voltage(rdev, voltage->voltage, SET_VOLTAGE_TYPE_ASIC_VDDC);
 224                         rdev->pm.current_vddc = voltage->voltage;
 225                         DRM_DEBUG("Setting: vddc: %d\n", voltage->voltage);
 226                 }
 227                 /* 0xff01 is a flag rather then an actual voltage */
 228                 if (voltage->vddci == 0xff01)
 229                         return;
 230                 if (voltage->vddci && (voltage->vddci != rdev->pm.current_vddci)) {
 231                         radeon_atom_set_voltage(rdev, voltage->vddci, SET_VOLTAGE_TYPE_ASIC_VDDCI);
 232                         rdev->pm.current_vddci = voltage->vddci;
 233                         DRM_DEBUG("Setting: vddci: %d\n", voltage->vddci);
 234                 }
 235         }
 236 }
 237
 238 void evergreen_pm_prepare(struct radeon_device *rdev)
 239 {
 240         struct drm_device *ddev = rdev->ddev;
 241         struct drm_crtc *crtc;
 242         struct radeon_crtc *radeon_crtc;
 243         u32 tmp;
 244
 245         /* disable any active CRTCs */
 246         list_for_each_entry(crtc, &ddev->mode_config.crtc_list, head) {
 247                 radeon_crtc = to_radeon_crtc(crtc);
 248                 if (radeon_crtc->enabled) {
 249                         tmp = RREG32(EVERGREEN_CRTC_CONTROL + radeon_crtc->crtc_offset);
 250                         tmp |= EVERGREEN_CRTC_DISP_READ_REQUEST_DISABLE;
 251                         WREG32(EVERGREEN_CRTC_CONTROL + radeon_crtc->crtc_offset, tmp);
 252                 }
 253         }
 254 }
 255
 256 void evergreen_pm_finish(struct radeon_device *rdev)
 257 {
 258         struct drm_device *ddev = rdev->ddev;
 259         struct drm_crtc *crtc;
 260         struct radeon_crtc *radeon_crtc;
 261         u32 tmp;
 262
 263         /* enable any active CRTCs */
 264         list_for_each_entry(crtc, &ddev->mode_config.crtc_list, head) {
 265                 radeon_crtc = to_radeon_crtc(crtc);
 266                 if (radeon_crtc->enabled) {
 267                         tmp = RREG32(EVERGREEN_CRTC_CONTROL + radeon_crtc->crtc_offset);
 268                         tmp &= ~EVERGREEN_CRTC_DISP_READ_REQUEST_DISABLE;
 269                         WREG32(EVERGREEN_CRTC_CONTROL + radeon_crtc->crtc_offset, tmp);
 270                 }
 271         }
 272 }
 273
 274 bool evergreen_hpd_sense(struct radeon_device *rdev, enum radeon_hpd_id hpd)
 275 {
 276         bool connected = false;
 277
 278         switch (hpd) {
 279         case RADEON_HPD_1:
 280                 if (RREG32(DC_HPD1_INT_STATUS) & DC_HPDx_SENSE)
 281                         connected = true;
 282                 break;
 283         case RADEON_HPD_2:
 284                 if (RREG32(DC_HPD2_INT_STATUS) & DC_HPDx_SENSE)
 285                         connected = true;
 286                 break;
 287         case RADEON_HPD_3:
 288                 if (RREG32(DC_HPD3_INT_STATUS) & DC_HPDx_SENSE)
 289                         connected = true;
 290                 break;
 291         case RADEON_HPD_4:
 292                 if (RREG32(DC_HPD4_INT_STATUS) & DC_HPDx_SENSE)
 293                         connected = true;
 294                 break;
 295         case RADEON_HPD_5:
 296                 if (RREG32(DC_HPD5_INT_STATUS) & DC_HPDx_SENSE)
 297                         connected = true;
 298                 break;
 299         case RADEON_HPD_6:
 300                 if (RREG32(DC_HPD6_INT_STATUS) & DC_HPDx_SENSE)
 301                         connected = true;
 302                         break;
 303         default:
 304                 break;
 305         }
 306
 307         return connected;
 308 }
 309
 310 void evergreen_hpd_set_polarity(struct radeon_device *rdev,
 311                                 enum radeon_hpd_id hpd)
 312 {
 313         u32 tmp;
 314         bool connected = evergreen_hpd_sense(rdev, hpd);
 315
 316         switch (hpd) {
 317         case RADEON_HPD_1:
 318                 tmp = RREG32(DC_HPD1_INT_CONTROL);
 319                 if (connected)
 320                         tmp &= ~DC_HPDx_INT_POLARITY;
 321                 else
 322                         tmp |= DC_HPDx_INT_POLARITY;
 323                 WREG32(DC_HPD1_INT_CONTROL, tmp);
 324                 break;
 325         case RADEON_HPD_2:
 326                 tmp = RREG32(DC_HPD2_INT_CONTROL);
 327                 if (connected)
 328                         tmp &= ~DC_HPDx_INT_POLARITY;
 329                 else
 330                         tmp |= DC_HPDx_INT_POLARITY;
 331                 WREG32(DC_HPD2_INT_CONTROL, tmp);
 332                 break;
 333         case RADEON_HPD_3:
 334                 tmp = RREG32(DC_HPD3_INT_CONTROL);
 335                 if (connected)
 336                         tmp &= ~DC_HPDx_INT_POLARITY;
 337                 else
 338                         tmp |= DC_HPDx_INT_POLARITY;
 339                 WREG32(DC_HPD3_INT_CONTROL, tmp);
 340                 break;
 341         case RADEON_HPD_4:
 342                 tmp = RREG32(DC_HPD4_INT_CONTROL);
 343                 if (connected)
 344                         tmp &= ~DC_HPDx_INT_POLARITY;
 345                 else
 346                         tmp |= DC_HPDx_INT_POLARITY;
 347                 WREG32(DC_HPD4_INT_CONTROL, tmp);
 348                 break;
 349         case RADEON_HPD_5:
 350                 tmp = RREG32(DC_HPD5_INT_CONTROL);
 351                 if (connected)
 352                         tmp &= ~DC_HPDx_INT_POLARITY;
 353                 else
 354                         tmp |= DC_HPDx_INT_POLARITY;
 355                 WREG32(DC_HPD5_INT_CONTROL, tmp);
 356                         break;
 357         case RADEON_HPD_6:
 358                 tmp = RREG32(DC_HPD6_INT_CONTROL);
 359                 if (connected)
 360                         tmp &= ~DC_HPDx_INT_POLARITY;
 361                 else
 362                         tmp |= DC_HPDx_INT_POLARITY;
 363                 WREG32(DC_HPD6_INT_CONTROL, tmp);
 364                 break;
 365         default:
 366                 break;
 367         }
 368 }
 369
 370 void evergreen_hpd_init(struct radeon_device *rdev)
 371 {
 372         struct drm_device *dev = rdev->ddev;
 373         struct drm_connector *connector;
 374         u32 tmp = DC_HPDx_CONNECTION_TIMER(0x9c4) |
 375                 DC_HPDx_RX_INT_TIMER(0xfa) | DC_HPDx_EN;
 376
 377         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
 378                 struct radeon_connector *radeon_connector = to_radeon_connector(connector);
 379                 switch (radeon_connector->hpd.hpd) {
 380                 case RADEON_HPD_1:
 381                         WREG32(DC_HPD1_CONTROL, tmp);
 382                         rdev->irq.hpd[0] = true;
 383                         break;
 384                 case RADEON_HPD_2:
 385                         WREG32(DC_HPD2_CONTROL, tmp);
 386                         rdev->irq.hpd[1] = true;
 387                         break;
 388                 case RADEON_HPD_3:
 389                         WREG32(DC_HPD3_CONTROL, tmp);
 390                         rdev->irq.hpd[2] = true;
 391                         break;
 392                 case RADEON_HPD_4:
 393                         WREG32(DC_HPD4_CONTROL, tmp);
 394                         rdev->irq.hpd[3] = true;
 395                         break;
 396                 case RADEON_HPD_5:
 397                         WREG32(DC_HPD5_CONTROL, tmp);
 398                         rdev->irq.hpd[4] = true;
 399                         break;
 400                 case RADEON_HPD_6:
 401                         WREG32(DC_HPD6_CONTROL, tmp);
 402                         rdev->irq.hpd[5] = true;
 403                         break;
 404                 default:
 405                         break;
 406                 }
 407                 radeon_hpd_set_polarity(rdev, radeon_connector->hpd.hpd);
 408         }
 409         if (rdev->irq.installed)
 410                 evergreen_irq_set(rdev);
 411 }
 412
 413 void evergreen_hpd_fini(struct radeon_device *rdev)
 414 {
 415         struct drm_device *dev = rdev->ddev;
 416         struct drm_connector *connector;
 417
 418         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
 419                 struct radeon_connector *radeon_connector = to_radeon_connector(connector);
 420                 switch (radeon_connector->hpd.hpd) {
 421                 case RADEON_HPD_1:
 422                         WREG32(DC_HPD1_CONTROL, 0);
 423                         rdev->irq.hpd[0] = false;
 424                         break;
 425                 case RADEON_HPD_2:
 426                         WREG32(DC_HPD2_CONTROL, 0);
 427                         rdev->irq.hpd[1] = false;
 428                         break;
 429                 case RADEON_HPD_3:
 430                         WREG32(DC_HPD3_CONTROL, 0);
 431                         rdev->irq.hpd[2] = false;
 432                         break;
 433                 case RADEON_HPD_4:
 434                         WREG32(DC_HPD4_CONTROL, 0);
 435                         rdev->irq.hpd[3] = false;
 436                         break;
 437                 case RADEON_HPD_5:
 438                         WREG32(DC_HPD5_CONTROL, 0);
 439                         rdev->irq.hpd[4] = false;
 440                         break;
 441                 case RADEON_HPD_6:
 442                         WREG32(DC_HPD6_CONTROL, 0);
 443                         rdev->irq.hpd[5] = false;
 444                         break;
 445                 default:
 446                         break;
 447                 }
 448         }
 449 }
 450
 451 /* watermark setup */
 452
 453 static u32 evergreen_line_buffer_adjust(struct radeon_device *rdev,
 454                                         struct radeon_crtc *radeon_crtc,
 455                                         struct drm_display_mode *mode,
 456                                         struct drm_display_mode *other_mode)
 457 {
 458         u32 tmp;
 459         /*
 460          * Line Buffer Setup
 461          * There are 3 line buffers, each one shared by 2 display controllers.
 462          * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
 463          * the display controllers.  The paritioning is done via one of four
 464          * preset allocations specified in bits 2:0:
 465          * first display controller
 466          *  0 - first half of lb (3840 * 2)
 467          *  1 - first 3/4 of lb (5760 * 2)
 468          *  2 - whole lb (7680 * 2), other crtc must be disabled
 469          *  3 - first 1/4 of lb (1920 * 2)
 470          * second display controller
 471          *  4 - second half of lb (3840 * 2)
 472          *  5 - second 3/4 of lb (5760 * 2)
 473          *  6 - whole lb (7680 * 2), other crtc must be disabled
 474          *  7 - last 1/4 of lb (1920 * 2)
 475          */
 476         /* this can get tricky if we have two large displays on a paired group
 477          * of crtcs.  Ideally for multiple large displays we'd assign them to
 478          * non-linked crtcs for maximum line buffer allocation.
 479          */
 480         if (radeon_crtc->base.enabled && mode) {
 481                 if (other_mode)
 482                         tmp = 0; /* 1/2 */
 483                 else
 484                         tmp = 2; /* whole */
 485         } else
 486                 tmp = 0;
 487
 488         /* second controller of the pair uses second half of the lb */
 489         if (radeon_crtc->crtc_id % 2)
 490                 tmp += 4;
 491         WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset, tmp);
 492
 493         if (radeon_crtc->base.enabled && mode) {
 494                 switch (tmp) {
 495                 case 0:
 496                 case 4:
 497                 default:
 498                         if (ASIC_IS_DCE5(rdev))
 499                                 return 4096 * 2;
 500                         else
 501                                 return 3840 * 2;
 502                 case 1:
 503                 case 5:
 504                         if (ASIC_IS_DCE5(rdev))
 505                                 return 6144 * 2;
 506                         else
 507                                 return 5760 * 2;
 508                 case 2:
 509                 case 6:
 510                         if (ASIC_IS_DCE5(rdev))
 511                                 return 8192 * 2;
 512                         else
 513                                 return 7680 * 2;
 514                 case 3:
 515                 case 7:
 516                         if (ASIC_IS_DCE5(rdev))
 517                                 return 2048 * 2;
 518                         else
 519                                 return 1920 * 2;
 520                 }
 521         }
 522
 523         /* controller not enabled, so no lb used */
 524         return 0;
 525 }
 526
 527 static u32 evergreen_get_number_of_dram_channels(struct radeon_device *rdev)
 528 {
 529         u32 tmp = RREG32(MC_SHARED_CHMAP);
 530
 531         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 532         case 0:
 533         default:
 534                 return 1;
 535         case 1:
 536                 return 2;
 537         case 2:
 538                 return 4;
 539         case 3:
 540                 return 8;
 541         }
 542 }
 543
 544 struct evergreen_wm_params {
 545         u32 dram_channels; /* number of dram channels */
 546         u32 yclk;          /* bandwidth per dram data pin in kHz */
 547         u32 sclk;          /* engine clock in kHz */
 548         u32 disp_clk;      /* display clock in kHz */
 549         u32 src_width;     /* viewport width */
 550         u32 active_time;   /* active display time in ns */
 551         u32 blank_time;    /* blank time in ns */
 552         bool interlaced;    /* mode is interlaced */
 553         fixed20_12 vsc;    /* vertical scale ratio */
 554         u32 num_heads;     /* number of active crtcs */
 555         u32 bytes_per_pixel; /* bytes per pixel display + overlay */
 556         u32 lb_size;       /* line buffer allocated to pipe */
 557         u32 vtaps;         /* vertical scaler taps */
 558 };
 559
 560 static u32 evergreen_dram_bandwidth(struct evergreen_wm_params *wm)
 561 {
 562         /* Calculate DRAM Bandwidth and the part allocated to display. */
 563         fixed20_12 dram_efficiency; /* 0.7 */
 564         fixed20_12 yclk, dram_channels, bandwidth;
 565         fixed20_12 a;
 566
 567         a.full = dfixed_const(1000);
 568         yclk.full = dfixed_const(wm->yclk);
 569         yclk.full = dfixed_div(yclk, a);
 570         dram_channels.full = dfixed_const(wm->dram_channels * 4);
 571         a.full = dfixed_const(10);
 572         dram_efficiency.full = dfixed_const(7);
 573         dram_efficiency.full = dfixed_div(dram_efficiency, a);
 574         bandwidth.full = dfixed_mul(dram_channels, yclk);
 575         bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
 576
 577         return dfixed_trunc(bandwidth);
 578 }
 579
 580 static u32 evergreen_dram_bandwidth_for_display(struct evergreen_wm_params *wm)
 581 {
 582         /* Calculate DRAM Bandwidth and the part allocated to display. */
 583         fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
 584         fixed20_12 yclk, dram_channels, bandwidth;
 585         fixed20_12 a;
 586
 587         a.full = dfixed_const(1000);
 588         yclk.full = dfixed_const(wm->yclk);
 589         yclk.full = dfixed_div(yclk, a);
 590         dram_channels.full = dfixed_const(wm->dram_channels * 4);
 591         a.full = dfixed_const(10);
 592         disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
 593         disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
 594         bandwidth.full = dfixed_mul(dram_channels, yclk);
 595         bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
 596
 597         return dfixed_trunc(bandwidth);
 598 }
 599
 600 static u32 evergreen_data_return_bandwidth(struct evergreen_wm_params *wm)
 601 {
 602         /* Calculate the display Data return Bandwidth */
 603         fixed20_12 return_efficiency; /* 0.8 */
 604         fixed20_12 sclk, bandwidth;
 605         fixed20_12 a;
 606
 607         a.full = dfixed_const(1000);
 608         sclk.full = dfixed_const(wm->sclk);
 609         sclk.full = dfixed_div(sclk, a);
 610         a.full = dfixed_const(10);
 611         return_efficiency.full = dfixed_const(8);
 612         return_efficiency.full = dfixed_div(return_efficiency, a);
 613         a.full = dfixed_const(32);
 614         bandwidth.full = dfixed_mul(a, sclk);
 615         bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
 616
 617         return dfixed_trunc(bandwidth);
 618 }
 619
 620 static u32 evergreen_dmif_request_bandwidth(struct evergreen_wm_params *wm)
 621 {
 622         /* Calculate the DMIF Request Bandwidth */
 623         fixed20_12 disp_clk_request_efficiency; /* 0.8 */
 624         fixed20_12 disp_clk, bandwidth;
 625         fixed20_12 a;
 626
 627         a.full = dfixed_const(1000);
 628         disp_clk.full = dfixed_const(wm->disp_clk);
 629         disp_clk.full = dfixed_div(disp_clk, a);
 630         a.full = dfixed_const(10);
 631         disp_clk_request_efficiency.full = dfixed_const(8);
 632         disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
 633         a.full = dfixed_const(32);
 634         bandwidth.full = dfixed_mul(a, disp_clk);
 635         bandwidth.full = dfixed_mul(bandwidth, disp_clk_request_efficiency);
 636
 637         return dfixed_trunc(bandwidth);
 638 }
 639
 640 static u32 evergreen_available_bandwidth(struct evergreen_wm_params *wm)
 641 {
 642         /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
 643         u32 dram_bandwidth = evergreen_dram_bandwidth(wm);
 644         u32 data_return_bandwidth = evergreen_data_return_bandwidth(wm);
 645         u32 dmif_req_bandwidth = evergreen_dmif_request_bandwidth(wm);
 646
 647         return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
 648 }
 649
 650 static u32 evergreen_average_bandwidth(struct evergreen_wm_params *wm)
 651 {
 652         /* Calculate the display mode Average Bandwidth
 653          * DisplayMode should contain the source and destination dimensions,
 654          * timing, etc.
 655          */
 656         fixed20_12 bpp;
 657         fixed20_12 line_time;
 658         fixed20_12 src_width;
 659         fixed20_12 bandwidth;
 660         fixed20_12 a;
 661
 662         a.full = dfixed_const(1000);
 663         line_time.full = dfixed_const(wm->active_time + wm->blank_time);
 664         line_time.full = dfixed_div(line_time, a);
 665         bpp.full = dfixed_const(wm->bytes_per_pixel);
 666         src_width.full = dfixed_const(wm->src_width);
 667         bandwidth.full = dfixed_mul(src_width, bpp);
 668         bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
 669         bandwidth.full = dfixed_div(bandwidth, line_time);
 670
 671         return dfixed_trunc(bandwidth);
 672 }
 673
 674 static u32 evergreen_latency_watermark(struct evergreen_wm_params *wm)
 675 {
 676         /* First calcualte the latency in ns */
 677         u32 mc_latency = 2000; /* 2000 ns. */
 678         u32 available_bandwidth = evergreen_available_bandwidth(wm);
 679         u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
 680         u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
 681         u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
 682         u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
 683                 (wm->num_heads * cursor_line_pair_return_time);
 684         u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
 685         u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
 686         fixed20_12 a, b, c;
 687
 688         if (wm->num_heads == 0)
 689                 return 0;
 690
 691         a.full = dfixed_const(2);
 692         b.full = dfixed_const(1);
 693         if ((wm->vsc.full > a.full) ||
 694             ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
 695             (wm->vtaps >= 5) ||
 696             ((wm->vsc.full >= a.full) && wm->interlaced))
 697                 max_src_lines_per_dst_line = 4;
 698         else
 699                 max_src_lines_per_dst_line = 2;
 700
 701         a.full = dfixed_const(available_bandwidth);
 702         b.full = dfixed_const(wm->num_heads);
 703         a.full = dfixed_div(a, b);
 704
 705         b.full = dfixed_const(1000);
 706         c.full = dfixed_const(wm->disp_clk);
 707         b.full = dfixed_div(c, b);
 708         c.full = dfixed_const(wm->bytes_per_pixel);
 709         b.full = dfixed_mul(b, c);
 710
 711         lb_fill_bw = min(dfixed_trunc(a), dfixed_trunc(b));
 712
 713         a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
 714         b.full = dfixed_const(1000);
 715         c.full = dfixed_const(lb_fill_bw);
 716         b.full = dfixed_div(c, b);
 717         a.full = dfixed_div(a, b);
 718         line_fill_time = dfixed_trunc(a);
 719
 720         if (line_fill_time < wm->active_time)
 721                 return latency;
 722         else
 723                 return latency + (line_fill_time - wm->active_time);
 724
 725 }
 726
 727 static bool evergreen_average_bandwidth_vs_dram_bandwidth_for_display(struct evergreen_wm_params *wm)
 728 {
 729         if (evergreen_average_bandwidth(wm) <=
 730             (evergreen_dram_bandwidth_for_display(wm) / wm->num_heads))
 731                 return true;
 732         else
 733                 return false;
 734 };
 735
 736 static bool evergreen_average_bandwidth_vs_available_bandwidth(struct evergreen_wm_params *wm)
 737 {
 738         if (evergreen_average_bandwidth(wm) <=
 739             (evergreen_available_bandwidth(wm) / wm->num_heads))
 740                 return true;
 741         else
 742                 return false;
 743 };
 744
 745 static bool evergreen_check_latency_hiding(struct evergreen_wm_params *wm)
 746 {
 747         u32 lb_partitions = wm->lb_size / wm->src_width;
 748         u32 line_time = wm->active_time + wm->blank_time;
 749         u32 latency_tolerant_lines;
 750         u32 latency_hiding;
 751         fixed20_12 a;
 752
 753         a.full = dfixed_const(1);
 754         if (wm->vsc.full > a.full)
 755                 latency_tolerant_lines = 1;
 756         else {
 757                 if (lb_partitions <= (wm->vtaps + 1))
 758                         latency_tolerant_lines = 1;
 759                 else
 760                         latency_tolerant_lines = 2;
 761         }
 762
 763         latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
 764
 765         if (evergreen_latency_watermark(wm) <= latency_hiding)
 766                 return true;
 767         else
 768                 return false;
 769 }
 770
 771 static void evergreen_program_watermarks(struct radeon_device *rdev,
 772                                          struct radeon_crtc *radeon_crtc,
 773                                          u32 lb_size, u32 num_heads)
 774 {
 775         struct drm_display_mode *mode = &radeon_crtc->base.mode;
 776         struct evergreen_wm_params wm;
 777         u32 pixel_period;
 778         u32 line_time = 0;
 779         u32 latency_watermark_a = 0, latency_watermark_b = 0;
 780         u32 priority_a_mark = 0, priority_b_mark = 0;
 781         u32 priority_a_cnt = PRIORITY_OFF;
 782         u32 priority_b_cnt = PRIORITY_OFF;
 783         u32 pipe_offset = radeon_crtc->crtc_id * 16;
 784         u32 tmp, arb_control3;
 785         fixed20_12 a, b, c;
 786
 787         if (radeon_crtc->base.enabled && num_heads && mode) {
 788                 pixel_period = 1000000 / (u32)mode->clock;
 789                 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
 790                 priority_a_cnt = 0;
 791                 priority_b_cnt = 0;
 792
 793                 wm.yclk = rdev->pm.current_mclk * 10;
 794                 wm.sclk = rdev->pm.current_sclk * 10;
 795                 wm.disp_clk = mode->clock;
 796                 wm.src_width = mode->crtc_hdisplay;
 797                 wm.active_time = mode->crtc_hdisplay * pixel_period;
 798                 wm.blank_time = line_time - wm.active_time;
 799                 wm.interlaced = false;
 800                 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
 801                         wm.interlaced = true;
 802                 wm.vsc = radeon_crtc->vsc;
 803                 wm.vtaps = 1;
 804                 if (radeon_crtc->rmx_type != RMX_OFF)
 805                         wm.vtaps = 2;
 806                 wm.bytes_per_pixel = 4; /* XXX: get this from fb config */
 807                 wm.lb_size = lb_size;
 808                 wm.dram_channels = evergreen_get_number_of_dram_channels(rdev);
 809                 wm.num_heads = num_heads;
 810
 811                 /* set for high clocks */
 812                 latency_watermark_a = min(evergreen_latency_watermark(&wm), (u32)65535);
 813                 /* set for low clocks */
 814                 /* wm.yclk = low clk; wm.sclk = low clk */
 815                 latency_watermark_b = min(evergreen_latency_watermark(&wm), (u32)65535);
 816
 817                 /* possibly force display priority to high */
 818                 /* should really do this at mode validation time... */
 819                 if (!evergreen_average_bandwidth_vs_dram_bandwidth_for_display(&wm) ||
 820                     !evergreen_average_bandwidth_vs_available_bandwidth(&wm) ||
 821                     !evergreen_check_latency_hiding(&wm) ||
 822                     (rdev->disp_priority == 2)) {
 823                         DRM_DEBUG_KMS("force priority to high\n");
 824                         priority_a_cnt |= PRIORITY_ALWAYS_ON;
 825                         priority_b_cnt |= PRIORITY_ALWAYS_ON;
 826                 }
 827
 828                 a.full = dfixed_const(1000);
 829                 b.full = dfixed_const(mode->clock);
 830                 b.full = dfixed_div(b, a);
 831                 c.full = dfixed_const(latency_watermark_a);
 832                 c.full = dfixed_mul(c, b);
 833                 c.full = dfixed_mul(c, radeon_crtc->hsc);
 834                 c.full = dfixed_div(c, a);
 835                 a.full = dfixed_const(16);
 836                 c.full = dfixed_div(c, a);
 837                 priority_a_mark = dfixed_trunc(c);
 838                 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
 839
 840                 a.full = dfixed_const(1000);
 841                 b.full = dfixed_const(mode->clock);
 842                 b.full = dfixed_div(b, a);
 843                 c.full = dfixed_const(latency_watermark_b);
 844                 c.full = dfixed_mul(c, b);
 845                 c.full = dfixed_mul(c, radeon_crtc->hsc);
 846                 c.full = dfixed_div(c, a);
 847                 a.full = dfixed_const(16);
 848                 c.full = dfixed_div(c, a);
 849                 priority_b_mark = dfixed_trunc(c);
 850                 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
 851         }
 852
 853         /* select wm A */
 854         arb_control3 = RREG32(PIPE0_ARBITRATION_CONTROL3 + pipe_offset);
 855         tmp = arb_control3;
 856         tmp &= ~LATENCY_WATERMARK_MASK(3);
 857         tmp |= LATENCY_WATERMARK_MASK(1);
 858         WREG32(PIPE0_ARBITRATION_CONTROL3 + pipe_offset, tmp);
 859         WREG32(PIPE0_LATENCY_CONTROL + pipe_offset,
 860                (LATENCY_LOW_WATERMARK(latency_watermark_a) |
 861                 LATENCY_HIGH_WATERMARK(line_time)));
 862         /* select wm B */
 863         tmp = RREG32(PIPE0_ARBITRATION_CONTROL3 + pipe_offset);
 864         tmp &= ~LATENCY_WATERMARK_MASK(3);
 865         tmp |= LATENCY_WATERMARK_MASK(2);
 866         WREG32(PIPE0_ARBITRATION_CONTROL3 + pipe_offset, tmp);
 867         WREG32(PIPE0_LATENCY_CONTROL + pipe_offset,
 868                (LATENCY_LOW_WATERMARK(latency_watermark_b) |
 869                 LATENCY_HIGH_WATERMARK(line_time)));
 870         /* restore original selection */
 871         WREG32(PIPE0_ARBITRATION_CONTROL3 + pipe_offset, arb_control3);
 872
 873         /* write the priority marks */
 874         WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
 875         WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
 876
 877 }
 878
 879 void evergreen_bandwidth_update(struct radeon_device *rdev)
 880 {
 881         struct drm_display_mode *mode0 = NULL;
 882         struct drm_display_mode *mode1 = NULL;
 883         u32 num_heads = 0, lb_size;
 884         int i;
 885
 886         radeon_update_display_priority(rdev);
 887
 888         for (i = 0; i < rdev->num_crtc; i++) {
 889                 if (rdev->mode_info.crtcs[i]->base.enabled)
 890                         num_heads++;
 891         }
 892         for (i = 0; i < rdev->num_crtc; i += 2) {
 893                 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
 894                 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
 895                 lb_size = evergreen_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
 896                 evergreen_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
 897                 lb_size = evergreen_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
 898                 evergreen_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
 899         }
 900 }
 901
 902 int evergreen_mc_wait_for_idle(struct radeon_device *rdev)
 903 {
 904         unsigned i;
 905         u32 tmp;
 906
 907         for (i = 0; i < rdev->usec_timeout; i++) {
 908                 /* read MC_STATUS */
 909                 tmp = RREG32(SRBM_STATUS) & 0x1F00;
 910                 if (!tmp)
 911                         return 0;
 912                 udelay(1);
 913         }
 914         return -1;
 915 }
 916
 917 /*
 918  * GART
 919  */
 920 void evergreen_pcie_gart_tlb_flush(struct radeon_device *rdev)
 921 {
 922         unsigned i;
 923         u32 tmp;
 924
 925         WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
 926
 927         WREG32(VM_CONTEXT0_REQUEST_RESPONSE, REQUEST_TYPE(1));
 928         for (i = 0; i < rdev->usec_timeout; i++) {
 929                 /* read MC_STATUS */
 930                 tmp = RREG32(VM_CONTEXT0_REQUEST_RESPONSE);
 931                 tmp = (tmp & RESPONSE_TYPE_MASK) >> RESPONSE_TYPE_SHIFT;
 932                 if (tmp == 2) {
 933                         printk(KERN_WARNING "[drm] r600 flush TLB failed\n");
 934                         return;
 935                 }
 936                 if (tmp) {
 937                         return;
 938                 }
 939                 udelay(1);
 940         }
 941 }
 942
 943 int evergreen_pcie_gart_enable(struct radeon_device *rdev)
 944 {
 945         u32 tmp;
 946         int r;
 947
 948         if (rdev->gart.robj == NULL) {
 949                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
 950                 return -EINVAL;
 951         }
 952         r = radeon_gart_table_vram_pin(rdev);
 953         if (r)
 954                 return r;
 955         radeon_gart_restore(rdev);
 956         /* Setup L2 cache */
 957         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 958                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 959                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 960         WREG32(VM_L2_CNTL2, 0);
 961         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 962         /* Setup TLB control */
 963         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 964                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 965                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 966                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 967         if (rdev->flags & RADEON_IS_IGP) {
 968                 WREG32(FUS_MC_VM_MD_L1_TLB0_CNTL, tmp);
 969                 WREG32(FUS_MC_VM_MD_L1_TLB1_CNTL, tmp);
 970                 WREG32(FUS_MC_VM_MD_L1_TLB2_CNTL, tmp);
 971         } else {
 972                 WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 973                 WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 974                 WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 975         }
 976         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 977         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 978         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 979         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 980         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 981         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
 982         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
 983         WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
 984                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
 985         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 986                         (u32)(rdev->dummy_page.addr >> 12));
 987         WREG32(VM_CONTEXT1_CNTL, 0);
 988
 989         evergreen_pcie_gart_tlb_flush(rdev);
 990         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 991                  (unsigned)(rdev->mc.gtt_size >> 20),
 992                  (unsigned long long)rdev->gart.table_addr);
 993         rdev->gart.ready = true;
 994         return 0;
 995 }
 996
 997 void evergreen_pcie_gart_disable(struct radeon_device *rdev)
 998 {
 999         u32 tmp;
1000
1001         /* Disable all tables */
1002         WREG32(VM_CONTEXT0_CNTL, 0);
1003         WREG32(VM_CONTEXT1_CNTL, 0);
1004
1005         /* Setup L2 cache */
1006         WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
1007                                 EFFECTIVE_L2_QUEUE_SIZE(7));
1008         WREG32(VM_L2_CNTL2, 0);
1009         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
1010         /* Setup TLB control */
1011         tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
1012         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
1013         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
1014         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
1015         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
1016         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
1017         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
1018         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
1019         radeon_gart_table_vram_unpin(rdev);
1020 }
1021
1022 void evergreen_pcie_gart_fini(struct radeon_device *rdev)
1023 {
1024         evergreen_pcie_gart_disable(rdev);
1025         radeon_gart_table_vram_free(rdev);
1026         radeon_gart_fini(rdev);
1027 }
1028
1029
1030 void evergreen_agp_enable(struct radeon_device *rdev)
1031 {
1032         u32 tmp;
1033
1034         /* Setup L2 cache */
1035         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
1036                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
1037                                 EFFECTIVE_L2_QUEUE_SIZE(7));
1038         WREG32(VM_L2_CNTL2, 0);
1039         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
1040         /* Setup TLB control */
1041         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
1042                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
1043                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
1044                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
1045         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
1046         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
1047         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
1048         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
1049         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
1050         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
1051         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
1052         WREG32(VM_CONTEXT0_CNTL, 0);
1053         WREG32(VM_CONTEXT1_CNTL, 0);
1054 }
1055
1056 void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save)
1057 {
1058         save->vga_control[0] = RREG32(D1VGA_CONTROL);
1059         save->vga_control[1] = RREG32(D2VGA_CONTROL);
1060         save->vga_render_control = RREG32(VGA_RENDER_CONTROL);
1061         save->vga_hdp_control = RREG32(VGA_HDP_CONTROL);
1062         save->crtc_control[0] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET);
1063         save->crtc_control[1] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET);
1064         if (rdev->num_crtc >= 4) {
1065                 save->vga_control[2] = RREG32(EVERGREEN_D3VGA_CONTROL);
1066                 save->vga_control[3] = RREG32(EVERGREEN_D4VGA_CONTROL);
1067                 save->crtc_control[2] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET);
1068                 save->crtc_control[3] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET);
1069         }
1070         if (rdev->num_crtc >= 6) {
1071                 save->vga_control[4] = RREG32(EVERGREEN_D5VGA_CONTROL);
1072                 save->vga_control[5] = RREG32(EVERGREEN_D6VGA_CONTROL);
1073                 save->crtc_control[4] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET);
1074                 save->crtc_control[5] = RREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET);
1075         }
1076
1077         /* Stop all video */
1078         WREG32(VGA_RENDER_CONTROL, 0);
1079         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC0_REGISTER_OFFSET, 1);
1080         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC1_REGISTER_OFFSET, 1);
1081         if (rdev->num_crtc >= 4) {
1082                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC2_REGISTER_OFFSET, 1);
1083                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC3_REGISTER_OFFSET, 1);
1084         }
1085         if (rdev->num_crtc >= 6) {
1086                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC4_REGISTER_OFFSET, 1);
1087                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC5_REGISTER_OFFSET, 1);
1088         }
1089         WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
1090         WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
1091         if (rdev->num_crtc >= 4) {
1092                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
1093                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
1094         }
1095         if (rdev->num_crtc >= 6) {
1096                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
1097                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
1098         }
1099         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
1100         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
1101         if (rdev->num_crtc >= 4) {
1102                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
1103                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
1104         }
1105         if (rdev->num_crtc >= 6) {
1106                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
1107                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
1108         }
1109
1110         WREG32(D1VGA_CONTROL, 0);
1111         WREG32(D2VGA_CONTROL, 0);
1112         if (rdev->num_crtc >= 4) {
1113                 WREG32(EVERGREEN_D3VGA_CONTROL, 0);
1114                 WREG32(EVERGREEN_D4VGA_CONTROL, 0);
1115         }
1116         if (rdev->num_crtc >= 6) {
1117                 WREG32(EVERGREEN_D5VGA_CONTROL, 0);
1118                 WREG32(EVERGREEN_D6VGA_CONTROL, 0);
1119         }
1120 }
1121
1122 void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save)
1123 {
1124         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC0_REGISTER_OFFSET,
1125                upper_32_bits(rdev->mc.vram_start));
1126         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC0_REGISTER_OFFSET,
1127                upper_32_bits(rdev->mc.vram_start));
1128         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC0_REGISTER_OFFSET,
1129                (u32)rdev->mc.vram_start);
1130         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC0_REGISTER_OFFSET,
1131                (u32)rdev->mc.vram_start);
1132
1133         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC1_REGISTER_OFFSET,
1134                upper_32_bits(rdev->mc.vram_start));
1135         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC1_REGISTER_OFFSET,
1136                upper_32_bits(rdev->mc.vram_start));
1137         WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC1_REGISTER_OFFSET,
1138                (u32)rdev->mc.vram_start);
1139         WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC1_REGISTER_OFFSET,
1140                (u32)rdev->mc.vram_start);
1141
1142         if (rdev->num_crtc >= 4) {
1143                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC2_REGISTER_OFFSET,
1144                        upper_32_bits(rdev->mc.vram_start));
1145                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC2_REGISTER_OFFSET,
1146                        upper_32_bits(rdev->mc.vram_start));
1147                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC2_REGISTER_OFFSET,
1148                        (u32)rdev->mc.vram_start);
1149                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC2_REGISTER_OFFSET,
1150                        (u32)rdev->mc.vram_start);
1151
1152                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC3_REGISTER_OFFSET,
1153                        upper_32_bits(rdev->mc.vram_start));
1154                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC3_REGISTER_OFFSET,
1155                        upper_32_bits(rdev->mc.vram_start));
1156                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC3_REGISTER_OFFSET,
1157                        (u32)rdev->mc.vram_start);
1158                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC3_REGISTER_OFFSET,
1159                        (u32)rdev->mc.vram_start);
1160         }
1161         if (rdev->num_crtc >= 6) {
1162                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC4_REGISTER_OFFSET,
1163                        upper_32_bits(rdev->mc.vram_start));
1164                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC4_REGISTER_OFFSET,
1165                        upper_32_bits(rdev->mc.vram_start));
1166                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC4_REGISTER_OFFSET,
1167                        (u32)rdev->mc.vram_start);
1168                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC4_REGISTER_OFFSET,
1169                        (u32)rdev->mc.vram_start);
1170
1171                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC5_REGISTER_OFFSET,
1172                        upper_32_bits(rdev->mc.vram_start));
1173                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS_HIGH + EVERGREEN_CRTC5_REGISTER_OFFSET,
1174                        upper_32_bits(rdev->mc.vram_start));
1175                 WREG32(EVERGREEN_GRPH_PRIMARY_SURFACE_ADDRESS + EVERGREEN_CRTC5_REGISTER_OFFSET,
1176                        (u32)rdev->mc.vram_start);
1177                 WREG32(EVERGREEN_GRPH_SECONDARY_SURFACE_ADDRESS + EVERGREEN_CRTC5_REGISTER_OFFSET,
1178                        (u32)rdev->mc.vram_start);
1179         }
1180
1181         WREG32(EVERGREEN_VGA_MEMORY_BASE_ADDRESS_HIGH, upper_32_bits(rdev->mc.vram_start));
1182         WREG32(EVERGREEN_VGA_MEMORY_BASE_ADDRESS, (u32)rdev->mc.vram_start);
1183         /* Unlock host access */
1184         WREG32(VGA_HDP_CONTROL, save->vga_hdp_control);
1185         mdelay(1);
1186         /* Restore video state */
1187         WREG32(D1VGA_CONTROL, save->vga_control[0]);
1188         WREG32(D2VGA_CONTROL, save->vga_control[1]);
1189         if (rdev->num_crtc >= 4) {
1190                 WREG32(EVERGREEN_D3VGA_CONTROL, save->vga_control[2]);
1191                 WREG32(EVERGREEN_D4VGA_CONTROL, save->vga_control[3]);
1192         }
1193         if (rdev->num_crtc >= 6) {
1194                 WREG32(EVERGREEN_D5VGA_CONTROL, save->vga_control[4]);
1195                 WREG32(EVERGREEN_D6VGA_CONTROL, save->vga_control[5]);
1196         }
1197         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC0_REGISTER_OFFSET, 1);
1198         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC1_REGISTER_OFFSET, 1);
1199         if (rdev->num_crtc >= 4) {
1200                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC2_REGISTER_OFFSET, 1);
1201                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC3_REGISTER_OFFSET, 1);
1202         }
1203         if (rdev->num_crtc >= 6) {
1204                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC4_REGISTER_OFFSET, 1);
1205                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC5_REGISTER_OFFSET, 1);
1206         }
1207         WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, save->crtc_control[0]);
1208         WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, save->crtc_control[1]);
1209         if (rdev->num_crtc >= 4) {
1210                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, save->crtc_control[2]);
1211                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, save->crtc_control[3]);
1212         }
1213         if (rdev->num_crtc >= 6) {
1214                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, save->crtc_control[4]);
1215                 WREG32(EVERGREEN_CRTC_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, save->crtc_control[5]);
1216         }
1217         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
1218         WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
1219         if (rdev->num_crtc >= 4) {
1220                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
1221                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
1222         }
1223         if (rdev->num_crtc >= 6) {
1224                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
1225                 WREG32(EVERGREEN_CRTC_UPDATE_LOCK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
1226         }
1227         WREG32(VGA_RENDER_CONTROL, save->vga_render_control);
1228 }
1229
1230 void evergreen_mc_program(struct radeon_device *rdev)
1231 {
1232         struct evergreen_mc_save save;
1233         u32 tmp;
1234         int i, j;
1235
1236         /* Initialize HDP */
1237         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
1238                 WREG32((0x2c14 + j), 0x00000000);
1239                 WREG32((0x2c18 + j), 0x00000000);
1240                 WREG32((0x2c1c + j), 0x00000000);
1241                 WREG32((0x2c20 + j), 0x00000000);
1242                 WREG32((0x2c24 + j), 0x00000000);
1243         }
1244         WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
1245
1246         evergreen_mc_stop(rdev, &save);
1247         if (evergreen_mc_wait_for_idle(rdev)) {
1248                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1249         }
1250         /* Lockout access through VGA aperture*/
1251         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
1252         /* Update configuration */
1253         if (rdev->flags & RADEON_IS_AGP) {
1254                 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
1255                         /* VRAM before AGP */
1256                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
1257                                 rdev->mc.vram_start >> 12);
1258                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
1259                                 rdev->mc.gtt_end >> 12);
1260                 } else {
1261                         /* VRAM after AGP */
1262                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
1263                                 rdev->mc.gtt_start >> 12);
1264                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
1265                                 rdev->mc.vram_end >> 12);
1266                 }
1267         } else {
1268                 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
1269                         rdev->mc.vram_start >> 12);
1270                 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
1271                         rdev->mc.vram_end >> 12);
1272         }
1273         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, rdev->vram_scratch.gpu_addr >> 12);
1274         if (rdev->flags & RADEON_IS_IGP) {
1275                 tmp = RREG32(MC_FUS_VM_FB_OFFSET) & 0x000FFFFF;
1276                 tmp |= ((rdev->mc.vram_end >> 20) & 0xF) << 24;
1277                 tmp |= ((rdev->mc.vram_start >> 20) & 0xF) << 20;
1278                 WREG32(MC_FUS_VM_FB_OFFSET, tmp);
1279         }
1280         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
1281         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
1282         WREG32(MC_VM_FB_LOCATION, tmp);
1283         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
1284         WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
1285         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
1286         if (rdev->flags & RADEON_IS_AGP) {
1287                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
1288                 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16);
1289                 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
1290         } else {
1291                 WREG32(MC_VM_AGP_BASE, 0);
1292                 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
1293                 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
1294         }
1295         if (evergreen_mc_wait_for_idle(rdev)) {
1296                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
1297         }
1298         evergreen_mc_resume(rdev, &save);
1299         /* we need to own VRAM, so turn off the VGA renderer here
1300          * to stop it overwriting our objects */
1301         rv515_vga_render_disable(rdev);
1302 }
1303
1304 /*
1305  * CP.
1306  */
1307 void evergreen_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
1308 {
1309         /* set to DX10/11 mode */
1310         radeon_ring_write(rdev, PACKET3(PACKET3_MODE_CONTROL, 0));
1311         radeon_ring_write(rdev, 1);
1312         /* FIXME: implement */
1313         radeon_ring_write(rdev, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
1314         radeon_ring_write(rdev,
1315 #ifdef __BIG_ENDIAN
1316                           (2 << 0) |
1317 #endif
1318                           (ib->gpu_addr & 0xFFFFFFFC));
1319         radeon_ring_write(rdev, upper_32_bits(ib->gpu_addr) & 0xFF);
1320         radeon_ring_write(rdev, ib->length_dw);
1321 }
1322
1323
1324 static int evergreen_cp_load_microcode(struct radeon_device *rdev)
1325 {
1326         const __be32 *fw_data;
1327         int i;
1328
1329         if (!rdev->me_fw || !rdev->pfp_fw)
1330                 return -EINVAL;
1331
1332         r700_cp_stop(rdev);
1333         WREG32(CP_RB_CNTL,
1334 #ifdef __BIG_ENDIAN
1335                BUF_SWAP_32BIT |
1336 #endif
1337                RB_NO_UPDATE | RB_BLKSZ(15) | RB_BUFSZ(3));
1338
1339         fw_data = (const __be32 *)rdev->pfp_fw->data;
1340         WREG32(CP_PFP_UCODE_ADDR, 0);
1341         for (i = 0; i < EVERGREEN_PFP_UCODE_SIZE; i++)
1342                 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
1343         WREG32(CP_PFP_UCODE_ADDR, 0);
1344
1345         fw_data = (const __be32 *)rdev->me_fw->data;
1346         WREG32(CP_ME_RAM_WADDR, 0);
1347         for (i = 0; i < EVERGREEN_PM4_UCODE_SIZE; i++)
1348                 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
1349
1350         WREG32(CP_PFP_UCODE_ADDR, 0);
1351         WREG32(CP_ME_RAM_WADDR, 0);
1352         WREG32(CP_ME_RAM_RADDR, 0);
1353         return 0;
1354 }
1355
1356 static int evergreen_cp_start(struct radeon_device *rdev)
1357 {
1358         int r, i;
1359         uint32_t cp_me;
1360
1361         r = radeon_ring_lock(rdev, 7);
1362         if (r) {
1363                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
1364                 return r;
1365         }
1366         radeon_ring_write(rdev, PACKET3(PACKET3_ME_INITIALIZE, 5));
1367         radeon_ring_write(rdev, 0x1);
1368         radeon_ring_write(rdev, 0x0);
1369         radeon_ring_write(rdev, rdev->config.evergreen.max_hw_contexts - 1);
1370         radeon_ring_write(rdev, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
1371         radeon_ring_write(rdev, 0);
1372         radeon_ring_write(rdev, 0);
1373         radeon_ring_unlock_commit(rdev);
1374
1375         cp_me = 0xff;
1376         WREG32(CP_ME_CNTL, cp_me);
1377
1378         r = radeon_ring_lock(rdev, evergreen_default_size + 19);
1379         if (r) {
1380                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
1381                 return r;
1382         }
1383
1384         /* setup clear context state */
1385         radeon_ring_write(rdev, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1386         radeon_ring_write(rdev, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
1387
1388         for (i = 0; i < evergreen_default_size; i++)
1389                 radeon_ring_write(rdev, evergreen_default_state[i]);
1390
1391         radeon_ring_write(rdev, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1392         radeon_ring_write(rdev, PACKET3_PREAMBLE_END_CLEAR_STATE);
1393
1394         /* set clear context state */
1395         radeon_ring_write(rdev, PACKET3(PACKET3_CLEAR_STATE, 0));
1396         radeon_ring_write(rdev, 0);
1397
1398         /* SQ_VTX_BASE_VTX_LOC */
1399         radeon_ring_write(rdev, 0xc0026f00);
1400         radeon_ring_write(rdev, 0x00000000);
1401         radeon_ring_write(rdev, 0x00000000);
1402         radeon_ring_write(rdev, 0x00000000);
1403
1404         /* Clear consts */
1405         radeon_ring_write(rdev, 0xc0036f00);
1406         radeon_ring_write(rdev, 0x00000bc4);
1407         radeon_ring_write(rdev, 0xffffffff);
1408         radeon_ring_write(rdev, 0xffffffff);
1409         radeon_ring_write(rdev, 0xffffffff);
1410
1411         radeon_ring_write(rdev, 0xc0026900);
1412         radeon_ring_write(rdev, 0x00000316);
1413         radeon_ring_write(rdev, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
1414         radeon_ring_write(rdev, 0x00000010); /*  */
1415
1416         radeon_ring_unlock_commit(rdev);
1417
1418         return 0;
1419 }
1420
1421 int evergreen_cp_resume(struct radeon_device *rdev)
1422 {
1423         u32 tmp;
1424         u32 rb_bufsz;
1425         int r;
1426
1427         /* Reset cp; if cp is reset, then PA, SH, VGT also need to be reset */
1428         WREG32(GRBM_SOFT_RESET, (SOFT_RESET_CP |
1429                                  SOFT_RESET_PA |
1430                                  SOFT_RESET_SH |
1431                                  SOFT_RESET_VGT |
1432                                  SOFT_RESET_SPI |
1433                                  SOFT_RESET_SX));
1434         RREG32(GRBM_SOFT_RESET);
1435         mdelay(15);
1436         WREG32(GRBM_SOFT_RESET, 0);
1437         RREG32(GRBM_SOFT_RESET);
1438
1439         /* Set ring buffer size */
1440         rb_bufsz = drm_order(rdev->cp.ring_size / 8);
1441         tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
1442 #ifdef __BIG_ENDIAN
1443         tmp |= BUF_SWAP_32BIT;
1444 #endif
1445         WREG32(CP_RB_CNTL, tmp);
1446         WREG32(CP_SEM_WAIT_TIMER, 0x4);
1447
1448         /* Set the write pointer delay */
1449         WREG32(CP_RB_WPTR_DELAY, 0);
1450
1451         /* Initialize the ring buffer's read and write pointers */
1452         WREG32(CP_RB_CNTL, tmp | RB_RPTR_WR_ENA);
1453         WREG32(CP_RB_RPTR_WR, 0);
1454         rdev->cp.wptr = 0;
1455         WREG32(CP_RB_WPTR, rdev->cp.wptr);
1456
1457         /* set the wb address wether it's enabled or not */
1458         WREG32(CP_RB_RPTR_ADDR,
1459                ((rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC));
1460         WREG32(CP_RB_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
1461         WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
1462
1463         if (rdev->wb.enabled)
1464                 WREG32(SCRATCH_UMSK, 0xff);
1465         else {
1466                 tmp |= RB_NO_UPDATE;
1467                 WREG32(SCRATCH_UMSK, 0);
1468         }
1469
1470         mdelay(1);
1471         WREG32(CP_RB_CNTL, tmp);
1472
1473         WREG32(CP_RB_BASE, rdev->cp.gpu_addr >> 8);
1474         WREG32(CP_DEBUG, (1 << 27) | (1 << 28));
1475
1476         rdev->cp.rptr = RREG32(CP_RB_RPTR);
1477
1478         evergreen_cp_start(rdev);
1479         rdev->cp.ready = true;
1480         r = radeon_ring_test(rdev);
1481         if (r) {
1482                 rdev->cp.ready = false;
1483                 return r;
1484         }
1485         return 0;
1486 }
1487
1488 /*
1489  * Core functions
1490  */
1491 static u32 evergreen_get_tile_pipe_to_backend_map(struct radeon_device *rdev,
1492                                                   u32 num_tile_pipes,
1493                                                   u32 num_backends,
1494                                                   u32 backend_disable_mask)
1495 {
1496         u32 backend_map = 0;
1497         u32 enabled_backends_mask = 0;
1498         u32 enabled_backends_count = 0;
1499         u32 cur_pipe;
1500         u32 swizzle_pipe[EVERGREEN_MAX_PIPES];
1501         u32 cur_backend = 0;
1502         u32 i;
1503         bool force_no_swizzle;
1504
1505         if (num_tile_pipes > EVERGREEN_MAX_PIPES)
1506                 num_tile_pipes = EVERGREEN_MAX_PIPES;
1507         if (num_tile_pipes < 1)
1508                 num_tile_pipes = 1;
1509         if (num_backends > EVERGREEN_MAX_BACKENDS)
1510                 num_backends = EVERGREEN_MAX_BACKENDS;
1511         if (num_backends < 1)
1512                 num_backends = 1;
1513
1514         for (i = 0; i < EVERGREEN_MAX_BACKENDS; ++i) {
1515                 if (((backend_disable_mask >> i) & 1) == 0) {
1516                         enabled_backends_mask |= (1 << i);
1517                         ++enabled_backends_count;
1518                 }
1519                 if (enabled_backends_count == num_backends)
1520                         break;
1521         }
1522
1523         if (enabled_backends_count == 0) {
1524                 enabled_backends_mask = 1;
1525                 enabled_backends_count = 1;
1526         }
1527
1528         if (enabled_backends_count != num_backends)
1529                 num_backends = enabled_backends_count;
1530
1531         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * EVERGREEN_MAX_PIPES);
1532         switch (rdev->family) {
1533         case CHIP_CEDAR:
1534         case CHIP_REDWOOD:
1535         case CHIP_PALM:
1536         case CHIP_SUMO:
1537         case CHIP_SUMO2:
1538         case CHIP_TURKS:
1539         case CHIP_CAICOS:
1540                 force_no_swizzle = false;
1541                 break;
1542         case CHIP_CYPRESS:
1543         case CHIP_HEMLOCK:
1544         case CHIP_JUNIPER:
1545         case CHIP_BARTS:
1546         default:
1547                 force_no_swizzle = true;
1548                 break;
1549         }
1550         if (force_no_swizzle) {
1551                 bool last_backend_enabled = false;
1552
1553                 force_no_swizzle = false;
1554                 for (i = 0; i < EVERGREEN_MAX_BACKENDS; ++i) {
1555                         if (((enabled_backends_mask >> i) & 1) == 1) {
1556                                 if (last_backend_enabled)
1557                                         force_no_swizzle = true;
1558                                 last_backend_enabled = true;
1559                         } else
1560                                 last_backend_enabled = false;
1561                 }
1562         }
1563
1564         switch (num_tile_pipes) {
1565         case 1:
1566         case 3:
1567         case 5:
1568         case 7:
1569                 DRM_ERROR("odd number of pipes!\n");
1570                 break;
1571         case 2:
1572                 swizzle_pipe[0] = 0;
1573                 swizzle_pipe[1] = 1;
1574                 break;
1575         case 4:
1576                 if (force_no_swizzle) {
1577                         swizzle_pipe[0] = 0;
1578                         swizzle_pipe[1] = 1;
1579                         swizzle_pipe[2] = 2;
1580                         swizzle_pipe[3] = 3;
1581                 } else {
1582                         swizzle_pipe[0] = 0;
1583                         swizzle_pipe[1] = 2;
1584                         swizzle_pipe[2] = 1;
1585                         swizzle_pipe[3] = 3;
1586                 }
1587                 break;
1588         case 6:
1589                 if (force_no_swizzle) {
1590                         swizzle_pipe[0] = 0;
1591                         swizzle_pipe[1] = 1;
1592                         swizzle_pipe[2] = 2;
1593                         swizzle_pipe[3] = 3;
1594                         swizzle_pipe[4] = 4;
1595                         swizzle_pipe[5] = 5;
1596                 } else {
1597                         swizzle_pipe[0] = 0;
1598                         swizzle_pipe[1] = 2;
1599                         swizzle_pipe[2] = 4;
1600                         swizzle_pipe[3] = 1;
1601                         swizzle_pipe[4] = 3;
1602                         swizzle_pipe[5] = 5;
1603                 }
1604                 break;
1605         case 8:
1606                 if (force_no_swizzle) {
1607                         swizzle_pipe[0] = 0;
1608                         swizzle_pipe[1] = 1;
1609                         swizzle_pipe[2] = 2;
1610                         swizzle_pipe[3] = 3;
1611                         swizzle_pipe[4] = 4;
1612                         swizzle_pipe[5] = 5;
1613                         swizzle_pipe[6] = 6;
1614                         swizzle_pipe[7] = 7;
1615                 } else {
1616                         swizzle_pipe[0] = 0;
1617                         swizzle_pipe[1] = 2;
1618                         swizzle_pipe[2] = 4;
1619                         swizzle_pipe[3] = 6;
1620                         swizzle_pipe[4] = 1;
1621                         swizzle_pipe[5] = 3;
1622                         swizzle_pipe[6] = 5;
1623                         swizzle_pipe[7] = 7;
1624                 }
1625                 break;
1626         }
1627
1628         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
1629                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
1630                         cur_backend = (cur_backend + 1) % EVERGREEN_MAX_BACKENDS;
1631
1632                 backend_map |= (((cur_backend & 0xf) << (swizzle_pipe[cur_pipe] * 4)));
1633
1634                 cur_backend = (cur_backend + 1) % EVERGREEN_MAX_BACKENDS;
1635         }
1636
1637         return backend_map;
1638 }
1639
1640 static void evergreen_gpu_init(struct radeon_device *rdev)
1641 {
1642         u32 cc_rb_backend_disable = 0;
1643         u32 cc_gc_shader_pipe_config;
1644         u32 gb_addr_config = 0;
1645         u32 mc_shared_chmap, mc_arb_ramcfg;
1646         u32 gb_backend_map;
1647         u32 grbm_gfx_index;
1648         u32 sx_debug_1;
1649         u32 smx_dc_ctl0;
1650         u32 sq_config;
1651         u32 sq_lds_resource_mgmt;
1652         u32 sq_gpr_resource_mgmt_1;
1653         u32 sq_gpr_resource_mgmt_2;
1654         u32 sq_gpr_resource_mgmt_3;
1655         u32 sq_thread_resource_mgmt;
1656         u32 sq_thread_resource_mgmt_2;
1657         u32 sq_stack_resource_mgmt_1;
1658         u32 sq_stack_resource_mgmt_2;
1659         u32 sq_stack_resource_mgmt_3;
1660         u32 vgt_cache_invalidation;
1661         u32 hdp_host_path_cntl, tmp;
1662         int i, j, num_shader_engines, ps_thread_count;
1663
1664         switch (rdev->family) {
1665         case CHIP_CYPRESS:
1666         case CHIP_HEMLOCK:
1667                 rdev->config.evergreen.num_ses = 2;
1668                 rdev->config.evergreen.max_pipes = 4;
1669                 rdev->config.evergreen.max_tile_pipes = 8;
1670                 rdev->config.evergreen.max_simds = 10;
1671                 rdev->config.evergreen.max_backends = 4 * rdev->config.evergreen.num_ses;
1672                 rdev->config.evergreen.max_gprs = 256;
1673                 rdev->config.evergreen.max_threads = 248;
1674                 rdev->config.evergreen.max_gs_threads = 32;
1675                 rdev->config.evergreen.max_stack_entries = 512;
1676                 rdev->config.evergreen.sx_num_of_sets = 4;
1677                 rdev->config.evergreen.sx_max_export_size = 256;
1678                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1679                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1680                 rdev->config.evergreen.max_hw_contexts = 8;
1681                 rdev->config.evergreen.sq_num_cf_insts = 2;
1682
1683                 rdev->config.evergreen.sc_prim_fifo_size = 0x100;
1684                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1685                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1686                 break;
1687         case CHIP_JUNIPER:
1688                 rdev->config.evergreen.num_ses = 1;
1689                 rdev->config.evergreen.max_pipes = 4;
1690                 rdev->config.evergreen.max_tile_pipes = 4;
1691                 rdev->config.evergreen.max_simds = 10;
1692                 rdev->config.evergreen.max_backends = 4 * rdev->config.evergreen.num_ses;
1693                 rdev->config.evergreen.max_gprs = 256;
1694                 rdev->config.evergreen.max_threads = 248;
1695                 rdev->config.evergreen.max_gs_threads = 32;
1696                 rdev->config.evergreen.max_stack_entries = 512;
1697                 rdev->config.evergreen.sx_num_of_sets = 4;
1698                 rdev->config.evergreen.sx_max_export_size = 256;
1699                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1700                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1701                 rdev->config.evergreen.max_hw_contexts = 8;
1702                 rdev->config.evergreen.sq_num_cf_insts = 2;
1703
1704                 rdev->config.evergreen.sc_prim_fifo_size = 0x100;
1705                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1706                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1707                 break;
1708         case CHIP_REDWOOD:
1709                 rdev->config.evergreen.num_ses = 1;
1710                 rdev->config.evergreen.max_pipes = 4;
1711                 rdev->config.evergreen.max_tile_pipes = 4;
1712                 rdev->config.evergreen.max_simds = 5;
1713                 rdev->config.evergreen.max_backends = 2 * rdev->config.evergreen.num_ses;
1714                 rdev->config.evergreen.max_gprs = 256;
1715                 rdev->config.evergreen.max_threads = 248;
1716                 rdev->config.evergreen.max_gs_threads = 32;
1717                 rdev->config.evergreen.max_stack_entries = 256;
1718                 rdev->config.evergreen.sx_num_of_sets = 4;
1719                 rdev->config.evergreen.sx_max_export_size = 256;
1720                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1721                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1722                 rdev->config.evergreen.max_hw_contexts = 8;
1723                 rdev->config.evergreen.sq_num_cf_insts = 2;
1724
1725                 rdev->config.evergreen.sc_prim_fifo_size = 0x100;
1726                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1727                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1728                 break;
1729         case CHIP_CEDAR:
1730         default:
1731                 rdev->config.evergreen.num_ses = 1;
1732                 rdev->config.evergreen.max_pipes = 2;
1733                 rdev->config.evergreen.max_tile_pipes = 2;
1734                 rdev->config.evergreen.max_simds = 2;
1735                 rdev->config.evergreen.max_backends = 1 * rdev->config.evergreen.num_ses;
1736                 rdev->config.evergreen.max_gprs = 256;
1737                 rdev->config.evergreen.max_threads = 192;
1738                 rdev->config.evergreen.max_gs_threads = 16;
1739                 rdev->config.evergreen.max_stack_entries = 256;
1740                 rdev->config.evergreen.sx_num_of_sets = 4;
1741                 rdev->config.evergreen.sx_max_export_size = 128;
1742                 rdev->config.evergreen.sx_max_export_pos_size = 32;
1743                 rdev->config.evergreen.sx_max_export_smx_size = 96;
1744                 rdev->config.evergreen.max_hw_contexts = 4;
1745                 rdev->config.evergreen.sq_num_cf_insts = 1;
1746
1747                 rdev->config.evergreen.sc_prim_fifo_size = 0x40;
1748                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1749                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1750                 break;
1751         case CHIP_PALM:
1752                 rdev->config.evergreen.num_ses = 1;
1753                 rdev->config.evergreen.max_pipes = 2;
1754                 rdev->config.evergreen.max_tile_pipes = 2;
1755                 rdev->config.evergreen.max_simds = 2;
1756                 rdev->config.evergreen.max_backends = 1 * rdev->config.evergreen.num_ses;
1757                 rdev->config.evergreen.max_gprs = 256;
1758                 rdev->config.evergreen.max_threads = 192;
1759                 rdev->config.evergreen.max_gs_threads = 16;
1760                 rdev->config.evergreen.max_stack_entries = 256;
1761                 rdev->config.evergreen.sx_num_of_sets = 4;
1762                 rdev->config.evergreen.sx_max_export_size = 128;
1763                 rdev->config.evergreen.sx_max_export_pos_size = 32;
1764                 rdev->config.evergreen.sx_max_export_smx_size = 96;
1765                 rdev->config.evergreen.max_hw_contexts = 4;
1766                 rdev->config.evergreen.sq_num_cf_insts = 1;
1767
1768                 rdev->config.evergreen.sc_prim_fifo_size = 0x40;
1769                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1770                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1771                 break;
1772         case CHIP_SUMO:
1773                 rdev->config.evergreen.num_ses = 1;
1774                 rdev->config.evergreen.max_pipes = 4;
1775                 rdev->config.evergreen.max_tile_pipes = 2;
1776                 if (rdev->pdev->device == 0x9648)
1777                         rdev->config.evergreen.max_simds = 3;
1778                 else if ((rdev->pdev->device == 0x9647) ||
1779                          (rdev->pdev->device == 0x964a))
1780                         rdev->config.evergreen.max_simds = 4;
1781                 else
1782                         rdev->config.evergreen.max_simds = 5;
1783                 rdev->config.evergreen.max_backends = 2 * rdev->config.evergreen.num_ses;
1784                 rdev->config.evergreen.max_gprs = 256;
1785                 rdev->config.evergreen.max_threads = 248;
1786                 rdev->config.evergreen.max_gs_threads = 32;
1787                 rdev->config.evergreen.max_stack_entries = 256;
1788                 rdev->config.evergreen.sx_num_of_sets = 4;
1789                 rdev->config.evergreen.sx_max_export_size = 256;
1790                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1791                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1792                 rdev->config.evergreen.max_hw_contexts = 8;
1793                 rdev->config.evergreen.sq_num_cf_insts = 2;
1794
1795                 rdev->config.evergreen.sc_prim_fifo_size = 0x40;
1796                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1797                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1798                 break;
1799         case CHIP_SUMO2:
1800                 rdev->config.evergreen.num_ses = 1;
1801                 rdev->config.evergreen.max_pipes = 4;
1802                 rdev->config.evergreen.max_tile_pipes = 4;
1803                 rdev->config.evergreen.max_simds = 2;
1804                 rdev->config.evergreen.max_backends = 1 * rdev->config.evergreen.num_ses;
1805                 rdev->config.evergreen.max_gprs = 256;
1806                 rdev->config.evergreen.max_threads = 248;
1807                 rdev->config.evergreen.max_gs_threads = 32;
1808                 rdev->config.evergreen.max_stack_entries = 512;
1809                 rdev->config.evergreen.sx_num_of_sets = 4;
1810                 rdev->config.evergreen.sx_max_export_size = 256;
1811                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1812                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1813                 rdev->config.evergreen.max_hw_contexts = 8;
1814                 rdev->config.evergreen.sq_num_cf_insts = 2;
1815
1816                 rdev->config.evergreen.sc_prim_fifo_size = 0x40;
1817                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1818                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1819                 break;
1820         case CHIP_BARTS:
1821                 rdev->config.evergreen.num_ses = 2;
1822                 rdev->config.evergreen.max_pipes = 4;
1823                 rdev->config.evergreen.max_tile_pipes = 8;
1824                 rdev->config.evergreen.max_simds = 7;
1825                 rdev->config.evergreen.max_backends = 4 * rdev->config.evergreen.num_ses;
1826                 rdev->config.evergreen.max_gprs = 256;
1827                 rdev->config.evergreen.max_threads = 248;
1828                 rdev->config.evergreen.max_gs_threads = 32;
1829                 rdev->config.evergreen.max_stack_entries = 512;
1830                 rdev->config.evergreen.sx_num_of_sets = 4;
1831                 rdev->config.evergreen.sx_max_export_size = 256;
1832                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1833                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1834                 rdev->config.evergreen.max_hw_contexts = 8;
1835                 rdev->config.evergreen.sq_num_cf_insts = 2;
1836
1837                 rdev->config.evergreen.sc_prim_fifo_size = 0x100;
1838                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1839                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1840                 break;
1841         case CHIP_TURKS:
1842                 rdev->config.evergreen.num_ses = 1;
1843                 rdev->config.evergreen.max_pipes = 4;
1844                 rdev->config.evergreen.max_tile_pipes = 4;
1845                 rdev->config.evergreen.max_simds = 6;
1846                 rdev->config.evergreen.max_backends = 2 * rdev->config.evergreen.num_ses;
1847                 rdev->config.evergreen.max_gprs = 256;
1848                 rdev->config.evergreen.max_threads = 248;
1849                 rdev->config.evergreen.max_gs_threads = 32;
1850                 rdev->config.evergreen.max_stack_entries = 256;
1851                 rdev->config.evergreen.sx_num_of_sets = 4;
1852                 rdev->config.evergreen.sx_max_export_size = 256;
1853                 rdev->config.evergreen.sx_max_export_pos_size = 64;
1854                 rdev->config.evergreen.sx_max_export_smx_size = 192;
1855                 rdev->config.evergreen.max_hw_contexts = 8;
1856                 rdev->config.evergreen.sq_num_cf_insts = 2;
1857
1858                 rdev->config.evergreen.sc_prim_fifo_size = 0x100;
1859                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1860                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1861                 break;
1862         case CHIP_CAICOS:
1863                 rdev->config.evergreen.num_ses = 1;
1864                 rdev->config.evergreen.max_pipes = 4;
1865                 rdev->config.evergreen.max_tile_pipes = 2;
1866                 rdev->config.evergreen.max_simds = 2;
1867                 rdev->config.evergreen.max_backends = 1 * rdev->config.evergreen.num_ses;
1868                 rdev->config.evergreen.max_gprs = 256;
1869                 rdev->config.evergreen.max_threads = 192;
1870                 rdev->config.evergreen.max_gs_threads = 16;
1871                 rdev->config.evergreen.max_stack_entries = 256;
1872                 rdev->config.evergreen.sx_num_of_sets = 4;
1873                 rdev->config.evergreen.sx_max_export_size = 128;
1874                 rdev->config.evergreen.sx_max_export_pos_size = 32;
1875                 rdev->config.evergreen.sx_max_export_smx_size = 96;
1876                 rdev->config.evergreen.max_hw_contexts = 4;
1877                 rdev->config.evergreen.sq_num_cf_insts = 1;
1878
1879                 rdev->config.evergreen.sc_prim_fifo_size = 0x40;
1880                 rdev->config.evergreen.sc_hiz_tile_fifo_size = 0x30;
1881                 rdev->config.evergreen.sc_earlyz_tile_fifo_size = 0x130;
1882                 break;
1883         }
1884
1885         /* Initialize HDP */
1886         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
1887                 WREG32((0x2c14 + j), 0x00000000);
1888                 WREG32((0x2c18 + j), 0x00000000);
1889                 WREG32((0x2c1c + j), 0x00000000);
1890                 WREG32((0x2c20 + j), 0x00000000);
1891                 WREG32((0x2c24 + j), 0x00000000);
1892         }
1893
1894         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
1895
1896         evergreen_fix_pci_max_read_req_size(rdev);
1897
1898         cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & ~2;
1899
1900         cc_gc_shader_pipe_config |=
1901                 INACTIVE_QD_PIPES((EVERGREEN_MAX_PIPES_MASK << rdev->config.evergreen.max_pipes)
1902                                   & EVERGREEN_MAX_PIPES_MASK);
1903         cc_gc_shader_pipe_config |=
1904                 INACTIVE_SIMDS((EVERGREEN_MAX_SIMDS_MASK << rdev->config.evergreen.max_simds)
1905                                & EVERGREEN_MAX_SIMDS_MASK);
1906
1907         cc_rb_backend_disable =
1908                 BACKEND_DISABLE((EVERGREEN_MAX_BACKENDS_MASK << rdev->config.evergreen.max_backends)
1909                                 & EVERGREEN_MAX_BACKENDS_MASK);
1910
1911
1912         mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
1913         if (rdev->flags & RADEON_IS_IGP)
1914                 mc_arb_ramcfg = RREG32(FUS_MC_ARB_RAMCFG);
1915         else
1916                 mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
1917
1918         switch (rdev->config.evergreen.max_tile_pipes) {
1919         case 1:
1920         default:
1921                 gb_addr_config |= NUM_PIPES(0);
1922                 break;
1923         case 2:
1924                 gb_addr_config |= NUM_PIPES(1);
1925                 break;
1926         case 4:
1927                 gb_addr_config |= NUM_PIPES(2);
1928                 break;
1929         case 8:
1930                 gb_addr_config |= NUM_PIPES(3);
1931                 break;
1932         }
1933
1934         gb_addr_config |= PIPE_INTERLEAVE_SIZE((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT);
1935         gb_addr_config |= BANK_INTERLEAVE_SIZE(0);
1936         gb_addr_config |= NUM_SHADER_ENGINES(rdev->config.evergreen.num_ses - 1);
1937         gb_addr_config |= SHADER_ENGINE_TILE_SIZE(1);
1938         gb_addr_config |= NUM_GPUS(0); /* Hemlock? */
1939         gb_addr_config |= MULTI_GPU_TILE_SIZE(2);
1940
1941         if (((mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT) > 2)
1942                 gb_addr_config |= ROW_SIZE(2);
1943         else
1944                 gb_addr_config |= ROW_SIZE((mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT);
1945
1946         if (rdev->ddev->pdev->device == 0x689e) {
1947                 u32 efuse_straps_4;
1948                 u32 efuse_straps_3;
1949                 u8 efuse_box_bit_131_124;
1950
1951                 WREG32(RCU_IND_INDEX, 0x204);
1952                 efuse_straps_4 = RREG32(RCU_IND_DATA);
1953                 WREG32(RCU_IND_INDEX, 0x203);
1954                 efuse_straps_3 = RREG32(RCU_IND_DATA);
1955                 efuse_box_bit_131_124 = (u8)(((efuse_straps_4 & 0xf) << 4) | ((efuse_straps_3 & 0xf0000000) >> 28));
1956
1957                 switch(efuse_box_bit_131_124) {
1958                 case 0x00:
1959                         gb_backend_map = 0x76543210;
1960                         break;
1961                 case 0x55:
1962                         gb_backend_map = 0x77553311;
1963                         break;
1964                 case 0x56:
1965                         gb_backend_map = 0x77553300;
1966                         break;
1967                 case 0x59:
1968                         gb_backend_map = 0x77552211;
1969                         break;
1970                 case 0x66:
1971                         gb_backend_map = 0x77443300;
1972                         break;
1973                 case 0x99:
1974                         gb_backend_map = 0x66552211;
1975                         break;
1976                 case 0x5a:
1977                         gb_backend_map = 0x77552200;
1978                         break;
1979                 case 0xaa:
1980                         gb_backend_map = 0x66442200;
1981                         break;
1982                 case 0x95:
1983                         gb_backend_map = 0x66553311;
1984                         break;
1985                 default:
1986                         DRM_ERROR("bad backend map, using default\n");
1987                         gb_backend_map =
1988                                 evergreen_get_tile_pipe_to_backend_map(rdev,
1989                                                                        rdev->config.evergreen.max_tile_pipes,
1990                                                                        rdev->config.evergreen.max_backends,
1991                                                                        ((EVERGREEN_MAX_BACKENDS_MASK <<
1992                                                                    rdev->config.evergreen.max_backends) &
1993                                                                         EVERGREEN_MAX_BACKENDS_MASK));
1994                         break;
1995                 }
1996         } else if (rdev->ddev->pdev->device == 0x68b9) {
1997                 u32 efuse_straps_3;
1998                 u8 efuse_box_bit_127_124;
1999
2000                 WREG32(RCU_IND_INDEX, 0x203);
2001                 efuse_straps_3 = RREG32(RCU_IND_DATA);
2002                 efuse_box_bit_127_124 = (u8)((efuse_straps_3 & 0xF0000000) >> 28);
2003
2004                 switch(efuse_box_bit_127_124) {
2005                 case 0x0:
2006                         gb_backend_map = 0x00003210;
2007                         break;
2008                 case 0x5:
2009                 case 0x6:
2010                 case 0x9:
2011                 case 0xa:
2012                         gb_backend_map = 0x00003311;
2013                         break;
2014                 default:
2015                         DRM_ERROR("bad backend map, using default\n");
2016                         gb_backend_map =
2017                                 evergreen_get_tile_pipe_to_backend_map(rdev,
2018                                                                        rdev->config.evergreen.max_tile_pipes,
2019                                                                        rdev->config.evergreen.max_backends,
2020                                                                        ((EVERGREEN_MAX_BACKENDS_MASK <<
2021                                                                    rdev->config.evergreen.max_backends) &
2022                                                                         EVERGREEN_MAX_BACKENDS_MASK));
2023                         break;
2024                 }
2025         } else {
2026                 switch (rdev->family) {
2027                 case CHIP_CYPRESS:
2028                 case CHIP_HEMLOCK:
2029                 case CHIP_BARTS:
2030                         gb_backend_map = 0x66442200;
2031                         break;
2032                 case CHIP_JUNIPER:
2033                         gb_backend_map = 0x00002200;
2034                         break;
2035                 default:
2036                         gb_backend_map =
2037                                 evergreen_get_tile_pipe_to_backend_map(rdev,
2038                                                                        rdev->config.evergreen.max_tile_pipes,
2039                                                                        rdev->config.evergreen.max_backends,
2040                                                                        ((EVERGREEN_MAX_BACKENDS_MASK <<
2041                                                                          rdev->config.evergreen.max_backends) &
2042                                                                         EVERGREEN_MAX_BACKENDS_MASK));
2043                 }
2044         }
2045
2046         /* setup tiling info dword.  gb_addr_config is not adequate since it does
2047          * not have bank info, so create a custom tiling dword.
2048          * bits 3:0   num_pipes
2049          * bits 7:4   num_banks
2050          * bits 11:8  group_size
2051          * bits 15:12 row_size
2052          */
2053         rdev->config.evergreen.tile_config = 0;
2054         switch (rdev->config.evergreen.max_tile_pipes) {
2055         case 1:
2056         default:
2057                 rdev->config.evergreen.tile_config |= (0 << 0);
2058                 break;
2059         case 2:
2060                 rdev->config.evergreen.tile_config |= (1 << 0);
2061                 break;
2062         case 4:
2063                 rdev->config.evergreen.tile_config |= (2 << 0);
2064                 break;
2065         case 8:
2066                 rdev->config.evergreen.tile_config |= (3 << 0);
2067                 break;
2068         }
2069         /* num banks is 8 on all fusion asics. 0 = 4, 1 = 8, 2 = 16 */
2070         if (rdev->flags & RADEON_IS_IGP)
2071                 rdev->config.evergreen.tile_config |= 1 << 4;
2072         else
2073                 rdev->config.evergreen.tile_config |=
2074                         ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) << 4;
2075         rdev->config.evergreen.tile_config |=
2076                 ((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT) << 8;
2077         rdev->config.evergreen.tile_config |=
2078                 ((gb_addr_config & 0x30000000) >> 28) << 12;
2079
2080         rdev->config.evergreen.backend_map = gb_backend_map;
2081         WREG32(GB_BACKEND_MAP, gb_backend_map);
2082         WREG32(GB_ADDR_CONFIG, gb_addr_config);
2083         WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
2084         WREG32(HDP_ADDR_CONFIG, gb_addr_config);
2085
2086         num_shader_engines = ((RREG32(GB_ADDR_CONFIG) & NUM_SHADER_ENGINES(3)) >> 12) + 1;
2087         grbm_gfx_index = INSTANCE_BROADCAST_WRITES;
2088
2089         for (i = 0; i < rdev->config.evergreen.num_ses; i++) {
2090                 u32 rb = cc_rb_backend_disable | (0xf0 << 16);
2091                 u32 sp = cc_gc_shader_pipe_config;
2092                 u32 gfx = grbm_gfx_index | SE_INDEX(i);
2093
2094                 if (i == num_shader_engines) {
2095                         rb |= BACKEND_DISABLE(EVERGREEN_MAX_BACKENDS_MASK);
2096                         sp |= INACTIVE_SIMDS(EVERGREEN_MAX_SIMDS_MASK);
2097                 }
2098
2099                 WREG32(GRBM_GFX_INDEX, gfx);
2100                 WREG32(RLC_GFX_INDEX, gfx);
2101
2102                 WREG32(CC_RB_BACKEND_DISABLE, rb);
2103                 WREG32(CC_SYS_RB_BACKEND_DISABLE, rb);
2104                 WREG32(GC_USER_RB_BACKEND_DISABLE, rb);
2105                 WREG32(CC_GC_SHADER_PIPE_CONFIG, sp);
2106         }
2107
2108         grbm_gfx_index |= SE_BROADCAST_WRITES;
2109         WREG32(GRBM_GFX_INDEX, grbm_gfx_index);
2110         WREG32(RLC_GFX_INDEX, grbm_gfx_index);
2111
2112         WREG32(CGTS_SYS_TCC_DISABLE, 0);
2113         WREG32(CGTS_TCC_DISABLE, 0);
2114         WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
2115         WREG32(CGTS_USER_TCC_DISABLE, 0);
2116
2117         /* set HW defaults for 3D engine */
2118         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
2119                                      ROQ_IB2_START(0x2b)));
2120
2121         WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
2122
2123         WREG32(TA_CNTL_AUX, (DISABLE_CUBE_ANISO |
2124                              SYNC_GRADIENT |
2125                              SYNC_WALKER |
2126                              SYNC_ALIGNER));
2127
2128         sx_debug_1 = RREG32(SX_DEBUG_1);
2129         sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
2130         WREG32(SX_DEBUG_1, sx_debug_1);
2131
2132
2133         smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
2134         smx_dc_ctl0 &= ~NUMBER_OF_SETS(0x1ff);
2135         smx_dc_ctl0 |= NUMBER_OF_SETS(rdev->config.evergreen.sx_num_of_sets);
2136         WREG32(SMX_DC_CTL0, smx_dc_ctl0);
2137
2138         WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.evergreen.sx_max_export_size / 4) - 1) |
2139                                         POSITION_BUFFER_SIZE((rdev->config.evergreen.sx_max_export_pos_size / 4) - 1) |
2140                                         SMX_BUFFER_SIZE((rdev->config.evergreen.sx_max_export_smx_size / 4) - 1)));
2141
2142         WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.evergreen.sc_prim_fifo_size) |
2143                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.evergreen.sc_hiz_tile_fifo_size) |
2144                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.evergreen.sc_earlyz_tile_fifo_size)));
2145
2146         WREG32(VGT_NUM_INSTANCES, 1);
2147         WREG32(SPI_CONFIG_CNTL, 0);
2148         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
2149         WREG32(CP_PERFMON_CNTL, 0);
2150
2151         WREG32(SQ_MS_FIFO_SIZES, (CACHE_FIFO_SIZE(16 * rdev->config.evergreen.sq_num_cf_insts) |
2152                                   FETCH_FIFO_HIWATER(0x4) |
2153                                   DONE_FIFO_HIWATER(0xe0) |
2154                                   ALU_UPDATE_FIFO_HIWATER(0x8)));
2155
2156         sq_config = RREG32(SQ_CONFIG);
2157         sq_config &= ~(PS_PRIO(3) |
2158                        VS_PRIO(3) |
2159                        GS_PRIO(3) |
2160                        ES_PRIO(3));
2161         sq_config |= (VC_ENABLE |
2162                       EXPORT_SRC_C |
2163                       PS_PRIO(0) |
2164                       VS_PRIO(1) |
2165                       GS_PRIO(2) |
2166                       ES_PRIO(3));
2167
2168         switch (rdev->family) {
2169         case CHIP_CEDAR:
2170         case CHIP_PALM:
2171         case CHIP_SUMO:
2172         case CHIP_SUMO2:
2173         case CHIP_CAICOS:
2174                 /* no vertex cache */
2175                 sq_config &= ~VC_ENABLE;
2176                 break;
2177         default:
2178                 break;
2179         }
2180
2181         sq_lds_resource_mgmt = RREG32(SQ_LDS_RESOURCE_MGMT);
2182
2183         sq_gpr_resource_mgmt_1 = NUM_PS_GPRS((rdev->config.evergreen.max_gprs - (4 * 2))* 12 / 32);
2184         sq_gpr_resource_mgmt_1 |= NUM_VS_GPRS((rdev->config.evergreen.max_gprs - (4 * 2)) * 6 / 32);
2185         sq_gpr_resource_mgmt_1 |= NUM_CLAUSE_TEMP_GPRS(4);
2186         sq_gpr_resource_mgmt_2 = NUM_GS_GPRS((rdev->config.evergreen.max_gprs - (4 * 2)) * 4 / 32);
2187         sq_gpr_resource_mgmt_2 |= NUM_ES_GPRS((rdev->config.evergreen.max_gprs - (4 * 2)) * 4 / 32);
2188         sq_gpr_resource_mgmt_3 = NUM_HS_GPRS((rdev->config.evergreen.max_gprs - (4 * 2)) * 3 / 32);
2189         sq_gpr_resource_mgmt_3 |= NUM_LS_GPRS((rdev->config.evergreen.max_gprs - (4 * 2)) * 3 / 32);
2190
2191         switch (rdev->family) {
2192         case CHIP_CEDAR:
2193         case CHIP_PALM:
2194         case CHIP_SUMO:
2195         case CHIP_SUMO2:
2196                 ps_thread_count = 96;
2197                 break;
2198         default:
2199                 ps_thread_count = 128;
2200                 break;
2201         }
2202
2203         sq_thread_resource_mgmt = NUM_PS_THREADS(ps_thread_count);
2204         sq_thread_resource_mgmt |= NUM_VS_THREADS((((rdev->config.evergreen.max_threads - ps_thread_count) / 6) / 8) * 8);
2205         sq_thread_resource_mgmt |= NUM_GS_THREADS((((rdev->config.evergreen.max_threads - ps_thread_count) / 6) / 8) * 8);
2206         sq_thread_resource_mgmt |= NUM_ES_THREADS((((rdev->config.evergreen.max_threads - ps_thread_count) / 6) / 8) * 8);
2207         sq_thread_resource_mgmt_2 = NUM_HS_THREADS((((rdev->config.evergreen.max_threads - ps_thread_count) / 6) / 8) * 8);
2208         sq_thread_resource_mgmt_2 |= NUM_LS_THREADS((((rdev->config.evergreen.max_threads - ps_thread_count) / 6) / 8) * 8);
2209
2210         sq_stack_resource_mgmt_1 = NUM_PS_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2211         sq_stack_resource_mgmt_1 |= NUM_VS_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2212         sq_stack_resource_mgmt_2 = NUM_GS_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2213         sq_stack_resource_mgmt_2 |= NUM_ES_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2214         sq_stack_resource_mgmt_3 = NUM_HS_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2215         sq_stack_resource_mgmt_3 |= NUM_LS_STACK_ENTRIES((rdev->config.evergreen.max_stack_entries * 1) / 6);
2216
2217         WREG32(SQ_CONFIG, sq_config);
2218         WREG32(SQ_GPR_RESOURCE_MGMT_1, sq_gpr_resource_mgmt_1);
2219         WREG32(SQ_GPR_RESOURCE_MGMT_2, sq_gpr_resource_mgmt_2);
2220         WREG32(SQ_GPR_RESOURCE_MGMT_3, sq_gpr_resource_mgmt_3);
2221         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
2222         WREG32(SQ_THREAD_RESOURCE_MGMT_2, sq_thread_resource_mgmt_2);
2223         WREG32(SQ_STACK_RESOURCE_MGMT_1, sq_stack_resource_mgmt_1);
2224         WREG32(SQ_STACK_RESOURCE_MGMT_2, sq_stack_resource_mgmt_2);
2225         WREG32(SQ_STACK_RESOURCE_MGMT_3, sq_stack_resource_mgmt_3);
2226         WREG32(SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, 0);
2227         WREG32(SQ_LDS_RESOURCE_MGMT, sq_lds_resource_mgmt);
2228
2229         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
2230                                           FORCE_EOV_MAX_REZ_CNT(255)));
2231
2232         switch (rdev->family) {
2233         case CHIP_CEDAR:
2234         case CHIP_PALM:
2235         case CHIP_SUMO:
2236         case CHIP_SUMO2:
2237         case CHIP_CAICOS:
2238                 vgt_cache_invalidation = CACHE_INVALIDATION(TC_ONLY);
2239                 break;
2240         default:
2241                 vgt_cache_invalidation = CACHE_INVALIDATION(VC_AND_TC);
2242                 break;
2243         }
2244         vgt_cache_invalidation |= AUTO_INVLD_EN(ES_AND_GS_AUTO);
2245         WREG32(VGT_CACHE_INVALIDATION, vgt_cache_invalidation);
2246
2247         WREG32(VGT_GS_VERTEX_REUSE, 16);
2248         WREG32(PA_SU_LINE_STIPPLE_VALUE, 0);
2249         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
2250
2251         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, 14);
2252         WREG32(VGT_OUT_DEALLOC_CNTL, 16);
2253
2254         WREG32(CB_PERF_CTR0_SEL_0, 0);
2255         WREG32(CB_PERF_CTR0_SEL_1, 0);
2256         WREG32(CB_PERF_CTR1_SEL_0, 0);
2257         WREG32(CB_PERF_CTR1_SEL_1, 0);
2258         WREG32(CB_PERF_CTR2_SEL_0, 0);
2259         WREG32(CB_PERF_CTR2_SEL_1, 0);
2260         WREG32(CB_PERF_CTR3_SEL_0, 0);
2261         WREG32(CB_PERF_CTR3_SEL_1, 0);
2262
2263         /* clear render buffer base addresses */
2264         WREG32(CB_COLOR0_BASE, 0);
2265         WREG32(CB_COLOR1_BASE, 0);
2266         WREG32(CB_COLOR2_BASE, 0);
2267         WREG32(CB_COLOR3_BASE, 0);
2268         WREG32(CB_COLOR4_BASE, 0);
2269         WREG32(CB_COLOR5_BASE, 0);
2270         WREG32(CB_COLOR6_BASE, 0);
2271         WREG32(CB_COLOR7_BASE, 0);
2272         WREG32(CB_COLOR8_BASE, 0);
2273         WREG32(CB_COLOR9_BASE, 0);
2274         WREG32(CB_COLOR10_BASE, 0);
2275         WREG32(CB_COLOR11_BASE, 0);
2276
2277         /* set the shader const cache sizes to 0 */
2278         for (i = SQ_ALU_CONST_BUFFER_SIZE_PS_0; i < 0x28200; i += 4)
2279                 WREG32(i, 0);
2280         for (i = SQ_ALU_CONST_BUFFER_SIZE_HS_0; i < 0x29000; i += 4)
2281                 WREG32(i, 0);
2282
2283         tmp = RREG32(HDP_MISC_CNTL);
2284         tmp |= HDP_FLUSH_INVALIDATE_CACHE;
2285         WREG32(HDP_MISC_CNTL, tmp);
2286
2287         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
2288         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
2289
2290         WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
2291
2292         udelay(50);
2293
2294 }
2295
2296 int evergreen_mc_init(struct radeon_device *rdev)
2297 {
2298         u32 tmp;
2299         int chansize, numchan;
2300
2301         /* Get VRAM informations */
2302         rdev->mc.vram_is_ddr = true;
2303         if (rdev->flags & RADEON_IS_IGP)
2304                 tmp = RREG32(FUS_MC_ARB_RAMCFG);
2305         else
2306                 tmp = RREG32(MC_ARB_RAMCFG);
2307         if (tmp & CHANSIZE_OVERRIDE) {
2308                 chansize = 16;
2309         } else if (tmp & CHANSIZE_MASK) {
2310                 chansize = 64;
2311         } else {
2312                 chansize = 32;
2313         }
2314         tmp = RREG32(MC_SHARED_CHMAP);
2315         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
2316         case 0:
2317         default:
2318                 numchan = 1;
2319                 break;
2320         case 1:
2321                 numchan = 2;
2322                 break;
2323         case 2:
2324                 numchan = 4;
2325                 break;
2326         case 3:
2327                 numchan = 8;
2328                 break;
2329         }
2330         rdev->mc.vram_width = numchan * chansize;
2331         /* Could aper size report 0 ? */
2332         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
2333         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
2334         /* Setup GPU memory space */
2335         if (rdev->flags & RADEON_IS_IGP) {
2336                 /* size in bytes on fusion */
2337                 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
2338                 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
2339         } else {
2340                 /* size in MB on evergreen */
2341                 rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
2342                 rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
2343         }
2344         rdev->mc.visible_vram_size = rdev->mc.aper_size;
2345         r700_vram_gtt_location(rdev, &rdev->mc);
2346         radeon_update_bandwidth_info(rdev);
2347
2348         return 0;
2349 }
2350
2351 bool evergreen_gpu_is_lockup(struct radeon_device *rdev)
2352 {
2353         u32 srbm_status;
2354         u32 grbm_status;
2355         u32 grbm_status_se0, grbm_status_se1;
2356         struct r100_gpu_lockup *lockup = &rdev->config.evergreen.lockup;
2357         int r;
2358
2359         srbm_status = RREG32(SRBM_STATUS);
2360         grbm_status = RREG32(GRBM_STATUS);
2361         grbm_status_se0 = RREG32(GRBM_STATUS_SE0);
2362         grbm_status_se1 = RREG32(GRBM_STATUS_SE1);
2363         if (!(grbm_status & GUI_ACTIVE)) {
2364                 r100_gpu_lockup_update(lockup, &rdev->cp);
2365                 return false;
2366         }
2367         /* force CP activities */
2368         r = radeon_ring_lock(rdev, 2);
2369         if (!r) {
2370                 /* PACKET2 NOP */
2371                 radeon_ring_write(rdev, 0x80000000);
2372                 radeon_ring_write(rdev, 0x80000000);
2373                 radeon_ring_unlock_commit(rdev);
2374         }
2375         rdev->cp.rptr = RREG32(CP_RB_RPTR);
2376         return r100_gpu_cp_is_lockup(rdev, lockup, &rdev->cp);
2377 }
2378
2379 static int evergreen_gpu_soft_reset(struct radeon_device *rdev)
2380 {
2381         struct evergreen_mc_save save;
2382         u32 grbm_reset = 0;
2383
2384         if (!(RREG32(GRBM_STATUS) & GUI_ACTIVE))
2385                 return 0;
2386
2387         dev_info(rdev->dev, "GPU softreset \n");
2388         dev_info(rdev->dev, "  GRBM_STATUS=0x%08X\n",
2389                 RREG32(GRBM_STATUS));
2390         dev_info(rdev->dev, "  GRBM_STATUS_SE0=0x%08X\n",
2391                 RREG32(GRBM_STATUS_SE0));
2392         dev_info(rdev->dev, "  GRBM_STATUS_SE1=0x%08X\n",
2393                 RREG32(GRBM_STATUS_SE1));
2394         dev_info(rdev->dev, "  SRBM_STATUS=0x%08X\n",
2395                 RREG32(SRBM_STATUS));
2396         evergreen_mc_stop(rdev, &save);
2397         if (evergreen_mc_wait_for_idle(rdev)) {
2398                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
2399         }
2400         /* Disable CP parsing/prefetching */
2401         WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT);
2402
2403         /* reset all the gfx blocks */
2404         grbm_reset = (SOFT_RESET_CP |
2405                       SOFT_RESET_CB |
2406                       SOFT_RESET_DB |
2407                       SOFT_RESET_PA |
2408                       SOFT_RESET_SC |
2409                       SOFT_RESET_SPI |
2410                       SOFT_RESET_SH |
2411                       SOFT_RESET_SX |
2412                       SOFT_RESET_TC |
2413                       SOFT_RESET_TA |
2414                       SOFT_RESET_VC |
2415                       SOFT_RESET_VGT);
2416
2417         dev_info(rdev->dev, "  GRBM_SOFT_RESET=0x%08X\n", grbm_reset);
2418         WREG32(GRBM_SOFT_RESET, grbm_reset);
2419         (void)RREG32(GRBM_SOFT_RESET);
2420         udelay(50);
2421         WREG32(GRBM_SOFT_RESET, 0);
2422         (void)RREG32(GRBM_SOFT_RESET);
2423         /* Wait a little for things to settle down */
2424         udelay(50);
2425         dev_info(rdev->dev, "  GRBM_STATUS=0x%08X\n",
2426                 RREG32(GRBM_STATUS));
2427         dev_info(rdev->dev, "  GRBM_STATUS_SE0=0x%08X\n",
2428                 RREG32(GRBM_STATUS_SE0));
2429         dev_info(rdev->dev, "  GRBM_STATUS_SE1=0x%08X\n",
2430                 RREG32(GRBM_STATUS_SE1));
2431         dev_info(rdev->dev, "  SRBM_STATUS=0x%08X\n",
2432                 RREG32(SRBM_STATUS));
2433         evergreen_mc_resume(rdev, &save);
2434         return 0;
2435 }
2436
2437 int evergreen_asic_reset(struct radeon_device *rdev)
2438 {
2439         return evergreen_gpu_soft_reset(rdev);
2440 }
2441
2442 /* Interrupts */
2443
2444 u32 evergreen_get_vblank_counter(struct radeon_device *rdev, int crtc)
2445 {
2446         switch (crtc) {
2447         case 0:
2448                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC0_REGISTER_OFFSET);
2449         case 1:
2450                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC1_REGISTER_OFFSET);
2451         case 2:
2452                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC2_REGISTER_OFFSET);
2453         case 3:
2454                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC3_REGISTER_OFFSET);
2455         case 4:
2456                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC4_REGISTER_OFFSET);
2457         case 5:
2458                 return RREG32(CRTC_STATUS_FRAME_COUNT + EVERGREEN_CRTC5_REGISTER_OFFSET);
2459         default:
2460                 return 0;
2461         }
2462 }
2463
2464 void evergreen_disable_interrupt_state(struct radeon_device *rdev)
2465 {
2466         u32 tmp;
2467
2468         WREG32(CP_INT_CNTL, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
2469         WREG32(GRBM_INT_CNTL, 0);
2470         WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
2471         WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
2472         if (rdev->num_crtc >= 4) {
2473                 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
2474                 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
2475         }
2476         if (rdev->num_crtc >= 6) {
2477                 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
2478                 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
2479         }
2480
2481         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
2482         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
2483         if (rdev->num_crtc >= 4) {
2484                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
2485                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
2486         }
2487         if (rdev->num_crtc >= 6) {
2488                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
2489                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
2490         }
2491
2492         WREG32(DACA_AUTODETECT_INT_CONTROL, 0);
2493         WREG32(DACB_AUTODETECT_INT_CONTROL, 0);
2494
2495         tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2496         WREG32(DC_HPD1_INT_CONTROL, tmp);
2497         tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2498         WREG32(DC_HPD2_INT_CONTROL, tmp);
2499         tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2500         WREG32(DC_HPD3_INT_CONTROL, tmp);
2501         tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2502         WREG32(DC_HPD4_INT_CONTROL, tmp);
2503         tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2504         WREG32(DC_HPD5_INT_CONTROL, tmp);
2505         tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
2506         WREG32(DC_HPD6_INT_CONTROL, tmp);
2507
2508 }
2509
2510 int evergreen_irq_set(struct radeon_device *rdev)
2511 {
2512         u32 cp_int_cntl = CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE;
2513         u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
2514         u32 hpd1, hpd2, hpd3, hpd4, hpd5, hpd6;
2515         u32 grbm_int_cntl = 0;
2516         u32 grph1 = 0, grph2 = 0, grph3 = 0, grph4 = 0, grph5 = 0, grph6 = 0;
2517
2518         if (!rdev->irq.installed) {
2519                 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
2520                 return -EINVAL;
2521         }
2522         /* don't enable anything if the ih is disabled */
2523         if (!rdev->ih.enabled) {
2524                 r600_disable_interrupts(rdev);
2525                 /* force the active interrupt state to all disabled */
2526                 evergreen_disable_interrupt_state(rdev);
2527                 return 0;
2528         }
2529
2530         hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
2531         hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
2532         hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
2533         hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
2534         hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
2535         hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
2536
2537         if (rdev->irq.sw_int) {
2538                 DRM_DEBUG("evergreen_irq_set: sw int\n");
2539                 cp_int_cntl |= RB_INT_ENABLE;
2540                 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
2541         }
2542         if (rdev->irq.crtc_vblank_int[0] ||
2543             rdev->irq.pflip[0]) {
2544                 DRM_DEBUG("evergreen_irq_set: vblank 0\n");
2545                 crtc1 |= VBLANK_INT_MASK;
2546         }
2547         if (rdev->irq.crtc_vblank_int[1] ||
2548             rdev->irq.pflip[1]) {
2549                 DRM_DEBUG("evergreen_irq_set: vblank 1\n");
2550                 crtc2 |= VBLANK_INT_MASK;
2551         }
2552         if (rdev->irq.crtc_vblank_int[2] ||
2553             rdev->irq.pflip[2]) {
2554                 DRM_DEBUG("evergreen_irq_set: vblank 2\n");
2555                 crtc3 |= VBLANK_INT_MASK;
2556         }
2557         if (rdev->irq.crtc_vblank_int[3] ||
2558             rdev->irq.pflip[3]) {
2559                 DRM_DEBUG("evergreen_irq_set: vblank 3\n");
2560                 crtc4 |= VBLANK_INT_MASK;
2561         }
2562         if (rdev->irq.crtc_vblank_int[4] ||
2563             rdev->irq.pflip[4]) {
2564                 DRM_DEBUG("evergreen_irq_set: vblank 4\n");
2565                 crtc5 |= VBLANK_INT_MASK;
2566         }
2567         if (rdev->irq.crtc_vblank_int[5] ||
2568             rdev->irq.pflip[5]) {
2569                 DRM_DEBUG("evergreen_irq_set: vblank 5\n");
2570                 crtc6 |= VBLANK_INT_MASK;
2571         }
2572         if (rdev->irq.hpd[0]) {
2573                 DRM_DEBUG("evergreen_irq_set: hpd 1\n");
2574                 hpd1 |= DC_HPDx_INT_EN;
2575         }
2576         if (rdev->irq.hpd[1]) {
2577                 DRM_DEBUG("evergreen_irq_set: hpd 2\n");
2578                 hpd2 |= DC_HPDx_INT_EN;
2579         }
2580         if (rdev->irq.hpd[2]) {
2581                 DRM_DEBUG("evergreen_irq_set: hpd 3\n");
2582                 hpd3 |= DC_HPDx_INT_EN;
2583         }
2584         if (rdev->irq.hpd[3]) {
2585                 DRM_DEBUG("evergreen_irq_set: hpd 4\n");
2586                 hpd4 |= DC_HPDx_INT_EN;
2587         }
2588         if (rdev->irq.hpd[4]) {
2589                 DRM_DEBUG("evergreen_irq_set: hpd 5\n");
2590                 hpd5 |= DC_HPDx_INT_EN;
2591         }
2592         if (rdev->irq.hpd[5]) {
2593                 DRM_DEBUG("evergreen_irq_set: hpd 6\n");
2594                 hpd6 |= DC_HPDx_INT_EN;
2595         }
2596         if (rdev->irq.gui_idle) {
2597                 DRM_DEBUG("gui idle\n");
2598                 grbm_int_cntl |= GUI_IDLE_INT_ENABLE;
2599         }
2600
2601         WREG32(CP_INT_CNTL, cp_int_cntl);
2602         WREG32(GRBM_INT_CNTL, grbm_int_cntl);
2603
2604         WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
2605         WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
2606         if (rdev->num_crtc >= 4) {
2607                 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
2608                 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
2609         }
2610         if (rdev->num_crtc >= 6) {
2611                 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
2612                 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
2613         }
2614
2615         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, grph1);
2616         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, grph2);
2617         if (rdev->num_crtc >= 4) {
2618                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, grph3);
2619                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, grph4);
2620         }
2621         if (rdev->num_crtc >= 6) {
2622                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, grph5);
2623                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, grph6);
2624         }
2625
2626         WREG32(DC_HPD1_INT_CONTROL, hpd1);
2627         WREG32(DC_HPD2_INT_CONTROL, hpd2);
2628         WREG32(DC_HPD3_INT_CONTROL, hpd3);
2629         WREG32(DC_HPD4_INT_CONTROL, hpd4);
2630         WREG32(DC_HPD5_INT_CONTROL, hpd5);
2631         WREG32(DC_HPD6_INT_CONTROL, hpd6);
2632
2633         return 0;
2634 }
2635
2636 static void evergreen_irq_ack(struct radeon_device *rdev)
2637 {
2638         u32 tmp;
2639
2640         rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
2641         rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
2642         rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
2643         rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
2644         rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
2645         rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
2646         rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
2647         rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
2648         if (rdev->num_crtc >= 4) {
2649                 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
2650                 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
2651         }
2652         if (rdev->num_crtc >= 6) {
2653                 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
2654                 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
2655         }
2656
2657         if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
2658                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2659         if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
2660                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2661         if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
2662                 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
2663         if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
2664                 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
2665         if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
2666                 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
2667         if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
2668                 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
2669
2670         if (rdev->num_crtc >= 4) {
2671                 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
2672                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2673                 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
2674                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2675                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
2676                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
2677                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
2678                         WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
2679                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
2680                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
2681                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
2682                         WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
2683         }
2684
2685         if (rdev->num_crtc >= 6) {
2686                 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
2687                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2688                 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
2689                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
2690                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
2691                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
2692                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
2693                         WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
2694                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
2695                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
2696                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
2697                         WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
2698         }
2699
2700         if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
2701                 tmp = RREG32(DC_HPD1_INT_CONTROL);
2702                 tmp |= DC_HPDx_INT_ACK;
2703                 WREG32(DC_HPD1_INT_CONTROL, tmp);
2704         }
2705         if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
2706                 tmp = RREG32(DC_HPD2_INT_CONTROL);
2707                 tmp |= DC_HPDx_INT_ACK;
2708                 WREG32(DC_HPD2_INT_CONTROL, tmp);
2709         }
2710         if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
2711                 tmp = RREG32(DC_HPD3_INT_CONTROL);
2712                 tmp |= DC_HPDx_INT_ACK;
2713                 WREG32(DC_HPD3_INT_CONTROL, tmp);
2714         }
2715         if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
2716                 tmp = RREG32(DC_HPD4_INT_CONTROL);
2717                 tmp |= DC_HPDx_INT_ACK;
2718                 WREG32(DC_HPD4_INT_CONTROL, tmp);
2719         }
2720         if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
2721                 tmp = RREG32(DC_HPD5_INT_CONTROL);
2722                 tmp |= DC_HPDx_INT_ACK;
2723                 WREG32(DC_HPD5_INT_CONTROL, tmp);
2724         }
2725         if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
2726                 tmp = RREG32(DC_HPD5_INT_CONTROL);
2727                 tmp |= DC_HPDx_INT_ACK;
2728                 WREG32(DC_HPD6_INT_CONTROL, tmp);
2729         }
2730 }
2731
2732 void evergreen_irq_disable(struct radeon_device *rdev)
2733 {
2734         r600_disable_interrupts(rdev);
2735         /* Wait and acknowledge irq */
2736         mdelay(1);
2737         evergreen_irq_ack(rdev);
2738         evergreen_disable_interrupt_state(rdev);
2739 }
2740
2741 void evergreen_irq_suspend(struct radeon_device *rdev)
2742 {
2743         evergreen_irq_disable(rdev);
2744         r600_rlc_stop(rdev);
2745 }
2746
2747 static u32 evergreen_get_ih_wptr(struct radeon_device *rdev)
2748 {
2749         u32 wptr, tmp;
2750
2751         if (rdev->wb.enabled)
2752                 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
2753         else
2754                 wptr = RREG32(IH_RB_WPTR);
2755
2756         if (wptr & RB_OVERFLOW) {
2757                 /* When a ring buffer overflow happen start parsing interrupt
2758                  * from the last not overwritten vector (wptr + 16). Hopefully
2759                  * this should allow us to catchup.
2760                  */
2761                 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
2762                         wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
2763                 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
2764                 tmp = RREG32(IH_RB_CNTL);
2765                 tmp |= IH_WPTR_OVERFLOW_CLEAR;
2766                 WREG32(IH_RB_CNTL, tmp);
2767         }
2768         return (wptr & rdev->ih.ptr_mask);
2769 }
2770
2771 int evergreen_irq_process(struct radeon_device *rdev)
2772 {
2773         u32 wptr;
2774         u32 rptr;
2775         u32 src_id, src_data;
2776         u32 ring_index;
2777         unsigned long flags;
2778         bool queue_hotplug = false;
2779
2780         if (!rdev->ih.enabled || rdev->shutdown)
2781                 return IRQ_NONE;
2782
2783         wptr = evergreen_get_ih_wptr(rdev);
2784         rptr = rdev->ih.rptr;
2785         DRM_DEBUG("r600_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
2786
2787         spin_lock_irqsave(&rdev->ih.lock, flags);
2788         if (rptr == wptr) {
2789                 spin_unlock_irqrestore(&rdev->ih.lock, flags);
2790                 return IRQ_NONE;
2791         }
2792 restart_ih:
2793         /* Order reading of wptr vs. reading of IH ring data */
2794         rmb();
2795
2796         /* display interrupts */
2797         evergreen_irq_ack(rdev);
2798
2799         rdev->ih.wptr = wptr;
2800         while (rptr != wptr) {
2801                 /* wptr/rptr are in bytes! */
2802                 ring_index = rptr / 4;
2803                 src_id =  le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
2804                 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
2805
2806                 switch (src_id) {
2807                 case 1: /* D1 vblank/vline */
2808                         switch (src_data) {
2809                         case 0: /* D1 vblank */
2810                                 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
2811                                         if (rdev->irq.crtc_vblank_int[0]) {
2812                                                 drm_handle_vblank(rdev->ddev, 0);
2813                                                 rdev->pm.vblank_sync = true;
2814                                                 wake_up(&rdev->irq.vblank_queue);
2815                                         }
2816                                         if (rdev->irq.pflip[0])
2817                                                 radeon_crtc_handle_flip(rdev, 0);
2818                                         rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
2819                                         DRM_DEBUG("IH: D1 vblank\n");
2820                                 }
2821                                 break;
2822                         case 1: /* D1 vline */
2823                                 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
2824                                         rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
2825                                         DRM_DEBUG("IH: D1 vline\n");
2826                                 }
2827                                 break;
2828                         default:
2829                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2830                                 break;
2831                         }
2832                         break;
2833                 case 2: /* D2 vblank/vline */
2834                         switch (src_data) {
2835                         case 0: /* D2 vblank */
2836                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
2837                                         if (rdev->irq.crtc_vblank_int[1]) {
2838                                                 drm_handle_vblank(rdev->ddev, 1);
2839                                                 rdev->pm.vblank_sync = true;
2840                                                 wake_up(&rdev->irq.vblank_queue);
2841                                         }
2842                                         if (rdev->irq.pflip[1])
2843                                                 radeon_crtc_handle_flip(rdev, 1);
2844                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
2845                                         DRM_DEBUG("IH: D2 vblank\n");
2846                                 }
2847                                 break;
2848                         case 1: /* D2 vline */
2849                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
2850                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
2851                                         DRM_DEBUG("IH: D2 vline\n");
2852                                 }
2853                                 break;
2854                         default:
2855                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2856                                 break;
2857                         }
2858                         break;
2859                 case 3: /* D3 vblank/vline */
2860                         switch (src_data) {
2861                         case 0: /* D3 vblank */
2862                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
2863                                         if (rdev->irq.crtc_vblank_int[2]) {
2864                                                 drm_handle_vblank(rdev->ddev, 2);
2865                                                 rdev->pm.vblank_sync = true;
2866                                                 wake_up(&rdev->irq.vblank_queue);
2867                                         }
2868                                         if (rdev->irq.pflip[2])
2869                                                 radeon_crtc_handle_flip(rdev, 2);
2870                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
2871                                         DRM_DEBUG("IH: D3 vblank\n");
2872                                 }
2873                                 break;
2874                         case 1: /* D3 vline */
2875                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
2876                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
2877                                         DRM_DEBUG("IH: D3 vline\n");
2878                                 }
2879                                 break;
2880                         default:
2881                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2882                                 break;
2883                         }
2884                         break;
2885                 case 4: /* D4 vblank/vline */
2886                         switch (src_data) {
2887                         case 0: /* D4 vblank */
2888                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
2889                                         if (rdev->irq.crtc_vblank_int[3]) {
2890                                                 drm_handle_vblank(rdev->ddev, 3);
2891                                                 rdev->pm.vblank_sync = true;
2892                                                 wake_up(&rdev->irq.vblank_queue);
2893                                         }
2894                                         if (rdev->irq.pflip[3])
2895                                                 radeon_crtc_handle_flip(rdev, 3);
2896                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
2897                                         DRM_DEBUG("IH: D4 vblank\n");
2898                                 }
2899                                 break;
2900                         case 1: /* D4 vline */
2901                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
2902                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
2903                                         DRM_DEBUG("IH: D4 vline\n");
2904                                 }
2905                                 break;
2906                         default:
2907                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2908                                 break;
2909                         }
2910                         break;
2911                 case 5: /* D5 vblank/vline */
2912                         switch (src_data) {
2913                         case 0: /* D5 vblank */
2914                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
2915                                         if (rdev->irq.crtc_vblank_int[4]) {
2916                                                 drm_handle_vblank(rdev->ddev, 4);
2917                                                 rdev->pm.vblank_sync = true;
2918                                                 wake_up(&rdev->irq.vblank_queue);
2919                                         }
2920                                         if (rdev->irq.pflip[4])
2921                                                 radeon_crtc_handle_flip(rdev, 4);
2922                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
2923                                         DRM_DEBUG("IH: D5 vblank\n");
2924                                 }
2925                                 break;
2926                         case 1: /* D5 vline */
2927                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
2928                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
2929                                         DRM_DEBUG("IH: D5 vline\n");
2930                                 }
2931                                 break;
2932                         default:
2933                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2934                                 break;
2935                         }
2936                         break;
2937                 case 6: /* D6 vblank/vline */
2938                         switch (src_data) {
2939                         case 0: /* D6 vblank */
2940                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
2941                                         if (rdev->irq.crtc_vblank_int[5]) {
2942                                                 drm_handle_vblank(rdev->ddev, 5);
2943                                                 rdev->pm.vblank_sync = true;
2944                                                 wake_up(&rdev->irq.vblank_queue);
2945                                         }
2946                                         if (rdev->irq.pflip[5])
2947                                                 radeon_crtc_handle_flip(rdev, 5);
2948                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
2949                                         DRM_DEBUG("IH: D6 vblank\n");
2950                                 }
2951                                 break;
2952                         case 1: /* D6 vline */
2953                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
2954                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
2955                                         DRM_DEBUG("IH: D6 vline\n");
2956                                 }
2957                                 break;
2958                         default:
2959                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
2960                                 break;
2961                         }
2962                         break;
2963                 case 42: /* HPD hotplug */
2964                         switch (src_data) {
2965                         case 0:
2966                                 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
2967                                         rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
2968                                         queue_hotplug = true;
2969                                         DRM_DEBUG("IH: HPD1\n");
2970                                 }
2971                                 break;
2972                         case 1:
2973                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
2974                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
2975                                         queue_hotplug = true;
2976                                         DRM_DEBUG("IH: HPD2\n");
2977                                 }
2978                                 break;
2979                         case 2:
2980                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
2981                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
2982                                         queue_hotplug = true;
2983                                         DRM_DEBUG("IH: HPD3\n");
2984                                 }
2985                                 break;
2986                         case 3:
2987                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
2988                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
2989                                         queue_hotplug = true;
2990                                         DRM_DEBUG("IH: HPD4\n");
2991                                 }
2992                                 break;
2993                         case 4:
2994                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
2995                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
2996                                         queue_hotplug = true;
2997                                         DRM_DEBUG("IH: HPD5\n");
2998                                 }
2999                                 break;
3000                         case 5:
3001                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
3002                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
3003                                         queue_hotplug = true;
3004                                         DRM_DEBUG("IH: HPD6\n");
3005                                 }
3006                                 break;
3007                         default:
3008                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3009                                 break;
3010                         }
3011                         break;
3012                 case 176: /* CP_INT in ring buffer */
3013                 case 177: /* CP_INT in IB1 */
3014                 case 178: /* CP_INT in IB2 */
3015                         DRM_DEBUG("IH: CP int: 0x%08x\n", src_data);
3016                         radeon_fence_process(rdev);
3017                         break;
3018                 case 181: /* CP EOP event */
3019                         DRM_DEBUG("IH: CP EOP\n");
3020                         radeon_fence_process(rdev);
3021                         break;
3022                 case 233: /* GUI IDLE */
3023                         DRM_DEBUG("IH: GUI idle\n");
3024                         rdev->pm.gui_idle = true;
3025                         wake_up(&rdev->irq.idle_queue);
3026                         break;
3027                 default:
3028                         DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3029                         break;
3030                 }
3031
3032                 /* wptr/rptr are in bytes! */
3033                 rptr += 16;
3034                 rptr &= rdev->ih.ptr_mask;
3035         }
3036         /* make sure wptr hasn't changed while processing */
3037         wptr = evergreen_get_ih_wptr(rdev);
3038         if (wptr != rdev->ih.wptr)
3039                 goto restart_ih;
3040         if (queue_hotplug)
3041                 schedule_work(&rdev->hotplug_work);
3042         rdev->ih.rptr = rptr;
3043         WREG32(IH_RB_RPTR, rdev->ih.rptr);
3044         spin_unlock_irqrestore(&rdev->ih.lock, flags);
3045         return IRQ_HANDLED;
3046 }
3047
3048 static int evergreen_startup(struct radeon_device *rdev)
3049 {
3050         int r;
3051
3052         /* enable pcie gen2 link */
3053         evergreen_pcie_gen2_enable(rdev);
3054
3055         if (ASIC_IS_DCE5(rdev)) {
3056                 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw || !rdev->mc_fw) {
3057                         r = ni_init_microcode(rdev);
3058                         if (r) {
3059                                 DRM_ERROR("Failed to load firmware!\n");
3060                                 return r;
3061                         }
3062                 }
3063                 r = ni_mc_load_microcode(rdev);
3064                 if (r) {
3065                         DRM_ERROR("Failed to load MC firmware!\n");
3066                         return r;
3067                 }
3068         } else {
3069                 if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
3070                         r = r600_init_microcode(rdev);
3071                         if (r) {
3072                                 DRM_ERROR("Failed to load firmware!\n");
3073                                 return r;
3074                         }
3075                 }
3076         }
3077
3078         r = r600_vram_scratch_init(rdev);
3079         if (r)
3080                 return r;
3081
3082         evergreen_mc_program(rdev);
3083         if (rdev->flags & RADEON_IS_AGP) {
3084                 evergreen_agp_enable(rdev);
3085         } else {
3086                 r = evergreen_pcie_gart_enable(rdev);
3087                 if (r)
3088                         return r;
3089         }
3090         evergreen_gpu_init(rdev);
3091
3092         r = evergreen_blit_init(rdev);
3093         if (r) {
3094                 r600_blit_fini(rdev);
3095                 rdev->asic->copy = NULL;
3096                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
3097         }
3098
3099         /* allocate wb buffer */
3100         r = radeon_wb_init(rdev);
3101         if (r)
3102                 return r;
3103
3104         /* Enable IRQ */
3105         r = r600_irq_init(rdev);
3106         if (r) {
3107                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
3108                 radeon_irq_kms_fini(rdev);
3109                 return r;
3110         }
3111         evergreen_irq_set(rdev);
3112
3113         r = radeon_ring_init(rdev, rdev->cp.ring_size);
3114         if (r)
3115                 return r;
3116         r = evergreen_cp_load_microcode(rdev);
3117         if (r)
3118                 return r;
3119         r = evergreen_cp_resume(rdev);
3120         if (r)
3121                 return r;
3122
3123         return 0;
3124 }
3125
3126 int evergreen_resume(struct radeon_device *rdev)
3127 {
3128         int r;
3129
3130         /* reset the asic, the gfx blocks are often in a bad state
3131          * after the driver is unloaded or after a resume
3132          */
3133         if (radeon_asic_reset(rdev))
3134                 dev_warn(rdev->dev, "GPU reset failed !\n");
3135         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
3136          * posting will perform necessary task to bring back GPU into good
3137          * shape.
3138          */
3139         /* post card */
3140         atom_asic_init(rdev->mode_info.atom_context);
3141
3142         r = evergreen_startup(rdev);
3143         if (r) {
3144                 DRM_ERROR("evergreen startup failed on resume\n");
3145                 return r;
3146         }
3147
3148         r = r600_ib_test(rdev);
3149         if (r) {
3150                 DRM_ERROR("radeon: failed testing IB (%d).\n", r);
3151                 return r;
3152         }
3153
3154         return r;
3155
3156 }
3157
3158 int evergreen_suspend(struct radeon_device *rdev)
3159 {
3160         /* FIXME: we should wait for ring to be empty */
3161         r700_cp_stop(rdev);
3162         rdev->cp.ready = false;
3163         evergreen_irq_suspend(rdev);
3164         radeon_wb_disable(rdev);
3165         evergreen_pcie_gart_disable(rdev);
3166         r600_blit_suspend(rdev);
3167
3168         return 0;
3169 }
3170
3171 /* Plan is to move initialization in that function and use
3172  * helper function so that radeon_device_init pretty much
3173  * do nothing more than calling asic specific function. This
3174  * should also allow to remove a bunch of callback function
3175  * like vram_info.
3176  */
3177 int evergreen_init(struct radeon_device *rdev)
3178 {
3179         int r;
3180
3181         /* This don't do much */
3182         r = radeon_gem_init(rdev);
3183         if (r)
3184                 return r;
3185         /* Read BIOS */
3186         if (!radeon_get_bios(rdev)) {
3187                 if (ASIC_IS_AVIVO(rdev))
3188                         return -EINVAL;
3189         }
3190         /* Must be an ATOMBIOS */
3191         if (!rdev->is_atom_bios) {
3192                 dev_err(rdev->dev, "Expecting atombios for evergreen GPU\n");
3193                 return -EINVAL;
3194         }
3195         r = radeon_atombios_init(rdev);
3196         if (r)
3197                 return r;
3198         /* reset the asic, the gfx blocks are often in a bad state
3199          * after the driver is unloaded or after a resume
3200          */
3201         if (radeon_asic_reset(rdev))
3202                 dev_warn(rdev->dev, "GPU reset failed !\n");
3203         /* Post card if necessary */
3204         if (!radeon_card_posted(rdev)) {
3205                 if (!rdev->bios) {
3206                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
3207                         return -EINVAL;
3208                 }
3209                 DRM_INFO("GPU not posted. posting now...\n");
3210                 atom_asic_init(rdev->mode_info.atom_context);
3211         }
3212         /* Initialize scratch registers */
3213         r600_scratch_init(rdev);
3214         /* Initialize surface registers */
3215         radeon_surface_init(rdev);
3216         /* Initialize clocks */
3217         radeon_get_clock_info(rdev->ddev);
3218         /* Fence driver */
3219         r = radeon_fence_driver_init(rdev);
3220         if (r)
3221                 return r;
3222         /* initialize AGP */
3223         if (rdev->flags & RADEON_IS_AGP) {
3224                 r = radeon_agp_init(rdev);
3225                 if (r)
3226                         radeon_agp_disable(rdev);
3227         }
3228         /* initialize memory controller */
3229         r = evergreen_mc_init(rdev);
3230         if (r)
3231                 return r;
3232         /* Memory manager */
3233         r = radeon_bo_init(rdev);
3234         if (r)
3235                 return r;
3236
3237         r = radeon_irq_kms_init(rdev);
3238         if (r)
3239                 return r;
3240
3241         rdev->cp.ring_obj = NULL;
3242         r600_ring_init(rdev, 1024 * 1024);
3243
3244         rdev->ih.ring_obj = NULL;
3245         r600_ih_ring_init(rdev, 64 * 1024);
3246
3247         r = r600_pcie_gart_init(rdev);
3248         if (r)
3249                 return r;
3250
3251         rdev->accel_working = true;
3252         r = evergreen_startup(rdev);
3253         if (r) {
3254                 dev_err(rdev->dev, "disabling GPU acceleration\n");
3255                 r700_cp_fini(rdev);
3256                 r600_irq_fini(rdev);
3257                 radeon_wb_fini(rdev);
3258                 radeon_irq_kms_fini(rdev);
3259                 evergreen_pcie_gart_fini(rdev);
3260                 rdev->accel_working = false;
3261         }
3262         if (rdev->accel_working) {
3263                 r = radeon_ib_pool_init(rdev);
3264                 if (r) {
3265                         DRM_ERROR("radeon: failed initializing IB pool (%d).\n", r);
3266                         rdev->accel_working = false;
3267                 }
3268                 r = r600_ib_test(rdev);
3269                 if (r) {
3270                         DRM_ERROR("radeon: failed testing IB (%d).\n", r);
3271                         rdev->accel_working = false;
3272                 }
3273         }
3274         return 0;
3275 }
3276
3277 void evergreen_fini(struct radeon_device *rdev)
3278 {
3279         r600_blit_fini(rdev);
3280         r700_cp_fini(rdev);
3281         r600_irq_fini(rdev);
3282         radeon_wb_fini(rdev);
3283         radeon_ib_pool_fini(rdev);
3284         radeon_irq_kms_fini(rdev);
3285         evergreen_pcie_gart_fini(rdev);
3286         r600_vram_scratch_fini(rdev);
3287         radeon_gem_fini(rdev);
3288         radeon_fence_driver_fini(rdev);
3289         radeon_agp_fini(rdev);
3290         radeon_bo_fini(rdev);
3291         radeon_atombios_fini(rdev);
3292         kfree(rdev->bios);
3293         rdev->bios = NULL;
3294 }
3295
3296 void evergreen_pcie_gen2_enable(struct radeon_device *rdev)
3297 {
3298         u32 link_width_cntl, speed_cntl;
3299
3300         if (radeon_pcie_gen2 == 0)
3301                 return;
3302
3303         if (rdev->flags & RADEON_IS_IGP)
3304                 return;
3305
3306         if (!(rdev->flags & RADEON_IS_PCIE))
3307                 return;
3308
3309         /* x2 cards have a special sequence */
3310         if (ASIC_IS_X2(rdev))
3311                 return;
3312
3313         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
3314         if ((speed_cntl & LC_OTHER_SIDE_EVER_SENT_GEN2) ||
3315             (speed_cntl & LC_OTHER_SIDE_SUPPORTS_GEN2)) {
3316
3317                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
3318                 link_width_cntl &= ~LC_UPCONFIGURE_DIS;
3319                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
3320
3321                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
3322                 speed_cntl &= ~LC_TARGET_LINK_SPEED_OVERRIDE_EN;
3323                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
3324
3325                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
3326                 speed_cntl |= LC_CLR_FAILED_SPD_CHANGE_CNT;
3327                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
3328
3329                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
3330                 speed_cntl &= ~LC_CLR_FAILED_SPD_CHANGE_CNT;
3331                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
3332
3333                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
3334                 speed_cntl |= LC_GEN2_EN_STRAP;
3335                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
3336
3337         } else {
3338                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
3339                 /* XXX: only disable it if gen1 bridge vendor == 0x111d or 0x1106 */
3340                 if (1)
3341                         link_width_cntl |= LC_UPCONFIGURE_DIS;
3342                 else
3343                         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
3344                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
3345         }
3346 }