drivers/gpu/drm/radeon/rv770.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/firmware.h>
  29 #include <linux/platform_device.h>
  30 #include <linux/slab.h>
  31 #include "drmP.h"
  32 #include "radeon.h"
  33 #include "radeon_asic.h"
  34 #include "radeon_drm.h"
  35 #include "rv770d.h"
  36 #include "atom.h"
  37 #include "avivod.h"
  38
  39 #define R700_PFP_UCODE_SIZE 848
  40 #define R700_PM4_UCODE_SIZE 1360
  41
  42 static void rv770_gpu_init(struct radeon_device *rdev);
  43 void rv770_fini(struct radeon_device *rdev);
  44 static void rv770_pcie_gen2_enable(struct radeon_device *rdev);
  45
  46 u32 rv770_page_flip(struct radeon_device *rdev, int crtc_id, u64 crtc_base)
  47 {
  48         struct radeon_crtc *radeon_crtc = rdev->mode_info.crtcs[crtc_id];
  49         u32 tmp = RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset);
  50         int i;
  51
  52         /* Lock the graphics update lock */
  53         tmp |= AVIVO_D1GRPH_UPDATE_LOCK;
  54         WREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
  55
  56         /* update the scanout addresses */
  57         if (radeon_crtc->crtc_id) {
  58                 WREG32(D2GRPH_SECONDARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  59                 WREG32(D2GRPH_PRIMARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  60         } else {
  61                 WREG32(D1GRPH_SECONDARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  62                 WREG32(D1GRPH_PRIMARY_SURFACE_ADDRESS_HIGH, upper_32_bits(crtc_base));
  63         }
  64         WREG32(D1GRPH_SECONDARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  65                (u32)crtc_base);
  66         WREG32(D1GRPH_PRIMARY_SURFACE_ADDRESS + radeon_crtc->crtc_offset,
  67                (u32)crtc_base);
  68
  69         /* Wait for update_pending to go high. */
  70         for (i = 0; i < rdev->usec_timeout; i++) {
  71                 if (RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset) & AVIVO_D1GRPH_SURFACE_UPDATE_PENDING)
  72                         break;
  73                 udelay(1);
  74         }
  75         DRM_DEBUG("Update pending now high. Unlocking vupdate_lock.\n");
  76
  77         /* Unlock the lock, so double-buffering can take place inside vblank */
  78         tmp &= ~AVIVO_D1GRPH_UPDATE_LOCK;
  79         WREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset, tmp);
  80
  81         /* Return current update_pending status: */
  82         return RREG32(AVIVO_D1GRPH_UPDATE + radeon_crtc->crtc_offset) & AVIVO_D1GRPH_SURFACE_UPDATE_PENDING;
  83 }
  84
  85 /* get temperature in millidegrees */
  86 int rv770_get_temp(struct radeon_device *rdev)
  87 {
  88         u32 temp = (RREG32(CG_MULT_THERMAL_STATUS) & ASIC_T_MASK) >>
  89                 ASIC_T_SHIFT;
  90         int actual_temp;
  91
  92         if (temp & 0x400)
  93                 actual_temp = -256;
  94         else if (temp & 0x200)
  95                 actual_temp = 255;
  96         else if (temp & 0x100) {
  97                 actual_temp = temp & 0x1ff;
  98                 actual_temp |= ~0x1ff;
  99         } else
 100                 actual_temp = temp & 0xff;
 101
 102         return (actual_temp * 1000) / 2;
 103 }
 104
 105 void rv770_pm_misc(struct radeon_device *rdev)
 106 {
 107         int req_ps_idx = rdev->pm.requested_power_state_index;
 108         int req_cm_idx = rdev->pm.requested_clock_mode_index;
 109         struct radeon_power_state *ps = &rdev->pm.power_state[req_ps_idx];
 110         struct radeon_voltage *voltage = &ps->clock_info[req_cm_idx].voltage;
 111
 112         if ((voltage->type == VOLTAGE_SW) && voltage->voltage) {
 113                 /* 0xff01 is a flag rather then an actual voltage */
 114                 if (voltage->voltage == 0xff01)
 115                         return;
 116                 if (voltage->voltage != rdev->pm.current_vddc) {
 117                         radeon_atom_set_voltage(rdev, voltage->voltage, SET_VOLTAGE_TYPE_ASIC_VDDC);
 118                         rdev->pm.current_vddc = voltage->voltage;
 119                         DRM_DEBUG("Setting: v: %d\n", voltage->voltage);
 120                 }
 121         }
 122 }
 123
 124 /*
 125  * GART
 126  */
 127 int rv770_pcie_gart_enable(struct radeon_device *rdev)
 128 {
 129         u32 tmp;
 130         int r, i;
 131
 132         if (rdev->gart.robj == NULL) {
 133                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
 134                 return -EINVAL;
 135         }
 136         r = radeon_gart_table_vram_pin(rdev);
 137         if (r)
 138                 return r;
 139         radeon_gart_restore(rdev);
 140         /* Setup L2 cache */
 141         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 142                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 143                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 144         WREG32(VM_L2_CNTL2, 0);
 145         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 146         /* Setup TLB control */
 147         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 148                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 149                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 150                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 151         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 152         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 153         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 154         if (rdev->family == CHIP_RV740)
 155                 WREG32(MC_VM_MD_L1_TLB3_CNTL, tmp);
 156         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 157         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 158         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 159         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 160         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 161         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
 162         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
 163         WREG32(VM_CONTEXT0_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
 164                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT);
 165         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 166                         (u32)(rdev->dummy_page.addr >> 12));
 167         for (i = 1; i < 7; i++)
 168                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 169
 170         r600_pcie_gart_tlb_flush(rdev);
 171         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 172                  (unsigned)(rdev->mc.gtt_size >> 20),
 173                  (unsigned long long)rdev->gart.table_addr);
 174         rdev->gart.ready = true;
 175         return 0;
 176 }
 177
 178 void rv770_pcie_gart_disable(struct radeon_device *rdev)
 179 {
 180         u32 tmp;
 181         int i;
 182
 183         /* Disable all tables */
 184         for (i = 0; i < 7; i++)
 185                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 186
 187         /* Setup L2 cache */
 188         WREG32(VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING |
 189                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 190         WREG32(VM_L2_CNTL2, 0);
 191         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 192         /* Setup TLB control */
 193         tmp = EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 194         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 195         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 196         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 197         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 198         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 199         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 200         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 201         radeon_gart_table_vram_unpin(rdev);
 202 }
 203
 204 void rv770_pcie_gart_fini(struct radeon_device *rdev)
 205 {
 206         radeon_gart_fini(rdev);
 207         rv770_pcie_gart_disable(rdev);
 208         radeon_gart_table_vram_free(rdev);
 209 }
 210
 211
 212 void rv770_agp_enable(struct radeon_device *rdev)
 213 {
 214         u32 tmp;
 215         int i;
 216
 217         /* Setup L2 cache */
 218         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE | ENABLE_L2_FRAGMENT_PROCESSING |
 219                                 ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 220                                 EFFECTIVE_L2_QUEUE_SIZE(7));
 221         WREG32(VM_L2_CNTL2, 0);
 222         WREG32(VM_L2_CNTL3, BANK_SELECT(0) | CACHE_UPDATE_MODE(2));
 223         /* Setup TLB control */
 224         tmp = ENABLE_L1_TLB | ENABLE_L1_FRAGMENT_PROCESSING |
 225                 SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 226                 SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 227                 EFFECTIVE_L1_TLB_SIZE(5) | EFFECTIVE_L1_QUEUE_SIZE(5);
 228         WREG32(MC_VM_MD_L1_TLB0_CNTL, tmp);
 229         WREG32(MC_VM_MD_L1_TLB1_CNTL, tmp);
 230         WREG32(MC_VM_MD_L1_TLB2_CNTL, tmp);
 231         WREG32(MC_VM_MB_L1_TLB0_CNTL, tmp);
 232         WREG32(MC_VM_MB_L1_TLB1_CNTL, tmp);
 233         WREG32(MC_VM_MB_L1_TLB2_CNTL, tmp);
 234         WREG32(MC_VM_MB_L1_TLB3_CNTL, tmp);
 235         for (i = 0; i < 7; i++)
 236                 WREG32(VM_CONTEXT0_CNTL + (i * 4), 0);
 237 }
 238
 239 static void rv770_mc_program(struct radeon_device *rdev)
 240 {
 241         struct rv515_mc_save save;
 242         u32 tmp;
 243         int i, j;
 244
 245         /* Initialize HDP */
 246         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
 247                 WREG32((0x2c14 + j), 0x00000000);
 248                 WREG32((0x2c18 + j), 0x00000000);
 249                 WREG32((0x2c1c + j), 0x00000000);
 250                 WREG32((0x2c20 + j), 0x00000000);
 251                 WREG32((0x2c24 + j), 0x00000000);
 252         }
 253         /* r7xx hw bug.  Read from HDP_DEBUG1 rather
 254          * than writing to HDP_REG_COHERENCY_FLUSH_CNTL
 255          */
 256         tmp = RREG32(HDP_DEBUG1);
 257
 258         rv515_mc_stop(rdev, &save);
 259         if (r600_mc_wait_for_idle(rdev)) {
 260                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 261         }
 262         /* Lockout access through VGA aperture*/
 263         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
 264         /* Update configuration */
 265         if (rdev->flags & RADEON_IS_AGP) {
 266                 if (rdev->mc.vram_start < rdev->mc.gtt_start) {
 267                         /* VRAM before AGP */
 268                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 269                                 rdev->mc.vram_start >> 12);
 270                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 271                                 rdev->mc.gtt_end >> 12);
 272                 } else {
 273                         /* VRAM after AGP */
 274                         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 275                                 rdev->mc.gtt_start >> 12);
 276                         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 277                                 rdev->mc.vram_end >> 12);
 278                 }
 279         } else {
 280                 WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
 281                         rdev->mc.vram_start >> 12);
 282                 WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 283                         rdev->mc.vram_end >> 12);
 284         }
 285         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, rdev->vram_scratch.gpu_addr >> 12);
 286         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
 287         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
 288         WREG32(MC_VM_FB_LOCATION, tmp);
 289         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
 290         WREG32(HDP_NONSURFACE_INFO, (2 << 7));
 291         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
 292         if (rdev->flags & RADEON_IS_AGP) {
 293                 WREG32(MC_VM_AGP_TOP, rdev->mc.gtt_end >> 16);
 294                 WREG32(MC_VM_AGP_BOT, rdev->mc.gtt_start >> 16);
 295                 WREG32(MC_VM_AGP_BASE, rdev->mc.agp_base >> 22);
 296         } else {
 297                 WREG32(MC_VM_AGP_BASE, 0);
 298                 WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
 299                 WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
 300         }
 301         if (r600_mc_wait_for_idle(rdev)) {
 302                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
 303         }
 304         rv515_mc_resume(rdev, &save);
 305         /* we need to own VRAM, so turn off the VGA renderer here
 306          * to stop it overwriting our objects */
 307         rv515_vga_render_disable(rdev);
 308 }
 309
 310
 311 /*
 312  * CP.
 313  */
 314 void r700_cp_stop(struct radeon_device *rdev)
 315 {
 316         radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
 317         WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT));
 318         WREG32(SCRATCH_UMSK, 0);
 319 }
 320
 321 static int rv770_cp_load_microcode(struct radeon_device *rdev)
 322 {
 323         const __be32 *fw_data;
 324         int i;
 325
 326         if (!rdev->me_fw || !rdev->pfp_fw)
 327                 return -EINVAL;
 328
 329         r700_cp_stop(rdev);
 330         WREG32(CP_RB_CNTL,
 331 #ifdef __BIG_ENDIAN
 332                BUF_SWAP_32BIT |
 333 #endif
 334                RB_NO_UPDATE | RB_BLKSZ(15) | RB_BUFSZ(3));
 335
 336         /* Reset cp */
 337         WREG32(GRBM_SOFT_RESET, SOFT_RESET_CP);
 338         RREG32(GRBM_SOFT_RESET);
 339         mdelay(15);
 340         WREG32(GRBM_SOFT_RESET, 0);
 341
 342         fw_data = (const __be32 *)rdev->pfp_fw->data;
 343         WREG32(CP_PFP_UCODE_ADDR, 0);
 344         for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 345                 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
 346         WREG32(CP_PFP_UCODE_ADDR, 0);
 347
 348         fw_data = (const __be32 *)rdev->me_fw->data;
 349         WREG32(CP_ME_RAM_WADDR, 0);
 350         for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 351                 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
 352
 353         WREG32(CP_PFP_UCODE_ADDR, 0);
 354         WREG32(CP_ME_RAM_WADDR, 0);
 355         WREG32(CP_ME_RAM_RADDR, 0);
 356         return 0;
 357 }
 358
 359 void r700_cp_fini(struct radeon_device *rdev)
 360 {
 361         r700_cp_stop(rdev);
 362         radeon_ring_fini(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
 363 }
 364
 365 /*
 366  * Core functions
 367  */
 368 static void rv770_gpu_init(struct radeon_device *rdev)
 369 {
 370         int i, j, num_qd_pipes;
 371         u32 ta_aux_cntl;
 372         u32 sx_debug_1;
 373         u32 smx_dc_ctl0;
 374         u32 db_debug3;
 375         u32 num_gs_verts_per_thread;
 376         u32 vgt_gs_per_es;
 377         u32 gs_prim_buffer_depth = 0;
 378         u32 sq_ms_fifo_sizes;
 379         u32 sq_config;
 380         u32 sq_thread_resource_mgmt;
 381         u32 hdp_host_path_cntl;
 382         u32 sq_dyn_gpr_size_simd_ab_0;
 383         u32 gb_tiling_config = 0;
 384         u32 cc_rb_backend_disable = 0;
 385         u32 cc_gc_shader_pipe_config = 0;
 386         u32 mc_arb_ramcfg;
 387         u32 db_debug4, tmp;
 388         u32 inactive_pipes, shader_pipe_config;
 389         u32 disabled_rb_mask;
 390         unsigned active_number;
 391
 392         /* setup chip specs */
 393         rdev->config.rv770.tiling_group_size = 256;
 394         switch (rdev->family) {
 395         case CHIP_RV770:
 396                 rdev->config.rv770.max_pipes = 4;
 397                 rdev->config.rv770.max_tile_pipes = 8;
 398                 rdev->config.rv770.max_simds = 10;
 399                 rdev->config.rv770.max_backends = 4;
 400                 rdev->config.rv770.max_gprs = 256;
 401                 rdev->config.rv770.max_threads = 248;
 402                 rdev->config.rv770.max_stack_entries = 512;
 403                 rdev->config.rv770.max_hw_contexts = 8;
 404                 rdev->config.rv770.max_gs_threads = 16 * 2;
 405                 rdev->config.rv770.sx_max_export_size = 128;
 406                 rdev->config.rv770.sx_max_export_pos_size = 16;
 407                 rdev->config.rv770.sx_max_export_smx_size = 112;
 408                 rdev->config.rv770.sq_num_cf_insts = 2;
 409
 410                 rdev->config.rv770.sx_num_of_sets = 7;
 411                 rdev->config.rv770.sc_prim_fifo_size = 0xF9;
 412                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 413                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 414                 break;
 415         case CHIP_RV730:
 416                 rdev->config.rv770.max_pipes = 2;
 417                 rdev->config.rv770.max_tile_pipes = 4;
 418                 rdev->config.rv770.max_simds = 8;
 419                 rdev->config.rv770.max_backends = 2;
 420                 rdev->config.rv770.max_gprs = 128;
 421                 rdev->config.rv770.max_threads = 248;
 422                 rdev->config.rv770.max_stack_entries = 256;
 423                 rdev->config.rv770.max_hw_contexts = 8;
 424                 rdev->config.rv770.max_gs_threads = 16 * 2;
 425                 rdev->config.rv770.sx_max_export_size = 256;
 426                 rdev->config.rv770.sx_max_export_pos_size = 32;
 427                 rdev->config.rv770.sx_max_export_smx_size = 224;
 428                 rdev->config.rv770.sq_num_cf_insts = 2;
 429
 430                 rdev->config.rv770.sx_num_of_sets = 7;
 431                 rdev->config.rv770.sc_prim_fifo_size = 0xf9;
 432                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 433                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 434                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 435                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 436                         rdev->config.rv770.sx_max_export_smx_size += 16;
 437                 }
 438                 break;
 439         case CHIP_RV710:
 440                 rdev->config.rv770.max_pipes = 2;
 441                 rdev->config.rv770.max_tile_pipes = 2;
 442                 rdev->config.rv770.max_simds = 2;
 443                 rdev->config.rv770.max_backends = 1;
 444                 rdev->config.rv770.max_gprs = 256;
 445                 rdev->config.rv770.max_threads = 192;
 446                 rdev->config.rv770.max_stack_entries = 256;
 447                 rdev->config.rv770.max_hw_contexts = 4;
 448                 rdev->config.rv770.max_gs_threads = 8 * 2;
 449                 rdev->config.rv770.sx_max_export_size = 128;
 450                 rdev->config.rv770.sx_max_export_pos_size = 16;
 451                 rdev->config.rv770.sx_max_export_smx_size = 112;
 452                 rdev->config.rv770.sq_num_cf_insts = 1;
 453
 454                 rdev->config.rv770.sx_num_of_sets = 7;
 455                 rdev->config.rv770.sc_prim_fifo_size = 0x40;
 456                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 457                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 458                 break;
 459         case CHIP_RV740:
 460                 rdev->config.rv770.max_pipes = 4;
 461                 rdev->config.rv770.max_tile_pipes = 4;
 462                 rdev->config.rv770.max_simds = 8;
 463                 rdev->config.rv770.max_backends = 4;
 464                 rdev->config.rv770.max_gprs = 256;
 465                 rdev->config.rv770.max_threads = 248;
 466                 rdev->config.rv770.max_stack_entries = 512;
 467                 rdev->config.rv770.max_hw_contexts = 8;
 468                 rdev->config.rv770.max_gs_threads = 16 * 2;
 469                 rdev->config.rv770.sx_max_export_size = 256;
 470                 rdev->config.rv770.sx_max_export_pos_size = 32;
 471                 rdev->config.rv770.sx_max_export_smx_size = 224;
 472                 rdev->config.rv770.sq_num_cf_insts = 2;
 473
 474                 rdev->config.rv770.sx_num_of_sets = 7;
 475                 rdev->config.rv770.sc_prim_fifo_size = 0x100;
 476                 rdev->config.rv770.sc_hiz_tile_fifo_size = 0x30;
 477                 rdev->config.rv770.sc_earlyz_tile_fifo_fize = 0x130;
 478
 479                 if (rdev->config.rv770.sx_max_export_pos_size > 16) {
 480                         rdev->config.rv770.sx_max_export_pos_size -= 16;
 481                         rdev->config.rv770.sx_max_export_smx_size += 16;
 482                 }
 483                 break;
 484         default:
 485                 break;
 486         }
 487
 488         /* Initialize HDP */
 489         j = 0;
 490         for (i = 0; i < 32; i++) {
 491                 WREG32((0x2c14 + j), 0x00000000);
 492                 WREG32((0x2c18 + j), 0x00000000);
 493                 WREG32((0x2c1c + j), 0x00000000);
 494                 WREG32((0x2c20 + j), 0x00000000);
 495                 WREG32((0x2c24 + j), 0x00000000);
 496                 j += 0x18;
 497         }
 498
 499         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
 500
 501         /* setup tiling, simd, pipe config */
 502         mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
 503
 504         shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG);
 505         inactive_pipes = (shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> INACTIVE_QD_PIPES_SHIFT;
 506         for (i = 0, tmp = 1, active_number = 0; i < R7XX_MAX_PIPES; i++) {
 507                 if (!(inactive_pipes & tmp)) {
 508                         active_number++;
 509                 }
 510                 tmp <<= 1;
 511         }
 512         if (active_number == 1) {
 513                 WREG32(SPI_CONFIG_CNTL, DISABLE_INTERP_1);
 514         } else {
 515                 WREG32(SPI_CONFIG_CNTL, 0);
 516         }
 517
 518         cc_rb_backend_disable = RREG32(CC_RB_BACKEND_DISABLE) & 0x00ff0000;
 519         tmp = R7XX_MAX_BACKENDS - r600_count_pipe_bits(cc_rb_backend_disable >> 16);
 520         if (tmp < rdev->config.rv770.max_backends) {
 521                 rdev->config.rv770.max_backends = tmp;
 522         }
 523
 524         cc_gc_shader_pipe_config = RREG32(CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
 525         tmp = R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config >> 8) & R7XX_MAX_PIPES_MASK);
 526         if (tmp < rdev->config.rv770.max_pipes) {
 527                 rdev->config.rv770.max_pipes = tmp;
 528         }
 529         tmp = R7XX_MAX_SIMDS - r600_count_pipe_bits((cc_gc_shader_pipe_config >> 16) & R7XX_MAX_SIMDS_MASK);
 530         if (tmp < rdev->config.rv770.max_simds) {
 531                 rdev->config.rv770.max_simds = tmp;
 532         }
 533
 534         switch (rdev->config.rv770.max_tile_pipes) {
 535         case 1:
 536         default:
 537                 gb_tiling_config = PIPE_TILING(0);
 538                 break;
 539         case 2:
 540                 gb_tiling_config = PIPE_TILING(1);
 541                 break;
 542         case 4:
 543                 gb_tiling_config = PIPE_TILING(2);
 544                 break;
 545         case 8:
 546                 gb_tiling_config = PIPE_TILING(3);
 547                 break;
 548         }
 549         rdev->config.rv770.tiling_npipes = rdev->config.rv770.max_tile_pipes;
 550
 551         disabled_rb_mask = (RREG32(CC_RB_BACKEND_DISABLE) >> 16) & R7XX_MAX_BACKENDS_MASK;
 552         tmp = (gb_tiling_config & PIPE_TILING__MASK) >> PIPE_TILING__SHIFT;
 553         tmp = r6xx_remap_render_backend(rdev, tmp, rdev->config.rv770.max_backends,
 554                                         R7XX_MAX_BACKENDS, disabled_rb_mask);
 555         gb_tiling_config |= tmp << 16;
 556         rdev->config.rv770.backend_map = tmp;
 557
 558         if (rdev->family == CHIP_RV770)
 559                 gb_tiling_config |= BANK_TILING(1);
 560         else {
 561                 if ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT)
 562                         gb_tiling_config |= BANK_TILING(1);
 563                 else
 564                         gb_tiling_config |= BANK_TILING(0);
 565         }
 566         rdev->config.rv770.tiling_nbanks = 4 << ((gb_tiling_config >> 4) & 0x3);
 567         gb_tiling_config |= GROUP_SIZE((mc_arb_ramcfg & BURSTLENGTH_MASK) >> BURSTLENGTH_SHIFT);
 568         if (((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT) > 3) {
 569                 gb_tiling_config |= ROW_TILING(3);
 570                 gb_tiling_config |= SAMPLE_SPLIT(3);
 571         } else {
 572                 gb_tiling_config |=
 573                         ROW_TILING(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 574                 gb_tiling_config |=
 575                         SAMPLE_SPLIT(((mc_arb_ramcfg & NOOFROWS_MASK) >> NOOFROWS_SHIFT));
 576         }
 577
 578         gb_tiling_config |= BANK_SWAPS(1);
 579         rdev->config.rv770.tile_config = gb_tiling_config;
 580
 581         WREG32(GB_TILING_CONFIG, gb_tiling_config);
 582         WREG32(DCP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 583         WREG32(HDP_TILING_CONFIG, (gb_tiling_config & 0xffff));
 584
 585         WREG32(CGTS_SYS_TCC_DISABLE, 0);
 586         WREG32(CGTS_TCC_DISABLE, 0);
 587         WREG32(CGTS_USER_SYS_TCC_DISABLE, 0);
 588         WREG32(CGTS_USER_TCC_DISABLE, 0);
 589
 590
 591         num_qd_pipes = R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & INACTIVE_QD_PIPES_MASK) >> 8);
 592         WREG32(VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & DEALLOC_DIST_MASK);
 593         WREG32(VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & VTX_REUSE_DEPTH_MASK);
 594
 595         /* set HW defaults for 3D engine */
 596         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
 597                                      ROQ_IB2_START(0x2b)));
 598
 599         WREG32(CP_MEQ_THRESHOLDS, STQ_SPLIT(0x30));
 600
 601         ta_aux_cntl = RREG32(TA_CNTL_AUX);
 602         WREG32(TA_CNTL_AUX, ta_aux_cntl | DISABLE_CUBE_ANISO);
 603
 604         sx_debug_1 = RREG32(SX_DEBUG_1);
 605         sx_debug_1 |= ENABLE_NEW_SMX_ADDRESS;
 606         WREG32(SX_DEBUG_1, sx_debug_1);
 607
 608         smx_dc_ctl0 = RREG32(SMX_DC_CTL0);
 609         smx_dc_ctl0 &= ~CACHE_DEPTH(0x1ff);
 610         smx_dc_ctl0 |= CACHE_DEPTH((rdev->config.rv770.sx_num_of_sets * 64) - 1);
 611         WREG32(SMX_DC_CTL0, smx_dc_ctl0);
 612
 613         if (rdev->family != CHIP_RV740)
 614                 WREG32(SMX_EVENT_CTL, (ES_FLUSH_CTL(4) |
 615                                        GS_FLUSH_CTL(4) |
 616                                        ACK_FLUSH_CTL(3) |
 617                                        SYNC_FLUSH_CTL));
 618
 619         db_debug3 = RREG32(DB_DEBUG3);
 620         db_debug3 &= ~DB_CLK_OFF_DELAY(0x1f);
 621         switch (rdev->family) {
 622         case CHIP_RV770:
 623         case CHIP_RV740:
 624                 db_debug3 |= DB_CLK_OFF_DELAY(0x1f);
 625                 break;
 626         case CHIP_RV710:
 627         case CHIP_RV730:
 628         default:
 629                 db_debug3 |= DB_CLK_OFF_DELAY(2);
 630                 break;
 631         }
 632         WREG32(DB_DEBUG3, db_debug3);
 633
 634         if (rdev->family != CHIP_RV770) {
 635                 db_debug4 = RREG32(DB_DEBUG4);
 636                 db_debug4 |= DISABLE_TILE_COVERED_FOR_PS_ITER;
 637                 WREG32(DB_DEBUG4, db_debug4);
 638         }
 639
 640         WREG32(SX_EXPORT_BUFFER_SIZES, (COLOR_BUFFER_SIZE((rdev->config.rv770.sx_max_export_size / 4) - 1) |
 641                                         POSITION_BUFFER_SIZE((rdev->config.rv770.sx_max_export_pos_size / 4) - 1) |
 642                                         SMX_BUFFER_SIZE((rdev->config.rv770.sx_max_export_smx_size / 4) - 1)));
 643
 644         WREG32(PA_SC_FIFO_SIZE, (SC_PRIM_FIFO_SIZE(rdev->config.rv770.sc_prim_fifo_size) |
 645                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_hiz_tile_fifo_size) |
 646                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.rv770.sc_earlyz_tile_fifo_fize)));
 647
 648         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 649
 650         WREG32(VGT_NUM_INSTANCES, 1);
 651
 652         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
 653
 654         WREG32(CP_PERFMON_CNTL, 0);
 655
 656         sq_ms_fifo_sizes = (CACHE_FIFO_SIZE(16 * rdev->config.rv770.sq_num_cf_insts) |
 657                             DONE_FIFO_HIWATER(0xe0) |
 658                             ALU_UPDATE_FIFO_HIWATER(0x8));
 659         switch (rdev->family) {
 660         case CHIP_RV770:
 661         case CHIP_RV730:
 662         case CHIP_RV710:
 663                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x1);
 664                 break;
 665         case CHIP_RV740:
 666         default:
 667                 sq_ms_fifo_sizes |= FETCH_FIFO_HIWATER(0x4);
 668                 break;
 669         }
 670         WREG32(SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 671
 672         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 673          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 674          */
 675         sq_config = RREG32(SQ_CONFIG);
 676         sq_config &= ~(PS_PRIO(3) |
 677                        VS_PRIO(3) |
 678                        GS_PRIO(3) |
 679                        ES_PRIO(3));
 680         sq_config |= (DX9_CONSTS |
 681                       VC_ENABLE |
 682                       EXPORT_SRC_C |
 683                       PS_PRIO(0) |
 684                       VS_PRIO(1) |
 685                       GS_PRIO(2) |
 686                       ES_PRIO(3));
 687         if (rdev->family == CHIP_RV710)
 688                 /* no vertex cache */
 689                 sq_config &= ~VC_ENABLE;
 690
 691         WREG32(SQ_CONFIG, sq_config);
 692
 693         WREG32(SQ_GPR_RESOURCE_MGMT_1,  (NUM_PS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 694                                          NUM_VS_GPRS((rdev->config.rv770.max_gprs * 24)/64) |
 695                                          NUM_CLAUSE_TEMP_GPRS(((rdev->config.rv770.max_gprs * 24)/64)/2)));
 696
 697         WREG32(SQ_GPR_RESOURCE_MGMT_2,  (NUM_GS_GPRS((rdev->config.rv770.max_gprs * 7)/64) |
 698                                          NUM_ES_GPRS((rdev->config.rv770.max_gprs * 7)/64)));
 699
 700         sq_thread_resource_mgmt = (NUM_PS_THREADS((rdev->config.rv770.max_threads * 4)/8) |
 701                                    NUM_VS_THREADS((rdev->config.rv770.max_threads * 2)/8) |
 702                                    NUM_ES_THREADS((rdev->config.rv770.max_threads * 1)/8));
 703         if (((rdev->config.rv770.max_threads * 1) / 8) > rdev->config.rv770.max_gs_threads)
 704                 sq_thread_resource_mgmt |= NUM_GS_THREADS(rdev->config.rv770.max_gs_threads);
 705         else
 706                 sq_thread_resource_mgmt |= NUM_GS_THREADS((rdev->config.rv770.max_gs_threads * 1)/8);
 707         WREG32(SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
 708
 709         WREG32(SQ_STACK_RESOURCE_MGMT_1, (NUM_PS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 710                                                      NUM_VS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 711
 712         WREG32(SQ_STACK_RESOURCE_MGMT_2, (NUM_GS_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4) |
 713                                                      NUM_ES_STACK_ENTRIES((rdev->config.rv770.max_stack_entries * 1)/4)));
 714
 715         sq_dyn_gpr_size_simd_ab_0 = (SIMDA_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 716                                      SIMDA_RING1((rdev->config.rv770.max_gprs * 38)/64) |
 717                                      SIMDB_RING0((rdev->config.rv770.max_gprs * 38)/64) |
 718                                      SIMDB_RING1((rdev->config.rv770.max_gprs * 38)/64));
 719
 720         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
 721         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
 722         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
 723         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
 724         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
 725         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
 726         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
 727         WREG32(SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
 728
 729         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
 730                                           FORCE_EOV_MAX_REZ_CNT(255)));
 731
 732         if (rdev->family == CHIP_RV710)
 733                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(TC_ONLY) |
 734                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 735         else
 736                 WREG32(VGT_CACHE_INVALIDATION, (CACHE_INVALIDATION(VC_AND_TC) |
 737                                                 AUTO_INVLD_EN(ES_AND_GS_AUTO)));
 738
 739         switch (rdev->family) {
 740         case CHIP_RV770:
 741         case CHIP_RV730:
 742         case CHIP_RV740:
 743                 gs_prim_buffer_depth = 384;
 744                 break;
 745         case CHIP_RV710:
 746                 gs_prim_buffer_depth = 128;
 747                 break;
 748         default:
 749                 break;
 750         }
 751
 752         num_gs_verts_per_thread = rdev->config.rv770.max_pipes * 16;
 753         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
 754         /* Max value for this is 256 */
 755         if (vgt_gs_per_es > 256)
 756                 vgt_gs_per_es = 256;
 757
 758         WREG32(VGT_ES_PER_GS, 128);
 759         WREG32(VGT_GS_PER_ES, vgt_gs_per_es);
 760         WREG32(VGT_GS_PER_VS, 2);
 761
 762         /* more default values. 2D/3D driver should adjust as needed */
 763         WREG32(VGT_GS_VERTEX_REUSE, 16);
 764         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
 765         WREG32(VGT_STRMOUT_EN, 0);
 766         WREG32(SX_MISC, 0);
 767         WREG32(PA_SC_MODE_CNTL, 0);
 768         WREG32(PA_SC_EDGERULE, 0xaaaaaaaa);
 769         WREG32(PA_SC_AA_CONFIG, 0);
 770         WREG32(PA_SC_CLIPRECT_RULE, 0xffff);
 771         WREG32(PA_SC_LINE_STIPPLE, 0);
 772         WREG32(SPI_INPUT_Z, 0);
 773         WREG32(SPI_PS_IN_CONTROL_0, NUM_INTERP(2));
 774         WREG32(CB_COLOR7_FRAG, 0);
 775
 776         /* clear render buffer base addresses */
 777         WREG32(CB_COLOR0_BASE, 0);
 778         WREG32(CB_COLOR1_BASE, 0);
 779         WREG32(CB_COLOR2_BASE, 0);
 780         WREG32(CB_COLOR3_BASE, 0);
 781         WREG32(CB_COLOR4_BASE, 0);
 782         WREG32(CB_COLOR5_BASE, 0);
 783         WREG32(CB_COLOR6_BASE, 0);
 784         WREG32(CB_COLOR7_BASE, 0);
 785
 786         WREG32(TCP_CNTL, 0);
 787
 788         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
 789         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
 790
 791         WREG32(PA_SC_MULTI_CHIP_CNTL, 0);
 792
 793         WREG32(PA_CL_ENHANCE, (CLIP_VTX_REORDER_ENA |
 794                                           NUM_CLIP_SEQ(3)));
 795
 796 }
 797
 798 void r700_vram_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc)
 799 {
 800         u64 size_bf, size_af;
 801
 802         if (mc->mc_vram_size > 0xE0000000) {
 803                 /* leave room for at least 512M GTT */
 804                 dev_warn(rdev->dev, "limiting VRAM\n");
 805                 mc->real_vram_size = 0xE0000000;
 806                 mc->mc_vram_size = 0xE0000000;
 807         }
 808         if (rdev->flags & RADEON_IS_AGP) {
 809                 size_bf = mc->gtt_start;
 810                 size_af = 0xFFFFFFFF - mc->gtt_end;
 811                 if (size_bf > size_af) {
 812                         if (mc->mc_vram_size > size_bf) {
 813                                 dev_warn(rdev->dev, "limiting VRAM\n");
 814                                 mc->real_vram_size = size_bf;
 815                                 mc->mc_vram_size = size_bf;
 816                         }
 817                         mc->vram_start = mc->gtt_start - mc->mc_vram_size;
 818                 } else {
 819                         if (mc->mc_vram_size > size_af) {
 820                                 dev_warn(rdev->dev, "limiting VRAM\n");
 821                                 mc->real_vram_size = size_af;
 822                                 mc->mc_vram_size = size_af;
 823                         }
 824                         mc->vram_start = mc->gtt_end + 1;
 825                 }
 826                 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
 827                 dev_info(rdev->dev, "VRAM: %lluM 0x%08llX - 0x%08llX (%lluM used)\n",
 828                                 mc->mc_vram_size >> 20, mc->vram_start,
 829                                 mc->vram_end, mc->real_vram_size >> 20);
 830         } else {
 831                 radeon_vram_location(rdev, &rdev->mc, 0);
 832                 rdev->mc.gtt_base_align = 0;
 833                 radeon_gtt_location(rdev, mc);
 834         }
 835 }
 836
 837 int rv770_mc_init(struct radeon_device *rdev)
 838 {
 839         u32 tmp;
 840         int chansize, numchan;
 841
 842         /* Get VRAM informations */
 843         rdev->mc.vram_is_ddr = true;
 844         tmp = RREG32(MC_ARB_RAMCFG);
 845         if (tmp & CHANSIZE_OVERRIDE) {
 846                 chansize = 16;
 847         } else if (tmp & CHANSIZE_MASK) {
 848                 chansize = 64;
 849         } else {
 850                 chansize = 32;
 851         }
 852         tmp = RREG32(MC_SHARED_CHMAP);
 853         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 854         case 0:
 855         default:
 856                 numchan = 1;
 857                 break;
 858         case 1:
 859                 numchan = 2;
 860                 break;
 861         case 2:
 862                 numchan = 4;
 863                 break;
 864         case 3:
 865                 numchan = 8;
 866                 break;
 867         }
 868         rdev->mc.vram_width = numchan * chansize;
 869         /* Could aper size report 0 ? */
 870         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
 871         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
 872         /* Setup GPU memory space */
 873         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE);
 874         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE);
 875         rdev->mc.visible_vram_size = rdev->mc.aper_size;
 876         r700_vram_gtt_location(rdev, &rdev->mc);
 877         radeon_update_bandwidth_info(rdev);
 878
 879         return 0;
 880 }
 881
 882 static int rv770_startup(struct radeon_device *rdev)
 883 {
 884         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
 885         int r;
 886
 887         /* enable pcie gen2 link */
 888         rv770_pcie_gen2_enable(rdev);
 889
 890         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->rlc_fw) {
 891                 r = r600_init_microcode(rdev);
 892                 if (r) {
 893                         DRM_ERROR("Failed to load firmware!\n");
 894                         return r;
 895                 }
 896         }
 897
 898         r = r600_vram_scratch_init(rdev);
 899         if (r)
 900                 return r;
 901
 902         rv770_mc_program(rdev);
 903         if (rdev->flags & RADEON_IS_AGP) {
 904                 rv770_agp_enable(rdev);
 905         } else {
 906                 r = rv770_pcie_gart_enable(rdev);
 907                 if (r)
 908                         return r;
 909         }
 910
 911         rv770_gpu_init(rdev);
 912         r = r600_blit_init(rdev);
 913         if (r) {
 914                 r600_blit_fini(rdev);
 915                 rdev->asic->copy.copy = NULL;
 916                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
 917         }
 918
 919         /* allocate wb buffer */
 920         r = radeon_wb_init(rdev);
 921         if (r)
 922                 return r;
 923
 924         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
 925         if (r) {
 926                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
 927                 return r;
 928         }
 929
 930         /* Enable IRQ */
 931         r = r600_irq_init(rdev);
 932         if (r) {
 933                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
 934                 radeon_irq_kms_fini(rdev);
 935                 return r;
 936         }
 937         r600_irq_set(rdev);
 938
 939         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
 940                              R600_CP_RB_RPTR, R600_CP_RB_WPTR,
 941                              0, 0xfffff, RADEON_CP_PACKET2);
 942         if (r)
 943                 return r;
 944         r = rv770_cp_load_microcode(rdev);
 945         if (r)
 946                 return r;
 947         r = r600_cp_resume(rdev);
 948         if (r)
 949                 return r;
 950
 951         r = radeon_ib_pool_start(rdev);
 952         if (r)
 953                 return r;
 954
 955         r = radeon_ib_ring_tests(rdev);
 956         if (r)
 957                 return r;
 958
 959         return 0;
 960 }
 961
 962 int rv770_resume(struct radeon_device *rdev)
 963 {
 964         int r;
 965
 966         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
 967          * posting will perform necessary task to bring back GPU into good
 968          * shape.
 969          */
 970         /* post card */
 971         atom_asic_init(rdev->mode_info.atom_context);
 972
 973         rdev->accel_working = true;
 974         r = rv770_startup(rdev);
 975         if (r) {
 976                 DRM_ERROR("r600 startup failed on resume\n");
 977                 rdev->accel_working = false;
 978                 return r;
 979         }
 980
 981         r = r600_audio_init(rdev);
 982         if (r) {
 983                 dev_err(rdev->dev, "radeon: audio init failed\n");
 984                 return r;
 985         }
 986
 987         return r;
 988
 989 }
 990
 991 int rv770_suspend(struct radeon_device *rdev)
 992 {
 993         r600_audio_fini(rdev);
 994         radeon_ib_pool_suspend(rdev);
 995         r600_blit_suspend(rdev);
 996         /* FIXME: we should wait for ring to be empty */
 997         r700_cp_stop(rdev);
 998         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
 999         r600_irq_suspend(rdev);
1000         radeon_wb_disable(rdev);
1001         rv770_pcie_gart_disable(rdev);
1002
1003         return 0;
1004 }
1005
1006 /* Plan is to move initialization in that function and use
1007  * helper function so that radeon_device_init pretty much
1008  * do nothing more than calling asic specific function. This
1009  * should also allow to remove a bunch of callback function
1010  * like vram_info.
1011  */
1012 int rv770_init(struct radeon_device *rdev)
1013 {
1014         int r;
1015
1016         /* Read BIOS */
1017         if (!radeon_get_bios(rdev)) {
1018                 if (ASIC_IS_AVIVO(rdev))
1019                         return -EINVAL;
1020         }
1021         /* Must be an ATOMBIOS */
1022         if (!rdev->is_atom_bios) {
1023                 dev_err(rdev->dev, "Expecting atombios for R600 GPU\n");
1024                 return -EINVAL;
1025         }
1026         r = radeon_atombios_init(rdev);
1027         if (r)
1028                 return r;
1029         /* Post card if necessary */
1030         if (!radeon_card_posted(rdev)) {
1031                 if (!rdev->bios) {
1032                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
1033                         return -EINVAL;
1034                 }
1035                 DRM_INFO("GPU not posted. posting now...\n");
1036                 atom_asic_init(rdev->mode_info.atom_context);
1037         }
1038         /* Initialize scratch registers */
1039         r600_scratch_init(rdev);
1040         /* Initialize surface registers */
1041         radeon_surface_init(rdev);
1042         /* Initialize clocks */
1043         radeon_get_clock_info(rdev->ddev);
1044         /* Fence driver */
1045         r = radeon_fence_driver_init(rdev);
1046         if (r)
1047                 return r;
1048         /* initialize AGP */
1049         if (rdev->flags & RADEON_IS_AGP) {
1050                 r = radeon_agp_init(rdev);
1051                 if (r)
1052                         radeon_agp_disable(rdev);
1053         }
1054         r = rv770_mc_init(rdev);
1055         if (r)
1056                 return r;
1057         /* Memory manager */
1058         r = radeon_bo_init(rdev);
1059         if (r)
1060                 return r;
1061
1062         r = radeon_irq_kms_init(rdev);
1063         if (r)
1064                 return r;
1065
1066         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ring_obj = NULL;
1067         r600_ring_init(rdev, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX], 1024 * 1024);
1068
1069         rdev->ih.ring_obj = NULL;
1070         r600_ih_ring_init(rdev, 64 * 1024);
1071
1072         r = r600_pcie_gart_init(rdev);
1073         if (r)
1074                 return r;
1075
1076         r = radeon_ib_pool_init(rdev);
1077         rdev->accel_working = true;
1078         if (r) {
1079                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
1080                 rdev->accel_working = false;
1081         }
1082
1083         r = rv770_startup(rdev);
1084         if (r) {
1085                 dev_err(rdev->dev, "disabling GPU acceleration\n");
1086                 r700_cp_fini(rdev);
1087                 r600_irq_fini(rdev);
1088                 radeon_wb_fini(rdev);
1089                 r100_ib_fini(rdev);
1090                 radeon_irq_kms_fini(rdev);
1091                 rv770_pcie_gart_fini(rdev);
1092                 rdev->accel_working = false;
1093         }
1094
1095         r = r600_audio_init(rdev);
1096         if (r) {
1097                 dev_err(rdev->dev, "radeon: audio init failed\n");
1098                 return r;
1099         }
1100
1101         return 0;
1102 }
1103
1104 void rv770_fini(struct radeon_device *rdev)
1105 {
1106         r600_blit_fini(rdev);
1107         r700_cp_fini(rdev);
1108         r600_irq_fini(rdev);
1109         radeon_wb_fini(rdev);
1110         r100_ib_fini(rdev);
1111         radeon_irq_kms_fini(rdev);
1112         rv770_pcie_gart_fini(rdev);
1113         r600_vram_scratch_fini(rdev);
1114         radeon_gem_fini(rdev);
1115         radeon_fence_driver_fini(rdev);
1116         radeon_agp_fini(rdev);
1117         radeon_bo_fini(rdev);
1118         radeon_atombios_fini(rdev);
1119         kfree(rdev->bios);
1120         rdev->bios = NULL;
1121 }
1122
1123 static void rv770_pcie_gen2_enable(struct radeon_device *rdev)
1124 {
1125         u32 link_width_cntl, lanes, speed_cntl, tmp;
1126         u16 link_cntl2;
1127
1128         if (radeon_pcie_gen2 == 0)
1129                 return;
1130
1131         if (rdev->flags & RADEON_IS_IGP)
1132                 return;
1133
1134         if (!(rdev->flags & RADEON_IS_PCIE))
1135                 return;
1136
1137         /* x2 cards have a special sequence */
1138         if (ASIC_IS_X2(rdev))
1139                 return;
1140
1141         /* advertise upconfig capability */
1142         link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1143         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
1144         WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1145         link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1146         if (link_width_cntl & LC_RENEGOTIATION_SUPPORT) {
1147                 lanes = (link_width_cntl & LC_LINK_WIDTH_RD_MASK) >> LC_LINK_WIDTH_RD_SHIFT;
1148                 link_width_cntl &= ~(LC_LINK_WIDTH_MASK |
1149                                      LC_RECONFIG_ARC_MISSING_ESCAPE);
1150                 link_width_cntl |= lanes | LC_RECONFIG_NOW |
1151                         LC_RENEGOTIATE_EN | LC_UPCONFIGURE_SUPPORT;
1152                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1153         } else {
1154                 link_width_cntl |= LC_UPCONFIGURE_DIS;
1155                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1156         }
1157
1158         speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1159         if ((speed_cntl & LC_OTHER_SIDE_EVER_SENT_GEN2) &&
1160             (speed_cntl & LC_OTHER_SIDE_SUPPORTS_GEN2)) {
1161
1162                 tmp = RREG32(0x541c);
1163                 WREG32(0x541c, tmp | 0x8);
1164                 WREG32(MM_CFGREGS_CNTL, MM_WR_TO_CFG_EN);
1165                 link_cntl2 = RREG16(0x4088);
1166                 link_cntl2 &= ~TARGET_LINK_SPEED_MASK;
1167                 link_cntl2 |= 0x2;
1168                 WREG16(0x4088, link_cntl2);
1169                 WREG32(MM_CFGREGS_CNTL, 0);
1170
1171                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1172                 speed_cntl &= ~LC_TARGET_LINK_SPEED_OVERRIDE_EN;
1173                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1174
1175                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1176                 speed_cntl |= LC_CLR_FAILED_SPD_CHANGE_CNT;
1177                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1178
1179                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1180                 speed_cntl &= ~LC_CLR_FAILED_SPD_CHANGE_CNT;
1181                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1182
1183                 speed_cntl = RREG32_PCIE_P(PCIE_LC_SPEED_CNTL);
1184                 speed_cntl |= LC_GEN2_EN_STRAP;
1185                 WREG32_PCIE_P(PCIE_LC_SPEED_CNTL, speed_cntl);
1186
1187         } else {
1188                 link_width_cntl = RREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL);
1189                 /* XXX: only disable it if gen1 bridge vendor == 0x111d or 0x1106 */
1190                 if (1)
1191                         link_width_cntl |= LC_UPCONFIGURE_DIS;
1192                 else
1193                         link_width_cntl &= ~LC_UPCONFIGURE_DIS;
1194                 WREG32_PCIE_P(PCIE_LC_LINK_WIDTH_CNTL, link_width_cntl);
1195         }
1196 }