SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/gpu/drm/radeon/r600_cp.c

   1 /*
   2  * Copyright 2008-2009 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * 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) AND/OR ITS SUPPLIERS 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 OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *     Dave Airlie <airlied@redhat.com>
  26  *     Alex Deucher <alexander.deucher@amd.com>
  27  */
  28
  29 #include "drmP.h"
  30 #include "drm.h"
  31 #include "radeon_drm.h"
  32 #include "radeon_drv.h"
  33
  34 #define PFP_UCODE_SIZE 576
  35 #define PM4_UCODE_SIZE 1792
  36 #define R700_PFP_UCODE_SIZE 848
  37 #define R700_PM4_UCODE_SIZE 1360
  38
  39 /* Firmware Names */
  40 MODULE_FIRMWARE("radeon/R600_pfp.bin");
  41 MODULE_FIRMWARE("radeon/R600_me.bin");
  42 MODULE_FIRMWARE("radeon/RV610_pfp.bin");
  43 MODULE_FIRMWARE("radeon/RV610_me.bin");
  44 MODULE_FIRMWARE("radeon/RV630_pfp.bin");
  45 MODULE_FIRMWARE("radeon/RV630_me.bin");
  46 MODULE_FIRMWARE("radeon/RV620_pfp.bin");
  47 MODULE_FIRMWARE("radeon/RV620_me.bin");
  48 MODULE_FIRMWARE("radeon/RV635_pfp.bin");
  49 MODULE_FIRMWARE("radeon/RV635_me.bin");
  50 MODULE_FIRMWARE("radeon/RV670_pfp.bin");
  51 MODULE_FIRMWARE("radeon/RV670_me.bin");
  52 MODULE_FIRMWARE("radeon/RS780_pfp.bin");
  53 MODULE_FIRMWARE("radeon/RS780_me.bin");
  54 MODULE_FIRMWARE("radeon/RV770_pfp.bin");
  55 MODULE_FIRMWARE("radeon/RV770_me.bin");
  56 MODULE_FIRMWARE("radeon/RV730_pfp.bin");
  57 MODULE_FIRMWARE("radeon/RV730_me.bin");
  58 MODULE_FIRMWARE("radeon/RV710_pfp.bin");
  59 MODULE_FIRMWARE("radeon/RV710_me.bin");
  60
  61
  62 int r600_cs_legacy(struct drm_device *dev, void *data, struct drm_file *filp,
  63                         unsigned family, u32 *ib, int *l);
  64 void r600_cs_legacy_init(void);
  65
  66
  67 # define ATI_PCIGART_PAGE_SIZE          4096    /**< PCI GART page size */
  68 # define ATI_PCIGART_PAGE_MASK          (~(ATI_PCIGART_PAGE_SIZE-1))
  69
  70 #define R600_PTE_VALID     (1 << 0)
  71 #define R600_PTE_SYSTEM    (1 << 1)
  72 #define R600_PTE_SNOOPED   (1 << 2)
  73 #define R600_PTE_READABLE  (1 << 5)
  74 #define R600_PTE_WRITEABLE (1 << 6)
  75
  76 /* MAX values used for gfx init */
  77 #define R6XX_MAX_SH_GPRS           256
  78 #define R6XX_MAX_TEMP_GPRS         16
  79 #define R6XX_MAX_SH_THREADS        256
  80 #define R6XX_MAX_SH_STACK_ENTRIES  4096
  81 #define R6XX_MAX_BACKENDS          8
  82 #define R6XX_MAX_BACKENDS_MASK     0xff
  83 #define R6XX_MAX_SIMDS             8
  84 #define R6XX_MAX_SIMDS_MASK        0xff
  85 #define R6XX_MAX_PIPES             8
  86 #define R6XX_MAX_PIPES_MASK        0xff
  87
  88 #define R7XX_MAX_SH_GPRS           256
  89 #define R7XX_MAX_TEMP_GPRS         16
  90 #define R7XX_MAX_SH_THREADS        256
  91 #define R7XX_MAX_SH_STACK_ENTRIES  4096
  92 #define R7XX_MAX_BACKENDS          8
  93 #define R7XX_MAX_BACKENDS_MASK     0xff
  94 #define R7XX_MAX_SIMDS             16
  95 #define R7XX_MAX_SIMDS_MASK        0xffff
  96 #define R7XX_MAX_PIPES             8
  97 #define R7XX_MAX_PIPES_MASK        0xff
  98
  99 static int r600_do_wait_for_fifo(drm_radeon_private_t *dev_priv, int entries)
 100 {
 101         int i;
 102
 103         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 104
 105         for (i = 0; i < dev_priv->usec_timeout; i++) {
 106                 int slots;
 107                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 108                         slots = (RADEON_READ(R600_GRBM_STATUS)
 109                                  & R700_CMDFIFO_AVAIL_MASK);
 110                 else
 111                         slots = (RADEON_READ(R600_GRBM_STATUS)
 112                                  & R600_CMDFIFO_AVAIL_MASK);
 113                 if (slots >= entries)
 114                         return 0;
 115                 DRM_UDELAY(1);
 116         }
 117         DRM_INFO("wait for fifo failed status : 0x%08X 0x%08X\n",
 118                  RADEON_READ(R600_GRBM_STATUS),
 119                  RADEON_READ(R600_GRBM_STATUS2));
 120
 121         return -EBUSY;
 122 }
 123
 124 static int r600_do_wait_for_idle(drm_radeon_private_t *dev_priv)
 125 {
 126         int i, ret;
 127
 128         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 129
 130         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 131                 ret = r600_do_wait_for_fifo(dev_priv, 8);
 132         else
 133                 ret = r600_do_wait_for_fifo(dev_priv, 16);
 134         if (ret)
 135                 return ret;
 136         for (i = 0; i < dev_priv->usec_timeout; i++) {
 137                 if (!(RADEON_READ(R600_GRBM_STATUS) & R600_GUI_ACTIVE))
 138                         return 0;
 139                 DRM_UDELAY(1);
 140         }
 141         DRM_INFO("wait idle failed status : 0x%08X 0x%08X\n",
 142                  RADEON_READ(R600_GRBM_STATUS),
 143                  RADEON_READ(R600_GRBM_STATUS2));
 144
 145         return -EBUSY;
 146 }
 147
 148 void r600_page_table_cleanup(struct drm_device *dev, struct drm_ati_pcigart_info *gart_info)
 149 {
 150         struct drm_sg_mem *entry = dev->sg;
 151         int max_pages;
 152         int pages;
 153         int i;
 154
 155         if (!entry)
 156                 return;
 157
 158         if (gart_info->bus_addr) {
 159                 max_pages = (gart_info->table_size / sizeof(u64));
 160                 pages = (entry->pages <= max_pages)
 161                   ? entry->pages : max_pages;
 162
 163                 for (i = 0; i < pages; i++) {
 164                         if (!entry->busaddr[i])
 165                                 break;
 166                         pci_unmap_page(dev->pdev, entry->busaddr[i],
 167                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 168                 }
 169                 if (gart_info->gart_table_location == DRM_ATI_GART_MAIN)
 170                         gart_info->bus_addr = 0;
 171         }
 172 }
 173
 174 /* R600 has page table setup */
 175 int r600_page_table_init(struct drm_device *dev)
 176 {
 177         drm_radeon_private_t *dev_priv = dev->dev_private;
 178         struct drm_ati_pcigart_info *gart_info = &dev_priv->gart_info;
 179         struct drm_local_map *map = &gart_info->mapping;
 180         struct drm_sg_mem *entry = dev->sg;
 181         int ret = 0;
 182         int i, j;
 183         int pages;
 184         u64 page_base;
 185         dma_addr_t entry_addr;
 186         int max_ati_pages, max_real_pages, gart_idx;
 187
 188         /* okay page table is available - lets rock */
 189         max_ati_pages = (gart_info->table_size / sizeof(u64));
 190         max_real_pages = max_ati_pages / (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE);
 191
 192         pages = (entry->pages <= max_real_pages) ?
 193                 entry->pages : max_real_pages;
 194
 195         memset_io((void __iomem *)map->handle, 0, max_ati_pages * sizeof(u64));
 196
 197         gart_idx = 0;
 198         for (i = 0; i < pages; i++) {
 199                 entry->busaddr[i] = pci_map_page(dev->pdev,
 200                                                  entry->pagelist[i], 0,
 201                                                  PAGE_SIZE,
 202                                                  PCI_DMA_BIDIRECTIONAL);
 203                 if (entry->busaddr[i] == 0) {
 204                         DRM_ERROR("unable to map PCIGART pages!\n");
 205                         r600_page_table_cleanup(dev, gart_info);
 206                         goto done;
 207                 }
 208                 entry_addr = entry->busaddr[i];
 209                 for (j = 0; j < (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE); j++) {
 210                         page_base = (u64) entry_addr & ATI_PCIGART_PAGE_MASK;
 211                         page_base |= R600_PTE_VALID | R600_PTE_SYSTEM | R600_PTE_SNOOPED;
 212                         page_base |= R600_PTE_READABLE | R600_PTE_WRITEABLE;
 213
 214                         DRM_WRITE64(map, gart_idx * sizeof(u64), page_base);
 215
 216                         gart_idx++;
 217
 218                         if ((i % 128) == 0)
 219                                 DRM_DEBUG("page entry %d: 0x%016llx\n",
 220                                     i, (unsigned long long)page_base);
 221                         entry_addr += ATI_PCIGART_PAGE_SIZE;
 222                 }
 223         }
 224         ret = 1;
 225 done:
 226         return ret;
 227 }
 228
 229 static void r600_vm_flush_gart_range(struct drm_device *dev)
 230 {
 231         drm_radeon_private_t *dev_priv = dev->dev_private;
 232         u32 resp, countdown = 1000;
 233         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 234         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 235         RADEON_WRITE(R600_VM_CONTEXT0_REQUEST_RESPONSE, 2);
 236
 237         do {
 238                 resp = RADEON_READ(R600_VM_CONTEXT0_REQUEST_RESPONSE);
 239                 countdown--;
 240                 DRM_UDELAY(1);
 241         } while (((resp & 0xf0) == 0) && countdown);
 242 }
 243
 244 static void r600_vm_init(struct drm_device *dev)
 245 {
 246         drm_radeon_private_t *dev_priv = dev->dev_private;
 247         /* initialise the VM to use the page table we constructed up there */
 248         u32 vm_c0, i;
 249         u32 mc_rd_a;
 250         u32 vm_l2_cntl, vm_l2_cntl3;
 251         /* okay set up the PCIE aperture type thingo */
 252         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 253         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 254         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 255
 256         /* setup MC RD a */
 257         mc_rd_a = R600_MCD_L1_TLB | R600_MCD_L1_FRAG_PROC | R600_MCD_SYSTEM_ACCESS_MODE_IN_SYS |
 258                 R600_MCD_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | R600_MCD_EFFECTIVE_L1_TLB_SIZE(5) |
 259                 R600_MCD_EFFECTIVE_L1_QUEUE_SIZE(5) | R600_MCD_WAIT_L2_QUERY;
 260
 261         RADEON_WRITE(R600_MCD_RD_A_CNTL, mc_rd_a);
 262         RADEON_WRITE(R600_MCD_RD_B_CNTL, mc_rd_a);
 263
 264         RADEON_WRITE(R600_MCD_WR_A_CNTL, mc_rd_a);
 265         RADEON_WRITE(R600_MCD_WR_B_CNTL, mc_rd_a);
 266
 267         RADEON_WRITE(R600_MCD_RD_GFX_CNTL, mc_rd_a);
 268         RADEON_WRITE(R600_MCD_WR_GFX_CNTL, mc_rd_a);
 269
 270         RADEON_WRITE(R600_MCD_RD_SYS_CNTL, mc_rd_a);
 271         RADEON_WRITE(R600_MCD_WR_SYS_CNTL, mc_rd_a);
 272
 273         RADEON_WRITE(R600_MCD_RD_HDP_CNTL, mc_rd_a | R600_MCD_L1_STRICT_ORDERING);
 274         RADEON_WRITE(R600_MCD_WR_HDP_CNTL, mc_rd_a /*| R600_MCD_L1_STRICT_ORDERING*/);
 275
 276         RADEON_WRITE(R600_MCD_RD_PDMA_CNTL, mc_rd_a);
 277         RADEON_WRITE(R600_MCD_WR_PDMA_CNTL, mc_rd_a);
 278
 279         RADEON_WRITE(R600_MCD_RD_SEM_CNTL, mc_rd_a | R600_MCD_SEMAPHORE_MODE);
 280         RADEON_WRITE(R600_MCD_WR_SEM_CNTL, mc_rd_a);
 281
 282         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 283         vm_l2_cntl |= R600_VM_L2_CNTL_QUEUE_SIZE(7);
 284         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 285
 286         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 287         vm_l2_cntl3 = (R600_VM_L2_CNTL3_BANK_SELECT_0(0) |
 288                        R600_VM_L2_CNTL3_BANK_SELECT_1(1) |
 289                        R600_VM_L2_CNTL3_CACHE_UPDATE_MODE(2));
 290         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 291
 292         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 293
 294         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 295
 296         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 297
 298         /* disable all other contexts */
 299         for (i = 1; i < 8; i++)
 300                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 301
 302         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 303         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 304         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 305
 306         r600_vm_flush_gart_range(dev);
 307 }
 308
 309 static int r600_cp_init_microcode(drm_radeon_private_t *dev_priv)
 310 {
 311         struct platform_device *pdev;
 312         const char *chip_name;
 313         size_t pfp_req_size, me_req_size;
 314         char fw_name[30];
 315         int err;
 316
 317         pdev = platform_device_register_simple("r600_cp", 0, NULL, 0);
 318         err = IS_ERR(pdev);
 319         if (err) {
 320                 printk(KERN_ERR "r600_cp: Failed to register firmware\n");
 321                 return -EINVAL;
 322         }
 323
 324         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
 325         case CHIP_R600:  chip_name = "R600";  break;
 326         case CHIP_RV610: chip_name = "RV610"; break;
 327         case CHIP_RV630: chip_name = "RV630"; break;
 328         case CHIP_RV620: chip_name = "RV620"; break;
 329         case CHIP_RV635: chip_name = "RV635"; break;
 330         case CHIP_RV670: chip_name = "RV670"; break;
 331         case CHIP_RS780:
 332         case CHIP_RS880: chip_name = "RS780"; break;
 333         case CHIP_RV770: chip_name = "RV770"; break;
 334         case CHIP_RV730:
 335         case CHIP_RV740: chip_name = "RV730"; break;
 336         case CHIP_RV710: chip_name = "RV710"; break;
 337         default:         BUG();
 338         }
 339
 340         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770) {
 341                 pfp_req_size = R700_PFP_UCODE_SIZE * 4;
 342                 me_req_size = R700_PM4_UCODE_SIZE * 4;
 343         } else {
 344                 pfp_req_size = PFP_UCODE_SIZE * 4;
 345                 me_req_size = PM4_UCODE_SIZE * 12;
 346         }
 347
 348         DRM_INFO("Loading %s CP Microcode\n", chip_name);
 349
 350         snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
 351         err = request_firmware(&dev_priv->pfp_fw, fw_name, &pdev->dev);
 352         if (err)
 353                 goto out;
 354         if (dev_priv->pfp_fw->size != pfp_req_size) {
 355                 printk(KERN_ERR
 356                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
 357                        dev_priv->pfp_fw->size, fw_name);
 358                 err = -EINVAL;
 359                 goto out;
 360         }
 361
 362         snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
 363         err = request_firmware(&dev_priv->me_fw, fw_name, &pdev->dev);
 364         if (err)
 365                 goto out;
 366         if (dev_priv->me_fw->size != me_req_size) {
 367                 printk(KERN_ERR
 368                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
 369                        dev_priv->me_fw->size, fw_name);
 370                 err = -EINVAL;
 371         }
 372 out:
 373         platform_device_unregister(pdev);
 374
 375         if (err) {
 376                 if (err != -EINVAL)
 377                         printk(KERN_ERR
 378                                "r600_cp: Failed to load firmware \"%s\"\n",
 379                                fw_name);
 380                 release_firmware(dev_priv->pfp_fw);
 381                 dev_priv->pfp_fw = NULL;
 382                 release_firmware(dev_priv->me_fw);
 383                 dev_priv->me_fw = NULL;
 384         }
 385         return err;
 386 }
 387
 388 static void r600_cp_load_microcode(drm_radeon_private_t *dev_priv)
 389 {
 390         const __be32 *fw_data;
 391         int i;
 392
 393         if (!dev_priv->me_fw || !dev_priv->pfp_fw)
 394                 return;
 395
 396         r600_do_cp_stop(dev_priv);
 397
 398         RADEON_WRITE(R600_CP_RB_CNTL,
 399                      R600_RB_NO_UPDATE |
 400                      R600_RB_BLKSZ(15) |
 401                      R600_RB_BUFSZ(3));
 402
 403         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 404         RADEON_READ(R600_GRBM_SOFT_RESET);
 405         DRM_UDELAY(15000);
 406         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 407
 408         fw_data = (const __be32 *)dev_priv->me_fw->data;
 409         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 410         for (i = 0; i < PM4_UCODE_SIZE * 3; i++)
 411                 RADEON_WRITE(R600_CP_ME_RAM_DATA,
 412                              be32_to_cpup(fw_data++));
 413
 414         fw_data = (const __be32 *)dev_priv->pfp_fw->data;
 415         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 416         for (i = 0; i < PFP_UCODE_SIZE; i++)
 417                 RADEON_WRITE(R600_CP_PFP_UCODE_DATA,
 418                              be32_to_cpup(fw_data++));
 419
 420         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 421         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 422         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 423
 424 }
 425
 426 static void r700_vm_init(struct drm_device *dev)
 427 {
 428         drm_radeon_private_t *dev_priv = dev->dev_private;
 429         /* initialise the VM to use the page table we constructed up there */
 430         u32 vm_c0, i;
 431         u32 mc_vm_md_l1;
 432         u32 vm_l2_cntl, vm_l2_cntl3;
 433         /* okay set up the PCIE aperture type thingo */
 434         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 435         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 436         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 437
 438         mc_vm_md_l1 = R700_ENABLE_L1_TLB |
 439             R700_ENABLE_L1_FRAGMENT_PROCESSING |
 440             R700_SYSTEM_ACCESS_MODE_IN_SYS |
 441             R700_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 442             R700_EFFECTIVE_L1_TLB_SIZE(5) |
 443             R700_EFFECTIVE_L1_QUEUE_SIZE(5);
 444
 445         RADEON_WRITE(R700_MC_VM_MD_L1_TLB0_CNTL, mc_vm_md_l1);
 446         RADEON_WRITE(R700_MC_VM_MD_L1_TLB1_CNTL, mc_vm_md_l1);
 447         RADEON_WRITE(R700_MC_VM_MD_L1_TLB2_CNTL, mc_vm_md_l1);
 448         RADEON_WRITE(R700_MC_VM_MB_L1_TLB0_CNTL, mc_vm_md_l1);
 449         RADEON_WRITE(R700_MC_VM_MB_L1_TLB1_CNTL, mc_vm_md_l1);
 450         RADEON_WRITE(R700_MC_VM_MB_L1_TLB2_CNTL, mc_vm_md_l1);
 451         RADEON_WRITE(R700_MC_VM_MB_L1_TLB3_CNTL, mc_vm_md_l1);
 452
 453         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 454         vm_l2_cntl |= R700_VM_L2_CNTL_QUEUE_SIZE(7);
 455         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 456
 457         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 458         vm_l2_cntl3 = R700_VM_L2_CNTL3_BANK_SELECT(0) | R700_VM_L2_CNTL3_CACHE_UPDATE_MODE(2);
 459         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 460
 461         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 462
 463         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 464
 465         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 466
 467         /* disable all other contexts */
 468         for (i = 1; i < 8; i++)
 469                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 470
 471         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 472         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 473         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 474
 475         r600_vm_flush_gart_range(dev);
 476 }
 477
 478 static void r700_cp_load_microcode(drm_radeon_private_t *dev_priv)
 479 {
 480         const __be32 *fw_data;
 481         int i;
 482
 483         if (!dev_priv->me_fw || !dev_priv->pfp_fw)
 484                 return;
 485
 486         r600_do_cp_stop(dev_priv);
 487
 488         RADEON_WRITE(R600_CP_RB_CNTL,
 489                      R600_RB_NO_UPDATE |
 490                      (15 << 8) |
 491                      (3 << 0));
 492
 493         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 494         RADEON_READ(R600_GRBM_SOFT_RESET);
 495         DRM_UDELAY(15000);
 496         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 497
 498         fw_data = (const __be32 *)dev_priv->pfp_fw->data;
 499         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 500         for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 501                 RADEON_WRITE(R600_CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
 502         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 503
 504         fw_data = (const __be32 *)dev_priv->me_fw->data;
 505         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 506         for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 507                 RADEON_WRITE(R600_CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
 508         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 509
 510         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 511         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 512         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 513
 514 }
 515
 516 static void r600_test_writeback(drm_radeon_private_t *dev_priv)
 517 {
 518         u32 tmp;
 519
 520         /* Start with assuming that writeback doesn't work */
 521         dev_priv->writeback_works = 0;
 522
 523         /* Writeback doesn't seem to work everywhere, test it here and possibly
 524          * enable it if it appears to work
 525          */
 526         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
 527
 528         RADEON_WRITE(R600_SCRATCH_REG1, 0xdeadbeef);
 529
 530         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 531                 u32 val;
 532
 533                 val = radeon_read_ring_rptr(dev_priv, R600_SCRATCHOFF(1));
 534                 if (val == 0xdeadbeef)
 535                         break;
 536                 DRM_UDELAY(1);
 537         }
 538
 539         if (tmp < dev_priv->usec_timeout) {
 540                 dev_priv->writeback_works = 1;
 541                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 542         } else {
 543                 dev_priv->writeback_works = 0;
 544                 DRM_INFO("writeback test failed\n");
 545         }
 546         if (radeon_no_wb == 1) {
 547                 dev_priv->writeback_works = 0;
 548                 DRM_INFO("writeback forced off\n");
 549         }
 550
 551         if (!dev_priv->writeback_works) {
 552                 /* Disable writeback to avoid unnecessary bus master transfer */
 553                 RADEON_WRITE(R600_CP_RB_CNTL, RADEON_READ(R600_CP_RB_CNTL) |
 554                              RADEON_RB_NO_UPDATE);
 555                 RADEON_WRITE(R600_SCRATCH_UMSK, 0);
 556         }
 557 }
 558
 559 int r600_do_engine_reset(struct drm_device *dev)
 560 {
 561         drm_radeon_private_t *dev_priv = dev->dev_private;
 562         u32 cp_ptr, cp_me_cntl, cp_rb_cntl;
 563
 564         DRM_INFO("Resetting GPU\n");
 565
 566         cp_ptr = RADEON_READ(R600_CP_RB_WPTR);
 567         cp_me_cntl = RADEON_READ(R600_CP_ME_CNTL);
 568         RADEON_WRITE(R600_CP_ME_CNTL, R600_CP_ME_HALT);
 569
 570         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0x7fff);
 571         RADEON_READ(R600_GRBM_SOFT_RESET);
 572         DRM_UDELAY(50);
 573         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 574         RADEON_READ(R600_GRBM_SOFT_RESET);
 575
 576         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
 577         cp_rb_cntl = RADEON_READ(R600_CP_RB_CNTL);
 578         RADEON_WRITE(R600_CP_RB_CNTL, R600_RB_RPTR_WR_ENA);
 579
 580         RADEON_WRITE(R600_CP_RB_RPTR_WR, cp_ptr);
 581         RADEON_WRITE(R600_CP_RB_WPTR, cp_ptr);
 582         RADEON_WRITE(R600_CP_RB_CNTL, cp_rb_cntl);
 583         RADEON_WRITE(R600_CP_ME_CNTL, cp_me_cntl);
 584
 585         /* Reset the CP ring */
 586         r600_do_cp_reset(dev_priv);
 587
 588         /* The CP is no longer running after an engine reset */
 589         dev_priv->cp_running = 0;
 590
 591         /* Reset any pending vertex, indirect buffers */
 592         radeon_freelist_reset(dev);
 593
 594         return 0;
 595
 596 }
 597
 598 static u32 r600_get_tile_pipe_to_backend_map(u32 num_tile_pipes,
 599                                              u32 num_backends,
 600                                              u32 backend_disable_mask)
 601 {
 602         u32 backend_map = 0;
 603         u32 enabled_backends_mask;
 604         u32 enabled_backends_count;
 605         u32 cur_pipe;
 606         u32 swizzle_pipe[R6XX_MAX_PIPES];
 607         u32 cur_backend;
 608         u32 i;
 609
 610         if (num_tile_pipes > R6XX_MAX_PIPES)
 611                 num_tile_pipes = R6XX_MAX_PIPES;
 612         if (num_tile_pipes < 1)
 613                 num_tile_pipes = 1;
 614         if (num_backends > R6XX_MAX_BACKENDS)
 615                 num_backends = R6XX_MAX_BACKENDS;
 616         if (num_backends < 1)
 617                 num_backends = 1;
 618
 619         enabled_backends_mask = 0;
 620         enabled_backends_count = 0;
 621         for (i = 0; i < R6XX_MAX_BACKENDS; ++i) {
 622                 if (((backend_disable_mask >> i) & 1) == 0) {
 623                         enabled_backends_mask |= (1 << i);
 624                         ++enabled_backends_count;
 625                 }
 626                 if (enabled_backends_count == num_backends)
 627                         break;
 628         }
 629
 630         if (enabled_backends_count == 0) {
 631                 enabled_backends_mask = 1;
 632                 enabled_backends_count = 1;
 633         }
 634
 635         if (enabled_backends_count != num_backends)
 636                 num_backends = enabled_backends_count;
 637
 638         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R6XX_MAX_PIPES);
 639         switch (num_tile_pipes) {
 640         case 1:
 641                 swizzle_pipe[0] = 0;
 642                 break;
 643         case 2:
 644                 swizzle_pipe[0] = 0;
 645                 swizzle_pipe[1] = 1;
 646                 break;
 647         case 3:
 648                 swizzle_pipe[0] = 0;
 649                 swizzle_pipe[1] = 1;
 650                 swizzle_pipe[2] = 2;
 651                 break;
 652         case 4:
 653                 swizzle_pipe[0] = 0;
 654                 swizzle_pipe[1] = 1;
 655                 swizzle_pipe[2] = 2;
 656                 swizzle_pipe[3] = 3;
 657                 break;
 658         case 5:
 659                 swizzle_pipe[0] = 0;
 660                 swizzle_pipe[1] = 1;
 661                 swizzle_pipe[2] = 2;
 662                 swizzle_pipe[3] = 3;
 663                 swizzle_pipe[4] = 4;
 664                 break;
 665         case 6:
 666                 swizzle_pipe[0] = 0;
 667                 swizzle_pipe[1] = 2;
 668                 swizzle_pipe[2] = 4;
 669                 swizzle_pipe[3] = 5;
 670                 swizzle_pipe[4] = 1;
 671                 swizzle_pipe[5] = 3;
 672                 break;
 673         case 7:
 674                 swizzle_pipe[0] = 0;
 675                 swizzle_pipe[1] = 2;
 676                 swizzle_pipe[2] = 4;
 677                 swizzle_pipe[3] = 6;
 678                 swizzle_pipe[4] = 1;
 679                 swizzle_pipe[5] = 3;
 680                 swizzle_pipe[6] = 5;
 681                 break;
 682         case 8:
 683                 swizzle_pipe[0] = 0;
 684                 swizzle_pipe[1] = 2;
 685                 swizzle_pipe[2] = 4;
 686                 swizzle_pipe[3] = 6;
 687                 swizzle_pipe[4] = 1;
 688                 swizzle_pipe[5] = 3;
 689                 swizzle_pipe[6] = 5;
 690                 swizzle_pipe[7] = 7;
 691                 break;
 692         }
 693
 694         cur_backend = 0;
 695         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
 696                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
 697                         cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 698
 699                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
 700
 701                 cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 702         }
 703
 704         return backend_map;
 705 }
 706
 707 static int r600_count_pipe_bits(uint32_t val)
 708 {
 709         int i, ret = 0;
 710         for (i = 0; i < 32; i++) {
 711                 ret += val & 1;
 712                 val >>= 1;
 713         }
 714         return ret;
 715 }
 716
 717 static void r600_gfx_init(struct drm_device *dev,
 718                           drm_radeon_private_t *dev_priv)
 719 {
 720         int i, j, num_qd_pipes;
 721         u32 sx_debug_1;
 722         u32 tc_cntl;
 723         u32 arb_pop;
 724         u32 num_gs_verts_per_thread;
 725         u32 vgt_gs_per_es;
 726         u32 gs_prim_buffer_depth = 0;
 727         u32 sq_ms_fifo_sizes;
 728         u32 sq_config;
 729         u32 sq_gpr_resource_mgmt_1 = 0;
 730         u32 sq_gpr_resource_mgmt_2 = 0;
 731         u32 sq_thread_resource_mgmt = 0;
 732         u32 sq_stack_resource_mgmt_1 = 0;
 733         u32 sq_stack_resource_mgmt_2 = 0;
 734         u32 hdp_host_path_cntl;
 735         u32 backend_map;
 736         u32 gb_tiling_config = 0;
 737         u32 cc_rb_backend_disable;
 738         u32 cc_gc_shader_pipe_config;
 739         u32 ramcfg;
 740
 741         /* setup chip specs */
 742         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
 743         case CHIP_R600:
 744                 dev_priv->r600_max_pipes = 4;
 745                 dev_priv->r600_max_tile_pipes = 8;
 746                 dev_priv->r600_max_simds = 4;
 747                 dev_priv->r600_max_backends = 4;
 748                 dev_priv->r600_max_gprs = 256;
 749                 dev_priv->r600_max_threads = 192;
 750                 dev_priv->r600_max_stack_entries = 256;
 751                 dev_priv->r600_max_hw_contexts = 8;
 752                 dev_priv->r600_max_gs_threads = 16;
 753                 dev_priv->r600_sx_max_export_size = 128;
 754                 dev_priv->r600_sx_max_export_pos_size = 16;
 755                 dev_priv->r600_sx_max_export_smx_size = 128;
 756                 dev_priv->r600_sq_num_cf_insts = 2;
 757                 break;
 758         case CHIP_RV630:
 759         case CHIP_RV635:
 760                 dev_priv->r600_max_pipes = 2;
 761                 dev_priv->r600_max_tile_pipes = 2;
 762                 dev_priv->r600_max_simds = 3;
 763                 dev_priv->r600_max_backends = 1;
 764                 dev_priv->r600_max_gprs = 128;
 765                 dev_priv->r600_max_threads = 192;
 766                 dev_priv->r600_max_stack_entries = 128;
 767                 dev_priv->r600_max_hw_contexts = 8;
 768                 dev_priv->r600_max_gs_threads = 4;
 769                 dev_priv->r600_sx_max_export_size = 128;
 770                 dev_priv->r600_sx_max_export_pos_size = 16;
 771                 dev_priv->r600_sx_max_export_smx_size = 128;
 772                 dev_priv->r600_sq_num_cf_insts = 2;
 773                 break;
 774         case CHIP_RV610:
 775         case CHIP_RS780:
 776         case CHIP_RS880:
 777         case CHIP_RV620:
 778                 dev_priv->r600_max_pipes = 1;
 779                 dev_priv->r600_max_tile_pipes = 1;
 780                 dev_priv->r600_max_simds = 2;
 781                 dev_priv->r600_max_backends = 1;
 782                 dev_priv->r600_max_gprs = 128;
 783                 dev_priv->r600_max_threads = 192;
 784                 dev_priv->r600_max_stack_entries = 128;
 785                 dev_priv->r600_max_hw_contexts = 4;
 786                 dev_priv->r600_max_gs_threads = 4;
 787                 dev_priv->r600_sx_max_export_size = 128;
 788                 dev_priv->r600_sx_max_export_pos_size = 16;
 789                 dev_priv->r600_sx_max_export_smx_size = 128;
 790                 dev_priv->r600_sq_num_cf_insts = 1;
 791                 break;
 792         case CHIP_RV670:
 793                 dev_priv->r600_max_pipes = 4;
 794                 dev_priv->r600_max_tile_pipes = 4;
 795                 dev_priv->r600_max_simds = 4;
 796                 dev_priv->r600_max_backends = 4;
 797                 dev_priv->r600_max_gprs = 192;
 798                 dev_priv->r600_max_threads = 192;
 799                 dev_priv->r600_max_stack_entries = 256;
 800                 dev_priv->r600_max_hw_contexts = 8;
 801                 dev_priv->r600_max_gs_threads = 16;
 802                 dev_priv->r600_sx_max_export_size = 128;
 803                 dev_priv->r600_sx_max_export_pos_size = 16;
 804                 dev_priv->r600_sx_max_export_smx_size = 128;
 805                 dev_priv->r600_sq_num_cf_insts = 2;
 806                 break;
 807         default:
 808                 break;
 809         }
 810
 811         /* Initialize HDP */
 812         j = 0;
 813         for (i = 0; i < 32; i++) {
 814                 RADEON_WRITE((0x2c14 + j), 0x00000000);
 815                 RADEON_WRITE((0x2c18 + j), 0x00000000);
 816                 RADEON_WRITE((0x2c1c + j), 0x00000000);
 817                 RADEON_WRITE((0x2c20 + j), 0x00000000);
 818                 RADEON_WRITE((0x2c24 + j), 0x00000000);
 819                 j += 0x18;
 820         }
 821
 822         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
 823
 824         /* setup tiling, simd, pipe config */
 825         ramcfg = RADEON_READ(R600_RAMCFG);
 826
 827         switch (dev_priv->r600_max_tile_pipes) {
 828         case 1:
 829                 gb_tiling_config |= R600_PIPE_TILING(0);
 830                 break;
 831         case 2:
 832                 gb_tiling_config |= R600_PIPE_TILING(1);
 833                 break;
 834         case 4:
 835                 gb_tiling_config |= R600_PIPE_TILING(2);
 836                 break;
 837         case 8:
 838                 gb_tiling_config |= R600_PIPE_TILING(3);
 839                 break;
 840         default:
 841                 break;
 842         }
 843
 844         gb_tiling_config |= R600_BANK_TILING((ramcfg >> R600_NOOFBANK_SHIFT) & R600_NOOFBANK_MASK);
 845
 846         gb_tiling_config |= R600_GROUP_SIZE(0);
 847
 848         if (((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK) > 3) {
 849                 gb_tiling_config |= R600_ROW_TILING(3);
 850                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
 851         } else {
 852                 gb_tiling_config |=
 853                         R600_ROW_TILING(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 854                 gb_tiling_config |=
 855                         R600_SAMPLE_SPLIT(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 856         }
 857
 858         gb_tiling_config |= R600_BANK_SWAPS(1);
 859
 860         cc_rb_backend_disable = RADEON_READ(R600_CC_RB_BACKEND_DISABLE) & 0x00ff0000;
 861         cc_rb_backend_disable |=
 862                 R600_BACKEND_DISABLE((R6XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R6XX_MAX_BACKENDS_MASK);
 863
 864         cc_gc_shader_pipe_config = RADEON_READ(R600_CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
 865         cc_gc_shader_pipe_config |=
 866                 R600_INACTIVE_QD_PIPES((R6XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R6XX_MAX_PIPES_MASK);
 867         cc_gc_shader_pipe_config |=
 868                 R600_INACTIVE_SIMDS((R6XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R6XX_MAX_SIMDS_MASK);
 869
 870         backend_map = r600_get_tile_pipe_to_backend_map(dev_priv->r600_max_tile_pipes,
 871                                                         (R6XX_MAX_BACKENDS -
 872                                                          r600_count_pipe_bits((cc_rb_backend_disable &
 873                                                                                R6XX_MAX_BACKENDS_MASK) >> 16)),
 874                                                         (cc_rb_backend_disable >> 16));
 875         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
 876
 877         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
 878         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 879         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 880         if (gb_tiling_config & 0xc0) {
 881                 dev_priv->r600_group_size = 512;
 882         } else {
 883                 dev_priv->r600_group_size = 256;
 884         }
 885         dev_priv->r600_npipes = 1 << ((gb_tiling_config >> 1) & 0x7);
 886         if (gb_tiling_config & 0x30) {
 887                 dev_priv->r600_nbanks = 8;
 888         } else {
 889                 dev_priv->r600_nbanks = 4;
 890         }
 891
 892         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 893         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 894         RADEON_WRITE(R600_GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 895
 896         num_qd_pipes =
 897                 R6XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK) >> 8);
 898         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
 899         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
 900
 901         /* set HW defaults for 3D engine */
 902         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
 903                                                 R600_ROQ_IB2_START(0x2b)));
 904
 905         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, (R600_MEQ_END(0x40) |
 906                                               R600_ROQ_END(0x40)));
 907
 908         RADEON_WRITE(R600_TA_CNTL_AUX, (R600_DISABLE_CUBE_ANISO |
 909                                         R600_SYNC_GRADIENT |
 910                                         R600_SYNC_WALKER |
 911                                         R600_SYNC_ALIGNER));
 912
 913         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670)
 914                 RADEON_WRITE(R600_ARB_GDEC_RD_CNTL, 0x00000021);
 915
 916         sx_debug_1 = RADEON_READ(R600_SX_DEBUG_1);
 917         sx_debug_1 |= R600_SMX_EVENT_RELEASE;
 918         if (((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_R600))
 919                 sx_debug_1 |= R600_ENABLE_NEW_SMX_ADDRESS;
 920         RADEON_WRITE(R600_SX_DEBUG_1, sx_debug_1);
 921
 922         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 923             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
 924             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 925             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 926             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 927             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
 928                 RADEON_WRITE(R600_DB_DEBUG, R600_PREZ_MUST_WAIT_FOR_POSTZ_DONE);
 929         else
 930                 RADEON_WRITE(R600_DB_DEBUG, 0);
 931
 932         RADEON_WRITE(R600_DB_WATERMARKS, (R600_DEPTH_FREE(4) |
 933                                           R600_DEPTH_FLUSH(16) |
 934                                           R600_DEPTH_PENDING_FREE(4) |
 935                                           R600_DEPTH_CACHELINE_FREE(16)));
 936         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
 937         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 0);
 938
 939         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
 940         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(0));
 941
 942         sq_ms_fifo_sizes = RADEON_READ(R600_SQ_MS_FIFO_SIZES);
 943         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 944             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 945             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 946             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 947                 sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(0xa) |
 948                                     R600_FETCH_FIFO_HIWATER(0xa) |
 949                                     R600_DONE_FIFO_HIWATER(0xe0) |
 950                                     R600_ALU_UPDATE_FIFO_HIWATER(0x8));
 951         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 952                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630)) {
 953                 sq_ms_fifo_sizes &= ~R600_DONE_FIFO_HIWATER(0xff);
 954                 sq_ms_fifo_sizes |= R600_DONE_FIFO_HIWATER(0x4);
 955         }
 956         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 957
 958         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 959          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 960          */
 961         sq_config = RADEON_READ(R600_SQ_CONFIG);
 962         sq_config &= ~(R600_PS_PRIO(3) |
 963                        R600_VS_PRIO(3) |
 964                        R600_GS_PRIO(3) |
 965                        R600_ES_PRIO(3));
 966         sq_config |= (R600_DX9_CONSTS |
 967                       R600_VC_ENABLE |
 968                       R600_PS_PRIO(0) |
 969                       R600_VS_PRIO(1) |
 970                       R600_GS_PRIO(2) |
 971                       R600_ES_PRIO(3));
 972
 973         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) {
 974                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(124) |
 975                                           R600_NUM_VS_GPRS(124) |
 976                                           R600_NUM_CLAUSE_TEMP_GPRS(4));
 977                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(0) |
 978                                           R600_NUM_ES_GPRS(0));
 979                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(136) |
 980                                            R600_NUM_VS_THREADS(48) |
 981                                            R600_NUM_GS_THREADS(4) |
 982                                            R600_NUM_ES_THREADS(4));
 983                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(128) |
 984                                             R600_NUM_VS_STACK_ENTRIES(128));
 985                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(0) |
 986                                             R600_NUM_ES_STACK_ENTRIES(0));
 987         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 988                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 989                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 990                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 991                 /* no vertex cache */
 992                 sq_config &= ~R600_VC_ENABLE;
 993
 994                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
 995                                           R600_NUM_VS_GPRS(44) |
 996                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
 997                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
 998                                           R600_NUM_ES_GPRS(17));
 999                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1000                                            R600_NUM_VS_THREADS(78) |
1001                                            R600_NUM_GS_THREADS(4) |
1002                                            R600_NUM_ES_THREADS(31));
1003                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1004                                             R600_NUM_VS_STACK_ENTRIES(40));
1005                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1006                                             R600_NUM_ES_STACK_ENTRIES(16));
1007         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
1008                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV635)) {
1009                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1010                                           R600_NUM_VS_GPRS(44) |
1011                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1012                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(18) |
1013                                           R600_NUM_ES_GPRS(18));
1014                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1015                                            R600_NUM_VS_THREADS(78) |
1016                                            R600_NUM_GS_THREADS(4) |
1017                                            R600_NUM_ES_THREADS(31));
1018                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1019                                             R600_NUM_VS_STACK_ENTRIES(40));
1020                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1021                                             R600_NUM_ES_STACK_ENTRIES(16));
1022         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670) {
1023                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1024                                           R600_NUM_VS_GPRS(44) |
1025                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1026                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
1027                                           R600_NUM_ES_GPRS(17));
1028                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1029                                            R600_NUM_VS_THREADS(78) |
1030                                            R600_NUM_GS_THREADS(4) |
1031                                            R600_NUM_ES_THREADS(31));
1032                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(64) |
1033                                             R600_NUM_VS_STACK_ENTRIES(64));
1034                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(64) |
1035                                             R600_NUM_ES_STACK_ENTRIES(64));
1036         }
1037
1038         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1039         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  sq_gpr_resource_mgmt_1);
1040         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  sq_gpr_resource_mgmt_2);
1041         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1042         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, sq_stack_resource_mgmt_1);
1043         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, sq_stack_resource_mgmt_2);
1044
1045         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
1046             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
1047             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
1048             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
1049                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_TC_ONLY));
1050         else
1051                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_VC_AND_TC));
1052
1053         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_2S, (R600_S0_X(0xc) |
1054                                                     R600_S0_Y(0x4) |
1055                                                     R600_S1_X(0x4) |
1056                                                     R600_S1_Y(0xc)));
1057         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_4S, (R600_S0_X(0xe) |
1058                                                     R600_S0_Y(0xe) |
1059                                                     R600_S1_X(0x2) |
1060                                                     R600_S1_Y(0x2) |
1061                                                     R600_S2_X(0xa) |
1062                                                     R600_S2_Y(0x6) |
1063                                                     R600_S3_X(0x6) |
1064                                                     R600_S3_Y(0xa)));
1065         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD0, (R600_S0_X(0xe) |
1066                                                         R600_S0_Y(0xb) |
1067                                                         R600_S1_X(0x4) |
1068                                                         R600_S1_Y(0xc) |
1069                                                         R600_S2_X(0x1) |
1070                                                         R600_S2_Y(0x6) |
1071                                                         R600_S3_X(0xa) |
1072                                                         R600_S3_Y(0xe)));
1073         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD1, (R600_S4_X(0x6) |
1074                                                         R600_S4_Y(0x1) |
1075                                                         R600_S5_X(0x0) |
1076                                                         R600_S5_Y(0x0) |
1077                                                         R600_S6_X(0xb) |
1078                                                         R600_S6_Y(0x4) |
1079                                                         R600_S7_X(0x7) |
1080                                                         R600_S7_Y(0x8)));
1081
1082
1083         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1084         case CHIP_R600:
1085         case CHIP_RV630:
1086         case CHIP_RV635:
1087                 gs_prim_buffer_depth = 0;
1088                 break;
1089         case CHIP_RV610:
1090         case CHIP_RS780:
1091         case CHIP_RS880:
1092         case CHIP_RV620:
1093                 gs_prim_buffer_depth = 32;
1094                 break;
1095         case CHIP_RV670:
1096                 gs_prim_buffer_depth = 128;
1097                 break;
1098         default:
1099                 break;
1100         }
1101
1102         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1103         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1104         /* Max value for this is 256 */
1105         if (vgt_gs_per_es > 256)
1106                 vgt_gs_per_es = 256;
1107
1108         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1109         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1110         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1111         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1112
1113         /* more default values. 2D/3D driver should adjust as needed */
1114         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1115         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1116         RADEON_WRITE(R600_SX_MISC, 0);
1117         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1118         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1119         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1120         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1121         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1122         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1123
1124         /* clear render buffer base addresses */
1125         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1126         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1127         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1128         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1129         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1130         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1131         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1132         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1133
1134         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1135         case CHIP_RV610:
1136         case CHIP_RS780:
1137         case CHIP_RS880:
1138         case CHIP_RV620:
1139                 tc_cntl = R600_TC_L2_SIZE(8);
1140                 break;
1141         case CHIP_RV630:
1142         case CHIP_RV635:
1143                 tc_cntl = R600_TC_L2_SIZE(4);
1144                 break;
1145         case CHIP_R600:
1146                 tc_cntl = R600_TC_L2_SIZE(0) | R600_L2_DISABLE_LATE_HIT;
1147                 break;
1148         default:
1149                 tc_cntl = R600_TC_L2_SIZE(0);
1150                 break;
1151         }
1152
1153         RADEON_WRITE(R600_TC_CNTL, tc_cntl);
1154
1155         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1156         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1157
1158         arb_pop = RADEON_READ(R600_ARB_POP);
1159         arb_pop |= R600_ENABLE_TC128;
1160         RADEON_WRITE(R600_ARB_POP, arb_pop);
1161
1162         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1163         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1164                                           R600_NUM_CLIP_SEQ(3)));
1165         RADEON_WRITE(R600_PA_SC_ENHANCE, R600_FORCE_EOV_MAX_CLK_CNT(4095));
1166
1167 }
1168
1169 static u32 r700_get_tile_pipe_to_backend_map(drm_radeon_private_t *dev_priv,
1170                                              u32 num_tile_pipes,
1171                                              u32 num_backends,
1172                                              u32 backend_disable_mask)
1173 {
1174         u32 backend_map = 0;
1175         u32 enabled_backends_mask;
1176         u32 enabled_backends_count;
1177         u32 cur_pipe;
1178         u32 swizzle_pipe[R7XX_MAX_PIPES];
1179         u32 cur_backend;
1180         u32 i;
1181         bool force_no_swizzle;
1182
1183         if (num_tile_pipes > R7XX_MAX_PIPES)
1184                 num_tile_pipes = R7XX_MAX_PIPES;
1185         if (num_tile_pipes < 1)
1186                 num_tile_pipes = 1;
1187         if (num_backends > R7XX_MAX_BACKENDS)
1188                 num_backends = R7XX_MAX_BACKENDS;
1189         if (num_backends < 1)
1190                 num_backends = 1;
1191
1192         enabled_backends_mask = 0;
1193         enabled_backends_count = 0;
1194         for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
1195                 if (((backend_disable_mask >> i) & 1) == 0) {
1196                         enabled_backends_mask |= (1 << i);
1197                         ++enabled_backends_count;
1198                 }
1199                 if (enabled_backends_count == num_backends)
1200                         break;
1201         }
1202
1203         if (enabled_backends_count == 0) {
1204                 enabled_backends_mask = 1;
1205                 enabled_backends_count = 1;
1206         }
1207
1208         if (enabled_backends_count != num_backends)
1209                 num_backends = enabled_backends_count;
1210
1211         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1212         case CHIP_RV770:
1213         case CHIP_RV730:
1214                 force_no_swizzle = false;
1215                 break;
1216         case CHIP_RV710:
1217         case CHIP_RV740:
1218         default:
1219                 force_no_swizzle = true;
1220                 break;
1221         }
1222
1223         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
1224         switch (num_tile_pipes) {
1225         case 1:
1226                 swizzle_pipe[0] = 0;
1227                 break;
1228         case 2:
1229                 swizzle_pipe[0] = 0;
1230                 swizzle_pipe[1] = 1;
1231                 break;
1232         case 3:
1233                 if (force_no_swizzle) {
1234                         swizzle_pipe[0] = 0;
1235                         swizzle_pipe[1] = 1;
1236                         swizzle_pipe[2] = 2;
1237                 } else {
1238                         swizzle_pipe[0] = 0;
1239                         swizzle_pipe[1] = 2;
1240                         swizzle_pipe[2] = 1;
1241                 }
1242                 break;
1243         case 4:
1244                 if (force_no_swizzle) {
1245                         swizzle_pipe[0] = 0;
1246                         swizzle_pipe[1] = 1;
1247                         swizzle_pipe[2] = 2;
1248                         swizzle_pipe[3] = 3;
1249                 } else {
1250                         swizzle_pipe[0] = 0;
1251                         swizzle_pipe[1] = 2;
1252                         swizzle_pipe[2] = 3;
1253                         swizzle_pipe[3] = 1;
1254                 }
1255                 break;
1256         case 5:
1257                 if (force_no_swizzle) {
1258                         swizzle_pipe[0] = 0;
1259                         swizzle_pipe[1] = 1;
1260                         swizzle_pipe[2] = 2;
1261                         swizzle_pipe[3] = 3;
1262                         swizzle_pipe[4] = 4;
1263                 } else {
1264                         swizzle_pipe[0] = 0;
1265                         swizzle_pipe[1] = 2;
1266                         swizzle_pipe[2] = 4;
1267                         swizzle_pipe[3] = 1;
1268                         swizzle_pipe[4] = 3;
1269                 }
1270                 break;
1271         case 6:
1272                 if (force_no_swizzle) {
1273                         swizzle_pipe[0] = 0;
1274                         swizzle_pipe[1] = 1;
1275                         swizzle_pipe[2] = 2;
1276                         swizzle_pipe[3] = 3;
1277                         swizzle_pipe[4] = 4;
1278                         swizzle_pipe[5] = 5;
1279                 } else {
1280                         swizzle_pipe[0] = 0;
1281                         swizzle_pipe[1] = 2;
1282                         swizzle_pipe[2] = 4;
1283                         swizzle_pipe[3] = 5;
1284                         swizzle_pipe[4] = 3;
1285                         swizzle_pipe[5] = 1;
1286                 }
1287                 break;
1288         case 7:
1289                 if (force_no_swizzle) {
1290                         swizzle_pipe[0] = 0;
1291                         swizzle_pipe[1] = 1;
1292                         swizzle_pipe[2] = 2;
1293                         swizzle_pipe[3] = 3;
1294                         swizzle_pipe[4] = 4;
1295                         swizzle_pipe[5] = 5;
1296                         swizzle_pipe[6] = 6;
1297                 } else {
1298                         swizzle_pipe[0] = 0;
1299                         swizzle_pipe[1] = 2;
1300                         swizzle_pipe[2] = 4;
1301                         swizzle_pipe[3] = 6;
1302                         swizzle_pipe[4] = 3;
1303                         swizzle_pipe[5] = 1;
1304                         swizzle_pipe[6] = 5;
1305                 }
1306                 break;
1307         case 8:
1308                 if (force_no_swizzle) {
1309                         swizzle_pipe[0] = 0;
1310                         swizzle_pipe[1] = 1;
1311                         swizzle_pipe[2] = 2;
1312                         swizzle_pipe[3] = 3;
1313                         swizzle_pipe[4] = 4;
1314                         swizzle_pipe[5] = 5;
1315                         swizzle_pipe[6] = 6;
1316                         swizzle_pipe[7] = 7;
1317                 } else {
1318                         swizzle_pipe[0] = 0;
1319                         swizzle_pipe[1] = 2;
1320                         swizzle_pipe[2] = 4;
1321                         swizzle_pipe[3] = 6;
1322                         swizzle_pipe[4] = 3;
1323                         swizzle_pipe[5] = 1;
1324                         swizzle_pipe[6] = 7;
1325                         swizzle_pipe[7] = 5;
1326                 }
1327                 break;
1328         }
1329
1330         cur_backend = 0;
1331         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
1332                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
1333                         cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1334
1335                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
1336
1337                 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1338         }
1339
1340         return backend_map;
1341 }
1342
1343 static void r700_gfx_init(struct drm_device *dev,
1344                           drm_radeon_private_t *dev_priv)
1345 {
1346         int i, j, num_qd_pipes;
1347         u32 ta_aux_cntl;
1348         u32 sx_debug_1;
1349         u32 smx_dc_ctl0;
1350         u32 db_debug3;
1351         u32 num_gs_verts_per_thread;
1352         u32 vgt_gs_per_es;
1353         u32 gs_prim_buffer_depth = 0;
1354         u32 sq_ms_fifo_sizes;
1355         u32 sq_config;
1356         u32 sq_thread_resource_mgmt;
1357         u32 hdp_host_path_cntl;
1358         u32 sq_dyn_gpr_size_simd_ab_0;
1359         u32 backend_map;
1360         u32 gb_tiling_config = 0;
1361         u32 cc_rb_backend_disable;
1362         u32 cc_gc_shader_pipe_config;
1363         u32 mc_arb_ramcfg;
1364         u32 db_debug4;
1365
1366         /* setup chip specs */
1367         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1368         case CHIP_RV770:
1369                 dev_priv->r600_max_pipes = 4;
1370                 dev_priv->r600_max_tile_pipes = 8;
1371                 dev_priv->r600_max_simds = 10;
1372                 dev_priv->r600_max_backends = 4;
1373                 dev_priv->r600_max_gprs = 256;
1374                 dev_priv->r600_max_threads = 248;
1375                 dev_priv->r600_max_stack_entries = 512;
1376                 dev_priv->r600_max_hw_contexts = 8;
1377                 dev_priv->r600_max_gs_threads = 16 * 2;
1378                 dev_priv->r600_sx_max_export_size = 128;
1379                 dev_priv->r600_sx_max_export_pos_size = 16;
1380                 dev_priv->r600_sx_max_export_smx_size = 112;
1381                 dev_priv->r600_sq_num_cf_insts = 2;
1382
1383                 dev_priv->r700_sx_num_of_sets = 7;
1384                 dev_priv->r700_sc_prim_fifo_size = 0xF9;
1385                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1386                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1387                 break;
1388         case CHIP_RV730:
1389                 dev_priv->r600_max_pipes = 2;
1390                 dev_priv->r600_max_tile_pipes = 4;
1391                 dev_priv->r600_max_simds = 8;
1392                 dev_priv->r600_max_backends = 2;
1393                 dev_priv->r600_max_gprs = 128;
1394                 dev_priv->r600_max_threads = 248;
1395                 dev_priv->r600_max_stack_entries = 256;
1396                 dev_priv->r600_max_hw_contexts = 8;
1397                 dev_priv->r600_max_gs_threads = 16 * 2;
1398                 dev_priv->r600_sx_max_export_size = 256;
1399                 dev_priv->r600_sx_max_export_pos_size = 32;
1400                 dev_priv->r600_sx_max_export_smx_size = 224;
1401                 dev_priv->r600_sq_num_cf_insts = 2;
1402
1403                 dev_priv->r700_sx_num_of_sets = 7;
1404                 dev_priv->r700_sc_prim_fifo_size = 0xf9;
1405                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1406                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1407                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1408                         dev_priv->r600_sx_max_export_pos_size -= 16;
1409                         dev_priv->r600_sx_max_export_smx_size += 16;
1410                 }
1411                 break;
1412         case CHIP_RV710:
1413                 dev_priv->r600_max_pipes = 2;
1414                 dev_priv->r600_max_tile_pipes = 2;
1415                 dev_priv->r600_max_simds = 2;
1416                 dev_priv->r600_max_backends = 1;
1417                 dev_priv->r600_max_gprs = 256;
1418                 dev_priv->r600_max_threads = 192;
1419                 dev_priv->r600_max_stack_entries = 256;
1420                 dev_priv->r600_max_hw_contexts = 4;
1421                 dev_priv->r600_max_gs_threads = 8 * 2;
1422                 dev_priv->r600_sx_max_export_size = 128;
1423                 dev_priv->r600_sx_max_export_pos_size = 16;
1424                 dev_priv->r600_sx_max_export_smx_size = 112;
1425                 dev_priv->r600_sq_num_cf_insts = 1;
1426
1427                 dev_priv->r700_sx_num_of_sets = 7;
1428                 dev_priv->r700_sc_prim_fifo_size = 0x40;
1429                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1430                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1431                 break;
1432         case CHIP_RV740:
1433                 dev_priv->r600_max_pipes = 4;
1434                 dev_priv->r600_max_tile_pipes = 4;
1435                 dev_priv->r600_max_simds = 8;
1436                 dev_priv->r600_max_backends = 4;
1437                 dev_priv->r600_max_gprs = 256;
1438                 dev_priv->r600_max_threads = 248;
1439                 dev_priv->r600_max_stack_entries = 512;
1440                 dev_priv->r600_max_hw_contexts = 8;
1441                 dev_priv->r600_max_gs_threads = 16 * 2;
1442                 dev_priv->r600_sx_max_export_size = 256;
1443                 dev_priv->r600_sx_max_export_pos_size = 32;
1444                 dev_priv->r600_sx_max_export_smx_size = 224;
1445                 dev_priv->r600_sq_num_cf_insts = 2;
1446
1447                 dev_priv->r700_sx_num_of_sets = 7;
1448                 dev_priv->r700_sc_prim_fifo_size = 0x100;
1449                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1450                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1451
1452                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1453                         dev_priv->r600_sx_max_export_pos_size -= 16;
1454                         dev_priv->r600_sx_max_export_smx_size += 16;
1455                 }
1456                 break;
1457         default:
1458                 break;
1459         }
1460
1461         /* Initialize HDP */
1462         j = 0;
1463         for (i = 0; i < 32; i++) {
1464                 RADEON_WRITE((0x2c14 + j), 0x00000000);
1465                 RADEON_WRITE((0x2c18 + j), 0x00000000);
1466                 RADEON_WRITE((0x2c1c + j), 0x00000000);
1467                 RADEON_WRITE((0x2c20 + j), 0x00000000);
1468                 RADEON_WRITE((0x2c24 + j), 0x00000000);
1469                 j += 0x18;
1470         }
1471
1472         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
1473
1474         /* setup tiling, simd, pipe config */
1475         mc_arb_ramcfg = RADEON_READ(R700_MC_ARB_RAMCFG);
1476
1477         switch (dev_priv->r600_max_tile_pipes) {
1478         case 1:
1479                 gb_tiling_config |= R600_PIPE_TILING(0);
1480                 break;
1481         case 2:
1482                 gb_tiling_config |= R600_PIPE_TILING(1);
1483                 break;
1484         case 4:
1485                 gb_tiling_config |= R600_PIPE_TILING(2);
1486                 break;
1487         case 8:
1488                 gb_tiling_config |= R600_PIPE_TILING(3);
1489                 break;
1490         default:
1491                 break;
1492         }
1493
1494         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV770)
1495                 gb_tiling_config |= R600_BANK_TILING(1);
1496         else
1497                 gb_tiling_config |= R600_BANK_TILING((mc_arb_ramcfg >> R700_NOOFBANK_SHIFT) & R700_NOOFBANK_MASK);
1498
1499         gb_tiling_config |= R600_GROUP_SIZE(0);
1500
1501         if (((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK) > 3) {
1502                 gb_tiling_config |= R600_ROW_TILING(3);
1503                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
1504         } else {
1505                 gb_tiling_config |=
1506                         R600_ROW_TILING(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1507                 gb_tiling_config |=
1508                         R600_SAMPLE_SPLIT(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1509         }
1510
1511         gb_tiling_config |= R600_BANK_SWAPS(1);
1512
1513         cc_rb_backend_disable = RADEON_READ(R600_CC_RB_BACKEND_DISABLE) & 0x00ff0000;
1514         cc_rb_backend_disable |=
1515                 R600_BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R7XX_MAX_BACKENDS_MASK);
1516
1517         cc_gc_shader_pipe_config = RADEON_READ(R600_CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
1518         cc_gc_shader_pipe_config |=
1519                 R600_INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R7XX_MAX_PIPES_MASK);
1520         cc_gc_shader_pipe_config |=
1521                 R600_INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R7XX_MAX_SIMDS_MASK);
1522
1523         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV740)
1524                 backend_map = 0x28;
1525         else
1526                 backend_map = r700_get_tile_pipe_to_backend_map(dev_priv,
1527                                                                 dev_priv->r600_max_tile_pipes,
1528                                                                 (R7XX_MAX_BACKENDS -
1529                                                                  r600_count_pipe_bits((cc_rb_backend_disable &
1530                                                                                        R7XX_MAX_BACKENDS_MASK) >> 16)),
1531                                                                 (cc_rb_backend_disable >> 16));
1532         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
1533
1534         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
1535         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1536         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1537         if (gb_tiling_config & 0xc0) {
1538                 dev_priv->r600_group_size = 512;
1539         } else {
1540                 dev_priv->r600_group_size = 256;
1541         }
1542         dev_priv->r600_npipes = 1 << ((gb_tiling_config >> 1) & 0x7);
1543         if (gb_tiling_config & 0x30) {
1544                 dev_priv->r600_nbanks = 8;
1545         } else {
1546                 dev_priv->r600_nbanks = 4;
1547         }
1548
1549         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
1550         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
1551
1552         RADEON_WRITE(R700_CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
1553         RADEON_WRITE(R700_CGTS_SYS_TCC_DISABLE, 0);
1554         RADEON_WRITE(R700_CGTS_TCC_DISABLE, 0);
1555
1556         num_qd_pipes =
1557                 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK) >> 8);
1558         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
1559         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
1560
1561         /* set HW defaults for 3D engine */
1562         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
1563                                                 R600_ROQ_IB2_START(0x2b)));
1564
1565         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, R700_STQ_SPLIT(0x30));
1566
1567         ta_aux_cntl = RADEON_READ(R600_TA_CNTL_AUX);
1568         RADEON_WRITE(R600_TA_CNTL_AUX, ta_aux_cntl | R600_DISABLE_CUBE_ANISO);
1569
1570         sx_debug_1 = RADEON_READ(R700_SX_DEBUG_1);
1571         sx_debug_1 |= R700_ENABLE_NEW_SMX_ADDRESS;
1572         RADEON_WRITE(R700_SX_DEBUG_1, sx_debug_1);
1573
1574         smx_dc_ctl0 = RADEON_READ(R600_SMX_DC_CTL0);
1575         smx_dc_ctl0 &= ~R700_CACHE_DEPTH(0x1ff);
1576         smx_dc_ctl0 |= R700_CACHE_DEPTH((dev_priv->r700_sx_num_of_sets * 64) - 1);
1577         RADEON_WRITE(R600_SMX_DC_CTL0, smx_dc_ctl0);
1578
1579         if ((dev_priv->flags & RADEON_FAMILY_MASK) != CHIP_RV740)
1580                 RADEON_WRITE(R700_SMX_EVENT_CTL, (R700_ES_FLUSH_CTL(4) |
1581                                                   R700_GS_FLUSH_CTL(4) |
1582                                                   R700_ACK_FLUSH_CTL(3) |
1583                                                   R700_SYNC_FLUSH_CTL));
1584
1585         db_debug3 = RADEON_READ(R700_DB_DEBUG3);
1586         db_debug3 &= ~R700_DB_CLK_OFF_DELAY(0x1f);
1587         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1588         case CHIP_RV770:
1589         case CHIP_RV740:
1590                 db_debug3 |= R700_DB_CLK_OFF_DELAY(0x1f);
1591                 break;
1592         case CHIP_RV710:
1593         case CHIP_RV730:
1594         default:
1595                 db_debug3 |= R700_DB_CLK_OFF_DELAY(2);
1596                 break;
1597         }
1598         RADEON_WRITE(R700_DB_DEBUG3, db_debug3);
1599
1600         if ((dev_priv->flags & RADEON_FAMILY_MASK) != CHIP_RV770) {
1601                 db_debug4 = RADEON_READ(RV700_DB_DEBUG4);
1602                 db_debug4 |= RV700_DISABLE_TILE_COVERED_FOR_PS_ITER;
1603                 RADEON_WRITE(RV700_DB_DEBUG4, db_debug4);
1604         }
1605
1606         RADEON_WRITE(R600_SX_EXPORT_BUFFER_SIZES, (R600_COLOR_BUFFER_SIZE((dev_priv->r600_sx_max_export_size / 4) - 1) |
1607                                                    R600_POSITION_BUFFER_SIZE((dev_priv->r600_sx_max_export_pos_size / 4) - 1) |
1608                                                    R600_SMX_BUFFER_SIZE((dev_priv->r600_sx_max_export_smx_size / 4) - 1)));
1609
1610         RADEON_WRITE(R700_PA_SC_FIFO_SIZE_R7XX, (R700_SC_PRIM_FIFO_SIZE(dev_priv->r700_sc_prim_fifo_size) |
1611                                                  R700_SC_HIZ_TILE_FIFO_SIZE(dev_priv->r700_sc_hiz_tile_fifo_size) |
1612                                                  R700_SC_EARLYZ_TILE_FIFO_SIZE(dev_priv->r700_sc_earlyz_tile_fifo_fize)));
1613
1614         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1615
1616         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 1);
1617
1618         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
1619
1620         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(4));
1621
1622         RADEON_WRITE(R600_CP_PERFMON_CNTL, 0);
1623
1624         sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(16 * dev_priv->r600_sq_num_cf_insts) |
1625                             R600_DONE_FIFO_HIWATER(0xe0) |
1626                             R600_ALU_UPDATE_FIFO_HIWATER(0x8));
1627         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1628         case CHIP_RV770:
1629         case CHIP_RV730:
1630         case CHIP_RV710:
1631                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x1);
1632                 break;
1633         case CHIP_RV740:
1634         default:
1635                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x4);
1636                 break;
1637         }
1638         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
1639
1640         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
1641          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
1642          */
1643         sq_config = RADEON_READ(R600_SQ_CONFIG);
1644         sq_config &= ~(R600_PS_PRIO(3) |
1645                        R600_VS_PRIO(3) |
1646                        R600_GS_PRIO(3) |
1647                        R600_ES_PRIO(3));
1648         sq_config |= (R600_DX9_CONSTS |
1649                       R600_VC_ENABLE |
1650                       R600_EXPORT_SRC_C |
1651                       R600_PS_PRIO(0) |
1652                       R600_VS_PRIO(1) |
1653                       R600_GS_PRIO(2) |
1654                       R600_ES_PRIO(3));
1655         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1656                 /* no vertex cache */
1657                 sq_config &= ~R600_VC_ENABLE;
1658
1659         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1660
1661         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  (R600_NUM_PS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1662                                                     R600_NUM_VS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1663                                                     R600_NUM_CLAUSE_TEMP_GPRS(((dev_priv->r600_max_gprs * 24)/64)/2)));
1664
1665         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  (R600_NUM_GS_GPRS((dev_priv->r600_max_gprs * 7)/64) |
1666                                                     R600_NUM_ES_GPRS((dev_priv->r600_max_gprs * 7)/64)));
1667
1668         sq_thread_resource_mgmt = (R600_NUM_PS_THREADS((dev_priv->r600_max_threads * 4)/8) |
1669                                    R600_NUM_VS_THREADS((dev_priv->r600_max_threads * 2)/8) |
1670                                    R600_NUM_ES_THREADS((dev_priv->r600_max_threads * 1)/8));
1671         if (((dev_priv->r600_max_threads * 1) / 8) > dev_priv->r600_max_gs_threads)
1672                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS(dev_priv->r600_max_gs_threads);
1673         else
1674                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS((dev_priv->r600_max_gs_threads * 1)/8);
1675         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1676
1677         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, (R600_NUM_PS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1678                                                      R600_NUM_VS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1679
1680         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, (R600_NUM_GS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1681                                                      R600_NUM_ES_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1682
1683         sq_dyn_gpr_size_simd_ab_0 = (R700_SIMDA_RING0((dev_priv->r600_max_gprs * 38)/64) |
1684                                      R700_SIMDA_RING1((dev_priv->r600_max_gprs * 38)/64) |
1685                                      R700_SIMDB_RING0((dev_priv->r600_max_gprs * 38)/64) |
1686                                      R700_SIMDB_RING1((dev_priv->r600_max_gprs * 38)/64));
1687
1688         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
1689         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
1690         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
1691         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
1692         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
1693         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
1694         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
1695         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
1696
1697         RADEON_WRITE(R700_PA_SC_FORCE_EOV_MAX_CNTS, (R700_FORCE_EOV_MAX_CLK_CNT(4095) |
1698                                                      R700_FORCE_EOV_MAX_REZ_CNT(255)));
1699
1700         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1701                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_TC_ONLY) |
1702                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1703         else
1704                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_VC_AND_TC) |
1705                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1706
1707         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1708         case CHIP_RV770:
1709         case CHIP_RV730:
1710         case CHIP_RV740:
1711                 gs_prim_buffer_depth = 384;
1712                 break;
1713         case CHIP_RV710:
1714                 gs_prim_buffer_depth = 128;
1715                 break;
1716         default:
1717                 break;
1718         }
1719
1720         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1721         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1722         /* Max value for this is 256 */
1723         if (vgt_gs_per_es > 256)
1724                 vgt_gs_per_es = 256;
1725
1726         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1727         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1728         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1729
1730         /* more default values. 2D/3D driver should adjust as needed */
1731         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1732         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1733         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1734         RADEON_WRITE(R600_SX_MISC, 0);
1735         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1736         RADEON_WRITE(R700_PA_SC_EDGERULE, 0xaaaaaaaa);
1737         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1738         RADEON_WRITE(R600_PA_SC_CLIPRECT_RULE, 0xffff);
1739         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1740         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1741         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1742         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1743
1744         /* clear render buffer base addresses */
1745         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1746         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1747         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1748         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1749         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1750         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1751         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1752         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1753
1754         RADEON_WRITE(R700_TCP_CNTL, 0);
1755
1756         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1757         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1758
1759         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1760
1761         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1762                                           R600_NUM_CLIP_SEQ(3)));
1763
1764 }
1765
1766 static void r600_cp_init_ring_buffer(struct drm_device *dev,
1767                                        drm_radeon_private_t *dev_priv,
1768                                        struct drm_file *file_priv)
1769 {
1770         struct drm_radeon_master_private *master_priv;
1771         u32 ring_start;
1772         u64 rptr_addr;
1773
1774         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
1775                 r700_gfx_init(dev, dev_priv);
1776         else
1777                 r600_gfx_init(dev, dev_priv);
1778
1779         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
1780         RADEON_READ(R600_GRBM_SOFT_RESET);
1781         DRM_UDELAY(15000);
1782         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
1783
1784
1785         /* Set ring buffer size */
1786 #ifdef __BIG_ENDIAN
1787         RADEON_WRITE(R600_CP_RB_CNTL,
1788                      RADEON_BUF_SWAP_32BIT |
1789                      RADEON_RB_NO_UPDATE |
1790                      (dev_priv->ring.rptr_update_l2qw << 8) |
1791                      dev_priv->ring.size_l2qw);
1792 #else
1793         RADEON_WRITE(R600_CP_RB_CNTL,
1794                      RADEON_RB_NO_UPDATE |
1795                      (dev_priv->ring.rptr_update_l2qw << 8) |
1796                      dev_priv->ring.size_l2qw);
1797 #endif
1798
1799         RADEON_WRITE(R600_CP_SEM_WAIT_TIMER, 0x4);
1800
1801         /* Set the write pointer delay */
1802         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
1803
1804 #ifdef __BIG_ENDIAN
1805         RADEON_WRITE(R600_CP_RB_CNTL,
1806                      RADEON_BUF_SWAP_32BIT |
1807                      RADEON_RB_NO_UPDATE |
1808                      RADEON_RB_RPTR_WR_ENA |
1809                      (dev_priv->ring.rptr_update_l2qw << 8) |
1810                      dev_priv->ring.size_l2qw);
1811 #else
1812         RADEON_WRITE(R600_CP_RB_CNTL,
1813                      RADEON_RB_NO_UPDATE |
1814                      RADEON_RB_RPTR_WR_ENA |
1815                      (dev_priv->ring.rptr_update_l2qw << 8) |
1816                      dev_priv->ring.size_l2qw);
1817 #endif
1818
1819         /* Initialize the ring buffer's read and write pointers */
1820         RADEON_WRITE(R600_CP_RB_RPTR_WR, 0);
1821         RADEON_WRITE(R600_CP_RB_WPTR, 0);
1822         SET_RING_HEAD(dev_priv, 0);
1823         dev_priv->ring.tail = 0;
1824
1825 #if __OS_HAS_AGP
1826         if (dev_priv->flags & RADEON_IS_AGP) {
1827                 rptr_addr = dev_priv->ring_rptr->offset
1828                         - dev->agp->base +
1829                         dev_priv->gart_vm_start;
1830         } else
1831 #endif
1832         {
1833                 rptr_addr = dev_priv->ring_rptr->offset
1834                         - ((unsigned long) dev->sg->virtual)
1835                         + dev_priv->gart_vm_start;
1836         }
1837         RADEON_WRITE(R600_CP_RB_RPTR_ADDR,
1838                      rptr_addr & 0xffffffff);
1839         RADEON_WRITE(R600_CP_RB_RPTR_ADDR_HI,
1840                      upper_32_bits(rptr_addr));
1841
1842 #ifdef __BIG_ENDIAN
1843         RADEON_WRITE(R600_CP_RB_CNTL,
1844                      RADEON_BUF_SWAP_32BIT |
1845                      (dev_priv->ring.rptr_update_l2qw << 8) |
1846                      dev_priv->ring.size_l2qw);
1847 #else
1848         RADEON_WRITE(R600_CP_RB_CNTL,
1849                      (dev_priv->ring.rptr_update_l2qw << 8) |
1850                      dev_priv->ring.size_l2qw);
1851 #endif
1852
1853 #if __OS_HAS_AGP
1854         if (dev_priv->flags & RADEON_IS_AGP) {
1855                 /* XXX */
1856                 radeon_write_agp_base(dev_priv, dev->agp->base);
1857
1858                 /* XXX */
1859                 radeon_write_agp_location(dev_priv,
1860                              (((dev_priv->gart_vm_start - 1 +
1861                                 dev_priv->gart_size) & 0xffff0000) |
1862                               (dev_priv->gart_vm_start >> 16)));
1863
1864                 ring_start = (dev_priv->cp_ring->offset
1865                               - dev->agp->base
1866                               + dev_priv->gart_vm_start);
1867         } else
1868 #endif
1869                 ring_start = (dev_priv->cp_ring->offset
1870                               - (unsigned long)dev->sg->virtual
1871                               + dev_priv->gart_vm_start);
1872
1873         RADEON_WRITE(R600_CP_RB_BASE, ring_start >> 8);
1874
1875         RADEON_WRITE(R600_CP_ME_CNTL, 0xff);
1876
1877         RADEON_WRITE(R600_CP_DEBUG, (1 << 27) | (1 << 28));
1878
1879         /* Initialize the scratch register pointer.  This will cause
1880          * the scratch register values to be written out to memory
1881          * whenever they are updated.
1882          *
1883          * We simply put this behind the ring read pointer, this works
1884          * with PCI GART as well as (whatever kind of) AGP GART
1885          */
1886         {
1887                 u64 scratch_addr;
1888
1889                 scratch_addr = RADEON_READ(R600_CP_RB_RPTR_ADDR);
1890                 scratch_addr |= ((u64)RADEON_READ(R600_CP_RB_RPTR_ADDR_HI)) << 32;
1891                 scratch_addr += R600_SCRATCH_REG_OFFSET;
1892                 scratch_addr >>= 8;
1893                 scratch_addr &= 0xffffffff;
1894
1895                 RADEON_WRITE(R600_SCRATCH_ADDR, (uint32_t)scratch_addr);
1896         }
1897
1898         RADEON_WRITE(R600_SCRATCH_UMSK, 0x7);
1899
1900         /* Turn on bus mastering */
1901         radeon_enable_bm(dev_priv);
1902
1903         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(0), 0);
1904         RADEON_WRITE(R600_LAST_FRAME_REG, 0);
1905
1906         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
1907         RADEON_WRITE(R600_LAST_DISPATCH_REG, 0);
1908
1909         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(2), 0);
1910         RADEON_WRITE(R600_LAST_CLEAR_REG, 0);
1911
1912         /* reset sarea copies of these */
1913         master_priv = file_priv->master->driver_priv;
1914         if (master_priv->sarea_priv) {
1915                 master_priv->sarea_priv->last_frame = 0;
1916                 master_priv->sarea_priv->last_dispatch = 0;
1917                 master_priv->sarea_priv->last_clear = 0;
1918         }
1919
1920         r600_do_wait_for_idle(dev_priv);
1921
1922 }
1923
1924 int r600_do_cleanup_cp(struct drm_device *dev)
1925 {
1926         drm_radeon_private_t *dev_priv = dev->dev_private;
1927         DRM_DEBUG("\n");
1928
1929         /* Make sure interrupts are disabled here because the uninstall ioctl
1930          * may not have been called from userspace and after dev_private
1931          * is freed, it's too late.
1932          */
1933         if (dev->irq_enabled)
1934                 drm_irq_uninstall(dev);
1935
1936 #if __OS_HAS_AGP
1937         if (dev_priv->flags & RADEON_IS_AGP) {
1938                 if (dev_priv->cp_ring != NULL) {
1939                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1940                         dev_priv->cp_ring = NULL;
1941                 }
1942                 if (dev_priv->ring_rptr != NULL) {
1943                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1944                         dev_priv->ring_rptr = NULL;
1945                 }
1946                 if (dev->agp_buffer_map != NULL) {
1947                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1948                         dev->agp_buffer_map = NULL;
1949                 }
1950         } else
1951 #endif
1952         {
1953
1954                 if (dev_priv->gart_info.bus_addr)
1955                         r600_page_table_cleanup(dev, &dev_priv->gart_info);
1956
1957                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB) {
1958                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1959                         dev_priv->gart_info.addr = NULL;
1960                 }
1961         }
1962         /* only clear to the start of flags */
1963         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1964
1965         return 0;
1966 }
1967
1968 int r600_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
1969                     struct drm_file *file_priv)
1970 {
1971         drm_radeon_private_t *dev_priv = dev->dev_private;
1972         struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1973
1974         DRM_DEBUG("\n");
1975
1976         mutex_init(&dev_priv->cs_mutex);
1977         r600_cs_legacy_init();
1978         /* if we require new memory map but we don't have it fail */
1979         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
1980                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
1981                 r600_do_cleanup_cp(dev);
1982                 return -EINVAL;
1983         }
1984
1985         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
1986                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
1987                 dev_priv->flags &= ~RADEON_IS_AGP;
1988                 /* The writeback test succeeds, but when writeback is enabled,
1989                  * the ring buffer read ptr update fails after first 128 bytes.
1990                  */
1991                 radeon_no_wb = 1;
1992         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
1993                  && !init->is_pci) {
1994                 DRM_DEBUG("Restoring AGP flag\n");
1995                 dev_priv->flags |= RADEON_IS_AGP;
1996         }
1997
1998         dev_priv->usec_timeout = init->usec_timeout;
1999         if (dev_priv->usec_timeout < 1 ||
2000             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
2001                 DRM_DEBUG("TIMEOUT problem!\n");
2002                 r600_do_cleanup_cp(dev);
2003                 return -EINVAL;
2004         }
2005
2006         /* Enable vblank on CRTC1 for older X servers
2007          */
2008         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
2009         dev_priv->do_boxes = 0;
2010         dev_priv->cp_mode = init->cp_mode;
2011
2012         /* We don't support anything other than bus-mastering ring mode,
2013          * but the ring can be in either AGP or PCI space for the ring
2014          * read pointer.
2015          */
2016         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
2017             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
2018                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
2019                 r600_do_cleanup_cp(dev);
2020                 return -EINVAL;
2021         }
2022
2023         switch (init->fb_bpp) {
2024         case 16:
2025                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
2026                 break;
2027         case 32:
2028         default:
2029                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
2030                 break;
2031         }
2032         dev_priv->front_offset = init->front_offset;
2033         dev_priv->front_pitch = init->front_pitch;
2034         dev_priv->back_offset = init->back_offset;
2035         dev_priv->back_pitch = init->back_pitch;
2036
2037         dev_priv->ring_offset = init->ring_offset;
2038         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
2039         dev_priv->buffers_offset = init->buffers_offset;
2040         dev_priv->gart_textures_offset = init->gart_textures_offset;
2041
2042         master_priv->sarea = drm_getsarea(dev);
2043         if (!master_priv->sarea) {
2044                 DRM_ERROR("could not find sarea!\n");
2045                 r600_do_cleanup_cp(dev);
2046                 return -EINVAL;
2047         }
2048
2049         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
2050         if (!dev_priv->cp_ring) {
2051                 DRM_ERROR("could not find cp ring region!\n");
2052                 r600_do_cleanup_cp(dev);
2053                 return -EINVAL;
2054         }
2055         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
2056         if (!dev_priv->ring_rptr) {
2057                 DRM_ERROR("could not find ring read pointer!\n");
2058                 r600_do_cleanup_cp(dev);
2059                 return -EINVAL;
2060         }
2061         dev->agp_buffer_token = init->buffers_offset;
2062         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
2063         if (!dev->agp_buffer_map) {
2064                 DRM_ERROR("could not find dma buffer region!\n");
2065                 r600_do_cleanup_cp(dev);
2066                 return -EINVAL;
2067         }
2068
2069         if (init->gart_textures_offset) {
2070                 dev_priv->gart_textures =
2071                     drm_core_findmap(dev, init->gart_textures_offset);
2072                 if (!dev_priv->gart_textures) {
2073                         DRM_ERROR("could not find GART texture region!\n");
2074                         r600_do_cleanup_cp(dev);
2075                         return -EINVAL;
2076                 }
2077         }
2078
2079 #if __OS_HAS_AGP
2080         /* XXX */
2081         if (dev_priv->flags & RADEON_IS_AGP) {
2082                 drm_core_ioremap_wc(dev_priv->cp_ring, dev);
2083                 drm_core_ioremap_wc(dev_priv->ring_rptr, dev);
2084                 drm_core_ioremap_wc(dev->agp_buffer_map, dev);
2085                 if (!dev_priv->cp_ring->handle ||
2086                     !dev_priv->ring_rptr->handle ||
2087                     !dev->agp_buffer_map->handle) {
2088                         DRM_ERROR("could not find ioremap agp regions!\n");
2089                         r600_do_cleanup_cp(dev);
2090                         return -EINVAL;
2091                 }
2092         } else
2093 #endif
2094         {
2095                 dev_priv->cp_ring->handle = (void *)(unsigned long)dev_priv->cp_ring->offset;
2096                 dev_priv->ring_rptr->handle =
2097                         (void *)(unsigned long)dev_priv->ring_rptr->offset;
2098                 dev->agp_buffer_map->handle =
2099                         (void *)(unsigned long)dev->agp_buffer_map->offset;
2100
2101                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
2102                           dev_priv->cp_ring->handle);
2103                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
2104                           dev_priv->ring_rptr->handle);
2105                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
2106                           dev->agp_buffer_map->handle);
2107         }
2108
2109         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 24;
2110         dev_priv->fb_size =
2111                 (((radeon_read_fb_location(dev_priv) & 0xffff0000u) << 8) + 0x1000000)
2112                 - dev_priv->fb_location;
2113
2114         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
2115                                         ((dev_priv->front_offset
2116                                           + dev_priv->fb_location) >> 10));
2117
2118         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
2119                                        ((dev_priv->back_offset
2120                                          + dev_priv->fb_location) >> 10));
2121
2122         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
2123                                         ((dev_priv->depth_offset
2124                                           + dev_priv->fb_location) >> 10));
2125
2126         dev_priv->gart_size = init->gart_size;
2127
2128         /* New let's set the memory map ... */
2129         if (dev_priv->new_memmap) {
2130                 u32 base = 0;
2131
2132                 DRM_INFO("Setting GART location based on new memory map\n");
2133
2134                 /* If using AGP, try to locate the AGP aperture at the same
2135                  * location in the card and on the bus, though we have to
2136                  * align it down.
2137                  */
2138 #if __OS_HAS_AGP
2139                 /* XXX */
2140                 if (dev_priv->flags & RADEON_IS_AGP) {
2141                         base = dev->agp->base;
2142                         /* Check if valid */
2143                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
2144                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
2145                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
2146                                          dev->agp->base);
2147                                 base = 0;
2148                         }
2149                 }
2150 #endif
2151                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
2152                 if (base == 0) {
2153                         base = dev_priv->fb_location + dev_priv->fb_size;
2154                         if (base < dev_priv->fb_location ||
2155                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
2156                                 base = dev_priv->fb_location
2157                                         - dev_priv->gart_size;
2158                 }
2159                 dev_priv->gart_vm_start = base & 0xffc00000u;
2160                 if (dev_priv->gart_vm_start != base)
2161                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
2162                                  base, dev_priv->gart_vm_start);
2163         }
2164
2165 #if __OS_HAS_AGP
2166         /* XXX */
2167         if (dev_priv->flags & RADEON_IS_AGP)
2168                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2169                                                  - dev->agp->base
2170                                                  + dev_priv->gart_vm_start);
2171         else
2172 #endif
2173                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2174                                                  - (unsigned long)dev->sg->virtual
2175                                                  + dev_priv->gart_vm_start);
2176
2177         DRM_DEBUG("fb 0x%08x size %d\n",
2178                   (unsigned int) dev_priv->fb_location,
2179                   (unsigned int) dev_priv->fb_size);
2180         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
2181         DRM_DEBUG("dev_priv->gart_vm_start 0x%08x\n",
2182                   (unsigned int) dev_priv->gart_vm_start);
2183         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%08lx\n",
2184                   dev_priv->gart_buffers_offset);
2185
2186         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
2187         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
2188                               + init->ring_size / sizeof(u32));
2189         dev_priv->ring.size = init->ring_size;
2190         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
2191
2192         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
2193         dev_priv->ring.rptr_update_l2qw = drm_order(/* init->rptr_update */ 4096 / 8);
2194
2195         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
2196         dev_priv->ring.fetch_size_l2ow = drm_order(/* init->fetch_size */ 32 / 16);
2197
2198         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
2199
2200         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
2201
2202 #if __OS_HAS_AGP
2203         if (dev_priv->flags & RADEON_IS_AGP) {
2204                 /* XXX turn off pcie gart */
2205         } else
2206 #endif
2207         {
2208                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
2209                 /* if we have an offset set from userspace */
2210                 if (!dev_priv->pcigart_offset_set) {
2211                         DRM_ERROR("Need gart offset from userspace\n");
2212                         r600_do_cleanup_cp(dev);
2213                         return -EINVAL;
2214                 }
2215
2216                 DRM_DEBUG("Using gart offset 0x%08lx\n", dev_priv->pcigart_offset);
2217
2218                 dev_priv->gart_info.bus_addr =
2219                         dev_priv->pcigart_offset + dev_priv->fb_location;
2220                 dev_priv->gart_info.mapping.offset =
2221                         dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
2222                 dev_priv->gart_info.mapping.size =
2223                         dev_priv->gart_info.table_size;
2224
2225                 drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
2226                 if (!dev_priv->gart_info.mapping.handle) {
2227                         DRM_ERROR("ioremap failed.\n");
2228                         r600_do_cleanup_cp(dev);
2229                         return -EINVAL;
2230                 }
2231
2232                 dev_priv->gart_info.addr =
2233                         dev_priv->gart_info.mapping.handle;
2234
2235                 DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
2236                           dev_priv->gart_info.addr,
2237                           dev_priv->pcigart_offset);
2238
2239                 if (!r600_page_table_init(dev)) {
2240                         DRM_ERROR("Failed to init GART table\n");
2241                         r600_do_cleanup_cp(dev);
2242                         return -EINVAL;
2243                 }
2244
2245                 if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2246                         r700_vm_init(dev);
2247                 else
2248                         r600_vm_init(dev);
2249         }
2250
2251         if (!dev_priv->me_fw || !dev_priv->pfp_fw) {
2252                 int err = r600_cp_init_microcode(dev_priv);
2253                 if (err) {
2254                         DRM_ERROR("Failed to load firmware!\n");
2255                         r600_do_cleanup_cp(dev);
2256                         return err;
2257                 }
2258         }
2259         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2260                 r700_cp_load_microcode(dev_priv);
2261         else
2262                 r600_cp_load_microcode(dev_priv);
2263
2264         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2265
2266         dev_priv->last_buf = 0;
2267
2268         r600_do_engine_reset(dev);
2269         r600_test_writeback(dev_priv);
2270
2271         return 0;
2272 }
2273
2274 int r600_do_resume_cp(struct drm_device *dev, struct drm_file *file_priv)
2275 {
2276         drm_radeon_private_t *dev_priv = dev->dev_private;
2277
2278         DRM_DEBUG("\n");
2279         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)) {
2280                 r700_vm_init(dev);
2281                 r700_cp_load_microcode(dev_priv);
2282         } else {
2283                 r600_vm_init(dev);
2284                 r600_cp_load_microcode(dev_priv);
2285         }
2286         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2287         r600_do_engine_reset(dev);
2288
2289         return 0;
2290 }
2291
2292 /* Wait for the CP to go idle.
2293  */
2294 int r600_do_cp_idle(drm_radeon_private_t *dev_priv)
2295 {
2296         RING_LOCALS;
2297         DRM_DEBUG("\n");
2298
2299         BEGIN_RING(5);
2300         OUT_RING(CP_PACKET3(R600_IT_EVENT_WRITE, 0));
2301         OUT_RING(R600_CACHE_FLUSH_AND_INV_EVENT);
2302         /* wait for 3D idle clean */
2303         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 1));
2304         OUT_RING((R600_WAIT_UNTIL - R600_SET_CONFIG_REG_OFFSET) >> 2);
2305         OUT_RING(RADEON_WAIT_3D_IDLE | RADEON_WAIT_3D_IDLECLEAN);
2306
2307         ADVANCE_RING();
2308         COMMIT_RING();
2309
2310         return r600_do_wait_for_idle(dev_priv);
2311 }
2312
2313 /* Start the Command Processor.
2314  */
2315 void r600_do_cp_start(drm_radeon_private_t *dev_priv)
2316 {
2317         u32 cp_me;
2318         RING_LOCALS;
2319         DRM_DEBUG("\n");
2320
2321         BEGIN_RING(7);
2322         OUT_RING(CP_PACKET3(R600_IT_ME_INITIALIZE, 5));
2323         OUT_RING(0x00000001);
2324         if (((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV770))
2325                 OUT_RING(0x00000003);
2326         else
2327                 OUT_RING(0x00000000);
2328         OUT_RING((dev_priv->r600_max_hw_contexts - 1));
2329         OUT_RING(R600_ME_INITIALIZE_DEVICE_ID(1));
2330         OUT_RING(0x00000000);
2331         OUT_RING(0x00000000);
2332         ADVANCE_RING();
2333         COMMIT_RING();
2334
2335         /* set the mux and reset the halt bit */
2336         cp_me = 0xff;
2337         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2338
2339         dev_priv->cp_running = 1;
2340
2341 }
2342
2343 void r600_do_cp_reset(drm_radeon_private_t *dev_priv)
2344 {
2345         u32 cur_read_ptr;
2346         DRM_DEBUG("\n");
2347
2348         cur_read_ptr = RADEON_READ(R600_CP_RB_RPTR);
2349         RADEON_WRITE(R600_CP_RB_WPTR, cur_read_ptr);
2350         SET_RING_HEAD(dev_priv, cur_read_ptr);
2351         dev_priv->ring.tail = cur_read_ptr;
2352 }
2353
2354 void r600_do_cp_stop(drm_radeon_private_t *dev_priv)
2355 {
2356         uint32_t cp_me;
2357
2358         DRM_DEBUG("\n");
2359
2360         cp_me = 0xff | R600_CP_ME_HALT;
2361
2362         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2363
2364         dev_priv->cp_running = 0;
2365 }
2366
2367 int r600_cp_dispatch_indirect(struct drm_device *dev,
2368                               struct drm_buf *buf, int start, int end)
2369 {
2370         drm_radeon_private_t *dev_priv = dev->dev_private;
2371         RING_LOCALS;
2372
2373         if (start != end) {
2374                 unsigned long offset = (dev_priv->gart_buffers_offset
2375                                         + buf->offset + start);
2376                 int dwords = (end - start + 3) / sizeof(u32);
2377
2378                 DRM_DEBUG("dwords:%d\n", dwords);
2379                 DRM_DEBUG("offset 0x%lx\n", offset);
2380
2381
2382                 /* Indirect buffer data must be a multiple of 16 dwords.
2383                  * pad the data with a Type-2 CP packet.
2384                  */
2385                 while (dwords & 0xf) {
2386                         u32 *data = (u32 *)
2387                             ((char *)dev->agp_buffer_map->handle
2388                              + buf->offset + start);
2389                         data[dwords++] = RADEON_CP_PACKET2;
2390                 }
2391
2392                 /* Fire off the indirect buffer */
2393                 BEGIN_RING(4);
2394                 OUT_RING(CP_PACKET3(R600_IT_INDIRECT_BUFFER, 2));
2395                 OUT_RING((offset & 0xfffffffc));
2396                 OUT_RING((upper_32_bits(offset) & 0xff));
2397                 OUT_RING(dwords);
2398                 ADVANCE_RING();
2399         }
2400
2401         return 0;
2402 }
2403
2404 void r600_cp_dispatch_swap(struct drm_device *dev, struct drm_file *file_priv)
2405 {
2406         drm_radeon_private_t *dev_priv = dev->dev_private;
2407         struct drm_master *master = file_priv->master;
2408         struct drm_radeon_master_private *master_priv = master->driver_priv;
2409         drm_radeon_sarea_t *sarea_priv = master_priv->sarea_priv;
2410         int nbox = sarea_priv->nbox;
2411         struct drm_clip_rect *pbox = sarea_priv->boxes;
2412         int i, cpp, src_pitch, dst_pitch;
2413         uint64_t src, dst;
2414         RING_LOCALS;
2415         DRM_DEBUG("\n");
2416
2417         if (dev_priv->color_fmt == RADEON_COLOR_FORMAT_ARGB8888)
2418                 cpp = 4;
2419         else
2420                 cpp = 2;
2421
2422         if (sarea_priv->pfCurrentPage == 0) {
2423                 src_pitch = dev_priv->back_pitch;
2424                 dst_pitch = dev_priv->front_pitch;
2425                 src = dev_priv->back_offset + dev_priv->fb_location;
2426                 dst = dev_priv->front_offset + dev_priv->fb_location;
2427         } else {
2428                 src_pitch = dev_priv->front_pitch;
2429                 dst_pitch = dev_priv->back_pitch;
2430                 src = dev_priv->front_offset + dev_priv->fb_location;
2431                 dst = dev_priv->back_offset + dev_priv->fb_location;
2432         }
2433
2434         if (r600_prepare_blit_copy(dev, file_priv)) {
2435                 DRM_ERROR("unable to allocate vertex buffer for swap buffer\n");
2436                 return;
2437         }
2438         for (i = 0; i < nbox; i++) {
2439                 int x = pbox[i].x1;
2440                 int y = pbox[i].y1;
2441                 int w = pbox[i].x2 - x;
2442                 int h = pbox[i].y2 - y;
2443
2444                 DRM_DEBUG("%d,%d-%d,%d\n", x, y, w, h);
2445
2446                 r600_blit_swap(dev,
2447                                src, dst,
2448                                x, y, x, y, w, h,
2449                                src_pitch, dst_pitch, cpp);
2450         }
2451         r600_done_blit_copy(dev);
2452
2453         /* Increment the frame counter.  The client-side 3D driver must
2454          * throttle the framerate by waiting for this value before
2455          * performing the swapbuffer ioctl.
2456          */
2457         sarea_priv->last_frame++;
2458
2459         BEGIN_RING(3);
2460         R600_FRAME_AGE(sarea_priv->last_frame);
2461         ADVANCE_RING();
2462 }
2463
2464 int r600_cp_dispatch_texture(struct drm_device *dev,
2465                              struct drm_file *file_priv,
2466                              drm_radeon_texture_t *tex,
2467                              drm_radeon_tex_image_t *image)
2468 {
2469         drm_radeon_private_t *dev_priv = dev->dev_private;
2470         struct drm_buf *buf;
2471         u32 *buffer;
2472         const u8 __user *data;
2473         int size, pass_size;
2474         u64 src_offset, dst_offset;
2475
2476         if (!radeon_check_offset(dev_priv, tex->offset)) {
2477                 DRM_ERROR("Invalid destination offset\n");
2478                 return -EINVAL;
2479         }
2480
2481         /* this might fail for zero-sized uploads - are those illegal? */
2482         if (!radeon_check_offset(dev_priv, tex->offset + tex->height * tex->pitch - 1)) {
2483                 DRM_ERROR("Invalid final destination offset\n");
2484                 return -EINVAL;
2485         }
2486
2487         size = tex->height * tex->pitch;
2488
2489         if (size == 0)
2490                 return 0;
2491
2492         dst_offset = tex->offset;
2493
2494         if (r600_prepare_blit_copy(dev, file_priv)) {
2495                 DRM_ERROR("unable to allocate vertex buffer for swap buffer\n");
2496                 return -EAGAIN;
2497         }
2498         do {
2499                 data = (const u8 __user *)image->data;
2500                 pass_size = size;
2501
2502                 buf = radeon_freelist_get(dev);
2503                 if (!buf) {
2504                         DRM_DEBUG("EAGAIN\n");
2505                         if (DRM_COPY_TO_USER(tex->image, image, sizeof(*image)))
2506                                 return -EFAULT;
2507                         return -EAGAIN;
2508                 }
2509
2510                 if (pass_size > buf->total)
2511                         pass_size = buf->total;
2512
2513                 /* Dispatch the indirect buffer.
2514                  */
2515                 buffer =
2516                     (u32 *) ((char *)dev->agp_buffer_map->handle + buf->offset);
2517
2518                 if (DRM_COPY_FROM_USER(buffer, data, pass_size)) {
2519                         DRM_ERROR("EFAULT on pad, %d bytes\n", pass_size);
2520                         return -EFAULT;
2521                 }
2522
2523                 buf->file_priv = file_priv;
2524                 buf->used = pass_size;
2525                 src_offset = dev_priv->gart_buffers_offset + buf->offset;
2526
2527                 r600_blit_copy(dev, src_offset, dst_offset, pass_size);
2528
2529                 radeon_cp_discard_buffer(dev, file_priv->master, buf);
2530
2531                 /* Update the input parameters for next time */
2532                 image->data = (const u8 __user *)image->data + pass_size;
2533                 dst_offset += pass_size;
2534                 size -= pass_size;
2535         } while (size > 0);
2536         r600_done_blit_copy(dev);
2537
2538         return 0;
2539 }
2540
2541 /*
2542  * Legacy cs ioctl
2543  */
2544 static u32 radeon_cs_id_get(struct drm_radeon_private *radeon)
2545 {
2546         /* FIXME: check if wrap affect last reported wrap & sequence */
2547         radeon->cs_id_scnt = (radeon->cs_id_scnt + 1) & 0x00FFFFFF;
2548         if (!radeon->cs_id_scnt) {
2549                 /* increment wrap counter */
2550                 radeon->cs_id_wcnt += 0x01000000;
2551                 /* valid sequence counter start at 1 */
2552                 radeon->cs_id_scnt = 1;
2553         }
2554         return (radeon->cs_id_scnt | radeon->cs_id_wcnt);
2555 }
2556
2557 static void r600_cs_id_emit(drm_radeon_private_t *dev_priv, u32 *id)
2558 {
2559         RING_LOCALS;
2560
2561         *id = radeon_cs_id_get(dev_priv);
2562
2563         /* SCRATCH 2 */
2564         BEGIN_RING(3);
2565         R600_CLEAR_AGE(*id);
2566         ADVANCE_RING();
2567         COMMIT_RING();
2568 }
2569
2570 static int r600_ib_get(struct drm_device *dev,
2571                         struct drm_file *fpriv,
2572                         struct drm_buf **buffer)
2573 {
2574         struct drm_buf *buf;
2575
2576         *buffer = NULL;
2577         buf = radeon_freelist_get(dev);
2578         if (!buf) {
2579                 return -EBUSY;
2580         }
2581         buf->file_priv = fpriv;
2582         *buffer = buf;
2583         return 0;
2584 }
2585
2586 static void r600_ib_free(struct drm_device *dev, struct drm_buf *buf,
2587                         struct drm_file *fpriv, int l, int r)
2588 {
2589         drm_radeon_private_t *dev_priv = dev->dev_private;
2590
2591         if (buf) {
2592                 if (!r)
2593                         r600_cp_dispatch_indirect(dev, buf, 0, l * 4);
2594                 radeon_cp_discard_buffer(dev, fpriv->master, buf);
2595                 COMMIT_RING();
2596         }
2597 }
2598
2599 int r600_cs_legacy_ioctl(struct drm_device *dev, void *data, struct drm_file *fpriv)
2600 {
2601         struct drm_radeon_private *dev_priv = dev->dev_private;
2602         struct drm_radeon_cs *cs = data;
2603         struct drm_buf *buf;
2604         unsigned family;
2605         int l, r = 0;
2606         u32 *ib, cs_id = 0;
2607
2608         if (dev_priv == NULL) {
2609                 DRM_ERROR("called with no initialization\n");
2610                 return -EINVAL;
2611         }
2612         family = dev_priv->flags & RADEON_FAMILY_MASK;
2613         if (family < CHIP_R600) {
2614                 DRM_ERROR("cs ioctl valid only for R6XX & R7XX in legacy mode\n");
2615                 return -EINVAL;
2616         }
2617         mutex_lock(&dev_priv->cs_mutex);
2618         /* get ib */
2619         r = r600_ib_get(dev, fpriv, &buf);
2620         if (r) {
2621                 DRM_ERROR("ib_get failed\n");
2622                 goto out;
2623         }
2624         ib = dev->agp_buffer_map->handle + buf->offset;
2625         /* now parse command stream */
2626         r = r600_cs_legacy(dev, data,  fpriv, family, ib, &l);
2627         if (r) {
2628                 goto out;
2629         }
2630
2631 out:
2632         r600_ib_free(dev, buf, fpriv, l, r);
2633         /* emit cs id sequence */
2634         r600_cs_id_emit(dev_priv, &cs_id);
2635         cs->cs_id = cs_id;
2636         mutex_unlock(&dev_priv->cs_mutex);
2637         return r;
2638 }
2639
2640 void r600_cs_legacy_get_tiling_conf(struct drm_device *dev, u32 *npipes, u32 *nbanks, u32 *group_size)
2641 {
2642         struct drm_radeon_private *dev_priv = dev->dev_private;
2643
2644         *npipes = dev_priv->r600_npipes;
2645         *nbanks = dev_priv->r600_nbanks;
2646         *group_size = dev_priv->r600_group_size;
2647 }