sys/dev/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  * ------------------------ This file is DEPRECATED! -------------------------
  29  *
  30  * $FreeBSD: head/sys/dev/drm2/radeon/r600_cp.c 254885 2013-08-25 19:37:15Z dumbbell $
  31  */
  32
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/linker.h>
  36 #include <sys/firmware.h>
  37 #include <linux/module.h>
  38
  39 #include <drm/drmP.h>
  40 #include <uapi_drm/radeon_drm.h>
  41 #include "radeon_drv.h"
  42 #include "r600_cp.h"
  43
  44 #define PFP_UCODE_SIZE 576
  45 #define PM4_UCODE_SIZE 1792
  46 #define R700_PFP_UCODE_SIZE 848
  47 #define R700_PM4_UCODE_SIZE 1360
  48
  49 /* Firmware Names */
  50 MODULE_FIRMWARE("radeon/R600_pfp.bin");
  51 MODULE_FIRMWARE("radeon/R600_me.bin");
  52 MODULE_FIRMWARE("radeon/RV610_pfp.bin");
  53 MODULE_FIRMWARE("radeon/RV610_me.bin");
  54 MODULE_FIRMWARE("radeon/RV630_pfp.bin");
  55 MODULE_FIRMWARE("radeon/RV630_me.bin");
  56 MODULE_FIRMWARE("radeon/RV620_pfp.bin");
  57 MODULE_FIRMWARE("radeon/RV620_me.bin");
  58 MODULE_FIRMWARE("radeon/RV635_pfp.bin");
  59 MODULE_FIRMWARE("radeon/RV635_me.bin");
  60 MODULE_FIRMWARE("radeon/RV670_pfp.bin");
  61 MODULE_FIRMWARE("radeon/RV670_me.bin");
  62 MODULE_FIRMWARE("radeon/RS780_pfp.bin");
  63 MODULE_FIRMWARE("radeon/RS780_me.bin");
  64 MODULE_FIRMWARE("radeon/RV770_pfp.bin");
  65 MODULE_FIRMWARE("radeon/RV770_me.bin");
  66 MODULE_FIRMWARE("radeon/RV730_pfp.bin");
  67 MODULE_FIRMWARE("radeon/RV730_me.bin");
  68 MODULE_FIRMWARE("radeon/RV710_pfp.bin");
  69 MODULE_FIRMWARE("radeon/RV710_me.bin");
  70
  71 # define ATI_PCIGART_PAGE_SIZE          4096    /**< PCI GART page size */
  72 # define ATI_PCIGART_PAGE_MASK          (~(ATI_PCIGART_PAGE_SIZE-1))
  73
  74 #define R600_PTE_VALID     (1 << 0)
  75 #define R600_PTE_SYSTEM    (1 << 1)
  76 #define R600_PTE_SNOOPED   (1 << 2)
  77 #define R600_PTE_READABLE  (1 << 5)
  78 #define R600_PTE_WRITEABLE (1 << 6)
  79
  80 /* MAX values used for gfx init */
  81 #define R6XX_MAX_SH_GPRS           256
  82 #define R6XX_MAX_TEMP_GPRS         16
  83 #define R6XX_MAX_SH_THREADS        256
  84 #define R6XX_MAX_SH_STACK_ENTRIES  4096
  85 #define R6XX_MAX_BACKENDS          8
  86 #define R6XX_MAX_BACKENDS_MASK     0xff
  87 #define R6XX_MAX_SIMDS             8
  88 #define R6XX_MAX_SIMDS_MASK        0xff
  89 #define R6XX_MAX_PIPES             8
  90 #define R6XX_MAX_PIPES_MASK        0xff
  91
  92 #define R7XX_MAX_SH_GPRS           256
  93 #define R7XX_MAX_TEMP_GPRS         16
  94 #define R7XX_MAX_SH_THREADS        256
  95 #define R7XX_MAX_SH_STACK_ENTRIES  4096
  96 #define R7XX_MAX_BACKENDS          8
  97 #define R7XX_MAX_BACKENDS_MASK     0xff
  98 #define R7XX_MAX_SIMDS             16
  99 #define R7XX_MAX_SIMDS_MASK        0xffff
 100 #define R7XX_MAX_PIPES             8
 101 #define R7XX_MAX_PIPES_MASK        0xff
 102
 103 static int r600_do_wait_for_fifo(drm_radeon_private_t *dev_priv, int entries)
 104 {
 105         int i;
 106
 107         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 108
 109         for (i = 0; i < dev_priv->usec_timeout; i++) {
 110                 int slots;
 111                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 112                         slots = (RADEON_READ(R600_GRBM_STATUS)
 113                                  & R700_CMDFIFO_AVAIL_MASK);
 114                 else
 115                         slots = (RADEON_READ(R600_GRBM_STATUS)
 116                                  & R600_CMDFIFO_AVAIL_MASK);
 117                 if (slots >= entries)
 118                         return 0;
 119                 DRM_UDELAY(1);
 120         }
 121         DRM_INFO("wait for fifo failed status : 0x%08X 0x%08X\n",
 122                  RADEON_READ(R600_GRBM_STATUS),
 123                  RADEON_READ(R600_GRBM_STATUS2));
 124
 125         return -EBUSY;
 126 }
 127
 128 static int r600_do_wait_for_idle(drm_radeon_private_t *dev_priv)
 129 {
 130         int i, ret;
 131
 132         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 133
 134         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 135                 ret = r600_do_wait_for_fifo(dev_priv, 8);
 136         else
 137                 ret = r600_do_wait_for_fifo(dev_priv, 16);
 138         if (ret)
 139                 return ret;
 140         for (i = 0; i < dev_priv->usec_timeout; i++) {
 141                 if (!(RADEON_READ(R600_GRBM_STATUS) & R600_GUI_ACTIVE))
 142                         return 0;
 143                 DRM_UDELAY(1);
 144         }
 145         DRM_INFO("wait idle failed status : 0x%08X 0x%08X\n",
 146                  RADEON_READ(R600_GRBM_STATUS),
 147                  RADEON_READ(R600_GRBM_STATUS2));
 148
 149         return -EBUSY;
 150 }
 151
 152 void r600_page_table_cleanup(struct drm_device *dev, struct drm_ati_pcigart_info *gart_info)
 153 {
 154         struct drm_sg_mem *entry = dev->sg;
 155 #ifdef __linux__
 156         int max_pages;
 157         int pages;
 158         int i;
 159 #endif
 160
 161         if (!entry)
 162                 return;
 163
 164         if (gart_info->bus_addr) {
 165 #ifdef __linux__
 166                 max_pages = (gart_info->table_size / sizeof(u64));
 167                 pages = (entry->pages <= max_pages)
 168                   ? entry->pages : max_pages;
 169
 170                 for (i = 0; i < pages; i++) {
 171                         if (!entry->busaddr[i])
 172                                 break;
 173                         pci_unmap_page(dev->pdev, entry->busaddr[i],
 174                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 175                 }
 176 #endif
 177                 if (gart_info->gart_table_location == DRM_ATI_GART_MAIN)
 178                         gart_info->bus_addr = 0;
 179         }
 180 }
 181
 182 /* R600 has page table setup */
 183 int r600_page_table_init(struct drm_device *dev)
 184 {
 185         drm_radeon_private_t *dev_priv = dev->dev_private;
 186         struct drm_ati_pcigart_info *gart_info = &dev_priv->gart_info;
 187         struct drm_local_map *map = &gart_info->mapping;
 188         struct drm_sg_mem *entry = dev->sg;
 189         int ret = 0;
 190         int i, j;
 191         int pages;
 192         u64 page_base;
 193         dma_addr_t entry_addr;
 194         int max_ati_pages, max_real_pages, gart_idx;
 195
 196         /* okay page table is available - lets rock */
 197         max_ati_pages = (gart_info->table_size / sizeof(u64));
 198         max_real_pages = max_ati_pages / (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE);
 199
 200         pages = (entry->pages <= max_real_pages) ?
 201                 entry->pages : max_real_pages;
 202
 203         memset_io((void __iomem *)map->handle, 0, max_ati_pages * sizeof(u64));
 204
 205         gart_idx = 0;
 206         for (i = 0; i < pages; i++) {
 207 #ifdef __linux__
 208                 entry->busaddr[i] = pci_map_page(dev->pdev,
 209                                                  entry->pagelist[i], 0,
 210                                                  PAGE_SIZE,
 211                                                  PCI_DMA_BIDIRECTIONAL);
 212                 if (pci_dma_mapping_error(dev->pdev, entry->busaddr[i])) {
 213                         DRM_ERROR("unable to map PCIGART pages!\n");
 214                         r600_page_table_cleanup(dev, gart_info);
 215                         goto done;
 216                 }
 217 #endif
 218                 entry_addr = entry->busaddr[i];
 219                 for (j = 0; j < (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE); j++) {
 220                         page_base = (u64) entry_addr & ATI_PCIGART_PAGE_MASK;
 221                         page_base |= R600_PTE_VALID | R600_PTE_SYSTEM | R600_PTE_SNOOPED;
 222                         page_base |= R600_PTE_READABLE | R600_PTE_WRITEABLE;
 223
 224                         DRM_WRITE64(map, gart_idx * sizeof(u64), page_base);
 225
 226                         gart_idx++;
 227
 228                         if ((i % 128) == 0)
 229                                 DRM_DEBUG("page entry %d: 0x%016llx\n",
 230                                     i, (unsigned long long)page_base);
 231                         entry_addr += ATI_PCIGART_PAGE_SIZE;
 232                 }
 233         }
 234         ret = 1;
 235 #ifdef __linux__
 236 done:
 237 #endif
 238         return ret;
 239 }
 240
 241 static void r600_vm_flush_gart_range(struct drm_device *dev)
 242 {
 243         drm_radeon_private_t *dev_priv = dev->dev_private;
 244         u32 resp, countdown = 1000;
 245         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 246         RADEON_WRITE(R600_VM_CONTEXT0_INVALIDATION_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 247         RADEON_WRITE(R600_VM_CONTEXT0_REQUEST_RESPONSE, 2);
 248
 249         do {
 250                 resp = RADEON_READ(R600_VM_CONTEXT0_REQUEST_RESPONSE);
 251                 countdown--;
 252                 DRM_UDELAY(1);
 253         } while (((resp & 0xf0) == 0) && countdown);
 254 }
 255
 256 static void r600_vm_init(struct drm_device *dev)
 257 {
 258         drm_radeon_private_t *dev_priv = dev->dev_private;
 259         /* initialise the VM to use the page table we constructed up there */
 260         u32 vm_c0, i;
 261         u32 mc_rd_a;
 262         u32 vm_l2_cntl, vm_l2_cntl3;
 263         /* okay set up the PCIE aperture type thingo */
 264         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 265         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 266         RADEON_WRITE(R600_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 267
 268         /* setup MC RD a */
 269         mc_rd_a = R600_MCD_L1_TLB | R600_MCD_L1_FRAG_PROC | R600_MCD_SYSTEM_ACCESS_MODE_IN_SYS |
 270                 R600_MCD_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU | R600_MCD_EFFECTIVE_L1_TLB_SIZE(5) |
 271                 R600_MCD_EFFECTIVE_L1_QUEUE_SIZE(5) | R600_MCD_WAIT_L2_QUERY;
 272
 273         RADEON_WRITE(R600_MCD_RD_A_CNTL, mc_rd_a);
 274         RADEON_WRITE(R600_MCD_RD_B_CNTL, mc_rd_a);
 275
 276         RADEON_WRITE(R600_MCD_WR_A_CNTL, mc_rd_a);
 277         RADEON_WRITE(R600_MCD_WR_B_CNTL, mc_rd_a);
 278
 279         RADEON_WRITE(R600_MCD_RD_GFX_CNTL, mc_rd_a);
 280         RADEON_WRITE(R600_MCD_WR_GFX_CNTL, mc_rd_a);
 281
 282         RADEON_WRITE(R600_MCD_RD_SYS_CNTL, mc_rd_a);
 283         RADEON_WRITE(R600_MCD_WR_SYS_CNTL, mc_rd_a);
 284
 285         RADEON_WRITE(R600_MCD_RD_HDP_CNTL, mc_rd_a | R600_MCD_L1_STRICT_ORDERING);
 286         RADEON_WRITE(R600_MCD_WR_HDP_CNTL, mc_rd_a /*| R600_MCD_L1_STRICT_ORDERING*/);
 287
 288         RADEON_WRITE(R600_MCD_RD_PDMA_CNTL, mc_rd_a);
 289         RADEON_WRITE(R600_MCD_WR_PDMA_CNTL, mc_rd_a);
 290
 291         RADEON_WRITE(R600_MCD_RD_SEM_CNTL, mc_rd_a | R600_MCD_SEMAPHORE_MODE);
 292         RADEON_WRITE(R600_MCD_WR_SEM_CNTL, mc_rd_a);
 293
 294         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 295         vm_l2_cntl |= R600_VM_L2_CNTL_QUEUE_SIZE(7);
 296         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 297
 298         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 299         vm_l2_cntl3 = (R600_VM_L2_CNTL3_BANK_SELECT_0(0) |
 300                        R600_VM_L2_CNTL3_BANK_SELECT_1(1) |
 301                        R600_VM_L2_CNTL3_CACHE_UPDATE_MODE(2));
 302         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 303
 304         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 305
 306         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 307
 308         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 309
 310         /* disable all other contexts */
 311         for (i = 1; i < 8; i++)
 312                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 313
 314         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 315         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 316         RADEON_WRITE(R600_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 317
 318         r600_vm_flush_gart_range(dev);
 319 }
 320
 321 static int r600_cp_init_microcode(drm_radeon_private_t *dev_priv)
 322 {
 323         struct platform_device *pdev;
 324         const char *chip_name;
 325         size_t pfp_req_size, me_req_size;
 326         char fw_name[30];
 327         int err;
 328
 329         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
 330         case CHIP_R600:  chip_name = "R600";  break;
 331         case CHIP_RV610: chip_name = "RV610"; break;
 332         case CHIP_RV630: chip_name = "RV630"; break;
 333         case CHIP_RV620: chip_name = "RV620"; break;
 334         case CHIP_RV635: chip_name = "RV635"; break;
 335         case CHIP_RV670: chip_name = "RV670"; break;
 336         case CHIP_RS780:
 337         case CHIP_RS880: chip_name = "RS780"; break;
 338         case CHIP_RV770: chip_name = "RV770"; break;
 339         case CHIP_RV730:
 340         case CHIP_RV740: chip_name = "RV730"; break;
 341         case CHIP_RV710: chip_name = "RV710"; break;
 342         default:         panic("%s: Unsupported family %d", __func__, dev_priv->flags & RADEON_FAMILY_MASK);
 343         }
 344
 345         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770) {
 346                 pfp_req_size = R700_PFP_UCODE_SIZE * 4;
 347                 me_req_size = R700_PM4_UCODE_SIZE * 4;
 348         } else {
 349                 pfp_req_size = PFP_UCODE_SIZE * 4;
 350                 me_req_size = PM4_UCODE_SIZE * 12;
 351         }
 352
 353         DRM_INFO("Loading %s CP Microcode\n", chip_name);
 354
 355         ksnprintf(fw_name, sizeof(fw_name), "radeonkmsfw_%s_pfp", chip_name);
 356         err = request_firmware(&dev_priv->pfp_fw, fw_name, &pdev->dev);
 357         if (err)
 358                 goto out;
 359         if (dev_priv->pfp_fw->datasize != pfp_req_size) {
 360                 DRM_ERROR(
 361                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
 362                        dev_priv->pfp_fw->datasize, fw_name);
 363                 err = -EINVAL;
 364                 goto out;
 365         }
 366
 367         ksnprintf(fw_name, sizeof(fw_name), "radeonkmsfw_%s_me", chip_name);
 368         err = request_firmware(&dev_priv->me_fw, fw_name, &pdev->dev);
 369         if (err)
 370                 goto out;
 371         if (dev_priv->me_fw->datasize != me_req_size) {
 372                 DRM_ERROR(
 373                        "r600_cp: Bogus length %zu in firmware \"%s\"\n",
 374                        dev_priv->me_fw->datasize, fw_name);
 375                 err = -EINVAL;
 376         }
 377 out:
 378         if (err) {
 379                 if (err != -EINVAL)
 380                         DRM_ERROR(
 381                                "r600_cp: Failed to load firmware \"%s\"\n",
 382                                fw_name);
 383                 release_firmware(dev_priv->pfp_fw);
 384                 dev_priv->pfp_fw = NULL;
 385                 release_firmware(dev_priv->me_fw);
 386                 dev_priv->me_fw = NULL;
 387         }
 388         return err;
 389 }
 390
 391 static void r600_cp_load_microcode(drm_radeon_private_t *dev_priv)
 392 {
 393         const __be32 *fw_data;
 394         int i;
 395
 396         if (!dev_priv->me_fw || !dev_priv->pfp_fw)
 397                 return;
 398
 399         r600_do_cp_stop(dev_priv);
 400
 401         RADEON_WRITE(R600_CP_RB_CNTL,
 402 #ifdef __BIG_ENDIAN
 403                      R600_BUF_SWAP_32BIT |
 404 #endif
 405                      R600_RB_NO_UPDATE |
 406                      R600_RB_BLKSZ(15) |
 407                      R600_RB_BUFSZ(3));
 408
 409         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 410         RADEON_READ(R600_GRBM_SOFT_RESET);
 411         DRM_MDELAY(15);
 412         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 413
 414         fw_data = (const __be32 *)dev_priv->me_fw->data;
 415         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 416         for (i = 0; i < PM4_UCODE_SIZE * 3; i++)
 417                 RADEON_WRITE(R600_CP_ME_RAM_DATA,
 418                              be32_to_cpup(fw_data++));
 419
 420         fw_data = (const __be32 *)dev_priv->pfp_fw->data;
 421         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 422         for (i = 0; i < PFP_UCODE_SIZE; i++)
 423                 RADEON_WRITE(R600_CP_PFP_UCODE_DATA,
 424                              be32_to_cpup(fw_data++));
 425
 426         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 427         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 428         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 429
 430 }
 431
 432 static void r700_vm_init(struct drm_device *dev)
 433 {
 434         drm_radeon_private_t *dev_priv = dev->dev_private;
 435         /* initialise the VM to use the page table we constructed up there */
 436         u32 vm_c0, i;
 437         u32 mc_vm_md_l1;
 438         u32 vm_l2_cntl, vm_l2_cntl3;
 439         /* okay set up the PCIE aperture type thingo */
 440         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_LOW_ADDR, dev_priv->gart_vm_start >> 12);
 441         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_HIGH_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 442         RADEON_WRITE(R700_MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 0);
 443
 444         mc_vm_md_l1 = R700_ENABLE_L1_TLB |
 445             R700_ENABLE_L1_FRAGMENT_PROCESSING |
 446             R700_SYSTEM_ACCESS_MODE_IN_SYS |
 447             R700_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU |
 448             R700_EFFECTIVE_L1_TLB_SIZE(5) |
 449             R700_EFFECTIVE_L1_QUEUE_SIZE(5);
 450
 451         RADEON_WRITE(R700_MC_VM_MD_L1_TLB0_CNTL, mc_vm_md_l1);
 452         RADEON_WRITE(R700_MC_VM_MD_L1_TLB1_CNTL, mc_vm_md_l1);
 453         RADEON_WRITE(R700_MC_VM_MD_L1_TLB2_CNTL, mc_vm_md_l1);
 454         RADEON_WRITE(R700_MC_VM_MB_L1_TLB0_CNTL, mc_vm_md_l1);
 455         RADEON_WRITE(R700_MC_VM_MB_L1_TLB1_CNTL, mc_vm_md_l1);
 456         RADEON_WRITE(R700_MC_VM_MB_L1_TLB2_CNTL, mc_vm_md_l1);
 457         RADEON_WRITE(R700_MC_VM_MB_L1_TLB3_CNTL, mc_vm_md_l1);
 458
 459         vm_l2_cntl = R600_VM_L2_CACHE_EN | R600_VM_L2_FRAG_PROC | R600_VM_ENABLE_PTE_CACHE_LRU_W;
 460         vm_l2_cntl |= R700_VM_L2_CNTL_QUEUE_SIZE(7);
 461         RADEON_WRITE(R600_VM_L2_CNTL, vm_l2_cntl);
 462
 463         RADEON_WRITE(R600_VM_L2_CNTL2, 0);
 464         vm_l2_cntl3 = R700_VM_L2_CNTL3_BANK_SELECT(0) | R700_VM_L2_CNTL3_CACHE_UPDATE_MODE(2);
 465         RADEON_WRITE(R600_VM_L2_CNTL3, vm_l2_cntl3);
 466
 467         vm_c0 = R600_VM_ENABLE_CONTEXT | R600_VM_PAGE_TABLE_DEPTH_FLAT;
 468
 469         RADEON_WRITE(R600_VM_CONTEXT0_CNTL, vm_c0);
 470
 471         vm_c0 &= ~R600_VM_ENABLE_CONTEXT;
 472
 473         /* disable all other contexts */
 474         for (i = 1; i < 8; i++)
 475                 RADEON_WRITE(R600_VM_CONTEXT0_CNTL + (i * 4), vm_c0);
 476
 477         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, dev_priv->gart_info.bus_addr >> 12);
 478         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_START_ADDR, dev_priv->gart_vm_start >> 12);
 479         RADEON_WRITE(R700_VM_CONTEXT0_PAGE_TABLE_END_ADDR, (dev_priv->gart_vm_start + dev_priv->gart_size - 1) >> 12);
 480
 481         r600_vm_flush_gart_range(dev);
 482 }
 483
 484 static void r700_cp_load_microcode(drm_radeon_private_t *dev_priv)
 485 {
 486         const __be32 *fw_data;
 487         int i;
 488
 489         if (!dev_priv->me_fw || !dev_priv->pfp_fw)
 490                 return;
 491
 492         r600_do_cp_stop(dev_priv);
 493
 494         RADEON_WRITE(R600_CP_RB_CNTL,
 495 #ifdef __BIG_ENDIAN
 496                      R600_BUF_SWAP_32BIT |
 497 #endif
 498                      R600_RB_NO_UPDATE |
 499                      R600_RB_BLKSZ(15) |
 500                      R600_RB_BUFSZ(3));
 501
 502         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
 503         RADEON_READ(R600_GRBM_SOFT_RESET);
 504         DRM_MDELAY(15);
 505         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 506
 507         fw_data = (const __be32 *)dev_priv->pfp_fw->data;
 508         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 509         for (i = 0; i < R700_PFP_UCODE_SIZE; i++)
 510                 RADEON_WRITE(R600_CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
 511         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 512
 513         fw_data = (const __be32 *)dev_priv->me_fw->data;
 514         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 515         for (i = 0; i < R700_PM4_UCODE_SIZE; i++)
 516                 RADEON_WRITE(R600_CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
 517         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 518
 519         RADEON_WRITE(R600_CP_PFP_UCODE_ADDR, 0);
 520         RADEON_WRITE(R600_CP_ME_RAM_WADDR, 0);
 521         RADEON_WRITE(R600_CP_ME_RAM_RADDR, 0);
 522
 523 }
 524
 525 static void r600_test_writeback(drm_radeon_private_t *dev_priv)
 526 {
 527         u32 tmp;
 528
 529         /* Start with assuming that writeback doesn't work */
 530         dev_priv->writeback_works = 0;
 531
 532         /* Writeback doesn't seem to work everywhere, test it here and possibly
 533          * enable it if it appears to work
 534          */
 535         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
 536
 537         RADEON_WRITE(R600_SCRATCH_REG1, 0xdeadbeef);
 538
 539         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 540                 u32 val;
 541
 542                 val = radeon_read_ring_rptr(dev_priv, R600_SCRATCHOFF(1));
 543                 if (val == 0xdeadbeef)
 544                         break;
 545                 DRM_UDELAY(1);
 546         }
 547
 548         if (tmp < dev_priv->usec_timeout) {
 549                 dev_priv->writeback_works = 1;
 550                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 551         } else {
 552                 dev_priv->writeback_works = 0;
 553                 DRM_INFO("writeback test failed\n");
 554         }
 555         if (radeon_no_wb == 1) {
 556                 dev_priv->writeback_works = 0;
 557                 DRM_INFO("writeback forced off\n");
 558         }
 559
 560         if (!dev_priv->writeback_works) {
 561                 /* Disable writeback to avoid unnecessary bus master transfer */
 562                 RADEON_WRITE(R600_CP_RB_CNTL,
 563 #ifdef __BIG_ENDIAN
 564                              R600_BUF_SWAP_32BIT |
 565 #endif
 566                              RADEON_READ(R600_CP_RB_CNTL) |
 567                              R600_RB_NO_UPDATE);
 568                 RADEON_WRITE(R600_SCRATCH_UMSK, 0);
 569         }
 570 }
 571
 572 int r600_do_engine_reset(struct drm_device *dev)
 573 {
 574         drm_radeon_private_t *dev_priv = dev->dev_private;
 575         u32 cp_ptr, cp_me_cntl, cp_rb_cntl;
 576
 577         DRM_INFO("Resetting GPU\n");
 578
 579         cp_ptr = RADEON_READ(R600_CP_RB_WPTR);
 580         cp_me_cntl = RADEON_READ(R600_CP_ME_CNTL);
 581         RADEON_WRITE(R600_CP_ME_CNTL, R600_CP_ME_HALT);
 582
 583         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0x7fff);
 584         RADEON_READ(R600_GRBM_SOFT_RESET);
 585         DRM_UDELAY(50);
 586         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
 587         RADEON_READ(R600_GRBM_SOFT_RESET);
 588
 589         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
 590         cp_rb_cntl = RADEON_READ(R600_CP_RB_CNTL);
 591         RADEON_WRITE(R600_CP_RB_CNTL,
 592 #ifdef __BIG_ENDIAN
 593                      R600_BUF_SWAP_32BIT |
 594 #endif
 595                      R600_RB_RPTR_WR_ENA);
 596
 597         RADEON_WRITE(R600_CP_RB_RPTR_WR, cp_ptr);
 598         RADEON_WRITE(R600_CP_RB_WPTR, cp_ptr);
 599         RADEON_WRITE(R600_CP_RB_CNTL, cp_rb_cntl);
 600         RADEON_WRITE(R600_CP_ME_CNTL, cp_me_cntl);
 601
 602         /* Reset the CP ring */
 603         r600_do_cp_reset(dev_priv);
 604
 605         /* The CP is no longer running after an engine reset */
 606         dev_priv->cp_running = 0;
 607
 608         /* Reset any pending vertex, indirect buffers */
 609         radeon_freelist_reset(dev);
 610
 611         return 0;
 612
 613 }
 614
 615 static u32 r600_get_tile_pipe_to_backend_map(u32 num_tile_pipes,
 616                                              u32 num_backends,
 617                                              u32 backend_disable_mask)
 618 {
 619         u32 backend_map = 0;
 620         u32 enabled_backends_mask;
 621         u32 enabled_backends_count;
 622         u32 cur_pipe;
 623         u32 swizzle_pipe[R6XX_MAX_PIPES];
 624         u32 cur_backend;
 625         u32 i;
 626
 627         if (num_tile_pipes > R6XX_MAX_PIPES)
 628                 num_tile_pipes = R6XX_MAX_PIPES;
 629         if (num_tile_pipes < 1)
 630                 num_tile_pipes = 1;
 631         if (num_backends > R6XX_MAX_BACKENDS)
 632                 num_backends = R6XX_MAX_BACKENDS;
 633         if (num_backends < 1)
 634                 num_backends = 1;
 635
 636         enabled_backends_mask = 0;
 637         enabled_backends_count = 0;
 638         for (i = 0; i < R6XX_MAX_BACKENDS; ++i) {
 639                 if (((backend_disable_mask >> i) & 1) == 0) {
 640                         enabled_backends_mask |= (1 << i);
 641                         ++enabled_backends_count;
 642                 }
 643                 if (enabled_backends_count == num_backends)
 644                         break;
 645         }
 646
 647         if (enabled_backends_count == 0) {
 648                 enabled_backends_mask = 1;
 649                 enabled_backends_count = 1;
 650         }
 651
 652         if (enabled_backends_count != num_backends)
 653                 num_backends = enabled_backends_count;
 654
 655         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R6XX_MAX_PIPES);
 656         switch (num_tile_pipes) {
 657         case 1:
 658                 swizzle_pipe[0] = 0;
 659                 break;
 660         case 2:
 661                 swizzle_pipe[0] = 0;
 662                 swizzle_pipe[1] = 1;
 663                 break;
 664         case 3:
 665                 swizzle_pipe[0] = 0;
 666                 swizzle_pipe[1] = 1;
 667                 swizzle_pipe[2] = 2;
 668                 break;
 669         case 4:
 670                 swizzle_pipe[0] = 0;
 671                 swizzle_pipe[1] = 1;
 672                 swizzle_pipe[2] = 2;
 673                 swizzle_pipe[3] = 3;
 674                 break;
 675         case 5:
 676                 swizzle_pipe[0] = 0;
 677                 swizzle_pipe[1] = 1;
 678                 swizzle_pipe[2] = 2;
 679                 swizzle_pipe[3] = 3;
 680                 swizzle_pipe[4] = 4;
 681                 break;
 682         case 6:
 683                 swizzle_pipe[0] = 0;
 684                 swizzle_pipe[1] = 2;
 685                 swizzle_pipe[2] = 4;
 686                 swizzle_pipe[3] = 5;
 687                 swizzle_pipe[4] = 1;
 688                 swizzle_pipe[5] = 3;
 689                 break;
 690         case 7:
 691                 swizzle_pipe[0] = 0;
 692                 swizzle_pipe[1] = 2;
 693                 swizzle_pipe[2] = 4;
 694                 swizzle_pipe[3] = 6;
 695                 swizzle_pipe[4] = 1;
 696                 swizzle_pipe[5] = 3;
 697                 swizzle_pipe[6] = 5;
 698                 break;
 699         case 8:
 700                 swizzle_pipe[0] = 0;
 701                 swizzle_pipe[1] = 2;
 702                 swizzle_pipe[2] = 4;
 703                 swizzle_pipe[3] = 6;
 704                 swizzle_pipe[4] = 1;
 705                 swizzle_pipe[5] = 3;
 706                 swizzle_pipe[6] = 5;
 707                 swizzle_pipe[7] = 7;
 708                 break;
 709         }
 710
 711         cur_backend = 0;
 712         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
 713                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
 714                         cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 715
 716                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
 717
 718                 cur_backend = (cur_backend + 1) % R6XX_MAX_BACKENDS;
 719         }
 720
 721         return backend_map;
 722 }
 723
 724 static int r600_count_pipe_bits(uint32_t val)
 725 {
 726         return hweight32(val);
 727 }
 728
 729 static void r600_gfx_init(struct drm_device *dev,
 730                           drm_radeon_private_t *dev_priv)
 731 {
 732         int i, j, num_qd_pipes;
 733         u32 sx_debug_1;
 734         u32 tc_cntl;
 735         u32 arb_pop;
 736         u32 num_gs_verts_per_thread;
 737         u32 vgt_gs_per_es;
 738         u32 gs_prim_buffer_depth = 0;
 739         u32 sq_ms_fifo_sizes;
 740         u32 sq_config;
 741         u32 sq_gpr_resource_mgmt_1 = 0;
 742         u32 sq_gpr_resource_mgmt_2 = 0;
 743         u32 sq_thread_resource_mgmt = 0;
 744         u32 sq_stack_resource_mgmt_1 = 0;
 745         u32 sq_stack_resource_mgmt_2 = 0;
 746         u32 hdp_host_path_cntl;
 747         u32 backend_map;
 748         u32 gb_tiling_config = 0;
 749         u32 cc_rb_backend_disable;
 750         u32 cc_gc_shader_pipe_config;
 751         u32 ramcfg;
 752
 753         /* setup chip specs */
 754         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
 755         case CHIP_R600:
 756                 dev_priv->r600_max_pipes = 4;
 757                 dev_priv->r600_max_tile_pipes = 8;
 758                 dev_priv->r600_max_simds = 4;
 759                 dev_priv->r600_max_backends = 4;
 760                 dev_priv->r600_max_gprs = 256;
 761                 dev_priv->r600_max_threads = 192;
 762                 dev_priv->r600_max_stack_entries = 256;
 763                 dev_priv->r600_max_hw_contexts = 8;
 764                 dev_priv->r600_max_gs_threads = 16;
 765                 dev_priv->r600_sx_max_export_size = 128;
 766                 dev_priv->r600_sx_max_export_pos_size = 16;
 767                 dev_priv->r600_sx_max_export_smx_size = 128;
 768                 dev_priv->r600_sq_num_cf_insts = 2;
 769                 break;
 770         case CHIP_RV630:
 771         case CHIP_RV635:
 772                 dev_priv->r600_max_pipes = 2;
 773                 dev_priv->r600_max_tile_pipes = 2;
 774                 dev_priv->r600_max_simds = 3;
 775                 dev_priv->r600_max_backends = 1;
 776                 dev_priv->r600_max_gprs = 128;
 777                 dev_priv->r600_max_threads = 192;
 778                 dev_priv->r600_max_stack_entries = 128;
 779                 dev_priv->r600_max_hw_contexts = 8;
 780                 dev_priv->r600_max_gs_threads = 4;
 781                 dev_priv->r600_sx_max_export_size = 128;
 782                 dev_priv->r600_sx_max_export_pos_size = 16;
 783                 dev_priv->r600_sx_max_export_smx_size = 128;
 784                 dev_priv->r600_sq_num_cf_insts = 2;
 785                 break;
 786         case CHIP_RV610:
 787         case CHIP_RS780:
 788         case CHIP_RS880:
 789         case CHIP_RV620:
 790                 dev_priv->r600_max_pipes = 1;
 791                 dev_priv->r600_max_tile_pipes = 1;
 792                 dev_priv->r600_max_simds = 2;
 793                 dev_priv->r600_max_backends = 1;
 794                 dev_priv->r600_max_gprs = 128;
 795                 dev_priv->r600_max_threads = 192;
 796                 dev_priv->r600_max_stack_entries = 128;
 797                 dev_priv->r600_max_hw_contexts = 4;
 798                 dev_priv->r600_max_gs_threads = 4;
 799                 dev_priv->r600_sx_max_export_size = 128;
 800                 dev_priv->r600_sx_max_export_pos_size = 16;
 801                 dev_priv->r600_sx_max_export_smx_size = 128;
 802                 dev_priv->r600_sq_num_cf_insts = 1;
 803                 break;
 804         case CHIP_RV670:
 805                 dev_priv->r600_max_pipes = 4;
 806                 dev_priv->r600_max_tile_pipes = 4;
 807                 dev_priv->r600_max_simds = 4;
 808                 dev_priv->r600_max_backends = 4;
 809                 dev_priv->r600_max_gprs = 192;
 810                 dev_priv->r600_max_threads = 192;
 811                 dev_priv->r600_max_stack_entries = 256;
 812                 dev_priv->r600_max_hw_contexts = 8;
 813                 dev_priv->r600_max_gs_threads = 16;
 814                 dev_priv->r600_sx_max_export_size = 128;
 815                 dev_priv->r600_sx_max_export_pos_size = 16;
 816                 dev_priv->r600_sx_max_export_smx_size = 128;
 817                 dev_priv->r600_sq_num_cf_insts = 2;
 818                 break;
 819         default:
 820                 break;
 821         }
 822
 823         /* Initialize HDP */
 824         j = 0;
 825         for (i = 0; i < 32; i++) {
 826                 RADEON_WRITE((0x2c14 + j), 0x00000000);
 827                 RADEON_WRITE((0x2c18 + j), 0x00000000);
 828                 RADEON_WRITE((0x2c1c + j), 0x00000000);
 829                 RADEON_WRITE((0x2c20 + j), 0x00000000);
 830                 RADEON_WRITE((0x2c24 + j), 0x00000000);
 831                 j += 0x18;
 832         }
 833
 834         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
 835
 836         /* setup tiling, simd, pipe config */
 837         ramcfg = RADEON_READ(R600_RAMCFG);
 838
 839         switch (dev_priv->r600_max_tile_pipes) {
 840         case 1:
 841                 gb_tiling_config |= R600_PIPE_TILING(0);
 842                 break;
 843         case 2:
 844                 gb_tiling_config |= R600_PIPE_TILING(1);
 845                 break;
 846         case 4:
 847                 gb_tiling_config |= R600_PIPE_TILING(2);
 848                 break;
 849         case 8:
 850                 gb_tiling_config |= R600_PIPE_TILING(3);
 851                 break;
 852         default:
 853                 break;
 854         }
 855
 856         gb_tiling_config |= R600_BANK_TILING((ramcfg >> R600_NOOFBANK_SHIFT) & R600_NOOFBANK_MASK);
 857
 858         gb_tiling_config |= R600_GROUP_SIZE(0);
 859
 860         if (((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK) > 3) {
 861                 gb_tiling_config |= R600_ROW_TILING(3);
 862                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
 863         } else {
 864                 gb_tiling_config |=
 865                         R600_ROW_TILING(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 866                 gb_tiling_config |=
 867                         R600_SAMPLE_SPLIT(((ramcfg >> R600_NOOFROWS_SHIFT) & R600_NOOFROWS_MASK));
 868         }
 869
 870         gb_tiling_config |= R600_BANK_SWAPS(1);
 871
 872         cc_rb_backend_disable = RADEON_READ(R600_CC_RB_BACKEND_DISABLE) & 0x00ff0000;
 873         cc_rb_backend_disable |=
 874                 R600_BACKEND_DISABLE((R6XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R6XX_MAX_BACKENDS_MASK);
 875
 876         cc_gc_shader_pipe_config = RADEON_READ(R600_CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
 877         cc_gc_shader_pipe_config |=
 878                 R600_INACTIVE_QD_PIPES((R6XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R6XX_MAX_PIPES_MASK);
 879         cc_gc_shader_pipe_config |=
 880                 R600_INACTIVE_SIMDS((R6XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R6XX_MAX_SIMDS_MASK);
 881
 882         backend_map = r600_get_tile_pipe_to_backend_map(dev_priv->r600_max_tile_pipes,
 883                                                         (R6XX_MAX_BACKENDS -
 884                                                          r600_count_pipe_bits((cc_rb_backend_disable &
 885                                                                                R6XX_MAX_BACKENDS_MASK) >> 16)),
 886                                                         (cc_rb_backend_disable >> 16));
 887         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
 888
 889         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
 890         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 891         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
 892         if (gb_tiling_config & 0xc0) {
 893                 dev_priv->r600_group_size = 512;
 894         } else {
 895                 dev_priv->r600_group_size = 256;
 896         }
 897         dev_priv->r600_npipes = 1 << ((gb_tiling_config >> 1) & 0x7);
 898         if (gb_tiling_config & 0x30) {
 899                 dev_priv->r600_nbanks = 8;
 900         } else {
 901                 dev_priv->r600_nbanks = 4;
 902         }
 903
 904         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
 905         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
 906         RADEON_WRITE(R600_GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
 907
 908         num_qd_pipes =
 909                 R6XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK) >> 8);
 910         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
 911         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
 912
 913         /* set HW defaults for 3D engine */
 914         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
 915                                                 R600_ROQ_IB2_START(0x2b)));
 916
 917         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, (R600_MEQ_END(0x40) |
 918                                               R600_ROQ_END(0x40)));
 919
 920         RADEON_WRITE(R600_TA_CNTL_AUX, (R600_DISABLE_CUBE_ANISO |
 921                                         R600_SYNC_GRADIENT |
 922                                         R600_SYNC_WALKER |
 923                                         R600_SYNC_ALIGNER));
 924
 925         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670)
 926                 RADEON_WRITE(R600_ARB_GDEC_RD_CNTL, 0x00000021);
 927
 928         sx_debug_1 = RADEON_READ(R600_SX_DEBUG_1);
 929         sx_debug_1 |= R600_SMX_EVENT_RELEASE;
 930         if (((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_R600))
 931                 sx_debug_1 |= R600_ENABLE_NEW_SMX_ADDRESS;
 932         RADEON_WRITE(R600_SX_DEBUG_1, sx_debug_1);
 933
 934         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 935             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
 936             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 937             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 938             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 939             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
 940                 RADEON_WRITE(R600_DB_DEBUG, R600_PREZ_MUST_WAIT_FOR_POSTZ_DONE);
 941         else
 942                 RADEON_WRITE(R600_DB_DEBUG, 0);
 943
 944         RADEON_WRITE(R600_DB_WATERMARKS, (R600_DEPTH_FREE(4) |
 945                                           R600_DEPTH_FLUSH(16) |
 946                                           R600_DEPTH_PENDING_FREE(4) |
 947                                           R600_DEPTH_CACHELINE_FREE(16)));
 948         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
 949         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 0);
 950
 951         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
 952         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(0));
 953
 954         sq_ms_fifo_sizes = RADEON_READ(R600_SQ_MS_FIFO_SIZES);
 955         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
 956             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
 957             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
 958             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
 959                 sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(0xa) |
 960                                     R600_FETCH_FIFO_HIWATER(0xa) |
 961                                     R600_DONE_FIFO_HIWATER(0xe0) |
 962                                     R600_ALU_UPDATE_FIFO_HIWATER(0x8));
 963         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) ||
 964                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630)) {
 965                 sq_ms_fifo_sizes &= ~R600_DONE_FIFO_HIWATER(0xff);
 966                 sq_ms_fifo_sizes |= R600_DONE_FIFO_HIWATER(0x4);
 967         }
 968         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
 969
 970         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
 971          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
 972          */
 973         sq_config = RADEON_READ(R600_SQ_CONFIG);
 974         sq_config &= ~(R600_PS_PRIO(3) |
 975                        R600_VS_PRIO(3) |
 976                        R600_GS_PRIO(3) |
 977                        R600_ES_PRIO(3));
 978         sq_config |= (R600_DX9_CONSTS |
 979                       R600_VC_ENABLE |
 980                       R600_PS_PRIO(0) |
 981                       R600_VS_PRIO(1) |
 982                       R600_GS_PRIO(2) |
 983                       R600_ES_PRIO(3));
 984
 985         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R600) {
 986                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(124) |
 987                                           R600_NUM_VS_GPRS(124) |
 988                                           R600_NUM_CLAUSE_TEMP_GPRS(4));
 989                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(0) |
 990                                           R600_NUM_ES_GPRS(0));
 991                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(136) |
 992                                            R600_NUM_VS_THREADS(48) |
 993                                            R600_NUM_GS_THREADS(4) |
 994                                            R600_NUM_ES_THREADS(4));
 995                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(128) |
 996                                             R600_NUM_VS_STACK_ENTRIES(128));
 997                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(0) |
 998                                             R600_NUM_ES_STACK_ENTRIES(0));
 999         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
1000                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
1001                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
1002                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880)) {
1003                 /* no vertex cache */
1004                 sq_config &= ~R600_VC_ENABLE;
1005
1006                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1007                                           R600_NUM_VS_GPRS(44) |
1008                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1009                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
1010                                           R600_NUM_ES_GPRS(17));
1011                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1012                                            R600_NUM_VS_THREADS(78) |
1013                                            R600_NUM_GS_THREADS(4) |
1014                                            R600_NUM_ES_THREADS(31));
1015                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1016                                             R600_NUM_VS_STACK_ENTRIES(40));
1017                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1018                                             R600_NUM_ES_STACK_ENTRIES(16));
1019         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV630) ||
1020                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV635)) {
1021                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1022                                           R600_NUM_VS_GPRS(44) |
1023                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1024                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(18) |
1025                                           R600_NUM_ES_GPRS(18));
1026                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1027                                            R600_NUM_VS_THREADS(78) |
1028                                            R600_NUM_GS_THREADS(4) |
1029                                            R600_NUM_ES_THREADS(31));
1030                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(40) |
1031                                             R600_NUM_VS_STACK_ENTRIES(40));
1032                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(32) |
1033                                             R600_NUM_ES_STACK_ENTRIES(16));
1034         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV670) {
1035                 sq_gpr_resource_mgmt_1 = (R600_NUM_PS_GPRS(44) |
1036                                           R600_NUM_VS_GPRS(44) |
1037                                           R600_NUM_CLAUSE_TEMP_GPRS(2));
1038                 sq_gpr_resource_mgmt_2 = (R600_NUM_GS_GPRS(17) |
1039                                           R600_NUM_ES_GPRS(17));
1040                 sq_thread_resource_mgmt = (R600_NUM_PS_THREADS(79) |
1041                                            R600_NUM_VS_THREADS(78) |
1042                                            R600_NUM_GS_THREADS(4) |
1043                                            R600_NUM_ES_THREADS(31));
1044                 sq_stack_resource_mgmt_1 = (R600_NUM_PS_STACK_ENTRIES(64) |
1045                                             R600_NUM_VS_STACK_ENTRIES(64));
1046                 sq_stack_resource_mgmt_2 = (R600_NUM_GS_STACK_ENTRIES(64) |
1047                                             R600_NUM_ES_STACK_ENTRIES(64));
1048         }
1049
1050         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1051         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  sq_gpr_resource_mgmt_1);
1052         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  sq_gpr_resource_mgmt_2);
1053         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1054         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, sq_stack_resource_mgmt_1);
1055         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, sq_stack_resource_mgmt_2);
1056
1057         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV610) ||
1058             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV620) ||
1059             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS780) ||
1060             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS880))
1061                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_TC_ONLY));
1062         else
1063                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, R600_CACHE_INVALIDATION(R600_VC_AND_TC));
1064
1065         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_2S, (R600_S0_X(0xc) |
1066                                                     R600_S0_Y(0x4) |
1067                                                     R600_S1_X(0x4) |
1068                                                     R600_S1_Y(0xc)));
1069         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_4S, (R600_S0_X(0xe) |
1070                                                     R600_S0_Y(0xe) |
1071                                                     R600_S1_X(0x2) |
1072                                                     R600_S1_Y(0x2) |
1073                                                     R600_S2_X(0xa) |
1074                                                     R600_S2_Y(0x6) |
1075                                                     R600_S3_X(0x6) |
1076                                                     R600_S3_Y(0xa)));
1077         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD0, (R600_S0_X(0xe) |
1078                                                         R600_S0_Y(0xb) |
1079                                                         R600_S1_X(0x4) |
1080                                                         R600_S1_Y(0xc) |
1081                                                         R600_S2_X(0x1) |
1082                                                         R600_S2_Y(0x6) |
1083                                                         R600_S3_X(0xa) |
1084                                                         R600_S3_Y(0xe)));
1085         RADEON_WRITE(R600_PA_SC_AA_SAMPLE_LOCS_8S_WD1, (R600_S4_X(0x6) |
1086                                                         R600_S4_Y(0x1) |
1087                                                         R600_S5_X(0x0) |
1088                                                         R600_S5_Y(0x0) |
1089                                                         R600_S6_X(0xb) |
1090                                                         R600_S6_Y(0x4) |
1091                                                         R600_S7_X(0x7) |
1092                                                         R600_S7_Y(0x8)));
1093
1094
1095         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1096         case CHIP_R600:
1097         case CHIP_RV630:
1098         case CHIP_RV635:
1099                 gs_prim_buffer_depth = 0;
1100                 break;
1101         case CHIP_RV610:
1102         case CHIP_RS780:
1103         case CHIP_RS880:
1104         case CHIP_RV620:
1105                 gs_prim_buffer_depth = 32;
1106                 break;
1107         case CHIP_RV670:
1108                 gs_prim_buffer_depth = 128;
1109                 break;
1110         default:
1111                 break;
1112         }
1113
1114         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1115         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1116         /* Max value for this is 256 */
1117         if (vgt_gs_per_es > 256)
1118                 vgt_gs_per_es = 256;
1119
1120         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1121         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1122         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1123         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1124
1125         /* more default values. 2D/3D driver should adjust as needed */
1126         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1127         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1128         RADEON_WRITE(R600_SX_MISC, 0);
1129         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1130         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1131         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1132         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1133         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1134         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1135
1136         /* clear render buffer base addresses */
1137         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1138         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1139         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1140         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1141         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1142         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1143         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1144         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1145
1146         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1147         case CHIP_RV610:
1148         case CHIP_RS780:
1149         case CHIP_RS880:
1150         case CHIP_RV620:
1151                 tc_cntl = R600_TC_L2_SIZE(8);
1152                 break;
1153         case CHIP_RV630:
1154         case CHIP_RV635:
1155                 tc_cntl = R600_TC_L2_SIZE(4);
1156                 break;
1157         case CHIP_R600:
1158                 tc_cntl = R600_TC_L2_SIZE(0) | R600_L2_DISABLE_LATE_HIT;
1159                 break;
1160         default:
1161                 tc_cntl = R600_TC_L2_SIZE(0);
1162                 break;
1163         }
1164
1165         RADEON_WRITE(R600_TC_CNTL, tc_cntl);
1166
1167         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1168         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1169
1170         arb_pop = RADEON_READ(R600_ARB_POP);
1171         arb_pop |= R600_ENABLE_TC128;
1172         RADEON_WRITE(R600_ARB_POP, arb_pop);
1173
1174         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1175         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1176                                           R600_NUM_CLIP_SEQ(3)));
1177         RADEON_WRITE(R600_PA_SC_ENHANCE, R600_FORCE_EOV_MAX_CLK_CNT(4095));
1178
1179 }
1180
1181 static u32 r700_get_tile_pipe_to_backend_map(drm_radeon_private_t *dev_priv,
1182                                              u32 num_tile_pipes,
1183                                              u32 num_backends,
1184                                              u32 backend_disable_mask)
1185 {
1186         u32 backend_map = 0;
1187         u32 enabled_backends_mask;
1188         u32 enabled_backends_count;
1189         u32 cur_pipe;
1190         u32 swizzle_pipe[R7XX_MAX_PIPES];
1191         u32 cur_backend;
1192         u32 i;
1193         bool force_no_swizzle;
1194
1195         if (num_tile_pipes > R7XX_MAX_PIPES)
1196                 num_tile_pipes = R7XX_MAX_PIPES;
1197         if (num_tile_pipes < 1)
1198                 num_tile_pipes = 1;
1199         if (num_backends > R7XX_MAX_BACKENDS)
1200                 num_backends = R7XX_MAX_BACKENDS;
1201         if (num_backends < 1)
1202                 num_backends = 1;
1203
1204         enabled_backends_mask = 0;
1205         enabled_backends_count = 0;
1206         for (i = 0; i < R7XX_MAX_BACKENDS; ++i) {
1207                 if (((backend_disable_mask >> i) & 1) == 0) {
1208                         enabled_backends_mask |= (1 << i);
1209                         ++enabled_backends_count;
1210                 }
1211                 if (enabled_backends_count == num_backends)
1212                         break;
1213         }
1214
1215         if (enabled_backends_count == 0) {
1216                 enabled_backends_mask = 1;
1217                 enabled_backends_count = 1;
1218         }
1219
1220         if (enabled_backends_count != num_backends)
1221                 num_backends = enabled_backends_count;
1222
1223         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1224         case CHIP_RV770:
1225         case CHIP_RV730:
1226                 force_no_swizzle = false;
1227                 break;
1228         case CHIP_RV710:
1229         case CHIP_RV740:
1230         default:
1231                 force_no_swizzle = true;
1232                 break;
1233         }
1234
1235         memset((uint8_t *)&swizzle_pipe[0], 0, sizeof(u32) * R7XX_MAX_PIPES);
1236         switch (num_tile_pipes) {
1237         case 1:
1238                 swizzle_pipe[0] = 0;
1239                 break;
1240         case 2:
1241                 swizzle_pipe[0] = 0;
1242                 swizzle_pipe[1] = 1;
1243                 break;
1244         case 3:
1245                 if (force_no_swizzle) {
1246                         swizzle_pipe[0] = 0;
1247                         swizzle_pipe[1] = 1;
1248                         swizzle_pipe[2] = 2;
1249                 } else {
1250                         swizzle_pipe[0] = 0;
1251                         swizzle_pipe[1] = 2;
1252                         swizzle_pipe[2] = 1;
1253                 }
1254                 break;
1255         case 4:
1256                 if (force_no_swizzle) {
1257                         swizzle_pipe[0] = 0;
1258                         swizzle_pipe[1] = 1;
1259                         swizzle_pipe[2] = 2;
1260                         swizzle_pipe[3] = 3;
1261                 } else {
1262                         swizzle_pipe[0] = 0;
1263                         swizzle_pipe[1] = 2;
1264                         swizzle_pipe[2] = 3;
1265                         swizzle_pipe[3] = 1;
1266                 }
1267                 break;
1268         case 5:
1269                 if (force_no_swizzle) {
1270                         swizzle_pipe[0] = 0;
1271                         swizzle_pipe[1] = 1;
1272                         swizzle_pipe[2] = 2;
1273                         swizzle_pipe[3] = 3;
1274                         swizzle_pipe[4] = 4;
1275                 } else {
1276                         swizzle_pipe[0] = 0;
1277                         swizzle_pipe[1] = 2;
1278                         swizzle_pipe[2] = 4;
1279                         swizzle_pipe[3] = 1;
1280                         swizzle_pipe[4] = 3;
1281                 }
1282                 break;
1283         case 6:
1284                 if (force_no_swizzle) {
1285                         swizzle_pipe[0] = 0;
1286                         swizzle_pipe[1] = 1;
1287                         swizzle_pipe[2] = 2;
1288                         swizzle_pipe[3] = 3;
1289                         swizzle_pipe[4] = 4;
1290                         swizzle_pipe[5] = 5;
1291                 } else {
1292                         swizzle_pipe[0] = 0;
1293                         swizzle_pipe[1] = 2;
1294                         swizzle_pipe[2] = 4;
1295                         swizzle_pipe[3] = 5;
1296                         swizzle_pipe[4] = 3;
1297                         swizzle_pipe[5] = 1;
1298                 }
1299                 break;
1300         case 7:
1301                 if (force_no_swizzle) {
1302                         swizzle_pipe[0] = 0;
1303                         swizzle_pipe[1] = 1;
1304                         swizzle_pipe[2] = 2;
1305                         swizzle_pipe[3] = 3;
1306                         swizzle_pipe[4] = 4;
1307                         swizzle_pipe[5] = 5;
1308                         swizzle_pipe[6] = 6;
1309                 } else {
1310                         swizzle_pipe[0] = 0;
1311                         swizzle_pipe[1] = 2;
1312                         swizzle_pipe[2] = 4;
1313                         swizzle_pipe[3] = 6;
1314                         swizzle_pipe[4] = 3;
1315                         swizzle_pipe[5] = 1;
1316                         swizzle_pipe[6] = 5;
1317                 }
1318                 break;
1319         case 8:
1320                 if (force_no_swizzle) {
1321                         swizzle_pipe[0] = 0;
1322                         swizzle_pipe[1] = 1;
1323                         swizzle_pipe[2] = 2;
1324                         swizzle_pipe[3] = 3;
1325                         swizzle_pipe[4] = 4;
1326                         swizzle_pipe[5] = 5;
1327                         swizzle_pipe[6] = 6;
1328                         swizzle_pipe[7] = 7;
1329                 } else {
1330                         swizzle_pipe[0] = 0;
1331                         swizzle_pipe[1] = 2;
1332                         swizzle_pipe[2] = 4;
1333                         swizzle_pipe[3] = 6;
1334                         swizzle_pipe[4] = 3;
1335                         swizzle_pipe[5] = 1;
1336                         swizzle_pipe[6] = 7;
1337                         swizzle_pipe[7] = 5;
1338                 }
1339                 break;
1340         }
1341
1342         cur_backend = 0;
1343         for (cur_pipe = 0; cur_pipe < num_tile_pipes; ++cur_pipe) {
1344                 while (((1 << cur_backend) & enabled_backends_mask) == 0)
1345                         cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1346
1347                 backend_map |= (u32)(((cur_backend & 3) << (swizzle_pipe[cur_pipe] * 2)));
1348
1349                 cur_backend = (cur_backend + 1) % R7XX_MAX_BACKENDS;
1350         }
1351
1352         return backend_map;
1353 }
1354
1355 static void r700_gfx_init(struct drm_device *dev,
1356                           drm_radeon_private_t *dev_priv)
1357 {
1358         int i, j, num_qd_pipes;
1359         u32 ta_aux_cntl;
1360         u32 sx_debug_1;
1361         u32 smx_dc_ctl0;
1362         u32 db_debug3;
1363         u32 num_gs_verts_per_thread;
1364         u32 vgt_gs_per_es;
1365         u32 gs_prim_buffer_depth = 0;
1366         u32 sq_ms_fifo_sizes;
1367         u32 sq_config;
1368         u32 sq_thread_resource_mgmt;
1369         u32 hdp_host_path_cntl;
1370         u32 sq_dyn_gpr_size_simd_ab_0;
1371         u32 backend_map;
1372         u32 gb_tiling_config = 0;
1373         u32 cc_rb_backend_disable;
1374         u32 cc_gc_shader_pipe_config;
1375         u32 mc_arb_ramcfg;
1376         u32 db_debug4;
1377
1378         /* setup chip specs */
1379         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1380         case CHIP_RV770:
1381                 dev_priv->r600_max_pipes = 4;
1382                 dev_priv->r600_max_tile_pipes = 8;
1383                 dev_priv->r600_max_simds = 10;
1384                 dev_priv->r600_max_backends = 4;
1385                 dev_priv->r600_max_gprs = 256;
1386                 dev_priv->r600_max_threads = 248;
1387                 dev_priv->r600_max_stack_entries = 512;
1388                 dev_priv->r600_max_hw_contexts = 8;
1389                 dev_priv->r600_max_gs_threads = 16 * 2;
1390                 dev_priv->r600_sx_max_export_size = 128;
1391                 dev_priv->r600_sx_max_export_pos_size = 16;
1392                 dev_priv->r600_sx_max_export_smx_size = 112;
1393                 dev_priv->r600_sq_num_cf_insts = 2;
1394
1395                 dev_priv->r700_sx_num_of_sets = 7;
1396                 dev_priv->r700_sc_prim_fifo_size = 0xF9;
1397                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1398                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1399                 break;
1400         case CHIP_RV730:
1401                 dev_priv->r600_max_pipes = 2;
1402                 dev_priv->r600_max_tile_pipes = 4;
1403                 dev_priv->r600_max_simds = 8;
1404                 dev_priv->r600_max_backends = 2;
1405                 dev_priv->r600_max_gprs = 128;
1406                 dev_priv->r600_max_threads = 248;
1407                 dev_priv->r600_max_stack_entries = 256;
1408                 dev_priv->r600_max_hw_contexts = 8;
1409                 dev_priv->r600_max_gs_threads = 16 * 2;
1410                 dev_priv->r600_sx_max_export_size = 256;
1411                 dev_priv->r600_sx_max_export_pos_size = 32;
1412                 dev_priv->r600_sx_max_export_smx_size = 224;
1413                 dev_priv->r600_sq_num_cf_insts = 2;
1414
1415                 dev_priv->r700_sx_num_of_sets = 7;
1416                 dev_priv->r700_sc_prim_fifo_size = 0xf9;
1417                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1418                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1419                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1420                         dev_priv->r600_sx_max_export_pos_size -= 16;
1421                         dev_priv->r600_sx_max_export_smx_size += 16;
1422                 }
1423                 break;
1424         case CHIP_RV710:
1425                 dev_priv->r600_max_pipes = 2;
1426                 dev_priv->r600_max_tile_pipes = 2;
1427                 dev_priv->r600_max_simds = 2;
1428                 dev_priv->r600_max_backends = 1;
1429                 dev_priv->r600_max_gprs = 256;
1430                 dev_priv->r600_max_threads = 192;
1431                 dev_priv->r600_max_stack_entries = 256;
1432                 dev_priv->r600_max_hw_contexts = 4;
1433                 dev_priv->r600_max_gs_threads = 8 * 2;
1434                 dev_priv->r600_sx_max_export_size = 128;
1435                 dev_priv->r600_sx_max_export_pos_size = 16;
1436                 dev_priv->r600_sx_max_export_smx_size = 112;
1437                 dev_priv->r600_sq_num_cf_insts = 1;
1438
1439                 dev_priv->r700_sx_num_of_sets = 7;
1440                 dev_priv->r700_sc_prim_fifo_size = 0x40;
1441                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1442                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1443                 break;
1444         case CHIP_RV740:
1445                 dev_priv->r600_max_pipes = 4;
1446                 dev_priv->r600_max_tile_pipes = 4;
1447                 dev_priv->r600_max_simds = 8;
1448                 dev_priv->r600_max_backends = 4;
1449                 dev_priv->r600_max_gprs = 256;
1450                 dev_priv->r600_max_threads = 248;
1451                 dev_priv->r600_max_stack_entries = 512;
1452                 dev_priv->r600_max_hw_contexts = 8;
1453                 dev_priv->r600_max_gs_threads = 16 * 2;
1454                 dev_priv->r600_sx_max_export_size = 256;
1455                 dev_priv->r600_sx_max_export_pos_size = 32;
1456                 dev_priv->r600_sx_max_export_smx_size = 224;
1457                 dev_priv->r600_sq_num_cf_insts = 2;
1458
1459                 dev_priv->r700_sx_num_of_sets = 7;
1460                 dev_priv->r700_sc_prim_fifo_size = 0x100;
1461                 dev_priv->r700_sc_hiz_tile_fifo_size = 0x30;
1462                 dev_priv->r700_sc_earlyz_tile_fifo_fize = 0x130;
1463
1464                 if (dev_priv->r600_sx_max_export_pos_size > 16) {
1465                         dev_priv->r600_sx_max_export_pos_size -= 16;
1466                         dev_priv->r600_sx_max_export_smx_size += 16;
1467                 }
1468                 break;
1469         default:
1470                 break;
1471         }
1472
1473         /* Initialize HDP */
1474         j = 0;
1475         for (i = 0; i < 32; i++) {
1476                 RADEON_WRITE((0x2c14 + j), 0x00000000);
1477                 RADEON_WRITE((0x2c18 + j), 0x00000000);
1478                 RADEON_WRITE((0x2c1c + j), 0x00000000);
1479                 RADEON_WRITE((0x2c20 + j), 0x00000000);
1480                 RADEON_WRITE((0x2c24 + j), 0x00000000);
1481                 j += 0x18;
1482         }
1483
1484         RADEON_WRITE(R600_GRBM_CNTL, R600_GRBM_READ_TIMEOUT(0xff));
1485
1486         /* setup tiling, simd, pipe config */
1487         mc_arb_ramcfg = RADEON_READ(R700_MC_ARB_RAMCFG);
1488
1489         switch (dev_priv->r600_max_tile_pipes) {
1490         case 1:
1491                 gb_tiling_config |= R600_PIPE_TILING(0);
1492                 break;
1493         case 2:
1494                 gb_tiling_config |= R600_PIPE_TILING(1);
1495                 break;
1496         case 4:
1497                 gb_tiling_config |= R600_PIPE_TILING(2);
1498                 break;
1499         case 8:
1500                 gb_tiling_config |= R600_PIPE_TILING(3);
1501                 break;
1502         default:
1503                 break;
1504         }
1505
1506         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV770)
1507                 gb_tiling_config |= R600_BANK_TILING(1);
1508         else
1509                 gb_tiling_config |= R600_BANK_TILING((mc_arb_ramcfg >> R700_NOOFBANK_SHIFT) & R700_NOOFBANK_MASK);
1510
1511         gb_tiling_config |= R600_GROUP_SIZE(0);
1512
1513         if (((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK) > 3) {
1514                 gb_tiling_config |= R600_ROW_TILING(3);
1515                 gb_tiling_config |= R600_SAMPLE_SPLIT(3);
1516         } else {
1517                 gb_tiling_config |=
1518                         R600_ROW_TILING(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1519                 gb_tiling_config |=
1520                         R600_SAMPLE_SPLIT(((mc_arb_ramcfg >> R700_NOOFROWS_SHIFT) & R700_NOOFROWS_MASK));
1521         }
1522
1523         gb_tiling_config |= R600_BANK_SWAPS(1);
1524
1525         cc_rb_backend_disable = RADEON_READ(R600_CC_RB_BACKEND_DISABLE) & 0x00ff0000;
1526         cc_rb_backend_disable |=
1527                 R600_BACKEND_DISABLE((R7XX_MAX_BACKENDS_MASK << dev_priv->r600_max_backends) & R7XX_MAX_BACKENDS_MASK);
1528
1529         cc_gc_shader_pipe_config = RADEON_READ(R600_CC_GC_SHADER_PIPE_CONFIG) & 0xffffff00;
1530         cc_gc_shader_pipe_config |=
1531                 R600_INACTIVE_QD_PIPES((R7XX_MAX_PIPES_MASK << dev_priv->r600_max_pipes) & R7XX_MAX_PIPES_MASK);
1532         cc_gc_shader_pipe_config |=
1533                 R600_INACTIVE_SIMDS((R7XX_MAX_SIMDS_MASK << dev_priv->r600_max_simds) & R7XX_MAX_SIMDS_MASK);
1534
1535         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV740)
1536                 backend_map = 0x28;
1537         else
1538                 backend_map = r700_get_tile_pipe_to_backend_map(dev_priv,
1539                                                                 dev_priv->r600_max_tile_pipes,
1540                                                                 (R7XX_MAX_BACKENDS -
1541                                                                  r600_count_pipe_bits((cc_rb_backend_disable &
1542                                                                                        R7XX_MAX_BACKENDS_MASK) >> 16)),
1543                                                                 (cc_rb_backend_disable >> 16));
1544         gb_tiling_config |= R600_BACKEND_MAP(backend_map);
1545
1546         RADEON_WRITE(R600_GB_TILING_CONFIG,      gb_tiling_config);
1547         RADEON_WRITE(R600_DCP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1548         RADEON_WRITE(R600_HDP_TILING_CONFIG,    (gb_tiling_config & 0xffff));
1549         if (gb_tiling_config & 0xc0) {
1550                 dev_priv->r600_group_size = 512;
1551         } else {
1552                 dev_priv->r600_group_size = 256;
1553         }
1554         dev_priv->r600_npipes = 1 << ((gb_tiling_config >> 1) & 0x7);
1555         if (gb_tiling_config & 0x30) {
1556                 dev_priv->r600_nbanks = 8;
1557         } else {
1558                 dev_priv->r600_nbanks = 4;
1559         }
1560
1561         RADEON_WRITE(R600_CC_RB_BACKEND_DISABLE,      cc_rb_backend_disable);
1562         RADEON_WRITE(R600_CC_GC_SHADER_PIPE_CONFIG,   cc_gc_shader_pipe_config);
1563         RADEON_WRITE(R600_GC_USER_SHADER_PIPE_CONFIG, cc_gc_shader_pipe_config);
1564
1565         RADEON_WRITE(R700_CC_SYS_RB_BACKEND_DISABLE, cc_rb_backend_disable);
1566         RADEON_WRITE(R700_CGTS_SYS_TCC_DISABLE, 0);
1567         RADEON_WRITE(R700_CGTS_TCC_DISABLE, 0);
1568         RADEON_WRITE(R700_CGTS_USER_SYS_TCC_DISABLE, 0);
1569         RADEON_WRITE(R700_CGTS_USER_TCC_DISABLE, 0);
1570
1571         num_qd_pipes =
1572                 R7XX_MAX_PIPES - r600_count_pipe_bits((cc_gc_shader_pipe_config & R600_INACTIVE_QD_PIPES_MASK) >> 8);
1573         RADEON_WRITE(R600_VGT_OUT_DEALLOC_CNTL, (num_qd_pipes * 4) & R600_DEALLOC_DIST_MASK);
1574         RADEON_WRITE(R600_VGT_VERTEX_REUSE_BLOCK_CNTL, ((num_qd_pipes * 4) - 2) & R600_VTX_REUSE_DEPTH_MASK);
1575
1576         /* set HW defaults for 3D engine */
1577         RADEON_WRITE(R600_CP_QUEUE_THRESHOLDS, (R600_ROQ_IB1_START(0x16) |
1578                                                 R600_ROQ_IB2_START(0x2b)));
1579
1580         RADEON_WRITE(R600_CP_MEQ_THRESHOLDS, R700_STQ_SPLIT(0x30));
1581
1582         ta_aux_cntl = RADEON_READ(R600_TA_CNTL_AUX);
1583         RADEON_WRITE(R600_TA_CNTL_AUX, ta_aux_cntl | R600_DISABLE_CUBE_ANISO);
1584
1585         sx_debug_1 = RADEON_READ(R700_SX_DEBUG_1);
1586         sx_debug_1 |= R700_ENABLE_NEW_SMX_ADDRESS;
1587         RADEON_WRITE(R700_SX_DEBUG_1, sx_debug_1);
1588
1589         smx_dc_ctl0 = RADEON_READ(R600_SMX_DC_CTL0);
1590         smx_dc_ctl0 &= ~R700_CACHE_DEPTH(0x1ff);
1591         smx_dc_ctl0 |= R700_CACHE_DEPTH((dev_priv->r700_sx_num_of_sets * 64) - 1);
1592         RADEON_WRITE(R600_SMX_DC_CTL0, smx_dc_ctl0);
1593
1594         if ((dev_priv->flags & RADEON_FAMILY_MASK) != CHIP_RV740)
1595                 RADEON_WRITE(R700_SMX_EVENT_CTL, (R700_ES_FLUSH_CTL(4) |
1596                                                   R700_GS_FLUSH_CTL(4) |
1597                                                   R700_ACK_FLUSH_CTL(3) |
1598                                                   R700_SYNC_FLUSH_CTL));
1599
1600         db_debug3 = RADEON_READ(R700_DB_DEBUG3);
1601         db_debug3 &= ~R700_DB_CLK_OFF_DELAY(0x1f);
1602         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1603         case CHIP_RV770:
1604         case CHIP_RV740:
1605                 db_debug3 |= R700_DB_CLK_OFF_DELAY(0x1f);
1606                 break;
1607         case CHIP_RV710:
1608         case CHIP_RV730:
1609         default:
1610                 db_debug3 |= R700_DB_CLK_OFF_DELAY(2);
1611                 break;
1612         }
1613         RADEON_WRITE(R700_DB_DEBUG3, db_debug3);
1614
1615         if ((dev_priv->flags & RADEON_FAMILY_MASK) != CHIP_RV770) {
1616                 db_debug4 = RADEON_READ(RV700_DB_DEBUG4);
1617                 db_debug4 |= RV700_DISABLE_TILE_COVERED_FOR_PS_ITER;
1618                 RADEON_WRITE(RV700_DB_DEBUG4, db_debug4);
1619         }
1620
1621         RADEON_WRITE(R600_SX_EXPORT_BUFFER_SIZES, (R600_COLOR_BUFFER_SIZE((dev_priv->r600_sx_max_export_size / 4) - 1) |
1622                                                    R600_POSITION_BUFFER_SIZE((dev_priv->r600_sx_max_export_pos_size / 4) - 1) |
1623                                                    R600_SMX_BUFFER_SIZE((dev_priv->r600_sx_max_export_smx_size / 4) - 1)));
1624
1625         RADEON_WRITE(R700_PA_SC_FIFO_SIZE_R7XX, (R700_SC_PRIM_FIFO_SIZE(dev_priv->r700_sc_prim_fifo_size) |
1626                                                  R700_SC_HIZ_TILE_FIFO_SIZE(dev_priv->r700_sc_hiz_tile_fifo_size) |
1627                                                  R700_SC_EARLYZ_TILE_FIFO_SIZE(dev_priv->r700_sc_earlyz_tile_fifo_fize)));
1628
1629         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1630
1631         RADEON_WRITE(R600_VGT_NUM_INSTANCES, 1);
1632
1633         RADEON_WRITE(R600_SPI_CONFIG_CNTL, R600_GPR_WRITE_PRIORITY(0));
1634
1635         RADEON_WRITE(R600_SPI_CONFIG_CNTL_1, R600_VTX_DONE_DELAY(4));
1636
1637         RADEON_WRITE(R600_CP_PERFMON_CNTL, 0);
1638
1639         sq_ms_fifo_sizes = (R600_CACHE_FIFO_SIZE(16 * dev_priv->r600_sq_num_cf_insts) |
1640                             R600_DONE_FIFO_HIWATER(0xe0) |
1641                             R600_ALU_UPDATE_FIFO_HIWATER(0x8));
1642         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1643         case CHIP_RV770:
1644         case CHIP_RV730:
1645         case CHIP_RV710:
1646                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x1);
1647                 break;
1648         case CHIP_RV740:
1649         default:
1650                 sq_ms_fifo_sizes |= R600_FETCH_FIFO_HIWATER(0x4);
1651                 break;
1652         }
1653         RADEON_WRITE(R600_SQ_MS_FIFO_SIZES, sq_ms_fifo_sizes);
1654
1655         /* SQ_CONFIG, SQ_GPR_RESOURCE_MGMT, SQ_THREAD_RESOURCE_MGMT, SQ_STACK_RESOURCE_MGMT
1656          * should be adjusted as needed by the 2D/3D drivers.  This just sets default values
1657          */
1658         sq_config = RADEON_READ(R600_SQ_CONFIG);
1659         sq_config &= ~(R600_PS_PRIO(3) |
1660                        R600_VS_PRIO(3) |
1661                        R600_GS_PRIO(3) |
1662                        R600_ES_PRIO(3));
1663         sq_config |= (R600_DX9_CONSTS |
1664                       R600_VC_ENABLE |
1665                       R600_EXPORT_SRC_C |
1666                       R600_PS_PRIO(0) |
1667                       R600_VS_PRIO(1) |
1668                       R600_GS_PRIO(2) |
1669                       R600_ES_PRIO(3));
1670         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1671                 /* no vertex cache */
1672                 sq_config &= ~R600_VC_ENABLE;
1673
1674         RADEON_WRITE(R600_SQ_CONFIG, sq_config);
1675
1676         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_1,  (R600_NUM_PS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1677                                                     R600_NUM_VS_GPRS((dev_priv->r600_max_gprs * 24)/64) |
1678                                                     R600_NUM_CLAUSE_TEMP_GPRS(((dev_priv->r600_max_gprs * 24)/64)/2)));
1679
1680         RADEON_WRITE(R600_SQ_GPR_RESOURCE_MGMT_2,  (R600_NUM_GS_GPRS((dev_priv->r600_max_gprs * 7)/64) |
1681                                                     R600_NUM_ES_GPRS((dev_priv->r600_max_gprs * 7)/64)));
1682
1683         sq_thread_resource_mgmt = (R600_NUM_PS_THREADS((dev_priv->r600_max_threads * 4)/8) |
1684                                    R600_NUM_VS_THREADS((dev_priv->r600_max_threads * 2)/8) |
1685                                    R600_NUM_ES_THREADS((dev_priv->r600_max_threads * 1)/8));
1686         if (((dev_priv->r600_max_threads * 1) / 8) > dev_priv->r600_max_gs_threads)
1687                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS(dev_priv->r600_max_gs_threads);
1688         else
1689                 sq_thread_resource_mgmt |= R600_NUM_GS_THREADS((dev_priv->r600_max_gs_threads * 1)/8);
1690         RADEON_WRITE(R600_SQ_THREAD_RESOURCE_MGMT, sq_thread_resource_mgmt);
1691
1692         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_1, (R600_NUM_PS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1693                                                      R600_NUM_VS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1694
1695         RADEON_WRITE(R600_SQ_STACK_RESOURCE_MGMT_2, (R600_NUM_GS_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4) |
1696                                                      R600_NUM_ES_STACK_ENTRIES((dev_priv->r600_max_stack_entries * 1)/4)));
1697
1698         sq_dyn_gpr_size_simd_ab_0 = (R700_SIMDA_RING0((dev_priv->r600_max_gprs * 38)/64) |
1699                                      R700_SIMDA_RING1((dev_priv->r600_max_gprs * 38)/64) |
1700                                      R700_SIMDB_RING0((dev_priv->r600_max_gprs * 38)/64) |
1701                                      R700_SIMDB_RING1((dev_priv->r600_max_gprs * 38)/64));
1702
1703         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_0, sq_dyn_gpr_size_simd_ab_0);
1704         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_1, sq_dyn_gpr_size_simd_ab_0);
1705         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_2, sq_dyn_gpr_size_simd_ab_0);
1706         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_3, sq_dyn_gpr_size_simd_ab_0);
1707         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_4, sq_dyn_gpr_size_simd_ab_0);
1708         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_5, sq_dyn_gpr_size_simd_ab_0);
1709         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_6, sq_dyn_gpr_size_simd_ab_0);
1710         RADEON_WRITE(R700_SQ_DYN_GPR_SIZE_SIMD_AB_7, sq_dyn_gpr_size_simd_ab_0);
1711
1712         RADEON_WRITE(R700_PA_SC_FORCE_EOV_MAX_CNTS, (R700_FORCE_EOV_MAX_CLK_CNT(4095) |
1713                                                      R700_FORCE_EOV_MAX_REZ_CNT(255)));
1714
1715         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV710)
1716                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_TC_ONLY) |
1717                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1718         else
1719                 RADEON_WRITE(R600_VGT_CACHE_INVALIDATION, (R600_CACHE_INVALIDATION(R600_VC_AND_TC) |
1720                                                            R700_AUTO_INVLD_EN(R700_ES_AND_GS_AUTO)));
1721
1722         switch (dev_priv->flags & RADEON_FAMILY_MASK) {
1723         case CHIP_RV770:
1724         case CHIP_RV730:
1725         case CHIP_RV740:
1726                 gs_prim_buffer_depth = 384;
1727                 break;
1728         case CHIP_RV710:
1729                 gs_prim_buffer_depth = 128;
1730                 break;
1731         default:
1732                 break;
1733         }
1734
1735         num_gs_verts_per_thread = dev_priv->r600_max_pipes * 16;
1736         vgt_gs_per_es = gs_prim_buffer_depth + num_gs_verts_per_thread;
1737         /* Max value for this is 256 */
1738         if (vgt_gs_per_es > 256)
1739                 vgt_gs_per_es = 256;
1740
1741         RADEON_WRITE(R600_VGT_ES_PER_GS, 128);
1742         RADEON_WRITE(R600_VGT_GS_PER_ES, vgt_gs_per_es);
1743         RADEON_WRITE(R600_VGT_GS_PER_VS, 2);
1744
1745         /* more default values. 2D/3D driver should adjust as needed */
1746         RADEON_WRITE(R600_VGT_GS_VERTEX_REUSE, 16);
1747         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE_STATE, 0);
1748         RADEON_WRITE(R600_VGT_STRMOUT_EN, 0);
1749         RADEON_WRITE(R600_SX_MISC, 0);
1750         RADEON_WRITE(R600_PA_SC_MODE_CNTL, 0);
1751         RADEON_WRITE(R700_PA_SC_EDGERULE, 0xaaaaaaaa);
1752         RADEON_WRITE(R600_PA_SC_AA_CONFIG, 0);
1753         RADEON_WRITE(R600_PA_SC_CLIPRECT_RULE, 0xffff);
1754         RADEON_WRITE(R600_PA_SC_LINE_STIPPLE, 0);
1755         RADEON_WRITE(R600_SPI_INPUT_Z, 0);
1756         RADEON_WRITE(R600_SPI_PS_IN_CONTROL_0, R600_NUM_INTERP(2));
1757         RADEON_WRITE(R600_CB_COLOR7_FRAG, 0);
1758
1759         /* clear render buffer base addresses */
1760         RADEON_WRITE(R600_CB_COLOR0_BASE, 0);
1761         RADEON_WRITE(R600_CB_COLOR1_BASE, 0);
1762         RADEON_WRITE(R600_CB_COLOR2_BASE, 0);
1763         RADEON_WRITE(R600_CB_COLOR3_BASE, 0);
1764         RADEON_WRITE(R600_CB_COLOR4_BASE, 0);
1765         RADEON_WRITE(R600_CB_COLOR5_BASE, 0);
1766         RADEON_WRITE(R600_CB_COLOR6_BASE, 0);
1767         RADEON_WRITE(R600_CB_COLOR7_BASE, 0);
1768
1769         RADEON_WRITE(R700_TCP_CNTL, 0);
1770
1771         hdp_host_path_cntl = RADEON_READ(R600_HDP_HOST_PATH_CNTL);
1772         RADEON_WRITE(R600_HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1773
1774         RADEON_WRITE(R600_PA_SC_MULTI_CHIP_CNTL, 0);
1775
1776         RADEON_WRITE(R600_PA_CL_ENHANCE, (R600_CLIP_VTX_REORDER_ENA |
1777                                           R600_NUM_CLIP_SEQ(3)));
1778
1779 }
1780
1781 static void r600_cp_init_ring_buffer(struct drm_device *dev,
1782                                        drm_radeon_private_t *dev_priv,
1783                                        struct drm_file *file_priv)
1784 {
1785         struct drm_radeon_master_private *master_priv;
1786         u32 ring_start;
1787         u64 rptr_addr;
1788
1789         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
1790                 r700_gfx_init(dev, dev_priv);
1791         else
1792                 r600_gfx_init(dev, dev_priv);
1793
1794         RADEON_WRITE(R600_GRBM_SOFT_RESET, R600_SOFT_RESET_CP);
1795         RADEON_READ(R600_GRBM_SOFT_RESET);
1796         DRM_MDELAY(15);
1797         RADEON_WRITE(R600_GRBM_SOFT_RESET, 0);
1798
1799
1800         /* Set ring buffer size */
1801 #ifdef __BIG_ENDIAN
1802         RADEON_WRITE(R600_CP_RB_CNTL,
1803                      R600_BUF_SWAP_32BIT |
1804                      R600_RB_NO_UPDATE |
1805                      (dev_priv->ring.rptr_update_l2qw << 8) |
1806                      dev_priv->ring.size_l2qw);
1807 #else
1808         RADEON_WRITE(R600_CP_RB_CNTL,
1809                      RADEON_RB_NO_UPDATE |
1810                      (dev_priv->ring.rptr_update_l2qw << 8) |
1811                      dev_priv->ring.size_l2qw);
1812 #endif
1813
1814         RADEON_WRITE(R600_CP_SEM_WAIT_TIMER, 0x0);
1815
1816         /* Set the write pointer delay */
1817         RADEON_WRITE(R600_CP_RB_WPTR_DELAY, 0);
1818
1819 #ifdef __BIG_ENDIAN
1820         RADEON_WRITE(R600_CP_RB_CNTL,
1821                      R600_BUF_SWAP_32BIT |
1822                      R600_RB_NO_UPDATE |
1823                      R600_RB_RPTR_WR_ENA |
1824                      (dev_priv->ring.rptr_update_l2qw << 8) |
1825                      dev_priv->ring.size_l2qw);
1826 #else
1827         RADEON_WRITE(R600_CP_RB_CNTL,
1828                      R600_RB_NO_UPDATE |
1829                      R600_RB_RPTR_WR_ENA |
1830                      (dev_priv->ring.rptr_update_l2qw << 8) |
1831                      dev_priv->ring.size_l2qw);
1832 #endif
1833
1834         /* Initialize the ring buffer's read and write pointers */
1835         RADEON_WRITE(R600_CP_RB_RPTR_WR, 0);
1836         RADEON_WRITE(R600_CP_RB_WPTR, 0);
1837         SET_RING_HEAD(dev_priv, 0);
1838         dev_priv->ring.tail = 0;
1839
1840 #if __OS_HAS_AGP
1841         if (dev_priv->flags & RADEON_IS_AGP) {
1842                 rptr_addr = dev_priv->ring_rptr->offset
1843                         - dev->agp->base +
1844                         dev_priv->gart_vm_start;
1845         } else
1846 #endif
1847         {
1848                 rptr_addr = dev_priv->ring_rptr->offset
1849                         - ((unsigned long) dev->sg->vaddr)
1850                         + dev_priv->gart_vm_start;
1851         }
1852         RADEON_WRITE(R600_CP_RB_RPTR_ADDR, (rptr_addr & 0xfffffffc));
1853         RADEON_WRITE(R600_CP_RB_RPTR_ADDR_HI, upper_32_bits(rptr_addr));
1854
1855 #ifdef __BIG_ENDIAN
1856         RADEON_WRITE(R600_CP_RB_CNTL,
1857                      RADEON_BUF_SWAP_32BIT |
1858                      (dev_priv->ring.rptr_update_l2qw << 8) |
1859                      dev_priv->ring.size_l2qw);
1860 #else
1861         RADEON_WRITE(R600_CP_RB_CNTL,
1862                      (dev_priv->ring.rptr_update_l2qw << 8) |
1863                      dev_priv->ring.size_l2qw);
1864 #endif
1865
1866 #if __OS_HAS_AGP
1867         if (dev_priv->flags & RADEON_IS_AGP) {
1868                 /* XXX */
1869                 radeon_write_agp_base(dev_priv, dev->agp->base);
1870
1871                 /* XXX */
1872                 radeon_write_agp_location(dev_priv,
1873                              (((dev_priv->gart_vm_start - 1 +
1874                                 dev_priv->gart_size) & 0xffff0000) |
1875                               (dev_priv->gart_vm_start >> 16)));
1876
1877                 ring_start = (dev_priv->cp_ring->offset
1878                               - dev->agp->base
1879                               + dev_priv->gart_vm_start);
1880         } else
1881 #endif
1882                 ring_start = (dev_priv->cp_ring->offset
1883                               - (unsigned long)dev->sg->vaddr>
1884                               + dev_priv->gart_vm_start);
1885
1886         RADEON_WRITE(R600_CP_RB_BASE, ring_start >> 8);
1887
1888         RADEON_WRITE(R600_CP_ME_CNTL, 0xff);
1889
1890         RADEON_WRITE(R600_CP_DEBUG, (1 << 27) | (1 << 28));
1891
1892         /* Initialize the scratch register pointer.  This will cause
1893          * the scratch register values to be written out to memory
1894          * whenever they are updated.
1895          *
1896          * We simply put this behind the ring read pointer, this works
1897          * with PCI GART as well as (whatever kind of) AGP GART
1898          */
1899         {
1900                 u64 scratch_addr;
1901
1902                 scratch_addr = RADEON_READ(R600_CP_RB_RPTR_ADDR) & 0xFFFFFFFC;
1903                 scratch_addr |= ((u64)RADEON_READ(R600_CP_RB_RPTR_ADDR_HI)) << 32;
1904                 scratch_addr += R600_SCRATCH_REG_OFFSET;
1905                 scratch_addr >>= 8;
1906                 scratch_addr &= 0xffffffff;
1907
1908                 RADEON_WRITE(R600_SCRATCH_ADDR, (uint32_t)scratch_addr);
1909         }
1910
1911         RADEON_WRITE(R600_SCRATCH_UMSK, 0x7);
1912
1913         /* Turn on bus mastering */
1914         radeon_enable_bm(dev_priv);
1915
1916         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(0), 0);
1917         RADEON_WRITE(R600_LAST_FRAME_REG, 0);
1918
1919         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(1), 0);
1920         RADEON_WRITE(R600_LAST_DISPATCH_REG, 0);
1921
1922         radeon_write_ring_rptr(dev_priv, R600_SCRATCHOFF(2), 0);
1923         RADEON_WRITE(R600_LAST_CLEAR_REG, 0);
1924
1925         /* reset sarea copies of these */
1926         master_priv = file_priv->masterp->driver_priv;
1927         if (master_priv->sarea_priv) {
1928                 master_priv->sarea_priv->last_frame = 0;
1929                 master_priv->sarea_priv->last_dispatch = 0;
1930                 master_priv->sarea_priv->last_clear = 0;
1931         }
1932
1933         r600_do_wait_for_idle(dev_priv);
1934
1935 }
1936
1937 int r600_do_cleanup_cp(struct drm_device *dev)
1938 {
1939         drm_radeon_private_t *dev_priv = dev->dev_private;
1940         DRM_DEBUG("\n");
1941
1942         /* Make sure interrupts are disabled here because the uninstall ioctl
1943          * may not have been called from userspace and after dev_private
1944          * is freed, it's too late.
1945          */
1946         if (dev->irq_enabled)
1947                 drm_irq_uninstall(dev);
1948
1949 #if __OS_HAS_AGP
1950         if (dev_priv->flags & RADEON_IS_AGP) {
1951                 if (dev_priv->cp_ring != NULL) {
1952                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1953                         dev_priv->cp_ring = NULL;
1954                 }
1955                 if (dev_priv->ring_rptr != NULL) {
1956                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1957                         dev_priv->ring_rptr = NULL;
1958                 }
1959                 if (dev->agp_buffer_map != NULL) {
1960                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1961                         dev->agp_buffer_map = NULL;
1962                 }
1963         } else
1964 #endif
1965         {
1966
1967                 if (dev_priv->gart_info.bus_addr)
1968                         r600_page_table_cleanup(dev, &dev_priv->gart_info);
1969
1970                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB) {
1971                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1972                         dev_priv->gart_info.addr = NULL;
1973                 }
1974         }
1975         /* only clear to the start of flags */
1976         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1977
1978         return 0;
1979 }
1980
1981 int r600_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
1982                     struct drm_file *file_priv)
1983 {
1984         drm_radeon_private_t *dev_priv = dev->dev_private;
1985         struct drm_radeon_master_private *master_priv = file_priv->masterp->driver_priv;
1986
1987         DRM_DEBUG("\n");
1988
1989         lockinit(&dev_priv->cs_mutex, "drm__radeon_private__cs_mutex", 0,
1990                  LK_CANRECURSE);
1991         r600_cs_legacy_init();
1992         /* if we require new memory map but we don't have it fail */
1993         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
1994                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
1995                 r600_do_cleanup_cp(dev);
1996                 return -EINVAL;
1997         }
1998
1999         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
2000                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
2001                 dev_priv->flags &= ~RADEON_IS_AGP;
2002                 /* The writeback test succeeds, but when writeback is enabled,
2003                  * the ring buffer read ptr update fails after first 128 bytes.
2004                  */
2005                 radeon_no_wb = 1;
2006         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
2007                  && !init->is_pci) {
2008                 DRM_DEBUG("Restoring AGP flag\n");
2009                 dev_priv->flags |= RADEON_IS_AGP;
2010         }
2011
2012         dev_priv->usec_timeout = init->usec_timeout;
2013         if (dev_priv->usec_timeout < 1 ||
2014             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
2015                 DRM_DEBUG("TIMEOUT problem!\n");
2016                 r600_do_cleanup_cp(dev);
2017                 return -EINVAL;
2018         }
2019
2020         /* Enable vblank on CRTC1 for older X servers
2021          */
2022         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
2023         dev_priv->do_boxes = 0;
2024         dev_priv->cp_mode = init->cp_mode;
2025
2026         /* We don't support anything other than bus-mastering ring mode,
2027          * but the ring can be in either AGP or PCI space for the ring
2028          * read pointer.
2029          */
2030         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
2031             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
2032                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
2033                 r600_do_cleanup_cp(dev);
2034                 return -EINVAL;
2035         }
2036
2037         switch (init->fb_bpp) {
2038         case 16:
2039                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
2040                 break;
2041         case 32:
2042         default:
2043                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
2044                 break;
2045         }
2046         dev_priv->front_offset = init->front_offset;
2047         dev_priv->front_pitch = init->front_pitch;
2048         dev_priv->back_offset = init->back_offset;
2049         dev_priv->back_pitch = init->back_pitch;
2050
2051         dev_priv->ring_offset = init->ring_offset;
2052         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
2053         dev_priv->buffers_offset = init->buffers_offset;
2054         dev_priv->gart_textures_offset = init->gart_textures_offset;
2055
2056         master_priv->sarea = drm_getsarea(dev);
2057         if (!master_priv->sarea) {
2058                 DRM_ERROR("could not find sarea!\n");
2059                 r600_do_cleanup_cp(dev);
2060                 return -EINVAL;
2061         }
2062
2063         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
2064         if (!dev_priv->cp_ring) {
2065                 DRM_ERROR("could not find cp ring region!\n");
2066                 r600_do_cleanup_cp(dev);
2067                 return -EINVAL;
2068         }
2069         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
2070         if (!dev_priv->ring_rptr) {
2071                 DRM_ERROR("could not find ring read pointer!\n");
2072                 r600_do_cleanup_cp(dev);
2073                 return -EINVAL;
2074         }
2075         dev->agp_buffer_token = init->buffers_offset;
2076         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
2077         if (!dev->agp_buffer_map) {
2078                 DRM_ERROR("could not find dma buffer region!\n");
2079                 r600_do_cleanup_cp(dev);
2080                 return -EINVAL;
2081         }
2082
2083         if (init->gart_textures_offset) {
2084                 dev_priv->gart_textures =
2085                     drm_core_findmap(dev, init->gart_textures_offset);
2086                 if (!dev_priv->gart_textures) {
2087                         DRM_ERROR("could not find GART texture region!\n");
2088                         r600_do_cleanup_cp(dev);
2089                         return -EINVAL;
2090                 }
2091         }
2092
2093 #if __OS_HAS_AGP
2094         /* XXX */
2095         if (dev_priv->flags & RADEON_IS_AGP) {
2096                 drm_core_ioremap_wc(dev_priv->cp_ring, dev);
2097                 drm_core_ioremap_wc(dev_priv->ring_rptr, dev);
2098                 drm_core_ioremap_wc(dev->agp_buffer_map, dev);
2099                 if (!dev_priv->cp_ring->handle ||
2100                     !dev_priv->ring_rptr->handle ||
2101                     !dev->agp_buffer_map->handle) {
2102                         DRM_ERROR("could not find ioremap agp regions!\n");
2103                         r600_do_cleanup_cp(dev);
2104                         return -EINVAL;
2105                 }
2106         } else
2107 #endif
2108         {
2109                 dev_priv->cp_ring->handle = (void *)(unsigned long)dev_priv->cp_ring->offset;
2110                 dev_priv->ring_rptr->handle =
2111                         (void *)(unsigned long)dev_priv->ring_rptr->offset;
2112                 dev->agp_buffer_map->handle =
2113                         (void *)(unsigned long)dev->agp_buffer_map->offset;
2114
2115                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
2116                           dev_priv->cp_ring->handle);
2117                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
2118                           dev_priv->ring_rptr->handle);
2119                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
2120                           dev->agp_buffer_map->handle);
2121         }
2122
2123         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 24;
2124         dev_priv->fb_size =
2125                 (((radeon_read_fb_location(dev_priv) & 0xffff0000u) << 8) + 0x1000000)
2126                 - dev_priv->fb_location;
2127
2128         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
2129                                         ((dev_priv->front_offset
2130                                           + dev_priv->fb_location) >> 10));
2131
2132         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
2133                                        ((dev_priv->back_offset
2134                                          + dev_priv->fb_location) >> 10));
2135
2136         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
2137                                         ((dev_priv->depth_offset
2138                                           + dev_priv->fb_location) >> 10));
2139
2140         dev_priv->gart_size = init->gart_size;
2141
2142         /* New let's set the memory map ... */
2143         if (dev_priv->new_memmap) {
2144                 u32 base = 0;
2145
2146                 DRM_INFO("Setting GART location based on new memory map\n");
2147
2148                 /* If using AGP, try to locate the AGP aperture at the same
2149                  * location in the card and on the bus, though we have to
2150                  * align it down.
2151                  */
2152 #if __OS_HAS_AGP
2153                 /* XXX */
2154                 if (dev_priv->flags & RADEON_IS_AGP) {
2155                         base = dev->agp->base;
2156                         /* Check if valid */
2157                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
2158                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
2159                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
2160                                          dev->agp->base);
2161                                 base = 0;
2162                         }
2163                 }
2164 #endif
2165                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
2166                 if (base == 0) {
2167                         base = dev_priv->fb_location + dev_priv->fb_size;
2168                         if (base < dev_priv->fb_location ||
2169                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
2170                                 base = dev_priv->fb_location
2171                                         - dev_priv->gart_size;
2172                 }
2173                 dev_priv->gart_vm_start = base & 0xffc00000u;
2174                 if (dev_priv->gart_vm_start != base)
2175                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
2176                                  base, dev_priv->gart_vm_start);
2177         }
2178
2179 #if __OS_HAS_AGP
2180         /* XXX */
2181         if (dev_priv->flags & RADEON_IS_AGP)
2182                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2183                                                  - dev->agp->base
2184                                                  + dev_priv->gart_vm_start);
2185         else
2186 #endif
2187                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
2188                                                  - (unsigned long)dev->sg->vaddr
2189                                                  + dev_priv->gart_vm_start);
2190
2191         DRM_DEBUG("fb 0x%08x size %d\n",
2192                   (unsigned int) dev_priv->fb_location,
2193                   (unsigned int) dev_priv->fb_size);
2194         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
2195         DRM_DEBUG("dev_priv->gart_vm_start 0x%08x\n",
2196                   (unsigned int) dev_priv->gart_vm_start);
2197         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%08lx\n",
2198                   dev_priv->gart_buffers_offset);
2199
2200         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
2201         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
2202                               + init->ring_size / sizeof(u32));
2203         dev_priv->ring.size = init->ring_size;
2204         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
2205
2206         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
2207         dev_priv->ring.rptr_update_l2qw = drm_order(/* init->rptr_update */ 4096 / 8);
2208
2209         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
2210         dev_priv->ring.fetch_size_l2ow = drm_order(/* init->fetch_size */ 32 / 16);
2211
2212         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
2213
2214         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
2215
2216 #if __OS_HAS_AGP
2217         if (dev_priv->flags & RADEON_IS_AGP) {
2218                 /* XXX turn off pcie gart */
2219         } else
2220 #endif
2221         {
2222                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
2223                 /* if we have an offset set from userspace */
2224                 if (!dev_priv->pcigart_offset_set) {
2225                         DRM_ERROR("Need gart offset from userspace\n");
2226                         r600_do_cleanup_cp(dev);
2227                         return -EINVAL;
2228                 }
2229
2230                 DRM_DEBUG("Using gart offset 0x%08lx\n", dev_priv->pcigart_offset);
2231
2232                 dev_priv->gart_info.bus_addr =
2233                         dev_priv->pcigart_offset + dev_priv->fb_location;
2234                 dev_priv->gart_info.mapping.offset =
2235                         dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
2236                 dev_priv->gart_info.mapping.size =
2237                         dev_priv->gart_info.table_size;
2238
2239                 drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
2240                 if (!dev_priv->gart_info.mapping.handle) {
2241                         DRM_ERROR("ioremap failed.\n");
2242                         r600_do_cleanup_cp(dev);
2243                         return -EINVAL;
2244                 }
2245
2246                 dev_priv->gart_info.addr =
2247                         dev_priv->gart_info.mapping.handle;
2248
2249                 DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
2250                           dev_priv->gart_info.addr,
2251                           dev_priv->pcigart_offset);
2252
2253                 if (!r600_page_table_init(dev)) {
2254                         DRM_ERROR("Failed to init GART table\n");
2255                         r600_do_cleanup_cp(dev);
2256                         return -EINVAL;
2257                 }
2258
2259                 if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2260                         r700_vm_init(dev);
2261                 else
2262                         r600_vm_init(dev);
2263         }
2264
2265         if (!dev_priv->me_fw || !dev_priv->pfp_fw) {
2266                 int err = r600_cp_init_microcode(dev_priv);
2267                 if (err) {
2268                         DRM_ERROR("Failed to load firmware!\n");
2269                         r600_do_cleanup_cp(dev);
2270                         return err;
2271                 }
2272         }
2273         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770))
2274                 r700_cp_load_microcode(dev_priv);
2275         else
2276                 r600_cp_load_microcode(dev_priv);
2277
2278         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2279
2280         dev_priv->last_buf = 0;
2281
2282         r600_do_engine_reset(dev);
2283         r600_test_writeback(dev_priv);
2284
2285         return 0;
2286 }
2287
2288 int r600_do_resume_cp(struct drm_device *dev, struct drm_file *file_priv)
2289 {
2290         drm_radeon_private_t *dev_priv = dev->dev_private;
2291
2292         DRM_DEBUG("\n");
2293         if (((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)) {
2294                 r700_vm_init(dev);
2295                 r700_cp_load_microcode(dev_priv);
2296         } else {
2297                 r600_vm_init(dev);
2298                 r600_cp_load_microcode(dev_priv);
2299         }
2300         r600_cp_init_ring_buffer(dev, dev_priv, file_priv);
2301         r600_do_engine_reset(dev);
2302
2303         return 0;
2304 }
2305
2306 /* Wait for the CP to go idle.
2307  */
2308 int r600_do_cp_idle(drm_radeon_private_t *dev_priv)
2309 {
2310         RING_LOCALS;
2311         DRM_DEBUG("\n");
2312
2313         BEGIN_RING(5);
2314         OUT_RING(CP_PACKET3(R600_IT_EVENT_WRITE, 0));
2315         OUT_RING(R600_CACHE_FLUSH_AND_INV_EVENT);
2316         /* wait for 3D idle clean */
2317         OUT_RING(CP_PACKET3(R600_IT_SET_CONFIG_REG, 1));
2318         OUT_RING((R600_WAIT_UNTIL - R600_SET_CONFIG_REG_OFFSET) >> 2);
2319         OUT_RING(RADEON_WAIT_3D_IDLE | RADEON_WAIT_3D_IDLECLEAN);
2320
2321         ADVANCE_RING();
2322         COMMIT_RING();
2323
2324         return r600_do_wait_for_idle(dev_priv);
2325 }
2326
2327 /* Start the Command Processor.
2328  */
2329 void r600_do_cp_start(drm_radeon_private_t *dev_priv)
2330 {
2331         u32 cp_me;
2332         RING_LOCALS;
2333         DRM_DEBUG("\n");
2334
2335         BEGIN_RING(7);
2336         OUT_RING(CP_PACKET3(R600_IT_ME_INITIALIZE, 5));
2337         OUT_RING(0x00000001);
2338         if (((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_RV770))
2339                 OUT_RING(0x00000003);
2340         else
2341                 OUT_RING(0x00000000);
2342         OUT_RING((dev_priv->r600_max_hw_contexts - 1));
2343         OUT_RING(R600_ME_INITIALIZE_DEVICE_ID(1));
2344         OUT_RING(0x00000000);
2345         OUT_RING(0x00000000);
2346         ADVANCE_RING();
2347         COMMIT_RING();
2348
2349         /* set the mux and reset the halt bit */
2350         cp_me = 0xff;
2351         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2352
2353         dev_priv->cp_running = 1;
2354
2355 }
2356
2357 void r600_do_cp_reset(drm_radeon_private_t *dev_priv)
2358 {
2359         u32 cur_read_ptr;
2360         DRM_DEBUG("\n");
2361
2362         cur_read_ptr = RADEON_READ(R600_CP_RB_RPTR);
2363         RADEON_WRITE(R600_CP_RB_WPTR, cur_read_ptr);
2364         SET_RING_HEAD(dev_priv, cur_read_ptr);
2365         dev_priv->ring.tail = cur_read_ptr;
2366 }
2367
2368 void r600_do_cp_stop(drm_radeon_private_t *dev_priv)
2369 {
2370         uint32_t cp_me;
2371
2372         DRM_DEBUG("\n");
2373
2374         cp_me = 0xff | R600_CP_ME_HALT;
2375
2376         RADEON_WRITE(R600_CP_ME_CNTL, cp_me);
2377
2378         dev_priv->cp_running = 0;
2379 }
2380
2381 int r600_cp_dispatch_indirect(struct drm_device *dev,
2382                               struct drm_buf *buf, int start, int end)
2383 {
2384         drm_radeon_private_t *dev_priv = dev->dev_private;
2385         RING_LOCALS;
2386
2387         if (start != end) {
2388                 unsigned long offset = (dev_priv->gart_buffers_offset
2389                                         + buf->offset + start);
2390                 int dwords = (end - start + 3) / sizeof(u32);
2391
2392                 DRM_DEBUG("dwords:%d\n", dwords);
2393                 DRM_DEBUG("offset 0x%lx\n", offset);
2394
2395
2396                 /* Indirect buffer data must be a multiple of 16 dwords.
2397                  * pad the data with a Type-2 CP packet.
2398                  */
2399                 while (dwords & 0xf) {
2400                         u32 *data = (u32 *)
2401                             ((char *)dev->agp_buffer_map->handle
2402                              + buf->offset + start);
2403                         data[dwords++] = RADEON_CP_PACKET2;
2404                 }
2405
2406                 /* Fire off the indirect buffer */
2407                 BEGIN_RING(4);
2408                 OUT_RING(CP_PACKET3(R600_IT_INDIRECT_BUFFER, 2));
2409                 OUT_RING((offset & 0xfffffffc));
2410                 OUT_RING((upper_32_bits(offset) & 0xff));
2411                 OUT_RING(dwords);
2412                 ADVANCE_RING();
2413         }
2414
2415         return 0;
2416 }
2417
2418 void r600_cp_dispatch_swap(struct drm_device *dev, struct drm_file *file_priv)
2419 {
2420         drm_radeon_private_t *dev_priv = dev->dev_private;
2421         struct drm_master *master = file_priv->masterp;
2422         struct drm_radeon_master_private *master_priv = master->driver_priv;
2423         drm_radeon_sarea_t *sarea_priv = master_priv->sarea_priv;
2424         int nbox = sarea_priv->nbox;
2425         struct drm_clip_rect *pbox = sarea_priv->boxes;
2426         int i, cpp, src_pitch, dst_pitch;
2427         uint64_t src, dst;
2428         RING_LOCALS;
2429         DRM_DEBUG("\n");
2430
2431         if (dev_priv->color_fmt == RADEON_COLOR_FORMAT_ARGB8888)
2432                 cpp = 4;
2433         else
2434                 cpp = 2;
2435
2436         if (sarea_priv->pfCurrentPage == 0) {
2437                 src_pitch = dev_priv->back_pitch;
2438                 dst_pitch = dev_priv->front_pitch;
2439                 src = dev_priv->back_offset + dev_priv->fb_location;
2440                 dst = dev_priv->front_offset + dev_priv->fb_location;
2441         } else {
2442                 src_pitch = dev_priv->front_pitch;
2443                 dst_pitch = dev_priv->back_pitch;
2444                 src = dev_priv->front_offset + dev_priv->fb_location;
2445                 dst = dev_priv->back_offset + dev_priv->fb_location;
2446         }
2447
2448         if (r600_prepare_blit_copy(dev, file_priv)) {
2449                 DRM_ERROR("unable to allocate vertex buffer for swap buffer\n");
2450                 return;
2451         }
2452         for (i = 0; i < nbox; i++) {
2453                 int x = pbox[i].x1;
2454                 int y = pbox[i].y1;
2455                 int w = pbox[i].x2 - x;
2456                 int h = pbox[i].y2 - y;
2457
2458                 DRM_DEBUG("%d,%d-%d,%d\n", x, y, w, h);
2459
2460                 r600_blit_swap(dev,
2461                                src, dst,
2462                                x, y, x, y, w, h,
2463                                src_pitch, dst_pitch, cpp);
2464         }
2465         r600_done_blit_copy(dev);
2466
2467         /* Increment the frame counter.  The client-side 3D driver must
2468          * throttle the framerate by waiting for this value before
2469          * performing the swapbuffer ioctl.
2470          */
2471         sarea_priv->last_frame++;
2472
2473         BEGIN_RING(3);
2474         R600_FRAME_AGE(sarea_priv->last_frame);
2475         ADVANCE_RING();
2476 }
2477
2478 int r600_cp_dispatch_texture(struct drm_device *dev,
2479                              struct drm_file *file_priv,
2480                              drm_radeon_texture_t *tex,
2481                              drm_radeon_tex_image_t *image)
2482 {
2483         drm_radeon_private_t *dev_priv = dev->dev_private;
2484         struct drm_buf *buf;
2485         u32 *buffer;
2486         const u8 __user *data;
2487         int size, pass_size;
2488         u64 src_offset, dst_offset;
2489
2490         if (!radeon_check_offset(dev_priv, tex->offset)) {
2491                 DRM_ERROR("Invalid destination offset\n");
2492                 return -EINVAL;
2493         }
2494
2495         /* this might fail for zero-sized uploads - are those illegal? */
2496         if (!radeon_check_offset(dev_priv, tex->offset + tex->height * tex->pitch - 1)) {
2497                 DRM_ERROR("Invalid final destination offset\n");
2498                 return -EINVAL;
2499         }
2500
2501         size = tex->height * tex->pitch;
2502
2503         if (size == 0)
2504                 return 0;
2505
2506         dst_offset = tex->offset;
2507
2508         if (r600_prepare_blit_copy(dev, file_priv)) {
2509                 DRM_ERROR("unable to allocate vertex buffer for swap buffer\n");
2510                 return -EAGAIN;
2511         }
2512         do {
2513                 data = (const u8 __user *)image->data;
2514                 pass_size = size;
2515
2516                 buf = radeon_freelist_get(dev);
2517                 if (!buf) {
2518                         DRM_DEBUG("EAGAIN\n");
2519                         if (DRM_COPY_TO_USER(tex->image, image, sizeof(*image)))
2520                                 return -EFAULT;
2521                         return -EAGAIN;
2522                 }
2523
2524                 if (pass_size > buf->total)
2525                         pass_size = buf->total;
2526
2527                 /* Dispatch the indirect buffer.
2528                  */
2529                 buffer =
2530                     (u32 *) ((char *)dev->agp_buffer_map->handle + buf->offset);
2531
2532                 if (DRM_COPY_FROM_USER(buffer, data, pass_size)) {
2533                         DRM_ERROR("EFAULT on pad, %d bytes\n", pass_size);
2534                         return -EFAULT;
2535                 }
2536
2537                 buf->file_priv = file_priv;
2538                 buf->used = pass_size;
2539                 src_offset = dev_priv->gart_buffers_offset + buf->offset;
2540
2541                 r600_blit_copy(dev, src_offset, dst_offset, pass_size);
2542
2543                 radeon_cp_discard_buffer(dev, file_priv->masterp, buf);
2544
2545                 /* Update the input parameters for next time */
2546                 image->data = (const u8 __user *)image->data + pass_size;
2547                 dst_offset += pass_size;
2548                 size -= pass_size;
2549         } while (size > 0);
2550         r600_done_blit_copy(dev);
2551
2552         return 0;
2553 }
2554
2555 /*
2556  * Legacy cs ioctl
2557  */
2558 static u32 radeon_cs_id_get(struct drm_radeon_private *radeon)
2559 {
2560         /* FIXME: check if wrap affect last reported wrap & sequence */
2561         radeon->cs_id_scnt = (radeon->cs_id_scnt + 1) & 0x00FFFFFF;
2562         if (!radeon->cs_id_scnt) {
2563                 /* increment wrap counter */
2564                 radeon->cs_id_wcnt += 0x01000000;
2565                 /* valid sequence counter start at 1 */
2566                 radeon->cs_id_scnt = 1;
2567         }
2568         return (radeon->cs_id_scnt | radeon->cs_id_wcnt);
2569 }
2570
2571 static void r600_cs_id_emit(drm_radeon_private_t *dev_priv, u32 *id)
2572 {
2573         RING_LOCALS;
2574
2575         *id = radeon_cs_id_get(dev_priv);
2576
2577         /* SCRATCH 2 */
2578         BEGIN_RING(3);
2579         R600_CLEAR_AGE(*id);
2580         ADVANCE_RING();
2581         COMMIT_RING();
2582 }
2583
2584 static int r600_ib_get(struct drm_device *dev,
2585                         struct drm_file *fpriv,
2586                         struct drm_buf **buffer)
2587 {
2588         struct drm_buf *buf;
2589
2590         *buffer = NULL;
2591         buf = radeon_freelist_get(dev);
2592         if (!buf) {
2593                 return -EBUSY;
2594         }
2595         buf->file_priv = fpriv;
2596         *buffer = buf;
2597         return 0;
2598 }
2599
2600 static void r600_ib_free(struct drm_device *dev, struct drm_buf *buf,
2601                         struct drm_file *fpriv, int l, int r)
2602 {
2603         drm_radeon_private_t *dev_priv = dev->dev_private;
2604
2605         if (buf) {
2606                 if (!r)
2607                         r600_cp_dispatch_indirect(dev, buf, 0, l * 4);
2608                 radeon_cp_discard_buffer(dev, fpriv->masterp, buf);
2609                 COMMIT_RING();
2610         }
2611 }
2612
2613 int r600_cs_legacy_ioctl(struct drm_device *dev, void *data, struct drm_file *fpriv)
2614 {
2615         struct drm_radeon_private *dev_priv = dev->dev_private;
2616         struct drm_radeon_cs *cs = data;
2617         struct drm_buf *buf;
2618         unsigned family;
2619         int l, r = 0;
2620         u32 *ib, cs_id = 0;
2621
2622         if (dev_priv == NULL) {
2623                 DRM_ERROR("called with no initialization\n");
2624                 return -EINVAL;
2625         }
2626         family = dev_priv->flags & RADEON_FAMILY_MASK;
2627         if (family < CHIP_R600) {
2628                 DRM_ERROR("cs ioctl valid only for R6XX & R7XX in legacy mode\n");
2629                 return -EINVAL;
2630         }
2631         lockmgr(&dev_priv->cs_mutex, LK_EXCLUSIVE);
2632         /* get ib */
2633         l = 0;
2634         r = r600_ib_get(dev, fpriv, &buf);
2635         if (r) {
2636                 DRM_ERROR("ib_get failed\n");
2637                 goto out;
2638         }
2639         ib = (u32 *)((uintptr_t)dev->agp_buffer_map->handle + buf->offset);
2640         /* now parse command stream */
2641         r = r600_cs_legacy(dev, data,  fpriv, family, ib, &l);
2642         if (r) {
2643                 goto out;
2644         }
2645
2646 out:
2647         r600_ib_free(dev, buf, fpriv, l, r);
2648         /* emit cs id sequence */
2649         r600_cs_id_emit(dev_priv, &cs_id);
2650         cs->cs_id = cs_id;
2651         lockmgr(&dev_priv->cs_mutex, LK_RELEASE);
2652         return r;
2653 }
2654
2655 void r600_cs_legacy_get_tiling_conf(struct drm_device *dev, u32 *npipes, u32 *nbanks, u32 *group_size)
2656 {
2657         struct drm_radeon_private *dev_priv = dev->dev_private;
2658
2659         *npipes = dev_priv->r600_npipes;
2660         *nbanks = dev_priv->r600_nbanks;
2661         *group_size = dev_priv->r600_group_size;
2662 }