drivers/gpu/drm/radeon/radeon_cp.c

   1 /* radeon_cp.c -- CP support for Radeon -*- linux-c -*- */
   2 /*
   3  * Copyright 2000 Precision Insight, Inc., Cedar Park, Texas.
   4  * Copyright 2000 VA Linux Systems, Inc., Fremont, California.
   5  * Copyright 2007 Advanced Micro Devices, Inc.
   6  * All Rights Reserved.
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  23  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  24  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  25  * DEALINGS IN THE SOFTWARE.
  26  *
  27  * Authors:
  28  *    Kevin E. Martin <martin@valinux.com>
  29  *    Gareth Hughes <gareth@valinux.com>
  30  */
  31
  32 #include "drmP.h"
  33 #include "drm.h"
  34 #include "radeon_drm.h"
  35 #include "radeon_drv.h"
  36 #include "r300_reg.h"
  37
  38 #include "radeon_microcode.h"
  39
  40 #define RADEON_FIFO_DEBUG       0
  41
  42 static int radeon_do_cleanup_cp(struct drm_device * dev);
  43
  44 static u32 R500_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  45 {
  46         u32 ret;
  47         RADEON_WRITE(R520_MC_IND_INDEX, 0x7f0000 | (addr & 0xff));
  48         ret = RADEON_READ(R520_MC_IND_DATA);
  49         RADEON_WRITE(R520_MC_IND_INDEX, 0);
  50         return ret;
  51 }
  52
  53 static u32 RS480_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  54 {
  55         u32 ret;
  56         RADEON_WRITE(RS480_NB_MC_INDEX, addr & 0xff);
  57         ret = RADEON_READ(RS480_NB_MC_DATA);
  58         RADEON_WRITE(RS480_NB_MC_INDEX, 0xff);
  59         return ret;
  60 }
  61
  62 static u32 RS690_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  63 {
  64         u32 ret;
  65         RADEON_WRITE(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
  66         ret = RADEON_READ(RS690_MC_DATA);
  67         RADEON_WRITE(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
  68         return ret;
  69 }
  70
  71 static u32 IGP_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  72 {
  73         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690)
  74                 return RS690_READ_MCIND(dev_priv, addr);
  75         else
  76                 return RS480_READ_MCIND(dev_priv, addr);
  77 }
  78
  79 u32 radeon_read_fb_location(drm_radeon_private_t *dev_priv)
  80 {
  81
  82         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  83                 return R500_READ_MCIND(dev_priv, RV515_MC_FB_LOCATION);
  84         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690)
  85                 return RS690_READ_MCIND(dev_priv, RS690_MC_FB_LOCATION);
  86         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
  87                 return R500_READ_MCIND(dev_priv, R520_MC_FB_LOCATION);
  88         else
  89                 return RADEON_READ(RADEON_MC_FB_LOCATION);
  90 }
  91
  92 static void radeon_write_fb_location(drm_radeon_private_t *dev_priv, u32 fb_loc)
  93 {
  94         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  95                 R500_WRITE_MCIND(RV515_MC_FB_LOCATION, fb_loc);
  96         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690)
  97                 RS690_WRITE_MCIND(RS690_MC_FB_LOCATION, fb_loc);
  98         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
  99                 R500_WRITE_MCIND(R520_MC_FB_LOCATION, fb_loc);
 100         else
 101                 RADEON_WRITE(RADEON_MC_FB_LOCATION, fb_loc);
 102 }
 103
 104 static void radeon_write_agp_location(drm_radeon_private_t *dev_priv, u32 agp_loc)
 105 {
 106         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 107                 R500_WRITE_MCIND(RV515_MC_AGP_LOCATION, agp_loc);
 108         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690)
 109                 RS690_WRITE_MCIND(RS690_MC_AGP_LOCATION, agp_loc);
 110         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 111                 R500_WRITE_MCIND(R520_MC_AGP_LOCATION, agp_loc);
 112         else
 113                 RADEON_WRITE(RADEON_MC_AGP_LOCATION, agp_loc);
 114 }
 115
 116 static void radeon_write_agp_base(drm_radeon_private_t *dev_priv, u64 agp_base)
 117 {
 118         u32 agp_base_hi = upper_32_bits(agp_base);
 119         u32 agp_base_lo = agp_base & 0xffffffff;
 120
 121         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) {
 122                 R500_WRITE_MCIND(RV515_MC_AGP_BASE, agp_base_lo);
 123                 R500_WRITE_MCIND(RV515_MC_AGP_BASE_2, agp_base_hi);
 124         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) {
 125                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE, agp_base_lo);
 126                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE_2, agp_base_hi);
 127         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515) {
 128                 R500_WRITE_MCIND(R520_MC_AGP_BASE, agp_base_lo);
 129                 R500_WRITE_MCIND(R520_MC_AGP_BASE_2, agp_base_hi);
 130         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480) {
 131                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 132                 RADEON_WRITE(RS480_AGP_BASE_2, 0);
 133         } else {
 134                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 135                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R200)
 136                         RADEON_WRITE(RADEON_AGP_BASE_2, agp_base_hi);
 137         }
 138 }
 139
 140 static int RADEON_READ_PLL(struct drm_device * dev, int addr)
 141 {
 142         drm_radeon_private_t *dev_priv = dev->dev_private;
 143
 144         RADEON_WRITE8(RADEON_CLOCK_CNTL_INDEX, addr & 0x1f);
 145         return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 146 }
 147
 148 static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 149 {
 150         RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 151         return RADEON_READ(RADEON_PCIE_DATA);
 152 }
 153
 154 #if RADEON_FIFO_DEBUG
 155 static void radeon_status(drm_radeon_private_t * dev_priv)
 156 {
 157         printk("%s:\n", __func__);
 158         printk("RBBM_STATUS = 0x%08x\n",
 159                (unsigned int)RADEON_READ(RADEON_RBBM_STATUS));
 160         printk("CP_RB_RTPR = 0x%08x\n",
 161                (unsigned int)RADEON_READ(RADEON_CP_RB_RPTR));
 162         printk("CP_RB_WTPR = 0x%08x\n",
 163                (unsigned int)RADEON_READ(RADEON_CP_RB_WPTR));
 164         printk("AIC_CNTL = 0x%08x\n",
 165                (unsigned int)RADEON_READ(RADEON_AIC_CNTL));
 166         printk("AIC_STAT = 0x%08x\n",
 167                (unsigned int)RADEON_READ(RADEON_AIC_STAT));
 168         printk("AIC_PT_BASE = 0x%08x\n",
 169                (unsigned int)RADEON_READ(RADEON_AIC_PT_BASE));
 170         printk("TLB_ADDR = 0x%08x\n",
 171                (unsigned int)RADEON_READ(RADEON_AIC_TLB_ADDR));
 172         printk("TLB_DATA = 0x%08x\n",
 173                (unsigned int)RADEON_READ(RADEON_AIC_TLB_DATA));
 174 }
 175 #endif
 176
 177 /* ================================================================
 178  * Engine, FIFO control
 179  */
 180
 181 static int radeon_do_pixcache_flush(drm_radeon_private_t * dev_priv)
 182 {
 183         u32 tmp;
 184         int i;
 185
 186         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 187
 188         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV280) {
 189                 tmp = RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT);
 190                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 191                 RADEON_WRITE(RADEON_RB3D_DSTCACHE_CTLSTAT, tmp);
 192
 193                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 194                         if (!(RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT)
 195                               & RADEON_RB3D_DC_BUSY)) {
 196                                 return 0;
 197                         }
 198                         DRM_UDELAY(1);
 199                 }
 200         } else {
 201                 /* 3D */
 202                 tmp = RADEON_READ(R300_RB3D_DSTCACHE_CTLSTAT);
 203                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 204                 RADEON_WRITE(R300_RB3D_DSTCACHE_CTLSTAT, tmp);
 205
 206                 /* 2D */
 207                 tmp = RADEON_READ(R300_DSTCACHE_CTLSTAT);
 208                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 209                 RADEON_WRITE(R300_DSTCACHE_CTLSTAT, tmp);
 210
 211                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 212                         if (!(RADEON_READ(R300_DSTCACHE_CTLSTAT)
 213                           & RADEON_RB3D_DC_BUSY)) {
 214                                 return 0;
 215                         }
 216                         DRM_UDELAY(1);
 217                 }
 218         }
 219
 220 #if RADEON_FIFO_DEBUG
 221         DRM_ERROR("failed!\n");
 222         radeon_status(dev_priv);
 223 #endif
 224         return -EBUSY;
 225 }
 226
 227 static int radeon_do_wait_for_fifo(drm_radeon_private_t * dev_priv, int entries)
 228 {
 229         int i;
 230
 231         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 232
 233         for (i = 0; i < dev_priv->usec_timeout; i++) {
 234                 int slots = (RADEON_READ(RADEON_RBBM_STATUS)
 235                              & RADEON_RBBM_FIFOCNT_MASK);
 236                 if (slots >= entries)
 237                         return 0;
 238                 DRM_UDELAY(1);
 239         }
 240
 241 #if RADEON_FIFO_DEBUG
 242         DRM_ERROR("failed!\n");
 243         radeon_status(dev_priv);
 244 #endif
 245         return -EBUSY;
 246 }
 247
 248 static int radeon_do_wait_for_idle(drm_radeon_private_t * dev_priv)
 249 {
 250         int i, ret;
 251
 252         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 253
 254         ret = radeon_do_wait_for_fifo(dev_priv, 64);
 255         if (ret)
 256                 return ret;
 257
 258         for (i = 0; i < dev_priv->usec_timeout; i++) {
 259                 if (!(RADEON_READ(RADEON_RBBM_STATUS)
 260                       & RADEON_RBBM_ACTIVE)) {
 261                         radeon_do_pixcache_flush(dev_priv);
 262                         return 0;
 263                 }
 264                 DRM_UDELAY(1);
 265         }
 266
 267 #if RADEON_FIFO_DEBUG
 268         DRM_ERROR("failed!\n");
 269         radeon_status(dev_priv);
 270 #endif
 271         return -EBUSY;
 272 }
 273
 274 static void radeon_init_pipes(drm_radeon_private_t *dev_priv)
 275 {
 276         uint32_t gb_tile_config, gb_pipe_sel = 0;
 277
 278         /* RS4xx/RS6xx/R4xx/R5xx */
 279         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R420) {
 280                 gb_pipe_sel = RADEON_READ(R400_GB_PIPE_SELECT);
 281                 dev_priv->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
 282         } else {
 283                 /* R3xx */
 284                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 285                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350)) {
 286                         dev_priv->num_gb_pipes = 2;
 287                 } else {
 288                         /* R3Vxx */
 289                         dev_priv->num_gb_pipes = 1;
 290                 }
 291         }
 292         DRM_INFO("Num pipes: %d\n", dev_priv->num_gb_pipes);
 293
 294         gb_tile_config = (R300_ENABLE_TILING | R300_TILE_SIZE_16 /*| R300_SUBPIXEL_1_16*/);
 295
 296         switch (dev_priv->num_gb_pipes) {
 297         case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
 298         case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
 299         case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
 300         default:
 301         case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
 302         }
 303
 304         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
 305                 RADEON_WRITE_PLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
 306                 RADEON_WRITE(R500_SU_REG_DEST, ((1 << dev_priv->num_gb_pipes) - 1));
 307         }
 308         RADEON_WRITE(R300_GB_TILE_CONFIG, gb_tile_config);
 309         radeon_do_wait_for_idle(dev_priv);
 310         RADEON_WRITE(R300_DST_PIPE_CONFIG, RADEON_READ(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
 311         RADEON_WRITE(R300_RB2D_DSTCACHE_MODE, (RADEON_READ(R300_RB2D_DSTCACHE_MODE) |
 312                                                R300_DC_AUTOFLUSH_ENABLE |
 313                                                R300_DC_DC_DISABLE_IGNORE_PE));
 314
 315
 316 }
 317
 318 /* ================================================================
 319  * CP control, initialization
 320  */
 321
 322 /* Load the microcode for the CP */
 323 static void radeon_cp_load_microcode(drm_radeon_private_t * dev_priv)
 324 {
 325         int i;
 326         DRM_DEBUG("\n");
 327
 328         radeon_do_wait_for_idle(dev_priv);
 329
 330         RADEON_WRITE(RADEON_CP_ME_RAM_ADDR, 0);
 331         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R100) ||
 332             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV100) ||
 333             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV200) ||
 334             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS100) ||
 335             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS200)) {
 336                 DRM_INFO("Loading R100 Microcode\n");
 337                 for (i = 0; i < 256; i++) {
 338                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 339                                      R100_cp_microcode[i][1]);
 340                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 341                                      R100_cp_microcode[i][0]);
 342                 }
 343         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R200) ||
 344                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV250) ||
 345                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV280) ||
 346                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS300)) {
 347                 DRM_INFO("Loading R200 Microcode\n");
 348                 for (i = 0; i < 256; i++) {
 349                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 350                                      R200_cp_microcode[i][1]);
 351                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 352                                      R200_cp_microcode[i][0]);
 353                 }
 354         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 355                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350) ||
 356                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV350) ||
 357                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV380) ||
 358                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 359                 DRM_INFO("Loading R300 Microcode\n");
 360                 for (i = 0; i < 256; i++) {
 361                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 362                                      R300_cp_microcode[i][1]);
 363                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 364                                      R300_cp_microcode[i][0]);
 365                 }
 366         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 367                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV410)) {
 368                 DRM_INFO("Loading R400 Microcode\n");
 369                 for (i = 0; i < 256; i++) {
 370                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 371                                      R420_cp_microcode[i][1]);
 372                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 373                                      R420_cp_microcode[i][0]);
 374                 }
 375         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) {
 376                 DRM_INFO("Loading RS690 Microcode\n");
 377                 for (i = 0; i < 256; i++) {
 378                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 379                                      RS690_cp_microcode[i][1]);
 380                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 381                                      RS690_cp_microcode[i][0]);
 382                 }
 383         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) ||
 384                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R520) ||
 385                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530) ||
 386                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R580) ||
 387                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV560) ||
 388                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV570)) {
 389                 DRM_INFO("Loading R500 Microcode\n");
 390                 for (i = 0; i < 256; i++) {
 391                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 392                                      R520_cp_microcode[i][1]);
 393                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 394                                      R520_cp_microcode[i][0]);
 395                 }
 396         }
 397 }
 398
 399 /* Flush any pending commands to the CP.  This should only be used just
 400  * prior to a wait for idle, as it informs the engine that the command
 401  * stream is ending.
 402  */
 403 static void radeon_do_cp_flush(drm_radeon_private_t * dev_priv)
 404 {
 405         DRM_DEBUG("\n");
 406 #if 0
 407         u32 tmp;
 408
 409         tmp = RADEON_READ(RADEON_CP_RB_WPTR) | (1 << 31);
 410         RADEON_WRITE(RADEON_CP_RB_WPTR, tmp);
 411 #endif
 412 }
 413
 414 /* Wait for the CP to go idle.
 415  */
 416 int radeon_do_cp_idle(drm_radeon_private_t * dev_priv)
 417 {
 418         RING_LOCALS;
 419         DRM_DEBUG("\n");
 420
 421         BEGIN_RING(6);
 422
 423         RADEON_PURGE_CACHE();
 424         RADEON_PURGE_ZCACHE();
 425         RADEON_WAIT_UNTIL_IDLE();
 426
 427         ADVANCE_RING();
 428         COMMIT_RING();
 429
 430         return radeon_do_wait_for_idle(dev_priv);
 431 }
 432
 433 /* Start the Command Processor.
 434  */
 435 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 436 {
 437         RING_LOCALS;
 438         DRM_DEBUG("\n");
 439
 440         radeon_do_wait_for_idle(dev_priv);
 441
 442         RADEON_WRITE(RADEON_CP_CSQ_CNTL, dev_priv->cp_mode);
 443
 444         dev_priv->cp_running = 1;
 445
 446         BEGIN_RING(6);
 447
 448         RADEON_PURGE_CACHE();
 449         RADEON_PURGE_ZCACHE();
 450         RADEON_WAIT_UNTIL_IDLE();
 451
 452         ADVANCE_RING();
 453         COMMIT_RING();
 454 }
 455
 456 /* Reset the Command Processor.  This will not flush any pending
 457  * commands, so you must wait for the CP command stream to complete
 458  * before calling this routine.
 459  */
 460 static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
 461 {
 462         u32 cur_read_ptr;
 463         DRM_DEBUG("\n");
 464
 465         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 466         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 467         SET_RING_HEAD(dev_priv, cur_read_ptr);
 468         dev_priv->ring.tail = cur_read_ptr;
 469 }
 470
 471 /* Stop the Command Processor.  This will not flush any pending
 472  * commands, so you must flush the command stream and wait for the CP
 473  * to go idle before calling this routine.
 474  */
 475 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 476 {
 477         DRM_DEBUG("\n");
 478
 479         RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 480
 481         dev_priv->cp_running = 0;
 482 }
 483
 484 /* Reset the engine.  This will stop the CP if it is running.
 485  */
 486 static int radeon_do_engine_reset(struct drm_device * dev)
 487 {
 488         drm_radeon_private_t *dev_priv = dev->dev_private;
 489         u32 clock_cntl_index = 0, mclk_cntl = 0, rbbm_soft_reset;
 490         DRM_DEBUG("\n");
 491
 492         radeon_do_pixcache_flush(dev_priv);
 493
 494         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 495                 /* may need something similar for newer chips */
 496                 clock_cntl_index = RADEON_READ(RADEON_CLOCK_CNTL_INDEX);
 497                 mclk_cntl = RADEON_READ_PLL(dev, RADEON_MCLK_CNTL);
 498
 499                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, (mclk_cntl |
 500                                                     RADEON_FORCEON_MCLKA |
 501                                                     RADEON_FORCEON_MCLKB |
 502                                                     RADEON_FORCEON_YCLKA |
 503                                                     RADEON_FORCEON_YCLKB |
 504                                                     RADEON_FORCEON_MC |
 505                                                     RADEON_FORCEON_AIC));
 506         }
 507
 508         rbbm_soft_reset = RADEON_READ(RADEON_RBBM_SOFT_RESET);
 509
 510         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset |
 511                                               RADEON_SOFT_RESET_CP |
 512                                               RADEON_SOFT_RESET_HI |
 513                                               RADEON_SOFT_RESET_SE |
 514                                               RADEON_SOFT_RESET_RE |
 515                                               RADEON_SOFT_RESET_PP |
 516                                               RADEON_SOFT_RESET_E2 |
 517                                               RADEON_SOFT_RESET_RB));
 518         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 519         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset &
 520                                               ~(RADEON_SOFT_RESET_CP |
 521                                                 RADEON_SOFT_RESET_HI |
 522                                                 RADEON_SOFT_RESET_SE |
 523                                                 RADEON_SOFT_RESET_RE |
 524                                                 RADEON_SOFT_RESET_PP |
 525                                                 RADEON_SOFT_RESET_E2 |
 526                                                 RADEON_SOFT_RESET_RB)));
 527         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 528
 529         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 530                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, mclk_cntl);
 531                 RADEON_WRITE(RADEON_CLOCK_CNTL_INDEX, clock_cntl_index);
 532                 RADEON_WRITE(RADEON_RBBM_SOFT_RESET, rbbm_soft_reset);
 533         }
 534
 535         /* setup the raster pipes */
 536         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R300)
 537             radeon_init_pipes(dev_priv);
 538
 539         /* Reset the CP ring */
 540         radeon_do_cp_reset(dev_priv);
 541
 542         /* The CP is no longer running after an engine reset */
 543         dev_priv->cp_running = 0;
 544
 545         /* Reset any pending vertex, indirect buffers */
 546         radeon_freelist_reset(dev);
 547
 548         return 0;
 549 }
 550
 551 static void radeon_cp_init_ring_buffer(struct drm_device * dev,
 552                                        drm_radeon_private_t * dev_priv)
 553 {
 554         u32 ring_start, cur_read_ptr;
 555         u32 tmp;
 556
 557         /* Initialize the memory controller. With new memory map, the fb location
 558          * is not changed, it should have been properly initialized already. Part
 559          * of the problem is that the code below is bogus, assuming the GART is
 560          * always appended to the fb which is not necessarily the case
 561          */
 562         if (!dev_priv->new_memmap)
 563                 radeon_write_fb_location(dev_priv,
 564                              ((dev_priv->gart_vm_start - 1) & 0xffff0000)
 565                              | (dev_priv->fb_location >> 16));
 566
 567 #if __OS_HAS_AGP
 568         if (dev_priv->flags & RADEON_IS_AGP) {
 569                 radeon_write_agp_base(dev_priv, dev->agp->base);
 570
 571                 radeon_write_agp_location(dev_priv,
 572                              (((dev_priv->gart_vm_start - 1 +
 573                                 dev_priv->gart_size) & 0xffff0000) |
 574                               (dev_priv->gart_vm_start >> 16)));
 575
 576                 ring_start = (dev_priv->cp_ring->offset
 577                               - dev->agp->base
 578                               + dev_priv->gart_vm_start);
 579         } else
 580 #endif
 581                 ring_start = (dev_priv->cp_ring->offset
 582                               - (unsigned long)dev->sg->virtual
 583                               + dev_priv->gart_vm_start);
 584
 585         RADEON_WRITE(RADEON_CP_RB_BASE, ring_start);
 586
 587         /* Set the write pointer delay */
 588         RADEON_WRITE(RADEON_CP_RB_WPTR_DELAY, 0);
 589
 590         /* Initialize the ring buffer's read and write pointers */
 591         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 592         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 593         SET_RING_HEAD(dev_priv, cur_read_ptr);
 594         dev_priv->ring.tail = cur_read_ptr;
 595
 596 #if __OS_HAS_AGP
 597         if (dev_priv->flags & RADEON_IS_AGP) {
 598                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 599                              dev_priv->ring_rptr->offset
 600                              - dev->agp->base + dev_priv->gart_vm_start);
 601         } else
 602 #endif
 603         {
 604                 struct drm_sg_mem *entry = dev->sg;
 605                 unsigned long tmp_ofs, page_ofs;
 606
 607                 tmp_ofs = dev_priv->ring_rptr->offset -
 608                                 (unsigned long)dev->sg->virtual;
 609                 page_ofs = tmp_ofs >> PAGE_SHIFT;
 610
 611                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR, entry->busaddr[page_ofs]);
 612                 DRM_DEBUG("ring rptr: offset=0x%08lx handle=0x%08lx\n",
 613                           (unsigned long)entry->busaddr[page_ofs],
 614                           entry->handle + tmp_ofs);
 615         }
 616
 617         /* Set ring buffer size */
 618 #ifdef __BIG_ENDIAN
 619         RADEON_WRITE(RADEON_CP_RB_CNTL,
 620                      RADEON_BUF_SWAP_32BIT |
 621                      (dev_priv->ring.fetch_size_l2ow << 18) |
 622                      (dev_priv->ring.rptr_update_l2qw << 8) |
 623                      dev_priv->ring.size_l2qw);
 624 #else
 625         RADEON_WRITE(RADEON_CP_RB_CNTL,
 626                      (dev_priv->ring.fetch_size_l2ow << 18) |
 627                      (dev_priv->ring.rptr_update_l2qw << 8) |
 628                      dev_priv->ring.size_l2qw);
 629 #endif
 630
 631         /* Start with assuming that writeback doesn't work */
 632         dev_priv->writeback_works = 0;
 633
 634         /* Initialize the scratch register pointer.  This will cause
 635          * the scratch register values to be written out to memory
 636          * whenever they are updated.
 637          *
 638          * We simply put this behind the ring read pointer, this works
 639          * with PCI GART as well as (whatever kind of) AGP GART
 640          */
 641         RADEON_WRITE(RADEON_SCRATCH_ADDR, RADEON_READ(RADEON_CP_RB_RPTR_ADDR)
 642                      + RADEON_SCRATCH_REG_OFFSET);
 643
 644         dev_priv->scratch = ((__volatile__ u32 *)
 645                              dev_priv->ring_rptr->handle +
 646                              (RADEON_SCRATCH_REG_OFFSET / sizeof(u32)));
 647
 648         RADEON_WRITE(RADEON_SCRATCH_UMSK, 0x7);
 649
 650         /* Turn on bus mastering */
 651         tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RADEON_BUS_MASTER_DIS;
 652         RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 653
 654         dev_priv->sarea_priv->last_frame = dev_priv->scratch[0] = 0;
 655         RADEON_WRITE(RADEON_LAST_FRAME_REG, dev_priv->sarea_priv->last_frame);
 656
 657         dev_priv->sarea_priv->last_dispatch = dev_priv->scratch[1] = 0;
 658         RADEON_WRITE(RADEON_LAST_DISPATCH_REG,
 659                      dev_priv->sarea_priv->last_dispatch);
 660
 661         dev_priv->sarea_priv->last_clear = dev_priv->scratch[2] = 0;
 662         RADEON_WRITE(RADEON_LAST_CLEAR_REG, dev_priv->sarea_priv->last_clear);
 663
 664         radeon_do_wait_for_idle(dev_priv);
 665
 666         /* Sync everything up */
 667         RADEON_WRITE(RADEON_ISYNC_CNTL,
 668                      (RADEON_ISYNC_ANY2D_IDLE3D |
 669                       RADEON_ISYNC_ANY3D_IDLE2D |
 670                       RADEON_ISYNC_WAIT_IDLEGUI |
 671                       RADEON_ISYNC_CPSCRATCH_IDLEGUI));
 672
 673 }
 674
 675 static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 676 {
 677         u32 tmp;
 678
 679         /* Writeback doesn't seem to work everywhere, test it here and possibly
 680          * enable it if it appears to work
 681          */
 682         DRM_WRITE32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1), 0);
 683         RADEON_WRITE(RADEON_SCRATCH_REG1, 0xdeadbeef);
 684
 685         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 686                 if (DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1)) ==
 687                     0xdeadbeef)
 688                         break;
 689                 DRM_UDELAY(1);
 690         }
 691
 692         if (tmp < dev_priv->usec_timeout) {
 693                 dev_priv->writeback_works = 1;
 694                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 695         } else {
 696                 dev_priv->writeback_works = 0;
 697                 DRM_INFO("writeback test failed\n");
 698         }
 699         if (radeon_no_wb == 1) {
 700                 dev_priv->writeback_works = 0;
 701                 DRM_INFO("writeback forced off\n");
 702         }
 703
 704         if (!dev_priv->writeback_works) {
 705                 /* Disable writeback to avoid unnecessary bus master transfer */
 706                 RADEON_WRITE(RADEON_CP_RB_CNTL, RADEON_READ(RADEON_CP_RB_CNTL) |
 707                              RADEON_RB_NO_UPDATE);
 708                 RADEON_WRITE(RADEON_SCRATCH_UMSK, 0);
 709         }
 710 }
 711
 712 /* Enable or disable IGP GART on the chip */
 713 static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
 714 {
 715         u32 temp;
 716
 717         if (on) {
 718                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 719                           dev_priv->gart_vm_start,
 720                           (long)dev_priv->gart_info.bus_addr,
 721                           dev_priv->gart_size);
 722
 723                 temp = IGP_READ_MCIND(dev_priv, RS480_MC_MISC_CNTL);
 724                 if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690)
 725                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, (RS480_GART_INDEX_REG_EN |
 726                                                              RS690_BLOCK_GFX_D3_EN));
 727                 else
 728                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, RS480_GART_INDEX_REG_EN);
 729
 730                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 731                                                                RS480_VA_SIZE_32MB));
 732
 733                 temp = IGP_READ_MCIND(dev_priv, RS480_GART_FEATURE_ID);
 734                 IGP_WRITE_MCIND(RS480_GART_FEATURE_ID, (RS480_HANG_EN |
 735                                                         RS480_TLB_ENABLE |
 736                                                         RS480_GTW_LAC_EN |
 737                                                         RS480_1LEVEL_GART));
 738
 739                 temp = dev_priv->gart_info.bus_addr & 0xfffff000;
 740                 temp |= (upper_32_bits(dev_priv->gart_info.bus_addr) & 0xff) << 4;
 741                 IGP_WRITE_MCIND(RS480_GART_BASE, temp);
 742
 743                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_MODE_CNTL);
 744                 IGP_WRITE_MCIND(RS480_AGP_MODE_CNTL, ((1 << RS480_REQ_TYPE_SNOOP_SHIFT) |
 745                                                       RS480_REQ_TYPE_SNOOP_DIS));
 746
 747                 radeon_write_agp_base(dev_priv, dev_priv->gart_vm_start);
 748
 749                 dev_priv->gart_size = 32*1024*1024;
 750                 temp = (((dev_priv->gart_vm_start - 1 + dev_priv->gart_size) &
 751                          0xffff0000) | (dev_priv->gart_vm_start >> 16));
 752
 753                 radeon_write_agp_location(dev_priv, temp);
 754
 755                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_ADDRESS_SPACE_SIZE);
 756                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 757                                                                RS480_VA_SIZE_32MB));
 758
 759                 do {
 760                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 761                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 762                                 break;
 763                         DRM_UDELAY(1);
 764                 } while (1);
 765
 766                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL,
 767                                 RS480_GART_CACHE_INVALIDATE);
 768
 769                 do {
 770                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 771                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 772                                 break;
 773                         DRM_UDELAY(1);
 774                 } while (1);
 775
 776                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL, 0);
 777         } else {
 778                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, 0);
 779         }
 780 }
 781
 782 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
 783 {
 784         u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
 785         if (on) {
 786
 787                 DRM_DEBUG("programming pcie %08X %08lX %08X\n",
 788                           dev_priv->gart_vm_start,
 789                           (long)dev_priv->gart_info.bus_addr,
 790                           dev_priv->gart_size);
 791                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_DISCARD_RD_ADDR_LO,
 792                                   dev_priv->gart_vm_start);
 793                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_BASE,
 794                                   dev_priv->gart_info.bus_addr);
 795                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_START_LO,
 796                                   dev_priv->gart_vm_start);
 797                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_END_LO,
 798                                   dev_priv->gart_vm_start +
 799                                   dev_priv->gart_size - 1);
 800
 801                 radeon_write_agp_location(dev_priv, 0xffffffc0); /* ?? */
 802
 803                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 804                                   RADEON_PCIE_TX_GART_EN);
 805         } else {
 806                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 807                                   tmp & ~RADEON_PCIE_TX_GART_EN);
 808         }
 809 }
 810
 811 /* Enable or disable PCI GART on the chip */
 812 static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
 813 {
 814         u32 tmp;
 815
 816         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 817             (dev_priv->flags & RADEON_IS_IGPGART)) {
 818                 radeon_set_igpgart(dev_priv, on);
 819                 return;
 820         }
 821
 822         if (dev_priv->flags & RADEON_IS_PCIE) {
 823                 radeon_set_pciegart(dev_priv, on);
 824                 return;
 825         }
 826
 827         tmp = RADEON_READ(RADEON_AIC_CNTL);
 828
 829         if (on) {
 830                 RADEON_WRITE(RADEON_AIC_CNTL,
 831                              tmp | RADEON_PCIGART_TRANSLATE_EN);
 832
 833                 /* set PCI GART page-table base address
 834                  */
 835                 RADEON_WRITE(RADEON_AIC_PT_BASE, dev_priv->gart_info.bus_addr);
 836
 837                 /* set address range for PCI address translate
 838                  */
 839                 RADEON_WRITE(RADEON_AIC_LO_ADDR, dev_priv->gart_vm_start);
 840                 RADEON_WRITE(RADEON_AIC_HI_ADDR, dev_priv->gart_vm_start
 841                              + dev_priv->gart_size - 1);
 842
 843                 /* Turn off AGP aperture -- is this required for PCI GART?
 844                  */
 845                 radeon_write_agp_location(dev_priv, 0xffffffc0);
 846                 RADEON_WRITE(RADEON_AGP_COMMAND, 0);    /* clear AGP_COMMAND */
 847         } else {
 848                 RADEON_WRITE(RADEON_AIC_CNTL,
 849                              tmp & ~RADEON_PCIGART_TRANSLATE_EN);
 850         }
 851 }
 852
 853 static int radeon_do_init_cp(struct drm_device * dev, drm_radeon_init_t * init)
 854 {
 855         drm_radeon_private_t *dev_priv = dev->dev_private;
 856
 857         DRM_DEBUG("\n");
 858
 859         /* if we require new memory map but we don't have it fail */
 860         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
 861                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
 862                 radeon_do_cleanup_cp(dev);
 863                 return -EINVAL;
 864         }
 865
 866         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
 867                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
 868                 dev_priv->flags &= ~RADEON_IS_AGP;
 869         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
 870                    && !init->is_pci) {
 871                 DRM_DEBUG("Restoring AGP flag\n");
 872                 dev_priv->flags |= RADEON_IS_AGP;
 873         }
 874
 875         if ((!(dev_priv->flags & RADEON_IS_AGP)) && !dev->sg) {
 876                 DRM_ERROR("PCI GART memory not allocated!\n");
 877                 radeon_do_cleanup_cp(dev);
 878                 return -EINVAL;
 879         }
 880
 881         dev_priv->usec_timeout = init->usec_timeout;
 882         if (dev_priv->usec_timeout < 1 ||
 883             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
 884                 DRM_DEBUG("TIMEOUT problem!\n");
 885                 radeon_do_cleanup_cp(dev);
 886                 return -EINVAL;
 887         }
 888
 889         /* Enable vblank on CRTC1 for older X servers
 890          */
 891         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
 892
 893         switch(init->func) {
 894         case RADEON_INIT_R200_CP:
 895                 dev_priv->microcode_version = UCODE_R200;
 896                 break;
 897         case RADEON_INIT_R300_CP:
 898                 dev_priv->microcode_version = UCODE_R300;
 899                 break;
 900         default:
 901                 dev_priv->microcode_version = UCODE_R100;
 902         }
 903
 904         dev_priv->do_boxes = 0;
 905         dev_priv->cp_mode = init->cp_mode;
 906
 907         /* We don't support anything other than bus-mastering ring mode,
 908          * but the ring can be in either AGP or PCI space for the ring
 909          * read pointer.
 910          */
 911         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
 912             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
 913                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
 914                 radeon_do_cleanup_cp(dev);
 915                 return -EINVAL;
 916         }
 917
 918         switch (init->fb_bpp) {
 919         case 16:
 920                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
 921                 break;
 922         case 32:
 923         default:
 924                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
 925                 break;
 926         }
 927         dev_priv->front_offset = init->front_offset;
 928         dev_priv->front_pitch = init->front_pitch;
 929         dev_priv->back_offset = init->back_offset;
 930         dev_priv->back_pitch = init->back_pitch;
 931
 932         switch (init->depth_bpp) {
 933         case 16:
 934                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_16BIT_INT_Z;
 935                 break;
 936         case 32:
 937         default:
 938                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_24BIT_INT_Z;
 939                 break;
 940         }
 941         dev_priv->depth_offset = init->depth_offset;
 942         dev_priv->depth_pitch = init->depth_pitch;
 943
 944         /* Hardware state for depth clears.  Remove this if/when we no
 945          * longer clear the depth buffer with a 3D rectangle.  Hard-code
 946          * all values to prevent unwanted 3D state from slipping through
 947          * and screwing with the clear operation.
 948          */
 949         dev_priv->depth_clear.rb3d_cntl = (RADEON_PLANE_MASK_ENABLE |
 950                                            (dev_priv->color_fmt << 10) |
 951                                            (dev_priv->microcode_version ==
 952                                             UCODE_R100 ? RADEON_ZBLOCK16 : 0));
 953
 954         dev_priv->depth_clear.rb3d_zstencilcntl =
 955             (dev_priv->depth_fmt |
 956              RADEON_Z_TEST_ALWAYS |
 957              RADEON_STENCIL_TEST_ALWAYS |
 958              RADEON_STENCIL_S_FAIL_REPLACE |
 959              RADEON_STENCIL_ZPASS_REPLACE |
 960              RADEON_STENCIL_ZFAIL_REPLACE | RADEON_Z_WRITE_ENABLE);
 961
 962         dev_priv->depth_clear.se_cntl = (RADEON_FFACE_CULL_CW |
 963                                          RADEON_BFACE_SOLID |
 964                                          RADEON_FFACE_SOLID |
 965                                          RADEON_FLAT_SHADE_VTX_LAST |
 966                                          RADEON_DIFFUSE_SHADE_FLAT |
 967                                          RADEON_ALPHA_SHADE_FLAT |
 968                                          RADEON_SPECULAR_SHADE_FLAT |
 969                                          RADEON_FOG_SHADE_FLAT |
 970                                          RADEON_VTX_PIX_CENTER_OGL |
 971                                          RADEON_ROUND_MODE_TRUNC |
 972                                          RADEON_ROUND_PREC_8TH_PIX);
 973
 974
 975         dev_priv->ring_offset = init->ring_offset;
 976         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
 977         dev_priv->buffers_offset = init->buffers_offset;
 978         dev_priv->gart_textures_offset = init->gart_textures_offset;
 979
 980         dev_priv->sarea = drm_getsarea(dev);
 981         if (!dev_priv->sarea) {
 982                 DRM_ERROR("could not find sarea!\n");
 983                 radeon_do_cleanup_cp(dev);
 984                 return -EINVAL;
 985         }
 986
 987         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
 988         if (!dev_priv->cp_ring) {
 989                 DRM_ERROR("could not find cp ring region!\n");
 990                 radeon_do_cleanup_cp(dev);
 991                 return -EINVAL;
 992         }
 993         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
 994         if (!dev_priv->ring_rptr) {
 995                 DRM_ERROR("could not find ring read pointer!\n");
 996                 radeon_do_cleanup_cp(dev);
 997                 return -EINVAL;
 998         }
 999         dev->agp_buffer_token = init->buffers_offset;
1000         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1001         if (!dev->agp_buffer_map) {
1002                 DRM_ERROR("could not find dma buffer region!\n");
1003                 radeon_do_cleanup_cp(dev);
1004                 return -EINVAL;
1005         }
1006
1007         if (init->gart_textures_offset) {
1008                 dev_priv->gart_textures =
1009                     drm_core_findmap(dev, init->gart_textures_offset);
1010                 if (!dev_priv->gart_textures) {
1011                         DRM_ERROR("could not find GART texture region!\n");
1012                         radeon_do_cleanup_cp(dev);
1013                         return -EINVAL;
1014                 }
1015         }
1016
1017         dev_priv->sarea_priv =
1018             (drm_radeon_sarea_t *) ((u8 *) dev_priv->sarea->handle +
1019                                     init->sarea_priv_offset);
1020
1021 #if __OS_HAS_AGP
1022         if (dev_priv->flags & RADEON_IS_AGP) {
1023                 drm_core_ioremap(dev_priv->cp_ring, dev);
1024                 drm_core_ioremap(dev_priv->ring_rptr, dev);
1025                 drm_core_ioremap(dev->agp_buffer_map, dev);
1026                 if (!dev_priv->cp_ring->handle ||
1027                     !dev_priv->ring_rptr->handle ||
1028                     !dev->agp_buffer_map->handle) {
1029                         DRM_ERROR("could not find ioremap agp regions!\n");
1030                         radeon_do_cleanup_cp(dev);
1031                         return -EINVAL;
1032                 }
1033         } else
1034 #endif
1035         {
1036                 dev_priv->cp_ring->handle = (void *)dev_priv->cp_ring->offset;
1037                 dev_priv->ring_rptr->handle =
1038                     (void *)dev_priv->ring_rptr->offset;
1039                 dev->agp_buffer_map->handle =
1040                     (void *)dev->agp_buffer_map->offset;
1041
1042                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
1043                           dev_priv->cp_ring->handle);
1044                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
1045                           dev_priv->ring_rptr->handle);
1046                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
1047                           dev->agp_buffer_map->handle);
1048         }
1049
1050         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 16;
1051         dev_priv->fb_size =
1052                 ((radeon_read_fb_location(dev_priv) & 0xffff0000u) + 0x10000)
1053                 - dev_priv->fb_location;
1054
1055         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
1056                                         ((dev_priv->front_offset
1057                                           + dev_priv->fb_location) >> 10));
1058
1059         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
1060                                        ((dev_priv->back_offset
1061                                          + dev_priv->fb_location) >> 10));
1062
1063         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
1064                                         ((dev_priv->depth_offset
1065                                           + dev_priv->fb_location) >> 10));
1066
1067         dev_priv->gart_size = init->gart_size;
1068
1069         /* New let's set the memory map ... */
1070         if (dev_priv->new_memmap) {
1071                 u32 base = 0;
1072
1073                 DRM_INFO("Setting GART location based on new memory map\n");
1074
1075                 /* If using AGP, try to locate the AGP aperture at the same
1076                  * location in the card and on the bus, though we have to
1077                  * align it down.
1078                  */
1079 #if __OS_HAS_AGP
1080                 if (dev_priv->flags & RADEON_IS_AGP) {
1081                         base = dev->agp->base;
1082                         /* Check if valid */
1083                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
1084                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
1085                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
1086                                          dev->agp->base);
1087                                 base = 0;
1088                         }
1089                 }
1090 #endif
1091                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
1092                 if (base == 0) {
1093                         base = dev_priv->fb_location + dev_priv->fb_size;
1094                         if (base < dev_priv->fb_location ||
1095                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
1096                                 base = dev_priv->fb_location
1097                                         - dev_priv->gart_size;
1098                 }
1099                 dev_priv->gart_vm_start = base & 0xffc00000u;
1100                 if (dev_priv->gart_vm_start != base)
1101                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
1102                                  base, dev_priv->gart_vm_start);
1103         } else {
1104                 DRM_INFO("Setting GART location based on old memory map\n");
1105                 dev_priv->gart_vm_start = dev_priv->fb_location +
1106                         RADEON_READ(RADEON_CONFIG_APER_SIZE);
1107         }
1108
1109 #if __OS_HAS_AGP
1110         if (dev_priv->flags & RADEON_IS_AGP)
1111                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1112                                                  - dev->agp->base
1113                                                  + dev_priv->gart_vm_start);
1114         else
1115 #endif
1116                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1117                                         - (unsigned long)dev->sg->virtual
1118                                         + dev_priv->gart_vm_start);
1119
1120         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
1121         DRM_DEBUG("dev_priv->gart_vm_start 0x%x\n", dev_priv->gart_vm_start);
1122         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%lx\n",
1123                   dev_priv->gart_buffers_offset);
1124
1125         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
1126         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
1127                               + init->ring_size / sizeof(u32));
1128         dev_priv->ring.size = init->ring_size;
1129         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
1130
1131         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
1132         dev_priv->ring.rptr_update_l2qw = drm_order( /* init->rptr_update */ 4096 / 8);
1133
1134         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
1135         dev_priv->ring.fetch_size_l2ow = drm_order( /* init->fetch_size */ 32 / 16);
1136         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
1137
1138         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
1139
1140 #if __OS_HAS_AGP
1141         if (dev_priv->flags & RADEON_IS_AGP) {
1142                 /* Turn off PCI GART */
1143                 radeon_set_pcigart(dev_priv, 0);
1144         } else
1145 #endif
1146         {
1147                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
1148                 /* if we have an offset set from userspace */
1149                 if (dev_priv->pcigart_offset_set) {
1150                         dev_priv->gart_info.bus_addr =
1151                             dev_priv->pcigart_offset + dev_priv->fb_location;
1152                         dev_priv->gart_info.mapping.offset =
1153                             dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
1154                         dev_priv->gart_info.mapping.size =
1155                             dev_priv->gart_info.table_size;
1156
1157                         drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
1158                         dev_priv->gart_info.addr =
1159                             dev_priv->gart_info.mapping.handle;
1160
1161                         if (dev_priv->flags & RADEON_IS_PCIE)
1162                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
1163                         else
1164                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1165                         dev_priv->gart_info.gart_table_location =
1166                             DRM_ATI_GART_FB;
1167
1168                         DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
1169                                   dev_priv->gart_info.addr,
1170                                   dev_priv->pcigart_offset);
1171                 } else {
1172                         if (dev_priv->flags & RADEON_IS_IGPGART)
1173                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
1174                         else
1175                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1176                         dev_priv->gart_info.gart_table_location =
1177                             DRM_ATI_GART_MAIN;
1178                         dev_priv->gart_info.addr = NULL;
1179                         dev_priv->gart_info.bus_addr = 0;
1180                         if (dev_priv->flags & RADEON_IS_PCIE) {
1181                                 DRM_ERROR
1182                                     ("Cannot use PCI Express without GART in FB memory\n");
1183                                 radeon_do_cleanup_cp(dev);
1184                                 return -EINVAL;
1185                         }
1186                 }
1187
1188                 if (!drm_ati_pcigart_init(dev, &dev_priv->gart_info)) {
1189                         DRM_ERROR("failed to init PCI GART!\n");
1190                         radeon_do_cleanup_cp(dev);
1191                         return -ENOMEM;
1192                 }
1193
1194                 /* Turn on PCI GART */
1195                 radeon_set_pcigart(dev_priv, 1);
1196         }
1197
1198         radeon_cp_load_microcode(dev_priv);
1199         radeon_cp_init_ring_buffer(dev, dev_priv);
1200
1201         dev_priv->last_buf = 0;
1202
1203         radeon_do_engine_reset(dev);
1204         radeon_test_writeback(dev_priv);
1205
1206         return 0;
1207 }
1208
1209 static int radeon_do_cleanup_cp(struct drm_device * dev)
1210 {
1211         drm_radeon_private_t *dev_priv = dev->dev_private;
1212         DRM_DEBUG("\n");
1213
1214         /* Make sure interrupts are disabled here because the uninstall ioctl
1215          * may not have been called from userspace and after dev_private
1216          * is freed, it's too late.
1217          */
1218         if (dev->irq_enabled)
1219                 drm_irq_uninstall(dev);
1220
1221 #if __OS_HAS_AGP
1222         if (dev_priv->flags & RADEON_IS_AGP) {
1223                 if (dev_priv->cp_ring != NULL) {
1224                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1225                         dev_priv->cp_ring = NULL;
1226                 }
1227                 if (dev_priv->ring_rptr != NULL) {
1228                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1229                         dev_priv->ring_rptr = NULL;
1230                 }
1231                 if (dev->agp_buffer_map != NULL) {
1232                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1233                         dev->agp_buffer_map = NULL;
1234                 }
1235         } else
1236 #endif
1237         {
1238
1239                 if (dev_priv->gart_info.bus_addr) {
1240                         /* Turn off PCI GART */
1241                         radeon_set_pcigart(dev_priv, 0);
1242                         if (!drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info))
1243                                 DRM_ERROR("failed to cleanup PCI GART!\n");
1244                 }
1245
1246                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB)
1247                 {
1248                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1249                         dev_priv->gart_info.addr = 0;
1250                 }
1251         }
1252         /* only clear to the start of flags */
1253         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1254
1255         return 0;
1256 }
1257
1258 /* This code will reinit the Radeon CP hardware after a resume from disc.
1259  * AFAIK, it would be very difficult to pickle the state at suspend time, so
1260  * here we make sure that all Radeon hardware initialisation is re-done without
1261  * affecting running applications.
1262  *
1263  * Charl P. Botha <http://cpbotha.net>
1264  */
1265 static int radeon_do_resume_cp(struct drm_device * dev)
1266 {
1267         drm_radeon_private_t *dev_priv = dev->dev_private;
1268
1269         if (!dev_priv) {
1270                 DRM_ERROR("Called with no initialization\n");
1271                 return -EINVAL;
1272         }
1273
1274         DRM_DEBUG("Starting radeon_do_resume_cp()\n");
1275
1276 #if __OS_HAS_AGP
1277         if (dev_priv->flags & RADEON_IS_AGP) {
1278                 /* Turn off PCI GART */
1279                 radeon_set_pcigart(dev_priv, 0);
1280         } else
1281 #endif
1282         {
1283                 /* Turn on PCI GART */
1284                 radeon_set_pcigart(dev_priv, 1);
1285         }
1286
1287         radeon_cp_load_microcode(dev_priv);
1288         radeon_cp_init_ring_buffer(dev, dev_priv);
1289
1290         radeon_do_engine_reset(dev);
1291         radeon_enable_interrupt(dev);
1292
1293         DRM_DEBUG("radeon_do_resume_cp() complete\n");
1294
1295         return 0;
1296 }
1297
1298 int radeon_cp_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
1299 {
1300         drm_radeon_init_t *init = data;
1301
1302         LOCK_TEST_WITH_RETURN(dev, file_priv);
1303
1304         if (init->func == RADEON_INIT_R300_CP)
1305                 r300_init_reg_flags(dev);
1306
1307         switch (init->func) {
1308         case RADEON_INIT_CP:
1309         case RADEON_INIT_R200_CP:
1310         case RADEON_INIT_R300_CP:
1311                 return radeon_do_init_cp(dev, init);
1312         case RADEON_CLEANUP_CP:
1313                 return radeon_do_cleanup_cp(dev);
1314         }
1315
1316         return -EINVAL;
1317 }
1318
1319 int radeon_cp_start(struct drm_device *dev, void *data, struct drm_file *file_priv)
1320 {
1321         drm_radeon_private_t *dev_priv = dev->dev_private;
1322         DRM_DEBUG("\n");
1323
1324         LOCK_TEST_WITH_RETURN(dev, file_priv);
1325
1326         if (dev_priv->cp_running) {
1327                 DRM_DEBUG("while CP running\n");
1328                 return 0;
1329         }
1330         if (dev_priv->cp_mode == RADEON_CSQ_PRIDIS_INDDIS) {
1331                 DRM_DEBUG("called with bogus CP mode (%d)\n",
1332                           dev_priv->cp_mode);
1333                 return 0;
1334         }
1335
1336         radeon_do_cp_start(dev_priv);
1337
1338         return 0;
1339 }
1340
1341 /* Stop the CP.  The engine must have been idled before calling this
1342  * routine.
1343  */
1344 int radeon_cp_stop(struct drm_device *dev, void *data, struct drm_file *file_priv)
1345 {
1346         drm_radeon_private_t *dev_priv = dev->dev_private;
1347         drm_radeon_cp_stop_t *stop = data;
1348         int ret;
1349         DRM_DEBUG("\n");
1350
1351         LOCK_TEST_WITH_RETURN(dev, file_priv);
1352
1353         if (!dev_priv->cp_running)
1354                 return 0;
1355
1356         /* Flush any pending CP commands.  This ensures any outstanding
1357          * commands are exectuted by the engine before we turn it off.
1358          */
1359         if (stop->flush) {
1360                 radeon_do_cp_flush(dev_priv);
1361         }
1362
1363         /* If we fail to make the engine go idle, we return an error
1364          * code so that the DRM ioctl wrapper can try again.
1365          */
1366         if (stop->idle) {
1367                 ret = radeon_do_cp_idle(dev_priv);
1368                 if (ret)
1369                         return ret;
1370         }
1371
1372         /* Finally, we can turn off the CP.  If the engine isn't idle,
1373          * we will get some dropped triangles as they won't be fully
1374          * rendered before the CP is shut down.
1375          */
1376         radeon_do_cp_stop(dev_priv);
1377
1378         /* Reset the engine */
1379         radeon_do_engine_reset(dev);
1380
1381         return 0;
1382 }
1383
1384 void radeon_do_release(struct drm_device * dev)
1385 {
1386         drm_radeon_private_t *dev_priv = dev->dev_private;
1387         int i, ret;
1388
1389         if (dev_priv) {
1390                 if (dev_priv->cp_running) {
1391                         /* Stop the cp */
1392                         while ((ret = radeon_do_cp_idle(dev_priv)) != 0) {
1393                                 DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1394 #ifdef __linux__
1395                                 schedule();
1396 #else
1397                                 tsleep(&ret, PZERO, "rdnrel", 1);
1398 #endif
1399                         }
1400                         radeon_do_cp_stop(dev_priv);
1401                         radeon_do_engine_reset(dev);
1402                 }
1403
1404                 /* Disable *all* interrupts */
1405                 if (dev_priv->mmio)     /* remove this after permanent addmaps */
1406                         RADEON_WRITE(RADEON_GEN_INT_CNTL, 0);
1407
1408                 if (dev_priv->mmio) {   /* remove all surfaces */
1409                         for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1410                                 RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, 0);
1411                                 RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND +
1412                                              16 * i, 0);
1413                                 RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND +
1414                                              16 * i, 0);
1415                         }
1416                 }
1417
1418                 /* Free memory heap structures */
1419                 radeon_mem_takedown(&(dev_priv->gart_heap));
1420                 radeon_mem_takedown(&(dev_priv->fb_heap));
1421
1422                 /* deallocate kernel resources */
1423                 radeon_do_cleanup_cp(dev);
1424         }
1425 }
1426
1427 /* Just reset the CP ring.  Called as part of an X Server engine reset.
1428  */
1429 int radeon_cp_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1430 {
1431         drm_radeon_private_t *dev_priv = dev->dev_private;
1432         DRM_DEBUG("\n");
1433
1434         LOCK_TEST_WITH_RETURN(dev, file_priv);
1435
1436         if (!dev_priv) {
1437                 DRM_DEBUG("called before init done\n");
1438                 return -EINVAL;
1439         }
1440
1441         radeon_do_cp_reset(dev_priv);
1442
1443         /* The CP is no longer running after an engine reset */
1444         dev_priv->cp_running = 0;
1445
1446         return 0;
1447 }
1448
1449 int radeon_cp_idle(struct drm_device *dev, void *data, struct drm_file *file_priv)
1450 {
1451         drm_radeon_private_t *dev_priv = dev->dev_private;
1452         DRM_DEBUG("\n");
1453
1454         LOCK_TEST_WITH_RETURN(dev, file_priv);
1455
1456         return radeon_do_cp_idle(dev_priv);
1457 }
1458
1459 /* Added by Charl P. Botha to call radeon_do_resume_cp().
1460  */
1461 int radeon_cp_resume(struct drm_device *dev, void *data, struct drm_file *file_priv)
1462 {
1463
1464         return radeon_do_resume_cp(dev);
1465 }
1466
1467 int radeon_engine_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1468 {
1469         DRM_DEBUG("\n");
1470
1471         LOCK_TEST_WITH_RETURN(dev, file_priv);
1472
1473         return radeon_do_engine_reset(dev);
1474 }
1475
1476 /* ================================================================
1477  * Fullscreen mode
1478  */
1479
1480 /* KW: Deprecated to say the least:
1481  */
1482 int radeon_fullscreen(struct drm_device *dev, void *data, struct drm_file *file_priv)
1483 {
1484         return 0;
1485 }
1486
1487 /* ================================================================
1488  * Freelist management
1489  */
1490
1491 /* Original comment: FIXME: ROTATE_BUFS is a hack to cycle through
1492  *   bufs until freelist code is used.  Note this hides a problem with
1493  *   the scratch register * (used to keep track of last buffer
1494  *   completed) being written to before * the last buffer has actually
1495  *   completed rendering.
1496  *
1497  * KW:  It's also a good way to find free buffers quickly.
1498  *
1499  * KW: Ideally this loop wouldn't exist, and freelist_get wouldn't
1500  * sleep.  However, bugs in older versions of radeon_accel.c mean that
1501  * we essentially have to do this, else old clients will break.
1502  *
1503  * However, it does leave open a potential deadlock where all the
1504  * buffers are held by other clients, which can't release them because
1505  * they can't get the lock.
1506  */
1507
1508 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1509 {
1510         struct drm_device_dma *dma = dev->dma;
1511         drm_radeon_private_t *dev_priv = dev->dev_private;
1512         drm_radeon_buf_priv_t *buf_priv;
1513         struct drm_buf *buf;
1514         int i, t;
1515         int start;
1516
1517         if (++dev_priv->last_buf >= dma->buf_count)
1518                 dev_priv->last_buf = 0;
1519
1520         start = dev_priv->last_buf;
1521
1522         for (t = 0; t < dev_priv->usec_timeout; t++) {
1523                 u32 done_age = GET_SCRATCH(1);
1524                 DRM_DEBUG("done_age = %d\n", done_age);
1525                 for (i = start; i < dma->buf_count; i++) {
1526                         buf = dma->buflist[i];
1527                         buf_priv = buf->dev_private;
1528                         if (buf->file_priv == NULL || (buf->pending &&
1529                                                        buf_priv->age <=
1530                                                        done_age)) {
1531                                 dev_priv->stats.requested_bufs++;
1532                                 buf->pending = 0;
1533                                 return buf;
1534                         }
1535                         start = 0;
1536                 }
1537
1538                 if (t) {
1539                         DRM_UDELAY(1);
1540                         dev_priv->stats.freelist_loops++;
1541                 }
1542         }
1543
1544         DRM_DEBUG("returning NULL!\n");
1545         return NULL;
1546 }
1547
1548 #if 0
1549 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1550 {
1551         struct drm_device_dma *dma = dev->dma;
1552         drm_radeon_private_t *dev_priv = dev->dev_private;
1553         drm_radeon_buf_priv_t *buf_priv;
1554         struct drm_buf *buf;
1555         int i, t;
1556         int start;
1557         u32 done_age = DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1));
1558
1559         if (++dev_priv->last_buf >= dma->buf_count)
1560                 dev_priv->last_buf = 0;
1561
1562         start = dev_priv->last_buf;
1563         dev_priv->stats.freelist_loops++;
1564
1565         for (t = 0; t < 2; t++) {
1566                 for (i = start; i < dma->buf_count; i++) {
1567                         buf = dma->buflist[i];
1568                         buf_priv = buf->dev_private;
1569                         if (buf->file_priv == 0 || (buf->pending &&
1570                                                     buf_priv->age <=
1571                                                     done_age)) {
1572                                 dev_priv->stats.requested_bufs++;
1573                                 buf->pending = 0;
1574                                 return buf;
1575                         }
1576                 }
1577                 start = 0;
1578         }
1579
1580         return NULL;
1581 }
1582 #endif
1583
1584 void radeon_freelist_reset(struct drm_device * dev)
1585 {
1586         struct drm_device_dma *dma = dev->dma;
1587         drm_radeon_private_t *dev_priv = dev->dev_private;
1588         int i;
1589
1590         dev_priv->last_buf = 0;
1591         for (i = 0; i < dma->buf_count; i++) {
1592                 struct drm_buf *buf = dma->buflist[i];
1593                 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
1594                 buf_priv->age = 0;
1595         }
1596 }
1597
1598 /* ================================================================
1599  * CP command submission
1600  */
1601
1602 int radeon_wait_ring(drm_radeon_private_t * dev_priv, int n)
1603 {
1604         drm_radeon_ring_buffer_t *ring = &dev_priv->ring;
1605         int i;
1606         u32 last_head = GET_RING_HEAD(dev_priv);
1607
1608         for (i = 0; i < dev_priv->usec_timeout; i++) {
1609                 u32 head = GET_RING_HEAD(dev_priv);
1610
1611                 ring->space = (head - ring->tail) * sizeof(u32);
1612                 if (ring->space <= 0)
1613                         ring->space += ring->size;
1614                 if (ring->space > n)
1615                         return 0;
1616
1617                 dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
1618
1619                 if (head != last_head)
1620                         i = 0;
1621                 last_head = head;
1622
1623                 DRM_UDELAY(1);
1624         }
1625
1626         /* FIXME: This return value is ignored in the BEGIN_RING macro! */
1627 #if RADEON_FIFO_DEBUG
1628         radeon_status(dev_priv);
1629         DRM_ERROR("failed!\n");
1630 #endif
1631         return -EBUSY;
1632 }
1633
1634 static int radeon_cp_get_buffers(struct drm_device *dev,
1635                                  struct drm_file *file_priv,
1636                                  struct drm_dma * d)
1637 {
1638         int i;
1639         struct drm_buf *buf;
1640
1641         for (i = d->granted_count; i < d->request_count; i++) {
1642                 buf = radeon_freelist_get(dev);
1643                 if (!buf)
1644                         return -EBUSY;  /* NOTE: broken client */
1645
1646                 buf->file_priv = file_priv;
1647
1648                 if (DRM_COPY_TO_USER(&d->request_indices[i], &buf->idx,
1649                                      sizeof(buf->idx)))
1650                         return -EFAULT;
1651                 if (DRM_COPY_TO_USER(&d->request_sizes[i], &buf->total,
1652                                      sizeof(buf->total)))
1653                         return -EFAULT;
1654
1655                 d->granted_count++;
1656         }
1657         return 0;
1658 }
1659
1660 int radeon_cp_buffers(struct drm_device *dev, void *data, struct drm_file *file_priv)
1661 {
1662         struct drm_device_dma *dma = dev->dma;
1663         int ret = 0;
1664         struct drm_dma *d = data;
1665
1666         LOCK_TEST_WITH_RETURN(dev, file_priv);
1667
1668         /* Please don't send us buffers.
1669          */
1670         if (d->send_count != 0) {
1671                 DRM_ERROR("Process %d trying to send %d buffers via drmDMA\n",
1672                           DRM_CURRENTPID, d->send_count);
1673                 return -EINVAL;
1674         }
1675
1676         /* We'll send you buffers.
1677          */
1678         if (d->request_count < 0 || d->request_count > dma->buf_count) {
1679                 DRM_ERROR("Process %d trying to get %d buffers (of %d max)\n",
1680                           DRM_CURRENTPID, d->request_count, dma->buf_count);
1681                 return -EINVAL;
1682         }
1683
1684         d->granted_count = 0;
1685
1686         if (d->request_count) {
1687                 ret = radeon_cp_get_buffers(dev, file_priv, d);
1688         }
1689
1690         return ret;
1691 }
1692
1693 int radeon_driver_load(struct drm_device *dev, unsigned long flags)
1694 {
1695         drm_radeon_private_t *dev_priv;
1696         int ret = 0;
1697
1698         dev_priv = drm_alloc(sizeof(drm_radeon_private_t), DRM_MEM_DRIVER);
1699         if (dev_priv == NULL)
1700                 return -ENOMEM;
1701
1702         memset(dev_priv, 0, sizeof(drm_radeon_private_t));
1703         dev->dev_private = (void *)dev_priv;
1704         dev_priv->flags = flags;
1705
1706         switch (flags & RADEON_FAMILY_MASK) {
1707         case CHIP_R100:
1708         case CHIP_RV200:
1709         case CHIP_R200:
1710         case CHIP_R300:
1711         case CHIP_R350:
1712         case CHIP_R420:
1713         case CHIP_RV410:
1714         case CHIP_RV515:
1715         case CHIP_R520:
1716         case CHIP_RV570:
1717         case CHIP_R580:
1718                 dev_priv->flags |= RADEON_HAS_HIERZ;
1719                 break;
1720         default:
1721                 /* all other chips have no hierarchical z buffer */
1722                 break;
1723         }
1724
1725         if (drm_device_is_agp(dev))
1726                 dev_priv->flags |= RADEON_IS_AGP;
1727         else if (drm_device_is_pcie(dev))
1728                 dev_priv->flags |= RADEON_IS_PCIE;
1729         else
1730                 dev_priv->flags |= RADEON_IS_PCI;
1731
1732         DRM_DEBUG("%s card detected\n",
1733                   ((dev_priv->flags & RADEON_IS_AGP) ? "AGP" : (((dev_priv->flags & RADEON_IS_PCIE) ? "PCIE" : "PCI"))));
1734         return ret;
1735 }
1736
1737 /* Create mappings for registers and framebuffer so userland doesn't necessarily
1738  * have to find them.
1739  */
1740 int radeon_driver_firstopen(struct drm_device *dev)
1741 {
1742         int ret;
1743         drm_local_map_t *map;
1744         drm_radeon_private_t *dev_priv = dev->dev_private;
1745
1746         dev_priv->gart_info.table_size = RADEON_PCIGART_TABLE_SIZE;
1747
1748         ret = drm_addmap(dev, drm_get_resource_start(dev, 2),
1749                          drm_get_resource_len(dev, 2), _DRM_REGISTERS,
1750                          _DRM_READ_ONLY, &dev_priv->mmio);
1751         if (ret != 0)
1752                 return ret;
1753
1754         dev_priv->fb_aper_offset = drm_get_resource_start(dev, 0);
1755         ret = drm_addmap(dev, dev_priv->fb_aper_offset,
1756                          drm_get_resource_len(dev, 0), _DRM_FRAME_BUFFER,
1757                          _DRM_WRITE_COMBINING, &map);
1758         if (ret != 0)
1759                 return ret;
1760
1761         return 0;
1762 }
1763
1764 int radeon_driver_unload(struct drm_device *dev)
1765 {
1766         drm_radeon_private_t *dev_priv = dev->dev_private;
1767
1768         DRM_DEBUG("\n");
1769         drm_free(dev_priv, sizeof(*dev_priv), DRM_MEM_DRIVER);
1770
1771         dev->dev_private = NULL;
1772         return 0;
1773 }