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 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv);
  44
  45 static u32 R500_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  46 {
  47         u32 ret;
  48         RADEON_WRITE(R520_MC_IND_INDEX, 0x7f0000 | (addr & 0xff));
  49         ret = RADEON_READ(R520_MC_IND_DATA);
  50         RADEON_WRITE(R520_MC_IND_INDEX, 0);
  51         return ret;
  52 }
  53
  54 static u32 RS480_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  55 {
  56         u32 ret;
  57         RADEON_WRITE(RS480_NB_MC_INDEX, addr & 0xff);
  58         ret = RADEON_READ(RS480_NB_MC_DATA);
  59         RADEON_WRITE(RS480_NB_MC_INDEX, 0xff);
  60         return ret;
  61 }
  62
  63 static u32 RS690_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  64 {
  65         u32 ret;
  66         RADEON_WRITE(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
  67         ret = RADEON_READ(RS690_MC_DATA);
  68         RADEON_WRITE(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
  69         return ret;
  70 }
  71
  72 static u32 IGP_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  73 {
  74         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
  75             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
  76                 return RS690_READ_MCIND(dev_priv, addr);
  77         else
  78                 return RS480_READ_MCIND(dev_priv, addr);
  79 }
  80
  81 u32 radeon_read_fb_location(drm_radeon_private_t *dev_priv)
  82 {
  83
  84         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  85                 return R500_READ_MCIND(dev_priv, RV515_MC_FB_LOCATION);
  86         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
  87                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
  88                 return RS690_READ_MCIND(dev_priv, RS690_MC_FB_LOCATION);
  89         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
  90                 return R500_READ_MCIND(dev_priv, R520_MC_FB_LOCATION);
  91         else
  92                 return RADEON_READ(RADEON_MC_FB_LOCATION);
  93 }
  94
  95 static void radeon_write_fb_location(drm_radeon_private_t *dev_priv, u32 fb_loc)
  96 {
  97         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  98                 R500_WRITE_MCIND(RV515_MC_FB_LOCATION, fb_loc);
  99         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 100                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 101                 RS690_WRITE_MCIND(RS690_MC_FB_LOCATION, fb_loc);
 102         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 103                 R500_WRITE_MCIND(R520_MC_FB_LOCATION, fb_loc);
 104         else
 105                 RADEON_WRITE(RADEON_MC_FB_LOCATION, fb_loc);
 106 }
 107
 108 static void radeon_write_agp_location(drm_radeon_private_t *dev_priv, u32 agp_loc)
 109 {
 110         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 111                 R500_WRITE_MCIND(RV515_MC_AGP_LOCATION, agp_loc);
 112         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 113                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 114                 RS690_WRITE_MCIND(RS690_MC_AGP_LOCATION, agp_loc);
 115         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 116                 R500_WRITE_MCIND(R520_MC_AGP_LOCATION, agp_loc);
 117         else
 118                 RADEON_WRITE(RADEON_MC_AGP_LOCATION, agp_loc);
 119 }
 120
 121 static void radeon_write_agp_base(drm_radeon_private_t *dev_priv, u64 agp_base)
 122 {
 123         u32 agp_base_hi = upper_32_bits(agp_base);
 124         u32 agp_base_lo = agp_base & 0xffffffff;
 125
 126         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) {
 127                 R500_WRITE_MCIND(RV515_MC_AGP_BASE, agp_base_lo);
 128                 R500_WRITE_MCIND(RV515_MC_AGP_BASE_2, agp_base_hi);
 129         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 130                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 131                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE, agp_base_lo);
 132                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE_2, agp_base_hi);
 133         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515) {
 134                 R500_WRITE_MCIND(R520_MC_AGP_BASE, agp_base_lo);
 135                 R500_WRITE_MCIND(R520_MC_AGP_BASE_2, agp_base_hi);
 136         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 137                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 138                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 139                 RADEON_WRITE(RS480_AGP_BASE_2, agp_base_hi);
 140         } else {
 141                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 142                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R200)
 143                         RADEON_WRITE(RADEON_AGP_BASE_2, agp_base_hi);
 144         }
 145 }
 146
 147 static int RADEON_READ_PLL(struct drm_device * dev, int addr)
 148 {
 149         drm_radeon_private_t *dev_priv = dev->dev_private;
 150
 151         RADEON_WRITE8(RADEON_CLOCK_CNTL_INDEX, addr & 0x1f);
 152         return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 153 }
 154
 155 static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 156 {
 157         RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 158         return RADEON_READ(RADEON_PCIE_DATA);
 159 }
 160
 161 #if RADEON_FIFO_DEBUG
 162 static void radeon_status(drm_radeon_private_t * dev_priv)
 163 {
 164         printk("%s:\n", __func__);
 165         printk("RBBM_STATUS = 0x%08x\n",
 166                (unsigned int)RADEON_READ(RADEON_RBBM_STATUS));
 167         printk("CP_RB_RTPR = 0x%08x\n",
 168                (unsigned int)RADEON_READ(RADEON_CP_RB_RPTR));
 169         printk("CP_RB_WTPR = 0x%08x\n",
 170                (unsigned int)RADEON_READ(RADEON_CP_RB_WPTR));
 171         printk("AIC_CNTL = 0x%08x\n",
 172                (unsigned int)RADEON_READ(RADEON_AIC_CNTL));
 173         printk("AIC_STAT = 0x%08x\n",
 174                (unsigned int)RADEON_READ(RADEON_AIC_STAT));
 175         printk("AIC_PT_BASE = 0x%08x\n",
 176                (unsigned int)RADEON_READ(RADEON_AIC_PT_BASE));
 177         printk("TLB_ADDR = 0x%08x\n",
 178                (unsigned int)RADEON_READ(RADEON_AIC_TLB_ADDR));
 179         printk("TLB_DATA = 0x%08x\n",
 180                (unsigned int)RADEON_READ(RADEON_AIC_TLB_DATA));
 181 }
 182 #endif
 183
 184 /* ================================================================
 185  * Engine, FIFO control
 186  */
 187
 188 static int radeon_do_pixcache_flush(drm_radeon_private_t * dev_priv)
 189 {
 190         u32 tmp;
 191         int i;
 192
 193         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 194
 195         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV280) {
 196                 tmp = RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT);
 197                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 198                 RADEON_WRITE(RADEON_RB3D_DSTCACHE_CTLSTAT, tmp);
 199
 200                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 201                         if (!(RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT)
 202                               & RADEON_RB3D_DC_BUSY)) {
 203                                 return 0;
 204                         }
 205                         DRM_UDELAY(1);
 206                 }
 207         } else {
 208                 /* don't flush or purge cache here or lockup */
 209                 return 0;
 210         }
 211
 212 #if RADEON_FIFO_DEBUG
 213         DRM_ERROR("failed!\n");
 214         radeon_status(dev_priv);
 215 #endif
 216         return -EBUSY;
 217 }
 218
 219 static int radeon_do_wait_for_fifo(drm_radeon_private_t * dev_priv, int entries)
 220 {
 221         int i;
 222
 223         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 224
 225         for (i = 0; i < dev_priv->usec_timeout; i++) {
 226                 int slots = (RADEON_READ(RADEON_RBBM_STATUS)
 227                              & RADEON_RBBM_FIFOCNT_MASK);
 228                 if (slots >= entries)
 229                         return 0;
 230                 DRM_UDELAY(1);
 231         }
 232         DRM_DEBUG("wait for fifo failed status : 0x%08X 0x%08X\n",
 233                  RADEON_READ(RADEON_RBBM_STATUS),
 234                  RADEON_READ(R300_VAP_CNTL_STATUS));
 235
 236 #if RADEON_FIFO_DEBUG
 237         DRM_ERROR("failed!\n");
 238         radeon_status(dev_priv);
 239 #endif
 240         return -EBUSY;
 241 }
 242
 243 static int radeon_do_wait_for_idle(drm_radeon_private_t * dev_priv)
 244 {
 245         int i, ret;
 246
 247         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 248
 249         ret = radeon_do_wait_for_fifo(dev_priv, 64);
 250         if (ret)
 251                 return ret;
 252
 253         for (i = 0; i < dev_priv->usec_timeout; i++) {
 254                 if (!(RADEON_READ(RADEON_RBBM_STATUS)
 255                       & RADEON_RBBM_ACTIVE)) {
 256                         radeon_do_pixcache_flush(dev_priv);
 257                         return 0;
 258                 }
 259                 DRM_UDELAY(1);
 260         }
 261         DRM_DEBUG("wait idle failed status : 0x%08X 0x%08X\n",
 262                  RADEON_READ(RADEON_RBBM_STATUS),
 263                  RADEON_READ(R300_VAP_CNTL_STATUS));
 264
 265 #if RADEON_FIFO_DEBUG
 266         DRM_ERROR("failed!\n");
 267         radeon_status(dev_priv);
 268 #endif
 269         return -EBUSY;
 270 }
 271
 272 static void radeon_init_pipes(drm_radeon_private_t *dev_priv)
 273 {
 274         uint32_t gb_tile_config, gb_pipe_sel = 0;
 275
 276         /* RS4xx/RS6xx/R4xx/R5xx */
 277         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R420) {
 278                 gb_pipe_sel = RADEON_READ(R400_GB_PIPE_SELECT);
 279                 dev_priv->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
 280         } else {
 281                 /* R3xx */
 282                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 283                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350)) {
 284                         dev_priv->num_gb_pipes = 2;
 285                 } else {
 286                         /* R3Vxx */
 287                         dev_priv->num_gb_pipes = 1;
 288                 }
 289         }
 290         DRM_INFO("Num pipes: %d\n", dev_priv->num_gb_pipes);
 291
 292         gb_tile_config = (R300_ENABLE_TILING | R300_TILE_SIZE_16 /*| R300_SUBPIXEL_1_16*/);
 293
 294         switch (dev_priv->num_gb_pipes) {
 295         case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
 296         case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
 297         case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
 298         default:
 299         case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
 300         }
 301
 302         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
 303                 RADEON_WRITE_PLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
 304                 RADEON_WRITE(R500_SU_REG_DEST, ((1 << dev_priv->num_gb_pipes) - 1));
 305         }
 306         RADEON_WRITE(R300_GB_TILE_CONFIG, gb_tile_config);
 307         radeon_do_wait_for_idle(dev_priv);
 308         RADEON_WRITE(R300_DST_PIPE_CONFIG, RADEON_READ(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
 309         RADEON_WRITE(R300_RB2D_DSTCACHE_MODE, (RADEON_READ(R300_RB2D_DSTCACHE_MODE) |
 310                                                R300_DC_AUTOFLUSH_ENABLE |
 311                                                R300_DC_DC_DISABLE_IGNORE_PE));
 312
 313
 314 }
 315
 316 /* ================================================================
 317  * CP control, initialization
 318  */
 319
 320 /* Load the microcode for the CP */
 321 static void radeon_cp_load_microcode(drm_radeon_private_t * dev_priv)
 322 {
 323         int i;
 324         DRM_DEBUG("\n");
 325
 326         radeon_do_wait_for_idle(dev_priv);
 327
 328         RADEON_WRITE(RADEON_CP_ME_RAM_ADDR, 0);
 329         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R100) ||
 330             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV100) ||
 331             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV200) ||
 332             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS100) ||
 333             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS200)) {
 334                 DRM_INFO("Loading R100 Microcode\n");
 335                 for (i = 0; i < 256; i++) {
 336                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 337                                      R100_cp_microcode[i][1]);
 338                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 339                                      R100_cp_microcode[i][0]);
 340                 }
 341         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R200) ||
 342                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV250) ||
 343                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV280) ||
 344                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS300)) {
 345                 DRM_INFO("Loading R200 Microcode\n");
 346                 for (i = 0; i < 256; i++) {
 347                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 348                                      R200_cp_microcode[i][1]);
 349                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 350                                      R200_cp_microcode[i][0]);
 351                 }
 352         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 353                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350) ||
 354                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV350) ||
 355                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV380) ||
 356                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 357                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 358                 DRM_INFO("Loading R300 Microcode\n");
 359                 for (i = 0; i < 256; i++) {
 360                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 361                                      R300_cp_microcode[i][1]);
 362                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 363                                      R300_cp_microcode[i][0]);
 364                 }
 365         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 366                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R423) ||
 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                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 377                 DRM_INFO("Loading RS690/RS740 Microcode\n");
 378                 for (i = 0; i < 256; i++) {
 379                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 380                                      RS690_cp_microcode[i][1]);
 381                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 382                                      RS690_cp_microcode[i][0]);
 383                 }
 384         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) ||
 385                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R520) ||
 386                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530) ||
 387                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R580) ||
 388                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV560) ||
 389                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV570)) {
 390                 DRM_INFO("Loading R500 Microcode\n");
 391                 for (i = 0; i < 256; i++) {
 392                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 393                                      R520_cp_microcode[i][1]);
 394                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 395                                      R520_cp_microcode[i][0]);
 396                 }
 397         }
 398 }
 399
 400 /* Flush any pending commands to the CP.  This should only be used just
 401  * prior to a wait for idle, as it informs the engine that the command
 402  * stream is ending.
 403  */
 404 static void radeon_do_cp_flush(drm_radeon_private_t * dev_priv)
 405 {
 406         DRM_DEBUG("\n");
 407 #if 0
 408         u32 tmp;
 409
 410         tmp = RADEON_READ(RADEON_CP_RB_WPTR) | (1 << 31);
 411         RADEON_WRITE(RADEON_CP_RB_WPTR, tmp);
 412 #endif
 413 }
 414
 415 /* Wait for the CP to go idle.
 416  */
 417 int radeon_do_cp_idle(drm_radeon_private_t * dev_priv)
 418 {
 419         RING_LOCALS;
 420         DRM_DEBUG("\n");
 421
 422         BEGIN_RING(6);
 423
 424         RADEON_PURGE_CACHE();
 425         RADEON_PURGE_ZCACHE();
 426         RADEON_WAIT_UNTIL_IDLE();
 427
 428         ADVANCE_RING();
 429         COMMIT_RING();
 430
 431         return radeon_do_wait_for_idle(dev_priv);
 432 }
 433
 434 /* Start the Command Processor.
 435  */
 436 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 437 {
 438         RING_LOCALS;
 439         DRM_DEBUG("\n");
 440
 441         radeon_do_wait_for_idle(dev_priv);
 442
 443         RADEON_WRITE(RADEON_CP_CSQ_CNTL, dev_priv->cp_mode);
 444
 445         dev_priv->cp_running = 1;
 446
 447         BEGIN_RING(8);
 448         /* isync can only be written through cp on r5xx write it here */
 449         OUT_RING(CP_PACKET0(RADEON_ISYNC_CNTL, 0));
 450         OUT_RING(RADEON_ISYNC_ANY2D_IDLE3D |
 451                  RADEON_ISYNC_ANY3D_IDLE2D |
 452                  RADEON_ISYNC_WAIT_IDLEGUI |
 453                  RADEON_ISYNC_CPSCRATCH_IDLEGUI);
 454         RADEON_PURGE_CACHE();
 455         RADEON_PURGE_ZCACHE();
 456         RADEON_WAIT_UNTIL_IDLE();
 457         ADVANCE_RING();
 458         COMMIT_RING();
 459
 460         dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
 461 }
 462
 463 /* Reset the Command Processor.  This will not flush any pending
 464  * commands, so you must wait for the CP command stream to complete
 465  * before calling this routine.
 466  */
 467 static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
 468 {
 469         u32 cur_read_ptr;
 470         DRM_DEBUG("\n");
 471
 472         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 473         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 474         SET_RING_HEAD(dev_priv, cur_read_ptr);
 475         dev_priv->ring.tail = cur_read_ptr;
 476 }
 477
 478 /* Stop the Command Processor.  This will not flush any pending
 479  * commands, so you must flush the command stream and wait for the CP
 480  * to go idle before calling this routine.
 481  */
 482 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 483 {
 484         DRM_DEBUG("\n");
 485
 486         RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 487
 488         dev_priv->cp_running = 0;
 489 }
 490
 491 /* Reset the engine.  This will stop the CP if it is running.
 492  */
 493 static int radeon_do_engine_reset(struct drm_device * dev)
 494 {
 495         drm_radeon_private_t *dev_priv = dev->dev_private;
 496         u32 clock_cntl_index = 0, mclk_cntl = 0, rbbm_soft_reset;
 497         DRM_DEBUG("\n");
 498
 499         radeon_do_pixcache_flush(dev_priv);
 500
 501         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 502                 /* may need something similar for newer chips */
 503                 clock_cntl_index = RADEON_READ(RADEON_CLOCK_CNTL_INDEX);
 504                 mclk_cntl = RADEON_READ_PLL(dev, RADEON_MCLK_CNTL);
 505
 506                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, (mclk_cntl |
 507                                                     RADEON_FORCEON_MCLKA |
 508                                                     RADEON_FORCEON_MCLKB |
 509                                                     RADEON_FORCEON_YCLKA |
 510                                                     RADEON_FORCEON_YCLKB |
 511                                                     RADEON_FORCEON_MC |
 512                                                     RADEON_FORCEON_AIC));
 513         }
 514
 515         rbbm_soft_reset = RADEON_READ(RADEON_RBBM_SOFT_RESET);
 516
 517         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset |
 518                                               RADEON_SOFT_RESET_CP |
 519                                               RADEON_SOFT_RESET_HI |
 520                                               RADEON_SOFT_RESET_SE |
 521                                               RADEON_SOFT_RESET_RE |
 522                                               RADEON_SOFT_RESET_PP |
 523                                               RADEON_SOFT_RESET_E2 |
 524                                               RADEON_SOFT_RESET_RB));
 525         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 526         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset &
 527                                               ~(RADEON_SOFT_RESET_CP |
 528                                                 RADEON_SOFT_RESET_HI |
 529                                                 RADEON_SOFT_RESET_SE |
 530                                                 RADEON_SOFT_RESET_RE |
 531                                                 RADEON_SOFT_RESET_PP |
 532                                                 RADEON_SOFT_RESET_E2 |
 533                                                 RADEON_SOFT_RESET_RB)));
 534         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 535
 536         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 537                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, mclk_cntl);
 538                 RADEON_WRITE(RADEON_CLOCK_CNTL_INDEX, clock_cntl_index);
 539                 RADEON_WRITE(RADEON_RBBM_SOFT_RESET, rbbm_soft_reset);
 540         }
 541
 542         /* setup the raster pipes */
 543         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R300)
 544             radeon_init_pipes(dev_priv);
 545
 546         /* Reset the CP ring */
 547         radeon_do_cp_reset(dev_priv);
 548
 549         /* The CP is no longer running after an engine reset */
 550         dev_priv->cp_running = 0;
 551
 552         /* Reset any pending vertex, indirect buffers */
 553         radeon_freelist_reset(dev);
 554
 555         return 0;
 556 }
 557
 558 static void radeon_cp_init_ring_buffer(struct drm_device * dev,
 559                                        drm_radeon_private_t * dev_priv)
 560 {
 561         u32 ring_start, cur_read_ptr;
 562         u32 tmp;
 563
 564         /* Initialize the memory controller. With new memory map, the fb location
 565          * is not changed, it should have been properly initialized already. Part
 566          * of the problem is that the code below is bogus, assuming the GART is
 567          * always appended to the fb which is not necessarily the case
 568          */
 569         if (!dev_priv->new_memmap)
 570                 radeon_write_fb_location(dev_priv,
 571                              ((dev_priv->gart_vm_start - 1) & 0xffff0000)
 572                              | (dev_priv->fb_location >> 16));
 573
 574 #if __OS_HAS_AGP
 575         if (dev_priv->flags & RADEON_IS_AGP) {
 576                 radeon_write_agp_base(dev_priv, dev->agp->base);
 577
 578                 radeon_write_agp_location(dev_priv,
 579                              (((dev_priv->gart_vm_start - 1 +
 580                                 dev_priv->gart_size) & 0xffff0000) |
 581                               (dev_priv->gart_vm_start >> 16)));
 582
 583                 ring_start = (dev_priv->cp_ring->offset
 584                               - dev->agp->base
 585                               + dev_priv->gart_vm_start);
 586         } else
 587 #endif
 588                 ring_start = (dev_priv->cp_ring->offset
 589                               - (unsigned long)dev->sg->virtual
 590                               + dev_priv->gart_vm_start);
 591
 592         RADEON_WRITE(RADEON_CP_RB_BASE, ring_start);
 593
 594         /* Set the write pointer delay */
 595         RADEON_WRITE(RADEON_CP_RB_WPTR_DELAY, 0);
 596
 597         /* Initialize the ring buffer's read and write pointers */
 598         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 599         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 600         SET_RING_HEAD(dev_priv, cur_read_ptr);
 601         dev_priv->ring.tail = cur_read_ptr;
 602
 603 #if __OS_HAS_AGP
 604         if (dev_priv->flags & RADEON_IS_AGP) {
 605                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 606                              dev_priv->ring_rptr->offset
 607                              - dev->agp->base + dev_priv->gart_vm_start);
 608         } else
 609 #endif
 610         {
 611                 struct drm_sg_mem *entry = dev->sg;
 612                 unsigned long tmp_ofs, page_ofs;
 613
 614                 tmp_ofs = dev_priv->ring_rptr->offset -
 615                                 (unsigned long)dev->sg->virtual;
 616                 page_ofs = tmp_ofs >> PAGE_SHIFT;
 617
 618                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR, entry->busaddr[page_ofs]);
 619                 DRM_DEBUG("ring rptr: offset=0x%08lx handle=0x%08lx\n",
 620                           (unsigned long)entry->busaddr[page_ofs],
 621                           entry->handle + tmp_ofs);
 622         }
 623
 624         /* Set ring buffer size */
 625 #ifdef __BIG_ENDIAN
 626         RADEON_WRITE(RADEON_CP_RB_CNTL,
 627                      RADEON_BUF_SWAP_32BIT |
 628                      (dev_priv->ring.fetch_size_l2ow << 18) |
 629                      (dev_priv->ring.rptr_update_l2qw << 8) |
 630                      dev_priv->ring.size_l2qw);
 631 #else
 632         RADEON_WRITE(RADEON_CP_RB_CNTL,
 633                      (dev_priv->ring.fetch_size_l2ow << 18) |
 634                      (dev_priv->ring.rptr_update_l2qw << 8) |
 635                      dev_priv->ring.size_l2qw);
 636 #endif
 637
 638
 639         /* Initialize the scratch register pointer.  This will cause
 640          * the scratch register values to be written out to memory
 641          * whenever they are updated.
 642          *
 643          * We simply put this behind the ring read pointer, this works
 644          * with PCI GART as well as (whatever kind of) AGP GART
 645          */
 646         RADEON_WRITE(RADEON_SCRATCH_ADDR, RADEON_READ(RADEON_CP_RB_RPTR_ADDR)
 647                      + RADEON_SCRATCH_REG_OFFSET);
 648
 649         dev_priv->scratch = ((__volatile__ u32 *)
 650                              dev_priv->ring_rptr->handle +
 651                              (RADEON_SCRATCH_REG_OFFSET / sizeof(u32)));
 652
 653         RADEON_WRITE(RADEON_SCRATCH_UMSK, 0x7);
 654
 655         /* Turn on bus mastering */
 656         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 657             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 658                 /* rs600/rs690/rs740 */
 659                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RS600_BUS_MASTER_DIS;
 660                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 661         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV350) ||
 662                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 663                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 664                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 665                 /* r1xx, r2xx, r300, r(v)350, r420/r481, rs400/rs480 */
 666                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RADEON_BUS_MASTER_DIS;
 667                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 668         } /* PCIE cards appears to not need this */
 669
 670         dev_priv->sarea_priv->last_frame = dev_priv->scratch[0] = 0;
 671         RADEON_WRITE(RADEON_LAST_FRAME_REG, dev_priv->sarea_priv->last_frame);
 672
 673         dev_priv->sarea_priv->last_dispatch = dev_priv->scratch[1] = 0;
 674         RADEON_WRITE(RADEON_LAST_DISPATCH_REG,
 675                      dev_priv->sarea_priv->last_dispatch);
 676
 677         dev_priv->sarea_priv->last_clear = dev_priv->scratch[2] = 0;
 678         RADEON_WRITE(RADEON_LAST_CLEAR_REG, dev_priv->sarea_priv->last_clear);
 679
 680         radeon_do_wait_for_idle(dev_priv);
 681
 682         /* Sync everything up */
 683         RADEON_WRITE(RADEON_ISYNC_CNTL,
 684                      (RADEON_ISYNC_ANY2D_IDLE3D |
 685                       RADEON_ISYNC_ANY3D_IDLE2D |
 686                       RADEON_ISYNC_WAIT_IDLEGUI |
 687                       RADEON_ISYNC_CPSCRATCH_IDLEGUI));
 688
 689 }
 690
 691 static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 692 {
 693         u32 tmp;
 694
 695         /* Start with assuming that writeback doesn't work */
 696         dev_priv->writeback_works = 0;
 697
 698         /* Writeback doesn't seem to work everywhere, test it here and possibly
 699          * enable it if it appears to work
 700          */
 701         DRM_WRITE32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1), 0);
 702         RADEON_WRITE(RADEON_SCRATCH_REG1, 0xdeadbeef);
 703
 704         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 705                 if (DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1)) ==
 706                     0xdeadbeef)
 707                         break;
 708                 DRM_UDELAY(1);
 709         }
 710
 711         if (tmp < dev_priv->usec_timeout) {
 712                 dev_priv->writeback_works = 1;
 713                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 714         } else {
 715                 dev_priv->writeback_works = 0;
 716                 DRM_INFO("writeback test failed\n");
 717         }
 718         if (radeon_no_wb == 1) {
 719                 dev_priv->writeback_works = 0;
 720                 DRM_INFO("writeback forced off\n");
 721         }
 722
 723         if (!dev_priv->writeback_works) {
 724                 /* Disable writeback to avoid unnecessary bus master transfer */
 725                 RADEON_WRITE(RADEON_CP_RB_CNTL, RADEON_READ(RADEON_CP_RB_CNTL) |
 726                              RADEON_RB_NO_UPDATE);
 727                 RADEON_WRITE(RADEON_SCRATCH_UMSK, 0);
 728         }
 729 }
 730
 731 /* Enable or disable IGP GART on the chip */
 732 static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
 733 {
 734         u32 temp;
 735
 736         if (on) {
 737                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 738                           dev_priv->gart_vm_start,
 739                           (long)dev_priv->gart_info.bus_addr,
 740                           dev_priv->gart_size);
 741
 742                 temp = IGP_READ_MCIND(dev_priv, RS480_MC_MISC_CNTL);
 743                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 744                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 745                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, (RS480_GART_INDEX_REG_EN |
 746                                                              RS690_BLOCK_GFX_D3_EN));
 747                 else
 748                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, RS480_GART_INDEX_REG_EN);
 749
 750                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 751                                                                RS480_VA_SIZE_32MB));
 752
 753                 temp = IGP_READ_MCIND(dev_priv, RS480_GART_FEATURE_ID);
 754                 IGP_WRITE_MCIND(RS480_GART_FEATURE_ID, (RS480_HANG_EN |
 755                                                         RS480_TLB_ENABLE |
 756                                                         RS480_GTW_LAC_EN |
 757                                                         RS480_1LEVEL_GART));
 758
 759                 temp = dev_priv->gart_info.bus_addr & 0xfffff000;
 760                 temp |= (upper_32_bits(dev_priv->gart_info.bus_addr) & 0xff) << 4;
 761                 IGP_WRITE_MCIND(RS480_GART_BASE, temp);
 762
 763                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_MODE_CNTL);
 764                 IGP_WRITE_MCIND(RS480_AGP_MODE_CNTL, ((1 << RS480_REQ_TYPE_SNOOP_SHIFT) |
 765                                                       RS480_REQ_TYPE_SNOOP_DIS));
 766
 767                 radeon_write_agp_base(dev_priv, dev_priv->gart_vm_start);
 768
 769                 dev_priv->gart_size = 32*1024*1024;
 770                 temp = (((dev_priv->gart_vm_start - 1 + dev_priv->gart_size) &
 771                          0xffff0000) | (dev_priv->gart_vm_start >> 16));
 772
 773                 radeon_write_agp_location(dev_priv, temp);
 774
 775                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_ADDRESS_SPACE_SIZE);
 776                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 777                                                                RS480_VA_SIZE_32MB));
 778
 779                 do {
 780                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 781                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 782                                 break;
 783                         DRM_UDELAY(1);
 784                 } while (1);
 785
 786                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL,
 787                                 RS480_GART_CACHE_INVALIDATE);
 788
 789                 do {
 790                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 791                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 792                                 break;
 793                         DRM_UDELAY(1);
 794                 } while (1);
 795
 796                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL, 0);
 797         } else {
 798                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, 0);
 799         }
 800 }
 801
 802 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
 803 {
 804         u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
 805         if (on) {
 806
 807                 DRM_DEBUG("programming pcie %08X %08lX %08X\n",
 808                           dev_priv->gart_vm_start,
 809                           (long)dev_priv->gart_info.bus_addr,
 810                           dev_priv->gart_size);
 811                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_DISCARD_RD_ADDR_LO,
 812                                   dev_priv->gart_vm_start);
 813                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_BASE,
 814                                   dev_priv->gart_info.bus_addr);
 815                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_START_LO,
 816                                   dev_priv->gart_vm_start);
 817                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_END_LO,
 818                                   dev_priv->gart_vm_start +
 819                                   dev_priv->gart_size - 1);
 820
 821                 radeon_write_agp_location(dev_priv, 0xffffffc0); /* ?? */
 822
 823                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 824                                   RADEON_PCIE_TX_GART_EN);
 825         } else {
 826                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 827                                   tmp & ~RADEON_PCIE_TX_GART_EN);
 828         }
 829 }
 830
 831 /* Enable or disable PCI GART on the chip */
 832 static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
 833 {
 834         u32 tmp;
 835
 836         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 837             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740) ||
 838             (dev_priv->flags & RADEON_IS_IGPGART)) {
 839                 radeon_set_igpgart(dev_priv, on);
 840                 return;
 841         }
 842
 843         if (dev_priv->flags & RADEON_IS_PCIE) {
 844                 radeon_set_pciegart(dev_priv, on);
 845                 return;
 846         }
 847
 848         tmp = RADEON_READ(RADEON_AIC_CNTL);
 849
 850         if (on) {
 851                 RADEON_WRITE(RADEON_AIC_CNTL,
 852                              tmp | RADEON_PCIGART_TRANSLATE_EN);
 853
 854                 /* set PCI GART page-table base address
 855                  */
 856                 RADEON_WRITE(RADEON_AIC_PT_BASE, dev_priv->gart_info.bus_addr);
 857
 858                 /* set address range for PCI address translate
 859                  */
 860                 RADEON_WRITE(RADEON_AIC_LO_ADDR, dev_priv->gart_vm_start);
 861                 RADEON_WRITE(RADEON_AIC_HI_ADDR, dev_priv->gart_vm_start
 862                              + dev_priv->gart_size - 1);
 863
 864                 /* Turn off AGP aperture -- is this required for PCI GART?
 865                  */
 866                 radeon_write_agp_location(dev_priv, 0xffffffc0);
 867                 RADEON_WRITE(RADEON_AGP_COMMAND, 0);    /* clear AGP_COMMAND */
 868         } else {
 869                 RADEON_WRITE(RADEON_AIC_CNTL,
 870                              tmp & ~RADEON_PCIGART_TRANSLATE_EN);
 871         }
 872 }
 873
 874 static int radeon_do_init_cp(struct drm_device * dev, drm_radeon_init_t * init)
 875 {
 876         drm_radeon_private_t *dev_priv = dev->dev_private;
 877
 878         DRM_DEBUG("\n");
 879
 880         /* if we require new memory map but we don't have it fail */
 881         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
 882                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
 883                 radeon_do_cleanup_cp(dev);
 884                 return -EINVAL;
 885         }
 886
 887         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
 888                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
 889                 dev_priv->flags &= ~RADEON_IS_AGP;
 890         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
 891                    && !init->is_pci) {
 892                 DRM_DEBUG("Restoring AGP flag\n");
 893                 dev_priv->flags |= RADEON_IS_AGP;
 894         }
 895
 896         if ((!(dev_priv->flags & RADEON_IS_AGP)) && !dev->sg) {
 897                 DRM_ERROR("PCI GART memory not allocated!\n");
 898                 radeon_do_cleanup_cp(dev);
 899                 return -EINVAL;
 900         }
 901
 902         dev_priv->usec_timeout = init->usec_timeout;
 903         if (dev_priv->usec_timeout < 1 ||
 904             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
 905                 DRM_DEBUG("TIMEOUT problem!\n");
 906                 radeon_do_cleanup_cp(dev);
 907                 return -EINVAL;
 908         }
 909
 910         /* Enable vblank on CRTC1 for older X servers
 911          */
 912         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
 913
 914         switch(init->func) {
 915         case RADEON_INIT_R200_CP:
 916                 dev_priv->microcode_version = UCODE_R200;
 917                 break;
 918         case RADEON_INIT_R300_CP:
 919                 dev_priv->microcode_version = UCODE_R300;
 920                 break;
 921         default:
 922                 dev_priv->microcode_version = UCODE_R100;
 923         }
 924
 925         dev_priv->do_boxes = 0;
 926         dev_priv->cp_mode = init->cp_mode;
 927
 928         /* We don't support anything other than bus-mastering ring mode,
 929          * but the ring can be in either AGP or PCI space for the ring
 930          * read pointer.
 931          */
 932         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
 933             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
 934                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
 935                 radeon_do_cleanup_cp(dev);
 936                 return -EINVAL;
 937         }
 938
 939         switch (init->fb_bpp) {
 940         case 16:
 941                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
 942                 break;
 943         case 32:
 944         default:
 945                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
 946                 break;
 947         }
 948         dev_priv->front_offset = init->front_offset;
 949         dev_priv->front_pitch = init->front_pitch;
 950         dev_priv->back_offset = init->back_offset;
 951         dev_priv->back_pitch = init->back_pitch;
 952
 953         switch (init->depth_bpp) {
 954         case 16:
 955                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_16BIT_INT_Z;
 956                 break;
 957         case 32:
 958         default:
 959                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_24BIT_INT_Z;
 960                 break;
 961         }
 962         dev_priv->depth_offset = init->depth_offset;
 963         dev_priv->depth_pitch = init->depth_pitch;
 964
 965         /* Hardware state for depth clears.  Remove this if/when we no
 966          * longer clear the depth buffer with a 3D rectangle.  Hard-code
 967          * all values to prevent unwanted 3D state from slipping through
 968          * and screwing with the clear operation.
 969          */
 970         dev_priv->depth_clear.rb3d_cntl = (RADEON_PLANE_MASK_ENABLE |
 971                                            (dev_priv->color_fmt << 10) |
 972                                            (dev_priv->microcode_version ==
 973                                             UCODE_R100 ? RADEON_ZBLOCK16 : 0));
 974
 975         dev_priv->depth_clear.rb3d_zstencilcntl =
 976             (dev_priv->depth_fmt |
 977              RADEON_Z_TEST_ALWAYS |
 978              RADEON_STENCIL_TEST_ALWAYS |
 979              RADEON_STENCIL_S_FAIL_REPLACE |
 980              RADEON_STENCIL_ZPASS_REPLACE |
 981              RADEON_STENCIL_ZFAIL_REPLACE | RADEON_Z_WRITE_ENABLE);
 982
 983         dev_priv->depth_clear.se_cntl = (RADEON_FFACE_CULL_CW |
 984                                          RADEON_BFACE_SOLID |
 985                                          RADEON_FFACE_SOLID |
 986                                          RADEON_FLAT_SHADE_VTX_LAST |
 987                                          RADEON_DIFFUSE_SHADE_FLAT |
 988                                          RADEON_ALPHA_SHADE_FLAT |
 989                                          RADEON_SPECULAR_SHADE_FLAT |
 990                                          RADEON_FOG_SHADE_FLAT |
 991                                          RADEON_VTX_PIX_CENTER_OGL |
 992                                          RADEON_ROUND_MODE_TRUNC |
 993                                          RADEON_ROUND_PREC_8TH_PIX);
 994
 995
 996         dev_priv->ring_offset = init->ring_offset;
 997         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
 998         dev_priv->buffers_offset = init->buffers_offset;
 999         dev_priv->gart_textures_offset = init->gart_textures_offset;
1000
1001         dev_priv->sarea = drm_getsarea(dev);
1002         if (!dev_priv->sarea) {
1003                 DRM_ERROR("could not find sarea!\n");
1004                 radeon_do_cleanup_cp(dev);
1005                 return -EINVAL;
1006         }
1007
1008         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
1009         if (!dev_priv->cp_ring) {
1010                 DRM_ERROR("could not find cp ring region!\n");
1011                 radeon_do_cleanup_cp(dev);
1012                 return -EINVAL;
1013         }
1014         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
1015         if (!dev_priv->ring_rptr) {
1016                 DRM_ERROR("could not find ring read pointer!\n");
1017                 radeon_do_cleanup_cp(dev);
1018                 return -EINVAL;
1019         }
1020         dev->agp_buffer_token = init->buffers_offset;
1021         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1022         if (!dev->agp_buffer_map) {
1023                 DRM_ERROR("could not find dma buffer region!\n");
1024                 radeon_do_cleanup_cp(dev);
1025                 return -EINVAL;
1026         }
1027
1028         if (init->gart_textures_offset) {
1029                 dev_priv->gart_textures =
1030                     drm_core_findmap(dev, init->gart_textures_offset);
1031                 if (!dev_priv->gart_textures) {
1032                         DRM_ERROR("could not find GART texture region!\n");
1033                         radeon_do_cleanup_cp(dev);
1034                         return -EINVAL;
1035                 }
1036         }
1037
1038         dev_priv->sarea_priv =
1039             (drm_radeon_sarea_t *) ((u8 *) dev_priv->sarea->handle +
1040                                     init->sarea_priv_offset);
1041
1042 #if __OS_HAS_AGP
1043         if (dev_priv->flags & RADEON_IS_AGP) {
1044                 drm_core_ioremap(dev_priv->cp_ring, dev);
1045                 drm_core_ioremap(dev_priv->ring_rptr, dev);
1046                 drm_core_ioremap(dev->agp_buffer_map, dev);
1047                 if (!dev_priv->cp_ring->handle ||
1048                     !dev_priv->ring_rptr->handle ||
1049                     !dev->agp_buffer_map->handle) {
1050                         DRM_ERROR("could not find ioremap agp regions!\n");
1051                         radeon_do_cleanup_cp(dev);
1052                         return -EINVAL;
1053                 }
1054         } else
1055 #endif
1056         {
1057                 dev_priv->cp_ring->handle = (void *)dev_priv->cp_ring->offset;
1058                 dev_priv->ring_rptr->handle =
1059                     (void *)dev_priv->ring_rptr->offset;
1060                 dev->agp_buffer_map->handle =
1061                     (void *)dev->agp_buffer_map->offset;
1062
1063                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
1064                           dev_priv->cp_ring->handle);
1065                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
1066                           dev_priv->ring_rptr->handle);
1067                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
1068                           dev->agp_buffer_map->handle);
1069         }
1070
1071         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 16;
1072         dev_priv->fb_size =
1073                 ((radeon_read_fb_location(dev_priv) & 0xffff0000u) + 0x10000)
1074                 - dev_priv->fb_location;
1075
1076         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
1077                                         ((dev_priv->front_offset
1078                                           + dev_priv->fb_location) >> 10));
1079
1080         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
1081                                        ((dev_priv->back_offset
1082                                          + dev_priv->fb_location) >> 10));
1083
1084         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
1085                                         ((dev_priv->depth_offset
1086                                           + dev_priv->fb_location) >> 10));
1087
1088         dev_priv->gart_size = init->gart_size;
1089
1090         /* New let's set the memory map ... */
1091         if (dev_priv->new_memmap) {
1092                 u32 base = 0;
1093
1094                 DRM_INFO("Setting GART location based on new memory map\n");
1095
1096                 /* If using AGP, try to locate the AGP aperture at the same
1097                  * location in the card and on the bus, though we have to
1098                  * align it down.
1099                  */
1100 #if __OS_HAS_AGP
1101                 if (dev_priv->flags & RADEON_IS_AGP) {
1102                         base = dev->agp->base;
1103                         /* Check if valid */
1104                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
1105                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
1106                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
1107                                          dev->agp->base);
1108                                 base = 0;
1109                         }
1110                 }
1111 #endif
1112                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
1113                 if (base == 0) {
1114                         base = dev_priv->fb_location + dev_priv->fb_size;
1115                         if (base < dev_priv->fb_location ||
1116                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
1117                                 base = dev_priv->fb_location
1118                                         - dev_priv->gart_size;
1119                 }
1120                 dev_priv->gart_vm_start = base & 0xffc00000u;
1121                 if (dev_priv->gart_vm_start != base)
1122                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
1123                                  base, dev_priv->gart_vm_start);
1124         } else {
1125                 DRM_INFO("Setting GART location based on old memory map\n");
1126                 dev_priv->gart_vm_start = dev_priv->fb_location +
1127                         RADEON_READ(RADEON_CONFIG_APER_SIZE);
1128         }
1129
1130 #if __OS_HAS_AGP
1131         if (dev_priv->flags & RADEON_IS_AGP)
1132                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1133                                                  - dev->agp->base
1134                                                  + dev_priv->gart_vm_start);
1135         else
1136 #endif
1137                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1138                                         - (unsigned long)dev->sg->virtual
1139                                         + dev_priv->gart_vm_start);
1140
1141         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
1142         DRM_DEBUG("dev_priv->gart_vm_start 0x%x\n", dev_priv->gart_vm_start);
1143         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%lx\n",
1144                   dev_priv->gart_buffers_offset);
1145
1146         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
1147         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
1148                               + init->ring_size / sizeof(u32));
1149         dev_priv->ring.size = init->ring_size;
1150         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
1151
1152         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
1153         dev_priv->ring.rptr_update_l2qw = drm_order( /* init->rptr_update */ 4096 / 8);
1154
1155         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
1156         dev_priv->ring.fetch_size_l2ow = drm_order( /* init->fetch_size */ 32 / 16);
1157         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
1158
1159         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
1160
1161 #if __OS_HAS_AGP
1162         if (dev_priv->flags & RADEON_IS_AGP) {
1163                 /* Turn off PCI GART */
1164                 radeon_set_pcigart(dev_priv, 0);
1165         } else
1166 #endif
1167         {
1168                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
1169                 /* if we have an offset set from userspace */
1170                 if (dev_priv->pcigart_offset_set) {
1171                         dev_priv->gart_info.bus_addr =
1172                             dev_priv->pcigart_offset + dev_priv->fb_location;
1173                         dev_priv->gart_info.mapping.offset =
1174                             dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
1175                         dev_priv->gart_info.mapping.size =
1176                             dev_priv->gart_info.table_size;
1177
1178                         drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
1179                         dev_priv->gart_info.addr =
1180                             dev_priv->gart_info.mapping.handle;
1181
1182                         if (dev_priv->flags & RADEON_IS_PCIE)
1183                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
1184                         else
1185                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1186                         dev_priv->gart_info.gart_table_location =
1187                             DRM_ATI_GART_FB;
1188
1189                         DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
1190                                   dev_priv->gart_info.addr,
1191                                   dev_priv->pcigart_offset);
1192                 } else {
1193                         if (dev_priv->flags & RADEON_IS_IGPGART)
1194                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
1195                         else
1196                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1197                         dev_priv->gart_info.gart_table_location =
1198                             DRM_ATI_GART_MAIN;
1199                         dev_priv->gart_info.addr = NULL;
1200                         dev_priv->gart_info.bus_addr = 0;
1201                         if (dev_priv->flags & RADEON_IS_PCIE) {
1202                                 DRM_ERROR
1203                                     ("Cannot use PCI Express without GART in FB memory\n");
1204                                 radeon_do_cleanup_cp(dev);
1205                                 return -EINVAL;
1206                         }
1207                 }
1208
1209                 if (!drm_ati_pcigart_init(dev, &dev_priv->gart_info)) {
1210                         DRM_ERROR("failed to init PCI GART!\n");
1211                         radeon_do_cleanup_cp(dev);
1212                         return -ENOMEM;
1213                 }
1214
1215                 /* Turn on PCI GART */
1216                 radeon_set_pcigart(dev_priv, 1);
1217         }
1218
1219         radeon_cp_load_microcode(dev_priv);
1220         radeon_cp_init_ring_buffer(dev, dev_priv);
1221
1222         dev_priv->last_buf = 0;
1223
1224         radeon_do_engine_reset(dev);
1225         radeon_test_writeback(dev_priv);
1226
1227         return 0;
1228 }
1229
1230 static int radeon_do_cleanup_cp(struct drm_device * dev)
1231 {
1232         drm_radeon_private_t *dev_priv = dev->dev_private;
1233         DRM_DEBUG("\n");
1234
1235         /* Make sure interrupts are disabled here because the uninstall ioctl
1236          * may not have been called from userspace and after dev_private
1237          * is freed, it's too late.
1238          */
1239         if (dev->irq_enabled)
1240                 drm_irq_uninstall(dev);
1241
1242 #if __OS_HAS_AGP
1243         if (dev_priv->flags & RADEON_IS_AGP) {
1244                 if (dev_priv->cp_ring != NULL) {
1245                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1246                         dev_priv->cp_ring = NULL;
1247                 }
1248                 if (dev_priv->ring_rptr != NULL) {
1249                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1250                         dev_priv->ring_rptr = NULL;
1251                 }
1252                 if (dev->agp_buffer_map != NULL) {
1253                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1254                         dev->agp_buffer_map = NULL;
1255                 }
1256         } else
1257 #endif
1258         {
1259
1260                 if (dev_priv->gart_info.bus_addr) {
1261                         /* Turn off PCI GART */
1262                         radeon_set_pcigart(dev_priv, 0);
1263                         if (!drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info))
1264                                 DRM_ERROR("failed to cleanup PCI GART!\n");
1265                 }
1266
1267                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB)
1268                 {
1269                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1270                         dev_priv->gart_info.addr = 0;
1271                 }
1272         }
1273         /* only clear to the start of flags */
1274         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1275
1276         return 0;
1277 }
1278
1279 /* This code will reinit the Radeon CP hardware after a resume from disc.
1280  * AFAIK, it would be very difficult to pickle the state at suspend time, so
1281  * here we make sure that all Radeon hardware initialisation is re-done without
1282  * affecting running applications.
1283  *
1284  * Charl P. Botha <http://cpbotha.net>
1285  */
1286 static int radeon_do_resume_cp(struct drm_device * dev)
1287 {
1288         drm_radeon_private_t *dev_priv = dev->dev_private;
1289
1290         if (!dev_priv) {
1291                 DRM_ERROR("Called with no initialization\n");
1292                 return -EINVAL;
1293         }
1294
1295         DRM_DEBUG("Starting radeon_do_resume_cp()\n");
1296
1297 #if __OS_HAS_AGP
1298         if (dev_priv->flags & RADEON_IS_AGP) {
1299                 /* Turn off PCI GART */
1300                 radeon_set_pcigart(dev_priv, 0);
1301         } else
1302 #endif
1303         {
1304                 /* Turn on PCI GART */
1305                 radeon_set_pcigart(dev_priv, 1);
1306         }
1307
1308         radeon_cp_load_microcode(dev_priv);
1309         radeon_cp_init_ring_buffer(dev, dev_priv);
1310
1311         radeon_do_engine_reset(dev);
1312         radeon_irq_set_state(dev, RADEON_SW_INT_ENABLE, 1);
1313
1314         DRM_DEBUG("radeon_do_resume_cp() complete\n");
1315
1316         return 0;
1317 }
1318
1319 int radeon_cp_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
1320 {
1321         drm_radeon_init_t *init = data;
1322
1323         LOCK_TEST_WITH_RETURN(dev, file_priv);
1324
1325         if (init->func == RADEON_INIT_R300_CP)
1326                 r300_init_reg_flags(dev);
1327
1328         switch (init->func) {
1329         case RADEON_INIT_CP:
1330         case RADEON_INIT_R200_CP:
1331         case RADEON_INIT_R300_CP:
1332                 return radeon_do_init_cp(dev, init);
1333         case RADEON_CLEANUP_CP:
1334                 return radeon_do_cleanup_cp(dev);
1335         }
1336
1337         return -EINVAL;
1338 }
1339
1340 int radeon_cp_start(struct drm_device *dev, void *data, struct drm_file *file_priv)
1341 {
1342         drm_radeon_private_t *dev_priv = dev->dev_private;
1343         DRM_DEBUG("\n");
1344
1345         LOCK_TEST_WITH_RETURN(dev, file_priv);
1346
1347         if (dev_priv->cp_running) {
1348                 DRM_DEBUG("while CP running\n");
1349                 return 0;
1350         }
1351         if (dev_priv->cp_mode == RADEON_CSQ_PRIDIS_INDDIS) {
1352                 DRM_DEBUG("called with bogus CP mode (%d)\n",
1353                           dev_priv->cp_mode);
1354                 return 0;
1355         }
1356
1357         radeon_do_cp_start(dev_priv);
1358
1359         return 0;
1360 }
1361
1362 /* Stop the CP.  The engine must have been idled before calling this
1363  * routine.
1364  */
1365 int radeon_cp_stop(struct drm_device *dev, void *data, struct drm_file *file_priv)
1366 {
1367         drm_radeon_private_t *dev_priv = dev->dev_private;
1368         drm_radeon_cp_stop_t *stop = data;
1369         int ret;
1370         DRM_DEBUG("\n");
1371
1372         LOCK_TEST_WITH_RETURN(dev, file_priv);
1373
1374         if (!dev_priv->cp_running)
1375                 return 0;
1376
1377         /* Flush any pending CP commands.  This ensures any outstanding
1378          * commands are exectuted by the engine before we turn it off.
1379          */
1380         if (stop->flush) {
1381                 radeon_do_cp_flush(dev_priv);
1382         }
1383
1384         /* If we fail to make the engine go idle, we return an error
1385          * code so that the DRM ioctl wrapper can try again.
1386          */
1387         if (stop->idle) {
1388                 ret = radeon_do_cp_idle(dev_priv);
1389                 if (ret)
1390                         return ret;
1391         }
1392
1393         /* Finally, we can turn off the CP.  If the engine isn't idle,
1394          * we will get some dropped triangles as they won't be fully
1395          * rendered before the CP is shut down.
1396          */
1397         radeon_do_cp_stop(dev_priv);
1398
1399         /* Reset the engine */
1400         radeon_do_engine_reset(dev);
1401
1402         return 0;
1403 }
1404
1405 void radeon_do_release(struct drm_device * dev)
1406 {
1407         drm_radeon_private_t *dev_priv = dev->dev_private;
1408         int i, ret;
1409
1410         if (dev_priv) {
1411                 if (dev_priv->cp_running) {
1412                         /* Stop the cp */
1413                         while ((ret = radeon_do_cp_idle(dev_priv)) != 0) {
1414                                 DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1415 #ifdef __linux__
1416                                 schedule();
1417 #else
1418                                 tsleep(&ret, PZERO, "rdnrel", 1);
1419 #endif
1420                         }
1421                         radeon_do_cp_stop(dev_priv);
1422                         radeon_do_engine_reset(dev);
1423                 }
1424
1425                 /* Disable *all* interrupts */
1426                 if (dev_priv->mmio)     /* remove this after permanent addmaps */
1427                         RADEON_WRITE(RADEON_GEN_INT_CNTL, 0);
1428
1429                 if (dev_priv->mmio) {   /* remove all surfaces */
1430                         for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1431                                 RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, 0);
1432                                 RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND +
1433                                              16 * i, 0);
1434                                 RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND +
1435                                              16 * i, 0);
1436                         }
1437                 }
1438
1439                 /* Free memory heap structures */
1440                 radeon_mem_takedown(&(dev_priv->gart_heap));
1441                 radeon_mem_takedown(&(dev_priv->fb_heap));
1442
1443                 /* deallocate kernel resources */
1444                 radeon_do_cleanup_cp(dev);
1445         }
1446 }
1447
1448 /* Just reset the CP ring.  Called as part of an X Server engine reset.
1449  */
1450 int radeon_cp_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1451 {
1452         drm_radeon_private_t *dev_priv = dev->dev_private;
1453         DRM_DEBUG("\n");
1454
1455         LOCK_TEST_WITH_RETURN(dev, file_priv);
1456
1457         if (!dev_priv) {
1458                 DRM_DEBUG("called before init done\n");
1459                 return -EINVAL;
1460         }
1461
1462         radeon_do_cp_reset(dev_priv);
1463
1464         /* The CP is no longer running after an engine reset */
1465         dev_priv->cp_running = 0;
1466
1467         return 0;
1468 }
1469
1470 int radeon_cp_idle(struct drm_device *dev, void *data, struct drm_file *file_priv)
1471 {
1472         drm_radeon_private_t *dev_priv = dev->dev_private;
1473         DRM_DEBUG("\n");
1474
1475         LOCK_TEST_WITH_RETURN(dev, file_priv);
1476
1477         return radeon_do_cp_idle(dev_priv);
1478 }
1479
1480 /* Added by Charl P. Botha to call radeon_do_resume_cp().
1481  */
1482 int radeon_cp_resume(struct drm_device *dev, void *data, struct drm_file *file_priv)
1483 {
1484
1485         return radeon_do_resume_cp(dev);
1486 }
1487
1488 int radeon_engine_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1489 {
1490         DRM_DEBUG("\n");
1491
1492         LOCK_TEST_WITH_RETURN(dev, file_priv);
1493
1494         return radeon_do_engine_reset(dev);
1495 }
1496
1497 /* ================================================================
1498  * Fullscreen mode
1499  */
1500
1501 /* KW: Deprecated to say the least:
1502  */
1503 int radeon_fullscreen(struct drm_device *dev, void *data, struct drm_file *file_priv)
1504 {
1505         return 0;
1506 }
1507
1508 /* ================================================================
1509  * Freelist management
1510  */
1511
1512 /* Original comment: FIXME: ROTATE_BUFS is a hack to cycle through
1513  *   bufs until freelist code is used.  Note this hides a problem with
1514  *   the scratch register * (used to keep track of last buffer
1515  *   completed) being written to before * the last buffer has actually
1516  *   completed rendering.
1517  *
1518  * KW:  It's also a good way to find free buffers quickly.
1519  *
1520  * KW: Ideally this loop wouldn't exist, and freelist_get wouldn't
1521  * sleep.  However, bugs in older versions of radeon_accel.c mean that
1522  * we essentially have to do this, else old clients will break.
1523  *
1524  * However, it does leave open a potential deadlock where all the
1525  * buffers are held by other clients, which can't release them because
1526  * they can't get the lock.
1527  */
1528
1529 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1530 {
1531         struct drm_device_dma *dma = dev->dma;
1532         drm_radeon_private_t *dev_priv = dev->dev_private;
1533         drm_radeon_buf_priv_t *buf_priv;
1534         struct drm_buf *buf;
1535         int i, t;
1536         int start;
1537
1538         if (++dev_priv->last_buf >= dma->buf_count)
1539                 dev_priv->last_buf = 0;
1540
1541         start = dev_priv->last_buf;
1542
1543         for (t = 0; t < dev_priv->usec_timeout; t++) {
1544                 u32 done_age = GET_SCRATCH(1);
1545                 DRM_DEBUG("done_age = %d\n", done_age);
1546                 for (i = start; i < dma->buf_count; i++) {
1547                         buf = dma->buflist[i];
1548                         buf_priv = buf->dev_private;
1549                         if (buf->file_priv == NULL || (buf->pending &&
1550                                                        buf_priv->age <=
1551                                                        done_age)) {
1552                                 dev_priv->stats.requested_bufs++;
1553                                 buf->pending = 0;
1554                                 return buf;
1555                         }
1556                         start = 0;
1557                 }
1558
1559                 if (t) {
1560                         DRM_UDELAY(1);
1561                         dev_priv->stats.freelist_loops++;
1562                 }
1563         }
1564
1565         DRM_DEBUG("returning NULL!\n");
1566         return NULL;
1567 }
1568
1569 #if 0
1570 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1571 {
1572         struct drm_device_dma *dma = dev->dma;
1573         drm_radeon_private_t *dev_priv = dev->dev_private;
1574         drm_radeon_buf_priv_t *buf_priv;
1575         struct drm_buf *buf;
1576         int i, t;
1577         int start;
1578         u32 done_age = DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1));
1579
1580         if (++dev_priv->last_buf >= dma->buf_count)
1581                 dev_priv->last_buf = 0;
1582
1583         start = dev_priv->last_buf;
1584         dev_priv->stats.freelist_loops++;
1585
1586         for (t = 0; t < 2; t++) {
1587                 for (i = start; i < dma->buf_count; i++) {
1588                         buf = dma->buflist[i];
1589                         buf_priv = buf->dev_private;
1590                         if (buf->file_priv == 0 || (buf->pending &&
1591                                                     buf_priv->age <=
1592                                                     done_age)) {
1593                                 dev_priv->stats.requested_bufs++;
1594                                 buf->pending = 0;
1595                                 return buf;
1596                         }
1597                 }
1598                 start = 0;
1599         }
1600
1601         return NULL;
1602 }
1603 #endif
1604
1605 void radeon_freelist_reset(struct drm_device * dev)
1606 {
1607         struct drm_device_dma *dma = dev->dma;
1608         drm_radeon_private_t *dev_priv = dev->dev_private;
1609         int i;
1610
1611         dev_priv->last_buf = 0;
1612         for (i = 0; i < dma->buf_count; i++) {
1613                 struct drm_buf *buf = dma->buflist[i];
1614                 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
1615                 buf_priv->age = 0;
1616         }
1617 }
1618
1619 /* ================================================================
1620  * CP command submission
1621  */
1622
1623 int radeon_wait_ring(drm_radeon_private_t * dev_priv, int n)
1624 {
1625         drm_radeon_ring_buffer_t *ring = &dev_priv->ring;
1626         int i;
1627         u32 last_head = GET_RING_HEAD(dev_priv);
1628
1629         for (i = 0; i < dev_priv->usec_timeout; i++) {
1630                 u32 head = GET_RING_HEAD(dev_priv);
1631
1632                 ring->space = (head - ring->tail) * sizeof(u32);
1633                 if (ring->space <= 0)
1634                         ring->space += ring->size;
1635                 if (ring->space > n)
1636                         return 0;
1637
1638                 dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
1639
1640                 if (head != last_head)
1641                         i = 0;
1642                 last_head = head;
1643
1644                 DRM_UDELAY(1);
1645         }
1646
1647         /* FIXME: This return value is ignored in the BEGIN_RING macro! */
1648 #if RADEON_FIFO_DEBUG
1649         radeon_status(dev_priv);
1650         DRM_ERROR("failed!\n");
1651 #endif
1652         return -EBUSY;
1653 }
1654
1655 static int radeon_cp_get_buffers(struct drm_device *dev,
1656                                  struct drm_file *file_priv,
1657                                  struct drm_dma * d)
1658 {
1659         int i;
1660         struct drm_buf *buf;
1661
1662         for (i = d->granted_count; i < d->request_count; i++) {
1663                 buf = radeon_freelist_get(dev);
1664                 if (!buf)
1665                         return -EBUSY;  /* NOTE: broken client */
1666
1667                 buf->file_priv = file_priv;
1668
1669                 if (DRM_COPY_TO_USER(&d->request_indices[i], &buf->idx,
1670                                      sizeof(buf->idx)))
1671                         return -EFAULT;
1672                 if (DRM_COPY_TO_USER(&d->request_sizes[i], &buf->total,
1673                                      sizeof(buf->total)))
1674                         return -EFAULT;
1675
1676                 d->granted_count++;
1677         }
1678         return 0;
1679 }
1680
1681 int radeon_cp_buffers(struct drm_device *dev, void *data, struct drm_file *file_priv)
1682 {
1683         struct drm_device_dma *dma = dev->dma;
1684         int ret = 0;
1685         struct drm_dma *d = data;
1686
1687         LOCK_TEST_WITH_RETURN(dev, file_priv);
1688
1689         /* Please don't send us buffers.
1690          */
1691         if (d->send_count != 0) {
1692                 DRM_ERROR("Process %d trying to send %d buffers via drmDMA\n",
1693                           DRM_CURRENTPID, d->send_count);
1694                 return -EINVAL;
1695         }
1696
1697         /* We'll send you buffers.
1698          */
1699         if (d->request_count < 0 || d->request_count > dma->buf_count) {
1700                 DRM_ERROR("Process %d trying to get %d buffers (of %d max)\n",
1701                           DRM_CURRENTPID, d->request_count, dma->buf_count);
1702                 return -EINVAL;
1703         }
1704
1705         d->granted_count = 0;
1706
1707         if (d->request_count) {
1708                 ret = radeon_cp_get_buffers(dev, file_priv, d);
1709         }
1710
1711         return ret;
1712 }
1713
1714 int radeon_driver_load(struct drm_device *dev, unsigned long flags)
1715 {
1716         drm_radeon_private_t *dev_priv;
1717         int ret = 0;
1718
1719         dev_priv = drm_alloc(sizeof(drm_radeon_private_t), DRM_MEM_DRIVER);
1720         if (dev_priv == NULL)
1721                 return -ENOMEM;
1722
1723         memset(dev_priv, 0, sizeof(drm_radeon_private_t));
1724         dev->dev_private = (void *)dev_priv;
1725         dev_priv->flags = flags;
1726
1727         switch (flags & RADEON_FAMILY_MASK) {
1728         case CHIP_R100:
1729         case CHIP_RV200:
1730         case CHIP_R200:
1731         case CHIP_R300:
1732         case CHIP_R350:
1733         case CHIP_R420:
1734         case CHIP_R423:
1735         case CHIP_RV410:
1736         case CHIP_RV515:
1737         case CHIP_R520:
1738         case CHIP_RV570:
1739         case CHIP_R580:
1740                 dev_priv->flags |= RADEON_HAS_HIERZ;
1741                 break;
1742         default:
1743                 /* all other chips have no hierarchical z buffer */
1744                 break;
1745         }
1746
1747         if (drm_device_is_agp(dev))
1748                 dev_priv->flags |= RADEON_IS_AGP;
1749         else if (drm_device_is_pcie(dev))
1750                 dev_priv->flags |= RADEON_IS_PCIE;
1751         else
1752                 dev_priv->flags |= RADEON_IS_PCI;
1753
1754         DRM_DEBUG("%s card detected\n",
1755                   ((dev_priv->flags & RADEON_IS_AGP) ? "AGP" : (((dev_priv->flags & RADEON_IS_PCIE) ? "PCIE" : "PCI"))));
1756         return ret;
1757 }
1758
1759 /* Create mappings for registers and framebuffer so userland doesn't necessarily
1760  * have to find them.
1761  */
1762 int radeon_driver_firstopen(struct drm_device *dev)
1763 {
1764         int ret;
1765         drm_local_map_t *map;
1766         drm_radeon_private_t *dev_priv = dev->dev_private;
1767
1768         dev_priv->gart_info.table_size = RADEON_PCIGART_TABLE_SIZE;
1769
1770         ret = drm_addmap(dev, drm_get_resource_start(dev, 2),
1771                          drm_get_resource_len(dev, 2), _DRM_REGISTERS,
1772                          _DRM_READ_ONLY, &dev_priv->mmio);
1773         if (ret != 0)
1774                 return ret;
1775
1776         dev_priv->fb_aper_offset = drm_get_resource_start(dev, 0);
1777         ret = drm_addmap(dev, dev_priv->fb_aper_offset,
1778                          drm_get_resource_len(dev, 0), _DRM_FRAME_BUFFER,
1779                          _DRM_WRITE_COMBINING, &map);
1780         if (ret != 0)
1781                 return ret;
1782
1783         return 0;
1784 }
1785
1786 int radeon_driver_unload(struct drm_device *dev)
1787 {
1788         drm_radeon_private_t *dev_priv = dev->dev_private;
1789
1790         DRM_DEBUG("\n");
1791         drm_free(dev_priv, sizeof(*dev_priv), DRM_MEM_DRIVER);
1792
1793         dev->dev_private = NULL;
1794         return 0;
1795 }