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 "drm_sarea.h"
  35 #include "radeon_drm.h"
  36 #include "radeon_drv.h"
  37 #include "r300_reg.h"
  38
  39 #include "radeon_microcode.h"
  40
  41 #define RADEON_FIFO_DEBUG       0
  42
  43 static int radeon_do_cleanup_cp(struct drm_device * dev);
  44 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv);
  45
  46 static u32 radeon_read_ring_rptr(drm_radeon_private_t *dev_priv, u32 off)
  47 {
  48         u32 val;
  49
  50         if (dev_priv->flags & RADEON_IS_AGP) {
  51                 val = DRM_READ32(dev_priv->ring_rptr, off);
  52         } else {
  53                 val = *(((volatile u32 *)
  54                          dev_priv->ring_rptr->handle) +
  55                         (off / sizeof(u32)));
  56                 val = le32_to_cpu(val);
  57         }
  58         return val;
  59 }
  60
  61 u32 radeon_get_ring_head(drm_radeon_private_t *dev_priv)
  62 {
  63         if (dev_priv->writeback_works)
  64                 return radeon_read_ring_rptr(dev_priv, 0);
  65         else
  66                 return RADEON_READ(RADEON_CP_RB_RPTR);
  67 }
  68
  69 static void radeon_write_ring_rptr(drm_radeon_private_t *dev_priv, u32 off, u32 val)
  70 {
  71         if (dev_priv->flags & RADEON_IS_AGP)
  72                 DRM_WRITE32(dev_priv->ring_rptr, off, val);
  73         else
  74                 *(((volatile u32 *) dev_priv->ring_rptr->handle) +
  75                   (off / sizeof(u32))) = cpu_to_le32(val);
  76 }
  77
  78 void radeon_set_ring_head(drm_radeon_private_t *dev_priv, u32 val)
  79 {
  80         radeon_write_ring_rptr(dev_priv, 0, val);
  81 }
  82
  83 u32 radeon_get_scratch(drm_radeon_private_t *dev_priv, int index)
  84 {
  85         if (dev_priv->writeback_works)
  86                 return radeon_read_ring_rptr(dev_priv,
  87                                              RADEON_SCRATCHOFF(index));
  88         else
  89                 return RADEON_READ(RADEON_SCRATCH_REG0 + 4*index);
  90 }
  91
  92 static u32 R500_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  93 {
  94         u32 ret;
  95         RADEON_WRITE(R520_MC_IND_INDEX, 0x7f0000 | (addr & 0xff));
  96         ret = RADEON_READ(R520_MC_IND_DATA);
  97         RADEON_WRITE(R520_MC_IND_INDEX, 0);
  98         return ret;
  99 }
 100
 101 static u32 RS480_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 102 {
 103         u32 ret;
 104         RADEON_WRITE(RS480_NB_MC_INDEX, addr & 0xff);
 105         ret = RADEON_READ(RS480_NB_MC_DATA);
 106         RADEON_WRITE(RS480_NB_MC_INDEX, 0xff);
 107         return ret;
 108 }
 109
 110 static u32 RS690_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 111 {
 112         u32 ret;
 113         RADEON_WRITE(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
 114         ret = RADEON_READ(RS690_MC_DATA);
 115         RADEON_WRITE(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
 116         return ret;
 117 }
 118
 119 static u32 IGP_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 120 {
 121         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 122             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 123                 return RS690_READ_MCIND(dev_priv, addr);
 124         else
 125                 return RS480_READ_MCIND(dev_priv, addr);
 126 }
 127
 128 u32 radeon_read_fb_location(drm_radeon_private_t *dev_priv)
 129 {
 130
 131         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 132                 return R500_READ_MCIND(dev_priv, RV515_MC_FB_LOCATION);
 133         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 134                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 135                 return RS690_READ_MCIND(dev_priv, RS690_MC_FB_LOCATION);
 136         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 137                 return R500_READ_MCIND(dev_priv, R520_MC_FB_LOCATION);
 138         else
 139                 return RADEON_READ(RADEON_MC_FB_LOCATION);
 140 }
 141
 142 static void radeon_write_fb_location(drm_radeon_private_t *dev_priv, u32 fb_loc)
 143 {
 144         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 145                 R500_WRITE_MCIND(RV515_MC_FB_LOCATION, fb_loc);
 146         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 147                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 148                 RS690_WRITE_MCIND(RS690_MC_FB_LOCATION, fb_loc);
 149         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 150                 R500_WRITE_MCIND(R520_MC_FB_LOCATION, fb_loc);
 151         else
 152                 RADEON_WRITE(RADEON_MC_FB_LOCATION, fb_loc);
 153 }
 154
 155 static void radeon_write_agp_location(drm_radeon_private_t *dev_priv, u32 agp_loc)
 156 {
 157         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 158                 R500_WRITE_MCIND(RV515_MC_AGP_LOCATION, agp_loc);
 159         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 160                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 161                 RS690_WRITE_MCIND(RS690_MC_AGP_LOCATION, agp_loc);
 162         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 163                 R500_WRITE_MCIND(R520_MC_AGP_LOCATION, agp_loc);
 164         else
 165                 RADEON_WRITE(RADEON_MC_AGP_LOCATION, agp_loc);
 166 }
 167
 168 static void radeon_write_agp_base(drm_radeon_private_t *dev_priv, u64 agp_base)
 169 {
 170         u32 agp_base_hi = upper_32_bits(agp_base);
 171         u32 agp_base_lo = agp_base & 0xffffffff;
 172
 173         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) {
 174                 R500_WRITE_MCIND(RV515_MC_AGP_BASE, agp_base_lo);
 175                 R500_WRITE_MCIND(RV515_MC_AGP_BASE_2, agp_base_hi);
 176         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 177                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 178                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE, agp_base_lo);
 179                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE_2, agp_base_hi);
 180         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515) {
 181                 R500_WRITE_MCIND(R520_MC_AGP_BASE, agp_base_lo);
 182                 R500_WRITE_MCIND(R520_MC_AGP_BASE_2, agp_base_hi);
 183         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 184                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 185                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 186                 RADEON_WRITE(RS480_AGP_BASE_2, agp_base_hi);
 187         } else {
 188                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 189                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R200)
 190                         RADEON_WRITE(RADEON_AGP_BASE_2, agp_base_hi);
 191         }
 192 }
 193
 194 static int RADEON_READ_PLL(struct drm_device * dev, int addr)
 195 {
 196         drm_radeon_private_t *dev_priv = dev->dev_private;
 197
 198         RADEON_WRITE8(RADEON_CLOCK_CNTL_INDEX, addr & 0x1f);
 199         return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 200 }
 201
 202 static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 203 {
 204         RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 205         return RADEON_READ(RADEON_PCIE_DATA);
 206 }
 207
 208 #if RADEON_FIFO_DEBUG
 209 static void radeon_status(drm_radeon_private_t * dev_priv)
 210 {
 211         printk("%s:\n", __func__);
 212         printk("RBBM_STATUS = 0x%08x\n",
 213                (unsigned int)RADEON_READ(RADEON_RBBM_STATUS));
 214         printk("CP_RB_RTPR = 0x%08x\n",
 215                (unsigned int)RADEON_READ(RADEON_CP_RB_RPTR));
 216         printk("CP_RB_WTPR = 0x%08x\n",
 217                (unsigned int)RADEON_READ(RADEON_CP_RB_WPTR));
 218         printk("AIC_CNTL = 0x%08x\n",
 219                (unsigned int)RADEON_READ(RADEON_AIC_CNTL));
 220         printk("AIC_STAT = 0x%08x\n",
 221                (unsigned int)RADEON_READ(RADEON_AIC_STAT));
 222         printk("AIC_PT_BASE = 0x%08x\n",
 223                (unsigned int)RADEON_READ(RADEON_AIC_PT_BASE));
 224         printk("TLB_ADDR = 0x%08x\n",
 225                (unsigned int)RADEON_READ(RADEON_AIC_TLB_ADDR));
 226         printk("TLB_DATA = 0x%08x\n",
 227                (unsigned int)RADEON_READ(RADEON_AIC_TLB_DATA));
 228 }
 229 #endif
 230
 231 /* ================================================================
 232  * Engine, FIFO control
 233  */
 234
 235 static int radeon_do_pixcache_flush(drm_radeon_private_t * dev_priv)
 236 {
 237         u32 tmp;
 238         int i;
 239
 240         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 241
 242         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV280) {
 243                 tmp = RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT);
 244                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 245                 RADEON_WRITE(RADEON_RB3D_DSTCACHE_CTLSTAT, tmp);
 246
 247                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 248                         if (!(RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT)
 249                               & RADEON_RB3D_DC_BUSY)) {
 250                                 return 0;
 251                         }
 252                         DRM_UDELAY(1);
 253                 }
 254         } else {
 255                 /* don't flush or purge cache here or lockup */
 256                 return 0;
 257         }
 258
 259 #if RADEON_FIFO_DEBUG
 260         DRM_ERROR("failed!\n");
 261         radeon_status(dev_priv);
 262 #endif
 263         return -EBUSY;
 264 }
 265
 266 static int radeon_do_wait_for_fifo(drm_radeon_private_t * dev_priv, int entries)
 267 {
 268         int i;
 269
 270         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 271
 272         for (i = 0; i < dev_priv->usec_timeout; i++) {
 273                 int slots = (RADEON_READ(RADEON_RBBM_STATUS)
 274                              & RADEON_RBBM_FIFOCNT_MASK);
 275                 if (slots >= entries)
 276                         return 0;
 277                 DRM_UDELAY(1);
 278         }
 279         DRM_DEBUG("wait for fifo failed status : 0x%08X 0x%08X\n",
 280                  RADEON_READ(RADEON_RBBM_STATUS),
 281                  RADEON_READ(R300_VAP_CNTL_STATUS));
 282
 283 #if RADEON_FIFO_DEBUG
 284         DRM_ERROR("failed!\n");
 285         radeon_status(dev_priv);
 286 #endif
 287         return -EBUSY;
 288 }
 289
 290 static int radeon_do_wait_for_idle(drm_radeon_private_t * dev_priv)
 291 {
 292         int i, ret;
 293
 294         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 295
 296         ret = radeon_do_wait_for_fifo(dev_priv, 64);
 297         if (ret)
 298                 return ret;
 299
 300         for (i = 0; i < dev_priv->usec_timeout; i++) {
 301                 if (!(RADEON_READ(RADEON_RBBM_STATUS)
 302                       & RADEON_RBBM_ACTIVE)) {
 303                         radeon_do_pixcache_flush(dev_priv);
 304                         return 0;
 305                 }
 306                 DRM_UDELAY(1);
 307         }
 308         DRM_DEBUG("wait idle failed status : 0x%08X 0x%08X\n",
 309                  RADEON_READ(RADEON_RBBM_STATUS),
 310                  RADEON_READ(R300_VAP_CNTL_STATUS));
 311
 312 #if RADEON_FIFO_DEBUG
 313         DRM_ERROR("failed!\n");
 314         radeon_status(dev_priv);
 315 #endif
 316         return -EBUSY;
 317 }
 318
 319 static void radeon_init_pipes(drm_radeon_private_t *dev_priv)
 320 {
 321         uint32_t gb_tile_config, gb_pipe_sel = 0;
 322
 323         /* RS4xx/RS6xx/R4xx/R5xx */
 324         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R420) {
 325                 gb_pipe_sel = RADEON_READ(R400_GB_PIPE_SELECT);
 326                 dev_priv->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
 327         } else {
 328                 /* R3xx */
 329                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 330                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350)) {
 331                         dev_priv->num_gb_pipes = 2;
 332                 } else {
 333                         /* R3Vxx */
 334                         dev_priv->num_gb_pipes = 1;
 335                 }
 336         }
 337         DRM_INFO("Num pipes: %d\n", dev_priv->num_gb_pipes);
 338
 339         gb_tile_config = (R300_ENABLE_TILING | R300_TILE_SIZE_16 /*| R300_SUBPIXEL_1_16*/);
 340
 341         switch (dev_priv->num_gb_pipes) {
 342         case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
 343         case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
 344         case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
 345         default:
 346         case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
 347         }
 348
 349         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
 350                 RADEON_WRITE_PLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
 351                 RADEON_WRITE(R500_SU_REG_DEST, ((1 << dev_priv->num_gb_pipes) - 1));
 352         }
 353         RADEON_WRITE(R300_GB_TILE_CONFIG, gb_tile_config);
 354         radeon_do_wait_for_idle(dev_priv);
 355         RADEON_WRITE(R300_DST_PIPE_CONFIG, RADEON_READ(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
 356         RADEON_WRITE(R300_RB2D_DSTCACHE_MODE, (RADEON_READ(R300_RB2D_DSTCACHE_MODE) |
 357                                                R300_DC_AUTOFLUSH_ENABLE |
 358                                                R300_DC_DC_DISABLE_IGNORE_PE));
 359
 360
 361 }
 362
 363 /* ================================================================
 364  * CP control, initialization
 365  */
 366
 367 /* Load the microcode for the CP */
 368 static void radeon_cp_load_microcode(drm_radeon_private_t * dev_priv)
 369 {
 370         int i;
 371         DRM_DEBUG("\n");
 372
 373         radeon_do_wait_for_idle(dev_priv);
 374
 375         RADEON_WRITE(RADEON_CP_ME_RAM_ADDR, 0);
 376         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R100) ||
 377             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV100) ||
 378             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV200) ||
 379             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS100) ||
 380             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS200)) {
 381                 DRM_INFO("Loading R100 Microcode\n");
 382                 for (i = 0; i < 256; i++) {
 383                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 384                                      R100_cp_microcode[i][1]);
 385                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 386                                      R100_cp_microcode[i][0]);
 387                 }
 388         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R200) ||
 389                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV250) ||
 390                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV280) ||
 391                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS300)) {
 392                 DRM_INFO("Loading R200 Microcode\n");
 393                 for (i = 0; i < 256; i++) {
 394                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 395                                      R200_cp_microcode[i][1]);
 396                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 397                                      R200_cp_microcode[i][0]);
 398                 }
 399         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 400                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350) ||
 401                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV350) ||
 402                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV380) ||
 403                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 404                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 405                 DRM_INFO("Loading R300 Microcode\n");
 406                 for (i = 0; i < 256; i++) {
 407                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 408                                      R300_cp_microcode[i][1]);
 409                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 410                                      R300_cp_microcode[i][0]);
 411                 }
 412         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 413                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R423) ||
 414                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV410)) {
 415                 DRM_INFO("Loading R400 Microcode\n");
 416                 for (i = 0; i < 256; i++) {
 417                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 418                                      R420_cp_microcode[i][1]);
 419                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 420                                      R420_cp_microcode[i][0]);
 421                 }
 422         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 423                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 424                 DRM_INFO("Loading RS690/RS740 Microcode\n");
 425                 for (i = 0; i < 256; i++) {
 426                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 427                                      RS690_cp_microcode[i][1]);
 428                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 429                                      RS690_cp_microcode[i][0]);
 430                 }
 431         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) ||
 432                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R520) ||
 433                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530) ||
 434                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R580) ||
 435                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV560) ||
 436                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV570)) {
 437                 DRM_INFO("Loading R500 Microcode\n");
 438                 for (i = 0; i < 256; i++) {
 439                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 440                                      R520_cp_microcode[i][1]);
 441                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 442                                      R520_cp_microcode[i][0]);
 443                 }
 444         }
 445 }
 446
 447 /* Flush any pending commands to the CP.  This should only be used just
 448  * prior to a wait for idle, as it informs the engine that the command
 449  * stream is ending.
 450  */
 451 static void radeon_do_cp_flush(drm_radeon_private_t * dev_priv)
 452 {
 453         DRM_DEBUG("\n");
 454 #if 0
 455         u32 tmp;
 456
 457         tmp = RADEON_READ(RADEON_CP_RB_WPTR) | (1 << 31);
 458         RADEON_WRITE(RADEON_CP_RB_WPTR, tmp);
 459 #endif
 460 }
 461
 462 /* Wait for the CP to go idle.
 463  */
 464 int radeon_do_cp_idle(drm_radeon_private_t * dev_priv)
 465 {
 466         RING_LOCALS;
 467         DRM_DEBUG("\n");
 468
 469         BEGIN_RING(6);
 470
 471         RADEON_PURGE_CACHE();
 472         RADEON_PURGE_ZCACHE();
 473         RADEON_WAIT_UNTIL_IDLE();
 474
 475         ADVANCE_RING();
 476         COMMIT_RING();
 477
 478         return radeon_do_wait_for_idle(dev_priv);
 479 }
 480
 481 /* Start the Command Processor.
 482  */
 483 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 484 {
 485         RING_LOCALS;
 486         DRM_DEBUG("\n");
 487
 488         radeon_do_wait_for_idle(dev_priv);
 489
 490         RADEON_WRITE(RADEON_CP_CSQ_CNTL, dev_priv->cp_mode);
 491
 492         dev_priv->cp_running = 1;
 493
 494         BEGIN_RING(8);
 495         /* isync can only be written through cp on r5xx write it here */
 496         OUT_RING(CP_PACKET0(RADEON_ISYNC_CNTL, 0));
 497         OUT_RING(RADEON_ISYNC_ANY2D_IDLE3D |
 498                  RADEON_ISYNC_ANY3D_IDLE2D |
 499                  RADEON_ISYNC_WAIT_IDLEGUI |
 500                  RADEON_ISYNC_CPSCRATCH_IDLEGUI);
 501         RADEON_PURGE_CACHE();
 502         RADEON_PURGE_ZCACHE();
 503         RADEON_WAIT_UNTIL_IDLE();
 504         ADVANCE_RING();
 505         COMMIT_RING();
 506
 507         dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
 508 }
 509
 510 /* Reset the Command Processor.  This will not flush any pending
 511  * commands, so you must wait for the CP command stream to complete
 512  * before calling this routine.
 513  */
 514 static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
 515 {
 516         u32 cur_read_ptr;
 517         DRM_DEBUG("\n");
 518
 519         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 520         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 521         SET_RING_HEAD(dev_priv, cur_read_ptr);
 522         dev_priv->ring.tail = cur_read_ptr;
 523 }
 524
 525 /* Stop the Command Processor.  This will not flush any pending
 526  * commands, so you must flush the command stream and wait for the CP
 527  * to go idle before calling this routine.
 528  */
 529 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 530 {
 531         DRM_DEBUG("\n");
 532
 533         RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 534
 535         dev_priv->cp_running = 0;
 536 }
 537
 538 /* Reset the engine.  This will stop the CP if it is running.
 539  */
 540 static int radeon_do_engine_reset(struct drm_device * dev)
 541 {
 542         drm_radeon_private_t *dev_priv = dev->dev_private;
 543         u32 clock_cntl_index = 0, mclk_cntl = 0, rbbm_soft_reset;
 544         DRM_DEBUG("\n");
 545
 546         radeon_do_pixcache_flush(dev_priv);
 547
 548         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 549                 /* may need something similar for newer chips */
 550                 clock_cntl_index = RADEON_READ(RADEON_CLOCK_CNTL_INDEX);
 551                 mclk_cntl = RADEON_READ_PLL(dev, RADEON_MCLK_CNTL);
 552
 553                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, (mclk_cntl |
 554                                                     RADEON_FORCEON_MCLKA |
 555                                                     RADEON_FORCEON_MCLKB |
 556                                                     RADEON_FORCEON_YCLKA |
 557                                                     RADEON_FORCEON_YCLKB |
 558                                                     RADEON_FORCEON_MC |
 559                                                     RADEON_FORCEON_AIC));
 560         }
 561
 562         rbbm_soft_reset = RADEON_READ(RADEON_RBBM_SOFT_RESET);
 563
 564         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset |
 565                                               RADEON_SOFT_RESET_CP |
 566                                               RADEON_SOFT_RESET_HI |
 567                                               RADEON_SOFT_RESET_SE |
 568                                               RADEON_SOFT_RESET_RE |
 569                                               RADEON_SOFT_RESET_PP |
 570                                               RADEON_SOFT_RESET_E2 |
 571                                               RADEON_SOFT_RESET_RB));
 572         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 573         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset &
 574                                               ~(RADEON_SOFT_RESET_CP |
 575                                                 RADEON_SOFT_RESET_HI |
 576                                                 RADEON_SOFT_RESET_SE |
 577                                                 RADEON_SOFT_RESET_RE |
 578                                                 RADEON_SOFT_RESET_PP |
 579                                                 RADEON_SOFT_RESET_E2 |
 580                                                 RADEON_SOFT_RESET_RB)));
 581         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 582
 583         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 584                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, mclk_cntl);
 585                 RADEON_WRITE(RADEON_CLOCK_CNTL_INDEX, clock_cntl_index);
 586                 RADEON_WRITE(RADEON_RBBM_SOFT_RESET, rbbm_soft_reset);
 587         }
 588
 589         /* setup the raster pipes */
 590         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R300)
 591             radeon_init_pipes(dev_priv);
 592
 593         /* Reset the CP ring */
 594         radeon_do_cp_reset(dev_priv);
 595
 596         /* The CP is no longer running after an engine reset */
 597         dev_priv->cp_running = 0;
 598
 599         /* Reset any pending vertex, indirect buffers */
 600         radeon_freelist_reset(dev);
 601
 602         return 0;
 603 }
 604
 605 static void radeon_cp_init_ring_buffer(struct drm_device * dev,
 606                                        drm_radeon_private_t *dev_priv,
 607                                        struct drm_file *file_priv)
 608 {
 609         struct drm_radeon_master_private *master_priv;
 610         u32 ring_start, cur_read_ptr;
 611         u32 tmp;
 612
 613         /* Initialize the memory controller. With new memory map, the fb location
 614          * is not changed, it should have been properly initialized already. Part
 615          * of the problem is that the code below is bogus, assuming the GART is
 616          * always appended to the fb which is not necessarily the case
 617          */
 618         if (!dev_priv->new_memmap)
 619                 radeon_write_fb_location(dev_priv,
 620                              ((dev_priv->gart_vm_start - 1) & 0xffff0000)
 621                              | (dev_priv->fb_location >> 16));
 622
 623 #if __OS_HAS_AGP
 624         if (dev_priv->flags & RADEON_IS_AGP) {
 625                 radeon_write_agp_base(dev_priv, dev->agp->base);
 626
 627                 radeon_write_agp_location(dev_priv,
 628                              (((dev_priv->gart_vm_start - 1 +
 629                                 dev_priv->gart_size) & 0xffff0000) |
 630                               (dev_priv->gart_vm_start >> 16)));
 631
 632                 ring_start = (dev_priv->cp_ring->offset
 633                               - dev->agp->base
 634                               + dev_priv->gart_vm_start);
 635         } else
 636 #endif
 637                 ring_start = (dev_priv->cp_ring->offset
 638                               - (unsigned long)dev->sg->virtual
 639                               + dev_priv->gart_vm_start);
 640
 641         RADEON_WRITE(RADEON_CP_RB_BASE, ring_start);
 642
 643         /* Set the write pointer delay */
 644         RADEON_WRITE(RADEON_CP_RB_WPTR_DELAY, 0);
 645
 646         /* Initialize the ring buffer's read and write pointers */
 647         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 648         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 649         SET_RING_HEAD(dev_priv, cur_read_ptr);
 650         dev_priv->ring.tail = cur_read_ptr;
 651
 652 #if __OS_HAS_AGP
 653         if (dev_priv->flags & RADEON_IS_AGP) {
 654                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 655                              dev_priv->ring_rptr->offset
 656                              - dev->agp->base + dev_priv->gart_vm_start);
 657         } else
 658 #endif
 659         {
 660                 struct drm_sg_mem *entry = dev->sg;
 661                 unsigned long tmp_ofs, page_ofs;
 662
 663                 tmp_ofs = dev_priv->ring_rptr->offset -
 664                                 (unsigned long)dev->sg->virtual;
 665                 page_ofs = tmp_ofs >> PAGE_SHIFT;
 666
 667                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR, entry->busaddr[page_ofs]);
 668                 DRM_DEBUG("ring rptr: offset=0x%08lx handle=0x%08lx\n",
 669                           (unsigned long)entry->busaddr[page_ofs],
 670                           entry->handle + tmp_ofs);
 671         }
 672
 673         /* Set ring buffer size */
 674 #ifdef __BIG_ENDIAN
 675         RADEON_WRITE(RADEON_CP_RB_CNTL,
 676                      RADEON_BUF_SWAP_32BIT |
 677                      (dev_priv->ring.fetch_size_l2ow << 18) |
 678                      (dev_priv->ring.rptr_update_l2qw << 8) |
 679                      dev_priv->ring.size_l2qw);
 680 #else
 681         RADEON_WRITE(RADEON_CP_RB_CNTL,
 682                      (dev_priv->ring.fetch_size_l2ow << 18) |
 683                      (dev_priv->ring.rptr_update_l2qw << 8) |
 684                      dev_priv->ring.size_l2qw);
 685 #endif
 686
 687
 688         /* Initialize the scratch register pointer.  This will cause
 689          * the scratch register values to be written out to memory
 690          * whenever they are updated.
 691          *
 692          * We simply put this behind the ring read pointer, this works
 693          * with PCI GART as well as (whatever kind of) AGP GART
 694          */
 695         RADEON_WRITE(RADEON_SCRATCH_ADDR, RADEON_READ(RADEON_CP_RB_RPTR_ADDR)
 696                      + RADEON_SCRATCH_REG_OFFSET);
 697
 698         RADEON_WRITE(RADEON_SCRATCH_UMSK, 0x7);
 699
 700         /* Turn on bus mastering */
 701         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 702             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 703                 /* rs600/rs690/rs740 */
 704                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RS600_BUS_MASTER_DIS;
 705                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 706         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV350) ||
 707                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 708                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 709                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 710                 /* r1xx, r2xx, r300, r(v)350, r420/r481, rs400/rs480 */
 711                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RADEON_BUS_MASTER_DIS;
 712                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 713         } /* PCIE cards appears to not need this */
 714
 715         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(0), 0);
 716         RADEON_WRITE(RADEON_LAST_FRAME_REG, 0);
 717
 718         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1), 0);
 719         RADEON_WRITE(RADEON_LAST_DISPATCH_REG, 0);
 720
 721         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(2), 0);
 722         RADEON_WRITE(RADEON_LAST_CLEAR_REG, 0);
 723
 724         /* reset sarea copies of these */
 725         master_priv = file_priv->master->driver_priv;
 726         if (master_priv->sarea_priv) {
 727                 master_priv->sarea_priv->last_frame = 0;
 728                 master_priv->sarea_priv->last_dispatch = 0;
 729                 master_priv->sarea_priv->last_clear = 0;
 730         }
 731
 732         radeon_do_wait_for_idle(dev_priv);
 733
 734         /* Sync everything up */
 735         RADEON_WRITE(RADEON_ISYNC_CNTL,
 736                      (RADEON_ISYNC_ANY2D_IDLE3D |
 737                       RADEON_ISYNC_ANY3D_IDLE2D |
 738                       RADEON_ISYNC_WAIT_IDLEGUI |
 739                       RADEON_ISYNC_CPSCRATCH_IDLEGUI));
 740
 741 }
 742
 743 static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 744 {
 745         u32 tmp;
 746
 747         /* Start with assuming that writeback doesn't work */
 748         dev_priv->writeback_works = 0;
 749
 750         /* Writeback doesn't seem to work everywhere, test it here and possibly
 751          * enable it if it appears to work
 752          */
 753         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1), 0);
 754
 755         RADEON_WRITE(RADEON_SCRATCH_REG1, 0xdeadbeef);
 756
 757         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 758                 u32 val;
 759
 760                 val = radeon_read_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1));
 761                 if (val == 0xdeadbeef)
 762                         break;
 763                 DRM_UDELAY(1);
 764         }
 765
 766         if (tmp < dev_priv->usec_timeout) {
 767                 dev_priv->writeback_works = 1;
 768                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 769         } else {
 770                 dev_priv->writeback_works = 0;
 771                 DRM_INFO("writeback test failed\n");
 772         }
 773         if (radeon_no_wb == 1) {
 774                 dev_priv->writeback_works = 0;
 775                 DRM_INFO("writeback forced off\n");
 776         }
 777
 778         if (!dev_priv->writeback_works) {
 779                 /* Disable writeback to avoid unnecessary bus master transfer */
 780                 RADEON_WRITE(RADEON_CP_RB_CNTL, RADEON_READ(RADEON_CP_RB_CNTL) |
 781                              RADEON_RB_NO_UPDATE);
 782                 RADEON_WRITE(RADEON_SCRATCH_UMSK, 0);
 783         }
 784 }
 785
 786 /* Enable or disable IGP GART on the chip */
 787 static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
 788 {
 789         u32 temp;
 790
 791         if (on) {
 792                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 793                           dev_priv->gart_vm_start,
 794                           (long)dev_priv->gart_info.bus_addr,
 795                           dev_priv->gart_size);
 796
 797                 temp = IGP_READ_MCIND(dev_priv, RS480_MC_MISC_CNTL);
 798                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 799                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 800                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, (RS480_GART_INDEX_REG_EN |
 801                                                              RS690_BLOCK_GFX_D3_EN));
 802                 else
 803                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, RS480_GART_INDEX_REG_EN);
 804
 805                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 806                                                                RS480_VA_SIZE_32MB));
 807
 808                 temp = IGP_READ_MCIND(dev_priv, RS480_GART_FEATURE_ID);
 809                 IGP_WRITE_MCIND(RS480_GART_FEATURE_ID, (RS480_HANG_EN |
 810                                                         RS480_TLB_ENABLE |
 811                                                         RS480_GTW_LAC_EN |
 812                                                         RS480_1LEVEL_GART));
 813
 814                 temp = dev_priv->gart_info.bus_addr & 0xfffff000;
 815                 temp |= (upper_32_bits(dev_priv->gart_info.bus_addr) & 0xff) << 4;
 816                 IGP_WRITE_MCIND(RS480_GART_BASE, temp);
 817
 818                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_MODE_CNTL);
 819                 IGP_WRITE_MCIND(RS480_AGP_MODE_CNTL, ((1 << RS480_REQ_TYPE_SNOOP_SHIFT) |
 820                                                       RS480_REQ_TYPE_SNOOP_DIS));
 821
 822                 radeon_write_agp_base(dev_priv, dev_priv->gart_vm_start);
 823
 824                 dev_priv->gart_size = 32*1024*1024;
 825                 temp = (((dev_priv->gart_vm_start - 1 + dev_priv->gart_size) &
 826                          0xffff0000) | (dev_priv->gart_vm_start >> 16));
 827
 828                 radeon_write_agp_location(dev_priv, temp);
 829
 830                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_ADDRESS_SPACE_SIZE);
 831                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 832                                                                RS480_VA_SIZE_32MB));
 833
 834                 do {
 835                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 836                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 837                                 break;
 838                         DRM_UDELAY(1);
 839                 } while (1);
 840
 841                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL,
 842                                 RS480_GART_CACHE_INVALIDATE);
 843
 844                 do {
 845                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 846                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 847                                 break;
 848                         DRM_UDELAY(1);
 849                 } while (1);
 850
 851                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL, 0);
 852         } else {
 853                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, 0);
 854         }
 855 }
 856
 857 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
 858 {
 859         u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
 860         if (on) {
 861
 862                 DRM_DEBUG("programming pcie %08X %08lX %08X\n",
 863                           dev_priv->gart_vm_start,
 864                           (long)dev_priv->gart_info.bus_addr,
 865                           dev_priv->gart_size);
 866                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_DISCARD_RD_ADDR_LO,
 867                                   dev_priv->gart_vm_start);
 868                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_BASE,
 869                                   dev_priv->gart_info.bus_addr);
 870                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_START_LO,
 871                                   dev_priv->gart_vm_start);
 872                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_END_LO,
 873                                   dev_priv->gart_vm_start +
 874                                   dev_priv->gart_size - 1);
 875
 876                 radeon_write_agp_location(dev_priv, 0xffffffc0); /* ?? */
 877
 878                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 879                                   RADEON_PCIE_TX_GART_EN);
 880         } else {
 881                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 882                                   tmp & ~RADEON_PCIE_TX_GART_EN);
 883         }
 884 }
 885
 886 /* Enable or disable PCI GART on the chip */
 887 static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
 888 {
 889         u32 tmp;
 890
 891         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 892             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740) ||
 893             (dev_priv->flags & RADEON_IS_IGPGART)) {
 894                 radeon_set_igpgart(dev_priv, on);
 895                 return;
 896         }
 897
 898         if (dev_priv->flags & RADEON_IS_PCIE) {
 899                 radeon_set_pciegart(dev_priv, on);
 900                 return;
 901         }
 902
 903         tmp = RADEON_READ(RADEON_AIC_CNTL);
 904
 905         if (on) {
 906                 RADEON_WRITE(RADEON_AIC_CNTL,
 907                              tmp | RADEON_PCIGART_TRANSLATE_EN);
 908
 909                 /* set PCI GART page-table base address
 910                  */
 911                 RADEON_WRITE(RADEON_AIC_PT_BASE, dev_priv->gart_info.bus_addr);
 912
 913                 /* set address range for PCI address translate
 914                  */
 915                 RADEON_WRITE(RADEON_AIC_LO_ADDR, dev_priv->gart_vm_start);
 916                 RADEON_WRITE(RADEON_AIC_HI_ADDR, dev_priv->gart_vm_start
 917                              + dev_priv->gart_size - 1);
 918
 919                 /* Turn off AGP aperture -- is this required for PCI GART?
 920                  */
 921                 radeon_write_agp_location(dev_priv, 0xffffffc0);
 922                 RADEON_WRITE(RADEON_AGP_COMMAND, 0);    /* clear AGP_COMMAND */
 923         } else {
 924                 RADEON_WRITE(RADEON_AIC_CNTL,
 925                              tmp & ~RADEON_PCIGART_TRANSLATE_EN);
 926         }
 927 }
 928
 929 static int radeon_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
 930                              struct drm_file *file_priv)
 931 {
 932         drm_radeon_private_t *dev_priv = dev->dev_private;
 933         struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
 934
 935         DRM_DEBUG("\n");
 936
 937         /* if we require new memory map but we don't have it fail */
 938         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
 939                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
 940                 radeon_do_cleanup_cp(dev);
 941                 return -EINVAL;
 942         }
 943
 944         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
 945                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
 946                 dev_priv->flags &= ~RADEON_IS_AGP;
 947         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
 948                    && !init->is_pci) {
 949                 DRM_DEBUG("Restoring AGP flag\n");
 950                 dev_priv->flags |= RADEON_IS_AGP;
 951         }
 952
 953         if ((!(dev_priv->flags & RADEON_IS_AGP)) && !dev->sg) {
 954                 DRM_ERROR("PCI GART memory not allocated!\n");
 955                 radeon_do_cleanup_cp(dev);
 956                 return -EINVAL;
 957         }
 958
 959         dev_priv->usec_timeout = init->usec_timeout;
 960         if (dev_priv->usec_timeout < 1 ||
 961             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
 962                 DRM_DEBUG("TIMEOUT problem!\n");
 963                 radeon_do_cleanup_cp(dev);
 964                 return -EINVAL;
 965         }
 966
 967         /* Enable vblank on CRTC1 for older X servers
 968          */
 969         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
 970
 971         switch(init->func) {
 972         case RADEON_INIT_R200_CP:
 973                 dev_priv->microcode_version = UCODE_R200;
 974                 break;
 975         case RADEON_INIT_R300_CP:
 976                 dev_priv->microcode_version = UCODE_R300;
 977                 break;
 978         default:
 979                 dev_priv->microcode_version = UCODE_R100;
 980         }
 981
 982         dev_priv->do_boxes = 0;
 983         dev_priv->cp_mode = init->cp_mode;
 984
 985         /* We don't support anything other than bus-mastering ring mode,
 986          * but the ring can be in either AGP or PCI space for the ring
 987          * read pointer.
 988          */
 989         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
 990             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
 991                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
 992                 radeon_do_cleanup_cp(dev);
 993                 return -EINVAL;
 994         }
 995
 996         switch (init->fb_bpp) {
 997         case 16:
 998                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
 999                 break;
1000         case 32:
1001         default:
1002                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
1003                 break;
1004         }
1005         dev_priv->front_offset = init->front_offset;
1006         dev_priv->front_pitch = init->front_pitch;
1007         dev_priv->back_offset = init->back_offset;
1008         dev_priv->back_pitch = init->back_pitch;
1009
1010         switch (init->depth_bpp) {
1011         case 16:
1012                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_16BIT_INT_Z;
1013                 break;
1014         case 32:
1015         default:
1016                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_24BIT_INT_Z;
1017                 break;
1018         }
1019         dev_priv->depth_offset = init->depth_offset;
1020         dev_priv->depth_pitch = init->depth_pitch;
1021
1022         /* Hardware state for depth clears.  Remove this if/when we no
1023          * longer clear the depth buffer with a 3D rectangle.  Hard-code
1024          * all values to prevent unwanted 3D state from slipping through
1025          * and screwing with the clear operation.
1026          */
1027         dev_priv->depth_clear.rb3d_cntl = (RADEON_PLANE_MASK_ENABLE |
1028                                            (dev_priv->color_fmt << 10) |
1029                                            (dev_priv->microcode_version ==
1030                                             UCODE_R100 ? RADEON_ZBLOCK16 : 0));
1031
1032         dev_priv->depth_clear.rb3d_zstencilcntl =
1033             (dev_priv->depth_fmt |
1034              RADEON_Z_TEST_ALWAYS |
1035              RADEON_STENCIL_TEST_ALWAYS |
1036              RADEON_STENCIL_S_FAIL_REPLACE |
1037              RADEON_STENCIL_ZPASS_REPLACE |
1038              RADEON_STENCIL_ZFAIL_REPLACE | RADEON_Z_WRITE_ENABLE);
1039
1040         dev_priv->depth_clear.se_cntl = (RADEON_FFACE_CULL_CW |
1041                                          RADEON_BFACE_SOLID |
1042                                          RADEON_FFACE_SOLID |
1043                                          RADEON_FLAT_SHADE_VTX_LAST |
1044                                          RADEON_DIFFUSE_SHADE_FLAT |
1045                                          RADEON_ALPHA_SHADE_FLAT |
1046                                          RADEON_SPECULAR_SHADE_FLAT |
1047                                          RADEON_FOG_SHADE_FLAT |
1048                                          RADEON_VTX_PIX_CENTER_OGL |
1049                                          RADEON_ROUND_MODE_TRUNC |
1050                                          RADEON_ROUND_PREC_8TH_PIX);
1051
1052
1053         dev_priv->ring_offset = init->ring_offset;
1054         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
1055         dev_priv->buffers_offset = init->buffers_offset;
1056         dev_priv->gart_textures_offset = init->gart_textures_offset;
1057
1058         master_priv->sarea = drm_getsarea(dev);
1059         if (!master_priv->sarea) {
1060                 DRM_ERROR("could not find sarea!\n");
1061                 radeon_do_cleanup_cp(dev);
1062                 return -EINVAL;
1063         }
1064
1065         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
1066         if (!dev_priv->cp_ring) {
1067                 DRM_ERROR("could not find cp ring region!\n");
1068                 radeon_do_cleanup_cp(dev);
1069                 return -EINVAL;
1070         }
1071         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
1072         if (!dev_priv->ring_rptr) {
1073                 DRM_ERROR("could not find ring read pointer!\n");
1074                 radeon_do_cleanup_cp(dev);
1075                 return -EINVAL;
1076         }
1077         dev->agp_buffer_token = init->buffers_offset;
1078         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1079         if (!dev->agp_buffer_map) {
1080                 DRM_ERROR("could not find dma buffer region!\n");
1081                 radeon_do_cleanup_cp(dev);
1082                 return -EINVAL;
1083         }
1084
1085         if (init->gart_textures_offset) {
1086                 dev_priv->gart_textures =
1087                     drm_core_findmap(dev, init->gart_textures_offset);
1088                 if (!dev_priv->gart_textures) {
1089                         DRM_ERROR("could not find GART texture region!\n");
1090                         radeon_do_cleanup_cp(dev);
1091                         return -EINVAL;
1092                 }
1093         }
1094
1095 #if __OS_HAS_AGP
1096         if (dev_priv->flags & RADEON_IS_AGP) {
1097                 drm_core_ioremap_wc(dev_priv->cp_ring, dev);
1098                 drm_core_ioremap_wc(dev_priv->ring_rptr, dev);
1099                 drm_core_ioremap_wc(dev->agp_buffer_map, dev);
1100                 if (!dev_priv->cp_ring->handle ||
1101                     !dev_priv->ring_rptr->handle ||
1102                     !dev->agp_buffer_map->handle) {
1103                         DRM_ERROR("could not find ioremap agp regions!\n");
1104                         radeon_do_cleanup_cp(dev);
1105                         return -EINVAL;
1106                 }
1107         } else
1108 #endif
1109         {
1110                 dev_priv->cp_ring->handle =
1111                         (void *)(unsigned long)dev_priv->cp_ring->offset;
1112                 dev_priv->ring_rptr->handle =
1113                         (void *)(unsigned long)dev_priv->ring_rptr->offset;
1114                 dev->agp_buffer_map->handle =
1115                         (void *)(unsigned long)dev->agp_buffer_map->offset;
1116
1117                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
1118                           dev_priv->cp_ring->handle);
1119                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
1120                           dev_priv->ring_rptr->handle);
1121                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
1122                           dev->agp_buffer_map->handle);
1123         }
1124
1125         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 16;
1126         dev_priv->fb_size =
1127                 ((radeon_read_fb_location(dev_priv) & 0xffff0000u) + 0x10000)
1128                 - dev_priv->fb_location;
1129
1130         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
1131                                         ((dev_priv->front_offset
1132                                           + dev_priv->fb_location) >> 10));
1133
1134         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
1135                                        ((dev_priv->back_offset
1136                                          + dev_priv->fb_location) >> 10));
1137
1138         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
1139                                         ((dev_priv->depth_offset
1140                                           + dev_priv->fb_location) >> 10));
1141
1142         dev_priv->gart_size = init->gart_size;
1143
1144         /* New let's set the memory map ... */
1145         if (dev_priv->new_memmap) {
1146                 u32 base = 0;
1147
1148                 DRM_INFO("Setting GART location based on new memory map\n");
1149
1150                 /* If using AGP, try to locate the AGP aperture at the same
1151                  * location in the card and on the bus, though we have to
1152                  * align it down.
1153                  */
1154 #if __OS_HAS_AGP
1155                 if (dev_priv->flags & RADEON_IS_AGP) {
1156                         base = dev->agp->base;
1157                         /* Check if valid */
1158                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
1159                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
1160                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
1161                                          dev->agp->base);
1162                                 base = 0;
1163                         }
1164                 }
1165 #endif
1166                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
1167                 if (base == 0) {
1168                         base = dev_priv->fb_location + dev_priv->fb_size;
1169                         if (base < dev_priv->fb_location ||
1170                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
1171                                 base = dev_priv->fb_location
1172                                         - dev_priv->gart_size;
1173                 }
1174                 dev_priv->gart_vm_start = base & 0xffc00000u;
1175                 if (dev_priv->gart_vm_start != base)
1176                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
1177                                  base, dev_priv->gart_vm_start);
1178         } else {
1179                 DRM_INFO("Setting GART location based on old memory map\n");
1180                 dev_priv->gart_vm_start = dev_priv->fb_location +
1181                         RADEON_READ(RADEON_CONFIG_APER_SIZE);
1182         }
1183
1184 #if __OS_HAS_AGP
1185         if (dev_priv->flags & RADEON_IS_AGP)
1186                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1187                                                  - dev->agp->base
1188                                                  + dev_priv->gart_vm_start);
1189         else
1190 #endif
1191                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1192                                         - (unsigned long)dev->sg->virtual
1193                                         + dev_priv->gart_vm_start);
1194
1195         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
1196         DRM_DEBUG("dev_priv->gart_vm_start 0x%x\n", dev_priv->gart_vm_start);
1197         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%lx\n",
1198                   dev_priv->gart_buffers_offset);
1199
1200         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
1201         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
1202                               + init->ring_size / sizeof(u32));
1203         dev_priv->ring.size = init->ring_size;
1204         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
1205
1206         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
1207         dev_priv->ring.rptr_update_l2qw = drm_order( /* init->rptr_update */ 4096 / 8);
1208
1209         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
1210         dev_priv->ring.fetch_size_l2ow = drm_order( /* init->fetch_size */ 32 / 16);
1211         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
1212
1213         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
1214
1215 #if __OS_HAS_AGP
1216         if (dev_priv->flags & RADEON_IS_AGP) {
1217                 /* Turn off PCI GART */
1218                 radeon_set_pcigart(dev_priv, 0);
1219         } else
1220 #endif
1221         {
1222                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
1223                 /* if we have an offset set from userspace */
1224                 if (dev_priv->pcigart_offset_set) {
1225                         dev_priv->gart_info.bus_addr =
1226                                 (resource_size_t)dev_priv->pcigart_offset + dev_priv->fb_location;
1227                         dev_priv->gart_info.mapping.offset =
1228                             dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
1229                         dev_priv->gart_info.mapping.size =
1230                             dev_priv->gart_info.table_size;
1231
1232                         drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
1233                         dev_priv->gart_info.addr =
1234                             dev_priv->gart_info.mapping.handle;
1235
1236                         if (dev_priv->flags & RADEON_IS_PCIE)
1237                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
1238                         else
1239                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1240                         dev_priv->gart_info.gart_table_location =
1241                             DRM_ATI_GART_FB;
1242
1243                         DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
1244                                   dev_priv->gart_info.addr,
1245                                   dev_priv->pcigart_offset);
1246                 } else {
1247                         if (dev_priv->flags & RADEON_IS_IGPGART)
1248                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
1249                         else
1250                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1251                         dev_priv->gart_info.gart_table_location =
1252                             DRM_ATI_GART_MAIN;
1253                         dev_priv->gart_info.addr = NULL;
1254                         dev_priv->gart_info.bus_addr = 0;
1255                         if (dev_priv->flags & RADEON_IS_PCIE) {
1256                                 DRM_ERROR
1257                                     ("Cannot use PCI Express without GART in FB memory\n");
1258                                 radeon_do_cleanup_cp(dev);
1259                                 return -EINVAL;
1260                         }
1261                 }
1262
1263                 if (!drm_ati_pcigart_init(dev, &dev_priv->gart_info)) {
1264                         DRM_ERROR("failed to init PCI GART!\n");
1265                         radeon_do_cleanup_cp(dev);
1266                         return -ENOMEM;
1267                 }
1268
1269                 /* Turn on PCI GART */
1270                 radeon_set_pcigart(dev_priv, 1);
1271         }
1272
1273         radeon_cp_load_microcode(dev_priv);
1274         radeon_cp_init_ring_buffer(dev, dev_priv, file_priv);
1275
1276         dev_priv->last_buf = 0;
1277
1278         radeon_do_engine_reset(dev);
1279         radeon_test_writeback(dev_priv);
1280
1281         return 0;
1282 }
1283
1284 static int radeon_do_cleanup_cp(struct drm_device * dev)
1285 {
1286         drm_radeon_private_t *dev_priv = dev->dev_private;
1287         DRM_DEBUG("\n");
1288
1289         /* Make sure interrupts are disabled here because the uninstall ioctl
1290          * may not have been called from userspace and after dev_private
1291          * is freed, it's too late.
1292          */
1293         if (dev->irq_enabled)
1294                 drm_irq_uninstall(dev);
1295
1296 #if __OS_HAS_AGP
1297         if (dev_priv->flags & RADEON_IS_AGP) {
1298                 if (dev_priv->cp_ring != NULL) {
1299                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1300                         dev_priv->cp_ring = NULL;
1301                 }
1302                 if (dev_priv->ring_rptr != NULL) {
1303                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1304                         dev_priv->ring_rptr = NULL;
1305                 }
1306                 if (dev->agp_buffer_map != NULL) {
1307                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1308                         dev->agp_buffer_map = NULL;
1309                 }
1310         } else
1311 #endif
1312         {
1313
1314                 if (dev_priv->gart_info.bus_addr) {
1315                         /* Turn off PCI GART */
1316                         radeon_set_pcigart(dev_priv, 0);
1317                         if (!drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info))
1318                                 DRM_ERROR("failed to cleanup PCI GART!\n");
1319                 }
1320
1321                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB)
1322                 {
1323                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1324                         dev_priv->gart_info.addr = 0;
1325                 }
1326         }
1327         /* only clear to the start of flags */
1328         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1329
1330         return 0;
1331 }
1332
1333 /* This code will reinit the Radeon CP hardware after a resume from disc.
1334  * AFAIK, it would be very difficult to pickle the state at suspend time, so
1335  * here we make sure that all Radeon hardware initialisation is re-done without
1336  * affecting running applications.
1337  *
1338  * Charl P. Botha <http://cpbotha.net>
1339  */
1340 static int radeon_do_resume_cp(struct drm_device *dev, struct drm_file *file_priv)
1341 {
1342         drm_radeon_private_t *dev_priv = dev->dev_private;
1343
1344         if (!dev_priv) {
1345                 DRM_ERROR("Called with no initialization\n");
1346                 return -EINVAL;
1347         }
1348
1349         DRM_DEBUG("Starting radeon_do_resume_cp()\n");
1350
1351 #if __OS_HAS_AGP
1352         if (dev_priv->flags & RADEON_IS_AGP) {
1353                 /* Turn off PCI GART */
1354                 radeon_set_pcigart(dev_priv, 0);
1355         } else
1356 #endif
1357         {
1358                 /* Turn on PCI GART */
1359                 radeon_set_pcigart(dev_priv, 1);
1360         }
1361
1362         radeon_cp_load_microcode(dev_priv);
1363         radeon_cp_init_ring_buffer(dev, dev_priv, file_priv);
1364
1365         radeon_do_engine_reset(dev);
1366         radeon_irq_set_state(dev, RADEON_SW_INT_ENABLE, 1);
1367
1368         DRM_DEBUG("radeon_do_resume_cp() complete\n");
1369
1370         return 0;
1371 }
1372
1373 int radeon_cp_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
1374 {
1375         drm_radeon_init_t *init = data;
1376
1377         LOCK_TEST_WITH_RETURN(dev, file_priv);
1378
1379         if (init->func == RADEON_INIT_R300_CP)
1380                 r300_init_reg_flags(dev);
1381
1382         switch (init->func) {
1383         case RADEON_INIT_CP:
1384         case RADEON_INIT_R200_CP:
1385         case RADEON_INIT_R300_CP:
1386                 return radeon_do_init_cp(dev, init, file_priv);
1387         case RADEON_CLEANUP_CP:
1388                 return radeon_do_cleanup_cp(dev);
1389         }
1390
1391         return -EINVAL;
1392 }
1393
1394 int radeon_cp_start(struct drm_device *dev, void *data, struct drm_file *file_priv)
1395 {
1396         drm_radeon_private_t *dev_priv = dev->dev_private;
1397         DRM_DEBUG("\n");
1398
1399         LOCK_TEST_WITH_RETURN(dev, file_priv);
1400
1401         if (dev_priv->cp_running) {
1402                 DRM_DEBUG("while CP running\n");
1403                 return 0;
1404         }
1405         if (dev_priv->cp_mode == RADEON_CSQ_PRIDIS_INDDIS) {
1406                 DRM_DEBUG("called with bogus CP mode (%d)\n",
1407                           dev_priv->cp_mode);
1408                 return 0;
1409         }
1410
1411         radeon_do_cp_start(dev_priv);
1412
1413         return 0;
1414 }
1415
1416 /* Stop the CP.  The engine must have been idled before calling this
1417  * routine.
1418  */
1419 int radeon_cp_stop(struct drm_device *dev, void *data, struct drm_file *file_priv)
1420 {
1421         drm_radeon_private_t *dev_priv = dev->dev_private;
1422         drm_radeon_cp_stop_t *stop = data;
1423         int ret;
1424         DRM_DEBUG("\n");
1425
1426         LOCK_TEST_WITH_RETURN(dev, file_priv);
1427
1428         if (!dev_priv->cp_running)
1429                 return 0;
1430
1431         /* Flush any pending CP commands.  This ensures any outstanding
1432          * commands are exectuted by the engine before we turn it off.
1433          */
1434         if (stop->flush) {
1435                 radeon_do_cp_flush(dev_priv);
1436         }
1437
1438         /* If we fail to make the engine go idle, we return an error
1439          * code so that the DRM ioctl wrapper can try again.
1440          */
1441         if (stop->idle) {
1442                 ret = radeon_do_cp_idle(dev_priv);
1443                 if (ret)
1444                         return ret;
1445         }
1446
1447         /* Finally, we can turn off the CP.  If the engine isn't idle,
1448          * we will get some dropped triangles as they won't be fully
1449          * rendered before the CP is shut down.
1450          */
1451         radeon_do_cp_stop(dev_priv);
1452
1453         /* Reset the engine */
1454         radeon_do_engine_reset(dev);
1455
1456         return 0;
1457 }
1458
1459 void radeon_do_release(struct drm_device * dev)
1460 {
1461         drm_radeon_private_t *dev_priv = dev->dev_private;
1462         int i, ret;
1463
1464         if (dev_priv) {
1465                 if (dev_priv->cp_running) {
1466                         /* Stop the cp */
1467                         while ((ret = radeon_do_cp_idle(dev_priv)) != 0) {
1468                                 DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1469 #ifdef __linux__
1470                                 schedule();
1471 #else
1472                                 tsleep(&ret, PZERO, "rdnrel", 1);
1473 #endif
1474                         }
1475                         radeon_do_cp_stop(dev_priv);
1476                         radeon_do_engine_reset(dev);
1477                 }
1478
1479                 /* Disable *all* interrupts */
1480                 if (dev_priv->mmio)     /* remove this after permanent addmaps */
1481                         RADEON_WRITE(RADEON_GEN_INT_CNTL, 0);
1482
1483                 if (dev_priv->mmio) {   /* remove all surfaces */
1484                         for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1485                                 RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, 0);
1486                                 RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND +
1487                                              16 * i, 0);
1488                                 RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND +
1489                                              16 * i, 0);
1490                         }
1491                 }
1492
1493                 /* Free memory heap structures */
1494                 radeon_mem_takedown(&(dev_priv->gart_heap));
1495                 radeon_mem_takedown(&(dev_priv->fb_heap));
1496
1497                 /* deallocate kernel resources */
1498                 radeon_do_cleanup_cp(dev);
1499         }
1500 }
1501
1502 /* Just reset the CP ring.  Called as part of an X Server engine reset.
1503  */
1504 int radeon_cp_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1505 {
1506         drm_radeon_private_t *dev_priv = dev->dev_private;
1507         DRM_DEBUG("\n");
1508
1509         LOCK_TEST_WITH_RETURN(dev, file_priv);
1510
1511         if (!dev_priv) {
1512                 DRM_DEBUG("called before init done\n");
1513                 return -EINVAL;
1514         }
1515
1516         radeon_do_cp_reset(dev_priv);
1517
1518         /* The CP is no longer running after an engine reset */
1519         dev_priv->cp_running = 0;
1520
1521         return 0;
1522 }
1523
1524 int radeon_cp_idle(struct drm_device *dev, void *data, struct drm_file *file_priv)
1525 {
1526         drm_radeon_private_t *dev_priv = dev->dev_private;
1527         DRM_DEBUG("\n");
1528
1529         LOCK_TEST_WITH_RETURN(dev, file_priv);
1530
1531         return radeon_do_cp_idle(dev_priv);
1532 }
1533
1534 /* Added by Charl P. Botha to call radeon_do_resume_cp().
1535  */
1536 int radeon_cp_resume(struct drm_device *dev, void *data, struct drm_file *file_priv)
1537 {
1538         return radeon_do_resume_cp(dev, file_priv);
1539 }
1540
1541 int radeon_engine_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1542 {
1543         DRM_DEBUG("\n");
1544
1545         LOCK_TEST_WITH_RETURN(dev, file_priv);
1546
1547         return radeon_do_engine_reset(dev);
1548 }
1549
1550 /* ================================================================
1551  * Fullscreen mode
1552  */
1553
1554 /* KW: Deprecated to say the least:
1555  */
1556 int radeon_fullscreen(struct drm_device *dev, void *data, struct drm_file *file_priv)
1557 {
1558         return 0;
1559 }
1560
1561 /* ================================================================
1562  * Freelist management
1563  */
1564
1565 /* Original comment: FIXME: ROTATE_BUFS is a hack to cycle through
1566  *   bufs until freelist code is used.  Note this hides a problem with
1567  *   the scratch register * (used to keep track of last buffer
1568  *   completed) being written to before * the last buffer has actually
1569  *   completed rendering.
1570  *
1571  * KW:  It's also a good way to find free buffers quickly.
1572  *
1573  * KW: Ideally this loop wouldn't exist, and freelist_get wouldn't
1574  * sleep.  However, bugs in older versions of radeon_accel.c mean that
1575  * we essentially have to do this, else old clients will break.
1576  *
1577  * However, it does leave open a potential deadlock where all the
1578  * buffers are held by other clients, which can't release them because
1579  * they can't get the lock.
1580  */
1581
1582 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1583 {
1584         struct drm_device_dma *dma = dev->dma;
1585         drm_radeon_private_t *dev_priv = dev->dev_private;
1586         drm_radeon_buf_priv_t *buf_priv;
1587         struct drm_buf *buf;
1588         int i, t;
1589         int start;
1590
1591         if (++dev_priv->last_buf >= dma->buf_count)
1592                 dev_priv->last_buf = 0;
1593
1594         start = dev_priv->last_buf;
1595
1596         for (t = 0; t < dev_priv->usec_timeout; t++) {
1597                 u32 done_age = GET_SCRATCH(dev_priv, 1);
1598                 DRM_DEBUG("done_age = %d\n", done_age);
1599                 for (i = start; i < dma->buf_count; i++) {
1600                         buf = dma->buflist[i];
1601                         buf_priv = buf->dev_private;
1602                         if (buf->file_priv == NULL || (buf->pending &&
1603                                                        buf_priv->age <=
1604                                                        done_age)) {
1605                                 dev_priv->stats.requested_bufs++;
1606                                 buf->pending = 0;
1607                                 return buf;
1608                         }
1609                         start = 0;
1610                 }
1611
1612                 if (t) {
1613                         DRM_UDELAY(1);
1614                         dev_priv->stats.freelist_loops++;
1615                 }
1616         }
1617
1618         DRM_DEBUG("returning NULL!\n");
1619         return NULL;
1620 }
1621
1622 #if 0
1623 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1624 {
1625         struct drm_device_dma *dma = dev->dma;
1626         drm_radeon_private_t *dev_priv = dev->dev_private;
1627         drm_radeon_buf_priv_t *buf_priv;
1628         struct drm_buf *buf;
1629         int i, t;
1630         int start;
1631         u32 done_age;
1632
1633         done_age = radeon_read_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1));
1634         if (++dev_priv->last_buf >= dma->buf_count)
1635                 dev_priv->last_buf = 0;
1636
1637         start = dev_priv->last_buf;
1638         dev_priv->stats.freelist_loops++;
1639
1640         for (t = 0; t < 2; t++) {
1641                 for (i = start; i < dma->buf_count; i++) {
1642                         buf = dma->buflist[i];
1643                         buf_priv = buf->dev_private;
1644                         if (buf->file_priv == 0 || (buf->pending &&
1645                                                     buf_priv->age <=
1646                                                     done_age)) {
1647                                 dev_priv->stats.requested_bufs++;
1648                                 buf->pending = 0;
1649                                 return buf;
1650                         }
1651                 }
1652                 start = 0;
1653         }
1654
1655         return NULL;
1656 }
1657 #endif
1658
1659 void radeon_freelist_reset(struct drm_device * dev)
1660 {
1661         struct drm_device_dma *dma = dev->dma;
1662         drm_radeon_private_t *dev_priv = dev->dev_private;
1663         int i;
1664
1665         dev_priv->last_buf = 0;
1666         for (i = 0; i < dma->buf_count; i++) {
1667                 struct drm_buf *buf = dma->buflist[i];
1668                 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
1669                 buf_priv->age = 0;
1670         }
1671 }
1672
1673 /* ================================================================
1674  * CP command submission
1675  */
1676
1677 int radeon_wait_ring(drm_radeon_private_t * dev_priv, int n)
1678 {
1679         drm_radeon_ring_buffer_t *ring = &dev_priv->ring;
1680         int i;
1681         u32 last_head = GET_RING_HEAD(dev_priv);
1682
1683         for (i = 0; i < dev_priv->usec_timeout; i++) {
1684                 u32 head = GET_RING_HEAD(dev_priv);
1685
1686                 ring->space = (head - ring->tail) * sizeof(u32);
1687                 if (ring->space <= 0)
1688                         ring->space += ring->size;
1689                 if (ring->space > n)
1690                         return 0;
1691
1692                 dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
1693
1694                 if (head != last_head)
1695                         i = 0;
1696                 last_head = head;
1697
1698                 DRM_UDELAY(1);
1699         }
1700
1701         /* FIXME: This return value is ignored in the BEGIN_RING macro! */
1702 #if RADEON_FIFO_DEBUG
1703         radeon_status(dev_priv);
1704         DRM_ERROR("failed!\n");
1705 #endif
1706         return -EBUSY;
1707 }
1708
1709 static int radeon_cp_get_buffers(struct drm_device *dev,
1710                                  struct drm_file *file_priv,
1711                                  struct drm_dma * d)
1712 {
1713         int i;
1714         struct drm_buf *buf;
1715
1716         for (i = d->granted_count; i < d->request_count; i++) {
1717                 buf = radeon_freelist_get(dev);
1718                 if (!buf)
1719                         return -EBUSY;  /* NOTE: broken client */
1720
1721                 buf->file_priv = file_priv;
1722
1723                 if (DRM_COPY_TO_USER(&d->request_indices[i], &buf->idx,
1724                                      sizeof(buf->idx)))
1725                         return -EFAULT;
1726                 if (DRM_COPY_TO_USER(&d->request_sizes[i], &buf->total,
1727                                      sizeof(buf->total)))
1728                         return -EFAULT;
1729
1730                 d->granted_count++;
1731         }
1732         return 0;
1733 }
1734
1735 int radeon_cp_buffers(struct drm_device *dev, void *data, struct drm_file *file_priv)
1736 {
1737         struct drm_device_dma *dma = dev->dma;
1738         int ret = 0;
1739         struct drm_dma *d = data;
1740
1741         LOCK_TEST_WITH_RETURN(dev, file_priv);
1742
1743         /* Please don't send us buffers.
1744          */
1745         if (d->send_count != 0) {
1746                 DRM_ERROR("Process %d trying to send %d buffers via drmDMA\n",
1747                           DRM_CURRENTPID, d->send_count);
1748                 return -EINVAL;
1749         }
1750
1751         /* We'll send you buffers.
1752          */
1753         if (d->request_count < 0 || d->request_count > dma->buf_count) {
1754                 DRM_ERROR("Process %d trying to get %d buffers (of %d max)\n",
1755                           DRM_CURRENTPID, d->request_count, dma->buf_count);
1756                 return -EINVAL;
1757         }
1758
1759         d->granted_count = 0;
1760
1761         if (d->request_count) {
1762                 ret = radeon_cp_get_buffers(dev, file_priv, d);
1763         }
1764
1765         return ret;
1766 }
1767
1768 int radeon_driver_load(struct drm_device *dev, unsigned long flags)
1769 {
1770         drm_radeon_private_t *dev_priv;
1771         int ret = 0;
1772
1773         dev_priv = drm_alloc(sizeof(drm_radeon_private_t), DRM_MEM_DRIVER);
1774         if (dev_priv == NULL)
1775                 return -ENOMEM;
1776
1777         memset(dev_priv, 0, sizeof(drm_radeon_private_t));
1778         dev->dev_private = (void *)dev_priv;
1779         dev_priv->flags = flags;
1780
1781         switch (flags & RADEON_FAMILY_MASK) {
1782         case CHIP_R100:
1783         case CHIP_RV200:
1784         case CHIP_R200:
1785         case CHIP_R300:
1786         case CHIP_R350:
1787         case CHIP_R420:
1788         case CHIP_R423:
1789         case CHIP_RV410:
1790         case CHIP_RV515:
1791         case CHIP_R520:
1792         case CHIP_RV570:
1793         case CHIP_R580:
1794                 dev_priv->flags |= RADEON_HAS_HIERZ;
1795                 break;
1796         default:
1797                 /* all other chips have no hierarchical z buffer */
1798                 break;
1799         }
1800
1801         if (drm_device_is_agp(dev))
1802                 dev_priv->flags |= RADEON_IS_AGP;
1803         else if (drm_device_is_pcie(dev))
1804                 dev_priv->flags |= RADEON_IS_PCIE;
1805         else
1806                 dev_priv->flags |= RADEON_IS_PCI;
1807
1808         ret = drm_addmap(dev, drm_get_resource_start(dev, 2),
1809                          drm_get_resource_len(dev, 2), _DRM_REGISTERS,
1810                          _DRM_READ_ONLY | _DRM_DRIVER, &dev_priv->mmio);
1811         if (ret != 0)
1812                 return ret;
1813
1814         ret = drm_vblank_init(dev, 2);
1815         if (ret) {
1816                 radeon_driver_unload(dev);
1817                 return ret;
1818         }
1819
1820         DRM_DEBUG("%s card detected\n",
1821                   ((dev_priv->flags & RADEON_IS_AGP) ? "AGP" : (((dev_priv->flags & RADEON_IS_PCIE) ? "PCIE" : "PCI"))));
1822         return ret;
1823 }
1824
1825 int radeon_master_create(struct drm_device *dev, struct drm_master *master)
1826 {
1827         struct drm_radeon_master_private *master_priv;
1828         unsigned long sareapage;
1829         int ret;
1830
1831         master_priv = drm_calloc(1, sizeof(*master_priv), DRM_MEM_DRIVER);
1832         if (!master_priv)
1833                 return -ENOMEM;
1834
1835         /* prebuild the SAREA */
1836         sareapage = max_t(unsigned long, SAREA_MAX, PAGE_SIZE);
1837         ret = drm_addmap(dev, 0, sareapage, _DRM_SHM, _DRM_CONTAINS_LOCK|_DRM_DRIVER,
1838                          &master_priv->sarea);
1839         if (ret) {
1840                 DRM_ERROR("SAREA setup failed\n");
1841                 return ret;
1842         }
1843         master_priv->sarea_priv = master_priv->sarea->handle + sizeof(struct drm_sarea);
1844         master_priv->sarea_priv->pfCurrentPage = 0;
1845
1846         master->driver_priv = master_priv;
1847         return 0;
1848 }
1849
1850 void radeon_master_destroy(struct drm_device *dev, struct drm_master *master)
1851 {
1852         struct drm_radeon_master_private *master_priv = master->driver_priv;
1853
1854         if (!master_priv)
1855                 return;
1856
1857         if (master_priv->sarea_priv &&
1858             master_priv->sarea_priv->pfCurrentPage != 0)
1859                 radeon_cp_dispatch_flip(dev, master);
1860
1861         master_priv->sarea_priv = NULL;
1862         if (master_priv->sarea)
1863                 drm_rmmap_locked(dev, master_priv->sarea);
1864
1865         drm_free(master_priv, sizeof(*master_priv), DRM_MEM_DRIVER);
1866
1867         master->driver_priv = NULL;
1868 }
1869
1870 /* Create mappings for registers and framebuffer so userland doesn't necessarily
1871  * have to find them.
1872  */
1873 int radeon_driver_firstopen(struct drm_device *dev)
1874 {
1875         int ret;
1876         drm_local_map_t *map;
1877         drm_radeon_private_t *dev_priv = dev->dev_private;
1878
1879         dev_priv->gart_info.table_size = RADEON_PCIGART_TABLE_SIZE;
1880
1881         dev_priv->fb_aper_offset = drm_get_resource_start(dev, 0);
1882         ret = drm_addmap(dev, dev_priv->fb_aper_offset,
1883                          drm_get_resource_len(dev, 0), _DRM_FRAME_BUFFER,
1884                          _DRM_WRITE_COMBINING, &map);
1885         if (ret != 0)
1886                 return ret;
1887
1888         return 0;
1889 }
1890
1891 int radeon_driver_unload(struct drm_device *dev)
1892 {
1893         drm_radeon_private_t *dev_priv = dev->dev_private;
1894
1895         DRM_DEBUG("\n");
1896
1897         drm_rmmap(dev, dev_priv->mmio);
1898
1899         drm_free(dev_priv, sizeof(*dev_priv), DRM_MEM_DRIVER);
1900
1901         dev->dev_private = NULL;
1902         return 0;
1903 }