git.oblomov.eu Git - linux-2.6/blob - 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 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                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
  67                         return RADEON_READ(R600_CP_RB_RPTR);
  68                 else
  69                         return RADEON_READ(RADEON_CP_RB_RPTR);
  70         }
  71 }
  72
  73 void radeon_write_ring_rptr(drm_radeon_private_t *dev_priv, u32 off, u32 val)
  74 {
  75         if (dev_priv->flags & RADEON_IS_AGP)
  76                 DRM_WRITE32(dev_priv->ring_rptr, off, val);
  77         else
  78                 *(((volatile u32 *) dev_priv->ring_rptr->handle) +
  79                   (off / sizeof(u32))) = cpu_to_le32(val);
  80 }
  81
  82 void radeon_set_ring_head(drm_radeon_private_t *dev_priv, u32 val)
  83 {
  84         radeon_write_ring_rptr(dev_priv, 0, val);
  85 }
  86
  87 u32 radeon_get_scratch(drm_radeon_private_t *dev_priv, int index)
  88 {
  89         if (dev_priv->writeback_works) {
  90                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
  91                         return radeon_read_ring_rptr(dev_priv,
  92                                                      R600_SCRATCHOFF(index));
  93                 else
  94                         return radeon_read_ring_rptr(dev_priv,
  95                                                      RADEON_SCRATCHOFF(index));
  96         } else {
  97                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
  98                         return RADEON_READ(R600_SCRATCH_REG0 + 4*index);
  99                 else
 100                         return RADEON_READ(RADEON_SCRATCH_REG0 + 4*index);
 101         }
 102 }
 103
 104 u32 RADEON_READ_MM(drm_radeon_private_t *dev_priv, int addr)
 105 {
 106         u32 ret;
 107
 108         if (addr < 0x10000)
 109                 ret = DRM_READ32(dev_priv->mmio, addr);
 110         else {
 111                 DRM_WRITE32(dev_priv->mmio, RADEON_MM_INDEX, addr);
 112                 ret = DRM_READ32(dev_priv->mmio, RADEON_MM_DATA);
 113         }
 114
 115         return ret;
 116 }
 117
 118 static u32 R500_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 119 {
 120         u32 ret;
 121         RADEON_WRITE(R520_MC_IND_INDEX, 0x7f0000 | (addr & 0xff));
 122         ret = RADEON_READ(R520_MC_IND_DATA);
 123         RADEON_WRITE(R520_MC_IND_INDEX, 0);
 124         return ret;
 125 }
 126
 127 static u32 RS480_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 128 {
 129         u32 ret;
 130         RADEON_WRITE(RS480_NB_MC_INDEX, addr & 0xff);
 131         ret = RADEON_READ(RS480_NB_MC_DATA);
 132         RADEON_WRITE(RS480_NB_MC_INDEX, 0xff);
 133         return ret;
 134 }
 135
 136 static u32 RS690_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 137 {
 138         u32 ret;
 139         RADEON_WRITE(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
 140         ret = RADEON_READ(RS690_MC_DATA);
 141         RADEON_WRITE(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
 142         return ret;
 143 }
 144
 145 static u32 RS600_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 146 {
 147         u32 ret;
 148         RADEON_WRITE(RS600_MC_INDEX, ((addr & RS600_MC_ADDR_MASK) |
 149                                       RS600_MC_IND_CITF_ARB0));
 150         ret = RADEON_READ(RS600_MC_DATA);
 151         return ret;
 152 }
 153
 154 static u32 IGP_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
 155 {
 156         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 157             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 158                 return RS690_READ_MCIND(dev_priv, addr);
 159         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
 160                 return RS600_READ_MCIND(dev_priv, addr);
 161         else
 162                 return RS480_READ_MCIND(dev_priv, addr);
 163 }
 164
 165 u32 radeon_read_fb_location(drm_radeon_private_t *dev_priv)
 166 {
 167
 168         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 169                 return RADEON_READ(R700_MC_VM_FB_LOCATION);
 170         else if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
 171                 return RADEON_READ(R600_MC_VM_FB_LOCATION);
 172         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 173                 return R500_READ_MCIND(dev_priv, RV515_MC_FB_LOCATION);
 174         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 175                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 176                 return RS690_READ_MCIND(dev_priv, RS690_MC_FB_LOCATION);
 177         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
 178                 return RS600_READ_MCIND(dev_priv, RS600_MC_FB_LOCATION);
 179         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 180                 return R500_READ_MCIND(dev_priv, R520_MC_FB_LOCATION);
 181         else
 182                 return RADEON_READ(RADEON_MC_FB_LOCATION);
 183 }
 184
 185 static void radeon_write_fb_location(drm_radeon_private_t *dev_priv, u32 fb_loc)
 186 {
 187         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 188                 RADEON_WRITE(R700_MC_VM_FB_LOCATION, fb_loc);
 189         else if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
 190                 RADEON_WRITE(R600_MC_VM_FB_LOCATION, fb_loc);
 191         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 192                 R500_WRITE_MCIND(RV515_MC_FB_LOCATION, fb_loc);
 193         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 194                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 195                 RS690_WRITE_MCIND(RS690_MC_FB_LOCATION, fb_loc);
 196         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
 197                 RS600_WRITE_MCIND(RS600_MC_FB_LOCATION, fb_loc);
 198         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 199                 R500_WRITE_MCIND(R520_MC_FB_LOCATION, fb_loc);
 200         else
 201                 RADEON_WRITE(RADEON_MC_FB_LOCATION, fb_loc);
 202 }
 203
 204 void radeon_write_agp_location(drm_radeon_private_t *dev_priv, u32 agp_loc)
 205 {
 206         /*R6xx/R7xx: AGP_TOP and BOT are actually 18 bits each */
 207         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770) {
 208                 RADEON_WRITE(R700_MC_VM_AGP_BOT, agp_loc & 0xffff); /* FIX ME */
 209                 RADEON_WRITE(R700_MC_VM_AGP_TOP, (agp_loc >> 16) & 0xffff);
 210         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600) {
 211                 RADEON_WRITE(R600_MC_VM_AGP_BOT, agp_loc & 0xffff); /* FIX ME */
 212                 RADEON_WRITE(R600_MC_VM_AGP_TOP, (agp_loc >> 16) & 0xffff);
 213         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 214                 R500_WRITE_MCIND(RV515_MC_AGP_LOCATION, agp_loc);
 215         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 216                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 217                 RS690_WRITE_MCIND(RS690_MC_AGP_LOCATION, agp_loc);
 218         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
 219                 RS600_WRITE_MCIND(RS600_MC_AGP_LOCATION, agp_loc);
 220         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 221                 R500_WRITE_MCIND(R520_MC_AGP_LOCATION, agp_loc);
 222         else
 223                 RADEON_WRITE(RADEON_MC_AGP_LOCATION, agp_loc);
 224 }
 225
 226 void radeon_write_agp_base(drm_radeon_private_t *dev_priv, u64 agp_base)
 227 {
 228         u32 agp_base_hi = upper_32_bits(agp_base);
 229         u32 agp_base_lo = agp_base & 0xffffffff;
 230         u32 r6xx_agp_base = (agp_base >> 22) & 0x3ffff;
 231
 232         /* R6xx/R7xx must be aligned to a 4MB boundry */
 233         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV770)
 234                 RADEON_WRITE(R700_MC_VM_AGP_BASE, r6xx_agp_base);
 235         else if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
 236                 RADEON_WRITE(R600_MC_VM_AGP_BASE, r6xx_agp_base);
 237         else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) {
 238                 R500_WRITE_MCIND(RV515_MC_AGP_BASE, agp_base_lo);
 239                 R500_WRITE_MCIND(RV515_MC_AGP_BASE_2, agp_base_hi);
 240         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 241                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 242                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE, agp_base_lo);
 243                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE_2, agp_base_hi);
 244         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600) {
 245                 RS600_WRITE_MCIND(RS600_AGP_BASE, agp_base_lo);
 246                 RS600_WRITE_MCIND(RS600_AGP_BASE_2, agp_base_hi);
 247         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515) {
 248                 R500_WRITE_MCIND(R520_MC_AGP_BASE, agp_base_lo);
 249                 R500_WRITE_MCIND(R520_MC_AGP_BASE_2, agp_base_hi);
 250         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 251                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 252                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 253                 RADEON_WRITE(RS480_AGP_BASE_2, agp_base_hi);
 254         } else {
 255                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 256                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R200)
 257                         RADEON_WRITE(RADEON_AGP_BASE_2, agp_base_hi);
 258         }
 259 }
 260
 261 void radeon_enable_bm(struct drm_radeon_private *dev_priv)
 262 {
 263         u32 tmp;
 264         /* Turn on bus mastering */
 265         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 266             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 267                 /* rs600/rs690/rs740 */
 268                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RS600_BUS_MASTER_DIS;
 269                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 270         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV350) ||
 271                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 272                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 273                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 274                 /* r1xx, r2xx, r300, r(v)350, r420/r481, rs400/rs480 */
 275                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RADEON_BUS_MASTER_DIS;
 276                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 277         } /* PCIE cards appears to not need this */
 278 }
 279
 280 static int RADEON_READ_PLL(struct drm_device * dev, int addr)
 281 {
 282         drm_radeon_private_t *dev_priv = dev->dev_private;
 283
 284         RADEON_WRITE8(RADEON_CLOCK_CNTL_INDEX, addr & 0x1f);
 285         return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 286 }
 287
 288 static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 289 {
 290         RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 291         return RADEON_READ(RADEON_PCIE_DATA);
 292 }
 293
 294 #if RADEON_FIFO_DEBUG
 295 static void radeon_status(drm_radeon_private_t * dev_priv)
 296 {
 297         printk("%s:\n", __func__);
 298         printk("RBBM_STATUS = 0x%08x\n",
 299                (unsigned int)RADEON_READ(RADEON_RBBM_STATUS));
 300         printk("CP_RB_RTPR = 0x%08x\n",
 301                (unsigned int)RADEON_READ(RADEON_CP_RB_RPTR));
 302         printk("CP_RB_WTPR = 0x%08x\n",
 303                (unsigned int)RADEON_READ(RADEON_CP_RB_WPTR));
 304         printk("AIC_CNTL = 0x%08x\n",
 305                (unsigned int)RADEON_READ(RADEON_AIC_CNTL));
 306         printk("AIC_STAT = 0x%08x\n",
 307                (unsigned int)RADEON_READ(RADEON_AIC_STAT));
 308         printk("AIC_PT_BASE = 0x%08x\n",
 309                (unsigned int)RADEON_READ(RADEON_AIC_PT_BASE));
 310         printk("TLB_ADDR = 0x%08x\n",
 311                (unsigned int)RADEON_READ(RADEON_AIC_TLB_ADDR));
 312         printk("TLB_DATA = 0x%08x\n",
 313                (unsigned int)RADEON_READ(RADEON_AIC_TLB_DATA));
 314 }
 315 #endif
 316
 317 /* ================================================================
 318  * Engine, FIFO control
 319  */
 320
 321 static int radeon_do_pixcache_flush(drm_radeon_private_t * dev_priv)
 322 {
 323         u32 tmp;
 324         int i;
 325
 326         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 327
 328         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV280) {
 329                 tmp = RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT);
 330                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 331                 RADEON_WRITE(RADEON_RB3D_DSTCACHE_CTLSTAT, tmp);
 332
 333                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 334                         if (!(RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT)
 335                               & RADEON_RB3D_DC_BUSY)) {
 336                                 return 0;
 337                         }
 338                         DRM_UDELAY(1);
 339                 }
 340         } else {
 341                 /* don't flush or purge cache here or lockup */
 342                 return 0;
 343         }
 344
 345 #if RADEON_FIFO_DEBUG
 346         DRM_ERROR("failed!\n");
 347         radeon_status(dev_priv);
 348 #endif
 349         return -EBUSY;
 350 }
 351
 352 static int radeon_do_wait_for_fifo(drm_radeon_private_t * dev_priv, int entries)
 353 {
 354         int i;
 355
 356         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 357
 358         for (i = 0; i < dev_priv->usec_timeout; i++) {
 359                 int slots = (RADEON_READ(RADEON_RBBM_STATUS)
 360                              & RADEON_RBBM_FIFOCNT_MASK);
 361                 if (slots >= entries)
 362                         return 0;
 363                 DRM_UDELAY(1);
 364         }
 365         DRM_DEBUG("wait for fifo failed status : 0x%08X 0x%08X\n",
 366                  RADEON_READ(RADEON_RBBM_STATUS),
 367                  RADEON_READ(R300_VAP_CNTL_STATUS));
 368
 369 #if RADEON_FIFO_DEBUG
 370         DRM_ERROR("failed!\n");
 371         radeon_status(dev_priv);
 372 #endif
 373         return -EBUSY;
 374 }
 375
 376 static int radeon_do_wait_for_idle(drm_radeon_private_t * dev_priv)
 377 {
 378         int i, ret;
 379
 380         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 381
 382         ret = radeon_do_wait_for_fifo(dev_priv, 64);
 383         if (ret)
 384                 return ret;
 385
 386         for (i = 0; i < dev_priv->usec_timeout; i++) {
 387                 if (!(RADEON_READ(RADEON_RBBM_STATUS)
 388                       & RADEON_RBBM_ACTIVE)) {
 389                         radeon_do_pixcache_flush(dev_priv);
 390                         return 0;
 391                 }
 392                 DRM_UDELAY(1);
 393         }
 394         DRM_DEBUG("wait idle failed status : 0x%08X 0x%08X\n",
 395                  RADEON_READ(RADEON_RBBM_STATUS),
 396                  RADEON_READ(R300_VAP_CNTL_STATUS));
 397
 398 #if RADEON_FIFO_DEBUG
 399         DRM_ERROR("failed!\n");
 400         radeon_status(dev_priv);
 401 #endif
 402         return -EBUSY;
 403 }
 404
 405 static void radeon_init_pipes(drm_radeon_private_t *dev_priv)
 406 {
 407         uint32_t gb_tile_config, gb_pipe_sel = 0;
 408
 409         /* RS4xx/RS6xx/R4xx/R5xx */
 410         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R420) {
 411                 gb_pipe_sel = RADEON_READ(R400_GB_PIPE_SELECT);
 412                 dev_priv->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
 413         } else {
 414                 /* R3xx */
 415                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 416                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350)) {
 417                         dev_priv->num_gb_pipes = 2;
 418                 } else {
 419                         /* R3Vxx */
 420                         dev_priv->num_gb_pipes = 1;
 421                 }
 422         }
 423         DRM_INFO("Num pipes: %d\n", dev_priv->num_gb_pipes);
 424
 425         gb_tile_config = (R300_ENABLE_TILING | R300_TILE_SIZE_16 /*| R300_SUBPIXEL_1_16*/);
 426
 427         switch (dev_priv->num_gb_pipes) {
 428         case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
 429         case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
 430         case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
 431         default:
 432         case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
 433         }
 434
 435         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
 436                 RADEON_WRITE_PLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
 437                 RADEON_WRITE(R300_SU_REG_DEST, ((1 << dev_priv->num_gb_pipes) - 1));
 438         }
 439         RADEON_WRITE(R300_GB_TILE_CONFIG, gb_tile_config);
 440         radeon_do_wait_for_idle(dev_priv);
 441         RADEON_WRITE(R300_DST_PIPE_CONFIG, RADEON_READ(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
 442         RADEON_WRITE(R300_RB2D_DSTCACHE_MODE, (RADEON_READ(R300_RB2D_DSTCACHE_MODE) |
 443                                                R300_DC_AUTOFLUSH_ENABLE |
 444                                                R300_DC_DC_DISABLE_IGNORE_PE));
 445
 446
 447 }
 448
 449 /* ================================================================
 450  * CP control, initialization
 451  */
 452
 453 /* Load the microcode for the CP */
 454 static void radeon_cp_load_microcode(drm_radeon_private_t * dev_priv)
 455 {
 456         int i;
 457         DRM_DEBUG("\n");
 458
 459         radeon_do_wait_for_idle(dev_priv);
 460
 461         RADEON_WRITE(RADEON_CP_ME_RAM_ADDR, 0);
 462         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R100) ||
 463             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV100) ||
 464             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV200) ||
 465             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS100) ||
 466             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS200)) {
 467                 DRM_INFO("Loading R100 Microcode\n");
 468                 for (i = 0; i < 256; i++) {
 469                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 470                                      R100_cp_microcode[i][1]);
 471                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 472                                      R100_cp_microcode[i][0]);
 473                 }
 474         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R200) ||
 475                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV250) ||
 476                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV280) ||
 477                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS300)) {
 478                 DRM_INFO("Loading R200 Microcode\n");
 479                 for (i = 0; i < 256; i++) {
 480                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 481                                      R200_cp_microcode[i][1]);
 482                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 483                                      R200_cp_microcode[i][0]);
 484                 }
 485         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 486                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350) ||
 487                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV350) ||
 488                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV380) ||
 489                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 490                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 491                 DRM_INFO("Loading R300 Microcode\n");
 492                 for (i = 0; i < 256; i++) {
 493                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 494                                      R300_cp_microcode[i][1]);
 495                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 496                                      R300_cp_microcode[i][0]);
 497                 }
 498         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 499                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R423) ||
 500                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV410)) {
 501                 DRM_INFO("Loading R400 Microcode\n");
 502                 for (i = 0; i < 256; i++) {
 503                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 504                                      R420_cp_microcode[i][1]);
 505                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 506                                      R420_cp_microcode[i][0]);
 507                 }
 508         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 509                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 510                 DRM_INFO("Loading RS690/RS740 Microcode\n");
 511                 for (i = 0; i < 256; i++) {
 512                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 513                                      RS690_cp_microcode[i][1]);
 514                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 515                                      RS690_cp_microcode[i][0]);
 516                 }
 517         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600) {
 518                 DRM_INFO("Loading RS600 Microcode\n");
 519                 for (i = 0; i < 256; i++) {
 520                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 521                                      RS600_cp_microcode[i][1]);
 522                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 523                                      RS600_cp_microcode[i][0]);
 524                 }
 525         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) ||
 526                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R520) ||
 527                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530) ||
 528                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R580) ||
 529                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV560) ||
 530                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV570)) {
 531                 DRM_INFO("Loading R500 Microcode\n");
 532                 for (i = 0; i < 256; i++) {
 533                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 534                                      R520_cp_microcode[i][1]);
 535                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 536                                      R520_cp_microcode[i][0]);
 537                 }
 538         }
 539 }
 540
 541 /* Flush any pending commands to the CP.  This should only be used just
 542  * prior to a wait for idle, as it informs the engine that the command
 543  * stream is ending.
 544  */
 545 static void radeon_do_cp_flush(drm_radeon_private_t * dev_priv)
 546 {
 547         DRM_DEBUG("\n");
 548 #if 0
 549         u32 tmp;
 550
 551         tmp = RADEON_READ(RADEON_CP_RB_WPTR) | (1 << 31);
 552         RADEON_WRITE(RADEON_CP_RB_WPTR, tmp);
 553 #endif
 554 }
 555
 556 /* Wait for the CP to go idle.
 557  */
 558 int radeon_do_cp_idle(drm_radeon_private_t * dev_priv)
 559 {
 560         RING_LOCALS;
 561         DRM_DEBUG("\n");
 562
 563         BEGIN_RING(6);
 564
 565         RADEON_PURGE_CACHE();
 566         RADEON_PURGE_ZCACHE();
 567         RADEON_WAIT_UNTIL_IDLE();
 568
 569         ADVANCE_RING();
 570         COMMIT_RING();
 571
 572         return radeon_do_wait_for_idle(dev_priv);
 573 }
 574
 575 /* Start the Command Processor.
 576  */
 577 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 578 {
 579         RING_LOCALS;
 580         DRM_DEBUG("\n");
 581
 582         radeon_do_wait_for_idle(dev_priv);
 583
 584         RADEON_WRITE(RADEON_CP_CSQ_CNTL, dev_priv->cp_mode);
 585
 586         dev_priv->cp_running = 1;
 587
 588         BEGIN_RING(8);
 589         /* isync can only be written through cp on r5xx write it here */
 590         OUT_RING(CP_PACKET0(RADEON_ISYNC_CNTL, 0));
 591         OUT_RING(RADEON_ISYNC_ANY2D_IDLE3D |
 592                  RADEON_ISYNC_ANY3D_IDLE2D |
 593                  RADEON_ISYNC_WAIT_IDLEGUI |
 594                  RADEON_ISYNC_CPSCRATCH_IDLEGUI);
 595         RADEON_PURGE_CACHE();
 596         RADEON_PURGE_ZCACHE();
 597         RADEON_WAIT_UNTIL_IDLE();
 598         ADVANCE_RING();
 599         COMMIT_RING();
 600
 601         dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
 602 }
 603
 604 /* Reset the Command Processor.  This will not flush any pending
 605  * commands, so you must wait for the CP command stream to complete
 606  * before calling this routine.
 607  */
 608 static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
 609 {
 610         u32 cur_read_ptr;
 611         DRM_DEBUG("\n");
 612
 613         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 614         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 615         SET_RING_HEAD(dev_priv, cur_read_ptr);
 616         dev_priv->ring.tail = cur_read_ptr;
 617 }
 618
 619 /* Stop the Command Processor.  This will not flush any pending
 620  * commands, so you must flush the command stream and wait for the CP
 621  * to go idle before calling this routine.
 622  */
 623 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 624 {
 625         DRM_DEBUG("\n");
 626
 627         RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 628
 629         dev_priv->cp_running = 0;
 630 }
 631
 632 /* Reset the engine.  This will stop the CP if it is running.
 633  */
 634 static int radeon_do_engine_reset(struct drm_device * dev)
 635 {
 636         drm_radeon_private_t *dev_priv = dev->dev_private;
 637         u32 clock_cntl_index = 0, mclk_cntl = 0, rbbm_soft_reset;
 638         DRM_DEBUG("\n");
 639
 640         radeon_do_pixcache_flush(dev_priv);
 641
 642         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 643                 /* may need something similar for newer chips */
 644                 clock_cntl_index = RADEON_READ(RADEON_CLOCK_CNTL_INDEX);
 645                 mclk_cntl = RADEON_READ_PLL(dev, RADEON_MCLK_CNTL);
 646
 647                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, (mclk_cntl |
 648                                                     RADEON_FORCEON_MCLKA |
 649                                                     RADEON_FORCEON_MCLKB |
 650                                                     RADEON_FORCEON_YCLKA |
 651                                                     RADEON_FORCEON_YCLKB |
 652                                                     RADEON_FORCEON_MC |
 653                                                     RADEON_FORCEON_AIC));
 654         }
 655
 656         rbbm_soft_reset = RADEON_READ(RADEON_RBBM_SOFT_RESET);
 657
 658         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset |
 659                                               RADEON_SOFT_RESET_CP |
 660                                               RADEON_SOFT_RESET_HI |
 661                                               RADEON_SOFT_RESET_SE |
 662                                               RADEON_SOFT_RESET_RE |
 663                                               RADEON_SOFT_RESET_PP |
 664                                               RADEON_SOFT_RESET_E2 |
 665                                               RADEON_SOFT_RESET_RB));
 666         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 667         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset &
 668                                               ~(RADEON_SOFT_RESET_CP |
 669                                                 RADEON_SOFT_RESET_HI |
 670                                                 RADEON_SOFT_RESET_SE |
 671                                                 RADEON_SOFT_RESET_RE |
 672                                                 RADEON_SOFT_RESET_PP |
 673                                                 RADEON_SOFT_RESET_E2 |
 674                                                 RADEON_SOFT_RESET_RB)));
 675         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 676
 677         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 678                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, mclk_cntl);
 679                 RADEON_WRITE(RADEON_CLOCK_CNTL_INDEX, clock_cntl_index);
 680                 RADEON_WRITE(RADEON_RBBM_SOFT_RESET, rbbm_soft_reset);
 681         }
 682
 683         /* setup the raster pipes */
 684         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R300)
 685             radeon_init_pipes(dev_priv);
 686
 687         /* Reset the CP ring */
 688         radeon_do_cp_reset(dev_priv);
 689
 690         /* The CP is no longer running after an engine reset */
 691         dev_priv->cp_running = 0;
 692
 693         /* Reset any pending vertex, indirect buffers */
 694         radeon_freelist_reset(dev);
 695
 696         return 0;
 697 }
 698
 699 static void radeon_cp_init_ring_buffer(struct drm_device * dev,
 700                                        drm_radeon_private_t *dev_priv,
 701                                        struct drm_file *file_priv)
 702 {
 703         struct drm_radeon_master_private *master_priv;
 704         u32 ring_start, cur_read_ptr;
 705
 706         /* Initialize the memory controller. With new memory map, the fb location
 707          * is not changed, it should have been properly initialized already. Part
 708          * of the problem is that the code below is bogus, assuming the GART is
 709          * always appended to the fb which is not necessarily the case
 710          */
 711         if (!dev_priv->new_memmap)
 712                 radeon_write_fb_location(dev_priv,
 713                              ((dev_priv->gart_vm_start - 1) & 0xffff0000)
 714                              | (dev_priv->fb_location >> 16));
 715
 716 #if __OS_HAS_AGP
 717         if (dev_priv->flags & RADEON_IS_AGP) {
 718                 radeon_write_agp_base(dev_priv, dev->agp->base);
 719
 720                 radeon_write_agp_location(dev_priv,
 721                              (((dev_priv->gart_vm_start - 1 +
 722                                 dev_priv->gart_size) & 0xffff0000) |
 723                               (dev_priv->gart_vm_start >> 16)));
 724
 725                 ring_start = (dev_priv->cp_ring->offset
 726                               - dev->agp->base
 727                               + dev_priv->gart_vm_start);
 728         } else
 729 #endif
 730                 ring_start = (dev_priv->cp_ring->offset
 731                               - (unsigned long)dev->sg->virtual
 732                               + dev_priv->gart_vm_start);
 733
 734         RADEON_WRITE(RADEON_CP_RB_BASE, ring_start);
 735
 736         /* Set the write pointer delay */
 737         RADEON_WRITE(RADEON_CP_RB_WPTR_DELAY, 0);
 738
 739         /* Initialize the ring buffer's read and write pointers */
 740         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 741         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 742         SET_RING_HEAD(dev_priv, cur_read_ptr);
 743         dev_priv->ring.tail = cur_read_ptr;
 744
 745 #if __OS_HAS_AGP
 746         if (dev_priv->flags & RADEON_IS_AGP) {
 747                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 748                              dev_priv->ring_rptr->offset
 749                              - dev->agp->base + dev_priv->gart_vm_start);
 750         } else
 751 #endif
 752         {
 753                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 754                              dev_priv->ring_rptr->offset
 755                              - ((unsigned long) dev->sg->virtual)
 756                              + dev_priv->gart_vm_start);
 757         }
 758
 759         /* Set ring buffer size */
 760 #ifdef __BIG_ENDIAN
 761         RADEON_WRITE(RADEON_CP_RB_CNTL,
 762                      RADEON_BUF_SWAP_32BIT |
 763                      (dev_priv->ring.fetch_size_l2ow << 18) |
 764                      (dev_priv->ring.rptr_update_l2qw << 8) |
 765                      dev_priv->ring.size_l2qw);
 766 #else
 767         RADEON_WRITE(RADEON_CP_RB_CNTL,
 768                      (dev_priv->ring.fetch_size_l2ow << 18) |
 769                      (dev_priv->ring.rptr_update_l2qw << 8) |
 770                      dev_priv->ring.size_l2qw);
 771 #endif
 772
 773
 774         /* Initialize the scratch register pointer.  This will cause
 775          * the scratch register values to be written out to memory
 776          * whenever they are updated.
 777          *
 778          * We simply put this behind the ring read pointer, this works
 779          * with PCI GART as well as (whatever kind of) AGP GART
 780          */
 781         RADEON_WRITE(RADEON_SCRATCH_ADDR, RADEON_READ(RADEON_CP_RB_RPTR_ADDR)
 782                      + RADEON_SCRATCH_REG_OFFSET);
 783
 784         RADEON_WRITE(RADEON_SCRATCH_UMSK, 0x7);
 785
 786         radeon_enable_bm(dev_priv);
 787
 788         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(0), 0);
 789         RADEON_WRITE(RADEON_LAST_FRAME_REG, 0);
 790
 791         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1), 0);
 792         RADEON_WRITE(RADEON_LAST_DISPATCH_REG, 0);
 793
 794         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(2), 0);
 795         RADEON_WRITE(RADEON_LAST_CLEAR_REG, 0);
 796
 797         /* reset sarea copies of these */
 798         master_priv = file_priv->master->driver_priv;
 799         if (master_priv->sarea_priv) {
 800                 master_priv->sarea_priv->last_frame = 0;
 801                 master_priv->sarea_priv->last_dispatch = 0;
 802                 master_priv->sarea_priv->last_clear = 0;
 803         }
 804
 805         radeon_do_wait_for_idle(dev_priv);
 806
 807         /* Sync everything up */
 808         RADEON_WRITE(RADEON_ISYNC_CNTL,
 809                      (RADEON_ISYNC_ANY2D_IDLE3D |
 810                       RADEON_ISYNC_ANY3D_IDLE2D |
 811                       RADEON_ISYNC_WAIT_IDLEGUI |
 812                       RADEON_ISYNC_CPSCRATCH_IDLEGUI));
 813
 814 }
 815
 816 static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 817 {
 818         u32 tmp;
 819
 820         /* Start with assuming that writeback doesn't work */
 821         dev_priv->writeback_works = 0;
 822
 823         /* Writeback doesn't seem to work everywhere, test it here and possibly
 824          * enable it if it appears to work
 825          */
 826         radeon_write_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1), 0);
 827
 828         RADEON_WRITE(RADEON_SCRATCH_REG1, 0xdeadbeef);
 829
 830         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 831                 u32 val;
 832
 833                 val = radeon_read_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1));
 834                 if (val == 0xdeadbeef)
 835                         break;
 836                 DRM_UDELAY(1);
 837         }
 838
 839         if (tmp < dev_priv->usec_timeout) {
 840                 dev_priv->writeback_works = 1;
 841                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 842         } else {
 843                 dev_priv->writeback_works = 0;
 844                 DRM_INFO("writeback test failed\n");
 845         }
 846         if (radeon_no_wb == 1) {
 847                 dev_priv->writeback_works = 0;
 848                 DRM_INFO("writeback forced off\n");
 849         }
 850
 851         if (!dev_priv->writeback_works) {
 852                 /* Disable writeback to avoid unnecessary bus master transfer */
 853                 RADEON_WRITE(RADEON_CP_RB_CNTL, RADEON_READ(RADEON_CP_RB_CNTL) |
 854                              RADEON_RB_NO_UPDATE);
 855                 RADEON_WRITE(RADEON_SCRATCH_UMSK, 0);
 856         }
 857 }
 858
 859 /* Enable or disable IGP GART on the chip */
 860 static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
 861 {
 862         u32 temp;
 863
 864         if (on) {
 865                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 866                           dev_priv->gart_vm_start,
 867                           (long)dev_priv->gart_info.bus_addr,
 868                           dev_priv->gart_size);
 869
 870                 temp = IGP_READ_MCIND(dev_priv, RS480_MC_MISC_CNTL);
 871                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 872                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 873                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, (RS480_GART_INDEX_REG_EN |
 874                                                              RS690_BLOCK_GFX_D3_EN));
 875                 else
 876                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, RS480_GART_INDEX_REG_EN);
 877
 878                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 879                                                                RS480_VA_SIZE_32MB));
 880
 881                 temp = IGP_READ_MCIND(dev_priv, RS480_GART_FEATURE_ID);
 882                 IGP_WRITE_MCIND(RS480_GART_FEATURE_ID, (RS480_HANG_EN |
 883                                                         RS480_TLB_ENABLE |
 884                                                         RS480_GTW_LAC_EN |
 885                                                         RS480_1LEVEL_GART));
 886
 887                 temp = dev_priv->gart_info.bus_addr & 0xfffff000;
 888                 temp |= (upper_32_bits(dev_priv->gart_info.bus_addr) & 0xff) << 4;
 889                 IGP_WRITE_MCIND(RS480_GART_BASE, temp);
 890
 891                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_MODE_CNTL);
 892                 IGP_WRITE_MCIND(RS480_AGP_MODE_CNTL, ((1 << RS480_REQ_TYPE_SNOOP_SHIFT) |
 893                                                       RS480_REQ_TYPE_SNOOP_DIS));
 894
 895                 radeon_write_agp_base(dev_priv, dev_priv->gart_vm_start);
 896
 897                 dev_priv->gart_size = 32*1024*1024;
 898                 temp = (((dev_priv->gart_vm_start - 1 + dev_priv->gart_size) &
 899                          0xffff0000) | (dev_priv->gart_vm_start >> 16));
 900
 901                 radeon_write_agp_location(dev_priv, temp);
 902
 903                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_ADDRESS_SPACE_SIZE);
 904                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 905                                                                RS480_VA_SIZE_32MB));
 906
 907                 do {
 908                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 909                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 910                                 break;
 911                         DRM_UDELAY(1);
 912                 } while (1);
 913
 914                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL,
 915                                 RS480_GART_CACHE_INVALIDATE);
 916
 917                 do {
 918                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 919                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 920                                 break;
 921                         DRM_UDELAY(1);
 922                 } while (1);
 923
 924                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL, 0);
 925         } else {
 926                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, 0);
 927         }
 928 }
 929
 930 /* Enable or disable IGP GART on the chip */
 931 static void rs600_set_igpgart(drm_radeon_private_t *dev_priv, int on)
 932 {
 933         u32 temp;
 934         int i;
 935
 936         if (on) {
 937                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 938                          dev_priv->gart_vm_start,
 939                          (long)dev_priv->gart_info.bus_addr,
 940                          dev_priv->gart_size);
 941
 942                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, (RS600_EFFECTIVE_L2_CACHE_SIZE(6) |
 943                                                     RS600_EFFECTIVE_L2_QUEUE_SIZE(6)));
 944
 945                 for (i = 0; i < 19; i++)
 946                         IGP_WRITE_MCIND(RS600_MC_PT0_CLIENT0_CNTL + i,
 947                                         (RS600_ENABLE_TRANSLATION_MODE_OVERRIDE |
 948                                          RS600_SYSTEM_ACCESS_MODE_IN_SYS |
 949                                          RS600_SYSTEM_APERTURE_UNMAPPED_ACCESS_PASSTHROUGH |
 950                                          RS600_EFFECTIVE_L1_CACHE_SIZE(3) |
 951                                          RS600_ENABLE_FRAGMENT_PROCESSING |
 952                                          RS600_EFFECTIVE_L1_QUEUE_SIZE(3)));
 953
 954                 IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_CNTL, (RS600_ENABLE_PAGE_TABLE |
 955                                                              RS600_PAGE_TABLE_TYPE_FLAT));
 956
 957                 /* disable all other contexts */
 958                 for (i = 1; i < 8; i++)
 959                         IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_CNTL + i, 0);
 960
 961                 /* setup the page table aperture */
 962                 IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_FLAT_BASE_ADDR,
 963                                 dev_priv->gart_info.bus_addr);
 964                 IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_FLAT_START_ADDR,
 965                                 dev_priv->gart_vm_start);
 966                 IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_FLAT_END_ADDR,
 967                                 (dev_priv->gart_vm_start + dev_priv->gart_size - 1));
 968                 IGP_WRITE_MCIND(RS600_MC_PT0_CONTEXT0_DEFAULT_READ_ADDR, 0);
 969
 970                 /* setup the system aperture */
 971                 IGP_WRITE_MCIND(RS600_MC_PT0_SYSTEM_APERTURE_LOW_ADDR,
 972                                 dev_priv->gart_vm_start);
 973                 IGP_WRITE_MCIND(RS600_MC_PT0_SYSTEM_APERTURE_HIGH_ADDR,
 974                                 (dev_priv->gart_vm_start + dev_priv->gart_size - 1));
 975
 976                 /* enable page tables */
 977                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_PT0_CNTL);
 978                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, (temp | RS600_ENABLE_PT));
 979
 980                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_CNTL1);
 981                 IGP_WRITE_MCIND(RS600_MC_CNTL1, (temp | RS600_ENABLE_PAGE_TABLES));
 982
 983                 /* invalidate the cache */
 984                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_PT0_CNTL);
 985
 986                 temp &= ~(RS600_INVALIDATE_ALL_L1_TLBS | RS600_INVALIDATE_L2_CACHE);
 987                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, temp);
 988                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_PT0_CNTL);
 989
 990                 temp |= RS600_INVALIDATE_ALL_L1_TLBS | RS600_INVALIDATE_L2_CACHE;
 991                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, temp);
 992                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_PT0_CNTL);
 993
 994                 temp &= ~(RS600_INVALIDATE_ALL_L1_TLBS | RS600_INVALIDATE_L2_CACHE);
 995                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, temp);
 996                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_PT0_CNTL);
 997
 998         } else {
 999                 IGP_WRITE_MCIND(RS600_MC_PT0_CNTL, 0);
1000                 temp = IGP_READ_MCIND(dev_priv, RS600_MC_CNTL1);
1001                 temp &= ~RS600_ENABLE_PAGE_TABLES;
1002                 IGP_WRITE_MCIND(RS600_MC_CNTL1, temp);
1003         }
1004 }
1005
1006 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
1007 {
1008         u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
1009         if (on) {
1010
1011                 DRM_DEBUG("programming pcie %08X %08lX %08X\n",
1012                           dev_priv->gart_vm_start,
1013                           (long)dev_priv->gart_info.bus_addr,
1014                           dev_priv->gart_size);
1015                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_DISCARD_RD_ADDR_LO,
1016                                   dev_priv->gart_vm_start);
1017                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_BASE,
1018                                   dev_priv->gart_info.bus_addr);
1019                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_START_LO,
1020                                   dev_priv->gart_vm_start);
1021                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_END_LO,
1022                                   dev_priv->gart_vm_start +
1023                                   dev_priv->gart_size - 1);
1024
1025                 radeon_write_agp_location(dev_priv, 0xffffffc0); /* ?? */
1026
1027                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
1028                                   RADEON_PCIE_TX_GART_EN);
1029         } else {
1030                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
1031                                   tmp & ~RADEON_PCIE_TX_GART_EN);
1032         }
1033 }
1034
1035 /* Enable or disable PCI GART on the chip */
1036 static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
1037 {
1038         u32 tmp;
1039
1040         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
1041             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740) ||
1042             (dev_priv->flags & RADEON_IS_IGPGART)) {
1043                 radeon_set_igpgart(dev_priv, on);
1044                 return;
1045         }
1046
1047         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600) {
1048                 rs600_set_igpgart(dev_priv, on);
1049                 return;
1050         }
1051
1052         if (dev_priv->flags & RADEON_IS_PCIE) {
1053                 radeon_set_pciegart(dev_priv, on);
1054                 return;
1055         }
1056
1057         tmp = RADEON_READ(RADEON_AIC_CNTL);
1058
1059         if (on) {
1060                 RADEON_WRITE(RADEON_AIC_CNTL,
1061                              tmp | RADEON_PCIGART_TRANSLATE_EN);
1062
1063                 /* set PCI GART page-table base address
1064                  */
1065                 RADEON_WRITE(RADEON_AIC_PT_BASE, dev_priv->gart_info.bus_addr);
1066
1067                 /* set address range for PCI address translate
1068                  */
1069                 RADEON_WRITE(RADEON_AIC_LO_ADDR, dev_priv->gart_vm_start);
1070                 RADEON_WRITE(RADEON_AIC_HI_ADDR, dev_priv->gart_vm_start
1071                              + dev_priv->gart_size - 1);
1072
1073                 /* Turn off AGP aperture -- is this required for PCI GART?
1074                  */
1075                 radeon_write_agp_location(dev_priv, 0xffffffc0);
1076                 RADEON_WRITE(RADEON_AGP_COMMAND, 0);    /* clear AGP_COMMAND */
1077         } else {
1078                 RADEON_WRITE(RADEON_AIC_CNTL,
1079                              tmp & ~RADEON_PCIGART_TRANSLATE_EN);
1080         }
1081 }
1082
1083 static int radeon_setup_pcigart_surface(drm_radeon_private_t *dev_priv)
1084 {
1085         struct drm_ati_pcigart_info *gart_info = &dev_priv->gart_info;
1086         struct radeon_virt_surface *vp;
1087         int i;
1088
1089         for (i = 0; i < RADEON_MAX_SURFACES * 2; i++) {
1090                 if (!dev_priv->virt_surfaces[i].file_priv ||
1091                     dev_priv->virt_surfaces[i].file_priv == PCIGART_FILE_PRIV)
1092                         break;
1093         }
1094         if (i >= 2 * RADEON_MAX_SURFACES)
1095                 return -ENOMEM;
1096         vp = &dev_priv->virt_surfaces[i];
1097
1098         for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1099                 struct radeon_surface *sp = &dev_priv->surfaces[i];
1100                 if (sp->refcount)
1101                         continue;
1102
1103                 vp->surface_index = i;
1104                 vp->lower = gart_info->bus_addr;
1105                 vp->upper = vp->lower + gart_info->table_size;
1106                 vp->flags = 0;
1107                 vp->file_priv = PCIGART_FILE_PRIV;
1108
1109                 sp->refcount = 1;
1110                 sp->lower = vp->lower;
1111                 sp->upper = vp->upper;
1112                 sp->flags = 0;
1113
1114                 RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, sp->flags);
1115                 RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND + 16 * i, sp->lower);
1116                 RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND + 16 * i, sp->upper);
1117                 return 0;
1118         }
1119
1120         return -ENOMEM;
1121 }
1122
1123 static int radeon_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
1124                              struct drm_file *file_priv)
1125 {
1126         drm_radeon_private_t *dev_priv = dev->dev_private;
1127         struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
1128
1129         DRM_DEBUG("\n");
1130
1131         /* if we require new memory map but we don't have it fail */
1132         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
1133                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
1134                 radeon_do_cleanup_cp(dev);
1135                 return -EINVAL;
1136         }
1137
1138         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
1139                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
1140                 dev_priv->flags &= ~RADEON_IS_AGP;
1141         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
1142                    && !init->is_pci) {
1143                 DRM_DEBUG("Restoring AGP flag\n");
1144                 dev_priv->flags |= RADEON_IS_AGP;
1145         }
1146
1147         if ((!(dev_priv->flags & RADEON_IS_AGP)) && !dev->sg) {
1148                 DRM_ERROR("PCI GART memory not allocated!\n");
1149                 radeon_do_cleanup_cp(dev);
1150                 return -EINVAL;
1151         }
1152
1153         dev_priv->usec_timeout = init->usec_timeout;
1154         if (dev_priv->usec_timeout < 1 ||
1155             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
1156                 DRM_DEBUG("TIMEOUT problem!\n");
1157                 radeon_do_cleanup_cp(dev);
1158                 return -EINVAL;
1159         }
1160
1161         /* Enable vblank on CRTC1 for older X servers
1162          */
1163         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
1164
1165         switch(init->func) {
1166         case RADEON_INIT_R200_CP:
1167                 dev_priv->microcode_version = UCODE_R200;
1168                 break;
1169         case RADEON_INIT_R300_CP:
1170                 dev_priv->microcode_version = UCODE_R300;
1171                 break;
1172         default:
1173                 dev_priv->microcode_version = UCODE_R100;
1174         }
1175
1176         dev_priv->do_boxes = 0;
1177         dev_priv->cp_mode = init->cp_mode;
1178
1179         /* We don't support anything other than bus-mastering ring mode,
1180          * but the ring can be in either AGP or PCI space for the ring
1181          * read pointer.
1182          */
1183         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
1184             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
1185                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
1186                 radeon_do_cleanup_cp(dev);
1187                 return -EINVAL;
1188         }
1189
1190         switch (init->fb_bpp) {
1191         case 16:
1192                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
1193                 break;
1194         case 32:
1195         default:
1196                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
1197                 break;
1198         }
1199         dev_priv->front_offset = init->front_offset;
1200         dev_priv->front_pitch = init->front_pitch;
1201         dev_priv->back_offset = init->back_offset;
1202         dev_priv->back_pitch = init->back_pitch;
1203
1204         switch (init->depth_bpp) {
1205         case 16:
1206                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_16BIT_INT_Z;
1207                 break;
1208         case 32:
1209         default:
1210                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_24BIT_INT_Z;
1211                 break;
1212         }
1213         dev_priv->depth_offset = init->depth_offset;
1214         dev_priv->depth_pitch = init->depth_pitch;
1215
1216         /* Hardware state for depth clears.  Remove this if/when we no
1217          * longer clear the depth buffer with a 3D rectangle.  Hard-code
1218          * all values to prevent unwanted 3D state from slipping through
1219          * and screwing with the clear operation.
1220          */
1221         dev_priv->depth_clear.rb3d_cntl = (RADEON_PLANE_MASK_ENABLE |
1222                                            (dev_priv->color_fmt << 10) |
1223                                            (dev_priv->microcode_version ==
1224                                             UCODE_R100 ? RADEON_ZBLOCK16 : 0));
1225
1226         dev_priv->depth_clear.rb3d_zstencilcntl =
1227             (dev_priv->depth_fmt |
1228              RADEON_Z_TEST_ALWAYS |
1229              RADEON_STENCIL_TEST_ALWAYS |
1230              RADEON_STENCIL_S_FAIL_REPLACE |
1231              RADEON_STENCIL_ZPASS_REPLACE |
1232              RADEON_STENCIL_ZFAIL_REPLACE | RADEON_Z_WRITE_ENABLE);
1233
1234         dev_priv->depth_clear.se_cntl = (RADEON_FFACE_CULL_CW |
1235                                          RADEON_BFACE_SOLID |
1236                                          RADEON_FFACE_SOLID |
1237                                          RADEON_FLAT_SHADE_VTX_LAST |
1238                                          RADEON_DIFFUSE_SHADE_FLAT |
1239                                          RADEON_ALPHA_SHADE_FLAT |
1240                                          RADEON_SPECULAR_SHADE_FLAT |
1241                                          RADEON_FOG_SHADE_FLAT |
1242                                          RADEON_VTX_PIX_CENTER_OGL |
1243                                          RADEON_ROUND_MODE_TRUNC |
1244                                          RADEON_ROUND_PREC_8TH_PIX);
1245
1246
1247         dev_priv->ring_offset = init->ring_offset;
1248         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
1249         dev_priv->buffers_offset = init->buffers_offset;
1250         dev_priv->gart_textures_offset = init->gart_textures_offset;
1251
1252         master_priv->sarea = drm_getsarea(dev);
1253         if (!master_priv->sarea) {
1254                 DRM_ERROR("could not find sarea!\n");
1255                 radeon_do_cleanup_cp(dev);
1256                 return -EINVAL;
1257         }
1258
1259         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
1260         if (!dev_priv->cp_ring) {
1261                 DRM_ERROR("could not find cp ring region!\n");
1262                 radeon_do_cleanup_cp(dev);
1263                 return -EINVAL;
1264         }
1265         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
1266         if (!dev_priv->ring_rptr) {
1267                 DRM_ERROR("could not find ring read pointer!\n");
1268                 radeon_do_cleanup_cp(dev);
1269                 return -EINVAL;
1270         }
1271         dev->agp_buffer_token = init->buffers_offset;
1272         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1273         if (!dev->agp_buffer_map) {
1274                 DRM_ERROR("could not find dma buffer region!\n");
1275                 radeon_do_cleanup_cp(dev);
1276                 return -EINVAL;
1277         }
1278
1279         if (init->gart_textures_offset) {
1280                 dev_priv->gart_textures =
1281                     drm_core_findmap(dev, init->gart_textures_offset);
1282                 if (!dev_priv->gart_textures) {
1283                         DRM_ERROR("could not find GART texture region!\n");
1284                         radeon_do_cleanup_cp(dev);
1285                         return -EINVAL;
1286                 }
1287         }
1288
1289 #if __OS_HAS_AGP
1290         if (dev_priv->flags & RADEON_IS_AGP) {
1291                 drm_core_ioremap_wc(dev_priv->cp_ring, dev);
1292                 drm_core_ioremap_wc(dev_priv->ring_rptr, dev);
1293                 drm_core_ioremap_wc(dev->agp_buffer_map, dev);
1294                 if (!dev_priv->cp_ring->handle ||
1295                     !dev_priv->ring_rptr->handle ||
1296                     !dev->agp_buffer_map->handle) {
1297                         DRM_ERROR("could not find ioremap agp regions!\n");
1298                         radeon_do_cleanup_cp(dev);
1299                         return -EINVAL;
1300                 }
1301         } else
1302 #endif
1303         {
1304                 dev_priv->cp_ring->handle =
1305                         (void *)(unsigned long)dev_priv->cp_ring->offset;
1306                 dev_priv->ring_rptr->handle =
1307                         (void *)(unsigned long)dev_priv->ring_rptr->offset;
1308                 dev->agp_buffer_map->handle =
1309                         (void *)(unsigned long)dev->agp_buffer_map->offset;
1310
1311                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
1312                           dev_priv->cp_ring->handle);
1313                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
1314                           dev_priv->ring_rptr->handle);
1315                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
1316                           dev->agp_buffer_map->handle);
1317         }
1318
1319         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 16;
1320         dev_priv->fb_size =
1321                 ((radeon_read_fb_location(dev_priv) & 0xffff0000u) + 0x10000)
1322                 - dev_priv->fb_location;
1323
1324         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
1325                                         ((dev_priv->front_offset
1326                                           + dev_priv->fb_location) >> 10));
1327
1328         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
1329                                        ((dev_priv->back_offset
1330                                          + dev_priv->fb_location) >> 10));
1331
1332         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
1333                                         ((dev_priv->depth_offset
1334                                           + dev_priv->fb_location) >> 10));
1335
1336         dev_priv->gart_size = init->gart_size;
1337
1338         /* New let's set the memory map ... */
1339         if (dev_priv->new_memmap) {
1340                 u32 base = 0;
1341
1342                 DRM_INFO("Setting GART location based on new memory map\n");
1343
1344                 /* If using AGP, try to locate the AGP aperture at the same
1345                  * location in the card and on the bus, though we have to
1346                  * align it down.
1347                  */
1348 #if __OS_HAS_AGP
1349                 if (dev_priv->flags & RADEON_IS_AGP) {
1350                         base = dev->agp->base;
1351                         /* Check if valid */
1352                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
1353                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
1354                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
1355                                          dev->agp->base);
1356                                 base = 0;
1357                         }
1358                 }
1359 #endif
1360                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
1361                 if (base == 0) {
1362                         base = dev_priv->fb_location + dev_priv->fb_size;
1363                         if (base < dev_priv->fb_location ||
1364                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
1365                                 base = dev_priv->fb_location
1366                                         - dev_priv->gart_size;
1367                 }
1368                 dev_priv->gart_vm_start = base & 0xffc00000u;
1369                 if (dev_priv->gart_vm_start != base)
1370                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
1371                                  base, dev_priv->gart_vm_start);
1372         } else {
1373                 DRM_INFO("Setting GART location based on old memory map\n");
1374                 dev_priv->gart_vm_start = dev_priv->fb_location +
1375                         RADEON_READ(RADEON_CONFIG_APER_SIZE);
1376         }
1377
1378 #if __OS_HAS_AGP
1379         if (dev_priv->flags & RADEON_IS_AGP)
1380                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1381                                                  - dev->agp->base
1382                                                  + dev_priv->gart_vm_start);
1383         else
1384 #endif
1385                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1386                                         - (unsigned long)dev->sg->virtual
1387                                         + dev_priv->gart_vm_start);
1388
1389         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
1390         DRM_DEBUG("dev_priv->gart_vm_start 0x%x\n", dev_priv->gart_vm_start);
1391         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%lx\n",
1392                   dev_priv->gart_buffers_offset);
1393
1394         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
1395         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
1396                               + init->ring_size / sizeof(u32));
1397         dev_priv->ring.size = init->ring_size;
1398         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
1399
1400         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
1401         dev_priv->ring.rptr_update_l2qw = drm_order( /* init->rptr_update */ 4096 / 8);
1402
1403         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
1404         dev_priv->ring.fetch_size_l2ow = drm_order( /* init->fetch_size */ 32 / 16);
1405         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
1406
1407         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
1408
1409 #if __OS_HAS_AGP
1410         if (dev_priv->flags & RADEON_IS_AGP) {
1411                 /* Turn off PCI GART */
1412                 radeon_set_pcigart(dev_priv, 0);
1413         } else
1414 #endif
1415         {
1416                 u32 sctrl;
1417                 int ret;
1418
1419                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
1420                 /* if we have an offset set from userspace */
1421                 if (dev_priv->pcigart_offset_set) {
1422                         dev_priv->gart_info.bus_addr =
1423                                 (resource_size_t)dev_priv->pcigart_offset + dev_priv->fb_location;
1424                         dev_priv->gart_info.mapping.offset =
1425                             dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
1426                         dev_priv->gart_info.mapping.size =
1427                             dev_priv->gart_info.table_size;
1428
1429                         drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
1430                         dev_priv->gart_info.addr =
1431                             dev_priv->gart_info.mapping.handle;
1432
1433                         if (dev_priv->flags & RADEON_IS_PCIE)
1434                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
1435                         else
1436                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1437                         dev_priv->gart_info.gart_table_location =
1438                             DRM_ATI_GART_FB;
1439
1440                         DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
1441                                   dev_priv->gart_info.addr,
1442                                   dev_priv->pcigart_offset);
1443                 } else {
1444                         if (dev_priv->flags & RADEON_IS_IGPGART)
1445                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
1446                         else
1447                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1448                         dev_priv->gart_info.gart_table_location =
1449                             DRM_ATI_GART_MAIN;
1450                         dev_priv->gart_info.addr = NULL;
1451                         dev_priv->gart_info.bus_addr = 0;
1452                         if (dev_priv->flags & RADEON_IS_PCIE) {
1453                                 DRM_ERROR
1454                                     ("Cannot use PCI Express without GART in FB memory\n");
1455                                 radeon_do_cleanup_cp(dev);
1456                                 return -EINVAL;
1457                         }
1458                 }
1459
1460                 sctrl = RADEON_READ(RADEON_SURFACE_CNTL);
1461                 RADEON_WRITE(RADEON_SURFACE_CNTL, 0);
1462                 if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
1463                         ret = r600_page_table_init(dev);
1464                 else
1465                         ret = drm_ati_pcigart_init(dev, &dev_priv->gart_info);
1466                 RADEON_WRITE(RADEON_SURFACE_CNTL, sctrl);
1467
1468                 if (!ret) {
1469                         DRM_ERROR("failed to init PCI GART!\n");
1470                         radeon_do_cleanup_cp(dev);
1471                         return -ENOMEM;
1472                 }
1473
1474                 ret = radeon_setup_pcigart_surface(dev_priv);
1475                 if (ret) {
1476                         DRM_ERROR("failed to setup GART surface!\n");
1477                         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
1478                                 r600_page_table_cleanup(dev, &dev_priv->gart_info);
1479                         else
1480                                 drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info);
1481                         radeon_do_cleanup_cp(dev);
1482                         return ret;
1483                 }
1484
1485                 /* Turn on PCI GART */
1486                 radeon_set_pcigart(dev_priv, 1);
1487         }
1488
1489         radeon_cp_load_microcode(dev_priv);
1490         radeon_cp_init_ring_buffer(dev, dev_priv, file_priv);
1491
1492         dev_priv->last_buf = 0;
1493
1494         radeon_do_engine_reset(dev);
1495         radeon_test_writeback(dev_priv);
1496
1497         return 0;
1498 }
1499
1500 static int radeon_do_cleanup_cp(struct drm_device * dev)
1501 {
1502         drm_radeon_private_t *dev_priv = dev->dev_private;
1503         DRM_DEBUG("\n");
1504
1505         /* Make sure interrupts are disabled here because the uninstall ioctl
1506          * may not have been called from userspace and after dev_private
1507          * is freed, it's too late.
1508          */
1509         if (dev->irq_enabled)
1510                 drm_irq_uninstall(dev);
1511
1512 #if __OS_HAS_AGP
1513         if (dev_priv->flags & RADEON_IS_AGP) {
1514                 if (dev_priv->cp_ring != NULL) {
1515                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1516                         dev_priv->cp_ring = NULL;
1517                 }
1518                 if (dev_priv->ring_rptr != NULL) {
1519                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1520                         dev_priv->ring_rptr = NULL;
1521                 }
1522                 if (dev->agp_buffer_map != NULL) {
1523                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1524                         dev->agp_buffer_map = NULL;
1525                 }
1526         } else
1527 #endif
1528         {
1529
1530                 if (dev_priv->gart_info.bus_addr) {
1531                         /* Turn off PCI GART */
1532                         radeon_set_pcigart(dev_priv, 0);
1533                         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS600)
1534                                 r600_page_table_cleanup(dev, &dev_priv->gart_info);
1535                         else {
1536                                 if (!drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info))
1537                                         DRM_ERROR("failed to cleanup PCI GART!\n");
1538                         }
1539                 }
1540
1541                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB)
1542                 {
1543                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1544                         dev_priv->gart_info.addr = NULL;
1545                 }
1546         }
1547         /* only clear to the start of flags */
1548         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1549
1550         return 0;
1551 }
1552
1553 /* This code will reinit the Radeon CP hardware after a resume from disc.
1554  * AFAIK, it would be very difficult to pickle the state at suspend time, so
1555  * here we make sure that all Radeon hardware initialisation is re-done without
1556  * affecting running applications.
1557  *
1558  * Charl P. Botha <http://cpbotha.net>
1559  */
1560 static int radeon_do_resume_cp(struct drm_device *dev, struct drm_file *file_priv)
1561 {
1562         drm_radeon_private_t *dev_priv = dev->dev_private;
1563
1564         if (!dev_priv) {
1565                 DRM_ERROR("Called with no initialization\n");
1566                 return -EINVAL;
1567         }
1568
1569         DRM_DEBUG("Starting radeon_do_resume_cp()\n");
1570
1571 #if __OS_HAS_AGP
1572         if (dev_priv->flags & RADEON_IS_AGP) {
1573                 /* Turn off PCI GART */
1574                 radeon_set_pcigart(dev_priv, 0);
1575         } else
1576 #endif
1577         {
1578                 /* Turn on PCI GART */
1579                 radeon_set_pcigart(dev_priv, 1);
1580         }
1581
1582         radeon_cp_load_microcode(dev_priv);
1583         radeon_cp_init_ring_buffer(dev, dev_priv, file_priv);
1584
1585         radeon_do_engine_reset(dev);
1586         radeon_irq_set_state(dev, RADEON_SW_INT_ENABLE, 1);
1587
1588         DRM_DEBUG("radeon_do_resume_cp() complete\n");
1589
1590         return 0;
1591 }
1592
1593 int radeon_cp_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
1594 {
1595         drm_radeon_private_t *dev_priv = dev->dev_private;
1596         drm_radeon_init_t *init = data;
1597
1598         LOCK_TEST_WITH_RETURN(dev, file_priv);
1599
1600         if (init->func == RADEON_INIT_R300_CP)
1601                 r300_init_reg_flags(dev);
1602
1603         switch (init->func) {
1604         case RADEON_INIT_CP:
1605         case RADEON_INIT_R200_CP:
1606         case RADEON_INIT_R300_CP:
1607                 return radeon_do_init_cp(dev, init, file_priv);
1608         case RADEON_INIT_R600_CP:
1609                 return r600_do_init_cp(dev, init, file_priv);
1610         case RADEON_CLEANUP_CP:
1611                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1612                         return r600_do_cleanup_cp(dev);
1613                 else
1614                         return radeon_do_cleanup_cp(dev);
1615         }
1616
1617         return -EINVAL;
1618 }
1619
1620 int radeon_cp_start(struct drm_device *dev, void *data, struct drm_file *file_priv)
1621 {
1622         drm_radeon_private_t *dev_priv = dev->dev_private;
1623         DRM_DEBUG("\n");
1624
1625         LOCK_TEST_WITH_RETURN(dev, file_priv);
1626
1627         if (dev_priv->cp_running) {
1628                 DRM_DEBUG("while CP running\n");
1629                 return 0;
1630         }
1631         if (dev_priv->cp_mode == RADEON_CSQ_PRIDIS_INDDIS) {
1632                 DRM_DEBUG("called with bogus CP mode (%d)\n",
1633                           dev_priv->cp_mode);
1634                 return 0;
1635         }
1636
1637         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1638                 r600_do_cp_start(dev_priv);
1639         else
1640                 radeon_do_cp_start(dev_priv);
1641
1642         return 0;
1643 }
1644
1645 /* Stop the CP.  The engine must have been idled before calling this
1646  * routine.
1647  */
1648 int radeon_cp_stop(struct drm_device *dev, void *data, struct drm_file *file_priv)
1649 {
1650         drm_radeon_private_t *dev_priv = dev->dev_private;
1651         drm_radeon_cp_stop_t *stop = data;
1652         int ret;
1653         DRM_DEBUG("\n");
1654
1655         LOCK_TEST_WITH_RETURN(dev, file_priv);
1656
1657         if (!dev_priv->cp_running)
1658                 return 0;
1659
1660         /* Flush any pending CP commands.  This ensures any outstanding
1661          * commands are exectuted by the engine before we turn it off.
1662          */
1663         if (stop->flush) {
1664                 radeon_do_cp_flush(dev_priv);
1665         }
1666
1667         /* If we fail to make the engine go idle, we return an error
1668          * code so that the DRM ioctl wrapper can try again.
1669          */
1670         if (stop->idle) {
1671                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1672                         ret = r600_do_cp_idle(dev_priv);
1673                 else
1674                         ret = radeon_do_cp_idle(dev_priv);
1675                 if (ret)
1676                         return ret;
1677         }
1678
1679         /* Finally, we can turn off the CP.  If the engine isn't idle,
1680          * we will get some dropped triangles as they won't be fully
1681          * rendered before the CP is shut down.
1682          */
1683         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1684                 r600_do_cp_stop(dev_priv);
1685         else
1686                 radeon_do_cp_stop(dev_priv);
1687
1688         /* Reset the engine */
1689         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1690                 r600_do_engine_reset(dev);
1691         else
1692                 radeon_do_engine_reset(dev);
1693
1694         return 0;
1695 }
1696
1697 void radeon_do_release(struct drm_device * dev)
1698 {
1699         drm_radeon_private_t *dev_priv = dev->dev_private;
1700         int i, ret;
1701
1702         if (dev_priv) {
1703                 if (dev_priv->cp_running) {
1704                         /* Stop the cp */
1705                         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600) {
1706                                 while ((ret = r600_do_cp_idle(dev_priv)) != 0) {
1707                                         DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1708 #ifdef __linux__
1709                                         schedule();
1710 #else
1711                                         tsleep(&ret, PZERO, "rdnrel", 1);
1712 #endif
1713                                 }
1714                         } else {
1715                                 while ((ret = radeon_do_cp_idle(dev_priv)) != 0) {
1716                                         DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1717 #ifdef __linux__
1718                                         schedule();
1719 #else
1720                                         tsleep(&ret, PZERO, "rdnrel", 1);
1721 #endif
1722                                 }
1723                         }
1724                         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600) {
1725                                 r600_do_cp_stop(dev_priv);
1726                                 r600_do_engine_reset(dev);
1727                         } else {
1728                                 radeon_do_cp_stop(dev_priv);
1729                                 radeon_do_engine_reset(dev);
1730                         }
1731                 }
1732
1733                 if ((dev_priv->flags & RADEON_FAMILY_MASK) < CHIP_R600) {
1734                         /* Disable *all* interrupts */
1735                         if (dev_priv->mmio)     /* remove this after permanent addmaps */
1736                                 RADEON_WRITE(RADEON_GEN_INT_CNTL, 0);
1737
1738                         if (dev_priv->mmio) {   /* remove all surfaces */
1739                                 for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1740                                         RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, 0);
1741                                         RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND +
1742                                                      16 * i, 0);
1743                                         RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND +
1744                                                      16 * i, 0);
1745                                 }
1746                         }
1747                 }
1748
1749                 /* Free memory heap structures */
1750                 radeon_mem_takedown(&(dev_priv->gart_heap));
1751                 radeon_mem_takedown(&(dev_priv->fb_heap));
1752
1753                 /* deallocate kernel resources */
1754                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1755                         r600_do_cleanup_cp(dev);
1756                 else
1757                         radeon_do_cleanup_cp(dev);
1758         }
1759 }
1760
1761 /* Just reset the CP ring.  Called as part of an X Server engine reset.
1762  */
1763 int radeon_cp_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1764 {
1765         drm_radeon_private_t *dev_priv = dev->dev_private;
1766         DRM_DEBUG("\n");
1767
1768         LOCK_TEST_WITH_RETURN(dev, file_priv);
1769
1770         if (!dev_priv) {
1771                 DRM_DEBUG("called before init done\n");
1772                 return -EINVAL;
1773         }
1774
1775         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1776                 r600_do_cp_reset(dev_priv);
1777         else
1778                 radeon_do_cp_reset(dev_priv);
1779
1780         /* The CP is no longer running after an engine reset */
1781         dev_priv->cp_running = 0;
1782
1783         return 0;
1784 }
1785
1786 int radeon_cp_idle(struct drm_device *dev, void *data, struct drm_file *file_priv)
1787 {
1788         drm_radeon_private_t *dev_priv = dev->dev_private;
1789         DRM_DEBUG("\n");
1790
1791         LOCK_TEST_WITH_RETURN(dev, file_priv);
1792
1793         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1794                 return r600_do_cp_idle(dev_priv);
1795         else
1796                 return radeon_do_cp_idle(dev_priv);
1797 }
1798
1799 /* Added by Charl P. Botha to call radeon_do_resume_cp().
1800  */
1801 int radeon_cp_resume(struct drm_device *dev, void *data, struct drm_file *file_priv)
1802 {
1803         drm_radeon_private_t *dev_priv = dev->dev_private;
1804         DRM_DEBUG("\n");
1805
1806         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1807                 return r600_do_resume_cp(dev, file_priv);
1808         else
1809                 return radeon_do_resume_cp(dev, file_priv);
1810 }
1811
1812 int radeon_engine_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1813 {
1814         drm_radeon_private_t *dev_priv = dev->dev_private;
1815         DRM_DEBUG("\n");
1816
1817         LOCK_TEST_WITH_RETURN(dev, file_priv);
1818
1819         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600)
1820                 return r600_do_engine_reset(dev);
1821         else
1822                 return radeon_do_engine_reset(dev);
1823 }
1824
1825 /* ================================================================
1826  * Fullscreen mode
1827  */
1828
1829 /* KW: Deprecated to say the least:
1830  */
1831 int radeon_fullscreen(struct drm_device *dev, void *data, struct drm_file *file_priv)
1832 {
1833         return 0;
1834 }
1835
1836 /* ================================================================
1837  * Freelist management
1838  */
1839
1840 /* Original comment: FIXME: ROTATE_BUFS is a hack to cycle through
1841  *   bufs until freelist code is used.  Note this hides a problem with
1842  *   the scratch register * (used to keep track of last buffer
1843  *   completed) being written to before * the last buffer has actually
1844  *   completed rendering.
1845  *
1846  * KW:  It's also a good way to find free buffers quickly.
1847  *
1848  * KW: Ideally this loop wouldn't exist, and freelist_get wouldn't
1849  * sleep.  However, bugs in older versions of radeon_accel.c mean that
1850  * we essentially have to do this, else old clients will break.
1851  *
1852  * However, it does leave open a potential deadlock where all the
1853  * buffers are held by other clients, which can't release them because
1854  * they can't get the lock.
1855  */
1856
1857 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1858 {
1859         struct drm_device_dma *dma = dev->dma;
1860         drm_radeon_private_t *dev_priv = dev->dev_private;
1861         drm_radeon_buf_priv_t *buf_priv;
1862         struct drm_buf *buf;
1863         int i, t;
1864         int start;
1865
1866         if (++dev_priv->last_buf >= dma->buf_count)
1867                 dev_priv->last_buf = 0;
1868
1869         start = dev_priv->last_buf;
1870
1871         for (t = 0; t < dev_priv->usec_timeout; t++) {
1872                 u32 done_age = GET_SCRATCH(dev_priv, 1);
1873                 DRM_DEBUG("done_age = %d\n", done_age);
1874                 for (i = start; i < dma->buf_count; i++) {
1875                         buf = dma->buflist[i];
1876                         buf_priv = buf->dev_private;
1877                         if (buf->file_priv == NULL || (buf->pending &&
1878                                                        buf_priv->age <=
1879                                                        done_age)) {
1880                                 dev_priv->stats.requested_bufs++;
1881                                 buf->pending = 0;
1882                                 return buf;
1883                         }
1884                         start = 0;
1885                 }
1886
1887                 if (t) {
1888                         DRM_UDELAY(1);
1889                         dev_priv->stats.freelist_loops++;
1890                 }
1891         }
1892
1893         DRM_DEBUG("returning NULL!\n");
1894         return NULL;
1895 }
1896
1897 #if 0
1898 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1899 {
1900         struct drm_device_dma *dma = dev->dma;
1901         drm_radeon_private_t *dev_priv = dev->dev_private;
1902         drm_radeon_buf_priv_t *buf_priv;
1903         struct drm_buf *buf;
1904         int i, t;
1905         int start;
1906         u32 done_age;
1907
1908         done_age = radeon_read_ring_rptr(dev_priv, RADEON_SCRATCHOFF(1));
1909         if (++dev_priv->last_buf >= dma->buf_count)
1910                 dev_priv->last_buf = 0;
1911
1912         start = dev_priv->last_buf;
1913         dev_priv->stats.freelist_loops++;
1914
1915         for (t = 0; t < 2; t++) {
1916                 for (i = start; i < dma->buf_count; i++) {
1917                         buf = dma->buflist[i];
1918                         buf_priv = buf->dev_private;
1919                         if (buf->file_priv == 0 || (buf->pending &&
1920                                                     buf_priv->age <=
1921                                                     done_age)) {
1922                                 dev_priv->stats.requested_bufs++;
1923                                 buf->pending = 0;
1924                                 return buf;
1925                         }
1926                 }
1927                 start = 0;
1928         }
1929
1930         return NULL;
1931 }
1932 #endif
1933
1934 void radeon_freelist_reset(struct drm_device * dev)
1935 {
1936         struct drm_device_dma *dma = dev->dma;
1937         drm_radeon_private_t *dev_priv = dev->dev_private;
1938         int i;
1939
1940         dev_priv->last_buf = 0;
1941         for (i = 0; i < dma->buf_count; i++) {
1942                 struct drm_buf *buf = dma->buflist[i];
1943                 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
1944                 buf_priv->age = 0;
1945         }
1946 }
1947
1948 /* ================================================================
1949  * CP command submission
1950  */
1951
1952 int radeon_wait_ring(drm_radeon_private_t * dev_priv, int n)
1953 {
1954         drm_radeon_ring_buffer_t *ring = &dev_priv->ring;
1955         int i;
1956         u32 last_head = GET_RING_HEAD(dev_priv);
1957
1958         for (i = 0; i < dev_priv->usec_timeout; i++) {
1959                 u32 head = GET_RING_HEAD(dev_priv);
1960
1961                 ring->space = (head - ring->tail) * sizeof(u32);
1962                 if (ring->space <= 0)
1963                         ring->space += ring->size;
1964                 if (ring->space > n)
1965                         return 0;
1966
1967                 dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
1968
1969                 if (head != last_head)
1970                         i = 0;
1971                 last_head = head;
1972
1973                 DRM_UDELAY(1);
1974         }
1975
1976         /* FIXME: This return value is ignored in the BEGIN_RING macro! */
1977 #if RADEON_FIFO_DEBUG
1978         radeon_status(dev_priv);
1979         DRM_ERROR("failed!\n");
1980 #endif
1981         return -EBUSY;
1982 }
1983
1984 static int radeon_cp_get_buffers(struct drm_device *dev,
1985                                  struct drm_file *file_priv,
1986                                  struct drm_dma * d)
1987 {
1988         int i;
1989         struct drm_buf *buf;
1990
1991         for (i = d->granted_count; i < d->request_count; i++) {
1992                 buf = radeon_freelist_get(dev);
1993                 if (!buf)
1994                         return -EBUSY;  /* NOTE: broken client */
1995
1996                 buf->file_priv = file_priv;
1997
1998                 if (DRM_COPY_TO_USER(&d->request_indices[i], &buf->idx,
1999                                      sizeof(buf->idx)))
2000                         return -EFAULT;
2001                 if (DRM_COPY_TO_USER(&d->request_sizes[i], &buf->total,
2002                                      sizeof(buf->total)))
2003                         return -EFAULT;
2004
2005                 d->granted_count++;
2006         }
2007         return 0;
2008 }
2009
2010 int radeon_cp_buffers(struct drm_device *dev, void *data, struct drm_file *file_priv)
2011 {
2012         struct drm_device_dma *dma = dev->dma;
2013         int ret = 0;
2014         struct drm_dma *d = data;
2015
2016         LOCK_TEST_WITH_RETURN(dev, file_priv);
2017
2018         /* Please don't send us buffers.
2019          */
2020         if (d->send_count != 0) {
2021                 DRM_ERROR("Process %d trying to send %d buffers via drmDMA\n",
2022                           DRM_CURRENTPID, d->send_count);
2023                 return -EINVAL;
2024         }
2025
2026         /* We'll send you buffers.
2027          */
2028         if (d->request_count < 0 || d->request_count > dma->buf_count) {
2029                 DRM_ERROR("Process %d trying to get %d buffers (of %d max)\n",
2030                           DRM_CURRENTPID, d->request_count, dma->buf_count);
2031                 return -EINVAL;
2032         }
2033
2034         d->granted_count = 0;
2035
2036         if (d->request_count) {
2037                 ret = radeon_cp_get_buffers(dev, file_priv, d);
2038         }
2039
2040         return ret;
2041 }
2042
2043 int radeon_driver_load(struct drm_device *dev, unsigned long flags)
2044 {
2045         drm_radeon_private_t *dev_priv;
2046         int ret = 0;
2047
2048         dev_priv = drm_alloc(sizeof(drm_radeon_private_t), DRM_MEM_DRIVER);
2049         if (dev_priv == NULL)
2050                 return -ENOMEM;
2051
2052         memset(dev_priv, 0, sizeof(drm_radeon_private_t));
2053         dev->dev_private = (void *)dev_priv;
2054         dev_priv->flags = flags;
2055
2056         switch (flags & RADEON_FAMILY_MASK) {
2057         case CHIP_R100:
2058         case CHIP_RV200:
2059         case CHIP_R200:
2060         case CHIP_R300:
2061         case CHIP_R350:
2062         case CHIP_R420:
2063         case CHIP_R423:
2064         case CHIP_RV410:
2065         case CHIP_RV515:
2066         case CHIP_R520:
2067         case CHIP_RV570:
2068         case CHIP_R580:
2069                 dev_priv->flags |= RADEON_HAS_HIERZ;
2070                 break;
2071         default:
2072                 /* all other chips have no hierarchical z buffer */
2073                 break;
2074         }
2075
2076         if (drm_device_is_agp(dev))
2077                 dev_priv->flags |= RADEON_IS_AGP;
2078         else if (drm_device_is_pcie(dev))
2079                 dev_priv->flags |= RADEON_IS_PCIE;
2080         else
2081                 dev_priv->flags |= RADEON_IS_PCI;
2082
2083         ret = drm_addmap(dev, drm_get_resource_start(dev, 2),
2084                          drm_get_resource_len(dev, 2), _DRM_REGISTERS,
2085                          _DRM_READ_ONLY | _DRM_DRIVER, &dev_priv->mmio);
2086         if (ret != 0)
2087                 return ret;
2088
2089         ret = drm_vblank_init(dev, 2);
2090         if (ret) {
2091                 radeon_driver_unload(dev);
2092                 return ret;
2093         }
2094
2095         DRM_DEBUG("%s card detected\n",
2096                   ((dev_priv->flags & RADEON_IS_AGP) ? "AGP" : (((dev_priv->flags & RADEON_IS_PCIE) ? "PCIE" : "PCI"))));
2097         return ret;
2098 }
2099
2100 int radeon_master_create(struct drm_device *dev, struct drm_master *master)
2101 {
2102         struct drm_radeon_master_private *master_priv;
2103         unsigned long sareapage;
2104         int ret;
2105
2106         master_priv = drm_calloc(1, sizeof(*master_priv), DRM_MEM_DRIVER);
2107         if (!master_priv)
2108                 return -ENOMEM;
2109
2110         /* prebuild the SAREA */
2111         sareapage = max_t(unsigned long, SAREA_MAX, PAGE_SIZE);
2112         ret = drm_addmap(dev, 0, sareapage, _DRM_SHM, _DRM_CONTAINS_LOCK|_DRM_DRIVER,
2113                          &master_priv->sarea);
2114         if (ret) {
2115                 DRM_ERROR("SAREA setup failed\n");
2116                 return ret;
2117         }
2118         master_priv->sarea_priv = master_priv->sarea->handle + sizeof(struct drm_sarea);
2119         master_priv->sarea_priv->pfCurrentPage = 0;
2120
2121         master->driver_priv = master_priv;
2122         return 0;
2123 }
2124
2125 void radeon_master_destroy(struct drm_device *dev, struct drm_master *master)
2126 {
2127         struct drm_radeon_master_private *master_priv = master->driver_priv;
2128
2129         if (!master_priv)
2130                 return;
2131
2132         if (master_priv->sarea_priv &&
2133             master_priv->sarea_priv->pfCurrentPage != 0)
2134                 radeon_cp_dispatch_flip(dev, master);
2135
2136         master_priv->sarea_priv = NULL;
2137         if (master_priv->sarea)
2138                 drm_rmmap_locked(dev, master_priv->sarea);
2139
2140         drm_free(master_priv, sizeof(*master_priv), DRM_MEM_DRIVER);
2141
2142         master->driver_priv = NULL;
2143 }
2144
2145 /* Create mappings for registers and framebuffer so userland doesn't necessarily
2146  * have to find them.
2147  */
2148 int radeon_driver_firstopen(struct drm_device *dev)
2149 {
2150         int ret;
2151         drm_local_map_t *map;
2152         drm_radeon_private_t *dev_priv = dev->dev_private;
2153
2154         dev_priv->gart_info.table_size = RADEON_PCIGART_TABLE_SIZE;
2155
2156         dev_priv->fb_aper_offset = drm_get_resource_start(dev, 0);
2157         ret = drm_addmap(dev, dev_priv->fb_aper_offset,
2158                          drm_get_resource_len(dev, 0), _DRM_FRAME_BUFFER,
2159                          _DRM_WRITE_COMBINING, &map);
2160         if (ret != 0)
2161                 return ret;
2162
2163         return 0;
2164 }
2165
2166 int radeon_driver_unload(struct drm_device *dev)
2167 {
2168         drm_radeon_private_t *dev_priv = dev->dev_private;
2169
2170         DRM_DEBUG("\n");
2171
2172         drm_rmmap(dev, dev_priv->mmio);
2173
2174         drm_free(dev_priv, sizeof(*dev_priv), DRM_MEM_DRIVER);
2175
2176         dev->dev_private = NULL;
2177         return 0;
2178 }
2179
2180 void radeon_commit_ring(drm_radeon_private_t *dev_priv)
2181 {
2182         int i;
2183         u32 *ring;
2184         int tail_aligned;
2185
2186         /* check if the ring is padded out to 16-dword alignment */
2187
2188         tail_aligned = dev_priv->ring.tail & 0xf;
2189         if (tail_aligned) {
2190                 int num_p2 = 16 - tail_aligned;
2191
2192                 ring = dev_priv->ring.start;
2193                 /* pad with some CP_PACKET2 */
2194                 for (i = 0; i < num_p2; i++)
2195                         ring[dev_priv->ring.tail + i] = CP_PACKET2();
2196
2197                 dev_priv->ring.tail += i;
2198
2199                 dev_priv->ring.space -= num_p2 * sizeof(u32);
2200         }
2201
2202         dev_priv->ring.tail &= dev_priv->ring.tail_mask;
2203
2204         DRM_MEMORYBARRIER();
2205         GET_RING_HEAD( dev_priv );
2206
2207         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R600) {
2208                 RADEON_WRITE(R600_CP_RB_WPTR, dev_priv->ring.tail);
2209                 /* read from PCI bus to ensure correct posting */
2210                 RADEON_READ(R600_CP_RB_RPTR);
2211         } else {
2212                 RADEON_WRITE(RADEON_CP_RB_WPTR, dev_priv->ring.tail);
2213                 /* read from PCI bus to ensure correct posting */
2214                 RADEON_READ(RADEON_CP_RB_RPTR);
2215         }
2216 }