drivers/gpu/drm/radeon/radeon_cp.c

   1 /* radeon_cp.c -- CP support for Radeon -*- linux-c -*- */
   2 /*
   3  * Copyright 2000 Precision Insight, Inc., Cedar Park, Texas.
   4  * Copyright 2000 VA Linux Systems, Inc., Fremont, California.
   5  * Copyright 2007 Advanced Micro Devices, Inc.
   6  * All Rights Reserved.
   7  *
   8  * Permission is hereby granted, free of charge, to any person obtaining a
   9  * copy of this software and associated documentation files (the "Software"),
  10  * to deal in the Software without restriction, including without limitation
  11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  12  * and/or sell copies of the Software, and to permit persons to whom the
  13  * Software is furnished to do so, subject to the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the next
  16  * paragraph) shall be included in all copies or substantial portions of the
  17  * Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  23  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  24  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  25  * DEALINGS IN THE SOFTWARE.
  26  *
  27  * Authors:
  28  *    Kevin E. Martin <martin@valinux.com>
  29  *    Gareth Hughes <gareth@valinux.com>
  30  */
  31
  32 #include "drmP.h"
  33 #include "drm.h"
  34 #include "drm_sarea.h"
  35 #include "radeon_drm.h"
  36 #include "radeon_drv.h"
  37 #include "r300_reg.h"
  38
  39 #include "radeon_microcode.h"
  40
  41 #define RADEON_FIFO_DEBUG       0
  42
  43 static int radeon_do_cleanup_cp(struct drm_device * dev);
  44 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv);
  45
  46 static u32 R500_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  47 {
  48         u32 ret;
  49         RADEON_WRITE(R520_MC_IND_INDEX, 0x7f0000 | (addr & 0xff));
  50         ret = RADEON_READ(R520_MC_IND_DATA);
  51         RADEON_WRITE(R520_MC_IND_INDEX, 0);
  52         return ret;
  53 }
  54
  55 static u32 RS480_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  56 {
  57         u32 ret;
  58         RADEON_WRITE(RS480_NB_MC_INDEX, addr & 0xff);
  59         ret = RADEON_READ(RS480_NB_MC_DATA);
  60         RADEON_WRITE(RS480_NB_MC_INDEX, 0xff);
  61         return ret;
  62 }
  63
  64 static u32 RS690_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  65 {
  66         u32 ret;
  67         RADEON_WRITE(RS690_MC_INDEX, (addr & RS690_MC_INDEX_MASK));
  68         ret = RADEON_READ(RS690_MC_DATA);
  69         RADEON_WRITE(RS690_MC_INDEX, RS690_MC_INDEX_MASK);
  70         return ret;
  71 }
  72
  73 static u32 IGP_READ_MCIND(drm_radeon_private_t *dev_priv, int addr)
  74 {
  75         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
  76             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
  77                 return RS690_READ_MCIND(dev_priv, addr);
  78         else
  79                 return RS480_READ_MCIND(dev_priv, addr);
  80 }
  81
  82 u32 radeon_read_fb_location(drm_radeon_private_t *dev_priv)
  83 {
  84
  85         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  86                 return R500_READ_MCIND(dev_priv, RV515_MC_FB_LOCATION);
  87         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
  88                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
  89                 return RS690_READ_MCIND(dev_priv, RS690_MC_FB_LOCATION);
  90         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
  91                 return R500_READ_MCIND(dev_priv, R520_MC_FB_LOCATION);
  92         else
  93                 return RADEON_READ(RADEON_MC_FB_LOCATION);
  94 }
  95
  96 static void radeon_write_fb_location(drm_radeon_private_t *dev_priv, u32 fb_loc)
  97 {
  98         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
  99                 R500_WRITE_MCIND(RV515_MC_FB_LOCATION, fb_loc);
 100         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 101                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 102                 RS690_WRITE_MCIND(RS690_MC_FB_LOCATION, fb_loc);
 103         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 104                 R500_WRITE_MCIND(R520_MC_FB_LOCATION, fb_loc);
 105         else
 106                 RADEON_WRITE(RADEON_MC_FB_LOCATION, fb_loc);
 107 }
 108
 109 static void radeon_write_agp_location(drm_radeon_private_t *dev_priv, u32 agp_loc)
 110 {
 111         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515)
 112                 R500_WRITE_MCIND(RV515_MC_AGP_LOCATION, agp_loc);
 113         else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 114                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 115                 RS690_WRITE_MCIND(RS690_MC_AGP_LOCATION, agp_loc);
 116         else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515)
 117                 R500_WRITE_MCIND(R520_MC_AGP_LOCATION, agp_loc);
 118         else
 119                 RADEON_WRITE(RADEON_MC_AGP_LOCATION, agp_loc);
 120 }
 121
 122 static void radeon_write_agp_base(drm_radeon_private_t *dev_priv, u64 agp_base)
 123 {
 124         u32 agp_base_hi = upper_32_bits(agp_base);
 125         u32 agp_base_lo = agp_base & 0xffffffff;
 126
 127         if ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) {
 128                 R500_WRITE_MCIND(RV515_MC_AGP_BASE, agp_base_lo);
 129                 R500_WRITE_MCIND(RV515_MC_AGP_BASE_2, agp_base_hi);
 130         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 131                  ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 132                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE, agp_base_lo);
 133                 RS690_WRITE_MCIND(RS690_MC_AGP_BASE_2, agp_base_hi);
 134         } else if ((dev_priv->flags & RADEON_FAMILY_MASK) > CHIP_RV515) {
 135                 R500_WRITE_MCIND(R520_MC_AGP_BASE, agp_base_lo);
 136                 R500_WRITE_MCIND(R520_MC_AGP_BASE_2, agp_base_hi);
 137         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 138                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 139                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 140                 RADEON_WRITE(RS480_AGP_BASE_2, agp_base_hi);
 141         } else {
 142                 RADEON_WRITE(RADEON_AGP_BASE, agp_base_lo);
 143                 if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R200)
 144                         RADEON_WRITE(RADEON_AGP_BASE_2, agp_base_hi);
 145         }
 146 }
 147
 148 static int RADEON_READ_PLL(struct drm_device * dev, int addr)
 149 {
 150         drm_radeon_private_t *dev_priv = dev->dev_private;
 151
 152         RADEON_WRITE8(RADEON_CLOCK_CNTL_INDEX, addr & 0x1f);
 153         return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 154 }
 155
 156 static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 157 {
 158         RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 159         return RADEON_READ(RADEON_PCIE_DATA);
 160 }
 161
 162 #if RADEON_FIFO_DEBUG
 163 static void radeon_status(drm_radeon_private_t * dev_priv)
 164 {
 165         printk("%s:\n", __func__);
 166         printk("RBBM_STATUS = 0x%08x\n",
 167                (unsigned int)RADEON_READ(RADEON_RBBM_STATUS));
 168         printk("CP_RB_RTPR = 0x%08x\n",
 169                (unsigned int)RADEON_READ(RADEON_CP_RB_RPTR));
 170         printk("CP_RB_WTPR = 0x%08x\n",
 171                (unsigned int)RADEON_READ(RADEON_CP_RB_WPTR));
 172         printk("AIC_CNTL = 0x%08x\n",
 173                (unsigned int)RADEON_READ(RADEON_AIC_CNTL));
 174         printk("AIC_STAT = 0x%08x\n",
 175                (unsigned int)RADEON_READ(RADEON_AIC_STAT));
 176         printk("AIC_PT_BASE = 0x%08x\n",
 177                (unsigned int)RADEON_READ(RADEON_AIC_PT_BASE));
 178         printk("TLB_ADDR = 0x%08x\n",
 179                (unsigned int)RADEON_READ(RADEON_AIC_TLB_ADDR));
 180         printk("TLB_DATA = 0x%08x\n",
 181                (unsigned int)RADEON_READ(RADEON_AIC_TLB_DATA));
 182 }
 183 #endif
 184
 185 /* ================================================================
 186  * Engine, FIFO control
 187  */
 188
 189 static int radeon_do_pixcache_flush(drm_radeon_private_t * dev_priv)
 190 {
 191         u32 tmp;
 192         int i;
 193
 194         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 195
 196         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV280) {
 197                 tmp = RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT);
 198                 tmp |= RADEON_RB3D_DC_FLUSH_ALL;
 199                 RADEON_WRITE(RADEON_RB3D_DSTCACHE_CTLSTAT, tmp);
 200
 201                 for (i = 0; i < dev_priv->usec_timeout; i++) {
 202                         if (!(RADEON_READ(RADEON_RB3D_DSTCACHE_CTLSTAT)
 203                               & RADEON_RB3D_DC_BUSY)) {
 204                                 return 0;
 205                         }
 206                         DRM_UDELAY(1);
 207                 }
 208         } else {
 209                 /* don't flush or purge cache here or lockup */
 210                 return 0;
 211         }
 212
 213 #if RADEON_FIFO_DEBUG
 214         DRM_ERROR("failed!\n");
 215         radeon_status(dev_priv);
 216 #endif
 217         return -EBUSY;
 218 }
 219
 220 static int radeon_do_wait_for_fifo(drm_radeon_private_t * dev_priv, int entries)
 221 {
 222         int i;
 223
 224         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 225
 226         for (i = 0; i < dev_priv->usec_timeout; i++) {
 227                 int slots = (RADEON_READ(RADEON_RBBM_STATUS)
 228                              & RADEON_RBBM_FIFOCNT_MASK);
 229                 if (slots >= entries)
 230                         return 0;
 231                 DRM_UDELAY(1);
 232         }
 233         DRM_DEBUG("wait for fifo failed status : 0x%08X 0x%08X\n",
 234                  RADEON_READ(RADEON_RBBM_STATUS),
 235                  RADEON_READ(R300_VAP_CNTL_STATUS));
 236
 237 #if RADEON_FIFO_DEBUG
 238         DRM_ERROR("failed!\n");
 239         radeon_status(dev_priv);
 240 #endif
 241         return -EBUSY;
 242 }
 243
 244 static int radeon_do_wait_for_idle(drm_radeon_private_t * dev_priv)
 245 {
 246         int i, ret;
 247
 248         dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
 249
 250         ret = radeon_do_wait_for_fifo(dev_priv, 64);
 251         if (ret)
 252                 return ret;
 253
 254         for (i = 0; i < dev_priv->usec_timeout; i++) {
 255                 if (!(RADEON_READ(RADEON_RBBM_STATUS)
 256                       & RADEON_RBBM_ACTIVE)) {
 257                         radeon_do_pixcache_flush(dev_priv);
 258                         return 0;
 259                 }
 260                 DRM_UDELAY(1);
 261         }
 262         DRM_DEBUG("wait idle failed status : 0x%08X 0x%08X\n",
 263                  RADEON_READ(RADEON_RBBM_STATUS),
 264                  RADEON_READ(R300_VAP_CNTL_STATUS));
 265
 266 #if RADEON_FIFO_DEBUG
 267         DRM_ERROR("failed!\n");
 268         radeon_status(dev_priv);
 269 #endif
 270         return -EBUSY;
 271 }
 272
 273 static void radeon_init_pipes(drm_radeon_private_t *dev_priv)
 274 {
 275         uint32_t gb_tile_config, gb_pipe_sel = 0;
 276
 277         /* RS4xx/RS6xx/R4xx/R5xx */
 278         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R420) {
 279                 gb_pipe_sel = RADEON_READ(R400_GB_PIPE_SELECT);
 280                 dev_priv->num_gb_pipes = ((gb_pipe_sel >> 12) & 0x3) + 1;
 281         } else {
 282                 /* R3xx */
 283                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 284                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350)) {
 285                         dev_priv->num_gb_pipes = 2;
 286                 } else {
 287                         /* R3Vxx */
 288                         dev_priv->num_gb_pipes = 1;
 289                 }
 290         }
 291         DRM_INFO("Num pipes: %d\n", dev_priv->num_gb_pipes);
 292
 293         gb_tile_config = (R300_ENABLE_TILING | R300_TILE_SIZE_16 /*| R300_SUBPIXEL_1_16*/);
 294
 295         switch (dev_priv->num_gb_pipes) {
 296         case 2: gb_tile_config |= R300_PIPE_COUNT_R300; break;
 297         case 3: gb_tile_config |= R300_PIPE_COUNT_R420_3P; break;
 298         case 4: gb_tile_config |= R300_PIPE_COUNT_R420; break;
 299         default:
 300         case 1: gb_tile_config |= R300_PIPE_COUNT_RV350; break;
 301         }
 302
 303         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_RV515) {
 304                 RADEON_WRITE_PLL(R500_DYN_SCLK_PWMEM_PIPE, (1 | ((gb_pipe_sel >> 8) & 0xf) << 4));
 305                 RADEON_WRITE(R500_SU_REG_DEST, ((1 << dev_priv->num_gb_pipes) - 1));
 306         }
 307         RADEON_WRITE(R300_GB_TILE_CONFIG, gb_tile_config);
 308         radeon_do_wait_for_idle(dev_priv);
 309         RADEON_WRITE(R300_DST_PIPE_CONFIG, RADEON_READ(R300_DST_PIPE_CONFIG) | R300_PIPE_AUTO_CONFIG);
 310         RADEON_WRITE(R300_RB2D_DSTCACHE_MODE, (RADEON_READ(R300_RB2D_DSTCACHE_MODE) |
 311                                                R300_DC_AUTOFLUSH_ENABLE |
 312                                                R300_DC_DC_DISABLE_IGNORE_PE));
 313
 314
 315 }
 316
 317 /* ================================================================
 318  * CP control, initialization
 319  */
 320
 321 /* Load the microcode for the CP */
 322 static void radeon_cp_load_microcode(drm_radeon_private_t * dev_priv)
 323 {
 324         int i;
 325         DRM_DEBUG("\n");
 326
 327         radeon_do_wait_for_idle(dev_priv);
 328
 329         RADEON_WRITE(RADEON_CP_ME_RAM_ADDR, 0);
 330         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R100) ||
 331             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV100) ||
 332             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV200) ||
 333             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS100) ||
 334             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS200)) {
 335                 DRM_INFO("Loading R100 Microcode\n");
 336                 for (i = 0; i < 256; i++) {
 337                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 338                                      R100_cp_microcode[i][1]);
 339                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 340                                      R100_cp_microcode[i][0]);
 341                 }
 342         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R200) ||
 343                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV250) ||
 344                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV280) ||
 345                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS300)) {
 346                 DRM_INFO("Loading R200 Microcode\n");
 347                 for (i = 0; i < 256; i++) {
 348                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 349                                      R200_cp_microcode[i][1]);
 350                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 351                                      R200_cp_microcode[i][0]);
 352                 }
 353         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R300) ||
 354                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R350) ||
 355                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV350) ||
 356                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV380) ||
 357                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 358                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 359                 DRM_INFO("Loading R300 Microcode\n");
 360                 for (i = 0; i < 256; i++) {
 361                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 362                                      R300_cp_microcode[i][1]);
 363                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 364                                      R300_cp_microcode[i][0]);
 365                 }
 366         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 367                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R423) ||
 368                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV410)) {
 369                 DRM_INFO("Loading R400 Microcode\n");
 370                 for (i = 0; i < 256; i++) {
 371                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 372                                      R420_cp_microcode[i][1]);
 373                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 374                                      R420_cp_microcode[i][0]);
 375                 }
 376         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 377                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 378                 DRM_INFO("Loading RS690/RS740 Microcode\n");
 379                 for (i = 0; i < 256; i++) {
 380                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 381                                      RS690_cp_microcode[i][1]);
 382                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 383                                      RS690_cp_microcode[i][0]);
 384                 }
 385         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV515) ||
 386                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R520) ||
 387                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV530) ||
 388                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R580) ||
 389                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV560) ||
 390                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RV570)) {
 391                 DRM_INFO("Loading R500 Microcode\n");
 392                 for (i = 0; i < 256; i++) {
 393                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAH,
 394                                      R520_cp_microcode[i][1]);
 395                         RADEON_WRITE(RADEON_CP_ME_RAM_DATAL,
 396                                      R520_cp_microcode[i][0]);
 397                 }
 398         }
 399 }
 400
 401 /* Flush any pending commands to the CP.  This should only be used just
 402  * prior to a wait for idle, as it informs the engine that the command
 403  * stream is ending.
 404  */
 405 static void radeon_do_cp_flush(drm_radeon_private_t * dev_priv)
 406 {
 407         DRM_DEBUG("\n");
 408 #if 0
 409         u32 tmp;
 410
 411         tmp = RADEON_READ(RADEON_CP_RB_WPTR) | (1 << 31);
 412         RADEON_WRITE(RADEON_CP_RB_WPTR, tmp);
 413 #endif
 414 }
 415
 416 /* Wait for the CP to go idle.
 417  */
 418 int radeon_do_cp_idle(drm_radeon_private_t * dev_priv)
 419 {
 420         RING_LOCALS;
 421         DRM_DEBUG("\n");
 422
 423         BEGIN_RING(6);
 424
 425         RADEON_PURGE_CACHE();
 426         RADEON_PURGE_ZCACHE();
 427         RADEON_WAIT_UNTIL_IDLE();
 428
 429         ADVANCE_RING();
 430         COMMIT_RING();
 431
 432         return radeon_do_wait_for_idle(dev_priv);
 433 }
 434
 435 /* Start the Command Processor.
 436  */
 437 static void radeon_do_cp_start(drm_radeon_private_t * dev_priv)
 438 {
 439         RING_LOCALS;
 440         DRM_DEBUG("\n");
 441
 442         radeon_do_wait_for_idle(dev_priv);
 443
 444         RADEON_WRITE(RADEON_CP_CSQ_CNTL, dev_priv->cp_mode);
 445
 446         dev_priv->cp_running = 1;
 447
 448         BEGIN_RING(8);
 449         /* isync can only be written through cp on r5xx write it here */
 450         OUT_RING(CP_PACKET0(RADEON_ISYNC_CNTL, 0));
 451         OUT_RING(RADEON_ISYNC_ANY2D_IDLE3D |
 452                  RADEON_ISYNC_ANY3D_IDLE2D |
 453                  RADEON_ISYNC_WAIT_IDLEGUI |
 454                  RADEON_ISYNC_CPSCRATCH_IDLEGUI);
 455         RADEON_PURGE_CACHE();
 456         RADEON_PURGE_ZCACHE();
 457         RADEON_WAIT_UNTIL_IDLE();
 458         ADVANCE_RING();
 459         COMMIT_RING();
 460
 461         dev_priv->track_flush |= RADEON_FLUSH_EMITED | RADEON_PURGE_EMITED;
 462 }
 463
 464 /* Reset the Command Processor.  This will not flush any pending
 465  * commands, so you must wait for the CP command stream to complete
 466  * before calling this routine.
 467  */
 468 static void radeon_do_cp_reset(drm_radeon_private_t * dev_priv)
 469 {
 470         u32 cur_read_ptr;
 471         DRM_DEBUG("\n");
 472
 473         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 474         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 475         SET_RING_HEAD(dev_priv, cur_read_ptr);
 476         dev_priv->ring.tail = cur_read_ptr;
 477 }
 478
 479 /* Stop the Command Processor.  This will not flush any pending
 480  * commands, so you must flush the command stream and wait for the CP
 481  * to go idle before calling this routine.
 482  */
 483 static void radeon_do_cp_stop(drm_radeon_private_t * dev_priv)
 484 {
 485         DRM_DEBUG("\n");
 486
 487         RADEON_WRITE(RADEON_CP_CSQ_CNTL, RADEON_CSQ_PRIDIS_INDDIS);
 488
 489         dev_priv->cp_running = 0;
 490 }
 491
 492 /* Reset the engine.  This will stop the CP if it is running.
 493  */
 494 static int radeon_do_engine_reset(struct drm_device * dev)
 495 {
 496         drm_radeon_private_t *dev_priv = dev->dev_private;
 497         u32 clock_cntl_index = 0, mclk_cntl = 0, rbbm_soft_reset;
 498         DRM_DEBUG("\n");
 499
 500         radeon_do_pixcache_flush(dev_priv);
 501
 502         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 503                 /* may need something similar for newer chips */
 504                 clock_cntl_index = RADEON_READ(RADEON_CLOCK_CNTL_INDEX);
 505                 mclk_cntl = RADEON_READ_PLL(dev, RADEON_MCLK_CNTL);
 506
 507                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, (mclk_cntl |
 508                                                     RADEON_FORCEON_MCLKA |
 509                                                     RADEON_FORCEON_MCLKB |
 510                                                     RADEON_FORCEON_YCLKA |
 511                                                     RADEON_FORCEON_YCLKB |
 512                                                     RADEON_FORCEON_MC |
 513                                                     RADEON_FORCEON_AIC));
 514         }
 515
 516         rbbm_soft_reset = RADEON_READ(RADEON_RBBM_SOFT_RESET);
 517
 518         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset |
 519                                               RADEON_SOFT_RESET_CP |
 520                                               RADEON_SOFT_RESET_HI |
 521                                               RADEON_SOFT_RESET_SE |
 522                                               RADEON_SOFT_RESET_RE |
 523                                               RADEON_SOFT_RESET_PP |
 524                                               RADEON_SOFT_RESET_E2 |
 525                                               RADEON_SOFT_RESET_RB));
 526         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 527         RADEON_WRITE(RADEON_RBBM_SOFT_RESET, (rbbm_soft_reset &
 528                                               ~(RADEON_SOFT_RESET_CP |
 529                                                 RADEON_SOFT_RESET_HI |
 530                                                 RADEON_SOFT_RESET_SE |
 531                                                 RADEON_SOFT_RESET_RE |
 532                                                 RADEON_SOFT_RESET_PP |
 533                                                 RADEON_SOFT_RESET_E2 |
 534                                                 RADEON_SOFT_RESET_RB)));
 535         RADEON_READ(RADEON_RBBM_SOFT_RESET);
 536
 537         if ((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV410) {
 538                 RADEON_WRITE_PLL(RADEON_MCLK_CNTL, mclk_cntl);
 539                 RADEON_WRITE(RADEON_CLOCK_CNTL_INDEX, clock_cntl_index);
 540                 RADEON_WRITE(RADEON_RBBM_SOFT_RESET, rbbm_soft_reset);
 541         }
 542
 543         /* setup the raster pipes */
 544         if ((dev_priv->flags & RADEON_FAMILY_MASK) >= CHIP_R300)
 545             radeon_init_pipes(dev_priv);
 546
 547         /* Reset the CP ring */
 548         radeon_do_cp_reset(dev_priv);
 549
 550         /* The CP is no longer running after an engine reset */
 551         dev_priv->cp_running = 0;
 552
 553         /* Reset any pending vertex, indirect buffers */
 554         radeon_freelist_reset(dev);
 555
 556         return 0;
 557 }
 558
 559 static void radeon_cp_init_ring_buffer(struct drm_device * dev,
 560                                        drm_radeon_private_t * dev_priv)
 561 {
 562         u32 ring_start, cur_read_ptr;
 563         u32 tmp;
 564
 565         /* Initialize the memory controller. With new memory map, the fb location
 566          * is not changed, it should have been properly initialized already. Part
 567          * of the problem is that the code below is bogus, assuming the GART is
 568          * always appended to the fb which is not necessarily the case
 569          */
 570         if (!dev_priv->new_memmap)
 571                 radeon_write_fb_location(dev_priv,
 572                              ((dev_priv->gart_vm_start - 1) & 0xffff0000)
 573                              | (dev_priv->fb_location >> 16));
 574
 575 #if __OS_HAS_AGP
 576         if (dev_priv->flags & RADEON_IS_AGP) {
 577                 radeon_write_agp_base(dev_priv, dev->agp->base);
 578
 579                 radeon_write_agp_location(dev_priv,
 580                              (((dev_priv->gart_vm_start - 1 +
 581                                 dev_priv->gart_size) & 0xffff0000) |
 582                               (dev_priv->gart_vm_start >> 16)));
 583
 584                 ring_start = (dev_priv->cp_ring->offset
 585                               - dev->agp->base
 586                               + dev_priv->gart_vm_start);
 587         } else
 588 #endif
 589                 ring_start = (dev_priv->cp_ring->offset
 590                               - (unsigned long)dev->sg->virtual
 591                               + dev_priv->gart_vm_start);
 592
 593         RADEON_WRITE(RADEON_CP_RB_BASE, ring_start);
 594
 595         /* Set the write pointer delay */
 596         RADEON_WRITE(RADEON_CP_RB_WPTR_DELAY, 0);
 597
 598         /* Initialize the ring buffer's read and write pointers */
 599         cur_read_ptr = RADEON_READ(RADEON_CP_RB_RPTR);
 600         RADEON_WRITE(RADEON_CP_RB_WPTR, cur_read_ptr);
 601         SET_RING_HEAD(dev_priv, cur_read_ptr);
 602         dev_priv->ring.tail = cur_read_ptr;
 603
 604 #if __OS_HAS_AGP
 605         if (dev_priv->flags & RADEON_IS_AGP) {
 606                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR,
 607                              dev_priv->ring_rptr->offset
 608                              - dev->agp->base + dev_priv->gart_vm_start);
 609         } else
 610 #endif
 611         {
 612                 struct drm_sg_mem *entry = dev->sg;
 613                 unsigned long tmp_ofs, page_ofs;
 614
 615                 tmp_ofs = dev_priv->ring_rptr->offset -
 616                                 (unsigned long)dev->sg->virtual;
 617                 page_ofs = tmp_ofs >> PAGE_SHIFT;
 618
 619                 RADEON_WRITE(RADEON_CP_RB_RPTR_ADDR, entry->busaddr[page_ofs]);
 620                 DRM_DEBUG("ring rptr: offset=0x%08lx handle=0x%08lx\n",
 621                           (unsigned long)entry->busaddr[page_ofs],
 622                           entry->handle + tmp_ofs);
 623         }
 624
 625         /* Set ring buffer size */
 626 #ifdef __BIG_ENDIAN
 627         RADEON_WRITE(RADEON_CP_RB_CNTL,
 628                      RADEON_BUF_SWAP_32BIT |
 629                      (dev_priv->ring.fetch_size_l2ow << 18) |
 630                      (dev_priv->ring.rptr_update_l2qw << 8) |
 631                      dev_priv->ring.size_l2qw);
 632 #else
 633         RADEON_WRITE(RADEON_CP_RB_CNTL,
 634                      (dev_priv->ring.fetch_size_l2ow << 18) |
 635                      (dev_priv->ring.rptr_update_l2qw << 8) |
 636                      dev_priv->ring.size_l2qw);
 637 #endif
 638
 639
 640         /* Initialize the scratch register pointer.  This will cause
 641          * the scratch register values to be written out to memory
 642          * whenever they are updated.
 643          *
 644          * We simply put this behind the ring read pointer, this works
 645          * with PCI GART as well as (whatever kind of) AGP GART
 646          */
 647         RADEON_WRITE(RADEON_SCRATCH_ADDR, RADEON_READ(RADEON_CP_RB_RPTR_ADDR)
 648                      + RADEON_SCRATCH_REG_OFFSET);
 649
 650         dev_priv->scratch = ((__volatile__ u32 *)
 651                              dev_priv->ring_rptr->handle +
 652                              (RADEON_SCRATCH_REG_OFFSET / sizeof(u32)));
 653
 654         RADEON_WRITE(RADEON_SCRATCH_UMSK, 0x7);
 655
 656         /* Turn on bus mastering */
 657         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 658             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740)) {
 659                 /* rs600/rs690/rs740 */
 660                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RS600_BUS_MASTER_DIS;
 661                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 662         } else if (((dev_priv->flags & RADEON_FAMILY_MASK) <= CHIP_RV350) ||
 663                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_R420) ||
 664                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS400) ||
 665                    ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS480)) {
 666                 /* r1xx, r2xx, r300, r(v)350, r420/r481, rs400/rs480 */
 667                 tmp = RADEON_READ(RADEON_BUS_CNTL) & ~RADEON_BUS_MASTER_DIS;
 668                 RADEON_WRITE(RADEON_BUS_CNTL, tmp);
 669         } /* PCIE cards appears to not need this */
 670
 671         dev_priv->scratch[0] = 0;
 672         RADEON_WRITE(RADEON_LAST_FRAME_REG, 0);
 673
 674         dev_priv->scratch[1] = 0;
 675         RADEON_WRITE(RADEON_LAST_DISPATCH_REG, 0);
 676
 677         dev_priv->scratch[2] = 0;
 678         RADEON_WRITE(RADEON_LAST_CLEAR_REG, 0);
 679
 680         radeon_do_wait_for_idle(dev_priv);
 681
 682         /* Sync everything up */
 683         RADEON_WRITE(RADEON_ISYNC_CNTL,
 684                      (RADEON_ISYNC_ANY2D_IDLE3D |
 685                       RADEON_ISYNC_ANY3D_IDLE2D |
 686                       RADEON_ISYNC_WAIT_IDLEGUI |
 687                       RADEON_ISYNC_CPSCRATCH_IDLEGUI));
 688
 689 }
 690
 691 static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 692 {
 693         u32 tmp;
 694
 695         /* Start with assuming that writeback doesn't work */
 696         dev_priv->writeback_works = 0;
 697
 698         /* Writeback doesn't seem to work everywhere, test it here and possibly
 699          * enable it if it appears to work
 700          */
 701         DRM_WRITE32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1), 0);
 702         RADEON_WRITE(RADEON_SCRATCH_REG1, 0xdeadbeef);
 703
 704         for (tmp = 0; tmp < dev_priv->usec_timeout; tmp++) {
 705                 if (DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1)) ==
 706                     0xdeadbeef)
 707                         break;
 708                 DRM_UDELAY(1);
 709         }
 710
 711         if (tmp < dev_priv->usec_timeout) {
 712                 dev_priv->writeback_works = 1;
 713                 DRM_INFO("writeback test succeeded in %d usecs\n", tmp);
 714         } else {
 715                 dev_priv->writeback_works = 0;
 716                 DRM_INFO("writeback test failed\n");
 717         }
 718         if (radeon_no_wb == 1) {
 719                 dev_priv->writeback_works = 0;
 720                 DRM_INFO("writeback forced off\n");
 721         }
 722
 723         if (!dev_priv->writeback_works) {
 724                 /* Disable writeback to avoid unnecessary bus master transfer */
 725                 RADEON_WRITE(RADEON_CP_RB_CNTL, RADEON_READ(RADEON_CP_RB_CNTL) |
 726                              RADEON_RB_NO_UPDATE);
 727                 RADEON_WRITE(RADEON_SCRATCH_UMSK, 0);
 728         }
 729 }
 730
 731 /* Enable or disable IGP GART on the chip */
 732 static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
 733 {
 734         u32 temp;
 735
 736         if (on) {
 737                 DRM_DEBUG("programming igp gart %08X %08lX %08X\n",
 738                           dev_priv->gart_vm_start,
 739                           (long)dev_priv->gart_info.bus_addr,
 740                           dev_priv->gart_size);
 741
 742                 temp = IGP_READ_MCIND(dev_priv, RS480_MC_MISC_CNTL);
 743                 if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 744                     ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740))
 745                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, (RS480_GART_INDEX_REG_EN |
 746                                                              RS690_BLOCK_GFX_D3_EN));
 747                 else
 748                         IGP_WRITE_MCIND(RS480_MC_MISC_CNTL, RS480_GART_INDEX_REG_EN);
 749
 750                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 751                                                                RS480_VA_SIZE_32MB));
 752
 753                 temp = IGP_READ_MCIND(dev_priv, RS480_GART_FEATURE_ID);
 754                 IGP_WRITE_MCIND(RS480_GART_FEATURE_ID, (RS480_HANG_EN |
 755                                                         RS480_TLB_ENABLE |
 756                                                         RS480_GTW_LAC_EN |
 757                                                         RS480_1LEVEL_GART));
 758
 759                 temp = dev_priv->gart_info.bus_addr & 0xfffff000;
 760                 temp |= (upper_32_bits(dev_priv->gart_info.bus_addr) & 0xff) << 4;
 761                 IGP_WRITE_MCIND(RS480_GART_BASE, temp);
 762
 763                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_MODE_CNTL);
 764                 IGP_WRITE_MCIND(RS480_AGP_MODE_CNTL, ((1 << RS480_REQ_TYPE_SNOOP_SHIFT) |
 765                                                       RS480_REQ_TYPE_SNOOP_DIS));
 766
 767                 radeon_write_agp_base(dev_priv, dev_priv->gart_vm_start);
 768
 769                 dev_priv->gart_size = 32*1024*1024;
 770                 temp = (((dev_priv->gart_vm_start - 1 + dev_priv->gart_size) &
 771                          0xffff0000) | (dev_priv->gart_vm_start >> 16));
 772
 773                 radeon_write_agp_location(dev_priv, temp);
 774
 775                 temp = IGP_READ_MCIND(dev_priv, RS480_AGP_ADDRESS_SPACE_SIZE);
 776                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, (RS480_GART_EN |
 777                                                                RS480_VA_SIZE_32MB));
 778
 779                 do {
 780                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 781                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 782                                 break;
 783                         DRM_UDELAY(1);
 784                 } while (1);
 785
 786                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL,
 787                                 RS480_GART_CACHE_INVALIDATE);
 788
 789                 do {
 790                         temp = IGP_READ_MCIND(dev_priv, RS480_GART_CACHE_CNTRL);
 791                         if ((temp & RS480_GART_CACHE_INVALIDATE) == 0)
 792                                 break;
 793                         DRM_UDELAY(1);
 794                 } while (1);
 795
 796                 IGP_WRITE_MCIND(RS480_GART_CACHE_CNTRL, 0);
 797         } else {
 798                 IGP_WRITE_MCIND(RS480_AGP_ADDRESS_SPACE_SIZE, 0);
 799         }
 800 }
 801
 802 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
 803 {
 804         u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
 805         if (on) {
 806
 807                 DRM_DEBUG("programming pcie %08X %08lX %08X\n",
 808                           dev_priv->gart_vm_start,
 809                           (long)dev_priv->gart_info.bus_addr,
 810                           dev_priv->gart_size);
 811                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_DISCARD_RD_ADDR_LO,
 812                                   dev_priv->gart_vm_start);
 813                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_BASE,
 814                                   dev_priv->gart_info.bus_addr);
 815                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_START_LO,
 816                                   dev_priv->gart_vm_start);
 817                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_END_LO,
 818                                   dev_priv->gart_vm_start +
 819                                   dev_priv->gart_size - 1);
 820
 821                 radeon_write_agp_location(dev_priv, 0xffffffc0); /* ?? */
 822
 823                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 824                                   RADEON_PCIE_TX_GART_EN);
 825         } else {
 826                 RADEON_WRITE_PCIE(RADEON_PCIE_TX_GART_CNTL,
 827                                   tmp & ~RADEON_PCIE_TX_GART_EN);
 828         }
 829 }
 830
 831 /* Enable or disable PCI GART on the chip */
 832 static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
 833 {
 834         u32 tmp;
 835
 836         if (((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS690) ||
 837             ((dev_priv->flags & RADEON_FAMILY_MASK) == CHIP_RS740) ||
 838             (dev_priv->flags & RADEON_IS_IGPGART)) {
 839                 radeon_set_igpgart(dev_priv, on);
 840                 return;
 841         }
 842
 843         if (dev_priv->flags & RADEON_IS_PCIE) {
 844                 radeon_set_pciegart(dev_priv, on);
 845                 return;
 846         }
 847
 848         tmp = RADEON_READ(RADEON_AIC_CNTL);
 849
 850         if (on) {
 851                 RADEON_WRITE(RADEON_AIC_CNTL,
 852                              tmp | RADEON_PCIGART_TRANSLATE_EN);
 853
 854                 /* set PCI GART page-table base address
 855                  */
 856                 RADEON_WRITE(RADEON_AIC_PT_BASE, dev_priv->gart_info.bus_addr);
 857
 858                 /* set address range for PCI address translate
 859                  */
 860                 RADEON_WRITE(RADEON_AIC_LO_ADDR, dev_priv->gart_vm_start);
 861                 RADEON_WRITE(RADEON_AIC_HI_ADDR, dev_priv->gart_vm_start
 862                              + dev_priv->gart_size - 1);
 863
 864                 /* Turn off AGP aperture -- is this required for PCI GART?
 865                  */
 866                 radeon_write_agp_location(dev_priv, 0xffffffc0);
 867                 RADEON_WRITE(RADEON_AGP_COMMAND, 0);    /* clear AGP_COMMAND */
 868         } else {
 869                 RADEON_WRITE(RADEON_AIC_CNTL,
 870                              tmp & ~RADEON_PCIGART_TRANSLATE_EN);
 871         }
 872 }
 873
 874 static int radeon_do_init_cp(struct drm_device *dev, drm_radeon_init_t *init,
 875                              struct drm_file *file_priv)
 876 {
 877         drm_radeon_private_t *dev_priv = dev->dev_private;
 878         struct drm_radeon_master_private *master_priv = file_priv->master->driver_priv;
 879
 880         DRM_DEBUG("\n");
 881
 882         /* if we require new memory map but we don't have it fail */
 883         if ((dev_priv->flags & RADEON_NEW_MEMMAP) && !dev_priv->new_memmap) {
 884                 DRM_ERROR("Cannot initialise DRM on this card\nThis card requires a new X.org DDX for 3D\n");
 885                 radeon_do_cleanup_cp(dev);
 886                 return -EINVAL;
 887         }
 888
 889         if (init->is_pci && (dev_priv->flags & RADEON_IS_AGP)) {
 890                 DRM_DEBUG("Forcing AGP card to PCI mode\n");
 891                 dev_priv->flags &= ~RADEON_IS_AGP;
 892         } else if (!(dev_priv->flags & (RADEON_IS_AGP | RADEON_IS_PCI | RADEON_IS_PCIE))
 893                    && !init->is_pci) {
 894                 DRM_DEBUG("Restoring AGP flag\n");
 895                 dev_priv->flags |= RADEON_IS_AGP;
 896         }
 897
 898         if ((!(dev_priv->flags & RADEON_IS_AGP)) && !dev->sg) {
 899                 DRM_ERROR("PCI GART memory not allocated!\n");
 900                 radeon_do_cleanup_cp(dev);
 901                 return -EINVAL;
 902         }
 903
 904         dev_priv->usec_timeout = init->usec_timeout;
 905         if (dev_priv->usec_timeout < 1 ||
 906             dev_priv->usec_timeout > RADEON_MAX_USEC_TIMEOUT) {
 907                 DRM_DEBUG("TIMEOUT problem!\n");
 908                 radeon_do_cleanup_cp(dev);
 909                 return -EINVAL;
 910         }
 911
 912         /* Enable vblank on CRTC1 for older X servers
 913          */
 914         dev_priv->vblank_crtc = DRM_RADEON_VBLANK_CRTC1;
 915
 916         switch(init->func) {
 917         case RADEON_INIT_R200_CP:
 918                 dev_priv->microcode_version = UCODE_R200;
 919                 break;
 920         case RADEON_INIT_R300_CP:
 921                 dev_priv->microcode_version = UCODE_R300;
 922                 break;
 923         default:
 924                 dev_priv->microcode_version = UCODE_R100;
 925         }
 926
 927         dev_priv->do_boxes = 0;
 928         dev_priv->cp_mode = init->cp_mode;
 929
 930         /* We don't support anything other than bus-mastering ring mode,
 931          * but the ring can be in either AGP or PCI space for the ring
 932          * read pointer.
 933          */
 934         if ((init->cp_mode != RADEON_CSQ_PRIBM_INDDIS) &&
 935             (init->cp_mode != RADEON_CSQ_PRIBM_INDBM)) {
 936                 DRM_DEBUG("BAD cp_mode (%x)!\n", init->cp_mode);
 937                 radeon_do_cleanup_cp(dev);
 938                 return -EINVAL;
 939         }
 940
 941         switch (init->fb_bpp) {
 942         case 16:
 943                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_RGB565;
 944                 break;
 945         case 32:
 946         default:
 947                 dev_priv->color_fmt = RADEON_COLOR_FORMAT_ARGB8888;
 948                 break;
 949         }
 950         dev_priv->front_offset = init->front_offset;
 951         dev_priv->front_pitch = init->front_pitch;
 952         dev_priv->back_offset = init->back_offset;
 953         dev_priv->back_pitch = init->back_pitch;
 954
 955         switch (init->depth_bpp) {
 956         case 16:
 957                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_16BIT_INT_Z;
 958                 break;
 959         case 32:
 960         default:
 961                 dev_priv->depth_fmt = RADEON_DEPTH_FORMAT_24BIT_INT_Z;
 962                 break;
 963         }
 964         dev_priv->depth_offset = init->depth_offset;
 965         dev_priv->depth_pitch = init->depth_pitch;
 966
 967         /* Hardware state for depth clears.  Remove this if/when we no
 968          * longer clear the depth buffer with a 3D rectangle.  Hard-code
 969          * all values to prevent unwanted 3D state from slipping through
 970          * and screwing with the clear operation.
 971          */
 972         dev_priv->depth_clear.rb3d_cntl = (RADEON_PLANE_MASK_ENABLE |
 973                                            (dev_priv->color_fmt << 10) |
 974                                            (dev_priv->microcode_version ==
 975                                             UCODE_R100 ? RADEON_ZBLOCK16 : 0));
 976
 977         dev_priv->depth_clear.rb3d_zstencilcntl =
 978             (dev_priv->depth_fmt |
 979              RADEON_Z_TEST_ALWAYS |
 980              RADEON_STENCIL_TEST_ALWAYS |
 981              RADEON_STENCIL_S_FAIL_REPLACE |
 982              RADEON_STENCIL_ZPASS_REPLACE |
 983              RADEON_STENCIL_ZFAIL_REPLACE | RADEON_Z_WRITE_ENABLE);
 984
 985         dev_priv->depth_clear.se_cntl = (RADEON_FFACE_CULL_CW |
 986                                          RADEON_BFACE_SOLID |
 987                                          RADEON_FFACE_SOLID |
 988                                          RADEON_FLAT_SHADE_VTX_LAST |
 989                                          RADEON_DIFFUSE_SHADE_FLAT |
 990                                          RADEON_ALPHA_SHADE_FLAT |
 991                                          RADEON_SPECULAR_SHADE_FLAT |
 992                                          RADEON_FOG_SHADE_FLAT |
 993                                          RADEON_VTX_PIX_CENTER_OGL |
 994                                          RADEON_ROUND_MODE_TRUNC |
 995                                          RADEON_ROUND_PREC_8TH_PIX);
 996
 997
 998         dev_priv->ring_offset = init->ring_offset;
 999         dev_priv->ring_rptr_offset = init->ring_rptr_offset;
1000         dev_priv->buffers_offset = init->buffers_offset;
1001         dev_priv->gart_textures_offset = init->gart_textures_offset;
1002
1003         master_priv->sarea = drm_getsarea(dev);
1004         if (!master_priv->sarea) {
1005                 DRM_ERROR("could not find sarea!\n");
1006                 radeon_do_cleanup_cp(dev);
1007                 return -EINVAL;
1008         }
1009
1010         dev_priv->cp_ring = drm_core_findmap(dev, init->ring_offset);
1011         if (!dev_priv->cp_ring) {
1012                 DRM_ERROR("could not find cp ring region!\n");
1013                 radeon_do_cleanup_cp(dev);
1014                 return -EINVAL;
1015         }
1016         dev_priv->ring_rptr = drm_core_findmap(dev, init->ring_rptr_offset);
1017         if (!dev_priv->ring_rptr) {
1018                 DRM_ERROR("could not find ring read pointer!\n");
1019                 radeon_do_cleanup_cp(dev);
1020                 return -EINVAL;
1021         }
1022         dev->agp_buffer_token = init->buffers_offset;
1023         dev->agp_buffer_map = drm_core_findmap(dev, init->buffers_offset);
1024         if (!dev->agp_buffer_map) {
1025                 DRM_ERROR("could not find dma buffer region!\n");
1026                 radeon_do_cleanup_cp(dev);
1027                 return -EINVAL;
1028         }
1029
1030         if (init->gart_textures_offset) {
1031                 dev_priv->gart_textures =
1032                     drm_core_findmap(dev, init->gart_textures_offset);
1033                 if (!dev_priv->gart_textures) {
1034                         DRM_ERROR("could not find GART texture region!\n");
1035                         radeon_do_cleanup_cp(dev);
1036                         return -EINVAL;
1037                 }
1038         }
1039
1040 #if __OS_HAS_AGP
1041         if (dev_priv->flags & RADEON_IS_AGP) {
1042                 drm_core_ioremap(dev_priv->cp_ring, dev);
1043                 drm_core_ioremap(dev_priv->ring_rptr, dev);
1044                 drm_core_ioremap(dev->agp_buffer_map, dev);
1045                 if (!dev_priv->cp_ring->handle ||
1046                     !dev_priv->ring_rptr->handle ||
1047                     !dev->agp_buffer_map->handle) {
1048                         DRM_ERROR("could not find ioremap agp regions!\n");
1049                         radeon_do_cleanup_cp(dev);
1050                         return -EINVAL;
1051                 }
1052         } else
1053 #endif
1054         {
1055                 dev_priv->cp_ring->handle = (void *)dev_priv->cp_ring->offset;
1056                 dev_priv->ring_rptr->handle =
1057                     (void *)dev_priv->ring_rptr->offset;
1058                 dev->agp_buffer_map->handle =
1059                     (void *)dev->agp_buffer_map->offset;
1060
1061                 DRM_DEBUG("dev_priv->cp_ring->handle %p\n",
1062                           dev_priv->cp_ring->handle);
1063                 DRM_DEBUG("dev_priv->ring_rptr->handle %p\n",
1064                           dev_priv->ring_rptr->handle);
1065                 DRM_DEBUG("dev->agp_buffer_map->handle %p\n",
1066                           dev->agp_buffer_map->handle);
1067         }
1068
1069         dev_priv->fb_location = (radeon_read_fb_location(dev_priv) & 0xffff) << 16;
1070         dev_priv->fb_size =
1071                 ((radeon_read_fb_location(dev_priv) & 0xffff0000u) + 0x10000)
1072                 - dev_priv->fb_location;
1073
1074         dev_priv->front_pitch_offset = (((dev_priv->front_pitch / 64) << 22) |
1075                                         ((dev_priv->front_offset
1076                                           + dev_priv->fb_location) >> 10));
1077
1078         dev_priv->back_pitch_offset = (((dev_priv->back_pitch / 64) << 22) |
1079                                        ((dev_priv->back_offset
1080                                          + dev_priv->fb_location) >> 10));
1081
1082         dev_priv->depth_pitch_offset = (((dev_priv->depth_pitch / 64) << 22) |
1083                                         ((dev_priv->depth_offset
1084                                           + dev_priv->fb_location) >> 10));
1085
1086         dev_priv->gart_size = init->gart_size;
1087
1088         /* New let's set the memory map ... */
1089         if (dev_priv->new_memmap) {
1090                 u32 base = 0;
1091
1092                 DRM_INFO("Setting GART location based on new memory map\n");
1093
1094                 /* If using AGP, try to locate the AGP aperture at the same
1095                  * location in the card and on the bus, though we have to
1096                  * align it down.
1097                  */
1098 #if __OS_HAS_AGP
1099                 if (dev_priv->flags & RADEON_IS_AGP) {
1100                         base = dev->agp->base;
1101                         /* Check if valid */
1102                         if ((base + dev_priv->gart_size - 1) >= dev_priv->fb_location &&
1103                             base < (dev_priv->fb_location + dev_priv->fb_size - 1)) {
1104                                 DRM_INFO("Can't use AGP base @0x%08lx, won't fit\n",
1105                                          dev->agp->base);
1106                                 base = 0;
1107                         }
1108                 }
1109 #endif
1110                 /* If not or if AGP is at 0 (Macs), try to put it elsewhere */
1111                 if (base == 0) {
1112                         base = dev_priv->fb_location + dev_priv->fb_size;
1113                         if (base < dev_priv->fb_location ||
1114                             ((base + dev_priv->gart_size) & 0xfffffffful) < base)
1115                                 base = dev_priv->fb_location
1116                                         - dev_priv->gart_size;
1117                 }
1118                 dev_priv->gart_vm_start = base & 0xffc00000u;
1119                 if (dev_priv->gart_vm_start != base)
1120                         DRM_INFO("GART aligned down from 0x%08x to 0x%08x\n",
1121                                  base, dev_priv->gart_vm_start);
1122         } else {
1123                 DRM_INFO("Setting GART location based on old memory map\n");
1124                 dev_priv->gart_vm_start = dev_priv->fb_location +
1125                         RADEON_READ(RADEON_CONFIG_APER_SIZE);
1126         }
1127
1128 #if __OS_HAS_AGP
1129         if (dev_priv->flags & RADEON_IS_AGP)
1130                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1131                                                  - dev->agp->base
1132                                                  + dev_priv->gart_vm_start);
1133         else
1134 #endif
1135                 dev_priv->gart_buffers_offset = (dev->agp_buffer_map->offset
1136                                         - (unsigned long)dev->sg->virtual
1137                                         + dev_priv->gart_vm_start);
1138
1139         DRM_DEBUG("dev_priv->gart_size %d\n", dev_priv->gart_size);
1140         DRM_DEBUG("dev_priv->gart_vm_start 0x%x\n", dev_priv->gart_vm_start);
1141         DRM_DEBUG("dev_priv->gart_buffers_offset 0x%lx\n",
1142                   dev_priv->gart_buffers_offset);
1143
1144         dev_priv->ring.start = (u32 *) dev_priv->cp_ring->handle;
1145         dev_priv->ring.end = ((u32 *) dev_priv->cp_ring->handle
1146                               + init->ring_size / sizeof(u32));
1147         dev_priv->ring.size = init->ring_size;
1148         dev_priv->ring.size_l2qw = drm_order(init->ring_size / 8);
1149
1150         dev_priv->ring.rptr_update = /* init->rptr_update */ 4096;
1151         dev_priv->ring.rptr_update_l2qw = drm_order( /* init->rptr_update */ 4096 / 8);
1152
1153         dev_priv->ring.fetch_size = /* init->fetch_size */ 32;
1154         dev_priv->ring.fetch_size_l2ow = drm_order( /* init->fetch_size */ 32 / 16);
1155         dev_priv->ring.tail_mask = (dev_priv->ring.size / sizeof(u32)) - 1;
1156
1157         dev_priv->ring.high_mark = RADEON_RING_HIGH_MARK;
1158
1159 #if __OS_HAS_AGP
1160         if (dev_priv->flags & RADEON_IS_AGP) {
1161                 /* Turn off PCI GART */
1162                 radeon_set_pcigart(dev_priv, 0);
1163         } else
1164 #endif
1165         {
1166                 dev_priv->gart_info.table_mask = DMA_BIT_MASK(32);
1167                 /* if we have an offset set from userspace */
1168                 if (dev_priv->pcigart_offset_set) {
1169                         dev_priv->gart_info.bus_addr =
1170                             dev_priv->pcigart_offset + dev_priv->fb_location;
1171                         dev_priv->gart_info.mapping.offset =
1172                             dev_priv->pcigart_offset + dev_priv->fb_aper_offset;
1173                         dev_priv->gart_info.mapping.size =
1174                             dev_priv->gart_info.table_size;
1175
1176                         drm_core_ioremap_wc(&dev_priv->gart_info.mapping, dev);
1177                         dev_priv->gart_info.addr =
1178                             dev_priv->gart_info.mapping.handle;
1179
1180                         if (dev_priv->flags & RADEON_IS_PCIE)
1181                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
1182                         else
1183                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1184                         dev_priv->gart_info.gart_table_location =
1185                             DRM_ATI_GART_FB;
1186
1187                         DRM_DEBUG("Setting phys_pci_gart to %p %08lX\n",
1188                                   dev_priv->gart_info.addr,
1189                                   dev_priv->pcigart_offset);
1190                 } else {
1191                         if (dev_priv->flags & RADEON_IS_IGPGART)
1192                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
1193                         else
1194                                 dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
1195                         dev_priv->gart_info.gart_table_location =
1196                             DRM_ATI_GART_MAIN;
1197                         dev_priv->gart_info.addr = NULL;
1198                         dev_priv->gart_info.bus_addr = 0;
1199                         if (dev_priv->flags & RADEON_IS_PCIE) {
1200                                 DRM_ERROR
1201                                     ("Cannot use PCI Express without GART in FB memory\n");
1202                                 radeon_do_cleanup_cp(dev);
1203                                 return -EINVAL;
1204                         }
1205                 }
1206
1207                 if (!drm_ati_pcigart_init(dev, &dev_priv->gart_info)) {
1208                         DRM_ERROR("failed to init PCI GART!\n");
1209                         radeon_do_cleanup_cp(dev);
1210                         return -ENOMEM;
1211                 }
1212
1213                 /* Turn on PCI GART */
1214                 radeon_set_pcigart(dev_priv, 1);
1215         }
1216
1217         radeon_cp_load_microcode(dev_priv);
1218         radeon_cp_init_ring_buffer(dev, dev_priv);
1219
1220         dev_priv->last_buf = 0;
1221
1222         radeon_do_engine_reset(dev);
1223         radeon_test_writeback(dev_priv);
1224
1225         return 0;
1226 }
1227
1228 static int radeon_do_cleanup_cp(struct drm_device * dev)
1229 {
1230         drm_radeon_private_t *dev_priv = dev->dev_private;
1231         DRM_DEBUG("\n");
1232
1233         /* Make sure interrupts are disabled here because the uninstall ioctl
1234          * may not have been called from userspace and after dev_private
1235          * is freed, it's too late.
1236          */
1237         if (dev->irq_enabled)
1238                 drm_irq_uninstall(dev);
1239
1240 #if __OS_HAS_AGP
1241         if (dev_priv->flags & RADEON_IS_AGP) {
1242                 if (dev_priv->cp_ring != NULL) {
1243                         drm_core_ioremapfree(dev_priv->cp_ring, dev);
1244                         dev_priv->cp_ring = NULL;
1245                 }
1246                 if (dev_priv->ring_rptr != NULL) {
1247                         drm_core_ioremapfree(dev_priv->ring_rptr, dev);
1248                         dev_priv->ring_rptr = NULL;
1249                 }
1250                 if (dev->agp_buffer_map != NULL) {
1251                         drm_core_ioremapfree(dev->agp_buffer_map, dev);
1252                         dev->agp_buffer_map = NULL;
1253                 }
1254         } else
1255 #endif
1256         {
1257
1258                 if (dev_priv->gart_info.bus_addr) {
1259                         /* Turn off PCI GART */
1260                         radeon_set_pcigart(dev_priv, 0);
1261                         if (!drm_ati_pcigart_cleanup(dev, &dev_priv->gart_info))
1262                                 DRM_ERROR("failed to cleanup PCI GART!\n");
1263                 }
1264
1265                 if (dev_priv->gart_info.gart_table_location == DRM_ATI_GART_FB)
1266                 {
1267                         drm_core_ioremapfree(&dev_priv->gart_info.mapping, dev);
1268                         dev_priv->gart_info.addr = 0;
1269                 }
1270         }
1271         /* only clear to the start of flags */
1272         memset(dev_priv, 0, offsetof(drm_radeon_private_t, flags));
1273
1274         return 0;
1275 }
1276
1277 /* This code will reinit the Radeon CP hardware after a resume from disc.
1278  * AFAIK, it would be very difficult to pickle the state at suspend time, so
1279  * here we make sure that all Radeon hardware initialisation is re-done without
1280  * affecting running applications.
1281  *
1282  * Charl P. Botha <http://cpbotha.net>
1283  */
1284 static int radeon_do_resume_cp(struct drm_device * dev)
1285 {
1286         drm_radeon_private_t *dev_priv = dev->dev_private;
1287
1288         if (!dev_priv) {
1289                 DRM_ERROR("Called with no initialization\n");
1290                 return -EINVAL;
1291         }
1292
1293         DRM_DEBUG("Starting radeon_do_resume_cp()\n");
1294
1295 #if __OS_HAS_AGP
1296         if (dev_priv->flags & RADEON_IS_AGP) {
1297                 /* Turn off PCI GART */
1298                 radeon_set_pcigart(dev_priv, 0);
1299         } else
1300 #endif
1301         {
1302                 /* Turn on PCI GART */
1303                 radeon_set_pcigart(dev_priv, 1);
1304         }
1305
1306         radeon_cp_load_microcode(dev_priv);
1307         radeon_cp_init_ring_buffer(dev, dev_priv);
1308
1309         radeon_do_engine_reset(dev);
1310         radeon_irq_set_state(dev, RADEON_SW_INT_ENABLE, 1);
1311
1312         DRM_DEBUG("radeon_do_resume_cp() complete\n");
1313
1314         return 0;
1315 }
1316
1317 int radeon_cp_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
1318 {
1319         drm_radeon_init_t *init = data;
1320
1321         LOCK_TEST_WITH_RETURN(dev, file_priv);
1322
1323         if (init->func == RADEON_INIT_R300_CP)
1324                 r300_init_reg_flags(dev);
1325
1326         switch (init->func) {
1327         case RADEON_INIT_CP:
1328         case RADEON_INIT_R200_CP:
1329         case RADEON_INIT_R300_CP:
1330                 return radeon_do_init_cp(dev, init, file_priv);
1331         case RADEON_CLEANUP_CP:
1332                 return radeon_do_cleanup_cp(dev);
1333         }
1334
1335         return -EINVAL;
1336 }
1337
1338 int radeon_cp_start(struct drm_device *dev, void *data, struct drm_file *file_priv)
1339 {
1340         drm_radeon_private_t *dev_priv = dev->dev_private;
1341         DRM_DEBUG("\n");
1342
1343         LOCK_TEST_WITH_RETURN(dev, file_priv);
1344
1345         if (dev_priv->cp_running) {
1346                 DRM_DEBUG("while CP running\n");
1347                 return 0;
1348         }
1349         if (dev_priv->cp_mode == RADEON_CSQ_PRIDIS_INDDIS) {
1350                 DRM_DEBUG("called with bogus CP mode (%d)\n",
1351                           dev_priv->cp_mode);
1352                 return 0;
1353         }
1354
1355         radeon_do_cp_start(dev_priv);
1356
1357         return 0;
1358 }
1359
1360 /* Stop the CP.  The engine must have been idled before calling this
1361  * routine.
1362  */
1363 int radeon_cp_stop(struct drm_device *dev, void *data, struct drm_file *file_priv)
1364 {
1365         drm_radeon_private_t *dev_priv = dev->dev_private;
1366         drm_radeon_cp_stop_t *stop = data;
1367         int ret;
1368         DRM_DEBUG("\n");
1369
1370         LOCK_TEST_WITH_RETURN(dev, file_priv);
1371
1372         if (!dev_priv->cp_running)
1373                 return 0;
1374
1375         /* Flush any pending CP commands.  This ensures any outstanding
1376          * commands are exectuted by the engine before we turn it off.
1377          */
1378         if (stop->flush) {
1379                 radeon_do_cp_flush(dev_priv);
1380         }
1381
1382         /* If we fail to make the engine go idle, we return an error
1383          * code so that the DRM ioctl wrapper can try again.
1384          */
1385         if (stop->idle) {
1386                 ret = radeon_do_cp_idle(dev_priv);
1387                 if (ret)
1388                         return ret;
1389         }
1390
1391         /* Finally, we can turn off the CP.  If the engine isn't idle,
1392          * we will get some dropped triangles as they won't be fully
1393          * rendered before the CP is shut down.
1394          */
1395         radeon_do_cp_stop(dev_priv);
1396
1397         /* Reset the engine */
1398         radeon_do_engine_reset(dev);
1399
1400         return 0;
1401 }
1402
1403 void radeon_do_release(struct drm_device * dev)
1404 {
1405         drm_radeon_private_t *dev_priv = dev->dev_private;
1406         int i, ret;
1407
1408         if (dev_priv) {
1409                 if (dev_priv->cp_running) {
1410                         /* Stop the cp */
1411                         while ((ret = radeon_do_cp_idle(dev_priv)) != 0) {
1412                                 DRM_DEBUG("radeon_do_cp_idle %d\n", ret);
1413 #ifdef __linux__
1414                                 schedule();
1415 #else
1416                                 tsleep(&ret, PZERO, "rdnrel", 1);
1417 #endif
1418                         }
1419                         radeon_do_cp_stop(dev_priv);
1420                         radeon_do_engine_reset(dev);
1421                 }
1422
1423                 /* Disable *all* interrupts */
1424                 if (dev_priv->mmio)     /* remove this after permanent addmaps */
1425                         RADEON_WRITE(RADEON_GEN_INT_CNTL, 0);
1426
1427                 if (dev_priv->mmio) {   /* remove all surfaces */
1428                         for (i = 0; i < RADEON_MAX_SURFACES; i++) {
1429                                 RADEON_WRITE(RADEON_SURFACE0_INFO + 16 * i, 0);
1430                                 RADEON_WRITE(RADEON_SURFACE0_LOWER_BOUND +
1431                                              16 * i, 0);
1432                                 RADEON_WRITE(RADEON_SURFACE0_UPPER_BOUND +
1433                                              16 * i, 0);
1434                         }
1435                 }
1436
1437                 /* Free memory heap structures */
1438                 radeon_mem_takedown(&(dev_priv->gart_heap));
1439                 radeon_mem_takedown(&(dev_priv->fb_heap));
1440
1441                 /* deallocate kernel resources */
1442                 radeon_do_cleanup_cp(dev);
1443         }
1444 }
1445
1446 /* Just reset the CP ring.  Called as part of an X Server engine reset.
1447  */
1448 int radeon_cp_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1449 {
1450         drm_radeon_private_t *dev_priv = dev->dev_private;
1451         DRM_DEBUG("\n");
1452
1453         LOCK_TEST_WITH_RETURN(dev, file_priv);
1454
1455         if (!dev_priv) {
1456                 DRM_DEBUG("called before init done\n");
1457                 return -EINVAL;
1458         }
1459
1460         radeon_do_cp_reset(dev_priv);
1461
1462         /* The CP is no longer running after an engine reset */
1463         dev_priv->cp_running = 0;
1464
1465         return 0;
1466 }
1467
1468 int radeon_cp_idle(struct drm_device *dev, void *data, struct drm_file *file_priv)
1469 {
1470         drm_radeon_private_t *dev_priv = dev->dev_private;
1471         DRM_DEBUG("\n");
1472
1473         LOCK_TEST_WITH_RETURN(dev, file_priv);
1474
1475         return radeon_do_cp_idle(dev_priv);
1476 }
1477
1478 /* Added by Charl P. Botha to call radeon_do_resume_cp().
1479  */
1480 int radeon_cp_resume(struct drm_device *dev, void *data, struct drm_file *file_priv)
1481 {
1482
1483         return radeon_do_resume_cp(dev);
1484 }
1485
1486 int radeon_engine_reset(struct drm_device *dev, void *data, struct drm_file *file_priv)
1487 {
1488         DRM_DEBUG("\n");
1489
1490         LOCK_TEST_WITH_RETURN(dev, file_priv);
1491
1492         return radeon_do_engine_reset(dev);
1493 }
1494
1495 /* ================================================================
1496  * Fullscreen mode
1497  */
1498
1499 /* KW: Deprecated to say the least:
1500  */
1501 int radeon_fullscreen(struct drm_device *dev, void *data, struct drm_file *file_priv)
1502 {
1503         return 0;
1504 }
1505
1506 /* ================================================================
1507  * Freelist management
1508  */
1509
1510 /* Original comment: FIXME: ROTATE_BUFS is a hack to cycle through
1511  *   bufs until freelist code is used.  Note this hides a problem with
1512  *   the scratch register * (used to keep track of last buffer
1513  *   completed) being written to before * the last buffer has actually
1514  *   completed rendering.
1515  *
1516  * KW:  It's also a good way to find free buffers quickly.
1517  *
1518  * KW: Ideally this loop wouldn't exist, and freelist_get wouldn't
1519  * sleep.  However, bugs in older versions of radeon_accel.c mean that
1520  * we essentially have to do this, else old clients will break.
1521  *
1522  * However, it does leave open a potential deadlock where all the
1523  * buffers are held by other clients, which can't release them because
1524  * they can't get the lock.
1525  */
1526
1527 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1528 {
1529         struct drm_device_dma *dma = dev->dma;
1530         drm_radeon_private_t *dev_priv = dev->dev_private;
1531         drm_radeon_buf_priv_t *buf_priv;
1532         struct drm_buf *buf;
1533         int i, t;
1534         int start;
1535
1536         if (++dev_priv->last_buf >= dma->buf_count)
1537                 dev_priv->last_buf = 0;
1538
1539         start = dev_priv->last_buf;
1540
1541         for (t = 0; t < dev_priv->usec_timeout; t++) {
1542                 u32 done_age = GET_SCRATCH(1);
1543                 DRM_DEBUG("done_age = %d\n", done_age);
1544                 for (i = start; i < dma->buf_count; i++) {
1545                         buf = dma->buflist[i];
1546                         buf_priv = buf->dev_private;
1547                         if (buf->file_priv == NULL || (buf->pending &&
1548                                                        buf_priv->age <=
1549                                                        done_age)) {
1550                                 dev_priv->stats.requested_bufs++;
1551                                 buf->pending = 0;
1552                                 return buf;
1553                         }
1554                         start = 0;
1555                 }
1556
1557                 if (t) {
1558                         DRM_UDELAY(1);
1559                         dev_priv->stats.freelist_loops++;
1560                 }
1561         }
1562
1563         DRM_DEBUG("returning NULL!\n");
1564         return NULL;
1565 }
1566
1567 #if 0
1568 struct drm_buf *radeon_freelist_get(struct drm_device * dev)
1569 {
1570         struct drm_device_dma *dma = dev->dma;
1571         drm_radeon_private_t *dev_priv = dev->dev_private;
1572         drm_radeon_buf_priv_t *buf_priv;
1573         struct drm_buf *buf;
1574         int i, t;
1575         int start;
1576         u32 done_age = DRM_READ32(dev_priv->ring_rptr, RADEON_SCRATCHOFF(1));
1577
1578         if (++dev_priv->last_buf >= dma->buf_count)
1579                 dev_priv->last_buf = 0;
1580
1581         start = dev_priv->last_buf;
1582         dev_priv->stats.freelist_loops++;
1583
1584         for (t = 0; t < 2; t++) {
1585                 for (i = start; i < dma->buf_count; i++) {
1586                         buf = dma->buflist[i];
1587                         buf_priv = buf->dev_private;
1588                         if (buf->file_priv == 0 || (buf->pending &&
1589                                                     buf_priv->age <=
1590                                                     done_age)) {
1591                                 dev_priv->stats.requested_bufs++;
1592                                 buf->pending = 0;
1593                                 return buf;
1594                         }
1595                 }
1596                 start = 0;
1597         }
1598
1599         return NULL;
1600 }
1601 #endif
1602
1603 void radeon_freelist_reset(struct drm_device * dev)
1604 {
1605         struct drm_device_dma *dma = dev->dma;
1606         drm_radeon_private_t *dev_priv = dev->dev_private;
1607         int i;
1608
1609         dev_priv->last_buf = 0;
1610         for (i = 0; i < dma->buf_count; i++) {
1611                 struct drm_buf *buf = dma->buflist[i];
1612                 drm_radeon_buf_priv_t *buf_priv = buf->dev_private;
1613                 buf_priv->age = 0;
1614         }
1615 }
1616
1617 /* ================================================================
1618  * CP command submission
1619  */
1620
1621 int radeon_wait_ring(drm_radeon_private_t * dev_priv, int n)
1622 {
1623         drm_radeon_ring_buffer_t *ring = &dev_priv->ring;
1624         int i;
1625         u32 last_head = GET_RING_HEAD(dev_priv);
1626
1627         for (i = 0; i < dev_priv->usec_timeout; i++) {
1628                 u32 head = GET_RING_HEAD(dev_priv);
1629
1630                 ring->space = (head - ring->tail) * sizeof(u32);
1631                 if (ring->space <= 0)
1632                         ring->space += ring->size;
1633                 if (ring->space > n)
1634                         return 0;
1635
1636                 dev_priv->stats.boxes |= RADEON_BOX_WAIT_IDLE;
1637
1638                 if (head != last_head)
1639                         i = 0;
1640                 last_head = head;
1641
1642                 DRM_UDELAY(1);
1643         }
1644
1645         /* FIXME: This return value is ignored in the BEGIN_RING macro! */
1646 #if RADEON_FIFO_DEBUG
1647         radeon_status(dev_priv);
1648         DRM_ERROR("failed!\n");
1649 #endif
1650         return -EBUSY;
1651 }
1652
1653 static int radeon_cp_get_buffers(struct drm_device *dev,
1654                                  struct drm_file *file_priv,
1655                                  struct drm_dma * d)
1656 {
1657         int i;
1658         struct drm_buf *buf;
1659
1660         for (i = d->granted_count; i < d->request_count; i++) {
1661                 buf = radeon_freelist_get(dev);
1662                 if (!buf)
1663                         return -EBUSY;  /* NOTE: broken client */
1664
1665                 buf->file_priv = file_priv;
1666
1667                 if (DRM_COPY_TO_USER(&d->request_indices[i], &buf->idx,
1668                                      sizeof(buf->idx)))
1669                         return -EFAULT;
1670                 if (DRM_COPY_TO_USER(&d->request_sizes[i], &buf->total,
1671                                      sizeof(buf->total)))
1672                         return -EFAULT;
1673
1674                 d->granted_count++;
1675         }
1676         return 0;
1677 }
1678
1679 int radeon_cp_buffers(struct drm_device *dev, void *data, struct drm_file *file_priv)
1680 {
1681         struct drm_device_dma *dma = dev->dma;
1682         int ret = 0;
1683         struct drm_dma *d = data;
1684
1685         LOCK_TEST_WITH_RETURN(dev, file_priv);
1686
1687         /* Please don't send us buffers.
1688          */
1689         if (d->send_count != 0) {
1690                 DRM_ERROR("Process %d trying to send %d buffers via drmDMA\n",
1691                           DRM_CURRENTPID, d->send_count);
1692                 return -EINVAL;
1693         }
1694
1695         /* We'll send you buffers.
1696          */
1697         if (d->request_count < 0 || d->request_count > dma->buf_count) {
1698                 DRM_ERROR("Process %d trying to get %d buffers (of %d max)\n",
1699                           DRM_CURRENTPID, d->request_count, dma->buf_count);
1700                 return -EINVAL;
1701         }
1702
1703         d->granted_count = 0;
1704
1705         if (d->request_count) {
1706                 ret = radeon_cp_get_buffers(dev, file_priv, d);
1707         }
1708
1709         return ret;
1710 }
1711
1712 int radeon_driver_load(struct drm_device *dev, unsigned long flags)
1713 {
1714         drm_radeon_private_t *dev_priv;
1715         int ret = 0;
1716
1717         dev_priv = drm_alloc(sizeof(drm_radeon_private_t), DRM_MEM_DRIVER);
1718         if (dev_priv == NULL)
1719                 return -ENOMEM;
1720
1721         memset(dev_priv, 0, sizeof(drm_radeon_private_t));
1722         dev->dev_private = (void *)dev_priv;
1723         dev_priv->flags = flags;
1724
1725         switch (flags & RADEON_FAMILY_MASK) {
1726         case CHIP_R100:
1727         case CHIP_RV200:
1728         case CHIP_R200:
1729         case CHIP_R300:
1730         case CHIP_R350:
1731         case CHIP_R420:
1732         case CHIP_R423:
1733         case CHIP_RV410:
1734         case CHIP_RV515:
1735         case CHIP_R520:
1736         case CHIP_RV570:
1737         case CHIP_R580:
1738                 dev_priv->flags |= RADEON_HAS_HIERZ;
1739                 break;
1740         default:
1741                 /* all other chips have no hierarchical z buffer */
1742                 break;
1743         }
1744
1745         if (drm_device_is_agp(dev))
1746                 dev_priv->flags |= RADEON_IS_AGP;
1747         else if (drm_device_is_pcie(dev))
1748                 dev_priv->flags |= RADEON_IS_PCIE;
1749         else
1750                 dev_priv->flags |= RADEON_IS_PCI;
1751
1752         ret = drm_addmap(dev, drm_get_resource_start(dev, 2),
1753                          drm_get_resource_len(dev, 2), _DRM_REGISTERS,
1754                          _DRM_READ_ONLY | _DRM_DRIVER, &dev_priv->mmio);
1755         if (ret != 0)
1756                 return ret;
1757
1758         ret = drm_vblank_init(dev, 2);
1759         if (ret) {
1760                 radeon_driver_unload(dev);
1761                 return ret;
1762         }
1763
1764         DRM_DEBUG("%s card detected\n",
1765                   ((dev_priv->flags & RADEON_IS_AGP) ? "AGP" : (((dev_priv->flags & RADEON_IS_PCIE) ? "PCIE" : "PCI"))));
1766         return ret;
1767 }
1768
1769 int radeon_master_create(struct drm_device *dev, struct drm_master *master)
1770 {
1771         struct drm_radeon_master_private *master_priv;
1772         unsigned long sareapage;
1773         int ret;
1774
1775         master_priv = drm_calloc(1, sizeof(*master_priv), DRM_MEM_DRIVER);
1776         if (!master_priv)
1777                 return -ENOMEM;
1778
1779         /* prebuild the SAREA */
1780         sareapage = max_t(unsigned long, SAREA_MAX, PAGE_SIZE);
1781         ret = drm_addmap(dev, 0, sareapage, _DRM_SHM, _DRM_CONTAINS_LOCK|_DRM_DRIVER,
1782                          &master_priv->sarea);
1783         if (ret) {
1784                 DRM_ERROR("SAREA setup failed\n");
1785                 return ret;
1786         }
1787         master_priv->sarea_priv = master_priv->sarea->handle + sizeof(struct drm_sarea);
1788         master_priv->sarea_priv->pfCurrentPage = 0;
1789
1790         master->driver_priv = master_priv;
1791         return 0;
1792 }
1793
1794 void radeon_master_destroy(struct drm_device *dev, struct drm_master *master)
1795 {
1796         struct drm_radeon_master_private *master_priv = master->driver_priv;
1797
1798         if (!master_priv)
1799                 return;
1800
1801         if (master_priv->sarea_priv &&
1802             master_priv->sarea_priv->pfCurrentPage != 0)
1803                 radeon_cp_dispatch_flip(dev, master);
1804
1805         master_priv->sarea_priv = NULL;
1806         if (master_priv->sarea)
1807                 drm_rmmap_locked(dev, master_priv->sarea);
1808
1809         drm_free(master_priv, sizeof(*master_priv), DRM_MEM_DRIVER);
1810
1811         master->driver_priv = NULL;
1812 }
1813
1814 /* Create mappings for registers and framebuffer so userland doesn't necessarily
1815  * have to find them.
1816  */
1817 int radeon_driver_firstopen(struct drm_device *dev)
1818 {
1819         int ret;
1820         drm_local_map_t *map;
1821         drm_radeon_private_t *dev_priv = dev->dev_private;
1822
1823         dev_priv->gart_info.table_size = RADEON_PCIGART_TABLE_SIZE;
1824
1825         dev_priv->fb_aper_offset = drm_get_resource_start(dev, 0);
1826         ret = drm_addmap(dev, dev_priv->fb_aper_offset,
1827                          drm_get_resource_len(dev, 0), _DRM_FRAME_BUFFER,
1828                          _DRM_WRITE_COMBINING, &map);
1829         if (ret != 0)
1830                 return ret;
1831
1832         return 0;
1833 }
1834
1835 int radeon_driver_unload(struct drm_device *dev)
1836 {
1837         drm_radeon_private_t *dev_priv = dev->dev_private;
1838
1839         DRM_DEBUG("\n");
1840
1841         drm_rmmap(dev, dev_priv->mmio);
1842
1843         drm_free(dev_priv, sizeof(*dev_priv), DRM_MEM_DRIVER);
1844
1845         dev->dev_private = NULL;
1846         return 0;
1847 }