shared-core/r128_state.c

   1 /* r128_state.c -- State support for r128 -*- linux-c -*-
   2  * Created: Thu Jan 27 02:53:43 2000 by gareth@valinux.com
   3  *
   4  * Copyright 2000 VA Linux Systems, Inc., Sunnyvale, California.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the next
  15  * paragraph) shall be included in all copies or substantial portions of the
  16  * Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  21  * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  22  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  23  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  24  * DEALINGS IN THE SOFTWARE.
  25  *
  26  * Authors:
  27  *    Gareth Hughes <gareth@valinux.com>
  28  */
  29
  30 #include "r128.h"
  31 #include "drmP.h"
  32 #include "drm.h"
  33 #include "r128_drm.h"
  34 #include "r128_drv.h"
  35
  36
  37 /* ================================================================
  38  * CCE hardware state programming functions
  39  */
  40
  41 static void r128_emit_clip_rects( drm_r128_private_t *dev_priv,
  42                                   drm_clip_rect_t *boxes, int count )
  43 {
  44         u32 aux_sc_cntl = 0x00000000;
  45         RING_LOCALS;
  46         DRM_DEBUG( "    %s\n", __FUNCTION__ );
  47
  48         BEGIN_RING( (count < 3? count: 3) * 5 + 2 );
  49
  50         if ( count >= 1 ) {
  51                 OUT_RING( CCE_PACKET0( R128_AUX1_SC_LEFT, 3 ) );
  52                 OUT_RING( boxes[0].x1 );
  53                 OUT_RING( boxes[0].x2 - 1 );
  54                 OUT_RING( boxes[0].y1 );
  55                 OUT_RING( boxes[0].y2 - 1 );
  56
  57                 aux_sc_cntl |= (R128_AUX1_SC_EN | R128_AUX1_SC_MODE_OR);
  58         }
  59         if ( count >= 2 ) {
  60                 OUT_RING( CCE_PACKET0( R128_AUX2_SC_LEFT, 3 ) );
  61                 OUT_RING( boxes[1].x1 );
  62                 OUT_RING( boxes[1].x2 - 1 );
  63                 OUT_RING( boxes[1].y1 );
  64                 OUT_RING( boxes[1].y2 - 1 );
  65
  66                 aux_sc_cntl |= (R128_AUX2_SC_EN | R128_AUX2_SC_MODE_OR);
  67         }
  68         if ( count >= 3 ) {
  69                 OUT_RING( CCE_PACKET0( R128_AUX3_SC_LEFT, 3 ) );
  70                 OUT_RING( boxes[2].x1 );
  71                 OUT_RING( boxes[2].x2 - 1 );
  72                 OUT_RING( boxes[2].y1 );
  73                 OUT_RING( boxes[2].y2 - 1 );
  74
  75                 aux_sc_cntl |= (R128_AUX3_SC_EN | R128_AUX3_SC_MODE_OR);
  76         }
  77
  78         OUT_RING( CCE_PACKET0( R128_AUX_SC_CNTL, 0 ) );
  79         OUT_RING( aux_sc_cntl );
  80
  81         ADVANCE_RING();
  82 }
  83
  84 static __inline__ void r128_emit_core( drm_r128_private_t *dev_priv )
  85 {
  86         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
  87         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
  88         RING_LOCALS;
  89         DRM_DEBUG( "    %s\n", __FUNCTION__ );
  90
  91         BEGIN_RING( 2 );
  92
  93         OUT_RING( CCE_PACKET0( R128_SCALE_3D_CNTL, 0 ) );
  94         OUT_RING( ctx->scale_3d_cntl );
  95
  96         ADVANCE_RING();
  97 }
  98
  99 static __inline__ void r128_emit_context( drm_r128_private_t *dev_priv )
 100 {
 101         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 102         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
 103         RING_LOCALS;
 104         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 105
 106         BEGIN_RING( 13 );
 107
 108         OUT_RING( CCE_PACKET0( R128_DST_PITCH_OFFSET_C, 11 ) );
 109         OUT_RING( ctx->dst_pitch_offset_c );
 110         OUT_RING( ctx->dp_gui_master_cntl_c );
 111         OUT_RING( ctx->sc_top_left_c );
 112         OUT_RING( ctx->sc_bottom_right_c );
 113         OUT_RING( ctx->z_offset_c );
 114         OUT_RING( ctx->z_pitch_c );
 115         OUT_RING( ctx->z_sten_cntl_c );
 116         OUT_RING( ctx->tex_cntl_c );
 117         OUT_RING( ctx->misc_3d_state_cntl_reg );
 118         OUT_RING( ctx->texture_clr_cmp_clr_c );
 119         OUT_RING( ctx->texture_clr_cmp_msk_c );
 120         OUT_RING( ctx->fog_color_c );
 121
 122         ADVANCE_RING();
 123 }
 124
 125 static __inline__ void r128_emit_setup( drm_r128_private_t *dev_priv )
 126 {
 127         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 128         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
 129         RING_LOCALS;
 130         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 131
 132         BEGIN_RING( 3 );
 133
 134         OUT_RING( CCE_PACKET1( R128_SETUP_CNTL, R128_PM4_VC_FPU_SETUP ) );
 135         OUT_RING( ctx->setup_cntl );
 136         OUT_RING( ctx->pm4_vc_fpu_setup );
 137
 138         ADVANCE_RING();
 139 }
 140
 141 static __inline__ void r128_emit_masks( drm_r128_private_t *dev_priv )
 142 {
 143         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 144         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
 145         RING_LOCALS;
 146         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 147
 148         BEGIN_RING( 5 );
 149
 150         OUT_RING( CCE_PACKET0( R128_DP_WRITE_MASK, 0 ) );
 151         OUT_RING( ctx->dp_write_mask );
 152
 153         OUT_RING( CCE_PACKET0( R128_STEN_REF_MASK_C, 1 ) );
 154         OUT_RING( ctx->sten_ref_mask_c );
 155         OUT_RING( ctx->plane_3d_mask_c );
 156
 157         ADVANCE_RING();
 158 }
 159
 160 static __inline__ void r128_emit_window( drm_r128_private_t *dev_priv )
 161 {
 162         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 163         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
 164         RING_LOCALS;
 165         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 166
 167         BEGIN_RING( 2 );
 168
 169         OUT_RING( CCE_PACKET0( R128_WINDOW_XY_OFFSET, 0 ) );
 170         OUT_RING( ctx->window_xy_offset );
 171
 172         ADVANCE_RING();
 173 }
 174
 175 static __inline__ void r128_emit_tex0( drm_r128_private_t *dev_priv )
 176 {
 177         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 178         drm_r128_context_regs_t *ctx = &sarea_priv->context_state;
 179         drm_r128_texture_regs_t *tex = &sarea_priv->tex_state[0];
 180         int i;
 181         RING_LOCALS;
 182         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 183
 184         BEGIN_RING( 7 + R128_MAX_TEXTURE_LEVELS );
 185
 186         OUT_RING( CCE_PACKET0( R128_PRIM_TEX_CNTL_C,
 187                                2 + R128_MAX_TEXTURE_LEVELS ) );
 188         OUT_RING( tex->tex_cntl );
 189         OUT_RING( tex->tex_combine_cntl );
 190         OUT_RING( ctx->tex_size_pitch_c );
 191         for ( i = 0 ; i < R128_MAX_TEXTURE_LEVELS ; i++ ) {
 192                 OUT_RING( tex->tex_offset[i] );
 193         }
 194
 195         OUT_RING( CCE_PACKET0( R128_CONSTANT_COLOR_C, 1 ) );
 196         OUT_RING( ctx->constant_color_c );
 197         OUT_RING( tex->tex_border_color );
 198
 199         ADVANCE_RING();
 200 }
 201
 202 static __inline__ void r128_emit_tex1( drm_r128_private_t *dev_priv )
 203 {
 204         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 205         drm_r128_texture_regs_t *tex = &sarea_priv->tex_state[1];
 206         int i;
 207         RING_LOCALS;
 208         DRM_DEBUG( "    %s\n", __FUNCTION__ );
 209
 210         BEGIN_RING( 5 + R128_MAX_TEXTURE_LEVELS );
 211
 212         OUT_RING( CCE_PACKET0( R128_SEC_TEX_CNTL_C,
 213                                1 + R128_MAX_TEXTURE_LEVELS ) );
 214         OUT_RING( tex->tex_cntl );
 215         OUT_RING( tex->tex_combine_cntl );
 216         for ( i = 0 ; i < R128_MAX_TEXTURE_LEVELS ; i++ ) {
 217                 OUT_RING( tex->tex_offset[i] );
 218         }
 219
 220         OUT_RING( CCE_PACKET0( R128_SEC_TEXTURE_BORDER_COLOR_C, 0 ) );
 221         OUT_RING( tex->tex_border_color );
 222
 223         ADVANCE_RING();
 224 }
 225
 226 static __inline__ void r128_emit_state( drm_r128_private_t *dev_priv )
 227 {
 228         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 229         unsigned int dirty = sarea_priv->dirty;
 230
 231         DRM_DEBUG( "%s: dirty=0x%08x\n", __FUNCTION__, dirty );
 232
 233         if ( dirty & R128_UPLOAD_CORE ) {
 234                 r128_emit_core( dev_priv );
 235                 sarea_priv->dirty &= ~R128_UPLOAD_CORE;
 236         }
 237
 238         if ( dirty & R128_UPLOAD_CONTEXT ) {
 239                 r128_emit_context( dev_priv );
 240                 sarea_priv->dirty &= ~R128_UPLOAD_CONTEXT;
 241         }
 242
 243         if ( dirty & R128_UPLOAD_SETUP ) {
 244                 r128_emit_setup( dev_priv );
 245                 sarea_priv->dirty &= ~R128_UPLOAD_SETUP;
 246         }
 247
 248         if ( dirty & R128_UPLOAD_MASKS ) {
 249                 r128_emit_masks( dev_priv );
 250                 sarea_priv->dirty &= ~R128_UPLOAD_MASKS;
 251         }
 252
 253         if ( dirty & R128_UPLOAD_WINDOW ) {
 254                 r128_emit_window( dev_priv );
 255                 sarea_priv->dirty &= ~R128_UPLOAD_WINDOW;
 256         }
 257
 258         if ( dirty & R128_UPLOAD_TEX0 ) {
 259                 r128_emit_tex0( dev_priv );
 260                 sarea_priv->dirty &= ~R128_UPLOAD_TEX0;
 261         }
 262
 263         if ( dirty & R128_UPLOAD_TEX1 ) {
 264                 r128_emit_tex1( dev_priv );
 265                 sarea_priv->dirty &= ~R128_UPLOAD_TEX1;
 266         }
 267
 268         /* Turn off the texture cache flushing */
 269         sarea_priv->context_state.tex_cntl_c &= ~R128_TEX_CACHE_FLUSH;
 270
 271         sarea_priv->dirty &= ~R128_REQUIRE_QUIESCENCE;
 272 }
 273
 274
 275 #if R128_PERFORMANCE_BOXES
 276 /* ================================================================
 277  * Performance monitoring functions
 278  */
 279
 280 static void r128_clear_box( drm_r128_private_t *dev_priv,
 281                             int x, int y, int w, int h,
 282                             int r, int g, int b )
 283 {
 284         u32 pitch, offset;
 285         u32 fb_bpp, color;
 286         RING_LOCALS;
 287
 288         switch ( dev_priv->fb_bpp ) {
 289         case 16:
 290                 fb_bpp = R128_GMC_DST_16BPP;
 291                 color = (((r & 0xf8) << 8) |
 292                          ((g & 0xfc) << 3) |
 293                          ((b & 0xf8) >> 3));
 294                 break;
 295         case 24:
 296                 fb_bpp = R128_GMC_DST_24BPP;
 297                 color = ((r << 16) | (g << 8) | b);
 298                 break;
 299         case 32:
 300                 fb_bpp = R128_GMC_DST_32BPP;
 301                 color = (((0xff) << 24) | (r << 16) | (g <<  8) | b);
 302                 break;
 303         default:
 304                 return;
 305         }
 306
 307         offset = dev_priv->back_offset;
 308         pitch = dev_priv->back_pitch >> 3;
 309
 310         BEGIN_RING( 6 );
 311
 312         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 313         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 314                   R128_GMC_BRUSH_SOLID_COLOR |
 315                   fb_bpp |
 316                   R128_GMC_SRC_DATATYPE_COLOR |
 317                   R128_ROP3_P |
 318                   R128_GMC_CLR_CMP_CNTL_DIS |
 319                   R128_GMC_AUX_CLIP_DIS );
 320
 321         OUT_RING( (pitch << 21) | (offset >> 5) );
 322         OUT_RING( color );
 323
 324         OUT_RING( (x << 16) | y );
 325         OUT_RING( (w << 16) | h );
 326
 327         ADVANCE_RING();
 328 }
 329
 330 static void r128_cce_performance_boxes( drm_r128_private_t *dev_priv )
 331 {
 332         if ( atomic_read( &dev_priv->idle_count ) == 0 ) {
 333                 r128_clear_box( dev_priv, 64, 4, 8, 8, 0, 255, 0 );
 334         } else {
 335                 atomic_set( &dev_priv->idle_count, 0 );
 336         }
 337 }
 338
 339 #endif
 340
 341
 342 /* ================================================================
 343  * CCE command dispatch functions
 344  */
 345
 346 static void r128_print_dirty( const char *msg, unsigned int flags )
 347 {
 348         DRM_INFO( "%s: (0x%x) %s%s%s%s%s%s%s%s%s\n",
 349                   msg,
 350                   flags,
 351                   (flags & R128_UPLOAD_CORE)        ? "core, " : "",
 352                   (flags & R128_UPLOAD_CONTEXT)     ? "context, " : "",
 353                   (flags & R128_UPLOAD_SETUP)       ? "setup, " : "",
 354                   (flags & R128_UPLOAD_TEX0)        ? "tex0, " : "",
 355                   (flags & R128_UPLOAD_TEX1)        ? "tex1, " : "",
 356                   (flags & R128_UPLOAD_MASKS)       ? "masks, " : "",
 357                   (flags & R128_UPLOAD_WINDOW)      ? "window, " : "",
 358                   (flags & R128_UPLOAD_CLIPRECTS)   ? "cliprects, " : "",
 359                   (flags & R128_REQUIRE_QUIESCENCE) ? "quiescence, " : "" );
 360 }
 361
 362 static void r128_cce_dispatch_clear( drm_device_t *dev,
 363                                      drm_r128_clear_t *clear )
 364 {
 365         drm_r128_private_t *dev_priv = dev->dev_private;
 366         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 367         int nbox = sarea_priv->nbox;
 368         drm_clip_rect_t *pbox = sarea_priv->boxes;
 369         unsigned int flags = clear->flags;
 370         int i;
 371         RING_LOCALS;
 372         DRM_DEBUG( "%s\n", __FUNCTION__ );
 373
 374         if ( dev_priv->page_flipping && dev_priv->current_page == 1 ) {
 375                 unsigned int tmp = flags;
 376
 377                 flags &= ~(R128_FRONT | R128_BACK);
 378                 if ( tmp & R128_FRONT ) flags |= R128_BACK;
 379                 if ( tmp & R128_BACK )  flags |= R128_FRONT;
 380         }
 381
 382         for ( i = 0 ; i < nbox ; i++ ) {
 383                 int x = pbox[i].x1;
 384                 int y = pbox[i].y1;
 385                 int w = pbox[i].x2 - x;
 386                 int h = pbox[i].y2 - y;
 387
 388                 DRM_DEBUG( "dispatch clear %d,%d-%d,%d flags 0x%x\n",
 389                            pbox[i].x1, pbox[i].y1, pbox[i].x2,
 390                            pbox[i].y2, flags );
 391
 392                 if ( flags & (R128_FRONT | R128_BACK) ) {
 393                         BEGIN_RING( 2 );
 394
 395                         OUT_RING( CCE_PACKET0( R128_DP_WRITE_MASK, 0 ) );
 396                         OUT_RING( clear->color_mask );
 397
 398                         ADVANCE_RING();
 399                 }
 400
 401                 if ( flags & R128_FRONT ) {
 402                         BEGIN_RING( 6 );
 403
 404                         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 405                         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 406                                   R128_GMC_BRUSH_SOLID_COLOR |
 407                                   (dev_priv->color_fmt << 8) |
 408                                   R128_GMC_SRC_DATATYPE_COLOR |
 409                                   R128_ROP3_P |
 410                                   R128_GMC_CLR_CMP_CNTL_DIS |
 411                                   R128_GMC_AUX_CLIP_DIS );
 412
 413                         OUT_RING( dev_priv->front_pitch_offset_c );
 414                         OUT_RING( clear->clear_color );
 415
 416                         OUT_RING( (x << 16) | y );
 417                         OUT_RING( (w << 16) | h );
 418
 419                         ADVANCE_RING();
 420                 }
 421
 422                 if ( flags & R128_BACK ) {
 423                         BEGIN_RING( 6 );
 424
 425                         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 426                         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 427                                   R128_GMC_BRUSH_SOLID_COLOR |
 428                                   (dev_priv->color_fmt << 8) |
 429                                   R128_GMC_SRC_DATATYPE_COLOR |
 430                                   R128_ROP3_P |
 431                                   R128_GMC_CLR_CMP_CNTL_DIS |
 432                                   R128_GMC_AUX_CLIP_DIS );
 433
 434                         OUT_RING( dev_priv->back_pitch_offset_c );
 435                         OUT_RING( clear->clear_color );
 436
 437                         OUT_RING( (x << 16) | y );
 438                         OUT_RING( (w << 16) | h );
 439
 440                         ADVANCE_RING();
 441                 }
 442
 443                 if ( flags & R128_DEPTH ) {
 444                         BEGIN_RING( 6 );
 445
 446                         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 447                         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 448                                   R128_GMC_BRUSH_SOLID_COLOR |
 449                                   (dev_priv->depth_fmt << 8) |
 450                                   R128_GMC_SRC_DATATYPE_COLOR |
 451                                   R128_ROP3_P |
 452                                   R128_GMC_CLR_CMP_CNTL_DIS |
 453                                   R128_GMC_AUX_CLIP_DIS |
 454                                   R128_GMC_WR_MSK_DIS );
 455
 456                         OUT_RING( dev_priv->depth_pitch_offset_c );
 457                         OUT_RING( clear->clear_depth );
 458
 459                         OUT_RING( (x << 16) | y );
 460                         OUT_RING( (w << 16) | h );
 461
 462                         ADVANCE_RING();
 463                 }
 464         }
 465 }
 466
 467 static void r128_cce_dispatch_swap( drm_device_t *dev )
 468 {
 469         drm_r128_private_t *dev_priv = dev->dev_private;
 470         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 471         int nbox = sarea_priv->nbox;
 472         drm_clip_rect_t *pbox = sarea_priv->boxes;
 473         int i;
 474         RING_LOCALS;
 475         DRM_DEBUG( "%s\n", __FUNCTION__ );
 476
 477 #if R128_PERFORMANCE_BOXES
 478         /* Do some trivial performance monitoring...
 479          */
 480         r128_cce_performance_boxes( dev_priv );
 481 #endif
 482
 483         for ( i = 0 ; i < nbox ; i++ ) {
 484                 int x = pbox[i].x1;
 485                 int y = pbox[i].y1;
 486                 int w = pbox[i].x2 - x;
 487                 int h = pbox[i].y2 - y;
 488
 489                 BEGIN_RING( 7 );
 490
 491                 OUT_RING( CCE_PACKET3( R128_CNTL_BITBLT_MULTI, 5 ) );
 492                 OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL |
 493                           R128_GMC_DST_PITCH_OFFSET_CNTL |
 494                           R128_GMC_BRUSH_NONE |
 495                           (dev_priv->color_fmt << 8) |
 496                           R128_GMC_SRC_DATATYPE_COLOR |
 497                           R128_ROP3_S |
 498                           R128_DP_SRC_SOURCE_MEMORY |
 499                           R128_GMC_CLR_CMP_CNTL_DIS |
 500                           R128_GMC_AUX_CLIP_DIS |
 501                           R128_GMC_WR_MSK_DIS );
 502
 503                 /* Make this work even if front & back are flipped:
 504                  */
 505                 if (dev_priv->current_page == 0) {
 506                         OUT_RING( dev_priv->back_pitch_offset_c );
 507                         OUT_RING( dev_priv->front_pitch_offset_c );
 508                 }
 509                 else {
 510                         OUT_RING( dev_priv->front_pitch_offset_c );
 511                         OUT_RING( dev_priv->back_pitch_offset_c );
 512                 }
 513
 514                 OUT_RING( (x << 16) | y );
 515                 OUT_RING( (x << 16) | y );
 516                 OUT_RING( (w << 16) | h );
 517
 518                 ADVANCE_RING();
 519         }
 520
 521         /* Increment the frame counter.  The client-side 3D driver must
 522          * throttle the framerate by waiting for this value before
 523          * performing the swapbuffer ioctl.
 524          */
 525         dev_priv->sarea_priv->last_frame++;
 526
 527         BEGIN_RING( 2 );
 528
 529         OUT_RING( CCE_PACKET0( R128_LAST_FRAME_REG, 0 ) );
 530         OUT_RING( dev_priv->sarea_priv->last_frame );
 531
 532         ADVANCE_RING();
 533 }
 534
 535 static void r128_cce_dispatch_flip( drm_device_t *dev )
 536 {
 537         drm_r128_private_t *dev_priv = dev->dev_private;
 538         RING_LOCALS;
 539         DRM_DEBUG("%s: page=%d pfCurrentPage=%d\n",
 540                 __FUNCTION__,
 541                 dev_priv->current_page,
 542                 dev_priv->sarea_priv->pfCurrentPage);
 543
 544 #if R128_PERFORMANCE_BOXES
 545         /* Do some trivial performance monitoring...
 546          */
 547         r128_cce_performance_boxes( dev_priv );
 548 #endif
 549
 550         BEGIN_RING( 4 );
 551
 552         R128_WAIT_UNTIL_PAGE_FLIPPED();
 553         OUT_RING( CCE_PACKET0( R128_CRTC_OFFSET, 0 ) );
 554
 555         if ( dev_priv->current_page == 0 ) {
 556                 OUT_RING( dev_priv->back_offset );
 557         } else {
 558                 OUT_RING( dev_priv->front_offset );
 559         }
 560
 561         ADVANCE_RING();
 562
 563         /* Increment the frame counter.  The client-side 3D driver must
 564          * throttle the framerate by waiting for this value before
 565          * performing the swapbuffer ioctl.
 566          */
 567         dev_priv->sarea_priv->last_frame++;
 568         dev_priv->sarea_priv->pfCurrentPage = dev_priv->current_page =
 569                                               1 - dev_priv->current_page;
 570
 571         BEGIN_RING( 2 );
 572
 573         OUT_RING( CCE_PACKET0( R128_LAST_FRAME_REG, 0 ) );
 574         OUT_RING( dev_priv->sarea_priv->last_frame );
 575
 576         ADVANCE_RING();
 577 }
 578
 579 static void r128_cce_dispatch_vertex( drm_device_t *dev,
 580                                       drm_buf_t *buf )
 581 {
 582         drm_r128_private_t *dev_priv = dev->dev_private;
 583         drm_r128_buf_priv_t *buf_priv = buf->dev_private;
 584         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 585         int format = sarea_priv->vc_format;
 586         int offset = buf->bus_address;
 587         int size = buf->used;
 588         int prim = buf_priv->prim;
 589         int i = 0;
 590         RING_LOCALS;
 591         DRM_DEBUG( "buf=%d nbox=%d\n", buf->idx, sarea_priv->nbox );
 592
 593         if ( 0 )
 594                 r128_print_dirty( "dispatch_vertex", sarea_priv->dirty );
 595
 596         if ( buf->used ) {
 597                 buf_priv->dispatched = 1;
 598
 599                 if ( sarea_priv->dirty & ~R128_UPLOAD_CLIPRECTS ) {
 600                         r128_emit_state( dev_priv );
 601                 }
 602
 603                 do {
 604                         /* Emit the next set of up to three cliprects */
 605                         if ( i < sarea_priv->nbox ) {
 606                                 r128_emit_clip_rects( dev_priv,
 607                                                       &sarea_priv->boxes[i],
 608                                                       sarea_priv->nbox - i );
 609                         }
 610
 611                         /* Emit the vertex buffer rendering commands */
 612                         BEGIN_RING( 5 );
 613
 614                         OUT_RING( CCE_PACKET3( R128_3D_RNDR_GEN_INDX_PRIM, 3 ) );
 615                         OUT_RING( offset );
 616                         OUT_RING( size );
 617                         OUT_RING( format );
 618                         OUT_RING( prim | R128_CCE_VC_CNTL_PRIM_WALK_LIST |
 619                                   (size << R128_CCE_VC_CNTL_NUM_SHIFT) );
 620
 621                         ADVANCE_RING();
 622
 623                         i += 3;
 624                 } while ( i < sarea_priv->nbox );
 625         }
 626
 627         if ( buf_priv->discard ) {
 628                 buf_priv->age = dev_priv->sarea_priv->last_dispatch;
 629
 630                 /* Emit the vertex buffer age */
 631                 BEGIN_RING( 2 );
 632
 633                 OUT_RING( CCE_PACKET0( R128_LAST_DISPATCH_REG, 0 ) );
 634                 OUT_RING( buf_priv->age );
 635
 636                 ADVANCE_RING();
 637
 638                 buf->pending = 1;
 639                 buf->used = 0;
 640                 /* FIXME: Check dispatched field */
 641                 buf_priv->dispatched = 0;
 642         }
 643
 644         dev_priv->sarea_priv->last_dispatch++;
 645
 646         sarea_priv->dirty &= ~R128_UPLOAD_CLIPRECTS;
 647         sarea_priv->nbox = 0;
 648 }
 649
 650 static void r128_cce_dispatch_indirect( drm_device_t *dev,
 651                                         drm_buf_t *buf,
 652                                         int start, int end )
 653 {
 654         drm_r128_private_t *dev_priv = dev->dev_private;
 655         drm_r128_buf_priv_t *buf_priv = buf->dev_private;
 656         RING_LOCALS;
 657         DRM_DEBUG( "indirect: buf=%d s=0x%x e=0x%x\n",
 658                    buf->idx, start, end );
 659
 660         if ( start != end ) {
 661                 int offset = buf->bus_address + start;
 662                 int dwords = (end - start + 3) / sizeof(u32);
 663
 664                 /* Indirect buffer data must be an even number of
 665                  * dwords, so if we've been given an odd number we must
 666                  * pad the data with a Type-2 CCE packet.
 667                  */
 668                 if ( dwords & 1 ) {
 669                         u32 *data = (u32 *)
 670                                 ((char *)dev_priv->buffers->handle
 671                                  + buf->offset + start);
 672                         data[dwords++] = cpu_to_le32( R128_CCE_PACKET2 );
 673                 }
 674
 675                 buf_priv->dispatched = 1;
 676
 677                 /* Fire off the indirect buffer */
 678                 BEGIN_RING( 3 );
 679
 680                 OUT_RING( CCE_PACKET0( R128_PM4_IW_INDOFF, 1 ) );
 681                 OUT_RING( offset );
 682                 OUT_RING( dwords );
 683
 684                 ADVANCE_RING();
 685         }
 686
 687         if ( buf_priv->discard ) {
 688                 buf_priv->age = dev_priv->sarea_priv->last_dispatch;
 689
 690                 /* Emit the indirect buffer age */
 691                 BEGIN_RING( 2 );
 692
 693                 OUT_RING( CCE_PACKET0( R128_LAST_DISPATCH_REG, 0 ) );
 694                 OUT_RING( buf_priv->age );
 695
 696                 ADVANCE_RING();
 697
 698                 buf->pending = 1;
 699                 buf->used = 0;
 700                 /* FIXME: Check dispatched field */
 701                 buf_priv->dispatched = 0;
 702         }
 703
 704         dev_priv->sarea_priv->last_dispatch++;
 705 }
 706
 707 static void r128_cce_dispatch_indices( drm_device_t *dev,
 708                                        drm_buf_t *buf,
 709                                        int start, int end,
 710                                        int count )
 711 {
 712         drm_r128_private_t *dev_priv = dev->dev_private;
 713         drm_r128_buf_priv_t *buf_priv = buf->dev_private;
 714         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
 715         int format = sarea_priv->vc_format;
 716         int offset = dev_priv->buffers->offset - dev_priv->cce_buffers_offset;
 717         int prim = buf_priv->prim;
 718         u32 *data;
 719         int dwords;
 720         int i = 0;
 721         RING_LOCALS;
 722         DRM_DEBUG( "indices: s=%d e=%d c=%d\n", start, end, count );
 723
 724         if ( 0 )
 725                 r128_print_dirty( "dispatch_indices", sarea_priv->dirty );
 726
 727         if ( start != end ) {
 728                 buf_priv->dispatched = 1;
 729
 730                 if ( sarea_priv->dirty & ~R128_UPLOAD_CLIPRECTS ) {
 731                         r128_emit_state( dev_priv );
 732                 }
 733
 734                 dwords = (end - start + 3) / sizeof(u32);
 735
 736                 data = (u32 *)((char *)dev_priv->buffers->handle
 737                                + buf->offset + start);
 738
 739                 data[0] = cpu_to_le32( CCE_PACKET3( R128_3D_RNDR_GEN_INDX_PRIM,
 740                                                     dwords-2 ) );
 741
 742                 data[1] = cpu_to_le32( offset );
 743                 data[2] = cpu_to_le32( R128_MAX_VB_VERTS );
 744                 data[3] = cpu_to_le32( format );
 745                 data[4] = cpu_to_le32( (prim | R128_CCE_VC_CNTL_PRIM_WALK_IND |
 746                                         (count << 16)) );
 747
 748                 if ( count & 0x1 ) {
 749 #ifdef __LITTLE_ENDIAN
 750                         data[dwords-1] &= 0x0000ffff;
 751 #else
 752                         data[dwords-1] &= 0xffff0000;
 753 #endif
 754                 }
 755
 756                 do {
 757                         /* Emit the next set of up to three cliprects */
 758                         if ( i < sarea_priv->nbox ) {
 759                                 r128_emit_clip_rects( dev_priv,
 760                                                       &sarea_priv->boxes[i],
 761                                                       sarea_priv->nbox - i );
 762                         }
 763
 764                         r128_cce_dispatch_indirect( dev, buf, start, end );
 765
 766                         i += 3;
 767                 } while ( i < sarea_priv->nbox );
 768         }
 769
 770         if ( buf_priv->discard ) {
 771                 buf_priv->age = dev_priv->sarea_priv->last_dispatch;
 772
 773                 /* Emit the vertex buffer age */
 774                 BEGIN_RING( 2 );
 775
 776                 OUT_RING( CCE_PACKET0( R128_LAST_DISPATCH_REG, 0 ) );
 777                 OUT_RING( buf_priv->age );
 778
 779                 ADVANCE_RING();
 780
 781                 buf->pending = 1;
 782                 /* FIXME: Check dispatched field */
 783                 buf_priv->dispatched = 0;
 784         }
 785
 786         dev_priv->sarea_priv->last_dispatch++;
 787
 788         sarea_priv->dirty &= ~R128_UPLOAD_CLIPRECTS;
 789         sarea_priv->nbox = 0;
 790 }
 791
 792 static int r128_cce_dispatch_blit( DRMFILE filp,
 793                                    drm_device_t *dev,
 794                                    drm_r128_blit_t *blit )
 795 {
 796         drm_r128_private_t *dev_priv = dev->dev_private;
 797         drm_device_dma_t *dma = dev->dma;
 798         drm_buf_t *buf;
 799         drm_r128_buf_priv_t *buf_priv;
 800         u32 *data;
 801         int dword_shift, dwords;
 802         RING_LOCALS;
 803         DRM_DEBUG( "\n" );
 804
 805         /* The compiler won't optimize away a division by a variable,
 806          * even if the only legal values are powers of two.  Thus, we'll
 807          * use a shift instead.
 808          */
 809         switch ( blit->format ) {
 810         case R128_DATATYPE_ARGB8888:
 811                 dword_shift = 0;
 812                 break;
 813         case R128_DATATYPE_ARGB1555:
 814         case R128_DATATYPE_RGB565:
 815         case R128_DATATYPE_ARGB4444:
 816         case R128_DATATYPE_YVYU422:
 817         case R128_DATATYPE_VYUY422:
 818                 dword_shift = 1;
 819                 break;
 820         case R128_DATATYPE_CI8:
 821         case R128_DATATYPE_RGB8:
 822                 dword_shift = 2;
 823                 break;
 824         default:
 825                 DRM_ERROR( "invalid blit format %d\n", blit->format );
 826                 return DRM_ERR(EINVAL);
 827         }
 828
 829         /* Flush the pixel cache, and mark the contents as Read Invalid.
 830          * This ensures no pixel data gets mixed up with the texture
 831          * data from the host data blit, otherwise part of the texture
 832          * image may be corrupted.
 833          */
 834         BEGIN_RING( 2 );
 835
 836         OUT_RING( CCE_PACKET0( R128_PC_GUI_CTLSTAT, 0 ) );
 837         OUT_RING( R128_PC_RI_GUI | R128_PC_FLUSH_GUI );
 838
 839         ADVANCE_RING();
 840
 841         /* Dispatch the indirect buffer.
 842          */
 843         buf = dma->buflist[blit->idx];
 844         buf_priv = buf->dev_private;
 845
 846         if ( buf->filp != filp ) {
 847                 DRM_ERROR( "process %d using buffer owned by %p\n",
 848                            DRM_CURRENTPID, buf->filp );
 849                 return DRM_ERR(EINVAL);
 850         }
 851         if ( buf->pending ) {
 852                 DRM_ERROR( "sending pending buffer %d\n", blit->idx );
 853                 return DRM_ERR(EINVAL);
 854         }
 855
 856         buf_priv->discard = 1;
 857
 858         dwords = (blit->width * blit->height) >> dword_shift;
 859
 860         data = (u32 *)((char *)dev_priv->buffers->handle + buf->offset);
 861
 862         data[0] = cpu_to_le32( CCE_PACKET3( R128_CNTL_HOSTDATA_BLT, dwords + 6 ) );
 863         data[1] = cpu_to_le32( (R128_GMC_DST_PITCH_OFFSET_CNTL |
 864                                 R128_GMC_BRUSH_NONE |
 865                                 (blit->format << 8) |
 866                                 R128_GMC_SRC_DATATYPE_COLOR |
 867                                 R128_ROP3_S |
 868                                 R128_DP_SRC_SOURCE_HOST_DATA |
 869                                 R128_GMC_CLR_CMP_CNTL_DIS |
 870                                 R128_GMC_AUX_CLIP_DIS |
 871                                 R128_GMC_WR_MSK_DIS) );
 872
 873         data[2] = cpu_to_le32( (blit->pitch << 21) | (blit->offset >> 5) );
 874         data[3] = cpu_to_le32( 0xffffffff );
 875         data[4] = cpu_to_le32( 0xffffffff );
 876         data[5] = cpu_to_le32( (blit->y << 16) | blit->x );
 877         data[6] = cpu_to_le32( (blit->height << 16) | blit->width );
 878         data[7] = cpu_to_le32( dwords );
 879
 880         buf->used = (dwords + 8) * sizeof(u32);
 881
 882         r128_cce_dispatch_indirect( dev, buf, 0, buf->used );
 883
 884         /* Flush the pixel cache after the blit completes.  This ensures
 885          * the texture data is written out to memory before rendering
 886          * continues.
 887          */
 888         BEGIN_RING( 2 );
 889
 890         OUT_RING( CCE_PACKET0( R128_PC_GUI_CTLSTAT, 0 ) );
 891         OUT_RING( R128_PC_FLUSH_GUI );
 892
 893         ADVANCE_RING();
 894
 895         return 0;
 896 }
 897
 898
 899 /* ================================================================
 900  * Tiled depth buffer management
 901  *
 902  * FIXME: These should all set the destination write mask for when we
 903  * have hardware stencil support.
 904  */
 905
 906 static int r128_cce_dispatch_write_span( drm_device_t *dev,
 907                                          drm_r128_depth_t *depth )
 908 {
 909         drm_r128_private_t *dev_priv = dev->dev_private;
 910         int count, x, y;
 911         u32 *buffer;
 912         u8 *mask;
 913         int i, buffer_size, mask_size;
 914         RING_LOCALS;
 915         DRM_DEBUG( "\n" );
 916
 917         count = depth->n;
 918         if (count > 4096 || count <= 0)
 919                 return -EMSGSIZE;
 920
 921         if ( DRM_COPY_FROM_USER( &x, depth->x, sizeof(x) ) ) {
 922                 return DRM_ERR(EFAULT);
 923         }
 924         if ( DRM_COPY_FROM_USER( &y, depth->y, sizeof(y) ) ) {
 925                 return DRM_ERR(EFAULT);
 926         }
 927
 928         buffer_size = depth->n * sizeof(u32);
 929         buffer = DRM_MALLOC( buffer_size );
 930         if ( buffer == NULL )
 931                 return DRM_ERR(ENOMEM);
 932         if ( DRM_COPY_FROM_USER( buffer, depth->buffer, buffer_size ) ) {
 933                 DRM_FREE( buffer, buffer_size);
 934                 return DRM_ERR(EFAULT);
 935         }
 936
 937         mask_size = depth->n * sizeof(u8);
 938         if ( depth->mask ) {
 939                 mask = DRM_MALLOC( mask_size );
 940                 if ( mask == NULL ) {
 941                         DRM_FREE( buffer, buffer_size );
 942                         return DRM_ERR(ENOMEM);
 943                 }
 944                 if ( DRM_COPY_FROM_USER( mask, depth->mask, mask_size ) ) {
 945                         DRM_FREE( buffer, buffer_size );
 946                         DRM_FREE( mask, mask_size );
 947                         return DRM_ERR(EFAULT);
 948                 }
 949
 950                 for ( i = 0 ; i < count ; i++, x++ ) {
 951                         if ( mask[i] ) {
 952                                 BEGIN_RING( 6 );
 953
 954                                 OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 955                                 OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 956                                           R128_GMC_BRUSH_SOLID_COLOR |
 957                                           (dev_priv->depth_fmt << 8) |
 958                                           R128_GMC_SRC_DATATYPE_COLOR |
 959                                           R128_ROP3_P |
 960                                           R128_GMC_CLR_CMP_CNTL_DIS |
 961                                           R128_GMC_WR_MSK_DIS );
 962
 963                                 OUT_RING( dev_priv->depth_pitch_offset_c );
 964                                 OUT_RING( buffer[i] );
 965
 966                                 OUT_RING( (x << 16) | y );
 967                                 OUT_RING( (1 << 16) | 1 );
 968
 969                                 ADVANCE_RING();
 970                         }
 971                 }
 972
 973                 DRM_FREE( mask, mask_size );
 974         } else {
 975                 for ( i = 0 ; i < count ; i++, x++ ) {
 976                         BEGIN_RING( 6 );
 977
 978                         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
 979                         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
 980                                   R128_GMC_BRUSH_SOLID_COLOR |
 981                                   (dev_priv->depth_fmt << 8) |
 982                                   R128_GMC_SRC_DATATYPE_COLOR |
 983                                   R128_ROP3_P |
 984                                   R128_GMC_CLR_CMP_CNTL_DIS |
 985                                   R128_GMC_WR_MSK_DIS );
 986
 987                         OUT_RING( dev_priv->depth_pitch_offset_c );
 988                         OUT_RING( buffer[i] );
 989
 990                         OUT_RING( (x << 16) | y );
 991                         OUT_RING( (1 << 16) | 1 );
 992
 993                         ADVANCE_RING();
 994                 }
 995         }
 996
 997         DRM_FREE( buffer, buffer_size );
 998
 999         return 0;
1000 }
1001
1002 static int r128_cce_dispatch_write_pixels( drm_device_t *dev,
1003                                            drm_r128_depth_t *depth )
1004 {
1005         drm_r128_private_t *dev_priv = dev->dev_private;
1006         int count, *x, *y;
1007         u32 *buffer;
1008         u8 *mask;
1009         int i, xbuf_size, ybuf_size, buffer_size, mask_size;
1010         RING_LOCALS;
1011         DRM_DEBUG( "\n" );
1012
1013         count = depth->n;
1014         if (count > 4096 || count <= 0)
1015                 return -EMSGSIZE;
1016
1017         xbuf_size = count * sizeof(*x);
1018         ybuf_size = count * sizeof(*y);
1019         x = DRM_MALLOC( xbuf_size );
1020         if ( x == NULL ) {
1021                 return DRM_ERR(ENOMEM);
1022         }
1023         y = DRM_MALLOC( ybuf_size );
1024         if ( y == NULL ) {
1025                 DRM_FREE( x, xbuf_size );
1026                 return DRM_ERR(ENOMEM);
1027         }
1028         if ( DRM_COPY_FROM_USER( x, depth->x, xbuf_size ) ) {
1029                 DRM_FREE( x, xbuf_size );
1030                 DRM_FREE( y, ybuf_size );
1031                 return DRM_ERR(EFAULT);
1032         }
1033         if ( DRM_COPY_FROM_USER( y, depth->y, xbuf_size ) ) {
1034                 DRM_FREE( x, xbuf_size );
1035                 DRM_FREE( y, ybuf_size );
1036                 return DRM_ERR(EFAULT);
1037         }
1038
1039         buffer_size = depth->n * sizeof(u32);
1040         buffer = DRM_MALLOC( buffer_size );
1041         if ( buffer == NULL ) {
1042                 DRM_FREE( x, xbuf_size );
1043                 DRM_FREE( y, ybuf_size );
1044                 return DRM_ERR(ENOMEM);
1045         }
1046         if ( DRM_COPY_FROM_USER( buffer, depth->buffer, buffer_size ) ) {
1047                 DRM_FREE( x, xbuf_size );
1048                 DRM_FREE( y, ybuf_size );
1049                 DRM_FREE( buffer, buffer_size );
1050                 return DRM_ERR(EFAULT);
1051         }
1052
1053         if ( depth->mask ) {
1054                 mask_size = depth->n * sizeof(u8);
1055                 mask = DRM_MALLOC( mask_size );
1056                 if ( mask == NULL ) {
1057                         DRM_FREE( x, xbuf_size );
1058                         DRM_FREE( y, ybuf_size );
1059                         DRM_FREE( buffer, buffer_size );
1060                         return DRM_ERR(ENOMEM);
1061                 }
1062                 if ( DRM_COPY_FROM_USER( mask, depth->mask, mask_size ) ) {
1063                         DRM_FREE( x, xbuf_size );
1064                         DRM_FREE( y, ybuf_size );
1065                         DRM_FREE( buffer, buffer_size );
1066                         DRM_FREE( mask, mask_size );
1067                         return DRM_ERR(EFAULT);
1068                 }
1069
1070                 for ( i = 0 ; i < count ; i++ ) {
1071                         if ( mask[i] ) {
1072                                 BEGIN_RING( 6 );
1073
1074                                 OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
1075                                 OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
1076                                           R128_GMC_BRUSH_SOLID_COLOR |
1077                                           (dev_priv->depth_fmt << 8) |
1078                                           R128_GMC_SRC_DATATYPE_COLOR |
1079                                           R128_ROP3_P |
1080                                           R128_GMC_CLR_CMP_CNTL_DIS |
1081                                           R128_GMC_WR_MSK_DIS );
1082
1083                                 OUT_RING( dev_priv->depth_pitch_offset_c );
1084                                 OUT_RING( buffer[i] );
1085
1086                                 OUT_RING( (x[i] << 16) | y[i] );
1087                                 OUT_RING( (1 << 16) | 1 );
1088
1089                                 ADVANCE_RING();
1090                         }
1091                 }
1092
1093                 DRM_FREE( mask, mask_size );
1094         } else {
1095                 for ( i = 0 ; i < count ; i++ ) {
1096                         BEGIN_RING( 6 );
1097
1098                         OUT_RING( CCE_PACKET3( R128_CNTL_PAINT_MULTI, 4 ) );
1099                         OUT_RING( R128_GMC_DST_PITCH_OFFSET_CNTL |
1100                                   R128_GMC_BRUSH_SOLID_COLOR |
1101                                   (dev_priv->depth_fmt << 8) |
1102                                   R128_GMC_SRC_DATATYPE_COLOR |
1103                                   R128_ROP3_P |
1104                                   R128_GMC_CLR_CMP_CNTL_DIS |
1105                                   R128_GMC_WR_MSK_DIS );
1106
1107                         OUT_RING( dev_priv->depth_pitch_offset_c );
1108                         OUT_RING( buffer[i] );
1109
1110                         OUT_RING( (x[i] << 16) | y[i] );
1111                         OUT_RING( (1 << 16) | 1 );
1112
1113                         ADVANCE_RING();
1114                 }
1115         }
1116
1117         DRM_FREE( x, xbuf_size );
1118         DRM_FREE( y, ybuf_size );
1119         DRM_FREE( buffer, buffer_size );
1120
1121         return 0;
1122 }
1123
1124 static int r128_cce_dispatch_read_span( drm_device_t *dev,
1125                                         drm_r128_depth_t *depth )
1126 {
1127         drm_r128_private_t *dev_priv = dev->dev_private;
1128         int count, x, y;
1129         RING_LOCALS;
1130         DRM_DEBUG( "\n" );
1131
1132         count = depth->n;
1133         if (count > 4096 || count <= 0)
1134                 return -EMSGSIZE;
1135
1136         if ( DRM_COPY_FROM_USER( &x, depth->x, sizeof(x) ) ) {
1137                 return DRM_ERR(EFAULT);
1138         }
1139         if ( DRM_COPY_FROM_USER( &y, depth->y, sizeof(y) ) ) {
1140                 return DRM_ERR(EFAULT);
1141         }
1142
1143         BEGIN_RING( 7 );
1144
1145         OUT_RING( CCE_PACKET3( R128_CNTL_BITBLT_MULTI, 5 ) );
1146         OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL |
1147                   R128_GMC_DST_PITCH_OFFSET_CNTL |
1148                   R128_GMC_BRUSH_NONE |
1149                   (dev_priv->depth_fmt << 8) |
1150                   R128_GMC_SRC_DATATYPE_COLOR |
1151                   R128_ROP3_S |
1152                   R128_DP_SRC_SOURCE_MEMORY |
1153                   R128_GMC_CLR_CMP_CNTL_DIS |
1154                   R128_GMC_WR_MSK_DIS );
1155
1156         OUT_RING( dev_priv->depth_pitch_offset_c );
1157         OUT_RING( dev_priv->span_pitch_offset_c );
1158
1159         OUT_RING( (x << 16) | y );
1160         OUT_RING( (0 << 16) | 0 );
1161         OUT_RING( (count << 16) | 1 );
1162
1163         ADVANCE_RING();
1164
1165         return 0;
1166 }
1167
1168 static int r128_cce_dispatch_read_pixels( drm_device_t *dev,
1169                                           drm_r128_depth_t *depth )
1170 {
1171         drm_r128_private_t *dev_priv = dev->dev_private;
1172         int count, *x, *y;
1173         int i, xbuf_size, ybuf_size;
1174         RING_LOCALS;
1175         DRM_DEBUG( "%s\n", __FUNCTION__ );
1176
1177         count = depth->n;
1178         if (count > 4096 || count <= 0)
1179                 return -EMSGSIZE;
1180
1181         if ( count > dev_priv->depth_pitch ) {
1182                 count = dev_priv->depth_pitch;
1183         }
1184
1185         xbuf_size = count * sizeof(*x);
1186         ybuf_size = count * sizeof(*y);
1187         x = DRM_MALLOC( xbuf_size );
1188         if ( x == NULL ) {
1189                 return DRM_ERR(ENOMEM);
1190         }
1191         y = DRM_MALLOC( ybuf_size );
1192         if ( y == NULL ) {
1193                 DRM_FREE( x, xbuf_size );
1194                 return DRM_ERR(ENOMEM);
1195         }
1196         if ( DRM_COPY_FROM_USER( x, depth->x, xbuf_size ) ) {
1197                 DRM_FREE( x, xbuf_size );
1198                 DRM_FREE( y, ybuf_size );
1199                 return DRM_ERR(EFAULT);
1200         }
1201         if ( DRM_COPY_FROM_USER( y, depth->y, ybuf_size ) ) {
1202                 DRM_FREE( x, xbuf_size );
1203                 DRM_FREE( y, ybuf_size );
1204                 return DRM_ERR(EFAULT);
1205         }
1206
1207         for ( i = 0 ; i < count ; i++ ) {
1208                 BEGIN_RING( 7 );
1209
1210                 OUT_RING( CCE_PACKET3( R128_CNTL_BITBLT_MULTI, 5 ) );
1211                 OUT_RING( R128_GMC_SRC_PITCH_OFFSET_CNTL |
1212                           R128_GMC_DST_PITCH_OFFSET_CNTL |
1213                           R128_GMC_BRUSH_NONE |
1214                           (dev_priv->depth_fmt << 8) |
1215                           R128_GMC_SRC_DATATYPE_COLOR |
1216                           R128_ROP3_S |
1217                           R128_DP_SRC_SOURCE_MEMORY |
1218                           R128_GMC_CLR_CMP_CNTL_DIS |
1219                           R128_GMC_WR_MSK_DIS );
1220
1221                 OUT_RING( dev_priv->depth_pitch_offset_c );
1222                 OUT_RING( dev_priv->span_pitch_offset_c );
1223
1224                 OUT_RING( (x[i] << 16) | y[i] );
1225                 OUT_RING( (i << 16) | 0 );
1226                 OUT_RING( (1 << 16) | 1 );
1227
1228                 ADVANCE_RING();
1229         }
1230
1231         DRM_FREE( x, xbuf_size );
1232         DRM_FREE( y, ybuf_size );
1233
1234         return 0;
1235 }
1236
1237
1238 /* ================================================================
1239  * Polygon stipple
1240  */
1241
1242 static void r128_cce_dispatch_stipple( drm_device_t *dev, u32 *stipple )
1243 {
1244         drm_r128_private_t *dev_priv = dev->dev_private;
1245         int i;
1246         RING_LOCALS;
1247         DRM_DEBUG( "%s\n", __FUNCTION__ );
1248
1249         BEGIN_RING( 33 );
1250
1251         OUT_RING( CCE_PACKET0( R128_BRUSH_DATA0, 31 ) );
1252         for ( i = 0 ; i < 32 ; i++ ) {
1253                 OUT_RING( stipple[i] );
1254         }
1255
1256         ADVANCE_RING();
1257 }
1258
1259
1260 /* ================================================================
1261  * IOCTL functions
1262  */
1263
1264 int r128_cce_clear( DRM_IOCTL_ARGS )
1265 {
1266         DRM_DEVICE;
1267         drm_r128_private_t *dev_priv = dev->dev_private;
1268         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
1269         drm_r128_clear_t clear;
1270         DRM_DEBUG( "\n" );
1271
1272         LOCK_TEST_WITH_RETURN( dev, filp );
1273
1274         DRM_COPY_FROM_USER_IOCTL( clear, (drm_r128_clear_t *) data,
1275                              sizeof(clear) );
1276
1277         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1278
1279         if ( sarea_priv->nbox > R128_NR_SAREA_CLIPRECTS )
1280                 sarea_priv->nbox = R128_NR_SAREA_CLIPRECTS;
1281
1282         r128_cce_dispatch_clear( dev, &clear );
1283         COMMIT_RING();
1284
1285         /* Make sure we restore the 3D state next time.
1286          */
1287         dev_priv->sarea_priv->dirty |= R128_UPLOAD_CONTEXT | R128_UPLOAD_MASKS;
1288
1289         return 0;
1290 }
1291
1292 static int r128_do_init_pageflip( drm_device_t *dev )
1293 {
1294         drm_r128_private_t *dev_priv = dev->dev_private;
1295         DRM_DEBUG( "\n" );
1296
1297         dev_priv->crtc_offset =      R128_READ( R128_CRTC_OFFSET );
1298         dev_priv->crtc_offset_cntl = R128_READ( R128_CRTC_OFFSET_CNTL );
1299
1300         R128_WRITE( R128_CRTC_OFFSET, dev_priv->front_offset );
1301         R128_WRITE( R128_CRTC_OFFSET_CNTL,
1302                     dev_priv->crtc_offset_cntl | R128_CRTC_OFFSET_FLIP_CNTL );
1303
1304         dev_priv->page_flipping = 1;
1305         dev_priv->current_page = 0;
1306         dev_priv->sarea_priv->pfCurrentPage = dev_priv->current_page;
1307
1308         return 0;
1309 }
1310
1311 int r128_do_cleanup_pageflip( drm_device_t *dev )
1312 {
1313         drm_r128_private_t *dev_priv = dev->dev_private;
1314         DRM_DEBUG( "\n" );
1315
1316         R128_WRITE( R128_CRTC_OFFSET,      dev_priv->crtc_offset );
1317         R128_WRITE( R128_CRTC_OFFSET_CNTL, dev_priv->crtc_offset_cntl );
1318
1319         if (dev_priv->current_page != 0) {
1320                 r128_cce_dispatch_flip( dev );
1321                 COMMIT_RING();
1322         }
1323
1324         dev_priv->page_flipping = 0;
1325         return 0;
1326 }
1327
1328 /* Swapping and flipping are different operations, need different ioctls.
1329  * They can & should be intermixed to support multiple 3d windows.
1330  */
1331
1332 int r128_cce_flip( DRM_IOCTL_ARGS )
1333 {
1334         DRM_DEVICE;
1335         drm_r128_private_t *dev_priv = dev->dev_private;
1336         DRM_DEBUG( "%s\n", __FUNCTION__ );
1337
1338         LOCK_TEST_WITH_RETURN( dev, filp );
1339
1340         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1341
1342         if (!dev_priv->page_flipping)
1343                 r128_do_init_pageflip( dev );
1344
1345         r128_cce_dispatch_flip( dev );
1346
1347         COMMIT_RING();
1348         return 0;
1349 }
1350
1351 int r128_cce_swap( DRM_IOCTL_ARGS )
1352 {
1353         DRM_DEVICE;
1354         drm_r128_private_t *dev_priv = dev->dev_private;
1355         drm_r128_sarea_t *sarea_priv = dev_priv->sarea_priv;
1356         DRM_DEBUG( "%s\n", __FUNCTION__ );
1357
1358         LOCK_TEST_WITH_RETURN( dev, filp );
1359
1360         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1361
1362         if ( sarea_priv->nbox > R128_NR_SAREA_CLIPRECTS )
1363                 sarea_priv->nbox = R128_NR_SAREA_CLIPRECTS;
1364
1365         r128_cce_dispatch_swap( dev );
1366         dev_priv->sarea_priv->dirty |= (R128_UPLOAD_CONTEXT |
1367                                         R128_UPLOAD_MASKS);
1368
1369         COMMIT_RING();
1370         return 0;
1371 }
1372
1373 int r128_cce_vertex( DRM_IOCTL_ARGS )
1374 {
1375         DRM_DEVICE;
1376         drm_r128_private_t *dev_priv = dev->dev_private;
1377         drm_device_dma_t *dma = dev->dma;
1378         drm_buf_t *buf;
1379         drm_r128_buf_priv_t *buf_priv;
1380         drm_r128_vertex_t vertex;
1381
1382         LOCK_TEST_WITH_RETURN( dev, filp );
1383
1384         if ( !dev_priv ) {
1385                 DRM_ERROR( "%s called with no initialization\n", __FUNCTION__ );
1386                 return DRM_ERR(EINVAL);
1387         }
1388
1389         DRM_COPY_FROM_USER_IOCTL( vertex, (drm_r128_vertex_t *) data,
1390                              sizeof(vertex) );
1391
1392         DRM_DEBUG( "pid=%d index=%d count=%d discard=%d\n",
1393                    DRM_CURRENTPID,
1394                    vertex.idx, vertex.count, vertex.discard );
1395
1396         if ( vertex.idx < 0 || vertex.idx >= dma->buf_count ) {
1397                 DRM_ERROR( "buffer index %d (of %d max)\n",
1398                            vertex.idx, dma->buf_count - 1 );
1399                 return DRM_ERR(EINVAL);
1400         }
1401         if ( vertex.prim < 0 ||
1402              vertex.prim > R128_CCE_VC_CNTL_PRIM_TYPE_TRI_TYPE2 ) {
1403                 DRM_ERROR( "buffer prim %d\n", vertex.prim );
1404                 return DRM_ERR(EINVAL);
1405         }
1406
1407         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1408         VB_AGE_TEST_WITH_RETURN( dev_priv );
1409
1410         buf = dma->buflist[vertex.idx];
1411         buf_priv = buf->dev_private;
1412
1413         if ( buf->filp != filp ) {
1414                 DRM_ERROR( "process %d using buffer owned by %p\n",
1415                            DRM_CURRENTPID, buf->filp );
1416                 return DRM_ERR(EINVAL);
1417         }
1418         if ( buf->pending ) {
1419                 DRM_ERROR( "sending pending buffer %d\n", vertex.idx );
1420                 return DRM_ERR(EINVAL);
1421         }
1422
1423         buf->used = vertex.count;
1424         buf_priv->prim = vertex.prim;
1425         buf_priv->discard = vertex.discard;
1426
1427         r128_cce_dispatch_vertex( dev, buf );
1428
1429         COMMIT_RING();
1430         return 0;
1431 }
1432
1433 int r128_cce_indices( DRM_IOCTL_ARGS )
1434 {
1435         DRM_DEVICE;
1436         drm_r128_private_t *dev_priv = dev->dev_private;
1437         drm_device_dma_t *dma = dev->dma;
1438         drm_buf_t *buf;
1439         drm_r128_buf_priv_t *buf_priv;
1440         drm_r128_indices_t elts;
1441         int count;
1442
1443         LOCK_TEST_WITH_RETURN( dev, filp );
1444
1445         if ( !dev_priv ) {
1446                 DRM_ERROR( "%s called with no initialization\n", __FUNCTION__ );
1447                 return DRM_ERR(EINVAL);
1448         }
1449
1450         DRM_COPY_FROM_USER_IOCTL( elts, (drm_r128_indices_t *) data,
1451                              sizeof(elts) );
1452
1453         DRM_DEBUG( "pid=%d buf=%d s=%d e=%d d=%d\n", DRM_CURRENTPID,
1454                    elts.idx, elts.start, elts.end, elts.discard );
1455
1456         if ( elts.idx < 0 || elts.idx >= dma->buf_count ) {
1457                 DRM_ERROR( "buffer index %d (of %d max)\n",
1458                            elts.idx, dma->buf_count - 1 );
1459                 return DRM_ERR(EINVAL);
1460         }
1461         if ( elts.prim < 0 ||
1462              elts.prim > R128_CCE_VC_CNTL_PRIM_TYPE_TRI_TYPE2 ) {
1463                 DRM_ERROR( "buffer prim %d\n", elts.prim );
1464                 return DRM_ERR(EINVAL);
1465         }
1466
1467         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1468         VB_AGE_TEST_WITH_RETURN( dev_priv );
1469
1470         buf = dma->buflist[elts.idx];
1471         buf_priv = buf->dev_private;
1472
1473         if ( buf->filp != filp ) {
1474                 DRM_ERROR( "process %d using buffer owned by %p\n",
1475                            DRM_CURRENTPID, buf->filp );
1476                 return DRM_ERR(EINVAL);
1477         }
1478         if ( buf->pending ) {
1479                 DRM_ERROR( "sending pending buffer %d\n", elts.idx );
1480                 return DRM_ERR(EINVAL);
1481         }
1482
1483         count = (elts.end - elts.start) / sizeof(u16);
1484         elts.start -= R128_INDEX_PRIM_OFFSET;
1485
1486         if ( elts.start & 0x7 ) {
1487                 DRM_ERROR( "misaligned buffer 0x%x\n", elts.start );
1488                 return DRM_ERR(EINVAL);
1489         }
1490         if ( elts.start < buf->used ) {
1491                 DRM_ERROR( "no header 0x%x - 0x%x\n", elts.start, buf->used );
1492                 return DRM_ERR(EINVAL);
1493         }
1494
1495         buf->used = elts.end;
1496         buf_priv->prim = elts.prim;
1497         buf_priv->discard = elts.discard;
1498
1499         r128_cce_dispatch_indices( dev, buf, elts.start, elts.end, count );
1500
1501         COMMIT_RING();
1502         return 0;
1503 }
1504
1505 int r128_cce_blit( DRM_IOCTL_ARGS )
1506 {
1507         DRM_DEVICE;
1508         drm_device_dma_t *dma = dev->dma;
1509         drm_r128_private_t *dev_priv = dev->dev_private;
1510         drm_r128_blit_t blit;
1511         int ret;
1512
1513         LOCK_TEST_WITH_RETURN( dev, filp );
1514
1515         DRM_COPY_FROM_USER_IOCTL( blit, (drm_r128_blit_t *) data,
1516                              sizeof(blit) );
1517
1518         DRM_DEBUG( "pid=%d index=%d\n", DRM_CURRENTPID, blit.idx );
1519
1520         if ( blit.idx < 0 || blit.idx >= dma->buf_count ) {
1521                 DRM_ERROR( "buffer index %d (of %d max)\n",
1522                            blit.idx, dma->buf_count - 1 );
1523                 return DRM_ERR(EINVAL);
1524         }
1525
1526         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1527         VB_AGE_TEST_WITH_RETURN( dev_priv );
1528
1529         ret = r128_cce_dispatch_blit( filp, dev, &blit );
1530
1531         COMMIT_RING();
1532         return ret;
1533 }
1534
1535 int r128_cce_depth( DRM_IOCTL_ARGS )
1536 {
1537         DRM_DEVICE;
1538         drm_r128_private_t *dev_priv = dev->dev_private;
1539         drm_r128_depth_t depth;
1540         int ret;
1541
1542         LOCK_TEST_WITH_RETURN( dev, filp );
1543
1544         DRM_COPY_FROM_USER_IOCTL( depth, (drm_r128_depth_t *) data,
1545                              sizeof(depth) );
1546
1547         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1548
1549         ret = DRM_ERR(EINVAL);
1550         switch ( depth.func ) {
1551         case R128_WRITE_SPAN:
1552                 ret = r128_cce_dispatch_write_span( dev, &depth );
1553         case R128_WRITE_PIXELS:
1554                 ret = r128_cce_dispatch_write_pixels( dev, &depth );
1555         case R128_READ_SPAN:
1556                 ret = r128_cce_dispatch_read_span( dev, &depth );
1557         case R128_READ_PIXELS:
1558                 ret = r128_cce_dispatch_read_pixels( dev, &depth );
1559         }
1560
1561         COMMIT_RING();
1562         return ret;
1563 }
1564
1565 int r128_cce_stipple( DRM_IOCTL_ARGS )
1566 {
1567         DRM_DEVICE;
1568         drm_r128_private_t *dev_priv = dev->dev_private;
1569         drm_r128_stipple_t stipple;
1570         u32 mask[32];
1571
1572         LOCK_TEST_WITH_RETURN( dev, filp );
1573
1574         DRM_COPY_FROM_USER_IOCTL( stipple, (drm_r128_stipple_t *) data,
1575                              sizeof(stipple) );
1576
1577         if ( DRM_COPY_FROM_USER( &mask, stipple.mask,
1578                              32 * sizeof(u32) ) )
1579                 return DRM_ERR( EFAULT );
1580
1581         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1582
1583         r128_cce_dispatch_stipple( dev, mask );
1584
1585         COMMIT_RING();
1586         return 0;
1587 }
1588
1589 int r128_cce_indirect( DRM_IOCTL_ARGS )
1590 {
1591         DRM_DEVICE;
1592         drm_r128_private_t *dev_priv = dev->dev_private;
1593         drm_device_dma_t *dma = dev->dma;
1594         drm_buf_t *buf;
1595         drm_r128_buf_priv_t *buf_priv;
1596         drm_r128_indirect_t indirect;
1597 #if 0
1598         RING_LOCALS;
1599 #endif
1600
1601         LOCK_TEST_WITH_RETURN( dev, filp );
1602
1603         if ( !dev_priv ) {
1604                 DRM_ERROR( "%s called with no initialization\n", __FUNCTION__ );
1605                 return DRM_ERR(EINVAL);
1606         }
1607
1608         DRM_COPY_FROM_USER_IOCTL( indirect, (drm_r128_indirect_t *) data,
1609                              sizeof(indirect) );
1610
1611         DRM_DEBUG( "indirect: idx=%d s=%d e=%d d=%d\n",
1612                    indirect.idx, indirect.start,
1613                    indirect.end, indirect.discard );
1614
1615         if ( indirect.idx < 0 || indirect.idx >= dma->buf_count ) {
1616                 DRM_ERROR( "buffer index %d (of %d max)\n",
1617                            indirect.idx, dma->buf_count - 1 );
1618                 return DRM_ERR(EINVAL);
1619         }
1620
1621         buf = dma->buflist[indirect.idx];
1622         buf_priv = buf->dev_private;
1623
1624         if ( buf->filp != filp ) {
1625                 DRM_ERROR( "process %d using buffer owned by %p\n",
1626                            DRM_CURRENTPID, buf->filp );
1627                 return DRM_ERR(EINVAL);
1628         }
1629         if ( buf->pending ) {
1630                 DRM_ERROR( "sending pending buffer %d\n", indirect.idx );
1631                 return DRM_ERR(EINVAL);
1632         }
1633
1634         if ( indirect.start < buf->used ) {
1635                 DRM_ERROR( "reusing indirect: start=0x%x actual=0x%x\n",
1636                            indirect.start, buf->used );
1637                 return DRM_ERR(EINVAL);
1638         }
1639
1640         RING_SPACE_TEST_WITH_RETURN( dev_priv );
1641         VB_AGE_TEST_WITH_RETURN( dev_priv );
1642
1643         buf->used = indirect.end;
1644         buf_priv->discard = indirect.discard;
1645
1646 #if 0
1647         /* Wait for the 3D stream to idle before the indirect buffer
1648          * containing 2D acceleration commands is processed.
1649          */
1650         BEGIN_RING( 2 );
1651         RADEON_WAIT_UNTIL_3D_IDLE();
1652         ADVANCE_RING();
1653 #endif
1654
1655         /* Dispatch the indirect buffer full of commands from the
1656          * X server.  This is insecure and is thus only available to
1657          * privileged clients.
1658          */
1659         r128_cce_dispatch_indirect( dev, buf, indirect.start, indirect.end );
1660
1661         COMMIT_RING();
1662         return 0;
1663 }
1664
1665 int r128_getparam( DRM_IOCTL_ARGS )
1666 {
1667         DRM_DEVICE;
1668         drm_r128_private_t *dev_priv = dev->dev_private;
1669         drm_r128_getparam_t param;
1670         int value;
1671
1672         if ( !dev_priv ) {
1673                 DRM_ERROR( "%s called with no initialization\n", __FUNCTION__ );
1674                 return DRM_ERR(EINVAL);
1675         }
1676
1677         DRM_COPY_FROM_USER_IOCTL( param, (drm_r128_getparam_t *)data,
1678                              sizeof(param) );
1679
1680         DRM_DEBUG( "pid=%d\n", DRM_CURRENTPID );
1681
1682         switch( param.param ) {
1683         case R128_PARAM_IRQ_NR:
1684                 value = dev->irq;
1685                 break;
1686         default:
1687                 return DRM_ERR(EINVAL);
1688         }
1689
1690         if ( DRM_COPY_TO_USER( param.value, &value, sizeof(int) ) ) {
1691                 DRM_ERROR( "copy_to_user\n" );
1692                 return DRM_ERR(EFAULT);
1693         }
1694
1695         return 0;
1696 }