drivers/gpu/drm/via/via_dma.c

   1 /* via_dma.c -- DMA support for the VIA Unichrome/Pro
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Copyright 2004 Digeo, Inc., Palo Alto, CA, U.S.A.
   7  * All Rights Reserved.
   8  *
   9  * Copyright 2004 The Unichrome project.
  10  * All Rights Reserved.
  11  *
  12  * Permission is hereby granted, free of charge, to any person obtaining a
  13  * copy of this software and associated documentation files (the "Software"),
  14  * to deal in the Software without restriction, including without limitation
  15  * the rights to use, copy, modify, merge, publish, distribute, sub license,
  16  * and/or sell copies of the Software, and to permit persons to whom the
  17  * Software is furnished to do so, subject to the following conditions:
  18  *
  19  * The above copyright notice and this permission notice (including the
  20  * next paragraph) shall be included in all copies or substantial portions
  21  * of the Software.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  24  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  25  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  26  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  27  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  28  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  29  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  30  *
  31  * Authors:
  32  *    Tungsten Graphics,
  33  *    Erdi Chen,
  34  *    Thomas Hellstrom.
  35  */
  36
  37 #include "drmP.h"
  38 #include "drm.h"
  39 #include "via_drm.h"
  40 #include "via_drv.h"
  41 #include "via_3d_reg.h"
  42
  43 #define CMDBUF_ALIGNMENT_SIZE   (0x100)
  44 #define CMDBUF_ALIGNMENT_MASK   (0x0ff)
  45
  46 /* defines for VIA 3D registers */
  47 #define VIA_REG_STATUS          0x400
  48 #define VIA_REG_TRANSET         0x43C
  49 #define VIA_REG_TRANSPACE       0x440
  50
  51 /* VIA_REG_STATUS(0x400): Engine Status */
  52 #define VIA_CMD_RGTR_BUSY       0x00000080      /* Command Regulator is busy */
  53 #define VIA_2D_ENG_BUSY         0x00000001      /* 2D Engine is busy */
  54 #define VIA_3D_ENG_BUSY         0x00000002      /* 3D Engine is busy */
  55 #define VIA_VR_QUEUE_BUSY       0x00020000      /* Virtual Queue is busy */
  56
  57 #define SetReg2DAGP(nReg, nData) {                              \
  58         *((uint32_t *)(vb)) = ((nReg) >> 2) | HALCYON_HEADER1;  \
  59         *((uint32_t *)(vb) + 1) = (nData);                      \
  60         vb = ((uint32_t *)vb) + 2;                              \
  61         dev_priv->dma_low +=8;                                  \
  62 }
  63
  64 #define via_flush_write_combine() DRM_MEMORYBARRIER()
  65
  66 #define VIA_OUT_RING_QW(w1,w2)                  \
  67         *vb++ = (w1);                           \
  68         *vb++ = (w2);                           \
  69         dev_priv->dma_low += 8;
  70
  71 static void via_cmdbuf_start(drm_via_private_t * dev_priv);
  72 static void via_cmdbuf_pause(drm_via_private_t * dev_priv);
  73 static void via_cmdbuf_reset(drm_via_private_t * dev_priv);
  74 static void via_cmdbuf_rewind(drm_via_private_t * dev_priv);
  75 static int via_wait_idle(drm_via_private_t * dev_priv);
  76 static void via_pad_cache(drm_via_private_t * dev_priv, int qwords);
  77
  78 /*
  79  * Free space in command buffer.
  80  */
  81
  82 static uint32_t via_cmdbuf_space(drm_via_private_t * dev_priv)
  83 {
  84         uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
  85         uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
  86
  87         return ((hw_addr <= dev_priv->dma_low) ?
  88                 (dev_priv->dma_high + hw_addr - dev_priv->dma_low) :
  89                 (hw_addr - dev_priv->dma_low));
  90 }
  91
  92 /*
  93  * How much does the command regulator lag behind?
  94  */
  95
  96 static uint32_t via_cmdbuf_lag(drm_via_private_t * dev_priv)
  97 {
  98         uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
  99         uint32_t hw_addr = *(dev_priv->hw_addr_ptr) - agp_base;
 100
 101         return ((hw_addr <= dev_priv->dma_low) ?
 102                 (dev_priv->dma_low - hw_addr) :
 103                 (dev_priv->dma_wrap + dev_priv->dma_low - hw_addr));
 104 }
 105
 106 /*
 107  * Check that the given size fits in the buffer, otherwise wait.
 108  */
 109
 110 static inline int
 111 via_cmdbuf_wait(drm_via_private_t * dev_priv, unsigned int size)
 112 {
 113         uint32_t agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
 114         uint32_t cur_addr, hw_addr, next_addr;
 115         volatile uint32_t *hw_addr_ptr;
 116         uint32_t count;
 117         hw_addr_ptr = dev_priv->hw_addr_ptr;
 118         cur_addr = dev_priv->dma_low;
 119         next_addr = cur_addr + size + 512 * 1024;
 120         count = 1000000;
 121         do {
 122                 hw_addr = *hw_addr_ptr - agp_base;
 123                 if (count-- == 0) {
 124                         DRM_ERROR
 125                             ("via_cmdbuf_wait timed out hw %x cur_addr %x next_addr %x\n",
 126                              hw_addr, cur_addr, next_addr);
 127                         return -1;
 128                 }
 129                 if  ((cur_addr < hw_addr) && (next_addr >= hw_addr))
 130                         msleep(1);
 131         } while ((cur_addr < hw_addr) && (next_addr >= hw_addr));
 132         return 0;
 133 }
 134
 135 /*
 136  * Checks whether buffer head has reach the end. Rewind the ring buffer
 137  * when necessary.
 138  *
 139  * Returns virtual pointer to ring buffer.
 140  */
 141
 142 static inline uint32_t *via_check_dma(drm_via_private_t * dev_priv,
 143                                       unsigned int size)
 144 {
 145         if ((dev_priv->dma_low + size + 4 * CMDBUF_ALIGNMENT_SIZE) >
 146             dev_priv->dma_high) {
 147                 via_cmdbuf_rewind(dev_priv);
 148         }
 149         if (via_cmdbuf_wait(dev_priv, size) != 0) {
 150                 return NULL;
 151         }
 152
 153         return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
 154 }
 155
 156 int via_dma_cleanup(struct drm_device * dev)
 157 {
 158         if (dev->dev_private) {
 159                 drm_via_private_t *dev_priv =
 160                     (drm_via_private_t *) dev->dev_private;
 161
 162                 if (dev_priv->ring.virtual_start) {
 163                         via_cmdbuf_reset(dev_priv);
 164
 165                         drm_core_ioremapfree(&dev_priv->ring.map, dev);
 166                         dev_priv->ring.virtual_start = NULL;
 167                 }
 168
 169         }
 170
 171         return 0;
 172 }
 173
 174 static int via_initialize(struct drm_device * dev,
 175                           drm_via_private_t * dev_priv,
 176                           drm_via_dma_init_t * init)
 177 {
 178         if (!dev_priv || !dev_priv->mmio) {
 179                 DRM_ERROR("via_dma_init called before via_map_init\n");
 180                 return -EFAULT;
 181         }
 182
 183         if (dev_priv->ring.virtual_start != NULL) {
 184                 DRM_ERROR("called again without calling cleanup\n");
 185                 return -EFAULT;
 186         }
 187
 188         if (!dev->agp || !dev->agp->base) {
 189                 DRM_ERROR("called with no agp memory available\n");
 190                 return -EFAULT;
 191         }
 192
 193         if (dev_priv->chipset == VIA_DX9_0) {
 194                 DRM_ERROR("AGP DMA is not supported on this chip\n");
 195                 return -EINVAL;
 196         }
 197
 198         dev_priv->ring.map.offset = dev->agp->base + init->offset;
 199         dev_priv->ring.map.size = init->size;
 200         dev_priv->ring.map.type = 0;
 201         dev_priv->ring.map.flags = 0;
 202         dev_priv->ring.map.mtrr = 0;
 203
 204         drm_core_ioremap(&dev_priv->ring.map, dev);
 205
 206         if (dev_priv->ring.map.handle == NULL) {
 207                 via_dma_cleanup(dev);
 208                 DRM_ERROR("can not ioremap virtual address for"
 209                           " ring buffer\n");
 210                 return -ENOMEM;
 211         }
 212
 213         dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
 214
 215         dev_priv->dma_ptr = dev_priv->ring.virtual_start;
 216         dev_priv->dma_low = 0;
 217         dev_priv->dma_high = init->size;
 218         dev_priv->dma_wrap = init->size;
 219         dev_priv->dma_offset = init->offset;
 220         dev_priv->last_pause_ptr = NULL;
 221         dev_priv->hw_addr_ptr =
 222                 (volatile uint32_t *)((char *)dev_priv->mmio->handle +
 223                 init->reg_pause_addr);
 224
 225         via_cmdbuf_start(dev_priv);
 226
 227         return 0;
 228 }
 229
 230 static int via_dma_init(struct drm_device *dev, void *data, struct drm_file *file_priv)
 231 {
 232         drm_via_private_t *dev_priv = (drm_via_private_t *) dev->dev_private;
 233         drm_via_dma_init_t *init = data;
 234         int retcode = 0;
 235
 236         switch (init->func) {
 237         case VIA_INIT_DMA:
 238                 if (!DRM_SUSER(DRM_CURPROC))
 239                         retcode = -EPERM;
 240                 else
 241                         retcode = via_initialize(dev, dev_priv, init);
 242                 break;
 243         case VIA_CLEANUP_DMA:
 244                 if (!DRM_SUSER(DRM_CURPROC))
 245                         retcode = -EPERM;
 246                 else
 247                         retcode = via_dma_cleanup(dev);
 248                 break;
 249         case VIA_DMA_INITIALIZED:
 250                 retcode = (dev_priv->ring.virtual_start != NULL) ?
 251                         0 : -EFAULT;
 252                 break;
 253         default:
 254                 retcode = -EINVAL;
 255                 break;
 256         }
 257
 258         return retcode;
 259 }
 260
 261 static int via_dispatch_cmdbuffer(struct drm_device * dev, drm_via_cmdbuffer_t * cmd)
 262 {
 263         drm_via_private_t *dev_priv;
 264         uint32_t *vb;
 265         int ret;
 266
 267         dev_priv = (drm_via_private_t *) dev->dev_private;
 268
 269         if (dev_priv->ring.virtual_start == NULL) {
 270                 DRM_ERROR("called without initializing AGP ring buffer.\n");
 271                 return -EFAULT;
 272         }
 273
 274         if (cmd->size > VIA_PCI_BUF_SIZE) {
 275                 return -ENOMEM;
 276         }
 277
 278         if (DRM_COPY_FROM_USER(dev_priv->pci_buf, cmd->buf, cmd->size))
 279                 return -EFAULT;
 280
 281         /*
 282          * Running this function on AGP memory is dead slow. Therefore
 283          * we run it on a temporary cacheable system memory buffer and
 284          * copy it to AGP memory when ready.
 285          */
 286
 287         if ((ret =
 288              via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
 289                                        cmd->size, dev, 1))) {
 290                 return ret;
 291         }
 292
 293         vb = via_check_dma(dev_priv, (cmd->size < 0x100) ? 0x102 : cmd->size);
 294         if (vb == NULL) {
 295                 return -EAGAIN;
 296         }
 297
 298         memcpy(vb, dev_priv->pci_buf, cmd->size);
 299
 300         dev_priv->dma_low += cmd->size;
 301
 302         /*
 303          * Small submissions somehow stalls the CPU. (AGP cache effects?)
 304          * pad to greater size.
 305          */
 306
 307         if (cmd->size < 0x100)
 308                 via_pad_cache(dev_priv, (0x100 - cmd->size) >> 3);
 309         via_cmdbuf_pause(dev_priv);
 310
 311         return 0;
 312 }
 313
 314 int via_driver_dma_quiescent(struct drm_device * dev)
 315 {
 316         drm_via_private_t *dev_priv = dev->dev_private;
 317
 318         if (!via_wait_idle(dev_priv)) {
 319                 return -EBUSY;
 320         }
 321         return 0;
 322 }
 323
 324 static int via_flush_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv)
 325 {
 326
 327         LOCK_TEST_WITH_RETURN(dev, file_priv);
 328
 329         return via_driver_dma_quiescent(dev);
 330 }
 331
 332 static int via_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
 333 {
 334         drm_via_cmdbuffer_t *cmdbuf = data;
 335         int ret;
 336
 337         LOCK_TEST_WITH_RETURN(dev, file_priv);
 338
 339         DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
 340
 341         ret = via_dispatch_cmdbuffer(dev, cmdbuf);
 342         if (ret) {
 343                 return ret;
 344         }
 345
 346         return 0;
 347 }
 348
 349 static int via_dispatch_pci_cmdbuffer(struct drm_device * dev,
 350                                       drm_via_cmdbuffer_t * cmd)
 351 {
 352         drm_via_private_t *dev_priv = dev->dev_private;
 353         int ret;
 354
 355         if (cmd->size > VIA_PCI_BUF_SIZE) {
 356                 return -ENOMEM;
 357         }
 358         if (DRM_COPY_FROM_USER(dev_priv->pci_buf, cmd->buf, cmd->size))
 359                 return -EFAULT;
 360
 361         if ((ret =
 362              via_verify_command_stream((uint32_t *) dev_priv->pci_buf,
 363                                        cmd->size, dev, 0))) {
 364                 return ret;
 365         }
 366
 367         ret =
 368             via_parse_command_stream(dev, (const uint32_t *)dev_priv->pci_buf,
 369                                      cmd->size);
 370         return ret;
 371 }
 372
 373 static int via_pci_cmdbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv)
 374 {
 375         drm_via_cmdbuffer_t *cmdbuf = data;
 376         int ret;
 377
 378         LOCK_TEST_WITH_RETURN(dev, file_priv);
 379
 380         DRM_DEBUG("buf %p size %lu\n", cmdbuf->buf, cmdbuf->size);
 381
 382         ret = via_dispatch_pci_cmdbuffer(dev, cmdbuf);
 383         if (ret) {
 384                 return ret;
 385         }
 386
 387         return 0;
 388 }
 389
 390 static inline uint32_t *via_align_buffer(drm_via_private_t * dev_priv,
 391                                          uint32_t * vb, int qw_count)
 392 {
 393         for (; qw_count > 0; --qw_count) {
 394                 VIA_OUT_RING_QW(HC_DUMMY, HC_DUMMY);
 395         }
 396         return vb;
 397 }
 398
 399 /*
 400  * This function is used internally by ring buffer management code.
 401  *
 402  * Returns virtual pointer to ring buffer.
 403  */
 404 static inline uint32_t *via_get_dma(drm_via_private_t * dev_priv)
 405 {
 406         return (uint32_t *) (dev_priv->dma_ptr + dev_priv->dma_low);
 407 }
 408
 409 /*
 410  * Hooks a segment of data into the tail of the ring-buffer by
 411  * modifying the pause address stored in the buffer itself. If
 412  * the regulator has already paused, restart it.
 413  */
 414 static int via_hook_segment(drm_via_private_t * dev_priv,
 415                             uint32_t pause_addr_hi, uint32_t pause_addr_lo,
 416                             int no_pci_fire)
 417 {
 418         int paused, count;
 419         volatile uint32_t *paused_at = dev_priv->last_pause_ptr;
 420         uint32_t reader,ptr;
 421         uint32_t diff;
 422
 423         paused = 0;
 424         via_flush_write_combine();
 425         (void) *(volatile uint32_t *)(via_get_dma(dev_priv) -1);
 426
 427         *paused_at = pause_addr_lo;
 428         via_flush_write_combine();
 429         (void) *paused_at;
 430
 431         reader = *(dev_priv->hw_addr_ptr);
 432         ptr = ((volatile char *)paused_at - dev_priv->dma_ptr) +
 433                 dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
 434
 435         dev_priv->last_pause_ptr = via_get_dma(dev_priv) - 1;
 436
 437         /*
 438          * If there is a possibility that the command reader will
 439          * miss the new pause address and pause on the old one,
 440          * In that case we need to program the new start address
 441          * using PCI.
 442          */
 443
 444         diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
 445         count = 10000000;
 446         while(diff == 0 && count--) {
 447                 paused = (VIA_READ(0x41c) & 0x80000000);
 448                 if (paused)
 449                         break;
 450                 reader = *(dev_priv->hw_addr_ptr);
 451                 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
 452         }
 453
 454         paused = VIA_READ(0x41c) & 0x80000000;
 455
 456         if (paused && !no_pci_fire) {
 457                 reader = *(dev_priv->hw_addr_ptr);
 458                 diff = (uint32_t) (ptr - reader) - dev_priv->dma_diff;
 459                 diff &= (dev_priv->dma_high - 1);
 460                 if (diff != 0 && diff < (dev_priv->dma_high >> 1)) {
 461                         DRM_ERROR("Paused at incorrect address. "
 462                                   "0x%08x, 0x%08x 0x%08x\n",
 463                                   ptr, reader, dev_priv->dma_diff);
 464                 } else if (diff == 0) {
 465                         /*
 466                          * There is a concern that these writes may stall the PCI bus
 467                          * if the GPU is not idle. However, idling the GPU first
 468                          * doesn't make a difference.
 469                          */
 470
 471                         VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
 472                         VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
 473                         VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
 474                         VIA_READ(VIA_REG_TRANSPACE);
 475                 }
 476         }
 477         return paused;
 478 }
 479
 480 static int via_wait_idle(drm_via_private_t * dev_priv)
 481 {
 482         int count = 10000000;
 483
 484         while (!(VIA_READ(VIA_REG_STATUS) & VIA_VR_QUEUE_BUSY) && count--);
 485
 486         while (count-- && (VIA_READ(VIA_REG_STATUS) &
 487                            (VIA_CMD_RGTR_BUSY | VIA_2D_ENG_BUSY |
 488                             VIA_3D_ENG_BUSY))) ;
 489         return count;
 490 }
 491
 492 static uint32_t *via_align_cmd(drm_via_private_t * dev_priv, uint32_t cmd_type,
 493                                uint32_t addr, uint32_t * cmd_addr_hi,
 494                                uint32_t * cmd_addr_lo, int skip_wait)
 495 {
 496         uint32_t agp_base;
 497         uint32_t cmd_addr, addr_lo, addr_hi;
 498         uint32_t *vb;
 499         uint32_t qw_pad_count;
 500
 501         if (!skip_wait)
 502                 via_cmdbuf_wait(dev_priv, 2 * CMDBUF_ALIGNMENT_SIZE);
 503
 504         vb = via_get_dma(dev_priv);
 505         VIA_OUT_RING_QW(HC_HEADER2 | ((VIA_REG_TRANSET >> 2) << 12) |
 506                         (VIA_REG_TRANSPACE >> 2), HC_ParaType_PreCR << 16);
 507         agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
 508         qw_pad_count = (CMDBUF_ALIGNMENT_SIZE >> 3) -
 509             ((dev_priv->dma_low & CMDBUF_ALIGNMENT_MASK) >> 3);
 510
 511         cmd_addr = (addr) ? addr :
 512             agp_base + dev_priv->dma_low - 8 + (qw_pad_count << 3);
 513         addr_lo = ((HC_SubA_HAGPBpL << 24) | (cmd_type & HC_HAGPBpID_MASK) |
 514                    (cmd_addr & HC_HAGPBpL_MASK));
 515         addr_hi = ((HC_SubA_HAGPBpH << 24) | (cmd_addr >> 24));
 516
 517         vb = via_align_buffer(dev_priv, vb, qw_pad_count - 1);
 518         VIA_OUT_RING_QW(*cmd_addr_hi = addr_hi, *cmd_addr_lo = addr_lo);
 519         return vb;
 520 }
 521
 522 static void via_cmdbuf_start(drm_via_private_t * dev_priv)
 523 {
 524         uint32_t pause_addr_lo, pause_addr_hi;
 525         uint32_t start_addr, start_addr_lo;
 526         uint32_t end_addr, end_addr_lo;
 527         uint32_t command;
 528         uint32_t agp_base;
 529         uint32_t ptr;
 530         uint32_t reader;
 531         int count;
 532
 533         dev_priv->dma_low = 0;
 534
 535         agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
 536         start_addr = agp_base;
 537         end_addr = agp_base + dev_priv->dma_high;
 538
 539         start_addr_lo = ((HC_SubA_HAGPBstL << 24) | (start_addr & 0xFFFFFF));
 540         end_addr_lo = ((HC_SubA_HAGPBendL << 24) | (end_addr & 0xFFFFFF));
 541         command = ((HC_SubA_HAGPCMNT << 24) | (start_addr >> 24) |
 542                    ((end_addr & 0xff000000) >> 16));
 543
 544         dev_priv->last_pause_ptr =
 545             via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0,
 546                           &pause_addr_hi, &pause_addr_lo, 1) - 1;
 547
 548         via_flush_write_combine();
 549         (void) *(volatile uint32_t *)dev_priv->last_pause_ptr;
 550
 551         VIA_WRITE(VIA_REG_TRANSET, (HC_ParaType_PreCR << 16));
 552         VIA_WRITE(VIA_REG_TRANSPACE, command);
 553         VIA_WRITE(VIA_REG_TRANSPACE, start_addr_lo);
 554         VIA_WRITE(VIA_REG_TRANSPACE, end_addr_lo);
 555
 556         VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_hi);
 557         VIA_WRITE(VIA_REG_TRANSPACE, pause_addr_lo);
 558         DRM_WRITEMEMORYBARRIER();
 559         VIA_WRITE(VIA_REG_TRANSPACE, command | HC_HAGPCMNT_MASK);
 560         VIA_READ(VIA_REG_TRANSPACE);
 561
 562         dev_priv->dma_diff = 0;
 563
 564         count = 10000000;
 565         while (!(VIA_READ(0x41c) & 0x80000000) && count--);
 566
 567         reader = *(dev_priv->hw_addr_ptr);
 568         ptr = ((volatile char *)dev_priv->last_pause_ptr - dev_priv->dma_ptr) +
 569             dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr + 4;
 570
 571         /*
 572          * This is the difference between where we tell the
 573          * command reader to pause and where it actually pauses.
 574          * This differs between hw implementation so we need to
 575          * detect it.
 576          */
 577
 578         dev_priv->dma_diff = ptr - reader;
 579 }
 580
 581 static void via_pad_cache(drm_via_private_t * dev_priv, int qwords)
 582 {
 583         uint32_t *vb;
 584
 585         via_cmdbuf_wait(dev_priv, qwords + 2);
 586         vb = via_get_dma(dev_priv);
 587         VIA_OUT_RING_QW(HC_HEADER2, HC_ParaType_NotTex << 16);
 588         via_align_buffer(dev_priv, vb, qwords);
 589 }
 590
 591 static inline void via_dummy_bitblt(drm_via_private_t * dev_priv)
 592 {
 593         uint32_t *vb = via_get_dma(dev_priv);
 594         SetReg2DAGP(0x0C, (0 | (0 << 16)));
 595         SetReg2DAGP(0x10, 0 | (0 << 16));
 596         SetReg2DAGP(0x0, 0x1 | 0x2000 | 0xAA000000);
 597 }
 598
 599 static void via_cmdbuf_jump(drm_via_private_t * dev_priv)
 600 {
 601         uint32_t agp_base;
 602         uint32_t pause_addr_lo, pause_addr_hi;
 603         uint32_t jump_addr_lo, jump_addr_hi;
 604         volatile uint32_t *last_pause_ptr;
 605         uint32_t dma_low_save1, dma_low_save2;
 606
 607         agp_base = dev_priv->dma_offset + (uint32_t) dev_priv->agpAddr;
 608         via_align_cmd(dev_priv, HC_HAGPBpID_JUMP, 0, &jump_addr_hi,
 609                       &jump_addr_lo, 0);
 610
 611         dev_priv->dma_wrap = dev_priv->dma_low;
 612
 613         /*
 614          * Wrap command buffer to the beginning.
 615          */
 616
 617         dev_priv->dma_low = 0;
 618         if (via_cmdbuf_wait(dev_priv, CMDBUF_ALIGNMENT_SIZE) != 0) {
 619                 DRM_ERROR("via_cmdbuf_jump failed\n");
 620         }
 621
 622         via_dummy_bitblt(dev_priv);
 623         via_dummy_bitblt(dev_priv);
 624
 625         last_pause_ptr =
 626             via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
 627                           &pause_addr_lo, 0) - 1;
 628         via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
 629                       &pause_addr_lo, 0);
 630
 631         *last_pause_ptr = pause_addr_lo;
 632         dma_low_save1 = dev_priv->dma_low;
 633
 634         /*
 635          * Now, set a trap that will pause the regulator if it tries to rerun the old
 636          * command buffer. (Which may happen if via_hook_segment detecs a command regulator pause
 637          * and reissues the jump command over PCI, while the regulator has already taken the jump
 638          * and actually paused at the current buffer end).
 639          * There appears to be no other way to detect this condition, since the hw_addr_pointer
 640          * does not seem to get updated immediately when a jump occurs.
 641          */
 642
 643         last_pause_ptr =
 644                 via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
 645                               &pause_addr_lo, 0) - 1;
 646         via_align_cmd(dev_priv, HC_HAGPBpID_PAUSE, 0, &pause_addr_hi,
 647                       &pause_addr_lo, 0);
 648         *last_pause_ptr = pause_addr_lo;
 649
 650         dma_low_save2 = dev_priv->dma_low;
 651         dev_priv->dma_low = dma_low_save1;
 652         via_hook_segment(dev_priv, jump_addr_hi, jump_addr_lo, 0);
 653         dev_priv->dma_low = dma_low_save2;
 654         via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
 655 }
 656
 657
 658 static void via_cmdbuf_rewind(drm_via_private_t * dev_priv)
 659 {
 660         via_cmdbuf_jump(dev_priv);
 661 }
 662
 663 static void via_cmdbuf_flush(drm_via_private_t * dev_priv, uint32_t cmd_type)
 664 {
 665         uint32_t pause_addr_lo, pause_addr_hi;
 666
 667         via_align_cmd(dev_priv, cmd_type, 0, &pause_addr_hi, &pause_addr_lo, 0);
 668         via_hook_segment(dev_priv, pause_addr_hi, pause_addr_lo, 0);
 669 }
 670
 671 static void via_cmdbuf_pause(drm_via_private_t * dev_priv)
 672 {
 673         via_cmdbuf_flush(dev_priv, HC_HAGPBpID_PAUSE);
 674 }
 675
 676 static void via_cmdbuf_reset(drm_via_private_t * dev_priv)
 677 {
 678         via_cmdbuf_flush(dev_priv, HC_HAGPBpID_STOP);
 679         via_wait_idle(dev_priv);
 680 }
 681
 682 /*
 683  * User interface to the space and lag functions.
 684  */
 685
 686 static int via_cmdbuf_size(struct drm_device *dev, void *data, struct drm_file *file_priv)
 687 {
 688         drm_via_cmdbuf_size_t *d_siz = data;
 689         int ret = 0;
 690         uint32_t tmp_size, count;
 691         drm_via_private_t *dev_priv;
 692
 693         DRM_DEBUG("\n");
 694         LOCK_TEST_WITH_RETURN(dev, file_priv);
 695
 696         dev_priv = (drm_via_private_t *) dev->dev_private;
 697
 698         if (dev_priv->ring.virtual_start == NULL) {
 699                 DRM_ERROR("called without initializing AGP ring buffer.\n");
 700                 return -EFAULT;
 701         }
 702
 703         count = 1000000;
 704         tmp_size = d_siz->size;
 705         switch (d_siz->func) {
 706         case VIA_CMDBUF_SPACE:
 707                 while (((tmp_size = via_cmdbuf_space(dev_priv)) < d_siz->size)
 708                        && count--) {
 709                         if (!d_siz->wait) {
 710                                 break;
 711                         }
 712                 }
 713                 if (!count) {
 714                         DRM_ERROR("VIA_CMDBUF_SPACE timed out.\n");
 715                         ret = -EAGAIN;
 716                 }
 717                 break;
 718         case VIA_CMDBUF_LAG:
 719                 while (((tmp_size = via_cmdbuf_lag(dev_priv)) > d_siz->size)
 720                        && count--) {
 721                         if (!d_siz->wait) {
 722                                 break;
 723                         }
 724                 }
 725                 if (!count) {
 726                         DRM_ERROR("VIA_CMDBUF_LAG timed out.\n");
 727                         ret = -EAGAIN;
 728                 }
 729                 break;
 730         default:
 731                 ret = -EFAULT;
 732         }
 733         d_siz->size = tmp_size;
 734
 735         return ret;
 736 }
 737
 738 struct drm_ioctl_desc via_ioctls[] = {
 739         DRM_IOCTL_DEF(DRM_VIA_ALLOCMEM, via_mem_alloc, DRM_AUTH),
 740         DRM_IOCTL_DEF(DRM_VIA_FREEMEM, via_mem_free, DRM_AUTH),
 741         DRM_IOCTL_DEF(DRM_VIA_AGP_INIT, via_agp_init, DRM_AUTH|DRM_MASTER),
 742         DRM_IOCTL_DEF(DRM_VIA_FB_INIT, via_fb_init, DRM_AUTH|DRM_MASTER),
 743         DRM_IOCTL_DEF(DRM_VIA_MAP_INIT, via_map_init, DRM_AUTH|DRM_MASTER),
 744         DRM_IOCTL_DEF(DRM_VIA_DEC_FUTEX, via_decoder_futex, DRM_AUTH),
 745         DRM_IOCTL_DEF(DRM_VIA_DMA_INIT, via_dma_init, DRM_AUTH),
 746         DRM_IOCTL_DEF(DRM_VIA_CMDBUFFER, via_cmdbuffer, DRM_AUTH),
 747         DRM_IOCTL_DEF(DRM_VIA_FLUSH, via_flush_ioctl, DRM_AUTH),
 748         DRM_IOCTL_DEF(DRM_VIA_PCICMD, via_pci_cmdbuffer, DRM_AUTH),
 749         DRM_IOCTL_DEF(DRM_VIA_CMDBUF_SIZE, via_cmdbuf_size, DRM_AUTH),
 750         DRM_IOCTL_DEF(DRM_VIA_WAIT_IRQ, via_wait_irq, DRM_AUTH),
 751         DRM_IOCTL_DEF(DRM_VIA_DMA_BLIT, via_dma_blit, DRM_AUTH),
 752         DRM_IOCTL_DEF(DRM_VIA_BLIT_SYNC, via_dma_blit_sync, DRM_AUTH)
 753 };
 754
 755 int via_max_ioctl = DRM_ARRAY_SIZE(via_ioctls);