git.oblomov.eu Git - nouveau/blob - src/nv_hw.c

   1 /***************************************************************************\
   2 |*                                                                           *|
   3 |*       Copyright 1993-2003 NVIDIA, Corporation.  All rights reserved.      *|
   4 |*                                                                           *|
   5 |*     NOTICE TO USER:   The source code  is copyrighted under  U.S. and     *|
   6 |*     international laws.  Users and possessors of this source code are     *|
   7 |*     hereby granted a nonexclusive,  royalty-free copyright license to     *|
   8 |*     use this code in individual and commercial software.                  *|
   9 |*                                                                           *|
  10 |*     Any use of this source code must include,  in the user documenta-     *|
  11 |*     tion and  internal comments to the code,  notices to the end user     *|
  12 |*     as follows:                                                           *|
  13 |*                                                                           *|
  14 |*       Copyright 1993-2003 NVIDIA, Corporation.  All rights reserved.      *|
  15 |*                                                                           *|
  16 |*     NVIDIA, CORPORATION MAKES NO REPRESENTATION ABOUT THE SUITABILITY     *|
  17 |*     OF  THIS SOURCE  CODE  FOR ANY PURPOSE.  IT IS  PROVIDED  "AS IS"     *|
  18 |*     WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.  NVIDIA, CORPOR-     *|
  19 |*     ATION DISCLAIMS ALL WARRANTIES  WITH REGARD  TO THIS SOURCE CODE,     *|
  20 |*     INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY, NONINFRINGE-     *|
  21 |*     MENT,  AND FITNESS  FOR A PARTICULAR PURPOSE.   IN NO EVENT SHALL     *|
  22 |*     NVIDIA, CORPORATION  BE LIABLE FOR ANY SPECIAL,  INDIRECT,  INCI-     *|
  23 |*     DENTAL, OR CONSEQUENTIAL DAMAGES,  OR ANY DAMAGES  WHATSOEVER RE-     *|
  24 |*     SULTING FROM LOSS OF USE,  DATA OR PROFITS,  WHETHER IN AN ACTION     *|
  25 |*     OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION,  ARISING OUT OF     *|
  26 |*     OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOURCE CODE.     *|
  27 |*                                                                           *|
  28 |*     U.S. Government  End  Users.   This source code  is a "commercial     *|
  29 |*     item,"  as that  term is  defined at  48 C.F.R. 2.101 (OCT 1995),     *|
  30 |*     consisting  of "commercial  computer  software"  and  "commercial     *|
  31 |*     computer  software  documentation,"  as such  terms  are  used in     *|
  32 |*     48 C.F.R. 12.212 (SEPT 1995)  and is provided to the U.S. Govern-     *|
  33 |*     ment only as  a commercial end item.   Consistent with  48 C.F.R.     *|
  34 |*     12.212 and  48 C.F.R. 227.7202-1 through  227.7202-4 (JUNE 1995),     *|
  35 |*     all U.S. Government End Users  acquire the source code  with only     *|
  36 |*     those rights set forth herein.                                        *|
  37 |*                                                                           *|
  38  \***************************************************************************/
  39 /* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/nv/nv_hw.c,v 1.21 2006/06/16 00:19:33 mvojkovi Exp $ */
  40
  41 #include "nv_include.h"
  42 #include "nv_local.h"
  43 #include "compiler.h"
  44
  45 uint8_t nvReadVGA(NVPtr pNv, uint8_t index)
  46 {
  47   volatile const uint8_t *ptr = pNv->cur_head ? pNv->PCIO1 : pNv->PCIO0;
  48   VGA_WR08(ptr, 0x03D4, index);
  49   return VGA_RD08(ptr, 0x03D5);
  50 }
  51
  52 void nvWriteVGA(NVPtr pNv, uint8_t index, uint8_t data)
  53 {
  54   volatile const uint8_t *ptr = pNv->cur_head ? pNv->PCIO1 : pNv->PCIO0;
  55   VGA_WR08(ptr, 0x03D4, index);
  56   VGA_WR08(ptr, 0x03D5, data);
  57 }
  58
  59 CARD32 nvReadRAMDAC(NVPtr pNv, uint8_t head, uint32_t ramdac_reg)
  60 {
  61   volatile const void *ptr = head ? pNv->PRAMDAC1 : pNv->PRAMDAC0;
  62   return MMIO_IN32(ptr, ramdac_reg);
  63 }
  64
  65 void nvWriteRAMDAC(NVPtr pNv, uint8_t head, uint32_t ramdac_reg, CARD32 val)
  66 {
  67   volatile const void *ptr = head ? pNv->PRAMDAC1 : pNv->PRAMDAC0;
  68   MMIO_OUT32(ptr, ramdac_reg, val);
  69 }
  70
  71 CARD32 nvReadCRTC(NVPtr pNv, uint8_t head, uint32_t reg)
  72 {
  73   volatile const void *ptr = head ? pNv->PCRTC1 : pNv->PCRTC0;
  74   return MMIO_IN32(ptr, reg);
  75 }
  76
  77 void nvWriteCRTC(NVPtr pNv, uint8_t head, uint32_t reg, CARD32 val)
  78 {
  79   volatile const void *ptr = head ? pNv->PCRTC1 : pNv->PCRTC0;
  80   MMIO_OUT32(ptr, reg, val);
  81 }
  82
  83 void NVLockUnlock (
  84     NVPtr pNv,
  85     Bool  Lock
  86 )
  87 {
  88     CARD8 cr11;
  89
  90     nvWriteVGA(pNv, 0x1f, Lock ? 0x99 : 0x57 );
  91
  92     cr11 = nvReadVGA(pNv, 0x11);
  93     if(Lock) cr11 |= 0x80;
  94     else cr11 &= ~0x80;
  95     nvWriteVGA(pNv, 0x11, cr11);
  96 }
  97
  98 int NVShowHideCursor (
  99     NVPtr pNv,
 100     int   ShowHide
 101 )
 102 {
 103     int current = pNv->CurrentState->cursor1;
 104
 105     pNv->CurrentState->cursor1 = (pNv->CurrentState->cursor1 & 0xFE) |
 106                                  (ShowHide & 0x01);
 107
 108     nvWriteVGA(pNv, 0x31, pNv->CurrentState->cursor1);
 109
 110     if(pNv->Architecture == NV_ARCH_40) {  /* HW bug */
 111        volatile CARD32 curpos = nvReadCurRAMDAC(pNv, NV_RAMDAC_CURSOR_POS);
 112        nvWriteCurRAMDAC(pNv, NV_RAMDAC_CURSOR_POS, curpos);
 113     }
 114
 115     return (current & 0x01);
 116 }
 117
 118 /****************************************************************************\
 119 *                                                                            *
 120 * The video arbitration routines calculate some "magic" numbers.  Fixes      *
 121 * the snow seen when accessing the framebuffer without it.                   *
 122 * It just works (I hope).                                                    *
 123 *                                                                            *
 124 \****************************************************************************/
 125
 126 typedef struct {
 127   int graphics_lwm;
 128   int video_lwm;
 129   int graphics_burst_size;
 130   int video_burst_size;
 131   int valid;
 132 } nv4_fifo_info;
 133
 134 typedef struct {
 135   int pclk_khz;
 136   int mclk_khz;
 137   int nvclk_khz;
 138   char mem_page_miss;
 139   char mem_latency;
 140   int memory_width;
 141   char enable_video;
 142   char gr_during_vid;
 143   char pix_bpp;
 144   char mem_aligned;
 145   char enable_mp;
 146 } nv4_sim_state;
 147
 148 typedef struct {
 149   int graphics_lwm;
 150   int video_lwm;
 151   int graphics_burst_size;
 152   int video_burst_size;
 153   int valid;
 154 } nv10_fifo_info;
 155
 156 typedef struct {
 157   int pclk_khz;
 158   int mclk_khz;
 159   int nvclk_khz;
 160   char mem_page_miss;
 161   char mem_latency;
 162   int memory_type;
 163   int memory_width;
 164   char enable_video;
 165   char gr_during_vid;
 166   char pix_bpp;
 167   char mem_aligned;
 168   char enable_mp;
 169 } nv10_sim_state;
 170
 171
 172 static void nvGetClocks(NVPtr pNv, unsigned int *MClk, unsigned int *NVClk)
 173 {
 174     unsigned int pll, N, M, MB, NB, P;
 175
 176     if(pNv->Architecture >= NV_ARCH_40) {
 177        pll = nvReadMC(pNv, 0x4020);
 178        P = (pll >> 16) & 0x07;
 179        pll = nvReadMC(pNv, 0x4024);
 180        M = pll & 0xFF;
 181        N = (pll >> 8) & 0xFF;
 182        if(((pNv->Chipset & 0xfff0) == CHIPSET_G71) ||
 183           ((pNv->Chipset & 0xfff0) == CHIPSET_G73))
 184        {
 185           MB = 1;
 186           NB = 1;
 187        } else {
 188           MB = (pll >> 16) & 0xFF;
 189           NB = (pll >> 24) & 0xFF;
 190        }
 191        *MClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 192
 193        pll = nvReadMC(pNv, 0x4000);
 194        P = (pll >> 16) & 0x07;
 195        pll = nvReadMC(pNv, 0x4004);
 196        M = pll & 0xFF;
 197        N = (pll >> 8) & 0xFF;
 198        MB = (pll >> 16) & 0xFF;
 199        NB = (pll >> 24) & 0xFF;
 200
 201        *NVClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 202     } else
 203     if(pNv->twoStagePLL) {
 204        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_MPLL);
 205        M = pll & 0xFF;
 206        N = (pll >> 8) & 0xFF;
 207        P = (pll >> 16) & 0x0F;
 208        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_MPLL_B);
 209        if(pll & 0x80000000) {
 210            MB = pll & 0xFF;
 211            NB = (pll >> 8) & 0xFF;
 212        } else {
 213            MB = 1;
 214            NB = 1;
 215        }
 216        *MClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 217
 218        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_NVPLL);
 219        M = pll & 0xFF;
 220        N = (pll >> 8) & 0xFF;
 221        P = (pll >> 16) & 0x0F;
 222        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_NVPLL_B);
 223        if(pll & 0x80000000) {
 224            MB = pll & 0xFF;
 225            NB = (pll >> 8) & 0xFF;
 226        } else {
 227            MB = 1;
 228            NB = 1;
 229        }
 230        *NVClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 231     } else
 232     if(((pNv->Chipset & 0x0ff0) == CHIPSET_NV30) ||
 233        ((pNv->Chipset & 0x0ff0) == CHIPSET_NV35))
 234     {
 235        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_MPLL);
 236        M = pll & 0x0F;
 237        N = (pll >> 8) & 0xFF;
 238        P = (pll >> 16) & 0x07;
 239        if(pll & 0x00000080) {
 240            MB = (pll >> 4) & 0x07;
 241            NB = (pll >> 19) & 0x1f;
 242        } else {
 243            MB = 1;
 244            NB = 1;
 245        }
 246        *MClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 247
 248        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_NVPLL);
 249        M = pll & 0x0F;
 250        N = (pll >> 8) & 0xFF;
 251        P = (pll >> 16) & 0x07;
 252        if(pll & 0x00000080) {
 253            MB = (pll >> 4) & 0x07;
 254            NB = (pll >> 19) & 0x1f;
 255        } else {
 256            MB = 1;
 257            NB = 1;
 258        }
 259        *NVClk = ((N * NB * pNv->CrystalFreqKHz) / (M * MB)) >> P;
 260     } else {
 261        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_MPLL);
 262        M = pll & 0xFF;
 263        N = (pll >> 8) & 0xFF;
 264        P = (pll >> 16) & 0x0F;
 265        *MClk = (N * pNv->CrystalFreqKHz / M) >> P;
 266
 267        pll = nvReadRAMDAC0(pNv, NV_RAMDAC_NVPLL);
 268        M = pll & 0xFF;
 269        N = (pll >> 8) & 0xFF;
 270        P = (pll >> 16) & 0x0F;
 271        *NVClk = (N * pNv->CrystalFreqKHz / M) >> P;
 272     }
 273
 274 #if 0
 275     ErrorF("NVClock = %i MHz, MEMClock = %i MHz\n", *NVClk/1000, *MClk/1000);
 276 #endif
 277 }
 278
 279
 280 static void nv4CalcArbitration (
 281     nv4_fifo_info *fifo,
 282     nv4_sim_state *arb
 283 )
 284 {
 285     int data, pagemiss, cas,width, video_enable, bpp;
 286     int nvclks, mclks, pclks, vpagemiss, crtpagemiss, vbs;
 287     int found, mclk_extra, mclk_loop, cbs, m1, p1;
 288     int mclk_freq, pclk_freq, nvclk_freq, mp_enable;
 289     int us_m, us_n, us_p, video_drain_rate, crtc_drain_rate;
 290     int vpm_us, us_video, vlwm, video_fill_us, cpm_us, us_crt,clwm;
 291
 292     fifo->valid = 1;
 293     pclk_freq = arb->pclk_khz;
 294     mclk_freq = arb->mclk_khz;
 295     nvclk_freq = arb->nvclk_khz;
 296     pagemiss = arb->mem_page_miss;
 297     cas = arb->mem_latency;
 298     width = arb->memory_width >> 6;
 299     video_enable = arb->enable_video;
 300     bpp = arb->pix_bpp;
 301     mp_enable = arb->enable_mp;
 302     clwm = 0;
 303     vlwm = 0;
 304     cbs = 128;
 305     pclks = 2;
 306     nvclks = 2;
 307     nvclks += 2;
 308     nvclks += 1;
 309     mclks = 5;
 310     mclks += 3;
 311     mclks += 1;
 312     mclks += cas;
 313     mclks += 1;
 314     mclks += 1;
 315     mclks += 1;
 316     mclks += 1;
 317     mclk_extra = 3;
 318     nvclks += 2;
 319     nvclks += 1;
 320     nvclks += 1;
 321     nvclks += 1;
 322     if (mp_enable)
 323         mclks+=4;
 324     nvclks += 0;
 325     pclks += 0;
 326     found = 0;
 327     vbs = 0;
 328     while (found != 1)
 329     {
 330         fifo->valid = 1;
 331         found = 1;
 332         mclk_loop = mclks+mclk_extra;
 333         us_m = mclk_loop *1000*1000 / mclk_freq;
 334         us_n = nvclks*1000*1000 / nvclk_freq;
 335         us_p = nvclks*1000*1000 / pclk_freq;
 336         if (video_enable)
 337         {
 338             video_drain_rate = pclk_freq * 2;
 339             crtc_drain_rate = pclk_freq * bpp/8;
 340             vpagemiss = 2;
 341             vpagemiss += 1;
 342             crtpagemiss = 2;
 343             vpm_us = (vpagemiss * pagemiss)*1000*1000/mclk_freq;
 344             if (nvclk_freq * 2 > mclk_freq * width)
 345                 video_fill_us = cbs*1000*1000 / 16 / nvclk_freq ;
 346             else
 347                 video_fill_us = cbs*1000*1000 / (8 * width) / mclk_freq;
 348             us_video = vpm_us + us_m + us_n + us_p + video_fill_us;
 349             vlwm = us_video * video_drain_rate/(1000*1000);
 350             vlwm++;
 351             vbs = 128;
 352             if (vlwm > 128) vbs = 64;
 353             if (vlwm > (256-64)) vbs = 32;
 354             if (nvclk_freq * 2 > mclk_freq * width)
 355                 video_fill_us = vbs *1000*1000/ 16 / nvclk_freq ;
 356             else
 357                 video_fill_us = vbs*1000*1000 / (8 * width) / mclk_freq;
 358             cpm_us = crtpagemiss  * pagemiss *1000*1000/ mclk_freq;
 359             us_crt =
 360             us_video
 361             +video_fill_us
 362             +cpm_us
 363             +us_m + us_n +us_p
 364             ;
 365             clwm = us_crt * crtc_drain_rate/(1000*1000);
 366             clwm++;
 367         }
 368         else
 369         {
 370             crtc_drain_rate = pclk_freq * bpp/8;
 371             crtpagemiss = 2;
 372             crtpagemiss += 1;
 373             cpm_us = crtpagemiss  * pagemiss *1000*1000/ mclk_freq;
 374             us_crt =  cpm_us + us_m + us_n + us_p ;
 375             clwm = us_crt * crtc_drain_rate/(1000*1000);
 376             clwm++;
 377         }
 378         m1 = clwm + cbs - 512;
 379         p1 = m1 * pclk_freq / mclk_freq;
 380         p1 = p1 * bpp / 8;
 381         if ((p1 < m1) && (m1 > 0))
 382         {
 383             fifo->valid = 0;
 384             found = 0;
 385             if (mclk_extra ==0)   found = 1;
 386             mclk_extra--;
 387         }
 388         else if (video_enable)
 389         {
 390             if ((clwm > 511) || (vlwm > 255))
 391             {
 392                 fifo->valid = 0;
 393                 found = 0;
 394                 if (mclk_extra ==0)   found = 1;
 395                 mclk_extra--;
 396             }
 397         }
 398         else
 399         {
 400             if (clwm > 519)
 401             {
 402                 fifo->valid = 0;
 403                 found = 0;
 404                 if (mclk_extra ==0)   found = 1;
 405                 mclk_extra--;
 406             }
 407         }
 408         if (clwm < 384) clwm = 384;
 409         if (vlwm < 128) vlwm = 128;
 410         data = (int)(clwm);
 411         fifo->graphics_lwm = data;
 412         fifo->graphics_burst_size = 128;
 413         data = (int)((vlwm+15));
 414         fifo->video_lwm = data;
 415         fifo->video_burst_size = vbs;
 416     }
 417 }
 418
 419 static void nv4UpdateArbitrationSettings (
 420     unsigned      VClk,
 421     unsigned      pixelDepth,
 422     unsigned     *burst,
 423     unsigned     *lwm,
 424     NVPtr        pNv
 425 )
 426 {
 427     nv4_fifo_info fifo_data;
 428     nv4_sim_state sim_data;
 429     unsigned int MClk, NVClk, cfg1;
 430
 431     nvGetClocks(pNv, &MClk, &NVClk);
 432
 433     cfg1 = nvReadFB(pNv, NV_PFB_CFG1);
 434     sim_data.pix_bpp        = (char)pixelDepth;
 435     sim_data.enable_video   = 0;
 436     sim_data.enable_mp      = 0;
 437     sim_data.memory_width   = (nvReadEXTDEV(pNv, 0x0000) & 0x10) ? 128 : 64;
 438     sim_data.mem_latency    = (char)cfg1 & 0x0F;
 439     sim_data.mem_aligned    = 1;
 440     sim_data.mem_page_miss  = (char)(((cfg1 >> 4) &0x0F) + ((cfg1 >> 31) & 0x01));
 441     sim_data.gr_during_vid  = 0;
 442     sim_data.pclk_khz       = VClk;
 443     sim_data.mclk_khz       = MClk;
 444     sim_data.nvclk_khz      = NVClk;
 445     nv4CalcArbitration(&fifo_data, &sim_data);
 446     if (fifo_data.valid)
 447     {
 448         int  b = fifo_data.graphics_burst_size >> 4;
 449         *burst = 0;
 450         while (b >>= 1) (*burst)++;
 451         *lwm   = fifo_data.graphics_lwm >> 3;
 452     }
 453 }
 454
 455 static void nv10CalcArbitration (
 456     nv10_fifo_info *fifo,
 457     nv10_sim_state *arb
 458 )
 459 {
 460     int data, pagemiss, width, video_enable, bpp;
 461     int nvclks, mclks, pclks, vpagemiss, crtpagemiss;
 462     int nvclk_fill;
 463     int found, mclk_extra, mclk_loop, cbs, m1;
 464     int mclk_freq, pclk_freq, nvclk_freq, mp_enable;
 465     int us_m, us_m_min, us_n, us_p, crtc_drain_rate;
 466     int vus_m;
 467     int vpm_us, us_video, cpm_us, us_crt,clwm;
 468     int clwm_rnd_down;
 469     int m2us, us_pipe_min, p1clk, p2;
 470     int min_mclk_extra;
 471     int us_min_mclk_extra;
 472
 473     fifo->valid = 1;
 474     pclk_freq = arb->pclk_khz; /* freq in KHz */
 475     mclk_freq = arb->mclk_khz;
 476     nvclk_freq = arb->nvclk_khz;
 477     pagemiss = arb->mem_page_miss;
 478     width = arb->memory_width/64;
 479     video_enable = arb->enable_video;
 480     bpp = arb->pix_bpp;
 481     mp_enable = arb->enable_mp;
 482     clwm = 0;
 483
 484     cbs = 512;
 485
 486     pclks = 4; /* lwm detect. */
 487
 488     nvclks = 3; /* lwm -> sync. */
 489     nvclks += 2; /* fbi bus cycles (1 req + 1 busy) */
 490
 491     mclks  = 1;   /* 2 edge sync.  may be very close to edge so just put one. */
 492
 493     mclks += 1;   /* arb_hp_req */
 494     mclks += 5;   /* ap_hp_req   tiling pipeline */
 495
 496     mclks += 2;    /* tc_req     latency fifo */
 497     mclks += 2;    /* fb_cas_n_  memory request to fbio block */
 498     mclks += 7;    /* sm_d_rdv   data returned from fbio block */
 499
 500     /* fb.rd.d.Put_gc   need to accumulate 256 bits for read */
 501     if (arb->memory_type == 0)
 502       if (arb->memory_width == 64) /* 64 bit bus */
 503         mclks += 4;
 504       else
 505         mclks += 2;
 506     else
 507       if (arb->memory_width == 64) /* 64 bit bus */
 508         mclks += 2;
 509       else
 510         mclks += 1;
 511
 512     if ((!video_enable) && (arb->memory_width == 128))
 513     {
 514       mclk_extra = (bpp == 32) ? 31 : 42; /* Margin of error */
 515       min_mclk_extra = 17;
 516     }
 517     else
 518     {
 519       mclk_extra = (bpp == 32) ? 8 : 4; /* Margin of error */
 520       /* mclk_extra = 4; */ /* Margin of error */
 521       min_mclk_extra = 18;
 522     }
 523
 524     nvclks += 1; /* 2 edge sync.  may be very close to edge so just put one. */
 525     nvclks += 1; /* fbi_d_rdv_n */
 526     nvclks += 1; /* Fbi_d_rdata */
 527     nvclks += 1; /* crtfifo load */
 528
 529     if(mp_enable)
 530       mclks+=4; /* Mp can get in with a burst of 8. */
 531     /* Extra clocks determined by heuristics */
 532
 533     nvclks += 0;
 534     pclks += 0;
 535     found = 0;
 536     while(found != 1) {
 537       fifo->valid = 1;
 538       found = 1;
 539       mclk_loop = mclks+mclk_extra;
 540       us_m = mclk_loop *1000*1000 / mclk_freq; /* Mclk latency in us */
 541       us_m_min = mclks * 1000*1000 / mclk_freq; /* Minimum Mclk latency in us */
 542       us_min_mclk_extra = min_mclk_extra *1000*1000 / mclk_freq;
 543       us_n = nvclks*1000*1000 / nvclk_freq;/* nvclk latency in us */
 544       us_p = pclks*1000*1000 / pclk_freq;/* nvclk latency in us */
 545       us_pipe_min = us_m_min + us_n + us_p;
 546
 547       vus_m = mclk_loop *1000*1000 / mclk_freq; /* Mclk latency in us */
 548
 549       if(video_enable) {
 550         crtc_drain_rate = pclk_freq * bpp/8; /* MB/s */
 551
 552         vpagemiss = 1; /* self generating page miss */
 553         vpagemiss += 1; /* One higher priority before */
 554
 555         crtpagemiss = 2; /* self generating page miss */
 556         if(mp_enable)
 557             crtpagemiss += 1; /* if MA0 conflict */
 558
 559         vpm_us = (vpagemiss * pagemiss)*1000*1000/mclk_freq;
 560
 561         us_video = vpm_us + vus_m; /* Video has separate read return path */
 562
 563         cpm_us = crtpagemiss  * pagemiss *1000*1000/ mclk_freq;
 564         us_crt =
 565           us_video  /* Wait for video */
 566           +cpm_us /* CRT Page miss */
 567           +us_m + us_n +us_p /* other latency */
 568           ;
 569
 570         clwm = us_crt * crtc_drain_rate/(1000*1000);
 571         clwm++; /* fixed point <= float_point - 1.  Fixes that */
 572       } else {
 573         crtc_drain_rate = pclk_freq * bpp/8; /* bpp * pclk/8 */
 574
 575         crtpagemiss = 1; /* self generating page miss */
 576         crtpagemiss += 1; /* MA0 page miss */
 577         if(mp_enable)
 578             crtpagemiss += 1; /* if MA0 conflict */
 579         cpm_us = crtpagemiss  * pagemiss *1000*1000/ mclk_freq;
 580         us_crt =  cpm_us + us_m + us_n + us_p ;
 581         clwm = us_crt * crtc_drain_rate/(1000*1000);
 582         clwm++; /* fixed point <= float_point - 1.  Fixes that */
 583
 584           /* Finally, a heuristic check when width == 64 bits */
 585           if(width == 1){
 586               nvclk_fill = nvclk_freq * 8;
 587               if(crtc_drain_rate * 100 >= nvclk_fill * 102)
 588                       clwm = 0xfff; /*Large number to fail */
 589
 590               else if(crtc_drain_rate * 100  >= nvclk_fill * 98) {
 591                   clwm = 1024;
 592                   cbs = 512;
 593               }
 594           }
 595       }
 596
 597
 598       /*
 599         Overfill check:
 600
 601         */
 602
 603       clwm_rnd_down = ((int)clwm/8)*8;
 604       if (clwm_rnd_down < clwm)
 605           clwm += 8;
 606
 607       m1 = clwm + cbs -  1024; /* Amount of overfill */
 608       m2us = us_pipe_min + us_min_mclk_extra;
 609
 610       /* pclk cycles to drain */
 611       p1clk = m2us * pclk_freq/(1000*1000);
 612       p2 = p1clk * bpp / 8; /* bytes drained. */
 613
 614       if((p2 < m1) && (m1 > 0)) {
 615           fifo->valid = 0;
 616           found = 0;
 617           if(min_mclk_extra == 0)   {
 618             if(cbs <= 32) {
 619               found = 1; /* Can't adjust anymore! */
 620             } else {
 621               cbs = cbs/2;  /* reduce the burst size */
 622             }
 623           } else {
 624             min_mclk_extra--;
 625           }
 626       } else {
 627         if (clwm > 1023){ /* Have some margin */
 628           fifo->valid = 0;
 629           found = 0;
 630           if(min_mclk_extra == 0)
 631               found = 1; /* Can't adjust anymore! */
 632           else
 633               min_mclk_extra--;
 634         }
 635       }
 636
 637       if(clwm < (1024-cbs+8)) clwm = 1024-cbs+8;
 638       data = (int)(clwm);
 639       /*  printf("CRT LWM: %f bytes, prog: 0x%x, bs: 256\n", clwm, data ); */
 640       fifo->graphics_lwm = data;   fifo->graphics_burst_size = cbs;
 641
 642       fifo->video_lwm = 1024;  fifo->video_burst_size = 512;
 643     }
 644 }
 645
 646 static void nv10UpdateArbitrationSettings (
 647     unsigned      VClk,
 648     unsigned      pixelDepth,
 649     unsigned     *burst,
 650     unsigned     *lwm,
 651     NVPtr        pNv
 652 )
 653 {
 654     nv10_fifo_info fifo_data;
 655     nv10_sim_state sim_data;
 656     unsigned int MClk, NVClk, cfg1;
 657
 658     nvGetClocks(pNv, &MClk, &NVClk);
 659
 660     cfg1 = nvReadFB(pNv, NV_PFB_CFG1);
 661     sim_data.pix_bpp        = (char)pixelDepth;
 662     sim_data.enable_video   = 1;
 663     sim_data.enable_mp      = 0;
 664     sim_data.memory_type    = (nvReadFB(pNv, NV_PFB_CFG0) & 0x01) ? 1 : 0;
 665     sim_data.memory_width   = (nvReadEXTDEV(pNv, 0x0000) & 0x10) ? 128 : 64;
 666     sim_data.mem_latency    = (char)cfg1 & 0x0F;
 667     sim_data.mem_aligned    = 1;
 668     sim_data.mem_page_miss  = (char)(((cfg1>>4) &0x0F) + ((cfg1>>31) & 0x01));
 669     sim_data.gr_during_vid  = 0;
 670     sim_data.pclk_khz       = VClk;
 671     sim_data.mclk_khz       = MClk;
 672     sim_data.nvclk_khz      = NVClk;
 673     nv10CalcArbitration(&fifo_data, &sim_data);
 674     if (fifo_data.valid) {
 675         int  b = fifo_data.graphics_burst_size >> 4;
 676         *burst = 0;
 677         while (b >>= 1) (*burst)++;
 678         *lwm   = fifo_data.graphics_lwm >> 3;
 679     }
 680 }
 681
 682
 683 static void nv30UpdateArbitrationSettings (
 684     NVPtr        pNv,
 685     unsigned     *burst,
 686     unsigned     *lwm
 687 )
 688 {
 689     unsigned int MClk, NVClk;
 690     unsigned int fifo_size, burst_size, graphics_lwm;
 691
 692     fifo_size = 2048;
 693     burst_size = 512;
 694     graphics_lwm = fifo_size - burst_size;
 695
 696     nvGetClocks(pNv, &MClk, &NVClk);
 697
 698     *burst = 0;
 699     burst_size >>= 5;
 700     while(burst_size >>= 1) (*burst)++;
 701     *lwm = graphics_lwm >> 3;
 702 }
 703
 704 static void nForceUpdateArbitrationSettings (
 705     unsigned      VClk,
 706     unsigned      pixelDepth,
 707     unsigned     *burst,
 708     unsigned     *lwm,
 709     NVPtr        pNv
 710 )
 711 {
 712     nv10_fifo_info fifo_data;
 713     nv10_sim_state sim_data;
 714     unsigned int M, N, P, pll, MClk, NVClk, memctrl;
 715
 716     if((pNv->Chipset & 0x0FF0) == CHIPSET_NFORCE) {
 717        unsigned int uMClkPostDiv;
 718
 719        uMClkPostDiv = (pciReadLong(pciTag(0, 0, 3), 0x6C) >> 8) & 0xf;
 720        if(!uMClkPostDiv) uMClkPostDiv = 4;
 721        MClk = 400000 / uMClkPostDiv;
 722     } else {
 723        MClk = pciReadLong(pciTag(0, 0, 5), 0x4C) / 1000;
 724     }
 725
 726     pll = nvReadRAMDAC0(pNv, NV_RAMDAC_NVPLL);
 727     M = (pll >> 0)  & 0xFF; N = (pll >> 8)  & 0xFF; P = (pll >> 16) & 0x0F;
 728     NVClk  = (N * pNv->CrystalFreqKHz / M) >> P;
 729     sim_data.pix_bpp        = (char)pixelDepth;
 730     sim_data.enable_video   = 0;
 731     sim_data.enable_mp      = 0;
 732     sim_data.memory_type    = (pciReadLong(pciTag(0, 0, 1), 0x7C) >> 12) & 1;
 733     sim_data.memory_width   = 64;
 734
 735     memctrl = pciReadLong(pciTag(0, 0, 3), 0x00) >> 16;
 736
 737     if((memctrl == 0x1A9) || (memctrl == 0x1AB) || (memctrl == 0x1ED)) {
 738         int dimm[3];
 739
 740         dimm[0] = (pciReadLong(pciTag(0, 0, 2), 0x40) >> 8) & 0x4F;
 741         dimm[1] = (pciReadLong(pciTag(0, 0, 2), 0x44) >> 8) & 0x4F;
 742         dimm[2] = (pciReadLong(pciTag(0, 0, 2), 0x48) >> 8) & 0x4F;
 743
 744         if((dimm[0] + dimm[1]) != dimm[2]) {
 745              ErrorF("WARNING: "
 746               "your nForce DIMMs are not arranged in optimal banks!\n");
 747         }
 748     }
 749
 750     sim_data.mem_latency    = 3;
 751     sim_data.mem_aligned    = 1;
 752     sim_data.mem_page_miss  = 10;
 753     sim_data.gr_during_vid  = 0;
 754     sim_data.pclk_khz       = VClk;
 755     sim_data.mclk_khz       = MClk;
 756     sim_data.nvclk_khz      = NVClk;
 757     nv10CalcArbitration(&fifo_data, &sim_data);
 758     if (fifo_data.valid)
 759     {
 760         int  b = fifo_data.graphics_burst_size >> 4;
 761         *burst = 0;
 762         while (b >>= 1) (*burst)++;
 763         *lwm   = fifo_data.graphics_lwm >> 3;
 764     }
 765 }
 766
 767
 768 /****************************************************************************\
 769 *                                                                            *
 770 *                          RIVA Mode State Routines                          *
 771 *                                                                            *
 772 \****************************************************************************/
 773
 774 /*
 775  * Calculate the Video Clock parameters for the PLL.
 776  */
 777 static void CalcVClock (
 778     int           clockIn,
 779     int          *clockOut,
 780     CARD32         *pllOut,
 781     NVPtr        pNv
 782 )
 783 {
 784     unsigned lowM, highM;
 785     unsigned DeltaNew, DeltaOld;
 786     unsigned VClk, Freq;
 787     unsigned M, N, P;
 788
 789     DeltaOld = 0xFFFFFFFF;
 790
 791     VClk = (unsigned)clockIn;
 792
 793     if (pNv->CrystalFreqKHz == 13500) {
 794         lowM  = 7;
 795         highM = 13;
 796     } else {
 797         lowM  = 8;
 798         highM = 14;
 799     }
 800
 801     for (P = 0; P <= 4; P++) {
 802         Freq = VClk << P;
 803         if ((Freq >= 128000) && (Freq <= 350000)) {
 804             for (M = lowM; M <= highM; M++) {
 805                 N = ((VClk << P) * M) / pNv->CrystalFreqKHz;
 806                 if(N <= 255) {
 807                     Freq = ((pNv->CrystalFreqKHz * N) / M) >> P;
 808                     if (Freq > VClk)
 809                         DeltaNew = Freq - VClk;
 810                     else
 811                         DeltaNew = VClk - Freq;
 812                     if (DeltaNew < DeltaOld) {
 813                         *pllOut   = (P << 16) | (N << 8) | M;
 814                         *clockOut = Freq;
 815                         DeltaOld  = DeltaNew;
 816                     }
 817                 }
 818             }
 819         }
 820     }
 821 }
 822
 823 static void CalcVClock2Stage (
 824     int           clockIn,
 825     int          *clockOut,
 826     CARD32         *pllOut,
 827     CARD32         *pllBOut,
 828     NVPtr        pNv
 829 )
 830 {
 831     unsigned DeltaNew, DeltaOld;
 832     unsigned VClk, Freq;
 833     unsigned M, N, P;
 834
 835     DeltaOld = 0xFFFFFFFF;
 836
 837     *pllBOut = 0x80000401;  /* fixed at x4 for now */
 838
 839     VClk = (unsigned)clockIn;
 840
 841     for (P = 0; P <= 6; P++) {
 842         Freq = VClk << P;
 843         if ((Freq >= 400000) && (Freq <= 1000000)) {
 844             for (M = 1; M <= 13; M++) {
 845                 N = ((VClk << P) * M) / (pNv->CrystalFreqKHz << 2);
 846                 if((N >= 5) && (N <= 255)) {
 847                     Freq = (((pNv->CrystalFreqKHz << 2) * N) / M) >> P;
 848                     if (Freq > VClk)
 849                         DeltaNew = Freq - VClk;
 850                     else
 851                         DeltaNew = VClk - Freq;
 852                     if (DeltaNew < DeltaOld) {
 853                         *pllOut   = (P << 16) | (N << 8) | M;
 854                         *clockOut = Freq;
 855                         DeltaOld  = DeltaNew;
 856                     }
 857                 }
 858             }
 859         }
 860     }
 861 }
 862
 863 /*
 864  * Calculate extended mode parameters (SVGA) and save in a
 865  * mode state structure.
 866  */
 867 void NVCalcStateExt (
 868     NVPtr pNv,
 869     RIVA_HW_STATE *state,
 870     int            bpp,
 871     int            width,
 872     int            hDisplaySize,
 873     int            height,
 874     int            dotClock,
 875     int            flags
 876 )
 877 {
 878     int pixelDepth, VClk;
 879         CARD32 CursorStart;
 880
 881     /*
 882      * Save mode parameters.
 883      */
 884     state->bpp    = bpp;    /* this is not bitsPerPixel, it's 8,15,16,32 */
 885     state->width  = width;
 886     state->height = height;
 887     /*
 888      * Extended RIVA registers.
 889      */
 890     pixelDepth = (bpp + 1)/8;
 891     if(pNv->twoStagePLL)
 892         CalcVClock2Stage(dotClock, &VClk, &state->pll, &state->pllB, pNv);
 893     else
 894         CalcVClock(dotClock, &VClk, &state->pll, pNv);
 895
 896     switch (pNv->Architecture)
 897     {
 898         case NV_ARCH_04:
 899             nv4UpdateArbitrationSettings(VClk,
 900                                          pixelDepth * 8,
 901                                         &(state->arbitration0),
 902                                         &(state->arbitration1),
 903                                          pNv);
 904             state->cursor0  = 0x00;
 905             state->cursor1  = 0xbC;
 906             if (flags & V_DBLSCAN)
 907                 state->cursor1 |= 2;
 908             state->cursor2  = 0x00000000;
 909             state->pllsel   = 0x10000700;
 910             state->config   = 0x00001114;
 911             state->general  = bpp == 16 ? 0x00101100 : 0x00100100;
 912             state->repaint1 = hDisplaySize < 1280 ? 0x04 : 0x00;
 913             break;
 914         case NV_ARCH_10:
 915         case NV_ARCH_20:
 916         case NV_ARCH_30:
 917         default:
 918             if(((pNv->Chipset & 0xfff0) == CHIPSET_C51) ||
 919                ((pNv->Chipset & 0xfff0) == CHIPSET_C512))
 920             {
 921                 state->arbitration0 = 128;
 922                 state->arbitration1 = 0x0480;
 923             } else
 924             if(((pNv->Chipset & 0xffff) == CHIPSET_NFORCE) ||
 925                ((pNv->Chipset & 0xffff) == CHIPSET_NFORCE2))
 926             {
 927                 nForceUpdateArbitrationSettings(VClk,
 928                                           pixelDepth * 8,
 929                                          &(state->arbitration0),
 930                                          &(state->arbitration1),
 931                                           pNv);
 932             } else if(pNv->Architecture < NV_ARCH_30) {
 933                 nv10UpdateArbitrationSettings(VClk,
 934                                           pixelDepth * 8,
 935                                          &(state->arbitration0),
 936                                          &(state->arbitration1),
 937                                           pNv);
 938             } else {
 939                 nv30UpdateArbitrationSettings(pNv,
 940                                          &(state->arbitration0),
 941                                          &(state->arbitration1));
 942             }
 943                         CursorStart = pNv->Cursor->offset - pNv->VRAMPhysical;
 944             state->cursor0  = 0x80 | (CursorStart >> 17);
 945             state->cursor1  = (CursorStart >> 11) << 2;
 946             state->cursor2  = CursorStart >> 24;
 947             if (flags & V_DBLSCAN)
 948                 state->cursor1 |= 2;
 949             state->pllsel   = 0x10000700;
 950             state->config   = nvReadFB(pNv, NV_PFB_CFG0);
 951             state->general  = bpp == 16 ? 0x00101100 : 0x00100100;
 952             state->repaint1 = hDisplaySize < 1280 ? 0x04 : 0x00;
 953             break;
 954     }
 955
 956     if(bpp != 8) /* DirectColor */
 957         state->general |= 0x00000030;
 958
 959     state->repaint0 = (((width / 8) * pixelDepth) & 0x700) >> 3;
 960     state->pixel    = (pixelDepth > 2) ? 3 : pixelDepth;
 961 }
 962
 963
 964 void NVLoadStateExt (
 965     ScrnInfoPtr pScrn,
 966     RIVA_HW_STATE *state
 967 )
 968 {
 969     NVPtr pNv = NVPTR(pScrn);
 970     int i, j;
 971     CARD32 temp;
 972
 973     if (!pNv->IRQ)
 974         nvWriteMC(pNv, 0x140, 0);
 975
 976     nvWriteTIMER(pNv, 0x0200, 0x00000008);
 977     nvWriteTIMER(pNv, 0x0210, 0x00000003);
 978     /*TODO: DRM handle PTIMER interrupts */
 979     nvWriteTIMER(pNv, 0x0140, 0x00000000);
 980     nvWriteTIMER(pNv, 0x0100, 0xFFFFFFFF);
 981
 982     /* begin surfaces */
 983     /* it seems those regions are equivalent to the radeon's SURFACEs. needs to go in-kernel just like the SURFACEs */
 984     if(pNv->Architecture == NV_ARCH_04) {
 985         nvWriteFB(pNv, NV_PFB_CFG0, state->config);
 986     } else
 987     if((pNv->Architecture < NV_ARCH_40) ||
 988        ((pNv->Chipset & 0xfff0) == CHIPSET_NV40))
 989     {
 990         for(i = 0; i < 8; i++) {
 991            nvWriteFB(pNv, (NV_PFB_TILE_NV10 + (i * 0x10)), 0);
 992            nvWriteFB(pNv, (NV_PFB_TILE_SIZE_NV10 + (i * 0x10)), pNv->VRAMPhysicalSize - 1);
 993         }
 994     } else {
 995         int regions = 12;
 996
 997         if(((pNv->Chipset & 0xfff0) == CHIPSET_G70) ||
 998            ((pNv->Chipset & 0xfff0) == CHIPSET_G71) ||
 999            ((pNv->Chipset & 0xfff0) == CHIPSET_G72) ||
1000            ((pNv->Chipset & 0xfff0) == CHIPSET_G73) ||
1001            ((pNv->Chipset & 0xfff0) == CHIPSET_C512))
1002         {
1003            regions = 15;
1004         }
1005
1006        for(i = 0; i < regions; i++) {
1007           nvWriteFB(pNv, (NV_PFB_TILE_NV40 + (i * 0x10)), 0);
1008           nvWriteFB(pNv, (NV_PFB_TILE_SIZE_NV40 + (i * 0x10)), pNv->VRAMPhysicalSize - 1);
1009        }
1010     }
1011     /* end of surfaces */
1012
1013     if(pNv->Architecture < NV_ARCH_10) {
1014        if((pNv->Chipset & 0x0fff) == CHIPSET_NV04) {
1015                    /*XXX: RAMIN access here, find out what it's for.
1016                         *     The DRM is handling RAMIN now
1017                         */
1018            pNv->PRAMIN[0x0824] |= 0x00020000;
1019            pNv->PRAMIN[0x0826] += pNv->VRAMPhysical;
1020        }
1021        nvWriteGRAPH(pNv, 0x0080, 0x000001FF);
1022        nvWriteGRAPH(pNv, 0x0080, 0x1230C000);
1023        nvWriteGRAPH(pNv, 0x0084, 0x72111101);
1024        nvWriteGRAPH(pNv, 0x0088, 0x11D5F071);
1025        nvWriteGRAPH(pNv, 0x008C, 0x0004FF31);
1026        nvWriteGRAPH(pNv, 0x008C, 0x4004FF31);
1027
1028        if (!pNv->IRQ) {
1029            nvWriteGRAPH(pNv, 0x140, 0x0);
1030            nvWriteGRAPH(pNv, 0x100, 0xFFFFFFFF);
1031        }
1032        nvWriteGRAPH(pNv, 0x0170, 0x10010100);
1033        nvWriteGRAPH(pNv, 0x0710, 0xFFFFFFFF);
1034        nvWriteGRAPH(pNv, 0x0720, 0x00000001);
1035
1036        nvWriteGRAPH(pNv, 0x0810, 0x00000000);
1037        nvWriteGRAPH(pNv, 0x0608, 0xFFFFFFFF);
1038     } else {
1039        nvWriteGRAPH(pNv, 0x0080, 0xFFFFFFFF);
1040        nvWriteGRAPH(pNv, 0x0080, 0x00000000);
1041
1042        if (!pNv->IRQ) {
1043            nvWriteGRAPH(pNv, 0x140, 0x0);
1044            nvWriteGRAPH(pNv, 0x100, 0xFFFFFFFF);
1045        }
1046        nvWriteGRAPH(pNv, 0x0144, 0x10010100);
1047        nvWriteGRAPH(pNv, 0x0714, 0xFFFFFFFF);
1048        nvWriteGRAPH(pNv, 0x0720, 0x00000001);
1049        temp = nvReadGRAPH(pNv, 0x710);
1050        nvWriteGRAPH(pNv, 0x0710, temp & 0x0007ff00);
1051        temp = nvReadGRAPH(pNv, 0x710);
1052        nvWriteGRAPH(pNv, 0x0710, temp | 0x00020100);
1053
1054        if(pNv->Architecture == NV_ARCH_10) {
1055            nvWriteGRAPH(pNv, 0x0084, 0x00118700);
1056            nvWriteGRAPH(pNv, 0x0088, 0x24E00810);
1057            nvWriteGRAPH(pNv, 0x008C, 0x55DE0030);
1058
1059            /* nv10 second surfaces */
1060            /* this is a copy of the surfaces. What is it for ? */
1061            for(i = 0; i < 32; i++)
1062              nvWriteGRAPH(pNv, 0x0B00 + (i*4), nvReadFB(pNv, NV_PFB_TILE_NV10 + (i*4)));
1063            /* end of nv10 second surfaces */
1064
1065            nvWriteGRAPH(pNv, 0x640, 0);
1066            nvWriteGRAPH(pNv, 0x644, 0);
1067            nvWriteGRAPH(pNv, 0x684, pNv->VRAMPhysicalSize - 1);
1068            nvWriteGRAPH(pNv, 0x688, pNv->VRAMPhysicalSize - 1);
1069
1070            nvWriteGRAPH(pNv, 0x0810, 0x00000000);
1071            nvWriteGRAPH(pNv, 0x0608, 0xFFFFFFFF);
1072        } else {
1073            if(pNv->Architecture >= NV_ARCH_40) {
1074               nvWriteGRAPH(pNv, 0x0084, 0x401287c0);
1075               nvWriteGRAPH(pNv, 0x008C, 0x60de8051);
1076               nvWriteGRAPH(pNv, 0x0090, 0x00008000);
1077               nvWriteGRAPH(pNv, 0x0610, 0x00be3c5f);
1078
1079               j = pNv->REGS[0x1540/4] & 0xff;
1080               if(j) {
1081                   for(i = 0; !(j & 1); j >>= 1, i++);
1082                   nvWriteGRAPH(pNv, 0x5000, i);
1083               }
1084
1085               if((pNv->Chipset & 0xfff0) == CHIPSET_NV40) {
1086                  nvWriteGRAPH(pNv, 0x09b0, 0x83280fff);
1087                  nvWriteGRAPH(pNv, 0x09b4, 0x000000a0);
1088               } else {
1089                  nvWriteGRAPH(pNv, 0x0820, 0x83280eff);
1090                  nvWriteGRAPH(pNv, 0x0824, 0x000000a0);
1091               }
1092
1093               switch(pNv->Chipset & 0xfff0) {
1094               case CHIPSET_NV40:
1095               case CHIPSET_NV45:
1096                  nvWriteGRAPH(pNv, 0x09b8, 0x0078e366);
1097                  nvWriteGRAPH(pNv, 0x09bc, 0x0000014c);
1098                  temp = nvReadFB(pNv, NV_PFB_CLOSE_PAGE2);
1099                  nvWriteFB(pNv, NV_PFB_CLOSE_PAGE2, temp & 0xffff7fff);
1100                  break;
1101               case CHIPSET_NV41:
1102               case 0x0120:
1103                  nvWriteGRAPH(pNv, 0x0828, 0x007596ff);
1104                  nvWriteGRAPH(pNv, 0x082C, 0x00000108);
1105                  break;
1106               case CHIPSET_NV44:
1107               case CHIPSET_G72:
1108               case CHIPSET_C51:
1109               case CHIPSET_C512:
1110                  nvWriteMC(pNv, 0x1700, nvReadFB(pNv, NV_PFB_020C));
1111                  nvWriteMC(pNv, 0x1704, 0);
1112                  nvWriteMC(pNv, 0x1708, 0);
1113                  nvWriteMC(pNv, 0x170C, nvReadFB(pNv, NV_PFB_020C));
1114                  nvWriteGRAPH(pNv, 0x0860, 0);
1115                  nvWriteGRAPH(pNv, 0x0864, 0);
1116                  temp = nvReadCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL);
1117                  nvWriteCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL, temp | 0x00100000);
1118                  break;
1119               case CHIPSET_NV43:
1120                  nvWriteGRAPH(pNv, 0x0828, 0x0072cb77);
1121                  nvWriteGRAPH(pNv, 0x082C, 0x00000108);
1122                  break;
1123               case CHIPSET_NV44A:
1124                  nvWriteGRAPH(pNv, 0x0860, 0);
1125                  nvWriteGRAPH(pNv, 0x0864, 0);
1126                  temp = nvReadCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL);
1127                  nvWriteCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL, temp | 0x00100000);
1128                  break;
1129               case CHIPSET_G70:
1130               case CHIPSET_G71:
1131               case CHIPSET_G73:
1132                  temp = nvReadCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL);
1133                  nvWriteCurRAMDAC(pNv, NV_RAMDAC_TEST_CONTROL, temp | 0x00100000);
1134                  nvWriteGRAPH(pNv, 0x0828, 0x07830610);
1135                  nvWriteGRAPH(pNv, 0x082C, 0x0000016A);
1136                  break;
1137               default:
1138                  break;
1139               };
1140
1141               nvWriteGRAPH(pNv, 0x0b38, 0x2ffff800);
1142               nvWriteGRAPH(pNv, 0x0b3c, 0x00006000);
1143               nvWriteGRAPH(pNv, 0x032C, 0x01000000);
1144            } else
1145            if(pNv->Architecture == NV_ARCH_30) {
1146               nvWriteGRAPH(pNv, 0x0084, 0x40108700);
1147               nvWriteGRAPH(pNv, 0x0890, 0x00140000);
1148               nvWriteGRAPH(pNv, 0x008C, 0xf00e0431);
1149               nvWriteGRAPH(pNv, 0x0090, 0x00008000);
1150               nvWriteGRAPH(pNv, 0x0610, 0xf04b1f36);
1151               nvWriteGRAPH(pNv, 0x0B80, 0x1002d888);
1152               nvWriteGRAPH(pNv, 0x0B88, 0x62ff007f);
1153            } else {
1154               nvWriteGRAPH(pNv, 0x0084, 0x00118700);
1155               nvWriteGRAPH(pNv, 0x008C, 0xF20E0431);
1156               nvWriteGRAPH(pNv, 0x0090, 0x00000000);
1157               nvWriteGRAPH(pNv, 0x009C, 0x00000040);
1158
1159               if((pNv->Chipset & 0x0ff0) >= CHIPSET_NV25) {
1160                  nvWriteGRAPH(pNv, 0x0890, 0x00080000);
1161                  nvWriteGRAPH(pNv, 0x0610, 0x304B1FB6);
1162                  nvWriteGRAPH(pNv, 0x0B80, 0x18B82880);
1163                  nvWriteGRAPH(pNv, 0x0B84, 0x44000000);
1164                  nvWriteGRAPH(pNv, 0x0098, 0x40000080);
1165                  nvWriteGRAPH(pNv, 0x0B88, 0x000000ff);
1166               } else {
1167                  nvWriteGRAPH(pNv, 0x0880, 0x00080000);
1168                  nvWriteGRAPH(pNv, 0x0094, 0x00000005);
1169                  nvWriteGRAPH(pNv, 0x0B80, 0x45CAA208);
1170                  nvWriteGRAPH(pNv, 0x0B84, 0x24000000);
1171                  nvWriteGRAPH(pNv, 0x0098, 0x00000040);
1172                  nvWriteGRAPH(pNv, 0x0750, 0x00E00038);
1173                  nvWriteGRAPH(pNv, 0x0754, 0x00000030);
1174                  nvWriteGRAPH(pNv, 0x0750, 0x00E10038);
1175                  nvWriteGRAPH(pNv, 0x0754, 0x00000030);
1176               }
1177            }
1178
1179            /* begin nv20+ secondr surfaces */
1180            /* again, a copy of the surfaces. */
1181            if((pNv->Architecture < NV_ARCH_40) ||
1182               ((pNv->Chipset & 0xfff0) == CHIPSET_NV40))
1183            {
1184               for(i = 0; i < 32; i++) {
1185                 nvWriteGRAPH(pNv, 0x0900 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV10 + i*4));
1186                 nvWriteGRAPH(pNv, 0x6900 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV10 + i*4));
1187               }
1188            } else {
1189               if(((pNv->Chipset & 0xfff0) == CHIPSET_G70) ||
1190                  ((pNv->Chipset & 0xfff0) == CHIPSET_G71) ||
1191                  ((pNv->Chipset & 0xfff0) == CHIPSET_G72) ||
1192                  ((pNv->Chipset & 0xfff0) == CHIPSET_G73) ||
1193                  ((pNv->Chipset & 0xfff0) == CHIPSET_C512))
1194               {
1195                  for(i = 0; i < 60; i++) {
1196                    nvWriteGRAPH(pNv, 0x0D00 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV40 + i*4));
1197                    nvWriteGRAPH(pNv, 0x6900 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV40 + i*4));
1198                  }
1199               } else {
1200                  for(i = 0; i < 48; i++) {
1201                    nvWriteGRAPH(pNv, 0x0900 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV40 + i*4));
1202                    if(((pNv->Chipset & 0xfff0) != CHIPSET_NV44) &&
1203                       ((pNv->Chipset & 0xfff0) != CHIPSET_NV44A) &&
1204                       ((pNv->Chipset & 0xfff0) != CHIPSET_C51))
1205                    {
1206                       nvWriteGRAPH(pNv, 0x6900 + i*4, nvReadFB(pNv, NV_PFB_TILE_NV40 + i*4));
1207                    }
1208                  }
1209               }
1210            }
1211            /* end nv20+ second surfaces */
1212
1213            /* begin RAM config */
1214            if(pNv->Architecture >= NV_ARCH_40) {
1215               if((pNv->Chipset & 0xfff0) == CHIPSET_NV40) {
1216                  nvWriteGRAPH(pNv, 0x09A4, nvReadFB(pNv, NV_PFB_CFG0));
1217                  nvWriteGRAPH(pNv, 0x09A8, nvReadFB(pNv, NV_PFB_CFG1));
1218                  nvWriteGRAPH(pNv, 0x69A4, nvReadFB(pNv, NV_PFB_CFG0));
1219                  nvWriteGRAPH(pNv, 0x69A8, nvReadFB(pNv, NV_PFB_CFG1));
1220
1221                  nvWriteGRAPH(pNv, 0x0820, 0);
1222                  nvWriteGRAPH(pNv, 0x0824, 0);
1223                  nvWriteGRAPH(pNv, 0x0864, pNv->VRAMPhysicalSize - 1);
1224                  nvWriteGRAPH(pNv, 0x0868, pNv->VRAMPhysicalSize - 1);
1225               } else {
1226                  if(((pNv->Chipset & 0xfff0) == CHIPSET_G70) ||
1227                     ((pNv->Chipset & 0xfff0) == CHIPSET_G71) ||
1228                     ((pNv->Chipset & 0xfff0) == CHIPSET_G72) ||
1229                     ((pNv->Chipset & 0xfff0) == CHIPSET_G73))
1230                  {
1231                     nvWriteGRAPH(pNv, 0x0DF0, nvReadFB(pNv, NV_PFB_CFG0));
1232                     nvWriteGRAPH(pNv, 0x0DF4, nvReadFB(pNv, NV_PFB_CFG1));
1233                  } else {
1234                     nvWriteGRAPH(pNv, 0x09F0, nvReadFB(pNv, NV_PFB_CFG0));
1235                     nvWriteGRAPH(pNv, 0x09F4, nvReadFB(pNv, NV_PFB_CFG1));
1236                  }
1237                  nvWriteGRAPH(pNv, 0x69F0, nvReadFB(pNv, NV_PFB_CFG0));
1238                  nvWriteGRAPH(pNv, 0x69F4, nvReadFB(pNv, NV_PFB_CFG1));
1239
1240                  nvWriteGRAPH(pNv, 0x0840, 0);
1241                  nvWriteGRAPH(pNv, 0x0844, 0);
1242                  nvWriteGRAPH(pNv, 0x08a0, pNv->VRAMPhysicalSize - 1);
1243                  nvWriteGRAPH(pNv, 0x08a4, pNv->VRAMPhysicalSize - 1);
1244               }
1245            } else {
1246               nvWriteGRAPH(pNv, 0x09A4, nvReadFB(pNv, NV_PFB_CFG0));
1247               nvWriteGRAPH(pNv, 0x09A8, nvReadFB(pNv, NV_PFB_CFG1));
1248               nvWriteGRAPH(pNv, 0x0750, 0x00EA0000);
1249               nvWriteGRAPH(pNv, 0x0754, nvReadFB(pNv, NV_PFB_CFG0));
1250               nvWriteGRAPH(pNv, 0x0750, 0x00EA0004);
1251               nvWriteGRAPH(pNv, 0x0754, nvReadFB(pNv, NV_PFB_CFG1));
1252
1253               nvWriteGRAPH(pNv, 0x0820, 0);
1254               nvWriteGRAPH(pNv, 0x0824, 0);
1255               nvWriteGRAPH(pNv, 0x0864, pNv->VRAMPhysicalSize - 1);
1256               nvWriteGRAPH(pNv, 0x0868, pNv->VRAMPhysicalSize - 1);
1257            }
1258            /* end of RAM config */
1259
1260            nvWriteGRAPH(pNv, 0x0B20, 0x00000000);
1261            nvWriteGRAPH(pNv, 0x0B04, 0xFFFFFFFF);
1262        }
1263     }
1264
1265     /* begin clipping values */
1266     nvWriteGRAPH(pNv, 0x053C, 0);
1267     nvWriteGRAPH(pNv, 0x0540, 0);
1268     nvWriteGRAPH(pNv, 0x0544, 0x00007FFF);
1269     nvWriteGRAPH(pNv, 0x0548, 0x00007FFF);
1270     /* end of clipping values */
1271
1272     if(pNv->Architecture >= NV_ARCH_10) {
1273         if(pNv->twoHeads) {
1274            nvWriteCRTC(pNv, 0, NV_CRTC_HEAD_CONFIG, state->head);
1275            nvWriteCRTC(pNv, 1, NV_CRTC_HEAD_CONFIG, state->head2);
1276         }
1277         temp = nvReadCurRAMDAC(pNv, NV_RAMDAC_0404);
1278         nvWriteCurRAMDAC(pNv, NV_RAMDAC_0404, temp | (1 << 25));
1279
1280         nvWriteMC(pNv, 0x8704, 1);
1281         nvWriteMC(pNv, 0x8140, 0);
1282         nvWriteMC(pNv, 0x8920, 0);
1283         nvWriteMC(pNv, 0x8924, 0);
1284         nvWriteMC(pNv, 0x8908, pNv->VRAMPhysicalSize - 1);
1285         nvWriteMC(pNv, 0x890C, pNv->VRAMPhysicalSize - 1);
1286         nvWriteMC(pNv, 0x1588, 0);
1287
1288         nvWriteCurCRTC(pNv, NV_CRTC_CURSOR_CONFIG, state->cursorConfig);
1289         nvWriteCurCRTC(pNv, NV_CRTC_0830, state->displayV - 3);
1290         nvWriteCurCRTC(pNv, NV_CRTC_0834, state->displayV - 1);
1291
1292         if(pNv->FlatPanel) {
1293            if((pNv->Chipset & 0x0ff0) == CHIPSET_NV11) {
1294                nvWriteCurRAMDAC(pNv, NV_RAMDAC_DITHER_NV11, state->dither);
1295            } else
1296            if(pNv->twoHeads) {
1297                nvWriteCurRAMDAC(pNv, NV_RAMDAC_FP_DITHER, state->dither);
1298            }
1299
1300            nvWriteVGA(pNv, 0x53, state->timingH);
1301            nvWriteVGA(pNv, 0x54, state->timingV);
1302            nvWriteVGA(pNv, 0x21, 0xfa);
1303         }
1304
1305         nvWriteVGA(pNv, 0x41, state->extra);
1306     }
1307
1308     nvWriteVGA(pNv, 0x19, state->repaint0);
1309     nvWriteVGA(pNv, 0x1A, state->repaint1);
1310     nvWriteVGA(pNv, 0x25, state->screen);
1311     nvWriteVGA(pNv, 0x28, state->pixel);
1312     nvWriteVGA(pNv, 0x2D, state->horiz);
1313     nvWriteVGA(pNv, 0x1C, state->fifo);
1314     nvWriteVGA(pNv, 0x1B, state->arbitration0);
1315     nvWriteVGA(pNv, 0x20, state->arbitration1);
1316     if(pNv->Architecture >= NV_ARCH_30) {
1317       nvWriteVGA(pNv, 0x47, state->arbitration1 >> 8);
1318     }
1319
1320     nvWriteVGA(pNv, 0x30, state->cursor0);
1321     nvWriteVGA(pNv, 0x31, state->cursor1);
1322     if(pNv->Architecture == NV_ARCH_40) {  /* HW bug */
1323        volatile CARD32 curpos = nvReadCurRAMDAC(pNv, NV_RAMDAC_CURSOR_POS);
1324        nvWriteCurRAMDAC(pNv, NV_RAMDAC_CURSOR_POS, curpos);
1325     }
1326     nvWriteVGA(pNv, 0x2F, state->cursor2);
1327     nvWriteVGA(pNv, 0x39, state->interlace);
1328
1329     if(!pNv->FlatPanel) {
1330        nvWriteRAMDAC0(pNv, NV_RAMDAC_PLL_SELECT, state->pllsel);
1331        nvWriteRAMDAC0(pNv, NV_RAMDAC_VPLL, state->vpll);
1332        if(pNv->twoHeads)
1333           nvWriteRAMDAC0(pNv, NV_RAMDAC_VPLL2, state->vpll2);
1334        if(pNv->twoStagePLL) {
1335           nvWriteRAMDAC0(pNv, NV_RAMDAC_VPLL_B, state->vpllB);
1336           nvWriteRAMDAC0(pNv, NV_RAMDAC_VPLL2_B, state->vpll2B);
1337        }
1338     } else {
1339        nvWriteCurRAMDAC(pNv, NV_RAMDAC_FP_CONTROL, state->scale);
1340        nvWriteCurRAMDAC(pNv, NV_RAMDAC_FP_HCRTC, state->crtcSync);
1341     }
1342     nvWriteCurRAMDAC(pNv, NV_RAMDAC_GENERAL_CONTROL, state->general);
1343
1344     nvWriteCurCRTC(pNv, NV_CRTC_INTR_EN_0, 0);
1345     nvWriteCurCRTC(pNv, NV_CRTC_INTR_0, NV_CRTC_INTR_VBLANK);
1346
1347     pNv->CurrentState = state;
1348 }
1349
1350 void NVUnloadStateExt
1351 (
1352     NVPtr pNv,
1353     RIVA_HW_STATE *state
1354 )
1355 {
1356     state->repaint0     = nvReadVGA(pNv, 0x19);
1357     state->repaint1     = nvReadVGA(pNv, 0x1A);
1358     state->screen       = nvReadVGA(pNv, 0x25);
1359     state->pixel        = nvReadVGA(pNv, 0x28);
1360     state->horiz        = nvReadVGA(pNv, 0x2D);
1361     state->fifo         = nvReadVGA(pNv, 0x1C);
1362     state->arbitration0 = nvReadVGA(pNv, 0x1B);
1363     state->arbitration1 = nvReadVGA(pNv, 0x20);
1364     if(pNv->Architecture >= NV_ARCH_30) {
1365        state->arbitration1 |= (nvReadVGA(pNv, 0x47) & 1) << 8;
1366     }
1367     state->cursor0      = nvReadVGA(pNv, 0x30);
1368     state->cursor1      = nvReadVGA(pNv, 0x31);
1369     state->cursor2      = nvReadVGA(pNv, 0x2F);
1370     state->interlace    = nvReadVGA(pNv, 0x39);
1371
1372     state->vpll         = nvReadRAMDAC0(pNv, NV_RAMDAC_VPLL);
1373     if(pNv->twoHeads)
1374        state->vpll2     = nvReadRAMDAC0(pNv, NV_RAMDAC_VPLL2);
1375     if(pNv->twoStagePLL) {
1376         state->vpllB    = nvReadRAMDAC0(pNv, NV_RAMDAC_VPLL_B);
1377         state->vpll2B   = nvReadRAMDAC0(pNv, NV_RAMDAC_VPLL2_B);
1378     }
1379     state->pllsel       = nvReadRAMDAC0(pNv, NV_RAMDAC_PLL_SELECT);
1380     state->general      = nvReadCurRAMDAC(pNv, NV_RAMDAC_GENERAL_CONTROL);
1381     state->scale        = nvReadCurRAMDAC(pNv, NV_RAMDAC_FP_CONTROL);
1382     state->config       = nvReadFB(pNv, NV_PFB_CFG0);
1383
1384     if(pNv->Architecture >= NV_ARCH_10) {
1385         if(pNv->twoHeads) {
1386            state->head     = nvReadCRTC(pNv, 0, NV_CRTC_HEAD_CONFIG);
1387            state->head2    = nvReadCRTC(pNv, 1, NV_CRTC_HEAD_CONFIG);
1388            state->crtcOwner = nvReadVGA(pNv, 0x44);
1389         }
1390         state->extra = nvReadVGA(pNv, 0x41);
1391
1392         state->cursorConfig = nvReadCurCRTC(pNv, NV_CRTC_CURSOR_CONFIG);
1393
1394         if((pNv->Chipset & 0x0ff0) == CHIPSET_NV11) {
1395            state->dither = nvReadCurRAMDAC(pNv, NV_RAMDAC_DITHER_NV11);
1396         } else
1397         if(pNv->twoHeads) {
1398             state->dither = nvReadCurRAMDAC(pNv, NV_RAMDAC_FP_DITHER);
1399         }
1400
1401         if(pNv->FlatPanel) {
1402            state->timingH = nvReadVGA(pNv, 0x53);
1403            state->timingV = nvReadVGA(pNv, 0x54);
1404         }
1405     }
1406
1407     if(pNv->FlatPanel) {
1408        state->crtcSync = nvReadCurRAMDAC(pNv, NV_RAMDAC_FP_HCRTC);
1409     }
1410 }
1411
1412 void NVSetStartAddress (
1413     NVPtr   pNv,
1414     CARD32 start
1415 )
1416 {
1417     nvWriteCurCRTC(pNv, NV_CRTC_START, start);
1418 }
1419
1420