drivers/gpu/drm/nouveau/nva3_pm.c

   1 /*
   2  * Copyright 2010 Red Hat Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Ben Skeggs
  23  */
  24
  25 #include "drmP.h"
  26 #include "nouveau_drv.h"
  27 #include "nouveau_bios.h"
  28 #include "nouveau_pm.h"
  29
  30 static u32 read_clk(struct drm_device *, int, bool);
  31 static u32 read_pll(struct drm_device *, int, u32);
  32
  33 static u32
  34 read_vco(struct drm_device *dev, int clk)
  35 {
  36         u32 sctl = nv_rd32(dev, 0x4120 + (clk * 4));
  37         if ((sctl & 0x00000030) != 0x00000030)
  38                 return read_pll(dev, 0x41, 0x00e820);
  39         return read_pll(dev, 0x42, 0x00e8a0);
  40 }
  41
  42 static u32
  43 read_clk(struct drm_device *dev, int clk, bool ignore_en)
  44 {
  45         struct drm_nouveau_private *dev_priv = dev->dev_private;
  46         u32 sctl, sdiv, sclk;
  47
  48         /* refclk for the 0xe8xx plls is a fixed frequency */
  49         if (clk >= 0x40) {
  50                 if (dev_priv->chipset == 0xaf) {
  51                         /* no joke.. seriously.. sigh.. */
  52                         return nv_rd32(dev, 0x00471c) * 1000;
  53                 }
  54
  55                 return dev_priv->crystal;
  56         }
  57
  58         sctl = nv_rd32(dev, 0x4120 + (clk * 4));
  59         if (!ignore_en && !(sctl & 0x00000100))
  60                 return 0;
  61
  62         switch (sctl & 0x00003000) {
  63         case 0x00000000:
  64                 return dev_priv->crystal;
  65         case 0x00002000:
  66                 if (sctl & 0x00000040)
  67                         return 108000;
  68                 return 100000;
  69         case 0x00003000:
  70                 sclk = read_vco(dev, clk);
  71                 sdiv = ((sctl & 0x003f0000) >> 16) + 2;
  72                 return (sclk * 2) / sdiv;
  73         default:
  74                 return 0;
  75         }
  76 }
  77
  78 static u32
  79 read_pll(struct drm_device *dev, int clk, u32 pll)
  80 {
  81         u32 ctrl = nv_rd32(dev, pll + 0);
  82         u32 sclk = 0, P = 1, N = 1, M = 1;
  83
  84         if (!(ctrl & 0x00000008)) {
  85                 if (ctrl & 0x00000001) {
  86                         u32 coef = nv_rd32(dev, pll + 4);
  87                         M = (coef & 0x000000ff) >> 0;
  88                         N = (coef & 0x0000ff00) >> 8;
  89                         P = (coef & 0x003f0000) >> 16;
  90
  91                         /* no post-divider on these.. */
  92                         if ((pll & 0x00ff00) == 0x00e800)
  93                                 P = 1;
  94
  95                         sclk = read_clk(dev, 0x00 + clk, false);
  96                 }
  97         } else {
  98                 sclk = read_clk(dev, 0x10 + clk, false);
  99         }
 100
 101         if (M * P)
 102                 return sclk * N / (M * P);
 103         return 0;
 104 }
 105
 106 struct creg {
 107         u32 clk;
 108         u32 pll;
 109 };
 110
 111 static int
 112 calc_clk(struct drm_device *dev, int clk, u32 pll, u32 khz, struct creg *reg)
 113 {
 114         struct pll_lims limits;
 115         u32 oclk, sclk, sdiv;
 116         int P, N, M, diff;
 117         int ret;
 118
 119         reg->pll = 0;
 120         reg->clk = 0;
 121         if (!khz) {
 122                 NV_DEBUG(dev, "no clock for 0x%04x/0x%02x\n", pll, clk);
 123                 return 0;
 124         }
 125
 126         switch (khz) {
 127         case 27000:
 128                 reg->clk = 0x00000100;
 129                 return khz;
 130         case 100000:
 131                 reg->clk = 0x00002100;
 132                 return khz;
 133         case 108000:
 134                 reg->clk = 0x00002140;
 135                 return khz;
 136         default:
 137                 sclk = read_vco(dev, clk);
 138                 sdiv = min((sclk * 2) / (khz - 2999), (u32)65);
 139                 /* if the clock has a PLL attached, and we can get a within
 140                  * [-2, 3) MHz of a divider, we'll disable the PLL and use
 141                  * the divider instead.
 142                  *
 143                  * divider can go as low as 2, limited here because NVIDIA
 144                  * and the VBIOS on my NVA8 seem to prefer using the PLL
 145                  * for 810MHz - is there a good reason?
 146                  */
 147                 if (sdiv > 4) {
 148                         oclk = (sclk * 2) / sdiv;
 149                         diff = khz - oclk;
 150                         if (!pll || (diff >= -2000 && diff < 3000)) {
 151                                 reg->clk = (((sdiv - 2) << 16) | 0x00003100);
 152                                 return oclk;
 153                         }
 154                 }
 155
 156                 if (!pll) {
 157                         NV_ERROR(dev, "bad freq %02x: %d %d\n", clk, khz, sclk);
 158                         return -ERANGE;
 159                 }
 160
 161                 break;
 162         }
 163
 164         ret = get_pll_limits(dev, pll, &limits);
 165         if (ret)
 166                 return ret;
 167
 168         limits.refclk = read_clk(dev, clk - 0x10, true);
 169         if (!limits.refclk)
 170                 return -EINVAL;
 171
 172         ret = nva3_calc_pll(dev, &limits, khz, &N, NULL, &M, &P);
 173         if (ret >= 0) {
 174                 reg->clk = nv_rd32(dev, 0x4120 + (clk * 4));
 175                 reg->pll = (P << 16) | (N << 8) | M;
 176         }
 177         return ret;
 178 }
 179
 180 static void
 181 prog_pll(struct drm_device *dev, int clk, u32 pll, struct creg *reg)
 182 {
 183         const u32 src0 = 0x004120 + (clk * 4);
 184         const u32 src1 = 0x004160 + (clk * 4);
 185         const u32 ctrl = pll + 0;
 186         const u32 coef = pll + 4;
 187
 188         if (!reg->clk && !reg->pll) {
 189                 NV_DEBUG(dev, "no clock for %02x\n", clk);
 190                 return;
 191         }
 192
 193         if (reg->pll) {
 194                 nv_mask(dev, src0, 0x00000101, 0x00000101);
 195                 nv_wr32(dev, coef, reg->pll);
 196                 nv_mask(dev, ctrl, 0x00000015, 0x00000015);
 197                 nv_mask(dev, ctrl, 0x00000010, 0x00000000);
 198                 nv_wait(dev, ctrl, 0x00020000, 0x00020000);
 199                 nv_mask(dev, ctrl, 0x00000010, 0x00000010);
 200                 nv_mask(dev, ctrl, 0x00000008, 0x00000000);
 201                 nv_mask(dev, src1, 0x00000100, 0x00000000);
 202                 nv_mask(dev, src1, 0x00000001, 0x00000000);
 203         } else {
 204                 nv_mask(dev, src1, 0x003f3141, 0x00000101 | reg->clk);
 205                 nv_mask(dev, ctrl, 0x00000018, 0x00000018);
 206                 udelay(20);
 207                 nv_mask(dev, ctrl, 0x00000001, 0x00000000);
 208                 nv_mask(dev, src0, 0x00000100, 0x00000000);
 209                 nv_mask(dev, src0, 0x00000001, 0x00000000);
 210         }
 211 }
 212
 213 static void
 214 prog_clk(struct drm_device *dev, int clk, struct creg *reg)
 215 {
 216         if (!reg->clk) {
 217                 NV_DEBUG(dev, "no clock for %02x\n", clk);
 218                 return;
 219         }
 220
 221         nv_mask(dev, 0x004120 + (clk * 4), 0x003f3141, 0x00000101 | reg->clk);
 222 }
 223
 224 int
 225 nva3_pm_clocks_get(struct drm_device *dev, struct nouveau_pm_level *perflvl)
 226 {
 227         perflvl->core   = read_pll(dev, 0x00, 0x4200);
 228         perflvl->shader = read_pll(dev, 0x01, 0x4220);
 229         perflvl->memory = read_pll(dev, 0x02, 0x4000);
 230         perflvl->unka0  = read_clk(dev, 0x20, false);
 231         perflvl->vdec   = read_clk(dev, 0x21, false);
 232         perflvl->daemon = read_clk(dev, 0x25, false);
 233         perflvl->copy   = perflvl->core;
 234         return 0;
 235 }
 236
 237 struct nva3_pm_state {
 238         struct nouveau_pm_level *perflvl;
 239
 240         struct creg nclk;
 241         struct creg sclk;
 242         struct creg vdec;
 243         struct creg unka0;
 244
 245         struct creg mclk;
 246         u8 *rammap;
 247         u8  rammap_ver;
 248         u8  rammap_len;
 249         u8 *ramcfg;
 250         u8  ramcfg_len;
 251 };
 252
 253 void *
 254 nva3_pm_clocks_pre(struct drm_device *dev, struct nouveau_pm_level *perflvl)
 255 {
 256         struct nva3_pm_state *info;
 257         u8 ramcfg_cnt;
 258         int ret;
 259
 260         info = kzalloc(sizeof(*info), GFP_KERNEL);
 261         if (!info)
 262                 return ERR_PTR(-ENOMEM);
 263
 264         ret = calc_clk(dev, 0x10, 0x4200, perflvl->core, &info->nclk);
 265         if (ret < 0)
 266                 goto out;
 267
 268         ret = calc_clk(dev, 0x11, 0x4220, perflvl->shader, &info->sclk);
 269         if (ret < 0)
 270                 goto out;
 271
 272         ret = calc_clk(dev, 0x12, 0x4000, perflvl->memory, &info->mclk);
 273         if (ret < 0)
 274                 goto out;
 275
 276         ret = calc_clk(dev, 0x20, 0x0000, perflvl->unka0, &info->unka0);
 277         if (ret < 0)
 278                 goto out;
 279
 280         ret = calc_clk(dev, 0x21, 0x0000, perflvl->vdec, &info->vdec);
 281         if (ret < 0)
 282                 goto out;
 283
 284         info->rammap = nouveau_perf_rammap(dev, perflvl->memory,
 285                                            &info->rammap_ver,
 286                                            &info->rammap_len,
 287                                            &ramcfg_cnt, &info->ramcfg_len);
 288         if (info->rammap_ver != 0x10 || info->rammap_len < 5)
 289                 info->rammap = NULL;
 290
 291         info->ramcfg = nouveau_perf_ramcfg(dev, perflvl->memory,
 292                                            &info->rammap_ver,
 293                                            &info->ramcfg_len);
 294         if (info->rammap_ver != 0x10)
 295                 info->ramcfg = NULL;
 296
 297         info->perflvl = perflvl;
 298 out:
 299         if (ret < 0) {
 300                 kfree(info);
 301                 info = ERR_PTR(ret);
 302         }
 303         return info;
 304 }
 305
 306 static bool
 307 nva3_pm_grcp_idle(void *data)
 308 {
 309         struct drm_device *dev = data;
 310
 311         if (!(nv_rd32(dev, 0x400304) & 0x00000001))
 312                 return true;
 313         if (nv_rd32(dev, 0x400308) == 0x0050001c)
 314                 return true;
 315         return false;
 316 }
 317
 318 static void
 319 mclk_precharge(struct nouveau_mem_exec_func *exec)
 320 {
 321         nv_wr32(exec->dev, 0x1002d4, 0x00000001);
 322 }
 323
 324 static void
 325 mclk_refresh(struct nouveau_mem_exec_func *exec)
 326 {
 327         nv_wr32(exec->dev, 0x1002d0, 0x00000001);
 328 }
 329
 330 static void
 331 mclk_refresh_auto(struct nouveau_mem_exec_func *exec, bool enable)
 332 {
 333         nv_wr32(exec->dev, 0x100210, enable ? 0x80000000 : 0x00000000);
 334 }
 335
 336 static void
 337 mclk_refresh_self(struct nouveau_mem_exec_func *exec, bool enable)
 338 {
 339         nv_wr32(exec->dev, 0x1002dc, enable ? 0x00000001 : 0x00000000);
 340 }
 341
 342 static void
 343 mclk_wait(struct nouveau_mem_exec_func *exec, u32 nsec)
 344 {
 345         udelay((nsec + 500) / 1000);
 346 }
 347
 348 static u32
 349 mclk_mrg(struct nouveau_mem_exec_func *exec, int mr)
 350 {
 351         if (mr <= 1)
 352                 return nv_rd32(exec->dev, 0x1002c0 + ((mr - 0) * 4));
 353         if (mr <= 3)
 354                 return nv_rd32(exec->dev, 0x1002e0 + ((mr - 2) * 4));
 355         return 0;
 356 }
 357
 358 static void
 359 mclk_mrs(struct nouveau_mem_exec_func *exec, int mr, u32 data)
 360 {
 361         struct drm_nouveau_private *dev_priv = exec->dev->dev_private;
 362
 363         if (mr <= 1) {
 364                 if (dev_priv->vram_rank_B)
 365                         nv_wr32(exec->dev, 0x1002c8 + ((mr - 0) * 4), data);
 366                 nv_wr32(exec->dev, 0x1002c0 + ((mr - 0) * 4), data);
 367         } else
 368         if (mr <= 3) {
 369                 if (dev_priv->vram_rank_B)
 370                         nv_wr32(exec->dev, 0x1002e8 + ((mr - 2) * 4), data);
 371                 nv_wr32(exec->dev, 0x1002e0 + ((mr - 2) * 4), data);
 372         }
 373 }
 374
 375 static void
 376 mclk_clock_set(struct nouveau_mem_exec_func *exec)
 377 {
 378         struct drm_device *dev = exec->dev;
 379         struct nva3_pm_state *info = exec->priv;
 380
 381         if (!info->mclk.pll) {
 382                 nv_mask(dev, 0x004168, 0x003f3040, info->mclk.clk);
 383                 nv_mask(dev, 0x004000, 0x00000008, 0x00000008);
 384                 nv_mask(dev, 0x1110e0, 0x00088000, 0x00088000);
 385                 nv_wr32(dev, 0x004018, 0x1000d000); /*XXX*/
 386         }
 387
 388         if (info->rammap) {
 389                 if (info->ramcfg && (info->rammap[4] & 0x08)) {
 390                         u32 unk5a0 = (ROM16(info->ramcfg[5]) << 8) |
 391                                       info->ramcfg[5];
 392                         u32 unk5a4 = ROM16(info->ramcfg[7]);
 393                         u32 unk804 = (info->ramcfg[9] & 0xf0) << 16 |
 394                                      (info->ramcfg[3] & 0x0f) << 16 |
 395                                      (info->ramcfg[9] & 0x0f) |
 396                                      0x80000000;
 397                         nv_wr32(dev, 0x1005a0, unk5a0);
 398                         nv_wr32(dev, 0x1005a4, unk5a4);
 399                         nv_wr32(dev, 0x10f804, unk804);
 400                         nv_mask(dev, 0x10053c, 0x00001000, 0x00000000);
 401                 } else {
 402                         nv_mask(dev, 0x10053c, 0x00001000, 0x00001000);
 403                         nv_mask(dev, 0x10f804, 0x80000000, 0x00000000);
 404                 }
 405         }
 406
 407         if (info->mclk.pll) {
 408                 nv_mask(dev, 0x1110e0, 0x00088000, 0x00011000);
 409                 nv_mask(dev, 0x004000, 0x00000008, 0x00000000);
 410         }
 411 }
 412
 413 static void
 414 mclk_timing_set(struct nouveau_mem_exec_func *exec)
 415 {
 416         struct drm_device *dev = exec->dev;
 417         struct nva3_pm_state *info = exec->priv;
 418         struct nouveau_pm_level *perflvl = info->perflvl;
 419         int i;
 420
 421         for (i = 0; i < 9; i++)
 422                 nv_wr32(dev, 0x100220 + (i * 4), perflvl->timing.reg[i]);
 423
 424         if (info->ramcfg) {
 425                 u32 data = (info->ramcfg[2] & 0x08) ? 0x00000000 : 0x00001000;
 426                 nv_mask(dev, 0x100200, 0x00001000, data);
 427         }
 428
 429         if (info->ramcfg) {
 430                 u32 unk714 = nv_rd32(dev, 0x100714) & ~0xf0000010;
 431                 u32 unk718 = nv_rd32(dev, 0x100718) & ~0x00000100;
 432                 u32 unk71c = nv_rd32(dev, 0x10071c) & ~0x00000100;
 433                 if ( (info->ramcfg[2] & 0x20))
 434                         unk714 |= 0xf0000000;
 435                 if (!(info->ramcfg[2] & 0x04))
 436                         unk714 |= 0x00000010;
 437                 nv_wr32(dev, 0x100714, unk714);
 438
 439                 if (info->ramcfg[2] & 0x01)
 440                         unk71c |= 0x00000100;
 441                 nv_wr32(dev, 0x10071c, unk71c);
 442
 443                 if (info->ramcfg[2] & 0x02)
 444                         unk718 |= 0x00000100;
 445                 nv_wr32(dev, 0x100718, unk718);
 446         }
 447 }
 448
 449 static void
 450 prog_mem(struct drm_device *dev, struct nva3_pm_state *info)
 451 {
 452         struct nouveau_mem_exec_func exec = {
 453                 .dev = dev,
 454                 .precharge = mclk_precharge,
 455                 .refresh = mclk_refresh,
 456                 .refresh_auto = mclk_refresh_auto,
 457                 .refresh_self = mclk_refresh_self,
 458                 .wait = mclk_wait,
 459                 .mrg = mclk_mrg,
 460                 .mrs = mclk_mrs,
 461                 .clock_set = mclk_clock_set,
 462                 .timing_set = mclk_timing_set,
 463                 .priv = info
 464         };
 465         u32 ctrl;
 466
 467         ctrl = nv_rd32(dev, 0x004000);
 468         if (ctrl & 0x00000008) {
 469                 if (info->mclk.pll) {
 470                         nv_mask(dev, 0x004128, 0x00000101, 0x00000101);
 471                         nv_wr32(dev, 0x004004, info->mclk.pll);
 472                         nv_wr32(dev, 0x004000, (ctrl |= 0x00000001));
 473                         nv_wr32(dev, 0x004000, (ctrl &= 0xffffffef));
 474                         nv_wait(dev, 0x004000, 0x00020000, 0x00020000);
 475                         nv_wr32(dev, 0x004000, (ctrl |= 0x00000010));
 476                         nv_wr32(dev, 0x004018, 0x00005000); /*XXX*/
 477                         nv_wr32(dev, 0x004000, (ctrl |= 0x00000004));
 478                 }
 479         } else {
 480                 if (!info->mclk.pll) {
 481                         nv_mask(dev, 0x004168, 0x003f3141,
 482                                                0x00000101 | info->mclk.clk);
 483                 }
 484         }
 485
 486         if (info->rammap && !(info->rammap[4] & 0x02))
 487                 nv_mask(dev, 0x100200, 0x00000800, 0x00000000);
 488         nv_wr32(dev, 0x611200, 0x00003300);
 489
 490         nouveau_mem_exec(&exec, info->perflvl);
 491
 492         nv_wr32(dev, 0x611200, 0x00003330);
 493         if (info->rammap && (info->rammap[4] & 0x02))
 494                 nv_mask(dev, 0x100200, 0x00000800, 0x00000800);
 495
 496         if (info->mclk.pll) {
 497                 nv_mask(dev, 0x004168, 0x00000001, 0x00000000);
 498                 nv_mask(dev, 0x004168, 0x00000100, 0x00000000);
 499         } else {
 500                 nv_mask(dev, 0x004000, 0x00000001, 0x00000000);
 501                 nv_mask(dev, 0x004128, 0x00000001, 0x00000000);
 502                 nv_mask(dev, 0x004128, 0x00000100, 0x00000000);
 503         }
 504 }
 505
 506 int
 507 nva3_pm_clocks_set(struct drm_device *dev, void *pre_state)
 508 {
 509         struct drm_nouveau_private *dev_priv = dev->dev_private;
 510         struct nva3_pm_state *info = pre_state;
 511         unsigned long flags;
 512         int ret = -EAGAIN;
 513
 514         /* prevent any new grctx switches from starting */
 515         spin_lock_irqsave(&dev_priv->context_switch_lock, flags);
 516         nv_wr32(dev, 0x400324, 0x00000000);
 517         nv_wr32(dev, 0x400328, 0x0050001c); /* wait flag 0x1c */
 518         /* wait for any pending grctx switches to complete */
 519         if (!nv_wait_cb(dev, nva3_pm_grcp_idle, dev)) {
 520                 NV_ERROR(dev, "pm: ctxprog didn't go idle\n");
 521                 goto cleanup;
 522         }
 523         /* freeze PFIFO */
 524         nv_mask(dev, 0x002504, 0x00000001, 0x00000001);
 525         if (!nv_wait(dev, 0x002504, 0x00000010, 0x00000010)) {
 526                 NV_ERROR(dev, "pm: fifo didn't go idle\n");
 527                 goto cleanup;
 528         }
 529
 530         prog_pll(dev, 0x00, 0x004200, &info->nclk);
 531         prog_pll(dev, 0x01, 0x004220, &info->sclk);
 532         prog_clk(dev, 0x20, &info->unka0);
 533         prog_clk(dev, 0x21, &info->vdec);
 534
 535         if (info->mclk.clk || info->mclk.pll)
 536                 prog_mem(dev, info);
 537
 538         ret = 0;
 539
 540 cleanup:
 541         /* unfreeze PFIFO */
 542         nv_mask(dev, 0x002504, 0x00000001, 0x00000000);
 543         /* restore ctxprog to normal */
 544         nv_wr32(dev, 0x400324, 0x00000000);
 545         nv_wr32(dev, 0x400328, 0x0070009c); /* set flag 0x1c */
 546         /* unblock it if necessary */
 547         if (nv_rd32(dev, 0x400308) == 0x0050001c)
 548                 nv_mask(dev, 0x400824, 0x10000000, 0x10000000);
 549         spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
 550         kfree(info);
 551         return ret;
 552 }