1 /* 2 * Copyright (c) 2016, NVIDIA CORPORATION. All rights reserved. 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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 20 * DEALINGS IN THE SOFTWARE. 21 */ 22 23 #include <subdev/clk.h> 24 #include <subdev/volt.h> 25 #include <subdev/timer.h> 26 #include <core/device.h> 27 #include <core/tegra.h> 28 29 #include "priv.h" 30 #include "gk20a.h" 31 32 #define GPCPLL_CFG_SYNC_MODE BIT(2) 33 34 #define BYPASSCTRL_SYS (SYS_GPCPLL_CFG_BASE + 0x340) 35 #define BYPASSCTRL_SYS_GPCPLL_SHIFT 0 36 #define BYPASSCTRL_SYS_GPCPLL_WIDTH 1 37 38 #define GPCPLL_CFG2_SDM_DIN_SHIFT 0 39 #define GPCPLL_CFG2_SDM_DIN_WIDTH 8 40 #define GPCPLL_CFG2_SDM_DIN_MASK \ 41 (MASK(GPCPLL_CFG2_SDM_DIN_WIDTH) << GPCPLL_CFG2_SDM_DIN_SHIFT) 42 #define GPCPLL_CFG2_SDM_DIN_NEW_SHIFT 8 43 #define GPCPLL_CFG2_SDM_DIN_NEW_WIDTH 15 44 #define GPCPLL_CFG2_SDM_DIN_NEW_MASK \ 45 (MASK(GPCPLL_CFG2_SDM_DIN_NEW_WIDTH) << GPCPLL_CFG2_SDM_DIN_NEW_SHIFT) 46 #define GPCPLL_CFG2_SETUP2_SHIFT 16 47 #define GPCPLL_CFG2_PLL_STEPA_SHIFT 24 48 49 #define GPCPLL_DVFS0 (SYS_GPCPLL_CFG_BASE + 0x10) 50 #define GPCPLL_DVFS0_DFS_COEFF_SHIFT 0 51 #define GPCPLL_DVFS0_DFS_COEFF_WIDTH 7 52 #define GPCPLL_DVFS0_DFS_COEFF_MASK \ 53 (MASK(GPCPLL_DVFS0_DFS_COEFF_WIDTH) << GPCPLL_DVFS0_DFS_COEFF_SHIFT) 54 #define GPCPLL_DVFS0_DFS_DET_MAX_SHIFT 8 55 #define GPCPLL_DVFS0_DFS_DET_MAX_WIDTH 7 56 #define GPCPLL_DVFS0_DFS_DET_MAX_MASK \ 57 (MASK(GPCPLL_DVFS0_DFS_DET_MAX_WIDTH) << GPCPLL_DVFS0_DFS_DET_MAX_SHIFT) 58 59 #define GPCPLL_DVFS1 (SYS_GPCPLL_CFG_BASE + 0x14) 60 #define GPCPLL_DVFS1_DFS_EXT_DET_SHIFT 0 61 #define GPCPLL_DVFS1_DFS_EXT_DET_WIDTH 7 62 #define GPCPLL_DVFS1_DFS_EXT_STRB_SHIFT 7 63 #define GPCPLL_DVFS1_DFS_EXT_STRB_WIDTH 1 64 #define GPCPLL_DVFS1_DFS_EXT_CAL_SHIFT 8 65 #define GPCPLL_DVFS1_DFS_EXT_CAL_WIDTH 7 66 #define GPCPLL_DVFS1_DFS_EXT_SEL_SHIFT 15 67 #define GPCPLL_DVFS1_DFS_EXT_SEL_WIDTH 1 68 #define GPCPLL_DVFS1_DFS_CTRL_SHIFT 16 69 #define GPCPLL_DVFS1_DFS_CTRL_WIDTH 12 70 #define GPCPLL_DVFS1_EN_SDM_SHIFT 28 71 #define GPCPLL_DVFS1_EN_SDM_WIDTH 1 72 #define GPCPLL_DVFS1_EN_SDM_BIT BIT(28) 73 #define GPCPLL_DVFS1_EN_DFS_SHIFT 29 74 #define GPCPLL_DVFS1_EN_DFS_WIDTH 1 75 #define GPCPLL_DVFS1_EN_DFS_BIT BIT(29) 76 #define GPCPLL_DVFS1_EN_DFS_CAL_SHIFT 30 77 #define GPCPLL_DVFS1_EN_DFS_CAL_WIDTH 1 78 #define GPCPLL_DVFS1_EN_DFS_CAL_BIT BIT(30) 79 #define GPCPLL_DVFS1_DFS_CAL_DONE_SHIFT 31 80 #define GPCPLL_DVFS1_DFS_CAL_DONE_WIDTH 1 81 #define GPCPLL_DVFS1_DFS_CAL_DONE_BIT BIT(31) 82 83 #define GPC_BCAST_GPCPLL_DVFS2 (GPC_BCAST_GPCPLL_CFG_BASE + 0x20) 84 #define GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT BIT(16) 85 86 #define GPCPLL_CFG3_PLL_DFS_TESTOUT_SHIFT 24 87 #define GPCPLL_CFG3_PLL_DFS_TESTOUT_WIDTH 7 88 89 #define DFS_DET_RANGE 6 /* -2^6 ... 2^6-1 */ 90 #define SDM_DIN_RANGE 12 /* -2^12 ... 2^12-1 */ 91 92 struct gm20b_clk_dvfs_params { 93 s32 coeff_slope; 94 s32 coeff_offs; 95 u32 vco_ctrl; 96 }; 97 98 static const struct gm20b_clk_dvfs_params gm20b_dvfs_params = { 99 .coeff_slope = -165230, 100 .coeff_offs = 214007, 101 .vco_ctrl = 0x7 << 3, 102 }; 103 104 /* 105 * base.n is now the *integer* part of the N factor. 106 * sdm_din contains n's decimal part. 107 */ 108 struct gm20b_pll { 109 struct gk20a_pll base; 110 u32 sdm_din; 111 }; 112 113 struct gm20b_clk_dvfs { 114 u32 dfs_coeff; 115 s32 dfs_det_max; 116 s32 dfs_ext_cal; 117 }; 118 119 struct gm20b_clk { 120 /* currently applied parameters */ 121 struct gk20a_clk base; 122 struct gm20b_clk_dvfs dvfs; 123 u32 uv; 124 125 /* new parameters to apply */ 126 struct gk20a_pll new_pll; 127 struct gm20b_clk_dvfs new_dvfs; 128 u32 new_uv; 129 130 const struct gm20b_clk_dvfs_params *dvfs_params; 131 132 /* fused parameters */ 133 s32 uvdet_slope; 134 s32 uvdet_offs; 135 136 /* safe frequency we can use at minimum voltage */ 137 u32 safe_fmax_vmin; 138 }; 139 #define gm20b_clk(p) container_of((gk20a_clk(p)), struct gm20b_clk, base) 140 141 static u32 pl_to_div(u32 pl) 142 { 143 return pl; 144 } 145 146 static u32 div_to_pl(u32 div) 147 { 148 return div; 149 } 150 151 static const struct gk20a_clk_pllg_params gm20b_pllg_params = { 152 .min_vco = 1300000, .max_vco = 2600000, 153 .min_u = 12000, .max_u = 38400, 154 .min_m = 1, .max_m = 255, 155 .min_n = 8, .max_n = 255, 156 .min_pl = 1, .max_pl = 31, 157 }; 158 159 static void 160 gm20b_pllg_read_mnp(struct gm20b_clk *clk, struct gm20b_pll *pll) 161 { 162 struct nvkm_subdev *subdev = &clk->base.base.subdev; 163 struct nvkm_device *device = subdev->device; 164 u32 val; 165 166 gk20a_pllg_read_mnp(&clk->base, &pll->base); 167 val = nvkm_rd32(device, GPCPLL_CFG2); 168 pll->sdm_din = (val >> GPCPLL_CFG2_SDM_DIN_SHIFT) & 169 MASK(GPCPLL_CFG2_SDM_DIN_WIDTH); 170 } 171 172 static void 173 gm20b_pllg_write_mnp(struct gm20b_clk *clk, const struct gm20b_pll *pll) 174 { 175 struct nvkm_device *device = clk->base.base.subdev.device; 176 177 nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_MASK, 178 pll->sdm_din << GPCPLL_CFG2_SDM_DIN_SHIFT); 179 gk20a_pllg_write_mnp(&clk->base, &pll->base); 180 } 181 182 /* 183 * Determine DFS_COEFF for the requested voltage. Always select external 184 * calibration override equal to the voltage, and set maximum detection 185 * limit "0" (to make sure that PLL output remains under F/V curve when 186 * voltage increases). 187 */ 188 static void 189 gm20b_dvfs_calc_det_coeff(struct gm20b_clk *clk, s32 uv, 190 struct gm20b_clk_dvfs *dvfs) 191 { 192 struct nvkm_subdev *subdev = &clk->base.base.subdev; 193 const struct gm20b_clk_dvfs_params *p = clk->dvfs_params; 194 u32 coeff; 195 /* Work with mv as uv would likely trigger an overflow */ 196 s32 mv = DIV_ROUND_CLOSEST(uv, 1000); 197 198 /* coeff = slope * voltage + offset */ 199 coeff = DIV_ROUND_CLOSEST(mv * p->coeff_slope, 1000) + p->coeff_offs; 200 coeff = DIV_ROUND_CLOSEST(coeff, 1000); 201 dvfs->dfs_coeff = min_t(u32, coeff, MASK(GPCPLL_DVFS0_DFS_COEFF_WIDTH)); 202 203 dvfs->dfs_ext_cal = DIV_ROUND_CLOSEST(uv - clk->uvdet_offs, 204 clk->uvdet_slope); 205 /* should never happen */ 206 if (abs(dvfs->dfs_ext_cal) >= BIT(DFS_DET_RANGE)) 207 nvkm_error(subdev, "dfs_ext_cal overflow!\n"); 208 209 dvfs->dfs_det_max = 0; 210 211 nvkm_debug(subdev, "%s uv: %d coeff: %x, ext_cal: %d, det_max: %d\n", 212 __func__, uv, dvfs->dfs_coeff, dvfs->dfs_ext_cal, 213 dvfs->dfs_det_max); 214 } 215 216 /* 217 * Solve equation for integer and fractional part of the effective NDIV: 218 * 219 * n_eff = n_int + 1/2 + (SDM_DIN / 2^(SDM_DIN_RANGE + 1)) + 220 * (DVFS_COEFF * DVFS_DET_DELTA) / 2^DFS_DET_RANGE 221 * 222 * The SDM_DIN LSB is finally shifted out, since it is not accessible by sw. 223 */ 224 static void 225 gm20b_dvfs_calc_ndiv(struct gm20b_clk *clk, u32 n_eff, u32 *n_int, u32 *sdm_din) 226 { 227 struct nvkm_subdev *subdev = &clk->base.base.subdev; 228 const struct gk20a_clk_pllg_params *p = clk->base.params; 229 u32 n; 230 s32 det_delta; 231 u32 rem, rem_range; 232 233 /* calculate current ext_cal and subtract previous one */ 234 det_delta = DIV_ROUND_CLOSEST(((s32)clk->uv) - clk->uvdet_offs, 235 clk->uvdet_slope); 236 det_delta -= clk->dvfs.dfs_ext_cal; 237 det_delta = min(det_delta, clk->dvfs.dfs_det_max); 238 det_delta *= clk->dvfs.dfs_coeff; 239 240 /* integer part of n */ 241 n = (n_eff << DFS_DET_RANGE) - det_delta; 242 /* should never happen! */ 243 if (n <= 0) { 244 nvkm_error(subdev, "ndiv <= 0 - setting to 1...\n"); 245 n = 1 << DFS_DET_RANGE; 246 } 247 if (n >> DFS_DET_RANGE > p->max_n) { 248 nvkm_error(subdev, "ndiv > max_n - setting to max_n...\n"); 249 n = p->max_n << DFS_DET_RANGE; 250 } 251 *n_int = n >> DFS_DET_RANGE; 252 253 /* fractional part of n */ 254 rem = ((u32)n) & MASK(DFS_DET_RANGE); 255 rem_range = SDM_DIN_RANGE + 1 - DFS_DET_RANGE; 256 /* subtract 2^SDM_DIN_RANGE to account for the 1/2 of the equation */ 257 rem = (rem << rem_range) - BIT(SDM_DIN_RANGE); 258 /* lose 8 LSB and clip - sdm_din only keeps the most significant byte */ 259 *sdm_din = (rem >> BITS_PER_BYTE) & MASK(GPCPLL_CFG2_SDM_DIN_WIDTH); 260 261 nvkm_debug(subdev, "%s n_eff: %d, n_int: %d, sdm_din: %d\n", __func__, 262 n_eff, *n_int, *sdm_din); 263 } 264 265 static int 266 gm20b_pllg_slide(struct gm20b_clk *clk, u32 n) 267 { 268 struct nvkm_subdev *subdev = &clk->base.base.subdev; 269 struct nvkm_device *device = subdev->device; 270 struct gm20b_pll pll; 271 u32 n_int, sdm_din; 272 int ret = 0; 273 274 /* calculate the new n_int/sdm_din for this n/uv */ 275 gm20b_dvfs_calc_ndiv(clk, n, &n_int, &sdm_din); 276 277 /* get old coefficients */ 278 gm20b_pllg_read_mnp(clk, &pll); 279 /* do nothing if NDIV is the same */ 280 if (n_int == pll.base.n && sdm_din == pll.sdm_din) 281 return 0; 282 283 /* pll slowdown mode */ 284 nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN, 285 BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT), 286 BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT)); 287 288 /* new ndiv ready for ramp */ 289 /* in DVFS mode SDM is updated via "new" field */ 290 nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_NEW_MASK, 291 sdm_din << GPCPLL_CFG2_SDM_DIN_NEW_SHIFT); 292 pll.base.n = n_int; 293 udelay(1); 294 gk20a_pllg_write_mnp(&clk->base, &pll.base); 295 296 /* dynamic ramp to new ndiv */ 297 udelay(1); 298 nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN, 299 BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT), 300 BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT)); 301 302 /* wait for ramping to complete */ 303 if (nvkm_wait_usec(device, 500, GPC_BCAST_NDIV_SLOWDOWN_DEBUG, 304 GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK, 305 GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK) < 0) 306 ret = -ETIMEDOUT; 307 308 /* in DVFS mode complete SDM update */ 309 nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_MASK, 310 sdm_din << GPCPLL_CFG2_SDM_DIN_SHIFT); 311 312 /* exit slowdown mode */ 313 nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN, 314 BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT) | 315 BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT), 0); 316 nvkm_rd32(device, GPCPLL_NDIV_SLOWDOWN); 317 318 return ret; 319 } 320 321 static int 322 gm20b_pllg_enable(struct gm20b_clk *clk) 323 { 324 struct nvkm_device *device = clk->base.base.subdev.device; 325 326 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, GPCPLL_CFG_ENABLE); 327 nvkm_rd32(device, GPCPLL_CFG); 328 329 /* In DVFS mode lock cannot be used - so just delay */ 330 udelay(40); 331 332 /* set SYNC_MODE for glitchless switch out of bypass */ 333 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_SYNC_MODE, 334 GPCPLL_CFG_SYNC_MODE); 335 nvkm_rd32(device, GPCPLL_CFG); 336 337 /* switch to VCO mode */ 338 nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT), 339 BIT(SEL_VCO_GPC2CLK_OUT_SHIFT)); 340 341 return 0; 342 } 343 344 static void 345 gm20b_pllg_disable(struct gm20b_clk *clk) 346 { 347 struct nvkm_device *device = clk->base.base.subdev.device; 348 349 /* put PLL in bypass before disabling it */ 350 nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT), 0); 351 352 /* clear SYNC_MODE before disabling PLL */ 353 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_SYNC_MODE, 0); 354 355 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, 0); 356 nvkm_rd32(device, GPCPLL_CFG); 357 } 358 359 static int 360 gm20b_pllg_program_mnp(struct gm20b_clk *clk, const struct gk20a_pll *pll) 361 { 362 struct nvkm_subdev *subdev = &clk->base.base.subdev; 363 struct nvkm_device *device = subdev->device; 364 struct gm20b_pll cur_pll; 365 u32 n_int, sdm_din; 366 /* if we only change pdiv, we can do a glitchless transition */ 367 bool pdiv_only; 368 int ret; 369 370 gm20b_dvfs_calc_ndiv(clk, pll->n, &n_int, &sdm_din); 371 gm20b_pllg_read_mnp(clk, &cur_pll); 372 pdiv_only = cur_pll.base.n == n_int && cur_pll.sdm_din == sdm_din && 373 cur_pll.base.m == pll->m; 374 375 /* need full sequence if clock not enabled yet */ 376 if (!gk20a_pllg_is_enabled(&clk->base)) 377 pdiv_only = false; 378 379 /* split VCO-to-bypass jump in half by setting out divider 1:2 */ 380 nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK, 381 GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT); 382 /* Intentional 2nd write to assure linear divider operation */ 383 nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK, 384 GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT); 385 nvkm_rd32(device, GPC2CLK_OUT); 386 udelay(2); 387 388 if (pdiv_only) { 389 u32 old = cur_pll.base.pl; 390 u32 new = pll->pl; 391 392 /* 393 * we can do a glitchless transition only if the old and new PL 394 * parameters share at least one bit set to 1. If this is not 395 * the case, calculate and program an interim PL that will allow 396 * us to respect that rule. 397 */ 398 if ((old & new) == 0) { 399 cur_pll.base.pl = min(old | BIT(ffs(new) - 1), 400 new | BIT(ffs(old) - 1)); 401 gk20a_pllg_write_mnp(&clk->base, &cur_pll.base); 402 } 403 404 cur_pll.base.pl = new; 405 gk20a_pllg_write_mnp(&clk->base, &cur_pll.base); 406 } else { 407 /* disable before programming if more than pdiv changes */ 408 gm20b_pllg_disable(clk); 409 410 cur_pll.base = *pll; 411 cur_pll.base.n = n_int; 412 cur_pll.sdm_din = sdm_din; 413 gm20b_pllg_write_mnp(clk, &cur_pll); 414 415 ret = gm20b_pllg_enable(clk); 416 if (ret) 417 return ret; 418 } 419 420 /* restore out divider 1:1 */ 421 udelay(2); 422 nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK, 423 GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT); 424 /* Intentional 2nd write to assure linear divider operation */ 425 nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK, 426 GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT); 427 nvkm_rd32(device, GPC2CLK_OUT); 428 429 return 0; 430 } 431 432 static int 433 gm20b_pllg_program_mnp_slide(struct gm20b_clk *clk, const struct gk20a_pll *pll) 434 { 435 struct gk20a_pll cur_pll; 436 int ret; 437 438 if (gk20a_pllg_is_enabled(&clk->base)) { 439 gk20a_pllg_read_mnp(&clk->base, &cur_pll); 440 441 /* just do NDIV slide if there is no change to M and PL */ 442 if (pll->m == cur_pll.m && pll->pl == cur_pll.pl) 443 return gm20b_pllg_slide(clk, pll->n); 444 445 /* slide down to current NDIV_LO */ 446 cur_pll.n = gk20a_pllg_n_lo(&clk->base, &cur_pll); 447 ret = gm20b_pllg_slide(clk, cur_pll.n); 448 if (ret) 449 return ret; 450 } 451 452 /* program MNP with the new clock parameters and new NDIV_LO */ 453 cur_pll = *pll; 454 cur_pll.n = gk20a_pllg_n_lo(&clk->base, &cur_pll); 455 ret = gm20b_pllg_program_mnp(clk, &cur_pll); 456 if (ret) 457 return ret; 458 459 /* slide up to new NDIV */ 460 return gm20b_pllg_slide(clk, pll->n); 461 } 462 463 static int 464 gm20b_clk_calc(struct nvkm_clk *base, struct nvkm_cstate *cstate) 465 { 466 struct gm20b_clk *clk = gm20b_clk(base); 467 struct nvkm_subdev *subdev = &base->subdev; 468 struct nvkm_volt *volt = base->subdev.device->volt; 469 int ret; 470 471 ret = gk20a_pllg_calc_mnp(&clk->base, cstate->domain[nv_clk_src_gpc] * 472 GK20A_CLK_GPC_MDIV, &clk->new_pll); 473 if (ret) 474 return ret; 475 476 clk->new_uv = volt->vid[cstate->voltage].uv; 477 gm20b_dvfs_calc_det_coeff(clk, clk->new_uv, &clk->new_dvfs); 478 479 nvkm_debug(subdev, "%s uv: %d uv\n", __func__, clk->new_uv); 480 481 return 0; 482 } 483 484 /* 485 * Compute PLL parameters that are always safe for the current voltage 486 */ 487 static void 488 gm20b_dvfs_calc_safe_pll(struct gm20b_clk *clk, struct gk20a_pll *pll) 489 { 490 u32 rate = gk20a_pllg_calc_rate(&clk->base, pll) / KHZ; 491 u32 parent_rate = clk->base.parent_rate / KHZ; 492 u32 nmin, nsafe; 493 494 /* remove a safe margin of 10% */ 495 if (rate > clk->safe_fmax_vmin) 496 rate = rate * (100 - 10) / 100; 497 498 /* gpc2clk */ 499 rate *= 2; 500 501 nmin = DIV_ROUND_UP(pll->m * clk->base.params->min_vco, parent_rate); 502 nsafe = pll->m * rate / (clk->base.parent_rate); 503 504 if (nsafe < nmin) { 505 pll->pl = DIV_ROUND_UP(nmin * parent_rate, pll->m * rate); 506 nsafe = nmin; 507 } 508 509 pll->n = nsafe; 510 } 511 512 static void 513 gm20b_dvfs_program_coeff(struct gm20b_clk *clk, u32 coeff) 514 { 515 struct nvkm_device *device = clk->base.base.subdev.device; 516 517 /* strobe to read external DFS coefficient */ 518 nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, 519 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 520 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT); 521 522 nvkm_mask(device, GPCPLL_DVFS0, GPCPLL_DVFS0_DFS_COEFF_MASK, 523 coeff << GPCPLL_DVFS0_DFS_COEFF_SHIFT); 524 525 udelay(1); 526 nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, 527 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 0); 528 } 529 530 static void 531 gm20b_dvfs_program_ext_cal(struct gm20b_clk *clk, u32 dfs_det_cal) 532 { 533 struct nvkm_device *device = clk->base.base.subdev.device; 534 u32 val; 535 536 nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, MASK(DFS_DET_RANGE + 1), 537 dfs_det_cal); 538 udelay(1); 539 540 val = nvkm_rd32(device, GPCPLL_DVFS1); 541 if (!(val & BIT(25))) { 542 /* Use external value to overwrite calibration value */ 543 val |= BIT(25) | BIT(16); 544 nvkm_wr32(device, GPCPLL_DVFS1, val); 545 } 546 } 547 548 static void 549 gm20b_dvfs_program_dfs_detection(struct gm20b_clk *clk, 550 struct gm20b_clk_dvfs *dvfs) 551 { 552 struct nvkm_device *device = clk->base.base.subdev.device; 553 554 /* strobe to read external DFS coefficient */ 555 nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, 556 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 557 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT); 558 559 nvkm_mask(device, GPCPLL_DVFS0, 560 GPCPLL_DVFS0_DFS_COEFF_MASK | GPCPLL_DVFS0_DFS_DET_MAX_MASK, 561 dvfs->dfs_coeff << GPCPLL_DVFS0_DFS_COEFF_SHIFT | 562 dvfs->dfs_det_max << GPCPLL_DVFS0_DFS_DET_MAX_SHIFT); 563 564 udelay(1); 565 nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, 566 GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 0); 567 568 gm20b_dvfs_program_ext_cal(clk, dvfs->dfs_ext_cal); 569 } 570 571 static int 572 gm20b_clk_prog(struct nvkm_clk *base) 573 { 574 struct gm20b_clk *clk = gm20b_clk(base); 575 u32 cur_freq; 576 int ret; 577 578 /* No change in DVFS settings? */ 579 if (clk->uv == clk->new_uv) 580 goto prog; 581 582 /* 583 * Interim step for changing DVFS detection settings: low enough 584 * frequency to be safe at at DVFS coeff = 0. 585 * 586 * 1. If voltage is increasing: 587 * - safe frequency target matches the lowest - old - frequency 588 * - DVFS settings are still old 589 * - Voltage already increased to new level by volt, but maximum 590 * detection limit assures PLL output remains under F/V curve 591 * 592 * 2. If voltage is decreasing: 593 * - safe frequency target matches the lowest - new - frequency 594 * - DVFS settings are still old 595 * - Voltage is also old, it will be lowered by volt afterwards 596 * 597 * Interim step can be skipped if old frequency is below safe minimum, 598 * i.e., it is low enough to be safe at any voltage in operating range 599 * with zero DVFS coefficient. 600 */ 601 cur_freq = nvkm_clk_read(&clk->base.base, nv_clk_src_gpc); 602 if (cur_freq > clk->safe_fmax_vmin) { 603 struct gk20a_pll pll_safe; 604 605 if (clk->uv < clk->new_uv) 606 /* voltage will raise: safe frequency is current one */ 607 pll_safe = clk->base.pll; 608 else 609 /* voltage will drop: safe frequency is new one */ 610 pll_safe = clk->new_pll; 611 612 gm20b_dvfs_calc_safe_pll(clk, &pll_safe); 613 ret = gm20b_pllg_program_mnp_slide(clk, &pll_safe); 614 if (ret) 615 return ret; 616 } 617 618 /* 619 * DVFS detection settings transition: 620 * - Set DVFS coefficient zero 621 * - Set calibration level to new voltage 622 * - Set DVFS coefficient to match new voltage 623 */ 624 gm20b_dvfs_program_coeff(clk, 0); 625 gm20b_dvfs_program_ext_cal(clk, clk->new_dvfs.dfs_ext_cal); 626 gm20b_dvfs_program_coeff(clk, clk->new_dvfs.dfs_coeff); 627 gm20b_dvfs_program_dfs_detection(clk, &clk->new_dvfs); 628 629 prog: 630 clk->uv = clk->new_uv; 631 clk->dvfs = clk->new_dvfs; 632 clk->base.pll = clk->new_pll; 633 634 return gm20b_pllg_program_mnp_slide(clk, &clk->base.pll); 635 } 636 637 static struct nvkm_pstate 638 gm20b_pstates[] = { 639 { 640 .base = { 641 .domain[nv_clk_src_gpc] = 76800, 642 .voltage = 0, 643 }, 644 }, 645 { 646 .base = { 647 .domain[nv_clk_src_gpc] = 153600, 648 .voltage = 1, 649 }, 650 }, 651 { 652 .base = { 653 .domain[nv_clk_src_gpc] = 230400, 654 .voltage = 2, 655 }, 656 }, 657 { 658 .base = { 659 .domain[nv_clk_src_gpc] = 307200, 660 .voltage = 3, 661 }, 662 }, 663 { 664 .base = { 665 .domain[nv_clk_src_gpc] = 384000, 666 .voltage = 4, 667 }, 668 }, 669 { 670 .base = { 671 .domain[nv_clk_src_gpc] = 460800, 672 .voltage = 5, 673 }, 674 }, 675 { 676 .base = { 677 .domain[nv_clk_src_gpc] = 537600, 678 .voltage = 6, 679 }, 680 }, 681 { 682 .base = { 683 .domain[nv_clk_src_gpc] = 614400, 684 .voltage = 7, 685 }, 686 }, 687 { 688 .base = { 689 .domain[nv_clk_src_gpc] = 691200, 690 .voltage = 8, 691 }, 692 }, 693 { 694 .base = { 695 .domain[nv_clk_src_gpc] = 768000, 696 .voltage = 9, 697 }, 698 }, 699 { 700 .base = { 701 .domain[nv_clk_src_gpc] = 844800, 702 .voltage = 10, 703 }, 704 }, 705 { 706 .base = { 707 .domain[nv_clk_src_gpc] = 921600, 708 .voltage = 11, 709 }, 710 }, 711 { 712 .base = { 713 .domain[nv_clk_src_gpc] = 998400, 714 .voltage = 12, 715 }, 716 }, 717 }; 718 719 static void 720 gm20b_clk_fini(struct nvkm_clk *base) 721 { 722 struct nvkm_device *device = base->subdev.device; 723 struct gm20b_clk *clk = gm20b_clk(base); 724 725 /* slide to VCO min */ 726 if (gk20a_pllg_is_enabled(&clk->base)) { 727 struct gk20a_pll pll; 728 u32 n_lo; 729 730 gk20a_pllg_read_mnp(&clk->base, &pll); 731 n_lo = gk20a_pllg_n_lo(&clk->base, &pll); 732 gm20b_pllg_slide(clk, n_lo); 733 } 734 735 gm20b_pllg_disable(clk); 736 737 /* set IDDQ */ 738 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 1); 739 } 740 741 static int 742 gm20b_clk_init_dvfs(struct gm20b_clk *clk) 743 { 744 struct nvkm_subdev *subdev = &clk->base.base.subdev; 745 struct nvkm_device *device = subdev->device; 746 bool fused = clk->uvdet_offs && clk->uvdet_slope; 747 static const s32 ADC_SLOPE_UV = 10000; /* default ADC detection slope */ 748 u32 data; 749 int ret; 750 751 /* Enable NA DVFS */ 752 nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_BIT, 753 GPCPLL_DVFS1_EN_DFS_BIT); 754 755 /* Set VCO_CTRL */ 756 if (clk->dvfs_params->vco_ctrl) 757 nvkm_mask(device, GPCPLL_CFG3, GPCPLL_CFG3_VCO_CTRL_MASK, 758 clk->dvfs_params->vco_ctrl << GPCPLL_CFG3_VCO_CTRL_SHIFT); 759 760 if (fused) { 761 /* Start internal calibration, but ignore results */ 762 nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_CAL_BIT, 763 GPCPLL_DVFS1_EN_DFS_CAL_BIT); 764 765 /* got uvdev parameters from fuse, skip calibration */ 766 goto calibrated; 767 } 768 769 /* 770 * If calibration parameters are not fused, start internal calibration, 771 * wait for completion, and use results along with default slope to 772 * calculate ADC offset during boot. 773 */ 774 nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_CAL_BIT, 775 GPCPLL_DVFS1_EN_DFS_CAL_BIT); 776 777 /* Wait for internal calibration done (spec < 2us). */ 778 ret = nvkm_wait_usec(device, 10, GPCPLL_DVFS1, 779 GPCPLL_DVFS1_DFS_CAL_DONE_BIT, 780 GPCPLL_DVFS1_DFS_CAL_DONE_BIT); 781 if (ret < 0) { 782 nvkm_error(subdev, "GPCPLL calibration timeout\n"); 783 return -ETIMEDOUT; 784 } 785 786 data = nvkm_rd32(device, GPCPLL_CFG3) >> 787 GPCPLL_CFG3_PLL_DFS_TESTOUT_SHIFT; 788 data &= MASK(GPCPLL_CFG3_PLL_DFS_TESTOUT_WIDTH); 789 790 clk->uvdet_slope = ADC_SLOPE_UV; 791 clk->uvdet_offs = ((s32)clk->uv) - data * ADC_SLOPE_UV; 792 793 nvkm_debug(subdev, "calibrated DVFS parameters: offs %d, slope %d\n", 794 clk->uvdet_offs, clk->uvdet_slope); 795 796 calibrated: 797 /* Compute and apply initial DVFS parameters */ 798 gm20b_dvfs_calc_det_coeff(clk, clk->uv, &clk->dvfs); 799 gm20b_dvfs_program_coeff(clk, 0); 800 gm20b_dvfs_program_ext_cal(clk, clk->dvfs.dfs_ext_cal); 801 gm20b_dvfs_program_coeff(clk, clk->dvfs.dfs_coeff); 802 gm20b_dvfs_program_dfs_detection(clk, &clk->new_dvfs); 803 804 return 0; 805 } 806 807 /* Forward declaration to detect speedo >=1 in gm20b_clk_init() */ 808 static const struct nvkm_clk_func gm20b_clk; 809 810 static int 811 gm20b_clk_init(struct nvkm_clk *base) 812 { 813 struct gk20a_clk *clk = gk20a_clk(base); 814 struct nvkm_subdev *subdev = &clk->base.subdev; 815 struct nvkm_device *device = subdev->device; 816 int ret; 817 u32 data; 818 819 /* get out from IDDQ */ 820 nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 0); 821 nvkm_rd32(device, GPCPLL_CFG); 822 udelay(5); 823 824 nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_INIT_MASK, 825 GPC2CLK_OUT_INIT_VAL); 826 827 /* Set the global bypass control to VCO */ 828 nvkm_mask(device, BYPASSCTRL_SYS, 829 MASK(BYPASSCTRL_SYS_GPCPLL_WIDTH) << BYPASSCTRL_SYS_GPCPLL_SHIFT, 830 0); 831 832 ret = gk20a_clk_setup_slide(clk); 833 if (ret) 834 return ret; 835 836 /* If not fused, set RAM SVOP PDP data 0x2, and enable fuse override */ 837 data = nvkm_rd32(device, 0x021944); 838 if (!(data & 0x3)) { 839 data |= 0x2; 840 nvkm_wr32(device, 0x021944, data); 841 842 data = nvkm_rd32(device, 0x021948); 843 data |= 0x1; 844 nvkm_wr32(device, 0x021948, data); 845 } 846 847 /* Disable idle slow down */ 848 nvkm_mask(device, 0x20160, 0x003f0000, 0x0); 849 850 /* speedo >= 1? */ 851 if (clk->base.func == &gm20b_clk) { 852 struct gm20b_clk *_clk = gm20b_clk(base); 853 struct nvkm_volt *volt = device->volt; 854 855 /* Get current voltage */ 856 _clk->uv = nvkm_volt_get(volt); 857 858 /* Initialize DVFS */ 859 ret = gm20b_clk_init_dvfs(_clk); 860 if (ret) 861 return ret; 862 } 863 864 /* Start with lowest frequency */ 865 base->func->calc(base, &base->func->pstates[0].base); 866 ret = base->func->prog(base); 867 if (ret) { 868 nvkm_error(subdev, "cannot initialize clock\n"); 869 return ret; 870 } 871 872 return 0; 873 } 874 875 static const struct nvkm_clk_func 876 gm20b_clk_speedo0 = { 877 .init = gm20b_clk_init, 878 .fini = gk20a_clk_fini, 879 .read = gk20a_clk_read, 880 .calc = gk20a_clk_calc, 881 .prog = gk20a_clk_prog, 882 .tidy = gk20a_clk_tidy, 883 .pstates = gm20b_pstates, 884 /* Speedo 0 only supports 12 voltages */ 885 .nr_pstates = ARRAY_SIZE(gm20b_pstates) - 1, 886 .domains = { 887 { nv_clk_src_crystal, 0xff }, 888 { nv_clk_src_gpc, 0xff, 0, "core", GK20A_CLK_GPC_MDIV }, 889 { nv_clk_src_max }, 890 }, 891 }; 892 893 static const struct nvkm_clk_func 894 gm20b_clk = { 895 .init = gm20b_clk_init, 896 .fini = gm20b_clk_fini, 897 .read = gk20a_clk_read, 898 .calc = gm20b_clk_calc, 899 .prog = gm20b_clk_prog, 900 .tidy = gk20a_clk_tidy, 901 .pstates = gm20b_pstates, 902 .nr_pstates = ARRAY_SIZE(gm20b_pstates), 903 .domains = { 904 { nv_clk_src_crystal, 0xff }, 905 { nv_clk_src_gpc, 0xff, 0, "core", GK20A_CLK_GPC_MDIV }, 906 { nv_clk_src_max }, 907 }, 908 }; 909 910 static int 911 gm20b_clk_new_speedo0(struct nvkm_device *device, int index, 912 struct nvkm_clk **pclk) 913 { 914 struct gk20a_clk *clk; 915 int ret; 916 917 clk = kzalloc(sizeof(*clk), GFP_KERNEL); 918 if (!clk) 919 return -ENOMEM; 920 *pclk = &clk->base; 921 922 ret = gk20a_clk_ctor(device, index, &gm20b_clk_speedo0, 923 &gm20b_pllg_params, clk); 924 925 clk->pl_to_div = pl_to_div; 926 clk->div_to_pl = div_to_pl; 927 928 return ret; 929 } 930 931 /* FUSE register */ 932 #define FUSE_RESERVED_CALIB0 0x204 933 #define FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_SHIFT 0 934 #define FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_WIDTH 4 935 #define FUSE_RESERVED_CALIB0_INTERCEPT_INT_SHIFT 4 936 #define FUSE_RESERVED_CALIB0_INTERCEPT_INT_WIDTH 10 937 #define FUSE_RESERVED_CALIB0_SLOPE_FRAC_SHIFT 14 938 #define FUSE_RESERVED_CALIB0_SLOPE_FRAC_WIDTH 10 939 #define FUSE_RESERVED_CALIB0_SLOPE_INT_SHIFT 24 940 #define FUSE_RESERVED_CALIB0_SLOPE_INT_WIDTH 6 941 #define FUSE_RESERVED_CALIB0_FUSE_REV_SHIFT 30 942 #define FUSE_RESERVED_CALIB0_FUSE_REV_WIDTH 2 943 944 static int 945 gm20b_clk_init_fused_params(struct gm20b_clk *clk) 946 { 947 struct nvkm_subdev *subdev = &clk->base.base.subdev; 948 u32 val = 0; 949 u32 rev = 0; 950 951 #if IS_ENABLED(CONFIG_ARCH_TEGRA) 952 tegra_fuse_readl(FUSE_RESERVED_CALIB0, &val); 953 rev = (val >> FUSE_RESERVED_CALIB0_FUSE_REV_SHIFT) & 954 MASK(FUSE_RESERVED_CALIB0_FUSE_REV_WIDTH); 955 #endif 956 957 /* No fused parameters, we will calibrate later */ 958 if (rev == 0) 959 return -EINVAL; 960 961 /* Integer part in mV + fractional part in uV */ 962 clk->uvdet_slope = ((val >> FUSE_RESERVED_CALIB0_SLOPE_INT_SHIFT) & 963 MASK(FUSE_RESERVED_CALIB0_SLOPE_INT_WIDTH)) * 1000 + 964 ((val >> FUSE_RESERVED_CALIB0_SLOPE_FRAC_SHIFT) & 965 MASK(FUSE_RESERVED_CALIB0_SLOPE_FRAC_WIDTH)); 966 967 /* Integer part in mV + fractional part in 100uV */ 968 clk->uvdet_offs = ((val >> FUSE_RESERVED_CALIB0_INTERCEPT_INT_SHIFT) & 969 MASK(FUSE_RESERVED_CALIB0_INTERCEPT_INT_WIDTH)) * 1000 + 970 ((val >> FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_SHIFT) & 971 MASK(FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_WIDTH)) * 100; 972 973 nvkm_debug(subdev, "fused calibration data: slope %d, offs %d\n", 974 clk->uvdet_slope, clk->uvdet_offs); 975 return 0; 976 } 977 978 static int 979 gm20b_clk_init_safe_fmax(struct gm20b_clk *clk) 980 { 981 struct nvkm_subdev *subdev = &clk->base.base.subdev; 982 struct nvkm_volt *volt = subdev->device->volt; 983 struct nvkm_pstate *pstates = clk->base.base.func->pstates; 984 int nr_pstates = clk->base.base.func->nr_pstates; 985 int vmin, id = 0; 986 u32 fmax = 0; 987 int i; 988 989 /* find lowest voltage we can use */ 990 vmin = volt->vid[0].uv; 991 for (i = 1; i < volt->vid_nr; i++) { 992 if (volt->vid[i].uv <= vmin) { 993 vmin = volt->vid[i].uv; 994 id = volt->vid[i].vid; 995 } 996 } 997 998 /* find max frequency at this voltage */ 999 for (i = 0; i < nr_pstates; i++) 1000 if (pstates[i].base.voltage == id) 1001 fmax = max(fmax, 1002 pstates[i].base.domain[nv_clk_src_gpc]); 1003 1004 if (!fmax) { 1005 nvkm_error(subdev, "failed to evaluate safe fmax\n"); 1006 return -EINVAL; 1007 } 1008 1009 /* we are safe at 90% of the max frequency */ 1010 clk->safe_fmax_vmin = fmax * (100 - 10) / 100; 1011 nvkm_debug(subdev, "safe fmax @ vmin = %u Khz\n", clk->safe_fmax_vmin); 1012 1013 return 0; 1014 } 1015 1016 int 1017 gm20b_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk) 1018 { 1019 struct nvkm_device_tegra *tdev = device->func->tegra(device); 1020 struct gm20b_clk *clk; 1021 struct nvkm_subdev *subdev; 1022 struct gk20a_clk_pllg_params *clk_params; 1023 int ret; 1024 1025 /* Speedo 0 GPUs cannot use noise-aware PLL */ 1026 if (tdev->gpu_speedo_id == 0) 1027 return gm20b_clk_new_speedo0(device, index, pclk); 1028 1029 /* Speedo >= 1, use NAPLL */ 1030 clk = kzalloc(sizeof(*clk) + sizeof(*clk_params), GFP_KERNEL); 1031 if (!clk) 1032 return -ENOMEM; 1033 *pclk = &clk->base.base; 1034 subdev = &clk->base.base.subdev; 1035 1036 /* duplicate the clock parameters since we will patch them below */ 1037 clk_params = (void *) (clk + 1); 1038 *clk_params = gm20b_pllg_params; 1039 ret = gk20a_clk_ctor(device, index, &gm20b_clk, clk_params, 1040 &clk->base); 1041 if (ret) 1042 return ret; 1043 1044 /* 1045 * NAPLL can only work with max_u, clamp the m range so 1046 * gk20a_pllg_calc_mnp always uses it 1047 */ 1048 clk_params->max_m = clk_params->min_m = DIV_ROUND_UP(clk_params->max_u, 1049 (clk->base.parent_rate / KHZ)); 1050 if (clk_params->max_m == 0) { 1051 nvkm_warn(subdev, "cannot use NAPLL, using legacy clock...\n"); 1052 kfree(clk); 1053 return gm20b_clk_new_speedo0(device, index, pclk); 1054 } 1055 1056 clk->base.pl_to_div = pl_to_div; 1057 clk->base.div_to_pl = div_to_pl; 1058 1059 clk->dvfs_params = &gm20b_dvfs_params; 1060 1061 ret = gm20b_clk_init_fused_params(clk); 1062 /* 1063 * we will calibrate during init - should never happen on 1064 * prod parts 1065 */ 1066 if (ret) 1067 nvkm_warn(subdev, "no fused calibration parameters\n"); 1068 1069 ret = gm20b_clk_init_safe_fmax(clk); 1070 if (ret) 1071 return ret; 1072 1073 return 0; 1074 } 1075