1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2019 Intel Corporation 4 */ 5 6 #include <drm/i915_drm.h> 7 8 #include "i915_drv.h" 9 #include "intel_breadcrumbs.h" 10 #include "intel_gt.h" 11 #include "intel_gt_clock_utils.h" 12 #include "intel_gt_irq.h" 13 #include "intel_gt_pm_irq.h" 14 #include "intel_rps.h" 15 #include "intel_sideband.h" 16 #include "../../../platform/x86/intel_ips.h" 17 18 #define BUSY_MAX_EI 20u /* ms */ 19 20 /* 21 * Lock protecting IPS related data structures 22 */ 23 static DEFINE_SPINLOCK(mchdev_lock); 24 25 static struct intel_gt *rps_to_gt(struct intel_rps *rps) 26 { 27 return container_of(rps, struct intel_gt, rps); 28 } 29 30 static struct drm_i915_private *rps_to_i915(struct intel_rps *rps) 31 { 32 return rps_to_gt(rps)->i915; 33 } 34 35 static struct intel_uncore *rps_to_uncore(struct intel_rps *rps) 36 { 37 return rps_to_gt(rps)->uncore; 38 } 39 40 static struct intel_guc_slpc *rps_to_slpc(struct intel_rps *rps) 41 { 42 struct intel_gt *gt = rps_to_gt(rps); 43 44 return >->uc.guc.slpc; 45 } 46 47 static bool rps_uses_slpc(struct intel_rps *rps) 48 { 49 struct intel_gt *gt = rps_to_gt(rps); 50 51 return intel_uc_uses_guc_slpc(>->uc); 52 } 53 54 static u32 rps_pm_sanitize_mask(struct intel_rps *rps, u32 mask) 55 { 56 return mask & ~rps->pm_intrmsk_mbz; 57 } 58 59 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val) 60 { 61 intel_uncore_write_fw(uncore, reg, val); 62 } 63 64 static void rps_timer(struct timer_list *t) 65 { 66 struct intel_rps *rps = from_timer(rps, t, timer); 67 struct intel_engine_cs *engine; 68 ktime_t dt, last, timestamp; 69 enum intel_engine_id id; 70 s64 max_busy[3] = {}; 71 72 timestamp = 0; 73 for_each_engine(engine, rps_to_gt(rps), id) { 74 s64 busy; 75 int i; 76 77 dt = intel_engine_get_busy_time(engine, ×tamp); 78 last = engine->stats.rps; 79 engine->stats.rps = dt; 80 81 busy = ktime_to_ns(ktime_sub(dt, last)); 82 for (i = 0; i < ARRAY_SIZE(max_busy); i++) { 83 if (busy > max_busy[i]) 84 swap(busy, max_busy[i]); 85 } 86 } 87 last = rps->pm_timestamp; 88 rps->pm_timestamp = timestamp; 89 90 if (intel_rps_is_active(rps)) { 91 s64 busy; 92 int i; 93 94 dt = ktime_sub(timestamp, last); 95 96 /* 97 * Our goal is to evaluate each engine independently, so we run 98 * at the lowest clocks required to sustain the heaviest 99 * workload. However, a task may be split into sequential 100 * dependent operations across a set of engines, such that 101 * the independent contributions do not account for high load, 102 * but overall the task is GPU bound. For example, consider 103 * video decode on vcs followed by colour post-processing 104 * on vecs, followed by general post-processing on rcs. 105 * Since multi-engines being active does imply a single 106 * continuous workload across all engines, we hedge our 107 * bets by only contributing a factor of the distributed 108 * load into our busyness calculation. 109 */ 110 busy = max_busy[0]; 111 for (i = 1; i < ARRAY_SIZE(max_busy); i++) { 112 if (!max_busy[i]) 113 break; 114 115 busy += div_u64(max_busy[i], 1 << i); 116 } 117 GT_TRACE(rps_to_gt(rps), 118 "busy:%lld [%d%%], max:[%lld, %lld, %lld], interval:%d\n", 119 busy, (int)div64_u64(100 * busy, dt), 120 max_busy[0], max_busy[1], max_busy[2], 121 rps->pm_interval); 122 123 if (100 * busy > rps->power.up_threshold * dt && 124 rps->cur_freq < rps->max_freq_softlimit) { 125 rps->pm_iir |= GEN6_PM_RP_UP_THRESHOLD; 126 rps->pm_interval = 1; 127 schedule_work(&rps->work); 128 } else if (100 * busy < rps->power.down_threshold * dt && 129 rps->cur_freq > rps->min_freq_softlimit) { 130 rps->pm_iir |= GEN6_PM_RP_DOWN_THRESHOLD; 131 rps->pm_interval = 1; 132 schedule_work(&rps->work); 133 } else { 134 rps->last_adj = 0; 135 } 136 137 mod_timer(&rps->timer, 138 jiffies + msecs_to_jiffies(rps->pm_interval)); 139 rps->pm_interval = min(rps->pm_interval * 2, BUSY_MAX_EI); 140 } 141 } 142 143 static void rps_start_timer(struct intel_rps *rps) 144 { 145 rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp); 146 rps->pm_interval = 1; 147 mod_timer(&rps->timer, jiffies + 1); 148 } 149 150 static void rps_stop_timer(struct intel_rps *rps) 151 { 152 del_timer_sync(&rps->timer); 153 rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp); 154 cancel_work_sync(&rps->work); 155 } 156 157 static u32 rps_pm_mask(struct intel_rps *rps, u8 val) 158 { 159 u32 mask = 0; 160 161 /* We use UP_EI_EXPIRED interrupts for both up/down in manual mode */ 162 if (val > rps->min_freq_softlimit) 163 mask |= (GEN6_PM_RP_UP_EI_EXPIRED | 164 GEN6_PM_RP_DOWN_THRESHOLD | 165 GEN6_PM_RP_DOWN_TIMEOUT); 166 167 if (val < rps->max_freq_softlimit) 168 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD; 169 170 mask &= rps->pm_events; 171 172 return rps_pm_sanitize_mask(rps, ~mask); 173 } 174 175 static void rps_reset_ei(struct intel_rps *rps) 176 { 177 memset(&rps->ei, 0, sizeof(rps->ei)); 178 } 179 180 static void rps_enable_interrupts(struct intel_rps *rps) 181 { 182 struct intel_gt *gt = rps_to_gt(rps); 183 184 GEM_BUG_ON(rps_uses_slpc(rps)); 185 186 GT_TRACE(gt, "interrupts:on rps->pm_events: %x, rps_pm_mask:%x\n", 187 rps->pm_events, rps_pm_mask(rps, rps->last_freq)); 188 189 rps_reset_ei(rps); 190 191 spin_lock_irq(>->irq_lock); 192 gen6_gt_pm_enable_irq(gt, rps->pm_events); 193 spin_unlock_irq(>->irq_lock); 194 195 intel_uncore_write(gt->uncore, 196 GEN6_PMINTRMSK, rps_pm_mask(rps, rps->last_freq)); 197 } 198 199 static void gen6_rps_reset_interrupts(struct intel_rps *rps) 200 { 201 gen6_gt_pm_reset_iir(rps_to_gt(rps), GEN6_PM_RPS_EVENTS); 202 } 203 204 static void gen11_rps_reset_interrupts(struct intel_rps *rps) 205 { 206 while (gen11_gt_reset_one_iir(rps_to_gt(rps), 0, GEN11_GTPM)) 207 ; 208 } 209 210 static void rps_reset_interrupts(struct intel_rps *rps) 211 { 212 struct intel_gt *gt = rps_to_gt(rps); 213 214 spin_lock_irq(>->irq_lock); 215 if (GRAPHICS_VER(gt->i915) >= 11) 216 gen11_rps_reset_interrupts(rps); 217 else 218 gen6_rps_reset_interrupts(rps); 219 220 rps->pm_iir = 0; 221 spin_unlock_irq(>->irq_lock); 222 } 223 224 static void rps_disable_interrupts(struct intel_rps *rps) 225 { 226 struct intel_gt *gt = rps_to_gt(rps); 227 228 intel_uncore_write(gt->uncore, 229 GEN6_PMINTRMSK, rps_pm_sanitize_mask(rps, ~0u)); 230 231 spin_lock_irq(>->irq_lock); 232 gen6_gt_pm_disable_irq(gt, GEN6_PM_RPS_EVENTS); 233 spin_unlock_irq(>->irq_lock); 234 235 intel_synchronize_irq(gt->i915); 236 237 /* 238 * Now that we will not be generating any more work, flush any 239 * outstanding tasks. As we are called on the RPS idle path, 240 * we will reset the GPU to minimum frequencies, so the current 241 * state of the worker can be discarded. 242 */ 243 cancel_work_sync(&rps->work); 244 245 rps_reset_interrupts(rps); 246 GT_TRACE(gt, "interrupts:off\n"); 247 } 248 249 static const struct cparams { 250 u16 i; 251 u16 t; 252 u16 m; 253 u16 c; 254 } cparams[] = { 255 { 1, 1333, 301, 28664 }, 256 { 1, 1066, 294, 24460 }, 257 { 1, 800, 294, 25192 }, 258 { 0, 1333, 276, 27605 }, 259 { 0, 1066, 276, 27605 }, 260 { 0, 800, 231, 23784 }, 261 }; 262 263 static void gen5_rps_init(struct intel_rps *rps) 264 { 265 struct drm_i915_private *i915 = rps_to_i915(rps); 266 struct intel_uncore *uncore = rps_to_uncore(rps); 267 u8 fmax, fmin, fstart; 268 u32 rgvmodectl; 269 int c_m, i; 270 271 if (i915->fsb_freq <= 3200) 272 c_m = 0; 273 else if (i915->fsb_freq <= 4800) 274 c_m = 1; 275 else 276 c_m = 2; 277 278 for (i = 0; i < ARRAY_SIZE(cparams); i++) { 279 if (cparams[i].i == c_m && cparams[i].t == i915->mem_freq) { 280 rps->ips.m = cparams[i].m; 281 rps->ips.c = cparams[i].c; 282 break; 283 } 284 } 285 286 rgvmodectl = intel_uncore_read(uncore, MEMMODECTL); 287 288 /* Set up min, max, and cur for interrupt handling */ 289 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT; 290 fmin = (rgvmodectl & MEMMODE_FMIN_MASK); 291 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >> 292 MEMMODE_FSTART_SHIFT; 293 drm_dbg(&i915->drm, "fmax: %d, fmin: %d, fstart: %d\n", 294 fmax, fmin, fstart); 295 296 rps->min_freq = fmax; 297 rps->efficient_freq = fstart; 298 rps->max_freq = fmin; 299 } 300 301 static unsigned long 302 __ips_chipset_val(struct intel_ips *ips) 303 { 304 struct intel_uncore *uncore = 305 rps_to_uncore(container_of(ips, struct intel_rps, ips)); 306 unsigned long now = jiffies_to_msecs(jiffies), dt; 307 unsigned long result; 308 u64 total, delta; 309 310 lockdep_assert_held(&mchdev_lock); 311 312 /* 313 * Prevent division-by-zero if we are asking too fast. 314 * Also, we don't get interesting results if we are polling 315 * faster than once in 10ms, so just return the saved value 316 * in such cases. 317 */ 318 dt = now - ips->last_time1; 319 if (dt <= 10) 320 return ips->chipset_power; 321 322 /* FIXME: handle per-counter overflow */ 323 total = intel_uncore_read(uncore, DMIEC); 324 total += intel_uncore_read(uncore, DDREC); 325 total += intel_uncore_read(uncore, CSIEC); 326 327 delta = total - ips->last_count1; 328 329 result = div_u64(div_u64(ips->m * delta, dt) + ips->c, 10); 330 331 ips->last_count1 = total; 332 ips->last_time1 = now; 333 334 ips->chipset_power = result; 335 336 return result; 337 } 338 339 static unsigned long ips_mch_val(struct intel_uncore *uncore) 340 { 341 unsigned int m, x, b; 342 u32 tsfs; 343 344 tsfs = intel_uncore_read(uncore, TSFS); 345 x = intel_uncore_read8(uncore, TR1); 346 347 b = tsfs & TSFS_INTR_MASK; 348 m = (tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT; 349 350 return m * x / 127 - b; 351 } 352 353 static int _pxvid_to_vd(u8 pxvid) 354 { 355 if (pxvid == 0) 356 return 0; 357 358 if (pxvid >= 8 && pxvid < 31) 359 pxvid = 31; 360 361 return (pxvid + 2) * 125; 362 } 363 364 static u32 pvid_to_extvid(struct drm_i915_private *i915, u8 pxvid) 365 { 366 const int vd = _pxvid_to_vd(pxvid); 367 368 if (INTEL_INFO(i915)->is_mobile) 369 return max(vd - 1125, 0); 370 371 return vd; 372 } 373 374 static void __gen5_ips_update(struct intel_ips *ips) 375 { 376 struct intel_uncore *uncore = 377 rps_to_uncore(container_of(ips, struct intel_rps, ips)); 378 u64 now, delta, dt; 379 u32 count; 380 381 lockdep_assert_held(&mchdev_lock); 382 383 now = ktime_get_raw_ns(); 384 dt = now - ips->last_time2; 385 do_div(dt, NSEC_PER_MSEC); 386 387 /* Don't divide by 0 */ 388 if (dt <= 10) 389 return; 390 391 count = intel_uncore_read(uncore, GFXEC); 392 delta = count - ips->last_count2; 393 394 ips->last_count2 = count; 395 ips->last_time2 = now; 396 397 /* More magic constants... */ 398 ips->gfx_power = div_u64(delta * 1181, dt * 10); 399 } 400 401 static void gen5_rps_update(struct intel_rps *rps) 402 { 403 spin_lock_irq(&mchdev_lock); 404 __gen5_ips_update(&rps->ips); 405 spin_unlock_irq(&mchdev_lock); 406 } 407 408 static unsigned int gen5_invert_freq(struct intel_rps *rps, 409 unsigned int val) 410 { 411 /* Invert the frequency bin into an ips delay */ 412 val = rps->max_freq - val; 413 val = rps->min_freq + val; 414 415 return val; 416 } 417 418 static int __gen5_rps_set(struct intel_rps *rps, u8 val) 419 { 420 struct intel_uncore *uncore = rps_to_uncore(rps); 421 u16 rgvswctl; 422 423 lockdep_assert_held(&mchdev_lock); 424 425 rgvswctl = intel_uncore_read16(uncore, MEMSWCTL); 426 if (rgvswctl & MEMCTL_CMD_STS) { 427 DRM_DEBUG("gpu busy, RCS change rejected\n"); 428 return -EBUSY; /* still busy with another command */ 429 } 430 431 /* Invert the frequency bin into an ips delay */ 432 val = gen5_invert_freq(rps, val); 433 434 rgvswctl = 435 (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) | 436 (val << MEMCTL_FREQ_SHIFT) | 437 MEMCTL_SFCAVM; 438 intel_uncore_write16(uncore, MEMSWCTL, rgvswctl); 439 intel_uncore_posting_read16(uncore, MEMSWCTL); 440 441 rgvswctl |= MEMCTL_CMD_STS; 442 intel_uncore_write16(uncore, MEMSWCTL, rgvswctl); 443 444 return 0; 445 } 446 447 static int gen5_rps_set(struct intel_rps *rps, u8 val) 448 { 449 int err; 450 451 spin_lock_irq(&mchdev_lock); 452 err = __gen5_rps_set(rps, val); 453 spin_unlock_irq(&mchdev_lock); 454 455 return err; 456 } 457 458 static unsigned long intel_pxfreq(u32 vidfreq) 459 { 460 int div = (vidfreq & 0x3f0000) >> 16; 461 int post = (vidfreq & 0x3000) >> 12; 462 int pre = (vidfreq & 0x7); 463 464 if (!pre) 465 return 0; 466 467 return div * 133333 / (pre << post); 468 } 469 470 static unsigned int init_emon(struct intel_uncore *uncore) 471 { 472 u8 pxw[16]; 473 int i; 474 475 /* Disable to program */ 476 intel_uncore_write(uncore, ECR, 0); 477 intel_uncore_posting_read(uncore, ECR); 478 479 /* Program energy weights for various events */ 480 intel_uncore_write(uncore, SDEW, 0x15040d00); 481 intel_uncore_write(uncore, CSIEW0, 0x007f0000); 482 intel_uncore_write(uncore, CSIEW1, 0x1e220004); 483 intel_uncore_write(uncore, CSIEW2, 0x04000004); 484 485 for (i = 0; i < 5; i++) 486 intel_uncore_write(uncore, PEW(i), 0); 487 for (i = 0; i < 3; i++) 488 intel_uncore_write(uncore, DEW(i), 0); 489 490 /* Program P-state weights to account for frequency power adjustment */ 491 for (i = 0; i < 16; i++) { 492 u32 pxvidfreq = intel_uncore_read(uncore, PXVFREQ(i)); 493 unsigned int freq = intel_pxfreq(pxvidfreq); 494 unsigned int vid = 495 (pxvidfreq & PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT; 496 unsigned int val; 497 498 val = vid * vid * freq / 1000 * 255; 499 val /= 127 * 127 * 900; 500 501 pxw[i] = val; 502 } 503 /* Render standby states get 0 weight */ 504 pxw[14] = 0; 505 pxw[15] = 0; 506 507 for (i = 0; i < 4; i++) { 508 intel_uncore_write(uncore, PXW(i), 509 pxw[i * 4 + 0] << 24 | 510 pxw[i * 4 + 1] << 16 | 511 pxw[i * 4 + 2] << 8 | 512 pxw[i * 4 + 3] << 0); 513 } 514 515 /* Adjust magic regs to magic values (more experimental results) */ 516 intel_uncore_write(uncore, OGW0, 0); 517 intel_uncore_write(uncore, OGW1, 0); 518 intel_uncore_write(uncore, EG0, 0x00007f00); 519 intel_uncore_write(uncore, EG1, 0x0000000e); 520 intel_uncore_write(uncore, EG2, 0x000e0000); 521 intel_uncore_write(uncore, EG3, 0x68000300); 522 intel_uncore_write(uncore, EG4, 0x42000000); 523 intel_uncore_write(uncore, EG5, 0x00140031); 524 intel_uncore_write(uncore, EG6, 0); 525 intel_uncore_write(uncore, EG7, 0); 526 527 for (i = 0; i < 8; i++) 528 intel_uncore_write(uncore, PXWL(i), 0); 529 530 /* Enable PMON + select events */ 531 intel_uncore_write(uncore, ECR, 0x80000019); 532 533 return intel_uncore_read(uncore, LCFUSE02) & LCFUSE_HIV_MASK; 534 } 535 536 static bool gen5_rps_enable(struct intel_rps *rps) 537 { 538 struct drm_i915_private *i915 = rps_to_i915(rps); 539 struct intel_uncore *uncore = rps_to_uncore(rps); 540 u8 fstart, vstart; 541 u32 rgvmodectl; 542 543 spin_lock_irq(&mchdev_lock); 544 545 rgvmodectl = intel_uncore_read(uncore, MEMMODECTL); 546 547 /* Enable temp reporting */ 548 intel_uncore_write16(uncore, PMMISC, 549 intel_uncore_read16(uncore, PMMISC) | MCPPCE_EN); 550 intel_uncore_write16(uncore, TSC1, 551 intel_uncore_read16(uncore, TSC1) | TSE); 552 553 /* 100ms RC evaluation intervals */ 554 intel_uncore_write(uncore, RCUPEI, 100000); 555 intel_uncore_write(uncore, RCDNEI, 100000); 556 557 /* Set max/min thresholds to 90ms and 80ms respectively */ 558 intel_uncore_write(uncore, RCBMAXAVG, 90000); 559 intel_uncore_write(uncore, RCBMINAVG, 80000); 560 561 intel_uncore_write(uncore, MEMIHYST, 1); 562 563 /* Set up min, max, and cur for interrupt handling */ 564 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >> 565 MEMMODE_FSTART_SHIFT; 566 567 vstart = (intel_uncore_read(uncore, PXVFREQ(fstart)) & 568 PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT; 569 570 intel_uncore_write(uncore, 571 MEMINTREN, 572 MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN); 573 574 intel_uncore_write(uncore, VIDSTART, vstart); 575 intel_uncore_posting_read(uncore, VIDSTART); 576 577 rgvmodectl |= MEMMODE_SWMODE_EN; 578 intel_uncore_write(uncore, MEMMODECTL, rgvmodectl); 579 580 if (wait_for_atomic((intel_uncore_read(uncore, MEMSWCTL) & 581 MEMCTL_CMD_STS) == 0, 10)) 582 drm_err(&uncore->i915->drm, 583 "stuck trying to change perf mode\n"); 584 mdelay(1); 585 586 __gen5_rps_set(rps, rps->cur_freq); 587 588 rps->ips.last_count1 = intel_uncore_read(uncore, DMIEC); 589 rps->ips.last_count1 += intel_uncore_read(uncore, DDREC); 590 rps->ips.last_count1 += intel_uncore_read(uncore, CSIEC); 591 rps->ips.last_time1 = jiffies_to_msecs(jiffies); 592 593 rps->ips.last_count2 = intel_uncore_read(uncore, GFXEC); 594 rps->ips.last_time2 = ktime_get_raw_ns(); 595 596 spin_lock(&i915->irq_lock); 597 ilk_enable_display_irq(i915, DE_PCU_EVENT); 598 spin_unlock(&i915->irq_lock); 599 600 spin_unlock_irq(&mchdev_lock); 601 602 rps->ips.corr = init_emon(uncore); 603 604 return true; 605 } 606 607 static void gen5_rps_disable(struct intel_rps *rps) 608 { 609 struct drm_i915_private *i915 = rps_to_i915(rps); 610 struct intel_uncore *uncore = rps_to_uncore(rps); 611 u16 rgvswctl; 612 613 spin_lock_irq(&mchdev_lock); 614 615 spin_lock(&i915->irq_lock); 616 ilk_disable_display_irq(i915, DE_PCU_EVENT); 617 spin_unlock(&i915->irq_lock); 618 619 rgvswctl = intel_uncore_read16(uncore, MEMSWCTL); 620 621 /* Ack interrupts, disable EFC interrupt */ 622 intel_uncore_write(uncore, MEMINTREN, 623 intel_uncore_read(uncore, MEMINTREN) & 624 ~MEMINT_EVAL_CHG_EN); 625 intel_uncore_write(uncore, MEMINTRSTS, MEMINT_EVAL_CHG); 626 627 /* Go back to the starting frequency */ 628 __gen5_rps_set(rps, rps->idle_freq); 629 mdelay(1); 630 rgvswctl |= MEMCTL_CMD_STS; 631 intel_uncore_write(uncore, MEMSWCTL, rgvswctl); 632 mdelay(1); 633 634 spin_unlock_irq(&mchdev_lock); 635 } 636 637 static u32 rps_limits(struct intel_rps *rps, u8 val) 638 { 639 u32 limits; 640 641 /* 642 * Only set the down limit when we've reached the lowest level to avoid 643 * getting more interrupts, otherwise leave this clear. This prevents a 644 * race in the hw when coming out of rc6: There's a tiny window where 645 * the hw runs at the minimal clock before selecting the desired 646 * frequency, if the down threshold expires in that window we will not 647 * receive a down interrupt. 648 */ 649 if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) { 650 limits = rps->max_freq_softlimit << 23; 651 if (val <= rps->min_freq_softlimit) 652 limits |= rps->min_freq_softlimit << 14; 653 } else { 654 limits = rps->max_freq_softlimit << 24; 655 if (val <= rps->min_freq_softlimit) 656 limits |= rps->min_freq_softlimit << 16; 657 } 658 659 return limits; 660 } 661 662 static void rps_set_power(struct intel_rps *rps, int new_power) 663 { 664 struct intel_gt *gt = rps_to_gt(rps); 665 struct intel_uncore *uncore = gt->uncore; 666 u32 threshold_up = 0, threshold_down = 0; /* in % */ 667 u32 ei_up = 0, ei_down = 0; 668 669 lockdep_assert_held(&rps->power.mutex); 670 671 if (new_power == rps->power.mode) 672 return; 673 674 threshold_up = 95; 675 threshold_down = 85; 676 677 /* Note the units here are not exactly 1us, but 1280ns. */ 678 switch (new_power) { 679 case LOW_POWER: 680 ei_up = 16000; 681 ei_down = 32000; 682 break; 683 684 case BETWEEN: 685 ei_up = 13000; 686 ei_down = 32000; 687 break; 688 689 case HIGH_POWER: 690 ei_up = 10000; 691 ei_down = 32000; 692 break; 693 } 694 695 /* When byt can survive without system hang with dynamic 696 * sw freq adjustments, this restriction can be lifted. 697 */ 698 if (IS_VALLEYVIEW(gt->i915)) 699 goto skip_hw_write; 700 701 GT_TRACE(gt, 702 "changing power mode [%d], up %d%% @ %dus, down %d%% @ %dus\n", 703 new_power, threshold_up, ei_up, threshold_down, ei_down); 704 705 set(uncore, GEN6_RP_UP_EI, 706 intel_gt_ns_to_pm_interval(gt, ei_up * 1000)); 707 set(uncore, GEN6_RP_UP_THRESHOLD, 708 intel_gt_ns_to_pm_interval(gt, ei_up * threshold_up * 10)); 709 710 set(uncore, GEN6_RP_DOWN_EI, 711 intel_gt_ns_to_pm_interval(gt, ei_down * 1000)); 712 set(uncore, GEN6_RP_DOWN_THRESHOLD, 713 intel_gt_ns_to_pm_interval(gt, ei_down * threshold_down * 10)); 714 715 set(uncore, GEN6_RP_CONTROL, 716 (GRAPHICS_VER(gt->i915) > 9 ? 0 : GEN6_RP_MEDIA_TURBO) | 717 GEN6_RP_MEDIA_HW_NORMAL_MODE | 718 GEN6_RP_MEDIA_IS_GFX | 719 GEN6_RP_ENABLE | 720 GEN6_RP_UP_BUSY_AVG | 721 GEN6_RP_DOWN_IDLE_AVG); 722 723 skip_hw_write: 724 rps->power.mode = new_power; 725 rps->power.up_threshold = threshold_up; 726 rps->power.down_threshold = threshold_down; 727 } 728 729 static void gen6_rps_set_thresholds(struct intel_rps *rps, u8 val) 730 { 731 int new_power; 732 733 new_power = rps->power.mode; 734 switch (rps->power.mode) { 735 case LOW_POWER: 736 if (val > rps->efficient_freq + 1 && 737 val > rps->cur_freq) 738 new_power = BETWEEN; 739 break; 740 741 case BETWEEN: 742 if (val <= rps->efficient_freq && 743 val < rps->cur_freq) 744 new_power = LOW_POWER; 745 else if (val >= rps->rp0_freq && 746 val > rps->cur_freq) 747 new_power = HIGH_POWER; 748 break; 749 750 case HIGH_POWER: 751 if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 && 752 val < rps->cur_freq) 753 new_power = BETWEEN; 754 break; 755 } 756 /* Max/min bins are special */ 757 if (val <= rps->min_freq_softlimit) 758 new_power = LOW_POWER; 759 if (val >= rps->max_freq_softlimit) 760 new_power = HIGH_POWER; 761 762 mutex_lock(&rps->power.mutex); 763 if (rps->power.interactive) 764 new_power = HIGH_POWER; 765 rps_set_power(rps, new_power); 766 mutex_unlock(&rps->power.mutex); 767 } 768 769 void intel_rps_mark_interactive(struct intel_rps *rps, bool interactive) 770 { 771 GT_TRACE(rps_to_gt(rps), "mark interactive: %s\n", yesno(interactive)); 772 773 mutex_lock(&rps->power.mutex); 774 if (interactive) { 775 if (!rps->power.interactive++ && intel_rps_is_active(rps)) 776 rps_set_power(rps, HIGH_POWER); 777 } else { 778 GEM_BUG_ON(!rps->power.interactive); 779 rps->power.interactive--; 780 } 781 mutex_unlock(&rps->power.mutex); 782 } 783 784 static int gen6_rps_set(struct intel_rps *rps, u8 val) 785 { 786 struct intel_uncore *uncore = rps_to_uncore(rps); 787 struct drm_i915_private *i915 = rps_to_i915(rps); 788 u32 swreq; 789 790 GEM_BUG_ON(rps_uses_slpc(rps)); 791 792 if (GRAPHICS_VER(i915) >= 9) 793 swreq = GEN9_FREQUENCY(val); 794 else if (IS_HASWELL(i915) || IS_BROADWELL(i915)) 795 swreq = HSW_FREQUENCY(val); 796 else 797 swreq = (GEN6_FREQUENCY(val) | 798 GEN6_OFFSET(0) | 799 GEN6_AGGRESSIVE_TURBO); 800 set(uncore, GEN6_RPNSWREQ, swreq); 801 802 GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d, swreq:%x\n", 803 val, intel_gpu_freq(rps, val), swreq); 804 805 return 0; 806 } 807 808 static int vlv_rps_set(struct intel_rps *rps, u8 val) 809 { 810 struct drm_i915_private *i915 = rps_to_i915(rps); 811 int err; 812 813 vlv_punit_get(i915); 814 err = vlv_punit_write(i915, PUNIT_REG_GPU_FREQ_REQ, val); 815 vlv_punit_put(i915); 816 817 GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d\n", 818 val, intel_gpu_freq(rps, val)); 819 820 return err; 821 } 822 823 static int rps_set(struct intel_rps *rps, u8 val, bool update) 824 { 825 struct drm_i915_private *i915 = rps_to_i915(rps); 826 int err; 827 828 if (val == rps->last_freq) 829 return 0; 830 831 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) 832 err = vlv_rps_set(rps, val); 833 else if (GRAPHICS_VER(i915) >= 6) 834 err = gen6_rps_set(rps, val); 835 else 836 err = gen5_rps_set(rps, val); 837 if (err) 838 return err; 839 840 if (update && GRAPHICS_VER(i915) >= 6) 841 gen6_rps_set_thresholds(rps, val); 842 rps->last_freq = val; 843 844 return 0; 845 } 846 847 void intel_rps_unpark(struct intel_rps *rps) 848 { 849 if (!intel_rps_is_enabled(rps)) 850 return; 851 852 GT_TRACE(rps_to_gt(rps), "unpark:%x\n", rps->cur_freq); 853 854 /* 855 * Use the user's desired frequency as a guide, but for better 856 * performance, jump directly to RPe as our starting frequency. 857 */ 858 mutex_lock(&rps->lock); 859 860 intel_rps_set_active(rps); 861 intel_rps_set(rps, 862 clamp(rps->cur_freq, 863 rps->min_freq_softlimit, 864 rps->max_freq_softlimit)); 865 866 mutex_unlock(&rps->lock); 867 868 rps->pm_iir = 0; 869 if (intel_rps_has_interrupts(rps)) 870 rps_enable_interrupts(rps); 871 if (intel_rps_uses_timer(rps)) 872 rps_start_timer(rps); 873 874 if (GRAPHICS_VER(rps_to_i915(rps)) == 5) 875 gen5_rps_update(rps); 876 } 877 878 void intel_rps_park(struct intel_rps *rps) 879 { 880 int adj; 881 882 if (!intel_rps_is_enabled(rps)) 883 return; 884 885 GEM_BUG_ON(atomic_read(&rps->num_waiters)); 886 887 if (!intel_rps_clear_active(rps)) 888 return; 889 890 if (intel_rps_uses_timer(rps)) 891 rps_stop_timer(rps); 892 if (intel_rps_has_interrupts(rps)) 893 rps_disable_interrupts(rps); 894 895 if (rps->last_freq <= rps->idle_freq) 896 return; 897 898 /* 899 * The punit delays the write of the frequency and voltage until it 900 * determines the GPU is awake. During normal usage we don't want to 901 * waste power changing the frequency if the GPU is sleeping (rc6). 902 * However, the GPU and driver is now idle and we do not want to delay 903 * switching to minimum voltage (reducing power whilst idle) as we do 904 * not expect to be woken in the near future and so must flush the 905 * change by waking the device. 906 * 907 * We choose to take the media powerwell (either would do to trick the 908 * punit into committing the voltage change) as that takes a lot less 909 * power than the render powerwell. 910 */ 911 intel_uncore_forcewake_get(rps_to_uncore(rps), FORCEWAKE_MEDIA); 912 rps_set(rps, rps->idle_freq, false); 913 intel_uncore_forcewake_put(rps_to_uncore(rps), FORCEWAKE_MEDIA); 914 915 /* 916 * Since we will try and restart from the previously requested 917 * frequency on unparking, treat this idle point as a downclock 918 * interrupt and reduce the frequency for resume. If we park/unpark 919 * more frequently than the rps worker can run, we will not respond 920 * to any EI and never see a change in frequency. 921 * 922 * (Note we accommodate Cherryview's limitation of only using an 923 * even bin by applying it to all.) 924 */ 925 adj = rps->last_adj; 926 if (adj < 0) 927 adj *= 2; 928 else /* CHV needs even encode values */ 929 adj = -2; 930 rps->last_adj = adj; 931 rps->cur_freq = max_t(int, rps->cur_freq + adj, rps->min_freq); 932 if (rps->cur_freq < rps->efficient_freq) { 933 rps->cur_freq = rps->efficient_freq; 934 rps->last_adj = 0; 935 } 936 937 GT_TRACE(rps_to_gt(rps), "park:%x\n", rps->cur_freq); 938 } 939 940 void intel_rps_boost(struct i915_request *rq) 941 { 942 if (i915_request_signaled(rq) || i915_request_has_waitboost(rq)) 943 return; 944 945 /* Serializes with i915_request_retire() */ 946 if (!test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) { 947 struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps; 948 949 if (atomic_fetch_inc(&rps->num_waiters)) 950 return; 951 952 if (!intel_rps_is_active(rps)) 953 return; 954 955 GT_TRACE(rps_to_gt(rps), "boost fence:%llx:%llx\n", 956 rq->fence.context, rq->fence.seqno); 957 958 if (READ_ONCE(rps->cur_freq) < rps->boost_freq) 959 schedule_work(&rps->work); 960 961 WRITE_ONCE(rps->boosts, rps->boosts + 1); /* debug only */ 962 } 963 } 964 965 int intel_rps_set(struct intel_rps *rps, u8 val) 966 { 967 int err; 968 969 lockdep_assert_held(&rps->lock); 970 GEM_BUG_ON(val > rps->max_freq); 971 GEM_BUG_ON(val < rps->min_freq); 972 973 if (intel_rps_is_active(rps)) { 974 err = rps_set(rps, val, true); 975 if (err) 976 return err; 977 978 /* 979 * Make sure we continue to get interrupts 980 * until we hit the minimum or maximum frequencies. 981 */ 982 if (intel_rps_has_interrupts(rps)) { 983 struct intel_uncore *uncore = rps_to_uncore(rps); 984 985 set(uncore, 986 GEN6_RP_INTERRUPT_LIMITS, rps_limits(rps, val)); 987 988 set(uncore, GEN6_PMINTRMSK, rps_pm_mask(rps, val)); 989 } 990 } 991 992 rps->cur_freq = val; 993 return 0; 994 } 995 996 static void gen6_rps_init(struct intel_rps *rps) 997 { 998 struct drm_i915_private *i915 = rps_to_i915(rps); 999 struct intel_uncore *uncore = rps_to_uncore(rps); 1000 1001 /* All of these values are in units of 50MHz */ 1002 1003 /* static values from HW: RP0 > RP1 > RPn (min_freq) */ 1004 if (IS_GEN9_LP(i915)) { 1005 u32 rp_state_cap = intel_uncore_read(uncore, BXT_RP_STATE_CAP); 1006 1007 rps->rp0_freq = (rp_state_cap >> 16) & 0xff; 1008 rps->rp1_freq = (rp_state_cap >> 8) & 0xff; 1009 rps->min_freq = (rp_state_cap >> 0) & 0xff; 1010 } else { 1011 u32 rp_state_cap = intel_uncore_read(uncore, GEN6_RP_STATE_CAP); 1012 1013 rps->rp0_freq = (rp_state_cap >> 0) & 0xff; 1014 rps->rp1_freq = (rp_state_cap >> 8) & 0xff; 1015 rps->min_freq = (rp_state_cap >> 16) & 0xff; 1016 } 1017 1018 /* hw_max = RP0 until we check for overclocking */ 1019 rps->max_freq = rps->rp0_freq; 1020 1021 rps->efficient_freq = rps->rp1_freq; 1022 if (IS_HASWELL(i915) || IS_BROADWELL(i915) || 1023 IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) { 1024 u32 ddcc_status = 0; 1025 1026 if (sandybridge_pcode_read(i915, 1027 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL, 1028 &ddcc_status, NULL) == 0) 1029 rps->efficient_freq = 1030 clamp_t(u8, 1031 (ddcc_status >> 8) & 0xff, 1032 rps->min_freq, 1033 rps->max_freq); 1034 } 1035 1036 if (IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) { 1037 /* Store the frequency values in 16.66 MHZ units, which is 1038 * the natural hardware unit for SKL 1039 */ 1040 rps->rp0_freq *= GEN9_FREQ_SCALER; 1041 rps->rp1_freq *= GEN9_FREQ_SCALER; 1042 rps->min_freq *= GEN9_FREQ_SCALER; 1043 rps->max_freq *= GEN9_FREQ_SCALER; 1044 rps->efficient_freq *= GEN9_FREQ_SCALER; 1045 } 1046 } 1047 1048 static bool rps_reset(struct intel_rps *rps) 1049 { 1050 struct drm_i915_private *i915 = rps_to_i915(rps); 1051 1052 /* force a reset */ 1053 rps->power.mode = -1; 1054 rps->last_freq = -1; 1055 1056 if (rps_set(rps, rps->min_freq, true)) { 1057 drm_err(&i915->drm, "Failed to reset RPS to initial values\n"); 1058 return false; 1059 } 1060 1061 rps->cur_freq = rps->min_freq; 1062 return true; 1063 } 1064 1065 /* See the Gen9_GT_PM_Programming_Guide doc for the below */ 1066 static bool gen9_rps_enable(struct intel_rps *rps) 1067 { 1068 struct intel_gt *gt = rps_to_gt(rps); 1069 struct intel_uncore *uncore = gt->uncore; 1070 1071 /* Program defaults and thresholds for RPS */ 1072 if (GRAPHICS_VER(gt->i915) == 9) 1073 intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ, 1074 GEN9_FREQUENCY(rps->rp1_freq)); 1075 1076 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 0xa); 1077 1078 rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD; 1079 1080 return rps_reset(rps); 1081 } 1082 1083 static bool gen8_rps_enable(struct intel_rps *rps) 1084 { 1085 struct intel_uncore *uncore = rps_to_uncore(rps); 1086 1087 intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ, 1088 HSW_FREQUENCY(rps->rp1_freq)); 1089 1090 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1091 1092 rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD; 1093 1094 return rps_reset(rps); 1095 } 1096 1097 static bool gen6_rps_enable(struct intel_rps *rps) 1098 { 1099 struct intel_uncore *uncore = rps_to_uncore(rps); 1100 1101 /* Power down if completely idle for over 50ms */ 1102 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 50000); 1103 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1104 1105 rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD | 1106 GEN6_PM_RP_DOWN_THRESHOLD | 1107 GEN6_PM_RP_DOWN_TIMEOUT); 1108 1109 return rps_reset(rps); 1110 } 1111 1112 static int chv_rps_max_freq(struct intel_rps *rps) 1113 { 1114 struct drm_i915_private *i915 = rps_to_i915(rps); 1115 struct intel_gt *gt = rps_to_gt(rps); 1116 u32 val; 1117 1118 val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE); 1119 1120 switch (gt->info.sseu.eu_total) { 1121 case 8: 1122 /* (2 * 4) config */ 1123 val >>= FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT; 1124 break; 1125 case 12: 1126 /* (2 * 6) config */ 1127 val >>= FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT; 1128 break; 1129 case 16: 1130 /* (2 * 8) config */ 1131 default: 1132 /* Setting (2 * 8) Min RP0 for any other combination */ 1133 val >>= FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT; 1134 break; 1135 } 1136 1137 return val & FB_GFX_FREQ_FUSE_MASK; 1138 } 1139 1140 static int chv_rps_rpe_freq(struct intel_rps *rps) 1141 { 1142 struct drm_i915_private *i915 = rps_to_i915(rps); 1143 u32 val; 1144 1145 val = vlv_punit_read(i915, PUNIT_GPU_DUTYCYCLE_REG); 1146 val >>= PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT; 1147 1148 return val & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK; 1149 } 1150 1151 static int chv_rps_guar_freq(struct intel_rps *rps) 1152 { 1153 struct drm_i915_private *i915 = rps_to_i915(rps); 1154 u32 val; 1155 1156 val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE); 1157 1158 return val & FB_GFX_FREQ_FUSE_MASK; 1159 } 1160 1161 static u32 chv_rps_min_freq(struct intel_rps *rps) 1162 { 1163 struct drm_i915_private *i915 = rps_to_i915(rps); 1164 u32 val; 1165 1166 val = vlv_punit_read(i915, FB_GFX_FMIN_AT_VMIN_FUSE); 1167 val >>= FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT; 1168 1169 return val & FB_GFX_FREQ_FUSE_MASK; 1170 } 1171 1172 static bool chv_rps_enable(struct intel_rps *rps) 1173 { 1174 struct intel_uncore *uncore = rps_to_uncore(rps); 1175 struct drm_i915_private *i915 = rps_to_i915(rps); 1176 u32 val; 1177 1178 /* 1: Program defaults and thresholds for RPS*/ 1179 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000); 1180 intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400); 1181 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000); 1182 intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000); 1183 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000); 1184 1185 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1186 1187 /* 2: Enable RPS */ 1188 intel_uncore_write_fw(uncore, GEN6_RP_CONTROL, 1189 GEN6_RP_MEDIA_HW_NORMAL_MODE | 1190 GEN6_RP_MEDIA_IS_GFX | 1191 GEN6_RP_ENABLE | 1192 GEN6_RP_UP_BUSY_AVG | 1193 GEN6_RP_DOWN_IDLE_AVG); 1194 1195 rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD | 1196 GEN6_PM_RP_DOWN_THRESHOLD | 1197 GEN6_PM_RP_DOWN_TIMEOUT); 1198 1199 /* Setting Fixed Bias */ 1200 vlv_punit_get(i915); 1201 1202 val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | CHV_BIAS_CPU_50_SOC_50; 1203 vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val); 1204 1205 val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS); 1206 1207 vlv_punit_put(i915); 1208 1209 /* RPS code assumes GPLL is used */ 1210 drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0, 1211 "GPLL not enabled\n"); 1212 1213 drm_dbg(&i915->drm, "GPLL enabled? %s\n", yesno(val & GPLLENABLE)); 1214 drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val); 1215 1216 return rps_reset(rps); 1217 } 1218 1219 static int vlv_rps_guar_freq(struct intel_rps *rps) 1220 { 1221 struct drm_i915_private *i915 = rps_to_i915(rps); 1222 u32 val, rp1; 1223 1224 val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE); 1225 1226 rp1 = val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK; 1227 rp1 >>= FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT; 1228 1229 return rp1; 1230 } 1231 1232 static int vlv_rps_max_freq(struct intel_rps *rps) 1233 { 1234 struct drm_i915_private *i915 = rps_to_i915(rps); 1235 u32 val, rp0; 1236 1237 val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE); 1238 1239 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT; 1240 /* Clamp to max */ 1241 rp0 = min_t(u32, rp0, 0xea); 1242 1243 return rp0; 1244 } 1245 1246 static int vlv_rps_rpe_freq(struct intel_rps *rps) 1247 { 1248 struct drm_i915_private *i915 = rps_to_i915(rps); 1249 u32 val, rpe; 1250 1251 val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_LO); 1252 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT; 1253 val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_HI); 1254 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5; 1255 1256 return rpe; 1257 } 1258 1259 static int vlv_rps_min_freq(struct intel_rps *rps) 1260 { 1261 struct drm_i915_private *i915 = rps_to_i915(rps); 1262 u32 val; 1263 1264 val = vlv_punit_read(i915, PUNIT_REG_GPU_LFM) & 0xff; 1265 /* 1266 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value 1267 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on 1268 * a BYT-M B0 the above register contains 0xbf. Moreover when setting 1269 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0 1270 * to make sure it matches what Punit accepts. 1271 */ 1272 return max_t(u32, val, 0xc0); 1273 } 1274 1275 static bool vlv_rps_enable(struct intel_rps *rps) 1276 { 1277 struct intel_uncore *uncore = rps_to_uncore(rps); 1278 struct drm_i915_private *i915 = rps_to_i915(rps); 1279 u32 val; 1280 1281 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000); 1282 intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400); 1283 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000); 1284 intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000); 1285 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000); 1286 1287 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1288 1289 intel_uncore_write_fw(uncore, GEN6_RP_CONTROL, 1290 GEN6_RP_MEDIA_TURBO | 1291 GEN6_RP_MEDIA_HW_NORMAL_MODE | 1292 GEN6_RP_MEDIA_IS_GFX | 1293 GEN6_RP_ENABLE | 1294 GEN6_RP_UP_BUSY_AVG | 1295 GEN6_RP_DOWN_IDLE_CONT); 1296 1297 /* WaGsvRC0ResidencyMethod:vlv */ 1298 rps->pm_events = GEN6_PM_RP_UP_EI_EXPIRED; 1299 1300 vlv_punit_get(i915); 1301 1302 /* Setting Fixed Bias */ 1303 val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | VLV_BIAS_CPU_125_SOC_875; 1304 vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val); 1305 1306 val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS); 1307 1308 vlv_punit_put(i915); 1309 1310 /* RPS code assumes GPLL is used */ 1311 drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0, 1312 "GPLL not enabled\n"); 1313 1314 drm_dbg(&i915->drm, "GPLL enabled? %s\n", yesno(val & GPLLENABLE)); 1315 drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val); 1316 1317 return rps_reset(rps); 1318 } 1319 1320 static unsigned long __ips_gfx_val(struct intel_ips *ips) 1321 { 1322 struct intel_rps *rps = container_of(ips, typeof(*rps), ips); 1323 struct intel_uncore *uncore = rps_to_uncore(rps); 1324 unsigned int t, state1, state2; 1325 u32 pxvid, ext_v; 1326 u64 corr, corr2; 1327 1328 lockdep_assert_held(&mchdev_lock); 1329 1330 pxvid = intel_uncore_read(uncore, PXVFREQ(rps->cur_freq)); 1331 pxvid = (pxvid >> 24) & 0x7f; 1332 ext_v = pvid_to_extvid(rps_to_i915(rps), pxvid); 1333 1334 state1 = ext_v; 1335 1336 /* Revel in the empirically derived constants */ 1337 1338 /* Correction factor in 1/100000 units */ 1339 t = ips_mch_val(uncore); 1340 if (t > 80) 1341 corr = t * 2349 + 135940; 1342 else if (t >= 50) 1343 corr = t * 964 + 29317; 1344 else /* < 50 */ 1345 corr = t * 301 + 1004; 1346 1347 corr = div_u64(corr * 150142 * state1, 10000) - 78642; 1348 corr2 = div_u64(corr, 100000) * ips->corr; 1349 1350 state2 = div_u64(corr2 * state1, 10000); 1351 state2 /= 100; /* convert to mW */ 1352 1353 __gen5_ips_update(ips); 1354 1355 return ips->gfx_power + state2; 1356 } 1357 1358 static bool has_busy_stats(struct intel_rps *rps) 1359 { 1360 struct intel_engine_cs *engine; 1361 enum intel_engine_id id; 1362 1363 for_each_engine(engine, rps_to_gt(rps), id) { 1364 if (!intel_engine_supports_stats(engine)) 1365 return false; 1366 } 1367 1368 return true; 1369 } 1370 1371 void intel_rps_enable(struct intel_rps *rps) 1372 { 1373 struct drm_i915_private *i915 = rps_to_i915(rps); 1374 struct intel_uncore *uncore = rps_to_uncore(rps); 1375 bool enabled = false; 1376 1377 if (!HAS_RPS(i915)) 1378 return; 1379 1380 if (rps_uses_slpc(rps)) 1381 return; 1382 1383 intel_gt_check_clock_frequency(rps_to_gt(rps)); 1384 1385 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 1386 if (rps->max_freq <= rps->min_freq) 1387 /* leave disabled, no room for dynamic reclocking */; 1388 else if (IS_CHERRYVIEW(i915)) 1389 enabled = chv_rps_enable(rps); 1390 else if (IS_VALLEYVIEW(i915)) 1391 enabled = vlv_rps_enable(rps); 1392 else if (GRAPHICS_VER(i915) >= 9) 1393 enabled = gen9_rps_enable(rps); 1394 else if (GRAPHICS_VER(i915) >= 8) 1395 enabled = gen8_rps_enable(rps); 1396 else if (GRAPHICS_VER(i915) >= 6) 1397 enabled = gen6_rps_enable(rps); 1398 else if (IS_IRONLAKE_M(i915)) 1399 enabled = gen5_rps_enable(rps); 1400 else 1401 MISSING_CASE(GRAPHICS_VER(i915)); 1402 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 1403 if (!enabled) 1404 return; 1405 1406 GT_TRACE(rps_to_gt(rps), 1407 "min:%x, max:%x, freq:[%d, %d]\n", 1408 rps->min_freq, rps->max_freq, 1409 intel_gpu_freq(rps, rps->min_freq), 1410 intel_gpu_freq(rps, rps->max_freq)); 1411 1412 GEM_BUG_ON(rps->max_freq < rps->min_freq); 1413 GEM_BUG_ON(rps->idle_freq > rps->max_freq); 1414 1415 GEM_BUG_ON(rps->efficient_freq < rps->min_freq); 1416 GEM_BUG_ON(rps->efficient_freq > rps->max_freq); 1417 1418 if (has_busy_stats(rps)) 1419 intel_rps_set_timer(rps); 1420 else if (GRAPHICS_VER(i915) >= 6) 1421 intel_rps_set_interrupts(rps); 1422 else 1423 /* Ironlake currently uses intel_ips.ko */ {} 1424 1425 intel_rps_set_enabled(rps); 1426 } 1427 1428 static void gen6_rps_disable(struct intel_rps *rps) 1429 { 1430 set(rps_to_uncore(rps), GEN6_RP_CONTROL, 0); 1431 } 1432 1433 void intel_rps_disable(struct intel_rps *rps) 1434 { 1435 struct drm_i915_private *i915 = rps_to_i915(rps); 1436 1437 intel_rps_clear_enabled(rps); 1438 intel_rps_clear_interrupts(rps); 1439 intel_rps_clear_timer(rps); 1440 1441 if (GRAPHICS_VER(i915) >= 6) 1442 gen6_rps_disable(rps); 1443 else if (IS_IRONLAKE_M(i915)) 1444 gen5_rps_disable(rps); 1445 } 1446 1447 static int byt_gpu_freq(struct intel_rps *rps, int val) 1448 { 1449 /* 1450 * N = val - 0xb7 1451 * Slow = Fast = GPLL ref * N 1452 */ 1453 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000); 1454 } 1455 1456 static int byt_freq_opcode(struct intel_rps *rps, int val) 1457 { 1458 return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7; 1459 } 1460 1461 static int chv_gpu_freq(struct intel_rps *rps, int val) 1462 { 1463 /* 1464 * N = val / 2 1465 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2 1466 */ 1467 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000); 1468 } 1469 1470 static int chv_freq_opcode(struct intel_rps *rps, int val) 1471 { 1472 /* CHV needs even values */ 1473 return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2; 1474 } 1475 1476 int intel_gpu_freq(struct intel_rps *rps, int val) 1477 { 1478 struct drm_i915_private *i915 = rps_to_i915(rps); 1479 1480 if (GRAPHICS_VER(i915) >= 9) 1481 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER, 1482 GEN9_FREQ_SCALER); 1483 else if (IS_CHERRYVIEW(i915)) 1484 return chv_gpu_freq(rps, val); 1485 else if (IS_VALLEYVIEW(i915)) 1486 return byt_gpu_freq(rps, val); 1487 else if (GRAPHICS_VER(i915) >= 6) 1488 return val * GT_FREQUENCY_MULTIPLIER; 1489 else 1490 return val; 1491 } 1492 1493 int intel_freq_opcode(struct intel_rps *rps, int val) 1494 { 1495 struct drm_i915_private *i915 = rps_to_i915(rps); 1496 1497 if (GRAPHICS_VER(i915) >= 9) 1498 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER, 1499 GT_FREQUENCY_MULTIPLIER); 1500 else if (IS_CHERRYVIEW(i915)) 1501 return chv_freq_opcode(rps, val); 1502 else if (IS_VALLEYVIEW(i915)) 1503 return byt_freq_opcode(rps, val); 1504 else if (GRAPHICS_VER(i915) >= 6) 1505 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER); 1506 else 1507 return val; 1508 } 1509 1510 static void vlv_init_gpll_ref_freq(struct intel_rps *rps) 1511 { 1512 struct drm_i915_private *i915 = rps_to_i915(rps); 1513 1514 rps->gpll_ref_freq = 1515 vlv_get_cck_clock(i915, "GPLL ref", 1516 CCK_GPLL_CLOCK_CONTROL, 1517 i915->czclk_freq); 1518 1519 drm_dbg(&i915->drm, "GPLL reference freq: %d kHz\n", 1520 rps->gpll_ref_freq); 1521 } 1522 1523 static void vlv_rps_init(struct intel_rps *rps) 1524 { 1525 struct drm_i915_private *i915 = rps_to_i915(rps); 1526 u32 val; 1527 1528 vlv_iosf_sb_get(i915, 1529 BIT(VLV_IOSF_SB_PUNIT) | 1530 BIT(VLV_IOSF_SB_NC) | 1531 BIT(VLV_IOSF_SB_CCK)); 1532 1533 vlv_init_gpll_ref_freq(rps); 1534 1535 val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS); 1536 switch ((val >> 6) & 3) { 1537 case 0: 1538 case 1: 1539 i915->mem_freq = 800; 1540 break; 1541 case 2: 1542 i915->mem_freq = 1066; 1543 break; 1544 case 3: 1545 i915->mem_freq = 1333; 1546 break; 1547 } 1548 drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq); 1549 1550 rps->max_freq = vlv_rps_max_freq(rps); 1551 rps->rp0_freq = rps->max_freq; 1552 drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n", 1553 intel_gpu_freq(rps, rps->max_freq), rps->max_freq); 1554 1555 rps->efficient_freq = vlv_rps_rpe_freq(rps); 1556 drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n", 1557 intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq); 1558 1559 rps->rp1_freq = vlv_rps_guar_freq(rps); 1560 drm_dbg(&i915->drm, "RP1(Guar Freq) GPU freq: %d MHz (%u)\n", 1561 intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq); 1562 1563 rps->min_freq = vlv_rps_min_freq(rps); 1564 drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n", 1565 intel_gpu_freq(rps, rps->min_freq), rps->min_freq); 1566 1567 vlv_iosf_sb_put(i915, 1568 BIT(VLV_IOSF_SB_PUNIT) | 1569 BIT(VLV_IOSF_SB_NC) | 1570 BIT(VLV_IOSF_SB_CCK)); 1571 } 1572 1573 static void chv_rps_init(struct intel_rps *rps) 1574 { 1575 struct drm_i915_private *i915 = rps_to_i915(rps); 1576 u32 val; 1577 1578 vlv_iosf_sb_get(i915, 1579 BIT(VLV_IOSF_SB_PUNIT) | 1580 BIT(VLV_IOSF_SB_NC) | 1581 BIT(VLV_IOSF_SB_CCK)); 1582 1583 vlv_init_gpll_ref_freq(rps); 1584 1585 val = vlv_cck_read(i915, CCK_FUSE_REG); 1586 1587 switch ((val >> 2) & 0x7) { 1588 case 3: 1589 i915->mem_freq = 2000; 1590 break; 1591 default: 1592 i915->mem_freq = 1600; 1593 break; 1594 } 1595 drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq); 1596 1597 rps->max_freq = chv_rps_max_freq(rps); 1598 rps->rp0_freq = rps->max_freq; 1599 drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n", 1600 intel_gpu_freq(rps, rps->max_freq), rps->max_freq); 1601 1602 rps->efficient_freq = chv_rps_rpe_freq(rps); 1603 drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n", 1604 intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq); 1605 1606 rps->rp1_freq = chv_rps_guar_freq(rps); 1607 drm_dbg(&i915->drm, "RP1(Guar) GPU freq: %d MHz (%u)\n", 1608 intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq); 1609 1610 rps->min_freq = chv_rps_min_freq(rps); 1611 drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n", 1612 intel_gpu_freq(rps, rps->min_freq), rps->min_freq); 1613 1614 vlv_iosf_sb_put(i915, 1615 BIT(VLV_IOSF_SB_PUNIT) | 1616 BIT(VLV_IOSF_SB_NC) | 1617 BIT(VLV_IOSF_SB_CCK)); 1618 1619 drm_WARN_ONCE(&i915->drm, (rps->max_freq | rps->efficient_freq | 1620 rps->rp1_freq | rps->min_freq) & 1, 1621 "Odd GPU freq values\n"); 1622 } 1623 1624 static void vlv_c0_read(struct intel_uncore *uncore, struct intel_rps_ei *ei) 1625 { 1626 ei->ktime = ktime_get_raw(); 1627 ei->render_c0 = intel_uncore_read(uncore, VLV_RENDER_C0_COUNT); 1628 ei->media_c0 = intel_uncore_read(uncore, VLV_MEDIA_C0_COUNT); 1629 } 1630 1631 static u32 vlv_wa_c0_ei(struct intel_rps *rps, u32 pm_iir) 1632 { 1633 struct intel_uncore *uncore = rps_to_uncore(rps); 1634 const struct intel_rps_ei *prev = &rps->ei; 1635 struct intel_rps_ei now; 1636 u32 events = 0; 1637 1638 if ((pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) == 0) 1639 return 0; 1640 1641 vlv_c0_read(uncore, &now); 1642 1643 if (prev->ktime) { 1644 u64 time, c0; 1645 u32 render, media; 1646 1647 time = ktime_us_delta(now.ktime, prev->ktime); 1648 1649 time *= rps_to_i915(rps)->czclk_freq; 1650 1651 /* Workload can be split between render + media, 1652 * e.g. SwapBuffers being blitted in X after being rendered in 1653 * mesa. To account for this we need to combine both engines 1654 * into our activity counter. 1655 */ 1656 render = now.render_c0 - prev->render_c0; 1657 media = now.media_c0 - prev->media_c0; 1658 c0 = max(render, media); 1659 c0 *= 1000 * 100 << 8; /* to usecs and scale to threshold% */ 1660 1661 if (c0 > time * rps->power.up_threshold) 1662 events = GEN6_PM_RP_UP_THRESHOLD; 1663 else if (c0 < time * rps->power.down_threshold) 1664 events = GEN6_PM_RP_DOWN_THRESHOLD; 1665 } 1666 1667 rps->ei = now; 1668 return events; 1669 } 1670 1671 static void rps_work(struct work_struct *work) 1672 { 1673 struct intel_rps *rps = container_of(work, typeof(*rps), work); 1674 struct intel_gt *gt = rps_to_gt(rps); 1675 struct drm_i915_private *i915 = rps_to_i915(rps); 1676 bool client_boost = false; 1677 int new_freq, adj, min, max; 1678 u32 pm_iir = 0; 1679 1680 spin_lock_irq(>->irq_lock); 1681 pm_iir = fetch_and_zero(&rps->pm_iir) & rps->pm_events; 1682 client_boost = atomic_read(&rps->num_waiters); 1683 spin_unlock_irq(>->irq_lock); 1684 1685 /* Make sure we didn't queue anything we're not going to process. */ 1686 if (!pm_iir && !client_boost) 1687 goto out; 1688 1689 mutex_lock(&rps->lock); 1690 if (!intel_rps_is_active(rps)) { 1691 mutex_unlock(&rps->lock); 1692 return; 1693 } 1694 1695 pm_iir |= vlv_wa_c0_ei(rps, pm_iir); 1696 1697 adj = rps->last_adj; 1698 new_freq = rps->cur_freq; 1699 min = rps->min_freq_softlimit; 1700 max = rps->max_freq_softlimit; 1701 if (client_boost) 1702 max = rps->max_freq; 1703 1704 GT_TRACE(gt, 1705 "pm_iir:%x, client_boost:%s, last:%d, cur:%x, min:%x, max:%x\n", 1706 pm_iir, yesno(client_boost), 1707 adj, new_freq, min, max); 1708 1709 if (client_boost && new_freq < rps->boost_freq) { 1710 new_freq = rps->boost_freq; 1711 adj = 0; 1712 } else if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) { 1713 if (adj > 0) 1714 adj *= 2; 1715 else /* CHV needs even encode values */ 1716 adj = IS_CHERRYVIEW(gt->i915) ? 2 : 1; 1717 1718 if (new_freq >= rps->max_freq_softlimit) 1719 adj = 0; 1720 } else if (client_boost) { 1721 adj = 0; 1722 } else if (pm_iir & GEN6_PM_RP_DOWN_TIMEOUT) { 1723 if (rps->cur_freq > rps->efficient_freq) 1724 new_freq = rps->efficient_freq; 1725 else if (rps->cur_freq > rps->min_freq_softlimit) 1726 new_freq = rps->min_freq_softlimit; 1727 adj = 0; 1728 } else if (pm_iir & GEN6_PM_RP_DOWN_THRESHOLD) { 1729 if (adj < 0) 1730 adj *= 2; 1731 else /* CHV needs even encode values */ 1732 adj = IS_CHERRYVIEW(gt->i915) ? -2 : -1; 1733 1734 if (new_freq <= rps->min_freq_softlimit) 1735 adj = 0; 1736 } else { /* unknown event */ 1737 adj = 0; 1738 } 1739 1740 /* 1741 * sysfs frequency limits may have snuck in while 1742 * servicing the interrupt 1743 */ 1744 new_freq += adj; 1745 new_freq = clamp_t(int, new_freq, min, max); 1746 1747 if (intel_rps_set(rps, new_freq)) { 1748 drm_dbg(&i915->drm, "Failed to set new GPU frequency\n"); 1749 adj = 0; 1750 } 1751 rps->last_adj = adj; 1752 1753 mutex_unlock(&rps->lock); 1754 1755 out: 1756 spin_lock_irq(>->irq_lock); 1757 gen6_gt_pm_unmask_irq(gt, rps->pm_events); 1758 spin_unlock_irq(>->irq_lock); 1759 } 1760 1761 void gen11_rps_irq_handler(struct intel_rps *rps, u32 pm_iir) 1762 { 1763 struct intel_gt *gt = rps_to_gt(rps); 1764 const u32 events = rps->pm_events & pm_iir; 1765 1766 lockdep_assert_held(>->irq_lock); 1767 1768 if (unlikely(!events)) 1769 return; 1770 1771 GT_TRACE(gt, "irq events:%x\n", events); 1772 1773 gen6_gt_pm_mask_irq(gt, events); 1774 1775 rps->pm_iir |= events; 1776 schedule_work(&rps->work); 1777 } 1778 1779 void gen6_rps_irq_handler(struct intel_rps *rps, u32 pm_iir) 1780 { 1781 struct intel_gt *gt = rps_to_gt(rps); 1782 u32 events; 1783 1784 events = pm_iir & rps->pm_events; 1785 if (events) { 1786 spin_lock(>->irq_lock); 1787 1788 GT_TRACE(gt, "irq events:%x\n", events); 1789 1790 gen6_gt_pm_mask_irq(gt, events); 1791 rps->pm_iir |= events; 1792 1793 schedule_work(&rps->work); 1794 spin_unlock(>->irq_lock); 1795 } 1796 1797 if (GRAPHICS_VER(gt->i915) >= 8) 1798 return; 1799 1800 if (pm_iir & PM_VEBOX_USER_INTERRUPT) 1801 intel_engine_cs_irq(gt->engine[VECS0], pm_iir >> 10); 1802 1803 if (pm_iir & PM_VEBOX_CS_ERROR_INTERRUPT) 1804 DRM_DEBUG("Command parser error, pm_iir 0x%08x\n", pm_iir); 1805 } 1806 1807 void gen5_rps_irq_handler(struct intel_rps *rps) 1808 { 1809 struct intel_uncore *uncore = rps_to_uncore(rps); 1810 u32 busy_up, busy_down, max_avg, min_avg; 1811 u8 new_freq; 1812 1813 spin_lock(&mchdev_lock); 1814 1815 intel_uncore_write16(uncore, 1816 MEMINTRSTS, 1817 intel_uncore_read(uncore, MEMINTRSTS)); 1818 1819 intel_uncore_write16(uncore, MEMINTRSTS, MEMINT_EVAL_CHG); 1820 busy_up = intel_uncore_read(uncore, RCPREVBSYTUPAVG); 1821 busy_down = intel_uncore_read(uncore, RCPREVBSYTDNAVG); 1822 max_avg = intel_uncore_read(uncore, RCBMAXAVG); 1823 min_avg = intel_uncore_read(uncore, RCBMINAVG); 1824 1825 /* Handle RCS change request from hw */ 1826 new_freq = rps->cur_freq; 1827 if (busy_up > max_avg) 1828 new_freq++; 1829 else if (busy_down < min_avg) 1830 new_freq--; 1831 new_freq = clamp(new_freq, 1832 rps->min_freq_softlimit, 1833 rps->max_freq_softlimit); 1834 1835 if (new_freq != rps->cur_freq && !__gen5_rps_set(rps, new_freq)) 1836 rps->cur_freq = new_freq; 1837 1838 spin_unlock(&mchdev_lock); 1839 } 1840 1841 void intel_rps_init_early(struct intel_rps *rps) 1842 { 1843 mutex_init(&rps->lock); 1844 mutex_init(&rps->power.mutex); 1845 1846 INIT_WORK(&rps->work, rps_work); 1847 timer_setup(&rps->timer, rps_timer, 0); 1848 1849 atomic_set(&rps->num_waiters, 0); 1850 } 1851 1852 void intel_rps_init(struct intel_rps *rps) 1853 { 1854 struct drm_i915_private *i915 = rps_to_i915(rps); 1855 1856 if (rps_uses_slpc(rps)) 1857 return; 1858 1859 if (IS_CHERRYVIEW(i915)) 1860 chv_rps_init(rps); 1861 else if (IS_VALLEYVIEW(i915)) 1862 vlv_rps_init(rps); 1863 else if (GRAPHICS_VER(i915) >= 6) 1864 gen6_rps_init(rps); 1865 else if (IS_IRONLAKE_M(i915)) 1866 gen5_rps_init(rps); 1867 1868 /* Derive initial user preferences/limits from the hardware limits */ 1869 rps->max_freq_softlimit = rps->max_freq; 1870 rps->min_freq_softlimit = rps->min_freq; 1871 1872 /* After setting max-softlimit, find the overclock max freq */ 1873 if (GRAPHICS_VER(i915) == 6 || IS_IVYBRIDGE(i915) || IS_HASWELL(i915)) { 1874 u32 params = 0; 1875 1876 sandybridge_pcode_read(i915, GEN6_READ_OC_PARAMS, 1877 ¶ms, NULL); 1878 if (params & BIT(31)) { /* OC supported */ 1879 drm_dbg(&i915->drm, 1880 "Overclocking supported, max: %dMHz, overclock: %dMHz\n", 1881 (rps->max_freq & 0xff) * 50, 1882 (params & 0xff) * 50); 1883 rps->max_freq = params & 0xff; 1884 } 1885 } 1886 1887 /* Finally allow us to boost to max by default */ 1888 rps->boost_freq = rps->max_freq; 1889 rps->idle_freq = rps->min_freq; 1890 1891 /* Start in the middle, from here we will autotune based on workload */ 1892 rps->cur_freq = rps->efficient_freq; 1893 1894 rps->pm_intrmsk_mbz = 0; 1895 1896 /* 1897 * SNB,IVB,HSW can while VLV,CHV may hard hang on looping batchbuffer 1898 * if GEN6_PM_UP_EI_EXPIRED is masked. 1899 * 1900 * TODO: verify if this can be reproduced on VLV,CHV. 1901 */ 1902 if (GRAPHICS_VER(i915) <= 7) 1903 rps->pm_intrmsk_mbz |= GEN6_PM_RP_UP_EI_EXPIRED; 1904 1905 if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) < 11) 1906 rps->pm_intrmsk_mbz |= GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC; 1907 1908 /* GuC needs ARAT expired interrupt unmasked */ 1909 if (intel_uc_uses_guc_submission(&rps_to_gt(rps)->uc)) 1910 rps->pm_intrmsk_mbz |= ARAT_EXPIRED_INTRMSK; 1911 } 1912 1913 void intel_rps_sanitize(struct intel_rps *rps) 1914 { 1915 if (rps_uses_slpc(rps)) 1916 return; 1917 1918 if (GRAPHICS_VER(rps_to_i915(rps)) >= 6) 1919 rps_disable_interrupts(rps); 1920 } 1921 1922 u32 intel_rps_get_cagf(struct intel_rps *rps, u32 rpstat) 1923 { 1924 struct drm_i915_private *i915 = rps_to_i915(rps); 1925 u32 cagf; 1926 1927 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) 1928 cagf = (rpstat >> 8) & 0xff; 1929 else if (GRAPHICS_VER(i915) >= 9) 1930 cagf = (rpstat & GEN9_CAGF_MASK) >> GEN9_CAGF_SHIFT; 1931 else if (IS_HASWELL(i915) || IS_BROADWELL(i915)) 1932 cagf = (rpstat & HSW_CAGF_MASK) >> HSW_CAGF_SHIFT; 1933 else if (GRAPHICS_VER(i915) >= 6) 1934 cagf = (rpstat & GEN6_CAGF_MASK) >> GEN6_CAGF_SHIFT; 1935 else 1936 cagf = gen5_invert_freq(rps, (rpstat & MEMSTAT_PSTATE_MASK) >> 1937 MEMSTAT_PSTATE_SHIFT); 1938 1939 return cagf; 1940 } 1941 1942 static u32 read_cagf(struct intel_rps *rps) 1943 { 1944 struct drm_i915_private *i915 = rps_to_i915(rps); 1945 struct intel_uncore *uncore = rps_to_uncore(rps); 1946 u32 freq; 1947 1948 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) { 1949 vlv_punit_get(i915); 1950 freq = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS); 1951 vlv_punit_put(i915); 1952 } else if (GRAPHICS_VER(i915) >= 6) { 1953 freq = intel_uncore_read(uncore, GEN6_RPSTAT1); 1954 } else { 1955 freq = intel_uncore_read(uncore, MEMSTAT_ILK); 1956 } 1957 1958 return intel_rps_get_cagf(rps, freq); 1959 } 1960 1961 u32 intel_rps_read_actual_frequency(struct intel_rps *rps) 1962 { 1963 struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm; 1964 intel_wakeref_t wakeref; 1965 u32 freq = 0; 1966 1967 with_intel_runtime_pm_if_in_use(rpm, wakeref) 1968 freq = intel_gpu_freq(rps, read_cagf(rps)); 1969 1970 return freq; 1971 } 1972 1973 u32 intel_rps_read_punit_req(struct intel_rps *rps) 1974 { 1975 struct intel_uncore *uncore = rps_to_uncore(rps); 1976 1977 return intel_uncore_read(uncore, GEN6_RPNSWREQ); 1978 } 1979 1980 static u32 intel_rps_get_req(u32 pureq) 1981 { 1982 u32 req = pureq >> GEN9_SW_REQ_UNSLICE_RATIO_SHIFT; 1983 1984 return req; 1985 } 1986 1987 u32 intel_rps_read_punit_req_frequency(struct intel_rps *rps) 1988 { 1989 u32 freq = intel_rps_get_req(intel_rps_read_punit_req(rps)); 1990 1991 return intel_gpu_freq(rps, freq); 1992 } 1993 1994 u32 intel_rps_get_requested_frequency(struct intel_rps *rps) 1995 { 1996 if (rps_uses_slpc(rps)) 1997 return intel_rps_read_punit_req_frequency(rps); 1998 else 1999 return intel_gpu_freq(rps, rps->cur_freq); 2000 } 2001 2002 u32 intel_rps_get_max_frequency(struct intel_rps *rps) 2003 { 2004 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2005 2006 if (rps_uses_slpc(rps)) 2007 return slpc->max_freq_softlimit; 2008 else 2009 return intel_gpu_freq(rps, rps->max_freq_softlimit); 2010 } 2011 2012 u32 intel_rps_get_rp0_frequency(struct intel_rps *rps) 2013 { 2014 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2015 2016 if (rps_uses_slpc(rps)) 2017 return slpc->rp0_freq; 2018 else 2019 return intel_gpu_freq(rps, rps->rp0_freq); 2020 } 2021 2022 u32 intel_rps_get_rp1_frequency(struct intel_rps *rps) 2023 { 2024 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2025 2026 if (rps_uses_slpc(rps)) 2027 return slpc->rp1_freq; 2028 else 2029 return intel_gpu_freq(rps, rps->rp1_freq); 2030 } 2031 2032 u32 intel_rps_get_rpn_frequency(struct intel_rps *rps) 2033 { 2034 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2035 2036 if (rps_uses_slpc(rps)) 2037 return slpc->min_freq; 2038 else 2039 return intel_gpu_freq(rps, rps->min_freq); 2040 } 2041 2042 static int set_max_freq(struct intel_rps *rps, u32 val) 2043 { 2044 struct drm_i915_private *i915 = rps_to_i915(rps); 2045 int ret = 0; 2046 2047 mutex_lock(&rps->lock); 2048 2049 val = intel_freq_opcode(rps, val); 2050 if (val < rps->min_freq || 2051 val > rps->max_freq || 2052 val < rps->min_freq_softlimit) { 2053 ret = -EINVAL; 2054 goto unlock; 2055 } 2056 2057 if (val > rps->rp0_freq) 2058 drm_dbg(&i915->drm, "User requested overclocking to %d\n", 2059 intel_gpu_freq(rps, val)); 2060 2061 rps->max_freq_softlimit = val; 2062 2063 val = clamp_t(int, rps->cur_freq, 2064 rps->min_freq_softlimit, 2065 rps->max_freq_softlimit); 2066 2067 /* 2068 * We still need *_set_rps to process the new max_delay and 2069 * update the interrupt limits and PMINTRMSK even though 2070 * frequency request may be unchanged. 2071 */ 2072 intel_rps_set(rps, val); 2073 2074 unlock: 2075 mutex_unlock(&rps->lock); 2076 2077 return ret; 2078 } 2079 2080 int intel_rps_set_max_frequency(struct intel_rps *rps, u32 val) 2081 { 2082 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2083 2084 if (rps_uses_slpc(rps)) 2085 return intel_guc_slpc_set_max_freq(slpc, val); 2086 else 2087 return set_max_freq(rps, val); 2088 } 2089 2090 u32 intel_rps_get_min_frequency(struct intel_rps *rps) 2091 { 2092 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2093 2094 if (rps_uses_slpc(rps)) 2095 return slpc->min_freq_softlimit; 2096 else 2097 return intel_gpu_freq(rps, rps->min_freq_softlimit); 2098 } 2099 2100 static int set_min_freq(struct intel_rps *rps, u32 val) 2101 { 2102 int ret = 0; 2103 2104 mutex_lock(&rps->lock); 2105 2106 val = intel_freq_opcode(rps, val); 2107 if (val < rps->min_freq || 2108 val > rps->max_freq || 2109 val > rps->max_freq_softlimit) { 2110 ret = -EINVAL; 2111 goto unlock; 2112 } 2113 2114 rps->min_freq_softlimit = val; 2115 2116 val = clamp_t(int, rps->cur_freq, 2117 rps->min_freq_softlimit, 2118 rps->max_freq_softlimit); 2119 2120 /* 2121 * We still need *_set_rps to process the new min_delay and 2122 * update the interrupt limits and PMINTRMSK even though 2123 * frequency request may be unchanged. 2124 */ 2125 intel_rps_set(rps, val); 2126 2127 unlock: 2128 mutex_unlock(&rps->lock); 2129 2130 return ret; 2131 } 2132 2133 int intel_rps_set_min_frequency(struct intel_rps *rps, u32 val) 2134 { 2135 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2136 2137 if (rps_uses_slpc(rps)) 2138 return intel_guc_slpc_set_min_freq(slpc, val); 2139 else 2140 return set_min_freq(rps, val); 2141 } 2142 2143 /* External interface for intel_ips.ko */ 2144 2145 static struct drm_i915_private __rcu *ips_mchdev; 2146 2147 /** 2148 * Tells the intel_ips driver that the i915 driver is now loaded, if 2149 * IPS got loaded first. 2150 * 2151 * This awkward dance is so that neither module has to depend on the 2152 * other in order for IPS to do the appropriate communication of 2153 * GPU turbo limits to i915. 2154 */ 2155 static void 2156 ips_ping_for_i915_load(void) 2157 { 2158 void (*link)(void); 2159 2160 link = symbol_get(ips_link_to_i915_driver); 2161 if (link) { 2162 link(); 2163 symbol_put(ips_link_to_i915_driver); 2164 } 2165 } 2166 2167 void intel_rps_driver_register(struct intel_rps *rps) 2168 { 2169 struct intel_gt *gt = rps_to_gt(rps); 2170 2171 /* 2172 * We only register the i915 ips part with intel-ips once everything is 2173 * set up, to avoid intel-ips sneaking in and reading bogus values. 2174 */ 2175 if (GRAPHICS_VER(gt->i915) == 5) { 2176 GEM_BUG_ON(ips_mchdev); 2177 rcu_assign_pointer(ips_mchdev, gt->i915); 2178 ips_ping_for_i915_load(); 2179 } 2180 } 2181 2182 void intel_rps_driver_unregister(struct intel_rps *rps) 2183 { 2184 if (rcu_access_pointer(ips_mchdev) == rps_to_i915(rps)) 2185 rcu_assign_pointer(ips_mchdev, NULL); 2186 } 2187 2188 static struct drm_i915_private *mchdev_get(void) 2189 { 2190 struct drm_i915_private *i915; 2191 2192 rcu_read_lock(); 2193 i915 = rcu_dereference(ips_mchdev); 2194 if (i915 && !kref_get_unless_zero(&i915->drm.ref)) 2195 i915 = NULL; 2196 rcu_read_unlock(); 2197 2198 return i915; 2199 } 2200 2201 /** 2202 * i915_read_mch_val - return value for IPS use 2203 * 2204 * Calculate and return a value for the IPS driver to use when deciding whether 2205 * we have thermal and power headroom to increase CPU or GPU power budget. 2206 */ 2207 unsigned long i915_read_mch_val(void) 2208 { 2209 struct drm_i915_private *i915; 2210 unsigned long chipset_val = 0; 2211 unsigned long graphics_val = 0; 2212 intel_wakeref_t wakeref; 2213 2214 i915 = mchdev_get(); 2215 if (!i915) 2216 return 0; 2217 2218 with_intel_runtime_pm(&i915->runtime_pm, wakeref) { 2219 struct intel_ips *ips = &i915->gt.rps.ips; 2220 2221 spin_lock_irq(&mchdev_lock); 2222 chipset_val = __ips_chipset_val(ips); 2223 graphics_val = __ips_gfx_val(ips); 2224 spin_unlock_irq(&mchdev_lock); 2225 } 2226 2227 drm_dev_put(&i915->drm); 2228 return chipset_val + graphics_val; 2229 } 2230 EXPORT_SYMBOL_GPL(i915_read_mch_val); 2231 2232 /** 2233 * i915_gpu_raise - raise GPU frequency limit 2234 * 2235 * Raise the limit; IPS indicates we have thermal headroom. 2236 */ 2237 bool i915_gpu_raise(void) 2238 { 2239 struct drm_i915_private *i915; 2240 struct intel_rps *rps; 2241 2242 i915 = mchdev_get(); 2243 if (!i915) 2244 return false; 2245 2246 rps = &i915->gt.rps; 2247 2248 spin_lock_irq(&mchdev_lock); 2249 if (rps->max_freq_softlimit < rps->max_freq) 2250 rps->max_freq_softlimit++; 2251 spin_unlock_irq(&mchdev_lock); 2252 2253 drm_dev_put(&i915->drm); 2254 return true; 2255 } 2256 EXPORT_SYMBOL_GPL(i915_gpu_raise); 2257 2258 /** 2259 * i915_gpu_lower - lower GPU frequency limit 2260 * 2261 * IPS indicates we're close to a thermal limit, so throttle back the GPU 2262 * frequency maximum. 2263 */ 2264 bool i915_gpu_lower(void) 2265 { 2266 struct drm_i915_private *i915; 2267 struct intel_rps *rps; 2268 2269 i915 = mchdev_get(); 2270 if (!i915) 2271 return false; 2272 2273 rps = &i915->gt.rps; 2274 2275 spin_lock_irq(&mchdev_lock); 2276 if (rps->max_freq_softlimit > rps->min_freq) 2277 rps->max_freq_softlimit--; 2278 spin_unlock_irq(&mchdev_lock); 2279 2280 drm_dev_put(&i915->drm); 2281 return true; 2282 } 2283 EXPORT_SYMBOL_GPL(i915_gpu_lower); 2284 2285 /** 2286 * i915_gpu_busy - indicate GPU business to IPS 2287 * 2288 * Tell the IPS driver whether or not the GPU is busy. 2289 */ 2290 bool i915_gpu_busy(void) 2291 { 2292 struct drm_i915_private *i915; 2293 bool ret; 2294 2295 i915 = mchdev_get(); 2296 if (!i915) 2297 return false; 2298 2299 ret = i915->gt.awake; 2300 2301 drm_dev_put(&i915->drm); 2302 return ret; 2303 } 2304 EXPORT_SYMBOL_GPL(i915_gpu_busy); 2305 2306 /** 2307 * i915_gpu_turbo_disable - disable graphics turbo 2308 * 2309 * Disable graphics turbo by resetting the max frequency and setting the 2310 * current frequency to the default. 2311 */ 2312 bool i915_gpu_turbo_disable(void) 2313 { 2314 struct drm_i915_private *i915; 2315 struct intel_rps *rps; 2316 bool ret; 2317 2318 i915 = mchdev_get(); 2319 if (!i915) 2320 return false; 2321 2322 rps = &i915->gt.rps; 2323 2324 spin_lock_irq(&mchdev_lock); 2325 rps->max_freq_softlimit = rps->min_freq; 2326 ret = !__gen5_rps_set(&i915->gt.rps, rps->min_freq); 2327 spin_unlock_irq(&mchdev_lock); 2328 2329 drm_dev_put(&i915->drm); 2330 return ret; 2331 } 2332 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable); 2333 2334 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2335 #include "selftest_rps.c" 2336 #include "selftest_slpc.c" 2337 #endif 2338