1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2019 Intel Corporation 4 */ 5 6 #include <linux/string_helpers.h> 7 8 #include <drm/intel/i915_drm.h> 9 10 #include "display/intel_display.h" 11 #include "display/intel_display_rps.h" 12 #include "soc/intel_dram.h" 13 14 #include "i915_drv.h" 15 #include "i915_irq.h" 16 #include "i915_reg.h" 17 #include "i915_wait_util.h" 18 #include "intel_breadcrumbs.h" 19 #include "intel_gt.h" 20 #include "intel_gt_clock_utils.h" 21 #include "intel_gt_irq.h" 22 #include "intel_gt_pm.h" 23 #include "intel_gt_pm_irq.h" 24 #include "intel_gt_print.h" 25 #include "intel_gt_regs.h" 26 #include "intel_mchbar_regs.h" 27 #include "intel_pcode.h" 28 #include "intel_rps.h" 29 #include "vlv_iosf_sb.h" 30 #include "../../../platform/x86/intel_ips.h" 31 32 #define BUSY_MAX_EI 20u /* ms */ 33 34 /* 35 * Lock protecting IPS related data structures 36 */ 37 static DEFINE_SPINLOCK(mchdev_lock); 38 39 static struct intel_gt *rps_to_gt(struct intel_rps *rps) 40 { 41 return container_of(rps, struct intel_gt, rps); 42 } 43 44 static struct drm_i915_private *rps_to_i915(struct intel_rps *rps) 45 { 46 return rps_to_gt(rps)->i915; 47 } 48 49 static struct intel_uncore *rps_to_uncore(struct intel_rps *rps) 50 { 51 return rps_to_gt(rps)->uncore; 52 } 53 54 static struct intel_guc_slpc *rps_to_slpc(struct intel_rps *rps) 55 { 56 struct intel_gt *gt = rps_to_gt(rps); 57 58 return >_to_guc(gt)->slpc; 59 } 60 61 static bool rps_uses_slpc(struct intel_rps *rps) 62 { 63 struct intel_gt *gt = rps_to_gt(rps); 64 65 return intel_uc_uses_guc_slpc(>->uc); 66 } 67 68 static u32 rps_pm_sanitize_mask(struct intel_rps *rps, u32 mask) 69 { 70 return mask & ~rps->pm_intrmsk_mbz; 71 } 72 73 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val) 74 { 75 intel_uncore_write_fw(uncore, reg, val); 76 } 77 78 static void rps_timer(struct timer_list *t) 79 { 80 struct intel_rps *rps = timer_container_of(rps, t, timer); 81 struct intel_gt *gt = rps_to_gt(rps); 82 struct intel_engine_cs *engine; 83 ktime_t dt, last, timestamp; 84 enum intel_engine_id id; 85 s64 max_busy[3] = {}; 86 87 timestamp = 0; 88 for_each_engine(engine, gt, id) { 89 s64 busy; 90 int i; 91 92 dt = intel_engine_get_busy_time(engine, ×tamp); 93 last = engine->stats.rps; 94 engine->stats.rps = dt; 95 96 busy = ktime_to_ns(ktime_sub(dt, last)); 97 for (i = 0; i < ARRAY_SIZE(max_busy); i++) { 98 if (busy > max_busy[i]) 99 swap(busy, max_busy[i]); 100 } 101 } 102 last = rps->pm_timestamp; 103 rps->pm_timestamp = timestamp; 104 105 if (intel_rps_is_active(rps)) { 106 s64 busy; 107 int i; 108 109 dt = ktime_sub(timestamp, last); 110 111 /* 112 * Our goal is to evaluate each engine independently, so we run 113 * at the lowest clocks required to sustain the heaviest 114 * workload. However, a task may be split into sequential 115 * dependent operations across a set of engines, such that 116 * the independent contributions do not account for high load, 117 * but overall the task is GPU bound. For example, consider 118 * video decode on vcs followed by colour post-processing 119 * on vecs, followed by general post-processing on rcs. 120 * Since multi-engines being active does imply a single 121 * continuous workload across all engines, we hedge our 122 * bets by only contributing a factor of the distributed 123 * load into our busyness calculation. 124 */ 125 busy = max_busy[0]; 126 for (i = 1; i < ARRAY_SIZE(max_busy); i++) { 127 if (!max_busy[i]) 128 break; 129 130 busy += div_u64(max_busy[i], 1 << i); 131 } 132 GT_TRACE(gt, 133 "busy:%lld [%d%%], max:[%lld, %lld, %lld], interval:%d\n", 134 busy, (int)div64_u64(100 * busy, dt), 135 max_busy[0], max_busy[1], max_busy[2], 136 rps->pm_interval); 137 138 if (100 * busy > rps->power.up_threshold * dt && 139 rps->cur_freq < rps->max_freq_softlimit) { 140 rps->pm_iir |= GEN6_PM_RP_UP_THRESHOLD; 141 rps->pm_interval = 1; 142 queue_work(gt->i915->unordered_wq, &rps->work); 143 } else if (100 * busy < rps->power.down_threshold * dt && 144 rps->cur_freq > rps->min_freq_softlimit) { 145 rps->pm_iir |= GEN6_PM_RP_DOWN_THRESHOLD; 146 rps->pm_interval = 1; 147 queue_work(gt->i915->unordered_wq, &rps->work); 148 } else { 149 rps->last_adj = 0; 150 } 151 152 mod_timer(&rps->timer, 153 jiffies + msecs_to_jiffies(rps->pm_interval)); 154 rps->pm_interval = min(rps->pm_interval * 2, BUSY_MAX_EI); 155 } 156 } 157 158 static void rps_start_timer(struct intel_rps *rps) 159 { 160 rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp); 161 rps->pm_interval = 1; 162 mod_timer(&rps->timer, jiffies + 1); 163 } 164 165 static void rps_stop_timer(struct intel_rps *rps) 166 { 167 timer_delete_sync(&rps->timer); 168 rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp); 169 cancel_work_sync(&rps->work); 170 } 171 172 static u32 rps_pm_mask(struct intel_rps *rps, u8 val) 173 { 174 u32 mask = 0; 175 176 /* We use UP_EI_EXPIRED interrupts for both up/down in manual mode */ 177 if (val > rps->min_freq_softlimit) 178 mask |= (GEN6_PM_RP_UP_EI_EXPIRED | 179 GEN6_PM_RP_DOWN_THRESHOLD | 180 GEN6_PM_RP_DOWN_TIMEOUT); 181 182 if (val < rps->max_freq_softlimit) 183 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD; 184 185 mask &= rps->pm_events; 186 187 return rps_pm_sanitize_mask(rps, ~mask); 188 } 189 190 static void rps_reset_ei(struct intel_rps *rps) 191 { 192 memset(&rps->ei, 0, sizeof(rps->ei)); 193 } 194 195 static void rps_enable_interrupts(struct intel_rps *rps) 196 { 197 struct intel_gt *gt = rps_to_gt(rps); 198 199 GEM_BUG_ON(rps_uses_slpc(rps)); 200 201 GT_TRACE(gt, "interrupts:on rps->pm_events: %x, rps_pm_mask:%x\n", 202 rps->pm_events, rps_pm_mask(rps, rps->last_freq)); 203 204 rps_reset_ei(rps); 205 206 spin_lock_irq(gt->irq_lock); 207 gen6_gt_pm_enable_irq(gt, rps->pm_events); 208 spin_unlock_irq(gt->irq_lock); 209 210 intel_uncore_write(gt->uncore, 211 GEN6_PMINTRMSK, rps_pm_mask(rps, rps->last_freq)); 212 } 213 214 static void gen6_rps_reset_interrupts(struct intel_rps *rps) 215 { 216 gen6_gt_pm_reset_iir(rps_to_gt(rps), GEN6_PM_RPS_EVENTS); 217 } 218 219 static void gen11_rps_reset_interrupts(struct intel_rps *rps) 220 { 221 while (gen11_gt_reset_one_iir(rps_to_gt(rps), 0, GEN11_GTPM)) 222 ; 223 } 224 225 static void rps_reset_interrupts(struct intel_rps *rps) 226 { 227 struct intel_gt *gt = rps_to_gt(rps); 228 229 spin_lock_irq(gt->irq_lock); 230 if (GRAPHICS_VER(gt->i915) >= 11) 231 gen11_rps_reset_interrupts(rps); 232 else 233 gen6_rps_reset_interrupts(rps); 234 235 rps->pm_iir = 0; 236 spin_unlock_irq(gt->irq_lock); 237 } 238 239 static void rps_disable_interrupts(struct intel_rps *rps) 240 { 241 struct intel_gt *gt = rps_to_gt(rps); 242 243 intel_uncore_write(gt->uncore, 244 GEN6_PMINTRMSK, rps_pm_sanitize_mask(rps, ~0u)); 245 246 spin_lock_irq(gt->irq_lock); 247 gen6_gt_pm_disable_irq(gt, GEN6_PM_RPS_EVENTS); 248 spin_unlock_irq(gt->irq_lock); 249 250 intel_synchronize_irq(gt->i915); 251 252 /* 253 * Now that we will not be generating any more work, flush any 254 * outstanding tasks. As we are called on the RPS idle path, 255 * we will reset the GPU to minimum frequencies, so the current 256 * state of the worker can be discarded. 257 */ 258 cancel_work_sync(&rps->work); 259 260 rps_reset_interrupts(rps); 261 GT_TRACE(gt, "interrupts:off\n"); 262 } 263 264 static const struct cparams { 265 u16 i; 266 u16 t; 267 u16 m; 268 u16 c; 269 } cparams[] = { 270 { 1, 1333, 301, 28664 }, 271 { 1, 1067, 294, 24460 }, 272 { 1, 800, 294, 25192 }, 273 { 0, 1333, 276, 27605 }, 274 { 0, 1067, 276, 27605 }, 275 { 0, 800, 231, 23784 }, 276 }; 277 278 static void gen5_rps_init(struct intel_rps *rps) 279 { 280 struct drm_i915_private *i915 = rps_to_i915(rps); 281 struct intel_uncore *uncore = rps_to_uncore(rps); 282 unsigned int fsb_freq, mem_freq; 283 u8 fmax, fmin, fstart; 284 u32 rgvmodectl; 285 int c_m, i; 286 287 fsb_freq = intel_fsb_freq(i915); 288 mem_freq = intel_mem_freq(i915); 289 290 if (fsb_freq <= 3200000) 291 c_m = 0; 292 else if (fsb_freq <= 4800000) 293 c_m = 1; 294 else 295 c_m = 2; 296 297 for (i = 0; i < ARRAY_SIZE(cparams); i++) { 298 if (cparams[i].i == c_m && 299 cparams[i].t == DIV_ROUND_CLOSEST(mem_freq, 1000)) { 300 rps->ips.m = cparams[i].m; 301 rps->ips.c = cparams[i].c; 302 break; 303 } 304 } 305 306 rgvmodectl = intel_uncore_read(uncore, MEMMODECTL); 307 308 /* Set up min, max, and cur for interrupt handling */ 309 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT; 310 fmin = (rgvmodectl & MEMMODE_FMIN_MASK); 311 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >> 312 MEMMODE_FSTART_SHIFT; 313 drm_dbg(&i915->drm, "fmax: %d, fmin: %d, fstart: %d\n", 314 fmax, fmin, fstart); 315 316 rps->min_freq = fmax; 317 rps->efficient_freq = fstart; 318 rps->max_freq = fmin; 319 } 320 321 static unsigned long 322 __ips_chipset_val(struct intel_ips *ips) 323 { 324 struct intel_uncore *uncore = 325 rps_to_uncore(container_of(ips, struct intel_rps, ips)); 326 unsigned long now = jiffies_to_msecs(jiffies), dt; 327 unsigned long result; 328 u64 total, delta; 329 330 lockdep_assert_held(&mchdev_lock); 331 332 /* 333 * Prevent division-by-zero if we are asking too fast. 334 * Also, we don't get interesting results if we are polling 335 * faster than once in 10ms, so just return the saved value 336 * in such cases. 337 */ 338 dt = now - ips->last_time1; 339 if (dt <= 10) 340 return ips->chipset_power; 341 342 /* FIXME: handle per-counter overflow */ 343 total = intel_uncore_read(uncore, DMIEC); 344 total += intel_uncore_read(uncore, DDREC); 345 total += intel_uncore_read(uncore, CSIEC); 346 347 delta = total - ips->last_count1; 348 349 result = div_u64(div_u64(ips->m * delta, dt) + ips->c, 10); 350 351 ips->last_count1 = total; 352 ips->last_time1 = now; 353 354 ips->chipset_power = result; 355 356 return result; 357 } 358 359 static unsigned long ips_mch_val(struct intel_uncore *uncore) 360 { 361 unsigned int m, x, b; 362 u32 tsfs; 363 364 tsfs = intel_uncore_read(uncore, TSFS); 365 x = intel_uncore_read8(uncore, TR1); 366 367 b = tsfs & TSFS_INTR_MASK; 368 m = (tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT; 369 370 return m * x / 127 - b; 371 } 372 373 static int _pxvid_to_vd(u8 pxvid) 374 { 375 if (pxvid == 0) 376 return 0; 377 378 if (pxvid >= 8 && pxvid < 31) 379 pxvid = 31; 380 381 return (pxvid + 2) * 125; 382 } 383 384 static u32 pvid_to_extvid(struct drm_i915_private *i915, u8 pxvid) 385 { 386 const int vd = _pxvid_to_vd(pxvid); 387 388 if (INTEL_INFO(i915)->is_mobile) 389 return max(vd - 1125, 0); 390 391 return vd; 392 } 393 394 static void __gen5_ips_update(struct intel_ips *ips) 395 { 396 struct intel_uncore *uncore = 397 rps_to_uncore(container_of(ips, struct intel_rps, ips)); 398 u64 now, delta, dt; 399 u32 count; 400 401 lockdep_assert_held(&mchdev_lock); 402 403 now = ktime_get_raw_ns(); 404 dt = now - ips->last_time2; 405 do_div(dt, NSEC_PER_MSEC); 406 407 /* Don't divide by 0 */ 408 if (dt <= 10) 409 return; 410 411 count = intel_uncore_read(uncore, GFXEC); 412 delta = count - ips->last_count2; 413 414 ips->last_count2 = count; 415 ips->last_time2 = now; 416 417 /* More magic constants... */ 418 ips->gfx_power = div_u64(delta * 1181, dt * 10); 419 } 420 421 static void gen5_rps_update(struct intel_rps *rps) 422 { 423 spin_lock_irq(&mchdev_lock); 424 __gen5_ips_update(&rps->ips); 425 spin_unlock_irq(&mchdev_lock); 426 } 427 428 static unsigned int gen5_invert_freq(struct intel_rps *rps, 429 unsigned int val) 430 { 431 /* Invert the frequency bin into an ips delay */ 432 val = rps->max_freq - val; 433 val = rps->min_freq + val; 434 435 return val; 436 } 437 438 static int __gen5_rps_set(struct intel_rps *rps, u8 val) 439 { 440 struct intel_uncore *uncore = rps_to_uncore(rps); 441 u16 rgvswctl; 442 443 lockdep_assert_held(&mchdev_lock); 444 445 rgvswctl = intel_uncore_read16(uncore, MEMSWCTL); 446 if (rgvswctl & MEMCTL_CMD_STS) { 447 drm_dbg(&rps_to_i915(rps)->drm, 448 "gpu busy, RCS change rejected\n"); 449 return -EBUSY; /* still busy with another command */ 450 } 451 452 /* Invert the frequency bin into an ips delay */ 453 val = gen5_invert_freq(rps, val); 454 455 rgvswctl = 456 (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) | 457 (val << MEMCTL_FREQ_SHIFT) | 458 MEMCTL_SFCAVM; 459 intel_uncore_write16(uncore, MEMSWCTL, rgvswctl); 460 intel_uncore_posting_read16(uncore, MEMSWCTL); 461 462 rgvswctl |= MEMCTL_CMD_STS; 463 intel_uncore_write16(uncore, MEMSWCTL, rgvswctl); 464 465 return 0; 466 } 467 468 static int gen5_rps_set(struct intel_rps *rps, u8 val) 469 { 470 int err; 471 472 spin_lock_irq(&mchdev_lock); 473 err = __gen5_rps_set(rps, val); 474 spin_unlock_irq(&mchdev_lock); 475 476 return err; 477 } 478 479 static unsigned long intel_pxfreq(u32 vidfreq) 480 { 481 int div = (vidfreq & 0x3f0000) >> 16; 482 int post = (vidfreq & 0x3000) >> 12; 483 int pre = (vidfreq & 0x7); 484 485 if (!pre) 486 return 0; 487 488 return div * 133333 / (pre << post); 489 } 490 491 static unsigned int init_emon(struct intel_uncore *uncore) 492 { 493 u8 pxw[16]; 494 int i; 495 496 /* Disable to program */ 497 intel_uncore_write(uncore, ECR, 0); 498 intel_uncore_posting_read(uncore, ECR); 499 500 /* Program energy weights for various events */ 501 intel_uncore_write(uncore, SDEW, 0x15040d00); 502 intel_uncore_write(uncore, CSIEW0, 0x007f0000); 503 intel_uncore_write(uncore, CSIEW1, 0x1e220004); 504 intel_uncore_write(uncore, CSIEW2, 0x04000004); 505 506 for (i = 0; i < 5; i++) 507 intel_uncore_write(uncore, PEW(i), 0); 508 for (i = 0; i < 3; i++) 509 intel_uncore_write(uncore, DEW(i), 0); 510 511 /* Program P-state weights to account for frequency power adjustment */ 512 for (i = 0; i < 16; i++) { 513 u32 pxvidfreq = intel_uncore_read(uncore, PXVFREQ(i)); 514 unsigned int freq = intel_pxfreq(pxvidfreq); 515 unsigned int vid = 516 (pxvidfreq & PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT; 517 unsigned int val; 518 519 val = vid * vid * freq / 1000 * 255; 520 val /= 127 * 127 * 900; 521 522 pxw[i] = val; 523 } 524 /* Render standby states get 0 weight */ 525 pxw[14] = 0; 526 pxw[15] = 0; 527 528 for (i = 0; i < 4; i++) { 529 intel_uncore_write(uncore, PXW(i), 530 pxw[i * 4 + 0] << 24 | 531 pxw[i * 4 + 1] << 16 | 532 pxw[i * 4 + 2] << 8 | 533 pxw[i * 4 + 3] << 0); 534 } 535 536 /* Adjust magic regs to magic values (more experimental results) */ 537 intel_uncore_write(uncore, OGW0, 0); 538 intel_uncore_write(uncore, OGW1, 0); 539 intel_uncore_write(uncore, EG0, 0x00007f00); 540 intel_uncore_write(uncore, EG1, 0x0000000e); 541 intel_uncore_write(uncore, EG2, 0x000e0000); 542 intel_uncore_write(uncore, EG3, 0x68000300); 543 intel_uncore_write(uncore, EG4, 0x42000000); 544 intel_uncore_write(uncore, EG5, 0x00140031); 545 intel_uncore_write(uncore, EG6, 0); 546 intel_uncore_write(uncore, EG7, 0); 547 548 for (i = 0; i < 8; i++) 549 intel_uncore_write(uncore, PXWL(i), 0); 550 551 /* Enable PMON + select events */ 552 intel_uncore_write(uncore, ECR, 0x80000019); 553 554 return intel_uncore_read(uncore, LCFUSE02) & LCFUSE_HIV_MASK; 555 } 556 557 static bool gen5_rps_enable(struct intel_rps *rps) 558 { 559 struct drm_i915_private *i915 = rps_to_i915(rps); 560 struct intel_display *display = i915->display; 561 struct intel_uncore *uncore = rps_to_uncore(rps); 562 u8 fstart, vstart; 563 u32 rgvmodectl; 564 565 spin_lock_irq(&mchdev_lock); 566 567 rgvmodectl = intel_uncore_read(uncore, MEMMODECTL); 568 569 /* Enable temp reporting */ 570 intel_uncore_write16(uncore, PMMISC, 571 intel_uncore_read16(uncore, PMMISC) | MCPPCE_EN); 572 intel_uncore_write16(uncore, TSC1, 573 intel_uncore_read16(uncore, TSC1) | TSE); 574 575 /* 100ms RC evaluation intervals */ 576 intel_uncore_write(uncore, RCUPEI, 100000); 577 intel_uncore_write(uncore, RCDNEI, 100000); 578 579 /* Set max/min thresholds to 90ms and 80ms respectively */ 580 intel_uncore_write(uncore, RCBMAXAVG, 90000); 581 intel_uncore_write(uncore, RCBMINAVG, 80000); 582 583 intel_uncore_write(uncore, MEMIHYST, 1); 584 585 /* Set up min, max, and cur for interrupt handling */ 586 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >> 587 MEMMODE_FSTART_SHIFT; 588 589 vstart = (intel_uncore_read(uncore, PXVFREQ(fstart)) & 590 PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT; 591 592 intel_uncore_write(uncore, 593 MEMINTREN, 594 MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN); 595 596 intel_uncore_write(uncore, VIDSTART, vstart); 597 intel_uncore_posting_read(uncore, VIDSTART); 598 599 rgvmodectl |= MEMMODE_SWMODE_EN; 600 intel_uncore_write(uncore, MEMMODECTL, rgvmodectl); 601 602 if (wait_for_atomic((intel_uncore_read(uncore, MEMSWCTL) & 603 MEMCTL_CMD_STS) == 0, 10)) 604 drm_err(&uncore->i915->drm, 605 "stuck trying to change perf mode\n"); 606 mdelay(1); 607 608 __gen5_rps_set(rps, rps->cur_freq); 609 610 rps->ips.last_count1 = intel_uncore_read(uncore, DMIEC); 611 rps->ips.last_count1 += intel_uncore_read(uncore, DDREC); 612 rps->ips.last_count1 += intel_uncore_read(uncore, CSIEC); 613 rps->ips.last_time1 = jiffies_to_msecs(jiffies); 614 615 rps->ips.last_count2 = intel_uncore_read(uncore, GFXEC); 616 rps->ips.last_time2 = ktime_get_raw_ns(); 617 618 ilk_display_rps_enable(display); 619 620 spin_unlock_irq(&mchdev_lock); 621 622 rps->ips.corr = init_emon(uncore); 623 624 return true; 625 } 626 627 static void gen5_rps_disable(struct intel_rps *rps) 628 { 629 struct drm_i915_private *i915 = rps_to_i915(rps); 630 struct intel_display *display = i915->display; 631 struct intel_uncore *uncore = rps_to_uncore(rps); 632 u16 rgvswctl; 633 634 spin_lock_irq(&mchdev_lock); 635 636 ilk_display_rps_disable(display); 637 638 rgvswctl = intel_uncore_read16(uncore, MEMSWCTL); 639 640 /* Ack interrupts, disable EFC interrupt */ 641 intel_uncore_rmw(uncore, MEMINTREN, MEMINT_EVAL_CHG_EN, 0); 642 intel_uncore_write(uncore, MEMINTRSTS, MEMINT_EVAL_CHG); 643 644 /* Go back to the starting frequency */ 645 __gen5_rps_set(rps, rps->idle_freq); 646 mdelay(1); 647 rgvswctl |= MEMCTL_CMD_STS; 648 intel_uncore_write(uncore, MEMSWCTL, rgvswctl); 649 mdelay(1); 650 651 spin_unlock_irq(&mchdev_lock); 652 } 653 654 static u32 rps_limits(struct intel_rps *rps, u8 val) 655 { 656 u32 limits; 657 658 /* 659 * Only set the down limit when we've reached the lowest level to avoid 660 * getting more interrupts, otherwise leave this clear. This prevents a 661 * race in the hw when coming out of rc6: There's a tiny window where 662 * the hw runs at the minimal clock before selecting the desired 663 * frequency, if the down threshold expires in that window we will not 664 * receive a down interrupt. 665 */ 666 if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) { 667 limits = rps->max_freq_softlimit << 23; 668 if (val <= rps->min_freq_softlimit) 669 limits |= rps->min_freq_softlimit << 14; 670 } else { 671 limits = rps->max_freq_softlimit << 24; 672 if (val <= rps->min_freq_softlimit) 673 limits |= rps->min_freq_softlimit << 16; 674 } 675 676 return limits; 677 } 678 679 static void rps_set_power(struct intel_rps *rps, int new_power) 680 { 681 struct intel_gt *gt = rps_to_gt(rps); 682 struct intel_uncore *uncore = gt->uncore; 683 u32 ei_up = 0, ei_down = 0; 684 685 lockdep_assert_held(&rps->power.mutex); 686 687 if (new_power == rps->power.mode) 688 return; 689 690 /* Note the units here are not exactly 1us, but 1280ns. */ 691 switch (new_power) { 692 case LOW_POWER: 693 ei_up = 16000; 694 ei_down = 32000; 695 break; 696 697 case BETWEEN: 698 ei_up = 13000; 699 ei_down = 32000; 700 break; 701 702 case HIGH_POWER: 703 ei_up = 10000; 704 ei_down = 32000; 705 break; 706 } 707 708 /* When byt can survive without system hang with dynamic 709 * sw freq adjustments, this restriction can be lifted. 710 */ 711 if (IS_VALLEYVIEW(gt->i915)) 712 goto skip_hw_write; 713 714 GT_TRACE(gt, 715 "changing power mode [%d], up %d%% @ %dus, down %d%% @ %dus\n", 716 new_power, 717 rps->power.up_threshold, ei_up, 718 rps->power.down_threshold, ei_down); 719 720 set(uncore, GEN6_RP_UP_EI, 721 intel_gt_ns_to_pm_interval(gt, ei_up * 1000)); 722 set(uncore, GEN6_RP_UP_THRESHOLD, 723 intel_gt_ns_to_pm_interval(gt, 724 ei_up * rps->power.up_threshold * 10)); 725 726 set(uncore, GEN6_RP_DOWN_EI, 727 intel_gt_ns_to_pm_interval(gt, ei_down * 1000)); 728 set(uncore, GEN6_RP_DOWN_THRESHOLD, 729 intel_gt_ns_to_pm_interval(gt, 730 ei_down * 731 rps->power.down_threshold * 10)); 732 733 set(uncore, GEN6_RP_CONTROL, 734 (GRAPHICS_VER(gt->i915) > 9 ? 0 : GEN6_RP_MEDIA_TURBO) | 735 GEN6_RP_MEDIA_HW_NORMAL_MODE | 736 GEN6_RP_MEDIA_IS_GFX | 737 GEN6_RP_ENABLE | 738 GEN6_RP_UP_BUSY_AVG | 739 GEN6_RP_DOWN_IDLE_AVG); 740 741 skip_hw_write: 742 rps->power.mode = new_power; 743 } 744 745 static void gen6_rps_set_thresholds(struct intel_rps *rps, u8 val) 746 { 747 int new_power; 748 749 new_power = rps->power.mode; 750 switch (rps->power.mode) { 751 case LOW_POWER: 752 if (val > rps->efficient_freq + 1 && 753 val > rps->cur_freq) 754 new_power = BETWEEN; 755 break; 756 757 case BETWEEN: 758 if (val <= rps->efficient_freq && 759 val < rps->cur_freq) 760 new_power = LOW_POWER; 761 else if (val >= rps->rp0_freq && 762 val > rps->cur_freq) 763 new_power = HIGH_POWER; 764 break; 765 766 case HIGH_POWER: 767 if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 && 768 val < rps->cur_freq) 769 new_power = BETWEEN; 770 break; 771 } 772 /* Max/min bins are special */ 773 if (val <= rps->min_freq_softlimit) 774 new_power = LOW_POWER; 775 if (val >= rps->max_freq_softlimit) 776 new_power = HIGH_POWER; 777 778 mutex_lock(&rps->power.mutex); 779 if (rps->power.interactive) 780 new_power = HIGH_POWER; 781 rps_set_power(rps, new_power); 782 mutex_unlock(&rps->power.mutex); 783 } 784 785 void intel_rps_mark_interactive(struct intel_rps *rps, bool interactive) 786 { 787 GT_TRACE(rps_to_gt(rps), "mark interactive: %s\n", 788 str_yes_no(interactive)); 789 790 mutex_lock(&rps->power.mutex); 791 if (interactive) { 792 if (!rps->power.interactive++ && intel_rps_is_active(rps)) 793 rps_set_power(rps, HIGH_POWER); 794 } else { 795 GEM_BUG_ON(!rps->power.interactive); 796 rps->power.interactive--; 797 } 798 mutex_unlock(&rps->power.mutex); 799 } 800 801 static int gen6_rps_set(struct intel_rps *rps, u8 val) 802 { 803 struct intel_uncore *uncore = rps_to_uncore(rps); 804 struct drm_i915_private *i915 = rps_to_i915(rps); 805 u32 swreq; 806 807 GEM_BUG_ON(rps_uses_slpc(rps)); 808 809 if (GRAPHICS_VER(i915) >= 9) 810 swreq = GEN9_FREQUENCY(val); 811 else if (IS_HASWELL(i915) || IS_BROADWELL(i915)) 812 swreq = HSW_FREQUENCY(val); 813 else 814 swreq = (GEN6_FREQUENCY(val) | 815 GEN6_OFFSET(0) | 816 GEN6_AGGRESSIVE_TURBO); 817 set(uncore, GEN6_RPNSWREQ, swreq); 818 819 GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d, swreq:%x\n", 820 val, intel_gpu_freq(rps, val), swreq); 821 822 return 0; 823 } 824 825 static int vlv_rps_set(struct intel_rps *rps, u8 val) 826 { 827 struct drm_i915_private *i915 = rps_to_i915(rps); 828 int err; 829 830 vlv_iosf_sb_get(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 831 err = vlv_iosf_sb_write(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_REG_GPU_FREQ_REQ, val); 832 vlv_iosf_sb_put(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 833 834 GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d\n", 835 val, intel_gpu_freq(rps, val)); 836 837 return err; 838 } 839 840 static int rps_set(struct intel_rps *rps, u8 val, bool update) 841 { 842 struct drm_i915_private *i915 = rps_to_i915(rps); 843 int err; 844 845 if (val == rps->last_freq) 846 return 0; 847 848 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) 849 err = vlv_rps_set(rps, val); 850 else if (GRAPHICS_VER(i915) >= 6) 851 err = gen6_rps_set(rps, val); 852 else 853 err = gen5_rps_set(rps, val); 854 if (err) 855 return err; 856 857 if (update && GRAPHICS_VER(i915) >= 6) 858 gen6_rps_set_thresholds(rps, val); 859 rps->last_freq = val; 860 861 return 0; 862 } 863 864 void intel_rps_unpark(struct intel_rps *rps) 865 { 866 if (!intel_rps_is_enabled(rps)) 867 return; 868 869 GT_TRACE(rps_to_gt(rps), "unpark:%x\n", rps->cur_freq); 870 871 /* 872 * Use the user's desired frequency as a guide, but for better 873 * performance, jump directly to RPe as our starting frequency. 874 */ 875 mutex_lock(&rps->lock); 876 877 intel_rps_set_active(rps); 878 intel_rps_set(rps, 879 clamp(rps->cur_freq, 880 rps->min_freq_softlimit, 881 rps->max_freq_softlimit)); 882 883 mutex_unlock(&rps->lock); 884 885 rps->pm_iir = 0; 886 if (intel_rps_has_interrupts(rps)) 887 rps_enable_interrupts(rps); 888 if (intel_rps_uses_timer(rps)) 889 rps_start_timer(rps); 890 891 if (GRAPHICS_VER(rps_to_i915(rps)) == 5) 892 gen5_rps_update(rps); 893 } 894 895 void intel_rps_park(struct intel_rps *rps) 896 { 897 int adj; 898 899 if (!intel_rps_is_enabled(rps)) 900 return; 901 902 if (!intel_rps_clear_active(rps)) 903 return; 904 905 if (intel_rps_uses_timer(rps)) 906 rps_stop_timer(rps); 907 if (intel_rps_has_interrupts(rps)) 908 rps_disable_interrupts(rps); 909 910 if (rps->last_freq <= rps->idle_freq) 911 return; 912 913 /* 914 * The punit delays the write of the frequency and voltage until it 915 * determines the GPU is awake. During normal usage we don't want to 916 * waste power changing the frequency if the GPU is sleeping (rc6). 917 * However, the GPU and driver is now idle and we do not want to delay 918 * switching to minimum voltage (reducing power whilst idle) as we do 919 * not expect to be woken in the near future and so must flush the 920 * change by waking the device. 921 * 922 * We choose to take the media powerwell (either would do to trick the 923 * punit into committing the voltage change) as that takes a lot less 924 * power than the render powerwell. 925 */ 926 intel_uncore_forcewake_get(rps_to_uncore(rps), FORCEWAKE_MEDIA); 927 rps_set(rps, rps->idle_freq, false); 928 intel_uncore_forcewake_put(rps_to_uncore(rps), FORCEWAKE_MEDIA); 929 930 /* 931 * Since we will try and restart from the previously requested 932 * frequency on unparking, treat this idle point as a downclock 933 * interrupt and reduce the frequency for resume. If we park/unpark 934 * more frequently than the rps worker can run, we will not respond 935 * to any EI and never see a change in frequency. 936 * 937 * (Note we accommodate Cherryview's limitation of only using an 938 * even bin by applying it to all.) 939 */ 940 adj = rps->last_adj; 941 if (adj < 0) 942 adj *= 2; 943 else /* CHV needs even encode values */ 944 adj = -2; 945 rps->last_adj = adj; 946 rps->cur_freq = max_t(int, rps->cur_freq + adj, rps->min_freq); 947 if (rps->cur_freq < rps->efficient_freq) { 948 rps->cur_freq = rps->efficient_freq; 949 rps->last_adj = 0; 950 } 951 952 GT_TRACE(rps_to_gt(rps), "park:%x\n", rps->cur_freq); 953 } 954 955 u32 intel_rps_get_boost_frequency(struct intel_rps *rps) 956 { 957 struct intel_guc_slpc *slpc; 958 959 if (rps_uses_slpc(rps)) { 960 slpc = rps_to_slpc(rps); 961 962 return slpc->boost_freq; 963 } else { 964 return intel_gpu_freq(rps, rps->boost_freq); 965 } 966 } 967 968 static int rps_set_boost_freq(struct intel_rps *rps, u32 val) 969 { 970 bool boost = false; 971 972 /* Validate against (static) hardware limits */ 973 val = intel_freq_opcode(rps, val); 974 if (val < rps->min_freq || val > rps->max_freq) 975 return -EINVAL; 976 977 mutex_lock(&rps->lock); 978 if (val != rps->boost_freq) { 979 rps->boost_freq = val; 980 boost = atomic_read(&rps->num_waiters); 981 } 982 mutex_unlock(&rps->lock); 983 if (boost) 984 queue_work(rps_to_gt(rps)->i915->unordered_wq, &rps->work); 985 986 return 0; 987 } 988 989 int intel_rps_set_boost_frequency(struct intel_rps *rps, u32 freq) 990 { 991 struct intel_guc_slpc *slpc; 992 993 if (rps_uses_slpc(rps)) { 994 slpc = rps_to_slpc(rps); 995 996 return intel_guc_slpc_set_boost_freq(slpc, freq); 997 } else { 998 return rps_set_boost_freq(rps, freq); 999 } 1000 } 1001 1002 void intel_rps_dec_waiters(struct intel_rps *rps) 1003 { 1004 struct intel_guc_slpc *slpc; 1005 1006 if (rps_uses_slpc(rps)) { 1007 slpc = rps_to_slpc(rps); 1008 1009 /* Don't decrement num_waiters for req where increment was skipped */ 1010 if (slpc->power_profile == SLPC_POWER_PROFILES_POWER_SAVING) 1011 return; 1012 1013 intel_guc_slpc_dec_waiters(slpc); 1014 } else { 1015 atomic_dec(&rps->num_waiters); 1016 } 1017 } 1018 1019 void intel_rps_boost(struct i915_request *rq) 1020 { 1021 struct intel_guc_slpc *slpc; 1022 1023 if (i915_request_signaled(rq) || i915_request_has_waitboost(rq)) 1024 return; 1025 1026 /* Waitboost is not needed for contexts marked with a Freq hint */ 1027 if (test_bit(CONTEXT_LOW_LATENCY, &rq->context->flags)) 1028 return; 1029 1030 /* Serializes with i915_request_retire() */ 1031 if (!test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) { 1032 struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps; 1033 1034 if (rps_uses_slpc(rps)) { 1035 slpc = rps_to_slpc(rps); 1036 1037 /* Waitboost should not be done with power saving profile */ 1038 if (slpc->power_profile == SLPC_POWER_PROFILES_POWER_SAVING) 1039 return; 1040 1041 /* Return if old value is non zero */ 1042 if (!atomic_fetch_inc(&slpc->num_waiters)) { 1043 /* 1044 * Skip queuing boost work if frequency is already boosted, 1045 * but still increment num_waiters. 1046 */ 1047 if (slpc->min_freq_softlimit >= slpc->boost_freq) 1048 return; 1049 1050 GT_TRACE(rps_to_gt(rps), "boost fence:%llx:%llx\n", 1051 rq->fence.context, rq->fence.seqno); 1052 queue_work(rps_to_gt(rps)->i915->unordered_wq, 1053 &slpc->boost_work); 1054 } 1055 1056 return; 1057 } 1058 1059 if (atomic_fetch_inc(&rps->num_waiters)) 1060 return; 1061 1062 if (!intel_rps_is_active(rps)) 1063 return; 1064 1065 GT_TRACE(rps_to_gt(rps), "boost fence:%llx:%llx\n", 1066 rq->fence.context, rq->fence.seqno); 1067 1068 if (READ_ONCE(rps->cur_freq) < rps->boost_freq) 1069 queue_work(rps_to_gt(rps)->i915->unordered_wq, &rps->work); 1070 1071 WRITE_ONCE(rps->boosts, rps->boosts + 1); /* debug only */ 1072 } 1073 } 1074 1075 int intel_rps_set(struct intel_rps *rps, u8 val) 1076 { 1077 int err; 1078 1079 lockdep_assert_held(&rps->lock); 1080 GEM_BUG_ON(val > rps->max_freq); 1081 GEM_BUG_ON(val < rps->min_freq); 1082 1083 if (intel_rps_is_active(rps)) { 1084 err = rps_set(rps, val, true); 1085 if (err) 1086 return err; 1087 1088 /* 1089 * Make sure we continue to get interrupts 1090 * until we hit the minimum or maximum frequencies. 1091 */ 1092 if (intel_rps_has_interrupts(rps)) { 1093 struct intel_uncore *uncore = rps_to_uncore(rps); 1094 1095 set(uncore, 1096 GEN6_RP_INTERRUPT_LIMITS, rps_limits(rps, val)); 1097 1098 set(uncore, GEN6_PMINTRMSK, rps_pm_mask(rps, val)); 1099 } 1100 } 1101 1102 rps->cur_freq = val; 1103 return 0; 1104 } 1105 1106 static u32 intel_rps_read_state_cap(struct intel_rps *rps) 1107 { 1108 struct drm_i915_private *i915 = rps_to_i915(rps); 1109 struct intel_uncore *uncore = rps_to_uncore(rps); 1110 1111 if (IS_GEN9_LP(i915)) 1112 return intel_uncore_read(uncore, BXT_RP_STATE_CAP); 1113 else 1114 return intel_uncore_read(uncore, GEN6_RP_STATE_CAP); 1115 } 1116 1117 static void 1118 mtl_get_freq_caps(struct intel_rps *rps, struct intel_rps_freq_caps *caps) 1119 { 1120 struct intel_uncore *uncore = rps_to_uncore(rps); 1121 u32 rp_state_cap = rps_to_gt(rps)->type == GT_MEDIA ? 1122 intel_uncore_read(uncore, MTL_MEDIAP_STATE_CAP) : 1123 intel_uncore_read(uncore, MTL_RP_STATE_CAP); 1124 u32 rpe = rps_to_gt(rps)->type == GT_MEDIA ? 1125 intel_uncore_read(uncore, MTL_MPE_FREQUENCY) : 1126 intel_uncore_read(uncore, MTL_GT_RPE_FREQUENCY); 1127 1128 /* MTL values are in units of 16.67 MHz */ 1129 caps->rp0_freq = REG_FIELD_GET(MTL_RP0_CAP_MASK, rp_state_cap); 1130 caps->min_freq = REG_FIELD_GET(MTL_RPN_CAP_MASK, rp_state_cap); 1131 caps->rp1_freq = REG_FIELD_GET(MTL_RPE_MASK, rpe); 1132 } 1133 1134 static void 1135 __gen6_rps_get_freq_caps(struct intel_rps *rps, struct intel_rps_freq_caps *caps) 1136 { 1137 struct drm_i915_private *i915 = rps_to_i915(rps); 1138 u32 rp_state_cap; 1139 1140 rp_state_cap = intel_rps_read_state_cap(rps); 1141 1142 /* static values from HW: RP0 > RP1 > RPn (min_freq) */ 1143 if (IS_GEN9_LP(i915)) { 1144 caps->rp0_freq = (rp_state_cap >> 16) & 0xff; 1145 caps->rp1_freq = (rp_state_cap >> 8) & 0xff; 1146 caps->min_freq = (rp_state_cap >> 0) & 0xff; 1147 } else { 1148 caps->rp0_freq = (rp_state_cap >> 0) & 0xff; 1149 if (GRAPHICS_VER(i915) >= 10) 1150 caps->rp1_freq = REG_FIELD_GET(RPE_MASK, 1151 intel_uncore_read(to_gt(i915)->uncore, 1152 GEN10_FREQ_INFO_REC)); 1153 else 1154 caps->rp1_freq = (rp_state_cap >> 8) & 0xff; 1155 caps->min_freq = (rp_state_cap >> 16) & 0xff; 1156 } 1157 1158 if (IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) { 1159 /* 1160 * In this case rp_state_cap register reports frequencies in 1161 * units of 50 MHz. Convert these to the actual "hw unit", i.e. 1162 * units of 16.67 MHz 1163 */ 1164 caps->rp0_freq *= GEN9_FREQ_SCALER; 1165 caps->rp1_freq *= GEN9_FREQ_SCALER; 1166 caps->min_freq *= GEN9_FREQ_SCALER; 1167 } 1168 } 1169 1170 /** 1171 * gen6_rps_get_freq_caps - Get freq caps exposed by HW 1172 * @rps: the intel_rps structure 1173 * @caps: returned freq caps 1174 * 1175 * Returned "caps" frequencies should be converted to MHz using 1176 * intel_gpu_freq() 1177 */ 1178 void gen6_rps_get_freq_caps(struct intel_rps *rps, struct intel_rps_freq_caps *caps) 1179 { 1180 struct drm_i915_private *i915 = rps_to_i915(rps); 1181 1182 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70)) 1183 return mtl_get_freq_caps(rps, caps); 1184 else 1185 return __gen6_rps_get_freq_caps(rps, caps); 1186 } 1187 1188 static void gen6_rps_init(struct intel_rps *rps) 1189 { 1190 struct drm_i915_private *i915 = rps_to_i915(rps); 1191 struct intel_rps_freq_caps caps; 1192 1193 gen6_rps_get_freq_caps(rps, &caps); 1194 rps->rp0_freq = caps.rp0_freq; 1195 rps->rp1_freq = caps.rp1_freq; 1196 rps->min_freq = caps.min_freq; 1197 1198 /* hw_max = RP0 until we check for overclocking */ 1199 rps->max_freq = rps->rp0_freq; 1200 1201 rps->efficient_freq = rps->rp1_freq; 1202 if (IS_HASWELL(i915) || IS_BROADWELL(i915) || 1203 IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) { 1204 u32 ddcc_status = 0; 1205 u32 mult = 1; 1206 1207 if (IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) 1208 mult = GEN9_FREQ_SCALER; 1209 if (snb_pcode_read(rps_to_gt(rps)->uncore, 1210 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL, 1211 &ddcc_status, NULL) == 0) 1212 rps->efficient_freq = 1213 clamp_t(u32, 1214 ((ddcc_status >> 8) & 0xff) * mult, 1215 rps->min_freq, 1216 rps->max_freq); 1217 } 1218 } 1219 1220 static bool rps_reset(struct intel_rps *rps) 1221 { 1222 struct drm_i915_private *i915 = rps_to_i915(rps); 1223 1224 /* force a reset */ 1225 rps->power.mode = -1; 1226 rps->last_freq = -1; 1227 1228 if (rps_set(rps, rps->min_freq, true)) { 1229 drm_err(&i915->drm, "Failed to reset RPS to initial values\n"); 1230 return false; 1231 } 1232 1233 rps->cur_freq = rps->min_freq; 1234 return true; 1235 } 1236 1237 /* See the Gen9_GT_PM_Programming_Guide doc for the below */ 1238 static bool gen9_rps_enable(struct intel_rps *rps) 1239 { 1240 struct intel_gt *gt = rps_to_gt(rps); 1241 struct intel_uncore *uncore = gt->uncore; 1242 1243 /* Program defaults and thresholds for RPS */ 1244 if (GRAPHICS_VER(gt->i915) == 9) 1245 intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ, 1246 GEN9_FREQUENCY(rps->rp1_freq)); 1247 1248 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 0xa); 1249 1250 rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD; 1251 1252 return rps_reset(rps); 1253 } 1254 1255 static bool gen8_rps_enable(struct intel_rps *rps) 1256 { 1257 struct intel_uncore *uncore = rps_to_uncore(rps); 1258 1259 intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ, 1260 HSW_FREQUENCY(rps->rp1_freq)); 1261 1262 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1263 1264 rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD; 1265 1266 return rps_reset(rps); 1267 } 1268 1269 static bool gen6_rps_enable(struct intel_rps *rps) 1270 { 1271 struct intel_uncore *uncore = rps_to_uncore(rps); 1272 1273 /* Power down if completely idle for over 50ms */ 1274 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 50000); 1275 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1276 1277 rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD | 1278 GEN6_PM_RP_DOWN_THRESHOLD | 1279 GEN6_PM_RP_DOWN_TIMEOUT); 1280 1281 return rps_reset(rps); 1282 } 1283 1284 static int chv_rps_max_freq(struct intel_rps *rps) 1285 { 1286 struct drm_i915_private *i915 = rps_to_i915(rps); 1287 struct intel_gt *gt = rps_to_gt(rps); 1288 u32 val; 1289 1290 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, FB_GFX_FMAX_AT_VMAX_FUSE); 1291 1292 switch (gt->info.sseu.eu_total) { 1293 case 8: 1294 /* (2 * 4) config */ 1295 val >>= FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT; 1296 break; 1297 case 12: 1298 /* (2 * 6) config */ 1299 val >>= FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT; 1300 break; 1301 case 16: 1302 /* (2 * 8) config */ 1303 default: 1304 /* Setting (2 * 8) Min RP0 for any other combination */ 1305 val >>= FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT; 1306 break; 1307 } 1308 1309 return val & FB_GFX_FREQ_FUSE_MASK; 1310 } 1311 1312 static int chv_rps_rpe_freq(struct intel_rps *rps) 1313 { 1314 struct drm_i915_private *i915 = rps_to_i915(rps); 1315 u32 val; 1316 1317 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_GPU_DUTYCYCLE_REG); 1318 val >>= PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT; 1319 1320 return val & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK; 1321 } 1322 1323 static int chv_rps_guar_freq(struct intel_rps *rps) 1324 { 1325 struct drm_i915_private *i915 = rps_to_i915(rps); 1326 u32 val; 1327 1328 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, FB_GFX_FMAX_AT_VMAX_FUSE); 1329 1330 return val & FB_GFX_FREQ_FUSE_MASK; 1331 } 1332 1333 static u32 chv_rps_min_freq(struct intel_rps *rps) 1334 { 1335 struct drm_i915_private *i915 = rps_to_i915(rps); 1336 u32 val; 1337 1338 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, FB_GFX_FMIN_AT_VMIN_FUSE); 1339 val >>= FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT; 1340 1341 return val & FB_GFX_FREQ_FUSE_MASK; 1342 } 1343 1344 static bool chv_rps_enable(struct intel_rps *rps) 1345 { 1346 struct intel_uncore *uncore = rps_to_uncore(rps); 1347 struct drm_i915_private *i915 = rps_to_i915(rps); 1348 u32 val; 1349 1350 /* 1: Program defaults and thresholds for RPS*/ 1351 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000); 1352 intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400); 1353 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000); 1354 intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000); 1355 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000); 1356 1357 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1358 1359 /* 2: Enable RPS */ 1360 intel_uncore_write_fw(uncore, GEN6_RP_CONTROL, 1361 GEN6_RP_MEDIA_HW_NORMAL_MODE | 1362 GEN6_RP_MEDIA_IS_GFX | 1363 GEN6_RP_ENABLE | 1364 GEN6_RP_UP_BUSY_AVG | 1365 GEN6_RP_DOWN_IDLE_AVG); 1366 1367 rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD | 1368 GEN6_PM_RP_DOWN_THRESHOLD | 1369 GEN6_PM_RP_DOWN_TIMEOUT); 1370 1371 /* Setting Fixed Bias */ 1372 vlv_iosf_sb_get(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 1373 1374 val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | CHV_BIAS_CPU_50_SOC_50; 1375 vlv_iosf_sb_write(&i915->drm, VLV_IOSF_SB_PUNIT, VLV_TURBO_SOC_OVERRIDE, val); 1376 1377 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_REG_GPU_FREQ_STS); 1378 1379 vlv_iosf_sb_put(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 1380 1381 /* RPS code assumes GPLL is used */ 1382 drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0, 1383 "GPLL not enabled\n"); 1384 1385 drm_dbg(&i915->drm, "GPLL enabled? %s\n", 1386 str_yes_no(val & GPLLENABLE)); 1387 drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val); 1388 1389 return rps_reset(rps); 1390 } 1391 1392 static int vlv_rps_guar_freq(struct intel_rps *rps) 1393 { 1394 struct drm_i915_private *i915 = rps_to_i915(rps); 1395 u32 val, rp1; 1396 1397 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_NC, IOSF_NC_FB_GFX_FREQ_FUSE); 1398 1399 rp1 = val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK; 1400 rp1 >>= FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT; 1401 1402 return rp1; 1403 } 1404 1405 static int vlv_rps_max_freq(struct intel_rps *rps) 1406 { 1407 struct drm_i915_private *i915 = rps_to_i915(rps); 1408 u32 val, rp0; 1409 1410 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_NC, IOSF_NC_FB_GFX_FREQ_FUSE); 1411 1412 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT; 1413 /* Clamp to max */ 1414 rp0 = min_t(u32, rp0, 0xea); 1415 1416 return rp0; 1417 } 1418 1419 static int vlv_rps_rpe_freq(struct intel_rps *rps) 1420 { 1421 struct drm_i915_private *i915 = rps_to_i915(rps); 1422 u32 val, rpe; 1423 1424 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_NC, IOSF_NC_FB_GFX_FMAX_FUSE_LO); 1425 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT; 1426 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_NC, IOSF_NC_FB_GFX_FMAX_FUSE_HI); 1427 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5; 1428 1429 return rpe; 1430 } 1431 1432 static int vlv_rps_min_freq(struct intel_rps *rps) 1433 { 1434 struct drm_i915_private *i915 = rps_to_i915(rps); 1435 u32 val; 1436 1437 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_REG_GPU_LFM) & 0xff; 1438 /* 1439 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value 1440 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on 1441 * a BYT-M B0 the above register contains 0xbf. Moreover when setting 1442 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0 1443 * to make sure it matches what Punit accepts. 1444 */ 1445 return max_t(u32, val, 0xc0); 1446 } 1447 1448 static bool vlv_rps_enable(struct intel_rps *rps) 1449 { 1450 struct intel_uncore *uncore = rps_to_uncore(rps); 1451 struct drm_i915_private *i915 = rps_to_i915(rps); 1452 u32 val; 1453 1454 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000); 1455 intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400); 1456 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000); 1457 intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000); 1458 intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000); 1459 1460 intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10); 1461 1462 intel_uncore_write_fw(uncore, GEN6_RP_CONTROL, 1463 GEN6_RP_MEDIA_TURBO | 1464 GEN6_RP_MEDIA_HW_NORMAL_MODE | 1465 GEN6_RP_MEDIA_IS_GFX | 1466 GEN6_RP_ENABLE | 1467 GEN6_RP_UP_BUSY_AVG | 1468 GEN6_RP_DOWN_IDLE_CONT); 1469 1470 /* WaGsvRC0ResidencyMethod:vlv */ 1471 rps->pm_events = GEN6_PM_RP_UP_EI_EXPIRED; 1472 1473 vlv_iosf_sb_get(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 1474 1475 /* Setting Fixed Bias */ 1476 val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | VLV_BIAS_CPU_125_SOC_875; 1477 vlv_iosf_sb_write(&i915->drm, VLV_IOSF_SB_PUNIT, VLV_TURBO_SOC_OVERRIDE, val); 1478 1479 val = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_REG_GPU_FREQ_STS); 1480 1481 vlv_iosf_sb_put(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 1482 1483 /* RPS code assumes GPLL is used */ 1484 drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0, 1485 "GPLL not enabled\n"); 1486 1487 drm_dbg(&i915->drm, "GPLL enabled? %s\n", 1488 str_yes_no(val & GPLLENABLE)); 1489 drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val); 1490 1491 return rps_reset(rps); 1492 } 1493 1494 static unsigned long __ips_gfx_val(struct intel_ips *ips) 1495 { 1496 struct intel_rps *rps = container_of(ips, typeof(*rps), ips); 1497 struct intel_uncore *uncore = rps_to_uncore(rps); 1498 unsigned int t, state1, state2; 1499 u32 pxvid, ext_v; 1500 u64 corr, corr2; 1501 1502 lockdep_assert_held(&mchdev_lock); 1503 1504 pxvid = intel_uncore_read(uncore, PXVFREQ(rps->cur_freq)); 1505 pxvid = (pxvid >> 24) & 0x7f; 1506 ext_v = pvid_to_extvid(rps_to_i915(rps), pxvid); 1507 1508 state1 = ext_v; 1509 1510 /* Revel in the empirically derived constants */ 1511 1512 /* Correction factor in 1/100000 units */ 1513 t = ips_mch_val(uncore); 1514 if (t > 80) 1515 corr = t * 2349 + 135940; 1516 else if (t >= 50) 1517 corr = t * 964 + 29317; 1518 else /* < 50 */ 1519 corr = t * 301 + 1004; 1520 1521 corr = div_u64(corr * 150142 * state1, 10000) - 78642; 1522 corr2 = div_u64(corr, 100000) * ips->corr; 1523 1524 state2 = div_u64(corr2 * state1, 10000); 1525 state2 /= 100; /* convert to mW */ 1526 1527 __gen5_ips_update(ips); 1528 1529 return ips->gfx_power + state2; 1530 } 1531 1532 static bool has_busy_stats(struct intel_rps *rps) 1533 { 1534 struct intel_engine_cs *engine; 1535 enum intel_engine_id id; 1536 1537 for_each_engine(engine, rps_to_gt(rps), id) { 1538 if (!intel_engine_supports_stats(engine)) 1539 return false; 1540 } 1541 1542 return true; 1543 } 1544 1545 void intel_rps_enable(struct intel_rps *rps) 1546 { 1547 struct drm_i915_private *i915 = rps_to_i915(rps); 1548 struct intel_uncore *uncore = rps_to_uncore(rps); 1549 bool enabled = false; 1550 1551 if (!HAS_RPS(i915)) 1552 return; 1553 1554 if (rps_uses_slpc(rps)) 1555 return; 1556 1557 intel_gt_check_clock_frequency(rps_to_gt(rps)); 1558 1559 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 1560 if (rps->max_freq <= rps->min_freq) 1561 /* leave disabled, no room for dynamic reclocking */; 1562 else if (IS_CHERRYVIEW(i915)) 1563 enabled = chv_rps_enable(rps); 1564 else if (IS_VALLEYVIEW(i915)) 1565 enabled = vlv_rps_enable(rps); 1566 else if (GRAPHICS_VER(i915) >= 9) 1567 enabled = gen9_rps_enable(rps); 1568 else if (GRAPHICS_VER(i915) >= 8) 1569 enabled = gen8_rps_enable(rps); 1570 else if (GRAPHICS_VER(i915) >= 6) 1571 enabled = gen6_rps_enable(rps); 1572 else if (IS_IRONLAKE_M(i915)) 1573 enabled = gen5_rps_enable(rps); 1574 else 1575 MISSING_CASE(GRAPHICS_VER(i915)); 1576 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 1577 if (!enabled) 1578 return; 1579 1580 GT_TRACE(rps_to_gt(rps), 1581 "min:%x, max:%x, freq:[%d, %d], thresholds:[%u, %u]\n", 1582 rps->min_freq, rps->max_freq, 1583 intel_gpu_freq(rps, rps->min_freq), 1584 intel_gpu_freq(rps, rps->max_freq), 1585 rps->power.up_threshold, 1586 rps->power.down_threshold); 1587 1588 GEM_BUG_ON(rps->max_freq < rps->min_freq); 1589 GEM_BUG_ON(rps->idle_freq > rps->max_freq); 1590 1591 GEM_BUG_ON(rps->efficient_freq < rps->min_freq); 1592 GEM_BUG_ON(rps->efficient_freq > rps->max_freq); 1593 1594 if (has_busy_stats(rps)) 1595 intel_rps_set_timer(rps); 1596 else if (GRAPHICS_VER(i915) >= 6 && GRAPHICS_VER(i915) <= 11) 1597 intel_rps_set_interrupts(rps); 1598 else 1599 /* Ironlake currently uses intel_ips.ko */ {} 1600 1601 intel_rps_set_enabled(rps); 1602 } 1603 1604 static void gen6_rps_disable(struct intel_rps *rps) 1605 { 1606 set(rps_to_uncore(rps), GEN6_RP_CONTROL, 0); 1607 } 1608 1609 void intel_rps_disable(struct intel_rps *rps) 1610 { 1611 struct drm_i915_private *i915 = rps_to_i915(rps); 1612 1613 if (!intel_rps_is_enabled(rps)) 1614 return; 1615 1616 intel_rps_clear_enabled(rps); 1617 intel_rps_clear_interrupts(rps); 1618 intel_rps_clear_timer(rps); 1619 1620 if (GRAPHICS_VER(i915) >= 6) 1621 gen6_rps_disable(rps); 1622 else if (IS_IRONLAKE_M(i915)) 1623 gen5_rps_disable(rps); 1624 } 1625 1626 static int byt_gpu_freq(struct intel_rps *rps, int val) 1627 { 1628 /* 1629 * N = val - 0xb7 1630 * Slow = Fast = GPLL ref * N 1631 */ 1632 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000); 1633 } 1634 1635 static int byt_freq_opcode(struct intel_rps *rps, int val) 1636 { 1637 return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7; 1638 } 1639 1640 static int chv_gpu_freq(struct intel_rps *rps, int val) 1641 { 1642 /* 1643 * N = val / 2 1644 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2 1645 */ 1646 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000); 1647 } 1648 1649 static int chv_freq_opcode(struct intel_rps *rps, int val) 1650 { 1651 /* CHV needs even values */ 1652 return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2; 1653 } 1654 1655 int intel_gpu_freq(struct intel_rps *rps, int val) 1656 { 1657 struct drm_i915_private *i915 = rps_to_i915(rps); 1658 1659 if (GRAPHICS_VER(i915) >= 9) 1660 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER, 1661 GEN9_FREQ_SCALER); 1662 else if (IS_CHERRYVIEW(i915)) 1663 return chv_gpu_freq(rps, val); 1664 else if (IS_VALLEYVIEW(i915)) 1665 return byt_gpu_freq(rps, val); 1666 else if (GRAPHICS_VER(i915) >= 6) 1667 return val * GT_FREQUENCY_MULTIPLIER; 1668 else 1669 return val; 1670 } 1671 1672 int intel_freq_opcode(struct intel_rps *rps, int val) 1673 { 1674 struct drm_i915_private *i915 = rps_to_i915(rps); 1675 1676 if (GRAPHICS_VER(i915) >= 9) 1677 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER, 1678 GT_FREQUENCY_MULTIPLIER); 1679 else if (IS_CHERRYVIEW(i915)) 1680 return chv_freq_opcode(rps, val); 1681 else if (IS_VALLEYVIEW(i915)) 1682 return byt_freq_opcode(rps, val); 1683 else if (GRAPHICS_VER(i915) >= 6) 1684 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER); 1685 else 1686 return val; 1687 } 1688 1689 static void vlv_init_gpll_ref_freq(struct intel_rps *rps) 1690 { 1691 struct drm_i915_private *i915 = rps_to_i915(rps); 1692 1693 rps->gpll_ref_freq = 1694 vlv_get_cck_clock(&i915->drm, "GPLL ref", 1695 CCK_GPLL_CLOCK_CONTROL, 1696 i915->czclk_freq); 1697 1698 drm_dbg(&i915->drm, "GPLL reference freq: %d kHz\n", 1699 rps->gpll_ref_freq); 1700 } 1701 1702 static void vlv_rps_init(struct intel_rps *rps) 1703 { 1704 struct drm_i915_private *i915 = rps_to_i915(rps); 1705 1706 vlv_iosf_sb_get(&i915->drm, 1707 BIT(VLV_IOSF_SB_PUNIT) | 1708 BIT(VLV_IOSF_SB_NC) | 1709 BIT(VLV_IOSF_SB_CCK)); 1710 1711 vlv_init_gpll_ref_freq(rps); 1712 1713 rps->max_freq = vlv_rps_max_freq(rps); 1714 rps->rp0_freq = rps->max_freq; 1715 drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n", 1716 intel_gpu_freq(rps, rps->max_freq), rps->max_freq); 1717 1718 rps->efficient_freq = vlv_rps_rpe_freq(rps); 1719 drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n", 1720 intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq); 1721 1722 rps->rp1_freq = vlv_rps_guar_freq(rps); 1723 drm_dbg(&i915->drm, "RP1(Guar Freq) GPU freq: %d MHz (%u)\n", 1724 intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq); 1725 1726 rps->min_freq = vlv_rps_min_freq(rps); 1727 drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n", 1728 intel_gpu_freq(rps, rps->min_freq), rps->min_freq); 1729 1730 vlv_iosf_sb_put(&i915->drm, 1731 BIT(VLV_IOSF_SB_PUNIT) | 1732 BIT(VLV_IOSF_SB_NC) | 1733 BIT(VLV_IOSF_SB_CCK)); 1734 } 1735 1736 static void chv_rps_init(struct intel_rps *rps) 1737 { 1738 struct drm_i915_private *i915 = rps_to_i915(rps); 1739 1740 vlv_iosf_sb_get(&i915->drm, 1741 BIT(VLV_IOSF_SB_PUNIT) | 1742 BIT(VLV_IOSF_SB_NC) | 1743 BIT(VLV_IOSF_SB_CCK)); 1744 1745 vlv_init_gpll_ref_freq(rps); 1746 1747 rps->max_freq = chv_rps_max_freq(rps); 1748 rps->rp0_freq = rps->max_freq; 1749 drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n", 1750 intel_gpu_freq(rps, rps->max_freq), rps->max_freq); 1751 1752 rps->efficient_freq = chv_rps_rpe_freq(rps); 1753 drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n", 1754 intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq); 1755 1756 rps->rp1_freq = chv_rps_guar_freq(rps); 1757 drm_dbg(&i915->drm, "RP1(Guar) GPU freq: %d MHz (%u)\n", 1758 intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq); 1759 1760 rps->min_freq = chv_rps_min_freq(rps); 1761 drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n", 1762 intel_gpu_freq(rps, rps->min_freq), rps->min_freq); 1763 1764 vlv_iosf_sb_put(&i915->drm, 1765 BIT(VLV_IOSF_SB_PUNIT) | 1766 BIT(VLV_IOSF_SB_NC) | 1767 BIT(VLV_IOSF_SB_CCK)); 1768 1769 drm_WARN_ONCE(&i915->drm, (rps->max_freq | rps->efficient_freq | 1770 rps->rp1_freq | rps->min_freq) & 1, 1771 "Odd GPU freq values\n"); 1772 } 1773 1774 static void vlv_c0_read(struct intel_uncore *uncore, struct intel_rps_ei *ei) 1775 { 1776 ei->ktime = ktime_get_raw(); 1777 ei->render_c0 = intel_uncore_read(uncore, VLV_RENDER_C0_COUNT); 1778 ei->media_c0 = intel_uncore_read(uncore, VLV_MEDIA_C0_COUNT); 1779 } 1780 1781 static u32 vlv_wa_c0_ei(struct intel_rps *rps, u32 pm_iir) 1782 { 1783 struct intel_uncore *uncore = rps_to_uncore(rps); 1784 const struct intel_rps_ei *prev = &rps->ei; 1785 struct intel_rps_ei now; 1786 u32 events = 0; 1787 1788 if ((pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) == 0) 1789 return 0; 1790 1791 vlv_c0_read(uncore, &now); 1792 1793 if (prev->ktime) { 1794 u64 time, c0; 1795 u32 render, media; 1796 1797 time = ktime_us_delta(now.ktime, prev->ktime); 1798 1799 time *= rps_to_i915(rps)->czclk_freq; 1800 1801 /* Workload can be split between render + media, 1802 * e.g. SwapBuffers being blitted in X after being rendered in 1803 * mesa. To account for this we need to combine both engines 1804 * into our activity counter. 1805 */ 1806 render = now.render_c0 - prev->render_c0; 1807 media = now.media_c0 - prev->media_c0; 1808 c0 = max(render, media); 1809 c0 *= 1000 * 100 << 8; /* to usecs and scale to threshold% */ 1810 1811 if (c0 > time * rps->power.up_threshold) 1812 events = GEN6_PM_RP_UP_THRESHOLD; 1813 else if (c0 < time * rps->power.down_threshold) 1814 events = GEN6_PM_RP_DOWN_THRESHOLD; 1815 } 1816 1817 rps->ei = now; 1818 return events; 1819 } 1820 1821 static void rps_work(struct work_struct *work) 1822 { 1823 struct intel_rps *rps = container_of(work, typeof(*rps), work); 1824 struct intel_gt *gt = rps_to_gt(rps); 1825 struct drm_i915_private *i915 = rps_to_i915(rps); 1826 bool client_boost = false; 1827 int new_freq, adj, min, max; 1828 u32 pm_iir = 0; 1829 1830 spin_lock_irq(gt->irq_lock); 1831 pm_iir = fetch_and_zero(&rps->pm_iir) & rps->pm_events; 1832 client_boost = atomic_read(&rps->num_waiters); 1833 spin_unlock_irq(gt->irq_lock); 1834 1835 /* Make sure we didn't queue anything we're not going to process. */ 1836 if (!pm_iir && !client_boost) 1837 goto out; 1838 1839 mutex_lock(&rps->lock); 1840 if (!intel_rps_is_active(rps)) { 1841 mutex_unlock(&rps->lock); 1842 return; 1843 } 1844 1845 pm_iir |= vlv_wa_c0_ei(rps, pm_iir); 1846 1847 adj = rps->last_adj; 1848 new_freq = rps->cur_freq; 1849 min = rps->min_freq_softlimit; 1850 max = rps->max_freq_softlimit; 1851 if (client_boost) 1852 max = rps->max_freq; 1853 1854 GT_TRACE(gt, 1855 "pm_iir:%x, client_boost:%s, last:%d, cur:%x, min:%x, max:%x\n", 1856 pm_iir, str_yes_no(client_boost), 1857 adj, new_freq, min, max); 1858 1859 if (client_boost && new_freq < rps->boost_freq) { 1860 new_freq = rps->boost_freq; 1861 adj = 0; 1862 } else if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) { 1863 if (adj > 0) 1864 adj *= 2; 1865 else /* CHV needs even encode values */ 1866 adj = IS_CHERRYVIEW(gt->i915) ? 2 : 1; 1867 1868 if (new_freq >= rps->max_freq_softlimit) 1869 adj = 0; 1870 } else if (client_boost) { 1871 adj = 0; 1872 } else if (pm_iir & GEN6_PM_RP_DOWN_TIMEOUT) { 1873 if (rps->cur_freq > rps->efficient_freq) 1874 new_freq = rps->efficient_freq; 1875 else if (rps->cur_freq > rps->min_freq_softlimit) 1876 new_freq = rps->min_freq_softlimit; 1877 adj = 0; 1878 } else if (pm_iir & GEN6_PM_RP_DOWN_THRESHOLD) { 1879 if (adj < 0) 1880 adj *= 2; 1881 else /* CHV needs even encode values */ 1882 adj = IS_CHERRYVIEW(gt->i915) ? -2 : -1; 1883 1884 if (new_freq <= rps->min_freq_softlimit) 1885 adj = 0; 1886 } else { /* unknown event */ 1887 adj = 0; 1888 } 1889 1890 /* 1891 * sysfs frequency limits may have snuck in while 1892 * servicing the interrupt 1893 */ 1894 new_freq += adj; 1895 new_freq = clamp_t(int, new_freq, min, max); 1896 1897 if (intel_rps_set(rps, new_freq)) { 1898 drm_dbg(&i915->drm, "Failed to set new GPU frequency\n"); 1899 adj = 0; 1900 } 1901 rps->last_adj = adj; 1902 1903 mutex_unlock(&rps->lock); 1904 1905 out: 1906 spin_lock_irq(gt->irq_lock); 1907 gen6_gt_pm_unmask_irq(gt, rps->pm_events); 1908 spin_unlock_irq(gt->irq_lock); 1909 } 1910 1911 void gen11_rps_irq_handler(struct intel_rps *rps, u32 pm_iir) 1912 { 1913 struct intel_gt *gt = rps_to_gt(rps); 1914 const u32 events = rps->pm_events & pm_iir; 1915 1916 lockdep_assert_held(gt->irq_lock); 1917 1918 if (unlikely(!events)) 1919 return; 1920 1921 GT_TRACE(gt, "irq events:%x\n", events); 1922 1923 gen6_gt_pm_mask_irq(gt, events); 1924 1925 rps->pm_iir |= events; 1926 queue_work(gt->i915->unordered_wq, &rps->work); 1927 } 1928 1929 void gen6_rps_irq_handler(struct intel_rps *rps, u32 pm_iir) 1930 { 1931 struct intel_gt *gt = rps_to_gt(rps); 1932 u32 events; 1933 1934 events = pm_iir & rps->pm_events; 1935 if (events) { 1936 spin_lock(gt->irq_lock); 1937 1938 GT_TRACE(gt, "irq events:%x\n", events); 1939 1940 gen6_gt_pm_mask_irq(gt, events); 1941 rps->pm_iir |= events; 1942 1943 queue_work(gt->i915->unordered_wq, &rps->work); 1944 spin_unlock(gt->irq_lock); 1945 } 1946 1947 if (GRAPHICS_VER(gt->i915) >= 8) 1948 return; 1949 1950 if (pm_iir & PM_VEBOX_USER_INTERRUPT) 1951 intel_engine_cs_irq(gt->engine[VECS0], pm_iir >> 10); 1952 1953 if (pm_iir & PM_VEBOX_CS_ERROR_INTERRUPT) 1954 drm_dbg(&rps_to_i915(rps)->drm, 1955 "Command parser error, pm_iir 0x%08x\n", pm_iir); 1956 } 1957 1958 void gen5_rps_irq_handler(struct intel_rps *rps) 1959 { 1960 struct intel_uncore *uncore = rps_to_uncore(rps); 1961 u32 busy_up, busy_down, max_avg, min_avg; 1962 u8 new_freq; 1963 1964 spin_lock(&mchdev_lock); 1965 1966 intel_uncore_write16(uncore, 1967 MEMINTRSTS, 1968 intel_uncore_read(uncore, MEMINTRSTS)); 1969 1970 intel_uncore_write16(uncore, MEMINTRSTS, MEMINT_EVAL_CHG); 1971 busy_up = intel_uncore_read(uncore, RCPREVBSYTUPAVG); 1972 busy_down = intel_uncore_read(uncore, RCPREVBSYTDNAVG); 1973 max_avg = intel_uncore_read(uncore, RCBMAXAVG); 1974 min_avg = intel_uncore_read(uncore, RCBMINAVG); 1975 1976 /* Handle RCS change request from hw */ 1977 new_freq = rps->cur_freq; 1978 if (busy_up > max_avg) 1979 new_freq++; 1980 else if (busy_down < min_avg) 1981 new_freq--; 1982 new_freq = clamp(new_freq, 1983 rps->min_freq_softlimit, 1984 rps->max_freq_softlimit); 1985 1986 if (new_freq != rps->cur_freq && !__gen5_rps_set(rps, new_freq)) 1987 rps->cur_freq = new_freq; 1988 1989 spin_unlock(&mchdev_lock); 1990 } 1991 1992 void intel_rps_init_early(struct intel_rps *rps) 1993 { 1994 mutex_init(&rps->lock); 1995 mutex_init(&rps->power.mutex); 1996 1997 INIT_WORK(&rps->work, rps_work); 1998 timer_setup(&rps->timer, rps_timer, 0); 1999 2000 atomic_set(&rps->num_waiters, 0); 2001 } 2002 2003 void intel_rps_init(struct intel_rps *rps) 2004 { 2005 struct drm_i915_private *i915 = rps_to_i915(rps); 2006 2007 if (rps_uses_slpc(rps)) 2008 return; 2009 2010 if (IS_CHERRYVIEW(i915)) 2011 chv_rps_init(rps); 2012 else if (IS_VALLEYVIEW(i915)) 2013 vlv_rps_init(rps); 2014 else if (GRAPHICS_VER(i915) >= 6) 2015 gen6_rps_init(rps); 2016 else if (IS_IRONLAKE_M(i915)) 2017 gen5_rps_init(rps); 2018 2019 /* Derive initial user preferences/limits from the hardware limits */ 2020 rps->max_freq_softlimit = rps->max_freq; 2021 rps_to_gt(rps)->defaults.max_freq = rps->max_freq_softlimit; 2022 rps->min_freq_softlimit = rps->min_freq; 2023 rps_to_gt(rps)->defaults.min_freq = rps->min_freq_softlimit; 2024 2025 /* After setting max-softlimit, find the overclock max freq */ 2026 if (GRAPHICS_VER(i915) == 6 || IS_IVYBRIDGE(i915) || IS_HASWELL(i915)) { 2027 u32 params = 0; 2028 2029 snb_pcode_read(rps_to_gt(rps)->uncore, GEN6_READ_OC_PARAMS, ¶ms, NULL); 2030 if (params & BIT(31)) { /* OC supported */ 2031 drm_dbg(&i915->drm, 2032 "Overclocking supported, max: %dMHz, overclock: %dMHz\n", 2033 (rps->max_freq & 0xff) * 50, 2034 (params & 0xff) * 50); 2035 rps->max_freq = params & 0xff; 2036 } 2037 } 2038 2039 /* Set default thresholds in % */ 2040 rps->power.up_threshold = 95; 2041 rps_to_gt(rps)->defaults.rps_up_threshold = rps->power.up_threshold; 2042 rps->power.down_threshold = 85; 2043 rps_to_gt(rps)->defaults.rps_down_threshold = rps->power.down_threshold; 2044 2045 /* Finally allow us to boost to max by default */ 2046 rps->boost_freq = rps->max_freq; 2047 rps->idle_freq = rps->min_freq; 2048 2049 /* Start in the middle, from here we will autotune based on workload */ 2050 rps->cur_freq = rps->efficient_freq; 2051 2052 rps->pm_intrmsk_mbz = 0; 2053 2054 /* 2055 * SNB,IVB,HSW can while VLV,CHV may hard hang on looping batchbuffer 2056 * if GEN6_PM_UP_EI_EXPIRED is masked. 2057 * 2058 * TODO: verify if this can be reproduced on VLV,CHV. 2059 */ 2060 if (GRAPHICS_VER(i915) <= 7) 2061 rps->pm_intrmsk_mbz |= GEN6_PM_RP_UP_EI_EXPIRED; 2062 2063 if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) < 11) 2064 rps->pm_intrmsk_mbz |= GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC; 2065 2066 /* GuC needs ARAT expired interrupt unmasked */ 2067 if (intel_uc_uses_guc_submission(&rps_to_gt(rps)->uc)) 2068 rps->pm_intrmsk_mbz |= ARAT_EXPIRED_INTRMSK; 2069 } 2070 2071 void intel_rps_sanitize(struct intel_rps *rps) 2072 { 2073 if (rps_uses_slpc(rps)) 2074 return; 2075 2076 if (GRAPHICS_VER(rps_to_i915(rps)) >= 6) 2077 rps_disable_interrupts(rps); 2078 } 2079 2080 u32 intel_rps_read_rpstat(struct intel_rps *rps) 2081 { 2082 struct drm_i915_private *i915 = rps_to_i915(rps); 2083 i915_reg_t rpstat; 2084 2085 rpstat = (GRAPHICS_VER(i915) >= 12) ? GEN12_RPSTAT1 : GEN6_RPSTAT1; 2086 2087 return intel_uncore_read(rps_to_gt(rps)->uncore, rpstat); 2088 } 2089 2090 static u32 intel_rps_get_cagf(struct intel_rps *rps, u32 rpstat) 2091 { 2092 struct drm_i915_private *i915 = rps_to_i915(rps); 2093 u32 cagf; 2094 2095 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70)) 2096 cagf = REG_FIELD_GET(MTL_CAGF_MASK, rpstat); 2097 else if (GRAPHICS_VER(i915) >= 12) 2098 cagf = REG_FIELD_GET(GEN12_CAGF_MASK, rpstat); 2099 else if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) 2100 cagf = REG_FIELD_GET(RPE_MASK, rpstat); 2101 else if (GRAPHICS_VER(i915) >= 9) 2102 cagf = REG_FIELD_GET(GEN9_CAGF_MASK, rpstat); 2103 else if (IS_HASWELL(i915) || IS_BROADWELL(i915)) 2104 cagf = REG_FIELD_GET(HSW_CAGF_MASK, rpstat); 2105 else if (GRAPHICS_VER(i915) >= 6) 2106 cagf = REG_FIELD_GET(GEN6_CAGF_MASK, rpstat); 2107 else 2108 cagf = gen5_invert_freq(rps, REG_FIELD_GET(MEMSTAT_PSTATE_MASK, rpstat)); 2109 2110 return cagf; 2111 } 2112 2113 static u32 __read_cagf(struct intel_rps *rps, bool take_fw) 2114 { 2115 struct drm_i915_private *i915 = rps_to_i915(rps); 2116 struct intel_uncore *uncore = rps_to_uncore(rps); 2117 i915_reg_t r = INVALID_MMIO_REG; 2118 u32 freq; 2119 2120 /* 2121 * For Gen12+ reading freq from HW does not need a forcewake and 2122 * registers will return 0 freq when GT is in RC6 2123 */ 2124 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 70)) { 2125 r = MTL_MIRROR_TARGET_WP1; 2126 } else if (GRAPHICS_VER(i915) >= 12) { 2127 r = GEN12_RPSTAT1; 2128 } else if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) { 2129 vlv_iosf_sb_get(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 2130 freq = vlv_iosf_sb_read(&i915->drm, VLV_IOSF_SB_PUNIT, PUNIT_REG_GPU_FREQ_STS); 2131 vlv_iosf_sb_put(&i915->drm, BIT(VLV_IOSF_SB_PUNIT)); 2132 } else if (GRAPHICS_VER(i915) >= 6) { 2133 r = GEN6_RPSTAT1; 2134 } else { 2135 r = MEMSTAT_ILK; 2136 } 2137 2138 if (i915_mmio_reg_valid(r)) 2139 freq = take_fw ? intel_uncore_read(uncore, r) : intel_uncore_read_fw(uncore, r); 2140 2141 return intel_rps_get_cagf(rps, freq); 2142 } 2143 2144 static u32 read_cagf(struct intel_rps *rps) 2145 { 2146 return __read_cagf(rps, true); 2147 } 2148 2149 u32 intel_rps_read_actual_frequency(struct intel_rps *rps) 2150 { 2151 struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm; 2152 intel_wakeref_t wakeref; 2153 u32 freq = 0; 2154 2155 with_intel_runtime_pm_if_in_use(rpm, wakeref) 2156 freq = intel_gpu_freq(rps, read_cagf(rps)); 2157 2158 return freq; 2159 } 2160 2161 u32 intel_rps_read_actual_frequency_fw(struct intel_rps *rps) 2162 { 2163 return intel_gpu_freq(rps, __read_cagf(rps, false)); 2164 } 2165 2166 static u32 intel_rps_read_punit_req(struct intel_rps *rps) 2167 { 2168 struct intel_uncore *uncore = rps_to_uncore(rps); 2169 struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm; 2170 intel_wakeref_t wakeref; 2171 u32 freq = 0; 2172 2173 with_intel_runtime_pm_if_in_use(rpm, wakeref) 2174 freq = intel_uncore_read(uncore, GEN6_RPNSWREQ); 2175 2176 return freq; 2177 } 2178 2179 static u32 intel_rps_get_req(u32 pureq) 2180 { 2181 u32 req = pureq >> GEN9_SW_REQ_UNSLICE_RATIO_SHIFT; 2182 2183 return req; 2184 } 2185 2186 u32 intel_rps_read_punit_req_frequency(struct intel_rps *rps) 2187 { 2188 u32 freq = intel_rps_get_req(intel_rps_read_punit_req(rps)); 2189 2190 return intel_gpu_freq(rps, freq); 2191 } 2192 2193 u32 intel_rps_get_requested_frequency(struct intel_rps *rps) 2194 { 2195 if (rps_uses_slpc(rps)) 2196 return intel_rps_read_punit_req_frequency(rps); 2197 else 2198 return intel_gpu_freq(rps, rps->cur_freq); 2199 } 2200 2201 u32 intel_rps_get_max_frequency(struct intel_rps *rps) 2202 { 2203 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2204 2205 if (rps_uses_slpc(rps)) 2206 return slpc->max_freq_softlimit; 2207 else 2208 return intel_gpu_freq(rps, rps->max_freq_softlimit); 2209 } 2210 2211 /** 2212 * intel_rps_get_max_raw_freq - returns the max frequency in some raw format. 2213 * @rps: the intel_rps structure 2214 * 2215 * Returns the max frequency in a raw format. In newer platforms raw is in 2216 * units of 50 MHz. 2217 */ 2218 u32 intel_rps_get_max_raw_freq(struct intel_rps *rps) 2219 { 2220 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2221 u32 freq; 2222 2223 if (rps_uses_slpc(rps)) { 2224 return DIV_ROUND_CLOSEST(slpc->rp0_freq, 2225 GT_FREQUENCY_MULTIPLIER); 2226 } else { 2227 freq = rps->max_freq; 2228 if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) { 2229 /* Convert GT frequency to 50 MHz units */ 2230 freq /= GEN9_FREQ_SCALER; 2231 } 2232 return freq; 2233 } 2234 } 2235 2236 u32 intel_rps_get_rp0_frequency(struct intel_rps *rps) 2237 { 2238 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2239 2240 if (rps_uses_slpc(rps)) 2241 return slpc->rp0_freq; 2242 else 2243 return intel_gpu_freq(rps, rps->rp0_freq); 2244 } 2245 2246 u32 intel_rps_get_rp1_frequency(struct intel_rps *rps) 2247 { 2248 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2249 2250 if (rps_uses_slpc(rps)) 2251 return slpc->rp1_freq; 2252 else 2253 return intel_gpu_freq(rps, rps->rp1_freq); 2254 } 2255 2256 u32 intel_rps_get_rpn_frequency(struct intel_rps *rps) 2257 { 2258 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2259 2260 if (rps_uses_slpc(rps)) 2261 return slpc->min_freq; 2262 else 2263 return intel_gpu_freq(rps, rps->min_freq); 2264 } 2265 2266 static void rps_frequency_dump(struct intel_rps *rps, struct drm_printer *p) 2267 { 2268 struct intel_gt *gt = rps_to_gt(rps); 2269 struct drm_i915_private *i915 = gt->i915; 2270 struct intel_uncore *uncore = gt->uncore; 2271 struct intel_rps_freq_caps caps; 2272 u32 rp_state_limits; 2273 u32 gt_perf_status; 2274 u32 rpmodectl, rpinclimit, rpdeclimit; 2275 u32 rpstat, cagf, reqf; 2276 u32 rpcurupei, rpcurup, rpprevup; 2277 u32 rpcurdownei, rpcurdown, rpprevdown; 2278 u32 rpupei, rpupt, rpdownei, rpdownt; 2279 u32 pm_ier, pm_imr, pm_isr, pm_iir, pm_mask; 2280 2281 rp_state_limits = intel_uncore_read(uncore, GEN6_RP_STATE_LIMITS); 2282 gen6_rps_get_freq_caps(rps, &caps); 2283 if (IS_GEN9_LP(i915)) 2284 gt_perf_status = intel_uncore_read(uncore, BXT_GT_PERF_STATUS); 2285 else 2286 gt_perf_status = intel_uncore_read(uncore, GEN6_GT_PERF_STATUS); 2287 2288 /* RPSTAT1 is in the GT power well */ 2289 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 2290 2291 reqf = intel_uncore_read(uncore, GEN6_RPNSWREQ); 2292 if (GRAPHICS_VER(i915) >= 9) { 2293 reqf >>= 23; 2294 } else { 2295 reqf &= ~GEN6_TURBO_DISABLE; 2296 if (IS_HASWELL(i915) || IS_BROADWELL(i915)) 2297 reqf >>= 24; 2298 else 2299 reqf >>= 25; 2300 } 2301 reqf = intel_gpu_freq(rps, reqf); 2302 2303 rpmodectl = intel_uncore_read(uncore, GEN6_RP_CONTROL); 2304 rpinclimit = intel_uncore_read(uncore, GEN6_RP_UP_THRESHOLD); 2305 rpdeclimit = intel_uncore_read(uncore, GEN6_RP_DOWN_THRESHOLD); 2306 2307 rpstat = intel_rps_read_rpstat(rps); 2308 rpcurupei = intel_uncore_read(uncore, GEN6_RP_CUR_UP_EI) & GEN6_CURICONT_MASK; 2309 rpcurup = intel_uncore_read(uncore, GEN6_RP_CUR_UP) & GEN6_CURBSYTAVG_MASK; 2310 rpprevup = intel_uncore_read(uncore, GEN6_RP_PREV_UP) & GEN6_CURBSYTAVG_MASK; 2311 rpcurdownei = intel_uncore_read(uncore, GEN6_RP_CUR_DOWN_EI) & GEN6_CURIAVG_MASK; 2312 rpcurdown = intel_uncore_read(uncore, GEN6_RP_CUR_DOWN) & GEN6_CURBSYTAVG_MASK; 2313 rpprevdown = intel_uncore_read(uncore, GEN6_RP_PREV_DOWN) & GEN6_CURBSYTAVG_MASK; 2314 2315 rpupei = intel_uncore_read(uncore, GEN6_RP_UP_EI); 2316 rpupt = intel_uncore_read(uncore, GEN6_RP_UP_THRESHOLD); 2317 2318 rpdownei = intel_uncore_read(uncore, GEN6_RP_DOWN_EI); 2319 rpdownt = intel_uncore_read(uncore, GEN6_RP_DOWN_THRESHOLD); 2320 2321 cagf = intel_rps_read_actual_frequency(rps); 2322 2323 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 2324 2325 if (GRAPHICS_VER(i915) >= 11) { 2326 pm_ier = intel_uncore_read(uncore, GEN11_GPM_WGBOXPERF_INTR_ENABLE); 2327 pm_imr = intel_uncore_read(uncore, GEN11_GPM_WGBOXPERF_INTR_MASK); 2328 /* 2329 * The equivalent to the PM ISR & IIR cannot be read 2330 * without affecting the current state of the system 2331 */ 2332 pm_isr = 0; 2333 pm_iir = 0; 2334 } else if (GRAPHICS_VER(i915) >= 8) { 2335 pm_ier = intel_uncore_read(uncore, GEN8_GT_IER(2)); 2336 pm_imr = intel_uncore_read(uncore, GEN8_GT_IMR(2)); 2337 pm_isr = intel_uncore_read(uncore, GEN8_GT_ISR(2)); 2338 pm_iir = intel_uncore_read(uncore, GEN8_GT_IIR(2)); 2339 } else { 2340 pm_ier = intel_uncore_read(uncore, GEN6_PMIER); 2341 pm_imr = intel_uncore_read(uncore, GEN6_PMIMR); 2342 pm_isr = intel_uncore_read(uncore, GEN6_PMISR); 2343 pm_iir = intel_uncore_read(uncore, GEN6_PMIIR); 2344 } 2345 pm_mask = intel_uncore_read(uncore, GEN6_PMINTRMSK); 2346 2347 drm_printf(p, "Video Turbo Mode: %s\n", 2348 str_yes_no(rpmodectl & GEN6_RP_MEDIA_TURBO)); 2349 drm_printf(p, "HW control enabled: %s\n", 2350 str_yes_no(rpmodectl & GEN6_RP_ENABLE)); 2351 drm_printf(p, "SW control enabled: %s\n", 2352 str_yes_no((rpmodectl & GEN6_RP_MEDIA_MODE_MASK) == GEN6_RP_MEDIA_SW_MODE)); 2353 2354 drm_printf(p, "PM IER=0x%08x IMR=0x%08x, MASK=0x%08x\n", 2355 pm_ier, pm_imr, pm_mask); 2356 if (GRAPHICS_VER(i915) <= 10) 2357 drm_printf(p, "PM ISR=0x%08x IIR=0x%08x\n", 2358 pm_isr, pm_iir); 2359 drm_printf(p, "pm_intrmsk_mbz: 0x%08x\n", 2360 rps->pm_intrmsk_mbz); 2361 drm_printf(p, "GT_PERF_STATUS: 0x%08x\n", gt_perf_status); 2362 drm_printf(p, "Render p-state ratio: %d\n", 2363 (gt_perf_status & (GRAPHICS_VER(i915) >= 9 ? 0x1ff00 : 0xff00)) >> 8); 2364 drm_printf(p, "Render p-state VID: %d\n", 2365 gt_perf_status & 0xff); 2366 drm_printf(p, "Render p-state limit: %d\n", 2367 rp_state_limits & 0xff); 2368 drm_printf(p, "RPSTAT1: 0x%08x\n", rpstat); 2369 drm_printf(p, "RPMODECTL: 0x%08x\n", rpmodectl); 2370 drm_printf(p, "RPINCLIMIT: 0x%08x\n", rpinclimit); 2371 drm_printf(p, "RPDECLIMIT: 0x%08x\n", rpdeclimit); 2372 drm_printf(p, "RPNSWREQ: %dMHz\n", reqf); 2373 drm_printf(p, "CAGF: %dMHz\n", cagf); 2374 drm_printf(p, "RP CUR UP EI: %d (%lldns)\n", 2375 rpcurupei, 2376 intel_gt_pm_interval_to_ns(gt, rpcurupei)); 2377 drm_printf(p, "RP CUR UP: %d (%lldns)\n", 2378 rpcurup, intel_gt_pm_interval_to_ns(gt, rpcurup)); 2379 drm_printf(p, "RP PREV UP: %d (%lldns)\n", 2380 rpprevup, intel_gt_pm_interval_to_ns(gt, rpprevup)); 2381 drm_printf(p, "Up threshold: %d%%\n", 2382 rps->power.up_threshold); 2383 drm_printf(p, "RP UP EI: %d (%lldns)\n", 2384 rpupei, intel_gt_pm_interval_to_ns(gt, rpupei)); 2385 drm_printf(p, "RP UP THRESHOLD: %d (%lldns)\n", 2386 rpupt, intel_gt_pm_interval_to_ns(gt, rpupt)); 2387 2388 drm_printf(p, "RP CUR DOWN EI: %d (%lldns)\n", 2389 rpcurdownei, 2390 intel_gt_pm_interval_to_ns(gt, rpcurdownei)); 2391 drm_printf(p, "RP CUR DOWN: %d (%lldns)\n", 2392 rpcurdown, 2393 intel_gt_pm_interval_to_ns(gt, rpcurdown)); 2394 drm_printf(p, "RP PREV DOWN: %d (%lldns)\n", 2395 rpprevdown, 2396 intel_gt_pm_interval_to_ns(gt, rpprevdown)); 2397 drm_printf(p, "Down threshold: %d%%\n", 2398 rps->power.down_threshold); 2399 drm_printf(p, "RP DOWN EI: %d (%lldns)\n", 2400 rpdownei, intel_gt_pm_interval_to_ns(gt, rpdownei)); 2401 drm_printf(p, "RP DOWN THRESHOLD: %d (%lldns)\n", 2402 rpdownt, intel_gt_pm_interval_to_ns(gt, rpdownt)); 2403 2404 drm_printf(p, "Lowest (RPN) frequency: %dMHz\n", 2405 intel_gpu_freq(rps, caps.min_freq)); 2406 drm_printf(p, "Nominal (RP1) frequency: %dMHz\n", 2407 intel_gpu_freq(rps, caps.rp1_freq)); 2408 drm_printf(p, "Max non-overclocked (RP0) frequency: %dMHz\n", 2409 intel_gpu_freq(rps, caps.rp0_freq)); 2410 drm_printf(p, "Max overclocked frequency: %dMHz\n", 2411 intel_gpu_freq(rps, rps->max_freq)); 2412 2413 drm_printf(p, "Current freq: %d MHz\n", 2414 intel_gpu_freq(rps, rps->cur_freq)); 2415 drm_printf(p, "Actual freq: %d MHz\n", cagf); 2416 drm_printf(p, "Idle freq: %d MHz\n", 2417 intel_gpu_freq(rps, rps->idle_freq)); 2418 drm_printf(p, "Min freq: %d MHz\n", 2419 intel_gpu_freq(rps, rps->min_freq)); 2420 drm_printf(p, "Boost freq: %d MHz\n", 2421 intel_gpu_freq(rps, rps->boost_freq)); 2422 drm_printf(p, "Max freq: %d MHz\n", 2423 intel_gpu_freq(rps, rps->max_freq)); 2424 drm_printf(p, 2425 "efficient (RPe) frequency: %d MHz\n", 2426 intel_gpu_freq(rps, rps->efficient_freq)); 2427 } 2428 2429 static void slpc_frequency_dump(struct intel_rps *rps, struct drm_printer *p) 2430 { 2431 struct intel_gt *gt = rps_to_gt(rps); 2432 struct intel_uncore *uncore = gt->uncore; 2433 struct intel_rps_freq_caps caps; 2434 u32 pm_mask; 2435 2436 gen6_rps_get_freq_caps(rps, &caps); 2437 pm_mask = intel_uncore_read(uncore, GEN6_PMINTRMSK); 2438 2439 drm_printf(p, "PM MASK=0x%08x\n", pm_mask); 2440 drm_printf(p, "pm_intrmsk_mbz: 0x%08x\n", 2441 rps->pm_intrmsk_mbz); 2442 drm_printf(p, "RPSTAT1: 0x%08x\n", intel_rps_read_rpstat(rps)); 2443 drm_printf(p, "RPNSWREQ: %dMHz\n", intel_rps_get_requested_frequency(rps)); 2444 drm_printf(p, "Lowest (RPN) frequency: %dMHz\n", 2445 intel_gpu_freq(rps, caps.min_freq)); 2446 drm_printf(p, "Nominal (RP1) frequency: %dMHz\n", 2447 intel_gpu_freq(rps, caps.rp1_freq)); 2448 drm_printf(p, "Max non-overclocked (RP0) frequency: %dMHz\n", 2449 intel_gpu_freq(rps, caps.rp0_freq)); 2450 drm_printf(p, "Current freq: %d MHz\n", 2451 intel_rps_get_requested_frequency(rps)); 2452 drm_printf(p, "Actual freq: %d MHz\n", 2453 intel_rps_read_actual_frequency(rps)); 2454 drm_printf(p, "Min freq: %d MHz\n", 2455 intel_rps_get_min_frequency(rps)); 2456 drm_printf(p, "Boost freq: %d MHz\n", 2457 intel_rps_get_boost_frequency(rps)); 2458 drm_printf(p, "Max freq: %d MHz\n", 2459 intel_rps_get_max_frequency(rps)); 2460 drm_printf(p, 2461 "efficient (RPe) frequency: %d MHz\n", 2462 intel_gpu_freq(rps, caps.rp1_freq)); 2463 } 2464 2465 void gen6_rps_frequency_dump(struct intel_rps *rps, struct drm_printer *p) 2466 { 2467 if (rps_uses_slpc(rps)) 2468 return slpc_frequency_dump(rps, p); 2469 else 2470 return rps_frequency_dump(rps, p); 2471 } 2472 2473 static int set_max_freq(struct intel_rps *rps, u32 val) 2474 { 2475 struct drm_i915_private *i915 = rps_to_i915(rps); 2476 int ret = 0; 2477 2478 mutex_lock(&rps->lock); 2479 2480 val = intel_freq_opcode(rps, val); 2481 if (val < rps->min_freq || 2482 val > rps->max_freq || 2483 val < rps->min_freq_softlimit) { 2484 ret = -EINVAL; 2485 goto unlock; 2486 } 2487 2488 if (val > rps->rp0_freq) 2489 drm_dbg(&i915->drm, "User requested overclocking to %d\n", 2490 intel_gpu_freq(rps, val)); 2491 2492 rps->max_freq_softlimit = val; 2493 2494 val = clamp_t(int, rps->cur_freq, 2495 rps->min_freq_softlimit, 2496 rps->max_freq_softlimit); 2497 2498 /* 2499 * We still need *_set_rps to process the new max_delay and 2500 * update the interrupt limits and PMINTRMSK even though 2501 * frequency request may be unchanged. 2502 */ 2503 intel_rps_set(rps, val); 2504 2505 unlock: 2506 mutex_unlock(&rps->lock); 2507 2508 return ret; 2509 } 2510 2511 int intel_rps_set_max_frequency(struct intel_rps *rps, u32 val) 2512 { 2513 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2514 2515 if (rps_uses_slpc(rps)) 2516 return intel_guc_slpc_set_max_freq(slpc, val); 2517 else 2518 return set_max_freq(rps, val); 2519 } 2520 2521 u32 intel_rps_get_min_frequency(struct intel_rps *rps) 2522 { 2523 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2524 2525 if (rps_uses_slpc(rps)) 2526 return slpc->min_freq_softlimit; 2527 else 2528 return intel_gpu_freq(rps, rps->min_freq_softlimit); 2529 } 2530 2531 /** 2532 * intel_rps_get_min_raw_freq - returns the min frequency in some raw format. 2533 * @rps: the intel_rps structure 2534 * 2535 * Returns the min frequency in a raw format. In newer platforms raw is in 2536 * units of 50 MHz. 2537 */ 2538 u32 intel_rps_get_min_raw_freq(struct intel_rps *rps) 2539 { 2540 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2541 u32 freq; 2542 2543 if (rps_uses_slpc(rps)) { 2544 return DIV_ROUND_CLOSEST(slpc->min_freq, 2545 GT_FREQUENCY_MULTIPLIER); 2546 } else { 2547 freq = rps->min_freq; 2548 if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) { 2549 /* Convert GT frequency to 50 MHz units */ 2550 freq /= GEN9_FREQ_SCALER; 2551 } 2552 return freq; 2553 } 2554 } 2555 2556 static int set_min_freq(struct intel_rps *rps, u32 val) 2557 { 2558 int ret = 0; 2559 2560 mutex_lock(&rps->lock); 2561 2562 val = intel_freq_opcode(rps, val); 2563 if (val < rps->min_freq || 2564 val > rps->max_freq || 2565 val > rps->max_freq_softlimit) { 2566 ret = -EINVAL; 2567 goto unlock; 2568 } 2569 2570 rps->min_freq_softlimit = val; 2571 2572 val = clamp_t(int, rps->cur_freq, 2573 rps->min_freq_softlimit, 2574 rps->max_freq_softlimit); 2575 2576 /* 2577 * We still need *_set_rps to process the new min_delay and 2578 * update the interrupt limits and PMINTRMSK even though 2579 * frequency request may be unchanged. 2580 */ 2581 intel_rps_set(rps, val); 2582 2583 unlock: 2584 mutex_unlock(&rps->lock); 2585 2586 return ret; 2587 } 2588 2589 int intel_rps_set_min_frequency(struct intel_rps *rps, u32 val) 2590 { 2591 struct intel_guc_slpc *slpc = rps_to_slpc(rps); 2592 2593 if (rps_uses_slpc(rps)) 2594 return intel_guc_slpc_set_min_freq(slpc, val); 2595 else 2596 return set_min_freq(rps, val); 2597 } 2598 2599 u8 intel_rps_get_up_threshold(struct intel_rps *rps) 2600 { 2601 return rps->power.up_threshold; 2602 } 2603 2604 static int rps_set_threshold(struct intel_rps *rps, u8 *threshold, u8 val) 2605 { 2606 int ret; 2607 2608 if (val > 100) 2609 return -EINVAL; 2610 2611 ret = mutex_lock_interruptible(&rps->lock); 2612 if (ret) 2613 return ret; 2614 2615 if (*threshold == val) 2616 goto out_unlock; 2617 2618 *threshold = val; 2619 2620 /* Force reset. */ 2621 rps->last_freq = -1; 2622 mutex_lock(&rps->power.mutex); 2623 rps->power.mode = -1; 2624 mutex_unlock(&rps->power.mutex); 2625 2626 intel_rps_set(rps, clamp(rps->cur_freq, 2627 rps->min_freq_softlimit, 2628 rps->max_freq_softlimit)); 2629 2630 out_unlock: 2631 mutex_unlock(&rps->lock); 2632 2633 return ret; 2634 } 2635 2636 int intel_rps_set_up_threshold(struct intel_rps *rps, u8 threshold) 2637 { 2638 return rps_set_threshold(rps, &rps->power.up_threshold, threshold); 2639 } 2640 2641 u8 intel_rps_get_down_threshold(struct intel_rps *rps) 2642 { 2643 return rps->power.down_threshold; 2644 } 2645 2646 int intel_rps_set_down_threshold(struct intel_rps *rps, u8 threshold) 2647 { 2648 return rps_set_threshold(rps, &rps->power.down_threshold, threshold); 2649 } 2650 2651 static void intel_rps_set_manual(struct intel_rps *rps, bool enable) 2652 { 2653 struct intel_uncore *uncore = rps_to_uncore(rps); 2654 u32 state = enable ? GEN9_RPSWCTL_ENABLE : GEN9_RPSWCTL_DISABLE; 2655 2656 /* Allow punit to process software requests */ 2657 intel_uncore_write(uncore, GEN6_RP_CONTROL, state); 2658 } 2659 2660 void intel_rps_raise_unslice(struct intel_rps *rps) 2661 { 2662 struct intel_uncore *uncore = rps_to_uncore(rps); 2663 2664 mutex_lock(&rps->lock); 2665 2666 if (rps_uses_slpc(rps)) { 2667 /* RP limits have not been initialized yet for SLPC path */ 2668 struct intel_rps_freq_caps caps; 2669 2670 gen6_rps_get_freq_caps(rps, &caps); 2671 2672 intel_rps_set_manual(rps, true); 2673 intel_uncore_write(uncore, GEN6_RPNSWREQ, 2674 ((caps.rp0_freq << 2675 GEN9_SW_REQ_UNSLICE_RATIO_SHIFT) | 2676 GEN9_IGNORE_SLICE_RATIO)); 2677 intel_rps_set_manual(rps, false); 2678 } else { 2679 intel_rps_set(rps, rps->rp0_freq); 2680 } 2681 2682 mutex_unlock(&rps->lock); 2683 } 2684 2685 void intel_rps_lower_unslice(struct intel_rps *rps) 2686 { 2687 struct intel_uncore *uncore = rps_to_uncore(rps); 2688 2689 mutex_lock(&rps->lock); 2690 2691 if (rps_uses_slpc(rps)) { 2692 /* RP limits have not been initialized yet for SLPC path */ 2693 struct intel_rps_freq_caps caps; 2694 2695 gen6_rps_get_freq_caps(rps, &caps); 2696 2697 intel_rps_set_manual(rps, true); 2698 intel_uncore_write(uncore, GEN6_RPNSWREQ, 2699 ((caps.min_freq << 2700 GEN9_SW_REQ_UNSLICE_RATIO_SHIFT) | 2701 GEN9_IGNORE_SLICE_RATIO)); 2702 intel_rps_set_manual(rps, false); 2703 } else { 2704 intel_rps_set(rps, rps->min_freq); 2705 } 2706 2707 mutex_unlock(&rps->lock); 2708 } 2709 2710 static u32 rps_read_mmio(struct intel_rps *rps, i915_reg_t reg32) 2711 { 2712 struct intel_gt *gt = rps_to_gt(rps); 2713 intel_wakeref_t wakeref; 2714 u32 val; 2715 2716 with_intel_runtime_pm(gt->uncore->rpm, wakeref) 2717 val = intel_uncore_read(gt->uncore, reg32); 2718 2719 return val; 2720 } 2721 2722 bool rps_read_mask_mmio(struct intel_rps *rps, 2723 i915_reg_t reg32, u32 mask) 2724 { 2725 return rps_read_mmio(rps, reg32) & mask; 2726 } 2727 2728 /* External interface for intel_ips.ko */ 2729 2730 static struct drm_i915_private __rcu *ips_mchdev; 2731 2732 /* 2733 * Tells the intel_ips driver that the i915 driver is now loaded, if 2734 * IPS got loaded first. 2735 * 2736 * This awkward dance is so that neither module has to depend on the 2737 * other in order for IPS to do the appropriate communication of 2738 * GPU turbo limits to i915. 2739 */ 2740 static void 2741 ips_ping_for_i915_load(void) 2742 { 2743 void (*link)(void); 2744 2745 link = symbol_get(ips_link_to_i915_driver); 2746 if (link) { 2747 link(); 2748 symbol_put(ips_link_to_i915_driver); 2749 } 2750 } 2751 2752 void intel_rps_driver_register(struct intel_rps *rps) 2753 { 2754 struct intel_gt *gt = rps_to_gt(rps); 2755 2756 /* 2757 * We only register the i915 ips part with intel-ips once everything is 2758 * set up, to avoid intel-ips sneaking in and reading bogus values. 2759 */ 2760 if (GRAPHICS_VER(gt->i915) == 5) { 2761 GEM_BUG_ON(ips_mchdev); 2762 rcu_assign_pointer(ips_mchdev, gt->i915); 2763 ips_ping_for_i915_load(); 2764 } 2765 } 2766 2767 void intel_rps_driver_unregister(struct intel_rps *rps) 2768 { 2769 if (rcu_access_pointer(ips_mchdev) == rps_to_i915(rps)) 2770 rcu_assign_pointer(ips_mchdev, NULL); 2771 } 2772 2773 static struct drm_i915_private *mchdev_get(void) 2774 { 2775 struct drm_i915_private *i915; 2776 2777 rcu_read_lock(); 2778 i915 = rcu_dereference(ips_mchdev); 2779 if (i915 && !kref_get_unless_zero(&i915->drm.ref)) 2780 i915 = NULL; 2781 rcu_read_unlock(); 2782 2783 return i915; 2784 } 2785 2786 /** 2787 * i915_read_mch_val - return value for IPS use 2788 * 2789 * Calculate and return a value for the IPS driver to use when deciding whether 2790 * we have thermal and power headroom to increase CPU or GPU power budget. 2791 */ 2792 unsigned long i915_read_mch_val(void) 2793 { 2794 struct drm_i915_private *i915; 2795 unsigned long chipset_val = 0; 2796 unsigned long graphics_val = 0; 2797 intel_wakeref_t wakeref; 2798 2799 i915 = mchdev_get(); 2800 if (!i915) 2801 return 0; 2802 2803 with_intel_runtime_pm(&i915->runtime_pm, wakeref) { 2804 struct intel_ips *ips = &to_gt(i915)->rps.ips; 2805 2806 spin_lock_irq(&mchdev_lock); 2807 chipset_val = __ips_chipset_val(ips); 2808 graphics_val = __ips_gfx_val(ips); 2809 spin_unlock_irq(&mchdev_lock); 2810 } 2811 2812 drm_dev_put(&i915->drm); 2813 return chipset_val + graphics_val; 2814 } 2815 EXPORT_SYMBOL_GPL(i915_read_mch_val); 2816 2817 /** 2818 * i915_gpu_raise - raise GPU frequency limit 2819 * 2820 * Raise the limit; IPS indicates we have thermal headroom. 2821 */ 2822 bool i915_gpu_raise(void) 2823 { 2824 struct drm_i915_private *i915; 2825 struct intel_rps *rps; 2826 2827 i915 = mchdev_get(); 2828 if (!i915) 2829 return false; 2830 2831 rps = &to_gt(i915)->rps; 2832 2833 spin_lock_irq(&mchdev_lock); 2834 if (rps->max_freq_softlimit < rps->max_freq) 2835 rps->max_freq_softlimit++; 2836 spin_unlock_irq(&mchdev_lock); 2837 2838 drm_dev_put(&i915->drm); 2839 return true; 2840 } 2841 EXPORT_SYMBOL_GPL(i915_gpu_raise); 2842 2843 /** 2844 * i915_gpu_lower - lower GPU frequency limit 2845 * 2846 * IPS indicates we're close to a thermal limit, so throttle back the GPU 2847 * frequency maximum. 2848 */ 2849 bool i915_gpu_lower(void) 2850 { 2851 struct drm_i915_private *i915; 2852 struct intel_rps *rps; 2853 2854 i915 = mchdev_get(); 2855 if (!i915) 2856 return false; 2857 2858 rps = &to_gt(i915)->rps; 2859 2860 spin_lock_irq(&mchdev_lock); 2861 if (rps->max_freq_softlimit > rps->min_freq) 2862 rps->max_freq_softlimit--; 2863 spin_unlock_irq(&mchdev_lock); 2864 2865 drm_dev_put(&i915->drm); 2866 return true; 2867 } 2868 EXPORT_SYMBOL_GPL(i915_gpu_lower); 2869 2870 /** 2871 * i915_gpu_busy - indicate GPU business to IPS 2872 * 2873 * Tell the IPS driver whether or not the GPU is busy. 2874 */ 2875 bool i915_gpu_busy(void) 2876 { 2877 struct drm_i915_private *i915; 2878 bool ret; 2879 2880 i915 = mchdev_get(); 2881 if (!i915) 2882 return false; 2883 2884 ret = to_gt(i915)->awake; 2885 2886 drm_dev_put(&i915->drm); 2887 return ret; 2888 } 2889 EXPORT_SYMBOL_GPL(i915_gpu_busy); 2890 2891 /** 2892 * i915_gpu_turbo_disable - disable graphics turbo 2893 * 2894 * Disable graphics turbo by resetting the max frequency and setting the 2895 * current frequency to the default. 2896 */ 2897 bool i915_gpu_turbo_disable(void) 2898 { 2899 struct drm_i915_private *i915; 2900 struct intel_rps *rps; 2901 bool ret; 2902 2903 i915 = mchdev_get(); 2904 if (!i915) 2905 return false; 2906 2907 rps = &to_gt(i915)->rps; 2908 2909 spin_lock_irq(&mchdev_lock); 2910 rps->max_freq_softlimit = rps->min_freq; 2911 ret = !__gen5_rps_set(&to_gt(i915)->rps, rps->min_freq); 2912 spin_unlock_irq(&mchdev_lock); 2913 2914 drm_dev_put(&i915->drm); 2915 return ret; 2916 } 2917 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable); 2918 2919 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2920 #include "selftest_rps.c" 2921 #include "selftest_slpc.c" 2922 #endif 2923