1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2019 Intel Corporation 5 */ 6 7 #include <linux/pm_runtime.h> 8 9 #include "i915_drv.h" 10 #include "i915_vgpu.h" 11 #include "intel_gt.h" 12 #include "intel_gt_pm.h" 13 #include "intel_rc6.h" 14 #include "intel_sideband.h" 15 16 /** 17 * DOC: RC6 18 * 19 * RC6 is a special power stage which allows the GPU to enter an very 20 * low-voltage mode when idle, using down to 0V while at this stage. This 21 * stage is entered automatically when the GPU is idle when RC6 support is 22 * enabled, and as soon as new workload arises GPU wakes up automatically as 23 * well. 24 * 25 * There are different RC6 modes available in Intel GPU, which differentiate 26 * among each other with the latency required to enter and leave RC6 and 27 * voltage consumed by the GPU in different states. 28 * 29 * The combination of the following flags define which states GPU is allowed 30 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and 31 * RC6pp is deepest RC6. Their support by hardware varies according to the 32 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one 33 * which brings the most power savings; deeper states save more power, but 34 * require higher latency to switch to and wake up. 35 */ 36 37 static struct intel_gt *rc6_to_gt(struct intel_rc6 *rc6) 38 { 39 return container_of(rc6, struct intel_gt, rc6); 40 } 41 42 static struct intel_uncore *rc6_to_uncore(struct intel_rc6 *rc) 43 { 44 return rc6_to_gt(rc)->uncore; 45 } 46 47 static struct drm_i915_private *rc6_to_i915(struct intel_rc6 *rc) 48 { 49 return rc6_to_gt(rc)->i915; 50 } 51 52 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val) 53 { 54 intel_uncore_write_fw(uncore, reg, val); 55 } 56 57 static void gen11_rc6_enable(struct intel_rc6 *rc6) 58 { 59 struct intel_gt *gt = rc6_to_gt(rc6); 60 struct intel_uncore *uncore = gt->uncore; 61 struct intel_engine_cs *engine; 62 enum intel_engine_id id; 63 u32 pg_enable; 64 int i; 65 66 /* 2b: Program RC6 thresholds.*/ 67 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85); 68 set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150); 69 70 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 71 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 72 for_each_engine(engine, rc6_to_gt(rc6), id) 73 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 74 75 set(uncore, GUC_MAX_IDLE_COUNT, 0xA); 76 77 set(uncore, GEN6_RC_SLEEP, 0); 78 79 set(uncore, GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */ 80 81 /* 82 * 2c: Program Coarse Power Gating Policies. 83 * 84 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we 85 * use instead is a more conservative estimate for the maximum time 86 * it takes us to service a CS interrupt and submit a new ELSP - that 87 * is the time which the GPU is idle waiting for the CPU to select the 88 * next request to execute. If the idle hysteresis is less than that 89 * interrupt service latency, the hardware will automatically gate 90 * the power well and we will then incur the wake up cost on top of 91 * the service latency. A similar guide from plane_state is that we 92 * do not want the enable hysteresis to less than the wakeup latency. 93 * 94 * igt/gem_exec_nop/sequential provides a rough estimate for the 95 * service latency, and puts it under 10us for Icelake, similar to 96 * Broadwell+, To be conservative, we want to factor in a context 97 * switch on top (due to ksoftirqd). 98 */ 99 set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 60); 100 set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 60); 101 102 /* 3a: Enable RC6 */ 103 rc6->ctl_enable = 104 GEN6_RC_CTL_HW_ENABLE | 105 GEN6_RC_CTL_RC6_ENABLE | 106 GEN6_RC_CTL_EI_MODE(1); 107 108 pg_enable = 109 GEN9_RENDER_PG_ENABLE | 110 GEN9_MEDIA_PG_ENABLE | 111 GEN11_MEDIA_SAMPLER_PG_ENABLE; 112 113 if (INTEL_GEN(gt->i915) >= 12) { 114 for (i = 0; i < I915_MAX_VCS; i++) 115 if (HAS_ENGINE(gt, _VCS(i))) 116 pg_enable |= (VDN_HCP_POWERGATE_ENABLE(i) | 117 VDN_MFX_POWERGATE_ENABLE(i)); 118 } 119 120 set(uncore, GEN9_PG_ENABLE, pg_enable); 121 } 122 123 static void gen9_rc6_enable(struct intel_rc6 *rc6) 124 { 125 struct intel_uncore *uncore = rc6_to_uncore(rc6); 126 struct intel_engine_cs *engine; 127 enum intel_engine_id id; 128 129 /* 2b: Program RC6 thresholds.*/ 130 if (INTEL_GEN(rc6_to_i915(rc6)) >= 10) { 131 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85); 132 set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150); 133 } else if (IS_SKYLAKE(rc6_to_i915(rc6))) { 134 /* 135 * WaRsDoubleRc6WrlWithCoarsePowerGating:skl Doubling WRL only 136 * when CPG is enabled 137 */ 138 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16); 139 } else { 140 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16); 141 } 142 143 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 144 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 145 for_each_engine(engine, rc6_to_gt(rc6), id) 146 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 147 148 set(uncore, GUC_MAX_IDLE_COUNT, 0xA); 149 150 set(uncore, GEN6_RC_SLEEP, 0); 151 152 /* 153 * 2c: Program Coarse Power Gating Policies. 154 * 155 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we 156 * use instead is a more conservative estimate for the maximum time 157 * it takes us to service a CS interrupt and submit a new ELSP - that 158 * is the time which the GPU is idle waiting for the CPU to select the 159 * next request to execute. If the idle hysteresis is less than that 160 * interrupt service latency, the hardware will automatically gate 161 * the power well and we will then incur the wake up cost on top of 162 * the service latency. A similar guide from plane_state is that we 163 * do not want the enable hysteresis to less than the wakeup latency. 164 * 165 * igt/gem_exec_nop/sequential provides a rough estimate for the 166 * service latency, and puts it around 10us for Broadwell (and other 167 * big core) and around 40us for Broxton (and other low power cores). 168 * [Note that for legacy ringbuffer submission, this is less than 1us!] 169 * However, the wakeup latency on Broxton is closer to 100us. To be 170 * conservative, we have to factor in a context switch on top (due 171 * to ksoftirqd). 172 */ 173 set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 250); 174 set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 250); 175 176 /* 3a: Enable RC6 */ 177 set(uncore, GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */ 178 179 180 rc6->ctl_enable = 181 GEN6_RC_CTL_HW_ENABLE | 182 GEN6_RC_CTL_RC6_ENABLE | 183 GEN6_RC_CTL_EI_MODE(1); 184 185 /* 186 * WaRsDisableCoarsePowerGating:skl,cnl 187 * - Render/Media PG need to be disabled with RC6. 188 */ 189 if (!NEEDS_WaRsDisableCoarsePowerGating(rc6_to_i915(rc6))) 190 set(uncore, GEN9_PG_ENABLE, 191 GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE); 192 } 193 194 static void gen8_rc6_enable(struct intel_rc6 *rc6) 195 { 196 struct intel_uncore *uncore = rc6_to_uncore(rc6); 197 struct intel_engine_cs *engine; 198 enum intel_engine_id id; 199 200 /* 2b: Program RC6 thresholds.*/ 201 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16); 202 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 203 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 204 for_each_engine(engine, rc6_to_gt(rc6), id) 205 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 206 set(uncore, GEN6_RC_SLEEP, 0); 207 set(uncore, GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */ 208 209 /* 3: Enable RC6 */ 210 rc6->ctl_enable = 211 GEN6_RC_CTL_HW_ENABLE | 212 GEN7_RC_CTL_TO_MODE | 213 GEN6_RC_CTL_RC6_ENABLE; 214 } 215 216 static void gen6_rc6_enable(struct intel_rc6 *rc6) 217 { 218 struct intel_uncore *uncore = rc6_to_uncore(rc6); 219 struct drm_i915_private *i915 = rc6_to_i915(rc6); 220 struct intel_engine_cs *engine; 221 enum intel_engine_id id; 222 u32 rc6vids, rc6_mask; 223 int ret; 224 225 set(uncore, GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16); 226 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30); 227 set(uncore, GEN6_RC6pp_WAKE_RATE_LIMIT, 30); 228 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); 229 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); 230 231 for_each_engine(engine, rc6_to_gt(rc6), id) 232 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 233 234 set(uncore, GEN6_RC_SLEEP, 0); 235 set(uncore, GEN6_RC1e_THRESHOLD, 1000); 236 set(uncore, GEN6_RC6_THRESHOLD, 50000); 237 set(uncore, GEN6_RC6p_THRESHOLD, 150000); 238 set(uncore, GEN6_RC6pp_THRESHOLD, 64000); /* unused */ 239 240 /* We don't use those on Haswell */ 241 rc6_mask = GEN6_RC_CTL_RC6_ENABLE; 242 if (HAS_RC6p(i915)) 243 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE; 244 if (HAS_RC6pp(i915)) 245 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE; 246 rc6->ctl_enable = 247 rc6_mask | 248 GEN6_RC_CTL_EI_MODE(1) | 249 GEN6_RC_CTL_HW_ENABLE; 250 251 rc6vids = 0; 252 ret = sandybridge_pcode_read(i915, GEN6_PCODE_READ_RC6VIDS, 253 &rc6vids, NULL); 254 if (IS_GEN(i915, 6) && ret) { 255 drm_dbg(&i915->drm, "Couldn't check for BIOS workaround\n"); 256 } else if (IS_GEN(i915, 6) && 257 (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) { 258 drm_dbg(&i915->drm, 259 "You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n", 260 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450); 261 rc6vids &= 0xffff00; 262 rc6vids |= GEN6_ENCODE_RC6_VID(450); 263 ret = sandybridge_pcode_write(i915, GEN6_PCODE_WRITE_RC6VIDS, rc6vids); 264 if (ret) 265 drm_err(&i915->drm, 266 "Couldn't fix incorrect rc6 voltage\n"); 267 } 268 } 269 270 /* Check that the pcbr address is not empty. */ 271 static int chv_rc6_init(struct intel_rc6 *rc6) 272 { 273 struct intel_uncore *uncore = rc6_to_uncore(rc6); 274 struct drm_i915_private *i915 = rc6_to_i915(rc6); 275 resource_size_t pctx_paddr, paddr; 276 resource_size_t pctx_size = 32 * SZ_1K; 277 u32 pcbr; 278 279 pcbr = intel_uncore_read(uncore, VLV_PCBR); 280 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) { 281 drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n"); 282 paddr = i915->dsm.end + 1 - pctx_size; 283 GEM_BUG_ON(paddr > U32_MAX); 284 285 pctx_paddr = (paddr & ~4095); 286 intel_uncore_write(uncore, VLV_PCBR, pctx_paddr); 287 } 288 289 return 0; 290 } 291 292 static int vlv_rc6_init(struct intel_rc6 *rc6) 293 { 294 struct drm_i915_private *i915 = rc6_to_i915(rc6); 295 struct intel_uncore *uncore = rc6_to_uncore(rc6); 296 struct drm_i915_gem_object *pctx; 297 resource_size_t pctx_paddr; 298 resource_size_t pctx_size = 24 * SZ_1K; 299 u32 pcbr; 300 301 pcbr = intel_uncore_read(uncore, VLV_PCBR); 302 if (pcbr) { 303 /* BIOS set it up already, grab the pre-alloc'd space */ 304 resource_size_t pcbr_offset; 305 306 pcbr_offset = (pcbr & ~4095) - i915->dsm.start; 307 pctx = i915_gem_object_create_stolen_for_preallocated(i915, 308 pcbr_offset, 309 pctx_size); 310 if (IS_ERR(pctx)) 311 return PTR_ERR(pctx); 312 313 goto out; 314 } 315 316 drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n"); 317 318 /* 319 * From the Gunit register HAS: 320 * The Gfx driver is expected to program this register and ensure 321 * proper allocation within Gfx stolen memory. For example, this 322 * register should be programmed such than the PCBR range does not 323 * overlap with other ranges, such as the frame buffer, protected 324 * memory, or any other relevant ranges. 325 */ 326 pctx = i915_gem_object_create_stolen(i915, pctx_size); 327 if (IS_ERR(pctx)) { 328 drm_dbg(&i915->drm, 329 "not enough stolen space for PCTX, disabling\n"); 330 return PTR_ERR(pctx); 331 } 332 333 GEM_BUG_ON(range_overflows_end_t(u64, 334 i915->dsm.start, 335 pctx->stolen->start, 336 U32_MAX)); 337 pctx_paddr = i915->dsm.start + pctx->stolen->start; 338 intel_uncore_write(uncore, VLV_PCBR, pctx_paddr); 339 340 out: 341 rc6->pctx = pctx; 342 return 0; 343 } 344 345 static void chv_rc6_enable(struct intel_rc6 *rc6) 346 { 347 struct intel_uncore *uncore = rc6_to_uncore(rc6); 348 struct intel_engine_cs *engine; 349 enum intel_engine_id id; 350 351 /* 2a: Program RC6 thresholds.*/ 352 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16); 353 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 354 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 355 356 for_each_engine(engine, rc6_to_gt(rc6), id) 357 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 358 set(uncore, GEN6_RC_SLEEP, 0); 359 360 /* TO threshold set to 500 us (0x186 * 1.28 us) */ 361 set(uncore, GEN6_RC6_THRESHOLD, 0x186); 362 363 /* Allows RC6 residency counter to work */ 364 set(uncore, VLV_COUNTER_CONTROL, 365 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH | 366 VLV_MEDIA_RC6_COUNT_EN | 367 VLV_RENDER_RC6_COUNT_EN)); 368 369 /* 3: Enable RC6 */ 370 rc6->ctl_enable = GEN7_RC_CTL_TO_MODE; 371 } 372 373 static void vlv_rc6_enable(struct intel_rc6 *rc6) 374 { 375 struct intel_uncore *uncore = rc6_to_uncore(rc6); 376 struct intel_engine_cs *engine; 377 enum intel_engine_id id; 378 379 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000); 380 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); 381 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); 382 383 for_each_engine(engine, rc6_to_gt(rc6), id) 384 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 385 386 set(uncore, GEN6_RC6_THRESHOLD, 0x557); 387 388 /* Allows RC6 residency counter to work */ 389 set(uncore, VLV_COUNTER_CONTROL, 390 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH | 391 VLV_MEDIA_RC0_COUNT_EN | 392 VLV_RENDER_RC0_COUNT_EN | 393 VLV_MEDIA_RC6_COUNT_EN | 394 VLV_RENDER_RC6_COUNT_EN)); 395 396 rc6->ctl_enable = 397 GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL; 398 } 399 400 static bool bxt_check_bios_rc6_setup(struct intel_rc6 *rc6) 401 { 402 struct intel_uncore *uncore = rc6_to_uncore(rc6); 403 struct drm_i915_private *i915 = rc6_to_i915(rc6); 404 u32 rc6_ctx_base, rc_ctl, rc_sw_target; 405 bool enable_rc6 = true; 406 407 rc_ctl = intel_uncore_read(uncore, GEN6_RC_CONTROL); 408 rc_sw_target = intel_uncore_read(uncore, GEN6_RC_STATE); 409 rc_sw_target &= RC_SW_TARGET_STATE_MASK; 410 rc_sw_target >>= RC_SW_TARGET_STATE_SHIFT; 411 drm_dbg(&i915->drm, "BIOS enabled RC states: " 412 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n", 413 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE), 414 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE), 415 rc_sw_target); 416 417 if (!(intel_uncore_read(uncore, RC6_LOCATION) & RC6_CTX_IN_DRAM)) { 418 drm_dbg(&i915->drm, "RC6 Base location not set properly.\n"); 419 enable_rc6 = false; 420 } 421 422 /* 423 * The exact context size is not known for BXT, so assume a page size 424 * for this check. 425 */ 426 rc6_ctx_base = 427 intel_uncore_read(uncore, RC6_CTX_BASE) & RC6_CTX_BASE_MASK; 428 if (!(rc6_ctx_base >= i915->dsm_reserved.start && 429 rc6_ctx_base + PAGE_SIZE < i915->dsm_reserved.end)) { 430 drm_dbg(&i915->drm, "RC6 Base address not as expected.\n"); 431 enable_rc6 = false; 432 } 433 434 if (!((intel_uncore_read(uncore, PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1 && 435 (intel_uncore_read(uncore, PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1 && 436 (intel_uncore_read(uncore, PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1 && 437 (intel_uncore_read(uncore, PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1)) { 438 drm_dbg(&i915->drm, 439 "Engine Idle wait time not set properly.\n"); 440 enable_rc6 = false; 441 } 442 443 if (!intel_uncore_read(uncore, GEN8_PUSHBUS_CONTROL) || 444 !intel_uncore_read(uncore, GEN8_PUSHBUS_ENABLE) || 445 !intel_uncore_read(uncore, GEN8_PUSHBUS_SHIFT)) { 446 drm_dbg(&i915->drm, "Pushbus not setup properly.\n"); 447 enable_rc6 = false; 448 } 449 450 if (!intel_uncore_read(uncore, GEN6_GFXPAUSE)) { 451 drm_dbg(&i915->drm, "GFX pause not setup properly.\n"); 452 enable_rc6 = false; 453 } 454 455 if (!intel_uncore_read(uncore, GEN8_MISC_CTRL0)) { 456 drm_dbg(&i915->drm, "GPM control not setup properly.\n"); 457 enable_rc6 = false; 458 } 459 460 return enable_rc6; 461 } 462 463 static bool rc6_supported(struct intel_rc6 *rc6) 464 { 465 struct drm_i915_private *i915 = rc6_to_i915(rc6); 466 467 if (!HAS_RC6(i915)) 468 return false; 469 470 if (intel_vgpu_active(i915)) 471 return false; 472 473 if (is_mock_gt(rc6_to_gt(rc6))) 474 return false; 475 476 if (IS_GEN9_LP(i915) && !bxt_check_bios_rc6_setup(rc6)) { 477 drm_notice(&i915->drm, 478 "RC6 and powersaving disabled by BIOS\n"); 479 return false; 480 } 481 482 return true; 483 } 484 485 static void rpm_get(struct intel_rc6 *rc6) 486 { 487 GEM_BUG_ON(rc6->wakeref); 488 pm_runtime_get_sync(&rc6_to_i915(rc6)->drm.pdev->dev); 489 rc6->wakeref = true; 490 } 491 492 static void rpm_put(struct intel_rc6 *rc6) 493 { 494 GEM_BUG_ON(!rc6->wakeref); 495 pm_runtime_put(&rc6_to_i915(rc6)->drm.pdev->dev); 496 rc6->wakeref = false; 497 } 498 499 static bool pctx_corrupted(struct intel_rc6 *rc6) 500 { 501 struct drm_i915_private *i915 = rc6_to_i915(rc6); 502 503 if (!NEEDS_RC6_CTX_CORRUPTION_WA(i915)) 504 return false; 505 506 if (intel_uncore_read(rc6_to_uncore(rc6), GEN8_RC6_CTX_INFO)) 507 return false; 508 509 drm_notice(&i915->drm, 510 "RC6 context corruption, disabling runtime power management\n"); 511 return true; 512 } 513 514 static void __intel_rc6_disable(struct intel_rc6 *rc6) 515 { 516 struct drm_i915_private *i915 = rc6_to_i915(rc6); 517 struct intel_uncore *uncore = rc6_to_uncore(rc6); 518 519 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 520 if (INTEL_GEN(i915) >= 9) 521 set(uncore, GEN9_PG_ENABLE, 0); 522 set(uncore, GEN6_RC_CONTROL, 0); 523 set(uncore, GEN6_RC_STATE, 0); 524 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 525 } 526 527 void intel_rc6_init(struct intel_rc6 *rc6) 528 { 529 struct drm_i915_private *i915 = rc6_to_i915(rc6); 530 int err; 531 532 /* Disable runtime-pm until we can save the GPU state with rc6 pctx */ 533 rpm_get(rc6); 534 535 if (!rc6_supported(rc6)) 536 return; 537 538 if (IS_CHERRYVIEW(i915)) 539 err = chv_rc6_init(rc6); 540 else if (IS_VALLEYVIEW(i915)) 541 err = vlv_rc6_init(rc6); 542 else 543 err = 0; 544 545 /* Sanitize rc6, ensure it is disabled before we are ready. */ 546 __intel_rc6_disable(rc6); 547 548 rc6->supported = err == 0; 549 } 550 551 void intel_rc6_sanitize(struct intel_rc6 *rc6) 552 { 553 memset(rc6->prev_hw_residency, 0, sizeof(rc6->prev_hw_residency)); 554 555 if (rc6->enabled) { /* unbalanced suspend/resume */ 556 rpm_get(rc6); 557 rc6->enabled = false; 558 } 559 560 if (rc6->supported) 561 __intel_rc6_disable(rc6); 562 } 563 564 void intel_rc6_enable(struct intel_rc6 *rc6) 565 { 566 struct drm_i915_private *i915 = rc6_to_i915(rc6); 567 struct intel_uncore *uncore = rc6_to_uncore(rc6); 568 569 if (!rc6->supported) 570 return; 571 572 GEM_BUG_ON(rc6->enabled); 573 574 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 575 576 if (IS_CHERRYVIEW(i915)) 577 chv_rc6_enable(rc6); 578 else if (IS_VALLEYVIEW(i915)) 579 vlv_rc6_enable(rc6); 580 else if (INTEL_GEN(i915) >= 11) 581 gen11_rc6_enable(rc6); 582 else if (INTEL_GEN(i915) >= 9) 583 gen9_rc6_enable(rc6); 584 else if (IS_BROADWELL(i915)) 585 gen8_rc6_enable(rc6); 586 else if (INTEL_GEN(i915) >= 6) 587 gen6_rc6_enable(rc6); 588 589 rc6->manual = rc6->ctl_enable & GEN6_RC_CTL_RC6_ENABLE; 590 if (NEEDS_RC6_CTX_CORRUPTION_WA(i915)) 591 rc6->ctl_enable = 0; 592 593 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 594 595 if (unlikely(pctx_corrupted(rc6))) 596 return; 597 598 /* rc6 is ready, runtime-pm is go! */ 599 rpm_put(rc6); 600 rc6->enabled = true; 601 } 602 603 void intel_rc6_unpark(struct intel_rc6 *rc6) 604 { 605 struct intel_uncore *uncore = rc6_to_uncore(rc6); 606 607 if (!rc6->enabled) 608 return; 609 610 /* Restore HW timers for automatic RC6 entry while busy */ 611 set(uncore, GEN6_RC_CONTROL, rc6->ctl_enable); 612 } 613 614 void intel_rc6_park(struct intel_rc6 *rc6) 615 { 616 struct intel_uncore *uncore = rc6_to_uncore(rc6); 617 unsigned int target; 618 619 if (!rc6->enabled) 620 return; 621 622 if (unlikely(pctx_corrupted(rc6))) { 623 intel_rc6_disable(rc6); 624 return; 625 } 626 627 if (!rc6->manual) 628 return; 629 630 /* Turn off the HW timers and go directly to rc6 */ 631 set(uncore, GEN6_RC_CONTROL, GEN6_RC_CTL_RC6_ENABLE); 632 633 if (HAS_RC6pp(rc6_to_i915(rc6))) 634 target = 0x6; /* deepest rc6 */ 635 else if (HAS_RC6p(rc6_to_i915(rc6))) 636 target = 0x5; /* deep rc6 */ 637 else 638 target = 0x4; /* normal rc6 */ 639 set(uncore, GEN6_RC_STATE, target << RC_SW_TARGET_STATE_SHIFT); 640 } 641 642 void intel_rc6_disable(struct intel_rc6 *rc6) 643 { 644 if (!rc6->enabled) 645 return; 646 647 rpm_get(rc6); 648 rc6->enabled = false; 649 650 __intel_rc6_disable(rc6); 651 } 652 653 void intel_rc6_fini(struct intel_rc6 *rc6) 654 { 655 struct drm_i915_gem_object *pctx; 656 657 intel_rc6_disable(rc6); 658 659 pctx = fetch_and_zero(&rc6->pctx); 660 if (pctx) 661 i915_gem_object_put(pctx); 662 663 if (rc6->wakeref) 664 rpm_put(rc6); 665 } 666 667 static u64 vlv_residency_raw(struct intel_uncore *uncore, const i915_reg_t reg) 668 { 669 u32 lower, upper, tmp; 670 int loop = 2; 671 672 /* 673 * The register accessed do not need forcewake. We borrow 674 * uncore lock to prevent concurrent access to range reg. 675 */ 676 lockdep_assert_held(&uncore->lock); 677 678 /* 679 * vlv and chv residency counters are 40 bits in width. 680 * With a control bit, we can choose between upper or lower 681 * 32bit window into this counter. 682 * 683 * Although we always use the counter in high-range mode elsewhere, 684 * userspace may attempt to read the value before rc6 is initialised, 685 * before we have set the default VLV_COUNTER_CONTROL value. So always 686 * set the high bit to be safe. 687 */ 688 set(uncore, VLV_COUNTER_CONTROL, 689 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH)); 690 upper = intel_uncore_read_fw(uncore, reg); 691 do { 692 tmp = upper; 693 694 set(uncore, VLV_COUNTER_CONTROL, 695 _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH)); 696 lower = intel_uncore_read_fw(uncore, reg); 697 698 set(uncore, VLV_COUNTER_CONTROL, 699 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH)); 700 upper = intel_uncore_read_fw(uncore, reg); 701 } while (upper != tmp && --loop); 702 703 /* 704 * Everywhere else we always use VLV_COUNTER_CONTROL with the 705 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set 706 * now. 707 */ 708 709 return lower | (u64)upper << 8; 710 } 711 712 u64 intel_rc6_residency_ns(struct intel_rc6 *rc6, const i915_reg_t reg) 713 { 714 struct drm_i915_private *i915 = rc6_to_i915(rc6); 715 struct intel_uncore *uncore = rc6_to_uncore(rc6); 716 u64 time_hw, prev_hw, overflow_hw; 717 unsigned int fw_domains; 718 unsigned long flags; 719 unsigned int i; 720 u32 mul, div; 721 722 if (!rc6->supported) 723 return 0; 724 725 /* 726 * Store previous hw counter values for counter wrap-around handling. 727 * 728 * There are only four interesting registers and they live next to each 729 * other so we can use the relative address, compared to the smallest 730 * one as the index into driver storage. 731 */ 732 i = (i915_mmio_reg_offset(reg) - 733 i915_mmio_reg_offset(GEN6_GT_GFX_RC6_LOCKED)) / sizeof(u32); 734 if (drm_WARN_ON_ONCE(&i915->drm, i >= ARRAY_SIZE(rc6->cur_residency))) 735 return 0; 736 737 fw_domains = intel_uncore_forcewake_for_reg(uncore, reg, FW_REG_READ); 738 739 spin_lock_irqsave(&uncore->lock, flags); 740 intel_uncore_forcewake_get__locked(uncore, fw_domains); 741 742 /* On VLV and CHV, residency time is in CZ units rather than 1.28us */ 743 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) { 744 mul = 1000000; 745 div = i915->czclk_freq; 746 overflow_hw = BIT_ULL(40); 747 time_hw = vlv_residency_raw(uncore, reg); 748 } else { 749 /* 833.33ns units on Gen9LP, 1.28us elsewhere. */ 750 if (IS_GEN9_LP(i915)) { 751 mul = 10000; 752 div = 12; 753 } else { 754 mul = 1280; 755 div = 1; 756 } 757 758 overflow_hw = BIT_ULL(32); 759 time_hw = intel_uncore_read_fw(uncore, reg); 760 } 761 762 /* 763 * Counter wrap handling. 764 * 765 * But relying on a sufficient frequency of queries otherwise counters 766 * can still wrap. 767 */ 768 prev_hw = rc6->prev_hw_residency[i]; 769 rc6->prev_hw_residency[i] = time_hw; 770 771 /* RC6 delta from last sample. */ 772 if (time_hw >= prev_hw) 773 time_hw -= prev_hw; 774 else 775 time_hw += overflow_hw - prev_hw; 776 777 /* Add delta to RC6 extended raw driver copy. */ 778 time_hw += rc6->cur_residency[i]; 779 rc6->cur_residency[i] = time_hw; 780 781 intel_uncore_forcewake_put__locked(uncore, fw_domains); 782 spin_unlock_irqrestore(&uncore->lock, flags); 783 784 return mul_u64_u32_div(time_hw, mul, div); 785 } 786 787 u64 intel_rc6_residency_us(struct intel_rc6 *rc6, i915_reg_t reg) 788 { 789 return DIV_ROUND_UP_ULL(intel_rc6_residency_ns(rc6, reg), 1000); 790 } 791 792 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 793 #include "selftest_rc6.c" 794 #endif 795