1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2019 Intel Corporation 4 */ 5 6 #include <linux/pm_runtime.h> 7 #include <linux/string_helpers.h> 8 9 #include "gem/i915_gem_region.h" 10 #include "i915_drv.h" 11 #include "i915_reg.h" 12 #include "i915_vgpu.h" 13 #include "intel_engine_regs.h" 14 #include "intel_gt.h" 15 #include "intel_gt_pm.h" 16 #include "intel_gt_regs.h" 17 #include "intel_pcode.h" 18 #include "intel_rc6.h" 19 20 /** 21 * DOC: RC6 22 * 23 * RC6 is a special power stage which allows the GPU to enter an very 24 * low-voltage mode when idle, using down to 0V while at this stage. This 25 * stage is entered automatically when the GPU is idle when RC6 support is 26 * enabled, and as soon as new workload arises GPU wakes up automatically as 27 * well. 28 * 29 * There are different RC6 modes available in Intel GPU, which differentiate 30 * among each other with the latency required to enter and leave RC6 and 31 * voltage consumed by the GPU in different states. 32 * 33 * The combination of the following flags define which states GPU is allowed 34 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and 35 * RC6pp is deepest RC6. Their support by hardware varies according to the 36 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one 37 * which brings the most power savings; deeper states save more power, but 38 * require higher latency to switch to and wake up. 39 */ 40 41 static struct intel_gt *rc6_to_gt(struct intel_rc6 *rc6) 42 { 43 return container_of(rc6, struct intel_gt, rc6); 44 } 45 46 static struct intel_uncore *rc6_to_uncore(struct intel_rc6 *rc) 47 { 48 return rc6_to_gt(rc)->uncore; 49 } 50 51 static struct drm_i915_private *rc6_to_i915(struct intel_rc6 *rc) 52 { 53 return rc6_to_gt(rc)->i915; 54 } 55 56 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val) 57 { 58 intel_uncore_write_fw(uncore, reg, val); 59 } 60 61 static void gen11_rc6_enable(struct intel_rc6 *rc6) 62 { 63 struct intel_gt *gt = rc6_to_gt(rc6); 64 struct intel_uncore *uncore = gt->uncore; 65 struct intel_engine_cs *engine; 66 enum intel_engine_id id; 67 u32 pg_enable; 68 int i; 69 70 /* 71 * With GuCRC, these parameters are set by GuC 72 */ 73 if (!intel_uc_uses_guc_rc(>->uc)) { 74 /* 2b: Program RC6 thresholds.*/ 75 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85); 76 set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150); 77 78 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 79 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 80 for_each_engine(engine, rc6_to_gt(rc6), id) 81 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 82 83 set(uncore, GUC_MAX_IDLE_COUNT, 0xA); 84 85 set(uncore, GEN6_RC_SLEEP, 0); 86 87 set(uncore, GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */ 88 } 89 90 /* 91 * 2c: Program Coarse Power Gating Policies. 92 * 93 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we 94 * use instead is a more conservative estimate for the maximum time 95 * it takes us to service a CS interrupt and submit a new ELSP - that 96 * is the time which the GPU is idle waiting for the CPU to select the 97 * next request to execute. If the idle hysteresis is less than that 98 * interrupt service latency, the hardware will automatically gate 99 * the power well and we will then incur the wake up cost on top of 100 * the service latency. A similar guide from plane_state is that we 101 * do not want the enable hysteresis to less than the wakeup latency. 102 * 103 * igt/gem_exec_nop/sequential provides a rough estimate for the 104 * service latency, and puts it under 10us for Icelake, similar to 105 * Broadwell+, To be conservative, we want to factor in a context 106 * switch on top (due to ksoftirqd). 107 */ 108 set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 60); 109 set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 60); 110 111 /* 3a: Enable RC6 112 * 113 * With GuCRC, we do not enable bit 31 of RC_CTL, 114 * thus allowing GuC to control RC6 entry/exit fully instead. 115 * We will not set the HW ENABLE and EI bits 116 */ 117 if (!intel_guc_rc_enable(>->uc.guc)) 118 rc6->ctl_enable = GEN6_RC_CTL_RC6_ENABLE; 119 else 120 rc6->ctl_enable = 121 GEN6_RC_CTL_HW_ENABLE | 122 GEN6_RC_CTL_RC6_ENABLE | 123 GEN6_RC_CTL_EI_MODE(1); 124 125 /* Wa_16011777198 - Render powergating must remain disabled */ 126 if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_C0) || 127 IS_DG2_GRAPHICS_STEP(gt->i915, G11, STEP_A0, STEP_B0)) 128 pg_enable = 129 GEN9_MEDIA_PG_ENABLE | 130 GEN11_MEDIA_SAMPLER_PG_ENABLE; 131 else 132 pg_enable = 133 GEN9_RENDER_PG_ENABLE | 134 GEN9_MEDIA_PG_ENABLE | 135 GEN11_MEDIA_SAMPLER_PG_ENABLE; 136 137 if (GRAPHICS_VER(gt->i915) >= 12) { 138 for (i = 0; i < I915_MAX_VCS; i++) 139 if (HAS_ENGINE(gt, _VCS(i))) 140 pg_enable |= (VDN_HCP_POWERGATE_ENABLE(i) | 141 VDN_MFX_POWERGATE_ENABLE(i)); 142 } 143 144 set(uncore, GEN9_PG_ENABLE, pg_enable); 145 } 146 147 static void gen9_rc6_enable(struct intel_rc6 *rc6) 148 { 149 struct intel_uncore *uncore = rc6_to_uncore(rc6); 150 struct intel_engine_cs *engine; 151 enum intel_engine_id id; 152 153 /* 2b: Program RC6 thresholds.*/ 154 if (GRAPHICS_VER(rc6_to_i915(rc6)) >= 11) { 155 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85); 156 set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150); 157 } else if (IS_SKYLAKE(rc6_to_i915(rc6))) { 158 /* 159 * WaRsDoubleRc6WrlWithCoarsePowerGating:skl Doubling WRL only 160 * when CPG is enabled 161 */ 162 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16); 163 } else { 164 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16); 165 } 166 167 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 168 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 169 for_each_engine(engine, rc6_to_gt(rc6), id) 170 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 171 172 set(uncore, GUC_MAX_IDLE_COUNT, 0xA); 173 174 set(uncore, GEN6_RC_SLEEP, 0); 175 176 /* 177 * 2c: Program Coarse Power Gating Policies. 178 * 179 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we 180 * use instead is a more conservative estimate for the maximum time 181 * it takes us to service a CS interrupt and submit a new ELSP - that 182 * is the time which the GPU is idle waiting for the CPU to select the 183 * next request to execute. If the idle hysteresis is less than that 184 * interrupt service latency, the hardware will automatically gate 185 * the power well and we will then incur the wake up cost on top of 186 * the service latency. A similar guide from plane_state is that we 187 * do not want the enable hysteresis to less than the wakeup latency. 188 * 189 * igt/gem_exec_nop/sequential provides a rough estimate for the 190 * service latency, and puts it around 10us for Broadwell (and other 191 * big core) and around 40us for Broxton (and other low power cores). 192 * [Note that for legacy ringbuffer submission, this is less than 1us!] 193 * However, the wakeup latency on Broxton is closer to 100us. To be 194 * conservative, we have to factor in a context switch on top (due 195 * to ksoftirqd). 196 */ 197 set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 250); 198 set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 250); 199 200 /* 3a: Enable RC6 */ 201 set(uncore, GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */ 202 203 rc6->ctl_enable = 204 GEN6_RC_CTL_HW_ENABLE | 205 GEN6_RC_CTL_RC6_ENABLE | 206 GEN6_RC_CTL_EI_MODE(1); 207 208 /* 209 * WaRsDisableCoarsePowerGating:skl,cnl 210 * - Render/Media PG need to be disabled with RC6. 211 */ 212 if (!NEEDS_WaRsDisableCoarsePowerGating(rc6_to_i915(rc6))) 213 set(uncore, GEN9_PG_ENABLE, 214 GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE); 215 } 216 217 static void gen8_rc6_enable(struct intel_rc6 *rc6) 218 { 219 struct intel_uncore *uncore = rc6_to_uncore(rc6); 220 struct intel_engine_cs *engine; 221 enum intel_engine_id id; 222 223 /* 2b: Program RC6 thresholds.*/ 224 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16); 225 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 226 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 227 for_each_engine(engine, rc6_to_gt(rc6), id) 228 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 229 set(uncore, GEN6_RC_SLEEP, 0); 230 set(uncore, GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */ 231 232 /* 3: Enable RC6 */ 233 rc6->ctl_enable = 234 GEN6_RC_CTL_HW_ENABLE | 235 GEN7_RC_CTL_TO_MODE | 236 GEN6_RC_CTL_RC6_ENABLE; 237 } 238 239 static void gen6_rc6_enable(struct intel_rc6 *rc6) 240 { 241 struct intel_uncore *uncore = rc6_to_uncore(rc6); 242 struct drm_i915_private *i915 = rc6_to_i915(rc6); 243 struct intel_engine_cs *engine; 244 enum intel_engine_id id; 245 u32 rc6vids, rc6_mask; 246 int ret; 247 248 set(uncore, GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16); 249 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30); 250 set(uncore, GEN6_RC6pp_WAKE_RATE_LIMIT, 30); 251 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); 252 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); 253 254 for_each_engine(engine, rc6_to_gt(rc6), id) 255 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 256 257 set(uncore, GEN6_RC_SLEEP, 0); 258 set(uncore, GEN6_RC1e_THRESHOLD, 1000); 259 set(uncore, GEN6_RC6_THRESHOLD, 50000); 260 set(uncore, GEN6_RC6p_THRESHOLD, 150000); 261 set(uncore, GEN6_RC6pp_THRESHOLD, 64000); /* unused */ 262 263 /* We don't use those on Haswell */ 264 rc6_mask = GEN6_RC_CTL_RC6_ENABLE; 265 if (HAS_RC6p(i915)) 266 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE; 267 if (HAS_RC6pp(i915)) 268 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE; 269 rc6->ctl_enable = 270 rc6_mask | 271 GEN6_RC_CTL_EI_MODE(1) | 272 GEN6_RC_CTL_HW_ENABLE; 273 274 rc6vids = 0; 275 ret = snb_pcode_read(i915, GEN6_PCODE_READ_RC6VIDS, &rc6vids, NULL); 276 if (GRAPHICS_VER(i915) == 6 && ret) { 277 drm_dbg(&i915->drm, "Couldn't check for BIOS workaround\n"); 278 } else if (GRAPHICS_VER(i915) == 6 && 279 (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) { 280 drm_dbg(&i915->drm, 281 "You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n", 282 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450); 283 rc6vids &= 0xffff00; 284 rc6vids |= GEN6_ENCODE_RC6_VID(450); 285 ret = snb_pcode_write(i915, GEN6_PCODE_WRITE_RC6VIDS, rc6vids); 286 if (ret) 287 drm_err(&i915->drm, 288 "Couldn't fix incorrect rc6 voltage\n"); 289 } 290 } 291 292 /* Check that the pcbr address is not empty. */ 293 static int chv_rc6_init(struct intel_rc6 *rc6) 294 { 295 struct intel_uncore *uncore = rc6_to_uncore(rc6); 296 struct drm_i915_private *i915 = rc6_to_i915(rc6); 297 resource_size_t pctx_paddr, paddr; 298 resource_size_t pctx_size = 32 * SZ_1K; 299 u32 pcbr; 300 301 pcbr = intel_uncore_read(uncore, VLV_PCBR); 302 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) { 303 drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n"); 304 paddr = i915->dsm.end + 1 - pctx_size; 305 GEM_BUG_ON(paddr > U32_MAX); 306 307 pctx_paddr = (paddr & ~4095); 308 intel_uncore_write(uncore, VLV_PCBR, pctx_paddr); 309 } 310 311 return 0; 312 } 313 314 static int vlv_rc6_init(struct intel_rc6 *rc6) 315 { 316 struct drm_i915_private *i915 = rc6_to_i915(rc6); 317 struct intel_uncore *uncore = rc6_to_uncore(rc6); 318 struct drm_i915_gem_object *pctx; 319 resource_size_t pctx_paddr; 320 resource_size_t pctx_size = 24 * SZ_1K; 321 u32 pcbr; 322 323 pcbr = intel_uncore_read(uncore, VLV_PCBR); 324 if (pcbr) { 325 /* BIOS set it up already, grab the pre-alloc'd space */ 326 resource_size_t pcbr_offset; 327 328 pcbr_offset = (pcbr & ~4095) - i915->dsm.start; 329 pctx = i915_gem_object_create_region_at(i915->mm.stolen_region, 330 pcbr_offset, 331 pctx_size, 332 0); 333 if (IS_ERR(pctx)) 334 return PTR_ERR(pctx); 335 336 goto out; 337 } 338 339 drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n"); 340 341 /* 342 * From the Gunit register HAS: 343 * The Gfx driver is expected to program this register and ensure 344 * proper allocation within Gfx stolen memory. For example, this 345 * register should be programmed such than the PCBR range does not 346 * overlap with other ranges, such as the frame buffer, protected 347 * memory, or any other relevant ranges. 348 */ 349 pctx = i915_gem_object_create_stolen(i915, pctx_size); 350 if (IS_ERR(pctx)) { 351 drm_dbg(&i915->drm, 352 "not enough stolen space for PCTX, disabling\n"); 353 return PTR_ERR(pctx); 354 } 355 356 GEM_BUG_ON(range_overflows_end_t(u64, 357 i915->dsm.start, 358 pctx->stolen->start, 359 U32_MAX)); 360 pctx_paddr = i915->dsm.start + pctx->stolen->start; 361 intel_uncore_write(uncore, VLV_PCBR, pctx_paddr); 362 363 out: 364 rc6->pctx = pctx; 365 return 0; 366 } 367 368 static void chv_rc6_enable(struct intel_rc6 *rc6) 369 { 370 struct intel_uncore *uncore = rc6_to_uncore(rc6); 371 struct intel_engine_cs *engine; 372 enum intel_engine_id id; 373 374 /* 2a: Program RC6 thresholds.*/ 375 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16); 376 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */ 377 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */ 378 379 for_each_engine(engine, rc6_to_gt(rc6), id) 380 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 381 set(uncore, GEN6_RC_SLEEP, 0); 382 383 /* TO threshold set to 500 us (0x186 * 1.28 us) */ 384 set(uncore, GEN6_RC6_THRESHOLD, 0x186); 385 386 /* Allows RC6 residency counter to work */ 387 set(uncore, VLV_COUNTER_CONTROL, 388 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH | 389 VLV_MEDIA_RC6_COUNT_EN | 390 VLV_RENDER_RC6_COUNT_EN)); 391 392 /* 3: Enable RC6 */ 393 rc6->ctl_enable = GEN7_RC_CTL_TO_MODE; 394 } 395 396 static void vlv_rc6_enable(struct intel_rc6 *rc6) 397 { 398 struct intel_uncore *uncore = rc6_to_uncore(rc6); 399 struct intel_engine_cs *engine; 400 enum intel_engine_id id; 401 402 set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000); 403 set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); 404 set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); 405 406 for_each_engine(engine, rc6_to_gt(rc6), id) 407 set(uncore, RING_MAX_IDLE(engine->mmio_base), 10); 408 409 set(uncore, GEN6_RC6_THRESHOLD, 0x557); 410 411 /* Allows RC6 residency counter to work */ 412 set(uncore, VLV_COUNTER_CONTROL, 413 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH | 414 VLV_MEDIA_RC0_COUNT_EN | 415 VLV_RENDER_RC0_COUNT_EN | 416 VLV_MEDIA_RC6_COUNT_EN | 417 VLV_RENDER_RC6_COUNT_EN)); 418 419 rc6->ctl_enable = 420 GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL; 421 } 422 423 static bool bxt_check_bios_rc6_setup(struct intel_rc6 *rc6) 424 { 425 struct intel_uncore *uncore = rc6_to_uncore(rc6); 426 struct drm_i915_private *i915 = rc6_to_i915(rc6); 427 u32 rc6_ctx_base, rc_ctl, rc_sw_target; 428 bool enable_rc6 = true; 429 430 rc_ctl = intel_uncore_read(uncore, GEN6_RC_CONTROL); 431 rc_sw_target = intel_uncore_read(uncore, GEN6_RC_STATE); 432 rc_sw_target &= RC_SW_TARGET_STATE_MASK; 433 rc_sw_target >>= RC_SW_TARGET_STATE_SHIFT; 434 drm_dbg(&i915->drm, "BIOS enabled RC states: " 435 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n", 436 str_on_off(rc_ctl & GEN6_RC_CTL_HW_ENABLE), 437 str_on_off(rc_ctl & GEN6_RC_CTL_RC6_ENABLE), 438 rc_sw_target); 439 440 if (!(intel_uncore_read(uncore, RC6_LOCATION) & RC6_CTX_IN_DRAM)) { 441 drm_dbg(&i915->drm, "RC6 Base location not set properly.\n"); 442 enable_rc6 = false; 443 } 444 445 /* 446 * The exact context size is not known for BXT, so assume a page size 447 * for this check. 448 */ 449 rc6_ctx_base = 450 intel_uncore_read(uncore, RC6_CTX_BASE) & RC6_CTX_BASE_MASK; 451 if (!(rc6_ctx_base >= i915->dsm_reserved.start && 452 rc6_ctx_base + PAGE_SIZE < i915->dsm_reserved.end)) { 453 drm_dbg(&i915->drm, "RC6 Base address not as expected.\n"); 454 enable_rc6 = false; 455 } 456 457 if (!((intel_uncore_read(uncore, PWRCTX_MAXCNT(RENDER_RING_BASE)) & IDLE_TIME_MASK) > 1 && 458 (intel_uncore_read(uncore, PWRCTX_MAXCNT(GEN6_BSD_RING_BASE)) & IDLE_TIME_MASK) > 1 && 459 (intel_uncore_read(uncore, PWRCTX_MAXCNT(BLT_RING_BASE)) & IDLE_TIME_MASK) > 1 && 460 (intel_uncore_read(uncore, PWRCTX_MAXCNT(VEBOX_RING_BASE)) & IDLE_TIME_MASK) > 1)) { 461 drm_dbg(&i915->drm, 462 "Engine Idle wait time not set properly.\n"); 463 enable_rc6 = false; 464 } 465 466 if (!intel_uncore_read(uncore, GEN8_PUSHBUS_CONTROL) || 467 !intel_uncore_read(uncore, GEN8_PUSHBUS_ENABLE) || 468 !intel_uncore_read(uncore, GEN8_PUSHBUS_SHIFT)) { 469 drm_dbg(&i915->drm, "Pushbus not setup properly.\n"); 470 enable_rc6 = false; 471 } 472 473 if (!intel_uncore_read(uncore, GEN6_GFXPAUSE)) { 474 drm_dbg(&i915->drm, "GFX pause not setup properly.\n"); 475 enable_rc6 = false; 476 } 477 478 if (!intel_uncore_read(uncore, GEN8_MISC_CTRL0)) { 479 drm_dbg(&i915->drm, "GPM control not setup properly.\n"); 480 enable_rc6 = false; 481 } 482 483 return enable_rc6; 484 } 485 486 static bool rc6_supported(struct intel_rc6 *rc6) 487 { 488 struct drm_i915_private *i915 = rc6_to_i915(rc6); 489 490 if (!HAS_RC6(i915)) 491 return false; 492 493 if (intel_vgpu_active(i915)) 494 return false; 495 496 if (is_mock_gt(rc6_to_gt(rc6))) 497 return false; 498 499 if (IS_GEN9_LP(i915) && !bxt_check_bios_rc6_setup(rc6)) { 500 drm_notice(&i915->drm, 501 "RC6 and powersaving disabled by BIOS\n"); 502 return false; 503 } 504 505 return true; 506 } 507 508 static void rpm_get(struct intel_rc6 *rc6) 509 { 510 GEM_BUG_ON(rc6->wakeref); 511 pm_runtime_get_sync(rc6_to_i915(rc6)->drm.dev); 512 rc6->wakeref = true; 513 } 514 515 static void rpm_put(struct intel_rc6 *rc6) 516 { 517 GEM_BUG_ON(!rc6->wakeref); 518 pm_runtime_put(rc6_to_i915(rc6)->drm.dev); 519 rc6->wakeref = false; 520 } 521 522 static bool pctx_corrupted(struct intel_rc6 *rc6) 523 { 524 struct drm_i915_private *i915 = rc6_to_i915(rc6); 525 526 if (!NEEDS_RC6_CTX_CORRUPTION_WA(i915)) 527 return false; 528 529 if (intel_uncore_read(rc6_to_uncore(rc6), GEN8_RC6_CTX_INFO)) 530 return false; 531 532 drm_notice(&i915->drm, 533 "RC6 context corruption, disabling runtime power management\n"); 534 return true; 535 } 536 537 static void __intel_rc6_disable(struct intel_rc6 *rc6) 538 { 539 struct drm_i915_private *i915 = rc6_to_i915(rc6); 540 struct intel_uncore *uncore = rc6_to_uncore(rc6); 541 struct intel_gt *gt = rc6_to_gt(rc6); 542 543 /* Take control of RC6 back from GuC */ 544 intel_guc_rc_disable(>->uc.guc); 545 546 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 547 if (GRAPHICS_VER(i915) >= 9) 548 set(uncore, GEN9_PG_ENABLE, 0); 549 set(uncore, GEN6_RC_CONTROL, 0); 550 set(uncore, GEN6_RC_STATE, 0); 551 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 552 } 553 554 void intel_rc6_init(struct intel_rc6 *rc6) 555 { 556 struct drm_i915_private *i915 = rc6_to_i915(rc6); 557 int err; 558 559 /* Disable runtime-pm until we can save the GPU state with rc6 pctx */ 560 rpm_get(rc6); 561 562 if (!rc6_supported(rc6)) 563 return; 564 565 if (IS_CHERRYVIEW(i915)) 566 err = chv_rc6_init(rc6); 567 else if (IS_VALLEYVIEW(i915)) 568 err = vlv_rc6_init(rc6); 569 else 570 err = 0; 571 572 /* Sanitize rc6, ensure it is disabled before we are ready. */ 573 __intel_rc6_disable(rc6); 574 575 rc6->supported = err == 0; 576 } 577 578 void intel_rc6_sanitize(struct intel_rc6 *rc6) 579 { 580 memset(rc6->prev_hw_residency, 0, sizeof(rc6->prev_hw_residency)); 581 582 if (rc6->enabled) { /* unbalanced suspend/resume */ 583 rpm_get(rc6); 584 rc6->enabled = false; 585 } 586 587 if (rc6->supported) 588 __intel_rc6_disable(rc6); 589 } 590 591 void intel_rc6_enable(struct intel_rc6 *rc6) 592 { 593 struct drm_i915_private *i915 = rc6_to_i915(rc6); 594 struct intel_uncore *uncore = rc6_to_uncore(rc6); 595 596 if (!rc6->supported) 597 return; 598 599 GEM_BUG_ON(rc6->enabled); 600 601 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL); 602 603 if (IS_CHERRYVIEW(i915)) 604 chv_rc6_enable(rc6); 605 else if (IS_VALLEYVIEW(i915)) 606 vlv_rc6_enable(rc6); 607 else if (GRAPHICS_VER(i915) >= 11) 608 gen11_rc6_enable(rc6); 609 else if (GRAPHICS_VER(i915) >= 9) 610 gen9_rc6_enable(rc6); 611 else if (IS_BROADWELL(i915)) 612 gen8_rc6_enable(rc6); 613 else if (GRAPHICS_VER(i915) >= 6) 614 gen6_rc6_enable(rc6); 615 616 rc6->manual = rc6->ctl_enable & GEN6_RC_CTL_RC6_ENABLE; 617 if (NEEDS_RC6_CTX_CORRUPTION_WA(i915)) 618 rc6->ctl_enable = 0; 619 620 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL); 621 622 if (unlikely(pctx_corrupted(rc6))) 623 return; 624 625 /* rc6 is ready, runtime-pm is go! */ 626 rpm_put(rc6); 627 rc6->enabled = true; 628 } 629 630 void intel_rc6_unpark(struct intel_rc6 *rc6) 631 { 632 struct intel_uncore *uncore = rc6_to_uncore(rc6); 633 634 if (!rc6->enabled) 635 return; 636 637 /* Restore HW timers for automatic RC6 entry while busy */ 638 set(uncore, GEN6_RC_CONTROL, rc6->ctl_enable); 639 } 640 641 void intel_rc6_park(struct intel_rc6 *rc6) 642 { 643 struct intel_uncore *uncore = rc6_to_uncore(rc6); 644 unsigned int target; 645 646 if (!rc6->enabled) 647 return; 648 649 if (unlikely(pctx_corrupted(rc6))) { 650 intel_rc6_disable(rc6); 651 return; 652 } 653 654 if (!rc6->manual) 655 return; 656 657 /* Turn off the HW timers and go directly to rc6 */ 658 set(uncore, GEN6_RC_CONTROL, GEN6_RC_CTL_RC6_ENABLE); 659 660 if (HAS_RC6pp(rc6_to_i915(rc6))) 661 target = 0x6; /* deepest rc6 */ 662 else if (HAS_RC6p(rc6_to_i915(rc6))) 663 target = 0x5; /* deep rc6 */ 664 else 665 target = 0x4; /* normal rc6 */ 666 set(uncore, GEN6_RC_STATE, target << RC_SW_TARGET_STATE_SHIFT); 667 } 668 669 void intel_rc6_disable(struct intel_rc6 *rc6) 670 { 671 if (!rc6->enabled) 672 return; 673 674 rpm_get(rc6); 675 rc6->enabled = false; 676 677 __intel_rc6_disable(rc6); 678 } 679 680 void intel_rc6_fini(struct intel_rc6 *rc6) 681 { 682 struct drm_i915_gem_object *pctx; 683 684 intel_rc6_disable(rc6); 685 686 pctx = fetch_and_zero(&rc6->pctx); 687 if (pctx) 688 i915_gem_object_put(pctx); 689 690 if (rc6->wakeref) 691 rpm_put(rc6); 692 } 693 694 static u64 vlv_residency_raw(struct intel_uncore *uncore, const i915_reg_t reg) 695 { 696 u32 lower, upper, tmp; 697 int loop = 2; 698 699 /* 700 * The register accessed do not need forcewake. We borrow 701 * uncore lock to prevent concurrent access to range reg. 702 */ 703 lockdep_assert_held(&uncore->lock); 704 705 /* 706 * vlv and chv residency counters are 40 bits in width. 707 * With a control bit, we can choose between upper or lower 708 * 32bit window into this counter. 709 * 710 * Although we always use the counter in high-range mode elsewhere, 711 * userspace may attempt to read the value before rc6 is initialised, 712 * before we have set the default VLV_COUNTER_CONTROL value. So always 713 * set the high bit to be safe. 714 */ 715 set(uncore, VLV_COUNTER_CONTROL, 716 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH)); 717 upper = intel_uncore_read_fw(uncore, reg); 718 do { 719 tmp = upper; 720 721 set(uncore, VLV_COUNTER_CONTROL, 722 _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH)); 723 lower = intel_uncore_read_fw(uncore, reg); 724 725 set(uncore, VLV_COUNTER_CONTROL, 726 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH)); 727 upper = intel_uncore_read_fw(uncore, reg); 728 } while (upper != tmp && --loop); 729 730 /* 731 * Everywhere else we always use VLV_COUNTER_CONTROL with the 732 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set 733 * now. 734 */ 735 736 return lower | (u64)upper << 8; 737 } 738 739 u64 intel_rc6_residency_ns(struct intel_rc6 *rc6, const i915_reg_t reg) 740 { 741 struct drm_i915_private *i915 = rc6_to_i915(rc6); 742 struct intel_uncore *uncore = rc6_to_uncore(rc6); 743 u64 time_hw, prev_hw, overflow_hw; 744 unsigned int fw_domains; 745 unsigned long flags; 746 unsigned int i; 747 u32 mul, div; 748 749 if (!rc6->supported) 750 return 0; 751 752 /* 753 * Store previous hw counter values for counter wrap-around handling. 754 * 755 * There are only four interesting registers and they live next to each 756 * other so we can use the relative address, compared to the smallest 757 * one as the index into driver storage. 758 */ 759 i = (i915_mmio_reg_offset(reg) - 760 i915_mmio_reg_offset(GEN6_GT_GFX_RC6_LOCKED)) / sizeof(u32); 761 if (drm_WARN_ON_ONCE(&i915->drm, i >= ARRAY_SIZE(rc6->cur_residency))) 762 return 0; 763 764 fw_domains = intel_uncore_forcewake_for_reg(uncore, reg, FW_REG_READ); 765 766 spin_lock_irqsave(&uncore->lock, flags); 767 intel_uncore_forcewake_get__locked(uncore, fw_domains); 768 769 /* On VLV and CHV, residency time is in CZ units rather than 1.28us */ 770 if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) { 771 mul = 1000000; 772 div = i915->czclk_freq; 773 overflow_hw = BIT_ULL(40); 774 time_hw = vlv_residency_raw(uncore, reg); 775 } else { 776 /* 833.33ns units on Gen9LP, 1.28us elsewhere. */ 777 if (IS_GEN9_LP(i915)) { 778 mul = 10000; 779 div = 12; 780 } else { 781 mul = 1280; 782 div = 1; 783 } 784 785 overflow_hw = BIT_ULL(32); 786 time_hw = intel_uncore_read_fw(uncore, reg); 787 } 788 789 /* 790 * Counter wrap handling. 791 * 792 * But relying on a sufficient frequency of queries otherwise counters 793 * can still wrap. 794 */ 795 prev_hw = rc6->prev_hw_residency[i]; 796 rc6->prev_hw_residency[i] = time_hw; 797 798 /* RC6 delta from last sample. */ 799 if (time_hw >= prev_hw) 800 time_hw -= prev_hw; 801 else 802 time_hw += overflow_hw - prev_hw; 803 804 /* Add delta to RC6 extended raw driver copy. */ 805 time_hw += rc6->cur_residency[i]; 806 rc6->cur_residency[i] = time_hw; 807 808 intel_uncore_forcewake_put__locked(uncore, fw_domains); 809 spin_unlock_irqrestore(&uncore->lock, flags); 810 811 return mul_u64_u32_div(time_hw, mul, div); 812 } 813 814 u64 intel_rc6_residency_us(struct intel_rc6 *rc6, i915_reg_t reg) 815 { 816 return DIV_ROUND_UP_ULL(intel_rc6_residency_ns(rc6, reg), 1000); 817 } 818 819 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 820 #include "selftest_rc6.c" 821 #endif 822