xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_rc6.c (revision 63705da3)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <linux/pm_runtime.h>
7 
8 #include "i915_drv.h"
9 #include "i915_vgpu.h"
10 #include "intel_gt.h"
11 #include "intel_gt_pm.h"
12 #include "intel_pcode.h"
13 #include "intel_rc6.h"
14 
15 /**
16  * DOC: RC6
17  *
18  * RC6 is a special power stage which allows the GPU to enter an very
19  * low-voltage mode when idle, using down to 0V while at this stage.  This
20  * stage is entered automatically when the GPU is idle when RC6 support is
21  * enabled, and as soon as new workload arises GPU wakes up automatically as
22  * well.
23  *
24  * There are different RC6 modes available in Intel GPU, which differentiate
25  * among each other with the latency required to enter and leave RC6 and
26  * voltage consumed by the GPU in different states.
27  *
28  * The combination of the following flags define which states GPU is allowed
29  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
30  * RC6pp is deepest RC6. Their support by hardware varies according to the
31  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
32  * which brings the most power savings; deeper states save more power, but
33  * require higher latency to switch to and wake up.
34  */
35 
36 static struct intel_gt *rc6_to_gt(struct intel_rc6 *rc6)
37 {
38 	return container_of(rc6, struct intel_gt, rc6);
39 }
40 
41 static struct intel_uncore *rc6_to_uncore(struct intel_rc6 *rc)
42 {
43 	return rc6_to_gt(rc)->uncore;
44 }
45 
46 static struct drm_i915_private *rc6_to_i915(struct intel_rc6 *rc)
47 {
48 	return rc6_to_gt(rc)->i915;
49 }
50 
51 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val)
52 {
53 	intel_uncore_write_fw(uncore, reg, val);
54 }
55 
56 static void gen11_rc6_enable(struct intel_rc6 *rc6)
57 {
58 	struct intel_gt *gt = rc6_to_gt(rc6);
59 	struct intel_uncore *uncore = gt->uncore;
60 	struct intel_engine_cs *engine;
61 	enum intel_engine_id id;
62 	u32 pg_enable;
63 	int i;
64 
65 	/*
66 	 * With GuCRC, these parameters are set by GuC
67 	 */
68 	if (!intel_uc_uses_guc_rc(&gt->uc)) {
69 		/* 2b: Program RC6 thresholds.*/
70 		set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85);
71 		set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150);
72 
73 		set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
74 		set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
75 		for_each_engine(engine, rc6_to_gt(rc6), id)
76 			set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
77 
78 		set(uncore, GUC_MAX_IDLE_COUNT, 0xA);
79 
80 		set(uncore, GEN6_RC_SLEEP, 0);
81 
82 		set(uncore, GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
83 	}
84 
85 	/*
86 	 * 2c: Program Coarse Power Gating Policies.
87 	 *
88 	 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we
89 	 * use instead is a more conservative estimate for the maximum time
90 	 * it takes us to service a CS interrupt and submit a new ELSP - that
91 	 * is the time which the GPU is idle waiting for the CPU to select the
92 	 * next request to execute. If the idle hysteresis is less than that
93 	 * interrupt service latency, the hardware will automatically gate
94 	 * the power well and we will then incur the wake up cost on top of
95 	 * the service latency. A similar guide from plane_state is that we
96 	 * do not want the enable hysteresis to less than the wakeup latency.
97 	 *
98 	 * igt/gem_exec_nop/sequential provides a rough estimate for the
99 	 * service latency, and puts it under 10us for Icelake, similar to
100 	 * Broadwell+, To be conservative, we want to factor in a context
101 	 * switch on top (due to ksoftirqd).
102 	 */
103 	set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 60);
104 	set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 60);
105 
106 	/* 3a: Enable RC6
107 	 *
108 	 * With GuCRC, we do not enable bit 31 of RC_CTL,
109 	 * thus allowing GuC to control RC6 entry/exit fully instead.
110 	 * We will not set the HW ENABLE and EI bits
111 	 */
112 	if (!intel_guc_rc_enable(&gt->uc.guc))
113 		rc6->ctl_enable = GEN6_RC_CTL_RC6_ENABLE;
114 	else
115 		rc6->ctl_enable =
116 			GEN6_RC_CTL_HW_ENABLE |
117 			GEN6_RC_CTL_RC6_ENABLE |
118 			GEN6_RC_CTL_EI_MODE(1);
119 
120 	pg_enable =
121 		GEN9_RENDER_PG_ENABLE |
122 		GEN9_MEDIA_PG_ENABLE |
123 		GEN11_MEDIA_SAMPLER_PG_ENABLE;
124 
125 	if (GRAPHICS_VER(gt->i915) >= 12) {
126 		for (i = 0; i < I915_MAX_VCS; i++)
127 			if (HAS_ENGINE(gt, _VCS(i)))
128 				pg_enable |= (VDN_HCP_POWERGATE_ENABLE(i) |
129 					      VDN_MFX_POWERGATE_ENABLE(i));
130 	}
131 
132 	set(uncore, GEN9_PG_ENABLE, pg_enable);
133 }
134 
135 static void gen9_rc6_enable(struct intel_rc6 *rc6)
136 {
137 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
138 	struct intel_engine_cs *engine;
139 	enum intel_engine_id id;
140 
141 	/* 2b: Program RC6 thresholds.*/
142 	if (GRAPHICS_VER(rc6_to_i915(rc6)) >= 11) {
143 		set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85);
144 		set(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150);
145 	} else if (IS_SKYLAKE(rc6_to_i915(rc6))) {
146 		/*
147 		 * WaRsDoubleRc6WrlWithCoarsePowerGating:skl Doubling WRL only
148 		 * when CPG is enabled
149 		 */
150 		set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
151 	} else {
152 		set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
153 	}
154 
155 	set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
156 	set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
157 	for_each_engine(engine, rc6_to_gt(rc6), id)
158 		set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
159 
160 	set(uncore, GUC_MAX_IDLE_COUNT, 0xA);
161 
162 	set(uncore, GEN6_RC_SLEEP, 0);
163 
164 	/*
165 	 * 2c: Program Coarse Power Gating Policies.
166 	 *
167 	 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we
168 	 * use instead is a more conservative estimate for the maximum time
169 	 * it takes us to service a CS interrupt and submit a new ELSP - that
170 	 * is the time which the GPU is idle waiting for the CPU to select the
171 	 * next request to execute. If the idle hysteresis is less than that
172 	 * interrupt service latency, the hardware will automatically gate
173 	 * the power well and we will then incur the wake up cost on top of
174 	 * the service latency. A similar guide from plane_state is that we
175 	 * do not want the enable hysteresis to less than the wakeup latency.
176 	 *
177 	 * igt/gem_exec_nop/sequential provides a rough estimate for the
178 	 * service latency, and puts it around 10us for Broadwell (and other
179 	 * big core) and around 40us for Broxton (and other low power cores).
180 	 * [Note that for legacy ringbuffer submission, this is less than 1us!]
181 	 * However, the wakeup latency on Broxton is closer to 100us. To be
182 	 * conservative, we have to factor in a context switch on top (due
183 	 * to ksoftirqd).
184 	 */
185 	set(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 250);
186 	set(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 250);
187 
188 	/* 3a: Enable RC6 */
189 	set(uncore, GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
190 
191 	rc6->ctl_enable =
192 		GEN6_RC_CTL_HW_ENABLE |
193 		GEN6_RC_CTL_RC6_ENABLE |
194 		GEN6_RC_CTL_EI_MODE(1);
195 
196 	/*
197 	 * WaRsDisableCoarsePowerGating:skl,cnl
198 	 *   - Render/Media PG need to be disabled with RC6.
199 	 */
200 	if (!NEEDS_WaRsDisableCoarsePowerGating(rc6_to_i915(rc6)))
201 		set(uncore, GEN9_PG_ENABLE,
202 		    GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE);
203 }
204 
205 static void gen8_rc6_enable(struct intel_rc6 *rc6)
206 {
207 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
208 	struct intel_engine_cs *engine;
209 	enum intel_engine_id id;
210 
211 	/* 2b: Program RC6 thresholds.*/
212 	set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
213 	set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
214 	set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
215 	for_each_engine(engine, rc6_to_gt(rc6), id)
216 		set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
217 	set(uncore, GEN6_RC_SLEEP, 0);
218 	set(uncore, GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
219 
220 	/* 3: Enable RC6 */
221 	rc6->ctl_enable =
222 	    GEN6_RC_CTL_HW_ENABLE |
223 	    GEN7_RC_CTL_TO_MODE |
224 	    GEN6_RC_CTL_RC6_ENABLE;
225 }
226 
227 static void gen6_rc6_enable(struct intel_rc6 *rc6)
228 {
229 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
230 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
231 	struct intel_engine_cs *engine;
232 	enum intel_engine_id id;
233 	u32 rc6vids, rc6_mask;
234 	int ret;
235 
236 	set(uncore, GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
237 	set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
238 	set(uncore, GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
239 	set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000);
240 	set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25);
241 
242 	for_each_engine(engine, rc6_to_gt(rc6), id)
243 		set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
244 
245 	set(uncore, GEN6_RC_SLEEP, 0);
246 	set(uncore, GEN6_RC1e_THRESHOLD, 1000);
247 	set(uncore, GEN6_RC6_THRESHOLD, 50000);
248 	set(uncore, GEN6_RC6p_THRESHOLD, 150000);
249 	set(uncore, GEN6_RC6pp_THRESHOLD, 64000); /* unused */
250 
251 	/* We don't use those on Haswell */
252 	rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
253 	if (HAS_RC6p(i915))
254 		rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
255 	if (HAS_RC6pp(i915))
256 		rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
257 	rc6->ctl_enable =
258 	    rc6_mask |
259 	    GEN6_RC_CTL_EI_MODE(1) |
260 	    GEN6_RC_CTL_HW_ENABLE;
261 
262 	rc6vids = 0;
263 	ret = sandybridge_pcode_read(i915, GEN6_PCODE_READ_RC6VIDS,
264 				     &rc6vids, NULL);
265 	if (GRAPHICS_VER(i915) == 6 && ret) {
266 		drm_dbg(&i915->drm, "Couldn't check for BIOS workaround\n");
267 	} else if (GRAPHICS_VER(i915) == 6 &&
268 		   (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
269 		drm_dbg(&i915->drm,
270 			"You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
271 			GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
272 		rc6vids &= 0xffff00;
273 		rc6vids |= GEN6_ENCODE_RC6_VID(450);
274 		ret = sandybridge_pcode_write(i915, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
275 		if (ret)
276 			drm_err(&i915->drm,
277 				"Couldn't fix incorrect rc6 voltage\n");
278 	}
279 }
280 
281 /* Check that the pcbr address is not empty. */
282 static int chv_rc6_init(struct intel_rc6 *rc6)
283 {
284 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
285 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
286 	resource_size_t pctx_paddr, paddr;
287 	resource_size_t pctx_size = 32 * SZ_1K;
288 	u32 pcbr;
289 
290 	pcbr = intel_uncore_read(uncore, VLV_PCBR);
291 	if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
292 		drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n");
293 		paddr = i915->dsm.end + 1 - pctx_size;
294 		GEM_BUG_ON(paddr > U32_MAX);
295 
296 		pctx_paddr = (paddr & ~4095);
297 		intel_uncore_write(uncore, VLV_PCBR, pctx_paddr);
298 	}
299 
300 	return 0;
301 }
302 
303 static int vlv_rc6_init(struct intel_rc6 *rc6)
304 {
305 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
306 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
307 	struct drm_i915_gem_object *pctx;
308 	resource_size_t pctx_paddr;
309 	resource_size_t pctx_size = 24 * SZ_1K;
310 	u32 pcbr;
311 
312 	pcbr = intel_uncore_read(uncore, VLV_PCBR);
313 	if (pcbr) {
314 		/* BIOS set it up already, grab the pre-alloc'd space */
315 		resource_size_t pcbr_offset;
316 
317 		pcbr_offset = (pcbr & ~4095) - i915->dsm.start;
318 		pctx = i915_gem_object_create_stolen_for_preallocated(i915,
319 								      pcbr_offset,
320 								      pctx_size);
321 		if (IS_ERR(pctx))
322 			return PTR_ERR(pctx);
323 
324 		goto out;
325 	}
326 
327 	drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n");
328 
329 	/*
330 	 * From the Gunit register HAS:
331 	 * The Gfx driver is expected to program this register and ensure
332 	 * proper allocation within Gfx stolen memory.  For example, this
333 	 * register should be programmed such than the PCBR range does not
334 	 * overlap with other ranges, such as the frame buffer, protected
335 	 * memory, or any other relevant ranges.
336 	 */
337 	pctx = i915_gem_object_create_stolen(i915, pctx_size);
338 	if (IS_ERR(pctx)) {
339 		drm_dbg(&i915->drm,
340 			"not enough stolen space for PCTX, disabling\n");
341 		return PTR_ERR(pctx);
342 	}
343 
344 	GEM_BUG_ON(range_overflows_end_t(u64,
345 					 i915->dsm.start,
346 					 pctx->stolen->start,
347 					 U32_MAX));
348 	pctx_paddr = i915->dsm.start + pctx->stolen->start;
349 	intel_uncore_write(uncore, VLV_PCBR, pctx_paddr);
350 
351 out:
352 	rc6->pctx = pctx;
353 	return 0;
354 }
355 
356 static void chv_rc6_enable(struct intel_rc6 *rc6)
357 {
358 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
359 	struct intel_engine_cs *engine;
360 	enum intel_engine_id id;
361 
362 	/* 2a: Program RC6 thresholds.*/
363 	set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
364 	set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
365 	set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
366 
367 	for_each_engine(engine, rc6_to_gt(rc6), id)
368 		set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
369 	set(uncore, GEN6_RC_SLEEP, 0);
370 
371 	/* TO threshold set to 500 us (0x186 * 1.28 us) */
372 	set(uncore, GEN6_RC6_THRESHOLD, 0x186);
373 
374 	/* Allows RC6 residency counter to work */
375 	set(uncore, VLV_COUNTER_CONTROL,
376 	    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
377 			       VLV_MEDIA_RC6_COUNT_EN |
378 			       VLV_RENDER_RC6_COUNT_EN));
379 
380 	/* 3: Enable RC6 */
381 	rc6->ctl_enable = GEN7_RC_CTL_TO_MODE;
382 }
383 
384 static void vlv_rc6_enable(struct intel_rc6 *rc6)
385 {
386 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
387 	struct intel_engine_cs *engine;
388 	enum intel_engine_id id;
389 
390 	set(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
391 	set(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000);
392 	set(uncore, GEN6_RC_IDLE_HYSTERSIS, 25);
393 
394 	for_each_engine(engine, rc6_to_gt(rc6), id)
395 		set(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
396 
397 	set(uncore, GEN6_RC6_THRESHOLD, 0x557);
398 
399 	/* Allows RC6 residency counter to work */
400 	set(uncore, VLV_COUNTER_CONTROL,
401 	    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
402 			       VLV_MEDIA_RC0_COUNT_EN |
403 			       VLV_RENDER_RC0_COUNT_EN |
404 			       VLV_MEDIA_RC6_COUNT_EN |
405 			       VLV_RENDER_RC6_COUNT_EN));
406 
407 	rc6->ctl_enable =
408 	    GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
409 }
410 
411 static bool bxt_check_bios_rc6_setup(struct intel_rc6 *rc6)
412 {
413 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
414 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
415 	u32 rc6_ctx_base, rc_ctl, rc_sw_target;
416 	bool enable_rc6 = true;
417 
418 	rc_ctl = intel_uncore_read(uncore, GEN6_RC_CONTROL);
419 	rc_sw_target = intel_uncore_read(uncore, GEN6_RC_STATE);
420 	rc_sw_target &= RC_SW_TARGET_STATE_MASK;
421 	rc_sw_target >>= RC_SW_TARGET_STATE_SHIFT;
422 	drm_dbg(&i915->drm, "BIOS enabled RC states: "
423 			 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
424 			 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
425 			 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
426 			 rc_sw_target);
427 
428 	if (!(intel_uncore_read(uncore, RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
429 		drm_dbg(&i915->drm, "RC6 Base location not set properly.\n");
430 		enable_rc6 = false;
431 	}
432 
433 	/*
434 	 * The exact context size is not known for BXT, so assume a page size
435 	 * for this check.
436 	 */
437 	rc6_ctx_base =
438 		intel_uncore_read(uncore, RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
439 	if (!(rc6_ctx_base >= i915->dsm_reserved.start &&
440 	      rc6_ctx_base + PAGE_SIZE < i915->dsm_reserved.end)) {
441 		drm_dbg(&i915->drm, "RC6 Base address not as expected.\n");
442 		enable_rc6 = false;
443 	}
444 
445 	if (!((intel_uncore_read(uncore, PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1 &&
446 	      (intel_uncore_read(uncore, PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1 &&
447 	      (intel_uncore_read(uncore, PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1 &&
448 	      (intel_uncore_read(uncore, PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1)) {
449 		drm_dbg(&i915->drm,
450 			"Engine Idle wait time not set properly.\n");
451 		enable_rc6 = false;
452 	}
453 
454 	if (!intel_uncore_read(uncore, GEN8_PUSHBUS_CONTROL) ||
455 	    !intel_uncore_read(uncore, GEN8_PUSHBUS_ENABLE) ||
456 	    !intel_uncore_read(uncore, GEN8_PUSHBUS_SHIFT)) {
457 		drm_dbg(&i915->drm, "Pushbus not setup properly.\n");
458 		enable_rc6 = false;
459 	}
460 
461 	if (!intel_uncore_read(uncore, GEN6_GFXPAUSE)) {
462 		drm_dbg(&i915->drm, "GFX pause not setup properly.\n");
463 		enable_rc6 = false;
464 	}
465 
466 	if (!intel_uncore_read(uncore, GEN8_MISC_CTRL0)) {
467 		drm_dbg(&i915->drm, "GPM control not setup properly.\n");
468 		enable_rc6 = false;
469 	}
470 
471 	return enable_rc6;
472 }
473 
474 static bool rc6_supported(struct intel_rc6 *rc6)
475 {
476 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
477 
478 	if (!HAS_RC6(i915))
479 		return false;
480 
481 	if (intel_vgpu_active(i915))
482 		return false;
483 
484 	if (is_mock_gt(rc6_to_gt(rc6)))
485 		return false;
486 
487 	if (IS_GEN9_LP(i915) && !bxt_check_bios_rc6_setup(rc6)) {
488 		drm_notice(&i915->drm,
489 			   "RC6 and powersaving disabled by BIOS\n");
490 		return false;
491 	}
492 
493 	return true;
494 }
495 
496 static void rpm_get(struct intel_rc6 *rc6)
497 {
498 	GEM_BUG_ON(rc6->wakeref);
499 	pm_runtime_get_sync(rc6_to_i915(rc6)->drm.dev);
500 	rc6->wakeref = true;
501 }
502 
503 static void rpm_put(struct intel_rc6 *rc6)
504 {
505 	GEM_BUG_ON(!rc6->wakeref);
506 	pm_runtime_put(rc6_to_i915(rc6)->drm.dev);
507 	rc6->wakeref = false;
508 }
509 
510 static bool pctx_corrupted(struct intel_rc6 *rc6)
511 {
512 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
513 
514 	if (!NEEDS_RC6_CTX_CORRUPTION_WA(i915))
515 		return false;
516 
517 	if (intel_uncore_read(rc6_to_uncore(rc6), GEN8_RC6_CTX_INFO))
518 		return false;
519 
520 	drm_notice(&i915->drm,
521 		   "RC6 context corruption, disabling runtime power management\n");
522 	return true;
523 }
524 
525 static void __intel_rc6_disable(struct intel_rc6 *rc6)
526 {
527 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
528 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
529 	struct intel_gt *gt = rc6_to_gt(rc6);
530 
531 	/* Take control of RC6 back from GuC */
532 	intel_guc_rc_disable(&gt->uc.guc);
533 
534 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
535 	if (GRAPHICS_VER(i915) >= 9)
536 		set(uncore, GEN9_PG_ENABLE, 0);
537 	set(uncore, GEN6_RC_CONTROL, 0);
538 	set(uncore, GEN6_RC_STATE, 0);
539 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
540 }
541 
542 void intel_rc6_init(struct intel_rc6 *rc6)
543 {
544 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
545 	int err;
546 
547 	/* Disable runtime-pm until we can save the GPU state with rc6 pctx */
548 	rpm_get(rc6);
549 
550 	if (!rc6_supported(rc6))
551 		return;
552 
553 	if (IS_CHERRYVIEW(i915))
554 		err = chv_rc6_init(rc6);
555 	else if (IS_VALLEYVIEW(i915))
556 		err = vlv_rc6_init(rc6);
557 	else
558 		err = 0;
559 
560 	/* Sanitize rc6, ensure it is disabled before we are ready. */
561 	__intel_rc6_disable(rc6);
562 
563 	rc6->supported = err == 0;
564 }
565 
566 void intel_rc6_sanitize(struct intel_rc6 *rc6)
567 {
568 	memset(rc6->prev_hw_residency, 0, sizeof(rc6->prev_hw_residency));
569 
570 	if (rc6->enabled) { /* unbalanced suspend/resume */
571 		rpm_get(rc6);
572 		rc6->enabled = false;
573 	}
574 
575 	if (rc6->supported)
576 		__intel_rc6_disable(rc6);
577 }
578 
579 void intel_rc6_enable(struct intel_rc6 *rc6)
580 {
581 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
582 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
583 
584 	if (!rc6->supported)
585 		return;
586 
587 	GEM_BUG_ON(rc6->enabled);
588 
589 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
590 
591 	if (IS_CHERRYVIEW(i915))
592 		chv_rc6_enable(rc6);
593 	else if (IS_VALLEYVIEW(i915))
594 		vlv_rc6_enable(rc6);
595 	else if (GRAPHICS_VER(i915) >= 11)
596 		gen11_rc6_enable(rc6);
597 	else if (GRAPHICS_VER(i915) >= 9)
598 		gen9_rc6_enable(rc6);
599 	else if (IS_BROADWELL(i915))
600 		gen8_rc6_enable(rc6);
601 	else if (GRAPHICS_VER(i915) >= 6)
602 		gen6_rc6_enable(rc6);
603 
604 	rc6->manual = rc6->ctl_enable & GEN6_RC_CTL_RC6_ENABLE;
605 	if (NEEDS_RC6_CTX_CORRUPTION_WA(i915))
606 		rc6->ctl_enable = 0;
607 
608 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
609 
610 	if (unlikely(pctx_corrupted(rc6)))
611 		return;
612 
613 	/* rc6 is ready, runtime-pm is go! */
614 	rpm_put(rc6);
615 	rc6->enabled = true;
616 }
617 
618 void intel_rc6_unpark(struct intel_rc6 *rc6)
619 {
620 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
621 
622 	if (!rc6->enabled)
623 		return;
624 
625 	/* Restore HW timers for automatic RC6 entry while busy */
626 	set(uncore, GEN6_RC_CONTROL, rc6->ctl_enable);
627 }
628 
629 void intel_rc6_park(struct intel_rc6 *rc6)
630 {
631 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
632 	unsigned int target;
633 
634 	if (!rc6->enabled)
635 		return;
636 
637 	if (unlikely(pctx_corrupted(rc6))) {
638 		intel_rc6_disable(rc6);
639 		return;
640 	}
641 
642 	if (!rc6->manual)
643 		return;
644 
645 	/* Turn off the HW timers and go directly to rc6 */
646 	set(uncore, GEN6_RC_CONTROL, GEN6_RC_CTL_RC6_ENABLE);
647 
648 	if (HAS_RC6pp(rc6_to_i915(rc6)))
649 		target = 0x6; /* deepest rc6 */
650 	else if (HAS_RC6p(rc6_to_i915(rc6)))
651 		target = 0x5; /* deep rc6 */
652 	else
653 		target = 0x4; /* normal rc6 */
654 	set(uncore, GEN6_RC_STATE, target << RC_SW_TARGET_STATE_SHIFT);
655 }
656 
657 void intel_rc6_disable(struct intel_rc6 *rc6)
658 {
659 	if (!rc6->enabled)
660 		return;
661 
662 	rpm_get(rc6);
663 	rc6->enabled = false;
664 
665 	__intel_rc6_disable(rc6);
666 }
667 
668 void intel_rc6_fini(struct intel_rc6 *rc6)
669 {
670 	struct drm_i915_gem_object *pctx;
671 
672 	intel_rc6_disable(rc6);
673 
674 	pctx = fetch_and_zero(&rc6->pctx);
675 	if (pctx)
676 		i915_gem_object_put(pctx);
677 
678 	if (rc6->wakeref)
679 		rpm_put(rc6);
680 }
681 
682 static u64 vlv_residency_raw(struct intel_uncore *uncore, const i915_reg_t reg)
683 {
684 	u32 lower, upper, tmp;
685 	int loop = 2;
686 
687 	/*
688 	 * The register accessed do not need forcewake. We borrow
689 	 * uncore lock to prevent concurrent access to range reg.
690 	 */
691 	lockdep_assert_held(&uncore->lock);
692 
693 	/*
694 	 * vlv and chv residency counters are 40 bits in width.
695 	 * With a control bit, we can choose between upper or lower
696 	 * 32bit window into this counter.
697 	 *
698 	 * Although we always use the counter in high-range mode elsewhere,
699 	 * userspace may attempt to read the value before rc6 is initialised,
700 	 * before we have set the default VLV_COUNTER_CONTROL value. So always
701 	 * set the high bit to be safe.
702 	 */
703 	set(uncore, VLV_COUNTER_CONTROL,
704 	    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
705 	upper = intel_uncore_read_fw(uncore, reg);
706 	do {
707 		tmp = upper;
708 
709 		set(uncore, VLV_COUNTER_CONTROL,
710 		    _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH));
711 		lower = intel_uncore_read_fw(uncore, reg);
712 
713 		set(uncore, VLV_COUNTER_CONTROL,
714 		    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
715 		upper = intel_uncore_read_fw(uncore, reg);
716 	} while (upper != tmp && --loop);
717 
718 	/*
719 	 * Everywhere else we always use VLV_COUNTER_CONTROL with the
720 	 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
721 	 * now.
722 	 */
723 
724 	return lower | (u64)upper << 8;
725 }
726 
727 u64 intel_rc6_residency_ns(struct intel_rc6 *rc6, const i915_reg_t reg)
728 {
729 	struct drm_i915_private *i915 = rc6_to_i915(rc6);
730 	struct intel_uncore *uncore = rc6_to_uncore(rc6);
731 	u64 time_hw, prev_hw, overflow_hw;
732 	unsigned int fw_domains;
733 	unsigned long flags;
734 	unsigned int i;
735 	u32 mul, div;
736 
737 	if (!rc6->supported)
738 		return 0;
739 
740 	/*
741 	 * Store previous hw counter values for counter wrap-around handling.
742 	 *
743 	 * There are only four interesting registers and they live next to each
744 	 * other so we can use the relative address, compared to the smallest
745 	 * one as the index into driver storage.
746 	 */
747 	i = (i915_mmio_reg_offset(reg) -
748 	     i915_mmio_reg_offset(GEN6_GT_GFX_RC6_LOCKED)) / sizeof(u32);
749 	if (drm_WARN_ON_ONCE(&i915->drm, i >= ARRAY_SIZE(rc6->cur_residency)))
750 		return 0;
751 
752 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg, FW_REG_READ);
753 
754 	spin_lock_irqsave(&uncore->lock, flags);
755 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
756 
757 	/* On VLV and CHV, residency time is in CZ units rather than 1.28us */
758 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) {
759 		mul = 1000000;
760 		div = i915->czclk_freq;
761 		overflow_hw = BIT_ULL(40);
762 		time_hw = vlv_residency_raw(uncore, reg);
763 	} else {
764 		/* 833.33ns units on Gen9LP, 1.28us elsewhere. */
765 		if (IS_GEN9_LP(i915)) {
766 			mul = 10000;
767 			div = 12;
768 		} else {
769 			mul = 1280;
770 			div = 1;
771 		}
772 
773 		overflow_hw = BIT_ULL(32);
774 		time_hw = intel_uncore_read_fw(uncore, reg);
775 	}
776 
777 	/*
778 	 * Counter wrap handling.
779 	 *
780 	 * But relying on a sufficient frequency of queries otherwise counters
781 	 * can still wrap.
782 	 */
783 	prev_hw = rc6->prev_hw_residency[i];
784 	rc6->prev_hw_residency[i] = time_hw;
785 
786 	/* RC6 delta from last sample. */
787 	if (time_hw >= prev_hw)
788 		time_hw -= prev_hw;
789 	else
790 		time_hw += overflow_hw - prev_hw;
791 
792 	/* Add delta to RC6 extended raw driver copy. */
793 	time_hw += rc6->cur_residency[i];
794 	rc6->cur_residency[i] = time_hw;
795 
796 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
797 	spin_unlock_irqrestore(&uncore->lock, flags);
798 
799 	return mul_u64_u32_div(time_hw, mul, div);
800 }
801 
802 u64 intel_rc6_residency_us(struct intel_rc6 *rc6, i915_reg_t reg)
803 {
804 	return DIV_ROUND_UP_ULL(intel_rc6_residency_ns(rc6, reg), 1000);
805 }
806 
807 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
808 #include "selftest_rc6.c"
809 #endif
810