xref: /openbsd/sys/dev/pci/drm/i915/gt/intel_rps.c (revision d415bd75)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <linux/string_helpers.h>
7 
8 #include <drm/i915_drm.h>
9 
10 #include "i915_drv.h"
11 #include "i915_irq.h"
12 #include "intel_breadcrumbs.h"
13 #include "intel_gt.h"
14 #include "intel_gt_clock_utils.h"
15 #include "intel_gt_irq.h"
16 #include "intel_gt_pm_irq.h"
17 #include "intel_gt_regs.h"
18 #include "intel_mchbar_regs.h"
19 #include "intel_pcode.h"
20 #include "intel_rps.h"
21 #include "vlv_sideband.h"
22 #ifdef __linux__
23 #include "../../../platform/x86/intel_ips.h"
24 #endif
25 
26 #define BUSY_MAX_EI	20u /* ms */
27 
28 /*
29  * Lock protecting IPS related data structures
30  */
31 static DEFINE_SPINLOCK(mchdev_lock);
32 
33 static struct intel_gt *rps_to_gt(struct intel_rps *rps)
34 {
35 	return container_of(rps, struct intel_gt, rps);
36 }
37 
38 static struct drm_i915_private *rps_to_i915(struct intel_rps *rps)
39 {
40 	return rps_to_gt(rps)->i915;
41 }
42 
43 static struct intel_uncore *rps_to_uncore(struct intel_rps *rps)
44 {
45 	return rps_to_gt(rps)->uncore;
46 }
47 
48 static struct intel_guc_slpc *rps_to_slpc(struct intel_rps *rps)
49 {
50 	struct intel_gt *gt = rps_to_gt(rps);
51 
52 	return &gt->uc.guc.slpc;
53 }
54 
55 static bool rps_uses_slpc(struct intel_rps *rps)
56 {
57 	struct intel_gt *gt = rps_to_gt(rps);
58 
59 	return intel_uc_uses_guc_slpc(&gt->uc);
60 }
61 
62 static u32 rps_pm_sanitize_mask(struct intel_rps *rps, u32 mask)
63 {
64 	return mask & ~rps->pm_intrmsk_mbz;
65 }
66 
67 static void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val)
68 {
69 	intel_uncore_write_fw(uncore, reg, val);
70 }
71 
72 static void rps_timer(void *arg)
73 {
74 	struct intel_rps *rps = arg;
75 	struct intel_engine_cs *engine;
76 	ktime_t dt, last, timestamp;
77 	enum intel_engine_id id;
78 	s64 max_busy[3] = {};
79 
80 	timestamp = 0;
81 	for_each_engine(engine, rps_to_gt(rps), id) {
82 		s64 busy;
83 		int i;
84 
85 		dt = intel_engine_get_busy_time(engine, &timestamp);
86 		last = engine->stats.rps;
87 		engine->stats.rps = dt;
88 
89 		busy = ktime_to_ns(ktime_sub(dt, last));
90 		for (i = 0; i < ARRAY_SIZE(max_busy); i++) {
91 			if (busy > max_busy[i])
92 				swap(busy, max_busy[i]);
93 		}
94 	}
95 	last = rps->pm_timestamp;
96 	rps->pm_timestamp = timestamp;
97 
98 	if (intel_rps_is_active(rps)) {
99 		s64 busy;
100 		int i;
101 
102 		dt = ktime_sub(timestamp, last);
103 
104 		/*
105 		 * Our goal is to evaluate each engine independently, so we run
106 		 * at the lowest clocks required to sustain the heaviest
107 		 * workload. However, a task may be split into sequential
108 		 * dependent operations across a set of engines, such that
109 		 * the independent contributions do not account for high load,
110 		 * but overall the task is GPU bound. For example, consider
111 		 * video decode on vcs followed by colour post-processing
112 		 * on vecs, followed by general post-processing on rcs.
113 		 * Since multi-engines being active does imply a single
114 		 * continuous workload across all engines, we hedge our
115 		 * bets by only contributing a factor of the distributed
116 		 * load into our busyness calculation.
117 		 */
118 		busy = max_busy[0];
119 		for (i = 1; i < ARRAY_SIZE(max_busy); i++) {
120 			if (!max_busy[i])
121 				break;
122 
123 			busy += div_u64(max_busy[i], 1 << i);
124 		}
125 		GT_TRACE(rps_to_gt(rps),
126 			 "busy:%lld [%d%%], max:[%lld, %lld, %lld], interval:%d\n",
127 			 busy, (int)div64_u64(100 * busy, dt),
128 			 max_busy[0], max_busy[1], max_busy[2],
129 			 rps->pm_interval);
130 
131 		if (100 * busy > rps->power.up_threshold * dt &&
132 		    rps->cur_freq < rps->max_freq_softlimit) {
133 			rps->pm_iir |= GEN6_PM_RP_UP_THRESHOLD;
134 			rps->pm_interval = 1;
135 			schedule_work(&rps->work);
136 		} else if (100 * busy < rps->power.down_threshold * dt &&
137 			   rps->cur_freq > rps->min_freq_softlimit) {
138 			rps->pm_iir |= GEN6_PM_RP_DOWN_THRESHOLD;
139 			rps->pm_interval = 1;
140 			schedule_work(&rps->work);
141 		} else {
142 			rps->last_adj = 0;
143 		}
144 
145 		mod_timer(&rps->timer,
146 			  jiffies + msecs_to_jiffies(rps->pm_interval));
147 		rps->pm_interval = min(rps->pm_interval * 2, BUSY_MAX_EI);
148 	}
149 }
150 
151 static void rps_start_timer(struct intel_rps *rps)
152 {
153 	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
154 	rps->pm_interval = 1;
155 	mod_timer(&rps->timer, jiffies + 1);
156 }
157 
158 static void rps_stop_timer(struct intel_rps *rps)
159 {
160 	del_timer_sync(&rps->timer);
161 	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
162 	cancel_work_sync(&rps->work);
163 }
164 
165 static u32 rps_pm_mask(struct intel_rps *rps, u8 val)
166 {
167 	u32 mask = 0;
168 
169 	/* We use UP_EI_EXPIRED interrupts for both up/down in manual mode */
170 	if (val > rps->min_freq_softlimit)
171 		mask |= (GEN6_PM_RP_UP_EI_EXPIRED |
172 			 GEN6_PM_RP_DOWN_THRESHOLD |
173 			 GEN6_PM_RP_DOWN_TIMEOUT);
174 
175 	if (val < rps->max_freq_softlimit)
176 		mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
177 
178 	mask &= rps->pm_events;
179 
180 	return rps_pm_sanitize_mask(rps, ~mask);
181 }
182 
183 static void rps_reset_ei(struct intel_rps *rps)
184 {
185 	memset(&rps->ei, 0, sizeof(rps->ei));
186 }
187 
188 static void rps_enable_interrupts(struct intel_rps *rps)
189 {
190 	struct intel_gt *gt = rps_to_gt(rps);
191 
192 	GEM_BUG_ON(rps_uses_slpc(rps));
193 
194 	GT_TRACE(gt, "interrupts:on rps->pm_events: %x, rps_pm_mask:%x\n",
195 		 rps->pm_events, rps_pm_mask(rps, rps->last_freq));
196 
197 	rps_reset_ei(rps);
198 
199 	spin_lock_irq(gt->irq_lock);
200 	gen6_gt_pm_enable_irq(gt, rps->pm_events);
201 	spin_unlock_irq(gt->irq_lock);
202 
203 	intel_uncore_write(gt->uncore,
204 			   GEN6_PMINTRMSK, rps_pm_mask(rps, rps->last_freq));
205 }
206 
207 static void gen6_rps_reset_interrupts(struct intel_rps *rps)
208 {
209 	gen6_gt_pm_reset_iir(rps_to_gt(rps), GEN6_PM_RPS_EVENTS);
210 }
211 
212 static void gen11_rps_reset_interrupts(struct intel_rps *rps)
213 {
214 	while (gen11_gt_reset_one_iir(rps_to_gt(rps), 0, GEN11_GTPM))
215 		;
216 }
217 
218 static void rps_reset_interrupts(struct intel_rps *rps)
219 {
220 	struct intel_gt *gt = rps_to_gt(rps);
221 
222 	spin_lock_irq(gt->irq_lock);
223 	if (GRAPHICS_VER(gt->i915) >= 11)
224 		gen11_rps_reset_interrupts(rps);
225 	else
226 		gen6_rps_reset_interrupts(rps);
227 
228 	rps->pm_iir = 0;
229 	spin_unlock_irq(gt->irq_lock);
230 }
231 
232 static void rps_disable_interrupts(struct intel_rps *rps)
233 {
234 	struct intel_gt *gt = rps_to_gt(rps);
235 
236 	intel_uncore_write(gt->uncore,
237 			   GEN6_PMINTRMSK, rps_pm_sanitize_mask(rps, ~0u));
238 
239 	spin_lock_irq(gt->irq_lock);
240 	gen6_gt_pm_disable_irq(gt, GEN6_PM_RPS_EVENTS);
241 	spin_unlock_irq(gt->irq_lock);
242 
243 	intel_synchronize_irq(gt->i915);
244 
245 	/*
246 	 * Now that we will not be generating any more work, flush any
247 	 * outstanding tasks. As we are called on the RPS idle path,
248 	 * we will reset the GPU to minimum frequencies, so the current
249 	 * state of the worker can be discarded.
250 	 */
251 	cancel_work_sync(&rps->work);
252 
253 	rps_reset_interrupts(rps);
254 	GT_TRACE(gt, "interrupts:off\n");
255 }
256 
257 static const struct cparams {
258 	u16 i;
259 	u16 t;
260 	u16 m;
261 	u16 c;
262 } cparams[] = {
263 	{ 1, 1333, 301, 28664 },
264 	{ 1, 1066, 294, 24460 },
265 	{ 1, 800, 294, 25192 },
266 	{ 0, 1333, 276, 27605 },
267 	{ 0, 1066, 276, 27605 },
268 	{ 0, 800, 231, 23784 },
269 };
270 
271 static void gen5_rps_init(struct intel_rps *rps)
272 {
273 	struct drm_i915_private *i915 = rps_to_i915(rps);
274 	struct intel_uncore *uncore = rps_to_uncore(rps);
275 	u8 fmax, fmin, fstart;
276 	u32 rgvmodectl;
277 	int c_m, i;
278 
279 	if (i915->fsb_freq <= 3200)
280 		c_m = 0;
281 	else if (i915->fsb_freq <= 4800)
282 		c_m = 1;
283 	else
284 		c_m = 2;
285 
286 	for (i = 0; i < ARRAY_SIZE(cparams); i++) {
287 		if (cparams[i].i == c_m && cparams[i].t == i915->mem_freq) {
288 			rps->ips.m = cparams[i].m;
289 			rps->ips.c = cparams[i].c;
290 			break;
291 		}
292 	}
293 
294 	rgvmodectl = intel_uncore_read(uncore, MEMMODECTL);
295 
296 	/* Set up min, max, and cur for interrupt handling */
297 	fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
298 	fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
299 	fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
300 		MEMMODE_FSTART_SHIFT;
301 	drm_dbg(&i915->drm, "fmax: %d, fmin: %d, fstart: %d\n",
302 		fmax, fmin, fstart);
303 
304 	rps->min_freq = fmax;
305 	rps->efficient_freq = fstart;
306 	rps->max_freq = fmin;
307 }
308 
309 static unsigned long
310 __ips_chipset_val(struct intel_ips *ips)
311 {
312 	struct intel_uncore *uncore =
313 		rps_to_uncore(container_of(ips, struct intel_rps, ips));
314 	unsigned long now = jiffies_to_msecs(jiffies), dt;
315 	unsigned long result;
316 	u64 total, delta;
317 
318 	lockdep_assert_held(&mchdev_lock);
319 
320 	/*
321 	 * Prevent division-by-zero if we are asking too fast.
322 	 * Also, we don't get interesting results if we are polling
323 	 * faster than once in 10ms, so just return the saved value
324 	 * in such cases.
325 	 */
326 	dt = now - ips->last_time1;
327 	if (dt <= 10)
328 		return ips->chipset_power;
329 
330 	/* FIXME: handle per-counter overflow */
331 	total = intel_uncore_read(uncore, DMIEC);
332 	total += intel_uncore_read(uncore, DDREC);
333 	total += intel_uncore_read(uncore, CSIEC);
334 
335 	delta = total - ips->last_count1;
336 
337 	result = div_u64(div_u64(ips->m * delta, dt) + ips->c, 10);
338 
339 	ips->last_count1 = total;
340 	ips->last_time1 = now;
341 
342 	ips->chipset_power = result;
343 
344 	return result;
345 }
346 
347 static unsigned long ips_mch_val(struct intel_uncore *uncore)
348 {
349 	unsigned int m, x, b;
350 	u32 tsfs;
351 
352 	tsfs = intel_uncore_read(uncore, TSFS);
353 	x = intel_uncore_read8(uncore, TR1);
354 
355 	b = tsfs & TSFS_INTR_MASK;
356 	m = (tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT;
357 
358 	return m * x / 127 - b;
359 }
360 
361 static int _pxvid_to_vd(u8 pxvid)
362 {
363 	if (pxvid == 0)
364 		return 0;
365 
366 	if (pxvid >= 8 && pxvid < 31)
367 		pxvid = 31;
368 
369 	return (pxvid + 2) * 125;
370 }
371 
372 static u32 pvid_to_extvid(struct drm_i915_private *i915, u8 pxvid)
373 {
374 	const int vd = _pxvid_to_vd(pxvid);
375 
376 	if (INTEL_INFO(i915)->is_mobile)
377 		return max(vd - 1125, 0);
378 
379 	return vd;
380 }
381 
382 static void __gen5_ips_update(struct intel_ips *ips)
383 {
384 	struct intel_uncore *uncore =
385 		rps_to_uncore(container_of(ips, struct intel_rps, ips));
386 	u64 now, delta, dt;
387 	u32 count;
388 
389 	lockdep_assert_held(&mchdev_lock);
390 
391 	now = ktime_get_raw_ns();
392 	dt = now - ips->last_time2;
393 	do_div(dt, NSEC_PER_MSEC);
394 
395 	/* Don't divide by 0 */
396 	if (dt <= 10)
397 		return;
398 
399 	count = intel_uncore_read(uncore, GFXEC);
400 	delta = count - ips->last_count2;
401 
402 	ips->last_count2 = count;
403 	ips->last_time2 = now;
404 
405 	/* More magic constants... */
406 	ips->gfx_power = div_u64(delta * 1181, dt * 10);
407 }
408 
409 static void gen5_rps_update(struct intel_rps *rps)
410 {
411 	spin_lock_irq(&mchdev_lock);
412 	__gen5_ips_update(&rps->ips);
413 	spin_unlock_irq(&mchdev_lock);
414 }
415 
416 static unsigned int gen5_invert_freq(struct intel_rps *rps,
417 				     unsigned int val)
418 {
419 	/* Invert the frequency bin into an ips delay */
420 	val = rps->max_freq - val;
421 	val = rps->min_freq + val;
422 
423 	return val;
424 }
425 
426 static int __gen5_rps_set(struct intel_rps *rps, u8 val)
427 {
428 	struct intel_uncore *uncore = rps_to_uncore(rps);
429 	u16 rgvswctl;
430 
431 	lockdep_assert_held(&mchdev_lock);
432 
433 	rgvswctl = intel_uncore_read16(uncore, MEMSWCTL);
434 	if (rgvswctl & MEMCTL_CMD_STS) {
435 		DRM_DEBUG("gpu busy, RCS change rejected\n");
436 		return -EBUSY; /* still busy with another command */
437 	}
438 
439 	/* Invert the frequency bin into an ips delay */
440 	val = gen5_invert_freq(rps, val);
441 
442 	rgvswctl =
443 		(MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
444 		(val << MEMCTL_FREQ_SHIFT) |
445 		MEMCTL_SFCAVM;
446 	intel_uncore_write16(uncore, MEMSWCTL, rgvswctl);
447 	intel_uncore_posting_read16(uncore, MEMSWCTL);
448 
449 	rgvswctl |= MEMCTL_CMD_STS;
450 	intel_uncore_write16(uncore, MEMSWCTL, rgvswctl);
451 
452 	return 0;
453 }
454 
455 static int gen5_rps_set(struct intel_rps *rps, u8 val)
456 {
457 	int err;
458 
459 	spin_lock_irq(&mchdev_lock);
460 	err = __gen5_rps_set(rps, val);
461 	spin_unlock_irq(&mchdev_lock);
462 
463 	return err;
464 }
465 
466 static unsigned long intel_pxfreq(u32 vidfreq)
467 {
468 	int div = (vidfreq & 0x3f0000) >> 16;
469 	int post = (vidfreq & 0x3000) >> 12;
470 	int pre = (vidfreq & 0x7);
471 
472 	if (!pre)
473 		return 0;
474 
475 	return div * 133333 / (pre << post);
476 }
477 
478 static unsigned int init_emon(struct intel_uncore *uncore)
479 {
480 	u8 pxw[16];
481 	int i;
482 
483 	/* Disable to program */
484 	intel_uncore_write(uncore, ECR, 0);
485 	intel_uncore_posting_read(uncore, ECR);
486 
487 	/* Program energy weights for various events */
488 	intel_uncore_write(uncore, SDEW, 0x15040d00);
489 	intel_uncore_write(uncore, CSIEW0, 0x007f0000);
490 	intel_uncore_write(uncore, CSIEW1, 0x1e220004);
491 	intel_uncore_write(uncore, CSIEW2, 0x04000004);
492 
493 	for (i = 0; i < 5; i++)
494 		intel_uncore_write(uncore, PEW(i), 0);
495 	for (i = 0; i < 3; i++)
496 		intel_uncore_write(uncore, DEW(i), 0);
497 
498 	/* Program P-state weights to account for frequency power adjustment */
499 	for (i = 0; i < 16; i++) {
500 		u32 pxvidfreq = intel_uncore_read(uncore, PXVFREQ(i));
501 		unsigned int freq = intel_pxfreq(pxvidfreq);
502 		unsigned int vid =
503 			(pxvidfreq & PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT;
504 		unsigned int val;
505 
506 		val = vid * vid * freq / 1000 * 255;
507 		val /= 127 * 127 * 900;
508 
509 		pxw[i] = val;
510 	}
511 	/* Render standby states get 0 weight */
512 	pxw[14] = 0;
513 	pxw[15] = 0;
514 
515 	for (i = 0; i < 4; i++) {
516 		intel_uncore_write(uncore, PXW(i),
517 				   pxw[i * 4 + 0] << 24 |
518 				   pxw[i * 4 + 1] << 16 |
519 				   pxw[i * 4 + 2] <<  8 |
520 				   pxw[i * 4 + 3] <<  0);
521 	}
522 
523 	/* Adjust magic regs to magic values (more experimental results) */
524 	intel_uncore_write(uncore, OGW0, 0);
525 	intel_uncore_write(uncore, OGW1, 0);
526 	intel_uncore_write(uncore, EG0, 0x00007f00);
527 	intel_uncore_write(uncore, EG1, 0x0000000e);
528 	intel_uncore_write(uncore, EG2, 0x000e0000);
529 	intel_uncore_write(uncore, EG3, 0x68000300);
530 	intel_uncore_write(uncore, EG4, 0x42000000);
531 	intel_uncore_write(uncore, EG5, 0x00140031);
532 	intel_uncore_write(uncore, EG6, 0);
533 	intel_uncore_write(uncore, EG7, 0);
534 
535 	for (i = 0; i < 8; i++)
536 		intel_uncore_write(uncore, PXWL(i), 0);
537 
538 	/* Enable PMON + select events */
539 	intel_uncore_write(uncore, ECR, 0x80000019);
540 
541 	return intel_uncore_read(uncore, LCFUSE02) & LCFUSE_HIV_MASK;
542 }
543 
544 static bool gen5_rps_enable(struct intel_rps *rps)
545 {
546 	struct drm_i915_private *i915 = rps_to_i915(rps);
547 	struct intel_uncore *uncore = rps_to_uncore(rps);
548 	u8 fstart, vstart;
549 	u32 rgvmodectl;
550 
551 	spin_lock_irq(&mchdev_lock);
552 
553 	rgvmodectl = intel_uncore_read(uncore, MEMMODECTL);
554 
555 	/* Enable temp reporting */
556 	intel_uncore_write16(uncore, PMMISC,
557 			     intel_uncore_read16(uncore, PMMISC) | MCPPCE_EN);
558 	intel_uncore_write16(uncore, TSC1,
559 			     intel_uncore_read16(uncore, TSC1) | TSE);
560 
561 	/* 100ms RC evaluation intervals */
562 	intel_uncore_write(uncore, RCUPEI, 100000);
563 	intel_uncore_write(uncore, RCDNEI, 100000);
564 
565 	/* Set max/min thresholds to 90ms and 80ms respectively */
566 	intel_uncore_write(uncore, RCBMAXAVG, 90000);
567 	intel_uncore_write(uncore, RCBMINAVG, 80000);
568 
569 	intel_uncore_write(uncore, MEMIHYST, 1);
570 
571 	/* Set up min, max, and cur for interrupt handling */
572 	fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
573 		MEMMODE_FSTART_SHIFT;
574 
575 	vstart = (intel_uncore_read(uncore, PXVFREQ(fstart)) &
576 		  PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT;
577 
578 	intel_uncore_write(uncore,
579 			   MEMINTREN,
580 			   MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
581 
582 	intel_uncore_write(uncore, VIDSTART, vstart);
583 	intel_uncore_posting_read(uncore, VIDSTART);
584 
585 	rgvmodectl |= MEMMODE_SWMODE_EN;
586 	intel_uncore_write(uncore, MEMMODECTL, rgvmodectl);
587 
588 	if (wait_for_atomic((intel_uncore_read(uncore, MEMSWCTL) &
589 			     MEMCTL_CMD_STS) == 0, 10))
590 		drm_err(&uncore->i915->drm,
591 			"stuck trying to change perf mode\n");
592 	mdelay(1);
593 
594 	__gen5_rps_set(rps, rps->cur_freq);
595 
596 	rps->ips.last_count1 = intel_uncore_read(uncore, DMIEC);
597 	rps->ips.last_count1 += intel_uncore_read(uncore, DDREC);
598 	rps->ips.last_count1 += intel_uncore_read(uncore, CSIEC);
599 	rps->ips.last_time1 = jiffies_to_msecs(jiffies);
600 
601 	rps->ips.last_count2 = intel_uncore_read(uncore, GFXEC);
602 	rps->ips.last_time2 = ktime_get_raw_ns();
603 
604 	spin_lock(&i915->irq_lock);
605 	ilk_enable_display_irq(i915, DE_PCU_EVENT);
606 	spin_unlock(&i915->irq_lock);
607 
608 	spin_unlock_irq(&mchdev_lock);
609 
610 	rps->ips.corr = init_emon(uncore);
611 
612 	return true;
613 }
614 
615 static void gen5_rps_disable(struct intel_rps *rps)
616 {
617 	struct drm_i915_private *i915 = rps_to_i915(rps);
618 	struct intel_uncore *uncore = rps_to_uncore(rps);
619 	u16 rgvswctl;
620 
621 	spin_lock_irq(&mchdev_lock);
622 
623 	spin_lock(&i915->irq_lock);
624 	ilk_disable_display_irq(i915, DE_PCU_EVENT);
625 	spin_unlock(&i915->irq_lock);
626 
627 	rgvswctl = intel_uncore_read16(uncore, MEMSWCTL);
628 
629 	/* Ack interrupts, disable EFC interrupt */
630 	intel_uncore_write(uncore, MEMINTREN,
631 			   intel_uncore_read(uncore, MEMINTREN) &
632 			   ~MEMINT_EVAL_CHG_EN);
633 	intel_uncore_write(uncore, MEMINTRSTS, MEMINT_EVAL_CHG);
634 
635 	/* Go back to the starting frequency */
636 	__gen5_rps_set(rps, rps->idle_freq);
637 	mdelay(1);
638 	rgvswctl |= MEMCTL_CMD_STS;
639 	intel_uncore_write(uncore, MEMSWCTL, rgvswctl);
640 	mdelay(1);
641 
642 	spin_unlock_irq(&mchdev_lock);
643 }
644 
645 static u32 rps_limits(struct intel_rps *rps, u8 val)
646 {
647 	u32 limits;
648 
649 	/*
650 	 * Only set the down limit when we've reached the lowest level to avoid
651 	 * getting more interrupts, otherwise leave this clear. This prevents a
652 	 * race in the hw when coming out of rc6: There's a tiny window where
653 	 * the hw runs at the minimal clock before selecting the desired
654 	 * frequency, if the down threshold expires in that window we will not
655 	 * receive a down interrupt.
656 	 */
657 	if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) {
658 		limits = rps->max_freq_softlimit << 23;
659 		if (val <= rps->min_freq_softlimit)
660 			limits |= rps->min_freq_softlimit << 14;
661 	} else {
662 		limits = rps->max_freq_softlimit << 24;
663 		if (val <= rps->min_freq_softlimit)
664 			limits |= rps->min_freq_softlimit << 16;
665 	}
666 
667 	return limits;
668 }
669 
670 static void rps_set_power(struct intel_rps *rps, int new_power)
671 {
672 	struct intel_gt *gt = rps_to_gt(rps);
673 	struct intel_uncore *uncore = gt->uncore;
674 	u32 threshold_up = 0, threshold_down = 0; /* in % */
675 	u32 ei_up = 0, ei_down = 0;
676 
677 	lockdep_assert_held(&rps->power.mutex);
678 
679 	if (new_power == rps->power.mode)
680 		return;
681 
682 	threshold_up = 95;
683 	threshold_down = 85;
684 
685 	/* Note the units here are not exactly 1us, but 1280ns. */
686 	switch (new_power) {
687 	case LOW_POWER:
688 		ei_up = 16000;
689 		ei_down = 32000;
690 		break;
691 
692 	case BETWEEN:
693 		ei_up = 13000;
694 		ei_down = 32000;
695 		break;
696 
697 	case HIGH_POWER:
698 		ei_up = 10000;
699 		ei_down = 32000;
700 		break;
701 	}
702 
703 	/* When byt can survive without system hang with dynamic
704 	 * sw freq adjustments, this restriction can be lifted.
705 	 */
706 	if (IS_VALLEYVIEW(gt->i915))
707 		goto skip_hw_write;
708 
709 	GT_TRACE(gt,
710 		 "changing power mode [%d], up %d%% @ %dus, down %d%% @ %dus\n",
711 		 new_power, threshold_up, ei_up, threshold_down, ei_down);
712 
713 	set(uncore, GEN6_RP_UP_EI,
714 	    intel_gt_ns_to_pm_interval(gt, ei_up * 1000));
715 	set(uncore, GEN6_RP_UP_THRESHOLD,
716 	    intel_gt_ns_to_pm_interval(gt, ei_up * threshold_up * 10));
717 
718 	set(uncore, GEN6_RP_DOWN_EI,
719 	    intel_gt_ns_to_pm_interval(gt, ei_down * 1000));
720 	set(uncore, GEN6_RP_DOWN_THRESHOLD,
721 	    intel_gt_ns_to_pm_interval(gt, ei_down * threshold_down * 10));
722 
723 	set(uncore, GEN6_RP_CONTROL,
724 	    (GRAPHICS_VER(gt->i915) > 9 ? 0 : GEN6_RP_MEDIA_TURBO) |
725 	    GEN6_RP_MEDIA_HW_NORMAL_MODE |
726 	    GEN6_RP_MEDIA_IS_GFX |
727 	    GEN6_RP_ENABLE |
728 	    GEN6_RP_UP_BUSY_AVG |
729 	    GEN6_RP_DOWN_IDLE_AVG);
730 
731 skip_hw_write:
732 	rps->power.mode = new_power;
733 	rps->power.up_threshold = threshold_up;
734 	rps->power.down_threshold = threshold_down;
735 }
736 
737 static void gen6_rps_set_thresholds(struct intel_rps *rps, u8 val)
738 {
739 	int new_power;
740 
741 	new_power = rps->power.mode;
742 	switch (rps->power.mode) {
743 	case LOW_POWER:
744 		if (val > rps->efficient_freq + 1 &&
745 		    val > rps->cur_freq)
746 			new_power = BETWEEN;
747 		break;
748 
749 	case BETWEEN:
750 		if (val <= rps->efficient_freq &&
751 		    val < rps->cur_freq)
752 			new_power = LOW_POWER;
753 		else if (val >= rps->rp0_freq &&
754 			 val > rps->cur_freq)
755 			new_power = HIGH_POWER;
756 		break;
757 
758 	case HIGH_POWER:
759 		if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 &&
760 		    val < rps->cur_freq)
761 			new_power = BETWEEN;
762 		break;
763 	}
764 	/* Max/min bins are special */
765 	if (val <= rps->min_freq_softlimit)
766 		new_power = LOW_POWER;
767 	if (val >= rps->max_freq_softlimit)
768 		new_power = HIGH_POWER;
769 
770 	mutex_lock(&rps->power.mutex);
771 	if (rps->power.interactive)
772 		new_power = HIGH_POWER;
773 	rps_set_power(rps, new_power);
774 	mutex_unlock(&rps->power.mutex);
775 }
776 
777 void intel_rps_mark_interactive(struct intel_rps *rps, bool interactive)
778 {
779 	GT_TRACE(rps_to_gt(rps), "mark interactive: %s\n",
780 		 str_yes_no(interactive));
781 
782 	mutex_lock(&rps->power.mutex);
783 	if (interactive) {
784 		if (!rps->power.interactive++ && intel_rps_is_active(rps))
785 			rps_set_power(rps, HIGH_POWER);
786 	} else {
787 		GEM_BUG_ON(!rps->power.interactive);
788 		rps->power.interactive--;
789 	}
790 	mutex_unlock(&rps->power.mutex);
791 }
792 
793 static int gen6_rps_set(struct intel_rps *rps, u8 val)
794 {
795 	struct intel_uncore *uncore = rps_to_uncore(rps);
796 	struct drm_i915_private *i915 = rps_to_i915(rps);
797 	u32 swreq;
798 
799 	GEM_BUG_ON(rps_uses_slpc(rps));
800 
801 	if (GRAPHICS_VER(i915) >= 9)
802 		swreq = GEN9_FREQUENCY(val);
803 	else if (IS_HASWELL(i915) || IS_BROADWELL(i915))
804 		swreq = HSW_FREQUENCY(val);
805 	else
806 		swreq = (GEN6_FREQUENCY(val) |
807 			 GEN6_OFFSET(0) |
808 			 GEN6_AGGRESSIVE_TURBO);
809 	set(uncore, GEN6_RPNSWREQ, swreq);
810 
811 	GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d, swreq:%x\n",
812 		 val, intel_gpu_freq(rps, val), swreq);
813 
814 	return 0;
815 }
816 
817 static int vlv_rps_set(struct intel_rps *rps, u8 val)
818 {
819 	struct drm_i915_private *i915 = rps_to_i915(rps);
820 	int err;
821 
822 	vlv_punit_get(i915);
823 	err = vlv_punit_write(i915, PUNIT_REG_GPU_FREQ_REQ, val);
824 	vlv_punit_put(i915);
825 
826 	GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d\n",
827 		 val, intel_gpu_freq(rps, val));
828 
829 	return err;
830 }
831 
832 static int rps_set(struct intel_rps *rps, u8 val, bool update)
833 {
834 	struct drm_i915_private *i915 = rps_to_i915(rps);
835 	int err;
836 
837 	if (val == rps->last_freq)
838 		return 0;
839 
840 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
841 		err = vlv_rps_set(rps, val);
842 	else if (GRAPHICS_VER(i915) >= 6)
843 		err = gen6_rps_set(rps, val);
844 	else
845 		err = gen5_rps_set(rps, val);
846 	if (err)
847 		return err;
848 
849 	if (update && GRAPHICS_VER(i915) >= 6)
850 		gen6_rps_set_thresholds(rps, val);
851 	rps->last_freq = val;
852 
853 	return 0;
854 }
855 
856 void intel_rps_unpark(struct intel_rps *rps)
857 {
858 	if (!intel_rps_is_enabled(rps))
859 		return;
860 
861 	GT_TRACE(rps_to_gt(rps), "unpark:%x\n", rps->cur_freq);
862 
863 	/*
864 	 * Use the user's desired frequency as a guide, but for better
865 	 * performance, jump directly to RPe as our starting frequency.
866 	 */
867 	mutex_lock(&rps->lock);
868 
869 	intel_rps_set_active(rps);
870 	intel_rps_set(rps,
871 		      clamp(rps->cur_freq,
872 			    rps->min_freq_softlimit,
873 			    rps->max_freq_softlimit));
874 
875 	mutex_unlock(&rps->lock);
876 
877 	rps->pm_iir = 0;
878 	if (intel_rps_has_interrupts(rps))
879 		rps_enable_interrupts(rps);
880 	if (intel_rps_uses_timer(rps))
881 		rps_start_timer(rps);
882 
883 	if (GRAPHICS_VER(rps_to_i915(rps)) == 5)
884 		gen5_rps_update(rps);
885 }
886 
887 void intel_rps_park(struct intel_rps *rps)
888 {
889 	int adj;
890 
891 	if (!intel_rps_is_enabled(rps))
892 		return;
893 
894 	if (!intel_rps_clear_active(rps))
895 		return;
896 
897 	if (intel_rps_uses_timer(rps))
898 		rps_stop_timer(rps);
899 	if (intel_rps_has_interrupts(rps))
900 		rps_disable_interrupts(rps);
901 
902 	if (rps->last_freq <= rps->idle_freq)
903 		return;
904 
905 	/*
906 	 * The punit delays the write of the frequency and voltage until it
907 	 * determines the GPU is awake. During normal usage we don't want to
908 	 * waste power changing the frequency if the GPU is sleeping (rc6).
909 	 * However, the GPU and driver is now idle and we do not want to delay
910 	 * switching to minimum voltage (reducing power whilst idle) as we do
911 	 * not expect to be woken in the near future and so must flush the
912 	 * change by waking the device.
913 	 *
914 	 * We choose to take the media powerwell (either would do to trick the
915 	 * punit into committing the voltage change) as that takes a lot less
916 	 * power than the render powerwell.
917 	 */
918 	intel_uncore_forcewake_get(rps_to_uncore(rps), FORCEWAKE_MEDIA);
919 	rps_set(rps, rps->idle_freq, false);
920 	intel_uncore_forcewake_put(rps_to_uncore(rps), FORCEWAKE_MEDIA);
921 
922 	/*
923 	 * Since we will try and restart from the previously requested
924 	 * frequency on unparking, treat this idle point as a downclock
925 	 * interrupt and reduce the frequency for resume. If we park/unpark
926 	 * more frequently than the rps worker can run, we will not respond
927 	 * to any EI and never see a change in frequency.
928 	 *
929 	 * (Note we accommodate Cherryview's limitation of only using an
930 	 * even bin by applying it to all.)
931 	 */
932 	adj = rps->last_adj;
933 	if (adj < 0)
934 		adj *= 2;
935 	else /* CHV needs even encode values */
936 		adj = -2;
937 	rps->last_adj = adj;
938 	rps->cur_freq = max_t(int, rps->cur_freq + adj, rps->min_freq);
939 	if (rps->cur_freq < rps->efficient_freq) {
940 		rps->cur_freq = rps->efficient_freq;
941 		rps->last_adj = 0;
942 	}
943 
944 	GT_TRACE(rps_to_gt(rps), "park:%x\n", rps->cur_freq);
945 }
946 
947 u32 intel_rps_get_boost_frequency(struct intel_rps *rps)
948 {
949 	struct intel_guc_slpc *slpc;
950 
951 	if (rps_uses_slpc(rps)) {
952 		slpc = rps_to_slpc(rps);
953 
954 		return slpc->boost_freq;
955 	} else {
956 		return intel_gpu_freq(rps, rps->boost_freq);
957 	}
958 }
959 
960 static int rps_set_boost_freq(struct intel_rps *rps, u32 val)
961 {
962 	bool boost = false;
963 
964 	/* Validate against (static) hardware limits */
965 	val = intel_freq_opcode(rps, val);
966 	if (val < rps->min_freq || val > rps->max_freq)
967 		return -EINVAL;
968 
969 	mutex_lock(&rps->lock);
970 	if (val != rps->boost_freq) {
971 		rps->boost_freq = val;
972 		boost = atomic_read(&rps->num_waiters);
973 	}
974 	mutex_unlock(&rps->lock);
975 	if (boost)
976 		schedule_work(&rps->work);
977 
978 	return 0;
979 }
980 
981 int intel_rps_set_boost_frequency(struct intel_rps *rps, u32 freq)
982 {
983 	struct intel_guc_slpc *slpc;
984 
985 	if (rps_uses_slpc(rps)) {
986 		slpc = rps_to_slpc(rps);
987 
988 		return intel_guc_slpc_set_boost_freq(slpc, freq);
989 	} else {
990 		return rps_set_boost_freq(rps, freq);
991 	}
992 }
993 
994 void intel_rps_dec_waiters(struct intel_rps *rps)
995 {
996 	struct intel_guc_slpc *slpc;
997 
998 	if (rps_uses_slpc(rps)) {
999 		slpc = rps_to_slpc(rps);
1000 
1001 		intel_guc_slpc_dec_waiters(slpc);
1002 	} else {
1003 		atomic_dec(&rps->num_waiters);
1004 	}
1005 }
1006 
1007 void intel_rps_boost(struct i915_request *rq)
1008 {
1009 	struct intel_guc_slpc *slpc;
1010 
1011 	if (i915_request_signaled(rq) || i915_request_has_waitboost(rq))
1012 		return;
1013 
1014 	/* Serializes with i915_request_retire() */
1015 	if (!test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags)) {
1016 		struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps;
1017 
1018 		if (rps_uses_slpc(rps)) {
1019 			slpc = rps_to_slpc(rps);
1020 
1021 			/* Return if old value is non zero */
1022 			if (!atomic_fetch_inc(&slpc->num_waiters))
1023 				schedule_work(&slpc->boost_work);
1024 
1025 			return;
1026 		}
1027 
1028 		if (atomic_fetch_inc(&rps->num_waiters))
1029 			return;
1030 
1031 		if (!intel_rps_is_active(rps))
1032 			return;
1033 
1034 		GT_TRACE(rps_to_gt(rps), "boost fence:%llx:%llx\n",
1035 			 rq->fence.context, rq->fence.seqno);
1036 
1037 		if (READ_ONCE(rps->cur_freq) < rps->boost_freq)
1038 			schedule_work(&rps->work);
1039 
1040 		WRITE_ONCE(rps->boosts, rps->boosts + 1); /* debug only */
1041 	}
1042 }
1043 
1044 int intel_rps_set(struct intel_rps *rps, u8 val)
1045 {
1046 	int err;
1047 
1048 	lockdep_assert_held(&rps->lock);
1049 	GEM_BUG_ON(val > rps->max_freq);
1050 	GEM_BUG_ON(val < rps->min_freq);
1051 
1052 	if (intel_rps_is_active(rps)) {
1053 		err = rps_set(rps, val, true);
1054 		if (err)
1055 			return err;
1056 
1057 		/*
1058 		 * Make sure we continue to get interrupts
1059 		 * until we hit the minimum or maximum frequencies.
1060 		 */
1061 		if (intel_rps_has_interrupts(rps)) {
1062 			struct intel_uncore *uncore = rps_to_uncore(rps);
1063 
1064 			set(uncore,
1065 			    GEN6_RP_INTERRUPT_LIMITS, rps_limits(rps, val));
1066 
1067 			set(uncore, GEN6_PMINTRMSK, rps_pm_mask(rps, val));
1068 		}
1069 	}
1070 
1071 	rps->cur_freq = val;
1072 	return 0;
1073 }
1074 
1075 static u32 intel_rps_read_state_cap(struct intel_rps *rps)
1076 {
1077 	struct drm_i915_private *i915 = rps_to_i915(rps);
1078 	struct intel_uncore *uncore = rps_to_uncore(rps);
1079 
1080 	if (IS_PONTEVECCHIO(i915))
1081 		return intel_uncore_read(uncore, PVC_RP_STATE_CAP);
1082 	else if (IS_XEHPSDV(i915))
1083 		return intel_uncore_read(uncore, XEHPSDV_RP_STATE_CAP);
1084 	else if (IS_GEN9_LP(i915))
1085 		return intel_uncore_read(uncore, BXT_RP_STATE_CAP);
1086 	else
1087 		return intel_uncore_read(uncore, GEN6_RP_STATE_CAP);
1088 }
1089 
1090 /**
1091  * gen6_rps_get_freq_caps - Get freq caps exposed by HW
1092  * @rps: the intel_rps structure
1093  * @caps: returned freq caps
1094  *
1095  * Returned "caps" frequencies should be converted to MHz using
1096  * intel_gpu_freq()
1097  */
1098 void gen6_rps_get_freq_caps(struct intel_rps *rps, struct intel_rps_freq_caps *caps)
1099 {
1100 	struct drm_i915_private *i915 = rps_to_i915(rps);
1101 	u32 rp_state_cap;
1102 
1103 	rp_state_cap = intel_rps_read_state_cap(rps);
1104 
1105 	/* static values from HW: RP0 > RP1 > RPn (min_freq) */
1106 	if (IS_GEN9_LP(i915)) {
1107 		caps->rp0_freq = (rp_state_cap >> 16) & 0xff;
1108 		caps->rp1_freq = (rp_state_cap >>  8) & 0xff;
1109 		caps->min_freq = (rp_state_cap >>  0) & 0xff;
1110 	} else {
1111 		caps->rp0_freq = (rp_state_cap >>  0) & 0xff;
1112 		if (GRAPHICS_VER(i915) >= 10)
1113 			caps->rp1_freq = REG_FIELD_GET(RPE_MASK,
1114 						       intel_uncore_read(to_gt(i915)->uncore,
1115 						       GEN10_FREQ_INFO_REC));
1116 		else
1117 			caps->rp1_freq = (rp_state_cap >>  8) & 0xff;
1118 		caps->min_freq = (rp_state_cap >> 16) & 0xff;
1119 	}
1120 
1121 	if (IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) {
1122 		/*
1123 		 * In this case rp_state_cap register reports frequencies in
1124 		 * units of 50 MHz. Convert these to the actual "hw unit", i.e.
1125 		 * units of 16.67 MHz
1126 		 */
1127 		caps->rp0_freq *= GEN9_FREQ_SCALER;
1128 		caps->rp1_freq *= GEN9_FREQ_SCALER;
1129 		caps->min_freq *= GEN9_FREQ_SCALER;
1130 	}
1131 }
1132 
1133 static void gen6_rps_init(struct intel_rps *rps)
1134 {
1135 	struct drm_i915_private *i915 = rps_to_i915(rps);
1136 	struct intel_rps_freq_caps caps;
1137 
1138 	gen6_rps_get_freq_caps(rps, &caps);
1139 	rps->rp0_freq = caps.rp0_freq;
1140 	rps->rp1_freq = caps.rp1_freq;
1141 	rps->min_freq = caps.min_freq;
1142 
1143 	/* hw_max = RP0 until we check for overclocking */
1144 	rps->max_freq = rps->rp0_freq;
1145 
1146 	rps->efficient_freq = rps->rp1_freq;
1147 	if (IS_HASWELL(i915) || IS_BROADWELL(i915) ||
1148 	    IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11) {
1149 		u32 ddcc_status = 0;
1150 		u32 mult = 1;
1151 
1152 		if (IS_GEN9_BC(i915) || GRAPHICS_VER(i915) >= 11)
1153 			mult = GEN9_FREQ_SCALER;
1154 		if (snb_pcode_read(rps_to_gt(rps)->uncore,
1155 				   HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
1156 				   &ddcc_status, NULL) == 0)
1157 			rps->efficient_freq =
1158 				clamp_t(u32,
1159 					((ddcc_status >> 8) & 0xff) * mult,
1160 					rps->min_freq,
1161 					rps->max_freq);
1162 	}
1163 }
1164 
1165 static bool rps_reset(struct intel_rps *rps)
1166 {
1167 	struct drm_i915_private *i915 = rps_to_i915(rps);
1168 
1169 	/* force a reset */
1170 	rps->power.mode = -1;
1171 	rps->last_freq = -1;
1172 
1173 	if (rps_set(rps, rps->min_freq, true)) {
1174 		drm_err(&i915->drm, "Failed to reset RPS to initial values\n");
1175 		return false;
1176 	}
1177 
1178 	rps->cur_freq = rps->min_freq;
1179 	return true;
1180 }
1181 
1182 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
1183 static bool gen9_rps_enable(struct intel_rps *rps)
1184 {
1185 	struct intel_gt *gt = rps_to_gt(rps);
1186 	struct intel_uncore *uncore = gt->uncore;
1187 
1188 	/* Program defaults and thresholds for RPS */
1189 	if (GRAPHICS_VER(gt->i915) == 9)
1190 		intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ,
1191 				      GEN9_FREQUENCY(rps->rp1_freq));
1192 
1193 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 0xa);
1194 
1195 	rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD;
1196 
1197 	return rps_reset(rps);
1198 }
1199 
1200 static bool gen8_rps_enable(struct intel_rps *rps)
1201 {
1202 	struct intel_uncore *uncore = rps_to_uncore(rps);
1203 
1204 	intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ,
1205 			      HSW_FREQUENCY(rps->rp1_freq));
1206 
1207 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1208 
1209 	rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD;
1210 
1211 	return rps_reset(rps);
1212 }
1213 
1214 static bool gen6_rps_enable(struct intel_rps *rps)
1215 {
1216 	struct intel_uncore *uncore = rps_to_uncore(rps);
1217 
1218 	/* Power down if completely idle for over 50ms */
1219 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 50000);
1220 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1221 
1222 	rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD |
1223 			  GEN6_PM_RP_DOWN_THRESHOLD |
1224 			  GEN6_PM_RP_DOWN_TIMEOUT);
1225 
1226 	return rps_reset(rps);
1227 }
1228 
1229 static int chv_rps_max_freq(struct intel_rps *rps)
1230 {
1231 	struct drm_i915_private *i915 = rps_to_i915(rps);
1232 	struct intel_gt *gt = rps_to_gt(rps);
1233 	u32 val;
1234 
1235 	val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE);
1236 
1237 	switch (gt->info.sseu.eu_total) {
1238 	case 8:
1239 		/* (2 * 4) config */
1240 		val >>= FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT;
1241 		break;
1242 	case 12:
1243 		/* (2 * 6) config */
1244 		val >>= FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT;
1245 		break;
1246 	case 16:
1247 		/* (2 * 8) config */
1248 	default:
1249 		/* Setting (2 * 8) Min RP0 for any other combination */
1250 		val >>= FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT;
1251 		break;
1252 	}
1253 
1254 	return val & FB_GFX_FREQ_FUSE_MASK;
1255 }
1256 
1257 static int chv_rps_rpe_freq(struct intel_rps *rps)
1258 {
1259 	struct drm_i915_private *i915 = rps_to_i915(rps);
1260 	u32 val;
1261 
1262 	val = vlv_punit_read(i915, PUNIT_GPU_DUTYCYCLE_REG);
1263 	val >>= PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT;
1264 
1265 	return val & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
1266 }
1267 
1268 static int chv_rps_guar_freq(struct intel_rps *rps)
1269 {
1270 	struct drm_i915_private *i915 = rps_to_i915(rps);
1271 	u32 val;
1272 
1273 	val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE);
1274 
1275 	return val & FB_GFX_FREQ_FUSE_MASK;
1276 }
1277 
1278 static u32 chv_rps_min_freq(struct intel_rps *rps)
1279 {
1280 	struct drm_i915_private *i915 = rps_to_i915(rps);
1281 	u32 val;
1282 
1283 	val = vlv_punit_read(i915, FB_GFX_FMIN_AT_VMIN_FUSE);
1284 	val >>= FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT;
1285 
1286 	return val & FB_GFX_FREQ_FUSE_MASK;
1287 }
1288 
1289 static bool chv_rps_enable(struct intel_rps *rps)
1290 {
1291 	struct intel_uncore *uncore = rps_to_uncore(rps);
1292 	struct drm_i915_private *i915 = rps_to_i915(rps);
1293 	u32 val;
1294 
1295 	/* 1: Program defaults and thresholds for RPS*/
1296 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000);
1297 	intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400);
1298 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000);
1299 	intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000);
1300 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000);
1301 
1302 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1303 
1304 	/* 2: Enable RPS */
1305 	intel_uncore_write_fw(uncore, GEN6_RP_CONTROL,
1306 			      GEN6_RP_MEDIA_HW_NORMAL_MODE |
1307 			      GEN6_RP_MEDIA_IS_GFX |
1308 			      GEN6_RP_ENABLE |
1309 			      GEN6_RP_UP_BUSY_AVG |
1310 			      GEN6_RP_DOWN_IDLE_AVG);
1311 
1312 	rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD |
1313 			  GEN6_PM_RP_DOWN_THRESHOLD |
1314 			  GEN6_PM_RP_DOWN_TIMEOUT);
1315 
1316 	/* Setting Fixed Bias */
1317 	vlv_punit_get(i915);
1318 
1319 	val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | CHV_BIAS_CPU_50_SOC_50;
1320 	vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val);
1321 
1322 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1323 
1324 	vlv_punit_put(i915);
1325 
1326 	/* RPS code assumes GPLL is used */
1327 	drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0,
1328 		      "GPLL not enabled\n");
1329 
1330 	drm_dbg(&i915->drm, "GPLL enabled? %s\n",
1331 		str_yes_no(val & GPLLENABLE));
1332 	drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val);
1333 
1334 	return rps_reset(rps);
1335 }
1336 
1337 static int vlv_rps_guar_freq(struct intel_rps *rps)
1338 {
1339 	struct drm_i915_private *i915 = rps_to_i915(rps);
1340 	u32 val, rp1;
1341 
1342 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE);
1343 
1344 	rp1 = val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK;
1345 	rp1 >>= FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
1346 
1347 	return rp1;
1348 }
1349 
1350 static int vlv_rps_max_freq(struct intel_rps *rps)
1351 {
1352 	struct drm_i915_private *i915 = rps_to_i915(rps);
1353 	u32 val, rp0;
1354 
1355 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE);
1356 
1357 	rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
1358 	/* Clamp to max */
1359 	rp0 = min_t(u32, rp0, 0xea);
1360 
1361 	return rp0;
1362 }
1363 
1364 static int vlv_rps_rpe_freq(struct intel_rps *rps)
1365 {
1366 	struct drm_i915_private *i915 = rps_to_i915(rps);
1367 	u32 val, rpe;
1368 
1369 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
1370 	rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
1371 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
1372 	rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
1373 
1374 	return rpe;
1375 }
1376 
1377 static int vlv_rps_min_freq(struct intel_rps *rps)
1378 {
1379 	struct drm_i915_private *i915 = rps_to_i915(rps);
1380 	u32 val;
1381 
1382 	val = vlv_punit_read(i915, PUNIT_REG_GPU_LFM) & 0xff;
1383 	/*
1384 	 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
1385 	 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
1386 	 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
1387 	 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
1388 	 * to make sure it matches what Punit accepts.
1389 	 */
1390 	return max_t(u32, val, 0xc0);
1391 }
1392 
1393 static bool vlv_rps_enable(struct intel_rps *rps)
1394 {
1395 	struct intel_uncore *uncore = rps_to_uncore(rps);
1396 	struct drm_i915_private *i915 = rps_to_i915(rps);
1397 	u32 val;
1398 
1399 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000);
1400 	intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400);
1401 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000);
1402 	intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000);
1403 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000);
1404 
1405 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1406 
1407 	intel_uncore_write_fw(uncore, GEN6_RP_CONTROL,
1408 			      GEN6_RP_MEDIA_TURBO |
1409 			      GEN6_RP_MEDIA_HW_NORMAL_MODE |
1410 			      GEN6_RP_MEDIA_IS_GFX |
1411 			      GEN6_RP_ENABLE |
1412 			      GEN6_RP_UP_BUSY_AVG |
1413 			      GEN6_RP_DOWN_IDLE_CONT);
1414 
1415 	/* WaGsvRC0ResidencyMethod:vlv */
1416 	rps->pm_events = GEN6_PM_RP_UP_EI_EXPIRED;
1417 
1418 	vlv_punit_get(i915);
1419 
1420 	/* Setting Fixed Bias */
1421 	val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | VLV_BIAS_CPU_125_SOC_875;
1422 	vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val);
1423 
1424 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1425 
1426 	vlv_punit_put(i915);
1427 
1428 	/* RPS code assumes GPLL is used */
1429 	drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0,
1430 		      "GPLL not enabled\n");
1431 
1432 	drm_dbg(&i915->drm, "GPLL enabled? %s\n",
1433 		str_yes_no(val & GPLLENABLE));
1434 	drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val);
1435 
1436 	return rps_reset(rps);
1437 }
1438 
1439 static unsigned long __ips_gfx_val(struct intel_ips *ips)
1440 {
1441 	struct intel_rps *rps = container_of(ips, typeof(*rps), ips);
1442 	struct intel_uncore *uncore = rps_to_uncore(rps);
1443 	unsigned int t, state1, state2;
1444 	u32 pxvid, ext_v;
1445 	u64 corr, corr2;
1446 
1447 	lockdep_assert_held(&mchdev_lock);
1448 
1449 	pxvid = intel_uncore_read(uncore, PXVFREQ(rps->cur_freq));
1450 	pxvid = (pxvid >> 24) & 0x7f;
1451 	ext_v = pvid_to_extvid(rps_to_i915(rps), pxvid);
1452 
1453 	state1 = ext_v;
1454 
1455 	/* Revel in the empirically derived constants */
1456 
1457 	/* Correction factor in 1/100000 units */
1458 	t = ips_mch_val(uncore);
1459 	if (t > 80)
1460 		corr = t * 2349 + 135940;
1461 	else if (t >= 50)
1462 		corr = t * 964 + 29317;
1463 	else /* < 50 */
1464 		corr = t * 301 + 1004;
1465 
1466 	corr = div_u64(corr * 150142 * state1, 10000) - 78642;
1467 	corr2 = div_u64(corr, 100000) * ips->corr;
1468 
1469 	state2 = div_u64(corr2 * state1, 10000);
1470 	state2 /= 100; /* convert to mW */
1471 
1472 	__gen5_ips_update(ips);
1473 
1474 	return ips->gfx_power + state2;
1475 }
1476 
1477 static bool has_busy_stats(struct intel_rps *rps)
1478 {
1479 	struct intel_engine_cs *engine;
1480 	enum intel_engine_id id;
1481 
1482 	for_each_engine(engine, rps_to_gt(rps), id) {
1483 		if (!intel_engine_supports_stats(engine))
1484 			return false;
1485 	}
1486 
1487 	return true;
1488 }
1489 
1490 void intel_rps_enable(struct intel_rps *rps)
1491 {
1492 	struct drm_i915_private *i915 = rps_to_i915(rps);
1493 	struct intel_uncore *uncore = rps_to_uncore(rps);
1494 	bool enabled = false;
1495 
1496 	if (!HAS_RPS(i915))
1497 		return;
1498 
1499 	if (rps_uses_slpc(rps))
1500 		return;
1501 
1502 	intel_gt_check_clock_frequency(rps_to_gt(rps));
1503 
1504 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1505 	if (rps->max_freq <= rps->min_freq)
1506 		/* leave disabled, no room for dynamic reclocking */;
1507 	else if (IS_CHERRYVIEW(i915))
1508 		enabled = chv_rps_enable(rps);
1509 	else if (IS_VALLEYVIEW(i915))
1510 		enabled = vlv_rps_enable(rps);
1511 	else if (GRAPHICS_VER(i915) >= 9)
1512 		enabled = gen9_rps_enable(rps);
1513 	else if (GRAPHICS_VER(i915) >= 8)
1514 		enabled = gen8_rps_enable(rps);
1515 	else if (GRAPHICS_VER(i915) >= 6)
1516 		enabled = gen6_rps_enable(rps);
1517 	else if (IS_IRONLAKE_M(i915))
1518 		enabled = gen5_rps_enable(rps);
1519 	else
1520 		MISSING_CASE(GRAPHICS_VER(i915));
1521 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1522 	if (!enabled)
1523 		return;
1524 
1525 	GT_TRACE(rps_to_gt(rps),
1526 		 "min:%x, max:%x, freq:[%d, %d]\n",
1527 		 rps->min_freq, rps->max_freq,
1528 		 intel_gpu_freq(rps, rps->min_freq),
1529 		 intel_gpu_freq(rps, rps->max_freq));
1530 
1531 	GEM_BUG_ON(rps->max_freq < rps->min_freq);
1532 	GEM_BUG_ON(rps->idle_freq > rps->max_freq);
1533 
1534 	GEM_BUG_ON(rps->efficient_freq < rps->min_freq);
1535 	GEM_BUG_ON(rps->efficient_freq > rps->max_freq);
1536 
1537 	if (has_busy_stats(rps))
1538 		intel_rps_set_timer(rps);
1539 	else if (GRAPHICS_VER(i915) >= 6 && GRAPHICS_VER(i915) <= 11)
1540 		intel_rps_set_interrupts(rps);
1541 	else
1542 		/* Ironlake currently uses intel_ips.ko */ {}
1543 
1544 	intel_rps_set_enabled(rps);
1545 }
1546 
1547 static void gen6_rps_disable(struct intel_rps *rps)
1548 {
1549 	set(rps_to_uncore(rps), GEN6_RP_CONTROL, 0);
1550 }
1551 
1552 void intel_rps_disable(struct intel_rps *rps)
1553 {
1554 	struct drm_i915_private *i915 = rps_to_i915(rps);
1555 
1556 	if (!intel_rps_is_enabled(rps))
1557 		return;
1558 
1559 	intel_rps_clear_enabled(rps);
1560 	intel_rps_clear_interrupts(rps);
1561 	intel_rps_clear_timer(rps);
1562 
1563 	if (GRAPHICS_VER(i915) >= 6)
1564 		gen6_rps_disable(rps);
1565 	else if (IS_IRONLAKE_M(i915))
1566 		gen5_rps_disable(rps);
1567 }
1568 
1569 static int byt_gpu_freq(struct intel_rps *rps, int val)
1570 {
1571 	/*
1572 	 * N = val - 0xb7
1573 	 * Slow = Fast = GPLL ref * N
1574 	 */
1575 	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000);
1576 }
1577 
1578 static int byt_freq_opcode(struct intel_rps *rps, int val)
1579 {
1580 	return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7;
1581 }
1582 
1583 static int chv_gpu_freq(struct intel_rps *rps, int val)
1584 {
1585 	/*
1586 	 * N = val / 2
1587 	 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
1588 	 */
1589 	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000);
1590 }
1591 
1592 static int chv_freq_opcode(struct intel_rps *rps, int val)
1593 {
1594 	/* CHV needs even values */
1595 	return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2;
1596 }
1597 
1598 int intel_gpu_freq(struct intel_rps *rps, int val)
1599 {
1600 	struct drm_i915_private *i915 = rps_to_i915(rps);
1601 
1602 	if (GRAPHICS_VER(i915) >= 9)
1603 		return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
1604 					 GEN9_FREQ_SCALER);
1605 	else if (IS_CHERRYVIEW(i915))
1606 		return chv_gpu_freq(rps, val);
1607 	else if (IS_VALLEYVIEW(i915))
1608 		return byt_gpu_freq(rps, val);
1609 	else if (GRAPHICS_VER(i915) >= 6)
1610 		return val * GT_FREQUENCY_MULTIPLIER;
1611 	else
1612 		return val;
1613 }
1614 
1615 int intel_freq_opcode(struct intel_rps *rps, int val)
1616 {
1617 	struct drm_i915_private *i915 = rps_to_i915(rps);
1618 
1619 	if (GRAPHICS_VER(i915) >= 9)
1620 		return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
1621 					 GT_FREQUENCY_MULTIPLIER);
1622 	else if (IS_CHERRYVIEW(i915))
1623 		return chv_freq_opcode(rps, val);
1624 	else if (IS_VALLEYVIEW(i915))
1625 		return byt_freq_opcode(rps, val);
1626 	else if (GRAPHICS_VER(i915) >= 6)
1627 		return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
1628 	else
1629 		return val;
1630 }
1631 
1632 static void vlv_init_gpll_ref_freq(struct intel_rps *rps)
1633 {
1634 	struct drm_i915_private *i915 = rps_to_i915(rps);
1635 
1636 	rps->gpll_ref_freq =
1637 		vlv_get_cck_clock(i915, "GPLL ref",
1638 				  CCK_GPLL_CLOCK_CONTROL,
1639 				  i915->czclk_freq);
1640 
1641 	drm_dbg(&i915->drm, "GPLL reference freq: %d kHz\n",
1642 		rps->gpll_ref_freq);
1643 }
1644 
1645 static void vlv_rps_init(struct intel_rps *rps)
1646 {
1647 	struct drm_i915_private *i915 = rps_to_i915(rps);
1648 	u32 val;
1649 
1650 	vlv_iosf_sb_get(i915,
1651 			BIT(VLV_IOSF_SB_PUNIT) |
1652 			BIT(VLV_IOSF_SB_NC) |
1653 			BIT(VLV_IOSF_SB_CCK));
1654 
1655 	vlv_init_gpll_ref_freq(rps);
1656 
1657 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1658 	switch ((val >> 6) & 3) {
1659 	case 0:
1660 	case 1:
1661 		i915->mem_freq = 800;
1662 		break;
1663 	case 2:
1664 		i915->mem_freq = 1066;
1665 		break;
1666 	case 3:
1667 		i915->mem_freq = 1333;
1668 		break;
1669 	}
1670 	drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq);
1671 
1672 	rps->max_freq = vlv_rps_max_freq(rps);
1673 	rps->rp0_freq = rps->max_freq;
1674 	drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n",
1675 		intel_gpu_freq(rps, rps->max_freq), rps->max_freq);
1676 
1677 	rps->efficient_freq = vlv_rps_rpe_freq(rps);
1678 	drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n",
1679 		intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq);
1680 
1681 	rps->rp1_freq = vlv_rps_guar_freq(rps);
1682 	drm_dbg(&i915->drm, "RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
1683 		intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq);
1684 
1685 	rps->min_freq = vlv_rps_min_freq(rps);
1686 	drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n",
1687 		intel_gpu_freq(rps, rps->min_freq), rps->min_freq);
1688 
1689 	vlv_iosf_sb_put(i915,
1690 			BIT(VLV_IOSF_SB_PUNIT) |
1691 			BIT(VLV_IOSF_SB_NC) |
1692 			BIT(VLV_IOSF_SB_CCK));
1693 }
1694 
1695 static void chv_rps_init(struct intel_rps *rps)
1696 {
1697 	struct drm_i915_private *i915 = rps_to_i915(rps);
1698 	u32 val;
1699 
1700 	vlv_iosf_sb_get(i915,
1701 			BIT(VLV_IOSF_SB_PUNIT) |
1702 			BIT(VLV_IOSF_SB_NC) |
1703 			BIT(VLV_IOSF_SB_CCK));
1704 
1705 	vlv_init_gpll_ref_freq(rps);
1706 
1707 	val = vlv_cck_read(i915, CCK_FUSE_REG);
1708 
1709 	switch ((val >> 2) & 0x7) {
1710 	case 3:
1711 		i915->mem_freq = 2000;
1712 		break;
1713 	default:
1714 		i915->mem_freq = 1600;
1715 		break;
1716 	}
1717 	drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq);
1718 
1719 	rps->max_freq = chv_rps_max_freq(rps);
1720 	rps->rp0_freq = rps->max_freq;
1721 	drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n",
1722 		intel_gpu_freq(rps, rps->max_freq), rps->max_freq);
1723 
1724 	rps->efficient_freq = chv_rps_rpe_freq(rps);
1725 	drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n",
1726 		intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq);
1727 
1728 	rps->rp1_freq = chv_rps_guar_freq(rps);
1729 	drm_dbg(&i915->drm, "RP1(Guar) GPU freq: %d MHz (%u)\n",
1730 		intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq);
1731 
1732 	rps->min_freq = chv_rps_min_freq(rps);
1733 	drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n",
1734 		intel_gpu_freq(rps, rps->min_freq), rps->min_freq);
1735 
1736 	vlv_iosf_sb_put(i915,
1737 			BIT(VLV_IOSF_SB_PUNIT) |
1738 			BIT(VLV_IOSF_SB_NC) |
1739 			BIT(VLV_IOSF_SB_CCK));
1740 
1741 	drm_WARN_ONCE(&i915->drm, (rps->max_freq | rps->efficient_freq |
1742 				   rps->rp1_freq | rps->min_freq) & 1,
1743 		      "Odd GPU freq values\n");
1744 }
1745 
1746 static void vlv_c0_read(struct intel_uncore *uncore, struct intel_rps_ei *ei)
1747 {
1748 	ei->ktime = ktime_get_raw();
1749 	ei->render_c0 = intel_uncore_read(uncore, VLV_RENDER_C0_COUNT);
1750 	ei->media_c0 = intel_uncore_read(uncore, VLV_MEDIA_C0_COUNT);
1751 }
1752 
1753 static u32 vlv_wa_c0_ei(struct intel_rps *rps, u32 pm_iir)
1754 {
1755 	struct intel_uncore *uncore = rps_to_uncore(rps);
1756 	const struct intel_rps_ei *prev = &rps->ei;
1757 	struct intel_rps_ei now;
1758 	u32 events = 0;
1759 
1760 	if ((pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) == 0)
1761 		return 0;
1762 
1763 	vlv_c0_read(uncore, &now);
1764 
1765 #ifdef __linux__
1766 	if (prev->ktime) {
1767 #else
1768 	if (ktime_to_ns(prev->ktime)) {
1769 #endif
1770 		u64 time, c0;
1771 		u32 render, media;
1772 
1773 		time = ktime_us_delta(now.ktime, prev->ktime);
1774 
1775 		time *= rps_to_i915(rps)->czclk_freq;
1776 
1777 		/* Workload can be split between render + media,
1778 		 * e.g. SwapBuffers being blitted in X after being rendered in
1779 		 * mesa. To account for this we need to combine both engines
1780 		 * into our activity counter.
1781 		 */
1782 		render = now.render_c0 - prev->render_c0;
1783 		media = now.media_c0 - prev->media_c0;
1784 		c0 = max(render, media);
1785 		c0 *= 1000 * 100 << 8; /* to usecs and scale to threshold% */
1786 
1787 		if (c0 > time * rps->power.up_threshold)
1788 			events = GEN6_PM_RP_UP_THRESHOLD;
1789 		else if (c0 < time * rps->power.down_threshold)
1790 			events = GEN6_PM_RP_DOWN_THRESHOLD;
1791 	}
1792 
1793 	rps->ei = now;
1794 	return events;
1795 }
1796 
1797 static void rps_work(struct work_struct *work)
1798 {
1799 	struct intel_rps *rps = container_of(work, typeof(*rps), work);
1800 	struct intel_gt *gt = rps_to_gt(rps);
1801 	struct drm_i915_private *i915 = rps_to_i915(rps);
1802 	bool client_boost = false;
1803 	int new_freq, adj, min, max;
1804 	u32 pm_iir = 0;
1805 
1806 	spin_lock_irq(gt->irq_lock);
1807 	pm_iir = fetch_and_zero(&rps->pm_iir) & rps->pm_events;
1808 	client_boost = atomic_read(&rps->num_waiters);
1809 	spin_unlock_irq(gt->irq_lock);
1810 
1811 	/* Make sure we didn't queue anything we're not going to process. */
1812 	if (!pm_iir && !client_boost)
1813 		goto out;
1814 
1815 	mutex_lock(&rps->lock);
1816 	if (!intel_rps_is_active(rps)) {
1817 		mutex_unlock(&rps->lock);
1818 		return;
1819 	}
1820 
1821 	pm_iir |= vlv_wa_c0_ei(rps, pm_iir);
1822 
1823 	adj = rps->last_adj;
1824 	new_freq = rps->cur_freq;
1825 	min = rps->min_freq_softlimit;
1826 	max = rps->max_freq_softlimit;
1827 	if (client_boost)
1828 		max = rps->max_freq;
1829 
1830 	GT_TRACE(gt,
1831 		 "pm_iir:%x, client_boost:%s, last:%d, cur:%x, min:%x, max:%x\n",
1832 		 pm_iir, str_yes_no(client_boost),
1833 		 adj, new_freq, min, max);
1834 
1835 	if (client_boost && new_freq < rps->boost_freq) {
1836 		new_freq = rps->boost_freq;
1837 		adj = 0;
1838 	} else if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) {
1839 		if (adj > 0)
1840 			adj *= 2;
1841 		else /* CHV needs even encode values */
1842 			adj = IS_CHERRYVIEW(gt->i915) ? 2 : 1;
1843 
1844 		if (new_freq >= rps->max_freq_softlimit)
1845 			adj = 0;
1846 	} else if (client_boost) {
1847 		adj = 0;
1848 	} else if (pm_iir & GEN6_PM_RP_DOWN_TIMEOUT) {
1849 		if (rps->cur_freq > rps->efficient_freq)
1850 			new_freq = rps->efficient_freq;
1851 		else if (rps->cur_freq > rps->min_freq_softlimit)
1852 			new_freq = rps->min_freq_softlimit;
1853 		adj = 0;
1854 	} else if (pm_iir & GEN6_PM_RP_DOWN_THRESHOLD) {
1855 		if (adj < 0)
1856 			adj *= 2;
1857 		else /* CHV needs even encode values */
1858 			adj = IS_CHERRYVIEW(gt->i915) ? -2 : -1;
1859 
1860 		if (new_freq <= rps->min_freq_softlimit)
1861 			adj = 0;
1862 	} else { /* unknown event */
1863 		adj = 0;
1864 	}
1865 
1866 	/*
1867 	 * sysfs frequency limits may have snuck in while
1868 	 * servicing the interrupt
1869 	 */
1870 	new_freq += adj;
1871 	new_freq = clamp_t(int, new_freq, min, max);
1872 
1873 	if (intel_rps_set(rps, new_freq)) {
1874 		drm_dbg(&i915->drm, "Failed to set new GPU frequency\n");
1875 		adj = 0;
1876 	}
1877 	rps->last_adj = adj;
1878 
1879 	mutex_unlock(&rps->lock);
1880 
1881 out:
1882 	spin_lock_irq(gt->irq_lock);
1883 	gen6_gt_pm_unmask_irq(gt, rps->pm_events);
1884 	spin_unlock_irq(gt->irq_lock);
1885 }
1886 
1887 void gen11_rps_irq_handler(struct intel_rps *rps, u32 pm_iir)
1888 {
1889 	struct intel_gt *gt = rps_to_gt(rps);
1890 	const u32 events = rps->pm_events & pm_iir;
1891 
1892 	lockdep_assert_held(gt->irq_lock);
1893 
1894 	if (unlikely(!events))
1895 		return;
1896 
1897 	GT_TRACE(gt, "irq events:%x\n", events);
1898 
1899 	gen6_gt_pm_mask_irq(gt, events);
1900 
1901 	rps->pm_iir |= events;
1902 	schedule_work(&rps->work);
1903 }
1904 
1905 void gen6_rps_irq_handler(struct intel_rps *rps, u32 pm_iir)
1906 {
1907 	struct intel_gt *gt = rps_to_gt(rps);
1908 	u32 events;
1909 
1910 	events = pm_iir & rps->pm_events;
1911 	if (events) {
1912 		spin_lock(gt->irq_lock);
1913 
1914 		GT_TRACE(gt, "irq events:%x\n", events);
1915 
1916 		gen6_gt_pm_mask_irq(gt, events);
1917 		rps->pm_iir |= events;
1918 
1919 		schedule_work(&rps->work);
1920 		spin_unlock(gt->irq_lock);
1921 	}
1922 
1923 	if (GRAPHICS_VER(gt->i915) >= 8)
1924 		return;
1925 
1926 	if (pm_iir & PM_VEBOX_USER_INTERRUPT)
1927 		intel_engine_cs_irq(gt->engine[VECS0], pm_iir >> 10);
1928 
1929 	if (pm_iir & PM_VEBOX_CS_ERROR_INTERRUPT)
1930 		DRM_DEBUG("Command parser error, pm_iir 0x%08x\n", pm_iir);
1931 }
1932 
1933 void gen5_rps_irq_handler(struct intel_rps *rps)
1934 {
1935 	struct intel_uncore *uncore = rps_to_uncore(rps);
1936 	u32 busy_up, busy_down, max_avg, min_avg;
1937 	u8 new_freq;
1938 
1939 	spin_lock(&mchdev_lock);
1940 
1941 	intel_uncore_write16(uncore,
1942 			     MEMINTRSTS,
1943 			     intel_uncore_read(uncore, MEMINTRSTS));
1944 
1945 	intel_uncore_write16(uncore, MEMINTRSTS, MEMINT_EVAL_CHG);
1946 	busy_up = intel_uncore_read(uncore, RCPREVBSYTUPAVG);
1947 	busy_down = intel_uncore_read(uncore, RCPREVBSYTDNAVG);
1948 	max_avg = intel_uncore_read(uncore, RCBMAXAVG);
1949 	min_avg = intel_uncore_read(uncore, RCBMINAVG);
1950 
1951 	/* Handle RCS change request from hw */
1952 	new_freq = rps->cur_freq;
1953 	if (busy_up > max_avg)
1954 		new_freq++;
1955 	else if (busy_down < min_avg)
1956 		new_freq--;
1957 	new_freq = clamp(new_freq,
1958 			 rps->min_freq_softlimit,
1959 			 rps->max_freq_softlimit);
1960 
1961 	if (new_freq != rps->cur_freq && !__gen5_rps_set(rps, new_freq))
1962 		rps->cur_freq = new_freq;
1963 
1964 	spin_unlock(&mchdev_lock);
1965 }
1966 
1967 void intel_rps_init_early(struct intel_rps *rps)
1968 {
1969 	rw_init(&rps->lock, "rpslk");
1970 	rw_init(&rps->power.mutex, "rpspwr");
1971 
1972 	INIT_WORK(&rps->work, rps_work);
1973 #ifdef __linux__
1974 	timer_setup(&rps->timer, rps_timer, 0);
1975 #else
1976 	timeout_set(&rps->timer, rps_timer, rps);
1977 #endif
1978 
1979 	atomic_set(&rps->num_waiters, 0);
1980 }
1981 
1982 void intel_rps_init(struct intel_rps *rps)
1983 {
1984 	struct drm_i915_private *i915 = rps_to_i915(rps);
1985 
1986 	if (rps_uses_slpc(rps))
1987 		return;
1988 
1989 	if (IS_CHERRYVIEW(i915))
1990 		chv_rps_init(rps);
1991 	else if (IS_VALLEYVIEW(i915))
1992 		vlv_rps_init(rps);
1993 	else if (GRAPHICS_VER(i915) >= 6)
1994 		gen6_rps_init(rps);
1995 	else if (IS_IRONLAKE_M(i915))
1996 		gen5_rps_init(rps);
1997 
1998 	/* Derive initial user preferences/limits from the hardware limits */
1999 	rps->max_freq_softlimit = rps->max_freq;
2000 	rps_to_gt(rps)->defaults.max_freq = rps->max_freq_softlimit;
2001 	rps->min_freq_softlimit = rps->min_freq;
2002 	rps_to_gt(rps)->defaults.min_freq = rps->min_freq_softlimit;
2003 
2004 	/* After setting max-softlimit, find the overclock max freq */
2005 	if (GRAPHICS_VER(i915) == 6 || IS_IVYBRIDGE(i915) || IS_HASWELL(i915)) {
2006 		u32 params = 0;
2007 
2008 		snb_pcode_read(rps_to_gt(rps)->uncore, GEN6_READ_OC_PARAMS, &params, NULL);
2009 		if (params & BIT(31)) { /* OC supported */
2010 			drm_dbg(&i915->drm,
2011 				"Overclocking supported, max: %dMHz, overclock: %dMHz\n",
2012 				(rps->max_freq & 0xff) * 50,
2013 				(params & 0xff) * 50);
2014 			rps->max_freq = params & 0xff;
2015 		}
2016 	}
2017 
2018 	/* Finally allow us to boost to max by default */
2019 	rps->boost_freq = rps->max_freq;
2020 	rps->idle_freq = rps->min_freq;
2021 
2022 	/* Start in the middle, from here we will autotune based on workload */
2023 	rps->cur_freq = rps->efficient_freq;
2024 
2025 	rps->pm_intrmsk_mbz = 0;
2026 
2027 	/*
2028 	 * SNB,IVB,HSW can while VLV,CHV may hard hang on looping batchbuffer
2029 	 * if GEN6_PM_UP_EI_EXPIRED is masked.
2030 	 *
2031 	 * TODO: verify if this can be reproduced on VLV,CHV.
2032 	 */
2033 	if (GRAPHICS_VER(i915) <= 7)
2034 		rps->pm_intrmsk_mbz |= GEN6_PM_RP_UP_EI_EXPIRED;
2035 
2036 	if (GRAPHICS_VER(i915) >= 8 && GRAPHICS_VER(i915) < 11)
2037 		rps->pm_intrmsk_mbz |= GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC;
2038 
2039 	/* GuC needs ARAT expired interrupt unmasked */
2040 	if (intel_uc_uses_guc_submission(&rps_to_gt(rps)->uc))
2041 		rps->pm_intrmsk_mbz |= ARAT_EXPIRED_INTRMSK;
2042 }
2043 
2044 void intel_rps_sanitize(struct intel_rps *rps)
2045 {
2046 	if (rps_uses_slpc(rps))
2047 		return;
2048 
2049 	if (GRAPHICS_VER(rps_to_i915(rps)) >= 6)
2050 		rps_disable_interrupts(rps);
2051 }
2052 
2053 u32 intel_rps_get_cagf(struct intel_rps *rps, u32 rpstat)
2054 {
2055 	struct drm_i915_private *i915 = rps_to_i915(rps);
2056 	u32 cagf;
2057 
2058 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
2059 		cagf = (rpstat >> 8) & 0xff;
2060 	else if (GRAPHICS_VER(i915) >= 9)
2061 		cagf = (rpstat & GEN9_CAGF_MASK) >> GEN9_CAGF_SHIFT;
2062 	else if (IS_HASWELL(i915) || IS_BROADWELL(i915))
2063 		cagf = (rpstat & HSW_CAGF_MASK) >> HSW_CAGF_SHIFT;
2064 	else if (GRAPHICS_VER(i915) >= 6)
2065 		cagf = (rpstat & GEN6_CAGF_MASK) >> GEN6_CAGF_SHIFT;
2066 	else
2067 		cagf = gen5_invert_freq(rps, (rpstat & MEMSTAT_PSTATE_MASK) >>
2068 					MEMSTAT_PSTATE_SHIFT);
2069 
2070 	return cagf;
2071 }
2072 
2073 static u32 read_cagf(struct intel_rps *rps)
2074 {
2075 	struct drm_i915_private *i915 = rps_to_i915(rps);
2076 	struct intel_uncore *uncore = rps_to_uncore(rps);
2077 	u32 freq;
2078 
2079 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) {
2080 		vlv_punit_get(i915);
2081 		freq = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
2082 		vlv_punit_put(i915);
2083 	} else if (GRAPHICS_VER(i915) >= 6) {
2084 		freq = intel_uncore_read(uncore, GEN6_RPSTAT1);
2085 	} else {
2086 		freq = intel_uncore_read(uncore, MEMSTAT_ILK);
2087 	}
2088 
2089 	return intel_rps_get_cagf(rps, freq);
2090 }
2091 
2092 u32 intel_rps_read_actual_frequency(struct intel_rps *rps)
2093 {
2094 	struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm;
2095 	intel_wakeref_t wakeref;
2096 	u32 freq = 0;
2097 
2098 	with_intel_runtime_pm_if_in_use(rpm, wakeref)
2099 		freq = intel_gpu_freq(rps, read_cagf(rps));
2100 
2101 	return freq;
2102 }
2103 
2104 u32 intel_rps_read_punit_req(struct intel_rps *rps)
2105 {
2106 	struct intel_uncore *uncore = rps_to_uncore(rps);
2107 	struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm;
2108 	intel_wakeref_t wakeref;
2109 	u32 freq = 0;
2110 
2111 	with_intel_runtime_pm_if_in_use(rpm, wakeref)
2112 		freq = intel_uncore_read(uncore, GEN6_RPNSWREQ);
2113 
2114 	return freq;
2115 }
2116 
2117 static u32 intel_rps_get_req(u32 pureq)
2118 {
2119 	u32 req = pureq >> GEN9_SW_REQ_UNSLICE_RATIO_SHIFT;
2120 
2121 	return req;
2122 }
2123 
2124 u32 intel_rps_read_punit_req_frequency(struct intel_rps *rps)
2125 {
2126 	u32 freq = intel_rps_get_req(intel_rps_read_punit_req(rps));
2127 
2128 	return intel_gpu_freq(rps, freq);
2129 }
2130 
2131 u32 intel_rps_get_requested_frequency(struct intel_rps *rps)
2132 {
2133 	if (rps_uses_slpc(rps))
2134 		return intel_rps_read_punit_req_frequency(rps);
2135 	else
2136 		return intel_gpu_freq(rps, rps->cur_freq);
2137 }
2138 
2139 u32 intel_rps_get_max_frequency(struct intel_rps *rps)
2140 {
2141 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2142 
2143 	if (rps_uses_slpc(rps))
2144 		return slpc->max_freq_softlimit;
2145 	else
2146 		return intel_gpu_freq(rps, rps->max_freq_softlimit);
2147 }
2148 
2149 /**
2150  * intel_rps_get_max_raw_freq - returns the max frequency in some raw format.
2151  * @rps: the intel_rps structure
2152  *
2153  * Returns the max frequency in a raw format. In newer platforms raw is in
2154  * units of 50 MHz.
2155  */
2156 u32 intel_rps_get_max_raw_freq(struct intel_rps *rps)
2157 {
2158 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2159 	u32 freq;
2160 
2161 	if (rps_uses_slpc(rps)) {
2162 		return DIV_ROUND_CLOSEST(slpc->rp0_freq,
2163 					 GT_FREQUENCY_MULTIPLIER);
2164 	} else {
2165 		freq = rps->max_freq;
2166 		if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) {
2167 			/* Convert GT frequency to 50 MHz units */
2168 			freq /= GEN9_FREQ_SCALER;
2169 		}
2170 		return freq;
2171 	}
2172 }
2173 
2174 u32 intel_rps_get_rp0_frequency(struct intel_rps *rps)
2175 {
2176 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2177 
2178 	if (rps_uses_slpc(rps))
2179 		return slpc->rp0_freq;
2180 	else
2181 		return intel_gpu_freq(rps, rps->rp0_freq);
2182 }
2183 
2184 u32 intel_rps_get_rp1_frequency(struct intel_rps *rps)
2185 {
2186 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2187 
2188 	if (rps_uses_slpc(rps))
2189 		return slpc->rp1_freq;
2190 	else
2191 		return intel_gpu_freq(rps, rps->rp1_freq);
2192 }
2193 
2194 u32 intel_rps_get_rpn_frequency(struct intel_rps *rps)
2195 {
2196 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2197 
2198 	if (rps_uses_slpc(rps))
2199 		return slpc->min_freq;
2200 	else
2201 		return intel_gpu_freq(rps, rps->min_freq);
2202 }
2203 
2204 static int set_max_freq(struct intel_rps *rps, u32 val)
2205 {
2206 	struct drm_i915_private *i915 = rps_to_i915(rps);
2207 	int ret = 0;
2208 
2209 	mutex_lock(&rps->lock);
2210 
2211 	val = intel_freq_opcode(rps, val);
2212 	if (val < rps->min_freq ||
2213 	    val > rps->max_freq ||
2214 	    val < rps->min_freq_softlimit) {
2215 		ret = -EINVAL;
2216 		goto unlock;
2217 	}
2218 
2219 	if (val > rps->rp0_freq)
2220 		drm_dbg(&i915->drm, "User requested overclocking to %d\n",
2221 			intel_gpu_freq(rps, val));
2222 
2223 	rps->max_freq_softlimit = val;
2224 
2225 	val = clamp_t(int, rps->cur_freq,
2226 		      rps->min_freq_softlimit,
2227 		      rps->max_freq_softlimit);
2228 
2229 	/*
2230 	 * We still need *_set_rps to process the new max_delay and
2231 	 * update the interrupt limits and PMINTRMSK even though
2232 	 * frequency request may be unchanged.
2233 	 */
2234 	intel_rps_set(rps, val);
2235 
2236 unlock:
2237 	mutex_unlock(&rps->lock);
2238 
2239 	return ret;
2240 }
2241 
2242 int intel_rps_set_max_frequency(struct intel_rps *rps, u32 val)
2243 {
2244 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2245 
2246 	if (rps_uses_slpc(rps))
2247 		return intel_guc_slpc_set_max_freq(slpc, val);
2248 	else
2249 		return set_max_freq(rps, val);
2250 }
2251 
2252 u32 intel_rps_get_min_frequency(struct intel_rps *rps)
2253 {
2254 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2255 
2256 	if (rps_uses_slpc(rps))
2257 		return slpc->min_freq_softlimit;
2258 	else
2259 		return intel_gpu_freq(rps, rps->min_freq_softlimit);
2260 }
2261 
2262 /**
2263  * intel_rps_get_min_raw_freq - returns the min frequency in some raw format.
2264  * @rps: the intel_rps structure
2265  *
2266  * Returns the min frequency in a raw format. In newer platforms raw is in
2267  * units of 50 MHz.
2268  */
2269 u32 intel_rps_get_min_raw_freq(struct intel_rps *rps)
2270 {
2271 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2272 	u32 freq;
2273 
2274 	if (rps_uses_slpc(rps)) {
2275 		return DIV_ROUND_CLOSEST(slpc->min_freq,
2276 					 GT_FREQUENCY_MULTIPLIER);
2277 	} else {
2278 		freq = rps->min_freq;
2279 		if (GRAPHICS_VER(rps_to_i915(rps)) >= 9) {
2280 			/* Convert GT frequency to 50 MHz units */
2281 			freq /= GEN9_FREQ_SCALER;
2282 		}
2283 		return freq;
2284 	}
2285 }
2286 
2287 static int set_min_freq(struct intel_rps *rps, u32 val)
2288 {
2289 	int ret = 0;
2290 
2291 	mutex_lock(&rps->lock);
2292 
2293 	val = intel_freq_opcode(rps, val);
2294 	if (val < rps->min_freq ||
2295 	    val > rps->max_freq ||
2296 	    val > rps->max_freq_softlimit) {
2297 		ret = -EINVAL;
2298 		goto unlock;
2299 	}
2300 
2301 	rps->min_freq_softlimit = val;
2302 
2303 	val = clamp_t(int, rps->cur_freq,
2304 		      rps->min_freq_softlimit,
2305 		      rps->max_freq_softlimit);
2306 
2307 	/*
2308 	 * We still need *_set_rps to process the new min_delay and
2309 	 * update the interrupt limits and PMINTRMSK even though
2310 	 * frequency request may be unchanged.
2311 	 */
2312 	intel_rps_set(rps, val);
2313 
2314 unlock:
2315 	mutex_unlock(&rps->lock);
2316 
2317 	return ret;
2318 }
2319 
2320 int intel_rps_set_min_frequency(struct intel_rps *rps, u32 val)
2321 {
2322 	struct intel_guc_slpc *slpc = rps_to_slpc(rps);
2323 
2324 	if (rps_uses_slpc(rps))
2325 		return intel_guc_slpc_set_min_freq(slpc, val);
2326 	else
2327 		return set_min_freq(rps, val);
2328 }
2329 
2330 static void intel_rps_set_manual(struct intel_rps *rps, bool enable)
2331 {
2332 	struct intel_uncore *uncore = rps_to_uncore(rps);
2333 	u32 state = enable ? GEN9_RPSWCTL_ENABLE : GEN9_RPSWCTL_DISABLE;
2334 
2335 	/* Allow punit to process software requests */
2336 	intel_uncore_write(uncore, GEN6_RP_CONTROL, state);
2337 }
2338 
2339 void intel_rps_raise_unslice(struct intel_rps *rps)
2340 {
2341 	struct intel_uncore *uncore = rps_to_uncore(rps);
2342 
2343 	mutex_lock(&rps->lock);
2344 
2345 	if (rps_uses_slpc(rps)) {
2346 		/* RP limits have not been initialized yet for SLPC path */
2347 		struct intel_rps_freq_caps caps;
2348 
2349 		gen6_rps_get_freq_caps(rps, &caps);
2350 
2351 		intel_rps_set_manual(rps, true);
2352 		intel_uncore_write(uncore, GEN6_RPNSWREQ,
2353 				   ((caps.rp0_freq <<
2354 				   GEN9_SW_REQ_UNSLICE_RATIO_SHIFT) |
2355 				   GEN9_IGNORE_SLICE_RATIO));
2356 		intel_rps_set_manual(rps, false);
2357 	} else {
2358 		intel_rps_set(rps, rps->rp0_freq);
2359 	}
2360 
2361 	mutex_unlock(&rps->lock);
2362 }
2363 
2364 void intel_rps_lower_unslice(struct intel_rps *rps)
2365 {
2366 	struct intel_uncore *uncore = rps_to_uncore(rps);
2367 
2368 	mutex_lock(&rps->lock);
2369 
2370 	if (rps_uses_slpc(rps)) {
2371 		/* RP limits have not been initialized yet for SLPC path */
2372 		struct intel_rps_freq_caps caps;
2373 
2374 		gen6_rps_get_freq_caps(rps, &caps);
2375 
2376 		intel_rps_set_manual(rps, true);
2377 		intel_uncore_write(uncore, GEN6_RPNSWREQ,
2378 				   ((caps.min_freq <<
2379 				   GEN9_SW_REQ_UNSLICE_RATIO_SHIFT) |
2380 				   GEN9_IGNORE_SLICE_RATIO));
2381 		intel_rps_set_manual(rps, false);
2382 	} else {
2383 		intel_rps_set(rps, rps->min_freq);
2384 	}
2385 
2386 	mutex_unlock(&rps->lock);
2387 }
2388 
2389 static u32 rps_read_mmio(struct intel_rps *rps, i915_reg_t reg32)
2390 {
2391 	struct intel_gt *gt = rps_to_gt(rps);
2392 	intel_wakeref_t wakeref;
2393 	u32 val;
2394 
2395 	with_intel_runtime_pm(gt->uncore->rpm, wakeref)
2396 		val = intel_uncore_read(gt->uncore, reg32);
2397 
2398 	return val;
2399 }
2400 
2401 bool rps_read_mask_mmio(struct intel_rps *rps,
2402 			i915_reg_t reg32, u32 mask)
2403 {
2404 	return rps_read_mmio(rps, reg32) & mask;
2405 }
2406 
2407 /* External interface for intel_ips.ko */
2408 
2409 static struct drm_i915_private __rcu *ips_mchdev;
2410 
2411 /**
2412  * Tells the intel_ips driver that the i915 driver is now loaded, if
2413  * IPS got loaded first.
2414  *
2415  * This awkward dance is so that neither module has to depend on the
2416  * other in order for IPS to do the appropriate communication of
2417  * GPU turbo limits to i915.
2418  */
2419 static void
2420 ips_ping_for_i915_load(void)
2421 {
2422 #ifdef __linux__
2423 	void (*link)(void);
2424 
2425 	link = symbol_get(ips_link_to_i915_driver);
2426 	if (link) {
2427 		link();
2428 		symbol_put(ips_link_to_i915_driver);
2429 	}
2430 #endif
2431 }
2432 
2433 void intel_rps_driver_register(struct intel_rps *rps)
2434 {
2435 	struct intel_gt *gt = rps_to_gt(rps);
2436 
2437 	/*
2438 	 * We only register the i915 ips part with intel-ips once everything is
2439 	 * set up, to avoid intel-ips sneaking in and reading bogus values.
2440 	 */
2441 	if (GRAPHICS_VER(gt->i915) == 5) {
2442 		GEM_BUG_ON(ips_mchdev);
2443 		rcu_assign_pointer(ips_mchdev, gt->i915);
2444 		ips_ping_for_i915_load();
2445 	}
2446 }
2447 
2448 void intel_rps_driver_unregister(struct intel_rps *rps)
2449 {
2450 	if (rcu_access_pointer(ips_mchdev) == rps_to_i915(rps))
2451 		rcu_assign_pointer(ips_mchdev, NULL);
2452 }
2453 
2454 static struct drm_i915_private *mchdev_get(void)
2455 {
2456 	struct drm_i915_private *i915;
2457 
2458 	rcu_read_lock();
2459 	i915 = rcu_dereference(ips_mchdev);
2460 	if (i915 && !kref_get_unless_zero(&i915->drm.ref))
2461 		i915 = NULL;
2462 	rcu_read_unlock();
2463 
2464 	return i915;
2465 }
2466 
2467 /**
2468  * i915_read_mch_val - return value for IPS use
2469  *
2470  * Calculate and return a value for the IPS driver to use when deciding whether
2471  * we have thermal and power headroom to increase CPU or GPU power budget.
2472  */
2473 unsigned long i915_read_mch_val(void)
2474 {
2475 	struct drm_i915_private *i915;
2476 	unsigned long chipset_val = 0;
2477 	unsigned long graphics_val = 0;
2478 	intel_wakeref_t wakeref;
2479 
2480 	i915 = mchdev_get();
2481 	if (!i915)
2482 		return 0;
2483 
2484 	with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
2485 		struct intel_ips *ips = &to_gt(i915)->rps.ips;
2486 
2487 		spin_lock_irq(&mchdev_lock);
2488 		chipset_val = __ips_chipset_val(ips);
2489 		graphics_val = __ips_gfx_val(ips);
2490 		spin_unlock_irq(&mchdev_lock);
2491 	}
2492 
2493 	drm_dev_put(&i915->drm);
2494 	return chipset_val + graphics_val;
2495 }
2496 EXPORT_SYMBOL_GPL(i915_read_mch_val);
2497 
2498 /**
2499  * i915_gpu_raise - raise GPU frequency limit
2500  *
2501  * Raise the limit; IPS indicates we have thermal headroom.
2502  */
2503 bool i915_gpu_raise(void)
2504 {
2505 	struct drm_i915_private *i915;
2506 	struct intel_rps *rps;
2507 
2508 	i915 = mchdev_get();
2509 	if (!i915)
2510 		return false;
2511 
2512 	rps = &to_gt(i915)->rps;
2513 
2514 	spin_lock_irq(&mchdev_lock);
2515 	if (rps->max_freq_softlimit < rps->max_freq)
2516 		rps->max_freq_softlimit++;
2517 	spin_unlock_irq(&mchdev_lock);
2518 
2519 	drm_dev_put(&i915->drm);
2520 	return true;
2521 }
2522 EXPORT_SYMBOL_GPL(i915_gpu_raise);
2523 
2524 /**
2525  * i915_gpu_lower - lower GPU frequency limit
2526  *
2527  * IPS indicates we're close to a thermal limit, so throttle back the GPU
2528  * frequency maximum.
2529  */
2530 bool i915_gpu_lower(void)
2531 {
2532 	struct drm_i915_private *i915;
2533 	struct intel_rps *rps;
2534 
2535 	i915 = mchdev_get();
2536 	if (!i915)
2537 		return false;
2538 
2539 	rps = &to_gt(i915)->rps;
2540 
2541 	spin_lock_irq(&mchdev_lock);
2542 	if (rps->max_freq_softlimit > rps->min_freq)
2543 		rps->max_freq_softlimit--;
2544 	spin_unlock_irq(&mchdev_lock);
2545 
2546 	drm_dev_put(&i915->drm);
2547 	return true;
2548 }
2549 EXPORT_SYMBOL_GPL(i915_gpu_lower);
2550 
2551 /**
2552  * i915_gpu_busy - indicate GPU business to IPS
2553  *
2554  * Tell the IPS driver whether or not the GPU is busy.
2555  */
2556 bool i915_gpu_busy(void)
2557 {
2558 	struct drm_i915_private *i915;
2559 	bool ret;
2560 
2561 	i915 = mchdev_get();
2562 	if (!i915)
2563 		return false;
2564 
2565 	ret = to_gt(i915)->awake;
2566 
2567 	drm_dev_put(&i915->drm);
2568 	return ret;
2569 }
2570 EXPORT_SYMBOL_GPL(i915_gpu_busy);
2571 
2572 /**
2573  * i915_gpu_turbo_disable - disable graphics turbo
2574  *
2575  * Disable graphics turbo by resetting the max frequency and setting the
2576  * current frequency to the default.
2577  */
2578 bool i915_gpu_turbo_disable(void)
2579 {
2580 	struct drm_i915_private *i915;
2581 	struct intel_rps *rps;
2582 	bool ret;
2583 
2584 	i915 = mchdev_get();
2585 	if (!i915)
2586 		return false;
2587 
2588 	rps = &to_gt(i915)->rps;
2589 
2590 	spin_lock_irq(&mchdev_lock);
2591 	rps->max_freq_softlimit = rps->min_freq;
2592 	ret = !__gen5_rps_set(&to_gt(i915)->rps, rps->min_freq);
2593 	spin_unlock_irq(&mchdev_lock);
2594 
2595 	drm_dev_put(&i915->drm);
2596 	return ret;
2597 }
2598 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
2599 
2600 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2601 #include "selftest_rps.c"
2602 #include "selftest_slpc.c"
2603 #endif
2604