xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_rps.c (revision fb574682)
1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include <drm/i915_drm.h>
8 
9 #include "i915_drv.h"
10 #include "intel_gt.h"
11 #include "intel_gt_clock_utils.h"
12 #include "intel_gt_irq.h"
13 #include "intel_gt_pm_irq.h"
14 #include "intel_rps.h"
15 #include "intel_sideband.h"
16 #include "../../../platform/x86/intel_ips.h"
17 
18 #define BUSY_MAX_EI	20u /* ms */
19 
20 /*
21  * Lock protecting IPS related data structures
22  */
23 static DEFINE_SPINLOCK(mchdev_lock);
24 
25 static struct intel_gt *rps_to_gt(struct intel_rps *rps)
26 {
27 	return container_of(rps, struct intel_gt, rps);
28 }
29 
30 static struct drm_i915_private *rps_to_i915(struct intel_rps *rps)
31 {
32 	return rps_to_gt(rps)->i915;
33 }
34 
35 static struct intel_uncore *rps_to_uncore(struct intel_rps *rps)
36 {
37 	return rps_to_gt(rps)->uncore;
38 }
39 
40 static u32 rps_pm_sanitize_mask(struct intel_rps *rps, u32 mask)
41 {
42 	return mask & ~rps->pm_intrmsk_mbz;
43 }
44 
45 static inline void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val)
46 {
47 	intel_uncore_write_fw(uncore, reg, val);
48 }
49 
50 static void rps_timer(struct timer_list *t)
51 {
52 	struct intel_rps *rps = from_timer(rps, t, timer);
53 	struct intel_engine_cs *engine;
54 	enum intel_engine_id id;
55 	s64 max_busy[3] = {};
56 	ktime_t dt, last;
57 
58 	for_each_engine(engine, rps_to_gt(rps), id) {
59 		s64 busy;
60 		int i;
61 
62 		dt = intel_engine_get_busy_time(engine);
63 		last = engine->stats.rps;
64 		engine->stats.rps = dt;
65 
66 		busy = ktime_to_ns(ktime_sub(dt, last));
67 		for (i = 0; i < ARRAY_SIZE(max_busy); i++) {
68 			if (busy > max_busy[i])
69 				swap(busy, max_busy[i]);
70 		}
71 	}
72 
73 	dt = ktime_get();
74 	last = rps->pm_timestamp;
75 	rps->pm_timestamp = dt;
76 
77 	if (intel_rps_is_active(rps)) {
78 		s64 busy;
79 		int i;
80 
81 		dt = ktime_sub(dt, last);
82 
83 		/*
84 		 * Our goal is to evaluate each engine independently, so we run
85 		 * at the lowest clocks required to sustain the heaviest
86 		 * workload. However, a task may be split into sequential
87 		 * dependent operations across a set of engines, such that
88 		 * the independent contributions do not account for high load,
89 		 * but overall the task is GPU bound. For example, consider
90 		 * video decode on vcs followed by colour post-processing
91 		 * on vecs, followed by general post-processing on rcs.
92 		 * Since multi-engines being active does imply a single
93 		 * continuous workload across all engines, we hedge our
94 		 * bets by only contributing a factor of the distributed
95 		 * load into our busyness calculation.
96 		 */
97 		busy = max_busy[0];
98 		for (i = 1; i < ARRAY_SIZE(max_busy); i++) {
99 			if (!max_busy[i])
100 				break;
101 
102 			busy += div_u64(max_busy[i], 1 << i);
103 		}
104 		GT_TRACE(rps_to_gt(rps),
105 			 "busy:%lld [%d%%], max:[%lld, %lld, %lld], interval:%d\n",
106 			 busy, (int)div64_u64(100 * busy, dt),
107 			 max_busy[0], max_busy[1], max_busy[2],
108 			 rps->pm_interval);
109 
110 		if (100 * busy > rps->power.up_threshold * dt &&
111 		    rps->cur_freq < rps->max_freq_softlimit) {
112 			rps->pm_iir |= GEN6_PM_RP_UP_THRESHOLD;
113 			rps->pm_interval = 1;
114 			schedule_work(&rps->work);
115 		} else if (100 * busy < rps->power.down_threshold * dt &&
116 			   rps->cur_freq > rps->min_freq_softlimit) {
117 			rps->pm_iir |= GEN6_PM_RP_DOWN_THRESHOLD;
118 			rps->pm_interval = 1;
119 			schedule_work(&rps->work);
120 		} else {
121 			rps->last_adj = 0;
122 		}
123 
124 		mod_timer(&rps->timer,
125 			  jiffies + msecs_to_jiffies(rps->pm_interval));
126 		rps->pm_interval = min(rps->pm_interval * 2, BUSY_MAX_EI);
127 	}
128 }
129 
130 static void rps_start_timer(struct intel_rps *rps)
131 {
132 	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
133 	rps->pm_interval = 1;
134 	mod_timer(&rps->timer, jiffies + 1);
135 }
136 
137 static void rps_stop_timer(struct intel_rps *rps)
138 {
139 	del_timer_sync(&rps->timer);
140 	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
141 	cancel_work_sync(&rps->work);
142 }
143 
144 static u32 rps_pm_mask(struct intel_rps *rps, u8 val)
145 {
146 	u32 mask = 0;
147 
148 	/* We use UP_EI_EXPIRED interrupts for both up/down in manual mode */
149 	if (val > rps->min_freq_softlimit)
150 		mask |= (GEN6_PM_RP_UP_EI_EXPIRED |
151 			 GEN6_PM_RP_DOWN_THRESHOLD |
152 			 GEN6_PM_RP_DOWN_TIMEOUT);
153 
154 	if (val < rps->max_freq_softlimit)
155 		mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
156 
157 	mask &= rps->pm_events;
158 
159 	return rps_pm_sanitize_mask(rps, ~mask);
160 }
161 
162 static void rps_reset_ei(struct intel_rps *rps)
163 {
164 	memset(&rps->ei, 0, sizeof(rps->ei));
165 }
166 
167 static void rps_enable_interrupts(struct intel_rps *rps)
168 {
169 	struct intel_gt *gt = rps_to_gt(rps);
170 
171 	GT_TRACE(gt, "interrupts:on rps->pm_events: %x, rps_pm_mask:%x\n",
172 		 rps->pm_events, rps_pm_mask(rps, rps->last_freq));
173 
174 	rps_reset_ei(rps);
175 
176 	spin_lock_irq(&gt->irq_lock);
177 	gen6_gt_pm_enable_irq(gt, rps->pm_events);
178 	spin_unlock_irq(&gt->irq_lock);
179 
180 	intel_uncore_write(gt->uncore,
181 			   GEN6_PMINTRMSK, rps_pm_mask(rps, rps->last_freq));
182 }
183 
184 static void gen6_rps_reset_interrupts(struct intel_rps *rps)
185 {
186 	gen6_gt_pm_reset_iir(rps_to_gt(rps), GEN6_PM_RPS_EVENTS);
187 }
188 
189 static void gen11_rps_reset_interrupts(struct intel_rps *rps)
190 {
191 	while (gen11_gt_reset_one_iir(rps_to_gt(rps), 0, GEN11_GTPM))
192 		;
193 }
194 
195 static void rps_reset_interrupts(struct intel_rps *rps)
196 {
197 	struct intel_gt *gt = rps_to_gt(rps);
198 
199 	spin_lock_irq(&gt->irq_lock);
200 	if (INTEL_GEN(gt->i915) >= 11)
201 		gen11_rps_reset_interrupts(rps);
202 	else
203 		gen6_rps_reset_interrupts(rps);
204 
205 	rps->pm_iir = 0;
206 	spin_unlock_irq(&gt->irq_lock);
207 }
208 
209 static void rps_disable_interrupts(struct intel_rps *rps)
210 {
211 	struct intel_gt *gt = rps_to_gt(rps);
212 
213 	intel_uncore_write(gt->uncore,
214 			   GEN6_PMINTRMSK, rps_pm_sanitize_mask(rps, ~0u));
215 
216 	spin_lock_irq(&gt->irq_lock);
217 	gen6_gt_pm_disable_irq(gt, GEN6_PM_RPS_EVENTS);
218 	spin_unlock_irq(&gt->irq_lock);
219 
220 	intel_synchronize_irq(gt->i915);
221 
222 	/*
223 	 * Now that we will not be generating any more work, flush any
224 	 * outstanding tasks. As we are called on the RPS idle path,
225 	 * we will reset the GPU to minimum frequencies, so the current
226 	 * state of the worker can be discarded.
227 	 */
228 	cancel_work_sync(&rps->work);
229 
230 	rps_reset_interrupts(rps);
231 	GT_TRACE(gt, "interrupts:off\n");
232 }
233 
234 static const struct cparams {
235 	u16 i;
236 	u16 t;
237 	u16 m;
238 	u16 c;
239 } cparams[] = {
240 	{ 1, 1333, 301, 28664 },
241 	{ 1, 1066, 294, 24460 },
242 	{ 1, 800, 294, 25192 },
243 	{ 0, 1333, 276, 27605 },
244 	{ 0, 1066, 276, 27605 },
245 	{ 0, 800, 231, 23784 },
246 };
247 
248 static void gen5_rps_init(struct intel_rps *rps)
249 {
250 	struct drm_i915_private *i915 = rps_to_i915(rps);
251 	struct intel_uncore *uncore = rps_to_uncore(rps);
252 	u8 fmax, fmin, fstart;
253 	u32 rgvmodectl;
254 	int c_m, i;
255 
256 	if (i915->fsb_freq <= 3200)
257 		c_m = 0;
258 	else if (i915->fsb_freq <= 4800)
259 		c_m = 1;
260 	else
261 		c_m = 2;
262 
263 	for (i = 0; i < ARRAY_SIZE(cparams); i++) {
264 		if (cparams[i].i == c_m && cparams[i].t == i915->mem_freq) {
265 			rps->ips.m = cparams[i].m;
266 			rps->ips.c = cparams[i].c;
267 			break;
268 		}
269 	}
270 
271 	rgvmodectl = intel_uncore_read(uncore, MEMMODECTL);
272 
273 	/* Set up min, max, and cur for interrupt handling */
274 	fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
275 	fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
276 	fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
277 		MEMMODE_FSTART_SHIFT;
278 	drm_dbg(&i915->drm, "fmax: %d, fmin: %d, fstart: %d\n",
279 		fmax, fmin, fstart);
280 
281 	rps->min_freq = fmax;
282 	rps->efficient_freq = fstart;
283 	rps->max_freq = fmin;
284 }
285 
286 static unsigned long
287 __ips_chipset_val(struct intel_ips *ips)
288 {
289 	struct intel_uncore *uncore =
290 		rps_to_uncore(container_of(ips, struct intel_rps, ips));
291 	unsigned long now = jiffies_to_msecs(jiffies), dt;
292 	unsigned long result;
293 	u64 total, delta;
294 
295 	lockdep_assert_held(&mchdev_lock);
296 
297 	/*
298 	 * Prevent division-by-zero if we are asking too fast.
299 	 * Also, we don't get interesting results if we are polling
300 	 * faster than once in 10ms, so just return the saved value
301 	 * in such cases.
302 	 */
303 	dt = now - ips->last_time1;
304 	if (dt <= 10)
305 		return ips->chipset_power;
306 
307 	/* FIXME: handle per-counter overflow */
308 	total = intel_uncore_read(uncore, DMIEC);
309 	total += intel_uncore_read(uncore, DDREC);
310 	total += intel_uncore_read(uncore, CSIEC);
311 
312 	delta = total - ips->last_count1;
313 
314 	result = div_u64(div_u64(ips->m * delta, dt) + ips->c, 10);
315 
316 	ips->last_count1 = total;
317 	ips->last_time1 = now;
318 
319 	ips->chipset_power = result;
320 
321 	return result;
322 }
323 
324 static unsigned long ips_mch_val(struct intel_uncore *uncore)
325 {
326 	unsigned int m, x, b;
327 	u32 tsfs;
328 
329 	tsfs = intel_uncore_read(uncore, TSFS);
330 	x = intel_uncore_read8(uncore, TR1);
331 
332 	b = tsfs & TSFS_INTR_MASK;
333 	m = (tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT;
334 
335 	return m * x / 127 - b;
336 }
337 
338 static int _pxvid_to_vd(u8 pxvid)
339 {
340 	if (pxvid == 0)
341 		return 0;
342 
343 	if (pxvid >= 8 && pxvid < 31)
344 		pxvid = 31;
345 
346 	return (pxvid + 2) * 125;
347 }
348 
349 static u32 pvid_to_extvid(struct drm_i915_private *i915, u8 pxvid)
350 {
351 	const int vd = _pxvid_to_vd(pxvid);
352 
353 	if (INTEL_INFO(i915)->is_mobile)
354 		return max(vd - 1125, 0);
355 
356 	return vd;
357 }
358 
359 static void __gen5_ips_update(struct intel_ips *ips)
360 {
361 	struct intel_uncore *uncore =
362 		rps_to_uncore(container_of(ips, struct intel_rps, ips));
363 	u64 now, delta, dt;
364 	u32 count;
365 
366 	lockdep_assert_held(&mchdev_lock);
367 
368 	now = ktime_get_raw_ns();
369 	dt = now - ips->last_time2;
370 	do_div(dt, NSEC_PER_MSEC);
371 
372 	/* Don't divide by 0 */
373 	if (dt <= 10)
374 		return;
375 
376 	count = intel_uncore_read(uncore, GFXEC);
377 	delta = count - ips->last_count2;
378 
379 	ips->last_count2 = count;
380 	ips->last_time2 = now;
381 
382 	/* More magic constants... */
383 	ips->gfx_power = div_u64(delta * 1181, dt * 10);
384 }
385 
386 static void gen5_rps_update(struct intel_rps *rps)
387 {
388 	spin_lock_irq(&mchdev_lock);
389 	__gen5_ips_update(&rps->ips);
390 	spin_unlock_irq(&mchdev_lock);
391 }
392 
393 static bool gen5_rps_set(struct intel_rps *rps, u8 val)
394 {
395 	struct intel_uncore *uncore = rps_to_uncore(rps);
396 	u16 rgvswctl;
397 
398 	lockdep_assert_held(&mchdev_lock);
399 
400 	rgvswctl = intel_uncore_read16(uncore, MEMSWCTL);
401 	if (rgvswctl & MEMCTL_CMD_STS) {
402 		DRM_DEBUG("gpu busy, RCS change rejected\n");
403 		return false; /* still busy with another command */
404 	}
405 
406 	/* Invert the frequency bin into an ips delay */
407 	val = rps->max_freq - val;
408 	val = rps->min_freq + val;
409 
410 	rgvswctl =
411 		(MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
412 		(val << MEMCTL_FREQ_SHIFT) |
413 		MEMCTL_SFCAVM;
414 	intel_uncore_write16(uncore, MEMSWCTL, rgvswctl);
415 	intel_uncore_posting_read16(uncore, MEMSWCTL);
416 
417 	rgvswctl |= MEMCTL_CMD_STS;
418 	intel_uncore_write16(uncore, MEMSWCTL, rgvswctl);
419 
420 	return true;
421 }
422 
423 static unsigned long intel_pxfreq(u32 vidfreq)
424 {
425 	int div = (vidfreq & 0x3f0000) >> 16;
426 	int post = (vidfreq & 0x3000) >> 12;
427 	int pre = (vidfreq & 0x7);
428 
429 	if (!pre)
430 		return 0;
431 
432 	return div * 133333 / (pre << post);
433 }
434 
435 static unsigned int init_emon(struct intel_uncore *uncore)
436 {
437 	u8 pxw[16];
438 	int i;
439 
440 	/* Disable to program */
441 	intel_uncore_write(uncore, ECR, 0);
442 	intel_uncore_posting_read(uncore, ECR);
443 
444 	/* Program energy weights for various events */
445 	intel_uncore_write(uncore, SDEW, 0x15040d00);
446 	intel_uncore_write(uncore, CSIEW0, 0x007f0000);
447 	intel_uncore_write(uncore, CSIEW1, 0x1e220004);
448 	intel_uncore_write(uncore, CSIEW2, 0x04000004);
449 
450 	for (i = 0; i < 5; i++)
451 		intel_uncore_write(uncore, PEW(i), 0);
452 	for (i = 0; i < 3; i++)
453 		intel_uncore_write(uncore, DEW(i), 0);
454 
455 	/* Program P-state weights to account for frequency power adjustment */
456 	for (i = 0; i < 16; i++) {
457 		u32 pxvidfreq = intel_uncore_read(uncore, PXVFREQ(i));
458 		unsigned int freq = intel_pxfreq(pxvidfreq);
459 		unsigned int vid =
460 			(pxvidfreq & PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT;
461 		unsigned int val;
462 
463 		val = vid * vid * freq / 1000 * 255;
464 		val /= 127 * 127 * 900;
465 
466 		pxw[i] = val;
467 	}
468 	/* Render standby states get 0 weight */
469 	pxw[14] = 0;
470 	pxw[15] = 0;
471 
472 	for (i = 0; i < 4; i++) {
473 		intel_uncore_write(uncore, PXW(i),
474 				   pxw[i * 4 + 0] << 24 |
475 				   pxw[i * 4 + 1] << 16 |
476 				   pxw[i * 4 + 2] <<  8 |
477 				   pxw[i * 4 + 3] <<  0);
478 	}
479 
480 	/* Adjust magic regs to magic values (more experimental results) */
481 	intel_uncore_write(uncore, OGW0, 0);
482 	intel_uncore_write(uncore, OGW1, 0);
483 	intel_uncore_write(uncore, EG0, 0x00007f00);
484 	intel_uncore_write(uncore, EG1, 0x0000000e);
485 	intel_uncore_write(uncore, EG2, 0x000e0000);
486 	intel_uncore_write(uncore, EG3, 0x68000300);
487 	intel_uncore_write(uncore, EG4, 0x42000000);
488 	intel_uncore_write(uncore, EG5, 0x00140031);
489 	intel_uncore_write(uncore, EG6, 0);
490 	intel_uncore_write(uncore, EG7, 0);
491 
492 	for (i = 0; i < 8; i++)
493 		intel_uncore_write(uncore, PXWL(i), 0);
494 
495 	/* Enable PMON + select events */
496 	intel_uncore_write(uncore, ECR, 0x80000019);
497 
498 	return intel_uncore_read(uncore, LCFUSE02) & LCFUSE_HIV_MASK;
499 }
500 
501 static bool gen5_rps_enable(struct intel_rps *rps)
502 {
503 	struct intel_uncore *uncore = rps_to_uncore(rps);
504 	u8 fstart, vstart;
505 	u32 rgvmodectl;
506 
507 	spin_lock_irq(&mchdev_lock);
508 
509 	rgvmodectl = intel_uncore_read(uncore, MEMMODECTL);
510 
511 	/* Enable temp reporting */
512 	intel_uncore_write16(uncore, PMMISC,
513 			     intel_uncore_read16(uncore, PMMISC) | MCPPCE_EN);
514 	intel_uncore_write16(uncore, TSC1,
515 			     intel_uncore_read16(uncore, TSC1) | TSE);
516 
517 	/* 100ms RC evaluation intervals */
518 	intel_uncore_write(uncore, RCUPEI, 100000);
519 	intel_uncore_write(uncore, RCDNEI, 100000);
520 
521 	/* Set max/min thresholds to 90ms and 80ms respectively */
522 	intel_uncore_write(uncore, RCBMAXAVG, 90000);
523 	intel_uncore_write(uncore, RCBMINAVG, 80000);
524 
525 	intel_uncore_write(uncore, MEMIHYST, 1);
526 
527 	/* Set up min, max, and cur for interrupt handling */
528 	fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
529 		MEMMODE_FSTART_SHIFT;
530 
531 	vstart = (intel_uncore_read(uncore, PXVFREQ(fstart)) &
532 		  PXVFREQ_PX_MASK) >> PXVFREQ_PX_SHIFT;
533 
534 	intel_uncore_write(uncore,
535 			   MEMINTREN,
536 			   MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
537 
538 	intel_uncore_write(uncore, VIDSTART, vstart);
539 	intel_uncore_posting_read(uncore, VIDSTART);
540 
541 	rgvmodectl |= MEMMODE_SWMODE_EN;
542 	intel_uncore_write(uncore, MEMMODECTL, rgvmodectl);
543 
544 	if (wait_for_atomic((intel_uncore_read(uncore, MEMSWCTL) &
545 			     MEMCTL_CMD_STS) == 0, 10))
546 		drm_err(&uncore->i915->drm,
547 			"stuck trying to change perf mode\n");
548 	mdelay(1);
549 
550 	gen5_rps_set(rps, rps->cur_freq);
551 
552 	rps->ips.last_count1 = intel_uncore_read(uncore, DMIEC);
553 	rps->ips.last_count1 += intel_uncore_read(uncore, DDREC);
554 	rps->ips.last_count1 += intel_uncore_read(uncore, CSIEC);
555 	rps->ips.last_time1 = jiffies_to_msecs(jiffies);
556 
557 	rps->ips.last_count2 = intel_uncore_read(uncore, GFXEC);
558 	rps->ips.last_time2 = ktime_get_raw_ns();
559 
560 	spin_unlock_irq(&mchdev_lock);
561 
562 	rps->ips.corr = init_emon(uncore);
563 
564 	return true;
565 }
566 
567 static void gen5_rps_disable(struct intel_rps *rps)
568 {
569 	struct intel_uncore *uncore = rps_to_uncore(rps);
570 	u16 rgvswctl;
571 
572 	spin_lock_irq(&mchdev_lock);
573 
574 	rgvswctl = intel_uncore_read16(uncore, MEMSWCTL);
575 
576 	/* Ack interrupts, disable EFC interrupt */
577 	intel_uncore_write(uncore, MEMINTREN,
578 			   intel_uncore_read(uncore, MEMINTREN) &
579 			   ~MEMINT_EVAL_CHG_EN);
580 	intel_uncore_write(uncore, MEMINTRSTS, MEMINT_EVAL_CHG);
581 	intel_uncore_write(uncore, DEIER,
582 			   intel_uncore_read(uncore, DEIER) & ~DE_PCU_EVENT);
583 	intel_uncore_write(uncore, DEIIR, DE_PCU_EVENT);
584 	intel_uncore_write(uncore, DEIMR,
585 			   intel_uncore_read(uncore, DEIMR) | DE_PCU_EVENT);
586 
587 	/* Go back to the starting frequency */
588 	gen5_rps_set(rps, rps->idle_freq);
589 	mdelay(1);
590 	rgvswctl |= MEMCTL_CMD_STS;
591 	intel_uncore_write(uncore, MEMSWCTL, rgvswctl);
592 	mdelay(1);
593 
594 	spin_unlock_irq(&mchdev_lock);
595 }
596 
597 static u32 rps_limits(struct intel_rps *rps, u8 val)
598 {
599 	u32 limits;
600 
601 	/*
602 	 * Only set the down limit when we've reached the lowest level to avoid
603 	 * getting more interrupts, otherwise leave this clear. This prevents a
604 	 * race in the hw when coming out of rc6: There's a tiny window where
605 	 * the hw runs at the minimal clock before selecting the desired
606 	 * frequency, if the down threshold expires in that window we will not
607 	 * receive a down interrupt.
608 	 */
609 	if (INTEL_GEN(rps_to_i915(rps)) >= 9) {
610 		limits = rps->max_freq_softlimit << 23;
611 		if (val <= rps->min_freq_softlimit)
612 			limits |= rps->min_freq_softlimit << 14;
613 	} else {
614 		limits = rps->max_freq_softlimit << 24;
615 		if (val <= rps->min_freq_softlimit)
616 			limits |= rps->min_freq_softlimit << 16;
617 	}
618 
619 	return limits;
620 }
621 
622 static void rps_set_power(struct intel_rps *rps, int new_power)
623 {
624 	struct intel_gt *gt = rps_to_gt(rps);
625 	struct intel_uncore *uncore = gt->uncore;
626 	u32 threshold_up = 0, threshold_down = 0; /* in % */
627 	u32 ei_up = 0, ei_down = 0;
628 
629 	lockdep_assert_held(&rps->power.mutex);
630 
631 	if (new_power == rps->power.mode)
632 		return;
633 
634 	threshold_up = 95;
635 	threshold_down = 85;
636 
637 	/* Note the units here are not exactly 1us, but 1280ns. */
638 	switch (new_power) {
639 	case LOW_POWER:
640 		ei_up = 16000;
641 		ei_down = 32000;
642 		break;
643 
644 	case BETWEEN:
645 		ei_up = 13000;
646 		ei_down = 32000;
647 		break;
648 
649 	case HIGH_POWER:
650 		ei_up = 10000;
651 		ei_down = 32000;
652 		break;
653 	}
654 
655 	/* When byt can survive without system hang with dynamic
656 	 * sw freq adjustments, this restriction can be lifted.
657 	 */
658 	if (IS_VALLEYVIEW(gt->i915))
659 		goto skip_hw_write;
660 
661 	GT_TRACE(gt,
662 		 "changing power mode [%d], up %d%% @ %dus, down %d%% @ %dus\n",
663 		 new_power, threshold_up, ei_up, threshold_down, ei_down);
664 
665 	set(uncore, GEN6_RP_UP_EI,
666 	    intel_gt_ns_to_pm_interval(gt, ei_up * 1000));
667 	set(uncore, GEN6_RP_UP_THRESHOLD,
668 	    intel_gt_ns_to_pm_interval(gt, ei_up * threshold_up * 10));
669 
670 	set(uncore, GEN6_RP_DOWN_EI,
671 	    intel_gt_ns_to_pm_interval(gt, ei_down * 1000));
672 	set(uncore, GEN6_RP_DOWN_THRESHOLD,
673 	    intel_gt_ns_to_pm_interval(gt, ei_down * threshold_down * 10));
674 
675 	set(uncore, GEN6_RP_CONTROL,
676 	    (INTEL_GEN(gt->i915) > 9 ? 0 : GEN6_RP_MEDIA_TURBO) |
677 	    GEN6_RP_MEDIA_HW_NORMAL_MODE |
678 	    GEN6_RP_MEDIA_IS_GFX |
679 	    GEN6_RP_ENABLE |
680 	    GEN6_RP_UP_BUSY_AVG |
681 	    GEN6_RP_DOWN_IDLE_AVG);
682 
683 skip_hw_write:
684 	rps->power.mode = new_power;
685 	rps->power.up_threshold = threshold_up;
686 	rps->power.down_threshold = threshold_down;
687 }
688 
689 static void gen6_rps_set_thresholds(struct intel_rps *rps, u8 val)
690 {
691 	int new_power;
692 
693 	new_power = rps->power.mode;
694 	switch (rps->power.mode) {
695 	case LOW_POWER:
696 		if (val > rps->efficient_freq + 1 &&
697 		    val > rps->cur_freq)
698 			new_power = BETWEEN;
699 		break;
700 
701 	case BETWEEN:
702 		if (val <= rps->efficient_freq &&
703 		    val < rps->cur_freq)
704 			new_power = LOW_POWER;
705 		else if (val >= rps->rp0_freq &&
706 			 val > rps->cur_freq)
707 			new_power = HIGH_POWER;
708 		break;
709 
710 	case HIGH_POWER:
711 		if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 &&
712 		    val < rps->cur_freq)
713 			new_power = BETWEEN;
714 		break;
715 	}
716 	/* Max/min bins are special */
717 	if (val <= rps->min_freq_softlimit)
718 		new_power = LOW_POWER;
719 	if (val >= rps->max_freq_softlimit)
720 		new_power = HIGH_POWER;
721 
722 	mutex_lock(&rps->power.mutex);
723 	if (rps->power.interactive)
724 		new_power = HIGH_POWER;
725 	rps_set_power(rps, new_power);
726 	mutex_unlock(&rps->power.mutex);
727 }
728 
729 void intel_rps_mark_interactive(struct intel_rps *rps, bool interactive)
730 {
731 	GT_TRACE(rps_to_gt(rps), "mark interactive: %s\n", yesno(interactive));
732 
733 	mutex_lock(&rps->power.mutex);
734 	if (interactive) {
735 		if (!rps->power.interactive++ && intel_rps_is_active(rps))
736 			rps_set_power(rps, HIGH_POWER);
737 	} else {
738 		GEM_BUG_ON(!rps->power.interactive);
739 		rps->power.interactive--;
740 	}
741 	mutex_unlock(&rps->power.mutex);
742 }
743 
744 static int gen6_rps_set(struct intel_rps *rps, u8 val)
745 {
746 	struct intel_uncore *uncore = rps_to_uncore(rps);
747 	struct drm_i915_private *i915 = rps_to_i915(rps);
748 	u32 swreq;
749 
750 	if (INTEL_GEN(i915) >= 9)
751 		swreq = GEN9_FREQUENCY(val);
752 	else if (IS_HASWELL(i915) || IS_BROADWELL(i915))
753 		swreq = HSW_FREQUENCY(val);
754 	else
755 		swreq = (GEN6_FREQUENCY(val) |
756 			 GEN6_OFFSET(0) |
757 			 GEN6_AGGRESSIVE_TURBO);
758 	set(uncore, GEN6_RPNSWREQ, swreq);
759 
760 	GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d, swreq:%x\n",
761 		 val, intel_gpu_freq(rps, val), swreq);
762 
763 	return 0;
764 }
765 
766 static int vlv_rps_set(struct intel_rps *rps, u8 val)
767 {
768 	struct drm_i915_private *i915 = rps_to_i915(rps);
769 	int err;
770 
771 	vlv_punit_get(i915);
772 	err = vlv_punit_write(i915, PUNIT_REG_GPU_FREQ_REQ, val);
773 	vlv_punit_put(i915);
774 
775 	GT_TRACE(rps_to_gt(rps), "set val:%x, freq:%d\n",
776 		 val, intel_gpu_freq(rps, val));
777 
778 	return err;
779 }
780 
781 static int rps_set(struct intel_rps *rps, u8 val, bool update)
782 {
783 	struct drm_i915_private *i915 = rps_to_i915(rps);
784 	int err;
785 
786 	if (INTEL_GEN(i915) < 6)
787 		return 0;
788 
789 	if (val == rps->last_freq)
790 		return 0;
791 
792 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
793 		err = vlv_rps_set(rps, val);
794 	else
795 		err = gen6_rps_set(rps, val);
796 	if (err)
797 		return err;
798 
799 	if (update)
800 		gen6_rps_set_thresholds(rps, val);
801 	rps->last_freq = val;
802 
803 	return 0;
804 }
805 
806 void intel_rps_unpark(struct intel_rps *rps)
807 {
808 	if (!intel_rps_is_enabled(rps))
809 		return;
810 
811 	GT_TRACE(rps_to_gt(rps), "unpark:%x\n", rps->cur_freq);
812 
813 	/*
814 	 * Use the user's desired frequency as a guide, but for better
815 	 * performance, jump directly to RPe as our starting frequency.
816 	 */
817 	mutex_lock(&rps->lock);
818 
819 	intel_rps_set_active(rps);
820 	intel_rps_set(rps,
821 		      clamp(rps->cur_freq,
822 			    rps->min_freq_softlimit,
823 			    rps->max_freq_softlimit));
824 
825 	mutex_unlock(&rps->lock);
826 
827 	rps->pm_iir = 0;
828 	if (intel_rps_has_interrupts(rps))
829 		rps_enable_interrupts(rps);
830 	if (intel_rps_uses_timer(rps))
831 		rps_start_timer(rps);
832 
833 	if (IS_GEN(rps_to_i915(rps), 5))
834 		gen5_rps_update(rps);
835 }
836 
837 void intel_rps_park(struct intel_rps *rps)
838 {
839 	int adj;
840 
841 	if (!intel_rps_clear_active(rps))
842 		return;
843 
844 	if (intel_rps_uses_timer(rps))
845 		rps_stop_timer(rps);
846 	if (intel_rps_has_interrupts(rps))
847 		rps_disable_interrupts(rps);
848 
849 	if (rps->last_freq <= rps->idle_freq)
850 		return;
851 
852 	/*
853 	 * The punit delays the write of the frequency and voltage until it
854 	 * determines the GPU is awake. During normal usage we don't want to
855 	 * waste power changing the frequency if the GPU is sleeping (rc6).
856 	 * However, the GPU and driver is now idle and we do not want to delay
857 	 * switching to minimum voltage (reducing power whilst idle) as we do
858 	 * not expect to be woken in the near future and so must flush the
859 	 * change by waking the device.
860 	 *
861 	 * We choose to take the media powerwell (either would do to trick the
862 	 * punit into committing the voltage change) as that takes a lot less
863 	 * power than the render powerwell.
864 	 */
865 	intel_uncore_forcewake_get(rps_to_uncore(rps), FORCEWAKE_MEDIA);
866 	rps_set(rps, rps->idle_freq, false);
867 	intel_uncore_forcewake_put(rps_to_uncore(rps), FORCEWAKE_MEDIA);
868 
869 	/*
870 	 * Since we will try and restart from the previously requested
871 	 * frequency on unparking, treat this idle point as a downclock
872 	 * interrupt and reduce the frequency for resume. If we park/unpark
873 	 * more frequently than the rps worker can run, we will not respond
874 	 * to any EI and never see a change in frequency.
875 	 *
876 	 * (Note we accommodate Cherryview's limitation of only using an
877 	 * even bin by applying it to all.)
878 	 */
879 	adj = rps->last_adj;
880 	if (adj < 0)
881 		adj *= 2;
882 	else /* CHV needs even encode values */
883 		adj = -2;
884 	rps->last_adj = adj;
885 	rps->cur_freq = max_t(int, rps->cur_freq + adj, rps->min_freq);
886 
887 	GT_TRACE(rps_to_gt(rps), "park:%x\n", rps->cur_freq);
888 }
889 
890 void intel_rps_boost(struct i915_request *rq)
891 {
892 	struct intel_rps *rps = &READ_ONCE(rq->engine)->gt->rps;
893 	unsigned long flags;
894 
895 	if (i915_request_signaled(rq) || !intel_rps_is_active(rps))
896 		return;
897 
898 	/* Serializes with i915_request_retire() */
899 	spin_lock_irqsave(&rq->lock, flags);
900 	if (!i915_request_has_waitboost(rq) &&
901 	    !dma_fence_is_signaled_locked(&rq->fence)) {
902 		set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags);
903 
904 		GT_TRACE(rps_to_gt(rps), "boost fence:%llx:%llx\n",
905 			 rq->fence.context, rq->fence.seqno);
906 
907 		if (!atomic_fetch_inc(&rps->num_waiters) &&
908 		    READ_ONCE(rps->cur_freq) < rps->boost_freq)
909 			schedule_work(&rps->work);
910 
911 		atomic_inc(&rps->boosts);
912 	}
913 	spin_unlock_irqrestore(&rq->lock, flags);
914 }
915 
916 int intel_rps_set(struct intel_rps *rps, u8 val)
917 {
918 	int err;
919 
920 	lockdep_assert_held(&rps->lock);
921 	GEM_BUG_ON(val > rps->max_freq);
922 	GEM_BUG_ON(val < rps->min_freq);
923 
924 	if (intel_rps_is_active(rps)) {
925 		err = rps_set(rps, val, true);
926 		if (err)
927 			return err;
928 
929 		/*
930 		 * Make sure we continue to get interrupts
931 		 * until we hit the minimum or maximum frequencies.
932 		 */
933 		if (intel_rps_has_interrupts(rps)) {
934 			struct intel_uncore *uncore = rps_to_uncore(rps);
935 
936 			set(uncore,
937 			    GEN6_RP_INTERRUPT_LIMITS, rps_limits(rps, val));
938 
939 			set(uncore, GEN6_PMINTRMSK, rps_pm_mask(rps, val));
940 		}
941 	}
942 
943 	rps->cur_freq = val;
944 	return 0;
945 }
946 
947 static void gen6_rps_init(struct intel_rps *rps)
948 {
949 	struct drm_i915_private *i915 = rps_to_i915(rps);
950 	struct intel_uncore *uncore = rps_to_uncore(rps);
951 
952 	/* All of these values are in units of 50MHz */
953 
954 	/* static values from HW: RP0 > RP1 > RPn (min_freq) */
955 	if (IS_GEN9_LP(i915)) {
956 		u32 rp_state_cap = intel_uncore_read(uncore, BXT_RP_STATE_CAP);
957 
958 		rps->rp0_freq = (rp_state_cap >> 16) & 0xff;
959 		rps->rp1_freq = (rp_state_cap >>  8) & 0xff;
960 		rps->min_freq = (rp_state_cap >>  0) & 0xff;
961 	} else {
962 		u32 rp_state_cap = intel_uncore_read(uncore, GEN6_RP_STATE_CAP);
963 
964 		rps->rp0_freq = (rp_state_cap >>  0) & 0xff;
965 		rps->rp1_freq = (rp_state_cap >>  8) & 0xff;
966 		rps->min_freq = (rp_state_cap >> 16) & 0xff;
967 	}
968 
969 	/* hw_max = RP0 until we check for overclocking */
970 	rps->max_freq = rps->rp0_freq;
971 
972 	rps->efficient_freq = rps->rp1_freq;
973 	if (IS_HASWELL(i915) || IS_BROADWELL(i915) ||
974 	    IS_GEN9_BC(i915) || INTEL_GEN(i915) >= 10) {
975 		u32 ddcc_status = 0;
976 
977 		if (sandybridge_pcode_read(i915,
978 					   HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
979 					   &ddcc_status, NULL) == 0)
980 			rps->efficient_freq =
981 				clamp_t(u8,
982 					(ddcc_status >> 8) & 0xff,
983 					rps->min_freq,
984 					rps->max_freq);
985 	}
986 
987 	if (IS_GEN9_BC(i915) || INTEL_GEN(i915) >= 10) {
988 		/* Store the frequency values in 16.66 MHZ units, which is
989 		 * the natural hardware unit for SKL
990 		 */
991 		rps->rp0_freq *= GEN9_FREQ_SCALER;
992 		rps->rp1_freq *= GEN9_FREQ_SCALER;
993 		rps->min_freq *= GEN9_FREQ_SCALER;
994 		rps->max_freq *= GEN9_FREQ_SCALER;
995 		rps->efficient_freq *= GEN9_FREQ_SCALER;
996 	}
997 }
998 
999 static bool rps_reset(struct intel_rps *rps)
1000 {
1001 	struct drm_i915_private *i915 = rps_to_i915(rps);
1002 
1003 	/* force a reset */
1004 	rps->power.mode = -1;
1005 	rps->last_freq = -1;
1006 
1007 	if (rps_set(rps, rps->min_freq, true)) {
1008 		drm_err(&i915->drm, "Failed to reset RPS to initial values\n");
1009 		return false;
1010 	}
1011 
1012 	rps->cur_freq = rps->min_freq;
1013 	return true;
1014 }
1015 
1016 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
1017 static bool gen9_rps_enable(struct intel_rps *rps)
1018 {
1019 	struct intel_gt *gt = rps_to_gt(rps);
1020 	struct intel_uncore *uncore = gt->uncore;
1021 
1022 	/* Program defaults and thresholds for RPS */
1023 	if (IS_GEN(gt->i915, 9))
1024 		intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ,
1025 				      GEN9_FREQUENCY(rps->rp1_freq));
1026 
1027 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 0xa);
1028 
1029 	rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD;
1030 
1031 	return rps_reset(rps);
1032 }
1033 
1034 static bool gen8_rps_enable(struct intel_rps *rps)
1035 {
1036 	struct intel_uncore *uncore = rps_to_uncore(rps);
1037 
1038 	intel_uncore_write_fw(uncore, GEN6_RC_VIDEO_FREQ,
1039 			      HSW_FREQUENCY(rps->rp1_freq));
1040 
1041 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1042 
1043 	rps->pm_events = GEN6_PM_RP_UP_THRESHOLD | GEN6_PM_RP_DOWN_THRESHOLD;
1044 
1045 	return rps_reset(rps);
1046 }
1047 
1048 static bool gen6_rps_enable(struct intel_rps *rps)
1049 {
1050 	struct intel_uncore *uncore = rps_to_uncore(rps);
1051 
1052 	/* Power down if completely idle for over 50ms */
1053 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 50000);
1054 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1055 
1056 	rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD |
1057 			  GEN6_PM_RP_DOWN_THRESHOLD |
1058 			  GEN6_PM_RP_DOWN_TIMEOUT);
1059 
1060 	return rps_reset(rps);
1061 }
1062 
1063 static int chv_rps_max_freq(struct intel_rps *rps)
1064 {
1065 	struct drm_i915_private *i915 = rps_to_i915(rps);
1066 	u32 val;
1067 
1068 	val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE);
1069 
1070 	switch (RUNTIME_INFO(i915)->sseu.eu_total) {
1071 	case 8:
1072 		/* (2 * 4) config */
1073 		val >>= FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT;
1074 		break;
1075 	case 12:
1076 		/* (2 * 6) config */
1077 		val >>= FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT;
1078 		break;
1079 	case 16:
1080 		/* (2 * 8) config */
1081 	default:
1082 		/* Setting (2 * 8) Min RP0 for any other combination */
1083 		val >>= FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT;
1084 		break;
1085 	}
1086 
1087 	return val & FB_GFX_FREQ_FUSE_MASK;
1088 }
1089 
1090 static int chv_rps_rpe_freq(struct intel_rps *rps)
1091 {
1092 	struct drm_i915_private *i915 = rps_to_i915(rps);
1093 	u32 val;
1094 
1095 	val = vlv_punit_read(i915, PUNIT_GPU_DUTYCYCLE_REG);
1096 	val >>= PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT;
1097 
1098 	return val & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
1099 }
1100 
1101 static int chv_rps_guar_freq(struct intel_rps *rps)
1102 {
1103 	struct drm_i915_private *i915 = rps_to_i915(rps);
1104 	u32 val;
1105 
1106 	val = vlv_punit_read(i915, FB_GFX_FMAX_AT_VMAX_FUSE);
1107 
1108 	return val & FB_GFX_FREQ_FUSE_MASK;
1109 }
1110 
1111 static u32 chv_rps_min_freq(struct intel_rps *rps)
1112 {
1113 	struct drm_i915_private *i915 = rps_to_i915(rps);
1114 	u32 val;
1115 
1116 	val = vlv_punit_read(i915, FB_GFX_FMIN_AT_VMIN_FUSE);
1117 	val >>= FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT;
1118 
1119 	return val & FB_GFX_FREQ_FUSE_MASK;
1120 }
1121 
1122 static bool chv_rps_enable(struct intel_rps *rps)
1123 {
1124 	struct intel_uncore *uncore = rps_to_uncore(rps);
1125 	struct drm_i915_private *i915 = rps_to_i915(rps);
1126 	u32 val;
1127 
1128 	/* 1: Program defaults and thresholds for RPS*/
1129 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000);
1130 	intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400);
1131 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000);
1132 	intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000);
1133 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000);
1134 
1135 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1136 
1137 	/* 2: Enable RPS */
1138 	intel_uncore_write_fw(uncore, GEN6_RP_CONTROL,
1139 			      GEN6_RP_MEDIA_HW_NORMAL_MODE |
1140 			      GEN6_RP_MEDIA_IS_GFX |
1141 			      GEN6_RP_ENABLE |
1142 			      GEN6_RP_UP_BUSY_AVG |
1143 			      GEN6_RP_DOWN_IDLE_AVG);
1144 
1145 	rps->pm_events = (GEN6_PM_RP_UP_THRESHOLD |
1146 			  GEN6_PM_RP_DOWN_THRESHOLD |
1147 			  GEN6_PM_RP_DOWN_TIMEOUT);
1148 
1149 	/* Setting Fixed Bias */
1150 	vlv_punit_get(i915);
1151 
1152 	val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | CHV_BIAS_CPU_50_SOC_50;
1153 	vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val);
1154 
1155 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1156 
1157 	vlv_punit_put(i915);
1158 
1159 	/* RPS code assumes GPLL is used */
1160 	drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0,
1161 		      "GPLL not enabled\n");
1162 
1163 	drm_dbg(&i915->drm, "GPLL enabled? %s\n", yesno(val & GPLLENABLE));
1164 	drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val);
1165 
1166 	return rps_reset(rps);
1167 }
1168 
1169 static int vlv_rps_guar_freq(struct intel_rps *rps)
1170 {
1171 	struct drm_i915_private *i915 = rps_to_i915(rps);
1172 	u32 val, rp1;
1173 
1174 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE);
1175 
1176 	rp1 = val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK;
1177 	rp1 >>= FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
1178 
1179 	return rp1;
1180 }
1181 
1182 static int vlv_rps_max_freq(struct intel_rps *rps)
1183 {
1184 	struct drm_i915_private *i915 = rps_to_i915(rps);
1185 	u32 val, rp0;
1186 
1187 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FREQ_FUSE);
1188 
1189 	rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
1190 	/* Clamp to max */
1191 	rp0 = min_t(u32, rp0, 0xea);
1192 
1193 	return rp0;
1194 }
1195 
1196 static int vlv_rps_rpe_freq(struct intel_rps *rps)
1197 {
1198 	struct drm_i915_private *i915 = rps_to_i915(rps);
1199 	u32 val, rpe;
1200 
1201 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
1202 	rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
1203 	val = vlv_nc_read(i915, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
1204 	rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
1205 
1206 	return rpe;
1207 }
1208 
1209 static int vlv_rps_min_freq(struct intel_rps *rps)
1210 {
1211 	struct drm_i915_private *i915 = rps_to_i915(rps);
1212 	u32 val;
1213 
1214 	val = vlv_punit_read(i915, PUNIT_REG_GPU_LFM) & 0xff;
1215 	/*
1216 	 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
1217 	 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
1218 	 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
1219 	 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
1220 	 * to make sure it matches what Punit accepts.
1221 	 */
1222 	return max_t(u32, val, 0xc0);
1223 }
1224 
1225 static bool vlv_rps_enable(struct intel_rps *rps)
1226 {
1227 	struct intel_uncore *uncore = rps_to_uncore(rps);
1228 	struct drm_i915_private *i915 = rps_to_i915(rps);
1229 	u32 val;
1230 
1231 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_TIMEOUT, 1000000);
1232 	intel_uncore_write_fw(uncore, GEN6_RP_UP_THRESHOLD, 59400);
1233 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_THRESHOLD, 245000);
1234 	intel_uncore_write_fw(uncore, GEN6_RP_UP_EI, 66000);
1235 	intel_uncore_write_fw(uncore, GEN6_RP_DOWN_EI, 350000);
1236 
1237 	intel_uncore_write_fw(uncore, GEN6_RP_IDLE_HYSTERSIS, 10);
1238 
1239 	intel_uncore_write_fw(uncore, GEN6_RP_CONTROL,
1240 			      GEN6_RP_MEDIA_TURBO |
1241 			      GEN6_RP_MEDIA_HW_NORMAL_MODE |
1242 			      GEN6_RP_MEDIA_IS_GFX |
1243 			      GEN6_RP_ENABLE |
1244 			      GEN6_RP_UP_BUSY_AVG |
1245 			      GEN6_RP_DOWN_IDLE_CONT);
1246 
1247 	/* WaGsvRC0ResidencyMethod:vlv */
1248 	rps->pm_events = GEN6_PM_RP_UP_EI_EXPIRED;
1249 
1250 	vlv_punit_get(i915);
1251 
1252 	/* Setting Fixed Bias */
1253 	val = VLV_OVERRIDE_EN | VLV_SOC_TDP_EN | VLV_BIAS_CPU_125_SOC_875;
1254 	vlv_punit_write(i915, VLV_TURBO_SOC_OVERRIDE, val);
1255 
1256 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1257 
1258 	vlv_punit_put(i915);
1259 
1260 	/* RPS code assumes GPLL is used */
1261 	drm_WARN_ONCE(&i915->drm, (val & GPLLENABLE) == 0,
1262 		      "GPLL not enabled\n");
1263 
1264 	drm_dbg(&i915->drm, "GPLL enabled? %s\n", yesno(val & GPLLENABLE));
1265 	drm_dbg(&i915->drm, "GPU status: 0x%08x\n", val);
1266 
1267 	return rps_reset(rps);
1268 }
1269 
1270 static unsigned long __ips_gfx_val(struct intel_ips *ips)
1271 {
1272 	struct intel_rps *rps = container_of(ips, typeof(*rps), ips);
1273 	struct intel_uncore *uncore = rps_to_uncore(rps);
1274 	unsigned long t, corr, state1, corr2, state2;
1275 	u32 pxvid, ext_v;
1276 
1277 	lockdep_assert_held(&mchdev_lock);
1278 
1279 	pxvid = intel_uncore_read(uncore, PXVFREQ(rps->cur_freq));
1280 	pxvid = (pxvid >> 24) & 0x7f;
1281 	ext_v = pvid_to_extvid(rps_to_i915(rps), pxvid);
1282 
1283 	state1 = ext_v;
1284 
1285 	/* Revel in the empirically derived constants */
1286 
1287 	/* Correction factor in 1/100000 units */
1288 	t = ips_mch_val(uncore);
1289 	if (t > 80)
1290 		corr = t * 2349 + 135940;
1291 	else if (t >= 50)
1292 		corr = t * 964 + 29317;
1293 	else /* < 50 */
1294 		corr = t * 301 + 1004;
1295 
1296 	corr = corr * 150142 * state1 / 10000 - 78642;
1297 	corr /= 100000;
1298 	corr2 = corr * ips->corr;
1299 
1300 	state2 = corr2 * state1 / 10000;
1301 	state2 /= 100; /* convert to mW */
1302 
1303 	__gen5_ips_update(ips);
1304 
1305 	return ips->gfx_power + state2;
1306 }
1307 
1308 static bool has_busy_stats(struct intel_rps *rps)
1309 {
1310 	struct intel_engine_cs *engine;
1311 	enum intel_engine_id id;
1312 
1313 	for_each_engine(engine, rps_to_gt(rps), id) {
1314 		if (!intel_engine_supports_stats(engine))
1315 			return false;
1316 	}
1317 
1318 	return true;
1319 }
1320 
1321 void intel_rps_enable(struct intel_rps *rps)
1322 {
1323 	struct drm_i915_private *i915 = rps_to_i915(rps);
1324 	struct intel_uncore *uncore = rps_to_uncore(rps);
1325 	bool enabled = false;
1326 
1327 	if (!HAS_RPS(i915))
1328 		return;
1329 
1330 	intel_gt_check_clock_frequency(rps_to_gt(rps));
1331 
1332 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1333 	if (rps->max_freq <= rps->min_freq)
1334 		/* leave disabled, no room for dynamic reclocking */;
1335 	else if (IS_CHERRYVIEW(i915))
1336 		enabled = chv_rps_enable(rps);
1337 	else if (IS_VALLEYVIEW(i915))
1338 		enabled = vlv_rps_enable(rps);
1339 	else if (INTEL_GEN(i915) >= 9)
1340 		enabled = gen9_rps_enable(rps);
1341 	else if (INTEL_GEN(i915) >= 8)
1342 		enabled = gen8_rps_enable(rps);
1343 	else if (INTEL_GEN(i915) >= 6)
1344 		enabled = gen6_rps_enable(rps);
1345 	else if (IS_IRONLAKE_M(i915))
1346 		enabled = gen5_rps_enable(rps);
1347 	else
1348 		MISSING_CASE(INTEL_GEN(i915));
1349 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1350 	if (!enabled)
1351 		return;
1352 
1353 	GT_TRACE(rps_to_gt(rps),
1354 		 "min:%x, max:%x, freq:[%d, %d]\n",
1355 		 rps->min_freq, rps->max_freq,
1356 		 intel_gpu_freq(rps, rps->min_freq),
1357 		 intel_gpu_freq(rps, rps->max_freq));
1358 
1359 	GEM_BUG_ON(rps->max_freq < rps->min_freq);
1360 	GEM_BUG_ON(rps->idle_freq > rps->max_freq);
1361 
1362 	GEM_BUG_ON(rps->efficient_freq < rps->min_freq);
1363 	GEM_BUG_ON(rps->efficient_freq > rps->max_freq);
1364 
1365 	if (has_busy_stats(rps))
1366 		intel_rps_set_timer(rps);
1367 	else if (INTEL_GEN(i915) >= 6)
1368 		intel_rps_set_interrupts(rps);
1369 	else
1370 		/* Ironlake currently uses intel_ips.ko */ {}
1371 
1372 	intel_rps_set_enabled(rps);
1373 }
1374 
1375 static void gen6_rps_disable(struct intel_rps *rps)
1376 {
1377 	set(rps_to_uncore(rps), GEN6_RP_CONTROL, 0);
1378 }
1379 
1380 void intel_rps_disable(struct intel_rps *rps)
1381 {
1382 	struct drm_i915_private *i915 = rps_to_i915(rps);
1383 
1384 	intel_rps_clear_enabled(rps);
1385 	intel_rps_clear_interrupts(rps);
1386 	intel_rps_clear_timer(rps);
1387 
1388 	if (INTEL_GEN(i915) >= 6)
1389 		gen6_rps_disable(rps);
1390 	else if (IS_IRONLAKE_M(i915))
1391 		gen5_rps_disable(rps);
1392 }
1393 
1394 static int byt_gpu_freq(struct intel_rps *rps, int val)
1395 {
1396 	/*
1397 	 * N = val - 0xb7
1398 	 * Slow = Fast = GPLL ref * N
1399 	 */
1400 	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000);
1401 }
1402 
1403 static int byt_freq_opcode(struct intel_rps *rps, int val)
1404 {
1405 	return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7;
1406 }
1407 
1408 static int chv_gpu_freq(struct intel_rps *rps, int val)
1409 {
1410 	/*
1411 	 * N = val / 2
1412 	 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
1413 	 */
1414 	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000);
1415 }
1416 
1417 static int chv_freq_opcode(struct intel_rps *rps, int val)
1418 {
1419 	/* CHV needs even values */
1420 	return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2;
1421 }
1422 
1423 int intel_gpu_freq(struct intel_rps *rps, int val)
1424 {
1425 	struct drm_i915_private *i915 = rps_to_i915(rps);
1426 
1427 	if (INTEL_GEN(i915) >= 9)
1428 		return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
1429 					 GEN9_FREQ_SCALER);
1430 	else if (IS_CHERRYVIEW(i915))
1431 		return chv_gpu_freq(rps, val);
1432 	else if (IS_VALLEYVIEW(i915))
1433 		return byt_gpu_freq(rps, val);
1434 	else
1435 		return val * GT_FREQUENCY_MULTIPLIER;
1436 }
1437 
1438 int intel_freq_opcode(struct intel_rps *rps, int val)
1439 {
1440 	struct drm_i915_private *i915 = rps_to_i915(rps);
1441 
1442 	if (INTEL_GEN(i915) >= 9)
1443 		return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
1444 					 GT_FREQUENCY_MULTIPLIER);
1445 	else if (IS_CHERRYVIEW(i915))
1446 		return chv_freq_opcode(rps, val);
1447 	else if (IS_VALLEYVIEW(i915))
1448 		return byt_freq_opcode(rps, val);
1449 	else
1450 		return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
1451 }
1452 
1453 static void vlv_init_gpll_ref_freq(struct intel_rps *rps)
1454 {
1455 	struct drm_i915_private *i915 = rps_to_i915(rps);
1456 
1457 	rps->gpll_ref_freq =
1458 		vlv_get_cck_clock(i915, "GPLL ref",
1459 				  CCK_GPLL_CLOCK_CONTROL,
1460 				  i915->czclk_freq);
1461 
1462 	drm_dbg(&i915->drm, "GPLL reference freq: %d kHz\n",
1463 		rps->gpll_ref_freq);
1464 }
1465 
1466 static void vlv_rps_init(struct intel_rps *rps)
1467 {
1468 	struct drm_i915_private *i915 = rps_to_i915(rps);
1469 	u32 val;
1470 
1471 	vlv_iosf_sb_get(i915,
1472 			BIT(VLV_IOSF_SB_PUNIT) |
1473 			BIT(VLV_IOSF_SB_NC) |
1474 			BIT(VLV_IOSF_SB_CCK));
1475 
1476 	vlv_init_gpll_ref_freq(rps);
1477 
1478 	val = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1479 	switch ((val >> 6) & 3) {
1480 	case 0:
1481 	case 1:
1482 		i915->mem_freq = 800;
1483 		break;
1484 	case 2:
1485 		i915->mem_freq = 1066;
1486 		break;
1487 	case 3:
1488 		i915->mem_freq = 1333;
1489 		break;
1490 	}
1491 	drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq);
1492 
1493 	rps->max_freq = vlv_rps_max_freq(rps);
1494 	rps->rp0_freq = rps->max_freq;
1495 	drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n",
1496 		intel_gpu_freq(rps, rps->max_freq), rps->max_freq);
1497 
1498 	rps->efficient_freq = vlv_rps_rpe_freq(rps);
1499 	drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n",
1500 		intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq);
1501 
1502 	rps->rp1_freq = vlv_rps_guar_freq(rps);
1503 	drm_dbg(&i915->drm, "RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
1504 		intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq);
1505 
1506 	rps->min_freq = vlv_rps_min_freq(rps);
1507 	drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n",
1508 		intel_gpu_freq(rps, rps->min_freq), rps->min_freq);
1509 
1510 	vlv_iosf_sb_put(i915,
1511 			BIT(VLV_IOSF_SB_PUNIT) |
1512 			BIT(VLV_IOSF_SB_NC) |
1513 			BIT(VLV_IOSF_SB_CCK));
1514 }
1515 
1516 static void chv_rps_init(struct intel_rps *rps)
1517 {
1518 	struct drm_i915_private *i915 = rps_to_i915(rps);
1519 	u32 val;
1520 
1521 	vlv_iosf_sb_get(i915,
1522 			BIT(VLV_IOSF_SB_PUNIT) |
1523 			BIT(VLV_IOSF_SB_NC) |
1524 			BIT(VLV_IOSF_SB_CCK));
1525 
1526 	vlv_init_gpll_ref_freq(rps);
1527 
1528 	val = vlv_cck_read(i915, CCK_FUSE_REG);
1529 
1530 	switch ((val >> 2) & 0x7) {
1531 	case 3:
1532 		i915->mem_freq = 2000;
1533 		break;
1534 	default:
1535 		i915->mem_freq = 1600;
1536 		break;
1537 	}
1538 	drm_dbg(&i915->drm, "DDR speed: %d MHz\n", i915->mem_freq);
1539 
1540 	rps->max_freq = chv_rps_max_freq(rps);
1541 	rps->rp0_freq = rps->max_freq;
1542 	drm_dbg(&i915->drm, "max GPU freq: %d MHz (%u)\n",
1543 		intel_gpu_freq(rps, rps->max_freq), rps->max_freq);
1544 
1545 	rps->efficient_freq = chv_rps_rpe_freq(rps);
1546 	drm_dbg(&i915->drm, "RPe GPU freq: %d MHz (%u)\n",
1547 		intel_gpu_freq(rps, rps->efficient_freq), rps->efficient_freq);
1548 
1549 	rps->rp1_freq = chv_rps_guar_freq(rps);
1550 	drm_dbg(&i915->drm, "RP1(Guar) GPU freq: %d MHz (%u)\n",
1551 		intel_gpu_freq(rps, rps->rp1_freq), rps->rp1_freq);
1552 
1553 	rps->min_freq = chv_rps_min_freq(rps);
1554 	drm_dbg(&i915->drm, "min GPU freq: %d MHz (%u)\n",
1555 		intel_gpu_freq(rps, rps->min_freq), rps->min_freq);
1556 
1557 	vlv_iosf_sb_put(i915,
1558 			BIT(VLV_IOSF_SB_PUNIT) |
1559 			BIT(VLV_IOSF_SB_NC) |
1560 			BIT(VLV_IOSF_SB_CCK));
1561 
1562 	drm_WARN_ONCE(&i915->drm, (rps->max_freq | rps->efficient_freq |
1563 				   rps->rp1_freq | rps->min_freq) & 1,
1564 		      "Odd GPU freq values\n");
1565 }
1566 
1567 static void vlv_c0_read(struct intel_uncore *uncore, struct intel_rps_ei *ei)
1568 {
1569 	ei->ktime = ktime_get_raw();
1570 	ei->render_c0 = intel_uncore_read(uncore, VLV_RENDER_C0_COUNT);
1571 	ei->media_c0 = intel_uncore_read(uncore, VLV_MEDIA_C0_COUNT);
1572 }
1573 
1574 static u32 vlv_wa_c0_ei(struct intel_rps *rps, u32 pm_iir)
1575 {
1576 	struct intel_uncore *uncore = rps_to_uncore(rps);
1577 	const struct intel_rps_ei *prev = &rps->ei;
1578 	struct intel_rps_ei now;
1579 	u32 events = 0;
1580 
1581 	if ((pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) == 0)
1582 		return 0;
1583 
1584 	vlv_c0_read(uncore, &now);
1585 
1586 	if (prev->ktime) {
1587 		u64 time, c0;
1588 		u32 render, media;
1589 
1590 		time = ktime_us_delta(now.ktime, prev->ktime);
1591 
1592 		time *= rps_to_i915(rps)->czclk_freq;
1593 
1594 		/* Workload can be split between render + media,
1595 		 * e.g. SwapBuffers being blitted in X after being rendered in
1596 		 * mesa. To account for this we need to combine both engines
1597 		 * into our activity counter.
1598 		 */
1599 		render = now.render_c0 - prev->render_c0;
1600 		media = now.media_c0 - prev->media_c0;
1601 		c0 = max(render, media);
1602 		c0 *= 1000 * 100 << 8; /* to usecs and scale to threshold% */
1603 
1604 		if (c0 > time * rps->power.up_threshold)
1605 			events = GEN6_PM_RP_UP_THRESHOLD;
1606 		else if (c0 < time * rps->power.down_threshold)
1607 			events = GEN6_PM_RP_DOWN_THRESHOLD;
1608 	}
1609 
1610 	rps->ei = now;
1611 	return events;
1612 }
1613 
1614 static void rps_work(struct work_struct *work)
1615 {
1616 	struct intel_rps *rps = container_of(work, typeof(*rps), work);
1617 	struct intel_gt *gt = rps_to_gt(rps);
1618 	struct drm_i915_private *i915 = rps_to_i915(rps);
1619 	bool client_boost = false;
1620 	int new_freq, adj, min, max;
1621 	u32 pm_iir = 0;
1622 
1623 	spin_lock_irq(&gt->irq_lock);
1624 	pm_iir = fetch_and_zero(&rps->pm_iir) & rps->pm_events;
1625 	client_boost = atomic_read(&rps->num_waiters);
1626 	spin_unlock_irq(&gt->irq_lock);
1627 
1628 	/* Make sure we didn't queue anything we're not going to process. */
1629 	if (!pm_iir && !client_boost)
1630 		goto out;
1631 
1632 	mutex_lock(&rps->lock);
1633 	if (!intel_rps_is_active(rps)) {
1634 		mutex_unlock(&rps->lock);
1635 		return;
1636 	}
1637 
1638 	pm_iir |= vlv_wa_c0_ei(rps, pm_iir);
1639 
1640 	adj = rps->last_adj;
1641 	new_freq = rps->cur_freq;
1642 	min = rps->min_freq_softlimit;
1643 	max = rps->max_freq_softlimit;
1644 	if (client_boost)
1645 		max = rps->max_freq;
1646 
1647 	GT_TRACE(gt,
1648 		 "pm_iir:%x, client_boost:%s, last:%d, cur:%x, min:%x, max:%x\n",
1649 		 pm_iir, yesno(client_boost),
1650 		 adj, new_freq, min, max);
1651 
1652 	if (client_boost && new_freq < rps->boost_freq) {
1653 		new_freq = rps->boost_freq;
1654 		adj = 0;
1655 	} else if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) {
1656 		if (adj > 0)
1657 			adj *= 2;
1658 		else /* CHV needs even encode values */
1659 			adj = IS_CHERRYVIEW(gt->i915) ? 2 : 1;
1660 
1661 		if (new_freq >= rps->max_freq_softlimit)
1662 			adj = 0;
1663 	} else if (client_boost) {
1664 		adj = 0;
1665 	} else if (pm_iir & GEN6_PM_RP_DOWN_TIMEOUT) {
1666 		if (rps->cur_freq > rps->efficient_freq)
1667 			new_freq = rps->efficient_freq;
1668 		else if (rps->cur_freq > rps->min_freq_softlimit)
1669 			new_freq = rps->min_freq_softlimit;
1670 		adj = 0;
1671 	} else if (pm_iir & GEN6_PM_RP_DOWN_THRESHOLD) {
1672 		if (adj < 0)
1673 			adj *= 2;
1674 		else /* CHV needs even encode values */
1675 			adj = IS_CHERRYVIEW(gt->i915) ? -2 : -1;
1676 
1677 		if (new_freq <= rps->min_freq_softlimit)
1678 			adj = 0;
1679 	} else { /* unknown event */
1680 		adj = 0;
1681 	}
1682 
1683 	/*
1684 	 * sysfs frequency limits may have snuck in while
1685 	 * servicing the interrupt
1686 	 */
1687 	new_freq += adj;
1688 	new_freq = clamp_t(int, new_freq, min, max);
1689 
1690 	if (intel_rps_set(rps, new_freq)) {
1691 		drm_dbg(&i915->drm, "Failed to set new GPU frequency\n");
1692 		adj = 0;
1693 	}
1694 	rps->last_adj = adj;
1695 
1696 	mutex_unlock(&rps->lock);
1697 
1698 out:
1699 	spin_lock_irq(&gt->irq_lock);
1700 	gen6_gt_pm_unmask_irq(gt, rps->pm_events);
1701 	spin_unlock_irq(&gt->irq_lock);
1702 }
1703 
1704 void gen11_rps_irq_handler(struct intel_rps *rps, u32 pm_iir)
1705 {
1706 	struct intel_gt *gt = rps_to_gt(rps);
1707 	const u32 events = rps->pm_events & pm_iir;
1708 
1709 	lockdep_assert_held(&gt->irq_lock);
1710 
1711 	if (unlikely(!events))
1712 		return;
1713 
1714 	GT_TRACE(gt, "irq events:%x\n", events);
1715 
1716 	gen6_gt_pm_mask_irq(gt, events);
1717 
1718 	rps->pm_iir |= events;
1719 	schedule_work(&rps->work);
1720 }
1721 
1722 void gen6_rps_irq_handler(struct intel_rps *rps, u32 pm_iir)
1723 {
1724 	struct intel_gt *gt = rps_to_gt(rps);
1725 	u32 events;
1726 
1727 	events = pm_iir & rps->pm_events;
1728 	if (events) {
1729 		spin_lock(&gt->irq_lock);
1730 
1731 		GT_TRACE(gt, "irq events:%x\n", events);
1732 
1733 		gen6_gt_pm_mask_irq(gt, events);
1734 		rps->pm_iir |= events;
1735 
1736 		schedule_work(&rps->work);
1737 		spin_unlock(&gt->irq_lock);
1738 	}
1739 
1740 	if (INTEL_GEN(gt->i915) >= 8)
1741 		return;
1742 
1743 	if (pm_iir & PM_VEBOX_USER_INTERRUPT)
1744 		intel_engine_signal_breadcrumbs(gt->engine[VECS0]);
1745 
1746 	if (pm_iir & PM_VEBOX_CS_ERROR_INTERRUPT)
1747 		DRM_DEBUG("Command parser error, pm_iir 0x%08x\n", pm_iir);
1748 }
1749 
1750 void gen5_rps_irq_handler(struct intel_rps *rps)
1751 {
1752 	struct intel_uncore *uncore = rps_to_uncore(rps);
1753 	u32 busy_up, busy_down, max_avg, min_avg;
1754 	u8 new_freq;
1755 
1756 	spin_lock(&mchdev_lock);
1757 
1758 	intel_uncore_write16(uncore,
1759 			     MEMINTRSTS,
1760 			     intel_uncore_read(uncore, MEMINTRSTS));
1761 
1762 	intel_uncore_write16(uncore, MEMINTRSTS, MEMINT_EVAL_CHG);
1763 	busy_up = intel_uncore_read(uncore, RCPREVBSYTUPAVG);
1764 	busy_down = intel_uncore_read(uncore, RCPREVBSYTDNAVG);
1765 	max_avg = intel_uncore_read(uncore, RCBMAXAVG);
1766 	min_avg = intel_uncore_read(uncore, RCBMINAVG);
1767 
1768 	/* Handle RCS change request from hw */
1769 	new_freq = rps->cur_freq;
1770 	if (busy_up > max_avg)
1771 		new_freq++;
1772 	else if (busy_down < min_avg)
1773 		new_freq--;
1774 	new_freq = clamp(new_freq,
1775 			 rps->min_freq_softlimit,
1776 			 rps->max_freq_softlimit);
1777 
1778 	if (new_freq != rps->cur_freq && gen5_rps_set(rps, new_freq))
1779 		rps->cur_freq = new_freq;
1780 
1781 	spin_unlock(&mchdev_lock);
1782 }
1783 
1784 void intel_rps_init_early(struct intel_rps *rps)
1785 {
1786 	mutex_init(&rps->lock);
1787 	mutex_init(&rps->power.mutex);
1788 
1789 	INIT_WORK(&rps->work, rps_work);
1790 	timer_setup(&rps->timer, rps_timer, 0);
1791 
1792 	atomic_set(&rps->num_waiters, 0);
1793 }
1794 
1795 void intel_rps_init(struct intel_rps *rps)
1796 {
1797 	struct drm_i915_private *i915 = rps_to_i915(rps);
1798 
1799 	if (IS_CHERRYVIEW(i915))
1800 		chv_rps_init(rps);
1801 	else if (IS_VALLEYVIEW(i915))
1802 		vlv_rps_init(rps);
1803 	else if (INTEL_GEN(i915) >= 6)
1804 		gen6_rps_init(rps);
1805 	else if (IS_IRONLAKE_M(i915))
1806 		gen5_rps_init(rps);
1807 
1808 	/* Derive initial user preferences/limits from the hardware limits */
1809 	rps->max_freq_softlimit = rps->max_freq;
1810 	rps->min_freq_softlimit = rps->min_freq;
1811 
1812 	/* After setting max-softlimit, find the overclock max freq */
1813 	if (IS_GEN(i915, 6) || IS_IVYBRIDGE(i915) || IS_HASWELL(i915)) {
1814 		u32 params = 0;
1815 
1816 		sandybridge_pcode_read(i915, GEN6_READ_OC_PARAMS,
1817 				       &params, NULL);
1818 		if (params & BIT(31)) { /* OC supported */
1819 			drm_dbg(&i915->drm,
1820 				"Overclocking supported, max: %dMHz, overclock: %dMHz\n",
1821 				(rps->max_freq & 0xff) * 50,
1822 				(params & 0xff) * 50);
1823 			rps->max_freq = params & 0xff;
1824 		}
1825 	}
1826 
1827 	/* Finally allow us to boost to max by default */
1828 	rps->boost_freq = rps->max_freq;
1829 	rps->idle_freq = rps->min_freq;
1830 
1831 	/* Start in the middle, from here we will autotune based on workload */
1832 	rps->cur_freq = rps->efficient_freq;
1833 
1834 	rps->pm_intrmsk_mbz = 0;
1835 
1836 	/*
1837 	 * SNB,IVB,HSW can while VLV,CHV may hard hang on looping batchbuffer
1838 	 * if GEN6_PM_UP_EI_EXPIRED is masked.
1839 	 *
1840 	 * TODO: verify if this can be reproduced on VLV,CHV.
1841 	 */
1842 	if (INTEL_GEN(i915) <= 7)
1843 		rps->pm_intrmsk_mbz |= GEN6_PM_RP_UP_EI_EXPIRED;
1844 
1845 	if (INTEL_GEN(i915) >= 8 && INTEL_GEN(i915) < 11)
1846 		rps->pm_intrmsk_mbz |= GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC;
1847 }
1848 
1849 void intel_rps_sanitize(struct intel_rps *rps)
1850 {
1851 	if (INTEL_GEN(rps_to_i915(rps)) >= 6)
1852 		rps_disable_interrupts(rps);
1853 }
1854 
1855 u32 intel_rps_get_cagf(struct intel_rps *rps, u32 rpstat)
1856 {
1857 	struct drm_i915_private *i915 = rps_to_i915(rps);
1858 	u32 cagf;
1859 
1860 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
1861 		cagf = (rpstat >> 8) & 0xff;
1862 	else if (INTEL_GEN(i915) >= 9)
1863 		cagf = (rpstat & GEN9_CAGF_MASK) >> GEN9_CAGF_SHIFT;
1864 	else if (IS_HASWELL(i915) || IS_BROADWELL(i915))
1865 		cagf = (rpstat & HSW_CAGF_MASK) >> HSW_CAGF_SHIFT;
1866 	else
1867 		cagf = (rpstat & GEN6_CAGF_MASK) >> GEN6_CAGF_SHIFT;
1868 
1869 	return cagf;
1870 }
1871 
1872 static u32 read_cagf(struct intel_rps *rps)
1873 {
1874 	struct drm_i915_private *i915 = rps_to_i915(rps);
1875 	u32 freq;
1876 
1877 	if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) {
1878 		vlv_punit_get(i915);
1879 		freq = vlv_punit_read(i915, PUNIT_REG_GPU_FREQ_STS);
1880 		vlv_punit_put(i915);
1881 	} else {
1882 		freq = intel_uncore_read(rps_to_uncore(rps), GEN6_RPSTAT1);
1883 	}
1884 
1885 	return intel_rps_get_cagf(rps, freq);
1886 }
1887 
1888 u32 intel_rps_read_actual_frequency(struct intel_rps *rps)
1889 {
1890 	struct intel_runtime_pm *rpm = rps_to_uncore(rps)->rpm;
1891 	intel_wakeref_t wakeref;
1892 	u32 freq = 0;
1893 
1894 	with_intel_runtime_pm_if_in_use(rpm, wakeref)
1895 		freq = intel_gpu_freq(rps, read_cagf(rps));
1896 
1897 	return freq;
1898 }
1899 
1900 /* External interface for intel_ips.ko */
1901 
1902 static struct drm_i915_private __rcu *ips_mchdev;
1903 
1904 /**
1905  * Tells the intel_ips driver that the i915 driver is now loaded, if
1906  * IPS got loaded first.
1907  *
1908  * This awkward dance is so that neither module has to depend on the
1909  * other in order for IPS to do the appropriate communication of
1910  * GPU turbo limits to i915.
1911  */
1912 static void
1913 ips_ping_for_i915_load(void)
1914 {
1915 	void (*link)(void);
1916 
1917 	link = symbol_get(ips_link_to_i915_driver);
1918 	if (link) {
1919 		link();
1920 		symbol_put(ips_link_to_i915_driver);
1921 	}
1922 }
1923 
1924 void intel_rps_driver_register(struct intel_rps *rps)
1925 {
1926 	struct intel_gt *gt = rps_to_gt(rps);
1927 
1928 	/*
1929 	 * We only register the i915 ips part with intel-ips once everything is
1930 	 * set up, to avoid intel-ips sneaking in and reading bogus values.
1931 	 */
1932 	if (IS_GEN(gt->i915, 5)) {
1933 		GEM_BUG_ON(ips_mchdev);
1934 		rcu_assign_pointer(ips_mchdev, gt->i915);
1935 		ips_ping_for_i915_load();
1936 	}
1937 }
1938 
1939 void intel_rps_driver_unregister(struct intel_rps *rps)
1940 {
1941 	if (rcu_access_pointer(ips_mchdev) == rps_to_i915(rps))
1942 		rcu_assign_pointer(ips_mchdev, NULL);
1943 }
1944 
1945 static struct drm_i915_private *mchdev_get(void)
1946 {
1947 	struct drm_i915_private *i915;
1948 
1949 	rcu_read_lock();
1950 	i915 = rcu_dereference(ips_mchdev);
1951 	if (!kref_get_unless_zero(&i915->drm.ref))
1952 		i915 = NULL;
1953 	rcu_read_unlock();
1954 
1955 	return i915;
1956 }
1957 
1958 /**
1959  * i915_read_mch_val - return value for IPS use
1960  *
1961  * Calculate and return a value for the IPS driver to use when deciding whether
1962  * we have thermal and power headroom to increase CPU or GPU power budget.
1963  */
1964 unsigned long i915_read_mch_val(void)
1965 {
1966 	struct drm_i915_private *i915;
1967 	unsigned long chipset_val = 0;
1968 	unsigned long graphics_val = 0;
1969 	intel_wakeref_t wakeref;
1970 
1971 	i915 = mchdev_get();
1972 	if (!i915)
1973 		return 0;
1974 
1975 	with_intel_runtime_pm(&i915->runtime_pm, wakeref) {
1976 		struct intel_ips *ips = &i915->gt.rps.ips;
1977 
1978 		spin_lock_irq(&mchdev_lock);
1979 		chipset_val = __ips_chipset_val(ips);
1980 		graphics_val = __ips_gfx_val(ips);
1981 		spin_unlock_irq(&mchdev_lock);
1982 	}
1983 
1984 	drm_dev_put(&i915->drm);
1985 	return chipset_val + graphics_val;
1986 }
1987 EXPORT_SYMBOL_GPL(i915_read_mch_val);
1988 
1989 /**
1990  * i915_gpu_raise - raise GPU frequency limit
1991  *
1992  * Raise the limit; IPS indicates we have thermal headroom.
1993  */
1994 bool i915_gpu_raise(void)
1995 {
1996 	struct drm_i915_private *i915;
1997 	struct intel_rps *rps;
1998 
1999 	i915 = mchdev_get();
2000 	if (!i915)
2001 		return false;
2002 
2003 	rps = &i915->gt.rps;
2004 
2005 	spin_lock_irq(&mchdev_lock);
2006 	if (rps->max_freq_softlimit < rps->max_freq)
2007 		rps->max_freq_softlimit++;
2008 	spin_unlock_irq(&mchdev_lock);
2009 
2010 	drm_dev_put(&i915->drm);
2011 	return true;
2012 }
2013 EXPORT_SYMBOL_GPL(i915_gpu_raise);
2014 
2015 /**
2016  * i915_gpu_lower - lower GPU frequency limit
2017  *
2018  * IPS indicates we're close to a thermal limit, so throttle back the GPU
2019  * frequency maximum.
2020  */
2021 bool i915_gpu_lower(void)
2022 {
2023 	struct drm_i915_private *i915;
2024 	struct intel_rps *rps;
2025 
2026 	i915 = mchdev_get();
2027 	if (!i915)
2028 		return false;
2029 
2030 	rps = &i915->gt.rps;
2031 
2032 	spin_lock_irq(&mchdev_lock);
2033 	if (rps->max_freq_softlimit > rps->min_freq)
2034 		rps->max_freq_softlimit--;
2035 	spin_unlock_irq(&mchdev_lock);
2036 
2037 	drm_dev_put(&i915->drm);
2038 	return true;
2039 }
2040 EXPORT_SYMBOL_GPL(i915_gpu_lower);
2041 
2042 /**
2043  * i915_gpu_busy - indicate GPU business to IPS
2044  *
2045  * Tell the IPS driver whether or not the GPU is busy.
2046  */
2047 bool i915_gpu_busy(void)
2048 {
2049 	struct drm_i915_private *i915;
2050 	bool ret;
2051 
2052 	i915 = mchdev_get();
2053 	if (!i915)
2054 		return false;
2055 
2056 	ret = i915->gt.awake;
2057 
2058 	drm_dev_put(&i915->drm);
2059 	return ret;
2060 }
2061 EXPORT_SYMBOL_GPL(i915_gpu_busy);
2062 
2063 /**
2064  * i915_gpu_turbo_disable - disable graphics turbo
2065  *
2066  * Disable graphics turbo by resetting the max frequency and setting the
2067  * current frequency to the default.
2068  */
2069 bool i915_gpu_turbo_disable(void)
2070 {
2071 	struct drm_i915_private *i915;
2072 	struct intel_rps *rps;
2073 	bool ret;
2074 
2075 	i915 = mchdev_get();
2076 	if (!i915)
2077 		return false;
2078 
2079 	rps = &i915->gt.rps;
2080 
2081 	spin_lock_irq(&mchdev_lock);
2082 	rps->max_freq_softlimit = rps->min_freq;
2083 	ret = gen5_rps_set(&i915->gt.rps, rps->min_freq);
2084 	spin_unlock_irq(&mchdev_lock);
2085 
2086 	drm_dev_put(&i915->drm);
2087 	return ret;
2088 }
2089 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
2090 
2091 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2092 #include "selftest_rps.c"
2093 #endif
2094