1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <drm/drm_atomic_state_helper.h>
7 
8 #include "intel_atomic.h"
9 #include "intel_bw.h"
10 #include "intel_cdclk.h"
11 #include "intel_display_types.h"
12 #include "intel_pcode.h"
13 #include "intel_pm.h"
14 
15 /* Parameters for Qclk Geyserville (QGV) */
16 struct intel_qgv_point {
17 	u16 dclk, t_rp, t_rdpre, t_rc, t_ras, t_rcd;
18 };
19 
20 struct intel_psf_gv_point {
21 	u8 clk; /* clock in multiples of 16.6666 MHz */
22 };
23 
24 struct intel_qgv_info {
25 	struct intel_qgv_point points[I915_NUM_QGV_POINTS];
26 	struct intel_psf_gv_point psf_points[I915_NUM_PSF_GV_POINTS];
27 	u8 num_points;
28 	u8 num_psf_points;
29 	u8 t_bl;
30 	u8 max_numchannels;
31 	u8 channel_width;
32 	u8 deinterleave;
33 };
34 
35 static int dg1_mchbar_read_qgv_point_info(struct drm_i915_private *dev_priv,
36 					  struct intel_qgv_point *sp,
37 					  int point)
38 {
39 	u32 dclk_ratio, dclk_reference;
40 	u32 val;
41 
42 	val = intel_uncore_read(&dev_priv->uncore, SA_PERF_STATUS_0_0_0_MCHBAR_PC);
43 	dclk_ratio = REG_FIELD_GET(DG1_QCLK_RATIO_MASK, val);
44 	if (val & DG1_QCLK_REFERENCE)
45 		dclk_reference = 6; /* 6 * 16.666 MHz = 100 MHz */
46 	else
47 		dclk_reference = 8; /* 8 * 16.666 MHz = 133 MHz */
48 	sp->dclk = DIV_ROUND_UP((16667 * dclk_ratio * dclk_reference) + 500, 1000);
49 
50 	val = intel_uncore_read(&dev_priv->uncore, SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU);
51 	if (val & DG1_GEAR_TYPE)
52 		sp->dclk *= 2;
53 
54 	if (sp->dclk == 0)
55 		return -EINVAL;
56 
57 	val = intel_uncore_read(&dev_priv->uncore, MCHBAR_CH0_CR_TC_PRE_0_0_0_MCHBAR);
58 	sp->t_rp = REG_FIELD_GET(DG1_DRAM_T_RP_MASK, val);
59 	sp->t_rdpre = REG_FIELD_GET(DG1_DRAM_T_RDPRE_MASK, val);
60 
61 	val = intel_uncore_read(&dev_priv->uncore, MCHBAR_CH0_CR_TC_PRE_0_0_0_MCHBAR_HIGH);
62 	sp->t_rcd = REG_FIELD_GET(DG1_DRAM_T_RCD_MASK, val);
63 	sp->t_ras = REG_FIELD_GET(DG1_DRAM_T_RAS_MASK, val);
64 
65 	sp->t_rc = sp->t_rp + sp->t_ras;
66 
67 	return 0;
68 }
69 
70 static int icl_pcode_read_qgv_point_info(struct drm_i915_private *dev_priv,
71 					 struct intel_qgv_point *sp,
72 					 int point)
73 {
74 	u32 val = 0, val2 = 0;
75 	u16 dclk;
76 	int ret;
77 
78 	ret = sandybridge_pcode_read(dev_priv,
79 				     ICL_PCODE_MEM_SUBSYSYSTEM_INFO |
80 				     ICL_PCODE_MEM_SS_READ_QGV_POINT_INFO(point),
81 				     &val, &val2);
82 	if (ret)
83 		return ret;
84 
85 	dclk = val & 0xffff;
86 	sp->dclk = DIV_ROUND_UP((16667 * dclk) + (DISPLAY_VER(dev_priv) > 11 ? 500 : 0), 1000);
87 	sp->t_rp = (val & 0xff0000) >> 16;
88 	sp->t_rcd = (val & 0xff000000) >> 24;
89 
90 	sp->t_rdpre = val2 & 0xff;
91 	sp->t_ras = (val2 & 0xff00) >> 8;
92 
93 	sp->t_rc = sp->t_rp + sp->t_ras;
94 
95 	return 0;
96 }
97 
98 static int adls_pcode_read_psf_gv_point_info(struct drm_i915_private *dev_priv,
99 					    struct intel_psf_gv_point *points)
100 {
101 	u32 val = 0;
102 	int ret;
103 	int i;
104 
105 	ret = sandybridge_pcode_read(dev_priv,
106 				     ICL_PCODE_MEM_SUBSYSYSTEM_INFO |
107 				     ADL_PCODE_MEM_SS_READ_PSF_GV_INFO,
108 				     &val, NULL);
109 	if (ret)
110 		return ret;
111 
112 	for (i = 0; i < I915_NUM_PSF_GV_POINTS; i++) {
113 		points[i].clk = val & 0xff;
114 		val >>= 8;
115 	}
116 
117 	return 0;
118 }
119 
120 int icl_pcode_restrict_qgv_points(struct drm_i915_private *dev_priv,
121 				  u32 points_mask)
122 {
123 	int ret;
124 
125 	/* bspec says to keep retrying for at least 1 ms */
126 	ret = skl_pcode_request(dev_priv, ICL_PCODE_SAGV_DE_MEM_SS_CONFIG,
127 				points_mask,
128 				ICL_PCODE_POINTS_RESTRICTED_MASK,
129 				ICL_PCODE_POINTS_RESTRICTED,
130 				1);
131 
132 	if (ret < 0) {
133 		drm_err(&dev_priv->drm, "Failed to disable qgv points (%d) points: 0x%x\n", ret, points_mask);
134 		return ret;
135 	}
136 
137 	return 0;
138 }
139 
140 static int icl_get_qgv_points(struct drm_i915_private *dev_priv,
141 			      struct intel_qgv_info *qi,
142 			      bool is_y_tile)
143 {
144 	const struct dram_info *dram_info = &dev_priv->dram_info;
145 	int i, ret;
146 
147 	qi->num_points = dram_info->num_qgv_points;
148 	qi->num_psf_points = dram_info->num_psf_gv_points;
149 
150 	if (DISPLAY_VER(dev_priv) >= 12)
151 		switch (dram_info->type) {
152 		case INTEL_DRAM_DDR4:
153 			qi->t_bl = is_y_tile ? 8 : 4;
154 			qi->max_numchannels = 2;
155 			qi->channel_width = 64;
156 			qi->deinterleave = is_y_tile ? 1 : 2;
157 			break;
158 		case INTEL_DRAM_DDR5:
159 			qi->t_bl = is_y_tile ? 16 : 8;
160 			qi->max_numchannels = 4;
161 			qi->channel_width = 32;
162 			qi->deinterleave = is_y_tile ? 1 : 2;
163 			break;
164 		case INTEL_DRAM_LPDDR4:
165 			if (IS_ROCKETLAKE(dev_priv)) {
166 				qi->t_bl = 8;
167 				qi->max_numchannels = 4;
168 				qi->channel_width = 32;
169 				qi->deinterleave = 2;
170 				break;
171 			}
172 			fallthrough;
173 		case INTEL_DRAM_LPDDR5:
174 			qi->t_bl = 16;
175 			qi->max_numchannels = 8;
176 			qi->channel_width = 16;
177 			qi->deinterleave = is_y_tile ? 2 : 4;
178 			break;
179 		default:
180 			qi->t_bl = 16;
181 			qi->max_numchannels = 1;
182 			break;
183 		}
184 	else if (DISPLAY_VER(dev_priv) == 11) {
185 		qi->t_bl = dev_priv->dram_info.type == INTEL_DRAM_DDR4 ? 4 : 8;
186 		qi->max_numchannels = 1;
187 	}
188 
189 	if (drm_WARN_ON(&dev_priv->drm,
190 			qi->num_points > ARRAY_SIZE(qi->points)))
191 		qi->num_points = ARRAY_SIZE(qi->points);
192 
193 	for (i = 0; i < qi->num_points; i++) {
194 		struct intel_qgv_point *sp = &qi->points[i];
195 
196 		if (IS_DG1(dev_priv))
197 			ret = dg1_mchbar_read_qgv_point_info(dev_priv, sp, i);
198 		else
199 			ret = icl_pcode_read_qgv_point_info(dev_priv, sp, i);
200 
201 		if (ret)
202 			return ret;
203 
204 		drm_dbg_kms(&dev_priv->drm,
205 			    "QGV %d: DCLK=%d tRP=%d tRDPRE=%d tRAS=%d tRCD=%d tRC=%d\n",
206 			    i, sp->dclk, sp->t_rp, sp->t_rdpre, sp->t_ras,
207 			    sp->t_rcd, sp->t_rc);
208 	}
209 
210 	if (qi->num_psf_points > 0) {
211 		ret = adls_pcode_read_psf_gv_point_info(dev_priv, qi->psf_points);
212 		if (ret) {
213 			drm_err(&dev_priv->drm, "Failed to read PSF point data; PSF points will not be considered in bandwidth calculations.\n");
214 			qi->num_psf_points = 0;
215 		}
216 
217 		for (i = 0; i < qi->num_psf_points; i++)
218 			drm_dbg_kms(&dev_priv->drm,
219 				    "PSF GV %d: CLK=%d \n",
220 				    i, qi->psf_points[i].clk);
221 	}
222 
223 	return 0;
224 }
225 
226 static int adl_calc_psf_bw(int clk)
227 {
228 	/*
229 	 * clk is multiples of 16.666MHz (100/6)
230 	 * According to BSpec PSF GV bandwidth is
231 	 * calculated as BW = 64 * clk * 16.666Mhz
232 	 */
233 	return DIV_ROUND_CLOSEST(64 * clk * 100, 6);
234 }
235 
236 static int icl_sagv_max_dclk(const struct intel_qgv_info *qi)
237 {
238 	u16 dclk = 0;
239 	int i;
240 
241 	for (i = 0; i < qi->num_points; i++)
242 		dclk = max(dclk, qi->points[i].dclk);
243 
244 	return dclk;
245 }
246 
247 struct intel_sa_info {
248 	u16 displayrtids;
249 	u8 deburst, deprogbwlimit, derating;
250 };
251 
252 static const struct intel_sa_info icl_sa_info = {
253 	.deburst = 8,
254 	.deprogbwlimit = 25, /* GB/s */
255 	.displayrtids = 128,
256 	.derating = 10,
257 };
258 
259 static const struct intel_sa_info tgl_sa_info = {
260 	.deburst = 16,
261 	.deprogbwlimit = 34, /* GB/s */
262 	.displayrtids = 256,
263 	.derating = 10,
264 };
265 
266 static const struct intel_sa_info rkl_sa_info = {
267 	.deburst = 8,
268 	.deprogbwlimit = 20, /* GB/s */
269 	.displayrtids = 128,
270 	.derating = 10,
271 };
272 
273 static const struct intel_sa_info adls_sa_info = {
274 	.deburst = 16,
275 	.deprogbwlimit = 38, /* GB/s */
276 	.displayrtids = 256,
277 	.derating = 10,
278 };
279 
280 static const struct intel_sa_info adlp_sa_info = {
281 	.deburst = 16,
282 	.deprogbwlimit = 38, /* GB/s */
283 	.displayrtids = 256,
284 	.derating = 20,
285 };
286 
287 static int icl_get_bw_info(struct drm_i915_private *dev_priv, const struct intel_sa_info *sa)
288 {
289 	struct intel_qgv_info qi = {};
290 	bool is_y_tile = true; /* assume y tile may be used */
291 	int num_channels = max_t(u8, 1, dev_priv->dram_info.num_channels);
292 	int ipqdepth, ipqdepthpch = 16;
293 	int dclk_max;
294 	int maxdebw;
295 	int num_groups = ARRAY_SIZE(dev_priv->max_bw);
296 	int i, ret;
297 
298 	ret = icl_get_qgv_points(dev_priv, &qi, is_y_tile);
299 	if (ret) {
300 		drm_dbg_kms(&dev_priv->drm,
301 			    "Failed to get memory subsystem information, ignoring bandwidth limits");
302 		return ret;
303 	}
304 
305 	dclk_max = icl_sagv_max_dclk(&qi);
306 	maxdebw = min(sa->deprogbwlimit * 1000, dclk_max * 16 * 6 / 10);
307 	ipqdepth = min(ipqdepthpch, sa->displayrtids / num_channels);
308 	qi.deinterleave = DIV_ROUND_UP(num_channels, is_y_tile ? 4 : 2);
309 
310 	for (i = 0; i < num_groups; i++) {
311 		struct intel_bw_info *bi = &dev_priv->max_bw[i];
312 		int clpchgroup;
313 		int j;
314 
315 		clpchgroup = (sa->deburst * qi.deinterleave / num_channels) << i;
316 		bi->num_planes = (ipqdepth - clpchgroup) / clpchgroup + 1;
317 
318 		bi->num_qgv_points = qi.num_points;
319 		bi->num_psf_gv_points = qi.num_psf_points;
320 
321 		for (j = 0; j < qi.num_points; j++) {
322 			const struct intel_qgv_point *sp = &qi.points[j];
323 			int ct, bw;
324 
325 			/*
326 			 * Max row cycle time
327 			 *
328 			 * FIXME what is the logic behind the
329 			 * assumed burst length?
330 			 */
331 			ct = max_t(int, sp->t_rc, sp->t_rp + sp->t_rcd +
332 				   (clpchgroup - 1) * qi.t_bl + sp->t_rdpre);
333 			bw = DIV_ROUND_UP(sp->dclk * clpchgroup * 32 * num_channels, ct);
334 
335 			bi->deratedbw[j] = min(maxdebw,
336 					       bw * (100 - sa->derating) / 100);
337 
338 			drm_dbg_kms(&dev_priv->drm,
339 				    "BW%d / QGV %d: num_planes=%d deratedbw=%u\n",
340 				    i, j, bi->num_planes, bi->deratedbw[j]);
341 		}
342 	}
343 	/*
344 	 * In case if SAGV is disabled in BIOS, we always get 1
345 	 * SAGV point, but we can't send PCode commands to restrict it
346 	 * as it will fail and pointless anyway.
347 	 */
348 	if (qi.num_points == 1)
349 		dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
350 	else
351 		dev_priv->sagv_status = I915_SAGV_ENABLED;
352 
353 	return 0;
354 }
355 
356 static int tgl_get_bw_info(struct drm_i915_private *dev_priv, const struct intel_sa_info *sa)
357 {
358 	struct intel_qgv_info qi = {};
359 	const struct dram_info *dram_info = &dev_priv->dram_info;
360 	bool is_y_tile = true; /* assume y tile may be used */
361 	int num_channels = max_t(u8, 1, dev_priv->dram_info.num_channels);
362 	int ipqdepth, ipqdepthpch = 16;
363 	int dclk_max;
364 	int maxdebw, peakbw;
365 	int clperchgroup;
366 	int num_groups = ARRAY_SIZE(dev_priv->max_bw);
367 	int i, ret;
368 
369 	ret = icl_get_qgv_points(dev_priv, &qi, is_y_tile);
370 	if (ret) {
371 		drm_dbg_kms(&dev_priv->drm,
372 			    "Failed to get memory subsystem information, ignoring bandwidth limits");
373 		return ret;
374 	}
375 
376 	if (dram_info->type == INTEL_DRAM_LPDDR4 || dram_info->type == INTEL_DRAM_LPDDR5)
377 		num_channels *= 2;
378 
379 	qi.deinterleave = qi.deinterleave ? : DIV_ROUND_UP(num_channels, is_y_tile ? 4 : 2);
380 
381 	if (num_channels < qi.max_numchannels && DISPLAY_VER(dev_priv) >= 12)
382 		qi.deinterleave = max(DIV_ROUND_UP(qi.deinterleave, 2), 1);
383 
384 	if (DISPLAY_VER(dev_priv) > 11 && num_channels > qi.max_numchannels)
385 		drm_warn(&dev_priv->drm, "Number of channels exceeds max number of channels.");
386 	if (qi.max_numchannels != 0)
387 		num_channels = min_t(u8, num_channels, qi.max_numchannels);
388 
389 	dclk_max = icl_sagv_max_dclk(&qi);
390 
391 	peakbw = num_channels * DIV_ROUND_UP(qi.channel_width, 8) * dclk_max;
392 	maxdebw = min(sa->deprogbwlimit * 1000, peakbw * 6 / 10); /* 60% */
393 
394 	ipqdepth = min(ipqdepthpch, sa->displayrtids / num_channels);
395 	/*
396 	 * clperchgroup = 4kpagespermempage * clperchperblock,
397 	 * clperchperblock = 8 / num_channels * interleave
398 	 */
399 	clperchgroup = 4 * DIV_ROUND_UP(8, num_channels) * qi.deinterleave;
400 
401 	for (i = 0; i < num_groups; i++) {
402 		struct intel_bw_info *bi = &dev_priv->max_bw[i];
403 		struct intel_bw_info *bi_next;
404 		int clpchgroup;
405 		int j;
406 
407 		if (i < num_groups - 1)
408 			bi_next = &dev_priv->max_bw[i + 1];
409 
410 		clpchgroup = (sa->deburst * qi.deinterleave / num_channels) << i;
411 
412 		if (i < num_groups - 1 && clpchgroup < clperchgroup)
413 			bi_next->num_planes = (ipqdepth - clpchgroup) / clpchgroup + 1;
414 		else
415 			bi_next->num_planes = 0;
416 
417 		bi->num_qgv_points = qi.num_points;
418 		bi->num_psf_gv_points = qi.num_psf_points;
419 
420 		for (j = 0; j < qi.num_points; j++) {
421 			const struct intel_qgv_point *sp = &qi.points[j];
422 			int ct, bw;
423 
424 			/*
425 			 * Max row cycle time
426 			 *
427 			 * FIXME what is the logic behind the
428 			 * assumed burst length?
429 			 */
430 			ct = max_t(int, sp->t_rc, sp->t_rp + sp->t_rcd +
431 				   (clpchgroup - 1) * qi.t_bl + sp->t_rdpre);
432 			bw = DIV_ROUND_UP(sp->dclk * clpchgroup * 32 * num_channels, ct);
433 
434 			bi->deratedbw[j] = min(maxdebw,
435 					       bw * (100 - sa->derating) / 100);
436 
437 			drm_dbg_kms(&dev_priv->drm,
438 				    "BW%d / QGV %d: num_planes=%d deratedbw=%u\n",
439 				    i, j, bi->num_planes, bi->deratedbw[j]);
440 		}
441 
442 		for (j = 0; j < qi.num_psf_points; j++) {
443 			const struct intel_psf_gv_point *sp = &qi.psf_points[j];
444 
445 			bi->psf_bw[j] = adl_calc_psf_bw(sp->clk);
446 
447 			drm_dbg_kms(&dev_priv->drm,
448 				    "BW%d / PSF GV %d: num_planes=%d bw=%u\n",
449 				    i, j, bi->num_planes, bi->psf_bw[j]);
450 		}
451 	}
452 
453 	/*
454 	 * In case if SAGV is disabled in BIOS, we always get 1
455 	 * SAGV point, but we can't send PCode commands to restrict it
456 	 * as it will fail and pointless anyway.
457 	 */
458 	if (qi.num_points == 1)
459 		dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
460 	else
461 		dev_priv->sagv_status = I915_SAGV_ENABLED;
462 
463 	return 0;
464 }
465 
466 static void dg2_get_bw_info(struct drm_i915_private *i915)
467 {
468 	struct intel_bw_info *bi = &i915->max_bw[0];
469 
470 	/*
471 	 * DG2 doesn't have SAGV or QGV points, just a constant max bandwidth
472 	 * that doesn't depend on the number of planes enabled.  Create a
473 	 * single dummy QGV point to reflect that.  DG2-G10 platforms have a
474 	 * constant 50 GB/s bandwidth, whereas DG2-G11 platforms have 38 GB/s.
475 	 */
476 	bi->num_planes = 1;
477 	bi->num_qgv_points = 1;
478 	if (IS_DG2_G11(i915))
479 		bi->deratedbw[0] = 38000;
480 	else
481 		bi->deratedbw[0] = 50000;
482 
483 	i915->sagv_status = I915_SAGV_NOT_CONTROLLED;
484 }
485 
486 static unsigned int icl_max_bw(struct drm_i915_private *dev_priv,
487 			       int num_planes, int qgv_point)
488 {
489 	int i;
490 
491 	/*
492 	 * Let's return max bw for 0 planes
493 	 */
494 	num_planes = max(1, num_planes);
495 
496 	for (i = 0; i < ARRAY_SIZE(dev_priv->max_bw); i++) {
497 		const struct intel_bw_info *bi =
498 			&dev_priv->max_bw[i];
499 
500 		/*
501 		 * Pcode will not expose all QGV points when
502 		 * SAGV is forced to off/min/med/max.
503 		 */
504 		if (qgv_point >= bi->num_qgv_points)
505 			return UINT_MAX;
506 
507 		if (num_planes >= bi->num_planes)
508 			return bi->deratedbw[qgv_point];
509 	}
510 
511 	return 0;
512 }
513 
514 static unsigned int tgl_max_bw(struct drm_i915_private *dev_priv,
515 			       int num_planes, int qgv_point)
516 {
517 	int i;
518 
519 	/*
520 	 * Let's return max bw for 0 planes
521 	 */
522 	num_planes = max(1, num_planes);
523 
524 	for (i = ARRAY_SIZE(dev_priv->max_bw) - 1; i >= 0; i--) {
525 		const struct intel_bw_info *bi =
526 			&dev_priv->max_bw[i];
527 
528 		/*
529 		 * Pcode will not expose all QGV points when
530 		 * SAGV is forced to off/min/med/max.
531 		 */
532 		if (qgv_point >= bi->num_qgv_points)
533 			return UINT_MAX;
534 
535 		if (num_planes <= bi->num_planes)
536 			return bi->deratedbw[qgv_point];
537 	}
538 
539 	return dev_priv->max_bw[0].deratedbw[qgv_point];
540 }
541 
542 static unsigned int adl_psf_bw(struct drm_i915_private *dev_priv,
543 			       int psf_gv_point)
544 {
545 	const struct intel_bw_info *bi =
546 			&dev_priv->max_bw[0];
547 
548 	return bi->psf_bw[psf_gv_point];
549 }
550 
551 void intel_bw_init_hw(struct drm_i915_private *dev_priv)
552 {
553 	if (!HAS_DISPLAY(dev_priv))
554 		return;
555 
556 	if (IS_DG2(dev_priv))
557 		dg2_get_bw_info(dev_priv);
558 	else if (IS_ALDERLAKE_P(dev_priv))
559 		tgl_get_bw_info(dev_priv, &adlp_sa_info);
560 	else if (IS_ALDERLAKE_S(dev_priv))
561 		tgl_get_bw_info(dev_priv, &adls_sa_info);
562 	else if (IS_ROCKETLAKE(dev_priv))
563 		tgl_get_bw_info(dev_priv, &rkl_sa_info);
564 	else if (DISPLAY_VER(dev_priv) == 12)
565 		tgl_get_bw_info(dev_priv, &tgl_sa_info);
566 	else if (DISPLAY_VER(dev_priv) == 11)
567 		icl_get_bw_info(dev_priv, &icl_sa_info);
568 }
569 
570 static unsigned int intel_bw_crtc_num_active_planes(const struct intel_crtc_state *crtc_state)
571 {
572 	/*
573 	 * We assume cursors are small enough
574 	 * to not not cause bandwidth problems.
575 	 */
576 	return hweight8(crtc_state->active_planes & ~BIT(PLANE_CURSOR));
577 }
578 
579 static unsigned int intel_bw_crtc_data_rate(const struct intel_crtc_state *crtc_state)
580 {
581 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
582 	unsigned int data_rate = 0;
583 	enum plane_id plane_id;
584 
585 	for_each_plane_id_on_crtc(crtc, plane_id) {
586 		/*
587 		 * We assume cursors are small enough
588 		 * to not not cause bandwidth problems.
589 		 */
590 		if (plane_id == PLANE_CURSOR)
591 			continue;
592 
593 		data_rate += crtc_state->data_rate[plane_id];
594 	}
595 
596 	return data_rate;
597 }
598 
599 void intel_bw_crtc_update(struct intel_bw_state *bw_state,
600 			  const struct intel_crtc_state *crtc_state)
601 {
602 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
603 	struct drm_i915_private *i915 = to_i915(crtc->base.dev);
604 
605 	bw_state->data_rate[crtc->pipe] =
606 		intel_bw_crtc_data_rate(crtc_state);
607 	bw_state->num_active_planes[crtc->pipe] =
608 		intel_bw_crtc_num_active_planes(crtc_state);
609 
610 	drm_dbg_kms(&i915->drm, "pipe %c data rate %u num active planes %u\n",
611 		    pipe_name(crtc->pipe),
612 		    bw_state->data_rate[crtc->pipe],
613 		    bw_state->num_active_planes[crtc->pipe]);
614 }
615 
616 static unsigned int intel_bw_num_active_planes(struct drm_i915_private *dev_priv,
617 					       const struct intel_bw_state *bw_state)
618 {
619 	unsigned int num_active_planes = 0;
620 	enum pipe pipe;
621 
622 	for_each_pipe(dev_priv, pipe)
623 		num_active_planes += bw_state->num_active_planes[pipe];
624 
625 	return num_active_planes;
626 }
627 
628 static unsigned int intel_bw_data_rate(struct drm_i915_private *dev_priv,
629 				       const struct intel_bw_state *bw_state)
630 {
631 	unsigned int data_rate = 0;
632 	enum pipe pipe;
633 
634 	for_each_pipe(dev_priv, pipe)
635 		data_rate += bw_state->data_rate[pipe];
636 
637 	if (DISPLAY_VER(dev_priv) >= 13 && intel_vtd_active(dev_priv))
638 		data_rate = data_rate * 105 / 100;
639 
640 	return data_rate;
641 }
642 
643 struct intel_bw_state *
644 intel_atomic_get_old_bw_state(struct intel_atomic_state *state)
645 {
646 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
647 	struct intel_global_state *bw_state;
648 
649 	bw_state = intel_atomic_get_old_global_obj_state(state, &dev_priv->bw_obj);
650 
651 	return to_intel_bw_state(bw_state);
652 }
653 
654 struct intel_bw_state *
655 intel_atomic_get_new_bw_state(struct intel_atomic_state *state)
656 {
657 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
658 	struct intel_global_state *bw_state;
659 
660 	bw_state = intel_atomic_get_new_global_obj_state(state, &dev_priv->bw_obj);
661 
662 	return to_intel_bw_state(bw_state);
663 }
664 
665 struct intel_bw_state *
666 intel_atomic_get_bw_state(struct intel_atomic_state *state)
667 {
668 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
669 	struct intel_global_state *bw_state;
670 
671 	bw_state = intel_atomic_get_global_obj_state(state, &dev_priv->bw_obj);
672 	if (IS_ERR(bw_state))
673 		return ERR_CAST(bw_state);
674 
675 	return to_intel_bw_state(bw_state);
676 }
677 
678 int skl_bw_calc_min_cdclk(struct intel_atomic_state *state)
679 {
680 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
681 	struct intel_bw_state *new_bw_state = NULL;
682 	struct intel_bw_state *old_bw_state = NULL;
683 	const struct intel_crtc_state *crtc_state;
684 	struct intel_crtc *crtc;
685 	int max_bw = 0;
686 	enum pipe pipe;
687 	int i;
688 
689 	for_each_new_intel_crtc_in_state(state, crtc, crtc_state, i) {
690 		enum plane_id plane_id;
691 		struct intel_dbuf_bw *crtc_bw;
692 
693 		new_bw_state = intel_atomic_get_bw_state(state);
694 		if (IS_ERR(new_bw_state))
695 			return PTR_ERR(new_bw_state);
696 
697 		old_bw_state = intel_atomic_get_old_bw_state(state);
698 
699 		crtc_bw = &new_bw_state->dbuf_bw[crtc->pipe];
700 
701 		memset(&crtc_bw->used_bw, 0, sizeof(crtc_bw->used_bw));
702 
703 		if (!crtc_state->hw.active)
704 			continue;
705 
706 		for_each_plane_id_on_crtc(crtc, plane_id) {
707 			const struct skl_ddb_entry *plane_alloc =
708 				&crtc_state->wm.skl.plane_ddb_y[plane_id];
709 			const struct skl_ddb_entry *uv_plane_alloc =
710 				&crtc_state->wm.skl.plane_ddb_uv[plane_id];
711 			unsigned int data_rate = crtc_state->data_rate[plane_id];
712 			unsigned int dbuf_mask = 0;
713 			enum dbuf_slice slice;
714 
715 			dbuf_mask |= skl_ddb_dbuf_slice_mask(dev_priv, plane_alloc);
716 			dbuf_mask |= skl_ddb_dbuf_slice_mask(dev_priv, uv_plane_alloc);
717 
718 			/*
719 			 * FIXME: To calculate that more properly we probably
720 			 * need to to split per plane data_rate into data_rate_y
721 			 * and data_rate_uv for multiplanar formats in order not
722 			 * to get accounted those twice if they happen to reside
723 			 * on different slices.
724 			 * However for pre-icl this would work anyway because
725 			 * we have only single slice and for icl+ uv plane has
726 			 * non-zero data rate.
727 			 * So in worst case those calculation are a bit
728 			 * pessimistic, which shouldn't pose any significant
729 			 * problem anyway.
730 			 */
731 			for_each_dbuf_slice_in_mask(dev_priv, slice, dbuf_mask)
732 				crtc_bw->used_bw[slice] += data_rate;
733 		}
734 	}
735 
736 	if (!old_bw_state)
737 		return 0;
738 
739 	for_each_pipe(dev_priv, pipe) {
740 		struct intel_dbuf_bw *crtc_bw;
741 		enum dbuf_slice slice;
742 
743 		crtc_bw = &new_bw_state->dbuf_bw[pipe];
744 
745 		for_each_dbuf_slice(dev_priv, slice) {
746 			/*
747 			 * Current experimental observations show that contrary
748 			 * to BSpec we get underruns once we exceed 64 * CDCLK
749 			 * for slices in total.
750 			 * As a temporary measure in order not to keep CDCLK
751 			 * bumped up all the time we calculate CDCLK according
752 			 * to this formula for  overall bw consumed by slices.
753 			 */
754 			max_bw += crtc_bw->used_bw[slice];
755 		}
756 	}
757 
758 	new_bw_state->min_cdclk = max_bw / 64;
759 
760 	if (new_bw_state->min_cdclk != old_bw_state->min_cdclk) {
761 		int ret = intel_atomic_lock_global_state(&new_bw_state->base);
762 
763 		if (ret)
764 			return ret;
765 	}
766 
767 	return 0;
768 }
769 
770 int intel_bw_calc_min_cdclk(struct intel_atomic_state *state)
771 {
772 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
773 	struct intel_bw_state *new_bw_state = NULL;
774 	struct intel_bw_state *old_bw_state = NULL;
775 	const struct intel_crtc_state *crtc_state;
776 	struct intel_crtc *crtc;
777 	int min_cdclk = 0;
778 	enum pipe pipe;
779 	int i;
780 
781 	for_each_new_intel_crtc_in_state(state, crtc, crtc_state, i) {
782 		new_bw_state = intel_atomic_get_bw_state(state);
783 		if (IS_ERR(new_bw_state))
784 			return PTR_ERR(new_bw_state);
785 
786 		old_bw_state = intel_atomic_get_old_bw_state(state);
787 	}
788 
789 	if (!old_bw_state)
790 		return 0;
791 
792 	for_each_pipe(dev_priv, pipe) {
793 		struct intel_cdclk_state *cdclk_state;
794 
795 		cdclk_state = intel_atomic_get_new_cdclk_state(state);
796 		if (!cdclk_state)
797 			return 0;
798 
799 		min_cdclk = max(cdclk_state->min_cdclk[pipe], min_cdclk);
800 	}
801 
802 	new_bw_state->min_cdclk = min_cdclk;
803 
804 	if (new_bw_state->min_cdclk != old_bw_state->min_cdclk) {
805 		int ret = intel_atomic_lock_global_state(&new_bw_state->base);
806 
807 		if (ret)
808 			return ret;
809 	}
810 
811 	return 0;
812 }
813 
814 int intel_bw_atomic_check(struct intel_atomic_state *state)
815 {
816 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
817 	struct intel_crtc_state *new_crtc_state, *old_crtc_state;
818 	struct intel_bw_state *new_bw_state = NULL;
819 	const struct intel_bw_state *old_bw_state = NULL;
820 	unsigned int data_rate;
821 	unsigned int num_active_planes;
822 	struct intel_crtc *crtc;
823 	int i, ret;
824 	u32 allowed_points = 0;
825 	unsigned int max_bw_point = 0, max_bw = 0;
826 	unsigned int num_qgv_points = dev_priv->max_bw[0].num_qgv_points;
827 	unsigned int num_psf_gv_points = dev_priv->max_bw[0].num_psf_gv_points;
828 	bool changed = false;
829 	u32 mask = 0;
830 
831 	/* FIXME earlier gens need some checks too */
832 	if (DISPLAY_VER(dev_priv) < 11)
833 		return 0;
834 
835 	/*
836 	 * We can _not_ use the whole ADLS_QGV_PT_MASK here, as PCode rejects
837 	 * it with failure if we try masking any unadvertised points.
838 	 * So need to operate only with those returned from PCode.
839 	 */
840 	if (num_qgv_points > 0)
841 		mask |= REG_GENMASK(num_qgv_points - 1, 0);
842 
843 	if (num_psf_gv_points > 0)
844 		mask |= REG_GENMASK(num_psf_gv_points - 1, 0) << ADLS_PSF_PT_SHIFT;
845 
846 	for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
847 					    new_crtc_state, i) {
848 		unsigned int old_data_rate =
849 			intel_bw_crtc_data_rate(old_crtc_state);
850 		unsigned int new_data_rate =
851 			intel_bw_crtc_data_rate(new_crtc_state);
852 		unsigned int old_active_planes =
853 			intel_bw_crtc_num_active_planes(old_crtc_state);
854 		unsigned int new_active_planes =
855 			intel_bw_crtc_num_active_planes(new_crtc_state);
856 
857 		/*
858 		 * Avoid locking the bw state when
859 		 * nothing significant has changed.
860 		 */
861 		if (old_data_rate == new_data_rate &&
862 		    old_active_planes == new_active_planes)
863 			continue;
864 
865 		new_bw_state = intel_atomic_get_bw_state(state);
866 		if (IS_ERR(new_bw_state))
867 			return PTR_ERR(new_bw_state);
868 
869 		new_bw_state->data_rate[crtc->pipe] = new_data_rate;
870 		new_bw_state->num_active_planes[crtc->pipe] = new_active_planes;
871 
872 		changed = true;
873 
874 		drm_dbg_kms(&dev_priv->drm,
875 			    "pipe %c data rate %u num active planes %u\n",
876 			    pipe_name(crtc->pipe),
877 			    new_bw_state->data_rate[crtc->pipe],
878 			    new_bw_state->num_active_planes[crtc->pipe]);
879 	}
880 
881 	old_bw_state = intel_atomic_get_old_bw_state(state);
882 	new_bw_state = intel_atomic_get_new_bw_state(state);
883 
884 	if (new_bw_state &&
885 	    intel_can_enable_sagv(dev_priv, old_bw_state) !=
886 	    intel_can_enable_sagv(dev_priv, new_bw_state))
887 		changed = true;
888 
889 	/*
890 	 * If none of our inputs (data rates, number of active
891 	 * planes, SAGV yes/no) changed then nothing to do here.
892 	 */
893 	if (!changed)
894 		return 0;
895 
896 	ret = intel_atomic_lock_global_state(&new_bw_state->base);
897 	if (ret)
898 		return ret;
899 
900 	data_rate = intel_bw_data_rate(dev_priv, new_bw_state);
901 	data_rate = DIV_ROUND_UP(data_rate, 1000);
902 
903 	num_active_planes = intel_bw_num_active_planes(dev_priv, new_bw_state);
904 
905 	for (i = 0; i < num_qgv_points; i++) {
906 		unsigned int max_data_rate;
907 
908 		if (DISPLAY_VER(dev_priv) > 11)
909 			max_data_rate = tgl_max_bw(dev_priv, num_active_planes, i);
910 		else
911 			max_data_rate = icl_max_bw(dev_priv, num_active_planes, i);
912 		/*
913 		 * We need to know which qgv point gives us
914 		 * maximum bandwidth in order to disable SAGV
915 		 * if we find that we exceed SAGV block time
916 		 * with watermarks. By that moment we already
917 		 * have those, as it is calculated earlier in
918 		 * intel_atomic_check,
919 		 */
920 		if (max_data_rate > max_bw) {
921 			max_bw_point = i;
922 			max_bw = max_data_rate;
923 		}
924 		if (max_data_rate >= data_rate)
925 			allowed_points |= REG_FIELD_PREP(ADLS_QGV_PT_MASK, BIT(i));
926 
927 		drm_dbg_kms(&dev_priv->drm, "QGV point %d: max bw %d required %d\n",
928 			    i, max_data_rate, data_rate);
929 	}
930 
931 	for (i = 0; i < num_psf_gv_points; i++) {
932 		unsigned int max_data_rate = adl_psf_bw(dev_priv, i);
933 
934 		if (max_data_rate >= data_rate)
935 			allowed_points |= REG_FIELD_PREP(ADLS_PSF_PT_MASK, BIT(i));
936 
937 		drm_dbg_kms(&dev_priv->drm, "PSF GV point %d: max bw %d"
938 			    " required %d\n",
939 			    i, max_data_rate, data_rate);
940 	}
941 
942 	/*
943 	 * BSpec states that we always should have at least one allowed point
944 	 * left, so if we couldn't - simply reject the configuration for obvious
945 	 * reasons.
946 	 */
947 	if ((allowed_points & ADLS_QGV_PT_MASK) == 0) {
948 		drm_dbg_kms(&dev_priv->drm, "No QGV points provide sufficient memory"
949 			    " bandwidth %d for display configuration(%d active planes).\n",
950 			    data_rate, num_active_planes);
951 		return -EINVAL;
952 	}
953 
954 	if (num_psf_gv_points > 0) {
955 		if ((allowed_points & ADLS_PSF_PT_MASK) == 0) {
956 			drm_dbg_kms(&dev_priv->drm, "No PSF GV points provide sufficient memory"
957 				    " bandwidth %d for display configuration(%d active planes).\n",
958 				    data_rate, num_active_planes);
959 			return -EINVAL;
960 		}
961 	}
962 
963 	/*
964 	 * Leave only single point with highest bandwidth, if
965 	 * we can't enable SAGV due to the increased memory latency it may
966 	 * cause.
967 	 */
968 	if (!intel_can_enable_sagv(dev_priv, new_bw_state)) {
969 		allowed_points = BIT(max_bw_point);
970 		drm_dbg_kms(&dev_priv->drm, "No SAGV, using single QGV point %d\n",
971 			    max_bw_point);
972 	}
973 	/*
974 	 * We store the ones which need to be masked as that is what PCode
975 	 * actually accepts as a parameter.
976 	 */
977 	new_bw_state->qgv_points_mask = ~allowed_points & mask;
978 
979 	/*
980 	 * If the actual mask had changed we need to make sure that
981 	 * the commits are serialized(in case this is a nomodeset, nonblocking)
982 	 */
983 	if (new_bw_state->qgv_points_mask != old_bw_state->qgv_points_mask) {
984 		ret = intel_atomic_serialize_global_state(&new_bw_state->base);
985 		if (ret)
986 			return ret;
987 	}
988 
989 	return 0;
990 }
991 
992 static struct intel_global_state *
993 intel_bw_duplicate_state(struct intel_global_obj *obj)
994 {
995 	struct intel_bw_state *state;
996 
997 	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
998 	if (!state)
999 		return NULL;
1000 
1001 	return &state->base;
1002 }
1003 
1004 static void intel_bw_destroy_state(struct intel_global_obj *obj,
1005 				   struct intel_global_state *state)
1006 {
1007 	kfree(state);
1008 }
1009 
1010 static const struct intel_global_state_funcs intel_bw_funcs = {
1011 	.atomic_duplicate_state = intel_bw_duplicate_state,
1012 	.atomic_destroy_state = intel_bw_destroy_state,
1013 };
1014 
1015 int intel_bw_init(struct drm_i915_private *dev_priv)
1016 {
1017 	struct intel_bw_state *state;
1018 
1019 	state = kzalloc(sizeof(*state), GFP_KERNEL);
1020 	if (!state)
1021 		return -ENOMEM;
1022 
1023 	intel_atomic_global_obj_init(dev_priv, &dev_priv->bw_obj,
1024 				     &state->base, &intel_bw_funcs);
1025 
1026 	return 0;
1027 }
1028