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_bw.h"
9 #include "intel_display_types.h"
10 #include "intel_sideband.h"
11 
12 /* Parameters for Qclk Geyserville (QGV) */
13 struct intel_qgv_point {
14 	u16 dclk, t_rp, t_rdpre, t_rc, t_ras, t_rcd;
15 };
16 
17 struct intel_qgv_info {
18 	struct intel_qgv_point points[I915_NUM_QGV_POINTS];
19 	u8 num_points;
20 	u8 num_channels;
21 	u8 t_bl;
22 	enum intel_dram_type dram_type;
23 };
24 
25 static int icl_pcode_read_mem_global_info(struct drm_i915_private *dev_priv,
26 					  struct intel_qgv_info *qi)
27 {
28 	u32 val = 0;
29 	int ret;
30 
31 	ret = sandybridge_pcode_read(dev_priv,
32 				     ICL_PCODE_MEM_SUBSYSYSTEM_INFO |
33 				     ICL_PCODE_MEM_SS_READ_GLOBAL_INFO,
34 				     &val, NULL);
35 	if (ret)
36 		return ret;
37 
38 	if (IS_GEN(dev_priv, 12)) {
39 		switch (val & 0xf) {
40 		case 0:
41 			qi->dram_type = INTEL_DRAM_DDR4;
42 			break;
43 		case 3:
44 			qi->dram_type = INTEL_DRAM_LPDDR4;
45 			break;
46 		case 4:
47 			qi->dram_type = INTEL_DRAM_DDR3;
48 			break;
49 		case 5:
50 			qi->dram_type = INTEL_DRAM_LPDDR3;
51 			break;
52 		default:
53 			MISSING_CASE(val & 0xf);
54 			break;
55 		}
56 	} else if (IS_GEN(dev_priv, 11)) {
57 		switch (val & 0xf) {
58 		case 0:
59 			qi->dram_type = INTEL_DRAM_DDR4;
60 			break;
61 		case 1:
62 			qi->dram_type = INTEL_DRAM_DDR3;
63 			break;
64 		case 2:
65 			qi->dram_type = INTEL_DRAM_LPDDR3;
66 			break;
67 		case 3:
68 			qi->dram_type = INTEL_DRAM_LPDDR4;
69 			break;
70 		default:
71 			MISSING_CASE(val & 0xf);
72 			break;
73 		}
74 	} else {
75 		MISSING_CASE(INTEL_GEN(dev_priv));
76 		qi->dram_type = INTEL_DRAM_LPDDR3; /* Conservative default */
77 	}
78 
79 	qi->num_channels = (val & 0xf0) >> 4;
80 	qi->num_points = (val & 0xf00) >> 8;
81 
82 	if (IS_GEN(dev_priv, 12))
83 		qi->t_bl = qi->dram_type == INTEL_DRAM_DDR4 ? 4 : 16;
84 	else if (IS_GEN(dev_priv, 11))
85 		qi->t_bl = qi->dram_type == INTEL_DRAM_DDR4 ? 4 : 8;
86 
87 	return 0;
88 }
89 
90 static int icl_pcode_read_qgv_point_info(struct drm_i915_private *dev_priv,
91 					 struct intel_qgv_point *sp,
92 					 int point)
93 {
94 	u32 val = 0, val2 = 0;
95 	int ret;
96 
97 	ret = sandybridge_pcode_read(dev_priv,
98 				     ICL_PCODE_MEM_SUBSYSYSTEM_INFO |
99 				     ICL_PCODE_MEM_SS_READ_QGV_POINT_INFO(point),
100 				     &val, &val2);
101 	if (ret)
102 		return ret;
103 
104 	sp->dclk = val & 0xffff;
105 	sp->t_rp = (val & 0xff0000) >> 16;
106 	sp->t_rcd = (val & 0xff000000) >> 24;
107 
108 	sp->t_rdpre = val2 & 0xff;
109 	sp->t_ras = (val2 & 0xff00) >> 8;
110 
111 	sp->t_rc = sp->t_rp + sp->t_ras;
112 
113 	return 0;
114 }
115 
116 static int icl_get_qgv_points(struct drm_i915_private *dev_priv,
117 			      struct intel_qgv_info *qi)
118 {
119 	int i, ret;
120 
121 	ret = icl_pcode_read_mem_global_info(dev_priv, qi);
122 	if (ret)
123 		return ret;
124 
125 	if (WARN_ON(qi->num_points > ARRAY_SIZE(qi->points)))
126 		qi->num_points = ARRAY_SIZE(qi->points);
127 
128 	for (i = 0; i < qi->num_points; i++) {
129 		struct intel_qgv_point *sp = &qi->points[i];
130 
131 		ret = icl_pcode_read_qgv_point_info(dev_priv, sp, i);
132 		if (ret)
133 			return ret;
134 
135 		DRM_DEBUG_KMS("QGV %d: DCLK=%d tRP=%d tRDPRE=%d tRAS=%d tRCD=%d tRC=%d\n",
136 			      i, sp->dclk, sp->t_rp, sp->t_rdpre, sp->t_ras,
137 			      sp->t_rcd, sp->t_rc);
138 	}
139 
140 	return 0;
141 }
142 
143 static int icl_calc_bw(int dclk, int num, int den)
144 {
145 	/* multiples of 16.666MHz (100/6) */
146 	return DIV_ROUND_CLOSEST(num * dclk * 100, den * 6);
147 }
148 
149 static int icl_sagv_max_dclk(const struct intel_qgv_info *qi)
150 {
151 	u16 dclk = 0;
152 	int i;
153 
154 	for (i = 0; i < qi->num_points; i++)
155 		dclk = max(dclk, qi->points[i].dclk);
156 
157 	return dclk;
158 }
159 
160 struct intel_sa_info {
161 	u16 displayrtids;
162 	u8 deburst, deprogbwlimit;
163 };
164 
165 static const struct intel_sa_info icl_sa_info = {
166 	.deburst = 8,
167 	.deprogbwlimit = 25, /* GB/s */
168 	.displayrtids = 128,
169 };
170 
171 static const struct intel_sa_info tgl_sa_info = {
172 	.deburst = 16,
173 	.deprogbwlimit = 34, /* GB/s */
174 	.displayrtids = 256,
175 };
176 
177 static int icl_get_bw_info(struct drm_i915_private *dev_priv, const struct intel_sa_info *sa)
178 {
179 	struct intel_qgv_info qi = {};
180 	bool is_y_tile = true; /* assume y tile may be used */
181 	int num_channels;
182 	int deinterleave;
183 	int ipqdepth, ipqdepthpch;
184 	int dclk_max;
185 	int maxdebw;
186 	int i, ret;
187 
188 	ret = icl_get_qgv_points(dev_priv, &qi);
189 	if (ret) {
190 		DRM_DEBUG_KMS("Failed to get memory subsystem information, ignoring bandwidth limits");
191 		return ret;
192 	}
193 	num_channels = qi.num_channels;
194 
195 	deinterleave = DIV_ROUND_UP(num_channels, is_y_tile ? 4 : 2);
196 	dclk_max = icl_sagv_max_dclk(&qi);
197 
198 	ipqdepthpch = 16;
199 
200 	maxdebw = min(sa->deprogbwlimit * 1000,
201 		      icl_calc_bw(dclk_max, 16, 1) * 6 / 10); /* 60% */
202 	ipqdepth = min(ipqdepthpch, sa->displayrtids / num_channels);
203 
204 	for (i = 0; i < ARRAY_SIZE(dev_priv->max_bw); i++) {
205 		struct intel_bw_info *bi = &dev_priv->max_bw[i];
206 		int clpchgroup;
207 		int j;
208 
209 		clpchgroup = (sa->deburst * deinterleave / num_channels) << i;
210 		bi->num_planes = (ipqdepth - clpchgroup) / clpchgroup + 1;
211 
212 		bi->num_qgv_points = qi.num_points;
213 
214 		for (j = 0; j < qi.num_points; j++) {
215 			const struct intel_qgv_point *sp = &qi.points[j];
216 			int ct, bw;
217 
218 			/*
219 			 * Max row cycle time
220 			 *
221 			 * FIXME what is the logic behind the
222 			 * assumed burst length?
223 			 */
224 			ct = max_t(int, sp->t_rc, sp->t_rp + sp->t_rcd +
225 				   (clpchgroup - 1) * qi.t_bl + sp->t_rdpre);
226 			bw = icl_calc_bw(sp->dclk, clpchgroup * 32 * num_channels, ct);
227 
228 			bi->deratedbw[j] = min(maxdebw,
229 					       bw * 9 / 10); /* 90% */
230 
231 			DRM_DEBUG_KMS("BW%d / QGV %d: num_planes=%d deratedbw=%u\n",
232 				      i, j, bi->num_planes, bi->deratedbw[j]);
233 		}
234 
235 		if (bi->num_planes == 1)
236 			break;
237 	}
238 
239 	return 0;
240 }
241 
242 static unsigned int icl_max_bw(struct drm_i915_private *dev_priv,
243 			       int num_planes, int qgv_point)
244 {
245 	int i;
246 
247 	for (i = 0; i < ARRAY_SIZE(dev_priv->max_bw); i++) {
248 		const struct intel_bw_info *bi =
249 			&dev_priv->max_bw[i];
250 
251 		/*
252 		 * Pcode will not expose all QGV points when
253 		 * SAGV is forced to off/min/med/max.
254 		 */
255 		if (qgv_point >= bi->num_qgv_points)
256 			return UINT_MAX;
257 
258 		if (num_planes >= bi->num_planes)
259 			return bi->deratedbw[qgv_point];
260 	}
261 
262 	return 0;
263 }
264 
265 void intel_bw_init_hw(struct drm_i915_private *dev_priv)
266 {
267 	if (!HAS_DISPLAY(dev_priv))
268 		return;
269 
270 	if (IS_GEN(dev_priv, 12))
271 		icl_get_bw_info(dev_priv, &tgl_sa_info);
272 	else if (IS_GEN(dev_priv, 11))
273 		icl_get_bw_info(dev_priv, &icl_sa_info);
274 }
275 
276 static unsigned int intel_max_data_rate(struct drm_i915_private *dev_priv,
277 					int num_planes)
278 {
279 	if (INTEL_GEN(dev_priv) >= 11) {
280 		/*
281 		 * Any bw group has same amount of QGV points
282 		 */
283 		const struct intel_bw_info *bi =
284 			&dev_priv->max_bw[0];
285 		unsigned int min_bw = UINT_MAX;
286 		int i;
287 
288 		/*
289 		 * FIXME with SAGV disabled maybe we can assume
290 		 * point 1 will always be used? Seems to match
291 		 * the behaviour observed in the wild.
292 		 */
293 		for (i = 0; i < bi->num_qgv_points; i++) {
294 			unsigned int bw = icl_max_bw(dev_priv, num_planes, i);
295 
296 			min_bw = min(bw, min_bw);
297 		}
298 		return min_bw;
299 	} else {
300 		return UINT_MAX;
301 	}
302 }
303 
304 static unsigned int intel_bw_crtc_num_active_planes(const struct intel_crtc_state *crtc_state)
305 {
306 	/*
307 	 * We assume cursors are small enough
308 	 * to not not cause bandwidth problems.
309 	 */
310 	return hweight8(crtc_state->active_planes & ~BIT(PLANE_CURSOR));
311 }
312 
313 static unsigned int intel_bw_crtc_data_rate(const struct intel_crtc_state *crtc_state)
314 {
315 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
316 	unsigned int data_rate = 0;
317 	enum plane_id plane_id;
318 
319 	for_each_plane_id_on_crtc(crtc, plane_id) {
320 		/*
321 		 * We assume cursors are small enough
322 		 * to not not cause bandwidth problems.
323 		 */
324 		if (plane_id == PLANE_CURSOR)
325 			continue;
326 
327 		data_rate += crtc_state->data_rate[plane_id];
328 	}
329 
330 	return data_rate;
331 }
332 
333 void intel_bw_crtc_update(struct intel_bw_state *bw_state,
334 			  const struct intel_crtc_state *crtc_state)
335 {
336 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
337 
338 	bw_state->data_rate[crtc->pipe] =
339 		intel_bw_crtc_data_rate(crtc_state);
340 	bw_state->num_active_planes[crtc->pipe] =
341 		intel_bw_crtc_num_active_planes(crtc_state);
342 
343 	DRM_DEBUG_KMS("pipe %c data rate %u num active planes %u\n",
344 		      pipe_name(crtc->pipe),
345 		      bw_state->data_rate[crtc->pipe],
346 		      bw_state->num_active_planes[crtc->pipe]);
347 }
348 
349 static unsigned int intel_bw_num_active_planes(struct drm_i915_private *dev_priv,
350 					       const struct intel_bw_state *bw_state)
351 {
352 	unsigned int num_active_planes = 0;
353 	enum pipe pipe;
354 
355 	for_each_pipe(dev_priv, pipe)
356 		num_active_planes += bw_state->num_active_planes[pipe];
357 
358 	return num_active_planes;
359 }
360 
361 static unsigned int intel_bw_data_rate(struct drm_i915_private *dev_priv,
362 				       const struct intel_bw_state *bw_state)
363 {
364 	unsigned int data_rate = 0;
365 	enum pipe pipe;
366 
367 	for_each_pipe(dev_priv, pipe)
368 		data_rate += bw_state->data_rate[pipe];
369 
370 	return data_rate;
371 }
372 
373 static struct intel_bw_state *
374 intel_atomic_get_bw_state(struct intel_atomic_state *state)
375 {
376 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
377 	struct drm_private_state *bw_state;
378 
379 	bw_state = drm_atomic_get_private_obj_state(&state->base,
380 						    &dev_priv->bw_obj);
381 	if (IS_ERR(bw_state))
382 		return ERR_CAST(bw_state);
383 
384 	return to_intel_bw_state(bw_state);
385 }
386 
387 int intel_bw_atomic_check(struct intel_atomic_state *state)
388 {
389 	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
390 	struct intel_crtc_state *new_crtc_state, *old_crtc_state;
391 	struct intel_bw_state *bw_state = NULL;
392 	unsigned int data_rate, max_data_rate;
393 	unsigned int num_active_planes;
394 	struct intel_crtc *crtc;
395 	int i;
396 
397 	/* FIXME earlier gens need some checks too */
398 	if (INTEL_GEN(dev_priv) < 11)
399 		return 0;
400 
401 	for_each_oldnew_intel_crtc_in_state(state, crtc, old_crtc_state,
402 					    new_crtc_state, i) {
403 		unsigned int old_data_rate =
404 			intel_bw_crtc_data_rate(old_crtc_state);
405 		unsigned int new_data_rate =
406 			intel_bw_crtc_data_rate(new_crtc_state);
407 		unsigned int old_active_planes =
408 			intel_bw_crtc_num_active_planes(old_crtc_state);
409 		unsigned int new_active_planes =
410 			intel_bw_crtc_num_active_planes(new_crtc_state);
411 
412 		/*
413 		 * Avoid locking the bw state when
414 		 * nothing significant has changed.
415 		 */
416 		if (old_data_rate == new_data_rate &&
417 		    old_active_planes == new_active_planes)
418 			continue;
419 
420 		bw_state  = intel_atomic_get_bw_state(state);
421 		if (IS_ERR(bw_state))
422 			return PTR_ERR(bw_state);
423 
424 		bw_state->data_rate[crtc->pipe] = new_data_rate;
425 		bw_state->num_active_planes[crtc->pipe] = new_active_planes;
426 
427 		DRM_DEBUG_KMS("pipe %c data rate %u num active planes %u\n",
428 			      pipe_name(crtc->pipe),
429 			      bw_state->data_rate[crtc->pipe],
430 			      bw_state->num_active_planes[crtc->pipe]);
431 	}
432 
433 	if (!bw_state)
434 		return 0;
435 
436 	data_rate = intel_bw_data_rate(dev_priv, bw_state);
437 	num_active_planes = intel_bw_num_active_planes(dev_priv, bw_state);
438 
439 	max_data_rate = intel_max_data_rate(dev_priv, num_active_planes);
440 
441 	data_rate = DIV_ROUND_UP(data_rate, 1000);
442 
443 	if (data_rate > max_data_rate) {
444 		DRM_DEBUG_KMS("Bandwidth %u MB/s exceeds max available %d MB/s (%d active planes)\n",
445 			      data_rate, max_data_rate, num_active_planes);
446 		return -EINVAL;
447 	}
448 
449 	return 0;
450 }
451 
452 static struct drm_private_state *intel_bw_duplicate_state(struct drm_private_obj *obj)
453 {
454 	struct intel_bw_state *state;
455 
456 	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
457 	if (!state)
458 		return NULL;
459 
460 	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
461 
462 	return &state->base;
463 }
464 
465 static void intel_bw_destroy_state(struct drm_private_obj *obj,
466 				   struct drm_private_state *state)
467 {
468 	kfree(state);
469 }
470 
471 static const struct drm_private_state_funcs intel_bw_funcs = {
472 	.atomic_duplicate_state = intel_bw_duplicate_state,
473 	.atomic_destroy_state = intel_bw_destroy_state,
474 };
475 
476 int intel_bw_init(struct drm_i915_private *dev_priv)
477 {
478 	struct intel_bw_state *state;
479 
480 	state = kzalloc(sizeof(*state), GFP_KERNEL);
481 	if (!state)
482 		return -ENOMEM;
483 
484 	drm_atomic_private_obj_init(&dev_priv->drm, &dev_priv->bw_obj,
485 				    &state->base, &intel_bw_funcs);
486 
487 	return 0;
488 }
489 
490 void intel_bw_cleanup(struct drm_i915_private *dev_priv)
491 {
492 	drm_atomic_private_obj_fini(&dev_priv->bw_obj);
493 }
494