1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include <linux/string_helpers.h>
7 
8 #include "i915_drv.h"
9 #include "intel_engine_regs.h"
10 #include "intel_gt_regs.h"
11 #include "intel_sseu.h"
12 
intel_sseu_set_info(struct sseu_dev_info * sseu,u8 max_slices,u8 max_subslices,u8 max_eus_per_subslice)13 void intel_sseu_set_info(struct sseu_dev_info *sseu, u8 max_slices,
14 			 u8 max_subslices, u8 max_eus_per_subslice)
15 {
16 	sseu->max_slices = max_slices;
17 	sseu->max_subslices = max_subslices;
18 	sseu->max_eus_per_subslice = max_eus_per_subslice;
19 }
20 
21 unsigned int
intel_sseu_subslice_total(const struct sseu_dev_info * sseu)22 intel_sseu_subslice_total(const struct sseu_dev_info *sseu)
23 {
24 	unsigned int i, total = 0;
25 
26 	if (sseu->has_xehp_dss)
27 		return bitmap_weight(sseu->subslice_mask.xehp,
28 				     XEHP_BITMAP_BITS(sseu->subslice_mask));
29 
30 	for (i = 0; i < ARRAY_SIZE(sseu->subslice_mask.hsw); i++)
31 		total += hweight8(sseu->subslice_mask.hsw[i]);
32 
33 	return total;
34 }
35 
36 unsigned int
intel_sseu_get_hsw_subslices(const struct sseu_dev_info * sseu,u8 slice)37 intel_sseu_get_hsw_subslices(const struct sseu_dev_info *sseu, u8 slice)
38 {
39 	WARN_ON(sseu->has_xehp_dss);
40 	if (WARN_ON(slice >= sseu->max_slices))
41 		return 0;
42 
43 	return sseu->subslice_mask.hsw[slice];
44 }
45 
sseu_get_eus(const struct sseu_dev_info * sseu,int slice,int subslice)46 static u16 sseu_get_eus(const struct sseu_dev_info *sseu, int slice,
47 			int subslice)
48 {
49 	if (sseu->has_xehp_dss) {
50 		WARN_ON(slice > 0);
51 		return sseu->eu_mask.xehp[subslice];
52 	} else {
53 		return sseu->eu_mask.hsw[slice][subslice];
54 	}
55 }
56 
sseu_set_eus(struct sseu_dev_info * sseu,int slice,int subslice,u16 eu_mask)57 static void sseu_set_eus(struct sseu_dev_info *sseu, int slice, int subslice,
58 			 u16 eu_mask)
59 {
60 	GEM_WARN_ON(eu_mask && __fls(eu_mask) >= sseu->max_eus_per_subslice);
61 	if (sseu->has_xehp_dss) {
62 		GEM_WARN_ON(slice > 0);
63 		sseu->eu_mask.xehp[subslice] = eu_mask;
64 	} else {
65 		sseu->eu_mask.hsw[slice][subslice] = eu_mask;
66 	}
67 }
68 
compute_eu_total(const struct sseu_dev_info * sseu)69 static u16 compute_eu_total(const struct sseu_dev_info *sseu)
70 {
71 	int s, ss, total = 0;
72 
73 	for (s = 0; s < sseu->max_slices; s++)
74 		for (ss = 0; ss < sseu->max_subslices; ss++)
75 			if (sseu->has_xehp_dss)
76 				total += hweight16(sseu->eu_mask.xehp[ss]);
77 			else
78 				total += hweight16(sseu->eu_mask.hsw[s][ss]);
79 
80 	return total;
81 }
82 
83 /**
84  * intel_sseu_copy_eumask_to_user - Copy EU mask into a userspace buffer
85  * @to: Pointer to userspace buffer to copy to
86  * @sseu: SSEU structure containing EU mask to copy
87  *
88  * Copies the EU mask to a userspace buffer in the format expected by
89  * the query ioctl's topology queries.
90  *
91  * Returns the result of the copy_to_user() operation.
92  */
intel_sseu_copy_eumask_to_user(void __user * to,const struct sseu_dev_info * sseu)93 int intel_sseu_copy_eumask_to_user(void __user *to,
94 				   const struct sseu_dev_info *sseu)
95 {
96 	u8 eu_mask[GEN_SS_MASK_SIZE * GEN_MAX_EU_STRIDE] = {};
97 	int eu_stride = GEN_SSEU_STRIDE(sseu->max_eus_per_subslice);
98 	int len = sseu->max_slices * sseu->max_subslices * eu_stride;
99 	int s, ss, i;
100 
101 	for (s = 0; s < sseu->max_slices; s++) {
102 		for (ss = 0; ss < sseu->max_subslices; ss++) {
103 			int uapi_offset =
104 				s * sseu->max_subslices * eu_stride +
105 				ss * eu_stride;
106 			u16 mask = sseu_get_eus(sseu, s, ss);
107 
108 			for (i = 0; i < eu_stride; i++)
109 				eu_mask[uapi_offset + i] =
110 					(mask >> (BITS_PER_BYTE * i)) & 0xff;
111 		}
112 	}
113 
114 	return copy_to_user(to, eu_mask, len);
115 }
116 
117 /**
118  * intel_sseu_copy_ssmask_to_user - Copy subslice mask into a userspace buffer
119  * @to: Pointer to userspace buffer to copy to
120  * @sseu: SSEU structure containing subslice mask to copy
121  *
122  * Copies the subslice mask to a userspace buffer in the format expected by
123  * the query ioctl's topology queries.
124  *
125  * Returns the result of the copy_to_user() operation.
126  */
intel_sseu_copy_ssmask_to_user(void __user * to,const struct sseu_dev_info * sseu)127 int intel_sseu_copy_ssmask_to_user(void __user *to,
128 				   const struct sseu_dev_info *sseu)
129 {
130 	u8 ss_mask[GEN_SS_MASK_SIZE] = {};
131 	int ss_stride = GEN_SSEU_STRIDE(sseu->max_subslices);
132 	int len = sseu->max_slices * ss_stride;
133 	int s, ss, i;
134 
135 	for (s = 0; s < sseu->max_slices; s++) {
136 		for (ss = 0; ss < sseu->max_subslices; ss++) {
137 			i = s * ss_stride * BITS_PER_BYTE + ss;
138 
139 			if (!intel_sseu_has_subslice(sseu, s, ss))
140 				continue;
141 
142 			ss_mask[i / BITS_PER_BYTE] |= BIT(i % BITS_PER_BYTE);
143 		}
144 	}
145 
146 	return copy_to_user(to, ss_mask, len);
147 }
148 
gen11_compute_sseu_info(struct sseu_dev_info * sseu,u32 ss_en,u16 eu_en)149 static void gen11_compute_sseu_info(struct sseu_dev_info *sseu,
150 				    u32 ss_en, u16 eu_en)
151 {
152 	u32 valid_ss_mask = GENMASK(sseu->max_subslices - 1, 0);
153 	int ss;
154 
155 	sseu->slice_mask |= BIT(0);
156 	sseu->subslice_mask.hsw[0] = ss_en & valid_ss_mask;
157 
158 	for (ss = 0; ss < sseu->max_subslices; ss++)
159 		if (intel_sseu_has_subslice(sseu, 0, ss))
160 			sseu_set_eus(sseu, 0, ss, eu_en);
161 
162 	sseu->eu_per_subslice = hweight16(eu_en);
163 	sseu->eu_total = compute_eu_total(sseu);
164 }
165 
xehp_compute_sseu_info(struct sseu_dev_info * sseu,u16 eu_en)166 static void xehp_compute_sseu_info(struct sseu_dev_info *sseu,
167 				   u16 eu_en)
168 {
169 	int ss;
170 
171 	sseu->slice_mask |= BIT(0);
172 
173 	bitmap_or(sseu->subslice_mask.xehp,
174 		  sseu->compute_subslice_mask.xehp,
175 		  sseu->geometry_subslice_mask.xehp,
176 		  XEHP_BITMAP_BITS(sseu->subslice_mask));
177 
178 	for (ss = 0; ss < sseu->max_subslices; ss++)
179 		if (intel_sseu_has_subslice(sseu, 0, ss))
180 			sseu_set_eus(sseu, 0, ss, eu_en);
181 
182 	sseu->eu_per_subslice = hweight16(eu_en);
183 	sseu->eu_total = compute_eu_total(sseu);
184 }
185 
186 static void
xehp_load_dss_mask(struct intel_uncore * uncore,intel_sseu_ss_mask_t * ssmask,int numregs,...)187 xehp_load_dss_mask(struct intel_uncore *uncore,
188 		   intel_sseu_ss_mask_t *ssmask,
189 		   int numregs,
190 		   ...)
191 {
192 	va_list argp;
193 	u32 fuse_val[I915_MAX_SS_FUSE_REGS] = {};
194 	int i;
195 
196 	if (WARN_ON(numregs > I915_MAX_SS_FUSE_REGS))
197 		numregs = I915_MAX_SS_FUSE_REGS;
198 
199 	va_start(argp, numregs);
200 	for (i = 0; i < numregs; i++)
201 		fuse_val[i] = intel_uncore_read(uncore, va_arg(argp, i915_reg_t));
202 	va_end(argp);
203 
204 	bitmap_from_arr32(ssmask->xehp, fuse_val, numregs * 32);
205 }
206 
xehp_sseu_info_init(struct intel_gt * gt)207 static void xehp_sseu_info_init(struct intel_gt *gt)
208 {
209 	struct sseu_dev_info *sseu = &gt->info.sseu;
210 	struct intel_uncore *uncore = gt->uncore;
211 	u16 eu_en = 0;
212 	u8 eu_en_fuse;
213 	int num_compute_regs, num_geometry_regs;
214 	int eu;
215 
216 	if (IS_PONTEVECCHIO(gt->i915)) {
217 		num_geometry_regs = 0;
218 		num_compute_regs = 2;
219 	} else {
220 		num_geometry_regs = 1;
221 		num_compute_regs = 1;
222 	}
223 
224 	/*
225 	 * The concept of slice has been removed in Xe_HP.  To be compatible
226 	 * with prior generations, assume a single slice across the entire
227 	 * device. Then calculate out the DSS for each workload type within
228 	 * that software slice.
229 	 */
230 	intel_sseu_set_info(sseu, 1,
231 			    32 * max(num_geometry_regs, num_compute_regs),
232 			    HAS_ONE_EU_PER_FUSE_BIT(gt->i915) ? 8 : 16);
233 	sseu->has_xehp_dss = 1;
234 
235 	xehp_load_dss_mask(uncore, &sseu->geometry_subslice_mask,
236 			   num_geometry_regs,
237 			   GEN12_GT_GEOMETRY_DSS_ENABLE);
238 	xehp_load_dss_mask(uncore, &sseu->compute_subslice_mask,
239 			   num_compute_regs,
240 			   GEN12_GT_COMPUTE_DSS_ENABLE,
241 			   XEHPC_GT_COMPUTE_DSS_ENABLE_EXT);
242 
243 	eu_en_fuse = intel_uncore_read(uncore, XEHP_EU_ENABLE) & XEHP_EU_ENA_MASK;
244 
245 	if (HAS_ONE_EU_PER_FUSE_BIT(gt->i915))
246 		eu_en = eu_en_fuse;
247 	else
248 		for (eu = 0; eu < sseu->max_eus_per_subslice / 2; eu++)
249 			if (eu_en_fuse & BIT(eu))
250 				eu_en |= BIT(eu * 2) | BIT(eu * 2 + 1);
251 
252 	xehp_compute_sseu_info(sseu, eu_en);
253 }
254 
gen12_sseu_info_init(struct intel_gt * gt)255 static void gen12_sseu_info_init(struct intel_gt *gt)
256 {
257 	struct sseu_dev_info *sseu = &gt->info.sseu;
258 	struct intel_uncore *uncore = gt->uncore;
259 	u32 g_dss_en;
260 	u16 eu_en = 0;
261 	u8 eu_en_fuse;
262 	u8 s_en;
263 	int eu;
264 
265 	/*
266 	 * Gen12 has Dual-Subslices, which behave similarly to 2 gen11 SS.
267 	 * Instead of splitting these, provide userspace with an array
268 	 * of DSS to more closely represent the hardware resource.
269 	 */
270 	intel_sseu_set_info(sseu, 1, 6, 16);
271 
272 	/*
273 	 * Although gen12 architecture supported multiple slices, TGL, RKL,
274 	 * DG1, and ADL only had a single slice.
275 	 */
276 	s_en = intel_uncore_read(uncore, GEN11_GT_SLICE_ENABLE) &
277 		GEN11_GT_S_ENA_MASK;
278 	drm_WARN_ON(&gt->i915->drm, s_en != 0x1);
279 
280 	g_dss_en = intel_uncore_read(uncore, GEN12_GT_GEOMETRY_DSS_ENABLE);
281 
282 	/* one bit per pair of EUs */
283 	eu_en_fuse = ~(intel_uncore_read(uncore, GEN11_EU_DISABLE) &
284 		       GEN11_EU_DIS_MASK);
285 
286 	for (eu = 0; eu < sseu->max_eus_per_subslice / 2; eu++)
287 		if (eu_en_fuse & BIT(eu))
288 			eu_en |= BIT(eu * 2) | BIT(eu * 2 + 1);
289 
290 	gen11_compute_sseu_info(sseu, g_dss_en, eu_en);
291 
292 	/* TGL only supports slice-level power gating */
293 	sseu->has_slice_pg = 1;
294 }
295 
gen11_sseu_info_init(struct intel_gt * gt)296 static void gen11_sseu_info_init(struct intel_gt *gt)
297 {
298 	struct sseu_dev_info *sseu = &gt->info.sseu;
299 	struct intel_uncore *uncore = gt->uncore;
300 	u32 ss_en;
301 	u8 eu_en;
302 	u8 s_en;
303 
304 	if (IS_JSL_EHL(gt->i915))
305 		intel_sseu_set_info(sseu, 1, 4, 8);
306 	else
307 		intel_sseu_set_info(sseu, 1, 8, 8);
308 
309 	/*
310 	 * Although gen11 architecture supported multiple slices, ICL and
311 	 * EHL/JSL only had a single slice in practice.
312 	 */
313 	s_en = intel_uncore_read(uncore, GEN11_GT_SLICE_ENABLE) &
314 		GEN11_GT_S_ENA_MASK;
315 	drm_WARN_ON(&gt->i915->drm, s_en != 0x1);
316 
317 	ss_en = ~intel_uncore_read(uncore, GEN11_GT_SUBSLICE_DISABLE);
318 
319 	eu_en = ~(intel_uncore_read(uncore, GEN11_EU_DISABLE) &
320 		  GEN11_EU_DIS_MASK);
321 
322 	gen11_compute_sseu_info(sseu, ss_en, eu_en);
323 
324 	/* ICL has no power gating restrictions. */
325 	sseu->has_slice_pg = 1;
326 	sseu->has_subslice_pg = 1;
327 	sseu->has_eu_pg = 1;
328 }
329 
cherryview_sseu_info_init(struct intel_gt * gt)330 static void cherryview_sseu_info_init(struct intel_gt *gt)
331 {
332 	struct sseu_dev_info *sseu = &gt->info.sseu;
333 	u32 fuse;
334 
335 	fuse = intel_uncore_read(gt->uncore, CHV_FUSE_GT);
336 
337 	sseu->slice_mask = BIT(0);
338 	intel_sseu_set_info(sseu, 1, 2, 8);
339 
340 	if (!(fuse & CHV_FGT_DISABLE_SS0)) {
341 		u8 disabled_mask =
342 			((fuse & CHV_FGT_EU_DIS_SS0_R0_MASK) >>
343 			 CHV_FGT_EU_DIS_SS0_R0_SHIFT) |
344 			(((fuse & CHV_FGT_EU_DIS_SS0_R1_MASK) >>
345 			  CHV_FGT_EU_DIS_SS0_R1_SHIFT) << 4);
346 
347 		sseu->subslice_mask.hsw[0] |= BIT(0);
348 		sseu_set_eus(sseu, 0, 0, ~disabled_mask & 0xFF);
349 	}
350 
351 	if (!(fuse & CHV_FGT_DISABLE_SS1)) {
352 		u8 disabled_mask =
353 			((fuse & CHV_FGT_EU_DIS_SS1_R0_MASK) >>
354 			 CHV_FGT_EU_DIS_SS1_R0_SHIFT) |
355 			(((fuse & CHV_FGT_EU_DIS_SS1_R1_MASK) >>
356 			  CHV_FGT_EU_DIS_SS1_R1_SHIFT) << 4);
357 
358 		sseu->subslice_mask.hsw[0] |= BIT(1);
359 		sseu_set_eus(sseu, 0, 1, ~disabled_mask & 0xFF);
360 	}
361 
362 	sseu->eu_total = compute_eu_total(sseu);
363 
364 	/*
365 	 * CHV expected to always have a uniform distribution of EU
366 	 * across subslices.
367 	 */
368 	sseu->eu_per_subslice = intel_sseu_subslice_total(sseu) ?
369 		sseu->eu_total /
370 		intel_sseu_subslice_total(sseu) :
371 		0;
372 	/*
373 	 * CHV supports subslice power gating on devices with more than
374 	 * one subslice, and supports EU power gating on devices with
375 	 * more than one EU pair per subslice.
376 	 */
377 	sseu->has_slice_pg = 0;
378 	sseu->has_subslice_pg = intel_sseu_subslice_total(sseu) > 1;
379 	sseu->has_eu_pg = (sseu->eu_per_subslice > 2);
380 }
381 
gen9_sseu_info_init(struct intel_gt * gt)382 static void gen9_sseu_info_init(struct intel_gt *gt)
383 {
384 	struct drm_i915_private *i915 = gt->i915;
385 	struct sseu_dev_info *sseu = &gt->info.sseu;
386 	struct intel_uncore *uncore = gt->uncore;
387 	u32 fuse2, eu_disable, subslice_mask;
388 	const u8 eu_mask = 0xff;
389 	int s, ss;
390 
391 	fuse2 = intel_uncore_read(uncore, GEN8_FUSE2);
392 	sseu->slice_mask = (fuse2 & GEN8_F2_S_ENA_MASK) >> GEN8_F2_S_ENA_SHIFT;
393 
394 	/* BXT has a single slice and at most 3 subslices. */
395 	intel_sseu_set_info(sseu, IS_GEN9_LP(i915) ? 1 : 3,
396 			    IS_GEN9_LP(i915) ? 3 : 4, 8);
397 
398 	/*
399 	 * The subslice disable field is global, i.e. it applies
400 	 * to each of the enabled slices.
401 	 */
402 	subslice_mask = (1 << sseu->max_subslices) - 1;
403 	subslice_mask &= ~((fuse2 & GEN9_F2_SS_DIS_MASK) >>
404 			   GEN9_F2_SS_DIS_SHIFT);
405 
406 	/*
407 	 * Iterate through enabled slices and subslices to
408 	 * count the total enabled EU.
409 	 */
410 	for (s = 0; s < sseu->max_slices; s++) {
411 		if (!(sseu->slice_mask & BIT(s)))
412 			/* skip disabled slice */
413 			continue;
414 
415 		sseu->subslice_mask.hsw[s] = subslice_mask;
416 
417 		eu_disable = intel_uncore_read(uncore, GEN9_EU_DISABLE(s));
418 		for (ss = 0; ss < sseu->max_subslices; ss++) {
419 			int eu_per_ss;
420 			u8 eu_disabled_mask;
421 
422 			if (!intel_sseu_has_subslice(sseu, s, ss))
423 				/* skip disabled subslice */
424 				continue;
425 
426 			eu_disabled_mask = (eu_disable >> (ss * 8)) & eu_mask;
427 
428 			sseu_set_eus(sseu, s, ss, ~eu_disabled_mask & eu_mask);
429 
430 			eu_per_ss = sseu->max_eus_per_subslice -
431 				hweight8(eu_disabled_mask);
432 
433 			/*
434 			 * Record which subslice(s) has(have) 7 EUs. we
435 			 * can tune the hash used to spread work among
436 			 * subslices if they are unbalanced.
437 			 */
438 			if (eu_per_ss == 7)
439 				sseu->subslice_7eu[s] |= BIT(ss);
440 		}
441 	}
442 
443 	sseu->eu_total = compute_eu_total(sseu);
444 
445 	/*
446 	 * SKL is expected to always have a uniform distribution
447 	 * of EU across subslices with the exception that any one
448 	 * EU in any one subslice may be fused off for die
449 	 * recovery. BXT is expected to be perfectly uniform in EU
450 	 * distribution.
451 	 */
452 	sseu->eu_per_subslice =
453 		intel_sseu_subslice_total(sseu) ?
454 		DIV_ROUND_UP(sseu->eu_total, intel_sseu_subslice_total(sseu)) :
455 		0;
456 
457 	/*
458 	 * SKL+ supports slice power gating on devices with more than
459 	 * one slice, and supports EU power gating on devices with
460 	 * more than one EU pair per subslice. BXT+ supports subslice
461 	 * power gating on devices with more than one subslice, and
462 	 * supports EU power gating on devices with more than one EU
463 	 * pair per subslice.
464 	 */
465 	sseu->has_slice_pg =
466 		!IS_GEN9_LP(i915) && hweight8(sseu->slice_mask) > 1;
467 	sseu->has_subslice_pg =
468 		IS_GEN9_LP(i915) && intel_sseu_subslice_total(sseu) > 1;
469 	sseu->has_eu_pg = sseu->eu_per_subslice > 2;
470 
471 	if (IS_GEN9_LP(i915)) {
472 #define IS_SS_DISABLED(ss)	(!(sseu->subslice_mask.hsw[0] & BIT(ss)))
473 		RUNTIME_INFO(i915)->has_pooled_eu = hweight8(sseu->subslice_mask.hsw[0]) == 3;
474 
475 		sseu->min_eu_in_pool = 0;
476 		if (HAS_POOLED_EU(i915)) {
477 			if (IS_SS_DISABLED(2) || IS_SS_DISABLED(0))
478 				sseu->min_eu_in_pool = 3;
479 			else if (IS_SS_DISABLED(1))
480 				sseu->min_eu_in_pool = 6;
481 			else
482 				sseu->min_eu_in_pool = 9;
483 		}
484 #undef IS_SS_DISABLED
485 	}
486 }
487 
bdw_sseu_info_init(struct intel_gt * gt)488 static void bdw_sseu_info_init(struct intel_gt *gt)
489 {
490 	struct sseu_dev_info *sseu = &gt->info.sseu;
491 	struct intel_uncore *uncore = gt->uncore;
492 	int s, ss;
493 	u32 fuse2, subslice_mask, eu_disable[3]; /* s_max */
494 	u32 eu_disable0, eu_disable1, eu_disable2;
495 
496 	fuse2 = intel_uncore_read(uncore, GEN8_FUSE2);
497 	sseu->slice_mask = (fuse2 & GEN8_F2_S_ENA_MASK) >> GEN8_F2_S_ENA_SHIFT;
498 	intel_sseu_set_info(sseu, 3, 3, 8);
499 
500 	/*
501 	 * The subslice disable field is global, i.e. it applies
502 	 * to each of the enabled slices.
503 	 */
504 	subslice_mask = GENMASK(sseu->max_subslices - 1, 0);
505 	subslice_mask &= ~((fuse2 & GEN8_F2_SS_DIS_MASK) >>
506 			   GEN8_F2_SS_DIS_SHIFT);
507 	eu_disable0 = intel_uncore_read(uncore, GEN8_EU_DISABLE0);
508 	eu_disable1 = intel_uncore_read(uncore, GEN8_EU_DISABLE1);
509 	eu_disable2 = intel_uncore_read(uncore, GEN8_EU_DISABLE2);
510 	eu_disable[0] = eu_disable0 & GEN8_EU_DIS0_S0_MASK;
511 	eu_disable[1] = (eu_disable0 >> GEN8_EU_DIS0_S1_SHIFT) |
512 		((eu_disable1 & GEN8_EU_DIS1_S1_MASK) <<
513 		 (32 - GEN8_EU_DIS0_S1_SHIFT));
514 	eu_disable[2] = (eu_disable1 >> GEN8_EU_DIS1_S2_SHIFT) |
515 		((eu_disable2 & GEN8_EU_DIS2_S2_MASK) <<
516 		 (32 - GEN8_EU_DIS1_S2_SHIFT));
517 
518 	/*
519 	 * Iterate through enabled slices and subslices to
520 	 * count the total enabled EU.
521 	 */
522 	for (s = 0; s < sseu->max_slices; s++) {
523 		if (!(sseu->slice_mask & BIT(s)))
524 			/* skip disabled slice */
525 			continue;
526 
527 		sseu->subslice_mask.hsw[s] = subslice_mask;
528 
529 		for (ss = 0; ss < sseu->max_subslices; ss++) {
530 			u8 eu_disabled_mask;
531 			u32 n_disabled;
532 
533 			if (!intel_sseu_has_subslice(sseu, s, ss))
534 				/* skip disabled subslice */
535 				continue;
536 
537 			eu_disabled_mask =
538 				eu_disable[s] >> (ss * sseu->max_eus_per_subslice);
539 
540 			sseu_set_eus(sseu, s, ss, ~eu_disabled_mask & 0xFF);
541 
542 			n_disabled = hweight8(eu_disabled_mask);
543 
544 			/*
545 			 * Record which subslices have 7 EUs.
546 			 */
547 			if (sseu->max_eus_per_subslice - n_disabled == 7)
548 				sseu->subslice_7eu[s] |= 1 << ss;
549 		}
550 	}
551 
552 	sseu->eu_total = compute_eu_total(sseu);
553 
554 	/*
555 	 * BDW is expected to always have a uniform distribution of EU across
556 	 * subslices with the exception that any one EU in any one subslice may
557 	 * be fused off for die recovery.
558 	 */
559 	sseu->eu_per_subslice =
560 		intel_sseu_subslice_total(sseu) ?
561 		DIV_ROUND_UP(sseu->eu_total, intel_sseu_subslice_total(sseu)) :
562 		0;
563 
564 	/*
565 	 * BDW supports slice power gating on devices with more than
566 	 * one slice.
567 	 */
568 	sseu->has_slice_pg = hweight8(sseu->slice_mask) > 1;
569 	sseu->has_subslice_pg = 0;
570 	sseu->has_eu_pg = 0;
571 }
572 
hsw_sseu_info_init(struct intel_gt * gt)573 static void hsw_sseu_info_init(struct intel_gt *gt)
574 {
575 	struct drm_i915_private *i915 = gt->i915;
576 	struct sseu_dev_info *sseu = &gt->info.sseu;
577 	u32 fuse1;
578 	u8 subslice_mask = 0;
579 	int s, ss;
580 
581 	/*
582 	 * There isn't a register to tell us how many slices/subslices. We
583 	 * work off the PCI-ids here.
584 	 */
585 	switch (INTEL_INFO(i915)->gt) {
586 	default:
587 		MISSING_CASE(INTEL_INFO(i915)->gt);
588 		fallthrough;
589 	case 1:
590 		sseu->slice_mask = BIT(0);
591 		subslice_mask = BIT(0);
592 		break;
593 	case 2:
594 		sseu->slice_mask = BIT(0);
595 		subslice_mask = BIT(0) | BIT(1);
596 		break;
597 	case 3:
598 		sseu->slice_mask = BIT(0) | BIT(1);
599 		subslice_mask = BIT(0) | BIT(1);
600 		break;
601 	}
602 
603 	fuse1 = intel_uncore_read(gt->uncore, HSW_PAVP_FUSE1);
604 	switch (REG_FIELD_GET(HSW_F1_EU_DIS_MASK, fuse1)) {
605 	default:
606 		MISSING_CASE(REG_FIELD_GET(HSW_F1_EU_DIS_MASK, fuse1));
607 		fallthrough;
608 	case HSW_F1_EU_DIS_10EUS:
609 		sseu->eu_per_subslice = 10;
610 		break;
611 	case HSW_F1_EU_DIS_8EUS:
612 		sseu->eu_per_subslice = 8;
613 		break;
614 	case HSW_F1_EU_DIS_6EUS:
615 		sseu->eu_per_subslice = 6;
616 		break;
617 	}
618 
619 	intel_sseu_set_info(sseu, hweight8(sseu->slice_mask),
620 			    hweight8(subslice_mask),
621 			    sseu->eu_per_subslice);
622 
623 	for (s = 0; s < sseu->max_slices; s++) {
624 		sseu->subslice_mask.hsw[s] = subslice_mask;
625 
626 		for (ss = 0; ss < sseu->max_subslices; ss++) {
627 			sseu_set_eus(sseu, s, ss,
628 				     (1UL << sseu->eu_per_subslice) - 1);
629 		}
630 	}
631 
632 	sseu->eu_total = compute_eu_total(sseu);
633 
634 	/* No powergating for you. */
635 	sseu->has_slice_pg = 0;
636 	sseu->has_subslice_pg = 0;
637 	sseu->has_eu_pg = 0;
638 }
639 
intel_sseu_info_init(struct intel_gt * gt)640 void intel_sseu_info_init(struct intel_gt *gt)
641 {
642 	struct drm_i915_private *i915 = gt->i915;
643 
644 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50))
645 		xehp_sseu_info_init(gt);
646 	else if (GRAPHICS_VER(i915) >= 12)
647 		gen12_sseu_info_init(gt);
648 	else if (GRAPHICS_VER(i915) >= 11)
649 		gen11_sseu_info_init(gt);
650 	else if (GRAPHICS_VER(i915) >= 9)
651 		gen9_sseu_info_init(gt);
652 	else if (IS_BROADWELL(i915))
653 		bdw_sseu_info_init(gt);
654 	else if (IS_CHERRYVIEW(i915))
655 		cherryview_sseu_info_init(gt);
656 	else if (IS_HASWELL(i915))
657 		hsw_sseu_info_init(gt);
658 }
659 
intel_sseu_make_rpcs(struct intel_gt * gt,const struct intel_sseu * req_sseu)660 u32 intel_sseu_make_rpcs(struct intel_gt *gt,
661 			 const struct intel_sseu *req_sseu)
662 {
663 	struct drm_i915_private *i915 = gt->i915;
664 	const struct sseu_dev_info *sseu = &gt->info.sseu;
665 	bool subslice_pg = sseu->has_subslice_pg;
666 	u8 slices, subslices;
667 	u32 rpcs = 0;
668 
669 	/*
670 	 * No explicit RPCS request is needed to ensure full
671 	 * slice/subslice/EU enablement prior to Gen9.
672 	 */
673 	if (GRAPHICS_VER(i915) < 9)
674 		return 0;
675 
676 	/*
677 	 * If i915/perf is active, we want a stable powergating configuration
678 	 * on the system. Use the configuration pinned by i915/perf.
679 	 */
680 	if (gt->perf.exclusive_stream)
681 		req_sseu = &gt->perf.sseu;
682 
683 	slices = hweight8(req_sseu->slice_mask);
684 	subslices = hweight8(req_sseu->subslice_mask);
685 
686 	/*
687 	 * Since the SScount bitfield in GEN8_R_PWR_CLK_STATE is only three bits
688 	 * wide and Icelake has up to eight subslices, specfial programming is
689 	 * needed in order to correctly enable all subslices.
690 	 *
691 	 * According to documentation software must consider the configuration
692 	 * as 2x4x8 and hardware will translate this to 1x8x8.
693 	 *
694 	 * Furthemore, even though SScount is three bits, maximum documented
695 	 * value for it is four. From this some rules/restrictions follow:
696 	 *
697 	 * 1.
698 	 * If enabled subslice count is greater than four, two whole slices must
699 	 * be enabled instead.
700 	 *
701 	 * 2.
702 	 * When more than one slice is enabled, hardware ignores the subslice
703 	 * count altogether.
704 	 *
705 	 * From these restrictions it follows that it is not possible to enable
706 	 * a count of subslices between the SScount maximum of four restriction,
707 	 * and the maximum available number on a particular SKU. Either all
708 	 * subslices are enabled, or a count between one and four on the first
709 	 * slice.
710 	 */
711 	if (GRAPHICS_VER(i915) == 11 &&
712 	    slices == 1 &&
713 	    subslices > min_t(u8, 4, hweight8(sseu->subslice_mask.hsw[0]) / 2)) {
714 		GEM_BUG_ON(subslices & 1);
715 
716 		subslice_pg = false;
717 		slices *= 2;
718 	}
719 
720 	/*
721 	 * Starting in Gen9, render power gating can leave
722 	 * slice/subslice/EU in a partially enabled state. We
723 	 * must make an explicit request through RPCS for full
724 	 * enablement.
725 	 */
726 	if (sseu->has_slice_pg) {
727 		u32 mask, val = slices;
728 
729 		if (GRAPHICS_VER(i915) >= 11) {
730 			mask = GEN11_RPCS_S_CNT_MASK;
731 			val <<= GEN11_RPCS_S_CNT_SHIFT;
732 		} else {
733 			mask = GEN8_RPCS_S_CNT_MASK;
734 			val <<= GEN8_RPCS_S_CNT_SHIFT;
735 		}
736 
737 		GEM_BUG_ON(val & ~mask);
738 		val &= mask;
739 
740 		rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_S_CNT_ENABLE | val;
741 	}
742 
743 	if (subslice_pg) {
744 		u32 val = subslices;
745 
746 		val <<= GEN8_RPCS_SS_CNT_SHIFT;
747 
748 		GEM_BUG_ON(val & ~GEN8_RPCS_SS_CNT_MASK);
749 		val &= GEN8_RPCS_SS_CNT_MASK;
750 
751 		rpcs |= GEN8_RPCS_ENABLE | GEN8_RPCS_SS_CNT_ENABLE | val;
752 	}
753 
754 	if (sseu->has_eu_pg) {
755 		u32 val;
756 
757 		val = req_sseu->min_eus_per_subslice << GEN8_RPCS_EU_MIN_SHIFT;
758 		GEM_BUG_ON(val & ~GEN8_RPCS_EU_MIN_MASK);
759 		val &= GEN8_RPCS_EU_MIN_MASK;
760 
761 		rpcs |= val;
762 
763 		val = req_sseu->max_eus_per_subslice << GEN8_RPCS_EU_MAX_SHIFT;
764 		GEM_BUG_ON(val & ~GEN8_RPCS_EU_MAX_MASK);
765 		val &= GEN8_RPCS_EU_MAX_MASK;
766 
767 		rpcs |= val;
768 
769 		rpcs |= GEN8_RPCS_ENABLE;
770 	}
771 
772 	return rpcs;
773 }
774 
intel_sseu_dump(const struct sseu_dev_info * sseu,struct drm_printer * p)775 void intel_sseu_dump(const struct sseu_dev_info *sseu, struct drm_printer *p)
776 {
777 	int s;
778 
779 	if (sseu->has_xehp_dss) {
780 		drm_printf(p, "subslice total: %u\n",
781 			   intel_sseu_subslice_total(sseu));
782 		drm_printf(p, "geometry dss mask=%*pb\n",
783 			   XEHP_BITMAP_BITS(sseu->geometry_subslice_mask),
784 			   sseu->geometry_subslice_mask.xehp);
785 		drm_printf(p, "compute dss mask=%*pb\n",
786 			   XEHP_BITMAP_BITS(sseu->compute_subslice_mask),
787 			   sseu->compute_subslice_mask.xehp);
788 	} else {
789 		drm_printf(p, "slice total: %u, mask=%04x\n",
790 			   hweight8(sseu->slice_mask), sseu->slice_mask);
791 		drm_printf(p, "subslice total: %u\n",
792 			   intel_sseu_subslice_total(sseu));
793 
794 		for (s = 0; s < sseu->max_slices; s++) {
795 			u8 ss_mask = sseu->subslice_mask.hsw[s];
796 
797 			drm_printf(p, "slice%d: %u subslices, mask=%08x\n",
798 				   s, hweight8(ss_mask), ss_mask);
799 		}
800 	}
801 
802 	drm_printf(p, "EU total: %u\n", sseu->eu_total);
803 	drm_printf(p, "EU per subslice: %u\n", sseu->eu_per_subslice);
804 	drm_printf(p, "has slice power gating: %s\n",
805 		   str_yes_no(sseu->has_slice_pg));
806 	drm_printf(p, "has subslice power gating: %s\n",
807 		   str_yes_no(sseu->has_subslice_pg));
808 	drm_printf(p, "has EU power gating: %s\n",
809 		   str_yes_no(sseu->has_eu_pg));
810 }
811 
sseu_print_hsw_topology(const struct sseu_dev_info * sseu,struct drm_printer * p)812 static void sseu_print_hsw_topology(const struct sseu_dev_info *sseu,
813 				    struct drm_printer *p)
814 {
815 	int s, ss;
816 
817 	for (s = 0; s < sseu->max_slices; s++) {
818 		u8 ss_mask = sseu->subslice_mask.hsw[s];
819 
820 		drm_printf(p, "slice%d: %u subslice(s) (0x%08x):\n",
821 			   s, hweight8(ss_mask), ss_mask);
822 
823 		for (ss = 0; ss < sseu->max_subslices; ss++) {
824 			u16 enabled_eus = sseu_get_eus(sseu, s, ss);
825 
826 			drm_printf(p, "\tsubslice%d: %u EUs (0x%hx)\n",
827 				   ss, hweight16(enabled_eus), enabled_eus);
828 		}
829 	}
830 }
831 
sseu_print_xehp_topology(const struct sseu_dev_info * sseu,struct drm_printer * p)832 static void sseu_print_xehp_topology(const struct sseu_dev_info *sseu,
833 				     struct drm_printer *p)
834 {
835 	int dss;
836 
837 	for (dss = 0; dss < sseu->max_subslices; dss++) {
838 		u16 enabled_eus = sseu_get_eus(sseu, 0, dss);
839 
840 		drm_printf(p, "DSS_%02d: G:%3s C:%3s, %2u EUs (0x%04hx)\n", dss,
841 			   str_yes_no(test_bit(dss, sseu->geometry_subslice_mask.xehp)),
842 			   str_yes_no(test_bit(dss, sseu->compute_subslice_mask.xehp)),
843 			   hweight16(enabled_eus), enabled_eus);
844 	}
845 }
846 
intel_sseu_print_topology(struct drm_i915_private * i915,const struct sseu_dev_info * sseu,struct drm_printer * p)847 void intel_sseu_print_topology(struct drm_i915_private *i915,
848 			       const struct sseu_dev_info *sseu,
849 			       struct drm_printer *p)
850 {
851 	if (sseu->max_slices == 0) {
852 		drm_printf(p, "Unavailable\n");
853 	} else if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
854 		sseu_print_xehp_topology(sseu, p);
855 	} else {
856 		sseu_print_hsw_topology(sseu, p);
857 	}
858 }
859 
intel_sseu_print_ss_info(const char * type,const struct sseu_dev_info * sseu,struct seq_file * m)860 void intel_sseu_print_ss_info(const char *type,
861 			      const struct sseu_dev_info *sseu,
862 			      struct seq_file *m)
863 {
864 	int s;
865 
866 	if (sseu->has_xehp_dss) {
867 		seq_printf(m, "  %s Geometry DSS: %u\n", type,
868 			   bitmap_weight(sseu->geometry_subslice_mask.xehp,
869 					 XEHP_BITMAP_BITS(sseu->geometry_subslice_mask)));
870 		seq_printf(m, "  %s Compute DSS: %u\n", type,
871 			   bitmap_weight(sseu->compute_subslice_mask.xehp,
872 					 XEHP_BITMAP_BITS(sseu->compute_subslice_mask)));
873 	} else {
874 		for (s = 0; s < fls(sseu->slice_mask); s++)
875 			seq_printf(m, "  %s Slice%i subslices: %u\n", type,
876 				   s, hweight8(sseu->subslice_mask.hsw[s]));
877 	}
878 }
879 
intel_slicemask_from_xehp_dssmask(intel_sseu_ss_mask_t dss_mask,int dss_per_slice)880 u16 intel_slicemask_from_xehp_dssmask(intel_sseu_ss_mask_t dss_mask,
881 				      int dss_per_slice)
882 {
883 	intel_sseu_ss_mask_t per_slice_mask = {};
884 	unsigned long slice_mask = 0;
885 	int i;
886 
887 	WARN_ON(DIV_ROUND_UP(XEHP_BITMAP_BITS(dss_mask), dss_per_slice) >
888 		8 * sizeof(slice_mask));
889 
890 	bitmap_fill(per_slice_mask.xehp, dss_per_slice);
891 	for (i = 0; !bitmap_empty(dss_mask.xehp, XEHP_BITMAP_BITS(dss_mask)); i++) {
892 		if (bitmap_intersects(dss_mask.xehp, per_slice_mask.xehp, dss_per_slice))
893 			slice_mask |= BIT(i);
894 
895 		bitmap_shift_right(dss_mask.xehp, dss_mask.xehp, dss_per_slice,
896 				   XEHP_BITMAP_BITS(dss_mask));
897 	}
898 
899 	return slice_mask;
900 }
901