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 = >->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 = >->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(>->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 = >->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(>->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 = >->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 = >->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 = >->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 = >->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 = >->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 = >->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