1 // SPDX-License-Identifier: MIT
3 * Copyright © 2015 Intel Corporation
8 #include "intel_engine.h"
10 #include "intel_gt_mcr.h"
11 #include "intel_gt_regs.h"
12 #include "intel_mocs.h"
13 #include "intel_ring.h"
15 /* structures required */
16 struct drm_i915_mocs_entry {
22 struct drm_i915_mocs_table {
24 unsigned int n_entries;
25 const struct drm_i915_mocs_entry *table;
27 u8 wb_index; /* Only used on HAS_L3_CCS_READ() platforms */
28 u8 unused_entries_index;
31 /* Defines for the tables (XXX_MOCS_0 - XXX_MOCS_63) */
32 #define _LE_CACHEABILITY(value) ((value) << 0)
33 #define _LE_TGT_CACHE(value) ((value) << 2)
34 #define LE_LRUM(value) ((value) << 4)
35 #define LE_AOM(value) ((value) << 6)
36 #define LE_RSC(value) ((value) << 7)
37 #define LE_SCC(value) ((value) << 8)
38 #define LE_PFM(value) ((value) << 11)
39 #define LE_SCF(value) ((value) << 14)
40 #define LE_COS(value) ((value) << 15)
41 #define LE_SSE(value) ((value) << 17)
43 /* Defines for the tables (GLOB_MOCS_0 - GLOB_MOCS_16) */
44 #define _L4_CACHEABILITY(value) ((value) << 2)
45 #define IG_PAT(value) ((value) << 8)
47 /* Defines for the tables (LNCFMOCS0 - LNCFMOCS31) - two entries per word */
48 #define L3_ESC(value) ((value) << 0)
49 #define L3_SCC(value) ((value) << 1)
50 #define _L3_CACHEABILITY(value) ((value) << 4)
51 #define L3_GLBGO(value) ((value) << 6)
52 #define L3_LKUP(value) ((value) << 7)
55 #define GEN9_NUM_MOCS_ENTRIES 64 /* 63-64 are reserved, but configured. */
56 #define MTL_NUM_MOCS_ENTRIES 16
58 /* (e)LLC caching options */
60 * Note: LE_0_PAGETABLE works only up to Gen11; for newer gens it means
63 #define LE_0_PAGETABLE _LE_CACHEABILITY(0)
64 #define LE_1_UC _LE_CACHEABILITY(1)
65 #define LE_2_WT _LE_CACHEABILITY(2)
66 #define LE_3_WB _LE_CACHEABILITY(3)
69 #define LE_TC_0_PAGETABLE _LE_TGT_CACHE(0)
70 #define LE_TC_1_LLC _LE_TGT_CACHE(1)
71 #define LE_TC_2_LLC_ELLC _LE_TGT_CACHE(2)
72 #define LE_TC_3_LLC_ELLC_ALT _LE_TGT_CACHE(3)
74 /* L3 caching options */
75 #define L3_0_DIRECT _L3_CACHEABILITY(0)
76 #define L3_1_UC _L3_CACHEABILITY(1)
77 #define L3_2_RESERVED _L3_CACHEABILITY(2)
78 #define L3_3_WB _L3_CACHEABILITY(3)
80 /* L4 caching options */
81 #define L4_0_WB _L4_CACHEABILITY(0)
82 #define L4_1_WT _L4_CACHEABILITY(1)
83 #define L4_2_RESERVED _L4_CACHEABILITY(2)
84 #define L4_3_UC _L4_CACHEABILITY(3)
86 #define MOCS_ENTRY(__idx, __control_value, __l3cc_value) \
88 .control_value = __control_value, \
89 .l3cc_value = __l3cc_value, \
96 * These are the MOCS tables that are programmed across all the rings.
97 * The control value is programmed to all the rings that support the
98 * MOCS registers. While the l3cc_values are only programmed to the
99 * LNCFCMOCS0 - LNCFCMOCS32 registers.
101 * These tables are intended to be kept reasonably consistent across
102 * HW platforms, and for ICL+, be identical across OSes. To achieve
103 * that, for Icelake and above, list of entries is published as part
106 * Entries not part of the following tables are undefined as far as
107 * userspace is concerned and shouldn't be relied upon. For Gen < 12
108 * they will be initialized to PTE. Gen >= 12 don't have a setting for
109 * PTE and those platforms except TGL/RKL will be initialized L3 WB to
110 * catch accidental use of reserved and unused mocs indexes.
112 * The last few entries are reserved by the hardware. For ICL+ they
113 * should be initialized according to bspec and never used, for older
114 * platforms they should never be written to.
116 * NOTE1: These tables are part of bspec and defined as part of hardware
117 * interface for ICL+. For older platforms, they are part of kernel
118 * ABI. It is expected that, for specific hardware platform, existing
119 * entries will remain constant and the table will only be updated by
120 * adding new entries, filling unused positions.
122 * NOTE2: For GEN >= 12 except TGL and RKL, reserved and unspecified MOCS
123 * indices have been set to L3 WB. These reserved entries should never
124 * be used, they may be changed to low performant variants with better
125 * coherency in the future if more entries are needed.
126 * For TGL/RKL, all the unspecified MOCS indexes are mapped to L3 UC.
128 #define GEN9_MOCS_ENTRIES \
129 MOCS_ENTRY(I915_MOCS_UNCACHED, \
130 LE_1_UC | LE_TC_2_LLC_ELLC, \
132 MOCS_ENTRY(I915_MOCS_PTE, \
133 LE_0_PAGETABLE | LE_TC_0_PAGETABLE | LE_LRUM(3), \
136 static const struct drm_i915_mocs_entry skl_mocs_table[] = {
138 MOCS_ENTRY(I915_MOCS_CACHED,
139 LE_3_WB | LE_TC_2_LLC_ELLC | LE_LRUM(3),
144 * - used by the L3 for all of its evictions.
145 * Thus it is expected to allow LLC cacheability to enable coherent
146 * flows to be maintained.
147 * - used to force L3 uncachable cycles.
148 * Thus it is expected to make the surface L3 uncacheable.
151 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
155 /* NOTE: the LE_TGT_CACHE is not used on Broxton */
156 static const struct drm_i915_mocs_entry broxton_mocs_table[] = {
158 MOCS_ENTRY(I915_MOCS_CACHED,
159 LE_1_UC | LE_TC_2_LLC_ELLC | LE_LRUM(3),
163 #define GEN11_MOCS_ENTRIES \
164 /* Entries 0 and 1 are defined per-platform */ \
165 /* Base - L3 + LLC */ \
167 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
169 /* Base - Uncached */ \
171 LE_1_UC | LE_TC_1_LLC, \
175 LE_1_UC | LE_TC_1_LLC, \
179 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
183 LE_3_WB | LE_TC_1_LLC | LE_LRUM(1), \
185 /* Age 0 - L3 + LLC */ \
187 LE_3_WB | LE_TC_1_LLC | LE_LRUM(1), \
189 /* Age: Don't Chg. - LLC */ \
191 LE_3_WB | LE_TC_1_LLC | LE_LRUM(2), \
193 /* Age: Don't Chg. - L3 + LLC */ \
195 LE_3_WB | LE_TC_1_LLC | LE_LRUM(2), \
199 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_AOM(1), \
201 /* No AOM - L3 + LLC */ \
203 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_AOM(1), \
205 /* No AOM; Age 0 - LLC */ \
207 LE_3_WB | LE_TC_1_LLC | LE_LRUM(1) | LE_AOM(1), \
209 /* No AOM; Age 0 - L3 + LLC */ \
211 LE_3_WB | LE_TC_1_LLC | LE_LRUM(1) | LE_AOM(1), \
213 /* No AOM; Age:DC - LLC */ \
215 LE_3_WB | LE_TC_1_LLC | LE_LRUM(2) | LE_AOM(1), \
217 /* No AOM; Age:DC - L3 + LLC */ \
219 LE_3_WB | LE_TC_1_LLC | LE_LRUM(2) | LE_AOM(1), \
221 /* Bypass LLC - Uncached (EHL+) */ \
223 LE_1_UC | LE_TC_1_LLC | LE_SCF(1), \
225 /* Bypass LLC - L3 (Read-Only) (EHL+) */ \
227 LE_1_UC | LE_TC_1_LLC | LE_SCF(1), \
229 /* Self-Snoop - L3 + LLC */ \
231 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_SSE(3), \
233 /* Skip Caching - L3 + LLC(12.5%) */ \
235 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_SCC(7), \
237 /* Skip Caching - L3 + LLC(25%) */ \
239 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_SCC(3), \
241 /* Skip Caching - L3 + LLC(50%) */ \
243 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_SCC(1), \
245 /* Skip Caching - L3 + LLC(75%) */ \
247 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_RSC(1) | LE_SCC(3), \
249 /* Skip Caching - L3 + LLC(87.5%) */ \
251 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3) | LE_RSC(1) | LE_SCC(7), \
253 /* HW Reserved - SW program but never use */ \
255 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
257 /* HW Reserved - SW program but never use */ \
259 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3), \
262 static const struct drm_i915_mocs_entry tgl_mocs_table[] = {
265 * Reserved and unspecified MOCS indices have been set to (L3 + LCC).
266 * These reserved entries should never be used, they may be changed
267 * to low performant variants with better coherency in the future if
268 * more entries are needed. We are programming index I915_MOCS_PTE(1)
269 * only, __init_mocs_table() take care to program unused index with
272 MOCS_ENTRY(I915_MOCS_PTE,
273 LE_0_PAGETABLE | LE_TC_0_PAGETABLE,
277 /* Implicitly enable L1 - HDC:L1 + L3 + LLC */
279 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
281 /* Implicitly enable L1 - HDC:L1 + L3 */
283 LE_1_UC | LE_TC_1_LLC,
285 /* Implicitly enable L1 - HDC:L1 + LLC */
287 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
289 /* Implicitly enable L1 - HDC:L1 */
291 LE_1_UC | LE_TC_1_LLC,
293 /* HW Special Case (CCS) */
295 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
297 /* HW Special Case (Displayable) */
299 LE_1_UC | LE_TC_1_LLC,
303 static const struct drm_i915_mocs_entry icl_mocs_table[] = {
304 /* Base - Uncached (Deprecated) */
305 MOCS_ENTRY(I915_MOCS_UNCACHED,
306 LE_1_UC | LE_TC_1_LLC,
308 /* Base - L3 + LeCC:PAT (Deprecated) */
309 MOCS_ENTRY(I915_MOCS_PTE,
310 LE_0_PAGETABLE | LE_TC_0_PAGETABLE,
316 static const struct drm_i915_mocs_entry dg1_mocs_table[] = {
319 MOCS_ENTRY(1, 0, L3_1_UC),
321 MOCS_ENTRY(5, 0, L3_3_WB),
323 MOCS_ENTRY(6, 0, L3_ESC(1) | L3_SCC(1) | L3_3_WB),
325 MOCS_ENTRY(7, 0, L3_ESC(1) | L3_SCC(3) | L3_3_WB),
327 MOCS_ENTRY(8, 0, L3_ESC(1) | L3_SCC(7) | L3_3_WB),
330 MOCS_ENTRY(48, 0, L3_3_WB),
332 MOCS_ENTRY(49, 0, L3_1_UC),
335 MOCS_ENTRY(60, 0, L3_1_UC),
336 MOCS_ENTRY(61, 0, L3_1_UC),
337 MOCS_ENTRY(62, 0, L3_1_UC),
338 MOCS_ENTRY(63, 0, L3_1_UC),
341 static const struct drm_i915_mocs_entry gen12_mocs_table[] = {
343 /* Implicitly enable L1 - HDC:L1 + L3 + LLC */
345 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
347 /* Implicitly enable L1 - HDC:L1 + L3 */
349 LE_1_UC | LE_TC_1_LLC,
351 /* Implicitly enable L1 - HDC:L1 + LLC */
353 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
355 /* Implicitly enable L1 - HDC:L1 */
357 LE_1_UC | LE_TC_1_LLC,
359 /* HW Special Case (CCS) */
361 LE_3_WB | LE_TC_1_LLC | LE_LRUM(3),
363 /* HW Special Case (Displayable) */
365 LE_1_UC | LE_TC_1_LLC,
369 static const struct drm_i915_mocs_entry dg2_mocs_table[] = {
370 /* UC - Coherent; GO:L3 */
371 MOCS_ENTRY(0, 0, L3_1_UC | L3_LKUP(1)),
372 /* UC - Coherent; GO:Memory */
373 MOCS_ENTRY(1, 0, L3_1_UC | L3_GLBGO(1) | L3_LKUP(1)),
374 /* UC - Non-Coherent; GO:Memory */
375 MOCS_ENTRY(2, 0, L3_1_UC | L3_GLBGO(1)),
378 MOCS_ENTRY(3, 0, L3_3_WB | L3_LKUP(1)),
381 static const struct drm_i915_mocs_entry mtl_mocs_table[] = {
382 /* Error - Reserved for Non-Use */
385 L3_LKUP(1) | L3_3_WB),
386 /* Cached - L3 + L4 */
389 L3_LKUP(1) | L3_3_WB),
393 L3_LKUP(1) | L3_1_UC),
394 /* Uncached - GO:L3 */
397 L3_LKUP(1) | L3_1_UC),
401 L3_LKUP(1) | L3_GLBGO(1) | L3_1_UC),
402 /* Uncached - GO:Mem */
405 L3_LKUP(1) | L3_GLBGO(1) | L3_1_UC),
406 /* L4 - L3:NoLKUP; GO:L3 */
410 /* Uncached - L3:NoLKUP; GO:L3 */
414 /* L4 - L3:NoLKUP; GO:Mem */
417 L3_GLBGO(1) | L3_1_UC),
418 /* Uncached - L3:NoLKUP; GO:Mem */
421 L3_GLBGO(1) | L3_1_UC),
422 /* Display - L3; L4:WT */
425 L3_LKUP(1) | L3_3_WB),
426 /* CCS - Non-Displayable */
429 L3_GLBGO(1) | L3_1_UC),
433 HAS_GLOBAL_MOCS = BIT(0),
434 HAS_ENGINE_MOCS = BIT(1),
435 HAS_RENDER_L3CC = BIT(2),
438 static bool has_l3cc(const struct drm_i915_private *i915)
443 static bool has_global_mocs(const struct drm_i915_private *i915)
445 return HAS_GLOBAL_MOCS_REGISTERS(i915);
448 static bool has_mocs(const struct drm_i915_private *i915)
450 return !IS_DGFX(i915);
453 static unsigned int get_mocs_settings(struct drm_i915_private *i915,
454 struct drm_i915_mocs_table *table)
458 memset(table, 0, sizeof(struct drm_i915_mocs_table));
460 table->unused_entries_index = I915_MOCS_PTE;
461 if (IS_GFX_GT_IP_RANGE(to_gt(i915), IP_VER(12, 70), IP_VER(12, 74))) {
462 table->size = ARRAY_SIZE(mtl_mocs_table);
463 table->table = mtl_mocs_table;
464 table->n_entries = MTL_NUM_MOCS_ENTRIES;
466 table->unused_entries_index = 1;
467 } else if (IS_DG2(i915)) {
468 table->size = ARRAY_SIZE(dg2_mocs_table);
469 table->table = dg2_mocs_table;
471 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
472 table->unused_entries_index = 3;
473 } else if (IS_DG1(i915)) {
474 table->size = ARRAY_SIZE(dg1_mocs_table);
475 table->table = dg1_mocs_table;
477 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
479 table->unused_entries_index = 5;
480 } else if (IS_TIGERLAKE(i915) || IS_ROCKETLAKE(i915)) {
481 /* For TGL/RKL, Can't be changed now for ABI reasons */
482 table->size = ARRAY_SIZE(tgl_mocs_table);
483 table->table = tgl_mocs_table;
484 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
486 } else if (GRAPHICS_VER(i915) >= 12) {
487 table->size = ARRAY_SIZE(gen12_mocs_table);
488 table->table = gen12_mocs_table;
489 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
491 table->unused_entries_index = 2;
492 } else if (GRAPHICS_VER(i915) == 11) {
493 table->size = ARRAY_SIZE(icl_mocs_table);
494 table->table = icl_mocs_table;
495 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
496 } else if (IS_GEN9_BC(i915)) {
497 table->size = ARRAY_SIZE(skl_mocs_table);
498 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
499 table->table = skl_mocs_table;
500 } else if (IS_GEN9_LP(i915)) {
501 table->size = ARRAY_SIZE(broxton_mocs_table);
502 table->n_entries = GEN9_NUM_MOCS_ENTRIES;
503 table->table = broxton_mocs_table;
505 drm_WARN_ONCE(&i915->drm, GRAPHICS_VER(i915) >= 9,
506 "Platform that should have a MOCS table does not.\n");
510 if (GEM_DEBUG_WARN_ON(table->size > table->n_entries))
513 /* WaDisableSkipCaching:skl,bxt,kbl,glk */
514 if (GRAPHICS_VER(i915) == 9) {
517 for (i = 0; i < table->size; i++)
518 if (GEM_DEBUG_WARN_ON(table->table[i].l3cc_value &
519 (L3_ESC(1) | L3_SCC(0x7))))
524 if (has_mocs(i915)) {
525 if (has_global_mocs(i915))
526 flags |= HAS_GLOBAL_MOCS;
528 flags |= HAS_ENGINE_MOCS;
531 flags |= HAS_RENDER_L3CC;
537 * Get control_value from MOCS entry taking into account when it's not used
538 * then if unused_entries_index is non-zero then its value will be returned
539 * otherwise I915_MOCS_PTE's value is returned in this case.
541 static u32 get_entry_control(const struct drm_i915_mocs_table *table,
544 if (index < table->size && table->table[index].used)
545 return table->table[index].control_value;
546 return table->table[table->unused_entries_index].control_value;
549 #define for_each_mocs(mocs, t, i) \
551 i < (t)->n_entries ? (mocs = get_entry_control((t), i)), 1 : 0;\
554 static void __init_mocs_table(struct intel_uncore *uncore,
555 const struct drm_i915_mocs_table *table,
561 drm_WARN_ONCE(&uncore->i915->drm, !table->unused_entries_index,
562 "Unused entries index should have been defined\n");
563 for_each_mocs(mocs, table, i)
564 intel_uncore_write_fw(uncore, _MMIO(addr + i * 4), mocs);
567 static u32 mocs_offset(const struct intel_engine_cs *engine)
569 static const u32 offset[] = {
570 [RCS0] = __GEN9_RCS0_MOCS0,
571 [VCS0] = __GEN9_VCS0_MOCS0,
572 [VCS1] = __GEN9_VCS1_MOCS0,
573 [VECS0] = __GEN9_VECS0_MOCS0,
574 [BCS0] = __GEN9_BCS0_MOCS0,
575 [VCS2] = __GEN11_VCS2_MOCS0,
578 GEM_BUG_ON(engine->id >= ARRAY_SIZE(offset));
579 return offset[engine->id];
582 static void init_mocs_table(struct intel_engine_cs *engine,
583 const struct drm_i915_mocs_table *table)
585 __init_mocs_table(engine->uncore, table, mocs_offset(engine));
589 * Get l3cc_value from MOCS entry taking into account when it's not used
590 * then if unused_entries_index is not zero then its value will be returned
591 * otherwise I915_MOCS_PTE's value is returned in this case.
593 static u16 get_entry_l3cc(const struct drm_i915_mocs_table *table,
596 if (index < table->size && table->table[index].used)
597 return table->table[index].l3cc_value;
598 return table->table[table->unused_entries_index].l3cc_value;
601 static u32 l3cc_combine(u16 low, u16 high)
603 return low | (u32)high << 16;
606 #define for_each_l3cc(l3cc, t, i) \
608 i < ((t)->n_entries + 1) / 2 ? \
609 (l3cc = l3cc_combine(get_entry_l3cc((t), 2 * i), \
610 get_entry_l3cc((t), 2 * i + 1))), 1 : \
614 static void init_l3cc_table(struct intel_gt *gt,
615 const struct drm_i915_mocs_table *table)
621 intel_gt_mcr_lock(gt, &flags);
622 for_each_l3cc(l3cc, table, i)
623 if (GRAPHICS_VER_FULL(gt->i915) >= IP_VER(12, 55))
624 intel_gt_mcr_multicast_write_fw(gt, XEHP_LNCFCMOCS(i), l3cc);
626 intel_uncore_write_fw(gt->uncore, GEN9_LNCFCMOCS(i), l3cc);
627 intel_gt_mcr_unlock(gt, flags);
630 void intel_mocs_init_engine(struct intel_engine_cs *engine)
632 struct drm_i915_mocs_table table;
635 /* Called under a blanket forcewake */
636 assert_forcewakes_active(engine->uncore, FORCEWAKE_ALL);
638 flags = get_mocs_settings(engine->i915, &table);
642 /* Platforms with global MOCS do not need per-engine initialization. */
643 if (flags & HAS_ENGINE_MOCS)
644 init_mocs_table(engine, &table);
646 if (flags & HAS_RENDER_L3CC && engine->class == RENDER_CLASS)
647 init_l3cc_table(engine->gt, &table);
650 static u32 global_mocs_offset(void)
652 return i915_mmio_reg_offset(GEN12_GLOBAL_MOCS(0));
655 void intel_set_mocs_index(struct intel_gt *gt)
657 struct drm_i915_mocs_table table;
659 get_mocs_settings(gt->i915, &table);
660 gt->mocs.uc_index = table.uc_index;
661 if (HAS_L3_CCS_READ(gt->i915))
662 gt->mocs.wb_index = table.wb_index;
665 void intel_mocs_init(struct intel_gt *gt)
667 struct drm_i915_mocs_table table;
671 * LLC and eDRAM control values are not applicable to dgfx
673 flags = get_mocs_settings(gt->i915, &table);
674 if (flags & HAS_GLOBAL_MOCS)
675 __init_mocs_table(gt->uncore, &table, global_mocs_offset());
678 * Initialize the L3CC table as part of mocs initalization to make
679 * sure the LNCFCMOCSx registers are programmed for the subsequent
680 * memory transactions including guc transactions
682 if (flags & HAS_RENDER_L3CC)
683 init_l3cc_table(gt, &table);
686 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
687 #include "selftest_mocs.c"