1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014-2018 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "i915_reg.h"
8 #include "intel_context.h"
9 #include "intel_engine_pm.h"
10 #include "intel_engine_regs.h"
11 #include "intel_gpu_commands.h"
12 #include "intel_gt.h"
13 #include "intel_gt_mcr.h"
14 #include "intel_gt_regs.h"
15 #include "intel_ring.h"
16 #include "intel_workarounds.h"
17 
18 /**
19  * DOC: Hardware workarounds
20  *
21  * This file is intended as a central place to implement most [1]_ of the
22  * required workarounds for hardware to work as originally intended. They fall
23  * in five basic categories depending on how/when they are applied:
24  *
25  * - Workarounds that touch registers that are saved/restored to/from the HW
26  *   context image. The list is emitted (via Load Register Immediate commands)
27  *   everytime a new context is created.
28  * - GT workarounds. The list of these WAs is applied whenever these registers
29  *   revert to default values (on GPU reset, suspend/resume [2]_, etc..).
30  * - Display workarounds. The list is applied during display clock-gating
31  *   initialization.
32  * - Workarounds that whitelist a privileged register, so that UMDs can manage
33  *   them directly. This is just a special case of a MMMIO workaround (as we
34  *   write the list of these to/be-whitelisted registers to some special HW
35  *   registers).
36  * - Workaround batchbuffers, that get executed automatically by the hardware
37  *   on every HW context restore.
38  *
39  * .. [1] Please notice that there are other WAs that, due to their nature,
40  *    cannot be applied from a central place. Those are peppered around the rest
41  *    of the code, as needed.
42  *
43  * .. [2] Technically, some registers are powercontext saved & restored, so they
44  *    survive a suspend/resume. In practice, writing them again is not too
45  *    costly and simplifies things. We can revisit this in the future.
46  *
47  * Layout
48  * ~~~~~~
49  *
50  * Keep things in this file ordered by WA type, as per the above (context, GT,
51  * display, register whitelist, batchbuffer). Then, inside each type, keep the
52  * following order:
53  *
54  * - Infrastructure functions and macros
55  * - WAs per platform in standard gen/chrono order
56  * - Public functions to init or apply the given workaround type.
57  */
58 
59 static void wa_init_start(struct i915_wa_list *wal, const char *name, const char *engine_name)
60 {
61 	wal->name = name;
62 	wal->engine_name = engine_name;
63 }
64 
65 #define WA_LIST_CHUNK (1 << 4)
66 
67 static void wa_init_finish(struct i915_wa_list *wal)
68 {
69 	/* Trim unused entries. */
70 	if (!IS_ALIGNED(wal->count, WA_LIST_CHUNK)) {
71 		struct i915_wa *list = kmemdup(wal->list,
72 					       wal->count * sizeof(*list),
73 					       GFP_KERNEL);
74 
75 		if (list) {
76 			kfree(wal->list);
77 			wal->list = list;
78 		}
79 	}
80 
81 	if (!wal->count)
82 		return;
83 
84 	DRM_DEBUG_DRIVER("Initialized %u %s workarounds on %s\n",
85 			 wal->wa_count, wal->name, wal->engine_name);
86 }
87 
88 static void _wa_add(struct i915_wa_list *wal, const struct i915_wa *wa)
89 {
90 	unsigned int addr = i915_mmio_reg_offset(wa->reg);
91 	unsigned int start = 0, end = wal->count;
92 	const unsigned int grow = WA_LIST_CHUNK;
93 	struct i915_wa *wa_;
94 
95 	GEM_BUG_ON(!is_power_of_2(grow));
96 
97 	if (IS_ALIGNED(wal->count, grow)) { /* Either uninitialized or full. */
98 		struct i915_wa *list;
99 
100 		list = kmalloc_array(ALIGN(wal->count + 1, grow), sizeof(*wa),
101 				     GFP_KERNEL);
102 		if (!list) {
103 			DRM_ERROR("No space for workaround init!\n");
104 			return;
105 		}
106 
107 		if (wal->list) {
108 			memcpy(list, wal->list, sizeof(*wa) * wal->count);
109 			kfree(wal->list);
110 		}
111 
112 		wal->list = list;
113 	}
114 
115 	while (start < end) {
116 		unsigned int mid = start + (end - start) / 2;
117 
118 		if (i915_mmio_reg_offset(wal->list[mid].reg) < addr) {
119 			start = mid + 1;
120 		} else if (i915_mmio_reg_offset(wal->list[mid].reg) > addr) {
121 			end = mid;
122 		} else {
123 			wa_ = &wal->list[mid];
124 
125 			if ((wa->clr | wa_->clr) && !(wa->clr & ~wa_->clr)) {
126 				DRM_ERROR("Discarding overwritten w/a for reg %04x (clear: %08x, set: %08x)\n",
127 					  i915_mmio_reg_offset(wa_->reg),
128 					  wa_->clr, wa_->set);
129 
130 				wa_->set &= ~wa->clr;
131 			}
132 
133 			wal->wa_count++;
134 			wa_->set |= wa->set;
135 			wa_->clr |= wa->clr;
136 			wa_->read |= wa->read;
137 			return;
138 		}
139 	}
140 
141 	wal->wa_count++;
142 	wa_ = &wal->list[wal->count++];
143 	*wa_ = *wa;
144 
145 	while (wa_-- > wal->list) {
146 		GEM_BUG_ON(i915_mmio_reg_offset(wa_[0].reg) ==
147 			   i915_mmio_reg_offset(wa_[1].reg));
148 		if (i915_mmio_reg_offset(wa_[1].reg) >
149 		    i915_mmio_reg_offset(wa_[0].reg))
150 			break;
151 
152 		swap(wa_[1], wa_[0]);
153 	}
154 }
155 
156 static void wa_add(struct i915_wa_list *wal, i915_reg_t reg,
157 		   u32 clear, u32 set, u32 read_mask, bool masked_reg)
158 {
159 	struct i915_wa wa = {
160 		.reg  = reg,
161 		.clr  = clear,
162 		.set  = set,
163 		.read = read_mask,
164 		.masked_reg = masked_reg,
165 	};
166 
167 	_wa_add(wal, &wa);
168 }
169 
170 static void
171 wa_write_clr_set(struct i915_wa_list *wal, i915_reg_t reg, u32 clear, u32 set)
172 {
173 	wa_add(wal, reg, clear, set, clear, false);
174 }
175 
176 static void
177 wa_write(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
178 {
179 	wa_write_clr_set(wal, reg, ~0, set);
180 }
181 
182 static void
183 wa_write_or(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
184 {
185 	wa_write_clr_set(wal, reg, set, set);
186 }
187 
188 static void
189 wa_write_clr(struct i915_wa_list *wal, i915_reg_t reg, u32 clr)
190 {
191 	wa_write_clr_set(wal, reg, clr, 0);
192 }
193 
194 /*
195  * WA operations on "masked register". A masked register has the upper 16 bits
196  * documented as "masked" in b-spec. Its purpose is to allow writing to just a
197  * portion of the register without a rmw: you simply write in the upper 16 bits
198  * the mask of bits you are going to modify.
199  *
200  * The wa_masked_* family of functions already does the necessary operations to
201  * calculate the mask based on the parameters passed, so user only has to
202  * provide the lower 16 bits of that register.
203  */
204 
205 static void
206 wa_masked_en(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
207 {
208 	wa_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
209 }
210 
211 static void
212 wa_masked_dis(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
213 {
214 	wa_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
215 }
216 
217 static void
218 wa_masked_field_set(struct i915_wa_list *wal, i915_reg_t reg,
219 		    u32 mask, u32 val)
220 {
221 	wa_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
222 }
223 
224 static void gen6_ctx_workarounds_init(struct intel_engine_cs *engine,
225 				      struct i915_wa_list *wal)
226 {
227 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
228 }
229 
230 static void gen7_ctx_workarounds_init(struct intel_engine_cs *engine,
231 				      struct i915_wa_list *wal)
232 {
233 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
234 }
235 
236 static void gen8_ctx_workarounds_init(struct intel_engine_cs *engine,
237 				      struct i915_wa_list *wal)
238 {
239 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
240 
241 	/* WaDisableAsyncFlipPerfMode:bdw,chv */
242 	wa_masked_en(wal, RING_MI_MODE(RENDER_RING_BASE), ASYNC_FLIP_PERF_DISABLE);
243 
244 	/* WaDisablePartialInstShootdown:bdw,chv */
245 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
246 		     PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
247 
248 	/* Use Force Non-Coherent whenever executing a 3D context. This is a
249 	 * workaround for a possible hang in the unlikely event a TLB
250 	 * invalidation occurs during a PSD flush.
251 	 */
252 	/* WaForceEnableNonCoherent:bdw,chv */
253 	/* WaHdcDisableFetchWhenMasked:bdw,chv */
254 	wa_masked_en(wal, HDC_CHICKEN0,
255 		     HDC_DONOT_FETCH_MEM_WHEN_MASKED |
256 		     HDC_FORCE_NON_COHERENT);
257 
258 	/* From the Haswell PRM, Command Reference: Registers, CACHE_MODE_0:
259 	 * "The Hierarchical Z RAW Stall Optimization allows non-overlapping
260 	 *  polygons in the same 8x4 pixel/sample area to be processed without
261 	 *  stalling waiting for the earlier ones to write to Hierarchical Z
262 	 *  buffer."
263 	 *
264 	 * This optimization is off by default for BDW and CHV; turn it on.
265 	 */
266 	wa_masked_dis(wal, CACHE_MODE_0_GEN7, HIZ_RAW_STALL_OPT_DISABLE);
267 
268 	/* Wa4x4STCOptimizationDisable:bdw,chv */
269 	wa_masked_en(wal, CACHE_MODE_1, GEN8_4x4_STC_OPTIMIZATION_DISABLE);
270 
271 	/*
272 	 * BSpec recommends 8x4 when MSAA is used,
273 	 * however in practice 16x4 seems fastest.
274 	 *
275 	 * Note that PS/WM thread counts depend on the WIZ hashing
276 	 * disable bit, which we don't touch here, but it's good
277 	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
278 	 */
279 	wa_masked_field_set(wal, GEN7_GT_MODE,
280 			    GEN6_WIZ_HASHING_MASK,
281 			    GEN6_WIZ_HASHING_16x4);
282 }
283 
284 static void bdw_ctx_workarounds_init(struct intel_engine_cs *engine,
285 				     struct i915_wa_list *wal)
286 {
287 	struct drm_i915_private *i915 = engine->i915;
288 
289 	gen8_ctx_workarounds_init(engine, wal);
290 
291 	/* WaDisableThreadStallDopClockGating:bdw (pre-production) */
292 	wa_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
293 
294 	/* WaDisableDopClockGating:bdw
295 	 *
296 	 * Also see the related UCGTCL1 write in bdw_init_clock_gating()
297 	 * to disable EUTC clock gating.
298 	 */
299 	wa_masked_en(wal, GEN7_ROW_CHICKEN2,
300 		     DOP_CLOCK_GATING_DISABLE);
301 
302 	wa_masked_en(wal, HALF_SLICE_CHICKEN3,
303 		     GEN8_SAMPLER_POWER_BYPASS_DIS);
304 
305 	wa_masked_en(wal, HDC_CHICKEN0,
306 		     /* WaForceContextSaveRestoreNonCoherent:bdw */
307 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
308 		     /* WaDisableFenceDestinationToSLM:bdw (pre-prod) */
309 		     (IS_BDW_GT3(i915) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
310 }
311 
312 static void chv_ctx_workarounds_init(struct intel_engine_cs *engine,
313 				     struct i915_wa_list *wal)
314 {
315 	gen8_ctx_workarounds_init(engine, wal);
316 
317 	/* WaDisableThreadStallDopClockGating:chv */
318 	wa_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
319 
320 	/* Improve HiZ throughput on CHV. */
321 	wa_masked_en(wal, HIZ_CHICKEN, CHV_HZ_8X8_MODE_IN_1X);
322 }
323 
324 static void gen9_ctx_workarounds_init(struct intel_engine_cs *engine,
325 				      struct i915_wa_list *wal)
326 {
327 	struct drm_i915_private *i915 = engine->i915;
328 
329 	if (HAS_LLC(i915)) {
330 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
331 		 *
332 		 * Must match Display Engine. See
333 		 * WaCompressedResourceDisplayNewHashMode.
334 		 */
335 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
336 			     GEN9_PBE_COMPRESSED_HASH_SELECTION);
337 		wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
338 			     GEN9_SAMPLER_HASH_COMPRESSED_READ_ADDR);
339 	}
340 
341 	/* WaClearFlowControlGpgpuContextSave:skl,bxt,kbl,glk,cfl */
342 	/* WaDisablePartialInstShootdown:skl,bxt,kbl,glk,cfl */
343 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
344 		     FLOW_CONTROL_ENABLE |
345 		     PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
346 
347 	/* WaEnableYV12BugFixInHalfSliceChicken7:skl,bxt,kbl,glk,cfl */
348 	/* WaEnableSamplerGPGPUPreemptionSupport:skl,bxt,kbl,cfl */
349 	wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
350 		     GEN9_ENABLE_YV12_BUGFIX |
351 		     GEN9_ENABLE_GPGPU_PREEMPTION);
352 
353 	/* Wa4x4STCOptimizationDisable:skl,bxt,kbl,glk,cfl */
354 	/* WaDisablePartialResolveInVc:skl,bxt,kbl,cfl */
355 	wa_masked_en(wal, CACHE_MODE_1,
356 		     GEN8_4x4_STC_OPTIMIZATION_DISABLE |
357 		     GEN9_PARTIAL_RESOLVE_IN_VC_DISABLE);
358 
359 	/* WaCcsTlbPrefetchDisable:skl,bxt,kbl,glk,cfl */
360 	wa_masked_dis(wal, GEN9_HALF_SLICE_CHICKEN5,
361 		      GEN9_CCS_TLB_PREFETCH_ENABLE);
362 
363 	/* WaForceContextSaveRestoreNonCoherent:skl,bxt,kbl,cfl */
364 	wa_masked_en(wal, HDC_CHICKEN0,
365 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
366 		     HDC_FORCE_CSR_NON_COHERENT_OVR_DISABLE);
367 
368 	/* WaForceEnableNonCoherent and WaDisableHDCInvalidation are
369 	 * both tied to WaForceContextSaveRestoreNonCoherent
370 	 * in some hsds for skl. We keep the tie for all gen9. The
371 	 * documentation is a bit hazy and so we want to get common behaviour,
372 	 * even though there is no clear evidence we would need both on kbl/bxt.
373 	 * This area has been source of system hangs so we play it safe
374 	 * and mimic the skl regardless of what bspec says.
375 	 *
376 	 * Use Force Non-Coherent whenever executing a 3D context. This
377 	 * is a workaround for a possible hang in the unlikely event
378 	 * a TLB invalidation occurs during a PSD flush.
379 	 */
380 
381 	/* WaForceEnableNonCoherent:skl,bxt,kbl,cfl */
382 	wa_masked_en(wal, HDC_CHICKEN0,
383 		     HDC_FORCE_NON_COHERENT);
384 
385 	/* WaDisableSamplerPowerBypassForSOPingPong:skl,bxt,kbl,cfl */
386 	if (IS_SKYLAKE(i915) ||
387 	    IS_KABYLAKE(i915) ||
388 	    IS_COFFEELAKE(i915) ||
389 	    IS_COMETLAKE(i915))
390 		wa_masked_en(wal, HALF_SLICE_CHICKEN3,
391 			     GEN8_SAMPLER_POWER_BYPASS_DIS);
392 
393 	/* WaDisableSTUnitPowerOptimization:skl,bxt,kbl,glk,cfl */
394 	wa_masked_en(wal, HALF_SLICE_CHICKEN2, GEN8_ST_PO_DISABLE);
395 
396 	/*
397 	 * Supporting preemption with fine-granularity requires changes in the
398 	 * batch buffer programming. Since we can't break old userspace, we
399 	 * need to set our default preemption level to safe value. Userspace is
400 	 * still able to use more fine-grained preemption levels, since in
401 	 * WaEnablePreemptionGranularityControlByUMD we're whitelisting the
402 	 * per-ctx register. As such, WaDisable{3D,GPGPU}MidCmdPreemption are
403 	 * not real HW workarounds, but merely a way to start using preemption
404 	 * while maintaining old contract with userspace.
405 	 */
406 
407 	/* WaDisable3DMidCmdPreemption:skl,bxt,glk,cfl,[cnl] */
408 	wa_masked_dis(wal, GEN8_CS_CHICKEN1, GEN9_PREEMPT_3D_OBJECT_LEVEL);
409 
410 	/* WaDisableGPGPUMidCmdPreemption:skl,bxt,blk,cfl,[cnl] */
411 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
412 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
413 			    GEN9_PREEMPT_GPGPU_COMMAND_LEVEL);
414 
415 	/* WaClearHIZ_WM_CHICKEN3:bxt,glk */
416 	if (IS_GEN9_LP(i915))
417 		wa_masked_en(wal, GEN9_WM_CHICKEN3, GEN9_FACTOR_IN_CLR_VAL_HIZ);
418 }
419 
420 static void skl_tune_iz_hashing(struct intel_engine_cs *engine,
421 				struct i915_wa_list *wal)
422 {
423 	struct intel_gt *gt = engine->gt;
424 	u8 vals[3] = { 0, 0, 0 };
425 	unsigned int i;
426 
427 	for (i = 0; i < 3; i++) {
428 		u8 ss;
429 
430 		/*
431 		 * Only consider slices where one, and only one, subslice has 7
432 		 * EUs
433 		 */
434 		if (!is_power_of_2(gt->info.sseu.subslice_7eu[i]))
435 			continue;
436 
437 		/*
438 		 * subslice_7eu[i] != 0 (because of the check above) and
439 		 * ss_max == 4 (maximum number of subslices possible per slice)
440 		 *
441 		 * ->    0 <= ss <= 3;
442 		 */
443 		ss = ffs(gt->info.sseu.subslice_7eu[i]) - 1;
444 		vals[i] = 3 - ss;
445 	}
446 
447 	if (vals[0] == 0 && vals[1] == 0 && vals[2] == 0)
448 		return;
449 
450 	/* Tune IZ hashing. See intel_device_info_runtime_init() */
451 	wa_masked_field_set(wal, GEN7_GT_MODE,
452 			    GEN9_IZ_HASHING_MASK(2) |
453 			    GEN9_IZ_HASHING_MASK(1) |
454 			    GEN9_IZ_HASHING_MASK(0),
455 			    GEN9_IZ_HASHING(2, vals[2]) |
456 			    GEN9_IZ_HASHING(1, vals[1]) |
457 			    GEN9_IZ_HASHING(0, vals[0]));
458 }
459 
460 static void skl_ctx_workarounds_init(struct intel_engine_cs *engine,
461 				     struct i915_wa_list *wal)
462 {
463 	gen9_ctx_workarounds_init(engine, wal);
464 	skl_tune_iz_hashing(engine, wal);
465 }
466 
467 static void bxt_ctx_workarounds_init(struct intel_engine_cs *engine,
468 				     struct i915_wa_list *wal)
469 {
470 	gen9_ctx_workarounds_init(engine, wal);
471 
472 	/* WaDisableThreadStallDopClockGating:bxt */
473 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
474 		     STALL_DOP_GATING_DISABLE);
475 
476 	/* WaToEnableHwFixForPushConstHWBug:bxt */
477 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
478 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
479 }
480 
481 static void kbl_ctx_workarounds_init(struct intel_engine_cs *engine,
482 				     struct i915_wa_list *wal)
483 {
484 	struct drm_i915_private *i915 = engine->i915;
485 
486 	gen9_ctx_workarounds_init(engine, wal);
487 
488 	/* WaToEnableHwFixForPushConstHWBug:kbl */
489 	if (IS_KBL_GRAPHICS_STEP(i915, STEP_C0, STEP_FOREVER))
490 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
491 			     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
492 
493 	/* WaDisableSbeCacheDispatchPortSharing:kbl */
494 	wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
495 		     GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
496 }
497 
498 static void glk_ctx_workarounds_init(struct intel_engine_cs *engine,
499 				     struct i915_wa_list *wal)
500 {
501 	gen9_ctx_workarounds_init(engine, wal);
502 
503 	/* WaToEnableHwFixForPushConstHWBug:glk */
504 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
505 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
506 }
507 
508 static void cfl_ctx_workarounds_init(struct intel_engine_cs *engine,
509 				     struct i915_wa_list *wal)
510 {
511 	gen9_ctx_workarounds_init(engine, wal);
512 
513 	/* WaToEnableHwFixForPushConstHWBug:cfl */
514 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
515 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
516 
517 	/* WaDisableSbeCacheDispatchPortSharing:cfl */
518 	wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
519 		     GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
520 }
521 
522 static void icl_ctx_workarounds_init(struct intel_engine_cs *engine,
523 				     struct i915_wa_list *wal)
524 {
525 	/* Wa_1406697149 (WaDisableBankHangMode:icl) */
526 	wa_write(wal,
527 		 GEN8_L3CNTLREG,
528 		 intel_uncore_read(engine->uncore, GEN8_L3CNTLREG) |
529 		 GEN8_ERRDETBCTRL);
530 
531 	/* WaForceEnableNonCoherent:icl
532 	 * This is not the same workaround as in early Gen9 platforms, where
533 	 * lacking this could cause system hangs, but coherency performance
534 	 * overhead is high and only a few compute workloads really need it
535 	 * (the register is whitelisted in hardware now, so UMDs can opt in
536 	 * for coherency if they have a good reason).
537 	 */
538 	wa_masked_en(wal, ICL_HDC_MODE, HDC_FORCE_NON_COHERENT);
539 
540 	/* WaEnableFloatBlendOptimization:icl */
541 	wa_add(wal, GEN10_CACHE_MODE_SS, 0,
542 	       _MASKED_BIT_ENABLE(FLOAT_BLEND_OPTIMIZATION_ENABLE),
543 	       0 /* write-only, so skip validation */,
544 	       true);
545 
546 	/* WaDisableGPGPUMidThreadPreemption:icl */
547 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
548 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
549 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
550 
551 	/* allow headerless messages for preemptible GPGPU context */
552 	wa_masked_en(wal, GEN10_SAMPLER_MODE,
553 		     GEN11_SAMPLER_ENABLE_HEADLESS_MSG);
554 
555 	/* Wa_1604278689:icl,ehl */
556 	wa_write(wal, IVB_FBC_RT_BASE, 0xFFFFFFFF & ~ILK_FBC_RT_VALID);
557 	wa_write_clr_set(wal, IVB_FBC_RT_BASE_UPPER,
558 			 0, /* write-only register; skip validation */
559 			 0xFFFFFFFF);
560 
561 	/* Wa_1406306137:icl,ehl */
562 	wa_masked_en(wal, GEN9_ROW_CHICKEN4, GEN11_DIS_PICK_2ND_EU);
563 }
564 
565 /*
566  * These settings aren't actually workarounds, but general tuning settings that
567  * need to be programmed on dg2 platform.
568  */
569 static void dg2_ctx_gt_tuning_init(struct intel_engine_cs *engine,
570 				   struct i915_wa_list *wal)
571 {
572 	wa_masked_en(wal, CHICKEN_RASTER_2, TBIMR_FAST_CLIP);
573 	wa_write_clr_set(wal, GEN11_L3SQCREG5, L3_PWM_TIMER_INIT_VAL_MASK,
574 			 REG_FIELD_PREP(L3_PWM_TIMER_INIT_VAL_MASK, 0x7f));
575 	wa_add(wal,
576 	       FF_MODE2,
577 	       FF_MODE2_TDS_TIMER_MASK,
578 	       FF_MODE2_TDS_TIMER_128,
579 	       0, false);
580 }
581 
582 /*
583  * These settings aren't actually workarounds, but general tuning settings that
584  * need to be programmed on several platforms.
585  */
586 static void gen12_ctx_gt_tuning_init(struct intel_engine_cs *engine,
587 				     struct i915_wa_list *wal)
588 {
589 	/*
590 	 * Although some platforms refer to it as Wa_1604555607, we need to
591 	 * program it even on those that don't explicitly list that
592 	 * workaround.
593 	 *
594 	 * Note that the programming of this register is further modified
595 	 * according to the FF_MODE2 guidance given by Wa_1608008084:gen12.
596 	 * Wa_1608008084 tells us the FF_MODE2 register will return the wrong
597 	 * value when read. The default value for this register is zero for all
598 	 * fields and there are no bit masks. So instead of doing a RMW we
599 	 * should just write TDS timer value. For the same reason read
600 	 * verification is ignored.
601 	 */
602 	wa_add(wal,
603 	       FF_MODE2,
604 	       FF_MODE2_TDS_TIMER_MASK,
605 	       FF_MODE2_TDS_TIMER_128,
606 	       0, false);
607 }
608 
609 static void gen12_ctx_workarounds_init(struct intel_engine_cs *engine,
610 				       struct i915_wa_list *wal)
611 {
612 	gen12_ctx_gt_tuning_init(engine, wal);
613 
614 	/*
615 	 * Wa_1409142259:tgl,dg1,adl-p
616 	 * Wa_1409347922:tgl,dg1,adl-p
617 	 * Wa_1409252684:tgl,dg1,adl-p
618 	 * Wa_1409217633:tgl,dg1,adl-p
619 	 * Wa_1409207793:tgl,dg1,adl-p
620 	 * Wa_1409178076:tgl,dg1,adl-p
621 	 * Wa_1408979724:tgl,dg1,adl-p
622 	 * Wa_14010443199:tgl,rkl,dg1,adl-p
623 	 * Wa_14010698770:tgl,rkl,dg1,adl-s,adl-p
624 	 * Wa_1409342910:tgl,rkl,dg1,adl-s,adl-p
625 	 */
626 	wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
627 		     GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
628 
629 	/* WaDisableGPGPUMidThreadPreemption:gen12 */
630 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
631 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
632 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
633 
634 	/*
635 	 * Wa_16011163337
636 	 *
637 	 * Like in gen12_ctx_gt_tuning_init(), read verification is ignored due
638 	 * to Wa_1608008084.
639 	 */
640 	wa_add(wal,
641 	       FF_MODE2,
642 	       FF_MODE2_GS_TIMER_MASK,
643 	       FF_MODE2_GS_TIMER_224,
644 	       0, false);
645 }
646 
647 static void dg1_ctx_workarounds_init(struct intel_engine_cs *engine,
648 				     struct i915_wa_list *wal)
649 {
650 	gen12_ctx_workarounds_init(engine, wal);
651 
652 	/* Wa_1409044764 */
653 	wa_masked_dis(wal, GEN11_COMMON_SLICE_CHICKEN3,
654 		      DG1_FLOAT_POINT_BLEND_OPT_STRICT_MODE_EN);
655 
656 	/* Wa_22010493298 */
657 	wa_masked_en(wal, HIZ_CHICKEN,
658 		     DG1_HZ_READ_SUPPRESSION_OPTIMIZATION_DISABLE);
659 }
660 
661 static void dg2_ctx_workarounds_init(struct intel_engine_cs *engine,
662 				     struct i915_wa_list *wal)
663 {
664 	dg2_ctx_gt_tuning_init(engine, wal);
665 
666 	/* Wa_16011186671:dg2_g11 */
667 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0)) {
668 		wa_masked_dis(wal, VFLSKPD, DIS_MULT_MISS_RD_SQUASH);
669 		wa_masked_en(wal, VFLSKPD, DIS_OVER_FETCH_CACHE);
670 	}
671 
672 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0)) {
673 		/* Wa_14010469329:dg2_g10 */
674 		wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
675 			     XEHP_DUAL_SIMD8_SEQ_MERGE_DISABLE);
676 
677 		/*
678 		 * Wa_22010465075:dg2_g10
679 		 * Wa_22010613112:dg2_g10
680 		 * Wa_14010698770:dg2_g10
681 		 */
682 		wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
683 			     GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
684 	}
685 
686 	/* Wa_16013271637:dg2 */
687 	wa_masked_en(wal, SLICE_COMMON_ECO_CHICKEN1,
688 		     MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
689 
690 	/* Wa_14014947963:dg2 */
691 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_B0, STEP_FOREVER) ||
692 		IS_DG2_G11(engine->i915) || IS_DG2_G12(engine->i915))
693 		wa_masked_field_set(wal, VF_PREEMPTION, PREEMPTION_VERTEX_COUNT, 0x4000);
694 
695 	/* Wa_15010599737:dg2 */
696 	wa_masked_en(wal, CHICKEN_RASTER_1, DIS_SF_ROUND_NEAREST_EVEN);
697 }
698 
699 static void fakewa_disable_nestedbb_mode(struct intel_engine_cs *engine,
700 					 struct i915_wa_list *wal)
701 {
702 	/*
703 	 * This is a "fake" workaround defined by software to ensure we
704 	 * maintain reliable, backward-compatible behavior for userspace with
705 	 * regards to how nested MI_BATCH_BUFFER_START commands are handled.
706 	 *
707 	 * The per-context setting of MI_MODE[12] determines whether the bits
708 	 * of a nested MI_BATCH_BUFFER_START instruction should be interpreted
709 	 * in the traditional manner or whether they should instead use a new
710 	 * tgl+ meaning that breaks backward compatibility, but allows nesting
711 	 * into 3rd-level batchbuffers.  When this new capability was first
712 	 * added in TGL, it remained off by default unless a context
713 	 * intentionally opted in to the new behavior.  However Xe_HPG now
714 	 * flips this on by default and requires that we explicitly opt out if
715 	 * we don't want the new behavior.
716 	 *
717 	 * From a SW perspective, we want to maintain the backward-compatible
718 	 * behavior for userspace, so we'll apply a fake workaround to set it
719 	 * back to the legacy behavior on platforms where the hardware default
720 	 * is to break compatibility.  At the moment there is no Linux
721 	 * userspace that utilizes third-level batchbuffers, so this will avoid
722 	 * userspace from needing to make any changes.  using the legacy
723 	 * meaning is the correct thing to do.  If/when we have userspace
724 	 * consumers that want to utilize third-level batch nesting, we can
725 	 * provide a context parameter to allow them to opt-in.
726 	 */
727 	wa_masked_dis(wal, RING_MI_MODE(engine->mmio_base), TGL_NESTED_BB_EN);
728 }
729 
730 static void gen12_ctx_gt_mocs_init(struct intel_engine_cs *engine,
731 				   struct i915_wa_list *wal)
732 {
733 	u8 mocs;
734 
735 	/*
736 	 * Some blitter commands do not have a field for MOCS, those
737 	 * commands will use MOCS index pointed by BLIT_CCTL.
738 	 * BLIT_CCTL registers are needed to be programmed to un-cached.
739 	 */
740 	if (engine->class == COPY_ENGINE_CLASS) {
741 		mocs = engine->gt->mocs.uc_index;
742 		wa_write_clr_set(wal,
743 				 BLIT_CCTL(engine->mmio_base),
744 				 BLIT_CCTL_MASK,
745 				 BLIT_CCTL_MOCS(mocs, mocs));
746 	}
747 }
748 
749 /*
750  * gen12_ctx_gt_fake_wa_init() aren't programmingan official workaround
751  * defined by the hardware team, but it programming general context registers.
752  * Adding those context register programming in context workaround
753  * allow us to use the wa framework for proper application and validation.
754  */
755 static void
756 gen12_ctx_gt_fake_wa_init(struct intel_engine_cs *engine,
757 			  struct i915_wa_list *wal)
758 {
759 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
760 		fakewa_disable_nestedbb_mode(engine, wal);
761 
762 	gen12_ctx_gt_mocs_init(engine, wal);
763 }
764 
765 static void
766 __intel_engine_init_ctx_wa(struct intel_engine_cs *engine,
767 			   struct i915_wa_list *wal,
768 			   const char *name)
769 {
770 	struct drm_i915_private *i915 = engine->i915;
771 
772 	wa_init_start(wal, name, engine->name);
773 
774 	/* Applies to all engines */
775 	/*
776 	 * Fake workarounds are not the actual workaround but
777 	 * programming of context registers using workaround framework.
778 	 */
779 	if (GRAPHICS_VER(i915) >= 12)
780 		gen12_ctx_gt_fake_wa_init(engine, wal);
781 
782 	if (engine->class != RENDER_CLASS)
783 		goto done;
784 
785 	if (IS_PONTEVECCHIO(i915))
786 		; /* noop; none at this time */
787 	else if (IS_DG2(i915))
788 		dg2_ctx_workarounds_init(engine, wal);
789 	else if (IS_XEHPSDV(i915))
790 		; /* noop; none at this time */
791 	else if (IS_DG1(i915))
792 		dg1_ctx_workarounds_init(engine, wal);
793 	else if (GRAPHICS_VER(i915) == 12)
794 		gen12_ctx_workarounds_init(engine, wal);
795 	else if (GRAPHICS_VER(i915) == 11)
796 		icl_ctx_workarounds_init(engine, wal);
797 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
798 		cfl_ctx_workarounds_init(engine, wal);
799 	else if (IS_GEMINILAKE(i915))
800 		glk_ctx_workarounds_init(engine, wal);
801 	else if (IS_KABYLAKE(i915))
802 		kbl_ctx_workarounds_init(engine, wal);
803 	else if (IS_BROXTON(i915))
804 		bxt_ctx_workarounds_init(engine, wal);
805 	else if (IS_SKYLAKE(i915))
806 		skl_ctx_workarounds_init(engine, wal);
807 	else if (IS_CHERRYVIEW(i915))
808 		chv_ctx_workarounds_init(engine, wal);
809 	else if (IS_BROADWELL(i915))
810 		bdw_ctx_workarounds_init(engine, wal);
811 	else if (GRAPHICS_VER(i915) == 7)
812 		gen7_ctx_workarounds_init(engine, wal);
813 	else if (GRAPHICS_VER(i915) == 6)
814 		gen6_ctx_workarounds_init(engine, wal);
815 	else if (GRAPHICS_VER(i915) < 8)
816 		;
817 	else
818 		MISSING_CASE(GRAPHICS_VER(i915));
819 
820 done:
821 	wa_init_finish(wal);
822 }
823 
824 void intel_engine_init_ctx_wa(struct intel_engine_cs *engine)
825 {
826 	__intel_engine_init_ctx_wa(engine, &engine->ctx_wa_list, "context");
827 }
828 
829 int intel_engine_emit_ctx_wa(struct i915_request *rq)
830 {
831 	struct i915_wa_list *wal = &rq->engine->ctx_wa_list;
832 	struct i915_wa *wa;
833 	unsigned int i;
834 	u32 *cs;
835 	int ret;
836 
837 	if (wal->count == 0)
838 		return 0;
839 
840 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
841 	if (ret)
842 		return ret;
843 
844 	cs = intel_ring_begin(rq, (wal->count * 2 + 2));
845 	if (IS_ERR(cs))
846 		return PTR_ERR(cs);
847 
848 	*cs++ = MI_LOAD_REGISTER_IMM(wal->count);
849 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
850 		*cs++ = i915_mmio_reg_offset(wa->reg);
851 		*cs++ = wa->set;
852 	}
853 	*cs++ = MI_NOOP;
854 
855 	intel_ring_advance(rq, cs);
856 
857 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
858 	if (ret)
859 		return ret;
860 
861 	return 0;
862 }
863 
864 static void
865 gen4_gt_workarounds_init(struct intel_gt *gt,
866 			 struct i915_wa_list *wal)
867 {
868 	/* WaDisable_RenderCache_OperationalFlush:gen4,ilk */
869 	wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
870 }
871 
872 static void
873 g4x_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
874 {
875 	gen4_gt_workarounds_init(gt, wal);
876 
877 	/* WaDisableRenderCachePipelinedFlush:g4x,ilk */
878 	wa_masked_en(wal, CACHE_MODE_0, CM0_PIPELINED_RENDER_FLUSH_DISABLE);
879 }
880 
881 static void
882 ilk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
883 {
884 	g4x_gt_workarounds_init(gt, wal);
885 
886 	wa_masked_en(wal, _3D_CHICKEN2, _3D_CHICKEN2_WM_READ_PIPELINED);
887 }
888 
889 static void
890 snb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
891 {
892 }
893 
894 static void
895 ivb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
896 {
897 	/* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
898 	wa_masked_dis(wal,
899 		      GEN7_COMMON_SLICE_CHICKEN1,
900 		      GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
901 
902 	/* WaApplyL3ControlAndL3ChickenMode:ivb */
903 	wa_write(wal, GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
904 	wa_write(wal, GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
905 
906 	/* WaForceL3Serialization:ivb */
907 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
908 }
909 
910 static void
911 vlv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
912 {
913 	/* WaForceL3Serialization:vlv */
914 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
915 
916 	/*
917 	 * WaIncreaseL3CreditsForVLVB0:vlv
918 	 * This is the hardware default actually.
919 	 */
920 	wa_write(wal, GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
921 }
922 
923 static void
924 hsw_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
925 {
926 	/* L3 caching of data atomics doesn't work -- disable it. */
927 	wa_write(wal, HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
928 
929 	wa_add(wal,
930 	       HSW_ROW_CHICKEN3, 0,
931 	       _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
932 	       0 /* XXX does this reg exist? */, true);
933 
934 	/* WaVSRefCountFullforceMissDisable:hsw */
935 	wa_write_clr(wal, GEN7_FF_THREAD_MODE, GEN7_FF_VS_REF_CNT_FFME);
936 }
937 
938 static void
939 gen9_wa_init_mcr(struct drm_i915_private *i915, struct i915_wa_list *wal)
940 {
941 	const struct sseu_dev_info *sseu = &to_gt(i915)->info.sseu;
942 	unsigned int slice, subslice;
943 	u32 mcr, mcr_mask;
944 
945 	GEM_BUG_ON(GRAPHICS_VER(i915) != 9);
946 
947 	/*
948 	 * WaProgramMgsrForCorrectSliceSpecificMmioReads:gen9,glk,kbl,cml
949 	 * Before any MMIO read into slice/subslice specific registers, MCR
950 	 * packet control register needs to be programmed to point to any
951 	 * enabled s/ss pair. Otherwise, incorrect values will be returned.
952 	 * This means each subsequent MMIO read will be forwarded to an
953 	 * specific s/ss combination, but this is OK since these registers
954 	 * are consistent across s/ss in almost all cases. In the rare
955 	 * occasions, such as INSTDONE, where this value is dependent
956 	 * on s/ss combo, the read should be done with read_subslice_reg.
957 	 */
958 	slice = ffs(sseu->slice_mask) - 1;
959 	GEM_BUG_ON(slice >= ARRAY_SIZE(sseu->subslice_mask.hsw));
960 	subslice = ffs(intel_sseu_get_hsw_subslices(sseu, slice));
961 	GEM_BUG_ON(!subslice);
962 	subslice--;
963 
964 	/*
965 	 * We use GEN8_MCR..() macros to calculate the |mcr| value for
966 	 * Gen9 to address WaProgramMgsrForCorrectSliceSpecificMmioReads
967 	 */
968 	mcr = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
969 	mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
970 
971 	drm_dbg(&i915->drm, "MCR slice:%d/subslice:%d = %x\n", slice, subslice, mcr);
972 
973 	wa_write_clr_set(wal, GEN8_MCR_SELECTOR, mcr_mask, mcr);
974 }
975 
976 static void
977 gen9_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
978 {
979 	struct drm_i915_private *i915 = gt->i915;
980 
981 	/* WaProgramMgsrForCorrectSliceSpecificMmioReads:glk,kbl,cml,gen9 */
982 	gen9_wa_init_mcr(i915, wal);
983 
984 	/* WaDisableKillLogic:bxt,skl,kbl */
985 	if (!IS_COFFEELAKE(i915) && !IS_COMETLAKE(i915))
986 		wa_write_or(wal,
987 			    GAM_ECOCHK,
988 			    ECOCHK_DIS_TLB);
989 
990 	if (HAS_LLC(i915)) {
991 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
992 		 *
993 		 * Must match Display Engine. See
994 		 * WaCompressedResourceDisplayNewHashMode.
995 		 */
996 		wa_write_or(wal,
997 			    MMCD_MISC_CTRL,
998 			    MMCD_PCLA | MMCD_HOTSPOT_EN);
999 	}
1000 
1001 	/* WaDisableHDCInvalidation:skl,bxt,kbl,cfl */
1002 	wa_write_or(wal,
1003 		    GAM_ECOCHK,
1004 		    BDW_DISABLE_HDC_INVALIDATION);
1005 }
1006 
1007 static void
1008 skl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1009 {
1010 	gen9_gt_workarounds_init(gt, wal);
1011 
1012 	/* WaDisableGafsUnitClkGating:skl */
1013 	wa_write_or(wal,
1014 		    GEN7_UCGCTL4,
1015 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1016 
1017 	/* WaInPlaceDecompressionHang:skl */
1018 	if (IS_SKL_GRAPHICS_STEP(gt->i915, STEP_A0, STEP_H0))
1019 		wa_write_or(wal,
1020 			    GEN9_GAMT_ECO_REG_RW_IA,
1021 			    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1022 }
1023 
1024 static void
1025 kbl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1026 {
1027 	gen9_gt_workarounds_init(gt, wal);
1028 
1029 	/* WaDisableDynamicCreditSharing:kbl */
1030 	if (IS_KBL_GRAPHICS_STEP(gt->i915, 0, STEP_C0))
1031 		wa_write_or(wal,
1032 			    GAMT_CHKN_BIT_REG,
1033 			    GAMT_CHKN_DISABLE_DYNAMIC_CREDIT_SHARING);
1034 
1035 	/* WaDisableGafsUnitClkGating:kbl */
1036 	wa_write_or(wal,
1037 		    GEN7_UCGCTL4,
1038 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1039 
1040 	/* WaInPlaceDecompressionHang:kbl */
1041 	wa_write_or(wal,
1042 		    GEN9_GAMT_ECO_REG_RW_IA,
1043 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1044 }
1045 
1046 static void
1047 glk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1048 {
1049 	gen9_gt_workarounds_init(gt, wal);
1050 }
1051 
1052 static void
1053 cfl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1054 {
1055 	gen9_gt_workarounds_init(gt, wal);
1056 
1057 	/* WaDisableGafsUnitClkGating:cfl */
1058 	wa_write_or(wal,
1059 		    GEN7_UCGCTL4,
1060 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1061 
1062 	/* WaInPlaceDecompressionHang:cfl */
1063 	wa_write_or(wal,
1064 		    GEN9_GAMT_ECO_REG_RW_IA,
1065 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1066 }
1067 
1068 static void __set_mcr_steering(struct i915_wa_list *wal,
1069 			       i915_reg_t steering_reg,
1070 			       unsigned int slice, unsigned int subslice)
1071 {
1072 	u32 mcr, mcr_mask;
1073 
1074 	mcr = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1075 	mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1076 
1077 	wa_write_clr_set(wal, steering_reg, mcr_mask, mcr);
1078 }
1079 
1080 static void __add_mcr_wa(struct intel_gt *gt, struct i915_wa_list *wal,
1081 			 unsigned int slice, unsigned int subslice)
1082 {
1083 	struct drm_printer p = drm_debug_printer("MCR Steering:");
1084 
1085 	__set_mcr_steering(wal, GEN8_MCR_SELECTOR, slice, subslice);
1086 
1087 	gt->default_steering.groupid = slice;
1088 	gt->default_steering.instanceid = subslice;
1089 
1090 	if (drm_debug_enabled(DRM_UT_DRIVER))
1091 		intel_gt_mcr_report_steering(&p, gt, false);
1092 }
1093 
1094 static void
1095 icl_wa_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1096 {
1097 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1098 	unsigned int subslice;
1099 
1100 	GEM_BUG_ON(GRAPHICS_VER(gt->i915) < 11);
1101 	GEM_BUG_ON(hweight8(sseu->slice_mask) > 1);
1102 
1103 	/*
1104 	 * Although a platform may have subslices, we need to always steer
1105 	 * reads to the lowest instance that isn't fused off.  When Render
1106 	 * Power Gating is enabled, grabbing forcewake will only power up a
1107 	 * single subslice (the "minconfig") if there isn't a real workload
1108 	 * that needs to be run; this means that if we steer register reads to
1109 	 * one of the higher subslices, we run the risk of reading back 0's or
1110 	 * random garbage.
1111 	 */
1112 	subslice = __ffs(intel_sseu_get_hsw_subslices(sseu, 0));
1113 
1114 	/*
1115 	 * If the subslice we picked above also steers us to a valid L3 bank,
1116 	 * then we can just rely on the default steering and won't need to
1117 	 * worry about explicitly re-steering L3BANK reads later.
1118 	 */
1119 	if (gt->info.l3bank_mask & BIT(subslice))
1120 		gt->steering_table[L3BANK] = NULL;
1121 
1122 	__add_mcr_wa(gt, wal, 0, subslice);
1123 }
1124 
1125 static void
1126 xehp_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1127 {
1128 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1129 	unsigned long slice, subslice = 0, slice_mask = 0;
1130 	u32 lncf_mask = 0;
1131 	int i;
1132 
1133 	/*
1134 	 * On Xe_HP the steering increases in complexity. There are now several
1135 	 * more units that require steering and we're not guaranteed to be able
1136 	 * to find a common setting for all of them. These are:
1137 	 * - GSLICE (fusable)
1138 	 * - DSS (sub-unit within gslice; fusable)
1139 	 * - L3 Bank (fusable)
1140 	 * - MSLICE (fusable)
1141 	 * - LNCF (sub-unit within mslice; always present if mslice is present)
1142 	 *
1143 	 * We'll do our default/implicit steering based on GSLICE (in the
1144 	 * sliceid field) and DSS (in the subsliceid field).  If we can
1145 	 * find overlap between the valid MSLICE and/or LNCF values with
1146 	 * a suitable GSLICE, then we can just re-use the default value and
1147 	 * skip and explicit steering at runtime.
1148 	 *
1149 	 * We only need to look for overlap between GSLICE/MSLICE/LNCF to find
1150 	 * a valid sliceid value.  DSS steering is the only type of steering
1151 	 * that utilizes the 'subsliceid' bits.
1152 	 *
1153 	 * Also note that, even though the steering domain is called "GSlice"
1154 	 * and it is encoded in the register using the gslice format, the spec
1155 	 * says that the combined (geometry | compute) fuse should be used to
1156 	 * select the steering.
1157 	 */
1158 
1159 	/* Find the potential gslice candidates */
1160 	slice_mask = intel_slicemask_from_xehp_dssmask(sseu->subslice_mask,
1161 						       GEN_DSS_PER_GSLICE);
1162 
1163 	/*
1164 	 * Find the potential LNCF candidates.  Either LNCF within a valid
1165 	 * mslice is fine.
1166 	 */
1167 	for_each_set_bit(i, &gt->info.mslice_mask, GEN12_MAX_MSLICES)
1168 		lncf_mask |= (0x3 << (i * 2));
1169 
1170 	/*
1171 	 * Are there any sliceid values that work for both GSLICE and LNCF
1172 	 * steering?
1173 	 */
1174 	if (slice_mask & lncf_mask) {
1175 		slice_mask &= lncf_mask;
1176 		gt->steering_table[LNCF] = NULL;
1177 	}
1178 
1179 	/* How about sliceid values that also work for MSLICE steering? */
1180 	if (slice_mask & gt->info.mslice_mask) {
1181 		slice_mask &= gt->info.mslice_mask;
1182 		gt->steering_table[MSLICE] = NULL;
1183 	}
1184 
1185 	slice = __ffs(slice_mask);
1186 	subslice = intel_sseu_find_first_xehp_dss(sseu, GEN_DSS_PER_GSLICE, slice) %
1187 		GEN_DSS_PER_GSLICE;
1188 
1189 	__add_mcr_wa(gt, wal, slice, subslice);
1190 
1191 	/*
1192 	 * SQIDI ranges are special because they use different steering
1193 	 * registers than everything else we work with.  On XeHP SDV and
1194 	 * DG2-G10, any value in the steering registers will work fine since
1195 	 * all instances are present, but DG2-G11 only has SQIDI instances at
1196 	 * ID's 2 and 3, so we need to steer to one of those.  For simplicity
1197 	 * we'll just steer to a hardcoded "2" since that value will work
1198 	 * everywhere.
1199 	 */
1200 	__set_mcr_steering(wal, MCFG_MCR_SELECTOR, 0, 2);
1201 	__set_mcr_steering(wal, SF_MCR_SELECTOR, 0, 2);
1202 }
1203 
1204 static void
1205 pvc_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1206 {
1207 	unsigned int dss;
1208 
1209 	/*
1210 	 * Setup implicit steering for COMPUTE and DSS ranges to the first
1211 	 * non-fused-off DSS.  All other types of MCR registers will be
1212 	 * explicitly steered.
1213 	 */
1214 	dss = intel_sseu_find_first_xehp_dss(&gt->info.sseu, 0, 0);
1215 	__add_mcr_wa(gt, wal, dss / GEN_DSS_PER_CSLICE, dss % GEN_DSS_PER_CSLICE);
1216 }
1217 
1218 static void
1219 icl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1220 {
1221 	struct drm_i915_private *i915 = gt->i915;
1222 
1223 	icl_wa_init_mcr(gt, wal);
1224 
1225 	/* WaModifyGamTlbPartitioning:icl */
1226 	wa_write_clr_set(wal,
1227 			 GEN11_GACB_PERF_CTRL,
1228 			 GEN11_HASH_CTRL_MASK,
1229 			 GEN11_HASH_CTRL_BIT0 | GEN11_HASH_CTRL_BIT4);
1230 
1231 	/* Wa_1405766107:icl
1232 	 * Formerly known as WaCL2SFHalfMaxAlloc
1233 	 */
1234 	wa_write_or(wal,
1235 		    GEN11_LSN_UNSLCVC,
1236 		    GEN11_LSN_UNSLCVC_GAFS_HALF_SF_MAXALLOC |
1237 		    GEN11_LSN_UNSLCVC_GAFS_HALF_CL2_MAXALLOC);
1238 
1239 	/* Wa_220166154:icl
1240 	 * Formerly known as WaDisCtxReload
1241 	 */
1242 	wa_write_or(wal,
1243 		    GEN8_GAMW_ECO_DEV_RW_IA,
1244 		    GAMW_ECO_DEV_CTX_RELOAD_DISABLE);
1245 
1246 	/* Wa_1406463099:icl
1247 	 * Formerly known as WaGamTlbPendError
1248 	 */
1249 	wa_write_or(wal,
1250 		    GAMT_CHKN_BIT_REG,
1251 		    GAMT_CHKN_DISABLE_L3_COH_PIPE);
1252 
1253 	/* Wa_1407352427:icl,ehl */
1254 	wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1255 		    PSDUNIT_CLKGATE_DIS);
1256 
1257 	/* Wa_1406680159:icl,ehl */
1258 	wa_write_or(wal,
1259 		    SUBSLICE_UNIT_LEVEL_CLKGATE,
1260 		    GWUNIT_CLKGATE_DIS);
1261 
1262 	/* Wa_1607087056:icl,ehl,jsl */
1263 	if (IS_ICELAKE(i915) ||
1264 	    IS_JSL_EHL_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1265 		wa_write_or(wal,
1266 			    SLICE_UNIT_LEVEL_CLKGATE,
1267 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1268 
1269 	/*
1270 	 * This is not a documented workaround, but rather an optimization
1271 	 * to reduce sampler power.
1272 	 */
1273 	wa_write_clr(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1274 }
1275 
1276 /*
1277  * Though there are per-engine instances of these registers,
1278  * they retain their value through engine resets and should
1279  * only be provided on the GT workaround list rather than
1280  * the engine-specific workaround list.
1281  */
1282 static void
1283 wa_14011060649(struct intel_gt *gt, struct i915_wa_list *wal)
1284 {
1285 	struct intel_engine_cs *engine;
1286 	int id;
1287 
1288 	for_each_engine(engine, gt, id) {
1289 		if (engine->class != VIDEO_DECODE_CLASS ||
1290 		    (engine->instance % 2))
1291 			continue;
1292 
1293 		wa_write_or(wal, VDBOX_CGCTL3F10(engine->mmio_base),
1294 			    IECPUNIT_CLKGATE_DIS);
1295 	}
1296 }
1297 
1298 static void
1299 gen12_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1300 {
1301 	icl_wa_init_mcr(gt, wal);
1302 
1303 	/* Wa_14011060649:tgl,rkl,dg1,adl-s,adl-p */
1304 	wa_14011060649(gt, wal);
1305 
1306 	/* Wa_14011059788:tgl,rkl,adl-s,dg1,adl-p */
1307 	wa_write_or(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1308 }
1309 
1310 static void
1311 tgl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1312 {
1313 	struct drm_i915_private *i915 = gt->i915;
1314 
1315 	gen12_gt_workarounds_init(gt, wal);
1316 
1317 	/* Wa_1409420604:tgl */
1318 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1319 		wa_write_or(wal,
1320 			    SUBSLICE_UNIT_LEVEL_CLKGATE2,
1321 			    CPSSUNIT_CLKGATE_DIS);
1322 
1323 	/* Wa_1607087056:tgl also know as BUG:1409180338 */
1324 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1325 		wa_write_or(wal,
1326 			    SLICE_UNIT_LEVEL_CLKGATE,
1327 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1328 
1329 	/* Wa_1408615072:tgl[a0] */
1330 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1331 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1332 			    VSUNIT_CLKGATE_DIS_TGL);
1333 }
1334 
1335 static void
1336 dg1_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1337 {
1338 	struct drm_i915_private *i915 = gt->i915;
1339 
1340 	gen12_gt_workarounds_init(gt, wal);
1341 
1342 	/* Wa_1607087056:dg1 */
1343 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1344 		wa_write_or(wal,
1345 			    SLICE_UNIT_LEVEL_CLKGATE,
1346 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1347 
1348 	/* Wa_1409420604:dg1 */
1349 	if (IS_DG1(i915))
1350 		wa_write_or(wal,
1351 			    SUBSLICE_UNIT_LEVEL_CLKGATE2,
1352 			    CPSSUNIT_CLKGATE_DIS);
1353 
1354 	/* Wa_1408615072:dg1 */
1355 	/* Empirical testing shows this register is unaffected by engine reset. */
1356 	if (IS_DG1(i915))
1357 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1358 			    VSUNIT_CLKGATE_DIS_TGL);
1359 }
1360 
1361 static void
1362 xehpsdv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1363 {
1364 	struct drm_i915_private *i915 = gt->i915;
1365 
1366 	xehp_init_mcr(gt, wal);
1367 
1368 	/* Wa_1409757795:xehpsdv */
1369 	wa_write_or(wal, SCCGCTL94DC, CG3DDISURB);
1370 
1371 	/* Wa_16011155590:xehpsdv */
1372 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1373 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1374 			    TSGUNIT_CLKGATE_DIS);
1375 
1376 	/* Wa_14011780169:xehpsdv */
1377 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_B0, STEP_FOREVER)) {
1378 		wa_write_or(wal, UNSLCGCTL9440, GAMTLBOACS_CLKGATE_DIS |
1379 			    GAMTLBVDBOX7_CLKGATE_DIS |
1380 			    GAMTLBVDBOX6_CLKGATE_DIS |
1381 			    GAMTLBVDBOX5_CLKGATE_DIS |
1382 			    GAMTLBVDBOX4_CLKGATE_DIS |
1383 			    GAMTLBVDBOX3_CLKGATE_DIS |
1384 			    GAMTLBVDBOX2_CLKGATE_DIS |
1385 			    GAMTLBVDBOX1_CLKGATE_DIS |
1386 			    GAMTLBVDBOX0_CLKGATE_DIS |
1387 			    GAMTLBKCR_CLKGATE_DIS |
1388 			    GAMTLBGUC_CLKGATE_DIS |
1389 			    GAMTLBBLT_CLKGATE_DIS);
1390 		wa_write_or(wal, UNSLCGCTL9444, GAMTLBGFXA0_CLKGATE_DIS |
1391 			    GAMTLBGFXA1_CLKGATE_DIS |
1392 			    GAMTLBCOMPA0_CLKGATE_DIS |
1393 			    GAMTLBCOMPA1_CLKGATE_DIS |
1394 			    GAMTLBCOMPB0_CLKGATE_DIS |
1395 			    GAMTLBCOMPB1_CLKGATE_DIS |
1396 			    GAMTLBCOMPC0_CLKGATE_DIS |
1397 			    GAMTLBCOMPC1_CLKGATE_DIS |
1398 			    GAMTLBCOMPD0_CLKGATE_DIS |
1399 			    GAMTLBCOMPD1_CLKGATE_DIS |
1400 			    GAMTLBMERT_CLKGATE_DIS   |
1401 			    GAMTLBVEBOX3_CLKGATE_DIS |
1402 			    GAMTLBVEBOX2_CLKGATE_DIS |
1403 			    GAMTLBVEBOX1_CLKGATE_DIS |
1404 			    GAMTLBVEBOX0_CLKGATE_DIS);
1405 	}
1406 
1407 	/* Wa_16012725990:xehpsdv */
1408 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A1, STEP_FOREVER))
1409 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE, VFUNIT_CLKGATE_DIS);
1410 
1411 	/* Wa_14011060649:xehpsdv */
1412 	wa_14011060649(gt, wal);
1413 }
1414 
1415 static void
1416 dg2_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1417 {
1418 	struct intel_engine_cs *engine;
1419 	int id;
1420 
1421 	xehp_init_mcr(gt, wal);
1422 
1423 	/* Wa_14011060649:dg2 */
1424 	wa_14011060649(gt, wal);
1425 
1426 	/*
1427 	 * Although there are per-engine instances of these registers,
1428 	 * they technically exist outside the engine itself and are not
1429 	 * impacted by engine resets.  Furthermore, they're part of the
1430 	 * GuC blacklist so trying to treat them as engine workarounds
1431 	 * will result in GuC initialization failure and a wedged GPU.
1432 	 */
1433 	for_each_engine(engine, gt, id) {
1434 		if (engine->class != VIDEO_DECODE_CLASS)
1435 			continue;
1436 
1437 		/* Wa_16010515920:dg2_g10 */
1438 		if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_B0))
1439 			wa_write_or(wal, VDBOX_CGCTL3F18(engine->mmio_base),
1440 				    ALNUNIT_CLKGATE_DIS);
1441 	}
1442 
1443 	if (IS_DG2_G10(gt->i915)) {
1444 		/* Wa_22010523718:dg2 */
1445 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1446 			    CG3DDISCFEG_CLKGATE_DIS);
1447 
1448 		/* Wa_14011006942:dg2 */
1449 		wa_write_or(wal, SUBSLICE_UNIT_LEVEL_CLKGATE,
1450 			    DSS_ROUTER_CLKGATE_DIS);
1451 	}
1452 
1453 	if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_B0)) {
1454 		/* Wa_14010948348:dg2_g10 */
1455 		wa_write_or(wal, UNSLCGCTL9430, MSQDUNIT_CLKGATE_DIS);
1456 
1457 		/* Wa_14011037102:dg2_g10 */
1458 		wa_write_or(wal, UNSLCGCTL9444, LTCDD_CLKGATE_DIS);
1459 
1460 		/* Wa_14011371254:dg2_g10 */
1461 		wa_write_or(wal, SLICE_UNIT_LEVEL_CLKGATE, NODEDSS_CLKGATE_DIS);
1462 
1463 		/* Wa_14011431319:dg2_g10 */
1464 		wa_write_or(wal, UNSLCGCTL9440, GAMTLBOACS_CLKGATE_DIS |
1465 			    GAMTLBVDBOX7_CLKGATE_DIS |
1466 			    GAMTLBVDBOX6_CLKGATE_DIS |
1467 			    GAMTLBVDBOX5_CLKGATE_DIS |
1468 			    GAMTLBVDBOX4_CLKGATE_DIS |
1469 			    GAMTLBVDBOX3_CLKGATE_DIS |
1470 			    GAMTLBVDBOX2_CLKGATE_DIS |
1471 			    GAMTLBVDBOX1_CLKGATE_DIS |
1472 			    GAMTLBVDBOX0_CLKGATE_DIS |
1473 			    GAMTLBKCR_CLKGATE_DIS |
1474 			    GAMTLBGUC_CLKGATE_DIS |
1475 			    GAMTLBBLT_CLKGATE_DIS);
1476 		wa_write_or(wal, UNSLCGCTL9444, GAMTLBGFXA0_CLKGATE_DIS |
1477 			    GAMTLBGFXA1_CLKGATE_DIS |
1478 			    GAMTLBCOMPA0_CLKGATE_DIS |
1479 			    GAMTLBCOMPA1_CLKGATE_DIS |
1480 			    GAMTLBCOMPB0_CLKGATE_DIS |
1481 			    GAMTLBCOMPB1_CLKGATE_DIS |
1482 			    GAMTLBCOMPC0_CLKGATE_DIS |
1483 			    GAMTLBCOMPC1_CLKGATE_DIS |
1484 			    GAMTLBCOMPD0_CLKGATE_DIS |
1485 			    GAMTLBCOMPD1_CLKGATE_DIS |
1486 			    GAMTLBMERT_CLKGATE_DIS   |
1487 			    GAMTLBVEBOX3_CLKGATE_DIS |
1488 			    GAMTLBVEBOX2_CLKGATE_DIS |
1489 			    GAMTLBVEBOX1_CLKGATE_DIS |
1490 			    GAMTLBVEBOX0_CLKGATE_DIS);
1491 
1492 		/* Wa_14010569222:dg2_g10 */
1493 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1494 			    GAMEDIA_CLKGATE_DIS);
1495 
1496 		/* Wa_14011028019:dg2_g10 */
1497 		wa_write_or(wal, SSMCGCTL9530, RTFUNIT_CLKGATE_DIS);
1498 	}
1499 
1500 	/* Wa_14014830051:dg2 */
1501 	wa_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1502 
1503 	/*
1504 	 * The following are not actually "workarounds" but rather
1505 	 * recommended tuning settings documented in the bspec's
1506 	 * performance guide section.
1507 	 */
1508 	wa_write_or(wal, GEN12_SQCM, EN_32B_ACCESS);
1509 
1510 	/* Wa_14015795083 */
1511 	wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1512 }
1513 
1514 static void
1515 pvc_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1516 {
1517 	pvc_init_mcr(gt, wal);
1518 
1519 	/* Wa_14015795083 */
1520 	wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1521 }
1522 
1523 static void
1524 gt_init_workarounds(struct intel_gt *gt, struct i915_wa_list *wal)
1525 {
1526 	struct drm_i915_private *i915 = gt->i915;
1527 
1528 	if (IS_PONTEVECCHIO(i915))
1529 		pvc_gt_workarounds_init(gt, wal);
1530 	else if (IS_DG2(i915))
1531 		dg2_gt_workarounds_init(gt, wal);
1532 	else if (IS_XEHPSDV(i915))
1533 		xehpsdv_gt_workarounds_init(gt, wal);
1534 	else if (IS_DG1(i915))
1535 		dg1_gt_workarounds_init(gt, wal);
1536 	else if (IS_TIGERLAKE(i915))
1537 		tgl_gt_workarounds_init(gt, wal);
1538 	else if (GRAPHICS_VER(i915) == 12)
1539 		gen12_gt_workarounds_init(gt, wal);
1540 	else if (GRAPHICS_VER(i915) == 11)
1541 		icl_gt_workarounds_init(gt, wal);
1542 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
1543 		cfl_gt_workarounds_init(gt, wal);
1544 	else if (IS_GEMINILAKE(i915))
1545 		glk_gt_workarounds_init(gt, wal);
1546 	else if (IS_KABYLAKE(i915))
1547 		kbl_gt_workarounds_init(gt, wal);
1548 	else if (IS_BROXTON(i915))
1549 		gen9_gt_workarounds_init(gt, wal);
1550 	else if (IS_SKYLAKE(i915))
1551 		skl_gt_workarounds_init(gt, wal);
1552 	else if (IS_HASWELL(i915))
1553 		hsw_gt_workarounds_init(gt, wal);
1554 	else if (IS_VALLEYVIEW(i915))
1555 		vlv_gt_workarounds_init(gt, wal);
1556 	else if (IS_IVYBRIDGE(i915))
1557 		ivb_gt_workarounds_init(gt, wal);
1558 	else if (GRAPHICS_VER(i915) == 6)
1559 		snb_gt_workarounds_init(gt, wal);
1560 	else if (GRAPHICS_VER(i915) == 5)
1561 		ilk_gt_workarounds_init(gt, wal);
1562 	else if (IS_G4X(i915))
1563 		g4x_gt_workarounds_init(gt, wal);
1564 	else if (GRAPHICS_VER(i915) == 4)
1565 		gen4_gt_workarounds_init(gt, wal);
1566 	else if (GRAPHICS_VER(i915) <= 8)
1567 		;
1568 	else
1569 		MISSING_CASE(GRAPHICS_VER(i915));
1570 }
1571 
1572 void intel_gt_init_workarounds(struct intel_gt *gt)
1573 {
1574 	struct i915_wa_list *wal = &gt->wa_list;
1575 
1576 	wa_init_start(wal, "GT", "global");
1577 	gt_init_workarounds(gt, wal);
1578 	wa_init_finish(wal);
1579 }
1580 
1581 static enum forcewake_domains
1582 wal_get_fw_for_rmw(struct intel_uncore *uncore, const struct i915_wa_list *wal)
1583 {
1584 	enum forcewake_domains fw = 0;
1585 	struct i915_wa *wa;
1586 	unsigned int i;
1587 
1588 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1589 		fw |= intel_uncore_forcewake_for_reg(uncore,
1590 						     wa->reg,
1591 						     FW_REG_READ |
1592 						     FW_REG_WRITE);
1593 
1594 	return fw;
1595 }
1596 
1597 static bool
1598 wa_verify(const struct i915_wa *wa, u32 cur, const char *name, const char *from)
1599 {
1600 	if ((cur ^ wa->set) & wa->read) {
1601 		DRM_ERROR("%s workaround lost on %s! (reg[%x]=0x%x, relevant bits were 0x%x vs expected 0x%x)\n",
1602 			  name, from, i915_mmio_reg_offset(wa->reg),
1603 			  cur, cur & wa->read, wa->set & wa->read);
1604 
1605 		return false;
1606 	}
1607 
1608 	return true;
1609 }
1610 
1611 static void
1612 wa_list_apply(struct intel_gt *gt, const struct i915_wa_list *wal)
1613 {
1614 	struct intel_uncore *uncore = gt->uncore;
1615 	enum forcewake_domains fw;
1616 	unsigned long flags;
1617 	struct i915_wa *wa;
1618 	unsigned int i;
1619 
1620 	if (!wal->count)
1621 		return;
1622 
1623 	fw = wal_get_fw_for_rmw(uncore, wal);
1624 
1625 	spin_lock_irqsave(&uncore->lock, flags);
1626 	intel_uncore_forcewake_get__locked(uncore, fw);
1627 
1628 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
1629 		u32 val, old = 0;
1630 
1631 		/* open-coded rmw due to steering */
1632 		old = wa->clr ? intel_gt_mcr_read_any_fw(gt, wa->reg) : 0;
1633 		val = (old & ~wa->clr) | wa->set;
1634 		if (val != old || !wa->clr)
1635 			intel_uncore_write_fw(uncore, wa->reg, val);
1636 
1637 		if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
1638 			wa_verify(wa, intel_gt_mcr_read_any_fw(gt, wa->reg),
1639 				  wal->name, "application");
1640 	}
1641 
1642 	intel_uncore_forcewake_put__locked(uncore, fw);
1643 	spin_unlock_irqrestore(&uncore->lock, flags);
1644 }
1645 
1646 void intel_gt_apply_workarounds(struct intel_gt *gt)
1647 {
1648 	wa_list_apply(gt, &gt->wa_list);
1649 }
1650 
1651 static bool wa_list_verify(struct intel_gt *gt,
1652 			   const struct i915_wa_list *wal,
1653 			   const char *from)
1654 {
1655 	struct intel_uncore *uncore = gt->uncore;
1656 	struct i915_wa *wa;
1657 	enum forcewake_domains fw;
1658 	unsigned long flags;
1659 	unsigned int i;
1660 	bool ok = true;
1661 
1662 	fw = wal_get_fw_for_rmw(uncore, wal);
1663 
1664 	spin_lock_irqsave(&uncore->lock, flags);
1665 	intel_uncore_forcewake_get__locked(uncore, fw);
1666 
1667 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1668 		ok &= wa_verify(wa,
1669 				intel_gt_mcr_read_any_fw(gt, wa->reg),
1670 				wal->name, from);
1671 
1672 	intel_uncore_forcewake_put__locked(uncore, fw);
1673 	spin_unlock_irqrestore(&uncore->lock, flags);
1674 
1675 	return ok;
1676 }
1677 
1678 bool intel_gt_verify_workarounds(struct intel_gt *gt, const char *from)
1679 {
1680 	return wa_list_verify(gt, &gt->wa_list, from);
1681 }
1682 
1683 __maybe_unused
1684 static bool is_nonpriv_flags_valid(u32 flags)
1685 {
1686 	/* Check only valid flag bits are set */
1687 	if (flags & ~RING_FORCE_TO_NONPRIV_MASK_VALID)
1688 		return false;
1689 
1690 	/* NB: Only 3 out of 4 enum values are valid for access field */
1691 	if ((flags & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
1692 	    RING_FORCE_TO_NONPRIV_ACCESS_INVALID)
1693 		return false;
1694 
1695 	return true;
1696 }
1697 
1698 static void
1699 whitelist_reg_ext(struct i915_wa_list *wal, i915_reg_t reg, u32 flags)
1700 {
1701 	struct i915_wa wa = {
1702 		.reg = reg
1703 	};
1704 
1705 	if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1706 		return;
1707 
1708 	if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1709 		return;
1710 
1711 	wa.reg.reg |= flags;
1712 	_wa_add(wal, &wa);
1713 }
1714 
1715 static void
1716 whitelist_reg(struct i915_wa_list *wal, i915_reg_t reg)
1717 {
1718 	whitelist_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1719 }
1720 
1721 static void gen9_whitelist_build(struct i915_wa_list *w)
1722 {
1723 	/* WaVFEStateAfterPipeControlwithMediaStateClear:skl,bxt,glk,cfl */
1724 	whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1725 
1726 	/* WaEnablePreemptionGranularityControlByUMD:skl,bxt,kbl,cfl,[cnl] */
1727 	whitelist_reg(w, GEN8_CS_CHICKEN1);
1728 
1729 	/* WaAllowUMDToModifyHDCChicken1:skl,bxt,kbl,glk,cfl */
1730 	whitelist_reg(w, GEN8_HDC_CHICKEN1);
1731 
1732 	/* WaSendPushConstantsFromMMIO:skl,bxt */
1733 	whitelist_reg(w, COMMON_SLICE_CHICKEN2);
1734 }
1735 
1736 static void skl_whitelist_build(struct intel_engine_cs *engine)
1737 {
1738 	struct i915_wa_list *w = &engine->whitelist;
1739 
1740 	if (engine->class != RENDER_CLASS)
1741 		return;
1742 
1743 	gen9_whitelist_build(w);
1744 
1745 	/* WaDisableLSQCROPERFforOCL:skl */
1746 	whitelist_reg(w, GEN8_L3SQCREG4);
1747 }
1748 
1749 static void bxt_whitelist_build(struct intel_engine_cs *engine)
1750 {
1751 	if (engine->class != RENDER_CLASS)
1752 		return;
1753 
1754 	gen9_whitelist_build(&engine->whitelist);
1755 }
1756 
1757 static void kbl_whitelist_build(struct intel_engine_cs *engine)
1758 {
1759 	struct i915_wa_list *w = &engine->whitelist;
1760 
1761 	if (engine->class != RENDER_CLASS)
1762 		return;
1763 
1764 	gen9_whitelist_build(w);
1765 
1766 	/* WaDisableLSQCROPERFforOCL:kbl */
1767 	whitelist_reg(w, GEN8_L3SQCREG4);
1768 }
1769 
1770 static void glk_whitelist_build(struct intel_engine_cs *engine)
1771 {
1772 	struct i915_wa_list *w = &engine->whitelist;
1773 
1774 	if (engine->class != RENDER_CLASS)
1775 		return;
1776 
1777 	gen9_whitelist_build(w);
1778 
1779 	/* WA #0862: Userspace has to set "Barrier Mode" to avoid hangs. */
1780 	whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1781 }
1782 
1783 static void cfl_whitelist_build(struct intel_engine_cs *engine)
1784 {
1785 	struct i915_wa_list *w = &engine->whitelist;
1786 
1787 	if (engine->class != RENDER_CLASS)
1788 		return;
1789 
1790 	gen9_whitelist_build(w);
1791 
1792 	/*
1793 	 * WaAllowPMDepthAndInvocationCountAccessFromUMD:cfl,whl,cml,aml
1794 	 *
1795 	 * This covers 4 register which are next to one another :
1796 	 *   - PS_INVOCATION_COUNT
1797 	 *   - PS_INVOCATION_COUNT_UDW
1798 	 *   - PS_DEPTH_COUNT
1799 	 *   - PS_DEPTH_COUNT_UDW
1800 	 */
1801 	whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1802 			  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1803 			  RING_FORCE_TO_NONPRIV_RANGE_4);
1804 }
1805 
1806 static void allow_read_ctx_timestamp(struct intel_engine_cs *engine)
1807 {
1808 	struct i915_wa_list *w = &engine->whitelist;
1809 
1810 	if (engine->class != RENDER_CLASS)
1811 		whitelist_reg_ext(w,
1812 				  RING_CTX_TIMESTAMP(engine->mmio_base),
1813 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1814 }
1815 
1816 static void cml_whitelist_build(struct intel_engine_cs *engine)
1817 {
1818 	allow_read_ctx_timestamp(engine);
1819 
1820 	cfl_whitelist_build(engine);
1821 }
1822 
1823 static void icl_whitelist_build(struct intel_engine_cs *engine)
1824 {
1825 	struct i915_wa_list *w = &engine->whitelist;
1826 
1827 	allow_read_ctx_timestamp(engine);
1828 
1829 	switch (engine->class) {
1830 	case RENDER_CLASS:
1831 		/* WaAllowUMDToModifyHalfSliceChicken7:icl */
1832 		whitelist_reg(w, GEN9_HALF_SLICE_CHICKEN7);
1833 
1834 		/* WaAllowUMDToModifySamplerMode:icl */
1835 		whitelist_reg(w, GEN10_SAMPLER_MODE);
1836 
1837 		/* WaEnableStateCacheRedirectToCS:icl */
1838 		whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1839 
1840 		/*
1841 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:icl
1842 		 *
1843 		 * This covers 4 register which are next to one another :
1844 		 *   - PS_INVOCATION_COUNT
1845 		 *   - PS_INVOCATION_COUNT_UDW
1846 		 *   - PS_DEPTH_COUNT
1847 		 *   - PS_DEPTH_COUNT_UDW
1848 		 */
1849 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1850 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1851 				  RING_FORCE_TO_NONPRIV_RANGE_4);
1852 		break;
1853 
1854 	case VIDEO_DECODE_CLASS:
1855 		/* hucStatusRegOffset */
1856 		whitelist_reg_ext(w, _MMIO(0x2000 + engine->mmio_base),
1857 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1858 		/* hucUKernelHdrInfoRegOffset */
1859 		whitelist_reg_ext(w, _MMIO(0x2014 + engine->mmio_base),
1860 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1861 		/* hucStatus2RegOffset */
1862 		whitelist_reg_ext(w, _MMIO(0x23B0 + engine->mmio_base),
1863 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1864 		break;
1865 
1866 	default:
1867 		break;
1868 	}
1869 }
1870 
1871 static void tgl_whitelist_build(struct intel_engine_cs *engine)
1872 {
1873 	struct i915_wa_list *w = &engine->whitelist;
1874 
1875 	allow_read_ctx_timestamp(engine);
1876 
1877 	switch (engine->class) {
1878 	case RENDER_CLASS:
1879 		/*
1880 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:tgl
1881 		 * Wa_1408556865:tgl
1882 		 *
1883 		 * This covers 4 registers which are next to one another :
1884 		 *   - PS_INVOCATION_COUNT
1885 		 *   - PS_INVOCATION_COUNT_UDW
1886 		 *   - PS_DEPTH_COUNT
1887 		 *   - PS_DEPTH_COUNT_UDW
1888 		 */
1889 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1890 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1891 				  RING_FORCE_TO_NONPRIV_RANGE_4);
1892 
1893 		/*
1894 		 * Wa_1808121037:tgl
1895 		 * Wa_14012131227:dg1
1896 		 * Wa_1508744258:tgl,rkl,dg1,adl-s,adl-p
1897 		 */
1898 		whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
1899 
1900 		/* Wa_1806527549:tgl */
1901 		whitelist_reg(w, HIZ_CHICKEN);
1902 		break;
1903 	default:
1904 		break;
1905 	}
1906 }
1907 
1908 static void dg1_whitelist_build(struct intel_engine_cs *engine)
1909 {
1910 	struct i915_wa_list *w = &engine->whitelist;
1911 
1912 	tgl_whitelist_build(engine);
1913 
1914 	/* GEN:BUG:1409280441:dg1 */
1915 	if (IS_DG1_GRAPHICS_STEP(engine->i915, STEP_A0, STEP_B0) &&
1916 	    (engine->class == RENDER_CLASS ||
1917 	     engine->class == COPY_ENGINE_CLASS))
1918 		whitelist_reg_ext(w, RING_ID(engine->mmio_base),
1919 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1920 }
1921 
1922 static void xehpsdv_whitelist_build(struct intel_engine_cs *engine)
1923 {
1924 	allow_read_ctx_timestamp(engine);
1925 }
1926 
1927 static void dg2_whitelist_build(struct intel_engine_cs *engine)
1928 {
1929 	struct i915_wa_list *w = &engine->whitelist;
1930 
1931 	allow_read_ctx_timestamp(engine);
1932 
1933 	switch (engine->class) {
1934 	case RENDER_CLASS:
1935 		/*
1936 		 * Wa_1507100340:dg2_g10
1937 		 *
1938 		 * This covers 4 registers which are next to one another :
1939 		 *   - PS_INVOCATION_COUNT
1940 		 *   - PS_INVOCATION_COUNT_UDW
1941 		 *   - PS_DEPTH_COUNT
1942 		 *   - PS_DEPTH_COUNT_UDW
1943 		 */
1944 		if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0))
1945 			whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1946 					  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1947 					  RING_FORCE_TO_NONPRIV_RANGE_4);
1948 
1949 		break;
1950 	case COMPUTE_CLASS:
1951 		/* Wa_16011157294:dg2_g10 */
1952 		if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0))
1953 			whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1954 		break;
1955 	default:
1956 		break;
1957 	}
1958 }
1959 
1960 static void blacklist_trtt(struct intel_engine_cs *engine)
1961 {
1962 	struct i915_wa_list *w = &engine->whitelist;
1963 
1964 	/*
1965 	 * Prevent read/write access to [0x4400, 0x4600) which covers
1966 	 * the TRTT range across all engines. Note that normally userspace
1967 	 * cannot access the other engines' trtt control, but for simplicity
1968 	 * we cover the entire range on each engine.
1969 	 */
1970 	whitelist_reg_ext(w, _MMIO(0x4400),
1971 			  RING_FORCE_TO_NONPRIV_DENY |
1972 			  RING_FORCE_TO_NONPRIV_RANGE_64);
1973 	whitelist_reg_ext(w, _MMIO(0x4500),
1974 			  RING_FORCE_TO_NONPRIV_DENY |
1975 			  RING_FORCE_TO_NONPRIV_RANGE_64);
1976 }
1977 
1978 static void pvc_whitelist_build(struct intel_engine_cs *engine)
1979 {
1980 	allow_read_ctx_timestamp(engine);
1981 
1982 	/* Wa_16014440446:pvc */
1983 	blacklist_trtt(engine);
1984 }
1985 
1986 void intel_engine_init_whitelist(struct intel_engine_cs *engine)
1987 {
1988 	struct drm_i915_private *i915 = engine->i915;
1989 	struct i915_wa_list *w = &engine->whitelist;
1990 
1991 	wa_init_start(w, "whitelist", engine->name);
1992 
1993 	if (IS_PONTEVECCHIO(i915))
1994 		pvc_whitelist_build(engine);
1995 	else if (IS_DG2(i915))
1996 		dg2_whitelist_build(engine);
1997 	else if (IS_XEHPSDV(i915))
1998 		xehpsdv_whitelist_build(engine);
1999 	else if (IS_DG1(i915))
2000 		dg1_whitelist_build(engine);
2001 	else if (GRAPHICS_VER(i915) == 12)
2002 		tgl_whitelist_build(engine);
2003 	else if (GRAPHICS_VER(i915) == 11)
2004 		icl_whitelist_build(engine);
2005 	else if (IS_COMETLAKE(i915))
2006 		cml_whitelist_build(engine);
2007 	else if (IS_COFFEELAKE(i915))
2008 		cfl_whitelist_build(engine);
2009 	else if (IS_GEMINILAKE(i915))
2010 		glk_whitelist_build(engine);
2011 	else if (IS_KABYLAKE(i915))
2012 		kbl_whitelist_build(engine);
2013 	else if (IS_BROXTON(i915))
2014 		bxt_whitelist_build(engine);
2015 	else if (IS_SKYLAKE(i915))
2016 		skl_whitelist_build(engine);
2017 	else if (GRAPHICS_VER(i915) <= 8)
2018 		;
2019 	else
2020 		MISSING_CASE(GRAPHICS_VER(i915));
2021 
2022 	wa_init_finish(w);
2023 }
2024 
2025 void intel_engine_apply_whitelist(struct intel_engine_cs *engine)
2026 {
2027 	const struct i915_wa_list *wal = &engine->whitelist;
2028 	struct intel_uncore *uncore = engine->uncore;
2029 	const u32 base = engine->mmio_base;
2030 	struct i915_wa *wa;
2031 	unsigned int i;
2032 
2033 	if (!wal->count)
2034 		return;
2035 
2036 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
2037 		intel_uncore_write(uncore,
2038 				   RING_FORCE_TO_NONPRIV(base, i),
2039 				   i915_mmio_reg_offset(wa->reg));
2040 
2041 	/* And clear the rest just in case of garbage */
2042 	for (; i < RING_MAX_NONPRIV_SLOTS; i++)
2043 		intel_uncore_write(uncore,
2044 				   RING_FORCE_TO_NONPRIV(base, i),
2045 				   i915_mmio_reg_offset(RING_NOPID(base)));
2046 }
2047 
2048 /*
2049  * engine_fake_wa_init(), a place holder to program the registers
2050  * which are not part of an official workaround defined by the
2051  * hardware team.
2052  * Adding programming of those register inside workaround will
2053  * allow utilizing wa framework to proper application and verification.
2054  */
2055 static void
2056 engine_fake_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2057 {
2058 	u8 mocs_w, mocs_r;
2059 
2060 	/*
2061 	 * RING_CMD_CCTL specifies the default MOCS entry that will be used
2062 	 * by the command streamer when executing commands that don't have
2063 	 * a way to explicitly specify a MOCS setting.  The default should
2064 	 * usually reference whichever MOCS entry corresponds to uncached
2065 	 * behavior, although use of a WB cached entry is recommended by the
2066 	 * spec in certain circumstances on specific platforms.
2067 	 */
2068 	if (GRAPHICS_VER(engine->i915) >= 12) {
2069 		mocs_r = engine->gt->mocs.uc_index;
2070 		mocs_w = engine->gt->mocs.uc_index;
2071 
2072 		if (HAS_L3_CCS_READ(engine->i915) &&
2073 		    engine->class == COMPUTE_CLASS) {
2074 			mocs_r = engine->gt->mocs.wb_index;
2075 
2076 			/*
2077 			 * Even on the few platforms where MOCS 0 is a
2078 			 * legitimate table entry, it's never the correct
2079 			 * setting to use here; we can assume the MOCS init
2080 			 * just forgot to initialize wb_index.
2081 			 */
2082 			drm_WARN_ON(&engine->i915->drm, mocs_r == 0);
2083 		}
2084 
2085 		wa_masked_field_set(wal,
2086 				    RING_CMD_CCTL(engine->mmio_base),
2087 				    CMD_CCTL_MOCS_MASK,
2088 				    CMD_CCTL_MOCS_OVERRIDE(mocs_w, mocs_r));
2089 	}
2090 }
2091 
2092 static bool needs_wa_1308578152(struct intel_engine_cs *engine)
2093 {
2094 	return intel_sseu_find_first_xehp_dss(&engine->gt->info.sseu, 0, 0) >=
2095 		GEN_DSS_PER_GSLICE;
2096 }
2097 
2098 static void
2099 rcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2100 {
2101 	struct drm_i915_private *i915 = engine->i915;
2102 
2103 	if (IS_DG2(i915)) {
2104 		/* Wa_1509235366:dg2 */
2105 		wa_write_or(wal, GEN12_GAMCNTRL_CTRL, INVALIDATION_BROADCAST_MODE_DIS |
2106 			    GLOBAL_INVALIDATION_MODE);
2107 	}
2108 
2109 	if (IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0)) {
2110 		/* Wa_14013392000:dg2_g11 */
2111 		wa_masked_en(wal, GEN7_ROW_CHICKEN2, GEN12_ENABLE_LARGE_GRF_MODE);
2112 
2113 		/* Wa_16011620976:dg2_g11 */
2114 		wa_write_or(wal, LSC_CHICKEN_BIT_0_UDW, DIS_CHAIN_2XSIMD8);
2115 	}
2116 
2117 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_FOREVER) ||
2118 	    IS_DG2_G11(i915) || IS_DG2_G12(i915)) {
2119 		/* Wa_1509727124:dg2 */
2120 		wa_masked_en(wal, GEN10_SAMPLER_MODE,
2121 			     SC_DISABLE_POWER_OPTIMIZATION_EBB);
2122 	}
2123 
2124 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0) ||
2125 	    IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0)) {
2126 		/* Wa_14012419201:dg2 */
2127 		wa_masked_en(wal, GEN9_ROW_CHICKEN4,
2128 			     GEN12_DISABLE_HDR_PAST_PAYLOAD_HOLD_FIX);
2129 	}
2130 
2131 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) ||
2132 	    IS_DG2_G11(i915)) {
2133 		/*
2134 		 * Wa_22012826095:dg2
2135 		 * Wa_22013059131:dg2
2136 		 */
2137 		wa_write_clr_set(wal, LSC_CHICKEN_BIT_0_UDW,
2138 				 MAXREQS_PER_BANK,
2139 				 REG_FIELD_PREP(MAXREQS_PER_BANK, 2));
2140 
2141 		/* Wa_22013059131:dg2 */
2142 		wa_write_or(wal, LSC_CHICKEN_BIT_0,
2143 			    FORCE_1_SUB_MESSAGE_PER_FRAGMENT);
2144 	}
2145 
2146 	/* Wa_1308578152:dg2_g10 when first gslice is fused off */
2147 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) &&
2148 	    needs_wa_1308578152(engine)) {
2149 		wa_masked_dis(wal, GEN12_CS_DEBUG_MODE1_CCCSUNIT_BE_COMMON,
2150 			      GEN12_REPLAY_MODE_GRANULARITY);
2151 	}
2152 
2153 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_FOREVER) ||
2154 	    IS_DG2_G11(i915) || IS_DG2_G12(i915)) {
2155 		/* Wa_22013037850:dg2 */
2156 		wa_write_or(wal, LSC_CHICKEN_BIT_0_UDW,
2157 			    DISABLE_128B_EVICTION_COMMAND_UDW);
2158 
2159 		/* Wa_22012856258:dg2 */
2160 		wa_masked_en(wal, GEN7_ROW_CHICKEN2,
2161 			     GEN12_DISABLE_READ_SUPPRESSION);
2162 
2163 		/*
2164 		 * Wa_22010960976:dg2
2165 		 * Wa_14013347512:dg2
2166 		 */
2167 		wa_masked_dis(wal, GEN12_HDC_CHICKEN0,
2168 			      LSC_L1_FLUSH_CTL_3D_DATAPORT_FLUSH_EVENTS_MASK);
2169 	}
2170 
2171 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0)) {
2172 		/*
2173 		 * Wa_1608949956:dg2_g10
2174 		 * Wa_14010198302:dg2_g10
2175 		 */
2176 		wa_masked_en(wal, GEN8_ROW_CHICKEN,
2177 			     MDQ_ARBITRATION_MODE | UGM_BACKUP_MODE);
2178 
2179 		/*
2180 		 * Wa_14010918519:dg2_g10
2181 		 *
2182 		 * LSC_CHICKEN_BIT_0 always reads back as 0 is this stepping,
2183 		 * so ignoring verification.
2184 		 */
2185 		wa_add(wal, LSC_CHICKEN_BIT_0_UDW, 0,
2186 		       FORCE_SLM_FENCE_SCOPE_TO_TILE | FORCE_UGM_FENCE_SCOPE_TO_TILE,
2187 		       0, false);
2188 	}
2189 
2190 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0)) {
2191 		/* Wa_22010430635:dg2 */
2192 		wa_masked_en(wal,
2193 			     GEN9_ROW_CHICKEN4,
2194 			     GEN12_DISABLE_GRF_CLEAR);
2195 
2196 		/* Wa_14010648519:dg2 */
2197 		wa_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
2198 	}
2199 
2200 	/* Wa_14013202645:dg2 */
2201 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) ||
2202 	    IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0))
2203 		wa_write_or(wal, RT_CTRL, DIS_NULL_QUERY);
2204 
2205 	/* Wa_22012532006:dg2 */
2206 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_C0) ||
2207 	    IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0))
2208 		wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
2209 			     DG2_DISABLE_ROUND_ENABLE_ALLOW_FOR_SSLA);
2210 
2211 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0)) {
2212 		/* Wa_14010680813:dg2_g10 */
2213 		wa_write_or(wal, GEN12_GAMSTLB_CTRL, CONTROL_BLOCK_CLKGATE_DIS |
2214 			    EGRESS_BLOCK_CLKGATE_DIS | TAG_BLOCK_CLKGATE_DIS);
2215 	}
2216 
2217 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0) ||
2218 	    IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0)) {
2219 		/* Wa_14012362059:dg2 */
2220 		wa_write_or(wal, GEN12_MERT_MOD_CTRL, FORCE_MISS_FTLB);
2221 	}
2222 
2223 	if (IS_DG2_GRAPHICS_STEP(i915, G11, STEP_B0, STEP_FOREVER) ||
2224 	    IS_DG2_G10(i915)) {
2225 		/* Wa_22014600077:dg2 */
2226 		wa_add(wal, GEN10_CACHE_MODE_SS, 0,
2227 		       _MASKED_BIT_ENABLE(ENABLE_EU_COUNT_FOR_TDL_FLUSH),
2228 		       0 /* Wa_14012342262 :write-only reg, so skip
2229 			    verification */,
2230 		       true);
2231 	}
2232 
2233 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2234 	    IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)) {
2235 		/*
2236 		 * Wa_1607138336:tgl[a0],dg1[a0]
2237 		 * Wa_1607063988:tgl[a0],dg1[a0]
2238 		 */
2239 		wa_write_or(wal,
2240 			    GEN9_CTX_PREEMPT_REG,
2241 			    GEN12_DISABLE_POSH_BUSY_FF_DOP_CG);
2242 	}
2243 
2244 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)) {
2245 		/*
2246 		 * Wa_1606679103:tgl
2247 		 * (see also Wa_1606682166:icl)
2248 		 */
2249 		wa_write_or(wal,
2250 			    GEN7_SARCHKMD,
2251 			    GEN7_DISABLE_SAMPLER_PREFETCH);
2252 	}
2253 
2254 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) || IS_DG1(i915) ||
2255 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2256 		/* Wa_1606931601:tgl,rkl,dg1,adl-s,adl-p */
2257 		wa_masked_en(wal, GEN7_ROW_CHICKEN2, GEN12_DISABLE_EARLY_READ);
2258 
2259 		/*
2260 		 * Wa_1407928979:tgl A*
2261 		 * Wa_18011464164:tgl[B0+],dg1[B0+]
2262 		 * Wa_22010931296:tgl[B0+],dg1[B0+]
2263 		 * Wa_14010919138:rkl,dg1,adl-s,adl-p
2264 		 */
2265 		wa_write_or(wal, GEN7_FF_THREAD_MODE,
2266 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2267 	}
2268 
2269 	if (IS_ALDERLAKE_P(i915) || IS_DG2(i915) || IS_ALDERLAKE_S(i915) ||
2270 	    IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2271 		/*
2272 		 * Wa_1606700617:tgl,dg1,adl-p
2273 		 * Wa_22010271021:tgl,rkl,dg1,adl-s,adl-p
2274 		 * Wa_14010826681:tgl,dg1,rkl,adl-p
2275 		 * Wa_18019627453:dg2
2276 		 */
2277 		wa_masked_en(wal,
2278 			     GEN9_CS_DEBUG_MODE1,
2279 			     FF_DOP_CLOCK_GATE_DISABLE);
2280 	}
2281 
2282 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2283 	    IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2284 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2285 		/* Wa_1409804808:tgl,rkl,dg1[a0],adl-s,adl-p */
2286 		wa_masked_en(wal, GEN7_ROW_CHICKEN2,
2287 			     GEN12_PUSH_CONST_DEREF_HOLD_DIS);
2288 
2289 		/*
2290 		 * Wa_1409085225:tgl
2291 		 * Wa_14010229206:tgl,rkl,dg1[a0],adl-s,adl-p
2292 		 */
2293 		wa_masked_en(wal, GEN9_ROW_CHICKEN4, GEN12_DISABLE_TDL_PUSH);
2294 	}
2295 
2296 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2297 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) || IS_ALDERLAKE_P(i915)) {
2298 		/*
2299 		 * Wa_1607030317:tgl
2300 		 * Wa_1607186500:tgl
2301 		 * Wa_1607297627:tgl,rkl,dg1[a0],adlp
2302 		 *
2303 		 * On TGL and RKL there are multiple entries for this WA in the
2304 		 * BSpec; some indicate this is an A0-only WA, others indicate
2305 		 * it applies to all steppings so we trust the "all steppings."
2306 		 * For DG1 this only applies to A0.
2307 		 */
2308 		wa_masked_en(wal,
2309 			     RING_PSMI_CTL(RENDER_RING_BASE),
2310 			     GEN12_WAIT_FOR_EVENT_POWER_DOWN_DISABLE |
2311 			     GEN8_RC_SEMA_IDLE_MSG_DISABLE);
2312 	}
2313 
2314 	if (IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) ||
2315 	    IS_ALDERLAKE_S(i915) || IS_ALDERLAKE_P(i915)) {
2316 		/* Wa_1406941453:tgl,rkl,dg1,adl-s,adl-p */
2317 		wa_masked_en(wal,
2318 			     GEN10_SAMPLER_MODE,
2319 			     ENABLE_SMALLPL);
2320 	}
2321 
2322 	if (GRAPHICS_VER(i915) == 11) {
2323 		/* This is not an Wa. Enable for better image quality */
2324 		wa_masked_en(wal,
2325 			     _3D_CHICKEN3,
2326 			     _3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE);
2327 
2328 		/*
2329 		 * Wa_1405543622:icl
2330 		 * Formerly known as WaGAPZPriorityScheme
2331 		 */
2332 		wa_write_or(wal,
2333 			    GEN8_GARBCNTL,
2334 			    GEN11_ARBITRATION_PRIO_ORDER_MASK);
2335 
2336 		/*
2337 		 * Wa_1604223664:icl
2338 		 * Formerly known as WaL3BankAddressHashing
2339 		 */
2340 		wa_write_clr_set(wal,
2341 				 GEN8_GARBCNTL,
2342 				 GEN11_HASH_CTRL_EXCL_MASK,
2343 				 GEN11_HASH_CTRL_EXCL_BIT0);
2344 		wa_write_clr_set(wal,
2345 				 GEN11_GLBLINVL,
2346 				 GEN11_BANK_HASH_ADDR_EXCL_MASK,
2347 				 GEN11_BANK_HASH_ADDR_EXCL_BIT0);
2348 
2349 		/*
2350 		 * Wa_1405733216:icl
2351 		 * Formerly known as WaDisableCleanEvicts
2352 		 */
2353 		wa_write_or(wal,
2354 			    GEN8_L3SQCREG4,
2355 			    GEN11_LQSC_CLEAN_EVICT_DISABLE);
2356 
2357 		/* Wa_1606682166:icl */
2358 		wa_write_or(wal,
2359 			    GEN7_SARCHKMD,
2360 			    GEN7_DISABLE_SAMPLER_PREFETCH);
2361 
2362 		/* Wa_1409178092:icl */
2363 		wa_write_clr_set(wal,
2364 				 GEN11_SCRATCH2,
2365 				 GEN11_COHERENT_PARTIAL_WRITE_MERGE_ENABLE,
2366 				 0);
2367 
2368 		/* WaEnable32PlaneMode:icl */
2369 		wa_masked_en(wal, GEN9_CSFE_CHICKEN1_RCS,
2370 			     GEN11_ENABLE_32_PLANE_MODE);
2371 
2372 		/*
2373 		 * Wa_1408615072:icl,ehl  (vsunit)
2374 		 * Wa_1407596294:icl,ehl  (hsunit)
2375 		 */
2376 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
2377 			    VSUNIT_CLKGATE_DIS | HSUNIT_CLKGATE_DIS);
2378 
2379 		/*
2380 		 * Wa_1408767742:icl[a2..forever],ehl[all]
2381 		 * Wa_1605460711:icl[a0..c0]
2382 		 */
2383 		wa_write_or(wal,
2384 			    GEN7_FF_THREAD_MODE,
2385 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2386 
2387 		/* Wa_22010271021 */
2388 		wa_masked_en(wal,
2389 			     GEN9_CS_DEBUG_MODE1,
2390 			     FF_DOP_CLOCK_GATE_DISABLE);
2391 	}
2392 
2393 	if (HAS_PERCTX_PREEMPT_CTRL(i915)) {
2394 		/* FtrPerCtxtPreemptionGranularityControl:skl,bxt,kbl,cfl,cnl,icl,tgl */
2395 		wa_masked_en(wal,
2396 			     GEN7_FF_SLICE_CS_CHICKEN1,
2397 			     GEN9_FFSC_PERCTX_PREEMPT_CTRL);
2398 	}
2399 
2400 	if (IS_SKYLAKE(i915) ||
2401 	    IS_KABYLAKE(i915) ||
2402 	    IS_COFFEELAKE(i915) ||
2403 	    IS_COMETLAKE(i915)) {
2404 		/* WaEnableGapsTsvCreditFix:skl,kbl,cfl */
2405 		wa_write_or(wal,
2406 			    GEN8_GARBCNTL,
2407 			    GEN9_GAPS_TSV_CREDIT_DISABLE);
2408 	}
2409 
2410 	if (IS_BROXTON(i915)) {
2411 		/* WaDisablePooledEuLoadBalancingFix:bxt */
2412 		wa_masked_en(wal,
2413 			     FF_SLICE_CS_CHICKEN2,
2414 			     GEN9_POOLED_EU_LOAD_BALANCING_FIX_DISABLE);
2415 	}
2416 
2417 	if (GRAPHICS_VER(i915) == 9) {
2418 		/* WaContextSwitchWithConcurrentTLBInvalidate:skl,bxt,kbl,glk,cfl */
2419 		wa_masked_en(wal,
2420 			     GEN9_CSFE_CHICKEN1_RCS,
2421 			     GEN9_PREEMPT_GPGPU_SYNC_SWITCH_DISABLE);
2422 
2423 		/* WaEnableLbsSlaRetryTimerDecrement:skl,bxt,kbl,glk,cfl */
2424 		wa_write_or(wal,
2425 			    BDW_SCRATCH1,
2426 			    GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
2427 
2428 		/* WaProgramL3SqcReg1DefaultForPerf:bxt,glk */
2429 		if (IS_GEN9_LP(i915))
2430 			wa_write_clr_set(wal,
2431 					 GEN8_L3SQCREG1,
2432 					 L3_PRIO_CREDITS_MASK,
2433 					 L3_GENERAL_PRIO_CREDITS(62) |
2434 					 L3_HIGH_PRIO_CREDITS(2));
2435 
2436 		/* WaOCLCoherentLineFlush:skl,bxt,kbl,cfl */
2437 		wa_write_or(wal,
2438 			    GEN8_L3SQCREG4,
2439 			    GEN8_LQSC_FLUSH_COHERENT_LINES);
2440 
2441 		/* Disable atomics in L3 to prevent unrecoverable hangs */
2442 		wa_write_clr_set(wal, GEN9_SCRATCH_LNCF1,
2443 				 GEN9_LNCF_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2444 		wa_write_clr_set(wal, GEN8_L3SQCREG4,
2445 				 GEN8_LQSQ_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2446 		wa_write_clr_set(wal, GEN9_SCRATCH1,
2447 				 EVICTION_PERF_FIX_ENABLE, 0);
2448 	}
2449 
2450 	if (IS_HASWELL(i915)) {
2451 		/* WaSampleCChickenBitEnable:hsw */
2452 		wa_masked_en(wal,
2453 			     HALF_SLICE_CHICKEN3, HSW_SAMPLE_C_PERFORMANCE);
2454 
2455 		wa_masked_dis(wal,
2456 			      CACHE_MODE_0_GEN7,
2457 			      /* enable HiZ Raw Stall Optimization */
2458 			      HIZ_RAW_STALL_OPT_DISABLE);
2459 	}
2460 
2461 	if (IS_VALLEYVIEW(i915)) {
2462 		/* WaDisableEarlyCull:vlv */
2463 		wa_masked_en(wal,
2464 			     _3D_CHICKEN3,
2465 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2466 
2467 		/*
2468 		 * WaVSThreadDispatchOverride:ivb,vlv
2469 		 *
2470 		 * This actually overrides the dispatch
2471 		 * mode for all thread types.
2472 		 */
2473 		wa_write_clr_set(wal,
2474 				 GEN7_FF_THREAD_MODE,
2475 				 GEN7_FF_SCHED_MASK,
2476 				 GEN7_FF_TS_SCHED_HW |
2477 				 GEN7_FF_VS_SCHED_HW |
2478 				 GEN7_FF_DS_SCHED_HW);
2479 
2480 		/* WaPsdDispatchEnable:vlv */
2481 		/* WaDisablePSDDualDispatchEnable:vlv */
2482 		wa_masked_en(wal,
2483 			     GEN7_HALF_SLICE_CHICKEN1,
2484 			     GEN7_MAX_PS_THREAD_DEP |
2485 			     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2486 	}
2487 
2488 	if (IS_IVYBRIDGE(i915)) {
2489 		/* WaDisableEarlyCull:ivb */
2490 		wa_masked_en(wal,
2491 			     _3D_CHICKEN3,
2492 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2493 
2494 		if (0) { /* causes HiZ corruption on ivb:gt1 */
2495 			/* enable HiZ Raw Stall Optimization */
2496 			wa_masked_dis(wal,
2497 				      CACHE_MODE_0_GEN7,
2498 				      HIZ_RAW_STALL_OPT_DISABLE);
2499 		}
2500 
2501 		/*
2502 		 * WaVSThreadDispatchOverride:ivb,vlv
2503 		 *
2504 		 * This actually overrides the dispatch
2505 		 * mode for all thread types.
2506 		 */
2507 		wa_write_clr_set(wal,
2508 				 GEN7_FF_THREAD_MODE,
2509 				 GEN7_FF_SCHED_MASK,
2510 				 GEN7_FF_TS_SCHED_HW |
2511 				 GEN7_FF_VS_SCHED_HW |
2512 				 GEN7_FF_DS_SCHED_HW);
2513 
2514 		/* WaDisablePSDDualDispatchEnable:ivb */
2515 		if (IS_IVB_GT1(i915))
2516 			wa_masked_en(wal,
2517 				     GEN7_HALF_SLICE_CHICKEN1,
2518 				     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2519 	}
2520 
2521 	if (GRAPHICS_VER(i915) == 7) {
2522 		/* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
2523 		wa_masked_en(wal,
2524 			     RING_MODE_GEN7(RENDER_RING_BASE),
2525 			     GFX_TLB_INVALIDATE_EXPLICIT | GFX_REPLAY_MODE);
2526 
2527 		/* WaDisable_RenderCache_OperationalFlush:ivb,vlv,hsw */
2528 		wa_masked_dis(wal, CACHE_MODE_0_GEN7, RC_OP_FLUSH_ENABLE);
2529 
2530 		/*
2531 		 * BSpec says this must be set, even though
2532 		 * WaDisable4x2SubspanOptimization:ivb,hsw
2533 		 * WaDisable4x2SubspanOptimization isn't listed for VLV.
2534 		 */
2535 		wa_masked_en(wal,
2536 			     CACHE_MODE_1,
2537 			     PIXEL_SUBSPAN_COLLECT_OPT_DISABLE);
2538 
2539 		/*
2540 		 * BSpec recommends 8x4 when MSAA is used,
2541 		 * however in practice 16x4 seems fastest.
2542 		 *
2543 		 * Note that PS/WM thread counts depend on the WIZ hashing
2544 		 * disable bit, which we don't touch here, but it's good
2545 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2546 		 */
2547 		wa_masked_field_set(wal,
2548 				    GEN7_GT_MODE,
2549 				    GEN6_WIZ_HASHING_MASK,
2550 				    GEN6_WIZ_HASHING_16x4);
2551 	}
2552 
2553 	if (IS_GRAPHICS_VER(i915, 6, 7))
2554 		/*
2555 		 * We need to disable the AsyncFlip performance optimisations in
2556 		 * order to use MI_WAIT_FOR_EVENT within the CS. It should
2557 		 * already be programmed to '1' on all products.
2558 		 *
2559 		 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
2560 		 */
2561 		wa_masked_en(wal,
2562 			     RING_MI_MODE(RENDER_RING_BASE),
2563 			     ASYNC_FLIP_PERF_DISABLE);
2564 
2565 	if (GRAPHICS_VER(i915) == 6) {
2566 		/*
2567 		 * Required for the hardware to program scanline values for
2568 		 * waiting
2569 		 * WaEnableFlushTlbInvalidationMode:snb
2570 		 */
2571 		wa_masked_en(wal,
2572 			     GFX_MODE,
2573 			     GFX_TLB_INVALIDATE_EXPLICIT);
2574 
2575 		/* WaDisableHiZPlanesWhenMSAAEnabled:snb */
2576 		wa_masked_en(wal,
2577 			     _3D_CHICKEN,
2578 			     _3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB);
2579 
2580 		wa_masked_en(wal,
2581 			     _3D_CHICKEN3,
2582 			     /* WaStripsFansDisableFastClipPerformanceFix:snb */
2583 			     _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL |
2584 			     /*
2585 			      * Bspec says:
2586 			      * "This bit must be set if 3DSTATE_CLIP clip mode is set
2587 			      * to normal and 3DSTATE_SF number of SF output attributes
2588 			      * is more than 16."
2589 			      */
2590 			     _3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH);
2591 
2592 		/*
2593 		 * BSpec recommends 8x4 when MSAA is used,
2594 		 * however in practice 16x4 seems fastest.
2595 		 *
2596 		 * Note that PS/WM thread counts depend on the WIZ hashing
2597 		 * disable bit, which we don't touch here, but it's good
2598 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2599 		 */
2600 		wa_masked_field_set(wal,
2601 				    GEN6_GT_MODE,
2602 				    GEN6_WIZ_HASHING_MASK,
2603 				    GEN6_WIZ_HASHING_16x4);
2604 
2605 		/* WaDisable_RenderCache_OperationalFlush:snb */
2606 		wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
2607 
2608 		/*
2609 		 * From the Sandybridge PRM, volume 1 part 3, page 24:
2610 		 * "If this bit is set, STCunit will have LRA as replacement
2611 		 *  policy. [...] This bit must be reset. LRA replacement
2612 		 *  policy is not supported."
2613 		 */
2614 		wa_masked_dis(wal,
2615 			      CACHE_MODE_0,
2616 			      CM0_STC_EVICT_DISABLE_LRA_SNB);
2617 	}
2618 
2619 	if (IS_GRAPHICS_VER(i915, 4, 6))
2620 		/* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
2621 		wa_add(wal, RING_MI_MODE(RENDER_RING_BASE),
2622 		       0, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH),
2623 		       /* XXX bit doesn't stick on Broadwater */
2624 		       IS_I965G(i915) ? 0 : VS_TIMER_DISPATCH, true);
2625 
2626 	if (GRAPHICS_VER(i915) == 4)
2627 		/*
2628 		 * Disable CONSTANT_BUFFER before it is loaded from the context
2629 		 * image. For as it is loaded, it is executed and the stored
2630 		 * address may no longer be valid, leading to a GPU hang.
2631 		 *
2632 		 * This imposes the requirement that userspace reload their
2633 		 * CONSTANT_BUFFER on every batch, fortunately a requirement
2634 		 * they are already accustomed to from before contexts were
2635 		 * enabled.
2636 		 */
2637 		wa_add(wal, ECOSKPD(RENDER_RING_BASE),
2638 		       0, _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE),
2639 		       0 /* XXX bit doesn't stick on Broadwater */,
2640 		       true);
2641 }
2642 
2643 static void
2644 xcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2645 {
2646 	struct drm_i915_private *i915 = engine->i915;
2647 
2648 	/* WaKBLVECSSemaphoreWaitPoll:kbl */
2649 	if (IS_KBL_GRAPHICS_STEP(i915, STEP_A0, STEP_F0)) {
2650 		wa_write(wal,
2651 			 RING_SEMA_WAIT_POLL(engine->mmio_base),
2652 			 1);
2653 	}
2654 }
2655 
2656 static void
2657 ccs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2658 {
2659 	if (IS_PVC_CT_STEP(engine->i915, STEP_A0, STEP_C0)) {
2660 		/* Wa_14014999345:pvc */
2661 		wa_masked_en(wal, GEN10_CACHE_MODE_SS, DISABLE_ECC);
2662 	}
2663 }
2664 
2665 /*
2666  * The bspec performance guide has recommended MMIO tuning settings.  These
2667  * aren't truly "workarounds" but we want to program them with the same
2668  * workaround infrastructure to ensure that they're automatically added to
2669  * the GuC save/restore lists, re-applied at the right times, and checked for
2670  * any conflicting programming requested by real workarounds.
2671  *
2672  * Programming settings should be added here only if their registers are not
2673  * part of an engine's register state context.  If a register is part of a
2674  * context, then any tuning settings should be programmed in an appropriate
2675  * function invoked by __intel_engine_init_ctx_wa().
2676  */
2677 static void
2678 add_render_compute_tuning_settings(struct drm_i915_private *i915,
2679 				   struct i915_wa_list *wal)
2680 {
2681 	if (IS_PONTEVECCHIO(i915)) {
2682 		wa_write(wal, XEHPC_L3SCRUB,
2683 			 SCRUB_CL_DWNGRADE_SHARED | SCRUB_RATE_4B_PER_CLK);
2684 	}
2685 
2686 	if (IS_DG2(i915)) {
2687 		wa_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
2688 		wa_write_clr_set(wal, RT_CTRL, STACKID_CTRL, STACKID_CTRL_512);
2689 		wa_write_clr_set(wal, DRAW_WATERMARK, VERT_WM_VAL,
2690 				 REG_FIELD_PREP(VERT_WM_VAL, 0x3FF));
2691 
2692 		/*
2693 		 * This is also listed as Wa_22012654132 for certain DG2
2694 		 * steppings, but the tuning setting programming is a superset
2695 		 * since it applies to all DG2 variants and steppings.
2696 		 *
2697 		 * Note that register 0xE420 is write-only and cannot be read
2698 		 * back for verification on DG2 (due to Wa_14012342262), so
2699 		 * we need to explicitly skip the readback.
2700 		 */
2701 		wa_add(wal, GEN10_CACHE_MODE_SS, 0,
2702 		       _MASKED_BIT_ENABLE(ENABLE_PREFETCH_INTO_IC),
2703 		       0 /* write-only, so skip validation */,
2704 		       true);
2705 	}
2706 }
2707 
2708 /*
2709  * The workarounds in this function apply to shared registers in
2710  * the general render reset domain that aren't tied to a
2711  * specific engine.  Since all render+compute engines get reset
2712  * together, and the contents of these registers are lost during
2713  * the shared render domain reset, we'll define such workarounds
2714  * here and then add them to just a single RCS or CCS engine's
2715  * workaround list (whichever engine has the XXXX flag).
2716  */
2717 static void
2718 general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2719 {
2720 	struct drm_i915_private *i915 = engine->i915;
2721 
2722 	add_render_compute_tuning_settings(i915, wal);
2723 
2724 	if (IS_PONTEVECCHIO(i915)) {
2725 		/* Wa_16016694945 */
2726 		wa_masked_en(wal, XEHPC_LNCFMISCCFGREG0, XEHPC_OVRLSCCC);
2727 	}
2728 
2729 	if (IS_XEHPSDV(i915)) {
2730 		/* Wa_1409954639 */
2731 		wa_masked_en(wal,
2732 			     GEN8_ROW_CHICKEN,
2733 			     SYSTOLIC_DOP_CLOCK_GATING_DIS);
2734 
2735 		/* Wa_1607196519 */
2736 		wa_masked_en(wal,
2737 			     GEN9_ROW_CHICKEN4,
2738 			     GEN12_DISABLE_GRF_CLEAR);
2739 
2740 		/* Wa_14010670810:xehpsdv */
2741 		wa_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
2742 
2743 		/* Wa_14010449647:xehpsdv */
2744 		wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
2745 			     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2746 
2747 		/* Wa_18011725039:xehpsdv */
2748 		if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A1, STEP_B0)) {
2749 			wa_masked_dis(wal, MLTICTXCTL, TDONRENDER);
2750 			wa_write_or(wal, L3SQCREG1_CCS0, FLUSHALLNONCOH);
2751 		}
2752 
2753 		/* Wa_14012362059:xehpsdv */
2754 		wa_write_or(wal, GEN12_MERT_MOD_CTRL, FORCE_MISS_FTLB);
2755 
2756 		/* Wa_14014368820:xehpsdv */
2757 		wa_write_or(wal, GEN12_GAMCNTRL_CTRL, INVALIDATION_BROADCAST_MODE_DIS |
2758 				GLOBAL_INVALIDATION_MODE);
2759 	}
2760 
2761 	if (IS_DG2(i915) || IS_PONTEVECCHIO(i915)) {
2762 		/* Wa_14015227452:dg2,pvc */
2763 		wa_masked_en(wal, GEN9_ROW_CHICKEN4, XEHP_DIS_BBL_SYSPIPE);
2764 
2765 		/* Wa_22014226127:dg2,pvc */
2766 		wa_write_or(wal, LSC_CHICKEN_BIT_0, DISABLE_D8_D16_COASLESCE);
2767 
2768 		/* Wa_16015675438:dg2,pvc */
2769 		wa_masked_en(wal, FF_SLICE_CS_CHICKEN2, GEN12_PERF_FIX_BALANCING_CFE_DISABLE);
2770 
2771 		/* Wa_18018781329:dg2,pvc */
2772 		wa_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
2773 		wa_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
2774 		wa_write_or(wal, VDBX_MOD_CTRL, FORCE_MISS_FTLB);
2775 		wa_write_or(wal, VEBX_MOD_CTRL, FORCE_MISS_FTLB);
2776 	}
2777 }
2778 
2779 static void
2780 engine_init_workarounds(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2781 {
2782 	if (I915_SELFTEST_ONLY(GRAPHICS_VER(engine->i915) < 4))
2783 		return;
2784 
2785 	engine_fake_wa_init(engine, wal);
2786 
2787 	/*
2788 	 * These are common workarounds that just need to applied
2789 	 * to a single RCS/CCS engine's workaround list since
2790 	 * they're reset as part of the general render domain reset.
2791 	 */
2792 	if (engine->flags & I915_ENGINE_FIRST_RENDER_COMPUTE)
2793 		general_render_compute_wa_init(engine, wal);
2794 
2795 	if (engine->class == COMPUTE_CLASS)
2796 		ccs_engine_wa_init(engine, wal);
2797 	else if (engine->class == RENDER_CLASS)
2798 		rcs_engine_wa_init(engine, wal);
2799 	else
2800 		xcs_engine_wa_init(engine, wal);
2801 }
2802 
2803 void intel_engine_init_workarounds(struct intel_engine_cs *engine)
2804 {
2805 	struct i915_wa_list *wal = &engine->wa_list;
2806 
2807 	if (GRAPHICS_VER(engine->i915) < 4)
2808 		return;
2809 
2810 	wa_init_start(wal, "engine", engine->name);
2811 	engine_init_workarounds(engine, wal);
2812 	wa_init_finish(wal);
2813 }
2814 
2815 void intel_engine_apply_workarounds(struct intel_engine_cs *engine)
2816 {
2817 	wa_list_apply(engine->gt, &engine->wa_list);
2818 }
2819 
2820 static const struct i915_range mcr_ranges_gen8[] = {
2821 	{ .start = 0x5500, .end = 0x55ff },
2822 	{ .start = 0x7000, .end = 0x7fff },
2823 	{ .start = 0x9400, .end = 0x97ff },
2824 	{ .start = 0xb000, .end = 0xb3ff },
2825 	{ .start = 0xe000, .end = 0xe7ff },
2826 	{},
2827 };
2828 
2829 static const struct i915_range mcr_ranges_gen12[] = {
2830 	{ .start =  0x8150, .end =  0x815f },
2831 	{ .start =  0x9520, .end =  0x955f },
2832 	{ .start =  0xb100, .end =  0xb3ff },
2833 	{ .start =  0xde80, .end =  0xe8ff },
2834 	{ .start = 0x24a00, .end = 0x24a7f },
2835 	{},
2836 };
2837 
2838 static const struct i915_range mcr_ranges_xehp[] = {
2839 	{ .start =  0x4000, .end =  0x4aff },
2840 	{ .start =  0x5200, .end =  0x52ff },
2841 	{ .start =  0x5400, .end =  0x7fff },
2842 	{ .start =  0x8140, .end =  0x815f },
2843 	{ .start =  0x8c80, .end =  0x8dff },
2844 	{ .start =  0x94d0, .end =  0x955f },
2845 	{ .start =  0x9680, .end =  0x96ff },
2846 	{ .start =  0xb000, .end =  0xb3ff },
2847 	{ .start =  0xc800, .end =  0xcfff },
2848 	{ .start =  0xd800, .end =  0xd8ff },
2849 	{ .start =  0xdc00, .end =  0xffff },
2850 	{ .start = 0x17000, .end = 0x17fff },
2851 	{ .start = 0x24a00, .end = 0x24a7f },
2852 	{},
2853 };
2854 
2855 static bool mcr_range(struct drm_i915_private *i915, u32 offset)
2856 {
2857 	const struct i915_range *mcr_ranges;
2858 	int i;
2859 
2860 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50))
2861 		mcr_ranges = mcr_ranges_xehp;
2862 	else if (GRAPHICS_VER(i915) >= 12)
2863 		mcr_ranges = mcr_ranges_gen12;
2864 	else if (GRAPHICS_VER(i915) >= 8)
2865 		mcr_ranges = mcr_ranges_gen8;
2866 	else
2867 		return false;
2868 
2869 	/*
2870 	 * Registers in these ranges are affected by the MCR selector
2871 	 * which only controls CPU initiated MMIO. Routing does not
2872 	 * work for CS access so we cannot verify them on this path.
2873 	 */
2874 	for (i = 0; mcr_ranges[i].start; i++)
2875 		if (offset >= mcr_ranges[i].start &&
2876 		    offset <= mcr_ranges[i].end)
2877 			return true;
2878 
2879 	return false;
2880 }
2881 
2882 static int
2883 wa_list_srm(struct i915_request *rq,
2884 	    const struct i915_wa_list *wal,
2885 	    struct i915_vma *vma)
2886 {
2887 	struct drm_i915_private *i915 = rq->engine->i915;
2888 	unsigned int i, count = 0;
2889 	const struct i915_wa *wa;
2890 	u32 srm, *cs;
2891 
2892 	srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
2893 	if (GRAPHICS_VER(i915) >= 8)
2894 		srm++;
2895 
2896 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2897 		if (!mcr_range(i915, i915_mmio_reg_offset(wa->reg)))
2898 			count++;
2899 	}
2900 
2901 	cs = intel_ring_begin(rq, 4 * count);
2902 	if (IS_ERR(cs))
2903 		return PTR_ERR(cs);
2904 
2905 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2906 		u32 offset = i915_mmio_reg_offset(wa->reg);
2907 
2908 		if (mcr_range(i915, offset))
2909 			continue;
2910 
2911 		*cs++ = srm;
2912 		*cs++ = offset;
2913 		*cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
2914 		*cs++ = 0;
2915 	}
2916 	intel_ring_advance(rq, cs);
2917 
2918 	return 0;
2919 }
2920 
2921 static int engine_wa_list_verify(struct intel_context *ce,
2922 				 const struct i915_wa_list * const wal,
2923 				 const char *from)
2924 {
2925 	const struct i915_wa *wa;
2926 	struct i915_request *rq;
2927 	struct i915_vma *vma;
2928 	struct i915_gem_ww_ctx ww;
2929 	unsigned int i;
2930 	u32 *results;
2931 	int err;
2932 
2933 	if (!wal->count)
2934 		return 0;
2935 
2936 	vma = __vm_create_scratch_for_read(&ce->engine->gt->ggtt->vm,
2937 					   wal->count * sizeof(u32));
2938 	if (IS_ERR(vma))
2939 		return PTR_ERR(vma);
2940 
2941 	intel_engine_pm_get(ce->engine);
2942 	i915_gem_ww_ctx_init(&ww, false);
2943 retry:
2944 	err = i915_gem_object_lock(vma->obj, &ww);
2945 	if (err == 0)
2946 		err = intel_context_pin_ww(ce, &ww);
2947 	if (err)
2948 		goto err_pm;
2949 
2950 	err = i915_vma_pin_ww(vma, &ww, 0, 0,
2951 			   i915_vma_is_ggtt(vma) ? PIN_GLOBAL : PIN_USER);
2952 	if (err)
2953 		goto err_unpin;
2954 
2955 	rq = i915_request_create(ce);
2956 	if (IS_ERR(rq)) {
2957 		err = PTR_ERR(rq);
2958 		goto err_vma;
2959 	}
2960 
2961 	err = i915_request_await_object(rq, vma->obj, true);
2962 	if (err == 0)
2963 		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
2964 	if (err == 0)
2965 		err = wa_list_srm(rq, wal, vma);
2966 
2967 	i915_request_get(rq);
2968 	if (err)
2969 		i915_request_set_error_once(rq, err);
2970 	i915_request_add(rq);
2971 
2972 	if (err)
2973 		goto err_rq;
2974 
2975 	if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2976 		err = -ETIME;
2977 		goto err_rq;
2978 	}
2979 
2980 	results = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
2981 	if (IS_ERR(results)) {
2982 		err = PTR_ERR(results);
2983 		goto err_rq;
2984 	}
2985 
2986 	err = 0;
2987 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2988 		if (mcr_range(rq->engine->i915, i915_mmio_reg_offset(wa->reg)))
2989 			continue;
2990 
2991 		if (!wa_verify(wa, results[i], wal->name, from))
2992 			err = -ENXIO;
2993 	}
2994 
2995 	i915_gem_object_unpin_map(vma->obj);
2996 
2997 err_rq:
2998 	i915_request_put(rq);
2999 err_vma:
3000 	i915_vma_unpin(vma);
3001 err_unpin:
3002 	intel_context_unpin(ce);
3003 err_pm:
3004 	if (err == -EDEADLK) {
3005 		err = i915_gem_ww_ctx_backoff(&ww);
3006 		if (!err)
3007 			goto retry;
3008 	}
3009 	i915_gem_ww_ctx_fini(&ww);
3010 	intel_engine_pm_put(ce->engine);
3011 	i915_vma_put(vma);
3012 	return err;
3013 }
3014 
3015 int intel_engine_verify_workarounds(struct intel_engine_cs *engine,
3016 				    const char *from)
3017 {
3018 	return engine_wa_list_verify(engine->kernel_context,
3019 				     &engine->wa_list,
3020 				     from);
3021 }
3022 
3023 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3024 #include "selftest_workarounds.c"
3025 #endif
3026