1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014-2018 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "intel_context.h"
8 #include "intel_engine_pm.h"
9 #include "intel_engine_regs.h"
10 #include "intel_gpu_commands.h"
11 #include "intel_gt.h"
12 #include "intel_gt_mcr.h"
13 #include "intel_gt_regs.h"
14 #include "intel_ring.h"
15 #include "intel_workarounds.h"
16 
17 /**
18  * DOC: Hardware workarounds
19  *
20  * This file is intended as a central place to implement most [1]_ of the
21  * required workarounds for hardware to work as originally intended. They fall
22  * in five basic categories depending on how/when they are applied:
23  *
24  * - Workarounds that touch registers that are saved/restored to/from the HW
25  *   context image. The list is emitted (via Load Register Immediate commands)
26  *   everytime a new context is created.
27  * - GT workarounds. The list of these WAs is applied whenever these registers
28  *   revert to default values (on GPU reset, suspend/resume [2]_, etc..).
29  * - Display workarounds. The list is applied during display clock-gating
30  *   initialization.
31  * - Workarounds that whitelist a privileged register, so that UMDs can manage
32  *   them directly. This is just a special case of a MMMIO workaround (as we
33  *   write the list of these to/be-whitelisted registers to some special HW
34  *   registers).
35  * - Workaround batchbuffers, that get executed automatically by the hardware
36  *   on every HW context restore.
37  *
38  * .. [1] Please notice that there are other WAs that, due to their nature,
39  *    cannot be applied from a central place. Those are peppered around the rest
40  *    of the code, as needed.
41  *
42  * .. [2] Technically, some registers are powercontext saved & restored, so they
43  *    survive a suspend/resume. In practice, writing them again is not too
44  *    costly and simplifies things. We can revisit this in the future.
45  *
46  * Layout
47  * ~~~~~~
48  *
49  * Keep things in this file ordered by WA type, as per the above (context, GT,
50  * display, register whitelist, batchbuffer). Then, inside each type, keep the
51  * following order:
52  *
53  * - Infrastructure functions and macros
54  * - WAs per platform in standard gen/chrono order
55  * - Public functions to init or apply the given workaround type.
56  */
57 
58 static void wa_init_start(struct i915_wa_list *wal, const char *name, const char *engine_name)
59 {
60 	wal->name = name;
61 	wal->engine_name = engine_name;
62 }
63 
64 #define WA_LIST_CHUNK (1 << 4)
65 
66 static void wa_init_finish(struct i915_wa_list *wal)
67 {
68 	/* Trim unused entries. */
69 	if (!IS_ALIGNED(wal->count, WA_LIST_CHUNK)) {
70 		struct i915_wa *list = kmemdup(wal->list,
71 					       wal->count * sizeof(*list),
72 					       GFP_KERNEL);
73 
74 		if (list) {
75 			kfree(wal->list);
76 			wal->list = list;
77 		}
78 	}
79 
80 	if (!wal->count)
81 		return;
82 
83 	DRM_DEBUG_DRIVER("Initialized %u %s workarounds on %s\n",
84 			 wal->wa_count, wal->name, wal->engine_name);
85 }
86 
87 static void _wa_add(struct i915_wa_list *wal, const struct i915_wa *wa)
88 {
89 	unsigned int addr = i915_mmio_reg_offset(wa->reg);
90 	unsigned int start = 0, end = wal->count;
91 	const unsigned int grow = WA_LIST_CHUNK;
92 	struct i915_wa *wa_;
93 
94 	GEM_BUG_ON(!is_power_of_2(grow));
95 
96 	if (IS_ALIGNED(wal->count, grow)) { /* Either uninitialized or full. */
97 		struct i915_wa *list;
98 
99 		list = kmalloc_array(ALIGN(wal->count + 1, grow), sizeof(*wa),
100 				     GFP_KERNEL);
101 		if (!list) {
102 			DRM_ERROR("No space for workaround init!\n");
103 			return;
104 		}
105 
106 		if (wal->list) {
107 			memcpy(list, wal->list, sizeof(*wa) * wal->count);
108 			kfree(wal->list);
109 		}
110 
111 		wal->list = list;
112 	}
113 
114 	while (start < end) {
115 		unsigned int mid = start + (end - start) / 2;
116 
117 		if (i915_mmio_reg_offset(wal->list[mid].reg) < addr) {
118 			start = mid + 1;
119 		} else if (i915_mmio_reg_offset(wal->list[mid].reg) > addr) {
120 			end = mid;
121 		} else {
122 			wa_ = &wal->list[mid];
123 
124 			if ((wa->clr | wa_->clr) && !(wa->clr & ~wa_->clr)) {
125 				DRM_ERROR("Discarding overwritten w/a for reg %04x (clear: %08x, set: %08x)\n",
126 					  i915_mmio_reg_offset(wa_->reg),
127 					  wa_->clr, wa_->set);
128 
129 				wa_->set &= ~wa->clr;
130 			}
131 
132 			wal->wa_count++;
133 			wa_->set |= wa->set;
134 			wa_->clr |= wa->clr;
135 			wa_->read |= wa->read;
136 			return;
137 		}
138 	}
139 
140 	wal->wa_count++;
141 	wa_ = &wal->list[wal->count++];
142 	*wa_ = *wa;
143 
144 	while (wa_-- > wal->list) {
145 		GEM_BUG_ON(i915_mmio_reg_offset(wa_[0].reg) ==
146 			   i915_mmio_reg_offset(wa_[1].reg));
147 		if (i915_mmio_reg_offset(wa_[1].reg) >
148 		    i915_mmio_reg_offset(wa_[0].reg))
149 			break;
150 
151 		swap(wa_[1], wa_[0]);
152 	}
153 }
154 
155 static void wa_add(struct i915_wa_list *wal, i915_reg_t reg,
156 		   u32 clear, u32 set, u32 read_mask, bool masked_reg)
157 {
158 	struct i915_wa wa = {
159 		.reg  = reg,
160 		.clr  = clear,
161 		.set  = set,
162 		.read = read_mask,
163 		.masked_reg = masked_reg,
164 	};
165 
166 	_wa_add(wal, &wa);
167 }
168 
169 static void
170 wa_write_clr_set(struct i915_wa_list *wal, i915_reg_t reg, u32 clear, u32 set)
171 {
172 	wa_add(wal, reg, clear, set, clear, false);
173 }
174 
175 static void
176 wa_write(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
177 {
178 	wa_write_clr_set(wal, reg, ~0, set);
179 }
180 
181 static void
182 wa_write_or(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
183 {
184 	wa_write_clr_set(wal, reg, set, set);
185 }
186 
187 static void
188 wa_write_clr(struct i915_wa_list *wal, i915_reg_t reg, u32 clr)
189 {
190 	wa_write_clr_set(wal, reg, clr, 0);
191 }
192 
193 /*
194  * WA operations on "masked register". A masked register has the upper 16 bits
195  * documented as "masked" in b-spec. Its purpose is to allow writing to just a
196  * portion of the register without a rmw: you simply write in the upper 16 bits
197  * the mask of bits you are going to modify.
198  *
199  * The wa_masked_* family of functions already does the necessary operations to
200  * calculate the mask based on the parameters passed, so user only has to
201  * provide the lower 16 bits of that register.
202  */
203 
204 static void
205 wa_masked_en(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
206 {
207 	wa_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
208 }
209 
210 static void
211 wa_masked_dis(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
212 {
213 	wa_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
214 }
215 
216 static void
217 wa_masked_field_set(struct i915_wa_list *wal, i915_reg_t reg,
218 		    u32 mask, u32 val)
219 {
220 	wa_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
221 }
222 
223 static void gen6_ctx_workarounds_init(struct intel_engine_cs *engine,
224 				      struct i915_wa_list *wal)
225 {
226 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
227 }
228 
229 static void gen7_ctx_workarounds_init(struct intel_engine_cs *engine,
230 				      struct i915_wa_list *wal)
231 {
232 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
233 }
234 
235 static void gen8_ctx_workarounds_init(struct intel_engine_cs *engine,
236 				      struct i915_wa_list *wal)
237 {
238 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
239 
240 	/* WaDisableAsyncFlipPerfMode:bdw,chv */
241 	wa_masked_en(wal, RING_MI_MODE(RENDER_RING_BASE), ASYNC_FLIP_PERF_DISABLE);
242 
243 	/* WaDisablePartialInstShootdown:bdw,chv */
244 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
245 		     PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
246 
247 	/* Use Force Non-Coherent whenever executing a 3D context. This is a
248 	 * workaround for a possible hang in the unlikely event a TLB
249 	 * invalidation occurs during a PSD flush.
250 	 */
251 	/* WaForceEnableNonCoherent:bdw,chv */
252 	/* WaHdcDisableFetchWhenMasked:bdw,chv */
253 	wa_masked_en(wal, HDC_CHICKEN0,
254 		     HDC_DONOT_FETCH_MEM_WHEN_MASKED |
255 		     HDC_FORCE_NON_COHERENT);
256 
257 	/* From the Haswell PRM, Command Reference: Registers, CACHE_MODE_0:
258 	 * "The Hierarchical Z RAW Stall Optimization allows non-overlapping
259 	 *  polygons in the same 8x4 pixel/sample area to be processed without
260 	 *  stalling waiting for the earlier ones to write to Hierarchical Z
261 	 *  buffer."
262 	 *
263 	 * This optimization is off by default for BDW and CHV; turn it on.
264 	 */
265 	wa_masked_dis(wal, CACHE_MODE_0_GEN7, HIZ_RAW_STALL_OPT_DISABLE);
266 
267 	/* Wa4x4STCOptimizationDisable:bdw,chv */
268 	wa_masked_en(wal, CACHE_MODE_1, GEN8_4x4_STC_OPTIMIZATION_DISABLE);
269 
270 	/*
271 	 * BSpec recommends 8x4 when MSAA is used,
272 	 * however in practice 16x4 seems fastest.
273 	 *
274 	 * Note that PS/WM thread counts depend on the WIZ hashing
275 	 * disable bit, which we don't touch here, but it's good
276 	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
277 	 */
278 	wa_masked_field_set(wal, GEN7_GT_MODE,
279 			    GEN6_WIZ_HASHING_MASK,
280 			    GEN6_WIZ_HASHING_16x4);
281 }
282 
283 static void bdw_ctx_workarounds_init(struct intel_engine_cs *engine,
284 				     struct i915_wa_list *wal)
285 {
286 	struct drm_i915_private *i915 = engine->i915;
287 
288 	gen8_ctx_workarounds_init(engine, wal);
289 
290 	/* WaDisableThreadStallDopClockGating:bdw (pre-production) */
291 	wa_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
292 
293 	/* WaDisableDopClockGating:bdw
294 	 *
295 	 * Also see the related UCGTCL1 write in bdw_init_clock_gating()
296 	 * to disable EUTC clock gating.
297 	 */
298 	wa_masked_en(wal, GEN7_ROW_CHICKEN2,
299 		     DOP_CLOCK_GATING_DISABLE);
300 
301 	wa_masked_en(wal, HALF_SLICE_CHICKEN3,
302 		     GEN8_SAMPLER_POWER_BYPASS_DIS);
303 
304 	wa_masked_en(wal, HDC_CHICKEN0,
305 		     /* WaForceContextSaveRestoreNonCoherent:bdw */
306 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
307 		     /* WaDisableFenceDestinationToSLM:bdw (pre-prod) */
308 		     (IS_BDW_GT3(i915) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
309 }
310 
311 static void chv_ctx_workarounds_init(struct intel_engine_cs *engine,
312 				     struct i915_wa_list *wal)
313 {
314 	gen8_ctx_workarounds_init(engine, wal);
315 
316 	/* WaDisableThreadStallDopClockGating:chv */
317 	wa_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
318 
319 	/* Improve HiZ throughput on CHV. */
320 	wa_masked_en(wal, HIZ_CHICKEN, CHV_HZ_8X8_MODE_IN_1X);
321 }
322 
323 static void gen9_ctx_workarounds_init(struct intel_engine_cs *engine,
324 				      struct i915_wa_list *wal)
325 {
326 	struct drm_i915_private *i915 = engine->i915;
327 
328 	if (HAS_LLC(i915)) {
329 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
330 		 *
331 		 * Must match Display Engine. See
332 		 * WaCompressedResourceDisplayNewHashMode.
333 		 */
334 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
335 			     GEN9_PBE_COMPRESSED_HASH_SELECTION);
336 		wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
337 			     GEN9_SAMPLER_HASH_COMPRESSED_READ_ADDR);
338 	}
339 
340 	/* WaClearFlowControlGpgpuContextSave:skl,bxt,kbl,glk,cfl */
341 	/* WaDisablePartialInstShootdown:skl,bxt,kbl,glk,cfl */
342 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
343 		     FLOW_CONTROL_ENABLE |
344 		     PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
345 
346 	/* WaEnableYV12BugFixInHalfSliceChicken7:skl,bxt,kbl,glk,cfl */
347 	/* WaEnableSamplerGPGPUPreemptionSupport:skl,bxt,kbl,cfl */
348 	wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
349 		     GEN9_ENABLE_YV12_BUGFIX |
350 		     GEN9_ENABLE_GPGPU_PREEMPTION);
351 
352 	/* Wa4x4STCOptimizationDisable:skl,bxt,kbl,glk,cfl */
353 	/* WaDisablePartialResolveInVc:skl,bxt,kbl,cfl */
354 	wa_masked_en(wal, CACHE_MODE_1,
355 		     GEN8_4x4_STC_OPTIMIZATION_DISABLE |
356 		     GEN9_PARTIAL_RESOLVE_IN_VC_DISABLE);
357 
358 	/* WaCcsTlbPrefetchDisable:skl,bxt,kbl,glk,cfl */
359 	wa_masked_dis(wal, GEN9_HALF_SLICE_CHICKEN5,
360 		      GEN9_CCS_TLB_PREFETCH_ENABLE);
361 
362 	/* WaForceContextSaveRestoreNonCoherent:skl,bxt,kbl,cfl */
363 	wa_masked_en(wal, HDC_CHICKEN0,
364 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
365 		     HDC_FORCE_CSR_NON_COHERENT_OVR_DISABLE);
366 
367 	/* WaForceEnableNonCoherent and WaDisableHDCInvalidation are
368 	 * both tied to WaForceContextSaveRestoreNonCoherent
369 	 * in some hsds for skl. We keep the tie for all gen9. The
370 	 * documentation is a bit hazy and so we want to get common behaviour,
371 	 * even though there is no clear evidence we would need both on kbl/bxt.
372 	 * This area has been source of system hangs so we play it safe
373 	 * and mimic the skl regardless of what bspec says.
374 	 *
375 	 * Use Force Non-Coherent whenever executing a 3D context. This
376 	 * is a workaround for a possible hang in the unlikely event
377 	 * a TLB invalidation occurs during a PSD flush.
378 	 */
379 
380 	/* WaForceEnableNonCoherent:skl,bxt,kbl,cfl */
381 	wa_masked_en(wal, HDC_CHICKEN0,
382 		     HDC_FORCE_NON_COHERENT);
383 
384 	/* WaDisableSamplerPowerBypassForSOPingPong:skl,bxt,kbl,cfl */
385 	if (IS_SKYLAKE(i915) ||
386 	    IS_KABYLAKE(i915) ||
387 	    IS_COFFEELAKE(i915) ||
388 	    IS_COMETLAKE(i915))
389 		wa_masked_en(wal, HALF_SLICE_CHICKEN3,
390 			     GEN8_SAMPLER_POWER_BYPASS_DIS);
391 
392 	/* WaDisableSTUnitPowerOptimization:skl,bxt,kbl,glk,cfl */
393 	wa_masked_en(wal, HALF_SLICE_CHICKEN2, GEN8_ST_PO_DISABLE);
394 
395 	/*
396 	 * Supporting preemption with fine-granularity requires changes in the
397 	 * batch buffer programming. Since we can't break old userspace, we
398 	 * need to set our default preemption level to safe value. Userspace is
399 	 * still able to use more fine-grained preemption levels, since in
400 	 * WaEnablePreemptionGranularityControlByUMD we're whitelisting the
401 	 * per-ctx register. As such, WaDisable{3D,GPGPU}MidCmdPreemption are
402 	 * not real HW workarounds, but merely a way to start using preemption
403 	 * while maintaining old contract with userspace.
404 	 */
405 
406 	/* WaDisable3DMidCmdPreemption:skl,bxt,glk,cfl,[cnl] */
407 	wa_masked_dis(wal, GEN8_CS_CHICKEN1, GEN9_PREEMPT_3D_OBJECT_LEVEL);
408 
409 	/* WaDisableGPGPUMidCmdPreemption:skl,bxt,blk,cfl,[cnl] */
410 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
411 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
412 			    GEN9_PREEMPT_GPGPU_COMMAND_LEVEL);
413 
414 	/* WaClearHIZ_WM_CHICKEN3:bxt,glk */
415 	if (IS_GEN9_LP(i915))
416 		wa_masked_en(wal, GEN9_WM_CHICKEN3, GEN9_FACTOR_IN_CLR_VAL_HIZ);
417 }
418 
419 static void skl_tune_iz_hashing(struct intel_engine_cs *engine,
420 				struct i915_wa_list *wal)
421 {
422 	struct intel_gt *gt = engine->gt;
423 	u8 vals[3] = { 0, 0, 0 };
424 	unsigned int i;
425 
426 	for (i = 0; i < 3; i++) {
427 		u8 ss;
428 
429 		/*
430 		 * Only consider slices where one, and only one, subslice has 7
431 		 * EUs
432 		 */
433 		if (!is_power_of_2(gt->info.sseu.subslice_7eu[i]))
434 			continue;
435 
436 		/*
437 		 * subslice_7eu[i] != 0 (because of the check above) and
438 		 * ss_max == 4 (maximum number of subslices possible per slice)
439 		 *
440 		 * ->    0 <= ss <= 3;
441 		 */
442 		ss = ffs(gt->info.sseu.subslice_7eu[i]) - 1;
443 		vals[i] = 3 - ss;
444 	}
445 
446 	if (vals[0] == 0 && vals[1] == 0 && vals[2] == 0)
447 		return;
448 
449 	/* Tune IZ hashing. See intel_device_info_runtime_init() */
450 	wa_masked_field_set(wal, GEN7_GT_MODE,
451 			    GEN9_IZ_HASHING_MASK(2) |
452 			    GEN9_IZ_HASHING_MASK(1) |
453 			    GEN9_IZ_HASHING_MASK(0),
454 			    GEN9_IZ_HASHING(2, vals[2]) |
455 			    GEN9_IZ_HASHING(1, vals[1]) |
456 			    GEN9_IZ_HASHING(0, vals[0]));
457 }
458 
459 static void skl_ctx_workarounds_init(struct intel_engine_cs *engine,
460 				     struct i915_wa_list *wal)
461 {
462 	gen9_ctx_workarounds_init(engine, wal);
463 	skl_tune_iz_hashing(engine, wal);
464 }
465 
466 static void bxt_ctx_workarounds_init(struct intel_engine_cs *engine,
467 				     struct i915_wa_list *wal)
468 {
469 	gen9_ctx_workarounds_init(engine, wal);
470 
471 	/* WaDisableThreadStallDopClockGating:bxt */
472 	wa_masked_en(wal, GEN8_ROW_CHICKEN,
473 		     STALL_DOP_GATING_DISABLE);
474 
475 	/* WaToEnableHwFixForPushConstHWBug:bxt */
476 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
477 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
478 }
479 
480 static void kbl_ctx_workarounds_init(struct intel_engine_cs *engine,
481 				     struct i915_wa_list *wal)
482 {
483 	struct drm_i915_private *i915 = engine->i915;
484 
485 	gen9_ctx_workarounds_init(engine, wal);
486 
487 	/* WaToEnableHwFixForPushConstHWBug:kbl */
488 	if (IS_KBL_GRAPHICS_STEP(i915, STEP_C0, STEP_FOREVER))
489 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
490 			     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
491 
492 	/* WaDisableSbeCacheDispatchPortSharing:kbl */
493 	wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
494 		     GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
495 }
496 
497 static void glk_ctx_workarounds_init(struct intel_engine_cs *engine,
498 				     struct i915_wa_list *wal)
499 {
500 	gen9_ctx_workarounds_init(engine, wal);
501 
502 	/* WaToEnableHwFixForPushConstHWBug:glk */
503 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
504 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
505 }
506 
507 static void cfl_ctx_workarounds_init(struct intel_engine_cs *engine,
508 				     struct i915_wa_list *wal)
509 {
510 	gen9_ctx_workarounds_init(engine, wal);
511 
512 	/* WaToEnableHwFixForPushConstHWBug:cfl */
513 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
514 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
515 
516 	/* WaDisableSbeCacheDispatchPortSharing:cfl */
517 	wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
518 		     GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
519 }
520 
521 static void icl_ctx_workarounds_init(struct intel_engine_cs *engine,
522 				     struct i915_wa_list *wal)
523 {
524 	/* Wa_1406697149 (WaDisableBankHangMode:icl) */
525 	wa_write(wal,
526 		 GEN8_L3CNTLREG,
527 		 intel_uncore_read(engine->uncore, GEN8_L3CNTLREG) |
528 		 GEN8_ERRDETBCTRL);
529 
530 	/* WaForceEnableNonCoherent:icl
531 	 * This is not the same workaround as in early Gen9 platforms, where
532 	 * lacking this could cause system hangs, but coherency performance
533 	 * overhead is high and only a few compute workloads really need it
534 	 * (the register is whitelisted in hardware now, so UMDs can opt in
535 	 * for coherency if they have a good reason).
536 	 */
537 	wa_masked_en(wal, ICL_HDC_MODE, HDC_FORCE_NON_COHERENT);
538 
539 	/* WaEnableFloatBlendOptimization:icl */
540 	wa_add(wal, GEN10_CACHE_MODE_SS, 0,
541 	       _MASKED_BIT_ENABLE(FLOAT_BLEND_OPTIMIZATION_ENABLE),
542 	       0 /* write-only, so skip validation */,
543 	       true);
544 
545 	/* WaDisableGPGPUMidThreadPreemption:icl */
546 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
547 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
548 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
549 
550 	/* allow headerless messages for preemptible GPGPU context */
551 	wa_masked_en(wal, GEN10_SAMPLER_MODE,
552 		     GEN11_SAMPLER_ENABLE_HEADLESS_MSG);
553 
554 	/* Wa_1604278689:icl,ehl */
555 	wa_write(wal, IVB_FBC_RT_BASE, 0xFFFFFFFF & ~ILK_FBC_RT_VALID);
556 	wa_write_clr_set(wal, IVB_FBC_RT_BASE_UPPER,
557 			 0, /* write-only register; skip validation */
558 			 0xFFFFFFFF);
559 
560 	/* Wa_1406306137:icl,ehl */
561 	wa_masked_en(wal, GEN9_ROW_CHICKEN4, GEN11_DIS_PICK_2ND_EU);
562 }
563 
564 /*
565  * These settings aren't actually workarounds, but general tuning settings that
566  * need to be programmed on dg2 platform.
567  */
568 static void dg2_ctx_gt_tuning_init(struct intel_engine_cs *engine,
569 				   struct i915_wa_list *wal)
570 {
571 	wa_masked_en(wal, CHICKEN_RASTER_2, TBIMR_FAST_CLIP);
572 	wa_write_clr_set(wal, GEN11_L3SQCREG5, L3_PWM_TIMER_INIT_VAL_MASK,
573 			 REG_FIELD_PREP(L3_PWM_TIMER_INIT_VAL_MASK, 0x7f));
574 	wa_add(wal,
575 	       FF_MODE2,
576 	       FF_MODE2_TDS_TIMER_MASK,
577 	       FF_MODE2_TDS_TIMER_128,
578 	       0, false);
579 }
580 
581 /*
582  * These settings aren't actually workarounds, but general tuning settings that
583  * need to be programmed on several platforms.
584  */
585 static void gen12_ctx_gt_tuning_init(struct intel_engine_cs *engine,
586 				     struct i915_wa_list *wal)
587 {
588 	/*
589 	 * Although some platforms refer to it as Wa_1604555607, we need to
590 	 * program it even on those that don't explicitly list that
591 	 * workaround.
592 	 *
593 	 * Note that the programming of this register is further modified
594 	 * according to the FF_MODE2 guidance given by Wa_1608008084:gen12.
595 	 * Wa_1608008084 tells us the FF_MODE2 register will return the wrong
596 	 * value when read. The default value for this register is zero for all
597 	 * fields and there are no bit masks. So instead of doing a RMW we
598 	 * should just write TDS timer value. For the same reason read
599 	 * verification is ignored.
600 	 */
601 	wa_add(wal,
602 	       FF_MODE2,
603 	       FF_MODE2_TDS_TIMER_MASK,
604 	       FF_MODE2_TDS_TIMER_128,
605 	       0, false);
606 }
607 
608 static void gen12_ctx_workarounds_init(struct intel_engine_cs *engine,
609 				       struct i915_wa_list *wal)
610 {
611 	gen12_ctx_gt_tuning_init(engine, wal);
612 
613 	/*
614 	 * Wa_1409142259:tgl,dg1,adl-p
615 	 * Wa_1409347922:tgl,dg1,adl-p
616 	 * Wa_1409252684:tgl,dg1,adl-p
617 	 * Wa_1409217633:tgl,dg1,adl-p
618 	 * Wa_1409207793:tgl,dg1,adl-p
619 	 * Wa_1409178076:tgl,dg1,adl-p
620 	 * Wa_1408979724:tgl,dg1,adl-p
621 	 * Wa_14010443199:tgl,rkl,dg1,adl-p
622 	 * Wa_14010698770:tgl,rkl,dg1,adl-s,adl-p
623 	 * Wa_1409342910:tgl,rkl,dg1,adl-s,adl-p
624 	 */
625 	wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
626 		     GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
627 
628 	/* WaDisableGPGPUMidThreadPreemption:gen12 */
629 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
630 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
631 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
632 
633 	/*
634 	 * Wa_16011163337
635 	 *
636 	 * Like in gen12_ctx_gt_tuning_init(), read verification is ignored due
637 	 * to Wa_1608008084.
638 	 */
639 	wa_add(wal,
640 	       FF_MODE2,
641 	       FF_MODE2_GS_TIMER_MASK,
642 	       FF_MODE2_GS_TIMER_224,
643 	       0, false);
644 }
645 
646 static void dg1_ctx_workarounds_init(struct intel_engine_cs *engine,
647 				     struct i915_wa_list *wal)
648 {
649 	gen12_ctx_workarounds_init(engine, wal);
650 
651 	/* Wa_1409044764 */
652 	wa_masked_dis(wal, GEN11_COMMON_SLICE_CHICKEN3,
653 		      DG1_FLOAT_POINT_BLEND_OPT_STRICT_MODE_EN);
654 
655 	/* Wa_22010493298 */
656 	wa_masked_en(wal, HIZ_CHICKEN,
657 		     DG1_HZ_READ_SUPPRESSION_OPTIMIZATION_DISABLE);
658 }
659 
660 static void dg2_ctx_workarounds_init(struct intel_engine_cs *engine,
661 				     struct i915_wa_list *wal)
662 {
663 	dg2_ctx_gt_tuning_init(engine, wal);
664 
665 	/* Wa_16011186671:dg2_g11 */
666 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0)) {
667 		wa_masked_dis(wal, VFLSKPD, DIS_MULT_MISS_RD_SQUASH);
668 		wa_masked_en(wal, VFLSKPD, DIS_OVER_FETCH_CACHE);
669 	}
670 
671 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0)) {
672 		/* Wa_14010469329:dg2_g10 */
673 		wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
674 			     XEHP_DUAL_SIMD8_SEQ_MERGE_DISABLE);
675 
676 		/*
677 		 * Wa_22010465075:dg2_g10
678 		 * Wa_22010613112:dg2_g10
679 		 * Wa_14010698770:dg2_g10
680 		 */
681 		wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
682 			     GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
683 	}
684 
685 	/* Wa_16013271637:dg2 */
686 	wa_masked_en(wal, SLICE_COMMON_ECO_CHICKEN1,
687 		     MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
688 
689 	/* Wa_14014947963:dg2 */
690 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_B0, STEP_FOREVER) ||
691 		IS_DG2_G11(engine->i915) || IS_DG2_G12(engine->i915))
692 		wa_masked_field_set(wal, VF_PREEMPTION, PREEMPTION_VERTEX_COUNT, 0x4000);
693 
694 	/* Wa_15010599737:dg2 */
695 	wa_masked_en(wal, CHICKEN_RASTER_1, DIS_SF_ROUND_NEAREST_EVEN);
696 }
697 
698 static void fakewa_disable_nestedbb_mode(struct intel_engine_cs *engine,
699 					 struct i915_wa_list *wal)
700 {
701 	/*
702 	 * This is a "fake" workaround defined by software to ensure we
703 	 * maintain reliable, backward-compatible behavior for userspace with
704 	 * regards to how nested MI_BATCH_BUFFER_START commands are handled.
705 	 *
706 	 * The per-context setting of MI_MODE[12] determines whether the bits
707 	 * of a nested MI_BATCH_BUFFER_START instruction should be interpreted
708 	 * in the traditional manner or whether they should instead use a new
709 	 * tgl+ meaning that breaks backward compatibility, but allows nesting
710 	 * into 3rd-level batchbuffers.  When this new capability was first
711 	 * added in TGL, it remained off by default unless a context
712 	 * intentionally opted in to the new behavior.  However Xe_HPG now
713 	 * flips this on by default and requires that we explicitly opt out if
714 	 * we don't want the new behavior.
715 	 *
716 	 * From a SW perspective, we want to maintain the backward-compatible
717 	 * behavior for userspace, so we'll apply a fake workaround to set it
718 	 * back to the legacy behavior on platforms where the hardware default
719 	 * is to break compatibility.  At the moment there is no Linux
720 	 * userspace that utilizes third-level batchbuffers, so this will avoid
721 	 * userspace from needing to make any changes.  using the legacy
722 	 * meaning is the correct thing to do.  If/when we have userspace
723 	 * consumers that want to utilize third-level batch nesting, we can
724 	 * provide a context parameter to allow them to opt-in.
725 	 */
726 	wa_masked_dis(wal, RING_MI_MODE(engine->mmio_base), TGL_NESTED_BB_EN);
727 }
728 
729 static void gen12_ctx_gt_mocs_init(struct intel_engine_cs *engine,
730 				   struct i915_wa_list *wal)
731 {
732 	u8 mocs;
733 
734 	/*
735 	 * Some blitter commands do not have a field for MOCS, those
736 	 * commands will use MOCS index pointed by BLIT_CCTL.
737 	 * BLIT_CCTL registers are needed to be programmed to un-cached.
738 	 */
739 	if (engine->class == COPY_ENGINE_CLASS) {
740 		mocs = engine->gt->mocs.uc_index;
741 		wa_write_clr_set(wal,
742 				 BLIT_CCTL(engine->mmio_base),
743 				 BLIT_CCTL_MASK,
744 				 BLIT_CCTL_MOCS(mocs, mocs));
745 	}
746 }
747 
748 /*
749  * gen12_ctx_gt_fake_wa_init() aren't programmingan official workaround
750  * defined by the hardware team, but it programming general context registers.
751  * Adding those context register programming in context workaround
752  * allow us to use the wa framework for proper application and validation.
753  */
754 static void
755 gen12_ctx_gt_fake_wa_init(struct intel_engine_cs *engine,
756 			  struct i915_wa_list *wal)
757 {
758 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
759 		fakewa_disable_nestedbb_mode(engine, wal);
760 
761 	gen12_ctx_gt_mocs_init(engine, wal);
762 }
763 
764 static void
765 __intel_engine_init_ctx_wa(struct intel_engine_cs *engine,
766 			   struct i915_wa_list *wal,
767 			   const char *name)
768 {
769 	struct drm_i915_private *i915 = engine->i915;
770 
771 	wa_init_start(wal, name, engine->name);
772 
773 	/* Applies to all engines */
774 	/*
775 	 * Fake workarounds are not the actual workaround but
776 	 * programming of context registers using workaround framework.
777 	 */
778 	if (GRAPHICS_VER(i915) >= 12)
779 		gen12_ctx_gt_fake_wa_init(engine, wal);
780 
781 	if (engine->class != RENDER_CLASS)
782 		goto done;
783 
784 	if (IS_PONTEVECCHIO(i915))
785 		; /* noop; none at this time */
786 	else if (IS_DG2(i915))
787 		dg2_ctx_workarounds_init(engine, wal);
788 	else if (IS_XEHPSDV(i915))
789 		; /* noop; none at this time */
790 	else if (IS_DG1(i915))
791 		dg1_ctx_workarounds_init(engine, wal);
792 	else if (GRAPHICS_VER(i915) == 12)
793 		gen12_ctx_workarounds_init(engine, wal);
794 	else if (GRAPHICS_VER(i915) == 11)
795 		icl_ctx_workarounds_init(engine, wal);
796 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
797 		cfl_ctx_workarounds_init(engine, wal);
798 	else if (IS_GEMINILAKE(i915))
799 		glk_ctx_workarounds_init(engine, wal);
800 	else if (IS_KABYLAKE(i915))
801 		kbl_ctx_workarounds_init(engine, wal);
802 	else if (IS_BROXTON(i915))
803 		bxt_ctx_workarounds_init(engine, wal);
804 	else if (IS_SKYLAKE(i915))
805 		skl_ctx_workarounds_init(engine, wal);
806 	else if (IS_CHERRYVIEW(i915))
807 		chv_ctx_workarounds_init(engine, wal);
808 	else if (IS_BROADWELL(i915))
809 		bdw_ctx_workarounds_init(engine, wal);
810 	else if (GRAPHICS_VER(i915) == 7)
811 		gen7_ctx_workarounds_init(engine, wal);
812 	else if (GRAPHICS_VER(i915) == 6)
813 		gen6_ctx_workarounds_init(engine, wal);
814 	else if (GRAPHICS_VER(i915) < 8)
815 		;
816 	else
817 		MISSING_CASE(GRAPHICS_VER(i915));
818 
819 done:
820 	wa_init_finish(wal);
821 }
822 
823 void intel_engine_init_ctx_wa(struct intel_engine_cs *engine)
824 {
825 	__intel_engine_init_ctx_wa(engine, &engine->ctx_wa_list, "context");
826 }
827 
828 int intel_engine_emit_ctx_wa(struct i915_request *rq)
829 {
830 	struct i915_wa_list *wal = &rq->engine->ctx_wa_list;
831 	struct i915_wa *wa;
832 	unsigned int i;
833 	u32 *cs;
834 	int ret;
835 
836 	if (wal->count == 0)
837 		return 0;
838 
839 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
840 	if (ret)
841 		return ret;
842 
843 	cs = intel_ring_begin(rq, (wal->count * 2 + 2));
844 	if (IS_ERR(cs))
845 		return PTR_ERR(cs);
846 
847 	*cs++ = MI_LOAD_REGISTER_IMM(wal->count);
848 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
849 		*cs++ = i915_mmio_reg_offset(wa->reg);
850 		*cs++ = wa->set;
851 	}
852 	*cs++ = MI_NOOP;
853 
854 	intel_ring_advance(rq, cs);
855 
856 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
857 	if (ret)
858 		return ret;
859 
860 	return 0;
861 }
862 
863 static void
864 gen4_gt_workarounds_init(struct intel_gt *gt,
865 			 struct i915_wa_list *wal)
866 {
867 	/* WaDisable_RenderCache_OperationalFlush:gen4,ilk */
868 	wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
869 }
870 
871 static void
872 g4x_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
873 {
874 	gen4_gt_workarounds_init(gt, wal);
875 
876 	/* WaDisableRenderCachePipelinedFlush:g4x,ilk */
877 	wa_masked_en(wal, CACHE_MODE_0, CM0_PIPELINED_RENDER_FLUSH_DISABLE);
878 }
879 
880 static void
881 ilk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
882 {
883 	g4x_gt_workarounds_init(gt, wal);
884 
885 	wa_masked_en(wal, _3D_CHICKEN2, _3D_CHICKEN2_WM_READ_PIPELINED);
886 }
887 
888 static void
889 snb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
890 {
891 }
892 
893 static void
894 ivb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
895 {
896 	/* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
897 	wa_masked_dis(wal,
898 		      GEN7_COMMON_SLICE_CHICKEN1,
899 		      GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
900 
901 	/* WaApplyL3ControlAndL3ChickenMode:ivb */
902 	wa_write(wal, GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
903 	wa_write(wal, GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
904 
905 	/* WaForceL3Serialization:ivb */
906 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
907 }
908 
909 static void
910 vlv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
911 {
912 	/* WaForceL3Serialization:vlv */
913 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
914 
915 	/*
916 	 * WaIncreaseL3CreditsForVLVB0:vlv
917 	 * This is the hardware default actually.
918 	 */
919 	wa_write(wal, GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
920 }
921 
922 static void
923 hsw_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
924 {
925 	/* L3 caching of data atomics doesn't work -- disable it. */
926 	wa_write(wal, HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
927 
928 	wa_add(wal,
929 	       HSW_ROW_CHICKEN3, 0,
930 	       _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
931 	       0 /* XXX does this reg exist? */, true);
932 
933 	/* WaVSRefCountFullforceMissDisable:hsw */
934 	wa_write_clr(wal, GEN7_FF_THREAD_MODE, GEN7_FF_VS_REF_CNT_FFME);
935 }
936 
937 static void
938 gen9_wa_init_mcr(struct drm_i915_private *i915, struct i915_wa_list *wal)
939 {
940 	const struct sseu_dev_info *sseu = &to_gt(i915)->info.sseu;
941 	unsigned int slice, subslice;
942 	u32 mcr, mcr_mask;
943 
944 	GEM_BUG_ON(GRAPHICS_VER(i915) != 9);
945 
946 	/*
947 	 * WaProgramMgsrForCorrectSliceSpecificMmioReads:gen9,glk,kbl,cml
948 	 * Before any MMIO read into slice/subslice specific registers, MCR
949 	 * packet control register needs to be programmed to point to any
950 	 * enabled s/ss pair. Otherwise, incorrect values will be returned.
951 	 * This means each subsequent MMIO read will be forwarded to an
952 	 * specific s/ss combination, but this is OK since these registers
953 	 * are consistent across s/ss in almost all cases. In the rare
954 	 * occasions, such as INSTDONE, where this value is dependent
955 	 * on s/ss combo, the read should be done with read_subslice_reg.
956 	 */
957 	slice = ffs(sseu->slice_mask) - 1;
958 	GEM_BUG_ON(slice >= ARRAY_SIZE(sseu->subslice_mask.hsw));
959 	subslice = ffs(intel_sseu_get_hsw_subslices(sseu, slice));
960 	GEM_BUG_ON(!subslice);
961 	subslice--;
962 
963 	/*
964 	 * We use GEN8_MCR..() macros to calculate the |mcr| value for
965 	 * Gen9 to address WaProgramMgsrForCorrectSliceSpecificMmioReads
966 	 */
967 	mcr = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
968 	mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
969 
970 	drm_dbg(&i915->drm, "MCR slice:%d/subslice:%d = %x\n", slice, subslice, mcr);
971 
972 	wa_write_clr_set(wal, GEN8_MCR_SELECTOR, mcr_mask, mcr);
973 }
974 
975 static void
976 gen9_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
977 {
978 	struct drm_i915_private *i915 = gt->i915;
979 
980 	/* WaProgramMgsrForCorrectSliceSpecificMmioReads:glk,kbl,cml,gen9 */
981 	gen9_wa_init_mcr(i915, wal);
982 
983 	/* WaDisableKillLogic:bxt,skl,kbl */
984 	if (!IS_COFFEELAKE(i915) && !IS_COMETLAKE(i915))
985 		wa_write_or(wal,
986 			    GAM_ECOCHK,
987 			    ECOCHK_DIS_TLB);
988 
989 	if (HAS_LLC(i915)) {
990 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
991 		 *
992 		 * Must match Display Engine. See
993 		 * WaCompressedResourceDisplayNewHashMode.
994 		 */
995 		wa_write_or(wal,
996 			    MMCD_MISC_CTRL,
997 			    MMCD_PCLA | MMCD_HOTSPOT_EN);
998 	}
999 
1000 	/* WaDisableHDCInvalidation:skl,bxt,kbl,cfl */
1001 	wa_write_or(wal,
1002 		    GAM_ECOCHK,
1003 		    BDW_DISABLE_HDC_INVALIDATION);
1004 }
1005 
1006 static void
1007 skl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1008 {
1009 	gen9_gt_workarounds_init(gt, wal);
1010 
1011 	/* WaDisableGafsUnitClkGating:skl */
1012 	wa_write_or(wal,
1013 		    GEN7_UCGCTL4,
1014 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1015 
1016 	/* WaInPlaceDecompressionHang:skl */
1017 	if (IS_SKL_GRAPHICS_STEP(gt->i915, STEP_A0, STEP_H0))
1018 		wa_write_or(wal,
1019 			    GEN9_GAMT_ECO_REG_RW_IA,
1020 			    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1021 }
1022 
1023 static void
1024 kbl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1025 {
1026 	gen9_gt_workarounds_init(gt, wal);
1027 
1028 	/* WaDisableDynamicCreditSharing:kbl */
1029 	if (IS_KBL_GRAPHICS_STEP(gt->i915, 0, STEP_C0))
1030 		wa_write_or(wal,
1031 			    GAMT_CHKN_BIT_REG,
1032 			    GAMT_CHKN_DISABLE_DYNAMIC_CREDIT_SHARING);
1033 
1034 	/* WaDisableGafsUnitClkGating:kbl */
1035 	wa_write_or(wal,
1036 		    GEN7_UCGCTL4,
1037 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1038 
1039 	/* WaInPlaceDecompressionHang:kbl */
1040 	wa_write_or(wal,
1041 		    GEN9_GAMT_ECO_REG_RW_IA,
1042 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1043 }
1044 
1045 static void
1046 glk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1047 {
1048 	gen9_gt_workarounds_init(gt, wal);
1049 }
1050 
1051 static void
1052 cfl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1053 {
1054 	gen9_gt_workarounds_init(gt, wal);
1055 
1056 	/* WaDisableGafsUnitClkGating:cfl */
1057 	wa_write_or(wal,
1058 		    GEN7_UCGCTL4,
1059 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1060 
1061 	/* WaInPlaceDecompressionHang:cfl */
1062 	wa_write_or(wal,
1063 		    GEN9_GAMT_ECO_REG_RW_IA,
1064 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1065 }
1066 
1067 static void __set_mcr_steering(struct i915_wa_list *wal,
1068 			       i915_reg_t steering_reg,
1069 			       unsigned int slice, unsigned int subslice)
1070 {
1071 	u32 mcr, mcr_mask;
1072 
1073 	mcr = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1074 	mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1075 
1076 	wa_write_clr_set(wal, steering_reg, mcr_mask, mcr);
1077 }
1078 
1079 static void __add_mcr_wa(struct intel_gt *gt, struct i915_wa_list *wal,
1080 			 unsigned int slice, unsigned int subslice)
1081 {
1082 	struct drm_printer p = drm_debug_printer("MCR Steering:");
1083 
1084 	__set_mcr_steering(wal, GEN8_MCR_SELECTOR, slice, subslice);
1085 
1086 	gt->default_steering.groupid = slice;
1087 	gt->default_steering.instanceid = subslice;
1088 
1089 	if (drm_debug_enabled(DRM_UT_DRIVER))
1090 		intel_gt_mcr_report_steering(&p, gt, false);
1091 }
1092 
1093 static void
1094 icl_wa_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1095 {
1096 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1097 	unsigned int subslice;
1098 
1099 	GEM_BUG_ON(GRAPHICS_VER(gt->i915) < 11);
1100 	GEM_BUG_ON(hweight8(sseu->slice_mask) > 1);
1101 
1102 	/*
1103 	 * Although a platform may have subslices, we need to always steer
1104 	 * reads to the lowest instance that isn't fused off.  When Render
1105 	 * Power Gating is enabled, grabbing forcewake will only power up a
1106 	 * single subslice (the "minconfig") if there isn't a real workload
1107 	 * that needs to be run; this means that if we steer register reads to
1108 	 * one of the higher subslices, we run the risk of reading back 0's or
1109 	 * random garbage.
1110 	 */
1111 	subslice = __ffs(intel_sseu_get_hsw_subslices(sseu, 0));
1112 
1113 	/*
1114 	 * If the subslice we picked above also steers us to a valid L3 bank,
1115 	 * then we can just rely on the default steering and won't need to
1116 	 * worry about explicitly re-steering L3BANK reads later.
1117 	 */
1118 	if (gt->info.l3bank_mask & BIT(subslice))
1119 		gt->steering_table[L3BANK] = NULL;
1120 
1121 	__add_mcr_wa(gt, wal, 0, subslice);
1122 }
1123 
1124 static void
1125 xehp_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1126 {
1127 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1128 	unsigned long slice, subslice = 0, slice_mask = 0;
1129 	u32 lncf_mask = 0;
1130 	int i;
1131 
1132 	/*
1133 	 * On Xe_HP the steering increases in complexity. There are now several
1134 	 * more units that require steering and we're not guaranteed to be able
1135 	 * to find a common setting for all of them. These are:
1136 	 * - GSLICE (fusable)
1137 	 * - DSS (sub-unit within gslice; fusable)
1138 	 * - L3 Bank (fusable)
1139 	 * - MSLICE (fusable)
1140 	 * - LNCF (sub-unit within mslice; always present if mslice is present)
1141 	 *
1142 	 * We'll do our default/implicit steering based on GSLICE (in the
1143 	 * sliceid field) and DSS (in the subsliceid field).  If we can
1144 	 * find overlap between the valid MSLICE and/or LNCF values with
1145 	 * a suitable GSLICE, then we can just re-use the default value and
1146 	 * skip and explicit steering at runtime.
1147 	 *
1148 	 * We only need to look for overlap between GSLICE/MSLICE/LNCF to find
1149 	 * a valid sliceid value.  DSS steering is the only type of steering
1150 	 * that utilizes the 'subsliceid' bits.
1151 	 *
1152 	 * Also note that, even though the steering domain is called "GSlice"
1153 	 * and it is encoded in the register using the gslice format, the spec
1154 	 * says that the combined (geometry | compute) fuse should be used to
1155 	 * select the steering.
1156 	 */
1157 
1158 	/* Find the potential gslice candidates */
1159 	slice_mask = intel_slicemask_from_xehp_dssmask(sseu->subslice_mask,
1160 						       GEN_DSS_PER_GSLICE);
1161 
1162 	/*
1163 	 * Find the potential LNCF candidates.  Either LNCF within a valid
1164 	 * mslice is fine.
1165 	 */
1166 	for_each_set_bit(i, &gt->info.mslice_mask, GEN12_MAX_MSLICES)
1167 		lncf_mask |= (0x3 << (i * 2));
1168 
1169 	/*
1170 	 * Are there any sliceid values that work for both GSLICE and LNCF
1171 	 * steering?
1172 	 */
1173 	if (slice_mask & lncf_mask) {
1174 		slice_mask &= lncf_mask;
1175 		gt->steering_table[LNCF] = NULL;
1176 	}
1177 
1178 	/* How about sliceid values that also work for MSLICE steering? */
1179 	if (slice_mask & gt->info.mslice_mask) {
1180 		slice_mask &= gt->info.mslice_mask;
1181 		gt->steering_table[MSLICE] = NULL;
1182 	}
1183 
1184 	slice = __ffs(slice_mask);
1185 	subslice = intel_sseu_find_first_xehp_dss(sseu, GEN_DSS_PER_GSLICE, slice) %
1186 		GEN_DSS_PER_GSLICE;
1187 
1188 	__add_mcr_wa(gt, wal, slice, subslice);
1189 
1190 	/*
1191 	 * SQIDI ranges are special because they use different steering
1192 	 * registers than everything else we work with.  On XeHP SDV and
1193 	 * DG2-G10, any value in the steering registers will work fine since
1194 	 * all instances are present, but DG2-G11 only has SQIDI instances at
1195 	 * ID's 2 and 3, so we need to steer to one of those.  For simplicity
1196 	 * we'll just steer to a hardcoded "2" since that value will work
1197 	 * everywhere.
1198 	 */
1199 	__set_mcr_steering(wal, MCFG_MCR_SELECTOR, 0, 2);
1200 	__set_mcr_steering(wal, SF_MCR_SELECTOR, 0, 2);
1201 }
1202 
1203 static void
1204 pvc_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1205 {
1206 	unsigned int dss;
1207 
1208 	/*
1209 	 * Setup implicit steering for COMPUTE and DSS ranges to the first
1210 	 * non-fused-off DSS.  All other types of MCR registers will be
1211 	 * explicitly steered.
1212 	 */
1213 	dss = intel_sseu_find_first_xehp_dss(&gt->info.sseu, 0, 0);
1214 	__add_mcr_wa(gt, wal, dss / GEN_DSS_PER_CSLICE, dss % GEN_DSS_PER_CSLICE);
1215 }
1216 
1217 static void
1218 icl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1219 {
1220 	struct drm_i915_private *i915 = gt->i915;
1221 
1222 	icl_wa_init_mcr(gt, wal);
1223 
1224 	/* WaModifyGamTlbPartitioning:icl */
1225 	wa_write_clr_set(wal,
1226 			 GEN11_GACB_PERF_CTRL,
1227 			 GEN11_HASH_CTRL_MASK,
1228 			 GEN11_HASH_CTRL_BIT0 | GEN11_HASH_CTRL_BIT4);
1229 
1230 	/* Wa_1405766107:icl
1231 	 * Formerly known as WaCL2SFHalfMaxAlloc
1232 	 */
1233 	wa_write_or(wal,
1234 		    GEN11_LSN_UNSLCVC,
1235 		    GEN11_LSN_UNSLCVC_GAFS_HALF_SF_MAXALLOC |
1236 		    GEN11_LSN_UNSLCVC_GAFS_HALF_CL2_MAXALLOC);
1237 
1238 	/* Wa_220166154:icl
1239 	 * Formerly known as WaDisCtxReload
1240 	 */
1241 	wa_write_or(wal,
1242 		    GEN8_GAMW_ECO_DEV_RW_IA,
1243 		    GAMW_ECO_DEV_CTX_RELOAD_DISABLE);
1244 
1245 	/* Wa_1406463099:icl
1246 	 * Formerly known as WaGamTlbPendError
1247 	 */
1248 	wa_write_or(wal,
1249 		    GAMT_CHKN_BIT_REG,
1250 		    GAMT_CHKN_DISABLE_L3_COH_PIPE);
1251 
1252 	/* Wa_1407352427:icl,ehl */
1253 	wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1254 		    PSDUNIT_CLKGATE_DIS);
1255 
1256 	/* Wa_1406680159:icl,ehl */
1257 	wa_write_or(wal,
1258 		    SUBSLICE_UNIT_LEVEL_CLKGATE,
1259 		    GWUNIT_CLKGATE_DIS);
1260 
1261 	/* Wa_1607087056:icl,ehl,jsl */
1262 	if (IS_ICELAKE(i915) ||
1263 	    IS_JSL_EHL_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1264 		wa_write_or(wal,
1265 			    SLICE_UNIT_LEVEL_CLKGATE,
1266 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1267 
1268 	/*
1269 	 * This is not a documented workaround, but rather an optimization
1270 	 * to reduce sampler power.
1271 	 */
1272 	wa_write_clr(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1273 }
1274 
1275 /*
1276  * Though there are per-engine instances of these registers,
1277  * they retain their value through engine resets and should
1278  * only be provided on the GT workaround list rather than
1279  * the engine-specific workaround list.
1280  */
1281 static void
1282 wa_14011060649(struct intel_gt *gt, struct i915_wa_list *wal)
1283 {
1284 	struct intel_engine_cs *engine;
1285 	int id;
1286 
1287 	for_each_engine(engine, gt, id) {
1288 		if (engine->class != VIDEO_DECODE_CLASS ||
1289 		    (engine->instance % 2))
1290 			continue;
1291 
1292 		wa_write_or(wal, VDBOX_CGCTL3F10(engine->mmio_base),
1293 			    IECPUNIT_CLKGATE_DIS);
1294 	}
1295 }
1296 
1297 static void
1298 gen12_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1299 {
1300 	icl_wa_init_mcr(gt, wal);
1301 
1302 	/* Wa_14011060649:tgl,rkl,dg1,adl-s,adl-p */
1303 	wa_14011060649(gt, wal);
1304 
1305 	/* Wa_14011059788:tgl,rkl,adl-s,dg1,adl-p */
1306 	wa_write_or(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1307 }
1308 
1309 static void
1310 tgl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1311 {
1312 	struct drm_i915_private *i915 = gt->i915;
1313 
1314 	gen12_gt_workarounds_init(gt, wal);
1315 
1316 	/* Wa_1409420604:tgl */
1317 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1318 		wa_write_or(wal,
1319 			    SUBSLICE_UNIT_LEVEL_CLKGATE2,
1320 			    CPSSUNIT_CLKGATE_DIS);
1321 
1322 	/* Wa_1607087056:tgl also know as BUG:1409180338 */
1323 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1324 		wa_write_or(wal,
1325 			    SLICE_UNIT_LEVEL_CLKGATE,
1326 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1327 
1328 	/* Wa_1408615072:tgl[a0] */
1329 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1330 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1331 			    VSUNIT_CLKGATE_DIS_TGL);
1332 }
1333 
1334 static void
1335 dg1_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1336 {
1337 	struct drm_i915_private *i915 = gt->i915;
1338 
1339 	gen12_gt_workarounds_init(gt, wal);
1340 
1341 	/* Wa_1607087056:dg1 */
1342 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1343 		wa_write_or(wal,
1344 			    SLICE_UNIT_LEVEL_CLKGATE,
1345 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1346 
1347 	/* Wa_1409420604:dg1 */
1348 	if (IS_DG1(i915))
1349 		wa_write_or(wal,
1350 			    SUBSLICE_UNIT_LEVEL_CLKGATE2,
1351 			    CPSSUNIT_CLKGATE_DIS);
1352 
1353 	/* Wa_1408615072:dg1 */
1354 	/* Empirical testing shows this register is unaffected by engine reset. */
1355 	if (IS_DG1(i915))
1356 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1357 			    VSUNIT_CLKGATE_DIS_TGL);
1358 }
1359 
1360 static void
1361 xehpsdv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1362 {
1363 	struct drm_i915_private *i915 = gt->i915;
1364 
1365 	xehp_init_mcr(gt, wal);
1366 
1367 	/* Wa_1409757795:xehpsdv */
1368 	wa_write_or(wal, SCCGCTL94DC, CG3DDISURB);
1369 
1370 	/* Wa_16011155590:xehpsdv */
1371 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A0, STEP_B0))
1372 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1373 			    TSGUNIT_CLKGATE_DIS);
1374 
1375 	/* Wa_14011780169:xehpsdv */
1376 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_B0, STEP_FOREVER)) {
1377 		wa_write_or(wal, UNSLCGCTL9440, GAMTLBOACS_CLKGATE_DIS |
1378 			    GAMTLBVDBOX7_CLKGATE_DIS |
1379 			    GAMTLBVDBOX6_CLKGATE_DIS |
1380 			    GAMTLBVDBOX5_CLKGATE_DIS |
1381 			    GAMTLBVDBOX4_CLKGATE_DIS |
1382 			    GAMTLBVDBOX3_CLKGATE_DIS |
1383 			    GAMTLBVDBOX2_CLKGATE_DIS |
1384 			    GAMTLBVDBOX1_CLKGATE_DIS |
1385 			    GAMTLBVDBOX0_CLKGATE_DIS |
1386 			    GAMTLBKCR_CLKGATE_DIS |
1387 			    GAMTLBGUC_CLKGATE_DIS |
1388 			    GAMTLBBLT_CLKGATE_DIS);
1389 		wa_write_or(wal, UNSLCGCTL9444, GAMTLBGFXA0_CLKGATE_DIS |
1390 			    GAMTLBGFXA1_CLKGATE_DIS |
1391 			    GAMTLBCOMPA0_CLKGATE_DIS |
1392 			    GAMTLBCOMPA1_CLKGATE_DIS |
1393 			    GAMTLBCOMPB0_CLKGATE_DIS |
1394 			    GAMTLBCOMPB1_CLKGATE_DIS |
1395 			    GAMTLBCOMPC0_CLKGATE_DIS |
1396 			    GAMTLBCOMPC1_CLKGATE_DIS |
1397 			    GAMTLBCOMPD0_CLKGATE_DIS |
1398 			    GAMTLBCOMPD1_CLKGATE_DIS |
1399 			    GAMTLBMERT_CLKGATE_DIS   |
1400 			    GAMTLBVEBOX3_CLKGATE_DIS |
1401 			    GAMTLBVEBOX2_CLKGATE_DIS |
1402 			    GAMTLBVEBOX1_CLKGATE_DIS |
1403 			    GAMTLBVEBOX0_CLKGATE_DIS);
1404 	}
1405 
1406 	/* Wa_16012725990:xehpsdv */
1407 	if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A1, STEP_FOREVER))
1408 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE, VFUNIT_CLKGATE_DIS);
1409 
1410 	/* Wa_14011060649:xehpsdv */
1411 	wa_14011060649(gt, wal);
1412 }
1413 
1414 static void
1415 dg2_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1416 {
1417 	struct intel_engine_cs *engine;
1418 	int id;
1419 
1420 	xehp_init_mcr(gt, wal);
1421 
1422 	/* Wa_14011060649:dg2 */
1423 	wa_14011060649(gt, wal);
1424 
1425 	/*
1426 	 * Although there are per-engine instances of these registers,
1427 	 * they technically exist outside the engine itself and are not
1428 	 * impacted by engine resets.  Furthermore, they're part of the
1429 	 * GuC blacklist so trying to treat them as engine workarounds
1430 	 * will result in GuC initialization failure and a wedged GPU.
1431 	 */
1432 	for_each_engine(engine, gt, id) {
1433 		if (engine->class != VIDEO_DECODE_CLASS)
1434 			continue;
1435 
1436 		/* Wa_16010515920:dg2_g10 */
1437 		if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_B0))
1438 			wa_write_or(wal, VDBOX_CGCTL3F18(engine->mmio_base),
1439 				    ALNUNIT_CLKGATE_DIS);
1440 	}
1441 
1442 	if (IS_DG2_G10(gt->i915)) {
1443 		/* Wa_22010523718:dg2 */
1444 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1445 			    CG3DDISCFEG_CLKGATE_DIS);
1446 
1447 		/* Wa_14011006942:dg2 */
1448 		wa_write_or(wal, SUBSLICE_UNIT_LEVEL_CLKGATE,
1449 			    DSS_ROUTER_CLKGATE_DIS);
1450 	}
1451 
1452 	if (IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_B0)) {
1453 		/* Wa_14010948348:dg2_g10 */
1454 		wa_write_or(wal, UNSLCGCTL9430, MSQDUNIT_CLKGATE_DIS);
1455 
1456 		/* Wa_14011037102:dg2_g10 */
1457 		wa_write_or(wal, UNSLCGCTL9444, LTCDD_CLKGATE_DIS);
1458 
1459 		/* Wa_14011371254:dg2_g10 */
1460 		wa_write_or(wal, SLICE_UNIT_LEVEL_CLKGATE, NODEDSS_CLKGATE_DIS);
1461 
1462 		/* Wa_14011431319:dg2_g10 */
1463 		wa_write_or(wal, UNSLCGCTL9440, GAMTLBOACS_CLKGATE_DIS |
1464 			    GAMTLBVDBOX7_CLKGATE_DIS |
1465 			    GAMTLBVDBOX6_CLKGATE_DIS |
1466 			    GAMTLBVDBOX5_CLKGATE_DIS |
1467 			    GAMTLBVDBOX4_CLKGATE_DIS |
1468 			    GAMTLBVDBOX3_CLKGATE_DIS |
1469 			    GAMTLBVDBOX2_CLKGATE_DIS |
1470 			    GAMTLBVDBOX1_CLKGATE_DIS |
1471 			    GAMTLBVDBOX0_CLKGATE_DIS |
1472 			    GAMTLBKCR_CLKGATE_DIS |
1473 			    GAMTLBGUC_CLKGATE_DIS |
1474 			    GAMTLBBLT_CLKGATE_DIS);
1475 		wa_write_or(wal, UNSLCGCTL9444, GAMTLBGFXA0_CLKGATE_DIS |
1476 			    GAMTLBGFXA1_CLKGATE_DIS |
1477 			    GAMTLBCOMPA0_CLKGATE_DIS |
1478 			    GAMTLBCOMPA1_CLKGATE_DIS |
1479 			    GAMTLBCOMPB0_CLKGATE_DIS |
1480 			    GAMTLBCOMPB1_CLKGATE_DIS |
1481 			    GAMTLBCOMPC0_CLKGATE_DIS |
1482 			    GAMTLBCOMPC1_CLKGATE_DIS |
1483 			    GAMTLBCOMPD0_CLKGATE_DIS |
1484 			    GAMTLBCOMPD1_CLKGATE_DIS |
1485 			    GAMTLBMERT_CLKGATE_DIS   |
1486 			    GAMTLBVEBOX3_CLKGATE_DIS |
1487 			    GAMTLBVEBOX2_CLKGATE_DIS |
1488 			    GAMTLBVEBOX1_CLKGATE_DIS |
1489 			    GAMTLBVEBOX0_CLKGATE_DIS);
1490 
1491 		/* Wa_14010569222:dg2_g10 */
1492 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1493 			    GAMEDIA_CLKGATE_DIS);
1494 
1495 		/* Wa_14011028019:dg2_g10 */
1496 		wa_write_or(wal, SSMCGCTL9530, RTFUNIT_CLKGATE_DIS);
1497 	}
1498 
1499 	/* Wa_14014830051:dg2 */
1500 	wa_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1501 
1502 	/*
1503 	 * The following are not actually "workarounds" but rather
1504 	 * recommended tuning settings documented in the bspec's
1505 	 * performance guide section.
1506 	 */
1507 	wa_write_or(wal, GEN12_SQCM, EN_32B_ACCESS);
1508 
1509 	/* Wa_14015795083 */
1510 	wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1511 }
1512 
1513 static void
1514 pvc_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1515 {
1516 	pvc_init_mcr(gt, wal);
1517 
1518 	/* Wa_14015795083 */
1519 	wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1520 }
1521 
1522 static void
1523 gt_init_workarounds(struct intel_gt *gt, struct i915_wa_list *wal)
1524 {
1525 	struct drm_i915_private *i915 = gt->i915;
1526 
1527 	if (IS_PONTEVECCHIO(i915))
1528 		pvc_gt_workarounds_init(gt, wal);
1529 	else if (IS_DG2(i915))
1530 		dg2_gt_workarounds_init(gt, wal);
1531 	else if (IS_XEHPSDV(i915))
1532 		xehpsdv_gt_workarounds_init(gt, wal);
1533 	else if (IS_DG1(i915))
1534 		dg1_gt_workarounds_init(gt, wal);
1535 	else if (IS_TIGERLAKE(i915))
1536 		tgl_gt_workarounds_init(gt, wal);
1537 	else if (GRAPHICS_VER(i915) == 12)
1538 		gen12_gt_workarounds_init(gt, wal);
1539 	else if (GRAPHICS_VER(i915) == 11)
1540 		icl_gt_workarounds_init(gt, wal);
1541 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
1542 		cfl_gt_workarounds_init(gt, wal);
1543 	else if (IS_GEMINILAKE(i915))
1544 		glk_gt_workarounds_init(gt, wal);
1545 	else if (IS_KABYLAKE(i915))
1546 		kbl_gt_workarounds_init(gt, wal);
1547 	else if (IS_BROXTON(i915))
1548 		gen9_gt_workarounds_init(gt, wal);
1549 	else if (IS_SKYLAKE(i915))
1550 		skl_gt_workarounds_init(gt, wal);
1551 	else if (IS_HASWELL(i915))
1552 		hsw_gt_workarounds_init(gt, wal);
1553 	else if (IS_VALLEYVIEW(i915))
1554 		vlv_gt_workarounds_init(gt, wal);
1555 	else if (IS_IVYBRIDGE(i915))
1556 		ivb_gt_workarounds_init(gt, wal);
1557 	else if (GRAPHICS_VER(i915) == 6)
1558 		snb_gt_workarounds_init(gt, wal);
1559 	else if (GRAPHICS_VER(i915) == 5)
1560 		ilk_gt_workarounds_init(gt, wal);
1561 	else if (IS_G4X(i915))
1562 		g4x_gt_workarounds_init(gt, wal);
1563 	else if (GRAPHICS_VER(i915) == 4)
1564 		gen4_gt_workarounds_init(gt, wal);
1565 	else if (GRAPHICS_VER(i915) <= 8)
1566 		;
1567 	else
1568 		MISSING_CASE(GRAPHICS_VER(i915));
1569 }
1570 
1571 void intel_gt_init_workarounds(struct intel_gt *gt)
1572 {
1573 	struct i915_wa_list *wal = &gt->wa_list;
1574 
1575 	wa_init_start(wal, "GT", "global");
1576 	gt_init_workarounds(gt, wal);
1577 	wa_init_finish(wal);
1578 }
1579 
1580 static enum forcewake_domains
1581 wal_get_fw_for_rmw(struct intel_uncore *uncore, const struct i915_wa_list *wal)
1582 {
1583 	enum forcewake_domains fw = 0;
1584 	struct i915_wa *wa;
1585 	unsigned int i;
1586 
1587 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1588 		fw |= intel_uncore_forcewake_for_reg(uncore,
1589 						     wa->reg,
1590 						     FW_REG_READ |
1591 						     FW_REG_WRITE);
1592 
1593 	return fw;
1594 }
1595 
1596 static bool
1597 wa_verify(const struct i915_wa *wa, u32 cur, const char *name, const char *from)
1598 {
1599 	if ((cur ^ wa->set) & wa->read) {
1600 		DRM_ERROR("%s workaround lost on %s! (reg[%x]=0x%x, relevant bits were 0x%x vs expected 0x%x)\n",
1601 			  name, from, i915_mmio_reg_offset(wa->reg),
1602 			  cur, cur & wa->read, wa->set & wa->read);
1603 
1604 		return false;
1605 	}
1606 
1607 	return true;
1608 }
1609 
1610 static void
1611 wa_list_apply(struct intel_gt *gt, const struct i915_wa_list *wal)
1612 {
1613 	struct intel_uncore *uncore = gt->uncore;
1614 	enum forcewake_domains fw;
1615 	unsigned long flags;
1616 	struct i915_wa *wa;
1617 	unsigned int i;
1618 
1619 	if (!wal->count)
1620 		return;
1621 
1622 	fw = wal_get_fw_for_rmw(uncore, wal);
1623 
1624 	spin_lock_irqsave(&uncore->lock, flags);
1625 	intel_uncore_forcewake_get__locked(uncore, fw);
1626 
1627 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
1628 		u32 val, old = 0;
1629 
1630 		/* open-coded rmw due to steering */
1631 		old = wa->clr ? intel_gt_mcr_read_any_fw(gt, wa->reg) : 0;
1632 		val = (old & ~wa->clr) | wa->set;
1633 		if (val != old || !wa->clr)
1634 			intel_uncore_write_fw(uncore, wa->reg, val);
1635 
1636 		if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
1637 			wa_verify(wa, intel_gt_mcr_read_any_fw(gt, wa->reg),
1638 				  wal->name, "application");
1639 	}
1640 
1641 	intel_uncore_forcewake_put__locked(uncore, fw);
1642 	spin_unlock_irqrestore(&uncore->lock, flags);
1643 }
1644 
1645 void intel_gt_apply_workarounds(struct intel_gt *gt)
1646 {
1647 	wa_list_apply(gt, &gt->wa_list);
1648 }
1649 
1650 static bool wa_list_verify(struct intel_gt *gt,
1651 			   const struct i915_wa_list *wal,
1652 			   const char *from)
1653 {
1654 	struct intel_uncore *uncore = gt->uncore;
1655 	struct i915_wa *wa;
1656 	enum forcewake_domains fw;
1657 	unsigned long flags;
1658 	unsigned int i;
1659 	bool ok = true;
1660 
1661 	fw = wal_get_fw_for_rmw(uncore, wal);
1662 
1663 	spin_lock_irqsave(&uncore->lock, flags);
1664 	intel_uncore_forcewake_get__locked(uncore, fw);
1665 
1666 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1667 		ok &= wa_verify(wa,
1668 				intel_gt_mcr_read_any_fw(gt, wa->reg),
1669 				wal->name, from);
1670 
1671 	intel_uncore_forcewake_put__locked(uncore, fw);
1672 	spin_unlock_irqrestore(&uncore->lock, flags);
1673 
1674 	return ok;
1675 }
1676 
1677 bool intel_gt_verify_workarounds(struct intel_gt *gt, const char *from)
1678 {
1679 	return wa_list_verify(gt, &gt->wa_list, from);
1680 }
1681 
1682 __maybe_unused
1683 static bool is_nonpriv_flags_valid(u32 flags)
1684 {
1685 	/* Check only valid flag bits are set */
1686 	if (flags & ~RING_FORCE_TO_NONPRIV_MASK_VALID)
1687 		return false;
1688 
1689 	/* NB: Only 3 out of 4 enum values are valid for access field */
1690 	if ((flags & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
1691 	    RING_FORCE_TO_NONPRIV_ACCESS_INVALID)
1692 		return false;
1693 
1694 	return true;
1695 }
1696 
1697 static void
1698 whitelist_reg_ext(struct i915_wa_list *wal, i915_reg_t reg, u32 flags)
1699 {
1700 	struct i915_wa wa = {
1701 		.reg = reg
1702 	};
1703 
1704 	if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1705 		return;
1706 
1707 	if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1708 		return;
1709 
1710 	wa.reg.reg |= flags;
1711 	_wa_add(wal, &wa);
1712 }
1713 
1714 static void
1715 whitelist_reg(struct i915_wa_list *wal, i915_reg_t reg)
1716 {
1717 	whitelist_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1718 }
1719 
1720 static void gen9_whitelist_build(struct i915_wa_list *w)
1721 {
1722 	/* WaVFEStateAfterPipeControlwithMediaStateClear:skl,bxt,glk,cfl */
1723 	whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1724 
1725 	/* WaEnablePreemptionGranularityControlByUMD:skl,bxt,kbl,cfl,[cnl] */
1726 	whitelist_reg(w, GEN8_CS_CHICKEN1);
1727 
1728 	/* WaAllowUMDToModifyHDCChicken1:skl,bxt,kbl,glk,cfl */
1729 	whitelist_reg(w, GEN8_HDC_CHICKEN1);
1730 
1731 	/* WaSendPushConstantsFromMMIO:skl,bxt */
1732 	whitelist_reg(w, COMMON_SLICE_CHICKEN2);
1733 }
1734 
1735 static void skl_whitelist_build(struct intel_engine_cs *engine)
1736 {
1737 	struct i915_wa_list *w = &engine->whitelist;
1738 
1739 	if (engine->class != RENDER_CLASS)
1740 		return;
1741 
1742 	gen9_whitelist_build(w);
1743 
1744 	/* WaDisableLSQCROPERFforOCL:skl */
1745 	whitelist_reg(w, GEN8_L3SQCREG4);
1746 }
1747 
1748 static void bxt_whitelist_build(struct intel_engine_cs *engine)
1749 {
1750 	if (engine->class != RENDER_CLASS)
1751 		return;
1752 
1753 	gen9_whitelist_build(&engine->whitelist);
1754 }
1755 
1756 static void kbl_whitelist_build(struct intel_engine_cs *engine)
1757 {
1758 	struct i915_wa_list *w = &engine->whitelist;
1759 
1760 	if (engine->class != RENDER_CLASS)
1761 		return;
1762 
1763 	gen9_whitelist_build(w);
1764 
1765 	/* WaDisableLSQCROPERFforOCL:kbl */
1766 	whitelist_reg(w, GEN8_L3SQCREG4);
1767 }
1768 
1769 static void glk_whitelist_build(struct intel_engine_cs *engine)
1770 {
1771 	struct i915_wa_list *w = &engine->whitelist;
1772 
1773 	if (engine->class != RENDER_CLASS)
1774 		return;
1775 
1776 	gen9_whitelist_build(w);
1777 
1778 	/* WA #0862: Userspace has to set "Barrier Mode" to avoid hangs. */
1779 	whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1780 }
1781 
1782 static void cfl_whitelist_build(struct intel_engine_cs *engine)
1783 {
1784 	struct i915_wa_list *w = &engine->whitelist;
1785 
1786 	if (engine->class != RENDER_CLASS)
1787 		return;
1788 
1789 	gen9_whitelist_build(w);
1790 
1791 	/*
1792 	 * WaAllowPMDepthAndInvocationCountAccessFromUMD:cfl,whl,cml,aml
1793 	 *
1794 	 * This covers 4 register which are next to one another :
1795 	 *   - PS_INVOCATION_COUNT
1796 	 *   - PS_INVOCATION_COUNT_UDW
1797 	 *   - PS_DEPTH_COUNT
1798 	 *   - PS_DEPTH_COUNT_UDW
1799 	 */
1800 	whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1801 			  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1802 			  RING_FORCE_TO_NONPRIV_RANGE_4);
1803 }
1804 
1805 static void allow_read_ctx_timestamp(struct intel_engine_cs *engine)
1806 {
1807 	struct i915_wa_list *w = &engine->whitelist;
1808 
1809 	if (engine->class != RENDER_CLASS)
1810 		whitelist_reg_ext(w,
1811 				  RING_CTX_TIMESTAMP(engine->mmio_base),
1812 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1813 }
1814 
1815 static void cml_whitelist_build(struct intel_engine_cs *engine)
1816 {
1817 	allow_read_ctx_timestamp(engine);
1818 
1819 	cfl_whitelist_build(engine);
1820 }
1821 
1822 static void icl_whitelist_build(struct intel_engine_cs *engine)
1823 {
1824 	struct i915_wa_list *w = &engine->whitelist;
1825 
1826 	allow_read_ctx_timestamp(engine);
1827 
1828 	switch (engine->class) {
1829 	case RENDER_CLASS:
1830 		/* WaAllowUMDToModifyHalfSliceChicken7:icl */
1831 		whitelist_reg(w, GEN9_HALF_SLICE_CHICKEN7);
1832 
1833 		/* WaAllowUMDToModifySamplerMode:icl */
1834 		whitelist_reg(w, GEN10_SAMPLER_MODE);
1835 
1836 		/* WaEnableStateCacheRedirectToCS:icl */
1837 		whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1838 
1839 		/*
1840 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:icl
1841 		 *
1842 		 * This covers 4 register which are next to one another :
1843 		 *   - PS_INVOCATION_COUNT
1844 		 *   - PS_INVOCATION_COUNT_UDW
1845 		 *   - PS_DEPTH_COUNT
1846 		 *   - PS_DEPTH_COUNT_UDW
1847 		 */
1848 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1849 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1850 				  RING_FORCE_TO_NONPRIV_RANGE_4);
1851 		break;
1852 
1853 	case VIDEO_DECODE_CLASS:
1854 		/* hucStatusRegOffset */
1855 		whitelist_reg_ext(w, _MMIO(0x2000 + engine->mmio_base),
1856 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1857 		/* hucUKernelHdrInfoRegOffset */
1858 		whitelist_reg_ext(w, _MMIO(0x2014 + engine->mmio_base),
1859 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1860 		/* hucStatus2RegOffset */
1861 		whitelist_reg_ext(w, _MMIO(0x23B0 + engine->mmio_base),
1862 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1863 		break;
1864 
1865 	default:
1866 		break;
1867 	}
1868 }
1869 
1870 static void tgl_whitelist_build(struct intel_engine_cs *engine)
1871 {
1872 	struct i915_wa_list *w = &engine->whitelist;
1873 
1874 	allow_read_ctx_timestamp(engine);
1875 
1876 	switch (engine->class) {
1877 	case RENDER_CLASS:
1878 		/*
1879 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:tgl
1880 		 * Wa_1408556865:tgl
1881 		 *
1882 		 * This covers 4 registers which are next to one another :
1883 		 *   - PS_INVOCATION_COUNT
1884 		 *   - PS_INVOCATION_COUNT_UDW
1885 		 *   - PS_DEPTH_COUNT
1886 		 *   - PS_DEPTH_COUNT_UDW
1887 		 */
1888 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1889 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1890 				  RING_FORCE_TO_NONPRIV_RANGE_4);
1891 
1892 		/*
1893 		 * Wa_1808121037:tgl
1894 		 * Wa_14012131227:dg1
1895 		 * Wa_1508744258:tgl,rkl,dg1,adl-s,adl-p
1896 		 */
1897 		whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
1898 
1899 		/* Wa_1806527549:tgl */
1900 		whitelist_reg(w, HIZ_CHICKEN);
1901 		break;
1902 	default:
1903 		break;
1904 	}
1905 }
1906 
1907 static void dg1_whitelist_build(struct intel_engine_cs *engine)
1908 {
1909 	struct i915_wa_list *w = &engine->whitelist;
1910 
1911 	tgl_whitelist_build(engine);
1912 
1913 	/* GEN:BUG:1409280441:dg1 */
1914 	if (IS_DG1_GRAPHICS_STEP(engine->i915, STEP_A0, STEP_B0) &&
1915 	    (engine->class == RENDER_CLASS ||
1916 	     engine->class == COPY_ENGINE_CLASS))
1917 		whitelist_reg_ext(w, RING_ID(engine->mmio_base),
1918 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1919 }
1920 
1921 static void xehpsdv_whitelist_build(struct intel_engine_cs *engine)
1922 {
1923 	allow_read_ctx_timestamp(engine);
1924 }
1925 
1926 static void dg2_whitelist_build(struct intel_engine_cs *engine)
1927 {
1928 	struct i915_wa_list *w = &engine->whitelist;
1929 
1930 	allow_read_ctx_timestamp(engine);
1931 
1932 	switch (engine->class) {
1933 	case RENDER_CLASS:
1934 		/*
1935 		 * Wa_1507100340:dg2_g10
1936 		 *
1937 		 * This covers 4 registers which are next to one another :
1938 		 *   - PS_INVOCATION_COUNT
1939 		 *   - PS_INVOCATION_COUNT_UDW
1940 		 *   - PS_DEPTH_COUNT
1941 		 *   - PS_DEPTH_COUNT_UDW
1942 		 */
1943 		if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0))
1944 			whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1945 					  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1946 					  RING_FORCE_TO_NONPRIV_RANGE_4);
1947 
1948 		break;
1949 	case COMPUTE_CLASS:
1950 		/* Wa_16011157294:dg2_g10 */
1951 		if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0))
1952 			whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1953 		break;
1954 	default:
1955 		break;
1956 	}
1957 }
1958 
1959 static void blacklist_trtt(struct intel_engine_cs *engine)
1960 {
1961 	struct i915_wa_list *w = &engine->whitelist;
1962 
1963 	/*
1964 	 * Prevent read/write access to [0x4400, 0x4600) which covers
1965 	 * the TRTT range across all engines. Note that normally userspace
1966 	 * cannot access the other engines' trtt control, but for simplicity
1967 	 * we cover the entire range on each engine.
1968 	 */
1969 	whitelist_reg_ext(w, _MMIO(0x4400),
1970 			  RING_FORCE_TO_NONPRIV_DENY |
1971 			  RING_FORCE_TO_NONPRIV_RANGE_64);
1972 	whitelist_reg_ext(w, _MMIO(0x4500),
1973 			  RING_FORCE_TO_NONPRIV_DENY |
1974 			  RING_FORCE_TO_NONPRIV_RANGE_64);
1975 }
1976 
1977 static void pvc_whitelist_build(struct intel_engine_cs *engine)
1978 {
1979 	allow_read_ctx_timestamp(engine);
1980 
1981 	/* Wa_16014440446:pvc */
1982 	blacklist_trtt(engine);
1983 }
1984 
1985 void intel_engine_init_whitelist(struct intel_engine_cs *engine)
1986 {
1987 	struct drm_i915_private *i915 = engine->i915;
1988 	struct i915_wa_list *w = &engine->whitelist;
1989 
1990 	wa_init_start(w, "whitelist", engine->name);
1991 
1992 	if (IS_PONTEVECCHIO(i915))
1993 		pvc_whitelist_build(engine);
1994 	else if (IS_DG2(i915))
1995 		dg2_whitelist_build(engine);
1996 	else if (IS_XEHPSDV(i915))
1997 		xehpsdv_whitelist_build(engine);
1998 	else if (IS_DG1(i915))
1999 		dg1_whitelist_build(engine);
2000 	else if (GRAPHICS_VER(i915) == 12)
2001 		tgl_whitelist_build(engine);
2002 	else if (GRAPHICS_VER(i915) == 11)
2003 		icl_whitelist_build(engine);
2004 	else if (IS_COMETLAKE(i915))
2005 		cml_whitelist_build(engine);
2006 	else if (IS_COFFEELAKE(i915))
2007 		cfl_whitelist_build(engine);
2008 	else if (IS_GEMINILAKE(i915))
2009 		glk_whitelist_build(engine);
2010 	else if (IS_KABYLAKE(i915))
2011 		kbl_whitelist_build(engine);
2012 	else if (IS_BROXTON(i915))
2013 		bxt_whitelist_build(engine);
2014 	else if (IS_SKYLAKE(i915))
2015 		skl_whitelist_build(engine);
2016 	else if (GRAPHICS_VER(i915) <= 8)
2017 		;
2018 	else
2019 		MISSING_CASE(GRAPHICS_VER(i915));
2020 
2021 	wa_init_finish(w);
2022 }
2023 
2024 void intel_engine_apply_whitelist(struct intel_engine_cs *engine)
2025 {
2026 	const struct i915_wa_list *wal = &engine->whitelist;
2027 	struct intel_uncore *uncore = engine->uncore;
2028 	const u32 base = engine->mmio_base;
2029 	struct i915_wa *wa;
2030 	unsigned int i;
2031 
2032 	if (!wal->count)
2033 		return;
2034 
2035 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
2036 		intel_uncore_write(uncore,
2037 				   RING_FORCE_TO_NONPRIV(base, i),
2038 				   i915_mmio_reg_offset(wa->reg));
2039 
2040 	/* And clear the rest just in case of garbage */
2041 	for (; i < RING_MAX_NONPRIV_SLOTS; i++)
2042 		intel_uncore_write(uncore,
2043 				   RING_FORCE_TO_NONPRIV(base, i),
2044 				   i915_mmio_reg_offset(RING_NOPID(base)));
2045 }
2046 
2047 /*
2048  * engine_fake_wa_init(), a place holder to program the registers
2049  * which are not part of an official workaround defined by the
2050  * hardware team.
2051  * Adding programming of those register inside workaround will
2052  * allow utilizing wa framework to proper application and verification.
2053  */
2054 static void
2055 engine_fake_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2056 {
2057 	u8 mocs_w, mocs_r;
2058 
2059 	/*
2060 	 * RING_CMD_CCTL specifies the default MOCS entry that will be used
2061 	 * by the command streamer when executing commands that don't have
2062 	 * a way to explicitly specify a MOCS setting.  The default should
2063 	 * usually reference whichever MOCS entry corresponds to uncached
2064 	 * behavior, although use of a WB cached entry is recommended by the
2065 	 * spec in certain circumstances on specific platforms.
2066 	 */
2067 	if (GRAPHICS_VER(engine->i915) >= 12) {
2068 		mocs_r = engine->gt->mocs.uc_index;
2069 		mocs_w = engine->gt->mocs.uc_index;
2070 
2071 		if (HAS_L3_CCS_READ(engine->i915) &&
2072 		    engine->class == COMPUTE_CLASS) {
2073 			mocs_r = engine->gt->mocs.wb_index;
2074 
2075 			/*
2076 			 * Even on the few platforms where MOCS 0 is a
2077 			 * legitimate table entry, it's never the correct
2078 			 * setting to use here; we can assume the MOCS init
2079 			 * just forgot to initialize wb_index.
2080 			 */
2081 			drm_WARN_ON(&engine->i915->drm, mocs_r == 0);
2082 		}
2083 
2084 		wa_masked_field_set(wal,
2085 				    RING_CMD_CCTL(engine->mmio_base),
2086 				    CMD_CCTL_MOCS_MASK,
2087 				    CMD_CCTL_MOCS_OVERRIDE(mocs_w, mocs_r));
2088 	}
2089 }
2090 
2091 static bool needs_wa_1308578152(struct intel_engine_cs *engine)
2092 {
2093 	return intel_sseu_find_first_xehp_dss(&engine->gt->info.sseu, 0, 0) >=
2094 		GEN_DSS_PER_GSLICE;
2095 }
2096 
2097 static void
2098 rcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2099 {
2100 	struct drm_i915_private *i915 = engine->i915;
2101 
2102 	if (IS_DG2(i915)) {
2103 		/* Wa_1509235366:dg2 */
2104 		wa_write_or(wal, GEN12_GAMCNTRL_CTRL, INVALIDATION_BROADCAST_MODE_DIS |
2105 			    GLOBAL_INVALIDATION_MODE);
2106 	}
2107 
2108 	if (IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0)) {
2109 		/* Wa_14013392000:dg2_g11 */
2110 		wa_masked_en(wal, GEN7_ROW_CHICKEN2, GEN12_ENABLE_LARGE_GRF_MODE);
2111 
2112 		/* Wa_16011620976:dg2_g11 */
2113 		wa_write_or(wal, LSC_CHICKEN_BIT_0_UDW, DIS_CHAIN_2XSIMD8);
2114 	}
2115 
2116 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_FOREVER) ||
2117 	    IS_DG2_G11(i915) || IS_DG2_G12(i915)) {
2118 		/* Wa_1509727124:dg2 */
2119 		wa_masked_en(wal, GEN10_SAMPLER_MODE,
2120 			     SC_DISABLE_POWER_OPTIMIZATION_EBB);
2121 	}
2122 
2123 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0) ||
2124 	    IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0)) {
2125 		/* Wa_14012419201:dg2 */
2126 		wa_masked_en(wal, GEN9_ROW_CHICKEN4,
2127 			     GEN12_DISABLE_HDR_PAST_PAYLOAD_HOLD_FIX);
2128 	}
2129 
2130 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) ||
2131 	    IS_DG2_G11(i915)) {
2132 		/*
2133 		 * Wa_22012826095:dg2
2134 		 * Wa_22013059131:dg2
2135 		 */
2136 		wa_write_clr_set(wal, LSC_CHICKEN_BIT_0_UDW,
2137 				 MAXREQS_PER_BANK,
2138 				 REG_FIELD_PREP(MAXREQS_PER_BANK, 2));
2139 
2140 		/* Wa_22013059131:dg2 */
2141 		wa_write_or(wal, LSC_CHICKEN_BIT_0,
2142 			    FORCE_1_SUB_MESSAGE_PER_FRAGMENT);
2143 	}
2144 
2145 	/* Wa_1308578152:dg2_g10 when first gslice is fused off */
2146 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) &&
2147 	    needs_wa_1308578152(engine)) {
2148 		wa_masked_dis(wal, GEN12_CS_DEBUG_MODE1_CCCSUNIT_BE_COMMON,
2149 			      GEN12_REPLAY_MODE_GRANULARITY);
2150 	}
2151 
2152 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_FOREVER) ||
2153 	    IS_DG2_G11(i915) || IS_DG2_G12(i915)) {
2154 		/* Wa_22013037850:dg2 */
2155 		wa_write_or(wal, LSC_CHICKEN_BIT_0_UDW,
2156 			    DISABLE_128B_EVICTION_COMMAND_UDW);
2157 
2158 		/* Wa_22012856258:dg2 */
2159 		wa_masked_en(wal, GEN7_ROW_CHICKEN2,
2160 			     GEN12_DISABLE_READ_SUPPRESSION);
2161 
2162 		/*
2163 		 * Wa_22010960976:dg2
2164 		 * Wa_14013347512:dg2
2165 		 */
2166 		wa_masked_dis(wal, GEN12_HDC_CHICKEN0,
2167 			      LSC_L1_FLUSH_CTL_3D_DATAPORT_FLUSH_EVENTS_MASK);
2168 	}
2169 
2170 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0)) {
2171 		/*
2172 		 * Wa_1608949956:dg2_g10
2173 		 * Wa_14010198302:dg2_g10
2174 		 */
2175 		wa_masked_en(wal, GEN8_ROW_CHICKEN,
2176 			     MDQ_ARBITRATION_MODE | UGM_BACKUP_MODE);
2177 
2178 		/*
2179 		 * Wa_14010918519:dg2_g10
2180 		 *
2181 		 * LSC_CHICKEN_BIT_0 always reads back as 0 is this stepping,
2182 		 * so ignoring verification.
2183 		 */
2184 		wa_add(wal, LSC_CHICKEN_BIT_0_UDW, 0,
2185 		       FORCE_SLM_FENCE_SCOPE_TO_TILE | FORCE_UGM_FENCE_SCOPE_TO_TILE,
2186 		       0, false);
2187 	}
2188 
2189 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_A0, STEP_B0)) {
2190 		/* Wa_22010430635:dg2 */
2191 		wa_masked_en(wal,
2192 			     GEN9_ROW_CHICKEN4,
2193 			     GEN12_DISABLE_GRF_CLEAR);
2194 
2195 		/* Wa_14010648519:dg2 */
2196 		wa_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
2197 	}
2198 
2199 	/* Wa_14013202645:dg2 */
2200 	if (IS_DG2_GRAPHICS_STEP(i915, G10, STEP_B0, STEP_C0) ||
2201 	    IS_DG2_GRAPHICS_STEP(i915, G11, STEP_A0, STEP_B0))
2202 		wa_write_or(wal, RT_CTRL, DIS_NULL_QUERY);
2203 
2204 	/* Wa_22012532006:dg2 */
2205 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_C0) ||
2206 	    IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0))
2207 		wa_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
2208 			     DG2_DISABLE_ROUND_ENABLE_ALLOW_FOR_SSLA);
2209 
2210 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0)) {
2211 		/* Wa_14010680813:dg2_g10 */
2212 		wa_write_or(wal, GEN12_GAMSTLB_CTRL, CONTROL_BLOCK_CLKGATE_DIS |
2213 			    EGRESS_BLOCK_CLKGATE_DIS | TAG_BLOCK_CLKGATE_DIS);
2214 	}
2215 
2216 	if (IS_DG2_GRAPHICS_STEP(engine->i915, G10, STEP_A0, STEP_B0) ||
2217 	    IS_DG2_GRAPHICS_STEP(engine->i915, G11, STEP_A0, STEP_B0)) {
2218 		/* Wa_14012362059:dg2 */
2219 		wa_write_or(wal, GEN12_MERT_MOD_CTRL, FORCE_MISS_FTLB);
2220 	}
2221 
2222 	if (IS_DG2_GRAPHICS_STEP(i915, G11, STEP_B0, STEP_FOREVER) ||
2223 	    IS_DG2_G10(i915)) {
2224 		/* Wa_22014600077:dg2 */
2225 		wa_add(wal, GEN10_CACHE_MODE_SS, 0,
2226 		       _MASKED_BIT_ENABLE(ENABLE_EU_COUNT_FOR_TDL_FLUSH),
2227 		       0 /* Wa_14012342262 :write-only reg, so skip
2228 			    verification */,
2229 		       true);
2230 	}
2231 
2232 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2233 	    IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)) {
2234 		/*
2235 		 * Wa_1607138336:tgl[a0],dg1[a0]
2236 		 * Wa_1607063988:tgl[a0],dg1[a0]
2237 		 */
2238 		wa_write_or(wal,
2239 			    GEN9_CTX_PREEMPT_REG,
2240 			    GEN12_DISABLE_POSH_BUSY_FF_DOP_CG);
2241 	}
2242 
2243 	if (IS_TGL_UY_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)) {
2244 		/*
2245 		 * Wa_1606679103:tgl
2246 		 * (see also Wa_1606682166:icl)
2247 		 */
2248 		wa_write_or(wal,
2249 			    GEN7_SARCHKMD,
2250 			    GEN7_DISABLE_SAMPLER_PREFETCH);
2251 	}
2252 
2253 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) || IS_DG1(i915) ||
2254 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2255 		/* Wa_1606931601:tgl,rkl,dg1,adl-s,adl-p */
2256 		wa_masked_en(wal, GEN7_ROW_CHICKEN2, GEN12_DISABLE_EARLY_READ);
2257 
2258 		/*
2259 		 * Wa_1407928979:tgl A*
2260 		 * Wa_18011464164:tgl[B0+],dg1[B0+]
2261 		 * Wa_22010931296:tgl[B0+],dg1[B0+]
2262 		 * Wa_14010919138:rkl,dg1,adl-s,adl-p
2263 		 */
2264 		wa_write_or(wal, GEN7_FF_THREAD_MODE,
2265 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2266 	}
2267 
2268 	if (IS_ALDERLAKE_P(i915) || IS_DG2(i915) || IS_ALDERLAKE_S(i915) ||
2269 	    IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2270 		/*
2271 		 * Wa_1606700617:tgl,dg1,adl-p
2272 		 * Wa_22010271021:tgl,rkl,dg1,adl-s,adl-p
2273 		 * Wa_14010826681:tgl,dg1,rkl,adl-p
2274 		 * Wa_18019627453:dg2
2275 		 */
2276 		wa_masked_en(wal,
2277 			     GEN9_CS_DEBUG_MODE1,
2278 			     FF_DOP_CLOCK_GATE_DISABLE);
2279 	}
2280 
2281 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2282 	    IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2283 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2284 		/* Wa_1409804808:tgl,rkl,dg1[a0],adl-s,adl-p */
2285 		wa_masked_en(wal, GEN7_ROW_CHICKEN2,
2286 			     GEN12_PUSH_CONST_DEREF_HOLD_DIS);
2287 
2288 		/*
2289 		 * Wa_1409085225:tgl
2290 		 * Wa_14010229206:tgl,rkl,dg1[a0],adl-s,adl-p
2291 		 */
2292 		wa_masked_en(wal, GEN9_ROW_CHICKEN4, GEN12_DISABLE_TDL_PUSH);
2293 	}
2294 
2295 	if (IS_DG1_GRAPHICS_STEP(i915, STEP_A0, STEP_B0) ||
2296 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) || IS_ALDERLAKE_P(i915)) {
2297 		/*
2298 		 * Wa_1607030317:tgl
2299 		 * Wa_1607186500:tgl
2300 		 * Wa_1607297627:tgl,rkl,dg1[a0],adlp
2301 		 *
2302 		 * On TGL and RKL there are multiple entries for this WA in the
2303 		 * BSpec; some indicate this is an A0-only WA, others indicate
2304 		 * it applies to all steppings so we trust the "all steppings."
2305 		 * For DG1 this only applies to A0.
2306 		 */
2307 		wa_masked_en(wal,
2308 			     RING_PSMI_CTL(RENDER_RING_BASE),
2309 			     GEN12_WAIT_FOR_EVENT_POWER_DOWN_DISABLE |
2310 			     GEN8_RC_SEMA_IDLE_MSG_DISABLE);
2311 	}
2312 
2313 	if (IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) ||
2314 	    IS_ALDERLAKE_S(i915) || IS_ALDERLAKE_P(i915)) {
2315 		/* Wa_1406941453:tgl,rkl,dg1,adl-s,adl-p */
2316 		wa_masked_en(wal,
2317 			     GEN10_SAMPLER_MODE,
2318 			     ENABLE_SMALLPL);
2319 	}
2320 
2321 	if (GRAPHICS_VER(i915) == 11) {
2322 		/* This is not an Wa. Enable for better image quality */
2323 		wa_masked_en(wal,
2324 			     _3D_CHICKEN3,
2325 			     _3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE);
2326 
2327 		/*
2328 		 * Wa_1405543622:icl
2329 		 * Formerly known as WaGAPZPriorityScheme
2330 		 */
2331 		wa_write_or(wal,
2332 			    GEN8_GARBCNTL,
2333 			    GEN11_ARBITRATION_PRIO_ORDER_MASK);
2334 
2335 		/*
2336 		 * Wa_1604223664:icl
2337 		 * Formerly known as WaL3BankAddressHashing
2338 		 */
2339 		wa_write_clr_set(wal,
2340 				 GEN8_GARBCNTL,
2341 				 GEN11_HASH_CTRL_EXCL_MASK,
2342 				 GEN11_HASH_CTRL_EXCL_BIT0);
2343 		wa_write_clr_set(wal,
2344 				 GEN11_GLBLINVL,
2345 				 GEN11_BANK_HASH_ADDR_EXCL_MASK,
2346 				 GEN11_BANK_HASH_ADDR_EXCL_BIT0);
2347 
2348 		/*
2349 		 * Wa_1405733216:icl
2350 		 * Formerly known as WaDisableCleanEvicts
2351 		 */
2352 		wa_write_or(wal,
2353 			    GEN8_L3SQCREG4,
2354 			    GEN11_LQSC_CLEAN_EVICT_DISABLE);
2355 
2356 		/* Wa_1606682166:icl */
2357 		wa_write_or(wal,
2358 			    GEN7_SARCHKMD,
2359 			    GEN7_DISABLE_SAMPLER_PREFETCH);
2360 
2361 		/* Wa_1409178092:icl */
2362 		wa_write_clr_set(wal,
2363 				 GEN11_SCRATCH2,
2364 				 GEN11_COHERENT_PARTIAL_WRITE_MERGE_ENABLE,
2365 				 0);
2366 
2367 		/* WaEnable32PlaneMode:icl */
2368 		wa_masked_en(wal, GEN9_CSFE_CHICKEN1_RCS,
2369 			     GEN11_ENABLE_32_PLANE_MODE);
2370 
2371 		/*
2372 		 * Wa_1408615072:icl,ehl  (vsunit)
2373 		 * Wa_1407596294:icl,ehl  (hsunit)
2374 		 */
2375 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
2376 			    VSUNIT_CLKGATE_DIS | HSUNIT_CLKGATE_DIS);
2377 
2378 		/*
2379 		 * Wa_1408767742:icl[a2..forever],ehl[all]
2380 		 * Wa_1605460711:icl[a0..c0]
2381 		 */
2382 		wa_write_or(wal,
2383 			    GEN7_FF_THREAD_MODE,
2384 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2385 
2386 		/* Wa_22010271021 */
2387 		wa_masked_en(wal,
2388 			     GEN9_CS_DEBUG_MODE1,
2389 			     FF_DOP_CLOCK_GATE_DISABLE);
2390 	}
2391 
2392 	if (IS_GRAPHICS_VER(i915, 9, 12)) {
2393 		/* FtrPerCtxtPreemptionGranularityControl:skl,bxt,kbl,cfl,cnl,icl,tgl */
2394 		wa_masked_en(wal,
2395 			     GEN7_FF_SLICE_CS_CHICKEN1,
2396 			     GEN9_FFSC_PERCTX_PREEMPT_CTRL);
2397 	}
2398 
2399 	if (IS_SKYLAKE(i915) ||
2400 	    IS_KABYLAKE(i915) ||
2401 	    IS_COFFEELAKE(i915) ||
2402 	    IS_COMETLAKE(i915)) {
2403 		/* WaEnableGapsTsvCreditFix:skl,kbl,cfl */
2404 		wa_write_or(wal,
2405 			    GEN8_GARBCNTL,
2406 			    GEN9_GAPS_TSV_CREDIT_DISABLE);
2407 	}
2408 
2409 	if (IS_BROXTON(i915)) {
2410 		/* WaDisablePooledEuLoadBalancingFix:bxt */
2411 		wa_masked_en(wal,
2412 			     FF_SLICE_CS_CHICKEN2,
2413 			     GEN9_POOLED_EU_LOAD_BALANCING_FIX_DISABLE);
2414 	}
2415 
2416 	if (GRAPHICS_VER(i915) == 9) {
2417 		/* WaContextSwitchWithConcurrentTLBInvalidate:skl,bxt,kbl,glk,cfl */
2418 		wa_masked_en(wal,
2419 			     GEN9_CSFE_CHICKEN1_RCS,
2420 			     GEN9_PREEMPT_GPGPU_SYNC_SWITCH_DISABLE);
2421 
2422 		/* WaEnableLbsSlaRetryTimerDecrement:skl,bxt,kbl,glk,cfl */
2423 		wa_write_or(wal,
2424 			    BDW_SCRATCH1,
2425 			    GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
2426 
2427 		/* WaProgramL3SqcReg1DefaultForPerf:bxt,glk */
2428 		if (IS_GEN9_LP(i915))
2429 			wa_write_clr_set(wal,
2430 					 GEN8_L3SQCREG1,
2431 					 L3_PRIO_CREDITS_MASK,
2432 					 L3_GENERAL_PRIO_CREDITS(62) |
2433 					 L3_HIGH_PRIO_CREDITS(2));
2434 
2435 		/* WaOCLCoherentLineFlush:skl,bxt,kbl,cfl */
2436 		wa_write_or(wal,
2437 			    GEN8_L3SQCREG4,
2438 			    GEN8_LQSC_FLUSH_COHERENT_LINES);
2439 
2440 		/* Disable atomics in L3 to prevent unrecoverable hangs */
2441 		wa_write_clr_set(wal, GEN9_SCRATCH_LNCF1,
2442 				 GEN9_LNCF_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2443 		wa_write_clr_set(wal, GEN8_L3SQCREG4,
2444 				 GEN8_LQSQ_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2445 		wa_write_clr_set(wal, GEN9_SCRATCH1,
2446 				 EVICTION_PERF_FIX_ENABLE, 0);
2447 	}
2448 
2449 	if (IS_HASWELL(i915)) {
2450 		/* WaSampleCChickenBitEnable:hsw */
2451 		wa_masked_en(wal,
2452 			     HALF_SLICE_CHICKEN3, HSW_SAMPLE_C_PERFORMANCE);
2453 
2454 		wa_masked_dis(wal,
2455 			      CACHE_MODE_0_GEN7,
2456 			      /* enable HiZ Raw Stall Optimization */
2457 			      HIZ_RAW_STALL_OPT_DISABLE);
2458 	}
2459 
2460 	if (IS_VALLEYVIEW(i915)) {
2461 		/* WaDisableEarlyCull:vlv */
2462 		wa_masked_en(wal,
2463 			     _3D_CHICKEN3,
2464 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2465 
2466 		/*
2467 		 * WaVSThreadDispatchOverride:ivb,vlv
2468 		 *
2469 		 * This actually overrides the dispatch
2470 		 * mode for all thread types.
2471 		 */
2472 		wa_write_clr_set(wal,
2473 				 GEN7_FF_THREAD_MODE,
2474 				 GEN7_FF_SCHED_MASK,
2475 				 GEN7_FF_TS_SCHED_HW |
2476 				 GEN7_FF_VS_SCHED_HW |
2477 				 GEN7_FF_DS_SCHED_HW);
2478 
2479 		/* WaPsdDispatchEnable:vlv */
2480 		/* WaDisablePSDDualDispatchEnable:vlv */
2481 		wa_masked_en(wal,
2482 			     GEN7_HALF_SLICE_CHICKEN1,
2483 			     GEN7_MAX_PS_THREAD_DEP |
2484 			     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2485 	}
2486 
2487 	if (IS_IVYBRIDGE(i915)) {
2488 		/* WaDisableEarlyCull:ivb */
2489 		wa_masked_en(wal,
2490 			     _3D_CHICKEN3,
2491 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2492 
2493 		if (0) { /* causes HiZ corruption on ivb:gt1 */
2494 			/* enable HiZ Raw Stall Optimization */
2495 			wa_masked_dis(wal,
2496 				      CACHE_MODE_0_GEN7,
2497 				      HIZ_RAW_STALL_OPT_DISABLE);
2498 		}
2499 
2500 		/*
2501 		 * WaVSThreadDispatchOverride:ivb,vlv
2502 		 *
2503 		 * This actually overrides the dispatch
2504 		 * mode for all thread types.
2505 		 */
2506 		wa_write_clr_set(wal,
2507 				 GEN7_FF_THREAD_MODE,
2508 				 GEN7_FF_SCHED_MASK,
2509 				 GEN7_FF_TS_SCHED_HW |
2510 				 GEN7_FF_VS_SCHED_HW |
2511 				 GEN7_FF_DS_SCHED_HW);
2512 
2513 		/* WaDisablePSDDualDispatchEnable:ivb */
2514 		if (IS_IVB_GT1(i915))
2515 			wa_masked_en(wal,
2516 				     GEN7_HALF_SLICE_CHICKEN1,
2517 				     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2518 	}
2519 
2520 	if (GRAPHICS_VER(i915) == 7) {
2521 		/* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
2522 		wa_masked_en(wal,
2523 			     RING_MODE_GEN7(RENDER_RING_BASE),
2524 			     GFX_TLB_INVALIDATE_EXPLICIT | GFX_REPLAY_MODE);
2525 
2526 		/* WaDisable_RenderCache_OperationalFlush:ivb,vlv,hsw */
2527 		wa_masked_dis(wal, CACHE_MODE_0_GEN7, RC_OP_FLUSH_ENABLE);
2528 
2529 		/*
2530 		 * BSpec says this must be set, even though
2531 		 * WaDisable4x2SubspanOptimization:ivb,hsw
2532 		 * WaDisable4x2SubspanOptimization isn't listed for VLV.
2533 		 */
2534 		wa_masked_en(wal,
2535 			     CACHE_MODE_1,
2536 			     PIXEL_SUBSPAN_COLLECT_OPT_DISABLE);
2537 
2538 		/*
2539 		 * BSpec recommends 8x4 when MSAA is used,
2540 		 * however in practice 16x4 seems fastest.
2541 		 *
2542 		 * Note that PS/WM thread counts depend on the WIZ hashing
2543 		 * disable bit, which we don't touch here, but it's good
2544 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2545 		 */
2546 		wa_masked_field_set(wal,
2547 				    GEN7_GT_MODE,
2548 				    GEN6_WIZ_HASHING_MASK,
2549 				    GEN6_WIZ_HASHING_16x4);
2550 	}
2551 
2552 	if (IS_GRAPHICS_VER(i915, 6, 7))
2553 		/*
2554 		 * We need to disable the AsyncFlip performance optimisations in
2555 		 * order to use MI_WAIT_FOR_EVENT within the CS. It should
2556 		 * already be programmed to '1' on all products.
2557 		 *
2558 		 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
2559 		 */
2560 		wa_masked_en(wal,
2561 			     RING_MI_MODE(RENDER_RING_BASE),
2562 			     ASYNC_FLIP_PERF_DISABLE);
2563 
2564 	if (GRAPHICS_VER(i915) == 6) {
2565 		/*
2566 		 * Required for the hardware to program scanline values for
2567 		 * waiting
2568 		 * WaEnableFlushTlbInvalidationMode:snb
2569 		 */
2570 		wa_masked_en(wal,
2571 			     GFX_MODE,
2572 			     GFX_TLB_INVALIDATE_EXPLICIT);
2573 
2574 		/* WaDisableHiZPlanesWhenMSAAEnabled:snb */
2575 		wa_masked_en(wal,
2576 			     _3D_CHICKEN,
2577 			     _3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB);
2578 
2579 		wa_masked_en(wal,
2580 			     _3D_CHICKEN3,
2581 			     /* WaStripsFansDisableFastClipPerformanceFix:snb */
2582 			     _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL |
2583 			     /*
2584 			      * Bspec says:
2585 			      * "This bit must be set if 3DSTATE_CLIP clip mode is set
2586 			      * to normal and 3DSTATE_SF number of SF output attributes
2587 			      * is more than 16."
2588 			      */
2589 			     _3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH);
2590 
2591 		/*
2592 		 * BSpec recommends 8x4 when MSAA is used,
2593 		 * however in practice 16x4 seems fastest.
2594 		 *
2595 		 * Note that PS/WM thread counts depend on the WIZ hashing
2596 		 * disable bit, which we don't touch here, but it's good
2597 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2598 		 */
2599 		wa_masked_field_set(wal,
2600 				    GEN6_GT_MODE,
2601 				    GEN6_WIZ_HASHING_MASK,
2602 				    GEN6_WIZ_HASHING_16x4);
2603 
2604 		/* WaDisable_RenderCache_OperationalFlush:snb */
2605 		wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
2606 
2607 		/*
2608 		 * From the Sandybridge PRM, volume 1 part 3, page 24:
2609 		 * "If this bit is set, STCunit will have LRA as replacement
2610 		 *  policy. [...] This bit must be reset. LRA replacement
2611 		 *  policy is not supported."
2612 		 */
2613 		wa_masked_dis(wal,
2614 			      CACHE_MODE_0,
2615 			      CM0_STC_EVICT_DISABLE_LRA_SNB);
2616 	}
2617 
2618 	if (IS_GRAPHICS_VER(i915, 4, 6))
2619 		/* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
2620 		wa_add(wal, RING_MI_MODE(RENDER_RING_BASE),
2621 		       0, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH),
2622 		       /* XXX bit doesn't stick on Broadwater */
2623 		       IS_I965G(i915) ? 0 : VS_TIMER_DISPATCH, true);
2624 
2625 	if (GRAPHICS_VER(i915) == 4)
2626 		/*
2627 		 * Disable CONSTANT_BUFFER before it is loaded from the context
2628 		 * image. For as it is loaded, it is executed and the stored
2629 		 * address may no longer be valid, leading to a GPU hang.
2630 		 *
2631 		 * This imposes the requirement that userspace reload their
2632 		 * CONSTANT_BUFFER on every batch, fortunately a requirement
2633 		 * they are already accustomed to from before contexts were
2634 		 * enabled.
2635 		 */
2636 		wa_add(wal, ECOSKPD(RENDER_RING_BASE),
2637 		       0, _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE),
2638 		       0 /* XXX bit doesn't stick on Broadwater */,
2639 		       true);
2640 }
2641 
2642 static void
2643 xcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2644 {
2645 	struct drm_i915_private *i915 = engine->i915;
2646 
2647 	/* WaKBLVECSSemaphoreWaitPoll:kbl */
2648 	if (IS_KBL_GRAPHICS_STEP(i915, STEP_A0, STEP_F0)) {
2649 		wa_write(wal,
2650 			 RING_SEMA_WAIT_POLL(engine->mmio_base),
2651 			 1);
2652 	}
2653 }
2654 
2655 static void
2656 ccs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2657 {
2658 	if (IS_PVC_CT_STEP(engine->i915, STEP_A0, STEP_C0)) {
2659 		/* Wa_14014999345:pvc */
2660 		wa_masked_en(wal, GEN10_CACHE_MODE_SS, DISABLE_ECC);
2661 	}
2662 }
2663 
2664 /*
2665  * The bspec performance guide has recommended MMIO tuning settings.  These
2666  * aren't truly "workarounds" but we want to program them with the same
2667  * workaround infrastructure to ensure that they're automatically added to
2668  * the GuC save/restore lists, re-applied at the right times, and checked for
2669  * any conflicting programming requested by real workarounds.
2670  *
2671  * Programming settings should be added here only if their registers are not
2672  * part of an engine's register state context.  If a register is part of a
2673  * context, then any tuning settings should be programmed in an appropriate
2674  * function invoked by __intel_engine_init_ctx_wa().
2675  */
2676 static void
2677 add_render_compute_tuning_settings(struct drm_i915_private *i915,
2678 				   struct i915_wa_list *wal)
2679 {
2680 	if (IS_PONTEVECCHIO(i915)) {
2681 		wa_write(wal, XEHPC_L3SCRUB,
2682 			 SCRUB_CL_DWNGRADE_SHARED | SCRUB_RATE_4B_PER_CLK);
2683 	}
2684 
2685 	if (IS_DG2(i915)) {
2686 		wa_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
2687 		wa_write_clr_set(wal, RT_CTRL, STACKID_CTRL, STACKID_CTRL_512);
2688 
2689 		/*
2690 		 * This is also listed as Wa_22012654132 for certain DG2
2691 		 * steppings, but the tuning setting programming is a superset
2692 		 * since it applies to all DG2 variants and steppings.
2693 		 *
2694 		 * Note that register 0xE420 is write-only and cannot be read
2695 		 * back for verification on DG2 (due to Wa_14012342262), so
2696 		 * we need to explicitly skip the readback.
2697 		 */
2698 		wa_add(wal, GEN10_CACHE_MODE_SS, 0,
2699 		       _MASKED_BIT_ENABLE(ENABLE_PREFETCH_INTO_IC),
2700 		       0 /* write-only, so skip validation */,
2701 		       true);
2702 	}
2703 
2704 	/*
2705 	 * This tuning setting proves beneficial only on ATS-M designs; the
2706 	 * default "age based" setting is optimal on regular DG2 and other
2707 	 * platforms.
2708 	 */
2709 	if (INTEL_INFO(i915)->tuning_thread_rr_after_dep)
2710 		wa_masked_field_set(wal, GEN9_ROW_CHICKEN4, THREAD_EX_ARB_MODE,
2711 				    THREAD_EX_ARB_MODE_RR_AFTER_DEP);
2712 }
2713 
2714 /*
2715  * The workarounds in this function apply to shared registers in
2716  * the general render reset domain that aren't tied to a
2717  * specific engine.  Since all render+compute engines get reset
2718  * together, and the contents of these registers are lost during
2719  * the shared render domain reset, we'll define such workarounds
2720  * here and then add them to just a single RCS or CCS engine's
2721  * workaround list (whichever engine has the XXXX flag).
2722  */
2723 static void
2724 general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2725 {
2726 	struct drm_i915_private *i915 = engine->i915;
2727 
2728 	add_render_compute_tuning_settings(i915, wal);
2729 
2730 	if (IS_PONTEVECCHIO(i915)) {
2731 		/* Wa_16016694945 */
2732 		wa_masked_en(wal, XEHPC_LNCFMISCCFGREG0, XEHPC_OVRLSCCC);
2733 	}
2734 
2735 	if (IS_XEHPSDV(i915)) {
2736 		/* Wa_1409954639 */
2737 		wa_masked_en(wal,
2738 			     GEN8_ROW_CHICKEN,
2739 			     SYSTOLIC_DOP_CLOCK_GATING_DIS);
2740 
2741 		/* Wa_1607196519 */
2742 		wa_masked_en(wal,
2743 			     GEN9_ROW_CHICKEN4,
2744 			     GEN12_DISABLE_GRF_CLEAR);
2745 
2746 		/* Wa_14010670810:xehpsdv */
2747 		wa_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
2748 
2749 		/* Wa_14010449647:xehpsdv */
2750 		wa_masked_en(wal, GEN7_HALF_SLICE_CHICKEN1,
2751 			     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2752 
2753 		/* Wa_18011725039:xehpsdv */
2754 		if (IS_XEHPSDV_GRAPHICS_STEP(i915, STEP_A1, STEP_B0)) {
2755 			wa_masked_dis(wal, MLTICTXCTL, TDONRENDER);
2756 			wa_write_or(wal, L3SQCREG1_CCS0, FLUSHALLNONCOH);
2757 		}
2758 
2759 		/* Wa_14012362059:xehpsdv */
2760 		wa_write_or(wal, GEN12_MERT_MOD_CTRL, FORCE_MISS_FTLB);
2761 
2762 		/* Wa_14014368820:xehpsdv */
2763 		wa_write_or(wal, GEN12_GAMCNTRL_CTRL, INVALIDATION_BROADCAST_MODE_DIS |
2764 				GLOBAL_INVALIDATION_MODE);
2765 	}
2766 
2767 	if (IS_DG2(i915) || IS_PONTEVECCHIO(i915)) {
2768 		/* Wa_14015227452:dg2,pvc */
2769 		wa_masked_en(wal, GEN9_ROW_CHICKEN4, XEHP_DIS_BBL_SYSPIPE);
2770 
2771 		/* Wa_22014226127:dg2,pvc */
2772 		wa_write_or(wal, LSC_CHICKEN_BIT_0, DISABLE_D8_D16_COASLESCE);
2773 
2774 		/* Wa_16015675438:dg2,pvc */
2775 		wa_masked_en(wal, FF_SLICE_CS_CHICKEN2, GEN12_PERF_FIX_BALANCING_CFE_DISABLE);
2776 
2777 		/* Wa_18018781329:dg2,pvc */
2778 		wa_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
2779 		wa_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
2780 		wa_write_or(wal, VDBX_MOD_CTRL, FORCE_MISS_FTLB);
2781 		wa_write_or(wal, VEBX_MOD_CTRL, FORCE_MISS_FTLB);
2782 	}
2783 }
2784 
2785 static void
2786 engine_init_workarounds(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2787 {
2788 	if (I915_SELFTEST_ONLY(GRAPHICS_VER(engine->i915) < 4))
2789 		return;
2790 
2791 	engine_fake_wa_init(engine, wal);
2792 
2793 	/*
2794 	 * These are common workarounds that just need to applied
2795 	 * to a single RCS/CCS engine's workaround list since
2796 	 * they're reset as part of the general render domain reset.
2797 	 */
2798 	if (engine->flags & I915_ENGINE_FIRST_RENDER_COMPUTE)
2799 		general_render_compute_wa_init(engine, wal);
2800 
2801 	if (engine->class == COMPUTE_CLASS)
2802 		ccs_engine_wa_init(engine, wal);
2803 	else if (engine->class == RENDER_CLASS)
2804 		rcs_engine_wa_init(engine, wal);
2805 	else
2806 		xcs_engine_wa_init(engine, wal);
2807 }
2808 
2809 void intel_engine_init_workarounds(struct intel_engine_cs *engine)
2810 {
2811 	struct i915_wa_list *wal = &engine->wa_list;
2812 
2813 	if (GRAPHICS_VER(engine->i915) < 4)
2814 		return;
2815 
2816 	wa_init_start(wal, "engine", engine->name);
2817 	engine_init_workarounds(engine, wal);
2818 	wa_init_finish(wal);
2819 }
2820 
2821 void intel_engine_apply_workarounds(struct intel_engine_cs *engine)
2822 {
2823 	wa_list_apply(engine->gt, &engine->wa_list);
2824 }
2825 
2826 static const struct i915_range mcr_ranges_gen8[] = {
2827 	{ .start = 0x5500, .end = 0x55ff },
2828 	{ .start = 0x7000, .end = 0x7fff },
2829 	{ .start = 0x9400, .end = 0x97ff },
2830 	{ .start = 0xb000, .end = 0xb3ff },
2831 	{ .start = 0xe000, .end = 0xe7ff },
2832 	{},
2833 };
2834 
2835 static const struct i915_range mcr_ranges_gen12[] = {
2836 	{ .start =  0x8150, .end =  0x815f },
2837 	{ .start =  0x9520, .end =  0x955f },
2838 	{ .start =  0xb100, .end =  0xb3ff },
2839 	{ .start =  0xde80, .end =  0xe8ff },
2840 	{ .start = 0x24a00, .end = 0x24a7f },
2841 	{},
2842 };
2843 
2844 static const struct i915_range mcr_ranges_xehp[] = {
2845 	{ .start =  0x4000, .end =  0x4aff },
2846 	{ .start =  0x5200, .end =  0x52ff },
2847 	{ .start =  0x5400, .end =  0x7fff },
2848 	{ .start =  0x8140, .end =  0x815f },
2849 	{ .start =  0x8c80, .end =  0x8dff },
2850 	{ .start =  0x94d0, .end =  0x955f },
2851 	{ .start =  0x9680, .end =  0x96ff },
2852 	{ .start =  0xb000, .end =  0xb3ff },
2853 	{ .start =  0xc800, .end =  0xcfff },
2854 	{ .start =  0xd800, .end =  0xd8ff },
2855 	{ .start =  0xdc00, .end =  0xffff },
2856 	{ .start = 0x17000, .end = 0x17fff },
2857 	{ .start = 0x24a00, .end = 0x24a7f },
2858 	{},
2859 };
2860 
2861 static bool mcr_range(struct drm_i915_private *i915, u32 offset)
2862 {
2863 	const struct i915_range *mcr_ranges;
2864 	int i;
2865 
2866 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50))
2867 		mcr_ranges = mcr_ranges_xehp;
2868 	else if (GRAPHICS_VER(i915) >= 12)
2869 		mcr_ranges = mcr_ranges_gen12;
2870 	else if (GRAPHICS_VER(i915) >= 8)
2871 		mcr_ranges = mcr_ranges_gen8;
2872 	else
2873 		return false;
2874 
2875 	/*
2876 	 * Registers in these ranges are affected by the MCR selector
2877 	 * which only controls CPU initiated MMIO. Routing does not
2878 	 * work for CS access so we cannot verify them on this path.
2879 	 */
2880 	for (i = 0; mcr_ranges[i].start; i++)
2881 		if (offset >= mcr_ranges[i].start &&
2882 		    offset <= mcr_ranges[i].end)
2883 			return true;
2884 
2885 	return false;
2886 }
2887 
2888 static int
2889 wa_list_srm(struct i915_request *rq,
2890 	    const struct i915_wa_list *wal,
2891 	    struct i915_vma *vma)
2892 {
2893 	struct drm_i915_private *i915 = rq->engine->i915;
2894 	unsigned int i, count = 0;
2895 	const struct i915_wa *wa;
2896 	u32 srm, *cs;
2897 
2898 	srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
2899 	if (GRAPHICS_VER(i915) >= 8)
2900 		srm++;
2901 
2902 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2903 		if (!mcr_range(i915, i915_mmio_reg_offset(wa->reg)))
2904 			count++;
2905 	}
2906 
2907 	cs = intel_ring_begin(rq, 4 * count);
2908 	if (IS_ERR(cs))
2909 		return PTR_ERR(cs);
2910 
2911 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2912 		u32 offset = i915_mmio_reg_offset(wa->reg);
2913 
2914 		if (mcr_range(i915, offset))
2915 			continue;
2916 
2917 		*cs++ = srm;
2918 		*cs++ = offset;
2919 		*cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
2920 		*cs++ = 0;
2921 	}
2922 	intel_ring_advance(rq, cs);
2923 
2924 	return 0;
2925 }
2926 
2927 static int engine_wa_list_verify(struct intel_context *ce,
2928 				 const struct i915_wa_list * const wal,
2929 				 const char *from)
2930 {
2931 	const struct i915_wa *wa;
2932 	struct i915_request *rq;
2933 	struct i915_vma *vma;
2934 	struct i915_gem_ww_ctx ww;
2935 	unsigned int i;
2936 	u32 *results;
2937 	int err;
2938 
2939 	if (!wal->count)
2940 		return 0;
2941 
2942 	vma = __vm_create_scratch_for_read(&ce->engine->gt->ggtt->vm,
2943 					   wal->count * sizeof(u32));
2944 	if (IS_ERR(vma))
2945 		return PTR_ERR(vma);
2946 
2947 	intel_engine_pm_get(ce->engine);
2948 	i915_gem_ww_ctx_init(&ww, false);
2949 retry:
2950 	err = i915_gem_object_lock(vma->obj, &ww);
2951 	if (err == 0)
2952 		err = intel_context_pin_ww(ce, &ww);
2953 	if (err)
2954 		goto err_pm;
2955 
2956 	err = i915_vma_pin_ww(vma, &ww, 0, 0,
2957 			   i915_vma_is_ggtt(vma) ? PIN_GLOBAL : PIN_USER);
2958 	if (err)
2959 		goto err_unpin;
2960 
2961 	rq = i915_request_create(ce);
2962 	if (IS_ERR(rq)) {
2963 		err = PTR_ERR(rq);
2964 		goto err_vma;
2965 	}
2966 
2967 	err = i915_request_await_object(rq, vma->obj, true);
2968 	if (err == 0)
2969 		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
2970 	if (err == 0)
2971 		err = wa_list_srm(rq, wal, vma);
2972 
2973 	i915_request_get(rq);
2974 	if (err)
2975 		i915_request_set_error_once(rq, err);
2976 	i915_request_add(rq);
2977 
2978 	if (err)
2979 		goto err_rq;
2980 
2981 	if (i915_request_wait(rq, 0, HZ / 5) < 0) {
2982 		err = -ETIME;
2983 		goto err_rq;
2984 	}
2985 
2986 	results = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
2987 	if (IS_ERR(results)) {
2988 		err = PTR_ERR(results);
2989 		goto err_rq;
2990 	}
2991 
2992 	err = 0;
2993 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2994 		if (mcr_range(rq->engine->i915, i915_mmio_reg_offset(wa->reg)))
2995 			continue;
2996 
2997 		if (!wa_verify(wa, results[i], wal->name, from))
2998 			err = -ENXIO;
2999 	}
3000 
3001 	i915_gem_object_unpin_map(vma->obj);
3002 
3003 err_rq:
3004 	i915_request_put(rq);
3005 err_vma:
3006 	i915_vma_unpin(vma);
3007 err_unpin:
3008 	intel_context_unpin(ce);
3009 err_pm:
3010 	if (err == -EDEADLK) {
3011 		err = i915_gem_ww_ctx_backoff(&ww);
3012 		if (!err)
3013 			goto retry;
3014 	}
3015 	i915_gem_ww_ctx_fini(&ww);
3016 	intel_engine_pm_put(ce->engine);
3017 	i915_vma_put(vma);
3018 	return err;
3019 }
3020 
3021 int intel_engine_verify_workarounds(struct intel_engine_cs *engine,
3022 				    const char *from)
3023 {
3024 	return engine_wa_list_verify(engine->kernel_context,
3025 				     &engine->wa_list,
3026 				     from);
3027 }
3028 
3029 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3030 #include "selftest_workarounds.c"
3031 #endif
3032