xref: /linux/drivers/gpu/drm/i915/gt/intel_workarounds.c (revision 3027ce13e04eee76539ca65c2cb1028a01c8c508)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014-2018 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "i915_reg.h"
8 #include "intel_context.h"
9 #include "intel_engine_pm.h"
10 #include "intel_engine_regs.h"
11 #include "intel_gpu_commands.h"
12 #include "intel_gt.h"
13 #include "intel_gt_ccs_mode.h"
14 #include "intel_gt_mcr.h"
15 #include "intel_gt_print.h"
16 #include "intel_gt_regs.h"
17 #include "intel_ring.h"
18 #include "intel_workarounds.h"
19 
20 /**
21  * DOC: Hardware workarounds
22  *
23  * Hardware workarounds are register programming documented to be executed in
24  * the driver that fall outside of the normal programming sequences for a
25  * platform. There are some basic categories of workarounds, depending on
26  * how/when they are applied:
27  *
28  * - Context workarounds: workarounds that touch registers that are
29  *   saved/restored to/from the HW context image. The list is emitted (via Load
30  *   Register Immediate commands) once when initializing the device and saved in
31  *   the default context. That default context is then used on every context
32  *   creation to have a "primed golden context", i.e. a context image that
33  *   already contains the changes needed to all the registers.
34  *
35  *   Context workarounds should be implemented in the \*_ctx_workarounds_init()
36  *   variants respective to the targeted platforms.
37  *
38  * - Engine workarounds: the list of these WAs is applied whenever the specific
39  *   engine is reset. It's also possible that a set of engine classes share a
40  *   common power domain and they are reset together. This happens on some
41  *   platforms with render and compute engines. In this case (at least) one of
42  *   them need to keeep the workaround programming: the approach taken in the
43  *   driver is to tie those workarounds to the first compute/render engine that
44  *   is registered.  When executing with GuC submission, engine resets are
45  *   outside of kernel driver control, hence the list of registers involved in
46  *   written once, on engine initialization, and then passed to GuC, that
47  *   saves/restores their values before/after the reset takes place. See
48  *   ``drivers/gpu/drm/i915/gt/uc/intel_guc_ads.c`` for reference.
49  *
50  *   Workarounds for registers specific to RCS and CCS should be implemented in
51  *   rcs_engine_wa_init() and ccs_engine_wa_init(), respectively; those for
52  *   registers belonging to BCS, VCS or VECS should be implemented in
53  *   xcs_engine_wa_init(). Workarounds for registers not belonging to a specific
54  *   engine's MMIO range but that are part of of the common RCS/CCS reset domain
55  *   should be implemented in general_render_compute_wa_init(). The settings
56  *   about the CCS load balancing should be added in ccs_engine_wa_mode().
57  *
58  * - GT workarounds: the list of these WAs is applied whenever these registers
59  *   revert to their default values: on GPU reset, suspend/resume [1]_, etc.
60  *
61  *   GT workarounds should be implemented in the \*_gt_workarounds_init()
62  *   variants respective to the targeted platforms.
63  *
64  * - Register whitelist: some workarounds need to be implemented in userspace,
65  *   but need to touch privileged registers. The whitelist in the kernel
66  *   instructs the hardware to allow the access to happen. From the kernel side,
67  *   this is just a special case of a MMIO workaround (as we write the list of
68  *   these to/be-whitelisted registers to some special HW registers).
69  *
70  *   Register whitelisting should be done in the \*_whitelist_build() variants
71  *   respective to the targeted platforms.
72  *
73  * - Workaround batchbuffers: buffers that get executed automatically by the
74  *   hardware on every HW context restore. These buffers are created and
75  *   programmed in the default context so the hardware always go through those
76  *   programming sequences when switching contexts. The support for workaround
77  *   batchbuffers is enabled these hardware mechanisms:
78  *
79  *   #. INDIRECT_CTX: A batchbuffer and an offset are provided in the default
80  *      context, pointing the hardware to jump to that location when that offset
81  *      is reached in the context restore. Workaround batchbuffer in the driver
82  *      currently uses this mechanism for all platforms.
83  *
84  *   #. BB_PER_CTX_PTR: A batchbuffer is provided in the default context,
85  *      pointing the hardware to a buffer to continue executing after the
86  *      engine registers are restored in a context restore sequence. This is
87  *      currently not used in the driver.
88  *
89  * - Other:  There are WAs that, due to their nature, cannot be applied from a
90  *   central place. Those are peppered around the rest of the code, as needed.
91  *   Workarounds related to the display IP are the main example.
92  *
93  * .. [1] Technically, some registers are powercontext saved & restored, so they
94  *    survive a suspend/resume. In practice, writing them again is not too
95  *    costly and simplifies things, so it's the approach taken in the driver.
96  */
97 
98 static void wa_init_start(struct i915_wa_list *wal, struct intel_gt *gt,
99 			  const char *name, const char *engine_name)
100 {
101 	wal->gt = gt;
102 	wal->name = name;
103 	wal->engine_name = engine_name;
104 }
105 
106 #define WA_LIST_CHUNK (1 << 4)
107 
108 static void wa_init_finish(struct i915_wa_list *wal)
109 {
110 	/* Trim unused entries. */
111 	if (!IS_ALIGNED(wal->count, WA_LIST_CHUNK)) {
112 		struct i915_wa *list = kmemdup(wal->list,
113 					       wal->count * sizeof(*list),
114 					       GFP_KERNEL);
115 
116 		if (list) {
117 			kfree(wal->list);
118 			wal->list = list;
119 		}
120 	}
121 
122 	if (!wal->count)
123 		return;
124 
125 	gt_dbg(wal->gt, "Initialized %u %s workarounds on %s\n",
126 	       wal->wa_count, wal->name, wal->engine_name);
127 }
128 
129 static enum forcewake_domains
130 wal_get_fw_for_rmw(struct intel_uncore *uncore, const struct i915_wa_list *wal)
131 {
132 	enum forcewake_domains fw = 0;
133 	struct i915_wa *wa;
134 	unsigned int i;
135 
136 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
137 		fw |= intel_uncore_forcewake_for_reg(uncore,
138 						     wa->reg,
139 						     FW_REG_READ |
140 						     FW_REG_WRITE);
141 
142 	return fw;
143 }
144 
145 static void _wa_add(struct i915_wa_list *wal, const struct i915_wa *wa)
146 {
147 	unsigned int addr = i915_mmio_reg_offset(wa->reg);
148 	struct drm_i915_private *i915 = wal->gt->i915;
149 	unsigned int start = 0, end = wal->count;
150 	const unsigned int grow = WA_LIST_CHUNK;
151 	struct i915_wa *wa_;
152 
153 	GEM_BUG_ON(!is_power_of_2(grow));
154 
155 	if (IS_ALIGNED(wal->count, grow)) { /* Either uninitialized or full. */
156 		struct i915_wa *list;
157 
158 		list = kmalloc_array(ALIGN(wal->count + 1, grow), sizeof(*wa),
159 				     GFP_KERNEL);
160 		if (!list) {
161 			drm_err(&i915->drm, "No space for workaround init!\n");
162 			return;
163 		}
164 
165 		if (wal->list) {
166 			memcpy(list, wal->list, sizeof(*wa) * wal->count);
167 			kfree(wal->list);
168 		}
169 
170 		wal->list = list;
171 	}
172 
173 	while (start < end) {
174 		unsigned int mid = start + (end - start) / 2;
175 
176 		if (i915_mmio_reg_offset(wal->list[mid].reg) < addr) {
177 			start = mid + 1;
178 		} else if (i915_mmio_reg_offset(wal->list[mid].reg) > addr) {
179 			end = mid;
180 		} else {
181 			wa_ = &wal->list[mid];
182 
183 			if ((wa->clr | wa_->clr) && !(wa->clr & ~wa_->clr)) {
184 				drm_err(&i915->drm,
185 					"Discarding overwritten w/a for reg %04x (clear: %08x, set: %08x)\n",
186 					i915_mmio_reg_offset(wa_->reg),
187 					wa_->clr, wa_->set);
188 
189 				wa_->set &= ~wa->clr;
190 			}
191 
192 			wal->wa_count++;
193 			wa_->set |= wa->set;
194 			wa_->clr |= wa->clr;
195 			wa_->read |= wa->read;
196 			return;
197 		}
198 	}
199 
200 	wal->wa_count++;
201 	wa_ = &wal->list[wal->count++];
202 	*wa_ = *wa;
203 
204 	while (wa_-- > wal->list) {
205 		GEM_BUG_ON(i915_mmio_reg_offset(wa_[0].reg) ==
206 			   i915_mmio_reg_offset(wa_[1].reg));
207 		if (i915_mmio_reg_offset(wa_[1].reg) >
208 		    i915_mmio_reg_offset(wa_[0].reg))
209 			break;
210 
211 		swap(wa_[1], wa_[0]);
212 	}
213 }
214 
215 static void wa_add(struct i915_wa_list *wal, i915_reg_t reg,
216 		   u32 clear, u32 set, u32 read_mask, bool masked_reg)
217 {
218 	struct i915_wa wa = {
219 		.reg  = reg,
220 		.clr  = clear,
221 		.set  = set,
222 		.read = read_mask,
223 		.masked_reg = masked_reg,
224 	};
225 
226 	_wa_add(wal, &wa);
227 }
228 
229 static void wa_mcr_add(struct i915_wa_list *wal, i915_mcr_reg_t reg,
230 		       u32 clear, u32 set, u32 read_mask, bool masked_reg)
231 {
232 	struct i915_wa wa = {
233 		.mcr_reg = reg,
234 		.clr  = clear,
235 		.set  = set,
236 		.read = read_mask,
237 		.masked_reg = masked_reg,
238 		.is_mcr = 1,
239 	};
240 
241 	_wa_add(wal, &wa);
242 }
243 
244 static void
245 wa_write_clr_set(struct i915_wa_list *wal, i915_reg_t reg, u32 clear, u32 set)
246 {
247 	wa_add(wal, reg, clear, set, clear | set, false);
248 }
249 
250 static void
251 wa_mcr_write_clr_set(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 clear, u32 set)
252 {
253 	wa_mcr_add(wal, reg, clear, set, clear | set, false);
254 }
255 
256 static void
257 wa_write(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
258 {
259 	wa_write_clr_set(wal, reg, ~0, set);
260 }
261 
262 static void
263 wa_write_or(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
264 {
265 	wa_write_clr_set(wal, reg, set, set);
266 }
267 
268 static void
269 wa_mcr_write_or(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 set)
270 {
271 	wa_mcr_write_clr_set(wal, reg, set, set);
272 }
273 
274 static void
275 wa_write_clr(struct i915_wa_list *wal, i915_reg_t reg, u32 clr)
276 {
277 	wa_write_clr_set(wal, reg, clr, 0);
278 }
279 
280 static void
281 wa_mcr_write_clr(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 clr)
282 {
283 	wa_mcr_write_clr_set(wal, reg, clr, 0);
284 }
285 
286 /*
287  * WA operations on "masked register". A masked register has the upper 16 bits
288  * documented as "masked" in b-spec. Its purpose is to allow writing to just a
289  * portion of the register without a rmw: you simply write in the upper 16 bits
290  * the mask of bits you are going to modify.
291  *
292  * The wa_masked_* family of functions already does the necessary operations to
293  * calculate the mask based on the parameters passed, so user only has to
294  * provide the lower 16 bits of that register.
295  */
296 
297 static void
298 wa_masked_en(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
299 {
300 	wa_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
301 }
302 
303 static void
304 wa_mcr_masked_en(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 val)
305 {
306 	wa_mcr_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
307 }
308 
309 static void
310 wa_masked_dis(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
311 {
312 	wa_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
313 }
314 
315 static void
316 wa_mcr_masked_dis(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 val)
317 {
318 	wa_mcr_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
319 }
320 
321 static void
322 wa_masked_field_set(struct i915_wa_list *wal, i915_reg_t reg,
323 		    u32 mask, u32 val)
324 {
325 	wa_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
326 }
327 
328 static void
329 wa_mcr_masked_field_set(struct i915_wa_list *wal, i915_mcr_reg_t reg,
330 			u32 mask, u32 val)
331 {
332 	wa_mcr_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
333 }
334 
335 static void gen6_ctx_workarounds_init(struct intel_engine_cs *engine,
336 				      struct i915_wa_list *wal)
337 {
338 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
339 }
340 
341 static void gen7_ctx_workarounds_init(struct intel_engine_cs *engine,
342 				      struct i915_wa_list *wal)
343 {
344 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
345 }
346 
347 static void gen8_ctx_workarounds_init(struct intel_engine_cs *engine,
348 				      struct i915_wa_list *wal)
349 {
350 	wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
351 
352 	/* WaDisableAsyncFlipPerfMode:bdw,chv */
353 	wa_masked_en(wal, RING_MI_MODE(RENDER_RING_BASE), ASYNC_FLIP_PERF_DISABLE);
354 
355 	/* WaDisablePartialInstShootdown:bdw,chv */
356 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
357 			 PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
358 
359 	/* Use Force Non-Coherent whenever executing a 3D context. This is a
360 	 * workaround for a possible hang in the unlikely event a TLB
361 	 * invalidation occurs during a PSD flush.
362 	 */
363 	/* WaForceEnableNonCoherent:bdw,chv */
364 	/* WaHdcDisableFetchWhenMasked:bdw,chv */
365 	wa_masked_en(wal, HDC_CHICKEN0,
366 		     HDC_DONOT_FETCH_MEM_WHEN_MASKED |
367 		     HDC_FORCE_NON_COHERENT);
368 
369 	/* From the Haswell PRM, Command Reference: Registers, CACHE_MODE_0:
370 	 * "The Hierarchical Z RAW Stall Optimization allows non-overlapping
371 	 *  polygons in the same 8x4 pixel/sample area to be processed without
372 	 *  stalling waiting for the earlier ones to write to Hierarchical Z
373 	 *  buffer."
374 	 *
375 	 * This optimization is off by default for BDW and CHV; turn it on.
376 	 */
377 	wa_masked_dis(wal, CACHE_MODE_0_GEN7, HIZ_RAW_STALL_OPT_DISABLE);
378 
379 	/* Wa4x4STCOptimizationDisable:bdw,chv */
380 	wa_masked_en(wal, CACHE_MODE_1, GEN8_4x4_STC_OPTIMIZATION_DISABLE);
381 
382 	/*
383 	 * BSpec recommends 8x4 when MSAA is used,
384 	 * however in practice 16x4 seems fastest.
385 	 *
386 	 * Note that PS/WM thread counts depend on the WIZ hashing
387 	 * disable bit, which we don't touch here, but it's good
388 	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
389 	 */
390 	wa_masked_field_set(wal, GEN7_GT_MODE,
391 			    GEN6_WIZ_HASHING_MASK,
392 			    GEN6_WIZ_HASHING_16x4);
393 }
394 
395 static void bdw_ctx_workarounds_init(struct intel_engine_cs *engine,
396 				     struct i915_wa_list *wal)
397 {
398 	struct drm_i915_private *i915 = engine->i915;
399 
400 	gen8_ctx_workarounds_init(engine, wal);
401 
402 	/* WaDisableThreadStallDopClockGating:bdw (pre-production) */
403 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
404 
405 	/* WaDisableDopClockGating:bdw
406 	 *
407 	 * Also see the related UCGTCL1 write in bdw_init_clock_gating()
408 	 * to disable EUTC clock gating.
409 	 */
410 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
411 			 DOP_CLOCK_GATING_DISABLE);
412 
413 	wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN3,
414 			 GEN8_SAMPLER_POWER_BYPASS_DIS);
415 
416 	wa_masked_en(wal, HDC_CHICKEN0,
417 		     /* WaForceContextSaveRestoreNonCoherent:bdw */
418 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
419 		     /* WaDisableFenceDestinationToSLM:bdw (pre-prod) */
420 		     (IS_BROADWELL_GT3(i915) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
421 }
422 
423 static void chv_ctx_workarounds_init(struct intel_engine_cs *engine,
424 				     struct i915_wa_list *wal)
425 {
426 	gen8_ctx_workarounds_init(engine, wal);
427 
428 	/* WaDisableThreadStallDopClockGating:chv */
429 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
430 
431 	/* Improve HiZ throughput on CHV. */
432 	wa_masked_en(wal, HIZ_CHICKEN, CHV_HZ_8X8_MODE_IN_1X);
433 }
434 
435 static void gen9_ctx_workarounds_init(struct intel_engine_cs *engine,
436 				      struct i915_wa_list *wal)
437 {
438 	struct drm_i915_private *i915 = engine->i915;
439 
440 	if (HAS_LLC(i915)) {
441 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
442 		 *
443 		 * Must match Display Engine. See
444 		 * WaCompressedResourceDisplayNewHashMode.
445 		 */
446 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
447 			     GEN9_PBE_COMPRESSED_HASH_SELECTION);
448 		wa_mcr_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
449 				 GEN9_SAMPLER_HASH_COMPRESSED_READ_ADDR);
450 	}
451 
452 	/* WaClearFlowControlGpgpuContextSave:skl,bxt,kbl,glk,cfl */
453 	/* WaDisablePartialInstShootdown:skl,bxt,kbl,glk,cfl */
454 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
455 			 FLOW_CONTROL_ENABLE |
456 			 PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
457 
458 	/* WaEnableYV12BugFixInHalfSliceChicken7:skl,bxt,kbl,glk,cfl */
459 	/* WaEnableSamplerGPGPUPreemptionSupport:skl,bxt,kbl,cfl */
460 	wa_mcr_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
461 			 GEN9_ENABLE_YV12_BUGFIX |
462 			 GEN9_ENABLE_GPGPU_PREEMPTION);
463 
464 	/* Wa4x4STCOptimizationDisable:skl,bxt,kbl,glk,cfl */
465 	/* WaDisablePartialResolveInVc:skl,bxt,kbl,cfl */
466 	wa_masked_en(wal, CACHE_MODE_1,
467 		     GEN8_4x4_STC_OPTIMIZATION_DISABLE |
468 		     GEN9_PARTIAL_RESOLVE_IN_VC_DISABLE);
469 
470 	/* WaCcsTlbPrefetchDisable:skl,bxt,kbl,glk,cfl */
471 	wa_mcr_masked_dis(wal, GEN9_HALF_SLICE_CHICKEN5,
472 			  GEN9_CCS_TLB_PREFETCH_ENABLE);
473 
474 	/* WaForceContextSaveRestoreNonCoherent:skl,bxt,kbl,cfl */
475 	wa_masked_en(wal, HDC_CHICKEN0,
476 		     HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
477 		     HDC_FORCE_CSR_NON_COHERENT_OVR_DISABLE);
478 
479 	/* WaForceEnableNonCoherent and WaDisableHDCInvalidation are
480 	 * both tied to WaForceContextSaveRestoreNonCoherent
481 	 * in some hsds for skl. We keep the tie for all gen9. The
482 	 * documentation is a bit hazy and so we want to get common behaviour,
483 	 * even though there is no clear evidence we would need both on kbl/bxt.
484 	 * This area has been source of system hangs so we play it safe
485 	 * and mimic the skl regardless of what bspec says.
486 	 *
487 	 * Use Force Non-Coherent whenever executing a 3D context. This
488 	 * is a workaround for a possible hang in the unlikely event
489 	 * a TLB invalidation occurs during a PSD flush.
490 	 */
491 
492 	/* WaForceEnableNonCoherent:skl,bxt,kbl,cfl */
493 	wa_masked_en(wal, HDC_CHICKEN0,
494 		     HDC_FORCE_NON_COHERENT);
495 
496 	/* WaDisableSamplerPowerBypassForSOPingPong:skl,bxt,kbl,cfl */
497 	if (IS_SKYLAKE(i915) ||
498 	    IS_KABYLAKE(i915) ||
499 	    IS_COFFEELAKE(i915) ||
500 	    IS_COMETLAKE(i915))
501 		wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN3,
502 				 GEN8_SAMPLER_POWER_BYPASS_DIS);
503 
504 	/* WaDisableSTUnitPowerOptimization:skl,bxt,kbl,glk,cfl */
505 	wa_mcr_masked_en(wal, HALF_SLICE_CHICKEN2, GEN8_ST_PO_DISABLE);
506 
507 	/*
508 	 * Supporting preemption with fine-granularity requires changes in the
509 	 * batch buffer programming. Since we can't break old userspace, we
510 	 * need to set our default preemption level to safe value. Userspace is
511 	 * still able to use more fine-grained preemption levels, since in
512 	 * WaEnablePreemptionGranularityControlByUMD we're whitelisting the
513 	 * per-ctx register. As such, WaDisable{3D,GPGPU}MidCmdPreemption are
514 	 * not real HW workarounds, but merely a way to start using preemption
515 	 * while maintaining old contract with userspace.
516 	 */
517 
518 	/* WaDisable3DMidCmdPreemption:skl,bxt,glk,cfl,[cnl] */
519 	wa_masked_dis(wal, GEN8_CS_CHICKEN1, GEN9_PREEMPT_3D_OBJECT_LEVEL);
520 
521 	/* WaDisableGPGPUMidCmdPreemption:skl,bxt,blk,cfl,[cnl] */
522 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
523 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
524 			    GEN9_PREEMPT_GPGPU_COMMAND_LEVEL);
525 
526 	/* WaClearHIZ_WM_CHICKEN3:bxt,glk */
527 	if (IS_GEN9_LP(i915))
528 		wa_masked_en(wal, GEN9_WM_CHICKEN3, GEN9_FACTOR_IN_CLR_VAL_HIZ);
529 }
530 
531 static void skl_tune_iz_hashing(struct intel_engine_cs *engine,
532 				struct i915_wa_list *wal)
533 {
534 	struct intel_gt *gt = engine->gt;
535 	u8 vals[3] = { 0, 0, 0 };
536 	unsigned int i;
537 
538 	for (i = 0; i < 3; i++) {
539 		u8 ss;
540 
541 		/*
542 		 * Only consider slices where one, and only one, subslice has 7
543 		 * EUs
544 		 */
545 		if (!is_power_of_2(gt->info.sseu.subslice_7eu[i]))
546 			continue;
547 
548 		/*
549 		 * subslice_7eu[i] != 0 (because of the check above) and
550 		 * ss_max == 4 (maximum number of subslices possible per slice)
551 		 *
552 		 * ->    0 <= ss <= 3;
553 		 */
554 		ss = ffs(gt->info.sseu.subslice_7eu[i]) - 1;
555 		vals[i] = 3 - ss;
556 	}
557 
558 	if (vals[0] == 0 && vals[1] == 0 && vals[2] == 0)
559 		return;
560 
561 	/* Tune IZ hashing. See intel_device_info_runtime_init() */
562 	wa_masked_field_set(wal, GEN7_GT_MODE,
563 			    GEN9_IZ_HASHING_MASK(2) |
564 			    GEN9_IZ_HASHING_MASK(1) |
565 			    GEN9_IZ_HASHING_MASK(0),
566 			    GEN9_IZ_HASHING(2, vals[2]) |
567 			    GEN9_IZ_HASHING(1, vals[1]) |
568 			    GEN9_IZ_HASHING(0, vals[0]));
569 }
570 
571 static void skl_ctx_workarounds_init(struct intel_engine_cs *engine,
572 				     struct i915_wa_list *wal)
573 {
574 	gen9_ctx_workarounds_init(engine, wal);
575 	skl_tune_iz_hashing(engine, wal);
576 }
577 
578 static void bxt_ctx_workarounds_init(struct intel_engine_cs *engine,
579 				     struct i915_wa_list *wal)
580 {
581 	gen9_ctx_workarounds_init(engine, wal);
582 
583 	/* WaDisableThreadStallDopClockGating:bxt */
584 	wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
585 			 STALL_DOP_GATING_DISABLE);
586 
587 	/* WaToEnableHwFixForPushConstHWBug:bxt */
588 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
589 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
590 }
591 
592 static void kbl_ctx_workarounds_init(struct intel_engine_cs *engine,
593 				     struct i915_wa_list *wal)
594 {
595 	struct drm_i915_private *i915 = engine->i915;
596 
597 	gen9_ctx_workarounds_init(engine, wal);
598 
599 	/* WaToEnableHwFixForPushConstHWBug:kbl */
600 	if (IS_KABYLAKE(i915) && IS_GRAPHICS_STEP(i915, STEP_C0, STEP_FOREVER))
601 		wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
602 			     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
603 
604 	/* WaDisableSbeCacheDispatchPortSharing:kbl */
605 	wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN1,
606 			 GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
607 }
608 
609 static void glk_ctx_workarounds_init(struct intel_engine_cs *engine,
610 				     struct i915_wa_list *wal)
611 {
612 	gen9_ctx_workarounds_init(engine, wal);
613 
614 	/* WaToEnableHwFixForPushConstHWBug:glk */
615 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
616 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
617 }
618 
619 static void cfl_ctx_workarounds_init(struct intel_engine_cs *engine,
620 				     struct i915_wa_list *wal)
621 {
622 	gen9_ctx_workarounds_init(engine, wal);
623 
624 	/* WaToEnableHwFixForPushConstHWBug:cfl */
625 	wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
626 		     GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
627 
628 	/* WaDisableSbeCacheDispatchPortSharing:cfl */
629 	wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN1,
630 			 GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
631 }
632 
633 static void icl_ctx_workarounds_init(struct intel_engine_cs *engine,
634 				     struct i915_wa_list *wal)
635 {
636 	/* Wa_1406697149 (WaDisableBankHangMode:icl) */
637 	wa_write(wal, GEN8_L3CNTLREG, GEN8_ERRDETBCTRL);
638 
639 	/* WaForceEnableNonCoherent:icl
640 	 * This is not the same workaround as in early Gen9 platforms, where
641 	 * lacking this could cause system hangs, but coherency performance
642 	 * overhead is high and only a few compute workloads really need it
643 	 * (the register is whitelisted in hardware now, so UMDs can opt in
644 	 * for coherency if they have a good reason).
645 	 */
646 	wa_mcr_masked_en(wal, ICL_HDC_MODE, HDC_FORCE_NON_COHERENT);
647 
648 	/* WaEnableFloatBlendOptimization:icl */
649 	wa_mcr_add(wal, GEN10_CACHE_MODE_SS, 0,
650 		   _MASKED_BIT_ENABLE(FLOAT_BLEND_OPTIMIZATION_ENABLE),
651 		   0 /* write-only, so skip validation */,
652 		   true);
653 
654 	/* WaDisableGPGPUMidThreadPreemption:icl */
655 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
656 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
657 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
658 
659 	/* allow headerless messages for preemptible GPGPU context */
660 	wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
661 			 GEN11_SAMPLER_ENABLE_HEADLESS_MSG);
662 
663 	/* Wa_1604278689:icl,ehl */
664 	wa_write(wal, IVB_FBC_RT_BASE, 0xFFFFFFFF & ~ILK_FBC_RT_VALID);
665 	wa_write_clr_set(wal, IVB_FBC_RT_BASE_UPPER,
666 			 0,
667 			 0xFFFFFFFF);
668 
669 	/* Wa_1406306137:icl,ehl */
670 	wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, GEN11_DIS_PICK_2ND_EU);
671 }
672 
673 /*
674  * These settings aren't actually workarounds, but general tuning settings that
675  * need to be programmed on dg2 platform.
676  */
677 static void dg2_ctx_gt_tuning_init(struct intel_engine_cs *engine,
678 				   struct i915_wa_list *wal)
679 {
680 	wa_mcr_masked_en(wal, CHICKEN_RASTER_2, TBIMR_FAST_CLIP);
681 	wa_mcr_write_clr_set(wal, XEHP_L3SQCREG5, L3_PWM_TIMER_INIT_VAL_MASK,
682 			     REG_FIELD_PREP(L3_PWM_TIMER_INIT_VAL_MASK, 0x7f));
683 	wa_mcr_write_clr_set(wal, XEHP_FF_MODE2, FF_MODE2_TDS_TIMER_MASK,
684 			     FF_MODE2_TDS_TIMER_128);
685 }
686 
687 static void gen12_ctx_workarounds_init(struct intel_engine_cs *engine,
688 				       struct i915_wa_list *wal)
689 {
690 	struct drm_i915_private *i915 = engine->i915;
691 
692 	/*
693 	 * Wa_1409142259:tgl,dg1,adl-p
694 	 * Wa_1409347922:tgl,dg1,adl-p
695 	 * Wa_1409252684:tgl,dg1,adl-p
696 	 * Wa_1409217633:tgl,dg1,adl-p
697 	 * Wa_1409207793:tgl,dg1,adl-p
698 	 * Wa_1409178076:tgl,dg1,adl-p
699 	 * Wa_1408979724:tgl,dg1,adl-p
700 	 * Wa_14010443199:tgl,rkl,dg1,adl-p
701 	 * Wa_14010698770:tgl,rkl,dg1,adl-s,adl-p
702 	 * Wa_1409342910:tgl,rkl,dg1,adl-s,adl-p
703 	 */
704 	wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
705 		     GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
706 
707 	/* WaDisableGPGPUMidThreadPreemption:gen12 */
708 	wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
709 			    GEN9_PREEMPT_GPGPU_LEVEL_MASK,
710 			    GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
711 
712 	/*
713 	 * Wa_16011163337 - GS_TIMER
714 	 *
715 	 * TDS_TIMER: Although some platforms refer to it as Wa_1604555607, we
716 	 * need to program it even on those that don't explicitly list that
717 	 * workaround.
718 	 *
719 	 * Note that the programming of GEN12_FF_MODE2 is further modified
720 	 * according to the FF_MODE2 guidance given by Wa_1608008084.
721 	 * Wa_1608008084 tells us the FF_MODE2 register will return the wrong
722 	 * value when read from the CPU.
723 	 *
724 	 * The default value for this register is zero for all fields.
725 	 * So instead of doing a RMW we should just write the desired values
726 	 * for TDS and GS timers. Note that since the readback can't be trusted,
727 	 * the clear mask is just set to ~0 to make sure other bits are not
728 	 * inadvertently set. For the same reason read verification is ignored.
729 	 */
730 	wa_add(wal,
731 	       GEN12_FF_MODE2,
732 	       ~0,
733 	       FF_MODE2_TDS_TIMER_128 | FF_MODE2_GS_TIMER_224,
734 	       0, false);
735 
736 	if (!IS_DG1(i915)) {
737 		/* Wa_1806527549 */
738 		wa_masked_en(wal, HIZ_CHICKEN, HZ_DEPTH_TEST_LE_GE_OPT_DISABLE);
739 
740 		/* Wa_1606376872 */
741 		wa_masked_en(wal, COMMON_SLICE_CHICKEN4, DISABLE_TDC_LOAD_BALANCING_CALC);
742 	}
743 }
744 
745 static void dg1_ctx_workarounds_init(struct intel_engine_cs *engine,
746 				     struct i915_wa_list *wal)
747 {
748 	gen12_ctx_workarounds_init(engine, wal);
749 
750 	/* Wa_1409044764 */
751 	wa_masked_dis(wal, GEN11_COMMON_SLICE_CHICKEN3,
752 		      DG1_FLOAT_POINT_BLEND_OPT_STRICT_MODE_EN);
753 
754 	/* Wa_22010493298 */
755 	wa_masked_en(wal, HIZ_CHICKEN,
756 		     DG1_HZ_READ_SUPPRESSION_OPTIMIZATION_DISABLE);
757 }
758 
759 static void dg2_ctx_workarounds_init(struct intel_engine_cs *engine,
760 				     struct i915_wa_list *wal)
761 {
762 	dg2_ctx_gt_tuning_init(engine, wal);
763 
764 	/* Wa_16013271637:dg2 */
765 	wa_mcr_masked_en(wal, XEHP_SLICE_COMMON_ECO_CHICKEN1,
766 			 MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
767 
768 	/* Wa_14014947963:dg2 */
769 	wa_masked_field_set(wal, VF_PREEMPTION, PREEMPTION_VERTEX_COUNT, 0x4000);
770 
771 	/* Wa_18018764978:dg2 */
772 	wa_mcr_masked_en(wal, XEHP_PSS_MODE2, SCOREBOARD_STALL_FLUSH_CONTROL);
773 
774 	/* Wa_18019271663:dg2 */
775 	wa_masked_en(wal, CACHE_MODE_1, MSAA_OPTIMIZATION_REDUC_DISABLE);
776 
777 	/* Wa_14019877138:dg2 */
778 	wa_mcr_masked_en(wal, XEHP_PSS_CHICKEN, FD_END_COLLECT);
779 }
780 
781 static void xelpg_ctx_gt_tuning_init(struct intel_engine_cs *engine,
782 				     struct i915_wa_list *wal)
783 {
784 	struct intel_gt *gt = engine->gt;
785 
786 	dg2_ctx_gt_tuning_init(engine, wal);
787 
788 	/*
789 	 * Due to Wa_16014892111, the DRAW_WATERMARK tuning must be done in
790 	 * gen12_emit_indirect_ctx_rcs() rather than here on some early
791 	 * steppings.
792 	 */
793 	if (!(IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
794 	      IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)))
795 		wa_add(wal, DRAW_WATERMARK, VERT_WM_VAL, 0x3FF, 0, false);
796 }
797 
798 static void xelpg_ctx_workarounds_init(struct intel_engine_cs *engine,
799 				       struct i915_wa_list *wal)
800 {
801 	struct intel_gt *gt = engine->gt;
802 
803 	xelpg_ctx_gt_tuning_init(engine, wal);
804 
805 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
806 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
807 		/* Wa_14014947963 */
808 		wa_masked_field_set(wal, VF_PREEMPTION,
809 				    PREEMPTION_VERTEX_COUNT, 0x4000);
810 
811 		/* Wa_16013271637 */
812 		wa_mcr_masked_en(wal, XEHP_SLICE_COMMON_ECO_CHICKEN1,
813 				 MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
814 
815 		/* Wa_18019627453 */
816 		wa_mcr_masked_en(wal, VFLSKPD, VF_PREFETCH_TLB_DIS);
817 
818 		/* Wa_18018764978 */
819 		wa_mcr_masked_en(wal, XEHP_PSS_MODE2, SCOREBOARD_STALL_FLUSH_CONTROL);
820 	}
821 
822 	/* Wa_18019271663 */
823 	wa_masked_en(wal, CACHE_MODE_1, MSAA_OPTIMIZATION_REDUC_DISABLE);
824 
825 	/* Wa_14019877138 */
826 	wa_mcr_masked_en(wal, XEHP_PSS_CHICKEN, FD_END_COLLECT);
827 }
828 
829 static void fakewa_disable_nestedbb_mode(struct intel_engine_cs *engine,
830 					 struct i915_wa_list *wal)
831 {
832 	/*
833 	 * This is a "fake" workaround defined by software to ensure we
834 	 * maintain reliable, backward-compatible behavior for userspace with
835 	 * regards to how nested MI_BATCH_BUFFER_START commands are handled.
836 	 *
837 	 * The per-context setting of MI_MODE[12] determines whether the bits
838 	 * of a nested MI_BATCH_BUFFER_START instruction should be interpreted
839 	 * in the traditional manner or whether they should instead use a new
840 	 * tgl+ meaning that breaks backward compatibility, but allows nesting
841 	 * into 3rd-level batchbuffers.  When this new capability was first
842 	 * added in TGL, it remained off by default unless a context
843 	 * intentionally opted in to the new behavior.  However Xe_HPG now
844 	 * flips this on by default and requires that we explicitly opt out if
845 	 * we don't want the new behavior.
846 	 *
847 	 * From a SW perspective, we want to maintain the backward-compatible
848 	 * behavior for userspace, so we'll apply a fake workaround to set it
849 	 * back to the legacy behavior on platforms where the hardware default
850 	 * is to break compatibility.  At the moment there is no Linux
851 	 * userspace that utilizes third-level batchbuffers, so this will avoid
852 	 * userspace from needing to make any changes.  using the legacy
853 	 * meaning is the correct thing to do.  If/when we have userspace
854 	 * consumers that want to utilize third-level batch nesting, we can
855 	 * provide a context parameter to allow them to opt-in.
856 	 */
857 	wa_masked_dis(wal, RING_MI_MODE(engine->mmio_base), TGL_NESTED_BB_EN);
858 }
859 
860 static void gen12_ctx_gt_mocs_init(struct intel_engine_cs *engine,
861 				   struct i915_wa_list *wal)
862 {
863 	u8 mocs;
864 
865 	/*
866 	 * Some blitter commands do not have a field for MOCS, those
867 	 * commands will use MOCS index pointed by BLIT_CCTL.
868 	 * BLIT_CCTL registers are needed to be programmed to un-cached.
869 	 */
870 	if (engine->class == COPY_ENGINE_CLASS) {
871 		mocs = engine->gt->mocs.uc_index;
872 		wa_write_clr_set(wal,
873 				 BLIT_CCTL(engine->mmio_base),
874 				 BLIT_CCTL_MASK,
875 				 BLIT_CCTL_MOCS(mocs, mocs));
876 	}
877 }
878 
879 /*
880  * gen12_ctx_gt_fake_wa_init() aren't programmingan official workaround
881  * defined by the hardware team, but it programming general context registers.
882  * Adding those context register programming in context workaround
883  * allow us to use the wa framework for proper application and validation.
884  */
885 static void
886 gen12_ctx_gt_fake_wa_init(struct intel_engine_cs *engine,
887 			  struct i915_wa_list *wal)
888 {
889 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
890 		fakewa_disable_nestedbb_mode(engine, wal);
891 
892 	gen12_ctx_gt_mocs_init(engine, wal);
893 }
894 
895 static void
896 __intel_engine_init_ctx_wa(struct intel_engine_cs *engine,
897 			   struct i915_wa_list *wal,
898 			   const char *name)
899 {
900 	struct drm_i915_private *i915 = engine->i915;
901 
902 	wa_init_start(wal, engine->gt, name, engine->name);
903 
904 	/* Applies to all engines */
905 	/*
906 	 * Fake workarounds are not the actual workaround but
907 	 * programming of context registers using workaround framework.
908 	 */
909 	if (GRAPHICS_VER(i915) >= 12)
910 		gen12_ctx_gt_fake_wa_init(engine, wal);
911 
912 	if (engine->class != RENDER_CLASS)
913 		goto done;
914 
915 	if (IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 70), IP_VER(12, 74)))
916 		xelpg_ctx_workarounds_init(engine, wal);
917 	else if (IS_DG2(i915))
918 		dg2_ctx_workarounds_init(engine, wal);
919 	else if (IS_DG1(i915))
920 		dg1_ctx_workarounds_init(engine, wal);
921 	else if (GRAPHICS_VER(i915) == 12)
922 		gen12_ctx_workarounds_init(engine, wal);
923 	else if (GRAPHICS_VER(i915) == 11)
924 		icl_ctx_workarounds_init(engine, wal);
925 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
926 		cfl_ctx_workarounds_init(engine, wal);
927 	else if (IS_GEMINILAKE(i915))
928 		glk_ctx_workarounds_init(engine, wal);
929 	else if (IS_KABYLAKE(i915))
930 		kbl_ctx_workarounds_init(engine, wal);
931 	else if (IS_BROXTON(i915))
932 		bxt_ctx_workarounds_init(engine, wal);
933 	else if (IS_SKYLAKE(i915))
934 		skl_ctx_workarounds_init(engine, wal);
935 	else if (IS_CHERRYVIEW(i915))
936 		chv_ctx_workarounds_init(engine, wal);
937 	else if (IS_BROADWELL(i915))
938 		bdw_ctx_workarounds_init(engine, wal);
939 	else if (GRAPHICS_VER(i915) == 7)
940 		gen7_ctx_workarounds_init(engine, wal);
941 	else if (GRAPHICS_VER(i915) == 6)
942 		gen6_ctx_workarounds_init(engine, wal);
943 	else if (GRAPHICS_VER(i915) < 8)
944 		;
945 	else
946 		MISSING_CASE(GRAPHICS_VER(i915));
947 
948 done:
949 	wa_init_finish(wal);
950 }
951 
952 void intel_engine_init_ctx_wa(struct intel_engine_cs *engine)
953 {
954 	__intel_engine_init_ctx_wa(engine, &engine->ctx_wa_list, "context");
955 }
956 
957 int intel_engine_emit_ctx_wa(struct i915_request *rq)
958 {
959 	struct i915_wa_list *wal = &rq->engine->ctx_wa_list;
960 	struct intel_uncore *uncore = rq->engine->uncore;
961 	enum forcewake_domains fw;
962 	unsigned long flags;
963 	struct i915_wa *wa;
964 	unsigned int i;
965 	u32 *cs;
966 	int ret;
967 
968 	if (wal->count == 0)
969 		return 0;
970 
971 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
972 	if (ret)
973 		return ret;
974 
975 	cs = intel_ring_begin(rq, (wal->count * 2 + 2));
976 	if (IS_ERR(cs))
977 		return PTR_ERR(cs);
978 
979 	fw = wal_get_fw_for_rmw(uncore, wal);
980 
981 	intel_gt_mcr_lock(wal->gt, &flags);
982 	spin_lock(&uncore->lock);
983 	intel_uncore_forcewake_get__locked(uncore, fw);
984 
985 	*cs++ = MI_LOAD_REGISTER_IMM(wal->count);
986 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
987 		u32 val;
988 
989 		/* Skip reading the register if it's not really needed */
990 		if (wa->masked_reg || (wa->clr | wa->set) == U32_MAX) {
991 			val = wa->set;
992 		} else {
993 			val = wa->is_mcr ?
994 				intel_gt_mcr_read_any_fw(wal->gt, wa->mcr_reg) :
995 				intel_uncore_read_fw(uncore, wa->reg);
996 			val &= ~wa->clr;
997 			val |= wa->set;
998 		}
999 
1000 		*cs++ = i915_mmio_reg_offset(wa->reg);
1001 		*cs++ = val;
1002 	}
1003 	*cs++ = MI_NOOP;
1004 
1005 	intel_uncore_forcewake_put__locked(uncore, fw);
1006 	spin_unlock(&uncore->lock);
1007 	intel_gt_mcr_unlock(wal->gt, flags);
1008 
1009 	intel_ring_advance(rq, cs);
1010 
1011 	ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
1012 	if (ret)
1013 		return ret;
1014 
1015 	return 0;
1016 }
1017 
1018 static void
1019 gen4_gt_workarounds_init(struct intel_gt *gt,
1020 			 struct i915_wa_list *wal)
1021 {
1022 	/* WaDisable_RenderCache_OperationalFlush:gen4,ilk */
1023 	wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
1024 }
1025 
1026 static void
1027 g4x_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1028 {
1029 	gen4_gt_workarounds_init(gt, wal);
1030 
1031 	/* WaDisableRenderCachePipelinedFlush:g4x,ilk */
1032 	wa_masked_en(wal, CACHE_MODE_0, CM0_PIPELINED_RENDER_FLUSH_DISABLE);
1033 }
1034 
1035 static void
1036 ilk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1037 {
1038 	g4x_gt_workarounds_init(gt, wal);
1039 
1040 	wa_masked_en(wal, _3D_CHICKEN2, _3D_CHICKEN2_WM_READ_PIPELINED);
1041 }
1042 
1043 static void
1044 snb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1045 {
1046 }
1047 
1048 static void
1049 ivb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1050 {
1051 	/* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
1052 	wa_masked_dis(wal,
1053 		      GEN7_COMMON_SLICE_CHICKEN1,
1054 		      GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
1055 
1056 	/* WaApplyL3ControlAndL3ChickenMode:ivb */
1057 	wa_write(wal, GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
1058 	wa_write(wal, GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
1059 
1060 	/* WaForceL3Serialization:ivb */
1061 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
1062 }
1063 
1064 static void
1065 vlv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1066 {
1067 	/* WaForceL3Serialization:vlv */
1068 	wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
1069 
1070 	/*
1071 	 * WaIncreaseL3CreditsForVLVB0:vlv
1072 	 * This is the hardware default actually.
1073 	 */
1074 	wa_write(wal, GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
1075 }
1076 
1077 static void
1078 hsw_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1079 {
1080 	/* L3 caching of data atomics doesn't work -- disable it. */
1081 	wa_write(wal, HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
1082 
1083 	wa_add(wal,
1084 	       HSW_ROW_CHICKEN3, 0,
1085 	       _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
1086 	       0 /* XXX does this reg exist? */, true);
1087 
1088 	/* WaVSRefCountFullforceMissDisable:hsw */
1089 	wa_write_clr(wal, GEN7_FF_THREAD_MODE, GEN7_FF_VS_REF_CNT_FFME);
1090 }
1091 
1092 static void
1093 gen9_wa_init_mcr(struct drm_i915_private *i915, struct i915_wa_list *wal)
1094 {
1095 	const struct sseu_dev_info *sseu = &to_gt(i915)->info.sseu;
1096 	unsigned int slice, subslice;
1097 	u32 mcr, mcr_mask;
1098 
1099 	GEM_BUG_ON(GRAPHICS_VER(i915) != 9);
1100 
1101 	/*
1102 	 * WaProgramMgsrForCorrectSliceSpecificMmioReads:gen9,glk,kbl,cml
1103 	 * Before any MMIO read into slice/subslice specific registers, MCR
1104 	 * packet control register needs to be programmed to point to any
1105 	 * enabled s/ss pair. Otherwise, incorrect values will be returned.
1106 	 * This means each subsequent MMIO read will be forwarded to an
1107 	 * specific s/ss combination, but this is OK since these registers
1108 	 * are consistent across s/ss in almost all cases. In the rare
1109 	 * occasions, such as INSTDONE, where this value is dependent
1110 	 * on s/ss combo, the read should be done with read_subslice_reg.
1111 	 */
1112 	slice = ffs(sseu->slice_mask) - 1;
1113 	GEM_BUG_ON(slice >= ARRAY_SIZE(sseu->subslice_mask.hsw));
1114 	subslice = ffs(intel_sseu_get_hsw_subslices(sseu, slice));
1115 	GEM_BUG_ON(!subslice);
1116 	subslice--;
1117 
1118 	/*
1119 	 * We use GEN8_MCR..() macros to calculate the |mcr| value for
1120 	 * Gen9 to address WaProgramMgsrForCorrectSliceSpecificMmioReads
1121 	 */
1122 	mcr = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
1123 	mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
1124 
1125 	drm_dbg(&i915->drm, "MCR slice:%d/subslice:%d = %x\n", slice, subslice, mcr);
1126 
1127 	wa_write_clr_set(wal, GEN8_MCR_SELECTOR, mcr_mask, mcr);
1128 }
1129 
1130 static void
1131 gen9_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1132 {
1133 	struct drm_i915_private *i915 = gt->i915;
1134 
1135 	/* WaProgramMgsrForCorrectSliceSpecificMmioReads:glk,kbl,cml,gen9 */
1136 	gen9_wa_init_mcr(i915, wal);
1137 
1138 	/* WaDisableKillLogic:bxt,skl,kbl */
1139 	if (!IS_COFFEELAKE(i915) && !IS_COMETLAKE(i915))
1140 		wa_write_or(wal,
1141 			    GAM_ECOCHK,
1142 			    ECOCHK_DIS_TLB);
1143 
1144 	if (HAS_LLC(i915)) {
1145 		/* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
1146 		 *
1147 		 * Must match Display Engine. See
1148 		 * WaCompressedResourceDisplayNewHashMode.
1149 		 */
1150 		wa_write_or(wal,
1151 			    MMCD_MISC_CTRL,
1152 			    MMCD_PCLA | MMCD_HOTSPOT_EN);
1153 	}
1154 
1155 	/* WaDisableHDCInvalidation:skl,bxt,kbl,cfl */
1156 	wa_write_or(wal,
1157 		    GAM_ECOCHK,
1158 		    BDW_DISABLE_HDC_INVALIDATION);
1159 }
1160 
1161 static void
1162 skl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1163 {
1164 	gen9_gt_workarounds_init(gt, wal);
1165 
1166 	/* WaDisableGafsUnitClkGating:skl */
1167 	wa_write_or(wal,
1168 		    GEN7_UCGCTL4,
1169 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1170 
1171 	/* WaInPlaceDecompressionHang:skl */
1172 	if (IS_SKYLAKE(gt->i915) && IS_GRAPHICS_STEP(gt->i915, STEP_A0, STEP_H0))
1173 		wa_write_or(wal,
1174 			    GEN9_GAMT_ECO_REG_RW_IA,
1175 			    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1176 }
1177 
1178 static void
1179 kbl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1180 {
1181 	gen9_gt_workarounds_init(gt, wal);
1182 
1183 	/* WaDisableDynamicCreditSharing:kbl */
1184 	if (IS_KABYLAKE(gt->i915) && IS_GRAPHICS_STEP(gt->i915, 0, STEP_C0))
1185 		wa_write_or(wal,
1186 			    GAMT_CHKN_BIT_REG,
1187 			    GAMT_CHKN_DISABLE_DYNAMIC_CREDIT_SHARING);
1188 
1189 	/* WaDisableGafsUnitClkGating:kbl */
1190 	wa_write_or(wal,
1191 		    GEN7_UCGCTL4,
1192 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1193 
1194 	/* WaInPlaceDecompressionHang:kbl */
1195 	wa_write_or(wal,
1196 		    GEN9_GAMT_ECO_REG_RW_IA,
1197 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1198 }
1199 
1200 static void
1201 glk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1202 {
1203 	gen9_gt_workarounds_init(gt, wal);
1204 }
1205 
1206 static void
1207 cfl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1208 {
1209 	gen9_gt_workarounds_init(gt, wal);
1210 
1211 	/* WaDisableGafsUnitClkGating:cfl */
1212 	wa_write_or(wal,
1213 		    GEN7_UCGCTL4,
1214 		    GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1215 
1216 	/* WaInPlaceDecompressionHang:cfl */
1217 	wa_write_or(wal,
1218 		    GEN9_GAMT_ECO_REG_RW_IA,
1219 		    GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1220 }
1221 
1222 static void __set_mcr_steering(struct i915_wa_list *wal,
1223 			       i915_reg_t steering_reg,
1224 			       unsigned int slice, unsigned int subslice)
1225 {
1226 	u32 mcr, mcr_mask;
1227 
1228 	mcr = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1229 	mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1230 
1231 	wa_write_clr_set(wal, steering_reg, mcr_mask, mcr);
1232 }
1233 
1234 static void debug_dump_steering(struct intel_gt *gt)
1235 {
1236 	struct drm_printer p = drm_dbg_printer(&gt->i915->drm, DRM_UT_DRIVER,
1237 					       "MCR Steering:");
1238 
1239 	if (drm_debug_enabled(DRM_UT_DRIVER))
1240 		intel_gt_mcr_report_steering(&p, gt, false);
1241 }
1242 
1243 static void __add_mcr_wa(struct intel_gt *gt, struct i915_wa_list *wal,
1244 			 unsigned int slice, unsigned int subslice)
1245 {
1246 	__set_mcr_steering(wal, GEN8_MCR_SELECTOR, slice, subslice);
1247 
1248 	gt->default_steering.groupid = slice;
1249 	gt->default_steering.instanceid = subslice;
1250 
1251 	debug_dump_steering(gt);
1252 }
1253 
1254 static void
1255 icl_wa_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1256 {
1257 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1258 	unsigned int subslice;
1259 
1260 	GEM_BUG_ON(GRAPHICS_VER(gt->i915) < 11);
1261 	GEM_BUG_ON(hweight8(sseu->slice_mask) > 1);
1262 
1263 	/*
1264 	 * Although a platform may have subslices, we need to always steer
1265 	 * reads to the lowest instance that isn't fused off.  When Render
1266 	 * Power Gating is enabled, grabbing forcewake will only power up a
1267 	 * single subslice (the "minconfig") if there isn't a real workload
1268 	 * that needs to be run; this means that if we steer register reads to
1269 	 * one of the higher subslices, we run the risk of reading back 0's or
1270 	 * random garbage.
1271 	 */
1272 	subslice = __ffs(intel_sseu_get_hsw_subslices(sseu, 0));
1273 
1274 	/*
1275 	 * If the subslice we picked above also steers us to a valid L3 bank,
1276 	 * then we can just rely on the default steering and won't need to
1277 	 * worry about explicitly re-steering L3BANK reads later.
1278 	 */
1279 	if (gt->info.l3bank_mask & BIT(subslice))
1280 		gt->steering_table[L3BANK] = NULL;
1281 
1282 	__add_mcr_wa(gt, wal, 0, subslice);
1283 }
1284 
1285 static void
1286 xehp_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1287 {
1288 	const struct sseu_dev_info *sseu = &gt->info.sseu;
1289 	unsigned long slice, subslice = 0, slice_mask = 0;
1290 	u32 lncf_mask = 0;
1291 	int i;
1292 
1293 	/*
1294 	 * On Xe_HP the steering increases in complexity. There are now several
1295 	 * more units that require steering and we're not guaranteed to be able
1296 	 * to find a common setting for all of them. These are:
1297 	 * - GSLICE (fusable)
1298 	 * - DSS (sub-unit within gslice; fusable)
1299 	 * - L3 Bank (fusable)
1300 	 * - MSLICE (fusable)
1301 	 * - LNCF (sub-unit within mslice; always present if mslice is present)
1302 	 *
1303 	 * We'll do our default/implicit steering based on GSLICE (in the
1304 	 * sliceid field) and DSS (in the subsliceid field).  If we can
1305 	 * find overlap between the valid MSLICE and/or LNCF values with
1306 	 * a suitable GSLICE, then we can just re-use the default value and
1307 	 * skip and explicit steering at runtime.
1308 	 *
1309 	 * We only need to look for overlap between GSLICE/MSLICE/LNCF to find
1310 	 * a valid sliceid value.  DSS steering is the only type of steering
1311 	 * that utilizes the 'subsliceid' bits.
1312 	 *
1313 	 * Also note that, even though the steering domain is called "GSlice"
1314 	 * and it is encoded in the register using the gslice format, the spec
1315 	 * says that the combined (geometry | compute) fuse should be used to
1316 	 * select the steering.
1317 	 */
1318 
1319 	/* Find the potential gslice candidates */
1320 	slice_mask = intel_slicemask_from_xehp_dssmask(sseu->subslice_mask,
1321 						       GEN_DSS_PER_GSLICE);
1322 
1323 	/*
1324 	 * Find the potential LNCF candidates.  Either LNCF within a valid
1325 	 * mslice is fine.
1326 	 */
1327 	for_each_set_bit(i, &gt->info.mslice_mask, GEN12_MAX_MSLICES)
1328 		lncf_mask |= (0x3 << (i * 2));
1329 
1330 	/*
1331 	 * Are there any sliceid values that work for both GSLICE and LNCF
1332 	 * steering?
1333 	 */
1334 	if (slice_mask & lncf_mask) {
1335 		slice_mask &= lncf_mask;
1336 		gt->steering_table[LNCF] = NULL;
1337 	}
1338 
1339 	/* How about sliceid values that also work for MSLICE steering? */
1340 	if (slice_mask & gt->info.mslice_mask) {
1341 		slice_mask &= gt->info.mslice_mask;
1342 		gt->steering_table[MSLICE] = NULL;
1343 	}
1344 
1345 	slice = __ffs(slice_mask);
1346 	subslice = intel_sseu_find_first_xehp_dss(sseu, GEN_DSS_PER_GSLICE, slice) %
1347 		GEN_DSS_PER_GSLICE;
1348 
1349 	__add_mcr_wa(gt, wal, slice, subslice);
1350 
1351 	/*
1352 	 * SQIDI ranges are special because they use different steering
1353 	 * registers than everything else we work with.  On XeHP SDV and
1354 	 * DG2-G10, any value in the steering registers will work fine since
1355 	 * all instances are present, but DG2-G11 only has SQIDI instances at
1356 	 * ID's 2 and 3, so we need to steer to one of those.  For simplicity
1357 	 * we'll just steer to a hardcoded "2" since that value will work
1358 	 * everywhere.
1359 	 */
1360 	__set_mcr_steering(wal, MCFG_MCR_SELECTOR, 0, 2);
1361 	__set_mcr_steering(wal, SF_MCR_SELECTOR, 0, 2);
1362 
1363 	/*
1364 	 * On DG2, GAM registers have a dedicated steering control register
1365 	 * and must always be programmed to a hardcoded groupid of "1."
1366 	 */
1367 	if (IS_DG2(gt->i915))
1368 		__set_mcr_steering(wal, GAM_MCR_SELECTOR, 1, 0);
1369 }
1370 
1371 static void
1372 icl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1373 {
1374 	struct drm_i915_private *i915 = gt->i915;
1375 
1376 	icl_wa_init_mcr(gt, wal);
1377 
1378 	/* WaModifyGamTlbPartitioning:icl */
1379 	wa_write_clr_set(wal,
1380 			 GEN11_GACB_PERF_CTRL,
1381 			 GEN11_HASH_CTRL_MASK,
1382 			 GEN11_HASH_CTRL_BIT0 | GEN11_HASH_CTRL_BIT4);
1383 
1384 	/* Wa_1405766107:icl
1385 	 * Formerly known as WaCL2SFHalfMaxAlloc
1386 	 */
1387 	wa_write_or(wal,
1388 		    GEN11_LSN_UNSLCVC,
1389 		    GEN11_LSN_UNSLCVC_GAFS_HALF_SF_MAXALLOC |
1390 		    GEN11_LSN_UNSLCVC_GAFS_HALF_CL2_MAXALLOC);
1391 
1392 	/* Wa_220166154:icl
1393 	 * Formerly known as WaDisCtxReload
1394 	 */
1395 	wa_write_or(wal,
1396 		    GEN8_GAMW_ECO_DEV_RW_IA,
1397 		    GAMW_ECO_DEV_CTX_RELOAD_DISABLE);
1398 
1399 	/* Wa_1406463099:icl
1400 	 * Formerly known as WaGamTlbPendError
1401 	 */
1402 	wa_write_or(wal,
1403 		    GAMT_CHKN_BIT_REG,
1404 		    GAMT_CHKN_DISABLE_L3_COH_PIPE);
1405 
1406 	/*
1407 	 * Wa_1408615072:icl,ehl  (vsunit)
1408 	 * Wa_1407596294:icl,ehl  (hsunit)
1409 	 */
1410 	wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1411 		    VSUNIT_CLKGATE_DIS | HSUNIT_CLKGATE_DIS);
1412 
1413 	/* Wa_1407352427:icl,ehl */
1414 	wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1415 		    PSDUNIT_CLKGATE_DIS);
1416 
1417 	/* Wa_1406680159:icl,ehl */
1418 	wa_mcr_write_or(wal,
1419 			GEN11_SUBSLICE_UNIT_LEVEL_CLKGATE,
1420 			GWUNIT_CLKGATE_DIS);
1421 
1422 	/* Wa_1607087056:icl,ehl,jsl */
1423 	if (IS_ICELAKE(i915) ||
1424 		((IS_JASPERLAKE(i915) || IS_ELKHARTLAKE(i915)) &&
1425 		IS_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)))
1426 		wa_write_or(wal,
1427 			    GEN11_SLICE_UNIT_LEVEL_CLKGATE,
1428 			    L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1429 
1430 	/*
1431 	 * This is not a documented workaround, but rather an optimization
1432 	 * to reduce sampler power.
1433 	 */
1434 	wa_mcr_write_clr(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1435 }
1436 
1437 /*
1438  * Though there are per-engine instances of these registers,
1439  * they retain their value through engine resets and should
1440  * only be provided on the GT workaround list rather than
1441  * the engine-specific workaround list.
1442  */
1443 static void
1444 wa_14011060649(struct intel_gt *gt, struct i915_wa_list *wal)
1445 {
1446 	struct intel_engine_cs *engine;
1447 	int id;
1448 
1449 	for_each_engine(engine, gt, id) {
1450 		if (engine->class != VIDEO_DECODE_CLASS ||
1451 		    (engine->instance % 2))
1452 			continue;
1453 
1454 		wa_write_or(wal, VDBOX_CGCTL3F10(engine->mmio_base),
1455 			    IECPUNIT_CLKGATE_DIS);
1456 	}
1457 }
1458 
1459 static void
1460 gen12_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1461 {
1462 	icl_wa_init_mcr(gt, wal);
1463 
1464 	/* Wa_14011060649:tgl,rkl,dg1,adl-s,adl-p */
1465 	wa_14011060649(gt, wal);
1466 
1467 	/* Wa_14011059788:tgl,rkl,adl-s,dg1,adl-p */
1468 	wa_mcr_write_or(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1469 
1470 	/*
1471 	 * Wa_14015795083
1472 	 *
1473 	 * Firmware on some gen12 platforms locks the MISCCPCTL register,
1474 	 * preventing i915 from modifying it for this workaround.  Skip the
1475 	 * readback verification for this workaround on debug builds; if the
1476 	 * workaround doesn't stick due to firmware behavior, it's not an error
1477 	 * that we want CI to flag.
1478 	 */
1479 	wa_add(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE,
1480 	       0, 0, false);
1481 }
1482 
1483 static void
1484 dg1_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1485 {
1486 	gen12_gt_workarounds_init(gt, wal);
1487 
1488 	/* Wa_1409420604:dg1 */
1489 	wa_mcr_write_or(wal, SUBSLICE_UNIT_LEVEL_CLKGATE2,
1490 			CPSSUNIT_CLKGATE_DIS);
1491 
1492 	/* Wa_1408615072:dg1 */
1493 	/* Empirical testing shows this register is unaffected by engine reset. */
1494 	wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2, VSUNIT_CLKGATE_DIS_TGL);
1495 }
1496 
1497 static void
1498 dg2_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1499 {
1500 	xehp_init_mcr(gt, wal);
1501 
1502 	/* Wa_14011060649:dg2 */
1503 	wa_14011060649(gt, wal);
1504 
1505 	if (IS_DG2_G10(gt->i915)) {
1506 		/* Wa_22010523718:dg2 */
1507 		wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1508 			    CG3DDISCFEG_CLKGATE_DIS);
1509 
1510 		/* Wa_14011006942:dg2 */
1511 		wa_mcr_write_or(wal, GEN11_SUBSLICE_UNIT_LEVEL_CLKGATE,
1512 				DSS_ROUTER_CLKGATE_DIS);
1513 	}
1514 
1515 	/* Wa_14014830051:dg2 */
1516 	wa_mcr_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1517 
1518 	/*
1519 	 * Wa_14015795083
1520 	 * Skip verification for possibly locked register.
1521 	 */
1522 	wa_add(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE,
1523 	       0, 0, false);
1524 
1525 	/* Wa_18018781329 */
1526 	wa_mcr_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
1527 	wa_mcr_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
1528 	wa_mcr_write_or(wal, XEHP_VDBX_MOD_CTRL, FORCE_MISS_FTLB);
1529 	wa_mcr_write_or(wal, XEHP_VEBX_MOD_CTRL, FORCE_MISS_FTLB);
1530 
1531 	/* Wa_1509235366:dg2 */
1532 	wa_mcr_write_or(wal, XEHP_GAMCNTRL_CTRL,
1533 			INVALIDATION_BROADCAST_MODE_DIS | GLOBAL_INVALIDATION_MODE);
1534 
1535 	/* Wa_14010648519:dg2 */
1536 	wa_mcr_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
1537 }
1538 
1539 static void
1540 xelpg_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1541 {
1542 	/* Wa_14018575942 / Wa_18018781329 */
1543 	wa_mcr_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
1544 	wa_mcr_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
1545 
1546 	/* Wa_22016670082 */
1547 	wa_write_or(wal, GEN12_SQCNT1, GEN12_STRICT_RAR_ENABLE);
1548 
1549 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
1550 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
1551 		/* Wa_14014830051 */
1552 		wa_mcr_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1553 
1554 		/* Wa_14015795083 */
1555 		wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1556 	}
1557 
1558 	/*
1559 	 * Unlike older platforms, we no longer setup implicit steering here;
1560 	 * all MCR accesses are explicitly steered.
1561 	 */
1562 	debug_dump_steering(gt);
1563 }
1564 
1565 static void
1566 wa_16021867713(struct intel_gt *gt, struct i915_wa_list *wal)
1567 {
1568 	struct intel_engine_cs *engine;
1569 	int id;
1570 
1571 	for_each_engine(engine, gt, id)
1572 		if (engine->class == VIDEO_DECODE_CLASS)
1573 			wa_write_or(wal, VDBOX_CGCTL3F1C(engine->mmio_base),
1574 				    MFXPIPE_CLKGATE_DIS);
1575 }
1576 
1577 static void
1578 xelpmp_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1579 {
1580 	wa_16021867713(gt, wal);
1581 
1582 	/*
1583 	 * Wa_14018778641
1584 	 * Wa_18018781329
1585 	 *
1586 	 * Note that although these registers are MCR on the primary
1587 	 * GT, the media GT's versions are regular singleton registers.
1588 	 */
1589 	wa_write_or(wal, XELPMP_GSC_MOD_CTRL, FORCE_MISS_FTLB);
1590 
1591 	/* Wa_22016670082 */
1592 	wa_write_or(wal, GEN12_SQCNT1, GEN12_STRICT_RAR_ENABLE);
1593 
1594 	debug_dump_steering(gt);
1595 }
1596 
1597 /*
1598  * The bspec performance guide has recommended MMIO tuning settings.  These
1599  * aren't truly "workarounds" but we want to program them through the
1600  * workaround infrastructure to make sure they're (re)applied at the proper
1601  * times.
1602  *
1603  * The programming in this function is for settings that persist through
1604  * engine resets and also are not part of any engine's register state context.
1605  * I.e., settings that only need to be re-applied in the event of a full GT
1606  * reset.
1607  */
1608 static void gt_tuning_settings(struct intel_gt *gt, struct i915_wa_list *wal)
1609 {
1610 	if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74))) {
1611 		wa_mcr_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
1612 		wa_mcr_write_or(wal, XEHP_SQCM, EN_32B_ACCESS);
1613 	}
1614 
1615 	if (IS_DG2(gt->i915)) {
1616 		wa_mcr_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
1617 		wa_mcr_write_or(wal, XEHP_SQCM, EN_32B_ACCESS);
1618 	}
1619 }
1620 
1621 static void
1622 gt_init_workarounds(struct intel_gt *gt, struct i915_wa_list *wal)
1623 {
1624 	struct drm_i915_private *i915 = gt->i915;
1625 
1626 	gt_tuning_settings(gt, wal);
1627 
1628 	if (gt->type == GT_MEDIA) {
1629 		if (MEDIA_VER_FULL(i915) == IP_VER(13, 0))
1630 			xelpmp_gt_workarounds_init(gt, wal);
1631 		else
1632 			MISSING_CASE(MEDIA_VER_FULL(i915));
1633 
1634 		return;
1635 	}
1636 
1637 	if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74)))
1638 		xelpg_gt_workarounds_init(gt, wal);
1639 	else if (IS_DG2(i915))
1640 		dg2_gt_workarounds_init(gt, wal);
1641 	else if (IS_DG1(i915))
1642 		dg1_gt_workarounds_init(gt, wal);
1643 	else if (GRAPHICS_VER(i915) == 12)
1644 		gen12_gt_workarounds_init(gt, wal);
1645 	else if (GRAPHICS_VER(i915) == 11)
1646 		icl_gt_workarounds_init(gt, wal);
1647 	else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
1648 		cfl_gt_workarounds_init(gt, wal);
1649 	else if (IS_GEMINILAKE(i915))
1650 		glk_gt_workarounds_init(gt, wal);
1651 	else if (IS_KABYLAKE(i915))
1652 		kbl_gt_workarounds_init(gt, wal);
1653 	else if (IS_BROXTON(i915))
1654 		gen9_gt_workarounds_init(gt, wal);
1655 	else if (IS_SKYLAKE(i915))
1656 		skl_gt_workarounds_init(gt, wal);
1657 	else if (IS_HASWELL(i915))
1658 		hsw_gt_workarounds_init(gt, wal);
1659 	else if (IS_VALLEYVIEW(i915))
1660 		vlv_gt_workarounds_init(gt, wal);
1661 	else if (IS_IVYBRIDGE(i915))
1662 		ivb_gt_workarounds_init(gt, wal);
1663 	else if (GRAPHICS_VER(i915) == 6)
1664 		snb_gt_workarounds_init(gt, wal);
1665 	else if (GRAPHICS_VER(i915) == 5)
1666 		ilk_gt_workarounds_init(gt, wal);
1667 	else if (IS_G4X(i915))
1668 		g4x_gt_workarounds_init(gt, wal);
1669 	else if (GRAPHICS_VER(i915) == 4)
1670 		gen4_gt_workarounds_init(gt, wal);
1671 	else if (GRAPHICS_VER(i915) <= 8)
1672 		;
1673 	else
1674 		MISSING_CASE(GRAPHICS_VER(i915));
1675 }
1676 
1677 void intel_gt_init_workarounds(struct intel_gt *gt)
1678 {
1679 	struct i915_wa_list *wal = &gt->wa_list;
1680 
1681 	wa_init_start(wal, gt, "GT", "global");
1682 	gt_init_workarounds(gt, wal);
1683 	wa_init_finish(wal);
1684 }
1685 
1686 static bool
1687 wa_verify(struct intel_gt *gt, const struct i915_wa *wa, u32 cur,
1688 	  const char *name, const char *from)
1689 {
1690 	if ((cur ^ wa->set) & wa->read) {
1691 		gt_err(gt,
1692 		       "%s workaround lost on %s! (reg[%x]=0x%x, relevant bits were 0x%x vs expected 0x%x)\n",
1693 		       name, from, i915_mmio_reg_offset(wa->reg),
1694 		       cur, cur & wa->read, wa->set & wa->read);
1695 
1696 		return false;
1697 	}
1698 
1699 	return true;
1700 }
1701 
1702 static void wa_list_apply(const struct i915_wa_list *wal)
1703 {
1704 	struct intel_gt *gt = wal->gt;
1705 	struct intel_uncore *uncore = gt->uncore;
1706 	enum forcewake_domains fw;
1707 	unsigned long flags;
1708 	struct i915_wa *wa;
1709 	unsigned int i;
1710 
1711 	if (!wal->count)
1712 		return;
1713 
1714 	fw = wal_get_fw_for_rmw(uncore, wal);
1715 
1716 	intel_gt_mcr_lock(gt, &flags);
1717 	spin_lock(&uncore->lock);
1718 	intel_uncore_forcewake_get__locked(uncore, fw);
1719 
1720 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
1721 		u32 val, old = 0;
1722 
1723 		/* open-coded rmw due to steering */
1724 		if (wa->clr)
1725 			old = wa->is_mcr ?
1726 				intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1727 				intel_uncore_read_fw(uncore, wa->reg);
1728 		val = (old & ~wa->clr) | wa->set;
1729 		if (val != old || !wa->clr) {
1730 			if (wa->is_mcr)
1731 				intel_gt_mcr_multicast_write_fw(gt, wa->mcr_reg, val);
1732 			else
1733 				intel_uncore_write_fw(uncore, wa->reg, val);
1734 		}
1735 
1736 		if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) {
1737 			u32 val = wa->is_mcr ?
1738 				intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1739 				intel_uncore_read_fw(uncore, wa->reg);
1740 
1741 			wa_verify(gt, wa, val, wal->name, "application");
1742 		}
1743 	}
1744 
1745 	intel_uncore_forcewake_put__locked(uncore, fw);
1746 	spin_unlock(&uncore->lock);
1747 	intel_gt_mcr_unlock(gt, flags);
1748 }
1749 
1750 void intel_gt_apply_workarounds(struct intel_gt *gt)
1751 {
1752 	wa_list_apply(&gt->wa_list);
1753 }
1754 
1755 static bool wa_list_verify(struct intel_gt *gt,
1756 			   const struct i915_wa_list *wal,
1757 			   const char *from)
1758 {
1759 	struct intel_uncore *uncore = gt->uncore;
1760 	struct i915_wa *wa;
1761 	enum forcewake_domains fw;
1762 	unsigned long flags;
1763 	unsigned int i;
1764 	bool ok = true;
1765 
1766 	fw = wal_get_fw_for_rmw(uncore, wal);
1767 
1768 	intel_gt_mcr_lock(gt, &flags);
1769 	spin_lock(&uncore->lock);
1770 	intel_uncore_forcewake_get__locked(uncore, fw);
1771 
1772 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1773 		ok &= wa_verify(wal->gt, wa, wa->is_mcr ?
1774 				intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1775 				intel_uncore_read_fw(uncore, wa->reg),
1776 				wal->name, from);
1777 
1778 	intel_uncore_forcewake_put__locked(uncore, fw);
1779 	spin_unlock(&uncore->lock);
1780 	intel_gt_mcr_unlock(gt, flags);
1781 
1782 	return ok;
1783 }
1784 
1785 bool intel_gt_verify_workarounds(struct intel_gt *gt, const char *from)
1786 {
1787 	return wa_list_verify(gt, &gt->wa_list, from);
1788 }
1789 
1790 __maybe_unused
1791 static bool is_nonpriv_flags_valid(u32 flags)
1792 {
1793 	/* Check only valid flag bits are set */
1794 	if (flags & ~RING_FORCE_TO_NONPRIV_MASK_VALID)
1795 		return false;
1796 
1797 	/* NB: Only 3 out of 4 enum values are valid for access field */
1798 	if ((flags & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
1799 	    RING_FORCE_TO_NONPRIV_ACCESS_INVALID)
1800 		return false;
1801 
1802 	return true;
1803 }
1804 
1805 static void
1806 whitelist_reg_ext(struct i915_wa_list *wal, i915_reg_t reg, u32 flags)
1807 {
1808 	struct i915_wa wa = {
1809 		.reg = reg
1810 	};
1811 
1812 	if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1813 		return;
1814 
1815 	if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1816 		return;
1817 
1818 	wa.reg.reg |= flags;
1819 	_wa_add(wal, &wa);
1820 }
1821 
1822 static void
1823 whitelist_mcr_reg_ext(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 flags)
1824 {
1825 	struct i915_wa wa = {
1826 		.mcr_reg = reg,
1827 		.is_mcr = 1,
1828 	};
1829 
1830 	if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1831 		return;
1832 
1833 	if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1834 		return;
1835 
1836 	wa.mcr_reg.reg |= flags;
1837 	_wa_add(wal, &wa);
1838 }
1839 
1840 static void
1841 whitelist_reg(struct i915_wa_list *wal, i915_reg_t reg)
1842 {
1843 	whitelist_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1844 }
1845 
1846 static void
1847 whitelist_mcr_reg(struct i915_wa_list *wal, i915_mcr_reg_t reg)
1848 {
1849 	whitelist_mcr_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1850 }
1851 
1852 static void gen9_whitelist_build(struct i915_wa_list *w)
1853 {
1854 	/* WaVFEStateAfterPipeControlwithMediaStateClear:skl,bxt,glk,cfl */
1855 	whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1856 
1857 	/* WaEnablePreemptionGranularityControlByUMD:skl,bxt,kbl,cfl,[cnl] */
1858 	whitelist_reg(w, GEN8_CS_CHICKEN1);
1859 
1860 	/* WaAllowUMDToModifyHDCChicken1:skl,bxt,kbl,glk,cfl */
1861 	whitelist_reg(w, GEN8_HDC_CHICKEN1);
1862 
1863 	/* WaSendPushConstantsFromMMIO:skl,bxt */
1864 	whitelist_reg(w, COMMON_SLICE_CHICKEN2);
1865 }
1866 
1867 static void skl_whitelist_build(struct intel_engine_cs *engine)
1868 {
1869 	struct i915_wa_list *w = &engine->whitelist;
1870 
1871 	if (engine->class != RENDER_CLASS)
1872 		return;
1873 
1874 	gen9_whitelist_build(w);
1875 
1876 	/* WaDisableLSQCROPERFforOCL:skl */
1877 	whitelist_mcr_reg(w, GEN8_L3SQCREG4);
1878 }
1879 
1880 static void bxt_whitelist_build(struct intel_engine_cs *engine)
1881 {
1882 	if (engine->class != RENDER_CLASS)
1883 		return;
1884 
1885 	gen9_whitelist_build(&engine->whitelist);
1886 }
1887 
1888 static void kbl_whitelist_build(struct intel_engine_cs *engine)
1889 {
1890 	struct i915_wa_list *w = &engine->whitelist;
1891 
1892 	if (engine->class != RENDER_CLASS)
1893 		return;
1894 
1895 	gen9_whitelist_build(w);
1896 
1897 	/* WaDisableLSQCROPERFforOCL:kbl */
1898 	whitelist_mcr_reg(w, GEN8_L3SQCREG4);
1899 }
1900 
1901 static void glk_whitelist_build(struct intel_engine_cs *engine)
1902 {
1903 	struct i915_wa_list *w = &engine->whitelist;
1904 
1905 	if (engine->class != RENDER_CLASS)
1906 		return;
1907 
1908 	gen9_whitelist_build(w);
1909 
1910 	/* WA #0862: Userspace has to set "Barrier Mode" to avoid hangs. */
1911 	whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1912 }
1913 
1914 static void cfl_whitelist_build(struct intel_engine_cs *engine)
1915 {
1916 	struct i915_wa_list *w = &engine->whitelist;
1917 
1918 	if (engine->class != RENDER_CLASS)
1919 		return;
1920 
1921 	gen9_whitelist_build(w);
1922 
1923 	/*
1924 	 * WaAllowPMDepthAndInvocationCountAccessFromUMD:cfl,whl,cml,aml
1925 	 *
1926 	 * This covers 4 register which are next to one another :
1927 	 *   - PS_INVOCATION_COUNT
1928 	 *   - PS_INVOCATION_COUNT_UDW
1929 	 *   - PS_DEPTH_COUNT
1930 	 *   - PS_DEPTH_COUNT_UDW
1931 	 */
1932 	whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1933 			  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1934 			  RING_FORCE_TO_NONPRIV_RANGE_4);
1935 }
1936 
1937 static void allow_read_ctx_timestamp(struct intel_engine_cs *engine)
1938 {
1939 	struct i915_wa_list *w = &engine->whitelist;
1940 
1941 	if (engine->class != RENDER_CLASS)
1942 		whitelist_reg_ext(w,
1943 				  RING_CTX_TIMESTAMP(engine->mmio_base),
1944 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1945 }
1946 
1947 static void cml_whitelist_build(struct intel_engine_cs *engine)
1948 {
1949 	allow_read_ctx_timestamp(engine);
1950 
1951 	cfl_whitelist_build(engine);
1952 }
1953 
1954 static void icl_whitelist_build(struct intel_engine_cs *engine)
1955 {
1956 	struct i915_wa_list *w = &engine->whitelist;
1957 
1958 	allow_read_ctx_timestamp(engine);
1959 
1960 	switch (engine->class) {
1961 	case RENDER_CLASS:
1962 		/* WaAllowUMDToModifyHalfSliceChicken7:icl */
1963 		whitelist_mcr_reg(w, GEN9_HALF_SLICE_CHICKEN7);
1964 
1965 		/* WaAllowUMDToModifySamplerMode:icl */
1966 		whitelist_mcr_reg(w, GEN10_SAMPLER_MODE);
1967 
1968 		/* WaEnableStateCacheRedirectToCS:icl */
1969 		whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1970 
1971 		/*
1972 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:icl
1973 		 *
1974 		 * This covers 4 register which are next to one another :
1975 		 *   - PS_INVOCATION_COUNT
1976 		 *   - PS_INVOCATION_COUNT_UDW
1977 		 *   - PS_DEPTH_COUNT
1978 		 *   - PS_DEPTH_COUNT_UDW
1979 		 */
1980 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1981 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
1982 				  RING_FORCE_TO_NONPRIV_RANGE_4);
1983 		break;
1984 
1985 	case VIDEO_DECODE_CLASS:
1986 		/* hucStatusRegOffset */
1987 		whitelist_reg_ext(w, _MMIO(0x2000 + engine->mmio_base),
1988 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1989 		/* hucUKernelHdrInfoRegOffset */
1990 		whitelist_reg_ext(w, _MMIO(0x2014 + engine->mmio_base),
1991 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1992 		/* hucStatus2RegOffset */
1993 		whitelist_reg_ext(w, _MMIO(0x23B0 + engine->mmio_base),
1994 				  RING_FORCE_TO_NONPRIV_ACCESS_RD);
1995 		break;
1996 
1997 	default:
1998 		break;
1999 	}
2000 }
2001 
2002 static void tgl_whitelist_build(struct intel_engine_cs *engine)
2003 {
2004 	struct i915_wa_list *w = &engine->whitelist;
2005 
2006 	allow_read_ctx_timestamp(engine);
2007 
2008 	switch (engine->class) {
2009 	case RENDER_CLASS:
2010 		/*
2011 		 * WaAllowPMDepthAndInvocationCountAccessFromUMD:tgl
2012 		 * Wa_1408556865:tgl
2013 		 *
2014 		 * This covers 4 registers which are next to one another :
2015 		 *   - PS_INVOCATION_COUNT
2016 		 *   - PS_INVOCATION_COUNT_UDW
2017 		 *   - PS_DEPTH_COUNT
2018 		 *   - PS_DEPTH_COUNT_UDW
2019 		 */
2020 		whitelist_reg_ext(w, PS_INVOCATION_COUNT,
2021 				  RING_FORCE_TO_NONPRIV_ACCESS_RD |
2022 				  RING_FORCE_TO_NONPRIV_RANGE_4);
2023 
2024 		/*
2025 		 * Wa_1808121037:tgl
2026 		 * Wa_14012131227:dg1
2027 		 * Wa_1508744258:tgl,rkl,dg1,adl-s,adl-p
2028 		 */
2029 		whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
2030 
2031 		/* Wa_1806527549:tgl */
2032 		whitelist_reg(w, HIZ_CHICKEN);
2033 
2034 		/* Required by recommended tuning setting (not a workaround) */
2035 		whitelist_reg(w, GEN11_COMMON_SLICE_CHICKEN3);
2036 
2037 		break;
2038 	default:
2039 		break;
2040 	}
2041 }
2042 
2043 static void dg2_whitelist_build(struct intel_engine_cs *engine)
2044 {
2045 	struct i915_wa_list *w = &engine->whitelist;
2046 
2047 	switch (engine->class) {
2048 	case RENDER_CLASS:
2049 		/* Required by recommended tuning setting (not a workaround) */
2050 		whitelist_mcr_reg(w, XEHP_COMMON_SLICE_CHICKEN3);
2051 
2052 		break;
2053 	default:
2054 		break;
2055 	}
2056 }
2057 
2058 static void xelpg_whitelist_build(struct intel_engine_cs *engine)
2059 {
2060 	struct i915_wa_list *w = &engine->whitelist;
2061 
2062 	switch (engine->class) {
2063 	case RENDER_CLASS:
2064 		/* Required by recommended tuning setting (not a workaround) */
2065 		whitelist_mcr_reg(w, XEHP_COMMON_SLICE_CHICKEN3);
2066 
2067 		break;
2068 	default:
2069 		break;
2070 	}
2071 }
2072 
2073 void intel_engine_init_whitelist(struct intel_engine_cs *engine)
2074 {
2075 	struct drm_i915_private *i915 = engine->i915;
2076 	struct i915_wa_list *w = &engine->whitelist;
2077 
2078 	wa_init_start(w, engine->gt, "whitelist", engine->name);
2079 
2080 	if (engine->gt->type == GT_MEDIA)
2081 		; /* none yet */
2082 	else if (IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 70), IP_VER(12, 74)))
2083 		xelpg_whitelist_build(engine);
2084 	else if (IS_DG2(i915))
2085 		dg2_whitelist_build(engine);
2086 	else if (GRAPHICS_VER(i915) == 12)
2087 		tgl_whitelist_build(engine);
2088 	else if (GRAPHICS_VER(i915) == 11)
2089 		icl_whitelist_build(engine);
2090 	else if (IS_COMETLAKE(i915))
2091 		cml_whitelist_build(engine);
2092 	else if (IS_COFFEELAKE(i915))
2093 		cfl_whitelist_build(engine);
2094 	else if (IS_GEMINILAKE(i915))
2095 		glk_whitelist_build(engine);
2096 	else if (IS_KABYLAKE(i915))
2097 		kbl_whitelist_build(engine);
2098 	else if (IS_BROXTON(i915))
2099 		bxt_whitelist_build(engine);
2100 	else if (IS_SKYLAKE(i915))
2101 		skl_whitelist_build(engine);
2102 	else if (GRAPHICS_VER(i915) <= 8)
2103 		;
2104 	else
2105 		MISSING_CASE(GRAPHICS_VER(i915));
2106 
2107 	wa_init_finish(w);
2108 }
2109 
2110 void intel_engine_apply_whitelist(struct intel_engine_cs *engine)
2111 {
2112 	const struct i915_wa_list *wal = &engine->whitelist;
2113 	struct intel_uncore *uncore = engine->uncore;
2114 	const u32 base = engine->mmio_base;
2115 	struct i915_wa *wa;
2116 	unsigned int i;
2117 
2118 	if (!wal->count)
2119 		return;
2120 
2121 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
2122 		intel_uncore_write(uncore,
2123 				   RING_FORCE_TO_NONPRIV(base, i),
2124 				   i915_mmio_reg_offset(wa->reg));
2125 
2126 	/* And clear the rest just in case of garbage */
2127 	for (; i < RING_MAX_NONPRIV_SLOTS; i++)
2128 		intel_uncore_write(uncore,
2129 				   RING_FORCE_TO_NONPRIV(base, i),
2130 				   i915_mmio_reg_offset(RING_NOPID(base)));
2131 }
2132 
2133 /*
2134  * engine_fake_wa_init(), a place holder to program the registers
2135  * which are not part of an official workaround defined by the
2136  * hardware team.
2137  * Adding programming of those register inside workaround will
2138  * allow utilizing wa framework to proper application and verification.
2139  */
2140 static void
2141 engine_fake_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2142 {
2143 	u8 mocs_w, mocs_r;
2144 
2145 	/*
2146 	 * RING_CMD_CCTL specifies the default MOCS entry that will be used
2147 	 * by the command streamer when executing commands that don't have
2148 	 * a way to explicitly specify a MOCS setting.  The default should
2149 	 * usually reference whichever MOCS entry corresponds to uncached
2150 	 * behavior, although use of a WB cached entry is recommended by the
2151 	 * spec in certain circumstances on specific platforms.
2152 	 */
2153 	if (GRAPHICS_VER(engine->i915) >= 12) {
2154 		mocs_r = engine->gt->mocs.uc_index;
2155 		mocs_w = engine->gt->mocs.uc_index;
2156 
2157 		if (HAS_L3_CCS_READ(engine->i915) &&
2158 		    engine->class == COMPUTE_CLASS) {
2159 			mocs_r = engine->gt->mocs.wb_index;
2160 
2161 			/*
2162 			 * Even on the few platforms where MOCS 0 is a
2163 			 * legitimate table entry, it's never the correct
2164 			 * setting to use here; we can assume the MOCS init
2165 			 * just forgot to initialize wb_index.
2166 			 */
2167 			drm_WARN_ON(&engine->i915->drm, mocs_r == 0);
2168 		}
2169 
2170 		wa_masked_field_set(wal,
2171 				    RING_CMD_CCTL(engine->mmio_base),
2172 				    CMD_CCTL_MOCS_MASK,
2173 				    CMD_CCTL_MOCS_OVERRIDE(mocs_w, mocs_r));
2174 	}
2175 }
2176 
2177 static void
2178 rcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2179 {
2180 	struct drm_i915_private *i915 = engine->i915;
2181 	struct intel_gt *gt = engine->gt;
2182 
2183 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2184 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
2185 		/* Wa_22014600077 */
2186 		wa_mcr_masked_en(wal, GEN10_CACHE_MODE_SS,
2187 				 ENABLE_EU_COUNT_FOR_TDL_FLUSH);
2188 	}
2189 
2190 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2191 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2192 	    IS_DG2(i915)) {
2193 		/* Wa_1509727124 */
2194 		wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
2195 				 SC_DISABLE_POWER_OPTIMIZATION_EBB);
2196 	}
2197 
2198 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2199 	    IS_DG2(i915)) {
2200 		/* Wa_22012856258 */
2201 		wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
2202 				 GEN12_DISABLE_READ_SUPPRESSION);
2203 	}
2204 
2205 	if (IS_DG2(i915)) {
2206 		/*
2207 		 * Wa_22010960976:dg2
2208 		 * Wa_14013347512:dg2
2209 		 */
2210 		wa_mcr_masked_dis(wal, XEHP_HDC_CHICKEN0,
2211 				  LSC_L1_FLUSH_CTL_3D_DATAPORT_FLUSH_EVENTS_MASK);
2212 	}
2213 
2214 	if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 71)) ||
2215 	    IS_DG2(i915)) {
2216 		/* Wa_14015150844 */
2217 		wa_mcr_add(wal, XEHP_HDC_CHICKEN0, 0,
2218 			   _MASKED_BIT_ENABLE(DIS_ATOMIC_CHAINING_TYPED_WRITES),
2219 			   0, true);
2220 	}
2221 
2222 	if (IS_DG2(i915) || IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2223 	    IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2224 		/*
2225 		 * Wa_1606700617:tgl,dg1,adl-p
2226 		 * Wa_22010271021:tgl,rkl,dg1,adl-s,adl-p
2227 		 * Wa_14010826681:tgl,dg1,rkl,adl-p
2228 		 * Wa_18019627453:dg2
2229 		 */
2230 		wa_masked_en(wal,
2231 			     GEN9_CS_DEBUG_MODE1,
2232 			     FF_DOP_CLOCK_GATE_DISABLE);
2233 	}
2234 
2235 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) || IS_DG1(i915) ||
2236 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2237 		/* Wa_1606931601:tgl,rkl,dg1,adl-s,adl-p */
2238 		wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2, GEN12_DISABLE_EARLY_READ);
2239 
2240 		/*
2241 		 * Wa_1407928979:tgl A*
2242 		 * Wa_18011464164:tgl[B0+],dg1[B0+]
2243 		 * Wa_22010931296:tgl[B0+],dg1[B0+]
2244 		 * Wa_14010919138:rkl,dg1,adl-s,adl-p
2245 		 */
2246 		wa_write_or(wal, GEN7_FF_THREAD_MODE,
2247 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2248 
2249 		/* Wa_1406941453:tgl,rkl,dg1,adl-s,adl-p */
2250 		wa_mcr_masked_en(wal,
2251 				 GEN10_SAMPLER_MODE,
2252 				 ENABLE_SMALLPL);
2253 	}
2254 
2255 	if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2256 	    IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2257 		/* Wa_1409804808 */
2258 		wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
2259 				 GEN12_PUSH_CONST_DEREF_HOLD_DIS);
2260 
2261 		/* Wa_14010229206 */
2262 		wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, GEN12_DISABLE_TDL_PUSH);
2263 	}
2264 
2265 	if (IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) || IS_ALDERLAKE_P(i915)) {
2266 		/*
2267 		 * Wa_1607297627
2268 		 *
2269 		 * On TGL and RKL there are multiple entries for this WA in the
2270 		 * BSpec; some indicate this is an A0-only WA, others indicate
2271 		 * it applies to all steppings so we trust the "all steppings."
2272 		 */
2273 		wa_masked_en(wal,
2274 			     RING_PSMI_CTL(RENDER_RING_BASE),
2275 			     GEN12_WAIT_FOR_EVENT_POWER_DOWN_DISABLE |
2276 			     GEN8_RC_SEMA_IDLE_MSG_DISABLE);
2277 	}
2278 
2279 	if (GRAPHICS_VER(i915) == 11) {
2280 		/* This is not an Wa. Enable for better image quality */
2281 		wa_masked_en(wal,
2282 			     _3D_CHICKEN3,
2283 			     _3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE);
2284 
2285 		/*
2286 		 * Wa_1405543622:icl
2287 		 * Formerly known as WaGAPZPriorityScheme
2288 		 */
2289 		wa_write_or(wal,
2290 			    GEN8_GARBCNTL,
2291 			    GEN11_ARBITRATION_PRIO_ORDER_MASK);
2292 
2293 		/*
2294 		 * Wa_1604223664:icl
2295 		 * Formerly known as WaL3BankAddressHashing
2296 		 */
2297 		wa_write_clr_set(wal,
2298 				 GEN8_GARBCNTL,
2299 				 GEN11_HASH_CTRL_EXCL_MASK,
2300 				 GEN11_HASH_CTRL_EXCL_BIT0);
2301 		wa_write_clr_set(wal,
2302 				 GEN11_GLBLINVL,
2303 				 GEN11_BANK_HASH_ADDR_EXCL_MASK,
2304 				 GEN11_BANK_HASH_ADDR_EXCL_BIT0);
2305 
2306 		/*
2307 		 * Wa_1405733216:icl
2308 		 * Formerly known as WaDisableCleanEvicts
2309 		 */
2310 		wa_mcr_write_or(wal,
2311 				GEN8_L3SQCREG4,
2312 				GEN11_LQSC_CLEAN_EVICT_DISABLE);
2313 
2314 		/* Wa_1606682166:icl */
2315 		wa_write_or(wal,
2316 			    GEN7_SARCHKMD,
2317 			    GEN7_DISABLE_SAMPLER_PREFETCH);
2318 
2319 		/* Wa_1409178092:icl */
2320 		wa_mcr_write_clr_set(wal,
2321 				     GEN11_SCRATCH2,
2322 				     GEN11_COHERENT_PARTIAL_WRITE_MERGE_ENABLE,
2323 				     0);
2324 
2325 		/* WaEnable32PlaneMode:icl */
2326 		wa_masked_en(wal, GEN9_CSFE_CHICKEN1_RCS,
2327 			     GEN11_ENABLE_32_PLANE_MODE);
2328 
2329 		/*
2330 		 * Wa_1408767742:icl[a2..forever],ehl[all]
2331 		 * Wa_1605460711:icl[a0..c0]
2332 		 */
2333 		wa_write_or(wal,
2334 			    GEN7_FF_THREAD_MODE,
2335 			    GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2336 
2337 		/* Wa_22010271021 */
2338 		wa_masked_en(wal,
2339 			     GEN9_CS_DEBUG_MODE1,
2340 			     FF_DOP_CLOCK_GATE_DISABLE);
2341 	}
2342 
2343 	/*
2344 	 * Intel platforms that support fine-grained preemption (i.e., gen9 and
2345 	 * beyond) allow the kernel-mode driver to choose between two different
2346 	 * options for controlling preemption granularity and behavior.
2347 	 *
2348 	 * Option 1 (hardware default):
2349 	 *   Preemption settings are controlled in a global manner via
2350 	 *   kernel-only register CS_DEBUG_MODE1 (0x20EC).  Any granularity
2351 	 *   and settings chosen by the kernel-mode driver will apply to all
2352 	 *   userspace clients.
2353 	 *
2354 	 * Option 2:
2355 	 *   Preemption settings are controlled on a per-context basis via
2356 	 *   register CS_CHICKEN1 (0x2580).  CS_CHICKEN1 is saved/restored on
2357 	 *   context switch and is writable by userspace (e.g., via
2358 	 *   MI_LOAD_REGISTER_IMMEDIATE instructions placed in a batch buffer)
2359 	 *   which allows different userspace drivers/clients to select
2360 	 *   different settings, or to change those settings on the fly in
2361 	 *   response to runtime needs.  This option was known by name
2362 	 *   "FtrPerCtxtPreemptionGranularityControl" at one time, although
2363 	 *   that name is somewhat misleading as other non-granularity
2364 	 *   preemption settings are also impacted by this decision.
2365 	 *
2366 	 * On Linux, our policy has always been to let userspace drivers
2367 	 * control preemption granularity/settings (Option 2).  This was
2368 	 * originally mandatory on gen9 to prevent ABI breakage (old gen9
2369 	 * userspace developed before object-level preemption was enabled would
2370 	 * not behave well if i915 were to go with Option 1 and enable that
2371 	 * preemption in a global manner).  On gen9 each context would have
2372 	 * object-level preemption disabled by default (see
2373 	 * WaDisable3DMidCmdPreemption in gen9_ctx_workarounds_init), but
2374 	 * userspace drivers could opt-in to object-level preemption as they
2375 	 * saw fit.  For post-gen9 platforms, we continue to utilize Option 2;
2376 	 * even though it is no longer necessary for ABI compatibility when
2377 	 * enabling a new platform, it does ensure that userspace will be able
2378 	 * to implement any workarounds that show up requiring temporary
2379 	 * adjustments to preemption behavior at runtime.
2380 	 *
2381 	 * Notes/Workarounds:
2382 	 *  - Wa_14015141709:  On DG2 and early steppings of MTL,
2383 	 *      CS_CHICKEN1[0] does not disable object-level preemption as
2384 	 *      it is supposed to (nor does CS_DEBUG_MODE1[0] if we had been
2385 	 *      using Option 1).  Effectively this means userspace is unable
2386 	 *      to disable object-level preemption on these platforms/steppings
2387 	 *      despite the setting here.
2388 	 *
2389 	 *  - Wa_16013994831:  May require that userspace program
2390 	 *      CS_CHICKEN1[10] when certain runtime conditions are true.
2391 	 *      Userspace requires Option 2 to be in effect for their update of
2392 	 *      CS_CHICKEN1[10] to be effective.
2393 	 *
2394 	 * Other workarounds may appear in the future that will also require
2395 	 * Option 2 behavior to allow proper userspace implementation.
2396 	 */
2397 	if (GRAPHICS_VER(i915) >= 9)
2398 		wa_masked_en(wal,
2399 			     GEN7_FF_SLICE_CS_CHICKEN1,
2400 			     GEN9_FFSC_PERCTX_PREEMPT_CTRL);
2401 
2402 	if (IS_SKYLAKE(i915) ||
2403 	    IS_KABYLAKE(i915) ||
2404 	    IS_COFFEELAKE(i915) ||
2405 	    IS_COMETLAKE(i915)) {
2406 		/* WaEnableGapsTsvCreditFix:skl,kbl,cfl */
2407 		wa_write_or(wal,
2408 			    GEN8_GARBCNTL,
2409 			    GEN9_GAPS_TSV_CREDIT_DISABLE);
2410 	}
2411 
2412 	if (IS_BROXTON(i915)) {
2413 		/* WaDisablePooledEuLoadBalancingFix:bxt */
2414 		wa_masked_en(wal,
2415 			     FF_SLICE_CS_CHICKEN2,
2416 			     GEN9_POOLED_EU_LOAD_BALANCING_FIX_DISABLE);
2417 	}
2418 
2419 	if (GRAPHICS_VER(i915) == 9) {
2420 		/* WaContextSwitchWithConcurrentTLBInvalidate:skl,bxt,kbl,glk,cfl */
2421 		wa_masked_en(wal,
2422 			     GEN9_CSFE_CHICKEN1_RCS,
2423 			     GEN9_PREEMPT_GPGPU_SYNC_SWITCH_DISABLE);
2424 
2425 		/* WaEnableLbsSlaRetryTimerDecrement:skl,bxt,kbl,glk,cfl */
2426 		wa_mcr_write_or(wal,
2427 				BDW_SCRATCH1,
2428 				GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
2429 
2430 		/* WaProgramL3SqcReg1DefaultForPerf:bxt,glk */
2431 		if (IS_GEN9_LP(i915))
2432 			wa_mcr_write_clr_set(wal,
2433 					     GEN8_L3SQCREG1,
2434 					     L3_PRIO_CREDITS_MASK,
2435 					     L3_GENERAL_PRIO_CREDITS(62) |
2436 					     L3_HIGH_PRIO_CREDITS(2));
2437 
2438 		/* WaOCLCoherentLineFlush:skl,bxt,kbl,cfl */
2439 		wa_mcr_write_or(wal,
2440 				GEN8_L3SQCREG4,
2441 				GEN8_LQSC_FLUSH_COHERENT_LINES);
2442 
2443 		/* Disable atomics in L3 to prevent unrecoverable hangs */
2444 		wa_write_clr_set(wal, GEN9_SCRATCH_LNCF1,
2445 				 GEN9_LNCF_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2446 		wa_mcr_write_clr_set(wal, GEN8_L3SQCREG4,
2447 				     GEN8_LQSQ_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2448 		wa_mcr_write_clr_set(wal, GEN9_SCRATCH1,
2449 				     EVICTION_PERF_FIX_ENABLE, 0);
2450 	}
2451 
2452 	if (IS_HASWELL(i915)) {
2453 		/* WaSampleCChickenBitEnable:hsw */
2454 		wa_masked_en(wal,
2455 			     HSW_HALF_SLICE_CHICKEN3, HSW_SAMPLE_C_PERFORMANCE);
2456 
2457 		wa_masked_dis(wal,
2458 			      CACHE_MODE_0_GEN7,
2459 			      /* enable HiZ Raw Stall Optimization */
2460 			      HIZ_RAW_STALL_OPT_DISABLE);
2461 	}
2462 
2463 	if (IS_VALLEYVIEW(i915)) {
2464 		/* WaDisableEarlyCull:vlv */
2465 		wa_masked_en(wal,
2466 			     _3D_CHICKEN3,
2467 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2468 
2469 		/*
2470 		 * WaVSThreadDispatchOverride:ivb,vlv
2471 		 *
2472 		 * This actually overrides the dispatch
2473 		 * mode for all thread types.
2474 		 */
2475 		wa_write_clr_set(wal,
2476 				 GEN7_FF_THREAD_MODE,
2477 				 GEN7_FF_SCHED_MASK,
2478 				 GEN7_FF_TS_SCHED_HW |
2479 				 GEN7_FF_VS_SCHED_HW |
2480 				 GEN7_FF_DS_SCHED_HW);
2481 
2482 		/* WaPsdDispatchEnable:vlv */
2483 		/* WaDisablePSDDualDispatchEnable:vlv */
2484 		wa_masked_en(wal,
2485 			     GEN7_HALF_SLICE_CHICKEN1,
2486 			     GEN7_MAX_PS_THREAD_DEP |
2487 			     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2488 	}
2489 
2490 	if (IS_IVYBRIDGE(i915)) {
2491 		/* WaDisableEarlyCull:ivb */
2492 		wa_masked_en(wal,
2493 			     _3D_CHICKEN3,
2494 			     _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2495 
2496 		if (0) { /* causes HiZ corruption on ivb:gt1 */
2497 			/* enable HiZ Raw Stall Optimization */
2498 			wa_masked_dis(wal,
2499 				      CACHE_MODE_0_GEN7,
2500 				      HIZ_RAW_STALL_OPT_DISABLE);
2501 		}
2502 
2503 		/*
2504 		 * WaVSThreadDispatchOverride:ivb,vlv
2505 		 *
2506 		 * This actually overrides the dispatch
2507 		 * mode for all thread types.
2508 		 */
2509 		wa_write_clr_set(wal,
2510 				 GEN7_FF_THREAD_MODE,
2511 				 GEN7_FF_SCHED_MASK,
2512 				 GEN7_FF_TS_SCHED_HW |
2513 				 GEN7_FF_VS_SCHED_HW |
2514 				 GEN7_FF_DS_SCHED_HW);
2515 
2516 		/* WaDisablePSDDualDispatchEnable:ivb */
2517 		if (IS_IVB_GT1(i915))
2518 			wa_masked_en(wal,
2519 				     GEN7_HALF_SLICE_CHICKEN1,
2520 				     GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2521 	}
2522 
2523 	if (GRAPHICS_VER(i915) == 7) {
2524 		/* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
2525 		wa_masked_en(wal,
2526 			     RING_MODE_GEN7(RENDER_RING_BASE),
2527 			     GFX_TLB_INVALIDATE_EXPLICIT | GFX_REPLAY_MODE);
2528 
2529 		/* WaDisable_RenderCache_OperationalFlush:ivb,vlv,hsw */
2530 		wa_masked_dis(wal, CACHE_MODE_0_GEN7, RC_OP_FLUSH_ENABLE);
2531 
2532 		/*
2533 		 * BSpec says this must be set, even though
2534 		 * WaDisable4x2SubspanOptimization:ivb,hsw
2535 		 * WaDisable4x2SubspanOptimization isn't listed for VLV.
2536 		 */
2537 		wa_masked_en(wal,
2538 			     CACHE_MODE_1,
2539 			     PIXEL_SUBSPAN_COLLECT_OPT_DISABLE);
2540 
2541 		/*
2542 		 * BSpec recommends 8x4 when MSAA is used,
2543 		 * however in practice 16x4 seems fastest.
2544 		 *
2545 		 * Note that PS/WM thread counts depend on the WIZ hashing
2546 		 * disable bit, which we don't touch here, but it's good
2547 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2548 		 */
2549 		wa_masked_field_set(wal,
2550 				    GEN7_GT_MODE,
2551 				    GEN6_WIZ_HASHING_MASK,
2552 				    GEN6_WIZ_HASHING_16x4);
2553 	}
2554 
2555 	if (IS_GRAPHICS_VER(i915, 6, 7))
2556 		/*
2557 		 * We need to disable the AsyncFlip performance optimisations in
2558 		 * order to use MI_WAIT_FOR_EVENT within the CS. It should
2559 		 * already be programmed to '1' on all products.
2560 		 *
2561 		 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
2562 		 */
2563 		wa_masked_en(wal,
2564 			     RING_MI_MODE(RENDER_RING_BASE),
2565 			     ASYNC_FLIP_PERF_DISABLE);
2566 
2567 	if (GRAPHICS_VER(i915) == 6) {
2568 		/*
2569 		 * Required for the hardware to program scanline values for
2570 		 * waiting
2571 		 * WaEnableFlushTlbInvalidationMode:snb
2572 		 */
2573 		wa_masked_en(wal,
2574 			     GFX_MODE,
2575 			     GFX_TLB_INVALIDATE_EXPLICIT);
2576 
2577 		/* WaDisableHiZPlanesWhenMSAAEnabled:snb */
2578 		wa_masked_en(wal,
2579 			     _3D_CHICKEN,
2580 			     _3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB);
2581 
2582 		wa_masked_en(wal,
2583 			     _3D_CHICKEN3,
2584 			     /* WaStripsFansDisableFastClipPerformanceFix:snb */
2585 			     _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL |
2586 			     /*
2587 			      * Bspec says:
2588 			      * "This bit must be set if 3DSTATE_CLIP clip mode is set
2589 			      * to normal and 3DSTATE_SF number of SF output attributes
2590 			      * is more than 16."
2591 			      */
2592 			     _3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH);
2593 
2594 		/*
2595 		 * BSpec recommends 8x4 when MSAA is used,
2596 		 * however in practice 16x4 seems fastest.
2597 		 *
2598 		 * Note that PS/WM thread counts depend on the WIZ hashing
2599 		 * disable bit, which we don't touch here, but it's good
2600 		 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2601 		 */
2602 		wa_masked_field_set(wal,
2603 				    GEN6_GT_MODE,
2604 				    GEN6_WIZ_HASHING_MASK,
2605 				    GEN6_WIZ_HASHING_16x4);
2606 
2607 		/* WaDisable_RenderCache_OperationalFlush:snb */
2608 		wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
2609 
2610 		/*
2611 		 * From the Sandybridge PRM, volume 1 part 3, page 24:
2612 		 * "If this bit is set, STCunit will have LRA as replacement
2613 		 *  policy. [...] This bit must be reset. LRA replacement
2614 		 *  policy is not supported."
2615 		 */
2616 		wa_masked_dis(wal,
2617 			      CACHE_MODE_0,
2618 			      CM0_STC_EVICT_DISABLE_LRA_SNB);
2619 	}
2620 
2621 	if (IS_GRAPHICS_VER(i915, 4, 6))
2622 		/* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
2623 		wa_add(wal, RING_MI_MODE(RENDER_RING_BASE),
2624 		       0, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH),
2625 		       /* XXX bit doesn't stick on Broadwater */
2626 		       IS_I965G(i915) ? 0 : VS_TIMER_DISPATCH, true);
2627 
2628 	if (GRAPHICS_VER(i915) == 4)
2629 		/*
2630 		 * Disable CONSTANT_BUFFER before it is loaded from the context
2631 		 * image. For as it is loaded, it is executed and the stored
2632 		 * address may no longer be valid, leading to a GPU hang.
2633 		 *
2634 		 * This imposes the requirement that userspace reload their
2635 		 * CONSTANT_BUFFER on every batch, fortunately a requirement
2636 		 * they are already accustomed to from before contexts were
2637 		 * enabled.
2638 		 */
2639 		wa_add(wal, ECOSKPD(RENDER_RING_BASE),
2640 		       0, _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE),
2641 		       0 /* XXX bit doesn't stick on Broadwater */,
2642 		       true);
2643 }
2644 
2645 static void
2646 xcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2647 {
2648 	struct drm_i915_private *i915 = engine->i915;
2649 
2650 	/* WaKBLVECSSemaphoreWaitPoll:kbl */
2651 	if (IS_KABYLAKE(i915) && IS_GRAPHICS_STEP(i915, STEP_A0, STEP_F0)) {
2652 		wa_write(wal,
2653 			 RING_SEMA_WAIT_POLL(engine->mmio_base),
2654 			 1);
2655 	}
2656 	/* Wa_16018031267, Wa_16018063123 */
2657 	if (NEEDS_FASTCOLOR_BLT_WABB(engine))
2658 		wa_masked_field_set(wal, ECOSKPD(engine->mmio_base),
2659 				    XEHP_BLITTER_SCHEDULING_MODE_MASK,
2660 				    XEHP_BLITTER_ROUND_ROBIN_MODE);
2661 }
2662 
2663 static void
2664 ccs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2665 {
2666 	/* boilerplate for any CCS engine workaround */
2667 }
2668 
2669 /*
2670  * The bspec performance guide has recommended MMIO tuning settings.  These
2671  * aren't truly "workarounds" but we want to program them with the same
2672  * workaround infrastructure to ensure that they're automatically added to
2673  * the GuC save/restore lists, re-applied at the right times, and checked for
2674  * any conflicting programming requested by real workarounds.
2675  *
2676  * Programming settings should be added here only if their registers are not
2677  * part of an engine's register state context.  If a register is part of a
2678  * context, then any tuning settings should be programmed in an appropriate
2679  * function invoked by __intel_engine_init_ctx_wa().
2680  */
2681 static void
2682 add_render_compute_tuning_settings(struct intel_gt *gt,
2683 				   struct i915_wa_list *wal)
2684 {
2685 	struct drm_i915_private *i915 = gt->i915;
2686 
2687 	if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74)) || IS_DG2(i915))
2688 		wa_mcr_write_clr_set(wal, RT_CTRL, STACKID_CTRL, STACKID_CTRL_512);
2689 
2690 	/*
2691 	 * This tuning setting proves beneficial only on ATS-M designs; the
2692 	 * default "age based" setting is optimal on regular DG2 and other
2693 	 * platforms.
2694 	 */
2695 	if (INTEL_INFO(i915)->tuning_thread_rr_after_dep)
2696 		wa_mcr_masked_field_set(wal, GEN9_ROW_CHICKEN4, THREAD_EX_ARB_MODE,
2697 					THREAD_EX_ARB_MODE_RR_AFTER_DEP);
2698 
2699 	if (GRAPHICS_VER(i915) == 12 && GRAPHICS_VER_FULL(i915) < IP_VER(12, 55))
2700 		wa_write_clr(wal, GEN8_GARBCNTL, GEN12_BUS_HASH_CTL_BIT_EXC);
2701 }
2702 
2703 static void ccs_engine_wa_mode(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2704 {
2705 	struct intel_gt *gt = engine->gt;
2706 
2707 	if (!IS_DG2(gt->i915))
2708 		return;
2709 
2710 	/*
2711 	 * Wa_14019159160: This workaround, along with others, leads to
2712 	 * significant challenges in utilizing load balancing among the
2713 	 * CCS slices. Consequently, an architectural decision has been
2714 	 * made to completely disable automatic CCS load balancing.
2715 	 */
2716 	wa_masked_en(wal, GEN12_RCU_MODE, XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE);
2717 
2718 	/*
2719 	 * After having disabled automatic load balancing we need to
2720 	 * assign all slices to a single CCS. We will call it CCS mode 1
2721 	 */
2722 	intel_gt_apply_ccs_mode(gt);
2723 }
2724 
2725 /*
2726  * The workarounds in this function apply to shared registers in
2727  * the general render reset domain that aren't tied to a
2728  * specific engine.  Since all render+compute engines get reset
2729  * together, and the contents of these registers are lost during
2730  * the shared render domain reset, we'll define such workarounds
2731  * here and then add them to just a single RCS or CCS engine's
2732  * workaround list (whichever engine has the XXXX flag).
2733  */
2734 static void
2735 general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2736 {
2737 	struct drm_i915_private *i915 = engine->i915;
2738 	struct intel_gt *gt = engine->gt;
2739 
2740 	add_render_compute_tuning_settings(gt, wal);
2741 
2742 	if (GRAPHICS_VER(i915) >= 11) {
2743 		/* This is not a Wa (although referred to as
2744 		 * WaSetInidrectStateOverride in places), this allows
2745 		 * applications that reference sampler states through
2746 		 * the BindlessSamplerStateBaseAddress to have their
2747 		 * border color relative to DynamicStateBaseAddress
2748 		 * rather than BindlessSamplerStateBaseAddress.
2749 		 *
2750 		 * Otherwise SAMPLER_STATE border colors have to be
2751 		 * copied in multiple heaps (DynamicStateBaseAddress &
2752 		 * BindlessSamplerStateBaseAddress)
2753 		 *
2754 		 * BSpec: 46052
2755 		 */
2756 		wa_mcr_masked_en(wal,
2757 				 GEN10_SAMPLER_MODE,
2758 				 GEN11_INDIRECT_STATE_BASE_ADDR_OVERRIDE);
2759 	}
2760 
2761 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_B0, STEP_FOREVER) ||
2762 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_B0, STEP_FOREVER) ||
2763 	    IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 74), IP_VER(12, 74)))
2764 		/* Wa_14017856879 */
2765 		wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN3, MTL_DISABLE_FIX_FOR_EOT_FLUSH);
2766 
2767 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2768 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0))
2769 		/*
2770 		 * Wa_14017066071
2771 		 * Wa_14017654203
2772 		 */
2773 		wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
2774 				 MTL_DISABLE_SAMPLER_SC_OOO);
2775 
2776 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0))
2777 		/* Wa_22015279794 */
2778 		wa_mcr_masked_en(wal, GEN10_CACHE_MODE_SS,
2779 				 DISABLE_PREFETCH_INTO_IC);
2780 
2781 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2782 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2783 	    IS_DG2(i915)) {
2784 		/* Wa_22013037850 */
2785 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW,
2786 				DISABLE_128B_EVICTION_COMMAND_UDW);
2787 
2788 		/* Wa_18017747507 */
2789 		wa_masked_en(wal, VFG_PREEMPTION_CHICKEN, POLYGON_TRIFAN_LINELOOP_DISABLE);
2790 	}
2791 
2792 	if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2793 	    IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2794 	    IS_DG2(i915)) {
2795 		/* Wa_22014226127 */
2796 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0, DISABLE_D8_D16_COASLESCE);
2797 	}
2798 
2799 	if (IS_DG2(i915)) {
2800 		/* Wa_14015227452:dg2,pvc */
2801 		wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, XEHP_DIS_BBL_SYSPIPE);
2802 
2803 		/* Wa_16015675438:dg2,pvc */
2804 		wa_masked_en(wal, FF_SLICE_CS_CHICKEN2, GEN12_PERF_FIX_BALANCING_CFE_DISABLE);
2805 
2806 		/*
2807 		 * Wa_16011620976:dg2_g11
2808 		 * Wa_22015475538:dg2
2809 		 */
2810 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, DIS_CHAIN_2XSIMD8);
2811 
2812 		/* Wa_18028616096 */
2813 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, UGM_FRAGMENT_THRESHOLD_TO_3);
2814 	}
2815 
2816 	if (IS_DG2_G11(i915)) {
2817 		/*
2818 		 * Wa_22012826095:dg2
2819 		 * Wa_22013059131:dg2
2820 		 */
2821 		wa_mcr_write_clr_set(wal, LSC_CHICKEN_BIT_0_UDW,
2822 				     MAXREQS_PER_BANK,
2823 				     REG_FIELD_PREP(MAXREQS_PER_BANK, 2));
2824 
2825 		/* Wa_22013059131:dg2 */
2826 		wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0,
2827 				FORCE_1_SUB_MESSAGE_PER_FRAGMENT);
2828 
2829 		/*
2830 		 * Wa_22012654132
2831 		 *
2832 		 * Note that register 0xE420 is write-only and cannot be read
2833 		 * back for verification on DG2 (due to Wa_14012342262), so
2834 		 * we need to explicitly skip the readback.
2835 		 */
2836 		wa_mcr_add(wal, GEN10_CACHE_MODE_SS, 0,
2837 			   _MASKED_BIT_ENABLE(ENABLE_PREFETCH_INTO_IC),
2838 			   0 /* write-only, so skip validation */,
2839 			   true);
2840 	}
2841 }
2842 
2843 static void
2844 engine_init_workarounds(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2845 {
2846 	if (GRAPHICS_VER(engine->i915) < 4)
2847 		return;
2848 
2849 	engine_fake_wa_init(engine, wal);
2850 
2851 	/*
2852 	 * These are common workarounds that just need to applied
2853 	 * to a single RCS/CCS engine's workaround list since
2854 	 * they're reset as part of the general render domain reset.
2855 	 */
2856 	if (engine->flags & I915_ENGINE_FIRST_RENDER_COMPUTE) {
2857 		general_render_compute_wa_init(engine, wal);
2858 		ccs_engine_wa_mode(engine, wal);
2859 	}
2860 
2861 	if (engine->class == COMPUTE_CLASS)
2862 		ccs_engine_wa_init(engine, wal);
2863 	else if (engine->class == RENDER_CLASS)
2864 		rcs_engine_wa_init(engine, wal);
2865 	else
2866 		xcs_engine_wa_init(engine, wal);
2867 }
2868 
2869 void intel_engine_init_workarounds(struct intel_engine_cs *engine)
2870 {
2871 	struct i915_wa_list *wal = &engine->wa_list;
2872 
2873 	wa_init_start(wal, engine->gt, "engine", engine->name);
2874 	engine_init_workarounds(engine, wal);
2875 	wa_init_finish(wal);
2876 }
2877 
2878 void intel_engine_apply_workarounds(struct intel_engine_cs *engine)
2879 {
2880 	wa_list_apply(&engine->wa_list);
2881 }
2882 
2883 static const struct i915_range mcr_ranges_gen8[] = {
2884 	{ .start = 0x5500, .end = 0x55ff },
2885 	{ .start = 0x7000, .end = 0x7fff },
2886 	{ .start = 0x9400, .end = 0x97ff },
2887 	{ .start = 0xb000, .end = 0xb3ff },
2888 	{ .start = 0xe000, .end = 0xe7ff },
2889 	{},
2890 };
2891 
2892 static const struct i915_range mcr_ranges_gen12[] = {
2893 	{ .start =  0x8150, .end =  0x815f },
2894 	{ .start =  0x9520, .end =  0x955f },
2895 	{ .start =  0xb100, .end =  0xb3ff },
2896 	{ .start =  0xde80, .end =  0xe8ff },
2897 	{ .start = 0x24a00, .end = 0x24a7f },
2898 	{},
2899 };
2900 
2901 static const struct i915_range mcr_ranges_xehp[] = {
2902 	{ .start =  0x4000, .end =  0x4aff },
2903 	{ .start =  0x5200, .end =  0x52ff },
2904 	{ .start =  0x5400, .end =  0x7fff },
2905 	{ .start =  0x8140, .end =  0x815f },
2906 	{ .start =  0x8c80, .end =  0x8dff },
2907 	{ .start =  0x94d0, .end =  0x955f },
2908 	{ .start =  0x9680, .end =  0x96ff },
2909 	{ .start =  0xb000, .end =  0xb3ff },
2910 	{ .start =  0xc800, .end =  0xcfff },
2911 	{ .start =  0xd800, .end =  0xd8ff },
2912 	{ .start =  0xdc00, .end =  0xffff },
2913 	{ .start = 0x17000, .end = 0x17fff },
2914 	{ .start = 0x24a00, .end = 0x24a7f },
2915 	{},
2916 };
2917 
2918 static bool mcr_range(struct drm_i915_private *i915, u32 offset)
2919 {
2920 	const struct i915_range *mcr_ranges;
2921 	int i;
2922 
2923 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55))
2924 		mcr_ranges = mcr_ranges_xehp;
2925 	else if (GRAPHICS_VER(i915) >= 12)
2926 		mcr_ranges = mcr_ranges_gen12;
2927 	else if (GRAPHICS_VER(i915) >= 8)
2928 		mcr_ranges = mcr_ranges_gen8;
2929 	else
2930 		return false;
2931 
2932 	/*
2933 	 * Registers in these ranges are affected by the MCR selector
2934 	 * which only controls CPU initiated MMIO. Routing does not
2935 	 * work for CS access so we cannot verify them on this path.
2936 	 */
2937 	for (i = 0; mcr_ranges[i].start; i++)
2938 		if (offset >= mcr_ranges[i].start &&
2939 		    offset <= mcr_ranges[i].end)
2940 			return true;
2941 
2942 	return false;
2943 }
2944 
2945 static int
2946 wa_list_srm(struct i915_request *rq,
2947 	    const struct i915_wa_list *wal,
2948 	    struct i915_vma *vma)
2949 {
2950 	struct drm_i915_private *i915 = rq->i915;
2951 	unsigned int i, count = 0;
2952 	const struct i915_wa *wa;
2953 	u32 srm, *cs;
2954 
2955 	srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
2956 	if (GRAPHICS_VER(i915) >= 8)
2957 		srm++;
2958 
2959 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2960 		if (!mcr_range(i915, i915_mmio_reg_offset(wa->reg)))
2961 			count++;
2962 	}
2963 
2964 	cs = intel_ring_begin(rq, 4 * count);
2965 	if (IS_ERR(cs))
2966 		return PTR_ERR(cs);
2967 
2968 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2969 		u32 offset = i915_mmio_reg_offset(wa->reg);
2970 
2971 		if (mcr_range(i915, offset))
2972 			continue;
2973 
2974 		*cs++ = srm;
2975 		*cs++ = offset;
2976 		*cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
2977 		*cs++ = 0;
2978 	}
2979 	intel_ring_advance(rq, cs);
2980 
2981 	return 0;
2982 }
2983 
2984 static int engine_wa_list_verify(struct intel_context *ce,
2985 				 const struct i915_wa_list * const wal,
2986 				 const char *from)
2987 {
2988 	const struct i915_wa *wa;
2989 	struct i915_request *rq;
2990 	struct i915_vma *vma;
2991 	struct i915_gem_ww_ctx ww;
2992 	unsigned int i;
2993 	u32 *results;
2994 	int err;
2995 
2996 	if (!wal->count)
2997 		return 0;
2998 
2999 	vma = __vm_create_scratch_for_read(&ce->engine->gt->ggtt->vm,
3000 					   wal->count * sizeof(u32));
3001 	if (IS_ERR(vma))
3002 		return PTR_ERR(vma);
3003 
3004 	intel_engine_pm_get(ce->engine);
3005 	i915_gem_ww_ctx_init(&ww, false);
3006 retry:
3007 	err = i915_gem_object_lock(vma->obj, &ww);
3008 	if (err == 0)
3009 		err = intel_context_pin_ww(ce, &ww);
3010 	if (err)
3011 		goto err_pm;
3012 
3013 	err = i915_vma_pin_ww(vma, &ww, 0, 0,
3014 			   i915_vma_is_ggtt(vma) ? PIN_GLOBAL : PIN_USER);
3015 	if (err)
3016 		goto err_unpin;
3017 
3018 	rq = i915_request_create(ce);
3019 	if (IS_ERR(rq)) {
3020 		err = PTR_ERR(rq);
3021 		goto err_vma;
3022 	}
3023 
3024 	err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
3025 	if (err == 0)
3026 		err = wa_list_srm(rq, wal, vma);
3027 
3028 	i915_request_get(rq);
3029 	if (err)
3030 		i915_request_set_error_once(rq, err);
3031 	i915_request_add(rq);
3032 
3033 	if (err)
3034 		goto err_rq;
3035 
3036 	if (i915_request_wait(rq, 0, HZ / 5) < 0) {
3037 		err = -ETIME;
3038 		goto err_rq;
3039 	}
3040 
3041 	results = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
3042 	if (IS_ERR(results)) {
3043 		err = PTR_ERR(results);
3044 		goto err_rq;
3045 	}
3046 
3047 	err = 0;
3048 	for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
3049 		if (mcr_range(rq->i915, i915_mmio_reg_offset(wa->reg)))
3050 			continue;
3051 
3052 		if (!wa_verify(wal->gt, wa, results[i], wal->name, from))
3053 			err = -ENXIO;
3054 	}
3055 
3056 	i915_gem_object_unpin_map(vma->obj);
3057 
3058 err_rq:
3059 	i915_request_put(rq);
3060 err_vma:
3061 	i915_vma_unpin(vma);
3062 err_unpin:
3063 	intel_context_unpin(ce);
3064 err_pm:
3065 	if (err == -EDEADLK) {
3066 		err = i915_gem_ww_ctx_backoff(&ww);
3067 		if (!err)
3068 			goto retry;
3069 	}
3070 	i915_gem_ww_ctx_fini(&ww);
3071 	intel_engine_pm_put(ce->engine);
3072 	i915_vma_put(vma);
3073 	return err;
3074 }
3075 
3076 int intel_engine_verify_workarounds(struct intel_engine_cs *engine,
3077 				    const char *from)
3078 {
3079 	return engine_wa_list_verify(engine->kernel_context,
3080 				     &engine->wa_list,
3081 				     from);
3082 }
3083 
3084 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3085 #include "selftest_workarounds.c"
3086 #endif
3087