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 #include "display/intel_fbc_regs.h"
21
22 /**
23 * DOC: Hardware workarounds
24 *
25 * Hardware workarounds are register programming documented to be executed in
26 * the driver that fall outside of the normal programming sequences for a
27 * platform. There are some basic categories of workarounds, depending on
28 * how/when they are applied:
29 *
30 * - Context workarounds: workarounds that touch registers that are
31 * saved/restored to/from the HW context image. The list is emitted (via Load
32 * Register Immediate commands) once when initializing the device and saved in
33 * the default context. That default context is then used on every context
34 * creation to have a "primed golden context", i.e. a context image that
35 * already contains the changes needed to all the registers.
36 *
37 * Context workarounds should be implemented in the \*_ctx_workarounds_init()
38 * variants respective to the targeted platforms.
39 *
40 * - Engine workarounds: the list of these WAs is applied whenever the specific
41 * engine is reset. It's also possible that a set of engine classes share a
42 * common power domain and they are reset together. This happens on some
43 * platforms with render and compute engines. In this case (at least) one of
44 * them need to keeep the workaround programming: the approach taken in the
45 * driver is to tie those workarounds to the first compute/render engine that
46 * is registered. When executing with GuC submission, engine resets are
47 * outside of kernel driver control, hence the list of registers involved in
48 * written once, on engine initialization, and then passed to GuC, that
49 * saves/restores their values before/after the reset takes place. See
50 * ``drivers/gpu/drm/i915/gt/uc/intel_guc_ads.c`` for reference.
51 *
52 * Workarounds for registers specific to RCS and CCS should be implemented in
53 * rcs_engine_wa_init() and ccs_engine_wa_init(), respectively; those for
54 * registers belonging to BCS, VCS or VECS should be implemented in
55 * xcs_engine_wa_init(). Workarounds for registers not belonging to a specific
56 * engine's MMIO range but that are part of of the common RCS/CCS reset domain
57 * should be implemented in general_render_compute_wa_init(). The settings
58 * about the CCS load balancing should be added in ccs_engine_wa_mode().
59 *
60 * - GT workarounds: the list of these WAs is applied whenever these registers
61 * revert to their default values: on GPU reset, suspend/resume [1]_, etc.
62 *
63 * GT workarounds should be implemented in the \*_gt_workarounds_init()
64 * variants respective to the targeted platforms.
65 *
66 * - Register whitelist: some workarounds need to be implemented in userspace,
67 * but need to touch privileged registers. The whitelist in the kernel
68 * instructs the hardware to allow the access to happen. From the kernel side,
69 * this is just a special case of a MMIO workaround (as we write the list of
70 * these to/be-whitelisted registers to some special HW registers).
71 *
72 * Register whitelisting should be done in the \*_whitelist_build() variants
73 * respective to the targeted platforms.
74 *
75 * - Workaround batchbuffers: buffers that get executed automatically by the
76 * hardware on every HW context restore. These buffers are created and
77 * programmed in the default context so the hardware always go through those
78 * programming sequences when switching contexts. The support for workaround
79 * batchbuffers is enabled these hardware mechanisms:
80 *
81 * #. INDIRECT_CTX: A batchbuffer and an offset are provided in the default
82 * context, pointing the hardware to jump to that location when that offset
83 * is reached in the context restore. Workaround batchbuffer in the driver
84 * currently uses this mechanism for all platforms.
85 *
86 * #. BB_PER_CTX_PTR: A batchbuffer is provided in the default context,
87 * pointing the hardware to a buffer to continue executing after the
88 * engine registers are restored in a context restore sequence. This is
89 * currently not used in the driver.
90 *
91 * - Other: There are WAs that, due to their nature, cannot be applied from a
92 * central place. Those are peppered around the rest of the code, as needed.
93 * Workarounds related to the display IP are the main example.
94 *
95 * .. [1] Technically, some registers are powercontext saved & restored, so they
96 * survive a suspend/resume. In practice, writing them again is not too
97 * costly and simplifies things, so it's the approach taken in the driver.
98 */
99
wa_init_start(struct i915_wa_list * wal,struct intel_gt * gt,const char * name,const char * engine_name)100 static void wa_init_start(struct i915_wa_list *wal, struct intel_gt *gt,
101 const char *name, const char *engine_name)
102 {
103 wal->gt = gt;
104 wal->name = name;
105 wal->engine_name = engine_name;
106 }
107
108 #define WA_LIST_CHUNK (1 << 4)
109
wa_init_finish(struct i915_wa_list * wal)110 static void wa_init_finish(struct i915_wa_list *wal)
111 {
112 /* Trim unused entries. */
113 if (!IS_ALIGNED(wal->count, WA_LIST_CHUNK)) {
114 struct i915_wa *list = kmemdup_array(wal->list, wal->count,
115 sizeof(*list), GFP_KERNEL);
116
117 if (list) {
118 kfree(wal->list);
119 wal->list = list;
120 }
121 }
122
123 if (!wal->count)
124 return;
125
126 gt_dbg(wal->gt, "Initialized %u %s workarounds on %s\n",
127 wal->wa_count, wal->name, wal->engine_name);
128 }
129
130 static enum forcewake_domains
wal_get_fw_for_rmw(struct intel_uncore * uncore,const struct i915_wa_list * wal)131 wal_get_fw_for_rmw(struct intel_uncore *uncore, const struct i915_wa_list *wal)
132 {
133 enum forcewake_domains fw = 0;
134 struct i915_wa *wa;
135 unsigned int i;
136
137 for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
138 fw |= intel_uncore_forcewake_for_reg(uncore,
139 wa->reg,
140 FW_REG_READ |
141 FW_REG_WRITE);
142
143 return fw;
144 }
145
_wa_add(struct i915_wa_list * wal,const struct i915_wa * wa)146 static void _wa_add(struct i915_wa_list *wal, const struct i915_wa *wa)
147 {
148 unsigned int addr = i915_mmio_reg_offset(wa->reg);
149 struct drm_i915_private *i915 = wal->gt->i915;
150 unsigned int start = 0, end = wal->count;
151 const unsigned int grow = WA_LIST_CHUNK;
152 struct i915_wa *wa_;
153
154 GEM_BUG_ON(!is_power_of_2(grow));
155
156 if (IS_ALIGNED(wal->count, grow)) { /* Either uninitialized or full. */
157 struct i915_wa *list;
158
159 list = kmalloc_array(ALIGN(wal->count + 1, grow), sizeof(*wa),
160 GFP_KERNEL);
161 if (!list) {
162 drm_err(&i915->drm, "No space for workaround init!\n");
163 return;
164 }
165
166 if (wal->list) {
167 memcpy(list, wal->list, sizeof(*wa) * wal->count);
168 kfree(wal->list);
169 }
170
171 wal->list = list;
172 }
173
174 while (start < end) {
175 unsigned int mid = start + (end - start) / 2;
176
177 if (i915_mmio_reg_offset(wal->list[mid].reg) < addr) {
178 start = mid + 1;
179 } else if (i915_mmio_reg_offset(wal->list[mid].reg) > addr) {
180 end = mid;
181 } else {
182 wa_ = &wal->list[mid];
183
184 if ((wa->clr | wa_->clr) && !(wa->clr & ~wa_->clr)) {
185 drm_err(&i915->drm,
186 "Discarding overwritten w/a for reg %04x (clear: %08x, set: %08x)\n",
187 i915_mmio_reg_offset(wa_->reg),
188 wa_->clr, wa_->set);
189
190 wa_->set &= ~wa->clr;
191 }
192
193 wal->wa_count++;
194 wa_->set |= wa->set;
195 wa_->clr |= wa->clr;
196 wa_->read |= wa->read;
197 return;
198 }
199 }
200
201 wal->wa_count++;
202 wa_ = &wal->list[wal->count++];
203 *wa_ = *wa;
204
205 while (wa_-- > wal->list) {
206 GEM_BUG_ON(i915_mmio_reg_offset(wa_[0].reg) ==
207 i915_mmio_reg_offset(wa_[1].reg));
208 if (i915_mmio_reg_offset(wa_[1].reg) >
209 i915_mmio_reg_offset(wa_[0].reg))
210 break;
211
212 swap(wa_[1], wa_[0]);
213 }
214 }
215
wa_add(struct i915_wa_list * wal,i915_reg_t reg,u32 clear,u32 set,u32 read_mask,bool masked_reg)216 static void wa_add(struct i915_wa_list *wal, i915_reg_t reg,
217 u32 clear, u32 set, u32 read_mask, bool masked_reg)
218 {
219 struct i915_wa wa = {
220 .reg = reg,
221 .clr = clear,
222 .set = set,
223 .read = read_mask,
224 .masked_reg = masked_reg,
225 };
226
227 _wa_add(wal, &wa);
228 }
229
wa_mcr_add(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 clear,u32 set,u32 read_mask,bool masked_reg)230 static void wa_mcr_add(struct i915_wa_list *wal, i915_mcr_reg_t reg,
231 u32 clear, u32 set, u32 read_mask, bool masked_reg)
232 {
233 struct i915_wa wa = {
234 .mcr_reg = reg,
235 .clr = clear,
236 .set = set,
237 .read = read_mask,
238 .masked_reg = masked_reg,
239 .is_mcr = 1,
240 };
241
242 _wa_add(wal, &wa);
243 }
244
245 static void
wa_write_clr_set(struct i915_wa_list * wal,i915_reg_t reg,u32 clear,u32 set)246 wa_write_clr_set(struct i915_wa_list *wal, i915_reg_t reg, u32 clear, u32 set)
247 {
248 wa_add(wal, reg, clear, set, clear | set, false);
249 }
250
251 static void
wa_mcr_write_clr_set(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 clear,u32 set)252 wa_mcr_write_clr_set(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 clear, u32 set)
253 {
254 wa_mcr_add(wal, reg, clear, set, clear | set, false);
255 }
256
257 static void
wa_write(struct i915_wa_list * wal,i915_reg_t reg,u32 set)258 wa_write(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
259 {
260 wa_write_clr_set(wal, reg, ~0, set);
261 }
262
263 static void
wa_write_or(struct i915_wa_list * wal,i915_reg_t reg,u32 set)264 wa_write_or(struct i915_wa_list *wal, i915_reg_t reg, u32 set)
265 {
266 wa_write_clr_set(wal, reg, set, set);
267 }
268
269 static void
wa_mcr_write_or(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 set)270 wa_mcr_write_or(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 set)
271 {
272 wa_mcr_write_clr_set(wal, reg, set, set);
273 }
274
275 static void
wa_write_clr(struct i915_wa_list * wal,i915_reg_t reg,u32 clr)276 wa_write_clr(struct i915_wa_list *wal, i915_reg_t reg, u32 clr)
277 {
278 wa_write_clr_set(wal, reg, clr, 0);
279 }
280
281 static void
wa_mcr_write_clr(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 clr)282 wa_mcr_write_clr(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 clr)
283 {
284 wa_mcr_write_clr_set(wal, reg, clr, 0);
285 }
286
287 /*
288 * WA operations on "masked register". A masked register has the upper 16 bits
289 * documented as "masked" in b-spec. Its purpose is to allow writing to just a
290 * portion of the register without a rmw: you simply write in the upper 16 bits
291 * the mask of bits you are going to modify.
292 *
293 * The wa_masked_* family of functions already does the necessary operations to
294 * calculate the mask based on the parameters passed, so user only has to
295 * provide the lower 16 bits of that register.
296 */
297
298 static void
wa_masked_en(struct i915_wa_list * wal,i915_reg_t reg,u32 val)299 wa_masked_en(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
300 {
301 wa_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
302 }
303
304 static void
wa_mcr_masked_en(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 val)305 wa_mcr_masked_en(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 val)
306 {
307 wa_mcr_add(wal, reg, 0, _MASKED_BIT_ENABLE(val), val, true);
308 }
309
310 static void
wa_masked_dis(struct i915_wa_list * wal,i915_reg_t reg,u32 val)311 wa_masked_dis(struct i915_wa_list *wal, i915_reg_t reg, u32 val)
312 {
313 wa_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
314 }
315
316 static void
wa_mcr_masked_dis(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 val)317 wa_mcr_masked_dis(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 val)
318 {
319 wa_mcr_add(wal, reg, 0, _MASKED_BIT_DISABLE(val), val, true);
320 }
321
322 static void
wa_masked_field_set(struct i915_wa_list * wal,i915_reg_t reg,u32 mask,u32 val)323 wa_masked_field_set(struct i915_wa_list *wal, i915_reg_t reg,
324 u32 mask, u32 val)
325 {
326 wa_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
327 }
328
329 static void
wa_mcr_masked_field_set(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 mask,u32 val)330 wa_mcr_masked_field_set(struct i915_wa_list *wal, i915_mcr_reg_t reg,
331 u32 mask, u32 val)
332 {
333 wa_mcr_add(wal, reg, 0, _MASKED_FIELD(mask, val), mask, true);
334 }
335
gen6_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)336 static void gen6_ctx_workarounds_init(struct intel_engine_cs *engine,
337 struct i915_wa_list *wal)
338 {
339 wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
340 }
341
gen7_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)342 static void gen7_ctx_workarounds_init(struct intel_engine_cs *engine,
343 struct i915_wa_list *wal)
344 {
345 wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
346 }
347
gen8_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)348 static void gen8_ctx_workarounds_init(struct intel_engine_cs *engine,
349 struct i915_wa_list *wal)
350 {
351 wa_masked_en(wal, INSTPM, INSTPM_FORCE_ORDERING);
352
353 /* WaDisableAsyncFlipPerfMode:bdw,chv */
354 wa_masked_en(wal, RING_MI_MODE(RENDER_RING_BASE), ASYNC_FLIP_PERF_DISABLE);
355
356 /* WaDisablePartialInstShootdown:bdw,chv */
357 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
358 PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
359
360 /* Use Force Non-Coherent whenever executing a 3D context. This is a
361 * workaround for a possible hang in the unlikely event a TLB
362 * invalidation occurs during a PSD flush.
363 */
364 /* WaForceEnableNonCoherent:bdw,chv */
365 /* WaHdcDisableFetchWhenMasked:bdw,chv */
366 wa_masked_en(wal, HDC_CHICKEN0,
367 HDC_DONOT_FETCH_MEM_WHEN_MASKED |
368 HDC_FORCE_NON_COHERENT);
369
370 /* From the Haswell PRM, Command Reference: Registers, CACHE_MODE_0:
371 * "The Hierarchical Z RAW Stall Optimization allows non-overlapping
372 * polygons in the same 8x4 pixel/sample area to be processed without
373 * stalling waiting for the earlier ones to write to Hierarchical Z
374 * buffer."
375 *
376 * This optimization is off by default for BDW and CHV; turn it on.
377 */
378 wa_masked_dis(wal, CACHE_MODE_0_GEN7, HIZ_RAW_STALL_OPT_DISABLE);
379
380 /* Wa4x4STCOptimizationDisable:bdw,chv */
381 wa_masked_en(wal, CACHE_MODE_1, GEN8_4x4_STC_OPTIMIZATION_DISABLE);
382
383 /*
384 * BSpec recommends 8x4 when MSAA is used,
385 * however in practice 16x4 seems fastest.
386 *
387 * Note that PS/WM thread counts depend on the WIZ hashing
388 * disable bit, which we don't touch here, but it's good
389 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
390 */
391 wa_masked_field_set(wal, GEN7_GT_MODE,
392 GEN6_WIZ_HASHING_MASK,
393 GEN6_WIZ_HASHING_16x4);
394 }
395
bdw_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)396 static void bdw_ctx_workarounds_init(struct intel_engine_cs *engine,
397 struct i915_wa_list *wal)
398 {
399 struct drm_i915_private *i915 = engine->i915;
400
401 gen8_ctx_workarounds_init(engine, wal);
402
403 /* WaDisableThreadStallDopClockGating:bdw (pre-production) */
404 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
405
406 /* WaDisableDopClockGating:bdw
407 *
408 * Also see the related UCGTCL1 write in bdw_init_clock_gating()
409 * to disable EUTC clock gating.
410 */
411 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
412 DOP_CLOCK_GATING_DISABLE);
413
414 wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN3,
415 GEN8_SAMPLER_POWER_BYPASS_DIS);
416
417 wa_masked_en(wal, HDC_CHICKEN0,
418 /* WaForceContextSaveRestoreNonCoherent:bdw */
419 HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
420 /* WaDisableFenceDestinationToSLM:bdw (pre-prod) */
421 (IS_BROADWELL_GT3(i915) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
422 }
423
chv_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)424 static void chv_ctx_workarounds_init(struct intel_engine_cs *engine,
425 struct i915_wa_list *wal)
426 {
427 gen8_ctx_workarounds_init(engine, wal);
428
429 /* WaDisableThreadStallDopClockGating:chv */
430 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
431
432 /* Improve HiZ throughput on CHV. */
433 wa_masked_en(wal, HIZ_CHICKEN, CHV_HZ_8X8_MODE_IN_1X);
434 }
435
gen9_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)436 static void gen9_ctx_workarounds_init(struct intel_engine_cs *engine,
437 struct i915_wa_list *wal)
438 {
439 struct drm_i915_private *i915 = engine->i915;
440
441 if (HAS_LLC(i915)) {
442 /* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
443 *
444 * Must match Display Engine. See
445 * WaCompressedResourceDisplayNewHashMode.
446 */
447 wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
448 GEN9_PBE_COMPRESSED_HASH_SELECTION);
449 wa_mcr_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
450 GEN9_SAMPLER_HASH_COMPRESSED_READ_ADDR);
451 }
452
453 /* WaClearFlowControlGpgpuContextSave:skl,bxt,kbl,glk,cfl */
454 /* WaDisablePartialInstShootdown:skl,bxt,kbl,glk,cfl */
455 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
456 FLOW_CONTROL_ENABLE |
457 PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
458
459 /* WaEnableYV12BugFixInHalfSliceChicken7:skl,bxt,kbl,glk,cfl */
460 /* WaEnableSamplerGPGPUPreemptionSupport:skl,bxt,kbl,cfl */
461 wa_mcr_masked_en(wal, GEN9_HALF_SLICE_CHICKEN7,
462 GEN9_ENABLE_YV12_BUGFIX |
463 GEN9_ENABLE_GPGPU_PREEMPTION);
464
465 /* Wa4x4STCOptimizationDisable:skl,bxt,kbl,glk,cfl */
466 /* WaDisablePartialResolveInVc:skl,bxt,kbl,cfl */
467 wa_masked_en(wal, CACHE_MODE_1,
468 GEN8_4x4_STC_OPTIMIZATION_DISABLE |
469 GEN9_PARTIAL_RESOLVE_IN_VC_DISABLE);
470
471 /* WaCcsTlbPrefetchDisable:skl,bxt,kbl,glk,cfl */
472 wa_mcr_masked_dis(wal, GEN9_HALF_SLICE_CHICKEN5,
473 GEN9_CCS_TLB_PREFETCH_ENABLE);
474
475 /* WaForceContextSaveRestoreNonCoherent:skl,bxt,kbl,cfl */
476 wa_masked_en(wal, HDC_CHICKEN0,
477 HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
478 HDC_FORCE_CSR_NON_COHERENT_OVR_DISABLE);
479
480 /* WaForceEnableNonCoherent and WaDisableHDCInvalidation are
481 * both tied to WaForceContextSaveRestoreNonCoherent
482 * in some hsds for skl. We keep the tie for all gen9. The
483 * documentation is a bit hazy and so we want to get common behaviour,
484 * even though there is no clear evidence we would need both on kbl/bxt.
485 * This area has been source of system hangs so we play it safe
486 * and mimic the skl regardless of what bspec says.
487 *
488 * Use Force Non-Coherent whenever executing a 3D context. This
489 * is a workaround for a possible hang in the unlikely event
490 * a TLB invalidation occurs during a PSD flush.
491 */
492
493 /* WaForceEnableNonCoherent:skl,bxt,kbl,cfl */
494 wa_masked_en(wal, HDC_CHICKEN0,
495 HDC_FORCE_NON_COHERENT);
496
497 /* WaDisableSamplerPowerBypassForSOPingPong:skl,bxt,kbl,cfl */
498 if (IS_SKYLAKE(i915) ||
499 IS_KABYLAKE(i915) ||
500 IS_COFFEELAKE(i915) ||
501 IS_COMETLAKE(i915))
502 wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN3,
503 GEN8_SAMPLER_POWER_BYPASS_DIS);
504
505 /* WaDisableSTUnitPowerOptimization:skl,bxt,kbl,glk,cfl */
506 wa_mcr_masked_en(wal, HALF_SLICE_CHICKEN2, GEN8_ST_PO_DISABLE);
507
508 /*
509 * Supporting preemption with fine-granularity requires changes in the
510 * batch buffer programming. Since we can't break old userspace, we
511 * need to set our default preemption level to safe value. Userspace is
512 * still able to use more fine-grained preemption levels, since in
513 * WaEnablePreemptionGranularityControlByUMD we're whitelisting the
514 * per-ctx register. As such, WaDisable{3D,GPGPU}MidCmdPreemption are
515 * not real HW workarounds, but merely a way to start using preemption
516 * while maintaining old contract with userspace.
517 */
518
519 /* WaDisable3DMidCmdPreemption:skl,bxt,glk,cfl,[cnl] */
520 wa_masked_dis(wal, GEN8_CS_CHICKEN1, GEN9_PREEMPT_3D_OBJECT_LEVEL);
521
522 /* WaDisableGPGPUMidCmdPreemption:skl,bxt,blk,cfl,[cnl] */
523 wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
524 GEN9_PREEMPT_GPGPU_LEVEL_MASK,
525 GEN9_PREEMPT_GPGPU_COMMAND_LEVEL);
526
527 /* WaClearHIZ_WM_CHICKEN3:bxt,glk */
528 if (IS_GEN9_LP(i915))
529 wa_masked_en(wal, GEN9_WM_CHICKEN3, GEN9_FACTOR_IN_CLR_VAL_HIZ);
530 }
531
skl_tune_iz_hashing(struct intel_engine_cs * engine,struct i915_wa_list * wal)532 static void skl_tune_iz_hashing(struct intel_engine_cs *engine,
533 struct i915_wa_list *wal)
534 {
535 struct intel_gt *gt = engine->gt;
536 u8 vals[3] = { 0, 0, 0 };
537 unsigned int i;
538
539 for (i = 0; i < 3; i++) {
540 u8 ss;
541
542 /*
543 * Only consider slices where one, and only one, subslice has 7
544 * EUs
545 */
546 if (!is_power_of_2(gt->info.sseu.subslice_7eu[i]))
547 continue;
548
549 /*
550 * subslice_7eu[i] != 0 (because of the check above) and
551 * ss_max == 4 (maximum number of subslices possible per slice)
552 *
553 * -> 0 <= ss <= 3;
554 */
555 ss = ffs(gt->info.sseu.subslice_7eu[i]) - 1;
556 vals[i] = 3 - ss;
557 }
558
559 if (vals[0] == 0 && vals[1] == 0 && vals[2] == 0)
560 return;
561
562 /* Tune IZ hashing. See intel_device_info_runtime_init() */
563 wa_masked_field_set(wal, GEN7_GT_MODE,
564 GEN9_IZ_HASHING_MASK(2) |
565 GEN9_IZ_HASHING_MASK(1) |
566 GEN9_IZ_HASHING_MASK(0),
567 GEN9_IZ_HASHING(2, vals[2]) |
568 GEN9_IZ_HASHING(1, vals[1]) |
569 GEN9_IZ_HASHING(0, vals[0]));
570 }
571
skl_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)572 static void skl_ctx_workarounds_init(struct intel_engine_cs *engine,
573 struct i915_wa_list *wal)
574 {
575 gen9_ctx_workarounds_init(engine, wal);
576 skl_tune_iz_hashing(engine, wal);
577 }
578
bxt_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)579 static void bxt_ctx_workarounds_init(struct intel_engine_cs *engine,
580 struct i915_wa_list *wal)
581 {
582 gen9_ctx_workarounds_init(engine, wal);
583
584 /* WaDisableThreadStallDopClockGating:bxt */
585 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN,
586 STALL_DOP_GATING_DISABLE);
587
588 /* WaToEnableHwFixForPushConstHWBug:bxt */
589 wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
590 GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
591 }
592
kbl_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)593 static void kbl_ctx_workarounds_init(struct intel_engine_cs *engine,
594 struct i915_wa_list *wal)
595 {
596 struct drm_i915_private *i915 = engine->i915;
597
598 gen9_ctx_workarounds_init(engine, wal);
599
600 /* WaToEnableHwFixForPushConstHWBug:kbl */
601 if (IS_KABYLAKE(i915) && IS_GRAPHICS_STEP(i915, STEP_C0, STEP_FOREVER))
602 wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
603 GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
604
605 /* WaDisableSbeCacheDispatchPortSharing:kbl */
606 wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN1,
607 GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
608 }
609
glk_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)610 static void glk_ctx_workarounds_init(struct intel_engine_cs *engine,
611 struct i915_wa_list *wal)
612 {
613 gen9_ctx_workarounds_init(engine, wal);
614
615 /* WaToEnableHwFixForPushConstHWBug:glk */
616 wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
617 GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
618 }
619
cfl_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)620 static void cfl_ctx_workarounds_init(struct intel_engine_cs *engine,
621 struct i915_wa_list *wal)
622 {
623 gen9_ctx_workarounds_init(engine, wal);
624
625 /* WaToEnableHwFixForPushConstHWBug:cfl */
626 wa_masked_en(wal, COMMON_SLICE_CHICKEN2,
627 GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
628
629 /* WaDisableSbeCacheDispatchPortSharing:cfl */
630 wa_mcr_masked_en(wal, GEN8_HALF_SLICE_CHICKEN1,
631 GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
632 }
633
icl_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)634 static void icl_ctx_workarounds_init(struct intel_engine_cs *engine,
635 struct i915_wa_list *wal)
636 {
637 /* Wa_1406697149 (WaDisableBankHangMode:icl) */
638 wa_write(wal, GEN8_L3CNTLREG, GEN8_ERRDETBCTRL);
639
640 /* WaForceEnableNonCoherent:icl
641 * This is not the same workaround as in early Gen9 platforms, where
642 * lacking this could cause system hangs, but coherency performance
643 * overhead is high and only a few compute workloads really need it
644 * (the register is whitelisted in hardware now, so UMDs can opt in
645 * for coherency if they have a good reason).
646 */
647 wa_mcr_masked_en(wal, ICL_HDC_MODE, HDC_FORCE_NON_COHERENT);
648
649 /* WaEnableFloatBlendOptimization:icl */
650 wa_mcr_add(wal, GEN10_CACHE_MODE_SS, 0,
651 _MASKED_BIT_ENABLE(FLOAT_BLEND_OPTIMIZATION_ENABLE),
652 0 /* write-only, so skip validation */,
653 true);
654
655 /* WaDisableGPGPUMidThreadPreemption:icl */
656 wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
657 GEN9_PREEMPT_GPGPU_LEVEL_MASK,
658 GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
659
660 /* allow headerless messages for preemptible GPGPU context */
661 wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
662 GEN11_SAMPLER_ENABLE_HEADLESS_MSG);
663
664 /* Wa_1604278689:icl,ehl */
665 wa_write(wal, IVB_FBC_RT_BASE, 0xFFFFFFFF & ~ILK_FBC_RT_VALID);
666 wa_write_clr_set(wal, IVB_FBC_RT_BASE_UPPER,
667 0,
668 0xFFFFFFFF);
669
670 /* Wa_1406306137:icl,ehl */
671 wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, GEN11_DIS_PICK_2ND_EU);
672 }
673
674 /*
675 * These settings aren't actually workarounds, but general tuning settings that
676 * need to be programmed on dg2 platform.
677 */
dg2_ctx_gt_tuning_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)678 static void dg2_ctx_gt_tuning_init(struct intel_engine_cs *engine,
679 struct i915_wa_list *wal)
680 {
681 wa_mcr_masked_en(wal, CHICKEN_RASTER_2, TBIMR_FAST_CLIP);
682 wa_mcr_write_clr_set(wal, XEHP_L3SQCREG5, L3_PWM_TIMER_INIT_VAL_MASK,
683 REG_FIELD_PREP(L3_PWM_TIMER_INIT_VAL_MASK, 0x7f));
684 wa_mcr_write_clr_set(wal, XEHP_FF_MODE2, FF_MODE2_TDS_TIMER_MASK,
685 FF_MODE2_TDS_TIMER_128);
686 }
687
gen12_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)688 static void gen12_ctx_workarounds_init(struct intel_engine_cs *engine,
689 struct i915_wa_list *wal)
690 {
691 struct drm_i915_private *i915 = engine->i915;
692
693 /*
694 * Wa_1409142259:tgl,dg1,adl-p
695 * Wa_1409347922:tgl,dg1,adl-p
696 * Wa_1409252684:tgl,dg1,adl-p
697 * Wa_1409217633:tgl,dg1,adl-p
698 * Wa_1409207793:tgl,dg1,adl-p
699 * Wa_1409178076:tgl,dg1,adl-p
700 * Wa_1408979724:tgl,dg1,adl-p
701 * Wa_14010443199:tgl,rkl,dg1,adl-p
702 * Wa_14010698770:tgl,rkl,dg1,adl-s,adl-p
703 * Wa_1409342910:tgl,rkl,dg1,adl-s,adl-p
704 */
705 wa_masked_en(wal, GEN11_COMMON_SLICE_CHICKEN3,
706 GEN12_DISABLE_CPS_AWARE_COLOR_PIPE);
707
708 /* WaDisableGPGPUMidThreadPreemption:gen12 */
709 wa_masked_field_set(wal, GEN8_CS_CHICKEN1,
710 GEN9_PREEMPT_GPGPU_LEVEL_MASK,
711 GEN9_PREEMPT_GPGPU_THREAD_GROUP_LEVEL);
712
713 /*
714 * Wa_16011163337 - GS_TIMER
715 *
716 * TDS_TIMER: Although some platforms refer to it as Wa_1604555607, we
717 * need to program it even on those that don't explicitly list that
718 * workaround.
719 *
720 * Note that the programming of GEN12_FF_MODE2 is further modified
721 * according to the FF_MODE2 guidance given by Wa_1608008084.
722 * Wa_1608008084 tells us the FF_MODE2 register will return the wrong
723 * value when read from the CPU.
724 *
725 * The default value for this register is zero for all fields.
726 * So instead of doing a RMW we should just write the desired values
727 * for TDS and GS timers. Note that since the readback can't be trusted,
728 * the clear mask is just set to ~0 to make sure other bits are not
729 * inadvertently set. For the same reason read verification is ignored.
730 */
731 wa_add(wal,
732 GEN12_FF_MODE2,
733 ~0,
734 FF_MODE2_TDS_TIMER_128 | FF_MODE2_GS_TIMER_224,
735 0, false);
736
737 if (!IS_DG1(i915)) {
738 /* Wa_1806527549 */
739 wa_masked_en(wal, HIZ_CHICKEN, HZ_DEPTH_TEST_LE_GE_OPT_DISABLE);
740
741 /* Wa_1606376872 */
742 wa_masked_en(wal, COMMON_SLICE_CHICKEN4, DISABLE_TDC_LOAD_BALANCING_CALC);
743 }
744 }
745
dg1_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)746 static void dg1_ctx_workarounds_init(struct intel_engine_cs *engine,
747 struct i915_wa_list *wal)
748 {
749 gen12_ctx_workarounds_init(engine, wal);
750
751 /* Wa_1409044764 */
752 wa_masked_dis(wal, GEN11_COMMON_SLICE_CHICKEN3,
753 DG1_FLOAT_POINT_BLEND_OPT_STRICT_MODE_EN);
754
755 /* Wa_22010493298 */
756 wa_masked_en(wal, HIZ_CHICKEN,
757 DG1_HZ_READ_SUPPRESSION_OPTIMIZATION_DISABLE);
758 }
759
dg2_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)760 static void dg2_ctx_workarounds_init(struct intel_engine_cs *engine,
761 struct i915_wa_list *wal)
762 {
763 dg2_ctx_gt_tuning_init(engine, wal);
764
765 /* Wa_16013271637:dg2 */
766 wa_mcr_masked_en(wal, XEHP_SLICE_COMMON_ECO_CHICKEN1,
767 MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
768
769 /* Wa_14014947963:dg2 */
770 wa_masked_field_set(wal, VF_PREEMPTION, PREEMPTION_VERTEX_COUNT, 0x4000);
771
772 /* Wa_18018764978:dg2 */
773 wa_mcr_masked_en(wal, XEHP_PSS_MODE2, SCOREBOARD_STALL_FLUSH_CONTROL);
774
775 /* Wa_18019271663:dg2 */
776 wa_masked_en(wal, CACHE_MODE_1, MSAA_OPTIMIZATION_REDUC_DISABLE);
777
778 /* Wa_14019877138:dg2 */
779 wa_mcr_masked_en(wal, XEHP_PSS_CHICKEN, FD_END_COLLECT);
780 }
781
xelpg_ctx_gt_tuning_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)782 static void xelpg_ctx_gt_tuning_init(struct intel_engine_cs *engine,
783 struct i915_wa_list *wal)
784 {
785 struct intel_gt *gt = engine->gt;
786
787 dg2_ctx_gt_tuning_init(engine, wal);
788
789 /*
790 * Due to Wa_16014892111, the DRAW_WATERMARK tuning must be done in
791 * gen12_emit_indirect_ctx_rcs() rather than here on some early
792 * steppings.
793 */
794 if (!(IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
795 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)))
796 wa_add(wal, DRAW_WATERMARK, VERT_WM_VAL, 0x3FF, 0, false);
797 }
798
xelpg_ctx_workarounds_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)799 static void xelpg_ctx_workarounds_init(struct intel_engine_cs *engine,
800 struct i915_wa_list *wal)
801 {
802 struct intel_gt *gt = engine->gt;
803
804 xelpg_ctx_gt_tuning_init(engine, wal);
805
806 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
807 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
808 /* Wa_14014947963 */
809 wa_masked_field_set(wal, VF_PREEMPTION,
810 PREEMPTION_VERTEX_COUNT, 0x4000);
811
812 /* Wa_16013271637 */
813 wa_mcr_masked_en(wal, XEHP_SLICE_COMMON_ECO_CHICKEN1,
814 MSC_MSAA_REODER_BUF_BYPASS_DISABLE);
815
816 /* Wa_18019627453 */
817 wa_mcr_masked_en(wal, VFLSKPD, VF_PREFETCH_TLB_DIS);
818
819 /* Wa_18018764978 */
820 wa_mcr_masked_en(wal, XEHP_PSS_MODE2, SCOREBOARD_STALL_FLUSH_CONTROL);
821 }
822
823 /* Wa_18019271663 */
824 wa_masked_en(wal, CACHE_MODE_1, MSAA_OPTIMIZATION_REDUC_DISABLE);
825
826 /* Wa_14019877138 */
827 wa_mcr_masked_en(wal, XEHP_PSS_CHICKEN, FD_END_COLLECT);
828 }
829
fakewa_disable_nestedbb_mode(struct intel_engine_cs * engine,struct i915_wa_list * wal)830 static void fakewa_disable_nestedbb_mode(struct intel_engine_cs *engine,
831 struct i915_wa_list *wal)
832 {
833 /*
834 * This is a "fake" workaround defined by software to ensure we
835 * maintain reliable, backward-compatible behavior for userspace with
836 * regards to how nested MI_BATCH_BUFFER_START commands are handled.
837 *
838 * The per-context setting of MI_MODE[12] determines whether the bits
839 * of a nested MI_BATCH_BUFFER_START instruction should be interpreted
840 * in the traditional manner or whether they should instead use a new
841 * tgl+ meaning that breaks backward compatibility, but allows nesting
842 * into 3rd-level batchbuffers. When this new capability was first
843 * added in TGL, it remained off by default unless a context
844 * intentionally opted in to the new behavior. However Xe_HPG now
845 * flips this on by default and requires that we explicitly opt out if
846 * we don't want the new behavior.
847 *
848 * From a SW perspective, we want to maintain the backward-compatible
849 * behavior for userspace, so we'll apply a fake workaround to set it
850 * back to the legacy behavior on platforms where the hardware default
851 * is to break compatibility. At the moment there is no Linux
852 * userspace that utilizes third-level batchbuffers, so this will avoid
853 * userspace from needing to make any changes. using the legacy
854 * meaning is the correct thing to do. If/when we have userspace
855 * consumers that want to utilize third-level batch nesting, we can
856 * provide a context parameter to allow them to opt-in.
857 */
858 wa_masked_dis(wal, RING_MI_MODE(engine->mmio_base), TGL_NESTED_BB_EN);
859 }
860
gen12_ctx_gt_mocs_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)861 static void gen12_ctx_gt_mocs_init(struct intel_engine_cs *engine,
862 struct i915_wa_list *wal)
863 {
864 u8 mocs;
865
866 /*
867 * Some blitter commands do not have a field for MOCS, those
868 * commands will use MOCS index pointed by BLIT_CCTL.
869 * BLIT_CCTL registers are needed to be programmed to un-cached.
870 */
871 if (engine->class == COPY_ENGINE_CLASS) {
872 mocs = engine->gt->mocs.uc_index;
873 wa_write_clr_set(wal,
874 BLIT_CCTL(engine->mmio_base),
875 BLIT_CCTL_MASK,
876 BLIT_CCTL_MOCS(mocs, mocs));
877 }
878 }
879
880 /*
881 * gen12_ctx_gt_fake_wa_init() aren't programmingan official workaround
882 * defined by the hardware team, but it programming general context registers.
883 * Adding those context register programming in context workaround
884 * allow us to use the wa framework for proper application and validation.
885 */
886 static void
gen12_ctx_gt_fake_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)887 gen12_ctx_gt_fake_wa_init(struct intel_engine_cs *engine,
888 struct i915_wa_list *wal)
889 {
890 if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
891 fakewa_disable_nestedbb_mode(engine, wal);
892
893 gen12_ctx_gt_mocs_init(engine, wal);
894 }
895
896 static void
__intel_engine_init_ctx_wa(struct intel_engine_cs * engine,struct i915_wa_list * wal,const char * name)897 __intel_engine_init_ctx_wa(struct intel_engine_cs *engine,
898 struct i915_wa_list *wal,
899 const char *name)
900 {
901 struct drm_i915_private *i915 = engine->i915;
902
903 wa_init_start(wal, engine->gt, name, engine->name);
904
905 /* Applies to all engines */
906 /*
907 * Fake workarounds are not the actual workaround but
908 * programming of context registers using workaround framework.
909 */
910 if (GRAPHICS_VER(i915) >= 12)
911 gen12_ctx_gt_fake_wa_init(engine, wal);
912
913 if (engine->class != RENDER_CLASS)
914 goto done;
915
916 if (IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 70), IP_VER(12, 74)))
917 xelpg_ctx_workarounds_init(engine, wal);
918 else if (IS_DG2(i915))
919 dg2_ctx_workarounds_init(engine, wal);
920 else if (IS_DG1(i915))
921 dg1_ctx_workarounds_init(engine, wal);
922 else if (GRAPHICS_VER(i915) == 12)
923 gen12_ctx_workarounds_init(engine, wal);
924 else if (GRAPHICS_VER(i915) == 11)
925 icl_ctx_workarounds_init(engine, wal);
926 else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
927 cfl_ctx_workarounds_init(engine, wal);
928 else if (IS_GEMINILAKE(i915))
929 glk_ctx_workarounds_init(engine, wal);
930 else if (IS_KABYLAKE(i915))
931 kbl_ctx_workarounds_init(engine, wal);
932 else if (IS_BROXTON(i915))
933 bxt_ctx_workarounds_init(engine, wal);
934 else if (IS_SKYLAKE(i915))
935 skl_ctx_workarounds_init(engine, wal);
936 else if (IS_CHERRYVIEW(i915))
937 chv_ctx_workarounds_init(engine, wal);
938 else if (IS_BROADWELL(i915))
939 bdw_ctx_workarounds_init(engine, wal);
940 else if (GRAPHICS_VER(i915) == 7)
941 gen7_ctx_workarounds_init(engine, wal);
942 else if (GRAPHICS_VER(i915) == 6)
943 gen6_ctx_workarounds_init(engine, wal);
944 else if (GRAPHICS_VER(i915) < 8)
945 ;
946 else
947 MISSING_CASE(GRAPHICS_VER(i915));
948
949 done:
950 wa_init_finish(wal);
951 }
952
intel_engine_init_ctx_wa(struct intel_engine_cs * engine)953 void intel_engine_init_ctx_wa(struct intel_engine_cs *engine)
954 {
955 __intel_engine_init_ctx_wa(engine, &engine->ctx_wa_list, "context");
956 }
957
intel_engine_emit_ctx_wa(struct i915_request * rq)958 int intel_engine_emit_ctx_wa(struct i915_request *rq)
959 {
960 struct i915_wa_list *wal = &rq->engine->ctx_wa_list;
961 struct intel_uncore *uncore = rq->engine->uncore;
962 enum forcewake_domains fw;
963 unsigned long flags;
964 struct i915_wa *wa;
965 unsigned int i;
966 u32 *cs;
967 int ret;
968
969 if (wal->count == 0)
970 return 0;
971
972 ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
973 if (ret)
974 return ret;
975
976 if ((IS_GFX_GT_IP_RANGE(rq->engine->gt, IP_VER(12, 70), IP_VER(12, 74)) ||
977 IS_DG2(rq->i915)) && rq->engine->class == RENDER_CLASS)
978 cs = intel_ring_begin(rq, (wal->count * 2 + 6));
979 else
980 cs = intel_ring_begin(rq, (wal->count * 2 + 2));
981
982 if (IS_ERR(cs))
983 return PTR_ERR(cs);
984
985 fw = wal_get_fw_for_rmw(uncore, wal);
986
987 intel_gt_mcr_lock(wal->gt, &flags);
988 spin_lock(&uncore->lock);
989 intel_uncore_forcewake_get__locked(uncore, fw);
990
991 *cs++ = MI_LOAD_REGISTER_IMM(wal->count);
992 for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
993 u32 val;
994
995 /* Skip reading the register if it's not really needed */
996 if (wa->masked_reg || (wa->clr | wa->set) == U32_MAX) {
997 val = wa->set;
998 } else {
999 val = wa->is_mcr ?
1000 intel_gt_mcr_read_any_fw(wal->gt, wa->mcr_reg) :
1001 intel_uncore_read_fw(uncore, wa->reg);
1002 val &= ~wa->clr;
1003 val |= wa->set;
1004 }
1005
1006 *cs++ = i915_mmio_reg_offset(wa->reg);
1007 *cs++ = val;
1008 }
1009 *cs++ = MI_NOOP;
1010
1011 /* Wa_14019789679 */
1012 if ((IS_GFX_GT_IP_RANGE(rq->engine->gt, IP_VER(12, 70), IP_VER(12, 74)) ||
1013 IS_DG2(rq->i915)) && rq->engine->class == RENDER_CLASS) {
1014 *cs++ = CMD_3DSTATE_MESH_CONTROL;
1015 *cs++ = 0;
1016 *cs++ = 0;
1017 *cs++ = MI_NOOP;
1018 }
1019
1020 intel_uncore_forcewake_put__locked(uncore, fw);
1021 spin_unlock(&uncore->lock);
1022 intel_gt_mcr_unlock(wal->gt, flags);
1023
1024 intel_ring_advance(rq, cs);
1025
1026 ret = rq->engine->emit_flush(rq, EMIT_BARRIER);
1027 if (ret)
1028 return ret;
1029
1030 return 0;
1031 }
1032
1033 static void
gen4_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1034 gen4_gt_workarounds_init(struct intel_gt *gt,
1035 struct i915_wa_list *wal)
1036 {
1037 /* WaDisable_RenderCache_OperationalFlush:gen4,ilk */
1038 wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
1039 }
1040
1041 static void
g4x_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1042 g4x_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1043 {
1044 gen4_gt_workarounds_init(gt, wal);
1045
1046 /* WaDisableRenderCachePipelinedFlush:g4x,ilk */
1047 wa_masked_en(wal, CACHE_MODE_0, CM0_PIPELINED_RENDER_FLUSH_DISABLE);
1048 }
1049
1050 static void
ilk_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1051 ilk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1052 {
1053 g4x_gt_workarounds_init(gt, wal);
1054
1055 wa_masked_en(wal, _3D_CHICKEN2, _3D_CHICKEN2_WM_READ_PIPELINED);
1056 }
1057
1058 static void
snb_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1059 snb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1060 {
1061 }
1062
1063 static void
ivb_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1064 ivb_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1065 {
1066 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
1067 wa_masked_dis(wal,
1068 GEN7_COMMON_SLICE_CHICKEN1,
1069 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
1070
1071 /* WaApplyL3ControlAndL3ChickenMode:ivb */
1072 wa_write(wal, GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
1073 wa_write(wal, GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
1074
1075 /* WaForceL3Serialization:ivb */
1076 wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
1077 }
1078
1079 static void
vlv_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1080 vlv_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1081 {
1082 /* WaForceL3Serialization:vlv */
1083 wa_write_clr(wal, GEN7_L3SQCREG4, L3SQ_URB_READ_CAM_MATCH_DISABLE);
1084
1085 /*
1086 * WaIncreaseL3CreditsForVLVB0:vlv
1087 * This is the hardware default actually.
1088 */
1089 wa_write(wal, GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
1090 }
1091
1092 static void
hsw_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1093 hsw_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1094 {
1095 /* L3 caching of data atomics doesn't work -- disable it. */
1096 wa_write(wal, HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
1097
1098 wa_add(wal,
1099 HSW_ROW_CHICKEN3, 0,
1100 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE),
1101 0 /* XXX does this reg exist? */, true);
1102
1103 /* WaVSRefCountFullforceMissDisable:hsw */
1104 wa_write_clr(wal, GEN7_FF_THREAD_MODE, GEN7_FF_VS_REF_CNT_FFME);
1105 }
1106
1107 static void
gen9_wa_init_mcr(struct drm_i915_private * i915,struct i915_wa_list * wal)1108 gen9_wa_init_mcr(struct drm_i915_private *i915, struct i915_wa_list *wal)
1109 {
1110 const struct sseu_dev_info *sseu = &to_gt(i915)->info.sseu;
1111 unsigned int slice, subslice;
1112 u32 mcr, mcr_mask;
1113
1114 GEM_BUG_ON(GRAPHICS_VER(i915) != 9);
1115
1116 /*
1117 * WaProgramMgsrForCorrectSliceSpecificMmioReads:gen9,glk,kbl,cml
1118 * Before any MMIO read into slice/subslice specific registers, MCR
1119 * packet control register needs to be programmed to point to any
1120 * enabled s/ss pair. Otherwise, incorrect values will be returned.
1121 * This means each subsequent MMIO read will be forwarded to an
1122 * specific s/ss combination, but this is OK since these registers
1123 * are consistent across s/ss in almost all cases. In the rare
1124 * occasions, such as INSTDONE, where this value is dependent
1125 * on s/ss combo, the read should be done with read_subslice_reg.
1126 */
1127 slice = ffs(sseu->slice_mask) - 1;
1128 GEM_BUG_ON(slice >= ARRAY_SIZE(sseu->subslice_mask.hsw));
1129 subslice = ffs(intel_sseu_get_hsw_subslices(sseu, slice));
1130 GEM_BUG_ON(!subslice);
1131 subslice--;
1132
1133 /*
1134 * We use GEN8_MCR..() macros to calculate the |mcr| value for
1135 * Gen9 to address WaProgramMgsrForCorrectSliceSpecificMmioReads
1136 */
1137 mcr = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
1138 mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
1139
1140 drm_dbg(&i915->drm, "MCR slice:%d/subslice:%d = %x\n", slice, subslice, mcr);
1141
1142 wa_write_clr_set(wal, GEN8_MCR_SELECTOR, mcr_mask, mcr);
1143 }
1144
1145 static void
gen9_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1146 gen9_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1147 {
1148 struct drm_i915_private *i915 = gt->i915;
1149
1150 /* WaProgramMgsrForCorrectSliceSpecificMmioReads:glk,kbl,cml,gen9 */
1151 gen9_wa_init_mcr(i915, wal);
1152
1153 /* WaDisableKillLogic:bxt,skl,kbl */
1154 if (!IS_COFFEELAKE(i915) && !IS_COMETLAKE(i915))
1155 wa_write_or(wal,
1156 GAM_ECOCHK,
1157 ECOCHK_DIS_TLB);
1158
1159 if (HAS_LLC(i915)) {
1160 /* WaCompressedResourceSamplerPbeMediaNewHashMode:skl,kbl
1161 *
1162 * Must match Display Engine. See
1163 * WaCompressedResourceDisplayNewHashMode.
1164 */
1165 wa_write_or(wal,
1166 MMCD_MISC_CTRL,
1167 MMCD_PCLA | MMCD_HOTSPOT_EN);
1168 }
1169
1170 /* WaDisableHDCInvalidation:skl,bxt,kbl,cfl */
1171 wa_write_or(wal,
1172 GAM_ECOCHK,
1173 BDW_DISABLE_HDC_INVALIDATION);
1174 }
1175
1176 static void
skl_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1177 skl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1178 {
1179 gen9_gt_workarounds_init(gt, wal);
1180
1181 /* WaDisableGafsUnitClkGating:skl */
1182 wa_write_or(wal,
1183 GEN7_UCGCTL4,
1184 GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1185
1186 /* WaInPlaceDecompressionHang:skl */
1187 if (IS_SKYLAKE(gt->i915) && IS_GRAPHICS_STEP(gt->i915, STEP_A0, STEP_H0))
1188 wa_write_or(wal,
1189 GEN9_GAMT_ECO_REG_RW_IA,
1190 GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1191 }
1192
1193 static void
kbl_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1194 kbl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1195 {
1196 gen9_gt_workarounds_init(gt, wal);
1197
1198 /* WaDisableDynamicCreditSharing:kbl */
1199 if (IS_KABYLAKE(gt->i915) && IS_GRAPHICS_STEP(gt->i915, 0, STEP_C0))
1200 wa_write_or(wal,
1201 GAMT_CHKN_BIT_REG,
1202 GAMT_CHKN_DISABLE_DYNAMIC_CREDIT_SHARING);
1203
1204 /* WaDisableGafsUnitClkGating:kbl */
1205 wa_write_or(wal,
1206 GEN7_UCGCTL4,
1207 GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1208
1209 /* WaInPlaceDecompressionHang:kbl */
1210 wa_write_or(wal,
1211 GEN9_GAMT_ECO_REG_RW_IA,
1212 GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1213 }
1214
1215 static void
glk_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1216 glk_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1217 {
1218 gen9_gt_workarounds_init(gt, wal);
1219 }
1220
1221 static void
cfl_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1222 cfl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1223 {
1224 gen9_gt_workarounds_init(gt, wal);
1225
1226 /* WaDisableGafsUnitClkGating:cfl */
1227 wa_write_or(wal,
1228 GEN7_UCGCTL4,
1229 GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
1230
1231 /* WaInPlaceDecompressionHang:cfl */
1232 wa_write_or(wal,
1233 GEN9_GAMT_ECO_REG_RW_IA,
1234 GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
1235 }
1236
__set_mcr_steering(struct i915_wa_list * wal,i915_reg_t steering_reg,unsigned int slice,unsigned int subslice)1237 static void __set_mcr_steering(struct i915_wa_list *wal,
1238 i915_reg_t steering_reg,
1239 unsigned int slice, unsigned int subslice)
1240 {
1241 u32 mcr, mcr_mask;
1242
1243 mcr = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1244 mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1245
1246 wa_write_clr_set(wal, steering_reg, mcr_mask, mcr);
1247 }
1248
debug_dump_steering(struct intel_gt * gt)1249 static void debug_dump_steering(struct intel_gt *gt)
1250 {
1251 struct drm_printer p = drm_dbg_printer(>->i915->drm, DRM_UT_DRIVER,
1252 "MCR Steering:");
1253
1254 if (drm_debug_enabled(DRM_UT_DRIVER))
1255 intel_gt_mcr_report_steering(&p, gt, false);
1256 }
1257
__add_mcr_wa(struct intel_gt * gt,struct i915_wa_list * wal,unsigned int slice,unsigned int subslice)1258 static void __add_mcr_wa(struct intel_gt *gt, struct i915_wa_list *wal,
1259 unsigned int slice, unsigned int subslice)
1260 {
1261 __set_mcr_steering(wal, GEN8_MCR_SELECTOR, slice, subslice);
1262
1263 gt->default_steering.groupid = slice;
1264 gt->default_steering.instanceid = subslice;
1265
1266 debug_dump_steering(gt);
1267 }
1268
1269 static void
icl_wa_init_mcr(struct intel_gt * gt,struct i915_wa_list * wal)1270 icl_wa_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1271 {
1272 const struct sseu_dev_info *sseu = >->info.sseu;
1273 unsigned int subslice;
1274
1275 GEM_BUG_ON(GRAPHICS_VER(gt->i915) < 11);
1276 GEM_BUG_ON(hweight8(sseu->slice_mask) > 1);
1277
1278 /*
1279 * Although a platform may have subslices, we need to always steer
1280 * reads to the lowest instance that isn't fused off. When Render
1281 * Power Gating is enabled, grabbing forcewake will only power up a
1282 * single subslice (the "minconfig") if there isn't a real workload
1283 * that needs to be run; this means that if we steer register reads to
1284 * one of the higher subslices, we run the risk of reading back 0's or
1285 * random garbage.
1286 */
1287 subslice = __ffs(intel_sseu_get_hsw_subslices(sseu, 0));
1288
1289 /*
1290 * If the subslice we picked above also steers us to a valid L3 bank,
1291 * then we can just rely on the default steering and won't need to
1292 * worry about explicitly re-steering L3BANK reads later.
1293 */
1294 if (gt->info.l3bank_mask & BIT(subslice))
1295 gt->steering_table[L3BANK] = NULL;
1296
1297 __add_mcr_wa(gt, wal, 0, subslice);
1298 }
1299
1300 static void
xehp_init_mcr(struct intel_gt * gt,struct i915_wa_list * wal)1301 xehp_init_mcr(struct intel_gt *gt, struct i915_wa_list *wal)
1302 {
1303 const struct sseu_dev_info *sseu = >->info.sseu;
1304 unsigned long slice, subslice = 0, slice_mask = 0;
1305 u32 lncf_mask = 0;
1306 int i;
1307
1308 /*
1309 * On Xe_HP the steering increases in complexity. There are now several
1310 * more units that require steering and we're not guaranteed to be able
1311 * to find a common setting for all of them. These are:
1312 * - GSLICE (fusable)
1313 * - DSS (sub-unit within gslice; fusable)
1314 * - L3 Bank (fusable)
1315 * - MSLICE (fusable)
1316 * - LNCF (sub-unit within mslice; always present if mslice is present)
1317 *
1318 * We'll do our default/implicit steering based on GSLICE (in the
1319 * sliceid field) and DSS (in the subsliceid field). If we can
1320 * find overlap between the valid MSLICE and/or LNCF values with
1321 * a suitable GSLICE, then we can just re-use the default value and
1322 * skip and explicit steering at runtime.
1323 *
1324 * We only need to look for overlap between GSLICE/MSLICE/LNCF to find
1325 * a valid sliceid value. DSS steering is the only type of steering
1326 * that utilizes the 'subsliceid' bits.
1327 *
1328 * Also note that, even though the steering domain is called "GSlice"
1329 * and it is encoded in the register using the gslice format, the spec
1330 * says that the combined (geometry | compute) fuse should be used to
1331 * select the steering.
1332 */
1333
1334 /* Find the potential gslice candidates */
1335 slice_mask = intel_slicemask_from_xehp_dssmask(sseu->subslice_mask,
1336 GEN_DSS_PER_GSLICE);
1337
1338 /*
1339 * Find the potential LNCF candidates. Either LNCF within a valid
1340 * mslice is fine.
1341 */
1342 for_each_set_bit(i, >->info.mslice_mask, GEN12_MAX_MSLICES)
1343 lncf_mask |= (0x3 << (i * 2));
1344
1345 /*
1346 * Are there any sliceid values that work for both GSLICE and LNCF
1347 * steering?
1348 */
1349 if (slice_mask & lncf_mask) {
1350 slice_mask &= lncf_mask;
1351 gt->steering_table[LNCF] = NULL;
1352 }
1353
1354 /* How about sliceid values that also work for MSLICE steering? */
1355 if (slice_mask & gt->info.mslice_mask) {
1356 slice_mask &= gt->info.mslice_mask;
1357 gt->steering_table[MSLICE] = NULL;
1358 }
1359
1360 slice = __ffs(slice_mask);
1361 subslice = intel_sseu_find_first_xehp_dss(sseu, GEN_DSS_PER_GSLICE, slice) %
1362 GEN_DSS_PER_GSLICE;
1363
1364 __add_mcr_wa(gt, wal, slice, subslice);
1365
1366 /*
1367 * SQIDI ranges are special because they use different steering
1368 * registers than everything else we work with. On XeHP SDV and
1369 * DG2-G10, any value in the steering registers will work fine since
1370 * all instances are present, but DG2-G11 only has SQIDI instances at
1371 * ID's 2 and 3, so we need to steer to one of those. For simplicity
1372 * we'll just steer to a hardcoded "2" since that value will work
1373 * everywhere.
1374 */
1375 __set_mcr_steering(wal, MCFG_MCR_SELECTOR, 0, 2);
1376 __set_mcr_steering(wal, SF_MCR_SELECTOR, 0, 2);
1377
1378 /*
1379 * On DG2, GAM registers have a dedicated steering control register
1380 * and must always be programmed to a hardcoded groupid of "1."
1381 */
1382 if (IS_DG2(gt->i915))
1383 __set_mcr_steering(wal, GAM_MCR_SELECTOR, 1, 0);
1384 }
1385
1386 static void
icl_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1387 icl_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1388 {
1389 struct drm_i915_private *i915 = gt->i915;
1390
1391 icl_wa_init_mcr(gt, wal);
1392
1393 /* WaModifyGamTlbPartitioning:icl */
1394 wa_write_clr_set(wal,
1395 GEN11_GACB_PERF_CTRL,
1396 GEN11_HASH_CTRL_MASK,
1397 GEN11_HASH_CTRL_BIT0 | GEN11_HASH_CTRL_BIT4);
1398
1399 /* Wa_1405766107:icl
1400 * Formerly known as WaCL2SFHalfMaxAlloc
1401 */
1402 wa_write_or(wal,
1403 GEN11_LSN_UNSLCVC,
1404 GEN11_LSN_UNSLCVC_GAFS_HALF_SF_MAXALLOC |
1405 GEN11_LSN_UNSLCVC_GAFS_HALF_CL2_MAXALLOC);
1406
1407 /* Wa_220166154:icl
1408 * Formerly known as WaDisCtxReload
1409 */
1410 wa_write_or(wal,
1411 GEN8_GAMW_ECO_DEV_RW_IA,
1412 GAMW_ECO_DEV_CTX_RELOAD_DISABLE);
1413
1414 /* Wa_1406463099:icl
1415 * Formerly known as WaGamTlbPendError
1416 */
1417 wa_write_or(wal,
1418 GAMT_CHKN_BIT_REG,
1419 GAMT_CHKN_DISABLE_L3_COH_PIPE);
1420
1421 /*
1422 * Wa_1408615072:icl,ehl (vsunit)
1423 * Wa_1407596294:icl,ehl (hsunit)
1424 */
1425 wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1426 VSUNIT_CLKGATE_DIS | HSUNIT_CLKGATE_DIS);
1427
1428 /* Wa_1407352427:icl,ehl */
1429 wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2,
1430 PSDUNIT_CLKGATE_DIS);
1431
1432 /* Wa_1406680159:icl,ehl */
1433 wa_mcr_write_or(wal,
1434 GEN11_SUBSLICE_UNIT_LEVEL_CLKGATE,
1435 GWUNIT_CLKGATE_DIS);
1436
1437 /* Wa_1607087056:icl,ehl,jsl */
1438 if (IS_ICELAKE(i915) ||
1439 ((IS_JASPERLAKE(i915) || IS_ELKHARTLAKE(i915)) &&
1440 IS_GRAPHICS_STEP(i915, STEP_A0, STEP_B0)))
1441 wa_write_or(wal,
1442 GEN11_SLICE_UNIT_LEVEL_CLKGATE,
1443 L3_CLKGATE_DIS | L3_CR2X_CLKGATE_DIS);
1444
1445 /*
1446 * This is not a documented workaround, but rather an optimization
1447 * to reduce sampler power.
1448 */
1449 wa_mcr_write_clr(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1450 }
1451
1452 /*
1453 * Though there are per-engine instances of these registers,
1454 * they retain their value through engine resets and should
1455 * only be provided on the GT workaround list rather than
1456 * the engine-specific workaround list.
1457 */
1458 static void
wa_14011060649(struct intel_gt * gt,struct i915_wa_list * wal)1459 wa_14011060649(struct intel_gt *gt, struct i915_wa_list *wal)
1460 {
1461 struct intel_engine_cs *engine;
1462 int id;
1463
1464 for_each_engine(engine, gt, id) {
1465 if (engine->class != VIDEO_DECODE_CLASS ||
1466 (engine->instance % 2))
1467 continue;
1468
1469 wa_write_or(wal, VDBOX_CGCTL3F10(engine->mmio_base),
1470 IECPUNIT_CLKGATE_DIS);
1471 }
1472 }
1473
1474 static void
gen12_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1475 gen12_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1476 {
1477 icl_wa_init_mcr(gt, wal);
1478
1479 /* Wa_14011060649:tgl,rkl,dg1,adl-s,adl-p */
1480 wa_14011060649(gt, wal);
1481
1482 /* Wa_14011059788:tgl,rkl,adl-s,dg1,adl-p */
1483 wa_mcr_write_or(wal, GEN10_DFR_RATIO_EN_AND_CHICKEN, DFR_DISABLE);
1484
1485 /*
1486 * Wa_14015795083
1487 *
1488 * Firmware on some gen12 platforms locks the MISCCPCTL register,
1489 * preventing i915 from modifying it for this workaround. Skip the
1490 * readback verification for this workaround on debug builds; if the
1491 * workaround doesn't stick due to firmware behavior, it's not an error
1492 * that we want CI to flag.
1493 */
1494 wa_add(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE,
1495 0, 0, false);
1496 }
1497
1498 static void
dg1_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1499 dg1_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1500 {
1501 gen12_gt_workarounds_init(gt, wal);
1502
1503 /* Wa_1409420604:dg1 */
1504 wa_mcr_write_or(wal, SUBSLICE_UNIT_LEVEL_CLKGATE2,
1505 CPSSUNIT_CLKGATE_DIS);
1506
1507 /* Wa_1408615072:dg1 */
1508 /* Empirical testing shows this register is unaffected by engine reset. */
1509 wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE2, VSUNIT_CLKGATE_DIS_TGL);
1510 }
1511
1512 static void
dg2_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1513 dg2_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1514 {
1515 xehp_init_mcr(gt, wal);
1516
1517 /* Wa_14011060649:dg2 */
1518 wa_14011060649(gt, wal);
1519
1520 if (IS_DG2_G10(gt->i915)) {
1521 /* Wa_22010523718:dg2 */
1522 wa_write_or(wal, UNSLICE_UNIT_LEVEL_CLKGATE,
1523 CG3DDISCFEG_CLKGATE_DIS);
1524
1525 /* Wa_14011006942:dg2 */
1526 wa_mcr_write_or(wal, GEN11_SUBSLICE_UNIT_LEVEL_CLKGATE,
1527 DSS_ROUTER_CLKGATE_DIS);
1528 }
1529
1530 /* Wa_14014830051:dg2 */
1531 wa_mcr_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1532
1533 /*
1534 * Wa_14015795083
1535 * Skip verification for possibly locked register.
1536 */
1537 wa_add(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE,
1538 0, 0, false);
1539
1540 /* Wa_18018781329 */
1541 wa_mcr_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
1542 wa_mcr_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
1543 wa_mcr_write_or(wal, XEHP_VDBX_MOD_CTRL, FORCE_MISS_FTLB);
1544 wa_mcr_write_or(wal, XEHP_VEBX_MOD_CTRL, FORCE_MISS_FTLB);
1545
1546 /* Wa_1509235366:dg2 */
1547 wa_mcr_write_or(wal, XEHP_GAMCNTRL_CTRL,
1548 INVALIDATION_BROADCAST_MODE_DIS | GLOBAL_INVALIDATION_MODE);
1549
1550 /* Wa_14010648519:dg2 */
1551 wa_mcr_write_or(wal, XEHP_L3NODEARBCFG, XEHP_LNESPARE);
1552 }
1553
1554 static void
xelpg_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1555 xelpg_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1556 {
1557 /* Wa_14018575942 / Wa_18018781329 */
1558 wa_mcr_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
1559 wa_mcr_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
1560
1561 /* Wa_22016670082 */
1562 wa_write_or(wal, GEN12_SQCNT1, GEN12_STRICT_RAR_ENABLE);
1563
1564 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
1565 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
1566 /* Wa_14014830051 */
1567 wa_mcr_write_clr(wal, SARB_CHICKEN1, COMP_CKN_IN);
1568
1569 /* Wa_14015795083 */
1570 wa_write_clr(wal, GEN7_MISCCPCTL, GEN12_DOP_CLOCK_GATE_RENDER_ENABLE);
1571 }
1572
1573 /*
1574 * Unlike older platforms, we no longer setup implicit steering here;
1575 * all MCR accesses are explicitly steered.
1576 */
1577 debug_dump_steering(gt);
1578 }
1579
1580 static void
wa_16021867713(struct intel_gt * gt,struct i915_wa_list * wal)1581 wa_16021867713(struct intel_gt *gt, struct i915_wa_list *wal)
1582 {
1583 struct intel_engine_cs *engine;
1584 int id;
1585
1586 for_each_engine(engine, gt, id)
1587 if (engine->class == VIDEO_DECODE_CLASS)
1588 wa_write_or(wal, VDBOX_CGCTL3F1C(engine->mmio_base),
1589 MFXPIPE_CLKGATE_DIS);
1590 }
1591
1592 static void
xelpmp_gt_workarounds_init(struct intel_gt * gt,struct i915_wa_list * wal)1593 xelpmp_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
1594 {
1595 wa_16021867713(gt, wal);
1596
1597 /*
1598 * Wa_14018778641
1599 * Wa_18018781329
1600 *
1601 * Note that although these registers are MCR on the primary
1602 * GT, the media GT's versions are regular singleton registers.
1603 */
1604 wa_write_or(wal, XELPMP_GSC_MOD_CTRL, FORCE_MISS_FTLB);
1605
1606 /*
1607 * Wa_14018575942
1608 *
1609 * Issue is seen on media KPI test running on VDBOX engine
1610 * especially VP9 encoding WLs
1611 */
1612 wa_write_or(wal, XELPMP_VDBX_MOD_CTRL, FORCE_MISS_FTLB);
1613
1614 /* Wa_22016670082 */
1615 wa_write_or(wal, GEN12_SQCNT1, GEN12_STRICT_RAR_ENABLE);
1616
1617 debug_dump_steering(gt);
1618 }
1619
1620 /*
1621 * The bspec performance guide has recommended MMIO tuning settings. These
1622 * aren't truly "workarounds" but we want to program them through the
1623 * workaround infrastructure to make sure they're (re)applied at the proper
1624 * times.
1625 *
1626 * The programming in this function is for settings that persist through
1627 * engine resets and also are not part of any engine's register state context.
1628 * I.e., settings that only need to be re-applied in the event of a full GT
1629 * reset.
1630 */
gt_tuning_settings(struct intel_gt * gt,struct i915_wa_list * wal)1631 static void gt_tuning_settings(struct intel_gt *gt, struct i915_wa_list *wal)
1632 {
1633 if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74))) {
1634 wa_mcr_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
1635 wa_mcr_write_or(wal, XEHP_SQCM, EN_32B_ACCESS);
1636 }
1637
1638 if (IS_DG2(gt->i915)) {
1639 wa_mcr_write_or(wal, XEHP_L3SCQREG7, BLEND_FILL_CACHING_OPT_DIS);
1640 wa_mcr_write_or(wal, XEHP_SQCM, EN_32B_ACCESS);
1641 }
1642 }
1643
1644 static void
gt_init_workarounds(struct intel_gt * gt,struct i915_wa_list * wal)1645 gt_init_workarounds(struct intel_gt *gt, struct i915_wa_list *wal)
1646 {
1647 struct drm_i915_private *i915 = gt->i915;
1648
1649 gt_tuning_settings(gt, wal);
1650
1651 if (gt->type == GT_MEDIA) {
1652 if (MEDIA_VER_FULL(i915) == IP_VER(13, 0))
1653 xelpmp_gt_workarounds_init(gt, wal);
1654 else
1655 MISSING_CASE(MEDIA_VER_FULL(i915));
1656
1657 return;
1658 }
1659
1660 if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74)))
1661 xelpg_gt_workarounds_init(gt, wal);
1662 else if (IS_DG2(i915))
1663 dg2_gt_workarounds_init(gt, wal);
1664 else if (IS_DG1(i915))
1665 dg1_gt_workarounds_init(gt, wal);
1666 else if (GRAPHICS_VER(i915) == 12)
1667 gen12_gt_workarounds_init(gt, wal);
1668 else if (GRAPHICS_VER(i915) == 11)
1669 icl_gt_workarounds_init(gt, wal);
1670 else if (IS_COFFEELAKE(i915) || IS_COMETLAKE(i915))
1671 cfl_gt_workarounds_init(gt, wal);
1672 else if (IS_GEMINILAKE(i915))
1673 glk_gt_workarounds_init(gt, wal);
1674 else if (IS_KABYLAKE(i915))
1675 kbl_gt_workarounds_init(gt, wal);
1676 else if (IS_BROXTON(i915))
1677 gen9_gt_workarounds_init(gt, wal);
1678 else if (IS_SKYLAKE(i915))
1679 skl_gt_workarounds_init(gt, wal);
1680 else if (IS_HASWELL(i915))
1681 hsw_gt_workarounds_init(gt, wal);
1682 else if (IS_VALLEYVIEW(i915))
1683 vlv_gt_workarounds_init(gt, wal);
1684 else if (IS_IVYBRIDGE(i915))
1685 ivb_gt_workarounds_init(gt, wal);
1686 else if (GRAPHICS_VER(i915) == 6)
1687 snb_gt_workarounds_init(gt, wal);
1688 else if (GRAPHICS_VER(i915) == 5)
1689 ilk_gt_workarounds_init(gt, wal);
1690 else if (IS_G4X(i915))
1691 g4x_gt_workarounds_init(gt, wal);
1692 else if (GRAPHICS_VER(i915) == 4)
1693 gen4_gt_workarounds_init(gt, wal);
1694 else if (GRAPHICS_VER(i915) <= 8)
1695 ;
1696 else
1697 MISSING_CASE(GRAPHICS_VER(i915));
1698 }
1699
intel_gt_init_workarounds(struct intel_gt * gt)1700 void intel_gt_init_workarounds(struct intel_gt *gt)
1701 {
1702 struct i915_wa_list *wal = >->wa_list;
1703
1704 wa_init_start(wal, gt, "GT", "global");
1705 gt_init_workarounds(gt, wal);
1706 wa_init_finish(wal);
1707 }
1708
1709 static bool
wa_verify(struct intel_gt * gt,const struct i915_wa * wa,u32 cur,const char * name,const char * from)1710 wa_verify(struct intel_gt *gt, const struct i915_wa *wa, u32 cur,
1711 const char *name, const char *from)
1712 {
1713 if ((cur ^ wa->set) & wa->read) {
1714 gt_err(gt,
1715 "%s workaround lost on %s! (reg[%x]=0x%x, relevant bits were 0x%x vs expected 0x%x)\n",
1716 name, from, i915_mmio_reg_offset(wa->reg),
1717 cur, cur & wa->read, wa->set & wa->read);
1718
1719 return false;
1720 }
1721
1722 return true;
1723 }
1724
wa_list_apply(const struct i915_wa_list * wal)1725 static void wa_list_apply(const struct i915_wa_list *wal)
1726 {
1727 struct intel_gt *gt = wal->gt;
1728 struct intel_uncore *uncore = gt->uncore;
1729 enum forcewake_domains fw;
1730 unsigned long flags;
1731 struct i915_wa *wa;
1732 unsigned int i;
1733
1734 if (!wal->count)
1735 return;
1736
1737 fw = wal_get_fw_for_rmw(uncore, wal);
1738
1739 intel_gt_mcr_lock(gt, &flags);
1740 spin_lock(&uncore->lock);
1741 intel_uncore_forcewake_get__locked(uncore, fw);
1742
1743 for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
1744 u32 val, old = 0;
1745
1746 /* open-coded rmw due to steering */
1747 if (wa->clr)
1748 old = wa->is_mcr ?
1749 intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1750 intel_uncore_read_fw(uncore, wa->reg);
1751 val = (old & ~wa->clr) | wa->set;
1752 if (val != old || !wa->clr) {
1753 if (wa->is_mcr)
1754 intel_gt_mcr_multicast_write_fw(gt, wa->mcr_reg, val);
1755 else
1756 intel_uncore_write_fw(uncore, wa->reg, val);
1757 }
1758
1759 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) {
1760 u32 val = wa->is_mcr ?
1761 intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1762 intel_uncore_read_fw(uncore, wa->reg);
1763
1764 wa_verify(gt, wa, val, wal->name, "application");
1765 }
1766 }
1767
1768 intel_uncore_forcewake_put__locked(uncore, fw);
1769 spin_unlock(&uncore->lock);
1770 intel_gt_mcr_unlock(gt, flags);
1771 }
1772
intel_gt_apply_workarounds(struct intel_gt * gt)1773 void intel_gt_apply_workarounds(struct intel_gt *gt)
1774 {
1775 wa_list_apply(>->wa_list);
1776 }
1777
wa_list_verify(struct intel_gt * gt,const struct i915_wa_list * wal,const char * from)1778 static bool wa_list_verify(struct intel_gt *gt,
1779 const struct i915_wa_list *wal,
1780 const char *from)
1781 {
1782 struct intel_uncore *uncore = gt->uncore;
1783 struct i915_wa *wa;
1784 enum forcewake_domains fw;
1785 unsigned long flags;
1786 unsigned int i;
1787 bool ok = true;
1788
1789 fw = wal_get_fw_for_rmw(uncore, wal);
1790
1791 intel_gt_mcr_lock(gt, &flags);
1792 spin_lock(&uncore->lock);
1793 intel_uncore_forcewake_get__locked(uncore, fw);
1794
1795 for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
1796 ok &= wa_verify(wal->gt, wa, wa->is_mcr ?
1797 intel_gt_mcr_read_any_fw(gt, wa->mcr_reg) :
1798 intel_uncore_read_fw(uncore, wa->reg),
1799 wal->name, from);
1800
1801 intel_uncore_forcewake_put__locked(uncore, fw);
1802 spin_unlock(&uncore->lock);
1803 intel_gt_mcr_unlock(gt, flags);
1804
1805 return ok;
1806 }
1807
intel_gt_verify_workarounds(struct intel_gt * gt,const char * from)1808 bool intel_gt_verify_workarounds(struct intel_gt *gt, const char *from)
1809 {
1810 return wa_list_verify(gt, >->wa_list, from);
1811 }
1812
1813 __maybe_unused
is_nonpriv_flags_valid(u32 flags)1814 static bool is_nonpriv_flags_valid(u32 flags)
1815 {
1816 /* Check only valid flag bits are set */
1817 if (flags & ~RING_FORCE_TO_NONPRIV_MASK_VALID)
1818 return false;
1819
1820 /* NB: Only 3 out of 4 enum values are valid for access field */
1821 if ((flags & RING_FORCE_TO_NONPRIV_ACCESS_MASK) ==
1822 RING_FORCE_TO_NONPRIV_ACCESS_INVALID)
1823 return false;
1824
1825 return true;
1826 }
1827
1828 static void
whitelist_reg_ext(struct i915_wa_list * wal,i915_reg_t reg,u32 flags)1829 whitelist_reg_ext(struct i915_wa_list *wal, i915_reg_t reg, u32 flags)
1830 {
1831 struct i915_wa wa = {
1832 .reg = reg
1833 };
1834
1835 if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1836 return;
1837
1838 if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1839 return;
1840
1841 wa.reg.reg |= flags;
1842 _wa_add(wal, &wa);
1843 }
1844
1845 static void
whitelist_mcr_reg_ext(struct i915_wa_list * wal,i915_mcr_reg_t reg,u32 flags)1846 whitelist_mcr_reg_ext(struct i915_wa_list *wal, i915_mcr_reg_t reg, u32 flags)
1847 {
1848 struct i915_wa wa = {
1849 .mcr_reg = reg,
1850 .is_mcr = 1,
1851 };
1852
1853 if (GEM_DEBUG_WARN_ON(wal->count >= RING_MAX_NONPRIV_SLOTS))
1854 return;
1855
1856 if (GEM_DEBUG_WARN_ON(!is_nonpriv_flags_valid(flags)))
1857 return;
1858
1859 wa.mcr_reg.reg |= flags;
1860 _wa_add(wal, &wa);
1861 }
1862
1863 static void
whitelist_reg(struct i915_wa_list * wal,i915_reg_t reg)1864 whitelist_reg(struct i915_wa_list *wal, i915_reg_t reg)
1865 {
1866 whitelist_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1867 }
1868
1869 static void
whitelist_mcr_reg(struct i915_wa_list * wal,i915_mcr_reg_t reg)1870 whitelist_mcr_reg(struct i915_wa_list *wal, i915_mcr_reg_t reg)
1871 {
1872 whitelist_mcr_reg_ext(wal, reg, RING_FORCE_TO_NONPRIV_ACCESS_RW);
1873 }
1874
gen9_whitelist_build(struct i915_wa_list * w)1875 static void gen9_whitelist_build(struct i915_wa_list *w)
1876 {
1877 /* WaVFEStateAfterPipeControlwithMediaStateClear:skl,bxt,glk,cfl */
1878 whitelist_reg(w, GEN9_CTX_PREEMPT_REG);
1879
1880 /* WaEnablePreemptionGranularityControlByUMD:skl,bxt,kbl,cfl,[cnl] */
1881 whitelist_reg(w, GEN8_CS_CHICKEN1);
1882
1883 /* WaAllowUMDToModifyHDCChicken1:skl,bxt,kbl,glk,cfl */
1884 whitelist_reg(w, GEN8_HDC_CHICKEN1);
1885
1886 /* WaSendPushConstantsFromMMIO:skl,bxt */
1887 whitelist_reg(w, COMMON_SLICE_CHICKEN2);
1888 }
1889
skl_whitelist_build(struct intel_engine_cs * engine)1890 static void skl_whitelist_build(struct intel_engine_cs *engine)
1891 {
1892 struct i915_wa_list *w = &engine->whitelist;
1893
1894 if (engine->class != RENDER_CLASS)
1895 return;
1896
1897 gen9_whitelist_build(w);
1898
1899 /* WaDisableLSQCROPERFforOCL:skl */
1900 whitelist_mcr_reg(w, GEN8_L3SQCREG4);
1901 }
1902
bxt_whitelist_build(struct intel_engine_cs * engine)1903 static void bxt_whitelist_build(struct intel_engine_cs *engine)
1904 {
1905 if (engine->class != RENDER_CLASS)
1906 return;
1907
1908 gen9_whitelist_build(&engine->whitelist);
1909 }
1910
kbl_whitelist_build(struct intel_engine_cs * engine)1911 static void kbl_whitelist_build(struct intel_engine_cs *engine)
1912 {
1913 struct i915_wa_list *w = &engine->whitelist;
1914
1915 if (engine->class != RENDER_CLASS)
1916 return;
1917
1918 gen9_whitelist_build(w);
1919
1920 /* WaDisableLSQCROPERFforOCL:kbl */
1921 whitelist_mcr_reg(w, GEN8_L3SQCREG4);
1922 }
1923
glk_whitelist_build(struct intel_engine_cs * engine)1924 static void glk_whitelist_build(struct intel_engine_cs *engine)
1925 {
1926 struct i915_wa_list *w = &engine->whitelist;
1927
1928 if (engine->class != RENDER_CLASS)
1929 return;
1930
1931 gen9_whitelist_build(w);
1932
1933 /* WA #0862: Userspace has to set "Barrier Mode" to avoid hangs. */
1934 whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1935 }
1936
cfl_whitelist_build(struct intel_engine_cs * engine)1937 static void cfl_whitelist_build(struct intel_engine_cs *engine)
1938 {
1939 struct i915_wa_list *w = &engine->whitelist;
1940
1941 if (engine->class != RENDER_CLASS)
1942 return;
1943
1944 gen9_whitelist_build(w);
1945
1946 /*
1947 * WaAllowPMDepthAndInvocationCountAccessFromUMD:cfl,whl,cml,aml
1948 *
1949 * This covers 4 register which are next to one another :
1950 * - PS_INVOCATION_COUNT
1951 * - PS_INVOCATION_COUNT_UDW
1952 * - PS_DEPTH_COUNT
1953 * - PS_DEPTH_COUNT_UDW
1954 */
1955 whitelist_reg_ext(w, PS_INVOCATION_COUNT,
1956 RING_FORCE_TO_NONPRIV_ACCESS_RD |
1957 RING_FORCE_TO_NONPRIV_RANGE_4);
1958 }
1959
allow_read_ctx_timestamp(struct intel_engine_cs * engine)1960 static void allow_read_ctx_timestamp(struct intel_engine_cs *engine)
1961 {
1962 struct i915_wa_list *w = &engine->whitelist;
1963
1964 if (engine->class != RENDER_CLASS)
1965 whitelist_reg_ext(w,
1966 RING_CTX_TIMESTAMP(engine->mmio_base),
1967 RING_FORCE_TO_NONPRIV_ACCESS_RD);
1968 }
1969
cml_whitelist_build(struct intel_engine_cs * engine)1970 static void cml_whitelist_build(struct intel_engine_cs *engine)
1971 {
1972 allow_read_ctx_timestamp(engine);
1973
1974 cfl_whitelist_build(engine);
1975 }
1976
icl_whitelist_build(struct intel_engine_cs * engine)1977 static void icl_whitelist_build(struct intel_engine_cs *engine)
1978 {
1979 struct i915_wa_list *w = &engine->whitelist;
1980
1981 allow_read_ctx_timestamp(engine);
1982
1983 switch (engine->class) {
1984 case RENDER_CLASS:
1985 /* WaAllowUMDToModifyHalfSliceChicken7:icl */
1986 whitelist_mcr_reg(w, GEN9_HALF_SLICE_CHICKEN7);
1987
1988 /* WaAllowUMDToModifySamplerMode:icl */
1989 whitelist_mcr_reg(w, GEN10_SAMPLER_MODE);
1990
1991 /* WaEnableStateCacheRedirectToCS:icl */
1992 whitelist_reg(w, GEN9_SLICE_COMMON_ECO_CHICKEN1);
1993
1994 /*
1995 * WaAllowPMDepthAndInvocationCountAccessFromUMD:icl
1996 *
1997 * This covers 4 register which are next to one another :
1998 * - PS_INVOCATION_COUNT
1999 * - PS_INVOCATION_COUNT_UDW
2000 * - PS_DEPTH_COUNT
2001 * - PS_DEPTH_COUNT_UDW
2002 */
2003 whitelist_reg_ext(w, PS_INVOCATION_COUNT,
2004 RING_FORCE_TO_NONPRIV_ACCESS_RD |
2005 RING_FORCE_TO_NONPRIV_RANGE_4);
2006 break;
2007
2008 case VIDEO_DECODE_CLASS:
2009 /* hucStatusRegOffset */
2010 whitelist_reg_ext(w, _MMIO(0x2000 + engine->mmio_base),
2011 RING_FORCE_TO_NONPRIV_ACCESS_RD);
2012 /* hucUKernelHdrInfoRegOffset */
2013 whitelist_reg_ext(w, _MMIO(0x2014 + engine->mmio_base),
2014 RING_FORCE_TO_NONPRIV_ACCESS_RD);
2015 /* hucStatus2RegOffset */
2016 whitelist_reg_ext(w, _MMIO(0x23B0 + engine->mmio_base),
2017 RING_FORCE_TO_NONPRIV_ACCESS_RD);
2018 break;
2019
2020 default:
2021 break;
2022 }
2023 }
2024
tgl_whitelist_build(struct intel_engine_cs * engine)2025 static void tgl_whitelist_build(struct intel_engine_cs *engine)
2026 {
2027 struct i915_wa_list *w = &engine->whitelist;
2028
2029 allow_read_ctx_timestamp(engine);
2030
2031 switch (engine->class) {
2032 case RENDER_CLASS:
2033 /*
2034 * WaAllowPMDepthAndInvocationCountAccessFromUMD:tgl
2035 * Wa_1408556865:tgl
2036 *
2037 * This covers 4 registers which are next to one another :
2038 * - PS_INVOCATION_COUNT
2039 * - PS_INVOCATION_COUNT_UDW
2040 * - PS_DEPTH_COUNT
2041 * - PS_DEPTH_COUNT_UDW
2042 */
2043 whitelist_reg_ext(w, PS_INVOCATION_COUNT,
2044 RING_FORCE_TO_NONPRIV_ACCESS_RD |
2045 RING_FORCE_TO_NONPRIV_RANGE_4);
2046
2047 /*
2048 * Wa_1808121037:tgl
2049 * Wa_14012131227:dg1
2050 * Wa_1508744258:tgl,rkl,dg1,adl-s,adl-p
2051 */
2052 whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
2053
2054 /* Wa_1806527549:tgl */
2055 whitelist_reg(w, HIZ_CHICKEN);
2056
2057 /* Required by recommended tuning setting (not a workaround) */
2058 whitelist_reg(w, GEN11_COMMON_SLICE_CHICKEN3);
2059
2060 break;
2061 default:
2062 break;
2063 }
2064 }
2065
dg2_whitelist_build(struct intel_engine_cs * engine)2066 static void dg2_whitelist_build(struct intel_engine_cs *engine)
2067 {
2068 struct i915_wa_list *w = &engine->whitelist;
2069
2070 switch (engine->class) {
2071 case RENDER_CLASS:
2072 /* Required by recommended tuning setting (not a workaround) */
2073 whitelist_mcr_reg(w, XEHP_COMMON_SLICE_CHICKEN3);
2074 whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
2075 break;
2076 default:
2077 break;
2078 }
2079 }
2080
xelpg_whitelist_build(struct intel_engine_cs * engine)2081 static void xelpg_whitelist_build(struct intel_engine_cs *engine)
2082 {
2083 struct i915_wa_list *w = &engine->whitelist;
2084
2085 switch (engine->class) {
2086 case RENDER_CLASS:
2087 /* Required by recommended tuning setting (not a workaround) */
2088 whitelist_mcr_reg(w, XEHP_COMMON_SLICE_CHICKEN3);
2089 whitelist_reg(w, GEN7_COMMON_SLICE_CHICKEN1);
2090 break;
2091 default:
2092 break;
2093 }
2094 }
2095
intel_engine_init_whitelist(struct intel_engine_cs * engine)2096 void intel_engine_init_whitelist(struct intel_engine_cs *engine)
2097 {
2098 struct drm_i915_private *i915 = engine->i915;
2099 struct i915_wa_list *w = &engine->whitelist;
2100
2101 wa_init_start(w, engine->gt, "whitelist", engine->name);
2102
2103 if (engine->gt->type == GT_MEDIA)
2104 ; /* none yet */
2105 else if (IS_GFX_GT_IP_RANGE(engine->gt, IP_VER(12, 70), IP_VER(12, 74)))
2106 xelpg_whitelist_build(engine);
2107 else if (IS_DG2(i915))
2108 dg2_whitelist_build(engine);
2109 else if (GRAPHICS_VER(i915) == 12)
2110 tgl_whitelist_build(engine);
2111 else if (GRAPHICS_VER(i915) == 11)
2112 icl_whitelist_build(engine);
2113 else if (IS_COMETLAKE(i915))
2114 cml_whitelist_build(engine);
2115 else if (IS_COFFEELAKE(i915))
2116 cfl_whitelist_build(engine);
2117 else if (IS_GEMINILAKE(i915))
2118 glk_whitelist_build(engine);
2119 else if (IS_KABYLAKE(i915))
2120 kbl_whitelist_build(engine);
2121 else if (IS_BROXTON(i915))
2122 bxt_whitelist_build(engine);
2123 else if (IS_SKYLAKE(i915))
2124 skl_whitelist_build(engine);
2125 else if (GRAPHICS_VER(i915) <= 8)
2126 ;
2127 else
2128 MISSING_CASE(GRAPHICS_VER(i915));
2129
2130 wa_init_finish(w);
2131 }
2132
intel_engine_apply_whitelist(struct intel_engine_cs * engine)2133 void intel_engine_apply_whitelist(struct intel_engine_cs *engine)
2134 {
2135 const struct i915_wa_list *wal = &engine->whitelist;
2136 struct intel_uncore *uncore = engine->uncore;
2137 const u32 base = engine->mmio_base;
2138 struct i915_wa *wa;
2139 unsigned int i;
2140
2141 if (!wal->count)
2142 return;
2143
2144 for (i = 0, wa = wal->list; i < wal->count; i++, wa++)
2145 intel_uncore_write(uncore,
2146 RING_FORCE_TO_NONPRIV(base, i),
2147 i915_mmio_reg_offset(wa->reg));
2148
2149 /* And clear the rest just in case of garbage */
2150 for (; i < RING_MAX_NONPRIV_SLOTS; i++)
2151 intel_uncore_write(uncore,
2152 RING_FORCE_TO_NONPRIV(base, i),
2153 i915_mmio_reg_offset(RING_NOPID(base)));
2154 }
2155
2156 /*
2157 * engine_fake_wa_init(), a place holder to program the registers
2158 * which are not part of an official workaround defined by the
2159 * hardware team.
2160 * Adding programming of those register inside workaround will
2161 * allow utilizing wa framework to proper application and verification.
2162 */
2163 static void
engine_fake_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)2164 engine_fake_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2165 {
2166 u8 mocs_w, mocs_r;
2167
2168 /*
2169 * RING_CMD_CCTL specifies the default MOCS entry that will be used
2170 * by the command streamer when executing commands that don't have
2171 * a way to explicitly specify a MOCS setting. The default should
2172 * usually reference whichever MOCS entry corresponds to uncached
2173 * behavior, although use of a WB cached entry is recommended by the
2174 * spec in certain circumstances on specific platforms.
2175 */
2176 if (GRAPHICS_VER(engine->i915) >= 12) {
2177 mocs_r = engine->gt->mocs.uc_index;
2178 mocs_w = engine->gt->mocs.uc_index;
2179
2180 if (HAS_L3_CCS_READ(engine->i915) &&
2181 engine->class == COMPUTE_CLASS) {
2182 mocs_r = engine->gt->mocs.wb_index;
2183
2184 /*
2185 * Even on the few platforms where MOCS 0 is a
2186 * legitimate table entry, it's never the correct
2187 * setting to use here; we can assume the MOCS init
2188 * just forgot to initialize wb_index.
2189 */
2190 drm_WARN_ON(&engine->i915->drm, mocs_r == 0);
2191 }
2192
2193 wa_masked_field_set(wal,
2194 RING_CMD_CCTL(engine->mmio_base),
2195 CMD_CCTL_MOCS_MASK,
2196 CMD_CCTL_MOCS_OVERRIDE(mocs_w, mocs_r));
2197 }
2198 }
2199
2200 static void
rcs_engine_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)2201 rcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2202 {
2203 struct drm_i915_private *i915 = engine->i915;
2204 struct intel_gt *gt = engine->gt;
2205
2206 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2207 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0)) {
2208 /* Wa_22014600077 */
2209 wa_mcr_masked_en(wal, GEN10_CACHE_MODE_SS,
2210 ENABLE_EU_COUNT_FOR_TDL_FLUSH);
2211 }
2212
2213 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2214 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2215 IS_DG2(i915)) {
2216 /* Wa_1509727124 */
2217 wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
2218 SC_DISABLE_POWER_OPTIMIZATION_EBB);
2219 }
2220
2221 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2222 IS_DG2(i915)) {
2223 /* Wa_22012856258 */
2224 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
2225 GEN12_DISABLE_READ_SUPPRESSION);
2226 }
2227
2228 if (IS_DG2(i915)) {
2229 /*
2230 * Wa_22010960976:dg2
2231 * Wa_14013347512:dg2
2232 */
2233 wa_mcr_masked_dis(wal, XEHP_HDC_CHICKEN0,
2234 LSC_L1_FLUSH_CTL_3D_DATAPORT_FLUSH_EVENTS_MASK);
2235 }
2236
2237 if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 71)) ||
2238 IS_DG2(i915)) {
2239 /* Wa_14015150844 */
2240 wa_mcr_add(wal, XEHP_HDC_CHICKEN0, 0,
2241 _MASKED_BIT_ENABLE(DIS_ATOMIC_CHAINING_TYPED_WRITES),
2242 0, true);
2243 }
2244
2245 if (IS_DG2(i915) || IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2246 IS_DG1(i915) || IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2247 /*
2248 * Wa_1606700617:tgl,dg1,adl-p
2249 * Wa_22010271021:tgl,rkl,dg1,adl-s,adl-p
2250 * Wa_14010826681:tgl,dg1,rkl,adl-p
2251 * Wa_18019627453:dg2
2252 */
2253 wa_masked_en(wal,
2254 GEN9_CS_DEBUG_MODE1,
2255 FF_DOP_CLOCK_GATE_DISABLE);
2256 }
2257
2258 if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) || IS_DG1(i915) ||
2259 IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2260 /* Wa_1606931601:tgl,rkl,dg1,adl-s,adl-p */
2261 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2, GEN12_DISABLE_EARLY_READ);
2262
2263 /*
2264 * Wa_1407928979:tgl A*
2265 * Wa_18011464164:tgl[B0+],dg1[B0+]
2266 * Wa_22010931296:tgl[B0+],dg1[B0+]
2267 * Wa_14010919138:rkl,dg1,adl-s,adl-p
2268 */
2269 wa_write_or(wal, GEN7_FF_THREAD_MODE,
2270 GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2271
2272 /* Wa_1406941453:tgl,rkl,dg1,adl-s,adl-p */
2273 wa_mcr_masked_en(wal,
2274 GEN10_SAMPLER_MODE,
2275 ENABLE_SMALLPL);
2276 }
2277
2278 if (IS_ALDERLAKE_P(i915) || IS_ALDERLAKE_S(i915) ||
2279 IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915)) {
2280 /* Wa_1409804808 */
2281 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2,
2282 GEN12_PUSH_CONST_DEREF_HOLD_DIS);
2283
2284 /* Wa_14010229206 */
2285 wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, GEN12_DISABLE_TDL_PUSH);
2286 }
2287
2288 if (IS_ROCKETLAKE(i915) || IS_TIGERLAKE(i915) || IS_ALDERLAKE_P(i915)) {
2289 /*
2290 * Wa_1607297627
2291 *
2292 * On TGL and RKL there are multiple entries for this WA in the
2293 * BSpec; some indicate this is an A0-only WA, others indicate
2294 * it applies to all steppings so we trust the "all steppings."
2295 */
2296 wa_masked_en(wal,
2297 RING_PSMI_CTL(RENDER_RING_BASE),
2298 GEN12_WAIT_FOR_EVENT_POWER_DOWN_DISABLE |
2299 GEN8_RC_SEMA_IDLE_MSG_DISABLE);
2300 }
2301
2302 if (GRAPHICS_VER(i915) == 11) {
2303 /* This is not an Wa. Enable for better image quality */
2304 wa_masked_en(wal,
2305 _3D_CHICKEN3,
2306 _3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE);
2307
2308 /*
2309 * Wa_1405543622:icl
2310 * Formerly known as WaGAPZPriorityScheme
2311 */
2312 wa_write_or(wal,
2313 GEN8_GARBCNTL,
2314 GEN11_ARBITRATION_PRIO_ORDER_MASK);
2315
2316 /*
2317 * Wa_1604223664:icl
2318 * Formerly known as WaL3BankAddressHashing
2319 */
2320 wa_write_clr_set(wal,
2321 GEN8_GARBCNTL,
2322 GEN11_HASH_CTRL_EXCL_MASK,
2323 GEN11_HASH_CTRL_EXCL_BIT0);
2324 wa_write_clr_set(wal,
2325 GEN11_GLBLINVL,
2326 GEN11_BANK_HASH_ADDR_EXCL_MASK,
2327 GEN11_BANK_HASH_ADDR_EXCL_BIT0);
2328
2329 /*
2330 * Wa_1405733216:icl
2331 * Formerly known as WaDisableCleanEvicts
2332 */
2333 wa_mcr_write_or(wal,
2334 GEN8_L3SQCREG4,
2335 GEN11_LQSC_CLEAN_EVICT_DISABLE);
2336
2337 /* Wa_1606682166:icl */
2338 wa_write_or(wal,
2339 GEN7_SARCHKMD,
2340 GEN7_DISABLE_SAMPLER_PREFETCH);
2341
2342 /* Wa_1409178092:icl */
2343 wa_mcr_write_clr_set(wal,
2344 GEN11_SCRATCH2,
2345 GEN11_COHERENT_PARTIAL_WRITE_MERGE_ENABLE,
2346 0);
2347
2348 /* WaEnable32PlaneMode:icl */
2349 wa_masked_en(wal, GEN9_CSFE_CHICKEN1_RCS,
2350 GEN11_ENABLE_32_PLANE_MODE);
2351
2352 /*
2353 * Wa_1408767742:icl[a2..forever],ehl[all]
2354 * Wa_1605460711:icl[a0..c0]
2355 */
2356 wa_write_or(wal,
2357 GEN7_FF_THREAD_MODE,
2358 GEN12_FF_TESSELATION_DOP_GATE_DISABLE);
2359
2360 /* Wa_22010271021 */
2361 wa_masked_en(wal,
2362 GEN9_CS_DEBUG_MODE1,
2363 FF_DOP_CLOCK_GATE_DISABLE);
2364 }
2365
2366 /*
2367 * Intel platforms that support fine-grained preemption (i.e., gen9 and
2368 * beyond) allow the kernel-mode driver to choose between two different
2369 * options for controlling preemption granularity and behavior.
2370 *
2371 * Option 1 (hardware default):
2372 * Preemption settings are controlled in a global manner via
2373 * kernel-only register CS_DEBUG_MODE1 (0x20EC). Any granularity
2374 * and settings chosen by the kernel-mode driver will apply to all
2375 * userspace clients.
2376 *
2377 * Option 2:
2378 * Preemption settings are controlled on a per-context basis via
2379 * register CS_CHICKEN1 (0x2580). CS_CHICKEN1 is saved/restored on
2380 * context switch and is writable by userspace (e.g., via
2381 * MI_LOAD_REGISTER_IMMEDIATE instructions placed in a batch buffer)
2382 * which allows different userspace drivers/clients to select
2383 * different settings, or to change those settings on the fly in
2384 * response to runtime needs. This option was known by name
2385 * "FtrPerCtxtPreemptionGranularityControl" at one time, although
2386 * that name is somewhat misleading as other non-granularity
2387 * preemption settings are also impacted by this decision.
2388 *
2389 * On Linux, our policy has always been to let userspace drivers
2390 * control preemption granularity/settings (Option 2). This was
2391 * originally mandatory on gen9 to prevent ABI breakage (old gen9
2392 * userspace developed before object-level preemption was enabled would
2393 * not behave well if i915 were to go with Option 1 and enable that
2394 * preemption in a global manner). On gen9 each context would have
2395 * object-level preemption disabled by default (see
2396 * WaDisable3DMidCmdPreemption in gen9_ctx_workarounds_init), but
2397 * userspace drivers could opt-in to object-level preemption as they
2398 * saw fit. For post-gen9 platforms, we continue to utilize Option 2;
2399 * even though it is no longer necessary for ABI compatibility when
2400 * enabling a new platform, it does ensure that userspace will be able
2401 * to implement any workarounds that show up requiring temporary
2402 * adjustments to preemption behavior at runtime.
2403 *
2404 * Notes/Workarounds:
2405 * - Wa_14015141709: On DG2 and early steppings of MTL,
2406 * CS_CHICKEN1[0] does not disable object-level preemption as
2407 * it is supposed to (nor does CS_DEBUG_MODE1[0] if we had been
2408 * using Option 1). Effectively this means userspace is unable
2409 * to disable object-level preemption on these platforms/steppings
2410 * despite the setting here.
2411 *
2412 * - Wa_16013994831: May require that userspace program
2413 * CS_CHICKEN1[10] when certain runtime conditions are true.
2414 * Userspace requires Option 2 to be in effect for their update of
2415 * CS_CHICKEN1[10] to be effective.
2416 *
2417 * Other workarounds may appear in the future that will also require
2418 * Option 2 behavior to allow proper userspace implementation.
2419 */
2420 if (GRAPHICS_VER(i915) >= 9)
2421 wa_masked_en(wal,
2422 GEN7_FF_SLICE_CS_CHICKEN1,
2423 GEN9_FFSC_PERCTX_PREEMPT_CTRL);
2424
2425 if (IS_SKYLAKE(i915) ||
2426 IS_KABYLAKE(i915) ||
2427 IS_COFFEELAKE(i915) ||
2428 IS_COMETLAKE(i915)) {
2429 /* WaEnableGapsTsvCreditFix:skl,kbl,cfl */
2430 wa_write_or(wal,
2431 GEN8_GARBCNTL,
2432 GEN9_GAPS_TSV_CREDIT_DISABLE);
2433 }
2434
2435 if (IS_BROXTON(i915)) {
2436 /* WaDisablePooledEuLoadBalancingFix:bxt */
2437 wa_masked_en(wal,
2438 FF_SLICE_CS_CHICKEN2,
2439 GEN9_POOLED_EU_LOAD_BALANCING_FIX_DISABLE);
2440 }
2441
2442 if (GRAPHICS_VER(i915) == 9) {
2443 /* WaContextSwitchWithConcurrentTLBInvalidate:skl,bxt,kbl,glk,cfl */
2444 wa_masked_en(wal,
2445 GEN9_CSFE_CHICKEN1_RCS,
2446 GEN9_PREEMPT_GPGPU_SYNC_SWITCH_DISABLE);
2447
2448 /* WaEnableLbsSlaRetryTimerDecrement:skl,bxt,kbl,glk,cfl */
2449 wa_mcr_write_or(wal,
2450 BDW_SCRATCH1,
2451 GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
2452
2453 /* WaProgramL3SqcReg1DefaultForPerf:bxt,glk */
2454 if (IS_GEN9_LP(i915))
2455 wa_mcr_write_clr_set(wal,
2456 GEN8_L3SQCREG1,
2457 L3_PRIO_CREDITS_MASK,
2458 L3_GENERAL_PRIO_CREDITS(62) |
2459 L3_HIGH_PRIO_CREDITS(2));
2460
2461 /* WaOCLCoherentLineFlush:skl,bxt,kbl,cfl */
2462 wa_mcr_write_or(wal,
2463 GEN8_L3SQCREG4,
2464 GEN8_LQSC_FLUSH_COHERENT_LINES);
2465
2466 /* Disable atomics in L3 to prevent unrecoverable hangs */
2467 wa_write_clr_set(wal, GEN9_SCRATCH_LNCF1,
2468 GEN9_LNCF_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2469 wa_mcr_write_clr_set(wal, GEN8_L3SQCREG4,
2470 GEN8_LQSQ_NONIA_COHERENT_ATOMICS_ENABLE, 0);
2471 wa_mcr_write_clr_set(wal, GEN9_SCRATCH1,
2472 EVICTION_PERF_FIX_ENABLE, 0);
2473 }
2474
2475 if (IS_HASWELL(i915)) {
2476 /* WaSampleCChickenBitEnable:hsw */
2477 wa_masked_en(wal,
2478 HSW_HALF_SLICE_CHICKEN3, HSW_SAMPLE_C_PERFORMANCE);
2479
2480 wa_masked_dis(wal,
2481 CACHE_MODE_0_GEN7,
2482 /* enable HiZ Raw Stall Optimization */
2483 HIZ_RAW_STALL_OPT_DISABLE);
2484 }
2485
2486 if (IS_VALLEYVIEW(i915)) {
2487 /* WaDisableEarlyCull:vlv */
2488 wa_masked_en(wal,
2489 _3D_CHICKEN3,
2490 _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2491
2492 /*
2493 * WaVSThreadDispatchOverride:ivb,vlv
2494 *
2495 * This actually overrides the dispatch
2496 * mode for all thread types.
2497 */
2498 wa_write_clr_set(wal,
2499 GEN7_FF_THREAD_MODE,
2500 GEN7_FF_SCHED_MASK,
2501 GEN7_FF_TS_SCHED_HW |
2502 GEN7_FF_VS_SCHED_HW |
2503 GEN7_FF_DS_SCHED_HW);
2504
2505 /* WaPsdDispatchEnable:vlv */
2506 /* WaDisablePSDDualDispatchEnable:vlv */
2507 wa_masked_en(wal,
2508 GEN7_HALF_SLICE_CHICKEN1,
2509 GEN7_MAX_PS_THREAD_DEP |
2510 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2511 }
2512
2513 if (IS_IVYBRIDGE(i915)) {
2514 /* WaDisableEarlyCull:ivb */
2515 wa_masked_en(wal,
2516 _3D_CHICKEN3,
2517 _3D_CHICKEN_SF_DISABLE_OBJEND_CULL);
2518
2519 if (0) { /* causes HiZ corruption on ivb:gt1 */
2520 /* enable HiZ Raw Stall Optimization */
2521 wa_masked_dis(wal,
2522 CACHE_MODE_0_GEN7,
2523 HIZ_RAW_STALL_OPT_DISABLE);
2524 }
2525
2526 /*
2527 * WaVSThreadDispatchOverride:ivb,vlv
2528 *
2529 * This actually overrides the dispatch
2530 * mode for all thread types.
2531 */
2532 wa_write_clr_set(wal,
2533 GEN7_FF_THREAD_MODE,
2534 GEN7_FF_SCHED_MASK,
2535 GEN7_FF_TS_SCHED_HW |
2536 GEN7_FF_VS_SCHED_HW |
2537 GEN7_FF_DS_SCHED_HW);
2538
2539 /* WaDisablePSDDualDispatchEnable:ivb */
2540 if (IS_IVB_GT1(i915))
2541 wa_masked_en(wal,
2542 GEN7_HALF_SLICE_CHICKEN1,
2543 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE);
2544 }
2545
2546 if (GRAPHICS_VER(i915) == 7) {
2547 /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
2548 wa_masked_en(wal,
2549 RING_MODE_GEN7(RENDER_RING_BASE),
2550 GFX_TLB_INVALIDATE_EXPLICIT | GFX_REPLAY_MODE);
2551
2552 /* WaDisable_RenderCache_OperationalFlush:ivb,vlv,hsw */
2553 wa_masked_dis(wal, CACHE_MODE_0_GEN7, RC_OP_FLUSH_ENABLE);
2554
2555 /*
2556 * BSpec says this must be set, even though
2557 * WaDisable4x2SubspanOptimization:ivb,hsw
2558 * WaDisable4x2SubspanOptimization isn't listed for VLV.
2559 */
2560 wa_masked_en(wal,
2561 CACHE_MODE_1,
2562 PIXEL_SUBSPAN_COLLECT_OPT_DISABLE);
2563
2564 /*
2565 * BSpec recommends 8x4 when MSAA is used,
2566 * however in practice 16x4 seems fastest.
2567 *
2568 * Note that PS/WM thread counts depend on the WIZ hashing
2569 * disable bit, which we don't touch here, but it's good
2570 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2571 */
2572 wa_masked_field_set(wal,
2573 GEN7_GT_MODE,
2574 GEN6_WIZ_HASHING_MASK,
2575 GEN6_WIZ_HASHING_16x4);
2576 }
2577
2578 if (IS_GRAPHICS_VER(i915, 6, 7))
2579 /*
2580 * We need to disable the AsyncFlip performance optimisations in
2581 * order to use MI_WAIT_FOR_EVENT within the CS. It should
2582 * already be programmed to '1' on all products.
2583 *
2584 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
2585 */
2586 wa_masked_en(wal,
2587 RING_MI_MODE(RENDER_RING_BASE),
2588 ASYNC_FLIP_PERF_DISABLE);
2589
2590 if (GRAPHICS_VER(i915) == 6) {
2591 /*
2592 * Required for the hardware to program scanline values for
2593 * waiting
2594 * WaEnableFlushTlbInvalidationMode:snb
2595 */
2596 wa_masked_en(wal,
2597 GFX_MODE,
2598 GFX_TLB_INVALIDATE_EXPLICIT);
2599
2600 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
2601 wa_masked_en(wal,
2602 _3D_CHICKEN,
2603 _3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB);
2604
2605 wa_masked_en(wal,
2606 _3D_CHICKEN3,
2607 /* WaStripsFansDisableFastClipPerformanceFix:snb */
2608 _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL |
2609 /*
2610 * Bspec says:
2611 * "This bit must be set if 3DSTATE_CLIP clip mode is set
2612 * to normal and 3DSTATE_SF number of SF output attributes
2613 * is more than 16."
2614 */
2615 _3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH);
2616
2617 /*
2618 * BSpec recommends 8x4 when MSAA is used,
2619 * however in practice 16x4 seems fastest.
2620 *
2621 * Note that PS/WM thread counts depend on the WIZ hashing
2622 * disable bit, which we don't touch here, but it's good
2623 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
2624 */
2625 wa_masked_field_set(wal,
2626 GEN6_GT_MODE,
2627 GEN6_WIZ_HASHING_MASK,
2628 GEN6_WIZ_HASHING_16x4);
2629
2630 /* WaDisable_RenderCache_OperationalFlush:snb */
2631 wa_masked_dis(wal, CACHE_MODE_0, RC_OP_FLUSH_ENABLE);
2632
2633 /*
2634 * From the Sandybridge PRM, volume 1 part 3, page 24:
2635 * "If this bit is set, STCunit will have LRA as replacement
2636 * policy. [...] This bit must be reset. LRA replacement
2637 * policy is not supported."
2638 */
2639 wa_masked_dis(wal,
2640 CACHE_MODE_0,
2641 CM0_STC_EVICT_DISABLE_LRA_SNB);
2642 }
2643
2644 if (IS_GRAPHICS_VER(i915, 4, 6))
2645 /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
2646 wa_add(wal, RING_MI_MODE(RENDER_RING_BASE),
2647 0, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH),
2648 /* XXX bit doesn't stick on Broadwater */
2649 IS_I965G(i915) ? 0 : VS_TIMER_DISPATCH, true);
2650
2651 if (GRAPHICS_VER(i915) == 4)
2652 /*
2653 * Disable CONSTANT_BUFFER before it is loaded from the context
2654 * image. For as it is loaded, it is executed and the stored
2655 * address may no longer be valid, leading to a GPU hang.
2656 *
2657 * This imposes the requirement that userspace reload their
2658 * CONSTANT_BUFFER on every batch, fortunately a requirement
2659 * they are already accustomed to from before contexts were
2660 * enabled.
2661 */
2662 wa_add(wal, ECOSKPD(RENDER_RING_BASE),
2663 0, _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE),
2664 0 /* XXX bit doesn't stick on Broadwater */,
2665 true);
2666 }
2667
2668 static void
xcs_engine_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)2669 xcs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2670 {
2671 struct drm_i915_private *i915 = engine->i915;
2672
2673 /* WaKBLVECSSemaphoreWaitPoll:kbl */
2674 if (IS_KABYLAKE(i915) && IS_GRAPHICS_STEP(i915, STEP_A0, STEP_F0)) {
2675 wa_write(wal,
2676 RING_SEMA_WAIT_POLL(engine->mmio_base),
2677 1);
2678 }
2679 /* Wa_16018031267, Wa_16018063123 */
2680 if (NEEDS_FASTCOLOR_BLT_WABB(engine))
2681 wa_masked_field_set(wal, ECOSKPD(engine->mmio_base),
2682 XEHP_BLITTER_SCHEDULING_MODE_MASK,
2683 XEHP_BLITTER_ROUND_ROBIN_MODE);
2684 }
2685
2686 static void
ccs_engine_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)2687 ccs_engine_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2688 {
2689 /* boilerplate for any CCS engine workaround */
2690 }
2691
2692 /*
2693 * The bspec performance guide has recommended MMIO tuning settings. These
2694 * aren't truly "workarounds" but we want to program them with the same
2695 * workaround infrastructure to ensure that they're automatically added to
2696 * the GuC save/restore lists, re-applied at the right times, and checked for
2697 * any conflicting programming requested by real workarounds.
2698 *
2699 * Programming settings should be added here only if their registers are not
2700 * part of an engine's register state context. If a register is part of a
2701 * context, then any tuning settings should be programmed in an appropriate
2702 * function invoked by __intel_engine_init_ctx_wa().
2703 */
2704 static void
add_render_compute_tuning_settings(struct intel_gt * gt,struct i915_wa_list * wal)2705 add_render_compute_tuning_settings(struct intel_gt *gt,
2706 struct i915_wa_list *wal)
2707 {
2708 struct drm_i915_private *i915 = gt->i915;
2709
2710 if (IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 70), IP_VER(12, 74)) || IS_DG2(i915))
2711 wa_mcr_write_clr_set(wal, RT_CTRL, STACKID_CTRL, STACKID_CTRL_512);
2712
2713 /*
2714 * This tuning setting proves beneficial only on ATS-M designs; the
2715 * default "age based" setting is optimal on regular DG2 and other
2716 * platforms.
2717 */
2718 if (INTEL_INFO(i915)->tuning_thread_rr_after_dep)
2719 wa_mcr_masked_field_set(wal, GEN9_ROW_CHICKEN4, THREAD_EX_ARB_MODE,
2720 THREAD_EX_ARB_MODE_RR_AFTER_DEP);
2721
2722 if (GRAPHICS_VER(i915) == 12 && GRAPHICS_VER_FULL(i915) < IP_VER(12, 55))
2723 wa_write_clr(wal, GEN8_GARBCNTL, GEN12_BUS_HASH_CTL_BIT_EXC);
2724 }
2725
ccs_engine_wa_mode(struct intel_engine_cs * engine,struct i915_wa_list * wal)2726 static void ccs_engine_wa_mode(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2727 {
2728 struct intel_gt *gt = engine->gt;
2729 u32 mode;
2730
2731 if (!IS_DG2(gt->i915))
2732 return;
2733
2734 /*
2735 * Wa_14019159160: This workaround, along with others, leads to
2736 * significant challenges in utilizing load balancing among the
2737 * CCS slices. Consequently, an architectural decision has been
2738 * made to completely disable automatic CCS load balancing.
2739 */
2740 wa_masked_en(wal, GEN12_RCU_MODE, XEHP_RCU_MODE_FIXED_SLICE_CCS_MODE);
2741
2742 /*
2743 * After having disabled automatic load balancing we need to
2744 * assign all slices to a single CCS. We will call it CCS mode 1
2745 */
2746 mode = intel_gt_apply_ccs_mode(gt);
2747 wa_masked_en(wal, XEHP_CCS_MODE, mode);
2748 }
2749
2750 /*
2751 * The workarounds in this function apply to shared registers in
2752 * the general render reset domain that aren't tied to a
2753 * specific engine. Since all render+compute engines get reset
2754 * together, and the contents of these registers are lost during
2755 * the shared render domain reset, we'll define such workarounds
2756 * here and then add them to just a single RCS or CCS engine's
2757 * workaround list (whichever engine has the XXXX flag).
2758 */
2759 static void
general_render_compute_wa_init(struct intel_engine_cs * engine,struct i915_wa_list * wal)2760 general_render_compute_wa_init(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2761 {
2762 struct drm_i915_private *i915 = engine->i915;
2763 struct intel_gt *gt = engine->gt;
2764
2765 add_render_compute_tuning_settings(gt, wal);
2766
2767 if (GRAPHICS_VER(i915) >= 11) {
2768 /* This is not a Wa (although referred to as
2769 * WaSetInidrectStateOverride in places), this allows
2770 * applications that reference sampler states through
2771 * the BindlessSamplerStateBaseAddress to have their
2772 * border color relative to DynamicStateBaseAddress
2773 * rather than BindlessSamplerStateBaseAddress.
2774 *
2775 * Otherwise SAMPLER_STATE border colors have to be
2776 * copied in multiple heaps (DynamicStateBaseAddress &
2777 * BindlessSamplerStateBaseAddress)
2778 *
2779 * BSpec: 46052
2780 */
2781 wa_mcr_masked_en(wal,
2782 GEN10_SAMPLER_MODE,
2783 GEN11_INDIRECT_STATE_BASE_ADDR_OVERRIDE);
2784 }
2785
2786 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_B0, STEP_FOREVER) ||
2787 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_B0, STEP_FOREVER) ||
2788 IS_GFX_GT_IP_RANGE(gt, IP_VER(12, 74), IP_VER(12, 74))) {
2789 /* Wa_14017856879 */
2790 wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN3, MTL_DISABLE_FIX_FOR_EOT_FLUSH);
2791
2792 /* Wa_14020495402 */
2793 wa_mcr_masked_en(wal, GEN8_ROW_CHICKEN2, XELPG_DISABLE_TDL_SVHS_GATING);
2794 }
2795
2796 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2797 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0))
2798 /*
2799 * Wa_14017066071
2800 * Wa_14017654203
2801 */
2802 wa_mcr_masked_en(wal, GEN10_SAMPLER_MODE,
2803 MTL_DISABLE_SAMPLER_SC_OOO);
2804
2805 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0))
2806 /* Wa_22015279794 */
2807 wa_mcr_masked_en(wal, GEN10_CACHE_MODE_SS,
2808 DISABLE_PREFETCH_INTO_IC);
2809
2810 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2811 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2812 IS_DG2(i915)) {
2813 /* Wa_22013037850 */
2814 wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW,
2815 DISABLE_128B_EVICTION_COMMAND_UDW);
2816
2817 /* Wa_18017747507 */
2818 wa_masked_en(wal, VFG_PREEMPTION_CHICKEN, POLYGON_TRIFAN_LINELOOP_DISABLE);
2819 }
2820
2821 if (IS_GFX_GT_IP_STEP(gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
2822 IS_GFX_GT_IP_STEP(gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
2823 IS_DG2(i915)) {
2824 /* Wa_22014226127 */
2825 wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0, DISABLE_D8_D16_COASLESCE);
2826 }
2827
2828 if (IS_DG2(i915)) {
2829 /* Wa_14015227452:dg2,pvc */
2830 wa_mcr_masked_en(wal, GEN9_ROW_CHICKEN4, XEHP_DIS_BBL_SYSPIPE);
2831
2832 /*
2833 * Wa_16011620976:dg2_g11
2834 * Wa_22015475538:dg2
2835 */
2836 wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, DIS_CHAIN_2XSIMD8);
2837
2838 /* Wa_18028616096 */
2839 wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0_UDW, UGM_FRAGMENT_THRESHOLD_TO_3);
2840 }
2841
2842 if (IS_DG2_G11(i915)) {
2843 /*
2844 * Wa_22012826095:dg2
2845 * Wa_22013059131:dg2
2846 */
2847 wa_mcr_write_clr_set(wal, LSC_CHICKEN_BIT_0_UDW,
2848 MAXREQS_PER_BANK,
2849 REG_FIELD_PREP(MAXREQS_PER_BANK, 2));
2850
2851 /* Wa_22013059131:dg2 */
2852 wa_mcr_write_or(wal, LSC_CHICKEN_BIT_0,
2853 FORCE_1_SUB_MESSAGE_PER_FRAGMENT);
2854
2855 /*
2856 * Wa_22012654132
2857 *
2858 * Note that register 0xE420 is write-only and cannot be read
2859 * back for verification on DG2 (due to Wa_14012342262), so
2860 * we need to explicitly skip the readback.
2861 */
2862 wa_mcr_add(wal, GEN10_CACHE_MODE_SS, 0,
2863 _MASKED_BIT_ENABLE(ENABLE_PREFETCH_INTO_IC),
2864 0 /* write-only, so skip validation */,
2865 true);
2866 }
2867 }
2868
2869 static void
engine_init_workarounds(struct intel_engine_cs * engine,struct i915_wa_list * wal)2870 engine_init_workarounds(struct intel_engine_cs *engine, struct i915_wa_list *wal)
2871 {
2872 if (GRAPHICS_VER(engine->i915) < 4)
2873 return;
2874
2875 engine_fake_wa_init(engine, wal);
2876
2877 /*
2878 * These are common workarounds that just need to applied
2879 * to a single RCS/CCS engine's workaround list since
2880 * they're reset as part of the general render domain reset.
2881 */
2882 if (engine->flags & I915_ENGINE_FIRST_RENDER_COMPUTE) {
2883 general_render_compute_wa_init(engine, wal);
2884 ccs_engine_wa_mode(engine, wal);
2885 }
2886
2887 if (engine->class == COMPUTE_CLASS)
2888 ccs_engine_wa_init(engine, wal);
2889 else if (engine->class == RENDER_CLASS)
2890 rcs_engine_wa_init(engine, wal);
2891 else
2892 xcs_engine_wa_init(engine, wal);
2893 }
2894
intel_engine_init_workarounds(struct intel_engine_cs * engine)2895 void intel_engine_init_workarounds(struct intel_engine_cs *engine)
2896 {
2897 struct i915_wa_list *wal = &engine->wa_list;
2898
2899 wa_init_start(wal, engine->gt, "engine", engine->name);
2900 engine_init_workarounds(engine, wal);
2901 wa_init_finish(wal);
2902 }
2903
intel_engine_apply_workarounds(struct intel_engine_cs * engine)2904 void intel_engine_apply_workarounds(struct intel_engine_cs *engine)
2905 {
2906 wa_list_apply(&engine->wa_list);
2907 }
2908
2909 static const struct i915_range mcr_ranges_gen8[] = {
2910 { .start = 0x5500, .end = 0x55ff },
2911 { .start = 0x7000, .end = 0x7fff },
2912 { .start = 0x9400, .end = 0x97ff },
2913 { .start = 0xb000, .end = 0xb3ff },
2914 { .start = 0xe000, .end = 0xe7ff },
2915 {},
2916 };
2917
2918 static const struct i915_range mcr_ranges_gen12[] = {
2919 { .start = 0x8150, .end = 0x815f },
2920 { .start = 0x9520, .end = 0x955f },
2921 { .start = 0xb100, .end = 0xb3ff },
2922 { .start = 0xde80, .end = 0xe8ff },
2923 { .start = 0x24a00, .end = 0x24a7f },
2924 {},
2925 };
2926
2927 static const struct i915_range mcr_ranges_xehp[] = {
2928 { .start = 0x4000, .end = 0x4aff },
2929 { .start = 0x5200, .end = 0x52ff },
2930 { .start = 0x5400, .end = 0x7fff },
2931 { .start = 0x8140, .end = 0x815f },
2932 { .start = 0x8c80, .end = 0x8dff },
2933 { .start = 0x94d0, .end = 0x955f },
2934 { .start = 0x9680, .end = 0x96ff },
2935 { .start = 0xb000, .end = 0xb3ff },
2936 { .start = 0xc800, .end = 0xcfff },
2937 { .start = 0xd800, .end = 0xd8ff },
2938 { .start = 0xdc00, .end = 0xffff },
2939 { .start = 0x17000, .end = 0x17fff },
2940 { .start = 0x24a00, .end = 0x24a7f },
2941 {},
2942 };
2943
mcr_range(struct drm_i915_private * i915,u32 offset)2944 static bool mcr_range(struct drm_i915_private *i915, u32 offset)
2945 {
2946 const struct i915_range *mcr_ranges;
2947 int i;
2948
2949 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55))
2950 mcr_ranges = mcr_ranges_xehp;
2951 else if (GRAPHICS_VER(i915) >= 12)
2952 mcr_ranges = mcr_ranges_gen12;
2953 else if (GRAPHICS_VER(i915) >= 8)
2954 mcr_ranges = mcr_ranges_gen8;
2955 else
2956 return false;
2957
2958 /*
2959 * Registers in these ranges are affected by the MCR selector
2960 * which only controls CPU initiated MMIO. Routing does not
2961 * work for CS access so we cannot verify them on this path.
2962 */
2963 for (i = 0; mcr_ranges[i].start; i++)
2964 if (offset >= mcr_ranges[i].start &&
2965 offset <= mcr_ranges[i].end)
2966 return true;
2967
2968 return false;
2969 }
2970
2971 static int
wa_list_srm(struct i915_request * rq,const struct i915_wa_list * wal,struct i915_vma * vma)2972 wa_list_srm(struct i915_request *rq,
2973 const struct i915_wa_list *wal,
2974 struct i915_vma *vma)
2975 {
2976 struct drm_i915_private *i915 = rq->i915;
2977 unsigned int i, count = 0;
2978 const struct i915_wa *wa;
2979 u32 srm, *cs;
2980
2981 srm = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
2982 if (GRAPHICS_VER(i915) >= 8)
2983 srm++;
2984
2985 for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2986 if (!mcr_range(i915, i915_mmio_reg_offset(wa->reg)))
2987 count++;
2988 }
2989
2990 cs = intel_ring_begin(rq, 4 * count);
2991 if (IS_ERR(cs))
2992 return PTR_ERR(cs);
2993
2994 for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
2995 u32 offset = i915_mmio_reg_offset(wa->reg);
2996
2997 if (mcr_range(i915, offset))
2998 continue;
2999
3000 *cs++ = srm;
3001 *cs++ = offset;
3002 *cs++ = i915_ggtt_offset(vma) + sizeof(u32) * i;
3003 *cs++ = 0;
3004 }
3005 intel_ring_advance(rq, cs);
3006
3007 return 0;
3008 }
3009
engine_wa_list_verify(struct intel_context * ce,const struct i915_wa_list * const wal,const char * from)3010 static int engine_wa_list_verify(struct intel_context *ce,
3011 const struct i915_wa_list * const wal,
3012 const char *from)
3013 {
3014 const struct i915_wa *wa;
3015 struct i915_request *rq;
3016 struct i915_vma *vma;
3017 struct i915_gem_ww_ctx ww;
3018 unsigned int i;
3019 u32 *results;
3020 int err;
3021
3022 if (!wal->count)
3023 return 0;
3024
3025 vma = __vm_create_scratch_for_read(&ce->engine->gt->ggtt->vm,
3026 wal->count * sizeof(u32));
3027 if (IS_ERR(vma))
3028 return PTR_ERR(vma);
3029
3030 intel_engine_pm_get(ce->engine);
3031 i915_gem_ww_ctx_init(&ww, false);
3032 retry:
3033 err = i915_gem_object_lock(vma->obj, &ww);
3034 if (err == 0)
3035 err = intel_context_pin_ww(ce, &ww);
3036 if (err)
3037 goto err_pm;
3038
3039 err = i915_vma_pin_ww(vma, &ww, 0, 0,
3040 i915_vma_is_ggtt(vma) ? PIN_GLOBAL : PIN_USER);
3041 if (err)
3042 goto err_unpin;
3043
3044 rq = i915_request_create(ce);
3045 if (IS_ERR(rq)) {
3046 err = PTR_ERR(rq);
3047 goto err_vma;
3048 }
3049
3050 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
3051 if (err == 0)
3052 err = wa_list_srm(rq, wal, vma);
3053
3054 i915_request_get(rq);
3055 if (err)
3056 i915_request_set_error_once(rq, err);
3057 i915_request_add(rq);
3058
3059 if (err)
3060 goto err_rq;
3061
3062 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
3063 err = -ETIME;
3064 goto err_rq;
3065 }
3066
3067 results = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
3068 if (IS_ERR(results)) {
3069 err = PTR_ERR(results);
3070 goto err_rq;
3071 }
3072
3073 err = 0;
3074 for (i = 0, wa = wal->list; i < wal->count; i++, wa++) {
3075 if (mcr_range(rq->i915, i915_mmio_reg_offset(wa->reg)))
3076 continue;
3077
3078 if (!wa_verify(wal->gt, wa, results[i], wal->name, from))
3079 err = -ENXIO;
3080 }
3081
3082 i915_gem_object_unpin_map(vma->obj);
3083
3084 err_rq:
3085 i915_request_put(rq);
3086 err_vma:
3087 i915_vma_unpin(vma);
3088 err_unpin:
3089 intel_context_unpin(ce);
3090 err_pm:
3091 if (err == -EDEADLK) {
3092 err = i915_gem_ww_ctx_backoff(&ww);
3093 if (!err)
3094 goto retry;
3095 }
3096 i915_gem_ww_ctx_fini(&ww);
3097 intel_engine_pm_put(ce->engine);
3098 i915_vma_put(vma);
3099 return err;
3100 }
3101
intel_engine_verify_workarounds(struct intel_engine_cs * engine,const char * from)3102 int intel_engine_verify_workarounds(struct intel_engine_cs *engine,
3103 const char *from)
3104 {
3105 return engine_wa_list_verify(engine->kernel_context,
3106 &engine->wa_list,
3107 from);
3108 }
3109
3110 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
3111 #include "selftest_workarounds.c"
3112 #endif
3113