1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2019 Intel Corporation
4 */
5
6 #include "debugfs_gt.h"
7
8 #include "gem/i915_gem_lmem.h"
9 #include "i915_drv.h"
10 #include "intel_context.h"
11 #include "intel_gt.h"
12 #include "intel_gt_buffer_pool.h"
13 #include "intel_gt_clock_utils.h"
14 #include "intel_gt_pm.h"
15 #include "intel_gt_requests.h"
16 #include "intel_migrate.h"
17 #include "intel_mocs.h"
18 #include "intel_rc6.h"
19 #include "intel_renderstate.h"
20 #include "intel_rps.h"
21 #include "intel_uncore.h"
22 #include "intel_pm.h"
23 #include "shmem_utils.h"
24
intel_gt_init_early(struct intel_gt * gt,struct drm_i915_private * i915)25 void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915)
26 {
27 gt->i915 = i915;
28 gt->uncore = &i915->uncore;
29
30 spin_lock_init(>->irq_lock);
31
32 INIT_LIST_HEAD(>->closed_vma);
33 spin_lock_init(>->closed_lock);
34
35 init_llist_head(>->watchdog.list);
36 INIT_WORK(>->watchdog.work, intel_gt_watchdog_work);
37
38 intel_gt_init_buffer_pool(gt);
39 intel_gt_init_reset(gt);
40 intel_gt_init_requests(gt);
41 intel_gt_init_timelines(gt);
42 intel_gt_pm_init_early(gt);
43
44 intel_uc_init_early(>->uc);
45 intel_rps_init_early(>->rps);
46 }
47
intel_gt_probe_lmem(struct intel_gt * gt)48 int intel_gt_probe_lmem(struct intel_gt *gt)
49 {
50 struct drm_i915_private *i915 = gt->i915;
51 struct intel_memory_region *mem;
52 int id;
53 int err;
54
55 mem = intel_gt_setup_lmem(gt);
56 if (mem == ERR_PTR(-ENODEV))
57 mem = intel_gt_setup_fake_lmem(gt);
58 if (IS_ERR(mem)) {
59 err = PTR_ERR(mem);
60 if (err == -ENODEV)
61 return 0;
62
63 drm_err(&i915->drm,
64 "Failed to setup region(%d) type=%d\n",
65 err, INTEL_MEMORY_LOCAL);
66 return err;
67 }
68
69 id = INTEL_REGION_LMEM;
70
71 mem->id = id;
72
73 intel_memory_region_set_name(mem, "local%u", mem->instance);
74
75 GEM_BUG_ON(!HAS_REGION(i915, id));
76 GEM_BUG_ON(i915->mm.regions[id]);
77 i915->mm.regions[id] = mem;
78
79 return 0;
80 }
81
intel_gt_init_hw_early(struct intel_gt * gt,struct i915_ggtt * ggtt)82 void intel_gt_init_hw_early(struct intel_gt *gt, struct i915_ggtt *ggtt)
83 {
84 gt->ggtt = ggtt;
85 }
86
87 static const struct intel_mmio_range icl_l3bank_steering_table[] = {
88 { 0x00B100, 0x00B3FF },
89 {},
90 };
91
92 static const struct intel_mmio_range xehpsdv_mslice_steering_table[] = {
93 { 0x004000, 0x004AFF },
94 { 0x00C800, 0x00CFFF },
95 { 0x00DD00, 0x00DDFF },
96 { 0x00E900, 0x00FFFF }, /* 0xEA00 - OxEFFF is unused */
97 {},
98 };
99
100 static const struct intel_mmio_range xehpsdv_lncf_steering_table[] = {
101 { 0x00B000, 0x00B0FF },
102 { 0x00D800, 0x00D8FF },
103 {},
104 };
105
106 static const struct intel_mmio_range dg2_lncf_steering_table[] = {
107 { 0x00B000, 0x00B0FF },
108 { 0x00D880, 0x00D8FF },
109 {},
110 };
111
slicemask(struct intel_gt * gt,int count)112 static u16 slicemask(struct intel_gt *gt, int count)
113 {
114 u64 dss_mask = intel_sseu_get_subslices(>->info.sseu, 0);
115
116 return intel_slicemask_from_dssmask(dss_mask, count);
117 }
118
intel_gt_init_mmio(struct intel_gt * gt)119 int intel_gt_init_mmio(struct intel_gt *gt)
120 {
121 struct drm_i915_private *i915 = gt->i915;
122
123 intel_gt_init_clock_frequency(gt);
124
125 intel_uc_init_mmio(>->uc);
126 intel_sseu_info_init(gt);
127
128 /*
129 * An mslice is unavailable only if both the meml3 for the slice is
130 * disabled *and* all of the DSS in the slice (quadrant) are disabled.
131 */
132 if (HAS_MSLICES(i915))
133 gt->info.mslice_mask =
134 slicemask(gt, GEN_DSS_PER_MSLICE) |
135 (intel_uncore_read(gt->uncore, GEN10_MIRROR_FUSE3) &
136 GEN12_MEML3_EN_MASK);
137
138 if (IS_DG2(i915)) {
139 gt->steering_table[MSLICE] = xehpsdv_mslice_steering_table;
140 gt->steering_table[LNCF] = dg2_lncf_steering_table;
141 } else if (IS_XEHPSDV(i915)) {
142 gt->steering_table[MSLICE] = xehpsdv_mslice_steering_table;
143 gt->steering_table[LNCF] = xehpsdv_lncf_steering_table;
144 } else if (GRAPHICS_VER(i915) >= 11 &&
145 GRAPHICS_VER_FULL(i915) < IP_VER(12, 50)) {
146 gt->steering_table[L3BANK] = icl_l3bank_steering_table;
147 gt->info.l3bank_mask =
148 ~intel_uncore_read(gt->uncore, GEN10_MIRROR_FUSE3) &
149 GEN10_L3BANK_MASK;
150 } else if (HAS_MSLICES(i915)) {
151 MISSING_CASE(INTEL_INFO(i915)->platform);
152 }
153
154 return intel_engines_init_mmio(gt);
155 }
156
init_unused_ring(struct intel_gt * gt,u32 base)157 static void init_unused_ring(struct intel_gt *gt, u32 base)
158 {
159 struct intel_uncore *uncore = gt->uncore;
160
161 intel_uncore_write(uncore, RING_CTL(base), 0);
162 intel_uncore_write(uncore, RING_HEAD(base), 0);
163 intel_uncore_write(uncore, RING_TAIL(base), 0);
164 intel_uncore_write(uncore, RING_START(base), 0);
165 }
166
init_unused_rings(struct intel_gt * gt)167 static void init_unused_rings(struct intel_gt *gt)
168 {
169 struct drm_i915_private *i915 = gt->i915;
170
171 if (IS_I830(i915)) {
172 init_unused_ring(gt, PRB1_BASE);
173 init_unused_ring(gt, SRB0_BASE);
174 init_unused_ring(gt, SRB1_BASE);
175 init_unused_ring(gt, SRB2_BASE);
176 init_unused_ring(gt, SRB3_BASE);
177 } else if (GRAPHICS_VER(i915) == 2) {
178 init_unused_ring(gt, SRB0_BASE);
179 init_unused_ring(gt, SRB1_BASE);
180 } else if (GRAPHICS_VER(i915) == 3) {
181 init_unused_ring(gt, PRB1_BASE);
182 init_unused_ring(gt, PRB2_BASE);
183 }
184 }
185
intel_gt_init_hw(struct intel_gt * gt)186 int intel_gt_init_hw(struct intel_gt *gt)
187 {
188 struct drm_i915_private *i915 = gt->i915;
189 struct intel_uncore *uncore = gt->uncore;
190 int ret;
191
192 gt->last_init_time = ktime_get();
193
194 /* Double layer security blanket, see i915_gem_init() */
195 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
196
197 if (HAS_EDRAM(i915) && GRAPHICS_VER(i915) < 9)
198 intel_uncore_rmw(uncore, HSW_IDICR, 0, IDIHASHMSK(0xf));
199
200 if (IS_HASWELL(i915))
201 intel_uncore_write(uncore,
202 MI_PREDICATE_RESULT_2,
203 IS_HSW_GT3(i915) ?
204 LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
205
206 /* Apply the GT workarounds... */
207 intel_gt_apply_workarounds(gt);
208 /* ...and determine whether they are sticking. */
209 intel_gt_verify_workarounds(gt, "init");
210
211 intel_gt_init_swizzling(gt);
212
213 /*
214 * At least 830 can leave some of the unused rings
215 * "active" (ie. head != tail) after resume which
216 * will prevent c3 entry. Makes sure all unused rings
217 * are totally idle.
218 */
219 init_unused_rings(gt);
220
221 ret = i915_ppgtt_init_hw(gt);
222 if (ret) {
223 DRM_ERROR("Enabling PPGTT failed (%d)\n", ret);
224 goto out;
225 }
226
227 /* We can't enable contexts until all firmware is loaded */
228 ret = intel_uc_init_hw(>->uc);
229 if (ret) {
230 i915_probe_error(i915, "Enabling uc failed (%d)\n", ret);
231 goto out;
232 }
233
234 intel_mocs_init(gt);
235
236 out:
237 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
238 return ret;
239 }
240
rmw_set(struct intel_uncore * uncore,i915_reg_t reg,u32 set)241 static void rmw_set(struct intel_uncore *uncore, i915_reg_t reg, u32 set)
242 {
243 intel_uncore_rmw(uncore, reg, 0, set);
244 }
245
rmw_clear(struct intel_uncore * uncore,i915_reg_t reg,u32 clr)246 static void rmw_clear(struct intel_uncore *uncore, i915_reg_t reg, u32 clr)
247 {
248 intel_uncore_rmw(uncore, reg, clr, 0);
249 }
250
clear_register(struct intel_uncore * uncore,i915_reg_t reg)251 static void clear_register(struct intel_uncore *uncore, i915_reg_t reg)
252 {
253 intel_uncore_rmw(uncore, reg, 0, 0);
254 }
255
gen6_clear_engine_error_register(struct intel_engine_cs * engine)256 static void gen6_clear_engine_error_register(struct intel_engine_cs *engine)
257 {
258 GEN6_RING_FAULT_REG_RMW(engine, RING_FAULT_VALID, 0);
259 GEN6_RING_FAULT_REG_POSTING_READ(engine);
260 }
261
262 void
intel_gt_clear_error_registers(struct intel_gt * gt,intel_engine_mask_t engine_mask)263 intel_gt_clear_error_registers(struct intel_gt *gt,
264 intel_engine_mask_t engine_mask)
265 {
266 struct drm_i915_private *i915 = gt->i915;
267 struct intel_uncore *uncore = gt->uncore;
268 u32 eir;
269
270 if (GRAPHICS_VER(i915) != 2)
271 clear_register(uncore, PGTBL_ER);
272
273 if (GRAPHICS_VER(i915) < 4)
274 clear_register(uncore, IPEIR(RENDER_RING_BASE));
275 else
276 clear_register(uncore, IPEIR_I965);
277
278 clear_register(uncore, EIR);
279 eir = intel_uncore_read(uncore, EIR);
280 if (eir) {
281 /*
282 * some errors might have become stuck,
283 * mask them.
284 */
285 DRM_DEBUG_DRIVER("EIR stuck: 0x%08x, masking\n", eir);
286 rmw_set(uncore, EMR, eir);
287 intel_uncore_write(uncore, GEN2_IIR,
288 I915_MASTER_ERROR_INTERRUPT);
289 }
290
291 if (GRAPHICS_VER(i915) >= 12) {
292 rmw_clear(uncore, GEN12_RING_FAULT_REG, RING_FAULT_VALID);
293 intel_uncore_posting_read(uncore, GEN12_RING_FAULT_REG);
294 } else if (GRAPHICS_VER(i915) >= 8) {
295 rmw_clear(uncore, GEN8_RING_FAULT_REG, RING_FAULT_VALID);
296 intel_uncore_posting_read(uncore, GEN8_RING_FAULT_REG);
297 } else if (GRAPHICS_VER(i915) >= 6) {
298 struct intel_engine_cs *engine;
299 enum intel_engine_id id;
300
301 for_each_engine_masked(engine, gt, engine_mask, id)
302 gen6_clear_engine_error_register(engine);
303 }
304 }
305
gen6_check_faults(struct intel_gt * gt)306 static void gen6_check_faults(struct intel_gt *gt)
307 {
308 struct intel_engine_cs *engine;
309 enum intel_engine_id id;
310 u32 fault;
311
312 for_each_engine(engine, gt, id) {
313 fault = GEN6_RING_FAULT_REG_READ(engine);
314 if (fault & RING_FAULT_VALID) {
315 drm_dbg(&engine->i915->drm, "Unexpected fault\n"
316 "\tAddr: 0x%08lx\n"
317 "\tAddress space: %s\n"
318 "\tSource ID: %d\n"
319 "\tType: %d\n",
320 fault & PAGE_MASK,
321 fault & RING_FAULT_GTTSEL_MASK ?
322 "GGTT" : "PPGTT",
323 RING_FAULT_SRCID(fault),
324 RING_FAULT_FAULT_TYPE(fault));
325 }
326 }
327 }
328
gen8_check_faults(struct intel_gt * gt)329 static void gen8_check_faults(struct intel_gt *gt)
330 {
331 struct intel_uncore *uncore = gt->uncore;
332 i915_reg_t fault_reg, fault_data0_reg, fault_data1_reg;
333 u32 fault;
334
335 if (GRAPHICS_VER(gt->i915) >= 12) {
336 fault_reg = GEN12_RING_FAULT_REG;
337 fault_data0_reg = GEN12_FAULT_TLB_DATA0;
338 fault_data1_reg = GEN12_FAULT_TLB_DATA1;
339 } else {
340 fault_reg = GEN8_RING_FAULT_REG;
341 fault_data0_reg = GEN8_FAULT_TLB_DATA0;
342 fault_data1_reg = GEN8_FAULT_TLB_DATA1;
343 }
344
345 fault = intel_uncore_read(uncore, fault_reg);
346 if (fault & RING_FAULT_VALID) {
347 u32 fault_data0, fault_data1;
348 u64 fault_addr;
349
350 fault_data0 = intel_uncore_read(uncore, fault_data0_reg);
351 fault_data1 = intel_uncore_read(uncore, fault_data1_reg);
352
353 fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) |
354 ((u64)fault_data0 << 12);
355
356 drm_dbg(&uncore->i915->drm, "Unexpected fault\n"
357 "\tAddr: 0x%08x_%08x\n"
358 "\tAddress space: %s\n"
359 "\tEngine ID: %d\n"
360 "\tSource ID: %d\n"
361 "\tType: %d\n",
362 upper_32_bits(fault_addr), lower_32_bits(fault_addr),
363 fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT",
364 GEN8_RING_FAULT_ENGINE_ID(fault),
365 RING_FAULT_SRCID(fault),
366 RING_FAULT_FAULT_TYPE(fault));
367 }
368 }
369
intel_gt_check_and_clear_faults(struct intel_gt * gt)370 void intel_gt_check_and_clear_faults(struct intel_gt *gt)
371 {
372 struct drm_i915_private *i915 = gt->i915;
373
374 /* From GEN8 onwards we only have one 'All Engine Fault Register' */
375 if (GRAPHICS_VER(i915) >= 8)
376 gen8_check_faults(gt);
377 else if (GRAPHICS_VER(i915) >= 6)
378 gen6_check_faults(gt);
379 else
380 return;
381
382 intel_gt_clear_error_registers(gt, ALL_ENGINES);
383 }
384
intel_gt_flush_ggtt_writes(struct intel_gt * gt)385 void intel_gt_flush_ggtt_writes(struct intel_gt *gt)
386 {
387 struct intel_uncore *uncore = gt->uncore;
388 intel_wakeref_t wakeref;
389
390 /*
391 * No actual flushing is required for the GTT write domain for reads
392 * from the GTT domain. Writes to it "immediately" go to main memory
393 * as far as we know, so there's no chipset flush. It also doesn't
394 * land in the GPU render cache.
395 *
396 * However, we do have to enforce the order so that all writes through
397 * the GTT land before any writes to the device, such as updates to
398 * the GATT itself.
399 *
400 * We also have to wait a bit for the writes to land from the GTT.
401 * An uncached read (i.e. mmio) seems to be ideal for the round-trip
402 * timing. This issue has only been observed when switching quickly
403 * between GTT writes and CPU reads from inside the kernel on recent hw,
404 * and it appears to only affect discrete GTT blocks (i.e. on LLC
405 * system agents we cannot reproduce this behaviour, until Cannonlake
406 * that was!).
407 */
408
409 wmb();
410
411 if (INTEL_INFO(gt->i915)->has_coherent_ggtt)
412 return;
413
414 intel_gt_chipset_flush(gt);
415
416 with_intel_runtime_pm_if_in_use(uncore->rpm, wakeref) {
417 unsigned long flags;
418
419 spin_lock_irqsave(&uncore->lock, flags);
420 intel_uncore_posting_read_fw(uncore,
421 RING_HEAD(RENDER_RING_BASE));
422 spin_unlock_irqrestore(&uncore->lock, flags);
423 }
424 }
425
intel_gt_chipset_flush(struct intel_gt * gt)426 void intel_gt_chipset_flush(struct intel_gt *gt)
427 {
428 wmb();
429 if (GRAPHICS_VER(gt->i915) < 6)
430 intel_gtt_chipset_flush();
431 }
432
intel_gt_driver_register(struct intel_gt * gt)433 void intel_gt_driver_register(struct intel_gt *gt)
434 {
435 intel_rps_driver_register(>->rps);
436
437 debugfs_gt_register(gt);
438 }
439
intel_gt_init_scratch(struct intel_gt * gt,unsigned int size)440 static int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size)
441 {
442 struct drm_i915_private *i915 = gt->i915;
443 struct drm_i915_gem_object *obj;
444 struct i915_vma *vma;
445 int ret;
446
447 obj = i915_gem_object_create_lmem(i915, size, I915_BO_ALLOC_VOLATILE);
448 if (IS_ERR(obj))
449 obj = i915_gem_object_create_stolen(i915, size);
450 if (IS_ERR(obj))
451 obj = i915_gem_object_create_internal(i915, size);
452 if (IS_ERR(obj)) {
453 drm_err(&i915->drm, "Failed to allocate scratch page\n");
454 return PTR_ERR(obj);
455 }
456
457 vma = i915_vma_instance(obj, >->ggtt->vm, NULL);
458 if (IS_ERR(vma)) {
459 ret = PTR_ERR(vma);
460 goto err_unref;
461 }
462
463 ret = i915_ggtt_pin(vma, NULL, 0, PIN_HIGH);
464 if (ret)
465 goto err_unref;
466
467 gt->scratch = i915_vma_make_unshrinkable(vma);
468
469 return 0;
470
471 err_unref:
472 i915_gem_object_put(obj);
473 return ret;
474 }
475
intel_gt_fini_scratch(struct intel_gt * gt)476 static void intel_gt_fini_scratch(struct intel_gt *gt)
477 {
478 i915_vma_unpin_and_release(>->scratch, 0);
479 }
480
kernel_vm(struct intel_gt * gt)481 static struct i915_address_space *kernel_vm(struct intel_gt *gt)
482 {
483 if (INTEL_PPGTT(gt->i915) > INTEL_PPGTT_ALIASING)
484 return &i915_ppgtt_create(gt)->vm;
485 else
486 return i915_vm_get(>->ggtt->vm);
487 }
488
__engines_record_defaults(struct intel_gt * gt)489 static int __engines_record_defaults(struct intel_gt *gt)
490 {
491 struct i915_request *requests[I915_NUM_ENGINES] = {};
492 struct intel_engine_cs *engine;
493 enum intel_engine_id id;
494 int err = 0;
495
496 /*
497 * As we reset the gpu during very early sanitisation, the current
498 * register state on the GPU should reflect its defaults values.
499 * We load a context onto the hw (with restore-inhibit), then switch
500 * over to a second context to save that default register state. We
501 * can then prime every new context with that state so they all start
502 * from the same default HW values.
503 */
504
505 for_each_engine(engine, gt, id) {
506 struct intel_renderstate so;
507 struct intel_context *ce;
508 struct i915_request *rq;
509
510 /* We must be able to switch to something! */
511 GEM_BUG_ON(!engine->kernel_context);
512
513 ce = intel_context_create(engine);
514 if (IS_ERR(ce)) {
515 err = PTR_ERR(ce);
516 goto out;
517 }
518
519 err = intel_renderstate_init(&so, ce);
520 if (err)
521 goto err;
522
523 rq = i915_request_create(ce);
524 if (IS_ERR(rq)) {
525 err = PTR_ERR(rq);
526 goto err_fini;
527 }
528
529 err = intel_engine_emit_ctx_wa(rq);
530 if (err)
531 goto err_rq;
532
533 err = intel_renderstate_emit(&so, rq);
534 if (err)
535 goto err_rq;
536
537 err_rq:
538 requests[id] = i915_request_get(rq);
539 i915_request_add(rq);
540 err_fini:
541 intel_renderstate_fini(&so, ce);
542 err:
543 if (err) {
544 intel_context_put(ce);
545 goto out;
546 }
547 }
548
549 /* Flush the default context image to memory, and enable powersaving. */
550 if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME) {
551 err = -EIO;
552 goto out;
553 }
554
555 for (id = 0; id < ARRAY_SIZE(requests); id++) {
556 struct i915_request *rq;
557 struct file *state;
558
559 rq = requests[id];
560 if (!rq)
561 continue;
562
563 if (rq->fence.error) {
564 err = -EIO;
565 goto out;
566 }
567
568 GEM_BUG_ON(!test_bit(CONTEXT_ALLOC_BIT, &rq->context->flags));
569 if (!rq->context->state)
570 continue;
571
572 /* Keep a copy of the state's backing pages; free the obj */
573 state = shmem_create_from_object(rq->context->state->obj);
574 if (IS_ERR(state)) {
575 err = PTR_ERR(state);
576 goto out;
577 }
578 rq->engine->default_state = state;
579 }
580
581 out:
582 /*
583 * If we have to abandon now, we expect the engines to be idle
584 * and ready to be torn-down. The quickest way we can accomplish
585 * this is by declaring ourselves wedged.
586 */
587 if (err)
588 intel_gt_set_wedged(gt);
589
590 for (id = 0; id < ARRAY_SIZE(requests); id++) {
591 struct intel_context *ce;
592 struct i915_request *rq;
593
594 rq = requests[id];
595 if (!rq)
596 continue;
597
598 ce = rq->context;
599 i915_request_put(rq);
600 intel_context_put(ce);
601 }
602 return err;
603 }
604
__engines_verify_workarounds(struct intel_gt * gt)605 static int __engines_verify_workarounds(struct intel_gt *gt)
606 {
607 struct intel_engine_cs *engine;
608 enum intel_engine_id id;
609 int err = 0;
610
611 if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
612 return 0;
613
614 for_each_engine(engine, gt, id) {
615 if (intel_engine_verify_workarounds(engine, "load"))
616 err = -EIO;
617 }
618
619 /* Flush and restore the kernel context for safety */
620 if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME)
621 err = -EIO;
622
623 return err;
624 }
625
__intel_gt_disable(struct intel_gt * gt)626 static void __intel_gt_disable(struct intel_gt *gt)
627 {
628 intel_gt_set_wedged_on_fini(gt);
629
630 intel_gt_suspend_prepare(gt);
631 intel_gt_suspend_late(gt);
632
633 GEM_BUG_ON(intel_gt_pm_is_awake(gt));
634 }
635
intel_gt_wait_for_idle(struct intel_gt * gt,long timeout)636 int intel_gt_wait_for_idle(struct intel_gt *gt, long timeout)
637 {
638 long remaining_timeout;
639
640 /* If the device is asleep, we have no requests outstanding */
641 if (!intel_gt_pm_is_awake(gt))
642 return 0;
643
644 while ((timeout = intel_gt_retire_requests_timeout(gt, timeout,
645 &remaining_timeout)) > 0) {
646 cond_resched();
647 if (signal_pending(current))
648 return -EINTR;
649 }
650
651 return timeout ? timeout : intel_uc_wait_for_idle(>->uc,
652 remaining_timeout);
653 }
654
intel_gt_init(struct intel_gt * gt)655 int intel_gt_init(struct intel_gt *gt)
656 {
657 int err;
658
659 err = i915_inject_probe_error(gt->i915, -ENODEV);
660 if (err)
661 return err;
662
663 /*
664 * This is just a security blanket to placate dragons.
665 * On some systems, we very sporadically observe that the first TLBs
666 * used by the CS may be stale, despite us poking the TLB reset. If
667 * we hold the forcewake during initialisation these problems
668 * just magically go away.
669 */
670 intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL);
671
672 err = intel_gt_init_scratch(gt,
673 GRAPHICS_VER(gt->i915) == 2 ? SZ_256K : SZ_4K);
674 if (err)
675 goto out_fw;
676
677 intel_gt_pm_init(gt);
678
679 gt->vm = kernel_vm(gt);
680 if (!gt->vm) {
681 err = -ENOMEM;
682 goto err_pm;
683 }
684
685 err = intel_engines_init(gt);
686 if (err)
687 goto err_engines;
688
689 err = intel_uc_init(>->uc);
690 if (err)
691 goto err_engines;
692
693 err = intel_gt_resume(gt);
694 if (err)
695 goto err_uc_init;
696
697 err = __engines_record_defaults(gt);
698 if (err)
699 goto err_gt;
700
701 err = __engines_verify_workarounds(gt);
702 if (err)
703 goto err_gt;
704
705 intel_uc_init_late(>->uc);
706
707 err = i915_inject_probe_error(gt->i915, -EIO);
708 if (err)
709 goto err_gt;
710
711 intel_migrate_init(>->migrate, gt);
712
713 goto out_fw;
714 err_gt:
715 __intel_gt_disable(gt);
716 intel_uc_fini_hw(>->uc);
717 err_uc_init:
718 intel_uc_fini(>->uc);
719 err_engines:
720 intel_engines_release(gt);
721 i915_vm_put(fetch_and_zero(>->vm));
722 err_pm:
723 intel_gt_pm_fini(gt);
724 intel_gt_fini_scratch(gt);
725 out_fw:
726 if (err)
727 intel_gt_set_wedged_on_init(gt);
728 intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL);
729 return err;
730 }
731
intel_gt_driver_remove(struct intel_gt * gt)732 void intel_gt_driver_remove(struct intel_gt *gt)
733 {
734 __intel_gt_disable(gt);
735
736 intel_migrate_fini(>->migrate);
737 intel_uc_driver_remove(>->uc);
738
739 intel_engines_release(gt);
740 }
741
intel_gt_driver_unregister(struct intel_gt * gt)742 void intel_gt_driver_unregister(struct intel_gt *gt)
743 {
744 intel_wakeref_t wakeref;
745
746 intel_rps_driver_unregister(>->rps);
747
748 /*
749 * Upon unregistering the device to prevent any new users, cancel
750 * all in-flight requests so that we can quickly unbind the active
751 * resources.
752 */
753 intel_gt_set_wedged(gt);
754
755 /* Scrub all HW state upon release */
756 with_intel_runtime_pm(gt->uncore->rpm, wakeref)
757 __intel_gt_reset(gt, ALL_ENGINES);
758 }
759
intel_gt_driver_release(struct intel_gt * gt)760 void intel_gt_driver_release(struct intel_gt *gt)
761 {
762 struct i915_address_space *vm;
763
764 vm = fetch_and_zero(>->vm);
765 if (vm) /* FIXME being called twice on error paths :( */
766 i915_vm_put(vm);
767
768 intel_gt_pm_fini(gt);
769 intel_gt_fini_scratch(gt);
770 intel_gt_fini_buffer_pool(gt);
771 }
772
intel_gt_driver_late_release(struct intel_gt * gt)773 void intel_gt_driver_late_release(struct intel_gt *gt)
774 {
775 /* We need to wait for inflight RCU frees to release their grip */
776 rcu_barrier();
777
778 intel_uc_driver_late_release(>->uc);
779 intel_gt_fini_requests(gt);
780 intel_gt_fini_reset(gt);
781 intel_gt_fini_timelines(gt);
782 intel_engines_free(gt);
783 }
784
785 /**
786 * intel_gt_reg_needs_read_steering - determine whether a register read
787 * requires explicit steering
788 * @gt: GT structure
789 * @reg: the register to check steering requirements for
790 * @type: type of multicast steering to check
791 *
792 * Determines whether @reg needs explicit steering of a specific type for
793 * reads.
794 *
795 * Returns false if @reg does not belong to a register range of the given
796 * steering type, or if the default (subslice-based) steering IDs are suitable
797 * for @type steering too.
798 */
intel_gt_reg_needs_read_steering(struct intel_gt * gt,i915_reg_t reg,enum intel_steering_type type)799 static bool intel_gt_reg_needs_read_steering(struct intel_gt *gt,
800 i915_reg_t reg,
801 enum intel_steering_type type)
802 {
803 const u32 offset = i915_mmio_reg_offset(reg);
804 const struct intel_mmio_range *entry;
805
806 if (likely(!intel_gt_needs_read_steering(gt, type)))
807 return false;
808
809 for (entry = gt->steering_table[type]; entry->end; entry++) {
810 if (offset >= entry->start && offset <= entry->end)
811 return true;
812 }
813
814 return false;
815 }
816
817 /**
818 * intel_gt_get_valid_steering - determines valid IDs for a class of MCR steering
819 * @gt: GT structure
820 * @type: multicast register type
821 * @sliceid: Slice ID returned
822 * @subsliceid: Subslice ID returned
823 *
824 * Determines sliceid and subsliceid values that will steer reads
825 * of a specific multicast register class to a valid value.
826 */
intel_gt_get_valid_steering(struct intel_gt * gt,enum intel_steering_type type,u8 * sliceid,u8 * subsliceid)827 static void intel_gt_get_valid_steering(struct intel_gt *gt,
828 enum intel_steering_type type,
829 u8 *sliceid, u8 *subsliceid)
830 {
831 switch (type) {
832 case L3BANK:
833 GEM_DEBUG_WARN_ON(!gt->info.l3bank_mask); /* should be impossible! */
834
835 *sliceid = 0; /* unused */
836 *subsliceid = __ffs(gt->info.l3bank_mask);
837 break;
838 case MSLICE:
839 GEM_DEBUG_WARN_ON(!gt->info.mslice_mask); /* should be impossible! */
840
841 *sliceid = __ffs(gt->info.mslice_mask);
842 *subsliceid = 0; /* unused */
843 break;
844 case LNCF:
845 GEM_DEBUG_WARN_ON(!gt->info.mslice_mask); /* should be impossible! */
846
847 /*
848 * An LNCF is always present if its mslice is present, so we
849 * can safely just steer to LNCF 0 in all cases.
850 */
851 *sliceid = __ffs(gt->info.mslice_mask) << 1;
852 *subsliceid = 0; /* unused */
853 break;
854 default:
855 MISSING_CASE(type);
856 *sliceid = 0;
857 *subsliceid = 0;
858 }
859 }
860
861 /**
862 * intel_gt_read_register_fw - reads a GT register with support for multicast
863 * @gt: GT structure
864 * @reg: register to read
865 *
866 * This function will read a GT register. If the register is a multicast
867 * register, the read will be steered to a valid instance (i.e., one that
868 * isn't fused off or powered down by power gating).
869 *
870 * Returns the value from a valid instance of @reg.
871 */
intel_gt_read_register_fw(struct intel_gt * gt,i915_reg_t reg)872 u32 intel_gt_read_register_fw(struct intel_gt *gt, i915_reg_t reg)
873 {
874 int type;
875 u8 sliceid, subsliceid;
876
877 for (type = 0; type < NUM_STEERING_TYPES; type++) {
878 if (intel_gt_reg_needs_read_steering(gt, reg, type)) {
879 intel_gt_get_valid_steering(gt, type, &sliceid,
880 &subsliceid);
881 return intel_uncore_read_with_mcr_steering_fw(gt->uncore,
882 reg,
883 sliceid,
884 subsliceid);
885 }
886 }
887
888 return intel_uncore_read_fw(gt->uncore, reg);
889 }
890
intel_gt_info_print(const struct intel_gt_info * info,struct drm_printer * p)891 void intel_gt_info_print(const struct intel_gt_info *info,
892 struct drm_printer *p)
893 {
894 drm_printf(p, "available engines: %x\n", info->engine_mask);
895
896 intel_sseu_dump(&info->sseu, p);
897 }
898