1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2017-2018 Intel Corporation
4 */
5
6 #include <linux/prime_numbers.h>
7 #include <linux/string_helpers.h>
8
9 #include "intel_context.h"
10 #include "intel_engine_heartbeat.h"
11 #include "intel_engine_pm.h"
12 #include "intel_engine_regs.h"
13 #include "intel_gpu_commands.h"
14 #include "intel_gt.h"
15 #include "intel_gt_requests.h"
16 #include "intel_ring.h"
17 #include "selftest_engine_heartbeat.h"
18
19 #include "../selftests/i915_random.h"
20 #include "../i915_selftest.h"
21
22 #include "selftests/igt_flush_test.h"
23 #include "selftests/lib_sw_fence.h"
24 #include "selftests/mock_gem_device.h"
25 #include "selftests/mock_timeline.h"
26
hwsp_page(struct intel_timeline * tl)27 static struct vm_page *hwsp_page(struct intel_timeline *tl)
28 {
29 struct drm_i915_gem_object *obj = tl->hwsp_ggtt->obj;
30
31 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
32 return sg_page(obj->mm.pages->sgl);
33 }
34
hwsp_cacheline(struct intel_timeline * tl)35 static unsigned long hwsp_cacheline(struct intel_timeline *tl)
36 {
37 unsigned long address = (unsigned long)page_address(hwsp_page(tl));
38
39 return (address + offset_in_page(tl->hwsp_offset)) / TIMELINE_SEQNO_BYTES;
40 }
41
selftest_tl_pin(struct intel_timeline * tl)42 static int selftest_tl_pin(struct intel_timeline *tl)
43 {
44 struct i915_gem_ww_ctx ww;
45 int err;
46
47 i915_gem_ww_ctx_init(&ww, false);
48 retry:
49 err = i915_gem_object_lock(tl->hwsp_ggtt->obj, &ww);
50 if (!err)
51 err = intel_timeline_pin(tl, &ww);
52
53 if (err == -EDEADLK) {
54 err = i915_gem_ww_ctx_backoff(&ww);
55 if (!err)
56 goto retry;
57 }
58 i915_gem_ww_ctx_fini(&ww);
59 return err;
60 }
61
62 /* Only half of seqno's are usable, see __intel_timeline_get_seqno() */
63 #define CACHELINES_PER_PAGE (PAGE_SIZE / TIMELINE_SEQNO_BYTES / 2)
64
65 struct mock_hwsp_freelist {
66 struct intel_gt *gt;
67 struct radix_tree_root cachelines;
68 struct intel_timeline **history;
69 unsigned long count, max;
70 struct rnd_state prng;
71 };
72
73 enum {
74 SHUFFLE = BIT(0),
75 };
76
__mock_hwsp_record(struct mock_hwsp_freelist * state,unsigned int idx,struct intel_timeline * tl)77 static void __mock_hwsp_record(struct mock_hwsp_freelist *state,
78 unsigned int idx,
79 struct intel_timeline *tl)
80 {
81 tl = xchg(&state->history[idx], tl);
82 if (tl) {
83 radix_tree_delete(&state->cachelines, hwsp_cacheline(tl));
84 intel_timeline_unpin(tl);
85 intel_timeline_put(tl);
86 }
87 }
88
__mock_hwsp_timeline(struct mock_hwsp_freelist * state,unsigned int count,unsigned int flags)89 static int __mock_hwsp_timeline(struct mock_hwsp_freelist *state,
90 unsigned int count,
91 unsigned int flags)
92 {
93 struct intel_timeline *tl;
94 unsigned int idx;
95
96 while (count--) {
97 unsigned long cacheline;
98 int err;
99
100 tl = intel_timeline_create(state->gt);
101 if (IS_ERR(tl))
102 return PTR_ERR(tl);
103
104 err = selftest_tl_pin(tl);
105 if (err) {
106 intel_timeline_put(tl);
107 return err;
108 }
109
110 cacheline = hwsp_cacheline(tl);
111 err = radix_tree_insert(&state->cachelines, cacheline, tl);
112 if (err) {
113 if (err == -EEXIST) {
114 pr_err("HWSP cacheline %lu already used; duplicate allocation!\n",
115 cacheline);
116 }
117 intel_timeline_unpin(tl);
118 intel_timeline_put(tl);
119 return err;
120 }
121
122 idx = state->count++ % state->max;
123 __mock_hwsp_record(state, idx, tl);
124 }
125
126 if (flags & SHUFFLE)
127 i915_prandom_shuffle(state->history,
128 sizeof(*state->history),
129 min(state->count, state->max),
130 &state->prng);
131
132 count = i915_prandom_u32_max_state(min(state->count, state->max),
133 &state->prng);
134 while (count--) {
135 idx = --state->count % state->max;
136 __mock_hwsp_record(state, idx, NULL);
137 }
138
139 return 0;
140 }
141
mock_hwsp_freelist(void * arg)142 static int mock_hwsp_freelist(void *arg)
143 {
144 struct mock_hwsp_freelist state;
145 struct drm_i915_private *i915;
146 const struct {
147 const char *name;
148 unsigned int flags;
149 } phases[] = {
150 { "linear", 0 },
151 { "shuffled", SHUFFLE },
152 { },
153 }, *p;
154 unsigned int na;
155 int err = 0;
156
157 i915 = mock_gem_device();
158 if (!i915)
159 return -ENOMEM;
160
161 INIT_RADIX_TREE(&state.cachelines, GFP_KERNEL);
162 state.prng = I915_RND_STATE_INITIALIZER(i915_selftest.random_seed);
163
164 state.gt = to_gt(i915);
165
166 /*
167 * Create a bunch of timelines and check that their HWSP do not overlap.
168 * Free some, and try again.
169 */
170
171 state.max = PAGE_SIZE / sizeof(*state.history);
172 state.count = 0;
173 state.history = kcalloc(state.max, sizeof(*state.history), GFP_KERNEL);
174 if (!state.history) {
175 err = -ENOMEM;
176 goto err_put;
177 }
178
179 for (p = phases; p->name; p++) {
180 pr_debug("%s(%s)\n", __func__, p->name);
181 for_each_prime_number_from(na, 1, 2 * CACHELINES_PER_PAGE) {
182 err = __mock_hwsp_timeline(&state, na, p->flags);
183 if (err)
184 goto out;
185 }
186 }
187
188 out:
189 for (na = 0; na < state.max; na++)
190 __mock_hwsp_record(&state, na, NULL);
191 kfree(state.history);
192 err_put:
193 mock_destroy_device(i915);
194 return err;
195 }
196
197 struct __igt_sync {
198 const char *name;
199 u32 seqno;
200 bool expected;
201 bool set;
202 };
203
__igt_sync(struct intel_timeline * tl,u64 ctx,const struct __igt_sync * p,const char * name)204 static int __igt_sync(struct intel_timeline *tl,
205 u64 ctx,
206 const struct __igt_sync *p,
207 const char *name)
208 {
209 int ret;
210
211 if (__intel_timeline_sync_is_later(tl, ctx, p->seqno) != p->expected) {
212 pr_err("%s: %s(ctx=%llu, seqno=%u) expected passed %s but failed\n",
213 name, p->name, ctx, p->seqno, str_yes_no(p->expected));
214 return -EINVAL;
215 }
216
217 if (p->set) {
218 ret = __intel_timeline_sync_set(tl, ctx, p->seqno);
219 if (ret)
220 return ret;
221 }
222
223 return 0;
224 }
225
igt_sync(void * arg)226 static int igt_sync(void *arg)
227 {
228 const struct __igt_sync pass[] = {
229 { "unset", 0, false, false },
230 { "new", 0, false, true },
231 { "0a", 0, true, true },
232 { "1a", 1, false, true },
233 { "1b", 1, true, true },
234 { "0b", 0, true, false },
235 { "2a", 2, false, true },
236 { "4", 4, false, true },
237 { "INT_MAX", INT_MAX, false, true },
238 { "INT_MAX-1", INT_MAX-1, true, false },
239 { "INT_MAX+1", (u32)INT_MAX+1, false, true },
240 { "INT_MAX", INT_MAX, true, false },
241 { "UINT_MAX", UINT_MAX, false, true },
242 { "wrap", 0, false, true },
243 { "unwrap", UINT_MAX, true, false },
244 {},
245 }, *p;
246 struct intel_timeline tl;
247 int order, offset;
248 int ret = -ENODEV;
249
250 mock_timeline_init(&tl, 0);
251 for (p = pass; p->name; p++) {
252 for (order = 1; order < 64; order++) {
253 for (offset = -1; offset <= (order > 1); offset++) {
254 u64 ctx = BIT_ULL(order) + offset;
255
256 ret = __igt_sync(&tl, ctx, p, "1");
257 if (ret)
258 goto out;
259 }
260 }
261 }
262 mock_timeline_fini(&tl);
263
264 mock_timeline_init(&tl, 0);
265 for (order = 1; order < 64; order++) {
266 for (offset = -1; offset <= (order > 1); offset++) {
267 u64 ctx = BIT_ULL(order) + offset;
268
269 for (p = pass; p->name; p++) {
270 ret = __igt_sync(&tl, ctx, p, "2");
271 if (ret)
272 goto out;
273 }
274 }
275 }
276
277 out:
278 mock_timeline_fini(&tl);
279 return ret;
280 }
281
random_engine(struct rnd_state * rnd)282 static unsigned int random_engine(struct rnd_state *rnd)
283 {
284 return i915_prandom_u32_max_state(I915_NUM_ENGINES, rnd);
285 }
286
bench_sync(void * arg)287 static int bench_sync(void *arg)
288 {
289 struct rnd_state prng;
290 struct intel_timeline tl;
291 unsigned long end_time, count;
292 u64 prng32_1M;
293 ktime_t kt;
294 int order, last_order;
295
296 mock_timeline_init(&tl, 0);
297
298 /* Lookups from cache are very fast and so the random number generation
299 * and the loop itself becomes a significant factor in the per-iteration
300 * timings. We try to compensate the results by measuring the overhead
301 * of the prng and subtract it from the reported results.
302 */
303 prandom_seed_state(&prng, i915_selftest.random_seed);
304 count = 0;
305 kt = ktime_get();
306 end_time = jiffies + HZ/10;
307 do {
308 u32 x;
309
310 /* Make sure the compiler doesn't optimise away the prng call */
311 WRITE_ONCE(x, prandom_u32_state(&prng));
312
313 count++;
314 } while (!time_after(jiffies, end_time));
315 kt = ktime_sub(ktime_get(), kt);
316 pr_debug("%s: %lu random evaluations, %lluns/prng\n",
317 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
318 prng32_1M = div64_ul(ktime_to_ns(kt) << 20, count);
319
320 /* Benchmark (only) setting random context ids */
321 prandom_seed_state(&prng, i915_selftest.random_seed);
322 count = 0;
323 kt = ktime_get();
324 end_time = jiffies + HZ/10;
325 do {
326 u64 id = i915_prandom_u64_state(&prng);
327
328 __intel_timeline_sync_set(&tl, id, 0);
329 count++;
330 } while (!time_after(jiffies, end_time));
331 kt = ktime_sub(ktime_get(), kt);
332 kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
333 pr_info("%s: %lu random insertions, %lluns/insert\n",
334 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
335
336 /* Benchmark looking up the exact same context ids as we just set */
337 prandom_seed_state(&prng, i915_selftest.random_seed);
338 end_time = count;
339 kt = ktime_get();
340 while (end_time--) {
341 u64 id = i915_prandom_u64_state(&prng);
342
343 if (!__intel_timeline_sync_is_later(&tl, id, 0)) {
344 mock_timeline_fini(&tl);
345 pr_err("Lookup of %llu failed\n", id);
346 return -EINVAL;
347 }
348 }
349 kt = ktime_sub(ktime_get(), kt);
350 kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
351 pr_info("%s: %lu random lookups, %lluns/lookup\n",
352 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
353
354 mock_timeline_fini(&tl);
355 cond_resched();
356
357 mock_timeline_init(&tl, 0);
358
359 /* Benchmark setting the first N (in order) contexts */
360 count = 0;
361 kt = ktime_get();
362 end_time = jiffies + HZ/10;
363 do {
364 __intel_timeline_sync_set(&tl, count++, 0);
365 } while (!time_after(jiffies, end_time));
366 kt = ktime_sub(ktime_get(), kt);
367 pr_info("%s: %lu in-order insertions, %lluns/insert\n",
368 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
369
370 /* Benchmark looking up the exact same context ids as we just set */
371 end_time = count;
372 kt = ktime_get();
373 while (end_time--) {
374 if (!__intel_timeline_sync_is_later(&tl, end_time, 0)) {
375 pr_err("Lookup of %lu failed\n", end_time);
376 mock_timeline_fini(&tl);
377 return -EINVAL;
378 }
379 }
380 kt = ktime_sub(ktime_get(), kt);
381 pr_info("%s: %lu in-order lookups, %lluns/lookup\n",
382 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
383
384 mock_timeline_fini(&tl);
385 cond_resched();
386
387 mock_timeline_init(&tl, 0);
388
389 /* Benchmark searching for a random context id and maybe changing it */
390 prandom_seed_state(&prng, i915_selftest.random_seed);
391 count = 0;
392 kt = ktime_get();
393 end_time = jiffies + HZ/10;
394 do {
395 u32 id = random_engine(&prng);
396 u32 seqno = prandom_u32_state(&prng);
397
398 if (!__intel_timeline_sync_is_later(&tl, id, seqno))
399 __intel_timeline_sync_set(&tl, id, seqno);
400
401 count++;
402 } while (!time_after(jiffies, end_time));
403 kt = ktime_sub(ktime_get(), kt);
404 kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
405 pr_info("%s: %lu repeated insert/lookups, %lluns/op\n",
406 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
407 mock_timeline_fini(&tl);
408 cond_resched();
409
410 /* Benchmark searching for a known context id and changing the seqno */
411 for (last_order = 1, order = 1; order < 32;
412 ({ int tmp = last_order; last_order = order; order += tmp; })) {
413 unsigned int mask = BIT(order) - 1;
414
415 mock_timeline_init(&tl, 0);
416
417 count = 0;
418 kt = ktime_get();
419 end_time = jiffies + HZ/10;
420 do {
421 /* Without assuming too many details of the underlying
422 * implementation, try to identify its phase-changes
423 * (if any)!
424 */
425 u64 id = (u64)(count & mask) << order;
426
427 __intel_timeline_sync_is_later(&tl, id, 0);
428 __intel_timeline_sync_set(&tl, id, 0);
429
430 count++;
431 } while (!time_after(jiffies, end_time));
432 kt = ktime_sub(ktime_get(), kt);
433 pr_info("%s: %lu cyclic/%d insert/lookups, %lluns/op\n",
434 __func__, count, order,
435 (long long)div64_ul(ktime_to_ns(kt), count));
436 mock_timeline_fini(&tl);
437 cond_resched();
438 }
439
440 return 0;
441 }
442
intel_timeline_mock_selftests(void)443 int intel_timeline_mock_selftests(void)
444 {
445 static const struct i915_subtest tests[] = {
446 SUBTEST(mock_hwsp_freelist),
447 SUBTEST(igt_sync),
448 SUBTEST(bench_sync),
449 };
450
451 return i915_subtests(tests, NULL);
452 }
453
emit_ggtt_store_dw(struct i915_request * rq,u32 addr,u32 value)454 static int emit_ggtt_store_dw(struct i915_request *rq, u32 addr, u32 value)
455 {
456 u32 *cs;
457
458 cs = intel_ring_begin(rq, 4);
459 if (IS_ERR(cs))
460 return PTR_ERR(cs);
461
462 if (GRAPHICS_VER(rq->i915) >= 8) {
463 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
464 *cs++ = addr;
465 *cs++ = 0;
466 *cs++ = value;
467 } else if (GRAPHICS_VER(rq->i915) >= 4) {
468 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
469 *cs++ = 0;
470 *cs++ = addr;
471 *cs++ = value;
472 } else {
473 *cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
474 *cs++ = addr;
475 *cs++ = value;
476 *cs++ = MI_NOOP;
477 }
478
479 intel_ring_advance(rq, cs);
480
481 return 0;
482 }
483
484 static struct i915_request *
checked_tl_write(struct intel_timeline * tl,struct intel_engine_cs * engine,u32 value)485 checked_tl_write(struct intel_timeline *tl, struct intel_engine_cs *engine, u32 value)
486 {
487 struct i915_request *rq;
488 int err;
489
490 err = selftest_tl_pin(tl);
491 if (err) {
492 rq = ERR_PTR(err);
493 goto out;
494 }
495
496 if (READ_ONCE(*tl->hwsp_seqno) != tl->seqno) {
497 pr_err("Timeline created with incorrect breadcrumb, found %x, expected %x\n",
498 *tl->hwsp_seqno, tl->seqno);
499 intel_timeline_unpin(tl);
500 return ERR_PTR(-EINVAL);
501 }
502
503 rq = intel_engine_create_kernel_request(engine);
504 if (IS_ERR(rq))
505 goto out_unpin;
506
507 i915_request_get(rq);
508
509 err = emit_ggtt_store_dw(rq, tl->hwsp_offset, value);
510 i915_request_add(rq);
511 if (err) {
512 i915_request_put(rq);
513 rq = ERR_PTR(err);
514 }
515
516 out_unpin:
517 intel_timeline_unpin(tl);
518 out:
519 if (IS_ERR(rq))
520 pr_err("Failed to write to timeline!\n");
521 return rq;
522 }
523
live_hwsp_engine(void * arg)524 static int live_hwsp_engine(void *arg)
525 {
526 #define NUM_TIMELINES 4096
527 struct intel_gt *gt = arg;
528 struct intel_timeline **timelines;
529 struct intel_engine_cs *engine;
530 enum intel_engine_id id;
531 unsigned long count, n;
532 int err = 0;
533
534 /*
535 * Create a bunch of timelines and check we can write
536 * independently to each of their breadcrumb slots.
537 */
538
539 timelines = kvmalloc_array(NUM_TIMELINES * I915_NUM_ENGINES,
540 sizeof(*timelines),
541 GFP_KERNEL);
542 if (!timelines)
543 return -ENOMEM;
544
545 count = 0;
546 for_each_engine(engine, gt, id) {
547 if (!intel_engine_can_store_dword(engine))
548 continue;
549
550 intel_engine_pm_get(engine);
551
552 for (n = 0; n < NUM_TIMELINES; n++) {
553 struct intel_timeline *tl;
554 struct i915_request *rq;
555
556 tl = intel_timeline_create(gt);
557 if (IS_ERR(tl)) {
558 err = PTR_ERR(tl);
559 break;
560 }
561
562 rq = checked_tl_write(tl, engine, count);
563 if (IS_ERR(rq)) {
564 intel_timeline_put(tl);
565 err = PTR_ERR(rq);
566 break;
567 }
568
569 timelines[count++] = tl;
570 i915_request_put(rq);
571 }
572
573 intel_engine_pm_put(engine);
574 if (err)
575 break;
576 }
577
578 if (igt_flush_test(gt->i915))
579 err = -EIO;
580
581 for (n = 0; n < count; n++) {
582 struct intel_timeline *tl = timelines[n];
583
584 if (!err && READ_ONCE(*tl->hwsp_seqno) != n) {
585 GEM_TRACE_ERR("Invalid seqno:%lu stored in timeline %llu @ %x, found 0x%x\n",
586 n, tl->fence_context, tl->hwsp_offset, *tl->hwsp_seqno);
587 GEM_TRACE_DUMP();
588 err = -EINVAL;
589 }
590 intel_timeline_put(tl);
591 }
592
593 kvfree(timelines);
594 return err;
595 #undef NUM_TIMELINES
596 }
597
live_hwsp_alternate(void * arg)598 static int live_hwsp_alternate(void *arg)
599 {
600 #define NUM_TIMELINES 4096
601 struct intel_gt *gt = arg;
602 struct intel_timeline **timelines;
603 struct intel_engine_cs *engine;
604 enum intel_engine_id id;
605 unsigned long count, n;
606 int err = 0;
607
608 /*
609 * Create a bunch of timelines and check we can write
610 * independently to each of their breadcrumb slots with adjacent
611 * engines.
612 */
613
614 timelines = kvmalloc_array(NUM_TIMELINES * I915_NUM_ENGINES,
615 sizeof(*timelines),
616 GFP_KERNEL);
617 if (!timelines)
618 return -ENOMEM;
619
620 count = 0;
621 for (n = 0; n < NUM_TIMELINES; n++) {
622 for_each_engine(engine, gt, id) {
623 struct intel_timeline *tl;
624 struct i915_request *rq;
625
626 if (!intel_engine_can_store_dword(engine))
627 continue;
628
629 tl = intel_timeline_create(gt);
630 if (IS_ERR(tl)) {
631 err = PTR_ERR(tl);
632 goto out;
633 }
634
635 intel_engine_pm_get(engine);
636 rq = checked_tl_write(tl, engine, count);
637 intel_engine_pm_put(engine);
638 if (IS_ERR(rq)) {
639 intel_timeline_put(tl);
640 err = PTR_ERR(rq);
641 goto out;
642 }
643
644 timelines[count++] = tl;
645 i915_request_put(rq);
646 }
647 }
648
649 out:
650 if (igt_flush_test(gt->i915))
651 err = -EIO;
652
653 for (n = 0; n < count; n++) {
654 struct intel_timeline *tl = timelines[n];
655
656 if (!err && READ_ONCE(*tl->hwsp_seqno) != n) {
657 GEM_TRACE_ERR("Invalid seqno:%lu stored in timeline %llu @ %x, found 0x%x\n",
658 n, tl->fence_context, tl->hwsp_offset, *tl->hwsp_seqno);
659 GEM_TRACE_DUMP();
660 err = -EINVAL;
661 }
662 intel_timeline_put(tl);
663 }
664
665 kvfree(timelines);
666 return err;
667 #undef NUM_TIMELINES
668 }
669
live_hwsp_wrap(void * arg)670 static int live_hwsp_wrap(void *arg)
671 {
672 struct intel_gt *gt = arg;
673 struct intel_engine_cs *engine;
674 struct intel_timeline *tl;
675 enum intel_engine_id id;
676 int err = 0;
677
678 /*
679 * Across a seqno wrap, we need to keep the old cacheline alive for
680 * foreign GPU references.
681 */
682
683 tl = intel_timeline_create(gt);
684 if (IS_ERR(tl))
685 return PTR_ERR(tl);
686
687 if (!tl->has_initial_breadcrumb)
688 goto out_free;
689
690 err = selftest_tl_pin(tl);
691 if (err)
692 goto out_free;
693
694 for_each_engine(engine, gt, id) {
695 const u32 *hwsp_seqno[2];
696 struct i915_request *rq;
697 u32 seqno[2];
698
699 if (!intel_engine_can_store_dword(engine))
700 continue;
701
702 rq = intel_engine_create_kernel_request(engine);
703 if (IS_ERR(rq)) {
704 err = PTR_ERR(rq);
705 goto out;
706 }
707
708 tl->seqno = -4u;
709
710 mutex_lock_nested(&tl->mutex, SINGLE_DEPTH_NESTING);
711 err = intel_timeline_get_seqno(tl, rq, &seqno[0]);
712 mutex_unlock(&tl->mutex);
713 if (err) {
714 i915_request_add(rq);
715 goto out;
716 }
717 pr_debug("seqno[0]:%08x, hwsp_offset:%08x\n",
718 seqno[0], tl->hwsp_offset);
719
720 err = emit_ggtt_store_dw(rq, tl->hwsp_offset, seqno[0]);
721 if (err) {
722 i915_request_add(rq);
723 goto out;
724 }
725 hwsp_seqno[0] = tl->hwsp_seqno;
726
727 mutex_lock_nested(&tl->mutex, SINGLE_DEPTH_NESTING);
728 err = intel_timeline_get_seqno(tl, rq, &seqno[1]);
729 mutex_unlock(&tl->mutex);
730 if (err) {
731 i915_request_add(rq);
732 goto out;
733 }
734 pr_debug("seqno[1]:%08x, hwsp_offset:%08x\n",
735 seqno[1], tl->hwsp_offset);
736
737 err = emit_ggtt_store_dw(rq, tl->hwsp_offset, seqno[1]);
738 if (err) {
739 i915_request_add(rq);
740 goto out;
741 }
742 hwsp_seqno[1] = tl->hwsp_seqno;
743
744 /* With wrap should come a new hwsp */
745 GEM_BUG_ON(seqno[1] >= seqno[0]);
746 GEM_BUG_ON(hwsp_seqno[0] == hwsp_seqno[1]);
747
748 i915_request_add(rq);
749
750 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
751 pr_err("Wait for timeline writes timed out!\n");
752 err = -EIO;
753 goto out;
754 }
755
756 if (READ_ONCE(*hwsp_seqno[0]) != seqno[0] ||
757 READ_ONCE(*hwsp_seqno[1]) != seqno[1]) {
758 pr_err("Bad timeline values: found (%x, %x), expected (%x, %x)\n",
759 *hwsp_seqno[0], *hwsp_seqno[1],
760 seqno[0], seqno[1]);
761 err = -EINVAL;
762 goto out;
763 }
764
765 intel_gt_retire_requests(gt); /* recycle HWSP */
766 }
767
768 out:
769 if (igt_flush_test(gt->i915))
770 err = -EIO;
771
772 intel_timeline_unpin(tl);
773 out_free:
774 intel_timeline_put(tl);
775 return err;
776 }
777
emit_read_hwsp(struct i915_request * rq,u32 seqno,u32 hwsp,u32 * addr)778 static int emit_read_hwsp(struct i915_request *rq,
779 u32 seqno, u32 hwsp,
780 u32 *addr)
781 {
782 const u32 gpr = i915_mmio_reg_offset(GEN8_RING_CS_GPR(rq->engine->mmio_base, 0));
783 u32 *cs;
784
785 cs = intel_ring_begin(rq, 12);
786 if (IS_ERR(cs))
787 return PTR_ERR(cs);
788
789 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
790 *cs++ = *addr;
791 *cs++ = 0;
792 *cs++ = seqno;
793 *addr += 4;
794
795 *cs++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_USE_GGTT;
796 *cs++ = gpr;
797 *cs++ = hwsp;
798 *cs++ = 0;
799
800 *cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
801 *cs++ = gpr;
802 *cs++ = *addr;
803 *cs++ = 0;
804 *addr += 4;
805
806 intel_ring_advance(rq, cs);
807
808 return 0;
809 }
810
811 struct hwsp_watcher {
812 struct i915_vma *vma;
813 struct i915_request *rq;
814 u32 addr;
815 u32 *map;
816 };
817
cmp_lt(u32 a,u32 b)818 static bool cmp_lt(u32 a, u32 b)
819 {
820 return a < b;
821 }
822
cmp_gte(u32 a,u32 b)823 static bool cmp_gte(u32 a, u32 b)
824 {
825 return a >= b;
826 }
827
setup_watcher(struct hwsp_watcher * w,struct intel_gt * gt,struct intel_timeline * tl)828 static int setup_watcher(struct hwsp_watcher *w, struct intel_gt *gt,
829 struct intel_timeline *tl)
830 {
831 struct drm_i915_gem_object *obj;
832 struct i915_vma *vma;
833
834 obj = i915_gem_object_create_internal(gt->i915, SZ_2M);
835 if (IS_ERR(obj))
836 return PTR_ERR(obj);
837
838 /* keep the same cache settings as timeline */
839 i915_gem_object_set_pat_index(obj, tl->hwsp_ggtt->obj->pat_index);
840 w->map = i915_gem_object_pin_map_unlocked(obj,
841 page_unmask_bits(tl->hwsp_ggtt->obj->mm.mapping));
842 if (IS_ERR(w->map)) {
843 i915_gem_object_put(obj);
844 return PTR_ERR(w->map);
845 }
846
847 vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, 0);
848 if (IS_ERR(vma)) {
849 i915_gem_object_put(obj);
850 return PTR_ERR(vma);
851 }
852
853 w->vma = vma;
854 w->addr = i915_ggtt_offset(vma);
855 return 0;
856 }
857
switch_tl_lock(struct i915_request * from,struct i915_request * to)858 static void switch_tl_lock(struct i915_request *from, struct i915_request *to)
859 {
860 /* some light mutex juggling required; think co-routines */
861
862 if (from) {
863 lockdep_unpin_lock(&from->context->timeline->mutex, from->cookie);
864 mutex_unlock(&from->context->timeline->mutex);
865 }
866
867 if (to) {
868 mutex_lock(&to->context->timeline->mutex);
869 to->cookie = lockdep_pin_lock(&to->context->timeline->mutex);
870 }
871 }
872
create_watcher(struct hwsp_watcher * w,struct intel_engine_cs * engine,int ringsz)873 static int create_watcher(struct hwsp_watcher *w,
874 struct intel_engine_cs *engine,
875 int ringsz)
876 {
877 struct intel_context *ce;
878
879 ce = intel_context_create(engine);
880 if (IS_ERR(ce))
881 return PTR_ERR(ce);
882
883 ce->ring_size = ringsz;
884 w->rq = intel_context_create_request(ce);
885 intel_context_put(ce);
886 if (IS_ERR(w->rq))
887 return PTR_ERR(w->rq);
888
889 w->addr = i915_ggtt_offset(w->vma);
890
891 switch_tl_lock(w->rq, NULL);
892
893 return 0;
894 }
895
check_watcher(struct hwsp_watcher * w,const char * name,bool (* op)(u32 hwsp,u32 seqno))896 static int check_watcher(struct hwsp_watcher *w, const char *name,
897 bool (*op)(u32 hwsp, u32 seqno))
898 {
899 struct i915_request *rq = fetch_and_zero(&w->rq);
900 u32 offset, end;
901 int err;
902
903 GEM_BUG_ON(w->addr - i915_ggtt_offset(w->vma) > w->vma->size);
904
905 i915_request_get(rq);
906 switch_tl_lock(NULL, rq);
907 i915_request_add(rq);
908
909 if (i915_request_wait(rq, 0, HZ) < 0) {
910 err = -ETIME;
911 goto out;
912 }
913
914 err = 0;
915 offset = 0;
916 end = (w->addr - i915_ggtt_offset(w->vma)) / sizeof(*w->map);
917 while (offset < end) {
918 if (!op(w->map[offset + 1], w->map[offset])) {
919 pr_err("Watcher '%s' found HWSP value %x for seqno %x\n",
920 name, w->map[offset + 1], w->map[offset]);
921 err = -EINVAL;
922 }
923
924 offset += 2;
925 }
926
927 out:
928 i915_request_put(rq);
929 return err;
930 }
931
cleanup_watcher(struct hwsp_watcher * w)932 static void cleanup_watcher(struct hwsp_watcher *w)
933 {
934 if (w->rq) {
935 switch_tl_lock(NULL, w->rq);
936
937 i915_request_add(w->rq);
938 }
939
940 i915_vma_unpin_and_release(&w->vma, I915_VMA_RELEASE_MAP);
941 }
942
retire_requests(struct intel_timeline * tl)943 static bool retire_requests(struct intel_timeline *tl)
944 {
945 struct i915_request *rq, *rn;
946
947 mutex_lock(&tl->mutex);
948 list_for_each_entry_safe(rq, rn, &tl->requests, link)
949 if (!i915_request_retire(rq))
950 break;
951 mutex_unlock(&tl->mutex);
952
953 return !i915_active_fence_isset(&tl->last_request);
954 }
955
wrap_timeline(struct i915_request * rq)956 static struct i915_request *wrap_timeline(struct i915_request *rq)
957 {
958 struct intel_context *ce = rq->context;
959 struct intel_timeline *tl = ce->timeline;
960 u32 seqno = rq->fence.seqno;
961
962 while (tl->seqno >= seqno) { /* Cause a wrap */
963 i915_request_put(rq);
964 rq = intel_context_create_request(ce);
965 if (IS_ERR(rq))
966 return rq;
967
968 i915_request_get(rq);
969 i915_request_add(rq);
970 }
971
972 i915_request_put(rq);
973 rq = i915_request_create(ce);
974 if (IS_ERR(rq))
975 return rq;
976
977 i915_request_get(rq);
978 i915_request_add(rq);
979
980 return rq;
981 }
982
live_hwsp_read(void * arg)983 static int live_hwsp_read(void *arg)
984 {
985 struct intel_gt *gt = arg;
986 struct hwsp_watcher watcher[2] = {};
987 struct intel_engine_cs *engine;
988 struct intel_timeline *tl;
989 enum intel_engine_id id;
990 int err = 0;
991 int i;
992
993 /*
994 * If we take a reference to the HWSP for reading on the GPU, that
995 * read may be arbitrarily delayed (either by foreign fence or
996 * priority saturation) and a wrap can happen within 30 minutes.
997 * When the GPU read is finally submitted it should be correct,
998 * even across multiple wraps.
999 */
1000
1001 if (GRAPHICS_VER(gt->i915) < 8) /* CS convenience [SRM/LRM] */
1002 return 0;
1003
1004 tl = intel_timeline_create(gt);
1005 if (IS_ERR(tl))
1006 return PTR_ERR(tl);
1007
1008 if (!tl->has_initial_breadcrumb)
1009 goto out_free;
1010
1011 selftest_tl_pin(tl);
1012
1013 for (i = 0; i < ARRAY_SIZE(watcher); i++) {
1014 err = setup_watcher(&watcher[i], gt, tl);
1015 if (err)
1016 goto out;
1017 }
1018
1019 for_each_engine(engine, gt, id) {
1020 struct intel_context *ce;
1021 unsigned long count = 0;
1022 IGT_TIMEOUT(end_time);
1023
1024 /* Create a request we can use for remote reading of the HWSP */
1025 err = create_watcher(&watcher[1], engine, SZ_512K);
1026 if (err)
1027 goto out;
1028
1029 do {
1030 struct i915_sw_fence *submit;
1031 struct i915_request *rq;
1032 u32 hwsp, dummy;
1033
1034 submit = heap_fence_create(GFP_KERNEL);
1035 if (!submit) {
1036 err = -ENOMEM;
1037 goto out;
1038 }
1039
1040 err = create_watcher(&watcher[0], engine, SZ_4K);
1041 if (err)
1042 goto out;
1043
1044 ce = intel_context_create(engine);
1045 if (IS_ERR(ce)) {
1046 err = PTR_ERR(ce);
1047 goto out;
1048 }
1049
1050 ce->timeline = intel_timeline_get(tl);
1051
1052 /* Ensure timeline is mapped, done during first pin */
1053 err = intel_context_pin(ce);
1054 if (err) {
1055 intel_context_put(ce);
1056 goto out;
1057 }
1058
1059 /*
1060 * Start at a new wrap, and set seqno right before another wrap,
1061 * saving 30 minutes of nops
1062 */
1063 tl->seqno = -12u + 2 * (count & 3);
1064 __intel_timeline_get_seqno(tl, &dummy);
1065
1066 rq = i915_request_create(ce);
1067 if (IS_ERR(rq)) {
1068 err = PTR_ERR(rq);
1069 intel_context_unpin(ce);
1070 intel_context_put(ce);
1071 goto out;
1072 }
1073
1074 err = i915_sw_fence_await_dma_fence(&rq->submit,
1075 &watcher[0].rq->fence, 0,
1076 GFP_KERNEL);
1077 if (err < 0) {
1078 i915_request_add(rq);
1079 intel_context_unpin(ce);
1080 intel_context_put(ce);
1081 goto out;
1082 }
1083
1084 switch_tl_lock(rq, watcher[0].rq);
1085 err = intel_timeline_read_hwsp(rq, watcher[0].rq, &hwsp);
1086 if (err == 0)
1087 err = emit_read_hwsp(watcher[0].rq, /* before */
1088 rq->fence.seqno, hwsp,
1089 &watcher[0].addr);
1090 switch_tl_lock(watcher[0].rq, rq);
1091 if (err) {
1092 i915_request_add(rq);
1093 intel_context_unpin(ce);
1094 intel_context_put(ce);
1095 goto out;
1096 }
1097
1098 switch_tl_lock(rq, watcher[1].rq);
1099 err = intel_timeline_read_hwsp(rq, watcher[1].rq, &hwsp);
1100 if (err == 0)
1101 err = emit_read_hwsp(watcher[1].rq, /* after */
1102 rq->fence.seqno, hwsp,
1103 &watcher[1].addr);
1104 switch_tl_lock(watcher[1].rq, rq);
1105 if (err) {
1106 i915_request_add(rq);
1107 intel_context_unpin(ce);
1108 intel_context_put(ce);
1109 goto out;
1110 }
1111
1112 i915_request_get(rq);
1113 i915_request_add(rq);
1114
1115 rq = wrap_timeline(rq);
1116 intel_context_unpin(ce);
1117 intel_context_put(ce);
1118 if (IS_ERR(rq)) {
1119 err = PTR_ERR(rq);
1120 goto out;
1121 }
1122
1123 err = i915_sw_fence_await_dma_fence(&watcher[1].rq->submit,
1124 &rq->fence, 0,
1125 GFP_KERNEL);
1126 if (err < 0) {
1127 i915_request_put(rq);
1128 goto out;
1129 }
1130
1131 err = check_watcher(&watcher[0], "before", cmp_lt);
1132 i915_sw_fence_commit(submit);
1133 heap_fence_put(submit);
1134 if (err) {
1135 i915_request_put(rq);
1136 goto out;
1137 }
1138 count++;
1139
1140 /* Flush the timeline before manually wrapping again */
1141 if (i915_request_wait(rq,
1142 I915_WAIT_INTERRUPTIBLE,
1143 HZ) < 0) {
1144 err = -ETIME;
1145 i915_request_put(rq);
1146 goto out;
1147 }
1148 retire_requests(tl);
1149 i915_request_put(rq);
1150
1151 /* Single requests are limited to half a ring at most */
1152 if (8 * watcher[1].rq->ring->emit >
1153 3 * watcher[1].rq->ring->size)
1154 break;
1155
1156 } while (!__igt_timeout(end_time, NULL) &&
1157 count < (PAGE_SIZE / TIMELINE_SEQNO_BYTES - 1) / 2);
1158
1159 pr_info("%s: simulated %lu wraps\n", engine->name, count);
1160 err = check_watcher(&watcher[1], "after", cmp_gte);
1161 if (err)
1162 goto out;
1163 }
1164
1165 out:
1166 for (i = 0; i < ARRAY_SIZE(watcher); i++)
1167 cleanup_watcher(&watcher[i]);
1168
1169 intel_timeline_unpin(tl);
1170
1171 if (igt_flush_test(gt->i915))
1172 err = -EIO;
1173
1174 out_free:
1175 intel_timeline_put(tl);
1176 return err;
1177 }
1178
live_hwsp_rollover_kernel(void * arg)1179 static int live_hwsp_rollover_kernel(void *arg)
1180 {
1181 struct intel_gt *gt = arg;
1182 struct intel_engine_cs *engine;
1183 enum intel_engine_id id;
1184 int err = 0;
1185
1186 /*
1187 * Run the host for long enough, and even the kernel context will
1188 * see a seqno rollover.
1189 */
1190
1191 for_each_engine(engine, gt, id) {
1192 struct intel_context *ce = engine->kernel_context;
1193 struct intel_timeline *tl = ce->timeline;
1194 struct i915_request *rq[3] = {};
1195 int i;
1196
1197 st_engine_heartbeat_disable(engine);
1198 if (intel_gt_wait_for_idle(gt, HZ / 2)) {
1199 err = -EIO;
1200 goto out;
1201 }
1202
1203 GEM_BUG_ON(i915_active_fence_isset(&tl->last_request));
1204 tl->seqno = -2u;
1205 WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
1206
1207 for (i = 0; i < ARRAY_SIZE(rq); i++) {
1208 struct i915_request *this;
1209
1210 this = i915_request_create(ce);
1211 if (IS_ERR(this)) {
1212 err = PTR_ERR(this);
1213 goto out;
1214 }
1215
1216 pr_debug("%s: create fence.seqnp:%d\n",
1217 engine->name,
1218 lower_32_bits(this->fence.seqno));
1219
1220 GEM_BUG_ON(rcu_access_pointer(this->timeline) != tl);
1221
1222 rq[i] = i915_request_get(this);
1223 i915_request_add(this);
1224 }
1225
1226 /* We expected a wrap! */
1227 GEM_BUG_ON(rq[2]->fence.seqno > rq[0]->fence.seqno);
1228
1229 if (i915_request_wait(rq[2], 0, HZ / 5) < 0) {
1230 pr_err("Wait for timeline wrap timed out!\n");
1231 err = -EIO;
1232 goto out;
1233 }
1234
1235 for (i = 0; i < ARRAY_SIZE(rq); i++) {
1236 if (!i915_request_completed(rq[i])) {
1237 pr_err("Pre-wrap request not completed!\n");
1238 err = -EINVAL;
1239 goto out;
1240 }
1241 }
1242
1243 out:
1244 for (i = 0; i < ARRAY_SIZE(rq); i++)
1245 i915_request_put(rq[i]);
1246 st_engine_heartbeat_enable(engine);
1247 if (err)
1248 break;
1249 }
1250
1251 if (igt_flush_test(gt->i915))
1252 err = -EIO;
1253
1254 return err;
1255 }
1256
live_hwsp_rollover_user(void * arg)1257 static int live_hwsp_rollover_user(void *arg)
1258 {
1259 struct intel_gt *gt = arg;
1260 struct intel_engine_cs *engine;
1261 enum intel_engine_id id;
1262 int err = 0;
1263
1264 /*
1265 * Simulate a long running user context, and force the seqno wrap
1266 * on the user's timeline.
1267 */
1268
1269 for_each_engine(engine, gt, id) {
1270 struct i915_request *rq[3] = {};
1271 struct intel_timeline *tl;
1272 struct intel_context *ce;
1273 int i;
1274
1275 ce = intel_context_create(engine);
1276 if (IS_ERR(ce))
1277 return PTR_ERR(ce);
1278
1279 err = intel_context_alloc_state(ce);
1280 if (err)
1281 goto out;
1282
1283 tl = ce->timeline;
1284 if (!tl->has_initial_breadcrumb)
1285 goto out;
1286
1287 err = intel_context_pin(ce);
1288 if (err)
1289 goto out;
1290
1291 tl->seqno = -4u;
1292 WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
1293
1294 for (i = 0; i < ARRAY_SIZE(rq); i++) {
1295 struct i915_request *this;
1296
1297 this = intel_context_create_request(ce);
1298 if (IS_ERR(this)) {
1299 err = PTR_ERR(this);
1300 goto out_unpin;
1301 }
1302
1303 pr_debug("%s: create fence.seqnp:%d\n",
1304 engine->name,
1305 lower_32_bits(this->fence.seqno));
1306
1307 GEM_BUG_ON(rcu_access_pointer(this->timeline) != tl);
1308
1309 rq[i] = i915_request_get(this);
1310 i915_request_add(this);
1311 }
1312
1313 /* We expected a wrap! */
1314 GEM_BUG_ON(rq[2]->fence.seqno > rq[0]->fence.seqno);
1315
1316 if (i915_request_wait(rq[2], 0, HZ / 5) < 0) {
1317 pr_err("Wait for timeline wrap timed out!\n");
1318 err = -EIO;
1319 goto out_unpin;
1320 }
1321
1322 for (i = 0; i < ARRAY_SIZE(rq); i++) {
1323 if (!i915_request_completed(rq[i])) {
1324 pr_err("Pre-wrap request not completed!\n");
1325 err = -EINVAL;
1326 goto out_unpin;
1327 }
1328 }
1329 out_unpin:
1330 intel_context_unpin(ce);
1331 out:
1332 for (i = 0; i < ARRAY_SIZE(rq); i++)
1333 i915_request_put(rq[i]);
1334 intel_context_put(ce);
1335 if (err)
1336 break;
1337 }
1338
1339 if (igt_flush_test(gt->i915))
1340 err = -EIO;
1341
1342 return err;
1343 }
1344
live_hwsp_recycle(void * arg)1345 static int live_hwsp_recycle(void *arg)
1346 {
1347 struct intel_gt *gt = arg;
1348 struct intel_engine_cs *engine;
1349 enum intel_engine_id id;
1350 unsigned long count;
1351 int err = 0;
1352
1353 /*
1354 * Check seqno writes into one timeline at a time. We expect to
1355 * recycle the breadcrumb slot between iterations and neither
1356 * want to confuse ourselves or the GPU.
1357 */
1358
1359 count = 0;
1360 for_each_engine(engine, gt, id) {
1361 IGT_TIMEOUT(end_time);
1362
1363 if (!intel_engine_can_store_dword(engine))
1364 continue;
1365
1366 intel_engine_pm_get(engine);
1367
1368 do {
1369 struct intel_timeline *tl;
1370 struct i915_request *rq;
1371
1372 tl = intel_timeline_create(gt);
1373 if (IS_ERR(tl)) {
1374 err = PTR_ERR(tl);
1375 break;
1376 }
1377
1378 rq = checked_tl_write(tl, engine, count);
1379 if (IS_ERR(rq)) {
1380 intel_timeline_put(tl);
1381 err = PTR_ERR(rq);
1382 break;
1383 }
1384
1385 if (i915_request_wait(rq, 0, HZ / 5) < 0) {
1386 pr_err("Wait for timeline writes timed out!\n");
1387 i915_request_put(rq);
1388 intel_timeline_put(tl);
1389 err = -EIO;
1390 break;
1391 }
1392
1393 if (READ_ONCE(*tl->hwsp_seqno) != count) {
1394 GEM_TRACE_ERR("Invalid seqno:%lu stored in timeline %llu @ %x found 0x%x\n",
1395 count, tl->fence_context,
1396 tl->hwsp_offset, *tl->hwsp_seqno);
1397 GEM_TRACE_DUMP();
1398 err = -EINVAL;
1399 }
1400
1401 i915_request_put(rq);
1402 intel_timeline_put(tl);
1403 count++;
1404
1405 if (err)
1406 break;
1407 } while (!__igt_timeout(end_time, NULL));
1408
1409 intel_engine_pm_put(engine);
1410 if (err)
1411 break;
1412 }
1413
1414 return err;
1415 }
1416
intel_timeline_live_selftests(struct drm_i915_private * i915)1417 int intel_timeline_live_selftests(struct drm_i915_private *i915)
1418 {
1419 static const struct i915_subtest tests[] = {
1420 SUBTEST(live_hwsp_recycle),
1421 SUBTEST(live_hwsp_engine),
1422 SUBTEST(live_hwsp_alternate),
1423 SUBTEST(live_hwsp_wrap),
1424 SUBTEST(live_hwsp_read),
1425 SUBTEST(live_hwsp_rollover_kernel),
1426 SUBTEST(live_hwsp_rollover_user),
1427 };
1428
1429 if (intel_gt_is_wedged(to_gt(i915)))
1430 return 0;
1431
1432 return intel_gt_live_subtests(tests, to_gt(i915));
1433 }
1434