1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2017 Intel Corporation 5 */ 6 7 #include <linux/prime_numbers.h> 8 9 #include "gem/i915_gem_pm.h" 10 #include "gt/intel_engine_pm.h" 11 #include "gt/intel_gt.h" 12 #include "gt/intel_gt_requests.h" 13 #include "gt/intel_reset.h" 14 #include "i915_selftest.h" 15 16 #include "gem/selftests/igt_gem_utils.h" 17 #include "selftests/i915_random.h" 18 #include "selftests/igt_flush_test.h" 19 #include "selftests/igt_live_test.h" 20 #include "selftests/igt_reset.h" 21 #include "selftests/igt_spinner.h" 22 #include "selftests/mock_drm.h" 23 #include "selftests/mock_gem_device.h" 24 25 #include "huge_gem_object.h" 26 #include "igt_gem_utils.h" 27 28 #define DW_PER_PAGE (PAGE_SIZE / sizeof(u32)) 29 30 static inline struct i915_address_space *ctx_vm(struct i915_gem_context *ctx) 31 { 32 /* single threaded, private ctx */ 33 return rcu_dereference_protected(ctx->vm, true); 34 } 35 36 static int live_nop_switch(void *arg) 37 { 38 const unsigned int nctx = 1024; 39 struct drm_i915_private *i915 = arg; 40 struct intel_engine_cs *engine; 41 struct i915_gem_context **ctx; 42 struct igt_live_test t; 43 struct file *file; 44 unsigned long n; 45 int err = -ENODEV; 46 47 /* 48 * Create as many contexts as we can feasibly get away with 49 * and check we can switch between them rapidly. 50 * 51 * Serves as very simple stress test for submission and HW switching 52 * between contexts. 53 */ 54 55 if (!DRIVER_CAPS(i915)->has_logical_contexts) 56 return 0; 57 58 file = mock_file(i915); 59 if (IS_ERR(file)) 60 return PTR_ERR(file); 61 62 ctx = kcalloc(nctx, sizeof(*ctx), GFP_KERNEL); 63 if (!ctx) { 64 err = -ENOMEM; 65 goto out_file; 66 } 67 68 for (n = 0; n < nctx; n++) { 69 ctx[n] = live_context(i915, file); 70 if (IS_ERR(ctx[n])) { 71 err = PTR_ERR(ctx[n]); 72 goto out_file; 73 } 74 } 75 76 for_each_uabi_engine(engine, i915) { 77 struct i915_request *rq = NULL; 78 unsigned long end_time, prime; 79 ktime_t times[2] = {}; 80 81 times[0] = ktime_get_raw(); 82 for (n = 0; n < nctx; n++) { 83 struct i915_request *this; 84 85 this = igt_request_alloc(ctx[n], engine); 86 if (IS_ERR(this)) { 87 err = PTR_ERR(this); 88 goto out_file; 89 } 90 if (rq) { 91 i915_request_await_dma_fence(this, &rq->fence); 92 i915_request_put(rq); 93 } 94 rq = i915_request_get(this); 95 i915_request_add(this); 96 } 97 if (i915_request_wait(rq, 0, HZ / 5) < 0) { 98 pr_err("Failed to populated %d contexts\n", nctx); 99 intel_gt_set_wedged(&i915->gt); 100 i915_request_put(rq); 101 err = -EIO; 102 goto out_file; 103 } 104 i915_request_put(rq); 105 106 times[1] = ktime_get_raw(); 107 108 pr_info("Populated %d contexts on %s in %lluns\n", 109 nctx, engine->name, ktime_to_ns(times[1] - times[0])); 110 111 err = igt_live_test_begin(&t, i915, __func__, engine->name); 112 if (err) 113 goto out_file; 114 115 end_time = jiffies + i915_selftest.timeout_jiffies; 116 for_each_prime_number_from(prime, 2, 8192) { 117 times[1] = ktime_get_raw(); 118 119 rq = NULL; 120 for (n = 0; n < prime; n++) { 121 struct i915_request *this; 122 123 this = igt_request_alloc(ctx[n % nctx], engine); 124 if (IS_ERR(this)) { 125 err = PTR_ERR(this); 126 goto out_file; 127 } 128 129 if (rq) { /* Force submission order */ 130 i915_request_await_dma_fence(this, &rq->fence); 131 i915_request_put(rq); 132 } 133 134 /* 135 * This space is left intentionally blank. 136 * 137 * We do not actually want to perform any 138 * action with this request, we just want 139 * to measure the latency in allocation 140 * and submission of our breadcrumbs - 141 * ensuring that the bare request is sufficient 142 * for the system to work (i.e. proper HEAD 143 * tracking of the rings, interrupt handling, 144 * etc). It also gives us the lowest bounds 145 * for latency. 146 */ 147 148 rq = i915_request_get(this); 149 i915_request_add(this); 150 } 151 GEM_BUG_ON(!rq); 152 if (i915_request_wait(rq, 0, HZ / 5) < 0) { 153 pr_err("Switching between %ld contexts timed out\n", 154 prime); 155 intel_gt_set_wedged(&i915->gt); 156 i915_request_put(rq); 157 break; 158 } 159 i915_request_put(rq); 160 161 times[1] = ktime_sub(ktime_get_raw(), times[1]); 162 if (prime == 2) 163 times[0] = times[1]; 164 165 if (__igt_timeout(end_time, NULL)) 166 break; 167 } 168 169 err = igt_live_test_end(&t); 170 if (err) 171 goto out_file; 172 173 pr_info("Switch latencies on %s: 1 = %lluns, %lu = %lluns\n", 174 engine->name, 175 ktime_to_ns(times[0]), 176 prime - 1, div64_u64(ktime_to_ns(times[1]), prime - 1)); 177 } 178 179 out_file: 180 fput(file); 181 return err; 182 } 183 184 struct parallel_switch { 185 struct task_struct *tsk; 186 struct intel_context *ce[2]; 187 }; 188 189 static int __live_parallel_switch1(void *data) 190 { 191 struct parallel_switch *arg = data; 192 IGT_TIMEOUT(end_time); 193 unsigned long count; 194 195 count = 0; 196 do { 197 struct i915_request *rq = NULL; 198 int err, n; 199 200 err = 0; 201 for (n = 0; !err && n < ARRAY_SIZE(arg->ce); n++) { 202 struct i915_request *prev = rq; 203 204 rq = i915_request_create(arg->ce[n]); 205 if (IS_ERR(rq)) { 206 i915_request_put(prev); 207 return PTR_ERR(rq); 208 } 209 210 i915_request_get(rq); 211 if (prev) { 212 err = i915_request_await_dma_fence(rq, &prev->fence); 213 i915_request_put(prev); 214 } 215 216 i915_request_add(rq); 217 } 218 if (i915_request_wait(rq, 0, HZ / 5) < 0) 219 err = -ETIME; 220 i915_request_put(rq); 221 if (err) 222 return err; 223 224 count++; 225 } while (!__igt_timeout(end_time, NULL)); 226 227 pr_info("%s: %lu switches (sync)\n", arg->ce[0]->engine->name, count); 228 return 0; 229 } 230 231 static int __live_parallel_switchN(void *data) 232 { 233 struct parallel_switch *arg = data; 234 struct i915_request *rq = NULL; 235 IGT_TIMEOUT(end_time); 236 unsigned long count; 237 int n; 238 239 count = 0; 240 do { 241 for (n = 0; n < ARRAY_SIZE(arg->ce); n++) { 242 struct i915_request *prev = rq; 243 int err = 0; 244 245 rq = i915_request_create(arg->ce[n]); 246 if (IS_ERR(rq)) { 247 i915_request_put(prev); 248 return PTR_ERR(rq); 249 } 250 251 i915_request_get(rq); 252 if (prev) { 253 err = i915_request_await_dma_fence(rq, &prev->fence); 254 i915_request_put(prev); 255 } 256 257 i915_request_add(rq); 258 if (err) { 259 i915_request_put(rq); 260 return err; 261 } 262 } 263 264 count++; 265 } while (!__igt_timeout(end_time, NULL)); 266 i915_request_put(rq); 267 268 pr_info("%s: %lu switches (many)\n", arg->ce[0]->engine->name, count); 269 return 0; 270 } 271 272 static int live_parallel_switch(void *arg) 273 { 274 struct drm_i915_private *i915 = arg; 275 static int (* const func[])(void *arg) = { 276 __live_parallel_switch1, 277 __live_parallel_switchN, 278 NULL, 279 }; 280 struct parallel_switch *data = NULL; 281 struct i915_gem_engines *engines; 282 struct i915_gem_engines_iter it; 283 int (* const *fn)(void *arg); 284 struct i915_gem_context *ctx; 285 struct intel_context *ce; 286 struct file *file; 287 int n, m, count; 288 int err = 0; 289 290 /* 291 * Check we can process switches on all engines simultaneously. 292 */ 293 294 if (!DRIVER_CAPS(i915)->has_logical_contexts) 295 return 0; 296 297 file = mock_file(i915); 298 if (IS_ERR(file)) 299 return PTR_ERR(file); 300 301 ctx = live_context(i915, file); 302 if (IS_ERR(ctx)) { 303 err = PTR_ERR(ctx); 304 goto out_file; 305 } 306 307 engines = i915_gem_context_lock_engines(ctx); 308 count = engines->num_engines; 309 310 data = kcalloc(count, sizeof(*data), GFP_KERNEL); 311 if (!data) { 312 i915_gem_context_unlock_engines(ctx); 313 err = -ENOMEM; 314 goto out_file; 315 } 316 317 m = 0; /* Use the first context as our template for the engines */ 318 for_each_gem_engine(ce, engines, it) { 319 err = intel_context_pin(ce); 320 if (err) { 321 i915_gem_context_unlock_engines(ctx); 322 goto out; 323 } 324 data[m++].ce[0] = intel_context_get(ce); 325 } 326 i915_gem_context_unlock_engines(ctx); 327 328 /* Clone the same set of engines into the other contexts */ 329 for (n = 1; n < ARRAY_SIZE(data->ce); n++) { 330 ctx = live_context(i915, file); 331 if (IS_ERR(ctx)) { 332 err = PTR_ERR(ctx); 333 goto out; 334 } 335 336 for (m = 0; m < count; m++) { 337 if (!data[m].ce[0]) 338 continue; 339 340 ce = intel_context_create(data[m].ce[0]->engine); 341 if (IS_ERR(ce)) 342 goto out; 343 344 err = intel_context_pin(ce); 345 if (err) { 346 intel_context_put(ce); 347 goto out; 348 } 349 350 data[m].ce[n] = ce; 351 } 352 } 353 354 for (fn = func; !err && *fn; fn++) { 355 struct igt_live_test t; 356 int n; 357 358 err = igt_live_test_begin(&t, i915, __func__, ""); 359 if (err) 360 break; 361 362 for (n = 0; n < count; n++) { 363 if (!data[n].ce[0]) 364 continue; 365 366 data[n].tsk = kthread_run(*fn, &data[n], 367 "igt/parallel:%s", 368 data[n].ce[0]->engine->name); 369 if (IS_ERR(data[n].tsk)) { 370 err = PTR_ERR(data[n].tsk); 371 break; 372 } 373 get_task_struct(data[n].tsk); 374 } 375 376 yield(); /* start all threads before we kthread_stop() */ 377 378 for (n = 0; n < count; n++) { 379 int status; 380 381 if (IS_ERR_OR_NULL(data[n].tsk)) 382 continue; 383 384 status = kthread_stop(data[n].tsk); 385 if (status && !err) 386 err = status; 387 388 put_task_struct(data[n].tsk); 389 data[n].tsk = NULL; 390 } 391 392 if (igt_live_test_end(&t)) 393 err = -EIO; 394 } 395 396 out: 397 for (n = 0; n < count; n++) { 398 for (m = 0; m < ARRAY_SIZE(data->ce); m++) { 399 if (!data[n].ce[m]) 400 continue; 401 402 intel_context_unpin(data[n].ce[m]); 403 intel_context_put(data[n].ce[m]); 404 } 405 } 406 kfree(data); 407 out_file: 408 fput(file); 409 return err; 410 } 411 412 static unsigned long real_page_count(struct drm_i915_gem_object *obj) 413 { 414 return huge_gem_object_phys_size(obj) >> PAGE_SHIFT; 415 } 416 417 static unsigned long fake_page_count(struct drm_i915_gem_object *obj) 418 { 419 return huge_gem_object_dma_size(obj) >> PAGE_SHIFT; 420 } 421 422 static int gpu_fill(struct intel_context *ce, 423 struct drm_i915_gem_object *obj, 424 unsigned int dw) 425 { 426 struct i915_vma *vma; 427 int err; 428 429 GEM_BUG_ON(obj->base.size > ce->vm->total); 430 GEM_BUG_ON(!intel_engine_can_store_dword(ce->engine)); 431 432 vma = i915_vma_instance(obj, ce->vm, NULL); 433 if (IS_ERR(vma)) 434 return PTR_ERR(vma); 435 436 err = i915_vma_pin(vma, 0, 0, PIN_HIGH | PIN_USER); 437 if (err) 438 return err; 439 440 /* 441 * Within the GTT the huge objects maps every page onto 442 * its 1024 real pages (using phys_pfn = dma_pfn % 1024). 443 * We set the nth dword within the page using the nth 444 * mapping via the GTT - this should exercise the GTT mapping 445 * whilst checking that each context provides a unique view 446 * into the object. 447 */ 448 err = igt_gpu_fill_dw(ce, vma, 449 (dw * real_page_count(obj)) << PAGE_SHIFT | 450 (dw * sizeof(u32)), 451 real_page_count(obj), 452 dw); 453 i915_vma_unpin(vma); 454 455 return err; 456 } 457 458 static int cpu_fill(struct drm_i915_gem_object *obj, u32 value) 459 { 460 const bool has_llc = HAS_LLC(to_i915(obj->base.dev)); 461 unsigned int n, m, need_flush; 462 int err; 463 464 err = i915_gem_object_prepare_write(obj, &need_flush); 465 if (err) 466 return err; 467 468 for (n = 0; n < real_page_count(obj); n++) { 469 u32 *map; 470 471 map = kmap_atomic(i915_gem_object_get_page(obj, n)); 472 for (m = 0; m < DW_PER_PAGE; m++) 473 map[m] = value; 474 if (!has_llc) 475 drm_clflush_virt_range(map, PAGE_SIZE); 476 kunmap_atomic(map); 477 } 478 479 i915_gem_object_finish_access(obj); 480 obj->read_domains = I915_GEM_DOMAIN_GTT | I915_GEM_DOMAIN_CPU; 481 obj->write_domain = 0; 482 return 0; 483 } 484 485 static noinline int cpu_check(struct drm_i915_gem_object *obj, 486 unsigned int idx, unsigned int max) 487 { 488 unsigned int n, m, needs_flush; 489 int err; 490 491 err = i915_gem_object_prepare_read(obj, &needs_flush); 492 if (err) 493 return err; 494 495 for (n = 0; n < real_page_count(obj); n++) { 496 u32 *map; 497 498 map = kmap_atomic(i915_gem_object_get_page(obj, n)); 499 if (needs_flush & CLFLUSH_BEFORE) 500 drm_clflush_virt_range(map, PAGE_SIZE); 501 502 for (m = 0; m < max; m++) { 503 if (map[m] != m) { 504 pr_err("%pS: Invalid value at object %d page %d/%ld, offset %d/%d: found %x expected %x\n", 505 __builtin_return_address(0), idx, 506 n, real_page_count(obj), m, max, 507 map[m], m); 508 err = -EINVAL; 509 goto out_unmap; 510 } 511 } 512 513 for (; m < DW_PER_PAGE; m++) { 514 if (map[m] != STACK_MAGIC) { 515 pr_err("%pS: Invalid value at object %d page %d, offset %d: found %x expected %x (uninitialised)\n", 516 __builtin_return_address(0), idx, n, m, 517 map[m], STACK_MAGIC); 518 err = -EINVAL; 519 goto out_unmap; 520 } 521 } 522 523 out_unmap: 524 kunmap_atomic(map); 525 if (err) 526 break; 527 } 528 529 i915_gem_object_finish_access(obj); 530 return err; 531 } 532 533 static int file_add_object(struct file *file, struct drm_i915_gem_object *obj) 534 { 535 int err; 536 537 GEM_BUG_ON(obj->base.handle_count); 538 539 /* tie the object to the drm_file for easy reaping */ 540 err = idr_alloc(&to_drm_file(file)->object_idr, 541 &obj->base, 1, 0, GFP_KERNEL); 542 if (err < 0) 543 return err; 544 545 i915_gem_object_get(obj); 546 obj->base.handle_count++; 547 return 0; 548 } 549 550 static struct drm_i915_gem_object * 551 create_test_object(struct i915_address_space *vm, 552 struct file *file, 553 struct list_head *objects) 554 { 555 struct drm_i915_gem_object *obj; 556 u64 size; 557 int err; 558 559 /* Keep in GEM's good graces */ 560 intel_gt_retire_requests(vm->gt); 561 562 size = min(vm->total / 2, 1024ull * DW_PER_PAGE * PAGE_SIZE); 563 size = round_down(size, DW_PER_PAGE * PAGE_SIZE); 564 565 obj = huge_gem_object(vm->i915, DW_PER_PAGE * PAGE_SIZE, size); 566 if (IS_ERR(obj)) 567 return obj; 568 569 err = file_add_object(file, obj); 570 i915_gem_object_put(obj); 571 if (err) 572 return ERR_PTR(err); 573 574 err = cpu_fill(obj, STACK_MAGIC); 575 if (err) { 576 pr_err("Failed to fill object with cpu, err=%d\n", 577 err); 578 return ERR_PTR(err); 579 } 580 581 list_add_tail(&obj->st_link, objects); 582 return obj; 583 } 584 585 static unsigned long max_dwords(struct drm_i915_gem_object *obj) 586 { 587 unsigned long npages = fake_page_count(obj); 588 589 GEM_BUG_ON(!IS_ALIGNED(npages, DW_PER_PAGE)); 590 return npages / DW_PER_PAGE; 591 } 592 593 static void throttle_release(struct i915_request **q, int count) 594 { 595 int i; 596 597 for (i = 0; i < count; i++) { 598 if (IS_ERR_OR_NULL(q[i])) 599 continue; 600 601 i915_request_put(fetch_and_zero(&q[i])); 602 } 603 } 604 605 static int throttle(struct intel_context *ce, 606 struct i915_request **q, int count) 607 { 608 int i; 609 610 if (!IS_ERR_OR_NULL(q[0])) { 611 if (i915_request_wait(q[0], 612 I915_WAIT_INTERRUPTIBLE, 613 MAX_SCHEDULE_TIMEOUT) < 0) 614 return -EINTR; 615 616 i915_request_put(q[0]); 617 } 618 619 for (i = 0; i < count - 1; i++) 620 q[i] = q[i + 1]; 621 622 q[i] = intel_context_create_request(ce); 623 if (IS_ERR(q[i])) 624 return PTR_ERR(q[i]); 625 626 i915_request_get(q[i]); 627 i915_request_add(q[i]); 628 629 return 0; 630 } 631 632 static int igt_ctx_exec(void *arg) 633 { 634 struct drm_i915_private *i915 = arg; 635 struct intel_engine_cs *engine; 636 int err = -ENODEV; 637 638 /* 639 * Create a few different contexts (with different mm) and write 640 * through each ctx/mm using the GPU making sure those writes end 641 * up in the expected pages of our obj. 642 */ 643 644 if (!DRIVER_CAPS(i915)->has_logical_contexts) 645 return 0; 646 647 for_each_uabi_engine(engine, i915) { 648 struct drm_i915_gem_object *obj = NULL; 649 unsigned long ncontexts, ndwords, dw; 650 struct i915_request *tq[5] = {}; 651 struct igt_live_test t; 652 IGT_TIMEOUT(end_time); 653 LIST_HEAD(objects); 654 struct file *file; 655 656 if (!intel_engine_can_store_dword(engine)) 657 continue; 658 659 if (!engine->context_size) 660 continue; /* No logical context support in HW */ 661 662 file = mock_file(i915); 663 if (IS_ERR(file)) 664 return PTR_ERR(file); 665 666 err = igt_live_test_begin(&t, i915, __func__, engine->name); 667 if (err) 668 goto out_file; 669 670 ncontexts = 0; 671 ndwords = 0; 672 dw = 0; 673 while (!time_after(jiffies, end_time)) { 674 struct i915_gem_context *ctx; 675 struct intel_context *ce; 676 677 ctx = kernel_context(i915); 678 if (IS_ERR(ctx)) { 679 err = PTR_ERR(ctx); 680 goto out_file; 681 } 682 683 ce = i915_gem_context_get_engine(ctx, engine->legacy_idx); 684 GEM_BUG_ON(IS_ERR(ce)); 685 686 if (!obj) { 687 obj = create_test_object(ce->vm, file, &objects); 688 if (IS_ERR(obj)) { 689 err = PTR_ERR(obj); 690 intel_context_put(ce); 691 kernel_context_close(ctx); 692 goto out_file; 693 } 694 } 695 696 err = gpu_fill(ce, obj, dw); 697 if (err) { 698 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n", 699 ndwords, dw, max_dwords(obj), 700 engine->name, 701 yesno(!!rcu_access_pointer(ctx->vm)), 702 err); 703 intel_context_put(ce); 704 kernel_context_close(ctx); 705 goto out_file; 706 } 707 708 err = throttle(ce, tq, ARRAY_SIZE(tq)); 709 if (err) { 710 intel_context_put(ce); 711 kernel_context_close(ctx); 712 goto out_file; 713 } 714 715 if (++dw == max_dwords(obj)) { 716 obj = NULL; 717 dw = 0; 718 } 719 720 ndwords++; 721 ncontexts++; 722 723 intel_context_put(ce); 724 kernel_context_close(ctx); 725 } 726 727 pr_info("Submitted %lu contexts to %s, filling %lu dwords\n", 728 ncontexts, engine->name, ndwords); 729 730 ncontexts = dw = 0; 731 list_for_each_entry(obj, &objects, st_link) { 732 unsigned int rem = 733 min_t(unsigned int, ndwords - dw, max_dwords(obj)); 734 735 err = cpu_check(obj, ncontexts++, rem); 736 if (err) 737 break; 738 739 dw += rem; 740 } 741 742 out_file: 743 throttle_release(tq, ARRAY_SIZE(tq)); 744 if (igt_live_test_end(&t)) 745 err = -EIO; 746 747 fput(file); 748 if (err) 749 return err; 750 751 i915_gem_drain_freed_objects(i915); 752 } 753 754 return 0; 755 } 756 757 static int igt_shared_ctx_exec(void *arg) 758 { 759 struct drm_i915_private *i915 = arg; 760 struct i915_request *tq[5] = {}; 761 struct i915_gem_context *parent; 762 struct intel_engine_cs *engine; 763 struct igt_live_test t; 764 struct file *file; 765 int err = 0; 766 767 /* 768 * Create a few different contexts with the same mm and write 769 * through each ctx using the GPU making sure those writes end 770 * up in the expected pages of our obj. 771 */ 772 if (!DRIVER_CAPS(i915)->has_logical_contexts) 773 return 0; 774 775 file = mock_file(i915); 776 if (IS_ERR(file)) 777 return PTR_ERR(file); 778 779 parent = live_context(i915, file); 780 if (IS_ERR(parent)) { 781 err = PTR_ERR(parent); 782 goto out_file; 783 } 784 785 if (!parent->vm) { /* not full-ppgtt; nothing to share */ 786 err = 0; 787 goto out_file; 788 } 789 790 err = igt_live_test_begin(&t, i915, __func__, ""); 791 if (err) 792 goto out_file; 793 794 for_each_uabi_engine(engine, i915) { 795 unsigned long ncontexts, ndwords, dw; 796 struct drm_i915_gem_object *obj = NULL; 797 IGT_TIMEOUT(end_time); 798 LIST_HEAD(objects); 799 800 if (!intel_engine_can_store_dword(engine)) 801 continue; 802 803 dw = 0; 804 ndwords = 0; 805 ncontexts = 0; 806 while (!time_after(jiffies, end_time)) { 807 struct i915_gem_context *ctx; 808 struct intel_context *ce; 809 810 ctx = kernel_context(i915); 811 if (IS_ERR(ctx)) { 812 err = PTR_ERR(ctx); 813 goto out_test; 814 } 815 816 mutex_lock(&ctx->mutex); 817 __assign_ppgtt(ctx, ctx_vm(parent)); 818 mutex_unlock(&ctx->mutex); 819 820 ce = i915_gem_context_get_engine(ctx, engine->legacy_idx); 821 GEM_BUG_ON(IS_ERR(ce)); 822 823 if (!obj) { 824 obj = create_test_object(ctx_vm(parent), 825 file, &objects); 826 if (IS_ERR(obj)) { 827 err = PTR_ERR(obj); 828 intel_context_put(ce); 829 kernel_context_close(ctx); 830 goto out_test; 831 } 832 } 833 834 err = gpu_fill(ce, obj, dw); 835 if (err) { 836 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n", 837 ndwords, dw, max_dwords(obj), 838 engine->name, 839 yesno(!!rcu_access_pointer(ctx->vm)), 840 err); 841 intel_context_put(ce); 842 kernel_context_close(ctx); 843 goto out_test; 844 } 845 846 err = throttle(ce, tq, ARRAY_SIZE(tq)); 847 if (err) { 848 intel_context_put(ce); 849 kernel_context_close(ctx); 850 goto out_test; 851 } 852 853 if (++dw == max_dwords(obj)) { 854 obj = NULL; 855 dw = 0; 856 } 857 858 ndwords++; 859 ncontexts++; 860 861 intel_context_put(ce); 862 kernel_context_close(ctx); 863 } 864 pr_info("Submitted %lu contexts to %s, filling %lu dwords\n", 865 ncontexts, engine->name, ndwords); 866 867 ncontexts = dw = 0; 868 list_for_each_entry(obj, &objects, st_link) { 869 unsigned int rem = 870 min_t(unsigned int, ndwords - dw, max_dwords(obj)); 871 872 err = cpu_check(obj, ncontexts++, rem); 873 if (err) 874 goto out_test; 875 876 dw += rem; 877 } 878 879 i915_gem_drain_freed_objects(i915); 880 } 881 out_test: 882 throttle_release(tq, ARRAY_SIZE(tq)); 883 if (igt_live_test_end(&t)) 884 err = -EIO; 885 out_file: 886 fput(file); 887 return err; 888 } 889 890 static struct i915_vma *rpcs_query_batch(struct i915_vma *vma) 891 { 892 struct drm_i915_gem_object *obj; 893 u32 *cmd; 894 int err; 895 896 if (INTEL_GEN(vma->vm->i915) < 8) 897 return ERR_PTR(-EINVAL); 898 899 obj = i915_gem_object_create_internal(vma->vm->i915, PAGE_SIZE); 900 if (IS_ERR(obj)) 901 return ERR_CAST(obj); 902 903 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB); 904 if (IS_ERR(cmd)) { 905 err = PTR_ERR(cmd); 906 goto err; 907 } 908 909 *cmd++ = MI_STORE_REGISTER_MEM_GEN8; 910 *cmd++ = i915_mmio_reg_offset(GEN8_R_PWR_CLK_STATE); 911 *cmd++ = lower_32_bits(vma->node.start); 912 *cmd++ = upper_32_bits(vma->node.start); 913 *cmd = MI_BATCH_BUFFER_END; 914 915 __i915_gem_object_flush_map(obj, 0, 64); 916 i915_gem_object_unpin_map(obj); 917 918 intel_gt_chipset_flush(vma->vm->gt); 919 920 vma = i915_vma_instance(obj, vma->vm, NULL); 921 if (IS_ERR(vma)) { 922 err = PTR_ERR(vma); 923 goto err; 924 } 925 926 err = i915_vma_pin(vma, 0, 0, PIN_USER); 927 if (err) 928 goto err; 929 930 return vma; 931 932 err: 933 i915_gem_object_put(obj); 934 return ERR_PTR(err); 935 } 936 937 static int 938 emit_rpcs_query(struct drm_i915_gem_object *obj, 939 struct intel_context *ce, 940 struct i915_request **rq_out) 941 { 942 struct i915_request *rq; 943 struct i915_vma *batch; 944 struct i915_vma *vma; 945 int err; 946 947 GEM_BUG_ON(!intel_engine_can_store_dword(ce->engine)); 948 949 vma = i915_vma_instance(obj, ce->vm, NULL); 950 if (IS_ERR(vma)) 951 return PTR_ERR(vma); 952 953 i915_gem_object_lock(obj); 954 err = i915_gem_object_set_to_gtt_domain(obj, false); 955 i915_gem_object_unlock(obj); 956 if (err) 957 return err; 958 959 err = i915_vma_pin(vma, 0, 0, PIN_USER); 960 if (err) 961 return err; 962 963 batch = rpcs_query_batch(vma); 964 if (IS_ERR(batch)) { 965 err = PTR_ERR(batch); 966 goto err_vma; 967 } 968 969 rq = i915_request_create(ce); 970 if (IS_ERR(rq)) { 971 err = PTR_ERR(rq); 972 goto err_batch; 973 } 974 975 err = rq->engine->emit_bb_start(rq, 976 batch->node.start, batch->node.size, 977 0); 978 if (err) 979 goto err_request; 980 981 i915_vma_lock(batch); 982 err = i915_request_await_object(rq, batch->obj, false); 983 if (err == 0) 984 err = i915_vma_move_to_active(batch, rq, 0); 985 i915_vma_unlock(batch); 986 if (err) 987 goto skip_request; 988 989 i915_vma_lock(vma); 990 err = i915_request_await_object(rq, vma->obj, true); 991 if (err == 0) 992 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE); 993 i915_vma_unlock(vma); 994 if (err) 995 goto skip_request; 996 997 i915_vma_unpin_and_release(&batch, 0); 998 i915_vma_unpin(vma); 999 1000 *rq_out = i915_request_get(rq); 1001 1002 i915_request_add(rq); 1003 1004 return 0; 1005 1006 skip_request: 1007 i915_request_skip(rq, err); 1008 err_request: 1009 i915_request_add(rq); 1010 err_batch: 1011 i915_vma_unpin_and_release(&batch, 0); 1012 err_vma: 1013 i915_vma_unpin(vma); 1014 1015 return err; 1016 } 1017 1018 #define TEST_IDLE BIT(0) 1019 #define TEST_BUSY BIT(1) 1020 #define TEST_RESET BIT(2) 1021 1022 static int 1023 __sseu_prepare(const char *name, 1024 unsigned int flags, 1025 struct intel_context *ce, 1026 struct igt_spinner **spin) 1027 { 1028 struct i915_request *rq; 1029 int ret; 1030 1031 *spin = NULL; 1032 if (!(flags & (TEST_BUSY | TEST_RESET))) 1033 return 0; 1034 1035 *spin = kzalloc(sizeof(**spin), GFP_KERNEL); 1036 if (!*spin) 1037 return -ENOMEM; 1038 1039 ret = igt_spinner_init(*spin, ce->engine->gt); 1040 if (ret) 1041 goto err_free; 1042 1043 rq = igt_spinner_create_request(*spin, ce, MI_NOOP); 1044 if (IS_ERR(rq)) { 1045 ret = PTR_ERR(rq); 1046 goto err_fini; 1047 } 1048 1049 i915_request_add(rq); 1050 1051 if (!igt_wait_for_spinner(*spin, rq)) { 1052 pr_err("%s: Spinner failed to start!\n", name); 1053 ret = -ETIMEDOUT; 1054 goto err_end; 1055 } 1056 1057 return 0; 1058 1059 err_end: 1060 igt_spinner_end(*spin); 1061 err_fini: 1062 igt_spinner_fini(*spin); 1063 err_free: 1064 kfree(fetch_and_zero(spin)); 1065 return ret; 1066 } 1067 1068 static int 1069 __read_slice_count(struct intel_context *ce, 1070 struct drm_i915_gem_object *obj, 1071 struct igt_spinner *spin, 1072 u32 *rpcs) 1073 { 1074 struct i915_request *rq = NULL; 1075 u32 s_mask, s_shift; 1076 unsigned int cnt; 1077 u32 *buf, val; 1078 long ret; 1079 1080 ret = emit_rpcs_query(obj, ce, &rq); 1081 if (ret) 1082 return ret; 1083 1084 if (spin) 1085 igt_spinner_end(spin); 1086 1087 ret = i915_request_wait(rq, 0, MAX_SCHEDULE_TIMEOUT); 1088 i915_request_put(rq); 1089 if (ret < 0) 1090 return ret; 1091 1092 buf = i915_gem_object_pin_map(obj, I915_MAP_WB); 1093 if (IS_ERR(buf)) { 1094 ret = PTR_ERR(buf); 1095 return ret; 1096 } 1097 1098 if (INTEL_GEN(ce->engine->i915) >= 11) { 1099 s_mask = GEN11_RPCS_S_CNT_MASK; 1100 s_shift = GEN11_RPCS_S_CNT_SHIFT; 1101 } else { 1102 s_mask = GEN8_RPCS_S_CNT_MASK; 1103 s_shift = GEN8_RPCS_S_CNT_SHIFT; 1104 } 1105 1106 val = *buf; 1107 cnt = (val & s_mask) >> s_shift; 1108 *rpcs = val; 1109 1110 i915_gem_object_unpin_map(obj); 1111 1112 return cnt; 1113 } 1114 1115 static int 1116 __check_rpcs(const char *name, u32 rpcs, int slices, unsigned int expected, 1117 const char *prefix, const char *suffix) 1118 { 1119 if (slices == expected) 1120 return 0; 1121 1122 if (slices < 0) { 1123 pr_err("%s: %s read slice count failed with %d%s\n", 1124 name, prefix, slices, suffix); 1125 return slices; 1126 } 1127 1128 pr_err("%s: %s slice count %d is not %u%s\n", 1129 name, prefix, slices, expected, suffix); 1130 1131 pr_info("RPCS=0x%x; %u%sx%u%s\n", 1132 rpcs, slices, 1133 (rpcs & GEN8_RPCS_S_CNT_ENABLE) ? "*" : "", 1134 (rpcs & GEN8_RPCS_SS_CNT_MASK) >> GEN8_RPCS_SS_CNT_SHIFT, 1135 (rpcs & GEN8_RPCS_SS_CNT_ENABLE) ? "*" : ""); 1136 1137 return -EINVAL; 1138 } 1139 1140 static int 1141 __sseu_finish(const char *name, 1142 unsigned int flags, 1143 struct intel_context *ce, 1144 struct drm_i915_gem_object *obj, 1145 unsigned int expected, 1146 struct igt_spinner *spin) 1147 { 1148 unsigned int slices = hweight32(ce->engine->sseu.slice_mask); 1149 u32 rpcs = 0; 1150 int ret = 0; 1151 1152 if (flags & TEST_RESET) { 1153 ret = intel_engine_reset(ce->engine, "sseu"); 1154 if (ret) 1155 goto out; 1156 } 1157 1158 ret = __read_slice_count(ce, obj, 1159 flags & TEST_RESET ? NULL : spin, &rpcs); 1160 ret = __check_rpcs(name, rpcs, ret, expected, "Context", "!"); 1161 if (ret) 1162 goto out; 1163 1164 ret = __read_slice_count(ce->engine->kernel_context, obj, NULL, &rpcs); 1165 ret = __check_rpcs(name, rpcs, ret, slices, "Kernel context", "!"); 1166 1167 out: 1168 if (spin) 1169 igt_spinner_end(spin); 1170 1171 if ((flags & TEST_IDLE) && ret == 0) { 1172 ret = igt_flush_test(ce->engine->i915); 1173 if (ret) 1174 return ret; 1175 1176 ret = __read_slice_count(ce, obj, NULL, &rpcs); 1177 ret = __check_rpcs(name, rpcs, ret, expected, 1178 "Context", " after idle!"); 1179 } 1180 1181 return ret; 1182 } 1183 1184 static int 1185 __sseu_test(const char *name, 1186 unsigned int flags, 1187 struct intel_context *ce, 1188 struct drm_i915_gem_object *obj, 1189 struct intel_sseu sseu) 1190 { 1191 struct igt_spinner *spin = NULL; 1192 int ret; 1193 1194 intel_engine_pm_get(ce->engine); 1195 1196 ret = __sseu_prepare(name, flags, ce, &spin); 1197 if (ret) 1198 goto out_pm; 1199 1200 ret = intel_context_reconfigure_sseu(ce, sseu); 1201 if (ret) 1202 goto out_spin; 1203 1204 ret = __sseu_finish(name, flags, ce, obj, 1205 hweight32(sseu.slice_mask), spin); 1206 1207 out_spin: 1208 if (spin) { 1209 igt_spinner_end(spin); 1210 igt_spinner_fini(spin); 1211 kfree(spin); 1212 } 1213 out_pm: 1214 intel_engine_pm_put(ce->engine); 1215 return ret; 1216 } 1217 1218 static int 1219 __igt_ctx_sseu(struct drm_i915_private *i915, 1220 const char *name, 1221 unsigned int flags) 1222 { 1223 struct drm_i915_gem_object *obj; 1224 int inst = 0; 1225 int ret = 0; 1226 1227 if (INTEL_GEN(i915) < 9 || !RUNTIME_INFO(i915)->sseu.has_slice_pg) 1228 return 0; 1229 1230 if (flags & TEST_RESET) 1231 igt_global_reset_lock(&i915->gt); 1232 1233 obj = i915_gem_object_create_internal(i915, PAGE_SIZE); 1234 if (IS_ERR(obj)) { 1235 ret = PTR_ERR(obj); 1236 goto out_unlock; 1237 } 1238 1239 do { 1240 struct intel_engine_cs *engine; 1241 struct intel_context *ce; 1242 struct intel_sseu pg_sseu; 1243 1244 engine = intel_engine_lookup_user(i915, 1245 I915_ENGINE_CLASS_RENDER, 1246 inst++); 1247 if (!engine) 1248 break; 1249 1250 if (hweight32(engine->sseu.slice_mask) < 2) 1251 continue; 1252 1253 /* 1254 * Gen11 VME friendly power-gated configuration with 1255 * half enabled sub-slices. 1256 */ 1257 pg_sseu = engine->sseu; 1258 pg_sseu.slice_mask = 1; 1259 pg_sseu.subslice_mask = 1260 ~(~0 << (hweight32(engine->sseu.subslice_mask) / 2)); 1261 1262 pr_info("%s: SSEU subtest '%s', flags=%x, def_slices=%u, pg_slices=%u\n", 1263 engine->name, name, flags, 1264 hweight32(engine->sseu.slice_mask), 1265 hweight32(pg_sseu.slice_mask)); 1266 1267 ce = intel_context_create(engine); 1268 if (IS_ERR(ce)) { 1269 ret = PTR_ERR(ce); 1270 goto out_put; 1271 } 1272 1273 ret = intel_context_pin(ce); 1274 if (ret) 1275 goto out_ce; 1276 1277 /* First set the default mask. */ 1278 ret = __sseu_test(name, flags, ce, obj, engine->sseu); 1279 if (ret) 1280 goto out_unpin; 1281 1282 /* Then set a power-gated configuration. */ 1283 ret = __sseu_test(name, flags, ce, obj, pg_sseu); 1284 if (ret) 1285 goto out_unpin; 1286 1287 /* Back to defaults. */ 1288 ret = __sseu_test(name, flags, ce, obj, engine->sseu); 1289 if (ret) 1290 goto out_unpin; 1291 1292 /* One last power-gated configuration for the road. */ 1293 ret = __sseu_test(name, flags, ce, obj, pg_sseu); 1294 if (ret) 1295 goto out_unpin; 1296 1297 out_unpin: 1298 intel_context_unpin(ce); 1299 out_ce: 1300 intel_context_put(ce); 1301 } while (!ret); 1302 1303 if (igt_flush_test(i915)) 1304 ret = -EIO; 1305 1306 out_put: 1307 i915_gem_object_put(obj); 1308 1309 out_unlock: 1310 if (flags & TEST_RESET) 1311 igt_global_reset_unlock(&i915->gt); 1312 1313 if (ret) 1314 pr_err("%s: Failed with %d!\n", name, ret); 1315 1316 return ret; 1317 } 1318 1319 static int igt_ctx_sseu(void *arg) 1320 { 1321 struct { 1322 const char *name; 1323 unsigned int flags; 1324 } *phase, phases[] = { 1325 { .name = "basic", .flags = 0 }, 1326 { .name = "idle", .flags = TEST_IDLE }, 1327 { .name = "busy", .flags = TEST_BUSY }, 1328 { .name = "busy-reset", .flags = TEST_BUSY | TEST_RESET }, 1329 { .name = "busy-idle", .flags = TEST_BUSY | TEST_IDLE }, 1330 { .name = "reset-idle", .flags = TEST_RESET | TEST_IDLE }, 1331 }; 1332 unsigned int i; 1333 int ret = 0; 1334 1335 for (i = 0, phase = phases; ret == 0 && i < ARRAY_SIZE(phases); 1336 i++, phase++) 1337 ret = __igt_ctx_sseu(arg, phase->name, phase->flags); 1338 1339 return ret; 1340 } 1341 1342 static int igt_ctx_readonly(void *arg) 1343 { 1344 struct drm_i915_private *i915 = arg; 1345 unsigned long idx, ndwords, dw, num_engines; 1346 struct drm_i915_gem_object *obj = NULL; 1347 struct i915_request *tq[5] = {}; 1348 struct i915_gem_engines_iter it; 1349 struct i915_address_space *vm; 1350 struct i915_gem_context *ctx; 1351 struct intel_context *ce; 1352 struct igt_live_test t; 1353 I915_RND_STATE(prng); 1354 IGT_TIMEOUT(end_time); 1355 LIST_HEAD(objects); 1356 struct file *file; 1357 int err = -ENODEV; 1358 1359 /* 1360 * Create a few read-only objects (with the occasional writable object) 1361 * and try to write into these object checking that the GPU discards 1362 * any write to a read-only object. 1363 */ 1364 1365 file = mock_file(i915); 1366 if (IS_ERR(file)) 1367 return PTR_ERR(file); 1368 1369 err = igt_live_test_begin(&t, i915, __func__, ""); 1370 if (err) 1371 goto out_file; 1372 1373 ctx = live_context(i915, file); 1374 if (IS_ERR(ctx)) { 1375 err = PTR_ERR(ctx); 1376 goto out_file; 1377 } 1378 1379 vm = ctx_vm(ctx) ?: &i915->ggtt.alias->vm; 1380 if (!vm || !vm->has_read_only) { 1381 err = 0; 1382 goto out_file; 1383 } 1384 1385 num_engines = 0; 1386 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) 1387 if (intel_engine_can_store_dword(ce->engine)) 1388 num_engines++; 1389 i915_gem_context_unlock_engines(ctx); 1390 1391 ndwords = 0; 1392 dw = 0; 1393 while (!time_after(jiffies, end_time)) { 1394 for_each_gem_engine(ce, 1395 i915_gem_context_lock_engines(ctx), it) { 1396 if (!intel_engine_can_store_dword(ce->engine)) 1397 continue; 1398 1399 if (!obj) { 1400 obj = create_test_object(ce->vm, file, &objects); 1401 if (IS_ERR(obj)) { 1402 err = PTR_ERR(obj); 1403 i915_gem_context_unlock_engines(ctx); 1404 goto out_file; 1405 } 1406 1407 if (prandom_u32_state(&prng) & 1) 1408 i915_gem_object_set_readonly(obj); 1409 } 1410 1411 err = gpu_fill(ce, obj, dw); 1412 if (err) { 1413 pr_err("Failed to fill dword %lu [%lu/%lu] with gpu (%s) [full-ppgtt? %s], err=%d\n", 1414 ndwords, dw, max_dwords(obj), 1415 ce->engine->name, 1416 yesno(!!ctx_vm(ctx)), 1417 err); 1418 i915_gem_context_unlock_engines(ctx); 1419 goto out_file; 1420 } 1421 1422 err = throttle(ce, tq, ARRAY_SIZE(tq)); 1423 if (err) { 1424 i915_gem_context_unlock_engines(ctx); 1425 goto out_file; 1426 } 1427 1428 if (++dw == max_dwords(obj)) { 1429 obj = NULL; 1430 dw = 0; 1431 } 1432 ndwords++; 1433 } 1434 i915_gem_context_unlock_engines(ctx); 1435 } 1436 pr_info("Submitted %lu dwords (across %lu engines)\n", 1437 ndwords, num_engines); 1438 1439 dw = 0; 1440 idx = 0; 1441 list_for_each_entry(obj, &objects, st_link) { 1442 unsigned int rem = 1443 min_t(unsigned int, ndwords - dw, max_dwords(obj)); 1444 unsigned int num_writes; 1445 1446 num_writes = rem; 1447 if (i915_gem_object_is_readonly(obj)) 1448 num_writes = 0; 1449 1450 err = cpu_check(obj, idx++, num_writes); 1451 if (err) 1452 break; 1453 1454 dw += rem; 1455 } 1456 1457 out_file: 1458 throttle_release(tq, ARRAY_SIZE(tq)); 1459 if (igt_live_test_end(&t)) 1460 err = -EIO; 1461 1462 fput(file); 1463 return err; 1464 } 1465 1466 static int check_scratch(struct i915_address_space *vm, u64 offset) 1467 { 1468 struct drm_mm_node *node = 1469 __drm_mm_interval_first(&vm->mm, 1470 offset, offset + sizeof(u32) - 1); 1471 if (!node || node->start > offset) 1472 return 0; 1473 1474 GEM_BUG_ON(offset >= node->start + node->size); 1475 1476 pr_err("Target offset 0x%08x_%08x overlaps with a node in the mm!\n", 1477 upper_32_bits(offset), lower_32_bits(offset)); 1478 return -EINVAL; 1479 } 1480 1481 static int write_to_scratch(struct i915_gem_context *ctx, 1482 struct intel_engine_cs *engine, 1483 u64 offset, u32 value) 1484 { 1485 struct drm_i915_private *i915 = ctx->i915; 1486 struct drm_i915_gem_object *obj; 1487 struct i915_address_space *vm; 1488 struct i915_request *rq; 1489 struct i915_vma *vma; 1490 u32 *cmd; 1491 int err; 1492 1493 GEM_BUG_ON(offset < I915_GTT_PAGE_SIZE); 1494 1495 obj = i915_gem_object_create_internal(i915, PAGE_SIZE); 1496 if (IS_ERR(obj)) 1497 return PTR_ERR(obj); 1498 1499 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB); 1500 if (IS_ERR(cmd)) { 1501 err = PTR_ERR(cmd); 1502 goto out; 1503 } 1504 1505 *cmd++ = MI_STORE_DWORD_IMM_GEN4; 1506 if (INTEL_GEN(i915) >= 8) { 1507 *cmd++ = lower_32_bits(offset); 1508 *cmd++ = upper_32_bits(offset); 1509 } else { 1510 *cmd++ = 0; 1511 *cmd++ = offset; 1512 } 1513 *cmd++ = value; 1514 *cmd = MI_BATCH_BUFFER_END; 1515 __i915_gem_object_flush_map(obj, 0, 64); 1516 i915_gem_object_unpin_map(obj); 1517 1518 intel_gt_chipset_flush(engine->gt); 1519 1520 vm = i915_gem_context_get_vm_rcu(ctx); 1521 vma = i915_vma_instance(obj, vm, NULL); 1522 if (IS_ERR(vma)) { 1523 err = PTR_ERR(vma); 1524 goto out_vm; 1525 } 1526 1527 err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_OFFSET_FIXED); 1528 if (err) 1529 goto out_vm; 1530 1531 err = check_scratch(vm, offset); 1532 if (err) 1533 goto err_unpin; 1534 1535 rq = igt_request_alloc(ctx, engine); 1536 if (IS_ERR(rq)) { 1537 err = PTR_ERR(rq); 1538 goto err_unpin; 1539 } 1540 1541 err = engine->emit_bb_start(rq, vma->node.start, vma->node.size, 0); 1542 if (err) 1543 goto err_request; 1544 1545 i915_vma_lock(vma); 1546 err = i915_request_await_object(rq, vma->obj, false); 1547 if (err == 0) 1548 err = i915_vma_move_to_active(vma, rq, 0); 1549 i915_vma_unlock(vma); 1550 if (err) 1551 goto skip_request; 1552 1553 i915_vma_unpin(vma); 1554 1555 i915_request_add(rq); 1556 1557 goto out_vm; 1558 skip_request: 1559 i915_request_skip(rq, err); 1560 err_request: 1561 i915_request_add(rq); 1562 err_unpin: 1563 i915_vma_unpin(vma); 1564 out_vm: 1565 i915_vm_put(vm); 1566 out: 1567 i915_gem_object_put(obj); 1568 return err; 1569 } 1570 1571 static int read_from_scratch(struct i915_gem_context *ctx, 1572 struct intel_engine_cs *engine, 1573 u64 offset, u32 *value) 1574 { 1575 struct drm_i915_private *i915 = ctx->i915; 1576 struct drm_i915_gem_object *obj; 1577 struct i915_address_space *vm; 1578 const u32 RCS_GPR0 = 0x2600; /* not all engines have their own GPR! */ 1579 const u32 result = 0x100; 1580 struct i915_request *rq; 1581 struct i915_vma *vma; 1582 u32 *cmd; 1583 int err; 1584 1585 GEM_BUG_ON(offset < I915_GTT_PAGE_SIZE); 1586 1587 obj = i915_gem_object_create_internal(i915, PAGE_SIZE); 1588 if (IS_ERR(obj)) 1589 return PTR_ERR(obj); 1590 1591 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB); 1592 if (IS_ERR(cmd)) { 1593 err = PTR_ERR(cmd); 1594 goto out; 1595 } 1596 1597 memset(cmd, POISON_INUSE, PAGE_SIZE); 1598 if (INTEL_GEN(i915) >= 8) { 1599 *cmd++ = MI_LOAD_REGISTER_MEM_GEN8; 1600 *cmd++ = RCS_GPR0; 1601 *cmd++ = lower_32_bits(offset); 1602 *cmd++ = upper_32_bits(offset); 1603 *cmd++ = MI_STORE_REGISTER_MEM_GEN8; 1604 *cmd++ = RCS_GPR0; 1605 *cmd++ = result; 1606 *cmd++ = 0; 1607 } else { 1608 *cmd++ = MI_LOAD_REGISTER_MEM; 1609 *cmd++ = RCS_GPR0; 1610 *cmd++ = offset; 1611 *cmd++ = MI_STORE_REGISTER_MEM; 1612 *cmd++ = RCS_GPR0; 1613 *cmd++ = result; 1614 } 1615 *cmd = MI_BATCH_BUFFER_END; 1616 1617 i915_gem_object_flush_map(obj); 1618 i915_gem_object_unpin_map(obj); 1619 1620 intel_gt_chipset_flush(engine->gt); 1621 1622 vm = i915_gem_context_get_vm_rcu(ctx); 1623 vma = i915_vma_instance(obj, vm, NULL); 1624 if (IS_ERR(vma)) { 1625 err = PTR_ERR(vma); 1626 goto out_vm; 1627 } 1628 1629 err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_OFFSET_FIXED); 1630 if (err) 1631 goto out_vm; 1632 1633 err = check_scratch(vm, offset); 1634 if (err) 1635 goto err_unpin; 1636 1637 rq = igt_request_alloc(ctx, engine); 1638 if (IS_ERR(rq)) { 1639 err = PTR_ERR(rq); 1640 goto err_unpin; 1641 } 1642 1643 err = engine->emit_bb_start(rq, vma->node.start, vma->node.size, 0); 1644 if (err) 1645 goto err_request; 1646 1647 i915_vma_lock(vma); 1648 err = i915_request_await_object(rq, vma->obj, true); 1649 if (err == 0) 1650 err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE); 1651 i915_vma_unlock(vma); 1652 if (err) 1653 goto skip_request; 1654 1655 i915_vma_unpin(vma); 1656 i915_vma_close(vma); 1657 1658 i915_request_add(rq); 1659 1660 i915_gem_object_lock(obj); 1661 err = i915_gem_object_set_to_cpu_domain(obj, false); 1662 i915_gem_object_unlock(obj); 1663 if (err) 1664 goto out_vm; 1665 1666 cmd = i915_gem_object_pin_map(obj, I915_MAP_WB); 1667 if (IS_ERR(cmd)) { 1668 err = PTR_ERR(cmd); 1669 goto out_vm; 1670 } 1671 1672 *value = cmd[result / sizeof(*cmd)]; 1673 i915_gem_object_unpin_map(obj); 1674 1675 goto out_vm; 1676 skip_request: 1677 i915_request_skip(rq, err); 1678 err_request: 1679 i915_request_add(rq); 1680 err_unpin: 1681 i915_vma_unpin(vma); 1682 out_vm: 1683 i915_vm_put(vm); 1684 out: 1685 i915_gem_object_put(obj); 1686 return err; 1687 } 1688 1689 static int igt_vm_isolation(void *arg) 1690 { 1691 struct drm_i915_private *i915 = arg; 1692 struct i915_gem_context *ctx_a, *ctx_b; 1693 unsigned long num_engines, count; 1694 struct intel_engine_cs *engine; 1695 struct igt_live_test t; 1696 I915_RND_STATE(prng); 1697 struct file *file; 1698 u64 vm_total; 1699 int err; 1700 1701 if (INTEL_GEN(i915) < 7) 1702 return 0; 1703 1704 /* 1705 * The simple goal here is that a write into one context is not 1706 * observed in a second (separate page tables and scratch). 1707 */ 1708 1709 file = mock_file(i915); 1710 if (IS_ERR(file)) 1711 return PTR_ERR(file); 1712 1713 err = igt_live_test_begin(&t, i915, __func__, ""); 1714 if (err) 1715 goto out_file; 1716 1717 ctx_a = live_context(i915, file); 1718 if (IS_ERR(ctx_a)) { 1719 err = PTR_ERR(ctx_a); 1720 goto out_file; 1721 } 1722 1723 ctx_b = live_context(i915, file); 1724 if (IS_ERR(ctx_b)) { 1725 err = PTR_ERR(ctx_b); 1726 goto out_file; 1727 } 1728 1729 /* We can only test vm isolation, if the vm are distinct */ 1730 if (ctx_vm(ctx_a) == ctx_vm(ctx_b)) 1731 goto out_file; 1732 1733 vm_total = ctx_vm(ctx_a)->total; 1734 GEM_BUG_ON(ctx_vm(ctx_b)->total != vm_total); 1735 vm_total -= I915_GTT_PAGE_SIZE; 1736 1737 count = 0; 1738 num_engines = 0; 1739 for_each_uabi_engine(engine, i915) { 1740 IGT_TIMEOUT(end_time); 1741 unsigned long this = 0; 1742 1743 if (!intel_engine_can_store_dword(engine)) 1744 continue; 1745 1746 while (!__igt_timeout(end_time, NULL)) { 1747 u32 value = 0xc5c5c5c5; 1748 u64 offset; 1749 1750 div64_u64_rem(i915_prandom_u64_state(&prng), 1751 vm_total, &offset); 1752 offset = round_down(offset, alignof_dword); 1753 offset += I915_GTT_PAGE_SIZE; 1754 1755 err = write_to_scratch(ctx_a, engine, 1756 offset, 0xdeadbeef); 1757 if (err == 0) 1758 err = read_from_scratch(ctx_b, engine, 1759 offset, &value); 1760 if (err) 1761 goto out_file; 1762 1763 if (value) { 1764 pr_err("%s: Read %08x from scratch (offset 0x%08x_%08x), after %lu reads!\n", 1765 engine->name, value, 1766 upper_32_bits(offset), 1767 lower_32_bits(offset), 1768 this); 1769 err = -EINVAL; 1770 goto out_file; 1771 } 1772 1773 this++; 1774 } 1775 count += this; 1776 num_engines++; 1777 } 1778 pr_info("Checked %lu scratch offsets across %lu engines\n", 1779 count, num_engines); 1780 1781 out_file: 1782 if (igt_live_test_end(&t)) 1783 err = -EIO; 1784 fput(file); 1785 return err; 1786 } 1787 1788 static bool skip_unused_engines(struct intel_context *ce, void *data) 1789 { 1790 return !ce->state; 1791 } 1792 1793 static void mock_barrier_task(void *data) 1794 { 1795 unsigned int *counter = data; 1796 1797 ++*counter; 1798 } 1799 1800 static int mock_context_barrier(void *arg) 1801 { 1802 #undef pr_fmt 1803 #define pr_fmt(x) "context_barrier_task():" # x 1804 struct drm_i915_private *i915 = arg; 1805 struct i915_gem_context *ctx; 1806 struct i915_request *rq; 1807 unsigned int counter; 1808 int err; 1809 1810 /* 1811 * The context barrier provides us with a callback after it emits 1812 * a request; useful for retiring old state after loading new. 1813 */ 1814 1815 ctx = mock_context(i915, "mock"); 1816 if (!ctx) 1817 return -ENOMEM; 1818 1819 counter = 0; 1820 err = context_barrier_task(ctx, 0, 1821 NULL, NULL, mock_barrier_task, &counter); 1822 if (err) { 1823 pr_err("Failed at line %d, err=%d\n", __LINE__, err); 1824 goto out; 1825 } 1826 if (counter == 0) { 1827 pr_err("Did not retire immediately with 0 engines\n"); 1828 err = -EINVAL; 1829 goto out; 1830 } 1831 1832 counter = 0; 1833 err = context_barrier_task(ctx, ALL_ENGINES, 1834 skip_unused_engines, 1835 NULL, 1836 mock_barrier_task, 1837 &counter); 1838 if (err) { 1839 pr_err("Failed at line %d, err=%d\n", __LINE__, err); 1840 goto out; 1841 } 1842 if (counter == 0) { 1843 pr_err("Did not retire immediately for all unused engines\n"); 1844 err = -EINVAL; 1845 goto out; 1846 } 1847 1848 rq = igt_request_alloc(ctx, i915->engine[RCS0]); 1849 if (IS_ERR(rq)) { 1850 pr_err("Request allocation failed!\n"); 1851 goto out; 1852 } 1853 i915_request_add(rq); 1854 1855 counter = 0; 1856 context_barrier_inject_fault = BIT(RCS0); 1857 err = context_barrier_task(ctx, ALL_ENGINES, 1858 NULL, NULL, mock_barrier_task, &counter); 1859 context_barrier_inject_fault = 0; 1860 if (err == -ENXIO) 1861 err = 0; 1862 else 1863 pr_err("Did not hit fault injection!\n"); 1864 if (counter != 0) { 1865 pr_err("Invoked callback on error!\n"); 1866 err = -EIO; 1867 } 1868 if (err) 1869 goto out; 1870 1871 counter = 0; 1872 err = context_barrier_task(ctx, ALL_ENGINES, 1873 skip_unused_engines, 1874 NULL, 1875 mock_barrier_task, 1876 &counter); 1877 if (err) { 1878 pr_err("Failed at line %d, err=%d\n", __LINE__, err); 1879 goto out; 1880 } 1881 mock_device_flush(i915); 1882 if (counter == 0) { 1883 pr_err("Did not retire on each active engines\n"); 1884 err = -EINVAL; 1885 goto out; 1886 } 1887 1888 out: 1889 mock_context_close(ctx); 1890 return err; 1891 #undef pr_fmt 1892 #define pr_fmt(x) x 1893 } 1894 1895 int i915_gem_context_mock_selftests(void) 1896 { 1897 static const struct i915_subtest tests[] = { 1898 SUBTEST(mock_context_barrier), 1899 }; 1900 struct drm_i915_private *i915; 1901 int err; 1902 1903 i915 = mock_gem_device(); 1904 if (!i915) 1905 return -ENOMEM; 1906 1907 err = i915_subtests(tests, i915); 1908 1909 drm_dev_put(&i915->drm); 1910 return err; 1911 } 1912 1913 int i915_gem_context_live_selftests(struct drm_i915_private *i915) 1914 { 1915 static const struct i915_subtest tests[] = { 1916 SUBTEST(live_nop_switch), 1917 SUBTEST(live_parallel_switch), 1918 SUBTEST(igt_ctx_exec), 1919 SUBTEST(igt_ctx_readonly), 1920 SUBTEST(igt_ctx_sseu), 1921 SUBTEST(igt_shared_ctx_exec), 1922 SUBTEST(igt_vm_isolation), 1923 }; 1924 1925 if (intel_gt_is_wedged(&i915->gt)) 1926 return 0; 1927 1928 return i915_live_subtests(tests, i915); 1929 } 1930