1 /* 2 * SPDX-License-Identifier: MIT 3 * 4 * Copyright © 2011-2012 Intel Corporation 5 */ 6 7 /* 8 * This file implements HW context support. On gen5+ a HW context consists of an 9 * opaque GPU object which is referenced at times of context saves and restores. 10 * With RC6 enabled, the context is also referenced as the GPU enters and exists 11 * from RC6 (GPU has it's own internal power context, except on gen5). Though 12 * something like a context does exist for the media ring, the code only 13 * supports contexts for the render ring. 14 * 15 * In software, there is a distinction between contexts created by the user, 16 * and the default HW context. The default HW context is used by GPU clients 17 * that do not request setup of their own hardware context. The default 18 * context's state is never restored to help prevent programming errors. This 19 * would happen if a client ran and piggy-backed off another clients GPU state. 20 * The default context only exists to give the GPU some offset to load as the 21 * current to invoke a save of the context we actually care about. In fact, the 22 * code could likely be constructed, albeit in a more complicated fashion, to 23 * never use the default context, though that limits the driver's ability to 24 * swap out, and/or destroy other contexts. 25 * 26 * All other contexts are created as a request by the GPU client. These contexts 27 * store GPU state, and thus allow GPU clients to not re-emit state (and 28 * potentially query certain state) at any time. The kernel driver makes 29 * certain that the appropriate commands are inserted. 30 * 31 * The context life cycle is semi-complicated in that context BOs may live 32 * longer than the context itself because of the way the hardware, and object 33 * tracking works. Below is a very crude representation of the state machine 34 * describing the context life. 35 * refcount pincount active 36 * S0: initial state 0 0 0 37 * S1: context created 1 0 0 38 * S2: context is currently running 2 1 X 39 * S3: GPU referenced, but not current 2 0 1 40 * S4: context is current, but destroyed 1 1 0 41 * S5: like S3, but destroyed 1 0 1 42 * 43 * The most common (but not all) transitions: 44 * S0->S1: client creates a context 45 * S1->S2: client submits execbuf with context 46 * S2->S3: other clients submits execbuf with context 47 * S3->S1: context object was retired 48 * S3->S2: clients submits another execbuf 49 * S2->S4: context destroy called with current context 50 * S3->S5->S0: destroy path 51 * S4->S5->S0: destroy path on current context 52 * 53 * There are two confusing terms used above: 54 * The "current context" means the context which is currently running on the 55 * GPU. The GPU has loaded its state already and has stored away the gtt 56 * offset of the BO. The GPU is not actively referencing the data at this 57 * offset, but it will on the next context switch. The only way to avoid this 58 * is to do a GPU reset. 59 * 60 * An "active context' is one which was previously the "current context" and is 61 * on the active list waiting for the next context switch to occur. Until this 62 * happens, the object must remain at the same gtt offset. It is therefore 63 * possible to destroy a context, but it is still active. 64 * 65 */ 66 67 #include <linux/highmem.h> 68 #include <linux/log2.h> 69 #include <linux/nospec.h> 70 71 #include <drm/drm_cache.h> 72 #include <drm/drm_syncobj.h> 73 74 #include "gt/gen6_ppgtt.h" 75 #include "gt/intel_context.h" 76 #include "gt/intel_context_param.h" 77 #include "gt/intel_engine_heartbeat.h" 78 #include "gt/intel_engine_user.h" 79 #include "gt/intel_gpu_commands.h" 80 #include "gt/intel_ring.h" 81 82 #include "pxp/intel_pxp.h" 83 84 #include "i915_file_private.h" 85 #include "i915_gem_context.h" 86 #include "i915_trace.h" 87 #include "i915_user_extensions.h" 88 89 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1 90 91 static struct kmem_cache *slab_luts; 92 93 struct i915_lut_handle *i915_lut_handle_alloc(void) 94 { 95 return kmem_cache_alloc(slab_luts, GFP_KERNEL); 96 } 97 98 void i915_lut_handle_free(struct i915_lut_handle *lut) 99 { 100 return kmem_cache_free(slab_luts, lut); 101 } 102 103 static void lut_close(struct i915_gem_context *ctx) 104 { 105 struct radix_tree_iter iter; 106 void __rcu **slot; 107 108 mutex_lock(&ctx->lut_mutex); 109 rcu_read_lock(); 110 radix_tree_for_each_slot(slot, &ctx->handles_vma, &iter, 0) { 111 struct i915_vma *vma = rcu_dereference_raw(*slot); 112 struct drm_i915_gem_object *obj = vma->obj; 113 struct i915_lut_handle *lut; 114 115 if (!kref_get_unless_zero(&obj->base.refcount)) 116 continue; 117 118 spin_lock(&obj->lut_lock); 119 list_for_each_entry(lut, &obj->lut_list, obj_link) { 120 if (lut->ctx != ctx) 121 continue; 122 123 if (lut->handle != iter.index) 124 continue; 125 126 list_del(&lut->obj_link); 127 break; 128 } 129 spin_unlock(&obj->lut_lock); 130 131 if (&lut->obj_link != &obj->lut_list) { 132 i915_lut_handle_free(lut); 133 radix_tree_iter_delete(&ctx->handles_vma, &iter, slot); 134 i915_vma_close(vma); 135 i915_gem_object_put(obj); 136 } 137 138 i915_gem_object_put(obj); 139 } 140 rcu_read_unlock(); 141 mutex_unlock(&ctx->lut_mutex); 142 } 143 144 static struct intel_context * 145 lookup_user_engine(struct i915_gem_context *ctx, 146 unsigned long flags, 147 const struct i915_engine_class_instance *ci) 148 #define LOOKUP_USER_INDEX BIT(0) 149 { 150 int idx; 151 152 if (!!(flags & LOOKUP_USER_INDEX) != i915_gem_context_user_engines(ctx)) 153 return ERR_PTR(-EINVAL); 154 155 if (!i915_gem_context_user_engines(ctx)) { 156 struct intel_engine_cs *engine; 157 158 engine = intel_engine_lookup_user(ctx->i915, 159 ci->engine_class, 160 ci->engine_instance); 161 if (!engine) 162 return ERR_PTR(-EINVAL); 163 164 idx = engine->legacy_idx; 165 } else { 166 idx = ci->engine_instance; 167 } 168 169 return i915_gem_context_get_engine(ctx, idx); 170 } 171 172 static int validate_priority(struct drm_i915_private *i915, 173 const struct drm_i915_gem_context_param *args) 174 { 175 s64 priority = args->value; 176 177 if (args->size) 178 return -EINVAL; 179 180 if (!(i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY)) 181 return -ENODEV; 182 183 if (priority > I915_CONTEXT_MAX_USER_PRIORITY || 184 priority < I915_CONTEXT_MIN_USER_PRIORITY) 185 return -EINVAL; 186 187 if (priority > I915_CONTEXT_DEFAULT_PRIORITY && 188 !capable(CAP_SYS_NICE)) 189 return -EPERM; 190 191 return 0; 192 } 193 194 static void proto_context_close(struct drm_i915_private *i915, 195 struct i915_gem_proto_context *pc) 196 { 197 int i; 198 199 if (pc->pxp_wakeref) 200 intel_runtime_pm_put(&i915->runtime_pm, pc->pxp_wakeref); 201 if (pc->vm) 202 i915_vm_put(pc->vm); 203 if (pc->user_engines) { 204 for (i = 0; i < pc->num_user_engines; i++) 205 kfree(pc->user_engines[i].siblings); 206 kfree(pc->user_engines); 207 } 208 kfree(pc); 209 } 210 211 static int proto_context_set_persistence(struct drm_i915_private *i915, 212 struct i915_gem_proto_context *pc, 213 bool persist) 214 { 215 if (persist) { 216 /* 217 * Only contexts that are short-lived [that will expire or be 218 * reset] are allowed to survive past termination. We require 219 * hangcheck to ensure that the persistent requests are healthy. 220 */ 221 if (!i915->params.enable_hangcheck) 222 return -EINVAL; 223 224 pc->user_flags |= BIT(UCONTEXT_PERSISTENCE); 225 } else { 226 /* To cancel a context we use "preempt-to-idle" */ 227 if (!(i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION)) 228 return -ENODEV; 229 230 /* 231 * If the cancel fails, we then need to reset, cleanly! 232 * 233 * If the per-engine reset fails, all hope is lost! We resort 234 * to a full GPU reset in that unlikely case, but realistically 235 * if the engine could not reset, the full reset does not fare 236 * much better. The damage has been done. 237 * 238 * However, if we cannot reset an engine by itself, we cannot 239 * cleanup a hanging persistent context without causing 240 * colateral damage, and we should not pretend we can by 241 * exposing the interface. 242 */ 243 if (!intel_has_reset_engine(to_gt(i915))) 244 return -ENODEV; 245 246 pc->user_flags &= ~BIT(UCONTEXT_PERSISTENCE); 247 } 248 249 return 0; 250 } 251 252 static int proto_context_set_protected(struct drm_i915_private *i915, 253 struct i915_gem_proto_context *pc, 254 bool protected) 255 { 256 int ret = 0; 257 258 if (!protected) { 259 pc->uses_protected_content = false; 260 } else if (!intel_pxp_is_enabled(i915->pxp)) { 261 ret = -ENODEV; 262 } else if ((pc->user_flags & BIT(UCONTEXT_RECOVERABLE)) || 263 !(pc->user_flags & BIT(UCONTEXT_BANNABLE))) { 264 ret = -EPERM; 265 } else { 266 pc->uses_protected_content = true; 267 268 /* 269 * protected context usage requires the PXP session to be up, 270 * which in turn requires the device to be active. 271 */ 272 pc->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm); 273 274 if (!intel_pxp_is_active(i915->pxp)) 275 ret = intel_pxp_start(i915->pxp); 276 } 277 278 return ret; 279 } 280 281 static struct i915_gem_proto_context * 282 proto_context_create(struct drm_i915_private *i915, unsigned int flags) 283 { 284 struct i915_gem_proto_context *pc, *err; 285 286 pc = kzalloc(sizeof(*pc), GFP_KERNEL); 287 if (!pc) 288 return ERR_PTR(-ENOMEM); 289 290 pc->num_user_engines = -1; 291 pc->user_engines = NULL; 292 pc->user_flags = BIT(UCONTEXT_BANNABLE) | 293 BIT(UCONTEXT_RECOVERABLE); 294 if (i915->params.enable_hangcheck) 295 pc->user_flags |= BIT(UCONTEXT_PERSISTENCE); 296 pc->sched.priority = I915_PRIORITY_NORMAL; 297 298 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) { 299 if (!HAS_EXECLISTS(i915)) { 300 err = ERR_PTR(-EINVAL); 301 goto proto_close; 302 } 303 pc->single_timeline = true; 304 } 305 306 return pc; 307 308 proto_close: 309 proto_context_close(i915, pc); 310 return err; 311 } 312 313 static int proto_context_register_locked(struct drm_i915_file_private *fpriv, 314 struct i915_gem_proto_context *pc, 315 u32 *id) 316 { 317 int ret; 318 void *old; 319 320 lockdep_assert_held(&fpriv->proto_context_lock); 321 322 ret = xa_alloc(&fpriv->context_xa, id, NULL, xa_limit_32b, GFP_KERNEL); 323 if (ret) 324 return ret; 325 326 old = xa_store(&fpriv->proto_context_xa, *id, pc, GFP_KERNEL); 327 if (xa_is_err(old)) { 328 xa_erase(&fpriv->context_xa, *id); 329 return xa_err(old); 330 } 331 WARN_ON(old); 332 333 return 0; 334 } 335 336 static int proto_context_register(struct drm_i915_file_private *fpriv, 337 struct i915_gem_proto_context *pc, 338 u32 *id) 339 { 340 int ret; 341 342 mutex_lock(&fpriv->proto_context_lock); 343 ret = proto_context_register_locked(fpriv, pc, id); 344 mutex_unlock(&fpriv->proto_context_lock); 345 346 return ret; 347 } 348 349 static struct i915_address_space * 350 i915_gem_vm_lookup(struct drm_i915_file_private *file_priv, u32 id) 351 { 352 struct i915_address_space *vm; 353 354 xa_lock(&file_priv->vm_xa); 355 vm = xa_load(&file_priv->vm_xa, id); 356 if (vm) 357 kref_get(&vm->ref); 358 xa_unlock(&file_priv->vm_xa); 359 360 return vm; 361 } 362 363 static int set_proto_ctx_vm(struct drm_i915_file_private *fpriv, 364 struct i915_gem_proto_context *pc, 365 const struct drm_i915_gem_context_param *args) 366 { 367 struct drm_i915_private *i915 = fpriv->i915; 368 struct i915_address_space *vm; 369 370 if (args->size) 371 return -EINVAL; 372 373 if (!HAS_FULL_PPGTT(i915)) 374 return -ENODEV; 375 376 if (upper_32_bits(args->value)) 377 return -ENOENT; 378 379 vm = i915_gem_vm_lookup(fpriv, args->value); 380 if (!vm) 381 return -ENOENT; 382 383 if (pc->vm) 384 i915_vm_put(pc->vm); 385 pc->vm = vm; 386 387 return 0; 388 } 389 390 struct set_proto_ctx_engines { 391 struct drm_i915_private *i915; 392 unsigned num_engines; 393 struct i915_gem_proto_engine *engines; 394 }; 395 396 static int 397 set_proto_ctx_engines_balance(struct i915_user_extension __user *base, 398 void *data) 399 { 400 struct i915_context_engines_load_balance __user *ext = 401 container_of_user(base, typeof(*ext), base); 402 const struct set_proto_ctx_engines *set = data; 403 struct drm_i915_private *i915 = set->i915; 404 struct intel_engine_cs **siblings; 405 u16 num_siblings, idx; 406 unsigned int n; 407 int err; 408 409 if (!HAS_EXECLISTS(i915)) 410 return -ENODEV; 411 412 if (get_user(idx, &ext->engine_index)) 413 return -EFAULT; 414 415 if (idx >= set->num_engines) { 416 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n", 417 idx, set->num_engines); 418 return -EINVAL; 419 } 420 421 idx = array_index_nospec(idx, set->num_engines); 422 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_INVALID) { 423 drm_dbg(&i915->drm, 424 "Invalid placement[%d], already occupied\n", idx); 425 return -EEXIST; 426 } 427 428 if (get_user(num_siblings, &ext->num_siblings)) 429 return -EFAULT; 430 431 err = check_user_mbz(&ext->flags); 432 if (err) 433 return err; 434 435 err = check_user_mbz(&ext->mbz64); 436 if (err) 437 return err; 438 439 if (num_siblings == 0) 440 return 0; 441 442 siblings = kmalloc_array(num_siblings, sizeof(*siblings), GFP_KERNEL); 443 if (!siblings) 444 return -ENOMEM; 445 446 for (n = 0; n < num_siblings; n++) { 447 struct i915_engine_class_instance ci; 448 449 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) { 450 err = -EFAULT; 451 goto err_siblings; 452 } 453 454 siblings[n] = intel_engine_lookup_user(i915, 455 ci.engine_class, 456 ci.engine_instance); 457 if (!siblings[n]) { 458 drm_dbg(&i915->drm, 459 "Invalid sibling[%d]: { class:%d, inst:%d }\n", 460 n, ci.engine_class, ci.engine_instance); 461 err = -EINVAL; 462 goto err_siblings; 463 } 464 } 465 466 if (num_siblings == 1) { 467 set->engines[idx].type = I915_GEM_ENGINE_TYPE_PHYSICAL; 468 set->engines[idx].engine = siblings[0]; 469 kfree(siblings); 470 } else { 471 set->engines[idx].type = I915_GEM_ENGINE_TYPE_BALANCED; 472 set->engines[idx].num_siblings = num_siblings; 473 set->engines[idx].siblings = siblings; 474 } 475 476 return 0; 477 478 err_siblings: 479 kfree(siblings); 480 481 return err; 482 } 483 484 static int 485 set_proto_ctx_engines_bond(struct i915_user_extension __user *base, void *data) 486 { 487 struct i915_context_engines_bond __user *ext = 488 container_of_user(base, typeof(*ext), base); 489 const struct set_proto_ctx_engines *set = data; 490 struct drm_i915_private *i915 = set->i915; 491 struct i915_engine_class_instance ci; 492 struct intel_engine_cs *master; 493 u16 idx, num_bonds; 494 int err, n; 495 496 if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915) && 497 !IS_ROCKETLAKE(i915) && !IS_ALDERLAKE_S(i915)) { 498 drm_dbg(&i915->drm, 499 "Bonding not supported on this platform\n"); 500 return -ENODEV; 501 } 502 503 if (get_user(idx, &ext->virtual_index)) 504 return -EFAULT; 505 506 if (idx >= set->num_engines) { 507 drm_dbg(&i915->drm, 508 "Invalid index for virtual engine: %d >= %d\n", 509 idx, set->num_engines); 510 return -EINVAL; 511 } 512 513 idx = array_index_nospec(idx, set->num_engines); 514 if (set->engines[idx].type == I915_GEM_ENGINE_TYPE_INVALID) { 515 drm_dbg(&i915->drm, "Invalid engine at %d\n", idx); 516 return -EINVAL; 517 } 518 519 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_PHYSICAL) { 520 drm_dbg(&i915->drm, 521 "Bonding with virtual engines not allowed\n"); 522 return -EINVAL; 523 } 524 525 err = check_user_mbz(&ext->flags); 526 if (err) 527 return err; 528 529 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) { 530 err = check_user_mbz(&ext->mbz64[n]); 531 if (err) 532 return err; 533 } 534 535 if (copy_from_user(&ci, &ext->master, sizeof(ci))) 536 return -EFAULT; 537 538 master = intel_engine_lookup_user(i915, 539 ci.engine_class, 540 ci.engine_instance); 541 if (!master) { 542 drm_dbg(&i915->drm, 543 "Unrecognised master engine: { class:%u, instance:%u }\n", 544 ci.engine_class, ci.engine_instance); 545 return -EINVAL; 546 } 547 548 if (intel_engine_uses_guc(master)) { 549 drm_dbg(&i915->drm, "bonding extension not supported with GuC submission"); 550 return -ENODEV; 551 } 552 553 if (get_user(num_bonds, &ext->num_bonds)) 554 return -EFAULT; 555 556 for (n = 0; n < num_bonds; n++) { 557 struct intel_engine_cs *bond; 558 559 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) 560 return -EFAULT; 561 562 bond = intel_engine_lookup_user(i915, 563 ci.engine_class, 564 ci.engine_instance); 565 if (!bond) { 566 drm_dbg(&i915->drm, 567 "Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n", 568 n, ci.engine_class, ci.engine_instance); 569 return -EINVAL; 570 } 571 } 572 573 return 0; 574 } 575 576 static int 577 set_proto_ctx_engines_parallel_submit(struct i915_user_extension __user *base, 578 void *data) 579 { 580 struct i915_context_engines_parallel_submit __user *ext = 581 container_of_user(base, typeof(*ext), base); 582 const struct set_proto_ctx_engines *set = data; 583 struct drm_i915_private *i915 = set->i915; 584 struct i915_engine_class_instance prev_engine; 585 u64 flags; 586 int err = 0, n, i, j; 587 u16 slot, width, num_siblings; 588 struct intel_engine_cs **siblings = NULL; 589 intel_engine_mask_t prev_mask; 590 591 if (get_user(slot, &ext->engine_index)) 592 return -EFAULT; 593 594 if (get_user(width, &ext->width)) 595 return -EFAULT; 596 597 if (get_user(num_siblings, &ext->num_siblings)) 598 return -EFAULT; 599 600 if (!intel_uc_uses_guc_submission(&to_gt(i915)->uc) && 601 num_siblings != 1) { 602 drm_dbg(&i915->drm, "Only 1 sibling (%d) supported in non-GuC mode\n", 603 num_siblings); 604 return -EINVAL; 605 } 606 607 if (slot >= set->num_engines) { 608 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n", 609 slot, set->num_engines); 610 return -EINVAL; 611 } 612 613 if (set->engines[slot].type != I915_GEM_ENGINE_TYPE_INVALID) { 614 drm_dbg(&i915->drm, 615 "Invalid placement[%d], already occupied\n", slot); 616 return -EINVAL; 617 } 618 619 if (get_user(flags, &ext->flags)) 620 return -EFAULT; 621 622 if (flags) { 623 drm_dbg(&i915->drm, "Unknown flags 0x%02llx", flags); 624 return -EINVAL; 625 } 626 627 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) { 628 err = check_user_mbz(&ext->mbz64[n]); 629 if (err) 630 return err; 631 } 632 633 if (width < 2) { 634 drm_dbg(&i915->drm, "Width (%d) < 2\n", width); 635 return -EINVAL; 636 } 637 638 if (num_siblings < 1) { 639 drm_dbg(&i915->drm, "Number siblings (%d) < 1\n", 640 num_siblings); 641 return -EINVAL; 642 } 643 644 siblings = kmalloc_array(num_siblings * width, 645 sizeof(*siblings), 646 GFP_KERNEL); 647 if (!siblings) 648 return -ENOMEM; 649 650 /* Create contexts / engines */ 651 for (i = 0; i < width; ++i) { 652 intel_engine_mask_t current_mask = 0; 653 654 for (j = 0; j < num_siblings; ++j) { 655 struct i915_engine_class_instance ci; 656 657 n = i * num_siblings + j; 658 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) { 659 err = -EFAULT; 660 goto out_err; 661 } 662 663 siblings[n] = 664 intel_engine_lookup_user(i915, ci.engine_class, 665 ci.engine_instance); 666 if (!siblings[n]) { 667 drm_dbg(&i915->drm, 668 "Invalid sibling[%d]: { class:%d, inst:%d }\n", 669 n, ci.engine_class, ci.engine_instance); 670 err = -EINVAL; 671 goto out_err; 672 } 673 674 /* 675 * We don't support breadcrumb handshake on these 676 * classes 677 */ 678 if (siblings[n]->class == RENDER_CLASS || 679 siblings[n]->class == COMPUTE_CLASS) { 680 err = -EINVAL; 681 goto out_err; 682 } 683 684 if (n) { 685 if (prev_engine.engine_class != 686 ci.engine_class) { 687 drm_dbg(&i915->drm, 688 "Mismatched class %d, %d\n", 689 prev_engine.engine_class, 690 ci.engine_class); 691 err = -EINVAL; 692 goto out_err; 693 } 694 } 695 696 prev_engine = ci; 697 current_mask |= siblings[n]->logical_mask; 698 } 699 700 if (i > 0) { 701 if (current_mask != prev_mask << 1) { 702 drm_dbg(&i915->drm, 703 "Non contiguous logical mask 0x%x, 0x%x\n", 704 prev_mask, current_mask); 705 err = -EINVAL; 706 goto out_err; 707 } 708 } 709 prev_mask = current_mask; 710 } 711 712 set->engines[slot].type = I915_GEM_ENGINE_TYPE_PARALLEL; 713 set->engines[slot].num_siblings = num_siblings; 714 set->engines[slot].width = width; 715 set->engines[slot].siblings = siblings; 716 717 return 0; 718 719 out_err: 720 kfree(siblings); 721 722 return err; 723 } 724 725 static const i915_user_extension_fn set_proto_ctx_engines_extensions[] = { 726 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_proto_ctx_engines_balance, 727 [I915_CONTEXT_ENGINES_EXT_BOND] = set_proto_ctx_engines_bond, 728 [I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT] = 729 set_proto_ctx_engines_parallel_submit, 730 }; 731 732 static int set_proto_ctx_engines(struct drm_i915_file_private *fpriv, 733 struct i915_gem_proto_context *pc, 734 const struct drm_i915_gem_context_param *args) 735 { 736 struct drm_i915_private *i915 = fpriv->i915; 737 struct set_proto_ctx_engines set = { .i915 = i915 }; 738 struct i915_context_param_engines __user *user = 739 u64_to_user_ptr(args->value); 740 unsigned int n; 741 u64 extensions; 742 int err; 743 744 if (pc->num_user_engines >= 0) { 745 drm_dbg(&i915->drm, "Cannot set engines twice"); 746 return -EINVAL; 747 } 748 749 if (args->size < sizeof(*user) || 750 !IS_ALIGNED(args->size - sizeof(*user), sizeof(*user->engines))) { 751 drm_dbg(&i915->drm, "Invalid size for engine array: %d\n", 752 args->size); 753 return -EINVAL; 754 } 755 756 set.num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines); 757 /* RING_MASK has no shift so we can use it directly here */ 758 if (set.num_engines > I915_EXEC_RING_MASK + 1) 759 return -EINVAL; 760 761 set.engines = kmalloc_array(set.num_engines, sizeof(*set.engines), GFP_KERNEL); 762 if (!set.engines) 763 return -ENOMEM; 764 765 for (n = 0; n < set.num_engines; n++) { 766 struct i915_engine_class_instance ci; 767 struct intel_engine_cs *engine; 768 769 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) { 770 kfree(set.engines); 771 return -EFAULT; 772 } 773 774 memset(&set.engines[n], 0, sizeof(set.engines[n])); 775 776 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID && 777 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE) 778 continue; 779 780 engine = intel_engine_lookup_user(i915, 781 ci.engine_class, 782 ci.engine_instance); 783 if (!engine) { 784 drm_dbg(&i915->drm, 785 "Invalid engine[%d]: { class:%d, instance:%d }\n", 786 n, ci.engine_class, ci.engine_instance); 787 kfree(set.engines); 788 return -ENOENT; 789 } 790 791 set.engines[n].type = I915_GEM_ENGINE_TYPE_PHYSICAL; 792 set.engines[n].engine = engine; 793 } 794 795 err = -EFAULT; 796 if (!get_user(extensions, &user->extensions)) 797 err = i915_user_extensions(u64_to_user_ptr(extensions), 798 set_proto_ctx_engines_extensions, 799 ARRAY_SIZE(set_proto_ctx_engines_extensions), 800 &set); 801 if (err) { 802 kfree(set.engines); 803 return err; 804 } 805 806 pc->num_user_engines = set.num_engines; 807 pc->user_engines = set.engines; 808 809 return 0; 810 } 811 812 static int set_proto_ctx_sseu(struct drm_i915_file_private *fpriv, 813 struct i915_gem_proto_context *pc, 814 struct drm_i915_gem_context_param *args) 815 { 816 struct drm_i915_private *i915 = fpriv->i915; 817 struct drm_i915_gem_context_param_sseu user_sseu; 818 struct intel_sseu *sseu; 819 int ret; 820 821 if (args->size < sizeof(user_sseu)) 822 return -EINVAL; 823 824 if (GRAPHICS_VER(i915) != 11) 825 return -ENODEV; 826 827 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 828 sizeof(user_sseu))) 829 return -EFAULT; 830 831 if (user_sseu.rsvd) 832 return -EINVAL; 833 834 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 835 return -EINVAL; 836 837 if (!!(user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) != (pc->num_user_engines >= 0)) 838 return -EINVAL; 839 840 if (pc->num_user_engines >= 0) { 841 int idx = user_sseu.engine.engine_instance; 842 struct i915_gem_proto_engine *pe; 843 844 if (idx >= pc->num_user_engines) 845 return -EINVAL; 846 847 pe = &pc->user_engines[idx]; 848 849 /* Only render engine supports RPCS configuration. */ 850 if (pe->engine->class != RENDER_CLASS) 851 return -EINVAL; 852 853 sseu = &pe->sseu; 854 } else { 855 /* Only render engine supports RPCS configuration. */ 856 if (user_sseu.engine.engine_class != I915_ENGINE_CLASS_RENDER) 857 return -EINVAL; 858 859 /* There is only one render engine */ 860 if (user_sseu.engine.engine_instance != 0) 861 return -EINVAL; 862 863 sseu = &pc->legacy_rcs_sseu; 864 } 865 866 ret = i915_gem_user_to_context_sseu(to_gt(i915), &user_sseu, sseu); 867 if (ret) 868 return ret; 869 870 args->size = sizeof(user_sseu); 871 872 return 0; 873 } 874 875 static int set_proto_ctx_param(struct drm_i915_file_private *fpriv, 876 struct i915_gem_proto_context *pc, 877 struct drm_i915_gem_context_param *args) 878 { 879 int ret = 0; 880 881 switch (args->param) { 882 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 883 if (args->size) 884 ret = -EINVAL; 885 else if (args->value) 886 pc->user_flags |= BIT(UCONTEXT_NO_ERROR_CAPTURE); 887 else 888 pc->user_flags &= ~BIT(UCONTEXT_NO_ERROR_CAPTURE); 889 break; 890 891 case I915_CONTEXT_PARAM_BANNABLE: 892 if (args->size) 893 ret = -EINVAL; 894 else if (!capable(CAP_SYS_ADMIN) && !args->value) 895 ret = -EPERM; 896 else if (args->value) 897 pc->user_flags |= BIT(UCONTEXT_BANNABLE); 898 else if (pc->uses_protected_content) 899 ret = -EPERM; 900 else 901 pc->user_flags &= ~BIT(UCONTEXT_BANNABLE); 902 break; 903 904 case I915_CONTEXT_PARAM_RECOVERABLE: 905 if (args->size) 906 ret = -EINVAL; 907 else if (!args->value) 908 pc->user_flags &= ~BIT(UCONTEXT_RECOVERABLE); 909 else if (pc->uses_protected_content) 910 ret = -EPERM; 911 else 912 pc->user_flags |= BIT(UCONTEXT_RECOVERABLE); 913 break; 914 915 case I915_CONTEXT_PARAM_PRIORITY: 916 ret = validate_priority(fpriv->i915, args); 917 if (!ret) 918 pc->sched.priority = args->value; 919 break; 920 921 case I915_CONTEXT_PARAM_SSEU: 922 ret = set_proto_ctx_sseu(fpriv, pc, args); 923 break; 924 925 case I915_CONTEXT_PARAM_VM: 926 ret = set_proto_ctx_vm(fpriv, pc, args); 927 break; 928 929 case I915_CONTEXT_PARAM_ENGINES: 930 ret = set_proto_ctx_engines(fpriv, pc, args); 931 break; 932 933 case I915_CONTEXT_PARAM_PERSISTENCE: 934 if (args->size) 935 ret = -EINVAL; 936 else 937 ret = proto_context_set_persistence(fpriv->i915, pc, 938 args->value); 939 break; 940 941 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 942 ret = proto_context_set_protected(fpriv->i915, pc, 943 args->value); 944 break; 945 946 case I915_CONTEXT_PARAM_NO_ZEROMAP: 947 case I915_CONTEXT_PARAM_BAN_PERIOD: 948 case I915_CONTEXT_PARAM_RINGSIZE: 949 default: 950 ret = -EINVAL; 951 break; 952 } 953 954 return ret; 955 } 956 957 static int intel_context_set_gem(struct intel_context *ce, 958 struct i915_gem_context *ctx, 959 struct intel_sseu sseu) 960 { 961 int ret = 0; 962 963 GEM_BUG_ON(rcu_access_pointer(ce->gem_context)); 964 RCU_INIT_POINTER(ce->gem_context, ctx); 965 966 GEM_BUG_ON(intel_context_is_pinned(ce)); 967 ce->ring_size = SZ_16K; 968 969 i915_vm_put(ce->vm); 970 ce->vm = i915_gem_context_get_eb_vm(ctx); 971 972 if (ctx->sched.priority >= I915_PRIORITY_NORMAL && 973 intel_engine_has_timeslices(ce->engine) && 974 intel_engine_has_semaphores(ce->engine)) 975 __set_bit(CONTEXT_USE_SEMAPHORES, &ce->flags); 976 977 if (CONFIG_DRM_I915_REQUEST_TIMEOUT && 978 ctx->i915->params.request_timeout_ms) { 979 unsigned int timeout_ms = ctx->i915->params.request_timeout_ms; 980 981 intel_context_set_watchdog_us(ce, (u64)timeout_ms * 1000); 982 } 983 984 /* A valid SSEU has no zero fields */ 985 if (sseu.slice_mask && !WARN_ON(ce->engine->class != RENDER_CLASS)) 986 ret = intel_context_reconfigure_sseu(ce, sseu); 987 988 return ret; 989 } 990 991 static void __unpin_engines(struct i915_gem_engines *e, unsigned int count) 992 { 993 while (count--) { 994 struct intel_context *ce = e->engines[count], *child; 995 996 if (!ce || !test_bit(CONTEXT_PERMA_PIN, &ce->flags)) 997 continue; 998 999 for_each_child(ce, child) 1000 intel_context_unpin(child); 1001 intel_context_unpin(ce); 1002 } 1003 } 1004 1005 static void unpin_engines(struct i915_gem_engines *e) 1006 { 1007 __unpin_engines(e, e->num_engines); 1008 } 1009 1010 static void __free_engines(struct i915_gem_engines *e, unsigned int count) 1011 { 1012 while (count--) { 1013 if (!e->engines[count]) 1014 continue; 1015 1016 intel_context_put(e->engines[count]); 1017 } 1018 kfree(e); 1019 } 1020 1021 static void free_engines(struct i915_gem_engines *e) 1022 { 1023 __free_engines(e, e->num_engines); 1024 } 1025 1026 static void free_engines_rcu(struct rcu_head *rcu) 1027 { 1028 struct i915_gem_engines *engines = 1029 container_of(rcu, struct i915_gem_engines, rcu); 1030 1031 i915_sw_fence_fini(&engines->fence); 1032 free_engines(engines); 1033 } 1034 1035 static void accumulate_runtime(struct i915_drm_client *client, 1036 struct i915_gem_engines *engines) 1037 { 1038 struct i915_gem_engines_iter it; 1039 struct intel_context *ce; 1040 1041 if (!client) 1042 return; 1043 1044 /* Transfer accumulated runtime to the parent GEM context. */ 1045 for_each_gem_engine(ce, engines, it) { 1046 unsigned int class = ce->engine->uabi_class; 1047 1048 GEM_BUG_ON(class >= ARRAY_SIZE(client->past_runtime)); 1049 atomic64_add(intel_context_get_total_runtime_ns(ce), 1050 &client->past_runtime[class]); 1051 } 1052 } 1053 1054 static int 1055 engines_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) 1056 { 1057 struct i915_gem_engines *engines = 1058 container_of(fence, typeof(*engines), fence); 1059 struct i915_gem_context *ctx = engines->ctx; 1060 1061 switch (state) { 1062 case FENCE_COMPLETE: 1063 if (!list_empty(&engines->link)) { 1064 unsigned long flags; 1065 1066 spin_lock_irqsave(&ctx->stale.lock, flags); 1067 list_del(&engines->link); 1068 spin_unlock_irqrestore(&ctx->stale.lock, flags); 1069 } 1070 accumulate_runtime(ctx->client, engines); 1071 i915_gem_context_put(ctx); 1072 1073 break; 1074 1075 case FENCE_FREE: 1076 init_rcu_head(&engines->rcu); 1077 call_rcu(&engines->rcu, free_engines_rcu); 1078 break; 1079 } 1080 1081 return NOTIFY_DONE; 1082 } 1083 1084 static struct i915_gem_engines *alloc_engines(unsigned int count) 1085 { 1086 struct i915_gem_engines *e; 1087 1088 e = kzalloc(struct_size(e, engines, count), GFP_KERNEL); 1089 if (!e) 1090 return NULL; 1091 1092 i915_sw_fence_init(&e->fence, engines_notify); 1093 return e; 1094 } 1095 1096 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx, 1097 struct intel_sseu rcs_sseu) 1098 { 1099 const unsigned int max = I915_NUM_ENGINES; 1100 struct intel_engine_cs *engine; 1101 struct i915_gem_engines *e, *err; 1102 1103 e = alloc_engines(max); 1104 if (!e) 1105 return ERR_PTR(-ENOMEM); 1106 1107 for_each_uabi_engine(engine, ctx->i915) { 1108 struct intel_context *ce; 1109 struct intel_sseu sseu = {}; 1110 int ret; 1111 1112 if (engine->legacy_idx == INVALID_ENGINE) 1113 continue; 1114 1115 GEM_BUG_ON(engine->legacy_idx >= max); 1116 GEM_BUG_ON(e->engines[engine->legacy_idx]); 1117 1118 ce = intel_context_create(engine); 1119 if (IS_ERR(ce)) { 1120 err = ERR_CAST(ce); 1121 goto free_engines; 1122 } 1123 1124 e->engines[engine->legacy_idx] = ce; 1125 e->num_engines = max(e->num_engines, engine->legacy_idx + 1); 1126 1127 if (engine->class == RENDER_CLASS) 1128 sseu = rcs_sseu; 1129 1130 ret = intel_context_set_gem(ce, ctx, sseu); 1131 if (ret) { 1132 err = ERR_PTR(ret); 1133 goto free_engines; 1134 } 1135 1136 } 1137 1138 return e; 1139 1140 free_engines: 1141 free_engines(e); 1142 return err; 1143 } 1144 1145 static int perma_pin_contexts(struct intel_context *ce) 1146 { 1147 struct intel_context *child; 1148 int i = 0, j = 0, ret; 1149 1150 GEM_BUG_ON(!intel_context_is_parent(ce)); 1151 1152 ret = intel_context_pin(ce); 1153 if (unlikely(ret)) 1154 return ret; 1155 1156 for_each_child(ce, child) { 1157 ret = intel_context_pin(child); 1158 if (unlikely(ret)) 1159 goto unwind; 1160 ++i; 1161 } 1162 1163 set_bit(CONTEXT_PERMA_PIN, &ce->flags); 1164 1165 return 0; 1166 1167 unwind: 1168 intel_context_unpin(ce); 1169 for_each_child(ce, child) { 1170 if (j++ < i) 1171 intel_context_unpin(child); 1172 else 1173 break; 1174 } 1175 1176 return ret; 1177 } 1178 1179 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx, 1180 unsigned int num_engines, 1181 struct i915_gem_proto_engine *pe) 1182 { 1183 struct i915_gem_engines *e, *err; 1184 unsigned int n; 1185 1186 e = alloc_engines(num_engines); 1187 if (!e) 1188 return ERR_PTR(-ENOMEM); 1189 e->num_engines = num_engines; 1190 1191 for (n = 0; n < num_engines; n++) { 1192 struct intel_context *ce, *child; 1193 int ret; 1194 1195 switch (pe[n].type) { 1196 case I915_GEM_ENGINE_TYPE_PHYSICAL: 1197 ce = intel_context_create(pe[n].engine); 1198 break; 1199 1200 case I915_GEM_ENGINE_TYPE_BALANCED: 1201 ce = intel_engine_create_virtual(pe[n].siblings, 1202 pe[n].num_siblings, 0); 1203 break; 1204 1205 case I915_GEM_ENGINE_TYPE_PARALLEL: 1206 ce = intel_engine_create_parallel(pe[n].siblings, 1207 pe[n].num_siblings, 1208 pe[n].width); 1209 break; 1210 1211 case I915_GEM_ENGINE_TYPE_INVALID: 1212 default: 1213 GEM_WARN_ON(pe[n].type != I915_GEM_ENGINE_TYPE_INVALID); 1214 continue; 1215 } 1216 1217 if (IS_ERR(ce)) { 1218 err = ERR_CAST(ce); 1219 goto free_engines; 1220 } 1221 1222 e->engines[n] = ce; 1223 1224 ret = intel_context_set_gem(ce, ctx, pe->sseu); 1225 if (ret) { 1226 err = ERR_PTR(ret); 1227 goto free_engines; 1228 } 1229 for_each_child(ce, child) { 1230 ret = intel_context_set_gem(child, ctx, pe->sseu); 1231 if (ret) { 1232 err = ERR_PTR(ret); 1233 goto free_engines; 1234 } 1235 } 1236 1237 /* 1238 * XXX: Must be done after calling intel_context_set_gem as that 1239 * function changes the ring size. The ring is allocated when 1240 * the context is pinned. If the ring size is changed after 1241 * allocation we have a mismatch of the ring size and will cause 1242 * the context to hang. Presumably with a bit of reordering we 1243 * could move the perma-pin step to the backend function 1244 * intel_engine_create_parallel. 1245 */ 1246 if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) { 1247 ret = perma_pin_contexts(ce); 1248 if (ret) { 1249 err = ERR_PTR(ret); 1250 goto free_engines; 1251 } 1252 } 1253 } 1254 1255 return e; 1256 1257 free_engines: 1258 free_engines(e); 1259 return err; 1260 } 1261 1262 static void i915_gem_context_release_work(struct work_struct *work) 1263 { 1264 struct i915_gem_context *ctx = container_of(work, typeof(*ctx), 1265 release_work); 1266 struct i915_address_space *vm; 1267 1268 trace_i915_context_free(ctx); 1269 GEM_BUG_ON(!i915_gem_context_is_closed(ctx)); 1270 1271 spin_lock(&ctx->i915->gem.contexts.lock); 1272 list_del(&ctx->link); 1273 spin_unlock(&ctx->i915->gem.contexts.lock); 1274 1275 if (ctx->syncobj) 1276 drm_syncobj_put(ctx->syncobj); 1277 1278 vm = ctx->vm; 1279 if (vm) 1280 i915_vm_put(vm); 1281 1282 if (ctx->pxp_wakeref) 1283 intel_runtime_pm_put(&ctx->i915->runtime_pm, ctx->pxp_wakeref); 1284 1285 if (ctx->client) 1286 i915_drm_client_put(ctx->client); 1287 1288 mutex_destroy(&ctx->engines_mutex); 1289 mutex_destroy(&ctx->lut_mutex); 1290 1291 put_pid(ctx->pid); 1292 mutex_destroy(&ctx->mutex); 1293 1294 kfree_rcu(ctx, rcu); 1295 } 1296 1297 void i915_gem_context_release(struct kref *ref) 1298 { 1299 struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref); 1300 1301 queue_work(ctx->i915->wq, &ctx->release_work); 1302 } 1303 1304 static inline struct i915_gem_engines * 1305 __context_engines_static(const struct i915_gem_context *ctx) 1306 { 1307 return rcu_dereference_protected(ctx->engines, true); 1308 } 1309 1310 static void __reset_context(struct i915_gem_context *ctx, 1311 struct intel_engine_cs *engine) 1312 { 1313 intel_gt_handle_error(engine->gt, engine->mask, 0, 1314 "context closure in %s", ctx->name); 1315 } 1316 1317 static bool __cancel_engine(struct intel_engine_cs *engine) 1318 { 1319 /* 1320 * Send a "high priority pulse" down the engine to cause the 1321 * current request to be momentarily preempted. (If it fails to 1322 * be preempted, it will be reset). As we have marked our context 1323 * as banned, any incomplete request, including any running, will 1324 * be skipped following the preemption. 1325 * 1326 * If there is no hangchecking (one of the reasons why we try to 1327 * cancel the context) and no forced preemption, there may be no 1328 * means by which we reset the GPU and evict the persistent hog. 1329 * Ergo if we are unable to inject a preemptive pulse that can 1330 * kill the banned context, we fallback to doing a local reset 1331 * instead. 1332 */ 1333 return intel_engine_pulse(engine) == 0; 1334 } 1335 1336 static struct intel_engine_cs *active_engine(struct intel_context *ce) 1337 { 1338 struct intel_engine_cs *engine = NULL; 1339 struct i915_request *rq; 1340 1341 if (intel_context_has_inflight(ce)) 1342 return intel_context_inflight(ce); 1343 1344 if (!ce->timeline) 1345 return NULL; 1346 1347 /* 1348 * rq->link is only SLAB_TYPESAFE_BY_RCU, we need to hold a reference 1349 * to the request to prevent it being transferred to a new timeline 1350 * (and onto a new timeline->requests list). 1351 */ 1352 rcu_read_lock(); 1353 list_for_each_entry_reverse(rq, &ce->timeline->requests, link) { 1354 bool found; 1355 1356 /* timeline is already completed upto this point? */ 1357 if (!i915_request_get_rcu(rq)) 1358 break; 1359 1360 /* Check with the backend if the request is inflight */ 1361 found = true; 1362 if (likely(rcu_access_pointer(rq->timeline) == ce->timeline)) 1363 found = i915_request_active_engine(rq, &engine); 1364 1365 i915_request_put(rq); 1366 if (found) 1367 break; 1368 } 1369 rcu_read_unlock(); 1370 1371 return engine; 1372 } 1373 1374 static void 1375 kill_engines(struct i915_gem_engines *engines, bool exit, bool persistent) 1376 { 1377 struct i915_gem_engines_iter it; 1378 struct intel_context *ce; 1379 1380 /* 1381 * Map the user's engine back to the actual engines; one virtual 1382 * engine will be mapped to multiple engines, and using ctx->engine[] 1383 * the same engine may be have multiple instances in the user's map. 1384 * However, we only care about pending requests, so only include 1385 * engines on which there are incomplete requests. 1386 */ 1387 for_each_gem_engine(ce, engines, it) { 1388 struct intel_engine_cs *engine; 1389 1390 if ((exit || !persistent) && intel_context_revoke(ce)) 1391 continue; /* Already marked. */ 1392 1393 /* 1394 * Check the current active state of this context; if we 1395 * are currently executing on the GPU we need to evict 1396 * ourselves. On the other hand, if we haven't yet been 1397 * submitted to the GPU or if everything is complete, 1398 * we have nothing to do. 1399 */ 1400 engine = active_engine(ce); 1401 1402 /* First attempt to gracefully cancel the context */ 1403 if (engine && !__cancel_engine(engine) && (exit || !persistent)) 1404 /* 1405 * If we are unable to send a preemptive pulse to bump 1406 * the context from the GPU, we have to resort to a full 1407 * reset. We hope the collateral damage is worth it. 1408 */ 1409 __reset_context(engines->ctx, engine); 1410 } 1411 } 1412 1413 static void kill_context(struct i915_gem_context *ctx) 1414 { 1415 struct i915_gem_engines *pos, *next; 1416 1417 spin_lock_irq(&ctx->stale.lock); 1418 GEM_BUG_ON(!i915_gem_context_is_closed(ctx)); 1419 list_for_each_entry_safe(pos, next, &ctx->stale.engines, link) { 1420 if (!i915_sw_fence_await(&pos->fence)) { 1421 list_del_init(&pos->link); 1422 continue; 1423 } 1424 1425 spin_unlock_irq(&ctx->stale.lock); 1426 1427 kill_engines(pos, !ctx->i915->params.enable_hangcheck, 1428 i915_gem_context_is_persistent(ctx)); 1429 1430 spin_lock_irq(&ctx->stale.lock); 1431 GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence)); 1432 list_safe_reset_next(pos, next, link); 1433 list_del_init(&pos->link); /* decouple from FENCE_COMPLETE */ 1434 1435 i915_sw_fence_complete(&pos->fence); 1436 } 1437 spin_unlock_irq(&ctx->stale.lock); 1438 } 1439 1440 static void engines_idle_release(struct i915_gem_context *ctx, 1441 struct i915_gem_engines *engines) 1442 { 1443 struct i915_gem_engines_iter it; 1444 struct intel_context *ce; 1445 1446 INIT_LIST_HEAD(&engines->link); 1447 1448 engines->ctx = i915_gem_context_get(ctx); 1449 1450 for_each_gem_engine(ce, engines, it) { 1451 int err; 1452 1453 /* serialises with execbuf */ 1454 intel_context_close(ce); 1455 if (!intel_context_pin_if_active(ce)) 1456 continue; 1457 1458 /* Wait until context is finally scheduled out and retired */ 1459 err = i915_sw_fence_await_active(&engines->fence, 1460 &ce->active, 1461 I915_ACTIVE_AWAIT_BARRIER); 1462 intel_context_unpin(ce); 1463 if (err) 1464 goto kill; 1465 } 1466 1467 spin_lock_irq(&ctx->stale.lock); 1468 if (!i915_gem_context_is_closed(ctx)) 1469 list_add_tail(&engines->link, &ctx->stale.engines); 1470 spin_unlock_irq(&ctx->stale.lock); 1471 1472 kill: 1473 if (list_empty(&engines->link)) /* raced, already closed */ 1474 kill_engines(engines, true, 1475 i915_gem_context_is_persistent(ctx)); 1476 1477 i915_sw_fence_commit(&engines->fence); 1478 } 1479 1480 static void set_closed_name(struct i915_gem_context *ctx) 1481 { 1482 char *s; 1483 1484 /* Replace '[]' with '<>' to indicate closed in debug prints */ 1485 1486 s = strrchr(ctx->name, '['); 1487 if (!s) 1488 return; 1489 1490 *s = '<'; 1491 1492 s = strchr(s + 1, ']'); 1493 if (s) 1494 *s = '>'; 1495 } 1496 1497 static void context_close(struct i915_gem_context *ctx) 1498 { 1499 struct i915_drm_client *client; 1500 1501 /* Flush any concurrent set_engines() */ 1502 mutex_lock(&ctx->engines_mutex); 1503 unpin_engines(__context_engines_static(ctx)); 1504 engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1)); 1505 i915_gem_context_set_closed(ctx); 1506 mutex_unlock(&ctx->engines_mutex); 1507 1508 mutex_lock(&ctx->mutex); 1509 1510 set_closed_name(ctx); 1511 1512 /* 1513 * The LUT uses the VMA as a backpointer to unref the object, 1514 * so we need to clear the LUT before we close all the VMA (inside 1515 * the ppgtt). 1516 */ 1517 lut_close(ctx); 1518 1519 ctx->file_priv = ERR_PTR(-EBADF); 1520 1521 client = ctx->client; 1522 if (client) { 1523 spin_lock(&client->ctx_lock); 1524 list_del_rcu(&ctx->client_link); 1525 spin_unlock(&client->ctx_lock); 1526 } 1527 1528 mutex_unlock(&ctx->mutex); 1529 1530 /* 1531 * If the user has disabled hangchecking, we can not be sure that 1532 * the batches will ever complete after the context is closed, 1533 * keeping the context and all resources pinned forever. So in this 1534 * case we opt to forcibly kill off all remaining requests on 1535 * context close. 1536 */ 1537 kill_context(ctx); 1538 1539 i915_gem_context_put(ctx); 1540 } 1541 1542 static int __context_set_persistence(struct i915_gem_context *ctx, bool state) 1543 { 1544 if (i915_gem_context_is_persistent(ctx) == state) 1545 return 0; 1546 1547 if (state) { 1548 /* 1549 * Only contexts that are short-lived [that will expire or be 1550 * reset] are allowed to survive past termination. We require 1551 * hangcheck to ensure that the persistent requests are healthy. 1552 */ 1553 if (!ctx->i915->params.enable_hangcheck) 1554 return -EINVAL; 1555 1556 i915_gem_context_set_persistence(ctx); 1557 } else { 1558 /* To cancel a context we use "preempt-to-idle" */ 1559 if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION)) 1560 return -ENODEV; 1561 1562 /* 1563 * If the cancel fails, we then need to reset, cleanly! 1564 * 1565 * If the per-engine reset fails, all hope is lost! We resort 1566 * to a full GPU reset in that unlikely case, but realistically 1567 * if the engine could not reset, the full reset does not fare 1568 * much better. The damage has been done. 1569 * 1570 * However, if we cannot reset an engine by itself, we cannot 1571 * cleanup a hanging persistent context without causing 1572 * colateral damage, and we should not pretend we can by 1573 * exposing the interface. 1574 */ 1575 if (!intel_has_reset_engine(to_gt(ctx->i915))) 1576 return -ENODEV; 1577 1578 i915_gem_context_clear_persistence(ctx); 1579 } 1580 1581 return 0; 1582 } 1583 1584 static struct i915_gem_context * 1585 i915_gem_create_context(struct drm_i915_private *i915, 1586 const struct i915_gem_proto_context *pc) 1587 { 1588 struct i915_gem_context *ctx; 1589 struct i915_address_space *vm = NULL; 1590 struct i915_gem_engines *e; 1591 int err; 1592 int i; 1593 1594 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 1595 if (!ctx) 1596 return ERR_PTR(-ENOMEM); 1597 1598 kref_init(&ctx->ref); 1599 ctx->i915 = i915; 1600 ctx->sched = pc->sched; 1601 mutex_init(&ctx->mutex); 1602 INIT_LIST_HEAD(&ctx->link); 1603 INIT_WORK(&ctx->release_work, i915_gem_context_release_work); 1604 1605 spin_lock_init(&ctx->stale.lock); 1606 INIT_LIST_HEAD(&ctx->stale.engines); 1607 1608 if (pc->vm) { 1609 vm = i915_vm_get(pc->vm); 1610 } else if (HAS_FULL_PPGTT(i915)) { 1611 struct i915_ppgtt *ppgtt; 1612 1613 ppgtt = i915_ppgtt_create(to_gt(i915), 0); 1614 if (IS_ERR(ppgtt)) { 1615 drm_dbg(&i915->drm, "PPGTT setup failed (%ld)\n", 1616 PTR_ERR(ppgtt)); 1617 err = PTR_ERR(ppgtt); 1618 goto err_ctx; 1619 } 1620 vm = &ppgtt->vm; 1621 } 1622 if (vm) 1623 ctx->vm = vm; 1624 1625 mutex_init(&ctx->engines_mutex); 1626 if (pc->num_user_engines >= 0) { 1627 i915_gem_context_set_user_engines(ctx); 1628 e = user_engines(ctx, pc->num_user_engines, pc->user_engines); 1629 } else { 1630 i915_gem_context_clear_user_engines(ctx); 1631 e = default_engines(ctx, pc->legacy_rcs_sseu); 1632 } 1633 if (IS_ERR(e)) { 1634 err = PTR_ERR(e); 1635 goto err_vm; 1636 } 1637 RCU_INIT_POINTER(ctx->engines, e); 1638 1639 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL); 1640 mutex_init(&ctx->lut_mutex); 1641 1642 /* NB: Mark all slices as needing a remap so that when the context first 1643 * loads it will restore whatever remap state already exists. If there 1644 * is no remap info, it will be a NOP. */ 1645 ctx->remap_slice = ALL_L3_SLICES(i915); 1646 1647 ctx->user_flags = pc->user_flags; 1648 1649 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++) 1650 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES; 1651 1652 if (pc->single_timeline) { 1653 err = drm_syncobj_create(&ctx->syncobj, 1654 DRM_SYNCOBJ_CREATE_SIGNALED, 1655 NULL); 1656 if (err) 1657 goto err_engines; 1658 } 1659 1660 if (pc->uses_protected_content) { 1661 ctx->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm); 1662 ctx->uses_protected_content = true; 1663 } 1664 1665 trace_i915_context_create(ctx); 1666 1667 return ctx; 1668 1669 err_engines: 1670 free_engines(e); 1671 err_vm: 1672 if (ctx->vm) 1673 i915_vm_put(ctx->vm); 1674 err_ctx: 1675 kfree(ctx); 1676 return ERR_PTR(err); 1677 } 1678 1679 static void init_contexts(struct i915_gem_contexts *gc) 1680 { 1681 spin_lock_init(&gc->lock); 1682 INIT_LIST_HEAD(&gc->list); 1683 } 1684 1685 void i915_gem_init__contexts(struct drm_i915_private *i915) 1686 { 1687 init_contexts(&i915->gem.contexts); 1688 } 1689 1690 /* 1691 * Note that this implicitly consumes the ctx reference, by placing 1692 * the ctx in the context_xa. 1693 */ 1694 static void gem_context_register(struct i915_gem_context *ctx, 1695 struct drm_i915_file_private *fpriv, 1696 u32 id) 1697 { 1698 struct drm_i915_private *i915 = ctx->i915; 1699 void *old; 1700 1701 ctx->file_priv = fpriv; 1702 1703 ctx->pid = get_task_pid(current, PIDTYPE_PID); 1704 ctx->client = i915_drm_client_get(fpriv->client); 1705 1706 snprintf(ctx->name, sizeof(ctx->name), "%s[%d]", 1707 current->comm, pid_nr(ctx->pid)); 1708 1709 spin_lock(&ctx->client->ctx_lock); 1710 list_add_tail_rcu(&ctx->client_link, &ctx->client->ctx_list); 1711 spin_unlock(&ctx->client->ctx_lock); 1712 1713 spin_lock(&i915->gem.contexts.lock); 1714 list_add_tail(&ctx->link, &i915->gem.contexts.list); 1715 spin_unlock(&i915->gem.contexts.lock); 1716 1717 /* And finally expose ourselves to userspace via the idr */ 1718 old = xa_store(&fpriv->context_xa, id, ctx, GFP_KERNEL); 1719 WARN_ON(old); 1720 } 1721 1722 int i915_gem_context_open(struct drm_i915_private *i915, 1723 struct drm_file *file) 1724 { 1725 struct drm_i915_file_private *file_priv = file->driver_priv; 1726 struct i915_gem_proto_context *pc; 1727 struct i915_gem_context *ctx; 1728 int err; 1729 1730 mutex_init(&file_priv->proto_context_lock); 1731 xa_init_flags(&file_priv->proto_context_xa, XA_FLAGS_ALLOC); 1732 1733 /* 0 reserved for the default context */ 1734 xa_init_flags(&file_priv->context_xa, XA_FLAGS_ALLOC1); 1735 1736 /* 0 reserved for invalid/unassigned ppgtt */ 1737 xa_init_flags(&file_priv->vm_xa, XA_FLAGS_ALLOC1); 1738 1739 pc = proto_context_create(i915, 0); 1740 if (IS_ERR(pc)) { 1741 err = PTR_ERR(pc); 1742 goto err; 1743 } 1744 1745 ctx = i915_gem_create_context(i915, pc); 1746 proto_context_close(i915, pc); 1747 if (IS_ERR(ctx)) { 1748 err = PTR_ERR(ctx); 1749 goto err; 1750 } 1751 1752 gem_context_register(ctx, file_priv, 0); 1753 1754 return 0; 1755 1756 err: 1757 xa_destroy(&file_priv->vm_xa); 1758 xa_destroy(&file_priv->context_xa); 1759 xa_destroy(&file_priv->proto_context_xa); 1760 mutex_destroy(&file_priv->proto_context_lock); 1761 return err; 1762 } 1763 1764 void i915_gem_context_close(struct drm_file *file) 1765 { 1766 struct drm_i915_file_private *file_priv = file->driver_priv; 1767 struct i915_gem_proto_context *pc; 1768 struct i915_address_space *vm; 1769 struct i915_gem_context *ctx; 1770 unsigned long idx; 1771 1772 xa_for_each(&file_priv->proto_context_xa, idx, pc) 1773 proto_context_close(file_priv->i915, pc); 1774 xa_destroy(&file_priv->proto_context_xa); 1775 mutex_destroy(&file_priv->proto_context_lock); 1776 1777 xa_for_each(&file_priv->context_xa, idx, ctx) 1778 context_close(ctx); 1779 xa_destroy(&file_priv->context_xa); 1780 1781 xa_for_each(&file_priv->vm_xa, idx, vm) 1782 i915_vm_put(vm); 1783 xa_destroy(&file_priv->vm_xa); 1784 } 1785 1786 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data, 1787 struct drm_file *file) 1788 { 1789 struct drm_i915_private *i915 = to_i915(dev); 1790 struct drm_i915_gem_vm_control *args = data; 1791 struct drm_i915_file_private *file_priv = file->driver_priv; 1792 struct i915_ppgtt *ppgtt; 1793 u32 id; 1794 int err; 1795 1796 if (!HAS_FULL_PPGTT(i915)) 1797 return -ENODEV; 1798 1799 if (args->flags) 1800 return -EINVAL; 1801 1802 ppgtt = i915_ppgtt_create(to_gt(i915), 0); 1803 if (IS_ERR(ppgtt)) 1804 return PTR_ERR(ppgtt); 1805 1806 if (args->extensions) { 1807 err = i915_user_extensions(u64_to_user_ptr(args->extensions), 1808 NULL, 0, 1809 ppgtt); 1810 if (err) 1811 goto err_put; 1812 } 1813 1814 err = xa_alloc(&file_priv->vm_xa, &id, &ppgtt->vm, 1815 xa_limit_32b, GFP_KERNEL); 1816 if (err) 1817 goto err_put; 1818 1819 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */ 1820 args->vm_id = id; 1821 return 0; 1822 1823 err_put: 1824 i915_vm_put(&ppgtt->vm); 1825 return err; 1826 } 1827 1828 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data, 1829 struct drm_file *file) 1830 { 1831 struct drm_i915_file_private *file_priv = file->driver_priv; 1832 struct drm_i915_gem_vm_control *args = data; 1833 struct i915_address_space *vm; 1834 1835 if (args->flags) 1836 return -EINVAL; 1837 1838 if (args->extensions) 1839 return -EINVAL; 1840 1841 vm = xa_erase(&file_priv->vm_xa, args->vm_id); 1842 if (!vm) 1843 return -ENOENT; 1844 1845 i915_vm_put(vm); 1846 return 0; 1847 } 1848 1849 static int get_ppgtt(struct drm_i915_file_private *file_priv, 1850 struct i915_gem_context *ctx, 1851 struct drm_i915_gem_context_param *args) 1852 { 1853 struct i915_address_space *vm; 1854 int err; 1855 u32 id; 1856 1857 if (!i915_gem_context_has_full_ppgtt(ctx)) 1858 return -ENODEV; 1859 1860 vm = ctx->vm; 1861 GEM_BUG_ON(!vm); 1862 1863 /* 1864 * Get a reference for the allocated handle. Once the handle is 1865 * visible in the vm_xa table, userspace could try to close it 1866 * from under our feet, so we need to hold the extra reference 1867 * first. 1868 */ 1869 i915_vm_get(vm); 1870 1871 err = xa_alloc(&file_priv->vm_xa, &id, vm, xa_limit_32b, GFP_KERNEL); 1872 if (err) { 1873 i915_vm_put(vm); 1874 return err; 1875 } 1876 1877 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */ 1878 args->value = id; 1879 args->size = 0; 1880 1881 return err; 1882 } 1883 1884 int 1885 i915_gem_user_to_context_sseu(struct intel_gt *gt, 1886 const struct drm_i915_gem_context_param_sseu *user, 1887 struct intel_sseu *context) 1888 { 1889 const struct sseu_dev_info *device = >->info.sseu; 1890 struct drm_i915_private *i915 = gt->i915; 1891 unsigned int dev_subslice_mask = intel_sseu_get_hsw_subslices(device, 0); 1892 1893 /* No zeros in any field. */ 1894 if (!user->slice_mask || !user->subslice_mask || 1895 !user->min_eus_per_subslice || !user->max_eus_per_subslice) 1896 return -EINVAL; 1897 1898 /* Max > min. */ 1899 if (user->max_eus_per_subslice < user->min_eus_per_subslice) 1900 return -EINVAL; 1901 1902 /* 1903 * Some future proofing on the types since the uAPI is wider than the 1904 * current internal implementation. 1905 */ 1906 if (overflows_type(user->slice_mask, context->slice_mask) || 1907 overflows_type(user->subslice_mask, context->subslice_mask) || 1908 overflows_type(user->min_eus_per_subslice, 1909 context->min_eus_per_subslice) || 1910 overflows_type(user->max_eus_per_subslice, 1911 context->max_eus_per_subslice)) 1912 return -EINVAL; 1913 1914 /* Check validity against hardware. */ 1915 if (user->slice_mask & ~device->slice_mask) 1916 return -EINVAL; 1917 1918 if (user->subslice_mask & ~dev_subslice_mask) 1919 return -EINVAL; 1920 1921 if (user->max_eus_per_subslice > device->max_eus_per_subslice) 1922 return -EINVAL; 1923 1924 context->slice_mask = user->slice_mask; 1925 context->subslice_mask = user->subslice_mask; 1926 context->min_eus_per_subslice = user->min_eus_per_subslice; 1927 context->max_eus_per_subslice = user->max_eus_per_subslice; 1928 1929 /* Part specific restrictions. */ 1930 if (GRAPHICS_VER(i915) == 11) { 1931 unsigned int hw_s = hweight8(device->slice_mask); 1932 unsigned int hw_ss_per_s = hweight8(dev_subslice_mask); 1933 unsigned int req_s = hweight8(context->slice_mask); 1934 unsigned int req_ss = hweight8(context->subslice_mask); 1935 1936 /* 1937 * Only full subslice enablement is possible if more than one 1938 * slice is turned on. 1939 */ 1940 if (req_s > 1 && req_ss != hw_ss_per_s) 1941 return -EINVAL; 1942 1943 /* 1944 * If more than four (SScount bitfield limit) subslices are 1945 * requested then the number has to be even. 1946 */ 1947 if (req_ss > 4 && (req_ss & 1)) 1948 return -EINVAL; 1949 1950 /* 1951 * If only one slice is enabled and subslice count is below the 1952 * device full enablement, it must be at most half of the all 1953 * available subslices. 1954 */ 1955 if (req_s == 1 && req_ss < hw_ss_per_s && 1956 req_ss > (hw_ss_per_s / 2)) 1957 return -EINVAL; 1958 1959 /* ABI restriction - VME use case only. */ 1960 1961 /* All slices or one slice only. */ 1962 if (req_s != 1 && req_s != hw_s) 1963 return -EINVAL; 1964 1965 /* 1966 * Half subslices or full enablement only when one slice is 1967 * enabled. 1968 */ 1969 if (req_s == 1 && 1970 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2))) 1971 return -EINVAL; 1972 1973 /* No EU configuration changes. */ 1974 if ((user->min_eus_per_subslice != 1975 device->max_eus_per_subslice) || 1976 (user->max_eus_per_subslice != 1977 device->max_eus_per_subslice)) 1978 return -EINVAL; 1979 } 1980 1981 return 0; 1982 } 1983 1984 static int set_sseu(struct i915_gem_context *ctx, 1985 struct drm_i915_gem_context_param *args) 1986 { 1987 struct drm_i915_private *i915 = ctx->i915; 1988 struct drm_i915_gem_context_param_sseu user_sseu; 1989 struct intel_context *ce; 1990 struct intel_sseu sseu; 1991 unsigned long lookup; 1992 int ret; 1993 1994 if (args->size < sizeof(user_sseu)) 1995 return -EINVAL; 1996 1997 if (GRAPHICS_VER(i915) != 11) 1998 return -ENODEV; 1999 2000 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 2001 sizeof(user_sseu))) 2002 return -EFAULT; 2003 2004 if (user_sseu.rsvd) 2005 return -EINVAL; 2006 2007 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 2008 return -EINVAL; 2009 2010 lookup = 0; 2011 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) 2012 lookup |= LOOKUP_USER_INDEX; 2013 2014 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine); 2015 if (IS_ERR(ce)) 2016 return PTR_ERR(ce); 2017 2018 /* Only render engine supports RPCS configuration. */ 2019 if (ce->engine->class != RENDER_CLASS) { 2020 ret = -ENODEV; 2021 goto out_ce; 2022 } 2023 2024 ret = i915_gem_user_to_context_sseu(ce->engine->gt, &user_sseu, &sseu); 2025 if (ret) 2026 goto out_ce; 2027 2028 ret = intel_context_reconfigure_sseu(ce, sseu); 2029 if (ret) 2030 goto out_ce; 2031 2032 args->size = sizeof(user_sseu); 2033 2034 out_ce: 2035 intel_context_put(ce); 2036 return ret; 2037 } 2038 2039 static int 2040 set_persistence(struct i915_gem_context *ctx, 2041 const struct drm_i915_gem_context_param *args) 2042 { 2043 if (args->size) 2044 return -EINVAL; 2045 2046 return __context_set_persistence(ctx, args->value); 2047 } 2048 2049 static int set_priority(struct i915_gem_context *ctx, 2050 const struct drm_i915_gem_context_param *args) 2051 { 2052 struct i915_gem_engines_iter it; 2053 struct intel_context *ce; 2054 int err; 2055 2056 err = validate_priority(ctx->i915, args); 2057 if (err) 2058 return err; 2059 2060 ctx->sched.priority = args->value; 2061 2062 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) { 2063 if (!intel_engine_has_timeslices(ce->engine)) 2064 continue; 2065 2066 if (ctx->sched.priority >= I915_PRIORITY_NORMAL && 2067 intel_engine_has_semaphores(ce->engine)) 2068 intel_context_set_use_semaphores(ce); 2069 else 2070 intel_context_clear_use_semaphores(ce); 2071 } 2072 i915_gem_context_unlock_engines(ctx); 2073 2074 return 0; 2075 } 2076 2077 static int get_protected(struct i915_gem_context *ctx, 2078 struct drm_i915_gem_context_param *args) 2079 { 2080 args->size = 0; 2081 args->value = i915_gem_context_uses_protected_content(ctx); 2082 2083 return 0; 2084 } 2085 2086 static int ctx_setparam(struct drm_i915_file_private *fpriv, 2087 struct i915_gem_context *ctx, 2088 struct drm_i915_gem_context_param *args) 2089 { 2090 int ret = 0; 2091 2092 switch (args->param) { 2093 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 2094 if (args->size) 2095 ret = -EINVAL; 2096 else if (args->value) 2097 i915_gem_context_set_no_error_capture(ctx); 2098 else 2099 i915_gem_context_clear_no_error_capture(ctx); 2100 break; 2101 2102 case I915_CONTEXT_PARAM_BANNABLE: 2103 if (args->size) 2104 ret = -EINVAL; 2105 else if (!capable(CAP_SYS_ADMIN) && !args->value) 2106 ret = -EPERM; 2107 else if (args->value) 2108 i915_gem_context_set_bannable(ctx); 2109 else if (i915_gem_context_uses_protected_content(ctx)) 2110 ret = -EPERM; /* can't clear this for protected contexts */ 2111 else 2112 i915_gem_context_clear_bannable(ctx); 2113 break; 2114 2115 case I915_CONTEXT_PARAM_RECOVERABLE: 2116 if (args->size) 2117 ret = -EINVAL; 2118 else if (!args->value) 2119 i915_gem_context_clear_recoverable(ctx); 2120 else if (i915_gem_context_uses_protected_content(ctx)) 2121 ret = -EPERM; /* can't set this for protected contexts */ 2122 else 2123 i915_gem_context_set_recoverable(ctx); 2124 break; 2125 2126 case I915_CONTEXT_PARAM_PRIORITY: 2127 ret = set_priority(ctx, args); 2128 break; 2129 2130 case I915_CONTEXT_PARAM_SSEU: 2131 ret = set_sseu(ctx, args); 2132 break; 2133 2134 case I915_CONTEXT_PARAM_PERSISTENCE: 2135 ret = set_persistence(ctx, args); 2136 break; 2137 2138 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 2139 case I915_CONTEXT_PARAM_NO_ZEROMAP: 2140 case I915_CONTEXT_PARAM_BAN_PERIOD: 2141 case I915_CONTEXT_PARAM_RINGSIZE: 2142 case I915_CONTEXT_PARAM_VM: 2143 case I915_CONTEXT_PARAM_ENGINES: 2144 default: 2145 ret = -EINVAL; 2146 break; 2147 } 2148 2149 return ret; 2150 } 2151 2152 struct create_ext { 2153 struct i915_gem_proto_context *pc; 2154 struct drm_i915_file_private *fpriv; 2155 }; 2156 2157 static int create_setparam(struct i915_user_extension __user *ext, void *data) 2158 { 2159 struct drm_i915_gem_context_create_ext_setparam local; 2160 const struct create_ext *arg = data; 2161 2162 if (copy_from_user(&local, ext, sizeof(local))) 2163 return -EFAULT; 2164 2165 if (local.param.ctx_id) 2166 return -EINVAL; 2167 2168 return set_proto_ctx_param(arg->fpriv, arg->pc, &local.param); 2169 } 2170 2171 static int invalid_ext(struct i915_user_extension __user *ext, void *data) 2172 { 2173 return -EINVAL; 2174 } 2175 2176 static const i915_user_extension_fn create_extensions[] = { 2177 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam, 2178 [I915_CONTEXT_CREATE_EXT_CLONE] = invalid_ext, 2179 }; 2180 2181 static bool client_is_banned(struct drm_i915_file_private *file_priv) 2182 { 2183 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED; 2184 } 2185 2186 static inline struct i915_gem_context * 2187 __context_lookup(struct drm_i915_file_private *file_priv, u32 id) 2188 { 2189 struct i915_gem_context *ctx; 2190 2191 rcu_read_lock(); 2192 ctx = xa_load(&file_priv->context_xa, id); 2193 if (ctx && !kref_get_unless_zero(&ctx->ref)) 2194 ctx = NULL; 2195 rcu_read_unlock(); 2196 2197 return ctx; 2198 } 2199 2200 static struct i915_gem_context * 2201 finalize_create_context_locked(struct drm_i915_file_private *file_priv, 2202 struct i915_gem_proto_context *pc, u32 id) 2203 { 2204 struct i915_gem_context *ctx; 2205 void *old; 2206 2207 lockdep_assert_held(&file_priv->proto_context_lock); 2208 2209 ctx = i915_gem_create_context(file_priv->i915, pc); 2210 if (IS_ERR(ctx)) 2211 return ctx; 2212 2213 /* 2214 * One for the xarray and one for the caller. We need to grab 2215 * the reference *prior* to making the ctx visble to userspace 2216 * in gem_context_register(), as at any point after that 2217 * userspace can try to race us with another thread destroying 2218 * the context under our feet. 2219 */ 2220 i915_gem_context_get(ctx); 2221 2222 gem_context_register(ctx, file_priv, id); 2223 2224 old = xa_erase(&file_priv->proto_context_xa, id); 2225 GEM_BUG_ON(old != pc); 2226 proto_context_close(file_priv->i915, pc); 2227 2228 return ctx; 2229 } 2230 2231 struct i915_gem_context * 2232 i915_gem_context_lookup(struct drm_i915_file_private *file_priv, u32 id) 2233 { 2234 struct i915_gem_proto_context *pc; 2235 struct i915_gem_context *ctx; 2236 2237 ctx = __context_lookup(file_priv, id); 2238 if (ctx) 2239 return ctx; 2240 2241 mutex_lock(&file_priv->proto_context_lock); 2242 /* Try one more time under the lock */ 2243 ctx = __context_lookup(file_priv, id); 2244 if (!ctx) { 2245 pc = xa_load(&file_priv->proto_context_xa, id); 2246 if (!pc) 2247 ctx = ERR_PTR(-ENOENT); 2248 else 2249 ctx = finalize_create_context_locked(file_priv, pc, id); 2250 } 2251 mutex_unlock(&file_priv->proto_context_lock); 2252 2253 return ctx; 2254 } 2255 2256 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data, 2257 struct drm_file *file) 2258 { 2259 struct drm_i915_private *i915 = to_i915(dev); 2260 struct drm_i915_gem_context_create_ext *args = data; 2261 struct create_ext ext_data; 2262 int ret; 2263 u32 id; 2264 2265 if (!DRIVER_CAPS(i915)->has_logical_contexts) 2266 return -ENODEV; 2267 2268 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN) 2269 return -EINVAL; 2270 2271 ret = intel_gt_terminally_wedged(to_gt(i915)); 2272 if (ret) 2273 return ret; 2274 2275 ext_data.fpriv = file->driver_priv; 2276 if (client_is_banned(ext_data.fpriv)) { 2277 drm_dbg(&i915->drm, 2278 "client %s[%d] banned from creating ctx\n", 2279 current->comm, task_pid_nr(current)); 2280 return -EIO; 2281 } 2282 2283 ext_data.pc = proto_context_create(i915, args->flags); 2284 if (IS_ERR(ext_data.pc)) 2285 return PTR_ERR(ext_data.pc); 2286 2287 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) { 2288 ret = i915_user_extensions(u64_to_user_ptr(args->extensions), 2289 create_extensions, 2290 ARRAY_SIZE(create_extensions), 2291 &ext_data); 2292 if (ret) 2293 goto err_pc; 2294 } 2295 2296 if (GRAPHICS_VER(i915) > 12) { 2297 struct i915_gem_context *ctx; 2298 2299 /* Get ourselves a context ID */ 2300 ret = xa_alloc(&ext_data.fpriv->context_xa, &id, NULL, 2301 xa_limit_32b, GFP_KERNEL); 2302 if (ret) 2303 goto err_pc; 2304 2305 ctx = i915_gem_create_context(i915, ext_data.pc); 2306 if (IS_ERR(ctx)) { 2307 ret = PTR_ERR(ctx); 2308 goto err_pc; 2309 } 2310 2311 proto_context_close(i915, ext_data.pc); 2312 gem_context_register(ctx, ext_data.fpriv, id); 2313 } else { 2314 ret = proto_context_register(ext_data.fpriv, ext_data.pc, &id); 2315 if (ret < 0) 2316 goto err_pc; 2317 } 2318 2319 args->ctx_id = id; 2320 2321 return 0; 2322 2323 err_pc: 2324 proto_context_close(i915, ext_data.pc); 2325 return ret; 2326 } 2327 2328 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data, 2329 struct drm_file *file) 2330 { 2331 struct drm_i915_gem_context_destroy *args = data; 2332 struct drm_i915_file_private *file_priv = file->driver_priv; 2333 struct i915_gem_proto_context *pc; 2334 struct i915_gem_context *ctx; 2335 2336 if (args->pad != 0) 2337 return -EINVAL; 2338 2339 if (!args->ctx_id) 2340 return -ENOENT; 2341 2342 /* We need to hold the proto-context lock here to prevent races 2343 * with finalize_create_context_locked(). 2344 */ 2345 mutex_lock(&file_priv->proto_context_lock); 2346 ctx = xa_erase(&file_priv->context_xa, args->ctx_id); 2347 pc = xa_erase(&file_priv->proto_context_xa, args->ctx_id); 2348 mutex_unlock(&file_priv->proto_context_lock); 2349 2350 if (!ctx && !pc) 2351 return -ENOENT; 2352 GEM_WARN_ON(ctx && pc); 2353 2354 if (pc) 2355 proto_context_close(file_priv->i915, pc); 2356 2357 if (ctx) 2358 context_close(ctx); 2359 2360 return 0; 2361 } 2362 2363 static int get_sseu(struct i915_gem_context *ctx, 2364 struct drm_i915_gem_context_param *args) 2365 { 2366 struct drm_i915_gem_context_param_sseu user_sseu; 2367 struct intel_context *ce; 2368 unsigned long lookup; 2369 int err; 2370 2371 if (args->size == 0) 2372 goto out; 2373 else if (args->size < sizeof(user_sseu)) 2374 return -EINVAL; 2375 2376 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 2377 sizeof(user_sseu))) 2378 return -EFAULT; 2379 2380 if (user_sseu.rsvd) 2381 return -EINVAL; 2382 2383 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 2384 return -EINVAL; 2385 2386 lookup = 0; 2387 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) 2388 lookup |= LOOKUP_USER_INDEX; 2389 2390 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine); 2391 if (IS_ERR(ce)) 2392 return PTR_ERR(ce); 2393 2394 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */ 2395 if (err) { 2396 intel_context_put(ce); 2397 return err; 2398 } 2399 2400 user_sseu.slice_mask = ce->sseu.slice_mask; 2401 user_sseu.subslice_mask = ce->sseu.subslice_mask; 2402 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice; 2403 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice; 2404 2405 intel_context_unlock_pinned(ce); 2406 intel_context_put(ce); 2407 2408 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu, 2409 sizeof(user_sseu))) 2410 return -EFAULT; 2411 2412 out: 2413 args->size = sizeof(user_sseu); 2414 2415 return 0; 2416 } 2417 2418 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data, 2419 struct drm_file *file) 2420 { 2421 struct drm_i915_file_private *file_priv = file->driver_priv; 2422 struct drm_i915_gem_context_param *args = data; 2423 struct i915_gem_context *ctx; 2424 struct i915_address_space *vm; 2425 int ret = 0; 2426 2427 ctx = i915_gem_context_lookup(file_priv, args->ctx_id); 2428 if (IS_ERR(ctx)) 2429 return PTR_ERR(ctx); 2430 2431 switch (args->param) { 2432 case I915_CONTEXT_PARAM_GTT_SIZE: 2433 args->size = 0; 2434 vm = i915_gem_context_get_eb_vm(ctx); 2435 args->value = vm->total; 2436 i915_vm_put(vm); 2437 2438 break; 2439 2440 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 2441 args->size = 0; 2442 args->value = i915_gem_context_no_error_capture(ctx); 2443 break; 2444 2445 case I915_CONTEXT_PARAM_BANNABLE: 2446 args->size = 0; 2447 args->value = i915_gem_context_is_bannable(ctx); 2448 break; 2449 2450 case I915_CONTEXT_PARAM_RECOVERABLE: 2451 args->size = 0; 2452 args->value = i915_gem_context_is_recoverable(ctx); 2453 break; 2454 2455 case I915_CONTEXT_PARAM_PRIORITY: 2456 args->size = 0; 2457 args->value = ctx->sched.priority; 2458 break; 2459 2460 case I915_CONTEXT_PARAM_SSEU: 2461 ret = get_sseu(ctx, args); 2462 break; 2463 2464 case I915_CONTEXT_PARAM_VM: 2465 ret = get_ppgtt(file_priv, ctx, args); 2466 break; 2467 2468 case I915_CONTEXT_PARAM_PERSISTENCE: 2469 args->size = 0; 2470 args->value = i915_gem_context_is_persistent(ctx); 2471 break; 2472 2473 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 2474 ret = get_protected(ctx, args); 2475 break; 2476 2477 case I915_CONTEXT_PARAM_NO_ZEROMAP: 2478 case I915_CONTEXT_PARAM_BAN_PERIOD: 2479 case I915_CONTEXT_PARAM_ENGINES: 2480 case I915_CONTEXT_PARAM_RINGSIZE: 2481 default: 2482 ret = -EINVAL; 2483 break; 2484 } 2485 2486 i915_gem_context_put(ctx); 2487 return ret; 2488 } 2489 2490 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data, 2491 struct drm_file *file) 2492 { 2493 struct drm_i915_file_private *file_priv = file->driver_priv; 2494 struct drm_i915_gem_context_param *args = data; 2495 struct i915_gem_proto_context *pc; 2496 struct i915_gem_context *ctx; 2497 int ret = 0; 2498 2499 mutex_lock(&file_priv->proto_context_lock); 2500 ctx = __context_lookup(file_priv, args->ctx_id); 2501 if (!ctx) { 2502 pc = xa_load(&file_priv->proto_context_xa, args->ctx_id); 2503 if (pc) { 2504 /* Contexts should be finalized inside 2505 * GEM_CONTEXT_CREATE starting with graphics 2506 * version 13. 2507 */ 2508 WARN_ON(GRAPHICS_VER(file_priv->i915) > 12); 2509 ret = set_proto_ctx_param(file_priv, pc, args); 2510 } else { 2511 ret = -ENOENT; 2512 } 2513 } 2514 mutex_unlock(&file_priv->proto_context_lock); 2515 2516 if (ctx) { 2517 ret = ctx_setparam(file_priv, ctx, args); 2518 i915_gem_context_put(ctx); 2519 } 2520 2521 return ret; 2522 } 2523 2524 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev, 2525 void *data, struct drm_file *file) 2526 { 2527 struct drm_i915_private *i915 = to_i915(dev); 2528 struct drm_i915_reset_stats *args = data; 2529 struct i915_gem_context *ctx; 2530 2531 if (args->flags || args->pad) 2532 return -EINVAL; 2533 2534 ctx = i915_gem_context_lookup(file->driver_priv, args->ctx_id); 2535 if (IS_ERR(ctx)) 2536 return PTR_ERR(ctx); 2537 2538 /* 2539 * We opt for unserialised reads here. This may result in tearing 2540 * in the extremely unlikely event of a GPU hang on this context 2541 * as we are querying them. If we need that extra layer of protection, 2542 * we should wrap the hangstats with a seqlock. 2543 */ 2544 2545 if (capable(CAP_SYS_ADMIN)) 2546 args->reset_count = i915_reset_count(&i915->gpu_error); 2547 else 2548 args->reset_count = 0; 2549 2550 args->batch_active = atomic_read(&ctx->guilty_count); 2551 args->batch_pending = atomic_read(&ctx->active_count); 2552 2553 i915_gem_context_put(ctx); 2554 return 0; 2555 } 2556 2557 /* GEM context-engines iterator: for_each_gem_engine() */ 2558 struct intel_context * 2559 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it) 2560 { 2561 const struct i915_gem_engines *e = it->engines; 2562 struct intel_context *ctx; 2563 2564 if (unlikely(!e)) 2565 return NULL; 2566 2567 do { 2568 if (it->idx >= e->num_engines) 2569 return NULL; 2570 2571 ctx = e->engines[it->idx++]; 2572 } while (!ctx); 2573 2574 return ctx; 2575 } 2576 2577 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2578 #include "selftests/mock_context.c" 2579 #include "selftests/i915_gem_context.c" 2580 #endif 2581 2582 void i915_gem_context_module_exit(void) 2583 { 2584 kmem_cache_destroy(slab_luts); 2585 } 2586 2587 int __init i915_gem_context_module_init(void) 2588 { 2589 slab_luts = KMEM_CACHE(i915_lut_handle, 0); 2590 if (!slab_luts) 2591 return -ENOMEM; 2592 2593 return 0; 2594 } 2595