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_file_private *fpriv, 283 struct drm_i915_private *i915, unsigned int flags) 284 { 285 struct i915_gem_proto_context *pc, *err; 286 287 pc = kzalloc(sizeof(*pc), GFP_KERNEL); 288 if (!pc) 289 return ERR_PTR(-ENOMEM); 290 291 pc->fpriv = fpriv; 292 pc->num_user_engines = -1; 293 pc->user_engines = NULL; 294 pc->user_flags = BIT(UCONTEXT_BANNABLE) | 295 BIT(UCONTEXT_RECOVERABLE); 296 if (i915->params.enable_hangcheck) 297 pc->user_flags |= BIT(UCONTEXT_PERSISTENCE); 298 pc->sched.priority = I915_PRIORITY_NORMAL; 299 300 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) { 301 if (!HAS_EXECLISTS(i915)) { 302 err = ERR_PTR(-EINVAL); 303 goto proto_close; 304 } 305 pc->single_timeline = true; 306 } 307 308 return pc; 309 310 proto_close: 311 proto_context_close(i915, pc); 312 return err; 313 } 314 315 static int proto_context_register_locked(struct drm_i915_file_private *fpriv, 316 struct i915_gem_proto_context *pc, 317 u32 *id) 318 { 319 int ret; 320 void *old; 321 322 lockdep_assert_held(&fpriv->proto_context_lock); 323 324 ret = xa_alloc(&fpriv->context_xa, id, NULL, xa_limit_32b, GFP_KERNEL); 325 if (ret) 326 return ret; 327 328 old = xa_store(&fpriv->proto_context_xa, *id, pc, GFP_KERNEL); 329 if (xa_is_err(old)) { 330 xa_erase(&fpriv->context_xa, *id); 331 return xa_err(old); 332 } 333 WARN_ON(old); 334 335 return 0; 336 } 337 338 static int proto_context_register(struct drm_i915_file_private *fpriv, 339 struct i915_gem_proto_context *pc, 340 u32 *id) 341 { 342 int ret; 343 344 mutex_lock(&fpriv->proto_context_lock); 345 ret = proto_context_register_locked(fpriv, pc, id); 346 mutex_unlock(&fpriv->proto_context_lock); 347 348 return ret; 349 } 350 351 static struct i915_address_space * 352 i915_gem_vm_lookup(struct drm_i915_file_private *file_priv, u32 id) 353 { 354 struct i915_address_space *vm; 355 356 xa_lock(&file_priv->vm_xa); 357 vm = xa_load(&file_priv->vm_xa, id); 358 if (vm) 359 kref_get(&vm->ref); 360 xa_unlock(&file_priv->vm_xa); 361 362 return vm; 363 } 364 365 static int set_proto_ctx_vm(struct drm_i915_file_private *fpriv, 366 struct i915_gem_proto_context *pc, 367 const struct drm_i915_gem_context_param *args) 368 { 369 struct drm_i915_private *i915 = fpriv->i915; 370 struct i915_address_space *vm; 371 372 if (args->size) 373 return -EINVAL; 374 375 if (!HAS_FULL_PPGTT(i915)) 376 return -ENODEV; 377 378 if (upper_32_bits(args->value)) 379 return -ENOENT; 380 381 vm = i915_gem_vm_lookup(fpriv, args->value); 382 if (!vm) 383 return -ENOENT; 384 385 if (pc->vm) 386 i915_vm_put(pc->vm); 387 pc->vm = vm; 388 389 return 0; 390 } 391 392 struct set_proto_ctx_engines { 393 struct drm_i915_private *i915; 394 unsigned num_engines; 395 struct i915_gem_proto_engine *engines; 396 }; 397 398 static int 399 set_proto_ctx_engines_balance(struct i915_user_extension __user *base, 400 void *data) 401 { 402 struct i915_context_engines_load_balance __user *ext = 403 container_of_user(base, typeof(*ext), base); 404 const struct set_proto_ctx_engines *set = data; 405 struct drm_i915_private *i915 = set->i915; 406 struct intel_engine_cs **siblings; 407 u16 num_siblings, idx; 408 unsigned int n; 409 int err; 410 411 if (!HAS_EXECLISTS(i915)) 412 return -ENODEV; 413 414 if (get_user(idx, &ext->engine_index)) 415 return -EFAULT; 416 417 if (idx >= set->num_engines) { 418 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n", 419 idx, set->num_engines); 420 return -EINVAL; 421 } 422 423 idx = array_index_nospec(idx, set->num_engines); 424 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_INVALID) { 425 drm_dbg(&i915->drm, 426 "Invalid placement[%d], already occupied\n", idx); 427 return -EEXIST; 428 } 429 430 if (get_user(num_siblings, &ext->num_siblings)) 431 return -EFAULT; 432 433 err = check_user_mbz(&ext->flags); 434 if (err) 435 return err; 436 437 err = check_user_mbz(&ext->mbz64); 438 if (err) 439 return err; 440 441 if (num_siblings == 0) 442 return 0; 443 444 siblings = kmalloc_array(num_siblings, sizeof(*siblings), GFP_KERNEL); 445 if (!siblings) 446 return -ENOMEM; 447 448 for (n = 0; n < num_siblings; n++) { 449 struct i915_engine_class_instance ci; 450 451 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) { 452 err = -EFAULT; 453 goto err_siblings; 454 } 455 456 siblings[n] = intel_engine_lookup_user(i915, 457 ci.engine_class, 458 ci.engine_instance); 459 if (!siblings[n]) { 460 drm_dbg(&i915->drm, 461 "Invalid sibling[%d]: { class:%d, inst:%d }\n", 462 n, ci.engine_class, ci.engine_instance); 463 err = -EINVAL; 464 goto err_siblings; 465 } 466 } 467 468 if (num_siblings == 1) { 469 set->engines[idx].type = I915_GEM_ENGINE_TYPE_PHYSICAL; 470 set->engines[idx].engine = siblings[0]; 471 kfree(siblings); 472 } else { 473 set->engines[idx].type = I915_GEM_ENGINE_TYPE_BALANCED; 474 set->engines[idx].num_siblings = num_siblings; 475 set->engines[idx].siblings = siblings; 476 } 477 478 return 0; 479 480 err_siblings: 481 kfree(siblings); 482 483 return err; 484 } 485 486 static int 487 set_proto_ctx_engines_bond(struct i915_user_extension __user *base, void *data) 488 { 489 struct i915_context_engines_bond __user *ext = 490 container_of_user(base, typeof(*ext), base); 491 const struct set_proto_ctx_engines *set = data; 492 struct drm_i915_private *i915 = set->i915; 493 struct i915_engine_class_instance ci; 494 struct intel_engine_cs *master; 495 u16 idx, num_bonds; 496 int err, n; 497 498 if (GRAPHICS_VER(i915) >= 12 && !IS_TIGERLAKE(i915) && 499 !IS_ROCKETLAKE(i915) && !IS_ALDERLAKE_S(i915)) { 500 drm_dbg(&i915->drm, 501 "Bonding not supported on this platform\n"); 502 return -ENODEV; 503 } 504 505 if (get_user(idx, &ext->virtual_index)) 506 return -EFAULT; 507 508 if (idx >= set->num_engines) { 509 drm_dbg(&i915->drm, 510 "Invalid index for virtual engine: %d >= %d\n", 511 idx, set->num_engines); 512 return -EINVAL; 513 } 514 515 idx = array_index_nospec(idx, set->num_engines); 516 if (set->engines[idx].type == I915_GEM_ENGINE_TYPE_INVALID) { 517 drm_dbg(&i915->drm, "Invalid engine at %d\n", idx); 518 return -EINVAL; 519 } 520 521 if (set->engines[idx].type != I915_GEM_ENGINE_TYPE_PHYSICAL) { 522 drm_dbg(&i915->drm, 523 "Bonding with virtual engines not allowed\n"); 524 return -EINVAL; 525 } 526 527 err = check_user_mbz(&ext->flags); 528 if (err) 529 return err; 530 531 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) { 532 err = check_user_mbz(&ext->mbz64[n]); 533 if (err) 534 return err; 535 } 536 537 if (copy_from_user(&ci, &ext->master, sizeof(ci))) 538 return -EFAULT; 539 540 master = intel_engine_lookup_user(i915, 541 ci.engine_class, 542 ci.engine_instance); 543 if (!master) { 544 drm_dbg(&i915->drm, 545 "Unrecognised master engine: { class:%u, instance:%u }\n", 546 ci.engine_class, ci.engine_instance); 547 return -EINVAL; 548 } 549 550 if (intel_engine_uses_guc(master)) { 551 drm_dbg(&i915->drm, "bonding extension not supported with GuC submission"); 552 return -ENODEV; 553 } 554 555 if (get_user(num_bonds, &ext->num_bonds)) 556 return -EFAULT; 557 558 for (n = 0; n < num_bonds; n++) { 559 struct intel_engine_cs *bond; 560 561 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) 562 return -EFAULT; 563 564 bond = intel_engine_lookup_user(i915, 565 ci.engine_class, 566 ci.engine_instance); 567 if (!bond) { 568 drm_dbg(&i915->drm, 569 "Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n", 570 n, ci.engine_class, ci.engine_instance); 571 return -EINVAL; 572 } 573 } 574 575 return 0; 576 } 577 578 static int 579 set_proto_ctx_engines_parallel_submit(struct i915_user_extension __user *base, 580 void *data) 581 { 582 struct i915_context_engines_parallel_submit __user *ext = 583 container_of_user(base, typeof(*ext), base); 584 const struct set_proto_ctx_engines *set = data; 585 struct drm_i915_private *i915 = set->i915; 586 struct i915_engine_class_instance prev_engine; 587 u64 flags; 588 int err = 0, n, i, j; 589 u16 slot, width, num_siblings; 590 struct intel_engine_cs **siblings = NULL; 591 intel_engine_mask_t prev_mask; 592 593 if (get_user(slot, &ext->engine_index)) 594 return -EFAULT; 595 596 if (get_user(width, &ext->width)) 597 return -EFAULT; 598 599 if (get_user(num_siblings, &ext->num_siblings)) 600 return -EFAULT; 601 602 if (!intel_uc_uses_guc_submission(&to_gt(i915)->uc) && 603 num_siblings != 1) { 604 drm_dbg(&i915->drm, "Only 1 sibling (%d) supported in non-GuC mode\n", 605 num_siblings); 606 return -EINVAL; 607 } 608 609 if (slot >= set->num_engines) { 610 drm_dbg(&i915->drm, "Invalid placement value, %d >= %d\n", 611 slot, set->num_engines); 612 return -EINVAL; 613 } 614 615 if (set->engines[slot].type != I915_GEM_ENGINE_TYPE_INVALID) { 616 drm_dbg(&i915->drm, 617 "Invalid placement[%d], already occupied\n", slot); 618 return -EINVAL; 619 } 620 621 if (get_user(flags, &ext->flags)) 622 return -EFAULT; 623 624 if (flags) { 625 drm_dbg(&i915->drm, "Unknown flags 0x%02llx", flags); 626 return -EINVAL; 627 } 628 629 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) { 630 err = check_user_mbz(&ext->mbz64[n]); 631 if (err) 632 return err; 633 } 634 635 if (width < 2) { 636 drm_dbg(&i915->drm, "Width (%d) < 2\n", width); 637 return -EINVAL; 638 } 639 640 if (num_siblings < 1) { 641 drm_dbg(&i915->drm, "Number siblings (%d) < 1\n", 642 num_siblings); 643 return -EINVAL; 644 } 645 646 siblings = kmalloc_array(num_siblings * width, 647 sizeof(*siblings), 648 GFP_KERNEL); 649 if (!siblings) 650 return -ENOMEM; 651 652 /* Create contexts / engines */ 653 for (i = 0; i < width; ++i) { 654 intel_engine_mask_t current_mask = 0; 655 656 for (j = 0; j < num_siblings; ++j) { 657 struct i915_engine_class_instance ci; 658 659 n = i * num_siblings + j; 660 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) { 661 err = -EFAULT; 662 goto out_err; 663 } 664 665 siblings[n] = 666 intel_engine_lookup_user(i915, ci.engine_class, 667 ci.engine_instance); 668 if (!siblings[n]) { 669 drm_dbg(&i915->drm, 670 "Invalid sibling[%d]: { class:%d, inst:%d }\n", 671 n, ci.engine_class, ci.engine_instance); 672 err = -EINVAL; 673 goto out_err; 674 } 675 676 /* 677 * We don't support breadcrumb handshake on these 678 * classes 679 */ 680 if (siblings[n]->class == RENDER_CLASS || 681 siblings[n]->class == COMPUTE_CLASS) { 682 err = -EINVAL; 683 goto out_err; 684 } 685 686 if (n) { 687 if (prev_engine.engine_class != 688 ci.engine_class) { 689 drm_dbg(&i915->drm, 690 "Mismatched class %d, %d\n", 691 prev_engine.engine_class, 692 ci.engine_class); 693 err = -EINVAL; 694 goto out_err; 695 } 696 } 697 698 prev_engine = ci; 699 current_mask |= siblings[n]->logical_mask; 700 } 701 702 if (i > 0) { 703 if (current_mask != prev_mask << 1) { 704 drm_dbg(&i915->drm, 705 "Non contiguous logical mask 0x%x, 0x%x\n", 706 prev_mask, current_mask); 707 err = -EINVAL; 708 goto out_err; 709 } 710 } 711 prev_mask = current_mask; 712 } 713 714 set->engines[slot].type = I915_GEM_ENGINE_TYPE_PARALLEL; 715 set->engines[slot].num_siblings = num_siblings; 716 set->engines[slot].width = width; 717 set->engines[slot].siblings = siblings; 718 719 return 0; 720 721 out_err: 722 kfree(siblings); 723 724 return err; 725 } 726 727 static const i915_user_extension_fn set_proto_ctx_engines_extensions[] = { 728 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_proto_ctx_engines_balance, 729 [I915_CONTEXT_ENGINES_EXT_BOND] = set_proto_ctx_engines_bond, 730 [I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT] = 731 set_proto_ctx_engines_parallel_submit, 732 }; 733 734 static int set_proto_ctx_engines(struct drm_i915_file_private *fpriv, 735 struct i915_gem_proto_context *pc, 736 const struct drm_i915_gem_context_param *args) 737 { 738 struct drm_i915_private *i915 = fpriv->i915; 739 struct set_proto_ctx_engines set = { .i915 = i915 }; 740 struct i915_context_param_engines __user *user = 741 u64_to_user_ptr(args->value); 742 unsigned int n; 743 u64 extensions; 744 int err; 745 746 if (pc->num_user_engines >= 0) { 747 drm_dbg(&i915->drm, "Cannot set engines twice"); 748 return -EINVAL; 749 } 750 751 if (args->size < sizeof(*user) || 752 !IS_ALIGNED(args->size - sizeof(*user), sizeof(*user->engines))) { 753 drm_dbg(&i915->drm, "Invalid size for engine array: %d\n", 754 args->size); 755 return -EINVAL; 756 } 757 758 set.num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines); 759 /* RING_MASK has no shift so we can use it directly here */ 760 if (set.num_engines > I915_EXEC_RING_MASK + 1) 761 return -EINVAL; 762 763 set.engines = kmalloc_array(set.num_engines, sizeof(*set.engines), GFP_KERNEL); 764 if (!set.engines) 765 return -ENOMEM; 766 767 for (n = 0; n < set.num_engines; n++) { 768 struct i915_engine_class_instance ci; 769 struct intel_engine_cs *engine; 770 771 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) { 772 kfree(set.engines); 773 return -EFAULT; 774 } 775 776 memset(&set.engines[n], 0, sizeof(set.engines[n])); 777 778 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID && 779 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE) 780 continue; 781 782 engine = intel_engine_lookup_user(i915, 783 ci.engine_class, 784 ci.engine_instance); 785 if (!engine) { 786 drm_dbg(&i915->drm, 787 "Invalid engine[%d]: { class:%d, instance:%d }\n", 788 n, ci.engine_class, ci.engine_instance); 789 kfree(set.engines); 790 return -ENOENT; 791 } 792 793 set.engines[n].type = I915_GEM_ENGINE_TYPE_PHYSICAL; 794 set.engines[n].engine = engine; 795 } 796 797 err = -EFAULT; 798 if (!get_user(extensions, &user->extensions)) 799 err = i915_user_extensions(u64_to_user_ptr(extensions), 800 set_proto_ctx_engines_extensions, 801 ARRAY_SIZE(set_proto_ctx_engines_extensions), 802 &set); 803 if (err) { 804 kfree(set.engines); 805 return err; 806 } 807 808 pc->num_user_engines = set.num_engines; 809 pc->user_engines = set.engines; 810 811 return 0; 812 } 813 814 static int set_proto_ctx_sseu(struct drm_i915_file_private *fpriv, 815 struct i915_gem_proto_context *pc, 816 struct drm_i915_gem_context_param *args) 817 { 818 struct drm_i915_private *i915 = fpriv->i915; 819 struct drm_i915_gem_context_param_sseu user_sseu; 820 struct intel_sseu *sseu; 821 int ret; 822 823 if (args->size < sizeof(user_sseu)) 824 return -EINVAL; 825 826 if (GRAPHICS_VER(i915) != 11) 827 return -ENODEV; 828 829 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 830 sizeof(user_sseu))) 831 return -EFAULT; 832 833 if (user_sseu.rsvd) 834 return -EINVAL; 835 836 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 837 return -EINVAL; 838 839 if (!!(user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) != (pc->num_user_engines >= 0)) 840 return -EINVAL; 841 842 if (pc->num_user_engines >= 0) { 843 int idx = user_sseu.engine.engine_instance; 844 struct i915_gem_proto_engine *pe; 845 846 if (idx >= pc->num_user_engines) 847 return -EINVAL; 848 849 idx = array_index_nospec(idx, pc->num_user_engines); 850 pe = &pc->user_engines[idx]; 851 852 /* Only render engine supports RPCS configuration. */ 853 if (pe->engine->class != RENDER_CLASS) 854 return -EINVAL; 855 856 sseu = &pe->sseu; 857 } else { 858 /* Only render engine supports RPCS configuration. */ 859 if (user_sseu.engine.engine_class != I915_ENGINE_CLASS_RENDER) 860 return -EINVAL; 861 862 /* There is only one render engine */ 863 if (user_sseu.engine.engine_instance != 0) 864 return -EINVAL; 865 866 sseu = &pc->legacy_rcs_sseu; 867 } 868 869 ret = i915_gem_user_to_context_sseu(to_gt(i915), &user_sseu, sseu); 870 if (ret) 871 return ret; 872 873 args->size = sizeof(user_sseu); 874 875 return 0; 876 } 877 878 static int set_proto_ctx_param(struct drm_i915_file_private *fpriv, 879 struct i915_gem_proto_context *pc, 880 struct drm_i915_gem_context_param *args) 881 { 882 int ret = 0; 883 884 switch (args->param) { 885 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 886 if (args->size) 887 ret = -EINVAL; 888 else if (args->value) 889 pc->user_flags |= BIT(UCONTEXT_NO_ERROR_CAPTURE); 890 else 891 pc->user_flags &= ~BIT(UCONTEXT_NO_ERROR_CAPTURE); 892 break; 893 894 case I915_CONTEXT_PARAM_BANNABLE: 895 if (args->size) 896 ret = -EINVAL; 897 else if (!capable(CAP_SYS_ADMIN) && !args->value) 898 ret = -EPERM; 899 else if (args->value) 900 pc->user_flags |= BIT(UCONTEXT_BANNABLE); 901 else if (pc->uses_protected_content) 902 ret = -EPERM; 903 else 904 pc->user_flags &= ~BIT(UCONTEXT_BANNABLE); 905 break; 906 907 case I915_CONTEXT_PARAM_RECOVERABLE: 908 if (args->size) 909 ret = -EINVAL; 910 else if (!args->value) 911 pc->user_flags &= ~BIT(UCONTEXT_RECOVERABLE); 912 else if (pc->uses_protected_content) 913 ret = -EPERM; 914 else 915 pc->user_flags |= BIT(UCONTEXT_RECOVERABLE); 916 break; 917 918 case I915_CONTEXT_PARAM_PRIORITY: 919 ret = validate_priority(fpriv->i915, args); 920 if (!ret) 921 pc->sched.priority = args->value; 922 break; 923 924 case I915_CONTEXT_PARAM_SSEU: 925 ret = set_proto_ctx_sseu(fpriv, pc, args); 926 break; 927 928 case I915_CONTEXT_PARAM_VM: 929 ret = set_proto_ctx_vm(fpriv, pc, args); 930 break; 931 932 case I915_CONTEXT_PARAM_ENGINES: 933 ret = set_proto_ctx_engines(fpriv, pc, args); 934 break; 935 936 case I915_CONTEXT_PARAM_PERSISTENCE: 937 if (args->size) 938 ret = -EINVAL; 939 else 940 ret = proto_context_set_persistence(fpriv->i915, pc, 941 args->value); 942 break; 943 944 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 945 ret = proto_context_set_protected(fpriv->i915, pc, 946 args->value); 947 break; 948 949 case I915_CONTEXT_PARAM_NO_ZEROMAP: 950 case I915_CONTEXT_PARAM_BAN_PERIOD: 951 case I915_CONTEXT_PARAM_RINGSIZE: 952 default: 953 ret = -EINVAL; 954 break; 955 } 956 957 return ret; 958 } 959 960 static int intel_context_set_gem(struct intel_context *ce, 961 struct i915_gem_context *ctx, 962 struct intel_sseu sseu) 963 { 964 int ret = 0; 965 966 GEM_BUG_ON(rcu_access_pointer(ce->gem_context)); 967 RCU_INIT_POINTER(ce->gem_context, ctx); 968 969 GEM_BUG_ON(intel_context_is_pinned(ce)); 970 971 if (ce->engine->class == COMPUTE_CLASS) 972 ce->ring_size = SZ_512K; 973 else 974 ce->ring_size = SZ_16K; 975 976 i915_vm_put(ce->vm); 977 ce->vm = i915_gem_context_get_eb_vm(ctx); 978 979 if (ctx->sched.priority >= I915_PRIORITY_NORMAL && 980 intel_engine_has_timeslices(ce->engine) && 981 intel_engine_has_semaphores(ce->engine)) 982 __set_bit(CONTEXT_USE_SEMAPHORES, &ce->flags); 983 984 if (CONFIG_DRM_I915_REQUEST_TIMEOUT && 985 ctx->i915->params.request_timeout_ms) { 986 unsigned int timeout_ms = ctx->i915->params.request_timeout_ms; 987 988 intel_context_set_watchdog_us(ce, (u64)timeout_ms * 1000); 989 } 990 991 /* A valid SSEU has no zero fields */ 992 if (sseu.slice_mask && !WARN_ON(ce->engine->class != RENDER_CLASS)) 993 ret = intel_context_reconfigure_sseu(ce, sseu); 994 995 return ret; 996 } 997 998 static void __unpin_engines(struct i915_gem_engines *e, unsigned int count) 999 { 1000 while (count--) { 1001 struct intel_context *ce = e->engines[count], *child; 1002 1003 if (!ce || !test_bit(CONTEXT_PERMA_PIN, &ce->flags)) 1004 continue; 1005 1006 for_each_child(ce, child) 1007 intel_context_unpin(child); 1008 intel_context_unpin(ce); 1009 } 1010 } 1011 1012 static void unpin_engines(struct i915_gem_engines *e) 1013 { 1014 __unpin_engines(e, e->num_engines); 1015 } 1016 1017 static void __free_engines(struct i915_gem_engines *e, unsigned int count) 1018 { 1019 while (count--) { 1020 if (!e->engines[count]) 1021 continue; 1022 1023 intel_context_put(e->engines[count]); 1024 } 1025 kfree(e); 1026 } 1027 1028 static void free_engines(struct i915_gem_engines *e) 1029 { 1030 __free_engines(e, e->num_engines); 1031 } 1032 1033 static void free_engines_rcu(struct rcu_head *rcu) 1034 { 1035 struct i915_gem_engines *engines = 1036 container_of(rcu, struct i915_gem_engines, rcu); 1037 1038 i915_sw_fence_fini(&engines->fence); 1039 free_engines(engines); 1040 } 1041 1042 static void accumulate_runtime(struct i915_drm_client *client, 1043 struct i915_gem_engines *engines) 1044 { 1045 struct i915_gem_engines_iter it; 1046 struct intel_context *ce; 1047 1048 if (!client) 1049 return; 1050 1051 /* Transfer accumulated runtime to the parent GEM context. */ 1052 for_each_gem_engine(ce, engines, it) { 1053 unsigned int class = ce->engine->uabi_class; 1054 1055 GEM_BUG_ON(class >= ARRAY_SIZE(client->past_runtime)); 1056 atomic64_add(intel_context_get_total_runtime_ns(ce), 1057 &client->past_runtime[class]); 1058 } 1059 } 1060 1061 static int 1062 engines_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) 1063 { 1064 struct i915_gem_engines *engines = 1065 container_of(fence, typeof(*engines), fence); 1066 struct i915_gem_context *ctx = engines->ctx; 1067 1068 switch (state) { 1069 case FENCE_COMPLETE: 1070 if (!list_empty(&engines->link)) { 1071 unsigned long flags; 1072 1073 spin_lock_irqsave(&ctx->stale.lock, flags); 1074 list_del(&engines->link); 1075 spin_unlock_irqrestore(&ctx->stale.lock, flags); 1076 } 1077 accumulate_runtime(ctx->client, engines); 1078 i915_gem_context_put(ctx); 1079 1080 break; 1081 1082 case FENCE_FREE: 1083 init_rcu_head(&engines->rcu); 1084 call_rcu(&engines->rcu, free_engines_rcu); 1085 break; 1086 } 1087 1088 return NOTIFY_DONE; 1089 } 1090 1091 static struct i915_gem_engines *alloc_engines(unsigned int count) 1092 { 1093 struct i915_gem_engines *e; 1094 1095 e = kzalloc(struct_size(e, engines, count), GFP_KERNEL); 1096 if (!e) 1097 return NULL; 1098 1099 i915_sw_fence_init(&e->fence, engines_notify); 1100 return e; 1101 } 1102 1103 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx, 1104 struct intel_sseu rcs_sseu) 1105 { 1106 const unsigned int max = I915_NUM_ENGINES; 1107 struct intel_engine_cs *engine; 1108 struct i915_gem_engines *e, *err; 1109 1110 e = alloc_engines(max); 1111 if (!e) 1112 return ERR_PTR(-ENOMEM); 1113 1114 for_each_uabi_engine(engine, ctx->i915) { 1115 struct intel_context *ce; 1116 struct intel_sseu sseu = {}; 1117 int ret; 1118 1119 if (engine->legacy_idx == INVALID_ENGINE) 1120 continue; 1121 1122 GEM_BUG_ON(engine->legacy_idx >= max); 1123 GEM_BUG_ON(e->engines[engine->legacy_idx]); 1124 1125 ce = intel_context_create(engine); 1126 if (IS_ERR(ce)) { 1127 err = ERR_CAST(ce); 1128 goto free_engines; 1129 } 1130 1131 e->engines[engine->legacy_idx] = ce; 1132 e->num_engines = max(e->num_engines, engine->legacy_idx + 1); 1133 1134 if (engine->class == RENDER_CLASS) 1135 sseu = rcs_sseu; 1136 1137 ret = intel_context_set_gem(ce, ctx, sseu); 1138 if (ret) { 1139 err = ERR_PTR(ret); 1140 goto free_engines; 1141 } 1142 1143 } 1144 1145 return e; 1146 1147 free_engines: 1148 free_engines(e); 1149 return err; 1150 } 1151 1152 static int perma_pin_contexts(struct intel_context *ce) 1153 { 1154 struct intel_context *child; 1155 int i = 0, j = 0, ret; 1156 1157 GEM_BUG_ON(!intel_context_is_parent(ce)); 1158 1159 ret = intel_context_pin(ce); 1160 if (unlikely(ret)) 1161 return ret; 1162 1163 for_each_child(ce, child) { 1164 ret = intel_context_pin(child); 1165 if (unlikely(ret)) 1166 goto unwind; 1167 ++i; 1168 } 1169 1170 set_bit(CONTEXT_PERMA_PIN, &ce->flags); 1171 1172 return 0; 1173 1174 unwind: 1175 intel_context_unpin(ce); 1176 for_each_child(ce, child) { 1177 if (j++ < i) 1178 intel_context_unpin(child); 1179 else 1180 break; 1181 } 1182 1183 return ret; 1184 } 1185 1186 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx, 1187 unsigned int num_engines, 1188 struct i915_gem_proto_engine *pe) 1189 { 1190 struct i915_gem_engines *e, *err; 1191 unsigned int n; 1192 1193 e = alloc_engines(num_engines); 1194 if (!e) 1195 return ERR_PTR(-ENOMEM); 1196 e->num_engines = num_engines; 1197 1198 for (n = 0; n < num_engines; n++) { 1199 struct intel_context *ce, *child; 1200 int ret; 1201 1202 switch (pe[n].type) { 1203 case I915_GEM_ENGINE_TYPE_PHYSICAL: 1204 ce = intel_context_create(pe[n].engine); 1205 break; 1206 1207 case I915_GEM_ENGINE_TYPE_BALANCED: 1208 ce = intel_engine_create_virtual(pe[n].siblings, 1209 pe[n].num_siblings, 0); 1210 break; 1211 1212 case I915_GEM_ENGINE_TYPE_PARALLEL: 1213 ce = intel_engine_create_parallel(pe[n].siblings, 1214 pe[n].num_siblings, 1215 pe[n].width); 1216 break; 1217 1218 case I915_GEM_ENGINE_TYPE_INVALID: 1219 default: 1220 GEM_WARN_ON(pe[n].type != I915_GEM_ENGINE_TYPE_INVALID); 1221 continue; 1222 } 1223 1224 if (IS_ERR(ce)) { 1225 err = ERR_CAST(ce); 1226 goto free_engines; 1227 } 1228 1229 e->engines[n] = ce; 1230 1231 ret = intel_context_set_gem(ce, ctx, pe->sseu); 1232 if (ret) { 1233 err = ERR_PTR(ret); 1234 goto free_engines; 1235 } 1236 for_each_child(ce, child) { 1237 ret = intel_context_set_gem(child, ctx, pe->sseu); 1238 if (ret) { 1239 err = ERR_PTR(ret); 1240 goto free_engines; 1241 } 1242 } 1243 1244 /* 1245 * XXX: Must be done after calling intel_context_set_gem as that 1246 * function changes the ring size. The ring is allocated when 1247 * the context is pinned. If the ring size is changed after 1248 * allocation we have a mismatch of the ring size and will cause 1249 * the context to hang. Presumably with a bit of reordering we 1250 * could move the perma-pin step to the backend function 1251 * intel_engine_create_parallel. 1252 */ 1253 if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) { 1254 ret = perma_pin_contexts(ce); 1255 if (ret) { 1256 err = ERR_PTR(ret); 1257 goto free_engines; 1258 } 1259 } 1260 } 1261 1262 return e; 1263 1264 free_engines: 1265 free_engines(e); 1266 return err; 1267 } 1268 1269 static void i915_gem_context_release_work(struct work_struct *work) 1270 { 1271 struct i915_gem_context *ctx = container_of(work, typeof(*ctx), 1272 release_work); 1273 struct i915_address_space *vm; 1274 1275 trace_i915_context_free(ctx); 1276 GEM_BUG_ON(!i915_gem_context_is_closed(ctx)); 1277 1278 spin_lock(&ctx->i915->gem.contexts.lock); 1279 list_del(&ctx->link); 1280 spin_unlock(&ctx->i915->gem.contexts.lock); 1281 1282 if (ctx->syncobj) 1283 drm_syncobj_put(ctx->syncobj); 1284 1285 vm = ctx->vm; 1286 if (vm) 1287 i915_vm_put(vm); 1288 1289 if (ctx->pxp_wakeref) 1290 intel_runtime_pm_put(&ctx->i915->runtime_pm, ctx->pxp_wakeref); 1291 1292 if (ctx->client) 1293 i915_drm_client_put(ctx->client); 1294 1295 mutex_destroy(&ctx->engines_mutex); 1296 mutex_destroy(&ctx->lut_mutex); 1297 1298 put_pid(ctx->pid); 1299 mutex_destroy(&ctx->mutex); 1300 1301 kfree_rcu(ctx, rcu); 1302 } 1303 1304 void i915_gem_context_release(struct kref *ref) 1305 { 1306 struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref); 1307 1308 queue_work(ctx->i915->wq, &ctx->release_work); 1309 } 1310 1311 static inline struct i915_gem_engines * 1312 __context_engines_static(const struct i915_gem_context *ctx) 1313 { 1314 return rcu_dereference_protected(ctx->engines, true); 1315 } 1316 1317 static void __reset_context(struct i915_gem_context *ctx, 1318 struct intel_engine_cs *engine) 1319 { 1320 intel_gt_handle_error(engine->gt, engine->mask, 0, 1321 "context closure in %s", ctx->name); 1322 } 1323 1324 static bool __cancel_engine(struct intel_engine_cs *engine) 1325 { 1326 /* 1327 * Send a "high priority pulse" down the engine to cause the 1328 * current request to be momentarily preempted. (If it fails to 1329 * be preempted, it will be reset). As we have marked our context 1330 * as banned, any incomplete request, including any running, will 1331 * be skipped following the preemption. 1332 * 1333 * If there is no hangchecking (one of the reasons why we try to 1334 * cancel the context) and no forced preemption, there may be no 1335 * means by which we reset the GPU and evict the persistent hog. 1336 * Ergo if we are unable to inject a preemptive pulse that can 1337 * kill the banned context, we fallback to doing a local reset 1338 * instead. 1339 */ 1340 return intel_engine_pulse(engine) == 0; 1341 } 1342 1343 static struct intel_engine_cs *active_engine(struct intel_context *ce) 1344 { 1345 struct intel_engine_cs *engine = NULL; 1346 struct i915_request *rq; 1347 1348 if (intel_context_has_inflight(ce)) 1349 return intel_context_inflight(ce); 1350 1351 if (!ce->timeline) 1352 return NULL; 1353 1354 /* 1355 * rq->link is only SLAB_TYPESAFE_BY_RCU, we need to hold a reference 1356 * to the request to prevent it being transferred to a new timeline 1357 * (and onto a new timeline->requests list). 1358 */ 1359 rcu_read_lock(); 1360 list_for_each_entry_reverse(rq, &ce->timeline->requests, link) { 1361 bool found; 1362 1363 /* timeline is already completed upto this point? */ 1364 if (!i915_request_get_rcu(rq)) 1365 break; 1366 1367 /* Check with the backend if the request is inflight */ 1368 found = true; 1369 if (likely(rcu_access_pointer(rq->timeline) == ce->timeline)) 1370 found = i915_request_active_engine(rq, &engine); 1371 1372 i915_request_put(rq); 1373 if (found) 1374 break; 1375 } 1376 rcu_read_unlock(); 1377 1378 return engine; 1379 } 1380 1381 static void 1382 kill_engines(struct i915_gem_engines *engines, bool exit, bool persistent) 1383 { 1384 struct i915_gem_engines_iter it; 1385 struct intel_context *ce; 1386 1387 /* 1388 * Map the user's engine back to the actual engines; one virtual 1389 * engine will be mapped to multiple engines, and using ctx->engine[] 1390 * the same engine may be have multiple instances in the user's map. 1391 * However, we only care about pending requests, so only include 1392 * engines on which there are incomplete requests. 1393 */ 1394 for_each_gem_engine(ce, engines, it) { 1395 struct intel_engine_cs *engine; 1396 1397 if ((exit || !persistent) && intel_context_revoke(ce)) 1398 continue; /* Already marked. */ 1399 1400 /* 1401 * Check the current active state of this context; if we 1402 * are currently executing on the GPU we need to evict 1403 * ourselves. On the other hand, if we haven't yet been 1404 * submitted to the GPU or if everything is complete, 1405 * we have nothing to do. 1406 */ 1407 engine = active_engine(ce); 1408 1409 /* First attempt to gracefully cancel the context */ 1410 if (engine && !__cancel_engine(engine) && (exit || !persistent)) 1411 /* 1412 * If we are unable to send a preemptive pulse to bump 1413 * the context from the GPU, we have to resort to a full 1414 * reset. We hope the collateral damage is worth it. 1415 */ 1416 __reset_context(engines->ctx, engine); 1417 } 1418 } 1419 1420 static void kill_context(struct i915_gem_context *ctx) 1421 { 1422 struct i915_gem_engines *pos, *next; 1423 1424 spin_lock_irq(&ctx->stale.lock); 1425 GEM_BUG_ON(!i915_gem_context_is_closed(ctx)); 1426 list_for_each_entry_safe(pos, next, &ctx->stale.engines, link) { 1427 if (!i915_sw_fence_await(&pos->fence)) { 1428 list_del_init(&pos->link); 1429 continue; 1430 } 1431 1432 spin_unlock_irq(&ctx->stale.lock); 1433 1434 kill_engines(pos, !ctx->i915->params.enable_hangcheck, 1435 i915_gem_context_is_persistent(ctx)); 1436 1437 spin_lock_irq(&ctx->stale.lock); 1438 GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence)); 1439 list_safe_reset_next(pos, next, link); 1440 list_del_init(&pos->link); /* decouple from FENCE_COMPLETE */ 1441 1442 i915_sw_fence_complete(&pos->fence); 1443 } 1444 spin_unlock_irq(&ctx->stale.lock); 1445 } 1446 1447 static void engines_idle_release(struct i915_gem_context *ctx, 1448 struct i915_gem_engines *engines) 1449 { 1450 struct i915_gem_engines_iter it; 1451 struct intel_context *ce; 1452 1453 INIT_LIST_HEAD(&engines->link); 1454 1455 engines->ctx = i915_gem_context_get(ctx); 1456 1457 for_each_gem_engine(ce, engines, it) { 1458 int err; 1459 1460 /* serialises with execbuf */ 1461 intel_context_close(ce); 1462 if (!intel_context_pin_if_active(ce)) 1463 continue; 1464 1465 /* Wait until context is finally scheduled out and retired */ 1466 err = i915_sw_fence_await_active(&engines->fence, 1467 &ce->active, 1468 I915_ACTIVE_AWAIT_BARRIER); 1469 intel_context_unpin(ce); 1470 if (err) 1471 goto kill; 1472 } 1473 1474 spin_lock_irq(&ctx->stale.lock); 1475 if (!i915_gem_context_is_closed(ctx)) 1476 list_add_tail(&engines->link, &ctx->stale.engines); 1477 spin_unlock_irq(&ctx->stale.lock); 1478 1479 kill: 1480 if (list_empty(&engines->link)) /* raced, already closed */ 1481 kill_engines(engines, true, 1482 i915_gem_context_is_persistent(ctx)); 1483 1484 i915_sw_fence_commit(&engines->fence); 1485 } 1486 1487 static void set_closed_name(struct i915_gem_context *ctx) 1488 { 1489 char *s; 1490 1491 /* Replace '[]' with '<>' to indicate closed in debug prints */ 1492 1493 s = strrchr(ctx->name, '['); 1494 if (!s) 1495 return; 1496 1497 *s = '<'; 1498 1499 s = strchr(s + 1, ']'); 1500 if (s) 1501 *s = '>'; 1502 } 1503 1504 static void context_close(struct i915_gem_context *ctx) 1505 { 1506 struct i915_drm_client *client; 1507 1508 /* Flush any concurrent set_engines() */ 1509 mutex_lock(&ctx->engines_mutex); 1510 unpin_engines(__context_engines_static(ctx)); 1511 engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1)); 1512 i915_gem_context_set_closed(ctx); 1513 mutex_unlock(&ctx->engines_mutex); 1514 1515 mutex_lock(&ctx->mutex); 1516 1517 set_closed_name(ctx); 1518 1519 /* 1520 * The LUT uses the VMA as a backpointer to unref the object, 1521 * so we need to clear the LUT before we close all the VMA (inside 1522 * the ppgtt). 1523 */ 1524 lut_close(ctx); 1525 1526 ctx->file_priv = ERR_PTR(-EBADF); 1527 1528 client = ctx->client; 1529 if (client) { 1530 spin_lock(&client->ctx_lock); 1531 list_del_rcu(&ctx->client_link); 1532 spin_unlock(&client->ctx_lock); 1533 } 1534 1535 mutex_unlock(&ctx->mutex); 1536 1537 /* 1538 * If the user has disabled hangchecking, we can not be sure that 1539 * the batches will ever complete after the context is closed, 1540 * keeping the context and all resources pinned forever. So in this 1541 * case we opt to forcibly kill off all remaining requests on 1542 * context close. 1543 */ 1544 kill_context(ctx); 1545 1546 i915_gem_context_put(ctx); 1547 } 1548 1549 static int __context_set_persistence(struct i915_gem_context *ctx, bool state) 1550 { 1551 if (i915_gem_context_is_persistent(ctx) == state) 1552 return 0; 1553 1554 if (state) { 1555 /* 1556 * Only contexts that are short-lived [that will expire or be 1557 * reset] are allowed to survive past termination. We require 1558 * hangcheck to ensure that the persistent requests are healthy. 1559 */ 1560 if (!ctx->i915->params.enable_hangcheck) 1561 return -EINVAL; 1562 1563 i915_gem_context_set_persistence(ctx); 1564 } else { 1565 /* To cancel a context we use "preempt-to-idle" */ 1566 if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION)) 1567 return -ENODEV; 1568 1569 /* 1570 * If the cancel fails, we then need to reset, cleanly! 1571 * 1572 * If the per-engine reset fails, all hope is lost! We resort 1573 * to a full GPU reset in that unlikely case, but realistically 1574 * if the engine could not reset, the full reset does not fare 1575 * much better. The damage has been done. 1576 * 1577 * However, if we cannot reset an engine by itself, we cannot 1578 * cleanup a hanging persistent context without causing 1579 * colateral damage, and we should not pretend we can by 1580 * exposing the interface. 1581 */ 1582 if (!intel_has_reset_engine(to_gt(ctx->i915))) 1583 return -ENODEV; 1584 1585 i915_gem_context_clear_persistence(ctx); 1586 } 1587 1588 return 0; 1589 } 1590 1591 static struct i915_gem_context * 1592 i915_gem_create_context(struct drm_i915_private *i915, 1593 const struct i915_gem_proto_context *pc) 1594 { 1595 struct i915_gem_context *ctx; 1596 struct i915_address_space *vm = NULL; 1597 struct i915_gem_engines *e; 1598 int err; 1599 int i; 1600 1601 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 1602 if (!ctx) 1603 return ERR_PTR(-ENOMEM); 1604 1605 kref_init(&ctx->ref); 1606 ctx->i915 = i915; 1607 ctx->sched = pc->sched; 1608 mutex_init(&ctx->mutex); 1609 INIT_LIST_HEAD(&ctx->link); 1610 INIT_WORK(&ctx->release_work, i915_gem_context_release_work); 1611 1612 spin_lock_init(&ctx->stale.lock); 1613 INIT_LIST_HEAD(&ctx->stale.engines); 1614 1615 if (pc->vm) { 1616 vm = i915_vm_get(pc->vm); 1617 } else if (HAS_FULL_PPGTT(i915)) { 1618 struct i915_ppgtt *ppgtt; 1619 1620 ppgtt = i915_ppgtt_create(to_gt(i915), 0); 1621 if (IS_ERR(ppgtt)) { 1622 drm_dbg(&i915->drm, "PPGTT setup failed (%ld)\n", 1623 PTR_ERR(ppgtt)); 1624 err = PTR_ERR(ppgtt); 1625 goto err_ctx; 1626 } 1627 ppgtt->vm.fpriv = pc->fpriv; 1628 vm = &ppgtt->vm; 1629 } 1630 if (vm) 1631 ctx->vm = vm; 1632 1633 mutex_init(&ctx->engines_mutex); 1634 if (pc->num_user_engines >= 0) { 1635 i915_gem_context_set_user_engines(ctx); 1636 e = user_engines(ctx, pc->num_user_engines, pc->user_engines); 1637 } else { 1638 i915_gem_context_clear_user_engines(ctx); 1639 e = default_engines(ctx, pc->legacy_rcs_sseu); 1640 } 1641 if (IS_ERR(e)) { 1642 err = PTR_ERR(e); 1643 goto err_vm; 1644 } 1645 RCU_INIT_POINTER(ctx->engines, e); 1646 1647 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL); 1648 mutex_init(&ctx->lut_mutex); 1649 1650 /* NB: Mark all slices as needing a remap so that when the context first 1651 * loads it will restore whatever remap state already exists. If there 1652 * is no remap info, it will be a NOP. */ 1653 ctx->remap_slice = ALL_L3_SLICES(i915); 1654 1655 ctx->user_flags = pc->user_flags; 1656 1657 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++) 1658 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES; 1659 1660 if (pc->single_timeline) { 1661 err = drm_syncobj_create(&ctx->syncobj, 1662 DRM_SYNCOBJ_CREATE_SIGNALED, 1663 NULL); 1664 if (err) 1665 goto err_engines; 1666 } 1667 1668 if (pc->uses_protected_content) { 1669 ctx->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm); 1670 ctx->uses_protected_content = true; 1671 } 1672 1673 trace_i915_context_create(ctx); 1674 1675 return ctx; 1676 1677 err_engines: 1678 free_engines(e); 1679 err_vm: 1680 if (ctx->vm) 1681 i915_vm_put(ctx->vm); 1682 err_ctx: 1683 kfree(ctx); 1684 return ERR_PTR(err); 1685 } 1686 1687 static void init_contexts(struct i915_gem_contexts *gc) 1688 { 1689 spin_lock_init(&gc->lock); 1690 INIT_LIST_HEAD(&gc->list); 1691 } 1692 1693 void i915_gem_init__contexts(struct drm_i915_private *i915) 1694 { 1695 init_contexts(&i915->gem.contexts); 1696 } 1697 1698 /* 1699 * Note that this implicitly consumes the ctx reference, by placing 1700 * the ctx in the context_xa. 1701 */ 1702 static void gem_context_register(struct i915_gem_context *ctx, 1703 struct drm_i915_file_private *fpriv, 1704 u32 id) 1705 { 1706 struct drm_i915_private *i915 = ctx->i915; 1707 void *old; 1708 1709 ctx->file_priv = fpriv; 1710 1711 ctx->pid = get_task_pid(current, PIDTYPE_PID); 1712 ctx->client = i915_drm_client_get(fpriv->client); 1713 1714 snprintf(ctx->name, sizeof(ctx->name), "%s[%d]", 1715 current->comm, pid_nr(ctx->pid)); 1716 1717 spin_lock(&ctx->client->ctx_lock); 1718 list_add_tail_rcu(&ctx->client_link, &ctx->client->ctx_list); 1719 spin_unlock(&ctx->client->ctx_lock); 1720 1721 spin_lock(&i915->gem.contexts.lock); 1722 list_add_tail(&ctx->link, &i915->gem.contexts.list); 1723 spin_unlock(&i915->gem.contexts.lock); 1724 1725 /* And finally expose ourselves to userspace via the idr */ 1726 old = xa_store(&fpriv->context_xa, id, ctx, GFP_KERNEL); 1727 WARN_ON(old); 1728 } 1729 1730 int i915_gem_context_open(struct drm_i915_private *i915, 1731 struct drm_file *file) 1732 { 1733 struct drm_i915_file_private *file_priv = file->driver_priv; 1734 struct i915_gem_proto_context *pc; 1735 struct i915_gem_context *ctx; 1736 int err; 1737 1738 mutex_init(&file_priv->proto_context_lock); 1739 xa_init_flags(&file_priv->proto_context_xa, XA_FLAGS_ALLOC); 1740 1741 /* 0 reserved for the default context */ 1742 xa_init_flags(&file_priv->context_xa, XA_FLAGS_ALLOC1); 1743 1744 /* 0 reserved for invalid/unassigned ppgtt */ 1745 xa_init_flags(&file_priv->vm_xa, XA_FLAGS_ALLOC1); 1746 1747 pc = proto_context_create(file_priv, i915, 0); 1748 if (IS_ERR(pc)) { 1749 err = PTR_ERR(pc); 1750 goto err; 1751 } 1752 1753 ctx = i915_gem_create_context(i915, pc); 1754 proto_context_close(i915, pc); 1755 if (IS_ERR(ctx)) { 1756 err = PTR_ERR(ctx); 1757 goto err; 1758 } 1759 1760 gem_context_register(ctx, file_priv, 0); 1761 1762 return 0; 1763 1764 err: 1765 xa_destroy(&file_priv->vm_xa); 1766 xa_destroy(&file_priv->context_xa); 1767 xa_destroy(&file_priv->proto_context_xa); 1768 mutex_destroy(&file_priv->proto_context_lock); 1769 return err; 1770 } 1771 1772 void i915_gem_context_close(struct drm_file *file) 1773 { 1774 struct drm_i915_file_private *file_priv = file->driver_priv; 1775 struct i915_gem_proto_context *pc; 1776 struct i915_address_space *vm; 1777 struct i915_gem_context *ctx; 1778 unsigned long idx; 1779 1780 xa_for_each(&file_priv->proto_context_xa, idx, pc) 1781 proto_context_close(file_priv->i915, pc); 1782 xa_destroy(&file_priv->proto_context_xa); 1783 mutex_destroy(&file_priv->proto_context_lock); 1784 1785 xa_for_each(&file_priv->context_xa, idx, ctx) 1786 context_close(ctx); 1787 xa_destroy(&file_priv->context_xa); 1788 1789 xa_for_each(&file_priv->vm_xa, idx, vm) 1790 i915_vm_put(vm); 1791 xa_destroy(&file_priv->vm_xa); 1792 } 1793 1794 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data, 1795 struct drm_file *file) 1796 { 1797 struct drm_i915_private *i915 = to_i915(dev); 1798 struct drm_i915_gem_vm_control *args = data; 1799 struct drm_i915_file_private *file_priv = file->driver_priv; 1800 struct i915_ppgtt *ppgtt; 1801 u32 id; 1802 int err; 1803 1804 if (!HAS_FULL_PPGTT(i915)) 1805 return -ENODEV; 1806 1807 if (args->flags) 1808 return -EINVAL; 1809 1810 ppgtt = i915_ppgtt_create(to_gt(i915), 0); 1811 if (IS_ERR(ppgtt)) 1812 return PTR_ERR(ppgtt); 1813 1814 if (args->extensions) { 1815 err = i915_user_extensions(u64_to_user_ptr(args->extensions), 1816 NULL, 0, 1817 ppgtt); 1818 if (err) 1819 goto err_put; 1820 } 1821 1822 err = xa_alloc(&file_priv->vm_xa, &id, &ppgtt->vm, 1823 xa_limit_32b, GFP_KERNEL); 1824 if (err) 1825 goto err_put; 1826 1827 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */ 1828 args->vm_id = id; 1829 ppgtt->vm.fpriv = file_priv; 1830 return 0; 1831 1832 err_put: 1833 i915_vm_put(&ppgtt->vm); 1834 return err; 1835 } 1836 1837 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data, 1838 struct drm_file *file) 1839 { 1840 struct drm_i915_file_private *file_priv = file->driver_priv; 1841 struct drm_i915_gem_vm_control *args = data; 1842 struct i915_address_space *vm; 1843 1844 if (args->flags) 1845 return -EINVAL; 1846 1847 if (args->extensions) 1848 return -EINVAL; 1849 1850 vm = xa_erase(&file_priv->vm_xa, args->vm_id); 1851 if (!vm) 1852 return -ENOENT; 1853 1854 i915_vm_put(vm); 1855 return 0; 1856 } 1857 1858 static int get_ppgtt(struct drm_i915_file_private *file_priv, 1859 struct i915_gem_context *ctx, 1860 struct drm_i915_gem_context_param *args) 1861 { 1862 struct i915_address_space *vm; 1863 int err; 1864 u32 id; 1865 1866 if (!i915_gem_context_has_full_ppgtt(ctx)) 1867 return -ENODEV; 1868 1869 vm = ctx->vm; 1870 GEM_BUG_ON(!vm); 1871 1872 /* 1873 * Get a reference for the allocated handle. Once the handle is 1874 * visible in the vm_xa table, userspace could try to close it 1875 * from under our feet, so we need to hold the extra reference 1876 * first. 1877 */ 1878 i915_vm_get(vm); 1879 1880 err = xa_alloc(&file_priv->vm_xa, &id, vm, xa_limit_32b, GFP_KERNEL); 1881 if (err) { 1882 i915_vm_put(vm); 1883 return err; 1884 } 1885 1886 GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */ 1887 args->value = id; 1888 args->size = 0; 1889 1890 return err; 1891 } 1892 1893 int 1894 i915_gem_user_to_context_sseu(struct intel_gt *gt, 1895 const struct drm_i915_gem_context_param_sseu *user, 1896 struct intel_sseu *context) 1897 { 1898 const struct sseu_dev_info *device = >->info.sseu; 1899 struct drm_i915_private *i915 = gt->i915; 1900 unsigned int dev_subslice_mask = intel_sseu_get_hsw_subslices(device, 0); 1901 1902 /* No zeros in any field. */ 1903 if (!user->slice_mask || !user->subslice_mask || 1904 !user->min_eus_per_subslice || !user->max_eus_per_subslice) 1905 return -EINVAL; 1906 1907 /* Max > min. */ 1908 if (user->max_eus_per_subslice < user->min_eus_per_subslice) 1909 return -EINVAL; 1910 1911 /* 1912 * Some future proofing on the types since the uAPI is wider than the 1913 * current internal implementation. 1914 */ 1915 if (overflows_type(user->slice_mask, context->slice_mask) || 1916 overflows_type(user->subslice_mask, context->subslice_mask) || 1917 overflows_type(user->min_eus_per_subslice, 1918 context->min_eus_per_subslice) || 1919 overflows_type(user->max_eus_per_subslice, 1920 context->max_eus_per_subslice)) 1921 return -EINVAL; 1922 1923 /* Check validity against hardware. */ 1924 if (user->slice_mask & ~device->slice_mask) 1925 return -EINVAL; 1926 1927 if (user->subslice_mask & ~dev_subslice_mask) 1928 return -EINVAL; 1929 1930 if (user->max_eus_per_subslice > device->max_eus_per_subslice) 1931 return -EINVAL; 1932 1933 context->slice_mask = user->slice_mask; 1934 context->subslice_mask = user->subslice_mask; 1935 context->min_eus_per_subslice = user->min_eus_per_subslice; 1936 context->max_eus_per_subslice = user->max_eus_per_subslice; 1937 1938 /* Part specific restrictions. */ 1939 if (GRAPHICS_VER(i915) == 11) { 1940 unsigned int hw_s = hweight8(device->slice_mask); 1941 unsigned int hw_ss_per_s = hweight8(dev_subslice_mask); 1942 unsigned int req_s = hweight8(context->slice_mask); 1943 unsigned int req_ss = hweight8(context->subslice_mask); 1944 1945 /* 1946 * Only full subslice enablement is possible if more than one 1947 * slice is turned on. 1948 */ 1949 if (req_s > 1 && req_ss != hw_ss_per_s) 1950 return -EINVAL; 1951 1952 /* 1953 * If more than four (SScount bitfield limit) subslices are 1954 * requested then the number has to be even. 1955 */ 1956 if (req_ss > 4 && (req_ss & 1)) 1957 return -EINVAL; 1958 1959 /* 1960 * If only one slice is enabled and subslice count is below the 1961 * device full enablement, it must be at most half of the all 1962 * available subslices. 1963 */ 1964 if (req_s == 1 && req_ss < hw_ss_per_s && 1965 req_ss > (hw_ss_per_s / 2)) 1966 return -EINVAL; 1967 1968 /* ABI restriction - VME use case only. */ 1969 1970 /* All slices or one slice only. */ 1971 if (req_s != 1 && req_s != hw_s) 1972 return -EINVAL; 1973 1974 /* 1975 * Half subslices or full enablement only when one slice is 1976 * enabled. 1977 */ 1978 if (req_s == 1 && 1979 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2))) 1980 return -EINVAL; 1981 1982 /* No EU configuration changes. */ 1983 if ((user->min_eus_per_subslice != 1984 device->max_eus_per_subslice) || 1985 (user->max_eus_per_subslice != 1986 device->max_eus_per_subslice)) 1987 return -EINVAL; 1988 } 1989 1990 return 0; 1991 } 1992 1993 static int set_sseu(struct i915_gem_context *ctx, 1994 struct drm_i915_gem_context_param *args) 1995 { 1996 struct drm_i915_private *i915 = ctx->i915; 1997 struct drm_i915_gem_context_param_sseu user_sseu; 1998 struct intel_context *ce; 1999 struct intel_sseu sseu; 2000 unsigned long lookup; 2001 int ret; 2002 2003 if (args->size < sizeof(user_sseu)) 2004 return -EINVAL; 2005 2006 if (GRAPHICS_VER(i915) != 11) 2007 return -ENODEV; 2008 2009 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 2010 sizeof(user_sseu))) 2011 return -EFAULT; 2012 2013 if (user_sseu.rsvd) 2014 return -EINVAL; 2015 2016 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 2017 return -EINVAL; 2018 2019 lookup = 0; 2020 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) 2021 lookup |= LOOKUP_USER_INDEX; 2022 2023 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine); 2024 if (IS_ERR(ce)) 2025 return PTR_ERR(ce); 2026 2027 /* Only render engine supports RPCS configuration. */ 2028 if (ce->engine->class != RENDER_CLASS) { 2029 ret = -ENODEV; 2030 goto out_ce; 2031 } 2032 2033 ret = i915_gem_user_to_context_sseu(ce->engine->gt, &user_sseu, &sseu); 2034 if (ret) 2035 goto out_ce; 2036 2037 ret = intel_context_reconfigure_sseu(ce, sseu); 2038 if (ret) 2039 goto out_ce; 2040 2041 args->size = sizeof(user_sseu); 2042 2043 out_ce: 2044 intel_context_put(ce); 2045 return ret; 2046 } 2047 2048 static int 2049 set_persistence(struct i915_gem_context *ctx, 2050 const struct drm_i915_gem_context_param *args) 2051 { 2052 if (args->size) 2053 return -EINVAL; 2054 2055 return __context_set_persistence(ctx, args->value); 2056 } 2057 2058 static int set_priority(struct i915_gem_context *ctx, 2059 const struct drm_i915_gem_context_param *args) 2060 { 2061 struct i915_gem_engines_iter it; 2062 struct intel_context *ce; 2063 int err; 2064 2065 err = validate_priority(ctx->i915, args); 2066 if (err) 2067 return err; 2068 2069 ctx->sched.priority = args->value; 2070 2071 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) { 2072 if (!intel_engine_has_timeslices(ce->engine)) 2073 continue; 2074 2075 if (ctx->sched.priority >= I915_PRIORITY_NORMAL && 2076 intel_engine_has_semaphores(ce->engine)) 2077 intel_context_set_use_semaphores(ce); 2078 else 2079 intel_context_clear_use_semaphores(ce); 2080 } 2081 i915_gem_context_unlock_engines(ctx); 2082 2083 return 0; 2084 } 2085 2086 static int get_protected(struct i915_gem_context *ctx, 2087 struct drm_i915_gem_context_param *args) 2088 { 2089 args->size = 0; 2090 args->value = i915_gem_context_uses_protected_content(ctx); 2091 2092 return 0; 2093 } 2094 2095 static int ctx_setparam(struct drm_i915_file_private *fpriv, 2096 struct i915_gem_context *ctx, 2097 struct drm_i915_gem_context_param *args) 2098 { 2099 int ret = 0; 2100 2101 switch (args->param) { 2102 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 2103 if (args->size) 2104 ret = -EINVAL; 2105 else if (args->value) 2106 i915_gem_context_set_no_error_capture(ctx); 2107 else 2108 i915_gem_context_clear_no_error_capture(ctx); 2109 break; 2110 2111 case I915_CONTEXT_PARAM_BANNABLE: 2112 if (args->size) 2113 ret = -EINVAL; 2114 else if (!capable(CAP_SYS_ADMIN) && !args->value) 2115 ret = -EPERM; 2116 else if (args->value) 2117 i915_gem_context_set_bannable(ctx); 2118 else if (i915_gem_context_uses_protected_content(ctx)) 2119 ret = -EPERM; /* can't clear this for protected contexts */ 2120 else 2121 i915_gem_context_clear_bannable(ctx); 2122 break; 2123 2124 case I915_CONTEXT_PARAM_RECOVERABLE: 2125 if (args->size) 2126 ret = -EINVAL; 2127 else if (!args->value) 2128 i915_gem_context_clear_recoverable(ctx); 2129 else if (i915_gem_context_uses_protected_content(ctx)) 2130 ret = -EPERM; /* can't set this for protected contexts */ 2131 else 2132 i915_gem_context_set_recoverable(ctx); 2133 break; 2134 2135 case I915_CONTEXT_PARAM_PRIORITY: 2136 ret = set_priority(ctx, args); 2137 break; 2138 2139 case I915_CONTEXT_PARAM_SSEU: 2140 ret = set_sseu(ctx, args); 2141 break; 2142 2143 case I915_CONTEXT_PARAM_PERSISTENCE: 2144 ret = set_persistence(ctx, args); 2145 break; 2146 2147 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 2148 case I915_CONTEXT_PARAM_NO_ZEROMAP: 2149 case I915_CONTEXT_PARAM_BAN_PERIOD: 2150 case I915_CONTEXT_PARAM_RINGSIZE: 2151 case I915_CONTEXT_PARAM_VM: 2152 case I915_CONTEXT_PARAM_ENGINES: 2153 default: 2154 ret = -EINVAL; 2155 break; 2156 } 2157 2158 return ret; 2159 } 2160 2161 struct create_ext { 2162 struct i915_gem_proto_context *pc; 2163 struct drm_i915_file_private *fpriv; 2164 }; 2165 2166 static int create_setparam(struct i915_user_extension __user *ext, void *data) 2167 { 2168 struct drm_i915_gem_context_create_ext_setparam local; 2169 const struct create_ext *arg = data; 2170 2171 if (copy_from_user(&local, ext, sizeof(local))) 2172 return -EFAULT; 2173 2174 if (local.param.ctx_id) 2175 return -EINVAL; 2176 2177 return set_proto_ctx_param(arg->fpriv, arg->pc, &local.param); 2178 } 2179 2180 static int invalid_ext(struct i915_user_extension __user *ext, void *data) 2181 { 2182 return -EINVAL; 2183 } 2184 2185 static const i915_user_extension_fn create_extensions[] = { 2186 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam, 2187 [I915_CONTEXT_CREATE_EXT_CLONE] = invalid_ext, 2188 }; 2189 2190 static bool client_is_banned(struct drm_i915_file_private *file_priv) 2191 { 2192 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED; 2193 } 2194 2195 static inline struct i915_gem_context * 2196 __context_lookup(struct drm_i915_file_private *file_priv, u32 id) 2197 { 2198 struct i915_gem_context *ctx; 2199 2200 rcu_read_lock(); 2201 ctx = xa_load(&file_priv->context_xa, id); 2202 if (ctx && !kref_get_unless_zero(&ctx->ref)) 2203 ctx = NULL; 2204 rcu_read_unlock(); 2205 2206 return ctx; 2207 } 2208 2209 static struct i915_gem_context * 2210 finalize_create_context_locked(struct drm_i915_file_private *file_priv, 2211 struct i915_gem_proto_context *pc, u32 id) 2212 { 2213 struct i915_gem_context *ctx; 2214 void *old; 2215 2216 lockdep_assert_held(&file_priv->proto_context_lock); 2217 2218 ctx = i915_gem_create_context(file_priv->i915, pc); 2219 if (IS_ERR(ctx)) 2220 return ctx; 2221 2222 /* 2223 * One for the xarray and one for the caller. We need to grab 2224 * the reference *prior* to making the ctx visble to userspace 2225 * in gem_context_register(), as at any point after that 2226 * userspace can try to race us with another thread destroying 2227 * the context under our feet. 2228 */ 2229 i915_gem_context_get(ctx); 2230 2231 gem_context_register(ctx, file_priv, id); 2232 2233 old = xa_erase(&file_priv->proto_context_xa, id); 2234 GEM_BUG_ON(old != pc); 2235 proto_context_close(file_priv->i915, pc); 2236 2237 return ctx; 2238 } 2239 2240 struct i915_gem_context * 2241 i915_gem_context_lookup(struct drm_i915_file_private *file_priv, u32 id) 2242 { 2243 struct i915_gem_proto_context *pc; 2244 struct i915_gem_context *ctx; 2245 2246 ctx = __context_lookup(file_priv, id); 2247 if (ctx) 2248 return ctx; 2249 2250 mutex_lock(&file_priv->proto_context_lock); 2251 /* Try one more time under the lock */ 2252 ctx = __context_lookup(file_priv, id); 2253 if (!ctx) { 2254 pc = xa_load(&file_priv->proto_context_xa, id); 2255 if (!pc) 2256 ctx = ERR_PTR(-ENOENT); 2257 else 2258 ctx = finalize_create_context_locked(file_priv, pc, id); 2259 } 2260 mutex_unlock(&file_priv->proto_context_lock); 2261 2262 return ctx; 2263 } 2264 2265 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data, 2266 struct drm_file *file) 2267 { 2268 struct drm_i915_private *i915 = to_i915(dev); 2269 struct drm_i915_gem_context_create_ext *args = data; 2270 struct create_ext ext_data; 2271 int ret; 2272 u32 id; 2273 2274 if (!DRIVER_CAPS(i915)->has_logical_contexts) 2275 return -ENODEV; 2276 2277 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN) 2278 return -EINVAL; 2279 2280 ret = intel_gt_terminally_wedged(to_gt(i915)); 2281 if (ret) 2282 return ret; 2283 2284 ext_data.fpriv = file->driver_priv; 2285 if (client_is_banned(ext_data.fpriv)) { 2286 drm_dbg(&i915->drm, 2287 "client %s[%d] banned from creating ctx\n", 2288 current->comm, task_pid_nr(current)); 2289 return -EIO; 2290 } 2291 2292 ext_data.pc = proto_context_create(file->driver_priv, i915, 2293 args->flags); 2294 if (IS_ERR(ext_data.pc)) 2295 return PTR_ERR(ext_data.pc); 2296 2297 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) { 2298 ret = i915_user_extensions(u64_to_user_ptr(args->extensions), 2299 create_extensions, 2300 ARRAY_SIZE(create_extensions), 2301 &ext_data); 2302 if (ret) 2303 goto err_pc; 2304 } 2305 2306 if (GRAPHICS_VER(i915) > 12) { 2307 struct i915_gem_context *ctx; 2308 2309 /* Get ourselves a context ID */ 2310 ret = xa_alloc(&ext_data.fpriv->context_xa, &id, NULL, 2311 xa_limit_32b, GFP_KERNEL); 2312 if (ret) 2313 goto err_pc; 2314 2315 ctx = i915_gem_create_context(i915, ext_data.pc); 2316 if (IS_ERR(ctx)) { 2317 ret = PTR_ERR(ctx); 2318 goto err_pc; 2319 } 2320 2321 proto_context_close(i915, ext_data.pc); 2322 gem_context_register(ctx, ext_data.fpriv, id); 2323 } else { 2324 ret = proto_context_register(ext_data.fpriv, ext_data.pc, &id); 2325 if (ret < 0) 2326 goto err_pc; 2327 } 2328 2329 args->ctx_id = id; 2330 2331 return 0; 2332 2333 err_pc: 2334 proto_context_close(i915, ext_data.pc); 2335 return ret; 2336 } 2337 2338 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data, 2339 struct drm_file *file) 2340 { 2341 struct drm_i915_gem_context_destroy *args = data; 2342 struct drm_i915_file_private *file_priv = file->driver_priv; 2343 struct i915_gem_proto_context *pc; 2344 struct i915_gem_context *ctx; 2345 2346 if (args->pad != 0) 2347 return -EINVAL; 2348 2349 if (!args->ctx_id) 2350 return -ENOENT; 2351 2352 /* We need to hold the proto-context lock here to prevent races 2353 * with finalize_create_context_locked(). 2354 */ 2355 mutex_lock(&file_priv->proto_context_lock); 2356 ctx = xa_erase(&file_priv->context_xa, args->ctx_id); 2357 pc = xa_erase(&file_priv->proto_context_xa, args->ctx_id); 2358 mutex_unlock(&file_priv->proto_context_lock); 2359 2360 if (!ctx && !pc) 2361 return -ENOENT; 2362 GEM_WARN_ON(ctx && pc); 2363 2364 if (pc) 2365 proto_context_close(file_priv->i915, pc); 2366 2367 if (ctx) 2368 context_close(ctx); 2369 2370 return 0; 2371 } 2372 2373 static int get_sseu(struct i915_gem_context *ctx, 2374 struct drm_i915_gem_context_param *args) 2375 { 2376 struct drm_i915_gem_context_param_sseu user_sseu; 2377 struct intel_context *ce; 2378 unsigned long lookup; 2379 int err; 2380 2381 if (args->size == 0) 2382 goto out; 2383 else if (args->size < sizeof(user_sseu)) 2384 return -EINVAL; 2385 2386 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value), 2387 sizeof(user_sseu))) 2388 return -EFAULT; 2389 2390 if (user_sseu.rsvd) 2391 return -EINVAL; 2392 2393 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)) 2394 return -EINVAL; 2395 2396 lookup = 0; 2397 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX) 2398 lookup |= LOOKUP_USER_INDEX; 2399 2400 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine); 2401 if (IS_ERR(ce)) 2402 return PTR_ERR(ce); 2403 2404 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */ 2405 if (err) { 2406 intel_context_put(ce); 2407 return err; 2408 } 2409 2410 user_sseu.slice_mask = ce->sseu.slice_mask; 2411 user_sseu.subslice_mask = ce->sseu.subslice_mask; 2412 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice; 2413 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice; 2414 2415 intel_context_unlock_pinned(ce); 2416 intel_context_put(ce); 2417 2418 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu, 2419 sizeof(user_sseu))) 2420 return -EFAULT; 2421 2422 out: 2423 args->size = sizeof(user_sseu); 2424 2425 return 0; 2426 } 2427 2428 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data, 2429 struct drm_file *file) 2430 { 2431 struct drm_i915_file_private *file_priv = file->driver_priv; 2432 struct drm_i915_gem_context_param *args = data; 2433 struct i915_gem_context *ctx; 2434 struct i915_address_space *vm; 2435 int ret = 0; 2436 2437 ctx = i915_gem_context_lookup(file_priv, args->ctx_id); 2438 if (IS_ERR(ctx)) 2439 return PTR_ERR(ctx); 2440 2441 switch (args->param) { 2442 case I915_CONTEXT_PARAM_GTT_SIZE: 2443 args->size = 0; 2444 vm = i915_gem_context_get_eb_vm(ctx); 2445 args->value = vm->total; 2446 i915_vm_put(vm); 2447 2448 break; 2449 2450 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE: 2451 args->size = 0; 2452 args->value = i915_gem_context_no_error_capture(ctx); 2453 break; 2454 2455 case I915_CONTEXT_PARAM_BANNABLE: 2456 args->size = 0; 2457 args->value = i915_gem_context_is_bannable(ctx); 2458 break; 2459 2460 case I915_CONTEXT_PARAM_RECOVERABLE: 2461 args->size = 0; 2462 args->value = i915_gem_context_is_recoverable(ctx); 2463 break; 2464 2465 case I915_CONTEXT_PARAM_PRIORITY: 2466 args->size = 0; 2467 args->value = ctx->sched.priority; 2468 break; 2469 2470 case I915_CONTEXT_PARAM_SSEU: 2471 ret = get_sseu(ctx, args); 2472 break; 2473 2474 case I915_CONTEXT_PARAM_VM: 2475 ret = get_ppgtt(file_priv, ctx, args); 2476 break; 2477 2478 case I915_CONTEXT_PARAM_PERSISTENCE: 2479 args->size = 0; 2480 args->value = i915_gem_context_is_persistent(ctx); 2481 break; 2482 2483 case I915_CONTEXT_PARAM_PROTECTED_CONTENT: 2484 ret = get_protected(ctx, args); 2485 break; 2486 2487 case I915_CONTEXT_PARAM_NO_ZEROMAP: 2488 case I915_CONTEXT_PARAM_BAN_PERIOD: 2489 case I915_CONTEXT_PARAM_ENGINES: 2490 case I915_CONTEXT_PARAM_RINGSIZE: 2491 default: 2492 ret = -EINVAL; 2493 break; 2494 } 2495 2496 i915_gem_context_put(ctx); 2497 return ret; 2498 } 2499 2500 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data, 2501 struct drm_file *file) 2502 { 2503 struct drm_i915_file_private *file_priv = file->driver_priv; 2504 struct drm_i915_gem_context_param *args = data; 2505 struct i915_gem_proto_context *pc; 2506 struct i915_gem_context *ctx; 2507 int ret = 0; 2508 2509 mutex_lock(&file_priv->proto_context_lock); 2510 ctx = __context_lookup(file_priv, args->ctx_id); 2511 if (!ctx) { 2512 pc = xa_load(&file_priv->proto_context_xa, args->ctx_id); 2513 if (pc) { 2514 /* Contexts should be finalized inside 2515 * GEM_CONTEXT_CREATE starting with graphics 2516 * version 13. 2517 */ 2518 WARN_ON(GRAPHICS_VER(file_priv->i915) > 12); 2519 ret = set_proto_ctx_param(file_priv, pc, args); 2520 } else { 2521 ret = -ENOENT; 2522 } 2523 } 2524 mutex_unlock(&file_priv->proto_context_lock); 2525 2526 if (ctx) { 2527 ret = ctx_setparam(file_priv, ctx, args); 2528 i915_gem_context_put(ctx); 2529 } 2530 2531 return ret; 2532 } 2533 2534 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev, 2535 void *data, struct drm_file *file) 2536 { 2537 struct drm_i915_private *i915 = to_i915(dev); 2538 struct drm_i915_reset_stats *args = data; 2539 struct i915_gem_context *ctx; 2540 2541 if (args->flags || args->pad) 2542 return -EINVAL; 2543 2544 ctx = i915_gem_context_lookup(file->driver_priv, args->ctx_id); 2545 if (IS_ERR(ctx)) 2546 return PTR_ERR(ctx); 2547 2548 /* 2549 * We opt for unserialised reads here. This may result in tearing 2550 * in the extremely unlikely event of a GPU hang on this context 2551 * as we are querying them. If we need that extra layer of protection, 2552 * we should wrap the hangstats with a seqlock. 2553 */ 2554 2555 if (capable(CAP_SYS_ADMIN)) 2556 args->reset_count = i915_reset_count(&i915->gpu_error); 2557 else 2558 args->reset_count = 0; 2559 2560 args->batch_active = atomic_read(&ctx->guilty_count); 2561 args->batch_pending = atomic_read(&ctx->active_count); 2562 2563 i915_gem_context_put(ctx); 2564 return 0; 2565 } 2566 2567 /* GEM context-engines iterator: for_each_gem_engine() */ 2568 struct intel_context * 2569 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it) 2570 { 2571 const struct i915_gem_engines *e = it->engines; 2572 struct intel_context *ctx; 2573 2574 if (unlikely(!e)) 2575 return NULL; 2576 2577 do { 2578 if (it->idx >= e->num_engines) 2579 return NULL; 2580 2581 ctx = e->engines[it->idx++]; 2582 } while (!ctx); 2583 2584 return ctx; 2585 } 2586 2587 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) 2588 #include "selftests/mock_context.c" 2589 #include "selftests/i915_gem_context.c" 2590 #endif 2591 2592 void i915_gem_context_module_exit(void) 2593 { 2594 kmem_cache_destroy(slab_luts); 2595 } 2596 2597 int __init i915_gem_context_module_init(void) 2598 { 2599 slab_luts = KMEM_CACHE(i915_lut_handle, 0); 2600 if (!slab_luts) 2601 return -ENOMEM; 2602 2603 return 0; 2604 } 2605