drivers/gpu/drm/i915/gem/i915_gem_context.c

   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(&to_gt(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(&to_gt(i915)->pxp))
 275                         ret = intel_pxp_start(&to_gt(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->dev_priv;
 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_DEBUG("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->dev_priv;
 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->dev_priv;
 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->dev_priv, 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->dev_priv, pc,
 938                                                             args->value);
 939                 break;
 940
 941         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
 942                 ret = proto_context_set_protected(fpriv->dev_priv, 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 struct intel_gt *gt = to_gt(ctx->i915);
1100         struct intel_engine_cs *engine;
1101         struct i915_gem_engines *e, *err;
1102         enum intel_engine_id id;
1103
1104         e = alloc_engines(I915_NUM_ENGINES);
1105         if (!e)
1106                 return ERR_PTR(-ENOMEM);
1107
1108         for_each_engine(engine, gt, id) {
1109                 struct intel_context *ce;
1110                 struct intel_sseu sseu = {};
1111                 int ret;
1112
1113                 if (engine->legacy_idx == INVALID_ENGINE)
1114                         continue;
1115
1116                 GEM_BUG_ON(engine->legacy_idx >= I915_NUM_ENGINES);
1117                 GEM_BUG_ON(e->engines[engine->legacy_idx]);
1118
1119                 ce = intel_context_create(engine);
1120                 if (IS_ERR(ce)) {
1121                         err = ERR_CAST(ce);
1122                         goto free_engines;
1123                 }
1124
1125                 e->engines[engine->legacy_idx] = ce;
1126                 e->num_engines = max(e->num_engines, engine->legacy_idx + 1);
1127
1128                 if (engine->class == RENDER_CLASS)
1129                         sseu = rcs_sseu;
1130
1131                 ret = intel_context_set_gem(ce, ctx, sseu);
1132                 if (ret) {
1133                         err = ERR_PTR(ret);
1134                         goto free_engines;
1135                 }
1136
1137         }
1138
1139         return e;
1140
1141 free_engines:
1142         free_engines(e);
1143         return err;
1144 }
1145
1146 static int perma_pin_contexts(struct intel_context *ce)
1147 {
1148         struct intel_context *child;
1149         int i = 0, j = 0, ret;
1150
1151         GEM_BUG_ON(!intel_context_is_parent(ce));
1152
1153         ret = intel_context_pin(ce);
1154         if (unlikely(ret))
1155                 return ret;
1156
1157         for_each_child(ce, child) {
1158                 ret = intel_context_pin(child);
1159                 if (unlikely(ret))
1160                         goto unwind;
1161                 ++i;
1162         }
1163
1164         set_bit(CONTEXT_PERMA_PIN, &ce->flags);
1165
1166         return 0;
1167
1168 unwind:
1169         intel_context_unpin(ce);
1170         for_each_child(ce, child) {
1171                 if (j++ < i)
1172                         intel_context_unpin(child);
1173                 else
1174                         break;
1175         }
1176
1177         return ret;
1178 }
1179
1180 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx,
1181                                              unsigned int num_engines,
1182                                              struct i915_gem_proto_engine *pe)
1183 {
1184         struct i915_gem_engines *e, *err;
1185         unsigned int n;
1186
1187         e = alloc_engines(num_engines);
1188         if (!e)
1189                 return ERR_PTR(-ENOMEM);
1190         e->num_engines = num_engines;
1191
1192         for (n = 0; n < num_engines; n++) {
1193                 struct intel_context *ce, *child;
1194                 int ret;
1195
1196                 switch (pe[n].type) {
1197                 case I915_GEM_ENGINE_TYPE_PHYSICAL:
1198                         ce = intel_context_create(pe[n].engine);
1199                         break;
1200
1201                 case I915_GEM_ENGINE_TYPE_BALANCED:
1202                         ce = intel_engine_create_virtual(pe[n].siblings,
1203                                                          pe[n].num_siblings, 0);
1204                         break;
1205
1206                 case I915_GEM_ENGINE_TYPE_PARALLEL:
1207                         ce = intel_engine_create_parallel(pe[n].siblings,
1208                                                           pe[n].num_siblings,
1209                                                           pe[n].width);
1210                         break;
1211
1212                 case I915_GEM_ENGINE_TYPE_INVALID:
1213                 default:
1214                         GEM_WARN_ON(pe[n].type != I915_GEM_ENGINE_TYPE_INVALID);
1215                         continue;
1216                 }
1217
1218                 if (IS_ERR(ce)) {
1219                         err = ERR_CAST(ce);
1220                         goto free_engines;
1221                 }
1222
1223                 e->engines[n] = ce;
1224
1225                 ret = intel_context_set_gem(ce, ctx, pe->sseu);
1226                 if (ret) {
1227                         err = ERR_PTR(ret);
1228                         goto free_engines;
1229                 }
1230                 for_each_child(ce, child) {
1231                         ret = intel_context_set_gem(child, ctx, pe->sseu);
1232                         if (ret) {
1233                                 err = ERR_PTR(ret);
1234                                 goto free_engines;
1235                         }
1236                 }
1237
1238                 /*
1239                  * XXX: Must be done after calling intel_context_set_gem as that
1240                  * function changes the ring size. The ring is allocated when
1241                  * the context is pinned. If the ring size is changed after
1242                  * allocation we have a mismatch of the ring size and will cause
1243                  * the context to hang. Presumably with a bit of reordering we
1244                  * could move the perma-pin step to the backend function
1245                  * intel_engine_create_parallel.
1246                  */
1247                 if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) {
1248                         ret = perma_pin_contexts(ce);
1249                         if (ret) {
1250                                 err = ERR_PTR(ret);
1251                                 goto free_engines;
1252                         }
1253                 }
1254         }
1255
1256         return e;
1257
1258 free_engines:
1259         free_engines(e);
1260         return err;
1261 }
1262
1263 static void i915_gem_context_release_work(struct work_struct *work)
1264 {
1265         struct i915_gem_context *ctx = container_of(work, typeof(*ctx),
1266                                                     release_work);
1267         struct i915_address_space *vm;
1268
1269         trace_i915_context_free(ctx);
1270         GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1271
1272         if (ctx->syncobj)
1273                 drm_syncobj_put(ctx->syncobj);
1274
1275         vm = ctx->vm;
1276         if (vm)
1277                 i915_vm_put(vm);
1278
1279         if (ctx->pxp_wakeref)
1280                 intel_runtime_pm_put(&ctx->i915->runtime_pm, ctx->pxp_wakeref);
1281
1282         if (ctx->client)
1283                 i915_drm_client_put(ctx->client);
1284
1285         mutex_destroy(&ctx->engines_mutex);
1286         mutex_destroy(&ctx->lut_mutex);
1287
1288         put_pid(ctx->pid);
1289         mutex_destroy(&ctx->mutex);
1290
1291         kfree_rcu(ctx, rcu);
1292 }
1293
1294 void i915_gem_context_release(struct kref *ref)
1295 {
1296         struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
1297
1298         queue_work(ctx->i915->wq, &ctx->release_work);
1299 }
1300
1301 static inline struct i915_gem_engines *
1302 __context_engines_static(const struct i915_gem_context *ctx)
1303 {
1304         return rcu_dereference_protected(ctx->engines, true);
1305 }
1306
1307 static void __reset_context(struct i915_gem_context *ctx,
1308                             struct intel_engine_cs *engine)
1309 {
1310         intel_gt_handle_error(engine->gt, engine->mask, 0,
1311                               "context closure in %s", ctx->name);
1312 }
1313
1314 static bool __cancel_engine(struct intel_engine_cs *engine)
1315 {
1316         /*
1317          * Send a "high priority pulse" down the engine to cause the
1318          * current request to be momentarily preempted. (If it fails to
1319          * be preempted, it will be reset). As we have marked our context
1320          * as banned, any incomplete request, including any running, will
1321          * be skipped following the preemption.
1322          *
1323          * If there is no hangchecking (one of the reasons why we try to
1324          * cancel the context) and no forced preemption, there may be no
1325          * means by which we reset the GPU and evict the persistent hog.
1326          * Ergo if we are unable to inject a preemptive pulse that can
1327          * kill the banned context, we fallback to doing a local reset
1328          * instead.
1329          */
1330         return intel_engine_pulse(engine) == 0;
1331 }
1332
1333 static struct intel_engine_cs *active_engine(struct intel_context *ce)
1334 {
1335         struct intel_engine_cs *engine = NULL;
1336         struct i915_request *rq;
1337
1338         if (intel_context_has_inflight(ce))
1339                 return intel_context_inflight(ce);
1340
1341         if (!ce->timeline)
1342                 return NULL;
1343
1344         /*
1345          * rq->link is only SLAB_TYPESAFE_BY_RCU, we need to hold a reference
1346          * to the request to prevent it being transferred to a new timeline
1347          * (and onto a new timeline->requests list).
1348          */
1349         rcu_read_lock();
1350         list_for_each_entry_reverse(rq, &ce->timeline->requests, link) {
1351                 bool found;
1352
1353                 /* timeline is already completed upto this point? */
1354                 if (!i915_request_get_rcu(rq))
1355                         break;
1356
1357                 /* Check with the backend if the request is inflight */
1358                 found = true;
1359                 if (likely(rcu_access_pointer(rq->timeline) == ce->timeline))
1360                         found = i915_request_active_engine(rq, &engine);
1361
1362                 i915_request_put(rq);
1363                 if (found)
1364                         break;
1365         }
1366         rcu_read_unlock();
1367
1368         return engine;
1369 }
1370
1371 static void
1372 kill_engines(struct i915_gem_engines *engines, bool exit, bool persistent)
1373 {
1374         struct i915_gem_engines_iter it;
1375         struct intel_context *ce;
1376
1377         /*
1378          * Map the user's engine back to the actual engines; one virtual
1379          * engine will be mapped to multiple engines, and using ctx->engine[]
1380          * the same engine may be have multiple instances in the user's map.
1381          * However, we only care about pending requests, so only include
1382          * engines on which there are incomplete requests.
1383          */
1384         for_each_gem_engine(ce, engines, it) {
1385                 struct intel_engine_cs *engine;
1386                 bool skip = false;
1387
1388                 if (exit)
1389                         skip = intel_context_set_exiting(ce);
1390                 else if (!persistent)
1391                         skip = intel_context_exit_nonpersistent(ce, NULL);
1392
1393                 if (skip)
1394                         continue; /* Already marked. */
1395
1396                 /*
1397                  * Check the current active state of this context; if we
1398                  * are currently executing on the GPU we need to evict
1399                  * ourselves. On the other hand, if we haven't yet been
1400                  * submitted to the GPU or if everything is complete,
1401                  * we have nothing to do.
1402                  */
1403                 engine = active_engine(ce);
1404
1405                 /* First attempt to gracefully cancel the context */
1406                 if (engine && !__cancel_engine(engine) && (exit || !persistent))
1407                         /*
1408                          * If we are unable to send a preemptive pulse to bump
1409                          * the context from the GPU, we have to resort to a full
1410                          * reset. We hope the collateral damage is worth it.
1411                          */
1412                         __reset_context(engines->ctx, engine);
1413         }
1414 }
1415
1416 static void kill_context(struct i915_gem_context *ctx)
1417 {
1418         struct i915_gem_engines *pos, *next;
1419
1420         spin_lock_irq(&ctx->stale.lock);
1421         GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1422         list_for_each_entry_safe(pos, next, &ctx->stale.engines, link) {
1423                 if (!i915_sw_fence_await(&pos->fence)) {
1424                         list_del_init(&pos->link);
1425                         continue;
1426                 }
1427
1428                 spin_unlock_irq(&ctx->stale.lock);
1429
1430                 kill_engines(pos, !ctx->i915->params.enable_hangcheck,
1431                              i915_gem_context_is_persistent(ctx));
1432
1433                 spin_lock_irq(&ctx->stale.lock);
1434                 GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence));
1435                 list_safe_reset_next(pos, next, link);
1436                 list_del_init(&pos->link); /* decouple from FENCE_COMPLETE */
1437
1438                 i915_sw_fence_complete(&pos->fence);
1439         }
1440         spin_unlock_irq(&ctx->stale.lock);
1441 }
1442
1443 static void engines_idle_release(struct i915_gem_context *ctx,
1444                                  struct i915_gem_engines *engines)
1445 {
1446         struct i915_gem_engines_iter it;
1447         struct intel_context *ce;
1448
1449         INIT_LIST_HEAD(&engines->link);
1450
1451         engines->ctx = i915_gem_context_get(ctx);
1452
1453         for_each_gem_engine(ce, engines, it) {
1454                 int err;
1455
1456                 /* serialises with execbuf */
1457                 set_bit(CONTEXT_CLOSED_BIT, &ce->flags);
1458                 if (!intel_context_pin_if_active(ce))
1459                         continue;
1460
1461                 /* Wait until context is finally scheduled out and retired */
1462                 err = i915_sw_fence_await_active(&engines->fence,
1463                                                  &ce->active,
1464                                                  I915_ACTIVE_AWAIT_BARRIER);
1465                 intel_context_unpin(ce);
1466                 if (err)
1467                         goto kill;
1468         }
1469
1470         spin_lock_irq(&ctx->stale.lock);
1471         if (!i915_gem_context_is_closed(ctx))
1472                 list_add_tail(&engines->link, &ctx->stale.engines);
1473         spin_unlock_irq(&ctx->stale.lock);
1474
1475 kill:
1476         if (list_empty(&engines->link)) /* raced, already closed */
1477                 kill_engines(engines, true,
1478                              i915_gem_context_is_persistent(ctx));
1479
1480         i915_sw_fence_commit(&engines->fence);
1481 }
1482
1483 static void set_closed_name(struct i915_gem_context *ctx)
1484 {
1485         char *s;
1486
1487         /* Replace '[]' with '<>' to indicate closed in debug prints */
1488
1489         s = strrchr(ctx->name, '[');
1490         if (!s)
1491                 return;
1492
1493         *s = '<';
1494
1495         s = strchr(s + 1, ']');
1496         if (s)
1497                 *s = '>';
1498 }
1499
1500 static void context_close(struct i915_gem_context *ctx)
1501 {
1502         struct i915_drm_client *client;
1503
1504         /* Flush any concurrent set_engines() */
1505         mutex_lock(&ctx->engines_mutex);
1506         unpin_engines(__context_engines_static(ctx));
1507         engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1));
1508         i915_gem_context_set_closed(ctx);
1509         mutex_unlock(&ctx->engines_mutex);
1510
1511         mutex_lock(&ctx->mutex);
1512
1513         set_closed_name(ctx);
1514
1515         /*
1516          * The LUT uses the VMA as a backpointer to unref the object,
1517          * so we need to clear the LUT before we close all the VMA (inside
1518          * the ppgtt).
1519          */
1520         lut_close(ctx);
1521
1522         ctx->file_priv = ERR_PTR(-EBADF);
1523
1524         spin_lock(&ctx->i915->gem.contexts.lock);
1525         list_del(&ctx->link);
1526         spin_unlock(&ctx->i915->gem.contexts.lock);
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                 vm = &ppgtt->vm;
1628         }
1629         if (vm)
1630                 ctx->vm = vm;
1631
1632         mutex_init(&ctx->engines_mutex);
1633         if (pc->num_user_engines >= 0) {
1634                 i915_gem_context_set_user_engines(ctx);
1635                 e = user_engines(ctx, pc->num_user_engines, pc->user_engines);
1636         } else {
1637                 i915_gem_context_clear_user_engines(ctx);
1638                 e = default_engines(ctx, pc->legacy_rcs_sseu);
1639         }
1640         if (IS_ERR(e)) {
1641                 err = PTR_ERR(e);
1642                 goto err_vm;
1643         }
1644         RCU_INIT_POINTER(ctx->engines, e);
1645
1646         INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
1647         mutex_init(&ctx->lut_mutex);
1648
1649         /* NB: Mark all slices as needing a remap so that when the context first
1650          * loads it will restore whatever remap state already exists. If there
1651          * is no remap info, it will be a NOP. */
1652         ctx->remap_slice = ALL_L3_SLICES(i915);
1653
1654         ctx->user_flags = pc->user_flags;
1655
1656         for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
1657                 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
1658
1659         if (pc->single_timeline) {
1660                 err = drm_syncobj_create(&ctx->syncobj,
1661                                          DRM_SYNCOBJ_CREATE_SIGNALED,
1662                                          NULL);
1663                 if (err)
1664                         goto err_engines;
1665         }
1666
1667         if (pc->uses_protected_content) {
1668                 ctx->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1669                 ctx->uses_protected_content = true;
1670         }
1671
1672         trace_i915_context_create(ctx);
1673
1674         return ctx;
1675
1676 err_engines:
1677         free_engines(e);
1678 err_vm:
1679         if (ctx->vm)
1680                 i915_vm_put(ctx->vm);
1681 err_ctx:
1682         kfree(ctx);
1683         return ERR_PTR(err);
1684 }
1685
1686 static void init_contexts(struct i915_gem_contexts *gc)
1687 {
1688         spin_lock_init(&gc->lock);
1689         INIT_LIST_HEAD(&gc->list);
1690 }
1691
1692 void i915_gem_init__contexts(struct drm_i915_private *i915)
1693 {
1694         init_contexts(&i915->gem.contexts);
1695 }
1696
1697 static void gem_context_register(struct i915_gem_context *ctx,
1698                                  struct drm_i915_file_private *fpriv,
1699                                  u32 id)
1700 {
1701         struct drm_i915_private *i915 = ctx->i915;
1702         void *old;
1703
1704         ctx->file_priv = fpriv;
1705
1706         ctx->pid = get_task_pid(current, PIDTYPE_PID);
1707         ctx->client = i915_drm_client_get(fpriv->client);
1708
1709         snprintf(ctx->name, sizeof(ctx->name), "%s[%d]",
1710                  current->comm, pid_nr(ctx->pid));
1711
1712         /* And finally expose ourselves to userspace via the idr */
1713         old = xa_store(&fpriv->context_xa, id, ctx, GFP_KERNEL);
1714         WARN_ON(old);
1715
1716         spin_lock(&ctx->client->ctx_lock);
1717         list_add_tail_rcu(&ctx->client_link, &ctx->client->ctx_list);
1718         spin_unlock(&ctx->client->ctx_lock);
1719
1720         spin_lock(&i915->gem.contexts.lock);
1721         list_add_tail(&ctx->link, &i915->gem.contexts.list);
1722         spin_unlock(&i915->gem.contexts.lock);
1723 }
1724
1725 int i915_gem_context_open(struct drm_i915_private *i915,
1726                           struct drm_file *file)
1727 {
1728         struct drm_i915_file_private *file_priv = file->driver_priv;
1729         struct i915_gem_proto_context *pc;
1730         struct i915_gem_context *ctx;
1731         int err;
1732
1733         mutex_init(&file_priv->proto_context_lock);
1734         xa_init_flags(&file_priv->proto_context_xa, XA_FLAGS_ALLOC);
1735
1736         /* 0 reserved for the default context */
1737         xa_init_flags(&file_priv->context_xa, XA_FLAGS_ALLOC1);
1738
1739         /* 0 reserved for invalid/unassigned ppgtt */
1740         xa_init_flags(&file_priv->vm_xa, XA_FLAGS_ALLOC1);
1741
1742         pc = proto_context_create(i915, 0);
1743         if (IS_ERR(pc)) {
1744                 err = PTR_ERR(pc);
1745                 goto err;
1746         }
1747
1748         ctx = i915_gem_create_context(i915, pc);
1749         proto_context_close(i915, pc);
1750         if (IS_ERR(ctx)) {
1751                 err = PTR_ERR(ctx);
1752                 goto err;
1753         }
1754
1755         gem_context_register(ctx, file_priv, 0);
1756
1757         return 0;
1758
1759 err:
1760         xa_destroy(&file_priv->vm_xa);
1761         xa_destroy(&file_priv->context_xa);
1762         xa_destroy(&file_priv->proto_context_xa);
1763         mutex_destroy(&file_priv->proto_context_lock);
1764         return err;
1765 }
1766
1767 void i915_gem_context_close(struct drm_file *file)
1768 {
1769         struct drm_i915_file_private *file_priv = file->driver_priv;
1770         struct i915_gem_proto_context *pc;
1771         struct i915_address_space *vm;
1772         struct i915_gem_context *ctx;
1773         unsigned long idx;
1774
1775         xa_for_each(&file_priv->proto_context_xa, idx, pc)
1776                 proto_context_close(file_priv->dev_priv, pc);
1777         xa_destroy(&file_priv->proto_context_xa);
1778         mutex_destroy(&file_priv->proto_context_lock);
1779
1780         xa_for_each(&file_priv->context_xa, idx, ctx)
1781                 context_close(ctx);
1782         xa_destroy(&file_priv->context_xa);
1783
1784         xa_for_each(&file_priv->vm_xa, idx, vm)
1785                 i915_vm_put(vm);
1786         xa_destroy(&file_priv->vm_xa);
1787 }
1788
1789 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
1790                              struct drm_file *file)
1791 {
1792         struct drm_i915_private *i915 = to_i915(dev);
1793         struct drm_i915_gem_vm_control *args = data;
1794         struct drm_i915_file_private *file_priv = file->driver_priv;
1795         struct i915_ppgtt *ppgtt;
1796         u32 id;
1797         int err;
1798
1799         if (!HAS_FULL_PPGTT(i915))
1800                 return -ENODEV;
1801
1802         if (args->flags)
1803                 return -EINVAL;
1804
1805         ppgtt = i915_ppgtt_create(to_gt(i915), 0);
1806         if (IS_ERR(ppgtt))
1807                 return PTR_ERR(ppgtt);
1808
1809         if (args->extensions) {
1810                 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
1811                                            NULL, 0,
1812                                            ppgtt);
1813                 if (err)
1814                         goto err_put;
1815         }
1816
1817         err = xa_alloc(&file_priv->vm_xa, &id, &ppgtt->vm,
1818                        xa_limit_32b, GFP_KERNEL);
1819         if (err)
1820                 goto err_put;
1821
1822         GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1823         args->vm_id = id;
1824         return 0;
1825
1826 err_put:
1827         i915_vm_put(&ppgtt->vm);
1828         return err;
1829 }
1830
1831 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
1832                               struct drm_file *file)
1833 {
1834         struct drm_i915_file_private *file_priv = file->driver_priv;
1835         struct drm_i915_gem_vm_control *args = data;
1836         struct i915_address_space *vm;
1837
1838         if (args->flags)
1839                 return -EINVAL;
1840
1841         if (args->extensions)
1842                 return -EINVAL;
1843
1844         vm = xa_erase(&file_priv->vm_xa, args->vm_id);
1845         if (!vm)
1846                 return -ENOENT;
1847
1848         i915_vm_put(vm);
1849         return 0;
1850 }
1851
1852 static int get_ppgtt(struct drm_i915_file_private *file_priv,
1853                      struct i915_gem_context *ctx,
1854                      struct drm_i915_gem_context_param *args)
1855 {
1856         struct i915_address_space *vm;
1857         int err;
1858         u32 id;
1859
1860         if (!i915_gem_context_has_full_ppgtt(ctx))
1861                 return -ENODEV;
1862
1863         vm = ctx->vm;
1864         GEM_BUG_ON(!vm);
1865
1866         err = xa_alloc(&file_priv->vm_xa, &id, vm, xa_limit_32b, GFP_KERNEL);
1867         if (err)
1868                 return err;
1869
1870         i915_vm_get(vm);
1871
1872         GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1873         args->value = id;
1874         args->size = 0;
1875
1876         return err;
1877 }
1878
1879 int
1880 i915_gem_user_to_context_sseu(struct intel_gt *gt,
1881                               const struct drm_i915_gem_context_param_sseu *user,
1882                               struct intel_sseu *context)
1883 {
1884         const struct sseu_dev_info *device = &gt->info.sseu;
1885         struct drm_i915_private *i915 = gt->i915;
1886         unsigned int dev_subslice_mask = intel_sseu_get_hsw_subslices(device, 0);
1887
1888         /* No zeros in any field. */
1889         if (!user->slice_mask || !user->subslice_mask ||
1890             !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1891                 return -EINVAL;
1892
1893         /* Max > min. */
1894         if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1895                 return -EINVAL;
1896
1897         /*
1898          * Some future proofing on the types since the uAPI is wider than the
1899          * current internal implementation.
1900          */
1901         if (overflows_type(user->slice_mask, context->slice_mask) ||
1902             overflows_type(user->subslice_mask, context->subslice_mask) ||
1903             overflows_type(user->min_eus_per_subslice,
1904                            context->min_eus_per_subslice) ||
1905             overflows_type(user->max_eus_per_subslice,
1906                            context->max_eus_per_subslice))
1907                 return -EINVAL;
1908
1909         /* Check validity against hardware. */
1910         if (user->slice_mask & ~device->slice_mask)
1911                 return -EINVAL;
1912
1913         if (user->subslice_mask & ~dev_subslice_mask)
1914                 return -EINVAL;
1915
1916         if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1917                 return -EINVAL;
1918
1919         context->slice_mask = user->slice_mask;
1920         context->subslice_mask = user->subslice_mask;
1921         context->min_eus_per_subslice = user->min_eus_per_subslice;
1922         context->max_eus_per_subslice = user->max_eus_per_subslice;
1923
1924         /* Part specific restrictions. */
1925         if (GRAPHICS_VER(i915) == 11) {
1926                 unsigned int hw_s = hweight8(device->slice_mask);
1927                 unsigned int hw_ss_per_s = hweight8(dev_subslice_mask);
1928                 unsigned int req_s = hweight8(context->slice_mask);
1929                 unsigned int req_ss = hweight8(context->subslice_mask);
1930
1931                 /*
1932                  * Only full subslice enablement is possible if more than one
1933                  * slice is turned on.
1934                  */
1935                 if (req_s > 1 && req_ss != hw_ss_per_s)
1936                         return -EINVAL;
1937
1938                 /*
1939                  * If more than four (SScount bitfield limit) subslices are
1940                  * requested then the number has to be even.
1941                  */
1942                 if (req_ss > 4 && (req_ss & 1))
1943                         return -EINVAL;
1944
1945                 /*
1946                  * If only one slice is enabled and subslice count is below the
1947                  * device full enablement, it must be at most half of the all
1948                  * available subslices.
1949                  */
1950                 if (req_s == 1 && req_ss < hw_ss_per_s &&
1951                     req_ss > (hw_ss_per_s / 2))
1952                         return -EINVAL;
1953
1954                 /* ABI restriction - VME use case only. */
1955
1956                 /* All slices or one slice only. */
1957                 if (req_s != 1 && req_s != hw_s)
1958                         return -EINVAL;
1959
1960                 /*
1961                  * Half subslices or full enablement only when one slice is
1962                  * enabled.
1963                  */
1964                 if (req_s == 1 &&
1965                     (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1966                         return -EINVAL;
1967
1968                 /* No EU configuration changes. */
1969                 if ((user->min_eus_per_subslice !=
1970                      device->max_eus_per_subslice) ||
1971                     (user->max_eus_per_subslice !=
1972                      device->max_eus_per_subslice))
1973                         return -EINVAL;
1974         }
1975
1976         return 0;
1977 }
1978
1979 static int set_sseu(struct i915_gem_context *ctx,
1980                     struct drm_i915_gem_context_param *args)
1981 {
1982         struct drm_i915_private *i915 = ctx->i915;
1983         struct drm_i915_gem_context_param_sseu user_sseu;
1984         struct intel_context *ce;
1985         struct intel_sseu sseu;
1986         unsigned long lookup;
1987         int ret;
1988
1989         if (args->size < sizeof(user_sseu))
1990                 return -EINVAL;
1991
1992         if (GRAPHICS_VER(i915) != 11)
1993                 return -ENODEV;
1994
1995         if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1996                            sizeof(user_sseu)))
1997                 return -EFAULT;
1998
1999         if (user_sseu.rsvd)
2000                 return -EINVAL;
2001
2002         if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2003                 return -EINVAL;
2004
2005         lookup = 0;
2006         if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2007                 lookup |= LOOKUP_USER_INDEX;
2008
2009         ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2010         if (IS_ERR(ce))
2011                 return PTR_ERR(ce);
2012
2013         /* Only render engine supports RPCS configuration. */
2014         if (ce->engine->class != RENDER_CLASS) {
2015                 ret = -ENODEV;
2016                 goto out_ce;
2017         }
2018
2019         ret = i915_gem_user_to_context_sseu(ce->engine->gt, &user_sseu, &sseu);
2020         if (ret)
2021                 goto out_ce;
2022
2023         ret = intel_context_reconfigure_sseu(ce, sseu);
2024         if (ret)
2025                 goto out_ce;
2026
2027         args->size = sizeof(user_sseu);
2028
2029 out_ce:
2030         intel_context_put(ce);
2031         return ret;
2032 }
2033
2034 static int
2035 set_persistence(struct i915_gem_context *ctx,
2036                 const struct drm_i915_gem_context_param *args)
2037 {
2038         if (args->size)
2039                 return -EINVAL;
2040
2041         return __context_set_persistence(ctx, args->value);
2042 }
2043
2044 static int set_priority(struct i915_gem_context *ctx,
2045                         const struct drm_i915_gem_context_param *args)
2046 {
2047         struct i915_gem_engines_iter it;
2048         struct intel_context *ce;
2049         int err;
2050
2051         err = validate_priority(ctx->i915, args);
2052         if (err)
2053                 return err;
2054
2055         ctx->sched.priority = args->value;
2056
2057         for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
2058                 if (!intel_engine_has_timeslices(ce->engine))
2059                         continue;
2060
2061                 if (ctx->sched.priority >= I915_PRIORITY_NORMAL &&
2062                     intel_engine_has_semaphores(ce->engine))
2063                         intel_context_set_use_semaphores(ce);
2064                 else
2065                         intel_context_clear_use_semaphores(ce);
2066         }
2067         i915_gem_context_unlock_engines(ctx);
2068
2069         return 0;
2070 }
2071
2072 static int get_protected(struct i915_gem_context *ctx,
2073                          struct drm_i915_gem_context_param *args)
2074 {
2075         args->size = 0;
2076         args->value = i915_gem_context_uses_protected_content(ctx);
2077
2078         return 0;
2079 }
2080
2081 static int ctx_setparam(struct drm_i915_file_private *fpriv,
2082                         struct i915_gem_context *ctx,
2083                         struct drm_i915_gem_context_param *args)
2084 {
2085         int ret = 0;
2086
2087         switch (args->param) {
2088         case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2089                 if (args->size)
2090                         ret = -EINVAL;
2091                 else if (args->value)
2092                         i915_gem_context_set_no_error_capture(ctx);
2093                 else
2094                         i915_gem_context_clear_no_error_capture(ctx);
2095                 break;
2096
2097         case I915_CONTEXT_PARAM_BANNABLE:
2098                 if (args->size)
2099                         ret = -EINVAL;
2100                 else if (!capable(CAP_SYS_ADMIN) && !args->value)
2101                         ret = -EPERM;
2102                 else if (args->value)
2103                         i915_gem_context_set_bannable(ctx);
2104                 else if (i915_gem_context_uses_protected_content(ctx))
2105                         ret = -EPERM; /* can't clear this for protected contexts */
2106                 else
2107                         i915_gem_context_clear_bannable(ctx);
2108                 break;
2109
2110         case I915_CONTEXT_PARAM_RECOVERABLE:
2111                 if (args->size)
2112                         ret = -EINVAL;
2113                 else if (!args->value)
2114                         i915_gem_context_clear_recoverable(ctx);
2115                 else if (i915_gem_context_uses_protected_content(ctx))
2116                         ret = -EPERM; /* can't set this for protected contexts */
2117                 else
2118                         i915_gem_context_set_recoverable(ctx);
2119                 break;
2120
2121         case I915_CONTEXT_PARAM_PRIORITY:
2122                 ret = set_priority(ctx, args);
2123                 break;
2124
2125         case I915_CONTEXT_PARAM_SSEU:
2126                 ret = set_sseu(ctx, args);
2127                 break;
2128
2129         case I915_CONTEXT_PARAM_PERSISTENCE:
2130                 ret = set_persistence(ctx, args);
2131                 break;
2132
2133         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2134         case I915_CONTEXT_PARAM_NO_ZEROMAP:
2135         case I915_CONTEXT_PARAM_BAN_PERIOD:
2136         case I915_CONTEXT_PARAM_RINGSIZE:
2137         case I915_CONTEXT_PARAM_VM:
2138         case I915_CONTEXT_PARAM_ENGINES:
2139         default:
2140                 ret = -EINVAL;
2141                 break;
2142         }
2143
2144         return ret;
2145 }
2146
2147 struct create_ext {
2148         struct i915_gem_proto_context *pc;
2149         struct drm_i915_file_private *fpriv;
2150 };
2151
2152 static int create_setparam(struct i915_user_extension __user *ext, void *data)
2153 {
2154         struct drm_i915_gem_context_create_ext_setparam local;
2155         const struct create_ext *arg = data;
2156
2157         if (copy_from_user(&local, ext, sizeof(local)))
2158                 return -EFAULT;
2159
2160         if (local.param.ctx_id)
2161                 return -EINVAL;
2162
2163         return set_proto_ctx_param(arg->fpriv, arg->pc, &local.param);
2164 }
2165
2166 static int invalid_ext(struct i915_user_extension __user *ext, void *data)
2167 {
2168         return -EINVAL;
2169 }
2170
2171 static const i915_user_extension_fn create_extensions[] = {
2172         [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2173         [I915_CONTEXT_CREATE_EXT_CLONE] = invalid_ext,
2174 };
2175
2176 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2177 {
2178         return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2179 }
2180
2181 static inline struct i915_gem_context *
2182 __context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2183 {
2184         struct i915_gem_context *ctx;
2185
2186         rcu_read_lock();
2187         ctx = xa_load(&file_priv->context_xa, id);
2188         if (ctx && !kref_get_unless_zero(&ctx->ref))
2189                 ctx = NULL;
2190         rcu_read_unlock();
2191
2192         return ctx;
2193 }
2194
2195 static struct i915_gem_context *
2196 finalize_create_context_locked(struct drm_i915_file_private *file_priv,
2197                                struct i915_gem_proto_context *pc, u32 id)
2198 {
2199         struct i915_gem_context *ctx;
2200         void *old;
2201
2202         lockdep_assert_held(&file_priv->proto_context_lock);
2203
2204         ctx = i915_gem_create_context(file_priv->dev_priv, pc);
2205         if (IS_ERR(ctx))
2206                 return ctx;
2207
2208         gem_context_register(ctx, file_priv, id);
2209
2210         old = xa_erase(&file_priv->proto_context_xa, id);
2211         GEM_BUG_ON(old != pc);
2212         proto_context_close(file_priv->dev_priv, pc);
2213
2214         /* One for the xarray and one for the caller */
2215         return i915_gem_context_get(ctx);
2216 }
2217
2218 struct i915_gem_context *
2219 i915_gem_context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2220 {
2221         struct i915_gem_proto_context *pc;
2222         struct i915_gem_context *ctx;
2223
2224         ctx = __context_lookup(file_priv, id);
2225         if (ctx)
2226                 return ctx;
2227
2228         mutex_lock(&file_priv->proto_context_lock);
2229         /* Try one more time under the lock */
2230         ctx = __context_lookup(file_priv, id);
2231         if (!ctx) {
2232                 pc = xa_load(&file_priv->proto_context_xa, id);
2233                 if (!pc)
2234                         ctx = ERR_PTR(-ENOENT);
2235                 else
2236                         ctx = finalize_create_context_locked(file_priv, pc, id);
2237         }
2238         mutex_unlock(&file_priv->proto_context_lock);
2239
2240         return ctx;
2241 }
2242
2243 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2244                                   struct drm_file *file)
2245 {
2246         struct drm_i915_private *i915 = to_i915(dev);
2247         struct drm_i915_gem_context_create_ext *args = data;
2248         struct create_ext ext_data;
2249         int ret;
2250         u32 id;
2251
2252         if (!DRIVER_CAPS(i915)->has_logical_contexts)
2253                 return -ENODEV;
2254
2255         if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2256                 return -EINVAL;
2257
2258         ret = intel_gt_terminally_wedged(to_gt(i915));
2259         if (ret)
2260                 return ret;
2261
2262         ext_data.fpriv = file->driver_priv;
2263         if (client_is_banned(ext_data.fpriv)) {
2264                 drm_dbg(&i915->drm,
2265                         "client %s[%d] banned from creating ctx\n",
2266                         current->comm, task_pid_nr(current));
2267                 return -EIO;
2268         }
2269
2270         ext_data.pc = proto_context_create(i915, args->flags);
2271         if (IS_ERR(ext_data.pc))
2272                 return PTR_ERR(ext_data.pc);
2273
2274         if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2275                 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2276                                            create_extensions,
2277                                            ARRAY_SIZE(create_extensions),
2278                                            &ext_data);
2279                 if (ret)
2280                         goto err_pc;
2281         }
2282
2283         if (GRAPHICS_VER(i915) > 12) {
2284                 struct i915_gem_context *ctx;
2285
2286                 /* Get ourselves a context ID */
2287                 ret = xa_alloc(&ext_data.fpriv->context_xa, &id, NULL,
2288                                xa_limit_32b, GFP_KERNEL);
2289                 if (ret)
2290                         goto err_pc;
2291
2292                 ctx = i915_gem_create_context(i915, ext_data.pc);
2293                 if (IS_ERR(ctx)) {
2294                         ret = PTR_ERR(ctx);
2295                         goto err_pc;
2296                 }
2297
2298                 proto_context_close(i915, ext_data.pc);
2299                 gem_context_register(ctx, ext_data.fpriv, id);
2300         } else {
2301                 ret = proto_context_register(ext_data.fpriv, ext_data.pc, &id);
2302                 if (ret < 0)
2303                         goto err_pc;
2304         }
2305
2306         args->ctx_id = id;
2307         drm_dbg(&i915->drm, "HW context %d created\n", args->ctx_id);
2308
2309         return 0;
2310
2311 err_pc:
2312         proto_context_close(i915, ext_data.pc);
2313         return ret;
2314 }
2315
2316 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2317                                    struct drm_file *file)
2318 {
2319         struct drm_i915_gem_context_destroy *args = data;
2320         struct drm_i915_file_private *file_priv = file->driver_priv;
2321         struct i915_gem_proto_context *pc;
2322         struct i915_gem_context *ctx;
2323
2324         if (args->pad != 0)
2325                 return -EINVAL;
2326
2327         if (!args->ctx_id)
2328                 return -ENOENT;
2329
2330         /* We need to hold the proto-context lock here to prevent races
2331          * with finalize_create_context_locked().
2332          */
2333         mutex_lock(&file_priv->proto_context_lock);
2334         ctx = xa_erase(&file_priv->context_xa, args->ctx_id);
2335         pc = xa_erase(&file_priv->proto_context_xa, args->ctx_id);
2336         mutex_unlock(&file_priv->proto_context_lock);
2337
2338         if (!ctx && !pc)
2339                 return -ENOENT;
2340         GEM_WARN_ON(ctx && pc);
2341
2342         if (pc)
2343                 proto_context_close(file_priv->dev_priv, pc);
2344
2345         if (ctx)
2346                 context_close(ctx);
2347
2348         return 0;
2349 }
2350
2351 static int get_sseu(struct i915_gem_context *ctx,
2352                     struct drm_i915_gem_context_param *args)
2353 {
2354         struct drm_i915_gem_context_param_sseu user_sseu;
2355         struct intel_context *ce;
2356         unsigned long lookup;
2357         int err;
2358
2359         if (args->size == 0)
2360                 goto out;
2361         else if (args->size < sizeof(user_sseu))
2362                 return -EINVAL;
2363
2364         if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2365                            sizeof(user_sseu)))
2366                 return -EFAULT;
2367
2368         if (user_sseu.rsvd)
2369                 return -EINVAL;
2370
2371         if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2372                 return -EINVAL;
2373
2374         lookup = 0;
2375         if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2376                 lookup |= LOOKUP_USER_INDEX;
2377
2378         ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2379         if (IS_ERR(ce))
2380                 return PTR_ERR(ce);
2381
2382         err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2383         if (err) {
2384                 intel_context_put(ce);
2385                 return err;
2386         }
2387
2388         user_sseu.slice_mask = ce->sseu.slice_mask;
2389         user_sseu.subslice_mask = ce->sseu.subslice_mask;
2390         user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2391         user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2392
2393         intel_context_unlock_pinned(ce);
2394         intel_context_put(ce);
2395
2396         if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2397                          sizeof(user_sseu)))
2398                 return -EFAULT;
2399
2400 out:
2401         args->size = sizeof(user_sseu);
2402
2403         return 0;
2404 }
2405
2406 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2407                                     struct drm_file *file)
2408 {
2409         struct drm_i915_file_private *file_priv = file->driver_priv;
2410         struct drm_i915_gem_context_param *args = data;
2411         struct i915_gem_context *ctx;
2412         struct i915_address_space *vm;
2413         int ret = 0;
2414
2415         ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2416         if (IS_ERR(ctx))
2417                 return PTR_ERR(ctx);
2418
2419         switch (args->param) {
2420         case I915_CONTEXT_PARAM_GTT_SIZE:
2421                 args->size = 0;
2422                 vm = i915_gem_context_get_eb_vm(ctx);
2423                 args->value = vm->total;
2424                 i915_vm_put(vm);
2425
2426                 break;
2427
2428         case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2429                 args->size = 0;
2430                 args->value = i915_gem_context_no_error_capture(ctx);
2431                 break;
2432
2433         case I915_CONTEXT_PARAM_BANNABLE:
2434                 args->size = 0;
2435                 args->value = i915_gem_context_is_bannable(ctx);
2436                 break;
2437
2438         case I915_CONTEXT_PARAM_RECOVERABLE:
2439                 args->size = 0;
2440                 args->value = i915_gem_context_is_recoverable(ctx);
2441                 break;
2442
2443         case I915_CONTEXT_PARAM_PRIORITY:
2444                 args->size = 0;
2445                 args->value = ctx->sched.priority;
2446                 break;
2447
2448         case I915_CONTEXT_PARAM_SSEU:
2449                 ret = get_sseu(ctx, args);
2450                 break;
2451
2452         case I915_CONTEXT_PARAM_VM:
2453                 ret = get_ppgtt(file_priv, ctx, args);
2454                 break;
2455
2456         case I915_CONTEXT_PARAM_PERSISTENCE:
2457                 args->size = 0;
2458                 args->value = i915_gem_context_is_persistent(ctx);
2459                 break;
2460
2461         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2462                 ret = get_protected(ctx, args);
2463                 break;
2464
2465         case I915_CONTEXT_PARAM_NO_ZEROMAP:
2466         case I915_CONTEXT_PARAM_BAN_PERIOD:
2467         case I915_CONTEXT_PARAM_ENGINES:
2468         case I915_CONTEXT_PARAM_RINGSIZE:
2469         default:
2470                 ret = -EINVAL;
2471                 break;
2472         }
2473
2474         i915_gem_context_put(ctx);
2475         return ret;
2476 }
2477
2478 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2479                                     struct drm_file *file)
2480 {
2481         struct drm_i915_file_private *file_priv = file->driver_priv;
2482         struct drm_i915_gem_context_param *args = data;
2483         struct i915_gem_proto_context *pc;
2484         struct i915_gem_context *ctx;
2485         int ret = 0;
2486
2487         mutex_lock(&file_priv->proto_context_lock);
2488         ctx = __context_lookup(file_priv, args->ctx_id);
2489         if (!ctx) {
2490                 pc = xa_load(&file_priv->proto_context_xa, args->ctx_id);
2491                 if (pc) {
2492                         /* Contexts should be finalized inside
2493                          * GEM_CONTEXT_CREATE starting with graphics
2494                          * version 13.
2495                          */
2496                         WARN_ON(GRAPHICS_VER(file_priv->dev_priv) > 12);
2497                         ret = set_proto_ctx_param(file_priv, pc, args);
2498                 } else {
2499                         ret = -ENOENT;
2500                 }
2501         }
2502         mutex_unlock(&file_priv->proto_context_lock);
2503
2504         if (ctx) {
2505                 ret = ctx_setparam(file_priv, ctx, args);
2506                 i915_gem_context_put(ctx);
2507         }
2508
2509         return ret;
2510 }
2511
2512 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2513                                        void *data, struct drm_file *file)
2514 {
2515         struct drm_i915_private *i915 = to_i915(dev);
2516         struct drm_i915_reset_stats *args = data;
2517         struct i915_gem_context *ctx;
2518
2519         if (args->flags || args->pad)
2520                 return -EINVAL;
2521
2522         ctx = i915_gem_context_lookup(file->driver_priv, args->ctx_id);
2523         if (IS_ERR(ctx))
2524                 return PTR_ERR(ctx);
2525
2526         /*
2527          * We opt for unserialised reads here. This may result in tearing
2528          * in the extremely unlikely event of a GPU hang on this context
2529          * as we are querying them. If we need that extra layer of protection,
2530          * we should wrap the hangstats with a seqlock.
2531          */
2532
2533         if (capable(CAP_SYS_ADMIN))
2534                 args->reset_count = i915_reset_count(&i915->gpu_error);
2535         else
2536                 args->reset_count = 0;
2537
2538         args->batch_active = atomic_read(&ctx->guilty_count);
2539         args->batch_pending = atomic_read(&ctx->active_count);
2540
2541         i915_gem_context_put(ctx);
2542         return 0;
2543 }
2544
2545 /* GEM context-engines iterator: for_each_gem_engine() */
2546 struct intel_context *
2547 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2548 {
2549         const struct i915_gem_engines *e = it->engines;
2550         struct intel_context *ctx;
2551
2552         if (unlikely(!e))
2553                 return NULL;
2554
2555         do {
2556                 if (it->idx >= e->num_engines)
2557                         return NULL;
2558
2559                 ctx = e->engines[it->idx++];
2560         } while (!ctx);
2561
2562         return ctx;
2563 }
2564
2565 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2566 #include "selftests/mock_context.c"
2567 #include "selftests/i915_gem_context.c"
2568 #endif
2569
2570 void i915_gem_context_module_exit(void)
2571 {
2572         kmem_cache_destroy(slab_luts);
2573 }
2574
2575 int __init i915_gem_context_module_init(void)
2576 {
2577         slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2578         if (!slab_luts)
2579                 return -ENOMEM;
2580
2581         return 0;
2582 }