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                 ret = proto_context_set_persistence(fpriv->dev_priv, pc,
 937                                                     args->value);
 938                 break;
 939
 940         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
 941                 ret = proto_context_set_protected(fpriv->dev_priv, pc,
 942                                                   args->value);
 943                 break;
 944
 945         case I915_CONTEXT_PARAM_NO_ZEROMAP:
 946         case I915_CONTEXT_PARAM_BAN_PERIOD:
 947         case I915_CONTEXT_PARAM_RINGSIZE:
 948         default:
 949                 ret = -EINVAL;
 950                 break;
 951         }
 952
 953         return ret;
 954 }
 955
 956 static int intel_context_set_gem(struct intel_context *ce,
 957                                  struct i915_gem_context *ctx,
 958                                  struct intel_sseu sseu)
 959 {
 960         int ret = 0;
 961
 962         GEM_BUG_ON(rcu_access_pointer(ce->gem_context));
 963         RCU_INIT_POINTER(ce->gem_context, ctx);
 964
 965         GEM_BUG_ON(intel_context_is_pinned(ce));
 966         ce->ring_size = SZ_16K;
 967
 968         i915_vm_put(ce->vm);
 969         ce->vm = i915_gem_context_get_eb_vm(ctx);
 970
 971         if (ctx->sched.priority >= I915_PRIORITY_NORMAL &&
 972             intel_engine_has_timeslices(ce->engine) &&
 973             intel_engine_has_semaphores(ce->engine))
 974                 __set_bit(CONTEXT_USE_SEMAPHORES, &ce->flags);
 975
 976         if (CONFIG_DRM_I915_REQUEST_TIMEOUT &&
 977             ctx->i915->params.request_timeout_ms) {
 978                 unsigned int timeout_ms = ctx->i915->params.request_timeout_ms;
 979
 980                 intel_context_set_watchdog_us(ce, (u64)timeout_ms * 1000);
 981         }
 982
 983         /* A valid SSEU has no zero fields */
 984         if (sseu.slice_mask && !WARN_ON(ce->engine->class != RENDER_CLASS))
 985                 ret = intel_context_reconfigure_sseu(ce, sseu);
 986
 987         return ret;
 988 }
 989
 990 static void __unpin_engines(struct i915_gem_engines *e, unsigned int count)
 991 {
 992         while (count--) {
 993                 struct intel_context *ce = e->engines[count], *child;
 994
 995                 if (!ce || !test_bit(CONTEXT_PERMA_PIN, &ce->flags))
 996                         continue;
 997
 998                 for_each_child(ce, child)
 999                         intel_context_unpin(child);
1000                 intel_context_unpin(ce);
1001         }
1002 }
1003
1004 static void unpin_engines(struct i915_gem_engines *e)
1005 {
1006         __unpin_engines(e, e->num_engines);
1007 }
1008
1009 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
1010 {
1011         while (count--) {
1012                 if (!e->engines[count])
1013                         continue;
1014
1015                 intel_context_put(e->engines[count]);
1016         }
1017         kfree(e);
1018 }
1019
1020 static void free_engines(struct i915_gem_engines *e)
1021 {
1022         __free_engines(e, e->num_engines);
1023 }
1024
1025 static void free_engines_rcu(struct rcu_head *rcu)
1026 {
1027         struct i915_gem_engines *engines =
1028                 container_of(rcu, struct i915_gem_engines, rcu);
1029
1030         i915_sw_fence_fini(&engines->fence);
1031         free_engines(engines);
1032 }
1033
1034 static void accumulate_runtime(struct i915_drm_client *client,
1035                                struct i915_gem_engines *engines)
1036 {
1037         struct i915_gem_engines_iter it;
1038         struct intel_context *ce;
1039
1040         if (!client)
1041                 return;
1042
1043         /* Transfer accumulated runtime to the parent GEM context. */
1044         for_each_gem_engine(ce, engines, it) {
1045                 unsigned int class = ce->engine->uabi_class;
1046
1047                 GEM_BUG_ON(class >= ARRAY_SIZE(client->past_runtime));
1048                 atomic64_add(intel_context_get_total_runtime_ns(ce),
1049                              &client->past_runtime[class]);
1050         }
1051 }
1052
1053 static int
1054 engines_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
1055 {
1056         struct i915_gem_engines *engines =
1057                 container_of(fence, typeof(*engines), fence);
1058         struct i915_gem_context *ctx = engines->ctx;
1059
1060         switch (state) {
1061         case FENCE_COMPLETE:
1062                 if (!list_empty(&engines->link)) {
1063                         unsigned long flags;
1064
1065                         spin_lock_irqsave(&ctx->stale.lock, flags);
1066                         list_del(&engines->link);
1067                         spin_unlock_irqrestore(&ctx->stale.lock, flags);
1068                 }
1069                 accumulate_runtime(ctx->client, engines);
1070                 i915_gem_context_put(ctx);
1071
1072                 break;
1073
1074         case FENCE_FREE:
1075                 init_rcu_head(&engines->rcu);
1076                 call_rcu(&engines->rcu, free_engines_rcu);
1077                 break;
1078         }
1079
1080         return NOTIFY_DONE;
1081 }
1082
1083 static struct i915_gem_engines *alloc_engines(unsigned int count)
1084 {
1085         struct i915_gem_engines *e;
1086
1087         e = kzalloc(struct_size(e, engines, count), GFP_KERNEL);
1088         if (!e)
1089                 return NULL;
1090
1091         i915_sw_fence_init(&e->fence, engines_notify);
1092         return e;
1093 }
1094
1095 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx,
1096                                                 struct intel_sseu rcs_sseu)
1097 {
1098         const struct intel_gt *gt = to_gt(ctx->i915);
1099         struct intel_engine_cs *engine;
1100         struct i915_gem_engines *e, *err;
1101         enum intel_engine_id id;
1102
1103         e = alloc_engines(I915_NUM_ENGINES);
1104         if (!e)
1105                 return ERR_PTR(-ENOMEM);
1106
1107         for_each_engine(engine, gt, id) {
1108                 struct intel_context *ce;
1109                 struct intel_sseu sseu = {};
1110                 int ret;
1111
1112                 if (engine->legacy_idx == INVALID_ENGINE)
1113                         continue;
1114
1115                 GEM_BUG_ON(engine->legacy_idx >= I915_NUM_ENGINES);
1116                 GEM_BUG_ON(e->engines[engine->legacy_idx]);
1117
1118                 ce = intel_context_create(engine);
1119                 if (IS_ERR(ce)) {
1120                         err = ERR_CAST(ce);
1121                         goto free_engines;
1122                 }
1123
1124                 e->engines[engine->legacy_idx] = ce;
1125                 e->num_engines = max(e->num_engines, engine->legacy_idx + 1);
1126
1127                 if (engine->class == RENDER_CLASS)
1128                         sseu = rcs_sseu;
1129
1130                 ret = intel_context_set_gem(ce, ctx, sseu);
1131                 if (ret) {
1132                         err = ERR_PTR(ret);
1133                         goto free_engines;
1134                 }
1135
1136         }
1137
1138         return e;
1139
1140 free_engines:
1141         free_engines(e);
1142         return err;
1143 }
1144
1145 static int perma_pin_contexts(struct intel_context *ce)
1146 {
1147         struct intel_context *child;
1148         int i = 0, j = 0, ret;
1149
1150         GEM_BUG_ON(!intel_context_is_parent(ce));
1151
1152         ret = intel_context_pin(ce);
1153         if (unlikely(ret))
1154                 return ret;
1155
1156         for_each_child(ce, child) {
1157                 ret = intel_context_pin(child);
1158                 if (unlikely(ret))
1159                         goto unwind;
1160                 ++i;
1161         }
1162
1163         set_bit(CONTEXT_PERMA_PIN, &ce->flags);
1164
1165         return 0;
1166
1167 unwind:
1168         intel_context_unpin(ce);
1169         for_each_child(ce, child) {
1170                 if (j++ < i)
1171                         intel_context_unpin(child);
1172                 else
1173                         break;
1174         }
1175
1176         return ret;
1177 }
1178
1179 static struct i915_gem_engines *user_engines(struct i915_gem_context *ctx,
1180                                              unsigned int num_engines,
1181                                              struct i915_gem_proto_engine *pe)
1182 {
1183         struct i915_gem_engines *e, *err;
1184         unsigned int n;
1185
1186         e = alloc_engines(num_engines);
1187         if (!e)
1188                 return ERR_PTR(-ENOMEM);
1189         e->num_engines = num_engines;
1190
1191         for (n = 0; n < num_engines; n++) {
1192                 struct intel_context *ce, *child;
1193                 int ret;
1194
1195                 switch (pe[n].type) {
1196                 case I915_GEM_ENGINE_TYPE_PHYSICAL:
1197                         ce = intel_context_create(pe[n].engine);
1198                         break;
1199
1200                 case I915_GEM_ENGINE_TYPE_BALANCED:
1201                         ce = intel_engine_create_virtual(pe[n].siblings,
1202                                                          pe[n].num_siblings, 0);
1203                         break;
1204
1205                 case I915_GEM_ENGINE_TYPE_PARALLEL:
1206                         ce = intel_engine_create_parallel(pe[n].siblings,
1207                                                           pe[n].num_siblings,
1208                                                           pe[n].width);
1209                         break;
1210
1211                 case I915_GEM_ENGINE_TYPE_INVALID:
1212                 default:
1213                         GEM_WARN_ON(pe[n].type != I915_GEM_ENGINE_TYPE_INVALID);
1214                         continue;
1215                 }
1216
1217                 if (IS_ERR(ce)) {
1218                         err = ERR_CAST(ce);
1219                         goto free_engines;
1220                 }
1221
1222                 e->engines[n] = ce;
1223
1224                 ret = intel_context_set_gem(ce, ctx, pe->sseu);
1225                 if (ret) {
1226                         err = ERR_PTR(ret);
1227                         goto free_engines;
1228                 }
1229                 for_each_child(ce, child) {
1230                         ret = intel_context_set_gem(child, ctx, pe->sseu);
1231                         if (ret) {
1232                                 err = ERR_PTR(ret);
1233                                 goto free_engines;
1234                         }
1235                 }
1236
1237                 /*
1238                  * XXX: Must be done after calling intel_context_set_gem as that
1239                  * function changes the ring size. The ring is allocated when
1240                  * the context is pinned. If the ring size is changed after
1241                  * allocation we have a mismatch of the ring size and will cause
1242                  * the context to hang. Presumably with a bit of reordering we
1243                  * could move the perma-pin step to the backend function
1244                  * intel_engine_create_parallel.
1245                  */
1246                 if (pe[n].type == I915_GEM_ENGINE_TYPE_PARALLEL) {
1247                         ret = perma_pin_contexts(ce);
1248                         if (ret) {
1249                                 err = ERR_PTR(ret);
1250                                 goto free_engines;
1251                         }
1252                 }
1253         }
1254
1255         return e;
1256
1257 free_engines:
1258         free_engines(e);
1259         return err;
1260 }
1261
1262 static void i915_gem_context_release_work(struct work_struct *work)
1263 {
1264         struct i915_gem_context *ctx = container_of(work, typeof(*ctx),
1265                                                     release_work);
1266         struct i915_address_space *vm;
1267
1268         trace_i915_context_free(ctx);
1269         GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1270
1271         if (ctx->syncobj)
1272                 drm_syncobj_put(ctx->syncobj);
1273
1274         vm = ctx->vm;
1275         if (vm)
1276                 i915_vm_put(vm);
1277
1278         if (ctx->pxp_wakeref)
1279                 intel_runtime_pm_put(&ctx->i915->runtime_pm, ctx->pxp_wakeref);
1280
1281         if (ctx->client)
1282                 i915_drm_client_put(ctx->client);
1283
1284         mutex_destroy(&ctx->engines_mutex);
1285         mutex_destroy(&ctx->lut_mutex);
1286
1287         put_pid(ctx->pid);
1288         mutex_destroy(&ctx->mutex);
1289
1290         kfree_rcu(ctx, rcu);
1291 }
1292
1293 void i915_gem_context_release(struct kref *ref)
1294 {
1295         struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
1296
1297         queue_work(ctx->i915->wq, &ctx->release_work);
1298 }
1299
1300 static inline struct i915_gem_engines *
1301 __context_engines_static(const struct i915_gem_context *ctx)
1302 {
1303         return rcu_dereference_protected(ctx->engines, true);
1304 }
1305
1306 static void __reset_context(struct i915_gem_context *ctx,
1307                             struct intel_engine_cs *engine)
1308 {
1309         intel_gt_handle_error(engine->gt, engine->mask, 0,
1310                               "context closure in %s", ctx->name);
1311 }
1312
1313 static bool __cancel_engine(struct intel_engine_cs *engine)
1314 {
1315         /*
1316          * Send a "high priority pulse" down the engine to cause the
1317          * current request to be momentarily preempted. (If it fails to
1318          * be preempted, it will be reset). As we have marked our context
1319          * as banned, any incomplete request, including any running, will
1320          * be skipped following the preemption.
1321          *
1322          * If there is no hangchecking (one of the reasons why we try to
1323          * cancel the context) and no forced preemption, there may be no
1324          * means by which we reset the GPU and evict the persistent hog.
1325          * Ergo if we are unable to inject a preemptive pulse that can
1326          * kill the banned context, we fallback to doing a local reset
1327          * instead.
1328          */
1329         return intel_engine_pulse(engine) == 0;
1330 }
1331
1332 static struct intel_engine_cs *active_engine(struct intel_context *ce)
1333 {
1334         struct intel_engine_cs *engine = NULL;
1335         struct i915_request *rq;
1336
1337         if (intel_context_has_inflight(ce))
1338                 return intel_context_inflight(ce);
1339
1340         if (!ce->timeline)
1341                 return NULL;
1342
1343         /*
1344          * rq->link is only SLAB_TYPESAFE_BY_RCU, we need to hold a reference
1345          * to the request to prevent it being transferred to a new timeline
1346          * (and onto a new timeline->requests list).
1347          */
1348         rcu_read_lock();
1349         list_for_each_entry_reverse(rq, &ce->timeline->requests, link) {
1350                 bool found;
1351
1352                 /* timeline is already completed upto this point? */
1353                 if (!i915_request_get_rcu(rq))
1354                         break;
1355
1356                 /* Check with the backend if the request is inflight */
1357                 found = true;
1358                 if (likely(rcu_access_pointer(rq->timeline) == ce->timeline))
1359                         found = i915_request_active_engine(rq, &engine);
1360
1361                 i915_request_put(rq);
1362                 if (found)
1363                         break;
1364         }
1365         rcu_read_unlock();
1366
1367         return engine;
1368 }
1369
1370 static void kill_engines(struct i915_gem_engines *engines, bool ban)
1371 {
1372         struct i915_gem_engines_iter it;
1373         struct intel_context *ce;
1374
1375         /*
1376          * Map the user's engine back to the actual engines; one virtual
1377          * engine will be mapped to multiple engines, and using ctx->engine[]
1378          * the same engine may be have multiple instances in the user's map.
1379          * However, we only care about pending requests, so only include
1380          * engines on which there are incomplete requests.
1381          */
1382         for_each_gem_engine(ce, engines, it) {
1383                 struct intel_engine_cs *engine;
1384
1385                 if (ban && intel_context_ban(ce, NULL))
1386                         continue;
1387
1388                 /*
1389                  * Check the current active state of this context; if we
1390                  * are currently executing on the GPU we need to evict
1391                  * ourselves. On the other hand, if we haven't yet been
1392                  * submitted to the GPU or if everything is complete,
1393                  * we have nothing to do.
1394                  */
1395                 engine = active_engine(ce);
1396
1397                 /* First attempt to gracefully cancel the context */
1398                 if (engine && !__cancel_engine(engine) && ban)
1399                         /*
1400                          * If we are unable to send a preemptive pulse to bump
1401                          * the context from the GPU, we have to resort to a full
1402                          * reset. We hope the collateral damage is worth it.
1403                          */
1404                         __reset_context(engines->ctx, engine);
1405         }
1406 }
1407
1408 static void kill_context(struct i915_gem_context *ctx)
1409 {
1410         bool ban = (!i915_gem_context_is_persistent(ctx) ||
1411                     !ctx->i915->params.enable_hangcheck);
1412         struct i915_gem_engines *pos, *next;
1413
1414         spin_lock_irq(&ctx->stale.lock);
1415         GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
1416         list_for_each_entry_safe(pos, next, &ctx->stale.engines, link) {
1417                 if (!i915_sw_fence_await(&pos->fence)) {
1418                         list_del_init(&pos->link);
1419                         continue;
1420                 }
1421
1422                 spin_unlock_irq(&ctx->stale.lock);
1423
1424                 kill_engines(pos, ban);
1425
1426                 spin_lock_irq(&ctx->stale.lock);
1427                 GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence));
1428                 list_safe_reset_next(pos, next, link);
1429                 list_del_init(&pos->link); /* decouple from FENCE_COMPLETE */
1430
1431                 i915_sw_fence_complete(&pos->fence);
1432         }
1433         spin_unlock_irq(&ctx->stale.lock);
1434 }
1435
1436 static void engines_idle_release(struct i915_gem_context *ctx,
1437                                  struct i915_gem_engines *engines)
1438 {
1439         struct i915_gem_engines_iter it;
1440         struct intel_context *ce;
1441
1442         INIT_LIST_HEAD(&engines->link);
1443
1444         engines->ctx = i915_gem_context_get(ctx);
1445
1446         for_each_gem_engine(ce, engines, it) {
1447                 int err;
1448
1449                 /* serialises with execbuf */
1450                 set_bit(CONTEXT_CLOSED_BIT, &ce->flags);
1451                 if (!intel_context_pin_if_active(ce))
1452                         continue;
1453
1454                 /* Wait until context is finally scheduled out and retired */
1455                 err = i915_sw_fence_await_active(&engines->fence,
1456                                                  &ce->active,
1457                                                  I915_ACTIVE_AWAIT_BARRIER);
1458                 intel_context_unpin(ce);
1459                 if (err)
1460                         goto kill;
1461         }
1462
1463         spin_lock_irq(&ctx->stale.lock);
1464         if (!i915_gem_context_is_closed(ctx))
1465                 list_add_tail(&engines->link, &ctx->stale.engines);
1466         spin_unlock_irq(&ctx->stale.lock);
1467
1468 kill:
1469         if (list_empty(&engines->link)) /* raced, already closed */
1470                 kill_engines(engines, true);
1471
1472         i915_sw_fence_commit(&engines->fence);
1473 }
1474
1475 static void set_closed_name(struct i915_gem_context *ctx)
1476 {
1477         char *s;
1478
1479         /* Replace '[]' with '<>' to indicate closed in debug prints */
1480
1481         s = strrchr(ctx->name, '[');
1482         if (!s)
1483                 return;
1484
1485         *s = '<';
1486
1487         s = strchr(s + 1, ']');
1488         if (s)
1489                 *s = '>';
1490 }
1491
1492 static void context_close(struct i915_gem_context *ctx)
1493 {
1494         struct i915_drm_client *client;
1495
1496         /* Flush any concurrent set_engines() */
1497         mutex_lock(&ctx->engines_mutex);
1498         unpin_engines(__context_engines_static(ctx));
1499         engines_idle_release(ctx, rcu_replace_pointer(ctx->engines, NULL, 1));
1500         i915_gem_context_set_closed(ctx);
1501         mutex_unlock(&ctx->engines_mutex);
1502
1503         mutex_lock(&ctx->mutex);
1504
1505         set_closed_name(ctx);
1506
1507         /*
1508          * The LUT uses the VMA as a backpointer to unref the object,
1509          * so we need to clear the LUT before we close all the VMA (inside
1510          * the ppgtt).
1511          */
1512         lut_close(ctx);
1513
1514         ctx->file_priv = ERR_PTR(-EBADF);
1515
1516         spin_lock(&ctx->i915->gem.contexts.lock);
1517         list_del(&ctx->link);
1518         spin_unlock(&ctx->i915->gem.contexts.lock);
1519
1520         client = ctx->client;
1521         if (client) {
1522                 spin_lock(&client->ctx_lock);
1523                 list_del_rcu(&ctx->client_link);
1524                 spin_unlock(&client->ctx_lock);
1525         }
1526
1527         mutex_unlock(&ctx->mutex);
1528
1529         /*
1530          * If the user has disabled hangchecking, we can not be sure that
1531          * the batches will ever complete after the context is closed,
1532          * keeping the context and all resources pinned forever. So in this
1533          * case we opt to forcibly kill off all remaining requests on
1534          * context close.
1535          */
1536         kill_context(ctx);
1537
1538         i915_gem_context_put(ctx);
1539 }
1540
1541 static int __context_set_persistence(struct i915_gem_context *ctx, bool state)
1542 {
1543         if (i915_gem_context_is_persistent(ctx) == state)
1544                 return 0;
1545
1546         if (state) {
1547                 /*
1548                  * Only contexts that are short-lived [that will expire or be
1549                  * reset] are allowed to survive past termination. We require
1550                  * hangcheck to ensure that the persistent requests are healthy.
1551                  */
1552                 if (!ctx->i915->params.enable_hangcheck)
1553                         return -EINVAL;
1554
1555                 i915_gem_context_set_persistence(ctx);
1556         } else {
1557                 /* To cancel a context we use "preempt-to-idle" */
1558                 if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
1559                         return -ENODEV;
1560
1561                 /*
1562                  * If the cancel fails, we then need to reset, cleanly!
1563                  *
1564                  * If the per-engine reset fails, all hope is lost! We resort
1565                  * to a full GPU reset in that unlikely case, but realistically
1566                  * if the engine could not reset, the full reset does not fare
1567                  * much better. The damage has been done.
1568                  *
1569                  * However, if we cannot reset an engine by itself, we cannot
1570                  * cleanup a hanging persistent context without causing
1571                  * colateral damage, and we should not pretend we can by
1572                  * exposing the interface.
1573                  */
1574                 if (!intel_has_reset_engine(to_gt(ctx->i915)))
1575                         return -ENODEV;
1576
1577                 i915_gem_context_clear_persistence(ctx);
1578         }
1579
1580         return 0;
1581 }
1582
1583 static struct i915_gem_context *
1584 i915_gem_create_context(struct drm_i915_private *i915,
1585                         const struct i915_gem_proto_context *pc)
1586 {
1587         struct i915_gem_context *ctx;
1588         struct i915_address_space *vm = NULL;
1589         struct i915_gem_engines *e;
1590         int err;
1591         int i;
1592
1593         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1594         if (!ctx)
1595                 return ERR_PTR(-ENOMEM);
1596
1597         kref_init(&ctx->ref);
1598         ctx->i915 = i915;
1599         ctx->sched = pc->sched;
1600         mutex_init(&ctx->mutex);
1601         INIT_LIST_HEAD(&ctx->link);
1602         INIT_WORK(&ctx->release_work, i915_gem_context_release_work);
1603
1604         spin_lock_init(&ctx->stale.lock);
1605         INIT_LIST_HEAD(&ctx->stale.engines);
1606
1607         if (pc->vm) {
1608                 vm = i915_vm_get(pc->vm);
1609         } else if (HAS_FULL_PPGTT(i915)) {
1610                 struct i915_ppgtt *ppgtt;
1611
1612                 ppgtt = i915_ppgtt_create(to_gt(i915), 0);
1613                 if (IS_ERR(ppgtt)) {
1614                         drm_dbg(&i915->drm, "PPGTT setup failed (%ld)\n",
1615                                 PTR_ERR(ppgtt));
1616                         err = PTR_ERR(ppgtt);
1617                         goto err_ctx;
1618                 }
1619                 vm = &ppgtt->vm;
1620         }
1621         if (vm)
1622                 ctx->vm = vm;
1623
1624         mutex_init(&ctx->engines_mutex);
1625         if (pc->num_user_engines >= 0) {
1626                 i915_gem_context_set_user_engines(ctx);
1627                 e = user_engines(ctx, pc->num_user_engines, pc->user_engines);
1628         } else {
1629                 i915_gem_context_clear_user_engines(ctx);
1630                 e = default_engines(ctx, pc->legacy_rcs_sseu);
1631         }
1632         if (IS_ERR(e)) {
1633                 err = PTR_ERR(e);
1634                 goto err_vm;
1635         }
1636         RCU_INIT_POINTER(ctx->engines, e);
1637
1638         INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
1639         mutex_init(&ctx->lut_mutex);
1640
1641         /* NB: Mark all slices as needing a remap so that when the context first
1642          * loads it will restore whatever remap state already exists. If there
1643          * is no remap info, it will be a NOP. */
1644         ctx->remap_slice = ALL_L3_SLICES(i915);
1645
1646         ctx->user_flags = pc->user_flags;
1647
1648         for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
1649                 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
1650
1651         if (pc->single_timeline) {
1652                 err = drm_syncobj_create(&ctx->syncobj,
1653                                          DRM_SYNCOBJ_CREATE_SIGNALED,
1654                                          NULL);
1655                 if (err)
1656                         goto err_engines;
1657         }
1658
1659         if (pc->uses_protected_content) {
1660                 ctx->pxp_wakeref = intel_runtime_pm_get(&i915->runtime_pm);
1661                 ctx->uses_protected_content = true;
1662         }
1663
1664         trace_i915_context_create(ctx);
1665
1666         return ctx;
1667
1668 err_engines:
1669         free_engines(e);
1670 err_vm:
1671         if (ctx->vm)
1672                 i915_vm_put(ctx->vm);
1673 err_ctx:
1674         kfree(ctx);
1675         return ERR_PTR(err);
1676 }
1677
1678 static void init_contexts(struct i915_gem_contexts *gc)
1679 {
1680         spin_lock_init(&gc->lock);
1681         INIT_LIST_HEAD(&gc->list);
1682 }
1683
1684 void i915_gem_init__contexts(struct drm_i915_private *i915)
1685 {
1686         init_contexts(&i915->gem.contexts);
1687 }
1688
1689 static void gem_context_register(struct i915_gem_context *ctx,
1690                                  struct drm_i915_file_private *fpriv,
1691                                  u32 id)
1692 {
1693         struct drm_i915_private *i915 = ctx->i915;
1694         void *old;
1695
1696         ctx->file_priv = fpriv;
1697
1698         ctx->pid = get_task_pid(current, PIDTYPE_PID);
1699         ctx->client = i915_drm_client_get(fpriv->client);
1700
1701         snprintf(ctx->name, sizeof(ctx->name), "%s[%d]",
1702                  current->comm, pid_nr(ctx->pid));
1703
1704         /* And finally expose ourselves to userspace via the idr */
1705         old = xa_store(&fpriv->context_xa, id, ctx, GFP_KERNEL);
1706         WARN_ON(old);
1707
1708         spin_lock(&ctx->client->ctx_lock);
1709         list_add_tail_rcu(&ctx->client_link, &ctx->client->ctx_list);
1710         spin_unlock(&ctx->client->ctx_lock);
1711
1712         spin_lock(&i915->gem.contexts.lock);
1713         list_add_tail(&ctx->link, &i915->gem.contexts.list);
1714         spin_unlock(&i915->gem.contexts.lock);
1715 }
1716
1717 int i915_gem_context_open(struct drm_i915_private *i915,
1718                           struct drm_file *file)
1719 {
1720         struct drm_i915_file_private *file_priv = file->driver_priv;
1721         struct i915_gem_proto_context *pc;
1722         struct i915_gem_context *ctx;
1723         int err;
1724
1725         mutex_init(&file_priv->proto_context_lock);
1726         xa_init_flags(&file_priv->proto_context_xa, XA_FLAGS_ALLOC);
1727
1728         /* 0 reserved for the default context */
1729         xa_init_flags(&file_priv->context_xa, XA_FLAGS_ALLOC1);
1730
1731         /* 0 reserved for invalid/unassigned ppgtt */
1732         xa_init_flags(&file_priv->vm_xa, XA_FLAGS_ALLOC1);
1733
1734         pc = proto_context_create(i915, 0);
1735         if (IS_ERR(pc)) {
1736                 err = PTR_ERR(pc);
1737                 goto err;
1738         }
1739
1740         ctx = i915_gem_create_context(i915, pc);
1741         proto_context_close(i915, pc);
1742         if (IS_ERR(ctx)) {
1743                 err = PTR_ERR(ctx);
1744                 goto err;
1745         }
1746
1747         gem_context_register(ctx, file_priv, 0);
1748
1749         return 0;
1750
1751 err:
1752         xa_destroy(&file_priv->vm_xa);
1753         xa_destroy(&file_priv->context_xa);
1754         xa_destroy(&file_priv->proto_context_xa);
1755         mutex_destroy(&file_priv->proto_context_lock);
1756         return err;
1757 }
1758
1759 void i915_gem_context_close(struct drm_file *file)
1760 {
1761         struct drm_i915_file_private *file_priv = file->driver_priv;
1762         struct i915_gem_proto_context *pc;
1763         struct i915_address_space *vm;
1764         struct i915_gem_context *ctx;
1765         unsigned long idx;
1766
1767         xa_for_each(&file_priv->proto_context_xa, idx, pc)
1768                 proto_context_close(file_priv->dev_priv, pc);
1769         xa_destroy(&file_priv->proto_context_xa);
1770         mutex_destroy(&file_priv->proto_context_lock);
1771
1772         xa_for_each(&file_priv->context_xa, idx, ctx)
1773                 context_close(ctx);
1774         xa_destroy(&file_priv->context_xa);
1775
1776         xa_for_each(&file_priv->vm_xa, idx, vm)
1777                 i915_vm_put(vm);
1778         xa_destroy(&file_priv->vm_xa);
1779 }
1780
1781 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
1782                              struct drm_file *file)
1783 {
1784         struct drm_i915_private *i915 = to_i915(dev);
1785         struct drm_i915_gem_vm_control *args = data;
1786         struct drm_i915_file_private *file_priv = file->driver_priv;
1787         struct i915_ppgtt *ppgtt;
1788         u32 id;
1789         int err;
1790
1791         if (!HAS_FULL_PPGTT(i915))
1792                 return -ENODEV;
1793
1794         if (args->flags)
1795                 return -EINVAL;
1796
1797         ppgtt = i915_ppgtt_create(to_gt(i915), 0);
1798         if (IS_ERR(ppgtt))
1799                 return PTR_ERR(ppgtt);
1800
1801         if (args->extensions) {
1802                 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
1803                                            NULL, 0,
1804                                            ppgtt);
1805                 if (err)
1806                         goto err_put;
1807         }
1808
1809         err = xa_alloc(&file_priv->vm_xa, &id, &ppgtt->vm,
1810                        xa_limit_32b, GFP_KERNEL);
1811         if (err)
1812                 goto err_put;
1813
1814         GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1815         args->vm_id = id;
1816         return 0;
1817
1818 err_put:
1819         i915_vm_put(&ppgtt->vm);
1820         return err;
1821 }
1822
1823 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
1824                               struct drm_file *file)
1825 {
1826         struct drm_i915_file_private *file_priv = file->driver_priv;
1827         struct drm_i915_gem_vm_control *args = data;
1828         struct i915_address_space *vm;
1829
1830         if (args->flags)
1831                 return -EINVAL;
1832
1833         if (args->extensions)
1834                 return -EINVAL;
1835
1836         vm = xa_erase(&file_priv->vm_xa, args->vm_id);
1837         if (!vm)
1838                 return -ENOENT;
1839
1840         i915_vm_put(vm);
1841         return 0;
1842 }
1843
1844 static int get_ppgtt(struct drm_i915_file_private *file_priv,
1845                      struct i915_gem_context *ctx,
1846                      struct drm_i915_gem_context_param *args)
1847 {
1848         struct i915_address_space *vm;
1849         int err;
1850         u32 id;
1851
1852         if (!i915_gem_context_has_full_ppgtt(ctx))
1853                 return -ENODEV;
1854
1855         vm = ctx->vm;
1856         GEM_BUG_ON(!vm);
1857
1858         err = xa_alloc(&file_priv->vm_xa, &id, vm, xa_limit_32b, GFP_KERNEL);
1859         if (err)
1860                 return err;
1861
1862         i915_vm_get(vm);
1863
1864         GEM_BUG_ON(id == 0); /* reserved for invalid/unassigned ppgtt */
1865         args->value = id;
1866         args->size = 0;
1867
1868         return err;
1869 }
1870
1871 int
1872 i915_gem_user_to_context_sseu(struct intel_gt *gt,
1873                               const struct drm_i915_gem_context_param_sseu *user,
1874                               struct intel_sseu *context)
1875 {
1876         const struct sseu_dev_info *device = &gt->info.sseu;
1877         struct drm_i915_private *i915 = gt->i915;
1878
1879         /* No zeros in any field. */
1880         if (!user->slice_mask || !user->subslice_mask ||
1881             !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1882                 return -EINVAL;
1883
1884         /* Max > min. */
1885         if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1886                 return -EINVAL;
1887
1888         /*
1889          * Some future proofing on the types since the uAPI is wider than the
1890          * current internal implementation.
1891          */
1892         if (overflows_type(user->slice_mask, context->slice_mask) ||
1893             overflows_type(user->subslice_mask, context->subslice_mask) ||
1894             overflows_type(user->min_eus_per_subslice,
1895                            context->min_eus_per_subslice) ||
1896             overflows_type(user->max_eus_per_subslice,
1897                            context->max_eus_per_subslice))
1898                 return -EINVAL;
1899
1900         /* Check validity against hardware. */
1901         if (user->slice_mask & ~device->slice_mask)
1902                 return -EINVAL;
1903
1904         if (user->subslice_mask & ~device->subslice_mask[0])
1905                 return -EINVAL;
1906
1907         if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1908                 return -EINVAL;
1909
1910         context->slice_mask = user->slice_mask;
1911         context->subslice_mask = user->subslice_mask;
1912         context->min_eus_per_subslice = user->min_eus_per_subslice;
1913         context->max_eus_per_subslice = user->max_eus_per_subslice;
1914
1915         /* Part specific restrictions. */
1916         if (GRAPHICS_VER(i915) == 11) {
1917                 unsigned int hw_s = hweight8(device->slice_mask);
1918                 unsigned int hw_ss_per_s = hweight8(device->subslice_mask[0]);
1919                 unsigned int req_s = hweight8(context->slice_mask);
1920                 unsigned int req_ss = hweight8(context->subslice_mask);
1921
1922                 /*
1923                  * Only full subslice enablement is possible if more than one
1924                  * slice is turned on.
1925                  */
1926                 if (req_s > 1 && req_ss != hw_ss_per_s)
1927                         return -EINVAL;
1928
1929                 /*
1930                  * If more than four (SScount bitfield limit) subslices are
1931                  * requested then the number has to be even.
1932                  */
1933                 if (req_ss > 4 && (req_ss & 1))
1934                         return -EINVAL;
1935
1936                 /*
1937                  * If only one slice is enabled and subslice count is below the
1938                  * device full enablement, it must be at most half of the all
1939                  * available subslices.
1940                  */
1941                 if (req_s == 1 && req_ss < hw_ss_per_s &&
1942                     req_ss > (hw_ss_per_s / 2))
1943                         return -EINVAL;
1944
1945                 /* ABI restriction - VME use case only. */
1946
1947                 /* All slices or one slice only. */
1948                 if (req_s != 1 && req_s != hw_s)
1949                         return -EINVAL;
1950
1951                 /*
1952                  * Half subslices or full enablement only when one slice is
1953                  * enabled.
1954                  */
1955                 if (req_s == 1 &&
1956                     (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1957                         return -EINVAL;
1958
1959                 /* No EU configuration changes. */
1960                 if ((user->min_eus_per_subslice !=
1961                      device->max_eus_per_subslice) ||
1962                     (user->max_eus_per_subslice !=
1963                      device->max_eus_per_subslice))
1964                         return -EINVAL;
1965         }
1966
1967         return 0;
1968 }
1969
1970 static int set_sseu(struct i915_gem_context *ctx,
1971                     struct drm_i915_gem_context_param *args)
1972 {
1973         struct drm_i915_private *i915 = ctx->i915;
1974         struct drm_i915_gem_context_param_sseu user_sseu;
1975         struct intel_context *ce;
1976         struct intel_sseu sseu;
1977         unsigned long lookup;
1978         int ret;
1979
1980         if (args->size < sizeof(user_sseu))
1981                 return -EINVAL;
1982
1983         if (GRAPHICS_VER(i915) != 11)
1984                 return -ENODEV;
1985
1986         if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1987                            sizeof(user_sseu)))
1988                 return -EFAULT;
1989
1990         if (user_sseu.rsvd)
1991                 return -EINVAL;
1992
1993         if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
1994                 return -EINVAL;
1995
1996         lookup = 0;
1997         if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
1998                 lookup |= LOOKUP_USER_INDEX;
1999
2000         ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2001         if (IS_ERR(ce))
2002                 return PTR_ERR(ce);
2003
2004         /* Only render engine supports RPCS configuration. */
2005         if (ce->engine->class != RENDER_CLASS) {
2006                 ret = -ENODEV;
2007                 goto out_ce;
2008         }
2009
2010         ret = i915_gem_user_to_context_sseu(ce->engine->gt, &user_sseu, &sseu);
2011         if (ret)
2012                 goto out_ce;
2013
2014         ret = intel_context_reconfigure_sseu(ce, sseu);
2015         if (ret)
2016                 goto out_ce;
2017
2018         args->size = sizeof(user_sseu);
2019
2020 out_ce:
2021         intel_context_put(ce);
2022         return ret;
2023 }
2024
2025 static int
2026 set_persistence(struct i915_gem_context *ctx,
2027                 const struct drm_i915_gem_context_param *args)
2028 {
2029         if (args->size)
2030                 return -EINVAL;
2031
2032         return __context_set_persistence(ctx, args->value);
2033 }
2034
2035 static int set_priority(struct i915_gem_context *ctx,
2036                         const struct drm_i915_gem_context_param *args)
2037 {
2038         struct i915_gem_engines_iter it;
2039         struct intel_context *ce;
2040         int err;
2041
2042         err = validate_priority(ctx->i915, args);
2043         if (err)
2044                 return err;
2045
2046         ctx->sched.priority = args->value;
2047
2048         for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
2049                 if (!intel_engine_has_timeslices(ce->engine))
2050                         continue;
2051
2052                 if (ctx->sched.priority >= I915_PRIORITY_NORMAL &&
2053                     intel_engine_has_semaphores(ce->engine))
2054                         intel_context_set_use_semaphores(ce);
2055                 else
2056                         intel_context_clear_use_semaphores(ce);
2057         }
2058         i915_gem_context_unlock_engines(ctx);
2059
2060         return 0;
2061 }
2062
2063 static int get_protected(struct i915_gem_context *ctx,
2064                          struct drm_i915_gem_context_param *args)
2065 {
2066         args->size = 0;
2067         args->value = i915_gem_context_uses_protected_content(ctx);
2068
2069         return 0;
2070 }
2071
2072 static int ctx_setparam(struct drm_i915_file_private *fpriv,
2073                         struct i915_gem_context *ctx,
2074                         struct drm_i915_gem_context_param *args)
2075 {
2076         int ret = 0;
2077
2078         switch (args->param) {
2079         case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2080                 if (args->size)
2081                         ret = -EINVAL;
2082                 else if (args->value)
2083                         i915_gem_context_set_no_error_capture(ctx);
2084                 else
2085                         i915_gem_context_clear_no_error_capture(ctx);
2086                 break;
2087
2088         case I915_CONTEXT_PARAM_BANNABLE:
2089                 if (args->size)
2090                         ret = -EINVAL;
2091                 else if (!capable(CAP_SYS_ADMIN) && !args->value)
2092                         ret = -EPERM;
2093                 else if (args->value)
2094                         i915_gem_context_set_bannable(ctx);
2095                 else if (i915_gem_context_uses_protected_content(ctx))
2096                         ret = -EPERM; /* can't clear this for protected contexts */
2097                 else
2098                         i915_gem_context_clear_bannable(ctx);
2099                 break;
2100
2101         case I915_CONTEXT_PARAM_RECOVERABLE:
2102                 if (args->size)
2103                         ret = -EINVAL;
2104                 else if (!args->value)
2105                         i915_gem_context_clear_recoverable(ctx);
2106                 else if (i915_gem_context_uses_protected_content(ctx))
2107                         ret = -EPERM; /* can't set this for protected contexts */
2108                 else
2109                         i915_gem_context_set_recoverable(ctx);
2110                 break;
2111
2112         case I915_CONTEXT_PARAM_PRIORITY:
2113                 ret = set_priority(ctx, args);
2114                 break;
2115
2116         case I915_CONTEXT_PARAM_SSEU:
2117                 ret = set_sseu(ctx, args);
2118                 break;
2119
2120         case I915_CONTEXT_PARAM_PERSISTENCE:
2121                 ret = set_persistence(ctx, args);
2122                 break;
2123
2124         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2125         case I915_CONTEXT_PARAM_NO_ZEROMAP:
2126         case I915_CONTEXT_PARAM_BAN_PERIOD:
2127         case I915_CONTEXT_PARAM_RINGSIZE:
2128         case I915_CONTEXT_PARAM_VM:
2129         case I915_CONTEXT_PARAM_ENGINES:
2130         default:
2131                 ret = -EINVAL;
2132                 break;
2133         }
2134
2135         return ret;
2136 }
2137
2138 struct create_ext {
2139         struct i915_gem_proto_context *pc;
2140         struct drm_i915_file_private *fpriv;
2141 };
2142
2143 static int create_setparam(struct i915_user_extension __user *ext, void *data)
2144 {
2145         struct drm_i915_gem_context_create_ext_setparam local;
2146         const struct create_ext *arg = data;
2147
2148         if (copy_from_user(&local, ext, sizeof(local)))
2149                 return -EFAULT;
2150
2151         if (local.param.ctx_id)
2152                 return -EINVAL;
2153
2154         return set_proto_ctx_param(arg->fpriv, arg->pc, &local.param);
2155 }
2156
2157 static int invalid_ext(struct i915_user_extension __user *ext, void *data)
2158 {
2159         return -EINVAL;
2160 }
2161
2162 static const i915_user_extension_fn create_extensions[] = {
2163         [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2164         [I915_CONTEXT_CREATE_EXT_CLONE] = invalid_ext,
2165 };
2166
2167 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2168 {
2169         return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2170 }
2171
2172 static inline struct i915_gem_context *
2173 __context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2174 {
2175         struct i915_gem_context *ctx;
2176
2177         rcu_read_lock();
2178         ctx = xa_load(&file_priv->context_xa, id);
2179         if (ctx && !kref_get_unless_zero(&ctx->ref))
2180                 ctx = NULL;
2181         rcu_read_unlock();
2182
2183         return ctx;
2184 }
2185
2186 static struct i915_gem_context *
2187 finalize_create_context_locked(struct drm_i915_file_private *file_priv,
2188                                struct i915_gem_proto_context *pc, u32 id)
2189 {
2190         struct i915_gem_context *ctx;
2191         void *old;
2192
2193         lockdep_assert_held(&file_priv->proto_context_lock);
2194
2195         ctx = i915_gem_create_context(file_priv->dev_priv, pc);
2196         if (IS_ERR(ctx))
2197                 return ctx;
2198
2199         gem_context_register(ctx, file_priv, id);
2200
2201         old = xa_erase(&file_priv->proto_context_xa, id);
2202         GEM_BUG_ON(old != pc);
2203         proto_context_close(file_priv->dev_priv, pc);
2204
2205         /* One for the xarray and one for the caller */
2206         return i915_gem_context_get(ctx);
2207 }
2208
2209 struct i915_gem_context *
2210 i915_gem_context_lookup(struct drm_i915_file_private *file_priv, u32 id)
2211 {
2212         struct i915_gem_proto_context *pc;
2213         struct i915_gem_context *ctx;
2214
2215         ctx = __context_lookup(file_priv, id);
2216         if (ctx)
2217                 return ctx;
2218
2219         mutex_lock(&file_priv->proto_context_lock);
2220         /* Try one more time under the lock */
2221         ctx = __context_lookup(file_priv, id);
2222         if (!ctx) {
2223                 pc = xa_load(&file_priv->proto_context_xa, id);
2224                 if (!pc)
2225                         ctx = ERR_PTR(-ENOENT);
2226                 else
2227                         ctx = finalize_create_context_locked(file_priv, pc, id);
2228         }
2229         mutex_unlock(&file_priv->proto_context_lock);
2230
2231         return ctx;
2232 }
2233
2234 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2235                                   struct drm_file *file)
2236 {
2237         struct drm_i915_private *i915 = to_i915(dev);
2238         struct drm_i915_gem_context_create_ext *args = data;
2239         struct create_ext ext_data;
2240         int ret;
2241         u32 id;
2242
2243         if (!DRIVER_CAPS(i915)->has_logical_contexts)
2244                 return -ENODEV;
2245
2246         if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2247                 return -EINVAL;
2248
2249         ret = intel_gt_terminally_wedged(to_gt(i915));
2250         if (ret)
2251                 return ret;
2252
2253         ext_data.fpriv = file->driver_priv;
2254         if (client_is_banned(ext_data.fpriv)) {
2255                 drm_dbg(&i915->drm,
2256                         "client %s[%d] banned from creating ctx\n",
2257                         current->comm, task_pid_nr(current));
2258                 return -EIO;
2259         }
2260
2261         ext_data.pc = proto_context_create(i915, args->flags);
2262         if (IS_ERR(ext_data.pc))
2263                 return PTR_ERR(ext_data.pc);
2264
2265         if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2266                 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2267                                            create_extensions,
2268                                            ARRAY_SIZE(create_extensions),
2269                                            &ext_data);
2270                 if (ret)
2271                         goto err_pc;
2272         }
2273
2274         if (GRAPHICS_VER(i915) > 12) {
2275                 struct i915_gem_context *ctx;
2276
2277                 /* Get ourselves a context ID */
2278                 ret = xa_alloc(&ext_data.fpriv->context_xa, &id, NULL,
2279                                xa_limit_32b, GFP_KERNEL);
2280                 if (ret)
2281                         goto err_pc;
2282
2283                 ctx = i915_gem_create_context(i915, ext_data.pc);
2284                 if (IS_ERR(ctx)) {
2285                         ret = PTR_ERR(ctx);
2286                         goto err_pc;
2287                 }
2288
2289                 proto_context_close(i915, ext_data.pc);
2290                 gem_context_register(ctx, ext_data.fpriv, id);
2291         } else {
2292                 ret = proto_context_register(ext_data.fpriv, ext_data.pc, &id);
2293                 if (ret < 0)
2294                         goto err_pc;
2295         }
2296
2297         args->ctx_id = id;
2298         drm_dbg(&i915->drm, "HW context %d created\n", args->ctx_id);
2299
2300         return 0;
2301
2302 err_pc:
2303         proto_context_close(i915, ext_data.pc);
2304         return ret;
2305 }
2306
2307 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2308                                    struct drm_file *file)
2309 {
2310         struct drm_i915_gem_context_destroy *args = data;
2311         struct drm_i915_file_private *file_priv = file->driver_priv;
2312         struct i915_gem_proto_context *pc;
2313         struct i915_gem_context *ctx;
2314
2315         if (args->pad != 0)
2316                 return -EINVAL;
2317
2318         if (!args->ctx_id)
2319                 return -ENOENT;
2320
2321         /* We need to hold the proto-context lock here to prevent races
2322          * with finalize_create_context_locked().
2323          */
2324         mutex_lock(&file_priv->proto_context_lock);
2325         ctx = xa_erase(&file_priv->context_xa, args->ctx_id);
2326         pc = xa_erase(&file_priv->proto_context_xa, args->ctx_id);
2327         mutex_unlock(&file_priv->proto_context_lock);
2328
2329         if (!ctx && !pc)
2330                 return -ENOENT;
2331         GEM_WARN_ON(ctx && pc);
2332
2333         if (pc)
2334                 proto_context_close(file_priv->dev_priv, pc);
2335
2336         if (ctx)
2337                 context_close(ctx);
2338
2339         return 0;
2340 }
2341
2342 static int get_sseu(struct i915_gem_context *ctx,
2343                     struct drm_i915_gem_context_param *args)
2344 {
2345         struct drm_i915_gem_context_param_sseu user_sseu;
2346         struct intel_context *ce;
2347         unsigned long lookup;
2348         int err;
2349
2350         if (args->size == 0)
2351                 goto out;
2352         else if (args->size < sizeof(user_sseu))
2353                 return -EINVAL;
2354
2355         if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2356                            sizeof(user_sseu)))
2357                 return -EFAULT;
2358
2359         if (user_sseu.rsvd)
2360                 return -EINVAL;
2361
2362         if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2363                 return -EINVAL;
2364
2365         lookup = 0;
2366         if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2367                 lookup |= LOOKUP_USER_INDEX;
2368
2369         ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2370         if (IS_ERR(ce))
2371                 return PTR_ERR(ce);
2372
2373         err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2374         if (err) {
2375                 intel_context_put(ce);
2376                 return err;
2377         }
2378
2379         user_sseu.slice_mask = ce->sseu.slice_mask;
2380         user_sseu.subslice_mask = ce->sseu.subslice_mask;
2381         user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2382         user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2383
2384         intel_context_unlock_pinned(ce);
2385         intel_context_put(ce);
2386
2387         if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2388                          sizeof(user_sseu)))
2389                 return -EFAULT;
2390
2391 out:
2392         args->size = sizeof(user_sseu);
2393
2394         return 0;
2395 }
2396
2397 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2398                                     struct drm_file *file)
2399 {
2400         struct drm_i915_file_private *file_priv = file->driver_priv;
2401         struct drm_i915_gem_context_param *args = data;
2402         struct i915_gem_context *ctx;
2403         struct i915_address_space *vm;
2404         int ret = 0;
2405
2406         ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2407         if (IS_ERR(ctx))
2408                 return PTR_ERR(ctx);
2409
2410         switch (args->param) {
2411         case I915_CONTEXT_PARAM_GTT_SIZE:
2412                 args->size = 0;
2413                 vm = i915_gem_context_get_eb_vm(ctx);
2414                 args->value = vm->total;
2415                 i915_vm_put(vm);
2416
2417                 break;
2418
2419         case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2420                 args->size = 0;
2421                 args->value = i915_gem_context_no_error_capture(ctx);
2422                 break;
2423
2424         case I915_CONTEXT_PARAM_BANNABLE:
2425                 args->size = 0;
2426                 args->value = i915_gem_context_is_bannable(ctx);
2427                 break;
2428
2429         case I915_CONTEXT_PARAM_RECOVERABLE:
2430                 args->size = 0;
2431                 args->value = i915_gem_context_is_recoverable(ctx);
2432                 break;
2433
2434         case I915_CONTEXT_PARAM_PRIORITY:
2435                 args->size = 0;
2436                 args->value = ctx->sched.priority;
2437                 break;
2438
2439         case I915_CONTEXT_PARAM_SSEU:
2440                 ret = get_sseu(ctx, args);
2441                 break;
2442
2443         case I915_CONTEXT_PARAM_VM:
2444                 ret = get_ppgtt(file_priv, ctx, args);
2445                 break;
2446
2447         case I915_CONTEXT_PARAM_PERSISTENCE:
2448                 args->size = 0;
2449                 args->value = i915_gem_context_is_persistent(ctx);
2450                 break;
2451
2452         case I915_CONTEXT_PARAM_PROTECTED_CONTENT:
2453                 ret = get_protected(ctx, args);
2454                 break;
2455
2456         case I915_CONTEXT_PARAM_NO_ZEROMAP:
2457         case I915_CONTEXT_PARAM_BAN_PERIOD:
2458         case I915_CONTEXT_PARAM_ENGINES:
2459         case I915_CONTEXT_PARAM_RINGSIZE:
2460         default:
2461                 ret = -EINVAL;
2462                 break;
2463         }
2464
2465         i915_gem_context_put(ctx);
2466         return ret;
2467 }
2468
2469 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2470                                     struct drm_file *file)
2471 {
2472         struct drm_i915_file_private *file_priv = file->driver_priv;
2473         struct drm_i915_gem_context_param *args = data;
2474         struct i915_gem_proto_context *pc;
2475         struct i915_gem_context *ctx;
2476         int ret = 0;
2477
2478         mutex_lock(&file_priv->proto_context_lock);
2479         ctx = __context_lookup(file_priv, args->ctx_id);
2480         if (!ctx) {
2481                 pc = xa_load(&file_priv->proto_context_xa, args->ctx_id);
2482                 if (pc) {
2483                         /* Contexts should be finalized inside
2484                          * GEM_CONTEXT_CREATE starting with graphics
2485                          * version 13.
2486                          */
2487                         WARN_ON(GRAPHICS_VER(file_priv->dev_priv) > 12);
2488                         ret = set_proto_ctx_param(file_priv, pc, args);
2489                 } else {
2490                         ret = -ENOENT;
2491                 }
2492         }
2493         mutex_unlock(&file_priv->proto_context_lock);
2494
2495         if (ctx) {
2496                 ret = ctx_setparam(file_priv, ctx, args);
2497                 i915_gem_context_put(ctx);
2498         }
2499
2500         return ret;
2501 }
2502
2503 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2504                                        void *data, struct drm_file *file)
2505 {
2506         struct drm_i915_private *i915 = to_i915(dev);
2507         struct drm_i915_reset_stats *args = data;
2508         struct i915_gem_context *ctx;
2509
2510         if (args->flags || args->pad)
2511                 return -EINVAL;
2512
2513         ctx = i915_gem_context_lookup(file->driver_priv, args->ctx_id);
2514         if (IS_ERR(ctx))
2515                 return PTR_ERR(ctx);
2516
2517         /*
2518          * We opt for unserialised reads here. This may result in tearing
2519          * in the extremely unlikely event of a GPU hang on this context
2520          * as we are querying them. If we need that extra layer of protection,
2521          * we should wrap the hangstats with a seqlock.
2522          */
2523
2524         if (capable(CAP_SYS_ADMIN))
2525                 args->reset_count = i915_reset_count(&i915->gpu_error);
2526         else
2527                 args->reset_count = 0;
2528
2529         args->batch_active = atomic_read(&ctx->guilty_count);
2530         args->batch_pending = atomic_read(&ctx->active_count);
2531
2532         i915_gem_context_put(ctx);
2533         return 0;
2534 }
2535
2536 /* GEM context-engines iterator: for_each_gem_engine() */
2537 struct intel_context *
2538 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2539 {
2540         const struct i915_gem_engines *e = it->engines;
2541         struct intel_context *ctx;
2542
2543         if (unlikely(!e))
2544                 return NULL;
2545
2546         do {
2547                 if (it->idx >= e->num_engines)
2548                         return NULL;
2549
2550                 ctx = e->engines[it->idx++];
2551         } while (!ctx);
2552
2553         return ctx;
2554 }
2555
2556 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2557 #include "selftests/mock_context.c"
2558 #include "selftests/i915_gem_context.c"
2559 #endif
2560
2561 void i915_gem_context_module_exit(void)
2562 {
2563         kmem_cache_destroy(slab_luts);
2564 }
2565
2566 int __init i915_gem_context_module_init(void)
2567 {
2568         slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2569         if (!slab_luts)
2570                 return -ENOMEM;
2571
2572         return 0;
2573 }