drivers/gpu/drm/i915/intel_engine_cs.c

   1 /*
   2  * Copyright © 2016 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <drm/drm_print.h>
  26
  27 #include "i915_drv.h"
  28 #include "i915_vgpu.h"
  29 #include "intel_ringbuffer.h"
  30 #include "intel_lrc.h"
  31
  32 /* Haswell does have the CXT_SIZE register however it does not appear to be
  33  * valid. Now, docs explain in dwords what is in the context object. The full
  34  * size is 70720 bytes, however, the power context and execlist context will
  35  * never be saved (power context is stored elsewhere, and execlists don't work
  36  * on HSW) - so the final size, including the extra state required for the
  37  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
  38  */
  39 #define HSW_CXT_TOTAL_SIZE              (17 * PAGE_SIZE)
  40
  41 #define DEFAULT_LR_CONTEXT_RENDER_SIZE  (22 * PAGE_SIZE)
  42 #define GEN8_LR_CONTEXT_RENDER_SIZE     (20 * PAGE_SIZE)
  43 #define GEN9_LR_CONTEXT_RENDER_SIZE     (22 * PAGE_SIZE)
  44 #define GEN10_LR_CONTEXT_RENDER_SIZE    (18 * PAGE_SIZE)
  45 #define GEN11_LR_CONTEXT_RENDER_SIZE    (14 * PAGE_SIZE)
  46
  47 #define GEN8_LR_CONTEXT_OTHER_SIZE      ( 2 * PAGE_SIZE)
  48
  49 struct engine_class_info {
  50         const char *name;
  51         int (*init_legacy)(struct intel_engine_cs *engine);
  52         int (*init_execlists)(struct intel_engine_cs *engine);
  53
  54         u8 uabi_class;
  55 };
  56
  57 static const struct engine_class_info intel_engine_classes[] = {
  58         [RENDER_CLASS] = {
  59                 .name = "rcs",
  60                 .init_execlists = logical_render_ring_init,
  61                 .init_legacy = intel_init_render_ring_buffer,
  62                 .uabi_class = I915_ENGINE_CLASS_RENDER,
  63         },
  64         [COPY_ENGINE_CLASS] = {
  65                 .name = "bcs",
  66                 .init_execlists = logical_xcs_ring_init,
  67                 .init_legacy = intel_init_blt_ring_buffer,
  68                 .uabi_class = I915_ENGINE_CLASS_COPY,
  69         },
  70         [VIDEO_DECODE_CLASS] = {
  71                 .name = "vcs",
  72                 .init_execlists = logical_xcs_ring_init,
  73                 .init_legacy = intel_init_bsd_ring_buffer,
  74                 .uabi_class = I915_ENGINE_CLASS_VIDEO,
  75         },
  76         [VIDEO_ENHANCEMENT_CLASS] = {
  77                 .name = "vecs",
  78                 .init_execlists = logical_xcs_ring_init,
  79                 .init_legacy = intel_init_vebox_ring_buffer,
  80                 .uabi_class = I915_ENGINE_CLASS_VIDEO_ENHANCE,
  81         },
  82 };
  83
  84 #define MAX_MMIO_BASES 3
  85 struct engine_info {
  86         unsigned int hw_id;
  87         unsigned int uabi_id;
  88         u8 class;
  89         u8 instance;
  90         /* mmio bases table *must* be sorted in reverse gen order */
  91         struct engine_mmio_base {
  92                 u32 gen : 8;
  93                 u32 base : 24;
  94         } mmio_bases[MAX_MMIO_BASES];
  95 };
  96
  97 static const struct engine_info intel_engines[] = {
  98         [RCS] = {
  99                 .hw_id = RCS_HW,
 100                 .uabi_id = I915_EXEC_RENDER,
 101                 .class = RENDER_CLASS,
 102                 .instance = 0,
 103                 .mmio_bases = {
 104                         { .gen = 1, .base = RENDER_RING_BASE }
 105                 },
 106         },
 107         [BCS] = {
 108                 .hw_id = BCS_HW,
 109                 .uabi_id = I915_EXEC_BLT,
 110                 .class = COPY_ENGINE_CLASS,
 111                 .instance = 0,
 112                 .mmio_bases = {
 113                         { .gen = 6, .base = BLT_RING_BASE }
 114                 },
 115         },
 116         [VCS] = {
 117                 .hw_id = VCS_HW,
 118                 .uabi_id = I915_EXEC_BSD,
 119                 .class = VIDEO_DECODE_CLASS,
 120                 .instance = 0,
 121                 .mmio_bases = {
 122                         { .gen = 11, .base = GEN11_BSD_RING_BASE },
 123                         { .gen = 6, .base = GEN6_BSD_RING_BASE },
 124                         { .gen = 4, .base = BSD_RING_BASE }
 125                 },
 126         },
 127         [VCS2] = {
 128                 .hw_id = VCS2_HW,
 129                 .uabi_id = I915_EXEC_BSD,
 130                 .class = VIDEO_DECODE_CLASS,
 131                 .instance = 1,
 132                 .mmio_bases = {
 133                         { .gen = 11, .base = GEN11_BSD2_RING_BASE },
 134                         { .gen = 8, .base = GEN8_BSD2_RING_BASE }
 135                 },
 136         },
 137         [VCS3] = {
 138                 .hw_id = VCS3_HW,
 139                 .uabi_id = I915_EXEC_BSD,
 140                 .class = VIDEO_DECODE_CLASS,
 141                 .instance = 2,
 142                 .mmio_bases = {
 143                         { .gen = 11, .base = GEN11_BSD3_RING_BASE }
 144                 },
 145         },
 146         [VCS4] = {
 147                 .hw_id = VCS4_HW,
 148                 .uabi_id = I915_EXEC_BSD,
 149                 .class = VIDEO_DECODE_CLASS,
 150                 .instance = 3,
 151                 .mmio_bases = {
 152                         { .gen = 11, .base = GEN11_BSD4_RING_BASE }
 153                 },
 154         },
 155         [VECS] = {
 156                 .hw_id = VECS_HW,
 157                 .uabi_id = I915_EXEC_VEBOX,
 158                 .class = VIDEO_ENHANCEMENT_CLASS,
 159                 .instance = 0,
 160                 .mmio_bases = {
 161                         { .gen = 11, .base = GEN11_VEBOX_RING_BASE },
 162                         { .gen = 7, .base = VEBOX_RING_BASE }
 163                 },
 164         },
 165         [VECS2] = {
 166                 .hw_id = VECS2_HW,
 167                 .uabi_id = I915_EXEC_VEBOX,
 168                 .class = VIDEO_ENHANCEMENT_CLASS,
 169                 .instance = 1,
 170                 .mmio_bases = {
 171                         { .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
 172                 },
 173         },
 174 };
 175
 176 /**
 177  * ___intel_engine_context_size() - return the size of the context for an engine
 178  * @dev_priv: i915 device private
 179  * @class: engine class
 180  *
 181  * Each engine class may require a different amount of space for a context
 182  * image.
 183  *
 184  * Return: size (in bytes) of an engine class specific context image
 185  *
 186  * Note: this size includes the HWSP, which is part of the context image
 187  * in LRC mode, but does not include the "shared data page" used with
 188  * GuC submission. The caller should account for this if using the GuC.
 189  */
 190 static u32
 191 __intel_engine_context_size(struct drm_i915_private *dev_priv, u8 class)
 192 {
 193         u32 cxt_size;
 194
 195         BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
 196
 197         switch (class) {
 198         case RENDER_CLASS:
 199                 switch (INTEL_GEN(dev_priv)) {
 200                 default:
 201                         MISSING_CASE(INTEL_GEN(dev_priv));
 202                         return DEFAULT_LR_CONTEXT_RENDER_SIZE;
 203                 case 11:
 204                         return GEN11_LR_CONTEXT_RENDER_SIZE;
 205                 case 10:
 206                         return GEN10_LR_CONTEXT_RENDER_SIZE;
 207                 case 9:
 208                         return GEN9_LR_CONTEXT_RENDER_SIZE;
 209                 case 8:
 210                         return GEN8_LR_CONTEXT_RENDER_SIZE;
 211                 case 7:
 212                         if (IS_HASWELL(dev_priv))
 213                                 return HSW_CXT_TOTAL_SIZE;
 214
 215                         cxt_size = I915_READ(GEN7_CXT_SIZE);
 216                         return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
 217                                         PAGE_SIZE);
 218                 case 6:
 219                         cxt_size = I915_READ(CXT_SIZE);
 220                         return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
 221                                         PAGE_SIZE);
 222                 case 5:
 223                 case 4:
 224                 case 3:
 225                 case 2:
 226                 /* For the special day when i810 gets merged. */
 227                 case 1:
 228                         return 0;
 229                 }
 230                 break;
 231         default:
 232                 MISSING_CASE(class);
 233         case VIDEO_DECODE_CLASS:
 234         case VIDEO_ENHANCEMENT_CLASS:
 235         case COPY_ENGINE_CLASS:
 236                 if (INTEL_GEN(dev_priv) < 8)
 237                         return 0;
 238                 return GEN8_LR_CONTEXT_OTHER_SIZE;
 239         }
 240 }
 241
 242 static u32 __engine_mmio_base(struct drm_i915_private *i915,
 243                               const struct engine_mmio_base *bases)
 244 {
 245         int i;
 246
 247         for (i = 0; i < MAX_MMIO_BASES; i++)
 248                 if (INTEL_GEN(i915) >= bases[i].gen)
 249                         break;
 250
 251         GEM_BUG_ON(i == MAX_MMIO_BASES);
 252         GEM_BUG_ON(!bases[i].base);
 253
 254         return bases[i].base;
 255 }
 256
 257 static void __sprint_engine_name(char *name, const struct engine_info *info)
 258 {
 259         WARN_ON(snprintf(name, INTEL_ENGINE_CS_MAX_NAME, "%s%u",
 260                          intel_engine_classes[info->class].name,
 261                          info->instance) >= INTEL_ENGINE_CS_MAX_NAME);
 262 }
 263
 264 static int
 265 intel_engine_setup(struct drm_i915_private *dev_priv,
 266                    enum intel_engine_id id)
 267 {
 268         const struct engine_info *info = &intel_engines[id];
 269         struct intel_engine_cs *engine;
 270
 271         GEM_BUG_ON(info->class >= ARRAY_SIZE(intel_engine_classes));
 272
 273         BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
 274         BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
 275
 276         if (GEM_WARN_ON(info->class > MAX_ENGINE_CLASS))
 277                 return -EINVAL;
 278
 279         if (GEM_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
 280                 return -EINVAL;
 281
 282         if (GEM_WARN_ON(dev_priv->engine_class[info->class][info->instance]))
 283                 return -EINVAL;
 284
 285         GEM_BUG_ON(dev_priv->engine[id]);
 286         engine = kzalloc(sizeof(*engine), GFP_KERNEL);
 287         if (!engine)
 288                 return -ENOMEM;
 289
 290         engine->id = id;
 291         engine->i915 = dev_priv;
 292         __sprint_engine_name(engine->name, info);
 293         engine->hw_id = engine->guc_id = info->hw_id;
 294         engine->mmio_base = __engine_mmio_base(dev_priv, info->mmio_bases);
 295         engine->class = info->class;
 296         engine->instance = info->instance;
 297
 298         engine->uabi_id = info->uabi_id;
 299         engine->uabi_class = intel_engine_classes[info->class].uabi_class;
 300
 301         engine->context_size = __intel_engine_context_size(dev_priv,
 302                                                            engine->class);
 303         if (WARN_ON(engine->context_size > BIT(20)))
 304                 engine->context_size = 0;
 305
 306         /* Nothing to do here, execute in order of dependencies */
 307         engine->schedule = NULL;
 308
 309         seqlock_init(&engine->stats.lock);
 310
 311         ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
 312
 313         dev_priv->engine_class[info->class][info->instance] = engine;
 314         dev_priv->engine[id] = engine;
 315         return 0;
 316 }
 317
 318 /**
 319  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
 320  * @dev_priv: i915 device private
 321  *
 322  * Return: non-zero if the initialization failed.
 323  */
 324 int intel_engines_init_mmio(struct drm_i915_private *dev_priv)
 325 {
 326         struct intel_device_info *device_info = mkwrite_device_info(dev_priv);
 327         const unsigned int ring_mask = INTEL_INFO(dev_priv)->ring_mask;
 328         struct intel_engine_cs *engine;
 329         enum intel_engine_id id;
 330         unsigned int mask = 0;
 331         unsigned int i;
 332         int err;
 333
 334         WARN_ON(ring_mask == 0);
 335         WARN_ON(ring_mask &
 336                 GENMASK(sizeof(mask) * BITS_PER_BYTE - 1, I915_NUM_ENGINES));
 337
 338         for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
 339                 if (!HAS_ENGINE(dev_priv, i))
 340                         continue;
 341
 342                 err = intel_engine_setup(dev_priv, i);
 343                 if (err)
 344                         goto cleanup;
 345
 346                 mask |= ENGINE_MASK(i);
 347         }
 348
 349         /*
 350          * Catch failures to update intel_engines table when the new engines
 351          * are added to the driver by a warning and disabling the forgotten
 352          * engines.
 353          */
 354         if (WARN_ON(mask != ring_mask))
 355                 device_info->ring_mask = mask;
 356
 357         /* We always presume we have at least RCS available for later probing */
 358         if (WARN_ON(!HAS_ENGINE(dev_priv, RCS))) {
 359                 err = -ENODEV;
 360                 goto cleanup;
 361         }
 362
 363         device_info->num_rings = hweight32(mask);
 364
 365         i915_check_and_clear_faults(dev_priv);
 366
 367         return 0;
 368
 369 cleanup:
 370         for_each_engine(engine, dev_priv, id)
 371                 kfree(engine);
 372         return err;
 373 }
 374
 375 /**
 376  * intel_engines_init() - init the Engine Command Streamers
 377  * @dev_priv: i915 device private
 378  *
 379  * Return: non-zero if the initialization failed.
 380  */
 381 int intel_engines_init(struct drm_i915_private *dev_priv)
 382 {
 383         struct intel_engine_cs *engine;
 384         enum intel_engine_id id, err_id;
 385         int err;
 386
 387         for_each_engine(engine, dev_priv, id) {
 388                 const struct engine_class_info *class_info =
 389                         &intel_engine_classes[engine->class];
 390                 int (*init)(struct intel_engine_cs *engine);
 391
 392                 if (HAS_EXECLISTS(dev_priv))
 393                         init = class_info->init_execlists;
 394                 else
 395                         init = class_info->init_legacy;
 396
 397                 err = -EINVAL;
 398                 err_id = id;
 399
 400                 if (GEM_WARN_ON(!init))
 401                         goto cleanup;
 402
 403                 err = init(engine);
 404                 if (err)
 405                         goto cleanup;
 406
 407                 GEM_BUG_ON(!engine->submit_request);
 408         }
 409
 410         return 0;
 411
 412 cleanup:
 413         for_each_engine(engine, dev_priv, id) {
 414                 if (id >= err_id) {
 415                         kfree(engine);
 416                         dev_priv->engine[id] = NULL;
 417                 } else {
 418                         dev_priv->gt.cleanup_engine(engine);
 419                 }
 420         }
 421         return err;
 422 }
 423
 424 void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno)
 425 {
 426         struct drm_i915_private *dev_priv = engine->i915;
 427
 428         /* Our semaphore implementation is strictly monotonic (i.e. we proceed
 429          * so long as the semaphore value in the register/page is greater
 430          * than the sync value), so whenever we reset the seqno,
 431          * so long as we reset the tracking semaphore value to 0, it will
 432          * always be before the next request's seqno. If we don't reset
 433          * the semaphore value, then when the seqno moves backwards all
 434          * future waits will complete instantly (causing rendering corruption).
 435          */
 436         if (IS_GEN6(dev_priv) || IS_GEN7(dev_priv)) {
 437                 I915_WRITE(RING_SYNC_0(engine->mmio_base), 0);
 438                 I915_WRITE(RING_SYNC_1(engine->mmio_base), 0);
 439                 if (HAS_VEBOX(dev_priv))
 440                         I915_WRITE(RING_SYNC_2(engine->mmio_base), 0);
 441         }
 442
 443         intel_write_status_page(engine, I915_GEM_HWS_INDEX, seqno);
 444         clear_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 445
 446         /* After manually advancing the seqno, fake the interrupt in case
 447          * there are any waiters for that seqno.
 448          */
 449         intel_engine_wakeup(engine);
 450
 451         GEM_BUG_ON(intel_engine_get_seqno(engine) != seqno);
 452 }
 453
 454 static void intel_engine_init_batch_pool(struct intel_engine_cs *engine)
 455 {
 456         i915_gem_batch_pool_init(&engine->batch_pool, engine);
 457 }
 458
 459 static bool csb_force_mmio(struct drm_i915_private *i915)
 460 {
 461         /* Older GVT emulation depends upon intercepting CSB mmio */
 462         if (intel_vgpu_active(i915) && !intel_vgpu_has_hwsp_emulation(i915))
 463                 return true;
 464
 465         return false;
 466 }
 467
 468 static void intel_engine_init_execlist(struct intel_engine_cs *engine)
 469 {
 470         struct intel_engine_execlists * const execlists = &engine->execlists;
 471
 472         execlists->csb_use_mmio = csb_force_mmio(engine->i915);
 473
 474         execlists->port_mask = 1;
 475         BUILD_BUG_ON_NOT_POWER_OF_2(execlists_num_ports(execlists));
 476         GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
 477
 478         execlists->queue_priority = INT_MIN;
 479         execlists->queue = RB_ROOT;
 480         execlists->first = NULL;
 481 }
 482
 483 /**
 484  * intel_engines_setup_common - setup engine state not requiring hw access
 485  * @engine: Engine to setup.
 486  *
 487  * Initializes @engine@ structure members shared between legacy and execlists
 488  * submission modes which do not require hardware access.
 489  *
 490  * Typically done early in the submission mode specific engine setup stage.
 491  */
 492 void intel_engine_setup_common(struct intel_engine_cs *engine)
 493 {
 494         i915_timeline_init(engine->i915, &engine->timeline, engine->name);
 495
 496         intel_engine_init_execlist(engine);
 497         intel_engine_init_hangcheck(engine);
 498         intel_engine_init_batch_pool(engine);
 499         intel_engine_init_cmd_parser(engine);
 500 }
 501
 502 int intel_engine_create_scratch(struct intel_engine_cs *engine, int size)
 503 {
 504         struct drm_i915_gem_object *obj;
 505         struct i915_vma *vma;
 506         int ret;
 507
 508         WARN_ON(engine->scratch);
 509
 510         obj = i915_gem_object_create_stolen(engine->i915, size);
 511         if (!obj)
 512                 obj = i915_gem_object_create_internal(engine->i915, size);
 513         if (IS_ERR(obj)) {
 514                 DRM_ERROR("Failed to allocate scratch page\n");
 515                 return PTR_ERR(obj);
 516         }
 517
 518         vma = i915_vma_instance(obj, &engine->i915->ggtt.base, NULL);
 519         if (IS_ERR(vma)) {
 520                 ret = PTR_ERR(vma);
 521                 goto err_unref;
 522         }
 523
 524         ret = i915_vma_pin(vma, 0, 4096, PIN_GLOBAL | PIN_HIGH);
 525         if (ret)
 526                 goto err_unref;
 527
 528         engine->scratch = vma;
 529         return 0;
 530
 531 err_unref:
 532         i915_gem_object_put(obj);
 533         return ret;
 534 }
 535
 536 static void intel_engine_cleanup_scratch(struct intel_engine_cs *engine)
 537 {
 538         i915_vma_unpin_and_release(&engine->scratch);
 539 }
 540
 541 static void cleanup_phys_status_page(struct intel_engine_cs *engine)
 542 {
 543         struct drm_i915_private *dev_priv = engine->i915;
 544
 545         if (!dev_priv->status_page_dmah)
 546                 return;
 547
 548         drm_pci_free(&dev_priv->drm, dev_priv->status_page_dmah);
 549         engine->status_page.page_addr = NULL;
 550 }
 551
 552 static void cleanup_status_page(struct intel_engine_cs *engine)
 553 {
 554         struct i915_vma *vma;
 555         struct drm_i915_gem_object *obj;
 556
 557         vma = fetch_and_zero(&engine->status_page.vma);
 558         if (!vma)
 559                 return;
 560
 561         obj = vma->obj;
 562
 563         i915_vma_unpin(vma);
 564         i915_vma_close(vma);
 565
 566         i915_gem_object_unpin_map(obj);
 567         __i915_gem_object_release_unless_active(obj);
 568 }
 569
 570 static int init_status_page(struct intel_engine_cs *engine)
 571 {
 572         struct drm_i915_gem_object *obj;
 573         struct i915_vma *vma;
 574         unsigned int flags;
 575         void *vaddr;
 576         int ret;
 577
 578         obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
 579         if (IS_ERR(obj)) {
 580                 DRM_ERROR("Failed to allocate status page\n");
 581                 return PTR_ERR(obj);
 582         }
 583
 584         ret = i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
 585         if (ret)
 586                 goto err;
 587
 588         vma = i915_vma_instance(obj, &engine->i915->ggtt.base, NULL);
 589         if (IS_ERR(vma)) {
 590                 ret = PTR_ERR(vma);
 591                 goto err;
 592         }
 593
 594         flags = PIN_GLOBAL;
 595         if (!HAS_LLC(engine->i915))
 596                 /* On g33, we cannot place HWS above 256MiB, so
 597                  * restrict its pinning to the low mappable arena.
 598                  * Though this restriction is not documented for
 599                  * gen4, gen5, or byt, they also behave similarly
 600                  * and hang if the HWS is placed at the top of the
 601                  * GTT. To generalise, it appears that all !llc
 602                  * platforms have issues with us placing the HWS
 603                  * above the mappable region (even though we never
 604                  * actually map it).
 605                  */
 606                 flags |= PIN_MAPPABLE;
 607         else
 608                 flags |= PIN_HIGH;
 609         ret = i915_vma_pin(vma, 0, 4096, flags);
 610         if (ret)
 611                 goto err;
 612
 613         vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
 614         if (IS_ERR(vaddr)) {
 615                 ret = PTR_ERR(vaddr);
 616                 goto err_unpin;
 617         }
 618
 619         engine->status_page.vma = vma;
 620         engine->status_page.ggtt_offset = i915_ggtt_offset(vma);
 621         engine->status_page.page_addr = memset(vaddr, 0, PAGE_SIZE);
 622         return 0;
 623
 624 err_unpin:
 625         i915_vma_unpin(vma);
 626 err:
 627         i915_gem_object_put(obj);
 628         return ret;
 629 }
 630
 631 static int init_phys_status_page(struct intel_engine_cs *engine)
 632 {
 633         struct drm_i915_private *dev_priv = engine->i915;
 634
 635         GEM_BUG_ON(engine->id != RCS);
 636
 637         dev_priv->status_page_dmah =
 638                 drm_pci_alloc(&dev_priv->drm, PAGE_SIZE, PAGE_SIZE);
 639         if (!dev_priv->status_page_dmah)
 640                 return -ENOMEM;
 641
 642         engine->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
 643         memset(engine->status_page.page_addr, 0, PAGE_SIZE);
 644
 645         return 0;
 646 }
 647
 648 /**
 649  * intel_engines_init_common - initialize cengine state which might require hw access
 650  * @engine: Engine to initialize.
 651  *
 652  * Initializes @engine@ structure members shared between legacy and execlists
 653  * submission modes which do require hardware access.
 654  *
 655  * Typcally done at later stages of submission mode specific engine setup.
 656  *
 657  * Returns zero on success or an error code on failure.
 658  */
 659 int intel_engine_init_common(struct intel_engine_cs *engine)
 660 {
 661         struct intel_ring *ring;
 662         int ret;
 663
 664         engine->set_default_submission(engine);
 665
 666         /* We may need to do things with the shrinker which
 667          * require us to immediately switch back to the default
 668          * context. This can cause a problem as pinning the
 669          * default context also requires GTT space which may not
 670          * be available. To avoid this we always pin the default
 671          * context.
 672          */
 673         ring = intel_context_pin(engine->i915->kernel_context, engine);
 674         if (IS_ERR(ring))
 675                 return PTR_ERR(ring);
 676
 677         /*
 678          * Similarly the preempt context must always be available so that
 679          * we can interrupt the engine at any time.
 680          */
 681         if (engine->i915->preempt_context) {
 682                 ring = intel_context_pin(engine->i915->preempt_context, engine);
 683                 if (IS_ERR(ring)) {
 684                         ret = PTR_ERR(ring);
 685                         goto err_unpin_kernel;
 686                 }
 687         }
 688
 689         ret = intel_engine_init_breadcrumbs(engine);
 690         if (ret)
 691                 goto err_unpin_preempt;
 692
 693         if (HWS_NEEDS_PHYSICAL(engine->i915))
 694                 ret = init_phys_status_page(engine);
 695         else
 696                 ret = init_status_page(engine);
 697         if (ret)
 698                 goto err_breadcrumbs;
 699
 700         return 0;
 701
 702 err_breadcrumbs:
 703         intel_engine_fini_breadcrumbs(engine);
 704 err_unpin_preempt:
 705         if (engine->i915->preempt_context)
 706                 intel_context_unpin(engine->i915->preempt_context, engine);
 707 err_unpin_kernel:
 708         intel_context_unpin(engine->i915->kernel_context, engine);
 709         return ret;
 710 }
 711
 712 /**
 713  * intel_engines_cleanup_common - cleans up the engine state created by
 714  *                                the common initiailizers.
 715  * @engine: Engine to cleanup.
 716  *
 717  * This cleans up everything created by the common helpers.
 718  */
 719 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
 720 {
 721         intel_engine_cleanup_scratch(engine);
 722
 723         if (HWS_NEEDS_PHYSICAL(engine->i915))
 724                 cleanup_phys_status_page(engine);
 725         else
 726                 cleanup_status_page(engine);
 727
 728         intel_engine_fini_breadcrumbs(engine);
 729         intel_engine_cleanup_cmd_parser(engine);
 730         i915_gem_batch_pool_fini(&engine->batch_pool);
 731
 732         if (engine->default_state)
 733                 i915_gem_object_put(engine->default_state);
 734
 735         if (engine->i915->preempt_context)
 736                 intel_context_unpin(engine->i915->preempt_context, engine);
 737         intel_context_unpin(engine->i915->kernel_context, engine);
 738
 739         i915_timeline_fini(&engine->timeline);
 740 }
 741
 742 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
 743 {
 744         struct drm_i915_private *dev_priv = engine->i915;
 745         u64 acthd;
 746
 747         if (INTEL_GEN(dev_priv) >= 8)
 748                 acthd = I915_READ64_2x32(RING_ACTHD(engine->mmio_base),
 749                                          RING_ACTHD_UDW(engine->mmio_base));
 750         else if (INTEL_GEN(dev_priv) >= 4)
 751                 acthd = I915_READ(RING_ACTHD(engine->mmio_base));
 752         else
 753                 acthd = I915_READ(ACTHD);
 754
 755         return acthd;
 756 }
 757
 758 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
 759 {
 760         struct drm_i915_private *dev_priv = engine->i915;
 761         u64 bbaddr;
 762
 763         if (INTEL_GEN(dev_priv) >= 8)
 764                 bbaddr = I915_READ64_2x32(RING_BBADDR(engine->mmio_base),
 765                                           RING_BBADDR_UDW(engine->mmio_base));
 766         else
 767                 bbaddr = I915_READ(RING_BBADDR(engine->mmio_base));
 768
 769         return bbaddr;
 770 }
 771
 772 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
 773 {
 774         switch (type) {
 775         case I915_CACHE_NONE: return " uncached";
 776         case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
 777         case I915_CACHE_L3_LLC: return " L3+LLC";
 778         case I915_CACHE_WT: return " WT";
 779         default: return "";
 780         }
 781 }
 782
 783 static inline uint32_t
 784 read_subslice_reg(struct drm_i915_private *dev_priv, int slice,
 785                   int subslice, i915_reg_t reg)
 786 {
 787         uint32_t mcr_slice_subslice_mask;
 788         uint32_t mcr_slice_subslice_select;
 789         uint32_t mcr;
 790         uint32_t ret;
 791         enum forcewake_domains fw_domains;
 792
 793         if (INTEL_GEN(dev_priv) >= 11) {
 794                 mcr_slice_subslice_mask = GEN11_MCR_SLICE_MASK |
 795                                           GEN11_MCR_SUBSLICE_MASK;
 796                 mcr_slice_subslice_select = GEN11_MCR_SLICE(slice) |
 797                                             GEN11_MCR_SUBSLICE(subslice);
 798         } else {
 799                 mcr_slice_subslice_mask = GEN8_MCR_SLICE_MASK |
 800                                           GEN8_MCR_SUBSLICE_MASK;
 801                 mcr_slice_subslice_select = GEN8_MCR_SLICE(slice) |
 802                                             GEN8_MCR_SUBSLICE(subslice);
 803         }
 804
 805         fw_domains = intel_uncore_forcewake_for_reg(dev_priv, reg,
 806                                                     FW_REG_READ);
 807         fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
 808                                                      GEN8_MCR_SELECTOR,
 809                                                      FW_REG_READ | FW_REG_WRITE);
 810
 811         spin_lock_irq(&dev_priv->uncore.lock);
 812         intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
 813
 814         mcr = I915_READ_FW(GEN8_MCR_SELECTOR);
 815         /*
 816          * The HW expects the slice and sublice selectors to be reset to 0
 817          * after reading out the registers.
 818          */
 819         WARN_ON_ONCE(mcr & mcr_slice_subslice_mask);
 820         mcr &= ~mcr_slice_subslice_mask;
 821         mcr |= mcr_slice_subslice_select;
 822         I915_WRITE_FW(GEN8_MCR_SELECTOR, mcr);
 823
 824         ret = I915_READ_FW(reg);
 825
 826         mcr &= ~mcr_slice_subslice_mask;
 827         I915_WRITE_FW(GEN8_MCR_SELECTOR, mcr);
 828
 829         intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
 830         spin_unlock_irq(&dev_priv->uncore.lock);
 831
 832         return ret;
 833 }
 834
 835 /* NB: please notice the memset */
 836 void intel_engine_get_instdone(struct intel_engine_cs *engine,
 837                                struct intel_instdone *instdone)
 838 {
 839         struct drm_i915_private *dev_priv = engine->i915;
 840         u32 mmio_base = engine->mmio_base;
 841         int slice;
 842         int subslice;
 843
 844         memset(instdone, 0, sizeof(*instdone));
 845
 846         switch (INTEL_GEN(dev_priv)) {
 847         default:
 848                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 849
 850                 if (engine->id != RCS)
 851                         break;
 852
 853                 instdone->slice_common = I915_READ(GEN7_SC_INSTDONE);
 854                 for_each_instdone_slice_subslice(dev_priv, slice, subslice) {
 855                         instdone->sampler[slice][subslice] =
 856                                 read_subslice_reg(dev_priv, slice, subslice,
 857                                                   GEN7_SAMPLER_INSTDONE);
 858                         instdone->row[slice][subslice] =
 859                                 read_subslice_reg(dev_priv, slice, subslice,
 860                                                   GEN7_ROW_INSTDONE);
 861                 }
 862                 break;
 863         case 7:
 864                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 865
 866                 if (engine->id != RCS)
 867                         break;
 868
 869                 instdone->slice_common = I915_READ(GEN7_SC_INSTDONE);
 870                 instdone->sampler[0][0] = I915_READ(GEN7_SAMPLER_INSTDONE);
 871                 instdone->row[0][0] = I915_READ(GEN7_ROW_INSTDONE);
 872
 873                 break;
 874         case 6:
 875         case 5:
 876         case 4:
 877                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 878
 879                 if (engine->id == RCS)
 880                         /* HACK: Using the wrong struct member */
 881                         instdone->slice_common = I915_READ(GEN4_INSTDONE1);
 882                 break;
 883         case 3:
 884         case 2:
 885                 instdone->instdone = I915_READ(GEN2_INSTDONE);
 886                 break;
 887         }
 888 }
 889
 890 static bool ring_is_idle(struct intel_engine_cs *engine)
 891 {
 892         struct drm_i915_private *dev_priv = engine->i915;
 893         bool idle = true;
 894
 895         /* If the whole device is asleep, the engine must be idle */
 896         if (!intel_runtime_pm_get_if_in_use(dev_priv))
 897                 return true;
 898
 899         /* First check that no commands are left in the ring */
 900         if ((I915_READ_HEAD(engine) & HEAD_ADDR) !=
 901             (I915_READ_TAIL(engine) & TAIL_ADDR))
 902                 idle = false;
 903
 904         /* No bit for gen2, so assume the CS parser is idle */
 905         if (INTEL_GEN(dev_priv) > 2 && !(I915_READ_MODE(engine) & MODE_IDLE))
 906                 idle = false;
 907
 908         intel_runtime_pm_put(dev_priv);
 909
 910         return idle;
 911 }
 912
 913 /**
 914  * intel_engine_is_idle() - Report if the engine has finished process all work
 915  * @engine: the intel_engine_cs
 916  *
 917  * Return true if there are no requests pending, nothing left to be submitted
 918  * to hardware, and that the engine is idle.
 919  */
 920 bool intel_engine_is_idle(struct intel_engine_cs *engine)
 921 {
 922         struct drm_i915_private *dev_priv = engine->i915;
 923
 924         /* More white lies, if wedged, hw state is inconsistent */
 925         if (i915_terminally_wedged(&dev_priv->gpu_error))
 926                 return true;
 927
 928         /* Any inflight/incomplete requests? */
 929         if (!i915_seqno_passed(intel_engine_get_seqno(engine),
 930                                intel_engine_last_submit(engine)))
 931                 return false;
 932
 933         if (I915_SELFTEST_ONLY(engine->breadcrumbs.mock))
 934                 return true;
 935
 936         /* Waiting to drain ELSP? */
 937         if (READ_ONCE(engine->execlists.active))
 938                 return false;
 939
 940         /* ELSP is empty, but there are ready requests? */
 941         if (READ_ONCE(engine->execlists.first))
 942                 return false;
 943
 944         /* Ring stopped? */
 945         if (!ring_is_idle(engine))
 946                 return false;
 947
 948         return true;
 949 }
 950
 951 bool intel_engines_are_idle(struct drm_i915_private *dev_priv)
 952 {
 953         struct intel_engine_cs *engine;
 954         enum intel_engine_id id;
 955
 956         /*
 957          * If the driver is wedged, HW state may be very inconsistent and
 958          * report that it is still busy, even though we have stopped using it.
 959          */
 960         if (i915_terminally_wedged(&dev_priv->gpu_error))
 961                 return true;
 962
 963         for_each_engine(engine, dev_priv, id) {
 964                 if (!intel_engine_is_idle(engine))
 965                         return false;
 966         }
 967
 968         return true;
 969 }
 970
 971 /**
 972  * intel_engine_has_kernel_context:
 973  * @engine: the engine
 974  *
 975  * Returns true if the last context to be executed on this engine, or has been
 976  * executed if the engine is already idle, is the kernel context
 977  * (#i915.kernel_context).
 978  */
 979 bool intel_engine_has_kernel_context(const struct intel_engine_cs *engine)
 980 {
 981         const struct i915_gem_context * const kernel_context =
 982                 engine->i915->kernel_context;
 983         struct i915_request *rq;
 984
 985         lockdep_assert_held(&engine->i915->drm.struct_mutex);
 986
 987         /*
 988          * Check the last context seen by the engine. If active, it will be
 989          * the last request that remains in the timeline. When idle, it is
 990          * the last executed context as tracked by retirement.
 991          */
 992         rq = __i915_gem_active_peek(&engine->timeline.last_request);
 993         if (rq)
 994                 return rq->ctx == kernel_context;
 995         else
 996                 return engine->last_retired_context == kernel_context;
 997 }
 998
 999 void intel_engines_reset_default_submission(struct drm_i915_private *i915)
1000 {
1001         struct intel_engine_cs *engine;
1002         enum intel_engine_id id;
1003
1004         for_each_engine(engine, i915, id)
1005                 engine->set_default_submission(engine);
1006 }
1007
1008 /**
1009  * intel_engines_park: called when the GT is transitioning from busy->idle
1010  * @i915: the i915 device
1011  *
1012  * The GT is now idle and about to go to sleep (maybe never to wake again?).
1013  * Time for us to tidy and put away our toys (release resources back to the
1014  * system).
1015  */
1016 void intel_engines_park(struct drm_i915_private *i915)
1017 {
1018         struct intel_engine_cs *engine;
1019         enum intel_engine_id id;
1020
1021         for_each_engine(engine, i915, id) {
1022                 /* Flush the residual irq tasklets first. */
1023                 intel_engine_disarm_breadcrumbs(engine);
1024                 tasklet_kill(&engine->execlists.tasklet);
1025
1026                 /*
1027                  * We are committed now to parking the engines, make sure there
1028                  * will be no more interrupts arriving later and the engines
1029                  * are truly idle.
1030                  */
1031                 if (wait_for(intel_engine_is_idle(engine), 10)) {
1032                         struct drm_printer p = drm_debug_printer(__func__);
1033
1034                         dev_err(i915->drm.dev,
1035                                 "%s is not idle before parking\n",
1036                                 engine->name);
1037                         intel_engine_dump(engine, &p, NULL);
1038                 }
1039
1040                 /* Must be reset upon idling, or we may miss the busy wakeup. */
1041                 GEM_BUG_ON(engine->execlists.queue_priority != INT_MIN);
1042
1043                 if (engine->park)
1044                         engine->park(engine);
1045
1046                 i915_gem_batch_pool_fini(&engine->batch_pool);
1047                 engine->execlists.no_priolist = false;
1048         }
1049 }
1050
1051 /**
1052  * intel_engines_unpark: called when the GT is transitioning from idle->busy
1053  * @i915: the i915 device
1054  *
1055  * The GT was idle and now about to fire up with some new user requests.
1056  */
1057 void intel_engines_unpark(struct drm_i915_private *i915)
1058 {
1059         struct intel_engine_cs *engine;
1060         enum intel_engine_id id;
1061
1062         for_each_engine(engine, i915, id) {
1063                 if (engine->unpark)
1064                         engine->unpark(engine);
1065
1066                 intel_engine_init_hangcheck(engine);
1067         }
1068 }
1069
1070 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1071 {
1072         switch (INTEL_GEN(engine->i915)) {
1073         case 2:
1074                 return false; /* uses physical not virtual addresses */
1075         case 3:
1076                 /* maybe only uses physical not virtual addresses */
1077                 return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1078         case 6:
1079                 return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1080         default:
1081                 return true;
1082         }
1083 }
1084
1085 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915)
1086 {
1087         struct intel_engine_cs *engine;
1088         enum intel_engine_id id;
1089         unsigned int which;
1090
1091         which = 0;
1092         for_each_engine(engine, i915, id)
1093                 if (engine->default_state)
1094                         which |= BIT(engine->uabi_class);
1095
1096         return which;
1097 }
1098
1099 static int print_sched_attr(struct drm_i915_private *i915,
1100                             const struct i915_sched_attr *attr,
1101                             char *buf, int x, int len)
1102 {
1103         if (attr->priority == I915_PRIORITY_INVALID)
1104                 return x;
1105
1106         x += snprintf(buf + x, len - x,
1107                       " prio=%d", attr->priority);
1108
1109         return x;
1110 }
1111
1112 static void print_request(struct drm_printer *m,
1113                           struct i915_request *rq,
1114                           const char *prefix)
1115 {
1116         const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1117         char buf[80];
1118         int x = 0;
1119
1120         x = print_sched_attr(rq->i915, &rq->sched.attr, buf, x, sizeof(buf));
1121
1122         drm_printf(m, "%s%x%s [%llx:%x]%s @ %dms: %s\n",
1123                    prefix,
1124                    rq->global_seqno,
1125                    i915_request_completed(rq) ? "!" : "",
1126                    rq->fence.context, rq->fence.seqno,
1127                    buf,
1128                    jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1129                    name);
1130 }
1131
1132 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1133 {
1134         const size_t rowsize = 8 * sizeof(u32);
1135         const void *prev = NULL;
1136         bool skip = false;
1137         size_t pos;
1138
1139         for (pos = 0; pos < len; pos += rowsize) {
1140                 char line[128];
1141
1142                 if (prev && !memcmp(prev, buf + pos, rowsize)) {
1143                         if (!skip) {
1144                                 drm_printf(m, "*\n");
1145                                 skip = true;
1146                         }
1147                         continue;
1148                 }
1149
1150                 WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1151                                                 rowsize, sizeof(u32),
1152                                                 line, sizeof(line),
1153                                                 false) >= sizeof(line));
1154                 drm_printf(m, "%08zx %s\n", pos, line);
1155
1156                 prev = buf + pos;
1157                 skip = false;
1158         }
1159 }
1160
1161 static void intel_engine_print_registers(const struct intel_engine_cs *engine,
1162                                          struct drm_printer *m)
1163 {
1164         struct drm_i915_private *dev_priv = engine->i915;
1165         const struct intel_engine_execlists * const execlists =
1166                 &engine->execlists;
1167         u64 addr;
1168
1169         drm_printf(m, "\tRING_START: 0x%08x\n",
1170                    I915_READ(RING_START(engine->mmio_base)));
1171         drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1172                    I915_READ(RING_HEAD(engine->mmio_base)) & HEAD_ADDR);
1173         drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1174                    I915_READ(RING_TAIL(engine->mmio_base)) & TAIL_ADDR);
1175         drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1176                    I915_READ(RING_CTL(engine->mmio_base)),
1177                    I915_READ(RING_CTL(engine->mmio_base)) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1178         if (INTEL_GEN(engine->i915) > 2) {
1179                 drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1180                            I915_READ(RING_MI_MODE(engine->mmio_base)),
1181                            I915_READ(RING_MI_MODE(engine->mmio_base)) & (MODE_IDLE) ? " [idle]" : "");
1182         }
1183
1184         if (INTEL_GEN(dev_priv) >= 6) {
1185                 drm_printf(m, "\tRING_IMR: %08x\n", I915_READ_IMR(engine));
1186         }
1187
1188         if (HAS_LEGACY_SEMAPHORES(dev_priv)) {
1189                 drm_printf(m, "\tSYNC_0: 0x%08x\n",
1190                            I915_READ(RING_SYNC_0(engine->mmio_base)));
1191                 drm_printf(m, "\tSYNC_1: 0x%08x\n",
1192                            I915_READ(RING_SYNC_1(engine->mmio_base)));
1193                 if (HAS_VEBOX(dev_priv))
1194                         drm_printf(m, "\tSYNC_2: 0x%08x\n",
1195                                    I915_READ(RING_SYNC_2(engine->mmio_base)));
1196         }
1197
1198         addr = intel_engine_get_active_head(engine);
1199         drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1200                    upper_32_bits(addr), lower_32_bits(addr));
1201         addr = intel_engine_get_last_batch_head(engine);
1202         drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1203                    upper_32_bits(addr), lower_32_bits(addr));
1204         if (INTEL_GEN(dev_priv) >= 8)
1205                 addr = I915_READ64_2x32(RING_DMA_FADD(engine->mmio_base),
1206                                         RING_DMA_FADD_UDW(engine->mmio_base));
1207         else if (INTEL_GEN(dev_priv) >= 4)
1208                 addr = I915_READ(RING_DMA_FADD(engine->mmio_base));
1209         else
1210                 addr = I915_READ(DMA_FADD_I8XX);
1211         drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1212                    upper_32_bits(addr), lower_32_bits(addr));
1213         if (INTEL_GEN(dev_priv) >= 4) {
1214                 drm_printf(m, "\tIPEIR: 0x%08x\n",
1215                            I915_READ(RING_IPEIR(engine->mmio_base)));
1216                 drm_printf(m, "\tIPEHR: 0x%08x\n",
1217                            I915_READ(RING_IPEHR(engine->mmio_base)));
1218         } else {
1219                 drm_printf(m, "\tIPEIR: 0x%08x\n", I915_READ(IPEIR));
1220                 drm_printf(m, "\tIPEHR: 0x%08x\n", I915_READ(IPEHR));
1221         }
1222
1223         if (HAS_EXECLISTS(dev_priv)) {
1224                 const u32 *hws = &engine->status_page.page_addr[I915_HWS_CSB_BUF0_INDEX];
1225                 u32 ptr, read, write;
1226                 unsigned int idx;
1227
1228                 drm_printf(m, "\tExeclist status: 0x%08x %08x\n",
1229                            I915_READ(RING_EXECLIST_STATUS_LO(engine)),
1230                            I915_READ(RING_EXECLIST_STATUS_HI(engine)));
1231
1232                 ptr = I915_READ(RING_CONTEXT_STATUS_PTR(engine));
1233                 read = GEN8_CSB_READ_PTR(ptr);
1234                 write = GEN8_CSB_WRITE_PTR(ptr);
1235                 drm_printf(m, "\tExeclist CSB read %d [%d cached], write %d [%d from hws], interrupt posted? %s, tasklet queued? %s (%s)\n",
1236                            read, execlists->csb_head,
1237                            write,
1238                            intel_read_status_page(engine, intel_hws_csb_write_index(engine->i915)),
1239                            yesno(test_bit(ENGINE_IRQ_EXECLIST,
1240                                           &engine->irq_posted)),
1241                            yesno(test_bit(TASKLET_STATE_SCHED,
1242                                           &engine->execlists.tasklet.state)),
1243                            enableddisabled(!atomic_read(&engine->execlists.tasklet.count)));
1244                 if (read >= GEN8_CSB_ENTRIES)
1245                         read = 0;
1246                 if (write >= GEN8_CSB_ENTRIES)
1247                         write = 0;
1248                 if (read > write)
1249                         write += GEN8_CSB_ENTRIES;
1250                 while (read < write) {
1251                         idx = ++read % GEN8_CSB_ENTRIES;
1252                         drm_printf(m, "\tExeclist CSB[%d]: 0x%08x [0x%08x in hwsp], context: %d [%d in hwsp]\n",
1253                                    idx,
1254                                    I915_READ(RING_CONTEXT_STATUS_BUF_LO(engine, idx)),
1255                                    hws[idx * 2],
1256                                    I915_READ(RING_CONTEXT_STATUS_BUF_HI(engine, idx)),
1257                                    hws[idx * 2 + 1]);
1258                 }
1259
1260                 rcu_read_lock();
1261                 for (idx = 0; idx < execlists_num_ports(execlists); idx++) {
1262                         struct i915_request *rq;
1263                         unsigned int count;
1264
1265                         rq = port_unpack(&execlists->port[idx], &count);
1266                         if (rq) {
1267                                 char hdr[80];
1268
1269                                 snprintf(hdr, sizeof(hdr),
1270                                          "\t\tELSP[%d] count=%d, ring->start=%08x, rq: ",
1271                                          idx, count,
1272                                          i915_ggtt_offset(rq->ring->vma));
1273                                 print_request(m, rq, hdr);
1274                         } else {
1275                                 drm_printf(m, "\t\tELSP[%d] idle\n", idx);
1276                         }
1277                 }
1278                 drm_printf(m, "\t\tHW active? 0x%x\n", execlists->active);
1279                 rcu_read_unlock();
1280         } else if (INTEL_GEN(dev_priv) > 6) {
1281                 drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1282                            I915_READ(RING_PP_DIR_BASE(engine)));
1283                 drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1284                            I915_READ(RING_PP_DIR_BASE_READ(engine)));
1285                 drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1286                            I915_READ(RING_PP_DIR_DCLV(engine)));
1287         }
1288 }
1289
1290 void intel_engine_dump(struct intel_engine_cs *engine,
1291                        struct drm_printer *m,
1292                        const char *header, ...)
1293 {
1294         const int MAX_REQUESTS_TO_SHOW = 8;
1295         struct intel_breadcrumbs * const b = &engine->breadcrumbs;
1296         const struct intel_engine_execlists * const execlists = &engine->execlists;
1297         struct i915_gpu_error * const error = &engine->i915->gpu_error;
1298         struct i915_request *rq, *last;
1299         struct rb_node *rb;
1300         int count;
1301
1302         if (header) {
1303                 va_list ap;
1304
1305                 va_start(ap, header);
1306                 drm_vprintf(m, header, &ap);
1307                 va_end(ap);
1308         }
1309
1310         if (i915_terminally_wedged(&engine->i915->gpu_error))
1311                 drm_printf(m, "*** WEDGED ***\n");
1312
1313         drm_printf(m, "\tcurrent seqno %x, last %x, hangcheck %x [%d ms]\n",
1314                    intel_engine_get_seqno(engine),
1315                    intel_engine_last_submit(engine),
1316                    engine->hangcheck.seqno,
1317                    jiffies_to_msecs(jiffies - engine->hangcheck.action_timestamp));
1318         drm_printf(m, "\tReset count: %d (global %d)\n",
1319                    i915_reset_engine_count(error, engine),
1320                    i915_reset_count(error));
1321
1322         rcu_read_lock();
1323
1324         drm_printf(m, "\tRequests:\n");
1325
1326         rq = list_first_entry(&engine->timeline.requests,
1327                               struct i915_request, link);
1328         if (&rq->link != &engine->timeline.requests)
1329                 print_request(m, rq, "\t\tfirst  ");
1330
1331         rq = list_last_entry(&engine->timeline.requests,
1332                              struct i915_request, link);
1333         if (&rq->link != &engine->timeline.requests)
1334                 print_request(m, rq, "\t\tlast   ");
1335
1336         rq = i915_gem_find_active_request(engine);
1337         if (rq) {
1338                 print_request(m, rq, "\t\tactive ");
1339                 drm_printf(m,
1340                            "\t\t[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]\n",
1341                            rq->head, rq->postfix, rq->tail,
1342                            rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1343                            rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1344                 drm_printf(m, "\t\tring->start:  0x%08x\n",
1345                            i915_ggtt_offset(rq->ring->vma));
1346                 drm_printf(m, "\t\tring->head:   0x%08x\n",
1347                            rq->ring->head);
1348                 drm_printf(m, "\t\tring->tail:   0x%08x\n",
1349                            rq->ring->tail);
1350                 drm_printf(m, "\t\tring->emit:   0x%08x\n",
1351                            rq->ring->emit);
1352                 drm_printf(m, "\t\tring->space:  0x%08x\n",
1353                            rq->ring->space);
1354         }
1355
1356         rcu_read_unlock();
1357
1358         if (intel_runtime_pm_get_if_in_use(engine->i915)) {
1359                 intel_engine_print_registers(engine, m);
1360                 intel_runtime_pm_put(engine->i915);
1361         } else {
1362                 drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1363         }
1364
1365         spin_lock_irq(&engine->timeline.lock);
1366
1367         last = NULL;
1368         count = 0;
1369         list_for_each_entry(rq, &engine->timeline.requests, link) {
1370                 if (count++ < MAX_REQUESTS_TO_SHOW - 1)
1371                         print_request(m, rq, "\t\tE ");
1372                 else
1373                         last = rq;
1374         }
1375         if (last) {
1376                 if (count > MAX_REQUESTS_TO_SHOW) {
1377                         drm_printf(m,
1378                                    "\t\t...skipping %d executing requests...\n",
1379                                    count - MAX_REQUESTS_TO_SHOW);
1380                 }
1381                 print_request(m, last, "\t\tE ");
1382         }
1383
1384         last = NULL;
1385         count = 0;
1386         drm_printf(m, "\t\tQueue priority: %d\n", execlists->queue_priority);
1387         for (rb = execlists->first; rb; rb = rb_next(rb)) {
1388                 struct i915_priolist *p =
1389                         rb_entry(rb, typeof(*p), node);
1390
1391                 list_for_each_entry(rq, &p->requests, sched.link) {
1392                         if (count++ < MAX_REQUESTS_TO_SHOW - 1)
1393                                 print_request(m, rq, "\t\tQ ");
1394                         else
1395                                 last = rq;
1396                 }
1397         }
1398         if (last) {
1399                 if (count > MAX_REQUESTS_TO_SHOW) {
1400                         drm_printf(m,
1401                                    "\t\t...skipping %d queued requests...\n",
1402                                    count - MAX_REQUESTS_TO_SHOW);
1403                 }
1404                 print_request(m, last, "\t\tQ ");
1405         }
1406
1407         spin_unlock_irq(&engine->timeline.lock);
1408
1409         spin_lock_irq(&b->rb_lock);
1410         for (rb = rb_first(&b->waiters); rb; rb = rb_next(rb)) {
1411                 struct intel_wait *w = rb_entry(rb, typeof(*w), node);
1412
1413                 drm_printf(m, "\t%s [%d] waiting for %x\n",
1414                            w->tsk->comm, w->tsk->pid, w->seqno);
1415         }
1416         spin_unlock_irq(&b->rb_lock);
1417
1418         drm_printf(m, "IRQ? 0x%lx (breadcrumbs? %s) (execlists? %s)\n",
1419                    engine->irq_posted,
1420                    yesno(test_bit(ENGINE_IRQ_BREADCRUMB,
1421                                   &engine->irq_posted)),
1422                    yesno(test_bit(ENGINE_IRQ_EXECLIST,
1423                                   &engine->irq_posted)));
1424
1425         drm_printf(m, "HWSP:\n");
1426         hexdump(m, engine->status_page.page_addr, PAGE_SIZE);
1427
1428         drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1429 }
1430
1431 static u8 user_class_map[] = {
1432         [I915_ENGINE_CLASS_RENDER] = RENDER_CLASS,
1433         [I915_ENGINE_CLASS_COPY] = COPY_ENGINE_CLASS,
1434         [I915_ENGINE_CLASS_VIDEO] = VIDEO_DECODE_CLASS,
1435         [I915_ENGINE_CLASS_VIDEO_ENHANCE] = VIDEO_ENHANCEMENT_CLASS,
1436 };
1437
1438 struct intel_engine_cs *
1439 intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance)
1440 {
1441         if (class >= ARRAY_SIZE(user_class_map))
1442                 return NULL;
1443
1444         class = user_class_map[class];
1445
1446         GEM_BUG_ON(class > MAX_ENGINE_CLASS);
1447
1448         if (instance > MAX_ENGINE_INSTANCE)
1449                 return NULL;
1450
1451         return i915->engine_class[class][instance];
1452 }
1453
1454 /**
1455  * intel_enable_engine_stats() - Enable engine busy tracking on engine
1456  * @engine: engine to enable stats collection
1457  *
1458  * Start collecting the engine busyness data for @engine.
1459  *
1460  * Returns 0 on success or a negative error code.
1461  */
1462 int intel_enable_engine_stats(struct intel_engine_cs *engine)
1463 {
1464         struct intel_engine_execlists *execlists = &engine->execlists;
1465         unsigned long flags;
1466         int err = 0;
1467
1468         if (!intel_engine_supports_stats(engine))
1469                 return -ENODEV;
1470
1471         tasklet_disable(&execlists->tasklet);
1472         write_seqlock_irqsave(&engine->stats.lock, flags);
1473
1474         if (unlikely(engine->stats.enabled == ~0)) {
1475                 err = -EBUSY;
1476                 goto unlock;
1477         }
1478
1479         if (engine->stats.enabled++ == 0) {
1480                 const struct execlist_port *port = execlists->port;
1481                 unsigned int num_ports = execlists_num_ports(execlists);
1482
1483                 engine->stats.enabled_at = ktime_get();
1484
1485                 /* XXX submission method oblivious? */
1486                 while (num_ports-- && port_isset(port)) {
1487                         engine->stats.active++;
1488                         port++;
1489                 }
1490
1491                 if (engine->stats.active)
1492                         engine->stats.start = engine->stats.enabled_at;
1493         }
1494
1495 unlock:
1496         write_sequnlock_irqrestore(&engine->stats.lock, flags);
1497         tasklet_enable(&execlists->tasklet);
1498
1499         return err;
1500 }
1501
1502 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
1503 {
1504         ktime_t total = engine->stats.total;
1505
1506         /*
1507          * If the engine is executing something at the moment
1508          * add it to the total.
1509          */
1510         if (engine->stats.active)
1511                 total = ktime_add(total,
1512                                   ktime_sub(ktime_get(), engine->stats.start));
1513
1514         return total;
1515 }
1516
1517 /**
1518  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1519  * @engine: engine to report on
1520  *
1521  * Returns accumulated time @engine was busy since engine stats were enabled.
1522  */
1523 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine)
1524 {
1525         unsigned int seq;
1526         ktime_t total;
1527
1528         do {
1529                 seq = read_seqbegin(&engine->stats.lock);
1530                 total = __intel_engine_get_busy_time(engine);
1531         } while (read_seqretry(&engine->stats.lock, seq));
1532
1533         return total;
1534 }
1535
1536 /**
1537  * intel_disable_engine_stats() - Disable engine busy tracking on engine
1538  * @engine: engine to disable stats collection
1539  *
1540  * Stops collecting the engine busyness data for @engine.
1541  */
1542 void intel_disable_engine_stats(struct intel_engine_cs *engine)
1543 {
1544         unsigned long flags;
1545
1546         if (!intel_engine_supports_stats(engine))
1547                 return;
1548
1549         write_seqlock_irqsave(&engine->stats.lock, flags);
1550         WARN_ON_ONCE(engine->stats.enabled == 0);
1551         if (--engine->stats.enabled == 0) {
1552                 engine->stats.total = __intel_engine_get_busy_time(engine);
1553                 engine->stats.active = 0;
1554         }
1555         write_sequnlock_irqrestore(&engine->stats.lock, flags);
1556 }
1557
1558 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1559 #include "selftests/mock_engine.c"
1560 #include "selftests/intel_engine_cs.c"
1561 #endif