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 "i915_drv.h"
  26 #include "intel_ringbuffer.h"
  27 #include "intel_lrc.h"
  28
  29 static const struct engine_info {
  30         const char *name;
  31         unsigned exec_id;
  32         enum intel_engine_hw_id hw_id;
  33         u32 mmio_base;
  34         unsigned irq_shift;
  35         int (*init_legacy)(struct intel_engine_cs *engine);
  36         int (*init_execlists)(struct intel_engine_cs *engine);
  37 } intel_engines[] = {
  38         [RCS] = {
  39                 .name = "render ring",
  40                 .exec_id = I915_EXEC_RENDER,
  41                 .hw_id = RCS_HW,
  42                 .mmio_base = RENDER_RING_BASE,
  43                 .irq_shift = GEN8_RCS_IRQ_SHIFT,
  44                 .init_execlists = logical_render_ring_init,
  45                 .init_legacy = intel_init_render_ring_buffer,
  46         },
  47         [BCS] = {
  48                 .name = "blitter ring",
  49                 .exec_id = I915_EXEC_BLT,
  50                 .hw_id = BCS_HW,
  51                 .mmio_base = BLT_RING_BASE,
  52                 .irq_shift = GEN8_BCS_IRQ_SHIFT,
  53                 .init_execlists = logical_xcs_ring_init,
  54                 .init_legacy = intel_init_blt_ring_buffer,
  55         },
  56         [VCS] = {
  57                 .name = "bsd ring",
  58                 .exec_id = I915_EXEC_BSD,
  59                 .hw_id = VCS_HW,
  60                 .mmio_base = GEN6_BSD_RING_BASE,
  61                 .irq_shift = GEN8_VCS1_IRQ_SHIFT,
  62                 .init_execlists = logical_xcs_ring_init,
  63                 .init_legacy = intel_init_bsd_ring_buffer,
  64         },
  65         [VCS2] = {
  66                 .name = "bsd2 ring",
  67                 .exec_id = I915_EXEC_BSD,
  68                 .hw_id = VCS2_HW,
  69                 .mmio_base = GEN8_BSD2_RING_BASE,
  70                 .irq_shift = GEN8_VCS2_IRQ_SHIFT,
  71                 .init_execlists = logical_xcs_ring_init,
  72                 .init_legacy = intel_init_bsd2_ring_buffer,
  73         },
  74         [VECS] = {
  75                 .name = "video enhancement ring",
  76                 .exec_id = I915_EXEC_VEBOX,
  77                 .hw_id = VECS_HW,
  78                 .mmio_base = VEBOX_RING_BASE,
  79                 .irq_shift = GEN8_VECS_IRQ_SHIFT,
  80                 .init_execlists = logical_xcs_ring_init,
  81                 .init_legacy = intel_init_vebox_ring_buffer,
  82         },
  83 };
  84
  85 static int
  86 intel_engine_setup(struct drm_i915_private *dev_priv,
  87                    enum intel_engine_id id)
  88 {
  89         const struct engine_info *info = &intel_engines[id];
  90         struct intel_engine_cs *engine;
  91
  92         GEM_BUG_ON(dev_priv->engine[id]);
  93         engine = kzalloc(sizeof(*engine), GFP_KERNEL);
  94         if (!engine)
  95                 return -ENOMEM;
  96
  97         engine->id = id;
  98         engine->i915 = dev_priv;
  99         engine->name = info->name;
 100         engine->exec_id = info->exec_id;
 101         engine->hw_id = engine->guc_id = info->hw_id;
 102         engine->mmio_base = info->mmio_base;
 103         engine->irq_shift = info->irq_shift;
 104
 105         /* Nothing to do here, execute in order of dependencies */
 106         engine->schedule = NULL;
 107
 108         dev_priv->engine[id] = engine;
 109         return 0;
 110 }
 111
 112 /**
 113  * intel_engines_init() - allocate, populate and init the Engine Command Streamers
 114  * @dev_priv: i915 device private
 115  *
 116  * Return: non-zero if the initialization failed.
 117  */
 118 int intel_engines_init(struct drm_i915_private *dev_priv)
 119 {
 120         struct intel_device_info *device_info = mkwrite_device_info(dev_priv);
 121         unsigned int ring_mask = INTEL_INFO(dev_priv)->ring_mask;
 122         unsigned int mask = 0;
 123         int (*init)(struct intel_engine_cs *engine);
 124         struct intel_engine_cs *engine;
 125         enum intel_engine_id id;
 126         unsigned int i;
 127         int ret;
 128
 129         WARN_ON(ring_mask == 0);
 130         WARN_ON(ring_mask &
 131                 GENMASK(sizeof(mask) * BITS_PER_BYTE - 1, I915_NUM_ENGINES));
 132
 133         for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
 134                 if (!HAS_ENGINE(dev_priv, i))
 135                         continue;
 136
 137                 if (i915.enable_execlists)
 138                         init = intel_engines[i].init_execlists;
 139                 else
 140                         init = intel_engines[i].init_legacy;
 141
 142                 if (!init)
 143                         continue;
 144
 145                 ret = intel_engine_setup(dev_priv, i);
 146                 if (ret)
 147                         goto cleanup;
 148
 149                 ret = init(dev_priv->engine[i]);
 150                 if (ret)
 151                         goto cleanup;
 152
 153                 mask |= ENGINE_MASK(i);
 154         }
 155
 156         /*
 157          * Catch failures to update intel_engines table when the new engines
 158          * are added to the driver by a warning and disabling the forgotten
 159          * engines.
 160          */
 161         if (WARN_ON(mask != ring_mask))
 162                 device_info->ring_mask = mask;
 163
 164         device_info->num_rings = hweight32(mask);
 165
 166         return 0;
 167
 168 cleanup:
 169         for_each_engine(engine, dev_priv, id) {
 170                 if (i915.enable_execlists)
 171                         intel_logical_ring_cleanup(engine);
 172                 else
 173                         intel_engine_cleanup(engine);
 174         }
 175
 176         return ret;
 177 }
 178
 179 void intel_engine_init_global_seqno(struct intel_engine_cs *engine, u32 seqno)
 180 {
 181         struct drm_i915_private *dev_priv = engine->i915;
 182
 183         /* Our semaphore implementation is strictly monotonic (i.e. we proceed
 184          * so long as the semaphore value in the register/page is greater
 185          * than the sync value), so whenever we reset the seqno,
 186          * so long as we reset the tracking semaphore value to 0, it will
 187          * always be before the next request's seqno. If we don't reset
 188          * the semaphore value, then when the seqno moves backwards all
 189          * future waits will complete instantly (causing rendering corruption).
 190          */
 191         if (IS_GEN6(dev_priv) || IS_GEN7(dev_priv)) {
 192                 I915_WRITE(RING_SYNC_0(engine->mmio_base), 0);
 193                 I915_WRITE(RING_SYNC_1(engine->mmio_base), 0);
 194                 if (HAS_VEBOX(dev_priv))
 195                         I915_WRITE(RING_SYNC_2(engine->mmio_base), 0);
 196         }
 197         if (dev_priv->semaphore) {
 198                 struct page *page = i915_vma_first_page(dev_priv->semaphore);
 199                 void *semaphores;
 200
 201                 /* Semaphores are in noncoherent memory, flush to be safe */
 202                 semaphores = kmap(page);
 203                 memset(semaphores + GEN8_SEMAPHORE_OFFSET(engine->id, 0),
 204                        0, I915_NUM_ENGINES * gen8_semaphore_seqno_size);
 205                 drm_clflush_virt_range(semaphores + GEN8_SEMAPHORE_OFFSET(engine->id, 0),
 206                                        I915_NUM_ENGINES * gen8_semaphore_seqno_size);
 207                 kunmap(page);
 208         }
 209
 210         intel_write_status_page(engine, I915_GEM_HWS_INDEX, seqno);
 211         if (engine->irq_seqno_barrier)
 212                 engine->irq_seqno_barrier(engine);
 213
 214         GEM_BUG_ON(i915_gem_active_isset(&engine->timeline->last_request));
 215         engine->timeline->last_submitted_seqno = seqno;
 216
 217         engine->hangcheck.seqno = seqno;
 218
 219         /* After manually advancing the seqno, fake the interrupt in case
 220          * there are any waiters for that seqno.
 221          */
 222         intel_engine_wakeup(engine);
 223 }
 224
 225 static void intel_engine_init_timeline(struct intel_engine_cs *engine)
 226 {
 227         engine->timeline = &engine->i915->gt.global_timeline.engine[engine->id];
 228 }
 229
 230 /**
 231  * intel_engines_setup_common - setup engine state not requiring hw access
 232  * @engine: Engine to setup.
 233  *
 234  * Initializes @engine@ structure members shared between legacy and execlists
 235  * submission modes which do not require hardware access.
 236  *
 237  * Typically done early in the submission mode specific engine setup stage.
 238  */
 239 void intel_engine_setup_common(struct intel_engine_cs *engine)
 240 {
 241         engine->execlist_queue = RB_ROOT;
 242         engine->execlist_first = NULL;
 243
 244         intel_engine_init_timeline(engine);
 245         intel_engine_init_hangcheck(engine);
 246         i915_gem_batch_pool_init(engine, &engine->batch_pool);
 247
 248         intel_engine_init_cmd_parser(engine);
 249 }
 250
 251 int intel_engine_create_scratch(struct intel_engine_cs *engine, int size)
 252 {
 253         struct drm_i915_gem_object *obj;
 254         struct i915_vma *vma;
 255         int ret;
 256
 257         WARN_ON(engine->scratch);
 258
 259         obj = i915_gem_object_create_stolen(engine->i915, size);
 260         if (!obj)
 261                 obj = i915_gem_object_create_internal(engine->i915, size);
 262         if (IS_ERR(obj)) {
 263                 DRM_ERROR("Failed to allocate scratch page\n");
 264                 return PTR_ERR(obj);
 265         }
 266
 267         vma = i915_vma_create(obj, &engine->i915->ggtt.base, NULL);
 268         if (IS_ERR(vma)) {
 269                 ret = PTR_ERR(vma);
 270                 goto err_unref;
 271         }
 272
 273         ret = i915_vma_pin(vma, 0, 4096, PIN_GLOBAL | PIN_HIGH);
 274         if (ret)
 275                 goto err_unref;
 276
 277         engine->scratch = vma;
 278         DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
 279                          engine->name, i915_ggtt_offset(vma));
 280         return 0;
 281
 282 err_unref:
 283         i915_gem_object_put(obj);
 284         return ret;
 285 }
 286
 287 static void intel_engine_cleanup_scratch(struct intel_engine_cs *engine)
 288 {
 289         i915_vma_unpin_and_release(&engine->scratch);
 290 }
 291
 292 /**
 293  * intel_engines_init_common - initialize cengine state which might require hw access
 294  * @engine: Engine to initialize.
 295  *
 296  * Initializes @engine@ structure members shared between legacy and execlists
 297  * submission modes which do require hardware access.
 298  *
 299  * Typcally done at later stages of submission mode specific engine setup.
 300  *
 301  * Returns zero on success or an error code on failure.
 302  */
 303 int intel_engine_init_common(struct intel_engine_cs *engine)
 304 {
 305         int ret;
 306
 307         /* We may need to do things with the shrinker which
 308          * require us to immediately switch back to the default
 309          * context. This can cause a problem as pinning the
 310          * default context also requires GTT space which may not
 311          * be available. To avoid this we always pin the default
 312          * context.
 313          */
 314         ret = engine->context_pin(engine, engine->i915->kernel_context);
 315         if (ret)
 316                 return ret;
 317
 318         ret = intel_engine_init_breadcrumbs(engine);
 319         if (ret)
 320                 goto err_unpin;
 321
 322         ret = i915_gem_render_state_init(engine);
 323         if (ret)
 324                 goto err_unpin;
 325
 326         return 0;
 327
 328 err_unpin:
 329         engine->context_unpin(engine, engine->i915->kernel_context);
 330         return ret;
 331 }
 332
 333 /**
 334  * intel_engines_cleanup_common - cleans up the engine state created by
 335  *                                the common initiailizers.
 336  * @engine: Engine to cleanup.
 337  *
 338  * This cleans up everything created by the common helpers.
 339  */
 340 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
 341 {
 342         intel_engine_cleanup_scratch(engine);
 343
 344         i915_gem_render_state_fini(engine);
 345         intel_engine_fini_breadcrumbs(engine);
 346         intel_engine_cleanup_cmd_parser(engine);
 347         i915_gem_batch_pool_fini(&engine->batch_pool);
 348
 349         engine->context_unpin(engine, engine->i915->kernel_context);
 350 }
 351
 352 u64 intel_engine_get_active_head(struct intel_engine_cs *engine)
 353 {
 354         struct drm_i915_private *dev_priv = engine->i915;
 355         u64 acthd;
 356
 357         if (INTEL_GEN(dev_priv) >= 8)
 358                 acthd = I915_READ64_2x32(RING_ACTHD(engine->mmio_base),
 359                                          RING_ACTHD_UDW(engine->mmio_base));
 360         else if (INTEL_GEN(dev_priv) >= 4)
 361                 acthd = I915_READ(RING_ACTHD(engine->mmio_base));
 362         else
 363                 acthd = I915_READ(ACTHD);
 364
 365         return acthd;
 366 }
 367
 368 u64 intel_engine_get_last_batch_head(struct intel_engine_cs *engine)
 369 {
 370         struct drm_i915_private *dev_priv = engine->i915;
 371         u64 bbaddr;
 372
 373         if (INTEL_GEN(dev_priv) >= 8)
 374                 bbaddr = I915_READ64_2x32(RING_BBADDR(engine->mmio_base),
 375                                           RING_BBADDR_UDW(engine->mmio_base));
 376         else
 377                 bbaddr = I915_READ(RING_BBADDR(engine->mmio_base));
 378
 379         return bbaddr;
 380 }
 381
 382 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
 383 {
 384         switch (type) {
 385         case I915_CACHE_NONE: return " uncached";
 386         case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
 387         case I915_CACHE_L3_LLC: return " L3+LLC";
 388         case I915_CACHE_WT: return " WT";
 389         default: return "";
 390         }
 391 }
 392
 393 static inline uint32_t
 394 read_subslice_reg(struct drm_i915_private *dev_priv, int slice,
 395                   int subslice, i915_reg_t reg)
 396 {
 397         uint32_t mcr;
 398         uint32_t ret;
 399         enum forcewake_domains fw_domains;
 400
 401         fw_domains = intel_uncore_forcewake_for_reg(dev_priv, reg,
 402                                                     FW_REG_READ);
 403         fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
 404                                                      GEN8_MCR_SELECTOR,
 405                                                      FW_REG_READ | FW_REG_WRITE);
 406
 407         spin_lock_irq(&dev_priv->uncore.lock);
 408         intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
 409
 410         mcr = I915_READ_FW(GEN8_MCR_SELECTOR);
 411         /*
 412          * The HW expects the slice and sublice selectors to be reset to 0
 413          * after reading out the registers.
 414          */
 415         WARN_ON_ONCE(mcr & (GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK));
 416         mcr &= ~(GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK);
 417         mcr |= GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
 418         I915_WRITE_FW(GEN8_MCR_SELECTOR, mcr);
 419
 420         ret = I915_READ_FW(reg);
 421
 422         mcr &= ~(GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK);
 423         I915_WRITE_FW(GEN8_MCR_SELECTOR, mcr);
 424
 425         intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
 426         spin_unlock_irq(&dev_priv->uncore.lock);
 427
 428         return ret;
 429 }
 430
 431 /* NB: please notice the memset */
 432 void intel_engine_get_instdone(struct intel_engine_cs *engine,
 433                                struct intel_instdone *instdone)
 434 {
 435         struct drm_i915_private *dev_priv = engine->i915;
 436         u32 mmio_base = engine->mmio_base;
 437         int slice;
 438         int subslice;
 439
 440         memset(instdone, 0, sizeof(*instdone));
 441
 442         switch (INTEL_GEN(dev_priv)) {
 443         default:
 444                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 445
 446                 if (engine->id != RCS)
 447                         break;
 448
 449                 instdone->slice_common = I915_READ(GEN7_SC_INSTDONE);
 450                 for_each_instdone_slice_subslice(dev_priv, slice, subslice) {
 451                         instdone->sampler[slice][subslice] =
 452                                 read_subslice_reg(dev_priv, slice, subslice,
 453                                                   GEN7_SAMPLER_INSTDONE);
 454                         instdone->row[slice][subslice] =
 455                                 read_subslice_reg(dev_priv, slice, subslice,
 456                                                   GEN7_ROW_INSTDONE);
 457                 }
 458                 break;
 459         case 7:
 460                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 461
 462                 if (engine->id != RCS)
 463                         break;
 464
 465                 instdone->slice_common = I915_READ(GEN7_SC_INSTDONE);
 466                 instdone->sampler[0][0] = I915_READ(GEN7_SAMPLER_INSTDONE);
 467                 instdone->row[0][0] = I915_READ(GEN7_ROW_INSTDONE);
 468
 469                 break;
 470         case 6:
 471         case 5:
 472         case 4:
 473                 instdone->instdone = I915_READ(RING_INSTDONE(mmio_base));
 474
 475                 if (engine->id == RCS)
 476                         /* HACK: Using the wrong struct member */
 477                         instdone->slice_common = I915_READ(GEN4_INSTDONE1);
 478                 break;
 479         case 3:
 480         case 2:
 481                 instdone->instdone = I915_READ(GEN2_INSTDONE);
 482                 break;
 483         }
 484 }