drivers/gpu/drm/i915/intel_lrc.c

   1 /*
   2  * Copyright © 2014 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  * Authors:
  24  *    Ben Widawsky <ben@bwidawsk.net>
  25  *    Michel Thierry <michel.thierry@intel.com>
  26  *    Thomas Daniel <thomas.daniel@intel.com>
  27  *    Oscar Mateo <oscar.mateo@intel.com>
  28  *
  29  */
  30
  31 /**
  32  * DOC: Logical Rings, Logical Ring Contexts and Execlists
  33  *
  34  * Motivation:
  35  * GEN8 brings an expansion of the HW contexts: "Logical Ring Contexts".
  36  * These expanded contexts enable a number of new abilities, especially
  37  * "Execlists" (also implemented in this file).
  38  *
  39  * One of the main differences with the legacy HW contexts is that logical
  40  * ring contexts incorporate many more things to the context's state, like
  41  * PDPs or ringbuffer control registers:
  42  *
  43  * The reason why PDPs are included in the context is straightforward: as
  44  * PPGTTs (per-process GTTs) are actually per-context, having the PDPs
  45  * contained there mean you don't need to do a ppgtt->switch_mm yourself,
  46  * instead, the GPU will do it for you on the context switch.
  47  *
  48  * But, what about the ringbuffer control registers (head, tail, etc..)?
  49  * shouldn't we just need a set of those per engine command streamer? This is
  50  * where the name "Logical Rings" starts to make sense: by virtualizing the
  51  * rings, the engine cs shifts to a new "ring buffer" with every context
  52  * switch. When you want to submit a workload to the GPU you: A) choose your
  53  * context, B) find its appropriate virtualized ring, C) write commands to it
  54  * and then, finally, D) tell the GPU to switch to that context.
  55  *
  56  * Instead of the legacy MI_SET_CONTEXT, the way you tell the GPU to switch
  57  * to a contexts is via a context execution list, ergo "Execlists".
  58  *
  59  * LRC implementation:
  60  * Regarding the creation of contexts, we have:
  61  *
  62  * - One global default context.
  63  * - One local default context for each opened fd.
  64  * - One local extra context for each context create ioctl call.
  65  *
  66  * Now that ringbuffers belong per-context (and not per-engine, like before)
  67  * and that contexts are uniquely tied to a given engine (and not reusable,
  68  * like before) we need:
  69  *
  70  * - One ringbuffer per-engine inside each context.
  71  * - One backing object per-engine inside each context.
  72  *
  73  * The global default context starts its life with these new objects fully
  74  * allocated and populated. The local default context for each opened fd is
  75  * more complex, because we don't know at creation time which engine is going
  76  * to use them. To handle this, we have implemented a deferred creation of LR
  77  * contexts:
  78  *
  79  * The local context starts its life as a hollow or blank holder, that only
  80  * gets populated for a given engine once we receive an execbuffer. If later
  81  * on we receive another execbuffer ioctl for the same context but a different
  82  * engine, we allocate/populate a new ringbuffer and context backing object and
  83  * so on.
  84  *
  85  * Finally, regarding local contexts created using the ioctl call: as they are
  86  * only allowed with the render ring, we can allocate & populate them right
  87  * away (no need to defer anything, at least for now).
  88  *
  89  * Execlists implementation:
  90  * Execlists are the new method by which, on gen8+ hardware, workloads are
  91  * submitted for execution (as opposed to the legacy, ringbuffer-based, method).
  92  * This method works as follows:
  93  *
  94  * When a request is committed, its commands (the BB start and any leading or
  95  * trailing commands, like the seqno breadcrumbs) are placed in the ringbuffer
  96  * for the appropriate context. The tail pointer in the hardware context is not
  97  * updated at this time, but instead, kept by the driver in the ringbuffer
  98  * structure. A structure representing this request is added to a request queue
  99  * for the appropriate engine: this structure contains a copy of the context's
 100  * tail after the request was written to the ring buffer and a pointer to the
 101  * context itself.
 102  *
 103  * If the engine's request queue was empty before the request was added, the
 104  * queue is processed immediately. Otherwise the queue will be processed during
 105  * a context switch interrupt. In any case, elements on the queue will get sent
 106  * (in pairs) to the GPU's ExecLists Submit Port (ELSP, for short) with a
 107  * globally unique 20-bits submission ID.
 108  *
 109  * When execution of a request completes, the GPU updates the context status
 110  * buffer with a context complete event and generates a context switch interrupt.
 111  * During the interrupt handling, the driver examines the events in the buffer:
 112  * for each context complete event, if the announced ID matches that on the head
 113  * of the request queue, then that request is retired and removed from the queue.
 114  *
 115  * After processing, if any requests were retired and the queue is not empty
 116  * then a new execution list can be submitted. The two requests at the front of
 117  * the queue are next to be submitted but since a context may not occur twice in
 118  * an execution list, if subsequent requests have the same ID as the first then
 119  * the two requests must be combined. This is done simply by discarding requests
 120  * at the head of the queue until either only one requests is left (in which case
 121  * we use a NULL second context) or the first two requests have unique IDs.
 122  *
 123  * By always executing the first two requests in the queue the driver ensures
 124  * that the GPU is kept as busy as possible. In the case where a single context
 125  * completes but a second context is still executing, the request for this second
 126  * context will be at the head of the queue when we remove the first one. This
 127  * request will then be resubmitted along with a new request for a different context,
 128  * which will cause the hardware to continue executing the second request and queue
 129  * the new request (the GPU detects the condition of a context getting preempted
 130  * with the same context and optimizes the context switch flow by not doing
 131  * preemption, but just sampling the new tail pointer).
 132  *
 133  */
 134
 135 #include <drm/drmP.h>
 136 #include <drm/i915_drm.h>
 137 #include "i915_drv.h"
 138
 139 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
 140 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
 141
 142 #define GEN8_LR_CONTEXT_ALIGN 4096
 143
 144 #define RING_EXECLIST_QFULL             (1 << 0x2)
 145 #define RING_EXECLIST1_VALID            (1 << 0x3)
 146 #define RING_EXECLIST0_VALID            (1 << 0x4)
 147 #define RING_EXECLIST_ACTIVE_STATUS     (3 << 0xE)
 148 #define RING_EXECLIST1_ACTIVE           (1 << 0x11)
 149 #define RING_EXECLIST0_ACTIVE           (1 << 0x12)
 150
 151 #define GEN8_CTX_STATUS_IDLE_ACTIVE     (1 << 0)
 152 #define GEN8_CTX_STATUS_PREEMPTED       (1 << 1)
 153 #define GEN8_CTX_STATUS_ELEMENT_SWITCH  (1 << 2)
 154 #define GEN8_CTX_STATUS_ACTIVE_IDLE     (1 << 3)
 155 #define GEN8_CTX_STATUS_COMPLETE        (1 << 4)
 156 #define GEN8_CTX_STATUS_LITE_RESTORE    (1 << 15)
 157
 158 #define CTX_LRI_HEADER_0                0x01
 159 #define CTX_CONTEXT_CONTROL             0x02
 160 #define CTX_RING_HEAD                   0x04
 161 #define CTX_RING_TAIL                   0x06
 162 #define CTX_RING_BUFFER_START           0x08
 163 #define CTX_RING_BUFFER_CONTROL         0x0a
 164 #define CTX_BB_HEAD_U                   0x0c
 165 #define CTX_BB_HEAD_L                   0x0e
 166 #define CTX_BB_STATE                    0x10
 167 #define CTX_SECOND_BB_HEAD_U            0x12
 168 #define CTX_SECOND_BB_HEAD_L            0x14
 169 #define CTX_SECOND_BB_STATE             0x16
 170 #define CTX_BB_PER_CTX_PTR              0x18
 171 #define CTX_RCS_INDIRECT_CTX            0x1a
 172 #define CTX_RCS_INDIRECT_CTX_OFFSET     0x1c
 173 #define CTX_LRI_HEADER_1                0x21
 174 #define CTX_CTX_TIMESTAMP               0x22
 175 #define CTX_PDP3_UDW                    0x24
 176 #define CTX_PDP3_LDW                    0x26
 177 #define CTX_PDP2_UDW                    0x28
 178 #define CTX_PDP2_LDW                    0x2a
 179 #define CTX_PDP1_UDW                    0x2c
 180 #define CTX_PDP1_LDW                    0x2e
 181 #define CTX_PDP0_UDW                    0x30
 182 #define CTX_PDP0_LDW                    0x32
 183 #define CTX_LRI_HEADER_2                0x41
 184 #define CTX_R_PWR_CLK_STATE             0x42
 185 #define CTX_GPGPU_CSR_BASE_ADDRESS      0x44
 186
 187 #define GEN8_CTX_VALID (1<<0)
 188 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
 189 #define GEN8_CTX_FORCE_RESTORE (1<<2)
 190 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
 191 #define GEN8_CTX_PRIVILEGE (1<<8)
 192 enum {
 193         ADVANCED_CONTEXT = 0,
 194         LEGACY_CONTEXT,
 195         ADVANCED_AD_CONTEXT,
 196         LEGACY_64B_CONTEXT
 197 };
 198 #define GEN8_CTX_MODE_SHIFT 3
 199 enum {
 200         FAULT_AND_HANG = 0,
 201         FAULT_AND_HALT, /* Debug only */
 202         FAULT_AND_STREAM,
 203         FAULT_AND_CONTINUE /* Unsupported */
 204 };
 205 #define GEN8_CTX_ID_SHIFT 32
 206
 207 /**
 208  * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
 209  * @dev: DRM device.
 210  * @enable_execlists: value of i915.enable_execlists module parameter.
 211  *
 212  * Only certain platforms support Execlists (the prerequisites being
 213  * support for Logical Ring Contexts and Aliasing PPGTT or better),
 214  * and only when enabled via module parameter.
 215  *
 216  * Return: 1 if Execlists is supported and has to be enabled.
 217  */
 218 int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists)
 219 {
 220         WARN_ON(i915.enable_ppgtt == -1);
 221
 222         if (enable_execlists == 0)
 223                 return 0;
 224
 225         if (HAS_LOGICAL_RING_CONTEXTS(dev) && USES_PPGTT(dev) &&
 226             i915.use_mmio_flip >= 0)
 227                 return 1;
 228
 229         return 0;
 230 }
 231
 232 /**
 233  * intel_execlists_ctx_id() - get the Execlists Context ID
 234  * @ctx_obj: Logical Ring Context backing object.
 235  *
 236  * Do not confuse with ctx->id! Unfortunately we have a name overload
 237  * here: the old context ID we pass to userspace as a handler so that
 238  * they can refer to a context, and the new context ID we pass to the
 239  * ELSP so that the GPU can inform us of the context status via
 240  * interrupts.
 241  *
 242  * Return: 20-bits globally unique context ID.
 243  */
 244 u32 intel_execlists_ctx_id(struct drm_i915_gem_object *ctx_obj)
 245 {
 246         u32 lrca = i915_gem_obj_ggtt_offset(ctx_obj);
 247
 248         /* LRCA is required to be 4K aligned so the more significant 20 bits
 249          * are globally unique */
 250         return lrca >> 12;
 251 }
 252
 253 static uint64_t execlists_ctx_descriptor(struct drm_i915_gem_object *ctx_obj)
 254 {
 255         uint64_t desc;
 256         uint64_t lrca = i915_gem_obj_ggtt_offset(ctx_obj);
 257
 258         WARN_ON(lrca & 0xFFFFFFFF00000FFFULL);
 259
 260         desc = GEN8_CTX_VALID;
 261         desc |= LEGACY_CONTEXT << GEN8_CTX_MODE_SHIFT;
 262         desc |= GEN8_CTX_L3LLC_COHERENT;
 263         desc |= GEN8_CTX_PRIVILEGE;
 264         desc |= lrca;
 265         desc |= (u64)intel_execlists_ctx_id(ctx_obj) << GEN8_CTX_ID_SHIFT;
 266
 267         /* TODO: WaDisableLiteRestore when we start using semaphore
 268          * signalling between Command Streamers */
 269         /* desc |= GEN8_CTX_FORCE_RESTORE; */
 270
 271         return desc;
 272 }
 273
 274 static void execlists_elsp_write(struct intel_engine_cs *ring,
 275                                  struct drm_i915_gem_object *ctx_obj0,
 276                                  struct drm_i915_gem_object *ctx_obj1)
 277 {
 278         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 279         uint64_t temp = 0;
 280         uint32_t desc[4];
 281         unsigned long flags;
 282
 283         /* XXX: You must always write both descriptors in the order below. */
 284         if (ctx_obj1)
 285                 temp = execlists_ctx_descriptor(ctx_obj1);
 286         else
 287                 temp = 0;
 288         desc[1] = (u32)(temp >> 32);
 289         desc[0] = (u32)temp;
 290
 291         temp = execlists_ctx_descriptor(ctx_obj0);
 292         desc[3] = (u32)(temp >> 32);
 293         desc[2] = (u32)temp;
 294
 295         /* Set Force Wakeup bit to prevent GT from entering C6 while ELSP writes
 296          * are in progress.
 297          *
 298          * The other problem is that we can't just call gen6_gt_force_wake_get()
 299          * because that function calls intel_runtime_pm_get(), which might sleep.
 300          * Instead, we do the runtime_pm_get/put when creating/destroying requests.
 301          */
 302         spin_lock_irqsave(&dev_priv->uncore.lock, flags);
 303         if (IS_CHERRYVIEW(dev_priv->dev)) {
 304                 if (dev_priv->uncore.fw_rendercount++ == 0)
 305                         dev_priv->uncore.funcs.force_wake_get(dev_priv,
 306                                                               FORCEWAKE_RENDER);
 307                 if (dev_priv->uncore.fw_mediacount++ == 0)
 308                         dev_priv->uncore.funcs.force_wake_get(dev_priv,
 309                                                               FORCEWAKE_MEDIA);
 310         } else {
 311                 if (dev_priv->uncore.forcewake_count++ == 0)
 312                         dev_priv->uncore.funcs.force_wake_get(dev_priv,
 313                                                               FORCEWAKE_ALL);
 314         }
 315         spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 316
 317         I915_WRITE(RING_ELSP(ring), desc[1]);
 318         I915_WRITE(RING_ELSP(ring), desc[0]);
 319         I915_WRITE(RING_ELSP(ring), desc[3]);
 320         /* The context is automatically loaded after the following */
 321         I915_WRITE(RING_ELSP(ring), desc[2]);
 322
 323         /* ELSP is a wo register, so use another nearby reg for posting instead */
 324         POSTING_READ(RING_EXECLIST_STATUS(ring));
 325
 326         /* Release Force Wakeup (see the big comment above). */
 327         spin_lock_irqsave(&dev_priv->uncore.lock, flags);
 328         if (IS_CHERRYVIEW(dev_priv->dev)) {
 329                 if (--dev_priv->uncore.fw_rendercount == 0)
 330                         dev_priv->uncore.funcs.force_wake_put(dev_priv,
 331                                                               FORCEWAKE_RENDER);
 332                 if (--dev_priv->uncore.fw_mediacount == 0)
 333                         dev_priv->uncore.funcs.force_wake_put(dev_priv,
 334                                                               FORCEWAKE_MEDIA);
 335         } else {
 336                 if (--dev_priv->uncore.forcewake_count == 0)
 337                         dev_priv->uncore.funcs.force_wake_put(dev_priv,
 338                                                               FORCEWAKE_ALL);
 339         }
 340
 341         spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 342 }
 343
 344 static int execlists_ctx_write_tail(struct drm_i915_gem_object *ctx_obj, u32 tail)
 345 {
 346         struct page *page;
 347         uint32_t *reg_state;
 348
 349         page = i915_gem_object_get_page(ctx_obj, 1);
 350         reg_state = kmap_atomic(page);
 351
 352         reg_state[CTX_RING_TAIL+1] = tail;
 353
 354         kunmap_atomic(reg_state);
 355
 356         return 0;
 357 }
 358
 359 static void execlists_submit_contexts(struct intel_engine_cs *ring,
 360                                       struct intel_context *to0, u32 tail0,
 361                                       struct intel_context *to1, u32 tail1)
 362 {
 363         struct drm_i915_gem_object *ctx_obj0;
 364         struct drm_i915_gem_object *ctx_obj1 = NULL;
 365
 366         ctx_obj0 = to0->engine[ring->id].state;
 367         BUG_ON(!ctx_obj0);
 368         WARN_ON(!i915_gem_obj_is_pinned(ctx_obj0));
 369
 370         execlists_ctx_write_tail(ctx_obj0, tail0);
 371
 372         if (to1) {
 373                 ctx_obj1 = to1->engine[ring->id].state;
 374                 BUG_ON(!ctx_obj1);
 375                 WARN_ON(!i915_gem_obj_is_pinned(ctx_obj1));
 376
 377                 execlists_ctx_write_tail(ctx_obj1, tail1);
 378         }
 379
 380         execlists_elsp_write(ring, ctx_obj0, ctx_obj1);
 381 }
 382
 383 static void execlists_context_unqueue(struct intel_engine_cs *ring)
 384 {
 385         struct intel_ctx_submit_request *req0 = NULL, *req1 = NULL;
 386         struct intel_ctx_submit_request *cursor = NULL, *tmp = NULL;
 387         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 388
 389         assert_spin_locked(&ring->execlist_lock);
 390
 391         if (list_empty(&ring->execlist_queue))
 392                 return;
 393
 394         /* Try to read in pairs */
 395         list_for_each_entry_safe(cursor, tmp, &ring->execlist_queue,
 396                                  execlist_link) {
 397                 if (!req0) {
 398                         req0 = cursor;
 399                 } else if (req0->ctx == cursor->ctx) {
 400                         /* Same ctx: ignore first request, as second request
 401                          * will update tail past first request's workload */
 402                         cursor->elsp_submitted = req0->elsp_submitted;
 403                         list_del(&req0->execlist_link);
 404                         queue_work(dev_priv->wq, &req0->work);
 405                         req0 = cursor;
 406                 } else {
 407                         req1 = cursor;
 408                         break;
 409                 }
 410         }
 411
 412         WARN_ON(req1 && req1->elsp_submitted);
 413
 414         execlists_submit_contexts(ring, req0->ctx, req0->tail,
 415                                   req1 ? req1->ctx : NULL,
 416                                   req1 ? req1->tail : 0);
 417
 418         req0->elsp_submitted++;
 419         if (req1)
 420                 req1->elsp_submitted++;
 421 }
 422
 423 static bool execlists_check_remove_request(struct intel_engine_cs *ring,
 424                                            u32 request_id)
 425 {
 426         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 427         struct intel_ctx_submit_request *head_req;
 428
 429         assert_spin_locked(&ring->execlist_lock);
 430
 431         head_req = list_first_entry_or_null(&ring->execlist_queue,
 432                                             struct intel_ctx_submit_request,
 433                                             execlist_link);
 434
 435         if (head_req != NULL) {
 436                 struct drm_i915_gem_object *ctx_obj =
 437                                 head_req->ctx->engine[ring->id].state;
 438                 if (intel_execlists_ctx_id(ctx_obj) == request_id) {
 439                         WARN(head_req->elsp_submitted == 0,
 440                              "Never submitted head request\n");
 441
 442                         if (--head_req->elsp_submitted <= 0) {
 443                                 list_del(&head_req->execlist_link);
 444                                 queue_work(dev_priv->wq, &head_req->work);
 445                                 return true;
 446                         }
 447                 }
 448         }
 449
 450         return false;
 451 }
 452
 453 /**
 454  * intel_execlists_handle_ctx_events() - handle Context Switch interrupts
 455  * @ring: Engine Command Streamer to handle.
 456  *
 457  * Check the unread Context Status Buffers and manage the submission of new
 458  * contexts to the ELSP accordingly.
 459  */
 460 void intel_execlists_handle_ctx_events(struct intel_engine_cs *ring)
 461 {
 462         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 463         u32 status_pointer;
 464         u8 read_pointer;
 465         u8 write_pointer;
 466         u32 status;
 467         u32 status_id;
 468         u32 submit_contexts = 0;
 469
 470         status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring));
 471
 472         read_pointer = ring->next_context_status_buffer;
 473         write_pointer = status_pointer & 0x07;
 474         if (read_pointer > write_pointer)
 475                 write_pointer += 6;
 476
 477         spin_lock(&ring->execlist_lock);
 478
 479         while (read_pointer < write_pointer) {
 480                 read_pointer++;
 481                 status = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
 482                                 (read_pointer % 6) * 8);
 483                 status_id = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
 484                                 (read_pointer % 6) * 8 + 4);
 485
 486                 if (status & GEN8_CTX_STATUS_PREEMPTED) {
 487                         if (status & GEN8_CTX_STATUS_LITE_RESTORE) {
 488                                 if (execlists_check_remove_request(ring, status_id))
 489                                         WARN(1, "Lite Restored request removed from queue\n");
 490                         } else
 491                                 WARN(1, "Preemption without Lite Restore\n");
 492                 }
 493
 494                  if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) ||
 495                      (status & GEN8_CTX_STATUS_ELEMENT_SWITCH)) {
 496                         if (execlists_check_remove_request(ring, status_id))
 497                                 submit_contexts++;
 498                 }
 499         }
 500
 501         if (submit_contexts != 0)
 502                 execlists_context_unqueue(ring);
 503
 504         spin_unlock(&ring->execlist_lock);
 505
 506         WARN(submit_contexts > 2, "More than two context complete events?\n");
 507         ring->next_context_status_buffer = write_pointer % 6;
 508
 509         I915_WRITE(RING_CONTEXT_STATUS_PTR(ring),
 510                    ((u32)ring->next_context_status_buffer & 0x07) << 8);
 511 }
 512
 513 static void execlists_free_request_task(struct work_struct *work)
 514 {
 515         struct intel_ctx_submit_request *req =
 516                 container_of(work, struct intel_ctx_submit_request, work);
 517         struct drm_device *dev = req->ring->dev;
 518         struct drm_i915_private *dev_priv = dev->dev_private;
 519
 520         intel_runtime_pm_put(dev_priv);
 521
 522         mutex_lock(&dev->struct_mutex);
 523         i915_gem_context_unreference(req->ctx);
 524         mutex_unlock(&dev->struct_mutex);
 525
 526         kfree(req);
 527 }
 528
 529 static int execlists_context_queue(struct intel_engine_cs *ring,
 530                                    struct intel_context *to,
 531                                    u32 tail)
 532 {
 533         struct intel_ctx_submit_request *req = NULL, *cursor;
 534         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 535         unsigned long flags;
 536         int num_elements = 0;
 537
 538         req = kzalloc(sizeof(*req), GFP_KERNEL);
 539         if (req == NULL)
 540                 return -ENOMEM;
 541         req->ctx = to;
 542         i915_gem_context_reference(req->ctx);
 543         req->ring = ring;
 544         req->tail = tail;
 545         INIT_WORK(&req->work, execlists_free_request_task);
 546
 547         intel_runtime_pm_get(dev_priv);
 548
 549         spin_lock_irqsave(&ring->execlist_lock, flags);
 550
 551         list_for_each_entry(cursor, &ring->execlist_queue, execlist_link)
 552                 if (++num_elements > 2)
 553                         break;
 554
 555         if (num_elements > 2) {
 556                 struct intel_ctx_submit_request *tail_req;
 557
 558                 tail_req = list_last_entry(&ring->execlist_queue,
 559                                            struct intel_ctx_submit_request,
 560                                            execlist_link);
 561
 562                 if (to == tail_req->ctx) {
 563                         WARN(tail_req->elsp_submitted != 0,
 564                              "More than 2 already-submitted reqs queued\n");
 565                         list_del(&tail_req->execlist_link);
 566                         queue_work(dev_priv->wq, &tail_req->work);
 567                 }
 568         }
 569
 570         list_add_tail(&req->execlist_link, &ring->execlist_queue);
 571         if (num_elements == 0)
 572                 execlists_context_unqueue(ring);
 573
 574         spin_unlock_irqrestore(&ring->execlist_lock, flags);
 575
 576         return 0;
 577 }
 578
 579 static int logical_ring_invalidate_all_caches(struct intel_ringbuffer *ringbuf)
 580 {
 581         struct intel_engine_cs *ring = ringbuf->ring;
 582         uint32_t flush_domains;
 583         int ret;
 584
 585         flush_domains = 0;
 586         if (ring->gpu_caches_dirty)
 587                 flush_domains = I915_GEM_GPU_DOMAINS;
 588
 589         ret = ring->emit_flush(ringbuf, I915_GEM_GPU_DOMAINS, flush_domains);
 590         if (ret)
 591                 return ret;
 592
 593         ring->gpu_caches_dirty = false;
 594         return 0;
 595 }
 596
 597 static int execlists_move_to_gpu(struct intel_ringbuffer *ringbuf,
 598                                  struct list_head *vmas)
 599 {
 600         struct intel_engine_cs *ring = ringbuf->ring;
 601         struct i915_vma *vma;
 602         uint32_t flush_domains = 0;
 603         bool flush_chipset = false;
 604         int ret;
 605
 606         list_for_each_entry(vma, vmas, exec_list) {
 607                 struct drm_i915_gem_object *obj = vma->obj;
 608
 609                 ret = i915_gem_object_sync(obj, ring);
 610                 if (ret)
 611                         return ret;
 612
 613                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 614                         flush_chipset |= i915_gem_clflush_object(obj, false);
 615
 616                 flush_domains |= obj->base.write_domain;
 617         }
 618
 619         if (flush_domains & I915_GEM_DOMAIN_GTT)
 620                 wmb();
 621
 622         /* Unconditionally invalidate gpu caches and ensure that we do flush
 623          * any residual writes from the previous batch.
 624          */
 625         return logical_ring_invalidate_all_caches(ringbuf);
 626 }
 627
 628 /**
 629  * execlists_submission() - submit a batchbuffer for execution, Execlists style
 630  * @dev: DRM device.
 631  * @file: DRM file.
 632  * @ring: Engine Command Streamer to submit to.
 633  * @ctx: Context to employ for this submission.
 634  * @args: execbuffer call arguments.
 635  * @vmas: list of vmas.
 636  * @batch_obj: the batchbuffer to submit.
 637  * @exec_start: batchbuffer start virtual address pointer.
 638  * @flags: translated execbuffer call flags.
 639  *
 640  * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
 641  * away the submission details of the execbuffer ioctl call.
 642  *
 643  * Return: non-zero if the submission fails.
 644  */
 645 int intel_execlists_submission(struct drm_device *dev, struct drm_file *file,
 646                                struct intel_engine_cs *ring,
 647                                struct intel_context *ctx,
 648                                struct drm_i915_gem_execbuffer2 *args,
 649                                struct list_head *vmas,
 650                                struct drm_i915_gem_object *batch_obj,
 651                                u64 exec_start, u32 flags)
 652 {
 653         struct drm_i915_private *dev_priv = dev->dev_private;
 654         struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
 655         int instp_mode;
 656         u32 instp_mask;
 657         int ret;
 658
 659         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
 660         instp_mask = I915_EXEC_CONSTANTS_MASK;
 661         switch (instp_mode) {
 662         case I915_EXEC_CONSTANTS_REL_GENERAL:
 663         case I915_EXEC_CONSTANTS_ABSOLUTE:
 664         case I915_EXEC_CONSTANTS_REL_SURFACE:
 665                 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
 666                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
 667                         return -EINVAL;
 668                 }
 669
 670                 if (instp_mode != dev_priv->relative_constants_mode) {
 671                         if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
 672                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
 673                                 return -EINVAL;
 674                         }
 675
 676                         /* The HW changed the meaning on this bit on gen6 */
 677                         instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
 678                 }
 679                 break;
 680         default:
 681                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
 682                 return -EINVAL;
 683         }
 684
 685         if (args->num_cliprects != 0) {
 686                 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
 687                 return -EINVAL;
 688         } else {
 689                 if (args->DR4 == 0xffffffff) {
 690                         DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
 691                         args->DR4 = 0;
 692                 }
 693
 694                 if (args->DR1 || args->DR4 || args->cliprects_ptr) {
 695                         DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
 696                         return -EINVAL;
 697                 }
 698         }
 699
 700         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
 701                 DRM_DEBUG("sol reset is gen7 only\n");
 702                 return -EINVAL;
 703         }
 704
 705         ret = execlists_move_to_gpu(ringbuf, vmas);
 706         if (ret)
 707                 return ret;
 708
 709         if (ring == &dev_priv->ring[RCS] &&
 710             instp_mode != dev_priv->relative_constants_mode) {
 711                 ret = intel_logical_ring_begin(ringbuf, 4);
 712                 if (ret)
 713                         return ret;
 714
 715                 intel_logical_ring_emit(ringbuf, MI_NOOP);
 716                 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
 717                 intel_logical_ring_emit(ringbuf, INSTPM);
 718                 intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
 719                 intel_logical_ring_advance(ringbuf);
 720
 721                 dev_priv->relative_constants_mode = instp_mode;
 722         }
 723
 724         ret = ring->emit_bb_start(ringbuf, exec_start, flags);
 725         if (ret)
 726                 return ret;
 727
 728         i915_gem_execbuffer_move_to_active(vmas, ring);
 729         i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
 730
 731         return 0;
 732 }
 733
 734 void intel_logical_ring_stop(struct intel_engine_cs *ring)
 735 {
 736         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 737         int ret;
 738
 739         if (!intel_ring_initialized(ring))
 740                 return;
 741
 742         ret = intel_ring_idle(ring);
 743         if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error))
 744                 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
 745                           ring->name, ret);
 746
 747         /* TODO: Is this correct with Execlists enabled? */
 748         I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING));
 749         if (wait_for_atomic((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) {
 750                 DRM_ERROR("%s :timed out trying to stop ring\n", ring->name);
 751                 return;
 752         }
 753         I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING));
 754 }
 755
 756 int logical_ring_flush_all_caches(struct intel_ringbuffer *ringbuf)
 757 {
 758         struct intel_engine_cs *ring = ringbuf->ring;
 759         int ret;
 760
 761         if (!ring->gpu_caches_dirty)
 762                 return 0;
 763
 764         ret = ring->emit_flush(ringbuf, 0, I915_GEM_GPU_DOMAINS);
 765         if (ret)
 766                 return ret;
 767
 768         ring->gpu_caches_dirty = false;
 769         return 0;
 770 }
 771
 772 /**
 773  * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
 774  * @ringbuf: Logical Ringbuffer to advance.
 775  *
 776  * The tail is updated in our logical ringbuffer struct, not in the actual context. What
 777  * really happens during submission is that the context and current tail will be placed
 778  * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
 779  * point, the tail *inside* the context is updated and the ELSP written to.
 780  */
 781 void intel_logical_ring_advance_and_submit(struct intel_ringbuffer *ringbuf)
 782 {
 783         struct intel_engine_cs *ring = ringbuf->ring;
 784         struct intel_context *ctx = ringbuf->FIXME_lrc_ctx;
 785
 786         intel_logical_ring_advance(ringbuf);
 787
 788         if (intel_ring_stopped(ring))
 789                 return;
 790
 791         execlists_context_queue(ring, ctx, ringbuf->tail);
 792 }
 793
 794 static int logical_ring_alloc_seqno(struct intel_engine_cs *ring,
 795                                     struct intel_context *ctx)
 796 {
 797         if (ring->outstanding_lazy_seqno)
 798                 return 0;
 799
 800         if (ring->preallocated_lazy_request == NULL) {
 801                 struct drm_i915_gem_request *request;
 802
 803                 request = kmalloc(sizeof(*request), GFP_KERNEL);
 804                 if (request == NULL)
 805                         return -ENOMEM;
 806
 807                 /* Hold a reference to the context this request belongs to
 808                  * (we will need it when the time comes to emit/retire the
 809                  * request).
 810                  */
 811                 request->ctx = ctx;
 812                 i915_gem_context_reference(request->ctx);
 813
 814                 ring->preallocated_lazy_request = request;
 815         }
 816
 817         return i915_gem_get_seqno(ring->dev, &ring->outstanding_lazy_seqno);
 818 }
 819
 820 static int logical_ring_wait_request(struct intel_ringbuffer *ringbuf,
 821                                      int bytes)
 822 {
 823         struct intel_engine_cs *ring = ringbuf->ring;
 824         struct drm_i915_gem_request *request;
 825         u32 seqno = 0;
 826         int ret;
 827
 828         if (ringbuf->last_retired_head != -1) {
 829                 ringbuf->head = ringbuf->last_retired_head;
 830                 ringbuf->last_retired_head = -1;
 831
 832                 ringbuf->space = intel_ring_space(ringbuf);
 833                 if (ringbuf->space >= bytes)
 834                         return 0;
 835         }
 836
 837         list_for_each_entry(request, &ring->request_list, list) {
 838                 if (__intel_ring_space(request->tail, ringbuf->tail,
 839                                        ringbuf->size) >= bytes) {
 840                         seqno = request->seqno;
 841                         break;
 842                 }
 843         }
 844
 845         if (seqno == 0)
 846                 return -ENOSPC;
 847
 848         ret = i915_wait_seqno(ring, seqno);
 849         if (ret)
 850                 return ret;
 851
 852         i915_gem_retire_requests_ring(ring);
 853         ringbuf->head = ringbuf->last_retired_head;
 854         ringbuf->last_retired_head = -1;
 855
 856         ringbuf->space = intel_ring_space(ringbuf);
 857         return 0;
 858 }
 859
 860 static int logical_ring_wait_for_space(struct intel_ringbuffer *ringbuf,
 861                                        int bytes)
 862 {
 863         struct intel_engine_cs *ring = ringbuf->ring;
 864         struct drm_device *dev = ring->dev;
 865         struct drm_i915_private *dev_priv = dev->dev_private;
 866         unsigned long end;
 867         int ret;
 868
 869         ret = logical_ring_wait_request(ringbuf, bytes);
 870         if (ret != -ENOSPC)
 871                 return ret;
 872
 873         /* Force the context submission in case we have been skipping it */
 874         intel_logical_ring_advance_and_submit(ringbuf);
 875
 876         /* With GEM the hangcheck timer should kick us out of the loop,
 877          * leaving it early runs the risk of corrupting GEM state (due
 878          * to running on almost untested codepaths). But on resume
 879          * timers don't work yet, so prevent a complete hang in that
 880          * case by choosing an insanely large timeout. */
 881         end = jiffies + 60 * HZ;
 882
 883         do {
 884                 ringbuf->head = I915_READ_HEAD(ring);
 885                 ringbuf->space = intel_ring_space(ringbuf);
 886                 if (ringbuf->space >= bytes) {
 887                         ret = 0;
 888                         break;
 889                 }
 890
 891                 msleep(1);
 892
 893                 if (dev_priv->mm.interruptible && signal_pending(current)) {
 894                         ret = -ERESTARTSYS;
 895                         break;
 896                 }
 897
 898                 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
 899                                            dev_priv->mm.interruptible);
 900                 if (ret)
 901                         break;
 902
 903                 if (time_after(jiffies, end)) {
 904                         ret = -EBUSY;
 905                         break;
 906                 }
 907         } while (1);
 908
 909         return ret;
 910 }
 911
 912 static int logical_ring_wrap_buffer(struct intel_ringbuffer *ringbuf)
 913 {
 914         uint32_t __iomem *virt;
 915         int rem = ringbuf->size - ringbuf->tail;
 916
 917         if (ringbuf->space < rem) {
 918                 int ret = logical_ring_wait_for_space(ringbuf, rem);
 919
 920                 if (ret)
 921                         return ret;
 922         }
 923
 924         virt = ringbuf->virtual_start + ringbuf->tail;
 925         rem /= 4;
 926         while (rem--)
 927                 iowrite32(MI_NOOP, virt++);
 928
 929         ringbuf->tail = 0;
 930         ringbuf->space = intel_ring_space(ringbuf);
 931
 932         return 0;
 933 }
 934
 935 static int logical_ring_prepare(struct intel_ringbuffer *ringbuf, int bytes)
 936 {
 937         int ret;
 938
 939         if (unlikely(ringbuf->tail + bytes > ringbuf->effective_size)) {
 940                 ret = logical_ring_wrap_buffer(ringbuf);
 941                 if (unlikely(ret))
 942                         return ret;
 943         }
 944
 945         if (unlikely(ringbuf->space < bytes)) {
 946                 ret = logical_ring_wait_for_space(ringbuf, bytes);
 947                 if (unlikely(ret))
 948                         return ret;
 949         }
 950
 951         return 0;
 952 }
 953
 954 /**
 955  * intel_logical_ring_begin() - prepare the logical ringbuffer to accept some commands
 956  *
 957  * @ringbuf: Logical ringbuffer.
 958  * @num_dwords: number of DWORDs that we plan to write to the ringbuffer.
 959  *
 960  * The ringbuffer might not be ready to accept the commands right away (maybe it needs to
 961  * be wrapped, or wait a bit for the tail to be updated). This function takes care of that
 962  * and also preallocates a request (every workload submission is still mediated through
 963  * requests, same as it did with legacy ringbuffer submission).
 964  *
 965  * Return: non-zero if the ringbuffer is not ready to be written to.
 966  */
 967 int intel_logical_ring_begin(struct intel_ringbuffer *ringbuf, int num_dwords)
 968 {
 969         struct intel_engine_cs *ring = ringbuf->ring;
 970         struct drm_device *dev = ring->dev;
 971         struct drm_i915_private *dev_priv = dev->dev_private;
 972         int ret;
 973
 974         ret = i915_gem_check_wedge(&dev_priv->gpu_error,
 975                                    dev_priv->mm.interruptible);
 976         if (ret)
 977                 return ret;
 978
 979         ret = logical_ring_prepare(ringbuf, num_dwords * sizeof(uint32_t));
 980         if (ret)
 981                 return ret;
 982
 983         /* Preallocate the olr before touching the ring */
 984         ret = logical_ring_alloc_seqno(ring, ringbuf->FIXME_lrc_ctx);
 985         if (ret)
 986                 return ret;
 987
 988         ringbuf->space -= num_dwords * sizeof(uint32_t);
 989         return 0;
 990 }
 991
 992 static int gen8_init_common_ring(struct intel_engine_cs *ring)
 993 {
 994         struct drm_device *dev = ring->dev;
 995         struct drm_i915_private *dev_priv = dev->dev_private;
 996
 997         I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
 998         I915_WRITE(RING_HWSTAM(ring->mmio_base), 0xffffffff);
 999
1000         I915_WRITE(RING_MODE_GEN7(ring),
1001                    _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1002                    _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1003         POSTING_READ(RING_MODE_GEN7(ring));
1004         DRM_DEBUG_DRIVER("Execlists enabled for %s\n", ring->name);
1005
1006         memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
1007
1008         return 0;
1009 }
1010
1011 static int gen8_init_render_ring(struct intel_engine_cs *ring)
1012 {
1013         struct drm_device *dev = ring->dev;
1014         struct drm_i915_private *dev_priv = dev->dev_private;
1015         int ret;
1016
1017         ret = gen8_init_common_ring(ring);
1018         if (ret)
1019                 return ret;
1020
1021         /* We need to disable the AsyncFlip performance optimisations in order
1022          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1023          * programmed to '1' on all products.
1024          *
1025          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1026          */
1027         I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1028
1029         ret = intel_init_pipe_control(ring);
1030         if (ret)
1031                 return ret;
1032
1033         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1034
1035         return ret;
1036 }
1037
1038 static int gen8_emit_bb_start(struct intel_ringbuffer *ringbuf,
1039                               u64 offset, unsigned flags)
1040 {
1041         bool ppgtt = !(flags & I915_DISPATCH_SECURE);
1042         int ret;
1043
1044         ret = intel_logical_ring_begin(ringbuf, 4);
1045         if (ret)
1046                 return ret;
1047
1048         /* FIXME(BDW): Address space and security selectors. */
1049         intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8));
1050         intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1051         intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1052         intel_logical_ring_emit(ringbuf, MI_NOOP);
1053         intel_logical_ring_advance(ringbuf);
1054
1055         return 0;
1056 }
1057
1058 static bool gen8_logical_ring_get_irq(struct intel_engine_cs *ring)
1059 {
1060         struct drm_device *dev = ring->dev;
1061         struct drm_i915_private *dev_priv = dev->dev_private;
1062         unsigned long flags;
1063
1064         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1065                 return false;
1066
1067         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1068         if (ring->irq_refcount++ == 0) {
1069                 I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
1070                 POSTING_READ(RING_IMR(ring->mmio_base));
1071         }
1072         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1073
1074         return true;
1075 }
1076
1077 static void gen8_logical_ring_put_irq(struct intel_engine_cs *ring)
1078 {
1079         struct drm_device *dev = ring->dev;
1080         struct drm_i915_private *dev_priv = dev->dev_private;
1081         unsigned long flags;
1082
1083         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1084         if (--ring->irq_refcount == 0) {
1085                 I915_WRITE_IMR(ring, ~ring->irq_keep_mask);
1086                 POSTING_READ(RING_IMR(ring->mmio_base));
1087         }
1088         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1089 }
1090
1091 static int gen8_emit_flush(struct intel_ringbuffer *ringbuf,
1092                            u32 invalidate_domains,
1093                            u32 unused)
1094 {
1095         struct intel_engine_cs *ring = ringbuf->ring;
1096         struct drm_device *dev = ring->dev;
1097         struct drm_i915_private *dev_priv = dev->dev_private;
1098         uint32_t cmd;
1099         int ret;
1100
1101         ret = intel_logical_ring_begin(ringbuf, 4);
1102         if (ret)
1103                 return ret;
1104
1105         cmd = MI_FLUSH_DW + 1;
1106
1107         if (ring == &dev_priv->ring[VCS]) {
1108                 if (invalidate_domains & I915_GEM_GPU_DOMAINS)
1109                         cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD |
1110                                 MI_FLUSH_DW_STORE_INDEX |
1111                                 MI_FLUSH_DW_OP_STOREDW;
1112         } else {
1113                 if (invalidate_domains & I915_GEM_DOMAIN_RENDER)
1114                         cmd |= MI_INVALIDATE_TLB | MI_FLUSH_DW_STORE_INDEX |
1115                                 MI_FLUSH_DW_OP_STOREDW;
1116         }
1117
1118         intel_logical_ring_emit(ringbuf, cmd);
1119         intel_logical_ring_emit(ringbuf,
1120                                 I915_GEM_HWS_SCRATCH_ADDR |
1121                                 MI_FLUSH_DW_USE_GTT);
1122         intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1123         intel_logical_ring_emit(ringbuf, 0); /* value */
1124         intel_logical_ring_advance(ringbuf);
1125
1126         return 0;
1127 }
1128
1129 static int gen8_emit_flush_render(struct intel_ringbuffer *ringbuf,
1130                                   u32 invalidate_domains,
1131                                   u32 flush_domains)
1132 {
1133         struct intel_engine_cs *ring = ringbuf->ring;
1134         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1135         u32 flags = 0;
1136         int ret;
1137
1138         flags |= PIPE_CONTROL_CS_STALL;
1139
1140         if (flush_domains) {
1141                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1142                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1143         }
1144
1145         if (invalidate_domains) {
1146                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1147                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1148                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1149                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1150                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1151                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1152                 flags |= PIPE_CONTROL_QW_WRITE;
1153                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1154         }
1155
1156         ret = intel_logical_ring_begin(ringbuf, 6);
1157         if (ret)
1158                 return ret;
1159
1160         intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1161         intel_logical_ring_emit(ringbuf, flags);
1162         intel_logical_ring_emit(ringbuf, scratch_addr);
1163         intel_logical_ring_emit(ringbuf, 0);
1164         intel_logical_ring_emit(ringbuf, 0);
1165         intel_logical_ring_emit(ringbuf, 0);
1166         intel_logical_ring_advance(ringbuf);
1167
1168         return 0;
1169 }
1170
1171 static u32 gen8_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1172 {
1173         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1174 }
1175
1176 static void gen8_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1177 {
1178         intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1179 }
1180
1181 static int gen8_emit_request(struct intel_ringbuffer *ringbuf)
1182 {
1183         struct intel_engine_cs *ring = ringbuf->ring;
1184         u32 cmd;
1185         int ret;
1186
1187         ret = intel_logical_ring_begin(ringbuf, 6);
1188         if (ret)
1189                 return ret;
1190
1191         cmd = MI_STORE_DWORD_IMM_GEN8;
1192         cmd |= MI_GLOBAL_GTT;
1193
1194         intel_logical_ring_emit(ringbuf, cmd);
1195         intel_logical_ring_emit(ringbuf,
1196                                 (ring->status_page.gfx_addr +
1197                                 (I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)));
1198         intel_logical_ring_emit(ringbuf, 0);
1199         intel_logical_ring_emit(ringbuf, ring->outstanding_lazy_seqno);
1200         intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1201         intel_logical_ring_emit(ringbuf, MI_NOOP);
1202         intel_logical_ring_advance_and_submit(ringbuf);
1203
1204         return 0;
1205 }
1206
1207 /**
1208  * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1209  *
1210  * @ring: Engine Command Streamer.
1211  *
1212  */
1213 void intel_logical_ring_cleanup(struct intel_engine_cs *ring)
1214 {
1215         struct drm_i915_private *dev_priv;
1216
1217         if (!intel_ring_initialized(ring))
1218                 return;
1219
1220         dev_priv = ring->dev->dev_private;
1221
1222         intel_logical_ring_stop(ring);
1223         WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0);
1224         ring->preallocated_lazy_request = NULL;
1225         ring->outstanding_lazy_seqno = 0;
1226
1227         if (ring->cleanup)
1228                 ring->cleanup(ring);
1229
1230         i915_cmd_parser_fini_ring(ring);
1231
1232         if (ring->status_page.obj) {
1233                 kunmap(sg_page(ring->status_page.obj->pages->sgl));
1234                 ring->status_page.obj = NULL;
1235         }
1236 }
1237
1238 static int logical_ring_init(struct drm_device *dev, struct intel_engine_cs *ring)
1239 {
1240         int ret;
1241
1242         /* Intentionally left blank. */
1243         ring->buffer = NULL;
1244
1245         ring->dev = dev;
1246         INIT_LIST_HEAD(&ring->active_list);
1247         INIT_LIST_HEAD(&ring->request_list);
1248         init_waitqueue_head(&ring->irq_queue);
1249
1250         INIT_LIST_HEAD(&ring->execlist_queue);
1251         spin_lock_init(&ring->execlist_lock);
1252         ring->next_context_status_buffer = 0;
1253
1254         ret = i915_cmd_parser_init_ring(ring);
1255         if (ret)
1256                 return ret;
1257
1258         if (ring->init) {
1259                 ret = ring->init(ring);
1260                 if (ret)
1261                         return ret;
1262         }
1263
1264         ret = intel_lr_context_deferred_create(ring->default_context, ring);
1265
1266         return ret;
1267 }
1268
1269 static int logical_render_ring_init(struct drm_device *dev)
1270 {
1271         struct drm_i915_private *dev_priv = dev->dev_private;
1272         struct intel_engine_cs *ring = &dev_priv->ring[RCS];
1273
1274         ring->name = "render ring";
1275         ring->id = RCS;
1276         ring->mmio_base = RENDER_RING_BASE;
1277         ring->irq_enable_mask =
1278                 GT_RENDER_USER_INTERRUPT << GEN8_RCS_IRQ_SHIFT;
1279         ring->irq_keep_mask =
1280                 GT_CONTEXT_SWITCH_INTERRUPT << GEN8_RCS_IRQ_SHIFT;
1281         if (HAS_L3_DPF(dev))
1282                 ring->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1283
1284         ring->init = gen8_init_render_ring;
1285         ring->cleanup = intel_fini_pipe_control;
1286         ring->get_seqno = gen8_get_seqno;
1287         ring->set_seqno = gen8_set_seqno;
1288         ring->emit_request = gen8_emit_request;
1289         ring->emit_flush = gen8_emit_flush_render;
1290         ring->irq_get = gen8_logical_ring_get_irq;
1291         ring->irq_put = gen8_logical_ring_put_irq;
1292         ring->emit_bb_start = gen8_emit_bb_start;
1293
1294         return logical_ring_init(dev, ring);
1295 }
1296
1297 static int logical_bsd_ring_init(struct drm_device *dev)
1298 {
1299         struct drm_i915_private *dev_priv = dev->dev_private;
1300         struct intel_engine_cs *ring = &dev_priv->ring[VCS];
1301
1302         ring->name = "bsd ring";
1303         ring->id = VCS;
1304         ring->mmio_base = GEN6_BSD_RING_BASE;
1305         ring->irq_enable_mask =
1306                 GT_RENDER_USER_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
1307         ring->irq_keep_mask =
1308                 GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
1309
1310         ring->init = gen8_init_common_ring;
1311         ring->get_seqno = gen8_get_seqno;
1312         ring->set_seqno = gen8_set_seqno;
1313         ring->emit_request = gen8_emit_request;
1314         ring->emit_flush = gen8_emit_flush;
1315         ring->irq_get = gen8_logical_ring_get_irq;
1316         ring->irq_put = gen8_logical_ring_put_irq;
1317         ring->emit_bb_start = gen8_emit_bb_start;
1318
1319         return logical_ring_init(dev, ring);
1320 }
1321
1322 static int logical_bsd2_ring_init(struct drm_device *dev)
1323 {
1324         struct drm_i915_private *dev_priv = dev->dev_private;
1325         struct intel_engine_cs *ring = &dev_priv->ring[VCS2];
1326
1327         ring->name = "bds2 ring";
1328         ring->id = VCS2;
1329         ring->mmio_base = GEN8_BSD2_RING_BASE;
1330         ring->irq_enable_mask =
1331                 GT_RENDER_USER_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
1332         ring->irq_keep_mask =
1333                 GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
1334
1335         ring->init = gen8_init_common_ring;
1336         ring->get_seqno = gen8_get_seqno;
1337         ring->set_seqno = gen8_set_seqno;
1338         ring->emit_request = gen8_emit_request;
1339         ring->emit_flush = gen8_emit_flush;
1340         ring->irq_get = gen8_logical_ring_get_irq;
1341         ring->irq_put = gen8_logical_ring_put_irq;
1342         ring->emit_bb_start = gen8_emit_bb_start;
1343
1344         return logical_ring_init(dev, ring);
1345 }
1346
1347 static int logical_blt_ring_init(struct drm_device *dev)
1348 {
1349         struct drm_i915_private *dev_priv = dev->dev_private;
1350         struct intel_engine_cs *ring = &dev_priv->ring[BCS];
1351
1352         ring->name = "blitter ring";
1353         ring->id = BCS;
1354         ring->mmio_base = BLT_RING_BASE;
1355         ring->irq_enable_mask =
1356                 GT_RENDER_USER_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
1357         ring->irq_keep_mask =
1358                 GT_CONTEXT_SWITCH_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
1359
1360         ring->init = gen8_init_common_ring;
1361         ring->get_seqno = gen8_get_seqno;
1362         ring->set_seqno = gen8_set_seqno;
1363         ring->emit_request = gen8_emit_request;
1364         ring->emit_flush = gen8_emit_flush;
1365         ring->irq_get = gen8_logical_ring_get_irq;
1366         ring->irq_put = gen8_logical_ring_put_irq;
1367         ring->emit_bb_start = gen8_emit_bb_start;
1368
1369         return logical_ring_init(dev, ring);
1370 }
1371
1372 static int logical_vebox_ring_init(struct drm_device *dev)
1373 {
1374         struct drm_i915_private *dev_priv = dev->dev_private;
1375         struct intel_engine_cs *ring = &dev_priv->ring[VECS];
1376
1377         ring->name = "video enhancement ring";
1378         ring->id = VECS;
1379         ring->mmio_base = VEBOX_RING_BASE;
1380         ring->irq_enable_mask =
1381                 GT_RENDER_USER_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
1382         ring->irq_keep_mask =
1383                 GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
1384
1385         ring->init = gen8_init_common_ring;
1386         ring->get_seqno = gen8_get_seqno;
1387         ring->set_seqno = gen8_set_seqno;
1388         ring->emit_request = gen8_emit_request;
1389         ring->emit_flush = gen8_emit_flush;
1390         ring->irq_get = gen8_logical_ring_get_irq;
1391         ring->irq_put = gen8_logical_ring_put_irq;
1392         ring->emit_bb_start = gen8_emit_bb_start;
1393
1394         return logical_ring_init(dev, ring);
1395 }
1396
1397 /**
1398  * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
1399  * @dev: DRM device.
1400  *
1401  * This function inits the engines for an Execlists submission style (the equivalent in the
1402  * legacy ringbuffer submission world would be i915_gem_init_rings). It does it only for
1403  * those engines that are present in the hardware.
1404  *
1405  * Return: non-zero if the initialization failed.
1406  */
1407 int intel_logical_rings_init(struct drm_device *dev)
1408 {
1409         struct drm_i915_private *dev_priv = dev->dev_private;
1410         int ret;
1411
1412         ret = logical_render_ring_init(dev);
1413         if (ret)
1414                 return ret;
1415
1416         if (HAS_BSD(dev)) {
1417                 ret = logical_bsd_ring_init(dev);
1418                 if (ret)
1419                         goto cleanup_render_ring;
1420         }
1421
1422         if (HAS_BLT(dev)) {
1423                 ret = logical_blt_ring_init(dev);
1424                 if (ret)
1425                         goto cleanup_bsd_ring;
1426         }
1427
1428         if (HAS_VEBOX(dev)) {
1429                 ret = logical_vebox_ring_init(dev);
1430                 if (ret)
1431                         goto cleanup_blt_ring;
1432         }
1433
1434         if (HAS_BSD2(dev)) {
1435                 ret = logical_bsd2_ring_init(dev);
1436                 if (ret)
1437                         goto cleanup_vebox_ring;
1438         }
1439
1440         ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
1441         if (ret)
1442                 goto cleanup_bsd2_ring;
1443
1444         return 0;
1445
1446 cleanup_bsd2_ring:
1447         intel_logical_ring_cleanup(&dev_priv->ring[VCS2]);
1448 cleanup_vebox_ring:
1449         intel_logical_ring_cleanup(&dev_priv->ring[VECS]);
1450 cleanup_blt_ring:
1451         intel_logical_ring_cleanup(&dev_priv->ring[BCS]);
1452 cleanup_bsd_ring:
1453         intel_logical_ring_cleanup(&dev_priv->ring[VCS]);
1454 cleanup_render_ring:
1455         intel_logical_ring_cleanup(&dev_priv->ring[RCS]);
1456
1457         return ret;
1458 }
1459
1460 int intel_lr_context_render_state_init(struct intel_engine_cs *ring,
1461                                        struct intel_context *ctx)
1462 {
1463         struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1464         struct render_state so;
1465         struct drm_i915_file_private *file_priv = ctx->file_priv;
1466         struct drm_file *file = file_priv ? file_priv->file : NULL;
1467         int ret;
1468
1469         ret = i915_gem_render_state_prepare(ring, &so);
1470         if (ret)
1471                 return ret;
1472
1473         if (so.rodata == NULL)
1474                 return 0;
1475
1476         ret = ring->emit_bb_start(ringbuf,
1477                         so.ggtt_offset,
1478                         I915_DISPATCH_SECURE);
1479         if (ret)
1480                 goto out;
1481
1482         i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), ring);
1483
1484         ret = __i915_add_request(ring, file, so.obj, NULL);
1485         /* intel_logical_ring_add_request moves object to inactive if it
1486          * fails */
1487 out:
1488         i915_gem_render_state_fini(&so);
1489         return ret;
1490 }
1491
1492 static int
1493 populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_obj,
1494                     struct intel_engine_cs *ring, struct intel_ringbuffer *ringbuf)
1495 {
1496         struct drm_device *dev = ring->dev;
1497         struct drm_i915_private *dev_priv = dev->dev_private;
1498         struct drm_i915_gem_object *ring_obj = ringbuf->obj;
1499         struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
1500         struct page *page;
1501         uint32_t *reg_state;
1502         int ret;
1503
1504         if (!ppgtt)
1505                 ppgtt = dev_priv->mm.aliasing_ppgtt;
1506
1507         ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
1508         if (ret) {
1509                 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
1510                 return ret;
1511         }
1512
1513         ret = i915_gem_object_get_pages(ctx_obj);
1514         if (ret) {
1515                 DRM_DEBUG_DRIVER("Could not get object pages\n");
1516                 return ret;
1517         }
1518
1519         i915_gem_object_pin_pages(ctx_obj);
1520
1521         /* The second page of the context object contains some fields which must
1522          * be set up prior to the first execution. */
1523         page = i915_gem_object_get_page(ctx_obj, 1);
1524         reg_state = kmap_atomic(page);
1525
1526         /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
1527          * commands followed by (reg, value) pairs. The values we are setting here are
1528          * only for the first context restore: on a subsequent save, the GPU will
1529          * recreate this batchbuffer with new values (including all the missing
1530          * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
1531         if (ring->id == RCS)
1532                 reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(14);
1533         else
1534                 reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(11);
1535         reg_state[CTX_LRI_HEADER_0] |= MI_LRI_FORCE_POSTED;
1536         reg_state[CTX_CONTEXT_CONTROL] = RING_CONTEXT_CONTROL(ring);
1537         reg_state[CTX_CONTEXT_CONTROL+1] =
1538                         _MASKED_BIT_ENABLE((1<<3) | MI_RESTORE_INHIBIT);
1539         reg_state[CTX_RING_HEAD] = RING_HEAD(ring->mmio_base);
1540         reg_state[CTX_RING_HEAD+1] = 0;
1541         reg_state[CTX_RING_TAIL] = RING_TAIL(ring->mmio_base);
1542         reg_state[CTX_RING_TAIL+1] = 0;
1543         reg_state[CTX_RING_BUFFER_START] = RING_START(ring->mmio_base);
1544         reg_state[CTX_RING_BUFFER_START+1] = i915_gem_obj_ggtt_offset(ring_obj);
1545         reg_state[CTX_RING_BUFFER_CONTROL] = RING_CTL(ring->mmio_base);
1546         reg_state[CTX_RING_BUFFER_CONTROL+1] =
1547                         ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID;
1548         reg_state[CTX_BB_HEAD_U] = ring->mmio_base + 0x168;
1549         reg_state[CTX_BB_HEAD_U+1] = 0;
1550         reg_state[CTX_BB_HEAD_L] = ring->mmio_base + 0x140;
1551         reg_state[CTX_BB_HEAD_L+1] = 0;
1552         reg_state[CTX_BB_STATE] = ring->mmio_base + 0x110;
1553         reg_state[CTX_BB_STATE+1] = (1<<5);
1554         reg_state[CTX_SECOND_BB_HEAD_U] = ring->mmio_base + 0x11c;
1555         reg_state[CTX_SECOND_BB_HEAD_U+1] = 0;
1556         reg_state[CTX_SECOND_BB_HEAD_L] = ring->mmio_base + 0x114;
1557         reg_state[CTX_SECOND_BB_HEAD_L+1] = 0;
1558         reg_state[CTX_SECOND_BB_STATE] = ring->mmio_base + 0x118;
1559         reg_state[CTX_SECOND_BB_STATE+1] = 0;
1560         if (ring->id == RCS) {
1561                 /* TODO: according to BSpec, the register state context
1562                  * for CHV does not have these. OTOH, these registers do
1563                  * exist in CHV. I'm waiting for a clarification */
1564                 reg_state[CTX_BB_PER_CTX_PTR] = ring->mmio_base + 0x1c0;
1565                 reg_state[CTX_BB_PER_CTX_PTR+1] = 0;
1566                 reg_state[CTX_RCS_INDIRECT_CTX] = ring->mmio_base + 0x1c4;
1567                 reg_state[CTX_RCS_INDIRECT_CTX+1] = 0;
1568                 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET] = ring->mmio_base + 0x1c8;
1569                 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] = 0;
1570         }
1571         reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9);
1572         reg_state[CTX_LRI_HEADER_1] |= MI_LRI_FORCE_POSTED;
1573         reg_state[CTX_CTX_TIMESTAMP] = ring->mmio_base + 0x3a8;
1574         reg_state[CTX_CTX_TIMESTAMP+1] = 0;
1575         reg_state[CTX_PDP3_UDW] = GEN8_RING_PDP_UDW(ring, 3);
1576         reg_state[CTX_PDP3_LDW] = GEN8_RING_PDP_LDW(ring, 3);
1577         reg_state[CTX_PDP2_UDW] = GEN8_RING_PDP_UDW(ring, 2);
1578         reg_state[CTX_PDP2_LDW] = GEN8_RING_PDP_LDW(ring, 2);
1579         reg_state[CTX_PDP1_UDW] = GEN8_RING_PDP_UDW(ring, 1);
1580         reg_state[CTX_PDP1_LDW] = GEN8_RING_PDP_LDW(ring, 1);
1581         reg_state[CTX_PDP0_UDW] = GEN8_RING_PDP_UDW(ring, 0);
1582         reg_state[CTX_PDP0_LDW] = GEN8_RING_PDP_LDW(ring, 0);
1583         reg_state[CTX_PDP3_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[3]);
1584         reg_state[CTX_PDP3_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[3]);
1585         reg_state[CTX_PDP2_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[2]);
1586         reg_state[CTX_PDP2_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[2]);
1587         reg_state[CTX_PDP1_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[1]);
1588         reg_state[CTX_PDP1_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[1]);
1589         reg_state[CTX_PDP0_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[0]);
1590         reg_state[CTX_PDP0_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[0]);
1591         if (ring->id == RCS) {
1592                 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
1593                 reg_state[CTX_R_PWR_CLK_STATE] = 0x20c8;
1594                 reg_state[CTX_R_PWR_CLK_STATE+1] = 0;
1595         }
1596
1597         kunmap_atomic(reg_state);
1598
1599         ctx_obj->dirty = 1;
1600         set_page_dirty(page);
1601         i915_gem_object_unpin_pages(ctx_obj);
1602
1603         return 0;
1604 }
1605
1606 /**
1607  * intel_lr_context_free() - free the LRC specific bits of a context
1608  * @ctx: the LR context to free.
1609  *
1610  * The real context freeing is done in i915_gem_context_free: this only
1611  * takes care of the bits that are LRC related: the per-engine backing
1612  * objects and the logical ringbuffer.
1613  */
1614 void intel_lr_context_free(struct intel_context *ctx)
1615 {
1616         int i;
1617
1618         for (i = 0; i < I915_NUM_RINGS; i++) {
1619                 struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state;
1620                 struct intel_ringbuffer *ringbuf = ctx->engine[i].ringbuf;
1621
1622                 if (ctx_obj) {
1623                         intel_destroy_ringbuffer_obj(ringbuf);
1624                         kfree(ringbuf);
1625                         i915_gem_object_ggtt_unpin(ctx_obj);
1626                         drm_gem_object_unreference(&ctx_obj->base);
1627                 }
1628         }
1629 }
1630
1631 static uint32_t get_lr_context_size(struct intel_engine_cs *ring)
1632 {
1633         int ret = 0;
1634
1635         WARN_ON(INTEL_INFO(ring->dev)->gen != 8);
1636
1637         switch (ring->id) {
1638         case RCS:
1639                 ret = GEN8_LR_CONTEXT_RENDER_SIZE;
1640                 break;
1641         case VCS:
1642         case BCS:
1643         case VECS:
1644         case VCS2:
1645                 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
1646                 break;
1647         }
1648
1649         return ret;
1650 }
1651
1652 static int lrc_setup_hardware_status_page(struct intel_engine_cs *ring,
1653                 struct drm_i915_gem_object *default_ctx_obj)
1654 {
1655         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1656
1657         /* The status page is offset 0 from the default context object
1658          * in LRC mode. */
1659         ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(default_ctx_obj);
1660         ring->status_page.page_addr =
1661                         kmap(sg_page(default_ctx_obj->pages->sgl));
1662         if (ring->status_page.page_addr == NULL)
1663                 return -ENOMEM;
1664         ring->status_page.obj = default_ctx_obj;
1665
1666         I915_WRITE(RING_HWS_PGA(ring->mmio_base),
1667                         (u32)ring->status_page.gfx_addr);
1668         POSTING_READ(RING_HWS_PGA(ring->mmio_base));
1669
1670         return 0;
1671 }
1672
1673 /**
1674  * intel_lr_context_deferred_create() - create the LRC specific bits of a context
1675  * @ctx: LR context to create.
1676  * @ring: engine to be used with the context.
1677  *
1678  * This function can be called more than once, with different engines, if we plan
1679  * to use the context with them. The context backing objects and the ringbuffers
1680  * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
1681  * the creation is a deferred call: it's better to make sure first that we need to use
1682  * a given ring with the context.
1683  *
1684  * Return: non-zero on error.
1685  */
1686 int intel_lr_context_deferred_create(struct intel_context *ctx,
1687                                      struct intel_engine_cs *ring)
1688 {
1689         struct drm_device *dev = ring->dev;
1690         struct drm_i915_gem_object *ctx_obj;
1691         uint32_t context_size;
1692         struct intel_ringbuffer *ringbuf;
1693         int ret;
1694
1695         WARN_ON(ctx->legacy_hw_ctx.rcs_state != NULL);
1696         if (ctx->engine[ring->id].state)
1697                 return 0;
1698
1699         context_size = round_up(get_lr_context_size(ring), 4096);
1700
1701         ctx_obj = i915_gem_alloc_context_obj(dev, context_size);
1702         if (IS_ERR(ctx_obj)) {
1703                 ret = PTR_ERR(ctx_obj);
1704                 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed: %d\n", ret);
1705                 return ret;
1706         }
1707
1708         ret = i915_gem_obj_ggtt_pin(ctx_obj, GEN8_LR_CONTEXT_ALIGN, 0);
1709         if (ret) {
1710                 DRM_DEBUG_DRIVER("Pin LRC backing obj failed: %d\n", ret);
1711                 drm_gem_object_unreference(&ctx_obj->base);
1712                 return ret;
1713         }
1714
1715         ringbuf = kzalloc(sizeof(*ringbuf), GFP_KERNEL);
1716         if (!ringbuf) {
1717                 DRM_DEBUG_DRIVER("Failed to allocate ringbuffer %s\n",
1718                                 ring->name);
1719                 i915_gem_object_ggtt_unpin(ctx_obj);
1720                 drm_gem_object_unreference(&ctx_obj->base);
1721                 ret = -ENOMEM;
1722                 return ret;
1723         }
1724
1725         ringbuf->ring = ring;
1726         ringbuf->FIXME_lrc_ctx = ctx;
1727
1728         ringbuf->size = 32 * PAGE_SIZE;
1729         ringbuf->effective_size = ringbuf->size;
1730         ringbuf->head = 0;
1731         ringbuf->tail = 0;
1732         ringbuf->space = ringbuf->size;
1733         ringbuf->last_retired_head = -1;
1734
1735         /* TODO: For now we put this in the mappable region so that we can reuse
1736          * the existing ringbuffer code which ioremaps it. When we start
1737          * creating many contexts, this will no longer work and we must switch
1738          * to a kmapish interface.
1739          */
1740         ret = intel_alloc_ringbuffer_obj(dev, ringbuf);
1741         if (ret) {
1742                 DRM_DEBUG_DRIVER("Failed to allocate ringbuffer obj %s: %d\n",
1743                                 ring->name, ret);
1744                 goto error;
1745         }
1746
1747         ret = populate_lr_context(ctx, ctx_obj, ring, ringbuf);
1748         if (ret) {
1749                 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
1750                 intel_destroy_ringbuffer_obj(ringbuf);
1751                 goto error;
1752         }
1753
1754         ctx->engine[ring->id].ringbuf = ringbuf;
1755         ctx->engine[ring->id].state = ctx_obj;
1756
1757         if (ctx == ring->default_context) {
1758                 ret = lrc_setup_hardware_status_page(ring, ctx_obj);
1759                 if (ret) {
1760                         DRM_ERROR("Failed to setup hardware status page\n");
1761                         goto error;
1762                 }
1763         }
1764
1765         if (ring->id == RCS && !ctx->rcs_initialized) {
1766                 ret = intel_lr_context_render_state_init(ring, ctx);
1767                 if (ret) {
1768                         DRM_ERROR("Init render state failed: %d\n", ret);
1769                         ctx->engine[ring->id].ringbuf = NULL;
1770                         ctx->engine[ring->id].state = NULL;
1771                         intel_destroy_ringbuffer_obj(ringbuf);
1772                         goto error;
1773                 }
1774                 ctx->rcs_initialized = true;
1775         }
1776
1777         return 0;
1778
1779 error:
1780         kfree(ringbuf);
1781         i915_gem_object_ggtt_unpin(ctx_obj);
1782         drm_gem_object_unreference(&ctx_obj->base);
1783         return ret;
1784 }