drivers/gpu/drm/i915/intel_ringbuffer.c

   1 /*
   2  * Copyright © 2008-2010 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  *    Eric Anholt <eric@anholt.net>
  25  *    Zou Nan hai <nanhai.zou@intel.com>
  26  *    Xiang Hai hao<haihao.xiang@intel.com>
  27  *
  28  */
  29
  30 #include "drmP.h"
  31 #include "drm.h"
  32 #include "i915_drv.h"
  33 #include "i915_drm.h"
  34 #include "i915_trace.h"
  35 #include "intel_drv.h"
  36
  37 static u32 i915_gem_get_seqno(struct drm_device *dev)
  38 {
  39         drm_i915_private_t *dev_priv = dev->dev_private;
  40         u32 seqno;
  41
  42         seqno = dev_priv->next_seqno;
  43
  44         /* reserve 0 for non-seqno */
  45         if (++dev_priv->next_seqno == 0)
  46                 dev_priv->next_seqno = 1;
  47
  48         return seqno;
  49 }
  50
  51 static void
  52 render_ring_flush(struct intel_ring_buffer *ring,
  53                   u32   invalidate_domains,
  54                   u32   flush_domains)
  55 {
  56         struct drm_device *dev = ring->dev;
  57         drm_i915_private_t *dev_priv = dev->dev_private;
  58         u32 cmd;
  59
  60 #if WATCH_EXEC
  61         DRM_INFO("%s: invalidate %08x flush %08x\n", __func__,
  62                   invalidate_domains, flush_domains);
  63 #endif
  64
  65         trace_i915_gem_request_flush(dev, dev_priv->next_seqno,
  66                                      invalidate_domains, flush_domains);
  67
  68         if ((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) {
  69                 /*
  70                  * read/write caches:
  71                  *
  72                  * I915_GEM_DOMAIN_RENDER is always invalidated, but is
  73                  * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
  74                  * also flushed at 2d versus 3d pipeline switches.
  75                  *
  76                  * read-only caches:
  77                  *
  78                  * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
  79                  * MI_READ_FLUSH is set, and is always flushed on 965.
  80                  *
  81                  * I915_GEM_DOMAIN_COMMAND may not exist?
  82                  *
  83                  * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
  84                  * invalidated when MI_EXE_FLUSH is set.
  85                  *
  86                  * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
  87                  * invalidated with every MI_FLUSH.
  88                  *
  89                  * TLBs:
  90                  *
  91                  * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
  92                  * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
  93                  * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
  94                  * are flushed at any MI_FLUSH.
  95                  */
  96
  97                 cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
  98                 if ((invalidate_domains|flush_domains) &
  99                     I915_GEM_DOMAIN_RENDER)
 100                         cmd &= ~MI_NO_WRITE_FLUSH;
 101                 if (INTEL_INFO(dev)->gen < 4) {
 102                         /*
 103                          * On the 965, the sampler cache always gets flushed
 104                          * and this bit is reserved.
 105                          */
 106                         if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
 107                                 cmd |= MI_READ_FLUSH;
 108                 }
 109                 if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
 110                         cmd |= MI_EXE_FLUSH;
 111
 112 #if WATCH_EXEC
 113                 DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd);
 114 #endif
 115                 if (intel_ring_begin(ring, 2) == 0) {
 116                         intel_ring_emit(ring, cmd);
 117                         intel_ring_emit(ring, MI_NOOP);
 118                         intel_ring_advance(ring);
 119                 }
 120         }
 121 }
 122
 123 static void ring_write_tail(struct intel_ring_buffer *ring,
 124                             u32 value)
 125 {
 126         drm_i915_private_t *dev_priv = ring->dev->dev_private;
 127         I915_WRITE_TAIL(ring, value);
 128 }
 129
 130 u32 intel_ring_get_active_head(struct intel_ring_buffer *ring)
 131 {
 132         drm_i915_private_t *dev_priv = ring->dev->dev_private;
 133         u32 acthd_reg = INTEL_INFO(ring->dev)->gen >= 4 ?
 134                         RING_ACTHD(ring->mmio_base) : ACTHD;
 135
 136         return I915_READ(acthd_reg);
 137 }
 138
 139 static int init_ring_common(struct intel_ring_buffer *ring)
 140 {
 141         drm_i915_private_t *dev_priv = ring->dev->dev_private;
 142         struct drm_i915_gem_object *obj_priv = to_intel_bo(ring->gem_object);
 143         u32 head;
 144
 145         /* Stop the ring if it's running. */
 146         I915_WRITE_CTL(ring, 0);
 147         I915_WRITE_HEAD(ring, 0);
 148         ring->write_tail(ring, 0);
 149
 150         /* Initialize the ring. */
 151         I915_WRITE_START(ring, obj_priv->gtt_offset);
 152         head = I915_READ_HEAD(ring) & HEAD_ADDR;
 153
 154         /* G45 ring initialization fails to reset head to zero */
 155         if (head != 0) {
 156                 DRM_ERROR("%s head not reset to zero "
 157                                 "ctl %08x head %08x tail %08x start %08x\n",
 158                                 ring->name,
 159                                 I915_READ_CTL(ring),
 160                                 I915_READ_HEAD(ring),
 161                                 I915_READ_TAIL(ring),
 162                                 I915_READ_START(ring));
 163
 164                 I915_WRITE_HEAD(ring, 0);
 165
 166                 DRM_ERROR("%s head forced to zero "
 167                                 "ctl %08x head %08x tail %08x start %08x\n",
 168                                 ring->name,
 169                                 I915_READ_CTL(ring),
 170                                 I915_READ_HEAD(ring),
 171                                 I915_READ_TAIL(ring),
 172                                 I915_READ_START(ring));
 173         }
 174
 175         I915_WRITE_CTL(ring,
 176                         ((ring->gem_object->size - PAGE_SIZE) & RING_NR_PAGES)
 177                         | RING_REPORT_64K | RING_VALID);
 178
 179         /* If the head is still not zero, the ring is dead */
 180         if ((I915_READ_CTL(ring) & RING_VALID) == 0 ||
 181             I915_READ_START(ring) != obj_priv->gtt_offset ||
 182             (I915_READ_HEAD(ring) & HEAD_ADDR) != 0) {
 183                 DRM_ERROR("%s initialization failed "
 184                                 "ctl %08x head %08x tail %08x start %08x\n",
 185                                 ring->name,
 186                                 I915_READ_CTL(ring),
 187                                 I915_READ_HEAD(ring),
 188                                 I915_READ_TAIL(ring),
 189                                 I915_READ_START(ring));
 190                 return -EIO;
 191         }
 192
 193         if (!drm_core_check_feature(ring->dev, DRIVER_MODESET))
 194                 i915_kernel_lost_context(ring->dev);
 195         else {
 196                 ring->head = I915_READ_HEAD(ring) & HEAD_ADDR;
 197                 ring->tail = I915_READ_TAIL(ring) & TAIL_ADDR;
 198                 ring->space = ring->head - (ring->tail + 8);
 199                 if (ring->space < 0)
 200                         ring->space += ring->size;
 201         }
 202         return 0;
 203 }
 204
 205 static int init_render_ring(struct intel_ring_buffer *ring)
 206 {
 207         struct drm_device *dev = ring->dev;
 208         int ret = init_ring_common(ring);
 209
 210         if (INTEL_INFO(dev)->gen > 3) {
 211                 drm_i915_private_t *dev_priv = dev->dev_private;
 212                 int mode = VS_TIMER_DISPATCH << 16 | VS_TIMER_DISPATCH;
 213                 if (IS_GEN6(dev))
 214                         mode |= MI_FLUSH_ENABLE << 16 | MI_FLUSH_ENABLE;
 215                 I915_WRITE(MI_MODE, mode);
 216         }
 217
 218         return ret;
 219 }
 220
 221 #define PIPE_CONTROL_FLUSH(ring__, addr__)                                      \
 222 do {                                                                    \
 223         intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE |           \
 224                  PIPE_CONTROL_DEPTH_STALL | 2);                         \
 225         intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT);                    \
 226         intel_ring_emit(ring__, 0);                                                     \
 227         intel_ring_emit(ring__, 0);                                                     \
 228 } while (0)
 229
 230 /**
 231  * Creates a new sequence number, emitting a write of it to the status page
 232  * plus an interrupt, which will trigger i915_user_interrupt_handler.
 233  *
 234  * Must be called with struct_lock held.
 235  *
 236  * Returned sequence numbers are nonzero on success.
 237  */
 238 static int
 239 render_ring_add_request(struct intel_ring_buffer *ring,
 240                         u32 *result)
 241 {
 242         struct drm_device *dev = ring->dev;
 243         drm_i915_private_t *dev_priv = dev->dev_private;
 244         u32 seqno = i915_gem_get_seqno(dev);
 245         int ret;
 246
 247         if (IS_GEN6(dev)) {
 248                 ret = intel_ring_begin(ring, 6);
 249                 if (ret)
 250                     return ret;
 251
 252                 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL | 3);
 253                 intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE |
 254                                 PIPE_CONTROL_WC_FLUSH | PIPE_CONTROL_IS_FLUSH |
 255                                 PIPE_CONTROL_NOTIFY);
 256                 intel_ring_emit(ring, dev_priv->seqno_gfx_addr | PIPE_CONTROL_GLOBAL_GTT);
 257                 intel_ring_emit(ring, seqno);
 258                 intel_ring_emit(ring, 0);
 259                 intel_ring_emit(ring, 0);
 260         } else if (HAS_PIPE_CONTROL(dev)) {
 261                 u32 scratch_addr = dev_priv->seqno_gfx_addr + 128;
 262
 263                 /*
 264                  * Workaround qword write incoherence by flushing the
 265                  * PIPE_NOTIFY buffers out to memory before requesting
 266                  * an interrupt.
 267                  */
 268                 ret = intel_ring_begin(ring, 32);
 269                 if (ret)
 270                         return ret;
 271
 272                 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE |
 273                                 PIPE_CONTROL_WC_FLUSH | PIPE_CONTROL_TC_FLUSH);
 274                 intel_ring_emit(ring, dev_priv->seqno_gfx_addr | PIPE_CONTROL_GLOBAL_GTT);
 275                 intel_ring_emit(ring, seqno);
 276                 intel_ring_emit(ring, 0);
 277                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 278                 scratch_addr += 128; /* write to separate cachelines */
 279                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 280                 scratch_addr += 128;
 281                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 282                 scratch_addr += 128;
 283                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 284                 scratch_addr += 128;
 285                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 286                 scratch_addr += 128;
 287                 PIPE_CONTROL_FLUSH(ring, scratch_addr);
 288                 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL | PIPE_CONTROL_QW_WRITE |
 289                                 PIPE_CONTROL_WC_FLUSH | PIPE_CONTROL_TC_FLUSH |
 290                                 PIPE_CONTROL_NOTIFY);
 291                 intel_ring_emit(ring, dev_priv->seqno_gfx_addr | PIPE_CONTROL_GLOBAL_GTT);
 292                 intel_ring_emit(ring, seqno);
 293                 intel_ring_emit(ring, 0);
 294         } else {
 295                 ret = intel_ring_begin(ring, 4);
 296                 if (ret)
 297                     return ret;
 298
 299                 intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
 300                 intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 301                 intel_ring_emit(ring, seqno);
 302
 303                 intel_ring_emit(ring, MI_USER_INTERRUPT);
 304         }
 305
 306         intel_ring_advance(ring);
 307         *result = seqno;
 308         return 0;
 309 }
 310
 311 static u32
 312 render_ring_get_seqno(struct intel_ring_buffer *ring)
 313 {
 314         struct drm_device *dev = ring->dev;
 315         drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
 316         if (HAS_PIPE_CONTROL(dev))
 317                 return ((volatile u32 *)(dev_priv->seqno_page))[0];
 318         else
 319                 return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
 320 }
 321
 322 static void
 323 render_ring_get_user_irq(struct intel_ring_buffer *ring)
 324 {
 325         struct drm_device *dev = ring->dev;
 326         drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
 327         unsigned long irqflags;
 328
 329         spin_lock_irqsave(&dev_priv->user_irq_lock, irqflags);
 330         if (dev->irq_enabled && (++ring->user_irq_refcount == 1)) {
 331                 if (HAS_PCH_SPLIT(dev))
 332                         ironlake_enable_graphics_irq(dev_priv, GT_PIPE_NOTIFY);
 333                 else
 334                         i915_enable_irq(dev_priv, I915_USER_INTERRUPT);
 335         }
 336         spin_unlock_irqrestore(&dev_priv->user_irq_lock, irqflags);
 337 }
 338
 339 static void
 340 render_ring_put_user_irq(struct intel_ring_buffer *ring)
 341 {
 342         struct drm_device *dev = ring->dev;
 343         drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
 344         unsigned long irqflags;
 345
 346         spin_lock_irqsave(&dev_priv->user_irq_lock, irqflags);
 347         BUG_ON(dev->irq_enabled && ring->user_irq_refcount <= 0);
 348         if (dev->irq_enabled && (--ring->user_irq_refcount == 0)) {
 349                 if (HAS_PCH_SPLIT(dev))
 350                         ironlake_disable_graphics_irq(dev_priv, GT_PIPE_NOTIFY);
 351                 else
 352                         i915_disable_irq(dev_priv, I915_USER_INTERRUPT);
 353         }
 354         spin_unlock_irqrestore(&dev_priv->user_irq_lock, irqflags);
 355 }
 356
 357 void intel_ring_setup_status_page(struct intel_ring_buffer *ring)
 358 {
 359         drm_i915_private_t *dev_priv = ring->dev->dev_private;
 360         u32 mmio = IS_GEN6(ring->dev) ?
 361                 RING_HWS_PGA_GEN6(ring->mmio_base) :
 362                 RING_HWS_PGA(ring->mmio_base);
 363         I915_WRITE(mmio, (u32)ring->status_page.gfx_addr);
 364         POSTING_READ(mmio);
 365 }
 366
 367 static void
 368 bsd_ring_flush(struct intel_ring_buffer *ring,
 369                u32     invalidate_domains,
 370                u32     flush_domains)
 371 {
 372         if (intel_ring_begin(ring, 2) == 0) {
 373                 intel_ring_emit(ring, MI_FLUSH);
 374                 intel_ring_emit(ring, MI_NOOP);
 375                 intel_ring_advance(ring);
 376         }
 377 }
 378
 379 static int
 380 ring_add_request(struct intel_ring_buffer *ring,
 381                  u32 *result)
 382 {
 383         u32 seqno;
 384         int ret;
 385
 386         ret = intel_ring_begin(ring, 4);
 387         if (ret)
 388                 return ret;
 389
 390         seqno = i915_gem_get_seqno(ring->dev);
 391
 392         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
 393         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 394         intel_ring_emit(ring, seqno);
 395         intel_ring_emit(ring, MI_USER_INTERRUPT);
 396         intel_ring_advance(ring);
 397
 398         DRM_DEBUG_DRIVER("%s %d\n", ring->name, seqno);
 399         *result = seqno;
 400         return 0;
 401 }
 402
 403 static void
 404 bsd_ring_get_user_irq(struct intel_ring_buffer *ring)
 405 {
 406         /* do nothing */
 407 }
 408 static void
 409 bsd_ring_put_user_irq(struct intel_ring_buffer *ring)
 410 {
 411         /* do nothing */
 412 }
 413
 414 static u32
 415 ring_status_page_get_seqno(struct intel_ring_buffer *ring)
 416 {
 417         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
 418 }
 419
 420 static int
 421 ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
 422                          struct drm_i915_gem_execbuffer2 *exec,
 423                          struct drm_clip_rect *cliprects,
 424                          uint64_t exec_offset)
 425 {
 426         uint32_t exec_start;
 427         int ret;
 428
 429         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
 430
 431         ret = intel_ring_begin(ring, 2);
 432         if (ret)
 433                 return ret;
 434
 435         intel_ring_emit(ring,
 436                         MI_BATCH_BUFFER_START |
 437                         (2 << 6) |
 438                         MI_BATCH_NON_SECURE_I965);
 439         intel_ring_emit(ring, exec_start);
 440         intel_ring_advance(ring);
 441
 442         return 0;
 443 }
 444
 445 static int
 446 render_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
 447                                 struct drm_i915_gem_execbuffer2 *exec,
 448                                 struct drm_clip_rect *cliprects,
 449                                 uint64_t exec_offset)
 450 {
 451         struct drm_device *dev = ring->dev;
 452         drm_i915_private_t *dev_priv = dev->dev_private;
 453         int nbox = exec->num_cliprects;
 454         uint32_t exec_start, exec_len;
 455         int i, count, ret;
 456
 457         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
 458         exec_len = (uint32_t) exec->batch_len;
 459
 460         trace_i915_gem_request_submit(dev, dev_priv->next_seqno + 1);
 461
 462         count = nbox ? nbox : 1;
 463         for (i = 0; i < count; i++) {
 464                 if (i < nbox) {
 465                         ret = i915_emit_box(dev, cliprects, i,
 466                                             exec->DR1, exec->DR4);
 467                         if (ret)
 468                                 return ret;
 469                 }
 470
 471                 if (IS_I830(dev) || IS_845G(dev)) {
 472                         ret = intel_ring_begin(ring, 4);
 473                         if (ret)
 474                                 return ret;
 475
 476                         intel_ring_emit(ring, MI_BATCH_BUFFER);
 477                         intel_ring_emit(ring, exec_start | MI_BATCH_NON_SECURE);
 478                         intel_ring_emit(ring, exec_start + exec_len - 4);
 479                         intel_ring_emit(ring, 0);
 480                 } else {
 481                         ret = intel_ring_begin(ring, 2);
 482                         if (ret)
 483                                 return ret;
 484
 485                         if (INTEL_INFO(dev)->gen >= 4) {
 486                                 intel_ring_emit(ring,
 487                                                 MI_BATCH_BUFFER_START | (2 << 6)
 488                                                 | MI_BATCH_NON_SECURE_I965);
 489                                 intel_ring_emit(ring, exec_start);
 490                         } else {
 491                                 intel_ring_emit(ring, MI_BATCH_BUFFER_START
 492                                                 | (2 << 6));
 493                                 intel_ring_emit(ring, exec_start |
 494                                                 MI_BATCH_NON_SECURE);
 495                         }
 496                 }
 497                 intel_ring_advance(ring);
 498         }
 499
 500         if (IS_G4X(dev) || IS_GEN5(dev)) {
 501                 if (intel_ring_begin(ring, 2) == 0) {
 502                         intel_ring_emit(ring, MI_FLUSH |
 503                                         MI_NO_WRITE_FLUSH |
 504                                         MI_INVALIDATE_ISP );
 505                         intel_ring_emit(ring, MI_NOOP);
 506                         intel_ring_advance(ring);
 507                 }
 508         }
 509         /* XXX breadcrumb */
 510
 511         return 0;
 512 }
 513
 514 static void cleanup_status_page(struct intel_ring_buffer *ring)
 515 {
 516         drm_i915_private_t *dev_priv = ring->dev->dev_private;
 517         struct drm_gem_object *obj;
 518         struct drm_i915_gem_object *obj_priv;
 519
 520         obj = ring->status_page.obj;
 521         if (obj == NULL)
 522                 return;
 523         obj_priv = to_intel_bo(obj);
 524
 525         kunmap(obj_priv->pages[0]);
 526         i915_gem_object_unpin(obj);
 527         drm_gem_object_unreference(obj);
 528         ring->status_page.obj = NULL;
 529
 530         memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
 531 }
 532
 533 static int init_status_page(struct intel_ring_buffer *ring)
 534 {
 535         struct drm_device *dev = ring->dev;
 536         drm_i915_private_t *dev_priv = dev->dev_private;
 537         struct drm_gem_object *obj;
 538         struct drm_i915_gem_object *obj_priv;
 539         int ret;
 540
 541         obj = i915_gem_alloc_object(dev, 4096);
 542         if (obj == NULL) {
 543                 DRM_ERROR("Failed to allocate status page\n");
 544                 ret = -ENOMEM;
 545                 goto err;
 546         }
 547         obj_priv = to_intel_bo(obj);
 548         obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
 549
 550         ret = i915_gem_object_pin(obj, 4096, true);
 551         if (ret != 0) {
 552                 goto err_unref;
 553         }
 554
 555         ring->status_page.gfx_addr = obj_priv->gtt_offset;
 556         ring->status_page.page_addr = kmap(obj_priv->pages[0]);
 557         if (ring->status_page.page_addr == NULL) {
 558                 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
 559                 goto err_unpin;
 560         }
 561         ring->status_page.obj = obj;
 562         memset(ring->status_page.page_addr, 0, PAGE_SIZE);
 563
 564         intel_ring_setup_status_page(ring);
 565         DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
 566                         ring->name, ring->status_page.gfx_addr);
 567
 568         return 0;
 569
 570 err_unpin:
 571         i915_gem_object_unpin(obj);
 572 err_unref:
 573         drm_gem_object_unreference(obj);
 574 err:
 575         return ret;
 576 }
 577
 578 int intel_init_ring_buffer(struct drm_device *dev,
 579                            struct intel_ring_buffer *ring)
 580 {
 581         struct drm_i915_gem_object *obj_priv;
 582         struct drm_gem_object *obj;
 583         int ret;
 584
 585         ring->dev = dev;
 586         INIT_LIST_HEAD(&ring->active_list);
 587         INIT_LIST_HEAD(&ring->request_list);
 588         INIT_LIST_HEAD(&ring->gpu_write_list);
 589
 590         if (I915_NEED_GFX_HWS(dev)) {
 591                 ret = init_status_page(ring);
 592                 if (ret)
 593                         return ret;
 594         }
 595
 596         obj = i915_gem_alloc_object(dev, ring->size);
 597         if (obj == NULL) {
 598                 DRM_ERROR("Failed to allocate ringbuffer\n");
 599                 ret = -ENOMEM;
 600                 goto err_hws;
 601         }
 602
 603         ring->gem_object = obj;
 604
 605         ret = i915_gem_object_pin(obj, PAGE_SIZE, true);
 606         if (ret)
 607                 goto err_unref;
 608
 609         obj_priv = to_intel_bo(obj);
 610         ring->map.size = ring->size;
 611         ring->map.offset = dev->agp->base + obj_priv->gtt_offset;
 612         ring->map.type = 0;
 613         ring->map.flags = 0;
 614         ring->map.mtrr = 0;
 615
 616         drm_core_ioremap_wc(&ring->map, dev);
 617         if (ring->map.handle == NULL) {
 618                 DRM_ERROR("Failed to map ringbuffer.\n");
 619                 ret = -EINVAL;
 620                 goto err_unpin;
 621         }
 622
 623         ring->virtual_start = ring->map.handle;
 624         ret = ring->init(ring);
 625         if (ret)
 626                 goto err_unmap;
 627
 628         return 0;
 629
 630 err_unmap:
 631         drm_core_ioremapfree(&ring->map, dev);
 632 err_unpin:
 633         i915_gem_object_unpin(obj);
 634 err_unref:
 635         drm_gem_object_unreference(obj);
 636         ring->gem_object = NULL;
 637 err_hws:
 638         cleanup_status_page(ring);
 639         return ret;
 640 }
 641
 642 void intel_cleanup_ring_buffer(struct intel_ring_buffer *ring)
 643 {
 644         struct drm_i915_private *dev_priv;
 645         int ret;
 646
 647         if (ring->gem_object == NULL)
 648                 return;
 649
 650         /* Disable the ring buffer. The ring must be idle at this point */
 651         dev_priv = ring->dev->dev_private;
 652         ret = intel_wait_ring_buffer(ring, ring->size - 8);
 653         I915_WRITE_CTL(ring, 0);
 654
 655         drm_core_ioremapfree(&ring->map, ring->dev);
 656
 657         i915_gem_object_unpin(ring->gem_object);
 658         drm_gem_object_unreference(ring->gem_object);
 659         ring->gem_object = NULL;
 660
 661         if (ring->cleanup)
 662                 ring->cleanup(ring);
 663
 664         cleanup_status_page(ring);
 665 }
 666
 667 static int intel_wrap_ring_buffer(struct intel_ring_buffer *ring)
 668 {
 669         unsigned int *virt;
 670         int rem;
 671         rem = ring->size - ring->tail;
 672
 673         if (ring->space < rem) {
 674                 int ret = intel_wait_ring_buffer(ring, rem);
 675                 if (ret)
 676                         return ret;
 677         }
 678
 679         virt = (unsigned int *)(ring->virtual_start + ring->tail);
 680         rem /= 8;
 681         while (rem--) {
 682                 *virt++ = MI_NOOP;
 683                 *virt++ = MI_NOOP;
 684         }
 685
 686         ring->tail = 0;
 687         ring->space = ring->head - 8;
 688
 689         return 0;
 690 }
 691
 692 int intel_wait_ring_buffer(struct intel_ring_buffer *ring, int n)
 693 {
 694         struct drm_device *dev = ring->dev;
 695         drm_i915_private_t *dev_priv = dev->dev_private;
 696         unsigned long end;
 697         u32 head;
 698
 699         head = intel_read_status_page(ring, 4);
 700         if (head) {
 701                 ring->head = head & HEAD_ADDR;
 702                 ring->space = ring->head - (ring->tail + 8);
 703                 if (ring->space < 0)
 704                         ring->space += ring->size;
 705                 if (ring->space >= n)
 706                         return 0;
 707         }
 708
 709         trace_i915_ring_wait_begin (dev);
 710         end = jiffies + 3 * HZ;
 711         do {
 712                 ring->head = I915_READ_HEAD(ring) & HEAD_ADDR;
 713                 ring->space = ring->head - (ring->tail + 8);
 714                 if (ring->space < 0)
 715                         ring->space += ring->size;
 716                 if (ring->space >= n) {
 717                         trace_i915_ring_wait_end(dev);
 718                         return 0;
 719                 }
 720
 721                 if (dev->primary->master) {
 722                         struct drm_i915_master_private *master_priv = dev->primary->master->driver_priv;
 723                         if (master_priv->sarea_priv)
 724                                 master_priv->sarea_priv->perf_boxes |= I915_BOX_WAIT;
 725                 }
 726
 727                 msleep(1);
 728                 if (atomic_read(&dev_priv->mm.wedged))
 729                         return -EAGAIN;
 730         } while (!time_after(jiffies, end));
 731         trace_i915_ring_wait_end (dev);
 732         return -EBUSY;
 733 }
 734
 735 int intel_ring_begin(struct intel_ring_buffer *ring,
 736                      int num_dwords)
 737 {
 738         int n = 4*num_dwords;
 739         int ret;
 740
 741         if (unlikely(ring->tail + n > ring->size)) {
 742                 ret = intel_wrap_ring_buffer(ring);
 743                 if (unlikely(ret))
 744                         return ret;
 745         }
 746
 747         if (unlikely(ring->space < n)) {
 748                 ret = intel_wait_ring_buffer(ring, n);
 749                 if (unlikely(ret))
 750                         return ret;
 751         }
 752
 753         ring->space -= n;
 754         return 0;
 755 }
 756
 757 void intel_ring_advance(struct intel_ring_buffer *ring)
 758 {
 759         ring->tail &= ring->size - 1;
 760         ring->write_tail(ring, ring->tail);
 761 }
 762
 763 static const struct intel_ring_buffer render_ring = {
 764         .name                   = "render ring",
 765         .id                     = RING_RENDER,
 766         .mmio_base              = RENDER_RING_BASE,
 767         .size                   = 32 * PAGE_SIZE,
 768         .init                   = init_render_ring,
 769         .write_tail             = ring_write_tail,
 770         .flush                  = render_ring_flush,
 771         .add_request            = render_ring_add_request,
 772         .get_seqno              = render_ring_get_seqno,
 773         .user_irq_get           = render_ring_get_user_irq,
 774         .user_irq_put           = render_ring_put_user_irq,
 775         .dispatch_execbuffer    = render_ring_dispatch_execbuffer,
 776 };
 777
 778 /* ring buffer for bit-stream decoder */
 779
 780 static const struct intel_ring_buffer bsd_ring = {
 781         .name                   = "bsd ring",
 782         .id                     = RING_BSD,
 783         .mmio_base              = BSD_RING_BASE,
 784         .size                   = 32 * PAGE_SIZE,
 785         .init                   = init_ring_common,
 786         .write_tail             = ring_write_tail,
 787         .flush                  = bsd_ring_flush,
 788         .add_request            = ring_add_request,
 789         .get_seqno              = ring_status_page_get_seqno,
 790         .user_irq_get           = bsd_ring_get_user_irq,
 791         .user_irq_put           = bsd_ring_put_user_irq,
 792         .dispatch_execbuffer    = ring_dispatch_execbuffer,
 793 };
 794
 795
 796 static void gen6_bsd_ring_write_tail(struct intel_ring_buffer *ring,
 797                                      u32 value)
 798 {
 799        drm_i915_private_t *dev_priv = ring->dev->dev_private;
 800
 801        /* Every tail move must follow the sequence below */
 802        I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
 803                GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_MODIFY_MASK |
 804                GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_DISABLE);
 805        I915_WRITE(GEN6_BSD_RNCID, 0x0);
 806
 807        if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL) &
 808                                GEN6_BSD_SLEEP_PSMI_CONTROL_IDLE_INDICATOR) == 0,
 809                        50))
 810                DRM_ERROR("timed out waiting for IDLE Indicator\n");
 811
 812        I915_WRITE_TAIL(ring, value);
 813        I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
 814                GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_MODIFY_MASK |
 815                GEN6_BSD_SLEEP_PSMI_CONTROL_RC_ILDL_MESSAGE_ENABLE);
 816 }
 817
 818 static void gen6_ring_flush(struct intel_ring_buffer *ring,
 819                             u32 invalidate_domains,
 820                             u32 flush_domains)
 821 {
 822         if (intel_ring_begin(ring, 4) == 0) {
 823                 intel_ring_emit(ring, MI_FLUSH_DW);
 824                 intel_ring_emit(ring, 0);
 825                 intel_ring_emit(ring, 0);
 826                 intel_ring_emit(ring, 0);
 827                 intel_ring_advance(ring);
 828         }
 829 }
 830
 831 static int
 832 gen6_ring_dispatch_execbuffer(struct intel_ring_buffer *ring,
 833                               struct drm_i915_gem_execbuffer2 *exec,
 834                               struct drm_clip_rect *cliprects,
 835                               uint64_t exec_offset)
 836 {
 837        uint32_t exec_start;
 838        int ret;
 839
 840        exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
 841
 842        ret = intel_ring_begin(ring, 2);
 843        if (ret)
 844                return ret;
 845
 846        intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_NON_SECURE_I965);
 847        /* bit0-7 is the length on GEN6+ */
 848        intel_ring_emit(ring, exec_start);
 849        intel_ring_advance(ring);
 850
 851        return 0;
 852 }
 853
 854 /* ring buffer for Video Codec for Gen6+ */
 855 static const struct intel_ring_buffer gen6_bsd_ring = {
 856        .name                    = "gen6 bsd ring",
 857        .id                      = RING_BSD,
 858        .mmio_base               = GEN6_BSD_RING_BASE,
 859        .size                    = 32 * PAGE_SIZE,
 860        .init                    = init_ring_common,
 861        .write_tail              = gen6_bsd_ring_write_tail,
 862        .flush                   = gen6_ring_flush,
 863        .add_request             = ring_add_request,
 864        .get_seqno               = ring_status_page_get_seqno,
 865        .user_irq_get            = bsd_ring_get_user_irq,
 866        .user_irq_put            = bsd_ring_put_user_irq,
 867        .dispatch_execbuffer     = gen6_ring_dispatch_execbuffer,
 868 };
 869
 870 /* Blitter support (SandyBridge+) */
 871
 872 static void
 873 blt_ring_get_user_irq(struct intel_ring_buffer *ring)
 874 {
 875         /* do nothing */
 876 }
 877 static void
 878 blt_ring_put_user_irq(struct intel_ring_buffer *ring)
 879 {
 880         /* do nothing */
 881 }
 882
 883
 884 /* Workaround for some stepping of SNB,
 885  * each time when BLT engine ring tail moved,
 886  * the first command in the ring to be parsed
 887  * should be MI_BATCH_BUFFER_START
 888  */
 889 #define NEED_BLT_WORKAROUND(dev) \
 890         (IS_GEN6(dev) && (dev->pdev->revision < 8))
 891
 892 static inline struct drm_i915_gem_object *
 893 to_blt_workaround(struct intel_ring_buffer *ring)
 894 {
 895         return ring->private;
 896 }
 897
 898 static int blt_ring_init(struct intel_ring_buffer *ring)
 899 {
 900         if (NEED_BLT_WORKAROUND(ring->dev)) {
 901                 struct drm_i915_gem_object *obj;
 902                 u32 *ptr;
 903                 int ret;
 904
 905                 obj = to_intel_bo(i915_gem_alloc_object(ring->dev, 4096));
 906                 if (obj == NULL)
 907                         return -ENOMEM;
 908
 909                 ret = i915_gem_object_pin(&obj->base, 4096, true);
 910                 if (ret) {
 911                         drm_gem_object_unreference(&obj->base);
 912                         return ret;
 913                 }
 914
 915                 ptr = kmap(obj->pages[0]);
 916                 *ptr++ = MI_BATCH_BUFFER_END;
 917                 *ptr++ = MI_NOOP;
 918                 kunmap(obj->pages[0]);
 919
 920                 ret = i915_gem_object_set_to_gtt_domain(&obj->base, false);
 921                 if (ret) {
 922                         i915_gem_object_unpin(&obj->base);
 923                         drm_gem_object_unreference(&obj->base);
 924                         return ret;
 925                 }
 926
 927                 ring->private = obj;
 928         }
 929
 930         return init_ring_common(ring);
 931 }
 932
 933 static int blt_ring_begin(struct intel_ring_buffer *ring,
 934                           int num_dwords)
 935 {
 936         if (ring->private) {
 937                 int ret = intel_ring_begin(ring, num_dwords+2);
 938                 if (ret)
 939                         return ret;
 940
 941                 intel_ring_emit(ring, MI_BATCH_BUFFER_START);
 942                 intel_ring_emit(ring, to_blt_workaround(ring)->gtt_offset);
 943
 944                 return 0;
 945         } else
 946                 return intel_ring_begin(ring, 4);
 947 }
 948
 949 static void blt_ring_flush(struct intel_ring_buffer *ring,
 950                            u32 invalidate_domains,
 951                            u32 flush_domains)
 952 {
 953         if (blt_ring_begin(ring, 4) == 0) {
 954                 intel_ring_emit(ring, MI_FLUSH_DW);
 955                 intel_ring_emit(ring, 0);
 956                 intel_ring_emit(ring, 0);
 957                 intel_ring_emit(ring, 0);
 958                 intel_ring_advance(ring);
 959         }
 960 }
 961
 962 static int
 963 blt_ring_add_request(struct intel_ring_buffer *ring,
 964                      u32 *result)
 965 {
 966         u32 seqno;
 967         int ret;
 968
 969         ret = blt_ring_begin(ring, 4);
 970         if (ret)
 971                 return ret;
 972
 973         seqno = i915_gem_get_seqno(ring->dev);
 974
 975         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
 976         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 977         intel_ring_emit(ring, seqno);
 978         intel_ring_emit(ring, MI_USER_INTERRUPT);
 979         intel_ring_advance(ring);
 980
 981         DRM_DEBUG_DRIVER("%s %d\n", ring->name, seqno);
 982         *result = seqno;
 983         return 0;
 984 }
 985
 986 static void blt_ring_cleanup(struct intel_ring_buffer *ring)
 987 {
 988         if (!ring->private)
 989                 return;
 990
 991         i915_gem_object_unpin(ring->private);
 992         drm_gem_object_unreference(ring->private);
 993         ring->private = NULL;
 994 }
 995
 996 static const struct intel_ring_buffer gen6_blt_ring = {
 997        .name                    = "blt ring",
 998        .id                      = RING_BLT,
 999        .mmio_base               = BLT_RING_BASE,
1000        .size                    = 32 * PAGE_SIZE,
1001        .init                    = blt_ring_init,
1002        .write_tail              = ring_write_tail,
1003        .flush                   = blt_ring_flush,
1004        .add_request             = blt_ring_add_request,
1005        .get_seqno               = ring_status_page_get_seqno,
1006        .user_irq_get            = blt_ring_get_user_irq,
1007        .user_irq_put            = blt_ring_put_user_irq,
1008        .dispatch_execbuffer     = gen6_ring_dispatch_execbuffer,
1009        .cleanup                 = blt_ring_cleanup,
1010 };
1011
1012 int intel_init_render_ring_buffer(struct drm_device *dev)
1013 {
1014         drm_i915_private_t *dev_priv = dev->dev_private;
1015
1016         dev_priv->render_ring = render_ring;
1017
1018         if (!I915_NEED_GFX_HWS(dev)) {
1019                 dev_priv->render_ring.status_page.page_addr
1020                         = dev_priv->status_page_dmah->vaddr;
1021                 memset(dev_priv->render_ring.status_page.page_addr,
1022                                 0, PAGE_SIZE);
1023         }
1024
1025         return intel_init_ring_buffer(dev, &dev_priv->render_ring);
1026 }
1027
1028 int intel_init_bsd_ring_buffer(struct drm_device *dev)
1029 {
1030         drm_i915_private_t *dev_priv = dev->dev_private;
1031
1032         if (IS_GEN6(dev))
1033                 dev_priv->bsd_ring = gen6_bsd_ring;
1034         else
1035                 dev_priv->bsd_ring = bsd_ring;
1036
1037         return intel_init_ring_buffer(dev, &dev_priv->bsd_ring);
1038 }
1039
1040 int intel_init_blt_ring_buffer(struct drm_device *dev)
1041 {
1042         drm_i915_private_t *dev_priv = dev->dev_private;
1043
1044         dev_priv->blt_ring = gen6_blt_ring;
1045
1046         return intel_init_ring_buffer(dev, &dev_priv->blt_ring);
1047 }