drivers/gpu/drm/i915/i915_gem.c

   1 /*
   2  * Copyright © 2008 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  *
  26  */
  27
  28 #include "drmP.h"
  29 #include "drm.h"
  30 #include "i915_drm.h"
  31 #include "i915_drv.h"
  32 #include "i915_trace.h"
  33 #include "intel_drv.h"
  34 #include <linux/slab.h>
  35 #include <linux/swap.h>
  36 #include <linux/pci.h>
  37 #include <linux/intel-gtt.h>
  38
  39 static uint32_t i915_gem_get_gtt_alignment(struct drm_gem_object *obj);
  40
  41 static int i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj,
  42                                                   bool pipelined);
  43 static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj);
  44 static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj);
  45 static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj,
  46                                              int write);
  47 static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
  48                                                      uint64_t offset,
  49                                                      uint64_t size);
  50 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj);
  51 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj,
  52                                           bool interruptible);
  53 static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj,
  54                                            unsigned alignment);
  55 static void i915_gem_clear_fence_reg(struct drm_gem_object *obj);
  56 static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
  57                                 struct drm_i915_gem_pwrite *args,
  58                                 struct drm_file *file_priv);
  59 static void i915_gem_free_object_tail(struct drm_gem_object *obj);
  60
  61 static int
  62 i915_gem_object_get_pages(struct drm_gem_object *obj,
  63                           gfp_t gfpmask);
  64
  65 static void
  66 i915_gem_object_put_pages(struct drm_gem_object *obj);
  67
  68 static LIST_HEAD(shrink_list);
  69 static DEFINE_SPINLOCK(shrink_list_lock);
  70
  71 /* some bookkeeping */
  72 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
  73                                   size_t size)
  74 {
  75         dev_priv->mm.object_count++;
  76         dev_priv->mm.object_memory += size;
  77 }
  78
  79 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
  80                                      size_t size)
  81 {
  82         dev_priv->mm.object_count--;
  83         dev_priv->mm.object_memory -= size;
  84 }
  85
  86 static void i915_gem_info_add_gtt(struct drm_i915_private *dev_priv,
  87                                   size_t size)
  88 {
  89         dev_priv->mm.gtt_count++;
  90         dev_priv->mm.gtt_memory += size;
  91 }
  92
  93 static void i915_gem_info_remove_gtt(struct drm_i915_private *dev_priv,
  94                                      size_t size)
  95 {
  96         dev_priv->mm.gtt_count--;
  97         dev_priv->mm.gtt_memory -= size;
  98 }
  99
 100 static void i915_gem_info_add_pin(struct drm_i915_private *dev_priv,
 101                                   size_t size)
 102 {
 103         dev_priv->mm.pin_count++;
 104         dev_priv->mm.pin_memory += size;
 105 }
 106
 107 static void i915_gem_info_remove_pin(struct drm_i915_private *dev_priv,
 108                                      size_t size)
 109 {
 110         dev_priv->mm.pin_count--;
 111         dev_priv->mm.pin_memory -= size;
 112 }
 113
 114 int
 115 i915_gem_check_is_wedged(struct drm_device *dev)
 116 {
 117         struct drm_i915_private *dev_priv = dev->dev_private;
 118         struct completion *x = &dev_priv->error_completion;
 119         unsigned long flags;
 120         int ret;
 121
 122         if (!atomic_read(&dev_priv->mm.wedged))
 123                 return 0;
 124
 125         ret = wait_for_completion_interruptible(x);
 126         if (ret)
 127                 return ret;
 128
 129         /* Success, we reset the GPU! */
 130         if (!atomic_read(&dev_priv->mm.wedged))
 131                 return 0;
 132
 133         /* GPU is hung, bump the completion count to account for
 134          * the token we just consumed so that we never hit zero and
 135          * end up waiting upon a subsequent completion event that
 136          * will never happen.
 137          */
 138         spin_lock_irqsave(&x->wait.lock, flags);
 139         x->done++;
 140         spin_unlock_irqrestore(&x->wait.lock, flags);
 141         return -EIO;
 142 }
 143
 144 static int i915_mutex_lock_interruptible(struct drm_device *dev)
 145 {
 146         struct drm_i915_private *dev_priv = dev->dev_private;
 147         int ret;
 148
 149         ret = i915_gem_check_is_wedged(dev);
 150         if (ret)
 151                 return ret;
 152
 153         ret = mutex_lock_interruptible(&dev->struct_mutex);
 154         if (ret)
 155                 return ret;
 156
 157         if (atomic_read(&dev_priv->mm.wedged)) {
 158                 mutex_unlock(&dev->struct_mutex);
 159                 return -EAGAIN;
 160         }
 161
 162         WARN_ON(i915_verify_lists(dev));
 163         return 0;
 164 }
 165
 166 static inline bool
 167 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj_priv)
 168 {
 169         return obj_priv->gtt_space &&
 170                 !obj_priv->active &&
 171                 obj_priv->pin_count == 0;
 172 }
 173
 174 int i915_gem_do_init(struct drm_device *dev,
 175                      unsigned long start,
 176                      unsigned long end)
 177 {
 178         drm_i915_private_t *dev_priv = dev->dev_private;
 179
 180         if (start >= end ||
 181             (start & (PAGE_SIZE - 1)) != 0 ||
 182             (end & (PAGE_SIZE - 1)) != 0) {
 183                 return -EINVAL;
 184         }
 185
 186         drm_mm_init(&dev_priv->mm.gtt_space, start,
 187                     end - start);
 188
 189         dev_priv->mm.gtt_total = end - start;
 190
 191         return 0;
 192 }
 193
 194 int
 195 i915_gem_init_ioctl(struct drm_device *dev, void *data,
 196                     struct drm_file *file_priv)
 197 {
 198         struct drm_i915_gem_init *args = data;
 199         int ret;
 200
 201         mutex_lock(&dev->struct_mutex);
 202         ret = i915_gem_do_init(dev, args->gtt_start, args->gtt_end);
 203         mutex_unlock(&dev->struct_mutex);
 204
 205         return ret;
 206 }
 207
 208 int
 209 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
 210                             struct drm_file *file_priv)
 211 {
 212         struct drm_i915_private *dev_priv = dev->dev_private;
 213         struct drm_i915_gem_get_aperture *args = data;
 214
 215         if (!(dev->driver->driver_features & DRIVER_GEM))
 216                 return -ENODEV;
 217
 218         mutex_lock(&dev->struct_mutex);
 219         args->aper_size = dev_priv->mm.gtt_total;
 220         args->aper_available_size = args->aper_size - dev_priv->mm.pin_memory;
 221         mutex_unlock(&dev->struct_mutex);
 222
 223         return 0;
 224 }
 225
 226
 227 /**
 228  * Creates a new mm object and returns a handle to it.
 229  */
 230 int
 231 i915_gem_create_ioctl(struct drm_device *dev, void *data,
 232                       struct drm_file *file_priv)
 233 {
 234         struct drm_i915_gem_create *args = data;
 235         struct drm_gem_object *obj;
 236         int ret;
 237         u32 handle;
 238
 239         args->size = roundup(args->size, PAGE_SIZE);
 240
 241         /* Allocate the new object */
 242         obj = i915_gem_alloc_object(dev, args->size);
 243         if (obj == NULL)
 244                 return -ENOMEM;
 245
 246         ret = drm_gem_handle_create(file_priv, obj, &handle);
 247         if (ret) {
 248                 drm_gem_object_release(obj);
 249                 i915_gem_info_remove_obj(dev->dev_private, obj->size);
 250                 kfree(obj);
 251                 return ret;
 252         }
 253
 254         /* drop reference from allocate - handle holds it now */
 255         drm_gem_object_unreference(obj);
 256         trace_i915_gem_object_create(obj);
 257
 258         args->handle = handle;
 259         return 0;
 260 }
 261
 262 static inline int
 263 fast_shmem_read(struct page **pages,
 264                 loff_t page_base, int page_offset,
 265                 char __user *data,
 266                 int length)
 267 {
 268         char *vaddr;
 269         int ret;
 270
 271         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
 272         ret = __copy_to_user_inatomic(data, vaddr + page_offset, length);
 273         kunmap_atomic(vaddr, KM_USER0);
 274
 275         return ret;
 276 }
 277
 278 static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj)
 279 {
 280         drm_i915_private_t *dev_priv = obj->dev->dev_private;
 281         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 282
 283         return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
 284                 obj_priv->tiling_mode != I915_TILING_NONE;
 285 }
 286
 287 static inline void
 288 slow_shmem_copy(struct page *dst_page,
 289                 int dst_offset,
 290                 struct page *src_page,
 291                 int src_offset,
 292                 int length)
 293 {
 294         char *dst_vaddr, *src_vaddr;
 295
 296         dst_vaddr = kmap(dst_page);
 297         src_vaddr = kmap(src_page);
 298
 299         memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
 300
 301         kunmap(src_page);
 302         kunmap(dst_page);
 303 }
 304
 305 static inline void
 306 slow_shmem_bit17_copy(struct page *gpu_page,
 307                       int gpu_offset,
 308                       struct page *cpu_page,
 309                       int cpu_offset,
 310                       int length,
 311                       int is_read)
 312 {
 313         char *gpu_vaddr, *cpu_vaddr;
 314
 315         /* Use the unswizzled path if this page isn't affected. */
 316         if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
 317                 if (is_read)
 318                         return slow_shmem_copy(cpu_page, cpu_offset,
 319                                                gpu_page, gpu_offset, length);
 320                 else
 321                         return slow_shmem_copy(gpu_page, gpu_offset,
 322                                                cpu_page, cpu_offset, length);
 323         }
 324
 325         gpu_vaddr = kmap(gpu_page);
 326         cpu_vaddr = kmap(cpu_page);
 327
 328         /* Copy the data, XORing A6 with A17 (1). The user already knows he's
 329          * XORing with the other bits (A9 for Y, A9 and A10 for X)
 330          */
 331         while (length > 0) {
 332                 int cacheline_end = ALIGN(gpu_offset + 1, 64);
 333                 int this_length = min(cacheline_end - gpu_offset, length);
 334                 int swizzled_gpu_offset = gpu_offset ^ 64;
 335
 336                 if (is_read) {
 337                         memcpy(cpu_vaddr + cpu_offset,
 338                                gpu_vaddr + swizzled_gpu_offset,
 339                                this_length);
 340                 } else {
 341                         memcpy(gpu_vaddr + swizzled_gpu_offset,
 342                                cpu_vaddr + cpu_offset,
 343                                this_length);
 344                 }
 345                 cpu_offset += this_length;
 346                 gpu_offset += this_length;
 347                 length -= this_length;
 348         }
 349
 350         kunmap(cpu_page);
 351         kunmap(gpu_page);
 352 }
 353
 354 /**
 355  * This is the fast shmem pread path, which attempts to copy_from_user directly
 356  * from the backing pages of the object to the user's address space.  On a
 357  * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
 358  */
 359 static int
 360 i915_gem_shmem_pread_fast(struct drm_device *dev, struct drm_gem_object *obj,
 361                           struct drm_i915_gem_pread *args,
 362                           struct drm_file *file_priv)
 363 {
 364         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 365         ssize_t remain;
 366         loff_t offset, page_base;
 367         char __user *user_data;
 368         int page_offset, page_length;
 369
 370         user_data = (char __user *) (uintptr_t) args->data_ptr;
 371         remain = args->size;
 372
 373         obj_priv = to_intel_bo(obj);
 374         offset = args->offset;
 375
 376         while (remain > 0) {
 377                 /* Operation in this page
 378                  *
 379                  * page_base = page offset within aperture
 380                  * page_offset = offset within page
 381                  * page_length = bytes to copy for this page
 382                  */
 383                 page_base = (offset & ~(PAGE_SIZE-1));
 384                 page_offset = offset & (PAGE_SIZE-1);
 385                 page_length = remain;
 386                 if ((page_offset + remain) > PAGE_SIZE)
 387                         page_length = PAGE_SIZE - page_offset;
 388
 389                 if (fast_shmem_read(obj_priv->pages,
 390                                     page_base, page_offset,
 391                                     user_data, page_length))
 392                         return -EFAULT;
 393
 394                 remain -= page_length;
 395                 user_data += page_length;
 396                 offset += page_length;
 397         }
 398
 399         return 0;
 400 }
 401
 402 static int
 403 i915_gem_object_get_pages_or_evict(struct drm_gem_object *obj)
 404 {
 405         int ret;
 406
 407         ret = i915_gem_object_get_pages(obj, __GFP_NORETRY | __GFP_NOWARN);
 408
 409         /* If we've insufficient memory to map in the pages, attempt
 410          * to make some space by throwing out some old buffers.
 411          */
 412         if (ret == -ENOMEM) {
 413                 struct drm_device *dev = obj->dev;
 414
 415                 ret = i915_gem_evict_something(dev, obj->size,
 416                                                i915_gem_get_gtt_alignment(obj));
 417                 if (ret)
 418                         return ret;
 419
 420                 ret = i915_gem_object_get_pages(obj, 0);
 421         }
 422
 423         return ret;
 424 }
 425
 426 /**
 427  * This is the fallback shmem pread path, which allocates temporary storage
 428  * in kernel space to copy_to_user into outside of the struct_mutex, so we
 429  * can copy out of the object's backing pages while holding the struct mutex
 430  * and not take page faults.
 431  */
 432 static int
 433 i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj,
 434                           struct drm_i915_gem_pread *args,
 435                           struct drm_file *file_priv)
 436 {
 437         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 438         struct mm_struct *mm = current->mm;
 439         struct page **user_pages;
 440         ssize_t remain;
 441         loff_t offset, pinned_pages, i;
 442         loff_t first_data_page, last_data_page, num_pages;
 443         int shmem_page_index, shmem_page_offset;
 444         int data_page_index,  data_page_offset;
 445         int page_length;
 446         int ret;
 447         uint64_t data_ptr = args->data_ptr;
 448         int do_bit17_swizzling;
 449
 450         remain = args->size;
 451
 452         /* Pin the user pages containing the data.  We can't fault while
 453          * holding the struct mutex, yet we want to hold it while
 454          * dereferencing the user data.
 455          */
 456         first_data_page = data_ptr / PAGE_SIZE;
 457         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 458         num_pages = last_data_page - first_data_page + 1;
 459
 460         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 461         if (user_pages == NULL)
 462                 return -ENOMEM;
 463
 464         mutex_unlock(&dev->struct_mutex);
 465         down_read(&mm->mmap_sem);
 466         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 467                                       num_pages, 1, 0, user_pages, NULL);
 468         up_read(&mm->mmap_sem);
 469         mutex_lock(&dev->struct_mutex);
 470         if (pinned_pages < num_pages) {
 471                 ret = -EFAULT;
 472                 goto out;
 473         }
 474
 475         ret = i915_gem_object_set_cpu_read_domain_range(obj,
 476                                                         args->offset,
 477                                                         args->size);
 478         if (ret)
 479                 goto out;
 480
 481         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 482
 483         obj_priv = to_intel_bo(obj);
 484         offset = args->offset;
 485
 486         while (remain > 0) {
 487                 /* Operation in this page
 488                  *
 489                  * shmem_page_index = page number within shmem file
 490                  * shmem_page_offset = offset within page in shmem file
 491                  * data_page_index = page number in get_user_pages return
 492                  * data_page_offset = offset with data_page_index page.
 493                  * page_length = bytes to copy for this page
 494                  */
 495                 shmem_page_index = offset / PAGE_SIZE;
 496                 shmem_page_offset = offset & ~PAGE_MASK;
 497                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 498                 data_page_offset = data_ptr & ~PAGE_MASK;
 499
 500                 page_length = remain;
 501                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 502                         page_length = PAGE_SIZE - shmem_page_offset;
 503                 if ((data_page_offset + page_length) > PAGE_SIZE)
 504                         page_length = PAGE_SIZE - data_page_offset;
 505
 506                 if (do_bit17_swizzling) {
 507                         slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
 508                                               shmem_page_offset,
 509                                               user_pages[data_page_index],
 510                                               data_page_offset,
 511                                               page_length,
 512                                               1);
 513                 } else {
 514                         slow_shmem_copy(user_pages[data_page_index],
 515                                         data_page_offset,
 516                                         obj_priv->pages[shmem_page_index],
 517                                         shmem_page_offset,
 518                                         page_length);
 519                 }
 520
 521                 remain -= page_length;
 522                 data_ptr += page_length;
 523                 offset += page_length;
 524         }
 525
 526 out:
 527         for (i = 0; i < pinned_pages; i++) {
 528                 SetPageDirty(user_pages[i]);
 529                 page_cache_release(user_pages[i]);
 530         }
 531         drm_free_large(user_pages);
 532
 533         return ret;
 534 }
 535
 536 /**
 537  * Reads data from the object referenced by handle.
 538  *
 539  * On error, the contents of *data are undefined.
 540  */
 541 int
 542 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 543                      struct drm_file *file_priv)
 544 {
 545         struct drm_i915_gem_pread *args = data;
 546         struct drm_gem_object *obj;
 547         struct drm_i915_gem_object *obj_priv;
 548         int ret = 0;
 549
 550         ret = i915_mutex_lock_interruptible(dev);
 551         if (ret)
 552                 return ret;
 553
 554         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
 555         if (obj == NULL) {
 556                 ret = -ENOENT;
 557                 goto unlock;
 558         }
 559         obj_priv = to_intel_bo(obj);
 560
 561         /* Bounds check source.  */
 562         if (args->offset > obj->size || args->size > obj->size - args->offset) {
 563                 ret = -EINVAL;
 564                 goto out;
 565         }
 566
 567         if (args->size == 0)
 568                 goto out;
 569
 570         if (!access_ok(VERIFY_WRITE,
 571                        (char __user *)(uintptr_t)args->data_ptr,
 572                        args->size)) {
 573                 ret = -EFAULT;
 574                 goto out;
 575         }
 576
 577         ret = fault_in_pages_writeable((char __user *)(uintptr_t)args->data_ptr,
 578                                        args->size);
 579         if (ret) {
 580                 ret = -EFAULT;
 581                 goto out;
 582         }
 583
 584         ret = i915_gem_object_get_pages_or_evict(obj);
 585         if (ret)
 586                 goto out;
 587
 588         ret = i915_gem_object_set_cpu_read_domain_range(obj,
 589                                                         args->offset,
 590                                                         args->size);
 591         if (ret)
 592                 goto out_put;
 593
 594         ret = -EFAULT;
 595         if (!i915_gem_object_needs_bit17_swizzle(obj))
 596                 ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv);
 597         if (ret == -EFAULT)
 598                 ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv);
 599
 600 out_put:
 601         i915_gem_object_put_pages(obj);
 602 out:
 603         drm_gem_object_unreference(obj);
 604 unlock:
 605         mutex_unlock(&dev->struct_mutex);
 606         return ret;
 607 }
 608
 609 /* This is the fast write path which cannot handle
 610  * page faults in the source data
 611  */
 612
 613 static inline int
 614 fast_user_write(struct io_mapping *mapping,
 615                 loff_t page_base, int page_offset,
 616                 char __user *user_data,
 617                 int length)
 618 {
 619         char *vaddr_atomic;
 620         unsigned long unwritten;
 621
 622         vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base, KM_USER0);
 623         unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
 624                                                       user_data, length);
 625         io_mapping_unmap_atomic(vaddr_atomic, KM_USER0);
 626         return unwritten;
 627 }
 628
 629 /* Here's the write path which can sleep for
 630  * page faults
 631  */
 632
 633 static inline void
 634 slow_kernel_write(struct io_mapping *mapping,
 635                   loff_t gtt_base, int gtt_offset,
 636                   struct page *user_page, int user_offset,
 637                   int length)
 638 {
 639         char __iomem *dst_vaddr;
 640         char *src_vaddr;
 641
 642         dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
 643         src_vaddr = kmap(user_page);
 644
 645         memcpy_toio(dst_vaddr + gtt_offset,
 646                     src_vaddr + user_offset,
 647                     length);
 648
 649         kunmap(user_page);
 650         io_mapping_unmap(dst_vaddr);
 651 }
 652
 653 static inline int
 654 fast_shmem_write(struct page **pages,
 655                  loff_t page_base, int page_offset,
 656                  char __user *data,
 657                  int length)
 658 {
 659         char *vaddr;
 660         int ret;
 661
 662         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
 663         ret = __copy_from_user_inatomic(vaddr + page_offset, data, length);
 664         kunmap_atomic(vaddr, KM_USER0);
 665
 666         return ret;
 667 }
 668
 669 /**
 670  * This is the fast pwrite path, where we copy the data directly from the
 671  * user into the GTT, uncached.
 672  */
 673 static int
 674 i915_gem_gtt_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
 675                          struct drm_i915_gem_pwrite *args,
 676                          struct drm_file *file_priv)
 677 {
 678         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 679         drm_i915_private_t *dev_priv = dev->dev_private;
 680         ssize_t remain;
 681         loff_t offset, page_base;
 682         char __user *user_data;
 683         int page_offset, page_length;
 684
 685         user_data = (char __user *) (uintptr_t) args->data_ptr;
 686         remain = args->size;
 687
 688         obj_priv = to_intel_bo(obj);
 689         offset = obj_priv->gtt_offset + args->offset;
 690
 691         while (remain > 0) {
 692                 /* Operation in this page
 693                  *
 694                  * page_base = page offset within aperture
 695                  * page_offset = offset within page
 696                  * page_length = bytes to copy for this page
 697                  */
 698                 page_base = (offset & ~(PAGE_SIZE-1));
 699                 page_offset = offset & (PAGE_SIZE-1);
 700                 page_length = remain;
 701                 if ((page_offset + remain) > PAGE_SIZE)
 702                         page_length = PAGE_SIZE - page_offset;
 703
 704                 /* If we get a fault while copying data, then (presumably) our
 705                  * source page isn't available.  Return the error and we'll
 706                  * retry in the slow path.
 707                  */
 708                 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
 709                                     page_offset, user_data, page_length))
 710
 711                         return -EFAULT;
 712
 713                 remain -= page_length;
 714                 user_data += page_length;
 715                 offset += page_length;
 716         }
 717
 718         return 0;
 719 }
 720
 721 /**
 722  * This is the fallback GTT pwrite path, which uses get_user_pages to pin
 723  * the memory and maps it using kmap_atomic for copying.
 724  *
 725  * This code resulted in x11perf -rgb10text consuming about 10% more CPU
 726  * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
 727  */
 728 static int
 729 i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
 730                          struct drm_i915_gem_pwrite *args,
 731                          struct drm_file *file_priv)
 732 {
 733         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 734         drm_i915_private_t *dev_priv = dev->dev_private;
 735         ssize_t remain;
 736         loff_t gtt_page_base, offset;
 737         loff_t first_data_page, last_data_page, num_pages;
 738         loff_t pinned_pages, i;
 739         struct page **user_pages;
 740         struct mm_struct *mm = current->mm;
 741         int gtt_page_offset, data_page_offset, data_page_index, page_length;
 742         int ret;
 743         uint64_t data_ptr = args->data_ptr;
 744
 745         remain = args->size;
 746
 747         /* Pin the user pages containing the data.  We can't fault while
 748          * holding the struct mutex, and all of the pwrite implementations
 749          * want to hold it while dereferencing the user data.
 750          */
 751         first_data_page = data_ptr / PAGE_SIZE;
 752         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 753         num_pages = last_data_page - first_data_page + 1;
 754
 755         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 756         if (user_pages == NULL)
 757                 return -ENOMEM;
 758
 759         mutex_unlock(&dev->struct_mutex);
 760         down_read(&mm->mmap_sem);
 761         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 762                                       num_pages, 0, 0, user_pages, NULL);
 763         up_read(&mm->mmap_sem);
 764         mutex_lock(&dev->struct_mutex);
 765         if (pinned_pages < num_pages) {
 766                 ret = -EFAULT;
 767                 goto out_unpin_pages;
 768         }
 769
 770         ret = i915_gem_object_set_to_gtt_domain(obj, 1);
 771         if (ret)
 772                 goto out_unpin_pages;
 773
 774         obj_priv = to_intel_bo(obj);
 775         offset = obj_priv->gtt_offset + args->offset;
 776
 777         while (remain > 0) {
 778                 /* Operation in this page
 779                  *
 780                  * gtt_page_base = page offset within aperture
 781                  * gtt_page_offset = offset within page in aperture
 782                  * data_page_index = page number in get_user_pages return
 783                  * data_page_offset = offset with data_page_index page.
 784                  * page_length = bytes to copy for this page
 785                  */
 786                 gtt_page_base = offset & PAGE_MASK;
 787                 gtt_page_offset = offset & ~PAGE_MASK;
 788                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 789                 data_page_offset = data_ptr & ~PAGE_MASK;
 790
 791                 page_length = remain;
 792                 if ((gtt_page_offset + page_length) > PAGE_SIZE)
 793                         page_length = PAGE_SIZE - gtt_page_offset;
 794                 if ((data_page_offset + page_length) > PAGE_SIZE)
 795                         page_length = PAGE_SIZE - data_page_offset;
 796
 797                 slow_kernel_write(dev_priv->mm.gtt_mapping,
 798                                   gtt_page_base, gtt_page_offset,
 799                                   user_pages[data_page_index],
 800                                   data_page_offset,
 801                                   page_length);
 802
 803                 remain -= page_length;
 804                 offset += page_length;
 805                 data_ptr += page_length;
 806         }
 807
 808 out_unpin_pages:
 809         for (i = 0; i < pinned_pages; i++)
 810                 page_cache_release(user_pages[i]);
 811         drm_free_large(user_pages);
 812
 813         return ret;
 814 }
 815
 816 /**
 817  * This is the fast shmem pwrite path, which attempts to directly
 818  * copy_from_user into the kmapped pages backing the object.
 819  */
 820 static int
 821 i915_gem_shmem_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
 822                            struct drm_i915_gem_pwrite *args,
 823                            struct drm_file *file_priv)
 824 {
 825         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 826         ssize_t remain;
 827         loff_t offset, page_base;
 828         char __user *user_data;
 829         int page_offset, page_length;
 830
 831         user_data = (char __user *) (uintptr_t) args->data_ptr;
 832         remain = args->size;
 833
 834         obj_priv = to_intel_bo(obj);
 835         offset = args->offset;
 836         obj_priv->dirty = 1;
 837
 838         while (remain > 0) {
 839                 /* Operation in this page
 840                  *
 841                  * page_base = page offset within aperture
 842                  * page_offset = offset within page
 843                  * page_length = bytes to copy for this page
 844                  */
 845                 page_base = (offset & ~(PAGE_SIZE-1));
 846                 page_offset = offset & (PAGE_SIZE-1);
 847                 page_length = remain;
 848                 if ((page_offset + remain) > PAGE_SIZE)
 849                         page_length = PAGE_SIZE - page_offset;
 850
 851                 if (fast_shmem_write(obj_priv->pages,
 852                                        page_base, page_offset,
 853                                        user_data, page_length))
 854                         return -EFAULT;
 855
 856                 remain -= page_length;
 857                 user_data += page_length;
 858                 offset += page_length;
 859         }
 860
 861         return 0;
 862 }
 863
 864 /**
 865  * This is the fallback shmem pwrite path, which uses get_user_pages to pin
 866  * the memory and maps it using kmap_atomic for copying.
 867  *
 868  * This avoids taking mmap_sem for faulting on the user's address while the
 869  * struct_mutex is held.
 870  */
 871 static int
 872 i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
 873                            struct drm_i915_gem_pwrite *args,
 874                            struct drm_file *file_priv)
 875 {
 876         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
 877         struct mm_struct *mm = current->mm;
 878         struct page **user_pages;
 879         ssize_t remain;
 880         loff_t offset, pinned_pages, i;
 881         loff_t first_data_page, last_data_page, num_pages;
 882         int shmem_page_index, shmem_page_offset;
 883         int data_page_index,  data_page_offset;
 884         int page_length;
 885         int ret;
 886         uint64_t data_ptr = args->data_ptr;
 887         int do_bit17_swizzling;
 888
 889         remain = args->size;
 890
 891         /* Pin the user pages containing the data.  We can't fault while
 892          * holding the struct mutex, and all of the pwrite implementations
 893          * want to hold it while dereferencing the user data.
 894          */
 895         first_data_page = data_ptr / PAGE_SIZE;
 896         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 897         num_pages = last_data_page - first_data_page + 1;
 898
 899         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 900         if (user_pages == NULL)
 901                 return -ENOMEM;
 902
 903         mutex_unlock(&dev->struct_mutex);
 904         down_read(&mm->mmap_sem);
 905         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 906                                       num_pages, 0, 0, user_pages, NULL);
 907         up_read(&mm->mmap_sem);
 908         mutex_lock(&dev->struct_mutex);
 909         if (pinned_pages < num_pages) {
 910                 ret = -EFAULT;
 911                 goto out;
 912         }
 913
 914         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
 915         if (ret)
 916                 goto out;
 917
 918         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 919
 920         obj_priv = to_intel_bo(obj);
 921         offset = args->offset;
 922         obj_priv->dirty = 1;
 923
 924         while (remain > 0) {
 925                 /* Operation in this page
 926                  *
 927                  * shmem_page_index = page number within shmem file
 928                  * shmem_page_offset = offset within page in shmem file
 929                  * data_page_index = page number in get_user_pages return
 930                  * data_page_offset = offset with data_page_index page.
 931                  * page_length = bytes to copy for this page
 932                  */
 933                 shmem_page_index = offset / PAGE_SIZE;
 934                 shmem_page_offset = offset & ~PAGE_MASK;
 935                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 936                 data_page_offset = data_ptr & ~PAGE_MASK;
 937
 938                 page_length = remain;
 939                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 940                         page_length = PAGE_SIZE - shmem_page_offset;
 941                 if ((data_page_offset + page_length) > PAGE_SIZE)
 942                         page_length = PAGE_SIZE - data_page_offset;
 943
 944                 if (do_bit17_swizzling) {
 945                         slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
 946                                               shmem_page_offset,
 947                                               user_pages[data_page_index],
 948                                               data_page_offset,
 949                                               page_length,
 950                                               0);
 951                 } else {
 952                         slow_shmem_copy(obj_priv->pages[shmem_page_index],
 953                                         shmem_page_offset,
 954                                         user_pages[data_page_index],
 955                                         data_page_offset,
 956                                         page_length);
 957                 }
 958
 959                 remain -= page_length;
 960                 data_ptr += page_length;
 961                 offset += page_length;
 962         }
 963
 964 out:
 965         for (i = 0; i < pinned_pages; i++)
 966                 page_cache_release(user_pages[i]);
 967         drm_free_large(user_pages);
 968
 969         return ret;
 970 }
 971
 972 /**
 973  * Writes data to the object referenced by handle.
 974  *
 975  * On error, the contents of the buffer that were to be modified are undefined.
 976  */
 977 int
 978 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 979                       struct drm_file *file)
 980 {
 981         struct drm_i915_gem_pwrite *args = data;
 982         struct drm_gem_object *obj;
 983         struct drm_i915_gem_object *obj_priv;
 984         int ret = 0;
 985
 986         ret = i915_mutex_lock_interruptible(dev);
 987         if (ret)
 988                 return ret;
 989
 990         obj = drm_gem_object_lookup(dev, file, args->handle);
 991         if (obj == NULL) {
 992                 ret = -ENOENT;
 993                 goto unlock;
 994         }
 995         obj_priv = to_intel_bo(obj);
 996
 997
 998         /* Bounds check destination. */
 999         if (args->offset > obj->size || args->size > obj->size - args->offset) {
1000                 ret = -EINVAL;
1001                 goto out;
1002         }
1003
1004         if (args->size == 0)
1005                 goto out;
1006
1007         if (!access_ok(VERIFY_READ,
1008                        (char __user *)(uintptr_t)args->data_ptr,
1009                        args->size)) {
1010                 ret = -EFAULT;
1011                 goto out;
1012         }
1013
1014         ret = fault_in_pages_readable((char __user *)(uintptr_t)args->data_ptr,
1015                                       args->size);
1016         if (ret) {
1017                 ret = -EFAULT;
1018                 goto out;
1019         }
1020
1021         /* We can only do the GTT pwrite on untiled buffers, as otherwise
1022          * it would end up going through the fenced access, and we'll get
1023          * different detiling behavior between reading and writing.
1024          * pread/pwrite currently are reading and writing from the CPU
1025          * perspective, requiring manual detiling by the client.
1026          */
1027         if (obj_priv->phys_obj)
1028                 ret = i915_gem_phys_pwrite(dev, obj, args, file);
1029         else if (obj_priv->tiling_mode == I915_TILING_NONE &&
1030                  obj_priv->gtt_space &&
1031                  obj->write_domain != I915_GEM_DOMAIN_CPU) {
1032                 ret = i915_gem_object_pin(obj, 0);
1033                 if (ret)
1034                         goto out;
1035
1036                 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
1037                 if (ret)
1038                         goto out_unpin;
1039
1040                 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1041                 if (ret == -EFAULT)
1042                         ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file);
1043
1044 out_unpin:
1045                 i915_gem_object_unpin(obj);
1046         } else {
1047                 ret = i915_gem_object_get_pages_or_evict(obj);
1048                 if (ret)
1049                         goto out;
1050
1051                 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1052                 if (ret)
1053                         goto out_put;
1054
1055                 ret = -EFAULT;
1056                 if (!i915_gem_object_needs_bit17_swizzle(obj))
1057                         ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file);
1058                 if (ret == -EFAULT)
1059                         ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file);
1060
1061 out_put:
1062                 i915_gem_object_put_pages(obj);
1063         }
1064
1065 out:
1066         drm_gem_object_unreference(obj);
1067 unlock:
1068         mutex_unlock(&dev->struct_mutex);
1069         return ret;
1070 }
1071
1072 /**
1073  * Called when user space prepares to use an object with the CPU, either
1074  * through the mmap ioctl's mapping or a GTT mapping.
1075  */
1076 int
1077 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1078                           struct drm_file *file_priv)
1079 {
1080         struct drm_i915_private *dev_priv = dev->dev_private;
1081         struct drm_i915_gem_set_domain *args = data;
1082         struct drm_gem_object *obj;
1083         struct drm_i915_gem_object *obj_priv;
1084         uint32_t read_domains = args->read_domains;
1085         uint32_t write_domain = args->write_domain;
1086         int ret;
1087
1088         if (!(dev->driver->driver_features & DRIVER_GEM))
1089                 return -ENODEV;
1090
1091         /* Only handle setting domains to types used by the CPU. */
1092         if (write_domain & I915_GEM_GPU_DOMAINS)
1093                 return -EINVAL;
1094
1095         if (read_domains & I915_GEM_GPU_DOMAINS)
1096                 return -EINVAL;
1097
1098         /* Having something in the write domain implies it's in the read
1099          * domain, and only that read domain.  Enforce that in the request.
1100          */
1101         if (write_domain != 0 && read_domains != write_domain)
1102                 return -EINVAL;
1103
1104         ret = i915_mutex_lock_interruptible(dev);
1105         if (ret)
1106                 return ret;
1107
1108         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1109         if (obj == NULL) {
1110                 ret = -ENOENT;
1111                 goto unlock;
1112         }
1113         obj_priv = to_intel_bo(obj);
1114
1115         intel_mark_busy(dev, obj);
1116
1117         if (read_domains & I915_GEM_DOMAIN_GTT) {
1118                 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1119
1120                 /* Update the LRU on the fence for the CPU access that's
1121                  * about to occur.
1122                  */
1123                 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1124                         struct drm_i915_fence_reg *reg =
1125                                 &dev_priv->fence_regs[obj_priv->fence_reg];
1126                         list_move_tail(&reg->lru_list,
1127                                        &dev_priv->mm.fence_list);
1128                 }
1129
1130                 /* Silently promote "you're not bound, there was nothing to do"
1131                  * to success, since the client was just asking us to
1132                  * make sure everything was done.
1133                  */
1134                 if (ret == -EINVAL)
1135                         ret = 0;
1136         } else {
1137                 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1138         }
1139
1140         /* Maintain LRU order of "inactive" objects */
1141         if (ret == 0 && i915_gem_object_is_inactive(obj_priv))
1142                 list_move_tail(&obj_priv->mm_list, &dev_priv->mm.inactive_list);
1143
1144         drm_gem_object_unreference(obj);
1145 unlock:
1146         mutex_unlock(&dev->struct_mutex);
1147         return ret;
1148 }
1149
1150 /**
1151  * Called when user space has done writes to this buffer
1152  */
1153 int
1154 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1155                       struct drm_file *file_priv)
1156 {
1157         struct drm_i915_gem_sw_finish *args = data;
1158         struct drm_gem_object *obj;
1159         int ret = 0;
1160
1161         if (!(dev->driver->driver_features & DRIVER_GEM))
1162                 return -ENODEV;
1163
1164         ret = i915_mutex_lock_interruptible(dev);
1165         if (ret)
1166                 return ret;
1167
1168         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1169         if (obj == NULL) {
1170                 ret = -ENOENT;
1171                 goto unlock;
1172         }
1173
1174         /* Pinned buffers may be scanout, so flush the cache */
1175         if (to_intel_bo(obj)->pin_count)
1176                 i915_gem_object_flush_cpu_write_domain(obj);
1177
1178         drm_gem_object_unreference(obj);
1179 unlock:
1180         mutex_unlock(&dev->struct_mutex);
1181         return ret;
1182 }
1183
1184 /**
1185  * Maps the contents of an object, returning the address it is mapped
1186  * into.
1187  *
1188  * While the mapping holds a reference on the contents of the object, it doesn't
1189  * imply a ref on the object itself.
1190  */
1191 int
1192 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1193                    struct drm_file *file_priv)
1194 {
1195         struct drm_i915_gem_mmap *args = data;
1196         struct drm_gem_object *obj;
1197         loff_t offset;
1198         unsigned long addr;
1199
1200         if (!(dev->driver->driver_features & DRIVER_GEM))
1201                 return -ENODEV;
1202
1203         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1204         if (obj == NULL)
1205                 return -ENOENT;
1206
1207         offset = args->offset;
1208
1209         down_write(&current->mm->mmap_sem);
1210         addr = do_mmap(obj->filp, 0, args->size,
1211                        PROT_READ | PROT_WRITE, MAP_SHARED,
1212                        args->offset);
1213         up_write(&current->mm->mmap_sem);
1214         drm_gem_object_unreference_unlocked(obj);
1215         if (IS_ERR((void *)addr))
1216                 return addr;
1217
1218         args->addr_ptr = (uint64_t) addr;
1219
1220         return 0;
1221 }
1222
1223 /**
1224  * i915_gem_fault - fault a page into the GTT
1225  * vma: VMA in question
1226  * vmf: fault info
1227  *
1228  * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1229  * from userspace.  The fault handler takes care of binding the object to
1230  * the GTT (if needed), allocating and programming a fence register (again,
1231  * only if needed based on whether the old reg is still valid or the object
1232  * is tiled) and inserting a new PTE into the faulting process.
1233  *
1234  * Note that the faulting process may involve evicting existing objects
1235  * from the GTT and/or fence registers to make room.  So performance may
1236  * suffer if the GTT working set is large or there are few fence registers
1237  * left.
1238  */
1239 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1240 {
1241         struct drm_gem_object *obj = vma->vm_private_data;
1242         struct drm_device *dev = obj->dev;
1243         drm_i915_private_t *dev_priv = dev->dev_private;
1244         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1245         pgoff_t page_offset;
1246         unsigned long pfn;
1247         int ret = 0;
1248         bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1249
1250         /* We don't use vmf->pgoff since that has the fake offset */
1251         page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1252                 PAGE_SHIFT;
1253
1254         /* Now bind it into the GTT if needed */
1255         mutex_lock(&dev->struct_mutex);
1256         if (!obj_priv->gtt_space) {
1257                 ret = i915_gem_object_bind_to_gtt(obj, 0);
1258                 if (ret)
1259                         goto unlock;
1260
1261                 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1262                 if (ret)
1263                         goto unlock;
1264         }
1265
1266         /* Need a new fence register? */
1267         if (obj_priv->tiling_mode != I915_TILING_NONE) {
1268                 ret = i915_gem_object_get_fence_reg(obj, true);
1269                 if (ret)
1270                         goto unlock;
1271         }
1272
1273         if (i915_gem_object_is_inactive(obj_priv))
1274                 list_move_tail(&obj_priv->mm_list, &dev_priv->mm.inactive_list);
1275
1276         pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) +
1277                 page_offset;
1278
1279         /* Finally, remap it using the new GTT offset */
1280         ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1281 unlock:
1282         mutex_unlock(&dev->struct_mutex);
1283
1284         switch (ret) {
1285         case 0:
1286         case -ERESTARTSYS:
1287                 return VM_FAULT_NOPAGE;
1288         case -ENOMEM:
1289         case -EAGAIN:
1290                 return VM_FAULT_OOM;
1291         default:
1292                 return VM_FAULT_SIGBUS;
1293         }
1294 }
1295
1296 /**
1297  * i915_gem_create_mmap_offset - create a fake mmap offset for an object
1298  * @obj: obj in question
1299  *
1300  * GEM memory mapping works by handing back to userspace a fake mmap offset
1301  * it can use in a subsequent mmap(2) call.  The DRM core code then looks
1302  * up the object based on the offset and sets up the various memory mapping
1303  * structures.
1304  *
1305  * This routine allocates and attaches a fake offset for @obj.
1306  */
1307 static int
1308 i915_gem_create_mmap_offset(struct drm_gem_object *obj)
1309 {
1310         struct drm_device *dev = obj->dev;
1311         struct drm_gem_mm *mm = dev->mm_private;
1312         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1313         struct drm_map_list *list;
1314         struct drm_local_map *map;
1315         int ret = 0;
1316
1317         /* Set the object up for mmap'ing */
1318         list = &obj->map_list;
1319         list->map = kzalloc(sizeof(struct drm_map_list), GFP_KERNEL);
1320         if (!list->map)
1321                 return -ENOMEM;
1322
1323         map = list->map;
1324         map->type = _DRM_GEM;
1325         map->size = obj->size;
1326         map->handle = obj;
1327
1328         /* Get a DRM GEM mmap offset allocated... */
1329         list->file_offset_node = drm_mm_search_free(&mm->offset_manager,
1330                                                     obj->size / PAGE_SIZE, 0, 0);
1331         if (!list->file_offset_node) {
1332                 DRM_ERROR("failed to allocate offset for bo %d\n", obj->name);
1333                 ret = -ENOSPC;
1334                 goto out_free_list;
1335         }
1336
1337         list->file_offset_node = drm_mm_get_block(list->file_offset_node,
1338                                                   obj->size / PAGE_SIZE, 0);
1339         if (!list->file_offset_node) {
1340                 ret = -ENOMEM;
1341                 goto out_free_list;
1342         }
1343
1344         list->hash.key = list->file_offset_node->start;
1345         ret = drm_ht_insert_item(&mm->offset_hash, &list->hash);
1346         if (ret) {
1347                 DRM_ERROR("failed to add to map hash\n");
1348                 goto out_free_mm;
1349         }
1350
1351         /* By now we should be all set, any drm_mmap request on the offset
1352          * below will get to our mmap & fault handler */
1353         obj_priv->mmap_offset = ((uint64_t) list->hash.key) << PAGE_SHIFT;
1354
1355         return 0;
1356
1357 out_free_mm:
1358         drm_mm_put_block(list->file_offset_node);
1359 out_free_list:
1360         kfree(list->map);
1361
1362         return ret;
1363 }
1364
1365 /**
1366  * i915_gem_release_mmap - remove physical page mappings
1367  * @obj: obj in question
1368  *
1369  * Preserve the reservation of the mmapping with the DRM core code, but
1370  * relinquish ownership of the pages back to the system.
1371  *
1372  * It is vital that we remove the page mapping if we have mapped a tiled
1373  * object through the GTT and then lose the fence register due to
1374  * resource pressure. Similarly if the object has been moved out of the
1375  * aperture, than pages mapped into userspace must be revoked. Removing the
1376  * mapping will then trigger a page fault on the next user access, allowing
1377  * fixup by i915_gem_fault().
1378  */
1379 void
1380 i915_gem_release_mmap(struct drm_gem_object *obj)
1381 {
1382         struct drm_device *dev = obj->dev;
1383         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1384
1385         if (dev->dev_mapping)
1386                 unmap_mapping_range(dev->dev_mapping,
1387                                     obj_priv->mmap_offset, obj->size, 1);
1388 }
1389
1390 static void
1391 i915_gem_free_mmap_offset(struct drm_gem_object *obj)
1392 {
1393         struct drm_device *dev = obj->dev;
1394         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1395         struct drm_gem_mm *mm = dev->mm_private;
1396         struct drm_map_list *list;
1397
1398         list = &obj->map_list;
1399         drm_ht_remove_item(&mm->offset_hash, &list->hash);
1400
1401         if (list->file_offset_node) {
1402                 drm_mm_put_block(list->file_offset_node);
1403                 list->file_offset_node = NULL;
1404         }
1405
1406         if (list->map) {
1407                 kfree(list->map);
1408                 list->map = NULL;
1409         }
1410
1411         obj_priv->mmap_offset = 0;
1412 }
1413
1414 /**
1415  * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1416  * @obj: object to check
1417  *
1418  * Return the required GTT alignment for an object, taking into account
1419  * potential fence register mapping if needed.
1420  */
1421 static uint32_t
1422 i915_gem_get_gtt_alignment(struct drm_gem_object *obj)
1423 {
1424         struct drm_device *dev = obj->dev;
1425         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1426         int start, i;
1427
1428         /*
1429          * Minimum alignment is 4k (GTT page size), but might be greater
1430          * if a fence register is needed for the object.
1431          */
1432         if (INTEL_INFO(dev)->gen >= 4 || obj_priv->tiling_mode == I915_TILING_NONE)
1433                 return 4096;
1434
1435         /*
1436          * Previous chips need to be aligned to the size of the smallest
1437          * fence register that can contain the object.
1438          */
1439         if (INTEL_INFO(dev)->gen == 3)
1440                 start = 1024*1024;
1441         else
1442                 start = 512*1024;
1443
1444         for (i = start; i < obj->size; i <<= 1)
1445                 ;
1446
1447         return i;
1448 }
1449
1450 /**
1451  * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1452  * @dev: DRM device
1453  * @data: GTT mapping ioctl data
1454  * @file_priv: GEM object info
1455  *
1456  * Simply returns the fake offset to userspace so it can mmap it.
1457  * The mmap call will end up in drm_gem_mmap(), which will set things
1458  * up so we can get faults in the handler above.
1459  *
1460  * The fault handler will take care of binding the object into the GTT
1461  * (since it may have been evicted to make room for something), allocating
1462  * a fence register, and mapping the appropriate aperture address into
1463  * userspace.
1464  */
1465 int
1466 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1467                         struct drm_file *file_priv)
1468 {
1469         struct drm_i915_gem_mmap_gtt *args = data;
1470         struct drm_gem_object *obj;
1471         struct drm_i915_gem_object *obj_priv;
1472         int ret;
1473
1474         if (!(dev->driver->driver_features & DRIVER_GEM))
1475                 return -ENODEV;
1476
1477         ret = i915_mutex_lock_interruptible(dev);
1478         if (ret)
1479                 return ret;
1480
1481         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1482         if (obj == NULL) {
1483                 ret = -ENOENT;
1484                 goto unlock;
1485         }
1486         obj_priv = to_intel_bo(obj);
1487
1488         if (obj_priv->madv != I915_MADV_WILLNEED) {
1489                 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1490                 ret = -EINVAL;
1491                 goto out;
1492         }
1493
1494         if (!obj_priv->mmap_offset) {
1495                 ret = i915_gem_create_mmap_offset(obj);
1496                 if (ret)
1497                         goto out;
1498         }
1499
1500         args->offset = obj_priv->mmap_offset;
1501
1502         /*
1503          * Pull it into the GTT so that we have a page list (makes the
1504          * initial fault faster and any subsequent flushing possible).
1505          */
1506         if (!obj_priv->agp_mem) {
1507                 ret = i915_gem_object_bind_to_gtt(obj, 0);
1508                 if (ret)
1509                         goto out;
1510         }
1511
1512 out:
1513         drm_gem_object_unreference(obj);
1514 unlock:
1515         mutex_unlock(&dev->struct_mutex);
1516         return ret;
1517 }
1518
1519 static void
1520 i915_gem_object_put_pages(struct drm_gem_object *obj)
1521 {
1522         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1523         int page_count = obj->size / PAGE_SIZE;
1524         int i;
1525
1526         BUG_ON(obj_priv->pages_refcount == 0);
1527         BUG_ON(obj_priv->madv == __I915_MADV_PURGED);
1528
1529         if (--obj_priv->pages_refcount != 0)
1530                 return;
1531
1532         if (obj_priv->tiling_mode != I915_TILING_NONE)
1533                 i915_gem_object_save_bit_17_swizzle(obj);
1534
1535         if (obj_priv->madv == I915_MADV_DONTNEED)
1536                 obj_priv->dirty = 0;
1537
1538         for (i = 0; i < page_count; i++) {
1539                 if (obj_priv->dirty)
1540                         set_page_dirty(obj_priv->pages[i]);
1541
1542                 if (obj_priv->madv == I915_MADV_WILLNEED)
1543                         mark_page_accessed(obj_priv->pages[i]);
1544
1545                 page_cache_release(obj_priv->pages[i]);
1546         }
1547         obj_priv->dirty = 0;
1548
1549         drm_free_large(obj_priv->pages);
1550         obj_priv->pages = NULL;
1551 }
1552
1553 static uint32_t
1554 i915_gem_next_request_seqno(struct drm_device *dev,
1555                             struct intel_ring_buffer *ring)
1556 {
1557         drm_i915_private_t *dev_priv = dev->dev_private;
1558
1559         ring->outstanding_lazy_request = true;
1560         return dev_priv->next_seqno;
1561 }
1562
1563 static void
1564 i915_gem_object_move_to_active(struct drm_gem_object *obj,
1565                                struct intel_ring_buffer *ring)
1566 {
1567         struct drm_device *dev = obj->dev;
1568         struct drm_i915_private *dev_priv = dev->dev_private;
1569         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1570         uint32_t seqno = i915_gem_next_request_seqno(dev, ring);
1571
1572         BUG_ON(ring == NULL);
1573         obj_priv->ring = ring;
1574
1575         /* Add a reference if we're newly entering the active list. */
1576         if (!obj_priv->active) {
1577                 drm_gem_object_reference(obj);
1578                 obj_priv->active = 1;
1579         }
1580
1581         /* Move from whatever list we were on to the tail of execution. */
1582         list_move_tail(&obj_priv->mm_list, &dev_priv->mm.active_list);
1583         list_move_tail(&obj_priv->ring_list, &ring->active_list);
1584         obj_priv->last_rendering_seqno = seqno;
1585 }
1586
1587 static void
1588 i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
1589 {
1590         struct drm_device *dev = obj->dev;
1591         drm_i915_private_t *dev_priv = dev->dev_private;
1592         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1593
1594         BUG_ON(!obj_priv->active);
1595         list_move_tail(&obj_priv->mm_list, &dev_priv->mm.flushing_list);
1596         list_del_init(&obj_priv->ring_list);
1597         obj_priv->last_rendering_seqno = 0;
1598 }
1599
1600 /* Immediately discard the backing storage */
1601 static void
1602 i915_gem_object_truncate(struct drm_gem_object *obj)
1603 {
1604         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1605         struct inode *inode;
1606
1607         /* Our goal here is to return as much of the memory as
1608          * is possible back to the system as we are called from OOM.
1609          * To do this we must instruct the shmfs to drop all of its
1610          * backing pages, *now*. Here we mirror the actions taken
1611          * when by shmem_delete_inode() to release the backing store.
1612          */
1613         inode = obj->filp->f_path.dentry->d_inode;
1614         truncate_inode_pages(inode->i_mapping, 0);
1615         if (inode->i_op->truncate_range)
1616                 inode->i_op->truncate_range(inode, 0, (loff_t)-1);
1617
1618         obj_priv->madv = __I915_MADV_PURGED;
1619 }
1620
1621 static inline int
1622 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv)
1623 {
1624         return obj_priv->madv == I915_MADV_DONTNEED;
1625 }
1626
1627 static void
1628 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
1629 {
1630         struct drm_device *dev = obj->dev;
1631         drm_i915_private_t *dev_priv = dev->dev_private;
1632         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1633
1634         if (obj_priv->pin_count != 0)
1635                 list_move_tail(&obj_priv->mm_list, &dev_priv->mm.pinned_list);
1636         else
1637                 list_move_tail(&obj_priv->mm_list, &dev_priv->mm.inactive_list);
1638         list_del_init(&obj_priv->ring_list);
1639
1640         BUG_ON(!list_empty(&obj_priv->gpu_write_list));
1641
1642         obj_priv->last_rendering_seqno = 0;
1643         obj_priv->ring = NULL;
1644         if (obj_priv->active) {
1645                 obj_priv->active = 0;
1646                 drm_gem_object_unreference(obj);
1647         }
1648         WARN_ON(i915_verify_lists(dev));
1649 }
1650
1651 static void
1652 i915_gem_process_flushing_list(struct drm_device *dev,
1653                                uint32_t flush_domains,
1654                                struct intel_ring_buffer *ring)
1655 {
1656         drm_i915_private_t *dev_priv = dev->dev_private;
1657         struct drm_i915_gem_object *obj_priv, *next;
1658
1659         list_for_each_entry_safe(obj_priv, next,
1660                                  &ring->gpu_write_list,
1661                                  gpu_write_list) {
1662                 struct drm_gem_object *obj = &obj_priv->base;
1663
1664                 if (obj->write_domain & flush_domains) {
1665                         uint32_t old_write_domain = obj->write_domain;
1666
1667                         obj->write_domain = 0;
1668                         list_del_init(&obj_priv->gpu_write_list);
1669                         i915_gem_object_move_to_active(obj, ring);
1670
1671                         /* update the fence lru list */
1672                         if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1673                                 struct drm_i915_fence_reg *reg =
1674                                         &dev_priv->fence_regs[obj_priv->fence_reg];
1675                                 list_move_tail(&reg->lru_list,
1676                                                 &dev_priv->mm.fence_list);
1677                         }
1678
1679                         trace_i915_gem_object_change_domain(obj,
1680                                                             obj->read_domains,
1681                                                             old_write_domain);
1682                 }
1683         }
1684 }
1685
1686 uint32_t
1687 i915_add_request(struct drm_device *dev,
1688                  struct drm_file *file,
1689                  struct drm_i915_gem_request *request,
1690                  struct intel_ring_buffer *ring)
1691 {
1692         drm_i915_private_t *dev_priv = dev->dev_private;
1693         struct drm_i915_file_private *file_priv = NULL;
1694         uint32_t seqno;
1695         int was_empty;
1696
1697         if (file != NULL)
1698                 file_priv = file->driver_priv;
1699
1700         if (request == NULL) {
1701                 request = kzalloc(sizeof(*request), GFP_KERNEL);
1702                 if (request == NULL)
1703                         return 0;
1704         }
1705
1706         seqno = ring->add_request(dev, ring, 0);
1707         ring->outstanding_lazy_request = false;
1708
1709         request->seqno = seqno;
1710         request->ring = ring;
1711         request->emitted_jiffies = jiffies;
1712         was_empty = list_empty(&ring->request_list);
1713         list_add_tail(&request->list, &ring->request_list);
1714
1715         if (file_priv) {
1716                 spin_lock(&file_priv->mm.lock);
1717                 request->file_priv = file_priv;
1718                 list_add_tail(&request->client_list,
1719                               &file_priv->mm.request_list);
1720                 spin_unlock(&file_priv->mm.lock);
1721         }
1722
1723         if (!dev_priv->mm.suspended) {
1724                 mod_timer(&dev_priv->hangcheck_timer,
1725                           jiffies + msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1726                 if (was_empty)
1727                         queue_delayed_work(dev_priv->wq,
1728                                            &dev_priv->mm.retire_work, HZ);
1729         }
1730         return seqno;
1731 }
1732
1733 /**
1734  * Command execution barrier
1735  *
1736  * Ensures that all commands in the ring are finished
1737  * before signalling the CPU
1738  */
1739 static void
1740 i915_retire_commands(struct drm_device *dev, struct intel_ring_buffer *ring)
1741 {
1742         uint32_t flush_domains = 0;
1743
1744         /* The sampler always gets flushed on i965 (sigh) */
1745         if (INTEL_INFO(dev)->gen >= 4)
1746                 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
1747
1748         ring->flush(dev, ring,
1749                         I915_GEM_DOMAIN_COMMAND, flush_domains);
1750 }
1751
1752 static inline void
1753 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1754 {
1755         struct drm_i915_file_private *file_priv = request->file_priv;
1756
1757         if (!file_priv)
1758                 return;
1759
1760         spin_lock(&file_priv->mm.lock);
1761         list_del(&request->client_list);
1762         request->file_priv = NULL;
1763         spin_unlock(&file_priv->mm.lock);
1764 }
1765
1766 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1767                                       struct intel_ring_buffer *ring)
1768 {
1769         while (!list_empty(&ring->request_list)) {
1770                 struct drm_i915_gem_request *request;
1771
1772                 request = list_first_entry(&ring->request_list,
1773                                            struct drm_i915_gem_request,
1774                                            list);
1775
1776                 list_del(&request->list);
1777                 i915_gem_request_remove_from_client(request);
1778                 kfree(request);
1779         }
1780
1781         while (!list_empty(&ring->active_list)) {
1782                 struct drm_i915_gem_object *obj_priv;
1783
1784                 obj_priv = list_first_entry(&ring->active_list,
1785                                             struct drm_i915_gem_object,
1786                                             ring_list);
1787
1788                 obj_priv->base.write_domain = 0;
1789                 list_del_init(&obj_priv->gpu_write_list);
1790                 i915_gem_object_move_to_inactive(&obj_priv->base);
1791         }
1792 }
1793
1794 void i915_gem_reset(struct drm_device *dev)
1795 {
1796         struct drm_i915_private *dev_priv = dev->dev_private;
1797         struct drm_i915_gem_object *obj_priv;
1798         int i;
1799
1800         i915_gem_reset_ring_lists(dev_priv, &dev_priv->render_ring);
1801         i915_gem_reset_ring_lists(dev_priv, &dev_priv->bsd_ring);
1802         i915_gem_reset_ring_lists(dev_priv, &dev_priv->blt_ring);
1803
1804         /* Remove anything from the flushing lists. The GPU cache is likely
1805          * to be lost on reset along with the data, so simply move the
1806          * lost bo to the inactive list.
1807          */
1808         while (!list_empty(&dev_priv->mm.flushing_list)) {
1809                 obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
1810                                             struct drm_i915_gem_object,
1811                                             mm_list);
1812
1813                 obj_priv->base.write_domain = 0;
1814                 list_del_init(&obj_priv->gpu_write_list);
1815                 i915_gem_object_move_to_inactive(&obj_priv->base);
1816         }
1817
1818         /* Move everything out of the GPU domains to ensure we do any
1819          * necessary invalidation upon reuse.
1820          */
1821         list_for_each_entry(obj_priv,
1822                             &dev_priv->mm.inactive_list,
1823                             mm_list)
1824         {
1825                 obj_priv->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1826         }
1827
1828         /* The fence registers are invalidated so clear them out */
1829         for (i = 0; i < 16; i++) {
1830                 struct drm_i915_fence_reg *reg;
1831
1832                 reg = &dev_priv->fence_regs[i];
1833                 if (!reg->obj)
1834                         continue;
1835
1836                 i915_gem_clear_fence_reg(reg->obj);
1837         }
1838 }
1839
1840 /**
1841  * This function clears the request list as sequence numbers are passed.
1842  */
1843 static void
1844 i915_gem_retire_requests_ring(struct drm_device *dev,
1845                               struct intel_ring_buffer *ring)
1846 {
1847         drm_i915_private_t *dev_priv = dev->dev_private;
1848         uint32_t seqno;
1849
1850         if (!ring->status_page.page_addr ||
1851             list_empty(&ring->request_list))
1852                 return;
1853
1854         WARN_ON(i915_verify_lists(dev));
1855
1856         seqno = ring->get_seqno(dev, ring);
1857         while (!list_empty(&ring->request_list)) {
1858                 struct drm_i915_gem_request *request;
1859
1860                 request = list_first_entry(&ring->request_list,
1861                                            struct drm_i915_gem_request,
1862                                            list);
1863
1864                 if (!i915_seqno_passed(seqno, request->seqno))
1865                         break;
1866
1867                 trace_i915_gem_request_retire(dev, request->seqno);
1868
1869                 list_del(&request->list);
1870                 i915_gem_request_remove_from_client(request);
1871                 kfree(request);
1872         }
1873
1874         /* Move any buffers on the active list that are no longer referenced
1875          * by the ringbuffer to the flushing/inactive lists as appropriate.
1876          */
1877         while (!list_empty(&ring->active_list)) {
1878                 struct drm_gem_object *obj;
1879                 struct drm_i915_gem_object *obj_priv;
1880
1881                 obj_priv = list_first_entry(&ring->active_list,
1882                                             struct drm_i915_gem_object,
1883                                             ring_list);
1884
1885                 if (!i915_seqno_passed(seqno, obj_priv->last_rendering_seqno))
1886                         break;
1887
1888                 obj = &obj_priv->base;
1889                 if (obj->write_domain != 0)
1890                         i915_gem_object_move_to_flushing(obj);
1891                 else
1892                         i915_gem_object_move_to_inactive(obj);
1893         }
1894
1895         if (unlikely (dev_priv->trace_irq_seqno &&
1896                       i915_seqno_passed(dev_priv->trace_irq_seqno, seqno))) {
1897                 ring->user_irq_put(dev, ring);
1898                 dev_priv->trace_irq_seqno = 0;
1899         }
1900
1901         WARN_ON(i915_verify_lists(dev));
1902 }
1903
1904 void
1905 i915_gem_retire_requests(struct drm_device *dev)
1906 {
1907         drm_i915_private_t *dev_priv = dev->dev_private;
1908
1909         if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1910             struct drm_i915_gem_object *obj_priv, *tmp;
1911
1912             /* We must be careful that during unbind() we do not
1913              * accidentally infinitely recurse into retire requests.
1914              * Currently:
1915              *   retire -> free -> unbind -> wait -> retire_ring
1916              */
1917             list_for_each_entry_safe(obj_priv, tmp,
1918                                      &dev_priv->mm.deferred_free_list,
1919                                      mm_list)
1920                     i915_gem_free_object_tail(&obj_priv->base);
1921         }
1922
1923         i915_gem_retire_requests_ring(dev, &dev_priv->render_ring);
1924         i915_gem_retire_requests_ring(dev, &dev_priv->bsd_ring);
1925         i915_gem_retire_requests_ring(dev, &dev_priv->blt_ring);
1926 }
1927
1928 static void
1929 i915_gem_retire_work_handler(struct work_struct *work)
1930 {
1931         drm_i915_private_t *dev_priv;
1932         struct drm_device *dev;
1933
1934         dev_priv = container_of(work, drm_i915_private_t,
1935                                 mm.retire_work.work);
1936         dev = dev_priv->dev;
1937
1938         /* Come back later if the device is busy... */
1939         if (!mutex_trylock(&dev->struct_mutex)) {
1940                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1941                 return;
1942         }
1943
1944         i915_gem_retire_requests(dev);
1945
1946         if (!dev_priv->mm.suspended &&
1947                 (!list_empty(&dev_priv->render_ring.request_list) ||
1948                  !list_empty(&dev_priv->bsd_ring.request_list) ||
1949                  !list_empty(&dev_priv->blt_ring.request_list)))
1950                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1951         mutex_unlock(&dev->struct_mutex);
1952 }
1953
1954 int
1955 i915_do_wait_request(struct drm_device *dev, uint32_t seqno,
1956                      bool interruptible, struct intel_ring_buffer *ring)
1957 {
1958         drm_i915_private_t *dev_priv = dev->dev_private;
1959         u32 ier;
1960         int ret = 0;
1961
1962         BUG_ON(seqno == 0);
1963
1964         if (atomic_read(&dev_priv->mm.wedged))
1965                 return -EAGAIN;
1966
1967         if (ring->outstanding_lazy_request) {
1968                 seqno = i915_add_request(dev, NULL, NULL, ring);
1969                 if (seqno == 0)
1970                         return -ENOMEM;
1971         }
1972         BUG_ON(seqno == dev_priv->next_seqno);
1973
1974         if (!i915_seqno_passed(ring->get_seqno(dev, ring), seqno)) {
1975                 if (HAS_PCH_SPLIT(dev))
1976                         ier = I915_READ(DEIER) | I915_READ(GTIER);
1977                 else
1978                         ier = I915_READ(IER);
1979                 if (!ier) {
1980                         DRM_ERROR("something (likely vbetool) disabled "
1981                                   "interrupts, re-enabling\n");
1982                         i915_driver_irq_preinstall(dev);
1983                         i915_driver_irq_postinstall(dev);
1984                 }
1985
1986                 trace_i915_gem_request_wait_begin(dev, seqno);
1987
1988                 ring->waiting_gem_seqno = seqno;
1989                 ring->user_irq_get(dev, ring);
1990                 if (interruptible)
1991                         ret = wait_event_interruptible(ring->irq_queue,
1992                                 i915_seqno_passed(
1993                                         ring->get_seqno(dev, ring), seqno)
1994                                 || atomic_read(&dev_priv->mm.wedged));
1995                 else
1996                         wait_event(ring->irq_queue,
1997                                 i915_seqno_passed(
1998                                         ring->get_seqno(dev, ring), seqno)
1999                                 || atomic_read(&dev_priv->mm.wedged));
2000
2001                 ring->user_irq_put(dev, ring);
2002                 ring->waiting_gem_seqno = 0;
2003
2004                 trace_i915_gem_request_wait_end(dev, seqno);
2005         }
2006         if (atomic_read(&dev_priv->mm.wedged))
2007                 ret = -EAGAIN;
2008
2009         if (ret && ret != -ERESTARTSYS)
2010                 DRM_ERROR("%s returns %d (awaiting %d at %d, next %d)\n",
2011                           __func__, ret, seqno, ring->get_seqno(dev, ring),
2012                           dev_priv->next_seqno);
2013
2014         /* Directly dispatch request retiring.  While we have the work queue
2015          * to handle this, the waiter on a request often wants an associated
2016          * buffer to have made it to the inactive list, and we would need
2017          * a separate wait queue to handle that.
2018          */
2019         if (ret == 0)
2020                 i915_gem_retire_requests_ring(dev, ring);
2021
2022         return ret;
2023 }
2024
2025 /**
2026  * Waits for a sequence number to be signaled, and cleans up the
2027  * request and object lists appropriately for that event.
2028  */
2029 static int
2030 i915_wait_request(struct drm_device *dev, uint32_t seqno,
2031                   struct intel_ring_buffer *ring)
2032 {
2033         return i915_do_wait_request(dev, seqno, 1, ring);
2034 }
2035
2036 static void
2037 i915_gem_flush_ring(struct drm_device *dev,
2038                     struct drm_file *file_priv,
2039                     struct intel_ring_buffer *ring,
2040                     uint32_t invalidate_domains,
2041                     uint32_t flush_domains)
2042 {
2043         ring->flush(dev, ring, invalidate_domains, flush_domains);
2044         i915_gem_process_flushing_list(dev, flush_domains, ring);
2045 }
2046
2047 static void
2048 i915_gem_flush(struct drm_device *dev,
2049                struct drm_file *file_priv,
2050                uint32_t invalidate_domains,
2051                uint32_t flush_domains,
2052                uint32_t flush_rings)
2053 {
2054         drm_i915_private_t *dev_priv = dev->dev_private;
2055
2056         if (flush_domains & I915_GEM_DOMAIN_CPU)
2057                 drm_agp_chipset_flush(dev);
2058
2059         if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
2060                 if (flush_rings & RING_RENDER)
2061                         i915_gem_flush_ring(dev, file_priv,
2062                                             &dev_priv->render_ring,
2063                                             invalidate_domains, flush_domains);
2064                 if (flush_rings & RING_BSD)
2065                         i915_gem_flush_ring(dev, file_priv,
2066                                             &dev_priv->bsd_ring,
2067                                             invalidate_domains, flush_domains);
2068                 if (flush_rings & RING_BLT)
2069                         i915_gem_flush_ring(dev, file_priv,
2070                                             &dev_priv->blt_ring,
2071                                             invalidate_domains, flush_domains);
2072         }
2073 }
2074
2075 /**
2076  * Ensures that all rendering to the object has completed and the object is
2077  * safe to unbind from the GTT or access from the CPU.
2078  */
2079 static int
2080 i915_gem_object_wait_rendering(struct drm_gem_object *obj,
2081                                bool interruptible)
2082 {
2083         struct drm_device *dev = obj->dev;
2084         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2085         int ret;
2086
2087         /* This function only exists to support waiting for existing rendering,
2088          * not for emitting required flushes.
2089          */
2090         BUG_ON((obj->write_domain & I915_GEM_GPU_DOMAINS) != 0);
2091
2092         /* If there is rendering queued on the buffer being evicted, wait for
2093          * it.
2094          */
2095         if (obj_priv->active) {
2096                 ret = i915_do_wait_request(dev,
2097                                            obj_priv->last_rendering_seqno,
2098                                            interruptible,
2099                                            obj_priv->ring);
2100                 if (ret)
2101                         return ret;
2102         }
2103
2104         return 0;
2105 }
2106
2107 /**
2108  * Unbinds an object from the GTT aperture.
2109  */
2110 int
2111 i915_gem_object_unbind(struct drm_gem_object *obj)
2112 {
2113         struct drm_device *dev = obj->dev;
2114         struct drm_i915_private *dev_priv = dev->dev_private;
2115         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2116         int ret = 0;
2117
2118         if (obj_priv->gtt_space == NULL)
2119                 return 0;
2120
2121         if (obj_priv->pin_count != 0) {
2122                 DRM_ERROR("Attempting to unbind pinned buffer\n");
2123                 return -EINVAL;
2124         }
2125
2126         /* blow away mappings if mapped through GTT */
2127         i915_gem_release_mmap(obj);
2128
2129         /* Move the object to the CPU domain to ensure that
2130          * any possible CPU writes while it's not in the GTT
2131          * are flushed when we go to remap it. This will
2132          * also ensure that all pending GPU writes are finished
2133          * before we unbind.
2134          */
2135         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2136         if (ret == -ERESTARTSYS)
2137                 return ret;
2138         /* Continue on if we fail due to EIO, the GPU is hung so we
2139          * should be safe and we need to cleanup or else we might
2140          * cause memory corruption through use-after-free.
2141          */
2142         if (ret) {
2143                 i915_gem_clflush_object(obj);
2144                 obj->read_domains = obj->write_domain = I915_GEM_DOMAIN_CPU;
2145         }
2146
2147         /* release the fence reg _after_ flushing */
2148         if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
2149                 i915_gem_clear_fence_reg(obj);
2150
2151         drm_unbind_agp(obj_priv->agp_mem);
2152         drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
2153
2154         i915_gem_object_put_pages(obj);
2155         BUG_ON(obj_priv->pages_refcount);
2156
2157         i915_gem_info_remove_gtt(dev_priv, obj->size);
2158         list_del_init(&obj_priv->mm_list);
2159
2160         drm_mm_put_block(obj_priv->gtt_space);
2161         obj_priv->gtt_space = NULL;
2162         obj_priv->gtt_offset = 0;
2163
2164         if (i915_gem_object_is_purgeable(obj_priv))
2165                 i915_gem_object_truncate(obj);
2166
2167         trace_i915_gem_object_unbind(obj);
2168
2169         return ret;
2170 }
2171
2172 static int i915_ring_idle(struct drm_device *dev,
2173                           struct intel_ring_buffer *ring)
2174 {
2175         if (list_empty(&ring->gpu_write_list))
2176                 return 0;
2177
2178         i915_gem_flush_ring(dev, NULL, ring,
2179                             I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2180         return i915_wait_request(dev,
2181                                  i915_gem_next_request_seqno(dev, ring),
2182                                  ring);
2183 }
2184
2185 int
2186 i915_gpu_idle(struct drm_device *dev)
2187 {
2188         drm_i915_private_t *dev_priv = dev->dev_private;
2189         bool lists_empty;
2190         int ret;
2191
2192         lists_empty = (list_empty(&dev_priv->mm.flushing_list) &&
2193                        list_empty(&dev_priv->render_ring.active_list) &&
2194                        list_empty(&dev_priv->bsd_ring.active_list) &&
2195                        list_empty(&dev_priv->blt_ring.active_list));
2196         if (lists_empty)
2197                 return 0;
2198
2199         /* Flush everything onto the inactive list. */
2200         ret = i915_ring_idle(dev, &dev_priv->render_ring);
2201         if (ret)
2202                 return ret;
2203
2204         ret = i915_ring_idle(dev, &dev_priv->bsd_ring);
2205         if (ret)
2206                 return ret;
2207
2208         ret = i915_ring_idle(dev, &dev_priv->blt_ring);
2209         if (ret)
2210                 return ret;
2211
2212         return 0;
2213 }
2214
2215 static int
2216 i915_gem_object_get_pages(struct drm_gem_object *obj,
2217                           gfp_t gfpmask)
2218 {
2219         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2220         int page_count, i;
2221         struct address_space *mapping;
2222         struct inode *inode;
2223         struct page *page;
2224
2225         BUG_ON(obj_priv->pages_refcount
2226                         == DRM_I915_GEM_OBJECT_MAX_PAGES_REFCOUNT);
2227
2228         if (obj_priv->pages_refcount++ != 0)
2229                 return 0;
2230
2231         /* Get the list of pages out of our struct file.  They'll be pinned
2232          * at this point until we release them.
2233          */
2234         page_count = obj->size / PAGE_SIZE;
2235         BUG_ON(obj_priv->pages != NULL);
2236         obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *));
2237         if (obj_priv->pages == NULL) {
2238                 obj_priv->pages_refcount--;
2239                 return -ENOMEM;
2240         }
2241
2242         inode = obj->filp->f_path.dentry->d_inode;
2243         mapping = inode->i_mapping;
2244         for (i = 0; i < page_count; i++) {
2245                 page = read_cache_page_gfp(mapping, i,
2246                                            GFP_HIGHUSER |
2247                                            __GFP_COLD |
2248                                            __GFP_RECLAIMABLE |
2249                                            gfpmask);
2250                 if (IS_ERR(page))
2251                         goto err_pages;
2252
2253                 obj_priv->pages[i] = page;
2254         }
2255
2256         if (obj_priv->tiling_mode != I915_TILING_NONE)
2257                 i915_gem_object_do_bit_17_swizzle(obj);
2258
2259         return 0;
2260
2261 err_pages:
2262         while (i--)
2263                 page_cache_release(obj_priv->pages[i]);
2264
2265         drm_free_large(obj_priv->pages);
2266         obj_priv->pages = NULL;
2267         obj_priv->pages_refcount--;
2268         return PTR_ERR(page);
2269 }
2270
2271 static void sandybridge_write_fence_reg(struct drm_i915_fence_reg *reg)
2272 {
2273         struct drm_gem_object *obj = reg->obj;
2274         struct drm_device *dev = obj->dev;
2275         drm_i915_private_t *dev_priv = dev->dev_private;
2276         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2277         int regnum = obj_priv->fence_reg;
2278         uint64_t val;
2279
2280         val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2281                     0xfffff000) << 32;
2282         val |= obj_priv->gtt_offset & 0xfffff000;
2283         val |= (uint64_t)((obj_priv->stride / 128) - 1) <<
2284                 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2285
2286         if (obj_priv->tiling_mode == I915_TILING_Y)
2287                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2288         val |= I965_FENCE_REG_VALID;
2289
2290         I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + (regnum * 8), val);
2291 }
2292
2293 static void i965_write_fence_reg(struct drm_i915_fence_reg *reg)
2294 {
2295         struct drm_gem_object *obj = reg->obj;
2296         struct drm_device *dev = obj->dev;
2297         drm_i915_private_t *dev_priv = dev->dev_private;
2298         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2299         int regnum = obj_priv->fence_reg;
2300         uint64_t val;
2301
2302         val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2303                     0xfffff000) << 32;
2304         val |= obj_priv->gtt_offset & 0xfffff000;
2305         val |= ((obj_priv->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2306         if (obj_priv->tiling_mode == I915_TILING_Y)
2307                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2308         val |= I965_FENCE_REG_VALID;
2309
2310         I915_WRITE64(FENCE_REG_965_0 + (regnum * 8), val);
2311 }
2312
2313 static void i915_write_fence_reg(struct drm_i915_fence_reg *reg)
2314 {
2315         struct drm_gem_object *obj = reg->obj;
2316         struct drm_device *dev = obj->dev;
2317         drm_i915_private_t *dev_priv = dev->dev_private;
2318         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2319         int regnum = obj_priv->fence_reg;
2320         int tile_width;
2321         uint32_t fence_reg, val;
2322         uint32_t pitch_val;
2323
2324         if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) ||
2325             (obj_priv->gtt_offset & (obj->size - 1))) {
2326                 WARN(1, "%s: object 0x%08x not 1M or size (0x%zx) aligned\n",
2327                      __func__, obj_priv->gtt_offset, obj->size);
2328                 return;
2329         }
2330
2331         if (obj_priv->tiling_mode == I915_TILING_Y &&
2332             HAS_128_BYTE_Y_TILING(dev))
2333                 tile_width = 128;
2334         else
2335                 tile_width = 512;
2336
2337         /* Note: pitch better be a power of two tile widths */
2338         pitch_val = obj_priv->stride / tile_width;
2339         pitch_val = ffs(pitch_val) - 1;
2340
2341         if (obj_priv->tiling_mode == I915_TILING_Y &&
2342             HAS_128_BYTE_Y_TILING(dev))
2343                 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2344         else
2345                 WARN_ON(pitch_val > I915_FENCE_MAX_PITCH_VAL);
2346
2347         val = obj_priv->gtt_offset;
2348         if (obj_priv->tiling_mode == I915_TILING_Y)
2349                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2350         val |= I915_FENCE_SIZE_BITS(obj->size);
2351         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2352         val |= I830_FENCE_REG_VALID;
2353
2354         if (regnum < 8)
2355                 fence_reg = FENCE_REG_830_0 + (regnum * 4);
2356         else
2357                 fence_reg = FENCE_REG_945_8 + ((regnum - 8) * 4);
2358         I915_WRITE(fence_reg, val);
2359 }
2360
2361 static void i830_write_fence_reg(struct drm_i915_fence_reg *reg)
2362 {
2363         struct drm_gem_object *obj = reg->obj;
2364         struct drm_device *dev = obj->dev;
2365         drm_i915_private_t *dev_priv = dev->dev_private;
2366         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2367         int regnum = obj_priv->fence_reg;
2368         uint32_t val;
2369         uint32_t pitch_val;
2370         uint32_t fence_size_bits;
2371
2372         if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) ||
2373             (obj_priv->gtt_offset & (obj->size - 1))) {
2374                 WARN(1, "%s: object 0x%08x not 512K or size aligned\n",
2375                      __func__, obj_priv->gtt_offset);
2376                 return;
2377         }
2378
2379         pitch_val = obj_priv->stride / 128;
2380         pitch_val = ffs(pitch_val) - 1;
2381         WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2382
2383         val = obj_priv->gtt_offset;
2384         if (obj_priv->tiling_mode == I915_TILING_Y)
2385                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2386         fence_size_bits = I830_FENCE_SIZE_BITS(obj->size);
2387         WARN_ON(fence_size_bits & ~0x00000f00);
2388         val |= fence_size_bits;
2389         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2390         val |= I830_FENCE_REG_VALID;
2391
2392         I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val);
2393 }
2394
2395 static int i915_find_fence_reg(struct drm_device *dev,
2396                                bool interruptible)
2397 {
2398         struct drm_i915_fence_reg *reg = NULL;
2399         struct drm_i915_gem_object *obj_priv = NULL;
2400         struct drm_i915_private *dev_priv = dev->dev_private;
2401         struct drm_gem_object *obj = NULL;
2402         int i, avail, ret;
2403
2404         /* First try to find a free reg */
2405         avail = 0;
2406         for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2407                 reg = &dev_priv->fence_regs[i];
2408                 if (!reg->obj)
2409                         return i;
2410
2411                 obj_priv = to_intel_bo(reg->obj);
2412                 if (!obj_priv->pin_count)
2413                     avail++;
2414         }
2415
2416         if (avail == 0)
2417                 return -ENOSPC;
2418
2419         /* None available, try to steal one or wait for a user to finish */
2420         i = I915_FENCE_REG_NONE;
2421         list_for_each_entry(reg, &dev_priv->mm.fence_list,
2422                             lru_list) {
2423                 obj = reg->obj;
2424                 obj_priv = to_intel_bo(obj);
2425
2426                 if (obj_priv->pin_count)
2427                         continue;
2428
2429                 /* found one! */
2430                 i = obj_priv->fence_reg;
2431                 break;
2432         }
2433
2434         BUG_ON(i == I915_FENCE_REG_NONE);
2435
2436         /* We only have a reference on obj from the active list. put_fence_reg
2437          * might drop that one, causing a use-after-free in it. So hold a
2438          * private reference to obj like the other callers of put_fence_reg
2439          * (set_tiling ioctl) do. */
2440         drm_gem_object_reference(obj);
2441         ret = i915_gem_object_put_fence_reg(obj, interruptible);
2442         drm_gem_object_unreference(obj);
2443         if (ret != 0)
2444                 return ret;
2445
2446         return i;
2447 }
2448
2449 /**
2450  * i915_gem_object_get_fence_reg - set up a fence reg for an object
2451  * @obj: object to map through a fence reg
2452  *
2453  * When mapping objects through the GTT, userspace wants to be able to write
2454  * to them without having to worry about swizzling if the object is tiled.
2455  *
2456  * This function walks the fence regs looking for a free one for @obj,
2457  * stealing one if it can't find any.
2458  *
2459  * It then sets up the reg based on the object's properties: address, pitch
2460  * and tiling format.
2461  */
2462 int
2463 i915_gem_object_get_fence_reg(struct drm_gem_object *obj,
2464                               bool interruptible)
2465 {
2466         struct drm_device *dev = obj->dev;
2467         struct drm_i915_private *dev_priv = dev->dev_private;
2468         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2469         struct drm_i915_fence_reg *reg = NULL;
2470         int ret;
2471
2472         /* Just update our place in the LRU if our fence is getting used. */
2473         if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
2474                 reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2475                 list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2476                 return 0;
2477         }
2478
2479         switch (obj_priv->tiling_mode) {
2480         case I915_TILING_NONE:
2481                 WARN(1, "allocating a fence for non-tiled object?\n");
2482                 break;
2483         case I915_TILING_X:
2484                 if (!obj_priv->stride)
2485                         return -EINVAL;
2486                 WARN((obj_priv->stride & (512 - 1)),
2487                      "object 0x%08x is X tiled but has non-512B pitch\n",
2488                      obj_priv->gtt_offset);
2489                 break;
2490         case I915_TILING_Y:
2491                 if (!obj_priv->stride)
2492                         return -EINVAL;
2493                 WARN((obj_priv->stride & (128 - 1)),
2494                      "object 0x%08x is Y tiled but has non-128B pitch\n",
2495                      obj_priv->gtt_offset);
2496                 break;
2497         }
2498
2499         ret = i915_find_fence_reg(dev, interruptible);
2500         if (ret < 0)
2501                 return ret;
2502
2503         obj_priv->fence_reg = ret;
2504         reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2505         list_add_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2506
2507         reg->obj = obj;
2508
2509         switch (INTEL_INFO(dev)->gen) {
2510         case 6:
2511                 sandybridge_write_fence_reg(reg);
2512                 break;
2513         case 5:
2514         case 4:
2515                 i965_write_fence_reg(reg);
2516                 break;
2517         case 3:
2518                 i915_write_fence_reg(reg);
2519                 break;
2520         case 2:
2521                 i830_write_fence_reg(reg);
2522                 break;
2523         }
2524
2525         trace_i915_gem_object_get_fence(obj, obj_priv->fence_reg,
2526                         obj_priv->tiling_mode);
2527
2528         return 0;
2529 }
2530
2531 /**
2532  * i915_gem_clear_fence_reg - clear out fence register info
2533  * @obj: object to clear
2534  *
2535  * Zeroes out the fence register itself and clears out the associated
2536  * data structures in dev_priv and obj_priv.
2537  */
2538 static void
2539 i915_gem_clear_fence_reg(struct drm_gem_object *obj)
2540 {
2541         struct drm_device *dev = obj->dev;
2542         drm_i915_private_t *dev_priv = dev->dev_private;
2543         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2544         struct drm_i915_fence_reg *reg =
2545                 &dev_priv->fence_regs[obj_priv->fence_reg];
2546         uint32_t fence_reg;
2547
2548         switch (INTEL_INFO(dev)->gen) {
2549         case 6:
2550                 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 +
2551                              (obj_priv->fence_reg * 8), 0);
2552                 break;
2553         case 5:
2554         case 4:
2555                 I915_WRITE64(FENCE_REG_965_0 + (obj_priv->fence_reg * 8), 0);
2556                 break;
2557         case 3:
2558                 if (obj_priv->fence_reg >= 8)
2559                         fence_reg = FENCE_REG_945_8 + (obj_priv->fence_reg - 8) * 4;
2560                 else
2561         case 2:
2562                         fence_reg = FENCE_REG_830_0 + obj_priv->fence_reg * 4;
2563
2564                 I915_WRITE(fence_reg, 0);
2565                 break;
2566         }
2567
2568         reg->obj = NULL;
2569         obj_priv->fence_reg = I915_FENCE_REG_NONE;
2570         list_del_init(&reg->lru_list);
2571 }
2572
2573 /**
2574  * i915_gem_object_put_fence_reg - waits on outstanding fenced access
2575  * to the buffer to finish, and then resets the fence register.
2576  * @obj: tiled object holding a fence register.
2577  * @bool: whether the wait upon the fence is interruptible
2578  *
2579  * Zeroes out the fence register itself and clears out the associated
2580  * data structures in dev_priv and obj_priv.
2581  */
2582 int
2583 i915_gem_object_put_fence_reg(struct drm_gem_object *obj,
2584                               bool interruptible)
2585 {
2586         struct drm_device *dev = obj->dev;
2587         struct drm_i915_private *dev_priv = dev->dev_private;
2588         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2589         struct drm_i915_fence_reg *reg;
2590
2591         if (obj_priv->fence_reg == I915_FENCE_REG_NONE)
2592                 return 0;
2593
2594         /* If we've changed tiling, GTT-mappings of the object
2595          * need to re-fault to ensure that the correct fence register
2596          * setup is in place.
2597          */
2598         i915_gem_release_mmap(obj);
2599
2600         /* On the i915, GPU access to tiled buffers is via a fence,
2601          * therefore we must wait for any outstanding access to complete
2602          * before clearing the fence.
2603          */
2604         reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2605         if (reg->gpu) {
2606                 int ret;
2607
2608                 ret = i915_gem_object_flush_gpu_write_domain(obj, true);
2609                 if (ret)
2610                         return ret;
2611
2612                 ret = i915_gem_object_wait_rendering(obj, interruptible);
2613                 if (ret)
2614                         return ret;
2615
2616                 reg->gpu = false;
2617         }
2618
2619         i915_gem_object_flush_gtt_write_domain(obj);
2620         i915_gem_clear_fence_reg(obj);
2621
2622         return 0;
2623 }
2624
2625 /**
2626  * Finds free space in the GTT aperture and binds the object there.
2627  */
2628 static int
2629 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
2630 {
2631         struct drm_device *dev = obj->dev;
2632         drm_i915_private_t *dev_priv = dev->dev_private;
2633         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2634         struct drm_mm_node *free_space;
2635         gfp_t gfpmask =  __GFP_NORETRY | __GFP_NOWARN;
2636         int ret;
2637
2638         if (obj_priv->madv != I915_MADV_WILLNEED) {
2639                 DRM_ERROR("Attempting to bind a purgeable object\n");
2640                 return -EINVAL;
2641         }
2642
2643         if (alignment == 0)
2644                 alignment = i915_gem_get_gtt_alignment(obj);
2645         if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) {
2646                 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2647                 return -EINVAL;
2648         }
2649
2650         /* If the object is bigger than the entire aperture, reject it early
2651          * before evicting everything in a vain attempt to find space.
2652          */
2653         if (obj->size > dev_priv->mm.gtt_total) {
2654                 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2655                 return -E2BIG;
2656         }
2657
2658  search_free:
2659         free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2660                                         obj->size, alignment, 0);
2661         if (free_space != NULL)
2662                 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
2663                                                        alignment);
2664         if (obj_priv->gtt_space == NULL) {
2665                 /* If the gtt is empty and we're still having trouble
2666                  * fitting our object in, we're out of memory.
2667                  */
2668                 ret = i915_gem_evict_something(dev, obj->size, alignment);
2669                 if (ret)
2670                         return ret;
2671
2672                 goto search_free;
2673         }
2674
2675         ret = i915_gem_object_get_pages(obj, gfpmask);
2676         if (ret) {
2677                 drm_mm_put_block(obj_priv->gtt_space);
2678                 obj_priv->gtt_space = NULL;
2679
2680                 if (ret == -ENOMEM) {
2681                         /* first try to clear up some space from the GTT */
2682                         ret = i915_gem_evict_something(dev, obj->size,
2683                                                        alignment);
2684                         if (ret) {
2685                                 /* now try to shrink everyone else */
2686                                 if (gfpmask) {
2687                                         gfpmask = 0;
2688                                         goto search_free;
2689                                 }
2690
2691                                 return ret;
2692                         }
2693
2694                         goto search_free;
2695                 }
2696
2697                 return ret;
2698         }
2699
2700         /* Create an AGP memory structure pointing at our pages, and bind it
2701          * into the GTT.
2702          */
2703         obj_priv->agp_mem = drm_agp_bind_pages(dev,
2704                                                obj_priv->pages,
2705                                                obj->size >> PAGE_SHIFT,
2706                                                obj_priv->gtt_space->start,
2707                                                obj_priv->agp_type);
2708         if (obj_priv->agp_mem == NULL) {
2709                 i915_gem_object_put_pages(obj);
2710                 drm_mm_put_block(obj_priv->gtt_space);
2711                 obj_priv->gtt_space = NULL;
2712
2713                 ret = i915_gem_evict_something(dev, obj->size, alignment);
2714                 if (ret)
2715                         return ret;
2716
2717                 goto search_free;
2718         }
2719
2720         /* keep track of bounds object by adding it to the inactive list */
2721         list_add_tail(&obj_priv->mm_list, &dev_priv->mm.inactive_list);
2722         i915_gem_info_add_gtt(dev_priv, obj->size);
2723
2724         /* Assert that the object is not currently in any GPU domain. As it
2725          * wasn't in the GTT, there shouldn't be any way it could have been in
2726          * a GPU cache
2727          */
2728         BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2729         BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2730
2731         obj_priv->gtt_offset = obj_priv->gtt_space->start;
2732         trace_i915_gem_object_bind(obj, obj_priv->gtt_offset);
2733
2734         return 0;
2735 }
2736
2737 void
2738 i915_gem_clflush_object(struct drm_gem_object *obj)
2739 {
2740         struct drm_i915_gem_object      *obj_priv = to_intel_bo(obj);
2741
2742         /* If we don't have a page list set up, then we're not pinned
2743          * to GPU, and we can ignore the cache flush because it'll happen
2744          * again at bind time.
2745          */
2746         if (obj_priv->pages == NULL)
2747                 return;
2748
2749         trace_i915_gem_object_clflush(obj);
2750
2751         drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE);
2752 }
2753
2754 /** Flushes any GPU write domain for the object if it's dirty. */
2755 static int
2756 i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj,
2757                                        bool pipelined)
2758 {
2759         struct drm_device *dev = obj->dev;
2760         uint32_t old_write_domain;
2761
2762         if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
2763                 return 0;
2764
2765         /* Queue the GPU write cache flushing we need. */
2766         old_write_domain = obj->write_domain;
2767         i915_gem_flush_ring(dev, NULL,
2768                             to_intel_bo(obj)->ring,
2769                             0, obj->write_domain);
2770         BUG_ON(obj->write_domain);
2771
2772         trace_i915_gem_object_change_domain(obj,
2773                                             obj->read_domains,
2774                                             old_write_domain);
2775
2776         if (pipelined)
2777                 return 0;
2778
2779         return i915_gem_object_wait_rendering(obj, true);
2780 }
2781
2782 /** Flushes the GTT write domain for the object if it's dirty. */
2783 static void
2784 i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj)
2785 {
2786         uint32_t old_write_domain;
2787
2788         if (obj->write_domain != I915_GEM_DOMAIN_GTT)
2789                 return;
2790
2791         /* No actual flushing is required for the GTT write domain.   Writes
2792          * to it immediately go to main memory as far as we know, so there's
2793          * no chipset flush.  It also doesn't land in render cache.
2794          */
2795         old_write_domain = obj->write_domain;
2796         obj->write_domain = 0;
2797
2798         trace_i915_gem_object_change_domain(obj,
2799                                             obj->read_domains,
2800                                             old_write_domain);
2801 }
2802
2803 /** Flushes the CPU write domain for the object if it's dirty. */
2804 static void
2805 i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj)
2806 {
2807         struct drm_device *dev = obj->dev;
2808         uint32_t old_write_domain;
2809
2810         if (obj->write_domain != I915_GEM_DOMAIN_CPU)
2811                 return;
2812
2813         i915_gem_clflush_object(obj);
2814         drm_agp_chipset_flush(dev);
2815         old_write_domain = obj->write_domain;
2816         obj->write_domain = 0;
2817
2818         trace_i915_gem_object_change_domain(obj,
2819                                             obj->read_domains,
2820                                             old_write_domain);
2821 }
2822
2823 /**
2824  * Moves a single object to the GTT read, and possibly write domain.
2825  *
2826  * This function returns when the move is complete, including waiting on
2827  * flushes to occur.
2828  */
2829 int
2830 i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
2831 {
2832         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2833         uint32_t old_write_domain, old_read_domains;
2834         int ret;
2835
2836         /* Not valid to be called on unbound objects. */
2837         if (obj_priv->gtt_space == NULL)
2838                 return -EINVAL;
2839
2840         ret = i915_gem_object_flush_gpu_write_domain(obj, false);
2841         if (ret != 0)
2842                 return ret;
2843
2844         i915_gem_object_flush_cpu_write_domain(obj);
2845
2846         if (write) {
2847                 ret = i915_gem_object_wait_rendering(obj, true);
2848                 if (ret)
2849                         return ret;
2850         }
2851
2852         old_write_domain = obj->write_domain;
2853         old_read_domains = obj->read_domains;
2854
2855         /* It should now be out of any other write domains, and we can update
2856          * the domain values for our changes.
2857          */
2858         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2859         obj->read_domains |= I915_GEM_DOMAIN_GTT;
2860         if (write) {
2861                 obj->read_domains = I915_GEM_DOMAIN_GTT;
2862                 obj->write_domain = I915_GEM_DOMAIN_GTT;
2863                 obj_priv->dirty = 1;
2864         }
2865
2866         trace_i915_gem_object_change_domain(obj,
2867                                             old_read_domains,
2868                                             old_write_domain);
2869
2870         return 0;
2871 }
2872
2873 /*
2874  * Prepare buffer for display plane. Use uninterruptible for possible flush
2875  * wait, as in modesetting process we're not supposed to be interrupted.
2876  */
2877 int
2878 i915_gem_object_set_to_display_plane(struct drm_gem_object *obj,
2879                                      bool pipelined)
2880 {
2881         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2882         uint32_t old_read_domains;
2883         int ret;
2884
2885         /* Not valid to be called on unbound objects. */
2886         if (obj_priv->gtt_space == NULL)
2887                 return -EINVAL;
2888
2889         ret = i915_gem_object_flush_gpu_write_domain(obj, true);
2890         if (ret)
2891                 return ret;
2892
2893         /* Currently, we are always called from an non-interruptible context. */
2894         if (!pipelined) {
2895                 ret = i915_gem_object_wait_rendering(obj, false);
2896                 if (ret)
2897                         return ret;
2898         }
2899
2900         i915_gem_object_flush_cpu_write_domain(obj);
2901
2902         old_read_domains = obj->read_domains;
2903         obj->read_domains |= I915_GEM_DOMAIN_GTT;
2904
2905         trace_i915_gem_object_change_domain(obj,
2906                                             old_read_domains,
2907                                             obj->write_domain);
2908
2909         return 0;
2910 }
2911
2912 /**
2913  * Moves a single object to the CPU read, and possibly write domain.
2914  *
2915  * This function returns when the move is complete, including waiting on
2916  * flushes to occur.
2917  */
2918 static int
2919 i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
2920 {
2921         uint32_t old_write_domain, old_read_domains;
2922         int ret;
2923
2924         ret = i915_gem_object_flush_gpu_write_domain(obj, false);
2925         if (ret != 0)
2926                 return ret;
2927
2928         i915_gem_object_flush_gtt_write_domain(obj);
2929
2930         /* If we have a partially-valid cache of the object in the CPU,
2931          * finish invalidating it and free the per-page flags.
2932          */
2933         i915_gem_object_set_to_full_cpu_read_domain(obj);
2934
2935         if (write) {
2936                 ret = i915_gem_object_wait_rendering(obj, true);
2937                 if (ret)
2938                         return ret;
2939         }
2940
2941         old_write_domain = obj->write_domain;
2942         old_read_domains = obj->read_domains;
2943
2944         /* Flush the CPU cache if it's still invalid. */
2945         if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
2946                 i915_gem_clflush_object(obj);
2947
2948                 obj->read_domains |= I915_GEM_DOMAIN_CPU;
2949         }
2950
2951         /* It should now be out of any other write domains, and we can update
2952          * the domain values for our changes.
2953          */
2954         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2955
2956         /* If we're writing through the CPU, then the GPU read domains will
2957          * need to be invalidated at next use.
2958          */
2959         if (write) {
2960                 obj->read_domains = I915_GEM_DOMAIN_CPU;
2961                 obj->write_domain = I915_GEM_DOMAIN_CPU;
2962         }
2963
2964         trace_i915_gem_object_change_domain(obj,
2965                                             old_read_domains,
2966                                             old_write_domain);
2967
2968         return 0;
2969 }
2970
2971 /*
2972  * Set the next domain for the specified object. This
2973  * may not actually perform the necessary flushing/invaliding though,
2974  * as that may want to be batched with other set_domain operations
2975  *
2976  * This is (we hope) the only really tricky part of gem. The goal
2977  * is fairly simple -- track which caches hold bits of the object
2978  * and make sure they remain coherent. A few concrete examples may
2979  * help to explain how it works. For shorthand, we use the notation
2980  * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
2981  * a pair of read and write domain masks.
2982  *
2983  * Case 1: the batch buffer
2984  *
2985  *      1. Allocated
2986  *      2. Written by CPU
2987  *      3. Mapped to GTT
2988  *      4. Read by GPU
2989  *      5. Unmapped from GTT
2990  *      6. Freed
2991  *
2992  *      Let's take these a step at a time
2993  *
2994  *      1. Allocated
2995  *              Pages allocated from the kernel may still have
2996  *              cache contents, so we set them to (CPU, CPU) always.
2997  *      2. Written by CPU (using pwrite)
2998  *              The pwrite function calls set_domain (CPU, CPU) and
2999  *              this function does nothing (as nothing changes)
3000  *      3. Mapped by GTT
3001  *              This function asserts that the object is not
3002  *              currently in any GPU-based read or write domains
3003  *      4. Read by GPU
3004  *              i915_gem_execbuffer calls set_domain (COMMAND, 0).
3005  *              As write_domain is zero, this function adds in the
3006  *              current read domains (CPU+COMMAND, 0).
3007  *              flush_domains is set to CPU.
3008  *              invalidate_domains is set to COMMAND
3009  *              clflush is run to get data out of the CPU caches
3010  *              then i915_dev_set_domain calls i915_gem_flush to
3011  *              emit an MI_FLUSH and drm_agp_chipset_flush
3012  *      5. Unmapped from GTT
3013  *              i915_gem_object_unbind calls set_domain (CPU, CPU)
3014  *              flush_domains and invalidate_domains end up both zero
3015  *              so no flushing/invalidating happens
3016  *      6. Freed
3017  *              yay, done
3018  *
3019  * Case 2: The shared render buffer
3020  *
3021  *      1. Allocated
3022  *      2. Mapped to GTT
3023  *      3. Read/written by GPU
3024  *      4. set_domain to (CPU,CPU)
3025  *      5. Read/written by CPU
3026  *      6. Read/written by GPU
3027  *
3028  *      1. Allocated
3029  *              Same as last example, (CPU, CPU)
3030  *      2. Mapped to GTT
3031  *              Nothing changes (assertions find that it is not in the GPU)
3032  *      3. Read/written by GPU
3033  *              execbuffer calls set_domain (RENDER, RENDER)
3034  *              flush_domains gets CPU
3035  *              invalidate_domains gets GPU
3036  *              clflush (obj)
3037  *              MI_FLUSH and drm_agp_chipset_flush
3038  *      4. set_domain (CPU, CPU)
3039  *              flush_domains gets GPU
3040  *              invalidate_domains gets CPU
3041  *              wait_rendering (obj) to make sure all drawing is complete.
3042  *              This will include an MI_FLUSH to get the data from GPU
3043  *              to memory
3044  *              clflush (obj) to invalidate the CPU cache
3045  *              Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
3046  *      5. Read/written by CPU
3047  *              cache lines are loaded and dirtied
3048  *      6. Read written by GPU
3049  *              Same as last GPU access
3050  *
3051  * Case 3: The constant buffer
3052  *
3053  *      1. Allocated
3054  *      2. Written by CPU
3055  *      3. Read by GPU
3056  *      4. Updated (written) by CPU again
3057  *      5. Read by GPU
3058  *
3059  *      1. Allocated
3060  *              (CPU, CPU)
3061  *      2. Written by CPU
3062  *              (CPU, CPU)
3063  *      3. Read by GPU
3064  *              (CPU+RENDER, 0)
3065  *              flush_domains = CPU
3066  *              invalidate_domains = RENDER
3067  *              clflush (obj)
3068  *              MI_FLUSH
3069  *              drm_agp_chipset_flush
3070  *      4. Updated (written) by CPU again
3071  *              (CPU, CPU)
3072  *              flush_domains = 0 (no previous write domain)
3073  *              invalidate_domains = 0 (no new read domains)
3074  *      5. Read by GPU
3075  *              (CPU+RENDER, 0)
3076  *              flush_domains = CPU
3077  *              invalidate_domains = RENDER
3078  *              clflush (obj)
3079  *              MI_FLUSH
3080  *              drm_agp_chipset_flush
3081  */
3082 static void
3083 i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj,
3084                                   struct intel_ring_buffer *ring)
3085 {
3086         struct drm_device               *dev = obj->dev;
3087         struct drm_i915_private         *dev_priv = dev->dev_private;
3088         struct drm_i915_gem_object      *obj_priv = to_intel_bo(obj);
3089         uint32_t                        invalidate_domains = 0;
3090         uint32_t                        flush_domains = 0;
3091         uint32_t                        old_read_domains;
3092
3093         intel_mark_busy(dev, obj);
3094
3095         /*
3096          * If the object isn't moving to a new write domain,
3097          * let the object stay in multiple read domains
3098          */
3099         if (obj->pending_write_domain == 0)
3100                 obj->pending_read_domains |= obj->read_domains;
3101         else
3102                 obj_priv->dirty = 1;
3103
3104         /*
3105          * Flush the current write domain if
3106          * the new read domains don't match. Invalidate
3107          * any read domains which differ from the old
3108          * write domain
3109          */
3110         if (obj->write_domain &&
3111             obj->write_domain != obj->pending_read_domains) {
3112                 flush_domains |= obj->write_domain;
3113                 invalidate_domains |=
3114                         obj->pending_read_domains & ~obj->write_domain;
3115         }
3116         /*
3117          * Invalidate any read caches which may have
3118          * stale data. That is, any new read domains.
3119          */
3120         invalidate_domains |= obj->pending_read_domains & ~obj->read_domains;
3121         if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
3122                 i915_gem_clflush_object(obj);
3123
3124         old_read_domains = obj->read_domains;
3125
3126         /* The actual obj->write_domain will be updated with
3127          * pending_write_domain after we emit the accumulated flush for all
3128          * of our domain changes in execbuffers (which clears objects'
3129          * write_domains).  So if we have a current write domain that we
3130          * aren't changing, set pending_write_domain to that.
3131          */
3132         if (flush_domains == 0 && obj->pending_write_domain == 0)
3133                 obj->pending_write_domain = obj->write_domain;
3134         obj->read_domains = obj->pending_read_domains;
3135
3136         dev->invalidate_domains |= invalidate_domains;
3137         dev->flush_domains |= flush_domains;
3138         if (flush_domains & I915_GEM_GPU_DOMAINS)
3139                 dev_priv->mm.flush_rings |= obj_priv->ring->id;
3140         if (invalidate_domains & I915_GEM_GPU_DOMAINS)
3141                 dev_priv->mm.flush_rings |= ring->id;
3142
3143         trace_i915_gem_object_change_domain(obj,
3144                                             old_read_domains,
3145                                             obj->write_domain);
3146 }
3147
3148 /**
3149  * Moves the object from a partially CPU read to a full one.
3150  *
3151  * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3152  * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3153  */
3154 static void
3155 i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj)
3156 {
3157         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3158
3159         if (!obj_priv->page_cpu_valid)
3160                 return;
3161
3162         /* If we're partially in the CPU read domain, finish moving it in.
3163          */
3164         if (obj->read_domains & I915_GEM_DOMAIN_CPU) {
3165                 int i;
3166
3167                 for (i = 0; i <= (obj->size - 1) / PAGE_SIZE; i++) {
3168                         if (obj_priv->page_cpu_valid[i])
3169                                 continue;
3170                         drm_clflush_pages(obj_priv->pages + i, 1);
3171                 }
3172         }
3173
3174         /* Free the page_cpu_valid mappings which are now stale, whether
3175          * or not we've got I915_GEM_DOMAIN_CPU.
3176          */
3177         kfree(obj_priv->page_cpu_valid);
3178         obj_priv->page_cpu_valid = NULL;
3179 }
3180
3181 /**
3182  * Set the CPU read domain on a range of the object.
3183  *
3184  * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3185  * not entirely valid.  The page_cpu_valid member of the object flags which
3186  * pages have been flushed, and will be respected by
3187  * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3188  * of the whole object.
3189  *
3190  * This function returns when the move is complete, including waiting on
3191  * flushes to occur.
3192  */
3193 static int
3194 i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
3195                                           uint64_t offset, uint64_t size)
3196 {
3197         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3198         uint32_t old_read_domains;
3199         int i, ret;
3200
3201         if (offset == 0 && size == obj->size)
3202                 return i915_gem_object_set_to_cpu_domain(obj, 0);
3203
3204         ret = i915_gem_object_flush_gpu_write_domain(obj, false);
3205         if (ret != 0)
3206                 return ret;
3207         i915_gem_object_flush_gtt_write_domain(obj);
3208
3209         /* If we're already fully in the CPU read domain, we're done. */
3210         if (obj_priv->page_cpu_valid == NULL &&
3211             (obj->read_domains & I915_GEM_DOMAIN_CPU) != 0)
3212                 return 0;
3213
3214         /* Otherwise, create/clear the per-page CPU read domain flag if we're
3215          * newly adding I915_GEM_DOMAIN_CPU
3216          */
3217         if (obj_priv->page_cpu_valid == NULL) {
3218                 obj_priv->page_cpu_valid = kzalloc(obj->size / PAGE_SIZE,
3219                                                    GFP_KERNEL);
3220                 if (obj_priv->page_cpu_valid == NULL)
3221                         return -ENOMEM;
3222         } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0)
3223                 memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
3224
3225         /* Flush the cache on any pages that are still invalid from the CPU's
3226          * perspective.
3227          */
3228         for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3229              i++) {
3230                 if (obj_priv->page_cpu_valid[i])
3231                         continue;
3232
3233                 drm_clflush_pages(obj_priv->pages + i, 1);
3234
3235                 obj_priv->page_cpu_valid[i] = 1;
3236         }
3237
3238         /* It should now be out of any other write domains, and we can update
3239          * the domain values for our changes.
3240          */
3241         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3242
3243         old_read_domains = obj->read_domains;
3244         obj->read_domains |= I915_GEM_DOMAIN_CPU;
3245
3246         trace_i915_gem_object_change_domain(obj,
3247                                             old_read_domains,
3248                                             obj->write_domain);
3249
3250         return 0;
3251 }
3252
3253 /**
3254  * Pin an object to the GTT and evaluate the relocations landing in it.
3255  */
3256 static int
3257 i915_gem_execbuffer_relocate(struct drm_i915_gem_object *obj,
3258                              struct drm_file *file_priv,
3259                              struct drm_i915_gem_exec_object2 *entry)
3260 {
3261         struct drm_device *dev = obj->base.dev;
3262         drm_i915_private_t *dev_priv = dev->dev_private;
3263         struct drm_i915_gem_relocation_entry __user *user_relocs;
3264         struct drm_gem_object *target_obj = NULL;
3265         uint32_t target_handle = 0;
3266         int i, ret = 0;
3267
3268         user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
3269         for (i = 0; i < entry->relocation_count; i++) {
3270                 struct drm_i915_gem_relocation_entry reloc;
3271                 uint32_t target_offset;
3272
3273                 if (__copy_from_user_inatomic(&reloc,
3274                                               user_relocs+i,
3275                                               sizeof(reloc))) {
3276                         ret = -EFAULT;
3277                         break;
3278                 }
3279
3280                 if (reloc.target_handle != target_handle) {
3281                         drm_gem_object_unreference(target_obj);
3282
3283                         target_obj = drm_gem_object_lookup(dev, file_priv,
3284                                                            reloc.target_handle);
3285                         if (target_obj == NULL) {
3286                                 ret = -ENOENT;
3287                                 break;
3288                         }
3289
3290                         target_handle = reloc.target_handle;
3291                 }
3292                 target_offset = to_intel_bo(target_obj)->gtt_offset;
3293
3294 #if WATCH_RELOC
3295                 DRM_INFO("%s: obj %p offset %08x target %d "
3296                          "read %08x write %08x gtt %08x "
3297                          "presumed %08x delta %08x\n",
3298                          __func__,
3299                          obj,
3300                          (int) reloc.offset,
3301                          (int) reloc.target_handle,
3302                          (int) reloc.read_domains,
3303                          (int) reloc.write_domain,
3304                          (int) target_offset,
3305                          (int) reloc.presumed_offset,
3306                          reloc.delta);
3307 #endif
3308
3309                 /* The target buffer should have appeared before us in the
3310                  * exec_object list, so it should have a GTT space bound by now.
3311                  */
3312                 if (target_offset == 0) {
3313                         DRM_ERROR("No GTT space found for object %d\n",
3314                                   reloc.target_handle);
3315                         ret = -EINVAL;
3316                         break;
3317                 }
3318
3319                 /* Validate that the target is in a valid r/w GPU domain */
3320                 if (reloc.write_domain & (reloc.write_domain - 1)) {
3321                         DRM_ERROR("reloc with multiple write domains: "
3322                                   "obj %p target %d offset %d "
3323                                   "read %08x write %08x",
3324                                   obj, reloc.target_handle,
3325                                   (int) reloc.offset,
3326                                   reloc.read_domains,
3327                                   reloc.write_domain);
3328                         ret = -EINVAL;
3329                         break;
3330                 }
3331                 if (reloc.write_domain & I915_GEM_DOMAIN_CPU ||
3332                     reloc.read_domains & I915_GEM_DOMAIN_CPU) {
3333                         DRM_ERROR("reloc with read/write CPU domains: "
3334                                   "obj %p target %d offset %d "
3335                                   "read %08x write %08x",
3336                                   obj, reloc.target_handle,
3337                                   (int) reloc.offset,
3338                                   reloc.read_domains,
3339                                   reloc.write_domain);
3340                         ret = -EINVAL;
3341                         break;
3342                 }
3343                 if (reloc.write_domain && target_obj->pending_write_domain &&
3344                     reloc.write_domain != target_obj->pending_write_domain) {
3345                         DRM_ERROR("Write domain conflict: "
3346                                   "obj %p target %d offset %d "
3347                                   "new %08x old %08x\n",
3348                                   obj, reloc.target_handle,
3349                                   (int) reloc.offset,
3350                                   reloc.write_domain,
3351                                   target_obj->pending_write_domain);
3352                         ret = -EINVAL;
3353                         break;
3354                 }
3355
3356                 target_obj->pending_read_domains |= reloc.read_domains;
3357                 target_obj->pending_write_domain |= reloc.write_domain;
3358
3359                 /* If the relocation already has the right value in it, no
3360                  * more work needs to be done.
3361                  */
3362                 if (target_offset == reloc.presumed_offset)
3363                         continue;
3364
3365                 /* Check that the relocation address is valid... */
3366                 if (reloc.offset > obj->base.size - 4) {
3367                         DRM_ERROR("Relocation beyond object bounds: "
3368                                   "obj %p target %d offset %d size %d.\n",
3369                                   obj, reloc.target_handle,
3370                                   (int) reloc.offset, (int) obj->base.size);
3371                         ret = -EINVAL;
3372                         break;
3373                 }
3374                 if (reloc.offset & 3) {
3375                         DRM_ERROR("Relocation not 4-byte aligned: "
3376                                   "obj %p target %d offset %d.\n",
3377                                   obj, reloc.target_handle,
3378                                   (int) reloc.offset);
3379                         ret = -EINVAL;
3380                         break;
3381                 }
3382
3383                 /* and points to somewhere within the target object. */
3384                 if (reloc.delta >= target_obj->size) {
3385                         DRM_ERROR("Relocation beyond target object bounds: "
3386                                   "obj %p target %d delta %d size %d.\n",
3387                                   obj, reloc.target_handle,
3388                                   (int) reloc.delta, (int) target_obj->size);
3389                         ret = -EINVAL;
3390                         break;
3391                 }
3392
3393                 reloc.delta += target_offset;
3394                 if (obj->base.write_domain == I915_GEM_DOMAIN_CPU) {
3395                         uint32_t page_offset = reloc.offset & ~PAGE_MASK;
3396                         char *vaddr;
3397
3398                         vaddr = kmap_atomic(obj->pages[reloc.offset >> PAGE_SHIFT], KM_USER0);
3399                         *(uint32_t *)(vaddr + page_offset) = reloc.delta;
3400                         kunmap_atomic(vaddr, KM_USER0);
3401                 } else {
3402                         uint32_t __iomem *reloc_entry;
3403                         void __iomem *reloc_page;
3404
3405                         ret = i915_gem_object_set_to_gtt_domain(&obj->base, 1);
3406                         if (ret)
3407                                 break;
3408
3409                         /* Map the page containing the relocation we're going to perform.  */
3410                         reloc.offset += obj->gtt_offset;
3411                         reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
3412                                                               reloc.offset & PAGE_MASK,
3413                                                               KM_USER0);
3414                         reloc_entry = (uint32_t __iomem *)
3415                                 (reloc_page + (reloc.offset & ~PAGE_MASK));
3416                         iowrite32(reloc.delta, reloc_entry);
3417                         io_mapping_unmap_atomic(reloc_page, KM_USER0);
3418                 }
3419
3420                 /* and update the user's relocation entry */
3421                 reloc.presumed_offset = target_offset;
3422                 if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset,
3423                                               &reloc.presumed_offset,
3424                                               sizeof(reloc.presumed_offset))) {
3425                     ret = -EFAULT;
3426                     break;
3427                 }
3428         }
3429
3430         drm_gem_object_unreference(target_obj);
3431         return ret;
3432 }
3433
3434 static int
3435 i915_gem_execbuffer_pin(struct drm_device *dev,
3436                         struct drm_file *file,
3437                         struct drm_gem_object **object_list,
3438                         struct drm_i915_gem_exec_object2 *exec_list,
3439                         int count)
3440 {
3441         struct drm_i915_private *dev_priv = dev->dev_private;
3442         int ret, i, retry;
3443
3444         /* attempt to pin all of the buffers into the GTT */
3445         for (retry = 0; retry < 2; retry++) {
3446                 ret = 0;
3447                 for (i = 0; i < count; i++) {
3448                         struct drm_i915_gem_exec_object2 *entry = &exec_list[i];
3449                         struct drm_i915_gem_object *obj= to_intel_bo(object_list[i]);
3450                         bool need_fence =
3451                                 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
3452                                 obj->tiling_mode != I915_TILING_NONE;
3453
3454                         /* Check fence reg constraints and rebind if necessary */
3455                         if (need_fence &&
3456                             !i915_gem_object_fence_offset_ok(&obj->base,
3457                                                              obj->tiling_mode)) {
3458                                 ret = i915_gem_object_unbind(&obj->base);
3459                                 if (ret)
3460                                         break;
3461                         }
3462
3463                         ret = i915_gem_object_pin(&obj->base, entry->alignment);
3464                         if (ret)
3465                                 break;
3466
3467                         /*
3468                          * Pre-965 chips need a fence register set up in order
3469                          * to properly handle blits to/from tiled surfaces.
3470                          */
3471                         if (need_fence) {
3472                                 ret = i915_gem_object_get_fence_reg(&obj->base, true);
3473                                 if (ret) {
3474                                         i915_gem_object_unpin(&obj->base);
3475                                         break;
3476                                 }
3477
3478                                 dev_priv->fence_regs[obj->fence_reg].gpu = true;
3479                         }
3480
3481                         entry->offset = obj->gtt_offset;
3482                 }
3483
3484                 while (i--)
3485                         i915_gem_object_unpin(object_list[i]);
3486
3487                 if (ret == 0)
3488                         break;
3489
3490                 if (ret != -ENOSPC || retry)
3491                         return ret;
3492
3493                 ret = i915_gem_evict_everything(dev);
3494                 if (ret)
3495                         return ret;
3496         }
3497
3498         return 0;
3499 }
3500
3501 /* Throttle our rendering by waiting until the ring has completed our requests
3502  * emitted over 20 msec ago.
3503  *
3504  * Note that if we were to use the current jiffies each time around the loop,
3505  * we wouldn't escape the function with any frames outstanding if the time to
3506  * render a frame was over 20ms.
3507  *
3508  * This should get us reasonable parallelism between CPU and GPU but also
3509  * relatively low latency when blocking on a particular request to finish.
3510  */
3511 static int
3512 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
3513 {
3514         struct drm_i915_private *dev_priv = dev->dev_private;
3515         struct drm_i915_file_private *file_priv = file->driver_priv;
3516         unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3517         struct drm_i915_gem_request *request;
3518         struct intel_ring_buffer *ring = NULL;
3519         u32 seqno = 0;
3520         int ret;
3521
3522         spin_lock(&file_priv->mm.lock);
3523         list_for_each_entry(request, &file_priv->mm.request_list, client_list) {
3524                 if (time_after_eq(request->emitted_jiffies, recent_enough))
3525                         break;
3526
3527                 ring = request->ring;
3528                 seqno = request->seqno;
3529         }
3530         spin_unlock(&file_priv->mm.lock);
3531
3532         if (seqno == 0)
3533                 return 0;
3534
3535         ret = 0;
3536         if (!i915_seqno_passed(ring->get_seqno(dev, ring), seqno)) {
3537                 /* And wait for the seqno passing without holding any locks and
3538                  * causing extra latency for others. This is safe as the irq
3539                  * generation is designed to be run atomically and so is
3540                  * lockless.
3541                  */
3542                 ring->user_irq_get(dev, ring);
3543                 ret = wait_event_interruptible(ring->irq_queue,
3544                                                i915_seqno_passed(ring->get_seqno(dev, ring), seqno)
3545                                                || atomic_read(&dev_priv->mm.wedged));
3546                 ring->user_irq_put(dev, ring);
3547
3548                 if (ret == 0 && atomic_read(&dev_priv->mm.wedged))
3549                         ret = -EIO;
3550         }
3551
3552         if (ret == 0)
3553                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
3554
3555         return ret;
3556 }
3557
3558 static int
3559 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec,
3560                           uint64_t exec_offset)
3561 {
3562         uint32_t exec_start, exec_len;
3563
3564         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3565         exec_len = (uint32_t) exec->batch_len;
3566
3567         if ((exec_start | exec_len) & 0x7)
3568                 return -EINVAL;
3569
3570         if (!exec_start)
3571                 return -EINVAL;
3572
3573         return 0;
3574 }
3575
3576 static int
3577 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
3578                    int count)
3579 {
3580         int i;
3581
3582         for (i = 0; i < count; i++) {
3583                 char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
3584                 size_t length = exec[i].relocation_count * sizeof(struct drm_i915_gem_relocation_entry);
3585
3586                 if (!access_ok(VERIFY_READ, ptr, length))
3587                         return -EFAULT;
3588
3589                 /* we may also need to update the presumed offsets */
3590                 if (!access_ok(VERIFY_WRITE, ptr, length))
3591                         return -EFAULT;
3592
3593                 if (fault_in_pages_readable(ptr, length))
3594                         return -EFAULT;
3595         }
3596
3597         return 0;
3598 }
3599
3600 static int
3601 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
3602                        struct drm_file *file,
3603                        struct drm_i915_gem_execbuffer2 *args,
3604                        struct drm_i915_gem_exec_object2 *exec_list)
3605 {
3606         drm_i915_private_t *dev_priv = dev->dev_private;
3607         struct drm_gem_object **object_list = NULL;
3608         struct drm_gem_object *batch_obj;
3609         struct drm_i915_gem_object *obj_priv;
3610         struct drm_clip_rect *cliprects = NULL;
3611         struct drm_i915_gem_request *request = NULL;
3612         int ret, i, flips;
3613         uint64_t exec_offset;
3614
3615         struct intel_ring_buffer *ring = NULL;
3616
3617         ret = i915_gem_check_is_wedged(dev);
3618         if (ret)
3619                 return ret;
3620
3621         ret = validate_exec_list(exec_list, args->buffer_count);
3622         if (ret)
3623                 return ret;
3624
3625 #if WATCH_EXEC
3626         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3627                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3628 #endif
3629         switch (args->flags & I915_EXEC_RING_MASK) {
3630         case I915_EXEC_DEFAULT:
3631         case I915_EXEC_RENDER:
3632                 ring = &dev_priv->render_ring;
3633                 break;
3634         case I915_EXEC_BSD:
3635                 if (!HAS_BSD(dev)) {
3636                         DRM_ERROR("execbuf with invalid ring (BSD)\n");
3637                         return -EINVAL;
3638                 }
3639                 ring = &dev_priv->bsd_ring;
3640                 break;
3641         case I915_EXEC_BLT:
3642                 if (!HAS_BLT(dev)) {
3643                         DRM_ERROR("execbuf with invalid ring (BLT)\n");
3644                         return -EINVAL;
3645                 }
3646                 ring = &dev_priv->blt_ring;
3647                 break;
3648         default:
3649                 DRM_ERROR("execbuf with unknown ring: %d\n",
3650                           (int)(args->flags & I915_EXEC_RING_MASK));
3651                 return -EINVAL;
3652         }
3653
3654         if (args->buffer_count < 1) {
3655                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3656                 return -EINVAL;
3657         }
3658         object_list = drm_malloc_ab(sizeof(*object_list), args->buffer_count);
3659         if (object_list == NULL) {
3660                 DRM_ERROR("Failed to allocate object list for %d buffers\n",
3661                           args->buffer_count);
3662                 ret = -ENOMEM;
3663                 goto pre_mutex_err;
3664         }
3665
3666         if (args->num_cliprects != 0) {
3667                 cliprects = kcalloc(args->num_cliprects, sizeof(*cliprects),
3668                                     GFP_KERNEL);
3669                 if (cliprects == NULL) {
3670                         ret = -ENOMEM;
3671                         goto pre_mutex_err;
3672                 }
3673
3674                 ret = copy_from_user(cliprects,
3675                                      (struct drm_clip_rect __user *)
3676                                      (uintptr_t) args->cliprects_ptr,
3677                                      sizeof(*cliprects) * args->num_cliprects);
3678                 if (ret != 0) {
3679                         DRM_ERROR("copy %d cliprects failed: %d\n",
3680                                   args->num_cliprects, ret);
3681                         ret = -EFAULT;
3682                         goto pre_mutex_err;
3683                 }
3684         }
3685
3686         request = kzalloc(sizeof(*request), GFP_KERNEL);
3687         if (request == NULL) {
3688                 ret = -ENOMEM;
3689                 goto pre_mutex_err;
3690         }
3691
3692         ret = i915_mutex_lock_interruptible(dev);
3693         if (ret)
3694                 goto pre_mutex_err;
3695
3696         if (dev_priv->mm.suspended) {
3697                 mutex_unlock(&dev->struct_mutex);
3698                 ret = -EBUSY;
3699                 goto pre_mutex_err;
3700         }
3701
3702         /* Look up object handles */
3703         for (i = 0; i < args->buffer_count; i++) {
3704                 object_list[i] = drm_gem_object_lookup(dev, file,
3705                                                        exec_list[i].handle);
3706                 if (object_list[i] == NULL) {
3707                         DRM_ERROR("Invalid object handle %d at index %d\n",
3708                                    exec_list[i].handle, i);
3709                         /* prevent error path from reading uninitialized data */
3710                         args->buffer_count = i + 1;
3711                         ret = -ENOENT;
3712                         goto err;
3713                 }
3714
3715                 obj_priv = to_intel_bo(object_list[i]);
3716                 if (obj_priv->in_execbuffer) {
3717                         DRM_ERROR("Object %p appears more than once in object list\n",
3718                                    object_list[i]);
3719                         /* prevent error path from reading uninitialized data */
3720                         args->buffer_count = i + 1;
3721                         ret = -EINVAL;
3722                         goto err;
3723                 }
3724                 obj_priv->in_execbuffer = true;
3725         }
3726
3727         /* Move the objects en-masse into the GTT, evicting if necessary. */
3728         ret = i915_gem_execbuffer_pin(dev, file,
3729                                       object_list, exec_list,
3730                                       args->buffer_count);
3731         if (ret)
3732                 goto err;
3733
3734         /* The objects are in their final locations, apply the relocations. */
3735         for (i = 0; i < args->buffer_count; i++) {
3736                 struct drm_i915_gem_object *obj = to_intel_bo(object_list[i]);
3737                 obj->base.pending_read_domains = 0;
3738                 obj->base.pending_write_domain = 0;
3739                 ret = i915_gem_execbuffer_relocate(obj, file, &exec_list[i]);
3740                 if (ret)
3741                         goto err;
3742         }
3743
3744         /* Set the pending read domains for the batch buffer to COMMAND */
3745         batch_obj = object_list[args->buffer_count-1];
3746         if (batch_obj->pending_write_domain) {
3747                 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
3748                 ret = -EINVAL;
3749                 goto err;
3750         }
3751         batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
3752
3753         /* Sanity check the batch buffer */
3754         exec_offset = to_intel_bo(batch_obj)->gtt_offset;
3755         ret = i915_gem_check_execbuffer(args, exec_offset);
3756         if (ret != 0) {
3757                 DRM_ERROR("execbuf with invalid offset/length\n");
3758                 goto err;
3759         }
3760
3761         /* Zero the global flush/invalidate flags. These
3762          * will be modified as new domains are computed
3763          * for each object
3764          */
3765         dev->invalidate_domains = 0;
3766         dev->flush_domains = 0;
3767         dev_priv->mm.flush_rings = 0;
3768
3769         for (i = 0; i < args->buffer_count; i++) {
3770                 struct drm_gem_object *obj = object_list[i];
3771
3772                 /* Compute new gpu domains and update invalidate/flush */
3773                 i915_gem_object_set_to_gpu_domain(obj, ring);
3774         }
3775
3776         if (dev->invalidate_domains | dev->flush_domains) {
3777 #if WATCH_EXEC
3778                 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
3779                           __func__,
3780                          dev->invalidate_domains,
3781                          dev->flush_domains);
3782 #endif
3783                 i915_gem_flush(dev, file,
3784                                dev->invalidate_domains,
3785                                dev->flush_domains,
3786                                dev_priv->mm.flush_rings);
3787         }
3788
3789         for (i = 0; i < args->buffer_count; i++) {
3790                 struct drm_gem_object *obj = object_list[i];
3791                 uint32_t old_write_domain = obj->write_domain;
3792                 obj->write_domain = obj->pending_write_domain;
3793                 trace_i915_gem_object_change_domain(obj,
3794                                                     obj->read_domains,
3795                                                     old_write_domain);
3796         }
3797
3798 #if WATCH_COHERENCY
3799         for (i = 0; i < args->buffer_count; i++) {
3800                 i915_gem_object_check_coherency(object_list[i],
3801                                                 exec_list[i].handle);
3802         }
3803 #endif
3804
3805 #if WATCH_EXEC
3806         i915_gem_dump_object(batch_obj,
3807                               args->batch_len,
3808                               __func__,
3809                               ~0);
3810 #endif
3811
3812         /* Check for any pending flips. As we only maintain a flip queue depth
3813          * of 1, we can simply insert a WAIT for the next display flip prior
3814          * to executing the batch and avoid stalling the CPU.
3815          */
3816         flips = 0;
3817         for (i = 0; i < args->buffer_count; i++) {
3818                 if (object_list[i]->write_domain)
3819                         flips |= atomic_read(&to_intel_bo(object_list[i])->pending_flip);
3820         }
3821         if (flips) {
3822                 int plane, flip_mask;
3823
3824                 for (plane = 0; flips >> plane; plane++) {
3825                         if (((flips >> plane) & 1) == 0)
3826                                 continue;
3827
3828                         if (plane)
3829                                 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
3830                         else
3831                                 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
3832
3833                         intel_ring_begin(dev, ring, 2);
3834                         intel_ring_emit(dev, ring,
3835                                         MI_WAIT_FOR_EVENT | flip_mask);
3836                         intel_ring_emit(dev, ring, MI_NOOP);
3837                         intel_ring_advance(dev, ring);
3838                 }
3839         }
3840
3841         /* Exec the batchbuffer */
3842         ret = ring->dispatch_gem_execbuffer(dev, ring, args,
3843                                             cliprects, exec_offset);
3844         if (ret) {
3845                 DRM_ERROR("dispatch failed %d\n", ret);
3846                 goto err;
3847         }
3848
3849         /*
3850          * Ensure that the commands in the batch buffer are
3851          * finished before the interrupt fires
3852          */
3853         i915_retire_commands(dev, ring);
3854
3855         for (i = 0; i < args->buffer_count; i++) {
3856                 struct drm_gem_object *obj = object_list[i];
3857
3858                 i915_gem_object_move_to_active(obj, ring);
3859                 if (obj->write_domain)
3860                         list_move_tail(&to_intel_bo(obj)->gpu_write_list,
3861                                        &ring->gpu_write_list);
3862         }
3863
3864         i915_add_request(dev, file, request, ring);
3865         request = NULL;
3866
3867 err:
3868         for (i = 0; i < args->buffer_count; i++) {
3869                 if (object_list[i]) {
3870                         obj_priv = to_intel_bo(object_list[i]);
3871                         obj_priv->in_execbuffer = false;
3872                 }
3873                 drm_gem_object_unreference(object_list[i]);
3874         }
3875
3876         mutex_unlock(&dev->struct_mutex);
3877
3878 pre_mutex_err:
3879         drm_free_large(object_list);
3880         kfree(cliprects);
3881         kfree(request);
3882
3883         return ret;
3884 }
3885
3886 /*
3887  * Legacy execbuffer just creates an exec2 list from the original exec object
3888  * list array and passes it to the real function.
3889  */
3890 int
3891 i915_gem_execbuffer(struct drm_device *dev, void *data,
3892                     struct drm_file *file_priv)
3893 {
3894         struct drm_i915_gem_execbuffer *args = data;
3895         struct drm_i915_gem_execbuffer2 exec2;
3896         struct drm_i915_gem_exec_object *exec_list = NULL;
3897         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
3898         int ret, i;
3899
3900 #if WATCH_EXEC
3901         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3902                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3903 #endif
3904
3905         if (args->buffer_count < 1) {
3906                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3907                 return -EINVAL;
3908         }
3909
3910         /* Copy in the exec list from userland */
3911         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
3912         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
3913         if (exec_list == NULL || exec2_list == NULL) {
3914                 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
3915                           args->buffer_count);
3916                 drm_free_large(exec_list);
3917                 drm_free_large(exec2_list);
3918                 return -ENOMEM;
3919         }
3920         ret = copy_from_user(exec_list,
3921                              (struct drm_i915_relocation_entry __user *)
3922                              (uintptr_t) args->buffers_ptr,
3923                              sizeof(*exec_list) * args->buffer_count);
3924         if (ret != 0) {
3925                 DRM_ERROR("copy %d exec entries failed %d\n",
3926                           args->buffer_count, ret);
3927                 drm_free_large(exec_list);
3928                 drm_free_large(exec2_list);
3929                 return -EFAULT;
3930         }
3931
3932         for (i = 0; i < args->buffer_count; i++) {
3933                 exec2_list[i].handle = exec_list[i].handle;
3934                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
3935                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
3936                 exec2_list[i].alignment = exec_list[i].alignment;
3937                 exec2_list[i].offset = exec_list[i].offset;
3938                 if (INTEL_INFO(dev)->gen < 4)
3939                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
3940                 else
3941                         exec2_list[i].flags = 0;
3942         }
3943
3944         exec2.buffers_ptr = args->buffers_ptr;
3945         exec2.buffer_count = args->buffer_count;
3946         exec2.batch_start_offset = args->batch_start_offset;
3947         exec2.batch_len = args->batch_len;
3948         exec2.DR1 = args->DR1;
3949         exec2.DR4 = args->DR4;
3950         exec2.num_cliprects = args->num_cliprects;
3951         exec2.cliprects_ptr = args->cliprects_ptr;
3952         exec2.flags = I915_EXEC_RENDER;
3953
3954         ret = i915_gem_do_execbuffer(dev, data, file_priv, &exec2, exec2_list);
3955         if (!ret) {
3956                 /* Copy the new buffer offsets back to the user's exec list. */
3957                 for (i = 0; i < args->buffer_count; i++)
3958                         exec_list[i].offset = exec2_list[i].offset;
3959                 /* ... and back out to userspace */
3960                 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
3961                                    (uintptr_t) args->buffers_ptr,
3962                                    exec_list,
3963                                    sizeof(*exec_list) * args->buffer_count);
3964                 if (ret) {
3965                         ret = -EFAULT;
3966                         DRM_ERROR("failed to copy %d exec entries "
3967                                   "back to user (%d)\n",
3968                                   args->buffer_count, ret);
3969                 }
3970         }
3971
3972         drm_free_large(exec_list);
3973         drm_free_large(exec2_list);
3974         return ret;
3975 }
3976
3977 int
3978 i915_gem_execbuffer2(struct drm_device *dev, void *data,
3979                      struct drm_file *file_priv)
3980 {
3981         struct drm_i915_gem_execbuffer2 *args = data;
3982         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
3983         int ret;
3984
3985 #if WATCH_EXEC
3986         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3987                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3988 #endif
3989
3990         if (args->buffer_count < 1) {
3991                 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
3992                 return -EINVAL;
3993         }
3994
3995         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
3996         if (exec2_list == NULL) {
3997                 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
3998                           args->buffer_count);
3999                 return -ENOMEM;
4000         }
4001         ret = copy_from_user(exec2_list,
4002                              (struct drm_i915_relocation_entry __user *)
4003                              (uintptr_t) args->buffers_ptr,
4004                              sizeof(*exec2_list) * args->buffer_count);
4005         if (ret != 0) {
4006                 DRM_ERROR("copy %d exec entries failed %d\n",
4007                           args->buffer_count, ret);
4008                 drm_free_large(exec2_list);
4009                 return -EFAULT;
4010         }
4011
4012         ret = i915_gem_do_execbuffer(dev, data, file_priv, args, exec2_list);
4013         if (!ret) {
4014                 /* Copy the new buffer offsets back to the user's exec list. */
4015                 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4016                                    (uintptr_t) args->buffers_ptr,
4017                                    exec2_list,
4018                                    sizeof(*exec2_list) * args->buffer_count);
4019                 if (ret) {
4020                         ret = -EFAULT;
4021                         DRM_ERROR("failed to copy %d exec entries "
4022                                   "back to user (%d)\n",
4023                                   args->buffer_count, ret);
4024                 }
4025         }
4026
4027         drm_free_large(exec2_list);
4028         return ret;
4029 }
4030
4031 int
4032 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
4033 {
4034         struct drm_device *dev = obj->dev;
4035         struct drm_i915_private *dev_priv = dev->dev_private;
4036         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4037         int ret;
4038
4039         BUG_ON(obj_priv->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
4040         WARN_ON(i915_verify_lists(dev));
4041
4042         if (obj_priv->gtt_space != NULL) {
4043                 if (alignment == 0)
4044                         alignment = i915_gem_get_gtt_alignment(obj);
4045                 if (obj_priv->gtt_offset & (alignment - 1)) {
4046                         WARN(obj_priv->pin_count,
4047                              "bo is already pinned with incorrect alignment:"
4048                              " offset=%x, req.alignment=%x\n",
4049                              obj_priv->gtt_offset, alignment);
4050                         ret = i915_gem_object_unbind(obj);
4051                         if (ret)
4052                                 return ret;
4053                 }
4054         }
4055
4056         if (obj_priv->gtt_space == NULL) {
4057                 ret = i915_gem_object_bind_to_gtt(obj, alignment);
4058                 if (ret)
4059                         return ret;
4060         }
4061
4062         obj_priv->pin_count++;
4063
4064         /* If the object is not active and not pending a flush,
4065          * remove it from the inactive list
4066          */
4067         if (obj_priv->pin_count == 1) {
4068                 i915_gem_info_add_pin(dev_priv, obj->size);
4069                 if (!obj_priv->active)
4070                         list_move_tail(&obj_priv->mm_list,
4071                                        &dev_priv->mm.pinned_list);
4072         }
4073
4074         WARN_ON(i915_verify_lists(dev));
4075         return 0;
4076 }
4077
4078 void
4079 i915_gem_object_unpin(struct drm_gem_object *obj)
4080 {
4081         struct drm_device *dev = obj->dev;
4082         drm_i915_private_t *dev_priv = dev->dev_private;
4083         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4084
4085         WARN_ON(i915_verify_lists(dev));
4086         obj_priv->pin_count--;
4087         BUG_ON(obj_priv->pin_count < 0);
4088         BUG_ON(obj_priv->gtt_space == NULL);
4089
4090         /* If the object is no longer pinned, and is
4091          * neither active nor being flushed, then stick it on
4092          * the inactive list
4093          */
4094         if (obj_priv->pin_count == 0) {
4095                 if (!obj_priv->active)
4096                         list_move_tail(&obj_priv->mm_list,
4097                                        &dev_priv->mm.inactive_list);
4098                 i915_gem_info_remove_pin(dev_priv, obj->size);
4099         }
4100         WARN_ON(i915_verify_lists(dev));
4101 }
4102
4103 int
4104 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4105                    struct drm_file *file_priv)
4106 {
4107         struct drm_i915_gem_pin *args = data;
4108         struct drm_gem_object *obj;
4109         struct drm_i915_gem_object *obj_priv;
4110         int ret;
4111
4112         ret = i915_mutex_lock_interruptible(dev);
4113         if (ret)
4114                 return ret;
4115
4116         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4117         if (obj == NULL) {
4118                 ret = -ENOENT;
4119                 goto unlock;
4120         }
4121         obj_priv = to_intel_bo(obj);
4122
4123         if (obj_priv->madv != I915_MADV_WILLNEED) {
4124                 DRM_ERROR("Attempting to pin a purgeable buffer\n");
4125                 ret = -EINVAL;
4126                 goto out;
4127         }
4128
4129         if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) {
4130                 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
4131                           args->handle);
4132                 ret = -EINVAL;
4133                 goto out;
4134         }
4135
4136         obj_priv->user_pin_count++;
4137         obj_priv->pin_filp = file_priv;
4138         if (obj_priv->user_pin_count == 1) {
4139                 ret = i915_gem_object_pin(obj, args->alignment);
4140                 if (ret)
4141                         goto out;
4142         }
4143
4144         /* XXX - flush the CPU caches for pinned objects
4145          * as the X server doesn't manage domains yet
4146          */
4147         i915_gem_object_flush_cpu_write_domain(obj);
4148         args->offset = obj_priv->gtt_offset;
4149 out:
4150         drm_gem_object_unreference(obj);
4151 unlock:
4152         mutex_unlock(&dev->struct_mutex);
4153         return ret;
4154 }
4155
4156 int
4157 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4158                      struct drm_file *file_priv)
4159 {
4160         struct drm_i915_gem_pin *args = data;
4161         struct drm_gem_object *obj;
4162         struct drm_i915_gem_object *obj_priv;
4163         int ret;
4164
4165         ret = i915_mutex_lock_interruptible(dev);
4166         if (ret)
4167                 return ret;
4168
4169         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4170         if (obj == NULL) {
4171                 ret = -ENOENT;
4172                 goto unlock;
4173         }
4174         obj_priv = to_intel_bo(obj);
4175
4176         if (obj_priv->pin_filp != file_priv) {
4177                 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4178                           args->handle);
4179                 ret = -EINVAL;
4180                 goto out;
4181         }
4182         obj_priv->user_pin_count--;
4183         if (obj_priv->user_pin_count == 0) {
4184                 obj_priv->pin_filp = NULL;
4185                 i915_gem_object_unpin(obj);
4186         }
4187
4188 out:
4189         drm_gem_object_unreference(obj);
4190 unlock:
4191         mutex_unlock(&dev->struct_mutex);
4192         return ret;
4193 }
4194
4195 int
4196 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4197                     struct drm_file *file_priv)
4198 {
4199         struct drm_i915_gem_busy *args = data;
4200         struct drm_gem_object *obj;
4201         struct drm_i915_gem_object *obj_priv;
4202         int ret;
4203
4204         ret = i915_mutex_lock_interruptible(dev);
4205         if (ret)
4206                 return ret;
4207
4208         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4209         if (obj == NULL) {
4210                 ret = -ENOENT;
4211                 goto unlock;
4212         }
4213         obj_priv = to_intel_bo(obj);
4214
4215         /* Count all active objects as busy, even if they are currently not used
4216          * by the gpu. Users of this interface expect objects to eventually
4217          * become non-busy without any further actions, therefore emit any
4218          * necessary flushes here.
4219          */
4220         args->busy = obj_priv->active;
4221         if (args->busy) {
4222                 /* Unconditionally flush objects, even when the gpu still uses this
4223                  * object. Userspace calling this function indicates that it wants to
4224                  * use this buffer rather sooner than later, so issuing the required
4225                  * flush earlier is beneficial.
4226                  */
4227                 if (obj->write_domain & I915_GEM_GPU_DOMAINS)
4228                         i915_gem_flush_ring(dev, file_priv,
4229                                             obj_priv->ring,
4230                                             0, obj->write_domain);
4231
4232                 /* Update the active list for the hardware's current position.
4233                  * Otherwise this only updates on a delayed timer or when irqs
4234                  * are actually unmasked, and our working set ends up being
4235                  * larger than required.
4236                  */
4237                 i915_gem_retire_requests_ring(dev, obj_priv->ring);
4238
4239                 args->busy = obj_priv->active;
4240         }
4241
4242         drm_gem_object_unreference(obj);
4243 unlock:
4244         mutex_unlock(&dev->struct_mutex);
4245         return ret;
4246 }
4247
4248 int
4249 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4250                         struct drm_file *file_priv)
4251 {
4252     return i915_gem_ring_throttle(dev, file_priv);
4253 }
4254
4255 int
4256 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4257                        struct drm_file *file_priv)
4258 {
4259         struct drm_i915_gem_madvise *args = data;
4260         struct drm_gem_object *obj;
4261         struct drm_i915_gem_object *obj_priv;
4262         int ret;
4263
4264         switch (args->madv) {
4265         case I915_MADV_DONTNEED:
4266         case I915_MADV_WILLNEED:
4267             break;
4268         default:
4269             return -EINVAL;
4270         }
4271
4272         ret = i915_mutex_lock_interruptible(dev);
4273         if (ret)
4274                 return ret;
4275
4276         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4277         if (obj == NULL) {
4278                 ret = -ENOENT;
4279                 goto unlock;
4280         }
4281         obj_priv = to_intel_bo(obj);
4282
4283         if (obj_priv->pin_count) {
4284                 ret = -EINVAL;
4285                 goto out;
4286         }
4287
4288         if (obj_priv->madv != __I915_MADV_PURGED)
4289                 obj_priv->madv = args->madv;
4290
4291         /* if the object is no longer bound, discard its backing storage */
4292         if (i915_gem_object_is_purgeable(obj_priv) &&
4293             obj_priv->gtt_space == NULL)
4294                 i915_gem_object_truncate(obj);
4295
4296         args->retained = obj_priv->madv != __I915_MADV_PURGED;
4297
4298 out:
4299         drm_gem_object_unreference(obj);
4300 unlock:
4301         mutex_unlock(&dev->struct_mutex);
4302         return ret;
4303 }
4304
4305 struct drm_gem_object * i915_gem_alloc_object(struct drm_device *dev,
4306                                               size_t size)
4307 {
4308         struct drm_i915_private *dev_priv = dev->dev_private;
4309         struct drm_i915_gem_object *obj;
4310
4311         obj = kzalloc(sizeof(*obj), GFP_KERNEL);
4312         if (obj == NULL)
4313                 return NULL;
4314
4315         if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4316                 kfree(obj);
4317                 return NULL;
4318         }
4319
4320         i915_gem_info_add_obj(dev_priv, size);
4321
4322         obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4323         obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4324
4325         obj->agp_type = AGP_USER_MEMORY;
4326         obj->base.driver_private = NULL;
4327         obj->fence_reg = I915_FENCE_REG_NONE;
4328         INIT_LIST_HEAD(&obj->mm_list);
4329         INIT_LIST_HEAD(&obj->ring_list);
4330         INIT_LIST_HEAD(&obj->gpu_write_list);
4331         obj->madv = I915_MADV_WILLNEED;
4332
4333         return &obj->base;
4334 }
4335
4336 int i915_gem_init_object(struct drm_gem_object *obj)
4337 {
4338         BUG();
4339
4340         return 0;
4341 }
4342
4343 static void i915_gem_free_object_tail(struct drm_gem_object *obj)
4344 {
4345         struct drm_device *dev = obj->dev;
4346         drm_i915_private_t *dev_priv = dev->dev_private;
4347         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4348         int ret;
4349
4350         ret = i915_gem_object_unbind(obj);
4351         if (ret == -ERESTARTSYS) {
4352                 list_move(&obj_priv->mm_list,
4353                           &dev_priv->mm.deferred_free_list);
4354                 return;
4355         }
4356
4357         if (obj_priv->mmap_offset)
4358                 i915_gem_free_mmap_offset(obj);
4359
4360         drm_gem_object_release(obj);
4361         i915_gem_info_remove_obj(dev_priv, obj->size);
4362
4363         kfree(obj_priv->page_cpu_valid);
4364         kfree(obj_priv->bit_17);
4365         kfree(obj_priv);
4366 }
4367
4368 void i915_gem_free_object(struct drm_gem_object *obj)
4369 {
4370         struct drm_device *dev = obj->dev;
4371         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4372
4373         trace_i915_gem_object_destroy(obj);
4374
4375         while (obj_priv->pin_count > 0)
4376                 i915_gem_object_unpin(obj);
4377
4378         if (obj_priv->phys_obj)
4379                 i915_gem_detach_phys_object(dev, obj);
4380
4381         i915_gem_free_object_tail(obj);
4382 }
4383
4384 int
4385 i915_gem_idle(struct drm_device *dev)
4386 {
4387         drm_i915_private_t *dev_priv = dev->dev_private;
4388         int ret;
4389
4390         mutex_lock(&dev->struct_mutex);
4391
4392         if (dev_priv->mm.suspended) {
4393                 mutex_unlock(&dev->struct_mutex);
4394                 return 0;
4395         }
4396
4397         ret = i915_gpu_idle(dev);
4398         if (ret) {
4399                 mutex_unlock(&dev->struct_mutex);
4400                 return ret;
4401         }
4402
4403         /* Under UMS, be paranoid and evict. */
4404         if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
4405                 ret = i915_gem_evict_inactive(dev);
4406                 if (ret) {
4407                         mutex_unlock(&dev->struct_mutex);
4408                         return ret;
4409                 }
4410         }
4411
4412         /* Hack!  Don't let anybody do execbuf while we don't control the chip.
4413          * We need to replace this with a semaphore, or something.
4414          * And not confound mm.suspended!
4415          */
4416         dev_priv->mm.suspended = 1;
4417         del_timer_sync(&dev_priv->hangcheck_timer);
4418
4419         i915_kernel_lost_context(dev);
4420         i915_gem_cleanup_ringbuffer(dev);
4421
4422         mutex_unlock(&dev->struct_mutex);
4423
4424         /* Cancel the retire work handler, which should be idle now. */
4425         cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4426
4427         return 0;
4428 }
4429
4430 /*
4431  * 965+ support PIPE_CONTROL commands, which provide finer grained control
4432  * over cache flushing.
4433  */
4434 static int
4435 i915_gem_init_pipe_control(struct drm_device *dev)
4436 {
4437         drm_i915_private_t *dev_priv = dev->dev_private;
4438         struct drm_gem_object *obj;
4439         struct drm_i915_gem_object *obj_priv;
4440         int ret;
4441
4442         obj = i915_gem_alloc_object(dev, 4096);
4443         if (obj == NULL) {
4444                 DRM_ERROR("Failed to allocate seqno page\n");
4445                 ret = -ENOMEM;
4446                 goto err;
4447         }
4448         obj_priv = to_intel_bo(obj);
4449         obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
4450
4451         ret = i915_gem_object_pin(obj, 4096);
4452         if (ret)
4453                 goto err_unref;
4454
4455         dev_priv->seqno_gfx_addr = obj_priv->gtt_offset;
4456         dev_priv->seqno_page =  kmap(obj_priv->pages[0]);
4457         if (dev_priv->seqno_page == NULL)
4458                 goto err_unpin;
4459
4460         dev_priv->seqno_obj = obj;
4461         memset(dev_priv->seqno_page, 0, PAGE_SIZE);
4462
4463         return 0;
4464
4465 err_unpin:
4466         i915_gem_object_unpin(obj);
4467 err_unref:
4468         drm_gem_object_unreference(obj);
4469 err:
4470         return ret;
4471 }
4472
4473
4474 static void
4475 i915_gem_cleanup_pipe_control(struct drm_device *dev)
4476 {
4477         drm_i915_private_t *dev_priv = dev->dev_private;
4478         struct drm_gem_object *obj;
4479         struct drm_i915_gem_object *obj_priv;
4480
4481         obj = dev_priv->seqno_obj;
4482         obj_priv = to_intel_bo(obj);
4483         kunmap(obj_priv->pages[0]);
4484         i915_gem_object_unpin(obj);
4485         drm_gem_object_unreference(obj);
4486         dev_priv->seqno_obj = NULL;
4487
4488         dev_priv->seqno_page = NULL;
4489 }
4490
4491 int
4492 i915_gem_init_ringbuffer(struct drm_device *dev)
4493 {
4494         drm_i915_private_t *dev_priv = dev->dev_private;
4495         int ret;
4496
4497         if (HAS_PIPE_CONTROL(dev)) {
4498                 ret = i915_gem_init_pipe_control(dev);
4499                 if (ret)
4500                         return ret;
4501         }
4502
4503         ret = intel_init_render_ring_buffer(dev);
4504         if (ret)
4505                 goto cleanup_pipe_control;
4506
4507         if (HAS_BSD(dev)) {
4508                 ret = intel_init_bsd_ring_buffer(dev);
4509                 if (ret)
4510                         goto cleanup_render_ring;
4511         }
4512
4513         if (HAS_BLT(dev)) {
4514                 ret = intel_init_blt_ring_buffer(dev);
4515                 if (ret)
4516                         goto cleanup_bsd_ring;
4517         }
4518
4519         dev_priv->next_seqno = 1;
4520
4521         return 0;
4522
4523 cleanup_bsd_ring:
4524         intel_cleanup_ring_buffer(dev, &dev_priv->bsd_ring);
4525 cleanup_render_ring:
4526         intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4527 cleanup_pipe_control:
4528         if (HAS_PIPE_CONTROL(dev))
4529                 i915_gem_cleanup_pipe_control(dev);
4530         return ret;
4531 }
4532
4533 void
4534 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4535 {
4536         drm_i915_private_t *dev_priv = dev->dev_private;
4537
4538         intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4539         intel_cleanup_ring_buffer(dev, &dev_priv->bsd_ring);
4540         intel_cleanup_ring_buffer(dev, &dev_priv->blt_ring);
4541         if (HAS_PIPE_CONTROL(dev))
4542                 i915_gem_cleanup_pipe_control(dev);
4543 }
4544
4545 int
4546 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4547                        struct drm_file *file_priv)
4548 {
4549         drm_i915_private_t *dev_priv = dev->dev_private;
4550         int ret;
4551
4552         if (drm_core_check_feature(dev, DRIVER_MODESET))
4553                 return 0;
4554
4555         if (atomic_read(&dev_priv->mm.wedged)) {
4556                 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4557                 atomic_set(&dev_priv->mm.wedged, 0);
4558         }
4559
4560         mutex_lock(&dev->struct_mutex);
4561         dev_priv->mm.suspended = 0;
4562
4563         ret = i915_gem_init_ringbuffer(dev);
4564         if (ret != 0) {
4565                 mutex_unlock(&dev->struct_mutex);
4566                 return ret;
4567         }
4568
4569         BUG_ON(!list_empty(&dev_priv->mm.active_list));
4570         BUG_ON(!list_empty(&dev_priv->render_ring.active_list));
4571         BUG_ON(!list_empty(&dev_priv->bsd_ring.active_list));
4572         BUG_ON(!list_empty(&dev_priv->blt_ring.active_list));
4573         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
4574         BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
4575         BUG_ON(!list_empty(&dev_priv->render_ring.request_list));
4576         BUG_ON(!list_empty(&dev_priv->bsd_ring.request_list));
4577         BUG_ON(!list_empty(&dev_priv->blt_ring.request_list));
4578         mutex_unlock(&dev->struct_mutex);
4579
4580         ret = drm_irq_install(dev);
4581         if (ret)
4582                 goto cleanup_ringbuffer;
4583
4584         return 0;
4585
4586 cleanup_ringbuffer:
4587         mutex_lock(&dev->struct_mutex);
4588         i915_gem_cleanup_ringbuffer(dev);
4589         dev_priv->mm.suspended = 1;
4590         mutex_unlock(&dev->struct_mutex);
4591
4592         return ret;
4593 }
4594
4595 int
4596 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4597                        struct drm_file *file_priv)
4598 {
4599         if (drm_core_check_feature(dev, DRIVER_MODESET))
4600                 return 0;
4601
4602         drm_irq_uninstall(dev);
4603         return i915_gem_idle(dev);
4604 }
4605
4606 void
4607 i915_gem_lastclose(struct drm_device *dev)
4608 {
4609         int ret;
4610
4611         if (drm_core_check_feature(dev, DRIVER_MODESET))
4612                 return;
4613
4614         ret = i915_gem_idle(dev);
4615         if (ret)
4616                 DRM_ERROR("failed to idle hardware: %d\n", ret);
4617 }
4618
4619 static void
4620 init_ring_lists(struct intel_ring_buffer *ring)
4621 {
4622         INIT_LIST_HEAD(&ring->active_list);
4623         INIT_LIST_HEAD(&ring->request_list);
4624         INIT_LIST_HEAD(&ring->gpu_write_list);
4625 }
4626
4627 void
4628 i915_gem_load(struct drm_device *dev)
4629 {
4630         int i;
4631         drm_i915_private_t *dev_priv = dev->dev_private;
4632
4633         INIT_LIST_HEAD(&dev_priv->mm.active_list);
4634         INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
4635         INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4636         INIT_LIST_HEAD(&dev_priv->mm.pinned_list);
4637         INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4638         INIT_LIST_HEAD(&dev_priv->mm.deferred_free_list);
4639         init_ring_lists(&dev_priv->render_ring);
4640         init_ring_lists(&dev_priv->bsd_ring);
4641         init_ring_lists(&dev_priv->blt_ring);
4642         for (i = 0; i < 16; i++)
4643                 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4644         INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4645                           i915_gem_retire_work_handler);
4646         init_completion(&dev_priv->error_completion);
4647         spin_lock(&shrink_list_lock);
4648         list_add(&dev_priv->mm.shrink_list, &shrink_list);
4649         spin_unlock(&shrink_list_lock);
4650
4651         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4652         if (IS_GEN3(dev)) {
4653                 u32 tmp = I915_READ(MI_ARB_STATE);
4654                 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
4655                         /* arb state is a masked write, so set bit + bit in mask */
4656                         tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
4657                         I915_WRITE(MI_ARB_STATE, tmp);
4658                 }
4659         }
4660
4661         /* Old X drivers will take 0-2 for front, back, depth buffers */
4662         if (!drm_core_check_feature(dev, DRIVER_MODESET))
4663                 dev_priv->fence_reg_start = 3;
4664
4665         if (INTEL_INFO(dev)->gen >= 4 || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4666                 dev_priv->num_fence_regs = 16;
4667         else
4668                 dev_priv->num_fence_regs = 8;
4669
4670         /* Initialize fence registers to zero */
4671         switch (INTEL_INFO(dev)->gen) {
4672         case 6:
4673                 for (i = 0; i < 16; i++)
4674                         I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + (i * 8), 0);
4675                 break;
4676         case 5:
4677         case 4:
4678                 for (i = 0; i < 16; i++)
4679                         I915_WRITE64(FENCE_REG_965_0 + (i * 8), 0);
4680                 break;
4681         case 3:
4682                 if (IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4683                         for (i = 0; i < 8; i++)
4684                                 I915_WRITE(FENCE_REG_945_8 + (i * 4), 0);
4685         case 2:
4686                 for (i = 0; i < 8; i++)
4687                         I915_WRITE(FENCE_REG_830_0 + (i * 4), 0);
4688                 break;
4689         }
4690         i915_gem_detect_bit_6_swizzle(dev);
4691         init_waitqueue_head(&dev_priv->pending_flip_queue);
4692 }
4693
4694 /*
4695  * Create a physically contiguous memory object for this object
4696  * e.g. for cursor + overlay regs
4697  */
4698 static int i915_gem_init_phys_object(struct drm_device *dev,
4699                                      int id, int size, int align)
4700 {
4701         drm_i915_private_t *dev_priv = dev->dev_private;
4702         struct drm_i915_gem_phys_object *phys_obj;
4703         int ret;
4704
4705         if (dev_priv->mm.phys_objs[id - 1] || !size)
4706                 return 0;
4707
4708         phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4709         if (!phys_obj)
4710                 return -ENOMEM;
4711
4712         phys_obj->id = id;
4713
4714         phys_obj->handle = drm_pci_alloc(dev, size, align);
4715         if (!phys_obj->handle) {
4716                 ret = -ENOMEM;
4717                 goto kfree_obj;
4718         }
4719 #ifdef CONFIG_X86
4720         set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4721 #endif
4722
4723         dev_priv->mm.phys_objs[id - 1] = phys_obj;
4724
4725         return 0;
4726 kfree_obj:
4727         kfree(phys_obj);
4728         return ret;
4729 }
4730
4731 static void i915_gem_free_phys_object(struct drm_device *dev, int id)
4732 {
4733         drm_i915_private_t *dev_priv = dev->dev_private;
4734         struct drm_i915_gem_phys_object *phys_obj;
4735
4736         if (!dev_priv->mm.phys_objs[id - 1])
4737                 return;
4738
4739         phys_obj = dev_priv->mm.phys_objs[id - 1];
4740         if (phys_obj->cur_obj) {
4741                 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4742         }
4743
4744 #ifdef CONFIG_X86
4745         set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4746 #endif
4747         drm_pci_free(dev, phys_obj->handle);
4748         kfree(phys_obj);
4749         dev_priv->mm.phys_objs[id - 1] = NULL;
4750 }
4751
4752 void i915_gem_free_all_phys_object(struct drm_device *dev)
4753 {
4754         int i;
4755
4756         for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4757                 i915_gem_free_phys_object(dev, i);
4758 }
4759
4760 void i915_gem_detach_phys_object(struct drm_device *dev,
4761                                  struct drm_gem_object *obj)
4762 {
4763         struct drm_i915_gem_object *obj_priv;
4764         int i;
4765         int ret;
4766         int page_count;
4767
4768         obj_priv = to_intel_bo(obj);
4769         if (!obj_priv->phys_obj)
4770                 return;
4771
4772         ret = i915_gem_object_get_pages(obj, 0);
4773         if (ret)
4774                 goto out;
4775
4776         page_count = obj->size / PAGE_SIZE;
4777
4778         for (i = 0; i < page_count; i++) {
4779                 char *dst = kmap_atomic(obj_priv->pages[i], KM_USER0);
4780                 char *src = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4781
4782                 memcpy(dst, src, PAGE_SIZE);
4783                 kunmap_atomic(dst, KM_USER0);
4784         }
4785         drm_clflush_pages(obj_priv->pages, page_count);
4786         drm_agp_chipset_flush(dev);
4787
4788         i915_gem_object_put_pages(obj);
4789 out:
4790         obj_priv->phys_obj->cur_obj = NULL;
4791         obj_priv->phys_obj = NULL;
4792 }
4793
4794 int
4795 i915_gem_attach_phys_object(struct drm_device *dev,
4796                             struct drm_gem_object *obj,
4797                             int id,
4798                             int align)
4799 {
4800         drm_i915_private_t *dev_priv = dev->dev_private;
4801         struct drm_i915_gem_object *obj_priv;
4802         int ret = 0;
4803         int page_count;
4804         int i;
4805
4806         if (id > I915_MAX_PHYS_OBJECT)
4807                 return -EINVAL;
4808
4809         obj_priv = to_intel_bo(obj);
4810
4811         if (obj_priv->phys_obj) {
4812                 if (obj_priv->phys_obj->id == id)
4813                         return 0;
4814                 i915_gem_detach_phys_object(dev, obj);
4815         }
4816
4817         /* create a new object */
4818         if (!dev_priv->mm.phys_objs[id - 1]) {
4819                 ret = i915_gem_init_phys_object(dev, id,
4820                                                 obj->size, align);
4821                 if (ret) {
4822                         DRM_ERROR("failed to init phys object %d size: %zu\n", id, obj->size);
4823                         goto out;
4824                 }
4825         }
4826
4827         /* bind to the object */
4828         obj_priv->phys_obj = dev_priv->mm.phys_objs[id - 1];
4829         obj_priv->phys_obj->cur_obj = obj;
4830
4831         ret = i915_gem_object_get_pages(obj, 0);
4832         if (ret) {
4833                 DRM_ERROR("failed to get page list\n");
4834                 goto out;
4835         }
4836
4837         page_count = obj->size / PAGE_SIZE;
4838
4839         for (i = 0; i < page_count; i++) {
4840                 char *src = kmap_atomic(obj_priv->pages[i], KM_USER0);
4841                 char *dst = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4842
4843                 memcpy(dst, src, PAGE_SIZE);
4844                 kunmap_atomic(src, KM_USER0);
4845         }
4846
4847         i915_gem_object_put_pages(obj);
4848
4849         return 0;
4850 out:
4851         return ret;
4852 }
4853
4854 static int
4855 i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
4856                      struct drm_i915_gem_pwrite *args,
4857                      struct drm_file *file_priv)
4858 {
4859         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4860         void *obj_addr;
4861         int ret;
4862         char __user *user_data;
4863
4864         user_data = (char __user *) (uintptr_t) args->data_ptr;
4865         obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
4866
4867         DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
4868         ret = copy_from_user(obj_addr, user_data, args->size);
4869         if (ret)
4870                 return -EFAULT;
4871
4872         drm_agp_chipset_flush(dev);
4873         return 0;
4874 }
4875
4876 void i915_gem_release(struct drm_device *dev, struct drm_file *file)
4877 {
4878         struct drm_i915_file_private *file_priv = file->driver_priv;
4879
4880         /* Clean up our request list when the client is going away, so that
4881          * later retire_requests won't dereference our soon-to-be-gone
4882          * file_priv.
4883          */
4884         spin_lock(&file_priv->mm.lock);
4885         while (!list_empty(&file_priv->mm.request_list)) {
4886                 struct drm_i915_gem_request *request;
4887
4888                 request = list_first_entry(&file_priv->mm.request_list,
4889                                            struct drm_i915_gem_request,
4890                                            client_list);
4891                 list_del(&request->client_list);
4892                 request->file_priv = NULL;
4893         }
4894         spin_unlock(&file_priv->mm.lock);
4895 }
4896
4897 static int
4898 i915_gpu_is_active(struct drm_device *dev)
4899 {
4900         drm_i915_private_t *dev_priv = dev->dev_private;
4901         int lists_empty;
4902
4903         lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4904                       list_empty(&dev_priv->render_ring.active_list) &&
4905                       list_empty(&dev_priv->bsd_ring.active_list) &&
4906                       list_empty(&dev_priv->blt_ring.active_list);
4907
4908         return !lists_empty;
4909 }
4910
4911 static int
4912 i915_gem_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
4913 {
4914         drm_i915_private_t *dev_priv, *next_dev;
4915         struct drm_i915_gem_object *obj_priv, *next_obj;
4916         int cnt = 0;
4917         int would_deadlock = 1;
4918
4919         /* "fast-path" to count number of available objects */
4920         if (nr_to_scan == 0) {
4921                 spin_lock(&shrink_list_lock);
4922                 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
4923                         struct drm_device *dev = dev_priv->dev;
4924
4925                         if (mutex_trylock(&dev->struct_mutex)) {
4926                                 list_for_each_entry(obj_priv,
4927                                                     &dev_priv->mm.inactive_list,
4928                                                     mm_list)
4929                                         cnt++;
4930                                 mutex_unlock(&dev->struct_mutex);
4931                         }
4932                 }
4933                 spin_unlock(&shrink_list_lock);
4934
4935                 return (cnt / 100) * sysctl_vfs_cache_pressure;
4936         }
4937
4938         spin_lock(&shrink_list_lock);
4939
4940 rescan:
4941         /* first scan for clean buffers */
4942         list_for_each_entry_safe(dev_priv, next_dev,
4943                                  &shrink_list, mm.shrink_list) {
4944                 struct drm_device *dev = dev_priv->dev;
4945
4946                 if (! mutex_trylock(&dev->struct_mutex))
4947                         continue;
4948
4949                 spin_unlock(&shrink_list_lock);
4950                 i915_gem_retire_requests(dev);
4951
4952                 list_for_each_entry_safe(obj_priv, next_obj,
4953                                          &dev_priv->mm.inactive_list,
4954                                          mm_list) {
4955                         if (i915_gem_object_is_purgeable(obj_priv)) {
4956                                 i915_gem_object_unbind(&obj_priv->base);
4957                                 if (--nr_to_scan <= 0)
4958                                         break;
4959                         }
4960                 }
4961
4962                 spin_lock(&shrink_list_lock);
4963                 mutex_unlock(&dev->struct_mutex);
4964
4965                 would_deadlock = 0;
4966
4967                 if (nr_to_scan <= 0)
4968                         break;
4969         }
4970
4971         /* second pass, evict/count anything still on the inactive list */
4972         list_for_each_entry_safe(dev_priv, next_dev,
4973                                  &shrink_list, mm.shrink_list) {
4974                 struct drm_device *dev = dev_priv->dev;
4975
4976                 if (! mutex_trylock(&dev->struct_mutex))
4977                         continue;
4978
4979                 spin_unlock(&shrink_list_lock);
4980
4981                 list_for_each_entry_safe(obj_priv, next_obj,
4982                                          &dev_priv->mm.inactive_list,
4983                                          mm_list) {
4984                         if (nr_to_scan > 0) {
4985                                 i915_gem_object_unbind(&obj_priv->base);
4986                                 nr_to_scan--;
4987                         } else
4988                                 cnt++;
4989                 }
4990
4991                 spin_lock(&shrink_list_lock);
4992                 mutex_unlock(&dev->struct_mutex);
4993
4994                 would_deadlock = 0;
4995         }
4996
4997         if (nr_to_scan) {
4998                 int active = 0;
4999
5000                 /*
5001                  * We are desperate for pages, so as a last resort, wait
5002                  * for the GPU to finish and discard whatever we can.
5003                  * This has a dramatic impact to reduce the number of
5004                  * OOM-killer events whilst running the GPU aggressively.
5005                  */
5006                 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5007                         struct drm_device *dev = dev_priv->dev;
5008
5009                         if (!mutex_trylock(&dev->struct_mutex))
5010                                 continue;
5011
5012                         spin_unlock(&shrink_list_lock);
5013
5014                         if (i915_gpu_is_active(dev)) {
5015                                 i915_gpu_idle(dev);
5016                                 active++;
5017                         }
5018
5019                         spin_lock(&shrink_list_lock);
5020                         mutex_unlock(&dev->struct_mutex);
5021                 }
5022
5023                 if (active)
5024                         goto rescan;
5025         }
5026
5027         spin_unlock(&shrink_list_lock);
5028
5029         if (would_deadlock)
5030                 return -1;
5031         else if (cnt > 0)
5032                 return (cnt / 100) * sysctl_vfs_cache_pressure;
5033         else
5034                 return 0;
5035 }
5036
5037 static struct shrinker shrinker = {
5038         .shrink = i915_gem_shrink,
5039         .seeks = DEFAULT_SEEKS,
5040 };
5041
5042 __init void
5043 i915_gem_shrinker_init(void)
5044 {
5045     register_shrinker(&shrinker);
5046 }
5047
5048 __exit void
5049 i915_gem_shrinker_exit(void)
5050 {
5051     unregister_shrinker(&shrinker);
5052 }