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/shmem_fs.h>
  35 #include <linux/slab.h>
  36 #include <linux/swap.h>
  37 #include <linux/pci.h>
  38
  39 static __must_check int i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj);
  40 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
  41 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
  42 static __must_check int i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj,
  43                                                           bool write);
  44 static __must_check int i915_gem_object_set_cpu_read_domain_range(struct drm_i915_gem_object *obj,
  45                                                                   uint64_t offset,
  46                                                                   uint64_t size);
  47 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_i915_gem_object *obj);
  48 static __must_check int i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
  49                                                     unsigned alignment,
  50                                                     bool map_and_fenceable);
  51 static void i915_gem_clear_fence_reg(struct drm_device *dev,
  52                                      struct drm_i915_fence_reg *reg);
  53 static int i915_gem_phys_pwrite(struct drm_device *dev,
  54                                 struct drm_i915_gem_object *obj,
  55                                 struct drm_i915_gem_pwrite *args,
  56                                 struct drm_file *file);
  57 static void i915_gem_free_object_tail(struct drm_i915_gem_object *obj);
  58
  59 static int i915_gem_inactive_shrink(struct shrinker *shrinker,
  60                                     struct shrink_control *sc);
  61
  62 /* some bookkeeping */
  63 static void i915_gem_info_add_obj(struct drm_i915_private *dev_priv,
  64                                   size_t size)
  65 {
  66         dev_priv->mm.object_count++;
  67         dev_priv->mm.object_memory += size;
  68 }
  69
  70 static void i915_gem_info_remove_obj(struct drm_i915_private *dev_priv,
  71                                      size_t size)
  72 {
  73         dev_priv->mm.object_count--;
  74         dev_priv->mm.object_memory -= size;
  75 }
  76
  77 static int
  78 i915_gem_wait_for_error(struct drm_device *dev)
  79 {
  80         struct drm_i915_private *dev_priv = dev->dev_private;
  81         struct completion *x = &dev_priv->error_completion;
  82         unsigned long flags;
  83         int ret;
  84
  85         if (!atomic_read(&dev_priv->mm.wedged))
  86                 return 0;
  87
  88         ret = wait_for_completion_interruptible(x);
  89         if (ret)
  90                 return ret;
  91
  92         if (atomic_read(&dev_priv->mm.wedged)) {
  93                 /* GPU is hung, bump the completion count to account for
  94                  * the token we just consumed so that we never hit zero and
  95                  * end up waiting upon a subsequent completion event that
  96                  * will never happen.
  97                  */
  98                 spin_lock_irqsave(&x->wait.lock, flags);
  99                 x->done++;
 100                 spin_unlock_irqrestore(&x->wait.lock, flags);
 101         }
 102         return 0;
 103 }
 104
 105 int i915_mutex_lock_interruptible(struct drm_device *dev)
 106 {
 107         int ret;
 108
 109         ret = i915_gem_wait_for_error(dev);
 110         if (ret)
 111                 return ret;
 112
 113         ret = mutex_lock_interruptible(&dev->struct_mutex);
 114         if (ret)
 115                 return ret;
 116
 117         WARN_ON(i915_verify_lists(dev));
 118         return 0;
 119 }
 120
 121 static inline bool
 122 i915_gem_object_is_inactive(struct drm_i915_gem_object *obj)
 123 {
 124         return obj->gtt_space && !obj->active && obj->pin_count == 0;
 125 }
 126
 127 void i915_gem_do_init(struct drm_device *dev,
 128                       unsigned long start,
 129                       unsigned long mappable_end,
 130                       unsigned long end)
 131 {
 132         drm_i915_private_t *dev_priv = dev->dev_private;
 133
 134         drm_mm_init(&dev_priv->mm.gtt_space, start, end - start);
 135
 136         dev_priv->mm.gtt_start = start;
 137         dev_priv->mm.gtt_mappable_end = mappable_end;
 138         dev_priv->mm.gtt_end = end;
 139         dev_priv->mm.gtt_total = end - start;
 140         dev_priv->mm.mappable_gtt_total = min(end, mappable_end) - start;
 141
 142         /* Take over this portion of the GTT */
 143         intel_gtt_clear_range(start / PAGE_SIZE, (end-start) / PAGE_SIZE);
 144 }
 145
 146 int
 147 i915_gem_init_ioctl(struct drm_device *dev, void *data,
 148                     struct drm_file *file)
 149 {
 150         struct drm_i915_gem_init *args = data;
 151
 152         if (args->gtt_start >= args->gtt_end ||
 153             (args->gtt_end | args->gtt_start) & (PAGE_SIZE - 1))
 154                 return -EINVAL;
 155
 156         mutex_lock(&dev->struct_mutex);
 157         i915_gem_do_init(dev, args->gtt_start, args->gtt_end, args->gtt_end);
 158         mutex_unlock(&dev->struct_mutex);
 159
 160         return 0;
 161 }
 162
 163 int
 164 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
 165                             struct drm_file *file)
 166 {
 167         struct drm_i915_private *dev_priv = dev->dev_private;
 168         struct drm_i915_gem_get_aperture *args = data;
 169         struct drm_i915_gem_object *obj;
 170         size_t pinned;
 171
 172         if (!(dev->driver->driver_features & DRIVER_GEM))
 173                 return -ENODEV;
 174
 175         pinned = 0;
 176         mutex_lock(&dev->struct_mutex);
 177         list_for_each_entry(obj, &dev_priv->mm.pinned_list, mm_list)
 178                 pinned += obj->gtt_space->size;
 179         mutex_unlock(&dev->struct_mutex);
 180
 181         args->aper_size = dev_priv->mm.gtt_total;
 182         args->aper_available_size = args->aper_size - pinned;
 183
 184         return 0;
 185 }
 186
 187 static int
 188 i915_gem_create(struct drm_file *file,
 189                 struct drm_device *dev,
 190                 uint64_t size,
 191                 uint32_t *handle_p)
 192 {
 193         struct drm_i915_gem_object *obj;
 194         int ret;
 195         u32 handle;
 196
 197         size = roundup(size, PAGE_SIZE);
 198         if (size == 0)
 199                 return -EINVAL;
 200
 201         /* Allocate the new object */
 202         obj = i915_gem_alloc_object(dev, size);
 203         if (obj == NULL)
 204                 return -ENOMEM;
 205
 206         ret = drm_gem_handle_create(file, &obj->base, &handle);
 207         if (ret) {
 208                 drm_gem_object_release(&obj->base);
 209                 i915_gem_info_remove_obj(dev->dev_private, obj->base.size);
 210                 kfree(obj);
 211                 return ret;
 212         }
 213
 214         /* drop reference from allocate - handle holds it now */
 215         drm_gem_object_unreference(&obj->base);
 216         trace_i915_gem_object_create(obj);
 217
 218         *handle_p = handle;
 219         return 0;
 220 }
 221
 222 int
 223 i915_gem_dumb_create(struct drm_file *file,
 224                      struct drm_device *dev,
 225                      struct drm_mode_create_dumb *args)
 226 {
 227         /* have to work out size/pitch and return them */
 228         args->pitch = ALIGN(args->width * ((args->bpp + 7) / 8), 64);
 229         args->size = args->pitch * args->height;
 230         return i915_gem_create(file, dev,
 231                                args->size, &args->handle);
 232 }
 233
 234 int i915_gem_dumb_destroy(struct drm_file *file,
 235                           struct drm_device *dev,
 236                           uint32_t handle)
 237 {
 238         return drm_gem_handle_delete(file, handle);
 239 }
 240
 241 /**
 242  * Creates a new mm object and returns a handle to it.
 243  */
 244 int
 245 i915_gem_create_ioctl(struct drm_device *dev, void *data,
 246                       struct drm_file *file)
 247 {
 248         struct drm_i915_gem_create *args = data;
 249         return i915_gem_create(file, dev,
 250                                args->size, &args->handle);
 251 }
 252
 253 static int i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
 254 {
 255         drm_i915_private_t *dev_priv = obj->base.dev->dev_private;
 256
 257         return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
 258                 obj->tiling_mode != I915_TILING_NONE;
 259 }
 260
 261 static inline void
 262 slow_shmem_copy(struct page *dst_page,
 263                 int dst_offset,
 264                 struct page *src_page,
 265                 int src_offset,
 266                 int length)
 267 {
 268         char *dst_vaddr, *src_vaddr;
 269
 270         dst_vaddr = kmap(dst_page);
 271         src_vaddr = kmap(src_page);
 272
 273         memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
 274
 275         kunmap(src_page);
 276         kunmap(dst_page);
 277 }
 278
 279 static inline void
 280 slow_shmem_bit17_copy(struct page *gpu_page,
 281                       int gpu_offset,
 282                       struct page *cpu_page,
 283                       int cpu_offset,
 284                       int length,
 285                       int is_read)
 286 {
 287         char *gpu_vaddr, *cpu_vaddr;
 288
 289         /* Use the unswizzled path if this page isn't affected. */
 290         if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
 291                 if (is_read)
 292                         return slow_shmem_copy(cpu_page, cpu_offset,
 293                                                gpu_page, gpu_offset, length);
 294                 else
 295                         return slow_shmem_copy(gpu_page, gpu_offset,
 296                                                cpu_page, cpu_offset, length);
 297         }
 298
 299         gpu_vaddr = kmap(gpu_page);
 300         cpu_vaddr = kmap(cpu_page);
 301
 302         /* Copy the data, XORing A6 with A17 (1). The user already knows he's
 303          * XORing with the other bits (A9 for Y, A9 and A10 for X)
 304          */
 305         while (length > 0) {
 306                 int cacheline_end = ALIGN(gpu_offset + 1, 64);
 307                 int this_length = min(cacheline_end - gpu_offset, length);
 308                 int swizzled_gpu_offset = gpu_offset ^ 64;
 309
 310                 if (is_read) {
 311                         memcpy(cpu_vaddr + cpu_offset,
 312                                gpu_vaddr + swizzled_gpu_offset,
 313                                this_length);
 314                 } else {
 315                         memcpy(gpu_vaddr + swizzled_gpu_offset,
 316                                cpu_vaddr + cpu_offset,
 317                                this_length);
 318                 }
 319                 cpu_offset += this_length;
 320                 gpu_offset += this_length;
 321                 length -= this_length;
 322         }
 323
 324         kunmap(cpu_page);
 325         kunmap(gpu_page);
 326 }
 327
 328 /**
 329  * This is the fast shmem pread path, which attempts to copy_from_user directly
 330  * from the backing pages of the object to the user's address space.  On a
 331  * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
 332  */
 333 static int
 334 i915_gem_shmem_pread_fast(struct drm_device *dev,
 335                           struct drm_i915_gem_object *obj,
 336                           struct drm_i915_gem_pread *args,
 337                           struct drm_file *file)
 338 {
 339         struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
 340         ssize_t remain;
 341         loff_t offset;
 342         char __user *user_data;
 343         int page_offset, page_length;
 344
 345         user_data = (char __user *) (uintptr_t) args->data_ptr;
 346         remain = args->size;
 347
 348         offset = args->offset;
 349
 350         while (remain > 0) {
 351                 struct page *page;
 352                 char *vaddr;
 353                 int ret;
 354
 355                 /* Operation in this page
 356                  *
 357                  * page_offset = offset within page
 358                  * page_length = bytes to copy for this page
 359                  */
 360                 page_offset = offset_in_page(offset);
 361                 page_length = remain;
 362                 if ((page_offset + remain) > PAGE_SIZE)
 363                         page_length = PAGE_SIZE - page_offset;
 364
 365                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 366                 if (IS_ERR(page))
 367                         return PTR_ERR(page);
 368
 369                 vaddr = kmap_atomic(page);
 370                 ret = __copy_to_user_inatomic(user_data,
 371                                               vaddr + page_offset,
 372                                               page_length);
 373                 kunmap_atomic(vaddr);
 374
 375                 mark_page_accessed(page);
 376                 page_cache_release(page);
 377                 if (ret)
 378                         return -EFAULT;
 379
 380                 remain -= page_length;
 381                 user_data += page_length;
 382                 offset += page_length;
 383         }
 384
 385         return 0;
 386 }
 387
 388 /**
 389  * This is the fallback shmem pread path, which allocates temporary storage
 390  * in kernel space to copy_to_user into outside of the struct_mutex, so we
 391  * can copy out of the object's backing pages while holding the struct mutex
 392  * and not take page faults.
 393  */
 394 static int
 395 i915_gem_shmem_pread_slow(struct drm_device *dev,
 396                           struct drm_i915_gem_object *obj,
 397                           struct drm_i915_gem_pread *args,
 398                           struct drm_file *file)
 399 {
 400         struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
 401         struct mm_struct *mm = current->mm;
 402         struct page **user_pages;
 403         ssize_t remain;
 404         loff_t offset, pinned_pages, i;
 405         loff_t first_data_page, last_data_page, num_pages;
 406         int shmem_page_offset;
 407         int data_page_index, data_page_offset;
 408         int page_length;
 409         int ret;
 410         uint64_t data_ptr = args->data_ptr;
 411         int do_bit17_swizzling;
 412
 413         remain = args->size;
 414
 415         /* Pin the user pages containing the data.  We can't fault while
 416          * holding the struct mutex, yet we want to hold it while
 417          * dereferencing the user data.
 418          */
 419         first_data_page = data_ptr / PAGE_SIZE;
 420         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 421         num_pages = last_data_page - first_data_page + 1;
 422
 423         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 424         if (user_pages == NULL)
 425                 return -ENOMEM;
 426
 427         mutex_unlock(&dev->struct_mutex);
 428         down_read(&mm->mmap_sem);
 429         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 430                                       num_pages, 1, 0, user_pages, NULL);
 431         up_read(&mm->mmap_sem);
 432         mutex_lock(&dev->struct_mutex);
 433         if (pinned_pages < num_pages) {
 434                 ret = -EFAULT;
 435                 goto out;
 436         }
 437
 438         ret = i915_gem_object_set_cpu_read_domain_range(obj,
 439                                                         args->offset,
 440                                                         args->size);
 441         if (ret)
 442                 goto out;
 443
 444         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 445
 446         offset = args->offset;
 447
 448         while (remain > 0) {
 449                 struct page *page;
 450
 451                 /* Operation in this page
 452                  *
 453                  * shmem_page_offset = offset within page in shmem file
 454                  * data_page_index = page number in get_user_pages return
 455                  * data_page_offset = offset with data_page_index page.
 456                  * page_length = bytes to copy for this page
 457                  */
 458                 shmem_page_offset = offset_in_page(offset);
 459                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 460                 data_page_offset = offset_in_page(data_ptr);
 461
 462                 page_length = remain;
 463                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 464                         page_length = PAGE_SIZE - shmem_page_offset;
 465                 if ((data_page_offset + page_length) > PAGE_SIZE)
 466                         page_length = PAGE_SIZE - data_page_offset;
 467
 468                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 469                 if (IS_ERR(page)) {
 470                         ret = PTR_ERR(page);
 471                         goto out;
 472                 }
 473
 474                 if (do_bit17_swizzling) {
 475                         slow_shmem_bit17_copy(page,
 476                                               shmem_page_offset,
 477                                               user_pages[data_page_index],
 478                                               data_page_offset,
 479                                               page_length,
 480                                               1);
 481                 } else {
 482                         slow_shmem_copy(user_pages[data_page_index],
 483                                         data_page_offset,
 484                                         page,
 485                                         shmem_page_offset,
 486                                         page_length);
 487                 }
 488
 489                 mark_page_accessed(page);
 490                 page_cache_release(page);
 491
 492                 remain -= page_length;
 493                 data_ptr += page_length;
 494                 offset += page_length;
 495         }
 496
 497 out:
 498         for (i = 0; i < pinned_pages; i++) {
 499                 SetPageDirty(user_pages[i]);
 500                 mark_page_accessed(user_pages[i]);
 501                 page_cache_release(user_pages[i]);
 502         }
 503         drm_free_large(user_pages);
 504
 505         return ret;
 506 }
 507
 508 /**
 509  * Reads data from the object referenced by handle.
 510  *
 511  * On error, the contents of *data are undefined.
 512  */
 513 int
 514 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 515                      struct drm_file *file)
 516 {
 517         struct drm_i915_gem_pread *args = data;
 518         struct drm_i915_gem_object *obj;
 519         int ret = 0;
 520
 521         if (args->size == 0)
 522                 return 0;
 523
 524         if (!access_ok(VERIFY_WRITE,
 525                        (char __user *)(uintptr_t)args->data_ptr,
 526                        args->size))
 527                 return -EFAULT;
 528
 529         ret = fault_in_pages_writeable((char __user *)(uintptr_t)args->data_ptr,
 530                                        args->size);
 531         if (ret)
 532                 return -EFAULT;
 533
 534         ret = i915_mutex_lock_interruptible(dev);
 535         if (ret)
 536                 return ret;
 537
 538         obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
 539         if (&obj->base == NULL) {
 540                 ret = -ENOENT;
 541                 goto unlock;
 542         }
 543
 544         /* Bounds check source.  */
 545         if (args->offset > obj->base.size ||
 546             args->size > obj->base.size - args->offset) {
 547                 ret = -EINVAL;
 548                 goto out;
 549         }
 550
 551         trace_i915_gem_object_pread(obj, args->offset, args->size);
 552
 553         ret = i915_gem_object_set_cpu_read_domain_range(obj,
 554                                                         args->offset,
 555                                                         args->size);
 556         if (ret)
 557                 goto out;
 558
 559         ret = -EFAULT;
 560         if (!i915_gem_object_needs_bit17_swizzle(obj))
 561                 ret = i915_gem_shmem_pread_fast(dev, obj, args, file);
 562         if (ret == -EFAULT)
 563                 ret = i915_gem_shmem_pread_slow(dev, obj, args, file);
 564
 565 out:
 566         drm_gem_object_unreference(&obj->base);
 567 unlock:
 568         mutex_unlock(&dev->struct_mutex);
 569         return ret;
 570 }
 571
 572 /* This is the fast write path which cannot handle
 573  * page faults in the source data
 574  */
 575
 576 static inline int
 577 fast_user_write(struct io_mapping *mapping,
 578                 loff_t page_base, int page_offset,
 579                 char __user *user_data,
 580                 int length)
 581 {
 582         char *vaddr_atomic;
 583         unsigned long unwritten;
 584
 585         vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
 586         unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
 587                                                       user_data, length);
 588         io_mapping_unmap_atomic(vaddr_atomic);
 589         return unwritten;
 590 }
 591
 592 /* Here's the write path which can sleep for
 593  * page faults
 594  */
 595
 596 static inline void
 597 slow_kernel_write(struct io_mapping *mapping,
 598                   loff_t gtt_base, int gtt_offset,
 599                   struct page *user_page, int user_offset,
 600                   int length)
 601 {
 602         char __iomem *dst_vaddr;
 603         char *src_vaddr;
 604
 605         dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
 606         src_vaddr = kmap(user_page);
 607
 608         memcpy_toio(dst_vaddr + gtt_offset,
 609                     src_vaddr + user_offset,
 610                     length);
 611
 612         kunmap(user_page);
 613         io_mapping_unmap(dst_vaddr);
 614 }
 615
 616 /**
 617  * This is the fast pwrite path, where we copy the data directly from the
 618  * user into the GTT, uncached.
 619  */
 620 static int
 621 i915_gem_gtt_pwrite_fast(struct drm_device *dev,
 622                          struct drm_i915_gem_object *obj,
 623                          struct drm_i915_gem_pwrite *args,
 624                          struct drm_file *file)
 625 {
 626         drm_i915_private_t *dev_priv = dev->dev_private;
 627         ssize_t remain;
 628         loff_t offset, page_base;
 629         char __user *user_data;
 630         int page_offset, page_length;
 631
 632         user_data = (char __user *) (uintptr_t) args->data_ptr;
 633         remain = args->size;
 634
 635         offset = obj->gtt_offset + args->offset;
 636
 637         while (remain > 0) {
 638                 /* Operation in this page
 639                  *
 640                  * page_base = page offset within aperture
 641                  * page_offset = offset within page
 642                  * page_length = bytes to copy for this page
 643                  */
 644                 page_base = offset & PAGE_MASK;
 645                 page_offset = offset_in_page(offset);
 646                 page_length = remain;
 647                 if ((page_offset + remain) > PAGE_SIZE)
 648                         page_length = PAGE_SIZE - page_offset;
 649
 650                 /* If we get a fault while copying data, then (presumably) our
 651                  * source page isn't available.  Return the error and we'll
 652                  * retry in the slow path.
 653                  */
 654                 if (fast_user_write(dev_priv->mm.gtt_mapping, page_base,
 655                                     page_offset, user_data, page_length))
 656                         return -EFAULT;
 657
 658                 remain -= page_length;
 659                 user_data += page_length;
 660                 offset += page_length;
 661         }
 662
 663         return 0;
 664 }
 665
 666 /**
 667  * This is the fallback GTT pwrite path, which uses get_user_pages to pin
 668  * the memory and maps it using kmap_atomic for copying.
 669  *
 670  * This code resulted in x11perf -rgb10text consuming about 10% more CPU
 671  * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
 672  */
 673 static int
 674 i915_gem_gtt_pwrite_slow(struct drm_device *dev,
 675                          struct drm_i915_gem_object *obj,
 676                          struct drm_i915_gem_pwrite *args,
 677                          struct drm_file *file)
 678 {
 679         drm_i915_private_t *dev_priv = dev->dev_private;
 680         ssize_t remain;
 681         loff_t gtt_page_base, offset;
 682         loff_t first_data_page, last_data_page, num_pages;
 683         loff_t pinned_pages, i;
 684         struct page **user_pages;
 685         struct mm_struct *mm = current->mm;
 686         int gtt_page_offset, data_page_offset, data_page_index, page_length;
 687         int ret;
 688         uint64_t data_ptr = args->data_ptr;
 689
 690         remain = args->size;
 691
 692         /* Pin the user pages containing the data.  We can't fault while
 693          * holding the struct mutex, and all of the pwrite implementations
 694          * want to hold it while dereferencing the user data.
 695          */
 696         first_data_page = data_ptr / PAGE_SIZE;
 697         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 698         num_pages = last_data_page - first_data_page + 1;
 699
 700         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 701         if (user_pages == NULL)
 702                 return -ENOMEM;
 703
 704         mutex_unlock(&dev->struct_mutex);
 705         down_read(&mm->mmap_sem);
 706         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 707                                       num_pages, 0, 0, user_pages, NULL);
 708         up_read(&mm->mmap_sem);
 709         mutex_lock(&dev->struct_mutex);
 710         if (pinned_pages < num_pages) {
 711                 ret = -EFAULT;
 712                 goto out_unpin_pages;
 713         }
 714
 715         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 716         if (ret)
 717                 goto out_unpin_pages;
 718
 719         ret = i915_gem_object_put_fence(obj);
 720         if (ret)
 721                 goto out_unpin_pages;
 722
 723         offset = obj->gtt_offset + args->offset;
 724
 725         while (remain > 0) {
 726                 /* Operation in this page
 727                  *
 728                  * gtt_page_base = page offset within aperture
 729                  * gtt_page_offset = offset within page in aperture
 730                  * data_page_index = page number in get_user_pages return
 731                  * data_page_offset = offset with data_page_index page.
 732                  * page_length = bytes to copy for this page
 733                  */
 734                 gtt_page_base = offset & PAGE_MASK;
 735                 gtt_page_offset = offset_in_page(offset);
 736                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 737                 data_page_offset = offset_in_page(data_ptr);
 738
 739                 page_length = remain;
 740                 if ((gtt_page_offset + page_length) > PAGE_SIZE)
 741                         page_length = PAGE_SIZE - gtt_page_offset;
 742                 if ((data_page_offset + page_length) > PAGE_SIZE)
 743                         page_length = PAGE_SIZE - data_page_offset;
 744
 745                 slow_kernel_write(dev_priv->mm.gtt_mapping,
 746                                   gtt_page_base, gtt_page_offset,
 747                                   user_pages[data_page_index],
 748                                   data_page_offset,
 749                                   page_length);
 750
 751                 remain -= page_length;
 752                 offset += page_length;
 753                 data_ptr += page_length;
 754         }
 755
 756 out_unpin_pages:
 757         for (i = 0; i < pinned_pages; i++)
 758                 page_cache_release(user_pages[i]);
 759         drm_free_large(user_pages);
 760
 761         return ret;
 762 }
 763
 764 /**
 765  * This is the fast shmem pwrite path, which attempts to directly
 766  * copy_from_user into the kmapped pages backing the object.
 767  */
 768 static int
 769 i915_gem_shmem_pwrite_fast(struct drm_device *dev,
 770                            struct drm_i915_gem_object *obj,
 771                            struct drm_i915_gem_pwrite *args,
 772                            struct drm_file *file)
 773 {
 774         struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
 775         ssize_t remain;
 776         loff_t offset;
 777         char __user *user_data;
 778         int page_offset, page_length;
 779
 780         user_data = (char __user *) (uintptr_t) args->data_ptr;
 781         remain = args->size;
 782
 783         offset = args->offset;
 784         obj->dirty = 1;
 785
 786         while (remain > 0) {
 787                 struct page *page;
 788                 char *vaddr;
 789                 int ret;
 790
 791                 /* Operation in this page
 792                  *
 793                  * page_offset = offset within page
 794                  * page_length = bytes to copy for this page
 795                  */
 796                 page_offset = offset_in_page(offset);
 797                 page_length = remain;
 798                 if ((page_offset + remain) > PAGE_SIZE)
 799                         page_length = PAGE_SIZE - page_offset;
 800
 801                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 802                 if (IS_ERR(page))
 803                         return PTR_ERR(page);
 804
 805                 vaddr = kmap_atomic(page);
 806                 ret = __copy_from_user_inatomic(vaddr + page_offset,
 807                                                 user_data,
 808                                                 page_length);
 809                 kunmap_atomic(vaddr);
 810
 811                 set_page_dirty(page);
 812                 mark_page_accessed(page);
 813                 page_cache_release(page);
 814
 815                 /* If we get a fault while copying data, then (presumably) our
 816                  * source page isn't available.  Return the error and we'll
 817                  * retry in the slow path.
 818                  */
 819                 if (ret)
 820                         return -EFAULT;
 821
 822                 remain -= page_length;
 823                 user_data += page_length;
 824                 offset += page_length;
 825         }
 826
 827         return 0;
 828 }
 829
 830 /**
 831  * This is the fallback shmem pwrite path, which uses get_user_pages to pin
 832  * the memory and maps it using kmap_atomic for copying.
 833  *
 834  * This avoids taking mmap_sem for faulting on the user's address while the
 835  * struct_mutex is held.
 836  */
 837 static int
 838 i915_gem_shmem_pwrite_slow(struct drm_device *dev,
 839                            struct drm_i915_gem_object *obj,
 840                            struct drm_i915_gem_pwrite *args,
 841                            struct drm_file *file)
 842 {
 843         struct address_space *mapping = obj->base.filp->f_path.dentry->d_inode->i_mapping;
 844         struct mm_struct *mm = current->mm;
 845         struct page **user_pages;
 846         ssize_t remain;
 847         loff_t offset, pinned_pages, i;
 848         loff_t first_data_page, last_data_page, num_pages;
 849         int shmem_page_offset;
 850         int data_page_index,  data_page_offset;
 851         int page_length;
 852         int ret;
 853         uint64_t data_ptr = args->data_ptr;
 854         int do_bit17_swizzling;
 855
 856         remain = args->size;
 857
 858         /* Pin the user pages containing the data.  We can't fault while
 859          * holding the struct mutex, and all of the pwrite implementations
 860          * want to hold it while dereferencing the user data.
 861          */
 862         first_data_page = data_ptr / PAGE_SIZE;
 863         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 864         num_pages = last_data_page - first_data_page + 1;
 865
 866         user_pages = drm_malloc_ab(num_pages, sizeof(struct page *));
 867         if (user_pages == NULL)
 868                 return -ENOMEM;
 869
 870         mutex_unlock(&dev->struct_mutex);
 871         down_read(&mm->mmap_sem);
 872         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 873                                       num_pages, 0, 0, user_pages, NULL);
 874         up_read(&mm->mmap_sem);
 875         mutex_lock(&dev->struct_mutex);
 876         if (pinned_pages < num_pages) {
 877                 ret = -EFAULT;
 878                 goto out;
 879         }
 880
 881         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
 882         if (ret)
 883                 goto out;
 884
 885         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 886
 887         offset = args->offset;
 888         obj->dirty = 1;
 889
 890         while (remain > 0) {
 891                 struct page *page;
 892
 893                 /* Operation in this page
 894                  *
 895                  * shmem_page_offset = offset within page in shmem file
 896                  * data_page_index = page number in get_user_pages return
 897                  * data_page_offset = offset with data_page_index page.
 898                  * page_length = bytes to copy for this page
 899                  */
 900                 shmem_page_offset = offset_in_page(offset);
 901                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 902                 data_page_offset = offset_in_page(data_ptr);
 903
 904                 page_length = remain;
 905                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 906                         page_length = PAGE_SIZE - shmem_page_offset;
 907                 if ((data_page_offset + page_length) > PAGE_SIZE)
 908                         page_length = PAGE_SIZE - data_page_offset;
 909
 910                 page = shmem_read_mapping_page(mapping, offset >> PAGE_SHIFT);
 911                 if (IS_ERR(page)) {
 912                         ret = PTR_ERR(page);
 913                         goto out;
 914                 }
 915
 916                 if (do_bit17_swizzling) {
 917                         slow_shmem_bit17_copy(page,
 918                                               shmem_page_offset,
 919                                               user_pages[data_page_index],
 920                                               data_page_offset,
 921                                               page_length,
 922                                               0);
 923                 } else {
 924                         slow_shmem_copy(page,
 925                                         shmem_page_offset,
 926                                         user_pages[data_page_index],
 927                                         data_page_offset,
 928                                         page_length);
 929                 }
 930
 931                 set_page_dirty(page);
 932                 mark_page_accessed(page);
 933                 page_cache_release(page);
 934
 935                 remain -= page_length;
 936                 data_ptr += page_length;
 937                 offset += page_length;
 938         }
 939
 940 out:
 941         for (i = 0; i < pinned_pages; i++)
 942                 page_cache_release(user_pages[i]);
 943         drm_free_large(user_pages);
 944
 945         return ret;
 946 }
 947
 948 /**
 949  * Writes data to the object referenced by handle.
 950  *
 951  * On error, the contents of the buffer that were to be modified are undefined.
 952  */
 953 int
 954 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 955                       struct drm_file *file)
 956 {
 957         struct drm_i915_gem_pwrite *args = data;
 958         struct drm_i915_gem_object *obj;
 959         int ret;
 960
 961         if (args->size == 0)
 962                 return 0;
 963
 964         if (!access_ok(VERIFY_READ,
 965                        (char __user *)(uintptr_t)args->data_ptr,
 966                        args->size))
 967                 return -EFAULT;
 968
 969         ret = fault_in_pages_readable((char __user *)(uintptr_t)args->data_ptr,
 970                                       args->size);
 971         if (ret)
 972                 return -EFAULT;
 973
 974         ret = i915_mutex_lock_interruptible(dev);
 975         if (ret)
 976                 return ret;
 977
 978         obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
 979         if (&obj->base == NULL) {
 980                 ret = -ENOENT;
 981                 goto unlock;
 982         }
 983
 984         /* Bounds check destination. */
 985         if (args->offset > obj->base.size ||
 986             args->size > obj->base.size - args->offset) {
 987                 ret = -EINVAL;
 988                 goto out;
 989         }
 990
 991         trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 992
 993         /* We can only do the GTT pwrite on untiled buffers, as otherwise
 994          * it would end up going through the fenced access, and we'll get
 995          * different detiling behavior between reading and writing.
 996          * pread/pwrite currently are reading and writing from the CPU
 997          * perspective, requiring manual detiling by the client.
 998          */
 999         if (obj->phys_obj)
1000                 ret = i915_gem_phys_pwrite(dev, obj, args, file);
1001         else if (obj->gtt_space &&
1002                  obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
1003                 ret = i915_gem_object_pin(obj, 0, true);
1004                 if (ret)
1005                         goto out;
1006
1007                 ret = i915_gem_object_set_to_gtt_domain(obj, true);
1008                 if (ret)
1009                         goto out_unpin;
1010
1011                 ret = i915_gem_object_put_fence(obj);
1012                 if (ret)
1013                         goto out_unpin;
1014
1015                 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file);
1016                 if (ret == -EFAULT)
1017                         ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file);
1018
1019 out_unpin:
1020                 i915_gem_object_unpin(obj);
1021         } else {
1022                 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1023                 if (ret)
1024                         goto out;
1025
1026                 ret = -EFAULT;
1027                 if (!i915_gem_object_needs_bit17_swizzle(obj))
1028                         ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file);
1029                 if (ret == -EFAULT)
1030                         ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file);
1031         }
1032
1033 out:
1034         drm_gem_object_unreference(&obj->base);
1035 unlock:
1036         mutex_unlock(&dev->struct_mutex);
1037         return ret;
1038 }
1039
1040 /**
1041  * Called when user space prepares to use an object with the CPU, either
1042  * through the mmap ioctl's mapping or a GTT mapping.
1043  */
1044 int
1045 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
1046                           struct drm_file *file)
1047 {
1048         struct drm_i915_gem_set_domain *args = data;
1049         struct drm_i915_gem_object *obj;
1050         uint32_t read_domains = args->read_domains;
1051         uint32_t write_domain = args->write_domain;
1052         int ret;
1053
1054         if (!(dev->driver->driver_features & DRIVER_GEM))
1055                 return -ENODEV;
1056
1057         /* Only handle setting domains to types used by the CPU. */
1058         if (write_domain & I915_GEM_GPU_DOMAINS)
1059                 return -EINVAL;
1060
1061         if (read_domains & I915_GEM_GPU_DOMAINS)
1062                 return -EINVAL;
1063
1064         /* Having something in the write domain implies it's in the read
1065          * domain, and only that read domain.  Enforce that in the request.
1066          */
1067         if (write_domain != 0 && read_domains != write_domain)
1068                 return -EINVAL;
1069
1070         ret = i915_mutex_lock_interruptible(dev);
1071         if (ret)
1072                 return ret;
1073
1074         obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1075         if (&obj->base == NULL) {
1076                 ret = -ENOENT;
1077                 goto unlock;
1078         }
1079
1080         if (read_domains & I915_GEM_DOMAIN_GTT) {
1081                 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1082
1083                 /* Silently promote "you're not bound, there was nothing to do"
1084                  * to success, since the client was just asking us to
1085                  * make sure everything was done.
1086                  */
1087                 if (ret == -EINVAL)
1088                         ret = 0;
1089         } else {
1090                 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1091         }
1092
1093         drm_gem_object_unreference(&obj->base);
1094 unlock:
1095         mutex_unlock(&dev->struct_mutex);
1096         return ret;
1097 }
1098
1099 /**
1100  * Called when user space has done writes to this buffer
1101  */
1102 int
1103 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1104                          struct drm_file *file)
1105 {
1106         struct drm_i915_gem_sw_finish *args = data;
1107         struct drm_i915_gem_object *obj;
1108         int ret = 0;
1109
1110         if (!(dev->driver->driver_features & DRIVER_GEM))
1111                 return -ENODEV;
1112
1113         ret = i915_mutex_lock_interruptible(dev);
1114         if (ret)
1115                 return ret;
1116
1117         obj = to_intel_bo(drm_gem_object_lookup(dev, file, args->handle));
1118         if (&obj->base == NULL) {
1119                 ret = -ENOENT;
1120                 goto unlock;
1121         }
1122
1123         /* Pinned buffers may be scanout, so flush the cache */
1124         if (obj->pin_count)
1125                 i915_gem_object_flush_cpu_write_domain(obj);
1126
1127         drm_gem_object_unreference(&obj->base);
1128 unlock:
1129         mutex_unlock(&dev->struct_mutex);
1130         return ret;
1131 }
1132
1133 /**
1134  * Maps the contents of an object, returning the address it is mapped
1135  * into.
1136  *
1137  * While the mapping holds a reference on the contents of the object, it doesn't
1138  * imply a ref on the object itself.
1139  */
1140 int
1141 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1142                     struct drm_file *file)
1143 {
1144         struct drm_i915_private *dev_priv = dev->dev_private;
1145         struct drm_i915_gem_mmap *args = data;
1146         struct drm_gem_object *obj;
1147         unsigned long addr;
1148
1149         if (!(dev->driver->driver_features & DRIVER_GEM))
1150                 return -ENODEV;
1151
1152         obj = drm_gem_object_lookup(dev, file, args->handle);
1153         if (obj == NULL)
1154                 return -ENOENT;
1155
1156         if (obj->size > dev_priv->mm.gtt_mappable_end) {
1157                 drm_gem_object_unreference_unlocked(obj);
1158                 return -E2BIG;
1159         }
1160
1161         down_write(&current->mm->mmap_sem);
1162         addr = do_mmap(obj->filp, 0, args->size,
1163                        PROT_READ | PROT_WRITE, MAP_SHARED,
1164                        args->offset);
1165         up_write(&current->mm->mmap_sem);
1166         drm_gem_object_unreference_unlocked(obj);
1167         if (IS_ERR((void *)addr))
1168                 return addr;
1169
1170         args->addr_ptr = (uint64_t) addr;
1171
1172         return 0;
1173 }
1174
1175 /**
1176  * i915_gem_fault - fault a page into the GTT
1177  * vma: VMA in question
1178  * vmf: fault info
1179  *
1180  * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1181  * from userspace.  The fault handler takes care of binding the object to
1182  * the GTT (if needed), allocating and programming a fence register (again,
1183  * only if needed based on whether the old reg is still valid or the object
1184  * is tiled) and inserting a new PTE into the faulting process.
1185  *
1186  * Note that the faulting process may involve evicting existing objects
1187  * from the GTT and/or fence registers to make room.  So performance may
1188  * suffer if the GTT working set is large or there are few fence registers
1189  * left.
1190  */
1191 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1192 {
1193         struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
1194         struct drm_device *dev = obj->base.dev;
1195         drm_i915_private_t *dev_priv = dev->dev_private;
1196         pgoff_t page_offset;
1197         unsigned long pfn;
1198         int ret = 0;
1199         bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1200
1201         /* We don't use vmf->pgoff since that has the fake offset */
1202         page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1203                 PAGE_SHIFT;
1204
1205         ret = i915_mutex_lock_interruptible(dev);
1206         if (ret)
1207                 goto out;
1208
1209         trace_i915_gem_object_fault(obj, page_offset, true, write);
1210
1211         /* Now bind it into the GTT if needed */
1212         if (!obj->map_and_fenceable) {
1213                 ret = i915_gem_object_unbind(obj);
1214                 if (ret)
1215                         goto unlock;
1216         }
1217         if (!obj->gtt_space) {
1218                 ret = i915_gem_object_bind_to_gtt(obj, 0, true);
1219                 if (ret)
1220                         goto unlock;
1221
1222                 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1223                 if (ret)
1224                         goto unlock;
1225         }
1226
1227         if (obj->tiling_mode == I915_TILING_NONE)
1228                 ret = i915_gem_object_put_fence(obj);
1229         else
1230                 ret = i915_gem_object_get_fence(obj, NULL);
1231         if (ret)
1232                 goto unlock;
1233
1234         if (i915_gem_object_is_inactive(obj))
1235                 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1236
1237         obj->fault_mappable = true;
1238
1239         pfn = ((dev->agp->base + obj->gtt_offset) >> PAGE_SHIFT) +
1240                 page_offset;
1241
1242         /* Finally, remap it using the new GTT offset */
1243         ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1244 unlock:
1245         mutex_unlock(&dev->struct_mutex);
1246 out:
1247         switch (ret) {
1248         case -EIO:
1249         case -EAGAIN:
1250                 /* Give the error handler a chance to run and move the
1251                  * objects off the GPU active list. Next time we service the
1252                  * fault, we should be able to transition the page into the
1253                  * GTT without touching the GPU (and so avoid further
1254                  * EIO/EGAIN). If the GPU is wedged, then there is no issue
1255                  * with coherency, just lost writes.
1256                  */
1257                 set_need_resched();
1258         case 0:
1259         case -ERESTARTSYS:
1260         case -EINTR:
1261                 return VM_FAULT_NOPAGE;
1262         case -ENOMEM:
1263                 return VM_FAULT_OOM;
1264         default:
1265                 return VM_FAULT_SIGBUS;
1266         }
1267 }
1268
1269 /**
1270  * i915_gem_release_mmap - remove physical page mappings
1271  * @obj: obj in question
1272  *
1273  * Preserve the reservation of the mmapping with the DRM core code, but
1274  * relinquish ownership of the pages back to the system.
1275  *
1276  * It is vital that we remove the page mapping if we have mapped a tiled
1277  * object through the GTT and then lose the fence register due to
1278  * resource pressure. Similarly if the object has been moved out of the
1279  * aperture, than pages mapped into userspace must be revoked. Removing the
1280  * mapping will then trigger a page fault on the next user access, allowing
1281  * fixup by i915_gem_fault().
1282  */
1283 void
1284 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
1285 {
1286         if (!obj->fault_mappable)
1287                 return;
1288
1289         if (obj->base.dev->dev_mapping)
1290                 unmap_mapping_range(obj->base.dev->dev_mapping,
1291                                     (loff_t)obj->base.map_list.hash.key<<PAGE_SHIFT,
1292                                     obj->base.size, 1);
1293
1294         obj->fault_mappable = false;
1295 }
1296
1297 static uint32_t
1298 i915_gem_get_gtt_size(struct drm_device *dev, uint32_t size, int tiling_mode)
1299 {
1300         uint32_t gtt_size;
1301
1302         if (INTEL_INFO(dev)->gen >= 4 ||
1303             tiling_mode == I915_TILING_NONE)
1304                 return size;
1305
1306         /* Previous chips need a power-of-two fence region when tiling */
1307         if (INTEL_INFO(dev)->gen == 3)
1308                 gtt_size = 1024*1024;
1309         else
1310                 gtt_size = 512*1024;
1311
1312         while (gtt_size < size)
1313                 gtt_size <<= 1;
1314
1315         return gtt_size;
1316 }
1317
1318 /**
1319  * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1320  * @obj: object to check
1321  *
1322  * Return the required GTT alignment for an object, taking into account
1323  * potential fence register mapping.
1324  */
1325 static uint32_t
1326 i915_gem_get_gtt_alignment(struct drm_device *dev,
1327                            uint32_t size,
1328                            int tiling_mode)
1329 {
1330         /*
1331          * Minimum alignment is 4k (GTT page size), but might be greater
1332          * if a fence register is needed for the object.
1333          */
1334         if (INTEL_INFO(dev)->gen >= 4 ||
1335             tiling_mode == I915_TILING_NONE)
1336                 return 4096;
1337
1338         /*
1339          * Previous chips need to be aligned to the size of the smallest
1340          * fence register that can contain the object.
1341          */
1342         return i915_gem_get_gtt_size(dev, size, tiling_mode);
1343 }
1344
1345 /**
1346  * i915_gem_get_unfenced_gtt_alignment - return required GTT alignment for an
1347  *                                       unfenced object
1348  * @dev: the device
1349  * @size: size of the object
1350  * @tiling_mode: tiling mode of the object
1351  *
1352  * Return the required GTT alignment for an object, only taking into account
1353  * unfenced tiled surface requirements.
1354  */
1355 uint32_t
1356 i915_gem_get_unfenced_gtt_alignment(struct drm_device *dev,
1357                                     uint32_t size,
1358                                     int tiling_mode)
1359 {
1360         /*
1361          * Minimum alignment is 4k (GTT page size) for sane hw.
1362          */
1363         if (INTEL_INFO(dev)->gen >= 4 || IS_G33(dev) ||
1364             tiling_mode == I915_TILING_NONE)
1365                 return 4096;
1366
1367         /* Previous hardware however needs to be aligned to a power-of-two
1368          * tile height. The simplest method for determining this is to reuse
1369          * the power-of-tile object size.
1370          */
1371         return i915_gem_get_gtt_size(dev, size, tiling_mode);
1372 }
1373
1374 int
1375 i915_gem_mmap_gtt(struct drm_file *file,
1376                   struct drm_device *dev,
1377                   uint32_t handle,
1378                   uint64_t *offset)
1379 {
1380         struct drm_i915_private *dev_priv = dev->dev_private;
1381         struct drm_i915_gem_object *obj;
1382         int ret;
1383
1384         if (!(dev->driver->driver_features & DRIVER_GEM))
1385                 return -ENODEV;
1386
1387         ret = i915_mutex_lock_interruptible(dev);
1388         if (ret)
1389                 return ret;
1390
1391         obj = to_intel_bo(drm_gem_object_lookup(dev, file, handle));
1392         if (&obj->base == NULL) {
1393                 ret = -ENOENT;
1394                 goto unlock;
1395         }
1396
1397         if (obj->base.size > dev_priv->mm.gtt_mappable_end) {
1398                 ret = -E2BIG;
1399                 goto out;
1400         }
1401
1402         if (obj->madv != I915_MADV_WILLNEED) {
1403                 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1404                 ret = -EINVAL;
1405                 goto out;
1406         }
1407
1408         if (!obj->base.map_list.map) {
1409                 ret = drm_gem_create_mmap_offset(&obj->base);
1410                 if (ret)
1411                         goto out;
1412         }
1413
1414         *offset = (u64)obj->base.map_list.hash.key << PAGE_SHIFT;
1415
1416 out:
1417         drm_gem_object_unreference(&obj->base);
1418 unlock:
1419         mutex_unlock(&dev->struct_mutex);
1420         return ret;
1421 }
1422
1423 /**
1424  * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1425  * @dev: DRM device
1426  * @data: GTT mapping ioctl data
1427  * @file: GEM object info
1428  *
1429  * Simply returns the fake offset to userspace so it can mmap it.
1430  * The mmap call will end up in drm_gem_mmap(), which will set things
1431  * up so we can get faults in the handler above.
1432  *
1433  * The fault handler will take care of binding the object into the GTT
1434  * (since it may have been evicted to make room for something), allocating
1435  * a fence register, and mapping the appropriate aperture address into
1436  * userspace.
1437  */
1438 int
1439 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1440                         struct drm_file *file)
1441 {
1442         struct drm_i915_gem_mmap_gtt *args = data;
1443
1444         if (!(dev->driver->driver_features & DRIVER_GEM))
1445                 return -ENODEV;
1446
1447         return i915_gem_mmap_gtt(file, dev, args->handle, &args->offset);
1448 }
1449
1450
1451 static int
1452 i915_gem_object_get_pages_gtt(struct drm_i915_gem_object *obj,
1453                               gfp_t gfpmask)
1454 {
1455         int page_count, i;
1456         struct address_space *mapping;
1457         struct inode *inode;
1458         struct page *page;
1459
1460         /* Get the list of pages out of our struct file.  They'll be pinned
1461          * at this point until we release them.
1462          */
1463         page_count = obj->base.size / PAGE_SIZE;
1464         BUG_ON(obj->pages != NULL);
1465         obj->pages = drm_malloc_ab(page_count, sizeof(struct page *));
1466         if (obj->pages == NULL)
1467                 return -ENOMEM;
1468
1469         inode = obj->base.filp->f_path.dentry->d_inode;
1470         mapping = inode->i_mapping;
1471         gfpmask |= mapping_gfp_mask(mapping);
1472
1473         for (i = 0; i < page_count; i++) {
1474                 page = shmem_read_mapping_page_gfp(mapping, i, gfpmask);
1475                 if (IS_ERR(page))
1476                         goto err_pages;
1477
1478                 obj->pages[i] = page;
1479         }
1480
1481         if (i915_gem_object_needs_bit17_swizzle(obj))
1482                 i915_gem_object_do_bit_17_swizzle(obj);
1483
1484         return 0;
1485
1486 err_pages:
1487         while (i--)
1488                 page_cache_release(obj->pages[i]);
1489
1490         drm_free_large(obj->pages);
1491         obj->pages = NULL;
1492         return PTR_ERR(page);
1493 }
1494
1495 static void
1496 i915_gem_object_put_pages_gtt(struct drm_i915_gem_object *obj)
1497 {
1498         int page_count = obj->base.size / PAGE_SIZE;
1499         int i;
1500
1501         BUG_ON(obj->madv == __I915_MADV_PURGED);
1502
1503         if (i915_gem_object_needs_bit17_swizzle(obj))
1504                 i915_gem_object_save_bit_17_swizzle(obj);
1505
1506         if (obj->madv == I915_MADV_DONTNEED)
1507                 obj->dirty = 0;
1508
1509         for (i = 0; i < page_count; i++) {
1510                 if (obj->dirty)
1511                         set_page_dirty(obj->pages[i]);
1512
1513                 if (obj->madv == I915_MADV_WILLNEED)
1514                         mark_page_accessed(obj->pages[i]);
1515
1516                 page_cache_release(obj->pages[i]);
1517         }
1518         obj->dirty = 0;
1519
1520         drm_free_large(obj->pages);
1521         obj->pages = NULL;
1522 }
1523
1524 void
1525 i915_gem_object_move_to_active(struct drm_i915_gem_object *obj,
1526                                struct intel_ring_buffer *ring,
1527                                u32 seqno)
1528 {
1529         struct drm_device *dev = obj->base.dev;
1530         struct drm_i915_private *dev_priv = dev->dev_private;
1531
1532         BUG_ON(ring == NULL);
1533         obj->ring = ring;
1534
1535         /* Add a reference if we're newly entering the active list. */
1536         if (!obj->active) {
1537                 drm_gem_object_reference(&obj->base);
1538                 obj->active = 1;
1539         }
1540
1541         /* Move from whatever list we were on to the tail of execution. */
1542         list_move_tail(&obj->mm_list, &dev_priv->mm.active_list);
1543         list_move_tail(&obj->ring_list, &ring->active_list);
1544
1545         obj->last_rendering_seqno = seqno;
1546         if (obj->fenced_gpu_access) {
1547                 struct drm_i915_fence_reg *reg;
1548
1549                 BUG_ON(obj->fence_reg == I915_FENCE_REG_NONE);
1550
1551                 obj->last_fenced_seqno = seqno;
1552                 obj->last_fenced_ring = ring;
1553
1554                 reg = &dev_priv->fence_regs[obj->fence_reg];
1555                 list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
1556         }
1557 }
1558
1559 static void
1560 i915_gem_object_move_off_active(struct drm_i915_gem_object *obj)
1561 {
1562         list_del_init(&obj->ring_list);
1563         obj->last_rendering_seqno = 0;
1564 }
1565
1566 static void
1567 i915_gem_object_move_to_flushing(struct drm_i915_gem_object *obj)
1568 {
1569         struct drm_device *dev = obj->base.dev;
1570         drm_i915_private_t *dev_priv = dev->dev_private;
1571
1572         BUG_ON(!obj->active);
1573         list_move_tail(&obj->mm_list, &dev_priv->mm.flushing_list);
1574
1575         i915_gem_object_move_off_active(obj);
1576 }
1577
1578 static void
1579 i915_gem_object_move_to_inactive(struct drm_i915_gem_object *obj)
1580 {
1581         struct drm_device *dev = obj->base.dev;
1582         struct drm_i915_private *dev_priv = dev->dev_private;
1583
1584         if (obj->pin_count != 0)
1585                 list_move_tail(&obj->mm_list, &dev_priv->mm.pinned_list);
1586         else
1587                 list_move_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
1588
1589         BUG_ON(!list_empty(&obj->gpu_write_list));
1590         BUG_ON(!obj->active);
1591         obj->ring = NULL;
1592
1593         i915_gem_object_move_off_active(obj);
1594         obj->fenced_gpu_access = false;
1595
1596         obj->active = 0;
1597         obj->pending_gpu_write = false;
1598         drm_gem_object_unreference(&obj->base);
1599
1600         WARN_ON(i915_verify_lists(dev));
1601 }
1602
1603 /* Immediately discard the backing storage */
1604 static void
1605 i915_gem_object_truncate(struct drm_i915_gem_object *obj)
1606 {
1607         struct inode *inode;
1608
1609         /* Our goal here is to return as much of the memory as
1610          * is possible back to the system as we are called from OOM.
1611          * To do this we must instruct the shmfs to drop all of its
1612          * backing pages, *now*.
1613          */
1614         inode = obj->base.filp->f_path.dentry->d_inode;
1615         shmem_truncate_range(inode, 0, (loff_t)-1);
1616
1617         obj->madv = __I915_MADV_PURGED;
1618 }
1619
1620 static inline int
1621 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj)
1622 {
1623         return obj->madv == I915_MADV_DONTNEED;
1624 }
1625
1626 static void
1627 i915_gem_process_flushing_list(struct intel_ring_buffer *ring,
1628                                uint32_t flush_domains)
1629 {
1630         struct drm_i915_gem_object *obj, *next;
1631
1632         list_for_each_entry_safe(obj, next,
1633                                  &ring->gpu_write_list,
1634                                  gpu_write_list) {
1635                 if (obj->base.write_domain & flush_domains) {
1636                         uint32_t old_write_domain = obj->base.write_domain;
1637
1638                         obj->base.write_domain = 0;
1639                         list_del_init(&obj->gpu_write_list);
1640                         i915_gem_object_move_to_active(obj, ring,
1641                                                        i915_gem_next_request_seqno(ring));
1642
1643                         trace_i915_gem_object_change_domain(obj,
1644                                                             obj->base.read_domains,
1645                                                             old_write_domain);
1646                 }
1647         }
1648 }
1649
1650 int
1651 i915_add_request(struct intel_ring_buffer *ring,
1652                  struct drm_file *file,
1653                  struct drm_i915_gem_request *request)
1654 {
1655         drm_i915_private_t *dev_priv = ring->dev->dev_private;
1656         uint32_t seqno;
1657         int was_empty;
1658         int ret;
1659
1660         BUG_ON(request == NULL);
1661
1662         ret = ring->add_request(ring, &seqno);
1663         if (ret)
1664             return ret;
1665
1666         trace_i915_gem_request_add(ring, seqno);
1667
1668         request->seqno = seqno;
1669         request->ring = ring;
1670         request->emitted_jiffies = jiffies;
1671         was_empty = list_empty(&ring->request_list);
1672         list_add_tail(&request->list, &ring->request_list);
1673
1674         if (file) {
1675                 struct drm_i915_file_private *file_priv = file->driver_priv;
1676
1677                 spin_lock(&file_priv->mm.lock);
1678                 request->file_priv = file_priv;
1679                 list_add_tail(&request->client_list,
1680                               &file_priv->mm.request_list);
1681                 spin_unlock(&file_priv->mm.lock);
1682         }
1683
1684         ring->outstanding_lazy_request = false;
1685
1686         if (!dev_priv->mm.suspended) {
1687                 if (i915_enable_hangcheck) {
1688                         mod_timer(&dev_priv->hangcheck_timer,
1689                                   jiffies +
1690                                   msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD));
1691                 }
1692                 if (was_empty)
1693                         queue_delayed_work(dev_priv->wq,
1694                                            &dev_priv->mm.retire_work, HZ);
1695         }
1696         return 0;
1697 }
1698
1699 static inline void
1700 i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
1701 {
1702         struct drm_i915_file_private *file_priv = request->file_priv;
1703
1704         if (!file_priv)
1705                 return;
1706
1707         spin_lock(&file_priv->mm.lock);
1708         if (request->file_priv) {
1709                 list_del(&request->client_list);
1710                 request->file_priv = NULL;
1711         }
1712         spin_unlock(&file_priv->mm.lock);
1713 }
1714
1715 static void i915_gem_reset_ring_lists(struct drm_i915_private *dev_priv,
1716                                       struct intel_ring_buffer *ring)
1717 {
1718         while (!list_empty(&ring->request_list)) {
1719                 struct drm_i915_gem_request *request;
1720
1721                 request = list_first_entry(&ring->request_list,
1722                                            struct drm_i915_gem_request,
1723                                            list);
1724
1725                 list_del(&request->list);
1726                 i915_gem_request_remove_from_client(request);
1727                 kfree(request);
1728         }
1729
1730         while (!list_empty(&ring->active_list)) {
1731                 struct drm_i915_gem_object *obj;
1732
1733                 obj = list_first_entry(&ring->active_list,
1734                                        struct drm_i915_gem_object,
1735                                        ring_list);
1736
1737                 obj->base.write_domain = 0;
1738                 list_del_init(&obj->gpu_write_list);
1739                 i915_gem_object_move_to_inactive(obj);
1740         }
1741 }
1742
1743 static void i915_gem_reset_fences(struct drm_device *dev)
1744 {
1745         struct drm_i915_private *dev_priv = dev->dev_private;
1746         int i;
1747
1748         for (i = 0; i < dev_priv->num_fence_regs; i++) {
1749                 struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
1750                 struct drm_i915_gem_object *obj = reg->obj;
1751
1752                 if (!obj)
1753                         continue;
1754
1755                 if (obj->tiling_mode)
1756                         i915_gem_release_mmap(obj);
1757
1758                 reg->obj->fence_reg = I915_FENCE_REG_NONE;
1759                 reg->obj->fenced_gpu_access = false;
1760                 reg->obj->last_fenced_seqno = 0;
1761                 reg->obj->last_fenced_ring = NULL;
1762                 i915_gem_clear_fence_reg(dev, reg);
1763         }
1764 }
1765
1766 void i915_gem_reset(struct drm_device *dev)
1767 {
1768         struct drm_i915_private *dev_priv = dev->dev_private;
1769         struct drm_i915_gem_object *obj;
1770         int i;
1771
1772         for (i = 0; i < I915_NUM_RINGS; i++)
1773                 i915_gem_reset_ring_lists(dev_priv, &dev_priv->ring[i]);
1774
1775         /* Remove anything from the flushing lists. The GPU cache is likely
1776          * to be lost on reset along with the data, so simply move the
1777          * lost bo to the inactive list.
1778          */
1779         while (!list_empty(&dev_priv->mm.flushing_list)) {
1780                 obj = list_first_entry(&dev_priv->mm.flushing_list,
1781                                       struct drm_i915_gem_object,
1782                                       mm_list);
1783
1784                 obj->base.write_domain = 0;
1785                 list_del_init(&obj->gpu_write_list);
1786                 i915_gem_object_move_to_inactive(obj);
1787         }
1788
1789         /* Move everything out of the GPU domains to ensure we do any
1790          * necessary invalidation upon reuse.
1791          */
1792         list_for_each_entry(obj,
1793                             &dev_priv->mm.inactive_list,
1794                             mm_list)
1795         {
1796                 obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
1797         }
1798
1799         /* The fence registers are invalidated so clear them out */
1800         i915_gem_reset_fences(dev);
1801 }
1802
1803 /**
1804  * This function clears the request list as sequence numbers are passed.
1805  */
1806 static void
1807 i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
1808 {
1809         uint32_t seqno;
1810         int i;
1811
1812         if (list_empty(&ring->request_list))
1813                 return;
1814
1815         WARN_ON(i915_verify_lists(ring->dev));
1816
1817         seqno = ring->get_seqno(ring);
1818
1819         for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++)
1820                 if (seqno >= ring->sync_seqno[i])
1821                         ring->sync_seqno[i] = 0;
1822
1823         while (!list_empty(&ring->request_list)) {
1824                 struct drm_i915_gem_request *request;
1825
1826                 request = list_first_entry(&ring->request_list,
1827                                            struct drm_i915_gem_request,
1828                                            list);
1829
1830                 if (!i915_seqno_passed(seqno, request->seqno))
1831                         break;
1832
1833                 trace_i915_gem_request_retire(ring, request->seqno);
1834
1835                 list_del(&request->list);
1836                 i915_gem_request_remove_from_client(request);
1837                 kfree(request);
1838         }
1839
1840         /* Move any buffers on the active list that are no longer referenced
1841          * by the ringbuffer to the flushing/inactive lists as appropriate.
1842          */
1843         while (!list_empty(&ring->active_list)) {
1844                 struct drm_i915_gem_object *obj;
1845
1846                 obj = list_first_entry(&ring->active_list,
1847                                       struct drm_i915_gem_object,
1848                                       ring_list);
1849
1850                 if (!i915_seqno_passed(seqno, obj->last_rendering_seqno))
1851                         break;
1852
1853                 if (obj->base.write_domain != 0)
1854                         i915_gem_object_move_to_flushing(obj);
1855                 else
1856                         i915_gem_object_move_to_inactive(obj);
1857         }
1858
1859         if (unlikely(ring->trace_irq_seqno &&
1860                      i915_seqno_passed(seqno, ring->trace_irq_seqno))) {
1861                 ring->irq_put(ring);
1862                 ring->trace_irq_seqno = 0;
1863         }
1864
1865         WARN_ON(i915_verify_lists(ring->dev));
1866 }
1867
1868 void
1869 i915_gem_retire_requests(struct drm_device *dev)
1870 {
1871         drm_i915_private_t *dev_priv = dev->dev_private;
1872         int i;
1873
1874         if (!list_empty(&dev_priv->mm.deferred_free_list)) {
1875             struct drm_i915_gem_object *obj, *next;
1876
1877             /* We must be careful that during unbind() we do not
1878              * accidentally infinitely recurse into retire requests.
1879              * Currently:
1880              *   retire -> free -> unbind -> wait -> retire_ring
1881              */
1882             list_for_each_entry_safe(obj, next,
1883                                      &dev_priv->mm.deferred_free_list,
1884                                      mm_list)
1885                     i915_gem_free_object_tail(obj);
1886         }
1887
1888         for (i = 0; i < I915_NUM_RINGS; i++)
1889                 i915_gem_retire_requests_ring(&dev_priv->ring[i]);
1890 }
1891
1892 static void
1893 i915_gem_retire_work_handler(struct work_struct *work)
1894 {
1895         drm_i915_private_t *dev_priv;
1896         struct drm_device *dev;
1897         bool idle;
1898         int i;
1899
1900         dev_priv = container_of(work, drm_i915_private_t,
1901                                 mm.retire_work.work);
1902         dev = dev_priv->dev;
1903
1904         /* Come back later if the device is busy... */
1905         if (!mutex_trylock(&dev->struct_mutex)) {
1906                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1907                 return;
1908         }
1909
1910         i915_gem_retire_requests(dev);
1911
1912         /* Send a periodic flush down the ring so we don't hold onto GEM
1913          * objects indefinitely.
1914          */
1915         idle = true;
1916         for (i = 0; i < I915_NUM_RINGS; i++) {
1917                 struct intel_ring_buffer *ring = &dev_priv->ring[i];
1918
1919                 if (!list_empty(&ring->gpu_write_list)) {
1920                         struct drm_i915_gem_request *request;
1921                         int ret;
1922
1923                         ret = i915_gem_flush_ring(ring,
1924                                                   0, I915_GEM_GPU_DOMAINS);
1925                         request = kzalloc(sizeof(*request), GFP_KERNEL);
1926                         if (ret || request == NULL ||
1927                             i915_add_request(ring, NULL, request))
1928                             kfree(request);
1929                 }
1930
1931                 idle &= list_empty(&ring->request_list);
1932         }
1933
1934         if (!dev_priv->mm.suspended && !idle)
1935                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1936
1937         mutex_unlock(&dev->struct_mutex);
1938 }
1939
1940 /**
1941  * Waits for a sequence number to be signaled, and cleans up the
1942  * request and object lists appropriately for that event.
1943  */
1944 int
1945 i915_wait_request(struct intel_ring_buffer *ring,
1946                   uint32_t seqno)
1947 {
1948         drm_i915_private_t *dev_priv = ring->dev->dev_private;
1949         u32 ier;
1950         int ret = 0;
1951
1952         BUG_ON(seqno == 0);
1953
1954         if (atomic_read(&dev_priv->mm.wedged)) {
1955                 struct completion *x = &dev_priv->error_completion;
1956                 bool recovery_complete;
1957                 unsigned long flags;
1958
1959                 /* Give the error handler a chance to run. */
1960                 spin_lock_irqsave(&x->wait.lock, flags);
1961                 recovery_complete = x->done > 0;
1962                 spin_unlock_irqrestore(&x->wait.lock, flags);
1963
1964                 return recovery_complete ? -EIO : -EAGAIN;
1965         }
1966
1967         if (seqno == ring->outstanding_lazy_request) {
1968                 struct drm_i915_gem_request *request;
1969
1970                 request = kzalloc(sizeof(*request), GFP_KERNEL);
1971                 if (request == NULL)
1972                         return -ENOMEM;
1973
1974                 ret = i915_add_request(ring, NULL, request);
1975                 if (ret) {
1976                         kfree(request);
1977                         return ret;
1978                 }
1979
1980                 seqno = request->seqno;
1981         }
1982
1983         if (!i915_seqno_passed(ring->get_seqno(ring), seqno)) {
1984                 if (HAS_PCH_SPLIT(ring->dev))
1985                         ier = I915_READ(DEIER) | I915_READ(GTIER);
1986                 else
1987                         ier = I915_READ(IER);
1988                 if (!ier) {
1989                         DRM_ERROR("something (likely vbetool) disabled "
1990                                   "interrupts, re-enabling\n");
1991                         ring->dev->driver->irq_preinstall(ring->dev);
1992                         ring->dev->driver->irq_postinstall(ring->dev);
1993                 }
1994
1995                 trace_i915_gem_request_wait_begin(ring, seqno);
1996
1997                 ring->waiting_seqno = seqno;
1998                 if (ring->irq_get(ring)) {
1999                         if (dev_priv->mm.interruptible)
2000                                 ret = wait_event_interruptible(ring->irq_queue,
2001                                                                i915_seqno_passed(ring->get_seqno(ring), seqno)
2002                                                                || atomic_read(&dev_priv->mm.wedged));
2003                         else
2004                                 wait_event(ring->irq_queue,
2005                                            i915_seqno_passed(ring->get_seqno(ring), seqno)
2006                                            || atomic_read(&dev_priv->mm.wedged));
2007
2008                         ring->irq_put(ring);
2009                 } else if (wait_for(i915_seqno_passed(ring->get_seqno(ring),
2010                                                       seqno) ||
2011                                     atomic_read(&dev_priv->mm.wedged), 3000))
2012                         ret = -EBUSY;
2013                 ring->waiting_seqno = 0;
2014
2015                 trace_i915_gem_request_wait_end(ring, seqno);
2016         }
2017         if (atomic_read(&dev_priv->mm.wedged))
2018                 ret = -EAGAIN;
2019
2020         if (ret && ret != -ERESTARTSYS)
2021                 DRM_ERROR("%s returns %d (awaiting %d at %d, next %d)\n",
2022                           __func__, ret, seqno, ring->get_seqno(ring),
2023                           dev_priv->next_seqno);
2024
2025         /* Directly dispatch request retiring.  While we have the work queue
2026          * to handle this, the waiter on a request often wants an associated
2027          * buffer to have made it to the inactive list, and we would need
2028          * a separate wait queue to handle that.
2029          */
2030         if (ret == 0)
2031                 i915_gem_retire_requests_ring(ring);
2032
2033         return ret;
2034 }
2035
2036 /**
2037  * Ensures that all rendering to the object has completed and the object is
2038  * safe to unbind from the GTT or access from the CPU.
2039  */
2040 int
2041 i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj)
2042 {
2043         int ret;
2044
2045         /* This function only exists to support waiting for existing rendering,
2046          * not for emitting required flushes.
2047          */
2048         BUG_ON((obj->base.write_domain & I915_GEM_GPU_DOMAINS) != 0);
2049
2050         /* If there is rendering queued on the buffer being evicted, wait for
2051          * it.
2052          */
2053         if (obj->active) {
2054                 ret = i915_wait_request(obj->ring, obj->last_rendering_seqno);
2055                 if (ret)
2056                         return ret;
2057         }
2058
2059         return 0;
2060 }
2061
2062 static void i915_gem_object_finish_gtt(struct drm_i915_gem_object *obj)
2063 {
2064         u32 old_write_domain, old_read_domains;
2065
2066         /* Act a barrier for all accesses through the GTT */
2067         mb();
2068
2069         /* Force a pagefault for domain tracking on next user access */
2070         i915_gem_release_mmap(obj);
2071
2072         if ((obj->base.read_domains & I915_GEM_DOMAIN_GTT) == 0)
2073                 return;
2074
2075         old_read_domains = obj->base.read_domains;
2076         old_write_domain = obj->base.write_domain;
2077
2078         obj->base.read_domains &= ~I915_GEM_DOMAIN_GTT;
2079         obj->base.write_domain &= ~I915_GEM_DOMAIN_GTT;
2080
2081         trace_i915_gem_object_change_domain(obj,
2082                                             old_read_domains,
2083                                             old_write_domain);
2084 }
2085
2086 /**
2087  * Unbinds an object from the GTT aperture.
2088  */
2089 int
2090 i915_gem_object_unbind(struct drm_i915_gem_object *obj)
2091 {
2092         int ret = 0;
2093
2094         if (obj->gtt_space == NULL)
2095                 return 0;
2096
2097         if (obj->pin_count != 0) {
2098                 DRM_ERROR("Attempting to unbind pinned buffer\n");
2099                 return -EINVAL;
2100         }
2101
2102         ret = i915_gem_object_finish_gpu(obj);
2103         if (ret == -ERESTARTSYS)
2104                 return ret;
2105         /* Continue on if we fail due to EIO, the GPU is hung so we
2106          * should be safe and we need to cleanup or else we might
2107          * cause memory corruption through use-after-free.
2108          */
2109
2110         i915_gem_object_finish_gtt(obj);
2111
2112         /* Move the object to the CPU domain to ensure that
2113          * any possible CPU writes while it's not in the GTT
2114          * are flushed when we go to remap it.
2115          */
2116         if (ret == 0)
2117                 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
2118         if (ret == -ERESTARTSYS)
2119                 return ret;
2120         if (ret) {
2121                 /* In the event of a disaster, abandon all caches and
2122                  * hope for the best.
2123                  */
2124                 i915_gem_clflush_object(obj);
2125                 obj->base.read_domains = obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2126         }
2127
2128         /* release the fence reg _after_ flushing */
2129         ret = i915_gem_object_put_fence(obj);
2130         if (ret == -ERESTARTSYS)
2131                 return ret;
2132
2133         trace_i915_gem_object_unbind(obj);
2134
2135         i915_gem_gtt_unbind_object(obj);
2136         i915_gem_object_put_pages_gtt(obj);
2137
2138         list_del_init(&obj->gtt_list);
2139         list_del_init(&obj->mm_list);
2140         /* Avoid an unnecessary call to unbind on rebind. */
2141         obj->map_and_fenceable = true;
2142
2143         drm_mm_put_block(obj->gtt_space);
2144         obj->gtt_space = NULL;
2145         obj->gtt_offset = 0;
2146
2147         if (i915_gem_object_is_purgeable(obj))
2148                 i915_gem_object_truncate(obj);
2149
2150         return ret;
2151 }
2152
2153 int
2154 i915_gem_flush_ring(struct intel_ring_buffer *ring,
2155                     uint32_t invalidate_domains,
2156                     uint32_t flush_domains)
2157 {
2158         int ret;
2159
2160         if (((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) == 0)
2161                 return 0;
2162
2163         trace_i915_gem_ring_flush(ring, invalidate_domains, flush_domains);
2164
2165         ret = ring->flush(ring, invalidate_domains, flush_domains);
2166         if (ret)
2167                 return ret;
2168
2169         if (flush_domains & I915_GEM_GPU_DOMAINS)
2170                 i915_gem_process_flushing_list(ring, flush_domains);
2171
2172         return 0;
2173 }
2174
2175 static int i915_ring_idle(struct intel_ring_buffer *ring)
2176 {
2177         int ret;
2178
2179         if (list_empty(&ring->gpu_write_list) && list_empty(&ring->active_list))
2180                 return 0;
2181
2182         if (!list_empty(&ring->gpu_write_list)) {
2183                 ret = i915_gem_flush_ring(ring,
2184                                     I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2185                 if (ret)
2186                         return ret;
2187         }
2188
2189         return i915_wait_request(ring, i915_gem_next_request_seqno(ring));
2190 }
2191
2192 int
2193 i915_gpu_idle(struct drm_device *dev)
2194 {
2195         drm_i915_private_t *dev_priv = dev->dev_private;
2196         int ret, i;
2197
2198         /* Flush everything onto the inactive list. */
2199         for (i = 0; i < I915_NUM_RINGS; i++) {
2200                 ret = i915_ring_idle(&dev_priv->ring[i]);
2201                 if (ret)
2202                         return ret;
2203         }
2204
2205         return 0;
2206 }
2207
2208 static int sandybridge_write_fence_reg(struct drm_i915_gem_object *obj,
2209                                        struct intel_ring_buffer *pipelined)
2210 {
2211         struct drm_device *dev = obj->base.dev;
2212         drm_i915_private_t *dev_priv = dev->dev_private;
2213         u32 size = obj->gtt_space->size;
2214         int regnum = obj->fence_reg;
2215         uint64_t val;
2216
2217         val = (uint64_t)((obj->gtt_offset + size - 4096) &
2218                          0xfffff000) << 32;
2219         val |= obj->gtt_offset & 0xfffff000;
2220         val |= (uint64_t)((obj->stride / 128) - 1) <<
2221                 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2222
2223         if (obj->tiling_mode == I915_TILING_Y)
2224                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2225         val |= I965_FENCE_REG_VALID;
2226
2227         if (pipelined) {
2228                 int ret = intel_ring_begin(pipelined, 6);
2229                 if (ret)
2230                         return ret;
2231
2232                 intel_ring_emit(pipelined, MI_NOOP);
2233                 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2234                 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8);
2235                 intel_ring_emit(pipelined, (u32)val);
2236                 intel_ring_emit(pipelined, FENCE_REG_SANDYBRIDGE_0 + regnum*8 + 4);
2237                 intel_ring_emit(pipelined, (u32)(val >> 32));
2238                 intel_ring_advance(pipelined);
2239         } else
2240                 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + regnum * 8, val);
2241
2242         return 0;
2243 }
2244
2245 static int i965_write_fence_reg(struct drm_i915_gem_object *obj,
2246                                 struct intel_ring_buffer *pipelined)
2247 {
2248         struct drm_device *dev = obj->base.dev;
2249         drm_i915_private_t *dev_priv = dev->dev_private;
2250         u32 size = obj->gtt_space->size;
2251         int regnum = obj->fence_reg;
2252         uint64_t val;
2253
2254         val = (uint64_t)((obj->gtt_offset + size - 4096) &
2255                     0xfffff000) << 32;
2256         val |= obj->gtt_offset & 0xfffff000;
2257         val |= ((obj->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2258         if (obj->tiling_mode == I915_TILING_Y)
2259                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2260         val |= I965_FENCE_REG_VALID;
2261
2262         if (pipelined) {
2263                 int ret = intel_ring_begin(pipelined, 6);
2264                 if (ret)
2265                         return ret;
2266
2267                 intel_ring_emit(pipelined, MI_NOOP);
2268                 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(2));
2269                 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8);
2270                 intel_ring_emit(pipelined, (u32)val);
2271                 intel_ring_emit(pipelined, FENCE_REG_965_0 + regnum*8 + 4);
2272                 intel_ring_emit(pipelined, (u32)(val >> 32));
2273                 intel_ring_advance(pipelined);
2274         } else
2275                 I915_WRITE64(FENCE_REG_965_0 + regnum * 8, val);
2276
2277         return 0;
2278 }
2279
2280 static int i915_write_fence_reg(struct drm_i915_gem_object *obj,
2281                                 struct intel_ring_buffer *pipelined)
2282 {
2283         struct drm_device *dev = obj->base.dev;
2284         drm_i915_private_t *dev_priv = dev->dev_private;
2285         u32 size = obj->gtt_space->size;
2286         u32 fence_reg, val, pitch_val;
2287         int tile_width;
2288
2289         if (WARN((obj->gtt_offset & ~I915_FENCE_START_MASK) ||
2290                  (size & -size) != size ||
2291                  (obj->gtt_offset & (size - 1)),
2292                  "object 0x%08x [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
2293                  obj->gtt_offset, obj->map_and_fenceable, size))
2294                 return -EINVAL;
2295
2296         if (obj->tiling_mode == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
2297                 tile_width = 128;
2298         else
2299                 tile_width = 512;
2300
2301         /* Note: pitch better be a power of two tile widths */
2302         pitch_val = obj->stride / tile_width;
2303         pitch_val = ffs(pitch_val) - 1;
2304
2305         val = obj->gtt_offset;
2306         if (obj->tiling_mode == I915_TILING_Y)
2307                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2308         val |= I915_FENCE_SIZE_BITS(size);
2309         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2310         val |= I830_FENCE_REG_VALID;
2311
2312         fence_reg = obj->fence_reg;
2313         if (fence_reg < 8)
2314                 fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2315         else
2316                 fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2317
2318         if (pipelined) {
2319                 int ret = intel_ring_begin(pipelined, 4);
2320                 if (ret)
2321                         return ret;
2322
2323                 intel_ring_emit(pipelined, MI_NOOP);
2324                 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2325                 intel_ring_emit(pipelined, fence_reg);
2326                 intel_ring_emit(pipelined, val);
2327                 intel_ring_advance(pipelined);
2328         } else
2329                 I915_WRITE(fence_reg, val);
2330
2331         return 0;
2332 }
2333
2334 static int i830_write_fence_reg(struct drm_i915_gem_object *obj,
2335                                 struct intel_ring_buffer *pipelined)
2336 {
2337         struct drm_device *dev = obj->base.dev;
2338         drm_i915_private_t *dev_priv = dev->dev_private;
2339         u32 size = obj->gtt_space->size;
2340         int regnum = obj->fence_reg;
2341         uint32_t val;
2342         uint32_t pitch_val;
2343
2344         if (WARN((obj->gtt_offset & ~I830_FENCE_START_MASK) ||
2345                  (size & -size) != size ||
2346                  (obj->gtt_offset & (size - 1)),
2347                  "object 0x%08x not 512K or pot-size 0x%08x aligned\n",
2348                  obj->gtt_offset, size))
2349                 return -EINVAL;
2350
2351         pitch_val = obj->stride / 128;
2352         pitch_val = ffs(pitch_val) - 1;
2353
2354         val = obj->gtt_offset;
2355         if (obj->tiling_mode == I915_TILING_Y)
2356                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2357         val |= I830_FENCE_SIZE_BITS(size);
2358         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2359         val |= I830_FENCE_REG_VALID;
2360
2361         if (pipelined) {
2362                 int ret = intel_ring_begin(pipelined, 4);
2363                 if (ret)
2364                         return ret;
2365
2366                 intel_ring_emit(pipelined, MI_NOOP);
2367                 intel_ring_emit(pipelined, MI_LOAD_REGISTER_IMM(1));
2368                 intel_ring_emit(pipelined, FENCE_REG_830_0 + regnum*4);
2369                 intel_ring_emit(pipelined, val);
2370                 intel_ring_advance(pipelined);
2371         } else
2372                 I915_WRITE(FENCE_REG_830_0 + regnum * 4, val);
2373
2374         return 0;
2375 }
2376
2377 static bool ring_passed_seqno(struct intel_ring_buffer *ring, u32 seqno)
2378 {
2379         return i915_seqno_passed(ring->get_seqno(ring), seqno);
2380 }
2381
2382 static int
2383 i915_gem_object_flush_fence(struct drm_i915_gem_object *obj,
2384                             struct intel_ring_buffer *pipelined)
2385 {
2386         int ret;
2387
2388         if (obj->fenced_gpu_access) {
2389                 if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
2390                         ret = i915_gem_flush_ring(obj->last_fenced_ring,
2391                                                   0, obj->base.write_domain);
2392                         if (ret)
2393                                 return ret;
2394                 }
2395
2396                 obj->fenced_gpu_access = false;
2397         }
2398
2399         if (obj->last_fenced_seqno && pipelined != obj->last_fenced_ring) {
2400                 if (!ring_passed_seqno(obj->last_fenced_ring,
2401                                        obj->last_fenced_seqno)) {
2402                         ret = i915_wait_request(obj->last_fenced_ring,
2403                                                 obj->last_fenced_seqno);
2404                         if (ret)
2405                                 return ret;
2406                 }
2407
2408                 obj->last_fenced_seqno = 0;
2409                 obj->last_fenced_ring = NULL;
2410         }
2411
2412         /* Ensure that all CPU reads are completed before installing a fence
2413          * and all writes before removing the fence.
2414          */
2415         if (obj->base.read_domains & I915_GEM_DOMAIN_GTT)
2416                 mb();
2417
2418         return 0;
2419 }
2420
2421 int
2422 i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
2423 {
2424         int ret;
2425
2426         if (obj->tiling_mode)
2427                 i915_gem_release_mmap(obj);
2428
2429         ret = i915_gem_object_flush_fence(obj, NULL);
2430         if (ret)
2431                 return ret;
2432
2433         if (obj->fence_reg != I915_FENCE_REG_NONE) {
2434                 struct drm_i915_private *dev_priv = obj->base.dev->dev_private;
2435                 i915_gem_clear_fence_reg(obj->base.dev,
2436                                          &dev_priv->fence_regs[obj->fence_reg]);
2437
2438                 obj->fence_reg = I915_FENCE_REG_NONE;
2439         }
2440
2441         return 0;
2442 }
2443
2444 static struct drm_i915_fence_reg *
2445 i915_find_fence_reg(struct drm_device *dev,
2446                     struct intel_ring_buffer *pipelined)
2447 {
2448         struct drm_i915_private *dev_priv = dev->dev_private;
2449         struct drm_i915_fence_reg *reg, *first, *avail;
2450         int i;
2451
2452         /* First try to find a free reg */
2453         avail = NULL;
2454         for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2455                 reg = &dev_priv->fence_regs[i];
2456                 if (!reg->obj)
2457                         return reg;
2458
2459                 if (!reg->obj->pin_count)
2460                         avail = reg;
2461         }
2462
2463         if (avail == NULL)
2464                 return NULL;
2465
2466         /* None available, try to steal one or wait for a user to finish */
2467         avail = first = NULL;
2468         list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
2469                 if (reg->obj->pin_count)
2470                         continue;
2471
2472                 if (first == NULL)
2473                         first = reg;
2474
2475                 if (!pipelined ||
2476                     !reg->obj->last_fenced_ring ||
2477                     reg->obj->last_fenced_ring == pipelined) {
2478                         avail = reg;
2479                         break;
2480                 }
2481         }
2482
2483         if (avail == NULL)
2484                 avail = first;
2485
2486         return avail;
2487 }
2488
2489 /**
2490  * i915_gem_object_get_fence - set up a fence reg for an object
2491  * @obj: object to map through a fence reg
2492  * @pipelined: ring on which to queue the change, or NULL for CPU access
2493  * @interruptible: must we wait uninterruptibly for the register to retire?
2494  *
2495  * When mapping objects through the GTT, userspace wants to be able to write
2496  * to them without having to worry about swizzling if the object is tiled.
2497  *
2498  * This function walks the fence regs looking for a free one for @obj,
2499  * stealing one if it can't find any.
2500  *
2501  * It then sets up the reg based on the object's properties: address, pitch
2502  * and tiling format.
2503  */
2504 int
2505 i915_gem_object_get_fence(struct drm_i915_gem_object *obj,
2506                           struct intel_ring_buffer *pipelined)
2507 {
2508         struct drm_device *dev = obj->base.dev;
2509         struct drm_i915_private *dev_priv = dev->dev_private;
2510         struct drm_i915_fence_reg *reg;
2511         int ret;
2512
2513         /* XXX disable pipelining. There are bugs. Shocking. */
2514         pipelined = NULL;
2515
2516         /* Just update our place in the LRU if our fence is getting reused. */
2517         if (obj->fence_reg != I915_FENCE_REG_NONE) {
2518                 reg = &dev_priv->fence_regs[obj->fence_reg];
2519                 list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2520
2521                 if (obj->tiling_changed) {
2522                         ret = i915_gem_object_flush_fence(obj, pipelined);
2523                         if (ret)
2524                                 return ret;
2525
2526                         if (!obj->fenced_gpu_access && !obj->last_fenced_seqno)
2527                                 pipelined = NULL;
2528
2529                         if (pipelined) {
2530                                 reg->setup_seqno =
2531                                         i915_gem_next_request_seqno(pipelined);
2532                                 obj->last_fenced_seqno = reg->setup_seqno;
2533                                 obj->last_fenced_ring = pipelined;
2534                         }
2535
2536                         goto update;
2537                 }
2538
2539                 if (!pipelined) {
2540                         if (reg->setup_seqno) {
2541                                 if (!ring_passed_seqno(obj->last_fenced_ring,
2542                                                        reg->setup_seqno)) {
2543                                         ret = i915_wait_request(obj->last_fenced_ring,
2544                                                                 reg->setup_seqno);
2545                                         if (ret)
2546                                                 return ret;
2547                                 }
2548
2549                                 reg->setup_seqno = 0;
2550                         }
2551                 } else if (obj->last_fenced_ring &&
2552                            obj->last_fenced_ring != pipelined) {
2553                         ret = i915_gem_object_flush_fence(obj, pipelined);
2554                         if (ret)
2555                                 return ret;
2556                 }
2557
2558                 return 0;
2559         }
2560
2561         reg = i915_find_fence_reg(dev, pipelined);
2562         if (reg == NULL)
2563                 return -ENOSPC;
2564
2565         ret = i915_gem_object_flush_fence(obj, pipelined);
2566         if (ret)
2567                 return ret;
2568
2569         if (reg->obj) {
2570                 struct drm_i915_gem_object *old = reg->obj;
2571
2572                 drm_gem_object_reference(&old->base);
2573
2574                 if (old->tiling_mode)
2575                         i915_gem_release_mmap(old);
2576
2577                 ret = i915_gem_object_flush_fence(old, pipelined);
2578                 if (ret) {
2579                         drm_gem_object_unreference(&old->base);
2580                         return ret;
2581                 }
2582
2583                 if (old->last_fenced_seqno == 0 && obj->last_fenced_seqno == 0)
2584                         pipelined = NULL;
2585
2586                 old->fence_reg = I915_FENCE_REG_NONE;
2587                 old->last_fenced_ring = pipelined;
2588                 old->last_fenced_seqno =
2589                         pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2590
2591                 drm_gem_object_unreference(&old->base);
2592         } else if (obj->last_fenced_seqno == 0)
2593                 pipelined = NULL;
2594
2595         reg->obj = obj;
2596         list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2597         obj->fence_reg = reg - dev_priv->fence_regs;
2598         obj->last_fenced_ring = pipelined;
2599
2600         reg->setup_seqno =
2601                 pipelined ? i915_gem_next_request_seqno(pipelined) : 0;
2602         obj->last_fenced_seqno = reg->setup_seqno;
2603
2604 update:
2605         obj->tiling_changed = false;
2606         switch (INTEL_INFO(dev)->gen) {
2607         case 7:
2608         case 6:
2609                 ret = sandybridge_write_fence_reg(obj, pipelined);
2610                 break;
2611         case 5:
2612         case 4:
2613                 ret = i965_write_fence_reg(obj, pipelined);
2614                 break;
2615         case 3:
2616                 ret = i915_write_fence_reg(obj, pipelined);
2617                 break;
2618         case 2:
2619                 ret = i830_write_fence_reg(obj, pipelined);
2620                 break;
2621         }
2622
2623         return ret;
2624 }
2625
2626 /**
2627  * i915_gem_clear_fence_reg - clear out fence register info
2628  * @obj: object to clear
2629  *
2630  * Zeroes out the fence register itself and clears out the associated
2631  * data structures in dev_priv and obj.
2632  */
2633 static void
2634 i915_gem_clear_fence_reg(struct drm_device *dev,
2635                          struct drm_i915_fence_reg *reg)
2636 {
2637         drm_i915_private_t *dev_priv = dev->dev_private;
2638         uint32_t fence_reg = reg - dev_priv->fence_regs;
2639
2640         switch (INTEL_INFO(dev)->gen) {
2641         case 7:
2642         case 6:
2643                 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + fence_reg*8, 0);
2644                 break;
2645         case 5:
2646         case 4:
2647                 I915_WRITE64(FENCE_REG_965_0 + fence_reg*8, 0);
2648                 break;
2649         case 3:
2650                 if (fence_reg >= 8)
2651                         fence_reg = FENCE_REG_945_8 + (fence_reg - 8) * 4;
2652                 else
2653         case 2:
2654                         fence_reg = FENCE_REG_830_0 + fence_reg * 4;
2655
2656                 I915_WRITE(fence_reg, 0);
2657                 break;
2658         }
2659
2660         list_del_init(&reg->lru_list);
2661         reg->obj = NULL;
2662         reg->setup_seqno = 0;
2663 }
2664
2665 /**
2666  * Finds free space in the GTT aperture and binds the object there.
2667  */
2668 static int
2669 i915_gem_object_bind_to_gtt(struct drm_i915_gem_object *obj,
2670                             unsigned alignment,
2671                             bool map_and_fenceable)
2672 {
2673         struct drm_device *dev = obj->base.dev;
2674         drm_i915_private_t *dev_priv = dev->dev_private;
2675         struct drm_mm_node *free_space;
2676         gfp_t gfpmask = __GFP_NORETRY | __GFP_NOWARN;
2677         u32 size, fence_size, fence_alignment, unfenced_alignment;
2678         bool mappable, fenceable;
2679         int ret;
2680
2681         if (obj->madv != I915_MADV_WILLNEED) {
2682                 DRM_ERROR("Attempting to bind a purgeable object\n");
2683                 return -EINVAL;
2684         }
2685
2686         fence_size = i915_gem_get_gtt_size(dev,
2687                                            obj->base.size,
2688                                            obj->tiling_mode);
2689         fence_alignment = i915_gem_get_gtt_alignment(dev,
2690                                                      obj->base.size,
2691                                                      obj->tiling_mode);
2692         unfenced_alignment =
2693                 i915_gem_get_unfenced_gtt_alignment(dev,
2694                                                     obj->base.size,
2695                                                     obj->tiling_mode);
2696
2697         if (alignment == 0)
2698                 alignment = map_and_fenceable ? fence_alignment :
2699                                                 unfenced_alignment;
2700         if (map_and_fenceable && alignment & (fence_alignment - 1)) {
2701                 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2702                 return -EINVAL;
2703         }
2704
2705         size = map_and_fenceable ? fence_size : obj->base.size;
2706
2707         /* If the object is bigger than the entire aperture, reject it early
2708          * before evicting everything in a vain attempt to find space.
2709          */
2710         if (obj->base.size >
2711             (map_and_fenceable ? dev_priv->mm.gtt_mappable_end : dev_priv->mm.gtt_total)) {
2712                 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2713                 return -E2BIG;
2714         }
2715
2716  search_free:
2717         if (map_and_fenceable)
2718                 free_space =
2719                         drm_mm_search_free_in_range(&dev_priv->mm.gtt_space,
2720                                                     size, alignment, 0,
2721                                                     dev_priv->mm.gtt_mappable_end,
2722                                                     0);
2723         else
2724                 free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2725                                                 size, alignment, 0);
2726
2727         if (free_space != NULL) {
2728                 if (map_and_fenceable)
2729                         obj->gtt_space =
2730                                 drm_mm_get_block_range_generic(free_space,
2731                                                                size, alignment, 0,
2732                                                                dev_priv->mm.gtt_mappable_end,
2733                                                                0);
2734                 else
2735                         obj->gtt_space =
2736                                 drm_mm_get_block(free_space, size, alignment);
2737         }
2738         if (obj->gtt_space == NULL) {
2739                 /* If the gtt is empty and we're still having trouble
2740                  * fitting our object in, we're out of memory.
2741                  */
2742                 ret = i915_gem_evict_something(dev, size, alignment,
2743                                                map_and_fenceable);
2744                 if (ret)
2745                         return ret;
2746
2747                 goto search_free;
2748         }
2749
2750         ret = i915_gem_object_get_pages_gtt(obj, gfpmask);
2751         if (ret) {
2752                 drm_mm_put_block(obj->gtt_space);
2753                 obj->gtt_space = NULL;
2754
2755                 if (ret == -ENOMEM) {
2756                         /* first try to reclaim some memory by clearing the GTT */
2757                         ret = i915_gem_evict_everything(dev, false);
2758                         if (ret) {
2759                                 /* now try to shrink everyone else */
2760                                 if (gfpmask) {
2761                                         gfpmask = 0;
2762                                         goto search_free;
2763                                 }
2764
2765                                 return -ENOMEM;
2766                         }
2767
2768                         goto search_free;
2769                 }
2770
2771                 return ret;
2772         }
2773
2774         ret = i915_gem_gtt_bind_object(obj);
2775         if (ret) {
2776                 i915_gem_object_put_pages_gtt(obj);
2777                 drm_mm_put_block(obj->gtt_space);
2778                 obj->gtt_space = NULL;
2779
2780                 if (i915_gem_evict_everything(dev, false))
2781                         return ret;
2782
2783                 goto search_free;
2784         }
2785
2786         list_add_tail(&obj->gtt_list, &dev_priv->mm.gtt_list);
2787         list_add_tail(&obj->mm_list, &dev_priv->mm.inactive_list);
2788
2789         /* Assert that the object is not currently in any GPU domain. As it
2790          * wasn't in the GTT, there shouldn't be any way it could have been in
2791          * a GPU cache
2792          */
2793         BUG_ON(obj->base.read_domains & I915_GEM_GPU_DOMAINS);
2794         BUG_ON(obj->base.write_domain & I915_GEM_GPU_DOMAINS);
2795
2796         obj->gtt_offset = obj->gtt_space->start;
2797
2798         fenceable =
2799                 obj->gtt_space->size == fence_size &&
2800                 (obj->gtt_space->start & (fence_alignment - 1)) == 0;
2801
2802         mappable =
2803                 obj->gtt_offset + obj->base.size <= dev_priv->mm.gtt_mappable_end;
2804
2805         obj->map_and_fenceable = mappable && fenceable;
2806
2807         trace_i915_gem_object_bind(obj, map_and_fenceable);
2808         return 0;
2809 }
2810
2811 void
2812 i915_gem_clflush_object(struct drm_i915_gem_object *obj)
2813 {
2814         /* If we don't have a page list set up, then we're not pinned
2815          * to GPU, and we can ignore the cache flush because it'll happen
2816          * again at bind time.
2817          */
2818         if (obj->pages == NULL)
2819                 return;
2820
2821         /* If the GPU is snooping the contents of the CPU cache,
2822          * we do not need to manually clear the CPU cache lines.  However,
2823          * the caches are only snooped when the render cache is
2824          * flushed/invalidated.  As we always have to emit invalidations
2825          * and flushes when moving into and out of the RENDER domain, correct
2826          * snooping behaviour occurs naturally as the result of our domain
2827          * tracking.
2828          */
2829         if (obj->cache_level != I915_CACHE_NONE)
2830                 return;
2831
2832         trace_i915_gem_object_clflush(obj);
2833
2834         drm_clflush_pages(obj->pages, obj->base.size / PAGE_SIZE);
2835 }
2836
2837 /** Flushes any GPU write domain for the object if it's dirty. */
2838 static int
2839 i915_gem_object_flush_gpu_write_domain(struct drm_i915_gem_object *obj)
2840 {
2841         if ((obj->base.write_domain & I915_GEM_GPU_DOMAINS) == 0)
2842                 return 0;
2843
2844         /* Queue the GPU write cache flushing we need. */
2845         return i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
2846 }
2847
2848 /** Flushes the GTT write domain for the object if it's dirty. */
2849 static void
2850 i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj)
2851 {
2852         uint32_t old_write_domain;
2853
2854         if (obj->base.write_domain != I915_GEM_DOMAIN_GTT)
2855                 return;
2856
2857         /* No actual flushing is required for the GTT write domain.  Writes
2858          * to it immediately go to main memory as far as we know, so there's
2859          * no chipset flush.  It also doesn't land in render cache.
2860          *
2861          * However, we do have to enforce the order so that all writes through
2862          * the GTT land before any writes to the device, such as updates to
2863          * the GATT itself.
2864          */
2865         wmb();
2866
2867         old_write_domain = obj->base.write_domain;
2868         obj->base.write_domain = 0;
2869
2870         trace_i915_gem_object_change_domain(obj,
2871                                             obj->base.read_domains,
2872                                             old_write_domain);
2873 }
2874
2875 /** Flushes the CPU write domain for the object if it's dirty. */
2876 static void
2877 i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj)
2878 {
2879         uint32_t old_write_domain;
2880
2881         if (obj->base.write_domain != I915_GEM_DOMAIN_CPU)
2882                 return;
2883
2884         i915_gem_clflush_object(obj);
2885         intel_gtt_chipset_flush();
2886         old_write_domain = obj->base.write_domain;
2887         obj->base.write_domain = 0;
2888
2889         trace_i915_gem_object_change_domain(obj,
2890                                             obj->base.read_domains,
2891                                             old_write_domain);
2892 }
2893
2894 /**
2895  * Moves a single object to the GTT read, and possibly write domain.
2896  *
2897  * This function returns when the move is complete, including waiting on
2898  * flushes to occur.
2899  */
2900 int
2901 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
2902 {
2903         uint32_t old_write_domain, old_read_domains;
2904         int ret;
2905
2906         /* Not valid to be called on unbound objects. */
2907         if (obj->gtt_space == NULL)
2908                 return -EINVAL;
2909
2910         if (obj->base.write_domain == I915_GEM_DOMAIN_GTT)
2911                 return 0;
2912
2913         ret = i915_gem_object_flush_gpu_write_domain(obj);
2914         if (ret)
2915                 return ret;
2916
2917         if (obj->pending_gpu_write || write) {
2918                 ret = i915_gem_object_wait_rendering(obj);
2919                 if (ret)
2920                         return ret;
2921         }
2922
2923         i915_gem_object_flush_cpu_write_domain(obj);
2924
2925         old_write_domain = obj->base.write_domain;
2926         old_read_domains = obj->base.read_domains;
2927
2928         /* It should now be out of any other write domains, and we can update
2929          * the domain values for our changes.
2930          */
2931         BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2932         obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
2933         if (write) {
2934                 obj->base.read_domains = I915_GEM_DOMAIN_GTT;
2935                 obj->base.write_domain = I915_GEM_DOMAIN_GTT;
2936                 obj->dirty = 1;
2937         }
2938
2939         trace_i915_gem_object_change_domain(obj,
2940                                             old_read_domains,
2941                                             old_write_domain);
2942
2943         return 0;
2944 }
2945
2946 int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
2947                                     enum i915_cache_level cache_level)
2948 {
2949         int ret;
2950
2951         if (obj->cache_level == cache_level)
2952                 return 0;
2953
2954         if (obj->pin_count) {
2955                 DRM_DEBUG("can not change the cache level of pinned objects\n");
2956                 return -EBUSY;
2957         }
2958
2959         if (obj->gtt_space) {
2960                 ret = i915_gem_object_finish_gpu(obj);
2961                 if (ret)
2962                         return ret;
2963
2964                 i915_gem_object_finish_gtt(obj);
2965
2966                 /* Before SandyBridge, you could not use tiling or fence
2967                  * registers with snooped memory, so relinquish any fences
2968                  * currently pointing to our region in the aperture.
2969                  */
2970                 if (INTEL_INFO(obj->base.dev)->gen < 6) {
2971                         ret = i915_gem_object_put_fence(obj);
2972                         if (ret)
2973                                 return ret;
2974                 }
2975
2976                 i915_gem_gtt_rebind_object(obj, cache_level);
2977         }
2978
2979         if (cache_level == I915_CACHE_NONE) {
2980                 u32 old_read_domains, old_write_domain;
2981
2982                 /* If we're coming from LLC cached, then we haven't
2983                  * actually been tracking whether the data is in the
2984                  * CPU cache or not, since we only allow one bit set
2985                  * in obj->write_domain and have been skipping the clflushes.
2986                  * Just set it to the CPU cache for now.
2987                  */
2988                 WARN_ON(obj->base.write_domain & ~I915_GEM_DOMAIN_CPU);
2989                 WARN_ON(obj->base.read_domains & ~I915_GEM_DOMAIN_CPU);
2990
2991                 old_read_domains = obj->base.read_domains;
2992                 old_write_domain = obj->base.write_domain;
2993
2994                 obj->base.read_domains = I915_GEM_DOMAIN_CPU;
2995                 obj->base.write_domain = I915_GEM_DOMAIN_CPU;
2996
2997                 trace_i915_gem_object_change_domain(obj,
2998                                                     old_read_domains,
2999                                                     old_write_domain);
3000         }
3001
3002         obj->cache_level = cache_level;
3003         return 0;
3004 }
3005
3006 /*
3007  * Prepare buffer for display plane (scanout, cursors, etc).
3008  * Can be called from an uninterruptible phase (modesetting) and allows
3009  * any flushes to be pipelined (for pageflips).
3010  *
3011  * For the display plane, we want to be in the GTT but out of any write
3012  * domains. So in many ways this looks like set_to_gtt_domain() apart from the
3013  * ability to pipeline the waits, pinning and any additional subtleties
3014  * that may differentiate the display plane from ordinary buffers.
3015  */
3016 int
3017 i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
3018                                      u32 alignment,
3019                                      struct intel_ring_buffer *pipelined)
3020 {
3021         u32 old_read_domains, old_write_domain;
3022         int ret;
3023
3024         ret = i915_gem_object_flush_gpu_write_domain(obj);
3025         if (ret)
3026                 return ret;
3027
3028         if (pipelined != obj->ring) {
3029                 ret = i915_gem_object_wait_rendering(obj);
3030                 if (ret == -ERESTARTSYS)
3031                         return ret;
3032         }
3033
3034         /* The display engine is not coherent with the LLC cache on gen6.  As
3035          * a result, we make sure that the pinning that is about to occur is
3036          * done with uncached PTEs. This is lowest common denominator for all
3037          * chipsets.
3038          *
3039          * However for gen6+, we could do better by using the GFDT bit instead
3040          * of uncaching, which would allow us to flush all the LLC-cached data
3041          * with that bit in the PTE to main memory with just one PIPE_CONTROL.
3042          */
3043         ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
3044         if (ret)
3045                 return ret;
3046
3047         /* As the user may map the buffer once pinned in the display plane
3048          * (e.g. libkms for the bootup splash), we have to ensure that we
3049          * always use map_and_fenceable for all scanout buffers.
3050          */
3051         ret = i915_gem_object_pin(obj, alignment, true);
3052         if (ret)
3053                 return ret;
3054
3055         i915_gem_object_flush_cpu_write_domain(obj);
3056
3057         old_write_domain = obj->base.write_domain;
3058         old_read_domains = obj->base.read_domains;
3059
3060         /* It should now be out of any other write domains, and we can update
3061          * the domain values for our changes.
3062          */
3063         BUG_ON((obj->base.write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
3064         obj->base.read_domains |= I915_GEM_DOMAIN_GTT;
3065
3066         trace_i915_gem_object_change_domain(obj,
3067                                             old_read_domains,
3068                                             old_write_domain);
3069
3070         return 0;
3071 }
3072
3073 int
3074 i915_gem_object_finish_gpu(struct drm_i915_gem_object *obj)
3075 {
3076         int ret;
3077
3078         if ((obj->base.read_domains & I915_GEM_GPU_DOMAINS) == 0)
3079                 return 0;
3080
3081         if (obj->base.write_domain & I915_GEM_GPU_DOMAINS) {
3082                 ret = i915_gem_flush_ring(obj->ring, 0, obj->base.write_domain);
3083                 if (ret)
3084                         return ret;
3085         }
3086
3087         /* Ensure that we invalidate the GPU's caches and TLBs. */
3088         obj->base.read_domains &= ~I915_GEM_GPU_DOMAINS;
3089
3090         return i915_gem_object_wait_rendering(obj);
3091 }
3092
3093 /**
3094  * Moves a single object to the CPU read, and possibly write domain.