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 <linux/swap.h>
  33 #include <linux/pci.h>
  34
  35 #define I915_GEM_GPU_DOMAINS    (~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT))
  36
  37 static void i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj);
  38 static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj);
  39 static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj);
  40 static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj,
  41                                              int write);
  42 static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
  43                                                      uint64_t offset,
  44                                                      uint64_t size);
  45 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj);
  46 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
  47 static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj,
  48                                            unsigned alignment);
  49 static int i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write);
  50 static void i915_gem_clear_fence_reg(struct drm_gem_object *obj);
  51 static int i915_gem_evict_something(struct drm_device *dev);
  52 static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
  53                                 struct drm_i915_gem_pwrite *args,
  54                                 struct drm_file *file_priv);
  55
  56 int i915_gem_do_init(struct drm_device *dev, unsigned long start,
  57                      unsigned long end)
  58 {
  59         drm_i915_private_t *dev_priv = dev->dev_private;
  60
  61         if (start >= end ||
  62             (start & (PAGE_SIZE - 1)) != 0 ||
  63             (end & (PAGE_SIZE - 1)) != 0) {
  64                 return -EINVAL;
  65         }
  66
  67         drm_mm_init(&dev_priv->mm.gtt_space, start,
  68                     end - start);
  69
  70         dev->gtt_total = (uint32_t) (end - start);
  71
  72         return 0;
  73 }
  74
  75 int
  76 i915_gem_init_ioctl(struct drm_device *dev, void *data,
  77                     struct drm_file *file_priv)
  78 {
  79         struct drm_i915_gem_init *args = data;
  80         int ret;
  81
  82         mutex_lock(&dev->struct_mutex);
  83         ret = i915_gem_do_init(dev, args->gtt_start, args->gtt_end);
  84         mutex_unlock(&dev->struct_mutex);
  85
  86         return ret;
  87 }
  88
  89 int
  90 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
  91                             struct drm_file *file_priv)
  92 {
  93         struct drm_i915_gem_get_aperture *args = data;
  94
  95         if (!(dev->driver->driver_features & DRIVER_GEM))
  96                 return -ENODEV;
  97
  98         args->aper_size = dev->gtt_total;
  99         args->aper_available_size = (args->aper_size -
 100                                      atomic_read(&dev->pin_memory));
 101
 102         return 0;
 103 }
 104
 105
 106 /**
 107  * Creates a new mm object and returns a handle to it.
 108  */
 109 int
 110 i915_gem_create_ioctl(struct drm_device *dev, void *data,
 111                       struct drm_file *file_priv)
 112 {
 113         struct drm_i915_gem_create *args = data;
 114         struct drm_gem_object *obj;
 115         int handle, ret;
 116
 117         args->size = roundup(args->size, PAGE_SIZE);
 118
 119         /* Allocate the new object */
 120         obj = drm_gem_object_alloc(dev, args->size);
 121         if (obj == NULL)
 122                 return -ENOMEM;
 123
 124         ret = drm_gem_handle_create(file_priv, obj, &handle);
 125         mutex_lock(&dev->struct_mutex);
 126         drm_gem_object_handle_unreference(obj);
 127         mutex_unlock(&dev->struct_mutex);
 128
 129         if (ret)
 130                 return ret;
 131
 132         args->handle = handle;
 133
 134         return 0;
 135 }
 136
 137 static inline int
 138 fast_shmem_read(struct page **pages,
 139                 loff_t page_base, int page_offset,
 140                 char __user *data,
 141                 int length)
 142 {
 143         char __iomem *vaddr;
 144         int unwritten;
 145
 146         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
 147         if (vaddr == NULL)
 148                 return -ENOMEM;
 149         unwritten = __copy_to_user_inatomic(data, vaddr + page_offset, length);
 150         kunmap_atomic(vaddr, KM_USER0);
 151
 152         if (unwritten)
 153                 return -EFAULT;
 154
 155         return 0;
 156 }
 157
 158 static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj)
 159 {
 160         drm_i915_private_t *dev_priv = obj->dev->dev_private;
 161         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 162
 163         return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
 164                 obj_priv->tiling_mode != I915_TILING_NONE;
 165 }
 166
 167 static inline int
 168 slow_shmem_copy(struct page *dst_page,
 169                 int dst_offset,
 170                 struct page *src_page,
 171                 int src_offset,
 172                 int length)
 173 {
 174         char *dst_vaddr, *src_vaddr;
 175
 176         dst_vaddr = kmap_atomic(dst_page, KM_USER0);
 177         if (dst_vaddr == NULL)
 178                 return -ENOMEM;
 179
 180         src_vaddr = kmap_atomic(src_page, KM_USER1);
 181         if (src_vaddr == NULL) {
 182                 kunmap_atomic(dst_vaddr, KM_USER0);
 183                 return -ENOMEM;
 184         }
 185
 186         memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
 187
 188         kunmap_atomic(src_vaddr, KM_USER1);
 189         kunmap_atomic(dst_vaddr, KM_USER0);
 190
 191         return 0;
 192 }
 193
 194 static inline int
 195 slow_shmem_bit17_copy(struct page *gpu_page,
 196                       int gpu_offset,
 197                       struct page *cpu_page,
 198                       int cpu_offset,
 199                       int length,
 200                       int is_read)
 201 {
 202         char *gpu_vaddr, *cpu_vaddr;
 203
 204         /* Use the unswizzled path if this page isn't affected. */
 205         if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
 206                 if (is_read)
 207                         return slow_shmem_copy(cpu_page, cpu_offset,
 208                                                gpu_page, gpu_offset, length);
 209                 else
 210                         return slow_shmem_copy(gpu_page, gpu_offset,
 211                                                cpu_page, cpu_offset, length);
 212         }
 213
 214         gpu_vaddr = kmap_atomic(gpu_page, KM_USER0);
 215         if (gpu_vaddr == NULL)
 216                 return -ENOMEM;
 217
 218         cpu_vaddr = kmap_atomic(cpu_page, KM_USER1);
 219         if (cpu_vaddr == NULL) {
 220                 kunmap_atomic(gpu_vaddr, KM_USER0);
 221                 return -ENOMEM;
 222         }
 223
 224         /* Copy the data, XORing A6 with A17 (1). The user already knows he's
 225          * XORing with the other bits (A9 for Y, A9 and A10 for X)
 226          */
 227         while (length > 0) {
 228                 int cacheline_end = ALIGN(gpu_offset + 1, 64);
 229                 int this_length = min(cacheline_end - gpu_offset, length);
 230                 int swizzled_gpu_offset = gpu_offset ^ 64;
 231
 232                 if (is_read) {
 233                         memcpy(cpu_vaddr + cpu_offset,
 234                                gpu_vaddr + swizzled_gpu_offset,
 235                                this_length);
 236                 } else {
 237                         memcpy(gpu_vaddr + swizzled_gpu_offset,
 238                                cpu_vaddr + cpu_offset,
 239                                this_length);
 240                 }
 241                 cpu_offset += this_length;
 242                 gpu_offset += this_length;
 243                 length -= this_length;
 244         }
 245
 246         kunmap_atomic(cpu_vaddr, KM_USER1);
 247         kunmap_atomic(gpu_vaddr, KM_USER0);
 248
 249         return 0;
 250 }
 251
 252 /**
 253  * This is the fast shmem pread path, which attempts to copy_from_user directly
 254  * from the backing pages of the object to the user's address space.  On a
 255  * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
 256  */
 257 static int
 258 i915_gem_shmem_pread_fast(struct drm_device *dev, struct drm_gem_object *obj,
 259                           struct drm_i915_gem_pread *args,
 260                           struct drm_file *file_priv)
 261 {
 262         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 263         ssize_t remain;
 264         loff_t offset, page_base;
 265         char __user *user_data;
 266         int page_offset, page_length;
 267         int ret;
 268
 269         user_data = (char __user *) (uintptr_t) args->data_ptr;
 270         remain = args->size;
 271
 272         mutex_lock(&dev->struct_mutex);
 273
 274         ret = i915_gem_object_get_pages(obj);
 275         if (ret != 0)
 276                 goto fail_unlock;
 277
 278         ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
 279                                                         args->size);
 280         if (ret != 0)
 281                 goto fail_put_pages;
 282
 283         obj_priv = obj->driver_private;
 284         offset = args->offset;
 285
 286         while (remain > 0) {
 287                 /* Operation in this page
 288                  *
 289                  * page_base = page offset within aperture
 290                  * page_offset = offset within page
 291                  * page_length = bytes to copy for this page
 292                  */
 293                 page_base = (offset & ~(PAGE_SIZE-1));
 294                 page_offset = offset & (PAGE_SIZE-1);
 295                 page_length = remain;
 296                 if ((page_offset + remain) > PAGE_SIZE)
 297                         page_length = PAGE_SIZE - page_offset;
 298
 299                 ret = fast_shmem_read(obj_priv->pages,
 300                                       page_base, page_offset,
 301                                       user_data, page_length);
 302                 if (ret)
 303                         goto fail_put_pages;
 304
 305                 remain -= page_length;
 306                 user_data += page_length;
 307                 offset += page_length;
 308         }
 309
 310 fail_put_pages:
 311         i915_gem_object_put_pages(obj);
 312 fail_unlock:
 313         mutex_unlock(&dev->struct_mutex);
 314
 315         return ret;
 316 }
 317
 318 /**
 319  * This is the fallback shmem pread path, which allocates temporary storage
 320  * in kernel space to copy_to_user into outside of the struct_mutex, so we
 321  * can copy out of the object's backing pages while holding the struct mutex
 322  * and not take page faults.
 323  */
 324 static int
 325 i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj,
 326                           struct drm_i915_gem_pread *args,
 327                           struct drm_file *file_priv)
 328 {
 329         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 330         struct mm_struct *mm = current->mm;
 331         struct page **user_pages;
 332         ssize_t remain;
 333         loff_t offset, pinned_pages, i;
 334         loff_t first_data_page, last_data_page, num_pages;
 335         int shmem_page_index, shmem_page_offset;
 336         int data_page_index,  data_page_offset;
 337         int page_length;
 338         int ret;
 339         uint64_t data_ptr = args->data_ptr;
 340         int do_bit17_swizzling;
 341
 342         remain = args->size;
 343
 344         /* Pin the user pages containing the data.  We can't fault while
 345          * holding the struct mutex, yet we want to hold it while
 346          * dereferencing the user data.
 347          */
 348         first_data_page = data_ptr / PAGE_SIZE;
 349         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 350         num_pages = last_data_page - first_data_page + 1;
 351
 352         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
 353         if (user_pages == NULL)
 354                 return -ENOMEM;
 355
 356         down_read(&mm->mmap_sem);
 357         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 358                                       num_pages, 1, 0, user_pages, NULL);
 359         up_read(&mm->mmap_sem);
 360         if (pinned_pages < num_pages) {
 361                 ret = -EFAULT;
 362                 goto fail_put_user_pages;
 363         }
 364
 365         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 366
 367         mutex_lock(&dev->struct_mutex);
 368
 369         ret = i915_gem_object_get_pages(obj);
 370         if (ret != 0)
 371                 goto fail_unlock;
 372
 373         ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
 374                                                         args->size);
 375         if (ret != 0)
 376                 goto fail_put_pages;
 377
 378         obj_priv = obj->driver_private;
 379         offset = args->offset;
 380
 381         while (remain > 0) {
 382                 /* Operation in this page
 383                  *
 384                  * shmem_page_index = page number within shmem file
 385                  * shmem_page_offset = offset within page in shmem file
 386                  * data_page_index = page number in get_user_pages return
 387                  * data_page_offset = offset with data_page_index page.
 388                  * page_length = bytes to copy for this page
 389                  */
 390                 shmem_page_index = offset / PAGE_SIZE;
 391                 shmem_page_offset = offset & ~PAGE_MASK;
 392                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 393                 data_page_offset = data_ptr & ~PAGE_MASK;
 394
 395                 page_length = remain;
 396                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 397                         page_length = PAGE_SIZE - shmem_page_offset;
 398                 if ((data_page_offset + page_length) > PAGE_SIZE)
 399                         page_length = PAGE_SIZE - data_page_offset;
 400
 401                 if (do_bit17_swizzling) {
 402                         ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
 403                                                     shmem_page_offset,
 404                                                     user_pages[data_page_index],
 405                                                     data_page_offset,
 406                                                     page_length,
 407                                                     1);
 408                 } else {
 409                         ret = slow_shmem_copy(user_pages[data_page_index],
 410                                               data_page_offset,
 411                                               obj_priv->pages[shmem_page_index],
 412                                               shmem_page_offset,
 413                                               page_length);
 414                 }
 415                 if (ret)
 416                         goto fail_put_pages;
 417
 418                 remain -= page_length;
 419                 data_ptr += page_length;
 420                 offset += page_length;
 421         }
 422
 423 fail_put_pages:
 424         i915_gem_object_put_pages(obj);
 425 fail_unlock:
 426         mutex_unlock(&dev->struct_mutex);
 427 fail_put_user_pages:
 428         for (i = 0; i < pinned_pages; i++) {
 429                 SetPageDirty(user_pages[i]);
 430                 page_cache_release(user_pages[i]);
 431         }
 432         drm_free_large(user_pages);
 433
 434         return ret;
 435 }
 436
 437 /**
 438  * Reads data from the object referenced by handle.
 439  *
 440  * On error, the contents of *data are undefined.
 441  */
 442 int
 443 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 444                      struct drm_file *file_priv)
 445 {
 446         struct drm_i915_gem_pread *args = data;
 447         struct drm_gem_object *obj;
 448         struct drm_i915_gem_object *obj_priv;
 449         int ret;
 450
 451         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
 452         if (obj == NULL)
 453                 return -EBADF;
 454         obj_priv = obj->driver_private;
 455
 456         /* Bounds check source.
 457          *
 458          * XXX: This could use review for overflow issues...
 459          */
 460         if (args->offset > obj->size || args->size > obj->size ||
 461             args->offset + args->size > obj->size) {
 462                 drm_gem_object_unreference(obj);
 463                 return -EINVAL;
 464         }
 465
 466         if (i915_gem_object_needs_bit17_swizzle(obj)) {
 467                 ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv);
 468         } else {
 469                 ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv);
 470                 if (ret != 0)
 471                         ret = i915_gem_shmem_pread_slow(dev, obj, args,
 472                                                         file_priv);
 473         }
 474
 475         drm_gem_object_unreference(obj);
 476
 477         return ret;
 478 }
 479
 480 /* This is the fast write path which cannot handle
 481  * page faults in the source data
 482  */
 483
 484 static inline int
 485 fast_user_write(struct io_mapping *mapping,
 486                 loff_t page_base, int page_offset,
 487                 char __user *user_data,
 488                 int length)
 489 {
 490         char *vaddr_atomic;
 491         unsigned long unwritten;
 492
 493         vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
 494         unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
 495                                                       user_data, length);
 496         io_mapping_unmap_atomic(vaddr_atomic);
 497         if (unwritten)
 498                 return -EFAULT;
 499         return 0;
 500 }
 501
 502 /* Here's the write path which can sleep for
 503  * page faults
 504  */
 505
 506 static inline int
 507 slow_kernel_write(struct io_mapping *mapping,
 508                   loff_t gtt_base, int gtt_offset,
 509                   struct page *user_page, int user_offset,
 510                   int length)
 511 {
 512         char *src_vaddr, *dst_vaddr;
 513         unsigned long unwritten;
 514
 515         dst_vaddr = io_mapping_map_atomic_wc(mapping, gtt_base);
 516         src_vaddr = kmap_atomic(user_page, KM_USER1);
 517         unwritten = __copy_from_user_inatomic_nocache(dst_vaddr + gtt_offset,
 518                                                       src_vaddr + user_offset,
 519                                                       length);
 520         kunmap_atomic(src_vaddr, KM_USER1);
 521         io_mapping_unmap_atomic(dst_vaddr);
 522         if (unwritten)
 523                 return -EFAULT;
 524         return 0;
 525 }
 526
 527 static inline int
 528 fast_shmem_write(struct page **pages,
 529                  loff_t page_base, int page_offset,
 530                  char __user *data,
 531                  int length)
 532 {
 533         char __iomem *vaddr;
 534         unsigned long unwritten;
 535
 536         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
 537         if (vaddr == NULL)
 538                 return -ENOMEM;
 539         unwritten = __copy_from_user_inatomic(vaddr + page_offset, data, length);
 540         kunmap_atomic(vaddr, KM_USER0);
 541
 542         if (unwritten)
 543                 return -EFAULT;
 544         return 0;
 545 }
 546
 547 /**
 548  * This is the fast pwrite path, where we copy the data directly from the
 549  * user into the GTT, uncached.
 550  */
 551 static int
 552 i915_gem_gtt_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
 553                          struct drm_i915_gem_pwrite *args,
 554                          struct drm_file *file_priv)
 555 {
 556         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 557         drm_i915_private_t *dev_priv = dev->dev_private;
 558         ssize_t remain;
 559         loff_t offset, page_base;
 560         char __user *user_data;
 561         int page_offset, page_length;
 562         int ret;
 563
 564         user_data = (char __user *) (uintptr_t) args->data_ptr;
 565         remain = args->size;
 566         if (!access_ok(VERIFY_READ, user_data, remain))
 567                 return -EFAULT;
 568
 569
 570         mutex_lock(&dev->struct_mutex);
 571         ret = i915_gem_object_pin(obj, 0);
 572         if (ret) {
 573                 mutex_unlock(&dev->struct_mutex);
 574                 return ret;
 575         }
 576         ret = i915_gem_object_set_to_gtt_domain(obj, 1);
 577         if (ret)
 578                 goto fail;
 579
 580         obj_priv = obj->driver_private;
 581         offset = obj_priv->gtt_offset + args->offset;
 582
 583         while (remain > 0) {
 584                 /* Operation in this page
 585                  *
 586                  * page_base = page offset within aperture
 587                  * page_offset = offset within page
 588                  * page_length = bytes to copy for this page
 589                  */
 590                 page_base = (offset & ~(PAGE_SIZE-1));
 591                 page_offset = offset & (PAGE_SIZE-1);
 592                 page_length = remain;
 593                 if ((page_offset + remain) > PAGE_SIZE)
 594                         page_length = PAGE_SIZE - page_offset;
 595
 596                 ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
 597                                        page_offset, user_data, page_length);
 598
 599                 /* If we get a fault while copying data, then (presumably) our
 600                  * source page isn't available.  Return the error and we'll
 601                  * retry in the slow path.
 602                  */
 603                 if (ret)
 604                         goto fail;
 605
 606                 remain -= page_length;
 607                 user_data += page_length;
 608                 offset += page_length;
 609         }
 610
 611 fail:
 612         i915_gem_object_unpin(obj);
 613         mutex_unlock(&dev->struct_mutex);
 614
 615         return ret;
 616 }
 617
 618 /**
 619  * This is the fallback GTT pwrite path, which uses get_user_pages to pin
 620  * the memory and maps it using kmap_atomic for copying.
 621  *
 622  * This code resulted in x11perf -rgb10text consuming about 10% more CPU
 623  * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
 624  */
 625 static int
 626 i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
 627                          struct drm_i915_gem_pwrite *args,
 628                          struct drm_file *file_priv)
 629 {
 630         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 631         drm_i915_private_t *dev_priv = dev->dev_private;
 632         ssize_t remain;
 633         loff_t gtt_page_base, offset;
 634         loff_t first_data_page, last_data_page, num_pages;
 635         loff_t pinned_pages, i;
 636         struct page **user_pages;
 637         struct mm_struct *mm = current->mm;
 638         int gtt_page_offset, data_page_offset, data_page_index, page_length;
 639         int ret;
 640         uint64_t data_ptr = args->data_ptr;
 641
 642         remain = args->size;
 643
 644         /* Pin the user pages containing the data.  We can't fault while
 645          * holding the struct mutex, and all of the pwrite implementations
 646          * want to hold it while dereferencing the user data.
 647          */
 648         first_data_page = data_ptr / PAGE_SIZE;
 649         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 650         num_pages = last_data_page - first_data_page + 1;
 651
 652         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
 653         if (user_pages == NULL)
 654                 return -ENOMEM;
 655
 656         down_read(&mm->mmap_sem);
 657         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 658                                       num_pages, 0, 0, user_pages, NULL);
 659         up_read(&mm->mmap_sem);
 660         if (pinned_pages < num_pages) {
 661                 ret = -EFAULT;
 662                 goto out_unpin_pages;
 663         }
 664
 665         mutex_lock(&dev->struct_mutex);
 666         ret = i915_gem_object_pin(obj, 0);
 667         if (ret)
 668                 goto out_unlock;
 669
 670         ret = i915_gem_object_set_to_gtt_domain(obj, 1);
 671         if (ret)
 672                 goto out_unpin_object;
 673
 674         obj_priv = obj->driver_private;
 675         offset = obj_priv->gtt_offset + args->offset;
 676
 677         while (remain > 0) {
 678                 /* Operation in this page
 679                  *
 680                  * gtt_page_base = page offset within aperture
 681                  * gtt_page_offset = offset within page in aperture
 682                  * data_page_index = page number in get_user_pages return
 683                  * data_page_offset = offset with data_page_index page.
 684                  * page_length = bytes to copy for this page
 685                  */
 686                 gtt_page_base = offset & PAGE_MASK;
 687                 gtt_page_offset = offset & ~PAGE_MASK;
 688                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 689                 data_page_offset = data_ptr & ~PAGE_MASK;
 690
 691                 page_length = remain;
 692                 if ((gtt_page_offset + page_length) > PAGE_SIZE)
 693                         page_length = PAGE_SIZE - gtt_page_offset;
 694                 if ((data_page_offset + page_length) > PAGE_SIZE)
 695                         page_length = PAGE_SIZE - data_page_offset;
 696
 697                 ret = slow_kernel_write(dev_priv->mm.gtt_mapping,
 698                                         gtt_page_base, gtt_page_offset,
 699                                         user_pages[data_page_index],
 700                                         data_page_offset,
 701                                         page_length);
 702
 703                 /* If we get a fault while copying data, then (presumably) our
 704                  * source page isn't available.  Return the error and we'll
 705                  * retry in the slow path.
 706                  */
 707                 if (ret)
 708                         goto out_unpin_object;
 709
 710                 remain -= page_length;
 711                 offset += page_length;
 712                 data_ptr += page_length;
 713         }
 714
 715 out_unpin_object:
 716         i915_gem_object_unpin(obj);
 717 out_unlock:
 718         mutex_unlock(&dev->struct_mutex);
 719 out_unpin_pages:
 720         for (i = 0; i < pinned_pages; i++)
 721                 page_cache_release(user_pages[i]);
 722         drm_free_large(user_pages);
 723
 724         return ret;
 725 }
 726
 727 /**
 728  * This is the fast shmem pwrite path, which attempts to directly
 729  * copy_from_user into the kmapped pages backing the object.
 730  */
 731 static int
 732 i915_gem_shmem_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
 733                            struct drm_i915_gem_pwrite *args,
 734                            struct drm_file *file_priv)
 735 {
 736         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 737         ssize_t remain;
 738         loff_t offset, page_base;
 739         char __user *user_data;
 740         int page_offset, page_length;
 741         int ret;
 742
 743         user_data = (char __user *) (uintptr_t) args->data_ptr;
 744         remain = args->size;
 745
 746         mutex_lock(&dev->struct_mutex);
 747
 748         ret = i915_gem_object_get_pages(obj);
 749         if (ret != 0)
 750                 goto fail_unlock;
 751
 752         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
 753         if (ret != 0)
 754                 goto fail_put_pages;
 755
 756         obj_priv = obj->driver_private;
 757         offset = args->offset;
 758         obj_priv->dirty = 1;
 759
 760         while (remain > 0) {
 761                 /* Operation in this page
 762                  *
 763                  * page_base = page offset within aperture
 764                  * page_offset = offset within page
 765                  * page_length = bytes to copy for this page
 766                  */
 767                 page_base = (offset & ~(PAGE_SIZE-1));
 768                 page_offset = offset & (PAGE_SIZE-1);
 769                 page_length = remain;
 770                 if ((page_offset + remain) > PAGE_SIZE)
 771                         page_length = PAGE_SIZE - page_offset;
 772
 773                 ret = fast_shmem_write(obj_priv->pages,
 774                                        page_base, page_offset,
 775                                        user_data, page_length);
 776                 if (ret)
 777                         goto fail_put_pages;
 778
 779                 remain -= page_length;
 780                 user_data += page_length;
 781                 offset += page_length;
 782         }
 783
 784 fail_put_pages:
 785         i915_gem_object_put_pages(obj);
 786 fail_unlock:
 787         mutex_unlock(&dev->struct_mutex);
 788
 789         return ret;
 790 }
 791
 792 /**
 793  * This is the fallback shmem pwrite path, which uses get_user_pages to pin
 794  * the memory and maps it using kmap_atomic for copying.
 795  *
 796  * This avoids taking mmap_sem for faulting on the user's address while the
 797  * struct_mutex is held.
 798  */
 799 static int
 800 i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
 801                            struct drm_i915_gem_pwrite *args,
 802                            struct drm_file *file_priv)
 803 {
 804         struct drm_i915_gem_object *obj_priv = obj->driver_private;
 805         struct mm_struct *mm = current->mm;
 806         struct page **user_pages;
 807         ssize_t remain;
 808         loff_t offset, pinned_pages, i;
 809         loff_t first_data_page, last_data_page, num_pages;
 810         int shmem_page_index, shmem_page_offset;
 811         int data_page_index,  data_page_offset;
 812         int page_length;
 813         int ret;
 814         uint64_t data_ptr = args->data_ptr;
 815         int do_bit17_swizzling;
 816
 817         remain = args->size;
 818
 819         /* Pin the user pages containing the data.  We can't fault while
 820          * holding the struct mutex, and all of the pwrite implementations
 821          * want to hold it while dereferencing the user data.
 822          */
 823         first_data_page = data_ptr / PAGE_SIZE;
 824         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
 825         num_pages = last_data_page - first_data_page + 1;
 826
 827         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
 828         if (user_pages == NULL)
 829                 return -ENOMEM;
 830
 831         down_read(&mm->mmap_sem);
 832         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
 833                                       num_pages, 0, 0, user_pages, NULL);
 834         up_read(&mm->mmap_sem);
 835         if (pinned_pages < num_pages) {
 836                 ret = -EFAULT;
 837                 goto fail_put_user_pages;
 838         }
 839
 840         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
 841
 842         mutex_lock(&dev->struct_mutex);
 843
 844         ret = i915_gem_object_get_pages(obj);
 845         if (ret != 0)
 846                 goto fail_unlock;
 847
 848         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
 849         if (ret != 0)
 850                 goto fail_put_pages;
 851
 852         obj_priv = obj->driver_private;
 853         offset = args->offset;
 854         obj_priv->dirty = 1;
 855
 856         while (remain > 0) {
 857                 /* Operation in this page
 858                  *
 859                  * shmem_page_index = page number within shmem file
 860                  * shmem_page_offset = offset within page in shmem file
 861                  * data_page_index = page number in get_user_pages return
 862                  * data_page_offset = offset with data_page_index page.
 863                  * page_length = bytes to copy for this page
 864                  */
 865                 shmem_page_index = offset / PAGE_SIZE;
 866                 shmem_page_offset = offset & ~PAGE_MASK;
 867                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
 868                 data_page_offset = data_ptr & ~PAGE_MASK;
 869
 870                 page_length = remain;
 871                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
 872                         page_length = PAGE_SIZE - shmem_page_offset;
 873                 if ((data_page_offset + page_length) > PAGE_SIZE)
 874                         page_length = PAGE_SIZE - data_page_offset;
 875
 876                 if (do_bit17_swizzling) {
 877                         ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
 878                                                     shmem_page_offset,
 879                                                     user_pages[data_page_index],
 880                                                     data_page_offset,
 881                                                     page_length,
 882                                                     0);
 883                 } else {
 884                         ret = slow_shmem_copy(obj_priv->pages[shmem_page_index],
 885                                               shmem_page_offset,
 886                                               user_pages[data_page_index],
 887                                               data_page_offset,
 888                                               page_length);
 889                 }
 890                 if (ret)
 891                         goto fail_put_pages;
 892
 893                 remain -= page_length;
 894                 data_ptr += page_length;
 895                 offset += page_length;
 896         }
 897
 898 fail_put_pages:
 899         i915_gem_object_put_pages(obj);
 900 fail_unlock:
 901         mutex_unlock(&dev->struct_mutex);
 902 fail_put_user_pages:
 903         for (i = 0; i < pinned_pages; i++)
 904                 page_cache_release(user_pages[i]);
 905         drm_free_large(user_pages);
 906
 907         return ret;
 908 }
 909
 910 /**
 911  * Writes data to the object referenced by handle.
 912  *
 913  * On error, the contents of the buffer that were to be modified are undefined.
 914  */
 915 int
 916 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 917                       struct drm_file *file_priv)
 918 {
 919         struct drm_i915_gem_pwrite *args = data;
 920         struct drm_gem_object *obj;
 921         struct drm_i915_gem_object *obj_priv;
 922         int ret = 0;
 923
 924         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
 925         if (obj == NULL)
 926                 return -EBADF;
 927         obj_priv = obj->driver_private;
 928
 929         /* Bounds check destination.
 930          *
 931          * XXX: This could use review for overflow issues...
 932          */
 933         if (args->offset > obj->size || args->size > obj->size ||
 934             args->offset + args->size > obj->size) {
 935                 drm_gem_object_unreference(obj);
 936                 return -EINVAL;
 937         }
 938
 939         /* We can only do the GTT pwrite on untiled buffers, as otherwise
 940          * it would end up going through the fenced access, and we'll get
 941          * different detiling behavior between reading and writing.
 942          * pread/pwrite currently are reading and writing from the CPU
 943          * perspective, requiring manual detiling by the client.
 944          */
 945         if (obj_priv->phys_obj)
 946                 ret = i915_gem_phys_pwrite(dev, obj, args, file_priv);
 947         else if (obj_priv->tiling_mode == I915_TILING_NONE &&
 948                  dev->gtt_total != 0) {
 949                 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file_priv);
 950                 if (ret == -EFAULT) {
 951                         ret = i915_gem_gtt_pwrite_slow(dev, obj, args,
 952                                                        file_priv);
 953                 }
 954         } else if (i915_gem_object_needs_bit17_swizzle(obj)) {
 955                 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file_priv);
 956         } else {
 957                 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file_priv);
 958                 if (ret == -EFAULT) {
 959                         ret = i915_gem_shmem_pwrite_slow(dev, obj, args,
 960                                                          file_priv);
 961                 }
 962         }
 963
 964 #if WATCH_PWRITE
 965         if (ret)
 966                 DRM_INFO("pwrite failed %d\n", ret);
 967 #endif
 968
 969         drm_gem_object_unreference(obj);
 970
 971         return ret;
 972 }
 973
 974 /**
 975  * Called when user space prepares to use an object with the CPU, either
 976  * through the mmap ioctl's mapping or a GTT mapping.
 977  */
 978 int
 979 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
 980                           struct drm_file *file_priv)
 981 {
 982         struct drm_i915_gem_set_domain *args = data;
 983         struct drm_gem_object *obj;
 984         uint32_t read_domains = args->read_domains;
 985         uint32_t write_domain = args->write_domain;
 986         int ret;
 987
 988         if (!(dev->driver->driver_features & DRIVER_GEM))
 989                 return -ENODEV;
 990
 991         /* Only handle setting domains to types used by the CPU. */
 992         if (write_domain & I915_GEM_GPU_DOMAINS)
 993                 return -EINVAL;
 994
 995         if (read_domains & I915_GEM_GPU_DOMAINS)
 996                 return -EINVAL;
 997
 998         /* Having something in the write domain implies it's in the read
 999          * domain, and only that read domain.  Enforce that in the request.
1000          */
1001         if (write_domain != 0 && read_domains != write_domain)
1002                 return -EINVAL;
1003
1004         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1005         if (obj == NULL)
1006                 return -EBADF;
1007
1008         mutex_lock(&dev->struct_mutex);
1009 #if WATCH_BUF
1010         DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n",
1011                  obj, obj->size, read_domains, write_domain);
1012 #endif
1013         if (read_domains & I915_GEM_DOMAIN_GTT) {
1014                 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1015
1016                 /* Silently promote "you're not bound, there was nothing to do"
1017                  * to success, since the client was just asking us to
1018                  * make sure everything was done.
1019                  */
1020                 if (ret == -EINVAL)
1021                         ret = 0;
1022         } else {
1023                 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1024         }
1025
1026         drm_gem_object_unreference(obj);
1027         mutex_unlock(&dev->struct_mutex);
1028         return ret;
1029 }
1030
1031 /**
1032  * Called when user space has done writes to this buffer
1033  */
1034 int
1035 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1036                       struct drm_file *file_priv)
1037 {
1038         struct drm_i915_gem_sw_finish *args = data;
1039         struct drm_gem_object *obj;
1040         struct drm_i915_gem_object *obj_priv;
1041         int ret = 0;
1042
1043         if (!(dev->driver->driver_features & DRIVER_GEM))
1044                 return -ENODEV;
1045
1046         mutex_lock(&dev->struct_mutex);
1047         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1048         if (obj == NULL) {
1049                 mutex_unlock(&dev->struct_mutex);
1050                 return -EBADF;
1051         }
1052
1053 #if WATCH_BUF
1054         DRM_INFO("%s: sw_finish %d (%p %d)\n",
1055                  __func__, args->handle, obj, obj->size);
1056 #endif
1057         obj_priv = obj->driver_private;
1058
1059         /* Pinned buffers may be scanout, so flush the cache */
1060         if (obj_priv->pin_count)
1061                 i915_gem_object_flush_cpu_write_domain(obj);
1062
1063         drm_gem_object_unreference(obj);
1064         mutex_unlock(&dev->struct_mutex);
1065         return ret;
1066 }
1067
1068 /**
1069  * Maps the contents of an object, returning the address it is mapped
1070  * into.
1071  *
1072  * While the mapping holds a reference on the contents of the object, it doesn't
1073  * imply a ref on the object itself.
1074  */
1075 int
1076 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1077                    struct drm_file *file_priv)
1078 {
1079         struct drm_i915_gem_mmap *args = data;
1080         struct drm_gem_object *obj;
1081         loff_t offset;
1082         unsigned long addr;
1083
1084         if (!(dev->driver->driver_features & DRIVER_GEM))
1085                 return -ENODEV;
1086
1087         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1088         if (obj == NULL)
1089                 return -EBADF;
1090
1091         offset = args->offset;
1092
1093         down_write(&current->mm->mmap_sem);
1094         addr = do_mmap(obj->filp, 0, args->size,
1095                        PROT_READ | PROT_WRITE, MAP_SHARED,
1096                        args->offset);
1097         up_write(&current->mm->mmap_sem);
1098         mutex_lock(&dev->struct_mutex);
1099         drm_gem_object_unreference(obj);
1100         mutex_unlock(&dev->struct_mutex);
1101         if (IS_ERR((void *)addr))
1102                 return addr;
1103
1104         args->addr_ptr = (uint64_t) addr;
1105
1106         return 0;
1107 }
1108
1109 /**
1110  * i915_gem_fault - fault a page into the GTT
1111  * vma: VMA in question
1112  * vmf: fault info
1113  *
1114  * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1115  * from userspace.  The fault handler takes care of binding the object to
1116  * the GTT (if needed), allocating and programming a fence register (again,
1117  * only if needed based on whether the old reg is still valid or the object
1118  * is tiled) and inserting a new PTE into the faulting process.
1119  *
1120  * Note that the faulting process may involve evicting existing objects
1121  * from the GTT and/or fence registers to make room.  So performance may
1122  * suffer if the GTT working set is large or there are few fence registers
1123  * left.
1124  */
1125 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1126 {
1127         struct drm_gem_object *obj = vma->vm_private_data;
1128         struct drm_device *dev = obj->dev;
1129         struct drm_i915_private *dev_priv = dev->dev_private;
1130         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1131         pgoff_t page_offset;
1132         unsigned long pfn;
1133         int ret = 0;
1134         bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1135
1136         /* We don't use vmf->pgoff since that has the fake offset */
1137         page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1138                 PAGE_SHIFT;
1139
1140         /* Now bind it into the GTT if needed */
1141         mutex_lock(&dev->struct_mutex);
1142         if (!obj_priv->gtt_space) {
1143                 ret = i915_gem_object_bind_to_gtt(obj, obj_priv->gtt_alignment);
1144                 if (ret) {
1145                         mutex_unlock(&dev->struct_mutex);
1146                         return VM_FAULT_SIGBUS;
1147                 }
1148
1149                 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1150                 if (ret) {
1151                         mutex_unlock(&dev->struct_mutex);
1152                         return VM_FAULT_SIGBUS;
1153                 }
1154
1155                 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1156         }
1157
1158         /* Need a new fence register? */
1159         if (obj_priv->fence_reg == I915_FENCE_REG_NONE &&
1160             obj_priv->tiling_mode != I915_TILING_NONE) {
1161                 ret = i915_gem_object_get_fence_reg(obj, write);
1162                 if (ret) {
1163                         mutex_unlock(&dev->struct_mutex);
1164                         return VM_FAULT_SIGBUS;
1165                 }
1166         }
1167
1168         pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) +
1169                 page_offset;
1170
1171         /* Finally, remap it using the new GTT offset */
1172         ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1173
1174         mutex_unlock(&dev->struct_mutex);
1175
1176         switch (ret) {
1177         case -ENOMEM:
1178         case -EAGAIN:
1179                 return VM_FAULT_OOM;
1180         case -EFAULT:
1181         case -EINVAL:
1182                 return VM_FAULT_SIGBUS;
1183         default:
1184                 return VM_FAULT_NOPAGE;
1185         }
1186 }
1187
1188 /**
1189  * i915_gem_create_mmap_offset - create a fake mmap offset for an object
1190  * @obj: obj in question
1191  *
1192  * GEM memory mapping works by handing back to userspace a fake mmap offset
1193  * it can use in a subsequent mmap(2) call.  The DRM core code then looks
1194  * up the object based on the offset and sets up the various memory mapping
1195  * structures.
1196  *
1197  * This routine allocates and attaches a fake offset for @obj.
1198  */
1199 static int
1200 i915_gem_create_mmap_offset(struct drm_gem_object *obj)
1201 {
1202         struct drm_device *dev = obj->dev;
1203         struct drm_gem_mm *mm = dev->mm_private;
1204         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1205         struct drm_map_list *list;
1206         struct drm_local_map *map;
1207         int ret = 0;
1208
1209         /* Set the object up for mmap'ing */
1210         list = &obj->map_list;
1211         list->map = drm_calloc(1, sizeof(struct drm_map_list),
1212                                DRM_MEM_DRIVER);
1213         if (!list->map)
1214                 return -ENOMEM;
1215
1216         map = list->map;
1217         map->type = _DRM_GEM;
1218         map->size = obj->size;
1219         map->handle = obj;
1220
1221         /* Get a DRM GEM mmap offset allocated... */
1222         list->file_offset_node = drm_mm_search_free(&mm->offset_manager,
1223                                                     obj->size / PAGE_SIZE, 0, 0);
1224         if (!list->file_offset_node) {
1225                 DRM_ERROR("failed to allocate offset for bo %d\n", obj->name);
1226                 ret = -ENOMEM;
1227                 goto out_free_list;
1228         }
1229
1230         list->file_offset_node = drm_mm_get_block(list->file_offset_node,
1231                                                   obj->size / PAGE_SIZE, 0);
1232         if (!list->file_offset_node) {
1233                 ret = -ENOMEM;
1234                 goto out_free_list;
1235         }
1236
1237         list->hash.key = list->file_offset_node->start;
1238         if (drm_ht_insert_item(&mm->offset_hash, &list->hash)) {
1239                 DRM_ERROR("failed to add to map hash\n");
1240                 goto out_free_mm;
1241         }
1242
1243         /* By now we should be all set, any drm_mmap request on the offset
1244          * below will get to our mmap & fault handler */
1245         obj_priv->mmap_offset = ((uint64_t) list->hash.key) << PAGE_SHIFT;
1246
1247         return 0;
1248
1249 out_free_mm:
1250         drm_mm_put_block(list->file_offset_node);
1251 out_free_list:
1252         drm_free(list->map, sizeof(struct drm_map_list), DRM_MEM_DRIVER);
1253
1254         return ret;
1255 }
1256
1257 static void
1258 i915_gem_free_mmap_offset(struct drm_gem_object *obj)
1259 {
1260         struct drm_device *dev = obj->dev;
1261         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1262         struct drm_gem_mm *mm = dev->mm_private;
1263         struct drm_map_list *list;
1264
1265         list = &obj->map_list;
1266         drm_ht_remove_item(&mm->offset_hash, &list->hash);
1267
1268         if (list->file_offset_node) {
1269                 drm_mm_put_block(list->file_offset_node);
1270                 list->file_offset_node = NULL;
1271         }
1272
1273         if (list->map) {
1274                 drm_free(list->map, sizeof(struct drm_map), DRM_MEM_DRIVER);
1275                 list->map = NULL;
1276         }
1277
1278         obj_priv->mmap_offset = 0;
1279 }
1280
1281 /**
1282  * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1283  * @obj: object to check
1284  *
1285  * Return the required GTT alignment for an object, taking into account
1286  * potential fence register mapping if needed.
1287  */
1288 static uint32_t
1289 i915_gem_get_gtt_alignment(struct drm_gem_object *obj)
1290 {
1291         struct drm_device *dev = obj->dev;
1292         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1293         int start, i;
1294
1295         /*
1296          * Minimum alignment is 4k (GTT page size), but might be greater
1297          * if a fence register is needed for the object.
1298          */
1299         if (IS_I965G(dev) || obj_priv->tiling_mode == I915_TILING_NONE)
1300                 return 4096;
1301
1302         /*
1303          * Previous chips need to be aligned to the size of the smallest
1304          * fence register that can contain the object.
1305          */
1306         if (IS_I9XX(dev))
1307                 start = 1024*1024;
1308         else
1309                 start = 512*1024;
1310
1311         for (i = start; i < obj->size; i <<= 1)
1312                 ;
1313
1314         return i;
1315 }
1316
1317 /**
1318  * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1319  * @dev: DRM device
1320  * @data: GTT mapping ioctl data
1321  * @file_priv: GEM object info
1322  *
1323  * Simply returns the fake offset to userspace so it can mmap it.
1324  * The mmap call will end up in drm_gem_mmap(), which will set things
1325  * up so we can get faults in the handler above.
1326  *
1327  * The fault handler will take care of binding the object into the GTT
1328  * (since it may have been evicted to make room for something), allocating
1329  * a fence register, and mapping the appropriate aperture address into
1330  * userspace.
1331  */
1332 int
1333 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1334                         struct drm_file *file_priv)
1335 {
1336         struct drm_i915_gem_mmap_gtt *args = data;
1337         struct drm_i915_private *dev_priv = dev->dev_private;
1338         struct drm_gem_object *obj;
1339         struct drm_i915_gem_object *obj_priv;
1340         int ret;
1341
1342         if (!(dev->driver->driver_features & DRIVER_GEM))
1343                 return -ENODEV;
1344
1345         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1346         if (obj == NULL)
1347                 return -EBADF;
1348
1349         mutex_lock(&dev->struct_mutex);
1350
1351         obj_priv = obj->driver_private;
1352
1353         if (!obj_priv->mmap_offset) {
1354                 ret = i915_gem_create_mmap_offset(obj);
1355                 if (ret) {
1356                         drm_gem_object_unreference(obj);
1357                         mutex_unlock(&dev->struct_mutex);
1358                         return ret;
1359                 }
1360         }
1361
1362         args->offset = obj_priv->mmap_offset;
1363
1364         obj_priv->gtt_alignment = i915_gem_get_gtt_alignment(obj);
1365
1366         /* Make sure the alignment is correct for fence regs etc */
1367         if (obj_priv->agp_mem &&
1368             (obj_priv->gtt_offset & (obj_priv->gtt_alignment - 1))) {
1369                 drm_gem_object_unreference(obj);
1370                 mutex_unlock(&dev->struct_mutex);
1371                 return -EINVAL;
1372         }
1373
1374         /*
1375          * Pull it into the GTT so that we have a page list (makes the
1376          * initial fault faster and any subsequent flushing possible).
1377          */
1378         if (!obj_priv->agp_mem) {
1379                 ret = i915_gem_object_bind_to_gtt(obj, obj_priv->gtt_alignment);
1380                 if (ret) {
1381                         drm_gem_object_unreference(obj);
1382                         mutex_unlock(&dev->struct_mutex);
1383                         return ret;
1384                 }
1385                 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1386         }
1387
1388         drm_gem_object_unreference(obj);
1389         mutex_unlock(&dev->struct_mutex);
1390
1391         return 0;
1392 }
1393
1394 void
1395 i915_gem_object_put_pages(struct drm_gem_object *obj)
1396 {
1397         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1398         int page_count = obj->size / PAGE_SIZE;
1399         int i;
1400
1401         BUG_ON(obj_priv->pages_refcount == 0);
1402
1403         if (--obj_priv->pages_refcount != 0)
1404                 return;
1405
1406         if (obj_priv->tiling_mode != I915_TILING_NONE)
1407                 i915_gem_object_save_bit_17_swizzle(obj);
1408
1409         for (i = 0; i < page_count; i++)
1410                 if (obj_priv->pages[i] != NULL) {
1411                         if (obj_priv->dirty)
1412                                 set_page_dirty(obj_priv->pages[i]);
1413                         mark_page_accessed(obj_priv->pages[i]);
1414                         page_cache_release(obj_priv->pages[i]);
1415                 }
1416         obj_priv->dirty = 0;
1417
1418         drm_free_large(obj_priv->pages);
1419         obj_priv->pages = NULL;
1420 }
1421
1422 static void
1423 i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno)
1424 {
1425         struct drm_device *dev = obj->dev;
1426         drm_i915_private_t *dev_priv = dev->dev_private;
1427         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1428
1429         /* Add a reference if we're newly entering the active list. */
1430         if (!obj_priv->active) {
1431                 drm_gem_object_reference(obj);
1432                 obj_priv->active = 1;
1433         }
1434         /* Move from whatever list we were on to the tail of execution. */
1435         spin_lock(&dev_priv->mm.active_list_lock);
1436         list_move_tail(&obj_priv->list,
1437                        &dev_priv->mm.active_list);
1438         spin_unlock(&dev_priv->mm.active_list_lock);
1439         obj_priv->last_rendering_seqno = seqno;
1440 }
1441
1442 static void
1443 i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
1444 {
1445         struct drm_device *dev = obj->dev;
1446         drm_i915_private_t *dev_priv = dev->dev_private;
1447         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1448
1449         BUG_ON(!obj_priv->active);
1450         list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list);
1451         obj_priv->last_rendering_seqno = 0;
1452 }
1453
1454 static void
1455 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
1456 {
1457         struct drm_device *dev = obj->dev;
1458         drm_i915_private_t *dev_priv = dev->dev_private;
1459         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1460
1461         i915_verify_inactive(dev, __FILE__, __LINE__);
1462         if (obj_priv->pin_count != 0)
1463                 list_del_init(&obj_priv->list);
1464         else
1465                 list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1466
1467         obj_priv->last_rendering_seqno = 0;
1468         if (obj_priv->active) {
1469                 obj_priv->active = 0;
1470                 drm_gem_object_unreference(obj);
1471         }
1472         i915_verify_inactive(dev, __FILE__, __LINE__);
1473 }
1474
1475 /**
1476  * Creates a new sequence number, emitting a write of it to the status page
1477  * plus an interrupt, which will trigger i915_user_interrupt_handler.
1478  *
1479  * Must be called with struct_lock held.
1480  *
1481  * Returned sequence numbers are nonzero on success.
1482  */
1483 static uint32_t
1484 i915_add_request(struct drm_device *dev, struct drm_file *file_priv,
1485                  uint32_t flush_domains)
1486 {
1487         drm_i915_private_t *dev_priv = dev->dev_private;
1488         struct drm_i915_file_private *i915_file_priv = NULL;
1489         struct drm_i915_gem_request *request;
1490         uint32_t seqno;
1491         int was_empty;
1492         RING_LOCALS;
1493
1494         if (file_priv != NULL)
1495                 i915_file_priv = file_priv->driver_priv;
1496
1497         request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER);
1498         if (request == NULL)
1499                 return 0;
1500
1501         /* Grab the seqno we're going to make this request be, and bump the
1502          * next (skipping 0 so it can be the reserved no-seqno value).
1503          */
1504         seqno = dev_priv->mm.next_gem_seqno;
1505         dev_priv->mm.next_gem_seqno++;
1506         if (dev_priv->mm.next_gem_seqno == 0)
1507                 dev_priv->mm.next_gem_seqno++;
1508
1509         BEGIN_LP_RING(4);
1510         OUT_RING(MI_STORE_DWORD_INDEX);
1511         OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
1512         OUT_RING(seqno);
1513
1514         OUT_RING(MI_USER_INTERRUPT);
1515         ADVANCE_LP_RING();
1516
1517         DRM_DEBUG("%d\n", seqno);
1518
1519         request->seqno = seqno;
1520         request->emitted_jiffies = jiffies;
1521         was_empty = list_empty(&dev_priv->mm.request_list);
1522         list_add_tail(&request->list, &dev_priv->mm.request_list);
1523         if (i915_file_priv) {
1524                 list_add_tail(&request->client_list,
1525                               &i915_file_priv->mm.request_list);
1526         } else {
1527                 INIT_LIST_HEAD(&request->client_list);
1528         }
1529
1530         /* Associate any objects on the flushing list matching the write
1531          * domain we're flushing with our flush.
1532          */
1533         if (flush_domains != 0) {
1534                 struct drm_i915_gem_object *obj_priv, *next;
1535
1536                 list_for_each_entry_safe(obj_priv, next,
1537                                          &dev_priv->mm.flushing_list, list) {
1538                         struct drm_gem_object *obj = obj_priv->obj;
1539
1540                         if ((obj->write_domain & flush_domains) ==
1541                             obj->write_domain) {
1542                                 obj->write_domain = 0;
1543                                 i915_gem_object_move_to_active(obj, seqno);
1544                         }
1545                 }
1546
1547         }
1548
1549         if (was_empty && !dev_priv->mm.suspended)
1550                 schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
1551         return seqno;
1552 }
1553
1554 /**
1555  * Command execution barrier
1556  *
1557  * Ensures that all commands in the ring are finished
1558  * before signalling the CPU
1559  */
1560 static uint32_t
1561 i915_retire_commands(struct drm_device *dev)
1562 {
1563         drm_i915_private_t *dev_priv = dev->dev_private;
1564         uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
1565         uint32_t flush_domains = 0;
1566         RING_LOCALS;
1567
1568         /* The sampler always gets flushed on i965 (sigh) */
1569         if (IS_I965G(dev))
1570                 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
1571         BEGIN_LP_RING(2);
1572         OUT_RING(cmd);
1573         OUT_RING(0); /* noop */
1574         ADVANCE_LP_RING();
1575         return flush_domains;
1576 }
1577
1578 /**
1579  * Moves buffers associated only with the given active seqno from the active
1580  * to inactive list, potentially freeing them.
1581  */
1582 static void
1583 i915_gem_retire_request(struct drm_device *dev,
1584                         struct drm_i915_gem_request *request)
1585 {
1586         drm_i915_private_t *dev_priv = dev->dev_private;
1587
1588         /* Move any buffers on the active list that are no longer referenced
1589          * by the ringbuffer to the flushing/inactive lists as appropriate.
1590          */
1591         spin_lock(&dev_priv->mm.active_list_lock);
1592         while (!list_empty(&dev_priv->mm.active_list)) {
1593                 struct drm_gem_object *obj;
1594                 struct drm_i915_gem_object *obj_priv;
1595
1596                 obj_priv = list_first_entry(&dev_priv->mm.active_list,
1597                                             struct drm_i915_gem_object,
1598                                             list);
1599                 obj = obj_priv->obj;
1600
1601                 /* If the seqno being retired doesn't match the oldest in the
1602                  * list, then the oldest in the list must still be newer than
1603                  * this seqno.
1604                  */
1605                 if (obj_priv->last_rendering_seqno != request->seqno)
1606                         goto out;
1607
1608 #if WATCH_LRU
1609                 DRM_INFO("%s: retire %d moves to inactive list %p\n",
1610                          __func__, request->seqno, obj);
1611 #endif
1612
1613                 if (obj->write_domain != 0)
1614                         i915_gem_object_move_to_flushing(obj);
1615                 else {
1616                         /* Take a reference on the object so it won't be
1617                          * freed while the spinlock is held.  The list
1618                          * protection for this spinlock is safe when breaking
1619                          * the lock like this since the next thing we do
1620                          * is just get the head of the list again.
1621                          */
1622                         drm_gem_object_reference(obj);
1623                         i915_gem_object_move_to_inactive(obj);
1624                         spin_unlock(&dev_priv->mm.active_list_lock);
1625                         drm_gem_object_unreference(obj);
1626                         spin_lock(&dev_priv->mm.active_list_lock);
1627                 }
1628         }
1629 out:
1630         spin_unlock(&dev_priv->mm.active_list_lock);
1631 }
1632
1633 /**
1634  * Returns true if seq1 is later than seq2.
1635  */
1636 static int
1637 i915_seqno_passed(uint32_t seq1, uint32_t seq2)
1638 {
1639         return (int32_t)(seq1 - seq2) >= 0;
1640 }
1641
1642 uint32_t
1643 i915_get_gem_seqno(struct drm_device *dev)
1644 {
1645         drm_i915_private_t *dev_priv = dev->dev_private;
1646
1647         return READ_HWSP(dev_priv, I915_GEM_HWS_INDEX);
1648 }
1649
1650 /**
1651  * This function clears the request list as sequence numbers are passed.
1652  */
1653 void
1654 i915_gem_retire_requests(struct drm_device *dev)
1655 {
1656         drm_i915_private_t *dev_priv = dev->dev_private;
1657         uint32_t seqno;
1658
1659         if (!dev_priv->hw_status_page)
1660                 return;
1661
1662         seqno = i915_get_gem_seqno(dev);
1663
1664         while (!list_empty(&dev_priv->mm.request_list)) {
1665                 struct drm_i915_gem_request *request;
1666                 uint32_t retiring_seqno;
1667
1668                 request = list_first_entry(&dev_priv->mm.request_list,
1669                                            struct drm_i915_gem_request,
1670                                            list);
1671                 retiring_seqno = request->seqno;
1672
1673                 if (i915_seqno_passed(seqno, retiring_seqno) ||
1674                     dev_priv->mm.wedged) {
1675                         i915_gem_retire_request(dev, request);
1676
1677                         list_del(&request->list);
1678                         list_del(&request->client_list);
1679                         drm_free(request, sizeof(*request), DRM_MEM_DRIVER);
1680                 } else
1681                         break;
1682         }
1683 }
1684
1685 void
1686 i915_gem_retire_work_handler(struct work_struct *work)
1687 {
1688         drm_i915_private_t *dev_priv;
1689         struct drm_device *dev;
1690
1691         dev_priv = container_of(work, drm_i915_private_t,
1692                                 mm.retire_work.work);
1693         dev = dev_priv->dev;
1694
1695         mutex_lock(&dev->struct_mutex);
1696         i915_gem_retire_requests(dev);
1697         if (!dev_priv->mm.suspended &&
1698             !list_empty(&dev_priv->mm.request_list))
1699                 schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
1700         mutex_unlock(&dev->struct_mutex);
1701 }
1702
1703 /**
1704  * Waits for a sequence number to be signaled, and cleans up the
1705  * request and object lists appropriately for that event.
1706  */
1707 static int
1708 i915_wait_request(struct drm_device *dev, uint32_t seqno)
1709 {
1710         drm_i915_private_t *dev_priv = dev->dev_private;
1711         u32 ier;
1712         int ret = 0;
1713
1714         BUG_ON(seqno == 0);
1715
1716         if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
1717                 if (IS_IGDNG(dev))
1718                         ier = I915_READ(DEIER) | I915_READ(GTIER);
1719                 else
1720                         ier = I915_READ(IER);
1721                 if (!ier) {
1722                         DRM_ERROR("something (likely vbetool) disabled "
1723                                   "interrupts, re-enabling\n");
1724                         i915_driver_irq_preinstall(dev);
1725                         i915_driver_irq_postinstall(dev);
1726                 }
1727
1728                 dev_priv->mm.waiting_gem_seqno = seqno;
1729                 i915_user_irq_get(dev);
1730                 ret = wait_event_interruptible(dev_priv->irq_queue,
1731                                                i915_seqno_passed(i915_get_gem_seqno(dev),
1732                                                                  seqno) ||
1733                                                dev_priv->mm.wedged);
1734                 i915_user_irq_put(dev);
1735                 dev_priv->mm.waiting_gem_seqno = 0;
1736         }
1737         if (dev_priv->mm.wedged)
1738                 ret = -EIO;
1739
1740         if (ret && ret != -ERESTARTSYS)
1741                 DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
1742                           __func__, ret, seqno, i915_get_gem_seqno(dev));
1743
1744         /* Directly dispatch request retiring.  While we have the work queue
1745          * to handle this, the waiter on a request often wants an associated
1746          * buffer to have made it to the inactive list, and we would need
1747          * a separate wait queue to handle that.
1748          */
1749         if (ret == 0)
1750                 i915_gem_retire_requests(dev);
1751
1752         return ret;
1753 }
1754
1755 static void
1756 i915_gem_flush(struct drm_device *dev,
1757                uint32_t invalidate_domains,
1758                uint32_t flush_domains)
1759 {
1760         drm_i915_private_t *dev_priv = dev->dev_private;
1761         uint32_t cmd;
1762         RING_LOCALS;
1763
1764 #if WATCH_EXEC
1765         DRM_INFO("%s: invalidate %08x flush %08x\n", __func__,
1766                   invalidate_domains, flush_domains);
1767 #endif
1768
1769         if (flush_domains & I915_GEM_DOMAIN_CPU)
1770                 drm_agp_chipset_flush(dev);
1771
1772         if ((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) {
1773                 /*
1774                  * read/write caches:
1775                  *
1776                  * I915_GEM_DOMAIN_RENDER is always invalidated, but is
1777                  * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
1778                  * also flushed at 2d versus 3d pipeline switches.
1779                  *
1780                  * read-only caches:
1781                  *
1782                  * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
1783                  * MI_READ_FLUSH is set, and is always flushed on 965.
1784                  *
1785                  * I915_GEM_DOMAIN_COMMAND may not exist?
1786                  *
1787                  * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
1788                  * invalidated when MI_EXE_FLUSH is set.
1789                  *
1790                  * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
1791                  * invalidated with every MI_FLUSH.
1792                  *
1793                  * TLBs:
1794                  *
1795                  * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
1796                  * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
1797                  * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
1798                  * are flushed at any MI_FLUSH.
1799                  */
1800
1801                 cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
1802                 if ((invalidate_domains|flush_domains) &
1803                     I915_GEM_DOMAIN_RENDER)
1804                         cmd &= ~MI_NO_WRITE_FLUSH;
1805                 if (!IS_I965G(dev)) {
1806                         /*
1807                          * On the 965, the sampler cache always gets flushed
1808                          * and this bit is reserved.
1809                          */
1810                         if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
1811                                 cmd |= MI_READ_FLUSH;
1812                 }
1813                 if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
1814                         cmd |= MI_EXE_FLUSH;
1815
1816 #if WATCH_EXEC
1817                 DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd);
1818 #endif
1819                 BEGIN_LP_RING(2);
1820                 OUT_RING(cmd);
1821                 OUT_RING(0); /* noop */
1822                 ADVANCE_LP_RING();
1823         }
1824 }
1825
1826 /**
1827  * Ensures that all rendering to the object has completed and the object is
1828  * safe to unbind from the GTT or access from the CPU.
1829  */
1830 static int
1831 i915_gem_object_wait_rendering(struct drm_gem_object *obj)
1832 {
1833         struct drm_device *dev = obj->dev;
1834         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1835         int ret;
1836
1837         /* This function only exists to support waiting for existing rendering,
1838          * not for emitting required flushes.
1839          */
1840         BUG_ON((obj->write_domain & I915_GEM_GPU_DOMAINS) != 0);
1841
1842         /* If there is rendering queued on the buffer being evicted, wait for
1843          * it.
1844          */
1845         if (obj_priv->active) {
1846 #if WATCH_BUF
1847                 DRM_INFO("%s: object %p wait for seqno %08x\n",
1848                           __func__, obj, obj_priv->last_rendering_seqno);
1849 #endif
1850                 ret = i915_wait_request(dev, obj_priv->last_rendering_seqno);
1851                 if (ret != 0)
1852                         return ret;
1853         }
1854
1855         return 0;
1856 }
1857
1858 /**
1859  * Unbinds an object from the GTT aperture.
1860  */
1861 int
1862 i915_gem_object_unbind(struct drm_gem_object *obj)
1863 {
1864         struct drm_device *dev = obj->dev;
1865         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1866         loff_t offset;
1867         int ret = 0;
1868
1869 #if WATCH_BUF
1870         DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
1871         DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
1872 #endif
1873         if (obj_priv->gtt_space == NULL)
1874                 return 0;
1875
1876         if (obj_priv->pin_count != 0) {
1877                 DRM_ERROR("Attempting to unbind pinned buffer\n");
1878                 return -EINVAL;
1879         }
1880
1881         /* Move the object to the CPU domain to ensure that
1882          * any possible CPU writes while it's not in the GTT
1883          * are flushed when we go to remap it. This will
1884          * also ensure that all pending GPU writes are finished
1885          * before we unbind.
1886          */
1887         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1888         if (ret) {
1889                 if (ret != -ERESTARTSYS)
1890                         DRM_ERROR("set_domain failed: %d\n", ret);
1891                 return ret;
1892         }
1893
1894         if (obj_priv->agp_mem != NULL) {
1895                 drm_unbind_agp(obj_priv->agp_mem);
1896                 drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
1897                 obj_priv->agp_mem = NULL;
1898         }
1899
1900         BUG_ON(obj_priv->active);
1901
1902         /* blow away mappings if mapped through GTT */
1903         offset = ((loff_t) obj->map_list.hash.key) << PAGE_SHIFT;
1904         if (dev->dev_mapping)
1905                 unmap_mapping_range(dev->dev_mapping, offset, obj->size, 1);
1906
1907         if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
1908                 i915_gem_clear_fence_reg(obj);
1909
1910         i915_gem_object_put_pages(obj);
1911
1912         if (obj_priv->gtt_space) {
1913                 atomic_dec(&dev->gtt_count);
1914                 atomic_sub(obj->size, &dev->gtt_memory);
1915
1916                 drm_mm_put_block(obj_priv->gtt_space);
1917                 obj_priv->gtt_space = NULL;
1918         }
1919
1920         /* Remove ourselves from the LRU list if present. */
1921         if (!list_empty(&obj_priv->list))
1922                 list_del_init(&obj_priv->list);
1923
1924         return 0;
1925 }
1926
1927 static int
1928 i915_gem_evict_something(struct drm_device *dev)
1929 {
1930         drm_i915_private_t *dev_priv = dev->dev_private;
1931         struct drm_gem_object *obj;
1932         struct drm_i915_gem_object *obj_priv;
1933         int ret = 0;
1934
1935         for (;;) {
1936                 /* If there's an inactive buffer available now, grab it
1937                  * and be done.
1938                  */
1939                 if (!list_empty(&dev_priv->mm.inactive_list)) {
1940                         obj_priv = list_first_entry(&dev_priv->mm.inactive_list,
1941                                                     struct drm_i915_gem_object,
1942                                                     list);
1943                         obj = obj_priv->obj;
1944                         BUG_ON(obj_priv->pin_count != 0);
1945 #if WATCH_LRU
1946                         DRM_INFO("%s: evicting %p\n", __func__, obj);
1947 #endif
1948                         BUG_ON(obj_priv->active);
1949
1950                         /* Wait on the rendering and unbind the buffer. */
1951                         ret = i915_gem_object_unbind(obj);
1952                         break;
1953                 }
1954
1955                 /* If we didn't get anything, but the ring is still processing
1956                  * things, wait for one of those things to finish and hopefully
1957                  * leave us a buffer to evict.
1958                  */
1959                 if (!list_empty(&dev_priv->mm.request_list)) {
1960                         struct drm_i915_gem_request *request;
1961
1962                         request = list_first_entry(&dev_priv->mm.request_list,
1963                                                    struct drm_i915_gem_request,
1964                                                    list);
1965
1966                         ret = i915_wait_request(dev, request->seqno);
1967                         if (ret)
1968                                 break;
1969
1970                         /* if waiting caused an object to become inactive,
1971                          * then loop around and wait for it. Otherwise, we
1972                          * assume that waiting freed and unbound something,
1973                          * so there should now be some space in the GTT
1974                          */
1975                         if (!list_empty(&dev_priv->mm.inactive_list))
1976                                 continue;
1977                         break;
1978                 }
1979
1980                 /* If we didn't have anything on the request list but there
1981                  * are buffers awaiting a flush, emit one and try again.
1982                  * When we wait on it, those buffers waiting for that flush
1983                  * will get moved to inactive.
1984                  */
1985                 if (!list_empty(&dev_priv->mm.flushing_list)) {
1986                         obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
1987                                                     struct drm_i915_gem_object,
1988                                                     list);
1989                         obj = obj_priv->obj;
1990
1991                         i915_gem_flush(dev,
1992                                        obj->write_domain,
1993                                        obj->write_domain);
1994                         i915_add_request(dev, NULL, obj->write_domain);
1995
1996                         obj = NULL;
1997                         continue;
1998                 }
1999
2000                 DRM_ERROR("inactive empty %d request empty %d "
2001                           "flushing empty %d\n",
2002                           list_empty(&dev_priv->mm.inactive_list),
2003                           list_empty(&dev_priv->mm.request_list),
2004                           list_empty(&dev_priv->mm.flushing_list));
2005                 /* If we didn't do any of the above, there's nothing to be done
2006                  * and we just can't fit it in.
2007                  */
2008                 return -ENOMEM;
2009         }
2010         return ret;
2011 }
2012
2013 static int
2014 i915_gem_evict_everything(struct drm_device *dev)
2015 {
2016         int ret;
2017
2018         for (;;) {
2019                 ret = i915_gem_evict_something(dev);
2020                 if (ret != 0)
2021                         break;
2022         }
2023         if (ret == -ENOMEM)
2024                 return 0;
2025         return ret;
2026 }
2027
2028 int
2029 i915_gem_object_get_pages(struct drm_gem_object *obj)
2030 {
2031         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2032         int page_count, i;
2033         struct address_space *mapping;
2034         struct inode *inode;
2035         struct page *page;
2036         int ret;
2037
2038         if (obj_priv->pages_refcount++ != 0)
2039                 return 0;
2040
2041         /* Get the list of pages out of our struct file.  They'll be pinned
2042          * at this point until we release them.
2043          */
2044         page_count = obj->size / PAGE_SIZE;
2045         BUG_ON(obj_priv->pages != NULL);
2046         obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *));
2047         if (obj_priv->pages == NULL) {
2048                 DRM_ERROR("Faled to allocate page list\n");
2049                 obj_priv->pages_refcount--;
2050                 return -ENOMEM;
2051         }
2052
2053         inode = obj->filp->f_path.dentry->d_inode;
2054         mapping = inode->i_mapping;
2055         for (i = 0; i < page_count; i++) {
2056                 page = read_mapping_page(mapping, i, NULL);
2057                 if (IS_ERR(page)) {
2058                         ret = PTR_ERR(page);
2059                         DRM_ERROR("read_mapping_page failed: %d\n", ret);
2060                         i915_gem_object_put_pages(obj);
2061                         return ret;
2062                 }
2063                 obj_priv->pages[i] = page;
2064         }
2065
2066         if (obj_priv->tiling_mode != I915_TILING_NONE)
2067                 i915_gem_object_do_bit_17_swizzle(obj);
2068
2069         return 0;
2070 }
2071
2072 static void i965_write_fence_reg(struct drm_i915_fence_reg *reg)
2073 {
2074         struct drm_gem_object *obj = reg->obj;
2075         struct drm_device *dev = obj->dev;
2076         drm_i915_private_t *dev_priv = dev->dev_private;
2077         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2078         int regnum = obj_priv->fence_reg;
2079         uint64_t val;
2080
2081         val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2082                     0xfffff000) << 32;
2083         val |= obj_priv->gtt_offset & 0xfffff000;
2084         val |= ((obj_priv->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2085         if (obj_priv->tiling_mode == I915_TILING_Y)
2086                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2087         val |= I965_FENCE_REG_VALID;
2088
2089         I915_WRITE64(FENCE_REG_965_0 + (regnum * 8), val);
2090 }
2091
2092 static void i915_write_fence_reg(struct drm_i915_fence_reg *reg)
2093 {
2094         struct drm_gem_object *obj = reg->obj;
2095         struct drm_device *dev = obj->dev;
2096         drm_i915_private_t *dev_priv = dev->dev_private;
2097         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2098         int regnum = obj_priv->fence_reg;
2099         int tile_width;
2100         uint32_t fence_reg, val;
2101         uint32_t pitch_val;
2102
2103         if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) ||
2104             (obj_priv->gtt_offset & (obj->size - 1))) {
2105                 WARN(1, "%s: object 0x%08x not 1M or size (0x%zx) aligned\n",
2106                      __func__, obj_priv->gtt_offset, obj->size);
2107                 return;
2108         }
2109
2110         if (obj_priv->tiling_mode == I915_TILING_Y &&
2111             HAS_128_BYTE_Y_TILING(dev))
2112                 tile_width = 128;
2113         else
2114                 tile_width = 512;
2115
2116         /* Note: pitch better be a power of two tile widths */
2117         pitch_val = obj_priv->stride / tile_width;
2118         pitch_val = ffs(pitch_val) - 1;
2119
2120         val = obj_priv->gtt_offset;
2121         if (obj_priv->tiling_mode == I915_TILING_Y)
2122                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2123         val |= I915_FENCE_SIZE_BITS(obj->size);
2124         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2125         val |= I830_FENCE_REG_VALID;
2126
2127         if (regnum < 8)
2128                 fence_reg = FENCE_REG_830_0 + (regnum * 4);
2129         else
2130                 fence_reg = FENCE_REG_945_8 + ((regnum - 8) * 4);
2131         I915_WRITE(fence_reg, val);
2132 }
2133
2134 static void i830_write_fence_reg(struct drm_i915_fence_reg *reg)
2135 {
2136         struct drm_gem_object *obj = reg->obj;
2137         struct drm_device *dev = obj->dev;
2138         drm_i915_private_t *dev_priv = dev->dev_private;
2139         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2140         int regnum = obj_priv->fence_reg;
2141         uint32_t val;
2142         uint32_t pitch_val;
2143         uint32_t fence_size_bits;
2144
2145         if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) ||
2146             (obj_priv->gtt_offset & (obj->size - 1))) {
2147                 WARN(1, "%s: object 0x%08x not 512K or size aligned\n",
2148                      __func__, obj_priv->gtt_offset);
2149                 return;
2150         }
2151
2152         pitch_val = obj_priv->stride / 128;
2153         pitch_val = ffs(pitch_val) - 1;
2154         WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2155
2156         val = obj_priv->gtt_offset;
2157         if (obj_priv->tiling_mode == I915_TILING_Y)
2158                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2159         fence_size_bits = I830_FENCE_SIZE_BITS(obj->size);
2160         WARN_ON(fence_size_bits & ~0x00000f00);
2161         val |= fence_size_bits;
2162         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2163         val |= I830_FENCE_REG_VALID;
2164
2165         I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val);
2166
2167 }
2168
2169 /**
2170  * i915_gem_object_get_fence_reg - set up a fence reg for an object
2171  * @obj: object to map through a fence reg
2172  * @write: object is about to be written
2173  *
2174  * When mapping objects through the GTT, userspace wants to be able to write
2175  * to them without having to worry about swizzling if the object is tiled.
2176  *
2177  * This function walks the fence regs looking for a free one for @obj,
2178  * stealing one if it can't find any.
2179  *
2180  * It then sets up the reg based on the object's properties: address, pitch
2181  * and tiling format.
2182  */
2183 static int
2184 i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write)
2185 {
2186         struct drm_device *dev = obj->dev;
2187         struct drm_i915_private *dev_priv = dev->dev_private;
2188         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2189         struct drm_i915_fence_reg *reg = NULL;
2190         struct drm_i915_gem_object *old_obj_priv = NULL;
2191         int i, ret, avail;
2192
2193         switch (obj_priv->tiling_mode) {
2194         case I915_TILING_NONE:
2195                 WARN(1, "allocating a fence for non-tiled object?\n");
2196                 break;
2197         case I915_TILING_X:
2198                 if (!obj_priv->stride)
2199                         return -EINVAL;
2200                 WARN((obj_priv->stride & (512 - 1)),
2201                      "object 0x%08x is X tiled but has non-512B pitch\n",
2202                      obj_priv->gtt_offset);
2203                 break;
2204         case I915_TILING_Y:
2205                 if (!obj_priv->stride)
2206                         return -EINVAL;
2207                 WARN((obj_priv->stride & (128 - 1)),
2208                      "object 0x%08x is Y tiled but has non-128B pitch\n",
2209                      obj_priv->gtt_offset);
2210                 break;
2211         }
2212
2213         /* First try to find a free reg */
2214 try_again:
2215         avail = 0;
2216         for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2217                 reg = &dev_priv->fence_regs[i];
2218                 if (!reg->obj)
2219                         break;
2220
2221                 old_obj_priv = reg->obj->driver_private;
2222                 if (!old_obj_priv->pin_count)
2223                     avail++;
2224         }
2225
2226         /* None available, try to steal one or wait for a user to finish */
2227         if (i == dev_priv->num_fence_regs) {
2228                 uint32_t seqno = dev_priv->mm.next_gem_seqno;
2229                 loff_t offset;
2230
2231                 if (avail == 0)
2232                         return -ENOMEM;
2233
2234                 for (i = dev_priv->fence_reg_start;
2235                      i < dev_priv->num_fence_regs; i++) {
2236                         uint32_t this_seqno;
2237
2238                         reg = &dev_priv->fence_regs[i];
2239                         old_obj_priv = reg->obj->driver_private;
2240
2241                         if (old_obj_priv->pin_count)
2242                                 continue;
2243
2244                         /* i915 uses fences for GPU access to tiled buffers */
2245                         if (IS_I965G(dev) || !old_obj_priv->active)
2246                                 break;
2247
2248                         /* find the seqno of the first available fence */
2249                         this_seqno = old_obj_priv->last_rendering_seqno;
2250                         if (this_seqno != 0 &&
2251                             reg->obj->write_domain == 0 &&
2252                             i915_seqno_passed(seqno, this_seqno))
2253                                 seqno = this_seqno;
2254                 }
2255
2256                 /*
2257                  * Now things get ugly... we have to wait for one of the
2258                  * objects to finish before trying again.
2259                  */
2260                 if (i == dev_priv->num_fence_regs) {
2261                         if (seqno == dev_priv->mm.next_gem_seqno) {
2262                                 i915_gem_flush(dev,
2263                                                I915_GEM_GPU_DOMAINS,
2264                                                I915_GEM_GPU_DOMAINS);
2265                                 seqno = i915_add_request(dev, NULL,
2266                                                          I915_GEM_GPU_DOMAINS);
2267                                 if (seqno == 0)
2268                                         return -ENOMEM;
2269                         }
2270
2271                         ret = i915_wait_request(dev, seqno);
2272                         if (ret)
2273                                 return ret;
2274                         goto try_again;
2275                 }
2276
2277                 /*
2278                  * Zap this virtual mapping so we can set up a fence again
2279                  * for this object next time we need it.
2280                  */
2281                 offset = ((loff_t) reg->obj->map_list.hash.key) << PAGE_SHIFT;
2282                 if (dev->dev_mapping)
2283                         unmap_mapping_range(dev->dev_mapping, offset,
2284                                             reg->obj->size, 1);
2285                 old_obj_priv->fence_reg = I915_FENCE_REG_NONE;
2286         }
2287
2288         obj_priv->fence_reg = i;
2289         reg->obj = obj;
2290
2291         if (IS_I965G(dev))
2292                 i965_write_fence_reg(reg);
2293         else if (IS_I9XX(dev))
2294                 i915_write_fence_reg(reg);
2295         else
2296                 i830_write_fence_reg(reg);
2297
2298         return 0;
2299 }
2300
2301 /**
2302  * i915_gem_clear_fence_reg - clear out fence register info
2303  * @obj: object to clear
2304  *
2305  * Zeroes out the fence register itself and clears out the associated
2306  * data structures in dev_priv and obj_priv.
2307  */
2308 static void
2309 i915_gem_clear_fence_reg(struct drm_gem_object *obj)
2310 {
2311         struct drm_device *dev = obj->dev;
2312         drm_i915_private_t *dev_priv = dev->dev_private;
2313         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2314
2315         if (IS_I965G(dev))
2316                 I915_WRITE64(FENCE_REG_965_0 + (obj_priv->fence_reg * 8), 0);
2317         else {
2318                 uint32_t fence_reg;
2319
2320                 if (obj_priv->fence_reg < 8)
2321                         fence_reg = FENCE_REG_830_0 + obj_priv->fence_reg * 4;
2322                 else
2323                         fence_reg = FENCE_REG_945_8 + (obj_priv->fence_reg -
2324                                                        8) * 4;
2325
2326                 I915_WRITE(fence_reg, 0);
2327         }
2328
2329         dev_priv->fence_regs[obj_priv->fence_reg].obj = NULL;
2330         obj_priv->fence_reg = I915_FENCE_REG_NONE;
2331 }
2332
2333 /**
2334  * Finds free space in the GTT aperture and binds the object there.
2335  */
2336 static int
2337 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
2338 {
2339         struct drm_device *dev = obj->dev;
2340         drm_i915_private_t *dev_priv = dev->dev_private;
2341         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2342         struct drm_mm_node *free_space;
2343         int page_count, ret;
2344
2345         if (dev_priv->mm.suspended)
2346                 return -EBUSY;
2347         if (alignment == 0)
2348                 alignment = i915_gem_get_gtt_alignment(obj);
2349         if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) {
2350                 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2351                 return -EINVAL;
2352         }
2353
2354  search_free:
2355         free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2356                                         obj->size, alignment, 0);
2357         if (free_space != NULL) {
2358                 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
2359                                                        alignment);
2360                 if (obj_priv->gtt_space != NULL) {
2361                         obj_priv->gtt_space->private = obj;
2362                         obj_priv->gtt_offset = obj_priv->gtt_space->start;
2363                 }
2364         }
2365         if (obj_priv->gtt_space == NULL) {
2366                 bool lists_empty;
2367
2368                 /* If the gtt is empty and we're still having trouble
2369                  * fitting our object in, we're out of memory.
2370                  */
2371 #if WATCH_LRU
2372                 DRM_INFO("%s: GTT full, evicting something\n", __func__);
2373 #endif
2374                 spin_lock(&dev_priv->mm.active_list_lock);
2375                 lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2376                                list_empty(&dev_priv->mm.flushing_list) &&
2377                                list_empty(&dev_priv->mm.active_list));
2378                 spin_unlock(&dev_priv->mm.active_list_lock);
2379                 if (lists_empty) {
2380                         DRM_ERROR("GTT full, but LRU list empty\n");
2381                         return -ENOMEM;
2382                 }
2383
2384                 ret = i915_gem_evict_something(dev);
2385                 if (ret != 0) {
2386                         if (ret != -ERESTARTSYS)
2387                                 DRM_ERROR("Failed to evict a buffer %d\n", ret);
2388                         return ret;
2389                 }
2390                 goto search_free;
2391         }
2392
2393 #if WATCH_BUF
2394         DRM_INFO("Binding object of size %d at 0x%08x\n",
2395                  obj->size, obj_priv->gtt_offset);
2396 #endif
2397         ret = i915_gem_object_get_pages(obj);
2398         if (ret) {
2399                 drm_mm_put_block(obj_priv->gtt_space);
2400                 obj_priv->gtt_space = NULL;
2401                 return ret;
2402         }
2403
2404         page_count = obj->size / PAGE_SIZE;
2405         /* Create an AGP memory structure pointing at our pages, and bind it
2406          * into the GTT.
2407          */
2408         obj_priv->agp_mem = drm_agp_bind_pages(dev,
2409                                                obj_priv->pages,
2410                                                page_count,
2411                                                obj_priv->gtt_offset,
2412                                                obj_priv->agp_type);
2413         if (obj_priv->agp_mem == NULL) {
2414                 i915_gem_object_put_pages(obj);
2415                 drm_mm_put_block(obj_priv->gtt_space);
2416                 obj_priv->gtt_space = NULL;
2417                 return -ENOMEM;
2418         }
2419         atomic_inc(&dev->gtt_count);
2420         atomic_add(obj->size, &dev->gtt_memory);
2421
2422         /* Assert that the object is not currently in any GPU domain. As it
2423          * wasn't in the GTT, there shouldn't be any way it could have been in
2424          * a GPU cache
2425          */
2426         BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2427         BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2428
2429         return 0;
2430 }
2431
2432 void
2433 i915_gem_clflush_object(struct drm_gem_object *obj)
2434 {
2435         struct drm_i915_gem_object      *obj_priv = obj->driver_private;
2436
2437         /* If we don't have a page list set up, then we're not pinned
2438          * to GPU, and we can ignore the cache flush because it'll happen
2439          * again at bind time.
2440          */
2441         if (obj_priv->pages == NULL)
2442                 return;
2443
2444         /* XXX: The 865 in particular appears to be weird in how it handles
2445          * cache flushing.  We haven't figured it out, but the
2446          * clflush+agp_chipset_flush doesn't appear to successfully get the
2447          * data visible to the PGU, while wbinvd + agp_chipset_flush does.
2448          */
2449         if (IS_I865G(obj->dev)) {
2450                 wbinvd();
2451                 return;
2452         }
2453
2454         drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE);
2455 }
2456
2457 /** Flushes any GPU write domain for the object if it's dirty. */
2458 static void
2459 i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj)
2460 {
2461         struct drm_device *dev = obj->dev;
2462         uint32_t seqno;
2463
2464         if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
2465                 return;
2466
2467         /* Queue the GPU write cache flushing we need. */
2468         i915_gem_flush(dev, 0, obj->write_domain);
2469         seqno = i915_add_request(dev, NULL, obj->write_domain);
2470         obj->write_domain = 0;
2471         i915_gem_object_move_to_active(obj, seqno);
2472 }
2473
2474 /** Flushes the GTT write domain for the object if it's dirty. */
2475 static void
2476 i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj)
2477 {
2478         if (obj->write_domain != I915_GEM_DOMAIN_GTT)
2479                 return;
2480
2481         /* No actual flushing is required for the GTT write domain.   Writes
2482          * to it immediately go to main memory as far as we know, so there's
2483          * no chipset flush.  It also doesn't land in render cache.
2484          */
2485         obj->write_domain = 0;
2486 }
2487
2488 /** Flushes the CPU write domain for the object if it's dirty. */
2489 static void
2490 i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj)
2491 {
2492         struct drm_device *dev = obj->dev;
2493
2494         if (obj->write_domain != I915_GEM_DOMAIN_CPU)
2495                 return;
2496
2497         i915_gem_clflush_object(obj);
2498         drm_agp_chipset_flush(dev);
2499         obj->write_domain = 0;
2500 }
2501
2502 /**
2503  * Moves a single object to the GTT read, and possibly write domain.
2504  *
2505  * This function returns when the move is complete, including waiting on
2506  * flushes to occur.
2507  */
2508 int
2509 i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
2510 {
2511         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2512         int ret;
2513
2514         /* Not valid to be called on unbound objects. */
2515         if (obj_priv->gtt_space == NULL)
2516                 return -EINVAL;
2517
2518         i915_gem_object_flush_gpu_write_domain(obj);
2519         /* Wait on any GPU rendering and flushing to occur. */
2520         ret = i915_gem_object_wait_rendering(obj);
2521         if (ret != 0)
2522                 return ret;
2523
2524         /* If we're writing through the GTT domain, then CPU and GPU caches
2525          * will need to be invalidated at next use.
2526          */
2527         if (write)
2528                 obj->read_domains &= I915_GEM_DOMAIN_GTT;
2529
2530         i915_gem_object_flush_cpu_write_domain(obj);
2531
2532         /* It should now be out of any other write domains, and we can update
2533          * the domain values for our changes.
2534          */
2535         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2536         obj->read_domains |= I915_GEM_DOMAIN_GTT;
2537         if (write) {
2538                 obj->write_domain = I915_GEM_DOMAIN_GTT;
2539                 obj_priv->dirty = 1;
2540         }
2541
2542         return 0;
2543 }
2544
2545 /**
2546  * Moves a single object to the CPU read, and possibly write domain.
2547  *
2548  * This function returns when the move is complete, including waiting on
2549  * flushes to occur.
2550  */
2551 static int
2552 i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
2553 {
2554         int ret;
2555
2556         i915_gem_object_flush_gpu_write_domain(obj);
2557         /* Wait on any GPU rendering and flushing to occur. */
2558         ret = i915_gem_object_wait_rendering(obj);
2559         if (ret != 0)
2560                 return ret;
2561
2562         i915_gem_object_flush_gtt_write_domain(obj);
2563
2564         /* If we have a partially-valid cache of the object in the CPU,
2565          * finish invalidating it and free the per-page flags.
2566          */
2567         i915_gem_object_set_to_full_cpu_read_domain(obj);
2568
2569         /* Flush the CPU cache if it's still invalid. */
2570         if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
2571                 i915_gem_clflush_object(obj);
2572
2573                 obj->read_domains |= I915_GEM_DOMAIN_CPU;
2574         }
2575
2576         /* It should now be out of any other write domains, and we can update
2577          * the domain values for our changes.
2578          */
2579         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2580
2581         /* If we're writing through the CPU, then the GPU read domains will
2582          * need to be invalidated at next use.
2583          */
2584         if (write) {
2585                 obj->read_domains &= I915_GEM_DOMAIN_CPU;
2586                 obj->write_domain = I915_GEM_DOMAIN_CPU;
2587         }
2588
2589         return 0;
2590 }
2591
2592 /*
2593  * Set the next domain for the specified object. This
2594  * may not actually perform the necessary flushing/invaliding though,
2595  * as that may want to be batched with other set_domain operations
2596  *
2597  * This is (we hope) the only really tricky part of gem. The goal
2598  * is fairly simple -- track which caches hold bits of the object
2599  * and make sure they remain coherent. A few concrete examples may
2600  * help to explain how it works. For shorthand, we use the notation
2601  * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
2602  * a pair of read and write domain masks.
2603  *
2604  * Case 1: the batch buffer
2605  *
2606  *      1. Allocated
2607  *      2. Written by CPU
2608  *      3. Mapped to GTT
2609  *      4. Read by GPU
2610  *      5. Unmapped from GTT
2611  *      6. Freed
2612  *
2613  *      Let's take these a step at a time
2614  *
2615  *      1. Allocated
2616  *              Pages allocated from the kernel may still have
2617  *              cache contents, so we set them to (CPU, CPU) always.
2618  *      2. Written by CPU (using pwrite)
2619  *              The pwrite function calls set_domain (CPU, CPU) and
2620  *              this function does nothing (as nothing changes)
2621  *      3. Mapped by GTT
2622  *              This function asserts that the object is not
2623  *              currently in any GPU-based read or write domains
2624  *      4. Read by GPU
2625  *              i915_gem_execbuffer calls set_domain (COMMAND, 0).
2626  *              As write_domain is zero, this function adds in the
2627  *              current read domains (CPU+COMMAND, 0).
2628  *              flush_domains is set to CPU.
2629  *              invalidate_domains is set to COMMAND
2630  *              clflush is run to get data out of the CPU caches
2631  *              then i915_dev_set_domain calls i915_gem_flush to
2632  *              emit an MI_FLUSH and drm_agp_chipset_flush
2633  *      5. Unmapped from GTT
2634  *              i915_gem_object_unbind calls set_domain (CPU, CPU)
2635  *              flush_domains and invalidate_domains end up both zero
2636  *              so no flushing/invalidating happens
2637  *      6. Freed
2638  *              yay, done
2639  *
2640  * Case 2: The shared render buffer
2641  *
2642  *      1. Allocated
2643  *      2. Mapped to GTT
2644  *      3. Read/written by GPU
2645  *      4. set_domain to (CPU,CPU)
2646  *      5. Read/written by CPU
2647  *      6. Read/written by GPU
2648  *
2649  *      1. Allocated
2650  *              Same as last example, (CPU, CPU)
2651  *      2. Mapped to GTT
2652  *              Nothing changes (assertions find that it is not in the GPU)
2653  *      3. Read/written by GPU
2654  *              execbuffer calls set_domain (RENDER, RENDER)
2655  *              flush_domains gets CPU
2656  *              invalidate_domains gets GPU
2657  *              clflush (obj)
2658  *              MI_FLUSH and drm_agp_chipset_flush
2659  *      4. set_domain (CPU, CPU)
2660  *              flush_domains gets GPU
2661  *              invalidate_domains gets CPU
2662  *              wait_rendering (obj) to make sure all drawing is complete.
2663  *              This will include an MI_FLUSH to get the data from GPU
2664  *              to memory
2665  *              clflush (obj) to invalidate the CPU cache
2666  *              Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
2667  *      5. Read/written by CPU
2668  *              cache lines are loaded and dirtied
2669  *      6. Read written by GPU
2670  *              Same as last GPU access
2671  *
2672  * Case 3: The constant buffer
2673  *
2674  *      1. Allocated
2675  *      2. Written by CPU
2676  *      3. Read by GPU
2677  *      4. Updated (written) by CPU again
2678  *      5. Read by GPU
2679  *
2680  *      1. Allocated
2681  *              (CPU, CPU)
2682  *      2. Written by CPU
2683  *              (CPU, CPU)
2684  *      3. Read by GPU
2685  *              (CPU+RENDER, 0)
2686  *              flush_domains = CPU
2687  *              invalidate_domains = RENDER
2688  *              clflush (obj)
2689  *              MI_FLUSH
2690  *              drm_agp_chipset_flush
2691  *      4. Updated (written) by CPU again
2692  *              (CPU, CPU)
2693  *              flush_domains = 0 (no previous write domain)
2694  *              invalidate_domains = 0 (no new read domains)
2695  *      5. Read by GPU
2696  *              (CPU+RENDER, 0)
2697  *              flush_domains = CPU
2698  *              invalidate_domains = RENDER
2699  *              clflush (obj)
2700  *              MI_FLUSH
2701  *              drm_agp_chipset_flush
2702  */
2703 static void
2704 i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj)
2705 {
2706         struct drm_device               *dev = obj->dev;
2707         struct drm_i915_gem_object      *obj_priv = obj->driver_private;
2708         uint32_t                        invalidate_domains = 0;
2709         uint32_t                        flush_domains = 0;
2710
2711         BUG_ON(obj->pending_read_domains & I915_GEM_DOMAIN_CPU);
2712         BUG_ON(obj->pending_write_domain == I915_GEM_DOMAIN_CPU);
2713
2714 #if WATCH_BUF
2715         DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
2716                  __func__, obj,
2717                  obj->read_domains, obj->pending_read_domains,
2718                  obj->write_domain, obj->pending_write_domain);
2719 #endif
2720         /*
2721          * If the object isn't moving to a new write domain,
2722          * let the object stay in multiple read domains
2723          */
2724         if (obj->pending_write_domain == 0)
2725                 obj->pending_read_domains |= obj->read_domains;
2726         else
2727                 obj_priv->dirty = 1;
2728
2729         /*
2730          * Flush the current write domain if
2731          * the new read domains don't match. Invalidate
2732          * any read domains which differ from the old
2733          * write domain
2734          */
2735         if (obj->write_domain &&
2736             obj->write_domain != obj->pending_read_domains) {
2737                 flush_domains |= obj->write_domain;
2738                 invalidate_domains |=
2739                         obj->pending_read_domains & ~obj->write_domain;
2740         }
2741         /*
2742          * Invalidate any read caches which may have
2743          * stale data. That is, any new read domains.
2744          */
2745         invalidate_domains |= obj->pending_read_domains & ~obj->read_domains;
2746         if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
2747 #if WATCH_BUF
2748                 DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
2749                          __func__, flush_domains, invalidate_domains);
2750 #endif
2751                 i915_gem_clflush_object(obj);
2752         }
2753
2754         /* The actual obj->write_domain will be updated with
2755          * pending_write_domain after we emit the accumulated flush for all
2756          * of our domain changes in execbuffers (which clears objects'
2757          * write_domains).  So if we have a current write domain that we
2758          * aren't changing, set pending_write_domain to that.
2759          */
2760         if (flush_domains == 0 && obj->pending_write_domain == 0)
2761                 obj->pending_write_domain = obj->write_domain;
2762         obj->read_domains = obj->pending_read_domains;
2763
2764         dev->invalidate_domains |= invalidate_domains;
2765         dev->flush_domains |= flush_domains;
2766 #if WATCH_BUF
2767         DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
2768                  __func__,
2769                  obj->read_domains, obj->write_domain,
2770                  dev->invalidate_domains, dev->flush_domains);
2771 #endif
2772 }
2773
2774 /**
2775  * Moves the object from a partially CPU read to a full one.
2776  *
2777  * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
2778  * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
2779  */
2780 static void
2781 i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj)
2782 {
2783         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2784
2785         if (!obj_priv->page_cpu_valid)
2786                 return;
2787
2788         /* If we're partially in the CPU read domain, finish moving it in.
2789          */
2790         if (obj->read_domains & I915_GEM_DOMAIN_CPU) {
2791                 int i;
2792
2793                 for (i = 0; i <= (obj->size - 1) / PAGE_SIZE; i++) {
2794                         if (obj_priv->page_cpu_valid[i])
2795                                 continue;
2796                         drm_clflush_pages(obj_priv->pages + i, 1);
2797                 }
2798         }
2799
2800         /* Free the page_cpu_valid mappings which are now stale, whether
2801          * or not we've got I915_GEM_DOMAIN_CPU.
2802          */
2803         drm_free(obj_priv->page_cpu_valid, obj->size / PAGE_SIZE,
2804                  DRM_MEM_DRIVER);
2805         obj_priv->page_cpu_valid = NULL;
2806 }
2807
2808 /**
2809  * Set the CPU read domain on a range of the object.
2810  *
2811  * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
2812  * not entirely valid.  The page_cpu_valid member of the object flags which
2813  * pages have been flushed, and will be respected by
2814  * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
2815  * of the whole object.
2816  *
2817  * This function returns when the move is complete, including waiting on
2818  * flushes to occur.
2819  */
2820 static int
2821 i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
2822                                           uint64_t offset, uint64_t size)
2823 {
2824         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2825         int i, ret;
2826
2827         if (offset == 0 && size == obj->size)
2828                 return i915_gem_object_set_to_cpu_domain(obj, 0);
2829
2830         i915_gem_object_flush_gpu_write_domain(obj);
2831         /* Wait on any GPU rendering and flushing to occur. */
2832         ret = i915_gem_object_wait_rendering(obj);
2833         if (ret != 0)
2834                 return ret;
2835         i915_gem_object_flush_gtt_write_domain(obj);
2836
2837         /* If we're already fully in the CPU read domain, we're done. */
2838         if (obj_priv->page_cpu_valid == NULL &&
2839             (obj->read_domains & I915_GEM_DOMAIN_CPU) != 0)
2840                 return 0;
2841
2842         /* Otherwise, create/clear the per-page CPU read domain flag if we're
2843          * newly adding I915_GEM_DOMAIN_CPU
2844          */
2845         if (obj_priv->page_cpu_valid == NULL) {
2846                 obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE,
2847                                                       DRM_MEM_DRIVER);
2848                 if (obj_priv->page_cpu_valid == NULL)
2849                         return -ENOMEM;
2850         } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0)
2851                 memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
2852
2853         /* Flush the cache on any pages that are still invalid from the CPU's
2854          * perspective.
2855          */
2856         for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
2857              i++) {
2858                 if (obj_priv->page_cpu_valid[i])
2859                         continue;
2860
2861                 drm_clflush_pages(obj_priv->pages + i, 1);
2862
2863                 obj_priv->page_cpu_valid[i] = 1;
2864         }
2865
2866         /* It should now be out of any other write domains, and we can update
2867          * the domain values for our changes.
2868          */
2869         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2870
2871         obj->read_domains |= I915_GEM_DOMAIN_CPU;
2872
2873         return 0;
2874 }
2875
2876 /**
2877  * Pin an object to the GTT and evaluate the relocations landing in it.
2878  */
2879 static int
2880 i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
2881                                  struct drm_file *file_priv,
2882                                  struct drm_i915_gem_exec_object *entry,
2883                                  struct drm_i915_gem_relocation_entry *relocs)
2884 {
2885         struct drm_device *dev = obj->dev;
2886         drm_i915_private_t *dev_priv = dev->dev_private;
2887         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2888         int i, ret;
2889         void __iomem *reloc_page;
2890
2891         /* Choose the GTT offset for our buffer and put it there. */
2892         ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
2893         if (ret)
2894                 return ret;
2895
2896         entry->offset = obj_priv->gtt_offset;
2897
2898         /* Apply the relocations, using the GTT aperture to avoid cache
2899          * flushing requirements.
2900          */
2901         for (i = 0; i < entry->relocation_count; i++) {
2902                 struct drm_i915_gem_relocation_entry *reloc= &relocs[i];
2903                 struct drm_gem_object *target_obj;
2904                 struct drm_i915_gem_object *target_obj_priv;
2905                 uint32_t reloc_val, reloc_offset;
2906                 uint32_t __iomem *reloc_entry;
2907
2908                 target_obj = drm_gem_object_lookup(obj->dev, file_priv,
2909                                                    reloc->target_handle);
2910                 if (target_obj == NULL) {
2911                         i915_gem_object_unpin(obj);
2912                         return -EBADF;
2913                 }
2914                 target_obj_priv = target_obj->driver_private;
2915
2916                 /* The target buffer should have appeared before us in the
2917                  * exec_object list, so it should have a GTT space bound by now.
2918                  */
2919                 if (target_obj_priv->gtt_space == NULL) {
2920                         DRM_ERROR("No GTT space found for object %d\n",
2921                                   reloc->target_handle);
2922                         drm_gem_object_unreference(target_obj);
2923                         i915_gem_object_unpin(obj);
2924                         return -EINVAL;
2925                 }
2926
2927                 if (reloc->offset > obj->size - 4) {
2928                         DRM_ERROR("Relocation beyond object bounds: "
2929                                   "obj %p target %d offset %d size %d.\n",
2930                                   obj, reloc->target_handle,
2931                                   (int) reloc->offset, (int) obj->size);
2932                         drm_gem_object_unreference(target_obj);
2933                         i915_gem_object_unpin(obj);
2934                         return -EINVAL;
2935                 }
2936                 if (reloc->offset & 3) {
2937                         DRM_ERROR("Relocation not 4-byte aligned: "
2938                                   "obj %p target %d offset %d.\n",
2939                                   obj, reloc->target_handle,
2940                                   (int) reloc->offset);
2941                         drm_gem_object_unreference(target_obj);
2942                         i915_gem_object_unpin(obj);
2943                         return -EINVAL;
2944                 }
2945
2946                 if (reloc->write_domain & I915_GEM_DOMAIN_CPU ||
2947                     reloc->read_domains & I915_GEM_DOMAIN_CPU) {
2948                         DRM_ERROR("reloc with read/write CPU domains: "
2949                                   "obj %p target %d offset %d "
2950                                   "read %08x write %08x",
2951                                   obj, reloc->target_handle,
2952                                   (int) reloc->offset,
2953                                   reloc->read_domains,
2954                                   reloc->write_domain);
2955                         drm_gem_object_unreference(target_obj);
2956                         i915_gem_object_unpin(obj);
2957                         return -EINVAL;
2958                 }
2959
2960                 if (reloc->write_domain && target_obj->pending_write_domain &&
2961                     reloc->write_domain != target_obj->pending_write_domain) {
2962                         DRM_ERROR("Write domain conflict: "
2963                                   "obj %p target %d offset %d "
2964                                   "new %08x old %08x\n",
2965                                   obj, reloc->target_handle,
2966                                   (int) reloc->offset,
2967                                   reloc->write_domain,
2968                                   target_obj->pending_write_domain);
2969                         drm_gem_object_unreference(target_obj);
2970                         i915_gem_object_unpin(obj);
2971                         return -EINVAL;
2972                 }
2973
2974 #if WATCH_RELOC
2975                 DRM_INFO("%s: obj %p offset %08x target %d "
2976                          "read %08x write %08x gtt %08x "
2977                          "presumed %08x delta %08x\n",
2978                          __func__,
2979                          obj,
2980                          (int) reloc->offset,
2981                          (int) reloc->target_handle,
2982                          (int) reloc->read_domains,
2983                          (int) reloc->write_domain,
2984                          (int) target_obj_priv->gtt_offset,
2985                          (int) reloc->presumed_offset,
2986                          reloc->delta);
2987 #endif
2988
2989                 target_obj->pending_read_domains |= reloc->read_domains;
2990                 target_obj->pending_write_domain |= reloc->write_domain;
2991
2992                 /* If the relocation already has the right value in it, no
2993                  * more work needs to be done.
2994                  */
2995                 if (target_obj_priv->gtt_offset == reloc->presumed_offset) {
2996                         drm_gem_object_unreference(target_obj);
2997                         continue;
2998                 }
2999
3000                 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
3001                 if (ret != 0) {
3002                         drm_gem_object_unreference(target_obj);
3003                         i915_gem_object_unpin(obj);
3004                         return -EINVAL;
3005                 }
3006
3007                 /* Map the page containing the relocation we're going to
3008                  * perform.
3009                  */
3010                 reloc_offset = obj_priv->gtt_offset + reloc->offset;
3011                 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
3012                                                       (reloc_offset &
3013                                                        ~(PAGE_SIZE - 1)));
3014                 reloc_entry = (uint32_t __iomem *)(reloc_page +
3015                                                    (reloc_offset & (PAGE_SIZE - 1)));
3016                 reloc_val = target_obj_priv->gtt_offset + reloc->delta;
3017
3018 #if WATCH_BUF
3019                 DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
3020                           obj, (unsigned int) reloc->offset,
3021                           readl(reloc_entry), reloc_val);
3022 #endif
3023                 writel(reloc_val, reloc_entry);
3024                 io_mapping_unmap_atomic(reloc_page);
3025
3026                 /* The updated presumed offset for this entry will be
3027                  * copied back out to the user.
3028                  */
3029                 reloc->presumed_offset = target_obj_priv->gtt_offset;
3030
3031                 drm_gem_object_unreference(target_obj);
3032         }
3033
3034 #if WATCH_BUF
3035         if (0)
3036                 i915_gem_dump_object(obj, 128, __func__, ~0);
3037 #endif
3038         return 0;
3039 }
3040
3041 /** Dispatch a batchbuffer to the ring
3042  */
3043 static int
3044 i915_dispatch_gem_execbuffer(struct drm_device *dev,
3045                               struct drm_i915_gem_execbuffer *exec,
3046                               struct drm_clip_rect *cliprects,
3047                               uint64_t exec_offset)
3048 {
3049         drm_i915_private_t *dev_priv = dev->dev_private;
3050         int nbox = exec->num_cliprects;
3051         int i = 0, count;
3052         uint32_t exec_start, exec_len;
3053         RING_LOCALS;
3054
3055         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3056         exec_len = (uint32_t) exec->batch_len;
3057
3058         count = nbox ? nbox : 1;
3059
3060         for (i = 0; i < count; i++) {
3061                 if (i < nbox) {
3062                         int ret = i915_emit_box(dev, cliprects, i,
3063                                                 exec->DR1, exec->DR4);
3064                         if (ret)
3065                                 return ret;
3066                 }
3067
3068                 if (IS_I830(dev) || IS_845G(dev)) {
3069                         BEGIN_LP_RING(4);
3070                         OUT_RING(MI_BATCH_BUFFER);
3071                         OUT_RING(exec_start | MI_BATCH_NON_SECURE);
3072                         OUT_RING(exec_start + exec_len - 4);
3073                         OUT_RING(0);
3074                         ADVANCE_LP_RING();
3075                 } else {
3076                         BEGIN_LP_RING(2);
3077                         if (IS_I965G(dev)) {
3078                                 OUT_RING(MI_BATCH_BUFFER_START |
3079                                          (2 << 6) |
3080                                          MI_BATCH_NON_SECURE_I965);
3081                                 OUT_RING(exec_start);
3082                         } else {
3083                                 OUT_RING(MI_BATCH_BUFFER_START |
3084                                          (2 << 6));
3085                                 OUT_RING(exec_start | MI_BATCH_NON_SECURE);
3086                         }
3087                         ADVANCE_LP_RING();
3088                 }
3089         }
3090
3091         /* XXX breadcrumb */
3092         return 0;
3093 }
3094
3095 /* Throttle our rendering by waiting until the ring has completed our requests
3096  * emitted over 20 msec ago.
3097  *
3098  * Note that if we were to use the current jiffies each time around the loop,
3099  * we wouldn't escape the function with any frames outstanding if the time to
3100  * render a frame was over 20ms.
3101  *
3102  * This should get us reasonable parallelism between CPU and GPU but also
3103  * relatively low latency when blocking on a particular request to finish.
3104  */
3105 static int
3106 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
3107 {
3108         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
3109         int ret = 0;
3110         unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3111
3112         mutex_lock(&dev->struct_mutex);
3113         while (!list_empty(&i915_file_priv->mm.request_list)) {
3114                 struct drm_i915_gem_request *request;
3115
3116                 request = list_first_entry(&i915_file_priv->mm.request_list,
3117                                            struct drm_i915_gem_request,
3118                                            client_list);
3119
3120                 if (time_after_eq(request->emitted_jiffies, recent_enough))
3121                         break;
3122
3123                 ret = i915_wait_request(dev, request->seqno);
3124                 if (ret != 0)
3125                         break;
3126         }
3127         mutex_unlock(&dev->struct_mutex);
3128
3129         return ret;
3130 }
3131
3132 static int
3133 i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object *exec_list,
3134                               uint32_t buffer_count,
3135                               struct drm_i915_gem_relocation_entry **relocs)
3136 {
3137         uint32_t reloc_count = 0, reloc_index = 0, i;
3138         int ret;
3139
3140         *relocs = NULL;
3141         for (i = 0; i < buffer_count; i++) {
3142                 if (reloc_count + exec_list[i].relocation_count < reloc_count)
3143                         return -EINVAL;
3144                 reloc_count += exec_list[i].relocation_count;
3145         }
3146
3147         *relocs = drm_calloc_large(reloc_count, sizeof(**relocs));
3148         if (*relocs == NULL)
3149                 return -ENOMEM;
3150
3151         for (i = 0; i < buffer_count; i++) {
3152                 struct drm_i915_gem_relocation_entry __user *user_relocs;
3153
3154                 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3155
3156                 ret = copy_from_user(&(*relocs)[reloc_index],
3157                                      user_relocs,
3158                                      exec_list[i].relocation_count *
3159                                      sizeof(**relocs));
3160                 if (ret != 0) {
3161                         drm_free_large(*relocs);
3162                         *relocs = NULL;
3163                         return -EFAULT;
3164                 }
3165
3166                 reloc_index += exec_list[i].relocation_count;
3167         }
3168
3169         return 0;
3170 }
3171
3172 static int
3173 i915_gem_put_relocs_to_user(struct drm_i915_gem_exec_object *exec_list,
3174                             uint32_t buffer_count,
3175                             struct drm_i915_gem_relocation_entry *relocs)
3176 {
3177         uint32_t reloc_count = 0, i;
3178         int ret = 0;
3179
3180         for (i = 0; i < buffer_count; i++) {
3181                 struct drm_i915_gem_relocation_entry __user *user_relocs;
3182                 int unwritten;
3183
3184                 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3185
3186                 unwritten = copy_to_user(user_relocs,
3187                                          &relocs[reloc_count],
3188                                          exec_list[i].relocation_count *
3189                                          sizeof(*relocs));
3190
3191                 if (unwritten) {
3192                         ret = -EFAULT;
3193                         goto err;
3194                 }
3195
3196                 reloc_count += exec_list[i].relocation_count;
3197         }
3198
3199 err:
3200         drm_free_large(relocs);
3201
3202         return ret;
3203 }
3204
3205 static int
3206 i915_gem_check_execbuffer (struct drm_i915_gem_execbuffer *exec,
3207                            uint64_t exec_offset)
3208 {
3209         uint32_t exec_start, exec_len;
3210
3211         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3212         exec_len = (uint32_t) exec->batch_len;
3213
3214         if ((exec_start | exec_len) & 0x7)
3215                 return -EINVAL;
3216
3217         if (!exec_start)
3218                 return -EINVAL;
3219
3220         return 0;
3221 }
3222
3223 int
3224 i915_gem_execbuffer(struct drm_device *dev, void *data,
3225                     struct drm_file *file_priv)
3226 {
3227         drm_i915_private_t *dev_priv = dev->dev_private;
3228         struct drm_i915_gem_execbuffer *args = data;
3229         struct drm_i915_gem_exec_object *exec_list = NULL;
3230         struct drm_gem_object **object_list = NULL;
3231         struct drm_gem_object *batch_obj;
3232         struct drm_i915_gem_object *obj_priv;
3233         struct drm_clip_rect *cliprects = NULL;
3234         struct drm_i915_gem_relocation_entry *relocs;
3235         int ret, ret2, i, pinned = 0;
3236         uint64_t exec_offset;
3237         uint32_t seqno, flush_domains, reloc_index;
3238         int pin_tries;
3239
3240 #if WATCH_EXEC
3241         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3242                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3243 #endif
3244
3245         if (args->buffer_count < 1) {
3246                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3247                 return -EINVAL;
3248         }
3249         /* Copy in the exec list from userland */
3250         exec_list = drm_calloc_large(sizeof(*exec_list), args->buffer_count);
3251         object_list = drm_calloc_large(sizeof(*object_list), args->buffer_count);
3252         if (exec_list == NULL || object_list == NULL) {
3253                 DRM_ERROR("Failed to allocate exec or object list "
3254                           "for %d buffers\n",
3255                           args->buffer_count);
3256                 ret = -ENOMEM;
3257                 goto pre_mutex_err;
3258         }
3259         ret = copy_from_user(exec_list,
3260                              (struct drm_i915_relocation_entry __user *)
3261                              (uintptr_t) args->buffers_ptr,
3262                              sizeof(*exec_list) * args->buffer_count);
3263         if (ret != 0) {
3264                 DRM_ERROR("copy %d exec entries failed %d\n",
3265                           args->buffer_count, ret);
3266                 goto pre_mutex_err;
3267         }
3268
3269         if (args->num_cliprects != 0) {
3270                 cliprects = drm_calloc(args->num_cliprects, sizeof(*cliprects),
3271                                        DRM_MEM_DRIVER);
3272                 if (cliprects == NULL)
3273                         goto pre_mutex_err;
3274
3275                 ret = copy_from_user(cliprects,
3276                                      (struct drm_clip_rect __user *)
3277                                      (uintptr_t) args->cliprects_ptr,
3278                                      sizeof(*cliprects) * args->num_cliprects);
3279                 if (ret != 0) {
3280                         DRM_ERROR("copy %d cliprects failed: %d\n",
3281                                   args->num_cliprects, ret);
3282                         goto pre_mutex_err;
3283                 }
3284         }
3285
3286         ret = i915_gem_get_relocs_from_user(exec_list, args->buffer_count,
3287                                             &relocs);
3288         if (ret != 0)
3289                 goto pre_mutex_err;
3290
3291         mutex_lock(&dev->struct_mutex);
3292
3293         i915_verify_inactive(dev, __FILE__, __LINE__);
3294
3295         if (dev_priv->mm.wedged) {
3296                 DRM_ERROR("Execbuf while wedged\n");
3297                 mutex_unlock(&dev->struct_mutex);
3298                 ret = -EIO;
3299                 goto pre_mutex_err;
3300         }
3301
3302         if (dev_priv->mm.suspended) {
3303                 DRM_ERROR("Execbuf while VT-switched.\n");
3304                 mutex_unlock(&dev->struct_mutex);
3305                 ret = -EBUSY;
3306                 goto pre_mutex_err;
3307         }
3308
3309         /* Look up object handles */
3310         for (i = 0; i < args->buffer_count; i++) {
3311                 object_list[i] = drm_gem_object_lookup(dev, file_priv,
3312                                                        exec_list[i].handle);
3313                 if (object_list[i] == NULL) {
3314                         DRM_ERROR("Invalid object handle %d at index %d\n",
3315                                    exec_list[i].handle, i);
3316                         ret = -EBADF;
3317                         goto err;
3318                 }
3319
3320                 obj_priv = object_list[i]->driver_private;
3321                 if (obj_priv->in_execbuffer) {
3322                         DRM_ERROR("Object %p appears more than once in object list\n",
3323                                    object_list[i]);
3324                         ret = -EBADF;
3325                         goto err;
3326                 }
3327                 obj_priv->in_execbuffer = true;
3328         }
3329
3330         /* Pin and relocate */
3331         for (pin_tries = 0; ; pin_tries++) {
3332                 ret = 0;
3333                 reloc_index = 0;
3334
3335                 for (i = 0; i < args->buffer_count; i++) {
3336                         object_list[i]->pending_read_domains = 0;
3337                         object_list[i]->pending_write_domain = 0;
3338                         ret = i915_gem_object_pin_and_relocate(object_list[i],
3339                                                                file_priv,
3340                                                                &exec_list[i],
3341                                                                &relocs[reloc_index]);
3342                         if (ret)
3343                                 break;
3344                         pinned = i + 1;
3345                         reloc_index += exec_list[i].relocation_count;
3346                 }
3347                 /* success */
3348                 if (ret == 0)
3349                         break;
3350
3351                 /* error other than GTT full, or we've already tried again */
3352                 if (ret != -ENOMEM || pin_tries >= 1) {
3353                         if (ret != -ERESTARTSYS)
3354                                 DRM_ERROR("Failed to pin buffers %d\n", ret);
3355                         goto err;
3356                 }
3357
3358                 /* unpin all of our buffers */
3359                 for (i = 0; i < pinned; i++)
3360                         i915_gem_object_unpin(object_list[i]);
3361                 pinned = 0;
3362
3363                 /* evict everyone we can from the aperture */
3364                 ret = i915_gem_evict_everything(dev);
3365                 if (ret)
3366                         goto err;
3367         }
3368
3369         /* Set the pending read domains for the batch buffer to COMMAND */
3370         batch_obj = object_list[args->buffer_count-1];
3371         if (batch_obj->pending_write_domain) {
3372                 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
3373                 ret = -EINVAL;
3374                 goto err;
3375         }
3376         batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
3377
3378         /* Sanity check the batch buffer, prior to moving objects */
3379         exec_offset = exec_list[args->buffer_count - 1].offset;
3380         ret = i915_gem_check_execbuffer (args, exec_offset);
3381         if (ret != 0) {
3382                 DRM_ERROR("execbuf with invalid offset/length\n");
3383                 goto err;
3384         }
3385
3386         i915_verify_inactive(dev, __FILE__, __LINE__);
3387
3388         /* Zero the global flush/invalidate flags. These
3389          * will be modified as new domains are computed
3390          * for each object
3391          */
3392         dev->invalidate_domains = 0;
3393         dev->flush_domains = 0;
3394
3395         for (i = 0; i < args->buffer_count; i++) {
3396                 struct drm_gem_object *obj = object_list[i];
3397
3398                 /* Compute new gpu domains and update invalidate/flush */
3399                 i915_gem_object_set_to_gpu_domain(obj);
3400         }
3401
3402         i915_verify_inactive(dev, __FILE__, __LINE__);
3403
3404         if (dev->invalidate_domains | dev->flush_domains) {
3405 #if WATCH_EXEC
3406                 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
3407                           __func__,
3408                          dev->invalidate_domains,
3409                          dev->flush_domains);
3410 #endif
3411                 i915_gem_flush(dev,
3412                                dev->invalidate_domains,
3413                                dev->flush_domains);
3414                 if (dev->flush_domains)
3415                         (void)i915_add_request(dev, file_priv,
3416                                                dev->flush_domains);
3417         }
3418
3419         for (i = 0; i < args->buffer_count; i++) {
3420                 struct drm_gem_object *obj = object_list[i];
3421
3422                 obj->write_domain = obj->pending_write_domain;
3423         }
3424
3425         i915_verify_inactive(dev, __FILE__, __LINE__);
3426
3427 #if WATCH_COHERENCY
3428         for (i = 0; i < args->buffer_count; i++) {
3429                 i915_gem_object_check_coherency(object_list[i],
3430                                                 exec_list[i].handle);
3431         }
3432 #endif
3433
3434 #if WATCH_EXEC
3435         i915_gem_dump_object(batch_obj,
3436                               args->batch_len,
3437                               __func__,
3438                               ~0);
3439 #endif
3440
3441         /* Exec the batchbuffer */
3442         ret = i915_dispatch_gem_execbuffer(dev, args, cliprects, exec_offset);
3443         if (ret) {
3444                 DRM_ERROR("dispatch failed %d\n", ret);
3445                 goto err;
3446         }
3447
3448         /*
3449          * Ensure that the commands in the batch buffer are
3450          * finished before the interrupt fires
3451          */
3452         flush_domains = i915_retire_commands(dev);
3453
3454         i915_verify_inactive(dev, __FILE__, __LINE__);
3455
3456         /*
3457          * Get a seqno representing the execution of the current buffer,
3458          * which we can wait on.  We would like to mitigate these interrupts,
3459          * likely by only creating seqnos occasionally (so that we have
3460          * *some* interrupts representing completion of buffers that we can
3461          * wait on when trying to clear up gtt space).
3462          */
3463         seqno = i915_add_request(dev, file_priv, flush_domains);
3464         BUG_ON(seqno == 0);
3465         for (i = 0; i < args->buffer_count; i++) {
3466                 struct drm_gem_object *obj = object_list[i];
3467
3468                 i915_gem_object_move_to_active(obj, seqno);
3469 #if WATCH_LRU
3470                 DRM_INFO("%s: move to exec list %p\n", __func__, obj);
3471 #endif
3472         }
3473 #if WATCH_LRU
3474         i915_dump_lru(dev, __func__);
3475 #endif
3476
3477         i915_verify_inactive(dev, __FILE__, __LINE__);
3478
3479 err:
3480         for (i = 0; i < pinned; i++)
3481                 i915_gem_object_unpin(object_list[i]);
3482
3483         for (i = 0; i < args->buffer_count; i++) {
3484                 if (object_list[i]) {
3485                         obj_priv = object_list[i]->driver_private;
3486                         obj_priv->in_execbuffer = false;
3487                 }
3488                 drm_gem_object_unreference(object_list[i]);
3489         }
3490
3491         mutex_unlock(&dev->struct_mutex);
3492
3493         if (!ret) {
3494                 /* Copy the new buffer offsets back to the user's exec list. */
3495                 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
3496                                    (uintptr_t) args->buffers_ptr,
3497                                    exec_list,
3498                                    sizeof(*exec_list) * args->buffer_count);
3499                 if (ret) {
3500                         ret = -EFAULT;
3501                         DRM_ERROR("failed to copy %d exec entries "
3502                                   "back to user (%d)\n",
3503                                   args->buffer_count, ret);
3504                 }
3505         }
3506
3507         /* Copy the updated relocations out regardless of current error
3508          * state.  Failure to update the relocs would mean that the next
3509          * time userland calls execbuf, it would do so with presumed offset
3510          * state that didn't match the actual object state.
3511          */
3512         ret2 = i915_gem_put_relocs_to_user(exec_list, args->buffer_count,
3513                                            relocs);
3514         if (ret2 != 0) {
3515                 DRM_ERROR("Failed to copy relocations back out: %d\n", ret2);
3516
3517                 if (ret == 0)
3518                         ret = ret2;
3519         }
3520
3521 pre_mutex_err:
3522         drm_free_large(object_list);
3523         drm_free_large(exec_list);
3524         drm_free(cliprects, sizeof(*cliprects) * args->num_cliprects,
3525                  DRM_MEM_DRIVER);
3526
3527         return ret;
3528 }
3529
3530 int
3531 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
3532 {
3533         struct drm_device *dev = obj->dev;
3534         struct drm_i915_gem_object *obj_priv = obj->driver_private;
3535         int ret;
3536
3537         i915_verify_inactive(dev, __FILE__, __LINE__);
3538         if (obj_priv->gtt_space == NULL) {
3539                 ret = i915_gem_object_bind_to_gtt(obj, alignment);
3540                 if (ret != 0) {
3541                         if (ret != -EBUSY && ret != -ERESTARTSYS)
3542                                 DRM_ERROR("Failure to bind: %d\n", ret);
3543                         return ret;
3544                 }
3545         }
3546         /*
3547          * Pre-965 chips need a fence register set up in order to
3548          * properly handle tiled surfaces.
3549          */
3550         if (!IS_I965G(dev) &&
3551             obj_priv->fence_reg == I915_FENCE_REG_NONE &&
3552             obj_priv->tiling_mode != I915_TILING_NONE) {
3553                 ret = i915_gem_object_get_fence_reg(obj, true);
3554                 if (ret != 0) {
3555                         if (ret != -EBUSY && ret != -ERESTARTSYS)
3556                                 DRM_ERROR("Failure to install fence: %d\n",
3557                                           ret);
3558                         return ret;
3559                 }
3560         }
3561         obj_priv->pin_count++;
3562
3563         /* If the object is not active and not pending a flush,
3564          * remove it from the inactive list
3565          */
3566         if (obj_priv->pin_count == 1) {
3567                 atomic_inc(&dev->pin_count);
3568                 atomic_add(obj->size, &dev->pin_memory);
3569                 if (!obj_priv->active &&
3570                     (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0 &&
3571                     !list_empty(&obj_priv->list))
3572                         list_del_init(&obj_priv->list);
3573         }
3574         i915_verify_inactive(dev, __FILE__, __LINE__);
3575
3576         return 0;
3577 }
3578
3579 void
3580 i915_gem_object_unpin(struct drm_gem_object *obj)
3581 {
3582         struct drm_device *dev = obj->dev;
3583         drm_i915_private_t *dev_priv = dev->dev_private;
3584         struct drm_i915_gem_object *obj_priv = obj->driver_private;
3585
3586         i915_verify_inactive(dev, __FILE__, __LINE__);
3587         obj_priv->pin_count--;
3588         BUG_ON(obj_priv->pin_count < 0);
3589         BUG_ON(obj_priv->gtt_space == NULL);
3590
3591         /* If the object is no longer pinned, and is
3592          * neither active nor being flushed, then stick it on
3593          * the inactive list
3594          */
3595         if (obj_priv->pin_count == 0) {
3596                 if (!obj_priv->active &&
3597                     (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
3598                         list_move_tail(&obj_priv->list,
3599                                        &dev_priv->mm.inactive_list);
3600                 atomic_dec(&dev->pin_count);
3601                 atomic_sub(obj->size, &dev->pin_memory);
3602         }
3603         i915_verify_inactive(dev, __FILE__, __LINE__);
3604 }
3605
3606 int
3607 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
3608                    struct drm_file *file_priv)
3609 {
3610         struct drm_i915_gem_pin *args = data;
3611         struct drm_gem_object *obj;
3612         struct drm_i915_gem_object *obj_priv;
3613         int ret;
3614
3615         mutex_lock(&dev->struct_mutex);
3616
3617         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
3618         if (obj == NULL) {
3619                 DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
3620                           args->handle);
3621                 mutex_unlock(&dev->struct_mutex);
3622                 return -EBADF;
3623         }
3624         obj_priv = obj->driver_private;
3625
3626         if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) {
3627                 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
3628                           args->handle);
3629                 drm_gem_object_unreference(obj);
3630                 mutex_unlock(&dev->struct_mutex);
3631                 return -EINVAL;
3632         }
3633
3634         obj_priv->user_pin_count++;
3635         obj_priv->pin_filp = file_priv;
3636         if (obj_priv->user_pin_count == 1) {
3637                 ret = i915_gem_object_pin(obj, args->alignment);
3638                 if (ret != 0) {
3639                         drm_gem_object_unreference(obj);
3640                         mutex_unlock(&dev->struct_mutex);
3641                         return ret;
3642                 }
3643         }
3644
3645         /* XXX - flush the CPU caches for pinned objects
3646          * as the X server doesn't manage domains yet
3647          */
3648         i915_gem_object_flush_cpu_write_domain(obj);
3649         args->offset = obj_priv->gtt_offset;
3650         drm_gem_object_unreference(obj);
3651         mutex_unlock(&dev->struct_mutex);
3652
3653         return 0;
3654 }
3655
3656 int
3657 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
3658                      struct drm_file *file_priv)
3659 {
3660         struct drm_i915_gem_pin *args = data;
3661         struct drm_gem_object *obj;
3662         struct drm_i915_gem_object *obj_priv;
3663
3664         mutex_lock(&dev->struct_mutex);
3665
3666         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
3667         if (obj == NULL) {
3668                 DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
3669                           args->handle);
3670                 mutex_unlock(&dev->struct_mutex);
3671                 return -EBADF;
3672         }
3673
3674         obj_priv = obj->driver_private;
3675         if (obj_priv->pin_filp != file_priv) {
3676                 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
3677                           args->handle);
3678                 drm_gem_object_unreference(obj);
3679                 mutex_unlock(&dev->struct_mutex);
3680                 return -EINVAL;
3681         }
3682         obj_priv->user_pin_count--;
3683         if (obj_priv->user_pin_count == 0) {
3684                 obj_priv->pin_filp = NULL;
3685                 i915_gem_object_unpin(obj);
3686         }
3687
3688         drm_gem_object_unreference(obj);
3689         mutex_unlock(&dev->struct_mutex);
3690         return 0;
3691 }
3692
3693 int
3694 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
3695                     struct drm_file *file_priv)
3696 {
3697         struct drm_i915_gem_busy *args = data;
3698         struct drm_gem_object *obj;
3699         struct drm_i915_gem_object *obj_priv;
3700
3701         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
3702         if (obj == NULL) {
3703                 DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
3704                           args->handle);
3705                 return -EBADF;
3706         }
3707
3708         mutex_lock(&dev->struct_mutex);
3709         /* Update the active list for the hardware's current position.
3710          * Otherwise this only updates on a delayed timer or when irqs are
3711          * actually unmasked, and our working set ends up being larger than
3712          * required.
3713          */
3714         i915_gem_retire_requests(dev);
3715
3716         obj_priv = obj->driver_private;
3717         /* Don't count being on the flushing list against the object being
3718          * done.  Otherwise, a buffer left on the flushing list but not getting
3719          * flushed (because nobody's flushing that domain) won't ever return
3720          * unbusy and get reused by libdrm's bo cache.  The other expected
3721          * consumer of this interface, OpenGL's occlusion queries, also specs
3722          * that the objects get unbusy "eventually" without any interference.
3723          */
3724         args->busy = obj_priv->active && obj_priv->last_rendering_seqno != 0;
3725
3726         drm_gem_object_unreference(obj);
3727         mutex_unlock(&dev->struct_mutex);
3728         return 0;
3729 }
3730
3731 int
3732 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
3733                         struct drm_file *file_priv)
3734 {
3735     return i915_gem_ring_throttle(dev, file_priv);
3736 }
3737
3738 int i915_gem_init_object(struct drm_gem_object *obj)
3739 {
3740         struct drm_i915_gem_object *obj_priv;
3741
3742         obj_priv = drm_calloc(1, sizeof(*obj_priv), DRM_MEM_DRIVER);
3743         if (obj_priv == NULL)
3744                 return -ENOMEM;
3745
3746         /*
3747          * We've just allocated pages from the kernel,
3748          * so they've just been written by the CPU with
3749          * zeros. They'll need to be clflushed before we
3750          * use them with the GPU.
3751          */
3752         obj->write_domain = I915_GEM_DOMAIN_CPU;
3753         obj->read_domains = I915_GEM_DOMAIN_CPU;
3754
3755         obj_priv->agp_type = AGP_USER_MEMORY;
3756
3757         obj->driver_private = obj_priv;
3758         obj_priv->obj = obj;
3759         obj_priv->fence_reg = I915_FENCE_REG_NONE;
3760         INIT_LIST_HEAD(&obj_priv->list);
3761
3762         return 0;
3763 }
3764
3765 void i915_gem_free_object(struct drm_gem_object *obj)
3766 {
3767         struct drm_device *dev = obj->dev;
3768         struct drm_i915_gem_object *obj_priv = obj->driver_private;
3769
3770         while (obj_priv->pin_count > 0)
3771                 i915_gem_object_unpin(obj);
3772
3773         if (obj_priv->phys_obj)
3774                 i915_gem_detach_phys_object(dev, obj);
3775
3776         i915_gem_object_unbind(obj);
3777
3778         i915_gem_free_mmap_offset(obj);
3779
3780         drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER);
3781         kfree(obj_priv->bit_17);
3782         drm_free(obj->driver_private, 1, DRM_MEM_DRIVER);
3783 }
3784
3785 /** Unbinds all objects that are on the given buffer list. */
3786 static int
3787 i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head)
3788 {
3789         struct drm_gem_object *obj;
3790         struct drm_i915_gem_object *obj_priv;
3791         int ret;
3792
3793         while (!list_empty(head)) {
3794                 obj_priv = list_first_entry(head,
3795                                             struct drm_i915_gem_object,
3796                                             list);
3797                 obj = obj_priv->obj;
3798
3799                 if (obj_priv->pin_count != 0) {
3800                         DRM_ERROR("Pinned object in unbind list\n");
3801                         mutex_unlock(&dev->struct_mutex);
3802                         return -EINVAL;
3803                 }
3804
3805                 ret = i915_gem_object_unbind(obj);
3806                 if (ret != 0) {
3807                         DRM_ERROR("Error unbinding object in LeaveVT: %d\n",
3808                                   ret);
3809                         mutex_unlock(&dev->struct_mutex);
3810                         return ret;
3811                 }
3812         }
3813
3814
3815         return 0;
3816 }
3817
3818 int
3819 i915_gem_idle(struct drm_device *dev)
3820 {
3821         drm_i915_private_t *dev_priv = dev->dev_private;
3822         uint32_t seqno, cur_seqno, last_seqno;
3823         int stuck, ret;
3824
3825         mutex_lock(&dev->struct_mutex);
3826
3827         if (dev_priv->mm.suspended || dev_priv->ring.ring_obj == NULL) {
3828                 mutex_unlock(&dev->struct_mutex);
3829                 return 0;
3830         }
3831
3832         /* Hack!  Don't let anybody do execbuf while we don't control the chip.
3833          * We need to replace this with a semaphore, or something.
3834          */
3835         dev_priv->mm.suspended = 1;
3836
3837         /* Cancel the retire work handler, wait for it to finish if running
3838          */
3839         mutex_unlock(&dev->struct_mutex);
3840         cancel_delayed_work_sync(&dev_priv->mm.retire_work);
3841         mutex_lock(&dev->struct_mutex);
3842
3843         i915_kernel_lost_context(dev);
3844
3845         /* Flush the GPU along with all non-CPU write domains
3846          */
3847         i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
3848         seqno = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS);
3849
3850         if (seqno == 0) {
3851                 mutex_unlock(&dev->struct_mutex);
3852                 return -ENOMEM;
3853         }
3854
3855         dev_priv->mm.waiting_gem_seqno = seqno;
3856         last_seqno = 0;
3857         stuck = 0;
3858         for (;;) {
3859                 cur_seqno = i915_get_gem_seqno(dev);
3860                 if (i915_seqno_passed(cur_seqno, seqno))
3861                         break;
3862                 if (last_seqno == cur_seqno) {
3863                         if (stuck++ > 100) {
3864                                 DRM_ERROR("hardware wedged\n");
3865                                 dev_priv->mm.wedged = 1;
3866                                 DRM_WAKEUP(&dev_priv->irq_queue);
3867                                 break;
3868                         }
3869                 }
3870                 msleep(10);
3871                 last_seqno = cur_seqno;
3872         }
3873         dev_priv->mm.waiting_gem_seqno = 0;
3874
3875         i915_gem_retire_requests(dev);
3876
3877         spin_lock(&dev_priv->mm.active_list_lock);
3878         if (!dev_priv->mm.wedged) {
3879                 /* Active and flushing should now be empty as we've
3880                  * waited for a sequence higher than any pending execbuffer
3881                  */
3882                 WARN_ON(!list_empty(&dev_priv->mm.active_list));
3883                 WARN_ON(!list_empty(&dev_priv->mm.flushing_list));
3884                 /* Request should now be empty as we've also waited
3885                  * for the last request in the list
3886                  */
3887                 WARN_ON(!list_empty(&dev_priv->mm.request_list));
3888         }
3889
3890         /* Empty the active and flushing lists to inactive.  If there's
3891          * anything left at this point, it means that we're wedged and
3892          * nothing good's going to happen by leaving them there.  So strip
3893          * the GPU domains and just stuff them onto inactive.
3894          */
3895         while (!list_empty(&dev_priv->mm.active_list)) {
3896                 struct drm_i915_gem_object *obj_priv;
3897
3898                 obj_priv = list_first_entry(&dev_priv->mm.active_list,
3899                                             struct drm_i915_gem_object,
3900                                             list);
3901                 obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
3902                 i915_gem_object_move_to_inactive(obj_priv->obj);
3903         }
3904         spin_unlock(&dev_priv->mm.active_list_lock);
3905
3906         while (!list_empty(&dev_priv->mm.flushing_list)) {
3907                 struct drm_i915_gem_object *obj_priv;
3908
3909                 obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
3910                                             struct drm_i915_gem_object,
3911                                             list);
3912                 obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS;
3913                 i915_gem_object_move_to_inactive(obj_priv->obj);
3914         }
3915
3916
3917         /* Move all inactive buffers out of the GTT. */
3918         ret = i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list);
3919         WARN_ON(!list_empty(&dev_priv->mm.inactive_list));
3920         if (ret) {
3921                 mutex_unlock(&dev->struct_mutex);
3922                 return ret;
3923         }
3924
3925         i915_gem_cleanup_ringbuffer(dev);
3926         mutex_unlock(&dev->struct_mutex);
3927
3928         return 0;
3929 }
3930
3931 static int
3932 i915_gem_init_hws(struct drm_device *dev)
3933 {
3934         drm_i915_private_t *dev_priv = dev->dev_private;
3935         struct drm_gem_object *obj;
3936         struct drm_i915_gem_object *obj_priv;
3937         int ret;
3938
3939         /* If we need a physical address for the status page, it's already
3940          * initialized at driver load time.
3941          */
3942         if (!I915_NEED_GFX_HWS(dev))
3943                 return 0;
3944
3945         obj = drm_gem_object_alloc(dev, 4096);
3946         if (obj == NULL) {
3947                 DRM_ERROR("Failed to allocate status page\n");
3948                 return -ENOMEM;
3949         }
3950         obj_priv = obj->driver_private;
3951         obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
3952
3953         ret = i915_gem_object_pin(obj, 4096);
3954         if (ret != 0) {
3955                 drm_gem_object_unreference(obj);
3956                 return ret;
3957         }
3958
3959         dev_priv->status_gfx_addr = obj_priv->gtt_offset;
3960
3961         dev_priv->hw_status_page = kmap(obj_priv->pages[0]);
3962         if (dev_priv->hw_status_page == NULL) {
3963                 DRM_ERROR("Failed to map status page.\n");
3964                 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
3965                 i915_gem_object_unpin(obj);
3966                 drm_gem_object_unreference(obj);
3967                 return -EINVAL;
3968         }
3969         dev_priv->hws_obj = obj;
3970         memset(dev_priv->hw_status_page, 0, PAGE_SIZE);
3971         I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr);
3972         I915_READ(HWS_PGA); /* posting read */
3973         DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr);
3974
3975         return 0;
3976 }
3977
3978 static void
3979 i915_gem_cleanup_hws(struct drm_device *dev)
3980 {
3981         drm_i915_private_t *dev_priv = dev->dev_private;
3982         struct drm_gem_object *obj;
3983         struct drm_i915_gem_object *obj_priv;
3984
3985         if (dev_priv->hws_obj == NULL)
3986                 return;
3987
3988         obj = dev_priv->hws_obj;
3989         obj_priv = obj->driver_private;
3990
3991         kunmap(obj_priv->pages[0]);
3992         i915_gem_object_unpin(obj);
3993         drm_gem_object_unreference(obj);
3994         dev_priv->hws_obj = NULL;
3995
3996         memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
3997         dev_priv->hw_status_page = NULL;
3998
3999         /* Write high address into HWS_PGA when disabling. */
4000         I915_WRITE(HWS_PGA, 0x1ffff000);
4001 }
4002
4003 int
4004 i915_gem_init_ringbuffer(struct drm_device *dev)
4005 {
4006         drm_i915_private_t *dev_priv = dev->dev_private;
4007         struct drm_gem_object *obj;
4008         struct drm_i915_gem_object *obj_priv;
4009         drm_i915_ring_buffer_t *ring = &dev_priv->ring;
4010         int ret;
4011         u32 head;
4012
4013         ret = i915_gem_init_hws(dev);
4014         if (ret != 0)
4015                 return ret;
4016
4017         obj = drm_gem_object_alloc(dev, 128 * 1024);
4018         if (obj == NULL) {
4019                 DRM_ERROR("Failed to allocate ringbuffer\n");
4020                 i915_gem_cleanup_hws(dev);
4021                 return -ENOMEM;
4022         }
4023         obj_priv = obj->driver_private;
4024
4025         ret = i915_gem_object_pin(obj, 4096);
4026         if (ret != 0) {
4027                 drm_gem_object_unreference(obj);
4028                 i915_gem_cleanup_hws(dev);
4029                 return ret;
4030         }
4031
4032         /* Set up the kernel mapping for the ring. */
4033         ring->Size = obj->size;
4034         ring->tail_mask = obj->size - 1;
4035
4036         ring->map.offset = dev->agp->base + obj_priv->gtt_offset;
4037         ring->map.size = obj->size;
4038         ring->map.type = 0;
4039         ring->map.flags = 0;
4040         ring->map.mtrr = 0;
4041
4042         drm_core_ioremap_wc(&ring->map, dev);
4043         if (ring->map.handle == NULL) {
4044                 DRM_ERROR("Failed to map ringbuffer.\n");
4045                 memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
4046                 i915_gem_object_unpin(obj);
4047                 drm_gem_object_unreference(obj);
4048                 i915_gem_cleanup_hws(dev);
4049                 return -EINVAL;
4050         }
4051         ring->ring_obj = obj;
4052         ring->virtual_start = ring->map.handle;
4053
4054         /* Stop the ring if it's running. */
4055         I915_WRITE(PRB0_CTL, 0);
4056         I915_WRITE(PRB0_TAIL, 0);
4057         I915_WRITE(PRB0_HEAD, 0);
4058
4059         /* Initialize the ring. */
4060         I915_WRITE(PRB0_START, obj_priv->gtt_offset);
4061         head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4062
4063         /* G45 ring initialization fails to reset head to zero */
4064         if (head != 0) {
4065                 DRM_ERROR("Ring head not reset to zero "
4066                           "ctl %08x head %08x tail %08x start %08x\n",
4067                           I915_READ(PRB0_CTL),
4068                           I915_READ(PRB0_HEAD),
4069                           I915_READ(PRB0_TAIL),
4070                           I915_READ(PRB0_START));
4071                 I915_WRITE(PRB0_HEAD, 0);
4072
4073                 DRM_ERROR("Ring head forced to zero "
4074                           "ctl %08x head %08x tail %08x start %08x\n",
4075                           I915_READ(PRB0_CTL),
4076                           I915_READ(PRB0_HEAD),
4077                           I915_READ(PRB0_TAIL),
4078                           I915_READ(PRB0_START));
4079         }
4080
4081         I915_WRITE(PRB0_CTL,
4082                    ((obj->size - 4096) & RING_NR_PAGES) |
4083                    RING_NO_REPORT |
4084                    RING_VALID);
4085
4086         head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4087
4088         /* If the head is still not zero, the ring is dead */
4089         if (head != 0) {
4090                 DRM_ERROR("Ring initialization failed "
4091                           "ctl %08x head %08x tail %08x start %08x\n",
4092                           I915_READ(PRB0_CTL),
4093                           I915_READ(PRB0_HEAD),
4094                           I915_READ(PRB0_TAIL),
4095                           I915_READ(PRB0_START));
4096                 return -EIO;
4097         }
4098
4099         /* Update our cache of the ring state */
4100         if (!drm_core_check_feature(dev, DRIVER_MODESET))
4101                 i915_kernel_lost_context(dev);
4102         else {
4103                 ring->head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
4104                 ring->tail = I915_READ(PRB0_TAIL) & TAIL_ADDR;
4105                 ring->space = ring->head - (ring->tail + 8);
4106                 if (ring->space < 0)
4107                         ring->space += ring->Size;
4108         }
4109
4110         return 0;
4111 }
4112
4113 void
4114 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4115 {
4116         drm_i915_private_t *dev_priv = dev->dev_private;
4117
4118         if (dev_priv->ring.ring_obj == NULL)
4119                 return;
4120
4121         drm_core_ioremapfree(&dev_priv->ring.map, dev);
4122
4123         i915_gem_object_unpin(dev_priv->ring.ring_obj);
4124         drm_gem_object_unreference(dev_priv->ring.ring_obj);
4125         dev_priv->ring.ring_obj = NULL;
4126         memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
4127
4128         i915_gem_cleanup_hws(dev);
4129 }
4130
4131 int
4132 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4133                        struct drm_file *file_priv)
4134 {
4135         drm_i915_private_t *dev_priv = dev->dev_private;
4136         int ret;
4137
4138         if (drm_core_check_feature(dev, DRIVER_MODESET))
4139                 return 0;
4140
4141         if (dev_priv->mm.wedged) {
4142                 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4143                 dev_priv->mm.wedged = 0;
4144         }
4145
4146         mutex_lock(&dev->struct_mutex);
4147         dev_priv->mm.suspended = 0;
4148
4149         ret = i915_gem_init_ringbuffer(dev);
4150         if (ret != 0) {
4151                 mutex_unlock(&dev->struct_mutex);
4152                 return ret;
4153         }
4154
4155         spin_lock(&dev_priv->mm.active_list_lock);
4156         BUG_ON(!list_empty(&dev_priv->mm.active_list));
4157         spin_unlock(&dev_priv->mm.active_list_lock);
4158
4159         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
4160         BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
4161         BUG_ON(!list_empty(&dev_priv->mm.request_list));
4162         mutex_unlock(&dev->struct_mutex);
4163
4164         drm_irq_install(dev);
4165
4166         return 0;
4167 }
4168
4169 int
4170 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4171                        struct drm_file *file_priv)
4172 {
4173         int ret;
4174
4175         if (drm_core_check_feature(dev, DRIVER_MODESET))
4176                 return 0;
4177
4178         ret = i915_gem_idle(dev);
4179         drm_irq_uninstall(dev);
4180
4181         return ret;
4182 }
4183
4184 void
4185 i915_gem_lastclose(struct drm_device *dev)
4186 {
4187         int ret;
4188
4189         if (drm_core_check_feature(dev, DRIVER_MODESET))
4190                 return;
4191
4192         ret = i915_gem_idle(dev);
4193         if (ret)
4194                 DRM_ERROR("failed to idle hardware: %d\n", ret);
4195 }
4196
4197 void
4198 i915_gem_load(struct drm_device *dev)
4199 {
4200         drm_i915_private_t *dev_priv = dev->dev_private;
4201
4202         spin_lock_init(&dev_priv->mm.active_list_lock);
4203         INIT_LIST_HEAD(&dev_priv->mm.active_list);
4204         INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
4205         INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4206         INIT_LIST_HEAD(&dev_priv->mm.request_list);
4207         INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4208                           i915_gem_retire_work_handler);
4209         dev_priv->mm.next_gem_seqno = 1;
4210
4211         /* Old X drivers will take 0-2 for front, back, depth buffers */
4212         dev_priv->fence_reg_start = 3;
4213
4214         if (IS_I965G(dev) || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4215                 dev_priv->num_fence_regs = 16;
4216         else
4217                 dev_priv->num_fence_regs = 8;
4218
4219         i915_gem_detect_bit_6_swizzle(dev);
4220 }
4221
4222 /*
4223  * Create a physically contiguous memory object for this object
4224  * e.g. for cursor + overlay regs
4225  */
4226 int i915_gem_init_phys_object(struct drm_device *dev,
4227                               int id, int size)
4228 {
4229         drm_i915_private_t *dev_priv = dev->dev_private;
4230         struct drm_i915_gem_phys_object *phys_obj;
4231         int ret;
4232
4233         if (dev_priv->mm.phys_objs[id - 1] || !size)
4234                 return 0;
4235
4236         phys_obj = drm_calloc(1, sizeof(struct drm_i915_gem_phys_object), DRM_MEM_DRIVER);
4237         if (!phys_obj)
4238                 return -ENOMEM;
4239
4240         phys_obj->id = id;
4241
4242         phys_obj->handle = drm_pci_alloc(dev, size, 0, 0xffffffff);
4243         if (!phys_obj->handle) {
4244                 ret = -ENOMEM;
4245                 goto kfree_obj;
4246         }
4247 #ifdef CONFIG_X86
4248         set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4249 #endif
4250
4251         dev_priv->mm.phys_objs[id - 1] = phys_obj;
4252
4253         return 0;
4254 kfree_obj:
4255         drm_free(phys_obj, sizeof(struct drm_i915_gem_phys_object), DRM_MEM_DRIVER);
4256         return ret;
4257 }
4258
4259 void i915_gem_free_phys_object(struct drm_device *dev, int id)
4260 {
4261         drm_i915_private_t *dev_priv = dev->dev_private;
4262         struct drm_i915_gem_phys_object *phys_obj;
4263
4264         if (!dev_priv->mm.phys_objs[id - 1])
4265                 return;
4266
4267         phys_obj = dev_priv->mm.phys_objs[id - 1];
4268         if (phys_obj->cur_obj) {
4269                 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4270         }
4271
4272 #ifdef CONFIG_X86
4273         set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4274 #endif
4275         drm_pci_free(dev, phys_obj->handle);
4276         kfree(phys_obj);
4277         dev_priv->mm.phys_objs[id - 1] = NULL;
4278 }
4279
4280 void i915_gem_free_all_phys_object(struct drm_device *dev)
4281 {
4282         int i;
4283
4284         for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4285                 i915_gem_free_phys_object(dev, i);
4286 }
4287
4288 void i915_gem_detach_phys_object(struct drm_device *dev,
4289                                  struct drm_gem_object *obj)
4290 {
4291         struct drm_i915_gem_object *obj_priv;
4292         int i;
4293         int ret;
4294         int page_count;
4295
4296         obj_priv = obj->driver_private;
4297         if (!obj_priv->phys_obj)
4298                 return;
4299
4300         ret = i915_gem_object_get_pages(obj);
4301         if (ret)
4302                 goto out;
4303
4304         page_count = obj->size / PAGE_SIZE;
4305
4306         for (i = 0; i < page_count; i++) {
4307                 char *dst = kmap_atomic(obj_priv->pages[i], KM_USER0);
4308                 char *src = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4309
4310                 memcpy(dst, src, PAGE_SIZE);
4311                 kunmap_atomic(dst, KM_USER0);
4312         }
4313         drm_clflush_pages(obj_priv->pages, page_count);
4314         drm_agp_chipset_flush(dev);
4315 out:
4316         obj_priv->phys_obj->cur_obj = NULL;
4317         obj_priv->phys_obj = NULL;
4318 }
4319
4320 int
4321 i915_gem_attach_phys_object(struct drm_device *dev,
4322                             struct drm_gem_object *obj, int id)
4323 {
4324         drm_i915_private_t *dev_priv = dev->dev_private;
4325         struct drm_i915_gem_object *obj_priv;
4326         int ret = 0;
4327         int page_count;
4328         int i;
4329
4330         if (id > I915_MAX_PHYS_OBJECT)
4331                 return -EINVAL;
4332
4333         obj_priv = obj->driver_private;
4334
4335         if (obj_priv->phys_obj) {
4336                 if (obj_priv->phys_obj->id == id)
4337                         return 0;
4338                 i915_gem_detach_phys_object(dev, obj);
4339         }
4340
4341
4342         /* create a new object */
4343         if (!dev_priv->mm.phys_objs[id - 1]) {
4344                 ret = i915_gem_init_phys_object(dev, id,
4345                                                 obj->size);
4346                 if (ret) {
4347                         DRM_ERROR("failed to init phys object %d size: %zu\n", id, obj->size);
4348                         goto out;
4349                 }
4350         }
4351
4352         /* bind to the object */
4353         obj_priv->phys_obj = dev_priv->mm.phys_objs[id - 1];
4354         obj_priv->phys_obj->cur_obj = obj;
4355
4356         ret = i915_gem_object_get_pages(obj);
4357         if (ret) {
4358                 DRM_ERROR("failed to get page list\n");
4359                 goto out;
4360         }
4361
4362         page_count = obj->size / PAGE_SIZE;
4363
4364         for (i = 0; i < page_count; i++) {
4365                 char *src = kmap_atomic(obj_priv->pages[i], KM_USER0);
4366                 char *dst = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4367
4368                 memcpy(dst, src, PAGE_SIZE);
4369                 kunmap_atomic(src, KM_USER0);
4370         }
4371
4372         return 0;
4373 out:
4374         return ret;
4375 }
4376
4377 static int
4378 i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
4379                      struct drm_i915_gem_pwrite *args,
4380                      struct drm_file *file_priv)
4381 {
4382         struct drm_i915_gem_object *obj_priv = obj->driver_private;
4383         void *obj_addr;
4384         int ret;
4385         char __user *user_data;
4386
4387         user_data = (char __user *) (uintptr_t) args->data_ptr;
4388         obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
4389
4390         DRM_DEBUG("obj_addr %p, %lld\n", obj_addr, args->size);
4391         ret = copy_from_user(obj_addr, user_data, args->size);
4392         if (ret)
4393                 return -EFAULT;
4394
4395         drm_agp_chipset_flush(dev);
4396         return 0;
4397 }
4398
4399 void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv)
4400 {
4401         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
4402
4403         /* Clean up our request list when the client is going away, so that
4404          * later retire_requests won't dereference our soon-to-be-gone
4405          * file_priv.
4406          */
4407         mutex_lock(&dev->struct_mutex);
4408         while (!list_empty(&i915_file_priv->mm.request_list))
4409                 list_del_init(i915_file_priv->mm.request_list.next);
4410         mutex_unlock(&dev->struct_mutex);
4411 }