Merge branch 'fix' of git://git.kernel.org/pub/scm/linux/kernel/git/ycmiao/pxa-linux-2.6
[linux-2.6.git] / drivers / gpu / drm / i915 / i915_gem.c
1 /*
2  * Copyright © 2008 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *
26  */
27
28 #include "drmP.h"
29 #include "drm.h"
30 #include "i915_drm.h"
31 #include "i915_drv.h"
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/slab.h>
35 #include <linux/swap.h>
36 #include <linux/pci.h>
37
38 static void i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj);
39 static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj);
40 static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj);
41 static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj,
42                                              int write);
43 static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
44                                                      uint64_t offset,
45                                                      uint64_t size);
46 static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj);
47 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
48 static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj,
49                                            unsigned alignment);
50 static void i915_gem_clear_fence_reg(struct drm_gem_object *obj);
51 static int i915_gem_evict_something(struct drm_device *dev, int min_size);
52 static int i915_gem_evict_from_inactive_list(struct drm_device *dev);
53 static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
54                                 struct drm_i915_gem_pwrite *args,
55                                 struct drm_file *file_priv);
56
57 static LIST_HEAD(shrink_list);
58 static DEFINE_SPINLOCK(shrink_list_lock);
59
60 int i915_gem_do_init(struct drm_device *dev, unsigned long start,
61                      unsigned long end)
62 {
63         drm_i915_private_t *dev_priv = dev->dev_private;
64
65         if (start >= end ||
66             (start & (PAGE_SIZE - 1)) != 0 ||
67             (end & (PAGE_SIZE - 1)) != 0) {
68                 return -EINVAL;
69         }
70
71         drm_mm_init(&dev_priv->mm.gtt_space, start,
72                     end - start);
73
74         dev->gtt_total = (uint32_t) (end - start);
75
76         return 0;
77 }
78
79 int
80 i915_gem_init_ioctl(struct drm_device *dev, void *data,
81                     struct drm_file *file_priv)
82 {
83         struct drm_i915_gem_init *args = data;
84         int ret;
85
86         mutex_lock(&dev->struct_mutex);
87         ret = i915_gem_do_init(dev, args->gtt_start, args->gtt_end);
88         mutex_unlock(&dev->struct_mutex);
89
90         return ret;
91 }
92
93 int
94 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
95                             struct drm_file *file_priv)
96 {
97         struct drm_i915_gem_get_aperture *args = data;
98
99         if (!(dev->driver->driver_features & DRIVER_GEM))
100                 return -ENODEV;
101
102         args->aper_size = dev->gtt_total;
103         args->aper_available_size = (args->aper_size -
104                                      atomic_read(&dev->pin_memory));
105
106         return 0;
107 }
108
109
110 /**
111  * Creates a new mm object and returns a handle to it.
112  */
113 int
114 i915_gem_create_ioctl(struct drm_device *dev, void *data,
115                       struct drm_file *file_priv)
116 {
117         struct drm_i915_gem_create *args = data;
118         struct drm_gem_object *obj;
119         int ret;
120         u32 handle;
121
122         args->size = roundup(args->size, PAGE_SIZE);
123
124         /* Allocate the new object */
125         obj = i915_gem_alloc_object(dev, args->size);
126         if (obj == NULL)
127                 return -ENOMEM;
128
129         ret = drm_gem_handle_create(file_priv, obj, &handle);
130         drm_gem_object_handle_unreference_unlocked(obj);
131
132         if (ret)
133                 return ret;
134
135         args->handle = handle;
136
137         return 0;
138 }
139
140 static inline int
141 fast_shmem_read(struct page **pages,
142                 loff_t page_base, int page_offset,
143                 char __user *data,
144                 int length)
145 {
146         char __iomem *vaddr;
147         int unwritten;
148
149         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
150         if (vaddr == NULL)
151                 return -ENOMEM;
152         unwritten = __copy_to_user_inatomic(data, vaddr + page_offset, length);
153         kunmap_atomic(vaddr, KM_USER0);
154
155         if (unwritten)
156                 return -EFAULT;
157
158         return 0;
159 }
160
161 static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj)
162 {
163         drm_i915_private_t *dev_priv = obj->dev->dev_private;
164         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
165
166         return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 &&
167                 obj_priv->tiling_mode != I915_TILING_NONE;
168 }
169
170 static inline void
171 slow_shmem_copy(struct page *dst_page,
172                 int dst_offset,
173                 struct page *src_page,
174                 int src_offset,
175                 int length)
176 {
177         char *dst_vaddr, *src_vaddr;
178
179         dst_vaddr = kmap(dst_page);
180         src_vaddr = kmap(src_page);
181
182         memcpy(dst_vaddr + dst_offset, src_vaddr + src_offset, length);
183
184         kunmap(src_page);
185         kunmap(dst_page);
186 }
187
188 static inline void
189 slow_shmem_bit17_copy(struct page *gpu_page,
190                       int gpu_offset,
191                       struct page *cpu_page,
192                       int cpu_offset,
193                       int length,
194                       int is_read)
195 {
196         char *gpu_vaddr, *cpu_vaddr;
197
198         /* Use the unswizzled path if this page isn't affected. */
199         if ((page_to_phys(gpu_page) & (1 << 17)) == 0) {
200                 if (is_read)
201                         return slow_shmem_copy(cpu_page, cpu_offset,
202                                                gpu_page, gpu_offset, length);
203                 else
204                         return slow_shmem_copy(gpu_page, gpu_offset,
205                                                cpu_page, cpu_offset, length);
206         }
207
208         gpu_vaddr = kmap(gpu_page);
209         cpu_vaddr = kmap(cpu_page);
210
211         /* Copy the data, XORing A6 with A17 (1). The user already knows he's
212          * XORing with the other bits (A9 for Y, A9 and A10 for X)
213          */
214         while (length > 0) {
215                 int cacheline_end = ALIGN(gpu_offset + 1, 64);
216                 int this_length = min(cacheline_end - gpu_offset, length);
217                 int swizzled_gpu_offset = gpu_offset ^ 64;
218
219                 if (is_read) {
220                         memcpy(cpu_vaddr + cpu_offset,
221                                gpu_vaddr + swizzled_gpu_offset,
222                                this_length);
223                 } else {
224                         memcpy(gpu_vaddr + swizzled_gpu_offset,
225                                cpu_vaddr + cpu_offset,
226                                this_length);
227                 }
228                 cpu_offset += this_length;
229                 gpu_offset += this_length;
230                 length -= this_length;
231         }
232
233         kunmap(cpu_page);
234         kunmap(gpu_page);
235 }
236
237 /**
238  * This is the fast shmem pread path, which attempts to copy_from_user directly
239  * from the backing pages of the object to the user's address space.  On a
240  * fault, it fails so we can fall back to i915_gem_shmem_pwrite_slow().
241  */
242 static int
243 i915_gem_shmem_pread_fast(struct drm_device *dev, struct drm_gem_object *obj,
244                           struct drm_i915_gem_pread *args,
245                           struct drm_file *file_priv)
246 {
247         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
248         ssize_t remain;
249         loff_t offset, page_base;
250         char __user *user_data;
251         int page_offset, page_length;
252         int ret;
253
254         user_data = (char __user *) (uintptr_t) args->data_ptr;
255         remain = args->size;
256
257         mutex_lock(&dev->struct_mutex);
258
259         ret = i915_gem_object_get_pages(obj, 0);
260         if (ret != 0)
261                 goto fail_unlock;
262
263         ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
264                                                         args->size);
265         if (ret != 0)
266                 goto fail_put_pages;
267
268         obj_priv = to_intel_bo(obj);
269         offset = args->offset;
270
271         while (remain > 0) {
272                 /* Operation in this page
273                  *
274                  * page_base = page offset within aperture
275                  * page_offset = offset within page
276                  * page_length = bytes to copy for this page
277                  */
278                 page_base = (offset & ~(PAGE_SIZE-1));
279                 page_offset = offset & (PAGE_SIZE-1);
280                 page_length = remain;
281                 if ((page_offset + remain) > PAGE_SIZE)
282                         page_length = PAGE_SIZE - page_offset;
283
284                 ret = fast_shmem_read(obj_priv->pages,
285                                       page_base, page_offset,
286                                       user_data, page_length);
287                 if (ret)
288                         goto fail_put_pages;
289
290                 remain -= page_length;
291                 user_data += page_length;
292                 offset += page_length;
293         }
294
295 fail_put_pages:
296         i915_gem_object_put_pages(obj);
297 fail_unlock:
298         mutex_unlock(&dev->struct_mutex);
299
300         return ret;
301 }
302
303 static int
304 i915_gem_object_get_pages_or_evict(struct drm_gem_object *obj)
305 {
306         int ret;
307
308         ret = i915_gem_object_get_pages(obj, __GFP_NORETRY | __GFP_NOWARN);
309
310         /* If we've insufficient memory to map in the pages, attempt
311          * to make some space by throwing out some old buffers.
312          */
313         if (ret == -ENOMEM) {
314                 struct drm_device *dev = obj->dev;
315
316                 ret = i915_gem_evict_something(dev, obj->size);
317                 if (ret)
318                         return ret;
319
320                 ret = i915_gem_object_get_pages(obj, 0);
321         }
322
323         return ret;
324 }
325
326 /**
327  * This is the fallback shmem pread path, which allocates temporary storage
328  * in kernel space to copy_to_user into outside of the struct_mutex, so we
329  * can copy out of the object's backing pages while holding the struct mutex
330  * and not take page faults.
331  */
332 static int
333 i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj,
334                           struct drm_i915_gem_pread *args,
335                           struct drm_file *file_priv)
336 {
337         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
338         struct mm_struct *mm = current->mm;
339         struct page **user_pages;
340         ssize_t remain;
341         loff_t offset, pinned_pages, i;
342         loff_t first_data_page, last_data_page, num_pages;
343         int shmem_page_index, shmem_page_offset;
344         int data_page_index,  data_page_offset;
345         int page_length;
346         int ret;
347         uint64_t data_ptr = args->data_ptr;
348         int do_bit17_swizzling;
349
350         remain = args->size;
351
352         /* Pin the user pages containing the data.  We can't fault while
353          * holding the struct mutex, yet we want to hold it while
354          * dereferencing the user data.
355          */
356         first_data_page = data_ptr / PAGE_SIZE;
357         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
358         num_pages = last_data_page - first_data_page + 1;
359
360         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
361         if (user_pages == NULL)
362                 return -ENOMEM;
363
364         down_read(&mm->mmap_sem);
365         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
366                                       num_pages, 1, 0, user_pages, NULL);
367         up_read(&mm->mmap_sem);
368         if (pinned_pages < num_pages) {
369                 ret = -EFAULT;
370                 goto fail_put_user_pages;
371         }
372
373         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
374
375         mutex_lock(&dev->struct_mutex);
376
377         ret = i915_gem_object_get_pages_or_evict(obj);
378         if (ret)
379                 goto fail_unlock;
380
381         ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset,
382                                                         args->size);
383         if (ret != 0)
384                 goto fail_put_pages;
385
386         obj_priv = to_intel_bo(obj);
387         offset = args->offset;
388
389         while (remain > 0) {
390                 /* Operation in this page
391                  *
392                  * shmem_page_index = page number within shmem file
393                  * shmem_page_offset = offset within page in shmem file
394                  * data_page_index = page number in get_user_pages return
395                  * data_page_offset = offset with data_page_index page.
396                  * page_length = bytes to copy for this page
397                  */
398                 shmem_page_index = offset / PAGE_SIZE;
399                 shmem_page_offset = offset & ~PAGE_MASK;
400                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
401                 data_page_offset = data_ptr & ~PAGE_MASK;
402
403                 page_length = remain;
404                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
405                         page_length = PAGE_SIZE - shmem_page_offset;
406                 if ((data_page_offset + page_length) > PAGE_SIZE)
407                         page_length = PAGE_SIZE - data_page_offset;
408
409                 if (do_bit17_swizzling) {
410                         slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
411                                               shmem_page_offset,
412                                               user_pages[data_page_index],
413                                               data_page_offset,
414                                               page_length,
415                                               1);
416                 } else {
417                         slow_shmem_copy(user_pages[data_page_index],
418                                         data_page_offset,
419                                         obj_priv->pages[shmem_page_index],
420                                         shmem_page_offset,
421                                         page_length);
422                 }
423
424                 remain -= page_length;
425                 data_ptr += page_length;
426                 offset += page_length;
427         }
428
429 fail_put_pages:
430         i915_gem_object_put_pages(obj);
431 fail_unlock:
432         mutex_unlock(&dev->struct_mutex);
433 fail_put_user_pages:
434         for (i = 0; i < pinned_pages; i++) {
435                 SetPageDirty(user_pages[i]);
436                 page_cache_release(user_pages[i]);
437         }
438         drm_free_large(user_pages);
439
440         return ret;
441 }
442
443 /**
444  * Reads data from the object referenced by handle.
445  *
446  * On error, the contents of *data are undefined.
447  */
448 int
449 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
450                      struct drm_file *file_priv)
451 {
452         struct drm_i915_gem_pread *args = data;
453         struct drm_gem_object *obj;
454         struct drm_i915_gem_object *obj_priv;
455         int ret;
456
457         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
458         if (obj == NULL)
459                 return -EBADF;
460         obj_priv = to_intel_bo(obj);
461
462         /* Bounds check source.
463          *
464          * XXX: This could use review for overflow issues...
465          */
466         if (args->offset > obj->size || args->size > obj->size ||
467             args->offset + args->size > obj->size) {
468                 drm_gem_object_unreference_unlocked(obj);
469                 return -EINVAL;
470         }
471
472         if (i915_gem_object_needs_bit17_swizzle(obj)) {
473                 ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv);
474         } else {
475                 ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv);
476                 if (ret != 0)
477                         ret = i915_gem_shmem_pread_slow(dev, obj, args,
478                                                         file_priv);
479         }
480
481         drm_gem_object_unreference_unlocked(obj);
482
483         return ret;
484 }
485
486 /* This is the fast write path which cannot handle
487  * page faults in the source data
488  */
489
490 static inline int
491 fast_user_write(struct io_mapping *mapping,
492                 loff_t page_base, int page_offset,
493                 char __user *user_data,
494                 int length)
495 {
496         char *vaddr_atomic;
497         unsigned long unwritten;
498
499         vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
500         unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
501                                                       user_data, length);
502         io_mapping_unmap_atomic(vaddr_atomic);
503         if (unwritten)
504                 return -EFAULT;
505         return 0;
506 }
507
508 /* Here's the write path which can sleep for
509  * page faults
510  */
511
512 static inline void
513 slow_kernel_write(struct io_mapping *mapping,
514                   loff_t gtt_base, int gtt_offset,
515                   struct page *user_page, int user_offset,
516                   int length)
517 {
518         char __iomem *dst_vaddr;
519         char *src_vaddr;
520
521         dst_vaddr = io_mapping_map_wc(mapping, gtt_base);
522         src_vaddr = kmap(user_page);
523
524         memcpy_toio(dst_vaddr + gtt_offset,
525                     src_vaddr + user_offset,
526                     length);
527
528         kunmap(user_page);
529         io_mapping_unmap(dst_vaddr);
530 }
531
532 static inline int
533 fast_shmem_write(struct page **pages,
534                  loff_t page_base, int page_offset,
535                  char __user *data,
536                  int length)
537 {
538         char __iomem *vaddr;
539         unsigned long unwritten;
540
541         vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0);
542         if (vaddr == NULL)
543                 return -ENOMEM;
544         unwritten = __copy_from_user_inatomic(vaddr + page_offset, data, length);
545         kunmap_atomic(vaddr, KM_USER0);
546
547         if (unwritten)
548                 return -EFAULT;
549         return 0;
550 }
551
552 /**
553  * This is the fast pwrite path, where we copy the data directly from the
554  * user into the GTT, uncached.
555  */
556 static int
557 i915_gem_gtt_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
558                          struct drm_i915_gem_pwrite *args,
559                          struct drm_file *file_priv)
560 {
561         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
562         drm_i915_private_t *dev_priv = dev->dev_private;
563         ssize_t remain;
564         loff_t offset, page_base;
565         char __user *user_data;
566         int page_offset, page_length;
567         int ret;
568
569         user_data = (char __user *) (uintptr_t) args->data_ptr;
570         remain = args->size;
571         if (!access_ok(VERIFY_READ, user_data, remain))
572                 return -EFAULT;
573
574
575         mutex_lock(&dev->struct_mutex);
576         ret = i915_gem_object_pin(obj, 0);
577         if (ret) {
578                 mutex_unlock(&dev->struct_mutex);
579                 return ret;
580         }
581         ret = i915_gem_object_set_to_gtt_domain(obj, 1);
582         if (ret)
583                 goto fail;
584
585         obj_priv = to_intel_bo(obj);
586         offset = obj_priv->gtt_offset + args->offset;
587
588         while (remain > 0) {
589                 /* Operation in this page
590                  *
591                  * page_base = page offset within aperture
592                  * page_offset = offset within page
593                  * page_length = bytes to copy for this page
594                  */
595                 page_base = (offset & ~(PAGE_SIZE-1));
596                 page_offset = offset & (PAGE_SIZE-1);
597                 page_length = remain;
598                 if ((page_offset + remain) > PAGE_SIZE)
599                         page_length = PAGE_SIZE - page_offset;
600
601                 ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
602                                        page_offset, user_data, page_length);
603
604                 /* If we get a fault while copying data, then (presumably) our
605                  * source page isn't available.  Return the error and we'll
606                  * retry in the slow path.
607                  */
608                 if (ret)
609                         goto fail;
610
611                 remain -= page_length;
612                 user_data += page_length;
613                 offset += page_length;
614         }
615
616 fail:
617         i915_gem_object_unpin(obj);
618         mutex_unlock(&dev->struct_mutex);
619
620         return ret;
621 }
622
623 /**
624  * This is the fallback GTT pwrite path, which uses get_user_pages to pin
625  * the memory and maps it using kmap_atomic for copying.
626  *
627  * This code resulted in x11perf -rgb10text consuming about 10% more CPU
628  * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit).
629  */
630 static int
631 i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
632                          struct drm_i915_gem_pwrite *args,
633                          struct drm_file *file_priv)
634 {
635         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
636         drm_i915_private_t *dev_priv = dev->dev_private;
637         ssize_t remain;
638         loff_t gtt_page_base, offset;
639         loff_t first_data_page, last_data_page, num_pages;
640         loff_t pinned_pages, i;
641         struct page **user_pages;
642         struct mm_struct *mm = current->mm;
643         int gtt_page_offset, data_page_offset, data_page_index, page_length;
644         int ret;
645         uint64_t data_ptr = args->data_ptr;
646
647         remain = args->size;
648
649         /* Pin the user pages containing the data.  We can't fault while
650          * holding the struct mutex, and all of the pwrite implementations
651          * want to hold it while dereferencing the user data.
652          */
653         first_data_page = data_ptr / PAGE_SIZE;
654         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
655         num_pages = last_data_page - first_data_page + 1;
656
657         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
658         if (user_pages == NULL)
659                 return -ENOMEM;
660
661         down_read(&mm->mmap_sem);
662         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
663                                       num_pages, 0, 0, user_pages, NULL);
664         up_read(&mm->mmap_sem);
665         if (pinned_pages < num_pages) {
666                 ret = -EFAULT;
667                 goto out_unpin_pages;
668         }
669
670         mutex_lock(&dev->struct_mutex);
671         ret = i915_gem_object_pin(obj, 0);
672         if (ret)
673                 goto out_unlock;
674
675         ret = i915_gem_object_set_to_gtt_domain(obj, 1);
676         if (ret)
677                 goto out_unpin_object;
678
679         obj_priv = to_intel_bo(obj);
680         offset = obj_priv->gtt_offset + args->offset;
681
682         while (remain > 0) {
683                 /* Operation in this page
684                  *
685                  * gtt_page_base = page offset within aperture
686                  * gtt_page_offset = offset within page in aperture
687                  * data_page_index = page number in get_user_pages return
688                  * data_page_offset = offset with data_page_index page.
689                  * page_length = bytes to copy for this page
690                  */
691                 gtt_page_base = offset & PAGE_MASK;
692                 gtt_page_offset = offset & ~PAGE_MASK;
693                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
694                 data_page_offset = data_ptr & ~PAGE_MASK;
695
696                 page_length = remain;
697                 if ((gtt_page_offset + page_length) > PAGE_SIZE)
698                         page_length = PAGE_SIZE - gtt_page_offset;
699                 if ((data_page_offset + page_length) > PAGE_SIZE)
700                         page_length = PAGE_SIZE - data_page_offset;
701
702                 slow_kernel_write(dev_priv->mm.gtt_mapping,
703                                   gtt_page_base, gtt_page_offset,
704                                   user_pages[data_page_index],
705                                   data_page_offset,
706                                   page_length);
707
708                 remain -= page_length;
709                 offset += page_length;
710                 data_ptr += page_length;
711         }
712
713 out_unpin_object:
714         i915_gem_object_unpin(obj);
715 out_unlock:
716         mutex_unlock(&dev->struct_mutex);
717 out_unpin_pages:
718         for (i = 0; i < pinned_pages; i++)
719                 page_cache_release(user_pages[i]);
720         drm_free_large(user_pages);
721
722         return ret;
723 }
724
725 /**
726  * This is the fast shmem pwrite path, which attempts to directly
727  * copy_from_user into the kmapped pages backing the object.
728  */
729 static int
730 i915_gem_shmem_pwrite_fast(struct drm_device *dev, struct drm_gem_object *obj,
731                            struct drm_i915_gem_pwrite *args,
732                            struct drm_file *file_priv)
733 {
734         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
735         ssize_t remain;
736         loff_t offset, page_base;
737         char __user *user_data;
738         int page_offset, page_length;
739         int ret;
740
741         user_data = (char __user *) (uintptr_t) args->data_ptr;
742         remain = args->size;
743
744         mutex_lock(&dev->struct_mutex);
745
746         ret = i915_gem_object_get_pages(obj, 0);
747         if (ret != 0)
748                 goto fail_unlock;
749
750         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
751         if (ret != 0)
752                 goto fail_put_pages;
753
754         obj_priv = to_intel_bo(obj);
755         offset = args->offset;
756         obj_priv->dirty = 1;
757
758         while (remain > 0) {
759                 /* Operation in this page
760                  *
761                  * page_base = page offset within aperture
762                  * page_offset = offset within page
763                  * page_length = bytes to copy for this page
764                  */
765                 page_base = (offset & ~(PAGE_SIZE-1));
766                 page_offset = offset & (PAGE_SIZE-1);
767                 page_length = remain;
768                 if ((page_offset + remain) > PAGE_SIZE)
769                         page_length = PAGE_SIZE - page_offset;
770
771                 ret = fast_shmem_write(obj_priv->pages,
772                                        page_base, page_offset,
773                                        user_data, page_length);
774                 if (ret)
775                         goto fail_put_pages;
776
777                 remain -= page_length;
778                 user_data += page_length;
779                 offset += page_length;
780         }
781
782 fail_put_pages:
783         i915_gem_object_put_pages(obj);
784 fail_unlock:
785         mutex_unlock(&dev->struct_mutex);
786
787         return ret;
788 }
789
790 /**
791  * This is the fallback shmem pwrite path, which uses get_user_pages to pin
792  * the memory and maps it using kmap_atomic for copying.
793  *
794  * This avoids taking mmap_sem for faulting on the user's address while the
795  * struct_mutex is held.
796  */
797 static int
798 i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj,
799                            struct drm_i915_gem_pwrite *args,
800                            struct drm_file *file_priv)
801 {
802         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
803         struct mm_struct *mm = current->mm;
804         struct page **user_pages;
805         ssize_t remain;
806         loff_t offset, pinned_pages, i;
807         loff_t first_data_page, last_data_page, num_pages;
808         int shmem_page_index, shmem_page_offset;
809         int data_page_index,  data_page_offset;
810         int page_length;
811         int ret;
812         uint64_t data_ptr = args->data_ptr;
813         int do_bit17_swizzling;
814
815         remain = args->size;
816
817         /* Pin the user pages containing the data.  We can't fault while
818          * holding the struct mutex, and all of the pwrite implementations
819          * want to hold it while dereferencing the user data.
820          */
821         first_data_page = data_ptr / PAGE_SIZE;
822         last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE;
823         num_pages = last_data_page - first_data_page + 1;
824
825         user_pages = drm_calloc_large(num_pages, sizeof(struct page *));
826         if (user_pages == NULL)
827                 return -ENOMEM;
828
829         down_read(&mm->mmap_sem);
830         pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr,
831                                       num_pages, 0, 0, user_pages, NULL);
832         up_read(&mm->mmap_sem);
833         if (pinned_pages < num_pages) {
834                 ret = -EFAULT;
835                 goto fail_put_user_pages;
836         }
837
838         do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj);
839
840         mutex_lock(&dev->struct_mutex);
841
842         ret = i915_gem_object_get_pages_or_evict(obj);
843         if (ret)
844                 goto fail_unlock;
845
846         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
847         if (ret != 0)
848                 goto fail_put_pages;
849
850         obj_priv = to_intel_bo(obj);
851         offset = args->offset;
852         obj_priv->dirty = 1;
853
854         while (remain > 0) {
855                 /* Operation in this page
856                  *
857                  * shmem_page_index = page number within shmem file
858                  * shmem_page_offset = offset within page in shmem file
859                  * data_page_index = page number in get_user_pages return
860                  * data_page_offset = offset with data_page_index page.
861                  * page_length = bytes to copy for this page
862                  */
863                 shmem_page_index = offset / PAGE_SIZE;
864                 shmem_page_offset = offset & ~PAGE_MASK;
865                 data_page_index = data_ptr / PAGE_SIZE - first_data_page;
866                 data_page_offset = data_ptr & ~PAGE_MASK;
867
868                 page_length = remain;
869                 if ((shmem_page_offset + page_length) > PAGE_SIZE)
870                         page_length = PAGE_SIZE - shmem_page_offset;
871                 if ((data_page_offset + page_length) > PAGE_SIZE)
872                         page_length = PAGE_SIZE - data_page_offset;
873
874                 if (do_bit17_swizzling) {
875                         slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index],
876                                               shmem_page_offset,
877                                               user_pages[data_page_index],
878                                               data_page_offset,
879                                               page_length,
880                                               0);
881                 } else {
882                         slow_shmem_copy(obj_priv->pages[shmem_page_index],
883                                         shmem_page_offset,
884                                         user_pages[data_page_index],
885                                         data_page_offset,
886                                         page_length);
887                 }
888
889                 remain -= page_length;
890                 data_ptr += page_length;
891                 offset += page_length;
892         }
893
894 fail_put_pages:
895         i915_gem_object_put_pages(obj);
896 fail_unlock:
897         mutex_unlock(&dev->struct_mutex);
898 fail_put_user_pages:
899         for (i = 0; i < pinned_pages; i++)
900                 page_cache_release(user_pages[i]);
901         drm_free_large(user_pages);
902
903         return ret;
904 }
905
906 /**
907  * Writes data to the object referenced by handle.
908  *
909  * On error, the contents of the buffer that were to be modified are undefined.
910  */
911 int
912 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
913                       struct drm_file *file_priv)
914 {
915         struct drm_i915_gem_pwrite *args = data;
916         struct drm_gem_object *obj;
917         struct drm_i915_gem_object *obj_priv;
918         int ret = 0;
919
920         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
921         if (obj == NULL)
922                 return -EBADF;
923         obj_priv = to_intel_bo(obj);
924
925         /* Bounds check destination.
926          *
927          * XXX: This could use review for overflow issues...
928          */
929         if (args->offset > obj->size || args->size > obj->size ||
930             args->offset + args->size > obj->size) {
931                 drm_gem_object_unreference_unlocked(obj);
932                 return -EINVAL;
933         }
934
935         /* We can only do the GTT pwrite on untiled buffers, as otherwise
936          * it would end up going through the fenced access, and we'll get
937          * different detiling behavior between reading and writing.
938          * pread/pwrite currently are reading and writing from the CPU
939          * perspective, requiring manual detiling by the client.
940          */
941         if (obj_priv->phys_obj)
942                 ret = i915_gem_phys_pwrite(dev, obj, args, file_priv);
943         else if (obj_priv->tiling_mode == I915_TILING_NONE &&
944                  dev->gtt_total != 0 &&
945                  obj->write_domain != I915_GEM_DOMAIN_CPU) {
946                 ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file_priv);
947                 if (ret == -EFAULT) {
948                         ret = i915_gem_gtt_pwrite_slow(dev, obj, args,
949                                                        file_priv);
950                 }
951         } else if (i915_gem_object_needs_bit17_swizzle(obj)) {
952                 ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file_priv);
953         } else {
954                 ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file_priv);
955                 if (ret == -EFAULT) {
956                         ret = i915_gem_shmem_pwrite_slow(dev, obj, args,
957                                                          file_priv);
958                 }
959         }
960
961 #if WATCH_PWRITE
962         if (ret)
963                 DRM_INFO("pwrite failed %d\n", ret);
964 #endif
965
966         drm_gem_object_unreference_unlocked(obj);
967
968         return ret;
969 }
970
971 /**
972  * Called when user space prepares to use an object with the CPU, either
973  * through the mmap ioctl's mapping or a GTT mapping.
974  */
975 int
976 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
977                           struct drm_file *file_priv)
978 {
979         struct drm_i915_private *dev_priv = dev->dev_private;
980         struct drm_i915_gem_set_domain *args = data;
981         struct drm_gem_object *obj;
982         struct drm_i915_gem_object *obj_priv;
983         uint32_t read_domains = args->read_domains;
984         uint32_t write_domain = args->write_domain;
985         int ret;
986
987         if (!(dev->driver->driver_features & DRIVER_GEM))
988                 return -ENODEV;
989
990         /* Only handle setting domains to types used by the CPU. */
991         if (write_domain & I915_GEM_GPU_DOMAINS)
992                 return -EINVAL;
993
994         if (read_domains & I915_GEM_GPU_DOMAINS)
995                 return -EINVAL;
996
997         /* Having something in the write domain implies it's in the read
998          * domain, and only that read domain.  Enforce that in the request.
999          */
1000         if (write_domain != 0 && read_domains != write_domain)
1001                 return -EINVAL;
1002
1003         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1004         if (obj == NULL)
1005                 return -EBADF;
1006         obj_priv = to_intel_bo(obj);
1007
1008         mutex_lock(&dev->struct_mutex);
1009
1010         intel_mark_busy(dev, obj);
1011
1012 #if WATCH_BUF
1013         DRM_INFO("set_domain_ioctl %p(%zd), %08x %08x\n",
1014                  obj, obj->size, read_domains, write_domain);
1015 #endif
1016         if (read_domains & I915_GEM_DOMAIN_GTT) {
1017                 ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0);
1018
1019                 /* Update the LRU on the fence for the CPU access that's
1020                  * about to occur.
1021                  */
1022                 if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1023                         struct drm_i915_fence_reg *reg =
1024                                 &dev_priv->fence_regs[obj_priv->fence_reg];
1025                         list_move_tail(&reg->lru_list,
1026                                        &dev_priv->mm.fence_list);
1027                 }
1028
1029                 /* Silently promote "you're not bound, there was nothing to do"
1030                  * to success, since the client was just asking us to
1031                  * make sure everything was done.
1032                  */
1033                 if (ret == -EINVAL)
1034                         ret = 0;
1035         } else {
1036                 ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
1037         }
1038
1039         drm_gem_object_unreference(obj);
1040         mutex_unlock(&dev->struct_mutex);
1041         return ret;
1042 }
1043
1044 /**
1045  * Called when user space has done writes to this buffer
1046  */
1047 int
1048 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
1049                       struct drm_file *file_priv)
1050 {
1051         struct drm_i915_gem_sw_finish *args = data;
1052         struct drm_gem_object *obj;
1053         struct drm_i915_gem_object *obj_priv;
1054         int ret = 0;
1055
1056         if (!(dev->driver->driver_features & DRIVER_GEM))
1057                 return -ENODEV;
1058
1059         mutex_lock(&dev->struct_mutex);
1060         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1061         if (obj == NULL) {
1062                 mutex_unlock(&dev->struct_mutex);
1063                 return -EBADF;
1064         }
1065
1066 #if WATCH_BUF
1067         DRM_INFO("%s: sw_finish %d (%p %zd)\n",
1068                  __func__, args->handle, obj, obj->size);
1069 #endif
1070         obj_priv = to_intel_bo(obj);
1071
1072         /* Pinned buffers may be scanout, so flush the cache */
1073         if (obj_priv->pin_count)
1074                 i915_gem_object_flush_cpu_write_domain(obj);
1075
1076         drm_gem_object_unreference(obj);
1077         mutex_unlock(&dev->struct_mutex);
1078         return ret;
1079 }
1080
1081 /**
1082  * Maps the contents of an object, returning the address it is mapped
1083  * into.
1084  *
1085  * While the mapping holds a reference on the contents of the object, it doesn't
1086  * imply a ref on the object itself.
1087  */
1088 int
1089 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
1090                    struct drm_file *file_priv)
1091 {
1092         struct drm_i915_gem_mmap *args = data;
1093         struct drm_gem_object *obj;
1094         loff_t offset;
1095         unsigned long addr;
1096
1097         if (!(dev->driver->driver_features & DRIVER_GEM))
1098                 return -ENODEV;
1099
1100         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1101         if (obj == NULL)
1102                 return -EBADF;
1103
1104         offset = args->offset;
1105
1106         down_write(&current->mm->mmap_sem);
1107         addr = do_mmap(obj->filp, 0, args->size,
1108                        PROT_READ | PROT_WRITE, MAP_SHARED,
1109                        args->offset);
1110         up_write(&current->mm->mmap_sem);
1111         drm_gem_object_unreference_unlocked(obj);
1112         if (IS_ERR((void *)addr))
1113                 return addr;
1114
1115         args->addr_ptr = (uint64_t) addr;
1116
1117         return 0;
1118 }
1119
1120 /**
1121  * i915_gem_fault - fault a page into the GTT
1122  * vma: VMA in question
1123  * vmf: fault info
1124  *
1125  * The fault handler is set up by drm_gem_mmap() when a object is GTT mapped
1126  * from userspace.  The fault handler takes care of binding the object to
1127  * the GTT (if needed), allocating and programming a fence register (again,
1128  * only if needed based on whether the old reg is still valid or the object
1129  * is tiled) and inserting a new PTE into the faulting process.
1130  *
1131  * Note that the faulting process may involve evicting existing objects
1132  * from the GTT and/or fence registers to make room.  So performance may
1133  * suffer if the GTT working set is large or there are few fence registers
1134  * left.
1135  */
1136 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1137 {
1138         struct drm_gem_object *obj = vma->vm_private_data;
1139         struct drm_device *dev = obj->dev;
1140         struct drm_i915_private *dev_priv = dev->dev_private;
1141         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1142         pgoff_t page_offset;
1143         unsigned long pfn;
1144         int ret = 0;
1145         bool write = !!(vmf->flags & FAULT_FLAG_WRITE);
1146
1147         /* We don't use vmf->pgoff since that has the fake offset */
1148         page_offset = ((unsigned long)vmf->virtual_address - vma->vm_start) >>
1149                 PAGE_SHIFT;
1150
1151         /* Now bind it into the GTT if needed */
1152         mutex_lock(&dev->struct_mutex);
1153         if (!obj_priv->gtt_space) {
1154                 ret = i915_gem_object_bind_to_gtt(obj, 0);
1155                 if (ret)
1156                         goto unlock;
1157
1158                 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1159
1160                 ret = i915_gem_object_set_to_gtt_domain(obj, write);
1161                 if (ret)
1162                         goto unlock;
1163         }
1164
1165         /* Need a new fence register? */
1166         if (obj_priv->tiling_mode != I915_TILING_NONE) {
1167                 ret = i915_gem_object_get_fence_reg(obj);
1168                 if (ret)
1169                         goto unlock;
1170         }
1171
1172         pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) +
1173                 page_offset;
1174
1175         /* Finally, remap it using the new GTT offset */
1176         ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
1177 unlock:
1178         mutex_unlock(&dev->struct_mutex);
1179
1180         switch (ret) {
1181         case 0:
1182         case -ERESTARTSYS:
1183                 return VM_FAULT_NOPAGE;
1184         case -ENOMEM:
1185         case -EAGAIN:
1186                 return VM_FAULT_OOM;
1187         default:
1188                 return VM_FAULT_SIGBUS;
1189         }
1190 }
1191
1192 /**
1193  * i915_gem_create_mmap_offset - create a fake mmap offset for an object
1194  * @obj: obj in question
1195  *
1196  * GEM memory mapping works by handing back to userspace a fake mmap offset
1197  * it can use in a subsequent mmap(2) call.  The DRM core code then looks
1198  * up the object based on the offset and sets up the various memory mapping
1199  * structures.
1200  *
1201  * This routine allocates and attaches a fake offset for @obj.
1202  */
1203 static int
1204 i915_gem_create_mmap_offset(struct drm_gem_object *obj)
1205 {
1206         struct drm_device *dev = obj->dev;
1207         struct drm_gem_mm *mm = dev->mm_private;
1208         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1209         struct drm_map_list *list;
1210         struct drm_local_map *map;
1211         int ret = 0;
1212
1213         /* Set the object up for mmap'ing */
1214         list = &obj->map_list;
1215         list->map = kzalloc(sizeof(struct drm_map_list), GFP_KERNEL);
1216         if (!list->map)
1217                 return -ENOMEM;
1218
1219         map = list->map;
1220         map->type = _DRM_GEM;
1221         map->size = obj->size;
1222         map->handle = obj;
1223
1224         /* Get a DRM GEM mmap offset allocated... */
1225         list->file_offset_node = drm_mm_search_free(&mm->offset_manager,
1226                                                     obj->size / PAGE_SIZE, 0, 0);
1227         if (!list->file_offset_node) {
1228                 DRM_ERROR("failed to allocate offset for bo %d\n", obj->name);
1229                 ret = -ENOMEM;
1230                 goto out_free_list;
1231         }
1232
1233         list->file_offset_node = drm_mm_get_block(list->file_offset_node,
1234                                                   obj->size / PAGE_SIZE, 0);
1235         if (!list->file_offset_node) {
1236                 ret = -ENOMEM;
1237                 goto out_free_list;
1238         }
1239
1240         list->hash.key = list->file_offset_node->start;
1241         if (drm_ht_insert_item(&mm->offset_hash, &list->hash)) {
1242                 DRM_ERROR("failed to add to map hash\n");
1243                 ret = -ENOMEM;
1244                 goto out_free_mm;
1245         }
1246
1247         /* By now we should be all set, any drm_mmap request on the offset
1248          * below will get to our mmap & fault handler */
1249         obj_priv->mmap_offset = ((uint64_t) list->hash.key) << PAGE_SHIFT;
1250
1251         return 0;
1252
1253 out_free_mm:
1254         drm_mm_put_block(list->file_offset_node);
1255 out_free_list:
1256         kfree(list->map);
1257
1258         return ret;
1259 }
1260
1261 /**
1262  * i915_gem_release_mmap - remove physical page mappings
1263  * @obj: obj in question
1264  *
1265  * Preserve the reservation of the mmapping with the DRM core code, but
1266  * relinquish ownership of the pages back to the system.
1267  *
1268  * It is vital that we remove the page mapping if we have mapped a tiled
1269  * object through the GTT and then lose the fence register due to
1270  * resource pressure. Similarly if the object has been moved out of the
1271  * aperture, than pages mapped into userspace must be revoked. Removing the
1272  * mapping will then trigger a page fault on the next user access, allowing
1273  * fixup by i915_gem_fault().
1274  */
1275 void
1276 i915_gem_release_mmap(struct drm_gem_object *obj)
1277 {
1278         struct drm_device *dev = obj->dev;
1279         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1280
1281         if (dev->dev_mapping)
1282                 unmap_mapping_range(dev->dev_mapping,
1283                                     obj_priv->mmap_offset, obj->size, 1);
1284 }
1285
1286 static void
1287 i915_gem_free_mmap_offset(struct drm_gem_object *obj)
1288 {
1289         struct drm_device *dev = obj->dev;
1290         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1291         struct drm_gem_mm *mm = dev->mm_private;
1292         struct drm_map_list *list;
1293
1294         list = &obj->map_list;
1295         drm_ht_remove_item(&mm->offset_hash, &list->hash);
1296
1297         if (list->file_offset_node) {
1298                 drm_mm_put_block(list->file_offset_node);
1299                 list->file_offset_node = NULL;
1300         }
1301
1302         if (list->map) {
1303                 kfree(list->map);
1304                 list->map = NULL;
1305         }
1306
1307         obj_priv->mmap_offset = 0;
1308 }
1309
1310 /**
1311  * i915_gem_get_gtt_alignment - return required GTT alignment for an object
1312  * @obj: object to check
1313  *
1314  * Return the required GTT alignment for an object, taking into account
1315  * potential fence register mapping if needed.
1316  */
1317 static uint32_t
1318 i915_gem_get_gtt_alignment(struct drm_gem_object *obj)
1319 {
1320         struct drm_device *dev = obj->dev;
1321         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1322         int start, i;
1323
1324         /*
1325          * Minimum alignment is 4k (GTT page size), but might be greater
1326          * if a fence register is needed for the object.
1327          */
1328         if (IS_I965G(dev) || obj_priv->tiling_mode == I915_TILING_NONE)
1329                 return 4096;
1330
1331         /*
1332          * Previous chips need to be aligned to the size of the smallest
1333          * fence register that can contain the object.
1334          */
1335         if (IS_I9XX(dev))
1336                 start = 1024*1024;
1337         else
1338                 start = 512*1024;
1339
1340         for (i = start; i < obj->size; i <<= 1)
1341                 ;
1342
1343         return i;
1344 }
1345
1346 /**
1347  * i915_gem_mmap_gtt_ioctl - prepare an object for GTT mmap'ing
1348  * @dev: DRM device
1349  * @data: GTT mapping ioctl data
1350  * @file_priv: GEM object info
1351  *
1352  * Simply returns the fake offset to userspace so it can mmap it.
1353  * The mmap call will end up in drm_gem_mmap(), which will set things
1354  * up so we can get faults in the handler above.
1355  *
1356  * The fault handler will take care of binding the object into the GTT
1357  * (since it may have been evicted to make room for something), allocating
1358  * a fence register, and mapping the appropriate aperture address into
1359  * userspace.
1360  */
1361 int
1362 i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data,
1363                         struct drm_file *file_priv)
1364 {
1365         struct drm_i915_gem_mmap_gtt *args = data;
1366         struct drm_i915_private *dev_priv = dev->dev_private;
1367         struct drm_gem_object *obj;
1368         struct drm_i915_gem_object *obj_priv;
1369         int ret;
1370
1371         if (!(dev->driver->driver_features & DRIVER_GEM))
1372                 return -ENODEV;
1373
1374         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
1375         if (obj == NULL)
1376                 return -EBADF;
1377
1378         mutex_lock(&dev->struct_mutex);
1379
1380         obj_priv = to_intel_bo(obj);
1381
1382         if (obj_priv->madv != I915_MADV_WILLNEED) {
1383                 DRM_ERROR("Attempting to mmap a purgeable buffer\n");
1384                 drm_gem_object_unreference(obj);
1385                 mutex_unlock(&dev->struct_mutex);
1386                 return -EINVAL;
1387         }
1388
1389
1390         if (!obj_priv->mmap_offset) {
1391                 ret = i915_gem_create_mmap_offset(obj);
1392                 if (ret) {
1393                         drm_gem_object_unreference(obj);
1394                         mutex_unlock(&dev->struct_mutex);
1395                         return ret;
1396                 }
1397         }
1398
1399         args->offset = obj_priv->mmap_offset;
1400
1401         /*
1402          * Pull it into the GTT so that we have a page list (makes the
1403          * initial fault faster and any subsequent flushing possible).
1404          */
1405         if (!obj_priv->agp_mem) {
1406                 ret = i915_gem_object_bind_to_gtt(obj, 0);
1407                 if (ret) {
1408                         drm_gem_object_unreference(obj);
1409                         mutex_unlock(&dev->struct_mutex);
1410                         return ret;
1411                 }
1412                 list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1413         }
1414
1415         drm_gem_object_unreference(obj);
1416         mutex_unlock(&dev->struct_mutex);
1417
1418         return 0;
1419 }
1420
1421 void
1422 i915_gem_object_put_pages(struct drm_gem_object *obj)
1423 {
1424         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1425         int page_count = obj->size / PAGE_SIZE;
1426         int i;
1427
1428         BUG_ON(obj_priv->pages_refcount == 0);
1429         BUG_ON(obj_priv->madv == __I915_MADV_PURGED);
1430
1431         if (--obj_priv->pages_refcount != 0)
1432                 return;
1433
1434         if (obj_priv->tiling_mode != I915_TILING_NONE)
1435                 i915_gem_object_save_bit_17_swizzle(obj);
1436
1437         if (obj_priv->madv == I915_MADV_DONTNEED)
1438                 obj_priv->dirty = 0;
1439
1440         for (i = 0; i < page_count; i++) {
1441                 if (obj_priv->dirty)
1442                         set_page_dirty(obj_priv->pages[i]);
1443
1444                 if (obj_priv->madv == I915_MADV_WILLNEED)
1445                         mark_page_accessed(obj_priv->pages[i]);
1446
1447                 page_cache_release(obj_priv->pages[i]);
1448         }
1449         obj_priv->dirty = 0;
1450
1451         drm_free_large(obj_priv->pages);
1452         obj_priv->pages = NULL;
1453 }
1454
1455 static void
1456 i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno,
1457                                struct intel_ring_buffer *ring)
1458 {
1459         struct drm_device *dev = obj->dev;
1460         drm_i915_private_t *dev_priv = dev->dev_private;
1461         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1462         BUG_ON(ring == NULL);
1463         obj_priv->ring = ring;
1464
1465         /* Add a reference if we're newly entering the active list. */
1466         if (!obj_priv->active) {
1467                 drm_gem_object_reference(obj);
1468                 obj_priv->active = 1;
1469         }
1470         /* Move from whatever list we were on to the tail of execution. */
1471         spin_lock(&dev_priv->mm.active_list_lock);
1472         list_move_tail(&obj_priv->list, &ring->active_list);
1473         spin_unlock(&dev_priv->mm.active_list_lock);
1474         obj_priv->last_rendering_seqno = seqno;
1475 }
1476
1477 static void
1478 i915_gem_object_move_to_flushing(struct drm_gem_object *obj)
1479 {
1480         struct drm_device *dev = obj->dev;
1481         drm_i915_private_t *dev_priv = dev->dev_private;
1482         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1483
1484         BUG_ON(!obj_priv->active);
1485         list_move_tail(&obj_priv->list, &dev_priv->mm.flushing_list);
1486         obj_priv->last_rendering_seqno = 0;
1487 }
1488
1489 /* Immediately discard the backing storage */
1490 static void
1491 i915_gem_object_truncate(struct drm_gem_object *obj)
1492 {
1493         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1494         struct inode *inode;
1495
1496         inode = obj->filp->f_path.dentry->d_inode;
1497         if (inode->i_op->truncate)
1498                 inode->i_op->truncate (inode);
1499
1500         obj_priv->madv = __I915_MADV_PURGED;
1501 }
1502
1503 static inline int
1504 i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv)
1505 {
1506         return obj_priv->madv == I915_MADV_DONTNEED;
1507 }
1508
1509 static void
1510 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
1511 {
1512         struct drm_device *dev = obj->dev;
1513         drm_i915_private_t *dev_priv = dev->dev_private;
1514         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1515
1516         i915_verify_inactive(dev, __FILE__, __LINE__);
1517         if (obj_priv->pin_count != 0)
1518                 list_del_init(&obj_priv->list);
1519         else
1520                 list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
1521
1522         BUG_ON(!list_empty(&obj_priv->gpu_write_list));
1523
1524         obj_priv->last_rendering_seqno = 0;
1525         obj_priv->ring = NULL;
1526         if (obj_priv->active) {
1527                 obj_priv->active = 0;
1528                 drm_gem_object_unreference(obj);
1529         }
1530         i915_verify_inactive(dev, __FILE__, __LINE__);
1531 }
1532
1533 static void
1534 i915_gem_process_flushing_list(struct drm_device *dev,
1535                                uint32_t flush_domains, uint32_t seqno,
1536                                struct intel_ring_buffer *ring)
1537 {
1538         drm_i915_private_t *dev_priv = dev->dev_private;
1539         struct drm_i915_gem_object *obj_priv, *next;
1540
1541         list_for_each_entry_safe(obj_priv, next,
1542                                  &dev_priv->mm.gpu_write_list,
1543                                  gpu_write_list) {
1544                 struct drm_gem_object *obj = &obj_priv->base;
1545
1546                 if ((obj->write_domain & flush_domains) ==
1547                     obj->write_domain &&
1548                     obj_priv->ring->ring_flag == ring->ring_flag) {
1549                         uint32_t old_write_domain = obj->write_domain;
1550
1551                         obj->write_domain = 0;
1552                         list_del_init(&obj_priv->gpu_write_list);
1553                         i915_gem_object_move_to_active(obj, seqno, ring);
1554
1555                         /* update the fence lru list */
1556                         if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
1557                                 struct drm_i915_fence_reg *reg =
1558                                         &dev_priv->fence_regs[obj_priv->fence_reg];
1559                                 list_move_tail(&reg->lru_list,
1560                                                 &dev_priv->mm.fence_list);
1561                         }
1562
1563                         trace_i915_gem_object_change_domain(obj,
1564                                                             obj->read_domains,
1565                                                             old_write_domain);
1566                 }
1567         }
1568 }
1569
1570 uint32_t
1571 i915_add_request(struct drm_device *dev, struct drm_file *file_priv,
1572                  uint32_t flush_domains, struct intel_ring_buffer *ring)
1573 {
1574         drm_i915_private_t *dev_priv = dev->dev_private;
1575         struct drm_i915_file_private *i915_file_priv = NULL;
1576         struct drm_i915_gem_request *request;
1577         uint32_t seqno;
1578         int was_empty;
1579
1580         if (file_priv != NULL)
1581                 i915_file_priv = file_priv->driver_priv;
1582
1583         request = kzalloc(sizeof(*request), GFP_KERNEL);
1584         if (request == NULL)
1585                 return 0;
1586
1587         seqno = ring->add_request(dev, ring, file_priv, flush_domains);
1588
1589         request->seqno = seqno;
1590         request->ring = ring;
1591         request->emitted_jiffies = jiffies;
1592         was_empty = list_empty(&ring->request_list);
1593         list_add_tail(&request->list, &ring->request_list);
1594
1595         if (i915_file_priv) {
1596                 list_add_tail(&request->client_list,
1597                               &i915_file_priv->mm.request_list);
1598         } else {
1599                 INIT_LIST_HEAD(&request->client_list);
1600         }
1601
1602         /* Associate any objects on the flushing list matching the write
1603          * domain we're flushing with our flush.
1604          */
1605         if (flush_domains != 0) 
1606                 i915_gem_process_flushing_list(dev, flush_domains, seqno, ring);
1607
1608         if (!dev_priv->mm.suspended) {
1609                 mod_timer(&dev_priv->hangcheck_timer, jiffies + DRM_I915_HANGCHECK_PERIOD);
1610                 if (was_empty)
1611                         queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1612         }
1613         return seqno;
1614 }
1615
1616 /**
1617  * Command execution barrier
1618  *
1619  * Ensures that all commands in the ring are finished
1620  * before signalling the CPU
1621  */
1622 static uint32_t
1623 i915_retire_commands(struct drm_device *dev, struct intel_ring_buffer *ring)
1624 {
1625         uint32_t flush_domains = 0;
1626
1627         /* The sampler always gets flushed on i965 (sigh) */
1628         if (IS_I965G(dev))
1629                 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
1630
1631         ring->flush(dev, ring,
1632                         I915_GEM_DOMAIN_COMMAND, flush_domains);
1633         return flush_domains;
1634 }
1635
1636 /**
1637  * Moves buffers associated only with the given active seqno from the active
1638  * to inactive list, potentially freeing them.
1639  */
1640 static void
1641 i915_gem_retire_request(struct drm_device *dev,
1642                         struct drm_i915_gem_request *request)
1643 {
1644         drm_i915_private_t *dev_priv = dev->dev_private;
1645
1646         trace_i915_gem_request_retire(dev, request->seqno);
1647
1648         /* Move any buffers on the active list that are no longer referenced
1649          * by the ringbuffer to the flushing/inactive lists as appropriate.
1650          */
1651         spin_lock(&dev_priv->mm.active_list_lock);
1652         while (!list_empty(&request->ring->active_list)) {
1653                 struct drm_gem_object *obj;
1654                 struct drm_i915_gem_object *obj_priv;
1655
1656                 obj_priv = list_first_entry(&request->ring->active_list,
1657                                             struct drm_i915_gem_object,
1658                                             list);
1659                 obj = &obj_priv->base;
1660
1661                 /* If the seqno being retired doesn't match the oldest in the
1662                  * list, then the oldest in the list must still be newer than
1663                  * this seqno.
1664                  */
1665                 if (obj_priv->last_rendering_seqno != request->seqno)
1666                         goto out;
1667
1668 #if WATCH_LRU
1669                 DRM_INFO("%s: retire %d moves to inactive list %p\n",
1670                          __func__, request->seqno, obj);
1671 #endif
1672
1673                 if (obj->write_domain != 0)
1674                         i915_gem_object_move_to_flushing(obj);
1675                 else {
1676                         /* Take a reference on the object so it won't be
1677                          * freed while the spinlock is held.  The list
1678                          * protection for this spinlock is safe when breaking
1679                          * the lock like this since the next thing we do
1680                          * is just get the head of the list again.
1681                          */
1682                         drm_gem_object_reference(obj);
1683                         i915_gem_object_move_to_inactive(obj);
1684                         spin_unlock(&dev_priv->mm.active_list_lock);
1685                         drm_gem_object_unreference(obj);
1686                         spin_lock(&dev_priv->mm.active_list_lock);
1687                 }
1688         }
1689 out:
1690         spin_unlock(&dev_priv->mm.active_list_lock);
1691 }
1692
1693 /**
1694  * Returns true if seq1 is later than seq2.
1695  */
1696 bool
1697 i915_seqno_passed(uint32_t seq1, uint32_t seq2)
1698 {
1699         return (int32_t)(seq1 - seq2) >= 0;
1700 }
1701
1702 uint32_t
1703 i915_get_gem_seqno(struct drm_device *dev,
1704                    struct intel_ring_buffer *ring)
1705 {
1706         return ring->get_gem_seqno(dev, ring);
1707 }
1708
1709 /**
1710  * This function clears the request list as sequence numbers are passed.
1711  */
1712 void
1713 i915_gem_retire_requests(struct drm_device *dev,
1714                 struct intel_ring_buffer *ring)
1715 {
1716         drm_i915_private_t *dev_priv = dev->dev_private;
1717         uint32_t seqno;
1718
1719         if (!ring->status_page.page_addr
1720                         || list_empty(&ring->request_list))
1721                 return;
1722
1723         seqno = i915_get_gem_seqno(dev, ring);
1724
1725         while (!list_empty(&ring->request_list)) {
1726                 struct drm_i915_gem_request *request;
1727                 uint32_t retiring_seqno;
1728
1729                 request = list_first_entry(&ring->request_list,
1730                                            struct drm_i915_gem_request,
1731                                            list);
1732                 retiring_seqno = request->seqno;
1733
1734                 if (i915_seqno_passed(seqno, retiring_seqno) ||
1735                     atomic_read(&dev_priv->mm.wedged)) {
1736                         i915_gem_retire_request(dev, request);
1737
1738                         list_del(&request->list);
1739                         list_del(&request->client_list);
1740                         kfree(request);
1741                 } else
1742                         break;
1743         }
1744
1745         if (unlikely (dev_priv->trace_irq_seqno &&
1746                       i915_seqno_passed(dev_priv->trace_irq_seqno, seqno))) {
1747
1748                 ring->user_irq_put(dev, ring);
1749                 dev_priv->trace_irq_seqno = 0;
1750         }
1751 }
1752
1753 void
1754 i915_gem_retire_work_handler(struct work_struct *work)
1755 {
1756         drm_i915_private_t *dev_priv;
1757         struct drm_device *dev;
1758
1759         dev_priv = container_of(work, drm_i915_private_t,
1760                                 mm.retire_work.work);
1761         dev = dev_priv->dev;
1762
1763         mutex_lock(&dev->struct_mutex);
1764         i915_gem_retire_requests(dev, &dev_priv->render_ring);
1765
1766         if (HAS_BSD(dev))
1767                 i915_gem_retire_requests(dev, &dev_priv->bsd_ring);
1768
1769         if (!dev_priv->mm.suspended &&
1770                 (!list_empty(&dev_priv->render_ring.request_list) ||
1771                         (HAS_BSD(dev) &&
1772                          !list_empty(&dev_priv->bsd_ring.request_list))))
1773                 queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ);
1774         mutex_unlock(&dev->struct_mutex);
1775 }
1776
1777 int
1778 i915_do_wait_request(struct drm_device *dev, uint32_t seqno,
1779                 int interruptible, struct intel_ring_buffer *ring)
1780 {
1781         drm_i915_private_t *dev_priv = dev->dev_private;
1782         u32 ier;
1783         int ret = 0;
1784
1785         BUG_ON(seqno == 0);
1786
1787         if (atomic_read(&dev_priv->mm.wedged))
1788                 return -EIO;
1789
1790         if (!i915_seqno_passed(ring->get_gem_seqno(dev, ring), seqno)) {
1791                 if (HAS_PCH_SPLIT(dev))
1792                         ier = I915_READ(DEIER) | I915_READ(GTIER);
1793                 else
1794                         ier = I915_READ(IER);
1795                 if (!ier) {
1796                         DRM_ERROR("something (likely vbetool) disabled "
1797                                   "interrupts, re-enabling\n");
1798                         i915_driver_irq_preinstall(dev);
1799                         i915_driver_irq_postinstall(dev);
1800                 }
1801
1802                 trace_i915_gem_request_wait_begin(dev, seqno);
1803
1804                 ring->waiting_gem_seqno = seqno;
1805                 ring->user_irq_get(dev, ring);
1806                 if (interruptible)
1807                         ret = wait_event_interruptible(ring->irq_queue,
1808                                 i915_seqno_passed(
1809                                         ring->get_gem_seqno(dev, ring), seqno)
1810                                 || atomic_read(&dev_priv->mm.wedged));
1811                 else
1812                         wait_event(ring->irq_queue,
1813                                 i915_seqno_passed(
1814                                         ring->get_gem_seqno(dev, ring), seqno)
1815                                 || atomic_read(&dev_priv->mm.wedged));
1816
1817                 ring->user_irq_put(dev, ring);
1818                 ring->waiting_gem_seqno = 0;
1819
1820                 trace_i915_gem_request_wait_end(dev, seqno);
1821         }
1822         if (atomic_read(&dev_priv->mm.wedged))
1823                 ret = -EIO;
1824
1825         if (ret && ret != -ERESTARTSYS)
1826                 DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
1827                           __func__, ret, seqno, ring->get_gem_seqno(dev, ring));
1828
1829         /* Directly dispatch request retiring.  While we have the work queue
1830          * to handle this, the waiter on a request often wants an associated
1831          * buffer to have made it to the inactive list, and we would need
1832          * a separate wait queue to handle that.
1833          */
1834         if (ret == 0)
1835                 i915_gem_retire_requests(dev, ring);
1836
1837         return ret;
1838 }
1839
1840 /**
1841  * Waits for a sequence number to be signaled, and cleans up the
1842  * request and object lists appropriately for that event.
1843  */
1844 static int
1845 i915_wait_request(struct drm_device *dev, uint32_t seqno,
1846                 struct intel_ring_buffer *ring)
1847 {
1848         return i915_do_wait_request(dev, seqno, 1, ring);
1849 }
1850
1851 static void
1852 i915_gem_flush(struct drm_device *dev,
1853                uint32_t invalidate_domains,
1854                uint32_t flush_domains)
1855 {
1856         drm_i915_private_t *dev_priv = dev->dev_private;
1857         if (flush_domains & I915_GEM_DOMAIN_CPU)
1858                 drm_agp_chipset_flush(dev);
1859         dev_priv->render_ring.flush(dev, &dev_priv->render_ring,
1860                         invalidate_domains,
1861                         flush_domains);
1862
1863         if (HAS_BSD(dev))
1864                 dev_priv->bsd_ring.flush(dev, &dev_priv->bsd_ring,
1865                                 invalidate_domains,
1866                                 flush_domains);
1867 }
1868
1869 static void
1870 i915_gem_flush_ring(struct drm_device *dev,
1871                uint32_t invalidate_domains,
1872                uint32_t flush_domains,
1873                struct intel_ring_buffer *ring)
1874 {
1875         if (flush_domains & I915_GEM_DOMAIN_CPU)
1876                 drm_agp_chipset_flush(dev);
1877         ring->flush(dev, ring,
1878                         invalidate_domains,
1879                         flush_domains);
1880 }
1881
1882 /**
1883  * Ensures that all rendering to the object has completed and the object is
1884  * safe to unbind from the GTT or access from the CPU.
1885  */
1886 static int
1887 i915_gem_object_wait_rendering(struct drm_gem_object *obj)
1888 {
1889         struct drm_device *dev = obj->dev;
1890         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1891         int ret;
1892
1893         /* This function only exists to support waiting for existing rendering,
1894          * not for emitting required flushes.
1895          */
1896         BUG_ON((obj->write_domain & I915_GEM_GPU_DOMAINS) != 0);
1897
1898         /* If there is rendering queued on the buffer being evicted, wait for
1899          * it.
1900          */
1901         if (obj_priv->active) {
1902 #if WATCH_BUF
1903                 DRM_INFO("%s: object %p wait for seqno %08x\n",
1904                           __func__, obj, obj_priv->last_rendering_seqno);
1905 #endif
1906                 ret = i915_wait_request(dev,
1907                                 obj_priv->last_rendering_seqno, obj_priv->ring);
1908                 if (ret != 0)
1909                         return ret;
1910         }
1911
1912         return 0;
1913 }
1914
1915 /**
1916  * Unbinds an object from the GTT aperture.
1917  */
1918 int
1919 i915_gem_object_unbind(struct drm_gem_object *obj)
1920 {
1921         struct drm_device *dev = obj->dev;
1922         drm_i915_private_t *dev_priv = dev->dev_private;
1923         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
1924         int ret = 0;
1925
1926 #if WATCH_BUF
1927         DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
1928         DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
1929 #endif
1930         if (obj_priv->gtt_space == NULL)
1931                 return 0;
1932
1933         if (obj_priv->pin_count != 0) {
1934                 DRM_ERROR("Attempting to unbind pinned buffer\n");
1935                 return -EINVAL;
1936         }
1937
1938         /* blow away mappings if mapped through GTT */
1939         i915_gem_release_mmap(obj);
1940
1941         /* Move the object to the CPU domain to ensure that
1942          * any possible CPU writes while it's not in the GTT
1943          * are flushed when we go to remap it. This will
1944          * also ensure that all pending GPU writes are finished
1945          * before we unbind.
1946          */
1947         ret = i915_gem_object_set_to_cpu_domain(obj, 1);
1948         if (ret) {
1949                 if (ret != -ERESTARTSYS)
1950                         DRM_ERROR("set_domain failed: %d\n", ret);
1951                 return ret;
1952         }
1953
1954         BUG_ON(obj_priv->active);
1955
1956         /* release the fence reg _after_ flushing */
1957         if (obj_priv->fence_reg != I915_FENCE_REG_NONE)
1958                 i915_gem_clear_fence_reg(obj);
1959
1960         if (obj_priv->agp_mem != NULL) {
1961                 drm_unbind_agp(obj_priv->agp_mem);
1962                 drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
1963                 obj_priv->agp_mem = NULL;
1964         }
1965
1966         i915_gem_object_put_pages(obj);
1967         BUG_ON(obj_priv->pages_refcount);
1968
1969         if (obj_priv->gtt_space) {
1970                 atomic_dec(&dev->gtt_count);
1971                 atomic_sub(obj->size, &dev->gtt_memory);
1972
1973                 drm_mm_put_block(obj_priv->gtt_space);
1974                 obj_priv->gtt_space = NULL;
1975         }
1976
1977         /* Remove ourselves from the LRU list if present. */
1978         spin_lock(&dev_priv->mm.active_list_lock);
1979         if (!list_empty(&obj_priv->list))
1980                 list_del_init(&obj_priv->list);
1981         spin_unlock(&dev_priv->mm.active_list_lock);
1982
1983         if (i915_gem_object_is_purgeable(obj_priv))
1984                 i915_gem_object_truncate(obj);
1985
1986         trace_i915_gem_object_unbind(obj);
1987
1988         return 0;
1989 }
1990
1991 static struct drm_gem_object *
1992 i915_gem_find_inactive_object(struct drm_device *dev, int min_size)
1993 {
1994         drm_i915_private_t *dev_priv = dev->dev_private;
1995         struct drm_i915_gem_object *obj_priv;
1996         struct drm_gem_object *best = NULL;
1997         struct drm_gem_object *first = NULL;
1998
1999         /* Try to find the smallest clean object */
2000         list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) {
2001                 struct drm_gem_object *obj = &obj_priv->base;
2002                 if (obj->size >= min_size) {
2003                         if ((!obj_priv->dirty ||
2004                              i915_gem_object_is_purgeable(obj_priv)) &&
2005                             (!best || obj->size < best->size)) {
2006                                 best = obj;
2007                                 if (best->size == min_size)
2008                                         return best;
2009                         }
2010                         if (!first)
2011                             first = obj;
2012                 }
2013         }
2014
2015         return best ? best : first;
2016 }
2017
2018 static int
2019 i915_gpu_idle(struct drm_device *dev)
2020 {
2021         drm_i915_private_t *dev_priv = dev->dev_private;
2022         bool lists_empty;
2023         uint32_t seqno1, seqno2;
2024         int ret;
2025
2026         spin_lock(&dev_priv->mm.active_list_lock);
2027         lists_empty = (list_empty(&dev_priv->mm.flushing_list) &&
2028                        list_empty(&dev_priv->render_ring.active_list) &&
2029                        (!HAS_BSD(dev) ||
2030                         list_empty(&dev_priv->bsd_ring.active_list)));
2031         spin_unlock(&dev_priv->mm.active_list_lock);
2032
2033         if (lists_empty)
2034                 return 0;
2035
2036         /* Flush everything onto the inactive list. */
2037         i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS);
2038         seqno1 = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS,
2039                         &dev_priv->render_ring);
2040         if (seqno1 == 0)
2041                 return -ENOMEM;
2042         ret = i915_wait_request(dev, seqno1, &dev_priv->render_ring);
2043
2044         if (HAS_BSD(dev)) {
2045                 seqno2 = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS,
2046                                 &dev_priv->bsd_ring);
2047                 if (seqno2 == 0)
2048                         return -ENOMEM;
2049
2050                 ret = i915_wait_request(dev, seqno2, &dev_priv->bsd_ring);
2051                 if (ret)
2052                         return ret;
2053         }
2054
2055
2056         return ret;
2057 }
2058
2059 static int
2060 i915_gem_evict_everything(struct drm_device *dev)
2061 {
2062         drm_i915_private_t *dev_priv = dev->dev_private;
2063         int ret;
2064         bool lists_empty;
2065
2066         spin_lock(&dev_priv->mm.active_list_lock);
2067         lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2068                        list_empty(&dev_priv->mm.flushing_list) &&
2069                        list_empty(&dev_priv->render_ring.active_list) &&
2070                        (!HAS_BSD(dev)
2071                         || list_empty(&dev_priv->bsd_ring.active_list)));
2072         spin_unlock(&dev_priv->mm.active_list_lock);
2073
2074         if (lists_empty)
2075                 return -ENOSPC;
2076
2077         /* Flush everything (on to the inactive lists) and evict */
2078         ret = i915_gpu_idle(dev);
2079         if (ret)
2080                 return ret;
2081
2082         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2083
2084         ret = i915_gem_evict_from_inactive_list(dev);
2085         if (ret)
2086                 return ret;
2087
2088         spin_lock(&dev_priv->mm.active_list_lock);
2089         lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
2090                        list_empty(&dev_priv->mm.flushing_list) &&
2091                        list_empty(&dev_priv->render_ring.active_list) &&
2092                        (!HAS_BSD(dev)
2093                         || list_empty(&dev_priv->bsd_ring.active_list)));
2094         spin_unlock(&dev_priv->mm.active_list_lock);
2095         BUG_ON(!lists_empty);
2096
2097         return 0;
2098 }
2099
2100 static int
2101 i915_gem_evict_something(struct drm_device *dev, int min_size)
2102 {
2103         drm_i915_private_t *dev_priv = dev->dev_private;
2104         struct drm_gem_object *obj;
2105         int ret;
2106
2107         struct intel_ring_buffer *render_ring = &dev_priv->render_ring;
2108         struct intel_ring_buffer *bsd_ring = &dev_priv->bsd_ring;
2109         for (;;) {
2110                 i915_gem_retire_requests(dev, render_ring);
2111
2112                 if (HAS_BSD(dev))
2113                         i915_gem_retire_requests(dev, bsd_ring);
2114
2115                 /* If there's an inactive buffer available now, grab it
2116                  * and be done.
2117                  */
2118                 obj = i915_gem_find_inactive_object(dev, min_size);
2119                 if (obj) {
2120                         struct drm_i915_gem_object *obj_priv;
2121
2122 #if WATCH_LRU
2123                         DRM_INFO("%s: evicting %p\n", __func__, obj);
2124 #endif
2125                         obj_priv = to_intel_bo(obj);
2126                         BUG_ON(obj_priv->pin_count != 0);
2127                         BUG_ON(obj_priv->active);
2128
2129                         /* Wait on the rendering and unbind the buffer. */
2130                         return i915_gem_object_unbind(obj);
2131                 }
2132
2133                 /* If we didn't get anything, but the ring is still processing
2134                  * things, wait for the next to finish and hopefully leave us
2135                  * a buffer to evict.
2136                  */
2137                 if (!list_empty(&render_ring->request_list)) {
2138                         struct drm_i915_gem_request *request;
2139
2140                         request = list_first_entry(&render_ring->request_list,
2141                                                    struct drm_i915_gem_request,
2142                                                    list);
2143
2144                         ret = i915_wait_request(dev,
2145                                         request->seqno, request->ring);
2146                         if (ret)
2147                                 return ret;
2148
2149                         continue;
2150                 }
2151
2152                 if (HAS_BSD(dev) && !list_empty(&bsd_ring->request_list)) {
2153                         struct drm_i915_gem_request *request;
2154
2155                         request = list_first_entry(&bsd_ring->request_list,
2156                                                    struct drm_i915_gem_request,
2157                                                    list);
2158
2159                         ret = i915_wait_request(dev,
2160                                         request->seqno, request->ring);
2161                         if (ret)
2162                                 return ret;
2163
2164                         continue;
2165                 }
2166
2167                 /* If we didn't have anything on the request list but there
2168                  * are buffers awaiting a flush, emit one and try again.
2169                  * When we wait on it, those buffers waiting for that flush
2170                  * will get moved to inactive.
2171                  */
2172                 if (!list_empty(&dev_priv->mm.flushing_list)) {
2173                         struct drm_i915_gem_object *obj_priv;
2174
2175                         /* Find an object that we can immediately reuse */
2176                         list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) {
2177                                 obj = &obj_priv->base;
2178                                 if (obj->size >= min_size)
2179                                         break;
2180
2181                                 obj = NULL;
2182                         }
2183
2184                         if (obj != NULL) {
2185                                 uint32_t seqno;
2186
2187                                 i915_gem_flush_ring(dev,
2188                                                obj->write_domain,
2189                                                obj->write_domain,
2190                                                obj_priv->ring);
2191                                 seqno = i915_add_request(dev, NULL,
2192                                                 obj->write_domain,
2193                                                 obj_priv->ring);
2194                                 if (seqno == 0)
2195                                         return -ENOMEM;
2196                                 continue;
2197                         }
2198                 }
2199
2200                 /* If we didn't do any of the above, there's no single buffer
2201                  * large enough to swap out for the new one, so just evict
2202                  * everything and start again. (This should be rare.)
2203                  */
2204                 if (!list_empty (&dev_priv->mm.inactive_list))
2205                         return i915_gem_evict_from_inactive_list(dev);
2206                 else
2207                         return i915_gem_evict_everything(dev);
2208         }
2209 }
2210
2211 int
2212 i915_gem_object_get_pages(struct drm_gem_object *obj,
2213                           gfp_t gfpmask)
2214 {
2215         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2216         int page_count, i;
2217         struct address_space *mapping;
2218         struct inode *inode;
2219         struct page *page;
2220
2221         BUG_ON(obj_priv->pages_refcount
2222                         == DRM_I915_GEM_OBJECT_MAX_PAGES_REFCOUNT);
2223
2224         if (obj_priv->pages_refcount++ != 0)
2225                 return 0;
2226
2227         /* Get the list of pages out of our struct file.  They'll be pinned
2228          * at this point until we release them.
2229          */
2230         page_count = obj->size / PAGE_SIZE;
2231         BUG_ON(obj_priv->pages != NULL);
2232         obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *));
2233         if (obj_priv->pages == NULL) {
2234                 obj_priv->pages_refcount--;
2235                 return -ENOMEM;
2236         }
2237
2238         inode = obj->filp->f_path.dentry->d_inode;
2239         mapping = inode->i_mapping;
2240         for (i = 0; i < page_count; i++) {
2241                 page = read_cache_page_gfp(mapping, i,
2242                                            GFP_HIGHUSER |
2243                                            __GFP_COLD |
2244                                            __GFP_RECLAIMABLE |
2245                                            gfpmask);
2246                 if (IS_ERR(page))
2247                         goto err_pages;
2248
2249                 obj_priv->pages[i] = page;
2250         }
2251
2252         if (obj_priv->tiling_mode != I915_TILING_NONE)
2253                 i915_gem_object_do_bit_17_swizzle(obj);
2254
2255         return 0;
2256
2257 err_pages:
2258         while (i--)
2259                 page_cache_release(obj_priv->pages[i]);
2260
2261         drm_free_large(obj_priv->pages);
2262         obj_priv->pages = NULL;
2263         obj_priv->pages_refcount--;
2264         return PTR_ERR(page);
2265 }
2266
2267 static void sandybridge_write_fence_reg(struct drm_i915_fence_reg *reg)
2268 {
2269         struct drm_gem_object *obj = reg->obj;
2270         struct drm_device *dev = obj->dev;
2271         drm_i915_private_t *dev_priv = dev->dev_private;
2272         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2273         int regnum = obj_priv->fence_reg;
2274         uint64_t val;
2275
2276         val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2277                     0xfffff000) << 32;
2278         val |= obj_priv->gtt_offset & 0xfffff000;
2279         val |= (uint64_t)((obj_priv->stride / 128) - 1) <<
2280                 SANDYBRIDGE_FENCE_PITCH_SHIFT;
2281
2282         if (obj_priv->tiling_mode == I915_TILING_Y)
2283                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2284         val |= I965_FENCE_REG_VALID;
2285
2286         I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 + (regnum * 8), val);
2287 }
2288
2289 static void i965_write_fence_reg(struct drm_i915_fence_reg *reg)
2290 {
2291         struct drm_gem_object *obj = reg->obj;
2292         struct drm_device *dev = obj->dev;
2293         drm_i915_private_t *dev_priv = dev->dev_private;
2294         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2295         int regnum = obj_priv->fence_reg;
2296         uint64_t val;
2297
2298         val = (uint64_t)((obj_priv->gtt_offset + obj->size - 4096) &
2299                     0xfffff000) << 32;
2300         val |= obj_priv->gtt_offset & 0xfffff000;
2301         val |= ((obj_priv->stride / 128) - 1) << I965_FENCE_PITCH_SHIFT;
2302         if (obj_priv->tiling_mode == I915_TILING_Y)
2303                 val |= 1 << I965_FENCE_TILING_Y_SHIFT;
2304         val |= I965_FENCE_REG_VALID;
2305
2306         I915_WRITE64(FENCE_REG_965_0 + (regnum * 8), val);
2307 }
2308
2309 static void i915_write_fence_reg(struct drm_i915_fence_reg *reg)
2310 {
2311         struct drm_gem_object *obj = reg->obj;
2312         struct drm_device *dev = obj->dev;
2313         drm_i915_private_t *dev_priv = dev->dev_private;
2314         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2315         int regnum = obj_priv->fence_reg;
2316         int tile_width;
2317         uint32_t fence_reg, val;
2318         uint32_t pitch_val;
2319
2320         if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) ||
2321             (obj_priv->gtt_offset & (obj->size - 1))) {
2322                 WARN(1, "%s: object 0x%08x not 1M or size (0x%zx) aligned\n",
2323                      __func__, obj_priv->gtt_offset, obj->size);
2324                 return;
2325         }
2326
2327         if (obj_priv->tiling_mode == I915_TILING_Y &&
2328             HAS_128_BYTE_Y_TILING(dev))
2329                 tile_width = 128;
2330         else
2331                 tile_width = 512;
2332
2333         /* Note: pitch better be a power of two tile widths */
2334         pitch_val = obj_priv->stride / tile_width;
2335         pitch_val = ffs(pitch_val) - 1;
2336
2337         if (obj_priv->tiling_mode == I915_TILING_Y &&
2338             HAS_128_BYTE_Y_TILING(dev))
2339                 WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2340         else
2341                 WARN_ON(pitch_val > I915_FENCE_MAX_PITCH_VAL);
2342
2343         val = obj_priv->gtt_offset;
2344         if (obj_priv->tiling_mode == I915_TILING_Y)
2345                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2346         val |= I915_FENCE_SIZE_BITS(obj->size);
2347         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2348         val |= I830_FENCE_REG_VALID;
2349
2350         if (regnum < 8)
2351                 fence_reg = FENCE_REG_830_0 + (regnum * 4);
2352         else
2353                 fence_reg = FENCE_REG_945_8 + ((regnum - 8) * 4);
2354         I915_WRITE(fence_reg, val);
2355 }
2356
2357 static void i830_write_fence_reg(struct drm_i915_fence_reg *reg)
2358 {
2359         struct drm_gem_object *obj = reg->obj;
2360         struct drm_device *dev = obj->dev;
2361         drm_i915_private_t *dev_priv = dev->dev_private;
2362         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2363         int regnum = obj_priv->fence_reg;
2364         uint32_t val;
2365         uint32_t pitch_val;
2366         uint32_t fence_size_bits;
2367
2368         if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) ||
2369             (obj_priv->gtt_offset & (obj->size - 1))) {
2370                 WARN(1, "%s: object 0x%08x not 512K or size aligned\n",
2371                      __func__, obj_priv->gtt_offset);
2372                 return;
2373         }
2374
2375         pitch_val = obj_priv->stride / 128;
2376         pitch_val = ffs(pitch_val) - 1;
2377         WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL);
2378
2379         val = obj_priv->gtt_offset;
2380         if (obj_priv->tiling_mode == I915_TILING_Y)
2381                 val |= 1 << I830_FENCE_TILING_Y_SHIFT;
2382         fence_size_bits = I830_FENCE_SIZE_BITS(obj->size);
2383         WARN_ON(fence_size_bits & ~0x00000f00);
2384         val |= fence_size_bits;
2385         val |= pitch_val << I830_FENCE_PITCH_SHIFT;
2386         val |= I830_FENCE_REG_VALID;
2387
2388         I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val);
2389 }
2390
2391 static int i915_find_fence_reg(struct drm_device *dev)
2392 {
2393         struct drm_i915_fence_reg *reg = NULL;
2394         struct drm_i915_gem_object *obj_priv = NULL;
2395         struct drm_i915_private *dev_priv = dev->dev_private;
2396         struct drm_gem_object *obj = NULL;
2397         int i, avail, ret;
2398
2399         /* First try to find a free reg */
2400         avail = 0;
2401         for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) {
2402                 reg = &dev_priv->fence_regs[i];
2403                 if (!reg->obj)
2404                         return i;
2405
2406                 obj_priv = to_intel_bo(reg->obj);
2407                 if (!obj_priv->pin_count)
2408                     avail++;
2409         }
2410
2411         if (avail == 0)
2412                 return -ENOSPC;
2413
2414         /* None available, try to steal one or wait for a user to finish */
2415         i = I915_FENCE_REG_NONE;
2416         list_for_each_entry(reg, &dev_priv->mm.fence_list,
2417                             lru_list) {
2418                 obj = reg->obj;
2419                 obj_priv = to_intel_bo(obj);
2420
2421                 if (obj_priv->pin_count)
2422                         continue;
2423
2424                 /* found one! */
2425                 i = obj_priv->fence_reg;
2426                 break;
2427         }
2428
2429         BUG_ON(i == I915_FENCE_REG_NONE);
2430
2431         /* We only have a reference on obj from the active list. put_fence_reg
2432          * might drop that one, causing a use-after-free in it. So hold a
2433          * private reference to obj like the other callers of put_fence_reg
2434          * (set_tiling ioctl) do. */
2435         drm_gem_object_reference(obj);
2436         ret = i915_gem_object_put_fence_reg(obj);
2437         drm_gem_object_unreference(obj);
2438         if (ret != 0)
2439                 return ret;
2440
2441         return i;
2442 }
2443
2444 /**
2445  * i915_gem_object_get_fence_reg - set up a fence reg for an object
2446  * @obj: object to map through a fence reg
2447  *
2448  * When mapping objects through the GTT, userspace wants to be able to write
2449  * to them without having to worry about swizzling if the object is tiled.
2450  *
2451  * This function walks the fence regs looking for a free one for @obj,
2452  * stealing one if it can't find any.
2453  *
2454  * It then sets up the reg based on the object's properties: address, pitch
2455  * and tiling format.
2456  */
2457 int
2458 i915_gem_object_get_fence_reg(struct drm_gem_object *obj)
2459 {
2460         struct drm_device *dev = obj->dev;
2461         struct drm_i915_private *dev_priv = dev->dev_private;
2462         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2463         struct drm_i915_fence_reg *reg = NULL;
2464         int ret;
2465
2466         /* Just update our place in the LRU if our fence is getting used. */
2467         if (obj_priv->fence_reg != I915_FENCE_REG_NONE) {
2468                 reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2469                 list_move_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2470                 return 0;
2471         }
2472
2473         switch (obj_priv->tiling_mode) {
2474         case I915_TILING_NONE:
2475                 WARN(1, "allocating a fence for non-tiled object?\n");
2476                 break;
2477         case I915_TILING_X:
2478                 if (!obj_priv->stride)
2479                         return -EINVAL;
2480                 WARN((obj_priv->stride & (512 - 1)),
2481                      "object 0x%08x is X tiled but has non-512B pitch\n",
2482                      obj_priv->gtt_offset);
2483                 break;
2484         case I915_TILING_Y:
2485                 if (!obj_priv->stride)
2486                         return -EINVAL;
2487                 WARN((obj_priv->stride & (128 - 1)),
2488                      "object 0x%08x is Y tiled but has non-128B pitch\n",
2489                      obj_priv->gtt_offset);
2490                 break;
2491         }
2492
2493         ret = i915_find_fence_reg(dev);
2494         if (ret < 0)
2495                 return ret;
2496
2497         obj_priv->fence_reg = ret;
2498         reg = &dev_priv->fence_regs[obj_priv->fence_reg];
2499         list_add_tail(&reg->lru_list, &dev_priv->mm.fence_list);
2500
2501         reg->obj = obj;
2502
2503         if (IS_GEN6(dev))
2504                 sandybridge_write_fence_reg(reg);
2505         else if (IS_I965G(dev))
2506                 i965_write_fence_reg(reg);
2507         else if (IS_I9XX(dev))
2508                 i915_write_fence_reg(reg);
2509         else
2510                 i830_write_fence_reg(reg);
2511
2512         trace_i915_gem_object_get_fence(obj, obj_priv->fence_reg,
2513                         obj_priv->tiling_mode);
2514
2515         return 0;
2516 }
2517
2518 /**
2519  * i915_gem_clear_fence_reg - clear out fence register info
2520  * @obj: object to clear
2521  *
2522  * Zeroes out the fence register itself and clears out the associated
2523  * data structures in dev_priv and obj_priv.
2524  */
2525 static void
2526 i915_gem_clear_fence_reg(struct drm_gem_object *obj)
2527 {
2528         struct drm_device *dev = obj->dev;
2529         drm_i915_private_t *dev_priv = dev->dev_private;
2530         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2531         struct drm_i915_fence_reg *reg =
2532                 &dev_priv->fence_regs[obj_priv->fence_reg];
2533
2534         if (IS_GEN6(dev)) {
2535                 I915_WRITE64(FENCE_REG_SANDYBRIDGE_0 +
2536                              (obj_priv->fence_reg * 8), 0);
2537         } else if (IS_I965G(dev)) {
2538                 I915_WRITE64(FENCE_REG_965_0 + (obj_priv->fence_reg * 8), 0);
2539         } else {
2540                 uint32_t fence_reg;
2541
2542                 if (obj_priv->fence_reg < 8)
2543                         fence_reg = FENCE_REG_830_0 + obj_priv->fence_reg * 4;
2544                 else
2545                         fence_reg = FENCE_REG_945_8 + (obj_priv->fence_reg -
2546                                                        8) * 4;
2547
2548                 I915_WRITE(fence_reg, 0);
2549         }
2550
2551         reg->obj = NULL;
2552         obj_priv->fence_reg = I915_FENCE_REG_NONE;
2553         list_del_init(&reg->lru_list);
2554 }
2555
2556 /**
2557  * i915_gem_object_put_fence_reg - waits on outstanding fenced access
2558  * to the buffer to finish, and then resets the fence register.
2559  * @obj: tiled object holding a fence register.
2560  *
2561  * Zeroes out the fence register itself and clears out the associated
2562  * data structures in dev_priv and obj_priv.
2563  */
2564 int
2565 i915_gem_object_put_fence_reg(struct drm_gem_object *obj)
2566 {
2567         struct drm_device *dev = obj->dev;
2568         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2569
2570         if (obj_priv->fence_reg == I915_FENCE_REG_NONE)
2571                 return 0;
2572
2573         /* If we've changed tiling, GTT-mappings of the object
2574          * need to re-fault to ensure that the correct fence register
2575          * setup is in place.
2576          */
2577         i915_gem_release_mmap(obj);
2578
2579         /* On the i915, GPU access to tiled buffers is via a fence,
2580          * therefore we must wait for any outstanding access to complete
2581          * before clearing the fence.
2582          */
2583         if (!IS_I965G(dev)) {
2584                 int ret;
2585
2586                 i915_gem_object_flush_gpu_write_domain(obj);
2587                 ret = i915_gem_object_wait_rendering(obj);
2588                 if (ret != 0)
2589                         return ret;
2590         }
2591
2592         i915_gem_object_flush_gtt_write_domain(obj);
2593         i915_gem_clear_fence_reg (obj);
2594
2595         return 0;
2596 }
2597
2598 /**
2599  * Finds free space in the GTT aperture and binds the object there.
2600  */
2601 static int
2602 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
2603 {
2604         struct drm_device *dev = obj->dev;
2605         drm_i915_private_t *dev_priv = dev->dev_private;
2606         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2607         struct drm_mm_node *free_space;
2608         gfp_t gfpmask =  __GFP_NORETRY | __GFP_NOWARN;
2609         int ret;
2610
2611         if (obj_priv->madv != I915_MADV_WILLNEED) {
2612                 DRM_ERROR("Attempting to bind a purgeable object\n");
2613                 return -EINVAL;
2614         }
2615
2616         if (alignment == 0)
2617                 alignment = i915_gem_get_gtt_alignment(obj);
2618         if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) {
2619                 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
2620                 return -EINVAL;
2621         }
2622
2623         /* If the object is bigger than the entire aperture, reject it early
2624          * before evicting everything in a vain attempt to find space.
2625          */
2626         if (obj->size > dev->gtt_total) {
2627                 DRM_ERROR("Attempting to bind an object larger than the aperture\n");
2628                 return -E2BIG;
2629         }
2630
2631  search_free:
2632         free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
2633                                         obj->size, alignment, 0);
2634         if (free_space != NULL) {
2635                 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
2636                                                        alignment);
2637                 if (obj_priv->gtt_space != NULL) {
2638                         obj_priv->gtt_space->private = obj;
2639                         obj_priv->gtt_offset = obj_priv->gtt_space->start;
2640                 }
2641         }
2642         if (obj_priv->gtt_space == NULL) {
2643                 /* If the gtt is empty and we're still having trouble
2644                  * fitting our object in, we're out of memory.
2645                  */
2646 #if WATCH_LRU
2647                 DRM_INFO("%s: GTT full, evicting something\n", __func__);
2648 #endif
2649                 ret = i915_gem_evict_something(dev, obj->size);
2650                 if (ret)
2651                         return ret;
2652
2653                 goto search_free;
2654         }
2655
2656 #if WATCH_BUF
2657         DRM_INFO("Binding object of size %zd at 0x%08x\n",
2658                  obj->size, obj_priv->gtt_offset);
2659 #endif
2660         ret = i915_gem_object_get_pages(obj, gfpmask);
2661         if (ret) {
2662                 drm_mm_put_block(obj_priv->gtt_space);
2663                 obj_priv->gtt_space = NULL;
2664
2665                 if (ret == -ENOMEM) {
2666                         /* first try to clear up some space from the GTT */
2667                         ret = i915_gem_evict_something(dev, obj->size);
2668                         if (ret) {
2669                                 /* now try to shrink everyone else */
2670                                 if (gfpmask) {
2671                                         gfpmask = 0;
2672                                         goto search_free;
2673                                 }
2674
2675                                 return ret;
2676                         }
2677
2678                         goto search_free;
2679                 }
2680
2681                 return ret;
2682         }
2683
2684         /* Create an AGP memory structure pointing at our pages, and bind it
2685          * into the GTT.
2686          */
2687         obj_priv->agp_mem = drm_agp_bind_pages(dev,
2688                                                obj_priv->pages,
2689                                                obj->size >> PAGE_SHIFT,
2690                                                obj_priv->gtt_offset,
2691                                                obj_priv->agp_type);
2692         if (obj_priv->agp_mem == NULL) {
2693                 i915_gem_object_put_pages(obj);
2694                 drm_mm_put_block(obj_priv->gtt_space);
2695                 obj_priv->gtt_space = NULL;
2696
2697                 ret = i915_gem_evict_something(dev, obj->size);
2698                 if (ret)
2699                         return ret;
2700
2701                 goto search_free;
2702         }
2703         atomic_inc(&dev->gtt_count);
2704         atomic_add(obj->size, &dev->gtt_memory);
2705
2706         /* Assert that the object is not currently in any GPU domain. As it
2707          * wasn't in the GTT, there shouldn't be any way it could have been in
2708          * a GPU cache
2709          */
2710         BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS);
2711         BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS);
2712
2713         trace_i915_gem_object_bind(obj, obj_priv->gtt_offset);
2714
2715         return 0;
2716 }
2717
2718 void
2719 i915_gem_clflush_object(struct drm_gem_object *obj)
2720 {
2721         struct drm_i915_gem_object      *obj_priv = to_intel_bo(obj);
2722
2723         /* If we don't have a page list set up, then we're not pinned
2724          * to GPU, and we can ignore the cache flush because it'll happen
2725          * again at bind time.
2726          */
2727         if (obj_priv->pages == NULL)
2728                 return;
2729
2730         trace_i915_gem_object_clflush(obj);
2731
2732         drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE);
2733 }
2734
2735 /** Flushes any GPU write domain for the object if it's dirty. */
2736 static void
2737 i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj)
2738 {
2739         struct drm_device *dev = obj->dev;
2740         uint32_t old_write_domain;
2741         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2742
2743         if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
2744                 return;
2745
2746         /* Queue the GPU write cache flushing we need. */
2747         old_write_domain = obj->write_domain;
2748         i915_gem_flush(dev, 0, obj->write_domain);
2749         (void) i915_add_request(dev, NULL, obj->write_domain, obj_priv->ring);
2750         BUG_ON(obj->write_domain);
2751
2752         trace_i915_gem_object_change_domain(obj,
2753                                             obj->read_domains,
2754                                             old_write_domain);
2755 }
2756
2757 /** Flushes the GTT write domain for the object if it's dirty. */
2758 static void
2759 i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj)
2760 {
2761         uint32_t old_write_domain;
2762
2763         if (obj->write_domain != I915_GEM_DOMAIN_GTT)
2764                 return;
2765
2766         /* No actual flushing is required for the GTT write domain.   Writes
2767          * to it immediately go to main memory as far as we know, so there's
2768          * no chipset flush.  It also doesn't land in render cache.
2769          */
2770         old_write_domain = obj->write_domain;
2771         obj->write_domain = 0;
2772
2773         trace_i915_gem_object_change_domain(obj,
2774                                             obj->read_domains,
2775                                             old_write_domain);
2776 }
2777
2778 /** Flushes the CPU write domain for the object if it's dirty. */
2779 static void
2780 i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj)
2781 {
2782         struct drm_device *dev = obj->dev;
2783         uint32_t old_write_domain;
2784
2785         if (obj->write_domain != I915_GEM_DOMAIN_CPU)
2786                 return;
2787
2788         i915_gem_clflush_object(obj);
2789         drm_agp_chipset_flush(dev);
2790         old_write_domain = obj->write_domain;
2791         obj->write_domain = 0;
2792
2793         trace_i915_gem_object_change_domain(obj,
2794                                             obj->read_domains,
2795                                             old_write_domain);
2796 }
2797
2798 void
2799 i915_gem_object_flush_write_domain(struct drm_gem_object *obj)
2800 {
2801         switch (obj->write_domain) {
2802         case I915_GEM_DOMAIN_GTT:
2803                 i915_gem_object_flush_gtt_write_domain(obj);
2804                 break;
2805         case I915_GEM_DOMAIN_CPU:
2806                 i915_gem_object_flush_cpu_write_domain(obj);
2807                 break;
2808         default:
2809                 i915_gem_object_flush_gpu_write_domain(obj);
2810                 break;
2811         }
2812 }
2813
2814 /**
2815  * Moves a single object to the GTT read, and possibly write domain.
2816  *
2817  * This function returns when the move is complete, including waiting on
2818  * flushes to occur.
2819  */
2820 int
2821 i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write)
2822 {
2823         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2824         uint32_t old_write_domain, old_read_domains;
2825         int ret;
2826
2827         /* Not valid to be called on unbound objects. */
2828         if (obj_priv->gtt_space == NULL)
2829                 return -EINVAL;
2830
2831         i915_gem_object_flush_gpu_write_domain(obj);
2832         /* Wait on any GPU rendering and flushing to occur. */
2833         ret = i915_gem_object_wait_rendering(obj);
2834         if (ret != 0)
2835                 return ret;
2836
2837         old_write_domain = obj->write_domain;
2838         old_read_domains = obj->read_domains;
2839
2840         /* If we're writing through the GTT domain, then CPU and GPU caches
2841          * will need to be invalidated at next use.
2842          */
2843         if (write)
2844                 obj->read_domains &= I915_GEM_DOMAIN_GTT;
2845
2846         i915_gem_object_flush_cpu_write_domain(obj);
2847
2848         /* It should now be out of any other write domains, and we can update
2849          * the domain values for our changes.
2850          */
2851         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2852         obj->read_domains |= I915_GEM_DOMAIN_GTT;
2853         if (write) {
2854                 obj->write_domain = I915_GEM_DOMAIN_GTT;
2855                 obj_priv->dirty = 1;
2856         }
2857
2858         trace_i915_gem_object_change_domain(obj,
2859                                             old_read_domains,
2860                                             old_write_domain);
2861
2862         return 0;
2863 }
2864
2865 /*
2866  * Prepare buffer for display plane. Use uninterruptible for possible flush
2867  * wait, as in modesetting process we're not supposed to be interrupted.
2868  */
2869 int
2870 i915_gem_object_set_to_display_plane(struct drm_gem_object *obj)
2871 {
2872         struct drm_device *dev = obj->dev;
2873         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
2874         uint32_t old_write_domain, old_read_domains;
2875         int ret;
2876
2877         /* Not valid to be called on unbound objects. */
2878         if (obj_priv->gtt_space == NULL)
2879                 return -EINVAL;
2880
2881         i915_gem_object_flush_gpu_write_domain(obj);
2882
2883         /* Wait on any GPU rendering and flushing to occur. */
2884         if (obj_priv->active) {
2885 #if WATCH_BUF
2886                 DRM_INFO("%s: object %p wait for seqno %08x\n",
2887                           __func__, obj, obj_priv->last_rendering_seqno);
2888 #endif
2889                 ret = i915_do_wait_request(dev,
2890                                 obj_priv->last_rendering_seqno,
2891                                 0,
2892                                 obj_priv->ring);
2893                 if (ret != 0)
2894                         return ret;
2895         }
2896
2897         i915_gem_object_flush_cpu_write_domain(obj);
2898
2899         old_write_domain = obj->write_domain;
2900         old_read_domains = obj->read_domains;
2901
2902         /* It should now be out of any other write domains, and we can update
2903          * the domain values for our changes.
2904          */
2905         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_GTT) != 0);
2906         obj->read_domains = I915_GEM_DOMAIN_GTT;
2907         obj->write_domain = I915_GEM_DOMAIN_GTT;
2908         obj_priv->dirty = 1;
2909
2910         trace_i915_gem_object_change_domain(obj,
2911                                             old_read_domains,
2912                                             old_write_domain);
2913
2914         return 0;
2915 }
2916
2917 /**
2918  * Moves a single object to the CPU read, and possibly write domain.
2919  *
2920  * This function returns when the move is complete, including waiting on
2921  * flushes to occur.
2922  */
2923 static int
2924 i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write)
2925 {
2926         uint32_t old_write_domain, old_read_domains;
2927         int ret;
2928
2929         i915_gem_object_flush_gpu_write_domain(obj);
2930         /* Wait on any GPU rendering and flushing to occur. */
2931         ret = i915_gem_object_wait_rendering(obj);
2932         if (ret != 0)
2933                 return ret;
2934
2935         i915_gem_object_flush_gtt_write_domain(obj);
2936
2937         /* If we have a partially-valid cache of the object in the CPU,
2938          * finish invalidating it and free the per-page flags.
2939          */
2940         i915_gem_object_set_to_full_cpu_read_domain(obj);
2941
2942         old_write_domain = obj->write_domain;
2943         old_read_domains = obj->read_domains;
2944
2945         /* Flush the CPU cache if it's still invalid. */
2946         if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) {
2947                 i915_gem_clflush_object(obj);
2948
2949                 obj->read_domains |= I915_GEM_DOMAIN_CPU;
2950         }
2951
2952         /* It should now be out of any other write domains, and we can update
2953          * the domain values for our changes.
2954          */
2955         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
2956
2957         /* If we're writing through the CPU, then the GPU read domains will
2958          * need to be invalidated at next use.
2959          */
2960         if (write) {
2961                 obj->read_domains &= I915_GEM_DOMAIN_CPU;
2962                 obj->write_domain = I915_GEM_DOMAIN_CPU;
2963         }
2964
2965         trace_i915_gem_object_change_domain(obj,
2966                                             old_read_domains,
2967                                             old_write_domain);
2968
2969         return 0;
2970 }
2971
2972 /*
2973  * Set the next domain for the specified object. This
2974  * may not actually perform the necessary flushing/invaliding though,
2975  * as that may want to be batched with other set_domain operations
2976  *
2977  * This is (we hope) the only really tricky part of gem. The goal
2978  * is fairly simple -- track which caches hold bits of the object
2979  * and make sure they remain coherent. A few concrete examples may
2980  * help to explain how it works. For shorthand, we use the notation
2981  * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
2982  * a pair of read and write domain masks.
2983  *
2984  * Case 1: the batch buffer
2985  *
2986  *      1. Allocated
2987  *      2. Written by CPU
2988  *      3. Mapped to GTT
2989  *      4. Read by GPU
2990  *      5. Unmapped from GTT
2991  *      6. Freed
2992  *
2993  *      Let's take these a step at a time
2994  *
2995  *      1. Allocated
2996  *              Pages allocated from the kernel may still have
2997  *              cache contents, so we set them to (CPU, CPU) always.
2998  *      2. Written by CPU (using pwrite)
2999  *              The pwrite function calls set_domain (CPU, CPU) and
3000  *              this function does nothing (as nothing changes)
3001  *      3. Mapped by GTT
3002  *              This function asserts that the object is not
3003  *              currently in any GPU-based read or write domains
3004  *      4. Read by GPU
3005  *              i915_gem_execbuffer calls set_domain (COMMAND, 0).
3006  *              As write_domain is zero, this function adds in the
3007  *              current read domains (CPU+COMMAND, 0).
3008  *              flush_domains is set to CPU.
3009  *              invalidate_domains is set to COMMAND
3010  *              clflush is run to get data out of the CPU caches
3011  *              then i915_dev_set_domain calls i915_gem_flush to
3012  *              emit an MI_FLUSH and drm_agp_chipset_flush
3013  *      5. Unmapped from GTT
3014  *              i915_gem_object_unbind calls set_domain (CPU, CPU)
3015  *              flush_domains and invalidate_domains end up both zero
3016  *              so no flushing/invalidating happens
3017  *      6. Freed
3018  *              yay, done
3019  *
3020  * Case 2: The shared render buffer
3021  *
3022  *      1. Allocated
3023  *      2. Mapped to GTT
3024  *      3. Read/written by GPU
3025  *      4. set_domain to (CPU,CPU)
3026  *      5. Read/written by CPU
3027  *      6. Read/written by GPU
3028  *
3029  *      1. Allocated
3030  *              Same as last example, (CPU, CPU)
3031  *      2. Mapped to GTT
3032  *              Nothing changes (assertions find that it is not in the GPU)
3033  *      3. Read/written by GPU
3034  *              execbuffer calls set_domain (RENDER, RENDER)
3035  *              flush_domains gets CPU
3036  *              invalidate_domains gets GPU
3037  *              clflush (obj)
3038  *              MI_FLUSH and drm_agp_chipset_flush
3039  *      4. set_domain (CPU, CPU)
3040  *              flush_domains gets GPU
3041  *              invalidate_domains gets CPU
3042  *              wait_rendering (obj) to make sure all drawing is complete.
3043  *              This will include an MI_FLUSH to get the data from GPU
3044  *              to memory
3045  *              clflush (obj) to invalidate the CPU cache
3046  *              Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
3047  *      5. Read/written by CPU
3048  *              cache lines are loaded and dirtied
3049  *      6. Read written by GPU
3050  *              Same as last GPU access
3051  *
3052  * Case 3: The constant buffer
3053  *
3054  *      1. Allocated
3055  *      2. Written by CPU
3056  *      3. Read by GPU
3057  *      4. Updated (written) by CPU again
3058  *      5. Read by GPU
3059  *
3060  *      1. Allocated
3061  *              (CPU, CPU)
3062  *      2. Written by CPU
3063  *              (CPU, CPU)
3064  *      3. Read by GPU
3065  *              (CPU+RENDER, 0)
3066  *              flush_domains = CPU
3067  *              invalidate_domains = RENDER
3068  *              clflush (obj)
3069  *              MI_FLUSH
3070  *              drm_agp_chipset_flush
3071  *      4. Updated (written) by CPU again
3072  *              (CPU, CPU)
3073  *              flush_domains = 0 (no previous write domain)
3074  *              invalidate_domains = 0 (no new read domains)
3075  *      5. Read by GPU
3076  *              (CPU+RENDER, 0)
3077  *              flush_domains = CPU
3078  *              invalidate_domains = RENDER
3079  *              clflush (obj)
3080  *              MI_FLUSH
3081  *              drm_agp_chipset_flush
3082  */
3083 static void
3084 i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj)
3085 {
3086         struct drm_device               *dev = obj->dev;
3087         struct drm_i915_gem_object      *obj_priv = to_intel_bo(obj);
3088         uint32_t                        invalidate_domains = 0;
3089         uint32_t                        flush_domains = 0;
3090         uint32_t                        old_read_domains;
3091
3092         BUG_ON(obj->pending_read_domains & I915_GEM_DOMAIN_CPU);
3093         BUG_ON(obj->pending_write_domain == I915_GEM_DOMAIN_CPU);
3094
3095         intel_mark_busy(dev, obj);
3096
3097 #if WATCH_BUF
3098         DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
3099                  __func__, obj,
3100                  obj->read_domains, obj->pending_read_domains,
3101                  obj->write_domain, obj->pending_write_domain);
3102 #endif
3103         /*
3104          * If the object isn't moving to a new write domain,
3105          * let the object stay in multiple read domains
3106          */
3107         if (obj->pending_write_domain == 0)
3108                 obj->pending_read_domains |= obj->read_domains;
3109         else
3110                 obj_priv->dirty = 1;
3111
3112         /*
3113          * Flush the current write domain if
3114          * the new read domains don't match. Invalidate
3115          * any read domains which differ from the old
3116          * write domain
3117          */
3118         if (obj->write_domain &&
3119             obj->write_domain != obj->pending_read_domains) {
3120                 flush_domains |= obj->write_domain;
3121                 invalidate_domains |=
3122                         obj->pending_read_domains & ~obj->write_domain;
3123         }
3124         /*
3125          * Invalidate any read caches which may have
3126          * stale data. That is, any new read domains.
3127          */
3128         invalidate_domains |= obj->pending_read_domains & ~obj->read_domains;
3129         if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
3130 #if WATCH_BUF
3131                 DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
3132                          __func__, flush_domains, invalidate_domains);
3133 #endif
3134                 i915_gem_clflush_object(obj);
3135         }
3136
3137         old_read_domains = obj->read_domains;
3138
3139         /* The actual obj->write_domain will be updated with
3140          * pending_write_domain after we emit the accumulated flush for all
3141          * of our domain changes in execbuffers (which clears objects'
3142          * write_domains).  So if we have a current write domain that we
3143          * aren't changing, set pending_write_domain to that.
3144          */
3145         if (flush_domains == 0 && obj->pending_write_domain == 0)
3146                 obj->pending_write_domain = obj->write_domain;
3147         obj->read_domains = obj->pending_read_domains;
3148
3149         dev->invalidate_domains |= invalidate_domains;
3150         dev->flush_domains |= flush_domains;
3151 #if WATCH_BUF
3152         DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
3153                  __func__,
3154                  obj->read_domains, obj->write_domain,
3155                  dev->invalidate_domains, dev->flush_domains);
3156 #endif
3157
3158         trace_i915_gem_object_change_domain(obj,
3159                                             old_read_domains,
3160                                             obj->write_domain);
3161 }
3162
3163 /**
3164  * Moves the object from a partially CPU read to a full one.
3165  *
3166  * Note that this only resolves i915_gem_object_set_cpu_read_domain_range(),
3167  * and doesn't handle transitioning from !(read_domains & I915_GEM_DOMAIN_CPU).
3168  */
3169 static void
3170 i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj)
3171 {
3172         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3173
3174         if (!obj_priv->page_cpu_valid)
3175                 return;
3176
3177         /* If we're partially in the CPU read domain, finish moving it in.
3178          */
3179         if (obj->read_domains & I915_GEM_DOMAIN_CPU) {
3180                 int i;
3181
3182                 for (i = 0; i <= (obj->size - 1) / PAGE_SIZE; i++) {
3183                         if (obj_priv->page_cpu_valid[i])
3184                                 continue;
3185                         drm_clflush_pages(obj_priv->pages + i, 1);
3186                 }
3187         }
3188
3189         /* Free the page_cpu_valid mappings which are now stale, whether
3190          * or not we've got I915_GEM_DOMAIN_CPU.
3191          */
3192         kfree(obj_priv->page_cpu_valid);
3193         obj_priv->page_cpu_valid = NULL;
3194 }
3195
3196 /**
3197  * Set the CPU read domain on a range of the object.
3198  *
3199  * The object ends up with I915_GEM_DOMAIN_CPU in its read flags although it's
3200  * not entirely valid.  The page_cpu_valid member of the object flags which
3201  * pages have been flushed, and will be respected by
3202  * i915_gem_object_set_to_cpu_domain() if it's called on to get a valid mapping
3203  * of the whole object.
3204  *
3205  * This function returns when the move is complete, including waiting on
3206  * flushes to occur.
3207  */
3208 static int
3209 i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj,
3210                                           uint64_t offset, uint64_t size)
3211 {
3212         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3213         uint32_t old_read_domains;
3214         int i, ret;
3215
3216         if (offset == 0 && size == obj->size)
3217                 return i915_gem_object_set_to_cpu_domain(obj, 0);
3218
3219         i915_gem_object_flush_gpu_write_domain(obj);
3220         /* Wait on any GPU rendering and flushing to occur. */
3221         ret = i915_gem_object_wait_rendering(obj);
3222         if (ret != 0)
3223                 return ret;
3224         i915_gem_object_flush_gtt_write_domain(obj);
3225
3226         /* If we're already fully in the CPU read domain, we're done. */
3227         if (obj_priv->page_cpu_valid == NULL &&
3228             (obj->read_domains & I915_GEM_DOMAIN_CPU) != 0)
3229                 return 0;
3230
3231         /* Otherwise, create/clear the per-page CPU read domain flag if we're
3232          * newly adding I915_GEM_DOMAIN_CPU
3233          */
3234         if (obj_priv->page_cpu_valid == NULL) {
3235                 obj_priv->page_cpu_valid = kzalloc(obj->size / PAGE_SIZE,
3236                                                    GFP_KERNEL);
3237                 if (obj_priv->page_cpu_valid == NULL)
3238                         return -ENOMEM;
3239         } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0)
3240                 memset(obj_priv->page_cpu_valid, 0, obj->size / PAGE_SIZE);
3241
3242         /* Flush the cache on any pages that are still invalid from the CPU's
3243          * perspective.
3244          */
3245         for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE;
3246              i++) {
3247                 if (obj_priv->page_cpu_valid[i])
3248                         continue;
3249
3250                 drm_clflush_pages(obj_priv->pages + i, 1);
3251
3252                 obj_priv->page_cpu_valid[i] = 1;
3253         }
3254
3255         /* It should now be out of any other write domains, and we can update
3256          * the domain values for our changes.
3257          */
3258         BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0);
3259
3260         old_read_domains = obj->read_domains;
3261         obj->read_domains |= I915_GEM_DOMAIN_CPU;
3262
3263         trace_i915_gem_object_change_domain(obj,
3264                                             old_read_domains,
3265                                             obj->write_domain);
3266
3267         return 0;
3268 }
3269
3270 /**
3271  * Pin an object to the GTT and evaluate the relocations landing in it.
3272  */
3273 static int
3274 i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
3275                                  struct drm_file *file_priv,
3276                                  struct drm_i915_gem_exec_object2 *entry,
3277                                  struct drm_i915_gem_relocation_entry *relocs)
3278 {
3279         struct drm_device *dev = obj->dev;
3280         drm_i915_private_t *dev_priv = dev->dev_private;
3281         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3282         int i, ret;
3283         void __iomem *reloc_page;
3284         bool need_fence;
3285
3286         need_fence = entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
3287                      obj_priv->tiling_mode != I915_TILING_NONE;
3288
3289         /* Check fence reg constraints and rebind if necessary */
3290         if (need_fence &&
3291             !i915_gem_object_fence_offset_ok(obj,
3292                                              obj_priv->tiling_mode)) {
3293                 ret = i915_gem_object_unbind(obj);
3294                 if (ret)
3295                         return ret;
3296         }
3297
3298         /* Choose the GTT offset for our buffer and put it there. */
3299         ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
3300         if (ret)
3301                 return ret;
3302
3303         /*
3304          * Pre-965 chips need a fence register set up in order to
3305          * properly handle blits to/from tiled surfaces.
3306          */
3307         if (need_fence) {
3308                 ret = i915_gem_object_get_fence_reg(obj);
3309                 if (ret != 0) {
3310                         i915_gem_object_unpin(obj);
3311                         return ret;
3312                 }
3313         }
3314
3315         entry->offset = obj_priv->gtt_offset;
3316
3317         /* Apply the relocations, using the GTT aperture to avoid cache
3318          * flushing requirements.
3319          */
3320         for (i = 0; i < entry->relocation_count; i++) {
3321                 struct drm_i915_gem_relocation_entry *reloc= &relocs[i];
3322                 struct drm_gem_object *target_obj;
3323                 struct drm_i915_gem_object *target_obj_priv;
3324                 uint32_t reloc_val, reloc_offset;
3325                 uint32_t __iomem *reloc_entry;
3326
3327                 target_obj = drm_gem_object_lookup(obj->dev, file_priv,
3328                                                    reloc->target_handle);
3329                 if (target_obj == NULL) {
3330                         i915_gem_object_unpin(obj);
3331                         return -EBADF;
3332                 }
3333                 target_obj_priv = to_intel_bo(target_obj);
3334
3335 #if WATCH_RELOC
3336                 DRM_INFO("%s: obj %p offset %08x target %d "
3337                          "read %08x write %08x gtt %08x "
3338                          "presumed %08x delta %08x\n",
3339                          __func__,
3340                          obj,
3341                          (int) reloc->offset,
3342                          (int) reloc->target_handle,
3343                          (int) reloc->read_domains,
3344                          (int) reloc->write_domain,
3345                          (int) target_obj_priv->gtt_offset,
3346                          (int) reloc->presumed_offset,
3347                          reloc->delta);
3348 #endif
3349
3350                 /* The target buffer should have appeared before us in the
3351                  * exec_object list, so it should have a GTT space bound by now.
3352                  */
3353                 if (target_obj_priv->gtt_space == NULL) {
3354                         DRM_ERROR("No GTT space found for object %d\n",
3355                                   reloc->target_handle);
3356                         drm_gem_object_unreference(target_obj);
3357                         i915_gem_object_unpin(obj);
3358                         return -EINVAL;
3359                 }
3360
3361                 /* Validate that the target is in a valid r/w GPU domain */
3362                 if (reloc->write_domain & (reloc->write_domain - 1)) {
3363                         DRM_ERROR("reloc with multiple write domains: "
3364                                   "obj %p target %d offset %d "
3365                                   "read %08x write %08x",
3366                                   obj, reloc->target_handle,
3367                                   (int) reloc->offset,
3368                                   reloc->read_domains,
3369                                   reloc->write_domain);
3370                         return -EINVAL;
3371                 }
3372                 if (reloc->write_domain & I915_GEM_DOMAIN_CPU ||
3373                     reloc->read_domains & I915_GEM_DOMAIN_CPU) {
3374                         DRM_ERROR("reloc with read/write CPU domains: "
3375                                   "obj %p target %d offset %d "
3376                                   "read %08x write %08x",
3377                                   obj, reloc->target_handle,
3378                                   (int) reloc->offset,
3379                                   reloc->read_domains,
3380                                   reloc->write_domain);
3381                         drm_gem_object_unreference(target_obj);
3382                         i915_gem_object_unpin(obj);
3383                         return -EINVAL;
3384                 }
3385                 if (reloc->write_domain && target_obj->pending_write_domain &&
3386                     reloc->write_domain != target_obj->pending_write_domain) {
3387                         DRM_ERROR("Write domain conflict: "
3388                                   "obj %p target %d offset %d "
3389                                   "new %08x old %08x\n",
3390                                   obj, reloc->target_handle,
3391                                   (int) reloc->offset,
3392                                   reloc->write_domain,
3393                                   target_obj->pending_write_domain);
3394                         drm_gem_object_unreference(target_obj);
3395                         i915_gem_object_unpin(obj);
3396                         return -EINVAL;
3397                 }
3398
3399                 target_obj->pending_read_domains |= reloc->read_domains;
3400                 target_obj->pending_write_domain |= reloc->write_domain;
3401
3402                 /* If the relocation already has the right value in it, no
3403                  * more work needs to be done.
3404                  */
3405                 if (target_obj_priv->gtt_offset == reloc->presumed_offset) {
3406                         drm_gem_object_unreference(target_obj);
3407                         continue;
3408                 }
3409
3410                 /* Check that the relocation address is valid... */
3411                 if (reloc->offset > obj->size - 4) {
3412                         DRM_ERROR("Relocation beyond object bounds: "
3413                                   "obj %p target %d offset %d size %d.\n",
3414                                   obj, reloc->target_handle,
3415                                   (int) reloc->offset, (int) obj->size);
3416                         drm_gem_object_unreference(target_obj);
3417                         i915_gem_object_unpin(obj);
3418                         return -EINVAL;
3419                 }
3420                 if (reloc->offset & 3) {
3421                         DRM_ERROR("Relocation not 4-byte aligned: "
3422                                   "obj %p target %d offset %d.\n",
3423                                   obj, reloc->target_handle,
3424                                   (int) reloc->offset);
3425                         drm_gem_object_unreference(target_obj);
3426                         i915_gem_object_unpin(obj);
3427                         return -EINVAL;
3428                 }
3429
3430                 /* and points to somewhere within the target object. */
3431                 if (reloc->delta >= target_obj->size) {
3432                         DRM_ERROR("Relocation beyond target object bounds: "
3433                                   "obj %p target %d delta %d size %d.\n",
3434                                   obj, reloc->target_handle,
3435                                   (int) reloc->delta, (int) target_obj->size);
3436                         drm_gem_object_unreference(target_obj);
3437                         i915_gem_object_unpin(obj);
3438                         return -EINVAL;
3439                 }
3440
3441                 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
3442                 if (ret != 0) {
3443                         drm_gem_object_unreference(target_obj);
3444                         i915_gem_object_unpin(obj);
3445                         return -EINVAL;
3446                 }
3447
3448                 /* Map the page containing the relocation we're going to
3449                  * perform.
3450                  */
3451                 reloc_offset = obj_priv->gtt_offset + reloc->offset;
3452                 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
3453                                                       (reloc_offset &
3454                                                        ~(PAGE_SIZE - 1)));
3455                 reloc_entry = (uint32_t __iomem *)(reloc_page +
3456                                                    (reloc_offset & (PAGE_SIZE - 1)));
3457                 reloc_val = target_obj_priv->gtt_offset + reloc->delta;
3458
3459 #if WATCH_BUF
3460                 DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
3461                           obj, (unsigned int) reloc->offset,
3462                           readl(reloc_entry), reloc_val);
3463 #endif
3464                 writel(reloc_val, reloc_entry);
3465                 io_mapping_unmap_atomic(reloc_page);
3466
3467                 /* The updated presumed offset for this entry will be
3468                  * copied back out to the user.
3469                  */
3470                 reloc->presumed_offset = target_obj_priv->gtt_offset;
3471
3472                 drm_gem_object_unreference(target_obj);
3473         }
3474
3475 #if WATCH_BUF
3476         if (0)
3477                 i915_gem_dump_object(obj, 128, __func__, ~0);
3478 #endif
3479         return 0;
3480 }
3481
3482 /* Throttle our rendering by waiting until the ring has completed our requests
3483  * emitted over 20 msec ago.
3484  *
3485  * Note that if we were to use the current jiffies each time around the loop,
3486  * we wouldn't escape the function with any frames outstanding if the time to
3487  * render a frame was over 20ms.
3488  *
3489  * This should get us reasonable parallelism between CPU and GPU but also
3490  * relatively low latency when blocking on a particular request to finish.
3491  */
3492 static int
3493 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
3494 {
3495         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
3496         int ret = 0;
3497         unsigned long recent_enough = jiffies - msecs_to_jiffies(20);
3498
3499         mutex_lock(&dev->struct_mutex);
3500         while (!list_empty(&i915_file_priv->mm.request_list)) {
3501                 struct drm_i915_gem_request *request;
3502
3503                 request = list_first_entry(&i915_file_priv->mm.request_list,
3504                                            struct drm_i915_gem_request,
3505                                            client_list);
3506
3507                 if (time_after_eq(request->emitted_jiffies, recent_enough))
3508                         break;
3509
3510                 ret = i915_wait_request(dev, request->seqno, request->ring);
3511                 if (ret != 0)
3512                         break;
3513         }
3514         mutex_unlock(&dev->struct_mutex);
3515
3516         return ret;
3517 }
3518
3519 static int
3520 i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object2 *exec_list,
3521                               uint32_t buffer_count,
3522                               struct drm_i915_gem_relocation_entry **relocs)
3523 {
3524         uint32_t reloc_count = 0, reloc_index = 0, i;
3525         int ret;
3526
3527         *relocs = NULL;
3528         for (i = 0; i < buffer_count; i++) {
3529                 if (reloc_count + exec_list[i].relocation_count < reloc_count)
3530                         return -EINVAL;
3531                 reloc_count += exec_list[i].relocation_count;
3532         }
3533
3534         *relocs = drm_calloc_large(reloc_count, sizeof(**relocs));
3535         if (*relocs == NULL) {
3536                 DRM_ERROR("failed to alloc relocs, count %d\n", reloc_count);
3537                 return -ENOMEM;
3538         }
3539
3540         for (i = 0; i < buffer_count; i++) {
3541                 struct drm_i915_gem_relocation_entry __user *user_relocs;
3542
3543                 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3544
3545                 ret = copy_from_user(&(*relocs)[reloc_index],
3546                                      user_relocs,
3547                                      exec_list[i].relocation_count *
3548                                      sizeof(**relocs));
3549                 if (ret != 0) {
3550                         drm_free_large(*relocs);
3551                         *relocs = NULL;
3552                         return -EFAULT;
3553                 }
3554
3555                 reloc_index += exec_list[i].relocation_count;
3556         }
3557
3558         return 0;
3559 }
3560
3561 static int
3562 i915_gem_put_relocs_to_user(struct drm_i915_gem_exec_object2 *exec_list,
3563                             uint32_t buffer_count,
3564                             struct drm_i915_gem_relocation_entry *relocs)
3565 {
3566         uint32_t reloc_count = 0, i;
3567         int ret = 0;
3568
3569         if (relocs == NULL)
3570             return 0;
3571
3572         for (i = 0; i < buffer_count; i++) {
3573                 struct drm_i915_gem_relocation_entry __user *user_relocs;
3574                 int unwritten;
3575
3576                 user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr;
3577
3578                 unwritten = copy_to_user(user_relocs,
3579                                          &relocs[reloc_count],
3580                                          exec_list[i].relocation_count *
3581                                          sizeof(*relocs));
3582
3583                 if (unwritten) {
3584                         ret = -EFAULT;
3585                         goto err;
3586                 }
3587
3588                 reloc_count += exec_list[i].relocation_count;
3589         }
3590
3591 err:
3592         drm_free_large(relocs);
3593
3594         return ret;
3595 }
3596
3597 static int
3598 i915_gem_check_execbuffer (struct drm_i915_gem_execbuffer2 *exec,
3599                            uint64_t exec_offset)
3600 {
3601         uint32_t exec_start, exec_len;
3602
3603         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
3604         exec_len = (uint32_t) exec->batch_len;
3605
3606         if ((exec_start | exec_len) & 0x7)
3607                 return -EINVAL;
3608
3609         if (!exec_start)
3610                 return -EINVAL;
3611
3612         return 0;
3613 }
3614
3615 static int
3616 i915_gem_wait_for_pending_flip(struct drm_device *dev,
3617                                struct drm_gem_object **object_list,
3618                                int count)
3619 {
3620         drm_i915_private_t *dev_priv = dev->dev_private;
3621         struct drm_i915_gem_object *obj_priv;
3622         DEFINE_WAIT(wait);
3623         int i, ret = 0;
3624
3625         for (;;) {
3626                 prepare_to_wait(&dev_priv->pending_flip_queue,
3627                                 &wait, TASK_INTERRUPTIBLE);
3628                 for (i = 0; i < count; i++) {
3629                         obj_priv = to_intel_bo(object_list[i]);
3630                         if (atomic_read(&obj_priv->pending_flip) > 0)
3631                                 break;
3632                 }
3633                 if (i == count)
3634                         break;
3635
3636                 if (!signal_pending(current)) {
3637                         mutex_unlock(&dev->struct_mutex);
3638                         schedule();
3639                         mutex_lock(&dev->struct_mutex);
3640                         continue;
3641                 }
3642                 ret = -ERESTARTSYS;
3643                 break;
3644         }
3645         finish_wait(&dev_priv->pending_flip_queue, &wait);
3646
3647         return ret;
3648 }
3649
3650 int
3651 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
3652                        struct drm_file *file_priv,
3653                        struct drm_i915_gem_execbuffer2 *args,
3654                        struct drm_i915_gem_exec_object2 *exec_list)
3655 {
3656         drm_i915_private_t *dev_priv = dev->dev_private;
3657         struct drm_gem_object **object_list = NULL;
3658         struct drm_gem_object *batch_obj;
3659         struct drm_i915_gem_object *obj_priv;
3660         struct drm_clip_rect *cliprects = NULL;
3661         struct drm_i915_gem_relocation_entry *relocs = NULL;
3662         int ret = 0, ret2, i, pinned = 0;
3663         uint64_t exec_offset;
3664         uint32_t seqno, flush_domains, reloc_index;
3665         int pin_tries, flips;
3666
3667         struct intel_ring_buffer *ring = NULL;
3668
3669 #if WATCH_EXEC
3670         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
3671                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
3672 #endif
3673         if (args->flags & I915_EXEC_BSD) {
3674                 if (!HAS_BSD(dev)) {
3675                         DRM_ERROR("execbuf with wrong flag\n");
3676                         return -EINVAL;
3677                 }
3678                 ring = &dev_priv->bsd_ring;
3679         } else {
3680                 ring = &dev_priv->render_ring;
3681         }
3682
3683
3684         if (args->buffer_count < 1) {
3685                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
3686                 return -EINVAL;
3687         }
3688         object_list = drm_malloc_ab(sizeof(*object_list), args->buffer_count);
3689         if (object_list == NULL) {
3690                 DRM_ERROR("Failed to allocate object list for %d buffers\n",
3691                           args->buffer_count);
3692                 ret = -ENOMEM;
3693                 goto pre_mutex_err;
3694         }
3695
3696         if (args->num_cliprects != 0) {
3697                 cliprects = kcalloc(args->num_cliprects, sizeof(*cliprects),
3698                                     GFP_KERNEL);
3699                 if (cliprects == NULL) {
3700                         ret = -ENOMEM;
3701                         goto pre_mutex_err;
3702                 }
3703
3704                 ret = copy_from_user(cliprects,
3705                                      (struct drm_clip_rect __user *)
3706                                      (uintptr_t) args->cliprects_ptr,
3707                                      sizeof(*cliprects) * args->num_cliprects);
3708                 if (ret != 0) {
3709                         DRM_ERROR("copy %d cliprects failed: %d\n",
3710                                   args->num_cliprects, ret);
3711                         goto pre_mutex_err;
3712                 }
3713         }
3714
3715         ret = i915_gem_get_relocs_from_user(exec_list, args->buffer_count,
3716                                             &relocs);
3717         if (ret != 0)
3718                 goto pre_mutex_err;
3719
3720         mutex_lock(&dev->struct_mutex);
3721
3722         i915_verify_inactive(dev, __FILE__, __LINE__);
3723
3724         if (atomic_read(&dev_priv->mm.wedged)) {
3725                 mutex_unlock(&dev->struct_mutex);
3726                 ret = -EIO;
3727                 goto pre_mutex_err;
3728         }
3729
3730         if (dev_priv->mm.suspended) {
3731                 mutex_unlock(&dev->struct_mutex);
3732                 ret = -EBUSY;
3733                 goto pre_mutex_err;
3734         }
3735
3736         /* Look up object handles */
3737         flips = 0;
3738         for (i = 0; i < args->buffer_count; i++) {
3739                 object_list[i] = drm_gem_object_lookup(dev, file_priv,
3740                                                        exec_list[i].handle);
3741                 if (object_list[i] == NULL) {
3742                         DRM_ERROR("Invalid object handle %d at index %d\n",
3743                                    exec_list[i].handle, i);
3744                         /* prevent error path from reading uninitialized data */
3745                         args->buffer_count = i + 1;
3746                         ret = -EBADF;
3747                         goto err;
3748                 }
3749
3750                 obj_priv = to_intel_bo(object_list[i]);
3751                 if (obj_priv->in_execbuffer) {
3752                         DRM_ERROR("Object %p appears more than once in object list\n",
3753                                    object_list[i]);
3754                         /* prevent error path from reading uninitialized data */
3755                         args->buffer_count = i + 1;
3756                         ret = -EBADF;
3757                         goto err;
3758                 }
3759                 obj_priv->in_execbuffer = true;
3760                 flips += atomic_read(&obj_priv->pending_flip);
3761         }
3762
3763         if (flips > 0) {
3764                 ret = i915_gem_wait_for_pending_flip(dev, object_list,
3765                                                      args->buffer_count);
3766                 if (ret)
3767                         goto err;
3768         }
3769
3770         /* Pin and relocate */
3771         for (pin_tries = 0; ; pin_tries++) {
3772                 ret = 0;
3773                 reloc_index = 0;
3774
3775                 for (i = 0; i < args->buffer_count; i++) {
3776                         object_list[i]->pending_read_domains = 0;
3777                         object_list[i]->pending_write_domain = 0;
3778                         ret = i915_gem_object_pin_and_relocate(object_list[i],
3779                                                                file_priv,
3780                                                                &exec_list[i],
3781                                                                &relocs[reloc_index]);
3782                         if (ret)
3783                                 break;
3784                         pinned = i + 1;
3785                         reloc_index += exec_list[i].relocation_count;
3786                 }
3787                 /* success */
3788                 if (ret == 0)
3789                         break;
3790
3791                 /* error other than GTT full, or we've already tried again */
3792                 if (ret != -ENOSPC || pin_tries >= 1) {
3793                         if (ret != -ERESTARTSYS) {
3794                                 unsigned long long total_size = 0;
3795                                 int num_fences = 0;
3796                                 for (i = 0; i < args->buffer_count; i++) {
3797                                         obj_priv = object_list[i]->driver_private;
3798
3799                                         total_size += object_list[i]->size;
3800                                         num_fences +=
3801                                                 exec_list[i].flags & EXEC_OBJECT_NEEDS_FENCE &&
3802                                                 obj_priv->tiling_mode != I915_TILING_NONE;
3803                                 }
3804                                 DRM_ERROR("Failed to pin buffer %d of %d, total %llu bytes, %d fences: %d\n",
3805                                           pinned+1, args->buffer_count,
3806                                           total_size, num_fences,
3807                                           ret);
3808                                 DRM_ERROR("%d objects [%d pinned], "
3809                                           "%d object bytes [%d pinned], "
3810                                           "%d/%d gtt bytes\n",
3811                                           atomic_read(&dev->object_count),
3812                                           atomic_read(&dev->pin_count),
3813                                           atomic_read(&dev->object_memory),
3814                                           atomic_read(&dev->pin_memory),
3815                                           atomic_read(&dev->gtt_memory),
3816                                           dev->gtt_total);
3817                         }
3818                         goto err;
3819                 }
3820
3821                 /* unpin all of our buffers */
3822                 for (i = 0; i < pinned; i++)
3823                         i915_gem_object_unpin(object_list[i]);
3824                 pinned = 0;
3825
3826                 /* evict everyone we can from the aperture */
3827                 ret = i915_gem_evict_everything(dev);
3828                 if (ret && ret != -ENOSPC)
3829                         goto err;
3830         }
3831
3832         /* Set the pending read domains for the batch buffer to COMMAND */
3833         batch_obj = object_list[args->buffer_count-1];
3834         if (batch_obj->pending_write_domain) {
3835                 DRM_ERROR("Attempting to use self-modifying batch buffer\n");
3836                 ret = -EINVAL;
3837                 goto err;
3838         }
3839         batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
3840
3841         /* Sanity check the batch buffer, prior to moving objects */
3842         exec_offset = exec_list[args->buffer_count - 1].offset;
3843         ret = i915_gem_check_execbuffer (args, exec_offset);
3844         if (ret != 0) {
3845                 DRM_ERROR("execbuf with invalid offset/length\n");
3846                 goto err;
3847         }
3848
3849         i915_verify_inactive(dev, __FILE__, __LINE__);
3850
3851         /* Zero the global flush/invalidate flags. These
3852          * will be modified as new domains are computed
3853          * for each object
3854          */
3855         dev->invalidate_domains = 0;
3856         dev->flush_domains = 0;
3857
3858         for (i = 0; i < args->buffer_count; i++) {
3859                 struct drm_gem_object *obj = object_list[i];
3860
3861                 /* Compute new gpu domains and update invalidate/flush */
3862                 i915_gem_object_set_to_gpu_domain(obj);
3863         }
3864
3865         i915_verify_inactive(dev, __FILE__, __LINE__);
3866
3867         if (dev->invalidate_domains | dev->flush_domains) {
3868 #if WATCH_EXEC
3869                 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
3870                           __func__,
3871                          dev->invalidate_domains,
3872                          dev->flush_domains);
3873 #endif
3874                 i915_gem_flush(dev,
3875                                dev->invalidate_domains,
3876                                dev->flush_domains);
3877                 if (dev->flush_domains & I915_GEM_GPU_DOMAINS) {
3878                         (void)i915_add_request(dev, file_priv,
3879                                         dev->flush_domains,
3880                                         &dev_priv->render_ring);
3881
3882                         if (HAS_BSD(dev))
3883                                 (void)i915_add_request(dev, file_priv,
3884                                                 dev->flush_domains,
3885                                                 &dev_priv->bsd_ring);
3886                 }
3887         }
3888
3889         for (i = 0; i < args->buffer_count; i++) {
3890                 struct drm_gem_object *obj = object_list[i];
3891                 struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
3892                 uint32_t old_write_domain = obj->write_domain;
3893
3894                 obj->write_domain = obj->pending_write_domain;
3895                 if (obj->write_domain)
3896                         list_move_tail(&obj_priv->gpu_write_list,
3897                                        &dev_priv->mm.gpu_write_list);
3898                 else
3899                         list_del_init(&obj_priv->gpu_write_list);
3900
3901                 trace_i915_gem_object_change_domain(obj,
3902                                                     obj->read_domains,
3903                                                     old_write_domain);
3904         }
3905
3906         i915_verify_inactive(dev, __FILE__, __LINE__);
3907
3908 #if WATCH_COHERENCY
3909         for (i = 0; i < args->buffer_count; i++) {
3910                 i915_gem_object_check_coherency(object_list[i],
3911                                                 exec_list[i].handle);
3912         }
3913 #endif
3914
3915 #if WATCH_EXEC
3916         i915_gem_dump_object(batch_obj,
3917                               args->batch_len,
3918                               __func__,
3919                               ~0);
3920 #endif
3921
3922         /* Exec the batchbuffer */
3923         ret = ring->dispatch_gem_execbuffer(dev, ring, args,
3924                         cliprects, exec_offset);
3925         if (ret) {
3926                 DRM_ERROR("dispatch failed %d\n", ret);
3927                 goto err;
3928         }
3929
3930         /*
3931          * Ensure that the commands in the batch buffer are
3932          * finished before the interrupt fires
3933          */
3934         flush_domains = i915_retire_commands(dev, ring);
3935
3936         i915_verify_inactive(dev, __FILE__, __LINE__);
3937
3938         /*
3939          * Get a seqno representing the execution of the current buffer,
3940          * which we can wait on.  We would like to mitigate these interrupts,
3941          * likely by only creating seqnos occasionally (so that we have
3942          * *some* interrupts representing completion of buffers that we can
3943          * wait on when trying to clear up gtt space).
3944          */
3945         seqno = i915_add_request(dev, file_priv, flush_domains, ring);
3946         BUG_ON(seqno == 0);
3947         for (i = 0; i < args->buffer_count; i++) {
3948                 struct drm_gem_object *obj = object_list[i];
3949                 obj_priv = to_intel_bo(obj);
3950
3951                 i915_gem_object_move_to_active(obj, seqno, ring);
3952 #if WATCH_LRU
3953                 DRM_INFO("%s: move to exec list %p\n", __func__, obj);
3954 #endif
3955         }
3956 #if WATCH_LRU
3957         i915_dump_lru(dev, __func__);
3958 #endif
3959
3960         i915_verify_inactive(dev, __FILE__, __LINE__);
3961
3962 err:
3963         for (i = 0; i < pinned; i++)
3964                 i915_gem_object_unpin(object_list[i]);
3965
3966         for (i = 0; i < args->buffer_count; i++) {
3967                 if (object_list[i]) {
3968                         obj_priv = to_intel_bo(object_list[i]);
3969                         obj_priv->in_execbuffer = false;
3970                 }
3971                 drm_gem_object_unreference(object_list[i]);
3972         }
3973
3974         mutex_unlock(&dev->struct_mutex);
3975
3976 pre_mutex_err:
3977         /* Copy the updated relocations out regardless of current error
3978          * state.  Failure to update the relocs would mean that the next
3979          * time userland calls execbuf, it would do so with presumed offset
3980          * state that didn't match the actual object state.
3981          */
3982         ret2 = i915_gem_put_relocs_to_user(exec_list, args->buffer_count,
3983                                            relocs);
3984         if (ret2 != 0) {
3985                 DRM_ERROR("Failed to copy relocations back out: %d\n", ret2);
3986
3987                 if (ret == 0)
3988                         ret = ret2;
3989         }
3990
3991         drm_free_large(object_list);
3992         kfree(cliprects);
3993
3994         return ret;
3995 }
3996
3997 /*
3998  * Legacy execbuffer just creates an exec2 list from the original exec object
3999  * list array and passes it to the real function.
4000  */
4001 int
4002 i915_gem_execbuffer(struct drm_device *dev, void *data,
4003                     struct drm_file *file_priv)
4004 {
4005         struct drm_i915_gem_execbuffer *args = data;
4006         struct drm_i915_gem_execbuffer2 exec2;
4007         struct drm_i915_gem_exec_object *exec_list = NULL;
4008         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4009         int ret, i;
4010
4011 #if WATCH_EXEC
4012         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4013                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4014 #endif
4015
4016         if (args->buffer_count < 1) {
4017                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
4018                 return -EINVAL;
4019         }
4020
4021         /* Copy in the exec list from userland */
4022         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
4023         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4024         if (exec_list == NULL || exec2_list == NULL) {
4025                 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4026                           args->buffer_count);
4027                 drm_free_large(exec_list);
4028                 drm_free_large(exec2_list);
4029                 return -ENOMEM;
4030         }
4031         ret = copy_from_user(exec_list,
4032                              (struct drm_i915_relocation_entry __user *)
4033                              (uintptr_t) args->buffers_ptr,
4034                              sizeof(*exec_list) * args->buffer_count);
4035         if (ret != 0) {
4036                 DRM_ERROR("copy %d exec entries failed %d\n",
4037                           args->buffer_count, ret);
4038                 drm_free_large(exec_list);
4039                 drm_free_large(exec2_list);
4040                 return -EFAULT;
4041         }
4042
4043         for (i = 0; i < args->buffer_count; i++) {
4044                 exec2_list[i].handle = exec_list[i].handle;
4045                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
4046                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
4047                 exec2_list[i].alignment = exec_list[i].alignment;
4048                 exec2_list[i].offset = exec_list[i].offset;
4049                 if (!IS_I965G(dev))
4050                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
4051                 else
4052                         exec2_list[i].flags = 0;
4053         }
4054
4055         exec2.buffers_ptr = args->buffers_ptr;
4056         exec2.buffer_count = args->buffer_count;
4057         exec2.batch_start_offset = args->batch_start_offset;
4058         exec2.batch_len = args->batch_len;
4059         exec2.DR1 = args->DR1;
4060         exec2.DR4 = args->DR4;
4061         exec2.num_cliprects = args->num_cliprects;
4062         exec2.cliprects_ptr = args->cliprects_ptr;
4063         exec2.flags = I915_EXEC_RENDER;
4064
4065         ret = i915_gem_do_execbuffer(dev, data, file_priv, &exec2, exec2_list);
4066         if (!ret) {
4067                 /* Copy the new buffer offsets back to the user's exec list. */
4068                 for (i = 0; i < args->buffer_count; i++)
4069                         exec_list[i].offset = exec2_list[i].offset;
4070                 /* ... and back out to userspace */
4071                 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4072                                    (uintptr_t) args->buffers_ptr,
4073                                    exec_list,
4074                                    sizeof(*exec_list) * args->buffer_count);
4075                 if (ret) {
4076                         ret = -EFAULT;
4077                         DRM_ERROR("failed to copy %d exec entries "
4078                                   "back to user (%d)\n",
4079                                   args->buffer_count, ret);
4080                 }
4081         }
4082
4083         drm_free_large(exec_list);
4084         drm_free_large(exec2_list);
4085         return ret;
4086 }
4087
4088 int
4089 i915_gem_execbuffer2(struct drm_device *dev, void *data,
4090                      struct drm_file *file_priv)
4091 {
4092         struct drm_i915_gem_execbuffer2 *args = data;
4093         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
4094         int ret;
4095
4096 #if WATCH_EXEC
4097         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
4098                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
4099 #endif
4100
4101         if (args->buffer_count < 1) {
4102                 DRM_ERROR("execbuf2 with %d buffers\n", args->buffer_count);
4103                 return -EINVAL;
4104         }
4105
4106         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
4107         if (exec2_list == NULL) {
4108                 DRM_ERROR("Failed to allocate exec list for %d buffers\n",
4109                           args->buffer_count);
4110                 return -ENOMEM;
4111         }
4112         ret = copy_from_user(exec2_list,
4113                              (struct drm_i915_relocation_entry __user *)
4114                              (uintptr_t) args->buffers_ptr,
4115                              sizeof(*exec2_list) * args->buffer_count);
4116         if (ret != 0) {
4117                 DRM_ERROR("copy %d exec entries failed %d\n",
4118                           args->buffer_count, ret);
4119                 drm_free_large(exec2_list);
4120                 return -EFAULT;
4121         }
4122
4123         ret = i915_gem_do_execbuffer(dev, data, file_priv, args, exec2_list);
4124         if (!ret) {
4125                 /* Copy the new buffer offsets back to the user's exec list. */
4126                 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
4127                                    (uintptr_t) args->buffers_ptr,
4128                                    exec2_list,
4129                                    sizeof(*exec2_list) * args->buffer_count);
4130                 if (ret) {
4131                         ret = -EFAULT;
4132                         DRM_ERROR("failed to copy %d exec entries "
4133                                   "back to user (%d)\n",
4134                                   args->buffer_count, ret);
4135                 }
4136         }
4137
4138         drm_free_large(exec2_list);
4139         return ret;
4140 }
4141
4142 int
4143 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
4144 {
4145         struct drm_device *dev = obj->dev;
4146         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4147         int ret;
4148
4149         BUG_ON(obj_priv->pin_count == DRM_I915_GEM_OBJECT_MAX_PIN_COUNT);
4150
4151         i915_verify_inactive(dev, __FILE__, __LINE__);
4152
4153         if (obj_priv->gtt_space != NULL) {
4154                 if (alignment == 0)
4155                         alignment = i915_gem_get_gtt_alignment(obj);
4156                 if (obj_priv->gtt_offset & (alignment - 1)) {
4157                         ret = i915_gem_object_unbind(obj);
4158                         if (ret)
4159                                 return ret;
4160                 }
4161         }
4162
4163         if (obj_priv->gtt_space == NULL) {
4164                 ret = i915_gem_object_bind_to_gtt(obj, alignment);
4165                 if (ret)
4166                         return ret;
4167         }
4168
4169         obj_priv->pin_count++;
4170
4171         /* If the object is not active and not pending a flush,
4172          * remove it from the inactive list
4173          */
4174         if (obj_priv->pin_count == 1) {
4175                 atomic_inc(&dev->pin_count);
4176                 atomic_add(obj->size, &dev->pin_memory);
4177                 if (!obj_priv->active &&
4178                     (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0 &&
4179                     !list_empty(&obj_priv->list))
4180                         list_del_init(&obj_priv->list);
4181         }
4182         i915_verify_inactive(dev, __FILE__, __LINE__);
4183
4184         return 0;
4185 }
4186
4187 void
4188 i915_gem_object_unpin(struct drm_gem_object *obj)
4189 {
4190         struct drm_device *dev = obj->dev;
4191         drm_i915_private_t *dev_priv = dev->dev_private;
4192         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4193
4194         i915_verify_inactive(dev, __FILE__, __LINE__);
4195         obj_priv->pin_count--;
4196         BUG_ON(obj_priv->pin_count < 0);
4197         BUG_ON(obj_priv->gtt_space == NULL);
4198
4199         /* If the object is no longer pinned, and is
4200          * neither active nor being flushed, then stick it on
4201          * the inactive list
4202          */
4203         if (obj_priv->pin_count == 0) {
4204                 if (!obj_priv->active &&
4205                     (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0)
4206                         list_move_tail(&obj_priv->list,
4207                                        &dev_priv->mm.inactive_list);
4208                 atomic_dec(&dev->pin_count);
4209                 atomic_sub(obj->size, &dev->pin_memory);
4210         }
4211         i915_verify_inactive(dev, __FILE__, __LINE__);
4212 }
4213
4214 int
4215 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
4216                    struct drm_file *file_priv)
4217 {
4218         struct drm_i915_gem_pin *args = data;
4219         struct drm_gem_object *obj;
4220         struct drm_i915_gem_object *obj_priv;
4221         int ret;
4222
4223         mutex_lock(&dev->struct_mutex);
4224
4225         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4226         if (obj == NULL) {
4227                 DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
4228                           args->handle);
4229                 mutex_unlock(&dev->struct_mutex);
4230                 return -EBADF;
4231         }
4232         obj_priv = to_intel_bo(obj);
4233
4234         if (obj_priv->madv != I915_MADV_WILLNEED) {
4235                 DRM_ERROR("Attempting to pin a purgeable buffer\n");
4236                 drm_gem_object_unreference(obj);
4237                 mutex_unlock(&dev->struct_mutex);
4238                 return -EINVAL;
4239         }
4240
4241         if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) {
4242                 DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n",
4243                           args->handle);
4244                 drm_gem_object_unreference(obj);
4245                 mutex_unlock(&dev->struct_mutex);
4246                 return -EINVAL;
4247         }
4248
4249         obj_priv->user_pin_count++;
4250         obj_priv->pin_filp = file_priv;
4251         if (obj_priv->user_pin_count == 1) {
4252                 ret = i915_gem_object_pin(obj, args->alignment);
4253                 if (ret != 0) {
4254                         drm_gem_object_unreference(obj);
4255                         mutex_unlock(&dev->struct_mutex);
4256                         return ret;
4257                 }
4258         }
4259
4260         /* XXX - flush the CPU caches for pinned objects
4261          * as the X server doesn't manage domains yet
4262          */
4263         i915_gem_object_flush_cpu_write_domain(obj);
4264         args->offset = obj_priv->gtt_offset;
4265         drm_gem_object_unreference(obj);
4266         mutex_unlock(&dev->struct_mutex);
4267
4268         return 0;
4269 }
4270
4271 int
4272 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
4273                      struct drm_file *file_priv)
4274 {
4275         struct drm_i915_gem_pin *args = data;
4276         struct drm_gem_object *obj;
4277         struct drm_i915_gem_object *obj_priv;
4278
4279         mutex_lock(&dev->struct_mutex);
4280
4281         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4282         if (obj == NULL) {
4283                 DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
4284                           args->handle);
4285                 mutex_unlock(&dev->struct_mutex);
4286                 return -EBADF;
4287         }
4288
4289         obj_priv = to_intel_bo(obj);
4290         if (obj_priv->pin_filp != file_priv) {
4291                 DRM_ERROR("Not pinned by caller in i915_gem_pin_ioctl(): %d\n",
4292                           args->handle);
4293                 drm_gem_object_unreference(obj);
4294                 mutex_unlock(&dev->struct_mutex);
4295                 return -EINVAL;
4296         }
4297         obj_priv->user_pin_count--;
4298         if (obj_priv->user_pin_count == 0) {
4299                 obj_priv->pin_filp = NULL;
4300                 i915_gem_object_unpin(obj);
4301         }
4302
4303         drm_gem_object_unreference(obj);
4304         mutex_unlock(&dev->struct_mutex);
4305         return 0;
4306 }
4307
4308 int
4309 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
4310                     struct drm_file *file_priv)
4311 {
4312         struct drm_i915_gem_busy *args = data;
4313         struct drm_gem_object *obj;
4314         struct drm_i915_gem_object *obj_priv;
4315         drm_i915_private_t *dev_priv = dev->dev_private;
4316
4317         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4318         if (obj == NULL) {
4319                 DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
4320                           args->handle);
4321                 return -EBADF;
4322         }
4323
4324         mutex_lock(&dev->struct_mutex);
4325         /* Update the active list for the hardware's current position.
4326          * Otherwise this only updates on a delayed timer or when irqs are
4327          * actually unmasked, and our working set ends up being larger than
4328          * required.
4329          */
4330         i915_gem_retire_requests(dev, &dev_priv->render_ring);
4331
4332         if (HAS_BSD(dev))
4333                 i915_gem_retire_requests(dev, &dev_priv->bsd_ring);
4334
4335         obj_priv = to_intel_bo(obj);
4336         /* Don't count being on the flushing list against the object being
4337          * done.  Otherwise, a buffer left on the flushing list but not getting
4338          * flushed (because nobody's flushing that domain) won't ever return
4339          * unbusy and get reused by libdrm's bo cache.  The other expected
4340          * consumer of this interface, OpenGL's occlusion queries, also specs
4341          * that the objects get unbusy "eventually" without any interference.
4342          */
4343         args->busy = obj_priv->active && obj_priv->last_rendering_seqno != 0;
4344
4345         drm_gem_object_unreference(obj);
4346         mutex_unlock(&dev->struct_mutex);
4347         return 0;
4348 }
4349
4350 int
4351 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
4352                         struct drm_file *file_priv)
4353 {
4354     return i915_gem_ring_throttle(dev, file_priv);
4355 }
4356
4357 int
4358 i915_gem_madvise_ioctl(struct drm_device *dev, void *data,
4359                        struct drm_file *file_priv)
4360 {
4361         struct drm_i915_gem_madvise *args = data;
4362         struct drm_gem_object *obj;
4363         struct drm_i915_gem_object *obj_priv;
4364
4365         switch (args->madv) {
4366         case I915_MADV_DONTNEED:
4367         case I915_MADV_WILLNEED:
4368             break;
4369         default:
4370             return -EINVAL;
4371         }
4372
4373         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
4374         if (obj == NULL) {
4375                 DRM_ERROR("Bad handle in i915_gem_madvise_ioctl(): %d\n",
4376                           args->handle);
4377                 return -EBADF;
4378         }
4379
4380         mutex_lock(&dev->struct_mutex);
4381         obj_priv = to_intel_bo(obj);
4382
4383         if (obj_priv->pin_count) {
4384                 drm_gem_object_unreference(obj);
4385                 mutex_unlock(&dev->struct_mutex);
4386
4387                 DRM_ERROR("Attempted i915_gem_madvise_ioctl() on a pinned object\n");
4388                 return -EINVAL;
4389         }
4390
4391         if (obj_priv->madv != __I915_MADV_PURGED)
4392                 obj_priv->madv = args->madv;
4393
4394         /* if the object is no longer bound, discard its backing storage */
4395         if (i915_gem_object_is_purgeable(obj_priv) &&
4396             obj_priv->gtt_space == NULL)
4397                 i915_gem_object_truncate(obj);
4398
4399         args->retained = obj_priv->madv != __I915_MADV_PURGED;
4400
4401         drm_gem_object_unreference(obj);
4402         mutex_unlock(&dev->struct_mutex);
4403
4404         return 0;
4405 }
4406
4407 struct drm_gem_object * i915_gem_alloc_object(struct drm_device *dev,
4408                                               size_t size)
4409 {
4410         struct drm_i915_gem_object *obj;
4411
4412         obj = kzalloc(sizeof(*obj), GFP_KERNEL);
4413         if (obj == NULL)
4414                 return NULL;
4415
4416         if (drm_gem_object_init(dev, &obj->base, size) != 0) {
4417                 kfree(obj);
4418                 return NULL;
4419         }
4420
4421         obj->base.write_domain = I915_GEM_DOMAIN_CPU;
4422         obj->base.read_domains = I915_GEM_DOMAIN_CPU;
4423
4424         obj->agp_type = AGP_USER_MEMORY;
4425         obj->base.driver_private = NULL;
4426         obj->fence_reg = I915_FENCE_REG_NONE;
4427         INIT_LIST_HEAD(&obj->list);
4428         INIT_LIST_HEAD(&obj->gpu_write_list);
4429         obj->madv = I915_MADV_WILLNEED;
4430
4431         trace_i915_gem_object_create(&obj->base);
4432
4433         return &obj->base;
4434 }
4435
4436 int i915_gem_init_object(struct drm_gem_object *obj)
4437 {
4438         BUG();
4439
4440         return 0;
4441 }
4442
4443 void i915_gem_free_object(struct drm_gem_object *obj)
4444 {
4445         struct drm_device *dev = obj->dev;
4446         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4447
4448         trace_i915_gem_object_destroy(obj);
4449
4450         while (obj_priv->pin_count > 0)
4451                 i915_gem_object_unpin(obj);
4452
4453         if (obj_priv->phys_obj)
4454                 i915_gem_detach_phys_object(dev, obj);
4455
4456         i915_gem_object_unbind(obj);
4457
4458         if (obj_priv->mmap_offset)
4459                 i915_gem_free_mmap_offset(obj);
4460
4461         drm_gem_object_release(obj);
4462
4463         kfree(obj_priv->page_cpu_valid);
4464         kfree(obj_priv->bit_17);
4465         kfree(obj_priv);
4466 }
4467
4468 /** Unbinds all inactive objects. */
4469 static int
4470 i915_gem_evict_from_inactive_list(struct drm_device *dev)
4471 {
4472         drm_i915_private_t *dev_priv = dev->dev_private;
4473
4474         while (!list_empty(&dev_priv->mm.inactive_list)) {
4475                 struct drm_gem_object *obj;
4476                 int ret;
4477
4478                 obj = &list_first_entry(&dev_priv->mm.inactive_list,
4479                                         struct drm_i915_gem_object,
4480                                         list)->base;
4481
4482                 ret = i915_gem_object_unbind(obj);
4483                 if (ret != 0) {
4484                         DRM_ERROR("Error unbinding object: %d\n", ret);
4485                         return ret;
4486                 }
4487         }
4488
4489         return 0;
4490 }
4491
4492 int
4493 i915_gem_idle(struct drm_device *dev)
4494 {
4495         drm_i915_private_t *dev_priv = dev->dev_private;
4496         int ret;
4497
4498         mutex_lock(&dev->struct_mutex);
4499
4500         if (dev_priv->mm.suspended ||
4501                         (dev_priv->render_ring.gem_object == NULL) ||
4502                         (HAS_BSD(dev) &&
4503                          dev_priv->bsd_ring.gem_object == NULL)) {
4504                 mutex_unlock(&dev->struct_mutex);
4505                 return 0;
4506         }
4507
4508         ret = i915_gpu_idle(dev);
4509         if (ret) {
4510                 mutex_unlock(&dev->struct_mutex);
4511                 return ret;
4512         }
4513
4514         /* Under UMS, be paranoid and evict. */
4515         if (!drm_core_check_feature(dev, DRIVER_MODESET)) {
4516                 ret = i915_gem_evict_from_inactive_list(dev);
4517                 if (ret) {
4518                         mutex_unlock(&dev->struct_mutex);
4519                         return ret;
4520                 }
4521         }
4522
4523         /* Hack!  Don't let anybody do execbuf while we don't control the chip.
4524          * We need to replace this with a semaphore, or something.
4525          * And not confound mm.suspended!
4526          */
4527         dev_priv->mm.suspended = 1;
4528         del_timer(&dev_priv->hangcheck_timer);
4529
4530         i915_kernel_lost_context(dev);
4531         i915_gem_cleanup_ringbuffer(dev);
4532
4533         mutex_unlock(&dev->struct_mutex);
4534
4535         /* Cancel the retire work handler, which should be idle now. */
4536         cancel_delayed_work_sync(&dev_priv->mm.retire_work);
4537
4538         return 0;
4539 }
4540
4541 /*
4542  * 965+ support PIPE_CONTROL commands, which provide finer grained control
4543  * over cache flushing.
4544  */
4545 static int
4546 i915_gem_init_pipe_control(struct drm_device *dev)
4547 {
4548         drm_i915_private_t *dev_priv = dev->dev_private;
4549         struct drm_gem_object *obj;
4550         struct drm_i915_gem_object *obj_priv;
4551         int ret;
4552
4553         obj = i915_gem_alloc_object(dev, 4096);
4554         if (obj == NULL) {
4555                 DRM_ERROR("Failed to allocate seqno page\n");
4556                 ret = -ENOMEM;
4557                 goto err;
4558         }
4559         obj_priv = to_intel_bo(obj);
4560         obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
4561
4562         ret = i915_gem_object_pin(obj, 4096);
4563         if (ret)
4564                 goto err_unref;
4565
4566         dev_priv->seqno_gfx_addr = obj_priv->gtt_offset;
4567         dev_priv->seqno_page =  kmap(obj_priv->pages[0]);
4568         if (dev_priv->seqno_page == NULL)
4569                 goto err_unpin;
4570
4571         dev_priv->seqno_obj = obj;
4572         memset(dev_priv->seqno_page, 0, PAGE_SIZE);
4573
4574         return 0;
4575
4576 err_unpin:
4577         i915_gem_object_unpin(obj);
4578 err_unref:
4579         drm_gem_object_unreference(obj);
4580 err:
4581         return ret;
4582 }
4583
4584
4585 static void
4586 i915_gem_cleanup_pipe_control(struct drm_device *dev)
4587 {
4588         drm_i915_private_t *dev_priv = dev->dev_private;
4589         struct drm_gem_object *obj;
4590         struct drm_i915_gem_object *obj_priv;
4591
4592         obj = dev_priv->seqno_obj;
4593         obj_priv = to_intel_bo(obj);
4594         kunmap(obj_priv->pages[0]);
4595         i915_gem_object_unpin(obj);
4596         drm_gem_object_unreference(obj);
4597         dev_priv->seqno_obj = NULL;
4598
4599         dev_priv->seqno_page = NULL;
4600 }
4601
4602 int
4603 i915_gem_init_ringbuffer(struct drm_device *dev)
4604 {
4605         drm_i915_private_t *dev_priv = dev->dev_private;
4606         int ret;
4607
4608         dev_priv->render_ring = render_ring;
4609
4610         if (!I915_NEED_GFX_HWS(dev)) {
4611                 dev_priv->render_ring.status_page.page_addr
4612                         = dev_priv->status_page_dmah->vaddr;
4613                 memset(dev_priv->render_ring.status_page.page_addr,
4614                                 0, PAGE_SIZE);
4615         }
4616
4617         if (HAS_PIPE_CONTROL(dev)) {
4618                 ret = i915_gem_init_pipe_control(dev);
4619                 if (ret)
4620                         return ret;
4621         }
4622
4623         ret = intel_init_ring_buffer(dev, &dev_priv->render_ring);
4624         if (ret)
4625                 goto cleanup_pipe_control;
4626
4627         if (HAS_BSD(dev)) {
4628                 dev_priv->bsd_ring = bsd_ring;
4629                 ret = intel_init_ring_buffer(dev, &dev_priv->bsd_ring);
4630                 if (ret)
4631                         goto cleanup_render_ring;
4632         }
4633
4634         return 0;
4635
4636 cleanup_render_ring:
4637         intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4638 cleanup_pipe_control:
4639         if (HAS_PIPE_CONTROL(dev))
4640                 i915_gem_cleanup_pipe_control(dev);
4641         return ret;
4642 }
4643
4644 void
4645 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
4646 {
4647         drm_i915_private_t *dev_priv = dev->dev_private;
4648
4649         intel_cleanup_ring_buffer(dev, &dev_priv->render_ring);
4650         if (HAS_BSD(dev))
4651                 intel_cleanup_ring_buffer(dev, &dev_priv->bsd_ring);
4652         if (HAS_PIPE_CONTROL(dev))
4653                 i915_gem_cleanup_pipe_control(dev);
4654 }
4655
4656 int
4657 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
4658                        struct drm_file *file_priv)
4659 {
4660         drm_i915_private_t *dev_priv = dev->dev_private;
4661         int ret;
4662
4663         if (drm_core_check_feature(dev, DRIVER_MODESET))
4664                 return 0;
4665
4666         if (atomic_read(&dev_priv->mm.wedged)) {
4667                 DRM_ERROR("Reenabling wedged hardware, good luck\n");
4668                 atomic_set(&dev_priv->mm.wedged, 0);
4669         }
4670
4671         mutex_lock(&dev->struct_mutex);
4672         dev_priv->mm.suspended = 0;
4673
4674         ret = i915_gem_init_ringbuffer(dev);
4675         if (ret != 0) {
4676                 mutex_unlock(&dev->struct_mutex);
4677                 return ret;
4678         }
4679
4680         spin_lock(&dev_priv->mm.active_list_lock);
4681         BUG_ON(!list_empty(&dev_priv->render_ring.active_list));
4682         BUG_ON(HAS_BSD(dev) && !list_empty(&dev_priv->bsd_ring.active_list));
4683         spin_unlock(&dev_priv->mm.active_list_lock);
4684
4685         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
4686         BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
4687         BUG_ON(!list_empty(&dev_priv->render_ring.request_list));
4688         BUG_ON(HAS_BSD(dev) && !list_empty(&dev_priv->bsd_ring.request_list));
4689         mutex_unlock(&dev->struct_mutex);
4690
4691         drm_irq_install(dev);
4692
4693         return 0;
4694 }
4695
4696 int
4697 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
4698                        struct drm_file *file_priv)
4699 {
4700         if (drm_core_check_feature(dev, DRIVER_MODESET))
4701                 return 0;
4702
4703         drm_irq_uninstall(dev);
4704         return i915_gem_idle(dev);
4705 }
4706
4707 void
4708 i915_gem_lastclose(struct drm_device *dev)
4709 {
4710         int ret;
4711
4712         if (drm_core_check_feature(dev, DRIVER_MODESET))
4713                 return;
4714
4715         ret = i915_gem_idle(dev);
4716         if (ret)
4717                 DRM_ERROR("failed to idle hardware: %d\n", ret);
4718 }
4719
4720 void
4721 i915_gem_load(struct drm_device *dev)
4722 {
4723         int i;
4724         drm_i915_private_t *dev_priv = dev->dev_private;
4725
4726         spin_lock_init(&dev_priv->mm.active_list_lock);
4727         INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
4728         INIT_LIST_HEAD(&dev_priv->mm.gpu_write_list);
4729         INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
4730         INIT_LIST_HEAD(&dev_priv->mm.fence_list);
4731         INIT_LIST_HEAD(&dev_priv->render_ring.active_list);
4732         INIT_LIST_HEAD(&dev_priv->render_ring.request_list);
4733         if (HAS_BSD(dev)) {
4734                 INIT_LIST_HEAD(&dev_priv->bsd_ring.active_list);
4735                 INIT_LIST_HEAD(&dev_priv->bsd_ring.request_list);
4736         }
4737         for (i = 0; i < 16; i++)
4738                 INIT_LIST_HEAD(&dev_priv->fence_regs[i].lru_list);
4739         INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
4740                           i915_gem_retire_work_handler);
4741         spin_lock(&shrink_list_lock);
4742         list_add(&dev_priv->mm.shrink_list, &shrink_list);
4743         spin_unlock(&shrink_list_lock);
4744
4745         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
4746         if (IS_GEN3(dev)) {
4747                 u32 tmp = I915_READ(MI_ARB_STATE);
4748                 if (!(tmp & MI_ARB_C3_LP_WRITE_ENABLE)) {
4749                         /* arb state is a masked write, so set bit + bit in mask */
4750                         tmp = MI_ARB_C3_LP_WRITE_ENABLE | (MI_ARB_C3_LP_WRITE_ENABLE << MI_ARB_MASK_SHIFT);
4751                         I915_WRITE(MI_ARB_STATE, tmp);
4752                 }
4753         }
4754
4755         /* Old X drivers will take 0-2 for front, back, depth buffers */
4756         if (!drm_core_check_feature(dev, DRIVER_MODESET))
4757                 dev_priv->fence_reg_start = 3;
4758
4759         if (IS_I965G(dev) || IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4760                 dev_priv->num_fence_regs = 16;
4761         else
4762                 dev_priv->num_fence_regs = 8;
4763
4764         /* Initialize fence registers to zero */
4765         if (IS_I965G(dev)) {
4766                 for (i = 0; i < 16; i++)
4767                         I915_WRITE64(FENCE_REG_965_0 + (i * 8), 0);
4768         } else {
4769                 for (i = 0; i < 8; i++)
4770                         I915_WRITE(FENCE_REG_830_0 + (i * 4), 0);
4771                 if (IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev))
4772                         for (i = 0; i < 8; i++)
4773                                 I915_WRITE(FENCE_REG_945_8 + (i * 4), 0);
4774         }
4775         i915_gem_detect_bit_6_swizzle(dev);
4776         init_waitqueue_head(&dev_priv->pending_flip_queue);
4777 }
4778
4779 /*
4780  * Create a physically contiguous memory object for this object
4781  * e.g. for cursor + overlay regs
4782  */
4783 int i915_gem_init_phys_object(struct drm_device *dev,
4784                               int id, int size)
4785 {
4786         drm_i915_private_t *dev_priv = dev->dev_private;
4787         struct drm_i915_gem_phys_object *phys_obj;
4788         int ret;
4789
4790         if (dev_priv->mm.phys_objs[id - 1] || !size)
4791                 return 0;
4792
4793         phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL);
4794         if (!phys_obj)
4795                 return -ENOMEM;
4796
4797         phys_obj->id = id;
4798
4799         phys_obj->handle = drm_pci_alloc(dev, size, 0);
4800         if (!phys_obj->handle) {
4801                 ret = -ENOMEM;
4802                 goto kfree_obj;
4803         }
4804 #ifdef CONFIG_X86
4805         set_memory_wc((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4806 #endif
4807
4808         dev_priv->mm.phys_objs[id - 1] = phys_obj;
4809
4810         return 0;
4811 kfree_obj:
4812         kfree(phys_obj);
4813         return ret;
4814 }
4815
4816 void i915_gem_free_phys_object(struct drm_device *dev, int id)
4817 {
4818         drm_i915_private_t *dev_priv = dev->dev_private;
4819         struct drm_i915_gem_phys_object *phys_obj;
4820
4821         if (!dev_priv->mm.phys_objs[id - 1])
4822                 return;
4823
4824         phys_obj = dev_priv->mm.phys_objs[id - 1];
4825         if (phys_obj->cur_obj) {
4826                 i915_gem_detach_phys_object(dev, phys_obj->cur_obj);
4827         }
4828
4829 #ifdef CONFIG_X86
4830         set_memory_wb((unsigned long)phys_obj->handle->vaddr, phys_obj->handle->size / PAGE_SIZE);
4831 #endif
4832         drm_pci_free(dev, phys_obj->handle);
4833         kfree(phys_obj);
4834         dev_priv->mm.phys_objs[id - 1] = NULL;
4835 }
4836
4837 void i915_gem_free_all_phys_object(struct drm_device *dev)
4838 {
4839         int i;
4840
4841         for (i = I915_GEM_PHYS_CURSOR_0; i <= I915_MAX_PHYS_OBJECT; i++)
4842                 i915_gem_free_phys_object(dev, i);
4843 }
4844
4845 void i915_gem_detach_phys_object(struct drm_device *dev,
4846                                  struct drm_gem_object *obj)
4847 {
4848         struct drm_i915_gem_object *obj_priv;
4849         int i;
4850         int ret;
4851         int page_count;
4852
4853         obj_priv = to_intel_bo(obj);
4854         if (!obj_priv->phys_obj)
4855                 return;
4856
4857         ret = i915_gem_object_get_pages(obj, 0);
4858         if (ret)
4859                 goto out;
4860
4861         page_count = obj->size / PAGE_SIZE;
4862
4863         for (i = 0; i < page_count; i++) {
4864                 char *dst = kmap_atomic(obj_priv->pages[i], KM_USER0);
4865                 char *src = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4866
4867                 memcpy(dst, src, PAGE_SIZE);
4868                 kunmap_atomic(dst, KM_USER0);
4869         }
4870         drm_clflush_pages(obj_priv->pages, page_count);
4871         drm_agp_chipset_flush(dev);
4872
4873         i915_gem_object_put_pages(obj);
4874 out:
4875         obj_priv->phys_obj->cur_obj = NULL;
4876         obj_priv->phys_obj = NULL;
4877 }
4878
4879 int
4880 i915_gem_attach_phys_object(struct drm_device *dev,
4881                             struct drm_gem_object *obj, int id)
4882 {
4883         drm_i915_private_t *dev_priv = dev->dev_private;
4884         struct drm_i915_gem_object *obj_priv;
4885         int ret = 0;
4886         int page_count;
4887         int i;
4888
4889         if (id > I915_MAX_PHYS_OBJECT)
4890                 return -EINVAL;
4891
4892         obj_priv = to_intel_bo(obj);
4893
4894         if (obj_priv->phys_obj) {
4895                 if (obj_priv->phys_obj->id == id)
4896                         return 0;
4897                 i915_gem_detach_phys_object(dev, obj);
4898         }
4899
4900
4901         /* create a new object */
4902         if (!dev_priv->mm.phys_objs[id - 1]) {
4903                 ret = i915_gem_init_phys_object(dev, id,
4904                                                 obj->size);
4905                 if (ret) {
4906                         DRM_ERROR("failed to init phys object %d size: %zu\n", id, obj->size);
4907                         goto out;
4908                 }
4909         }
4910
4911         /* bind to the object */
4912         obj_priv->phys_obj = dev_priv->mm.phys_objs[id - 1];
4913         obj_priv->phys_obj->cur_obj = obj;
4914
4915         ret = i915_gem_object_get_pages(obj, 0);
4916         if (ret) {
4917                 DRM_ERROR("failed to get page list\n");
4918                 goto out;
4919         }
4920
4921         page_count = obj->size / PAGE_SIZE;
4922
4923         for (i = 0; i < page_count; i++) {
4924                 char *src = kmap_atomic(obj_priv->pages[i], KM_USER0);
4925                 char *dst = obj_priv->phys_obj->handle->vaddr + (i * PAGE_SIZE);
4926
4927                 memcpy(dst, src, PAGE_SIZE);
4928                 kunmap_atomic(src, KM_USER0);
4929         }
4930
4931         i915_gem_object_put_pages(obj);
4932
4933         return 0;
4934 out:
4935         return ret;
4936 }
4937
4938 static int
4939 i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
4940                      struct drm_i915_gem_pwrite *args,
4941                      struct drm_file *file_priv)
4942 {
4943         struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
4944         void *obj_addr;
4945         int ret;
4946         char __user *user_data;
4947
4948         user_data = (char __user *) (uintptr_t) args->data_ptr;
4949         obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
4950
4951         DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
4952         ret = copy_from_user(obj_addr, user_data, args->size);
4953         if (ret)
4954                 return -EFAULT;
4955
4956         drm_agp_chipset_flush(dev);
4957         return 0;
4958 }
4959
4960 void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv)
4961 {
4962         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
4963
4964         /* Clean up our request list when the client is going away, so that
4965          * later retire_requests won't dereference our soon-to-be-gone
4966          * file_priv.
4967          */
4968         mutex_lock(&dev->struct_mutex);
4969         while (!list_empty(&i915_file_priv->mm.request_list))
4970                 list_del_init(i915_file_priv->mm.request_list.next);
4971         mutex_unlock(&dev->struct_mutex);
4972 }
4973
4974 static int
4975 i915_gpu_is_active(struct drm_device *dev)
4976 {
4977         drm_i915_private_t *dev_priv = dev->dev_private;
4978         int lists_empty;
4979
4980         spin_lock(&dev_priv->mm.active_list_lock);
4981         lists_empty = list_empty(&dev_priv->mm.flushing_list) &&
4982                       list_empty(&dev_priv->render_ring.active_list);
4983         if (HAS_BSD(dev))
4984                 lists_empty &= list_empty(&dev_priv->bsd_ring.active_list);
4985         spin_unlock(&dev_priv->mm.active_list_lock);
4986
4987         return !lists_empty;
4988 }
4989
4990 static int
4991 i915_gem_shrink(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
4992 {
4993         drm_i915_private_t *dev_priv, *next_dev;
4994         struct drm_i915_gem_object *obj_priv, *next_obj;
4995         int cnt = 0;
4996         int would_deadlock = 1;
4997
4998         /* "fast-path" to count number of available objects */
4999         if (nr_to_scan == 0) {
5000                 spin_lock(&shrink_list_lock);
5001                 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5002                         struct drm_device *dev = dev_priv->dev;
5003
5004                         if (mutex_trylock(&dev->struct_mutex)) {
5005                                 list_for_each_entry(obj_priv,
5006                                                     &dev_priv->mm.inactive_list,
5007                                                     list)
5008                                         cnt++;
5009                                 mutex_unlock(&dev->struct_mutex);
5010                         }
5011                 }
5012                 spin_unlock(&shrink_list_lock);
5013
5014                 return (cnt / 100) * sysctl_vfs_cache_pressure;
5015         }
5016
5017         spin_lock(&shrink_list_lock);
5018
5019 rescan:
5020         /* first scan for clean buffers */
5021         list_for_each_entry_safe(dev_priv, next_dev,
5022                                  &shrink_list, mm.shrink_list) {
5023                 struct drm_device *dev = dev_priv->dev;
5024
5025                 if (! mutex_trylock(&dev->struct_mutex))
5026                         continue;
5027
5028                 spin_unlock(&shrink_list_lock);
5029                 i915_gem_retire_requests(dev, &dev_priv->render_ring);
5030
5031                 if (HAS_BSD(dev))
5032                         i915_gem_retire_requests(dev, &dev_priv->bsd_ring);
5033
5034                 list_for_each_entry_safe(obj_priv, next_obj,
5035                                          &dev_priv->mm.inactive_list,
5036                                          list) {
5037                         if (i915_gem_object_is_purgeable(obj_priv)) {
5038                                 i915_gem_object_unbind(&obj_priv->base);
5039                                 if (--nr_to_scan <= 0)
5040                                         break;
5041                         }
5042                 }
5043
5044                 spin_lock(&shrink_list_lock);
5045                 mutex_unlock(&dev->struct_mutex);
5046
5047                 would_deadlock = 0;
5048
5049                 if (nr_to_scan <= 0)
5050                         break;
5051         }
5052
5053         /* second pass, evict/count anything still on the inactive list */
5054         list_for_each_entry_safe(dev_priv, next_dev,
5055                                  &shrink_list, mm.shrink_list) {
5056                 struct drm_device *dev = dev_priv->dev;
5057
5058                 if (! mutex_trylock(&dev->struct_mutex))
5059                         continue;
5060
5061                 spin_unlock(&shrink_list_lock);
5062
5063                 list_for_each_entry_safe(obj_priv, next_obj,
5064                                          &dev_priv->mm.inactive_list,
5065                                          list) {
5066                         if (nr_to_scan > 0) {
5067                                 i915_gem_object_unbind(&obj_priv->base);
5068                                 nr_to_scan--;
5069                         } else
5070                                 cnt++;
5071                 }
5072
5073                 spin_lock(&shrink_list_lock);
5074                 mutex_unlock(&dev->struct_mutex);
5075
5076                 would_deadlock = 0;
5077         }
5078
5079         if (nr_to_scan) {
5080                 int active = 0;
5081
5082                 /*
5083                  * We are desperate for pages, so as a last resort, wait
5084                  * for the GPU to finish and discard whatever we can.
5085                  * This has a dramatic impact to reduce the number of
5086                  * OOM-killer events whilst running the GPU aggressively.
5087                  */
5088                 list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) {
5089                         struct drm_device *dev = dev_priv->dev;
5090
5091                         if (!mutex_trylock(&dev->struct_mutex))
5092                                 continue;
5093
5094                         spin_unlock(&shrink_list_lock);
5095
5096                         if (i915_gpu_is_active(dev)) {
5097                                 i915_gpu_idle(dev);
5098                                 active++;
5099                         }
5100
5101                         spin_lock(&shrink_list_lock);
5102                         mutex_unlock(&dev->struct_mutex);
5103                 }
5104
5105                 if (active)
5106                         goto rescan;
5107         }
5108
5109         spin_unlock(&shrink_list_lock);
5110
5111         if (would_deadlock)
5112                 return -1;
5113         else if (cnt > 0)
5114                 return (cnt / 100) * sysctl_vfs_cache_pressure;
5115         else
5116                 return 0;
5117 }
5118
5119 static struct shrinker shrinker = {
5120         .shrink = i915_gem_shrink,
5121         .seeks = DEFAULT_SEEKS,
5122 };
5123
5124 __init void
5125 i915_gem_shrinker_init(void)
5126 {
5127     register_shrinker(&shrinker);
5128 }
5129
5130 __exit void
5131 i915_gem_shrinker_exit(void)
5132 {
5133     unregister_shrinker(&shrinker);
5134 }