]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - drivers/gpu/drm/i915/i915_gem.c
Merge branch 'misc' into release
[karo-tx-linux.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 <linux/swap.h>
33
34 static int
35 i915_gem_object_set_domain(struct drm_gem_object *obj,
36                             uint32_t read_domains,
37                             uint32_t write_domain);
38 static int
39 i915_gem_object_set_domain_range(struct drm_gem_object *obj,
40                                  uint64_t offset,
41                                  uint64_t size,
42                                  uint32_t read_domains,
43                                  uint32_t write_domain);
44 static int
45 i915_gem_set_domain(struct drm_gem_object *obj,
46                     struct drm_file *file_priv,
47                     uint32_t read_domains,
48                     uint32_t write_domain);
49 static int i915_gem_object_get_page_list(struct drm_gem_object *obj);
50 static void i915_gem_object_free_page_list(struct drm_gem_object *obj);
51 static int i915_gem_object_wait_rendering(struct drm_gem_object *obj);
52
53 static void
54 i915_gem_cleanup_ringbuffer(struct drm_device *dev);
55
56 int
57 i915_gem_init_ioctl(struct drm_device *dev, void *data,
58                     struct drm_file *file_priv)
59 {
60         drm_i915_private_t *dev_priv = dev->dev_private;
61         struct drm_i915_gem_init *args = data;
62
63         mutex_lock(&dev->struct_mutex);
64
65         if (args->gtt_start >= args->gtt_end ||
66             (args->gtt_start & (PAGE_SIZE - 1)) != 0 ||
67             (args->gtt_end & (PAGE_SIZE - 1)) != 0) {
68                 mutex_unlock(&dev->struct_mutex);
69                 return -EINVAL;
70         }
71
72         drm_mm_init(&dev_priv->mm.gtt_space, args->gtt_start,
73             args->gtt_end - args->gtt_start);
74
75         dev->gtt_total = (uint32_t) (args->gtt_end - args->gtt_start);
76
77         mutex_unlock(&dev->struct_mutex);
78
79         return 0;
80 }
81
82 int
83 i915_gem_get_aperture_ioctl(struct drm_device *dev, void *data,
84                             struct drm_file *file_priv)
85 {
86         drm_i915_private_t *dev_priv = dev->dev_private;
87         struct drm_i915_gem_get_aperture *args = data;
88         struct drm_i915_gem_object *obj_priv;
89
90         if (!(dev->driver->driver_features & DRIVER_GEM))
91                 return -ENODEV;
92
93         args->aper_size = dev->gtt_total;
94         args->aper_available_size = args->aper_size;
95
96         list_for_each_entry(obj_priv, &dev_priv->mm.active_list, list) {
97                 if (obj_priv->pin_count > 0)
98                         args->aper_available_size -= obj_priv->obj->size;
99         }
100
101         return 0;
102 }
103
104
105 /**
106  * Creates a new mm object and returns a handle to it.
107  */
108 int
109 i915_gem_create_ioctl(struct drm_device *dev, void *data,
110                       struct drm_file *file_priv)
111 {
112         struct drm_i915_gem_create *args = data;
113         struct drm_gem_object *obj;
114         int handle, ret;
115
116         args->size = roundup(args->size, PAGE_SIZE);
117
118         /* Allocate the new object */
119         obj = drm_gem_object_alloc(dev, args->size);
120         if (obj == NULL)
121                 return -ENOMEM;
122
123         ret = drm_gem_handle_create(file_priv, obj, &handle);
124         mutex_lock(&dev->struct_mutex);
125         drm_gem_object_handle_unreference(obj);
126         mutex_unlock(&dev->struct_mutex);
127
128         if (ret)
129                 return ret;
130
131         args->handle = handle;
132
133         return 0;
134 }
135
136 /**
137  * Reads data from the object referenced by handle.
138  *
139  * On error, the contents of *data are undefined.
140  */
141 int
142 i915_gem_pread_ioctl(struct drm_device *dev, void *data,
143                      struct drm_file *file_priv)
144 {
145         struct drm_i915_gem_pread *args = data;
146         struct drm_gem_object *obj;
147         struct drm_i915_gem_object *obj_priv;
148         ssize_t read;
149         loff_t offset;
150         int ret;
151
152         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
153         if (obj == NULL)
154                 return -EBADF;
155         obj_priv = obj->driver_private;
156
157         /* Bounds check source.
158          *
159          * XXX: This could use review for overflow issues...
160          */
161         if (args->offset > obj->size || args->size > obj->size ||
162             args->offset + args->size > obj->size) {
163                 drm_gem_object_unreference(obj);
164                 return -EINVAL;
165         }
166
167         mutex_lock(&dev->struct_mutex);
168
169         ret = i915_gem_object_set_domain_range(obj, args->offset, args->size,
170                                                I915_GEM_DOMAIN_CPU, 0);
171         if (ret != 0) {
172                 drm_gem_object_unreference(obj);
173                 mutex_unlock(&dev->struct_mutex);
174                 return ret;
175         }
176
177         offset = args->offset;
178
179         read = vfs_read(obj->filp, (char __user *)(uintptr_t)args->data_ptr,
180                         args->size, &offset);
181         if (read != args->size) {
182                 drm_gem_object_unreference(obj);
183                 mutex_unlock(&dev->struct_mutex);
184                 if (read < 0)
185                         return read;
186                 else
187                         return -EINVAL;
188         }
189
190         drm_gem_object_unreference(obj);
191         mutex_unlock(&dev->struct_mutex);
192
193         return 0;
194 }
195
196 /* This is the fast write path which cannot handle
197  * page faults in the source data
198  */
199
200 static inline int
201 fast_user_write(struct io_mapping *mapping,
202                 loff_t page_base, int page_offset,
203                 char __user *user_data,
204                 int length)
205 {
206         char *vaddr_atomic;
207         unsigned long unwritten;
208
209         vaddr_atomic = io_mapping_map_atomic_wc(mapping, page_base);
210         unwritten = __copy_from_user_inatomic_nocache(vaddr_atomic + page_offset,
211                                                       user_data, length);
212         io_mapping_unmap_atomic(vaddr_atomic);
213         if (unwritten)
214                 return -EFAULT;
215         return 0;
216 }
217
218 /* Here's the write path which can sleep for
219  * page faults
220  */
221
222 static inline int
223 slow_user_write(struct io_mapping *mapping,
224                 loff_t page_base, int page_offset,
225                 char __user *user_data,
226                 int length)
227 {
228         char __iomem *vaddr;
229         unsigned long unwritten;
230
231         vaddr = io_mapping_map_wc(mapping, page_base);
232         if (vaddr == NULL)
233                 return -EFAULT;
234         unwritten = __copy_from_user(vaddr + page_offset,
235                                      user_data, length);
236         io_mapping_unmap(vaddr);
237         if (unwritten)
238                 return -EFAULT;
239         return 0;
240 }
241
242 static int
243 i915_gem_gtt_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
244                     struct drm_i915_gem_pwrite *args,
245                     struct drm_file *file_priv)
246 {
247         struct drm_i915_gem_object *obj_priv = obj->driver_private;
248         drm_i915_private_t *dev_priv = dev->dev_private;
249         ssize_t remain;
250         loff_t offset, page_base;
251         char __user *user_data;
252         int page_offset, page_length;
253         int ret;
254
255         user_data = (char __user *) (uintptr_t) args->data_ptr;
256         remain = args->size;
257         if (!access_ok(VERIFY_READ, user_data, remain))
258                 return -EFAULT;
259
260
261         mutex_lock(&dev->struct_mutex);
262         ret = i915_gem_object_pin(obj, 0);
263         if (ret) {
264                 mutex_unlock(&dev->struct_mutex);
265                 return ret;
266         }
267         ret = i915_gem_set_domain(obj, file_priv,
268                                   I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
269         if (ret)
270                 goto fail;
271
272         obj_priv = obj->driver_private;
273         offset = obj_priv->gtt_offset + args->offset;
274         obj_priv->dirty = 1;
275
276         while (remain > 0) {
277                 /* Operation in this page
278                  *
279                  * page_base = page offset within aperture
280                  * page_offset = offset within page
281                  * page_length = bytes to copy for this page
282                  */
283                 page_base = (offset & ~(PAGE_SIZE-1));
284                 page_offset = offset & (PAGE_SIZE-1);
285                 page_length = remain;
286                 if ((page_offset + remain) > PAGE_SIZE)
287                         page_length = PAGE_SIZE - page_offset;
288
289                 ret = fast_user_write (dev_priv->mm.gtt_mapping, page_base,
290                                        page_offset, user_data, page_length);
291
292                 /* If we get a fault while copying data, then (presumably) our
293                  * source page isn't available. In this case, use the
294                  * non-atomic function
295                  */
296                 if (ret) {
297                         ret = slow_user_write (dev_priv->mm.gtt_mapping,
298                                                page_base, page_offset,
299                                                user_data, page_length);
300                         if (ret)
301                                 goto fail;
302                 }
303
304                 remain -= page_length;
305                 user_data += page_length;
306                 offset += page_length;
307         }
308
309 fail:
310         i915_gem_object_unpin(obj);
311         mutex_unlock(&dev->struct_mutex);
312
313         return ret;
314 }
315
316 static int
317 i915_gem_shmem_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
318                       struct drm_i915_gem_pwrite *args,
319                       struct drm_file *file_priv)
320 {
321         int ret;
322         loff_t offset;
323         ssize_t written;
324
325         mutex_lock(&dev->struct_mutex);
326
327         ret = i915_gem_set_domain(obj, file_priv,
328                                   I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
329         if (ret) {
330                 mutex_unlock(&dev->struct_mutex);
331                 return ret;
332         }
333
334         offset = args->offset;
335
336         written = vfs_write(obj->filp,
337                             (char __user *)(uintptr_t) args->data_ptr,
338                             args->size, &offset);
339         if (written != args->size) {
340                 mutex_unlock(&dev->struct_mutex);
341                 if (written < 0)
342                         return written;
343                 else
344                         return -EINVAL;
345         }
346
347         mutex_unlock(&dev->struct_mutex);
348
349         return 0;
350 }
351
352 /**
353  * Writes data to the object referenced by handle.
354  *
355  * On error, the contents of the buffer that were to be modified are undefined.
356  */
357 int
358 i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
359                       struct drm_file *file_priv)
360 {
361         struct drm_i915_gem_pwrite *args = data;
362         struct drm_gem_object *obj;
363         struct drm_i915_gem_object *obj_priv;
364         int ret = 0;
365
366         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
367         if (obj == NULL)
368                 return -EBADF;
369         obj_priv = obj->driver_private;
370
371         /* Bounds check destination.
372          *
373          * XXX: This could use review for overflow issues...
374          */
375         if (args->offset > obj->size || args->size > obj->size ||
376             args->offset + args->size > obj->size) {
377                 drm_gem_object_unreference(obj);
378                 return -EINVAL;
379         }
380
381         /* We can only do the GTT pwrite on untiled buffers, as otherwise
382          * it would end up going through the fenced access, and we'll get
383          * different detiling behavior between reading and writing.
384          * pread/pwrite currently are reading and writing from the CPU
385          * perspective, requiring manual detiling by the client.
386          */
387         if (obj_priv->tiling_mode == I915_TILING_NONE &&
388             dev->gtt_total != 0)
389                 ret = i915_gem_gtt_pwrite(dev, obj, args, file_priv);
390         else
391                 ret = i915_gem_shmem_pwrite(dev, obj, args, file_priv);
392
393 #if WATCH_PWRITE
394         if (ret)
395                 DRM_INFO("pwrite failed %d\n", ret);
396 #endif
397
398         drm_gem_object_unreference(obj);
399
400         return ret;
401 }
402
403 /**
404  * Called when user space prepares to use an object
405  */
406 int
407 i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
408                           struct drm_file *file_priv)
409 {
410         struct drm_i915_gem_set_domain *args = data;
411         struct drm_gem_object *obj;
412         int ret;
413
414         if (!(dev->driver->driver_features & DRIVER_GEM))
415                 return -ENODEV;
416
417         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
418         if (obj == NULL)
419                 return -EBADF;
420
421         mutex_lock(&dev->struct_mutex);
422 #if WATCH_BUF
423         DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n",
424                  obj, obj->size, args->read_domains, args->write_domain);
425 #endif
426         ret = i915_gem_set_domain(obj, file_priv,
427                                   args->read_domains, args->write_domain);
428         drm_gem_object_unreference(obj);
429         mutex_unlock(&dev->struct_mutex);
430         return ret;
431 }
432
433 /**
434  * Called when user space has done writes to this buffer
435  */
436 int
437 i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data,
438                       struct drm_file *file_priv)
439 {
440         struct drm_i915_gem_sw_finish *args = data;
441         struct drm_gem_object *obj;
442         struct drm_i915_gem_object *obj_priv;
443         int ret = 0;
444
445         if (!(dev->driver->driver_features & DRIVER_GEM))
446                 return -ENODEV;
447
448         mutex_lock(&dev->struct_mutex);
449         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
450         if (obj == NULL) {
451                 mutex_unlock(&dev->struct_mutex);
452                 return -EBADF;
453         }
454
455 #if WATCH_BUF
456         DRM_INFO("%s: sw_finish %d (%p %d)\n",
457                  __func__, args->handle, obj, obj->size);
458 #endif
459         obj_priv = obj->driver_private;
460
461         /* Pinned buffers may be scanout, so flush the cache */
462         if ((obj->write_domain & I915_GEM_DOMAIN_CPU) && obj_priv->pin_count) {
463                 i915_gem_clflush_object(obj);
464                 drm_agp_chipset_flush(dev);
465         }
466         drm_gem_object_unreference(obj);
467         mutex_unlock(&dev->struct_mutex);
468         return ret;
469 }
470
471 /**
472  * Maps the contents of an object, returning the address it is mapped
473  * into.
474  *
475  * While the mapping holds a reference on the contents of the object, it doesn't
476  * imply a ref on the object itself.
477  */
478 int
479 i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
480                    struct drm_file *file_priv)
481 {
482         struct drm_i915_gem_mmap *args = data;
483         struct drm_gem_object *obj;
484         loff_t offset;
485         unsigned long addr;
486
487         if (!(dev->driver->driver_features & DRIVER_GEM))
488                 return -ENODEV;
489
490         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
491         if (obj == NULL)
492                 return -EBADF;
493
494         offset = args->offset;
495
496         down_write(&current->mm->mmap_sem);
497         addr = do_mmap(obj->filp, 0, args->size,
498                        PROT_READ | PROT_WRITE, MAP_SHARED,
499                        args->offset);
500         up_write(&current->mm->mmap_sem);
501         mutex_lock(&dev->struct_mutex);
502         drm_gem_object_unreference(obj);
503         mutex_unlock(&dev->struct_mutex);
504         if (IS_ERR((void *)addr))
505                 return addr;
506
507         args->addr_ptr = (uint64_t) addr;
508
509         return 0;
510 }
511
512 static void
513 i915_gem_object_free_page_list(struct drm_gem_object *obj)
514 {
515         struct drm_i915_gem_object *obj_priv = obj->driver_private;
516         int page_count = obj->size / PAGE_SIZE;
517         int i;
518
519         if (obj_priv->page_list == NULL)
520                 return;
521
522
523         for (i = 0; i < page_count; i++)
524                 if (obj_priv->page_list[i] != NULL) {
525                         if (obj_priv->dirty)
526                                 set_page_dirty(obj_priv->page_list[i]);
527                         mark_page_accessed(obj_priv->page_list[i]);
528                         page_cache_release(obj_priv->page_list[i]);
529                 }
530         obj_priv->dirty = 0;
531
532         drm_free(obj_priv->page_list,
533                  page_count * sizeof(struct page *),
534                  DRM_MEM_DRIVER);
535         obj_priv->page_list = NULL;
536 }
537
538 static void
539 i915_gem_object_move_to_active(struct drm_gem_object *obj)
540 {
541         struct drm_device *dev = obj->dev;
542         drm_i915_private_t *dev_priv = dev->dev_private;
543         struct drm_i915_gem_object *obj_priv = obj->driver_private;
544
545         /* Add a reference if we're newly entering the active list. */
546         if (!obj_priv->active) {
547                 drm_gem_object_reference(obj);
548                 obj_priv->active = 1;
549         }
550         /* Move from whatever list we were on to the tail of execution. */
551         list_move_tail(&obj_priv->list,
552                        &dev_priv->mm.active_list);
553 }
554
555
556 static void
557 i915_gem_object_move_to_inactive(struct drm_gem_object *obj)
558 {
559         struct drm_device *dev = obj->dev;
560         drm_i915_private_t *dev_priv = dev->dev_private;
561         struct drm_i915_gem_object *obj_priv = obj->driver_private;
562
563         i915_verify_inactive(dev, __FILE__, __LINE__);
564         if (obj_priv->pin_count != 0)
565                 list_del_init(&obj_priv->list);
566         else
567                 list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list);
568
569         if (obj_priv->active) {
570                 obj_priv->active = 0;
571                 drm_gem_object_unreference(obj);
572         }
573         i915_verify_inactive(dev, __FILE__, __LINE__);
574 }
575
576 /**
577  * Creates a new sequence number, emitting a write of it to the status page
578  * plus an interrupt, which will trigger i915_user_interrupt_handler.
579  *
580  * Must be called with struct_lock held.
581  *
582  * Returned sequence numbers are nonzero on success.
583  */
584 static uint32_t
585 i915_add_request(struct drm_device *dev, uint32_t flush_domains)
586 {
587         drm_i915_private_t *dev_priv = dev->dev_private;
588         struct drm_i915_gem_request *request;
589         uint32_t seqno;
590         int was_empty;
591         RING_LOCALS;
592
593         request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER);
594         if (request == NULL)
595                 return 0;
596
597         /* Grab the seqno we're going to make this request be, and bump the
598          * next (skipping 0 so it can be the reserved no-seqno value).
599          */
600         seqno = dev_priv->mm.next_gem_seqno;
601         dev_priv->mm.next_gem_seqno++;
602         if (dev_priv->mm.next_gem_seqno == 0)
603                 dev_priv->mm.next_gem_seqno++;
604
605         BEGIN_LP_RING(4);
606         OUT_RING(MI_STORE_DWORD_INDEX);
607         OUT_RING(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
608         OUT_RING(seqno);
609
610         OUT_RING(MI_USER_INTERRUPT);
611         ADVANCE_LP_RING();
612
613         DRM_DEBUG("%d\n", seqno);
614
615         request->seqno = seqno;
616         request->emitted_jiffies = jiffies;
617         request->flush_domains = flush_domains;
618         was_empty = list_empty(&dev_priv->mm.request_list);
619         list_add_tail(&request->list, &dev_priv->mm.request_list);
620
621         if (was_empty && !dev_priv->mm.suspended)
622                 schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
623         return seqno;
624 }
625
626 /**
627  * Command execution barrier
628  *
629  * Ensures that all commands in the ring are finished
630  * before signalling the CPU
631  */
632 static uint32_t
633 i915_retire_commands(struct drm_device *dev)
634 {
635         drm_i915_private_t *dev_priv = dev->dev_private;
636         uint32_t cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
637         uint32_t flush_domains = 0;
638         RING_LOCALS;
639
640         /* The sampler always gets flushed on i965 (sigh) */
641         if (IS_I965G(dev))
642                 flush_domains |= I915_GEM_DOMAIN_SAMPLER;
643         BEGIN_LP_RING(2);
644         OUT_RING(cmd);
645         OUT_RING(0); /* noop */
646         ADVANCE_LP_RING();
647         return flush_domains;
648 }
649
650 /**
651  * Moves buffers associated only with the given active seqno from the active
652  * to inactive list, potentially freeing them.
653  */
654 static void
655 i915_gem_retire_request(struct drm_device *dev,
656                         struct drm_i915_gem_request *request)
657 {
658         drm_i915_private_t *dev_priv = dev->dev_private;
659
660         /* Move any buffers on the active list that are no longer referenced
661          * by the ringbuffer to the flushing/inactive lists as appropriate.
662          */
663         while (!list_empty(&dev_priv->mm.active_list)) {
664                 struct drm_gem_object *obj;
665                 struct drm_i915_gem_object *obj_priv;
666
667                 obj_priv = list_first_entry(&dev_priv->mm.active_list,
668                                             struct drm_i915_gem_object,
669                                             list);
670                 obj = obj_priv->obj;
671
672                 /* If the seqno being retired doesn't match the oldest in the
673                  * list, then the oldest in the list must still be newer than
674                  * this seqno.
675                  */
676                 if (obj_priv->last_rendering_seqno != request->seqno)
677                         return;
678 #if WATCH_LRU
679                 DRM_INFO("%s: retire %d moves to inactive list %p\n",
680                          __func__, request->seqno, obj);
681 #endif
682
683                 if (obj->write_domain != 0) {
684                         list_move_tail(&obj_priv->list,
685                                        &dev_priv->mm.flushing_list);
686                 } else {
687                         i915_gem_object_move_to_inactive(obj);
688                 }
689         }
690
691         if (request->flush_domains != 0) {
692                 struct drm_i915_gem_object *obj_priv, *next;
693
694                 /* Clear the write domain and activity from any buffers
695                  * that are just waiting for a flush matching the one retired.
696                  */
697                 list_for_each_entry_safe(obj_priv, next,
698                                          &dev_priv->mm.flushing_list, list) {
699                         struct drm_gem_object *obj = obj_priv->obj;
700
701                         if (obj->write_domain & request->flush_domains) {
702                                 obj->write_domain = 0;
703                                 i915_gem_object_move_to_inactive(obj);
704                         }
705                 }
706
707         }
708 }
709
710 /**
711  * Returns true if seq1 is later than seq2.
712  */
713 static int
714 i915_seqno_passed(uint32_t seq1, uint32_t seq2)
715 {
716         return (int32_t)(seq1 - seq2) >= 0;
717 }
718
719 uint32_t
720 i915_get_gem_seqno(struct drm_device *dev)
721 {
722         drm_i915_private_t *dev_priv = dev->dev_private;
723
724         return READ_HWSP(dev_priv, I915_GEM_HWS_INDEX);
725 }
726
727 /**
728  * This function clears the request list as sequence numbers are passed.
729  */
730 void
731 i915_gem_retire_requests(struct drm_device *dev)
732 {
733         drm_i915_private_t *dev_priv = dev->dev_private;
734         uint32_t seqno;
735
736         seqno = i915_get_gem_seqno(dev);
737
738         while (!list_empty(&dev_priv->mm.request_list)) {
739                 struct drm_i915_gem_request *request;
740                 uint32_t retiring_seqno;
741
742                 request = list_first_entry(&dev_priv->mm.request_list,
743                                            struct drm_i915_gem_request,
744                                            list);
745                 retiring_seqno = request->seqno;
746
747                 if (i915_seqno_passed(seqno, retiring_seqno) ||
748                     dev_priv->mm.wedged) {
749                         i915_gem_retire_request(dev, request);
750
751                         list_del(&request->list);
752                         drm_free(request, sizeof(*request), DRM_MEM_DRIVER);
753                 } else
754                         break;
755         }
756 }
757
758 void
759 i915_gem_retire_work_handler(struct work_struct *work)
760 {
761         drm_i915_private_t *dev_priv;
762         struct drm_device *dev;
763
764         dev_priv = container_of(work, drm_i915_private_t,
765                                 mm.retire_work.work);
766         dev = dev_priv->dev;
767
768         mutex_lock(&dev->struct_mutex);
769         i915_gem_retire_requests(dev);
770         if (!dev_priv->mm.suspended &&
771             !list_empty(&dev_priv->mm.request_list))
772                 schedule_delayed_work(&dev_priv->mm.retire_work, HZ);
773         mutex_unlock(&dev->struct_mutex);
774 }
775
776 /**
777  * Waits for a sequence number to be signaled, and cleans up the
778  * request and object lists appropriately for that event.
779  */
780 static int
781 i915_wait_request(struct drm_device *dev, uint32_t seqno)
782 {
783         drm_i915_private_t *dev_priv = dev->dev_private;
784         int ret = 0;
785
786         BUG_ON(seqno == 0);
787
788         if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
789                 dev_priv->mm.waiting_gem_seqno = seqno;
790                 i915_user_irq_get(dev);
791                 ret = wait_event_interruptible(dev_priv->irq_queue,
792                                                i915_seqno_passed(i915_get_gem_seqno(dev),
793                                                                  seqno) ||
794                                                dev_priv->mm.wedged);
795                 i915_user_irq_put(dev);
796                 dev_priv->mm.waiting_gem_seqno = 0;
797         }
798         if (dev_priv->mm.wedged)
799                 ret = -EIO;
800
801         if (ret && ret != -ERESTARTSYS)
802                 DRM_ERROR("%s returns %d (awaiting %d at %d)\n",
803                           __func__, ret, seqno, i915_get_gem_seqno(dev));
804
805         /* Directly dispatch request retiring.  While we have the work queue
806          * to handle this, the waiter on a request often wants an associated
807          * buffer to have made it to the inactive list, and we would need
808          * a separate wait queue to handle that.
809          */
810         if (ret == 0)
811                 i915_gem_retire_requests(dev);
812
813         return ret;
814 }
815
816 static void
817 i915_gem_flush(struct drm_device *dev,
818                uint32_t invalidate_domains,
819                uint32_t flush_domains)
820 {
821         drm_i915_private_t *dev_priv = dev->dev_private;
822         uint32_t cmd;
823         RING_LOCALS;
824
825 #if WATCH_EXEC
826         DRM_INFO("%s: invalidate %08x flush %08x\n", __func__,
827                   invalidate_domains, flush_domains);
828 #endif
829
830         if (flush_domains & I915_GEM_DOMAIN_CPU)
831                 drm_agp_chipset_flush(dev);
832
833         if ((invalidate_domains | flush_domains) & ~(I915_GEM_DOMAIN_CPU |
834                                                      I915_GEM_DOMAIN_GTT)) {
835                 /*
836                  * read/write caches:
837                  *
838                  * I915_GEM_DOMAIN_RENDER is always invalidated, but is
839                  * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
840                  * also flushed at 2d versus 3d pipeline switches.
841                  *
842                  * read-only caches:
843                  *
844                  * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
845                  * MI_READ_FLUSH is set, and is always flushed on 965.
846                  *
847                  * I915_GEM_DOMAIN_COMMAND may not exist?
848                  *
849                  * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
850                  * invalidated when MI_EXE_FLUSH is set.
851                  *
852                  * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
853                  * invalidated with every MI_FLUSH.
854                  *
855                  * TLBs:
856                  *
857                  * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
858                  * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
859                  * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
860                  * are flushed at any MI_FLUSH.
861                  */
862
863                 cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
864                 if ((invalidate_domains|flush_domains) &
865                     I915_GEM_DOMAIN_RENDER)
866                         cmd &= ~MI_NO_WRITE_FLUSH;
867                 if (!IS_I965G(dev)) {
868                         /*
869                          * On the 965, the sampler cache always gets flushed
870                          * and this bit is reserved.
871                          */
872                         if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
873                                 cmd |= MI_READ_FLUSH;
874                 }
875                 if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
876                         cmd |= MI_EXE_FLUSH;
877
878 #if WATCH_EXEC
879                 DRM_INFO("%s: queue flush %08x to ring\n", __func__, cmd);
880 #endif
881                 BEGIN_LP_RING(2);
882                 OUT_RING(cmd);
883                 OUT_RING(0); /* noop */
884                 ADVANCE_LP_RING();
885         }
886 }
887
888 /**
889  * Ensures that all rendering to the object has completed and the object is
890  * safe to unbind from the GTT or access from the CPU.
891  */
892 static int
893 i915_gem_object_wait_rendering(struct drm_gem_object *obj)
894 {
895         struct drm_device *dev = obj->dev;
896         struct drm_i915_gem_object *obj_priv = obj->driver_private;
897         int ret;
898
899         /* If there are writes queued to the buffer, flush and
900          * create a new seqno to wait for.
901          */
902         if (obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)) {
903                 uint32_t write_domain = obj->write_domain;
904 #if WATCH_BUF
905                 DRM_INFO("%s: flushing object %p from write domain %08x\n",
906                           __func__, obj, write_domain);
907 #endif
908                 i915_gem_flush(dev, 0, write_domain);
909
910                 i915_gem_object_move_to_active(obj);
911                 obj_priv->last_rendering_seqno = i915_add_request(dev,
912                                                                   write_domain);
913                 BUG_ON(obj_priv->last_rendering_seqno == 0);
914 #if WATCH_LRU
915                 DRM_INFO("%s: flush moves to exec list %p\n", __func__, obj);
916 #endif
917         }
918
919         /* If there is rendering queued on the buffer being evicted, wait for
920          * it.
921          */
922         if (obj_priv->active) {
923 #if WATCH_BUF
924                 DRM_INFO("%s: object %p wait for seqno %08x\n",
925                           __func__, obj, obj_priv->last_rendering_seqno);
926 #endif
927                 ret = i915_wait_request(dev, obj_priv->last_rendering_seqno);
928                 if (ret != 0)
929                         return ret;
930         }
931
932         return 0;
933 }
934
935 /**
936  * Unbinds an object from the GTT aperture.
937  */
938 static int
939 i915_gem_object_unbind(struct drm_gem_object *obj)
940 {
941         struct drm_device *dev = obj->dev;
942         struct drm_i915_gem_object *obj_priv = obj->driver_private;
943         int ret = 0;
944
945 #if WATCH_BUF
946         DRM_INFO("%s:%d %p\n", __func__, __LINE__, obj);
947         DRM_INFO("gtt_space %p\n", obj_priv->gtt_space);
948 #endif
949         if (obj_priv->gtt_space == NULL)
950                 return 0;
951
952         if (obj_priv->pin_count != 0) {
953                 DRM_ERROR("Attempting to unbind pinned buffer\n");
954                 return -EINVAL;
955         }
956
957         /* Wait for any rendering to complete
958          */
959         ret = i915_gem_object_wait_rendering(obj);
960         if (ret) {
961                 DRM_ERROR("wait_rendering failed: %d\n", ret);
962                 return ret;
963         }
964
965         /* Move the object to the CPU domain to ensure that
966          * any possible CPU writes while it's not in the GTT
967          * are flushed when we go to remap it. This will
968          * also ensure that all pending GPU writes are finished
969          * before we unbind.
970          */
971         ret = i915_gem_object_set_domain(obj, I915_GEM_DOMAIN_CPU,
972                                          I915_GEM_DOMAIN_CPU);
973         if (ret) {
974                 DRM_ERROR("set_domain failed: %d\n", ret);
975                 return ret;
976         }
977
978         if (obj_priv->agp_mem != NULL) {
979                 drm_unbind_agp(obj_priv->agp_mem);
980                 drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE);
981                 obj_priv->agp_mem = NULL;
982         }
983
984         BUG_ON(obj_priv->active);
985
986         i915_gem_object_free_page_list(obj);
987
988         if (obj_priv->gtt_space) {
989                 atomic_dec(&dev->gtt_count);
990                 atomic_sub(obj->size, &dev->gtt_memory);
991
992                 drm_mm_put_block(obj_priv->gtt_space);
993                 obj_priv->gtt_space = NULL;
994         }
995
996         /* Remove ourselves from the LRU list if present. */
997         if (!list_empty(&obj_priv->list))
998                 list_del_init(&obj_priv->list);
999
1000         return 0;
1001 }
1002
1003 static int
1004 i915_gem_evict_something(struct drm_device *dev)
1005 {
1006         drm_i915_private_t *dev_priv = dev->dev_private;
1007         struct drm_gem_object *obj;
1008         struct drm_i915_gem_object *obj_priv;
1009         int ret = 0;
1010
1011         for (;;) {
1012                 /* If there's an inactive buffer available now, grab it
1013                  * and be done.
1014                  */
1015                 if (!list_empty(&dev_priv->mm.inactive_list)) {
1016                         obj_priv = list_first_entry(&dev_priv->mm.inactive_list,
1017                                                     struct drm_i915_gem_object,
1018                                                     list);
1019                         obj = obj_priv->obj;
1020                         BUG_ON(obj_priv->pin_count != 0);
1021 #if WATCH_LRU
1022                         DRM_INFO("%s: evicting %p\n", __func__, obj);
1023 #endif
1024                         BUG_ON(obj_priv->active);
1025
1026                         /* Wait on the rendering and unbind the buffer. */
1027                         ret = i915_gem_object_unbind(obj);
1028                         break;
1029                 }
1030
1031                 /* If we didn't get anything, but the ring is still processing
1032                  * things, wait for one of those things to finish and hopefully
1033                  * leave us a buffer to evict.
1034                  */
1035                 if (!list_empty(&dev_priv->mm.request_list)) {
1036                         struct drm_i915_gem_request *request;
1037
1038                         request = list_first_entry(&dev_priv->mm.request_list,
1039                                                    struct drm_i915_gem_request,
1040                                                    list);
1041
1042                         ret = i915_wait_request(dev, request->seqno);
1043                         if (ret)
1044                                 break;
1045
1046                         /* if waiting caused an object to become inactive,
1047                          * then loop around and wait for it. Otherwise, we
1048                          * assume that waiting freed and unbound something,
1049                          * so there should now be some space in the GTT
1050                          */
1051                         if (!list_empty(&dev_priv->mm.inactive_list))
1052                                 continue;
1053                         break;
1054                 }
1055
1056                 /* If we didn't have anything on the request list but there
1057                  * are buffers awaiting a flush, emit one and try again.
1058                  * When we wait on it, those buffers waiting for that flush
1059                  * will get moved to inactive.
1060                  */
1061                 if (!list_empty(&dev_priv->mm.flushing_list)) {
1062                         obj_priv = list_first_entry(&dev_priv->mm.flushing_list,
1063                                                     struct drm_i915_gem_object,
1064                                                     list);
1065                         obj = obj_priv->obj;
1066
1067                         i915_gem_flush(dev,
1068                                        obj->write_domain,
1069                                        obj->write_domain);
1070                         i915_add_request(dev, obj->write_domain);
1071
1072                         obj = NULL;
1073                         continue;
1074                 }
1075
1076                 DRM_ERROR("inactive empty %d request empty %d "
1077                           "flushing empty %d\n",
1078                           list_empty(&dev_priv->mm.inactive_list),
1079                           list_empty(&dev_priv->mm.request_list),
1080                           list_empty(&dev_priv->mm.flushing_list));
1081                 /* If we didn't do any of the above, there's nothing to be done
1082                  * and we just can't fit it in.
1083                  */
1084                 return -ENOMEM;
1085         }
1086         return ret;
1087 }
1088
1089 static int
1090 i915_gem_object_get_page_list(struct drm_gem_object *obj)
1091 {
1092         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1093         int page_count, i;
1094         struct address_space *mapping;
1095         struct inode *inode;
1096         struct page *page;
1097         int ret;
1098
1099         if (obj_priv->page_list)
1100                 return 0;
1101
1102         /* Get the list of pages out of our struct file.  They'll be pinned
1103          * at this point until we release them.
1104          */
1105         page_count = obj->size / PAGE_SIZE;
1106         BUG_ON(obj_priv->page_list != NULL);
1107         obj_priv->page_list = drm_calloc(page_count, sizeof(struct page *),
1108                                          DRM_MEM_DRIVER);
1109         if (obj_priv->page_list == NULL) {
1110                 DRM_ERROR("Faled to allocate page list\n");
1111                 return -ENOMEM;
1112         }
1113
1114         inode = obj->filp->f_path.dentry->d_inode;
1115         mapping = inode->i_mapping;
1116         for (i = 0; i < page_count; i++) {
1117                 page = read_mapping_page(mapping, i, NULL);
1118                 if (IS_ERR(page)) {
1119                         ret = PTR_ERR(page);
1120                         DRM_ERROR("read_mapping_page failed: %d\n", ret);
1121                         i915_gem_object_free_page_list(obj);
1122                         return ret;
1123                 }
1124                 obj_priv->page_list[i] = page;
1125         }
1126         return 0;
1127 }
1128
1129 /**
1130  * Finds free space in the GTT aperture and binds the object there.
1131  */
1132 static int
1133 i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment)
1134 {
1135         struct drm_device *dev = obj->dev;
1136         drm_i915_private_t *dev_priv = dev->dev_private;
1137         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1138         struct drm_mm_node *free_space;
1139         int page_count, ret;
1140
1141         if (alignment == 0)
1142                 alignment = PAGE_SIZE;
1143         if (alignment & (PAGE_SIZE - 1)) {
1144                 DRM_ERROR("Invalid object alignment requested %u\n", alignment);
1145                 return -EINVAL;
1146         }
1147
1148  search_free:
1149         free_space = drm_mm_search_free(&dev_priv->mm.gtt_space,
1150                                         obj->size, alignment, 0);
1151         if (free_space != NULL) {
1152                 obj_priv->gtt_space = drm_mm_get_block(free_space, obj->size,
1153                                                        alignment);
1154                 if (obj_priv->gtt_space != NULL) {
1155                         obj_priv->gtt_space->private = obj;
1156                         obj_priv->gtt_offset = obj_priv->gtt_space->start;
1157                 }
1158         }
1159         if (obj_priv->gtt_space == NULL) {
1160                 /* If the gtt is empty and we're still having trouble
1161                  * fitting our object in, we're out of memory.
1162                  */
1163 #if WATCH_LRU
1164                 DRM_INFO("%s: GTT full, evicting something\n", __func__);
1165 #endif
1166                 if (list_empty(&dev_priv->mm.inactive_list) &&
1167                     list_empty(&dev_priv->mm.flushing_list) &&
1168                     list_empty(&dev_priv->mm.active_list)) {
1169                         DRM_ERROR("GTT full, but LRU list empty\n");
1170                         return -ENOMEM;
1171                 }
1172
1173                 ret = i915_gem_evict_something(dev);
1174                 if (ret != 0) {
1175                         DRM_ERROR("Failed to evict a buffer %d\n", ret);
1176                         return ret;
1177                 }
1178                 goto search_free;
1179         }
1180
1181 #if WATCH_BUF
1182         DRM_INFO("Binding object of size %d at 0x%08x\n",
1183                  obj->size, obj_priv->gtt_offset);
1184 #endif
1185         ret = i915_gem_object_get_page_list(obj);
1186         if (ret) {
1187                 drm_mm_put_block(obj_priv->gtt_space);
1188                 obj_priv->gtt_space = NULL;
1189                 return ret;
1190         }
1191
1192         page_count = obj->size / PAGE_SIZE;
1193         /* Create an AGP memory structure pointing at our pages, and bind it
1194          * into the GTT.
1195          */
1196         obj_priv->agp_mem = drm_agp_bind_pages(dev,
1197                                                obj_priv->page_list,
1198                                                page_count,
1199                                                obj_priv->gtt_offset,
1200                                                obj_priv->agp_type);
1201         if (obj_priv->agp_mem == NULL) {
1202                 i915_gem_object_free_page_list(obj);
1203                 drm_mm_put_block(obj_priv->gtt_space);
1204                 obj_priv->gtt_space = NULL;
1205                 return -ENOMEM;
1206         }
1207         atomic_inc(&dev->gtt_count);
1208         atomic_add(obj->size, &dev->gtt_memory);
1209
1210         /* Assert that the object is not currently in any GPU domain. As it
1211          * wasn't in the GTT, there shouldn't be any way it could have been in
1212          * a GPU cache
1213          */
1214         BUG_ON(obj->read_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
1215         BUG_ON(obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
1216
1217         return 0;
1218 }
1219
1220 void
1221 i915_gem_clflush_object(struct drm_gem_object *obj)
1222 {
1223         struct drm_i915_gem_object      *obj_priv = obj->driver_private;
1224
1225         /* If we don't have a page list set up, then we're not pinned
1226          * to GPU, and we can ignore the cache flush because it'll happen
1227          * again at bind time.
1228          */
1229         if (obj_priv->page_list == NULL)
1230                 return;
1231
1232         drm_clflush_pages(obj_priv->page_list, obj->size / PAGE_SIZE);
1233 }
1234
1235 /*
1236  * Set the next domain for the specified object. This
1237  * may not actually perform the necessary flushing/invaliding though,
1238  * as that may want to be batched with other set_domain operations
1239  *
1240  * This is (we hope) the only really tricky part of gem. The goal
1241  * is fairly simple -- track which caches hold bits of the object
1242  * and make sure they remain coherent. A few concrete examples may
1243  * help to explain how it works. For shorthand, we use the notation
1244  * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
1245  * a pair of read and write domain masks.
1246  *
1247  * Case 1: the batch buffer
1248  *
1249  *      1. Allocated
1250  *      2. Written by CPU
1251  *      3. Mapped to GTT
1252  *      4. Read by GPU
1253  *      5. Unmapped from GTT
1254  *      6. Freed
1255  *
1256  *      Let's take these a step at a time
1257  *
1258  *      1. Allocated
1259  *              Pages allocated from the kernel may still have
1260  *              cache contents, so we set them to (CPU, CPU) always.
1261  *      2. Written by CPU (using pwrite)
1262  *              The pwrite function calls set_domain (CPU, CPU) and
1263  *              this function does nothing (as nothing changes)
1264  *      3. Mapped by GTT
1265  *              This function asserts that the object is not
1266  *              currently in any GPU-based read or write domains
1267  *      4. Read by GPU
1268  *              i915_gem_execbuffer calls set_domain (COMMAND, 0).
1269  *              As write_domain is zero, this function adds in the
1270  *              current read domains (CPU+COMMAND, 0).
1271  *              flush_domains is set to CPU.
1272  *              invalidate_domains is set to COMMAND
1273  *              clflush is run to get data out of the CPU caches
1274  *              then i915_dev_set_domain calls i915_gem_flush to
1275  *              emit an MI_FLUSH and drm_agp_chipset_flush
1276  *      5. Unmapped from GTT
1277  *              i915_gem_object_unbind calls set_domain (CPU, CPU)
1278  *              flush_domains and invalidate_domains end up both zero
1279  *              so no flushing/invalidating happens
1280  *      6. Freed
1281  *              yay, done
1282  *
1283  * Case 2: The shared render buffer
1284  *
1285  *      1. Allocated
1286  *      2. Mapped to GTT
1287  *      3. Read/written by GPU
1288  *      4. set_domain to (CPU,CPU)
1289  *      5. Read/written by CPU
1290  *      6. Read/written by GPU
1291  *
1292  *      1. Allocated
1293  *              Same as last example, (CPU, CPU)
1294  *      2. Mapped to GTT
1295  *              Nothing changes (assertions find that it is not in the GPU)
1296  *      3. Read/written by GPU
1297  *              execbuffer calls set_domain (RENDER, RENDER)
1298  *              flush_domains gets CPU
1299  *              invalidate_domains gets GPU
1300  *              clflush (obj)
1301  *              MI_FLUSH and drm_agp_chipset_flush
1302  *      4. set_domain (CPU, CPU)
1303  *              flush_domains gets GPU
1304  *              invalidate_domains gets CPU
1305  *              wait_rendering (obj) to make sure all drawing is complete.
1306  *              This will include an MI_FLUSH to get the data from GPU
1307  *              to memory
1308  *              clflush (obj) to invalidate the CPU cache
1309  *              Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
1310  *      5. Read/written by CPU
1311  *              cache lines are loaded and dirtied
1312  *      6. Read written by GPU
1313  *              Same as last GPU access
1314  *
1315  * Case 3: The constant buffer
1316  *
1317  *      1. Allocated
1318  *      2. Written by CPU
1319  *      3. Read by GPU
1320  *      4. Updated (written) by CPU again
1321  *      5. Read by GPU
1322  *
1323  *      1. Allocated
1324  *              (CPU, CPU)
1325  *      2. Written by CPU
1326  *              (CPU, CPU)
1327  *      3. Read by GPU
1328  *              (CPU+RENDER, 0)
1329  *              flush_domains = CPU
1330  *              invalidate_domains = RENDER
1331  *              clflush (obj)
1332  *              MI_FLUSH
1333  *              drm_agp_chipset_flush
1334  *      4. Updated (written) by CPU again
1335  *              (CPU, CPU)
1336  *              flush_domains = 0 (no previous write domain)
1337  *              invalidate_domains = 0 (no new read domains)
1338  *      5. Read by GPU
1339  *              (CPU+RENDER, 0)
1340  *              flush_domains = CPU
1341  *              invalidate_domains = RENDER
1342  *              clflush (obj)
1343  *              MI_FLUSH
1344  *              drm_agp_chipset_flush
1345  */
1346 static int
1347 i915_gem_object_set_domain(struct drm_gem_object *obj,
1348                             uint32_t read_domains,
1349                             uint32_t write_domain)
1350 {
1351         struct drm_device               *dev = obj->dev;
1352         struct drm_i915_gem_object      *obj_priv = obj->driver_private;
1353         uint32_t                        invalidate_domains = 0;
1354         uint32_t                        flush_domains = 0;
1355         int                             ret;
1356
1357 #if WATCH_BUF
1358         DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n",
1359                  __func__, obj,
1360                  obj->read_domains, read_domains,
1361                  obj->write_domain, write_domain);
1362 #endif
1363         /*
1364          * If the object isn't moving to a new write domain,
1365          * let the object stay in multiple read domains
1366          */
1367         if (write_domain == 0)
1368                 read_domains |= obj->read_domains;
1369         else
1370                 obj_priv->dirty = 1;
1371
1372         /*
1373          * Flush the current write domain if
1374          * the new read domains don't match. Invalidate
1375          * any read domains which differ from the old
1376          * write domain
1377          */
1378         if (obj->write_domain && obj->write_domain != read_domains) {
1379                 flush_domains |= obj->write_domain;
1380                 invalidate_domains |= read_domains & ~obj->write_domain;
1381         }
1382         /*
1383          * Invalidate any read caches which may have
1384          * stale data. That is, any new read domains.
1385          */
1386         invalidate_domains |= read_domains & ~obj->read_domains;
1387         if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU) {
1388 #if WATCH_BUF
1389                 DRM_INFO("%s: CPU domain flush %08x invalidate %08x\n",
1390                          __func__, flush_domains, invalidate_domains);
1391 #endif
1392                 /*
1393                  * If we're invaliding the CPU cache and flushing a GPU cache,
1394                  * then pause for rendering so that the GPU caches will be
1395                  * flushed before the cpu cache is invalidated
1396                  */
1397                 if ((invalidate_domains & I915_GEM_DOMAIN_CPU) &&
1398                     (flush_domains & ~(I915_GEM_DOMAIN_CPU |
1399                                        I915_GEM_DOMAIN_GTT))) {
1400                         ret = i915_gem_object_wait_rendering(obj);
1401                         if (ret)
1402                                 return ret;
1403                 }
1404                 i915_gem_clflush_object(obj);
1405         }
1406
1407         if ((write_domain | flush_domains) != 0)
1408                 obj->write_domain = write_domain;
1409
1410         /* If we're invalidating the CPU domain, clear the per-page CPU
1411          * domain list as well.
1412          */
1413         if (obj_priv->page_cpu_valid != NULL &&
1414             (write_domain != 0 ||
1415              read_domains & I915_GEM_DOMAIN_CPU)) {
1416                 drm_free(obj_priv->page_cpu_valid, obj->size / PAGE_SIZE,
1417                          DRM_MEM_DRIVER);
1418                 obj_priv->page_cpu_valid = NULL;
1419         }
1420         obj->read_domains = read_domains;
1421
1422         dev->invalidate_domains |= invalidate_domains;
1423         dev->flush_domains |= flush_domains;
1424 #if WATCH_BUF
1425         DRM_INFO("%s: read %08x write %08x invalidate %08x flush %08x\n",
1426                  __func__,
1427                  obj->read_domains, obj->write_domain,
1428                  dev->invalidate_domains, dev->flush_domains);
1429 #endif
1430         return 0;
1431 }
1432
1433 /**
1434  * Set the read/write domain on a range of the object.
1435  *
1436  * Currently only implemented for CPU reads, otherwise drops to normal
1437  * i915_gem_object_set_domain().
1438  */
1439 static int
1440 i915_gem_object_set_domain_range(struct drm_gem_object *obj,
1441                                  uint64_t offset,
1442                                  uint64_t size,
1443                                  uint32_t read_domains,
1444                                  uint32_t write_domain)
1445 {
1446         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1447         int ret, i;
1448
1449         if (obj->read_domains & I915_GEM_DOMAIN_CPU)
1450                 return 0;
1451
1452         if (read_domains != I915_GEM_DOMAIN_CPU ||
1453             write_domain != 0)
1454                 return i915_gem_object_set_domain(obj,
1455                                                   read_domains, write_domain);
1456
1457         /* Wait on any GPU rendering to the object to be flushed. */
1458         if (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT)) {
1459                 ret = i915_gem_object_wait_rendering(obj);
1460                 if (ret)
1461                         return ret;
1462         }
1463
1464         if (obj_priv->page_cpu_valid == NULL) {
1465                 obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE,
1466                                                       DRM_MEM_DRIVER);
1467         }
1468
1469         /* Flush the cache on any pages that are still invalid from the CPU's
1470          * perspective.
1471          */
1472         for (i = offset / PAGE_SIZE; i <= (offset + size - 1) / PAGE_SIZE; i++) {
1473                 if (obj_priv->page_cpu_valid[i])
1474                         continue;
1475
1476                 drm_clflush_pages(obj_priv->page_list + i, 1);
1477
1478                 obj_priv->page_cpu_valid[i] = 1;
1479         }
1480
1481         return 0;
1482 }
1483
1484 /**
1485  * Once all of the objects have been set in the proper domain,
1486  * perform the necessary flush and invalidate operations.
1487  *
1488  * Returns the write domains flushed, for use in flush tracking.
1489  */
1490 static uint32_t
1491 i915_gem_dev_set_domain(struct drm_device *dev)
1492 {
1493         uint32_t flush_domains = dev->flush_domains;
1494
1495         /*
1496          * Now that all the buffers are synced to the proper domains,
1497          * flush and invalidate the collected domains
1498          */
1499         if (dev->invalidate_domains | dev->flush_domains) {
1500 #if WATCH_EXEC
1501                 DRM_INFO("%s: invalidate_domains %08x flush_domains %08x\n",
1502                           __func__,
1503                          dev->invalidate_domains,
1504                          dev->flush_domains);
1505 #endif
1506                 i915_gem_flush(dev,
1507                                dev->invalidate_domains,
1508                                dev->flush_domains);
1509                 dev->invalidate_domains = 0;
1510                 dev->flush_domains = 0;
1511         }
1512
1513         return flush_domains;
1514 }
1515
1516 /**
1517  * Pin an object to the GTT and evaluate the relocations landing in it.
1518  */
1519 static int
1520 i915_gem_object_pin_and_relocate(struct drm_gem_object *obj,
1521                                  struct drm_file *file_priv,
1522                                  struct drm_i915_gem_exec_object *entry)
1523 {
1524         struct drm_device *dev = obj->dev;
1525         drm_i915_private_t *dev_priv = dev->dev_private;
1526         struct drm_i915_gem_relocation_entry reloc;
1527         struct drm_i915_gem_relocation_entry __user *relocs;
1528         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1529         int i, ret;
1530         void __iomem *reloc_page;
1531
1532         /* Choose the GTT offset for our buffer and put it there. */
1533         ret = i915_gem_object_pin(obj, (uint32_t) entry->alignment);
1534         if (ret)
1535                 return ret;
1536
1537         entry->offset = obj_priv->gtt_offset;
1538
1539         relocs = (struct drm_i915_gem_relocation_entry __user *)
1540                  (uintptr_t) entry->relocs_ptr;
1541         /* Apply the relocations, using the GTT aperture to avoid cache
1542          * flushing requirements.
1543          */
1544         for (i = 0; i < entry->relocation_count; i++) {
1545                 struct drm_gem_object *target_obj;
1546                 struct drm_i915_gem_object *target_obj_priv;
1547                 uint32_t reloc_val, reloc_offset;
1548                 uint32_t __iomem *reloc_entry;
1549
1550                 ret = copy_from_user(&reloc, relocs + i, sizeof(reloc));
1551                 if (ret != 0) {
1552                         i915_gem_object_unpin(obj);
1553                         return ret;
1554                 }
1555
1556                 target_obj = drm_gem_object_lookup(obj->dev, file_priv,
1557                                                    reloc.target_handle);
1558                 if (target_obj == NULL) {
1559                         i915_gem_object_unpin(obj);
1560                         return -EBADF;
1561                 }
1562                 target_obj_priv = target_obj->driver_private;
1563
1564                 /* The target buffer should have appeared before us in the
1565                  * exec_object list, so it should have a GTT space bound by now.
1566                  */
1567                 if (target_obj_priv->gtt_space == NULL) {
1568                         DRM_ERROR("No GTT space found for object %d\n",
1569                                   reloc.target_handle);
1570                         drm_gem_object_unreference(target_obj);
1571                         i915_gem_object_unpin(obj);
1572                         return -EINVAL;
1573                 }
1574
1575                 if (reloc.offset > obj->size - 4) {
1576                         DRM_ERROR("Relocation beyond object bounds: "
1577                                   "obj %p target %d offset %d size %d.\n",
1578                                   obj, reloc.target_handle,
1579                                   (int) reloc.offset, (int) obj->size);
1580                         drm_gem_object_unreference(target_obj);
1581                         i915_gem_object_unpin(obj);
1582                         return -EINVAL;
1583                 }
1584                 if (reloc.offset & 3) {
1585                         DRM_ERROR("Relocation not 4-byte aligned: "
1586                                   "obj %p target %d offset %d.\n",
1587                                   obj, reloc.target_handle,
1588                                   (int) reloc.offset);
1589                         drm_gem_object_unreference(target_obj);
1590                         i915_gem_object_unpin(obj);
1591                         return -EINVAL;
1592                 }
1593
1594                 if (reloc.write_domain && target_obj->pending_write_domain &&
1595                     reloc.write_domain != target_obj->pending_write_domain) {
1596                         DRM_ERROR("Write domain conflict: "
1597                                   "obj %p target %d offset %d "
1598                                   "new %08x old %08x\n",
1599                                   obj, reloc.target_handle,
1600                                   (int) reloc.offset,
1601                                   reloc.write_domain,
1602                                   target_obj->pending_write_domain);
1603                         drm_gem_object_unreference(target_obj);
1604                         i915_gem_object_unpin(obj);
1605                         return -EINVAL;
1606                 }
1607
1608 #if WATCH_RELOC
1609                 DRM_INFO("%s: obj %p offset %08x target %d "
1610                          "read %08x write %08x gtt %08x "
1611                          "presumed %08x delta %08x\n",
1612                          __func__,
1613                          obj,
1614                          (int) reloc.offset,
1615                          (int) reloc.target_handle,
1616                          (int) reloc.read_domains,
1617                          (int) reloc.write_domain,
1618                          (int) target_obj_priv->gtt_offset,
1619                          (int) reloc.presumed_offset,
1620                          reloc.delta);
1621 #endif
1622
1623                 target_obj->pending_read_domains |= reloc.read_domains;
1624                 target_obj->pending_write_domain |= reloc.write_domain;
1625
1626                 /* If the relocation already has the right value in it, no
1627                  * more work needs to be done.
1628                  */
1629                 if (target_obj_priv->gtt_offset == reloc.presumed_offset) {
1630                         drm_gem_object_unreference(target_obj);
1631                         continue;
1632                 }
1633
1634                 /* Now that we're going to actually write some data in,
1635                  * make sure that any rendering using this buffer's contents
1636                  * is completed.
1637                  */
1638                 i915_gem_object_wait_rendering(obj);
1639
1640                 /* As we're writing through the gtt, flush
1641                  * any CPU writes before we write the relocations
1642                  */
1643                 if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
1644                         i915_gem_clflush_object(obj);
1645                         drm_agp_chipset_flush(dev);
1646                         obj->write_domain = 0;
1647                 }
1648
1649                 /* Map the page containing the relocation we're going to
1650                  * perform.
1651                  */
1652                 reloc_offset = obj_priv->gtt_offset + reloc.offset;
1653                 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
1654                                                       (reloc_offset &
1655                                                        ~(PAGE_SIZE - 1)));
1656                 reloc_entry = (uint32_t __iomem *)(reloc_page +
1657                                                    (reloc_offset & (PAGE_SIZE - 1)));
1658                 reloc_val = target_obj_priv->gtt_offset + reloc.delta;
1659
1660 #if WATCH_BUF
1661                 DRM_INFO("Applied relocation: %p@0x%08x %08x -> %08x\n",
1662                           obj, (unsigned int) reloc.offset,
1663                           readl(reloc_entry), reloc_val);
1664 #endif
1665                 writel(reloc_val, reloc_entry);
1666                 io_mapping_unmap_atomic(reloc_page);
1667
1668                 /* Write the updated presumed offset for this entry back out
1669                  * to the user.
1670                  */
1671                 reloc.presumed_offset = target_obj_priv->gtt_offset;
1672                 ret = copy_to_user(relocs + i, &reloc, sizeof(reloc));
1673                 if (ret != 0) {
1674                         drm_gem_object_unreference(target_obj);
1675                         i915_gem_object_unpin(obj);
1676                         return ret;
1677                 }
1678
1679                 drm_gem_object_unreference(target_obj);
1680         }
1681
1682 #if WATCH_BUF
1683         if (0)
1684                 i915_gem_dump_object(obj, 128, __func__, ~0);
1685 #endif
1686         return 0;
1687 }
1688
1689 /** Dispatch a batchbuffer to the ring
1690  */
1691 static int
1692 i915_dispatch_gem_execbuffer(struct drm_device *dev,
1693                               struct drm_i915_gem_execbuffer *exec,
1694                               uint64_t exec_offset)
1695 {
1696         drm_i915_private_t *dev_priv = dev->dev_private;
1697         struct drm_clip_rect __user *boxes = (struct drm_clip_rect __user *)
1698                                              (uintptr_t) exec->cliprects_ptr;
1699         int nbox = exec->num_cliprects;
1700         int i = 0, count;
1701         uint32_t        exec_start, exec_len;
1702         RING_LOCALS;
1703
1704         exec_start = (uint32_t) exec_offset + exec->batch_start_offset;
1705         exec_len = (uint32_t) exec->batch_len;
1706
1707         if ((exec_start | exec_len) & 0x7) {
1708                 DRM_ERROR("alignment\n");
1709                 return -EINVAL;
1710         }
1711
1712         if (!exec_start)
1713                 return -EINVAL;
1714
1715         count = nbox ? nbox : 1;
1716
1717         for (i = 0; i < count; i++) {
1718                 if (i < nbox) {
1719                         int ret = i915_emit_box(dev, boxes, i,
1720                                                 exec->DR1, exec->DR4);
1721                         if (ret)
1722                                 return ret;
1723                 }
1724
1725                 if (IS_I830(dev) || IS_845G(dev)) {
1726                         BEGIN_LP_RING(4);
1727                         OUT_RING(MI_BATCH_BUFFER);
1728                         OUT_RING(exec_start | MI_BATCH_NON_SECURE);
1729                         OUT_RING(exec_start + exec_len - 4);
1730                         OUT_RING(0);
1731                         ADVANCE_LP_RING();
1732                 } else {
1733                         BEGIN_LP_RING(2);
1734                         if (IS_I965G(dev)) {
1735                                 OUT_RING(MI_BATCH_BUFFER_START |
1736                                          (2 << 6) |
1737                                          MI_BATCH_NON_SECURE_I965);
1738                                 OUT_RING(exec_start);
1739                         } else {
1740                                 OUT_RING(MI_BATCH_BUFFER_START |
1741                                          (2 << 6));
1742                                 OUT_RING(exec_start | MI_BATCH_NON_SECURE);
1743                         }
1744                         ADVANCE_LP_RING();
1745                 }
1746         }
1747
1748         /* XXX breadcrumb */
1749         return 0;
1750 }
1751
1752 /* Throttle our rendering by waiting until the ring has completed our requests
1753  * emitted over 20 msec ago.
1754  *
1755  * This should get us reasonable parallelism between CPU and GPU but also
1756  * relatively low latency when blocking on a particular request to finish.
1757  */
1758 static int
1759 i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv)
1760 {
1761         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
1762         int ret = 0;
1763         uint32_t seqno;
1764
1765         mutex_lock(&dev->struct_mutex);
1766         seqno = i915_file_priv->mm.last_gem_throttle_seqno;
1767         i915_file_priv->mm.last_gem_throttle_seqno =
1768                 i915_file_priv->mm.last_gem_seqno;
1769         if (seqno)
1770                 ret = i915_wait_request(dev, seqno);
1771         mutex_unlock(&dev->struct_mutex);
1772         return ret;
1773 }
1774
1775 int
1776 i915_gem_execbuffer(struct drm_device *dev, void *data,
1777                     struct drm_file *file_priv)
1778 {
1779         drm_i915_private_t *dev_priv = dev->dev_private;
1780         struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv;
1781         struct drm_i915_gem_execbuffer *args = data;
1782         struct drm_i915_gem_exec_object *exec_list = NULL;
1783         struct drm_gem_object **object_list = NULL;
1784         struct drm_gem_object *batch_obj;
1785         int ret, i, pinned = 0;
1786         uint64_t exec_offset;
1787         uint32_t seqno, flush_domains;
1788
1789 #if WATCH_EXEC
1790         DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n",
1791                   (int) args->buffers_ptr, args->buffer_count, args->batch_len);
1792 #endif
1793
1794         if (args->buffer_count < 1) {
1795                 DRM_ERROR("execbuf with %d buffers\n", args->buffer_count);
1796                 return -EINVAL;
1797         }
1798         /* Copy in the exec list from userland */
1799         exec_list = drm_calloc(sizeof(*exec_list), args->buffer_count,
1800                                DRM_MEM_DRIVER);
1801         object_list = drm_calloc(sizeof(*object_list), args->buffer_count,
1802                                  DRM_MEM_DRIVER);
1803         if (exec_list == NULL || object_list == NULL) {
1804                 DRM_ERROR("Failed to allocate exec or object list "
1805                           "for %d buffers\n",
1806                           args->buffer_count);
1807                 ret = -ENOMEM;
1808                 goto pre_mutex_err;
1809         }
1810         ret = copy_from_user(exec_list,
1811                              (struct drm_i915_relocation_entry __user *)
1812                              (uintptr_t) args->buffers_ptr,
1813                              sizeof(*exec_list) * args->buffer_count);
1814         if (ret != 0) {
1815                 DRM_ERROR("copy %d exec entries failed %d\n",
1816                           args->buffer_count, ret);
1817                 goto pre_mutex_err;
1818         }
1819
1820         mutex_lock(&dev->struct_mutex);
1821
1822         i915_verify_inactive(dev, __FILE__, __LINE__);
1823
1824         if (dev_priv->mm.wedged) {
1825                 DRM_ERROR("Execbuf while wedged\n");
1826                 mutex_unlock(&dev->struct_mutex);
1827                 return -EIO;
1828         }
1829
1830         if (dev_priv->mm.suspended) {
1831                 DRM_ERROR("Execbuf while VT-switched.\n");
1832                 mutex_unlock(&dev->struct_mutex);
1833                 return -EBUSY;
1834         }
1835
1836         /* Zero the gloabl flush/invalidate flags. These
1837          * will be modified as each object is bound to the
1838          * gtt
1839          */
1840         dev->invalidate_domains = 0;
1841         dev->flush_domains = 0;
1842
1843         /* Look up object handles and perform the relocations */
1844         for (i = 0; i < args->buffer_count; i++) {
1845                 object_list[i] = drm_gem_object_lookup(dev, file_priv,
1846                                                        exec_list[i].handle);
1847                 if (object_list[i] == NULL) {
1848                         DRM_ERROR("Invalid object handle %d at index %d\n",
1849                                    exec_list[i].handle, i);
1850                         ret = -EBADF;
1851                         goto err;
1852                 }
1853
1854                 object_list[i]->pending_read_domains = 0;
1855                 object_list[i]->pending_write_domain = 0;
1856                 ret = i915_gem_object_pin_and_relocate(object_list[i],
1857                                                        file_priv,
1858                                                        &exec_list[i]);
1859                 if (ret) {
1860                         DRM_ERROR("object bind and relocate failed %d\n", ret);
1861                         goto err;
1862                 }
1863                 pinned = i + 1;
1864         }
1865
1866         /* Set the pending read domains for the batch buffer to COMMAND */
1867         batch_obj = object_list[args->buffer_count-1];
1868         batch_obj->pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1869         batch_obj->pending_write_domain = 0;
1870
1871         i915_verify_inactive(dev, __FILE__, __LINE__);
1872
1873         for (i = 0; i < args->buffer_count; i++) {
1874                 struct drm_gem_object *obj = object_list[i];
1875                 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1876
1877                 if (obj_priv->gtt_space == NULL) {
1878                         /* We evicted the buffer in the process of validating
1879                          * our set of buffers in.  We could try to recover by
1880                          * kicking them everything out and trying again from
1881                          * the start.
1882                          */
1883                         ret = -ENOMEM;
1884                         goto err;
1885                 }
1886
1887                 /* make sure all previous memory operations have passed */
1888                 ret = i915_gem_object_set_domain(obj,
1889                                                  obj->pending_read_domains,
1890                                                  obj->pending_write_domain);
1891                 if (ret)
1892                         goto err;
1893         }
1894
1895         i915_verify_inactive(dev, __FILE__, __LINE__);
1896
1897         /* Flush/invalidate caches and chipset buffer */
1898         flush_domains = i915_gem_dev_set_domain(dev);
1899
1900         i915_verify_inactive(dev, __FILE__, __LINE__);
1901
1902 #if WATCH_COHERENCY
1903         for (i = 0; i < args->buffer_count; i++) {
1904                 i915_gem_object_check_coherency(object_list[i],
1905                                                 exec_list[i].handle);
1906         }
1907 #endif
1908
1909         exec_offset = exec_list[args->buffer_count - 1].offset;
1910
1911 #if WATCH_EXEC
1912         i915_gem_dump_object(object_list[args->buffer_count - 1],
1913                               args->batch_len,
1914                               __func__,
1915                               ~0);
1916 #endif
1917
1918         (void)i915_add_request(dev, flush_domains);
1919
1920         /* Exec the batchbuffer */
1921         ret = i915_dispatch_gem_execbuffer(dev, args, exec_offset);
1922         if (ret) {
1923                 DRM_ERROR("dispatch failed %d\n", ret);
1924                 goto err;
1925         }
1926
1927         /*
1928          * Ensure that the commands in the batch buffer are
1929          * finished before the interrupt fires
1930          */
1931         flush_domains = i915_retire_commands(dev);
1932
1933         i915_verify_inactive(dev, __FILE__, __LINE__);
1934
1935         /*
1936          * Get a seqno representing the execution of the current buffer,
1937          * which we can wait on.  We would like to mitigate these interrupts,
1938          * likely by only creating seqnos occasionally (so that we have
1939          * *some* interrupts representing completion of buffers that we can
1940          * wait on when trying to clear up gtt space).
1941          */
1942         seqno = i915_add_request(dev, flush_domains);
1943         BUG_ON(seqno == 0);
1944         i915_file_priv->mm.last_gem_seqno = seqno;
1945         for (i = 0; i < args->buffer_count; i++) {
1946                 struct drm_gem_object *obj = object_list[i];
1947                 struct drm_i915_gem_object *obj_priv = obj->driver_private;
1948
1949                 i915_gem_object_move_to_active(obj);
1950                 obj_priv->last_rendering_seqno = seqno;
1951 #if WATCH_LRU
1952                 DRM_INFO("%s: move to exec list %p\n", __func__, obj);
1953 #endif
1954         }
1955 #if WATCH_LRU
1956         i915_dump_lru(dev, __func__);
1957 #endif
1958
1959         i915_verify_inactive(dev, __FILE__, __LINE__);
1960
1961         /* Copy the new buffer offsets back to the user's exec list. */
1962         ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1963                            (uintptr_t) args->buffers_ptr,
1964                            exec_list,
1965                            sizeof(*exec_list) * args->buffer_count);
1966         if (ret)
1967                 DRM_ERROR("failed to copy %d exec entries "
1968                           "back to user (%d)\n",
1969                            args->buffer_count, ret);
1970 err:
1971         if (object_list != NULL) {
1972                 for (i = 0; i < pinned; i++)
1973                         i915_gem_object_unpin(object_list[i]);
1974
1975                 for (i = 0; i < args->buffer_count; i++)
1976                         drm_gem_object_unreference(object_list[i]);
1977         }
1978         mutex_unlock(&dev->struct_mutex);
1979
1980 pre_mutex_err:
1981         drm_free(object_list, sizeof(*object_list) * args->buffer_count,
1982                  DRM_MEM_DRIVER);
1983         drm_free(exec_list, sizeof(*exec_list) * args->buffer_count,
1984                  DRM_MEM_DRIVER);
1985
1986         return ret;
1987 }
1988
1989 int
1990 i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment)
1991 {
1992         struct drm_device *dev = obj->dev;
1993         struct drm_i915_gem_object *obj_priv = obj->driver_private;
1994         int ret;
1995
1996         i915_verify_inactive(dev, __FILE__, __LINE__);
1997         if (obj_priv->gtt_space == NULL) {
1998                 ret = i915_gem_object_bind_to_gtt(obj, alignment);
1999                 if (ret != 0) {
2000                         DRM_ERROR("Failure to bind: %d", ret);
2001                         return ret;
2002                 }
2003         }
2004         obj_priv->pin_count++;
2005
2006         /* If the object is not active and not pending a flush,
2007          * remove it from the inactive list
2008          */
2009         if (obj_priv->pin_count == 1) {
2010                 atomic_inc(&dev->pin_count);
2011                 atomic_add(obj->size, &dev->pin_memory);
2012                 if (!obj_priv->active &&
2013                     (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
2014                                            I915_GEM_DOMAIN_GTT)) == 0 &&
2015                     !list_empty(&obj_priv->list))
2016                         list_del_init(&obj_priv->list);
2017         }
2018         i915_verify_inactive(dev, __FILE__, __LINE__);
2019
2020         return 0;
2021 }
2022
2023 void
2024 i915_gem_object_unpin(struct drm_gem_object *obj)
2025 {
2026         struct drm_device *dev = obj->dev;
2027         drm_i915_private_t *dev_priv = dev->dev_private;
2028         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2029
2030         i915_verify_inactive(dev, __FILE__, __LINE__);
2031         obj_priv->pin_count--;
2032         BUG_ON(obj_priv->pin_count < 0);
2033         BUG_ON(obj_priv->gtt_space == NULL);
2034
2035         /* If the object is no longer pinned, and is
2036          * neither active nor being flushed, then stick it on
2037          * the inactive list
2038          */
2039         if (obj_priv->pin_count == 0) {
2040                 if (!obj_priv->active &&
2041                     (obj->write_domain & ~(I915_GEM_DOMAIN_CPU |
2042                                            I915_GEM_DOMAIN_GTT)) == 0)
2043                         list_move_tail(&obj_priv->list,
2044                                        &dev_priv->mm.inactive_list);
2045                 atomic_dec(&dev->pin_count);
2046                 atomic_sub(obj->size, &dev->pin_memory);
2047         }
2048         i915_verify_inactive(dev, __FILE__, __LINE__);
2049 }
2050
2051 int
2052 i915_gem_pin_ioctl(struct drm_device *dev, void *data,
2053                    struct drm_file *file_priv)
2054 {
2055         struct drm_i915_gem_pin *args = data;
2056         struct drm_gem_object *obj;
2057         struct drm_i915_gem_object *obj_priv;
2058         int ret;
2059
2060         mutex_lock(&dev->struct_mutex);
2061
2062         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
2063         if (obj == NULL) {
2064                 DRM_ERROR("Bad handle in i915_gem_pin_ioctl(): %d\n",
2065                           args->handle);
2066                 mutex_unlock(&dev->struct_mutex);
2067                 return -EBADF;
2068         }
2069         obj_priv = obj->driver_private;
2070
2071         ret = i915_gem_object_pin(obj, args->alignment);
2072         if (ret != 0) {
2073                 drm_gem_object_unreference(obj);
2074                 mutex_unlock(&dev->struct_mutex);
2075                 return ret;
2076         }
2077
2078         /* XXX - flush the CPU caches for pinned objects
2079          * as the X server doesn't manage domains yet
2080          */
2081         if (obj->write_domain & I915_GEM_DOMAIN_CPU) {
2082                 i915_gem_clflush_object(obj);
2083                 drm_agp_chipset_flush(dev);
2084                 obj->write_domain = 0;
2085         }
2086         args->offset = obj_priv->gtt_offset;
2087         drm_gem_object_unreference(obj);
2088         mutex_unlock(&dev->struct_mutex);
2089
2090         return 0;
2091 }
2092
2093 int
2094 i915_gem_unpin_ioctl(struct drm_device *dev, void *data,
2095                      struct drm_file *file_priv)
2096 {
2097         struct drm_i915_gem_pin *args = data;
2098         struct drm_gem_object *obj;
2099
2100         mutex_lock(&dev->struct_mutex);
2101
2102         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
2103         if (obj == NULL) {
2104                 DRM_ERROR("Bad handle in i915_gem_unpin_ioctl(): %d\n",
2105                           args->handle);
2106                 mutex_unlock(&dev->struct_mutex);
2107                 return -EBADF;
2108         }
2109
2110         i915_gem_object_unpin(obj);
2111
2112         drm_gem_object_unreference(obj);
2113         mutex_unlock(&dev->struct_mutex);
2114         return 0;
2115 }
2116
2117 int
2118 i915_gem_busy_ioctl(struct drm_device *dev, void *data,
2119                     struct drm_file *file_priv)
2120 {
2121         struct drm_i915_gem_busy *args = data;
2122         struct drm_gem_object *obj;
2123         struct drm_i915_gem_object *obj_priv;
2124
2125         mutex_lock(&dev->struct_mutex);
2126         obj = drm_gem_object_lookup(dev, file_priv, args->handle);
2127         if (obj == NULL) {
2128                 DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n",
2129                           args->handle);
2130                 mutex_unlock(&dev->struct_mutex);
2131                 return -EBADF;
2132         }
2133
2134         obj_priv = obj->driver_private;
2135         args->busy = obj_priv->active;
2136
2137         drm_gem_object_unreference(obj);
2138         mutex_unlock(&dev->struct_mutex);
2139         return 0;
2140 }
2141
2142 int
2143 i915_gem_throttle_ioctl(struct drm_device *dev, void *data,
2144                         struct drm_file *file_priv)
2145 {
2146     return i915_gem_ring_throttle(dev, file_priv);
2147 }
2148
2149 int i915_gem_init_object(struct drm_gem_object *obj)
2150 {
2151         struct drm_i915_gem_object *obj_priv;
2152
2153         obj_priv = drm_calloc(1, sizeof(*obj_priv), DRM_MEM_DRIVER);
2154         if (obj_priv == NULL)
2155                 return -ENOMEM;
2156
2157         /*
2158          * We've just allocated pages from the kernel,
2159          * so they've just been written by the CPU with
2160          * zeros. They'll need to be clflushed before we
2161          * use them with the GPU.
2162          */
2163         obj->write_domain = I915_GEM_DOMAIN_CPU;
2164         obj->read_domains = I915_GEM_DOMAIN_CPU;
2165
2166         obj_priv->agp_type = AGP_USER_MEMORY;
2167
2168         obj->driver_private = obj_priv;
2169         obj_priv->obj = obj;
2170         INIT_LIST_HEAD(&obj_priv->list);
2171         return 0;
2172 }
2173
2174 void i915_gem_free_object(struct drm_gem_object *obj)
2175 {
2176         struct drm_i915_gem_object *obj_priv = obj->driver_private;
2177
2178         while (obj_priv->pin_count > 0)
2179                 i915_gem_object_unpin(obj);
2180
2181         i915_gem_object_unbind(obj);
2182
2183         drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER);
2184         drm_free(obj->driver_private, 1, DRM_MEM_DRIVER);
2185 }
2186
2187 static int
2188 i915_gem_set_domain(struct drm_gem_object *obj,
2189                     struct drm_file *file_priv,
2190                     uint32_t read_domains,
2191                     uint32_t write_domain)
2192 {
2193         struct drm_device *dev = obj->dev;
2194         int ret;
2195         uint32_t flush_domains;
2196
2197         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
2198
2199         ret = i915_gem_object_set_domain(obj, read_domains, write_domain);
2200         if (ret)
2201                 return ret;
2202         flush_domains = i915_gem_dev_set_domain(obj->dev);
2203
2204         if (flush_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT))
2205                 (void) i915_add_request(dev, flush_domains);
2206
2207         return 0;
2208 }
2209
2210 /** Unbinds all objects that are on the given buffer list. */
2211 static int
2212 i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head)
2213 {
2214         struct drm_gem_object *obj;
2215         struct drm_i915_gem_object *obj_priv;
2216         int ret;
2217
2218         while (!list_empty(head)) {
2219                 obj_priv = list_first_entry(head,
2220                                             struct drm_i915_gem_object,
2221                                             list);
2222                 obj = obj_priv->obj;
2223
2224                 if (obj_priv->pin_count != 0) {
2225                         DRM_ERROR("Pinned object in unbind list\n");
2226                         mutex_unlock(&dev->struct_mutex);
2227                         return -EINVAL;
2228                 }
2229
2230                 ret = i915_gem_object_unbind(obj);
2231                 if (ret != 0) {
2232                         DRM_ERROR("Error unbinding object in LeaveVT: %d\n",
2233                                   ret);
2234                         mutex_unlock(&dev->struct_mutex);
2235                         return ret;
2236                 }
2237         }
2238
2239
2240         return 0;
2241 }
2242
2243 static int
2244 i915_gem_idle(struct drm_device *dev)
2245 {
2246         drm_i915_private_t *dev_priv = dev->dev_private;
2247         uint32_t seqno, cur_seqno, last_seqno;
2248         int stuck, ret;
2249
2250         mutex_lock(&dev->struct_mutex);
2251
2252         if (dev_priv->mm.suspended || dev_priv->ring.ring_obj == NULL) {
2253                 mutex_unlock(&dev->struct_mutex);
2254                 return 0;
2255         }
2256
2257         /* Hack!  Don't let anybody do execbuf while we don't control the chip.
2258          * We need to replace this with a semaphore, or something.
2259          */
2260         dev_priv->mm.suspended = 1;
2261
2262         /* Cancel the retire work handler, wait for it to finish if running
2263          */
2264         mutex_unlock(&dev->struct_mutex);
2265         cancel_delayed_work_sync(&dev_priv->mm.retire_work);
2266         mutex_lock(&dev->struct_mutex);
2267
2268         i915_kernel_lost_context(dev);
2269
2270         /* Flush the GPU along with all non-CPU write domains
2271          */
2272         i915_gem_flush(dev, ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT),
2273                        ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT));
2274         seqno = i915_add_request(dev, ~(I915_GEM_DOMAIN_CPU |
2275                                         I915_GEM_DOMAIN_GTT));
2276
2277         if (seqno == 0) {
2278                 mutex_unlock(&dev->struct_mutex);
2279                 return -ENOMEM;
2280         }
2281
2282         dev_priv->mm.waiting_gem_seqno = seqno;
2283         last_seqno = 0;
2284         stuck = 0;
2285         for (;;) {
2286                 cur_seqno = i915_get_gem_seqno(dev);
2287                 if (i915_seqno_passed(cur_seqno, seqno))
2288                         break;
2289                 if (last_seqno == cur_seqno) {
2290                         if (stuck++ > 100) {
2291                                 DRM_ERROR("hardware wedged\n");
2292                                 dev_priv->mm.wedged = 1;
2293                                 DRM_WAKEUP(&dev_priv->irq_queue);
2294                                 break;
2295                         }
2296                 }
2297                 msleep(10);
2298                 last_seqno = cur_seqno;
2299         }
2300         dev_priv->mm.waiting_gem_seqno = 0;
2301
2302         i915_gem_retire_requests(dev);
2303
2304         /* Active and flushing should now be empty as we've
2305          * waited for a sequence higher than any pending execbuffer
2306          */
2307         BUG_ON(!list_empty(&dev_priv->mm.active_list));
2308         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2309
2310         /* Request should now be empty as we've also waited
2311          * for the last request in the list
2312          */
2313         BUG_ON(!list_empty(&dev_priv->mm.request_list));
2314
2315         /* Move all buffers out of the GTT. */
2316         ret = i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list);
2317         if (ret) {
2318                 mutex_unlock(&dev->struct_mutex);
2319                 return ret;
2320         }
2321
2322         BUG_ON(!list_empty(&dev_priv->mm.active_list));
2323         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2324         BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
2325         BUG_ON(!list_empty(&dev_priv->mm.request_list));
2326
2327         i915_gem_cleanup_ringbuffer(dev);
2328         mutex_unlock(&dev->struct_mutex);
2329
2330         return 0;
2331 }
2332
2333 static int
2334 i915_gem_init_hws(struct drm_device *dev)
2335 {
2336         drm_i915_private_t *dev_priv = dev->dev_private;
2337         struct drm_gem_object *obj;
2338         struct drm_i915_gem_object *obj_priv;
2339         int ret;
2340
2341         /* If we need a physical address for the status page, it's already
2342          * initialized at driver load time.
2343          */
2344         if (!I915_NEED_GFX_HWS(dev))
2345                 return 0;
2346
2347         obj = drm_gem_object_alloc(dev, 4096);
2348         if (obj == NULL) {
2349                 DRM_ERROR("Failed to allocate status page\n");
2350                 return -ENOMEM;
2351         }
2352         obj_priv = obj->driver_private;
2353         obj_priv->agp_type = AGP_USER_CACHED_MEMORY;
2354
2355         ret = i915_gem_object_pin(obj, 4096);
2356         if (ret != 0) {
2357                 drm_gem_object_unreference(obj);
2358                 return ret;
2359         }
2360
2361         dev_priv->status_gfx_addr = obj_priv->gtt_offset;
2362
2363         dev_priv->hw_status_page = kmap(obj_priv->page_list[0]);
2364         if (dev_priv->hw_status_page == NULL) {
2365                 DRM_ERROR("Failed to map status page.\n");
2366                 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
2367                 drm_gem_object_unreference(obj);
2368                 return -EINVAL;
2369         }
2370         dev_priv->hws_obj = obj;
2371         memset(dev_priv->hw_status_page, 0, PAGE_SIZE);
2372         I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr);
2373         I915_READ(HWS_PGA); /* posting read */
2374         DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr);
2375
2376         return 0;
2377 }
2378
2379 static int
2380 i915_gem_init_ringbuffer(struct drm_device *dev)
2381 {
2382         drm_i915_private_t *dev_priv = dev->dev_private;
2383         struct drm_gem_object *obj;
2384         struct drm_i915_gem_object *obj_priv;
2385         int ret;
2386         u32 head;
2387
2388         ret = i915_gem_init_hws(dev);
2389         if (ret != 0)
2390                 return ret;
2391
2392         obj = drm_gem_object_alloc(dev, 128 * 1024);
2393         if (obj == NULL) {
2394                 DRM_ERROR("Failed to allocate ringbuffer\n");
2395                 return -ENOMEM;
2396         }
2397         obj_priv = obj->driver_private;
2398
2399         ret = i915_gem_object_pin(obj, 4096);
2400         if (ret != 0) {
2401                 drm_gem_object_unreference(obj);
2402                 return ret;
2403         }
2404
2405         /* Set up the kernel mapping for the ring. */
2406         dev_priv->ring.Size = obj->size;
2407         dev_priv->ring.tail_mask = obj->size - 1;
2408
2409         dev_priv->ring.map.offset = dev->agp->base + obj_priv->gtt_offset;
2410         dev_priv->ring.map.size = obj->size;
2411         dev_priv->ring.map.type = 0;
2412         dev_priv->ring.map.flags = 0;
2413         dev_priv->ring.map.mtrr = 0;
2414
2415         drm_core_ioremap_wc(&dev_priv->ring.map, dev);
2416         if (dev_priv->ring.map.handle == NULL) {
2417                 DRM_ERROR("Failed to map ringbuffer.\n");
2418                 memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
2419                 drm_gem_object_unreference(obj);
2420                 return -EINVAL;
2421         }
2422         dev_priv->ring.ring_obj = obj;
2423         dev_priv->ring.virtual_start = dev_priv->ring.map.handle;
2424
2425         /* Stop the ring if it's running. */
2426         I915_WRITE(PRB0_CTL, 0);
2427         I915_WRITE(PRB0_TAIL, 0);
2428         I915_WRITE(PRB0_HEAD, 0);
2429
2430         /* Initialize the ring. */
2431         I915_WRITE(PRB0_START, obj_priv->gtt_offset);
2432         head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
2433
2434         /* G45 ring initialization fails to reset head to zero */
2435         if (head != 0) {
2436                 DRM_ERROR("Ring head not reset to zero "
2437                           "ctl %08x head %08x tail %08x start %08x\n",
2438                           I915_READ(PRB0_CTL),
2439                           I915_READ(PRB0_HEAD),
2440                           I915_READ(PRB0_TAIL),
2441                           I915_READ(PRB0_START));
2442                 I915_WRITE(PRB0_HEAD, 0);
2443
2444                 DRM_ERROR("Ring head forced to zero "
2445                           "ctl %08x head %08x tail %08x start %08x\n",
2446                           I915_READ(PRB0_CTL),
2447                           I915_READ(PRB0_HEAD),
2448                           I915_READ(PRB0_TAIL),
2449                           I915_READ(PRB0_START));
2450         }
2451
2452         I915_WRITE(PRB0_CTL,
2453                    ((obj->size - 4096) & RING_NR_PAGES) |
2454                    RING_NO_REPORT |
2455                    RING_VALID);
2456
2457         head = I915_READ(PRB0_HEAD) & HEAD_ADDR;
2458
2459         /* If the head is still not zero, the ring is dead */
2460         if (head != 0) {
2461                 DRM_ERROR("Ring initialization failed "
2462                           "ctl %08x head %08x tail %08x start %08x\n",
2463                           I915_READ(PRB0_CTL),
2464                           I915_READ(PRB0_HEAD),
2465                           I915_READ(PRB0_TAIL),
2466                           I915_READ(PRB0_START));
2467                 return -EIO;
2468         }
2469
2470         /* Update our cache of the ring state */
2471         i915_kernel_lost_context(dev);
2472
2473         return 0;
2474 }
2475
2476 static void
2477 i915_gem_cleanup_ringbuffer(struct drm_device *dev)
2478 {
2479         drm_i915_private_t *dev_priv = dev->dev_private;
2480
2481         if (dev_priv->ring.ring_obj == NULL)
2482                 return;
2483
2484         drm_core_ioremapfree(&dev_priv->ring.map, dev);
2485
2486         i915_gem_object_unpin(dev_priv->ring.ring_obj);
2487         drm_gem_object_unreference(dev_priv->ring.ring_obj);
2488         dev_priv->ring.ring_obj = NULL;
2489         memset(&dev_priv->ring, 0, sizeof(dev_priv->ring));
2490
2491         if (dev_priv->hws_obj != NULL) {
2492                 struct drm_gem_object *obj = dev_priv->hws_obj;
2493                 struct drm_i915_gem_object *obj_priv = obj->driver_private;
2494
2495                 kunmap(obj_priv->page_list[0]);
2496                 i915_gem_object_unpin(obj);
2497                 drm_gem_object_unreference(obj);
2498                 dev_priv->hws_obj = NULL;
2499                 memset(&dev_priv->hws_map, 0, sizeof(dev_priv->hws_map));
2500                 dev_priv->hw_status_page = NULL;
2501
2502                 /* Write high address into HWS_PGA when disabling. */
2503                 I915_WRITE(HWS_PGA, 0x1ffff000);
2504         }
2505 }
2506
2507 int
2508 i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
2509                        struct drm_file *file_priv)
2510 {
2511         drm_i915_private_t *dev_priv = dev->dev_private;
2512         int ret;
2513
2514         if (dev_priv->mm.wedged) {
2515                 DRM_ERROR("Reenabling wedged hardware, good luck\n");
2516                 dev_priv->mm.wedged = 0;
2517         }
2518
2519         ret = i915_gem_init_ringbuffer(dev);
2520         if (ret != 0)
2521                 return ret;
2522
2523         dev_priv->mm.gtt_mapping = io_mapping_create_wc(dev->agp->base,
2524                                                         dev->agp->agp_info.aper_size
2525                                                         * 1024 * 1024);
2526
2527         mutex_lock(&dev->struct_mutex);
2528         BUG_ON(!list_empty(&dev_priv->mm.active_list));
2529         BUG_ON(!list_empty(&dev_priv->mm.flushing_list));
2530         BUG_ON(!list_empty(&dev_priv->mm.inactive_list));
2531         BUG_ON(!list_empty(&dev_priv->mm.request_list));
2532         dev_priv->mm.suspended = 0;
2533         mutex_unlock(&dev->struct_mutex);
2534
2535         drm_irq_install(dev);
2536
2537         return 0;
2538 }
2539
2540 int
2541 i915_gem_leavevt_ioctl(struct drm_device *dev, void *data,
2542                        struct drm_file *file_priv)
2543 {
2544         drm_i915_private_t *dev_priv = dev->dev_private;
2545         int ret;
2546
2547         ret = i915_gem_idle(dev);
2548         drm_irq_uninstall(dev);
2549
2550         io_mapping_free(dev_priv->mm.gtt_mapping);
2551         return ret;
2552 }
2553
2554 void
2555 i915_gem_lastclose(struct drm_device *dev)
2556 {
2557         int ret;
2558
2559         ret = i915_gem_idle(dev);
2560         if (ret)
2561                 DRM_ERROR("failed to idle hardware: %d\n", ret);
2562 }
2563
2564 void
2565 i915_gem_load(struct drm_device *dev)
2566 {
2567         drm_i915_private_t *dev_priv = dev->dev_private;
2568
2569         INIT_LIST_HEAD(&dev_priv->mm.active_list);
2570         INIT_LIST_HEAD(&dev_priv->mm.flushing_list);
2571         INIT_LIST_HEAD(&dev_priv->mm.inactive_list);
2572         INIT_LIST_HEAD(&dev_priv->mm.request_list);
2573         INIT_DELAYED_WORK(&dev_priv->mm.retire_work,
2574                           i915_gem_retire_work_handler);
2575         dev_priv->mm.next_gem_seqno = 1;
2576
2577         i915_gem_detect_bit_6_swizzle(dev);
2578 }