Merge git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable
[sfrench/cifs-2.6.git] / drivers / gpu / drm / radeon / radeon_object.c
1 /*
2  * Copyright 2009 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30  *    Dave Airlie
31  */
32 #include <linux/list.h>
33 #include <drm/drmP.h>
34 #include "radeon_drm.h"
35 #include "radeon.h"
36
37
38 int radeon_ttm_init(struct radeon_device *rdev);
39 void radeon_ttm_fini(struct radeon_device *rdev);
40 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
41
42 /*
43  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
44  * function are calling it.
45  */
46
47 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
48 {
49         struct radeon_bo *bo;
50
51         bo = container_of(tbo, struct radeon_bo, tbo);
52         mutex_lock(&bo->rdev->gem.mutex);
53         list_del_init(&bo->list);
54         mutex_unlock(&bo->rdev->gem.mutex);
55         radeon_bo_clear_surface_reg(bo);
56         kfree(bo);
57 }
58
59 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
60 {
61         if (bo->destroy == &radeon_ttm_bo_destroy)
62                 return true;
63         return false;
64 }
65
66 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
67 {
68         u32 c = 0;
69
70         rbo->placement.fpfn = 0;
71         rbo->placement.lpfn = 0;
72         rbo->placement.placement = rbo->placements;
73         rbo->placement.busy_placement = rbo->placements;
74         if (domain & RADEON_GEM_DOMAIN_VRAM)
75                 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
76                                         TTM_PL_FLAG_VRAM;
77         if (domain & RADEON_GEM_DOMAIN_GTT)
78                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
79         if (domain & RADEON_GEM_DOMAIN_CPU)
80                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
81         if (!c)
82                 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
83         rbo->placement.num_placement = c;
84         rbo->placement.num_busy_placement = c;
85 }
86
87 int radeon_bo_create(struct radeon_device *rdev, struct drm_gem_object *gobj,
88                         unsigned long size, bool kernel, u32 domain,
89                         struct radeon_bo **bo_ptr)
90 {
91         struct radeon_bo *bo;
92         enum ttm_bo_type type;
93         int r;
94
95         if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
96                 rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
97         }
98         if (kernel) {
99                 type = ttm_bo_type_kernel;
100         } else {
101                 type = ttm_bo_type_device;
102         }
103         *bo_ptr = NULL;
104         bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
105         if (bo == NULL)
106                 return -ENOMEM;
107         bo->rdev = rdev;
108         bo->gobj = gobj;
109         bo->surface_reg = -1;
110         INIT_LIST_HEAD(&bo->list);
111
112         radeon_ttm_placement_from_domain(bo, domain);
113         /* Kernel allocation are uninterruptible */
114         r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
115                         &bo->placement, 0, 0, !kernel, NULL, size,
116                         &radeon_ttm_bo_destroy);
117         if (unlikely(r != 0)) {
118                 if (r != -ERESTARTSYS)
119                         dev_err(rdev->dev,
120                                 "object_init failed for (%lu, 0x%08X)\n",
121                                 size, domain);
122                 return r;
123         }
124         *bo_ptr = bo;
125         if (gobj) {
126                 mutex_lock(&bo->rdev->gem.mutex);
127                 list_add_tail(&bo->list, &rdev->gem.objects);
128                 mutex_unlock(&bo->rdev->gem.mutex);
129         }
130         return 0;
131 }
132
133 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
134 {
135         bool is_iomem;
136         int r;
137
138         if (bo->kptr) {
139                 if (ptr) {
140                         *ptr = bo->kptr;
141                 }
142                 return 0;
143         }
144         r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
145         if (r) {
146                 return r;
147         }
148         bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
149         if (ptr) {
150                 *ptr = bo->kptr;
151         }
152         radeon_bo_check_tiling(bo, 0, 0);
153         return 0;
154 }
155
156 void radeon_bo_kunmap(struct radeon_bo *bo)
157 {
158         if (bo->kptr == NULL)
159                 return;
160         bo->kptr = NULL;
161         radeon_bo_check_tiling(bo, 0, 0);
162         ttm_bo_kunmap(&bo->kmap);
163 }
164
165 void radeon_bo_unref(struct radeon_bo **bo)
166 {
167         struct ttm_buffer_object *tbo;
168
169         if ((*bo) == NULL)
170                 return;
171         tbo = &((*bo)->tbo);
172         ttm_bo_unref(&tbo);
173         if (tbo == NULL)
174                 *bo = NULL;
175 }
176
177 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
178 {
179         int r, i;
180
181         radeon_ttm_placement_from_domain(bo, domain);
182         if (bo->pin_count) {
183                 bo->pin_count++;
184                 if (gpu_addr)
185                         *gpu_addr = radeon_bo_gpu_offset(bo);
186                 return 0;
187         }
188         radeon_ttm_placement_from_domain(bo, domain);
189         for (i = 0; i < bo->placement.num_placement; i++)
190                 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
191         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
192         if (likely(r == 0)) {
193                 bo->pin_count = 1;
194                 if (gpu_addr != NULL)
195                         *gpu_addr = radeon_bo_gpu_offset(bo);
196         }
197         if (unlikely(r != 0))
198                 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
199         return r;
200 }
201
202 int radeon_bo_unpin(struct radeon_bo *bo)
203 {
204         int r, i;
205
206         if (!bo->pin_count) {
207                 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
208                 return 0;
209         }
210         bo->pin_count--;
211         if (bo->pin_count)
212                 return 0;
213         for (i = 0; i < bo->placement.num_placement; i++)
214                 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
215         r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
216         if (unlikely(r != 0))
217                 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
218         return r;
219 }
220
221 int radeon_bo_evict_vram(struct radeon_device *rdev)
222 {
223         /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
224         if (0 && (rdev->flags & RADEON_IS_IGP)) {
225                 if (rdev->mc.igp_sideport_enabled == false)
226                         /* Useless to evict on IGP chips */
227                         return 0;
228         }
229         return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
230 }
231
232 void radeon_bo_force_delete(struct radeon_device *rdev)
233 {
234         struct radeon_bo *bo, *n;
235         struct drm_gem_object *gobj;
236
237         if (list_empty(&rdev->gem.objects)) {
238                 return;
239         }
240         dev_err(rdev->dev, "Userspace still has active objects !\n");
241         list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
242                 mutex_lock(&rdev->ddev->struct_mutex);
243                 gobj = bo->gobj;
244                 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
245                         gobj, bo, (unsigned long)gobj->size,
246                         *((unsigned long *)&gobj->refcount));
247                 mutex_lock(&bo->rdev->gem.mutex);
248                 list_del_init(&bo->list);
249                 mutex_unlock(&bo->rdev->gem.mutex);
250                 radeon_bo_unref(&bo);
251                 gobj->driver_private = NULL;
252                 drm_gem_object_unreference(gobj);
253                 mutex_unlock(&rdev->ddev->struct_mutex);
254         }
255 }
256
257 int radeon_bo_init(struct radeon_device *rdev)
258 {
259         /* Add an MTRR for the VRAM */
260         rdev->mc.vram_mtrr = mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size,
261                         MTRR_TYPE_WRCOMB, 1);
262         DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
263                 rdev->mc.mc_vram_size >> 20,
264                 (unsigned long long)rdev->mc.aper_size >> 20);
265         DRM_INFO("RAM width %dbits %cDR\n",
266                         rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
267         return radeon_ttm_init(rdev);
268 }
269
270 void radeon_bo_fini(struct radeon_device *rdev)
271 {
272         radeon_ttm_fini(rdev);
273 }
274
275 void radeon_bo_list_add_object(struct radeon_bo_list *lobj,
276                                 struct list_head *head)
277 {
278         if (lobj->wdomain) {
279                 list_add(&lobj->list, head);
280         } else {
281                 list_add_tail(&lobj->list, head);
282         }
283 }
284
285 int radeon_bo_list_reserve(struct list_head *head)
286 {
287         struct radeon_bo_list *lobj;
288         int r;
289
290         list_for_each_entry(lobj, head, list){
291                 r = radeon_bo_reserve(lobj->bo, false);
292                 if (unlikely(r != 0))
293                         return r;
294         }
295         return 0;
296 }
297
298 void radeon_bo_list_unreserve(struct list_head *head)
299 {
300         struct radeon_bo_list *lobj;
301
302         list_for_each_entry(lobj, head, list) {
303                 /* only unreserve object we successfully reserved */
304                 if (radeon_bo_is_reserved(lobj->bo))
305                         radeon_bo_unreserve(lobj->bo);
306         }
307 }
308
309 int radeon_bo_list_validate(struct list_head *head, void *fence)
310 {
311         struct radeon_bo_list *lobj;
312         struct radeon_bo *bo;
313         struct radeon_fence *old_fence = NULL;
314         int r;
315
316         r = radeon_bo_list_reserve(head);
317         if (unlikely(r != 0)) {
318                 return r;
319         }
320         list_for_each_entry(lobj, head, list) {
321                 bo = lobj->bo;
322                 if (!bo->pin_count) {
323                         if (lobj->wdomain) {
324                                 radeon_ttm_placement_from_domain(bo,
325                                                                 lobj->wdomain);
326                         } else {
327                                 radeon_ttm_placement_from_domain(bo,
328                                                                 lobj->rdomain);
329                         }
330                         r = ttm_bo_validate(&bo->tbo, &bo->placement,
331                                                 true, false);
332                         if (unlikely(r))
333                                 return r;
334                 }
335                 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
336                 lobj->tiling_flags = bo->tiling_flags;
337                 if (fence) {
338                         old_fence = (struct radeon_fence *)bo->tbo.sync_obj;
339                         bo->tbo.sync_obj = radeon_fence_ref(fence);
340                         bo->tbo.sync_obj_arg = NULL;
341                 }
342                 if (old_fence) {
343                         radeon_fence_unref(&old_fence);
344                 }
345         }
346         return 0;
347 }
348
349 void radeon_bo_list_unvalidate(struct list_head *head, void *fence)
350 {
351         struct radeon_bo_list *lobj;
352         struct radeon_fence *old_fence;
353
354         if (fence)
355                 list_for_each_entry(lobj, head, list) {
356                         old_fence = to_radeon_fence(lobj->bo->tbo.sync_obj);
357                         if (old_fence == fence) {
358                                 lobj->bo->tbo.sync_obj = NULL;
359                                 radeon_fence_unref(&old_fence);
360                         }
361                 }
362         radeon_bo_list_unreserve(head);
363 }
364
365 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
366                              struct vm_area_struct *vma)
367 {
368         return ttm_fbdev_mmap(vma, &bo->tbo);
369 }
370
371 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
372 {
373         struct radeon_device *rdev = bo->rdev;
374         struct radeon_surface_reg *reg;
375         struct radeon_bo *old_object;
376         int steal;
377         int i;
378
379         BUG_ON(!atomic_read(&bo->tbo.reserved));
380
381         if (!bo->tiling_flags)
382                 return 0;
383
384         if (bo->surface_reg >= 0) {
385                 reg = &rdev->surface_regs[bo->surface_reg];
386                 i = bo->surface_reg;
387                 goto out;
388         }
389
390         steal = -1;
391         for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
392
393                 reg = &rdev->surface_regs[i];
394                 if (!reg->bo)
395                         break;
396
397                 old_object = reg->bo;
398                 if (old_object->pin_count == 0)
399                         steal = i;
400         }
401
402         /* if we are all out */
403         if (i == RADEON_GEM_MAX_SURFACES) {
404                 if (steal == -1)
405                         return -ENOMEM;
406                 /* find someone with a surface reg and nuke their BO */
407                 reg = &rdev->surface_regs[steal];
408                 old_object = reg->bo;
409                 /* blow away the mapping */
410                 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
411                 ttm_bo_unmap_virtual(&old_object->tbo);
412                 old_object->surface_reg = -1;
413                 i = steal;
414         }
415
416         bo->surface_reg = i;
417         reg->bo = bo;
418
419 out:
420         radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
421                                bo->tbo.mem.mm_node->start << PAGE_SHIFT,
422                                bo->tbo.num_pages << PAGE_SHIFT);
423         return 0;
424 }
425
426 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
427 {
428         struct radeon_device *rdev = bo->rdev;
429         struct radeon_surface_reg *reg;
430
431         if (bo->surface_reg == -1)
432                 return;
433
434         reg = &rdev->surface_regs[bo->surface_reg];
435         radeon_clear_surface_reg(rdev, bo->surface_reg);
436
437         reg->bo = NULL;
438         bo->surface_reg = -1;
439 }
440
441 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
442                                 uint32_t tiling_flags, uint32_t pitch)
443 {
444         int r;
445
446         r = radeon_bo_reserve(bo, false);
447         if (unlikely(r != 0))
448                 return r;
449         bo->tiling_flags = tiling_flags;
450         bo->pitch = pitch;
451         radeon_bo_unreserve(bo);
452         return 0;
453 }
454
455 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
456                                 uint32_t *tiling_flags,
457                                 uint32_t *pitch)
458 {
459         BUG_ON(!atomic_read(&bo->tbo.reserved));
460         if (tiling_flags)
461                 *tiling_flags = bo->tiling_flags;
462         if (pitch)
463                 *pitch = bo->pitch;
464 }
465
466 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
467                                 bool force_drop)
468 {
469         BUG_ON(!atomic_read(&bo->tbo.reserved));
470
471         if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
472                 return 0;
473
474         if (force_drop) {
475                 radeon_bo_clear_surface_reg(bo);
476                 return 0;
477         }
478
479         if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
480                 if (!has_moved)
481                         return 0;
482
483                 if (bo->surface_reg >= 0)
484                         radeon_bo_clear_surface_reg(bo);
485                 return 0;
486         }
487
488         if ((bo->surface_reg >= 0) && !has_moved)
489                 return 0;
490
491         return radeon_bo_get_surface_reg(bo);
492 }
493
494 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
495                            struct ttm_mem_reg *mem)
496 {
497         struct radeon_bo *rbo;
498         if (!radeon_ttm_bo_is_radeon_bo(bo))
499                 return;
500         rbo = container_of(bo, struct radeon_bo, tbo);
501         radeon_bo_check_tiling(rbo, 0, 1);
502 }
503
504 void radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
505 {
506         struct radeon_bo *rbo;
507         if (!radeon_ttm_bo_is_radeon_bo(bo))
508                 return;
509         rbo = container_of(bo, struct radeon_bo, tbo);
510         radeon_bo_check_tiling(rbo, 0, 0);
511 }