Loading...
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 <linux/slab.h>
34#include <drm/drmP.h>
35#include "radeon_drm.h"
36#include "radeon.h"
37#include "radeon_trace.h"
38
39
40int radeon_ttm_init(struct radeon_device *rdev);
41void radeon_ttm_fini(struct radeon_device *rdev);
42static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
43
44/*
45 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
46 * function are calling it.
47 */
48
49static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
50{
51 struct radeon_bo *bo;
52
53 bo = container_of(tbo, struct radeon_bo, tbo);
54 mutex_lock(&bo->rdev->gem.mutex);
55 list_del_init(&bo->list);
56 mutex_unlock(&bo->rdev->gem.mutex);
57 radeon_bo_clear_surface_reg(bo);
58 drm_gem_object_release(&bo->gem_base);
59 kfree(bo);
60}
61
62bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
63{
64 if (bo->destroy == &radeon_ttm_bo_destroy)
65 return true;
66 return false;
67}
68
69void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
70{
71 u32 c = 0;
72
73 rbo->placement.fpfn = 0;
74 rbo->placement.lpfn = 0;
75 rbo->placement.placement = rbo->placements;
76 rbo->placement.busy_placement = rbo->placements;
77 if (domain & RADEON_GEM_DOMAIN_VRAM)
78 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
79 TTM_PL_FLAG_VRAM;
80 if (domain & RADEON_GEM_DOMAIN_GTT)
81 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
82 if (domain & RADEON_GEM_DOMAIN_CPU)
83 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
84 if (!c)
85 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
86 rbo->placement.num_placement = c;
87 rbo->placement.num_busy_placement = c;
88}
89
90int radeon_bo_create(struct radeon_device *rdev,
91 unsigned long size, int byte_align, bool kernel, u32 domain,
92 struct radeon_bo **bo_ptr)
93{
94 struct radeon_bo *bo;
95 enum ttm_bo_type type;
96 unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
97 unsigned long max_size = 0;
98 int r;
99
100 size = ALIGN(size, PAGE_SIZE);
101
102 if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
103 rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
104 }
105 if (kernel) {
106 type = ttm_bo_type_kernel;
107 } else {
108 type = ttm_bo_type_device;
109 }
110 *bo_ptr = NULL;
111
112 /* maximun bo size is the minimun btw visible vram and gtt size */
113 max_size = min(rdev->mc.visible_vram_size, rdev->mc.gtt_size);
114 if ((page_align << PAGE_SHIFT) >= max_size) {
115 printk(KERN_WARNING "%s:%d alloc size %ldM bigger than %ldMb limit\n",
116 __func__, __LINE__, page_align >> (20 - PAGE_SHIFT), max_size >> 20);
117 return -ENOMEM;
118 }
119
120retry:
121 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
122 if (bo == NULL)
123 return -ENOMEM;
124 r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
125 if (unlikely(r)) {
126 kfree(bo);
127 return r;
128 }
129 bo->rdev = rdev;
130 bo->gem_base.driver_private = NULL;
131 bo->surface_reg = -1;
132 INIT_LIST_HEAD(&bo->list);
133 radeon_ttm_placement_from_domain(bo, domain);
134 /* Kernel allocation are uninterruptible */
135 mutex_lock(&rdev->vram_mutex);
136 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
137 &bo->placement, page_align, 0, !kernel, NULL, size,
138 &radeon_ttm_bo_destroy);
139 mutex_unlock(&rdev->vram_mutex);
140 if (unlikely(r != 0)) {
141 if (r != -ERESTARTSYS) {
142 if (domain == RADEON_GEM_DOMAIN_VRAM) {
143 domain |= RADEON_GEM_DOMAIN_GTT;
144 goto retry;
145 }
146 dev_err(rdev->dev,
147 "object_init failed for (%lu, 0x%08X)\n",
148 size, domain);
149 }
150 return r;
151 }
152 *bo_ptr = bo;
153
154 trace_radeon_bo_create(bo);
155
156 return 0;
157}
158
159int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
160{
161 bool is_iomem;
162 int r;
163
164 if (bo->kptr) {
165 if (ptr) {
166 *ptr = bo->kptr;
167 }
168 return 0;
169 }
170 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
171 if (r) {
172 return r;
173 }
174 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
175 if (ptr) {
176 *ptr = bo->kptr;
177 }
178 radeon_bo_check_tiling(bo, 0, 0);
179 return 0;
180}
181
182void radeon_bo_kunmap(struct radeon_bo *bo)
183{
184 if (bo->kptr == NULL)
185 return;
186 bo->kptr = NULL;
187 radeon_bo_check_tiling(bo, 0, 0);
188 ttm_bo_kunmap(&bo->kmap);
189}
190
191void radeon_bo_unref(struct radeon_bo **bo)
192{
193 struct ttm_buffer_object *tbo;
194 struct radeon_device *rdev;
195
196 if ((*bo) == NULL)
197 return;
198 rdev = (*bo)->rdev;
199 tbo = &((*bo)->tbo);
200 mutex_lock(&rdev->vram_mutex);
201 ttm_bo_unref(&tbo);
202 mutex_unlock(&rdev->vram_mutex);
203 if (tbo == NULL)
204 *bo = NULL;
205}
206
207int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
208{
209 int r, i;
210
211 if (bo->pin_count) {
212 bo->pin_count++;
213 if (gpu_addr)
214 *gpu_addr = radeon_bo_gpu_offset(bo);
215 return 0;
216 }
217 radeon_ttm_placement_from_domain(bo, domain);
218 if (domain == RADEON_GEM_DOMAIN_VRAM) {
219 /* force to pin into visible video ram */
220 bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
221 }
222 for (i = 0; i < bo->placement.num_placement; i++)
223 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
224 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
225 if (likely(r == 0)) {
226 bo->pin_count = 1;
227 if (gpu_addr != NULL)
228 *gpu_addr = radeon_bo_gpu_offset(bo);
229 }
230 if (unlikely(r != 0))
231 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
232 return r;
233}
234
235int radeon_bo_unpin(struct radeon_bo *bo)
236{
237 int r, i;
238
239 if (!bo->pin_count) {
240 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
241 return 0;
242 }
243 bo->pin_count--;
244 if (bo->pin_count)
245 return 0;
246 for (i = 0; i < bo->placement.num_placement; i++)
247 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
248 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
249 if (unlikely(r != 0))
250 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
251 return r;
252}
253
254int radeon_bo_evict_vram(struct radeon_device *rdev)
255{
256 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
257 if (0 && (rdev->flags & RADEON_IS_IGP)) {
258 if (rdev->mc.igp_sideport_enabled == false)
259 /* Useless to evict on IGP chips */
260 return 0;
261 }
262 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
263}
264
265void radeon_bo_force_delete(struct radeon_device *rdev)
266{
267 struct radeon_bo *bo, *n;
268
269 if (list_empty(&rdev->gem.objects)) {
270 return;
271 }
272 dev_err(rdev->dev, "Userspace still has active objects !\n");
273 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
274 mutex_lock(&rdev->ddev->struct_mutex);
275 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
276 &bo->gem_base, bo, (unsigned long)bo->gem_base.size,
277 *((unsigned long *)&bo->gem_base.refcount));
278 mutex_lock(&bo->rdev->gem.mutex);
279 list_del_init(&bo->list);
280 mutex_unlock(&bo->rdev->gem.mutex);
281 /* this should unref the ttm bo */
282 drm_gem_object_unreference(&bo->gem_base);
283 mutex_unlock(&rdev->ddev->struct_mutex);
284 }
285}
286
287int radeon_bo_init(struct radeon_device *rdev)
288{
289 /* Add an MTRR for the VRAM */
290 rdev->mc.vram_mtrr = mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size,
291 MTRR_TYPE_WRCOMB, 1);
292 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
293 rdev->mc.mc_vram_size >> 20,
294 (unsigned long long)rdev->mc.aper_size >> 20);
295 DRM_INFO("RAM width %dbits %cDR\n",
296 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
297 return radeon_ttm_init(rdev);
298}
299
300void radeon_bo_fini(struct radeon_device *rdev)
301{
302 radeon_ttm_fini(rdev);
303}
304
305void radeon_bo_list_add_object(struct radeon_bo_list *lobj,
306 struct list_head *head)
307{
308 if (lobj->wdomain) {
309 list_add(&lobj->tv.head, head);
310 } else {
311 list_add_tail(&lobj->tv.head, head);
312 }
313}
314
315int radeon_bo_list_validate(struct list_head *head)
316{
317 struct radeon_bo_list *lobj;
318 struct radeon_bo *bo;
319 u32 domain;
320 int r;
321
322 r = ttm_eu_reserve_buffers(head);
323 if (unlikely(r != 0)) {
324 return r;
325 }
326 list_for_each_entry(lobj, head, tv.head) {
327 bo = lobj->bo;
328 if (!bo->pin_count) {
329 domain = lobj->wdomain ? lobj->wdomain : lobj->rdomain;
330
331 retry:
332 radeon_ttm_placement_from_domain(bo, domain);
333 r = ttm_bo_validate(&bo->tbo, &bo->placement,
334 true, false, false);
335 if (unlikely(r)) {
336 if (r != -ERESTARTSYS && domain == RADEON_GEM_DOMAIN_VRAM) {
337 domain |= RADEON_GEM_DOMAIN_GTT;
338 goto retry;
339 }
340 return r;
341 }
342 }
343 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
344 lobj->tiling_flags = bo->tiling_flags;
345 }
346 return 0;
347}
348
349int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
350 struct vm_area_struct *vma)
351{
352 return ttm_fbdev_mmap(vma, &bo->tbo);
353}
354
355int radeon_bo_get_surface_reg(struct radeon_bo *bo)
356{
357 struct radeon_device *rdev = bo->rdev;
358 struct radeon_surface_reg *reg;
359 struct radeon_bo *old_object;
360 int steal;
361 int i;
362
363 BUG_ON(!atomic_read(&bo->tbo.reserved));
364
365 if (!bo->tiling_flags)
366 return 0;
367
368 if (bo->surface_reg >= 0) {
369 reg = &rdev->surface_regs[bo->surface_reg];
370 i = bo->surface_reg;
371 goto out;
372 }
373
374 steal = -1;
375 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
376
377 reg = &rdev->surface_regs[i];
378 if (!reg->bo)
379 break;
380
381 old_object = reg->bo;
382 if (old_object->pin_count == 0)
383 steal = i;
384 }
385
386 /* if we are all out */
387 if (i == RADEON_GEM_MAX_SURFACES) {
388 if (steal == -1)
389 return -ENOMEM;
390 /* find someone with a surface reg and nuke their BO */
391 reg = &rdev->surface_regs[steal];
392 old_object = reg->bo;
393 /* blow away the mapping */
394 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
395 ttm_bo_unmap_virtual(&old_object->tbo);
396 old_object->surface_reg = -1;
397 i = steal;
398 }
399
400 bo->surface_reg = i;
401 reg->bo = bo;
402
403out:
404 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
405 bo->tbo.mem.start << PAGE_SHIFT,
406 bo->tbo.num_pages << PAGE_SHIFT);
407 return 0;
408}
409
410static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
411{
412 struct radeon_device *rdev = bo->rdev;
413 struct radeon_surface_reg *reg;
414
415 if (bo->surface_reg == -1)
416 return;
417
418 reg = &rdev->surface_regs[bo->surface_reg];
419 radeon_clear_surface_reg(rdev, bo->surface_reg);
420
421 reg->bo = NULL;
422 bo->surface_reg = -1;
423}
424
425int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
426 uint32_t tiling_flags, uint32_t pitch)
427{
428 int r;
429
430 r = radeon_bo_reserve(bo, false);
431 if (unlikely(r != 0))
432 return r;
433 bo->tiling_flags = tiling_flags;
434 bo->pitch = pitch;
435 radeon_bo_unreserve(bo);
436 return 0;
437}
438
439void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
440 uint32_t *tiling_flags,
441 uint32_t *pitch)
442{
443 BUG_ON(!atomic_read(&bo->tbo.reserved));
444 if (tiling_flags)
445 *tiling_flags = bo->tiling_flags;
446 if (pitch)
447 *pitch = bo->pitch;
448}
449
450int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
451 bool force_drop)
452{
453 BUG_ON(!atomic_read(&bo->tbo.reserved));
454
455 if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
456 return 0;
457
458 if (force_drop) {
459 radeon_bo_clear_surface_reg(bo);
460 return 0;
461 }
462
463 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
464 if (!has_moved)
465 return 0;
466
467 if (bo->surface_reg >= 0)
468 radeon_bo_clear_surface_reg(bo);
469 return 0;
470 }
471
472 if ((bo->surface_reg >= 0) && !has_moved)
473 return 0;
474
475 return radeon_bo_get_surface_reg(bo);
476}
477
478void radeon_bo_move_notify(struct ttm_buffer_object *bo,
479 struct ttm_mem_reg *mem)
480{
481 struct radeon_bo *rbo;
482 if (!radeon_ttm_bo_is_radeon_bo(bo))
483 return;
484 rbo = container_of(bo, struct radeon_bo, tbo);
485 radeon_bo_check_tiling(rbo, 0, 1);
486}
487
488int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
489{
490 struct radeon_device *rdev;
491 struct radeon_bo *rbo;
492 unsigned long offset, size;
493 int r;
494
495 if (!radeon_ttm_bo_is_radeon_bo(bo))
496 return 0;
497 rbo = container_of(bo, struct radeon_bo, tbo);
498 radeon_bo_check_tiling(rbo, 0, 0);
499 rdev = rbo->rdev;
500 if (bo->mem.mem_type == TTM_PL_VRAM) {
501 size = bo->mem.num_pages << PAGE_SHIFT;
502 offset = bo->mem.start << PAGE_SHIFT;
503 if ((offset + size) > rdev->mc.visible_vram_size) {
504 /* hurrah the memory is not visible ! */
505 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
506 rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
507 r = ttm_bo_validate(bo, &rbo->placement, false, true, false);
508 if (unlikely(r != 0))
509 return r;
510 offset = bo->mem.start << PAGE_SHIFT;
511 /* this should not happen */
512 if ((offset + size) > rdev->mc.visible_vram_size)
513 return -EINVAL;
514 }
515 }
516 return 0;
517}
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 <linux/slab.h>
34#include <drm/drmP.h>
35#include <drm/radeon_drm.h>
36#include "radeon.h"
37#include "radeon_trace.h"
38
39
40int radeon_ttm_init(struct radeon_device *rdev);
41void radeon_ttm_fini(struct radeon_device *rdev);
42static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
43
44/*
45 * To exclude mutual BO access we rely on bo_reserve exclusion, as all
46 * function are calling it.
47 */
48
49static void radeon_bo_clear_va(struct radeon_bo *bo)
50{
51 struct radeon_bo_va *bo_va, *tmp;
52
53 list_for_each_entry_safe(bo_va, tmp, &bo->va, bo_list) {
54 /* remove from all vm address space */
55 radeon_vm_bo_rmv(bo->rdev, bo_va);
56 }
57}
58
59static void radeon_update_memory_usage(struct radeon_bo *bo,
60 unsigned mem_type, int sign)
61{
62 struct radeon_device *rdev = bo->rdev;
63 u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
64
65 switch (mem_type) {
66 case TTM_PL_TT:
67 if (sign > 0)
68 atomic64_add(size, &rdev->gtt_usage);
69 else
70 atomic64_sub(size, &rdev->gtt_usage);
71 break;
72 case TTM_PL_VRAM:
73 if (sign > 0)
74 atomic64_add(size, &rdev->vram_usage);
75 else
76 atomic64_sub(size, &rdev->vram_usage);
77 break;
78 }
79}
80
81static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
82{
83 struct radeon_bo *bo;
84
85 bo = container_of(tbo, struct radeon_bo, tbo);
86
87 radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
88
89 mutex_lock(&bo->rdev->gem.mutex);
90 list_del_init(&bo->list);
91 mutex_unlock(&bo->rdev->gem.mutex);
92 radeon_bo_clear_surface_reg(bo);
93 radeon_bo_clear_va(bo);
94 drm_gem_object_release(&bo->gem_base);
95 kfree(bo);
96}
97
98bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
99{
100 if (bo->destroy == &radeon_ttm_bo_destroy)
101 return true;
102 return false;
103}
104
105void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
106{
107 u32 c = 0, i;
108
109 rbo->placement.fpfn = 0;
110 rbo->placement.lpfn = 0;
111 rbo->placement.placement = rbo->placements;
112 rbo->placement.busy_placement = rbo->placements;
113 if (domain & RADEON_GEM_DOMAIN_VRAM)
114 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
115 TTM_PL_FLAG_VRAM;
116 if (domain & RADEON_GEM_DOMAIN_GTT) {
117 if (rbo->rdev->flags & RADEON_IS_AGP) {
118 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_TT;
119 } else {
120 rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_TT;
121 }
122 }
123 if (domain & RADEON_GEM_DOMAIN_CPU) {
124 if (rbo->rdev->flags & RADEON_IS_AGP) {
125 rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_SYSTEM;
126 } else {
127 rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM;
128 }
129 }
130 if (!c)
131 rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
132 rbo->placement.num_placement = c;
133 rbo->placement.num_busy_placement = c;
134
135 /*
136 * Use two-ended allocation depending on the buffer size to
137 * improve fragmentation quality.
138 * 512kb was measured as the most optimal number.
139 */
140 if (rbo->tbo.mem.size > 512 * 1024) {
141 for (i = 0; i < c; i++) {
142 rbo->placements[i] |= TTM_PL_FLAG_TOPDOWN;
143 }
144 }
145}
146
147int radeon_bo_create(struct radeon_device *rdev,
148 unsigned long size, int byte_align, bool kernel, u32 domain,
149 struct sg_table *sg, struct radeon_bo **bo_ptr)
150{
151 struct radeon_bo *bo;
152 enum ttm_bo_type type;
153 unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
154 size_t acc_size;
155 int r;
156
157 size = ALIGN(size, PAGE_SIZE);
158
159 if (kernel) {
160 type = ttm_bo_type_kernel;
161 } else if (sg) {
162 type = ttm_bo_type_sg;
163 } else {
164 type = ttm_bo_type_device;
165 }
166 *bo_ptr = NULL;
167
168 acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
169 sizeof(struct radeon_bo));
170
171 bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
172 if (bo == NULL)
173 return -ENOMEM;
174 r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
175 if (unlikely(r)) {
176 kfree(bo);
177 return r;
178 }
179 bo->rdev = rdev;
180 bo->surface_reg = -1;
181 INIT_LIST_HEAD(&bo->list);
182 INIT_LIST_HEAD(&bo->va);
183 bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
184 RADEON_GEM_DOMAIN_GTT |
185 RADEON_GEM_DOMAIN_CPU);
186 radeon_ttm_placement_from_domain(bo, domain);
187 /* Kernel allocation are uninterruptible */
188 down_read(&rdev->pm.mclk_lock);
189 r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
190 &bo->placement, page_align, !kernel, NULL,
191 acc_size, sg, &radeon_ttm_bo_destroy);
192 up_read(&rdev->pm.mclk_lock);
193 if (unlikely(r != 0)) {
194 return r;
195 }
196 *bo_ptr = bo;
197
198 trace_radeon_bo_create(bo);
199
200 return 0;
201}
202
203int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
204{
205 bool is_iomem;
206 int r;
207
208 if (bo->kptr) {
209 if (ptr) {
210 *ptr = bo->kptr;
211 }
212 return 0;
213 }
214 r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
215 if (r) {
216 return r;
217 }
218 bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
219 if (ptr) {
220 *ptr = bo->kptr;
221 }
222 radeon_bo_check_tiling(bo, 0, 0);
223 return 0;
224}
225
226void radeon_bo_kunmap(struct radeon_bo *bo)
227{
228 if (bo->kptr == NULL)
229 return;
230 bo->kptr = NULL;
231 radeon_bo_check_tiling(bo, 0, 0);
232 ttm_bo_kunmap(&bo->kmap);
233}
234
235void radeon_bo_unref(struct radeon_bo **bo)
236{
237 struct ttm_buffer_object *tbo;
238 struct radeon_device *rdev;
239
240 if ((*bo) == NULL)
241 return;
242 rdev = (*bo)->rdev;
243 tbo = &((*bo)->tbo);
244 down_read(&rdev->pm.mclk_lock);
245 ttm_bo_unref(&tbo);
246 up_read(&rdev->pm.mclk_lock);
247 if (tbo == NULL)
248 *bo = NULL;
249}
250
251int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
252 u64 *gpu_addr)
253{
254 int r, i;
255
256 if (bo->pin_count) {
257 bo->pin_count++;
258 if (gpu_addr)
259 *gpu_addr = radeon_bo_gpu_offset(bo);
260
261 if (max_offset != 0) {
262 u64 domain_start;
263
264 if (domain == RADEON_GEM_DOMAIN_VRAM)
265 domain_start = bo->rdev->mc.vram_start;
266 else
267 domain_start = bo->rdev->mc.gtt_start;
268 WARN_ON_ONCE(max_offset <
269 (radeon_bo_gpu_offset(bo) - domain_start));
270 }
271
272 return 0;
273 }
274 radeon_ttm_placement_from_domain(bo, domain);
275 if (domain == RADEON_GEM_DOMAIN_VRAM) {
276 /* force to pin into visible video ram */
277 bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
278 }
279 if (max_offset) {
280 u64 lpfn = max_offset >> PAGE_SHIFT;
281
282 if (!bo->placement.lpfn)
283 bo->placement.lpfn = bo->rdev->mc.gtt_size >> PAGE_SHIFT;
284
285 if (lpfn < bo->placement.lpfn)
286 bo->placement.lpfn = lpfn;
287 }
288 for (i = 0; i < bo->placement.num_placement; i++)
289 bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
290 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
291 if (likely(r == 0)) {
292 bo->pin_count = 1;
293 if (gpu_addr != NULL)
294 *gpu_addr = radeon_bo_gpu_offset(bo);
295 }
296 if (unlikely(r != 0))
297 dev_err(bo->rdev->dev, "%p pin failed\n", bo);
298 return r;
299}
300
301int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
302{
303 return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
304}
305
306int radeon_bo_unpin(struct radeon_bo *bo)
307{
308 int r, i;
309
310 if (!bo->pin_count) {
311 dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
312 return 0;
313 }
314 bo->pin_count--;
315 if (bo->pin_count)
316 return 0;
317 for (i = 0; i < bo->placement.num_placement; i++)
318 bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
319 r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
320 if (unlikely(r != 0))
321 dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
322 return r;
323}
324
325int radeon_bo_evict_vram(struct radeon_device *rdev)
326{
327 /* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
328 if (0 && (rdev->flags & RADEON_IS_IGP)) {
329 if (rdev->mc.igp_sideport_enabled == false)
330 /* Useless to evict on IGP chips */
331 return 0;
332 }
333 return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
334}
335
336void radeon_bo_force_delete(struct radeon_device *rdev)
337{
338 struct radeon_bo *bo, *n;
339
340 if (list_empty(&rdev->gem.objects)) {
341 return;
342 }
343 dev_err(rdev->dev, "Userspace still has active objects !\n");
344 list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
345 mutex_lock(&rdev->ddev->struct_mutex);
346 dev_err(rdev->dev, "%p %p %lu %lu force free\n",
347 &bo->gem_base, bo, (unsigned long)bo->gem_base.size,
348 *((unsigned long *)&bo->gem_base.refcount));
349 mutex_lock(&bo->rdev->gem.mutex);
350 list_del_init(&bo->list);
351 mutex_unlock(&bo->rdev->gem.mutex);
352 /* this should unref the ttm bo */
353 drm_gem_object_unreference(&bo->gem_base);
354 mutex_unlock(&rdev->ddev->struct_mutex);
355 }
356}
357
358int radeon_bo_init(struct radeon_device *rdev)
359{
360 /* Add an MTRR for the VRAM */
361 if (!rdev->fastfb_working) {
362 rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
363 rdev->mc.aper_size);
364 }
365 DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
366 rdev->mc.mc_vram_size >> 20,
367 (unsigned long long)rdev->mc.aper_size >> 20);
368 DRM_INFO("RAM width %dbits %cDR\n",
369 rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
370 return radeon_ttm_init(rdev);
371}
372
373void radeon_bo_fini(struct radeon_device *rdev)
374{
375 radeon_ttm_fini(rdev);
376 arch_phys_wc_del(rdev->mc.vram_mtrr);
377}
378
379/* Returns how many bytes TTM can move per IB.
380 */
381static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
382{
383 u64 real_vram_size = rdev->mc.real_vram_size;
384 u64 vram_usage = atomic64_read(&rdev->vram_usage);
385
386 /* This function is based on the current VRAM usage.
387 *
388 * - If all of VRAM is free, allow relocating the number of bytes that
389 * is equal to 1/4 of the size of VRAM for this IB.
390
391 * - If more than one half of VRAM is occupied, only allow relocating
392 * 1 MB of data for this IB.
393 *
394 * - From 0 to one half of used VRAM, the threshold decreases
395 * linearly.
396 * __________________
397 * 1/4 of -|\ |
398 * VRAM | \ |
399 * | \ |
400 * | \ |
401 * | \ |
402 * | \ |
403 * | \ |
404 * | \________|1 MB
405 * |----------------|
406 * VRAM 0 % 100 %
407 * used used
408 *
409 * Note: It's a threshold, not a limit. The threshold must be crossed
410 * for buffer relocations to stop, so any buffer of an arbitrary size
411 * can be moved as long as the threshold isn't crossed before
412 * the relocation takes place. We don't want to disable buffer
413 * relocations completely.
414 *
415 * The idea is that buffers should be placed in VRAM at creation time
416 * and TTM should only do a minimum number of relocations during
417 * command submission. In practice, you need to submit at least
418 * a dozen IBs to move all buffers to VRAM if they are in GTT.
419 *
420 * Also, things can get pretty crazy under memory pressure and actual
421 * VRAM usage can change a lot, so playing safe even at 50% does
422 * consistently increase performance.
423 */
424
425 u64 half_vram = real_vram_size >> 1;
426 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
427 u64 bytes_moved_threshold = half_free_vram >> 1;
428 return max(bytes_moved_threshold, 1024*1024ull);
429}
430
431int radeon_bo_list_validate(struct radeon_device *rdev,
432 struct ww_acquire_ctx *ticket,
433 struct list_head *head, int ring)
434{
435 struct radeon_cs_reloc *lobj;
436 struct radeon_bo *bo;
437 int r;
438 u64 bytes_moved = 0, initial_bytes_moved;
439 u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
440
441 r = ttm_eu_reserve_buffers(ticket, head);
442 if (unlikely(r != 0)) {
443 return r;
444 }
445
446 list_for_each_entry(lobj, head, tv.head) {
447 bo = lobj->robj;
448 if (!bo->pin_count) {
449 u32 domain = lobj->domain;
450 u32 current_domain =
451 radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
452
453 /* Check if this buffer will be moved and don't move it
454 * if we have moved too many buffers for this IB already.
455 *
456 * Note that this allows moving at least one buffer of
457 * any size, because it doesn't take the current "bo"
458 * into account. We don't want to disallow buffer moves
459 * completely.
460 */
461 if ((lobj->alt_domain & current_domain) != 0 &&
462 (domain & current_domain) == 0 && /* will be moved */
463 bytes_moved > bytes_moved_threshold) {
464 /* don't move it */
465 domain = current_domain;
466 }
467
468 retry:
469 radeon_ttm_placement_from_domain(bo, domain);
470 if (ring == R600_RING_TYPE_UVD_INDEX)
471 radeon_uvd_force_into_uvd_segment(bo);
472
473 initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
474 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
475 bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
476 initial_bytes_moved;
477
478 if (unlikely(r)) {
479 if (r != -ERESTARTSYS && domain != lobj->alt_domain) {
480 domain = lobj->alt_domain;
481 goto retry;
482 }
483 ttm_eu_backoff_reservation(ticket, head);
484 return r;
485 }
486 }
487 lobj->gpu_offset = radeon_bo_gpu_offset(bo);
488 lobj->tiling_flags = bo->tiling_flags;
489 }
490 return 0;
491}
492
493int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
494 struct vm_area_struct *vma)
495{
496 return ttm_fbdev_mmap(vma, &bo->tbo);
497}
498
499int radeon_bo_get_surface_reg(struct radeon_bo *bo)
500{
501 struct radeon_device *rdev = bo->rdev;
502 struct radeon_surface_reg *reg;
503 struct radeon_bo *old_object;
504 int steal;
505 int i;
506
507 lockdep_assert_held(&bo->tbo.resv->lock.base);
508
509 if (!bo->tiling_flags)
510 return 0;
511
512 if (bo->surface_reg >= 0) {
513 reg = &rdev->surface_regs[bo->surface_reg];
514 i = bo->surface_reg;
515 goto out;
516 }
517
518 steal = -1;
519 for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
520
521 reg = &rdev->surface_regs[i];
522 if (!reg->bo)
523 break;
524
525 old_object = reg->bo;
526 if (old_object->pin_count == 0)
527 steal = i;
528 }
529
530 /* if we are all out */
531 if (i == RADEON_GEM_MAX_SURFACES) {
532 if (steal == -1)
533 return -ENOMEM;
534 /* find someone with a surface reg and nuke their BO */
535 reg = &rdev->surface_regs[steal];
536 old_object = reg->bo;
537 /* blow away the mapping */
538 DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
539 ttm_bo_unmap_virtual(&old_object->tbo);
540 old_object->surface_reg = -1;
541 i = steal;
542 }
543
544 bo->surface_reg = i;
545 reg->bo = bo;
546
547out:
548 radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
549 bo->tbo.mem.start << PAGE_SHIFT,
550 bo->tbo.num_pages << PAGE_SHIFT);
551 return 0;
552}
553
554static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
555{
556 struct radeon_device *rdev = bo->rdev;
557 struct radeon_surface_reg *reg;
558
559 if (bo->surface_reg == -1)
560 return;
561
562 reg = &rdev->surface_regs[bo->surface_reg];
563 radeon_clear_surface_reg(rdev, bo->surface_reg);
564
565 reg->bo = NULL;
566 bo->surface_reg = -1;
567}
568
569int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
570 uint32_t tiling_flags, uint32_t pitch)
571{
572 struct radeon_device *rdev = bo->rdev;
573 int r;
574
575 if (rdev->family >= CHIP_CEDAR) {
576 unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
577
578 bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
579 bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
580 mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
581 tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
582 stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
583 switch (bankw) {
584 case 0:
585 case 1:
586 case 2:
587 case 4:
588 case 8:
589 break;
590 default:
591 return -EINVAL;
592 }
593 switch (bankh) {
594 case 0:
595 case 1:
596 case 2:
597 case 4:
598 case 8:
599 break;
600 default:
601 return -EINVAL;
602 }
603 switch (mtaspect) {
604 case 0:
605 case 1:
606 case 2:
607 case 4:
608 case 8:
609 break;
610 default:
611 return -EINVAL;
612 }
613 if (tilesplit > 6) {
614 return -EINVAL;
615 }
616 if (stilesplit > 6) {
617 return -EINVAL;
618 }
619 }
620 r = radeon_bo_reserve(bo, false);
621 if (unlikely(r != 0))
622 return r;
623 bo->tiling_flags = tiling_flags;
624 bo->pitch = pitch;
625 radeon_bo_unreserve(bo);
626 return 0;
627}
628
629void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
630 uint32_t *tiling_flags,
631 uint32_t *pitch)
632{
633 lockdep_assert_held(&bo->tbo.resv->lock.base);
634
635 if (tiling_flags)
636 *tiling_flags = bo->tiling_flags;
637 if (pitch)
638 *pitch = bo->pitch;
639}
640
641int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
642 bool force_drop)
643{
644 if (!force_drop)
645 lockdep_assert_held(&bo->tbo.resv->lock.base);
646
647 if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
648 return 0;
649
650 if (force_drop) {
651 radeon_bo_clear_surface_reg(bo);
652 return 0;
653 }
654
655 if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
656 if (!has_moved)
657 return 0;
658
659 if (bo->surface_reg >= 0)
660 radeon_bo_clear_surface_reg(bo);
661 return 0;
662 }
663
664 if ((bo->surface_reg >= 0) && !has_moved)
665 return 0;
666
667 return radeon_bo_get_surface_reg(bo);
668}
669
670void radeon_bo_move_notify(struct ttm_buffer_object *bo,
671 struct ttm_mem_reg *new_mem)
672{
673 struct radeon_bo *rbo;
674
675 if (!radeon_ttm_bo_is_radeon_bo(bo))
676 return;
677
678 rbo = container_of(bo, struct radeon_bo, tbo);
679 radeon_bo_check_tiling(rbo, 0, 1);
680 radeon_vm_bo_invalidate(rbo->rdev, rbo);
681
682 /* update statistics */
683 if (!new_mem)
684 return;
685
686 radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
687 radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
688}
689
690int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
691{
692 struct radeon_device *rdev;
693 struct radeon_bo *rbo;
694 unsigned long offset, size;
695 int r;
696
697 if (!radeon_ttm_bo_is_radeon_bo(bo))
698 return 0;
699 rbo = container_of(bo, struct radeon_bo, tbo);
700 radeon_bo_check_tiling(rbo, 0, 0);
701 rdev = rbo->rdev;
702 if (bo->mem.mem_type != TTM_PL_VRAM)
703 return 0;
704
705 size = bo->mem.num_pages << PAGE_SHIFT;
706 offset = bo->mem.start << PAGE_SHIFT;
707 if ((offset + size) <= rdev->mc.visible_vram_size)
708 return 0;
709
710 /* hurrah the memory is not visible ! */
711 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
712 rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
713 r = ttm_bo_validate(bo, &rbo->placement, false, false);
714 if (unlikely(r == -ENOMEM)) {
715 radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
716 return ttm_bo_validate(bo, &rbo->placement, false, false);
717 } else if (unlikely(r != 0)) {
718 return r;
719 }
720
721 offset = bo->mem.start << PAGE_SHIFT;
722 /* this should never happen */
723 if ((offset + size) > rdev->mc.visible_vram_size)
724 return -EINVAL;
725
726 return 0;
727}
728
729int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
730{
731 int r;
732
733 r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
734 if (unlikely(r != 0))
735 return r;
736 spin_lock(&bo->tbo.bdev->fence_lock);
737 if (mem_type)
738 *mem_type = bo->tbo.mem.mem_type;
739 if (bo->tbo.sync_obj)
740 r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
741 spin_unlock(&bo->tbo.bdev->fence_lock);
742 ttm_bo_unreserve(&bo->tbo);
743 return r;
744}