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 <drm/drm_cache.h>
37 #include "radeon.h"
38 #include "radeon_trace.h"
39 
40 
41 int radeon_ttm_init(struct radeon_device *rdev);
42 void radeon_ttm_fini(struct radeon_device *rdev);
43 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
44 
45 /*
46  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
47  * function are calling it.
48  */
49 
50 static void radeon_update_memory_usage(struct radeon_bo *bo,
51 				       unsigned mem_type, int sign)
52 {
53 	struct radeon_device *rdev = bo->rdev;
54 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
55 
56 	switch (mem_type) {
57 	case TTM_PL_TT:
58 		if (sign > 0)
59 			atomic64_add(size, &rdev->gtt_usage);
60 		else
61 			atomic64_sub(size, &rdev->gtt_usage);
62 		break;
63 	case TTM_PL_VRAM:
64 		if (sign > 0)
65 			atomic64_add(size, &rdev->vram_usage);
66 		else
67 			atomic64_sub(size, &rdev->vram_usage);
68 		break;
69 	}
70 }
71 
72 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
73 {
74 	struct radeon_bo *bo;
75 
76 	bo = container_of(tbo, struct radeon_bo, tbo);
77 
78 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
79 
80 	mutex_lock(&bo->rdev->gem.mutex);
81 	list_del_init(&bo->list);
82 	mutex_unlock(&bo->rdev->gem.mutex);
83 	radeon_bo_clear_surface_reg(bo);
84 	WARN_ON(!list_empty(&bo->va));
85 	drm_gem_object_release(&bo->gem_base);
86 	kfree(bo);
87 }
88 
89 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
90 {
91 	if (bo->destroy == &radeon_ttm_bo_destroy)
92 		return true;
93 	return false;
94 }
95 
96 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
97 {
98 	u32 c = 0, i;
99 
100 	rbo->placement.placement = rbo->placements;
101 	rbo->placement.busy_placement = rbo->placements;
102 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
103 		/* Try placing BOs which don't need CPU access outside of the
104 		 * CPU accessible part of VRAM
105 		 */
106 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
107 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
108 			rbo->placements[c].fpfn =
109 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
110 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
111 						     TTM_PL_FLAG_UNCACHED |
112 						     TTM_PL_FLAG_VRAM;
113 		}
114 
115 		rbo->placements[c].fpfn = 0;
116 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
117 					     TTM_PL_FLAG_UNCACHED |
118 					     TTM_PL_FLAG_VRAM;
119 	}
120 
121 	if (domain & RADEON_GEM_DOMAIN_GTT) {
122 		if (rbo->flags & RADEON_GEM_GTT_UC) {
123 			rbo->placements[c].fpfn = 0;
124 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
125 				TTM_PL_FLAG_TT;
126 
127 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
128 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
129 			rbo->placements[c].fpfn = 0;
130 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
131 				TTM_PL_FLAG_UNCACHED |
132 				TTM_PL_FLAG_TT;
133 		} else {
134 			rbo->placements[c].fpfn = 0;
135 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
136 						     TTM_PL_FLAG_TT;
137 		}
138 	}
139 
140 	if (domain & RADEON_GEM_DOMAIN_CPU) {
141 		if (rbo->flags & RADEON_GEM_GTT_UC) {
142 			rbo->placements[c].fpfn = 0;
143 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
144 				TTM_PL_FLAG_SYSTEM;
145 
146 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
147 		    rbo->rdev->flags & RADEON_IS_AGP) {
148 			rbo->placements[c].fpfn = 0;
149 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
150 				TTM_PL_FLAG_UNCACHED |
151 				TTM_PL_FLAG_SYSTEM;
152 		} else {
153 			rbo->placements[c].fpfn = 0;
154 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
155 						     TTM_PL_FLAG_SYSTEM;
156 		}
157 	}
158 	if (!c) {
159 		rbo->placements[c].fpfn = 0;
160 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
161 					     TTM_PL_FLAG_SYSTEM;
162 	}
163 
164 	rbo->placement.num_placement = c;
165 	rbo->placement.num_busy_placement = c;
166 
167 	for (i = 0; i < c; ++i) {
168 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
169 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
170 		    !rbo->placements[i].fpfn)
171 			rbo->placements[i].lpfn =
172 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
173 		else
174 			rbo->placements[i].lpfn = 0;
175 	}
176 }
177 
178 int radeon_bo_create(struct radeon_device *rdev,
179 		     unsigned long size, int byte_align, bool kernel,
180 		     u32 domain, u32 flags, struct sg_table *sg,
181 		     struct reservation_object *resv,
182 		     struct radeon_bo **bo_ptr)
183 {
184 	struct radeon_bo *bo;
185 	enum ttm_bo_type type;
186 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
187 	size_t acc_size;
188 	int r;
189 
190 	size = ALIGN(size, PAGE_SIZE);
191 
192 	if (kernel) {
193 		type = ttm_bo_type_kernel;
194 	} else if (sg) {
195 		type = ttm_bo_type_sg;
196 	} else {
197 		type = ttm_bo_type_device;
198 	}
199 	*bo_ptr = NULL;
200 
201 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
202 				       sizeof(struct radeon_bo));
203 
204 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
205 	if (bo == NULL)
206 		return -ENOMEM;
207 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
208 	if (unlikely(r)) {
209 		kfree(bo);
210 		return r;
211 	}
212 	bo->rdev = rdev;
213 	bo->surface_reg = -1;
214 	INIT_LIST_HEAD(&bo->list);
215 	INIT_LIST_HEAD(&bo->va);
216 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
217 				       RADEON_GEM_DOMAIN_GTT |
218 				       RADEON_GEM_DOMAIN_CPU);
219 
220 	bo->flags = flags;
221 	/* PCI GART is always snooped */
222 	if (!(rdev->flags & RADEON_IS_PCIE))
223 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
224 
225 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
226 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
227 	 */
228 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
229 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
230 
231 #ifdef CONFIG_X86_32
232 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
233 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
234 	 */
235 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
236 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
237 	/* Don't try to enable write-combining when it can't work, or things
238 	 * may be slow
239 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
240 	 */
241 
242 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
243 	 thanks to write-combining
244 
245 	if (bo->flags & RADEON_GEM_GTT_WC)
246 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
247 			      "better performance thanks to write-combining\n");
248 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
249 #else
250 	/* For architectures that don't support WC memory,
251 	 * mask out the WC flag from the BO
252 	 */
253 	if (!drm_arch_can_wc_memory())
254 		bo->flags &= ~RADEON_GEM_GTT_WC;
255 #endif
256 
257 	radeon_ttm_placement_from_domain(bo, domain);
258 	/* Kernel allocation are uninterruptible */
259 	down_read(&rdev->pm.mclk_lock);
260 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
261 			&bo->placement, page_align, !kernel, NULL,
262 			acc_size, sg, resv, &radeon_ttm_bo_destroy);
263 	up_read(&rdev->pm.mclk_lock);
264 	if (unlikely(r != 0)) {
265 		return r;
266 	}
267 	*bo_ptr = bo;
268 
269 	trace_radeon_bo_create(bo);
270 
271 	return 0;
272 }
273 
274 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
275 {
276 	bool is_iomem;
277 	int r;
278 
279 	if (bo->kptr) {
280 		if (ptr) {
281 			*ptr = bo->kptr;
282 		}
283 		return 0;
284 	}
285 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
286 	if (r) {
287 		return r;
288 	}
289 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
290 	if (ptr) {
291 		*ptr = bo->kptr;
292 	}
293 	radeon_bo_check_tiling(bo, 0, 0);
294 	return 0;
295 }
296 
297 void radeon_bo_kunmap(struct radeon_bo *bo)
298 {
299 	if (bo->kptr == NULL)
300 		return;
301 	bo->kptr = NULL;
302 	radeon_bo_check_tiling(bo, 0, 0);
303 	ttm_bo_kunmap(&bo->kmap);
304 }
305 
306 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
307 {
308 	if (bo == NULL)
309 		return NULL;
310 
311 	ttm_bo_reference(&bo->tbo);
312 	return bo;
313 }
314 
315 void radeon_bo_unref(struct radeon_bo **bo)
316 {
317 	struct ttm_buffer_object *tbo;
318 	struct radeon_device *rdev;
319 
320 	if ((*bo) == NULL)
321 		return;
322 	rdev = (*bo)->rdev;
323 	tbo = &((*bo)->tbo);
324 	ttm_bo_unref(&tbo);
325 	if (tbo == NULL)
326 		*bo = NULL;
327 }
328 
329 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
330 			     u64 *gpu_addr)
331 {
332 	int r, i;
333 
334 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
335 		return -EPERM;
336 
337 	if (bo->pin_count) {
338 		bo->pin_count++;
339 		if (gpu_addr)
340 			*gpu_addr = radeon_bo_gpu_offset(bo);
341 
342 		if (max_offset != 0) {
343 			u64 domain_start;
344 
345 			if (domain == RADEON_GEM_DOMAIN_VRAM)
346 				domain_start = bo->rdev->mc.vram_start;
347 			else
348 				domain_start = bo->rdev->mc.gtt_start;
349 			WARN_ON_ONCE(max_offset <
350 				     (radeon_bo_gpu_offset(bo) - domain_start));
351 		}
352 
353 		return 0;
354 	}
355 	radeon_ttm_placement_from_domain(bo, domain);
356 	for (i = 0; i < bo->placement.num_placement; i++) {
357 		/* force to pin into visible video ram */
358 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
359 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
360 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
361 			bo->placements[i].lpfn =
362 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
363 		else
364 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
365 
366 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
367 	}
368 
369 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
370 	if (likely(r == 0)) {
371 		bo->pin_count = 1;
372 		if (gpu_addr != NULL)
373 			*gpu_addr = radeon_bo_gpu_offset(bo);
374 		if (domain == RADEON_GEM_DOMAIN_VRAM)
375 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
376 		else
377 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
378 	} else {
379 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
380 	}
381 	return r;
382 }
383 
384 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
385 {
386 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
387 }
388 
389 int radeon_bo_unpin(struct radeon_bo *bo)
390 {
391 	int r, i;
392 
393 	if (!bo->pin_count) {
394 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
395 		return 0;
396 	}
397 	bo->pin_count--;
398 	if (bo->pin_count)
399 		return 0;
400 	for (i = 0; i < bo->placement.num_placement; i++) {
401 		bo->placements[i].lpfn = 0;
402 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
403 	}
404 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
405 	if (likely(r == 0)) {
406 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
407 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
408 		else
409 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
410 	} else {
411 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
412 	}
413 	return r;
414 }
415 
416 int radeon_bo_evict_vram(struct radeon_device *rdev)
417 {
418 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
419 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
420 		if (rdev->mc.igp_sideport_enabled == false)
421 			/* Useless to evict on IGP chips */
422 			return 0;
423 	}
424 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
425 }
426 
427 void radeon_bo_force_delete(struct radeon_device *rdev)
428 {
429 	struct radeon_bo *bo, *n;
430 
431 	if (list_empty(&rdev->gem.objects)) {
432 		return;
433 	}
434 	dev_err(rdev->dev, "Userspace still has active objects !\n");
435 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
436 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
437 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
438 			*((unsigned long *)&bo->gem_base.refcount));
439 		mutex_lock(&bo->rdev->gem.mutex);
440 		list_del_init(&bo->list);
441 		mutex_unlock(&bo->rdev->gem.mutex);
442 		/* this should unref the ttm bo */
443 		drm_gem_object_unreference_unlocked(&bo->gem_base);
444 	}
445 }
446 
447 int radeon_bo_init(struct radeon_device *rdev)
448 {
449 	/* Add an MTRR for the VRAM */
450 	if (!rdev->fastfb_working) {
451 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
452 						      rdev->mc.aper_size);
453 	}
454 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
455 		rdev->mc.mc_vram_size >> 20,
456 		(unsigned long long)rdev->mc.aper_size >> 20);
457 	DRM_INFO("RAM width %dbits %cDR\n",
458 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
459 	return radeon_ttm_init(rdev);
460 }
461 
462 void radeon_bo_fini(struct radeon_device *rdev)
463 {
464 	radeon_ttm_fini(rdev);
465 	arch_phys_wc_del(rdev->mc.vram_mtrr);
466 }
467 
468 /* Returns how many bytes TTM can move per IB.
469  */
470 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
471 {
472 	u64 real_vram_size = rdev->mc.real_vram_size;
473 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
474 
475 	/* This function is based on the current VRAM usage.
476 	 *
477 	 * - If all of VRAM is free, allow relocating the number of bytes that
478 	 *   is equal to 1/4 of the size of VRAM for this IB.
479 
480 	 * - If more than one half of VRAM is occupied, only allow relocating
481 	 *   1 MB of data for this IB.
482 	 *
483 	 * - From 0 to one half of used VRAM, the threshold decreases
484 	 *   linearly.
485 	 *         __________________
486 	 * 1/4 of -|\               |
487 	 * VRAM    | \              |
488 	 *         |  \             |
489 	 *         |   \            |
490 	 *         |    \           |
491 	 *         |     \          |
492 	 *         |      \         |
493 	 *         |       \________|1 MB
494 	 *         |----------------|
495 	 *    VRAM 0 %             100 %
496 	 *         used            used
497 	 *
498 	 * Note: It's a threshold, not a limit. The threshold must be crossed
499 	 * for buffer relocations to stop, so any buffer of an arbitrary size
500 	 * can be moved as long as the threshold isn't crossed before
501 	 * the relocation takes place. We don't want to disable buffer
502 	 * relocations completely.
503 	 *
504 	 * The idea is that buffers should be placed in VRAM at creation time
505 	 * and TTM should only do a minimum number of relocations during
506 	 * command submission. In practice, you need to submit at least
507 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
508 	 *
509 	 * Also, things can get pretty crazy under memory pressure and actual
510 	 * VRAM usage can change a lot, so playing safe even at 50% does
511 	 * consistently increase performance.
512 	 */
513 
514 	u64 half_vram = real_vram_size >> 1;
515 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
516 	u64 bytes_moved_threshold = half_free_vram >> 1;
517 	return max(bytes_moved_threshold, 1024*1024ull);
518 }
519 
520 int radeon_bo_list_validate(struct radeon_device *rdev,
521 			    struct ww_acquire_ctx *ticket,
522 			    struct list_head *head, int ring)
523 {
524 	struct radeon_bo_list *lobj;
525 	struct list_head duplicates;
526 	int r;
527 	u64 bytes_moved = 0, initial_bytes_moved;
528 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
529 
530 	INIT_LIST_HEAD(&duplicates);
531 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
532 	if (unlikely(r != 0)) {
533 		return r;
534 	}
535 
536 	list_for_each_entry(lobj, head, tv.head) {
537 		struct radeon_bo *bo = lobj->robj;
538 		if (!bo->pin_count) {
539 			u32 domain = lobj->prefered_domains;
540 			u32 allowed = lobj->allowed_domains;
541 			u32 current_domain =
542 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
543 
544 			/* Check if this buffer will be moved and don't move it
545 			 * if we have moved too many buffers for this IB already.
546 			 *
547 			 * Note that this allows moving at least one buffer of
548 			 * any size, because it doesn't take the current "bo"
549 			 * into account. We don't want to disallow buffer moves
550 			 * completely.
551 			 */
552 			if ((allowed & current_domain) != 0 &&
553 			    (domain & current_domain) == 0 && /* will be moved */
554 			    bytes_moved > bytes_moved_threshold) {
555 				/* don't move it */
556 				domain = current_domain;
557 			}
558 
559 		retry:
560 			radeon_ttm_placement_from_domain(bo, domain);
561 			if (ring == R600_RING_TYPE_UVD_INDEX)
562 				radeon_uvd_force_into_uvd_segment(bo, allowed);
563 
564 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
565 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
566 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
567 				       initial_bytes_moved;
568 
569 			if (unlikely(r)) {
570 				if (r != -ERESTARTSYS &&
571 				    domain != lobj->allowed_domains) {
572 					domain = lobj->allowed_domains;
573 					goto retry;
574 				}
575 				ttm_eu_backoff_reservation(ticket, head);
576 				return r;
577 			}
578 		}
579 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
580 		lobj->tiling_flags = bo->tiling_flags;
581 	}
582 
583 	list_for_each_entry(lobj, &duplicates, tv.head) {
584 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
585 		lobj->tiling_flags = lobj->robj->tiling_flags;
586 	}
587 
588 	return 0;
589 }
590 
591 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
592 {
593 	struct radeon_device *rdev = bo->rdev;
594 	struct radeon_surface_reg *reg;
595 	struct radeon_bo *old_object;
596 	int steal;
597 	int i;
598 
599 	lockdep_assert_held(&bo->tbo.resv->lock.base);
600 
601 	if (!bo->tiling_flags)
602 		return 0;
603 
604 	if (bo->surface_reg >= 0) {
605 		reg = &rdev->surface_regs[bo->surface_reg];
606 		i = bo->surface_reg;
607 		goto out;
608 	}
609 
610 	steal = -1;
611 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
612 
613 		reg = &rdev->surface_regs[i];
614 		if (!reg->bo)
615 			break;
616 
617 		old_object = reg->bo;
618 		if (old_object->pin_count == 0)
619 			steal = i;
620 	}
621 
622 	/* if we are all out */
623 	if (i == RADEON_GEM_MAX_SURFACES) {
624 		if (steal == -1)
625 			return -ENOMEM;
626 		/* find someone with a surface reg and nuke their BO */
627 		reg = &rdev->surface_regs[steal];
628 		old_object = reg->bo;
629 		/* blow away the mapping */
630 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
631 		ttm_bo_unmap_virtual(&old_object->tbo);
632 		old_object->surface_reg = -1;
633 		i = steal;
634 	}
635 
636 	bo->surface_reg = i;
637 	reg->bo = bo;
638 
639 out:
640 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
641 			       bo->tbo.mem.start << PAGE_SHIFT,
642 			       bo->tbo.num_pages << PAGE_SHIFT);
643 	return 0;
644 }
645 
646 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
647 {
648 	struct radeon_device *rdev = bo->rdev;
649 	struct radeon_surface_reg *reg;
650 
651 	if (bo->surface_reg == -1)
652 		return;
653 
654 	reg = &rdev->surface_regs[bo->surface_reg];
655 	radeon_clear_surface_reg(rdev, bo->surface_reg);
656 
657 	reg->bo = NULL;
658 	bo->surface_reg = -1;
659 }
660 
661 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
662 				uint32_t tiling_flags, uint32_t pitch)
663 {
664 	struct radeon_device *rdev = bo->rdev;
665 	int r;
666 
667 	if (rdev->family >= CHIP_CEDAR) {
668 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
669 
670 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
671 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
672 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
673 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
674 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
675 		switch (bankw) {
676 		case 0:
677 		case 1:
678 		case 2:
679 		case 4:
680 		case 8:
681 			break;
682 		default:
683 			return -EINVAL;
684 		}
685 		switch (bankh) {
686 		case 0:
687 		case 1:
688 		case 2:
689 		case 4:
690 		case 8:
691 			break;
692 		default:
693 			return -EINVAL;
694 		}
695 		switch (mtaspect) {
696 		case 0:
697 		case 1:
698 		case 2:
699 		case 4:
700 		case 8:
701 			break;
702 		default:
703 			return -EINVAL;
704 		}
705 		if (tilesplit > 6) {
706 			return -EINVAL;
707 		}
708 		if (stilesplit > 6) {
709 			return -EINVAL;
710 		}
711 	}
712 	r = radeon_bo_reserve(bo, false);
713 	if (unlikely(r != 0))
714 		return r;
715 	bo->tiling_flags = tiling_flags;
716 	bo->pitch = pitch;
717 	radeon_bo_unreserve(bo);
718 	return 0;
719 }
720 
721 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
722 				uint32_t *tiling_flags,
723 				uint32_t *pitch)
724 {
725 	lockdep_assert_held(&bo->tbo.resv->lock.base);
726 
727 	if (tiling_flags)
728 		*tiling_flags = bo->tiling_flags;
729 	if (pitch)
730 		*pitch = bo->pitch;
731 }
732 
733 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
734 				bool force_drop)
735 {
736 	if (!force_drop)
737 		lockdep_assert_held(&bo->tbo.resv->lock.base);
738 
739 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
740 		return 0;
741 
742 	if (force_drop) {
743 		radeon_bo_clear_surface_reg(bo);
744 		return 0;
745 	}
746 
747 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
748 		if (!has_moved)
749 			return 0;
750 
751 		if (bo->surface_reg >= 0)
752 			radeon_bo_clear_surface_reg(bo);
753 		return 0;
754 	}
755 
756 	if ((bo->surface_reg >= 0) && !has_moved)
757 		return 0;
758 
759 	return radeon_bo_get_surface_reg(bo);
760 }
761 
762 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
763 			   struct ttm_mem_reg *new_mem)
764 {
765 	struct radeon_bo *rbo;
766 
767 	if (!radeon_ttm_bo_is_radeon_bo(bo))
768 		return;
769 
770 	rbo = container_of(bo, struct radeon_bo, tbo);
771 	radeon_bo_check_tiling(rbo, 0, 1);
772 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
773 
774 	/* update statistics */
775 	if (!new_mem)
776 		return;
777 
778 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
779 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
780 }
781 
782 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
783 {
784 	struct radeon_device *rdev;
785 	struct radeon_bo *rbo;
786 	unsigned long offset, size, lpfn;
787 	int i, r;
788 
789 	if (!radeon_ttm_bo_is_radeon_bo(bo))
790 		return 0;
791 	rbo = container_of(bo, struct radeon_bo, tbo);
792 	radeon_bo_check_tiling(rbo, 0, 0);
793 	rdev = rbo->rdev;
794 	if (bo->mem.mem_type != TTM_PL_VRAM)
795 		return 0;
796 
797 	size = bo->mem.num_pages << PAGE_SHIFT;
798 	offset = bo->mem.start << PAGE_SHIFT;
799 	if ((offset + size) <= rdev->mc.visible_vram_size)
800 		return 0;
801 
802 	/* Can't move a pinned BO to visible VRAM */
803 	if (rbo->pin_count > 0)
804 		return -EINVAL;
805 
806 	/* hurrah the memory is not visible ! */
807 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
808 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
809 	for (i = 0; i < rbo->placement.num_placement; i++) {
810 		/* Force into visible VRAM */
811 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
812 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
813 			rbo->placements[i].lpfn = lpfn;
814 	}
815 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
816 	if (unlikely(r == -ENOMEM)) {
817 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
818 		return ttm_bo_validate(bo, &rbo->placement, false, false);
819 	} else if (unlikely(r != 0)) {
820 		return r;
821 	}
822 
823 	offset = bo->mem.start << PAGE_SHIFT;
824 	/* this should never happen */
825 	if ((offset + size) > rdev->mc.visible_vram_size)
826 		return -EINVAL;
827 
828 	return 0;
829 }
830 
831 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
832 {
833 	int r;
834 
835 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL);
836 	if (unlikely(r != 0))
837 		return r;
838 	if (mem_type)
839 		*mem_type = bo->tbo.mem.mem_type;
840 
841 	r = ttm_bo_wait(&bo->tbo, true, no_wait);
842 	ttm_bo_unreserve(&bo->tbo);
843 	return r;
844 }
845 
846 /**
847  * radeon_bo_fence - add fence to buffer object
848  *
849  * @bo: buffer object in question
850  * @fence: fence to add
851  * @shared: true if fence should be added shared
852  *
853  */
854 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
855 		     bool shared)
856 {
857 	struct reservation_object *resv = bo->tbo.resv;
858 
859 	if (shared)
860 		reservation_object_add_shared_fence(resv, &fence->base);
861 	else
862 		reservation_object_add_excl_fence(resv, &fence->base);
863 }
864