xref: /linux/drivers/gpu/drm/radeon/radeon_object.c (revision 957e3facd147510f2cf8780e38606f1d707f0e33)
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 
40 int radeon_ttm_init(struct radeon_device *rdev);
41 void radeon_ttm_fini(struct radeon_device *rdev);
42 static 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 
49 static void radeon_update_memory_usage(struct radeon_bo *bo,
50 				       unsigned mem_type, int sign)
51 {
52 	struct radeon_device *rdev = bo->rdev;
53 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
54 
55 	switch (mem_type) {
56 	case TTM_PL_TT:
57 		if (sign > 0)
58 			atomic64_add(size, &rdev->gtt_usage);
59 		else
60 			atomic64_sub(size, &rdev->gtt_usage);
61 		break;
62 	case TTM_PL_VRAM:
63 		if (sign > 0)
64 			atomic64_add(size, &rdev->vram_usage);
65 		else
66 			atomic64_sub(size, &rdev->vram_usage);
67 		break;
68 	}
69 }
70 
71 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
72 {
73 	struct radeon_bo *bo;
74 
75 	bo = container_of(tbo, struct radeon_bo, tbo);
76 
77 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
78 	radeon_mn_unregister(bo);
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 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
104 					     TTM_PL_FLAG_UNCACHED |
105 					     TTM_PL_FLAG_VRAM;
106 
107 	if (domain & RADEON_GEM_DOMAIN_GTT) {
108 		if (rbo->flags & RADEON_GEM_GTT_UC) {
109 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
110 				TTM_PL_FLAG_TT;
111 
112 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
113 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
114 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
115 				TTM_PL_FLAG_UNCACHED |
116 				TTM_PL_FLAG_TT;
117 		} else {
118 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
119 						     TTM_PL_FLAG_TT;
120 		}
121 	}
122 
123 	if (domain & RADEON_GEM_DOMAIN_CPU) {
124 		if (rbo->flags & RADEON_GEM_GTT_UC) {
125 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
126 				TTM_PL_FLAG_SYSTEM;
127 
128 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
129 		    rbo->rdev->flags & RADEON_IS_AGP) {
130 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
131 				TTM_PL_FLAG_UNCACHED |
132 				TTM_PL_FLAG_SYSTEM;
133 		} else {
134 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
135 						     TTM_PL_FLAG_SYSTEM;
136 		}
137 	}
138 	if (!c)
139 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
140 					     TTM_PL_FLAG_SYSTEM;
141 
142 	rbo->placement.num_placement = c;
143 	rbo->placement.num_busy_placement = c;
144 
145 	for (i = 0; i < c; ++i) {
146 		rbo->placements[i].fpfn = 0;
147 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
148 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM))
149 			rbo->placements[i].lpfn =
150 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
151 		else
152 			rbo->placements[i].lpfn = 0;
153 	}
154 
155 	/*
156 	 * Use two-ended allocation depending on the buffer size to
157 	 * improve fragmentation quality.
158 	 * 512kb was measured as the most optimal number.
159 	 */
160 	if (!((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
161 	      (rbo->placements[i].flags & TTM_PL_FLAG_VRAM)) &&
162 	    rbo->tbo.mem.size > 512 * 1024) {
163 		for (i = 0; i < c; i++) {
164 			rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
165 		}
166 	}
167 }
168 
169 int radeon_bo_create(struct radeon_device *rdev,
170 		     unsigned long size, int byte_align, bool kernel,
171 		     u32 domain, u32 flags, struct sg_table *sg,
172 		     struct reservation_object *resv,
173 		     struct radeon_bo **bo_ptr)
174 {
175 	struct radeon_bo *bo;
176 	enum ttm_bo_type type;
177 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
178 	size_t acc_size;
179 	int r;
180 
181 	size = ALIGN(size, PAGE_SIZE);
182 
183 	if (kernel) {
184 		type = ttm_bo_type_kernel;
185 	} else if (sg) {
186 		type = ttm_bo_type_sg;
187 	} else {
188 		type = ttm_bo_type_device;
189 	}
190 	*bo_ptr = NULL;
191 
192 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
193 				       sizeof(struct radeon_bo));
194 
195 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
196 	if (bo == NULL)
197 		return -ENOMEM;
198 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
199 	if (unlikely(r)) {
200 		kfree(bo);
201 		return r;
202 	}
203 	bo->rdev = rdev;
204 	bo->surface_reg = -1;
205 	INIT_LIST_HEAD(&bo->list);
206 	INIT_LIST_HEAD(&bo->va);
207 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
208 	                               RADEON_GEM_DOMAIN_GTT |
209 	                               RADEON_GEM_DOMAIN_CPU);
210 
211 	bo->flags = flags;
212 	/* PCI GART is always snooped */
213 	if (!(rdev->flags & RADEON_IS_PCIE))
214 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
215 
216 #ifdef CONFIG_X86_32
217 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
218 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
219 	 */
220 	bo->flags &= ~RADEON_GEM_GTT_WC;
221 #endif
222 
223 	radeon_ttm_placement_from_domain(bo, domain);
224 	/* Kernel allocation are uninterruptible */
225 	down_read(&rdev->pm.mclk_lock);
226 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
227 			&bo->placement, page_align, !kernel, NULL,
228 			acc_size, sg, resv, &radeon_ttm_bo_destroy);
229 	up_read(&rdev->pm.mclk_lock);
230 	if (unlikely(r != 0)) {
231 		return r;
232 	}
233 	*bo_ptr = bo;
234 
235 	trace_radeon_bo_create(bo);
236 
237 	return 0;
238 }
239 
240 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
241 {
242 	bool is_iomem;
243 	int r;
244 
245 	if (bo->kptr) {
246 		if (ptr) {
247 			*ptr = bo->kptr;
248 		}
249 		return 0;
250 	}
251 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
252 	if (r) {
253 		return r;
254 	}
255 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
256 	if (ptr) {
257 		*ptr = bo->kptr;
258 	}
259 	radeon_bo_check_tiling(bo, 0, 0);
260 	return 0;
261 }
262 
263 void radeon_bo_kunmap(struct radeon_bo *bo)
264 {
265 	if (bo->kptr == NULL)
266 		return;
267 	bo->kptr = NULL;
268 	radeon_bo_check_tiling(bo, 0, 0);
269 	ttm_bo_kunmap(&bo->kmap);
270 }
271 
272 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
273 {
274 	if (bo == NULL)
275 		return NULL;
276 
277 	ttm_bo_reference(&bo->tbo);
278 	return bo;
279 }
280 
281 void radeon_bo_unref(struct radeon_bo **bo)
282 {
283 	struct ttm_buffer_object *tbo;
284 	struct radeon_device *rdev;
285 
286 	if ((*bo) == NULL)
287 		return;
288 	rdev = (*bo)->rdev;
289 	tbo = &((*bo)->tbo);
290 	ttm_bo_unref(&tbo);
291 	if (tbo == NULL)
292 		*bo = NULL;
293 }
294 
295 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
296 			     u64 *gpu_addr)
297 {
298 	int r, i;
299 
300 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
301 		return -EPERM;
302 
303 	if (bo->pin_count) {
304 		bo->pin_count++;
305 		if (gpu_addr)
306 			*gpu_addr = radeon_bo_gpu_offset(bo);
307 
308 		if (max_offset != 0) {
309 			u64 domain_start;
310 
311 			if (domain == RADEON_GEM_DOMAIN_VRAM)
312 				domain_start = bo->rdev->mc.vram_start;
313 			else
314 				domain_start = bo->rdev->mc.gtt_start;
315 			WARN_ON_ONCE(max_offset <
316 				     (radeon_bo_gpu_offset(bo) - domain_start));
317 		}
318 
319 		return 0;
320 	}
321 	radeon_ttm_placement_from_domain(bo, domain);
322 	for (i = 0; i < bo->placement.num_placement; i++) {
323 		/* force to pin into visible video ram */
324 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
325 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
326 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
327 			bo->placements[i].lpfn =
328 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
329 		else
330 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
331 
332 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
333 	}
334 
335 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
336 	if (likely(r == 0)) {
337 		bo->pin_count = 1;
338 		if (gpu_addr != NULL)
339 			*gpu_addr = radeon_bo_gpu_offset(bo);
340 		if (domain == RADEON_GEM_DOMAIN_VRAM)
341 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
342 		else
343 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
344 	} else {
345 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
346 	}
347 	return r;
348 }
349 
350 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
351 {
352 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
353 }
354 
355 int radeon_bo_unpin(struct radeon_bo *bo)
356 {
357 	int r, i;
358 
359 	if (!bo->pin_count) {
360 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
361 		return 0;
362 	}
363 	bo->pin_count--;
364 	if (bo->pin_count)
365 		return 0;
366 	for (i = 0; i < bo->placement.num_placement; i++) {
367 		bo->placements[i].lpfn = 0;
368 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
369 	}
370 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
371 	if (likely(r == 0)) {
372 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
373 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
374 		else
375 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
376 	} else {
377 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
378 	}
379 	return r;
380 }
381 
382 int radeon_bo_evict_vram(struct radeon_device *rdev)
383 {
384 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
385 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
386 		if (rdev->mc.igp_sideport_enabled == false)
387 			/* Useless to evict on IGP chips */
388 			return 0;
389 	}
390 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
391 }
392 
393 void radeon_bo_force_delete(struct radeon_device *rdev)
394 {
395 	struct radeon_bo *bo, *n;
396 
397 	if (list_empty(&rdev->gem.objects)) {
398 		return;
399 	}
400 	dev_err(rdev->dev, "Userspace still has active objects !\n");
401 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
402 		mutex_lock(&rdev->ddev->struct_mutex);
403 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
404 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
405 			*((unsigned long *)&bo->gem_base.refcount));
406 		mutex_lock(&bo->rdev->gem.mutex);
407 		list_del_init(&bo->list);
408 		mutex_unlock(&bo->rdev->gem.mutex);
409 		/* this should unref the ttm bo */
410 		drm_gem_object_unreference(&bo->gem_base);
411 		mutex_unlock(&rdev->ddev->struct_mutex);
412 	}
413 }
414 
415 int radeon_bo_init(struct radeon_device *rdev)
416 {
417 	/* Add an MTRR for the VRAM */
418 	if (!rdev->fastfb_working) {
419 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
420 						      rdev->mc.aper_size);
421 	}
422 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
423 		rdev->mc.mc_vram_size >> 20,
424 		(unsigned long long)rdev->mc.aper_size >> 20);
425 	DRM_INFO("RAM width %dbits %cDR\n",
426 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
427 	return radeon_ttm_init(rdev);
428 }
429 
430 void radeon_bo_fini(struct radeon_device *rdev)
431 {
432 	radeon_ttm_fini(rdev);
433 	arch_phys_wc_del(rdev->mc.vram_mtrr);
434 }
435 
436 /* Returns how many bytes TTM can move per IB.
437  */
438 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
439 {
440 	u64 real_vram_size = rdev->mc.real_vram_size;
441 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
442 
443 	/* This function is based on the current VRAM usage.
444 	 *
445 	 * - If all of VRAM is free, allow relocating the number of bytes that
446 	 *   is equal to 1/4 of the size of VRAM for this IB.
447 
448 	 * - If more than one half of VRAM is occupied, only allow relocating
449 	 *   1 MB of data for this IB.
450 	 *
451 	 * - From 0 to one half of used VRAM, the threshold decreases
452 	 *   linearly.
453 	 *         __________________
454 	 * 1/4 of -|\               |
455 	 * VRAM    | \              |
456 	 *         |  \             |
457 	 *         |   \            |
458 	 *         |    \           |
459 	 *         |     \          |
460 	 *         |      \         |
461 	 *         |       \________|1 MB
462 	 *         |----------------|
463 	 *    VRAM 0 %             100 %
464 	 *         used            used
465 	 *
466 	 * Note: It's a threshold, not a limit. The threshold must be crossed
467 	 * for buffer relocations to stop, so any buffer of an arbitrary size
468 	 * can be moved as long as the threshold isn't crossed before
469 	 * the relocation takes place. We don't want to disable buffer
470 	 * relocations completely.
471 	 *
472 	 * The idea is that buffers should be placed in VRAM at creation time
473 	 * and TTM should only do a minimum number of relocations during
474 	 * command submission. In practice, you need to submit at least
475 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
476 	 *
477 	 * Also, things can get pretty crazy under memory pressure and actual
478 	 * VRAM usage can change a lot, so playing safe even at 50% does
479 	 * consistently increase performance.
480 	 */
481 
482 	u64 half_vram = real_vram_size >> 1;
483 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
484 	u64 bytes_moved_threshold = half_free_vram >> 1;
485 	return max(bytes_moved_threshold, 1024*1024ull);
486 }
487 
488 int radeon_bo_list_validate(struct radeon_device *rdev,
489 			    struct ww_acquire_ctx *ticket,
490 			    struct list_head *head, int ring)
491 {
492 	struct radeon_cs_reloc *lobj;
493 	struct radeon_bo *bo;
494 	int r;
495 	u64 bytes_moved = 0, initial_bytes_moved;
496 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
497 
498 	r = ttm_eu_reserve_buffers(ticket, head, true);
499 	if (unlikely(r != 0)) {
500 		return r;
501 	}
502 
503 	list_for_each_entry(lobj, head, tv.head) {
504 		bo = lobj->robj;
505 		if (!bo->pin_count) {
506 			u32 domain = lobj->prefered_domains;
507 			u32 allowed = lobj->allowed_domains;
508 			u32 current_domain =
509 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
510 
511 			/* Check if this buffer will be moved and don't move it
512 			 * if we have moved too many buffers for this IB already.
513 			 *
514 			 * Note that this allows moving at least one buffer of
515 			 * any size, because it doesn't take the current "bo"
516 			 * into account. We don't want to disallow buffer moves
517 			 * completely.
518 			 */
519 			if ((allowed & current_domain) != 0 &&
520 			    (domain & current_domain) == 0 && /* will be moved */
521 			    bytes_moved > bytes_moved_threshold) {
522 				/* don't move it */
523 				domain = current_domain;
524 			}
525 
526 		retry:
527 			radeon_ttm_placement_from_domain(bo, domain);
528 			if (ring == R600_RING_TYPE_UVD_INDEX)
529 				radeon_uvd_force_into_uvd_segment(bo, allowed);
530 
531 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
532 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
533 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
534 				       initial_bytes_moved;
535 
536 			if (unlikely(r)) {
537 				if (r != -ERESTARTSYS &&
538 				    domain != lobj->allowed_domains) {
539 					domain = lobj->allowed_domains;
540 					goto retry;
541 				}
542 				ttm_eu_backoff_reservation(ticket, head);
543 				return r;
544 			}
545 		}
546 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
547 		lobj->tiling_flags = bo->tiling_flags;
548 	}
549 	return 0;
550 }
551 
552 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
553 			     struct vm_area_struct *vma)
554 {
555 	return ttm_fbdev_mmap(vma, &bo->tbo);
556 }
557 
558 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
559 {
560 	struct radeon_device *rdev = bo->rdev;
561 	struct radeon_surface_reg *reg;
562 	struct radeon_bo *old_object;
563 	int steal;
564 	int i;
565 
566 	lockdep_assert_held(&bo->tbo.resv->lock.base);
567 
568 	if (!bo->tiling_flags)
569 		return 0;
570 
571 	if (bo->surface_reg >= 0) {
572 		reg = &rdev->surface_regs[bo->surface_reg];
573 		i = bo->surface_reg;
574 		goto out;
575 	}
576 
577 	steal = -1;
578 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
579 
580 		reg = &rdev->surface_regs[i];
581 		if (!reg->bo)
582 			break;
583 
584 		old_object = reg->bo;
585 		if (old_object->pin_count == 0)
586 			steal = i;
587 	}
588 
589 	/* if we are all out */
590 	if (i == RADEON_GEM_MAX_SURFACES) {
591 		if (steal == -1)
592 			return -ENOMEM;
593 		/* find someone with a surface reg and nuke their BO */
594 		reg = &rdev->surface_regs[steal];
595 		old_object = reg->bo;
596 		/* blow away the mapping */
597 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
598 		ttm_bo_unmap_virtual(&old_object->tbo);
599 		old_object->surface_reg = -1;
600 		i = steal;
601 	}
602 
603 	bo->surface_reg = i;
604 	reg->bo = bo;
605 
606 out:
607 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
608 			       bo->tbo.mem.start << PAGE_SHIFT,
609 			       bo->tbo.num_pages << PAGE_SHIFT);
610 	return 0;
611 }
612 
613 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
614 {
615 	struct radeon_device *rdev = bo->rdev;
616 	struct radeon_surface_reg *reg;
617 
618 	if (bo->surface_reg == -1)
619 		return;
620 
621 	reg = &rdev->surface_regs[bo->surface_reg];
622 	radeon_clear_surface_reg(rdev, bo->surface_reg);
623 
624 	reg->bo = NULL;
625 	bo->surface_reg = -1;
626 }
627 
628 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
629 				uint32_t tiling_flags, uint32_t pitch)
630 {
631 	struct radeon_device *rdev = bo->rdev;
632 	int r;
633 
634 	if (rdev->family >= CHIP_CEDAR) {
635 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
636 
637 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
638 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
639 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
640 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
641 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
642 		switch (bankw) {
643 		case 0:
644 		case 1:
645 		case 2:
646 		case 4:
647 		case 8:
648 			break;
649 		default:
650 			return -EINVAL;
651 		}
652 		switch (bankh) {
653 		case 0:
654 		case 1:
655 		case 2:
656 		case 4:
657 		case 8:
658 			break;
659 		default:
660 			return -EINVAL;
661 		}
662 		switch (mtaspect) {
663 		case 0:
664 		case 1:
665 		case 2:
666 		case 4:
667 		case 8:
668 			break;
669 		default:
670 			return -EINVAL;
671 		}
672 		if (tilesplit > 6) {
673 			return -EINVAL;
674 		}
675 		if (stilesplit > 6) {
676 			return -EINVAL;
677 		}
678 	}
679 	r = radeon_bo_reserve(bo, false);
680 	if (unlikely(r != 0))
681 		return r;
682 	bo->tiling_flags = tiling_flags;
683 	bo->pitch = pitch;
684 	radeon_bo_unreserve(bo);
685 	return 0;
686 }
687 
688 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
689 				uint32_t *tiling_flags,
690 				uint32_t *pitch)
691 {
692 	lockdep_assert_held(&bo->tbo.resv->lock.base);
693 
694 	if (tiling_flags)
695 		*tiling_flags = bo->tiling_flags;
696 	if (pitch)
697 		*pitch = bo->pitch;
698 }
699 
700 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
701 				bool force_drop)
702 {
703 	if (!force_drop)
704 		lockdep_assert_held(&bo->tbo.resv->lock.base);
705 
706 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
707 		return 0;
708 
709 	if (force_drop) {
710 		radeon_bo_clear_surface_reg(bo);
711 		return 0;
712 	}
713 
714 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
715 		if (!has_moved)
716 			return 0;
717 
718 		if (bo->surface_reg >= 0)
719 			radeon_bo_clear_surface_reg(bo);
720 		return 0;
721 	}
722 
723 	if ((bo->surface_reg >= 0) && !has_moved)
724 		return 0;
725 
726 	return radeon_bo_get_surface_reg(bo);
727 }
728 
729 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
730 			   struct ttm_mem_reg *new_mem)
731 {
732 	struct radeon_bo *rbo;
733 
734 	if (!radeon_ttm_bo_is_radeon_bo(bo))
735 		return;
736 
737 	rbo = container_of(bo, struct radeon_bo, tbo);
738 	radeon_bo_check_tiling(rbo, 0, 1);
739 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
740 
741 	/* update statistics */
742 	if (!new_mem)
743 		return;
744 
745 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
746 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
747 }
748 
749 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
750 {
751 	struct radeon_device *rdev;
752 	struct radeon_bo *rbo;
753 	unsigned long offset, size;
754 	int r;
755 
756 	if (!radeon_ttm_bo_is_radeon_bo(bo))
757 		return 0;
758 	rbo = container_of(bo, struct radeon_bo, tbo);
759 	radeon_bo_check_tiling(rbo, 0, 0);
760 	rdev = rbo->rdev;
761 	if (bo->mem.mem_type != TTM_PL_VRAM)
762 		return 0;
763 
764 	size = bo->mem.num_pages << PAGE_SHIFT;
765 	offset = bo->mem.start << PAGE_SHIFT;
766 	if ((offset + size) <= rdev->mc.visible_vram_size)
767 		return 0;
768 
769 	/* hurrah the memory is not visible ! */
770 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
771 	rbo->placements[0].lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
772 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
773 	if (unlikely(r == -ENOMEM)) {
774 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
775 		return ttm_bo_validate(bo, &rbo->placement, false, false);
776 	} else if (unlikely(r != 0)) {
777 		return r;
778 	}
779 
780 	offset = bo->mem.start << PAGE_SHIFT;
781 	/* this should never happen */
782 	if ((offset + size) > rdev->mc.visible_vram_size)
783 		return -EINVAL;
784 
785 	return 0;
786 }
787 
788 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
789 {
790 	int r;
791 
792 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
793 	if (unlikely(r != 0))
794 		return r;
795 	if (mem_type)
796 		*mem_type = bo->tbo.mem.mem_type;
797 
798 	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
799 	ttm_bo_unreserve(&bo->tbo);
800 	return r;
801 }
802