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