xref: /dflybsd-src/sys/dev/drm/radeon/radeon_object.c (revision e586f31ca9899b49a4fc156613d9ecd853defcec)
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 <drm/drmP.h>
33 #include <uapi_drm/radeon_drm.h>
34 #include "radeon.h"
35 #ifdef TRACE_TODO
36 #include "radeon_trace.h"
37 #endif
38 #include <linux/io.h>
39 
40 
41 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
42 
43 /*
44  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
45  * function are calling it.
46  */
47 
48 static void radeon_update_memory_usage(struct radeon_bo *bo,
49 				       unsigned mem_type, int sign)
50 {
51 	struct radeon_device *rdev = bo->rdev;
52 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
53 
54 	switch (mem_type) {
55 	case TTM_PL_TT:
56 		if (sign > 0)
57 			atomic64_add(size, &rdev->gtt_usage);
58 		else
59 			atomic64_sub(size, &rdev->gtt_usage);
60 		break;
61 	case TTM_PL_VRAM:
62 		if (sign > 0)
63 			atomic64_add(size, &rdev->vram_usage);
64 		else
65 			atomic64_sub(size, &rdev->vram_usage);
66 		break;
67 	}
68 }
69 
70 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
71 {
72 	struct radeon_bo *bo;
73 
74 	bo = container_of(tbo, struct radeon_bo, tbo);
75 
76 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
77 
78 	mutex_lock(&bo->rdev->gem.mutex);
79 	list_del_init(&bo->list);
80 	mutex_unlock(&bo->rdev->gem.mutex);
81 	radeon_bo_clear_surface_reg(bo);
82 	WARN_ON(!list_empty(&bo->va));
83 	drm_gem_object_release(&bo->gem_base);
84 	kfree(bo);
85 }
86 
87 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
88 {
89 	if (bo->destroy == &radeon_ttm_bo_destroy)
90 		return true;
91 	return false;
92 }
93 
94 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
95 {
96 	u32 c = 0, i;
97 
98 	rbo->placement.placement = rbo->placements;
99 	rbo->placement.busy_placement = rbo->placements;
100 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
101 		/* Try placing BOs which don't need CPU access outside of the
102 		 * CPU accessible part of VRAM
103 		 */
104 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
105 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
106 			rbo->placements[c].fpfn =
107 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
108 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
109 						     TTM_PL_FLAG_UNCACHED |
110 						     TTM_PL_FLAG_VRAM;
111 		}
112 
113 		rbo->placements[c].fpfn = 0;
114 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
115 					     TTM_PL_FLAG_UNCACHED |
116 					     TTM_PL_FLAG_VRAM;
117 	}
118 
119 	if (domain & RADEON_GEM_DOMAIN_GTT) {
120 		if (rbo->flags & RADEON_GEM_GTT_UC) {
121 			rbo->placements[c].fpfn = 0;
122 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
123 				TTM_PL_FLAG_TT;
124 
125 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
126 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
127 			rbo->placements[c].fpfn = 0;
128 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
129 				TTM_PL_FLAG_UNCACHED |
130 				TTM_PL_FLAG_TT;
131 		} else {
132 			rbo->placements[c].fpfn = 0;
133 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
134 						     TTM_PL_FLAG_TT;
135 		}
136 	}
137 
138 	if (domain & RADEON_GEM_DOMAIN_CPU) {
139 		if (rbo->flags & RADEON_GEM_GTT_UC) {
140 			rbo->placements[c].fpfn = 0;
141 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
142 				TTM_PL_FLAG_SYSTEM;
143 
144 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
145 		    rbo->rdev->flags & RADEON_IS_AGP) {
146 			rbo->placements[c].fpfn = 0;
147 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
148 				TTM_PL_FLAG_UNCACHED |
149 				TTM_PL_FLAG_SYSTEM;
150 		} else {
151 			rbo->placements[c].fpfn = 0;
152 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
153 						     TTM_PL_FLAG_SYSTEM;
154 		}
155 	}
156 	if (!c) {
157 		rbo->placements[c].fpfn = 0;
158  		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
159  					     TTM_PL_FLAG_SYSTEM;
160 	}
161 
162 	rbo->placement.num_placement = c;
163 	rbo->placement.num_busy_placement = c;
164 
165 	for (i = 0; i < c; ++i) {
166 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
167 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
168 		    !rbo->placements[i].fpfn)
169 			rbo->placements[i].lpfn =
170 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
171 		else
172 			rbo->placements[i].lpfn = 0;
173 	}
174 
175 #if 0
176 	/*
177 	 * Use two-ended allocation depending on the buffer size to
178 	 * improve fragmentation quality.
179 	 * 512kb was measured as the most optimal number.
180 	 */
181 	if (!((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
182 	      (rbo->placements[i].flags & TTM_PL_FLAG_VRAM)) &&
183 	    rbo->tbo.mem.size > 512 * 1024) {
184 		for (i = 0; i < c; i++) {
185 			rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
186 		}
187 	}
188 #endif
189 }
190 
191 int radeon_bo_create(struct radeon_device *rdev,
192 		     unsigned long size, int byte_align, bool kernel, u32 domain,
193 		     u32 flags, struct sg_table *sg, struct radeon_bo **bo_ptr)
194 {
195 	struct radeon_bo *bo;
196 	enum ttm_bo_type type;
197 	unsigned long page_align = roundup2(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 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
237 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
238 	 */
239 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
240 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
241 
242 /* DragonFly only supported on __x86_64__ and supports PAT */
243 #if !defined (__DragonFly__)
244 #ifdef CONFIG_X86_32
245 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
246 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
247 	 */
248 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
249 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
250 	/* Don't try to enable write-combining when it can't work, or things
251 	 * may be slow
252 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
253 	 */
254 
255 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
256 	 thanks to write-combining
257 
258 	if (bo->flags & RADEON_GEM_GTT_WC)
259 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
260 			      "better performance thanks to write-combining\n");
261 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
262 #else
263 	/* For architectures that don't support WC memory,
264 	 * mask out the WC flag from the BO
265 	 */
266 	if (!drm_arch_can_wc_memory())
267 		bo->flags &= ~RADEON_GEM_GTT_WC;
268 #endif
269 #endif /* __DragonFly__*/
270 
271 	radeon_ttm_placement_from_domain(bo, domain);
272 	/* Kernel allocation are uninterruptible */
273 	lockmgr(&rdev->pm.mclk_lock, LK_SHARED);
274 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
275 			&bo->placement, page_align, !kernel, NULL,
276 			acc_size, sg, &radeon_ttm_bo_destroy);
277 	lockmgr(&rdev->pm.mclk_lock, LK_RELEASE);
278 	if (unlikely(r != 0)) {
279 		return r;
280 	}
281 	*bo_ptr = bo;
282 
283 #ifdef TRACE_TODO
284 	trace_radeon_bo_create(bo);
285 #endif
286 
287 	return 0;
288 }
289 
290 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
291 {
292 	bool is_iomem;
293 	int r;
294 
295 	if (bo->kptr) {
296 		if (ptr) {
297 			*ptr = bo->kptr;
298 		}
299 		return 0;
300 	}
301 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
302 	if (r) {
303 		return r;
304 	}
305 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
306 	if (ptr) {
307 		*ptr = bo->kptr;
308 	}
309 	radeon_bo_check_tiling(bo, 0, 0);
310 	return 0;
311 }
312 
313 void radeon_bo_kunmap(struct radeon_bo *bo)
314 {
315 	if (bo->kptr == NULL)
316 		return;
317 	bo->kptr = NULL;
318 	radeon_bo_check_tiling(bo, 0, 0);
319 	ttm_bo_kunmap(&bo->kmap);
320 }
321 
322 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
323 {
324 	if (bo == NULL)
325 		return NULL;
326 
327 	ttm_bo_reference(&bo->tbo);
328 	return bo;
329 }
330 
331 void radeon_bo_unref(struct radeon_bo **bo)
332 {
333 	struct ttm_buffer_object *tbo;
334 	struct radeon_device *rdev;
335 	struct radeon_bo *rbo;
336 
337 	if ((rbo = *bo) == NULL)
338 		return;
339 	*bo = NULL;
340 	rdev = rbo->rdev;
341 	tbo = &rbo->tbo;
342 	ttm_bo_unref(&tbo);
343 }
344 
345 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
346 			     u64 *gpu_addr)
347 {
348 	int r, i;
349 
350 	if (bo->pin_count) {
351 		bo->pin_count++;
352 		if (gpu_addr)
353 			*gpu_addr = radeon_bo_gpu_offset(bo);
354 
355 		if (max_offset != 0) {
356 			u64 domain_start;
357 
358 			if (domain == RADEON_GEM_DOMAIN_VRAM)
359 				domain_start = bo->rdev->mc.vram_start;
360 			else
361 				domain_start = bo->rdev->mc.gtt_start;
362 			if (max_offset < (radeon_bo_gpu_offset(bo) - domain_start)) {
363 				DRM_ERROR("radeon_bo_pin_restricted: "
364 				    "max_offset(%ju) < "
365 				    "(radeon_bo_gpu_offset(%ju) - "
366 				    "domain_start(%ju)",
367 				    (uintmax_t)max_offset, (uintmax_t)radeon_bo_gpu_offset(bo),
368 				    (uintmax_t)domain_start);
369 			}
370 		}
371 
372 		return 0;
373 	}
374 	radeon_ttm_placement_from_domain(bo, domain);
375 	for (i = 0; i < bo->placement.num_placement; i++) {
376 		/* force to pin into visible video ram */
377 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
378 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
379 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
380 			bo->placements[i].lpfn =
381 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
382 		else
383 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
384 
385 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
386 	}
387 
388 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
389 	if (likely(r == 0)) {
390 		bo->pin_count = 1;
391 		if (gpu_addr != NULL)
392 			*gpu_addr = radeon_bo_gpu_offset(bo);
393 		if (domain == RADEON_GEM_DOMAIN_VRAM)
394 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
395 		else
396 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
397 	} else {
398 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
399 	}
400 	return r;
401 }
402 
403 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
404 {
405 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
406 }
407 
408 int radeon_bo_unpin(struct radeon_bo *bo)
409 {
410 	int r, i;
411 
412 	if (!bo->pin_count) {
413 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
414 		return 0;
415 	}
416 	bo->pin_count--;
417 	if (bo->pin_count)
418 		return 0;
419 	for (i = 0; i < bo->placement.num_placement; i++) {
420 		bo->placements[i].lpfn = 0;
421 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
422 	}
423 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
424 	if (likely(r == 0)) {
425 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
426 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
427 		else
428 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
429 	} else {
430 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
431 	}
432 	return r;
433 }
434 
435 int radeon_bo_evict_vram(struct radeon_device *rdev)
436 {
437 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
438 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
439 		if (rdev->mc.igp_sideport_enabled == false)
440 			/* Useless to evict on IGP chips */
441 			return 0;
442 	}
443 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
444 }
445 
446 void radeon_bo_force_delete(struct radeon_device *rdev)
447 {
448 	struct radeon_bo *bo, *n;
449 
450 	if (list_empty(&rdev->gem.objects)) {
451 		return;
452 	}
453 	dev_err(rdev->dev, "Userspace still has active objects !\n");
454 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
455 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
456 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
457 			*((unsigned long *)&bo->gem_base.refcount));
458 		mutex_lock(&bo->rdev->gem.mutex);
459 		list_del_init(&bo->list);
460 		mutex_unlock(&bo->rdev->gem.mutex);
461 		/* this should unref the ttm bo */
462 		drm_gem_object_unreference_unlocked(&bo->gem_base);
463 	}
464 }
465 
466 int radeon_bo_init(struct radeon_device *rdev)
467 {
468 	/* Add an MTRR for the VRAM */
469 	if (!rdev->fastfb_working) {
470 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
471 						      rdev->mc.aper_size);
472 	}
473 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
474 		rdev->mc.mc_vram_size >> 20,
475 		(unsigned long long)rdev->mc.aper_size >> 20);
476 	DRM_INFO("RAM width %dbits %cDR\n",
477 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
478 	return radeon_ttm_init(rdev);
479 }
480 
481 void radeon_bo_fini(struct radeon_device *rdev)
482 {
483 	radeon_ttm_fini(rdev);
484 	arch_phys_wc_del(rdev->mc.vram_mtrr);
485 }
486 
487 /* Returns how many bytes TTM can move per IB.
488  */
489 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
490 {
491 	u64 real_vram_size = rdev->mc.real_vram_size;
492 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
493 
494 	/* This function is based on the current VRAM usage.
495 	 *
496 	 * - If all of VRAM is free, allow relocating the number of bytes that
497 	 *   is equal to 1/4 of the size of VRAM for this IB.
498 
499 	 * - If more than one half of VRAM is occupied, only allow relocating
500 	 *   1 MB of data for this IB.
501 	 *
502 	 * - From 0 to one half of used VRAM, the threshold decreases
503 	 *   linearly.
504 	 *         __________________
505 	 * 1/4 of -|\               |
506 	 * VRAM    | \              |
507 	 *         |  \             |
508 	 *         |   \            |
509 	 *         |    \           |
510 	 *         |     \          |
511 	 *         |      \         |
512 	 *         |       \________|1 MB
513 	 *         |----------------|
514 	 *    VRAM 0 %             100 %
515 	 *         used            used
516 	 *
517 	 * Note: It's a threshold, not a limit. The threshold must be crossed
518 	 * for buffer relocations to stop, so any buffer of an arbitrary size
519 	 * can be moved as long as the threshold isn't crossed before
520 	 * the relocation takes place. We don't want to disable buffer
521 	 * relocations completely.
522 	 *
523 	 * The idea is that buffers should be placed in VRAM at creation time
524 	 * and TTM should only do a minimum number of relocations during
525 	 * command submission. In practice, you need to submit at least
526 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
527 	 *
528 	 * Also, things can get pretty crazy under memory pressure and actual
529 	 * VRAM usage can change a lot, so playing safe even at 50% does
530 	 * consistently increase performance.
531 	 */
532 
533 	u64 half_vram = real_vram_size >> 1;
534 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
535 	u64 bytes_moved_threshold = half_free_vram >> 1;
536 	return max(bytes_moved_threshold, 1024*1024ull);
537 }
538 
539 int radeon_bo_list_validate(struct radeon_device *rdev,
540 			    struct ww_acquire_ctx *ticket,
541 			    struct list_head *head, int ring)
542 {
543 	struct radeon_bo_list *lobj;
544 	struct radeon_bo *bo;
545 	int r;
546 	u64 bytes_moved = 0, initial_bytes_moved;
547 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
548 
549 	r = ttm_eu_reserve_buffers(ticket, head);
550 	if (unlikely(r != 0)) {
551 		return r;
552 	}
553 
554 	list_for_each_entry(lobj, head, tv.head) {
555 		bo = lobj->robj;
556 		if (!bo->pin_count) {
557 			u32 domain = lobj->prefered_domains;
558 			u32 allowed = lobj->allowed_domains;
559 			u32 current_domain =
560 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
561 
562 			/* Check if this buffer will be moved and don't move it
563 			 * if we have moved too many buffers for this IB already.
564 			 *
565 			 * Note that this allows moving at least one buffer of
566 			 * any size, because it doesn't take the current "bo"
567 			 * into account. We don't want to disallow buffer moves
568 			 * completely.
569 			 */
570 			if ((allowed & current_domain) != 0 &&
571 			    (domain & current_domain) == 0 && /* will be moved */
572 			    bytes_moved > bytes_moved_threshold) {
573 				/* don't move it */
574 				domain = current_domain;
575 			}
576 
577 		retry:
578 			radeon_ttm_placement_from_domain(bo, domain);
579 			if (ring == R600_RING_TYPE_UVD_INDEX)
580 				radeon_uvd_force_into_uvd_segment(bo, allowed);
581 
582 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
583 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
584 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
585 				       initial_bytes_moved;
586 
587 			if (unlikely(r)) {
588 				if (r != -ERESTARTSYS &&
589 				    domain != lobj->allowed_domains) {
590 					domain = lobj->allowed_domains;
591 					goto retry;
592 				}
593 				ttm_eu_backoff_reservation(ticket, head);
594 				return r;
595 			}
596 		}
597 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
598 		lobj->tiling_flags = bo->tiling_flags;
599 	}
600 	return 0;
601 }
602 
603 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
604 {
605 	struct radeon_device *rdev = bo->rdev;
606 	struct radeon_surface_reg *reg;
607 	struct radeon_bo *old_object;
608 	int steal;
609 	int i;
610 
611 	KASSERT(radeon_bo_is_reserved(bo),
612 	    ("radeon_bo_get_surface_reg: radeon_bo is not reserved"));
613 
614 	if (!bo->tiling_flags)
615 		return 0;
616 
617 	if (bo->surface_reg >= 0) {
618 		reg = &rdev->surface_regs[bo->surface_reg];
619 		i = bo->surface_reg;
620 		goto out;
621 	}
622 
623 	steal = -1;
624 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
625 
626 		reg = &rdev->surface_regs[i];
627 		if (!reg->bo)
628 			break;
629 
630 		old_object = reg->bo;
631 		if (old_object->pin_count == 0)
632 			steal = i;
633 	}
634 
635 	/* if we are all out */
636 	if (i == RADEON_GEM_MAX_SURFACES) {
637 		if (steal == -1)
638 			return -ENOMEM;
639 		/* find someone with a surface reg and nuke their BO */
640 		reg = &rdev->surface_regs[steal];
641 		old_object = reg->bo;
642 		/* blow away the mapping */
643 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
644 		ttm_bo_unmap_virtual(&old_object->tbo);
645 		old_object->surface_reg = -1;
646 		i = steal;
647 	}
648 
649 	bo->surface_reg = i;
650 	reg->bo = bo;
651 
652 out:
653 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
654 			       bo->tbo.mem.start << PAGE_SHIFT,
655 			       bo->tbo.num_pages << PAGE_SHIFT);
656 	return 0;
657 }
658 
659 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
660 {
661 	struct radeon_device *rdev = bo->rdev;
662 	struct radeon_surface_reg *reg;
663 
664 	if (bo->surface_reg == -1)
665 		return;
666 
667 	reg = &rdev->surface_regs[bo->surface_reg];
668 	radeon_clear_surface_reg(rdev, bo->surface_reg);
669 
670 	reg->bo = NULL;
671 	bo->surface_reg = -1;
672 }
673 
674 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
675 				uint32_t tiling_flags, uint32_t pitch)
676 {
677 	struct radeon_device *rdev = bo->rdev;
678 	int r;
679 
680 	if (rdev->family >= CHIP_CEDAR) {
681 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
682 
683 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
684 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
685 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
686 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
687 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
688 		switch (bankw) {
689 		case 0:
690 		case 1:
691 		case 2:
692 		case 4:
693 		case 8:
694 			break;
695 		default:
696 			return -EINVAL;
697 		}
698 		switch (bankh) {
699 		case 0:
700 		case 1:
701 		case 2:
702 		case 4:
703 		case 8:
704 			break;
705 		default:
706 			return -EINVAL;
707 		}
708 		switch (mtaspect) {
709 		case 0:
710 		case 1:
711 		case 2:
712 		case 4:
713 		case 8:
714 			break;
715 		default:
716 			return -EINVAL;
717 		}
718 		if (tilesplit > 6) {
719 			return -EINVAL;
720 		}
721 		if (stilesplit > 6) {
722 			return -EINVAL;
723 		}
724 	}
725 	r = radeon_bo_reserve(bo, false);
726 	if (unlikely(r != 0))
727 		return r;
728 	bo->tiling_flags = tiling_flags;
729 	bo->pitch = pitch;
730 	radeon_bo_unreserve(bo);
731 	return 0;
732 }
733 
734 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
735 				uint32_t *tiling_flags,
736 				uint32_t *pitch)
737 {
738 	KASSERT(radeon_bo_is_reserved(bo),
739 	    ("radeon_bo_get_tiling_flags: radeon_bo is not reserved"));
740 	if (tiling_flags)
741 		*tiling_flags = bo->tiling_flags;
742 	if (pitch)
743 		*pitch = bo->pitch;
744 }
745 
746 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
747 				bool force_drop)
748 {
749 	KASSERT((radeon_bo_is_reserved(bo) || force_drop),
750 	    ("radeon_bo_check_tiling: radeon_bo is not reserved && !force_drop"));
751 
752 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
753 		return 0;
754 
755 	if (force_drop) {
756 		radeon_bo_clear_surface_reg(bo);
757 		return 0;
758 	}
759 
760 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
761 		if (!has_moved)
762 			return 0;
763 
764 		if (bo->surface_reg >= 0)
765 			radeon_bo_clear_surface_reg(bo);
766 		return 0;
767 	}
768 
769 	if ((bo->surface_reg >= 0) && !has_moved)
770 		return 0;
771 
772 	return radeon_bo_get_surface_reg(bo);
773 }
774 
775 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
776 			   struct ttm_mem_reg *new_mem)
777 {
778 	struct radeon_bo *rbo;
779 
780 	if (!radeon_ttm_bo_is_radeon_bo(bo))
781 		return;
782 
783 	rbo = container_of(bo, struct radeon_bo, tbo);
784 	radeon_bo_check_tiling(rbo, 0, 1);
785 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
786 
787 	/* update statistics */
788 	if (!new_mem)
789 		return;
790 
791 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
792 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
793 }
794 
795 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
796 {
797 	struct radeon_device *rdev;
798 	struct radeon_bo *rbo;
799 	unsigned long offset, size, lpfn;
800 	int i, r;
801 
802 	if (!radeon_ttm_bo_is_radeon_bo(bo))
803 		return 0;
804 	rbo = container_of(bo, struct radeon_bo, tbo);
805 	radeon_bo_check_tiling(rbo, 0, 0);
806 	rdev = rbo->rdev;
807 	if (bo->mem.mem_type != TTM_PL_VRAM)
808 		return 0;
809 
810 	size = bo->mem.num_pages << PAGE_SHIFT;
811 	offset = bo->mem.start << PAGE_SHIFT;
812 	if ((offset + size) <= rdev->mc.visible_vram_size)
813 		return 0;
814 
815 	/* Can't move a pinned BO to visible VRAM */
816 	if (rbo->pin_count > 0)
817 		return -EINVAL;
818 
819 	/* hurrah the memory is not visible ! */
820 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
821 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
822 	for (i = 0; i < rbo->placement.num_placement; i++) {
823 		/* Force into visible VRAM */
824 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
825 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
826 			rbo->placements[i].lpfn = lpfn;
827 	}
828 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
829 	if (unlikely(r == -ENOMEM)) {
830 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
831 		return ttm_bo_validate(bo, &rbo->placement, false, false);
832 	} else if (unlikely(r != 0)) {
833 		return r;
834 	}
835 
836 	offset = bo->mem.start << PAGE_SHIFT;
837 	/* this should never happen */
838 	if ((offset + size) > rdev->mc.visible_vram_size)
839 		return -EINVAL;
840 
841 	return 0;
842 }
843 
844 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
845 {
846 	int r;
847 
848 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
849 	if (unlikely(r != 0))
850 		return r;
851 	lockmgr(&bo->tbo.bdev->fence_lock, LK_EXCLUSIVE);
852 	if (mem_type)
853 		*mem_type = bo->tbo.mem.mem_type;
854 	if (bo->tbo.sync_obj)
855 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
856 	lockmgr(&bo->tbo.bdev->fence_lock, LK_RELEASE);
857 	ttm_bo_unreserve(&bo->tbo);
858 	return r;
859 }
860