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_ONCE(!list_empty(&bo->va));
85 	if (bo->gem_base.import_attach)
86 		drm_prime_gem_destroy(&bo->gem_base, bo->tbo.sg);
87 	drm_gem_object_release(&bo->gem_base);
88 	kfree(bo);
89 }
90 
91 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
92 {
93 	if (bo->destroy == &radeon_ttm_bo_destroy)
94 		return true;
95 	return false;
96 }
97 
98 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
99 {
100 	u32 c = 0, i;
101 
102 	rbo->placement.placement = rbo->placements;
103 	rbo->placement.busy_placement = rbo->placements;
104 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
105 		/* Try placing BOs which don't need CPU access outside of the
106 		 * CPU accessible part of VRAM
107 		 */
108 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
109 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
110 			rbo->placements[c].fpfn =
111 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
112 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
113 						     TTM_PL_FLAG_UNCACHED |
114 						     TTM_PL_FLAG_VRAM;
115 		}
116 
117 		rbo->placements[c].fpfn = 0;
118 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
119 					     TTM_PL_FLAG_UNCACHED |
120 					     TTM_PL_FLAG_VRAM;
121 	}
122 
123 	if (domain & RADEON_GEM_DOMAIN_GTT) {
124 		if (rbo->flags & RADEON_GEM_GTT_UC) {
125 			rbo->placements[c].fpfn = 0;
126 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
127 				TTM_PL_FLAG_TT;
128 
129 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
130 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
131 			rbo->placements[c].fpfn = 0;
132 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
133 				TTM_PL_FLAG_UNCACHED |
134 				TTM_PL_FLAG_TT;
135 		} else {
136 			rbo->placements[c].fpfn = 0;
137 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
138 						     TTM_PL_FLAG_TT;
139 		}
140 	}
141 
142 	if (domain & RADEON_GEM_DOMAIN_CPU) {
143 		if (rbo->flags & RADEON_GEM_GTT_UC) {
144 			rbo->placements[c].fpfn = 0;
145 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
146 				TTM_PL_FLAG_SYSTEM;
147 
148 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
149 		    rbo->rdev->flags & RADEON_IS_AGP) {
150 			rbo->placements[c].fpfn = 0;
151 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
152 				TTM_PL_FLAG_UNCACHED |
153 				TTM_PL_FLAG_SYSTEM;
154 		} else {
155 			rbo->placements[c].fpfn = 0;
156 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
157 						     TTM_PL_FLAG_SYSTEM;
158 		}
159 	}
160 	if (!c) {
161 		rbo->placements[c].fpfn = 0;
162 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
163 					     TTM_PL_FLAG_SYSTEM;
164 	}
165 
166 	rbo->placement.num_placement = c;
167 	rbo->placement.num_busy_placement = c;
168 
169 	for (i = 0; i < c; ++i) {
170 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
171 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
172 		    !rbo->placements[i].fpfn)
173 			rbo->placements[i].lpfn =
174 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
175 		else
176 			rbo->placements[i].lpfn = 0;
177 	}
178 }
179 
180 int radeon_bo_create(struct radeon_device *rdev,
181 		     unsigned long size, int byte_align, bool kernel,
182 		     u32 domain, u32 flags, struct sg_table *sg,
183 		     struct reservation_object *resv,
184 		     struct radeon_bo **bo_ptr)
185 {
186 	struct radeon_bo *bo;
187 	enum ttm_bo_type type;
188 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
189 	size_t acc_size;
190 	int r;
191 
192 	size = ALIGN(size, PAGE_SIZE);
193 
194 	if (kernel) {
195 		type = ttm_bo_type_kernel;
196 	} else if (sg) {
197 		type = ttm_bo_type_sg;
198 	} else {
199 		type = ttm_bo_type_device;
200 	}
201 	*bo_ptr = NULL;
202 
203 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
204 				       sizeof(struct radeon_bo));
205 
206 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
207 	if (bo == NULL)
208 		return -ENOMEM;
209 	drm_gem_private_object_init(rdev->ddev, &bo->gem_base, size);
210 	bo->rdev = rdev;
211 	bo->surface_reg = -1;
212 	INIT_LIST_HEAD(&bo->list);
213 	INIT_LIST_HEAD(&bo->va);
214 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
215 				       RADEON_GEM_DOMAIN_GTT |
216 				       RADEON_GEM_DOMAIN_CPU);
217 
218 	bo->flags = flags;
219 	/* PCI GART is always snooped */
220 	if (!(rdev->flags & RADEON_IS_PCIE))
221 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
222 
223 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
224 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
225 	 */
226 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
227 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
228 
229 #ifdef CONFIG_X86_32
230 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
231 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
232 	 */
233 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
234 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
235 	/* Don't try to enable write-combining when it can't work, or things
236 	 * may be slow
237 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
238 	 */
239 #ifndef CONFIG_COMPILE_TEST
240 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
241 	 thanks to write-combining
242 #endif
243 
244 	if (bo->flags & RADEON_GEM_GTT_WC)
245 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
246 			      "better performance thanks to write-combining\n");
247 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
248 #else
249 	/* For architectures that don't support WC memory,
250 	 * mask out the WC flag from the BO
251 	 */
252 	if (!drm_arch_can_wc_memory())
253 		bo->flags &= ~RADEON_GEM_GTT_WC;
254 #endif
255 
256 	radeon_ttm_placement_from_domain(bo, domain);
257 	/* Kernel allocation are uninterruptible */
258 	down_read(&rdev->pm.mclk_lock);
259 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
260 			&bo->placement, page_align, !kernel, acc_size,
261 			sg, resv, &radeon_ttm_bo_destroy);
262 	up_read(&rdev->pm.mclk_lock);
263 	if (unlikely(r != 0)) {
264 		return r;
265 	}
266 	*bo_ptr = bo;
267 
268 	trace_radeon_bo_create(bo);
269 
270 	return 0;
271 }
272 
273 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
274 {
275 	bool is_iomem;
276 	int r;
277 
278 	if (bo->kptr) {
279 		if (ptr) {
280 			*ptr = bo->kptr;
281 		}
282 		return 0;
283 	}
284 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
285 	if (r) {
286 		return r;
287 	}
288 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
289 	if (ptr) {
290 		*ptr = bo->kptr;
291 	}
292 	radeon_bo_check_tiling(bo, 0, 0);
293 	return 0;
294 }
295 
296 void radeon_bo_kunmap(struct radeon_bo *bo)
297 {
298 	if (bo->kptr == NULL)
299 		return;
300 	bo->kptr = NULL;
301 	radeon_bo_check_tiling(bo, 0, 0);
302 	ttm_bo_kunmap(&bo->kmap);
303 }
304 
305 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
306 {
307 	if (bo == NULL)
308 		return NULL;
309 
310 	ttm_bo_get(&bo->tbo);
311 	return bo;
312 }
313 
314 void radeon_bo_unref(struct radeon_bo **bo)
315 {
316 	struct ttm_buffer_object *tbo;
317 	struct radeon_device *rdev;
318 
319 	if ((*bo) == NULL)
320 		return;
321 	rdev = (*bo)->rdev;
322 	tbo = &((*bo)->tbo);
323 	ttm_bo_put(tbo);
324 	*bo = NULL;
325 }
326 
327 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
328 			     u64 *gpu_addr)
329 {
330 	struct ttm_operation_ctx ctx = { false, false };
331 	int r, i;
332 
333 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
334 		return -EPERM;
335 
336 	if (bo->pin_count) {
337 		bo->pin_count++;
338 		if (gpu_addr)
339 			*gpu_addr = radeon_bo_gpu_offset(bo);
340 
341 		if (max_offset != 0) {
342 			u64 domain_start;
343 
344 			if (domain == RADEON_GEM_DOMAIN_VRAM)
345 				domain_start = bo->rdev->mc.vram_start;
346 			else
347 				domain_start = bo->rdev->mc.gtt_start;
348 			WARN_ON_ONCE(max_offset <
349 				     (radeon_bo_gpu_offset(bo) - domain_start));
350 		}
351 
352 		return 0;
353 	}
354 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
355 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
356 		return -EINVAL;
357 	}
358 
359 	radeon_ttm_placement_from_domain(bo, domain);
360 	for (i = 0; i < bo->placement.num_placement; i++) {
361 		/* force to pin into visible video ram */
362 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
363 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
364 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
365 			bo->placements[i].lpfn =
366 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
367 		else
368 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
369 
370 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
371 	}
372 
373 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
374 	if (likely(r == 0)) {
375 		bo->pin_count = 1;
376 		if (gpu_addr != NULL)
377 			*gpu_addr = radeon_bo_gpu_offset(bo);
378 		if (domain == RADEON_GEM_DOMAIN_VRAM)
379 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
380 		else
381 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
382 	} else {
383 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
384 	}
385 	return r;
386 }
387 
388 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
389 {
390 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
391 }
392 
393 int radeon_bo_unpin(struct radeon_bo *bo)
394 {
395 	struct ttm_operation_ctx ctx = { false, false };
396 	int r, i;
397 
398 	if (!bo->pin_count) {
399 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
400 		return 0;
401 	}
402 	bo->pin_count--;
403 	if (bo->pin_count)
404 		return 0;
405 	for (i = 0; i < bo->placement.num_placement; i++) {
406 		bo->placements[i].lpfn = 0;
407 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
408 	}
409 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
410 	if (likely(r == 0)) {
411 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
412 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
413 		else
414 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
415 	} else {
416 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
417 	}
418 	return r;
419 }
420 
421 int radeon_bo_evict_vram(struct radeon_device *rdev)
422 {
423 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
424 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
425 		if (rdev->mc.igp_sideport_enabled == false)
426 			/* Useless to evict on IGP chips */
427 			return 0;
428 	}
429 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
430 }
431 
432 void radeon_bo_force_delete(struct radeon_device *rdev)
433 {
434 	struct radeon_bo *bo, *n;
435 
436 	if (list_empty(&rdev->gem.objects)) {
437 		return;
438 	}
439 	dev_err(rdev->dev, "Userspace still has active objects !\n");
440 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
441 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
442 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
443 			*((unsigned long *)&bo->gem_base.refcount));
444 		mutex_lock(&bo->rdev->gem.mutex);
445 		list_del_init(&bo->list);
446 		mutex_unlock(&bo->rdev->gem.mutex);
447 		/* this should unref the ttm bo */
448 		drm_gem_object_put_unlocked(&bo->gem_base);
449 	}
450 }
451 
452 int radeon_bo_init(struct radeon_device *rdev)
453 {
454 	/* reserve PAT memory space to WC for VRAM */
455 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
456 				   rdev->mc.aper_size);
457 
458 	/* Add an MTRR for the VRAM */
459 	if (!rdev->fastfb_working) {
460 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
461 						      rdev->mc.aper_size);
462 	}
463 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
464 		rdev->mc.mc_vram_size >> 20,
465 		(unsigned long long)rdev->mc.aper_size >> 20);
466 	DRM_INFO("RAM width %dbits %cDR\n",
467 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
468 	return radeon_ttm_init(rdev);
469 }
470 
471 void radeon_bo_fini(struct radeon_device *rdev)
472 {
473 	radeon_ttm_fini(rdev);
474 	arch_phys_wc_del(rdev->mc.vram_mtrr);
475 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
476 }
477 
478 /* Returns how many bytes TTM can move per IB.
479  */
480 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
481 {
482 	u64 real_vram_size = rdev->mc.real_vram_size;
483 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
484 
485 	/* This function is based on the current VRAM usage.
486 	 *
487 	 * - If all of VRAM is free, allow relocating the number of bytes that
488 	 *   is equal to 1/4 of the size of VRAM for this IB.
489 
490 	 * - If more than one half of VRAM is occupied, only allow relocating
491 	 *   1 MB of data for this IB.
492 	 *
493 	 * - From 0 to one half of used VRAM, the threshold decreases
494 	 *   linearly.
495 	 *         __________________
496 	 * 1/4 of -|\               |
497 	 * VRAM    | \              |
498 	 *         |  \             |
499 	 *         |   \            |
500 	 *         |    \           |
501 	 *         |     \          |
502 	 *         |      \         |
503 	 *         |       \________|1 MB
504 	 *         |----------------|
505 	 *    VRAM 0 %             100 %
506 	 *         used            used
507 	 *
508 	 * Note: It's a threshold, not a limit. The threshold must be crossed
509 	 * for buffer relocations to stop, so any buffer of an arbitrary size
510 	 * can be moved as long as the threshold isn't crossed before
511 	 * the relocation takes place. We don't want to disable buffer
512 	 * relocations completely.
513 	 *
514 	 * The idea is that buffers should be placed in VRAM at creation time
515 	 * and TTM should only do a minimum number of relocations during
516 	 * command submission. In practice, you need to submit at least
517 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
518 	 *
519 	 * Also, things can get pretty crazy under memory pressure and actual
520 	 * VRAM usage can change a lot, so playing safe even at 50% does
521 	 * consistently increase performance.
522 	 */
523 
524 	u64 half_vram = real_vram_size >> 1;
525 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
526 	u64 bytes_moved_threshold = half_free_vram >> 1;
527 	return max(bytes_moved_threshold, 1024*1024ull);
528 }
529 
530 int radeon_bo_list_validate(struct radeon_device *rdev,
531 			    struct ww_acquire_ctx *ticket,
532 			    struct list_head *head, int ring)
533 {
534 	struct ttm_operation_ctx ctx = { true, false };
535 	struct radeon_bo_list *lobj;
536 	struct list_head duplicates;
537 	int r;
538 	u64 bytes_moved = 0, initial_bytes_moved;
539 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
540 
541 	INIT_LIST_HEAD(&duplicates);
542 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
543 	if (unlikely(r != 0)) {
544 		return r;
545 	}
546 
547 	list_for_each_entry(lobj, head, tv.head) {
548 		struct radeon_bo *bo = lobj->robj;
549 		if (!bo->pin_count) {
550 			u32 domain = lobj->preferred_domains;
551 			u32 allowed = lobj->allowed_domains;
552 			u32 current_domain =
553 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
554 
555 			/* Check if this buffer will be moved and don't move it
556 			 * if we have moved too many buffers for this IB already.
557 			 *
558 			 * Note that this allows moving at least one buffer of
559 			 * any size, because it doesn't take the current "bo"
560 			 * into account. We don't want to disallow buffer moves
561 			 * completely.
562 			 */
563 			if ((allowed & current_domain) != 0 &&
564 			    (domain & current_domain) == 0 && /* will be moved */
565 			    bytes_moved > bytes_moved_threshold) {
566 				/* don't move it */
567 				domain = current_domain;
568 			}
569 
570 		retry:
571 			radeon_ttm_placement_from_domain(bo, domain);
572 			if (ring == R600_RING_TYPE_UVD_INDEX)
573 				radeon_uvd_force_into_uvd_segment(bo, allowed);
574 
575 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
576 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
577 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
578 				       initial_bytes_moved;
579 
580 			if (unlikely(r)) {
581 				if (r != -ERESTARTSYS &&
582 				    domain != lobj->allowed_domains) {
583 					domain = lobj->allowed_domains;
584 					goto retry;
585 				}
586 				ttm_eu_backoff_reservation(ticket, head);
587 				return r;
588 			}
589 		}
590 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
591 		lobj->tiling_flags = bo->tiling_flags;
592 	}
593 
594 	list_for_each_entry(lobj, &duplicates, tv.head) {
595 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
596 		lobj->tiling_flags = lobj->robj->tiling_flags;
597 	}
598 
599 	return 0;
600 }
601 
602 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
603 {
604 	struct radeon_device *rdev = bo->rdev;
605 	struct radeon_surface_reg *reg;
606 	struct radeon_bo *old_object;
607 	int steal;
608 	int i;
609 
610 	lockdep_assert_held(&bo->tbo.resv->lock.base);
611 
612 	if (!bo->tiling_flags)
613 		return 0;
614 
615 	if (bo->surface_reg >= 0) {
616 		reg = &rdev->surface_regs[bo->surface_reg];
617 		i = bo->surface_reg;
618 		goto out;
619 	}
620 
621 	steal = -1;
622 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
623 
624 		reg = &rdev->surface_regs[i];
625 		if (!reg->bo)
626 			break;
627 
628 		old_object = reg->bo;
629 		if (old_object->pin_count == 0)
630 			steal = i;
631 	}
632 
633 	/* if we are all out */
634 	if (i == RADEON_GEM_MAX_SURFACES) {
635 		if (steal == -1)
636 			return -ENOMEM;
637 		/* find someone with a surface reg and nuke their BO */
638 		reg = &rdev->surface_regs[steal];
639 		old_object = reg->bo;
640 		/* blow away the mapping */
641 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
642 		ttm_bo_unmap_virtual(&old_object->tbo);
643 		old_object->surface_reg = -1;
644 		i = steal;
645 	}
646 
647 	bo->surface_reg = i;
648 	reg->bo = bo;
649 
650 out:
651 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
652 			       bo->tbo.mem.start << PAGE_SHIFT,
653 			       bo->tbo.num_pages << PAGE_SHIFT);
654 	return 0;
655 }
656 
657 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
658 {
659 	struct radeon_device *rdev = bo->rdev;
660 	struct radeon_surface_reg *reg;
661 
662 	if (bo->surface_reg == -1)
663 		return;
664 
665 	reg = &rdev->surface_regs[bo->surface_reg];
666 	radeon_clear_surface_reg(rdev, bo->surface_reg);
667 
668 	reg->bo = NULL;
669 	bo->surface_reg = -1;
670 }
671 
672 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
673 				uint32_t tiling_flags, uint32_t pitch)
674 {
675 	struct radeon_device *rdev = bo->rdev;
676 	int r;
677 
678 	if (rdev->family >= CHIP_CEDAR) {
679 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
680 
681 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
682 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
683 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
684 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
685 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
686 		switch (bankw) {
687 		case 0:
688 		case 1:
689 		case 2:
690 		case 4:
691 		case 8:
692 			break;
693 		default:
694 			return -EINVAL;
695 		}
696 		switch (bankh) {
697 		case 0:
698 		case 1:
699 		case 2:
700 		case 4:
701 		case 8:
702 			break;
703 		default:
704 			return -EINVAL;
705 		}
706 		switch (mtaspect) {
707 		case 0:
708 		case 1:
709 		case 2:
710 		case 4:
711 		case 8:
712 			break;
713 		default:
714 			return -EINVAL;
715 		}
716 		if (tilesplit > 6) {
717 			return -EINVAL;
718 		}
719 		if (stilesplit > 6) {
720 			return -EINVAL;
721 		}
722 	}
723 	r = radeon_bo_reserve(bo, false);
724 	if (unlikely(r != 0))
725 		return r;
726 	bo->tiling_flags = tiling_flags;
727 	bo->pitch = pitch;
728 	radeon_bo_unreserve(bo);
729 	return 0;
730 }
731 
732 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
733 				uint32_t *tiling_flags,
734 				uint32_t *pitch)
735 {
736 	lockdep_assert_held(&bo->tbo.resv->lock.base);
737 
738 	if (tiling_flags)
739 		*tiling_flags = bo->tiling_flags;
740 	if (pitch)
741 		*pitch = bo->pitch;
742 }
743 
744 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
745 				bool force_drop)
746 {
747 	if (!force_drop)
748 		lockdep_assert_held(&bo->tbo.resv->lock.base);
749 
750 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
751 		return 0;
752 
753 	if (force_drop) {
754 		radeon_bo_clear_surface_reg(bo);
755 		return 0;
756 	}
757 
758 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
759 		if (!has_moved)
760 			return 0;
761 
762 		if (bo->surface_reg >= 0)
763 			radeon_bo_clear_surface_reg(bo);
764 		return 0;
765 	}
766 
767 	if ((bo->surface_reg >= 0) && !has_moved)
768 		return 0;
769 
770 	return radeon_bo_get_surface_reg(bo);
771 }
772 
773 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
774 			   bool evict,
775 			   struct ttm_mem_reg *new_mem)
776 {
777 	struct radeon_bo *rbo;
778 
779 	if (!radeon_ttm_bo_is_radeon_bo(bo))
780 		return;
781 
782 	rbo = container_of(bo, struct radeon_bo, tbo);
783 	radeon_bo_check_tiling(rbo, 0, 1);
784 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
785 
786 	/* update statistics */
787 	if (!new_mem)
788 		return;
789 
790 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
791 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
792 }
793 
794 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
795 {
796 	struct ttm_operation_ctx ctx = { false, false };
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, &ctx);
829 	if (unlikely(r == -ENOMEM)) {
830 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
831 		return ttm_bo_validate(bo, &rbo->placement, &ctx);
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, NULL);
849 	if (unlikely(r != 0))
850 		return r;
851 	if (mem_type)
852 		*mem_type = bo->tbo.mem.mem_type;
853 
854 	r = ttm_bo_wait(&bo->tbo, true, no_wait);
855 	ttm_bo_unreserve(&bo->tbo);
856 	return r;
857 }
858 
859 /**
860  * radeon_bo_fence - add fence to buffer object
861  *
862  * @bo: buffer object in question
863  * @fence: fence to add
864  * @shared: true if fence should be added shared
865  *
866  */
867 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
868 		     bool shared)
869 {
870 	struct reservation_object *resv = bo->tbo.resv;
871 
872 	if (shared)
873 		reservation_object_add_shared_fence(resv, &fence->base);
874 	else
875 		reservation_object_add_excl_fence(resv, &fence->base);
876 }
877