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 
33 #include <linux/io.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36 
37 #include <drm/drm_cache.h>
38 #include <drm/drm_prime.h>
39 #include <drm/radeon_drm.h>
40 
41 #include "radeon.h"
42 #include "radeon_trace.h"
43 #include "radeon_ttm.h"
44 
45 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
46 
47 /*
48  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
49  * function are calling it.
50  */
51 
52 static void radeon_update_memory_usage(struct radeon_bo *bo,
53 				       unsigned mem_type, int sign)
54 {
55 	struct radeon_device *rdev = bo->rdev;
56 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
57 
58 	switch (mem_type) {
59 	case TTM_PL_TT:
60 		if (sign > 0)
61 			atomic64_add(size, &rdev->gtt_usage);
62 		else
63 			atomic64_sub(size, &rdev->gtt_usage);
64 		break;
65 	case TTM_PL_VRAM:
66 		if (sign > 0)
67 			atomic64_add(size, &rdev->vram_usage);
68 		else
69 			atomic64_sub(size, &rdev->vram_usage);
70 		break;
71 	}
72 }
73 
74 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
75 {
76 	struct radeon_bo *bo;
77 
78 	bo = container_of(tbo, struct radeon_bo, tbo);
79 
80 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
81 
82 	mutex_lock(&bo->rdev->gem.mutex);
83 	list_del_init(&bo->list);
84 	mutex_unlock(&bo->rdev->gem.mutex);
85 	radeon_bo_clear_surface_reg(bo);
86 	WARN_ON_ONCE(!list_empty(&bo->va));
87 	if (bo->tbo.base.import_attach)
88 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
89 	drm_gem_object_release(&bo->tbo.base);
90 	kfree(bo);
91 }
92 
93 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
94 {
95 	if (bo->destroy == &radeon_ttm_bo_destroy)
96 		return true;
97 	return false;
98 }
99 
100 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
101 {
102 	u32 c = 0, i;
103 
104 	rbo->placement.placement = rbo->placements;
105 	rbo->placement.busy_placement = rbo->placements;
106 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
107 		/* Try placing BOs which don't need CPU access outside of the
108 		 * CPU accessible part of VRAM
109 		 */
110 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
111 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
112 			rbo->placements[c].fpfn =
113 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
114 			rbo->placements[c].mem_type = TTM_PL_VRAM;
115 			rbo->placements[c++].flags = 0;
116 		}
117 
118 		rbo->placements[c].fpfn = 0;
119 		rbo->placements[c].mem_type = TTM_PL_VRAM;
120 		rbo->placements[c++].flags = 0;
121 	}
122 
123 	if (domain & RADEON_GEM_DOMAIN_GTT) {
124 		rbo->placements[c].fpfn = 0;
125 		rbo->placements[c].mem_type = TTM_PL_TT;
126 		rbo->placements[c++].flags = 0;
127 	}
128 
129 	if (domain & RADEON_GEM_DOMAIN_CPU) {
130 		rbo->placements[c].fpfn = 0;
131 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
132 		rbo->placements[c++].flags = 0;
133 	}
134 	if (!c) {
135 		rbo->placements[c].fpfn = 0;
136 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
137 		rbo->placements[c++].flags = 0;
138 	}
139 
140 	rbo->placement.num_placement = c;
141 	rbo->placement.num_busy_placement = c;
142 
143 	for (i = 0; i < c; ++i) {
144 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
145 		    (rbo->placements[i].mem_type == TTM_PL_VRAM) &&
146 		    !rbo->placements[i].fpfn)
147 			rbo->placements[i].lpfn =
148 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
149 		else
150 			rbo->placements[i].lpfn = 0;
151 	}
152 }
153 
154 int radeon_bo_create(struct radeon_device *rdev,
155 		     unsigned long size, int byte_align, bool kernel,
156 		     u32 domain, u32 flags, struct sg_table *sg,
157 		     struct dma_resv *resv,
158 		     struct radeon_bo **bo_ptr)
159 {
160 	struct radeon_bo *bo;
161 	enum ttm_bo_type type;
162 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
163 	size_t acc_size;
164 	int r;
165 
166 	size = ALIGN(size, PAGE_SIZE);
167 
168 	if (kernel) {
169 		type = ttm_bo_type_kernel;
170 	} else if (sg) {
171 		type = ttm_bo_type_sg;
172 	} else {
173 		type = ttm_bo_type_device;
174 	}
175 	*bo_ptr = NULL;
176 
177 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
178 				       sizeof(struct radeon_bo));
179 
180 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
181 	if (bo == NULL)
182 		return -ENOMEM;
183 	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
184 	bo->rdev = rdev;
185 	bo->surface_reg = -1;
186 	INIT_LIST_HEAD(&bo->list);
187 	INIT_LIST_HEAD(&bo->va);
188 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
189 				       RADEON_GEM_DOMAIN_GTT |
190 				       RADEON_GEM_DOMAIN_CPU);
191 
192 	bo->flags = flags;
193 	/* PCI GART is always snooped */
194 	if (!(rdev->flags & RADEON_IS_PCIE))
195 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
196 
197 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
198 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
199 	 */
200 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
201 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
202 
203 #ifdef CONFIG_X86_32
204 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
205 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
206 	 */
207 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
208 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
209 	/* Don't try to enable write-combining when it can't work, or things
210 	 * may be slow
211 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
212 	 */
213 #ifndef CONFIG_COMPILE_TEST
214 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
215 	 thanks to write-combining
216 #endif
217 
218 	if (bo->flags & RADEON_GEM_GTT_WC)
219 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
220 			      "better performance thanks to write-combining\n");
221 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
222 #else
223 	/* For architectures that don't support WC memory,
224 	 * mask out the WC flag from the BO
225 	 */
226 	if (!drm_arch_can_wc_memory())
227 		bo->flags &= ~RADEON_GEM_GTT_WC;
228 #endif
229 
230 	radeon_ttm_placement_from_domain(bo, domain);
231 	/* Kernel allocation are uninterruptible */
232 	down_read(&rdev->pm.mclk_lock);
233 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
234 			&bo->placement, page_align, !kernel, acc_size,
235 			sg, resv, &radeon_ttm_bo_destroy);
236 	up_read(&rdev->pm.mclk_lock);
237 	if (unlikely(r != 0)) {
238 		return r;
239 	}
240 	*bo_ptr = bo;
241 
242 	trace_radeon_bo_create(bo);
243 
244 	return 0;
245 }
246 
247 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
248 {
249 	bool is_iomem;
250 	int r;
251 
252 	if (bo->kptr) {
253 		if (ptr) {
254 			*ptr = bo->kptr;
255 		}
256 		return 0;
257 	}
258 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
259 	if (r) {
260 		return r;
261 	}
262 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
263 	if (ptr) {
264 		*ptr = bo->kptr;
265 	}
266 	radeon_bo_check_tiling(bo, 0, 0);
267 	return 0;
268 }
269 
270 void radeon_bo_kunmap(struct radeon_bo *bo)
271 {
272 	if (bo->kptr == NULL)
273 		return;
274 	bo->kptr = NULL;
275 	radeon_bo_check_tiling(bo, 0, 0);
276 	ttm_bo_kunmap(&bo->kmap);
277 }
278 
279 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
280 {
281 	if (bo == NULL)
282 		return NULL;
283 
284 	ttm_bo_get(&bo->tbo);
285 	return bo;
286 }
287 
288 void radeon_bo_unref(struct radeon_bo **bo)
289 {
290 	struct ttm_buffer_object *tbo;
291 
292 	if ((*bo) == NULL)
293 		return;
294 	tbo = &((*bo)->tbo);
295 	ttm_bo_put(tbo);
296 	*bo = NULL;
297 }
298 
299 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
300 			     u64 *gpu_addr)
301 {
302 	struct ttm_operation_ctx ctx = { false, false };
303 	int r, i;
304 
305 	if (radeon_ttm_tt_has_userptr(bo->rdev, bo->tbo.ttm))
306 		return -EPERM;
307 
308 	if (bo->tbo.pin_count) {
309 		ttm_bo_pin(&bo->tbo);
310 		if (gpu_addr)
311 			*gpu_addr = radeon_bo_gpu_offset(bo);
312 
313 		if (max_offset != 0) {
314 			u64 domain_start;
315 
316 			if (domain == RADEON_GEM_DOMAIN_VRAM)
317 				domain_start = bo->rdev->mc.vram_start;
318 			else
319 				domain_start = bo->rdev->mc.gtt_start;
320 			WARN_ON_ONCE(max_offset <
321 				     (radeon_bo_gpu_offset(bo) - domain_start));
322 		}
323 
324 		return 0;
325 	}
326 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
327 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
328 		return -EINVAL;
329 	}
330 
331 	radeon_ttm_placement_from_domain(bo, domain);
332 	for (i = 0; i < bo->placement.num_placement; i++) {
333 		/* force to pin into visible video ram */
334 		if ((bo->placements[i].mem_type == TTM_PL_VRAM) &&
335 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
336 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
337 			bo->placements[i].lpfn =
338 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
339 		else
340 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
341 	}
342 
343 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
344 	if (likely(r == 0)) {
345 		ttm_bo_pin(&bo->tbo);
346 		if (gpu_addr != NULL)
347 			*gpu_addr = radeon_bo_gpu_offset(bo);
348 		if (domain == RADEON_GEM_DOMAIN_VRAM)
349 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
350 		else
351 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
352 	} else {
353 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
354 	}
355 	return r;
356 }
357 
358 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
359 {
360 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
361 }
362 
363 void radeon_bo_unpin(struct radeon_bo *bo)
364 {
365 	ttm_bo_unpin(&bo->tbo);
366 	if (!bo->tbo.pin_count) {
367 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
368 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
369 		else
370 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
371 	}
372 }
373 
374 int radeon_bo_evict_vram(struct radeon_device *rdev)
375 {
376 	struct ttm_bo_device *bdev = &rdev->mman.bdev;
377 	struct ttm_resource_manager *man;
378 
379 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
380 #ifndef CONFIG_HIBERNATION
381 	if (rdev->flags & RADEON_IS_IGP) {
382 		if (rdev->mc.igp_sideport_enabled == false)
383 			/* Useless to evict on IGP chips */
384 			return 0;
385 	}
386 #endif
387 	man = ttm_manager_type(bdev, TTM_PL_VRAM);
388 	return ttm_resource_manager_evict_all(bdev, man);
389 }
390 
391 void radeon_bo_force_delete(struct radeon_device *rdev)
392 {
393 	struct radeon_bo *bo, *n;
394 
395 	if (list_empty(&rdev->gem.objects)) {
396 		return;
397 	}
398 	dev_err(rdev->dev, "Userspace still has active objects !\n");
399 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
400 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
401 			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
402 			*((unsigned long *)&bo->tbo.base.refcount));
403 		mutex_lock(&bo->rdev->gem.mutex);
404 		list_del_init(&bo->list);
405 		mutex_unlock(&bo->rdev->gem.mutex);
406 		/* this should unref the ttm bo */
407 		drm_gem_object_put(&bo->tbo.base);
408 	}
409 }
410 
411 int radeon_bo_init(struct radeon_device *rdev)
412 {
413 	/* reserve PAT memory space to WC for VRAM */
414 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
415 				   rdev->mc.aper_size);
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 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
435 }
436 
437 /* Returns how many bytes TTM can move per IB.
438  */
439 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
440 {
441 	u64 real_vram_size = rdev->mc.real_vram_size;
442 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
443 
444 	/* This function is based on the current VRAM usage.
445 	 *
446 	 * - If all of VRAM is free, allow relocating the number of bytes that
447 	 *   is equal to 1/4 of the size of VRAM for this IB.
448 
449 	 * - If more than one half of VRAM is occupied, only allow relocating
450 	 *   1 MB of data for this IB.
451 	 *
452 	 * - From 0 to one half of used VRAM, the threshold decreases
453 	 *   linearly.
454 	 *         __________________
455 	 * 1/4 of -|\               |
456 	 * VRAM    | \              |
457 	 *         |  \             |
458 	 *         |   \            |
459 	 *         |    \           |
460 	 *         |     \          |
461 	 *         |      \         |
462 	 *         |       \________|1 MB
463 	 *         |----------------|
464 	 *    VRAM 0 %             100 %
465 	 *         used            used
466 	 *
467 	 * Note: It's a threshold, not a limit. The threshold must be crossed
468 	 * for buffer relocations to stop, so any buffer of an arbitrary size
469 	 * can be moved as long as the threshold isn't crossed before
470 	 * the relocation takes place. We don't want to disable buffer
471 	 * relocations completely.
472 	 *
473 	 * The idea is that buffers should be placed in VRAM at creation time
474 	 * and TTM should only do a minimum number of relocations during
475 	 * command submission. In practice, you need to submit at least
476 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
477 	 *
478 	 * Also, things can get pretty crazy under memory pressure and actual
479 	 * VRAM usage can change a lot, so playing safe even at 50% does
480 	 * consistently increase performance.
481 	 */
482 
483 	u64 half_vram = real_vram_size >> 1;
484 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
485 	u64 bytes_moved_threshold = half_free_vram >> 1;
486 	return max(bytes_moved_threshold, 1024*1024ull);
487 }
488 
489 int radeon_bo_list_validate(struct radeon_device *rdev,
490 			    struct ww_acquire_ctx *ticket,
491 			    struct list_head *head, int ring)
492 {
493 	struct ttm_operation_ctx ctx = { true, false };
494 	struct radeon_bo_list *lobj;
495 	struct list_head duplicates;
496 	int r;
497 	u64 bytes_moved = 0, initial_bytes_moved;
498 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
499 
500 	INIT_LIST_HEAD(&duplicates);
501 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
502 	if (unlikely(r != 0)) {
503 		return r;
504 	}
505 
506 	list_for_each_entry(lobj, head, tv.head) {
507 		struct radeon_bo *bo = lobj->robj;
508 		if (!bo->tbo.pin_count) {
509 			u32 domain = lobj->preferred_domains;
510 			u32 allowed = lobj->allowed_domains;
511 			u32 current_domain =
512 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
513 
514 			/* Check if this buffer will be moved and don't move it
515 			 * if we have moved too many buffers for this IB already.
516 			 *
517 			 * Note that this allows moving at least one buffer of
518 			 * any size, because it doesn't take the current "bo"
519 			 * into account. We don't want to disallow buffer moves
520 			 * completely.
521 			 */
522 			if ((allowed & current_domain) != 0 &&
523 			    (domain & current_domain) == 0 && /* will be moved */
524 			    bytes_moved > bytes_moved_threshold) {
525 				/* don't move it */
526 				domain = current_domain;
527 			}
528 
529 		retry:
530 			radeon_ttm_placement_from_domain(bo, domain);
531 			if (ring == R600_RING_TYPE_UVD_INDEX)
532 				radeon_uvd_force_into_uvd_segment(bo, allowed);
533 
534 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
535 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
536 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
537 				       initial_bytes_moved;
538 
539 			if (unlikely(r)) {
540 				if (r != -ERESTARTSYS &&
541 				    domain != lobj->allowed_domains) {
542 					domain = lobj->allowed_domains;
543 					goto retry;
544 				}
545 				ttm_eu_backoff_reservation(ticket, head);
546 				return r;
547 			}
548 		}
549 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
550 		lobj->tiling_flags = bo->tiling_flags;
551 	}
552 
553 	list_for_each_entry(lobj, &duplicates, tv.head) {
554 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
555 		lobj->tiling_flags = lobj->robj->tiling_flags;
556 	}
557 
558 	return 0;
559 }
560 
561 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
562 {
563 	struct radeon_device *rdev = bo->rdev;
564 	struct radeon_surface_reg *reg;
565 	struct radeon_bo *old_object;
566 	int steal;
567 	int i;
568 
569 	dma_resv_assert_held(bo->tbo.base.resv);
570 
571 	if (!bo->tiling_flags)
572 		return 0;
573 
574 	if (bo->surface_reg >= 0) {
575 		reg = &rdev->surface_regs[bo->surface_reg];
576 		i = bo->surface_reg;
577 		goto out;
578 	}
579 
580 	steal = -1;
581 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
582 
583 		reg = &rdev->surface_regs[i];
584 		if (!reg->bo)
585 			break;
586 
587 		old_object = reg->bo;
588 		if (old_object->tbo.pin_count == 0)
589 			steal = i;
590 	}
591 
592 	/* if we are all out */
593 	if (i == RADEON_GEM_MAX_SURFACES) {
594 		if (steal == -1)
595 			return -ENOMEM;
596 		/* find someone with a surface reg and nuke their BO */
597 		reg = &rdev->surface_regs[steal];
598 		old_object = reg->bo;
599 		/* blow away the mapping */
600 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
601 		ttm_bo_unmap_virtual(&old_object->tbo);
602 		old_object->surface_reg = -1;
603 		i = steal;
604 	}
605 
606 	bo->surface_reg = i;
607 	reg->bo = bo;
608 
609 out:
610 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
611 			       bo->tbo.mem.start << PAGE_SHIFT,
612 			       bo->tbo.num_pages << PAGE_SHIFT);
613 	return 0;
614 }
615 
616 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
617 {
618 	struct radeon_device *rdev = bo->rdev;
619 	struct radeon_surface_reg *reg;
620 
621 	if (bo->surface_reg == -1)
622 		return;
623 
624 	reg = &rdev->surface_regs[bo->surface_reg];
625 	radeon_clear_surface_reg(rdev, bo->surface_reg);
626 
627 	reg->bo = NULL;
628 	bo->surface_reg = -1;
629 }
630 
631 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
632 				uint32_t tiling_flags, uint32_t pitch)
633 {
634 	struct radeon_device *rdev = bo->rdev;
635 	int r;
636 
637 	if (rdev->family >= CHIP_CEDAR) {
638 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
639 
640 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
641 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
642 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
643 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
644 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
645 		switch (bankw) {
646 		case 0:
647 		case 1:
648 		case 2:
649 		case 4:
650 		case 8:
651 			break;
652 		default:
653 			return -EINVAL;
654 		}
655 		switch (bankh) {
656 		case 0:
657 		case 1:
658 		case 2:
659 		case 4:
660 		case 8:
661 			break;
662 		default:
663 			return -EINVAL;
664 		}
665 		switch (mtaspect) {
666 		case 0:
667 		case 1:
668 		case 2:
669 		case 4:
670 		case 8:
671 			break;
672 		default:
673 			return -EINVAL;
674 		}
675 		if (tilesplit > 6) {
676 			return -EINVAL;
677 		}
678 		if (stilesplit > 6) {
679 			return -EINVAL;
680 		}
681 	}
682 	r = radeon_bo_reserve(bo, false);
683 	if (unlikely(r != 0))
684 		return r;
685 	bo->tiling_flags = tiling_flags;
686 	bo->pitch = pitch;
687 	radeon_bo_unreserve(bo);
688 	return 0;
689 }
690 
691 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
692 				uint32_t *tiling_flags,
693 				uint32_t *pitch)
694 {
695 	dma_resv_assert_held(bo->tbo.base.resv);
696 
697 	if (tiling_flags)
698 		*tiling_flags = bo->tiling_flags;
699 	if (pitch)
700 		*pitch = bo->pitch;
701 }
702 
703 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
704 				bool force_drop)
705 {
706 	if (!force_drop)
707 		dma_resv_assert_held(bo->tbo.base.resv);
708 
709 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
710 		return 0;
711 
712 	if (force_drop) {
713 		radeon_bo_clear_surface_reg(bo);
714 		return 0;
715 	}
716 
717 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
718 		if (!has_moved)
719 			return 0;
720 
721 		if (bo->surface_reg >= 0)
722 			radeon_bo_clear_surface_reg(bo);
723 		return 0;
724 	}
725 
726 	if ((bo->surface_reg >= 0) && !has_moved)
727 		return 0;
728 
729 	return radeon_bo_get_surface_reg(bo);
730 }
731 
732 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
733 			   bool evict,
734 			   struct ttm_resource *new_mem)
735 {
736 	struct radeon_bo *rbo;
737 
738 	if (!radeon_ttm_bo_is_radeon_bo(bo))
739 		return;
740 
741 	rbo = container_of(bo, struct radeon_bo, tbo);
742 	radeon_bo_check_tiling(rbo, 0, 1);
743 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
744 
745 	/* update statistics */
746 	if (!new_mem)
747 		return;
748 
749 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
750 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
751 }
752 
753 vm_fault_t radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
754 {
755 	struct ttm_operation_ctx ctx = { false, false };
756 	struct radeon_device *rdev;
757 	struct radeon_bo *rbo;
758 	unsigned long offset, size, lpfn;
759 	int i, r;
760 
761 	if (!radeon_ttm_bo_is_radeon_bo(bo))
762 		return 0;
763 	rbo = container_of(bo, struct radeon_bo, tbo);
764 	radeon_bo_check_tiling(rbo, 0, 0);
765 	rdev = rbo->rdev;
766 	if (bo->mem.mem_type != TTM_PL_VRAM)
767 		return 0;
768 
769 	size = bo->mem.num_pages << PAGE_SHIFT;
770 	offset = bo->mem.start << PAGE_SHIFT;
771 	if ((offset + size) <= rdev->mc.visible_vram_size)
772 		return 0;
773 
774 	/* Can't move a pinned BO to visible VRAM */
775 	if (rbo->tbo.pin_count > 0)
776 		return VM_FAULT_SIGBUS;
777 
778 	/* hurrah the memory is not visible ! */
779 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
780 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
781 	for (i = 0; i < rbo->placement.num_placement; i++) {
782 		/* Force into visible VRAM */
783 		if ((rbo->placements[i].mem_type == TTM_PL_VRAM) &&
784 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
785 			rbo->placements[i].lpfn = lpfn;
786 	}
787 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
788 	if (unlikely(r == -ENOMEM)) {
789 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
790 		r = ttm_bo_validate(bo, &rbo->placement, &ctx);
791 	} else if (likely(!r)) {
792 		offset = bo->mem.start << PAGE_SHIFT;
793 		/* this should never happen */
794 		if ((offset + size) > rdev->mc.visible_vram_size)
795 			return VM_FAULT_SIGBUS;
796 	}
797 
798 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
799 		return VM_FAULT_NOPAGE;
800 	else if (unlikely(r))
801 		return VM_FAULT_SIGBUS;
802 
803 	ttm_bo_move_to_lru_tail_unlocked(bo);
804 	return 0;
805 }
806 
807 /**
808  * radeon_bo_fence - add fence to buffer object
809  *
810  * @bo: buffer object in question
811  * @fence: fence to add
812  * @shared: true if fence should be added shared
813  *
814  */
815 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
816 		     bool shared)
817 {
818 	struct dma_resv *resv = bo->tbo.base.resv;
819 
820 	if (shared)
821 		dma_resv_add_shared_fence(resv, &fence->base);
822 	else
823 		dma_resv_add_excl_fence(resv, &fence->base);
824 }
825