xref: /linux/drivers/gpu/drm/radeon/radeon_object.c (revision 34f7c6e7d4396090692a09789db231e12cb4762b)
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_ttm_bo_destroy(struct ttm_buffer_object *tbo)
53 {
54 	struct radeon_bo *bo;
55 
56 	bo = container_of(tbo, struct radeon_bo, tbo);
57 
58 	mutex_lock(&bo->rdev->gem.mutex);
59 	list_del_init(&bo->list);
60 	mutex_unlock(&bo->rdev->gem.mutex);
61 	radeon_bo_clear_surface_reg(bo);
62 	WARN_ON_ONCE(!list_empty(&bo->va));
63 	if (bo->tbo.base.import_attach)
64 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
65 	drm_gem_object_release(&bo->tbo.base);
66 	kfree(bo);
67 }
68 
69 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
70 {
71 	if (bo->destroy == &radeon_ttm_bo_destroy)
72 		return true;
73 	return false;
74 }
75 
76 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
77 {
78 	u32 c = 0, i;
79 
80 	rbo->placement.placement = rbo->placements;
81 	rbo->placement.busy_placement = rbo->placements;
82 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
83 		/* Try placing BOs which don't need CPU access outside of the
84 		 * CPU accessible part of VRAM
85 		 */
86 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
87 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
88 			rbo->placements[c].fpfn =
89 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
90 			rbo->placements[c].mem_type = TTM_PL_VRAM;
91 			rbo->placements[c++].flags = 0;
92 		}
93 
94 		rbo->placements[c].fpfn = 0;
95 		rbo->placements[c].mem_type = TTM_PL_VRAM;
96 		rbo->placements[c++].flags = 0;
97 	}
98 
99 	if (domain & RADEON_GEM_DOMAIN_GTT) {
100 		rbo->placements[c].fpfn = 0;
101 		rbo->placements[c].mem_type = TTM_PL_TT;
102 		rbo->placements[c++].flags = 0;
103 	}
104 
105 	if (domain & RADEON_GEM_DOMAIN_CPU) {
106 		rbo->placements[c].fpfn = 0;
107 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
108 		rbo->placements[c++].flags = 0;
109 	}
110 	if (!c) {
111 		rbo->placements[c].fpfn = 0;
112 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
113 		rbo->placements[c++].flags = 0;
114 	}
115 
116 	rbo->placement.num_placement = c;
117 	rbo->placement.num_busy_placement = c;
118 
119 	for (i = 0; i < c; ++i) {
120 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
121 		    (rbo->placements[i].mem_type == TTM_PL_VRAM) &&
122 		    !rbo->placements[i].fpfn)
123 			rbo->placements[i].lpfn =
124 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
125 		else
126 			rbo->placements[i].lpfn = 0;
127 	}
128 }
129 
130 int radeon_bo_create(struct radeon_device *rdev,
131 		     unsigned long size, int byte_align, bool kernel,
132 		     u32 domain, u32 flags, struct sg_table *sg,
133 		     struct dma_resv *resv,
134 		     struct radeon_bo **bo_ptr)
135 {
136 	struct radeon_bo *bo;
137 	enum ttm_bo_type type;
138 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
139 	int r;
140 
141 	size = ALIGN(size, PAGE_SIZE);
142 
143 	if (kernel) {
144 		type = ttm_bo_type_kernel;
145 	} else if (sg) {
146 		type = ttm_bo_type_sg;
147 	} else {
148 		type = ttm_bo_type_device;
149 	}
150 	*bo_ptr = NULL;
151 
152 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
153 	if (bo == NULL)
154 		return -ENOMEM;
155 	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
156 	bo->rdev = rdev;
157 	bo->surface_reg = -1;
158 	INIT_LIST_HEAD(&bo->list);
159 	INIT_LIST_HEAD(&bo->va);
160 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
161 				       RADEON_GEM_DOMAIN_GTT |
162 				       RADEON_GEM_DOMAIN_CPU);
163 
164 	bo->flags = flags;
165 	/* PCI GART is always snooped */
166 	if (!(rdev->flags & RADEON_IS_PCIE))
167 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
168 
169 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
170 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
171 	 */
172 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
173 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
174 
175 #ifdef CONFIG_X86_32
176 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
177 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
178 	 */
179 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
180 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
181 	/* Don't try to enable write-combining when it can't work, or things
182 	 * may be slow
183 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
184 	 */
185 #ifndef CONFIG_COMPILE_TEST
186 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
187 	 thanks to write-combining
188 #endif
189 
190 	if (bo->flags & RADEON_GEM_GTT_WC)
191 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
192 			      "better performance thanks to write-combining\n");
193 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
194 #else
195 	/* For architectures that don't support WC memory,
196 	 * mask out the WC flag from the BO
197 	 */
198 	if (!drm_arch_can_wc_memory())
199 		bo->flags &= ~RADEON_GEM_GTT_WC;
200 #endif
201 
202 	radeon_ttm_placement_from_domain(bo, domain);
203 	/* Kernel allocation are uninterruptible */
204 	down_read(&rdev->pm.mclk_lock);
205 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
206 			&bo->placement, page_align, !kernel, sg, resv,
207 			&radeon_ttm_bo_destroy);
208 	up_read(&rdev->pm.mclk_lock);
209 	if (unlikely(r != 0)) {
210 		return r;
211 	}
212 	*bo_ptr = bo;
213 
214 	trace_radeon_bo_create(bo);
215 
216 	return 0;
217 }
218 
219 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
220 {
221 	bool is_iomem;
222 	int r;
223 
224 	if (bo->kptr) {
225 		if (ptr) {
226 			*ptr = bo->kptr;
227 		}
228 		return 0;
229 	}
230 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
231 	if (r) {
232 		return r;
233 	}
234 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
235 	if (ptr) {
236 		*ptr = bo->kptr;
237 	}
238 	radeon_bo_check_tiling(bo, 0, 0);
239 	return 0;
240 }
241 
242 void radeon_bo_kunmap(struct radeon_bo *bo)
243 {
244 	if (bo->kptr == NULL)
245 		return;
246 	bo->kptr = NULL;
247 	radeon_bo_check_tiling(bo, 0, 0);
248 	ttm_bo_kunmap(&bo->kmap);
249 }
250 
251 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
252 {
253 	if (bo == NULL)
254 		return NULL;
255 
256 	ttm_bo_get(&bo->tbo);
257 	return bo;
258 }
259 
260 void radeon_bo_unref(struct radeon_bo **bo)
261 {
262 	struct ttm_buffer_object *tbo;
263 
264 	if ((*bo) == NULL)
265 		return;
266 	tbo = &((*bo)->tbo);
267 	ttm_bo_put(tbo);
268 	*bo = NULL;
269 }
270 
271 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
272 			     u64 *gpu_addr)
273 {
274 	struct ttm_operation_ctx ctx = { false, false };
275 	int r, i;
276 
277 	if (radeon_ttm_tt_has_userptr(bo->rdev, bo->tbo.ttm))
278 		return -EPERM;
279 
280 	if (bo->tbo.pin_count) {
281 		ttm_bo_pin(&bo->tbo);
282 		if (gpu_addr)
283 			*gpu_addr = radeon_bo_gpu_offset(bo);
284 
285 		if (max_offset != 0) {
286 			u64 domain_start;
287 
288 			if (domain == RADEON_GEM_DOMAIN_VRAM)
289 				domain_start = bo->rdev->mc.vram_start;
290 			else
291 				domain_start = bo->rdev->mc.gtt_start;
292 			WARN_ON_ONCE(max_offset <
293 				     (radeon_bo_gpu_offset(bo) - domain_start));
294 		}
295 
296 		return 0;
297 	}
298 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
299 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
300 		return -EINVAL;
301 	}
302 
303 	radeon_ttm_placement_from_domain(bo, domain);
304 	for (i = 0; i < bo->placement.num_placement; i++) {
305 		/* force to pin into visible video ram */
306 		if ((bo->placements[i].mem_type == TTM_PL_VRAM) &&
307 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
308 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
309 			bo->placements[i].lpfn =
310 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
311 		else
312 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
313 	}
314 
315 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
316 	if (likely(r == 0)) {
317 		ttm_bo_pin(&bo->tbo);
318 		if (gpu_addr != NULL)
319 			*gpu_addr = radeon_bo_gpu_offset(bo);
320 		if (domain == RADEON_GEM_DOMAIN_VRAM)
321 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
322 		else
323 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
324 	} else {
325 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
326 	}
327 	return r;
328 }
329 
330 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
331 {
332 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
333 }
334 
335 void radeon_bo_unpin(struct radeon_bo *bo)
336 {
337 	ttm_bo_unpin(&bo->tbo);
338 	if (!bo->tbo.pin_count) {
339 		if (bo->tbo.resource->mem_type == TTM_PL_VRAM)
340 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
341 		else
342 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
343 	}
344 }
345 
346 int radeon_bo_evict_vram(struct radeon_device *rdev)
347 {
348 	struct ttm_device *bdev = &rdev->mman.bdev;
349 	struct ttm_resource_manager *man;
350 
351 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
352 #ifndef CONFIG_HIBERNATION
353 	if (rdev->flags & RADEON_IS_IGP) {
354 		if (rdev->mc.igp_sideport_enabled == false)
355 			/* Useless to evict on IGP chips */
356 			return 0;
357 	}
358 #endif
359 	man = ttm_manager_type(bdev, TTM_PL_VRAM);
360 	if (!man)
361 		return 0;
362 	return ttm_resource_manager_evict_all(bdev, man);
363 }
364 
365 void radeon_bo_force_delete(struct radeon_device *rdev)
366 {
367 	struct radeon_bo *bo, *n;
368 
369 	if (list_empty(&rdev->gem.objects)) {
370 		return;
371 	}
372 	dev_err(rdev->dev, "Userspace still has active objects !\n");
373 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
374 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
375 			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
376 			*((unsigned long *)&bo->tbo.base.refcount));
377 		mutex_lock(&bo->rdev->gem.mutex);
378 		list_del_init(&bo->list);
379 		mutex_unlock(&bo->rdev->gem.mutex);
380 		/* this should unref the ttm bo */
381 		drm_gem_object_put(&bo->tbo.base);
382 	}
383 }
384 
385 int radeon_bo_init(struct radeon_device *rdev)
386 {
387 	/* reserve PAT memory space to WC for VRAM */
388 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
389 				   rdev->mc.aper_size);
390 
391 	/* Add an MTRR for the VRAM */
392 	if (!rdev->fastfb_working) {
393 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
394 						      rdev->mc.aper_size);
395 	}
396 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
397 		rdev->mc.mc_vram_size >> 20,
398 		(unsigned long long)rdev->mc.aper_size >> 20);
399 	DRM_INFO("RAM width %dbits %cDR\n",
400 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
401 	return radeon_ttm_init(rdev);
402 }
403 
404 void radeon_bo_fini(struct radeon_device *rdev)
405 {
406 	radeon_ttm_fini(rdev);
407 	arch_phys_wc_del(rdev->mc.vram_mtrr);
408 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
409 }
410 
411 /* Returns how many bytes TTM can move per IB.
412  */
413 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
414 {
415 	u64 real_vram_size = rdev->mc.real_vram_size;
416 	struct ttm_resource_manager *man =
417 		ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
418 	u64 vram_usage = ttm_resource_manager_usage(man);
419 
420 	/* This function is based on the current VRAM usage.
421 	 *
422 	 * - If all of VRAM is free, allow relocating the number of bytes that
423 	 *   is equal to 1/4 of the size of VRAM for this IB.
424 
425 	 * - If more than one half of VRAM is occupied, only allow relocating
426 	 *   1 MB of data for this IB.
427 	 *
428 	 * - From 0 to one half of used VRAM, the threshold decreases
429 	 *   linearly.
430 	 *         __________________
431 	 * 1/4 of -|\               |
432 	 * VRAM    | \              |
433 	 *         |  \             |
434 	 *         |   \            |
435 	 *         |    \           |
436 	 *         |     \          |
437 	 *         |      \         |
438 	 *         |       \________|1 MB
439 	 *         |----------------|
440 	 *    VRAM 0 %             100 %
441 	 *         used            used
442 	 *
443 	 * Note: It's a threshold, not a limit. The threshold must be crossed
444 	 * for buffer relocations to stop, so any buffer of an arbitrary size
445 	 * can be moved as long as the threshold isn't crossed before
446 	 * the relocation takes place. We don't want to disable buffer
447 	 * relocations completely.
448 	 *
449 	 * The idea is that buffers should be placed in VRAM at creation time
450 	 * and TTM should only do a minimum number of relocations during
451 	 * command submission. In practice, you need to submit at least
452 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
453 	 *
454 	 * Also, things can get pretty crazy under memory pressure and actual
455 	 * VRAM usage can change a lot, so playing safe even at 50% does
456 	 * consistently increase performance.
457 	 */
458 
459 	u64 half_vram = real_vram_size >> 1;
460 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
461 	u64 bytes_moved_threshold = half_free_vram >> 1;
462 	return max(bytes_moved_threshold, 1024*1024ull);
463 }
464 
465 int radeon_bo_list_validate(struct radeon_device *rdev,
466 			    struct ww_acquire_ctx *ticket,
467 			    struct list_head *head, int ring)
468 {
469 	struct ttm_operation_ctx ctx = { true, false };
470 	struct radeon_bo_list *lobj;
471 	struct list_head duplicates;
472 	int r;
473 	u64 bytes_moved = 0, initial_bytes_moved;
474 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
475 
476 	INIT_LIST_HEAD(&duplicates);
477 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
478 	if (unlikely(r != 0)) {
479 		return r;
480 	}
481 
482 	list_for_each_entry(lobj, head, tv.head) {
483 		struct radeon_bo *bo = lobj->robj;
484 		if (!bo->tbo.pin_count) {
485 			u32 domain = lobj->preferred_domains;
486 			u32 allowed = lobj->allowed_domains;
487 			u32 current_domain =
488 				radeon_mem_type_to_domain(bo->tbo.resource->mem_type);
489 
490 			/* Check if this buffer will be moved and don't move it
491 			 * if we have moved too many buffers for this IB already.
492 			 *
493 			 * Note that this allows moving at least one buffer of
494 			 * any size, because it doesn't take the current "bo"
495 			 * into account. We don't want to disallow buffer moves
496 			 * completely.
497 			 */
498 			if ((allowed & current_domain) != 0 &&
499 			    (domain & current_domain) == 0 && /* will be moved */
500 			    bytes_moved > bytes_moved_threshold) {
501 				/* don't move it */
502 				domain = current_domain;
503 			}
504 
505 		retry:
506 			radeon_ttm_placement_from_domain(bo, domain);
507 			if (ring == R600_RING_TYPE_UVD_INDEX)
508 				radeon_uvd_force_into_uvd_segment(bo, allowed);
509 
510 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
511 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
512 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
513 				       initial_bytes_moved;
514 
515 			if (unlikely(r)) {
516 				if (r != -ERESTARTSYS &&
517 				    domain != lobj->allowed_domains) {
518 					domain = lobj->allowed_domains;
519 					goto retry;
520 				}
521 				ttm_eu_backoff_reservation(ticket, head);
522 				return r;
523 			}
524 		}
525 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
526 		lobj->tiling_flags = bo->tiling_flags;
527 	}
528 
529 	list_for_each_entry(lobj, &duplicates, tv.head) {
530 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
531 		lobj->tiling_flags = lobj->robj->tiling_flags;
532 	}
533 
534 	return 0;
535 }
536 
537 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
538 {
539 	struct radeon_device *rdev = bo->rdev;
540 	struct radeon_surface_reg *reg;
541 	struct radeon_bo *old_object;
542 	int steal;
543 	int i;
544 
545 	dma_resv_assert_held(bo->tbo.base.resv);
546 
547 	if (!bo->tiling_flags)
548 		return 0;
549 
550 	if (bo->surface_reg >= 0) {
551 		i = bo->surface_reg;
552 		goto out;
553 	}
554 
555 	steal = -1;
556 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
557 
558 		reg = &rdev->surface_regs[i];
559 		if (!reg->bo)
560 			break;
561 
562 		old_object = reg->bo;
563 		if (old_object->tbo.pin_count == 0)
564 			steal = i;
565 	}
566 
567 	/* if we are all out */
568 	if (i == RADEON_GEM_MAX_SURFACES) {
569 		if (steal == -1)
570 			return -ENOMEM;
571 		/* find someone with a surface reg and nuke their BO */
572 		reg = &rdev->surface_regs[steal];
573 		old_object = reg->bo;
574 		/* blow away the mapping */
575 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
576 		ttm_bo_unmap_virtual(&old_object->tbo);
577 		old_object->surface_reg = -1;
578 		i = steal;
579 	}
580 
581 	bo->surface_reg = i;
582 	reg->bo = bo;
583 
584 out:
585 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
586 			       bo->tbo.resource->start << PAGE_SHIFT,
587 			       bo->tbo.base.size);
588 	return 0;
589 }
590 
591 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
592 {
593 	struct radeon_device *rdev = bo->rdev;
594 	struct radeon_surface_reg *reg;
595 
596 	if (bo->surface_reg == -1)
597 		return;
598 
599 	reg = &rdev->surface_regs[bo->surface_reg];
600 	radeon_clear_surface_reg(rdev, bo->surface_reg);
601 
602 	reg->bo = NULL;
603 	bo->surface_reg = -1;
604 }
605 
606 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
607 				uint32_t tiling_flags, uint32_t pitch)
608 {
609 	struct radeon_device *rdev = bo->rdev;
610 	int r;
611 
612 	if (rdev->family >= CHIP_CEDAR) {
613 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
614 
615 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
616 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
617 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
618 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
619 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
620 		switch (bankw) {
621 		case 0:
622 		case 1:
623 		case 2:
624 		case 4:
625 		case 8:
626 			break;
627 		default:
628 			return -EINVAL;
629 		}
630 		switch (bankh) {
631 		case 0:
632 		case 1:
633 		case 2:
634 		case 4:
635 		case 8:
636 			break;
637 		default:
638 			return -EINVAL;
639 		}
640 		switch (mtaspect) {
641 		case 0:
642 		case 1:
643 		case 2:
644 		case 4:
645 		case 8:
646 			break;
647 		default:
648 			return -EINVAL;
649 		}
650 		if (tilesplit > 6) {
651 			return -EINVAL;
652 		}
653 		if (stilesplit > 6) {
654 			return -EINVAL;
655 		}
656 	}
657 	r = radeon_bo_reserve(bo, false);
658 	if (unlikely(r != 0))
659 		return r;
660 	bo->tiling_flags = tiling_flags;
661 	bo->pitch = pitch;
662 	radeon_bo_unreserve(bo);
663 	return 0;
664 }
665 
666 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
667 				uint32_t *tiling_flags,
668 				uint32_t *pitch)
669 {
670 	dma_resv_assert_held(bo->tbo.base.resv);
671 
672 	if (tiling_flags)
673 		*tiling_flags = bo->tiling_flags;
674 	if (pitch)
675 		*pitch = bo->pitch;
676 }
677 
678 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
679 				bool force_drop)
680 {
681 	if (!force_drop)
682 		dma_resv_assert_held(bo->tbo.base.resv);
683 
684 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
685 		return 0;
686 
687 	if (force_drop) {
688 		radeon_bo_clear_surface_reg(bo);
689 		return 0;
690 	}
691 
692 	if (bo->tbo.resource->mem_type != TTM_PL_VRAM) {
693 		if (!has_moved)
694 			return 0;
695 
696 		if (bo->surface_reg >= 0)
697 			radeon_bo_clear_surface_reg(bo);
698 		return 0;
699 	}
700 
701 	if ((bo->surface_reg >= 0) && !has_moved)
702 		return 0;
703 
704 	return radeon_bo_get_surface_reg(bo);
705 }
706 
707 void radeon_bo_move_notify(struct ttm_buffer_object *bo)
708 {
709 	struct radeon_bo *rbo;
710 
711 	if (!radeon_ttm_bo_is_radeon_bo(bo))
712 		return;
713 
714 	rbo = container_of(bo, struct radeon_bo, tbo);
715 	radeon_bo_check_tiling(rbo, 0, 1);
716 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
717 }
718 
719 vm_fault_t radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
720 {
721 	struct ttm_operation_ctx ctx = { false, false };
722 	struct radeon_device *rdev;
723 	struct radeon_bo *rbo;
724 	unsigned long offset, size, lpfn;
725 	int i, r;
726 
727 	if (!radeon_ttm_bo_is_radeon_bo(bo))
728 		return 0;
729 	rbo = container_of(bo, struct radeon_bo, tbo);
730 	radeon_bo_check_tiling(rbo, 0, 0);
731 	rdev = rbo->rdev;
732 	if (bo->resource->mem_type != TTM_PL_VRAM)
733 		return 0;
734 
735 	size = bo->resource->num_pages << PAGE_SHIFT;
736 	offset = bo->resource->start << PAGE_SHIFT;
737 	if ((offset + size) <= rdev->mc.visible_vram_size)
738 		return 0;
739 
740 	/* Can't move a pinned BO to visible VRAM */
741 	if (rbo->tbo.pin_count > 0)
742 		return VM_FAULT_SIGBUS;
743 
744 	/* hurrah the memory is not visible ! */
745 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
746 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
747 	for (i = 0; i < rbo->placement.num_placement; i++) {
748 		/* Force into visible VRAM */
749 		if ((rbo->placements[i].mem_type == TTM_PL_VRAM) &&
750 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
751 			rbo->placements[i].lpfn = lpfn;
752 	}
753 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
754 	if (unlikely(r == -ENOMEM)) {
755 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
756 		r = ttm_bo_validate(bo, &rbo->placement, &ctx);
757 	} else if (likely(!r)) {
758 		offset = bo->resource->start << PAGE_SHIFT;
759 		/* this should never happen */
760 		if ((offset + size) > rdev->mc.visible_vram_size)
761 			return VM_FAULT_SIGBUS;
762 	}
763 
764 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
765 		return VM_FAULT_NOPAGE;
766 	else if (unlikely(r))
767 		return VM_FAULT_SIGBUS;
768 
769 	ttm_bo_move_to_lru_tail_unlocked(bo);
770 	return 0;
771 }
772 
773 /**
774  * radeon_bo_fence - add fence to buffer object
775  *
776  * @bo: buffer object in question
777  * @fence: fence to add
778  * @shared: true if fence should be added shared
779  *
780  */
781 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
782 		     bool shared)
783 {
784 	struct dma_resv *resv = bo->tbo.base.resv;
785 
786 	if (shared)
787 		dma_resv_add_shared_fence(resv, &fence->base);
788 	else
789 		dma_resv_add_excl_fence(resv, &fence->base);
790 }
791