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 ttm_buffer_object *bo,
53 				       unsigned int mem_type, int sign)
54 {
55 	struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
56 
57 	switch (mem_type) {
58 	case TTM_PL_TT:
59 		if (sign > 0)
60 			atomic64_add(bo->base.size, &rdev->gtt_usage);
61 		else
62 			atomic64_sub(bo->base.size, &rdev->gtt_usage);
63 		break;
64 	case TTM_PL_VRAM:
65 		if (sign > 0)
66 			atomic64_add(bo->base.size, &rdev->vram_usage);
67 		else
68 			atomic64_sub(bo->base.size, &rdev->vram_usage);
69 		break;
70 	}
71 }
72 
73 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
74 {
75 	struct radeon_bo *bo;
76 
77 	bo = container_of(tbo, struct radeon_bo, tbo);
78 
79 	mutex_lock(&bo->rdev->gem.mutex);
80 	list_del_init(&bo->list);
81 	mutex_unlock(&bo->rdev->gem.mutex);
82 	radeon_bo_clear_surface_reg(bo);
83 	WARN_ON_ONCE(!list_empty(&bo->va));
84 	if (bo->tbo.base.import_attach)
85 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
86 	drm_gem_object_release(&bo->tbo.base);
87 	kfree(bo);
88 }
89 
90 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
91 {
92 	if (bo->destroy == &radeon_ttm_bo_destroy)
93 		return true;
94 	return false;
95 }
96 
97 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
98 {
99 	u32 c = 0, i;
100 
101 	rbo->placement.placement = rbo->placements;
102 	rbo->placement.busy_placement = rbo->placements;
103 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
104 		/* Try placing BOs which don't need CPU access outside of the
105 		 * CPU accessible part of VRAM
106 		 */
107 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
108 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
109 			rbo->placements[c].fpfn =
110 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
111 			rbo->placements[c].mem_type = TTM_PL_VRAM;
112 			rbo->placements[c++].flags = 0;
113 		}
114 
115 		rbo->placements[c].fpfn = 0;
116 		rbo->placements[c].mem_type = TTM_PL_VRAM;
117 		rbo->placements[c++].flags = 0;
118 	}
119 
120 	if (domain & RADEON_GEM_DOMAIN_GTT) {
121 		rbo->placements[c].fpfn = 0;
122 		rbo->placements[c].mem_type = TTM_PL_TT;
123 		rbo->placements[c++].flags = 0;
124 	}
125 
126 	if (domain & RADEON_GEM_DOMAIN_CPU) {
127 		rbo->placements[c].fpfn = 0;
128 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
129 		rbo->placements[c++].flags = 0;
130 	}
131 	if (!c) {
132 		rbo->placements[c].fpfn = 0;
133 		rbo->placements[c].mem_type = TTM_PL_SYSTEM;
134 		rbo->placements[c++].flags = 0;
135 	}
136 
137 	rbo->placement.num_placement = c;
138 	rbo->placement.num_busy_placement = c;
139 
140 	for (i = 0; i < c; ++i) {
141 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
142 		    (rbo->placements[i].mem_type == TTM_PL_VRAM) &&
143 		    !rbo->placements[i].fpfn)
144 			rbo->placements[i].lpfn =
145 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
146 		else
147 			rbo->placements[i].lpfn = 0;
148 	}
149 }
150 
151 int radeon_bo_create(struct radeon_device *rdev,
152 		     unsigned long size, int byte_align, bool kernel,
153 		     u32 domain, u32 flags, struct sg_table *sg,
154 		     struct dma_resv *resv,
155 		     struct radeon_bo **bo_ptr)
156 {
157 	struct radeon_bo *bo;
158 	enum ttm_bo_type type;
159 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
160 	int r;
161 
162 	size = ALIGN(size, PAGE_SIZE);
163 
164 	if (kernel) {
165 		type = ttm_bo_type_kernel;
166 	} else if (sg) {
167 		type = ttm_bo_type_sg;
168 	} else {
169 		type = ttm_bo_type_device;
170 	}
171 	*bo_ptr = NULL;
172 
173 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
174 	if (bo == NULL)
175 		return -ENOMEM;
176 	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
177 	bo->rdev = rdev;
178 	bo->surface_reg = -1;
179 	INIT_LIST_HEAD(&bo->list);
180 	INIT_LIST_HEAD(&bo->va);
181 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
182 				       RADEON_GEM_DOMAIN_GTT |
183 				       RADEON_GEM_DOMAIN_CPU);
184 
185 	bo->flags = flags;
186 	/* PCI GART is always snooped */
187 	if (!(rdev->flags & RADEON_IS_PCIE))
188 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
189 
190 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
191 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
192 	 */
193 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
194 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
195 
196 #ifdef CONFIG_X86_32
197 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
198 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
199 	 */
200 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
201 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
202 	/* Don't try to enable write-combining when it can't work, or things
203 	 * may be slow
204 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
205 	 */
206 #ifndef CONFIG_COMPILE_TEST
207 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
208 	 thanks to write-combining
209 #endif
210 
211 	if (bo->flags & RADEON_GEM_GTT_WC)
212 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
213 			      "better performance thanks to write-combining\n");
214 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
215 #else
216 	/* For architectures that don't support WC memory,
217 	 * mask out the WC flag from the BO
218 	 */
219 	if (!drm_arch_can_wc_memory())
220 		bo->flags &= ~RADEON_GEM_GTT_WC;
221 #endif
222 
223 	radeon_ttm_placement_from_domain(bo, domain);
224 	/* Kernel allocation are uninterruptible */
225 	down_read(&rdev->pm.mclk_lock);
226 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
227 			&bo->placement, page_align, !kernel, sg, resv,
228 			&radeon_ttm_bo_destroy);
229 	up_read(&rdev->pm.mclk_lock);
230 	if (unlikely(r != 0)) {
231 		return r;
232 	}
233 	*bo_ptr = bo;
234 
235 	trace_radeon_bo_create(bo);
236 
237 	return 0;
238 }
239 
240 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
241 {
242 	bool is_iomem;
243 	int r;
244 
245 	if (bo->kptr) {
246 		if (ptr) {
247 			*ptr = bo->kptr;
248 		}
249 		return 0;
250 	}
251 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.resource->num_pages, &bo->kmap);
252 	if (r) {
253 		return r;
254 	}
255 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
256 	if (ptr) {
257 		*ptr = bo->kptr;
258 	}
259 	radeon_bo_check_tiling(bo, 0, 0);
260 	return 0;
261 }
262 
263 void radeon_bo_kunmap(struct radeon_bo *bo)
264 {
265 	if (bo->kptr == NULL)
266 		return;
267 	bo->kptr = NULL;
268 	radeon_bo_check_tiling(bo, 0, 0);
269 	ttm_bo_kunmap(&bo->kmap);
270 }
271 
272 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
273 {
274 	if (bo == NULL)
275 		return NULL;
276 
277 	ttm_bo_get(&bo->tbo);
278 	return bo;
279 }
280 
281 void radeon_bo_unref(struct radeon_bo **bo)
282 {
283 	struct ttm_buffer_object *tbo;
284 
285 	if ((*bo) == NULL)
286 		return;
287 	tbo = &((*bo)->tbo);
288 	ttm_bo_put(tbo);
289 	*bo = NULL;
290 }
291 
292 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
293 			     u64 *gpu_addr)
294 {
295 	struct ttm_operation_ctx ctx = { false, false };
296 	int r, i;
297 
298 	if (radeon_ttm_tt_has_userptr(bo->rdev, bo->tbo.ttm))
299 		return -EPERM;
300 
301 	if (bo->tbo.pin_count) {
302 		ttm_bo_pin(&bo->tbo);
303 		if (gpu_addr)
304 			*gpu_addr = radeon_bo_gpu_offset(bo);
305 
306 		if (max_offset != 0) {
307 			u64 domain_start;
308 
309 			if (domain == RADEON_GEM_DOMAIN_VRAM)
310 				domain_start = bo->rdev->mc.vram_start;
311 			else
312 				domain_start = bo->rdev->mc.gtt_start;
313 			WARN_ON_ONCE(max_offset <
314 				     (radeon_bo_gpu_offset(bo) - domain_start));
315 		}
316 
317 		return 0;
318 	}
319 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
320 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
321 		return -EINVAL;
322 	}
323 
324 	radeon_ttm_placement_from_domain(bo, domain);
325 	for (i = 0; i < bo->placement.num_placement; i++) {
326 		/* force to pin into visible video ram */
327 		if ((bo->placements[i].mem_type == TTM_PL_VRAM) &&
328 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
329 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
330 			bo->placements[i].lpfn =
331 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
332 		else
333 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
334 	}
335 
336 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
337 	if (likely(r == 0)) {
338 		ttm_bo_pin(&bo->tbo);
339 		if (gpu_addr != NULL)
340 			*gpu_addr = radeon_bo_gpu_offset(bo);
341 		if (domain == RADEON_GEM_DOMAIN_VRAM)
342 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
343 		else
344 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
345 	} else {
346 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
347 	}
348 	return r;
349 }
350 
351 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
352 {
353 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
354 }
355 
356 void radeon_bo_unpin(struct radeon_bo *bo)
357 {
358 	ttm_bo_unpin(&bo->tbo);
359 	if (!bo->tbo.pin_count) {
360 		if (bo->tbo.resource->mem_type == TTM_PL_VRAM)
361 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
362 		else
363 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
364 	}
365 }
366 
367 int radeon_bo_evict_vram(struct radeon_device *rdev)
368 {
369 	struct ttm_device *bdev = &rdev->mman.bdev;
370 	struct ttm_resource_manager *man;
371 
372 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
373 #ifndef CONFIG_HIBERNATION
374 	if (rdev->flags & RADEON_IS_IGP) {
375 		if (rdev->mc.igp_sideport_enabled == false)
376 			/* Useless to evict on IGP chips */
377 			return 0;
378 	}
379 #endif
380 	man = ttm_manager_type(bdev, TTM_PL_VRAM);
381 	if (!man)
382 		return 0;
383 	return ttm_resource_manager_evict_all(bdev, man);
384 }
385 
386 void radeon_bo_force_delete(struct radeon_device *rdev)
387 {
388 	struct radeon_bo *bo, *n;
389 
390 	if (list_empty(&rdev->gem.objects)) {
391 		return;
392 	}
393 	dev_err(rdev->dev, "Userspace still has active objects !\n");
394 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
395 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
396 			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
397 			*((unsigned long *)&bo->tbo.base.refcount));
398 		mutex_lock(&bo->rdev->gem.mutex);
399 		list_del_init(&bo->list);
400 		mutex_unlock(&bo->rdev->gem.mutex);
401 		/* this should unref the ttm bo */
402 		drm_gem_object_put(&bo->tbo.base);
403 	}
404 }
405 
406 int radeon_bo_init(struct radeon_device *rdev)
407 {
408 	/* reserve PAT memory space to WC for VRAM */
409 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
410 				   rdev->mc.aper_size);
411 
412 	/* Add an MTRR for the VRAM */
413 	if (!rdev->fastfb_working) {
414 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
415 						      rdev->mc.aper_size);
416 	}
417 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
418 		rdev->mc.mc_vram_size >> 20,
419 		(unsigned long long)rdev->mc.aper_size >> 20);
420 	DRM_INFO("RAM width %dbits %cDR\n",
421 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
422 	return radeon_ttm_init(rdev);
423 }
424 
425 void radeon_bo_fini(struct radeon_device *rdev)
426 {
427 	radeon_ttm_fini(rdev);
428 	arch_phys_wc_del(rdev->mc.vram_mtrr);
429 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
430 }
431 
432 /* Returns how many bytes TTM can move per IB.
433  */
434 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
435 {
436 	u64 real_vram_size = rdev->mc.real_vram_size;
437 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
438 
439 	/* This function is based on the current VRAM usage.
440 	 *
441 	 * - If all of VRAM is free, allow relocating the number of bytes that
442 	 *   is equal to 1/4 of the size of VRAM for this IB.
443 
444 	 * - If more than one half of VRAM is occupied, only allow relocating
445 	 *   1 MB of data for this IB.
446 	 *
447 	 * - From 0 to one half of used VRAM, the threshold decreases
448 	 *   linearly.
449 	 *         __________________
450 	 * 1/4 of -|\               |
451 	 * VRAM    | \              |
452 	 *         |  \             |
453 	 *         |   \            |
454 	 *         |    \           |
455 	 *         |     \          |
456 	 *         |      \         |
457 	 *         |       \________|1 MB
458 	 *         |----------------|
459 	 *    VRAM 0 %             100 %
460 	 *         used            used
461 	 *
462 	 * Note: It's a threshold, not a limit. The threshold must be crossed
463 	 * for buffer relocations to stop, so any buffer of an arbitrary size
464 	 * can be moved as long as the threshold isn't crossed before
465 	 * the relocation takes place. We don't want to disable buffer
466 	 * relocations completely.
467 	 *
468 	 * The idea is that buffers should be placed in VRAM at creation time
469 	 * and TTM should only do a minimum number of relocations during
470 	 * command submission. In practice, you need to submit at least
471 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
472 	 *
473 	 * Also, things can get pretty crazy under memory pressure and actual
474 	 * VRAM usage can change a lot, so playing safe even at 50% does
475 	 * consistently increase performance.
476 	 */
477 
478 	u64 half_vram = real_vram_size >> 1;
479 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
480 	u64 bytes_moved_threshold = half_free_vram >> 1;
481 	return max(bytes_moved_threshold, 1024*1024ull);
482 }
483 
484 int radeon_bo_list_validate(struct radeon_device *rdev,
485 			    struct ww_acquire_ctx *ticket,
486 			    struct list_head *head, int ring)
487 {
488 	struct ttm_operation_ctx ctx = { true, false };
489 	struct radeon_bo_list *lobj;
490 	struct list_head duplicates;
491 	int r;
492 	u64 bytes_moved = 0, initial_bytes_moved;
493 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
494 
495 	INIT_LIST_HEAD(&duplicates);
496 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
497 	if (unlikely(r != 0)) {
498 		return r;
499 	}
500 
501 	list_for_each_entry(lobj, head, tv.head) {
502 		struct radeon_bo *bo = lobj->robj;
503 		if (!bo->tbo.pin_count) {
504 			u32 domain = lobj->preferred_domains;
505 			u32 allowed = lobj->allowed_domains;
506 			u32 current_domain =
507 				radeon_mem_type_to_domain(bo->tbo.resource->mem_type);
508 
509 			/* Check if this buffer will be moved and don't move it
510 			 * if we have moved too many buffers for this IB already.
511 			 *
512 			 * Note that this allows moving at least one buffer of
513 			 * any size, because it doesn't take the current "bo"
514 			 * into account. We don't want to disallow buffer moves
515 			 * completely.
516 			 */
517 			if ((allowed & current_domain) != 0 &&
518 			    (domain & current_domain) == 0 && /* will be moved */
519 			    bytes_moved > bytes_moved_threshold) {
520 				/* don't move it */
521 				domain = current_domain;
522 			}
523 
524 		retry:
525 			radeon_ttm_placement_from_domain(bo, domain);
526 			if (ring == R600_RING_TYPE_UVD_INDEX)
527 				radeon_uvd_force_into_uvd_segment(bo, allowed);
528 
529 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
530 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
531 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
532 				       initial_bytes_moved;
533 
534 			if (unlikely(r)) {
535 				if (r != -ERESTARTSYS &&
536 				    domain != lobj->allowed_domains) {
537 					domain = lobj->allowed_domains;
538 					goto retry;
539 				}
540 				ttm_eu_backoff_reservation(ticket, head);
541 				return r;
542 			}
543 		}
544 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
545 		lobj->tiling_flags = bo->tiling_flags;
546 	}
547 
548 	list_for_each_entry(lobj, &duplicates, tv.head) {
549 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
550 		lobj->tiling_flags = lobj->robj->tiling_flags;
551 	}
552 
553 	return 0;
554 }
555 
556 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
557 {
558 	struct radeon_device *rdev = bo->rdev;
559 	struct radeon_surface_reg *reg;
560 	struct radeon_bo *old_object;
561 	int steal;
562 	int i;
563 
564 	dma_resv_assert_held(bo->tbo.base.resv);
565 
566 	if (!bo->tiling_flags)
567 		return 0;
568 
569 	if (bo->surface_reg >= 0) {
570 		i = bo->surface_reg;
571 		goto out;
572 	}
573 
574 	steal = -1;
575 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
576 
577 		reg = &rdev->surface_regs[i];
578 		if (!reg->bo)
579 			break;
580 
581 		old_object = reg->bo;
582 		if (old_object->tbo.pin_count == 0)
583 			steal = i;
584 	}
585 
586 	/* if we are all out */
587 	if (i == RADEON_GEM_MAX_SURFACES) {
588 		if (steal == -1)
589 			return -ENOMEM;
590 		/* find someone with a surface reg and nuke their BO */
591 		reg = &rdev->surface_regs[steal];
592 		old_object = reg->bo;
593 		/* blow away the mapping */
594 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
595 		ttm_bo_unmap_virtual(&old_object->tbo);
596 		old_object->surface_reg = -1;
597 		i = steal;
598 	}
599 
600 	bo->surface_reg = i;
601 	reg->bo = bo;
602 
603 out:
604 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
605 			       bo->tbo.resource->start << PAGE_SHIFT,
606 			       bo->tbo.base.size);
607 	return 0;
608 }
609 
610 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
611 {
612 	struct radeon_device *rdev = bo->rdev;
613 	struct radeon_surface_reg *reg;
614 
615 	if (bo->surface_reg == -1)
616 		return;
617 
618 	reg = &rdev->surface_regs[bo->surface_reg];
619 	radeon_clear_surface_reg(rdev, bo->surface_reg);
620 
621 	reg->bo = NULL;
622 	bo->surface_reg = -1;
623 }
624 
625 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
626 				uint32_t tiling_flags, uint32_t pitch)
627 {
628 	struct radeon_device *rdev = bo->rdev;
629 	int r;
630 
631 	if (rdev->family >= CHIP_CEDAR) {
632 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
633 
634 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
635 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
636 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
637 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
638 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
639 		switch (bankw) {
640 		case 0:
641 		case 1:
642 		case 2:
643 		case 4:
644 		case 8:
645 			break;
646 		default:
647 			return -EINVAL;
648 		}
649 		switch (bankh) {
650 		case 0:
651 		case 1:
652 		case 2:
653 		case 4:
654 		case 8:
655 			break;
656 		default:
657 			return -EINVAL;
658 		}
659 		switch (mtaspect) {
660 		case 0:
661 		case 1:
662 		case 2:
663 		case 4:
664 		case 8:
665 			break;
666 		default:
667 			return -EINVAL;
668 		}
669 		if (tilesplit > 6) {
670 			return -EINVAL;
671 		}
672 		if (stilesplit > 6) {
673 			return -EINVAL;
674 		}
675 	}
676 	r = radeon_bo_reserve(bo, false);
677 	if (unlikely(r != 0))
678 		return r;
679 	bo->tiling_flags = tiling_flags;
680 	bo->pitch = pitch;
681 	radeon_bo_unreserve(bo);
682 	return 0;
683 }
684 
685 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
686 				uint32_t *tiling_flags,
687 				uint32_t *pitch)
688 {
689 	dma_resv_assert_held(bo->tbo.base.resv);
690 
691 	if (tiling_flags)
692 		*tiling_flags = bo->tiling_flags;
693 	if (pitch)
694 		*pitch = bo->pitch;
695 }
696 
697 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
698 				bool force_drop)
699 {
700 	if (!force_drop)
701 		dma_resv_assert_held(bo->tbo.base.resv);
702 
703 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
704 		return 0;
705 
706 	if (force_drop) {
707 		radeon_bo_clear_surface_reg(bo);
708 		return 0;
709 	}
710 
711 	if (bo->tbo.resource->mem_type != TTM_PL_VRAM) {
712 		if (!has_moved)
713 			return 0;
714 
715 		if (bo->surface_reg >= 0)
716 			radeon_bo_clear_surface_reg(bo);
717 		return 0;
718 	}
719 
720 	if ((bo->surface_reg >= 0) && !has_moved)
721 		return 0;
722 
723 	return radeon_bo_get_surface_reg(bo);
724 }
725 
726 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
727 			   unsigned int old_type,
728 			   struct ttm_resource *new_mem)
729 {
730 	struct radeon_bo *rbo;
731 
732 	radeon_update_memory_usage(bo, old_type, -1);
733 	if (new_mem)
734 		radeon_update_memory_usage(bo, new_mem->mem_type, 1);
735 
736 	if (!radeon_ttm_bo_is_radeon_bo(bo))
737 		return;
738 
739 	rbo = container_of(bo, struct radeon_bo, tbo);
740 	radeon_bo_check_tiling(rbo, 0, 1);
741 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
742 }
743 
744 vm_fault_t radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
745 {
746 	struct ttm_operation_ctx ctx = { false, false };
747 	struct radeon_device *rdev;
748 	struct radeon_bo *rbo;
749 	unsigned long offset, size, lpfn;
750 	int i, r;
751 
752 	if (!radeon_ttm_bo_is_radeon_bo(bo))
753 		return 0;
754 	rbo = container_of(bo, struct radeon_bo, tbo);
755 	radeon_bo_check_tiling(rbo, 0, 0);
756 	rdev = rbo->rdev;
757 	if (bo->resource->mem_type != TTM_PL_VRAM)
758 		return 0;
759 
760 	size = bo->resource->num_pages << PAGE_SHIFT;
761 	offset = bo->resource->start << PAGE_SHIFT;
762 	if ((offset + size) <= rdev->mc.visible_vram_size)
763 		return 0;
764 
765 	/* Can't move a pinned BO to visible VRAM */
766 	if (rbo->tbo.pin_count > 0)
767 		return VM_FAULT_SIGBUS;
768 
769 	/* hurrah the memory is not visible ! */
770 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
771 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
772 	for (i = 0; i < rbo->placement.num_placement; i++) {
773 		/* Force into visible VRAM */
774 		if ((rbo->placements[i].mem_type == TTM_PL_VRAM) &&
775 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
776 			rbo->placements[i].lpfn = lpfn;
777 	}
778 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
779 	if (unlikely(r == -ENOMEM)) {
780 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
781 		r = ttm_bo_validate(bo, &rbo->placement, &ctx);
782 	} else if (likely(!r)) {
783 		offset = bo->resource->start << PAGE_SHIFT;
784 		/* this should never happen */
785 		if ((offset + size) > rdev->mc.visible_vram_size)
786 			return VM_FAULT_SIGBUS;
787 	}
788 
789 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
790 		return VM_FAULT_NOPAGE;
791 	else if (unlikely(r))
792 		return VM_FAULT_SIGBUS;
793 
794 	ttm_bo_move_to_lru_tail_unlocked(bo);
795 	return 0;
796 }
797 
798 /**
799  * radeon_bo_fence - add fence to buffer object
800  *
801  * @bo: buffer object in question
802  * @fence: fence to add
803  * @shared: true if fence should be added shared
804  *
805  */
806 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
807 		     bool shared)
808 {
809 	struct dma_resv *resv = bo->tbo.base.resv;
810 
811 	if (shared)
812 		dma_resv_add_shared_fence(resv, &fence->base);
813 	else
814 		dma_resv_add_excl_fence(resv, &fence->base);
815 }
816