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 "radeon.h"
37 #include "radeon_trace.h"
38 
39 
40 int radeon_ttm_init(struct radeon_device *rdev);
41 void radeon_ttm_fini(struct radeon_device *rdev);
42 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
43 
44 /*
45  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
46  * function are calling it.
47  */
48 
49 static void radeon_update_memory_usage(struct radeon_bo *bo,
50 				       unsigned mem_type, int sign)
51 {
52 	struct radeon_device *rdev = bo->rdev;
53 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
54 
55 	switch (mem_type) {
56 	case TTM_PL_TT:
57 		if (sign > 0)
58 			atomic64_add(size, &rdev->gtt_usage);
59 		else
60 			atomic64_sub(size, &rdev->gtt_usage);
61 		break;
62 	case TTM_PL_VRAM:
63 		if (sign > 0)
64 			atomic64_add(size, &rdev->vram_usage);
65 		else
66 			atomic64_sub(size, &rdev->vram_usage);
67 		break;
68 	}
69 }
70 
71 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
72 {
73 	struct radeon_bo *bo;
74 
75 	bo = container_of(tbo, struct radeon_bo, tbo);
76 
77 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
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(!list_empty(&bo->va));
84 	drm_gem_object_release(&bo->gem_base);
85 	kfree(bo);
86 }
87 
88 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
89 {
90 	if (bo->destroy == &radeon_ttm_bo_destroy)
91 		return true;
92 	return false;
93 }
94 
95 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
96 {
97 	u32 c = 0, i;
98 
99 	rbo->placement.fpfn = 0;
100 	rbo->placement.lpfn = 0;
101 	rbo->placement.placement = rbo->placements;
102 	rbo->placement.busy_placement = rbo->placements;
103 	if (domain & RADEON_GEM_DOMAIN_VRAM)
104 		rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
105 					TTM_PL_FLAG_VRAM;
106 	if (domain & RADEON_GEM_DOMAIN_GTT) {
107 		if (rbo->flags & RADEON_GEM_GTT_UC) {
108 			rbo->placements[c++] = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_TT;
109 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
110 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
111 			rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
112 				TTM_PL_FLAG_TT;
113 		} else {
114 			rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_TT;
115 		}
116 	}
117 	if (domain & RADEON_GEM_DOMAIN_CPU) {
118 		if (rbo->flags & RADEON_GEM_GTT_UC) {
119 			rbo->placements[c++] = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_SYSTEM;
120 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
121 		    rbo->rdev->flags & RADEON_IS_AGP) {
122 			rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
123 				TTM_PL_FLAG_SYSTEM;
124 		} else {
125 			rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM;
126 		}
127 	}
128 	if (!c)
129 		rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
130 	rbo->placement.num_placement = c;
131 	rbo->placement.num_busy_placement = c;
132 
133 	/*
134 	 * Use two-ended allocation depending on the buffer size to
135 	 * improve fragmentation quality.
136 	 * 512kb was measured as the most optimal number.
137 	 */
138 	if (rbo->tbo.mem.size > 512 * 1024) {
139 		for (i = 0; i < c; i++) {
140 			rbo->placements[i] |= TTM_PL_FLAG_TOPDOWN;
141 		}
142 	}
143 }
144 
145 int radeon_bo_create(struct radeon_device *rdev,
146 		     unsigned long size, int byte_align, bool kernel, u32 domain,
147 		     u32 flags, struct sg_table *sg, struct radeon_bo **bo_ptr)
148 {
149 	struct radeon_bo *bo;
150 	enum ttm_bo_type type;
151 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
152 	size_t acc_size;
153 	int r;
154 
155 	size = ALIGN(size, PAGE_SIZE);
156 
157 	if (kernel) {
158 		type = ttm_bo_type_kernel;
159 	} else if (sg) {
160 		type = ttm_bo_type_sg;
161 	} else {
162 		type = ttm_bo_type_device;
163 	}
164 	*bo_ptr = NULL;
165 
166 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
167 				       sizeof(struct radeon_bo));
168 
169 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
170 	if (bo == NULL)
171 		return -ENOMEM;
172 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
173 	if (unlikely(r)) {
174 		kfree(bo);
175 		return r;
176 	}
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 	radeon_ttm_placement_from_domain(bo, domain);
191 	/* Kernel allocation are uninterruptible */
192 	down_read(&rdev->pm.mclk_lock);
193 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
194 			&bo->placement, page_align, !kernel, NULL,
195 			acc_size, sg, &radeon_ttm_bo_destroy);
196 	up_read(&rdev->pm.mclk_lock);
197 	if (unlikely(r != 0)) {
198 		return r;
199 	}
200 	*bo_ptr = bo;
201 
202 	trace_radeon_bo_create(bo);
203 
204 	return 0;
205 }
206 
207 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
208 {
209 	bool is_iomem;
210 	int r;
211 
212 	if (bo->kptr) {
213 		if (ptr) {
214 			*ptr = bo->kptr;
215 		}
216 		return 0;
217 	}
218 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
219 	if (r) {
220 		return r;
221 	}
222 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
223 	if (ptr) {
224 		*ptr = bo->kptr;
225 	}
226 	radeon_bo_check_tiling(bo, 0, 0);
227 	return 0;
228 }
229 
230 void radeon_bo_kunmap(struct radeon_bo *bo)
231 {
232 	if (bo->kptr == NULL)
233 		return;
234 	bo->kptr = NULL;
235 	radeon_bo_check_tiling(bo, 0, 0);
236 	ttm_bo_kunmap(&bo->kmap);
237 }
238 
239 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
240 {
241 	if (bo == NULL)
242 		return NULL;
243 
244 	ttm_bo_reference(&bo->tbo);
245 	return bo;
246 }
247 
248 void radeon_bo_unref(struct radeon_bo **bo)
249 {
250 	struct ttm_buffer_object *tbo;
251 	struct radeon_device *rdev;
252 
253 	if ((*bo) == NULL)
254 		return;
255 	rdev = (*bo)->rdev;
256 	tbo = &((*bo)->tbo);
257 	ttm_bo_unref(&tbo);
258 	if (tbo == NULL)
259 		*bo = NULL;
260 }
261 
262 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
263 			     u64 *gpu_addr)
264 {
265 	int r, i;
266 
267 	if (bo->pin_count) {
268 		bo->pin_count++;
269 		if (gpu_addr)
270 			*gpu_addr = radeon_bo_gpu_offset(bo);
271 
272 		if (max_offset != 0) {
273 			u64 domain_start;
274 
275 			if (domain == RADEON_GEM_DOMAIN_VRAM)
276 				domain_start = bo->rdev->mc.vram_start;
277 			else
278 				domain_start = bo->rdev->mc.gtt_start;
279 			WARN_ON_ONCE(max_offset <
280 				     (radeon_bo_gpu_offset(bo) - domain_start));
281 		}
282 
283 		return 0;
284 	}
285 	radeon_ttm_placement_from_domain(bo, domain);
286 	if (domain == RADEON_GEM_DOMAIN_VRAM) {
287 		/* force to pin into visible video ram */
288 		bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
289 	}
290 	if (max_offset) {
291 		u64 lpfn = max_offset >> PAGE_SHIFT;
292 
293 		if (!bo->placement.lpfn)
294 			bo->placement.lpfn = bo->rdev->mc.gtt_size >> PAGE_SHIFT;
295 
296 		if (lpfn < bo->placement.lpfn)
297 			bo->placement.lpfn = lpfn;
298 	}
299 	for (i = 0; i < bo->placement.num_placement; i++)
300 		bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
301 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
302 	if (likely(r == 0)) {
303 		bo->pin_count = 1;
304 		if (gpu_addr != NULL)
305 			*gpu_addr = radeon_bo_gpu_offset(bo);
306 		if (domain == RADEON_GEM_DOMAIN_VRAM)
307 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
308 		else
309 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
310 	} else {
311 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
312 	}
313 	return r;
314 }
315 
316 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
317 {
318 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
319 }
320 
321 int radeon_bo_unpin(struct radeon_bo *bo)
322 {
323 	int r, i;
324 
325 	if (!bo->pin_count) {
326 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
327 		return 0;
328 	}
329 	bo->pin_count--;
330 	if (bo->pin_count)
331 		return 0;
332 	for (i = 0; i < bo->placement.num_placement; i++)
333 		bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
334 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
335 	if (likely(r == 0)) {
336 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
337 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
338 		else
339 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
340 	} else {
341 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
342 	}
343 	return r;
344 }
345 
346 int radeon_bo_evict_vram(struct radeon_device *rdev)
347 {
348 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
349 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
350 		if (rdev->mc.igp_sideport_enabled == false)
351 			/* Useless to evict on IGP chips */
352 			return 0;
353 	}
354 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
355 }
356 
357 void radeon_bo_force_delete(struct radeon_device *rdev)
358 {
359 	struct radeon_bo *bo, *n;
360 
361 	if (list_empty(&rdev->gem.objects)) {
362 		return;
363 	}
364 	dev_err(rdev->dev, "Userspace still has active objects !\n");
365 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
366 		mutex_lock(&rdev->ddev->struct_mutex);
367 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
368 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
369 			*((unsigned long *)&bo->gem_base.refcount));
370 		mutex_lock(&bo->rdev->gem.mutex);
371 		list_del_init(&bo->list);
372 		mutex_unlock(&bo->rdev->gem.mutex);
373 		/* this should unref the ttm bo */
374 		drm_gem_object_unreference(&bo->gem_base);
375 		mutex_unlock(&rdev->ddev->struct_mutex);
376 	}
377 }
378 
379 int radeon_bo_init(struct radeon_device *rdev)
380 {
381 	/* Add an MTRR for the VRAM */
382 	if (!rdev->fastfb_working) {
383 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
384 						      rdev->mc.aper_size);
385 	}
386 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
387 		rdev->mc.mc_vram_size >> 20,
388 		(unsigned long long)rdev->mc.aper_size >> 20);
389 	DRM_INFO("RAM width %dbits %cDR\n",
390 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
391 	return radeon_ttm_init(rdev);
392 }
393 
394 void radeon_bo_fini(struct radeon_device *rdev)
395 {
396 	radeon_ttm_fini(rdev);
397 	arch_phys_wc_del(rdev->mc.vram_mtrr);
398 }
399 
400 /* Returns how many bytes TTM can move per IB.
401  */
402 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
403 {
404 	u64 real_vram_size = rdev->mc.real_vram_size;
405 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
406 
407 	/* This function is based on the current VRAM usage.
408 	 *
409 	 * - If all of VRAM is free, allow relocating the number of bytes that
410 	 *   is equal to 1/4 of the size of VRAM for this IB.
411 
412 	 * - If more than one half of VRAM is occupied, only allow relocating
413 	 *   1 MB of data for this IB.
414 	 *
415 	 * - From 0 to one half of used VRAM, the threshold decreases
416 	 *   linearly.
417 	 *         __________________
418 	 * 1/4 of -|\               |
419 	 * VRAM    | \              |
420 	 *         |  \             |
421 	 *         |   \            |
422 	 *         |    \           |
423 	 *         |     \          |
424 	 *         |      \         |
425 	 *         |       \________|1 MB
426 	 *         |----------------|
427 	 *    VRAM 0 %             100 %
428 	 *         used            used
429 	 *
430 	 * Note: It's a threshold, not a limit. The threshold must be crossed
431 	 * for buffer relocations to stop, so any buffer of an arbitrary size
432 	 * can be moved as long as the threshold isn't crossed before
433 	 * the relocation takes place. We don't want to disable buffer
434 	 * relocations completely.
435 	 *
436 	 * The idea is that buffers should be placed in VRAM at creation time
437 	 * and TTM should only do a minimum number of relocations during
438 	 * command submission. In practice, you need to submit at least
439 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
440 	 *
441 	 * Also, things can get pretty crazy under memory pressure and actual
442 	 * VRAM usage can change a lot, so playing safe even at 50% does
443 	 * consistently increase performance.
444 	 */
445 
446 	u64 half_vram = real_vram_size >> 1;
447 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
448 	u64 bytes_moved_threshold = half_free_vram >> 1;
449 	return max(bytes_moved_threshold, 1024*1024ull);
450 }
451 
452 int radeon_bo_list_validate(struct radeon_device *rdev,
453 			    struct ww_acquire_ctx *ticket,
454 			    struct list_head *head, int ring)
455 {
456 	struct radeon_cs_reloc *lobj;
457 	struct radeon_bo *bo;
458 	int r;
459 	u64 bytes_moved = 0, initial_bytes_moved;
460 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
461 
462 	r = ttm_eu_reserve_buffers(ticket, head);
463 	if (unlikely(r != 0)) {
464 		return r;
465 	}
466 
467 	list_for_each_entry(lobj, head, tv.head) {
468 		bo = lobj->robj;
469 		if (!bo->pin_count) {
470 			u32 domain = lobj->prefered_domains;
471 			u32 current_domain =
472 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
473 
474 			/* Check if this buffer will be moved and don't move it
475 			 * if we have moved too many buffers for this IB already.
476 			 *
477 			 * Note that this allows moving at least one buffer of
478 			 * any size, because it doesn't take the current "bo"
479 			 * into account. We don't want to disallow buffer moves
480 			 * completely.
481 			 */
482 			if ((lobj->allowed_domains & current_domain) != 0 &&
483 			    (domain & current_domain) == 0 && /* will be moved */
484 			    bytes_moved > bytes_moved_threshold) {
485 				/* don't move it */
486 				domain = current_domain;
487 			}
488 
489 		retry:
490 			radeon_ttm_placement_from_domain(bo, domain);
491 			if (ring == R600_RING_TYPE_UVD_INDEX)
492 				radeon_uvd_force_into_uvd_segment(bo);
493 
494 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
495 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
496 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
497 				       initial_bytes_moved;
498 
499 			if (unlikely(r)) {
500 				if (r != -ERESTARTSYS &&
501 				    domain != lobj->allowed_domains) {
502 					domain = lobj->allowed_domains;
503 					goto retry;
504 				}
505 				ttm_eu_backoff_reservation(ticket, head);
506 				return r;
507 			}
508 		}
509 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
510 		lobj->tiling_flags = bo->tiling_flags;
511 	}
512 	return 0;
513 }
514 
515 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
516 			     struct vm_area_struct *vma)
517 {
518 	return ttm_fbdev_mmap(vma, &bo->tbo);
519 }
520 
521 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
522 {
523 	struct radeon_device *rdev = bo->rdev;
524 	struct radeon_surface_reg *reg;
525 	struct radeon_bo *old_object;
526 	int steal;
527 	int i;
528 
529 	lockdep_assert_held(&bo->tbo.resv->lock.base);
530 
531 	if (!bo->tiling_flags)
532 		return 0;
533 
534 	if (bo->surface_reg >= 0) {
535 		reg = &rdev->surface_regs[bo->surface_reg];
536 		i = bo->surface_reg;
537 		goto out;
538 	}
539 
540 	steal = -1;
541 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
542 
543 		reg = &rdev->surface_regs[i];
544 		if (!reg->bo)
545 			break;
546 
547 		old_object = reg->bo;
548 		if (old_object->pin_count == 0)
549 			steal = i;
550 	}
551 
552 	/* if we are all out */
553 	if (i == RADEON_GEM_MAX_SURFACES) {
554 		if (steal == -1)
555 			return -ENOMEM;
556 		/* find someone with a surface reg and nuke their BO */
557 		reg = &rdev->surface_regs[steal];
558 		old_object = reg->bo;
559 		/* blow away the mapping */
560 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
561 		ttm_bo_unmap_virtual(&old_object->tbo);
562 		old_object->surface_reg = -1;
563 		i = steal;
564 	}
565 
566 	bo->surface_reg = i;
567 	reg->bo = bo;
568 
569 out:
570 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
571 			       bo->tbo.mem.start << PAGE_SHIFT,
572 			       bo->tbo.num_pages << PAGE_SHIFT);
573 	return 0;
574 }
575 
576 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
577 {
578 	struct radeon_device *rdev = bo->rdev;
579 	struct radeon_surface_reg *reg;
580 
581 	if (bo->surface_reg == -1)
582 		return;
583 
584 	reg = &rdev->surface_regs[bo->surface_reg];
585 	radeon_clear_surface_reg(rdev, bo->surface_reg);
586 
587 	reg->bo = NULL;
588 	bo->surface_reg = -1;
589 }
590 
591 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
592 				uint32_t tiling_flags, uint32_t pitch)
593 {
594 	struct radeon_device *rdev = bo->rdev;
595 	int r;
596 
597 	if (rdev->family >= CHIP_CEDAR) {
598 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
599 
600 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
601 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
602 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
603 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
604 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
605 		switch (bankw) {
606 		case 0:
607 		case 1:
608 		case 2:
609 		case 4:
610 		case 8:
611 			break;
612 		default:
613 			return -EINVAL;
614 		}
615 		switch (bankh) {
616 		case 0:
617 		case 1:
618 		case 2:
619 		case 4:
620 		case 8:
621 			break;
622 		default:
623 			return -EINVAL;
624 		}
625 		switch (mtaspect) {
626 		case 0:
627 		case 1:
628 		case 2:
629 		case 4:
630 		case 8:
631 			break;
632 		default:
633 			return -EINVAL;
634 		}
635 		if (tilesplit > 6) {
636 			return -EINVAL;
637 		}
638 		if (stilesplit > 6) {
639 			return -EINVAL;
640 		}
641 	}
642 	r = radeon_bo_reserve(bo, false);
643 	if (unlikely(r != 0))
644 		return r;
645 	bo->tiling_flags = tiling_flags;
646 	bo->pitch = pitch;
647 	radeon_bo_unreserve(bo);
648 	return 0;
649 }
650 
651 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
652 				uint32_t *tiling_flags,
653 				uint32_t *pitch)
654 {
655 	lockdep_assert_held(&bo->tbo.resv->lock.base);
656 
657 	if (tiling_flags)
658 		*tiling_flags = bo->tiling_flags;
659 	if (pitch)
660 		*pitch = bo->pitch;
661 }
662 
663 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
664 				bool force_drop)
665 {
666 	if (!force_drop)
667 		lockdep_assert_held(&bo->tbo.resv->lock.base);
668 
669 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
670 		return 0;
671 
672 	if (force_drop) {
673 		radeon_bo_clear_surface_reg(bo);
674 		return 0;
675 	}
676 
677 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
678 		if (!has_moved)
679 			return 0;
680 
681 		if (bo->surface_reg >= 0)
682 			radeon_bo_clear_surface_reg(bo);
683 		return 0;
684 	}
685 
686 	if ((bo->surface_reg >= 0) && !has_moved)
687 		return 0;
688 
689 	return radeon_bo_get_surface_reg(bo);
690 }
691 
692 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
693 			   struct ttm_mem_reg *new_mem)
694 {
695 	struct radeon_bo *rbo;
696 
697 	if (!radeon_ttm_bo_is_radeon_bo(bo))
698 		return;
699 
700 	rbo = container_of(bo, struct radeon_bo, tbo);
701 	radeon_bo_check_tiling(rbo, 0, 1);
702 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
703 
704 	/* update statistics */
705 	if (!new_mem)
706 		return;
707 
708 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
709 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
710 }
711 
712 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
713 {
714 	struct radeon_device *rdev;
715 	struct radeon_bo *rbo;
716 	unsigned long offset, size;
717 	int r;
718 
719 	if (!radeon_ttm_bo_is_radeon_bo(bo))
720 		return 0;
721 	rbo = container_of(bo, struct radeon_bo, tbo);
722 	radeon_bo_check_tiling(rbo, 0, 0);
723 	rdev = rbo->rdev;
724 	if (bo->mem.mem_type != TTM_PL_VRAM)
725 		return 0;
726 
727 	size = bo->mem.num_pages << PAGE_SHIFT;
728 	offset = bo->mem.start << PAGE_SHIFT;
729 	if ((offset + size) <= rdev->mc.visible_vram_size)
730 		return 0;
731 
732 	/* hurrah the memory is not visible ! */
733 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
734 	rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
735 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
736 	if (unlikely(r == -ENOMEM)) {
737 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
738 		return ttm_bo_validate(bo, &rbo->placement, false, false);
739 	} else if (unlikely(r != 0)) {
740 		return r;
741 	}
742 
743 	offset = bo->mem.start << PAGE_SHIFT;
744 	/* this should never happen */
745 	if ((offset + size) > rdev->mc.visible_vram_size)
746 		return -EINVAL;
747 
748 	return 0;
749 }
750 
751 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
752 {
753 	int r;
754 
755 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
756 	if (unlikely(r != 0))
757 		return r;
758 	spin_lock(&bo->tbo.bdev->fence_lock);
759 	if (mem_type)
760 		*mem_type = bo->tbo.mem.mem_type;
761 	if (bo->tbo.sync_obj)
762 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
763 	spin_unlock(&bo->tbo.bdev->fence_lock);
764 	ttm_bo_unreserve(&bo->tbo);
765 	return r;
766 }
767