xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/radeon/radeon_object.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
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_bo_clear_va(struct radeon_bo *bo)
50 {
51 	struct radeon_bo_va *bo_va, *tmp;
52 
53 	list_for_each_entry_safe(bo_va, tmp, &bo->va, bo_list) {
54 		/* remove from all vm address space */
55 		radeon_vm_bo_rmv(bo->rdev, bo_va);
56 	}
57 }
58 
59 static void radeon_update_memory_usage(struct radeon_bo *bo,
60 				       unsigned mem_type, int sign)
61 {
62 	struct radeon_device *rdev = bo->rdev;
63 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
64 
65 	switch (mem_type) {
66 	case TTM_PL_TT:
67 		if (sign > 0)
68 			atomic64_add(size, &rdev->gtt_usage);
69 		else
70 			atomic64_sub(size, &rdev->gtt_usage);
71 		break;
72 	case TTM_PL_VRAM:
73 		if (sign > 0)
74 			atomic64_add(size, &rdev->vram_usage);
75 		else
76 			atomic64_sub(size, &rdev->vram_usage);
77 		break;
78 	}
79 }
80 
81 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
82 {
83 	struct radeon_bo *bo;
84 
85 	bo = container_of(tbo, struct radeon_bo, tbo);
86 
87 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
88 
89 	mutex_lock(&bo->rdev->gem.mutex);
90 	list_del_init(&bo->list);
91 	mutex_unlock(&bo->rdev->gem.mutex);
92 	radeon_bo_clear_surface_reg(bo);
93 	radeon_bo_clear_va(bo);
94 	drm_gem_object_release(&bo->gem_base);
95 	kfree(bo);
96 }
97 
98 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
99 {
100 	if (bo->destroy == &radeon_ttm_bo_destroy)
101 		return true;
102 	return false;
103 }
104 
105 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
106 {
107 	u32 c = 0, i;
108 
109 	rbo->placement.fpfn = 0;
110 	rbo->placement.lpfn = 0;
111 	rbo->placement.placement = rbo->placements;
112 	rbo->placement.busy_placement = rbo->placements;
113 	if (domain & RADEON_GEM_DOMAIN_VRAM)
114 		rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
115 					TTM_PL_FLAG_VRAM;
116 	if (domain & RADEON_GEM_DOMAIN_GTT) {
117 		if (rbo->rdev->flags & RADEON_IS_AGP) {
118 			rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_TT;
119 		} else {
120 			rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_TT;
121 		}
122 	}
123 	if (domain & RADEON_GEM_DOMAIN_CPU) {
124 		if (rbo->rdev->flags & RADEON_IS_AGP) {
125 			rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_SYSTEM;
126 		} else {
127 			rbo->placements[c++] = TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM;
128 		}
129 	}
130 	if (!c)
131 		rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
132 	rbo->placement.num_placement = c;
133 	rbo->placement.num_busy_placement = c;
134 
135 	/*
136 	 * Use two-ended allocation depending on the buffer size to
137 	 * improve fragmentation quality.
138 	 * 512kb was measured as the most optimal number.
139 	 */
140 	if (rbo->tbo.mem.size > 512 * 1024) {
141 		for (i = 0; i < c; i++) {
142 			rbo->placements[i] |= TTM_PL_FLAG_TOPDOWN;
143 		}
144 	}
145 }
146 
147 int radeon_bo_create(struct radeon_device *rdev,
148 		     unsigned long size, int byte_align, bool kernel, u32 domain,
149 		     struct sg_table *sg, struct radeon_bo **bo_ptr)
150 {
151 	struct radeon_bo *bo;
152 	enum ttm_bo_type type;
153 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
154 	size_t acc_size;
155 	int r;
156 
157 #ifdef __NetBSD__		/* XXX ALIGN means something else.  */
158 	size = round_up(size, PAGE_SIZE);
159 #else
160 	size = ALIGN(size, PAGE_SIZE);
161 #endif
162 
163 	if (kernel) {
164 		type = ttm_bo_type_kernel;
165 	} else if (sg) {
166 		type = ttm_bo_type_sg;
167 	} else {
168 		type = ttm_bo_type_device;
169 	}
170 	*bo_ptr = NULL;
171 
172 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
173 				       sizeof(struct radeon_bo));
174 
175 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
176 	if (bo == NULL)
177 		return -ENOMEM;
178 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
179 	if (unlikely(r)) {
180 		kfree(bo);
181 		return r;
182 	}
183 	bo->rdev = rdev;
184 	bo->surface_reg = -1;
185 	INIT_LIST_HEAD(&bo->list);
186 	INIT_LIST_HEAD(&bo->va);
187 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
188 	                               RADEON_GEM_DOMAIN_GTT |
189 	                               RADEON_GEM_DOMAIN_CPU);
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 void radeon_bo_unref(struct radeon_bo **bo)
240 {
241 	struct ttm_buffer_object *tbo;
242 	struct radeon_device *rdev;
243 
244 	if ((*bo) == NULL)
245 		return;
246 	rdev = (*bo)->rdev;
247 	tbo = &((*bo)->tbo);
248 	down_read(&rdev->pm.mclk_lock);
249 	ttm_bo_unref(&tbo);
250 	up_read(&rdev->pm.mclk_lock);
251 	if (tbo == NULL)
252 		*bo = NULL;
253 }
254 
255 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
256 			     u64 *gpu_addr)
257 {
258 	int r, i;
259 
260 	if (bo->pin_count) {
261 		bo->pin_count++;
262 		if (gpu_addr)
263 			*gpu_addr = radeon_bo_gpu_offset(bo);
264 
265 		if (max_offset != 0) {
266 			u64 domain_start;
267 
268 			if (domain == RADEON_GEM_DOMAIN_VRAM)
269 				domain_start = bo->rdev->mc.vram_start;
270 			else
271 				domain_start = bo->rdev->mc.gtt_start;
272 			WARN_ON_ONCE(max_offset <
273 				     (radeon_bo_gpu_offset(bo) - domain_start));
274 		}
275 
276 		return 0;
277 	}
278 	radeon_ttm_placement_from_domain(bo, domain);
279 	if (domain == RADEON_GEM_DOMAIN_VRAM) {
280 		/* force to pin into visible video ram */
281 		bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
282 	}
283 	if (max_offset) {
284 		u64 lpfn = max_offset >> PAGE_SHIFT;
285 
286 		if (!bo->placement.lpfn)
287 			bo->placement.lpfn = bo->rdev->mc.gtt_size >> PAGE_SHIFT;
288 
289 		if (lpfn < bo->placement.lpfn)
290 			bo->placement.lpfn = lpfn;
291 	}
292 	for (i = 0; i < bo->placement.num_placement; i++)
293 		bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
294 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
295 	if (likely(r == 0)) {
296 		bo->pin_count = 1;
297 		if (gpu_addr != NULL)
298 			*gpu_addr = radeon_bo_gpu_offset(bo);
299 	}
300 	if (unlikely(r != 0))
301 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
302 	return r;
303 }
304 
305 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
306 {
307 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
308 }
309 
310 int radeon_bo_unpin(struct radeon_bo *bo)
311 {
312 	int r, i;
313 
314 	if (!bo->pin_count) {
315 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
316 		return 0;
317 	}
318 	bo->pin_count--;
319 	if (bo->pin_count)
320 		return 0;
321 	for (i = 0; i < bo->placement.num_placement; i++)
322 		bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
323 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
324 	if (unlikely(r != 0))
325 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
326 	return r;
327 }
328 
329 int radeon_bo_evict_vram(struct radeon_device *rdev)
330 {
331 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
332 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
333 		if (rdev->mc.igp_sideport_enabled == false)
334 			/* Useless to evict on IGP chips */
335 			return 0;
336 	}
337 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
338 }
339 
340 void radeon_bo_force_delete(struct radeon_device *rdev)
341 {
342 	struct radeon_bo *bo, *n;
343 
344 	if (list_empty(&rdev->gem.objects)) {
345 		return;
346 	}
347 	dev_err(rdev->dev, "Userspace still has active objects !\n");
348 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
349 		mutex_lock(&rdev->ddev->struct_mutex);
350 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
351 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
352 			*((unsigned long *)&bo->gem_base.refcount));
353 		mutex_lock(&bo->rdev->gem.mutex);
354 		list_del_init(&bo->list);
355 		mutex_unlock(&bo->rdev->gem.mutex);
356 		/* this should unref the ttm bo */
357 		drm_gem_object_unreference(&bo->gem_base);
358 		mutex_unlock(&rdev->ddev->struct_mutex);
359 	}
360 }
361 
362 int radeon_bo_init(struct radeon_device *rdev)
363 {
364 	/* Add an MTRR for the VRAM */
365 	if (!rdev->fastfb_working) {
366 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
367 						      rdev->mc.aper_size);
368 	}
369 #ifdef __NetBSD__
370 	if (rdev->mc.aper_base)
371 		pmap_pv_track(rdev->mc.aper_base, rdev->mc.aper_size);
372 #endif
373 	DRM_INFO("Detected VRAM RAM=%"PRIx64"M, BAR=%lluM\n",
374 		rdev->mc.mc_vram_size >> 20,
375 		(unsigned long long)rdev->mc.aper_size >> 20);
376 	DRM_INFO("RAM width %dbits %cDR\n",
377 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
378 	return radeon_ttm_init(rdev);
379 }
380 
381 void radeon_bo_fini(struct radeon_device *rdev)
382 {
383 	radeon_ttm_fini(rdev);
384 #ifdef __NetBSD__
385 	if (rdev->mc.aper_base)
386 		pmap_pv_untrack(rdev->mc.aper_base, rdev->mc.aper_size);
387 #endif
388 	arch_phys_wc_del(rdev->mc.vram_mtrr);
389 }
390 
391 /* Returns how many bytes TTM can move per IB.
392  */
393 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
394 {
395 	u64 real_vram_size = rdev->mc.real_vram_size;
396 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
397 
398 	/* This function is based on the current VRAM usage.
399 	 *
400 	 * - If all of VRAM is free, allow relocating the number of bytes that
401 	 *   is equal to 1/4 of the size of VRAM for this IB.
402 
403 	 * - If more than one half of VRAM is occupied, only allow relocating
404 	 *   1 MB of data for this IB.
405 	 *
406 	 * - From 0 to one half of used VRAM, the threshold decreases
407 	 *   linearly.
408 	 *         __________________
409 	 * 1/4 of -|\               |
410 	 * VRAM    | \              |
411 	 *         |  \             |
412 	 *         |   \            |
413 	 *         |    \           |
414 	 *         |     \          |
415 	 *         |      \         |
416 	 *         |       \________|1 MB
417 	 *         |----------------|
418 	 *    VRAM 0 %             100 %
419 	 *         used            used
420 	 *
421 	 * Note: It's a threshold, not a limit. The threshold must be crossed
422 	 * for buffer relocations to stop, so any buffer of an arbitrary size
423 	 * can be moved as long as the threshold isn't crossed before
424 	 * the relocation takes place. We don't want to disable buffer
425 	 * relocations completely.
426 	 *
427 	 * The idea is that buffers should be placed in VRAM at creation time
428 	 * and TTM should only do a minimum number of relocations during
429 	 * command submission. In practice, you need to submit at least
430 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
431 	 *
432 	 * Also, things can get pretty crazy under memory pressure and actual
433 	 * VRAM usage can change a lot, so playing safe even at 50% does
434 	 * consistently increase performance.
435 	 */
436 
437 	u64 half_vram = real_vram_size >> 1;
438 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
439 	u64 bytes_moved_threshold = half_free_vram >> 1;
440 	return max(bytes_moved_threshold, 1024*1024ull);
441 }
442 
443 int radeon_bo_list_validate(struct radeon_device *rdev,
444 			    struct ww_acquire_ctx *ticket,
445 			    struct list_head *head, int ring)
446 {
447 	struct radeon_cs_reloc *lobj;
448 	struct radeon_bo *bo;
449 	int r;
450 	u64 bytes_moved = 0, initial_bytes_moved;
451 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
452 
453 	r = ttm_eu_reserve_buffers(ticket, head);
454 	if (unlikely(r != 0)) {
455 		return r;
456 	}
457 
458 	list_for_each_entry(lobj, head, tv.head) {
459 		bo = lobj->robj;
460 		if (!bo->pin_count) {
461 			u32 domain = lobj->domain;
462 			u32 current_domain =
463 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
464 
465 			/* Check if this buffer will be moved and don't move it
466 			 * if we have moved too many buffers for this IB already.
467 			 *
468 			 * Note that this allows moving at least one buffer of
469 			 * any size, because it doesn't take the current "bo"
470 			 * into account. We don't want to disallow buffer moves
471 			 * completely.
472 			 */
473 			if ((lobj->alt_domain & current_domain) != 0 &&
474 			    (domain & current_domain) == 0 && /* will be moved */
475 			    bytes_moved > bytes_moved_threshold) {
476 				/* don't move it */
477 				domain = current_domain;
478 			}
479 
480 		retry:
481 			radeon_ttm_placement_from_domain(bo, domain);
482 			if (ring == R600_RING_TYPE_UVD_INDEX)
483 				radeon_uvd_force_into_uvd_segment(bo);
484 
485 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
486 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
487 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
488 				       initial_bytes_moved;
489 
490 			if (unlikely(r)) {
491 				if (r != -ERESTARTSYS && domain != lobj->alt_domain) {
492 					domain = lobj->alt_domain;
493 					goto retry;
494 				}
495 				ttm_eu_backoff_reservation(ticket, head);
496 				return r;
497 			}
498 		}
499 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
500 		lobj->tiling_flags = bo->tiling_flags;
501 	}
502 	return 0;
503 }
504 
505 #ifdef __NetBSD__
506 /* XXX Fill me in!  */
507 #else
508 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
509 			     struct vm_area_struct *vma)
510 {
511 	return ttm_fbdev_mmap(vma, &bo->tbo);
512 }
513 #endif
514 
515 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
516 {
517 	struct radeon_device *rdev = bo->rdev;
518 	struct radeon_surface_reg *reg;
519 	struct radeon_bo *old_object;
520 	int steal;
521 	int i;
522 
523 	lockdep_assert_held(&bo->tbo.resv->lock.base);
524 
525 	if (!bo->tiling_flags)
526 		return 0;
527 
528 	if (bo->surface_reg >= 0) {
529 		reg = &rdev->surface_regs[bo->surface_reg];
530 		i = bo->surface_reg;
531 		goto out;
532 	}
533 
534 	steal = -1;
535 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
536 
537 		reg = &rdev->surface_regs[i];
538 		if (!reg->bo)
539 			break;
540 
541 		old_object = reg->bo;
542 		if (old_object->pin_count == 0)
543 			steal = i;
544 	}
545 
546 	/* if we are all out */
547 	if (i == RADEON_GEM_MAX_SURFACES) {
548 		if (steal == -1)
549 			return -ENOMEM;
550 		/* find someone with a surface reg and nuke their BO */
551 		reg = &rdev->surface_regs[steal];
552 		old_object = reg->bo;
553 		/* blow away the mapping */
554 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
555 		ttm_bo_unmap_virtual(&old_object->tbo);
556 		old_object->surface_reg = -1;
557 		i = steal;
558 	}
559 
560 	bo->surface_reg = i;
561 	reg->bo = bo;
562 
563 out:
564 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
565 			       bo->tbo.mem.start << PAGE_SHIFT,
566 			       bo->tbo.num_pages << PAGE_SHIFT);
567 	return 0;
568 }
569 
570 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
571 {
572 	struct radeon_device *rdev = bo->rdev;
573 	struct radeon_surface_reg *reg;
574 
575 	if (bo->surface_reg == -1)
576 		return;
577 
578 	reg = &rdev->surface_regs[bo->surface_reg];
579 	radeon_clear_surface_reg(rdev, bo->surface_reg);
580 
581 	reg->bo = NULL;
582 	bo->surface_reg = -1;
583 }
584 
585 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
586 				uint32_t tiling_flags, uint32_t pitch)
587 {
588 	struct radeon_device *rdev = bo->rdev;
589 	int r;
590 
591 	if (rdev->family >= CHIP_CEDAR) {
592 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
593 
594 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
595 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
596 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
597 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
598 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
599 		switch (bankw) {
600 		case 0:
601 		case 1:
602 		case 2:
603 		case 4:
604 		case 8:
605 			break;
606 		default:
607 			return -EINVAL;
608 		}
609 		switch (bankh) {
610 		case 0:
611 		case 1:
612 		case 2:
613 		case 4:
614 		case 8:
615 			break;
616 		default:
617 			return -EINVAL;
618 		}
619 		switch (mtaspect) {
620 		case 0:
621 		case 1:
622 		case 2:
623 		case 4:
624 		case 8:
625 			break;
626 		default:
627 			return -EINVAL;
628 		}
629 		if (tilesplit > 6) {
630 			return -EINVAL;
631 		}
632 		if (stilesplit > 6) {
633 			return -EINVAL;
634 		}
635 	}
636 	r = radeon_bo_reserve(bo, false);
637 	if (unlikely(r != 0))
638 		return r;
639 	bo->tiling_flags = tiling_flags;
640 	bo->pitch = pitch;
641 	radeon_bo_unreserve(bo);
642 	return 0;
643 }
644 
645 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
646 				uint32_t *tiling_flags,
647 				uint32_t *pitch)
648 {
649 	lockdep_assert_held(&bo->tbo.resv->lock.base);
650 
651 	if (tiling_flags)
652 		*tiling_flags = bo->tiling_flags;
653 	if (pitch)
654 		*pitch = bo->pitch;
655 }
656 
657 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
658 				bool force_drop)
659 {
660 	if (!force_drop)
661 		lockdep_assert_held(&bo->tbo.resv->lock.base);
662 
663 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
664 		return 0;
665 
666 	if (force_drop) {
667 		radeon_bo_clear_surface_reg(bo);
668 		return 0;
669 	}
670 
671 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
672 		if (!has_moved)
673 			return 0;
674 
675 		if (bo->surface_reg >= 0)
676 			radeon_bo_clear_surface_reg(bo);
677 		return 0;
678 	}
679 
680 	if ((bo->surface_reg >= 0) && !has_moved)
681 		return 0;
682 
683 	return radeon_bo_get_surface_reg(bo);
684 }
685 
686 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
687 			   struct ttm_mem_reg *new_mem)
688 {
689 	struct radeon_bo *rbo;
690 
691 	if (!radeon_ttm_bo_is_radeon_bo(bo))
692 		return;
693 
694 	rbo = container_of(bo, struct radeon_bo, tbo);
695 	radeon_bo_check_tiling(rbo, 0, 1);
696 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
697 
698 	/* update statistics */
699 	if (!new_mem)
700 		return;
701 
702 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
703 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
704 }
705 
706 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
707 {
708 	struct radeon_device *rdev;
709 	struct radeon_bo *rbo;
710 	unsigned long offset, size;
711 	int r;
712 
713 	if (!radeon_ttm_bo_is_radeon_bo(bo))
714 		return 0;
715 	rbo = container_of(bo, struct radeon_bo, tbo);
716 	radeon_bo_check_tiling(rbo, 0, 0);
717 	rdev = rbo->rdev;
718 	if (bo->mem.mem_type != TTM_PL_VRAM)
719 		return 0;
720 
721 	size = bo->mem.num_pages << PAGE_SHIFT;
722 	offset = bo->mem.start << PAGE_SHIFT;
723 	if ((offset + size) <= rdev->mc.visible_vram_size)
724 		return 0;
725 
726 	/* hurrah the memory is not visible ! */
727 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
728 	rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
729 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
730 	if (unlikely(r == -ENOMEM)) {
731 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
732 		return ttm_bo_validate(bo, &rbo->placement, false, false);
733 	} else if (unlikely(r != 0)) {
734 		return r;
735 	}
736 
737 	offset = bo->mem.start << PAGE_SHIFT;
738 	/* this should never happen */
739 	if ((offset + size) > rdev->mc.visible_vram_size)
740 		return -EINVAL;
741 
742 	return 0;
743 }
744 
745 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
746 {
747 	int r;
748 
749 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
750 	if (unlikely(r != 0))
751 		return r;
752 	spin_lock(&bo->tbo.bdev->fence_lock);
753 	if (mem_type)
754 		*mem_type = bo->tbo.mem.mem_type;
755 	if (bo->tbo.sync_obj)
756 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
757 	spin_unlock(&bo->tbo.bdev->fence_lock);
758 	ttm_bo_unreserve(&bo->tbo);
759 	return r;
760 }
761