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 	radeon_mn_unregister(bo);
79 
80 	mutex_lock(&bo->rdev->gem.mutex);
81 	list_del_init(&bo->list);
82 	mutex_unlock(&bo->rdev->gem.mutex);
83 	radeon_bo_clear_surface_reg(bo);
84 	WARN_ON(!list_empty(&bo->va));
85 	drm_gem_object_release(&bo->gem_base);
86 	kfree(bo);
87 }
88 
89 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
90 {
91 	if (bo->destroy == &radeon_ttm_bo_destroy)
92 		return true;
93 	return false;
94 }
95 
96 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
97 {
98 	u32 c = 0, i;
99 
100 	rbo->placement.placement = rbo->placements;
101 	rbo->placement.busy_placement = rbo->placements;
102 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
103 		/* Try placing BOs which don't need CPU access outside of the
104 		 * CPU accessible part of VRAM
105 		 */
106 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
107 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
108 			rbo->placements[c].fpfn =
109 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
110 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
111 						     TTM_PL_FLAG_UNCACHED |
112 						     TTM_PL_FLAG_VRAM;
113 		}
114 
115 		rbo->placements[c].fpfn = 0;
116 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
117 					     TTM_PL_FLAG_UNCACHED |
118 					     TTM_PL_FLAG_VRAM;
119 	}
120 
121 	if (domain & RADEON_GEM_DOMAIN_GTT) {
122 		if (rbo->flags & RADEON_GEM_GTT_UC) {
123 			rbo->placements[c].fpfn = 0;
124 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
125 				TTM_PL_FLAG_TT;
126 
127 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
128 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
129 			rbo->placements[c].fpfn = 0;
130 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
131 				TTM_PL_FLAG_UNCACHED |
132 				TTM_PL_FLAG_TT;
133 		} else {
134 			rbo->placements[c].fpfn = 0;
135 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
136 						     TTM_PL_FLAG_TT;
137 		}
138 	}
139 
140 	if (domain & RADEON_GEM_DOMAIN_CPU) {
141 		if (rbo->flags & RADEON_GEM_GTT_UC) {
142 			rbo->placements[c].fpfn = 0;
143 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
144 				TTM_PL_FLAG_SYSTEM;
145 
146 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
147 		    rbo->rdev->flags & RADEON_IS_AGP) {
148 			rbo->placements[c].fpfn = 0;
149 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
150 				TTM_PL_FLAG_UNCACHED |
151 				TTM_PL_FLAG_SYSTEM;
152 		} else {
153 			rbo->placements[c].fpfn = 0;
154 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
155 						     TTM_PL_FLAG_SYSTEM;
156 		}
157 	}
158 	if (!c) {
159 		rbo->placements[c].fpfn = 0;
160 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
161 					     TTM_PL_FLAG_SYSTEM;
162 	}
163 
164 	rbo->placement.num_placement = c;
165 	rbo->placement.num_busy_placement = c;
166 
167 	for (i = 0; i < c; ++i) {
168 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
169 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
170 		    !rbo->placements[i].fpfn)
171 			rbo->placements[i].lpfn =
172 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
173 		else
174 			rbo->placements[i].lpfn = 0;
175 	}
176 
177 	/*
178 	 * Use two-ended allocation depending on the buffer size to
179 	 * improve fragmentation quality.
180 	 * 512kb was measured as the most optimal number.
181 	 */
182 	if (rbo->tbo.mem.size > 512 * 1024) {
183 		for (i = 0; i < c; i++) {
184 			rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
185 		}
186 	}
187 }
188 
189 int radeon_bo_create(struct radeon_device *rdev,
190 		     unsigned long size, int byte_align, bool kernel,
191 		     u32 domain, u32 flags, struct sg_table *sg,
192 		     struct reservation_object *resv,
193 		     struct radeon_bo **bo_ptr)
194 {
195 	struct radeon_bo *bo;
196 	enum ttm_bo_type type;
197 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
198 	size_t acc_size;
199 	int r;
200 
201 	size = ALIGN(size, PAGE_SIZE);
202 
203 	if (kernel) {
204 		type = ttm_bo_type_kernel;
205 	} else if (sg) {
206 		type = ttm_bo_type_sg;
207 	} else {
208 		type = ttm_bo_type_device;
209 	}
210 	*bo_ptr = NULL;
211 
212 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
213 				       sizeof(struct radeon_bo));
214 
215 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
216 	if (bo == NULL)
217 		return -ENOMEM;
218 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
219 	if (unlikely(r)) {
220 		kfree(bo);
221 		return r;
222 	}
223 	bo->rdev = rdev;
224 	bo->surface_reg = -1;
225 	INIT_LIST_HEAD(&bo->list);
226 	INIT_LIST_HEAD(&bo->va);
227 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
228 	                               RADEON_GEM_DOMAIN_GTT |
229 	                               RADEON_GEM_DOMAIN_CPU);
230 
231 	bo->flags = flags;
232 	/* PCI GART is always snooped */
233 	if (!(rdev->flags & RADEON_IS_PCIE))
234 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
235 
236 #ifdef CONFIG_X86_32
237 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
238 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
239 	 */
240 	bo->flags &= ~RADEON_GEM_GTT_WC;
241 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
242 	/* Don't try to enable write-combining when it can't work, or things
243 	 * may be slow
244 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
245 	 */
246 
247 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
248 	 thanks to write-combining
249 
250 	DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
251 		      "better performance thanks to write-combining\n");
252 	bo->flags &= ~RADEON_GEM_GTT_WC;
253 #endif
254 
255 	radeon_ttm_placement_from_domain(bo, domain);
256 	/* Kernel allocation are uninterruptible */
257 	down_read(&rdev->pm.mclk_lock);
258 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
259 			&bo->placement, page_align, !kernel, NULL,
260 			acc_size, sg, resv, &radeon_ttm_bo_destroy);
261 	up_read(&rdev->pm.mclk_lock);
262 	if (unlikely(r != 0)) {
263 		return r;
264 	}
265 	*bo_ptr = bo;
266 
267 	trace_radeon_bo_create(bo);
268 
269 	return 0;
270 }
271 
272 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
273 {
274 	bool is_iomem;
275 	int r;
276 
277 	if (bo->kptr) {
278 		if (ptr) {
279 			*ptr = bo->kptr;
280 		}
281 		return 0;
282 	}
283 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
284 	if (r) {
285 		return r;
286 	}
287 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
288 	if (ptr) {
289 		*ptr = bo->kptr;
290 	}
291 	radeon_bo_check_tiling(bo, 0, 0);
292 	return 0;
293 }
294 
295 void radeon_bo_kunmap(struct radeon_bo *bo)
296 {
297 	if (bo->kptr == NULL)
298 		return;
299 	bo->kptr = NULL;
300 	radeon_bo_check_tiling(bo, 0, 0);
301 	ttm_bo_kunmap(&bo->kmap);
302 }
303 
304 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
305 {
306 	if (bo == NULL)
307 		return NULL;
308 
309 	ttm_bo_reference(&bo->tbo);
310 	return bo;
311 }
312 
313 void radeon_bo_unref(struct radeon_bo **bo)
314 {
315 	struct ttm_buffer_object *tbo;
316 	struct radeon_device *rdev;
317 
318 	if ((*bo) == NULL)
319 		return;
320 	rdev = (*bo)->rdev;
321 	tbo = &((*bo)->tbo);
322 	ttm_bo_unref(&tbo);
323 	if (tbo == NULL)
324 		*bo = NULL;
325 }
326 
327 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
328 			     u64 *gpu_addr)
329 {
330 	int r, i;
331 
332 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
333 		return -EPERM;
334 
335 	if (bo->pin_count) {
336 		bo->pin_count++;
337 		if (gpu_addr)
338 			*gpu_addr = radeon_bo_gpu_offset(bo);
339 
340 		if (max_offset != 0) {
341 			u64 domain_start;
342 
343 			if (domain == RADEON_GEM_DOMAIN_VRAM)
344 				domain_start = bo->rdev->mc.vram_start;
345 			else
346 				domain_start = bo->rdev->mc.gtt_start;
347 			WARN_ON_ONCE(max_offset <
348 				     (radeon_bo_gpu_offset(bo) - domain_start));
349 		}
350 
351 		return 0;
352 	}
353 	radeon_ttm_placement_from_domain(bo, domain);
354 	for (i = 0; i < bo->placement.num_placement; i++) {
355 		/* force to pin into visible video ram */
356 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
357 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
358 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
359 			bo->placements[i].lpfn =
360 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
361 		else
362 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
363 
364 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
365 	}
366 
367 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
368 	if (likely(r == 0)) {
369 		bo->pin_count = 1;
370 		if (gpu_addr != NULL)
371 			*gpu_addr = radeon_bo_gpu_offset(bo);
372 		if (domain == RADEON_GEM_DOMAIN_VRAM)
373 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
374 		else
375 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
376 	} else {
377 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
378 	}
379 	return r;
380 }
381 
382 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
383 {
384 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
385 }
386 
387 int radeon_bo_unpin(struct radeon_bo *bo)
388 {
389 	int r, i;
390 
391 	if (!bo->pin_count) {
392 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
393 		return 0;
394 	}
395 	bo->pin_count--;
396 	if (bo->pin_count)
397 		return 0;
398 	for (i = 0; i < bo->placement.num_placement; i++) {
399 		bo->placements[i].lpfn = 0;
400 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
401 	}
402 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
403 	if (likely(r == 0)) {
404 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
405 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
406 		else
407 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
408 	} else {
409 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
410 	}
411 	return r;
412 }
413 
414 int radeon_bo_evict_vram(struct radeon_device *rdev)
415 {
416 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
417 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
418 		if (rdev->mc.igp_sideport_enabled == false)
419 			/* Useless to evict on IGP chips */
420 			return 0;
421 	}
422 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
423 }
424 
425 void radeon_bo_force_delete(struct radeon_device *rdev)
426 {
427 	struct radeon_bo *bo, *n;
428 
429 	if (list_empty(&rdev->gem.objects)) {
430 		return;
431 	}
432 	dev_err(rdev->dev, "Userspace still has active objects !\n");
433 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
434 		mutex_lock(&rdev->ddev->struct_mutex);
435 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
436 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
437 			*((unsigned long *)&bo->gem_base.refcount));
438 		mutex_lock(&bo->rdev->gem.mutex);
439 		list_del_init(&bo->list);
440 		mutex_unlock(&bo->rdev->gem.mutex);
441 		/* this should unref the ttm bo */
442 		drm_gem_object_unreference(&bo->gem_base);
443 		mutex_unlock(&rdev->ddev->struct_mutex);
444 	}
445 }
446 
447 int radeon_bo_init(struct radeon_device *rdev)
448 {
449 	/* Add an MTRR for the VRAM */
450 	if (!rdev->fastfb_working) {
451 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
452 						      rdev->mc.aper_size);
453 	}
454 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
455 		rdev->mc.mc_vram_size >> 20,
456 		(unsigned long long)rdev->mc.aper_size >> 20);
457 	DRM_INFO("RAM width %dbits %cDR\n",
458 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
459 	return radeon_ttm_init(rdev);
460 }
461 
462 void radeon_bo_fini(struct radeon_device *rdev)
463 {
464 	radeon_ttm_fini(rdev);
465 	arch_phys_wc_del(rdev->mc.vram_mtrr);
466 }
467 
468 /* Returns how many bytes TTM can move per IB.
469  */
470 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
471 {
472 	u64 real_vram_size = rdev->mc.real_vram_size;
473 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
474 
475 	/* This function is based on the current VRAM usage.
476 	 *
477 	 * - If all of VRAM is free, allow relocating the number of bytes that
478 	 *   is equal to 1/4 of the size of VRAM for this IB.
479 
480 	 * - If more than one half of VRAM is occupied, only allow relocating
481 	 *   1 MB of data for this IB.
482 	 *
483 	 * - From 0 to one half of used VRAM, the threshold decreases
484 	 *   linearly.
485 	 *         __________________
486 	 * 1/4 of -|\               |
487 	 * VRAM    | \              |
488 	 *         |  \             |
489 	 *         |   \            |
490 	 *         |    \           |
491 	 *         |     \          |
492 	 *         |      \         |
493 	 *         |       \________|1 MB
494 	 *         |----------------|
495 	 *    VRAM 0 %             100 %
496 	 *         used            used
497 	 *
498 	 * Note: It's a threshold, not a limit. The threshold must be crossed
499 	 * for buffer relocations to stop, so any buffer of an arbitrary size
500 	 * can be moved as long as the threshold isn't crossed before
501 	 * the relocation takes place. We don't want to disable buffer
502 	 * relocations completely.
503 	 *
504 	 * The idea is that buffers should be placed in VRAM at creation time
505 	 * and TTM should only do a minimum number of relocations during
506 	 * command submission. In practice, you need to submit at least
507 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
508 	 *
509 	 * Also, things can get pretty crazy under memory pressure and actual
510 	 * VRAM usage can change a lot, so playing safe even at 50% does
511 	 * consistently increase performance.
512 	 */
513 
514 	u64 half_vram = real_vram_size >> 1;
515 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
516 	u64 bytes_moved_threshold = half_free_vram >> 1;
517 	return max(bytes_moved_threshold, 1024*1024ull);
518 }
519 
520 int radeon_bo_list_validate(struct radeon_device *rdev,
521 			    struct ww_acquire_ctx *ticket,
522 			    struct list_head *head, int ring)
523 {
524 	struct radeon_bo_list *lobj;
525 	struct list_head duplicates;
526 	int r;
527 	u64 bytes_moved = 0, initial_bytes_moved;
528 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
529 
530 	INIT_LIST_HEAD(&duplicates);
531 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
532 	if (unlikely(r != 0)) {
533 		return r;
534 	}
535 
536 	list_for_each_entry(lobj, head, tv.head) {
537 		struct radeon_bo *bo = lobj->robj;
538 		if (!bo->pin_count) {
539 			u32 domain = lobj->prefered_domains;
540 			u32 allowed = lobj->allowed_domains;
541 			u32 current_domain =
542 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
543 
544 			/* Check if this buffer will be moved and don't move it
545 			 * if we have moved too many buffers for this IB already.
546 			 *
547 			 * Note that this allows moving at least one buffer of
548 			 * any size, because it doesn't take the current "bo"
549 			 * into account. We don't want to disallow buffer moves
550 			 * completely.
551 			 */
552 			if ((allowed & current_domain) != 0 &&
553 			    (domain & current_domain) == 0 && /* will be moved */
554 			    bytes_moved > bytes_moved_threshold) {
555 				/* don't move it */
556 				domain = current_domain;
557 			}
558 
559 		retry:
560 			radeon_ttm_placement_from_domain(bo, domain);
561 			if (ring == R600_RING_TYPE_UVD_INDEX)
562 				radeon_uvd_force_into_uvd_segment(bo, allowed);
563 
564 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
565 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
566 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
567 				       initial_bytes_moved;
568 
569 			if (unlikely(r)) {
570 				if (r != -ERESTARTSYS &&
571 				    domain != lobj->allowed_domains) {
572 					domain = lobj->allowed_domains;
573 					goto retry;
574 				}
575 				ttm_eu_backoff_reservation(ticket, head);
576 				return r;
577 			}
578 		}
579 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
580 		lobj->tiling_flags = bo->tiling_flags;
581 	}
582 
583 	list_for_each_entry(lobj, &duplicates, tv.head) {
584 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
585 		lobj->tiling_flags = lobj->robj->tiling_flags;
586 	}
587 
588 	return 0;
589 }
590 
591 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
592 {
593 	struct radeon_device *rdev = bo->rdev;
594 	struct radeon_surface_reg *reg;
595 	struct radeon_bo *old_object;
596 	int steal;
597 	int i;
598 
599 	lockdep_assert_held(&bo->tbo.resv->lock.base);
600 
601 	if (!bo->tiling_flags)
602 		return 0;
603 
604 	if (bo->surface_reg >= 0) {
605 		reg = &rdev->surface_regs[bo->surface_reg];
606 		i = bo->surface_reg;
607 		goto out;
608 	}
609 
610 	steal = -1;
611 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
612 
613 		reg = &rdev->surface_regs[i];
614 		if (!reg->bo)
615 			break;
616 
617 		old_object = reg->bo;
618 		if (old_object->pin_count == 0)
619 			steal = i;
620 	}
621 
622 	/* if we are all out */
623 	if (i == RADEON_GEM_MAX_SURFACES) {
624 		if (steal == -1)
625 			return -ENOMEM;
626 		/* find someone with a surface reg and nuke their BO */
627 		reg = &rdev->surface_regs[steal];
628 		old_object = reg->bo;
629 		/* blow away the mapping */
630 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
631 		ttm_bo_unmap_virtual(&old_object->tbo);
632 		old_object->surface_reg = -1;
633 		i = steal;
634 	}
635 
636 	bo->surface_reg = i;
637 	reg->bo = bo;
638 
639 out:
640 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
641 			       bo->tbo.mem.start << PAGE_SHIFT,
642 			       bo->tbo.num_pages << PAGE_SHIFT);
643 	return 0;
644 }
645 
646 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
647 {
648 	struct radeon_device *rdev = bo->rdev;
649 	struct radeon_surface_reg *reg;
650 
651 	if (bo->surface_reg == -1)
652 		return;
653 
654 	reg = &rdev->surface_regs[bo->surface_reg];
655 	radeon_clear_surface_reg(rdev, bo->surface_reg);
656 
657 	reg->bo = NULL;
658 	bo->surface_reg = -1;
659 }
660 
661 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
662 				uint32_t tiling_flags, uint32_t pitch)
663 {
664 	struct radeon_device *rdev = bo->rdev;
665 	int r;
666 
667 	if (rdev->family >= CHIP_CEDAR) {
668 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
669 
670 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
671 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
672 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
673 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
674 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
675 		switch (bankw) {
676 		case 0:
677 		case 1:
678 		case 2:
679 		case 4:
680 		case 8:
681 			break;
682 		default:
683 			return -EINVAL;
684 		}
685 		switch (bankh) {
686 		case 0:
687 		case 1:
688 		case 2:
689 		case 4:
690 		case 8:
691 			break;
692 		default:
693 			return -EINVAL;
694 		}
695 		switch (mtaspect) {
696 		case 0:
697 		case 1:
698 		case 2:
699 		case 4:
700 		case 8:
701 			break;
702 		default:
703 			return -EINVAL;
704 		}
705 		if (tilesplit > 6) {
706 			return -EINVAL;
707 		}
708 		if (stilesplit > 6) {
709 			return -EINVAL;
710 		}
711 	}
712 	r = radeon_bo_reserve(bo, false);
713 	if (unlikely(r != 0))
714 		return r;
715 	bo->tiling_flags = tiling_flags;
716 	bo->pitch = pitch;
717 	radeon_bo_unreserve(bo);
718 	return 0;
719 }
720 
721 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
722 				uint32_t *tiling_flags,
723 				uint32_t *pitch)
724 {
725 	lockdep_assert_held(&bo->tbo.resv->lock.base);
726 
727 	if (tiling_flags)
728 		*tiling_flags = bo->tiling_flags;
729 	if (pitch)
730 		*pitch = bo->pitch;
731 }
732 
733 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
734 				bool force_drop)
735 {
736 	if (!force_drop)
737 		lockdep_assert_held(&bo->tbo.resv->lock.base);
738 
739 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
740 		return 0;
741 
742 	if (force_drop) {
743 		radeon_bo_clear_surface_reg(bo);
744 		return 0;
745 	}
746 
747 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
748 		if (!has_moved)
749 			return 0;
750 
751 		if (bo->surface_reg >= 0)
752 			radeon_bo_clear_surface_reg(bo);
753 		return 0;
754 	}
755 
756 	if ((bo->surface_reg >= 0) && !has_moved)
757 		return 0;
758 
759 	return radeon_bo_get_surface_reg(bo);
760 }
761 
762 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
763 			   struct ttm_mem_reg *new_mem)
764 {
765 	struct radeon_bo *rbo;
766 
767 	if (!radeon_ttm_bo_is_radeon_bo(bo))
768 		return;
769 
770 	rbo = container_of(bo, struct radeon_bo, tbo);
771 	radeon_bo_check_tiling(rbo, 0, 1);
772 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
773 
774 	/* update statistics */
775 	if (!new_mem)
776 		return;
777 
778 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
779 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
780 }
781 
782 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
783 {
784 	struct radeon_device *rdev;
785 	struct radeon_bo *rbo;
786 	unsigned long offset, size, lpfn;
787 	int i, r;
788 
789 	if (!radeon_ttm_bo_is_radeon_bo(bo))
790 		return 0;
791 	rbo = container_of(bo, struct radeon_bo, tbo);
792 	radeon_bo_check_tiling(rbo, 0, 0);
793 	rdev = rbo->rdev;
794 	if (bo->mem.mem_type != TTM_PL_VRAM)
795 		return 0;
796 
797 	size = bo->mem.num_pages << PAGE_SHIFT;
798 	offset = bo->mem.start << PAGE_SHIFT;
799 	if ((offset + size) <= rdev->mc.visible_vram_size)
800 		return 0;
801 
802 	/* hurrah the memory is not visible ! */
803 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
804 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
805 	for (i = 0; i < rbo->placement.num_placement; i++) {
806 		/* Force into visible VRAM */
807 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
808 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
809 			rbo->placements[i].lpfn = lpfn;
810 	}
811 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
812 	if (unlikely(r == -ENOMEM)) {
813 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
814 		return ttm_bo_validate(bo, &rbo->placement, false, false);
815 	} else if (unlikely(r != 0)) {
816 		return r;
817 	}
818 
819 	offset = bo->mem.start << PAGE_SHIFT;
820 	/* this should never happen */
821 	if ((offset + size) > rdev->mc.visible_vram_size)
822 		return -EINVAL;
823 
824 	return 0;
825 }
826 
827 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
828 {
829 	int r;
830 
831 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
832 	if (unlikely(r != 0))
833 		return r;
834 	if (mem_type)
835 		*mem_type = bo->tbo.mem.mem_type;
836 
837 	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
838 	ttm_bo_unreserve(&bo->tbo);
839 	return r;
840 }
841 
842 /**
843  * radeon_bo_fence - add fence to buffer object
844  *
845  * @bo: buffer object in question
846  * @fence: fence to add
847  * @shared: true if fence should be added shared
848  *
849  */
850 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
851                      bool shared)
852 {
853 	struct reservation_object *resv = bo->tbo.resv;
854 
855 	if (shared)
856 		reservation_object_add_shared_fence(resv, &fence->base);
857 	else
858 		reservation_object_add_excl_fence(resv, &fence->base);
859 }
860