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