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