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 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
225 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
226 	 */
227 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
228 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
229 
230 #ifdef CONFIG_X86_32
231 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
232 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
233 	 */
234 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
235 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
236 	/* Don't try to enable write-combining when it can't work, or things
237 	 * may be slow
238 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
239 	 */
240 
241 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
242 	 thanks to write-combining
243 
244 	if (bo->flags & RADEON_GEM_GTT_WC)
245 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
246 			      "better performance thanks to write-combining\n");
247 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
248 #endif
249 
250 	radeon_ttm_placement_from_domain(bo, domain);
251 	/* Kernel allocation are uninterruptible */
252 	down_read(&rdev->pm.mclk_lock);
253 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
254 			&bo->placement, page_align, !kernel, NULL,
255 			acc_size, sg, resv, &radeon_ttm_bo_destroy);
256 	up_read(&rdev->pm.mclk_lock);
257 	if (unlikely(r != 0)) {
258 		return r;
259 	}
260 	*bo_ptr = bo;
261 
262 	trace_radeon_bo_create(bo);
263 
264 	return 0;
265 }
266 
267 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
268 {
269 	bool is_iomem;
270 	int r;
271 
272 	if (bo->kptr) {
273 		if (ptr) {
274 			*ptr = bo->kptr;
275 		}
276 		return 0;
277 	}
278 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
279 	if (r) {
280 		return r;
281 	}
282 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
283 	if (ptr) {
284 		*ptr = bo->kptr;
285 	}
286 	radeon_bo_check_tiling(bo, 0, 0);
287 	return 0;
288 }
289 
290 void radeon_bo_kunmap(struct radeon_bo *bo)
291 {
292 	if (bo->kptr == NULL)
293 		return;
294 	bo->kptr = NULL;
295 	radeon_bo_check_tiling(bo, 0, 0);
296 	ttm_bo_kunmap(&bo->kmap);
297 }
298 
299 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
300 {
301 	if (bo == NULL)
302 		return NULL;
303 
304 	ttm_bo_reference(&bo->tbo);
305 	return bo;
306 }
307 
308 void radeon_bo_unref(struct radeon_bo **bo)
309 {
310 	struct ttm_buffer_object *tbo;
311 	struct radeon_device *rdev;
312 
313 	if ((*bo) == NULL)
314 		return;
315 	rdev = (*bo)->rdev;
316 	tbo = &((*bo)->tbo);
317 	ttm_bo_unref(&tbo);
318 	if (tbo == NULL)
319 		*bo = NULL;
320 }
321 
322 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
323 			     u64 *gpu_addr)
324 {
325 	int r, i;
326 
327 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
328 		return -EPERM;
329 
330 	if (bo->pin_count) {
331 		bo->pin_count++;
332 		if (gpu_addr)
333 			*gpu_addr = radeon_bo_gpu_offset(bo);
334 
335 		if (max_offset != 0) {
336 			u64 domain_start;
337 
338 			if (domain == RADEON_GEM_DOMAIN_VRAM)
339 				domain_start = bo->rdev->mc.vram_start;
340 			else
341 				domain_start = bo->rdev->mc.gtt_start;
342 			WARN_ON_ONCE(max_offset <
343 				     (radeon_bo_gpu_offset(bo) - domain_start));
344 		}
345 
346 		return 0;
347 	}
348 	radeon_ttm_placement_from_domain(bo, domain);
349 	for (i = 0; i < bo->placement.num_placement; i++) {
350 		/* force to pin into visible video ram */
351 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
352 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
353 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
354 			bo->placements[i].lpfn =
355 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
356 		else
357 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
358 
359 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
360 	}
361 
362 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
363 	if (likely(r == 0)) {
364 		bo->pin_count = 1;
365 		if (gpu_addr != NULL)
366 			*gpu_addr = radeon_bo_gpu_offset(bo);
367 		if (domain == RADEON_GEM_DOMAIN_VRAM)
368 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
369 		else
370 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
371 	} else {
372 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
373 	}
374 	return r;
375 }
376 
377 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
378 {
379 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
380 }
381 
382 int radeon_bo_unpin(struct radeon_bo *bo)
383 {
384 	int r, i;
385 
386 	if (!bo->pin_count) {
387 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
388 		return 0;
389 	}
390 	bo->pin_count--;
391 	if (bo->pin_count)
392 		return 0;
393 	for (i = 0; i < bo->placement.num_placement; i++) {
394 		bo->placements[i].lpfn = 0;
395 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
396 	}
397 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
398 	if (likely(r == 0)) {
399 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
400 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
401 		else
402 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
403 	} else {
404 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
405 	}
406 	return r;
407 }
408 
409 int radeon_bo_evict_vram(struct radeon_device *rdev)
410 {
411 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
412 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
413 		if (rdev->mc.igp_sideport_enabled == false)
414 			/* Useless to evict on IGP chips */
415 			return 0;
416 	}
417 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
418 }
419 
420 void radeon_bo_force_delete(struct radeon_device *rdev)
421 {
422 	struct radeon_bo *bo, *n;
423 
424 	if (list_empty(&rdev->gem.objects)) {
425 		return;
426 	}
427 	dev_err(rdev->dev, "Userspace still has active objects !\n");
428 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
429 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
430 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
431 			*((unsigned long *)&bo->gem_base.refcount));
432 		mutex_lock(&bo->rdev->gem.mutex);
433 		list_del_init(&bo->list);
434 		mutex_unlock(&bo->rdev->gem.mutex);
435 		/* this should unref the ttm bo */
436 		drm_gem_object_unreference_unlocked(&bo->gem_base);
437 	}
438 }
439 
440 int radeon_bo_init(struct radeon_device *rdev)
441 {
442 	/* Add an MTRR for the VRAM */
443 	if (!rdev->fastfb_working) {
444 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
445 						      rdev->mc.aper_size);
446 	}
447 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
448 		rdev->mc.mc_vram_size >> 20,
449 		(unsigned long long)rdev->mc.aper_size >> 20);
450 	DRM_INFO("RAM width %dbits %cDR\n",
451 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
452 	return radeon_ttm_init(rdev);
453 }
454 
455 void radeon_bo_fini(struct radeon_device *rdev)
456 {
457 	radeon_ttm_fini(rdev);
458 	arch_phys_wc_del(rdev->mc.vram_mtrr);
459 }
460 
461 /* Returns how many bytes TTM can move per IB.
462  */
463 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
464 {
465 	u64 real_vram_size = rdev->mc.real_vram_size;
466 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
467 
468 	/* This function is based on the current VRAM usage.
469 	 *
470 	 * - If all of VRAM is free, allow relocating the number of bytes that
471 	 *   is equal to 1/4 of the size of VRAM for this IB.
472 
473 	 * - If more than one half of VRAM is occupied, only allow relocating
474 	 *   1 MB of data for this IB.
475 	 *
476 	 * - From 0 to one half of used VRAM, the threshold decreases
477 	 *   linearly.
478 	 *         __________________
479 	 * 1/4 of -|\               |
480 	 * VRAM    | \              |
481 	 *         |  \             |
482 	 *         |   \            |
483 	 *         |    \           |
484 	 *         |     \          |
485 	 *         |      \         |
486 	 *         |       \________|1 MB
487 	 *         |----------------|
488 	 *    VRAM 0 %             100 %
489 	 *         used            used
490 	 *
491 	 * Note: It's a threshold, not a limit. The threshold must be crossed
492 	 * for buffer relocations to stop, so any buffer of an arbitrary size
493 	 * can be moved as long as the threshold isn't crossed before
494 	 * the relocation takes place. We don't want to disable buffer
495 	 * relocations completely.
496 	 *
497 	 * The idea is that buffers should be placed in VRAM at creation time
498 	 * and TTM should only do a minimum number of relocations during
499 	 * command submission. In practice, you need to submit at least
500 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
501 	 *
502 	 * Also, things can get pretty crazy under memory pressure and actual
503 	 * VRAM usage can change a lot, so playing safe even at 50% does
504 	 * consistently increase performance.
505 	 */
506 
507 	u64 half_vram = real_vram_size >> 1;
508 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
509 	u64 bytes_moved_threshold = half_free_vram >> 1;
510 	return max(bytes_moved_threshold, 1024*1024ull);
511 }
512 
513 int radeon_bo_list_validate(struct radeon_device *rdev,
514 			    struct ww_acquire_ctx *ticket,
515 			    struct list_head *head, int ring)
516 {
517 	struct radeon_bo_list *lobj;
518 	struct list_head duplicates;
519 	int r;
520 	u64 bytes_moved = 0, initial_bytes_moved;
521 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
522 
523 	INIT_LIST_HEAD(&duplicates);
524 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
525 	if (unlikely(r != 0)) {
526 		return r;
527 	}
528 
529 	list_for_each_entry(lobj, head, tv.head) {
530 		struct radeon_bo *bo = lobj->robj;
531 		if (!bo->pin_count) {
532 			u32 domain = lobj->prefered_domains;
533 			u32 allowed = lobj->allowed_domains;
534 			u32 current_domain =
535 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
536 
537 			/* Check if this buffer will be moved and don't move it
538 			 * if we have moved too many buffers for this IB already.
539 			 *
540 			 * Note that this allows moving at least one buffer of
541 			 * any size, because it doesn't take the current "bo"
542 			 * into account. We don't want to disallow buffer moves
543 			 * completely.
544 			 */
545 			if ((allowed & current_domain) != 0 &&
546 			    (domain & current_domain) == 0 && /* will be moved */
547 			    bytes_moved > bytes_moved_threshold) {
548 				/* don't move it */
549 				domain = current_domain;
550 			}
551 
552 		retry:
553 			radeon_ttm_placement_from_domain(bo, domain);
554 			if (ring == R600_RING_TYPE_UVD_INDEX)
555 				radeon_uvd_force_into_uvd_segment(bo, allowed);
556 
557 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
558 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
559 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
560 				       initial_bytes_moved;
561 
562 			if (unlikely(r)) {
563 				if (r != -ERESTARTSYS &&
564 				    domain != lobj->allowed_domains) {
565 					domain = lobj->allowed_domains;
566 					goto retry;
567 				}
568 				ttm_eu_backoff_reservation(ticket, head);
569 				return r;
570 			}
571 		}
572 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
573 		lobj->tiling_flags = bo->tiling_flags;
574 	}
575 
576 	list_for_each_entry(lobj, &duplicates, tv.head) {
577 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
578 		lobj->tiling_flags = lobj->robj->tiling_flags;
579 	}
580 
581 	return 0;
582 }
583 
584 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
585 {
586 	struct radeon_device *rdev = bo->rdev;
587 	struct radeon_surface_reg *reg;
588 	struct radeon_bo *old_object;
589 	int steal;
590 	int i;
591 
592 	lockdep_assert_held(&bo->tbo.resv->lock.base);
593 
594 	if (!bo->tiling_flags)
595 		return 0;
596 
597 	if (bo->surface_reg >= 0) {
598 		reg = &rdev->surface_regs[bo->surface_reg];
599 		i = bo->surface_reg;
600 		goto out;
601 	}
602 
603 	steal = -1;
604 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
605 
606 		reg = &rdev->surface_regs[i];
607 		if (!reg->bo)
608 			break;
609 
610 		old_object = reg->bo;
611 		if (old_object->pin_count == 0)
612 			steal = i;
613 	}
614 
615 	/* if we are all out */
616 	if (i == RADEON_GEM_MAX_SURFACES) {
617 		if (steal == -1)
618 			return -ENOMEM;
619 		/* find someone with a surface reg and nuke their BO */
620 		reg = &rdev->surface_regs[steal];
621 		old_object = reg->bo;
622 		/* blow away the mapping */
623 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
624 		ttm_bo_unmap_virtual(&old_object->tbo);
625 		old_object->surface_reg = -1;
626 		i = steal;
627 	}
628 
629 	bo->surface_reg = i;
630 	reg->bo = bo;
631 
632 out:
633 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
634 			       bo->tbo.mem.start << PAGE_SHIFT,
635 			       bo->tbo.num_pages << PAGE_SHIFT);
636 	return 0;
637 }
638 
639 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
640 {
641 	struct radeon_device *rdev = bo->rdev;
642 	struct radeon_surface_reg *reg;
643 
644 	if (bo->surface_reg == -1)
645 		return;
646 
647 	reg = &rdev->surface_regs[bo->surface_reg];
648 	radeon_clear_surface_reg(rdev, bo->surface_reg);
649 
650 	reg->bo = NULL;
651 	bo->surface_reg = -1;
652 }
653 
654 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
655 				uint32_t tiling_flags, uint32_t pitch)
656 {
657 	struct radeon_device *rdev = bo->rdev;
658 	int r;
659 
660 	if (rdev->family >= CHIP_CEDAR) {
661 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
662 
663 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
664 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
665 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
666 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
667 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
668 		switch (bankw) {
669 		case 0:
670 		case 1:
671 		case 2:
672 		case 4:
673 		case 8:
674 			break;
675 		default:
676 			return -EINVAL;
677 		}
678 		switch (bankh) {
679 		case 0:
680 		case 1:
681 		case 2:
682 		case 4:
683 		case 8:
684 			break;
685 		default:
686 			return -EINVAL;
687 		}
688 		switch (mtaspect) {
689 		case 0:
690 		case 1:
691 		case 2:
692 		case 4:
693 		case 8:
694 			break;
695 		default:
696 			return -EINVAL;
697 		}
698 		if (tilesplit > 6) {
699 			return -EINVAL;
700 		}
701 		if (stilesplit > 6) {
702 			return -EINVAL;
703 		}
704 	}
705 	r = radeon_bo_reserve(bo, false);
706 	if (unlikely(r != 0))
707 		return r;
708 	bo->tiling_flags = tiling_flags;
709 	bo->pitch = pitch;
710 	radeon_bo_unreserve(bo);
711 	return 0;
712 }
713 
714 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
715 				uint32_t *tiling_flags,
716 				uint32_t *pitch)
717 {
718 	lockdep_assert_held(&bo->tbo.resv->lock.base);
719 
720 	if (tiling_flags)
721 		*tiling_flags = bo->tiling_flags;
722 	if (pitch)
723 		*pitch = bo->pitch;
724 }
725 
726 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
727 				bool force_drop)
728 {
729 	if (!force_drop)
730 		lockdep_assert_held(&bo->tbo.resv->lock.base);
731 
732 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
733 		return 0;
734 
735 	if (force_drop) {
736 		radeon_bo_clear_surface_reg(bo);
737 		return 0;
738 	}
739 
740 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
741 		if (!has_moved)
742 			return 0;
743 
744 		if (bo->surface_reg >= 0)
745 			radeon_bo_clear_surface_reg(bo);
746 		return 0;
747 	}
748 
749 	if ((bo->surface_reg >= 0) && !has_moved)
750 		return 0;
751 
752 	return radeon_bo_get_surface_reg(bo);
753 }
754 
755 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
756 			   struct ttm_mem_reg *new_mem)
757 {
758 	struct radeon_bo *rbo;
759 
760 	if (!radeon_ttm_bo_is_radeon_bo(bo))
761 		return;
762 
763 	rbo = container_of(bo, struct radeon_bo, tbo);
764 	radeon_bo_check_tiling(rbo, 0, 1);
765 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
766 
767 	/* update statistics */
768 	if (!new_mem)
769 		return;
770 
771 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
772 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
773 }
774 
775 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
776 {
777 	struct radeon_device *rdev;
778 	struct radeon_bo *rbo;
779 	unsigned long offset, size, lpfn;
780 	int i, r;
781 
782 	if (!radeon_ttm_bo_is_radeon_bo(bo))
783 		return 0;
784 	rbo = container_of(bo, struct radeon_bo, tbo);
785 	radeon_bo_check_tiling(rbo, 0, 0);
786 	rdev = rbo->rdev;
787 	if (bo->mem.mem_type != TTM_PL_VRAM)
788 		return 0;
789 
790 	size = bo->mem.num_pages << PAGE_SHIFT;
791 	offset = bo->mem.start << PAGE_SHIFT;
792 	if ((offset + size) <= rdev->mc.visible_vram_size)
793 		return 0;
794 
795 	/* hurrah the memory is not visible ! */
796 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
797 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
798 	for (i = 0; i < rbo->placement.num_placement; i++) {
799 		/* Force into visible VRAM */
800 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
801 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
802 			rbo->placements[i].lpfn = lpfn;
803 	}
804 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
805 	if (unlikely(r == -ENOMEM)) {
806 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
807 		return ttm_bo_validate(bo, &rbo->placement, false, false);
808 	} else if (unlikely(r != 0)) {
809 		return r;
810 	}
811 
812 	offset = bo->mem.start << PAGE_SHIFT;
813 	/* this should never happen */
814 	if ((offset + size) > rdev->mc.visible_vram_size)
815 		return -EINVAL;
816 
817 	return 0;
818 }
819 
820 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
821 {
822 	int r;
823 
824 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
825 	if (unlikely(r != 0))
826 		return r;
827 	if (mem_type)
828 		*mem_type = bo->tbo.mem.mem_type;
829 
830 	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
831 	ttm_bo_unreserve(&bo->tbo);
832 	return r;
833 }
834 
835 /**
836  * radeon_bo_fence - add fence to buffer object
837  *
838  * @bo: buffer object in question
839  * @fence: fence to add
840  * @shared: true if fence should be added shared
841  *
842  */
843 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
844                      bool shared)
845 {
846 	struct reservation_object *resv = bo->tbo.resv;
847 
848 	if (shared)
849 		reservation_object_add_shared_fence(resv, &fence->base);
850 	else
851 		reservation_object_add_excl_fence(resv, &fence->base);
852 }
853