xref: /dflybsd-src/sys/dev/drm/radeon/radeon_object.c (revision e17b1aed5714b34e3d31afcfde1a2f0d2b3f9b18)
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 <drm/drmP.h>
33 #include <drm/radeon_drm.h>
34 #include "radeon.h"
35 #ifdef TRACE_TODO
36 #include "radeon_trace.h"
37 #endif
38 #include <linux/io.h>
39 
40 
41 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
42 
43 /*
44  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
45  * function are calling it.
46  */
47 
48 static void radeon_update_memory_usage(struct radeon_bo *bo,
49 				       unsigned mem_type, int sign)
50 {
51 	struct radeon_device *rdev = bo->rdev;
52 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
53 
54 	switch (mem_type) {
55 	case TTM_PL_TT:
56 		if (sign > 0)
57 			atomic64_add(size, &rdev->gtt_usage);
58 		else
59 			atomic64_sub(size, &rdev->gtt_usage);
60 		break;
61 	case TTM_PL_VRAM:
62 		if (sign > 0)
63 			atomic64_add(size, &rdev->vram_usage);
64 		else
65 			atomic64_sub(size, &rdev->vram_usage);
66 		break;
67 	}
68 }
69 
70 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
71 {
72 	struct radeon_bo *bo;
73 
74 	bo = container_of(tbo, struct radeon_bo, tbo);
75 
76 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
77 	radeon_mn_unregister(bo);
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 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
103 					     TTM_PL_FLAG_UNCACHED |
104 					     TTM_PL_FLAG_VRAM;
105 
106 	if (domain & RADEON_GEM_DOMAIN_GTT) {
107 		if (rbo->flags & RADEON_GEM_GTT_UC) {
108 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
109 				TTM_PL_FLAG_TT;
110 
111 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
112 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
113 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
114 				TTM_PL_FLAG_UNCACHED |
115 				TTM_PL_FLAG_TT;
116 		} else {
117 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
118 						     TTM_PL_FLAG_TT;
119 		}
120 	}
121 
122 	if (domain & RADEON_GEM_DOMAIN_CPU) {
123 		if (rbo->flags & RADEON_GEM_GTT_UC) {
124 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
125 				TTM_PL_FLAG_SYSTEM;
126 
127 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
128 		    rbo->rdev->flags & RADEON_IS_AGP) {
129 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
130 				TTM_PL_FLAG_UNCACHED |
131 				TTM_PL_FLAG_SYSTEM;
132 		} else {
133 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
134 						     TTM_PL_FLAG_SYSTEM;
135 		}
136 	}
137 	if (!c)
138 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
139 					     TTM_PL_FLAG_SYSTEM;
140 
141 	rbo->placement.num_placement = c;
142 	rbo->placement.num_busy_placement = c;
143 
144 	for (i = 0; i < c; ++i) {
145 		rbo->placements[i].fpfn = 0;
146 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
147 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM))
148 			rbo->placements[i].lpfn =
149 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
150 		else
151 			rbo->placements[i].lpfn = 0;
152 	}
153 
154 	/*
155 	 * Use two-ended allocation depending on the buffer size to
156 	 * improve fragmentation quality.
157 	 * 512kb was measured as the most optimal number.
158 	 */
159 	if (!((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
160 	      (rbo->placements[i].flags & TTM_PL_FLAG_VRAM)) &&
161 	    rbo->tbo.mem.size > 512 * 1024) {
162 		for (i = 0; i < c; i++) {
163 			rbo->placements[i].flags |= TTM_PL_FLAG_TOPDOWN;
164 		}
165 	}
166 }
167 
168 int radeon_bo_create(struct radeon_device *rdev,
169 		     unsigned long size, int byte_align, bool kernel, u32 domain,
170 		     u32 flags, struct sg_table *sg, struct radeon_bo **bo_ptr)
171 {
172 	struct radeon_bo *bo;
173 	enum ttm_bo_type type;
174 	unsigned long page_align = roundup2(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
175 	size_t acc_size;
176 	int r;
177 
178 	size = ALIGN(size, PAGE_SIZE);
179 
180 	if (kernel) {
181 		type = ttm_bo_type_kernel;
182 	} else if (sg) {
183 		type = ttm_bo_type_sg;
184 	} else {
185 		type = ttm_bo_type_device;
186 	}
187 	*bo_ptr = NULL;
188 
189 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
190 				       sizeof(struct radeon_bo));
191 
192 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
193 	if (bo == NULL)
194 		return -ENOMEM;
195 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
196 	if (unlikely(r)) {
197 		kfree(bo);
198 		return r;
199 	}
200 	bo->rdev = rdev;
201 	bo->surface_reg = -1;
202 	INIT_LIST_HEAD(&bo->list);
203 	INIT_LIST_HEAD(&bo->va);
204 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
205 	                               RADEON_GEM_DOMAIN_GTT |
206 	                               RADEON_GEM_DOMAIN_CPU);
207 
208 	bo->flags = flags;
209 	/* PCI GART is always snooped */
210 	if (!(rdev->flags & RADEON_IS_PCIE))
211 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
212 
213 #ifdef CONFIG_X86_32
214 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
215 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
216 	 */
217 	bo->flags &= ~RADEON_GEM_GTT_WC;
218 #endif
219 
220 	radeon_ttm_placement_from_domain(bo, domain);
221 	/* Kernel allocation are uninterruptible */
222 	lockmgr(&rdev->pm.mclk_lock, LK_SHARED);
223 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
224 			&bo->placement, page_align, !kernel, NULL,
225 			acc_size, sg, NULL, &radeon_ttm_bo_destroy);
226 	up_read(&rdev->pm.mclk_lock);
227 	if (unlikely(r != 0)) {
228 		return r;
229 	}
230 	*bo_ptr = bo;
231 
232 #ifdef TRACE_TODO
233 	trace_radeon_bo_create(bo);
234 #endif
235 
236 	return 0;
237 }
238 
239 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
240 {
241 	bool is_iomem;
242 	int r;
243 
244 	if (bo->kptr) {
245 		if (ptr) {
246 			*ptr = bo->kptr;
247 		}
248 		return 0;
249 	}
250 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
251 	if (r) {
252 		return r;
253 	}
254 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
255 	if (ptr) {
256 		*ptr = bo->kptr;
257 	}
258 	radeon_bo_check_tiling(bo, 0, 0);
259 	return 0;
260 }
261 
262 void radeon_bo_kunmap(struct radeon_bo *bo)
263 {
264 	if (bo->kptr == NULL)
265 		return;
266 	bo->kptr = NULL;
267 	radeon_bo_check_tiling(bo, 0, 0);
268 	ttm_bo_kunmap(&bo->kmap);
269 }
270 
271 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
272 {
273 	if (bo == NULL)
274 		return NULL;
275 
276 	ttm_bo_reference(&bo->tbo);
277 	return bo;
278 }
279 
280 void radeon_bo_unref(struct radeon_bo **bo)
281 {
282 	struct ttm_buffer_object *tbo;
283 	struct radeon_device *rdev;
284 	struct radeon_bo *rbo;
285 
286 	if ((rbo = *bo) == NULL)
287 		return;
288 	*bo = NULL;
289 	rdev = rbo->rdev;
290 	tbo = &rbo->tbo;
291 	ttm_bo_unref(&tbo);
292 }
293 
294 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
295 			     u64 *gpu_addr)
296 {
297 	int r, i;
298 
299 	if (bo->pin_count) {
300 		bo->pin_count++;
301 		if (gpu_addr)
302 			*gpu_addr = radeon_bo_gpu_offset(bo);
303 
304 		if (max_offset != 0) {
305 			u64 domain_start;
306 
307 			if (domain == RADEON_GEM_DOMAIN_VRAM)
308 				domain_start = bo->rdev->mc.vram_start;
309 			else
310 				domain_start = bo->rdev->mc.gtt_start;
311 			if (max_offset < (radeon_bo_gpu_offset(bo) - domain_start)) {
312 				DRM_ERROR("radeon_bo_pin_restricted: "
313 				    "max_offset(%ju) < "
314 				    "(radeon_bo_gpu_offset(%ju) - "
315 				    "domain_start(%ju)",
316 				    (uintmax_t)max_offset, (uintmax_t)radeon_bo_gpu_offset(bo),
317 				    (uintmax_t)domain_start);
318 			}
319 		}
320 
321 		return 0;
322 	}
323 	radeon_ttm_placement_from_domain(bo, domain);
324 	for (i = 0; i < bo->placement.num_placement; i++) {
325 		/* force to pin into visible video ram */
326 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
327 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
328 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
329 			bo->placements[i].lpfn =
330 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
331 		else
332 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
333 
334 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
335 	}
336 
337 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
338 	if (likely(r == 0)) {
339 		bo->pin_count = 1;
340 		if (gpu_addr != NULL)
341 			*gpu_addr = radeon_bo_gpu_offset(bo);
342 		if (domain == RADEON_GEM_DOMAIN_VRAM)
343 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
344 		else
345 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
346 	} else {
347 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
348 	}
349 	return r;
350 }
351 
352 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
353 {
354 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
355 }
356 
357 int radeon_bo_unpin(struct radeon_bo *bo)
358 {
359 	int r, i;
360 
361 	if (!bo->pin_count) {
362 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
363 		return 0;
364 	}
365 	bo->pin_count--;
366 	if (bo->pin_count)
367 		return 0;
368 	for (i = 0; i < bo->placement.num_placement; i++) {
369 		bo->placements[i].lpfn = 0;
370 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
371 	}
372 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
373 	if (likely(r == 0)) {
374 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
375 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
376 		else
377 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
378 	} else {
379 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
380 	}
381 	return r;
382 }
383 
384 int radeon_bo_evict_vram(struct radeon_device *rdev)
385 {
386 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
387 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
388 		if (rdev->mc.igp_sideport_enabled == false)
389 			/* Useless to evict on IGP chips */
390 			return 0;
391 	}
392 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
393 }
394 
395 void radeon_bo_force_delete(struct radeon_device *rdev)
396 {
397 	struct radeon_bo *bo, *n;
398 
399 	if (list_empty(&rdev->gem.objects)) {
400 		return;
401 	}
402 	dev_err(rdev->dev, "Userspace still has active objects !\n");
403 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
404 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
405 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
406 			*((unsigned long *)&bo->gem_base.refcount));
407 		mutex_lock(&bo->rdev->gem.mutex);
408 		list_del_init(&bo->list);
409 		mutex_unlock(&bo->rdev->gem.mutex);
410 		/* this should unref the ttm bo */
411 		drm_gem_object_unreference(&bo->gem_base);
412 	}
413 }
414 
415 int radeon_bo_init(struct radeon_device *rdev)
416 {
417 	/* Add an MTRR for the VRAM */
418 	if (!rdev->fastfb_working) {
419 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
420 						      rdev->mc.aper_size);
421 	}
422 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
423 		rdev->mc.mc_vram_size >> 20,
424 		(unsigned long long)rdev->mc.aper_size >> 20);
425 	DRM_INFO("RAM width %dbits %cDR\n",
426 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
427 	return radeon_ttm_init(rdev);
428 }
429 
430 void radeon_bo_fini(struct radeon_device *rdev)
431 {
432 	radeon_ttm_fini(rdev);
433 	arch_phys_wc_del(rdev->mc.vram_mtrr);
434 }
435 
436 /* Returns how many bytes TTM can move per IB.
437  */
438 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
439 {
440 	u64 real_vram_size = rdev->mc.real_vram_size;
441 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
442 
443 	/* This function is based on the current VRAM usage.
444 	 *
445 	 * - If all of VRAM is free, allow relocating the number of bytes that
446 	 *   is equal to 1/4 of the size of VRAM for this IB.
447 
448 	 * - If more than one half of VRAM is occupied, only allow relocating
449 	 *   1 MB of data for this IB.
450 	 *
451 	 * - From 0 to one half of used VRAM, the threshold decreases
452 	 *   linearly.
453 	 *         __________________
454 	 * 1/4 of -|\               |
455 	 * VRAM    | \              |
456 	 *         |  \             |
457 	 *         |   \            |
458 	 *         |    \           |
459 	 *         |     \          |
460 	 *         |      \         |
461 	 *         |       \________|1 MB
462 	 *         |----------------|
463 	 *    VRAM 0 %             100 %
464 	 *         used            used
465 	 *
466 	 * Note: It's a threshold, not a limit. The threshold must be crossed
467 	 * for buffer relocations to stop, so any buffer of an arbitrary size
468 	 * can be moved as long as the threshold isn't crossed before
469 	 * the relocation takes place. We don't want to disable buffer
470 	 * relocations completely.
471 	 *
472 	 * The idea is that buffers should be placed in VRAM at creation time
473 	 * and TTM should only do a minimum number of relocations during
474 	 * command submission. In practice, you need to submit at least
475 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
476 	 *
477 	 * Also, things can get pretty crazy under memory pressure and actual
478 	 * VRAM usage can change a lot, so playing safe even at 50% does
479 	 * consistently increase performance.
480 	 */
481 
482 	u64 half_vram = real_vram_size >> 1;
483 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
484 	u64 bytes_moved_threshold = half_free_vram >> 1;
485 	return max(bytes_moved_threshold, 1024*1024ull);
486 }
487 
488 int radeon_bo_list_validate(struct radeon_device *rdev,
489 			    struct ww_acquire_ctx *ticket,
490 			    struct list_head *head, int ring)
491 {
492 	struct radeon_cs_reloc *lobj;
493 	struct radeon_bo *bo;
494 	int r;
495 	u64 bytes_moved = 0, initial_bytes_moved;
496 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
497 
498 	r = ttm_eu_reserve_buffers(ticket, head, true);
499 	if (unlikely(r != 0)) {
500 		return r;
501 	}
502 
503 	list_for_each_entry(lobj, head, tv.head) {
504 		bo = lobj->robj;
505 		if (!bo->pin_count) {
506 			u32 domain = lobj->prefered_domains;
507 			u32 allowed = lobj->allowed_domains;
508 			u32 current_domain =
509 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
510 
511 			/* Check if this buffer will be moved and don't move it
512 			 * if we have moved too many buffers for this IB already.
513 			 *
514 			 * Note that this allows moving at least one buffer of
515 			 * any size, because it doesn't take the current "bo"
516 			 * into account. We don't want to disallow buffer moves
517 			 * completely.
518 			 */
519 			if ((allowed & current_domain) != 0 &&
520 			    (domain & current_domain) == 0 && /* will be moved */
521 			    bytes_moved > bytes_moved_threshold) {
522 				/* don't move it */
523 				domain = current_domain;
524 			}
525 
526 		retry:
527 			radeon_ttm_placement_from_domain(bo, domain);
528 			if (ring == R600_RING_TYPE_UVD_INDEX)
529 				radeon_uvd_force_into_uvd_segment(bo, allowed);
530 
531 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
532 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
533 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
534 				       initial_bytes_moved;
535 
536 			if (unlikely(r)) {
537 				if (r != -ERESTARTSYS &&
538 				    domain != lobj->allowed_domains) {
539 					domain = lobj->allowed_domains;
540 					goto retry;
541 				}
542 				ttm_eu_backoff_reservation(ticket, head);
543 				return r;
544 			}
545 		}
546 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
547 		lobj->tiling_flags = bo->tiling_flags;
548 	}
549 	return 0;
550 }
551 
552 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
553 {
554 	struct radeon_device *rdev = bo->rdev;
555 	struct radeon_surface_reg *reg;
556 	struct radeon_bo *old_object;
557 	int steal;
558 	int i;
559 
560 	if (!bo->tiling_flags)
561 		return 0;
562 
563 	if (bo->surface_reg >= 0) {
564 		reg = &rdev->surface_regs[bo->surface_reg];
565 		i = bo->surface_reg;
566 		goto out;
567 	}
568 
569 	steal = -1;
570 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
571 
572 		reg = &rdev->surface_regs[i];
573 		if (!reg->bo)
574 			break;
575 
576 		old_object = reg->bo;
577 		if (old_object->pin_count == 0)
578 			steal = i;
579 	}
580 
581 	/* if we are all out */
582 	if (i == RADEON_GEM_MAX_SURFACES) {
583 		if (steal == -1)
584 			return -ENOMEM;
585 		/* find someone with a surface reg and nuke their BO */
586 		reg = &rdev->surface_regs[steal];
587 		old_object = reg->bo;
588 		/* blow away the mapping */
589 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
590 		ttm_bo_unmap_virtual(&old_object->tbo);
591 		old_object->surface_reg = -1;
592 		i = steal;
593 	}
594 
595 	bo->surface_reg = i;
596 	reg->bo = bo;
597 
598 out:
599 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
600 			       bo->tbo.mem.start << PAGE_SHIFT,
601 			       bo->tbo.num_pages << PAGE_SHIFT);
602 	return 0;
603 }
604 
605 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
606 {
607 	struct radeon_device *rdev = bo->rdev;
608 	struct radeon_surface_reg *reg;
609 
610 	if (bo->surface_reg == -1)
611 		return;
612 
613 	reg = &rdev->surface_regs[bo->surface_reg];
614 	radeon_clear_surface_reg(rdev, bo->surface_reg);
615 
616 	reg->bo = NULL;
617 	bo->surface_reg = -1;
618 }
619 
620 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
621 				uint32_t tiling_flags, uint32_t pitch)
622 {
623 	struct radeon_device *rdev = bo->rdev;
624 	int r;
625 
626 	if (rdev->family >= CHIP_CEDAR) {
627 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
628 
629 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
630 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
631 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
632 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
633 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
634 		switch (bankw) {
635 		case 0:
636 		case 1:
637 		case 2:
638 		case 4:
639 		case 8:
640 			break;
641 		default:
642 			return -EINVAL;
643 		}
644 		switch (bankh) {
645 		case 0:
646 		case 1:
647 		case 2:
648 		case 4:
649 		case 8:
650 			break;
651 		default:
652 			return -EINVAL;
653 		}
654 		switch (mtaspect) {
655 		case 0:
656 		case 1:
657 		case 2:
658 		case 4:
659 		case 8:
660 			break;
661 		default:
662 			return -EINVAL;
663 		}
664 		if (tilesplit > 6) {
665 			return -EINVAL;
666 		}
667 		if (stilesplit > 6) {
668 			return -EINVAL;
669 		}
670 	}
671 	r = radeon_bo_reserve(bo, false);
672 	if (unlikely(r != 0))
673 		return r;
674 	bo->tiling_flags = tiling_flags;
675 	bo->pitch = pitch;
676 	radeon_bo_unreserve(bo);
677 	return 0;
678 }
679 
680 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
681 				uint32_t *tiling_flags,
682 				uint32_t *pitch)
683 {
684 	if (tiling_flags)
685 		*tiling_flags = bo->tiling_flags;
686 	if (pitch)
687 		*pitch = bo->pitch;
688 }
689 
690 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
691 				bool force_drop)
692 {
693 
694 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
695 		return 0;
696 
697 	if (force_drop) {
698 		radeon_bo_clear_surface_reg(bo);
699 		return 0;
700 	}
701 
702 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
703 		if (!has_moved)
704 			return 0;
705 
706 		if (bo->surface_reg >= 0)
707 			radeon_bo_clear_surface_reg(bo);
708 		return 0;
709 	}
710 
711 	if ((bo->surface_reg >= 0) && !has_moved)
712 		return 0;
713 
714 	return radeon_bo_get_surface_reg(bo);
715 }
716 
717 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
718 			   struct ttm_mem_reg *new_mem)
719 {
720 	struct radeon_bo *rbo;
721 
722 	if (!radeon_ttm_bo_is_radeon_bo(bo))
723 		return;
724 
725 	rbo = container_of(bo, struct radeon_bo, tbo);
726 	radeon_bo_check_tiling(rbo, 0, 1);
727 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
728 
729 	/* update statistics */
730 	if (!new_mem)
731 		return;
732 
733 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
734 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
735 }
736 
737 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
738 {
739 	struct radeon_device *rdev;
740 	struct radeon_bo *rbo;
741 	unsigned long offset, size;
742 	int r;
743 
744 	if (!radeon_ttm_bo_is_radeon_bo(bo))
745 		return 0;
746 	rbo = container_of(bo, struct radeon_bo, tbo);
747 	radeon_bo_check_tiling(rbo, 0, 0);
748 	rdev = rbo->rdev;
749 	if (bo->mem.mem_type != TTM_PL_VRAM)
750 		return 0;
751 
752 	size = bo->mem.num_pages << PAGE_SHIFT;
753 	offset = bo->mem.start << PAGE_SHIFT;
754 	if ((offset + size) <= rdev->mc.visible_vram_size)
755 		return 0;
756 
757 	/* hurrah the memory is not visible ! */
758 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
759 	rbo->placements[0].lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
760 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
761 	if (unlikely(r == -ENOMEM)) {
762 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
763 		return ttm_bo_validate(bo, &rbo->placement, false, false);
764 	} else if (unlikely(r != 0)) {
765 		return r;
766 	}
767 
768 	offset = bo->mem.start << PAGE_SHIFT;
769 	/* this should never happen */
770 	if ((offset + size) > rdev->mc.visible_vram_size)
771 		return -EINVAL;
772 
773 	return 0;
774 }
775 
776 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
777 {
778 	int r;
779 
780 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
781 	if (unlikely(r != 0))
782 		return r;
783 	if (mem_type)
784 		*mem_type = bo->tbo.mem.mem_type;
785 
786 	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
787 	ttm_bo_unreserve(&bo->tbo);
788 	return r;
789 }
790 
791 
792 /**
793  * radeon_bo_reserve - reserve bo
794  * @bo:		bo structure
795  * @no_intr:	don't return -ERESTARTSYS on pending signal
796  *
797  * Returns:
798  * -ERESTARTSYS: A wait for the buffer to become unreserved was interrupted by
799  * a signal. Release all buffer reservations and return to user-space.
800  */
801 int radeon_bo_reserve(struct radeon_bo *bo, bool no_intr)
802 {
803 	int r;
804 
805 	r = ttm_bo_reserve(&bo->tbo, !no_intr, false, false, 0);
806 	if (unlikely(r != 0)) {
807 		if (r != -ERESTARTSYS)
808 			dev_err(bo->rdev->dev, "%p reserve failed\n", bo);
809 		return r;
810 	}
811 	return 0;
812 }
813