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 
33 #include <linux/io.h>
34 #include <linux/list.h>
35 #include <linux/slab.h>
36 
37 #include <drm/drm_cache.h>
38 #include <drm/drm_prime.h>
39 #include <drm/radeon_drm.h>
40 
41 #include "radeon.h"
42 #include "radeon_trace.h"
43 
44 int radeon_ttm_init(struct radeon_device *rdev);
45 void radeon_ttm_fini(struct radeon_device *rdev);
46 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
47 
48 /*
49  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
50  * function are calling it.
51  */
52 
53 static void radeon_update_memory_usage(struct radeon_bo *bo,
54 				       unsigned mem_type, int sign)
55 {
56 	struct radeon_device *rdev = bo->rdev;
57 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
58 
59 	switch (mem_type) {
60 	case TTM_PL_TT:
61 		if (sign > 0)
62 			atomic64_add(size, &rdev->gtt_usage);
63 		else
64 			atomic64_sub(size, &rdev->gtt_usage);
65 		break;
66 	case TTM_PL_VRAM:
67 		if (sign > 0)
68 			atomic64_add(size, &rdev->vram_usage);
69 		else
70 			atomic64_sub(size, &rdev->vram_usage);
71 		break;
72 	}
73 }
74 
75 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
76 {
77 	struct radeon_bo *bo;
78 
79 	bo = container_of(tbo, struct radeon_bo, tbo);
80 
81 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
82 
83 	mutex_lock(&bo->rdev->gem.mutex);
84 	list_del_init(&bo->list);
85 	mutex_unlock(&bo->rdev->gem.mutex);
86 	radeon_bo_clear_surface_reg(bo);
87 	WARN_ON_ONCE(!list_empty(&bo->va));
88 	if (bo->tbo.base.import_attach)
89 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
90 	drm_gem_object_release(&bo->tbo.base);
91 	kfree(bo);
92 }
93 
94 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
95 {
96 	if (bo->destroy == &radeon_ttm_bo_destroy)
97 		return true;
98 	return false;
99 }
100 
101 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
102 {
103 	u32 c = 0, i;
104 
105 	rbo->placement.placement = rbo->placements;
106 	rbo->placement.busy_placement = rbo->placements;
107 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
108 		/* Try placing BOs which don't need CPU access outside of the
109 		 * CPU accessible part of VRAM
110 		 */
111 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
112 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
113 			rbo->placements[c].fpfn =
114 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
115 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
116 						     TTM_PL_FLAG_UNCACHED |
117 						     TTM_PL_FLAG_VRAM;
118 		}
119 
120 		rbo->placements[c].fpfn = 0;
121 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
122 					     TTM_PL_FLAG_UNCACHED |
123 					     TTM_PL_FLAG_VRAM;
124 	}
125 
126 	if (domain & RADEON_GEM_DOMAIN_GTT) {
127 		if (rbo->flags & RADEON_GEM_GTT_UC) {
128 			rbo->placements[c].fpfn = 0;
129 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
130 				TTM_PL_FLAG_TT;
131 
132 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
133 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
134 			rbo->placements[c].fpfn = 0;
135 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
136 				TTM_PL_FLAG_UNCACHED |
137 				TTM_PL_FLAG_TT;
138 		} else {
139 			rbo->placements[c].fpfn = 0;
140 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
141 						     TTM_PL_FLAG_TT;
142 		}
143 	}
144 
145 	if (domain & RADEON_GEM_DOMAIN_CPU) {
146 		if (rbo->flags & RADEON_GEM_GTT_UC) {
147 			rbo->placements[c].fpfn = 0;
148 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
149 				TTM_PL_FLAG_SYSTEM;
150 
151 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
152 		    rbo->rdev->flags & RADEON_IS_AGP) {
153 			rbo->placements[c].fpfn = 0;
154 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
155 				TTM_PL_FLAG_UNCACHED |
156 				TTM_PL_FLAG_SYSTEM;
157 		} else {
158 			rbo->placements[c].fpfn = 0;
159 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
160 						     TTM_PL_FLAG_SYSTEM;
161 		}
162 	}
163 	if (!c) {
164 		rbo->placements[c].fpfn = 0;
165 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
166 					     TTM_PL_FLAG_SYSTEM;
167 	}
168 
169 	rbo->placement.num_placement = c;
170 	rbo->placement.num_busy_placement = c;
171 
172 	for (i = 0; i < c; ++i) {
173 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
174 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
175 		    !rbo->placements[i].fpfn)
176 			rbo->placements[i].lpfn =
177 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
178 		else
179 			rbo->placements[i].lpfn = 0;
180 	}
181 }
182 
183 int radeon_bo_create(struct radeon_device *rdev,
184 		     unsigned long size, int byte_align, bool kernel,
185 		     u32 domain, u32 flags, struct sg_table *sg,
186 		     struct dma_resv *resv,
187 		     struct radeon_bo **bo_ptr)
188 {
189 	struct radeon_bo *bo;
190 	enum ttm_bo_type type;
191 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
192 	size_t acc_size;
193 	int r;
194 
195 	size = ALIGN(size, PAGE_SIZE);
196 
197 	if (kernel) {
198 		type = ttm_bo_type_kernel;
199 	} else if (sg) {
200 		type = ttm_bo_type_sg;
201 	} else {
202 		type = ttm_bo_type_device;
203 	}
204 	*bo_ptr = NULL;
205 
206 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
207 				       sizeof(struct radeon_bo));
208 
209 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
210 	if (bo == NULL)
211 		return -ENOMEM;
212 	drm_gem_private_object_init(rdev->ddev, &bo->tbo.base, size);
213 	bo->rdev = rdev;
214 	bo->surface_reg = -1;
215 	INIT_LIST_HEAD(&bo->list);
216 	INIT_LIST_HEAD(&bo->va);
217 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
218 				       RADEON_GEM_DOMAIN_GTT |
219 				       RADEON_GEM_DOMAIN_CPU);
220 
221 	bo->flags = flags;
222 	/* PCI GART is always snooped */
223 	if (!(rdev->flags & RADEON_IS_PCIE))
224 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
225 
226 	/* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx
227 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=91268
228 	 */
229 	if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635)
230 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
231 
232 #ifdef CONFIG_X86_32
233 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
234 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
235 	 */
236 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
237 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
238 	/* Don't try to enable write-combining when it can't work, or things
239 	 * may be slow
240 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
241 	 */
242 #ifndef CONFIG_COMPILE_TEST
243 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
244 	 thanks to write-combining
245 #endif
246 
247 	if (bo->flags & RADEON_GEM_GTT_WC)
248 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
249 			      "better performance thanks to write-combining\n");
250 	bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
251 #else
252 	/* For architectures that don't support WC memory,
253 	 * mask out the WC flag from the BO
254 	 */
255 	if (!drm_arch_can_wc_memory())
256 		bo->flags &= ~RADEON_GEM_GTT_WC;
257 #endif
258 
259 	radeon_ttm_placement_from_domain(bo, domain);
260 	/* Kernel allocation are uninterruptible */
261 	down_read(&rdev->pm.mclk_lock);
262 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
263 			&bo->placement, page_align, !kernel, acc_size,
264 			sg, resv, &radeon_ttm_bo_destroy);
265 	up_read(&rdev->pm.mclk_lock);
266 	if (unlikely(r != 0)) {
267 		return r;
268 	}
269 	*bo_ptr = bo;
270 
271 	trace_radeon_bo_create(bo);
272 
273 	return 0;
274 }
275 
276 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
277 {
278 	bool is_iomem;
279 	int r;
280 
281 	if (bo->kptr) {
282 		if (ptr) {
283 			*ptr = bo->kptr;
284 		}
285 		return 0;
286 	}
287 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
288 	if (r) {
289 		return r;
290 	}
291 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
292 	if (ptr) {
293 		*ptr = bo->kptr;
294 	}
295 	radeon_bo_check_tiling(bo, 0, 0);
296 	return 0;
297 }
298 
299 void radeon_bo_kunmap(struct radeon_bo *bo)
300 {
301 	if (bo->kptr == NULL)
302 		return;
303 	bo->kptr = NULL;
304 	radeon_bo_check_tiling(bo, 0, 0);
305 	ttm_bo_kunmap(&bo->kmap);
306 }
307 
308 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
309 {
310 	if (bo == NULL)
311 		return NULL;
312 
313 	ttm_bo_get(&bo->tbo);
314 	return bo;
315 }
316 
317 void radeon_bo_unref(struct radeon_bo **bo)
318 {
319 	struct ttm_buffer_object *tbo;
320 
321 	if ((*bo) == NULL)
322 		return;
323 	tbo = &((*bo)->tbo);
324 	ttm_bo_put(tbo);
325 	*bo = NULL;
326 }
327 
328 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
329 			     u64 *gpu_addr)
330 {
331 	struct ttm_operation_ctx ctx = { false, false };
332 	int r, i;
333 
334 	if (radeon_ttm_tt_has_userptr(bo->tbo.ttm))
335 		return -EPERM;
336 
337 	if (bo->pin_count) {
338 		bo->pin_count++;
339 		if (gpu_addr)
340 			*gpu_addr = radeon_bo_gpu_offset(bo);
341 
342 		if (max_offset != 0) {
343 			u64 domain_start;
344 
345 			if (domain == RADEON_GEM_DOMAIN_VRAM)
346 				domain_start = bo->rdev->mc.vram_start;
347 			else
348 				domain_start = bo->rdev->mc.gtt_start;
349 			WARN_ON_ONCE(max_offset <
350 				     (radeon_bo_gpu_offset(bo) - domain_start));
351 		}
352 
353 		return 0;
354 	}
355 	if (bo->prime_shared_count && domain == RADEON_GEM_DOMAIN_VRAM) {
356 		/* A BO shared as a dma-buf cannot be sensibly migrated to VRAM */
357 		return -EINVAL;
358 	}
359 
360 	radeon_ttm_placement_from_domain(bo, domain);
361 	for (i = 0; i < bo->placement.num_placement; i++) {
362 		/* force to pin into visible video ram */
363 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
364 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
365 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
366 			bo->placements[i].lpfn =
367 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
368 		else
369 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
370 
371 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
372 	}
373 
374 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
375 	if (likely(r == 0)) {
376 		bo->pin_count = 1;
377 		if (gpu_addr != NULL)
378 			*gpu_addr = radeon_bo_gpu_offset(bo);
379 		if (domain == RADEON_GEM_DOMAIN_VRAM)
380 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
381 		else
382 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
383 	} else {
384 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
385 	}
386 	return r;
387 }
388 
389 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
390 {
391 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
392 }
393 
394 int radeon_bo_unpin(struct radeon_bo *bo)
395 {
396 	struct ttm_operation_ctx ctx = { false, false };
397 	int r, i;
398 
399 	if (!bo->pin_count) {
400 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
401 		return 0;
402 	}
403 	bo->pin_count--;
404 	if (bo->pin_count)
405 		return 0;
406 	for (i = 0; i < bo->placement.num_placement; i++) {
407 		bo->placements[i].lpfn = 0;
408 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
409 	}
410 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
411 	if (likely(r == 0)) {
412 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
413 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
414 		else
415 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
416 	} else {
417 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
418 	}
419 	return r;
420 }
421 
422 int radeon_bo_evict_vram(struct radeon_device *rdev)
423 {
424 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
425 #ifndef CONFIG_HIBERNATION
426 	if (rdev->flags & RADEON_IS_IGP) {
427 		if (rdev->mc.igp_sideport_enabled == false)
428 			/* Useless to evict on IGP chips */
429 			return 0;
430 	}
431 #endif
432 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
433 }
434 
435 void radeon_bo_force_delete(struct radeon_device *rdev)
436 {
437 	struct radeon_bo *bo, *n;
438 
439 	if (list_empty(&rdev->gem.objects)) {
440 		return;
441 	}
442 	dev_err(rdev->dev, "Userspace still has active objects !\n");
443 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
444 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
445 			&bo->tbo.base, bo, (unsigned long)bo->tbo.base.size,
446 			*((unsigned long *)&bo->tbo.base.refcount));
447 		mutex_lock(&bo->rdev->gem.mutex);
448 		list_del_init(&bo->list);
449 		mutex_unlock(&bo->rdev->gem.mutex);
450 		/* this should unref the ttm bo */
451 		drm_gem_object_put_unlocked(&bo->tbo.base);
452 	}
453 }
454 
455 int radeon_bo_init(struct radeon_device *rdev)
456 {
457 	/* reserve PAT memory space to WC for VRAM */
458 	arch_io_reserve_memtype_wc(rdev->mc.aper_base,
459 				   rdev->mc.aper_size);
460 
461 	/* Add an MTRR for the VRAM */
462 	if (!rdev->fastfb_working) {
463 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
464 						      rdev->mc.aper_size);
465 	}
466 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
467 		rdev->mc.mc_vram_size >> 20,
468 		(unsigned long long)rdev->mc.aper_size >> 20);
469 	DRM_INFO("RAM width %dbits %cDR\n",
470 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
471 	return radeon_ttm_init(rdev);
472 }
473 
474 void radeon_bo_fini(struct radeon_device *rdev)
475 {
476 	radeon_ttm_fini(rdev);
477 	arch_phys_wc_del(rdev->mc.vram_mtrr);
478 	arch_io_free_memtype_wc(rdev->mc.aper_base, rdev->mc.aper_size);
479 }
480 
481 /* Returns how many bytes TTM can move per IB.
482  */
483 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
484 {
485 	u64 real_vram_size = rdev->mc.real_vram_size;
486 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
487 
488 	/* This function is based on the current VRAM usage.
489 	 *
490 	 * - If all of VRAM is free, allow relocating the number of bytes that
491 	 *   is equal to 1/4 of the size of VRAM for this IB.
492 
493 	 * - If more than one half of VRAM is occupied, only allow relocating
494 	 *   1 MB of data for this IB.
495 	 *
496 	 * - From 0 to one half of used VRAM, the threshold decreases
497 	 *   linearly.
498 	 *         __________________
499 	 * 1/4 of -|\               |
500 	 * VRAM    | \              |
501 	 *         |  \             |
502 	 *         |   \            |
503 	 *         |    \           |
504 	 *         |     \          |
505 	 *         |      \         |
506 	 *         |       \________|1 MB
507 	 *         |----------------|
508 	 *    VRAM 0 %             100 %
509 	 *         used            used
510 	 *
511 	 * Note: It's a threshold, not a limit. The threshold must be crossed
512 	 * for buffer relocations to stop, so any buffer of an arbitrary size
513 	 * can be moved as long as the threshold isn't crossed before
514 	 * the relocation takes place. We don't want to disable buffer
515 	 * relocations completely.
516 	 *
517 	 * The idea is that buffers should be placed in VRAM at creation time
518 	 * and TTM should only do a minimum number of relocations during
519 	 * command submission. In practice, you need to submit at least
520 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
521 	 *
522 	 * Also, things can get pretty crazy under memory pressure and actual
523 	 * VRAM usage can change a lot, so playing safe even at 50% does
524 	 * consistently increase performance.
525 	 */
526 
527 	u64 half_vram = real_vram_size >> 1;
528 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
529 	u64 bytes_moved_threshold = half_free_vram >> 1;
530 	return max(bytes_moved_threshold, 1024*1024ull);
531 }
532 
533 int radeon_bo_list_validate(struct radeon_device *rdev,
534 			    struct ww_acquire_ctx *ticket,
535 			    struct list_head *head, int ring)
536 {
537 	struct ttm_operation_ctx ctx = { true, false };
538 	struct radeon_bo_list *lobj;
539 	struct list_head duplicates;
540 	int r;
541 	u64 bytes_moved = 0, initial_bytes_moved;
542 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
543 
544 	INIT_LIST_HEAD(&duplicates);
545 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
546 	if (unlikely(r != 0)) {
547 		return r;
548 	}
549 
550 	list_for_each_entry(lobj, head, tv.head) {
551 		struct radeon_bo *bo = lobj->robj;
552 		if (!bo->pin_count) {
553 			u32 domain = lobj->preferred_domains;
554 			u32 allowed = lobj->allowed_domains;
555 			u32 current_domain =
556 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
557 
558 			/* Check if this buffer will be moved and don't move it
559 			 * if we have moved too many buffers for this IB already.
560 			 *
561 			 * Note that this allows moving at least one buffer of
562 			 * any size, because it doesn't take the current "bo"
563 			 * into account. We don't want to disallow buffer moves
564 			 * completely.
565 			 */
566 			if ((allowed & current_domain) != 0 &&
567 			    (domain & current_domain) == 0 && /* will be moved */
568 			    bytes_moved > bytes_moved_threshold) {
569 				/* don't move it */
570 				domain = current_domain;
571 			}
572 
573 		retry:
574 			radeon_ttm_placement_from_domain(bo, domain);
575 			if (ring == R600_RING_TYPE_UVD_INDEX)
576 				radeon_uvd_force_into_uvd_segment(bo, allowed);
577 
578 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
579 			r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
580 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
581 				       initial_bytes_moved;
582 
583 			if (unlikely(r)) {
584 				if (r != -ERESTARTSYS &&
585 				    domain != lobj->allowed_domains) {
586 					domain = lobj->allowed_domains;
587 					goto retry;
588 				}
589 				ttm_eu_backoff_reservation(ticket, head);
590 				return r;
591 			}
592 		}
593 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
594 		lobj->tiling_flags = bo->tiling_flags;
595 	}
596 
597 	list_for_each_entry(lobj, &duplicates, tv.head) {
598 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
599 		lobj->tiling_flags = lobj->robj->tiling_flags;
600 	}
601 
602 	return 0;
603 }
604 
605 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
606 {
607 	struct radeon_device *rdev = bo->rdev;
608 	struct radeon_surface_reg *reg;
609 	struct radeon_bo *old_object;
610 	int steal;
611 	int i;
612 
613 	dma_resv_assert_held(bo->tbo.base.resv);
614 
615 	if (!bo->tiling_flags)
616 		return 0;
617 
618 	if (bo->surface_reg >= 0) {
619 		reg = &rdev->surface_regs[bo->surface_reg];
620 		i = bo->surface_reg;
621 		goto out;
622 	}
623 
624 	steal = -1;
625 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
626 
627 		reg = &rdev->surface_regs[i];
628 		if (!reg->bo)
629 			break;
630 
631 		old_object = reg->bo;
632 		if (old_object->pin_count == 0)
633 			steal = i;
634 	}
635 
636 	/* if we are all out */
637 	if (i == RADEON_GEM_MAX_SURFACES) {
638 		if (steal == -1)
639 			return -ENOMEM;
640 		/* find someone with a surface reg and nuke their BO */
641 		reg = &rdev->surface_regs[steal];
642 		old_object = reg->bo;
643 		/* blow away the mapping */
644 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
645 		ttm_bo_unmap_virtual(&old_object->tbo);
646 		old_object->surface_reg = -1;
647 		i = steal;
648 	}
649 
650 	bo->surface_reg = i;
651 	reg->bo = bo;
652 
653 out:
654 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
655 			       bo->tbo.mem.start << PAGE_SHIFT,
656 			       bo->tbo.num_pages << PAGE_SHIFT);
657 	return 0;
658 }
659 
660 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
661 {
662 	struct radeon_device *rdev = bo->rdev;
663 	struct radeon_surface_reg *reg;
664 
665 	if (bo->surface_reg == -1)
666 		return;
667 
668 	reg = &rdev->surface_regs[bo->surface_reg];
669 	radeon_clear_surface_reg(rdev, bo->surface_reg);
670 
671 	reg->bo = NULL;
672 	bo->surface_reg = -1;
673 }
674 
675 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
676 				uint32_t tiling_flags, uint32_t pitch)
677 {
678 	struct radeon_device *rdev = bo->rdev;
679 	int r;
680 
681 	if (rdev->family >= CHIP_CEDAR) {
682 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
683 
684 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
685 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
686 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
687 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
688 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
689 		switch (bankw) {
690 		case 0:
691 		case 1:
692 		case 2:
693 		case 4:
694 		case 8:
695 			break;
696 		default:
697 			return -EINVAL;
698 		}
699 		switch (bankh) {
700 		case 0:
701 		case 1:
702 		case 2:
703 		case 4:
704 		case 8:
705 			break;
706 		default:
707 			return -EINVAL;
708 		}
709 		switch (mtaspect) {
710 		case 0:
711 		case 1:
712 		case 2:
713 		case 4:
714 		case 8:
715 			break;
716 		default:
717 			return -EINVAL;
718 		}
719 		if (tilesplit > 6) {
720 			return -EINVAL;
721 		}
722 		if (stilesplit > 6) {
723 			return -EINVAL;
724 		}
725 	}
726 	r = radeon_bo_reserve(bo, false);
727 	if (unlikely(r != 0))
728 		return r;
729 	bo->tiling_flags = tiling_flags;
730 	bo->pitch = pitch;
731 	radeon_bo_unreserve(bo);
732 	return 0;
733 }
734 
735 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
736 				uint32_t *tiling_flags,
737 				uint32_t *pitch)
738 {
739 	dma_resv_assert_held(bo->tbo.base.resv);
740 
741 	if (tiling_flags)
742 		*tiling_flags = bo->tiling_flags;
743 	if (pitch)
744 		*pitch = bo->pitch;
745 }
746 
747 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
748 				bool force_drop)
749 {
750 	if (!force_drop)
751 		dma_resv_assert_held(bo->tbo.base.resv);
752 
753 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
754 		return 0;
755 
756 	if (force_drop) {
757 		radeon_bo_clear_surface_reg(bo);
758 		return 0;
759 	}
760 
761 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
762 		if (!has_moved)
763 			return 0;
764 
765 		if (bo->surface_reg >= 0)
766 			radeon_bo_clear_surface_reg(bo);
767 		return 0;
768 	}
769 
770 	if ((bo->surface_reg >= 0) && !has_moved)
771 		return 0;
772 
773 	return radeon_bo_get_surface_reg(bo);
774 }
775 
776 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
777 			   bool evict,
778 			   struct ttm_mem_reg *new_mem)
779 {
780 	struct radeon_bo *rbo;
781 
782 	if (!radeon_ttm_bo_is_radeon_bo(bo))
783 		return;
784 
785 	rbo = container_of(bo, struct radeon_bo, tbo);
786 	radeon_bo_check_tiling(rbo, 0, 1);
787 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
788 
789 	/* update statistics */
790 	if (!new_mem)
791 		return;
792 
793 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
794 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
795 }
796 
797 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
798 {
799 	struct ttm_operation_ctx ctx = { false, false };
800 	struct radeon_device *rdev;
801 	struct radeon_bo *rbo;
802 	unsigned long offset, size, lpfn;
803 	int i, r;
804 
805 	if (!radeon_ttm_bo_is_radeon_bo(bo))
806 		return 0;
807 	rbo = container_of(bo, struct radeon_bo, tbo);
808 	radeon_bo_check_tiling(rbo, 0, 0);
809 	rdev = rbo->rdev;
810 	if (bo->mem.mem_type != TTM_PL_VRAM)
811 		return 0;
812 
813 	size = bo->mem.num_pages << PAGE_SHIFT;
814 	offset = bo->mem.start << PAGE_SHIFT;
815 	if ((offset + size) <= rdev->mc.visible_vram_size)
816 		return 0;
817 
818 	/* Can't move a pinned BO to visible VRAM */
819 	if (rbo->pin_count > 0)
820 		return -EINVAL;
821 
822 	/* hurrah the memory is not visible ! */
823 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
824 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
825 	for (i = 0; i < rbo->placement.num_placement; i++) {
826 		/* Force into visible VRAM */
827 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
828 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
829 			rbo->placements[i].lpfn = lpfn;
830 	}
831 	r = ttm_bo_validate(bo, &rbo->placement, &ctx);
832 	if (unlikely(r == -ENOMEM)) {
833 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
834 		return ttm_bo_validate(bo, &rbo->placement, &ctx);
835 	} else if (unlikely(r != 0)) {
836 		return r;
837 	}
838 
839 	offset = bo->mem.start << PAGE_SHIFT;
840 	/* this should never happen */
841 	if ((offset + size) > rdev->mc.visible_vram_size)
842 		return -EINVAL;
843 
844 	return 0;
845 }
846 
847 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
848 {
849 	int r;
850 
851 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, NULL);
852 	if (unlikely(r != 0))
853 		return r;
854 	if (mem_type)
855 		*mem_type = bo->tbo.mem.mem_type;
856 
857 	r = ttm_bo_wait(&bo->tbo, true, no_wait);
858 	ttm_bo_unreserve(&bo->tbo);
859 	return r;
860 }
861 
862 /**
863  * radeon_bo_fence - add fence to buffer object
864  *
865  * @bo: buffer object in question
866  * @fence: fence to add
867  * @shared: true if fence should be added shared
868  *
869  */
870 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
871 		     bool shared)
872 {
873 	struct dma_resv *resv = bo->tbo.base.resv;
874 
875 	if (shared)
876 		dma_resv_add_shared_fence(resv, &fence->base);
877 	else
878 		dma_resv_add_excl_fence(resv, &fence->base);
879 }
880