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