xref: /dflybsd-src/sys/dev/drm/radeon/radeon_object.c (revision 7622fe528e67b227ddf51c26bf9c25a47405c8ad)
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 <drm/drmP.h>
34 #include <drm/radeon_drm.h>
35 #include "radeon.h"
36 #include "radeon_trace.h"
37 
38 
39 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
40 
41 /*
42  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
43  * function are calling it.
44  */
45 
46 static void radeon_update_memory_usage(struct radeon_bo *bo,
47 				       unsigned mem_type, int sign)
48 {
49 	struct radeon_device *rdev = bo->rdev;
50 	u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT;
51 
52 	switch (mem_type) {
53 	case TTM_PL_TT:
54 		if (sign > 0)
55 			atomic64_add(size, &rdev->gtt_usage);
56 		else
57 			atomic64_sub(size, &rdev->gtt_usage);
58 		break;
59 	case TTM_PL_VRAM:
60 		if (sign > 0)
61 			atomic64_add(size, &rdev->vram_usage);
62 		else
63 			atomic64_sub(size, &rdev->vram_usage);
64 		break;
65 	}
66 }
67 
68 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
69 {
70 	struct radeon_bo *bo;
71 
72 	bo = container_of(tbo, struct radeon_bo, tbo);
73 
74 	radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1);
75 	radeon_mn_unregister(bo);
76 
77 	mutex_lock(&bo->rdev->gem.mutex);
78 	list_del_init(&bo->list);
79 	mutex_unlock(&bo->rdev->gem.mutex);
80 	radeon_bo_clear_surface_reg(bo);
81 	WARN_ON(!list_empty(&bo->va));
82 	drm_gem_object_release(&bo->gem_base);
83 	kfree(bo);
84 }
85 
86 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
87 {
88 	if (bo->destroy == &radeon_ttm_bo_destroy)
89 		return true;
90 	return false;
91 }
92 
93 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
94 {
95 	u32 c = 0, i;
96 
97 	rbo->placement.placement = rbo->placements;
98 	rbo->placement.busy_placement = rbo->placements;
99 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
100 		/* Try placing BOs which don't need CPU access outside of the
101 		 * CPU accessible part of VRAM
102 		 */
103 		if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
104 		    rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) {
105 			rbo->placements[c].fpfn =
106 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
107 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
108 						     TTM_PL_FLAG_UNCACHED |
109 						     TTM_PL_FLAG_VRAM;
110 		}
111 
112 		rbo->placements[c].fpfn = 0;
113 		rbo->placements[c++].flags = TTM_PL_FLAG_WC |
114 					     TTM_PL_FLAG_UNCACHED |
115 					     TTM_PL_FLAG_VRAM;
116 	}
117 
118 	if (domain & RADEON_GEM_DOMAIN_GTT) {
119 		if (rbo->flags & RADEON_GEM_GTT_UC) {
120 			rbo->placements[c].fpfn = 0;
121 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
122 				TTM_PL_FLAG_TT;
123 
124 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
125 			   (rbo->rdev->flags & RADEON_IS_AGP)) {
126 			rbo->placements[c].fpfn = 0;
127 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
128 				TTM_PL_FLAG_UNCACHED |
129 				TTM_PL_FLAG_TT;
130 		} else {
131 			rbo->placements[c].fpfn = 0;
132 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
133 						     TTM_PL_FLAG_TT;
134 		}
135 	}
136 
137 	if (domain & RADEON_GEM_DOMAIN_CPU) {
138 		if (rbo->flags & RADEON_GEM_GTT_UC) {
139 			rbo->placements[c].fpfn = 0;
140 			rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED |
141 				TTM_PL_FLAG_SYSTEM;
142 
143 		} else if ((rbo->flags & RADEON_GEM_GTT_WC) ||
144 		    rbo->rdev->flags & RADEON_IS_AGP) {
145 			rbo->placements[c].fpfn = 0;
146 			rbo->placements[c++].flags = TTM_PL_FLAG_WC |
147 				TTM_PL_FLAG_UNCACHED |
148 				TTM_PL_FLAG_SYSTEM;
149 		} else {
150 			rbo->placements[c].fpfn = 0;
151 			rbo->placements[c++].flags = TTM_PL_FLAG_CACHED |
152 						     TTM_PL_FLAG_SYSTEM;
153 		}
154 	}
155 	if (!c) {
156 		rbo->placements[c].fpfn = 0;
157 		rbo->placements[c++].flags = TTM_PL_MASK_CACHING |
158 					     TTM_PL_FLAG_SYSTEM;
159 	}
160 
161 	rbo->placement.num_placement = c;
162 	rbo->placement.num_busy_placement = c;
163 
164 	for (i = 0; i < c; ++i) {
165 		if ((rbo->flags & RADEON_GEM_CPU_ACCESS) &&
166 		    (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
167 		    !rbo->placements[i].fpfn)
168 			rbo->placements[i].lpfn =
169 				rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
170 		else
171 			rbo->placements[i].lpfn = 0;
172 	}
173 }
174 
175 int radeon_bo_create(struct radeon_device *rdev,
176 		     unsigned long size, int byte_align, bool kernel,
177 		     u32 domain, u32 flags, struct sg_table *sg,
178 		     struct reservation_object *resv,
179 		     struct radeon_bo **bo_ptr)
180 {
181 	struct radeon_bo *bo;
182 	enum ttm_bo_type type;
183 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
184 	size_t acc_size;
185 	int r;
186 
187 	size = ALIGN(size, PAGE_SIZE);
188 
189 	if (kernel) {
190 		type = ttm_bo_type_kernel;
191 	} else if (sg) {
192 		type = ttm_bo_type_sg;
193 	} else {
194 		type = ttm_bo_type_device;
195 	}
196 	*bo_ptr = NULL;
197 
198 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
199 				       sizeof(struct radeon_bo));
200 
201 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
202 	if (bo == NULL)
203 		return -ENOMEM;
204 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
205 	if (unlikely(r)) {
206 		kfree(bo);
207 		return r;
208 	}
209 	bo->rdev = rdev;
210 	bo->surface_reg = -1;
211 	INIT_LIST_HEAD(&bo->list);
212 	INIT_LIST_HEAD(&bo->va);
213 	bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM |
214 	                               RADEON_GEM_DOMAIN_GTT |
215 	                               RADEON_GEM_DOMAIN_CPU);
216 
217 	bo->flags = flags;
218 	/* PCI GART is always snooped */
219 	if (!(rdev->flags & RADEON_IS_PCIE))
220 		bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC);
221 
222 #ifdef CONFIG_X86_32
223 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
224 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
225 	 */
226 	bo->flags &= ~RADEON_GEM_GTT_WC;
227 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
228 	/* Don't try to enable write-combining when it can't work, or things
229 	 * may be slow
230 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
231 	 */
232 
233 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
234 	 thanks to write-combining
235 
236 	DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
237 		      "better performance thanks to write-combining\n");
238 	bo->flags &= ~RADEON_GEM_GTT_WC;
239 #endif
240 
241 	radeon_ttm_placement_from_domain(bo, domain);
242 	/* Kernel allocation are uninterruptible */
243 	down_read(&rdev->pm.mclk_lock);
244 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
245 			&bo->placement, page_align, !kernel, NULL,
246 			acc_size, sg, resv, &radeon_ttm_bo_destroy);
247 	up_read(&rdev->pm.mclk_lock);
248 	if (unlikely(r != 0)) {
249 		return r;
250 	}
251 	*bo_ptr = bo;
252 
253 	trace_radeon_bo_create(bo);
254 
255 	return 0;
256 }
257 
258 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
259 {
260 	bool is_iomem;
261 	int r;
262 
263 	if (bo->kptr) {
264 		if (ptr) {
265 			*ptr = bo->kptr;
266 		}
267 		return 0;
268 	}
269 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
270 	if (r) {
271 		return r;
272 	}
273 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
274 	if (ptr) {
275 		*ptr = bo->kptr;
276 	}
277 	radeon_bo_check_tiling(bo, 0, 0);
278 	return 0;
279 }
280 
281 void radeon_bo_kunmap(struct radeon_bo *bo)
282 {
283 	if (bo->kptr == NULL)
284 		return;
285 	bo->kptr = NULL;
286 	radeon_bo_check_tiling(bo, 0, 0);
287 	ttm_bo_kunmap(&bo->kmap);
288 }
289 
290 struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo)
291 {
292 	if (bo == NULL)
293 		return NULL;
294 
295 	ttm_bo_reference(&bo->tbo);
296 	return bo;
297 }
298 
299 void radeon_bo_unref(struct radeon_bo **bo)
300 {
301 	struct ttm_buffer_object *tbo;
302 	struct radeon_device *rdev;
303 
304 	if ((*bo) == NULL)
305 		return;
306 	rdev = (*bo)->rdev;
307 	tbo = &((*bo)->tbo);
308 	ttm_bo_unref(&tbo);
309 	if (tbo == NULL)
310 		*bo = NULL;
311 }
312 
313 int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset,
314 			     u64 *gpu_addr)
315 {
316 	int r, i;
317 
318 	if (bo->pin_count) {
319 		bo->pin_count++;
320 		if (gpu_addr)
321 			*gpu_addr = radeon_bo_gpu_offset(bo);
322 
323 		if (max_offset != 0) {
324 			u64 domain_start;
325 
326 			if (domain == RADEON_GEM_DOMAIN_VRAM)
327 				domain_start = bo->rdev->mc.vram_start;
328 			else
329 				domain_start = bo->rdev->mc.gtt_start;
330 			WARN_ON_ONCE(max_offset <
331 				     (radeon_bo_gpu_offset(bo) - domain_start));
332 		}
333 
334 		return 0;
335 	}
336 	radeon_ttm_placement_from_domain(bo, domain);
337 	for (i = 0; i < bo->placement.num_placement; i++) {
338 		/* force to pin into visible video ram */
339 		if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
340 		    !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) &&
341 		    (!max_offset || max_offset > bo->rdev->mc.visible_vram_size))
342 			bo->placements[i].lpfn =
343 				bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
344 		else
345 			bo->placements[i].lpfn = max_offset >> PAGE_SHIFT;
346 
347 		bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT;
348 	}
349 
350 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
351 	if (likely(r == 0)) {
352 		bo->pin_count = 1;
353 		if (gpu_addr != NULL)
354 			*gpu_addr = radeon_bo_gpu_offset(bo);
355 		if (domain == RADEON_GEM_DOMAIN_VRAM)
356 			bo->rdev->vram_pin_size += radeon_bo_size(bo);
357 		else
358 			bo->rdev->gart_pin_size += radeon_bo_size(bo);
359 	} else {
360 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
361 	}
362 	return r;
363 }
364 
365 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
366 {
367 	return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr);
368 }
369 
370 int radeon_bo_unpin(struct radeon_bo *bo)
371 {
372 	int r, i;
373 
374 	if (!bo->pin_count) {
375 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
376 		return 0;
377 	}
378 	bo->pin_count--;
379 	if (bo->pin_count)
380 		return 0;
381 	for (i = 0; i < bo->placement.num_placement; i++) {
382 		bo->placements[i].lpfn = 0;
383 		bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT;
384 	}
385 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
386 	if (likely(r == 0)) {
387 		if (bo->tbo.mem.mem_type == TTM_PL_VRAM)
388 			bo->rdev->vram_pin_size -= radeon_bo_size(bo);
389 		else
390 			bo->rdev->gart_pin_size -= radeon_bo_size(bo);
391 	} else {
392 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
393 	}
394 	return r;
395 }
396 
397 int radeon_bo_evict_vram(struct radeon_device *rdev)
398 {
399 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
400 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
401 		if (rdev->mc.igp_sideport_enabled == false)
402 			/* Useless to evict on IGP chips */
403 			return 0;
404 	}
405 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
406 }
407 
408 void radeon_bo_force_delete(struct radeon_device *rdev)
409 {
410 	struct radeon_bo *bo, *n;
411 
412 	if (list_empty(&rdev->gem.objects)) {
413 		return;
414 	}
415 	dev_err(rdev->dev, "Userspace still has active objects !\n");
416 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
417 		mutex_lock(&rdev->ddev->struct_mutex);
418 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
419 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
420 			*((unsigned long *)&bo->gem_base.refcount));
421 		mutex_lock(&bo->rdev->gem.mutex);
422 		list_del_init(&bo->list);
423 		mutex_unlock(&bo->rdev->gem.mutex);
424 		/* this should unref the ttm bo */
425 		drm_gem_object_unreference(&bo->gem_base);
426 		mutex_unlock(&rdev->ddev->struct_mutex);
427 	}
428 }
429 
430 int radeon_bo_init(struct radeon_device *rdev)
431 {
432 	/* Add an MTRR for the VRAM */
433 	if (!rdev->fastfb_working) {
434 		rdev->mc.vram_mtrr = arch_phys_wc_add(rdev->mc.aper_base,
435 						      rdev->mc.aper_size);
436 	}
437 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
438 		rdev->mc.mc_vram_size >> 20,
439 		(unsigned long long)rdev->mc.aper_size >> 20);
440 	DRM_INFO("RAM width %dbits %cDR\n",
441 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
442 	return radeon_ttm_init(rdev);
443 }
444 
445 void radeon_bo_fini(struct radeon_device *rdev)
446 {
447 	radeon_ttm_fini(rdev);
448 	arch_phys_wc_del(rdev->mc.vram_mtrr);
449 }
450 
451 /* Returns how many bytes TTM can move per IB.
452  */
453 static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev)
454 {
455 	u64 real_vram_size = rdev->mc.real_vram_size;
456 	u64 vram_usage = atomic64_read(&rdev->vram_usage);
457 
458 	/* This function is based on the current VRAM usage.
459 	 *
460 	 * - If all of VRAM is free, allow relocating the number of bytes that
461 	 *   is equal to 1/4 of the size of VRAM for this IB.
462 
463 	 * - If more than one half of VRAM is occupied, only allow relocating
464 	 *   1 MB of data for this IB.
465 	 *
466 	 * - From 0 to one half of used VRAM, the threshold decreases
467 	 *   linearly.
468 	 *         __________________
469 	 * 1/4 of -|\               |
470 	 * VRAM    | \              |
471 	 *         |  \             |
472 	 *         |   \            |
473 	 *         |    \           |
474 	 *         |     \          |
475 	 *         |      \         |
476 	 *         |       \________|1 MB
477 	 *         |----------------|
478 	 *    VRAM 0 %             100 %
479 	 *         used            used
480 	 *
481 	 * Note: It's a threshold, not a limit. The threshold must be crossed
482 	 * for buffer relocations to stop, so any buffer of an arbitrary size
483 	 * can be moved as long as the threshold isn't crossed before
484 	 * the relocation takes place. We don't want to disable buffer
485 	 * relocations completely.
486 	 *
487 	 * The idea is that buffers should be placed in VRAM at creation time
488 	 * and TTM should only do a minimum number of relocations during
489 	 * command submission. In practice, you need to submit at least
490 	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
491 	 *
492 	 * Also, things can get pretty crazy under memory pressure and actual
493 	 * VRAM usage can change a lot, so playing safe even at 50% does
494 	 * consistently increase performance.
495 	 */
496 
497 	u64 half_vram = real_vram_size >> 1;
498 	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
499 	u64 bytes_moved_threshold = half_free_vram >> 1;
500 	return max(bytes_moved_threshold, 1024*1024ull);
501 }
502 
503 int radeon_bo_list_validate(struct radeon_device *rdev,
504 			    struct ww_acquire_ctx *ticket,
505 			    struct list_head *head, int ring)
506 {
507 	struct radeon_bo_list *lobj;
508 	struct list_head duplicates;
509 	int r;
510 	u64 bytes_moved = 0, initial_bytes_moved;
511 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
512 
513 	INIT_LIST_HEAD(&duplicates);
514 	r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates);
515 	if (unlikely(r != 0)) {
516 		return r;
517 	}
518 
519 	list_for_each_entry(lobj, head, tv.head) {
520 		struct radeon_bo *bo = lobj->robj;
521 		if (!bo->pin_count) {
522 			u32 domain = lobj->prefered_domains;
523 			u32 allowed = lobj->allowed_domains;
524 			u32 current_domain =
525 				radeon_mem_type_to_domain(bo->tbo.mem.mem_type);
526 
527 			/* Check if this buffer will be moved and don't move it
528 			 * if we have moved too many buffers for this IB already.
529 			 *
530 			 * Note that this allows moving at least one buffer of
531 			 * any size, because it doesn't take the current "bo"
532 			 * into account. We don't want to disallow buffer moves
533 			 * completely.
534 			 */
535 			if ((allowed & current_domain) != 0 &&
536 			    (domain & current_domain) == 0 && /* will be moved */
537 			    bytes_moved > bytes_moved_threshold) {
538 				/* don't move it */
539 				domain = current_domain;
540 			}
541 
542 		retry:
543 			radeon_ttm_placement_from_domain(bo, domain);
544 			if (ring == R600_RING_TYPE_UVD_INDEX)
545 				radeon_uvd_force_into_uvd_segment(bo, allowed);
546 
547 			initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved);
548 			r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
549 			bytes_moved += atomic64_read(&rdev->num_bytes_moved) -
550 				       initial_bytes_moved;
551 
552 			if (unlikely(r)) {
553 				if (r != -ERESTARTSYS &&
554 				    domain != lobj->allowed_domains) {
555 					domain = lobj->allowed_domains;
556 					goto retry;
557 				}
558 				ttm_eu_backoff_reservation(ticket, head);
559 				return r;
560 			}
561 		}
562 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
563 		lobj->tiling_flags = bo->tiling_flags;
564 	}
565 
566 	list_for_each_entry(lobj, &duplicates, tv.head) {
567 		lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj);
568 		lobj->tiling_flags = lobj->robj->tiling_flags;
569 	}
570 
571 	return 0;
572 }
573 
574 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
575 			     struct vm_area_struct *vma)
576 {
577 	return ttm_fbdev_mmap(vma, &bo->tbo);
578 }
579 
580 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
581 {
582 	struct radeon_device *rdev = bo->rdev;
583 	struct radeon_surface_reg *reg;
584 	struct radeon_bo *old_object;
585 	int steal;
586 	int i;
587 
588 	lockdep_assert_held(&bo->tbo.resv->lock.base);
589 
590 	if (!bo->tiling_flags)
591 		return 0;
592 
593 	if (bo->surface_reg >= 0) {
594 		reg = &rdev->surface_regs[bo->surface_reg];
595 		i = bo->surface_reg;
596 		goto out;
597 	}
598 
599 	steal = -1;
600 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
601 
602 		reg = &rdev->surface_regs[i];
603 		if (!reg->bo)
604 			break;
605 
606 		old_object = reg->bo;
607 		if (old_object->pin_count == 0)
608 			steal = i;
609 	}
610 
611 	/* if we are all out */
612 	if (i == RADEON_GEM_MAX_SURFACES) {
613 		if (steal == -1)
614 			return -ENOMEM;
615 		/* find someone with a surface reg and nuke their BO */
616 		reg = &rdev->surface_regs[steal];
617 		old_object = reg->bo;
618 		/* blow away the mapping */
619 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
620 		ttm_bo_unmap_virtual(&old_object->tbo);
621 		old_object->surface_reg = -1;
622 		i = steal;
623 	}
624 
625 	bo->surface_reg = i;
626 	reg->bo = bo;
627 
628 out:
629 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
630 			       bo->tbo.mem.start << PAGE_SHIFT,
631 			       bo->tbo.num_pages << PAGE_SHIFT);
632 	return 0;
633 }
634 
635 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
636 {
637 	struct radeon_device *rdev = bo->rdev;
638 	struct radeon_surface_reg *reg;
639 
640 	if (bo->surface_reg == -1)
641 		return;
642 
643 	reg = &rdev->surface_regs[bo->surface_reg];
644 	radeon_clear_surface_reg(rdev, bo->surface_reg);
645 
646 	reg->bo = NULL;
647 	bo->surface_reg = -1;
648 }
649 
650 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
651 				uint32_t tiling_flags, uint32_t pitch)
652 {
653 	struct radeon_device *rdev = bo->rdev;
654 	int r;
655 
656 	if (rdev->family >= CHIP_CEDAR) {
657 		unsigned bankw, bankh, mtaspect, tilesplit, stilesplit;
658 
659 		bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK;
660 		bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK;
661 		mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK;
662 		tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK;
663 		stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK;
664 		switch (bankw) {
665 		case 0:
666 		case 1:
667 		case 2:
668 		case 4:
669 		case 8:
670 			break;
671 		default:
672 			return -EINVAL;
673 		}
674 		switch (bankh) {
675 		case 0:
676 		case 1:
677 		case 2:
678 		case 4:
679 		case 8:
680 			break;
681 		default:
682 			return -EINVAL;
683 		}
684 		switch (mtaspect) {
685 		case 0:
686 		case 1:
687 		case 2:
688 		case 4:
689 		case 8:
690 			break;
691 		default:
692 			return -EINVAL;
693 		}
694 		if (tilesplit > 6) {
695 			return -EINVAL;
696 		}
697 		if (stilesplit > 6) {
698 			return -EINVAL;
699 		}
700 	}
701 	r = radeon_bo_reserve(bo, false);
702 	if (unlikely(r != 0))
703 		return r;
704 	bo->tiling_flags = tiling_flags;
705 	bo->pitch = pitch;
706 	radeon_bo_unreserve(bo);
707 	return 0;
708 }
709 
710 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
711 				uint32_t *tiling_flags,
712 				uint32_t *pitch)
713 {
714 	lockdep_assert_held(&bo->tbo.resv->lock.base);
715 
716 	if (tiling_flags)
717 		*tiling_flags = bo->tiling_flags;
718 	if (pitch)
719 		*pitch = bo->pitch;
720 }
721 
722 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
723 				bool force_drop)
724 {
725 	if (!force_drop)
726 		lockdep_assert_held(&bo->tbo.resv->lock.base);
727 
728 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
729 		return 0;
730 
731 	if (force_drop) {
732 		radeon_bo_clear_surface_reg(bo);
733 		return 0;
734 	}
735 
736 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
737 		if (!has_moved)
738 			return 0;
739 
740 		if (bo->surface_reg >= 0)
741 			radeon_bo_clear_surface_reg(bo);
742 		return 0;
743 	}
744 
745 	if ((bo->surface_reg >= 0) && !has_moved)
746 		return 0;
747 
748 	return radeon_bo_get_surface_reg(bo);
749 }
750 
751 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
752 			   struct ttm_mem_reg *new_mem)
753 {
754 	struct radeon_bo *rbo;
755 
756 	if (!radeon_ttm_bo_is_radeon_bo(bo))
757 		return;
758 
759 	rbo = container_of(bo, struct radeon_bo, tbo);
760 	radeon_bo_check_tiling(rbo, 0, 1);
761 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
762 
763 	/* update statistics */
764 	if (!new_mem)
765 		return;
766 
767 	radeon_update_memory_usage(rbo, bo->mem.mem_type, -1);
768 	radeon_update_memory_usage(rbo, new_mem->mem_type, 1);
769 }
770 
771 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
772 {
773 	struct radeon_device *rdev;
774 	struct radeon_bo *rbo;
775 	unsigned long offset, size, lpfn;
776 	int i, r;
777 
778 	if (!radeon_ttm_bo_is_radeon_bo(bo))
779 		return 0;
780 	rbo = container_of(bo, struct radeon_bo, tbo);
781 	radeon_bo_check_tiling(rbo, 0, 0);
782 	rdev = rbo->rdev;
783 	if (bo->mem.mem_type != TTM_PL_VRAM)
784 		return 0;
785 
786 	size = bo->mem.num_pages << PAGE_SHIFT;
787 	offset = bo->mem.start << PAGE_SHIFT;
788 	if ((offset + size) <= rdev->mc.visible_vram_size)
789 		return 0;
790 
791 	/* hurrah the memory is not visible ! */
792 	radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
793 	lpfn =	rdev->mc.visible_vram_size >> PAGE_SHIFT;
794 	for (i = 0; i < rbo->placement.num_placement; i++) {
795 		/* Force into visible VRAM */
796 		if ((rbo->placements[i].flags & TTM_PL_FLAG_VRAM) &&
797 		    (!rbo->placements[i].lpfn || rbo->placements[i].lpfn > lpfn))
798 			rbo->placements[i].lpfn = lpfn;
799 	}
800 	r = ttm_bo_validate(bo, &rbo->placement, false, false);
801 	if (unlikely(r == -ENOMEM)) {
802 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
803 		return ttm_bo_validate(bo, &rbo->placement, false, false);
804 	} else if (unlikely(r != 0)) {
805 		return r;
806 	}
807 
808 	offset = bo->mem.start << PAGE_SHIFT;
809 	/* this should never happen */
810 	if ((offset + size) > rdev->mc.visible_vram_size)
811 		return -EINVAL;
812 
813 	return 0;
814 }
815 
816 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
817 {
818 	int r;
819 
820 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
821 	if (unlikely(r != 0))
822 		return r;
823 	if (mem_type)
824 		*mem_type = bo->tbo.mem.mem_type;
825 
826 	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
827 	ttm_bo_unreserve(&bo->tbo);
828 	return r;
829 }
830 
831 /**
832  * radeon_bo_fence - add fence to buffer object
833  *
834  * @bo: buffer object in question
835  * @fence: fence to add
836  * @shared: true if fence should be added shared
837  *
838  */
839 void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence,
840                      bool shared)
841 {
842 	struct reservation_object *resv = bo->tbo.resv;
843 
844 	if (shared)
845 		reservation_object_add_shared_fence(resv, &fence->base);
846 	else
847 		reservation_object_add_excl_fence(resv, &fence->base);
848 }
849