xref: /linux/drivers/gpu/drm/radeon/radeon_object.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
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 "radeon_drm.h"
36 #include "radeon.h"
37 #include "radeon_trace.h"
38 
39 
40 int radeon_ttm_init(struct radeon_device *rdev);
41 void radeon_ttm_fini(struct radeon_device *rdev);
42 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo);
43 
44 /*
45  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
46  * function are calling it.
47  */
48 
49 void radeon_bo_clear_va(struct radeon_bo *bo)
50 {
51 	struct radeon_bo_va *bo_va, *tmp;
52 
53 	list_for_each_entry_safe(bo_va, tmp, &bo->va, bo_list) {
54 		/* remove from all vm address space */
55 		mutex_lock(&bo_va->vm->mutex);
56 		list_del(&bo_va->vm_list);
57 		mutex_unlock(&bo_va->vm->mutex);
58 		list_del(&bo_va->bo_list);
59 		kfree(bo_va);
60 	}
61 }
62 
63 static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo)
64 {
65 	struct radeon_bo *bo;
66 
67 	bo = container_of(tbo, struct radeon_bo, tbo);
68 	mutex_lock(&bo->rdev->gem.mutex);
69 	list_del_init(&bo->list);
70 	mutex_unlock(&bo->rdev->gem.mutex);
71 	radeon_bo_clear_surface_reg(bo);
72 	radeon_bo_clear_va(bo);
73 	drm_gem_object_release(&bo->gem_base);
74 	kfree(bo);
75 }
76 
77 bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo)
78 {
79 	if (bo->destroy == &radeon_ttm_bo_destroy)
80 		return true;
81 	return false;
82 }
83 
84 void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain)
85 {
86 	u32 c = 0;
87 
88 	rbo->placement.fpfn = 0;
89 	rbo->placement.lpfn = 0;
90 	rbo->placement.placement = rbo->placements;
91 	rbo->placement.busy_placement = rbo->placements;
92 	if (domain & RADEON_GEM_DOMAIN_VRAM)
93 		rbo->placements[c++] = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
94 					TTM_PL_FLAG_VRAM;
95 	if (domain & RADEON_GEM_DOMAIN_GTT)
96 		rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
97 	if (domain & RADEON_GEM_DOMAIN_CPU)
98 		rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
99 	if (!c)
100 		rbo->placements[c++] = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
101 	rbo->placement.num_placement = c;
102 	rbo->placement.num_busy_placement = c;
103 }
104 
105 int radeon_bo_create(struct radeon_device *rdev,
106 		     unsigned long size, int byte_align, bool kernel, u32 domain,
107 		     struct radeon_bo **bo_ptr)
108 {
109 	struct radeon_bo *bo;
110 	enum ttm_bo_type type;
111 	unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT;
112 	unsigned long max_size = 0;
113 	size_t acc_size;
114 	int r;
115 
116 	size = ALIGN(size, PAGE_SIZE);
117 
118 	if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
119 		rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
120 	}
121 	if (kernel) {
122 		type = ttm_bo_type_kernel;
123 	} else {
124 		type = ttm_bo_type_device;
125 	}
126 	*bo_ptr = NULL;
127 
128 	/* maximun bo size is the minimun btw visible vram and gtt size */
129 	max_size = min(rdev->mc.visible_vram_size, rdev->mc.gtt_size);
130 	if ((page_align << PAGE_SHIFT) >= max_size) {
131 		printk(KERN_WARNING "%s:%d alloc size %ldM bigger than %ldMb limit\n",
132 			__func__, __LINE__, page_align  >> (20 - PAGE_SHIFT), max_size >> 20);
133 		return -ENOMEM;
134 	}
135 
136 	acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size,
137 				       sizeof(struct radeon_bo));
138 
139 retry:
140 	bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL);
141 	if (bo == NULL)
142 		return -ENOMEM;
143 	r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size);
144 	if (unlikely(r)) {
145 		kfree(bo);
146 		return r;
147 	}
148 	bo->rdev = rdev;
149 	bo->gem_base.driver_private = NULL;
150 	bo->surface_reg = -1;
151 	INIT_LIST_HEAD(&bo->list);
152 	INIT_LIST_HEAD(&bo->va);
153 	radeon_ttm_placement_from_domain(bo, domain);
154 	/* Kernel allocation are uninterruptible */
155 	mutex_lock(&rdev->vram_mutex);
156 	r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type,
157 			&bo->placement, page_align, 0, !kernel, NULL,
158 			acc_size, &radeon_ttm_bo_destroy);
159 	mutex_unlock(&rdev->vram_mutex);
160 	if (unlikely(r != 0)) {
161 		if (r != -ERESTARTSYS) {
162 			if (domain == RADEON_GEM_DOMAIN_VRAM) {
163 				domain |= RADEON_GEM_DOMAIN_GTT;
164 				goto retry;
165 			}
166 			dev_err(rdev->dev,
167 				"object_init failed for (%lu, 0x%08X)\n",
168 				size, domain);
169 		}
170 		return r;
171 	}
172 	*bo_ptr = bo;
173 
174 	trace_radeon_bo_create(bo);
175 
176 	return 0;
177 }
178 
179 int radeon_bo_kmap(struct radeon_bo *bo, void **ptr)
180 {
181 	bool is_iomem;
182 	int r;
183 
184 	if (bo->kptr) {
185 		if (ptr) {
186 			*ptr = bo->kptr;
187 		}
188 		return 0;
189 	}
190 	r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap);
191 	if (r) {
192 		return r;
193 	}
194 	bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
195 	if (ptr) {
196 		*ptr = bo->kptr;
197 	}
198 	radeon_bo_check_tiling(bo, 0, 0);
199 	return 0;
200 }
201 
202 void radeon_bo_kunmap(struct radeon_bo *bo)
203 {
204 	if (bo->kptr == NULL)
205 		return;
206 	bo->kptr = NULL;
207 	radeon_bo_check_tiling(bo, 0, 0);
208 	ttm_bo_kunmap(&bo->kmap);
209 }
210 
211 void radeon_bo_unref(struct radeon_bo **bo)
212 {
213 	struct ttm_buffer_object *tbo;
214 	struct radeon_device *rdev;
215 
216 	if ((*bo) == NULL)
217 		return;
218 	rdev = (*bo)->rdev;
219 	tbo = &((*bo)->tbo);
220 	mutex_lock(&rdev->vram_mutex);
221 	ttm_bo_unref(&tbo);
222 	mutex_unlock(&rdev->vram_mutex);
223 	if (tbo == NULL)
224 		*bo = NULL;
225 }
226 
227 int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr)
228 {
229 	int r, i;
230 
231 	if (bo->pin_count) {
232 		bo->pin_count++;
233 		if (gpu_addr)
234 			*gpu_addr = radeon_bo_gpu_offset(bo);
235 		return 0;
236 	}
237 	radeon_ttm_placement_from_domain(bo, domain);
238 	if (domain == RADEON_GEM_DOMAIN_VRAM) {
239 		/* force to pin into visible video ram */
240 		bo->placement.lpfn = bo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
241 	}
242 	for (i = 0; i < bo->placement.num_placement; i++)
243 		bo->placements[i] |= TTM_PL_FLAG_NO_EVICT;
244 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
245 	if (likely(r == 0)) {
246 		bo->pin_count = 1;
247 		if (gpu_addr != NULL)
248 			*gpu_addr = radeon_bo_gpu_offset(bo);
249 	}
250 	if (unlikely(r != 0))
251 		dev_err(bo->rdev->dev, "%p pin failed\n", bo);
252 	return r;
253 }
254 
255 int radeon_bo_unpin(struct radeon_bo *bo)
256 {
257 	int r, i;
258 
259 	if (!bo->pin_count) {
260 		dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo);
261 		return 0;
262 	}
263 	bo->pin_count--;
264 	if (bo->pin_count)
265 		return 0;
266 	for (i = 0; i < bo->placement.num_placement; i++)
267 		bo->placements[i] &= ~TTM_PL_FLAG_NO_EVICT;
268 	r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false, false);
269 	if (unlikely(r != 0))
270 		dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo);
271 	return r;
272 }
273 
274 int radeon_bo_evict_vram(struct radeon_device *rdev)
275 {
276 	/* late 2.6.33 fix IGP hibernate - we need pm ops to do this correct */
277 	if (0 && (rdev->flags & RADEON_IS_IGP)) {
278 		if (rdev->mc.igp_sideport_enabled == false)
279 			/* Useless to evict on IGP chips */
280 			return 0;
281 	}
282 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
283 }
284 
285 void radeon_bo_force_delete(struct radeon_device *rdev)
286 {
287 	struct radeon_bo *bo, *n;
288 
289 	if (list_empty(&rdev->gem.objects)) {
290 		return;
291 	}
292 	dev_err(rdev->dev, "Userspace still has active objects !\n");
293 	list_for_each_entry_safe(bo, n, &rdev->gem.objects, list) {
294 		mutex_lock(&rdev->ddev->struct_mutex);
295 		dev_err(rdev->dev, "%p %p %lu %lu force free\n",
296 			&bo->gem_base, bo, (unsigned long)bo->gem_base.size,
297 			*((unsigned long *)&bo->gem_base.refcount));
298 		mutex_lock(&bo->rdev->gem.mutex);
299 		list_del_init(&bo->list);
300 		mutex_unlock(&bo->rdev->gem.mutex);
301 		/* this should unref the ttm bo */
302 		drm_gem_object_unreference(&bo->gem_base);
303 		mutex_unlock(&rdev->ddev->struct_mutex);
304 	}
305 }
306 
307 int radeon_bo_init(struct radeon_device *rdev)
308 {
309 	/* Add an MTRR for the VRAM */
310 	rdev->mc.vram_mtrr = mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size,
311 			MTRR_TYPE_WRCOMB, 1);
312 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
313 		rdev->mc.mc_vram_size >> 20,
314 		(unsigned long long)rdev->mc.aper_size >> 20);
315 	DRM_INFO("RAM width %dbits %cDR\n",
316 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
317 	return radeon_ttm_init(rdev);
318 }
319 
320 void radeon_bo_fini(struct radeon_device *rdev)
321 {
322 	radeon_ttm_fini(rdev);
323 }
324 
325 void radeon_bo_list_add_object(struct radeon_bo_list *lobj,
326 				struct list_head *head)
327 {
328 	if (lobj->wdomain) {
329 		list_add(&lobj->tv.head, head);
330 	} else {
331 		list_add_tail(&lobj->tv.head, head);
332 	}
333 }
334 
335 int radeon_bo_list_validate(struct list_head *head)
336 {
337 	struct radeon_bo_list *lobj;
338 	struct radeon_bo *bo;
339 	u32 domain;
340 	int r;
341 
342 	r = ttm_eu_reserve_buffers(head);
343 	if (unlikely(r != 0)) {
344 		return r;
345 	}
346 	list_for_each_entry(lobj, head, tv.head) {
347 		bo = lobj->bo;
348 		if (!bo->pin_count) {
349 			domain = lobj->wdomain ? lobj->wdomain : lobj->rdomain;
350 
351 		retry:
352 			radeon_ttm_placement_from_domain(bo, domain);
353 			r = ttm_bo_validate(&bo->tbo, &bo->placement,
354 						true, false, false);
355 			if (unlikely(r)) {
356 				if (r != -ERESTARTSYS && domain == RADEON_GEM_DOMAIN_VRAM) {
357 					domain |= RADEON_GEM_DOMAIN_GTT;
358 					goto retry;
359 				}
360 				return r;
361 			}
362 		}
363 		lobj->gpu_offset = radeon_bo_gpu_offset(bo);
364 		lobj->tiling_flags = bo->tiling_flags;
365 	}
366 	return 0;
367 }
368 
369 int radeon_bo_fbdev_mmap(struct radeon_bo *bo,
370 			     struct vm_area_struct *vma)
371 {
372 	return ttm_fbdev_mmap(vma, &bo->tbo);
373 }
374 
375 int radeon_bo_get_surface_reg(struct radeon_bo *bo)
376 {
377 	struct radeon_device *rdev = bo->rdev;
378 	struct radeon_surface_reg *reg;
379 	struct radeon_bo *old_object;
380 	int steal;
381 	int i;
382 
383 	BUG_ON(!atomic_read(&bo->tbo.reserved));
384 
385 	if (!bo->tiling_flags)
386 		return 0;
387 
388 	if (bo->surface_reg >= 0) {
389 		reg = &rdev->surface_regs[bo->surface_reg];
390 		i = bo->surface_reg;
391 		goto out;
392 	}
393 
394 	steal = -1;
395 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
396 
397 		reg = &rdev->surface_regs[i];
398 		if (!reg->bo)
399 			break;
400 
401 		old_object = reg->bo;
402 		if (old_object->pin_count == 0)
403 			steal = i;
404 	}
405 
406 	/* if we are all out */
407 	if (i == RADEON_GEM_MAX_SURFACES) {
408 		if (steal == -1)
409 			return -ENOMEM;
410 		/* find someone with a surface reg and nuke their BO */
411 		reg = &rdev->surface_regs[steal];
412 		old_object = reg->bo;
413 		/* blow away the mapping */
414 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
415 		ttm_bo_unmap_virtual(&old_object->tbo);
416 		old_object->surface_reg = -1;
417 		i = steal;
418 	}
419 
420 	bo->surface_reg = i;
421 	reg->bo = bo;
422 
423 out:
424 	radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch,
425 			       bo->tbo.mem.start << PAGE_SHIFT,
426 			       bo->tbo.num_pages << PAGE_SHIFT);
427 	return 0;
428 }
429 
430 static void radeon_bo_clear_surface_reg(struct radeon_bo *bo)
431 {
432 	struct radeon_device *rdev = bo->rdev;
433 	struct radeon_surface_reg *reg;
434 
435 	if (bo->surface_reg == -1)
436 		return;
437 
438 	reg = &rdev->surface_regs[bo->surface_reg];
439 	radeon_clear_surface_reg(rdev, bo->surface_reg);
440 
441 	reg->bo = NULL;
442 	bo->surface_reg = -1;
443 }
444 
445 int radeon_bo_set_tiling_flags(struct radeon_bo *bo,
446 				uint32_t tiling_flags, uint32_t pitch)
447 {
448 	int r;
449 
450 	r = radeon_bo_reserve(bo, false);
451 	if (unlikely(r != 0))
452 		return r;
453 	bo->tiling_flags = tiling_flags;
454 	bo->pitch = pitch;
455 	radeon_bo_unreserve(bo);
456 	return 0;
457 }
458 
459 void radeon_bo_get_tiling_flags(struct radeon_bo *bo,
460 				uint32_t *tiling_flags,
461 				uint32_t *pitch)
462 {
463 	BUG_ON(!atomic_read(&bo->tbo.reserved));
464 	if (tiling_flags)
465 		*tiling_flags = bo->tiling_flags;
466 	if (pitch)
467 		*pitch = bo->pitch;
468 }
469 
470 int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved,
471 				bool force_drop)
472 {
473 	BUG_ON(!atomic_read(&bo->tbo.reserved));
474 
475 	if (!(bo->tiling_flags & RADEON_TILING_SURFACE))
476 		return 0;
477 
478 	if (force_drop) {
479 		radeon_bo_clear_surface_reg(bo);
480 		return 0;
481 	}
482 
483 	if (bo->tbo.mem.mem_type != TTM_PL_VRAM) {
484 		if (!has_moved)
485 			return 0;
486 
487 		if (bo->surface_reg >= 0)
488 			radeon_bo_clear_surface_reg(bo);
489 		return 0;
490 	}
491 
492 	if ((bo->surface_reg >= 0) && !has_moved)
493 		return 0;
494 
495 	return radeon_bo_get_surface_reg(bo);
496 }
497 
498 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
499 			   struct ttm_mem_reg *mem)
500 {
501 	struct radeon_bo *rbo;
502 	if (!radeon_ttm_bo_is_radeon_bo(bo))
503 		return;
504 	rbo = container_of(bo, struct radeon_bo, tbo);
505 	radeon_bo_check_tiling(rbo, 0, 1);
506 	radeon_vm_bo_invalidate(rbo->rdev, rbo);
507 }
508 
509 int radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
510 {
511 	struct radeon_device *rdev;
512 	struct radeon_bo *rbo;
513 	unsigned long offset, size;
514 	int r;
515 
516 	if (!radeon_ttm_bo_is_radeon_bo(bo))
517 		return 0;
518 	rbo = container_of(bo, struct radeon_bo, tbo);
519 	radeon_bo_check_tiling(rbo, 0, 0);
520 	rdev = rbo->rdev;
521 	if (bo->mem.mem_type == TTM_PL_VRAM) {
522 		size = bo->mem.num_pages << PAGE_SHIFT;
523 		offset = bo->mem.start << PAGE_SHIFT;
524 		if ((offset + size) > rdev->mc.visible_vram_size) {
525 			/* hurrah the memory is not visible ! */
526 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM);
527 			rbo->placement.lpfn = rdev->mc.visible_vram_size >> PAGE_SHIFT;
528 			r = ttm_bo_validate(bo, &rbo->placement, false, true, false);
529 			if (unlikely(r != 0))
530 				return r;
531 			offset = bo->mem.start << PAGE_SHIFT;
532 			/* this should not happen */
533 			if ((offset + size) > rdev->mc.visible_vram_size)
534 				return -EINVAL;
535 		}
536 	}
537 	return 0;
538 }
539 
540 int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
541 {
542 	int r;
543 
544 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
545 	if (unlikely(r != 0))
546 		return r;
547 	spin_lock(&bo->tbo.bdev->fence_lock);
548 	if (mem_type)
549 		*mem_type = bo->tbo.mem.mem_type;
550 	if (bo->tbo.sync_obj)
551 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
552 	spin_unlock(&bo->tbo.bdev->fence_lock);
553 	ttm_bo_unreserve(&bo->tbo);
554 	return r;
555 }
556 
557 
558 /**
559  * radeon_bo_reserve - reserve bo
560  * @bo:		bo structure
561  * @no_wait:		don't sleep while trying to reserve (return -EBUSY)
562  *
563  * Returns:
564  * -EBUSY: buffer is busy and @no_wait is true
565  * -ERESTARTSYS: A wait for the buffer to become unreserved was interrupted by
566  * a signal. Release all buffer reservations and return to user-space.
567  */
568 int radeon_bo_reserve(struct radeon_bo *bo, bool no_wait)
569 {
570 	int r;
571 
572 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, 0);
573 	if (unlikely(r != 0)) {
574 		if (r != -ERESTARTSYS)
575 			dev_err(bo->rdev->dev, "%p reserve failed\n", bo);
576 		return r;
577 	}
578 	return 0;
579 }
580 
581 /* object have to be reserved */
582 struct radeon_bo_va *radeon_bo_va(struct radeon_bo *rbo, struct radeon_vm *vm)
583 {
584 	struct radeon_bo_va *bo_va;
585 
586 	list_for_each_entry(bo_va, &rbo->va, bo_list) {
587 		if (bo_va->vm == vm) {
588 			return bo_va;
589 		}
590 	}
591 	return NULL;
592 }
593