xref: /linux/drivers/gpu/drm/radeon/radeon_object.c (revision dfc349402de8e95f6a42e8341e9ea193b718eee3)
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 "radeon_drm.h"
35 #include "radeon.h"
36 
37 struct radeon_object {
38 	struct ttm_buffer_object	tobj;
39 	struct list_head		list;
40 	struct radeon_device		*rdev;
41 	struct drm_gem_object		*gobj;
42 	struct ttm_bo_kmap_obj		kmap;
43 	unsigned			pin_count;
44 	uint64_t			gpu_addr;
45 	void				*kptr;
46 	bool				is_iomem;
47 	uint32_t			tiling_flags;
48 	uint32_t			pitch;
49 	int				surface_reg;
50 };
51 
52 int radeon_ttm_init(struct radeon_device *rdev);
53 void radeon_ttm_fini(struct radeon_device *rdev);
54 
55 /*
56  * To exclude mutual BO access we rely on bo_reserve exclusion, as all
57  * function are calling it.
58  */
59 
60 static int radeon_object_reserve(struct radeon_object *robj, bool interruptible)
61 {
62 	return ttm_bo_reserve(&robj->tobj, interruptible, false, false, 0);
63 }
64 
65 static void radeon_object_unreserve(struct radeon_object *robj)
66 {
67 	ttm_bo_unreserve(&robj->tobj);
68 }
69 
70 static void radeon_ttm_object_object_destroy(struct ttm_buffer_object *tobj)
71 {
72 	struct radeon_object *robj;
73 
74 	robj = container_of(tobj, struct radeon_object, tobj);
75 	list_del_init(&robj->list);
76 	radeon_object_clear_surface_reg(robj);
77 	kfree(robj);
78 }
79 
80 static inline void radeon_object_gpu_addr(struct radeon_object *robj)
81 {
82 	/* Default gpu address */
83 	robj->gpu_addr = 0xFFFFFFFFFFFFFFFFULL;
84 	if (robj->tobj.mem.mm_node == NULL) {
85 		return;
86 	}
87 	robj->gpu_addr = ((u64)robj->tobj.mem.mm_node->start) << PAGE_SHIFT;
88 	switch (robj->tobj.mem.mem_type) {
89 	case TTM_PL_VRAM:
90 		robj->gpu_addr += (u64)robj->rdev->mc.vram_location;
91 		break;
92 	case TTM_PL_TT:
93 		robj->gpu_addr += (u64)robj->rdev->mc.gtt_location;
94 		break;
95 	default:
96 		DRM_ERROR("Unknown placement %d\n", robj->tobj.mem.mem_type);
97 		robj->gpu_addr = 0xFFFFFFFFFFFFFFFFULL;
98 		return;
99 	}
100 }
101 
102 static inline uint32_t radeon_object_flags_from_domain(uint32_t domain)
103 {
104 	uint32_t flags = 0;
105 	if (domain & RADEON_GEM_DOMAIN_VRAM) {
106 		flags |= TTM_PL_FLAG_VRAM | TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED;
107 	}
108 	if (domain & RADEON_GEM_DOMAIN_GTT) {
109 		flags |= TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
110 	}
111 	if (domain & RADEON_GEM_DOMAIN_CPU) {
112 		flags |= TTM_PL_FLAG_SYSTEM | TTM_PL_MASK_CACHING;
113 	}
114 	if (!flags) {
115 		flags |= TTM_PL_FLAG_SYSTEM | TTM_PL_MASK_CACHING;
116 	}
117 	return flags;
118 }
119 
120 int radeon_object_create(struct radeon_device *rdev,
121 			 struct drm_gem_object *gobj,
122 			 unsigned long size,
123 			 bool kernel,
124 			 uint32_t domain,
125 			 bool interruptible,
126 			 struct radeon_object **robj_ptr)
127 {
128 	struct radeon_object *robj;
129 	enum ttm_bo_type type;
130 	uint32_t flags;
131 	int r;
132 
133 	if (unlikely(rdev->mman.bdev.dev_mapping == NULL)) {
134 		rdev->mman.bdev.dev_mapping = rdev->ddev->dev_mapping;
135 	}
136 	if (kernel) {
137 		type = ttm_bo_type_kernel;
138 	} else {
139 		type = ttm_bo_type_device;
140 	}
141 	*robj_ptr = NULL;
142 	robj = kzalloc(sizeof(struct radeon_object), GFP_KERNEL);
143 	if (robj == NULL) {
144 		return -ENOMEM;
145 	}
146 	robj->rdev = rdev;
147 	robj->gobj = gobj;
148 	robj->surface_reg = -1;
149 	INIT_LIST_HEAD(&robj->list);
150 
151 	flags = radeon_object_flags_from_domain(domain);
152 	r = ttm_buffer_object_init(&rdev->mman.bdev, &robj->tobj, size, type, flags,
153 				   0, 0, false, NULL, size,
154 				   &radeon_ttm_object_object_destroy);
155 	if (unlikely(r != 0)) {
156 		/* ttm call radeon_ttm_object_object_destroy if error happen */
157 		DRM_ERROR("Failed to allocate TTM object (%ld, 0x%08X, %u)\n",
158 			  size, flags, 0);
159 		return r;
160 	}
161 	*robj_ptr = robj;
162 	if (gobj) {
163 		list_add_tail(&robj->list, &rdev->gem.objects);
164 	}
165 	return 0;
166 }
167 
168 int radeon_object_kmap(struct radeon_object *robj, void **ptr)
169 {
170 	int r;
171 
172 	spin_lock(&robj->tobj.lock);
173 	if (robj->kptr) {
174 		if (ptr) {
175 			*ptr = robj->kptr;
176 		}
177 		spin_unlock(&robj->tobj.lock);
178 		return 0;
179 	}
180 	spin_unlock(&robj->tobj.lock);
181 	r = ttm_bo_kmap(&robj->tobj, 0, robj->tobj.num_pages, &robj->kmap);
182 	if (r) {
183 		return r;
184 	}
185 	spin_lock(&robj->tobj.lock);
186 	robj->kptr = ttm_kmap_obj_virtual(&robj->kmap, &robj->is_iomem);
187 	spin_unlock(&robj->tobj.lock);
188 	if (ptr) {
189 		*ptr = robj->kptr;
190 	}
191 	radeon_object_check_tiling(robj, 0, 0);
192 	return 0;
193 }
194 
195 void radeon_object_kunmap(struct radeon_object *robj)
196 {
197 	spin_lock(&robj->tobj.lock);
198 	if (robj->kptr == NULL) {
199 		spin_unlock(&robj->tobj.lock);
200 		return;
201 	}
202 	robj->kptr = NULL;
203 	spin_unlock(&robj->tobj.lock);
204 	radeon_object_check_tiling(robj, 0, 0);
205 	ttm_bo_kunmap(&robj->kmap);
206 }
207 
208 void radeon_object_unref(struct radeon_object **robj)
209 {
210 	struct ttm_buffer_object *tobj;
211 
212 	if ((*robj) == NULL) {
213 		return;
214 	}
215 	tobj = &((*robj)->tobj);
216 	ttm_bo_unref(&tobj);
217 	if (tobj == NULL) {
218 		*robj = NULL;
219 	}
220 }
221 
222 int radeon_object_mmap(struct radeon_object *robj, uint64_t *offset)
223 {
224 	*offset = robj->tobj.addr_space_offset;
225 	return 0;
226 }
227 
228 int radeon_object_pin(struct radeon_object *robj, uint32_t domain,
229 		      uint64_t *gpu_addr)
230 {
231 	uint32_t flags;
232 	uint32_t tmp;
233 	int r;
234 
235 	flags = radeon_object_flags_from_domain(domain);
236 	spin_lock(&robj->tobj.lock);
237 	if (robj->pin_count) {
238 		robj->pin_count++;
239 		if (gpu_addr != NULL) {
240 			*gpu_addr = robj->gpu_addr;
241 		}
242 		spin_unlock(&robj->tobj.lock);
243 		return 0;
244 	}
245 	spin_unlock(&robj->tobj.lock);
246 	r = radeon_object_reserve(robj, false);
247 	if (unlikely(r != 0)) {
248 		DRM_ERROR("radeon: failed to reserve object for pinning it.\n");
249 		return r;
250 	}
251 	tmp = robj->tobj.mem.placement;
252 	ttm_flag_masked(&tmp, flags, TTM_PL_MASK_MEM);
253 	robj->tobj.proposed_placement = tmp | TTM_PL_FLAG_NO_EVICT | TTM_PL_MASK_CACHING;
254 	r = ttm_buffer_object_validate(&robj->tobj,
255 				       robj->tobj.proposed_placement,
256 				       false, false);
257 	radeon_object_gpu_addr(robj);
258 	if (gpu_addr != NULL) {
259 		*gpu_addr = robj->gpu_addr;
260 	}
261 	robj->pin_count = 1;
262 	if (unlikely(r != 0)) {
263 		DRM_ERROR("radeon: failed to pin object.\n");
264 	}
265 	radeon_object_unreserve(robj);
266 	return r;
267 }
268 
269 void radeon_object_unpin(struct radeon_object *robj)
270 {
271 	uint32_t flags;
272 	int r;
273 
274 	spin_lock(&robj->tobj.lock);
275 	if (!robj->pin_count) {
276 		spin_unlock(&robj->tobj.lock);
277 		printk(KERN_WARNING "Unpin not necessary for %p !\n", robj);
278 		return;
279 	}
280 	robj->pin_count--;
281 	if (robj->pin_count) {
282 		spin_unlock(&robj->tobj.lock);
283 		return;
284 	}
285 	spin_unlock(&robj->tobj.lock);
286 	r = radeon_object_reserve(robj, false);
287 	if (unlikely(r != 0)) {
288 		DRM_ERROR("radeon: failed to reserve object for unpinning it.\n");
289 		return;
290 	}
291 	flags = robj->tobj.mem.placement;
292 	robj->tobj.proposed_placement = flags & ~TTM_PL_FLAG_NO_EVICT;
293 	r = ttm_buffer_object_validate(&robj->tobj,
294 				       robj->tobj.proposed_placement,
295 				       false, false);
296 	if (unlikely(r != 0)) {
297 		DRM_ERROR("radeon: failed to unpin buffer.\n");
298 	}
299 	radeon_object_unreserve(robj);
300 }
301 
302 int radeon_object_wait(struct radeon_object *robj)
303 {
304 	int r = 0;
305 
306 	/* FIXME: should use block reservation instead */
307 	r = radeon_object_reserve(robj, true);
308 	if (unlikely(r != 0)) {
309 		DRM_ERROR("radeon: failed to reserve object for waiting.\n");
310 		return r;
311 	}
312 	spin_lock(&robj->tobj.lock);
313 	if (robj->tobj.sync_obj) {
314 		r = ttm_bo_wait(&robj->tobj, true, true, false);
315 	}
316 	spin_unlock(&robj->tobj.lock);
317 	radeon_object_unreserve(robj);
318 	return r;
319 }
320 
321 int radeon_object_busy_domain(struct radeon_object *robj, uint32_t *cur_placement)
322 {
323 	int r = 0;
324 
325 	r = radeon_object_reserve(robj, true);
326 	if (unlikely(r != 0)) {
327 		DRM_ERROR("radeon: failed to reserve object for waiting.\n");
328 		return r;
329 	}
330 	spin_lock(&robj->tobj.lock);
331 	*cur_placement = robj->tobj.mem.mem_type;
332 	if (robj->tobj.sync_obj) {
333 		r = ttm_bo_wait(&robj->tobj, true, true, true);
334 	}
335 	spin_unlock(&robj->tobj.lock);
336 	radeon_object_unreserve(robj);
337 	return r;
338 }
339 
340 int radeon_object_evict_vram(struct radeon_device *rdev)
341 {
342 	if (rdev->flags & RADEON_IS_IGP) {
343 		/* Useless to evict on IGP chips */
344 		return 0;
345 	}
346 	return ttm_bo_evict_mm(&rdev->mman.bdev, TTM_PL_VRAM);
347 }
348 
349 void radeon_object_force_delete(struct radeon_device *rdev)
350 {
351 	struct radeon_object *robj, *n;
352 	struct drm_gem_object *gobj;
353 
354 	if (list_empty(&rdev->gem.objects)) {
355 		return;
356 	}
357 	DRM_ERROR("Userspace still has active objects !\n");
358 	list_for_each_entry_safe(robj, n, &rdev->gem.objects, list) {
359 		mutex_lock(&rdev->ddev->struct_mutex);
360 		gobj = robj->gobj;
361 		DRM_ERROR("Force free for (%p,%p,%lu,%lu)\n",
362 			  gobj, robj, (unsigned long)gobj->size,
363 			  *((unsigned long *)&gobj->refcount));
364 		list_del_init(&robj->list);
365 		radeon_object_unref(&robj);
366 		gobj->driver_private = NULL;
367 		drm_gem_object_unreference(gobj);
368 		mutex_unlock(&rdev->ddev->struct_mutex);
369 	}
370 }
371 
372 int radeon_object_init(struct radeon_device *rdev)
373 {
374 	/* Add an MTRR for the VRAM */
375 	rdev->mc.vram_mtrr = mtrr_add(rdev->mc.aper_base, rdev->mc.aper_size,
376 			MTRR_TYPE_WRCOMB, 1);
377 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
378 		rdev->mc.mc_vram_size >> 20,
379 		(unsigned long long)rdev->mc.aper_size >> 20);
380 	DRM_INFO("RAM width %dbits %cDR\n",
381 			rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S');
382 	return radeon_ttm_init(rdev);
383 }
384 
385 void radeon_object_fini(struct radeon_device *rdev)
386 {
387 	radeon_ttm_fini(rdev);
388 }
389 
390 void radeon_object_list_add_object(struct radeon_object_list *lobj,
391 				   struct list_head *head)
392 {
393 	if (lobj->wdomain) {
394 		list_add(&lobj->list, head);
395 	} else {
396 		list_add_tail(&lobj->list, head);
397 	}
398 }
399 
400 int radeon_object_list_reserve(struct list_head *head)
401 {
402 	struct radeon_object_list *lobj;
403 	int r;
404 
405 	list_for_each_entry(lobj, head, list){
406 		if (!lobj->robj->pin_count) {
407 			r = radeon_object_reserve(lobj->robj, true);
408 			if (unlikely(r != 0)) {
409 				DRM_ERROR("radeon: failed to reserve object.\n");
410 				return r;
411 			}
412 		} else {
413 		}
414 	}
415 	return 0;
416 }
417 
418 void radeon_object_list_unreserve(struct list_head *head)
419 {
420 	struct radeon_object_list *lobj;
421 
422 	list_for_each_entry(lobj, head, list) {
423 		if (!lobj->robj->pin_count) {
424 			radeon_object_unreserve(lobj->robj);
425 		}
426 	}
427 }
428 
429 int radeon_object_list_validate(struct list_head *head, void *fence)
430 {
431 	struct radeon_object_list *lobj;
432 	struct radeon_object *robj;
433 	struct radeon_fence *old_fence = NULL;
434 	int r;
435 
436 	r = radeon_object_list_reserve(head);
437 	if (unlikely(r != 0)) {
438 		radeon_object_list_unreserve(head);
439 		return r;
440 	}
441 	list_for_each_entry(lobj, head, list) {
442 		robj = lobj->robj;
443 		if (!robj->pin_count) {
444 			if (lobj->wdomain) {
445 				robj->tobj.proposed_placement =
446 					radeon_object_flags_from_domain(lobj->wdomain);
447 			} else {
448 				robj->tobj.proposed_placement =
449 					radeon_object_flags_from_domain(lobj->rdomain);
450 			}
451 			r = ttm_buffer_object_validate(&robj->tobj,
452 						       robj->tobj.proposed_placement,
453 						       true, false);
454 			if (unlikely(r)) {
455 				DRM_ERROR("radeon: failed to validate.\n");
456 				return r;
457 			}
458 			radeon_object_gpu_addr(robj);
459 		}
460 		lobj->gpu_offset = robj->gpu_addr;
461 		lobj->tiling_flags = robj->tiling_flags;
462 		if (fence) {
463 			old_fence = (struct radeon_fence *)robj->tobj.sync_obj;
464 			robj->tobj.sync_obj = radeon_fence_ref(fence);
465 			robj->tobj.sync_obj_arg = NULL;
466 		}
467 		if (old_fence) {
468 			radeon_fence_unref(&old_fence);
469 		}
470 	}
471 	return 0;
472 }
473 
474 void radeon_object_list_unvalidate(struct list_head *head)
475 {
476 	struct radeon_object_list *lobj;
477 	struct radeon_fence *old_fence = NULL;
478 
479 	list_for_each_entry(lobj, head, list) {
480 		old_fence = (struct radeon_fence *)lobj->robj->tobj.sync_obj;
481 		lobj->robj->tobj.sync_obj = NULL;
482 		if (old_fence) {
483 			radeon_fence_unref(&old_fence);
484 		}
485 	}
486 	radeon_object_list_unreserve(head);
487 }
488 
489 void radeon_object_list_clean(struct list_head *head)
490 {
491 	radeon_object_list_unreserve(head);
492 }
493 
494 int radeon_object_fbdev_mmap(struct radeon_object *robj,
495 			     struct vm_area_struct *vma)
496 {
497 	return ttm_fbdev_mmap(vma, &robj->tobj);
498 }
499 
500 unsigned long radeon_object_size(struct radeon_object *robj)
501 {
502 	return robj->tobj.num_pages << PAGE_SHIFT;
503 }
504 
505 int radeon_object_get_surface_reg(struct radeon_object *robj)
506 {
507 	struct radeon_device *rdev = robj->rdev;
508 	struct radeon_surface_reg *reg;
509 	struct radeon_object *old_object;
510 	int steal;
511 	int i;
512 
513 	if (!robj->tiling_flags)
514 		return 0;
515 
516 	if (robj->surface_reg >= 0) {
517 		reg = &rdev->surface_regs[robj->surface_reg];
518 		i = robj->surface_reg;
519 		goto out;
520 	}
521 
522 	steal = -1;
523 	for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) {
524 
525 		reg = &rdev->surface_regs[i];
526 		if (!reg->robj)
527 			break;
528 
529 		old_object = reg->robj;
530 		if (old_object->pin_count == 0)
531 			steal = i;
532 	}
533 
534 	/* if we are all out */
535 	if (i == RADEON_GEM_MAX_SURFACES) {
536 		if (steal == -1)
537 			return -ENOMEM;
538 		/* find someone with a surface reg and nuke their BO */
539 		reg = &rdev->surface_regs[steal];
540 		old_object = reg->robj;
541 		/* blow away the mapping */
542 		DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object);
543 		ttm_bo_unmap_virtual(&old_object->tobj);
544 		old_object->surface_reg = -1;
545 		i = steal;
546 	}
547 
548 	robj->surface_reg = i;
549 	reg->robj = robj;
550 
551 out:
552 	radeon_set_surface_reg(rdev, i, robj->tiling_flags, robj->pitch,
553 			       robj->tobj.mem.mm_node->start << PAGE_SHIFT,
554 			       robj->tobj.num_pages << PAGE_SHIFT);
555 	return 0;
556 }
557 
558 void radeon_object_clear_surface_reg(struct radeon_object *robj)
559 {
560 	struct radeon_device *rdev = robj->rdev;
561 	struct radeon_surface_reg *reg;
562 
563 	if (robj->surface_reg == -1)
564 		return;
565 
566 	reg = &rdev->surface_regs[robj->surface_reg];
567 	radeon_clear_surface_reg(rdev, robj->surface_reg);
568 
569 	reg->robj = NULL;
570 	robj->surface_reg = -1;
571 }
572 
573 void radeon_object_set_tiling_flags(struct radeon_object *robj,
574 				    uint32_t tiling_flags, uint32_t pitch)
575 {
576 	robj->tiling_flags = tiling_flags;
577 	robj->pitch = pitch;
578 }
579 
580 void radeon_object_get_tiling_flags(struct radeon_object *robj,
581 				    uint32_t *tiling_flags,
582 				    uint32_t *pitch)
583 {
584 	if (tiling_flags)
585 		*tiling_flags = robj->tiling_flags;
586 	if (pitch)
587 		*pitch = robj->pitch;
588 }
589 
590 int radeon_object_check_tiling(struct radeon_object *robj, bool has_moved,
591 			       bool force_drop)
592 {
593 	if (!(robj->tiling_flags & RADEON_TILING_SURFACE))
594 		return 0;
595 
596 	if (force_drop) {
597 		radeon_object_clear_surface_reg(robj);
598 		return 0;
599 	}
600 
601 	if (robj->tobj.mem.mem_type != TTM_PL_VRAM) {
602 		if (!has_moved)
603 			return 0;
604 
605 		if (robj->surface_reg >= 0)
606 			radeon_object_clear_surface_reg(robj);
607 		return 0;
608 	}
609 
610 	if ((robj->surface_reg >= 0) && !has_moved)
611 		return 0;
612 
613 	return radeon_object_get_surface_reg(robj);
614 }
615 
616 void radeon_bo_move_notify(struct ttm_buffer_object *bo,
617 			  struct ttm_mem_reg *mem)
618 {
619 	struct radeon_object *robj = container_of(bo, struct radeon_object, tobj);
620 	radeon_object_check_tiling(robj, 0, 1);
621 }
622 
623 void radeon_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
624 {
625 	struct radeon_object *robj = container_of(bo, struct radeon_object, tobj);
626 	radeon_object_check_tiling(robj, 0, 0);
627 }
628