xref: /linux/drivers/gpu/drm/radeon/radeon_ttm.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
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 <ttm/ttm_bo_api.h>
33 #include <ttm/ttm_bo_driver.h>
34 #include <ttm/ttm_placement.h>
35 #include <ttm/ttm_module.h>
36 #include <ttm/ttm_page_alloc.h>
37 #include <drm/drmP.h>
38 #include <drm/radeon_drm.h>
39 #include <linux/seq_file.h>
40 #include <linux/slab.h>
41 #include <linux/swiotlb.h>
42 #include <linux/debugfs.h>
43 #include "radeon_reg.h"
44 #include "radeon.h"
45 
46 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
47 
48 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
49 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
50 
51 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
52 {
53 	struct radeon_mman *mman;
54 	struct radeon_device *rdev;
55 
56 	mman = container_of(bdev, struct radeon_mman, bdev);
57 	rdev = container_of(mman, struct radeon_device, mman);
58 	return rdev;
59 }
60 
61 
62 /*
63  * Global memory.
64  */
65 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
66 {
67 	return ttm_mem_global_init(ref->object);
68 }
69 
70 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
71 {
72 	ttm_mem_global_release(ref->object);
73 }
74 
75 static int radeon_ttm_global_init(struct radeon_device *rdev)
76 {
77 	struct drm_global_reference *global_ref;
78 	int r;
79 
80 	rdev->mman.mem_global_referenced = false;
81 	global_ref = &rdev->mman.mem_global_ref;
82 	global_ref->global_type = DRM_GLOBAL_TTM_MEM;
83 	global_ref->size = sizeof(struct ttm_mem_global);
84 	global_ref->init = &radeon_ttm_mem_global_init;
85 	global_ref->release = &radeon_ttm_mem_global_release;
86 	r = drm_global_item_ref(global_ref);
87 	if (r != 0) {
88 		DRM_ERROR("Failed setting up TTM memory accounting "
89 			  "subsystem.\n");
90 		return r;
91 	}
92 
93 	rdev->mman.bo_global_ref.mem_glob =
94 		rdev->mman.mem_global_ref.object;
95 	global_ref = &rdev->mman.bo_global_ref.ref;
96 	global_ref->global_type = DRM_GLOBAL_TTM_BO;
97 	global_ref->size = sizeof(struct ttm_bo_global);
98 	global_ref->init = &ttm_bo_global_init;
99 	global_ref->release = &ttm_bo_global_release;
100 	r = drm_global_item_ref(global_ref);
101 	if (r != 0) {
102 		DRM_ERROR("Failed setting up TTM BO subsystem.\n");
103 		drm_global_item_unref(&rdev->mman.mem_global_ref);
104 		return r;
105 	}
106 
107 	rdev->mman.mem_global_referenced = true;
108 	return 0;
109 }
110 
111 static void radeon_ttm_global_fini(struct radeon_device *rdev)
112 {
113 	if (rdev->mman.mem_global_referenced) {
114 		drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
115 		drm_global_item_unref(&rdev->mman.mem_global_ref);
116 		rdev->mman.mem_global_referenced = false;
117 	}
118 }
119 
120 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
121 {
122 	return 0;
123 }
124 
125 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
126 				struct ttm_mem_type_manager *man)
127 {
128 	struct radeon_device *rdev;
129 
130 	rdev = radeon_get_rdev(bdev);
131 
132 	switch (type) {
133 	case TTM_PL_SYSTEM:
134 		/* System memory */
135 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
136 		man->available_caching = TTM_PL_MASK_CACHING;
137 		man->default_caching = TTM_PL_FLAG_CACHED;
138 		break;
139 	case TTM_PL_TT:
140 		man->func = &ttm_bo_manager_func;
141 		man->gpu_offset = rdev->mc.gtt_start;
142 		man->available_caching = TTM_PL_MASK_CACHING;
143 		man->default_caching = TTM_PL_FLAG_CACHED;
144 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
145 #if __OS_HAS_AGP
146 		if (rdev->flags & RADEON_IS_AGP) {
147 			if (!rdev->ddev->agp) {
148 				DRM_ERROR("AGP is not enabled for memory type %u\n",
149 					  (unsigned)type);
150 				return -EINVAL;
151 			}
152 			if (!rdev->ddev->agp->cant_use_aperture)
153 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
154 			man->available_caching = TTM_PL_FLAG_UNCACHED |
155 						 TTM_PL_FLAG_WC;
156 			man->default_caching = TTM_PL_FLAG_WC;
157 		}
158 #endif
159 		break;
160 	case TTM_PL_VRAM:
161 		/* "On-card" video ram */
162 		man->func = &ttm_bo_manager_func;
163 		man->gpu_offset = rdev->mc.vram_start;
164 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
165 			     TTM_MEMTYPE_FLAG_MAPPABLE;
166 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
167 		man->default_caching = TTM_PL_FLAG_WC;
168 		break;
169 	default:
170 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
171 		return -EINVAL;
172 	}
173 	return 0;
174 }
175 
176 static void radeon_evict_flags(struct ttm_buffer_object *bo,
177 				struct ttm_placement *placement)
178 {
179 	struct radeon_bo *rbo;
180 	static u32 placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
181 
182 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
183 		placement->fpfn = 0;
184 		placement->lpfn = 0;
185 		placement->placement = &placements;
186 		placement->busy_placement = &placements;
187 		placement->num_placement = 1;
188 		placement->num_busy_placement = 1;
189 		return;
190 	}
191 	rbo = container_of(bo, struct radeon_bo, tbo);
192 	switch (bo->mem.mem_type) {
193 	case TTM_PL_VRAM:
194 		if (rbo->rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready == false)
195 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
196 		else
197 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
198 		break;
199 	case TTM_PL_TT:
200 	default:
201 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
202 	}
203 	*placement = rbo->placement;
204 }
205 
206 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
207 {
208 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
209 
210 	return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
211 }
212 
213 static void radeon_move_null(struct ttm_buffer_object *bo,
214 			     struct ttm_mem_reg *new_mem)
215 {
216 	struct ttm_mem_reg *old_mem = &bo->mem;
217 
218 	BUG_ON(old_mem->mm_node != NULL);
219 	*old_mem = *new_mem;
220 	new_mem->mm_node = NULL;
221 }
222 
223 static int radeon_move_blit(struct ttm_buffer_object *bo,
224 			bool evict, bool no_wait_gpu,
225 			struct ttm_mem_reg *new_mem,
226 			struct ttm_mem_reg *old_mem)
227 {
228 	struct radeon_device *rdev;
229 	uint64_t old_start, new_start;
230 	struct radeon_fence *fence;
231 	int r, ridx;
232 
233 	rdev = radeon_get_rdev(bo->bdev);
234 	ridx = radeon_copy_ring_index(rdev);
235 	old_start = old_mem->start << PAGE_SHIFT;
236 	new_start = new_mem->start << PAGE_SHIFT;
237 
238 	switch (old_mem->mem_type) {
239 	case TTM_PL_VRAM:
240 		old_start += rdev->mc.vram_start;
241 		break;
242 	case TTM_PL_TT:
243 		old_start += rdev->mc.gtt_start;
244 		break;
245 	default:
246 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
247 		return -EINVAL;
248 	}
249 	switch (new_mem->mem_type) {
250 	case TTM_PL_VRAM:
251 		new_start += rdev->mc.vram_start;
252 		break;
253 	case TTM_PL_TT:
254 		new_start += rdev->mc.gtt_start;
255 		break;
256 	default:
257 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
258 		return -EINVAL;
259 	}
260 	if (!rdev->ring[ridx].ready) {
261 		DRM_ERROR("Trying to move memory with ring turned off.\n");
262 		return -EINVAL;
263 	}
264 
265 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
266 
267 	/* sync other rings */
268 	fence = bo->sync_obj;
269 	r = radeon_copy(rdev, old_start, new_start,
270 			new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
271 			&fence);
272 	/* FIXME: handle copy error */
273 	r = ttm_bo_move_accel_cleanup(bo, (void *)fence,
274 				      evict, no_wait_gpu, new_mem);
275 	radeon_fence_unref(&fence);
276 	return r;
277 }
278 
279 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
280 				bool evict, bool interruptible,
281 				bool no_wait_gpu,
282 				struct ttm_mem_reg *new_mem)
283 {
284 	struct radeon_device *rdev;
285 	struct ttm_mem_reg *old_mem = &bo->mem;
286 	struct ttm_mem_reg tmp_mem;
287 	u32 placements;
288 	struct ttm_placement placement;
289 	int r;
290 
291 	rdev = radeon_get_rdev(bo->bdev);
292 	tmp_mem = *new_mem;
293 	tmp_mem.mm_node = NULL;
294 	placement.fpfn = 0;
295 	placement.lpfn = 0;
296 	placement.num_placement = 1;
297 	placement.placement = &placements;
298 	placement.num_busy_placement = 1;
299 	placement.busy_placement = &placements;
300 	placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
301 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
302 			     interruptible, no_wait_gpu);
303 	if (unlikely(r)) {
304 		return r;
305 	}
306 
307 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
308 	if (unlikely(r)) {
309 		goto out_cleanup;
310 	}
311 
312 	r = ttm_tt_bind(bo->ttm, &tmp_mem);
313 	if (unlikely(r)) {
314 		goto out_cleanup;
315 	}
316 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
317 	if (unlikely(r)) {
318 		goto out_cleanup;
319 	}
320 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
321 out_cleanup:
322 	ttm_bo_mem_put(bo, &tmp_mem);
323 	return r;
324 }
325 
326 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
327 				bool evict, bool interruptible,
328 				bool no_wait_gpu,
329 				struct ttm_mem_reg *new_mem)
330 {
331 	struct radeon_device *rdev;
332 	struct ttm_mem_reg *old_mem = &bo->mem;
333 	struct ttm_mem_reg tmp_mem;
334 	struct ttm_placement placement;
335 	u32 placements;
336 	int r;
337 
338 	rdev = radeon_get_rdev(bo->bdev);
339 	tmp_mem = *new_mem;
340 	tmp_mem.mm_node = NULL;
341 	placement.fpfn = 0;
342 	placement.lpfn = 0;
343 	placement.num_placement = 1;
344 	placement.placement = &placements;
345 	placement.num_busy_placement = 1;
346 	placement.busy_placement = &placements;
347 	placements = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
348 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
349 			     interruptible, no_wait_gpu);
350 	if (unlikely(r)) {
351 		return r;
352 	}
353 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
354 	if (unlikely(r)) {
355 		goto out_cleanup;
356 	}
357 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
358 	if (unlikely(r)) {
359 		goto out_cleanup;
360 	}
361 out_cleanup:
362 	ttm_bo_mem_put(bo, &tmp_mem);
363 	return r;
364 }
365 
366 static int radeon_bo_move(struct ttm_buffer_object *bo,
367 			bool evict, bool interruptible,
368 			bool no_wait_gpu,
369 			struct ttm_mem_reg *new_mem)
370 {
371 	struct radeon_device *rdev;
372 	struct ttm_mem_reg *old_mem = &bo->mem;
373 	int r;
374 
375 	rdev = radeon_get_rdev(bo->bdev);
376 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
377 		radeon_move_null(bo, new_mem);
378 		return 0;
379 	}
380 	if ((old_mem->mem_type == TTM_PL_TT &&
381 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
382 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
383 	     new_mem->mem_type == TTM_PL_TT)) {
384 		/* bind is enough */
385 		radeon_move_null(bo, new_mem);
386 		return 0;
387 	}
388 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
389 	    rdev->asic->copy.copy == NULL) {
390 		/* use memcpy */
391 		goto memcpy;
392 	}
393 
394 	if (old_mem->mem_type == TTM_PL_VRAM &&
395 	    new_mem->mem_type == TTM_PL_SYSTEM) {
396 		r = radeon_move_vram_ram(bo, evict, interruptible,
397 					no_wait_gpu, new_mem);
398 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
399 		   new_mem->mem_type == TTM_PL_VRAM) {
400 		r = radeon_move_ram_vram(bo, evict, interruptible,
401 					    no_wait_gpu, new_mem);
402 	} else {
403 		r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
404 	}
405 
406 	if (r) {
407 memcpy:
408 		r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
409 		if (r) {
410 			return r;
411 		}
412 	}
413 
414 	/* update statistics */
415 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
416 	return 0;
417 }
418 
419 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
420 {
421 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
422 	struct radeon_device *rdev = radeon_get_rdev(bdev);
423 
424 	mem->bus.addr = NULL;
425 	mem->bus.offset = 0;
426 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
427 	mem->bus.base = 0;
428 	mem->bus.is_iomem = false;
429 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
430 		return -EINVAL;
431 	switch (mem->mem_type) {
432 	case TTM_PL_SYSTEM:
433 		/* system memory */
434 		return 0;
435 	case TTM_PL_TT:
436 #if __OS_HAS_AGP
437 		if (rdev->flags & RADEON_IS_AGP) {
438 			/* RADEON_IS_AGP is set only if AGP is active */
439 			mem->bus.offset = mem->start << PAGE_SHIFT;
440 			mem->bus.base = rdev->mc.agp_base;
441 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
442 		}
443 #endif
444 		break;
445 	case TTM_PL_VRAM:
446 		mem->bus.offset = mem->start << PAGE_SHIFT;
447 		/* check if it's visible */
448 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
449 			return -EINVAL;
450 		mem->bus.base = rdev->mc.aper_base;
451 		mem->bus.is_iomem = true;
452 #ifdef __alpha__
453 		/*
454 		 * Alpha: use bus.addr to hold the ioremap() return,
455 		 * so we can modify bus.base below.
456 		 */
457 		if (mem->placement & TTM_PL_FLAG_WC)
458 			mem->bus.addr =
459 				ioremap_wc(mem->bus.base + mem->bus.offset,
460 					   mem->bus.size);
461 		else
462 			mem->bus.addr =
463 				ioremap_nocache(mem->bus.base + mem->bus.offset,
464 						mem->bus.size);
465 
466 		/*
467 		 * Alpha: Use just the bus offset plus
468 		 * the hose/domain memory base for bus.base.
469 		 * It then can be used to build PTEs for VRAM
470 		 * access, as done in ttm_bo_vm_fault().
471 		 */
472 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
473 			rdev->ddev->hose->dense_mem_base;
474 #endif
475 		break;
476 	default:
477 		return -EINVAL;
478 	}
479 	return 0;
480 }
481 
482 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
483 {
484 }
485 
486 static int radeon_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
487 {
488 	return radeon_fence_wait((struct radeon_fence *)sync_obj, interruptible);
489 }
490 
491 static int radeon_sync_obj_flush(void *sync_obj)
492 {
493 	return 0;
494 }
495 
496 static void radeon_sync_obj_unref(void **sync_obj)
497 {
498 	radeon_fence_unref((struct radeon_fence **)sync_obj);
499 }
500 
501 static void *radeon_sync_obj_ref(void *sync_obj)
502 {
503 	return radeon_fence_ref((struct radeon_fence *)sync_obj);
504 }
505 
506 static bool radeon_sync_obj_signaled(void *sync_obj)
507 {
508 	return radeon_fence_signaled((struct radeon_fence *)sync_obj);
509 }
510 
511 /*
512  * TTM backend functions.
513  */
514 struct radeon_ttm_tt {
515 	struct ttm_dma_tt		ttm;
516 	struct radeon_device		*rdev;
517 	u64				offset;
518 };
519 
520 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
521 				   struct ttm_mem_reg *bo_mem)
522 {
523 	struct radeon_ttm_tt *gtt = (void*)ttm;
524 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
525 		RADEON_GART_PAGE_WRITE;
526 	int r;
527 
528 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
529 	if (!ttm->num_pages) {
530 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
531 		     ttm->num_pages, bo_mem, ttm);
532 	}
533 	if (ttm->caching_state == tt_cached)
534 		flags |= RADEON_GART_PAGE_SNOOP;
535 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
536 			     ttm->pages, gtt->ttm.dma_address, flags);
537 	if (r) {
538 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
539 			  ttm->num_pages, (unsigned)gtt->offset);
540 		return r;
541 	}
542 	return 0;
543 }
544 
545 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
546 {
547 	struct radeon_ttm_tt *gtt = (void *)ttm;
548 
549 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
550 	return 0;
551 }
552 
553 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
554 {
555 	struct radeon_ttm_tt *gtt = (void *)ttm;
556 
557 	ttm_dma_tt_fini(&gtt->ttm);
558 	kfree(gtt);
559 }
560 
561 static struct ttm_backend_func radeon_backend_func = {
562 	.bind = &radeon_ttm_backend_bind,
563 	.unbind = &radeon_ttm_backend_unbind,
564 	.destroy = &radeon_ttm_backend_destroy,
565 };
566 
567 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
568 				    unsigned long size, uint32_t page_flags,
569 				    struct page *dummy_read_page)
570 {
571 	struct radeon_device *rdev;
572 	struct radeon_ttm_tt *gtt;
573 
574 	rdev = radeon_get_rdev(bdev);
575 #if __OS_HAS_AGP
576 	if (rdev->flags & RADEON_IS_AGP) {
577 		return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
578 					 size, page_flags, dummy_read_page);
579 	}
580 #endif
581 
582 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
583 	if (gtt == NULL) {
584 		return NULL;
585 	}
586 	gtt->ttm.ttm.func = &radeon_backend_func;
587 	gtt->rdev = rdev;
588 	if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
589 		kfree(gtt);
590 		return NULL;
591 	}
592 	return &gtt->ttm.ttm;
593 }
594 
595 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
596 {
597 	struct radeon_device *rdev;
598 	struct radeon_ttm_tt *gtt = (void *)ttm;
599 	unsigned i;
600 	int r;
601 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
602 
603 	if (ttm->state != tt_unpopulated)
604 		return 0;
605 
606 	if (slave && ttm->sg) {
607 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
608 						 gtt->ttm.dma_address, ttm->num_pages);
609 		ttm->state = tt_unbound;
610 		return 0;
611 	}
612 
613 	rdev = radeon_get_rdev(ttm->bdev);
614 #if __OS_HAS_AGP
615 	if (rdev->flags & RADEON_IS_AGP) {
616 		return ttm_agp_tt_populate(ttm);
617 	}
618 #endif
619 
620 #ifdef CONFIG_SWIOTLB
621 	if (swiotlb_nr_tbl()) {
622 		return ttm_dma_populate(&gtt->ttm, rdev->dev);
623 	}
624 #endif
625 
626 	r = ttm_pool_populate(ttm);
627 	if (r) {
628 		return r;
629 	}
630 
631 	for (i = 0; i < ttm->num_pages; i++) {
632 		gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
633 						       0, PAGE_SIZE,
634 						       PCI_DMA_BIDIRECTIONAL);
635 		if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
636 			while (--i) {
637 				pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
638 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
639 				gtt->ttm.dma_address[i] = 0;
640 			}
641 			ttm_pool_unpopulate(ttm);
642 			return -EFAULT;
643 		}
644 	}
645 	return 0;
646 }
647 
648 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
649 {
650 	struct radeon_device *rdev;
651 	struct radeon_ttm_tt *gtt = (void *)ttm;
652 	unsigned i;
653 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
654 
655 	if (slave)
656 		return;
657 
658 	rdev = radeon_get_rdev(ttm->bdev);
659 #if __OS_HAS_AGP
660 	if (rdev->flags & RADEON_IS_AGP) {
661 		ttm_agp_tt_unpopulate(ttm);
662 		return;
663 	}
664 #endif
665 
666 #ifdef CONFIG_SWIOTLB
667 	if (swiotlb_nr_tbl()) {
668 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
669 		return;
670 	}
671 #endif
672 
673 	for (i = 0; i < ttm->num_pages; i++) {
674 		if (gtt->ttm.dma_address[i]) {
675 			pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
676 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
677 		}
678 	}
679 
680 	ttm_pool_unpopulate(ttm);
681 }
682 
683 static struct ttm_bo_driver radeon_bo_driver = {
684 	.ttm_tt_create = &radeon_ttm_tt_create,
685 	.ttm_tt_populate = &radeon_ttm_tt_populate,
686 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
687 	.invalidate_caches = &radeon_invalidate_caches,
688 	.init_mem_type = &radeon_init_mem_type,
689 	.evict_flags = &radeon_evict_flags,
690 	.move = &radeon_bo_move,
691 	.verify_access = &radeon_verify_access,
692 	.sync_obj_signaled = &radeon_sync_obj_signaled,
693 	.sync_obj_wait = &radeon_sync_obj_wait,
694 	.sync_obj_flush = &radeon_sync_obj_flush,
695 	.sync_obj_unref = &radeon_sync_obj_unref,
696 	.sync_obj_ref = &radeon_sync_obj_ref,
697 	.move_notify = &radeon_bo_move_notify,
698 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
699 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
700 	.io_mem_free = &radeon_ttm_io_mem_free,
701 };
702 
703 int radeon_ttm_init(struct radeon_device *rdev)
704 {
705 	int r;
706 
707 	r = radeon_ttm_global_init(rdev);
708 	if (r) {
709 		return r;
710 	}
711 	/* No others user of address space so set it to 0 */
712 	r = ttm_bo_device_init(&rdev->mman.bdev,
713 			       rdev->mman.bo_global_ref.ref.object,
714 			       &radeon_bo_driver,
715 			       rdev->ddev->anon_inode->i_mapping,
716 			       DRM_FILE_PAGE_OFFSET,
717 			       rdev->need_dma32);
718 	if (r) {
719 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
720 		return r;
721 	}
722 	rdev->mman.initialized = true;
723 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
724 				rdev->mc.real_vram_size >> PAGE_SHIFT);
725 	if (r) {
726 		DRM_ERROR("Failed initializing VRAM heap.\n");
727 		return r;
728 	}
729 	/* Change the size here instead of the init above so only lpfn is affected */
730 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
731 
732 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
733 			     RADEON_GEM_DOMAIN_VRAM, 0,
734 			     NULL, &rdev->stollen_vga_memory);
735 	if (r) {
736 		return r;
737 	}
738 	r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
739 	if (r)
740 		return r;
741 	r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
742 	radeon_bo_unreserve(rdev->stollen_vga_memory);
743 	if (r) {
744 		radeon_bo_unref(&rdev->stollen_vga_memory);
745 		return r;
746 	}
747 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
748 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
749 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
750 				rdev->mc.gtt_size >> PAGE_SHIFT);
751 	if (r) {
752 		DRM_ERROR("Failed initializing GTT heap.\n");
753 		return r;
754 	}
755 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
756 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
757 
758 	r = radeon_ttm_debugfs_init(rdev);
759 	if (r) {
760 		DRM_ERROR("Failed to init debugfs\n");
761 		return r;
762 	}
763 	return 0;
764 }
765 
766 void radeon_ttm_fini(struct radeon_device *rdev)
767 {
768 	int r;
769 
770 	if (!rdev->mman.initialized)
771 		return;
772 	radeon_ttm_debugfs_fini(rdev);
773 	if (rdev->stollen_vga_memory) {
774 		r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
775 		if (r == 0) {
776 			radeon_bo_unpin(rdev->stollen_vga_memory);
777 			radeon_bo_unreserve(rdev->stollen_vga_memory);
778 		}
779 		radeon_bo_unref(&rdev->stollen_vga_memory);
780 	}
781 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
782 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
783 	ttm_bo_device_release(&rdev->mman.bdev);
784 	radeon_gart_fini(rdev);
785 	radeon_ttm_global_fini(rdev);
786 	rdev->mman.initialized = false;
787 	DRM_INFO("radeon: ttm finalized\n");
788 }
789 
790 /* this should only be called at bootup or when userspace
791  * isn't running */
792 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
793 {
794 	struct ttm_mem_type_manager *man;
795 
796 	if (!rdev->mman.initialized)
797 		return;
798 
799 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
800 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
801 	man->size = size >> PAGE_SHIFT;
802 }
803 
804 static struct vm_operations_struct radeon_ttm_vm_ops;
805 static const struct vm_operations_struct *ttm_vm_ops = NULL;
806 
807 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
808 {
809 	struct ttm_buffer_object *bo;
810 	struct radeon_device *rdev;
811 	int r;
812 
813 	bo = (struct ttm_buffer_object *)vma->vm_private_data;
814 	if (bo == NULL) {
815 		return VM_FAULT_NOPAGE;
816 	}
817 	rdev = radeon_get_rdev(bo->bdev);
818 	down_read(&rdev->pm.mclk_lock);
819 	r = ttm_vm_ops->fault(vma, vmf);
820 	up_read(&rdev->pm.mclk_lock);
821 	return r;
822 }
823 
824 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
825 {
826 	struct drm_file *file_priv;
827 	struct radeon_device *rdev;
828 	int r;
829 
830 	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
831 		return drm_mmap(filp, vma);
832 	}
833 
834 	file_priv = filp->private_data;
835 	rdev = file_priv->minor->dev->dev_private;
836 	if (rdev == NULL) {
837 		return -EINVAL;
838 	}
839 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
840 	if (unlikely(r != 0)) {
841 		return r;
842 	}
843 	if (unlikely(ttm_vm_ops == NULL)) {
844 		ttm_vm_ops = vma->vm_ops;
845 		radeon_ttm_vm_ops = *ttm_vm_ops;
846 		radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
847 	}
848 	vma->vm_ops = &radeon_ttm_vm_ops;
849 	return 0;
850 }
851 
852 #if defined(CONFIG_DEBUG_FS)
853 
854 static int radeon_mm_dump_table(struct seq_file *m, void *data)
855 {
856 	struct drm_info_node *node = (struct drm_info_node *)m->private;
857 	unsigned ttm_pl = *(int *)node->info_ent->data;
858 	struct drm_device *dev = node->minor->dev;
859 	struct radeon_device *rdev = dev->dev_private;
860 	struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
861 	int ret;
862 	struct ttm_bo_global *glob = rdev->mman.bdev.glob;
863 
864 	spin_lock(&glob->lru_lock);
865 	ret = drm_mm_dump_table(m, mm);
866 	spin_unlock(&glob->lru_lock);
867 	return ret;
868 }
869 
870 static int ttm_pl_vram = TTM_PL_VRAM;
871 static int ttm_pl_tt = TTM_PL_TT;
872 
873 static struct drm_info_list radeon_ttm_debugfs_list[] = {
874 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
875 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
876 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
877 #ifdef CONFIG_SWIOTLB
878 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
879 #endif
880 };
881 
882 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
883 {
884 	struct radeon_device *rdev = inode->i_private;
885 	i_size_write(inode, rdev->mc.mc_vram_size);
886 	filep->private_data = inode->i_private;
887 	return 0;
888 }
889 
890 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
891 				    size_t size, loff_t *pos)
892 {
893 	struct radeon_device *rdev = f->private_data;
894 	ssize_t result = 0;
895 	int r;
896 
897 	if (size & 0x3 || *pos & 0x3)
898 		return -EINVAL;
899 
900 	while (size) {
901 		unsigned long flags;
902 		uint32_t value;
903 
904 		if (*pos >= rdev->mc.mc_vram_size)
905 			return result;
906 
907 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
908 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
909 		if (rdev->family >= CHIP_CEDAR)
910 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
911 		value = RREG32(RADEON_MM_DATA);
912 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
913 
914 		r = put_user(value, (uint32_t *)buf);
915 		if (r)
916 			return r;
917 
918 		result += 4;
919 		buf += 4;
920 		*pos += 4;
921 		size -= 4;
922 	}
923 
924 	return result;
925 }
926 
927 static const struct file_operations radeon_ttm_vram_fops = {
928 	.owner = THIS_MODULE,
929 	.open = radeon_ttm_vram_open,
930 	.read = radeon_ttm_vram_read,
931 	.llseek = default_llseek
932 };
933 
934 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
935 {
936 	struct radeon_device *rdev = inode->i_private;
937 	i_size_write(inode, rdev->mc.gtt_size);
938 	filep->private_data = inode->i_private;
939 	return 0;
940 }
941 
942 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
943 				   size_t size, loff_t *pos)
944 {
945 	struct radeon_device *rdev = f->private_data;
946 	ssize_t result = 0;
947 	int r;
948 
949 	while (size) {
950 		loff_t p = *pos / PAGE_SIZE;
951 		unsigned off = *pos & ~PAGE_MASK;
952 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
953 		struct page *page;
954 		void *ptr;
955 
956 		if (p >= rdev->gart.num_cpu_pages)
957 			return result;
958 
959 		page = rdev->gart.pages[p];
960 		if (page) {
961 			ptr = kmap(page);
962 			ptr += off;
963 
964 			r = copy_to_user(buf, ptr, cur_size);
965 			kunmap(rdev->gart.pages[p]);
966 		} else
967 			r = clear_user(buf, cur_size);
968 
969 		if (r)
970 			return -EFAULT;
971 
972 		result += cur_size;
973 		buf += cur_size;
974 		*pos += cur_size;
975 		size -= cur_size;
976 	}
977 
978 	return result;
979 }
980 
981 static const struct file_operations radeon_ttm_gtt_fops = {
982 	.owner = THIS_MODULE,
983 	.open = radeon_ttm_gtt_open,
984 	.read = radeon_ttm_gtt_read,
985 	.llseek = default_llseek
986 };
987 
988 #endif
989 
990 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
991 {
992 #if defined(CONFIG_DEBUG_FS)
993 	unsigned count;
994 
995 	struct drm_minor *minor = rdev->ddev->primary;
996 	struct dentry *ent, *root = minor->debugfs_root;
997 
998 	ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
999 				  rdev, &radeon_ttm_vram_fops);
1000 	if (IS_ERR(ent))
1001 		return PTR_ERR(ent);
1002 	rdev->mman.vram = ent;
1003 
1004 	ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1005 				  rdev, &radeon_ttm_gtt_fops);
1006 	if (IS_ERR(ent))
1007 		return PTR_ERR(ent);
1008 	rdev->mman.gtt = ent;
1009 
1010 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1011 
1012 #ifdef CONFIG_SWIOTLB
1013 	if (!swiotlb_nr_tbl())
1014 		--count;
1015 #endif
1016 
1017 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1018 #else
1019 
1020 	return 0;
1021 #endif
1022 }
1023 
1024 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1025 {
1026 #if defined(CONFIG_DEBUG_FS)
1027 
1028 	debugfs_remove(rdev->mman.vram);
1029 	rdev->mman.vram = NULL;
1030 
1031 	debugfs_remove(rdev->mman.gtt);
1032 	rdev->mman.gtt = NULL;
1033 #endif
1034 }
1035