xref: /linux/drivers/gpu/drm/radeon/radeon_ttm.c (revision 2ba9268dd603d23e17643437b2246acb6844953b)
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/swap.h>
43 #include <linux/pagemap.h>
44 #include <linux/debugfs.h>
45 #include "radeon_reg.h"
46 #include "radeon.h"
47 
48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
49 
50 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
51 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
52 
53 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
54 {
55 	struct radeon_mman *mman;
56 	struct radeon_device *rdev;
57 
58 	mman = container_of(bdev, struct radeon_mman, bdev);
59 	rdev = container_of(mman, struct radeon_device, mman);
60 	return rdev;
61 }
62 
63 
64 /*
65  * Global memory.
66  */
67 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
68 {
69 	return ttm_mem_global_init(ref->object);
70 }
71 
72 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
73 {
74 	ttm_mem_global_release(ref->object);
75 }
76 
77 static int radeon_ttm_global_init(struct radeon_device *rdev)
78 {
79 	struct drm_global_reference *global_ref;
80 	int r;
81 
82 	rdev->mman.mem_global_referenced = false;
83 	global_ref = &rdev->mman.mem_global_ref;
84 	global_ref->global_type = DRM_GLOBAL_TTM_MEM;
85 	global_ref->size = sizeof(struct ttm_mem_global);
86 	global_ref->init = &radeon_ttm_mem_global_init;
87 	global_ref->release = &radeon_ttm_mem_global_release;
88 	r = drm_global_item_ref(global_ref);
89 	if (r != 0) {
90 		DRM_ERROR("Failed setting up TTM memory accounting "
91 			  "subsystem.\n");
92 		return r;
93 	}
94 
95 	rdev->mman.bo_global_ref.mem_glob =
96 		rdev->mman.mem_global_ref.object;
97 	global_ref = &rdev->mman.bo_global_ref.ref;
98 	global_ref->global_type = DRM_GLOBAL_TTM_BO;
99 	global_ref->size = sizeof(struct ttm_bo_global);
100 	global_ref->init = &ttm_bo_global_init;
101 	global_ref->release = &ttm_bo_global_release;
102 	r = drm_global_item_ref(global_ref);
103 	if (r != 0) {
104 		DRM_ERROR("Failed setting up TTM BO subsystem.\n");
105 		drm_global_item_unref(&rdev->mman.mem_global_ref);
106 		return r;
107 	}
108 
109 	rdev->mman.mem_global_referenced = true;
110 	return 0;
111 }
112 
113 static void radeon_ttm_global_fini(struct radeon_device *rdev)
114 {
115 	if (rdev->mman.mem_global_referenced) {
116 		drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
117 		drm_global_item_unref(&rdev->mman.mem_global_ref);
118 		rdev->mman.mem_global_referenced = false;
119 	}
120 }
121 
122 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
123 {
124 	return 0;
125 }
126 
127 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
128 				struct ttm_mem_type_manager *man)
129 {
130 	struct radeon_device *rdev;
131 
132 	rdev = radeon_get_rdev(bdev);
133 
134 	switch (type) {
135 	case TTM_PL_SYSTEM:
136 		/* System memory */
137 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
138 		man->available_caching = TTM_PL_MASK_CACHING;
139 		man->default_caching = TTM_PL_FLAG_CACHED;
140 		break;
141 	case TTM_PL_TT:
142 		man->func = &ttm_bo_manager_func;
143 		man->gpu_offset = rdev->mc.gtt_start;
144 		man->available_caching = TTM_PL_MASK_CACHING;
145 		man->default_caching = TTM_PL_FLAG_CACHED;
146 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
147 #if __OS_HAS_AGP
148 		if (rdev->flags & RADEON_IS_AGP) {
149 			if (!rdev->ddev->agp) {
150 				DRM_ERROR("AGP is not enabled for memory type %u\n",
151 					  (unsigned)type);
152 				return -EINVAL;
153 			}
154 			if (!rdev->ddev->agp->cant_use_aperture)
155 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
156 			man->available_caching = TTM_PL_FLAG_UNCACHED |
157 						 TTM_PL_FLAG_WC;
158 			man->default_caching = TTM_PL_FLAG_WC;
159 		}
160 #endif
161 		break;
162 	case TTM_PL_VRAM:
163 		/* "On-card" video ram */
164 		man->func = &ttm_bo_manager_func;
165 		man->gpu_offset = rdev->mc.vram_start;
166 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
167 			     TTM_MEMTYPE_FLAG_MAPPABLE;
168 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
169 		man->default_caching = TTM_PL_FLAG_WC;
170 		break;
171 	default:
172 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
173 		return -EINVAL;
174 	}
175 	return 0;
176 }
177 
178 static void radeon_evict_flags(struct ttm_buffer_object *bo,
179 				struct ttm_placement *placement)
180 {
181 	static struct ttm_place placements = {
182 		.fpfn = 0,
183 		.lpfn = 0,
184 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
185 	};
186 
187 	struct radeon_bo *rbo;
188 
189 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
190 		placement->placement = &placements;
191 		placement->busy_placement = &placements;
192 		placement->num_placement = 1;
193 		placement->num_busy_placement = 1;
194 		return;
195 	}
196 	rbo = container_of(bo, struct radeon_bo, tbo);
197 	switch (bo->mem.mem_type) {
198 	case TTM_PL_VRAM:
199 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
200 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
201 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
202 			 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
203 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
204 			int i;
205 
206 			/* Try evicting to the CPU inaccessible part of VRAM
207 			 * first, but only set GTT as busy placement, so this
208 			 * BO will be evicted to GTT rather than causing other
209 			 * BOs to be evicted from VRAM
210 			 */
211 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
212 							 RADEON_GEM_DOMAIN_GTT);
213 			rbo->placement.num_busy_placement = 0;
214 			for (i = 0; i < rbo->placement.num_placement; i++) {
215 				if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
216 					if (rbo->placements[0].fpfn < fpfn)
217 						rbo->placements[0].fpfn = fpfn;
218 				} else {
219 					rbo->placement.busy_placement =
220 						&rbo->placements[i];
221 					rbo->placement.num_busy_placement = 1;
222 				}
223 			}
224 		} else
225 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
226 		break;
227 	case TTM_PL_TT:
228 	default:
229 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
230 	}
231 	*placement = rbo->placement;
232 }
233 
234 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
235 {
236 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
237 
238 	return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
239 }
240 
241 static void radeon_move_null(struct ttm_buffer_object *bo,
242 			     struct ttm_mem_reg *new_mem)
243 {
244 	struct ttm_mem_reg *old_mem = &bo->mem;
245 
246 	BUG_ON(old_mem->mm_node != NULL);
247 	*old_mem = *new_mem;
248 	new_mem->mm_node = NULL;
249 }
250 
251 static int radeon_move_blit(struct ttm_buffer_object *bo,
252 			bool evict, bool no_wait_gpu,
253 			struct ttm_mem_reg *new_mem,
254 			struct ttm_mem_reg *old_mem)
255 {
256 	struct radeon_device *rdev;
257 	uint64_t old_start, new_start;
258 	struct radeon_fence *fence;
259 	unsigned num_pages;
260 	int r, ridx;
261 
262 	rdev = radeon_get_rdev(bo->bdev);
263 	ridx = radeon_copy_ring_index(rdev);
264 	old_start = old_mem->start << PAGE_SHIFT;
265 	new_start = new_mem->start << PAGE_SHIFT;
266 
267 	switch (old_mem->mem_type) {
268 	case TTM_PL_VRAM:
269 		old_start += rdev->mc.vram_start;
270 		break;
271 	case TTM_PL_TT:
272 		old_start += rdev->mc.gtt_start;
273 		break;
274 	default:
275 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
276 		return -EINVAL;
277 	}
278 	switch (new_mem->mem_type) {
279 	case TTM_PL_VRAM:
280 		new_start += rdev->mc.vram_start;
281 		break;
282 	case TTM_PL_TT:
283 		new_start += rdev->mc.gtt_start;
284 		break;
285 	default:
286 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
287 		return -EINVAL;
288 	}
289 	if (!rdev->ring[ridx].ready) {
290 		DRM_ERROR("Trying to move memory with ring turned off.\n");
291 		return -EINVAL;
292 	}
293 
294 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
295 
296 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
297 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->resv);
298 	if (IS_ERR(fence))
299 		return PTR_ERR(fence);
300 
301 	r = ttm_bo_move_accel_cleanup(bo, &fence->base,
302 				      evict, no_wait_gpu, new_mem);
303 	radeon_fence_unref(&fence);
304 	return r;
305 }
306 
307 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
308 				bool evict, bool interruptible,
309 				bool no_wait_gpu,
310 				struct ttm_mem_reg *new_mem)
311 {
312 	struct radeon_device *rdev;
313 	struct ttm_mem_reg *old_mem = &bo->mem;
314 	struct ttm_mem_reg tmp_mem;
315 	struct ttm_place placements;
316 	struct ttm_placement placement;
317 	int r;
318 
319 	rdev = radeon_get_rdev(bo->bdev);
320 	tmp_mem = *new_mem;
321 	tmp_mem.mm_node = NULL;
322 	placement.num_placement = 1;
323 	placement.placement = &placements;
324 	placement.num_busy_placement = 1;
325 	placement.busy_placement = &placements;
326 	placements.fpfn = 0;
327 	placements.lpfn = 0;
328 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
329 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
330 			     interruptible, no_wait_gpu);
331 	if (unlikely(r)) {
332 		return r;
333 	}
334 
335 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
336 	if (unlikely(r)) {
337 		goto out_cleanup;
338 	}
339 
340 	r = ttm_tt_bind(bo->ttm, &tmp_mem);
341 	if (unlikely(r)) {
342 		goto out_cleanup;
343 	}
344 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
345 	if (unlikely(r)) {
346 		goto out_cleanup;
347 	}
348 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
349 out_cleanup:
350 	ttm_bo_mem_put(bo, &tmp_mem);
351 	return r;
352 }
353 
354 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
355 				bool evict, bool interruptible,
356 				bool no_wait_gpu,
357 				struct ttm_mem_reg *new_mem)
358 {
359 	struct radeon_device *rdev;
360 	struct ttm_mem_reg *old_mem = &bo->mem;
361 	struct ttm_mem_reg tmp_mem;
362 	struct ttm_placement placement;
363 	struct ttm_place placements;
364 	int r;
365 
366 	rdev = radeon_get_rdev(bo->bdev);
367 	tmp_mem = *new_mem;
368 	tmp_mem.mm_node = NULL;
369 	placement.num_placement = 1;
370 	placement.placement = &placements;
371 	placement.num_busy_placement = 1;
372 	placement.busy_placement = &placements;
373 	placements.fpfn = 0;
374 	placements.lpfn = 0;
375 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
376 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
377 			     interruptible, no_wait_gpu);
378 	if (unlikely(r)) {
379 		return r;
380 	}
381 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
382 	if (unlikely(r)) {
383 		goto out_cleanup;
384 	}
385 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
386 	if (unlikely(r)) {
387 		goto out_cleanup;
388 	}
389 out_cleanup:
390 	ttm_bo_mem_put(bo, &tmp_mem);
391 	return r;
392 }
393 
394 static int radeon_bo_move(struct ttm_buffer_object *bo,
395 			bool evict, bool interruptible,
396 			bool no_wait_gpu,
397 			struct ttm_mem_reg *new_mem)
398 {
399 	struct radeon_device *rdev;
400 	struct ttm_mem_reg *old_mem = &bo->mem;
401 	int r;
402 
403 	rdev = radeon_get_rdev(bo->bdev);
404 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
405 		radeon_move_null(bo, new_mem);
406 		return 0;
407 	}
408 	if ((old_mem->mem_type == TTM_PL_TT &&
409 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
410 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
411 	     new_mem->mem_type == TTM_PL_TT)) {
412 		/* bind is enough */
413 		radeon_move_null(bo, new_mem);
414 		return 0;
415 	}
416 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
417 	    rdev->asic->copy.copy == NULL) {
418 		/* use memcpy */
419 		goto memcpy;
420 	}
421 
422 	if (old_mem->mem_type == TTM_PL_VRAM &&
423 	    new_mem->mem_type == TTM_PL_SYSTEM) {
424 		r = radeon_move_vram_ram(bo, evict, interruptible,
425 					no_wait_gpu, new_mem);
426 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
427 		   new_mem->mem_type == TTM_PL_VRAM) {
428 		r = radeon_move_ram_vram(bo, evict, interruptible,
429 					    no_wait_gpu, new_mem);
430 	} else {
431 		r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
432 	}
433 
434 	if (r) {
435 memcpy:
436 		r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
437 		if (r) {
438 			return r;
439 		}
440 	}
441 
442 	/* update statistics */
443 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
444 	return 0;
445 }
446 
447 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
448 {
449 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
450 	struct radeon_device *rdev = radeon_get_rdev(bdev);
451 
452 	mem->bus.addr = NULL;
453 	mem->bus.offset = 0;
454 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
455 	mem->bus.base = 0;
456 	mem->bus.is_iomem = false;
457 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
458 		return -EINVAL;
459 	switch (mem->mem_type) {
460 	case TTM_PL_SYSTEM:
461 		/* system memory */
462 		return 0;
463 	case TTM_PL_TT:
464 #if __OS_HAS_AGP
465 		if (rdev->flags & RADEON_IS_AGP) {
466 			/* RADEON_IS_AGP is set only if AGP is active */
467 			mem->bus.offset = mem->start << PAGE_SHIFT;
468 			mem->bus.base = rdev->mc.agp_base;
469 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
470 		}
471 #endif
472 		break;
473 	case TTM_PL_VRAM:
474 		mem->bus.offset = mem->start << PAGE_SHIFT;
475 		/* check if it's visible */
476 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
477 			return -EINVAL;
478 		mem->bus.base = rdev->mc.aper_base;
479 		mem->bus.is_iomem = true;
480 #ifdef __alpha__
481 		/*
482 		 * Alpha: use bus.addr to hold the ioremap() return,
483 		 * so we can modify bus.base below.
484 		 */
485 		if (mem->placement & TTM_PL_FLAG_WC)
486 			mem->bus.addr =
487 				ioremap_wc(mem->bus.base + mem->bus.offset,
488 					   mem->bus.size);
489 		else
490 			mem->bus.addr =
491 				ioremap_nocache(mem->bus.base + mem->bus.offset,
492 						mem->bus.size);
493 
494 		/*
495 		 * Alpha: Use just the bus offset plus
496 		 * the hose/domain memory base for bus.base.
497 		 * It then can be used to build PTEs for VRAM
498 		 * access, as done in ttm_bo_vm_fault().
499 		 */
500 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
501 			rdev->ddev->hose->dense_mem_base;
502 #endif
503 		break;
504 	default:
505 		return -EINVAL;
506 	}
507 	return 0;
508 }
509 
510 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
511 {
512 }
513 
514 /*
515  * TTM backend functions.
516  */
517 struct radeon_ttm_tt {
518 	struct ttm_dma_tt		ttm;
519 	struct radeon_device		*rdev;
520 	u64				offset;
521 
522 	uint64_t			userptr;
523 	struct mm_struct		*usermm;
524 	uint32_t			userflags;
525 };
526 
527 /* prepare the sg table with the user pages */
528 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
529 {
530 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
531 	struct radeon_ttm_tt *gtt = (void *)ttm;
532 	unsigned pinned = 0, nents;
533 	int r;
534 
535 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
536 	enum dma_data_direction direction = write ?
537 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
538 
539 	if (current->mm != gtt->usermm)
540 		return -EPERM;
541 
542 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
543 		/* check that we only pin down anonymous memory
544 		   to prevent problems with writeback */
545 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
546 		struct vm_area_struct *vma;
547 		vma = find_vma(gtt->usermm, gtt->userptr);
548 		if (!vma || vma->vm_file || vma->vm_end < end)
549 			return -EPERM;
550 	}
551 
552 	do {
553 		unsigned num_pages = ttm->num_pages - pinned;
554 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
555 		struct page **pages = ttm->pages + pinned;
556 
557 		r = get_user_pages(current, current->mm, userptr, num_pages,
558 				   write, 0, pages, NULL);
559 		if (r < 0)
560 			goto release_pages;
561 
562 		pinned += r;
563 
564 	} while (pinned < ttm->num_pages);
565 
566 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
567 				      ttm->num_pages << PAGE_SHIFT,
568 				      GFP_KERNEL);
569 	if (r)
570 		goto release_sg;
571 
572 	r = -ENOMEM;
573 	nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
574 	if (nents != ttm->sg->nents)
575 		goto release_sg;
576 
577 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
578 					 gtt->ttm.dma_address, ttm->num_pages);
579 
580 	return 0;
581 
582 release_sg:
583 	kfree(ttm->sg);
584 
585 release_pages:
586 	release_pages(ttm->pages, pinned, 0);
587 	return r;
588 }
589 
590 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
591 {
592 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
593 	struct radeon_ttm_tt *gtt = (void *)ttm;
594 	struct scatterlist *sg;
595 	int i;
596 
597 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
598 	enum dma_data_direction direction = write ?
599 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
600 
601 	/* double check that we don't free the table twice */
602 	if (!ttm->sg->sgl)
603 		return;
604 
605 	/* free the sg table and pages again */
606 	dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
607 
608 	for_each_sg(ttm->sg->sgl, sg, ttm->sg->nents, i) {
609 		struct page *page = sg_page(sg);
610 
611 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
612 			set_page_dirty(page);
613 
614 		mark_page_accessed(page);
615 		page_cache_release(page);
616 	}
617 
618 	sg_free_table(ttm->sg);
619 }
620 
621 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
622 				   struct ttm_mem_reg *bo_mem)
623 {
624 	struct radeon_ttm_tt *gtt = (void*)ttm;
625 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
626 		RADEON_GART_PAGE_WRITE;
627 	int r;
628 
629 	if (gtt->userptr) {
630 		radeon_ttm_tt_pin_userptr(ttm);
631 		flags &= ~RADEON_GART_PAGE_WRITE;
632 	}
633 
634 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
635 	if (!ttm->num_pages) {
636 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
637 		     ttm->num_pages, bo_mem, ttm);
638 	}
639 	if (ttm->caching_state == tt_cached)
640 		flags |= RADEON_GART_PAGE_SNOOP;
641 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
642 			     ttm->pages, gtt->ttm.dma_address, flags);
643 	if (r) {
644 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
645 			  ttm->num_pages, (unsigned)gtt->offset);
646 		return r;
647 	}
648 	return 0;
649 }
650 
651 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
652 {
653 	struct radeon_ttm_tt *gtt = (void *)ttm;
654 
655 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
656 
657 	if (gtt->userptr)
658 		radeon_ttm_tt_unpin_userptr(ttm);
659 
660 	return 0;
661 }
662 
663 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
664 {
665 	struct radeon_ttm_tt *gtt = (void *)ttm;
666 
667 	ttm_dma_tt_fini(&gtt->ttm);
668 	kfree(gtt);
669 }
670 
671 static struct ttm_backend_func radeon_backend_func = {
672 	.bind = &radeon_ttm_backend_bind,
673 	.unbind = &radeon_ttm_backend_unbind,
674 	.destroy = &radeon_ttm_backend_destroy,
675 };
676 
677 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
678 				    unsigned long size, uint32_t page_flags,
679 				    struct page *dummy_read_page)
680 {
681 	struct radeon_device *rdev;
682 	struct radeon_ttm_tt *gtt;
683 
684 	rdev = radeon_get_rdev(bdev);
685 #if __OS_HAS_AGP
686 	if (rdev->flags & RADEON_IS_AGP) {
687 		return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
688 					 size, page_flags, dummy_read_page);
689 	}
690 #endif
691 
692 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
693 	if (gtt == NULL) {
694 		return NULL;
695 	}
696 	gtt->ttm.ttm.func = &radeon_backend_func;
697 	gtt->rdev = rdev;
698 	if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
699 		kfree(gtt);
700 		return NULL;
701 	}
702 	return &gtt->ttm.ttm;
703 }
704 
705 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
706 {
707 	if (!ttm || ttm->func != &radeon_backend_func)
708 		return NULL;
709 	return (struct radeon_ttm_tt *)ttm;
710 }
711 
712 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
713 {
714 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
715 	struct radeon_device *rdev;
716 	unsigned i;
717 	int r;
718 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
719 
720 	if (ttm->state != tt_unpopulated)
721 		return 0;
722 
723 	if (gtt && gtt->userptr) {
724 		ttm->sg = kcalloc(1, sizeof(struct sg_table), GFP_KERNEL);
725 		if (!ttm->sg)
726 			return -ENOMEM;
727 
728 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
729 		ttm->state = tt_unbound;
730 		return 0;
731 	}
732 
733 	if (slave && ttm->sg) {
734 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
735 						 gtt->ttm.dma_address, ttm->num_pages);
736 		ttm->state = tt_unbound;
737 		return 0;
738 	}
739 
740 	rdev = radeon_get_rdev(ttm->bdev);
741 #if __OS_HAS_AGP
742 	if (rdev->flags & RADEON_IS_AGP) {
743 		return ttm_agp_tt_populate(ttm);
744 	}
745 #endif
746 
747 #ifdef CONFIG_SWIOTLB
748 	if (swiotlb_nr_tbl()) {
749 		return ttm_dma_populate(&gtt->ttm, rdev->dev);
750 	}
751 #endif
752 
753 	r = ttm_pool_populate(ttm);
754 	if (r) {
755 		return r;
756 	}
757 
758 	for (i = 0; i < ttm->num_pages; i++) {
759 		gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
760 						       0, PAGE_SIZE,
761 						       PCI_DMA_BIDIRECTIONAL);
762 		if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
763 			while (--i) {
764 				pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
765 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
766 				gtt->ttm.dma_address[i] = 0;
767 			}
768 			ttm_pool_unpopulate(ttm);
769 			return -EFAULT;
770 		}
771 	}
772 	return 0;
773 }
774 
775 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
776 {
777 	struct radeon_device *rdev;
778 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
779 	unsigned i;
780 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
781 
782 	if (gtt && gtt->userptr) {
783 		kfree(ttm->sg);
784 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
785 		return;
786 	}
787 
788 	if (slave)
789 		return;
790 
791 	rdev = radeon_get_rdev(ttm->bdev);
792 #if __OS_HAS_AGP
793 	if (rdev->flags & RADEON_IS_AGP) {
794 		ttm_agp_tt_unpopulate(ttm);
795 		return;
796 	}
797 #endif
798 
799 #ifdef CONFIG_SWIOTLB
800 	if (swiotlb_nr_tbl()) {
801 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
802 		return;
803 	}
804 #endif
805 
806 	for (i = 0; i < ttm->num_pages; i++) {
807 		if (gtt->ttm.dma_address[i]) {
808 			pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
809 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
810 		}
811 	}
812 
813 	ttm_pool_unpopulate(ttm);
814 }
815 
816 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
817 			      uint32_t flags)
818 {
819 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
820 
821 	if (gtt == NULL)
822 		return -EINVAL;
823 
824 	gtt->userptr = addr;
825 	gtt->usermm = current->mm;
826 	gtt->userflags = flags;
827 	return 0;
828 }
829 
830 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
831 {
832 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
833 
834 	if (gtt == NULL)
835 		return false;
836 
837 	return !!gtt->userptr;
838 }
839 
840 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
841 {
842 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
843 
844 	if (gtt == NULL)
845 		return false;
846 
847 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
848 }
849 
850 static struct ttm_bo_driver radeon_bo_driver = {
851 	.ttm_tt_create = &radeon_ttm_tt_create,
852 	.ttm_tt_populate = &radeon_ttm_tt_populate,
853 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
854 	.invalidate_caches = &radeon_invalidate_caches,
855 	.init_mem_type = &radeon_init_mem_type,
856 	.evict_flags = &radeon_evict_flags,
857 	.move = &radeon_bo_move,
858 	.verify_access = &radeon_verify_access,
859 	.move_notify = &radeon_bo_move_notify,
860 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
861 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
862 	.io_mem_free = &radeon_ttm_io_mem_free,
863 };
864 
865 int radeon_ttm_init(struct radeon_device *rdev)
866 {
867 	int r;
868 
869 	r = radeon_ttm_global_init(rdev);
870 	if (r) {
871 		return r;
872 	}
873 	/* No others user of address space so set it to 0 */
874 	r = ttm_bo_device_init(&rdev->mman.bdev,
875 			       rdev->mman.bo_global_ref.ref.object,
876 			       &radeon_bo_driver,
877 			       rdev->ddev->anon_inode->i_mapping,
878 			       DRM_FILE_PAGE_OFFSET,
879 			       rdev->need_dma32);
880 	if (r) {
881 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
882 		return r;
883 	}
884 	rdev->mman.initialized = true;
885 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
886 				rdev->mc.real_vram_size >> PAGE_SHIFT);
887 	if (r) {
888 		DRM_ERROR("Failed initializing VRAM heap.\n");
889 		return r;
890 	}
891 	/* Change the size here instead of the init above so only lpfn is affected */
892 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
893 
894 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
895 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
896 			     NULL, &rdev->stollen_vga_memory);
897 	if (r) {
898 		return r;
899 	}
900 	r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
901 	if (r)
902 		return r;
903 	r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
904 	radeon_bo_unreserve(rdev->stollen_vga_memory);
905 	if (r) {
906 		radeon_bo_unref(&rdev->stollen_vga_memory);
907 		return r;
908 	}
909 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
910 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
911 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
912 				rdev->mc.gtt_size >> PAGE_SHIFT);
913 	if (r) {
914 		DRM_ERROR("Failed initializing GTT heap.\n");
915 		return r;
916 	}
917 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
918 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
919 
920 	r = radeon_ttm_debugfs_init(rdev);
921 	if (r) {
922 		DRM_ERROR("Failed to init debugfs\n");
923 		return r;
924 	}
925 	return 0;
926 }
927 
928 void radeon_ttm_fini(struct radeon_device *rdev)
929 {
930 	int r;
931 
932 	if (!rdev->mman.initialized)
933 		return;
934 	radeon_ttm_debugfs_fini(rdev);
935 	if (rdev->stollen_vga_memory) {
936 		r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
937 		if (r == 0) {
938 			radeon_bo_unpin(rdev->stollen_vga_memory);
939 			radeon_bo_unreserve(rdev->stollen_vga_memory);
940 		}
941 		radeon_bo_unref(&rdev->stollen_vga_memory);
942 	}
943 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
944 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
945 	ttm_bo_device_release(&rdev->mman.bdev);
946 	radeon_gart_fini(rdev);
947 	radeon_ttm_global_fini(rdev);
948 	rdev->mman.initialized = false;
949 	DRM_INFO("radeon: ttm finalized\n");
950 }
951 
952 /* this should only be called at bootup or when userspace
953  * isn't running */
954 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
955 {
956 	struct ttm_mem_type_manager *man;
957 
958 	if (!rdev->mman.initialized)
959 		return;
960 
961 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
962 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
963 	man->size = size >> PAGE_SHIFT;
964 }
965 
966 static struct vm_operations_struct radeon_ttm_vm_ops;
967 static const struct vm_operations_struct *ttm_vm_ops = NULL;
968 
969 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
970 {
971 	struct ttm_buffer_object *bo;
972 	struct radeon_device *rdev;
973 	int r;
974 
975 	bo = (struct ttm_buffer_object *)vma->vm_private_data;
976 	if (bo == NULL) {
977 		return VM_FAULT_NOPAGE;
978 	}
979 	rdev = radeon_get_rdev(bo->bdev);
980 	down_read(&rdev->pm.mclk_lock);
981 	r = ttm_vm_ops->fault(vma, vmf);
982 	up_read(&rdev->pm.mclk_lock);
983 	return r;
984 }
985 
986 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
987 {
988 	struct drm_file *file_priv;
989 	struct radeon_device *rdev;
990 	int r;
991 
992 	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
993 		return -EINVAL;
994 	}
995 
996 	file_priv = filp->private_data;
997 	rdev = file_priv->minor->dev->dev_private;
998 	if (rdev == NULL) {
999 		return -EINVAL;
1000 	}
1001 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
1002 	if (unlikely(r != 0)) {
1003 		return r;
1004 	}
1005 	if (unlikely(ttm_vm_ops == NULL)) {
1006 		ttm_vm_ops = vma->vm_ops;
1007 		radeon_ttm_vm_ops = *ttm_vm_ops;
1008 		radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
1009 	}
1010 	vma->vm_ops = &radeon_ttm_vm_ops;
1011 	return 0;
1012 }
1013 
1014 #if defined(CONFIG_DEBUG_FS)
1015 
1016 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1017 {
1018 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1019 	unsigned ttm_pl = *(int *)node->info_ent->data;
1020 	struct drm_device *dev = node->minor->dev;
1021 	struct radeon_device *rdev = dev->dev_private;
1022 	struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
1023 	int ret;
1024 	struct ttm_bo_global *glob = rdev->mman.bdev.glob;
1025 
1026 	spin_lock(&glob->lru_lock);
1027 	ret = drm_mm_dump_table(m, mm);
1028 	spin_unlock(&glob->lru_lock);
1029 	return ret;
1030 }
1031 
1032 static int ttm_pl_vram = TTM_PL_VRAM;
1033 static int ttm_pl_tt = TTM_PL_TT;
1034 
1035 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1036 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1037 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1038 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1039 #ifdef CONFIG_SWIOTLB
1040 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1041 #endif
1042 };
1043 
1044 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1045 {
1046 	struct radeon_device *rdev = inode->i_private;
1047 	i_size_write(inode, rdev->mc.mc_vram_size);
1048 	filep->private_data = inode->i_private;
1049 	return 0;
1050 }
1051 
1052 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1053 				    size_t size, loff_t *pos)
1054 {
1055 	struct radeon_device *rdev = f->private_data;
1056 	ssize_t result = 0;
1057 	int r;
1058 
1059 	if (size & 0x3 || *pos & 0x3)
1060 		return -EINVAL;
1061 
1062 	while (size) {
1063 		unsigned long flags;
1064 		uint32_t value;
1065 
1066 		if (*pos >= rdev->mc.mc_vram_size)
1067 			return result;
1068 
1069 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1070 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1071 		if (rdev->family >= CHIP_CEDAR)
1072 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1073 		value = RREG32(RADEON_MM_DATA);
1074 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1075 
1076 		r = put_user(value, (uint32_t *)buf);
1077 		if (r)
1078 			return r;
1079 
1080 		result += 4;
1081 		buf += 4;
1082 		*pos += 4;
1083 		size -= 4;
1084 	}
1085 
1086 	return result;
1087 }
1088 
1089 static const struct file_operations radeon_ttm_vram_fops = {
1090 	.owner = THIS_MODULE,
1091 	.open = radeon_ttm_vram_open,
1092 	.read = radeon_ttm_vram_read,
1093 	.llseek = default_llseek
1094 };
1095 
1096 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1097 {
1098 	struct radeon_device *rdev = inode->i_private;
1099 	i_size_write(inode, rdev->mc.gtt_size);
1100 	filep->private_data = inode->i_private;
1101 	return 0;
1102 }
1103 
1104 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1105 				   size_t size, loff_t *pos)
1106 {
1107 	struct radeon_device *rdev = f->private_data;
1108 	ssize_t result = 0;
1109 	int r;
1110 
1111 	while (size) {
1112 		loff_t p = *pos / PAGE_SIZE;
1113 		unsigned off = *pos & ~PAGE_MASK;
1114 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1115 		struct page *page;
1116 		void *ptr;
1117 
1118 		if (p >= rdev->gart.num_cpu_pages)
1119 			return result;
1120 
1121 		page = rdev->gart.pages[p];
1122 		if (page) {
1123 			ptr = kmap(page);
1124 			ptr += off;
1125 
1126 			r = copy_to_user(buf, ptr, cur_size);
1127 			kunmap(rdev->gart.pages[p]);
1128 		} else
1129 			r = clear_user(buf, cur_size);
1130 
1131 		if (r)
1132 			return -EFAULT;
1133 
1134 		result += cur_size;
1135 		buf += cur_size;
1136 		*pos += cur_size;
1137 		size -= cur_size;
1138 	}
1139 
1140 	return result;
1141 }
1142 
1143 static const struct file_operations radeon_ttm_gtt_fops = {
1144 	.owner = THIS_MODULE,
1145 	.open = radeon_ttm_gtt_open,
1146 	.read = radeon_ttm_gtt_read,
1147 	.llseek = default_llseek
1148 };
1149 
1150 #endif
1151 
1152 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1153 {
1154 #if defined(CONFIG_DEBUG_FS)
1155 	unsigned count;
1156 
1157 	struct drm_minor *minor = rdev->ddev->primary;
1158 	struct dentry *ent, *root = minor->debugfs_root;
1159 
1160 	ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
1161 				  rdev, &radeon_ttm_vram_fops);
1162 	if (IS_ERR(ent))
1163 		return PTR_ERR(ent);
1164 	rdev->mman.vram = ent;
1165 
1166 	ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1167 				  rdev, &radeon_ttm_gtt_fops);
1168 	if (IS_ERR(ent))
1169 		return PTR_ERR(ent);
1170 	rdev->mman.gtt = ent;
1171 
1172 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1173 
1174 #ifdef CONFIG_SWIOTLB
1175 	if (!swiotlb_nr_tbl())
1176 		--count;
1177 #endif
1178 
1179 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1180 #else
1181 
1182 	return 0;
1183 #endif
1184 }
1185 
1186 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1187 {
1188 #if defined(CONFIG_DEBUG_FS)
1189 
1190 	debugfs_remove(rdev->mman.vram);
1191 	rdev->mman.vram = NULL;
1192 
1193 	debugfs_remove(rdev->mman.gtt);
1194 	rdev->mman.gtt = NULL;
1195 #endif
1196 }
1197