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 
33 #include <linux/dma-mapping.h>
34 #include <linux/pagemap.h>
35 #include <linux/pci.h>
36 #include <linux/seq_file.h>
37 #include <linux/slab.h>
38 #include <linux/swap.h>
39 #include <linux/swiotlb.h>
40 
41 #include <drm/drm_agpsupport.h>
42 #include <drm/drm_debugfs.h>
43 #include <drm/drm_device.h>
44 #include <drm/drm_file.h>
45 #include <drm/drm_prime.h>
46 #include <drm/radeon_drm.h>
47 #include <drm/ttm/ttm_bo_api.h>
48 #include <drm/ttm/ttm_bo_driver.h>
49 #include <drm/ttm/ttm_module.h>
50 #include <drm/ttm/ttm_page_alloc.h>
51 #include <drm/ttm/ttm_placement.h>
52 
53 #include "radeon_reg.h"
54 #include "radeon.h"
55 
56 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
57 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
58 
59 struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
60 {
61 	struct radeon_mman *mman;
62 	struct radeon_device *rdev;
63 
64 	mman = container_of(bdev, struct radeon_mman, bdev);
65 	rdev = container_of(mman, struct radeon_device, mman);
66 	return rdev;
67 }
68 
69 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
70 				struct ttm_mem_type_manager *man)
71 {
72 	struct radeon_device *rdev;
73 
74 	rdev = radeon_get_rdev(bdev);
75 
76 	switch (type) {
77 	case TTM_PL_SYSTEM:
78 		/* System memory */
79 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
80 		man->available_caching = TTM_PL_MASK_CACHING;
81 		man->default_caching = TTM_PL_FLAG_CACHED;
82 		break;
83 	case TTM_PL_TT:
84 		man->func = &ttm_bo_manager_func;
85 		man->available_caching = TTM_PL_MASK_CACHING;
86 		man->default_caching = TTM_PL_FLAG_CACHED;
87 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
88 #if IS_ENABLED(CONFIG_AGP)
89 		if (rdev->flags & RADEON_IS_AGP) {
90 			if (!rdev->ddev->agp) {
91 				DRM_ERROR("AGP is not enabled for memory type %u\n",
92 					  (unsigned)type);
93 				return -EINVAL;
94 			}
95 			if (!rdev->ddev->agp->cant_use_aperture)
96 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
97 			man->available_caching = TTM_PL_FLAG_UNCACHED |
98 						 TTM_PL_FLAG_WC;
99 			man->default_caching = TTM_PL_FLAG_WC;
100 		}
101 #endif
102 		break;
103 	case TTM_PL_VRAM:
104 		/* "On-card" video ram */
105 		man->func = &ttm_bo_manager_func;
106 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
107 			     TTM_MEMTYPE_FLAG_MAPPABLE;
108 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
109 		man->default_caching = TTM_PL_FLAG_WC;
110 		break;
111 	default:
112 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
113 		return -EINVAL;
114 	}
115 	return 0;
116 }
117 
118 static void radeon_evict_flags(struct ttm_buffer_object *bo,
119 				struct ttm_placement *placement)
120 {
121 	static const struct ttm_place placements = {
122 		.fpfn = 0,
123 		.lpfn = 0,
124 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
125 	};
126 
127 	struct radeon_bo *rbo;
128 
129 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
130 		placement->placement = &placements;
131 		placement->busy_placement = &placements;
132 		placement->num_placement = 1;
133 		placement->num_busy_placement = 1;
134 		return;
135 	}
136 	rbo = container_of(bo, struct radeon_bo, tbo);
137 	switch (bo->mem.mem_type) {
138 	case TTM_PL_VRAM:
139 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
140 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
141 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
142 			 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
143 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
144 			int i;
145 
146 			/* Try evicting to the CPU inaccessible part of VRAM
147 			 * first, but only set GTT as busy placement, so this
148 			 * BO will be evicted to GTT rather than causing other
149 			 * BOs to be evicted from VRAM
150 			 */
151 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
152 							 RADEON_GEM_DOMAIN_GTT);
153 			rbo->placement.num_busy_placement = 0;
154 			for (i = 0; i < rbo->placement.num_placement; i++) {
155 				if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
156 					if (rbo->placements[i].fpfn < fpfn)
157 						rbo->placements[i].fpfn = fpfn;
158 				} else {
159 					rbo->placement.busy_placement =
160 						&rbo->placements[i];
161 					rbo->placement.num_busy_placement = 1;
162 				}
163 			}
164 		} else
165 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
166 		break;
167 	case TTM_PL_TT:
168 	default:
169 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
170 	}
171 	*placement = rbo->placement;
172 }
173 
174 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
175 {
176 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
177 
178 	if (radeon_ttm_tt_has_userptr(bo->ttm))
179 		return -EPERM;
180 	return drm_vma_node_verify_access(&rbo->tbo.base.vma_node,
181 					  filp->private_data);
182 }
183 
184 static void radeon_move_null(struct ttm_buffer_object *bo,
185 			     struct ttm_mem_reg *new_mem)
186 {
187 	struct ttm_mem_reg *old_mem = &bo->mem;
188 
189 	BUG_ON(old_mem->mm_node != NULL);
190 	*old_mem = *new_mem;
191 	new_mem->mm_node = NULL;
192 }
193 
194 static int radeon_move_blit(struct ttm_buffer_object *bo,
195 			bool evict, bool no_wait_gpu,
196 			struct ttm_mem_reg *new_mem,
197 			struct ttm_mem_reg *old_mem)
198 {
199 	struct radeon_device *rdev;
200 	uint64_t old_start, new_start;
201 	struct radeon_fence *fence;
202 	unsigned num_pages;
203 	int r, ridx;
204 
205 	rdev = radeon_get_rdev(bo->bdev);
206 	ridx = radeon_copy_ring_index(rdev);
207 	old_start = (u64)old_mem->start << PAGE_SHIFT;
208 	new_start = (u64)new_mem->start << PAGE_SHIFT;
209 
210 	switch (old_mem->mem_type) {
211 	case TTM_PL_VRAM:
212 		old_start += rdev->mc.vram_start;
213 		break;
214 	case TTM_PL_TT:
215 		old_start += rdev->mc.gtt_start;
216 		break;
217 	default:
218 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
219 		return -EINVAL;
220 	}
221 	switch (new_mem->mem_type) {
222 	case TTM_PL_VRAM:
223 		new_start += rdev->mc.vram_start;
224 		break;
225 	case TTM_PL_TT:
226 		new_start += rdev->mc.gtt_start;
227 		break;
228 	default:
229 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
230 		return -EINVAL;
231 	}
232 	if (!rdev->ring[ridx].ready) {
233 		DRM_ERROR("Trying to move memory with ring turned off.\n");
234 		return -EINVAL;
235 	}
236 
237 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
238 
239 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
240 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
241 	if (IS_ERR(fence))
242 		return PTR_ERR(fence);
243 
244 	r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, new_mem);
245 	radeon_fence_unref(&fence);
246 	return r;
247 }
248 
249 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
250 				bool evict, bool interruptible,
251 				bool no_wait_gpu,
252 				struct ttm_mem_reg *new_mem)
253 {
254 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
255 	struct ttm_mem_reg *old_mem = &bo->mem;
256 	struct ttm_mem_reg tmp_mem;
257 	struct ttm_place placements;
258 	struct ttm_placement placement;
259 	int r;
260 
261 	tmp_mem = *new_mem;
262 	tmp_mem.mm_node = NULL;
263 	placement.num_placement = 1;
264 	placement.placement = &placements;
265 	placement.num_busy_placement = 1;
266 	placement.busy_placement = &placements;
267 	placements.fpfn = 0;
268 	placements.lpfn = 0;
269 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
270 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
271 	if (unlikely(r)) {
272 		return r;
273 	}
274 
275 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
276 	if (unlikely(r)) {
277 		goto out_cleanup;
278 	}
279 
280 	r = ttm_tt_bind(bo->ttm, &tmp_mem, &ctx);
281 	if (unlikely(r)) {
282 		goto out_cleanup;
283 	}
284 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
285 	if (unlikely(r)) {
286 		goto out_cleanup;
287 	}
288 	r = ttm_bo_move_ttm(bo, &ctx, new_mem);
289 out_cleanup:
290 	ttm_bo_mem_put(bo, &tmp_mem);
291 	return r;
292 }
293 
294 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
295 				bool evict, bool interruptible,
296 				bool no_wait_gpu,
297 				struct ttm_mem_reg *new_mem)
298 {
299 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
300 	struct ttm_mem_reg *old_mem = &bo->mem;
301 	struct ttm_mem_reg tmp_mem;
302 	struct ttm_placement placement;
303 	struct ttm_place placements;
304 	int r;
305 
306 	tmp_mem = *new_mem;
307 	tmp_mem.mm_node = NULL;
308 	placement.num_placement = 1;
309 	placement.placement = &placements;
310 	placement.num_busy_placement = 1;
311 	placement.busy_placement = &placements;
312 	placements.fpfn = 0;
313 	placements.lpfn = 0;
314 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
315 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
316 	if (unlikely(r)) {
317 		return r;
318 	}
319 	r = ttm_bo_move_ttm(bo, &ctx, &tmp_mem);
320 	if (unlikely(r)) {
321 		goto out_cleanup;
322 	}
323 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
324 	if (unlikely(r)) {
325 		goto out_cleanup;
326 	}
327 out_cleanup:
328 	ttm_bo_mem_put(bo, &tmp_mem);
329 	return r;
330 }
331 
332 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
333 			  struct ttm_operation_ctx *ctx,
334 			  struct ttm_mem_reg *new_mem)
335 {
336 	struct radeon_device *rdev;
337 	struct radeon_bo *rbo;
338 	struct ttm_mem_reg *old_mem = &bo->mem;
339 	int r;
340 
341 	r = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
342 	if (r)
343 		return r;
344 
345 	/* Can't move a pinned BO */
346 	rbo = container_of(bo, struct radeon_bo, tbo);
347 	if (WARN_ON_ONCE(rbo->pin_count > 0))
348 		return -EINVAL;
349 
350 	rdev = radeon_get_rdev(bo->bdev);
351 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
352 		radeon_move_null(bo, new_mem);
353 		return 0;
354 	}
355 	if ((old_mem->mem_type == TTM_PL_TT &&
356 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
357 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
358 	     new_mem->mem_type == TTM_PL_TT)) {
359 		/* bind is enough */
360 		radeon_move_null(bo, new_mem);
361 		return 0;
362 	}
363 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
364 	    rdev->asic->copy.copy == NULL) {
365 		/* use memcpy */
366 		goto memcpy;
367 	}
368 
369 	if (old_mem->mem_type == TTM_PL_VRAM &&
370 	    new_mem->mem_type == TTM_PL_SYSTEM) {
371 		r = radeon_move_vram_ram(bo, evict, ctx->interruptible,
372 					ctx->no_wait_gpu, new_mem);
373 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
374 		   new_mem->mem_type == TTM_PL_VRAM) {
375 		r = radeon_move_ram_vram(bo, evict, ctx->interruptible,
376 					    ctx->no_wait_gpu, new_mem);
377 	} else {
378 		r = radeon_move_blit(bo, evict, ctx->no_wait_gpu,
379 				     new_mem, old_mem);
380 	}
381 
382 	if (r) {
383 memcpy:
384 		r = ttm_bo_move_memcpy(bo, ctx, new_mem);
385 		if (r) {
386 			return r;
387 		}
388 	}
389 
390 	/* update statistics */
391 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
392 	return 0;
393 }
394 
395 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
396 {
397 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
398 	struct radeon_device *rdev = radeon_get_rdev(bdev);
399 
400 	mem->bus.addr = NULL;
401 	mem->bus.offset = 0;
402 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
403 	mem->bus.base = 0;
404 	mem->bus.is_iomem = false;
405 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
406 		return -EINVAL;
407 	switch (mem->mem_type) {
408 	case TTM_PL_SYSTEM:
409 		/* system memory */
410 		return 0;
411 	case TTM_PL_TT:
412 #if IS_ENABLED(CONFIG_AGP)
413 		if (rdev->flags & RADEON_IS_AGP) {
414 			/* RADEON_IS_AGP is set only if AGP is active */
415 			mem->bus.offset = mem->start << PAGE_SHIFT;
416 			mem->bus.base = rdev->mc.agp_base;
417 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
418 		}
419 #endif
420 		break;
421 	case TTM_PL_VRAM:
422 		mem->bus.offset = mem->start << PAGE_SHIFT;
423 		/* check if it's visible */
424 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
425 			return -EINVAL;
426 		mem->bus.base = rdev->mc.aper_base;
427 		mem->bus.is_iomem = true;
428 #ifdef __alpha__
429 		/*
430 		 * Alpha: use bus.addr to hold the ioremap() return,
431 		 * so we can modify bus.base below.
432 		 */
433 		if (mem->placement & TTM_PL_FLAG_WC)
434 			mem->bus.addr =
435 				ioremap_wc(mem->bus.base + mem->bus.offset,
436 					   mem->bus.size);
437 		else
438 			mem->bus.addr =
439 				ioremap(mem->bus.base + mem->bus.offset,
440 						mem->bus.size);
441 		if (!mem->bus.addr)
442 			return -ENOMEM;
443 
444 		/*
445 		 * Alpha: Use just the bus offset plus
446 		 * the hose/domain memory base for bus.base.
447 		 * It then can be used to build PTEs for VRAM
448 		 * access, as done in ttm_bo_vm_fault().
449 		 */
450 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
451 			rdev->ddev->hose->dense_mem_base;
452 #endif
453 		break;
454 	default:
455 		return -EINVAL;
456 	}
457 	return 0;
458 }
459 
460 /*
461  * TTM backend functions.
462  */
463 struct radeon_ttm_tt {
464 	struct ttm_dma_tt		ttm;
465 	struct radeon_device		*rdev;
466 	u64				offset;
467 
468 	uint64_t			userptr;
469 	struct mm_struct		*usermm;
470 	uint32_t			userflags;
471 };
472 
473 /* prepare the sg table with the user pages */
474 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
475 {
476 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
477 	struct radeon_ttm_tt *gtt = (void *)ttm;
478 	unsigned pinned = 0;
479 	int r;
480 
481 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
482 	enum dma_data_direction direction = write ?
483 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
484 
485 	if (current->mm != gtt->usermm)
486 		return -EPERM;
487 
488 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
489 		/* check that we only pin down anonymous memory
490 		   to prevent problems with writeback */
491 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
492 		struct vm_area_struct *vma;
493 		vma = find_vma(gtt->usermm, gtt->userptr);
494 		if (!vma || vma->vm_file || vma->vm_end < end)
495 			return -EPERM;
496 	}
497 
498 	do {
499 		unsigned num_pages = ttm->num_pages - pinned;
500 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
501 		struct page **pages = ttm->pages + pinned;
502 
503 		r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
504 				   pages, NULL);
505 		if (r < 0)
506 			goto release_pages;
507 
508 		pinned += r;
509 
510 	} while (pinned < ttm->num_pages);
511 
512 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
513 				      ttm->num_pages << PAGE_SHIFT,
514 				      GFP_KERNEL);
515 	if (r)
516 		goto release_sg;
517 
518 	r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
519 	if (r)
520 		goto release_sg;
521 
522 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
523 					 gtt->ttm.dma_address, ttm->num_pages);
524 
525 	return 0;
526 
527 release_sg:
528 	kfree(ttm->sg);
529 
530 release_pages:
531 	release_pages(ttm->pages, pinned);
532 	return r;
533 }
534 
535 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
536 {
537 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
538 	struct radeon_ttm_tt *gtt = (void *)ttm;
539 	struct sg_page_iter sg_iter;
540 
541 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
542 	enum dma_data_direction direction = write ?
543 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
544 
545 	/* double check that we don't free the table twice */
546 	if (!ttm->sg->sgl)
547 		return;
548 
549 	/* free the sg table and pages again */
550 	dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
551 
552 	for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
553 		struct page *page = sg_page_iter_page(&sg_iter);
554 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
555 			set_page_dirty(page);
556 
557 		mark_page_accessed(page);
558 		put_page(page);
559 	}
560 
561 	sg_free_table(ttm->sg);
562 }
563 
564 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
565 				   struct ttm_mem_reg *bo_mem)
566 {
567 	struct radeon_ttm_tt *gtt = (void*)ttm;
568 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
569 		RADEON_GART_PAGE_WRITE;
570 	int r;
571 
572 	if (gtt->userptr) {
573 		radeon_ttm_tt_pin_userptr(ttm);
574 		flags &= ~RADEON_GART_PAGE_WRITE;
575 	}
576 
577 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
578 	if (!ttm->num_pages) {
579 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
580 		     ttm->num_pages, bo_mem, ttm);
581 	}
582 	if (ttm->caching_state == tt_cached)
583 		flags |= RADEON_GART_PAGE_SNOOP;
584 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
585 			     ttm->pages, gtt->ttm.dma_address, flags);
586 	if (r) {
587 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
588 			  ttm->num_pages, (unsigned)gtt->offset);
589 		return r;
590 	}
591 	return 0;
592 }
593 
594 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
595 {
596 	struct radeon_ttm_tt *gtt = (void *)ttm;
597 
598 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
599 
600 	if (gtt->userptr)
601 		radeon_ttm_tt_unpin_userptr(ttm);
602 
603 	return 0;
604 }
605 
606 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
607 {
608 	struct radeon_ttm_tt *gtt = (void *)ttm;
609 
610 	ttm_dma_tt_fini(&gtt->ttm);
611 	kfree(gtt);
612 }
613 
614 static struct ttm_backend_func radeon_backend_func = {
615 	.bind = &radeon_ttm_backend_bind,
616 	.unbind = &radeon_ttm_backend_unbind,
617 	.destroy = &radeon_ttm_backend_destroy,
618 };
619 
620 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
621 					   uint32_t page_flags)
622 {
623 	struct radeon_device *rdev;
624 	struct radeon_ttm_tt *gtt;
625 
626 	rdev = radeon_get_rdev(bo->bdev);
627 #if IS_ENABLED(CONFIG_AGP)
628 	if (rdev->flags & RADEON_IS_AGP) {
629 		return ttm_agp_tt_create(bo, rdev->ddev->agp->bridge,
630 					 page_flags);
631 	}
632 #endif
633 
634 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
635 	if (gtt == NULL) {
636 		return NULL;
637 	}
638 	gtt->ttm.ttm.func = &radeon_backend_func;
639 	gtt->rdev = rdev;
640 	if (ttm_dma_tt_init(&gtt->ttm, bo, page_flags)) {
641 		kfree(gtt);
642 		return NULL;
643 	}
644 	return &gtt->ttm.ttm;
645 }
646 
647 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
648 {
649 	if (!ttm || ttm->func != &radeon_backend_func)
650 		return NULL;
651 	return (struct radeon_ttm_tt *)ttm;
652 }
653 
654 static int radeon_ttm_tt_populate(struct ttm_tt *ttm,
655 			struct ttm_operation_ctx *ctx)
656 {
657 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
658 	struct radeon_device *rdev;
659 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
660 
661 	if (gtt && gtt->userptr) {
662 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
663 		if (!ttm->sg)
664 			return -ENOMEM;
665 
666 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
667 		ttm->state = tt_unbound;
668 		return 0;
669 	}
670 
671 	if (slave && ttm->sg) {
672 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
673 						 gtt->ttm.dma_address, ttm->num_pages);
674 		ttm->state = tt_unbound;
675 		return 0;
676 	}
677 
678 	rdev = radeon_get_rdev(ttm->bdev);
679 #if IS_ENABLED(CONFIG_AGP)
680 	if (rdev->flags & RADEON_IS_AGP) {
681 		return ttm_agp_tt_populate(ttm, ctx);
682 	}
683 #endif
684 
685 #ifdef CONFIG_SWIOTLB
686 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
687 		return ttm_dma_populate(&gtt->ttm, rdev->dev, ctx);
688 	}
689 #endif
690 
691 	return ttm_populate_and_map_pages(rdev->dev, &gtt->ttm, ctx);
692 }
693 
694 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
695 {
696 	struct radeon_device *rdev;
697 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
698 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
699 
700 	if (gtt && gtt->userptr) {
701 		kfree(ttm->sg);
702 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
703 		return;
704 	}
705 
706 	if (slave)
707 		return;
708 
709 	rdev = radeon_get_rdev(ttm->bdev);
710 #if IS_ENABLED(CONFIG_AGP)
711 	if (rdev->flags & RADEON_IS_AGP) {
712 		ttm_agp_tt_unpopulate(ttm);
713 		return;
714 	}
715 #endif
716 
717 #ifdef CONFIG_SWIOTLB
718 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
719 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
720 		return;
721 	}
722 #endif
723 
724 	ttm_unmap_and_unpopulate_pages(rdev->dev, &gtt->ttm);
725 }
726 
727 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
728 			      uint32_t flags)
729 {
730 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
731 
732 	if (gtt == NULL)
733 		return -EINVAL;
734 
735 	gtt->userptr = addr;
736 	gtt->usermm = current->mm;
737 	gtt->userflags = flags;
738 	return 0;
739 }
740 
741 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
742 {
743 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
744 
745 	if (gtt == NULL)
746 		return false;
747 
748 	return !!gtt->userptr;
749 }
750 
751 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
752 {
753 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
754 
755 	if (gtt == NULL)
756 		return false;
757 
758 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
759 }
760 
761 static struct ttm_bo_driver radeon_bo_driver = {
762 	.ttm_tt_create = &radeon_ttm_tt_create,
763 	.ttm_tt_populate = &radeon_ttm_tt_populate,
764 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
765 	.init_mem_type = &radeon_init_mem_type,
766 	.eviction_valuable = ttm_bo_eviction_valuable,
767 	.evict_flags = &radeon_evict_flags,
768 	.move = &radeon_bo_move,
769 	.verify_access = &radeon_verify_access,
770 	.move_notify = &radeon_bo_move_notify,
771 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
772 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
773 };
774 
775 int radeon_ttm_init(struct radeon_device *rdev)
776 {
777 	int r;
778 
779 	/* No others user of address space so set it to 0 */
780 	r = ttm_bo_device_init(&rdev->mman.bdev,
781 			       &radeon_bo_driver,
782 			       rdev->ddev->anon_inode->i_mapping,
783 			       rdev->ddev->vma_offset_manager,
784 			       dma_addressing_limited(&rdev->pdev->dev));
785 	if (r) {
786 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
787 		return r;
788 	}
789 	rdev->mman.initialized = true;
790 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
791 				rdev->mc.real_vram_size >> PAGE_SHIFT);
792 	if (r) {
793 		DRM_ERROR("Failed initializing VRAM heap.\n");
794 		return r;
795 	}
796 	/* Change the size here instead of the init above so only lpfn is affected */
797 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
798 
799 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
800 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
801 			     NULL, &rdev->stolen_vga_memory);
802 	if (r) {
803 		return r;
804 	}
805 	r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
806 	if (r)
807 		return r;
808 	r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
809 	radeon_bo_unreserve(rdev->stolen_vga_memory);
810 	if (r) {
811 		radeon_bo_unref(&rdev->stolen_vga_memory);
812 		return r;
813 	}
814 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
815 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
816 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
817 				rdev->mc.gtt_size >> PAGE_SHIFT);
818 	if (r) {
819 		DRM_ERROR("Failed initializing GTT heap.\n");
820 		return r;
821 	}
822 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
823 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
824 
825 	r = radeon_ttm_debugfs_init(rdev);
826 	if (r) {
827 		DRM_ERROR("Failed to init debugfs\n");
828 		return r;
829 	}
830 	return 0;
831 }
832 
833 void radeon_ttm_fini(struct radeon_device *rdev)
834 {
835 	int r;
836 
837 	if (!rdev->mman.initialized)
838 		return;
839 	radeon_ttm_debugfs_fini(rdev);
840 	if (rdev->stolen_vga_memory) {
841 		r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
842 		if (r == 0) {
843 			radeon_bo_unpin(rdev->stolen_vga_memory);
844 			radeon_bo_unreserve(rdev->stolen_vga_memory);
845 		}
846 		radeon_bo_unref(&rdev->stolen_vga_memory);
847 	}
848 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
849 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
850 	ttm_bo_device_release(&rdev->mman.bdev);
851 	radeon_gart_fini(rdev);
852 	rdev->mman.initialized = false;
853 	DRM_INFO("radeon: ttm finalized\n");
854 }
855 
856 /* this should only be called at bootup or when userspace
857  * isn't running */
858 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
859 {
860 	struct ttm_mem_type_manager *man;
861 
862 	if (!rdev->mman.initialized)
863 		return;
864 
865 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
866 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
867 	man->size = size >> PAGE_SHIFT;
868 }
869 
870 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
871 {
872 	struct ttm_buffer_object *bo;
873 	struct radeon_device *rdev;
874 	vm_fault_t ret;
875 
876 	bo = (struct ttm_buffer_object *)vmf->vma->vm_private_data;
877 	if (bo == NULL)
878 		return VM_FAULT_NOPAGE;
879 
880 	rdev = radeon_get_rdev(bo->bdev);
881 	down_read(&rdev->pm.mclk_lock);
882 	ret = ttm_bo_vm_fault(vmf);
883 	up_read(&rdev->pm.mclk_lock);
884 	return ret;
885 }
886 
887 static struct vm_operations_struct radeon_ttm_vm_ops = {
888 	.fault = radeon_ttm_fault,
889 	.open = ttm_bo_vm_open,
890 	.close = ttm_bo_vm_close,
891 	.access = ttm_bo_vm_access
892 };
893 
894 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
895 {
896 	int r;
897 	struct drm_file *file_priv = filp->private_data;
898 	struct radeon_device *rdev = file_priv->minor->dev->dev_private;
899 
900 	if (rdev == NULL)
901 		return -EINVAL;
902 
903 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
904 	if (unlikely(r != 0))
905 		return r;
906 
907 	vma->vm_ops = &radeon_ttm_vm_ops;
908 	return 0;
909 }
910 
911 #if defined(CONFIG_DEBUG_FS)
912 
913 static int radeon_mm_dump_table(struct seq_file *m, void *data)
914 {
915 	struct drm_info_node *node = (struct drm_info_node *)m->private;
916 	unsigned ttm_pl = *(int*)node->info_ent->data;
917 	struct drm_device *dev = node->minor->dev;
918 	struct radeon_device *rdev = dev->dev_private;
919 	struct ttm_mem_type_manager *man = &rdev->mman.bdev.man[ttm_pl];
920 	struct drm_printer p = drm_seq_file_printer(m);
921 
922 	man->func->debug(man, &p);
923 	return 0;
924 }
925 
926 
927 static int ttm_pl_vram = TTM_PL_VRAM;
928 static int ttm_pl_tt = TTM_PL_TT;
929 
930 static struct drm_info_list radeon_ttm_debugfs_list[] = {
931 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
932 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
933 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
934 #ifdef CONFIG_SWIOTLB
935 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
936 #endif
937 };
938 
939 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
940 {
941 	struct radeon_device *rdev = inode->i_private;
942 	i_size_write(inode, rdev->mc.mc_vram_size);
943 	filep->private_data = inode->i_private;
944 	return 0;
945 }
946 
947 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
948 				    size_t size, loff_t *pos)
949 {
950 	struct radeon_device *rdev = f->private_data;
951 	ssize_t result = 0;
952 	int r;
953 
954 	if (size & 0x3 || *pos & 0x3)
955 		return -EINVAL;
956 
957 	while (size) {
958 		unsigned long flags;
959 		uint32_t value;
960 
961 		if (*pos >= rdev->mc.mc_vram_size)
962 			return result;
963 
964 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
965 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
966 		if (rdev->family >= CHIP_CEDAR)
967 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
968 		value = RREG32(RADEON_MM_DATA);
969 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
970 
971 		r = put_user(value, (uint32_t *)buf);
972 		if (r)
973 			return r;
974 
975 		result += 4;
976 		buf += 4;
977 		*pos += 4;
978 		size -= 4;
979 	}
980 
981 	return result;
982 }
983 
984 static const struct file_operations radeon_ttm_vram_fops = {
985 	.owner = THIS_MODULE,
986 	.open = radeon_ttm_vram_open,
987 	.read = radeon_ttm_vram_read,
988 	.llseek = default_llseek
989 };
990 
991 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
992 {
993 	struct radeon_device *rdev = inode->i_private;
994 	i_size_write(inode, rdev->mc.gtt_size);
995 	filep->private_data = inode->i_private;
996 	return 0;
997 }
998 
999 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1000 				   size_t size, loff_t *pos)
1001 {
1002 	struct radeon_device *rdev = f->private_data;
1003 	ssize_t result = 0;
1004 	int r;
1005 
1006 	while (size) {
1007 		loff_t p = *pos / PAGE_SIZE;
1008 		unsigned off = *pos & ~PAGE_MASK;
1009 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1010 		struct page *page;
1011 		void *ptr;
1012 
1013 		if (p >= rdev->gart.num_cpu_pages)
1014 			return result;
1015 
1016 		page = rdev->gart.pages[p];
1017 		if (page) {
1018 			ptr = kmap(page);
1019 			ptr += off;
1020 
1021 			r = copy_to_user(buf, ptr, cur_size);
1022 			kunmap(rdev->gart.pages[p]);
1023 		} else
1024 			r = clear_user(buf, cur_size);
1025 
1026 		if (r)
1027 			return -EFAULT;
1028 
1029 		result += cur_size;
1030 		buf += cur_size;
1031 		*pos += cur_size;
1032 		size -= cur_size;
1033 	}
1034 
1035 	return result;
1036 }
1037 
1038 static const struct file_operations radeon_ttm_gtt_fops = {
1039 	.owner = THIS_MODULE,
1040 	.open = radeon_ttm_gtt_open,
1041 	.read = radeon_ttm_gtt_read,
1042 	.llseek = default_llseek
1043 };
1044 
1045 #endif
1046 
1047 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1048 {
1049 #if defined(CONFIG_DEBUG_FS)
1050 	unsigned count;
1051 
1052 	struct drm_minor *minor = rdev->ddev->primary;
1053 	struct dentry *root = minor->debugfs_root;
1054 
1055 	rdev->mman.vram = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO,
1056 					      root, rdev,
1057 					      &radeon_ttm_vram_fops);
1058 
1059 	rdev->mman.gtt = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO,
1060 					     root, rdev, &radeon_ttm_gtt_fops);
1061 
1062 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1063 
1064 #ifdef CONFIG_SWIOTLB
1065 	if (!(rdev->need_swiotlb && swiotlb_nr_tbl()))
1066 		--count;
1067 #endif
1068 
1069 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1070 #else
1071 
1072 	return 0;
1073 #endif
1074 }
1075 
1076 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1077 {
1078 #if defined(CONFIG_DEBUG_FS)
1079 
1080 	debugfs_remove(rdev->mman.vram);
1081 	rdev->mman.vram = NULL;
1082 
1083 	debugfs_remove(rdev->mman.gtt);
1084 	rdev->mman.gtt = NULL;
1085 #endif
1086 }
1087