xref: /linux/drivers/gpu/drm/radeon/radeon_ttm.c (revision 0ad53fe3ae82443c74ff8cfd7bd13377cc1134a3)
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_device.h>
42 #include <drm/drm_file.h>
43 #include <drm/drm_prime.h>
44 #include <drm/radeon_drm.h>
45 #include <drm/ttm/ttm_bo_api.h>
46 #include <drm/ttm/ttm_bo_driver.h>
47 #include <drm/ttm/ttm_placement.h>
48 #include <drm/ttm/ttm_range_manager.h>
49 
50 #include "radeon_reg.h"
51 #include "radeon.h"
52 #include "radeon_ttm.h"
53 
54 static void radeon_ttm_debugfs_init(struct radeon_device *rdev);
55 
56 static int radeon_ttm_tt_bind(struct ttm_device *bdev, struct ttm_tt *ttm,
57 			      struct ttm_resource *bo_mem);
58 static void radeon_ttm_tt_unbind(struct ttm_device *bdev, struct ttm_tt *ttm);
59 
60 struct radeon_device *radeon_get_rdev(struct ttm_device *bdev)
61 {
62 	struct radeon_mman *mman;
63 	struct radeon_device *rdev;
64 
65 	mman = container_of(bdev, struct radeon_mman, bdev);
66 	rdev = container_of(mman, struct radeon_device, mman);
67 	return rdev;
68 }
69 
70 static int radeon_ttm_init_vram(struct radeon_device *rdev)
71 {
72 	return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_VRAM,
73 				  false, rdev->mc.real_vram_size >> PAGE_SHIFT);
74 }
75 
76 static int radeon_ttm_init_gtt(struct radeon_device *rdev)
77 {
78 	return ttm_range_man_init(&rdev->mman.bdev, TTM_PL_TT,
79 				  true, rdev->mc.gtt_size >> PAGE_SHIFT);
80 }
81 
82 static void radeon_evict_flags(struct ttm_buffer_object *bo,
83 				struct ttm_placement *placement)
84 {
85 	static const struct ttm_place placements = {
86 		.fpfn = 0,
87 		.lpfn = 0,
88 		.mem_type = TTM_PL_SYSTEM,
89 		.flags = 0
90 	};
91 
92 	struct radeon_bo *rbo;
93 
94 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
95 		placement->placement = &placements;
96 		placement->busy_placement = &placements;
97 		placement->num_placement = 1;
98 		placement->num_busy_placement = 1;
99 		return;
100 	}
101 	rbo = container_of(bo, struct radeon_bo, tbo);
102 	switch (bo->resource->mem_type) {
103 	case TTM_PL_VRAM:
104 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
105 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
106 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
107 			 bo->resource->start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
108 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
109 			int i;
110 
111 			/* Try evicting to the CPU inaccessible part of VRAM
112 			 * first, but only set GTT as busy placement, so this
113 			 * BO will be evicted to GTT rather than causing other
114 			 * BOs to be evicted from VRAM
115 			 */
116 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
117 							 RADEON_GEM_DOMAIN_GTT);
118 			rbo->placement.num_busy_placement = 0;
119 			for (i = 0; i < rbo->placement.num_placement; i++) {
120 				if (rbo->placements[i].mem_type == TTM_PL_VRAM) {
121 					if (rbo->placements[i].fpfn < fpfn)
122 						rbo->placements[i].fpfn = fpfn;
123 				} else {
124 					rbo->placement.busy_placement =
125 						&rbo->placements[i];
126 					rbo->placement.num_busy_placement = 1;
127 				}
128 			}
129 		} else
130 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
131 		break;
132 	case TTM_PL_TT:
133 	default:
134 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
135 	}
136 	*placement = rbo->placement;
137 }
138 
139 static int radeon_move_blit(struct ttm_buffer_object *bo,
140 			bool evict,
141 			struct ttm_resource *new_mem,
142 			struct ttm_resource *old_mem)
143 {
144 	struct radeon_device *rdev;
145 	uint64_t old_start, new_start;
146 	struct radeon_fence *fence;
147 	unsigned num_pages;
148 	int r, ridx;
149 
150 	rdev = radeon_get_rdev(bo->bdev);
151 	ridx = radeon_copy_ring_index(rdev);
152 	old_start = (u64)old_mem->start << PAGE_SHIFT;
153 	new_start = (u64)new_mem->start << PAGE_SHIFT;
154 
155 	switch (old_mem->mem_type) {
156 	case TTM_PL_VRAM:
157 		old_start += rdev->mc.vram_start;
158 		break;
159 	case TTM_PL_TT:
160 		old_start += rdev->mc.gtt_start;
161 		break;
162 	default:
163 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
164 		return -EINVAL;
165 	}
166 	switch (new_mem->mem_type) {
167 	case TTM_PL_VRAM:
168 		new_start += rdev->mc.vram_start;
169 		break;
170 	case TTM_PL_TT:
171 		new_start += rdev->mc.gtt_start;
172 		break;
173 	default:
174 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
175 		return -EINVAL;
176 	}
177 	if (!rdev->ring[ridx].ready) {
178 		DRM_ERROR("Trying to move memory with ring turned off.\n");
179 		return -EINVAL;
180 	}
181 
182 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
183 
184 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
185 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
186 	if (IS_ERR(fence))
187 		return PTR_ERR(fence);
188 
189 	r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, false, new_mem);
190 	radeon_fence_unref(&fence);
191 	return r;
192 }
193 
194 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
195 			  struct ttm_operation_ctx *ctx,
196 			  struct ttm_resource *new_mem,
197 			  struct ttm_place *hop)
198 {
199 	struct ttm_resource *old_mem = bo->resource;
200 	struct radeon_device *rdev;
201 	struct radeon_bo *rbo;
202 	int r, old_type;
203 
204 	if (new_mem->mem_type == TTM_PL_TT) {
205 		r = radeon_ttm_tt_bind(bo->bdev, bo->ttm, new_mem);
206 		if (r)
207 			return r;
208 	}
209 
210 	r = ttm_bo_wait_ctx(bo, ctx);
211 	if (r)
212 		return r;
213 
214 	/* Can't move a pinned BO */
215 	rbo = container_of(bo, struct radeon_bo, tbo);
216 	if (WARN_ON_ONCE(rbo->tbo.pin_count > 0))
217 		return -EINVAL;
218 
219 	/* Save old type for statistics update */
220 	old_type = old_mem->mem_type;
221 
222 	rdev = radeon_get_rdev(bo->bdev);
223 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
224 		ttm_bo_move_null(bo, new_mem);
225 		goto out;
226 	}
227 	if (old_mem->mem_type == TTM_PL_SYSTEM &&
228 	    new_mem->mem_type == TTM_PL_TT) {
229 		ttm_bo_move_null(bo, new_mem);
230 		goto out;
231 	}
232 
233 	if (old_mem->mem_type == TTM_PL_TT &&
234 	    new_mem->mem_type == TTM_PL_SYSTEM) {
235 		radeon_ttm_tt_unbind(bo->bdev, bo->ttm);
236 		ttm_resource_free(bo, &bo->resource);
237 		ttm_bo_assign_mem(bo, new_mem);
238 		goto out;
239 	}
240 	if (rdev->ring[radeon_copy_ring_index(rdev)].ready &&
241 	    rdev->asic->copy.copy != NULL) {
242 		if ((old_mem->mem_type == TTM_PL_SYSTEM &&
243 		     new_mem->mem_type == TTM_PL_VRAM) ||
244 		    (old_mem->mem_type == TTM_PL_VRAM &&
245 		     new_mem->mem_type == TTM_PL_SYSTEM)) {
246 			hop->fpfn = 0;
247 			hop->lpfn = 0;
248 			hop->mem_type = TTM_PL_TT;
249 			hop->flags = 0;
250 			return -EMULTIHOP;
251 		}
252 
253 		r = radeon_move_blit(bo, evict, new_mem, old_mem);
254 	} else {
255 		r = -ENODEV;
256 	}
257 
258 	if (r) {
259 		r = ttm_bo_move_memcpy(bo, ctx, new_mem);
260 		if (r)
261 			return r;
262 	}
263 
264 out:
265 	/* update statistics */
266 	atomic64_add(bo->base.size, &rdev->num_bytes_moved);
267 	radeon_bo_move_notify(bo, old_type, new_mem);
268 	return 0;
269 }
270 
271 static int radeon_ttm_io_mem_reserve(struct ttm_device *bdev, struct ttm_resource *mem)
272 {
273 	struct radeon_device *rdev = radeon_get_rdev(bdev);
274 	size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
275 
276 	switch (mem->mem_type) {
277 	case TTM_PL_SYSTEM:
278 		/* system memory */
279 		return 0;
280 	case TTM_PL_TT:
281 #if IS_ENABLED(CONFIG_AGP)
282 		if (rdev->flags & RADEON_IS_AGP) {
283 			/* RADEON_IS_AGP is set only if AGP is active */
284 			mem->bus.offset = (mem->start << PAGE_SHIFT) +
285 				rdev->mc.agp_base;
286 			mem->bus.is_iomem = !rdev->agp->cant_use_aperture;
287 			mem->bus.caching = ttm_write_combined;
288 		}
289 #endif
290 		break;
291 	case TTM_PL_VRAM:
292 		mem->bus.offset = mem->start << PAGE_SHIFT;
293 		/* check if it's visible */
294 		if ((mem->bus.offset + bus_size) > rdev->mc.visible_vram_size)
295 			return -EINVAL;
296 		mem->bus.offset += rdev->mc.aper_base;
297 		mem->bus.is_iomem = true;
298 		mem->bus.caching = ttm_write_combined;
299 #ifdef __alpha__
300 		/*
301 		 * Alpha: use bus.addr to hold the ioremap() return,
302 		 * so we can modify bus.base below.
303 		 */
304 		mem->bus.addr = ioremap_wc(mem->bus.offset, bus_size);
305 		if (!mem->bus.addr)
306 			return -ENOMEM;
307 
308 		/*
309 		 * Alpha: Use just the bus offset plus
310 		 * the hose/domain memory base for bus.base.
311 		 * It then can be used to build PTEs for VRAM
312 		 * access, as done in ttm_bo_vm_fault().
313 		 */
314 		mem->bus.offset = (mem->bus.offset & 0x0ffffffffUL) +
315 			rdev->hose->dense_mem_base;
316 #endif
317 		break;
318 	default:
319 		return -EINVAL;
320 	}
321 	return 0;
322 }
323 
324 /*
325  * TTM backend functions.
326  */
327 struct radeon_ttm_tt {
328 	struct ttm_tt		ttm;
329 	u64				offset;
330 
331 	uint64_t			userptr;
332 	struct mm_struct		*usermm;
333 	uint32_t			userflags;
334 	bool bound;
335 };
336 
337 /* prepare the sg table with the user pages */
338 static int radeon_ttm_tt_pin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm)
339 {
340 	struct radeon_device *rdev = radeon_get_rdev(bdev);
341 	struct radeon_ttm_tt *gtt = (void *)ttm;
342 	unsigned pinned = 0;
343 	int r;
344 
345 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
346 	enum dma_data_direction direction = write ?
347 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
348 
349 	if (current->mm != gtt->usermm)
350 		return -EPERM;
351 
352 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
353 		/* check that we only pin down anonymous memory
354 		   to prevent problems with writeback */
355 		unsigned long end = gtt->userptr + (u64)ttm->num_pages * PAGE_SIZE;
356 		struct vm_area_struct *vma;
357 		vma = find_vma(gtt->usermm, gtt->userptr);
358 		if (!vma || vma->vm_file || vma->vm_end < end)
359 			return -EPERM;
360 	}
361 
362 	do {
363 		unsigned num_pages = ttm->num_pages - pinned;
364 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
365 		struct page **pages = ttm->pages + pinned;
366 
367 		r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
368 				   pages, NULL);
369 		if (r < 0)
370 			goto release_pages;
371 
372 		pinned += r;
373 
374 	} while (pinned < ttm->num_pages);
375 
376 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
377 				      (u64)ttm->num_pages << PAGE_SHIFT,
378 				      GFP_KERNEL);
379 	if (r)
380 		goto release_sg;
381 
382 	r = dma_map_sgtable(rdev->dev, ttm->sg, direction, 0);
383 	if (r)
384 		goto release_sg;
385 
386 	drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
387 				       ttm->num_pages);
388 
389 	return 0;
390 
391 release_sg:
392 	kfree(ttm->sg);
393 
394 release_pages:
395 	release_pages(ttm->pages, pinned);
396 	return r;
397 }
398 
399 static void radeon_ttm_tt_unpin_userptr(struct ttm_device *bdev, struct ttm_tt *ttm)
400 {
401 	struct radeon_device *rdev = radeon_get_rdev(bdev);
402 	struct radeon_ttm_tt *gtt = (void *)ttm;
403 	struct sg_page_iter sg_iter;
404 
405 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
406 	enum dma_data_direction direction = write ?
407 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
408 
409 	/* double check that we don't free the table twice */
410 	if (!ttm->sg || !ttm->sg->sgl)
411 		return;
412 
413 	/* free the sg table and pages again */
414 	dma_unmap_sgtable(rdev->dev, ttm->sg, direction, 0);
415 
416 	for_each_sgtable_page(ttm->sg, &sg_iter, 0) {
417 		struct page *page = sg_page_iter_page(&sg_iter);
418 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
419 			set_page_dirty(page);
420 
421 		mark_page_accessed(page);
422 		put_page(page);
423 	}
424 
425 	sg_free_table(ttm->sg);
426 }
427 
428 static bool radeon_ttm_backend_is_bound(struct ttm_tt *ttm)
429 {
430 	struct radeon_ttm_tt *gtt = (void*)ttm;
431 
432 	return (gtt->bound);
433 }
434 
435 static int radeon_ttm_backend_bind(struct ttm_device *bdev,
436 				   struct ttm_tt *ttm,
437 				   struct ttm_resource *bo_mem)
438 {
439 	struct radeon_ttm_tt *gtt = (void*)ttm;
440 	struct radeon_device *rdev = radeon_get_rdev(bdev);
441 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
442 		RADEON_GART_PAGE_WRITE;
443 	int r;
444 
445 	if (gtt->bound)
446 		return 0;
447 
448 	if (gtt->userptr) {
449 		radeon_ttm_tt_pin_userptr(bdev, ttm);
450 		flags &= ~RADEON_GART_PAGE_WRITE;
451 	}
452 
453 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
454 	if (!ttm->num_pages) {
455 		WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
456 		     ttm->num_pages, bo_mem, ttm);
457 	}
458 	if (ttm->caching == ttm_cached)
459 		flags |= RADEON_GART_PAGE_SNOOP;
460 	r = radeon_gart_bind(rdev, gtt->offset, ttm->num_pages,
461 			     ttm->pages, gtt->ttm.dma_address, flags);
462 	if (r) {
463 		DRM_ERROR("failed to bind %u pages at 0x%08X\n",
464 			  ttm->num_pages, (unsigned)gtt->offset);
465 		return r;
466 	}
467 	gtt->bound = true;
468 	return 0;
469 }
470 
471 static void radeon_ttm_backend_unbind(struct ttm_device *bdev, struct ttm_tt *ttm)
472 {
473 	struct radeon_ttm_tt *gtt = (void *)ttm;
474 	struct radeon_device *rdev = radeon_get_rdev(bdev);
475 
476 	if (gtt->userptr)
477 		radeon_ttm_tt_unpin_userptr(bdev, ttm);
478 
479 	if (!gtt->bound)
480 		return;
481 
482 	radeon_gart_unbind(rdev, gtt->offset, ttm->num_pages);
483 
484 	gtt->bound = false;
485 }
486 
487 static void radeon_ttm_backend_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
488 {
489 	struct radeon_ttm_tt *gtt = (void *)ttm;
490 
491 	ttm_tt_fini(&gtt->ttm);
492 	kfree(gtt);
493 }
494 
495 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
496 					   uint32_t page_flags)
497 {
498 	struct radeon_ttm_tt *gtt;
499 	enum ttm_caching caching;
500 	struct radeon_bo *rbo;
501 #if IS_ENABLED(CONFIG_AGP)
502 	struct radeon_device *rdev = radeon_get_rdev(bo->bdev);
503 
504 	if (rdev->flags & RADEON_IS_AGP) {
505 		return ttm_agp_tt_create(bo, rdev->agp->bridge, page_flags);
506 	}
507 #endif
508 	rbo = container_of(bo, struct radeon_bo, tbo);
509 
510 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
511 	if (gtt == NULL) {
512 		return NULL;
513 	}
514 
515 	if (rbo->flags & RADEON_GEM_GTT_UC)
516 		caching = ttm_uncached;
517 	else if (rbo->flags & RADEON_GEM_GTT_WC)
518 		caching = ttm_write_combined;
519 	else
520 		caching = ttm_cached;
521 
522 	if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
523 		kfree(gtt);
524 		return NULL;
525 	}
526 	return &gtt->ttm;
527 }
528 
529 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct radeon_device *rdev,
530 						  struct ttm_tt *ttm)
531 {
532 #if IS_ENABLED(CONFIG_AGP)
533 	if (rdev->flags & RADEON_IS_AGP)
534 		return NULL;
535 #endif
536 
537 	if (!ttm)
538 		return NULL;
539 	return container_of(ttm, struct radeon_ttm_tt, ttm);
540 }
541 
542 static int radeon_ttm_tt_populate(struct ttm_device *bdev,
543 				  struct ttm_tt *ttm,
544 				  struct ttm_operation_ctx *ctx)
545 {
546 	struct radeon_device *rdev = radeon_get_rdev(bdev);
547 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
548 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
549 
550 	if (gtt && gtt->userptr) {
551 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
552 		if (!ttm->sg)
553 			return -ENOMEM;
554 
555 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
556 		return 0;
557 	}
558 
559 	if (slave && ttm->sg) {
560 		drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
561 					       ttm->num_pages);
562 		return 0;
563 	}
564 
565 	return ttm_pool_alloc(&rdev->mman.bdev.pool, ttm, ctx);
566 }
567 
568 static void radeon_ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
569 {
570 	struct radeon_device *rdev = radeon_get_rdev(bdev);
571 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
572 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
573 
574 	radeon_ttm_tt_unbind(bdev, ttm);
575 
576 	if (gtt && gtt->userptr) {
577 		kfree(ttm->sg);
578 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
579 		return;
580 	}
581 
582 	if (slave)
583 		return;
584 
585 	return ttm_pool_free(&rdev->mman.bdev.pool, ttm);
586 }
587 
588 int radeon_ttm_tt_set_userptr(struct radeon_device *rdev,
589 			      struct ttm_tt *ttm, uint64_t addr,
590 			      uint32_t flags)
591 {
592 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
593 
594 	if (gtt == NULL)
595 		return -EINVAL;
596 
597 	gtt->userptr = addr;
598 	gtt->usermm = current->mm;
599 	gtt->userflags = flags;
600 	return 0;
601 }
602 
603 bool radeon_ttm_tt_is_bound(struct ttm_device *bdev,
604 			    struct ttm_tt *ttm)
605 {
606 #if IS_ENABLED(CONFIG_AGP)
607 	struct radeon_device *rdev = radeon_get_rdev(bdev);
608 	if (rdev->flags & RADEON_IS_AGP)
609 		return ttm_agp_is_bound(ttm);
610 #endif
611 	return radeon_ttm_backend_is_bound(ttm);
612 }
613 
614 static int radeon_ttm_tt_bind(struct ttm_device *bdev,
615 			      struct ttm_tt *ttm,
616 			      struct ttm_resource *bo_mem)
617 {
618 #if IS_ENABLED(CONFIG_AGP)
619 	struct radeon_device *rdev = radeon_get_rdev(bdev);
620 #endif
621 
622 	if (!bo_mem)
623 		return -EINVAL;
624 #if IS_ENABLED(CONFIG_AGP)
625 	if (rdev->flags & RADEON_IS_AGP)
626 		return ttm_agp_bind(ttm, bo_mem);
627 #endif
628 
629 	return radeon_ttm_backend_bind(bdev, ttm, bo_mem);
630 }
631 
632 static void radeon_ttm_tt_unbind(struct ttm_device *bdev,
633 				 struct ttm_tt *ttm)
634 {
635 #if IS_ENABLED(CONFIG_AGP)
636 	struct radeon_device *rdev = radeon_get_rdev(bdev);
637 
638 	if (rdev->flags & RADEON_IS_AGP) {
639 		ttm_agp_unbind(ttm);
640 		return;
641 	}
642 #endif
643 	radeon_ttm_backend_unbind(bdev, ttm);
644 }
645 
646 static void radeon_ttm_tt_destroy(struct ttm_device *bdev,
647 				  struct ttm_tt *ttm)
648 {
649 #if IS_ENABLED(CONFIG_AGP)
650 	struct radeon_device *rdev = radeon_get_rdev(bdev);
651 
652 	if (rdev->flags & RADEON_IS_AGP) {
653 		ttm_agp_destroy(ttm);
654 		return;
655 	}
656 #endif
657 	radeon_ttm_backend_destroy(bdev, ttm);
658 }
659 
660 bool radeon_ttm_tt_has_userptr(struct radeon_device *rdev,
661 			       struct ttm_tt *ttm)
662 {
663 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
664 
665 	if (gtt == NULL)
666 		return false;
667 
668 	return !!gtt->userptr;
669 }
670 
671 bool radeon_ttm_tt_is_readonly(struct radeon_device *rdev,
672 			       struct ttm_tt *ttm)
673 {
674 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(rdev, ttm);
675 
676 	if (gtt == NULL)
677 		return false;
678 
679 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
680 }
681 
682 static void
683 radeon_bo_delete_mem_notify(struct ttm_buffer_object *bo)
684 {
685 	unsigned int old_type = TTM_PL_SYSTEM;
686 
687 	if (bo->resource)
688 		old_type = bo->resource->mem_type;
689 	radeon_bo_move_notify(bo, old_type, NULL);
690 }
691 
692 static struct ttm_device_funcs radeon_bo_driver = {
693 	.ttm_tt_create = &radeon_ttm_tt_create,
694 	.ttm_tt_populate = &radeon_ttm_tt_populate,
695 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
696 	.ttm_tt_destroy = &radeon_ttm_tt_destroy,
697 	.eviction_valuable = ttm_bo_eviction_valuable,
698 	.evict_flags = &radeon_evict_flags,
699 	.move = &radeon_bo_move,
700 	.delete_mem_notify = &radeon_bo_delete_mem_notify,
701 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
702 };
703 
704 int radeon_ttm_init(struct radeon_device *rdev)
705 {
706 	int r;
707 
708 	/* No others user of address space so set it to 0 */
709 	r = ttm_device_init(&rdev->mman.bdev, &radeon_bo_driver, rdev->dev,
710 			       rdev->ddev->anon_inode->i_mapping,
711 			       rdev->ddev->vma_offset_manager,
712 			       rdev->need_swiotlb,
713 			       dma_addressing_limited(&rdev->pdev->dev));
714 	if (r) {
715 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
716 		return r;
717 	}
718 	rdev->mman.initialized = true;
719 
720 	r = radeon_ttm_init_vram(rdev);
721 	if (r) {
722 		DRM_ERROR("Failed initializing VRAM heap.\n");
723 		return r;
724 	}
725 	/* Change the size here instead of the init above so only lpfn is affected */
726 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
727 
728 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
729 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
730 			     NULL, &rdev->stolen_vga_memory);
731 	if (r) {
732 		return r;
733 	}
734 	r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
735 	if (r)
736 		return r;
737 	r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
738 	radeon_bo_unreserve(rdev->stolen_vga_memory);
739 	if (r) {
740 		radeon_bo_unref(&rdev->stolen_vga_memory);
741 		return r;
742 	}
743 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
744 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
745 
746 	r = radeon_ttm_init_gtt(rdev);
747 	if (r) {
748 		DRM_ERROR("Failed initializing GTT heap.\n");
749 		return r;
750 	}
751 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
752 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
753 
754 	radeon_ttm_debugfs_init(rdev);
755 
756 	return 0;
757 }
758 
759 void radeon_ttm_fini(struct radeon_device *rdev)
760 {
761 	int r;
762 
763 	if (!rdev->mman.initialized)
764 		return;
765 
766 	if (rdev->stolen_vga_memory) {
767 		r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
768 		if (r == 0) {
769 			radeon_bo_unpin(rdev->stolen_vga_memory);
770 			radeon_bo_unreserve(rdev->stolen_vga_memory);
771 		}
772 		radeon_bo_unref(&rdev->stolen_vga_memory);
773 	}
774 	ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_VRAM);
775 	ttm_range_man_fini(&rdev->mman.bdev, TTM_PL_TT);
776 	ttm_device_fini(&rdev->mman.bdev);
777 	radeon_gart_fini(rdev);
778 	rdev->mman.initialized = false;
779 	DRM_INFO("radeon: ttm finalized\n");
780 }
781 
782 /* this should only be called at bootup or when userspace
783  * isn't running */
784 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
785 {
786 	struct ttm_resource_manager *man;
787 
788 	if (!rdev->mman.initialized)
789 		return;
790 
791 	man = ttm_manager_type(&rdev->mman.bdev, TTM_PL_VRAM);
792 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
793 	man->size = size >> PAGE_SHIFT;
794 }
795 
796 #if defined(CONFIG_DEBUG_FS)
797 
798 static int radeon_mm_vram_dump_table_show(struct seq_file *m, void *unused)
799 {
800 	struct radeon_device *rdev = (struct radeon_device *)m->private;
801 	struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev,
802 							    TTM_PL_VRAM);
803 	struct drm_printer p = drm_seq_file_printer(m);
804 
805 	man->func->debug(man, &p);
806 	return 0;
807 }
808 
809 static int radeon_ttm_page_pool_show(struct seq_file *m, void *data)
810 {
811 	struct radeon_device *rdev = (struct radeon_device *)m->private;
812 
813 	return ttm_pool_debugfs(&rdev->mman.bdev.pool, m);
814 }
815 
816 static int radeon_mm_gtt_dump_table_show(struct seq_file *m, void *unused)
817 {
818 	struct radeon_device *rdev = (struct radeon_device *)m->private;
819 	struct ttm_resource_manager *man = ttm_manager_type(&rdev->mman.bdev,
820 							    TTM_PL_TT);
821 	struct drm_printer p = drm_seq_file_printer(m);
822 
823 	man->func->debug(man, &p);
824 	return 0;
825 }
826 
827 DEFINE_SHOW_ATTRIBUTE(radeon_mm_vram_dump_table);
828 DEFINE_SHOW_ATTRIBUTE(radeon_mm_gtt_dump_table);
829 DEFINE_SHOW_ATTRIBUTE(radeon_ttm_page_pool);
830 
831 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
832 {
833 	struct radeon_device *rdev = inode->i_private;
834 	i_size_write(inode, rdev->mc.mc_vram_size);
835 	filep->private_data = inode->i_private;
836 	return 0;
837 }
838 
839 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
840 				    size_t size, loff_t *pos)
841 {
842 	struct radeon_device *rdev = f->private_data;
843 	ssize_t result = 0;
844 	int r;
845 
846 	if (size & 0x3 || *pos & 0x3)
847 		return -EINVAL;
848 
849 	while (size) {
850 		unsigned long flags;
851 		uint32_t value;
852 
853 		if (*pos >= rdev->mc.mc_vram_size)
854 			return result;
855 
856 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
857 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
858 		if (rdev->family >= CHIP_CEDAR)
859 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
860 		value = RREG32(RADEON_MM_DATA);
861 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
862 
863 		r = put_user(value, (uint32_t __user *)buf);
864 		if (r)
865 			return r;
866 
867 		result += 4;
868 		buf += 4;
869 		*pos += 4;
870 		size -= 4;
871 	}
872 
873 	return result;
874 }
875 
876 static const struct file_operations radeon_ttm_vram_fops = {
877 	.owner = THIS_MODULE,
878 	.open = radeon_ttm_vram_open,
879 	.read = radeon_ttm_vram_read,
880 	.llseek = default_llseek
881 };
882 
883 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
884 {
885 	struct radeon_device *rdev = inode->i_private;
886 	i_size_write(inode, rdev->mc.gtt_size);
887 	filep->private_data = inode->i_private;
888 	return 0;
889 }
890 
891 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
892 				   size_t size, loff_t *pos)
893 {
894 	struct radeon_device *rdev = f->private_data;
895 	ssize_t result = 0;
896 	int r;
897 
898 	while (size) {
899 		loff_t p = *pos / PAGE_SIZE;
900 		unsigned off = *pos & ~PAGE_MASK;
901 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
902 		struct page *page;
903 		void *ptr;
904 
905 		if (p >= rdev->gart.num_cpu_pages)
906 			return result;
907 
908 		page = rdev->gart.pages[p];
909 		if (page) {
910 			ptr = kmap(page);
911 			ptr += off;
912 
913 			r = copy_to_user(buf, ptr, cur_size);
914 			kunmap(rdev->gart.pages[p]);
915 		} else
916 			r = clear_user(buf, cur_size);
917 
918 		if (r)
919 			return -EFAULT;
920 
921 		result += cur_size;
922 		buf += cur_size;
923 		*pos += cur_size;
924 		size -= cur_size;
925 	}
926 
927 	return result;
928 }
929 
930 static const struct file_operations radeon_ttm_gtt_fops = {
931 	.owner = THIS_MODULE,
932 	.open = radeon_ttm_gtt_open,
933 	.read = radeon_ttm_gtt_read,
934 	.llseek = default_llseek
935 };
936 
937 #endif
938 
939 static void radeon_ttm_debugfs_init(struct radeon_device *rdev)
940 {
941 #if defined(CONFIG_DEBUG_FS)
942 	struct drm_minor *minor = rdev->ddev->primary;
943 	struct dentry *root = minor->debugfs_root;
944 
945 	debugfs_create_file("radeon_vram", 0444, root, rdev,
946 			    &radeon_ttm_vram_fops);
947 
948 	debugfs_create_file("radeon_gtt", 0444, root, rdev,
949 			    &radeon_ttm_gtt_fops);
950 
951 	debugfs_create_file("radeon_vram_mm", 0444, root, rdev,
952 			    &radeon_mm_vram_dump_table_fops);
953 	debugfs_create_file("radeon_gtt_mm", 0444, root, rdev,
954 			    &radeon_mm_gtt_dump_table_fops);
955 	debugfs_create_file("ttm_page_pool", 0444, root, rdev,
956 			    &radeon_ttm_page_pool_fops);
957 #endif
958 }
959