xref: /openbsd/sys/dev/pci/drm/radeon/radeon_ttm.c (revision 0608e4cb)
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 #ifdef __amd64__
57 #include "efifb.h"
58 #endif
59 
60 #if NEFIFB > 0
61 #include <machine/efifbvar.h>
62 #endif
63 
64 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
65 
66 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
67 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
68 
69 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
70 {
71 	struct radeon_mman *mman;
72 	struct radeon_device *rdev;
73 
74 	mman = container_of(bdev, struct radeon_mman, bdev);
75 	rdev = container_of(mman, struct radeon_device, mman);
76 	return rdev;
77 }
78 
79 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
80 				struct ttm_mem_type_manager *man)
81 {
82 	struct radeon_device *rdev;
83 
84 	rdev = radeon_get_rdev(bdev);
85 
86 	switch (type) {
87 	case TTM_PL_SYSTEM:
88 		/* System memory */
89 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
90 		man->available_caching = TTM_PL_MASK_CACHING;
91 		man->default_caching = TTM_PL_FLAG_CACHED;
92 		break;
93 	case TTM_PL_TT:
94 		man->func = &ttm_bo_manager_func;
95 		man->gpu_offset = rdev->mc.gtt_start;
96 		man->available_caching = TTM_PL_MASK_CACHING;
97 		man->default_caching = TTM_PL_FLAG_CACHED;
98 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
99 #if IS_ENABLED(CONFIG_AGP)
100 		if (rdev->flags & RADEON_IS_AGP) {
101 			if (!rdev->ddev->agp) {
102 				DRM_ERROR("AGP is not enabled for memory type %u\n",
103 					  (unsigned)type);
104 				return -EINVAL;
105 			}
106 			if (!rdev->ddev->agp->cant_use_aperture)
107 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
108 			man->available_caching = TTM_PL_FLAG_UNCACHED |
109 						 TTM_PL_FLAG_WC;
110 			man->default_caching = TTM_PL_FLAG_WC;
111 		}
112 #endif
113 		break;
114 	case TTM_PL_VRAM:
115 		/* "On-card" video ram */
116 		man->func = &ttm_bo_manager_func;
117 		man->gpu_offset = rdev->mc.vram_start;
118 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
119 			     TTM_MEMTYPE_FLAG_MAPPABLE;
120 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
121 		man->default_caching = TTM_PL_FLAG_WC;
122 		break;
123 	default:
124 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
125 		return -EINVAL;
126 	}
127 	return 0;
128 }
129 
130 static void radeon_evict_flags(struct ttm_buffer_object *bo,
131 				struct ttm_placement *placement)
132 {
133 	static const struct ttm_place placements = {
134 		.fpfn = 0,
135 		.lpfn = 0,
136 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
137 	};
138 
139 	struct radeon_bo *rbo;
140 
141 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
142 		placement->placement = &placements;
143 		placement->busy_placement = &placements;
144 		placement->num_placement = 1;
145 		placement->num_busy_placement = 1;
146 		return;
147 	}
148 	rbo = container_of(bo, struct radeon_bo, tbo);
149 	switch (bo->mem.mem_type) {
150 	case TTM_PL_VRAM:
151 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
152 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
153 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
154 			 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
155 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
156 			int i;
157 
158 			/* Try evicting to the CPU inaccessible part of VRAM
159 			 * first, but only set GTT as busy placement, so this
160 			 * BO will be evicted to GTT rather than causing other
161 			 * BOs to be evicted from VRAM
162 			 */
163 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
164 							 RADEON_GEM_DOMAIN_GTT);
165 			rbo->placement.num_busy_placement = 0;
166 			for (i = 0; i < rbo->placement.num_placement; i++) {
167 				if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
168 					if (rbo->placements[i].fpfn < fpfn)
169 						rbo->placements[i].fpfn = fpfn;
170 				} else {
171 					rbo->placement.busy_placement =
172 						&rbo->placements[i];
173 					rbo->placement.num_busy_placement = 1;
174 				}
175 			}
176 		} else
177 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
178 		break;
179 	case TTM_PL_TT:
180 	default:
181 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
182 	}
183 	*placement = rbo->placement;
184 }
185 
186 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
187 {
188 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
189 	struct drm_file *file_priv = (void *)filp;
190 
191 	if (radeon_ttm_tt_has_userptr(bo->ttm))
192 		return -EPERM;
193 	return drm_vma_node_verify_access(&rbo->tbo.base.vma_node, file_priv);
194 }
195 
196 static void radeon_move_null(struct ttm_buffer_object *bo,
197 			     struct ttm_mem_reg *new_mem)
198 {
199 	struct ttm_mem_reg *old_mem = &bo->mem;
200 
201 	BUG_ON(old_mem->mm_node != NULL);
202 	*old_mem = *new_mem;
203 	new_mem->mm_node = NULL;
204 }
205 
206 static int radeon_move_blit(struct ttm_buffer_object *bo,
207 			bool evict, bool no_wait_gpu,
208 			struct ttm_mem_reg *new_mem,
209 			struct ttm_mem_reg *old_mem)
210 {
211 	struct radeon_device *rdev;
212 	uint64_t old_start, new_start;
213 	struct radeon_fence *fence;
214 	unsigned num_pages;
215 	int r, ridx;
216 
217 	rdev = radeon_get_rdev(bo->bdev);
218 	ridx = radeon_copy_ring_index(rdev);
219 	old_start = (u64)old_mem->start << PAGE_SHIFT;
220 	new_start = (u64)new_mem->start << PAGE_SHIFT;
221 
222 	switch (old_mem->mem_type) {
223 	case TTM_PL_VRAM:
224 		old_start += rdev->mc.vram_start;
225 		break;
226 	case TTM_PL_TT:
227 		old_start += rdev->mc.gtt_start;
228 		break;
229 	default:
230 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
231 		return -EINVAL;
232 	}
233 	switch (new_mem->mem_type) {
234 	case TTM_PL_VRAM:
235 		new_start += rdev->mc.vram_start;
236 		break;
237 	case TTM_PL_TT:
238 		new_start += rdev->mc.gtt_start;
239 		break;
240 	default:
241 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
242 		return -EINVAL;
243 	}
244 	if (!rdev->ring[ridx].ready) {
245 		DRM_ERROR("Trying to move memory with ring turned off.\n");
246 		return -EINVAL;
247 	}
248 
249 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
250 
251 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
252 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->base.resv);
253 	if (IS_ERR(fence))
254 		return PTR_ERR(fence);
255 
256 	r = ttm_bo_move_accel_cleanup(bo, &fence->base, evict, new_mem);
257 	radeon_fence_unref(&fence);
258 	return r;
259 }
260 
261 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
262 				bool evict, bool interruptible,
263 				bool no_wait_gpu,
264 				struct ttm_mem_reg *new_mem)
265 {
266 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
267 	struct ttm_mem_reg *old_mem = &bo->mem;
268 	struct ttm_mem_reg tmp_mem;
269 	struct ttm_place placements;
270 	struct ttm_placement placement;
271 	int r;
272 
273 	tmp_mem = *new_mem;
274 	tmp_mem.mm_node = NULL;
275 	placement.num_placement = 1;
276 	placement.placement = &placements;
277 	placement.num_busy_placement = 1;
278 	placement.busy_placement = &placements;
279 	placements.fpfn = 0;
280 	placements.lpfn = 0;
281 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
282 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
283 	if (unlikely(r)) {
284 		return r;
285 	}
286 
287 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
288 	if (unlikely(r)) {
289 		goto out_cleanup;
290 	}
291 
292 	r = ttm_tt_bind(bo->ttm, &tmp_mem, &ctx);
293 	if (unlikely(r)) {
294 		goto out_cleanup;
295 	}
296 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
297 	if (unlikely(r)) {
298 		goto out_cleanup;
299 	}
300 	r = ttm_bo_move_ttm(bo, &ctx, new_mem);
301 out_cleanup:
302 	ttm_bo_mem_put(bo, &tmp_mem);
303 	return r;
304 }
305 
306 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
307 				bool evict, bool interruptible,
308 				bool no_wait_gpu,
309 				struct ttm_mem_reg *new_mem)
310 {
311 	struct ttm_operation_ctx ctx = { interruptible, no_wait_gpu };
312 	struct ttm_mem_reg *old_mem = &bo->mem;
313 	struct ttm_mem_reg tmp_mem;
314 	struct ttm_placement placement;
315 	struct ttm_place placements;
316 	int r;
317 
318 	tmp_mem = *new_mem;
319 	tmp_mem.mm_node = NULL;
320 	placement.num_placement = 1;
321 	placement.placement = &placements;
322 	placement.num_busy_placement = 1;
323 	placement.busy_placement = &placements;
324 	placements.fpfn = 0;
325 	placements.lpfn = 0;
326 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
327 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem, &ctx);
328 	if (unlikely(r)) {
329 		return r;
330 	}
331 	r = ttm_bo_move_ttm(bo, &ctx, &tmp_mem);
332 	if (unlikely(r)) {
333 		goto out_cleanup;
334 	}
335 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
336 	if (unlikely(r)) {
337 		goto out_cleanup;
338 	}
339 out_cleanup:
340 	ttm_bo_mem_put(bo, &tmp_mem);
341 	return r;
342 }
343 
344 static int radeon_bo_move(struct ttm_buffer_object *bo, bool evict,
345 			  struct ttm_operation_ctx *ctx,
346 			  struct ttm_mem_reg *new_mem)
347 {
348 	struct radeon_device *rdev;
349 	struct radeon_bo *rbo;
350 	struct ttm_mem_reg *old_mem = &bo->mem;
351 	int r;
352 
353 	r = ttm_bo_wait(bo, ctx->interruptible, ctx->no_wait_gpu);
354 	if (r)
355 		return r;
356 
357 	/* Can't move a pinned BO */
358 	rbo = container_of(bo, struct radeon_bo, tbo);
359 	if (WARN_ON_ONCE(rbo->pin_count > 0))
360 		return -EINVAL;
361 
362 	rdev = radeon_get_rdev(bo->bdev);
363 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
364 		radeon_move_null(bo, new_mem);
365 		return 0;
366 	}
367 	if ((old_mem->mem_type == TTM_PL_TT &&
368 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
369 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
370 	     new_mem->mem_type == TTM_PL_TT)) {
371 		/* bind is enough */
372 		radeon_move_null(bo, new_mem);
373 		return 0;
374 	}
375 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
376 	    rdev->asic->copy.copy == NULL) {
377 		/* use memcpy */
378 		goto memcpy;
379 	}
380 
381 	if (old_mem->mem_type == TTM_PL_VRAM &&
382 	    new_mem->mem_type == TTM_PL_SYSTEM) {
383 		r = radeon_move_vram_ram(bo, evict, ctx->interruptible,
384 					ctx->no_wait_gpu, new_mem);
385 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
386 		   new_mem->mem_type == TTM_PL_VRAM) {
387 		r = radeon_move_ram_vram(bo, evict, ctx->interruptible,
388 					    ctx->no_wait_gpu, new_mem);
389 	} else {
390 		r = radeon_move_blit(bo, evict, ctx->no_wait_gpu,
391 				     new_mem, old_mem);
392 	}
393 
394 	if (r) {
395 memcpy:
396 		r = ttm_bo_move_memcpy(bo, ctx, new_mem);
397 		if (r) {
398 			return r;
399 		}
400 	}
401 
402 	/* update statistics */
403 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
404 	return 0;
405 }
406 
407 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
408 {
409 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
410 	struct radeon_device *rdev = radeon_get_rdev(bdev);
411 
412 	mem->bus.addr = NULL;
413 	mem->bus.offset = 0;
414 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
415 	mem->bus.base = 0;
416 	mem->bus.is_iomem = false;
417 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
418 		return -EINVAL;
419 	switch (mem->mem_type) {
420 	case TTM_PL_SYSTEM:
421 		/* system memory */
422 		return 0;
423 	case TTM_PL_TT:
424 #if IS_ENABLED(CONFIG_AGP)
425 		if (rdev->flags & RADEON_IS_AGP) {
426 			/* RADEON_IS_AGP is set only if AGP is active */
427 			mem->bus.offset = mem->start << PAGE_SHIFT;
428 			mem->bus.base = rdev->mc.agp_base;
429 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
430 		}
431 #endif
432 		break;
433 	case TTM_PL_VRAM:
434 		mem->bus.offset = mem->start << PAGE_SHIFT;
435 		/* check if it's visible */
436 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
437 			return -EINVAL;
438 		mem->bus.base = rdev->mc.aper_base;
439 		mem->bus.is_iomem = true;
440 #ifdef __alpha__
441 		/*
442 		 * Alpha: use bus.addr to hold the ioremap() return,
443 		 * so we can modify bus.base below.
444 		 */
445 		if (mem->placement & TTM_PL_FLAG_WC)
446 			mem->bus.addr =
447 				ioremap_wc(mem->bus.base + mem->bus.offset,
448 					   mem->bus.size);
449 		else
450 			mem->bus.addr =
451 				ioremap(mem->bus.base + mem->bus.offset,
452 						mem->bus.size);
453 		if (!mem->bus.addr)
454 			return -ENOMEM;
455 
456 		/*
457 		 * Alpha: Use just the bus offset plus
458 		 * the hose/domain memory base for bus.base.
459 		 * It then can be used to build PTEs for VRAM
460 		 * access, as done in ttm_bo_vm_fault().
461 		 */
462 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
463 			rdev->ddev->hose->dense_mem_base;
464 #endif
465 		break;
466 	default:
467 		return -EINVAL;
468 	}
469 	return 0;
470 }
471 
472 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
473 {
474 }
475 
476 /*
477  * TTM backend functions.
478  */
479 struct radeon_ttm_tt {
480 	struct ttm_dma_tt		ttm;
481 	struct radeon_device		*rdev;
482 	u64				offset;
483 
484 	uint64_t			userptr;
485 	struct mm_struct		*usermm;
486 	uint32_t			userflags;
487 };
488 
489 /* prepare the sg table with the user pages */
490 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
491 {
492 	STUB();
493 	return -ENOSYS;
494 #ifdef notyet
495 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
496 	struct radeon_ttm_tt *gtt = (void *)ttm;
497 	unsigned pinned = 0, nents;
498 	int r;
499 
500 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
501 	enum dma_data_direction direction = write ?
502 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
503 
504 	if (current->mm != gtt->usermm)
505 		return -EPERM;
506 
507 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
508 		/* check that we only pin down anonymous memory
509 		   to prevent problems with writeback */
510 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
511 		struct vm_area_struct *vma;
512 		vma = find_vma(gtt->usermm, gtt->userptr);
513 		if (!vma || vma->vm_file || vma->vm_end < end)
514 			return -EPERM;
515 	}
516 
517 	do {
518 		unsigned num_pages = ttm->num_pages - pinned;
519 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
520 		struct vm_page **pages = ttm->pages + pinned;
521 
522 		r = get_user_pages(userptr, num_pages, write ? FOLL_WRITE : 0,
523 				   pages, NULL);
524 		if (r < 0)
525 			goto release_pages;
526 
527 		pinned += r;
528 
529 	} while (pinned < ttm->num_pages);
530 
531 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
532 				      ttm->num_pages << PAGE_SHIFT,
533 				      GFP_KERNEL);
534 	if (r)
535 		goto release_sg;
536 
537 	r = -ENOMEM;
538 	nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
539 	if (nents == 0)
540 		goto release_sg;
541 
542 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
543 					 gtt->ttm.dma_address, ttm->num_pages);
544 
545 	return 0;
546 
547 release_sg:
548 	kfree(ttm->sg);
549 
550 release_pages:
551 	release_pages(ttm->pages, pinned);
552 	return r;
553 #endif
554 }
555 
556 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
557 {
558 	STUB();
559 #ifdef notyet
560 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
561 	struct radeon_ttm_tt *gtt = (void *)ttm;
562 	struct sg_page_iter sg_iter;
563 
564 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
565 	enum dma_data_direction direction = write ?
566 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
567 
568 	/* double check that we don't free the table twice */
569 	if (!ttm->sg->sgl)
570 		return;
571 
572 	/* free the sg table and pages again */
573 	dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
574 
575 	for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
576 		struct vm_page *page = sg_page_iter_page(&sg_iter);
577 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
578 			set_page_dirty(page);
579 
580 		mark_page_accessed(page);
581 		put_page(page);
582 	}
583 
584 	sg_free_table(ttm->sg);
585 #endif
586 }
587 
588 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
589 				   struct ttm_mem_reg *bo_mem)
590 {
591 	struct radeon_ttm_tt *gtt = (void*)ttm;
592 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
593 		RADEON_GART_PAGE_WRITE;
594 	int r;
595 
596 	if (gtt->userptr) {
597 		radeon_ttm_tt_pin_userptr(ttm);
598 		flags &= ~RADEON_GART_PAGE_WRITE;
599 	}
600 
601 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
602 	if (!ttm->num_pages) {
603 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
604 		     ttm->num_pages, bo_mem, ttm);
605 	}
606 	if (ttm->caching_state == tt_cached)
607 		flags |= RADEON_GART_PAGE_SNOOP;
608 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
609 			     ttm->pages, gtt->ttm.dma_address, flags);
610 	if (r) {
611 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
612 			  ttm->num_pages, (unsigned)gtt->offset);
613 		return r;
614 	}
615 	return 0;
616 }
617 
618 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
619 {
620 	struct radeon_ttm_tt *gtt = (void *)ttm;
621 
622 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
623 
624 	if (gtt->userptr)
625 		radeon_ttm_tt_unpin_userptr(ttm);
626 
627 	return 0;
628 }
629 
630 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
631 {
632 	struct radeon_ttm_tt *gtt = (void *)ttm;
633 
634 	ttm_dma_tt_fini(&gtt->ttm);
635 	kfree(gtt);
636 }
637 
638 static struct ttm_backend_func radeon_backend_func = {
639 	.bind = &radeon_ttm_backend_bind,
640 	.unbind = &radeon_ttm_backend_unbind,
641 	.destroy = &radeon_ttm_backend_destroy,
642 };
643 
644 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_buffer_object *bo,
645 					   uint32_t page_flags)
646 {
647 	struct radeon_device *rdev;
648 	struct radeon_ttm_tt *gtt;
649 
650 	rdev = radeon_get_rdev(bo->bdev);
651 #if IS_ENABLED(CONFIG_AGP)
652 	if (rdev->flags & RADEON_IS_AGP) {
653 		return ttm_agp_tt_create(bo, rdev->ddev->agp,
654 					 page_flags);
655 	}
656 #endif
657 
658 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
659 	if (gtt == NULL) {
660 		return NULL;
661 	}
662 	gtt->ttm.ttm.func = &radeon_backend_func;
663 	gtt->rdev = rdev;
664 	if (ttm_dma_tt_init(&gtt->ttm, bo, page_flags)) {
665 		kfree(gtt);
666 		return NULL;
667 	}
668 	return &gtt->ttm.ttm;
669 }
670 
671 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
672 {
673 	if (!ttm || ttm->func != &radeon_backend_func)
674 		return NULL;
675 	return (struct radeon_ttm_tt *)ttm;
676 }
677 
678 static int radeon_ttm_tt_populate(struct ttm_tt *ttm,
679 			struct ttm_operation_ctx *ctx)
680 {
681 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
682 	struct radeon_device *rdev;
683 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
684 
685 	if (gtt && gtt->userptr) {
686 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
687 		if (!ttm->sg)
688 			return -ENOMEM;
689 
690 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
691 		ttm->state = tt_unbound;
692 		return 0;
693 	}
694 
695 	if (slave && ttm->sg) {
696 #ifdef notyet
697 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
698 						 gtt->ttm.dma_address, ttm->num_pages);
699 #endif
700 		ttm->state = tt_unbound;
701 		return 0;
702 	}
703 
704 	rdev = radeon_get_rdev(ttm->bdev);
705 #if IS_ENABLED(CONFIG_AGP)
706 	if (rdev->flags & RADEON_IS_AGP) {
707 		return ttm_agp_tt_populate(ttm, ctx);
708 	}
709 #endif
710 
711 #ifdef CONFIG_SWIOTLB
712 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
713 		return ttm_dma_populate(&gtt->ttm, rdev->dev, ctx);
714 	}
715 #endif
716 
717 	return ttm_populate_and_map_pages(rdev->dev, &gtt->ttm, ctx);
718 }
719 
720 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
721 {
722 	struct radeon_device *rdev;
723 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
724 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
725 
726 	if (gtt && gtt->userptr) {
727 		kfree(ttm->sg);
728 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
729 		return;
730 	}
731 
732 	if (slave)
733 		return;
734 
735 	rdev = radeon_get_rdev(ttm->bdev);
736 #if IS_ENABLED(CONFIG_AGP)
737 	if (rdev->flags & RADEON_IS_AGP) {
738 		ttm_agp_tt_unpopulate(ttm);
739 		return;
740 	}
741 #endif
742 
743 #ifdef CONFIG_SWIOTLB
744 	if (rdev->need_swiotlb && swiotlb_nr_tbl()) {
745 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
746 		return;
747 	}
748 #endif
749 
750 	ttm_unmap_and_unpopulate_pages(rdev->dev, &gtt->ttm);
751 }
752 
753 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
754 			      uint32_t flags)
755 {
756 	STUB();
757 	return -ENOSYS;
758 #if 0
759 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
760 
761 	if (gtt == NULL)
762 		return -EINVAL;
763 
764 	gtt->userptr = addr;
765 	gtt->usermm = current->mm;
766 	gtt->userflags = flags;
767 	return 0;
768 #endif
769 }
770 
771 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
772 {
773 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
774 
775 	if (gtt == NULL)
776 		return false;
777 
778 	return !!gtt->userptr;
779 }
780 
781 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
782 {
783 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
784 
785 	if (gtt == NULL)
786 		return false;
787 
788 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
789 }
790 
791 static struct ttm_bo_driver radeon_bo_driver = {
792 	.ttm_tt_create = &radeon_ttm_tt_create,
793 	.ttm_tt_populate = &radeon_ttm_tt_populate,
794 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
795 	.init_mem_type = &radeon_init_mem_type,
796 	.eviction_valuable = ttm_bo_eviction_valuable,
797 	.evict_flags = &radeon_evict_flags,
798 	.move = &radeon_bo_move,
799 	.verify_access = &radeon_verify_access,
800 	.move_notify = &radeon_bo_move_notify,
801 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
802 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
803 	.io_mem_free = &radeon_ttm_io_mem_free,
804 };
805 
806 int radeon_ttm_init(struct radeon_device *rdev)
807 {
808 	int r;
809 	unsigned long stolen_size = 0;
810 
811 #if NEFIFB > 0
812 	stolen_size = efifb_stolen();
813 #endif
814 	if (stolen_size == 0)
815 		stolen_size = 256 * 1024;
816 
817 	/* No others user of address space so set it to 0 */
818 #ifdef notyet
819 	r = ttm_bo_device_init(&rdev->mman.bdev,
820 			       &radeon_bo_driver,
821 			       rdev->ddev->anon_inode->i_mapping,
822 			       rdev->ddev->vma_offset_manager,
823 			       dma_addressing_limited(&rdev->pdev->dev));
824 #else
825 	r = ttm_bo_device_init(&rdev->mman.bdev,
826 			       &radeon_bo_driver,
827 			       /*rdev->ddev->anon_inode->i_mapping*/ NULL,
828 			       rdev->ddev->vma_offset_manager,
829 			       dma_addressing_limited(&rdev->pdev->dev));
830 #endif
831 	if (r) {
832 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
833 		return r;
834 	}
835 	rdev->mman.bdev.iot = rdev->iot;
836 	rdev->mman.bdev.memt = rdev->memt;
837 	rdev->mman.bdev.dmat = rdev->dmat;
838 	rdev->mman.initialized = true;
839 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
840 				rdev->mc.real_vram_size >> PAGE_SHIFT);
841 	if (r) {
842 		DRM_ERROR("Failed initializing VRAM heap.\n");
843 		return r;
844 	}
845 	/* Change the size here instead of the init above so only lpfn is affected */
846 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
847 
848 #ifdef __sparc64__
849 	r = radeon_bo_create(rdev, rdev->fb_offset, PAGE_SIZE, true,
850 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
851 			     NULL, &rdev->stolen_vga_memory);
852 #else
853 	r = radeon_bo_create(rdev, stolen_size, PAGE_SIZE, true,
854 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
855 			     NULL, &rdev->stolen_vga_memory);
856 #endif
857 	if (r) {
858 		return r;
859 	}
860 	r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
861 	if (r)
862 		return r;
863 	r = radeon_bo_pin(rdev->stolen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
864 	radeon_bo_unreserve(rdev->stolen_vga_memory);
865 	if (r) {
866 		radeon_bo_unref(&rdev->stolen_vga_memory);
867 		return r;
868 	}
869 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
870 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
871 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
872 				rdev->mc.gtt_size >> PAGE_SHIFT);
873 	if (r) {
874 		DRM_ERROR("Failed initializing GTT heap.\n");
875 		return r;
876 	}
877 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
878 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
879 
880 	r = radeon_ttm_debugfs_init(rdev);
881 	if (r) {
882 		DRM_ERROR("Failed to init debugfs\n");
883 		return r;
884 	}
885 	return 0;
886 }
887 
888 void radeon_ttm_fini(struct radeon_device *rdev)
889 {
890 	int r;
891 
892 	if (!rdev->mman.initialized)
893 		return;
894 	radeon_ttm_debugfs_fini(rdev);
895 	if (rdev->stolen_vga_memory) {
896 		r = radeon_bo_reserve(rdev->stolen_vga_memory, false);
897 		if (r == 0) {
898 			radeon_bo_unpin(rdev->stolen_vga_memory);
899 			radeon_bo_unreserve(rdev->stolen_vga_memory);
900 		}
901 		radeon_bo_unref(&rdev->stolen_vga_memory);
902 	}
903 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
904 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
905 	ttm_bo_device_release(&rdev->mman.bdev);
906 	radeon_gart_fini(rdev);
907 	rdev->mman.initialized = false;
908 	DRM_INFO("radeon: ttm finalized\n");
909 }
910 
911 /* this should only be called at bootup or when userspace
912  * isn't running */
913 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
914 {
915 	struct ttm_mem_type_manager *man;
916 
917 	if (!rdev->mman.initialized)
918 		return;
919 
920 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
921 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
922 	man->size = size >> PAGE_SHIFT;
923 }
924 
925 #ifdef __linux__
926 
927 static vm_fault_t radeon_ttm_fault(struct vm_fault *vmf)
928 {
929 	struct ttm_buffer_object *bo;
930 	struct radeon_device *rdev;
931 	vm_fault_t ret;
932 
933 	bo = (struct ttm_buffer_object *)vmf->vma->vm_private_data;
934 	if (bo == NULL)
935 		return VM_FAULT_NOPAGE;
936 
937 	rdev = radeon_get_rdev(bo->bdev);
938 	down_read(&rdev->pm.mclk_lock);
939 	ret = ttm_bo_vm_fault(vmf);
940 	up_read(&rdev->pm.mclk_lock);
941 	return ret;
942 }
943 
944 static struct vm_operations_struct radeon_ttm_vm_ops = {
945 	.fault = radeon_ttm_fault,
946 	.open = ttm_bo_vm_open,
947 	.close = ttm_bo_vm_close,
948 	.access = ttm_bo_vm_access
949 };
950 
951 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
952 {
953 	int r;
954 	struct drm_file *file_priv = filp->private_data;
955 	struct radeon_device *rdev = file_priv->minor->dev->dev_private;
956 
957 	if (rdev == NULL)
958 		return -EINVAL;
959 
960 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
961 	if (unlikely(r != 0))
962 		return r;
963 
964 	vma->vm_ops = &radeon_ttm_vm_ops;
965 	return 0;
966 }
967 #else
968 
969 static struct uvm_pagerops radeon_ttm_vm_ops;
970 static const struct uvm_pagerops *ttm_vm_ops = NULL;
971 
972 static int
973 radeon_ttm_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, vm_page_t *pps,
974     int npages, int centeridx, vm_fault_t fault_type,
975     vm_prot_t access_type, int flags)
976 {
977 	struct drm_gem_object *bo;
978 	struct radeon_device *rdev;
979 	int r;
980 
981 	bo = (struct drm_gem_object *)ufi->entry->object.uvm_obj;
982 	rdev = bo->dev->dev_private;
983 	down_read(&rdev->pm.mclk_lock);
984 	r = ttm_vm_ops->pgo_fault(ufi, vaddr, pps, npages, centeridx,
985 				  fault_type, access_type, flags);
986 	up_read(&rdev->pm.mclk_lock);
987 	return r;
988 }
989 
990 struct uvm_object *
991 radeon_mmap(struct file *filp, vm_prot_t accessprot, voff_t off,
992 	    vsize_t size)
993 {
994 	struct drm_file *file_priv = (void *)filp;
995 	struct radeon_device *rdev = file_priv->minor->dev->dev_private;
996 	struct uvm_object *uobj;
997 
998 	if (rdev == NULL)
999 		return NULL;
1000 
1001 	if (unlikely(off < DRM_FILE_PAGE_OFFSET))
1002 		return NULL;
1003 
1004 	uobj = ttm_bo_mmap(filp, off, size, &rdev->mman.bdev);
1005 	if (unlikely(uobj == NULL)) {
1006 		return NULL;
1007 	}
1008 	if (unlikely(ttm_vm_ops == NULL)) {
1009 		ttm_vm_ops = uobj->pgops;
1010 		radeon_ttm_vm_ops = *ttm_vm_ops;
1011 		radeon_ttm_vm_ops.pgo_fault = &radeon_ttm_fault;
1012 	}
1013 	uobj->pgops = &radeon_ttm_vm_ops;
1014 	return uobj;
1015 }
1016 #endif
1017 
1018 #if defined(CONFIG_DEBUG_FS)
1019 
1020 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1021 {
1022 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1023 	unsigned ttm_pl = *(int*)node->info_ent->data;
1024 	struct drm_device *dev = node->minor->dev;
1025 	struct radeon_device *rdev = dev->dev_private;
1026 	struct ttm_mem_type_manager *man = &rdev->mman.bdev.man[ttm_pl];
1027 	struct drm_printer p = drm_seq_file_printer(m);
1028 
1029 	man->func->debug(man, &p);
1030 	return 0;
1031 }
1032 
1033 
1034 static int ttm_pl_vram = TTM_PL_VRAM;
1035 static int ttm_pl_tt = TTM_PL_TT;
1036 
1037 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1038 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1039 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1040 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1041 #ifdef CONFIG_SWIOTLB
1042 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1043 #endif
1044 };
1045 
1046 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1047 {
1048 	struct radeon_device *rdev = inode->i_private;
1049 	i_size_write(inode, rdev->mc.mc_vram_size);
1050 	filep->private_data = inode->i_private;
1051 	return 0;
1052 }
1053 
1054 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1055 				    size_t size, loff_t *pos)
1056 {
1057 	struct radeon_device *rdev = f->private_data;
1058 	ssize_t result = 0;
1059 	int r;
1060 
1061 	if (size & 0x3 || *pos & 0x3)
1062 		return -EINVAL;
1063 
1064 	while (size) {
1065 		unsigned long flags;
1066 		uint32_t value;
1067 
1068 		if (*pos >= rdev->mc.mc_vram_size)
1069 			return result;
1070 
1071 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1072 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1073 		if (rdev->family >= CHIP_CEDAR)
1074 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1075 		value = RREG32(RADEON_MM_DATA);
1076 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1077 
1078 		r = put_user(value, (uint32_t *)buf);
1079 		if (r)
1080 			return r;
1081 
1082 		result += 4;
1083 		buf += 4;
1084 		*pos += 4;
1085 		size -= 4;
1086 	}
1087 
1088 	return result;
1089 }
1090 
1091 static const struct file_operations radeon_ttm_vram_fops = {
1092 	.owner = THIS_MODULE,
1093 	.open = radeon_ttm_vram_open,
1094 	.read = radeon_ttm_vram_read,
1095 	.llseek = default_llseek
1096 };
1097 
1098 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1099 {
1100 	struct radeon_device *rdev = inode->i_private;
1101 	i_size_write(inode, rdev->mc.gtt_size);
1102 	filep->private_data = inode->i_private;
1103 	return 0;
1104 }
1105 
1106 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1107 				   size_t size, loff_t *pos)
1108 {
1109 	struct radeon_device *rdev = f->private_data;
1110 	ssize_t result = 0;
1111 	int r;
1112 
1113 	while (size) {
1114 		loff_t p = *pos / PAGE_SIZE;
1115 		unsigned off = *pos & PAGE_MASK;
1116 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1117 		struct vm_page *page;
1118 		void *ptr;
1119 
1120 		if (p >= rdev->gart.num_cpu_pages)
1121 			return result;
1122 
1123 		page = rdev->gart.pages[p];
1124 		if (page) {
1125 			ptr = kmap(page);
1126 			ptr += off;
1127 
1128 			r = copy_to_user(buf, ptr, cur_size);
1129 			kunmap(ptr);
1130 		} else
1131 			r = clear_user(buf, cur_size);
1132 
1133 		if (r)
1134 			return -EFAULT;
1135 
1136 		result += cur_size;
1137 		buf += cur_size;
1138 		*pos += cur_size;
1139 		size -= cur_size;
1140 	}
1141 
1142 	return result;
1143 }
1144 
1145 static const struct file_operations radeon_ttm_gtt_fops = {
1146 	.owner = THIS_MODULE,
1147 	.open = radeon_ttm_gtt_open,
1148 	.read = radeon_ttm_gtt_read,
1149 	.llseek = default_llseek
1150 };
1151 
1152 #endif
1153 
1154 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1155 {
1156 #if defined(CONFIG_DEBUG_FS)
1157 	unsigned count;
1158 
1159 	struct drm_minor *minor = rdev->ddev->primary;
1160 	struct dentry *root = minor->debugfs_root;
1161 
1162 	rdev->mman.vram = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO,
1163 					      root, rdev,
1164 					      &radeon_ttm_vram_fops);
1165 
1166 	rdev->mman.gtt = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO,
1167 					     root, rdev, &radeon_ttm_gtt_fops);
1168 
1169 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1170 
1171 #ifdef CONFIG_SWIOTLB
1172 	if (!(rdev->need_swiotlb && swiotlb_nr_tbl()))
1173 		--count;
1174 #endif
1175 
1176 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1177 #else
1178 
1179 	return 0;
1180 #endif
1181 }
1182 
1183 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1184 {
1185 #if defined(CONFIG_DEBUG_FS)
1186 
1187 	debugfs_remove(rdev->mman.vram);
1188 	rdev->mman.vram = NULL;
1189 
1190 	debugfs_remove(rdev->mman.gtt);
1191 	rdev->mman.gtt = NULL;
1192 #endif
1193 }
1194