1 /*
2  * Copyright 2009 Jerome Glisse.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 /*
27  * Authors:
28  *    Jerome Glisse <glisse@freedesktop.org>
29  *    Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
30  *    Dave Airlie
31  */
32 #include <ttm/ttm_bo_api.h>
33 #include <ttm/ttm_bo_driver.h>
34 #include <ttm/ttm_placement.h>
35 #include <ttm/ttm_module.h>
36 #include <ttm/ttm_page_alloc.h>
37 #include <drm/drmP.h>
38 #include <drm/radeon_drm.h>
39 #include <linux/seq_file.h>
40 #include <linux/slab.h>
41 #include <linux/swiotlb.h>
42 #include <linux/swap.h>
43 #include <linux/pagemap.h>
44 #include <linux/debugfs.h>
45 #include "radeon_reg.h"
46 #include "radeon.h"
47 
48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
49 
50 static int radeon_ttm_debugfs_init(struct radeon_device *rdev);
51 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev);
52 
53 static struct radeon_device *radeon_get_rdev(struct ttm_bo_device *bdev)
54 {
55 	struct radeon_mman *mman;
56 	struct radeon_device *rdev;
57 
58 	mman = container_of(bdev, struct radeon_mman, bdev);
59 	rdev = container_of(mman, struct radeon_device, mman);
60 	return rdev;
61 }
62 
63 
64 /*
65  * Global memory.
66  */
67 static int radeon_ttm_mem_global_init(struct drm_global_reference *ref)
68 {
69 	return ttm_mem_global_init(ref->object);
70 }
71 
72 static void radeon_ttm_mem_global_release(struct drm_global_reference *ref)
73 {
74 	ttm_mem_global_release(ref->object);
75 }
76 
77 static int radeon_ttm_global_init(struct radeon_device *rdev)
78 {
79 	struct drm_global_reference *global_ref;
80 	int r;
81 
82 	rdev->mman.mem_global_referenced = false;
83 	global_ref = &rdev->mman.mem_global_ref;
84 	global_ref->global_type = DRM_GLOBAL_TTM_MEM;
85 	global_ref->size = sizeof(struct ttm_mem_global);
86 	global_ref->init = &radeon_ttm_mem_global_init;
87 	global_ref->release = &radeon_ttm_mem_global_release;
88 	r = drm_global_item_ref(global_ref);
89 	if (r != 0) {
90 		DRM_ERROR("Failed setting up TTM memory accounting "
91 			  "subsystem.\n");
92 		return r;
93 	}
94 
95 	rdev->mman.bo_global_ref.mem_glob =
96 		rdev->mman.mem_global_ref.object;
97 	global_ref = &rdev->mman.bo_global_ref.ref;
98 	global_ref->global_type = DRM_GLOBAL_TTM_BO;
99 	global_ref->size = sizeof(struct ttm_bo_global);
100 	global_ref->init = &ttm_bo_global_init;
101 	global_ref->release = &ttm_bo_global_release;
102 	r = drm_global_item_ref(global_ref);
103 	if (r != 0) {
104 		DRM_ERROR("Failed setting up TTM BO subsystem.\n");
105 		drm_global_item_unref(&rdev->mman.mem_global_ref);
106 		return r;
107 	}
108 
109 	rdev->mman.mem_global_referenced = true;
110 	return 0;
111 }
112 
113 static void radeon_ttm_global_fini(struct radeon_device *rdev)
114 {
115 	if (rdev->mman.mem_global_referenced) {
116 		drm_global_item_unref(&rdev->mman.bo_global_ref.ref);
117 		drm_global_item_unref(&rdev->mman.mem_global_ref);
118 		rdev->mman.mem_global_referenced = false;
119 	}
120 }
121 
122 static int radeon_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
123 {
124 	return 0;
125 }
126 
127 static int radeon_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
128 				struct ttm_mem_type_manager *man)
129 {
130 	struct radeon_device *rdev;
131 
132 	rdev = radeon_get_rdev(bdev);
133 
134 	switch (type) {
135 	case TTM_PL_SYSTEM:
136 		/* System memory */
137 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
138 		man->available_caching = TTM_PL_MASK_CACHING;
139 		man->default_caching = TTM_PL_FLAG_CACHED;
140 		break;
141 	case TTM_PL_TT:
142 		man->func = &ttm_bo_manager_func;
143 		man->gpu_offset = rdev->mc.gtt_start;
144 		man->available_caching = TTM_PL_MASK_CACHING;
145 		man->default_caching = TTM_PL_FLAG_CACHED;
146 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
147 #if IS_ENABLED(CONFIG_AGP)
148 		if (rdev->flags & RADEON_IS_AGP) {
149 			if (!rdev->ddev->agp) {
150 				DRM_ERROR("AGP is not enabled for memory type %u\n",
151 					  (unsigned)type);
152 				return -EINVAL;
153 			}
154 			if (!rdev->ddev->agp->cant_use_aperture)
155 				man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
156 			man->available_caching = TTM_PL_FLAG_UNCACHED |
157 						 TTM_PL_FLAG_WC;
158 			man->default_caching = TTM_PL_FLAG_WC;
159 		}
160 #endif
161 		break;
162 	case TTM_PL_VRAM:
163 		/* "On-card" video ram */
164 		man->func = &ttm_bo_manager_func;
165 		man->gpu_offset = rdev->mc.vram_start;
166 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
167 			     TTM_MEMTYPE_FLAG_MAPPABLE;
168 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
169 		man->default_caching = TTM_PL_FLAG_WC;
170 		break;
171 	default:
172 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
173 		return -EINVAL;
174 	}
175 	return 0;
176 }
177 
178 static void radeon_evict_flags(struct ttm_buffer_object *bo,
179 				struct ttm_placement *placement)
180 {
181 	static struct ttm_place placements = {
182 		.fpfn = 0,
183 		.lpfn = 0,
184 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
185 	};
186 
187 	struct radeon_bo *rbo;
188 
189 	if (!radeon_ttm_bo_is_radeon_bo(bo)) {
190 		placement->placement = &placements;
191 		placement->busy_placement = &placements;
192 		placement->num_placement = 1;
193 		placement->num_busy_placement = 1;
194 		return;
195 	}
196 	rbo = container_of(bo, struct radeon_bo, tbo);
197 	switch (bo->mem.mem_type) {
198 	case TTM_PL_VRAM:
199 		if (rbo->rdev->ring[radeon_copy_ring_index(rbo->rdev)].ready == false)
200 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
201 		else if (rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size &&
202 			 bo->mem.start < (rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT)) {
203 			unsigned fpfn = rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT;
204 			int i;
205 
206 			/* Try evicting to the CPU inaccessible part of VRAM
207 			 * first, but only set GTT as busy placement, so this
208 			 * BO will be evicted to GTT rather than causing other
209 			 * BOs to be evicted from VRAM
210 			 */
211 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_VRAM |
212 							 RADEON_GEM_DOMAIN_GTT);
213 			rbo->placement.num_busy_placement = 0;
214 			for (i = 0; i < rbo->placement.num_placement; i++) {
215 				if (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) {
216 					if (rbo->placements[0].fpfn < fpfn)
217 						rbo->placements[0].fpfn = fpfn;
218 				} else {
219 					rbo->placement.busy_placement =
220 						&rbo->placements[i];
221 					rbo->placement.num_busy_placement = 1;
222 				}
223 			}
224 		} else
225 			radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_GTT);
226 		break;
227 	case TTM_PL_TT:
228 	default:
229 		radeon_ttm_placement_from_domain(rbo, RADEON_GEM_DOMAIN_CPU);
230 	}
231 	*placement = rbo->placement;
232 }
233 
234 static int radeon_verify_access(struct ttm_buffer_object *bo, struct file *filp)
235 {
236 	struct radeon_bo *rbo = container_of(bo, struct radeon_bo, tbo);
237 
238 	if (radeon_ttm_tt_has_userptr(bo->ttm))
239 		return -EPERM;
240 	return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
241 }
242 
243 static void radeon_move_null(struct ttm_buffer_object *bo,
244 			     struct ttm_mem_reg *new_mem)
245 {
246 	struct ttm_mem_reg *old_mem = &bo->mem;
247 
248 	BUG_ON(old_mem->mm_node != NULL);
249 	*old_mem = *new_mem;
250 	new_mem->mm_node = NULL;
251 }
252 
253 static int radeon_move_blit(struct ttm_buffer_object *bo,
254 			bool evict, bool no_wait_gpu,
255 			struct ttm_mem_reg *new_mem,
256 			struct ttm_mem_reg *old_mem)
257 {
258 	struct radeon_device *rdev;
259 	uint64_t old_start, new_start;
260 	struct radeon_fence *fence;
261 	unsigned num_pages;
262 	int r, ridx;
263 
264 	rdev = radeon_get_rdev(bo->bdev);
265 	ridx = radeon_copy_ring_index(rdev);
266 	old_start = old_mem->start << PAGE_SHIFT;
267 	new_start = new_mem->start << PAGE_SHIFT;
268 
269 	switch (old_mem->mem_type) {
270 	case TTM_PL_VRAM:
271 		old_start += rdev->mc.vram_start;
272 		break;
273 	case TTM_PL_TT:
274 		old_start += rdev->mc.gtt_start;
275 		break;
276 	default:
277 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
278 		return -EINVAL;
279 	}
280 	switch (new_mem->mem_type) {
281 	case TTM_PL_VRAM:
282 		new_start += rdev->mc.vram_start;
283 		break;
284 	case TTM_PL_TT:
285 		new_start += rdev->mc.gtt_start;
286 		break;
287 	default:
288 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
289 		return -EINVAL;
290 	}
291 	if (!rdev->ring[ridx].ready) {
292 		DRM_ERROR("Trying to move memory with ring turned off.\n");
293 		return -EINVAL;
294 	}
295 
296 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
297 
298 	num_pages = new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE);
299 	fence = radeon_copy(rdev, old_start, new_start, num_pages, bo->resv);
300 	if (IS_ERR(fence))
301 		return PTR_ERR(fence);
302 
303 	r = ttm_bo_move_accel_cleanup(bo, &fence->base,
304 				      evict, no_wait_gpu, new_mem);
305 	radeon_fence_unref(&fence);
306 	return r;
307 }
308 
309 static int radeon_move_vram_ram(struct ttm_buffer_object *bo,
310 				bool evict, bool interruptible,
311 				bool no_wait_gpu,
312 				struct ttm_mem_reg *new_mem)
313 {
314 	struct radeon_device *rdev;
315 	struct ttm_mem_reg *old_mem = &bo->mem;
316 	struct ttm_mem_reg tmp_mem;
317 	struct ttm_place placements;
318 	struct ttm_placement placement;
319 	int r;
320 
321 	rdev = radeon_get_rdev(bo->bdev);
322 	tmp_mem = *new_mem;
323 	tmp_mem.mm_node = NULL;
324 	placement.num_placement = 1;
325 	placement.placement = &placements;
326 	placement.num_busy_placement = 1;
327 	placement.busy_placement = &placements;
328 	placements.fpfn = 0;
329 	placements.lpfn = 0;
330 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
331 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
332 			     interruptible, no_wait_gpu);
333 	if (unlikely(r)) {
334 		return r;
335 	}
336 
337 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
338 	if (unlikely(r)) {
339 		goto out_cleanup;
340 	}
341 
342 	r = ttm_tt_bind(bo->ttm, &tmp_mem);
343 	if (unlikely(r)) {
344 		goto out_cleanup;
345 	}
346 	r = radeon_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
347 	if (unlikely(r)) {
348 		goto out_cleanup;
349 	}
350 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
351 out_cleanup:
352 	ttm_bo_mem_put(bo, &tmp_mem);
353 	return r;
354 }
355 
356 static int radeon_move_ram_vram(struct ttm_buffer_object *bo,
357 				bool evict, bool interruptible,
358 				bool no_wait_gpu,
359 				struct ttm_mem_reg *new_mem)
360 {
361 	struct radeon_device *rdev;
362 	struct ttm_mem_reg *old_mem = &bo->mem;
363 	struct ttm_mem_reg tmp_mem;
364 	struct ttm_placement placement;
365 	struct ttm_place placements;
366 	int r;
367 
368 	rdev = radeon_get_rdev(bo->bdev);
369 	tmp_mem = *new_mem;
370 	tmp_mem.mm_node = NULL;
371 	placement.num_placement = 1;
372 	placement.placement = &placements;
373 	placement.num_busy_placement = 1;
374 	placement.busy_placement = &placements;
375 	placements.fpfn = 0;
376 	placements.lpfn = 0;
377 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
378 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
379 			     interruptible, no_wait_gpu);
380 	if (unlikely(r)) {
381 		return r;
382 	}
383 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
384 	if (unlikely(r)) {
385 		goto out_cleanup;
386 	}
387 	r = radeon_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
388 	if (unlikely(r)) {
389 		goto out_cleanup;
390 	}
391 out_cleanup:
392 	ttm_bo_mem_put(bo, &tmp_mem);
393 	return r;
394 }
395 
396 static int radeon_bo_move(struct ttm_buffer_object *bo,
397 			bool evict, bool interruptible,
398 			bool no_wait_gpu,
399 			struct ttm_mem_reg *new_mem)
400 {
401 	struct radeon_device *rdev;
402 	struct radeon_bo *rbo;
403 	struct ttm_mem_reg *old_mem = &bo->mem;
404 	int r;
405 
406 	/* Can't move a pinned BO */
407 	rbo = container_of(bo, struct radeon_bo, tbo);
408 	if (WARN_ON_ONCE(rbo->pin_count > 0))
409 		return -EINVAL;
410 
411 	rdev = radeon_get_rdev(bo->bdev);
412 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
413 		radeon_move_null(bo, new_mem);
414 		return 0;
415 	}
416 	if ((old_mem->mem_type == TTM_PL_TT &&
417 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
418 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
419 	     new_mem->mem_type == TTM_PL_TT)) {
420 		/* bind is enough */
421 		radeon_move_null(bo, new_mem);
422 		return 0;
423 	}
424 	if (!rdev->ring[radeon_copy_ring_index(rdev)].ready ||
425 	    rdev->asic->copy.copy == NULL) {
426 		/* use memcpy */
427 		goto memcpy;
428 	}
429 
430 	if (old_mem->mem_type == TTM_PL_VRAM &&
431 	    new_mem->mem_type == TTM_PL_SYSTEM) {
432 		r = radeon_move_vram_ram(bo, evict, interruptible,
433 					no_wait_gpu, new_mem);
434 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
435 		   new_mem->mem_type == TTM_PL_VRAM) {
436 		r = radeon_move_ram_vram(bo, evict, interruptible,
437 					    no_wait_gpu, new_mem);
438 	} else {
439 		r = radeon_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
440 	}
441 
442 	if (r) {
443 memcpy:
444 		r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
445 		if (r) {
446 			return r;
447 		}
448 	}
449 
450 	/* update statistics */
451 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &rdev->num_bytes_moved);
452 	return 0;
453 }
454 
455 static int radeon_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
456 {
457 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
458 	struct radeon_device *rdev = radeon_get_rdev(bdev);
459 
460 	mem->bus.addr = NULL;
461 	mem->bus.offset = 0;
462 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
463 	mem->bus.base = 0;
464 	mem->bus.is_iomem = false;
465 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
466 		return -EINVAL;
467 	switch (mem->mem_type) {
468 	case TTM_PL_SYSTEM:
469 		/* system memory */
470 		return 0;
471 	case TTM_PL_TT:
472 #if IS_ENABLED(CONFIG_AGP)
473 		if (rdev->flags & RADEON_IS_AGP) {
474 			/* RADEON_IS_AGP is set only if AGP is active */
475 			mem->bus.offset = mem->start << PAGE_SHIFT;
476 			mem->bus.base = rdev->mc.agp_base;
477 			mem->bus.is_iomem = !rdev->ddev->agp->cant_use_aperture;
478 		}
479 #endif
480 		break;
481 	case TTM_PL_VRAM:
482 		mem->bus.offset = mem->start << PAGE_SHIFT;
483 		/* check if it's visible */
484 		if ((mem->bus.offset + mem->bus.size) > rdev->mc.visible_vram_size)
485 			return -EINVAL;
486 		mem->bus.base = rdev->mc.aper_base;
487 		mem->bus.is_iomem = true;
488 #ifdef __alpha__
489 		/*
490 		 * Alpha: use bus.addr to hold the ioremap() return,
491 		 * so we can modify bus.base below.
492 		 */
493 		if (mem->placement & TTM_PL_FLAG_WC)
494 			mem->bus.addr =
495 				ioremap_wc(mem->bus.base + mem->bus.offset,
496 					   mem->bus.size);
497 		else
498 			mem->bus.addr =
499 				ioremap_nocache(mem->bus.base + mem->bus.offset,
500 						mem->bus.size);
501 
502 		/*
503 		 * Alpha: Use just the bus offset plus
504 		 * the hose/domain memory base for bus.base.
505 		 * It then can be used to build PTEs for VRAM
506 		 * access, as done in ttm_bo_vm_fault().
507 		 */
508 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
509 			rdev->ddev->hose->dense_mem_base;
510 #endif
511 		break;
512 	default:
513 		return -EINVAL;
514 	}
515 	return 0;
516 }
517 
518 static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
519 {
520 }
521 
522 /*
523  * TTM backend functions.
524  */
525 struct radeon_ttm_tt {
526 	struct ttm_dma_tt		ttm;
527 	struct radeon_device		*rdev;
528 	u64				offset;
529 
530 	uint64_t			userptr;
531 	struct mm_struct		*usermm;
532 	uint32_t			userflags;
533 };
534 
535 /* prepare the sg table with the user pages */
536 static int radeon_ttm_tt_pin_userptr(struct ttm_tt *ttm)
537 {
538 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
539 	struct radeon_ttm_tt *gtt = (void *)ttm;
540 	unsigned pinned = 0, nents;
541 	int r;
542 
543 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
544 	enum dma_data_direction direction = write ?
545 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
546 
547 	if (current->mm != gtt->usermm)
548 		return -EPERM;
549 
550 	if (gtt->userflags & RADEON_GEM_USERPTR_ANONONLY) {
551 		/* check that we only pin down anonymous memory
552 		   to prevent problems with writeback */
553 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
554 		struct vm_area_struct *vma;
555 		vma = find_vma(gtt->usermm, gtt->userptr);
556 		if (!vma || vma->vm_file || vma->vm_end < end)
557 			return -EPERM;
558 	}
559 
560 	do {
561 		unsigned num_pages = ttm->num_pages - pinned;
562 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
563 		struct page **pages = ttm->pages + pinned;
564 
565 		r = get_user_pages(userptr, num_pages, write, 0, pages, NULL);
566 		if (r < 0)
567 			goto release_pages;
568 
569 		pinned += r;
570 
571 	} while (pinned < ttm->num_pages);
572 
573 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
574 				      ttm->num_pages << PAGE_SHIFT,
575 				      GFP_KERNEL);
576 	if (r)
577 		goto release_sg;
578 
579 	r = -ENOMEM;
580 	nents = dma_map_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
581 	if (nents != ttm->sg->nents)
582 		goto release_sg;
583 
584 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
585 					 gtt->ttm.dma_address, ttm->num_pages);
586 
587 	return 0;
588 
589 release_sg:
590 	kfree(ttm->sg);
591 
592 release_pages:
593 	release_pages(ttm->pages, pinned, 0);
594 	return r;
595 }
596 
597 static void radeon_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
598 {
599 	struct radeon_device *rdev = radeon_get_rdev(ttm->bdev);
600 	struct radeon_ttm_tt *gtt = (void *)ttm;
601 	struct sg_page_iter sg_iter;
602 
603 	int write = !(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
604 	enum dma_data_direction direction = write ?
605 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
606 
607 	/* double check that we don't free the table twice */
608 	if (!ttm->sg->sgl)
609 		return;
610 
611 	/* free the sg table and pages again */
612 	dma_unmap_sg(rdev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
613 
614 	for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
615 		struct page *page = sg_page_iter_page(&sg_iter);
616 		if (!(gtt->userflags & RADEON_GEM_USERPTR_READONLY))
617 			set_page_dirty(page);
618 
619 		mark_page_accessed(page);
620 		put_page(page);
621 	}
622 
623 	sg_free_table(ttm->sg);
624 }
625 
626 static int radeon_ttm_backend_bind(struct ttm_tt *ttm,
627 				   struct ttm_mem_reg *bo_mem)
628 {
629 	struct radeon_ttm_tt *gtt = (void*)ttm;
630 	uint32_t flags = RADEON_GART_PAGE_VALID | RADEON_GART_PAGE_READ |
631 		RADEON_GART_PAGE_WRITE;
632 	int r;
633 
634 	if (gtt->userptr) {
635 		radeon_ttm_tt_pin_userptr(ttm);
636 		flags &= ~RADEON_GART_PAGE_WRITE;
637 	}
638 
639 	gtt->offset = (unsigned long)(bo_mem->start << PAGE_SHIFT);
640 	if (!ttm->num_pages) {
641 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
642 		     ttm->num_pages, bo_mem, ttm);
643 	}
644 	if (ttm->caching_state == tt_cached)
645 		flags |= RADEON_GART_PAGE_SNOOP;
646 	r = radeon_gart_bind(gtt->rdev, gtt->offset, ttm->num_pages,
647 			     ttm->pages, gtt->ttm.dma_address, flags);
648 	if (r) {
649 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
650 			  ttm->num_pages, (unsigned)gtt->offset);
651 		return r;
652 	}
653 	return 0;
654 }
655 
656 static int radeon_ttm_backend_unbind(struct ttm_tt *ttm)
657 {
658 	struct radeon_ttm_tt *gtt = (void *)ttm;
659 
660 	radeon_gart_unbind(gtt->rdev, gtt->offset, ttm->num_pages);
661 
662 	if (gtt->userptr)
663 		radeon_ttm_tt_unpin_userptr(ttm);
664 
665 	return 0;
666 }
667 
668 static void radeon_ttm_backend_destroy(struct ttm_tt *ttm)
669 {
670 	struct radeon_ttm_tt *gtt = (void *)ttm;
671 
672 	ttm_dma_tt_fini(&gtt->ttm);
673 	kfree(gtt);
674 }
675 
676 static struct ttm_backend_func radeon_backend_func = {
677 	.bind = &radeon_ttm_backend_bind,
678 	.unbind = &radeon_ttm_backend_unbind,
679 	.destroy = &radeon_ttm_backend_destroy,
680 };
681 
682 static struct ttm_tt *radeon_ttm_tt_create(struct ttm_bo_device *bdev,
683 				    unsigned long size, uint32_t page_flags,
684 				    struct page *dummy_read_page)
685 {
686 	struct radeon_device *rdev;
687 	struct radeon_ttm_tt *gtt;
688 
689 	rdev = radeon_get_rdev(bdev);
690 #if IS_ENABLED(CONFIG_AGP)
691 	if (rdev->flags & RADEON_IS_AGP) {
692 		return ttm_agp_tt_create(bdev, rdev->ddev->agp->bridge,
693 					 size, page_flags, dummy_read_page);
694 	}
695 #endif
696 
697 	gtt = kzalloc(sizeof(struct radeon_ttm_tt), GFP_KERNEL);
698 	if (gtt == NULL) {
699 		return NULL;
700 	}
701 	gtt->ttm.ttm.func = &radeon_backend_func;
702 	gtt->rdev = rdev;
703 	if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
704 		kfree(gtt);
705 		return NULL;
706 	}
707 	return &gtt->ttm.ttm;
708 }
709 
710 static struct radeon_ttm_tt *radeon_ttm_tt_to_gtt(struct ttm_tt *ttm)
711 {
712 	if (!ttm || ttm->func != &radeon_backend_func)
713 		return NULL;
714 	return (struct radeon_ttm_tt *)ttm;
715 }
716 
717 static int radeon_ttm_tt_populate(struct ttm_tt *ttm)
718 {
719 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
720 	struct radeon_device *rdev;
721 	unsigned i;
722 	int r;
723 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
724 
725 	if (ttm->state != tt_unpopulated)
726 		return 0;
727 
728 	if (gtt && gtt->userptr) {
729 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
730 		if (!ttm->sg)
731 			return -ENOMEM;
732 
733 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
734 		ttm->state = tt_unbound;
735 		return 0;
736 	}
737 
738 	if (slave && ttm->sg) {
739 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
740 						 gtt->ttm.dma_address, ttm->num_pages);
741 		ttm->state = tt_unbound;
742 		return 0;
743 	}
744 
745 	rdev = radeon_get_rdev(ttm->bdev);
746 #if IS_ENABLED(CONFIG_AGP)
747 	if (rdev->flags & RADEON_IS_AGP) {
748 		return ttm_agp_tt_populate(ttm);
749 	}
750 #endif
751 
752 #ifdef CONFIG_SWIOTLB
753 	if (swiotlb_nr_tbl()) {
754 		return ttm_dma_populate(&gtt->ttm, rdev->dev);
755 	}
756 #endif
757 
758 	r = ttm_pool_populate(ttm);
759 	if (r) {
760 		return r;
761 	}
762 
763 	for (i = 0; i < ttm->num_pages; i++) {
764 		gtt->ttm.dma_address[i] = pci_map_page(rdev->pdev, ttm->pages[i],
765 						       0, PAGE_SIZE,
766 						       PCI_DMA_BIDIRECTIONAL);
767 		if (pci_dma_mapping_error(rdev->pdev, gtt->ttm.dma_address[i])) {
768 			while (i--) {
769 				pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
770 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
771 				gtt->ttm.dma_address[i] = 0;
772 			}
773 			ttm_pool_unpopulate(ttm);
774 			return -EFAULT;
775 		}
776 	}
777 	return 0;
778 }
779 
780 static void radeon_ttm_tt_unpopulate(struct ttm_tt *ttm)
781 {
782 	struct radeon_device *rdev;
783 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
784 	unsigned i;
785 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
786 
787 	if (gtt && gtt->userptr) {
788 		kfree(ttm->sg);
789 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
790 		return;
791 	}
792 
793 	if (slave)
794 		return;
795 
796 	rdev = radeon_get_rdev(ttm->bdev);
797 #if IS_ENABLED(CONFIG_AGP)
798 	if (rdev->flags & RADEON_IS_AGP) {
799 		ttm_agp_tt_unpopulate(ttm);
800 		return;
801 	}
802 #endif
803 
804 #ifdef CONFIG_SWIOTLB
805 	if (swiotlb_nr_tbl()) {
806 		ttm_dma_unpopulate(&gtt->ttm, rdev->dev);
807 		return;
808 	}
809 #endif
810 
811 	for (i = 0; i < ttm->num_pages; i++) {
812 		if (gtt->ttm.dma_address[i]) {
813 			pci_unmap_page(rdev->pdev, gtt->ttm.dma_address[i],
814 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
815 		}
816 	}
817 
818 	ttm_pool_unpopulate(ttm);
819 }
820 
821 int radeon_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
822 			      uint32_t flags)
823 {
824 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
825 
826 	if (gtt == NULL)
827 		return -EINVAL;
828 
829 	gtt->userptr = addr;
830 	gtt->usermm = current->mm;
831 	gtt->userflags = flags;
832 	return 0;
833 }
834 
835 bool radeon_ttm_tt_has_userptr(struct ttm_tt *ttm)
836 {
837 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
838 
839 	if (gtt == NULL)
840 		return false;
841 
842 	return !!gtt->userptr;
843 }
844 
845 bool radeon_ttm_tt_is_readonly(struct ttm_tt *ttm)
846 {
847 	struct radeon_ttm_tt *gtt = radeon_ttm_tt_to_gtt(ttm);
848 
849 	if (gtt == NULL)
850 		return false;
851 
852 	return !!(gtt->userflags & RADEON_GEM_USERPTR_READONLY);
853 }
854 
855 static struct ttm_bo_driver radeon_bo_driver = {
856 	.ttm_tt_create = &radeon_ttm_tt_create,
857 	.ttm_tt_populate = &radeon_ttm_tt_populate,
858 	.ttm_tt_unpopulate = &radeon_ttm_tt_unpopulate,
859 	.invalidate_caches = &radeon_invalidate_caches,
860 	.init_mem_type = &radeon_init_mem_type,
861 	.evict_flags = &radeon_evict_flags,
862 	.move = &radeon_bo_move,
863 	.verify_access = &radeon_verify_access,
864 	.move_notify = &radeon_bo_move_notify,
865 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
866 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
867 	.io_mem_free = &radeon_ttm_io_mem_free,
868 	.lru_tail = &ttm_bo_default_lru_tail,
869 	.swap_lru_tail = &ttm_bo_default_swap_lru_tail,
870 };
871 
872 int radeon_ttm_init(struct radeon_device *rdev)
873 {
874 	int r;
875 
876 	r = radeon_ttm_global_init(rdev);
877 	if (r) {
878 		return r;
879 	}
880 	/* No others user of address space so set it to 0 */
881 	r = ttm_bo_device_init(&rdev->mman.bdev,
882 			       rdev->mman.bo_global_ref.ref.object,
883 			       &radeon_bo_driver,
884 			       rdev->ddev->anon_inode->i_mapping,
885 			       DRM_FILE_PAGE_OFFSET,
886 			       rdev->need_dma32);
887 	if (r) {
888 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
889 		return r;
890 	}
891 	rdev->mman.initialized = true;
892 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_VRAM,
893 				rdev->mc.real_vram_size >> PAGE_SHIFT);
894 	if (r) {
895 		DRM_ERROR("Failed initializing VRAM heap.\n");
896 		return r;
897 	}
898 	/* Change the size here instead of the init above so only lpfn is affected */
899 	radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
900 
901 	r = radeon_bo_create(rdev, 256 * 1024, PAGE_SIZE, true,
902 			     RADEON_GEM_DOMAIN_VRAM, 0, NULL,
903 			     NULL, &rdev->stollen_vga_memory);
904 	if (r) {
905 		return r;
906 	}
907 	r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
908 	if (r)
909 		return r;
910 	r = radeon_bo_pin(rdev->stollen_vga_memory, RADEON_GEM_DOMAIN_VRAM, NULL);
911 	radeon_bo_unreserve(rdev->stollen_vga_memory);
912 	if (r) {
913 		radeon_bo_unref(&rdev->stollen_vga_memory);
914 		return r;
915 	}
916 	DRM_INFO("radeon: %uM of VRAM memory ready\n",
917 		 (unsigned) (rdev->mc.real_vram_size / (1024 * 1024)));
918 	r = ttm_bo_init_mm(&rdev->mman.bdev, TTM_PL_TT,
919 				rdev->mc.gtt_size >> PAGE_SHIFT);
920 	if (r) {
921 		DRM_ERROR("Failed initializing GTT heap.\n");
922 		return r;
923 	}
924 	DRM_INFO("radeon: %uM of GTT memory ready.\n",
925 		 (unsigned)(rdev->mc.gtt_size / (1024 * 1024)));
926 
927 	r = radeon_ttm_debugfs_init(rdev);
928 	if (r) {
929 		DRM_ERROR("Failed to init debugfs\n");
930 		return r;
931 	}
932 	return 0;
933 }
934 
935 void radeon_ttm_fini(struct radeon_device *rdev)
936 {
937 	int r;
938 
939 	if (!rdev->mman.initialized)
940 		return;
941 	radeon_ttm_debugfs_fini(rdev);
942 	if (rdev->stollen_vga_memory) {
943 		r = radeon_bo_reserve(rdev->stollen_vga_memory, false);
944 		if (r == 0) {
945 			radeon_bo_unpin(rdev->stollen_vga_memory);
946 			radeon_bo_unreserve(rdev->stollen_vga_memory);
947 		}
948 		radeon_bo_unref(&rdev->stollen_vga_memory);
949 	}
950 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_VRAM);
951 	ttm_bo_clean_mm(&rdev->mman.bdev, TTM_PL_TT);
952 	ttm_bo_device_release(&rdev->mman.bdev);
953 	radeon_gart_fini(rdev);
954 	radeon_ttm_global_fini(rdev);
955 	rdev->mman.initialized = false;
956 	DRM_INFO("radeon: ttm finalized\n");
957 }
958 
959 /* this should only be called at bootup or when userspace
960  * isn't running */
961 void radeon_ttm_set_active_vram_size(struct radeon_device *rdev, u64 size)
962 {
963 	struct ttm_mem_type_manager *man;
964 
965 	if (!rdev->mman.initialized)
966 		return;
967 
968 	man = &rdev->mman.bdev.man[TTM_PL_VRAM];
969 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
970 	man->size = size >> PAGE_SHIFT;
971 }
972 
973 static struct vm_operations_struct radeon_ttm_vm_ops;
974 static const struct vm_operations_struct *ttm_vm_ops = NULL;
975 
976 static int radeon_ttm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
977 {
978 	struct ttm_buffer_object *bo;
979 	struct radeon_device *rdev;
980 	int r;
981 
982 	bo = (struct ttm_buffer_object *)vma->vm_private_data;
983 	if (bo == NULL) {
984 		return VM_FAULT_NOPAGE;
985 	}
986 	rdev = radeon_get_rdev(bo->bdev);
987 	down_read(&rdev->pm.mclk_lock);
988 	r = ttm_vm_ops->fault(vma, vmf);
989 	up_read(&rdev->pm.mclk_lock);
990 	return r;
991 }
992 
993 int radeon_mmap(struct file *filp, struct vm_area_struct *vma)
994 {
995 	struct drm_file *file_priv;
996 	struct radeon_device *rdev;
997 	int r;
998 
999 	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET)) {
1000 		return -EINVAL;
1001 	}
1002 
1003 	file_priv = filp->private_data;
1004 	rdev = file_priv->minor->dev->dev_private;
1005 	if (rdev == NULL) {
1006 		return -EINVAL;
1007 	}
1008 	r = ttm_bo_mmap(filp, vma, &rdev->mman.bdev);
1009 	if (unlikely(r != 0)) {
1010 		return r;
1011 	}
1012 	if (unlikely(ttm_vm_ops == NULL)) {
1013 		ttm_vm_ops = vma->vm_ops;
1014 		radeon_ttm_vm_ops = *ttm_vm_ops;
1015 		radeon_ttm_vm_ops.fault = &radeon_ttm_fault;
1016 	}
1017 	vma->vm_ops = &radeon_ttm_vm_ops;
1018 	return 0;
1019 }
1020 
1021 #if defined(CONFIG_DEBUG_FS)
1022 
1023 static int radeon_mm_dump_table(struct seq_file *m, void *data)
1024 {
1025 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1026 	unsigned ttm_pl = *(int *)node->info_ent->data;
1027 	struct drm_device *dev = node->minor->dev;
1028 	struct radeon_device *rdev = dev->dev_private;
1029 	struct drm_mm *mm = (struct drm_mm *)rdev->mman.bdev.man[ttm_pl].priv;
1030 	int ret;
1031 	struct ttm_bo_global *glob = rdev->mman.bdev.glob;
1032 
1033 	spin_lock(&glob->lru_lock);
1034 	ret = drm_mm_dump_table(m, mm);
1035 	spin_unlock(&glob->lru_lock);
1036 	return ret;
1037 }
1038 
1039 static int ttm_pl_vram = TTM_PL_VRAM;
1040 static int ttm_pl_tt = TTM_PL_TT;
1041 
1042 static struct drm_info_list radeon_ttm_debugfs_list[] = {
1043 	{"radeon_vram_mm", radeon_mm_dump_table, 0, &ttm_pl_vram},
1044 	{"radeon_gtt_mm", radeon_mm_dump_table, 0, &ttm_pl_tt},
1045 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1046 #ifdef CONFIG_SWIOTLB
1047 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1048 #endif
1049 };
1050 
1051 static int radeon_ttm_vram_open(struct inode *inode, struct file *filep)
1052 {
1053 	struct radeon_device *rdev = inode->i_private;
1054 	i_size_write(inode, rdev->mc.mc_vram_size);
1055 	filep->private_data = inode->i_private;
1056 	return 0;
1057 }
1058 
1059 static ssize_t radeon_ttm_vram_read(struct file *f, char __user *buf,
1060 				    size_t size, loff_t *pos)
1061 {
1062 	struct radeon_device *rdev = f->private_data;
1063 	ssize_t result = 0;
1064 	int r;
1065 
1066 	if (size & 0x3 || *pos & 0x3)
1067 		return -EINVAL;
1068 
1069 	while (size) {
1070 		unsigned long flags;
1071 		uint32_t value;
1072 
1073 		if (*pos >= rdev->mc.mc_vram_size)
1074 			return result;
1075 
1076 		spin_lock_irqsave(&rdev->mmio_idx_lock, flags);
1077 		WREG32(RADEON_MM_INDEX, ((uint32_t)*pos) | 0x80000000);
1078 		if (rdev->family >= CHIP_CEDAR)
1079 			WREG32(EVERGREEN_MM_INDEX_HI, *pos >> 31);
1080 		value = RREG32(RADEON_MM_DATA);
1081 		spin_unlock_irqrestore(&rdev->mmio_idx_lock, flags);
1082 
1083 		r = put_user(value, (uint32_t *)buf);
1084 		if (r)
1085 			return r;
1086 
1087 		result += 4;
1088 		buf += 4;
1089 		*pos += 4;
1090 		size -= 4;
1091 	}
1092 
1093 	return result;
1094 }
1095 
1096 static const struct file_operations radeon_ttm_vram_fops = {
1097 	.owner = THIS_MODULE,
1098 	.open = radeon_ttm_vram_open,
1099 	.read = radeon_ttm_vram_read,
1100 	.llseek = default_llseek
1101 };
1102 
1103 static int radeon_ttm_gtt_open(struct inode *inode, struct file *filep)
1104 {
1105 	struct radeon_device *rdev = inode->i_private;
1106 	i_size_write(inode, rdev->mc.gtt_size);
1107 	filep->private_data = inode->i_private;
1108 	return 0;
1109 }
1110 
1111 static ssize_t radeon_ttm_gtt_read(struct file *f, char __user *buf,
1112 				   size_t size, loff_t *pos)
1113 {
1114 	struct radeon_device *rdev = f->private_data;
1115 	ssize_t result = 0;
1116 	int r;
1117 
1118 	while (size) {
1119 		loff_t p = *pos / PAGE_SIZE;
1120 		unsigned off = *pos & ~PAGE_MASK;
1121 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1122 		struct page *page;
1123 		void *ptr;
1124 
1125 		if (p >= rdev->gart.num_cpu_pages)
1126 			return result;
1127 
1128 		page = rdev->gart.pages[p];
1129 		if (page) {
1130 			ptr = kmap(page);
1131 			ptr += off;
1132 
1133 			r = copy_to_user(buf, ptr, cur_size);
1134 			kunmap(rdev->gart.pages[p]);
1135 		} else
1136 			r = clear_user(buf, cur_size);
1137 
1138 		if (r)
1139 			return -EFAULT;
1140 
1141 		result += cur_size;
1142 		buf += cur_size;
1143 		*pos += cur_size;
1144 		size -= cur_size;
1145 	}
1146 
1147 	return result;
1148 }
1149 
1150 static const struct file_operations radeon_ttm_gtt_fops = {
1151 	.owner = THIS_MODULE,
1152 	.open = radeon_ttm_gtt_open,
1153 	.read = radeon_ttm_gtt_read,
1154 	.llseek = default_llseek
1155 };
1156 
1157 #endif
1158 
1159 static int radeon_ttm_debugfs_init(struct radeon_device *rdev)
1160 {
1161 #if defined(CONFIG_DEBUG_FS)
1162 	unsigned count;
1163 
1164 	struct drm_minor *minor = rdev->ddev->primary;
1165 	struct dentry *ent, *root = minor->debugfs_root;
1166 
1167 	ent = debugfs_create_file("radeon_vram", S_IFREG | S_IRUGO, root,
1168 				  rdev, &radeon_ttm_vram_fops);
1169 	if (IS_ERR(ent))
1170 		return PTR_ERR(ent);
1171 	rdev->mman.vram = ent;
1172 
1173 	ent = debugfs_create_file("radeon_gtt", S_IFREG | S_IRUGO, root,
1174 				  rdev, &radeon_ttm_gtt_fops);
1175 	if (IS_ERR(ent))
1176 		return PTR_ERR(ent);
1177 	rdev->mman.gtt = ent;
1178 
1179 	count = ARRAY_SIZE(radeon_ttm_debugfs_list);
1180 
1181 #ifdef CONFIG_SWIOTLB
1182 	if (!swiotlb_nr_tbl())
1183 		--count;
1184 #endif
1185 
1186 	return radeon_debugfs_add_files(rdev, radeon_ttm_debugfs_list, count);
1187 #else
1188 
1189 	return 0;
1190 #endif
1191 }
1192 
1193 static void radeon_ttm_debugfs_fini(struct radeon_device *rdev)
1194 {
1195 #if defined(CONFIG_DEBUG_FS)
1196 
1197 	debugfs_remove(rdev->mman.vram);
1198 	rdev->mman.vram = NULL;
1199 
1200 	debugfs_remove(rdev->mman.gtt);
1201 	rdev->mman.gtt = NULL;
1202 #endif
1203 }
1204