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 <ttm/ttm_memory.h>
38 #include <drm/drmP.h>
39 #include <drm/amdgpu_drm.h>
40 #include <linux/seq_file.h>
41 #include <linux/slab.h>
42 #include <linux/swiotlb.h>
43 #include <linux/swap.h>
44 #include <linux/pagemap.h>
45 #include <linux/debugfs.h>
46 #include "amdgpu.h"
47 #include "bif/bif_4_1_d.h"
48 
49 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
50 
51 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
52 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
53 
54 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev)
55 {
56 	struct amdgpu_mman *mman;
57 	struct amdgpu_device *adev;
58 
59 	mman = container_of(bdev, struct amdgpu_mman, bdev);
60 	adev = container_of(mman, struct amdgpu_device, mman);
61 	return adev;
62 }
63 
64 
65 /*
66  * Global memory.
67  */
68 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
69 {
70 	return ttm_mem_global_init(ref->object);
71 }
72 
73 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
74 {
75 	ttm_mem_global_release(ref->object);
76 }
77 
78 int amdgpu_ttm_global_init(struct amdgpu_device *adev)
79 {
80 	struct drm_global_reference *global_ref;
81 	struct amdgpu_ring *ring;
82 	struct amd_sched_rq *rq;
83 	int r;
84 
85 	adev->mman.mem_global_referenced = false;
86 	global_ref = &adev->mman.mem_global_ref;
87 	global_ref->global_type = DRM_GLOBAL_TTM_MEM;
88 	global_ref->size = sizeof(struct ttm_mem_global);
89 	global_ref->init = &amdgpu_ttm_mem_global_init;
90 	global_ref->release = &amdgpu_ttm_mem_global_release;
91 	r = drm_global_item_ref(global_ref);
92 	if (r) {
93 		DRM_ERROR("Failed setting up TTM memory accounting "
94 			  "subsystem.\n");
95 		goto error_mem;
96 	}
97 
98 	adev->mman.bo_global_ref.mem_glob =
99 		adev->mman.mem_global_ref.object;
100 	global_ref = &adev->mman.bo_global_ref.ref;
101 	global_ref->global_type = DRM_GLOBAL_TTM_BO;
102 	global_ref->size = sizeof(struct ttm_bo_global);
103 	global_ref->init = &ttm_bo_global_init;
104 	global_ref->release = &ttm_bo_global_release;
105 	r = drm_global_item_ref(global_ref);
106 	if (r) {
107 		DRM_ERROR("Failed setting up TTM BO subsystem.\n");
108 		goto error_bo;
109 	}
110 
111 	ring = adev->mman.buffer_funcs_ring;
112 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
113 	r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
114 				  rq, amdgpu_sched_jobs);
115 	if (r) {
116 		DRM_ERROR("Failed setting up TTM BO move run queue.\n");
117 		goto error_entity;
118 	}
119 
120 	adev->mman.mem_global_referenced = true;
121 
122 	return 0;
123 
124 error_entity:
125 	drm_global_item_unref(&adev->mman.bo_global_ref.ref);
126 error_bo:
127 	drm_global_item_unref(&adev->mman.mem_global_ref);
128 error_mem:
129 	return r;
130 }
131 
132 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
133 {
134 	if (adev->mman.mem_global_referenced) {
135 		amd_sched_entity_fini(adev->mman.entity.sched,
136 				      &adev->mman.entity);
137 		drm_global_item_unref(&adev->mman.bo_global_ref.ref);
138 		drm_global_item_unref(&adev->mman.mem_global_ref);
139 		adev->mman.mem_global_referenced = false;
140 	}
141 }
142 
143 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
144 {
145 	return 0;
146 }
147 
148 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
149 				struct ttm_mem_type_manager *man)
150 {
151 	struct amdgpu_device *adev;
152 
153 	adev = amdgpu_get_adev(bdev);
154 
155 	switch (type) {
156 	case TTM_PL_SYSTEM:
157 		/* System memory */
158 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
159 		man->available_caching = TTM_PL_MASK_CACHING;
160 		man->default_caching = TTM_PL_FLAG_CACHED;
161 		break;
162 	case TTM_PL_TT:
163 		man->func = &amdgpu_gtt_mgr_func;
164 		man->gpu_offset = adev->mc.gtt_start;
165 		man->available_caching = TTM_PL_MASK_CACHING;
166 		man->default_caching = TTM_PL_FLAG_CACHED;
167 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE | TTM_MEMTYPE_FLAG_CMA;
168 		break;
169 	case TTM_PL_VRAM:
170 		/* "On-card" video ram */
171 		man->func = &ttm_bo_manager_func;
172 		man->gpu_offset = adev->mc.vram_start;
173 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
174 			     TTM_MEMTYPE_FLAG_MAPPABLE;
175 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
176 		man->default_caching = TTM_PL_FLAG_WC;
177 		break;
178 	case AMDGPU_PL_GDS:
179 	case AMDGPU_PL_GWS:
180 	case AMDGPU_PL_OA:
181 		/* On-chip GDS memory*/
182 		man->func = &ttm_bo_manager_func;
183 		man->gpu_offset = 0;
184 		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
185 		man->available_caching = TTM_PL_FLAG_UNCACHED;
186 		man->default_caching = TTM_PL_FLAG_UNCACHED;
187 		break;
188 	default:
189 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
190 		return -EINVAL;
191 	}
192 	return 0;
193 }
194 
195 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
196 				struct ttm_placement *placement)
197 {
198 	struct amdgpu_bo *abo;
199 	static struct ttm_place placements = {
200 		.fpfn = 0,
201 		.lpfn = 0,
202 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
203 	};
204 	unsigned i;
205 
206 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
207 		placement->placement = &placements;
208 		placement->busy_placement = &placements;
209 		placement->num_placement = 1;
210 		placement->num_busy_placement = 1;
211 		return;
212 	}
213 	abo = container_of(bo, struct amdgpu_bo, tbo);
214 	switch (bo->mem.mem_type) {
215 	case TTM_PL_VRAM:
216 		if (abo->adev->mman.buffer_funcs_ring->ready == false) {
217 			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
218 		} else {
219 			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
220 			for (i = 0; i < abo->placement.num_placement; ++i) {
221 				if (!(abo->placements[i].flags &
222 				      TTM_PL_FLAG_TT))
223 					continue;
224 
225 				if (abo->placements[i].lpfn)
226 					continue;
227 
228 				/* set an upper limit to force directly
229 				 * allocating address space for the BO.
230 				 */
231 				abo->placements[i].lpfn =
232 					abo->adev->mc.gtt_size >> PAGE_SHIFT;
233 			}
234 		}
235 		break;
236 	case TTM_PL_TT:
237 	default:
238 		amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
239 	}
240 	*placement = abo->placement;
241 }
242 
243 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
244 {
245 	struct amdgpu_bo *abo = container_of(bo, struct amdgpu_bo, tbo);
246 
247 	if (amdgpu_ttm_tt_get_usermm(bo->ttm))
248 		return -EPERM;
249 	return drm_vma_node_verify_access(&abo->gem_base.vma_node,
250 					  filp->private_data);
251 }
252 
253 static void amdgpu_move_null(struct ttm_buffer_object *bo,
254 			     struct ttm_mem_reg *new_mem)
255 {
256 	struct ttm_mem_reg *old_mem = &bo->mem;
257 
258 	BUG_ON(old_mem->mm_node != NULL);
259 	*old_mem = *new_mem;
260 	new_mem->mm_node = NULL;
261 }
262 
263 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
264 			bool evict, bool no_wait_gpu,
265 			struct ttm_mem_reg *new_mem,
266 			struct ttm_mem_reg *old_mem)
267 {
268 	struct amdgpu_device *adev;
269 	struct amdgpu_ring *ring;
270 	uint64_t old_start, new_start;
271 	struct fence *fence;
272 	int r;
273 
274 	adev = amdgpu_get_adev(bo->bdev);
275 	ring = adev->mman.buffer_funcs_ring;
276 
277 	switch (old_mem->mem_type) {
278 	case TTM_PL_TT:
279 		r = amdgpu_ttm_bind(bo, old_mem);
280 		if (r)
281 			return r;
282 
283 	case TTM_PL_VRAM:
284 		old_start = (u64)old_mem->start << PAGE_SHIFT;
285 		old_start += bo->bdev->man[old_mem->mem_type].gpu_offset;
286 		break;
287 	default:
288 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
289 		return -EINVAL;
290 	}
291 	switch (new_mem->mem_type) {
292 	case TTM_PL_TT:
293 		r = amdgpu_ttm_bind(bo, new_mem);
294 		if (r)
295 			return r;
296 
297 	case TTM_PL_VRAM:
298 		new_start = (u64)new_mem->start << PAGE_SHIFT;
299 		new_start += bo->bdev->man[new_mem->mem_type].gpu_offset;
300 		break;
301 	default:
302 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
303 		return -EINVAL;
304 	}
305 	if (!ring->ready) {
306 		DRM_ERROR("Trying to move memory with ring turned off.\n");
307 		return -EINVAL;
308 	}
309 
310 	BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
311 
312 	r = amdgpu_copy_buffer(ring, old_start, new_start,
313 			       new_mem->num_pages * PAGE_SIZE, /* bytes */
314 			       bo->resv, &fence, false);
315 	if (r)
316 		return r;
317 
318 	r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
319 	fence_put(fence);
320 	return r;
321 }
322 
323 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
324 				bool evict, bool interruptible,
325 				bool no_wait_gpu,
326 				struct ttm_mem_reg *new_mem)
327 {
328 	struct amdgpu_device *adev;
329 	struct ttm_mem_reg *old_mem = &bo->mem;
330 	struct ttm_mem_reg tmp_mem;
331 	struct ttm_place placements;
332 	struct ttm_placement placement;
333 	int r;
334 
335 	adev = amdgpu_get_adev(bo->bdev);
336 	tmp_mem = *new_mem;
337 	tmp_mem.mm_node = NULL;
338 	placement.num_placement = 1;
339 	placement.placement = &placements;
340 	placement.num_busy_placement = 1;
341 	placement.busy_placement = &placements;
342 	placements.fpfn = 0;
343 	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
344 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
345 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
346 			     interruptible, no_wait_gpu);
347 	if (unlikely(r)) {
348 		return r;
349 	}
350 
351 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
352 	if (unlikely(r)) {
353 		goto out_cleanup;
354 	}
355 
356 	r = ttm_tt_bind(bo->ttm, &tmp_mem);
357 	if (unlikely(r)) {
358 		goto out_cleanup;
359 	}
360 	r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
361 	if (unlikely(r)) {
362 		goto out_cleanup;
363 	}
364 	r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, new_mem);
365 out_cleanup:
366 	ttm_bo_mem_put(bo, &tmp_mem);
367 	return r;
368 }
369 
370 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
371 				bool evict, bool interruptible,
372 				bool no_wait_gpu,
373 				struct ttm_mem_reg *new_mem)
374 {
375 	struct amdgpu_device *adev;
376 	struct ttm_mem_reg *old_mem = &bo->mem;
377 	struct ttm_mem_reg tmp_mem;
378 	struct ttm_placement placement;
379 	struct ttm_place placements;
380 	int r;
381 
382 	adev = amdgpu_get_adev(bo->bdev);
383 	tmp_mem = *new_mem;
384 	tmp_mem.mm_node = NULL;
385 	placement.num_placement = 1;
386 	placement.placement = &placements;
387 	placement.num_busy_placement = 1;
388 	placement.busy_placement = &placements;
389 	placements.fpfn = 0;
390 	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
391 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
392 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
393 			     interruptible, no_wait_gpu);
394 	if (unlikely(r)) {
395 		return r;
396 	}
397 	r = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, &tmp_mem);
398 	if (unlikely(r)) {
399 		goto out_cleanup;
400 	}
401 	r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
402 	if (unlikely(r)) {
403 		goto out_cleanup;
404 	}
405 out_cleanup:
406 	ttm_bo_mem_put(bo, &tmp_mem);
407 	return r;
408 }
409 
410 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
411 			bool evict, bool interruptible,
412 			bool no_wait_gpu,
413 			struct ttm_mem_reg *new_mem)
414 {
415 	struct amdgpu_device *adev;
416 	struct amdgpu_bo *abo;
417 	struct ttm_mem_reg *old_mem = &bo->mem;
418 	int r;
419 
420 	/* Can't move a pinned BO */
421 	abo = container_of(bo, struct amdgpu_bo, tbo);
422 	if (WARN_ON_ONCE(abo->pin_count > 0))
423 		return -EINVAL;
424 
425 	adev = amdgpu_get_adev(bo->bdev);
426 
427 	/* remember the eviction */
428 	if (evict)
429 		atomic64_inc(&adev->num_evictions);
430 
431 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
432 		amdgpu_move_null(bo, new_mem);
433 		return 0;
434 	}
435 	if ((old_mem->mem_type == TTM_PL_TT &&
436 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
437 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
438 	     new_mem->mem_type == TTM_PL_TT)) {
439 		/* bind is enough */
440 		amdgpu_move_null(bo, new_mem);
441 		return 0;
442 	}
443 	if (adev->mman.buffer_funcs == NULL ||
444 	    adev->mman.buffer_funcs_ring == NULL ||
445 	    !adev->mman.buffer_funcs_ring->ready) {
446 		/* use memcpy */
447 		goto memcpy;
448 	}
449 
450 	if (old_mem->mem_type == TTM_PL_VRAM &&
451 	    new_mem->mem_type == TTM_PL_SYSTEM) {
452 		r = amdgpu_move_vram_ram(bo, evict, interruptible,
453 					no_wait_gpu, new_mem);
454 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
455 		   new_mem->mem_type == TTM_PL_VRAM) {
456 		r = amdgpu_move_ram_vram(bo, evict, interruptible,
457 					    no_wait_gpu, new_mem);
458 	} else {
459 		r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
460 	}
461 
462 	if (r) {
463 memcpy:
464 		r = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, new_mem);
465 		if (r) {
466 			return r;
467 		}
468 	}
469 
470 	/* update statistics */
471 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
472 	return 0;
473 }
474 
475 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
476 {
477 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
478 	struct amdgpu_device *adev = amdgpu_get_adev(bdev);
479 
480 	mem->bus.addr = NULL;
481 	mem->bus.offset = 0;
482 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
483 	mem->bus.base = 0;
484 	mem->bus.is_iomem = false;
485 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
486 		return -EINVAL;
487 	switch (mem->mem_type) {
488 	case TTM_PL_SYSTEM:
489 		/* system memory */
490 		return 0;
491 	case TTM_PL_TT:
492 		break;
493 	case TTM_PL_VRAM:
494 		mem->bus.offset = mem->start << PAGE_SHIFT;
495 		/* check if it's visible */
496 		if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
497 			return -EINVAL;
498 		mem->bus.base = adev->mc.aper_base;
499 		mem->bus.is_iomem = true;
500 #ifdef __alpha__
501 		/*
502 		 * Alpha: use bus.addr to hold the ioremap() return,
503 		 * so we can modify bus.base below.
504 		 */
505 		if (mem->placement & TTM_PL_FLAG_WC)
506 			mem->bus.addr =
507 				ioremap_wc(mem->bus.base + mem->bus.offset,
508 					   mem->bus.size);
509 		else
510 			mem->bus.addr =
511 				ioremap_nocache(mem->bus.base + mem->bus.offset,
512 						mem->bus.size);
513 
514 		/*
515 		 * Alpha: Use just the bus offset plus
516 		 * the hose/domain memory base for bus.base.
517 		 * It then can be used to build PTEs for VRAM
518 		 * access, as done in ttm_bo_vm_fault().
519 		 */
520 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
521 			adev->ddev->hose->dense_mem_base;
522 #endif
523 		break;
524 	default:
525 		return -EINVAL;
526 	}
527 	return 0;
528 }
529 
530 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
531 {
532 }
533 
534 /*
535  * TTM backend functions.
536  */
537 struct amdgpu_ttm_gup_task_list {
538 	struct list_head	list;
539 	struct task_struct	*task;
540 };
541 
542 struct amdgpu_ttm_tt {
543 	struct ttm_dma_tt	ttm;
544 	struct amdgpu_device	*adev;
545 	u64			offset;
546 	uint64_t		userptr;
547 	struct mm_struct	*usermm;
548 	uint32_t		userflags;
549 	spinlock_t              guptasklock;
550 	struct list_head        guptasks;
551 	atomic_t		mmu_invalidations;
552 	struct list_head        list;
553 };
554 
555 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
556 {
557 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
558 	unsigned int flags = 0;
559 	unsigned pinned = 0;
560 	int r;
561 
562 	if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
563 		flags |= FOLL_WRITE;
564 
565 	if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
566 		/* check that we only use anonymous memory
567 		   to prevent problems with writeback */
568 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
569 		struct vm_area_struct *vma;
570 
571 		vma = find_vma(gtt->usermm, gtt->userptr);
572 		if (!vma || vma->vm_file || vma->vm_end < end)
573 			return -EPERM;
574 	}
575 
576 	do {
577 		unsigned num_pages = ttm->num_pages - pinned;
578 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
579 		struct page **p = pages + pinned;
580 		struct amdgpu_ttm_gup_task_list guptask;
581 
582 		guptask.task = current;
583 		spin_lock(&gtt->guptasklock);
584 		list_add(&guptask.list, &gtt->guptasks);
585 		spin_unlock(&gtt->guptasklock);
586 
587 		r = get_user_pages(userptr, num_pages, flags, p, NULL);
588 
589 		spin_lock(&gtt->guptasklock);
590 		list_del(&guptask.list);
591 		spin_unlock(&gtt->guptasklock);
592 
593 		if (r < 0)
594 			goto release_pages;
595 
596 		pinned += r;
597 
598 	} while (pinned < ttm->num_pages);
599 
600 	return 0;
601 
602 release_pages:
603 	release_pages(pages, pinned, 0);
604 	return r;
605 }
606 
607 /* prepare the sg table with the user pages */
608 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
609 {
610 	struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
611 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
612 	unsigned nents;
613 	int r;
614 
615 	int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
616 	enum dma_data_direction direction = write ?
617 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
618 
619 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
620 				      ttm->num_pages << PAGE_SHIFT,
621 				      GFP_KERNEL);
622 	if (r)
623 		goto release_sg;
624 
625 	r = -ENOMEM;
626 	nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
627 	if (nents != ttm->sg->nents)
628 		goto release_sg;
629 
630 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
631 					 gtt->ttm.dma_address, ttm->num_pages);
632 
633 	return 0;
634 
635 release_sg:
636 	kfree(ttm->sg);
637 	return r;
638 }
639 
640 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
641 {
642 	struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
643 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
644 	struct sg_page_iter sg_iter;
645 
646 	int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
647 	enum dma_data_direction direction = write ?
648 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
649 
650 	/* double check that we don't free the table twice */
651 	if (!ttm->sg->sgl)
652 		return;
653 
654 	/* free the sg table and pages again */
655 	dma_unmap_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
656 
657 	for_each_sg_page(ttm->sg->sgl, &sg_iter, ttm->sg->nents, 0) {
658 		struct page *page = sg_page_iter_page(&sg_iter);
659 		if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
660 			set_page_dirty(page);
661 
662 		mark_page_accessed(page);
663 		put_page(page);
664 	}
665 
666 	sg_free_table(ttm->sg);
667 }
668 
669 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
670 				   struct ttm_mem_reg *bo_mem)
671 {
672 	struct amdgpu_ttm_tt *gtt = (void*)ttm;
673 	int r;
674 
675 	if (gtt->userptr) {
676 		r = amdgpu_ttm_tt_pin_userptr(ttm);
677 		if (r) {
678 			DRM_ERROR("failed to pin userptr\n");
679 			return r;
680 		}
681 	}
682 	if (!ttm->num_pages) {
683 		WARN(1, "nothing to bind %lu pages for mreg %p back %p!\n",
684 		     ttm->num_pages, bo_mem, ttm);
685 	}
686 
687 	if (bo_mem->mem_type == AMDGPU_PL_GDS ||
688 	    bo_mem->mem_type == AMDGPU_PL_GWS ||
689 	    bo_mem->mem_type == AMDGPU_PL_OA)
690 		return -EINVAL;
691 
692 	return 0;
693 }
694 
695 bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
696 {
697 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
698 
699 	return gtt && !list_empty(&gtt->list);
700 }
701 
702 int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
703 {
704 	struct ttm_tt *ttm = bo->ttm;
705 	struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
706 	uint32_t flags;
707 	int r;
708 
709 	if (!ttm || amdgpu_ttm_is_bound(ttm))
710 		return 0;
711 
712 	r = amdgpu_gtt_mgr_alloc(&bo->bdev->man[TTM_PL_TT], bo,
713 				 NULL, bo_mem);
714 	if (r) {
715 		DRM_ERROR("Failed to allocate GTT address space (%d)\n", r);
716 		return r;
717 	}
718 
719 	flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
720 	gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
721 	r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
722 		ttm->pages, gtt->ttm.dma_address, flags);
723 
724 	if (r) {
725 		DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
726 			  ttm->num_pages, gtt->offset);
727 		return r;
728 	}
729 	spin_lock(&gtt->adev->gtt_list_lock);
730 	list_add_tail(&gtt->list, &gtt->adev->gtt_list);
731 	spin_unlock(&gtt->adev->gtt_list_lock);
732 	return 0;
733 }
734 
735 int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
736 {
737 	struct amdgpu_ttm_tt *gtt, *tmp;
738 	struct ttm_mem_reg bo_mem;
739 	uint32_t flags;
740 	int r;
741 
742 	bo_mem.mem_type = TTM_PL_TT;
743 	spin_lock(&adev->gtt_list_lock);
744 	list_for_each_entry_safe(gtt, tmp, &adev->gtt_list, list) {
745 		flags = amdgpu_ttm_tt_pte_flags(gtt->adev, &gtt->ttm.ttm, &bo_mem);
746 		r = amdgpu_gart_bind(adev, gtt->offset, gtt->ttm.ttm.num_pages,
747 				     gtt->ttm.ttm.pages, gtt->ttm.dma_address,
748 				     flags);
749 		if (r) {
750 			spin_unlock(&adev->gtt_list_lock);
751 			DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
752 				  gtt->ttm.ttm.num_pages, gtt->offset);
753 			return r;
754 		}
755 	}
756 	spin_unlock(&adev->gtt_list_lock);
757 	return 0;
758 }
759 
760 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
761 {
762 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
763 
764 	if (gtt->userptr)
765 		amdgpu_ttm_tt_unpin_userptr(ttm);
766 
767 	if (!amdgpu_ttm_is_bound(ttm))
768 		return 0;
769 
770 	/* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
771 	if (gtt->adev->gart.ready)
772 		amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
773 
774 	spin_lock(&gtt->adev->gtt_list_lock);
775 	list_del_init(&gtt->list);
776 	spin_unlock(&gtt->adev->gtt_list_lock);
777 
778 	return 0;
779 }
780 
781 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
782 {
783 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
784 
785 	ttm_dma_tt_fini(&gtt->ttm);
786 	kfree(gtt);
787 }
788 
789 static struct ttm_backend_func amdgpu_backend_func = {
790 	.bind = &amdgpu_ttm_backend_bind,
791 	.unbind = &amdgpu_ttm_backend_unbind,
792 	.destroy = &amdgpu_ttm_backend_destroy,
793 };
794 
795 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
796 				    unsigned long size, uint32_t page_flags,
797 				    struct page *dummy_read_page)
798 {
799 	struct amdgpu_device *adev;
800 	struct amdgpu_ttm_tt *gtt;
801 
802 	adev = amdgpu_get_adev(bdev);
803 
804 	gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
805 	if (gtt == NULL) {
806 		return NULL;
807 	}
808 	gtt->ttm.ttm.func = &amdgpu_backend_func;
809 	gtt->adev = adev;
810 	if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
811 		kfree(gtt);
812 		return NULL;
813 	}
814 	INIT_LIST_HEAD(&gtt->list);
815 	return &gtt->ttm.ttm;
816 }
817 
818 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
819 {
820 	struct amdgpu_device *adev;
821 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
822 	unsigned i;
823 	int r;
824 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
825 
826 	if (ttm->state != tt_unpopulated)
827 		return 0;
828 
829 	if (gtt && gtt->userptr) {
830 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
831 		if (!ttm->sg)
832 			return -ENOMEM;
833 
834 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
835 		ttm->state = tt_unbound;
836 		return 0;
837 	}
838 
839 	if (slave && ttm->sg) {
840 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
841 						 gtt->ttm.dma_address, ttm->num_pages);
842 		ttm->state = tt_unbound;
843 		return 0;
844 	}
845 
846 	adev = amdgpu_get_adev(ttm->bdev);
847 
848 #ifdef CONFIG_SWIOTLB
849 	if (swiotlb_nr_tbl()) {
850 		return ttm_dma_populate(&gtt->ttm, adev->dev);
851 	}
852 #endif
853 
854 	r = ttm_pool_populate(ttm);
855 	if (r) {
856 		return r;
857 	}
858 
859 	for (i = 0; i < ttm->num_pages; i++) {
860 		gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
861 						       0, PAGE_SIZE,
862 						       PCI_DMA_BIDIRECTIONAL);
863 		if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
864 			while (i--) {
865 				pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
866 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
867 				gtt->ttm.dma_address[i] = 0;
868 			}
869 			ttm_pool_unpopulate(ttm);
870 			return -EFAULT;
871 		}
872 	}
873 	return 0;
874 }
875 
876 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
877 {
878 	struct amdgpu_device *adev;
879 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
880 	unsigned i;
881 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
882 
883 	if (gtt && gtt->userptr) {
884 		kfree(ttm->sg);
885 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
886 		return;
887 	}
888 
889 	if (slave)
890 		return;
891 
892 	adev = amdgpu_get_adev(ttm->bdev);
893 
894 #ifdef CONFIG_SWIOTLB
895 	if (swiotlb_nr_tbl()) {
896 		ttm_dma_unpopulate(&gtt->ttm, adev->dev);
897 		return;
898 	}
899 #endif
900 
901 	for (i = 0; i < ttm->num_pages; i++) {
902 		if (gtt->ttm.dma_address[i]) {
903 			pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
904 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
905 		}
906 	}
907 
908 	ttm_pool_unpopulate(ttm);
909 }
910 
911 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
912 			      uint32_t flags)
913 {
914 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
915 
916 	if (gtt == NULL)
917 		return -EINVAL;
918 
919 	gtt->userptr = addr;
920 	gtt->usermm = current->mm;
921 	gtt->userflags = flags;
922 	spin_lock_init(&gtt->guptasklock);
923 	INIT_LIST_HEAD(&gtt->guptasks);
924 	atomic_set(&gtt->mmu_invalidations, 0);
925 
926 	return 0;
927 }
928 
929 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
930 {
931 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
932 
933 	if (gtt == NULL)
934 		return NULL;
935 
936 	return gtt->usermm;
937 }
938 
939 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
940 				  unsigned long end)
941 {
942 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
943 	struct amdgpu_ttm_gup_task_list *entry;
944 	unsigned long size;
945 
946 	if (gtt == NULL || !gtt->userptr)
947 		return false;
948 
949 	size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
950 	if (gtt->userptr > end || gtt->userptr + size <= start)
951 		return false;
952 
953 	spin_lock(&gtt->guptasklock);
954 	list_for_each_entry(entry, &gtt->guptasks, list) {
955 		if (entry->task == current) {
956 			spin_unlock(&gtt->guptasklock);
957 			return false;
958 		}
959 	}
960 	spin_unlock(&gtt->guptasklock);
961 
962 	atomic_inc(&gtt->mmu_invalidations);
963 
964 	return true;
965 }
966 
967 bool amdgpu_ttm_tt_userptr_invalidated(struct ttm_tt *ttm,
968 				       int *last_invalidated)
969 {
970 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
971 	int prev_invalidated = *last_invalidated;
972 
973 	*last_invalidated = atomic_read(&gtt->mmu_invalidations);
974 	return prev_invalidated != *last_invalidated;
975 }
976 
977 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
978 {
979 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
980 
981 	if (gtt == NULL)
982 		return false;
983 
984 	return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
985 }
986 
987 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
988 				 struct ttm_mem_reg *mem)
989 {
990 	uint32_t flags = 0;
991 
992 	if (mem && mem->mem_type != TTM_PL_SYSTEM)
993 		flags |= AMDGPU_PTE_VALID;
994 
995 	if (mem && mem->mem_type == TTM_PL_TT) {
996 		flags |= AMDGPU_PTE_SYSTEM;
997 
998 		if (ttm->caching_state == tt_cached)
999 			flags |= AMDGPU_PTE_SNOOPED;
1000 	}
1001 
1002 	if (adev->asic_type >= CHIP_TONGA)
1003 		flags |= AMDGPU_PTE_EXECUTABLE;
1004 
1005 	flags |= AMDGPU_PTE_READABLE;
1006 
1007 	if (!amdgpu_ttm_tt_is_readonly(ttm))
1008 		flags |= AMDGPU_PTE_WRITEABLE;
1009 
1010 	return flags;
1011 }
1012 
1013 static void amdgpu_ttm_lru_removal(struct ttm_buffer_object *tbo)
1014 {
1015 	struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
1016 	unsigned i, j;
1017 
1018 	for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
1019 		struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
1020 
1021 		for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1022 			if (&tbo->lru == lru->lru[j])
1023 				lru->lru[j] = tbo->lru.prev;
1024 
1025 		if (&tbo->swap == lru->swap_lru)
1026 			lru->swap_lru = tbo->swap.prev;
1027 	}
1028 }
1029 
1030 static struct amdgpu_mman_lru *amdgpu_ttm_lru(struct ttm_buffer_object *tbo)
1031 {
1032 	struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
1033 	unsigned log2_size = min(ilog2(tbo->num_pages),
1034 				 AMDGPU_TTM_LRU_SIZE - 1);
1035 
1036 	return &adev->mman.log2_size[log2_size];
1037 }
1038 
1039 static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
1040 {
1041 	struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
1042 	struct list_head *res = lru->lru[tbo->mem.mem_type];
1043 
1044 	lru->lru[tbo->mem.mem_type] = &tbo->lru;
1045 	while ((++lru)->lru[tbo->mem.mem_type] == res)
1046 		lru->lru[tbo->mem.mem_type] = &tbo->lru;
1047 
1048 	return res;
1049 }
1050 
1051 static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
1052 {
1053 	struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
1054 	struct list_head *res = lru->swap_lru;
1055 
1056 	lru->swap_lru = &tbo->swap;
1057 	while ((++lru)->swap_lru == res)
1058 		lru->swap_lru = &tbo->swap;
1059 
1060 	return res;
1061 }
1062 
1063 static struct ttm_bo_driver amdgpu_bo_driver = {
1064 	.ttm_tt_create = &amdgpu_ttm_tt_create,
1065 	.ttm_tt_populate = &amdgpu_ttm_tt_populate,
1066 	.ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1067 	.invalidate_caches = &amdgpu_invalidate_caches,
1068 	.init_mem_type = &amdgpu_init_mem_type,
1069 	.evict_flags = &amdgpu_evict_flags,
1070 	.move = &amdgpu_bo_move,
1071 	.verify_access = &amdgpu_verify_access,
1072 	.move_notify = &amdgpu_bo_move_notify,
1073 	.fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
1074 	.io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1075 	.io_mem_free = &amdgpu_ttm_io_mem_free,
1076 	.lru_removal = &amdgpu_ttm_lru_removal,
1077 	.lru_tail = &amdgpu_ttm_lru_tail,
1078 	.swap_lru_tail = &amdgpu_ttm_swap_lru_tail,
1079 };
1080 
1081 int amdgpu_ttm_init(struct amdgpu_device *adev)
1082 {
1083 	unsigned i, j;
1084 	int r;
1085 
1086 	/* No others user of address space so set it to 0 */
1087 	r = ttm_bo_device_init(&adev->mman.bdev,
1088 			       adev->mman.bo_global_ref.ref.object,
1089 			       &amdgpu_bo_driver,
1090 			       adev->ddev->anon_inode->i_mapping,
1091 			       DRM_FILE_PAGE_OFFSET,
1092 			       adev->need_dma32);
1093 	if (r) {
1094 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1095 		return r;
1096 	}
1097 
1098 	for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
1099 		struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
1100 
1101 		for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1102 			lru->lru[j] = &adev->mman.bdev.man[j].lru;
1103 		lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
1104 	}
1105 
1106 	for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
1107 		adev->mman.guard.lru[j] = NULL;
1108 	adev->mman.guard.swap_lru = NULL;
1109 
1110 	adev->mman.initialized = true;
1111 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
1112 				adev->mc.real_vram_size >> PAGE_SHIFT);
1113 	if (r) {
1114 		DRM_ERROR("Failed initializing VRAM heap.\n");
1115 		return r;
1116 	}
1117 	/* Change the size here instead of the init above so only lpfn is affected */
1118 	amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
1119 
1120 	r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true,
1121 			     AMDGPU_GEM_DOMAIN_VRAM,
1122 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
1123 			     NULL, NULL, &adev->stollen_vga_memory);
1124 	if (r) {
1125 		return r;
1126 	}
1127 	r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1128 	if (r)
1129 		return r;
1130 	r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL);
1131 	amdgpu_bo_unreserve(adev->stollen_vga_memory);
1132 	if (r) {
1133 		amdgpu_bo_unref(&adev->stollen_vga_memory);
1134 		return r;
1135 	}
1136 	DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1137 		 (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
1138 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT,
1139 				adev->mc.gtt_size >> PAGE_SHIFT);
1140 	if (r) {
1141 		DRM_ERROR("Failed initializing GTT heap.\n");
1142 		return r;
1143 	}
1144 	DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1145 		 (unsigned)(adev->mc.gtt_size / (1024 * 1024)));
1146 
1147 	adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
1148 	adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
1149 	adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
1150 	adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
1151 	adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
1152 	adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
1153 	adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
1154 	adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
1155 	adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
1156 	/* GDS Memory */
1157 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
1158 				adev->gds.mem.total_size >> PAGE_SHIFT);
1159 	if (r) {
1160 		DRM_ERROR("Failed initializing GDS heap.\n");
1161 		return r;
1162 	}
1163 
1164 	/* GWS */
1165 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
1166 				adev->gds.gws.total_size >> PAGE_SHIFT);
1167 	if (r) {
1168 		DRM_ERROR("Failed initializing gws heap.\n");
1169 		return r;
1170 	}
1171 
1172 	/* OA */
1173 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
1174 				adev->gds.oa.total_size >> PAGE_SHIFT);
1175 	if (r) {
1176 		DRM_ERROR("Failed initializing oa heap.\n");
1177 		return r;
1178 	}
1179 
1180 	r = amdgpu_ttm_debugfs_init(adev);
1181 	if (r) {
1182 		DRM_ERROR("Failed to init debugfs\n");
1183 		return r;
1184 	}
1185 	return 0;
1186 }
1187 
1188 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1189 {
1190 	int r;
1191 
1192 	if (!adev->mman.initialized)
1193 		return;
1194 	amdgpu_ttm_debugfs_fini(adev);
1195 	if (adev->stollen_vga_memory) {
1196 		r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
1197 		if (r == 0) {
1198 			amdgpu_bo_unpin(adev->stollen_vga_memory);
1199 			amdgpu_bo_unreserve(adev->stollen_vga_memory);
1200 		}
1201 		amdgpu_bo_unref(&adev->stollen_vga_memory);
1202 	}
1203 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
1204 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
1205 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
1206 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
1207 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
1208 	ttm_bo_device_release(&adev->mman.bdev);
1209 	amdgpu_gart_fini(adev);
1210 	amdgpu_ttm_global_fini(adev);
1211 	adev->mman.initialized = false;
1212 	DRM_INFO("amdgpu: ttm finalized\n");
1213 }
1214 
1215 /* this should only be called at bootup or when userspace
1216  * isn't running */
1217 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
1218 {
1219 	struct ttm_mem_type_manager *man;
1220 
1221 	if (!adev->mman.initialized)
1222 		return;
1223 
1224 	man = &adev->mman.bdev.man[TTM_PL_VRAM];
1225 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
1226 	man->size = size >> PAGE_SHIFT;
1227 }
1228 
1229 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1230 {
1231 	struct drm_file *file_priv;
1232 	struct amdgpu_device *adev;
1233 
1234 	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1235 		return -EINVAL;
1236 
1237 	file_priv = filp->private_data;
1238 	adev = file_priv->minor->dev->dev_private;
1239 	if (adev == NULL)
1240 		return -EINVAL;
1241 
1242 	return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1243 }
1244 
1245 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
1246 		       uint64_t src_offset,
1247 		       uint64_t dst_offset,
1248 		       uint32_t byte_count,
1249 		       struct reservation_object *resv,
1250 		       struct fence **fence, bool direct_submit)
1251 {
1252 	struct amdgpu_device *adev = ring->adev;
1253 	struct amdgpu_job *job;
1254 
1255 	uint32_t max_bytes;
1256 	unsigned num_loops, num_dw;
1257 	unsigned i;
1258 	int r;
1259 
1260 	max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1261 	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1262 	num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1263 
1264 	/* for IB padding */
1265 	while (num_dw & 0x7)
1266 		num_dw++;
1267 
1268 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1269 	if (r)
1270 		return r;
1271 
1272 	if (resv) {
1273 		r = amdgpu_sync_resv(adev, &job->sync, resv,
1274 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1275 		if (r) {
1276 			DRM_ERROR("sync failed (%d).\n", r);
1277 			goto error_free;
1278 		}
1279 	}
1280 
1281 	for (i = 0; i < num_loops; i++) {
1282 		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1283 
1284 		amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1285 					dst_offset, cur_size_in_bytes);
1286 
1287 		src_offset += cur_size_in_bytes;
1288 		dst_offset += cur_size_in_bytes;
1289 		byte_count -= cur_size_in_bytes;
1290 	}
1291 
1292 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1293 	WARN_ON(job->ibs[0].length_dw > num_dw);
1294 	if (direct_submit) {
1295 		r = amdgpu_ib_schedule(ring, job->num_ibs, job->ibs,
1296 				       NULL, NULL, fence);
1297 		job->fence = fence_get(*fence);
1298 		if (r)
1299 			DRM_ERROR("Error scheduling IBs (%d)\n", r);
1300 		amdgpu_job_free(job);
1301 	} else {
1302 		r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1303 				      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1304 		if (r)
1305 			goto error_free;
1306 	}
1307 
1308 	return r;
1309 
1310 error_free:
1311 	amdgpu_job_free(job);
1312 	return r;
1313 }
1314 
1315 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1316 		uint32_t src_data,
1317 		struct reservation_object *resv,
1318 		struct fence **fence)
1319 {
1320 	struct amdgpu_device *adev = bo->adev;
1321 	struct amdgpu_job *job;
1322 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1323 
1324 	uint32_t max_bytes, byte_count;
1325 	uint64_t dst_offset;
1326 	unsigned int num_loops, num_dw;
1327 	unsigned int i;
1328 	int r;
1329 
1330 	byte_count = bo->tbo.num_pages << PAGE_SHIFT;
1331 	max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1332 	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1333 	num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1334 
1335 	/* for IB padding */
1336 	while (num_dw & 0x7)
1337 		num_dw++;
1338 
1339 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
1340 	if (r)
1341 		return r;
1342 
1343 	if (resv) {
1344 		r = amdgpu_sync_resv(adev, &job->sync, resv,
1345 				AMDGPU_FENCE_OWNER_UNDEFINED);
1346 		if (r) {
1347 			DRM_ERROR("sync failed (%d).\n", r);
1348 			goto error_free;
1349 		}
1350 	}
1351 
1352 	dst_offset = bo->tbo.mem.start << PAGE_SHIFT;
1353 	for (i = 0; i < num_loops; i++) {
1354 		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1355 
1356 		amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
1357 				dst_offset, cur_size_in_bytes);
1358 
1359 		dst_offset += cur_size_in_bytes;
1360 		byte_count -= cur_size_in_bytes;
1361 	}
1362 
1363 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1364 	WARN_ON(job->ibs[0].length_dw > num_dw);
1365 	r = amdgpu_job_submit(job, ring, &adev->mman.entity,
1366 			AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1367 	if (r)
1368 		goto error_free;
1369 
1370 	return 0;
1371 
1372 error_free:
1373 	amdgpu_job_free(job);
1374 	return r;
1375 }
1376 
1377 #if defined(CONFIG_DEBUG_FS)
1378 
1379 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1380 {
1381 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1382 	unsigned ttm_pl = *(int *)node->info_ent->data;
1383 	struct drm_device *dev = node->minor->dev;
1384 	struct amdgpu_device *adev = dev->dev_private;
1385 	struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv;
1386 	int ret;
1387 	struct ttm_bo_global *glob = adev->mman.bdev.glob;
1388 
1389 	spin_lock(&glob->lru_lock);
1390 	ret = drm_mm_dump_table(m, mm);
1391 	spin_unlock(&glob->lru_lock);
1392 	if (ttm_pl == TTM_PL_VRAM)
1393 		seq_printf(m, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
1394 			   adev->mman.bdev.man[ttm_pl].size,
1395 			   (u64)atomic64_read(&adev->vram_usage) >> 20,
1396 			   (u64)atomic64_read(&adev->vram_vis_usage) >> 20);
1397 	return ret;
1398 }
1399 
1400 static int ttm_pl_vram = TTM_PL_VRAM;
1401 static int ttm_pl_tt = TTM_PL_TT;
1402 
1403 static const struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1404 	{"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1405 	{"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1406 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1407 #ifdef CONFIG_SWIOTLB
1408 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1409 #endif
1410 };
1411 
1412 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1413 				    size_t size, loff_t *pos)
1414 {
1415 	struct amdgpu_device *adev = f->f_inode->i_private;
1416 	ssize_t result = 0;
1417 	int r;
1418 
1419 	if (size & 0x3 || *pos & 0x3)
1420 		return -EINVAL;
1421 
1422 	while (size) {
1423 		unsigned long flags;
1424 		uint32_t value;
1425 
1426 		if (*pos >= adev->mc.mc_vram_size)
1427 			return result;
1428 
1429 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1430 		WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1431 		WREG32(mmMM_INDEX_HI, *pos >> 31);
1432 		value = RREG32(mmMM_DATA);
1433 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1434 
1435 		r = put_user(value, (uint32_t *)buf);
1436 		if (r)
1437 			return r;
1438 
1439 		result += 4;
1440 		buf += 4;
1441 		*pos += 4;
1442 		size -= 4;
1443 	}
1444 
1445 	return result;
1446 }
1447 
1448 static const struct file_operations amdgpu_ttm_vram_fops = {
1449 	.owner = THIS_MODULE,
1450 	.read = amdgpu_ttm_vram_read,
1451 	.llseek = default_llseek
1452 };
1453 
1454 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1455 
1456 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1457 				   size_t size, loff_t *pos)
1458 {
1459 	struct amdgpu_device *adev = f->f_inode->i_private;
1460 	ssize_t result = 0;
1461 	int r;
1462 
1463 	while (size) {
1464 		loff_t p = *pos / PAGE_SIZE;
1465 		unsigned off = *pos & ~PAGE_MASK;
1466 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1467 		struct page *page;
1468 		void *ptr;
1469 
1470 		if (p >= adev->gart.num_cpu_pages)
1471 			return result;
1472 
1473 		page = adev->gart.pages[p];
1474 		if (page) {
1475 			ptr = kmap(page);
1476 			ptr += off;
1477 
1478 			r = copy_to_user(buf, ptr, cur_size);
1479 			kunmap(adev->gart.pages[p]);
1480 		} else
1481 			r = clear_user(buf, cur_size);
1482 
1483 		if (r)
1484 			return -EFAULT;
1485 
1486 		result += cur_size;
1487 		buf += cur_size;
1488 		*pos += cur_size;
1489 		size -= cur_size;
1490 	}
1491 
1492 	return result;
1493 }
1494 
1495 static const struct file_operations amdgpu_ttm_gtt_fops = {
1496 	.owner = THIS_MODULE,
1497 	.read = amdgpu_ttm_gtt_read,
1498 	.llseek = default_llseek
1499 };
1500 
1501 #endif
1502 
1503 #endif
1504 
1505 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1506 {
1507 #if defined(CONFIG_DEBUG_FS)
1508 	unsigned count;
1509 
1510 	struct drm_minor *minor = adev->ddev->primary;
1511 	struct dentry *ent, *root = minor->debugfs_root;
1512 
1513 	ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1514 				  adev, &amdgpu_ttm_vram_fops);
1515 	if (IS_ERR(ent))
1516 		return PTR_ERR(ent);
1517 	i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1518 	adev->mman.vram = ent;
1519 
1520 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1521 	ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1522 				  adev, &amdgpu_ttm_gtt_fops);
1523 	if (IS_ERR(ent))
1524 		return PTR_ERR(ent);
1525 	i_size_write(ent->d_inode, adev->mc.gtt_size);
1526 	adev->mman.gtt = ent;
1527 
1528 #endif
1529 	count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1530 
1531 #ifdef CONFIG_SWIOTLB
1532 	if (!swiotlb_nr_tbl())
1533 		--count;
1534 #endif
1535 
1536 	return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1537 #else
1538 
1539 	return 0;
1540 #endif
1541 }
1542 
1543 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1544 {
1545 #if defined(CONFIG_DEBUG_FS)
1546 
1547 	debugfs_remove(adev->mman.vram);
1548 	adev->mman.vram = NULL;
1549 
1550 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
1551 	debugfs_remove(adev->mman.gtt);
1552 	adev->mman.gtt = NULL;
1553 #endif
1554 
1555 #endif
1556 }
1557 
1558 u64 amdgpu_ttm_get_gtt_mem_size(struct amdgpu_device *adev)
1559 {
1560 	return ttm_get_kernel_zone_memory_size(adev->mman.mem_global_ref.object);
1561 }
1562