• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/amdgpu_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 "amdgpu.h"
46 #include "bif/bif_4_1_d.h"
47 
48 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
49 
50 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
51 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
52 
amdgpu_get_adev(struct ttm_bo_device * bdev)53 static struct amdgpu_device *amdgpu_get_adev(struct ttm_bo_device *bdev)
54 {
55 	struct amdgpu_mman *mman;
56 	struct amdgpu_device *adev;
57 
58 	mman = container_of(bdev, struct amdgpu_mman, bdev);
59 	adev = container_of(mman, struct amdgpu_device, mman);
60 	return adev;
61 }
62 
63 
64 /*
65  * Global memory.
66  */
amdgpu_ttm_mem_global_init(struct drm_global_reference * ref)67 static int amdgpu_ttm_mem_global_init(struct drm_global_reference *ref)
68 {
69 	return ttm_mem_global_init(ref->object);
70 }
71 
amdgpu_ttm_mem_global_release(struct drm_global_reference * ref)72 static void amdgpu_ttm_mem_global_release(struct drm_global_reference *ref)
73 {
74 	ttm_mem_global_release(ref->object);
75 }
76 
amdgpu_ttm_global_init(struct amdgpu_device * adev)77 static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
78 {
79 	struct drm_global_reference *global_ref;
80 	int r;
81 
82 	adev->mman.mem_global_referenced = false;
83 	global_ref = &adev->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 = &amdgpu_ttm_mem_global_init;
87 	global_ref->release = &amdgpu_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 	adev->mman.bo_global_ref.mem_glob =
96 		adev->mman.mem_global_ref.object;
97 	global_ref = &adev->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(&adev->mman.mem_global_ref);
106 		return r;
107 	}
108 
109 	adev->mman.mem_global_referenced = true;
110 	return 0;
111 }
112 
amdgpu_ttm_global_fini(struct amdgpu_device * adev)113 static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
114 {
115 	if (adev->mman.mem_global_referenced) {
116 		drm_global_item_unref(&adev->mman.bo_global_ref.ref);
117 		drm_global_item_unref(&adev->mman.mem_global_ref);
118 		adev->mman.mem_global_referenced = false;
119 	}
120 }
121 
amdgpu_invalidate_caches(struct ttm_bo_device * bdev,uint32_t flags)122 static int amdgpu_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
123 {
124 	return 0;
125 }
126 
amdgpu_init_mem_type(struct ttm_bo_device * bdev,uint32_t type,struct ttm_mem_type_manager * man)127 static int amdgpu_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
128 				struct ttm_mem_type_manager *man)
129 {
130 	struct amdgpu_device *adev;
131 
132 	adev = amdgpu_get_adev(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 = adev->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 		break;
148 	case TTM_PL_VRAM:
149 		/* "On-card" video ram */
150 		man->func = &ttm_bo_manager_func;
151 		man->gpu_offset = adev->mc.vram_start;
152 		man->flags = TTM_MEMTYPE_FLAG_FIXED |
153 			     TTM_MEMTYPE_FLAG_MAPPABLE;
154 		man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
155 		man->default_caching = TTM_PL_FLAG_WC;
156 		break;
157 	case AMDGPU_PL_GDS:
158 	case AMDGPU_PL_GWS:
159 	case AMDGPU_PL_OA:
160 		/* On-chip GDS memory*/
161 		man->func = &ttm_bo_manager_func;
162 		man->gpu_offset = 0;
163 		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_CMA;
164 		man->available_caching = TTM_PL_FLAG_UNCACHED;
165 		man->default_caching = TTM_PL_FLAG_UNCACHED;
166 		break;
167 	default:
168 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
169 		return -EINVAL;
170 	}
171 	return 0;
172 }
173 
amdgpu_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)174 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
175 				struct ttm_placement *placement)
176 {
177 	struct amdgpu_bo *rbo;
178 	static struct ttm_place placements = {
179 		.fpfn = 0,
180 		.lpfn = 0,
181 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
182 	};
183 
184 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
185 		placement->placement = &placements;
186 		placement->busy_placement = &placements;
187 		placement->num_placement = 1;
188 		placement->num_busy_placement = 1;
189 		return;
190 	}
191 	rbo = container_of(bo, struct amdgpu_bo, tbo);
192 	switch (bo->mem.mem_type) {
193 	case TTM_PL_VRAM:
194 		if (rbo->adev->mman.buffer_funcs_ring->ready == false)
195 			amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
196 		else
197 			amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_GTT);
198 		break;
199 	case TTM_PL_TT:
200 	default:
201 		amdgpu_ttm_placement_from_domain(rbo, AMDGPU_GEM_DOMAIN_CPU);
202 	}
203 	*placement = rbo->placement;
204 }
205 
amdgpu_verify_access(struct ttm_buffer_object * bo,struct file * filp)206 static int amdgpu_verify_access(struct ttm_buffer_object *bo, struct file *filp)
207 {
208 	struct amdgpu_bo *rbo = container_of(bo, struct amdgpu_bo, tbo);
209 
210 	return drm_vma_node_verify_access(&rbo->gem_base.vma_node, filp);
211 }
212 
amdgpu_move_null(struct ttm_buffer_object * bo,struct ttm_mem_reg * new_mem)213 static void amdgpu_move_null(struct ttm_buffer_object *bo,
214 			     struct ttm_mem_reg *new_mem)
215 {
216 	struct ttm_mem_reg *old_mem = &bo->mem;
217 
218 	BUG_ON(old_mem->mm_node != NULL);
219 	*old_mem = *new_mem;
220 	new_mem->mm_node = NULL;
221 }
222 
amdgpu_move_blit(struct ttm_buffer_object * bo,bool evict,bool no_wait_gpu,struct ttm_mem_reg * new_mem,struct ttm_mem_reg * old_mem)223 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
224 			bool evict, bool no_wait_gpu,
225 			struct ttm_mem_reg *new_mem,
226 			struct ttm_mem_reg *old_mem)
227 {
228 	struct amdgpu_device *adev;
229 	struct amdgpu_ring *ring;
230 	uint64_t old_start, new_start;
231 	struct fence *fence;
232 	int r;
233 
234 	adev = amdgpu_get_adev(bo->bdev);
235 	ring = adev->mman.buffer_funcs_ring;
236 	old_start = (u64)old_mem->start << PAGE_SHIFT;
237 	new_start = (u64)new_mem->start << PAGE_SHIFT;
238 
239 	switch (old_mem->mem_type) {
240 	case TTM_PL_VRAM:
241 		old_start += adev->mc.vram_start;
242 		break;
243 	case TTM_PL_TT:
244 		old_start += adev->mc.gtt_start;
245 		break;
246 	default:
247 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
248 		return -EINVAL;
249 	}
250 	switch (new_mem->mem_type) {
251 	case TTM_PL_VRAM:
252 		new_start += adev->mc.vram_start;
253 		break;
254 	case TTM_PL_TT:
255 		new_start += adev->mc.gtt_start;
256 		break;
257 	default:
258 		DRM_ERROR("Unknown placement %d\n", old_mem->mem_type);
259 		return -EINVAL;
260 	}
261 	if (!ring->ready) {
262 		DRM_ERROR("Trying to move memory with ring turned off.\n");
263 		return -EINVAL;
264 	}
265 
266 	BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
267 
268 	r = amdgpu_copy_buffer(ring, old_start, new_start,
269 			       new_mem->num_pages * PAGE_SIZE, /* bytes */
270 			       bo->resv, &fence);
271 	/* FIXME: handle copy error */
272 	r = ttm_bo_move_accel_cleanup(bo, fence,
273 				      evict, no_wait_gpu, new_mem);
274 	fence_put(fence);
275 	return r;
276 }
277 
amdgpu_move_vram_ram(struct ttm_buffer_object * bo,bool evict,bool interruptible,bool no_wait_gpu,struct ttm_mem_reg * new_mem)278 static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
279 				bool evict, bool interruptible,
280 				bool no_wait_gpu,
281 				struct ttm_mem_reg *new_mem)
282 {
283 	struct amdgpu_device *adev;
284 	struct ttm_mem_reg *old_mem = &bo->mem;
285 	struct ttm_mem_reg tmp_mem;
286 	struct ttm_place placements;
287 	struct ttm_placement placement;
288 	int r;
289 
290 	adev = amdgpu_get_adev(bo->bdev);
291 	tmp_mem = *new_mem;
292 	tmp_mem.mm_node = NULL;
293 	placement.num_placement = 1;
294 	placement.placement = &placements;
295 	placement.num_busy_placement = 1;
296 	placement.busy_placement = &placements;
297 	placements.fpfn = 0;
298 	placements.lpfn = 0;
299 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
300 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
301 			     interruptible, no_wait_gpu);
302 	if (unlikely(r)) {
303 		return r;
304 	}
305 
306 	r = ttm_tt_set_placement_caching(bo->ttm, tmp_mem.placement);
307 	if (unlikely(r)) {
308 		goto out_cleanup;
309 	}
310 
311 	r = ttm_tt_bind(bo->ttm, &tmp_mem);
312 	if (unlikely(r)) {
313 		goto out_cleanup;
314 	}
315 	r = amdgpu_move_blit(bo, true, no_wait_gpu, &tmp_mem, old_mem);
316 	if (unlikely(r)) {
317 		goto out_cleanup;
318 	}
319 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
320 out_cleanup:
321 	ttm_bo_mem_put(bo, &tmp_mem);
322 	return r;
323 }
324 
amdgpu_move_ram_vram(struct ttm_buffer_object * bo,bool evict,bool interruptible,bool no_wait_gpu,struct ttm_mem_reg * new_mem)325 static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
326 				bool evict, bool interruptible,
327 				bool no_wait_gpu,
328 				struct ttm_mem_reg *new_mem)
329 {
330 	struct amdgpu_device *adev;
331 	struct ttm_mem_reg *old_mem = &bo->mem;
332 	struct ttm_mem_reg tmp_mem;
333 	struct ttm_placement placement;
334 	struct ttm_place placements;
335 	int r;
336 
337 	adev = amdgpu_get_adev(bo->bdev);
338 	tmp_mem = *new_mem;
339 	tmp_mem.mm_node = NULL;
340 	placement.num_placement = 1;
341 	placement.placement = &placements;
342 	placement.num_busy_placement = 1;
343 	placement.busy_placement = &placements;
344 	placements.fpfn = 0;
345 	placements.lpfn = 0;
346 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
347 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
348 			     interruptible, no_wait_gpu);
349 	if (unlikely(r)) {
350 		return r;
351 	}
352 	r = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
353 	if (unlikely(r)) {
354 		goto out_cleanup;
355 	}
356 	r = amdgpu_move_blit(bo, true, no_wait_gpu, new_mem, old_mem);
357 	if (unlikely(r)) {
358 		goto out_cleanup;
359 	}
360 out_cleanup:
361 	ttm_bo_mem_put(bo, &tmp_mem);
362 	return r;
363 }
364 
amdgpu_bo_move(struct ttm_buffer_object * bo,bool evict,bool interruptible,bool no_wait_gpu,struct ttm_mem_reg * new_mem)365 static int amdgpu_bo_move(struct ttm_buffer_object *bo,
366 			bool evict, bool interruptible,
367 			bool no_wait_gpu,
368 			struct ttm_mem_reg *new_mem)
369 {
370 	struct amdgpu_device *adev;
371 	struct ttm_mem_reg *old_mem = &bo->mem;
372 	int r;
373 
374 	adev = amdgpu_get_adev(bo->bdev);
375 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
376 		amdgpu_move_null(bo, new_mem);
377 		return 0;
378 	}
379 	if ((old_mem->mem_type == TTM_PL_TT &&
380 	     new_mem->mem_type == TTM_PL_SYSTEM) ||
381 	    (old_mem->mem_type == TTM_PL_SYSTEM &&
382 	     new_mem->mem_type == TTM_PL_TT)) {
383 		/* bind is enough */
384 		amdgpu_move_null(bo, new_mem);
385 		return 0;
386 	}
387 	if (adev->mman.buffer_funcs == NULL ||
388 	    adev->mman.buffer_funcs_ring == NULL ||
389 	    !adev->mman.buffer_funcs_ring->ready) {
390 		/* use memcpy */
391 		goto memcpy;
392 	}
393 
394 	if (old_mem->mem_type == TTM_PL_VRAM &&
395 	    new_mem->mem_type == TTM_PL_SYSTEM) {
396 		r = amdgpu_move_vram_ram(bo, evict, interruptible,
397 					no_wait_gpu, new_mem);
398 	} else if (old_mem->mem_type == TTM_PL_SYSTEM &&
399 		   new_mem->mem_type == TTM_PL_VRAM) {
400 		r = amdgpu_move_ram_vram(bo, evict, interruptible,
401 					    no_wait_gpu, new_mem);
402 	} else {
403 		r = amdgpu_move_blit(bo, evict, no_wait_gpu, new_mem, old_mem);
404 	}
405 
406 	if (r) {
407 memcpy:
408 		r = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
409 		if (r) {
410 			return r;
411 		}
412 	}
413 
414 	/* update statistics */
415 	atomic64_add((u64)bo->num_pages << PAGE_SHIFT, &adev->num_bytes_moved);
416 	return 0;
417 }
418 
amdgpu_ttm_io_mem_reserve(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)419 static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
420 {
421 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
422 	struct amdgpu_device *adev = amdgpu_get_adev(bdev);
423 
424 	mem->bus.addr = NULL;
425 	mem->bus.offset = 0;
426 	mem->bus.size = mem->num_pages << PAGE_SHIFT;
427 	mem->bus.base = 0;
428 	mem->bus.is_iomem = false;
429 	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
430 		return -EINVAL;
431 	switch (mem->mem_type) {
432 	case TTM_PL_SYSTEM:
433 		/* system memory */
434 		return 0;
435 	case TTM_PL_TT:
436 		break;
437 	case TTM_PL_VRAM:
438 		mem->bus.offset = mem->start << PAGE_SHIFT;
439 		/* check if it's visible */
440 		if ((mem->bus.offset + mem->bus.size) > adev->mc.visible_vram_size)
441 			return -EINVAL;
442 		mem->bus.base = adev->mc.aper_base;
443 		mem->bus.is_iomem = true;
444 #ifdef __alpha__
445 		/*
446 		 * Alpha: use bus.addr to hold the ioremap() return,
447 		 * so we can modify bus.base below.
448 		 */
449 		if (mem->placement & TTM_PL_FLAG_WC)
450 			mem->bus.addr =
451 				ioremap_wc(mem->bus.base + mem->bus.offset,
452 					   mem->bus.size);
453 		else
454 			mem->bus.addr =
455 				ioremap_nocache(mem->bus.base + mem->bus.offset,
456 						mem->bus.size);
457 
458 		/*
459 		 * Alpha: Use just the bus offset plus
460 		 * the hose/domain memory base for bus.base.
461 		 * It then can be used to build PTEs for VRAM
462 		 * access, as done in ttm_bo_vm_fault().
463 		 */
464 		mem->bus.base = (mem->bus.base & 0x0ffffffffUL) +
465 			adev->ddev->hose->dense_mem_base;
466 #endif
467 		break;
468 	default:
469 		return -EINVAL;
470 	}
471 	return 0;
472 }
473 
amdgpu_ttm_io_mem_free(struct ttm_bo_device * bdev,struct ttm_mem_reg * mem)474 static void amdgpu_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
475 {
476 }
477 
478 /*
479  * TTM backend functions.
480  */
481 struct amdgpu_ttm_tt {
482 	struct ttm_dma_tt		ttm;
483 	struct amdgpu_device		*adev;
484 	u64				offset;
485 	uint64_t			userptr;
486 	struct mm_struct		*usermm;
487 	uint32_t			userflags;
488 };
489 
490 /* prepare the sg table with the user pages */
amdgpu_ttm_tt_pin_userptr(struct ttm_tt * ttm)491 static int amdgpu_ttm_tt_pin_userptr(struct ttm_tt *ttm)
492 {
493 	struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
494 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
495 	unsigned pinned = 0, nents;
496 	int r;
497 
498 	int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
499 	unsigned int flags = 0;
500 	enum dma_data_direction direction = write ?
501 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
502 
503 	if (write)
504 		flags |= FOLL_WRITE;
505 
506 	if (current->mm != gtt->usermm)
507 		return -EPERM;
508 
509 	if (gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) {
510 		/* check that we only pin down anonymous memory
511 		   to prevent problems with writeback */
512 		unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
513 		struct vm_area_struct *vma;
514 
515 		vma = find_vma(gtt->usermm, gtt->userptr);
516 		if (!vma || vma->vm_file || vma->vm_end < end)
517 			return -EPERM;
518 	}
519 
520 	do {
521 		unsigned num_pages = ttm->num_pages - pinned;
522 		uint64_t userptr = gtt->userptr + pinned * PAGE_SIZE;
523 		struct page **pages = ttm->pages + pinned;
524 
525 		r = get_user_pages(current, current->mm, userptr, num_pages,
526 				   flags, pages, NULL);
527 		if (r < 0)
528 			goto release_pages;
529 
530 		pinned += r;
531 
532 	} while (pinned < ttm->num_pages);
533 
534 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
535 				      ttm->num_pages << PAGE_SHIFT,
536 				      GFP_KERNEL);
537 	if (r)
538 		goto release_sg;
539 
540 	r = -ENOMEM;
541 	nents = dma_map_sg(adev->dev, ttm->sg->sgl, ttm->sg->nents, direction);
542 	if (nents != ttm->sg->nents)
543 		goto release_sg;
544 
545 	drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
546 					 gtt->ttm.dma_address, ttm->num_pages);
547 
548 	return 0;
549 
550 release_sg:
551 	kfree(ttm->sg);
552 
553 release_pages:
554 	release_pages(ttm->pages, pinned, 0);
555 	return r;
556 }
557 
amdgpu_ttm_tt_unpin_userptr(struct ttm_tt * ttm)558 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
559 {
560 	struct amdgpu_device *adev = amdgpu_get_adev(ttm->bdev);
561 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
562 	struct sg_page_iter sg_iter;
563 
564 	int write = !(gtt->userflags & AMDGPU_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 || !ttm->sg->sgl)
570 		return;
571 
572 	/* free the sg table and pages again */
573 	dma_unmap_sg(adev->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 page *page = sg_page_iter_page(&sg_iter);
577 		if (!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY))
578 			set_page_dirty(page);
579 
580 		mark_page_accessed(page);
581 		page_cache_release(page);
582 	}
583 
584 	sg_free_table(ttm->sg);
585 }
586 
amdgpu_ttm_backend_bind(struct ttm_tt * ttm,struct ttm_mem_reg * bo_mem)587 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
588 				   struct ttm_mem_reg *bo_mem)
589 {
590 	struct amdgpu_ttm_tt *gtt = (void*)ttm;
591 	uint32_t flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
592 	int r;
593 
594 	if (gtt->userptr) {
595 		r = amdgpu_ttm_tt_pin_userptr(ttm);
596 		if (r) {
597 			DRM_ERROR("failed to pin userptr\n");
598 			return r;
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 
607 	if (bo_mem->mem_type == AMDGPU_PL_GDS ||
608 	    bo_mem->mem_type == AMDGPU_PL_GWS ||
609 	    bo_mem->mem_type == AMDGPU_PL_OA)
610 		return -EINVAL;
611 
612 	r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
613 		ttm->pages, gtt->ttm.dma_address, flags);
614 
615 	if (r) {
616 		DRM_ERROR("failed to bind %lu pages at 0x%08X\n",
617 			  ttm->num_pages, (unsigned)gtt->offset);
618 		return r;
619 	}
620 	return 0;
621 }
622 
amdgpu_ttm_backend_unbind(struct ttm_tt * ttm)623 static int amdgpu_ttm_backend_unbind(struct ttm_tt *ttm)
624 {
625 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
626 
627 	/* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
628 	if (gtt->adev->gart.ready)
629 		amdgpu_gart_unbind(gtt->adev, gtt->offset, ttm->num_pages);
630 
631 	if (gtt->userptr)
632 		amdgpu_ttm_tt_unpin_userptr(ttm);
633 
634 	return 0;
635 }
636 
amdgpu_ttm_backend_destroy(struct ttm_tt * ttm)637 static void amdgpu_ttm_backend_destroy(struct ttm_tt *ttm)
638 {
639 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
640 
641 	ttm_dma_tt_fini(&gtt->ttm);
642 	kfree(gtt);
643 }
644 
645 static struct ttm_backend_func amdgpu_backend_func = {
646 	.bind = &amdgpu_ttm_backend_bind,
647 	.unbind = &amdgpu_ttm_backend_unbind,
648 	.destroy = &amdgpu_ttm_backend_destroy,
649 };
650 
amdgpu_ttm_tt_create(struct ttm_bo_device * bdev,unsigned long size,uint32_t page_flags,struct page * dummy_read_page)651 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_bo_device *bdev,
652 				    unsigned long size, uint32_t page_flags,
653 				    struct page *dummy_read_page)
654 {
655 	struct amdgpu_device *adev;
656 	struct amdgpu_ttm_tt *gtt;
657 
658 	adev = amdgpu_get_adev(bdev);
659 
660 	gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
661 	if (gtt == NULL) {
662 		return NULL;
663 	}
664 	gtt->ttm.ttm.func = &amdgpu_backend_func;
665 	gtt->adev = adev;
666 	if (ttm_dma_tt_init(&gtt->ttm, bdev, size, page_flags, dummy_read_page)) {
667 		kfree(gtt);
668 		return NULL;
669 	}
670 	return &gtt->ttm.ttm;
671 }
672 
amdgpu_ttm_tt_populate(struct ttm_tt * ttm)673 static int amdgpu_ttm_tt_populate(struct ttm_tt *ttm)
674 {
675 	struct amdgpu_device *adev;
676 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
677 	unsigned i;
678 	int r;
679 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
680 
681 	if (ttm->state != tt_unpopulated)
682 		return 0;
683 
684 	if (gtt && gtt->userptr) {
685 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
686 		if (!ttm->sg)
687 			return -ENOMEM;
688 
689 		ttm->page_flags |= TTM_PAGE_FLAG_SG;
690 		ttm->state = tt_unbound;
691 		return 0;
692 	}
693 
694 	if (slave && ttm->sg) {
695 		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
696 						 gtt->ttm.dma_address, ttm->num_pages);
697 		ttm->state = tt_unbound;
698 		return 0;
699 	}
700 
701 	adev = amdgpu_get_adev(ttm->bdev);
702 
703 #ifdef CONFIG_SWIOTLB
704 	if (swiotlb_nr_tbl()) {
705 		return ttm_dma_populate(&gtt->ttm, adev->dev);
706 	}
707 #endif
708 
709 	r = ttm_pool_populate(ttm);
710 	if (r) {
711 		return r;
712 	}
713 
714 	for (i = 0; i < ttm->num_pages; i++) {
715 		gtt->ttm.dma_address[i] = pci_map_page(adev->pdev, ttm->pages[i],
716 						       0, PAGE_SIZE,
717 						       PCI_DMA_BIDIRECTIONAL);
718 		if (pci_dma_mapping_error(adev->pdev, gtt->ttm.dma_address[i])) {
719 			while (i--) {
720 				pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
721 					       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
722 				gtt->ttm.dma_address[i] = 0;
723 			}
724 			ttm_pool_unpopulate(ttm);
725 			return -EFAULT;
726 		}
727 	}
728 	return 0;
729 }
730 
amdgpu_ttm_tt_unpopulate(struct ttm_tt * ttm)731 static void amdgpu_ttm_tt_unpopulate(struct ttm_tt *ttm)
732 {
733 	struct amdgpu_device *adev;
734 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
735 	unsigned i;
736 	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
737 
738 	if (gtt && gtt->userptr) {
739 		kfree(ttm->sg);
740 		ttm->sg = NULL;
741 		ttm->page_flags &= ~TTM_PAGE_FLAG_SG;
742 		return;
743 	}
744 
745 	if (slave)
746 		return;
747 
748 	adev = amdgpu_get_adev(ttm->bdev);
749 
750 #ifdef CONFIG_SWIOTLB
751 	if (swiotlb_nr_tbl()) {
752 		ttm_dma_unpopulate(&gtt->ttm, adev->dev);
753 		return;
754 	}
755 #endif
756 
757 	for (i = 0; i < ttm->num_pages; i++) {
758 		if (gtt->ttm.dma_address[i]) {
759 			pci_unmap_page(adev->pdev, gtt->ttm.dma_address[i],
760 				       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
761 		}
762 	}
763 
764 	ttm_pool_unpopulate(ttm);
765 }
766 
amdgpu_ttm_tt_set_userptr(struct ttm_tt * ttm,uint64_t addr,uint32_t flags)767 int amdgpu_ttm_tt_set_userptr(struct ttm_tt *ttm, uint64_t addr,
768 			      uint32_t flags)
769 {
770 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
771 
772 	if (gtt == NULL)
773 		return -EINVAL;
774 
775 	gtt->userptr = addr;
776 	gtt->usermm = current->mm;
777 	gtt->userflags = flags;
778 	return 0;
779 }
780 
amdgpu_ttm_tt_has_userptr(struct ttm_tt * ttm)781 bool amdgpu_ttm_tt_has_userptr(struct ttm_tt *ttm)
782 {
783 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
784 
785 	if (gtt == NULL)
786 		return false;
787 
788 	return !!gtt->userptr;
789 }
790 
amdgpu_ttm_tt_affect_userptr(struct ttm_tt * ttm,unsigned long start,unsigned long end)791 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
792 				  unsigned long end)
793 {
794 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
795 	unsigned long size;
796 
797 	if (gtt == NULL)
798 		return false;
799 
800 	if (gtt->ttm.ttm.state != tt_bound || !gtt->userptr)
801 		return false;
802 
803 	size = (unsigned long)gtt->ttm.ttm.num_pages * PAGE_SIZE;
804 	if (gtt->userptr > end || gtt->userptr + size <= start)
805 		return false;
806 
807 	return true;
808 }
809 
amdgpu_ttm_tt_is_readonly(struct ttm_tt * ttm)810 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
811 {
812 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
813 
814 	if (gtt == NULL)
815 		return false;
816 
817 	return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
818 }
819 
amdgpu_ttm_tt_pte_flags(struct amdgpu_device * adev,struct ttm_tt * ttm,struct ttm_mem_reg * mem)820 uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
821 				 struct ttm_mem_reg *mem)
822 {
823 	uint32_t flags = 0;
824 
825 	if (mem && mem->mem_type != TTM_PL_SYSTEM)
826 		flags |= AMDGPU_PTE_VALID;
827 
828 	if (mem && mem->mem_type == TTM_PL_TT) {
829 		flags |= AMDGPU_PTE_SYSTEM;
830 
831 		if (ttm->caching_state == tt_cached)
832 			flags |= AMDGPU_PTE_SNOOPED;
833 	}
834 
835 	if (adev->asic_type >= CHIP_TONGA)
836 		flags |= AMDGPU_PTE_EXECUTABLE;
837 
838 	flags |= AMDGPU_PTE_READABLE;
839 
840 	if (!amdgpu_ttm_tt_is_readonly(ttm))
841 		flags |= AMDGPU_PTE_WRITEABLE;
842 
843 	return flags;
844 }
845 
846 static struct ttm_bo_driver amdgpu_bo_driver = {
847 	.ttm_tt_create = &amdgpu_ttm_tt_create,
848 	.ttm_tt_populate = &amdgpu_ttm_tt_populate,
849 	.ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
850 	.invalidate_caches = &amdgpu_invalidate_caches,
851 	.init_mem_type = &amdgpu_init_mem_type,
852 	.evict_flags = &amdgpu_evict_flags,
853 	.move = &amdgpu_bo_move,
854 	.verify_access = &amdgpu_verify_access,
855 	.move_notify = &amdgpu_bo_move_notify,
856 	.fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
857 	.io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
858 	.io_mem_free = &amdgpu_ttm_io_mem_free,
859 };
860 
amdgpu_ttm_init(struct amdgpu_device * adev)861 int amdgpu_ttm_init(struct amdgpu_device *adev)
862 {
863 	int r;
864 
865 	r = amdgpu_ttm_global_init(adev);
866 	if (r) {
867 		return r;
868 	}
869 	/* No others user of address space so set it to 0 */
870 	r = ttm_bo_device_init(&adev->mman.bdev,
871 			       adev->mman.bo_global_ref.ref.object,
872 			       &amdgpu_bo_driver,
873 			       adev->ddev->anon_inode->i_mapping,
874 			       DRM_FILE_PAGE_OFFSET,
875 			       adev->need_dma32);
876 	if (r) {
877 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
878 		return r;
879 	}
880 	adev->mman.initialized = true;
881 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
882 				adev->mc.real_vram_size >> PAGE_SHIFT);
883 	if (r) {
884 		DRM_ERROR("Failed initializing VRAM heap.\n");
885 		return r;
886 	}
887 	/* Change the size here instead of the init above so only lpfn is affected */
888 	amdgpu_ttm_set_active_vram_size(adev, adev->mc.visible_vram_size);
889 
890 	r = amdgpu_bo_create(adev, 256 * 1024, PAGE_SIZE, true,
891 			     AMDGPU_GEM_DOMAIN_VRAM,
892 			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
893 			     NULL, NULL, &adev->stollen_vga_memory);
894 	if (r) {
895 		return r;
896 	}
897 	r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
898 	if (r)
899 		return r;
900 	r = amdgpu_bo_pin(adev->stollen_vga_memory, AMDGPU_GEM_DOMAIN_VRAM, NULL);
901 	amdgpu_bo_unreserve(adev->stollen_vga_memory);
902 	if (r) {
903 		amdgpu_bo_unref(&adev->stollen_vga_memory);
904 		return r;
905 	}
906 	DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
907 		 (unsigned) (adev->mc.real_vram_size / (1024 * 1024)));
908 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_TT,
909 				adev->mc.gtt_size >> PAGE_SHIFT);
910 	if (r) {
911 		DRM_ERROR("Failed initializing GTT heap.\n");
912 		return r;
913 	}
914 	DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
915 		 (unsigned)(adev->mc.gtt_size / (1024 * 1024)));
916 
917 	adev->gds.mem.total_size = adev->gds.mem.total_size << AMDGPU_GDS_SHIFT;
918 	adev->gds.mem.gfx_partition_size = adev->gds.mem.gfx_partition_size << AMDGPU_GDS_SHIFT;
919 	adev->gds.mem.cs_partition_size = adev->gds.mem.cs_partition_size << AMDGPU_GDS_SHIFT;
920 	adev->gds.gws.total_size = adev->gds.gws.total_size << AMDGPU_GWS_SHIFT;
921 	adev->gds.gws.gfx_partition_size = adev->gds.gws.gfx_partition_size << AMDGPU_GWS_SHIFT;
922 	adev->gds.gws.cs_partition_size = adev->gds.gws.cs_partition_size << AMDGPU_GWS_SHIFT;
923 	adev->gds.oa.total_size = adev->gds.oa.total_size << AMDGPU_OA_SHIFT;
924 	adev->gds.oa.gfx_partition_size = adev->gds.oa.gfx_partition_size << AMDGPU_OA_SHIFT;
925 	adev->gds.oa.cs_partition_size = adev->gds.oa.cs_partition_size << AMDGPU_OA_SHIFT;
926 	/* GDS Memory */
927 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GDS,
928 				adev->gds.mem.total_size >> PAGE_SHIFT);
929 	if (r) {
930 		DRM_ERROR("Failed initializing GDS heap.\n");
931 		return r;
932 	}
933 
934 	/* GWS */
935 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_GWS,
936 				adev->gds.gws.total_size >> PAGE_SHIFT);
937 	if (r) {
938 		DRM_ERROR("Failed initializing gws heap.\n");
939 		return r;
940 	}
941 
942 	/* OA */
943 	r = ttm_bo_init_mm(&adev->mman.bdev, AMDGPU_PL_OA,
944 				adev->gds.oa.total_size >> PAGE_SHIFT);
945 	if (r) {
946 		DRM_ERROR("Failed initializing oa heap.\n");
947 		return r;
948 	}
949 
950 	r = amdgpu_ttm_debugfs_init(adev);
951 	if (r) {
952 		DRM_ERROR("Failed to init debugfs\n");
953 		return r;
954 	}
955 	return 0;
956 }
957 
amdgpu_ttm_fini(struct amdgpu_device * adev)958 void amdgpu_ttm_fini(struct amdgpu_device *adev)
959 {
960 	int r;
961 
962 	if (!adev->mman.initialized)
963 		return;
964 	amdgpu_ttm_debugfs_fini(adev);
965 	if (adev->stollen_vga_memory) {
966 		r = amdgpu_bo_reserve(adev->stollen_vga_memory, false);
967 		if (r == 0) {
968 			amdgpu_bo_unpin(adev->stollen_vga_memory);
969 			amdgpu_bo_unreserve(adev->stollen_vga_memory);
970 		}
971 		amdgpu_bo_unref(&adev->stollen_vga_memory);
972 	}
973 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
974 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
975 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GDS);
976 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_GWS);
977 	ttm_bo_clean_mm(&adev->mman.bdev, AMDGPU_PL_OA);
978 	ttm_bo_device_release(&adev->mman.bdev);
979 	amdgpu_gart_fini(adev);
980 	amdgpu_ttm_global_fini(adev);
981 	adev->mman.initialized = false;
982 	DRM_INFO("amdgpu: ttm finalized\n");
983 }
984 
985 /* this should only be called at bootup or when userspace
986  * isn't running */
amdgpu_ttm_set_active_vram_size(struct amdgpu_device * adev,u64 size)987 void amdgpu_ttm_set_active_vram_size(struct amdgpu_device *adev, u64 size)
988 {
989 	struct ttm_mem_type_manager *man;
990 
991 	if (!adev->mman.initialized)
992 		return;
993 
994 	man = &adev->mman.bdev.man[TTM_PL_VRAM];
995 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
996 	man->size = size >> PAGE_SHIFT;
997 }
998 
amdgpu_mmap(struct file * filp,struct vm_area_struct * vma)999 int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
1000 {
1001 	struct drm_file *file_priv;
1002 	struct amdgpu_device *adev;
1003 
1004 	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET))
1005 		return -EINVAL;
1006 
1007 	file_priv = filp->private_data;
1008 	adev = file_priv->minor->dev->dev_private;
1009 	if (adev == NULL)
1010 		return -EINVAL;
1011 
1012 	return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
1013 }
1014 
amdgpu_copy_buffer(struct amdgpu_ring * ring,uint64_t src_offset,uint64_t dst_offset,uint32_t byte_count,struct reservation_object * resv,struct fence ** fence)1015 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
1016 		       uint64_t src_offset,
1017 		       uint64_t dst_offset,
1018 		       uint32_t byte_count,
1019 		       struct reservation_object *resv,
1020 		       struct fence **fence)
1021 {
1022 	struct amdgpu_device *adev = ring->adev;
1023 	uint32_t max_bytes;
1024 	unsigned num_loops, num_dw;
1025 	struct amdgpu_ib *ib;
1026 	unsigned i;
1027 	int r;
1028 
1029 	max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1030 	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1031 	num_dw = num_loops * adev->mman.buffer_funcs->copy_num_dw;
1032 
1033 	/* for IB padding */
1034 	while (num_dw & 0x7)
1035 		num_dw++;
1036 
1037 	ib = kzalloc(sizeof(struct amdgpu_ib), GFP_KERNEL);
1038 	if (!ib)
1039 		return -ENOMEM;
1040 
1041 	r = amdgpu_ib_get(ring, NULL, num_dw * 4, ib);
1042 	if (r) {
1043 		kfree(ib);
1044 		return r;
1045 	}
1046 
1047 	ib->length_dw = 0;
1048 
1049 	if (resv) {
1050 		r = amdgpu_sync_resv(adev, &ib->sync, resv,
1051 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1052 		if (r) {
1053 			DRM_ERROR("sync failed (%d).\n", r);
1054 			goto error_free;
1055 		}
1056 	}
1057 
1058 	for (i = 0; i < num_loops; i++) {
1059 		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1060 
1061 		amdgpu_emit_copy_buffer(adev, ib, src_offset, dst_offset,
1062 					cur_size_in_bytes);
1063 
1064 		src_offset += cur_size_in_bytes;
1065 		dst_offset += cur_size_in_bytes;
1066 		byte_count -= cur_size_in_bytes;
1067 	}
1068 
1069 	amdgpu_vm_pad_ib(adev, ib);
1070 	WARN_ON(ib->length_dw > num_dw);
1071 	r = amdgpu_sched_ib_submit_kernel_helper(adev, ring, ib, 1,
1072 						 &amdgpu_vm_free_job,
1073 						 AMDGPU_FENCE_OWNER_UNDEFINED,
1074 						 fence);
1075 	if (r)
1076 		goto error_free;
1077 
1078 	if (!amdgpu_enable_scheduler) {
1079 		amdgpu_ib_free(adev, ib);
1080 		kfree(ib);
1081 	}
1082 	return 0;
1083 error_free:
1084 	amdgpu_ib_free(adev, ib);
1085 	kfree(ib);
1086 	return r;
1087 }
1088 
1089 #if defined(CONFIG_DEBUG_FS)
1090 
amdgpu_mm_dump_table(struct seq_file * m,void * data)1091 static int amdgpu_mm_dump_table(struct seq_file *m, void *data)
1092 {
1093 	struct drm_info_node *node = (struct drm_info_node *)m->private;
1094 	unsigned ttm_pl = *(int *)node->info_ent->data;
1095 	struct drm_device *dev = node->minor->dev;
1096 	struct amdgpu_device *adev = dev->dev_private;
1097 	struct drm_mm *mm = (struct drm_mm *)adev->mman.bdev.man[ttm_pl].priv;
1098 	int ret;
1099 	struct ttm_bo_global *glob = adev->mman.bdev.glob;
1100 
1101 	spin_lock(&glob->lru_lock);
1102 	ret = drm_mm_dump_table(m, mm);
1103 	spin_unlock(&glob->lru_lock);
1104 	if (ttm_pl == TTM_PL_VRAM)
1105 		seq_printf(m, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n",
1106 			   adev->mman.bdev.man[ttm_pl].size,
1107 			   (u64)atomic64_read(&adev->vram_usage) >> 20,
1108 			   (u64)atomic64_read(&adev->vram_vis_usage) >> 20);
1109 	return ret;
1110 }
1111 
1112 static int ttm_pl_vram = TTM_PL_VRAM;
1113 static int ttm_pl_tt = TTM_PL_TT;
1114 
1115 static struct drm_info_list amdgpu_ttm_debugfs_list[] = {
1116 	{"amdgpu_vram_mm", amdgpu_mm_dump_table, 0, &ttm_pl_vram},
1117 	{"amdgpu_gtt_mm", amdgpu_mm_dump_table, 0, &ttm_pl_tt},
1118 	{"ttm_page_pool", ttm_page_alloc_debugfs, 0, NULL},
1119 #ifdef CONFIG_SWIOTLB
1120 	{"ttm_dma_page_pool", ttm_dma_page_alloc_debugfs, 0, NULL}
1121 #endif
1122 };
1123 
amdgpu_ttm_vram_read(struct file * f,char __user * buf,size_t size,loff_t * pos)1124 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
1125 				    size_t size, loff_t *pos)
1126 {
1127 	struct amdgpu_device *adev = f->f_inode->i_private;
1128 	ssize_t result = 0;
1129 	int r;
1130 
1131 	if (size & 0x3 || *pos & 0x3)
1132 		return -EINVAL;
1133 
1134 	if (*pos >= adev->mc.mc_vram_size)
1135 		return -ENXIO;
1136 
1137 	while (size) {
1138 		unsigned long flags;
1139 		uint32_t value;
1140 
1141 		if (*pos >= adev->mc.mc_vram_size)
1142 			return result;
1143 
1144 		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
1145 		WREG32(mmMM_INDEX, ((uint32_t)*pos) | 0x80000000);
1146 		WREG32(mmMM_INDEX_HI, *pos >> 31);
1147 		value = RREG32(mmMM_DATA);
1148 		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
1149 
1150 		r = put_user(value, (uint32_t *)buf);
1151 		if (r)
1152 			return r;
1153 
1154 		result += 4;
1155 		buf += 4;
1156 		*pos += 4;
1157 		size -= 4;
1158 	}
1159 
1160 	return result;
1161 }
1162 
1163 static const struct file_operations amdgpu_ttm_vram_fops = {
1164 	.owner = THIS_MODULE,
1165 	.read = amdgpu_ttm_vram_read,
1166 	.llseek = default_llseek
1167 };
1168 
amdgpu_ttm_gtt_read(struct file * f,char __user * buf,size_t size,loff_t * pos)1169 static ssize_t amdgpu_ttm_gtt_read(struct file *f, char __user *buf,
1170 				   size_t size, loff_t *pos)
1171 {
1172 	struct amdgpu_device *adev = f->f_inode->i_private;
1173 	ssize_t result = 0;
1174 	int r;
1175 
1176 	while (size) {
1177 		loff_t p = *pos / PAGE_SIZE;
1178 		unsigned off = *pos & ~PAGE_MASK;
1179 		size_t cur_size = min_t(size_t, size, PAGE_SIZE - off);
1180 		struct page *page;
1181 		void *ptr;
1182 
1183 		if (p >= adev->gart.num_cpu_pages)
1184 			return result;
1185 
1186 		page = adev->gart.pages[p];
1187 		if (page) {
1188 			ptr = kmap(page);
1189 			ptr += off;
1190 
1191 			r = copy_to_user(buf, ptr, cur_size);
1192 			kunmap(adev->gart.pages[p]);
1193 		} else
1194 			r = clear_user(buf, cur_size);
1195 
1196 		if (r)
1197 			return -EFAULT;
1198 
1199 		result += cur_size;
1200 		buf += cur_size;
1201 		*pos += cur_size;
1202 		size -= cur_size;
1203 	}
1204 
1205 	return result;
1206 }
1207 
1208 static const struct file_operations amdgpu_ttm_gtt_fops = {
1209 	.owner = THIS_MODULE,
1210 	.read = amdgpu_ttm_gtt_read,
1211 	.llseek = default_llseek
1212 };
1213 
1214 #endif
1215 
amdgpu_ttm_debugfs_init(struct amdgpu_device * adev)1216 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
1217 {
1218 #if defined(CONFIG_DEBUG_FS)
1219 	unsigned count;
1220 
1221 	struct drm_minor *minor = adev->ddev->primary;
1222 	struct dentry *ent, *root = minor->debugfs_root;
1223 
1224 	ent = debugfs_create_file("amdgpu_vram", S_IFREG | S_IRUGO, root,
1225 				  adev, &amdgpu_ttm_vram_fops);
1226 	if (IS_ERR(ent))
1227 		return PTR_ERR(ent);
1228 	i_size_write(ent->d_inode, adev->mc.mc_vram_size);
1229 	adev->mman.vram = ent;
1230 
1231 	ent = debugfs_create_file("amdgpu_gtt", S_IFREG | S_IRUGO, root,
1232 				  adev, &amdgpu_ttm_gtt_fops);
1233 	if (IS_ERR(ent))
1234 		return PTR_ERR(ent);
1235 	i_size_write(ent->d_inode, adev->mc.gtt_size);
1236 	adev->mman.gtt = ent;
1237 
1238 	count = ARRAY_SIZE(amdgpu_ttm_debugfs_list);
1239 
1240 #ifdef CONFIG_SWIOTLB
1241 	if (!swiotlb_nr_tbl())
1242 		--count;
1243 #endif
1244 
1245 	return amdgpu_debugfs_add_files(adev, amdgpu_ttm_debugfs_list, count);
1246 #else
1247 
1248 	return 0;
1249 #endif
1250 }
1251 
amdgpu_ttm_debugfs_fini(struct amdgpu_device * adev)1252 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev)
1253 {
1254 #if defined(CONFIG_DEBUG_FS)
1255 
1256 	debugfs_remove(adev->mman.vram);
1257 	adev->mman.vram = NULL;
1258 
1259 	debugfs_remove(adev->mman.gtt);
1260 	adev->mman.gtt = NULL;
1261 #endif
1262 }
1263