• 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 
33 #include <linux/dma-mapping.h>
34 #include <linux/iommu.h>
35 #include <linux/pagemap.h>
36 #include <linux/sched/task.h>
37 #include <linux/sched/mm.h>
38 #include <linux/seq_file.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swiotlb.h>
42 #include <linux/dma-buf.h>
43 #include <linux/sizes.h>
44 
45 #include <drm/ttm/ttm_bo_api.h>
46 #include <drm/ttm/ttm_bo_driver.h>
47 #include <drm/ttm/ttm_placement.h>
48 #include <drm/ttm/ttm_range_manager.h>
49 
50 #include <drm/amdgpu_drm.h>
51 
52 #include "amdgpu.h"
53 #include "amdgpu_object.h"
54 #include "amdgpu_trace.h"
55 #include "amdgpu_amdkfd.h"
56 #include "amdgpu_sdma.h"
57 #include "amdgpu_ras.h"
58 #include "amdgpu_atomfirmware.h"
59 #include "amdgpu_res_cursor.h"
60 #include "bif/bif_4_1_d.h"
61 
62 #define AMDGPU_TTM_VRAM_MAX_DW_READ	(size_t)128
63 
64 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
65 				   struct ttm_tt *ttm,
66 				   struct ttm_resource *bo_mem);
67 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
68 				      struct ttm_tt *ttm);
69 
amdgpu_ttm_init_on_chip(struct amdgpu_device * adev,unsigned int type,uint64_t size_in_page)70 static int amdgpu_ttm_init_on_chip(struct amdgpu_device *adev,
71 				    unsigned int type,
72 				    uint64_t size_in_page)
73 {
74 	return ttm_range_man_init(&adev->mman.bdev, type,
75 				  false, size_in_page);
76 }
77 
78 /**
79  * amdgpu_evict_flags - Compute placement flags
80  *
81  * @bo: The buffer object to evict
82  * @placement: Possible destination(s) for evicted BO
83  *
84  * Fill in placement data when ttm_bo_evict() is called
85  */
amdgpu_evict_flags(struct ttm_buffer_object * bo,struct ttm_placement * placement)86 static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
87 				struct ttm_placement *placement)
88 {
89 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
90 	struct amdgpu_bo *abo;
91 	static const struct ttm_place placements = {
92 		.fpfn = 0,
93 		.lpfn = 0,
94 		.mem_type = TTM_PL_SYSTEM,
95 		.flags = 0
96 	};
97 
98 	/* Don't handle scatter gather BOs */
99 	if (bo->type == ttm_bo_type_sg) {
100 		placement->num_placement = 0;
101 		placement->num_busy_placement = 0;
102 		return;
103 	}
104 
105 	/* Object isn't an AMDGPU object so ignore */
106 	if (!amdgpu_bo_is_amdgpu_bo(bo)) {
107 		placement->placement = &placements;
108 		placement->busy_placement = &placements;
109 		placement->num_placement = 1;
110 		placement->num_busy_placement = 1;
111 		return;
112 	}
113 
114 	abo = ttm_to_amdgpu_bo(bo);
115 	if (abo->flags & AMDGPU_AMDKFD_CREATE_SVM_BO) {
116 		struct dma_fence *fence;
117 		struct dma_resv *resv = &bo->base._resv;
118 
119 		rcu_read_lock();
120 		fence = rcu_dereference(resv->fence_excl);
121 		if (fence && !fence->ops->signaled)
122 			dma_fence_enable_sw_signaling(fence);
123 
124 		placement->num_placement = 0;
125 		placement->num_busy_placement = 0;
126 		rcu_read_unlock();
127 		return;
128 	}
129 
130 	switch (bo->resource->mem_type) {
131 	case AMDGPU_PL_GDS:
132 	case AMDGPU_PL_GWS:
133 	case AMDGPU_PL_OA:
134 		placement->num_placement = 0;
135 		placement->num_busy_placement = 0;
136 		return;
137 
138 	case TTM_PL_VRAM:
139 		if (!adev->mman.buffer_funcs_enabled) {
140 			/* Move to system memory */
141 			amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
142 		} else if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
143 			   !(abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
144 			   amdgpu_bo_in_cpu_visible_vram(abo)) {
145 
146 			/* Try evicting to the CPU inaccessible part of VRAM
147 			 * first, but only set GTT as busy placement, so this
148 			 * BO will be evicted to GTT rather than causing other
149 			 * BOs to be evicted from VRAM
150 			 */
151 			amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
152 							AMDGPU_GEM_DOMAIN_GTT |
153 							AMDGPU_GEM_DOMAIN_CPU);
154 			abo->placements[0].fpfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
155 			abo->placements[0].lpfn = 0;
156 			abo->placement.busy_placement = &abo->placements[1];
157 			abo->placement.num_busy_placement = 1;
158 		} else {
159 			/* Move to GTT memory */
160 			amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT |
161 							AMDGPU_GEM_DOMAIN_CPU);
162 		}
163 		break;
164 	case TTM_PL_TT:
165 	case AMDGPU_PL_PREEMPT:
166 	default:
167 		amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
168 		break;
169 	}
170 	*placement = abo->placement;
171 }
172 
173 /**
174  * amdgpu_ttm_map_buffer - Map memory into the GART windows
175  * @bo: buffer object to map
176  * @mem: memory object to map
177  * @mm_cur: range to map
178  * @num_pages: number of pages to map
179  * @window: which GART window to use
180  * @ring: DMA ring to use for the copy
181  * @tmz: if we should setup a TMZ enabled mapping
182  * @addr: resulting address inside the MC address space
183  *
184  * Setup one of the GART windows to access a specific piece of memory or return
185  * the physical address for local memory.
186  */
amdgpu_ttm_map_buffer(struct ttm_buffer_object * bo,struct ttm_resource * mem,struct amdgpu_res_cursor * mm_cur,unsigned num_pages,unsigned window,struct amdgpu_ring * ring,bool tmz,uint64_t * addr)187 static int amdgpu_ttm_map_buffer(struct ttm_buffer_object *bo,
188 				 struct ttm_resource *mem,
189 				 struct amdgpu_res_cursor *mm_cur,
190 				 unsigned num_pages, unsigned window,
191 				 struct amdgpu_ring *ring, bool tmz,
192 				 uint64_t *addr)
193 {
194 	struct amdgpu_device *adev = ring->adev;
195 	struct amdgpu_job *job;
196 	unsigned num_dw, num_bytes;
197 	struct dma_fence *fence;
198 	uint64_t src_addr, dst_addr;
199 	void *cpu_addr;
200 	uint64_t flags;
201 	unsigned int i;
202 	int r;
203 
204 	BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
205 	       AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
206 	BUG_ON(mem->mem_type == AMDGPU_PL_PREEMPT);
207 
208 	/* Map only what can't be accessed directly */
209 	if (!tmz && mem->start != AMDGPU_BO_INVALID_OFFSET) {
210 		*addr = amdgpu_ttm_domain_start(adev, mem->mem_type) +
211 			mm_cur->start;
212 		return 0;
213 	}
214 
215 	*addr = adev->gmc.gart_start;
216 	*addr += (u64)window * AMDGPU_GTT_MAX_TRANSFER_SIZE *
217 		AMDGPU_GPU_PAGE_SIZE;
218 	*addr += mm_cur->start & ~PAGE_MASK;
219 
220 	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
221 	num_bytes = num_pages * 8 * AMDGPU_GPU_PAGES_IN_CPU_PAGE;
222 
223 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes,
224 				     AMDGPU_IB_POOL_DELAYED, &job);
225 	if (r)
226 		return r;
227 
228 	src_addr = num_dw * 4;
229 	src_addr += job->ibs[0].gpu_addr;
230 
231 	dst_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
232 	dst_addr += window * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
233 	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
234 				dst_addr, num_bytes, false);
235 
236 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
237 	WARN_ON(job->ibs[0].length_dw > num_dw);
238 
239 	flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, mem);
240 	if (tmz)
241 		flags |= AMDGPU_PTE_TMZ;
242 
243 	cpu_addr = &job->ibs[0].ptr[num_dw];
244 
245 	if (mem->mem_type == TTM_PL_TT) {
246 		dma_addr_t *dma_addr;
247 
248 		dma_addr = &bo->ttm->dma_address[mm_cur->start >> PAGE_SHIFT];
249 		r = amdgpu_gart_map(adev, 0, num_pages, dma_addr, flags,
250 				    cpu_addr);
251 		if (r)
252 			goto error_free;
253 	} else {
254 		dma_addr_t dma_address;
255 
256 		dma_address = mm_cur->start;
257 		dma_address += adev->vm_manager.vram_base_offset;
258 
259 		for (i = 0; i < num_pages; ++i) {
260 			r = amdgpu_gart_map(adev, i << PAGE_SHIFT, 1,
261 					    &dma_address, flags, cpu_addr);
262 			if (r)
263 				goto error_free;
264 
265 			dma_address += PAGE_SIZE;
266 		}
267 	}
268 
269 	r = amdgpu_job_submit(job, &adev->mman.entity,
270 			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
271 	if (r)
272 		goto error_free;
273 
274 	dma_fence_put(fence);
275 
276 	return r;
277 
278 error_free:
279 	amdgpu_job_free(job);
280 	return r;
281 }
282 
283 /**
284  * amdgpu_ttm_copy_mem_to_mem - Helper function for copy
285  * @adev: amdgpu device
286  * @src: buffer/address where to read from
287  * @dst: buffer/address where to write to
288  * @size: number of bytes to copy
289  * @tmz: if a secure copy should be used
290  * @resv: resv object to sync to
291  * @f: Returns the last fence if multiple jobs are submitted.
292  *
293  * The function copies @size bytes from {src->mem + src->offset} to
294  * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
295  * move and different for a BO to BO copy.
296  *
297  */
amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device * adev,const struct amdgpu_copy_mem * src,const struct amdgpu_copy_mem * dst,uint64_t size,bool tmz,struct dma_resv * resv,struct dma_fence ** f)298 int amdgpu_ttm_copy_mem_to_mem(struct amdgpu_device *adev,
299 			       const struct amdgpu_copy_mem *src,
300 			       const struct amdgpu_copy_mem *dst,
301 			       uint64_t size, bool tmz,
302 			       struct dma_resv *resv,
303 			       struct dma_fence **f)
304 {
305 	const uint32_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
306 					AMDGPU_GPU_PAGE_SIZE);
307 
308 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
309 	struct amdgpu_res_cursor src_mm, dst_mm;
310 	struct dma_fence *fence = NULL;
311 	int r = 0;
312 
313 	if (!adev->mman.buffer_funcs_enabled) {
314 		DRM_ERROR("Trying to move memory with ring turned off.\n");
315 		return -EINVAL;
316 	}
317 
318 	amdgpu_res_first(src->mem, src->offset, size, &src_mm);
319 	amdgpu_res_first(dst->mem, dst->offset, size, &dst_mm);
320 
321 	mutex_lock(&adev->mman.gtt_window_lock);
322 	while (src_mm.remaining) {
323 		uint32_t src_page_offset = src_mm.start & ~PAGE_MASK;
324 		uint32_t dst_page_offset = dst_mm.start & ~PAGE_MASK;
325 		struct dma_fence *next;
326 		uint32_t cur_size;
327 		uint64_t from, to;
328 
329 		/* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
330 		 * begins at an offset, then adjust the size accordingly
331 		 */
332 		cur_size = max(src_page_offset, dst_page_offset);
333 		cur_size = min(min3(src_mm.size, dst_mm.size, size),
334 			       (uint64_t)(GTT_MAX_BYTES - cur_size));
335 
336 		/* Map src to window 0 and dst to window 1. */
337 		r = amdgpu_ttm_map_buffer(src->bo, src->mem, &src_mm,
338 					  PFN_UP(cur_size + src_page_offset),
339 					  0, ring, tmz, &from);
340 		if (r)
341 			goto error;
342 
343 		r = amdgpu_ttm_map_buffer(dst->bo, dst->mem, &dst_mm,
344 					  PFN_UP(cur_size + dst_page_offset),
345 					  1, ring, tmz, &to);
346 		if (r)
347 			goto error;
348 
349 		r = amdgpu_copy_buffer(ring, from, to, cur_size,
350 				       resv, &next, false, true, tmz);
351 		if (r)
352 			goto error;
353 
354 		dma_fence_put(fence);
355 		fence = next;
356 
357 		amdgpu_res_next(&src_mm, cur_size);
358 		amdgpu_res_next(&dst_mm, cur_size);
359 	}
360 error:
361 	mutex_unlock(&adev->mman.gtt_window_lock);
362 	if (f)
363 		*f = dma_fence_get(fence);
364 	dma_fence_put(fence);
365 	return r;
366 }
367 
368 /*
369  * amdgpu_move_blit - Copy an entire buffer to another buffer
370  *
371  * This is a helper called by amdgpu_bo_move() and amdgpu_move_vram_ram() to
372  * help move buffers to and from VRAM.
373  */
amdgpu_move_blit(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem,struct ttm_resource * old_mem)374 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
375 			    bool evict,
376 			    struct ttm_resource *new_mem,
377 			    struct ttm_resource *old_mem)
378 {
379 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
380 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
381 	struct amdgpu_copy_mem src, dst;
382 	struct dma_fence *fence = NULL;
383 	int r;
384 
385 	src.bo = bo;
386 	dst.bo = bo;
387 	src.mem = old_mem;
388 	dst.mem = new_mem;
389 	src.offset = 0;
390 	dst.offset = 0;
391 
392 	r = amdgpu_ttm_copy_mem_to_mem(adev, &src, &dst,
393 				       new_mem->num_pages << PAGE_SHIFT,
394 				       amdgpu_bo_encrypted(abo),
395 				       bo->base.resv, &fence);
396 	if (r)
397 		goto error;
398 
399 	/* clear the space being freed */
400 	if (old_mem->mem_type == TTM_PL_VRAM &&
401 	    (abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE)) {
402 		struct dma_fence *wipe_fence = NULL;
403 
404 		r = amdgpu_fill_buffer(ttm_to_amdgpu_bo(bo), AMDGPU_POISON,
405 				       NULL, &wipe_fence);
406 		if (r) {
407 			goto error;
408 		} else if (wipe_fence) {
409 			dma_fence_put(fence);
410 			fence = wipe_fence;
411 		}
412 	}
413 
414 	/* Always block for VM page tables before committing the new location */
415 	if (bo->type == ttm_bo_type_kernel)
416 		r = ttm_bo_move_accel_cleanup(bo, fence, true, false, new_mem);
417 	else
418 		r = ttm_bo_move_accel_cleanup(bo, fence, evict, true, new_mem);
419 	dma_fence_put(fence);
420 	return r;
421 
422 error:
423 	if (fence)
424 		dma_fence_wait(fence, false);
425 	dma_fence_put(fence);
426 	return r;
427 }
428 
429 /*
430  * amdgpu_mem_visible - Check that memory can be accessed by ttm_bo_move_memcpy
431  *
432  * Called by amdgpu_bo_move()
433  */
amdgpu_mem_visible(struct amdgpu_device * adev,struct ttm_resource * mem)434 static bool amdgpu_mem_visible(struct amdgpu_device *adev,
435 			       struct ttm_resource *mem)
436 {
437 	uint64_t mem_size = (u64)mem->num_pages << PAGE_SHIFT;
438 	struct amdgpu_res_cursor cursor;
439 
440 	if (mem->mem_type == TTM_PL_SYSTEM ||
441 	    mem->mem_type == TTM_PL_TT)
442 		return true;
443 	if (mem->mem_type != TTM_PL_VRAM)
444 		return false;
445 
446 	amdgpu_res_first(mem, 0, mem_size, &cursor);
447 
448 	/* ttm_resource_ioremap only supports contiguous memory */
449 	if (cursor.size != mem_size)
450 		return false;
451 
452 	return cursor.start + cursor.size <= adev->gmc.visible_vram_size;
453 }
454 
455 /*
456  * amdgpu_bo_move - Move a buffer object to a new memory location
457  *
458  * Called by ttm_bo_handle_move_mem()
459  */
amdgpu_bo_move(struct ttm_buffer_object * bo,bool evict,struct ttm_operation_ctx * ctx,struct ttm_resource * new_mem,struct ttm_place * hop)460 static int amdgpu_bo_move(struct ttm_buffer_object *bo, bool evict,
461 			  struct ttm_operation_ctx *ctx,
462 			  struct ttm_resource *new_mem,
463 			  struct ttm_place *hop)
464 {
465 	struct amdgpu_device *adev;
466 	struct amdgpu_bo *abo;
467 	struct ttm_resource *old_mem = bo->resource;
468 	int r;
469 
470 	if (new_mem->mem_type == TTM_PL_TT ||
471 	    new_mem->mem_type == AMDGPU_PL_PREEMPT) {
472 		r = amdgpu_ttm_backend_bind(bo->bdev, bo->ttm, new_mem);
473 		if (r)
474 			return r;
475 	}
476 
477 	/* Can't move a pinned BO */
478 	abo = ttm_to_amdgpu_bo(bo);
479 	if (WARN_ON_ONCE(abo->tbo.pin_count > 0))
480 		return -EINVAL;
481 
482 	adev = amdgpu_ttm_adev(bo->bdev);
483 
484 	if (old_mem->mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
485 		ttm_bo_move_null(bo, new_mem);
486 		goto out;
487 	}
488 	if (old_mem->mem_type == TTM_PL_SYSTEM &&
489 	    (new_mem->mem_type == TTM_PL_TT ||
490 	     new_mem->mem_type == AMDGPU_PL_PREEMPT)) {
491 		ttm_bo_move_null(bo, new_mem);
492 		goto out;
493 	}
494 	if ((old_mem->mem_type == TTM_PL_TT ||
495 	     old_mem->mem_type == AMDGPU_PL_PREEMPT) &&
496 	    new_mem->mem_type == TTM_PL_SYSTEM) {
497 		r = ttm_bo_wait_ctx(bo, ctx);
498 		if (r)
499 			return r;
500 
501 		amdgpu_ttm_backend_unbind(bo->bdev, bo->ttm);
502 		ttm_resource_free(bo, &bo->resource);
503 		ttm_bo_assign_mem(bo, new_mem);
504 		goto out;
505 	}
506 
507 	if (old_mem->mem_type == AMDGPU_PL_GDS ||
508 	    old_mem->mem_type == AMDGPU_PL_GWS ||
509 	    old_mem->mem_type == AMDGPU_PL_OA ||
510 	    new_mem->mem_type == AMDGPU_PL_GDS ||
511 	    new_mem->mem_type == AMDGPU_PL_GWS ||
512 	    new_mem->mem_type == AMDGPU_PL_OA) {
513 		/* Nothing to save here */
514 		ttm_bo_move_null(bo, new_mem);
515 		goto out;
516 	}
517 
518 	if (bo->type == ttm_bo_type_device &&
519 	    new_mem->mem_type == TTM_PL_VRAM &&
520 	    old_mem->mem_type != TTM_PL_VRAM) {
521 		/* amdgpu_bo_fault_reserve_notify will re-set this if the CPU
522 		 * accesses the BO after it's moved.
523 		 */
524 		abo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
525 	}
526 
527 	if (adev->mman.buffer_funcs_enabled) {
528 		if (((old_mem->mem_type == TTM_PL_SYSTEM &&
529 		      new_mem->mem_type == TTM_PL_VRAM) ||
530 		     (old_mem->mem_type == TTM_PL_VRAM &&
531 		      new_mem->mem_type == TTM_PL_SYSTEM))) {
532 			hop->fpfn = 0;
533 			hop->lpfn = 0;
534 			hop->mem_type = TTM_PL_TT;
535 			hop->flags = TTM_PL_FLAG_TEMPORARY;
536 			return -EMULTIHOP;
537 		}
538 
539 		r = amdgpu_move_blit(bo, evict, new_mem, old_mem);
540 	} else {
541 		r = -ENODEV;
542 	}
543 
544 	if (r) {
545 		/* Check that all memory is CPU accessible */
546 		if (!amdgpu_mem_visible(adev, old_mem) ||
547 		    !amdgpu_mem_visible(adev, new_mem)) {
548 			pr_err("Move buffer fallback to memcpy unavailable\n");
549 			return r;
550 		}
551 
552 		r = ttm_bo_move_memcpy(bo, ctx, new_mem);
553 		if (r)
554 			return r;
555 	}
556 
557 	trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
558 out:
559 	/* update statistics */
560 	atomic64_add(bo->base.size, &adev->num_bytes_moved);
561 	amdgpu_bo_move_notify(bo, evict);
562 	return 0;
563 }
564 
565 /*
566  * amdgpu_ttm_io_mem_reserve - Reserve a block of memory during a fault
567  *
568  * Called by ttm_mem_io_reserve() ultimately via ttm_bo_vm_fault()
569  */
amdgpu_ttm_io_mem_reserve(struct ttm_device * bdev,struct ttm_resource * mem)570 static int amdgpu_ttm_io_mem_reserve(struct ttm_device *bdev,
571 				     struct ttm_resource *mem)
572 {
573 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
574 	size_t bus_size = (size_t)mem->num_pages << PAGE_SHIFT;
575 
576 	switch (mem->mem_type) {
577 	case TTM_PL_SYSTEM:
578 		/* system memory */
579 		return 0;
580 	case TTM_PL_TT:
581 	case AMDGPU_PL_PREEMPT:
582 		break;
583 	case TTM_PL_VRAM:
584 		mem->bus.offset = mem->start << PAGE_SHIFT;
585 		/* check if it's visible */
586 		if ((mem->bus.offset + bus_size) > adev->gmc.visible_vram_size)
587 			return -EINVAL;
588 
589 		if (adev->mman.aper_base_kaddr &&
590 		    mem->placement & TTM_PL_FLAG_CONTIGUOUS)
591 			mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
592 					mem->bus.offset;
593 
594 		mem->bus.offset += adev->gmc.aper_base;
595 		mem->bus.is_iomem = true;
596 		break;
597 	default:
598 		return -EINVAL;
599 	}
600 	return 0;
601 }
602 
amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object * bo,unsigned long page_offset)603 static unsigned long amdgpu_ttm_io_mem_pfn(struct ttm_buffer_object *bo,
604 					   unsigned long page_offset)
605 {
606 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
607 	struct amdgpu_res_cursor cursor;
608 
609 	amdgpu_res_first(bo->resource, (u64)page_offset << PAGE_SHIFT, 0,
610 			 &cursor);
611 	return (adev->gmc.aper_base + cursor.start) >> PAGE_SHIFT;
612 }
613 
614 /**
615  * amdgpu_ttm_domain_start - Returns GPU start address
616  * @adev: amdgpu device object
617  * @type: type of the memory
618  *
619  * Returns:
620  * GPU start address of a memory domain
621  */
622 
amdgpu_ttm_domain_start(struct amdgpu_device * adev,uint32_t type)623 uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type)
624 {
625 	switch (type) {
626 	case TTM_PL_TT:
627 		return adev->gmc.gart_start;
628 	case TTM_PL_VRAM:
629 		return adev->gmc.vram_start;
630 	}
631 
632 	return 0;
633 }
634 
635 /*
636  * TTM backend functions.
637  */
638 struct amdgpu_ttm_tt {
639 	struct ttm_tt	ttm;
640 	struct drm_gem_object	*gobj;
641 	u64			offset;
642 	uint64_t		userptr;
643 	struct task_struct	*usertask;
644 	uint32_t		userflags;
645 	bool			bound;
646 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
647 	struct hmm_range	*range;
648 #endif
649 };
650 
651 #ifdef CONFIG_DRM_AMDGPU_USERPTR
652 /*
653  * amdgpu_ttm_tt_get_user_pages - get device accessible pages that back user
654  * memory and start HMM tracking CPU page table update
655  *
656  * Calling function must call amdgpu_ttm_tt_userptr_range_done() once and only
657  * once afterwards to stop HMM tracking
658  */
amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo * bo,struct page ** pages)659 int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
660 {
661 	struct ttm_tt *ttm = bo->tbo.ttm;
662 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
663 	unsigned long start = gtt->userptr;
664 	struct vm_area_struct *vma;
665 	struct mm_struct *mm;
666 	bool readonly;
667 	int r = 0;
668 
669 	mm = bo->notifier.mm;
670 	if (unlikely(!mm)) {
671 		DRM_DEBUG_DRIVER("BO is not registered?\n");
672 		return -EFAULT;
673 	}
674 
675 	/* Another get_user_pages is running at the same time?? */
676 	if (WARN_ON(gtt->range))
677 		return -EFAULT;
678 
679 	if (!mmget_not_zero(mm)) /* Happens during process shutdown */
680 		return -ESRCH;
681 
682 	mmap_read_lock(mm);
683 	vma = vma_lookup(mm, start);
684 	if (unlikely(!vma)) {
685 		r = -EFAULT;
686 		goto out_unlock;
687 	}
688 	if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
689 		vma->vm_file)) {
690 		r = -EPERM;
691 		goto out_unlock;
692 	}
693 
694 	readonly = amdgpu_ttm_tt_is_readonly(ttm);
695 	r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start,
696 				       ttm->num_pages, &gtt->range, readonly,
697 				       true, NULL);
698 out_unlock:
699 	mmap_read_unlock(mm);
700 	mmput(mm);
701 
702 	return r;
703 }
704 
705 /*
706  * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
707  * Check if the pages backing this ttm range have been invalidated
708  *
709  * Returns: true if pages are still valid
710  */
amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt * ttm)711 bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
712 {
713 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
714 	bool r = false;
715 
716 	if (!gtt || !gtt->userptr)
717 		return false;
718 
719 	DRM_DEBUG_DRIVER("user_pages_done 0x%llx pages 0x%x\n",
720 		gtt->userptr, ttm->num_pages);
721 
722 	WARN_ONCE(!gtt->range || !gtt->range->hmm_pfns,
723 		"No user pages to check\n");
724 
725 	if (gtt->range) {
726 		/*
727 		 * FIXME: Must always hold notifier_lock for this, and must
728 		 * not ignore the return code.
729 		 */
730 		r = amdgpu_hmm_range_get_pages_done(gtt->range);
731 		gtt->range = NULL;
732 	}
733 
734 	return !r;
735 }
736 #endif
737 
738 /*
739  * amdgpu_ttm_tt_set_user_pages - Copy pages in, putting old pages as necessary.
740  *
741  * Called by amdgpu_cs_list_validate(). This creates the page list
742  * that backs user memory and will ultimately be mapped into the device
743  * address space.
744  */
amdgpu_ttm_tt_set_user_pages(struct ttm_tt * ttm,struct page ** pages)745 void amdgpu_ttm_tt_set_user_pages(struct ttm_tt *ttm, struct page **pages)
746 {
747 	unsigned long i;
748 
749 	for (i = 0; i < ttm->num_pages; ++i)
750 		ttm->pages[i] = pages ? pages[i] : NULL;
751 }
752 
753 /*
754  * amdgpu_ttm_tt_pin_userptr - prepare the sg table with the user pages
755  *
756  * Called by amdgpu_ttm_backend_bind()
757  **/
amdgpu_ttm_tt_pin_userptr(struct ttm_device * bdev,struct ttm_tt * ttm)758 static int amdgpu_ttm_tt_pin_userptr(struct ttm_device *bdev,
759 				     struct ttm_tt *ttm)
760 {
761 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
762 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
763 	int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
764 	enum dma_data_direction direction = write ?
765 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
766 	int r;
767 
768 	/* Allocate an SG array and squash pages into it */
769 	r = sg_alloc_table_from_pages(ttm->sg, ttm->pages, ttm->num_pages, 0,
770 				      (u64)ttm->num_pages << PAGE_SHIFT,
771 				      GFP_KERNEL);
772 	if (r)
773 		goto release_sg;
774 
775 	/* Map SG to device */
776 	r = dma_map_sgtable(adev->dev, ttm->sg, direction, 0);
777 	if (r)
778 		goto release_sg;
779 
780 	/* convert SG to linear array of pages and dma addresses */
781 	drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
782 				       ttm->num_pages);
783 
784 	return 0;
785 
786 release_sg:
787 	kfree(ttm->sg);
788 	ttm->sg = NULL;
789 	return r;
790 }
791 
792 /*
793  * amdgpu_ttm_tt_unpin_userptr - Unpin and unmap userptr pages
794  */
amdgpu_ttm_tt_unpin_userptr(struct ttm_device * bdev,struct ttm_tt * ttm)795 static void amdgpu_ttm_tt_unpin_userptr(struct ttm_device *bdev,
796 					struct ttm_tt *ttm)
797 {
798 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
799 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
800 	int write = !(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
801 	enum dma_data_direction direction = write ?
802 		DMA_BIDIRECTIONAL : DMA_TO_DEVICE;
803 
804 	/* double check that we don't free the table twice */
805 	if (!ttm->sg || !ttm->sg->sgl)
806 		return;
807 
808 	/* unmap the pages mapped to the device */
809 	dma_unmap_sgtable(adev->dev, ttm->sg, direction, 0);
810 	sg_free_table(ttm->sg);
811 
812 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
813 	if (gtt->range) {
814 		unsigned long i;
815 
816 		for (i = 0; i < ttm->num_pages; i++) {
817 			if (ttm->pages[i] !=
818 			    hmm_pfn_to_page(gtt->range->hmm_pfns[i]))
819 				break;
820 		}
821 
822 		WARN((i == ttm->num_pages), "Missing get_user_page_done\n");
823 	}
824 #endif
825 }
826 
amdgpu_ttm_gart_bind(struct amdgpu_device * adev,struct ttm_buffer_object * tbo,uint64_t flags)827 static int amdgpu_ttm_gart_bind(struct amdgpu_device *adev,
828 				struct ttm_buffer_object *tbo,
829 				uint64_t flags)
830 {
831 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(tbo);
832 	struct ttm_tt *ttm = tbo->ttm;
833 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
834 	int r;
835 
836 	if (amdgpu_bo_encrypted(abo))
837 		flags |= AMDGPU_PTE_TMZ;
838 
839 	if (abo->flags & AMDGPU_GEM_CREATE_CP_MQD_GFX9) {
840 		uint64_t page_idx = 1;
841 
842 		r = amdgpu_gart_bind(adev, gtt->offset, page_idx,
843 				gtt->ttm.dma_address, flags);
844 		if (r)
845 			goto gart_bind_fail;
846 
847 		/* The memory type of the first page defaults to UC. Now
848 		 * modify the memory type to NC from the second page of
849 		 * the BO onward.
850 		 */
851 		flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
852 		flags |= AMDGPU_PTE_MTYPE_VG10(AMDGPU_MTYPE_NC);
853 
854 		r = amdgpu_gart_bind(adev,
855 				gtt->offset + (page_idx << PAGE_SHIFT),
856 				ttm->num_pages - page_idx,
857 				&(gtt->ttm.dma_address[page_idx]), flags);
858 	} else {
859 		r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
860 				     gtt->ttm.dma_address, flags);
861 	}
862 
863 gart_bind_fail:
864 	if (r)
865 		DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
866 			  ttm->num_pages, gtt->offset);
867 
868 	return r;
869 }
870 
871 /*
872  * amdgpu_ttm_backend_bind - Bind GTT memory
873  *
874  * Called by ttm_tt_bind() on behalf of ttm_bo_handle_move_mem().
875  * This handles binding GTT memory to the device address space.
876  */
amdgpu_ttm_backend_bind(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_resource * bo_mem)877 static int amdgpu_ttm_backend_bind(struct ttm_device *bdev,
878 				   struct ttm_tt *ttm,
879 				   struct ttm_resource *bo_mem)
880 {
881 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
882 	struct amdgpu_ttm_tt *gtt = (void*)ttm;
883 	uint64_t flags;
884 	int r = 0;
885 
886 	if (!bo_mem)
887 		return -EINVAL;
888 
889 	if (gtt->bound)
890 		return 0;
891 
892 	if (gtt->userptr) {
893 		r = amdgpu_ttm_tt_pin_userptr(bdev, ttm);
894 		if (r) {
895 			DRM_ERROR("failed to pin userptr\n");
896 			return r;
897 		}
898 	} else if (ttm->page_flags & TTM_PAGE_FLAG_SG) {
899 		if (!ttm->sg) {
900 			struct dma_buf_attachment *attach;
901 			struct sg_table *sgt;
902 
903 			attach = gtt->gobj->import_attach;
904 			sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
905 			if (IS_ERR(sgt))
906 				return PTR_ERR(sgt);
907 
908 			ttm->sg = sgt;
909 		}
910 
911 		drm_prime_sg_to_dma_addr_array(ttm->sg, gtt->ttm.dma_address,
912 					       ttm->num_pages);
913 	}
914 
915 	if (!ttm->num_pages) {
916 		WARN(1, "nothing to bind %u pages for mreg %p back %p!\n",
917 		     ttm->num_pages, bo_mem, ttm);
918 	}
919 
920 	if (bo_mem->mem_type == AMDGPU_PL_GDS ||
921 	    bo_mem->mem_type == AMDGPU_PL_GWS ||
922 	    bo_mem->mem_type == AMDGPU_PL_OA)
923 		return -EINVAL;
924 
925 	if (bo_mem->mem_type != TTM_PL_TT ||
926 	    !amdgpu_gtt_mgr_has_gart_addr(bo_mem)) {
927 		gtt->offset = AMDGPU_BO_INVALID_OFFSET;
928 		return 0;
929 	}
930 
931 	/* compute PTE flags relevant to this BO memory */
932 	flags = amdgpu_ttm_tt_pte_flags(adev, ttm, bo_mem);
933 
934 	/* bind pages into GART page tables */
935 	gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
936 	r = amdgpu_gart_bind(adev, gtt->offset, ttm->num_pages,
937 		gtt->ttm.dma_address, flags);
938 
939 	if (r)
940 		DRM_ERROR("failed to bind %u pages at 0x%08llX\n",
941 			  ttm->num_pages, gtt->offset);
942 	gtt->bound = true;
943 	return r;
944 }
945 
946 /*
947  * amdgpu_ttm_alloc_gart - Make sure buffer object is accessible either
948  * through AGP or GART aperture.
949  *
950  * If bo is accessible through AGP aperture, then use AGP aperture
951  * to access bo; otherwise allocate logical space in GART aperture
952  * and map bo to GART aperture.
953  */
amdgpu_ttm_alloc_gart(struct ttm_buffer_object * bo)954 int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
955 {
956 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
957 	struct ttm_operation_ctx ctx = { false, false };
958 	struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
959 	struct ttm_placement placement;
960 	struct ttm_place placements;
961 	struct ttm_resource *tmp;
962 	uint64_t addr, flags;
963 	int r;
964 
965 	if (bo->resource->start != AMDGPU_BO_INVALID_OFFSET)
966 		return 0;
967 
968 	addr = amdgpu_gmc_agp_addr(bo);
969 	if (addr != AMDGPU_BO_INVALID_OFFSET) {
970 		bo->resource->start = addr >> PAGE_SHIFT;
971 		return 0;
972 	}
973 
974 	/* allocate GART space */
975 	placement.num_placement = 1;
976 	placement.placement = &placements;
977 	placement.num_busy_placement = 1;
978 	placement.busy_placement = &placements;
979 	placements.fpfn = 0;
980 	placements.lpfn = adev->gmc.gart_size >> PAGE_SHIFT;
981 	placements.mem_type = TTM_PL_TT;
982 	placements.flags = bo->resource->placement;
983 
984 	r = ttm_bo_mem_space(bo, &placement, &tmp, &ctx);
985 	if (unlikely(r))
986 		return r;
987 
988 	/* compute PTE flags for this buffer object */
989 	flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, tmp);
990 
991 	/* Bind pages */
992 	gtt->offset = (u64)tmp->start << PAGE_SHIFT;
993 	r = amdgpu_ttm_gart_bind(adev, bo, flags);
994 	if (unlikely(r)) {
995 		ttm_resource_free(bo, &tmp);
996 		return r;
997 	}
998 
999 	amdgpu_gart_invalidate_tlb(adev);
1000 	ttm_resource_free(bo, &bo->resource);
1001 	ttm_bo_assign_mem(bo, tmp);
1002 
1003 	return 0;
1004 }
1005 
1006 /*
1007  * amdgpu_ttm_recover_gart - Rebind GTT pages
1008  *
1009  * Called by amdgpu_gtt_mgr_recover() from amdgpu_device_reset() to
1010  * rebind GTT pages during a GPU reset.
1011  */
amdgpu_ttm_recover_gart(struct ttm_buffer_object * tbo)1012 int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
1013 {
1014 	struct amdgpu_device *adev = amdgpu_ttm_adev(tbo->bdev);
1015 	uint64_t flags;
1016 	int r;
1017 
1018 	if (!tbo->ttm)
1019 		return 0;
1020 
1021 	flags = amdgpu_ttm_tt_pte_flags(adev, tbo->ttm, tbo->resource);
1022 	r = amdgpu_ttm_gart_bind(adev, tbo, flags);
1023 
1024 	return r;
1025 }
1026 
1027 /*
1028  * amdgpu_ttm_backend_unbind - Unbind GTT mapped pages
1029  *
1030  * Called by ttm_tt_unbind() on behalf of ttm_bo_move_ttm() and
1031  * ttm_tt_destroy().
1032  */
amdgpu_ttm_backend_unbind(struct ttm_device * bdev,struct ttm_tt * ttm)1033 static void amdgpu_ttm_backend_unbind(struct ttm_device *bdev,
1034 				      struct ttm_tt *ttm)
1035 {
1036 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1037 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1038 	int r;
1039 
1040 	/* if the pages have userptr pinning then clear that first */
1041 	if (gtt->userptr) {
1042 		amdgpu_ttm_tt_unpin_userptr(bdev, ttm);
1043 	} else if (ttm->sg && gtt->gobj->import_attach) {
1044 		struct dma_buf_attachment *attach;
1045 
1046 		attach = gtt->gobj->import_attach;
1047 		dma_buf_unmap_attachment(attach, ttm->sg, DMA_BIDIRECTIONAL);
1048 		ttm->sg = NULL;
1049 	}
1050 
1051 	if (!gtt->bound)
1052 		return;
1053 
1054 	if (gtt->offset == AMDGPU_BO_INVALID_OFFSET)
1055 		return;
1056 
1057 	/* unbind shouldn't be done for GDS/GWS/OA in ttm_bo_clean_mm */
1058 	r = amdgpu_gart_unbind(adev, gtt->offset, ttm->num_pages);
1059 	if (r)
1060 		DRM_ERROR("failed to unbind %u pages at 0x%08llX\n",
1061 			  gtt->ttm.num_pages, gtt->offset);
1062 	gtt->bound = false;
1063 }
1064 
amdgpu_ttm_backend_destroy(struct ttm_device * bdev,struct ttm_tt * ttm)1065 static void amdgpu_ttm_backend_destroy(struct ttm_device *bdev,
1066 				       struct ttm_tt *ttm)
1067 {
1068 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1069 
1070 	amdgpu_ttm_backend_unbind(bdev, ttm);
1071 	ttm_tt_destroy_common(bdev, ttm);
1072 	if (gtt->usertask)
1073 		put_task_struct(gtt->usertask);
1074 
1075 	ttm_tt_fini(&gtt->ttm);
1076 	kfree(gtt);
1077 }
1078 
1079 /**
1080  * amdgpu_ttm_tt_create - Create a ttm_tt object for a given BO
1081  *
1082  * @bo: The buffer object to create a GTT ttm_tt object around
1083  * @page_flags: Page flags to be added to the ttm_tt object
1084  *
1085  * Called by ttm_tt_create().
1086  */
amdgpu_ttm_tt_create(struct ttm_buffer_object * bo,uint32_t page_flags)1087 static struct ttm_tt *amdgpu_ttm_tt_create(struct ttm_buffer_object *bo,
1088 					   uint32_t page_flags)
1089 {
1090 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1091 	struct amdgpu_ttm_tt *gtt;
1092 	enum ttm_caching caching;
1093 
1094 	gtt = kzalloc(sizeof(struct amdgpu_ttm_tt), GFP_KERNEL);
1095 	if (gtt == NULL) {
1096 		return NULL;
1097 	}
1098 	gtt->gobj = &bo->base;
1099 
1100 	if (abo->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
1101 		caching = ttm_write_combined;
1102 	else
1103 		caching = ttm_cached;
1104 
1105 	/* allocate space for the uninitialized page entries */
1106 	if (ttm_sg_tt_init(&gtt->ttm, bo, page_flags, caching)) {
1107 		kfree(gtt);
1108 		return NULL;
1109 	}
1110 	return &gtt->ttm;
1111 }
1112 
1113 /*
1114  * amdgpu_ttm_tt_populate - Map GTT pages visible to the device
1115  *
1116  * Map the pages of a ttm_tt object to an address space visible
1117  * to the underlying device.
1118  */
amdgpu_ttm_tt_populate(struct ttm_device * bdev,struct ttm_tt * ttm,struct ttm_operation_ctx * ctx)1119 static int amdgpu_ttm_tt_populate(struct ttm_device *bdev,
1120 				  struct ttm_tt *ttm,
1121 				  struct ttm_operation_ctx *ctx)
1122 {
1123 	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
1124 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1125 
1126 	/* user pages are bound by amdgpu_ttm_tt_pin_userptr() */
1127 	if (gtt->userptr) {
1128 		ttm->sg = kzalloc(sizeof(struct sg_table), GFP_KERNEL);
1129 		if (!ttm->sg)
1130 			return -ENOMEM;
1131 		return 0;
1132 	}
1133 
1134 	if (ttm->page_flags & TTM_PAGE_FLAG_SG)
1135 		return 0;
1136 
1137 	return ttm_pool_alloc(&adev->mman.bdev.pool, ttm, ctx);
1138 }
1139 
1140 /*
1141  * amdgpu_ttm_tt_unpopulate - unmap GTT pages and unpopulate page arrays
1142  *
1143  * Unmaps pages of a ttm_tt object from the device address space and
1144  * unpopulates the page array backing it.
1145  */
amdgpu_ttm_tt_unpopulate(struct ttm_device * bdev,struct ttm_tt * ttm)1146 static void amdgpu_ttm_tt_unpopulate(struct ttm_device *bdev,
1147 				     struct ttm_tt *ttm)
1148 {
1149 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1150 	struct amdgpu_device *adev;
1151 
1152 	if (gtt->userptr) {
1153 		amdgpu_ttm_tt_set_user_pages(ttm, NULL);
1154 		kfree(ttm->sg);
1155 		ttm->sg = NULL;
1156 		return;
1157 	}
1158 
1159 	if (ttm->page_flags & TTM_PAGE_FLAG_SG)
1160 		return;
1161 
1162 	adev = amdgpu_ttm_adev(bdev);
1163 	return ttm_pool_free(&adev->mman.bdev.pool, ttm);
1164 }
1165 
1166 /**
1167  * amdgpu_ttm_tt_set_userptr - Initialize userptr GTT ttm_tt for the current
1168  * task
1169  *
1170  * @bo: The ttm_buffer_object to bind this userptr to
1171  * @addr:  The address in the current tasks VM space to use
1172  * @flags: Requirements of userptr object.
1173  *
1174  * Called by amdgpu_gem_userptr_ioctl() to bind userptr pages
1175  * to current task
1176  */
amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object * bo,uint64_t addr,uint32_t flags)1177 int amdgpu_ttm_tt_set_userptr(struct ttm_buffer_object *bo,
1178 			      uint64_t addr, uint32_t flags)
1179 {
1180 	struct amdgpu_ttm_tt *gtt;
1181 
1182 	if (!bo->ttm) {
1183 		/* TODO: We want a separate TTM object type for userptrs */
1184 		bo->ttm = amdgpu_ttm_tt_create(bo, 0);
1185 		if (bo->ttm == NULL)
1186 			return -ENOMEM;
1187 	}
1188 
1189 	/* Set TTM_PAGE_FLAG_SG before populate but after create. */
1190 	bo->ttm->page_flags |= TTM_PAGE_FLAG_SG;
1191 
1192 	gtt = (void *)bo->ttm;
1193 	gtt->userptr = addr;
1194 	gtt->userflags = flags;
1195 
1196 	if (gtt->usertask)
1197 		put_task_struct(gtt->usertask);
1198 	gtt->usertask = current->group_leader;
1199 	get_task_struct(gtt->usertask);
1200 
1201 	return 0;
1202 }
1203 
1204 /*
1205  * amdgpu_ttm_tt_get_usermm - Return memory manager for ttm_tt object
1206  */
amdgpu_ttm_tt_get_usermm(struct ttm_tt * ttm)1207 struct mm_struct *amdgpu_ttm_tt_get_usermm(struct ttm_tt *ttm)
1208 {
1209 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1210 
1211 	if (gtt == NULL)
1212 		return NULL;
1213 
1214 	if (gtt->usertask == NULL)
1215 		return NULL;
1216 
1217 	return gtt->usertask->mm;
1218 }
1219 
1220 /*
1221  * amdgpu_ttm_tt_affect_userptr - Determine if a ttm_tt object lays inside an
1222  * address range for the current task.
1223  *
1224  */
amdgpu_ttm_tt_affect_userptr(struct ttm_tt * ttm,unsigned long start,unsigned long end)1225 bool amdgpu_ttm_tt_affect_userptr(struct ttm_tt *ttm, unsigned long start,
1226 				  unsigned long end)
1227 {
1228 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1229 	unsigned long size;
1230 
1231 	if (gtt == NULL || !gtt->userptr)
1232 		return false;
1233 
1234 	/* Return false if no part of the ttm_tt object lies within
1235 	 * the range
1236 	 */
1237 	size = (unsigned long)gtt->ttm.num_pages * PAGE_SIZE;
1238 	if (gtt->userptr > end || gtt->userptr + size <= start)
1239 		return false;
1240 
1241 	return true;
1242 }
1243 
1244 /*
1245  * amdgpu_ttm_tt_is_userptr - Have the pages backing by userptr?
1246  */
amdgpu_ttm_tt_is_userptr(struct ttm_tt * ttm)1247 bool amdgpu_ttm_tt_is_userptr(struct ttm_tt *ttm)
1248 {
1249 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1250 
1251 	if (gtt == NULL || !gtt->userptr)
1252 		return false;
1253 
1254 	return true;
1255 }
1256 
1257 /*
1258  * amdgpu_ttm_tt_is_readonly - Is the ttm_tt object read only?
1259  */
amdgpu_ttm_tt_is_readonly(struct ttm_tt * ttm)1260 bool amdgpu_ttm_tt_is_readonly(struct ttm_tt *ttm)
1261 {
1262 	struct amdgpu_ttm_tt *gtt = (void *)ttm;
1263 
1264 	if (gtt == NULL)
1265 		return false;
1266 
1267 	return !!(gtt->userflags & AMDGPU_GEM_USERPTR_READONLY);
1268 }
1269 
1270 /**
1271  * amdgpu_ttm_tt_pde_flags - Compute PDE flags for ttm_tt object
1272  *
1273  * @ttm: The ttm_tt object to compute the flags for
1274  * @mem: The memory registry backing this ttm_tt object
1275  *
1276  * Figure out the flags to use for a VM PDE (Page Directory Entry).
1277  */
amdgpu_ttm_tt_pde_flags(struct ttm_tt * ttm,struct ttm_resource * mem)1278 uint64_t amdgpu_ttm_tt_pde_flags(struct ttm_tt *ttm, struct ttm_resource *mem)
1279 {
1280 	uint64_t flags = 0;
1281 
1282 	if (mem && mem->mem_type != TTM_PL_SYSTEM)
1283 		flags |= AMDGPU_PTE_VALID;
1284 
1285 	if (mem && (mem->mem_type == TTM_PL_TT ||
1286 		    mem->mem_type == AMDGPU_PL_PREEMPT)) {
1287 		flags |= AMDGPU_PTE_SYSTEM;
1288 
1289 		if (ttm->caching == ttm_cached)
1290 			flags |= AMDGPU_PTE_SNOOPED;
1291 	}
1292 
1293 	if (mem && mem->mem_type == TTM_PL_VRAM &&
1294 			mem->bus.caching == ttm_cached)
1295 		flags |= AMDGPU_PTE_SNOOPED;
1296 
1297 	return flags;
1298 }
1299 
1300 /**
1301  * amdgpu_ttm_tt_pte_flags - Compute PTE flags for ttm_tt object
1302  *
1303  * @adev: amdgpu_device pointer
1304  * @ttm: The ttm_tt object to compute the flags for
1305  * @mem: The memory registry backing this ttm_tt object
1306  *
1307  * Figure out the flags to use for a VM PTE (Page Table Entry).
1308  */
amdgpu_ttm_tt_pte_flags(struct amdgpu_device * adev,struct ttm_tt * ttm,struct ttm_resource * mem)1309 uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
1310 				 struct ttm_resource *mem)
1311 {
1312 	uint64_t flags = amdgpu_ttm_tt_pde_flags(ttm, mem);
1313 
1314 	flags |= adev->gart.gart_pte_flags;
1315 	flags |= AMDGPU_PTE_READABLE;
1316 
1317 	if (!amdgpu_ttm_tt_is_readonly(ttm))
1318 		flags |= AMDGPU_PTE_WRITEABLE;
1319 
1320 	return flags;
1321 }
1322 
1323 /*
1324  * amdgpu_ttm_bo_eviction_valuable - Check to see if we can evict a buffer
1325  * object.
1326  *
1327  * Return true if eviction is sensible. Called by ttm_mem_evict_first() on
1328  * behalf of ttm_bo_mem_force_space() which tries to evict buffer objects until
1329  * it can find space for a new object and by ttm_bo_force_list_clean() which is
1330  * used to clean out a memory space.
1331  */
amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object * bo,const struct ttm_place * place)1332 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
1333 					    const struct ttm_place *place)
1334 {
1335 	unsigned long num_pages = bo->resource->num_pages;
1336 	struct amdgpu_res_cursor cursor;
1337 	struct dma_resv_list *flist;
1338 	struct dma_fence *f;
1339 	int i;
1340 
1341 	/* Swapout? */
1342 	if (bo->resource->mem_type == TTM_PL_SYSTEM)
1343 		return true;
1344 
1345 	if (bo->type == ttm_bo_type_kernel &&
1346 	    !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
1347 		return false;
1348 
1349 	/* If bo is a KFD BO, check if the bo belongs to the current process.
1350 	 * If true, then return false as any KFD process needs all its BOs to
1351 	 * be resident to run successfully
1352 	 */
1353 	flist = dma_resv_shared_list(bo->base.resv);
1354 	if (flist) {
1355 		for (i = 0; i < flist->shared_count; ++i) {
1356 			f = rcu_dereference_protected(flist->shared[i],
1357 				dma_resv_held(bo->base.resv));
1358 			if (amdkfd_fence_check_mm(f, current->mm))
1359 				return false;
1360 		}
1361 	}
1362 
1363 	switch (bo->resource->mem_type) {
1364 	case AMDGPU_PL_PREEMPT:
1365 		/* Preemptible BOs don't own system resources managed by the
1366 		 * driver (pages, VRAM, GART space). They point to resources
1367 		 * owned by someone else (e.g. pageable memory in user mode
1368 		 * or a DMABuf). They are used in a preemptible context so we
1369 		 * can guarantee no deadlocks and good QoS in case of MMU
1370 		 * notifiers or DMABuf move notifiers from the resource owner.
1371 		 */
1372 		return false;
1373 	case TTM_PL_TT:
1374 		if (amdgpu_bo_is_amdgpu_bo(bo) &&
1375 		    amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
1376 			return false;
1377 		return true;
1378 
1379 	case TTM_PL_VRAM:
1380 		/* Check each drm MM node individually */
1381 		amdgpu_res_first(bo->resource, 0, (u64)num_pages << PAGE_SHIFT,
1382 				 &cursor);
1383 		while (cursor.remaining) {
1384 			if (place->fpfn < PFN_DOWN(cursor.start + cursor.size)
1385 			    && !(place->lpfn &&
1386 				 place->lpfn <= PFN_DOWN(cursor.start)))
1387 				return true;
1388 
1389 			amdgpu_res_next(&cursor, cursor.size);
1390 		}
1391 		return false;
1392 
1393 	default:
1394 		break;
1395 	}
1396 
1397 	return ttm_bo_eviction_valuable(bo, place);
1398 }
1399 
amdgpu_ttm_vram_mm_access(struct amdgpu_device * adev,loff_t pos,void * buf,size_t size,bool write)1400 static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
1401 				      void *buf, size_t size, bool write)
1402 {
1403 	while (size) {
1404 		uint64_t aligned_pos = ALIGN_DOWN(pos, 4);
1405 		uint64_t bytes = 4 - (pos & 0x3);
1406 		uint32_t shift = (pos & 0x3) * 8;
1407 		uint32_t mask = 0xffffffff << shift;
1408 		uint32_t value = 0;
1409 
1410 		if (size < bytes) {
1411 			mask &= 0xffffffff >> (bytes - size) * 8;
1412 			bytes = size;
1413 		}
1414 
1415 		if (mask != 0xffffffff) {
1416 			amdgpu_device_mm_access(adev, aligned_pos, &value, 4, false);
1417 			if (write) {
1418 				value &= ~mask;
1419 				value |= (*(uint32_t *)buf << shift) & mask;
1420 				amdgpu_device_mm_access(adev, aligned_pos, &value, 4, true);
1421 			} else {
1422 				value = (value & mask) >> shift;
1423 				memcpy(buf, &value, bytes);
1424 			}
1425 		} else {
1426 			amdgpu_device_mm_access(adev, aligned_pos, buf, 4, write);
1427 		}
1428 
1429 		pos += bytes;
1430 		buf += bytes;
1431 		size -= bytes;
1432 	}
1433 }
1434 
1435 /**
1436  * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
1437  *
1438  * @bo:  The buffer object to read/write
1439  * @offset:  Offset into buffer object
1440  * @buf:  Secondary buffer to write/read from
1441  * @len: Length in bytes of access
1442  * @write:  true if writing
1443  *
1444  * This is used to access VRAM that backs a buffer object via MMIO
1445  * access for debugging purposes.
1446  */
amdgpu_ttm_access_memory(struct ttm_buffer_object * bo,unsigned long offset,void * buf,int len,int write)1447 static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
1448 				    unsigned long offset, void *buf, int len,
1449 				    int write)
1450 {
1451 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1452 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
1453 	struct amdgpu_res_cursor cursor;
1454 	int ret = 0;
1455 
1456 	if (bo->resource->mem_type != TTM_PL_VRAM)
1457 		return -EIO;
1458 
1459 	amdgpu_res_first(bo->resource, offset, len, &cursor);
1460 	while (cursor.remaining) {
1461 		size_t count, size = cursor.size;
1462 		loff_t pos = cursor.start;
1463 
1464 		count = amdgpu_device_aper_access(adev, pos, buf, size, write);
1465 		size -= count;
1466 		if (size) {
1467 			/* using MM to access rest vram and handle un-aligned address */
1468 			pos += count;
1469 			buf += count;
1470 			amdgpu_ttm_vram_mm_access(adev, pos, buf, size, write);
1471 		}
1472 
1473 		ret += cursor.size;
1474 		buf += cursor.size;
1475 		amdgpu_res_next(&cursor, cursor.size);
1476 	}
1477 
1478 	return ret;
1479 }
1480 
1481 static void
amdgpu_bo_delete_mem_notify(struct ttm_buffer_object * bo)1482 amdgpu_bo_delete_mem_notify(struct ttm_buffer_object *bo)
1483 {
1484 	amdgpu_bo_move_notify(bo, false);
1485 }
1486 
1487 static struct ttm_device_funcs amdgpu_bo_driver = {
1488 	.ttm_tt_create = &amdgpu_ttm_tt_create,
1489 	.ttm_tt_populate = &amdgpu_ttm_tt_populate,
1490 	.ttm_tt_unpopulate = &amdgpu_ttm_tt_unpopulate,
1491 	.ttm_tt_destroy = &amdgpu_ttm_backend_destroy,
1492 	.eviction_valuable = amdgpu_ttm_bo_eviction_valuable,
1493 	.evict_flags = &amdgpu_evict_flags,
1494 	.move = &amdgpu_bo_move,
1495 	.delete_mem_notify = &amdgpu_bo_delete_mem_notify,
1496 	.release_notify = &amdgpu_bo_release_notify,
1497 	.io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
1498 	.io_mem_pfn = amdgpu_ttm_io_mem_pfn,
1499 	.access_memory = &amdgpu_ttm_access_memory,
1500 	.del_from_lru_notify = &amdgpu_vm_del_from_lru_notify
1501 };
1502 
1503 /*
1504  * Firmware Reservation functions
1505  */
1506 /**
1507  * amdgpu_ttm_fw_reserve_vram_fini - free fw reserved vram
1508  *
1509  * @adev: amdgpu_device pointer
1510  *
1511  * free fw reserved vram if it has been reserved.
1512  */
amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device * adev)1513 static void amdgpu_ttm_fw_reserve_vram_fini(struct amdgpu_device *adev)
1514 {
1515 	amdgpu_bo_free_kernel(&adev->mman.fw_vram_usage_reserved_bo,
1516 		NULL, &adev->mman.fw_vram_usage_va);
1517 }
1518 
1519 /**
1520  * amdgpu_ttm_fw_reserve_vram_init - create bo vram reservation from fw
1521  *
1522  * @adev: amdgpu_device pointer
1523  *
1524  * create bo vram reservation from fw.
1525  */
amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device * adev)1526 static int amdgpu_ttm_fw_reserve_vram_init(struct amdgpu_device *adev)
1527 {
1528 	uint64_t vram_size = adev->gmc.visible_vram_size;
1529 
1530 	adev->mman.fw_vram_usage_va = NULL;
1531 	adev->mman.fw_vram_usage_reserved_bo = NULL;
1532 
1533 	if (adev->mman.fw_vram_usage_size == 0 ||
1534 	    adev->mman.fw_vram_usage_size > vram_size)
1535 		return 0;
1536 
1537 	return amdgpu_bo_create_kernel_at(adev,
1538 					  adev->mman.fw_vram_usage_start_offset,
1539 					  adev->mman.fw_vram_usage_size,
1540 					  AMDGPU_GEM_DOMAIN_VRAM,
1541 					  &adev->mman.fw_vram_usage_reserved_bo,
1542 					  &adev->mman.fw_vram_usage_va);
1543 }
1544 
1545 /*
1546  * Memoy training reservation functions
1547  */
1548 
1549 /**
1550  * amdgpu_ttm_training_reserve_vram_fini - free memory training reserved vram
1551  *
1552  * @adev: amdgpu_device pointer
1553  *
1554  * free memory training reserved vram if it has been reserved.
1555  */
amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device * adev)1556 static int amdgpu_ttm_training_reserve_vram_fini(struct amdgpu_device *adev)
1557 {
1558 	struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1559 
1560 	ctx->init = PSP_MEM_TRAIN_NOT_SUPPORT;
1561 	amdgpu_bo_free_kernel(&ctx->c2p_bo, NULL, NULL);
1562 	ctx->c2p_bo = NULL;
1563 
1564 	return 0;
1565 }
1566 
amdgpu_ttm_training_data_block_init(struct amdgpu_device * adev)1567 static void amdgpu_ttm_training_data_block_init(struct amdgpu_device *adev)
1568 {
1569 	struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1570 
1571 	memset(ctx, 0, sizeof(*ctx));
1572 
1573 	ctx->c2p_train_data_offset =
1574 		ALIGN((adev->gmc.mc_vram_size - adev->mman.discovery_tmr_size - SZ_1M), SZ_1M);
1575 	ctx->p2c_train_data_offset =
1576 		(adev->gmc.mc_vram_size - GDDR6_MEM_TRAINING_OFFSET);
1577 	ctx->train_data_size =
1578 		GDDR6_MEM_TRAINING_DATA_SIZE_IN_BYTES;
1579 
1580 	DRM_DEBUG("train_data_size:%llx,p2c_train_data_offset:%llx,c2p_train_data_offset:%llx.\n",
1581 			ctx->train_data_size,
1582 			ctx->p2c_train_data_offset,
1583 			ctx->c2p_train_data_offset);
1584 }
1585 
1586 /*
1587  * reserve TMR memory at the top of VRAM which holds
1588  * IP Discovery data and is protected by PSP.
1589  */
amdgpu_ttm_reserve_tmr(struct amdgpu_device * adev)1590 static int amdgpu_ttm_reserve_tmr(struct amdgpu_device *adev)
1591 {
1592 	int ret;
1593 	struct psp_memory_training_context *ctx = &adev->psp.mem_train_ctx;
1594 	bool mem_train_support = false;
1595 
1596 	if (!amdgpu_sriov_vf(adev)) {
1597 		if (amdgpu_atomfirmware_mem_training_supported(adev))
1598 			mem_train_support = true;
1599 		else
1600 			DRM_DEBUG("memory training does not support!\n");
1601 	}
1602 
1603 	/*
1604 	 * Query reserved tmr size through atom firmwareinfo for Sienna_Cichlid and onwards for all
1605 	 * the use cases (IP discovery/G6 memory training/profiling/diagnostic data.etc)
1606 	 *
1607 	 * Otherwise, fallback to legacy approach to check and reserve tmr block for ip
1608 	 * discovery data and G6 memory training data respectively
1609 	 */
1610 	adev->mman.discovery_tmr_size =
1611 		amdgpu_atomfirmware_get_fw_reserved_fb_size(adev);
1612 	if (!adev->mman.discovery_tmr_size)
1613 		adev->mman.discovery_tmr_size = DISCOVERY_TMR_OFFSET;
1614 
1615 	if (mem_train_support) {
1616 		/* reserve vram for mem train according to TMR location */
1617 		amdgpu_ttm_training_data_block_init(adev);
1618 		ret = amdgpu_bo_create_kernel_at(adev,
1619 					 ctx->c2p_train_data_offset,
1620 					 ctx->train_data_size,
1621 					 AMDGPU_GEM_DOMAIN_VRAM,
1622 					 &ctx->c2p_bo,
1623 					 NULL);
1624 		if (ret) {
1625 			DRM_ERROR("alloc c2p_bo failed(%d)!\n", ret);
1626 			amdgpu_ttm_training_reserve_vram_fini(adev);
1627 			return ret;
1628 		}
1629 		ctx->init = PSP_MEM_TRAIN_RESERVE_SUCCESS;
1630 	}
1631 
1632 	ret = amdgpu_bo_create_kernel_at(adev,
1633 				adev->gmc.real_vram_size - adev->mman.discovery_tmr_size,
1634 				adev->mman.discovery_tmr_size,
1635 				AMDGPU_GEM_DOMAIN_VRAM,
1636 				&adev->mman.discovery_memory,
1637 				NULL);
1638 	if (ret) {
1639 		DRM_ERROR("alloc tmr failed(%d)!\n", ret);
1640 		amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1641 		return ret;
1642 	}
1643 
1644 	return 0;
1645 }
1646 
1647 /*
1648  * amdgpu_ttm_init - Init the memory management (ttm) as well as various
1649  * gtt/vram related fields.
1650  *
1651  * This initializes all of the memory space pools that the TTM layer
1652  * will need such as the GTT space (system memory mapped to the device),
1653  * VRAM (on-board memory), and on-chip memories (GDS, GWS, OA) which
1654  * can be mapped per VMID.
1655  */
amdgpu_ttm_init(struct amdgpu_device * adev)1656 int amdgpu_ttm_init(struct amdgpu_device *adev)
1657 {
1658 	uint64_t gtt_size;
1659 	int r;
1660 	u64 vis_vram_limit;
1661 
1662 	mutex_init(&adev->mman.gtt_window_lock);
1663 
1664 	/* No others user of address space so set it to 0 */
1665 	r = ttm_device_init(&adev->mman.bdev, &amdgpu_bo_driver, adev->dev,
1666 			       adev_to_drm(adev)->anon_inode->i_mapping,
1667 			       adev_to_drm(adev)->vma_offset_manager,
1668 			       adev->need_swiotlb,
1669 			       dma_addressing_limited(adev->dev));
1670 	if (r) {
1671 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
1672 		return r;
1673 	}
1674 	adev->mman.initialized = true;
1675 
1676 	/* Initialize VRAM pool with all of VRAM divided into pages */
1677 	r = amdgpu_vram_mgr_init(adev);
1678 	if (r) {
1679 		DRM_ERROR("Failed initializing VRAM heap.\n");
1680 		return r;
1681 	}
1682 
1683 	/* Reduce size of CPU-visible VRAM if requested */
1684 	vis_vram_limit = (u64)amdgpu_vis_vram_limit * 1024 * 1024;
1685 	if (amdgpu_vis_vram_limit > 0 &&
1686 	    vis_vram_limit <= adev->gmc.visible_vram_size)
1687 		adev->gmc.visible_vram_size = vis_vram_limit;
1688 
1689 	/* Change the size here instead of the init above so only lpfn is affected */
1690 	amdgpu_ttm_set_buffer_funcs_status(adev, false);
1691 #ifdef CONFIG_64BIT
1692 #ifdef CONFIG_X86
1693 	if (adev->gmc.xgmi.connected_to_cpu)
1694 		adev->mman.aper_base_kaddr = ioremap_cache(adev->gmc.aper_base,
1695 				adev->gmc.visible_vram_size);
1696 
1697 	else
1698 #endif
1699 		adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
1700 				adev->gmc.visible_vram_size);
1701 #endif
1702 
1703 	/*
1704 	 *The reserved vram for firmware must be pinned to the specified
1705 	 *place on the VRAM, so reserve it early.
1706 	 */
1707 	r = amdgpu_ttm_fw_reserve_vram_init(adev);
1708 	if (r) {
1709 		return r;
1710 	}
1711 
1712 	/*
1713 	 * only NAVI10 and onwards ASIC support for IP discovery.
1714 	 * If IP discovery enabled, a block of memory should be
1715 	 * reserved for IP discovey.
1716 	 */
1717 	if (adev->mman.discovery_bin) {
1718 		r = amdgpu_ttm_reserve_tmr(adev);
1719 		if (r)
1720 			return r;
1721 	}
1722 
1723 	/* allocate memory as required for VGA
1724 	 * This is used for VGA emulation and pre-OS scanout buffers to
1725 	 * avoid display artifacts while transitioning between pre-OS
1726 	 * and driver.  */
1727 	r = amdgpu_bo_create_kernel_at(adev, 0, adev->mman.stolen_vga_size,
1728 				       AMDGPU_GEM_DOMAIN_VRAM,
1729 				       &adev->mman.stolen_vga_memory,
1730 				       NULL);
1731 	if (r)
1732 		return r;
1733 	r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_vga_size,
1734 				       adev->mman.stolen_extended_size,
1735 				       AMDGPU_GEM_DOMAIN_VRAM,
1736 				       &adev->mman.stolen_extended_memory,
1737 				       NULL);
1738 	if (r)
1739 		return r;
1740 	r = amdgpu_bo_create_kernel_at(adev, adev->mman.stolen_reserved_offset,
1741 				       adev->mman.stolen_reserved_size,
1742 				       AMDGPU_GEM_DOMAIN_VRAM,
1743 				       &adev->mman.stolen_reserved_memory,
1744 				       NULL);
1745 	if (r)
1746 		return r;
1747 
1748 	DRM_INFO("amdgpu: %uM of VRAM memory ready\n",
1749 		 (unsigned) (adev->gmc.real_vram_size / (1024 * 1024)));
1750 
1751 	/* Compute GTT size, either bsaed on 3/4th the size of RAM size
1752 	 * or whatever the user passed on module init */
1753 	if (amdgpu_gtt_size == -1) {
1754 		struct sysinfo si;
1755 
1756 		si_meminfo(&si);
1757 		gtt_size = min(max((AMDGPU_DEFAULT_GTT_SIZE_MB << 20),
1758 			       adev->gmc.mc_vram_size),
1759 			       ((uint64_t)si.totalram * si.mem_unit * 3/4));
1760 	}
1761 	else
1762 		gtt_size = (uint64_t)amdgpu_gtt_size << 20;
1763 
1764 	/* Initialize GTT memory pool */
1765 	r = amdgpu_gtt_mgr_init(adev, gtt_size);
1766 	if (r) {
1767 		DRM_ERROR("Failed initializing GTT heap.\n");
1768 		return r;
1769 	}
1770 	DRM_INFO("amdgpu: %uM of GTT memory ready.\n",
1771 		 (unsigned)(gtt_size / (1024 * 1024)));
1772 
1773 	/* Initialize preemptible memory pool */
1774 	r = amdgpu_preempt_mgr_init(adev);
1775 	if (r) {
1776 		DRM_ERROR("Failed initializing PREEMPT heap.\n");
1777 		return r;
1778 	}
1779 
1780 	/* Initialize various on-chip memory pools */
1781 	r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size);
1782 	if (r) {
1783 		DRM_ERROR("Failed initializing GDS heap.\n");
1784 		return r;
1785 	}
1786 
1787 	r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size);
1788 	if (r) {
1789 		DRM_ERROR("Failed initializing gws heap.\n");
1790 		return r;
1791 	}
1792 
1793 	r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size);
1794 	if (r) {
1795 		DRM_ERROR("Failed initializing oa heap.\n");
1796 		return r;
1797 	}
1798 
1799 	return 0;
1800 }
1801 
1802 /*
1803  * amdgpu_ttm_fini - De-initialize the TTM memory pools
1804  */
amdgpu_ttm_fini(struct amdgpu_device * adev)1805 void amdgpu_ttm_fini(struct amdgpu_device *adev)
1806 {
1807 	if (!adev->mman.initialized)
1808 		return;
1809 
1810 	amdgpu_ttm_training_reserve_vram_fini(adev);
1811 	/* return the stolen vga memory back to VRAM */
1812 	amdgpu_bo_free_kernel(&adev->mman.stolen_vga_memory, NULL, NULL);
1813 	amdgpu_bo_free_kernel(&adev->mman.stolen_extended_memory, NULL, NULL);
1814 	/* return the IP Discovery TMR memory back to VRAM */
1815 	amdgpu_bo_free_kernel(&adev->mman.discovery_memory, NULL, NULL);
1816 	if (adev->mman.stolen_reserved_size)
1817 		amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
1818 				      NULL, NULL);
1819 	amdgpu_ttm_fw_reserve_vram_fini(adev);
1820 
1821 	amdgpu_vram_mgr_fini(adev);
1822 	amdgpu_gtt_mgr_fini(adev);
1823 	amdgpu_preempt_mgr_fini(adev);
1824 	ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS);
1825 	ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GWS);
1826 	ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
1827 	ttm_device_fini(&adev->mman.bdev);
1828 	adev->mman.initialized = false;
1829 	DRM_INFO("amdgpu: ttm finalized\n");
1830 }
1831 
1832 /**
1833  * amdgpu_ttm_set_buffer_funcs_status - enable/disable use of buffer functions
1834  *
1835  * @adev: amdgpu_device pointer
1836  * @enable: true when we can use buffer functions.
1837  *
1838  * Enable/disable use of buffer functions during suspend/resume. This should
1839  * only be called at bootup or when userspace isn't running.
1840  */
amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device * adev,bool enable)1841 void amdgpu_ttm_set_buffer_funcs_status(struct amdgpu_device *adev, bool enable)
1842 {
1843 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
1844 	uint64_t size;
1845 	int r;
1846 
1847 	if (!adev->mman.initialized || amdgpu_in_reset(adev) ||
1848 	    adev->mman.buffer_funcs_enabled == enable)
1849 		return;
1850 
1851 	if (enable) {
1852 		struct amdgpu_ring *ring;
1853 		struct drm_gpu_scheduler *sched;
1854 
1855 		ring = adev->mman.buffer_funcs_ring;
1856 		sched = &ring->sched;
1857 		r = drm_sched_entity_init(&adev->mman.entity,
1858 					  DRM_SCHED_PRIORITY_KERNEL, &sched,
1859 					  1, NULL);
1860 		if (r) {
1861 			DRM_ERROR("Failed setting up TTM BO move entity (%d)\n",
1862 				  r);
1863 			return;
1864 		}
1865 	} else {
1866 		drm_sched_entity_destroy(&adev->mman.entity);
1867 		dma_fence_put(man->move);
1868 		man->move = NULL;
1869 	}
1870 
1871 	/* this just adjusts TTM size idea, which sets lpfn to the correct value */
1872 	if (enable)
1873 		size = adev->gmc.real_vram_size;
1874 	else
1875 		size = adev->gmc.visible_vram_size;
1876 	man->size = size >> PAGE_SHIFT;
1877 	adev->mman.buffer_funcs_enabled = enable;
1878 }
1879 
amdgpu_copy_buffer(struct amdgpu_ring * ring,uint64_t src_offset,uint64_t dst_offset,uint32_t byte_count,struct dma_resv * resv,struct dma_fence ** fence,bool direct_submit,bool vm_needs_flush,bool tmz)1880 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
1881 		       uint64_t dst_offset, uint32_t byte_count,
1882 		       struct dma_resv *resv,
1883 		       struct dma_fence **fence, bool direct_submit,
1884 		       bool vm_needs_flush, bool tmz)
1885 {
1886 	enum amdgpu_ib_pool_type pool = direct_submit ? AMDGPU_IB_POOL_DIRECT :
1887 		AMDGPU_IB_POOL_DELAYED;
1888 	struct amdgpu_device *adev = ring->adev;
1889 	struct amdgpu_job *job;
1890 
1891 	uint32_t max_bytes;
1892 	unsigned num_loops, num_dw;
1893 	unsigned i;
1894 	int r;
1895 
1896 	if (!direct_submit && !ring->sched.ready) {
1897 		DRM_ERROR("Trying to move memory with ring turned off.\n");
1898 		return -EINVAL;
1899 	}
1900 
1901 	max_bytes = adev->mman.buffer_funcs->copy_max_bytes;
1902 	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
1903 	num_dw = ALIGN(num_loops * adev->mman.buffer_funcs->copy_num_dw, 8);
1904 
1905 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, pool, &job);
1906 	if (r)
1907 		return r;
1908 
1909 	if (vm_needs_flush) {
1910 		job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gmc.pdb0_bo ?
1911 					adev->gmc.pdb0_bo : adev->gart.bo);
1912 		job->vm_needs_flush = true;
1913 	}
1914 	if (resv) {
1915 		r = amdgpu_sync_resv(adev, &job->sync, resv,
1916 				     AMDGPU_SYNC_ALWAYS,
1917 				     AMDGPU_FENCE_OWNER_UNDEFINED);
1918 		if (r) {
1919 			DRM_ERROR("sync failed (%d).\n", r);
1920 			goto error_free;
1921 		}
1922 	}
1923 
1924 	for (i = 0; i < num_loops; i++) {
1925 		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
1926 
1927 		amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_offset,
1928 					dst_offset, cur_size_in_bytes, tmz);
1929 
1930 		src_offset += cur_size_in_bytes;
1931 		dst_offset += cur_size_in_bytes;
1932 		byte_count -= cur_size_in_bytes;
1933 	}
1934 
1935 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
1936 	WARN_ON(job->ibs[0].length_dw > num_dw);
1937 	if (direct_submit)
1938 		r = amdgpu_job_submit_direct(job, ring, fence);
1939 	else
1940 		r = amdgpu_job_submit(job, &adev->mman.entity,
1941 				      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
1942 	if (r)
1943 		goto error_free;
1944 
1945 	return r;
1946 
1947 error_free:
1948 	amdgpu_job_free(job);
1949 	DRM_ERROR("Error scheduling IBs (%d)\n", r);
1950 	return r;
1951 }
1952 
amdgpu_fill_buffer(struct amdgpu_bo * bo,uint32_t src_data,struct dma_resv * resv,struct dma_fence ** fence)1953 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
1954 		       uint32_t src_data,
1955 		       struct dma_resv *resv,
1956 		       struct dma_fence **fence)
1957 {
1958 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1959 	uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
1960 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
1961 
1962 	struct amdgpu_res_cursor cursor;
1963 	unsigned int num_loops, num_dw;
1964 	uint64_t num_bytes;
1965 
1966 	struct amdgpu_job *job;
1967 	int r;
1968 
1969 	if (!adev->mman.buffer_funcs_enabled) {
1970 		DRM_ERROR("Trying to clear memory with ring turned off.\n");
1971 		return -EINVAL;
1972 	}
1973 
1974 	if (bo->tbo.resource->mem_type == AMDGPU_PL_PREEMPT) {
1975 		DRM_ERROR("Trying to clear preemptible memory.\n");
1976 		return -EINVAL;
1977 	}
1978 
1979 	if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1980 		r = amdgpu_ttm_alloc_gart(&bo->tbo);
1981 		if (r)
1982 			return r;
1983 	}
1984 
1985 	num_bytes = bo->tbo.resource->num_pages << PAGE_SHIFT;
1986 	num_loops = 0;
1987 
1988 	amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor);
1989 	while (cursor.remaining) {
1990 		num_loops += DIV_ROUND_UP_ULL(cursor.size, max_bytes);
1991 		amdgpu_res_next(&cursor, cursor.size);
1992 	}
1993 	num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
1994 
1995 	/* for IB padding */
1996 	num_dw += 64;
1997 
1998 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED,
1999 				     &job);
2000 	if (r)
2001 		return r;
2002 
2003 	if (resv) {
2004 		r = amdgpu_sync_resv(adev, &job->sync, resv,
2005 				     AMDGPU_SYNC_ALWAYS,
2006 				     AMDGPU_FENCE_OWNER_UNDEFINED);
2007 		if (r) {
2008 			DRM_ERROR("sync failed (%d).\n", r);
2009 			goto error_free;
2010 		}
2011 	}
2012 
2013 	amdgpu_res_first(bo->tbo.resource, 0, num_bytes, &cursor);
2014 	while (cursor.remaining) {
2015 		uint32_t cur_size = min_t(uint64_t, cursor.size, max_bytes);
2016 		uint64_t dst_addr = cursor.start;
2017 
2018 		dst_addr += amdgpu_ttm_domain_start(adev,
2019 						    bo->tbo.resource->mem_type);
2020 		amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data, dst_addr,
2021 					cur_size);
2022 
2023 		amdgpu_res_next(&cursor, cur_size);
2024 	}
2025 
2026 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
2027 	WARN_ON(job->ibs[0].length_dw > num_dw);
2028 	r = amdgpu_job_submit(job, &adev->mman.entity,
2029 			      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
2030 	if (r)
2031 		goto error_free;
2032 
2033 	return 0;
2034 
2035 error_free:
2036 	amdgpu_job_free(job);
2037 	return r;
2038 }
2039 
2040 /**
2041  * amdgpu_ttm_evict_resources - evict memory buffers
2042  * @adev: amdgpu device object
2043  * @mem_type: evicted BO's memory type
2044  *
2045  * Evicts all @mem_type buffers on the lru list of the memory type.
2046  *
2047  * Returns:
2048  * 0 for success or a negative error code on failure.
2049  */
amdgpu_ttm_evict_resources(struct amdgpu_device * adev,int mem_type)2050 int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type)
2051 {
2052 	struct ttm_resource_manager *man;
2053 
2054 	switch (mem_type) {
2055 	case TTM_PL_VRAM:
2056 	case TTM_PL_TT:
2057 	case AMDGPU_PL_GWS:
2058 	case AMDGPU_PL_GDS:
2059 	case AMDGPU_PL_OA:
2060 		man = ttm_manager_type(&adev->mman.bdev, mem_type);
2061 		break;
2062 	default:
2063 		DRM_ERROR("Trying to evict invalid memory type\n");
2064 		return -EINVAL;
2065 	}
2066 
2067 	return ttm_resource_manager_evict_all(&adev->mman.bdev, man);
2068 }
2069 
2070 #if defined(CONFIG_DEBUG_FS)
2071 
amdgpu_mm_vram_table_show(struct seq_file * m,void * unused)2072 static int amdgpu_mm_vram_table_show(struct seq_file *m, void *unused)
2073 {
2074 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2075 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2076 							    TTM_PL_VRAM);
2077 	struct drm_printer p = drm_seq_file_printer(m);
2078 
2079 	man->func->debug(man, &p);
2080 	return 0;
2081 }
2082 
amdgpu_ttm_page_pool_show(struct seq_file * m,void * unused)2083 static int amdgpu_ttm_page_pool_show(struct seq_file *m, void *unused)
2084 {
2085 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2086 
2087 	return ttm_pool_debugfs(&adev->mman.bdev.pool, m);
2088 }
2089 
amdgpu_mm_tt_table_show(struct seq_file * m,void * unused)2090 static int amdgpu_mm_tt_table_show(struct seq_file *m, void *unused)
2091 {
2092 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2093 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2094 							    TTM_PL_TT);
2095 	struct drm_printer p = drm_seq_file_printer(m);
2096 
2097 	man->func->debug(man, &p);
2098 	return 0;
2099 }
2100 
amdgpu_mm_gds_table_show(struct seq_file * m,void * unused)2101 static int amdgpu_mm_gds_table_show(struct seq_file *m, void *unused)
2102 {
2103 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2104 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2105 							    AMDGPU_PL_GDS);
2106 	struct drm_printer p = drm_seq_file_printer(m);
2107 
2108 	man->func->debug(man, &p);
2109 	return 0;
2110 }
2111 
amdgpu_mm_gws_table_show(struct seq_file * m,void * unused)2112 static int amdgpu_mm_gws_table_show(struct seq_file *m, void *unused)
2113 {
2114 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2115 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2116 							    AMDGPU_PL_GWS);
2117 	struct drm_printer p = drm_seq_file_printer(m);
2118 
2119 	man->func->debug(man, &p);
2120 	return 0;
2121 }
2122 
amdgpu_mm_oa_table_show(struct seq_file * m,void * unused)2123 static int amdgpu_mm_oa_table_show(struct seq_file *m, void *unused)
2124 {
2125 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
2126 	struct ttm_resource_manager *man = ttm_manager_type(&adev->mman.bdev,
2127 							    AMDGPU_PL_OA);
2128 	struct drm_printer p = drm_seq_file_printer(m);
2129 
2130 	man->func->debug(man, &p);
2131 	return 0;
2132 }
2133 
2134 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_vram_table);
2135 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_tt_table);
2136 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gds_table);
2137 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_gws_table);
2138 DEFINE_SHOW_ATTRIBUTE(amdgpu_mm_oa_table);
2139 DEFINE_SHOW_ATTRIBUTE(amdgpu_ttm_page_pool);
2140 
2141 /*
2142  * amdgpu_ttm_vram_read - Linear read access to VRAM
2143  *
2144  * Accesses VRAM via MMIO for debugging purposes.
2145  */
amdgpu_ttm_vram_read(struct file * f,char __user * buf,size_t size,loff_t * pos)2146 static ssize_t amdgpu_ttm_vram_read(struct file *f, char __user *buf,
2147 				    size_t size, loff_t *pos)
2148 {
2149 	struct amdgpu_device *adev = file_inode(f)->i_private;
2150 	ssize_t result = 0;
2151 
2152 	if (size & 0x3 || *pos & 0x3)
2153 		return -EINVAL;
2154 
2155 	if (*pos >= adev->gmc.mc_vram_size)
2156 		return -ENXIO;
2157 
2158 	size = min(size, (size_t)(adev->gmc.mc_vram_size - *pos));
2159 	while (size) {
2160 		size_t bytes = min(size, AMDGPU_TTM_VRAM_MAX_DW_READ * 4);
2161 		uint32_t value[AMDGPU_TTM_VRAM_MAX_DW_READ];
2162 
2163 		amdgpu_device_vram_access(adev, *pos, value, bytes, false);
2164 		if (copy_to_user(buf, value, bytes))
2165 			return -EFAULT;
2166 
2167 		result += bytes;
2168 		buf += bytes;
2169 		*pos += bytes;
2170 		size -= bytes;
2171 	}
2172 
2173 	return result;
2174 }
2175 
2176 /*
2177  * amdgpu_ttm_vram_write - Linear write access to VRAM
2178  *
2179  * Accesses VRAM via MMIO for debugging purposes.
2180  */
amdgpu_ttm_vram_write(struct file * f,const char __user * buf,size_t size,loff_t * pos)2181 static ssize_t amdgpu_ttm_vram_write(struct file *f, const char __user *buf,
2182 				    size_t size, loff_t *pos)
2183 {
2184 	struct amdgpu_device *adev = file_inode(f)->i_private;
2185 	ssize_t result = 0;
2186 	int r;
2187 
2188 	if (size & 0x3 || *pos & 0x3)
2189 		return -EINVAL;
2190 
2191 	if (*pos >= adev->gmc.mc_vram_size)
2192 		return -ENXIO;
2193 
2194 	while (size) {
2195 		uint32_t value;
2196 
2197 		if (*pos >= adev->gmc.mc_vram_size)
2198 			return result;
2199 
2200 		r = get_user(value, (uint32_t *)buf);
2201 		if (r)
2202 			return r;
2203 
2204 		amdgpu_device_mm_access(adev, *pos, &value, 4, true);
2205 
2206 		result += 4;
2207 		buf += 4;
2208 		*pos += 4;
2209 		size -= 4;
2210 	}
2211 
2212 	return result;
2213 }
2214 
2215 static const struct file_operations amdgpu_ttm_vram_fops = {
2216 	.owner = THIS_MODULE,
2217 	.read = amdgpu_ttm_vram_read,
2218 	.write = amdgpu_ttm_vram_write,
2219 	.llseek = default_llseek,
2220 };
2221 
2222 /*
2223  * amdgpu_iomem_read - Virtual read access to GPU mapped memory
2224  *
2225  * This function is used to read memory that has been mapped to the
2226  * GPU and the known addresses are not physical addresses but instead
2227  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2228  */
amdgpu_iomem_read(struct file * f,char __user * buf,size_t size,loff_t * pos)2229 static ssize_t amdgpu_iomem_read(struct file *f, char __user *buf,
2230 				 size_t size, loff_t *pos)
2231 {
2232 	struct amdgpu_device *adev = file_inode(f)->i_private;
2233 	struct iommu_domain *dom;
2234 	ssize_t result = 0;
2235 	int r;
2236 
2237 	/* retrieve the IOMMU domain if any for this device */
2238 	dom = iommu_get_domain_for_dev(adev->dev);
2239 
2240 	while (size) {
2241 		phys_addr_t addr = *pos & PAGE_MASK;
2242 		loff_t off = *pos & ~PAGE_MASK;
2243 		size_t bytes = PAGE_SIZE - off;
2244 		unsigned long pfn;
2245 		struct page *p;
2246 		void *ptr;
2247 
2248 		bytes = bytes < size ? bytes : size;
2249 
2250 		/* Translate the bus address to a physical address.  If
2251 		 * the domain is NULL it means there is no IOMMU active
2252 		 * and the address translation is the identity
2253 		 */
2254 		addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2255 
2256 		pfn = addr >> PAGE_SHIFT;
2257 		if (!pfn_valid(pfn))
2258 			return -EPERM;
2259 
2260 		p = pfn_to_page(pfn);
2261 		if (p->mapping != adev->mman.bdev.dev_mapping)
2262 			return -EPERM;
2263 
2264 		ptr = kmap(p);
2265 		r = copy_to_user(buf, ptr + off, bytes);
2266 		kunmap(p);
2267 		if (r)
2268 			return -EFAULT;
2269 
2270 		size -= bytes;
2271 		*pos += bytes;
2272 		result += bytes;
2273 	}
2274 
2275 	return result;
2276 }
2277 
2278 /*
2279  * amdgpu_iomem_write - Virtual write access to GPU mapped memory
2280  *
2281  * This function is used to write memory that has been mapped to the
2282  * GPU and the known addresses are not physical addresses but instead
2283  * bus addresses (e.g., what you'd put in an IB or ring buffer).
2284  */
amdgpu_iomem_write(struct file * f,const char __user * buf,size_t size,loff_t * pos)2285 static ssize_t amdgpu_iomem_write(struct file *f, const char __user *buf,
2286 				 size_t size, loff_t *pos)
2287 {
2288 	struct amdgpu_device *adev = file_inode(f)->i_private;
2289 	struct iommu_domain *dom;
2290 	ssize_t result = 0;
2291 	int r;
2292 
2293 	dom = iommu_get_domain_for_dev(adev->dev);
2294 
2295 	while (size) {
2296 		phys_addr_t addr = *pos & PAGE_MASK;
2297 		loff_t off = *pos & ~PAGE_MASK;
2298 		size_t bytes = PAGE_SIZE - off;
2299 		unsigned long pfn;
2300 		struct page *p;
2301 		void *ptr;
2302 
2303 		bytes = bytes < size ? bytes : size;
2304 
2305 		addr = dom ? iommu_iova_to_phys(dom, addr) : addr;
2306 
2307 		pfn = addr >> PAGE_SHIFT;
2308 		if (!pfn_valid(pfn))
2309 			return -EPERM;
2310 
2311 		p = pfn_to_page(pfn);
2312 		if (p->mapping != adev->mman.bdev.dev_mapping)
2313 			return -EPERM;
2314 
2315 		ptr = kmap(p);
2316 		r = copy_from_user(ptr + off, buf, bytes);
2317 		kunmap(p);
2318 		if (r)
2319 			return -EFAULT;
2320 
2321 		size -= bytes;
2322 		*pos += bytes;
2323 		result += bytes;
2324 	}
2325 
2326 	return result;
2327 }
2328 
2329 static const struct file_operations amdgpu_ttm_iomem_fops = {
2330 	.owner = THIS_MODULE,
2331 	.read = amdgpu_iomem_read,
2332 	.write = amdgpu_iomem_write,
2333 	.llseek = default_llseek
2334 };
2335 
2336 #endif
2337 
amdgpu_ttm_debugfs_init(struct amdgpu_device * adev)2338 void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev)
2339 {
2340 #if defined(CONFIG_DEBUG_FS)
2341 	struct drm_minor *minor = adev_to_drm(adev)->primary;
2342 	struct dentry *root = minor->debugfs_root;
2343 
2344 	debugfs_create_file_size("amdgpu_vram", 0444, root, adev,
2345 				 &amdgpu_ttm_vram_fops, adev->gmc.mc_vram_size);
2346 	debugfs_create_file("amdgpu_iomem", 0444, root, adev,
2347 			    &amdgpu_ttm_iomem_fops);
2348 	debugfs_create_file("amdgpu_vram_mm", 0444, root, adev,
2349 			    &amdgpu_mm_vram_table_fops);
2350 	debugfs_create_file("amdgpu_gtt_mm", 0444, root, adev,
2351 			    &amdgpu_mm_tt_table_fops);
2352 	debugfs_create_file("amdgpu_gds_mm", 0444, root, adev,
2353 			    &amdgpu_mm_gds_table_fops);
2354 	debugfs_create_file("amdgpu_gws_mm", 0444, root, adev,
2355 			    &amdgpu_mm_gws_table_fops);
2356 	debugfs_create_file("amdgpu_oa_mm", 0444, root, adev,
2357 			    &amdgpu_mm_oa_table_fops);
2358 	debugfs_create_file("ttm_page_pool", 0444, root, adev,
2359 			    &amdgpu_ttm_page_pool_fops);
2360 #endif
2361 }
2362