xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c (revision 220994d61cebfc04f071d69049127657c7e8191b)
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 <linux/list.h>
33 #include <linux/slab.h>
34 #include <linux/dma-buf.h>
35 #include <linux/export.h>
36 
37 #include <drm/drm_drv.h>
38 #include <drm/amdgpu_drm.h>
39 #include <drm/drm_cache.h>
40 #include "amdgpu.h"
41 #include "amdgpu_trace.h"
42 #include "amdgpu_amdkfd.h"
43 #include "amdgpu_vram_mgr.h"
44 #include "amdgpu_vm.h"
45 #include "amdgpu_dma_buf.h"
46 
47 /**
48  * DOC: amdgpu_object
49  *
50  * This defines the interfaces to operate on an &amdgpu_bo buffer object which
51  * represents memory used by driver (VRAM, system memory, etc.). The driver
52  * provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces
53  * to create/destroy/set buffer object which are then managed by the kernel TTM
54  * memory manager.
55  * The interfaces are also used internally by kernel clients, including gfx,
56  * uvd, etc. for kernel managed allocations used by the GPU.
57  *
58  */
59 
amdgpu_bo_destroy(struct ttm_buffer_object * tbo)60 static void amdgpu_bo_destroy(struct ttm_buffer_object *tbo)
61 {
62 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
63 
64 	amdgpu_bo_kunmap(bo);
65 
66 	if (drm_gem_is_imported(&bo->tbo.base))
67 		drm_prime_gem_destroy(&bo->tbo.base, bo->tbo.sg);
68 	drm_gem_object_release(&bo->tbo.base);
69 	amdgpu_bo_unref(&bo->parent);
70 	kvfree(bo);
71 }
72 
amdgpu_bo_user_destroy(struct ttm_buffer_object * tbo)73 static void amdgpu_bo_user_destroy(struct ttm_buffer_object *tbo)
74 {
75 	struct amdgpu_bo *bo = ttm_to_amdgpu_bo(tbo);
76 	struct amdgpu_bo_user *ubo;
77 
78 	ubo = to_amdgpu_bo_user(bo);
79 	kfree(ubo->metadata);
80 	amdgpu_bo_destroy(tbo);
81 }
82 
83 /**
84  * amdgpu_bo_is_amdgpu_bo - check if the buffer object is an &amdgpu_bo
85  * @bo: buffer object to be checked
86  *
87  * Uses destroy function associated with the object to determine if this is
88  * an &amdgpu_bo.
89  *
90  * Returns:
91  * true if the object belongs to &amdgpu_bo, false if not.
92  */
amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object * bo)93 bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
94 {
95 	if (bo->destroy == &amdgpu_bo_destroy ||
96 	    bo->destroy == &amdgpu_bo_user_destroy)
97 		return true;
98 
99 	return false;
100 }
101 
102 /**
103  * amdgpu_bo_placement_from_domain - set buffer's placement
104  * @abo: &amdgpu_bo buffer object whose placement is to be set
105  * @domain: requested domain
106  *
107  * Sets buffer's placement according to requested domain and the buffer's
108  * flags.
109  */
amdgpu_bo_placement_from_domain(struct amdgpu_bo * abo,u32 domain)110 void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)
111 {
112 	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
113 	struct ttm_placement *placement = &abo->placement;
114 	struct ttm_place *places = abo->placements;
115 	u64 flags = abo->flags;
116 	u32 c = 0;
117 
118 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
119 		unsigned int visible_pfn = adev->gmc.visible_vram_size >> PAGE_SHIFT;
120 		int8_t mem_id = KFD_XCP_MEM_ID(adev, abo->xcp_id);
121 
122 		if (adev->gmc.mem_partitions && mem_id >= 0) {
123 			places[c].fpfn = adev->gmc.mem_partitions[mem_id].range.fpfn;
124 			/*
125 			 * memory partition range lpfn is inclusive start + size - 1
126 			 * TTM place lpfn is exclusive start + size
127 			 */
128 			places[c].lpfn = adev->gmc.mem_partitions[mem_id].range.lpfn + 1;
129 		} else {
130 			places[c].fpfn = 0;
131 			places[c].lpfn = 0;
132 		}
133 		places[c].mem_type = TTM_PL_VRAM;
134 		places[c].flags = 0;
135 
136 		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
137 			places[c].lpfn = min_not_zero(places[c].lpfn, visible_pfn);
138 		else
139 			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
140 
141 		if (abo->tbo.type == ttm_bo_type_kernel &&
142 		    flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
143 			places[c].flags |= TTM_PL_FLAG_CONTIGUOUS;
144 
145 		c++;
146 	}
147 
148 	if (domain & AMDGPU_GEM_DOMAIN_DOORBELL) {
149 		places[c].fpfn = 0;
150 		places[c].lpfn = 0;
151 		places[c].mem_type = AMDGPU_PL_DOORBELL;
152 		places[c].flags = 0;
153 		c++;
154 	}
155 
156 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
157 		places[c].fpfn = 0;
158 		places[c].lpfn = 0;
159 		places[c].mem_type =
160 			abo->flags & AMDGPU_GEM_CREATE_PREEMPTIBLE ?
161 			AMDGPU_PL_PREEMPT : TTM_PL_TT;
162 		places[c].flags = 0;
163 		/*
164 		 * When GTT is just an alternative to VRAM make sure that we
165 		 * only use it as fallback and still try to fill up VRAM first.
166 		 */
167 		if (abo->tbo.resource && !(adev->flags & AMD_IS_APU) &&
168 		    domain & abo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM)
169 			places[c].flags |= TTM_PL_FLAG_FALLBACK;
170 		c++;
171 	}
172 
173 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
174 		places[c].fpfn = 0;
175 		places[c].lpfn = 0;
176 		places[c].mem_type = TTM_PL_SYSTEM;
177 		places[c].flags = 0;
178 		c++;
179 	}
180 
181 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
182 		places[c].fpfn = 0;
183 		places[c].lpfn = 0;
184 		places[c].mem_type = AMDGPU_PL_GDS;
185 		places[c].flags = 0;
186 		c++;
187 	}
188 
189 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
190 		places[c].fpfn = 0;
191 		places[c].lpfn = 0;
192 		places[c].mem_type = AMDGPU_PL_GWS;
193 		places[c].flags = 0;
194 		c++;
195 	}
196 
197 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
198 		places[c].fpfn = 0;
199 		places[c].lpfn = 0;
200 		places[c].mem_type = AMDGPU_PL_OA;
201 		places[c].flags = 0;
202 		c++;
203 	}
204 
205 	if (!c) {
206 		places[c].fpfn = 0;
207 		places[c].lpfn = 0;
208 		places[c].mem_type = TTM_PL_SYSTEM;
209 		places[c].flags = 0;
210 		c++;
211 	}
212 
213 	BUG_ON(c > AMDGPU_BO_MAX_PLACEMENTS);
214 
215 	placement->num_placement = c;
216 	placement->placement = places;
217 }
218 
219 /**
220  * amdgpu_bo_create_reserved - create reserved BO for kernel use
221  *
222  * @adev: amdgpu device object
223  * @size: size for the new BO
224  * @align: alignment for the new BO
225  * @domain: where to place it
226  * @bo_ptr: used to initialize BOs in structures
227  * @gpu_addr: GPU addr of the pinned BO
228  * @cpu_addr: optional CPU address mapping
229  *
230  * Allocates and pins a BO for kernel internal use, and returns it still
231  * reserved.
232  *
233  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
234  *
235  * Returns:
236  * 0 on success, negative error code otherwise.
237  */
amdgpu_bo_create_reserved(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)238 int amdgpu_bo_create_reserved(struct amdgpu_device *adev,
239 			      unsigned long size, int align,
240 			      u32 domain, struct amdgpu_bo **bo_ptr,
241 			      u64 *gpu_addr, void **cpu_addr)
242 {
243 	struct amdgpu_bo_param bp;
244 	bool free = false;
245 	int r;
246 
247 	if (!size) {
248 		amdgpu_bo_unref(bo_ptr);
249 		return 0;
250 	}
251 
252 	memset(&bp, 0, sizeof(bp));
253 	bp.size = size;
254 	bp.byte_align = align;
255 	bp.domain = domain;
256 	bp.flags = cpu_addr ? AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
257 		: AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
258 	bp.flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
259 	bp.type = ttm_bo_type_kernel;
260 	bp.resv = NULL;
261 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
262 
263 	if (!*bo_ptr) {
264 		r = amdgpu_bo_create(adev, &bp, bo_ptr);
265 		if (r) {
266 			dev_err(adev->dev, "(%d) failed to allocate kernel bo\n",
267 				r);
268 			return r;
269 		}
270 		free = true;
271 	}
272 
273 	r = amdgpu_bo_reserve(*bo_ptr, false);
274 	if (r) {
275 		dev_err(adev->dev, "(%d) failed to reserve kernel bo\n", r);
276 		goto error_free;
277 	}
278 
279 	r = amdgpu_bo_pin(*bo_ptr, domain);
280 	if (r) {
281 		dev_err(adev->dev, "(%d) kernel bo pin failed\n", r);
282 		goto error_unreserve;
283 	}
284 
285 	r = amdgpu_ttm_alloc_gart(&(*bo_ptr)->tbo);
286 	if (r) {
287 		dev_err(adev->dev, "%p bind failed\n", *bo_ptr);
288 		goto error_unpin;
289 	}
290 
291 	if (gpu_addr)
292 		*gpu_addr = amdgpu_bo_gpu_offset(*bo_ptr);
293 
294 	if (cpu_addr) {
295 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
296 		if (r) {
297 			dev_err(adev->dev, "(%d) kernel bo map failed\n", r);
298 			goto error_unpin;
299 		}
300 	}
301 
302 	return 0;
303 
304 error_unpin:
305 	amdgpu_bo_unpin(*bo_ptr);
306 error_unreserve:
307 	amdgpu_bo_unreserve(*bo_ptr);
308 
309 error_free:
310 	if (free)
311 		amdgpu_bo_unref(bo_ptr);
312 
313 	return r;
314 }
315 
316 /**
317  * amdgpu_bo_create_kernel - create BO for kernel use
318  *
319  * @adev: amdgpu device object
320  * @size: size for the new BO
321  * @align: alignment for the new BO
322  * @domain: where to place it
323  * @bo_ptr:  used to initialize BOs in structures
324  * @gpu_addr: GPU addr of the pinned BO
325  * @cpu_addr: optional CPU address mapping
326  *
327  * Allocates and pins a BO for kernel internal use.
328  *
329  * This function is exported to allow the V4L2 isp device
330  * external to drm device to create and access the kernel BO.
331  *
332  * Note: For bo_ptr new BO is only created if bo_ptr points to NULL.
333  *
334  * Returns:
335  * 0 on success, negative error code otherwise.
336  */
amdgpu_bo_create_kernel(struct amdgpu_device * adev,unsigned long size,int align,u32 domain,struct amdgpu_bo ** bo_ptr,u64 * gpu_addr,void ** cpu_addr)337 int amdgpu_bo_create_kernel(struct amdgpu_device *adev,
338 			    unsigned long size, int align,
339 			    u32 domain, struct amdgpu_bo **bo_ptr,
340 			    u64 *gpu_addr, void **cpu_addr)
341 {
342 	int r;
343 
344 	r = amdgpu_bo_create_reserved(adev, size, align, domain, bo_ptr,
345 				      gpu_addr, cpu_addr);
346 
347 	if (r)
348 		return r;
349 
350 	if (*bo_ptr)
351 		amdgpu_bo_unreserve(*bo_ptr);
352 
353 	return 0;
354 }
355 
356 /**
357  * amdgpu_bo_create_isp_user - create user BO for isp
358  *
359  * @adev: amdgpu device object
360  * @dma_buf: DMABUF handle for isp buffer
361  * @domain: where to place it
362  * @bo:  used to initialize BOs in structures
363  * @gpu_addr: GPU addr of the pinned BO
364  *
365  * Imports isp DMABUF to allocate and pin a user BO for isp internal use. It does
366  * GART alloc to generate gpu_addr for BO to make it accessible through the
367  * GART aperture for ISP HW.
368  *
369  * This function is exported to allow the V4L2 isp device external to drm device
370  * to create and access the isp user BO.
371  *
372  * Returns:
373  * 0 on success, negative error code otherwise.
374  */
amdgpu_bo_create_isp_user(struct amdgpu_device * adev,struct dma_buf * dma_buf,u32 domain,struct amdgpu_bo ** bo,u64 * gpu_addr)375 int amdgpu_bo_create_isp_user(struct amdgpu_device *adev,
376 			   struct dma_buf *dma_buf, u32 domain, struct amdgpu_bo **bo,
377 			   u64 *gpu_addr)
378 
379 {
380 	struct drm_gem_object *gem_obj;
381 	int r;
382 
383 	gem_obj = amdgpu_gem_prime_import(&adev->ddev, dma_buf);
384 	*bo = gem_to_amdgpu_bo(gem_obj);
385 	if (!(*bo)) {
386 		dev_err(adev->dev, "failed to get valid isp user bo\n");
387 		return -EINVAL;
388 	}
389 
390 	r = amdgpu_bo_reserve(*bo, false);
391 	if (r) {
392 		dev_err(adev->dev, "(%d) failed to reserve isp user bo\n", r);
393 		return r;
394 	}
395 
396 	r = amdgpu_bo_pin(*bo, domain);
397 	if (r) {
398 		dev_err(adev->dev, "(%d) isp user bo pin failed\n", r);
399 		goto error_unreserve;
400 	}
401 
402 	r = amdgpu_ttm_alloc_gart(&(*bo)->tbo);
403 	if (r) {
404 		dev_err(adev->dev, "%p bind failed\n", *bo);
405 		goto error_unpin;
406 	}
407 
408 	if (!WARN_ON(!gpu_addr))
409 		*gpu_addr = amdgpu_bo_gpu_offset(*bo);
410 
411 	amdgpu_bo_unreserve(*bo);
412 
413 	return 0;
414 
415 error_unpin:
416 	amdgpu_bo_unpin(*bo);
417 error_unreserve:
418 	amdgpu_bo_unreserve(*bo);
419 	amdgpu_bo_unref(bo);
420 
421 	return r;
422 }
423 
424 /**
425  * amdgpu_bo_create_kernel_at - create BO for kernel use at specific location
426  *
427  * @adev: amdgpu device object
428  * @offset: offset of the BO
429  * @size: size of the BO
430  * @bo_ptr:  used to initialize BOs in structures
431  * @cpu_addr: optional CPU address mapping
432  *
433  * Creates a kernel BO at a specific offset in VRAM.
434  *
435  * Returns:
436  * 0 on success, negative error code otherwise.
437  */
amdgpu_bo_create_kernel_at(struct amdgpu_device * adev,uint64_t offset,uint64_t size,struct amdgpu_bo ** bo_ptr,void ** cpu_addr)438 int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
439 			       uint64_t offset, uint64_t size,
440 			       struct amdgpu_bo **bo_ptr, void **cpu_addr)
441 {
442 	struct ttm_operation_ctx ctx = { false, false };
443 	unsigned int i;
444 	int r;
445 
446 	offset &= PAGE_MASK;
447 	size = ALIGN(size, PAGE_SIZE);
448 
449 	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
450 				      AMDGPU_GEM_DOMAIN_VRAM, bo_ptr, NULL,
451 				      cpu_addr);
452 	if (r)
453 		return r;
454 
455 	if ((*bo_ptr) == NULL)
456 		return 0;
457 
458 	/*
459 	 * Remove the original mem node and create a new one at the request
460 	 * position.
461 	 */
462 	if (cpu_addr)
463 		amdgpu_bo_kunmap(*bo_ptr);
464 
465 	ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
466 
467 	for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
468 		(*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
469 		(*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
470 	}
471 	r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
472 			     &(*bo_ptr)->tbo.resource, &ctx);
473 	if (r)
474 		goto error;
475 
476 	if (cpu_addr) {
477 		r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
478 		if (r)
479 			goto error;
480 	}
481 
482 	amdgpu_bo_unreserve(*bo_ptr);
483 	return 0;
484 
485 error:
486 	amdgpu_bo_unreserve(*bo_ptr);
487 	amdgpu_bo_unref(bo_ptr);
488 	return r;
489 }
490 
491 /**
492  * amdgpu_bo_free_kernel - free BO for kernel use
493  *
494  * @bo: amdgpu BO to free
495  * @gpu_addr: pointer to where the BO's GPU memory space address was stored
496  * @cpu_addr: pointer to where the BO's CPU memory space address was stored
497  *
498  * unmaps and unpin a BO for kernel internal use.
499  *
500  * This function is exported to allow the V4L2 isp device
501  * external to drm device to free the kernel BO.
502  */
amdgpu_bo_free_kernel(struct amdgpu_bo ** bo,u64 * gpu_addr,void ** cpu_addr)503 void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr,
504 			   void **cpu_addr)
505 {
506 	if (*bo == NULL)
507 		return;
508 
509 	WARN_ON(amdgpu_ttm_adev((*bo)->tbo.bdev)->in_suspend);
510 
511 	if (likely(amdgpu_bo_reserve(*bo, true) == 0)) {
512 		if (cpu_addr)
513 			amdgpu_bo_kunmap(*bo);
514 
515 		amdgpu_bo_unpin(*bo);
516 		amdgpu_bo_unreserve(*bo);
517 	}
518 	amdgpu_bo_unref(bo);
519 
520 	if (gpu_addr)
521 		*gpu_addr = 0;
522 
523 	if (cpu_addr)
524 		*cpu_addr = NULL;
525 }
526 
527 /**
528  * amdgpu_bo_free_isp_user - free BO for isp use
529  *
530  * @bo: amdgpu isp user BO to free
531  *
532  * unpin and unref BO for isp internal use.
533  *
534  * This function is exported to allow the V4L2 isp device
535  * external to drm device to free the isp user BO.
536  */
amdgpu_bo_free_isp_user(struct amdgpu_bo * bo)537 void amdgpu_bo_free_isp_user(struct amdgpu_bo *bo)
538 {
539 	if (bo == NULL)
540 		return;
541 
542 	if (amdgpu_bo_reserve(bo, true) == 0) {
543 		amdgpu_bo_unpin(bo);
544 		amdgpu_bo_unreserve(bo);
545 	}
546 	amdgpu_bo_unref(&bo);
547 }
548 
549 /* Validate bo size is bit bigger than the request domain */
amdgpu_bo_validate_size(struct amdgpu_device * adev,unsigned long size,u32 domain)550 static bool amdgpu_bo_validate_size(struct amdgpu_device *adev,
551 					  unsigned long size, u32 domain)
552 {
553 	struct ttm_resource_manager *man = NULL;
554 
555 	/*
556 	 * If GTT is part of requested domains the check must succeed to
557 	 * allow fall back to GTT.
558 	 */
559 	if (domain & AMDGPU_GEM_DOMAIN_GTT)
560 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_TT);
561 	else if (domain & AMDGPU_GEM_DOMAIN_VRAM)
562 		man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
563 	else
564 		return true;
565 
566 	if (!man) {
567 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
568 			WARN_ON_ONCE("GTT domain requested but GTT mem manager uninitialized");
569 		return false;
570 	}
571 
572 	/* TODO add more domains checks, such as AMDGPU_GEM_DOMAIN_CPU, _DOMAIN_DOORBELL */
573 	if (size < man->size)
574 		return true;
575 
576 	DRM_DEBUG("BO size %lu > total memory in domain: %llu\n", size, man->size);
577 	return false;
578 }
579 
amdgpu_bo_support_uswc(u64 bo_flags)580 bool amdgpu_bo_support_uswc(u64 bo_flags)
581 {
582 
583 #ifdef CONFIG_X86_32
584 	/* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit
585 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=84627
586 	 */
587 	return false;
588 #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT)
589 	/* Don't try to enable write-combining when it can't work, or things
590 	 * may be slow
591 	 * See https://bugs.freedesktop.org/show_bug.cgi?id=88758
592 	 */
593 
594 #ifndef CONFIG_COMPILE_TEST
595 #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \
596 	 thanks to write-combining
597 #endif
598 
599 	if (bo_flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
600 		DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for "
601 			      "better performance thanks to write-combining\n");
602 	return false;
603 #else
604 	/* For architectures that don't support WC memory,
605 	 * mask out the WC flag from the BO
606 	 */
607 	if (!drm_arch_can_wc_memory())
608 		return false;
609 
610 	return true;
611 #endif
612 }
613 
614 /**
615  * amdgpu_bo_create - create an &amdgpu_bo buffer object
616  * @adev: amdgpu device object
617  * @bp: parameters to be used for the buffer object
618  * @bo_ptr: pointer to the buffer object pointer
619  *
620  * Creates an &amdgpu_bo buffer object.
621  *
622  * Returns:
623  * 0 for success or a negative error code on failure.
624  */
amdgpu_bo_create(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo ** bo_ptr)625 int amdgpu_bo_create(struct amdgpu_device *adev,
626 			       struct amdgpu_bo_param *bp,
627 			       struct amdgpu_bo **bo_ptr)
628 {
629 	struct ttm_operation_ctx ctx = {
630 		.interruptible = (bp->type != ttm_bo_type_kernel),
631 		.no_wait_gpu = bp->no_wait_gpu,
632 		/* We opt to avoid OOM on system pages allocations */
633 		.gfp_retry_mayfail = true,
634 		.allow_res_evict = bp->type != ttm_bo_type_kernel,
635 		.resv = bp->resv
636 	};
637 	struct amdgpu_bo *bo;
638 	unsigned long page_align, size = bp->size;
639 	int r;
640 
641 	/* Note that GDS/GWS/OA allocates 1 page per byte/resource. */
642 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA)) {
643 		/* GWS and OA don't need any alignment. */
644 		page_align = bp->byte_align;
645 		size <<= PAGE_SHIFT;
646 
647 	} else if (bp->domain & AMDGPU_GEM_DOMAIN_GDS) {
648 		/* Both size and alignment must be a multiple of 4. */
649 		page_align = ALIGN(bp->byte_align, 4);
650 		size = ALIGN(size, 4) << PAGE_SHIFT;
651 	} else {
652 		/* Memory should be aligned at least to a page size. */
653 		page_align = ALIGN(bp->byte_align, PAGE_SIZE) >> PAGE_SHIFT;
654 		size = ALIGN(size, PAGE_SIZE);
655 	}
656 
657 	if (!amdgpu_bo_validate_size(adev, size, bp->domain))
658 		return -ENOMEM;
659 
660 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo));
661 
662 	*bo_ptr = NULL;
663 	bo = kvzalloc(bp->bo_ptr_size, GFP_KERNEL);
664 	if (bo == NULL)
665 		return -ENOMEM;
666 	drm_gem_private_object_init(adev_to_drm(adev), &bo->tbo.base, size);
667 	bo->tbo.base.funcs = &amdgpu_gem_object_funcs;
668 	bo->vm_bo = NULL;
669 	bo->preferred_domains = bp->preferred_domain ? bp->preferred_domain :
670 		bp->domain;
671 	bo->allowed_domains = bo->preferred_domains;
672 	if (bp->type != ttm_bo_type_kernel &&
673 	    !(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE) &&
674 	    bo->allowed_domains == AMDGPU_GEM_DOMAIN_VRAM)
675 		bo->allowed_domains |= AMDGPU_GEM_DOMAIN_GTT;
676 
677 	bo->flags = bp->flags;
678 
679 	if (adev->gmc.mem_partitions)
680 		/* For GPUs with spatial partitioning, bo->xcp_id=-1 means any partition */
681 		bo->xcp_id = bp->xcp_id_plus1 - 1;
682 	else
683 		/* For GPUs without spatial partitioning */
684 		bo->xcp_id = 0;
685 
686 	if (!amdgpu_bo_support_uswc(bo->flags))
687 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_GTT_USWC;
688 
689 	bo->tbo.bdev = &adev->mman.bdev;
690 	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
691 			  AMDGPU_GEM_DOMAIN_GDS))
692 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
693 	else
694 		amdgpu_bo_placement_from_domain(bo, bp->domain);
695 	if (bp->type == ttm_bo_type_kernel)
696 		bo->tbo.priority = 2;
697 	else if (!(bp->flags & AMDGPU_GEM_CREATE_DISCARDABLE))
698 		bo->tbo.priority = 1;
699 
700 	if (!bp->destroy)
701 		bp->destroy = &amdgpu_bo_destroy;
702 
703 	r = ttm_bo_init_reserved(&adev->mman.bdev, &bo->tbo, bp->type,
704 				 &bo->placement, page_align, &ctx,  NULL,
705 				 bp->resv, bp->destroy);
706 	if (unlikely(r != 0))
707 		return r;
708 
709 	if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
710 	    amdgpu_res_cpu_visible(adev, bo->tbo.resource))
711 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved,
712 					     ctx.bytes_moved);
713 	else
714 		amdgpu_cs_report_moved_bytes(adev, ctx.bytes_moved, 0);
715 
716 	if (bp->flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
717 	    bo->tbo.resource->mem_type == TTM_PL_VRAM) {
718 		struct dma_fence *fence;
719 
720 		r = amdgpu_ttm_clear_buffer(bo, bo->tbo.base.resv, &fence);
721 		if (unlikely(r))
722 			goto fail_unreserve;
723 
724 		dma_resv_add_fence(bo->tbo.base.resv, fence,
725 				   DMA_RESV_USAGE_KERNEL);
726 		dma_fence_put(fence);
727 	}
728 	if (!bp->resv)
729 		amdgpu_bo_unreserve(bo);
730 	*bo_ptr = bo;
731 
732 	trace_amdgpu_bo_create(bo);
733 
734 	/* Treat CPU_ACCESS_REQUIRED only as a hint if given by UMD */
735 	if (bp->type == ttm_bo_type_device)
736 		bo->flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
737 
738 	return 0;
739 
740 fail_unreserve:
741 	if (!bp->resv)
742 		dma_resv_unlock(bo->tbo.base.resv);
743 	amdgpu_bo_unref(&bo);
744 	return r;
745 }
746 
747 /**
748  * amdgpu_bo_create_user - create an &amdgpu_bo_user buffer object
749  * @adev: amdgpu device object
750  * @bp: parameters to be used for the buffer object
751  * @ubo_ptr: pointer to the buffer object pointer
752  *
753  * Create a BO to be used by user application;
754  *
755  * Returns:
756  * 0 for success or a negative error code on failure.
757  */
758 
amdgpu_bo_create_user(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_user ** ubo_ptr)759 int amdgpu_bo_create_user(struct amdgpu_device *adev,
760 			  struct amdgpu_bo_param *bp,
761 			  struct amdgpu_bo_user **ubo_ptr)
762 {
763 	struct amdgpu_bo *bo_ptr;
764 	int r;
765 
766 	bp->bo_ptr_size = sizeof(struct amdgpu_bo_user);
767 	bp->destroy = &amdgpu_bo_user_destroy;
768 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
769 	if (r)
770 		return r;
771 
772 	*ubo_ptr = to_amdgpu_bo_user(bo_ptr);
773 	return r;
774 }
775 
776 /**
777  * amdgpu_bo_create_vm - create an &amdgpu_bo_vm buffer object
778  * @adev: amdgpu device object
779  * @bp: parameters to be used for the buffer object
780  * @vmbo_ptr: pointer to the buffer object pointer
781  *
782  * Create a BO to be for GPUVM.
783  *
784  * Returns:
785  * 0 for success or a negative error code on failure.
786  */
787 
amdgpu_bo_create_vm(struct amdgpu_device * adev,struct amdgpu_bo_param * bp,struct amdgpu_bo_vm ** vmbo_ptr)788 int amdgpu_bo_create_vm(struct amdgpu_device *adev,
789 			struct amdgpu_bo_param *bp,
790 			struct amdgpu_bo_vm **vmbo_ptr)
791 {
792 	struct amdgpu_bo *bo_ptr;
793 	int r;
794 
795 	/* bo_ptr_size will be determined by the caller and it depends on
796 	 * num of amdgpu_vm_pt entries.
797 	 */
798 	BUG_ON(bp->bo_ptr_size < sizeof(struct amdgpu_bo_vm));
799 	r = amdgpu_bo_create(adev, bp, &bo_ptr);
800 	if (r)
801 		return r;
802 
803 	*vmbo_ptr = to_amdgpu_bo_vm(bo_ptr);
804 	return r;
805 }
806 
807 /**
808  * amdgpu_bo_kmap - map an &amdgpu_bo buffer object
809  * @bo: &amdgpu_bo buffer object to be mapped
810  * @ptr: kernel virtual address to be returned
811  *
812  * Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls
813  * amdgpu_bo_kptr() to get the kernel virtual address.
814  *
815  * Returns:
816  * 0 for success or a negative error code on failure.
817  */
amdgpu_bo_kmap(struct amdgpu_bo * bo,void ** ptr)818 int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)
819 {
820 	void *kptr;
821 	long r;
822 
823 	if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
824 		return -EPERM;
825 
826 	r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_KERNEL,
827 				  false, MAX_SCHEDULE_TIMEOUT);
828 	if (r < 0)
829 		return r;
830 
831 	kptr = amdgpu_bo_kptr(bo);
832 	if (kptr) {
833 		if (ptr)
834 			*ptr = kptr;
835 		return 0;
836 	}
837 
838 	r = ttm_bo_kmap(&bo->tbo, 0, PFN_UP(bo->tbo.base.size), &bo->kmap);
839 	if (r)
840 		return r;
841 
842 	if (ptr)
843 		*ptr = amdgpu_bo_kptr(bo);
844 
845 	return 0;
846 }
847 
848 /**
849  * amdgpu_bo_kptr - returns a kernel virtual address of the buffer object
850  * @bo: &amdgpu_bo buffer object
851  *
852  * Calls ttm_kmap_obj_virtual() to get the kernel virtual address
853  *
854  * Returns:
855  * the virtual address of a buffer object area.
856  */
amdgpu_bo_kptr(struct amdgpu_bo * bo)857 void *amdgpu_bo_kptr(struct amdgpu_bo *bo)
858 {
859 	bool is_iomem;
860 
861 	return ttm_kmap_obj_virtual(&bo->kmap, &is_iomem);
862 }
863 
864 /**
865  * amdgpu_bo_kunmap - unmap an &amdgpu_bo buffer object
866  * @bo: &amdgpu_bo buffer object to be unmapped
867  *
868  * Unmaps a kernel map set up by amdgpu_bo_kmap().
869  */
amdgpu_bo_kunmap(struct amdgpu_bo * bo)870 void amdgpu_bo_kunmap(struct amdgpu_bo *bo)
871 {
872 	if (bo->kmap.bo)
873 		ttm_bo_kunmap(&bo->kmap);
874 }
875 
876 /**
877  * amdgpu_bo_ref - reference an &amdgpu_bo buffer object
878  * @bo: &amdgpu_bo buffer object
879  *
880  * References the contained &ttm_buffer_object.
881  *
882  * Returns:
883  * a refcounted pointer to the &amdgpu_bo buffer object.
884  */
amdgpu_bo_ref(struct amdgpu_bo * bo)885 struct amdgpu_bo *amdgpu_bo_ref(struct amdgpu_bo *bo)
886 {
887 	if (bo == NULL)
888 		return NULL;
889 
890 	drm_gem_object_get(&bo->tbo.base);
891 	return bo;
892 }
893 
894 /**
895  * amdgpu_bo_unref - unreference an &amdgpu_bo buffer object
896  * @bo: &amdgpu_bo buffer object
897  *
898  * Unreferences the contained &ttm_buffer_object and clear the pointer
899  */
amdgpu_bo_unref(struct amdgpu_bo ** bo)900 void amdgpu_bo_unref(struct amdgpu_bo **bo)
901 {
902 	if ((*bo) == NULL)
903 		return;
904 
905 	drm_gem_object_put(&(*bo)->tbo.base);
906 	*bo = NULL;
907 }
908 
909 /**
910  * amdgpu_bo_pin - pin an &amdgpu_bo buffer object
911  * @bo: &amdgpu_bo buffer object to be pinned
912  * @domain: domain to be pinned to
913  *
914  * Pins the buffer object according to requested domain. If the memory is
915  * unbound gart memory, binds the pages into gart table. Adjusts pin_count and
916  * pin_size accordingly.
917  *
918  * Pinning means to lock pages in memory along with keeping them at a fixed
919  * offset. It is required when a buffer can not be moved, for example, when
920  * a display buffer is being scanned out.
921  *
922  * Returns:
923  * 0 for success or a negative error code on failure.
924  */
amdgpu_bo_pin(struct amdgpu_bo * bo,u32 domain)925 int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)
926 {
927 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
928 	struct ttm_operation_ctx ctx = { false, false };
929 	int r, i;
930 
931 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm))
932 		return -EPERM;
933 
934 	/* Check domain to be pinned to against preferred domains */
935 	if (bo->preferred_domains & domain)
936 		domain = bo->preferred_domains & domain;
937 
938 	/* A shared bo cannot be migrated to VRAM */
939 	if (drm_gem_is_imported(&bo->tbo.base)) {
940 		if (domain & AMDGPU_GEM_DOMAIN_GTT)
941 			domain = AMDGPU_GEM_DOMAIN_GTT;
942 		else
943 			return -EINVAL;
944 	}
945 
946 	if (bo->tbo.pin_count) {
947 		uint32_t mem_type = bo->tbo.resource->mem_type;
948 		uint32_t mem_flags = bo->tbo.resource->placement;
949 
950 		if (!(domain & amdgpu_mem_type_to_domain(mem_type)))
951 			return -EINVAL;
952 
953 		if ((mem_type == TTM_PL_VRAM) &&
954 		    (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS) &&
955 		    !(mem_flags & TTM_PL_FLAG_CONTIGUOUS))
956 			return -EINVAL;
957 
958 		ttm_bo_pin(&bo->tbo);
959 		return 0;
960 	}
961 
962 	/* This assumes only APU display buffers are pinned with (VRAM|GTT).
963 	 * See function amdgpu_display_supported_domains()
964 	 */
965 	domain = amdgpu_bo_get_preferred_domain(adev, domain);
966 
967 	if (drm_gem_is_imported(&bo->tbo.base))
968 		dma_buf_pin(bo->tbo.base.import_attach);
969 
970 	/* force to pin into visible video ram */
971 	if (!(bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS))
972 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
973 	amdgpu_bo_placement_from_domain(bo, domain);
974 	for (i = 0; i < bo->placement.num_placement; i++) {
975 		if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS &&
976 		    bo->placements[i].mem_type == TTM_PL_VRAM)
977 			bo->placements[i].flags |= TTM_PL_FLAG_CONTIGUOUS;
978 	}
979 
980 	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
981 	if (unlikely(r)) {
982 		dev_err(adev->dev, "%p pin failed\n", bo);
983 		goto error;
984 	}
985 
986 	ttm_bo_pin(&bo->tbo);
987 
988 	if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
989 		atomic64_add(amdgpu_bo_size(bo), &adev->vram_pin_size);
990 		atomic64_add(amdgpu_vram_mgr_bo_visible_size(bo),
991 			     &adev->visible_pin_size);
992 	} else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
993 		atomic64_add(amdgpu_bo_size(bo), &adev->gart_pin_size);
994 	}
995 
996 error:
997 	return r;
998 }
999 
1000 /**
1001  * amdgpu_bo_unpin - unpin an &amdgpu_bo buffer object
1002  * @bo: &amdgpu_bo buffer object to be unpinned
1003  *
1004  * Decreases the pin_count, and clears the flags if pin_count reaches 0.
1005  * Changes placement and pin size accordingly.
1006  *
1007  * Returns:
1008  * 0 for success or a negative error code on failure.
1009  */
amdgpu_bo_unpin(struct amdgpu_bo * bo)1010 void amdgpu_bo_unpin(struct amdgpu_bo *bo)
1011 {
1012 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1013 
1014 	ttm_bo_unpin(&bo->tbo);
1015 	if (bo->tbo.pin_count)
1016 		return;
1017 
1018 	if (drm_gem_is_imported(&bo->tbo.base))
1019 		dma_buf_unpin(bo->tbo.base.import_attach);
1020 
1021 	if (bo->tbo.resource->mem_type == TTM_PL_VRAM) {
1022 		atomic64_sub(amdgpu_bo_size(bo), &adev->vram_pin_size);
1023 		atomic64_sub(amdgpu_vram_mgr_bo_visible_size(bo),
1024 			     &adev->visible_pin_size);
1025 	} else if (bo->tbo.resource->mem_type == TTM_PL_TT) {
1026 		atomic64_sub(amdgpu_bo_size(bo), &adev->gart_pin_size);
1027 	}
1028 
1029 }
1030 
1031 static const char * const amdgpu_vram_names[] = {
1032 	"UNKNOWN",
1033 	"GDDR1",
1034 	"DDR2",
1035 	"GDDR3",
1036 	"GDDR4",
1037 	"GDDR5",
1038 	"HBM",
1039 	"DDR3",
1040 	"DDR4",
1041 	"GDDR6",
1042 	"DDR5",
1043 	"LPDDR4",
1044 	"LPDDR5",
1045 	"HBM3E"
1046 };
1047 
1048 /**
1049  * amdgpu_bo_init - initialize memory manager
1050  * @adev: amdgpu device object
1051  *
1052  * Calls amdgpu_ttm_init() to initialize amdgpu memory manager.
1053  *
1054  * Returns:
1055  * 0 for success or a negative error code on failure.
1056  */
amdgpu_bo_init(struct amdgpu_device * adev)1057 int amdgpu_bo_init(struct amdgpu_device *adev)
1058 {
1059 	/* On A+A platform, VRAM can be mapped as WB */
1060 	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1061 		/* reserve PAT memory space to WC for VRAM */
1062 		int r = arch_io_reserve_memtype_wc(adev->gmc.aper_base,
1063 				adev->gmc.aper_size);
1064 
1065 		if (r) {
1066 			DRM_ERROR("Unable to set WC memtype for the aperture base\n");
1067 			return r;
1068 		}
1069 
1070 		/* Add an MTRR for the VRAM */
1071 		adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
1072 				adev->gmc.aper_size);
1073 	}
1074 
1075 	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
1076 		 adev->gmc.mc_vram_size >> 20,
1077 		 (unsigned long long)adev->gmc.aper_size >> 20);
1078 	DRM_INFO("RAM width %dbits %s\n",
1079 		 adev->gmc.vram_width, amdgpu_vram_names[adev->gmc.vram_type]);
1080 	return amdgpu_ttm_init(adev);
1081 }
1082 
1083 /**
1084  * amdgpu_bo_fini - tear down memory manager
1085  * @adev: amdgpu device object
1086  *
1087  * Reverses amdgpu_bo_init() to tear down memory manager.
1088  */
amdgpu_bo_fini(struct amdgpu_device * adev)1089 void amdgpu_bo_fini(struct amdgpu_device *adev)
1090 {
1091 	int idx;
1092 
1093 	amdgpu_ttm_fini(adev);
1094 
1095 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
1096 		if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
1097 			arch_phys_wc_del(adev->gmc.vram_mtrr);
1098 			arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
1099 		}
1100 		drm_dev_exit(idx);
1101 	}
1102 }
1103 
1104 /**
1105  * amdgpu_bo_set_tiling_flags - set tiling flags
1106  * @bo: &amdgpu_bo buffer object
1107  * @tiling_flags: new flags
1108  *
1109  * Sets buffer object's tiling flags with the new one. Used by GEM ioctl or
1110  * kernel driver to set the tiling flags on a buffer.
1111  *
1112  * Returns:
1113  * 0 for success or a negative error code on failure.
1114  */
amdgpu_bo_set_tiling_flags(struct amdgpu_bo * bo,u64 tiling_flags)1115 int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)
1116 {
1117 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1118 	struct amdgpu_bo_user *ubo;
1119 
1120 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1121 	if (adev->family <= AMDGPU_FAMILY_CZ &&
1122 	    AMDGPU_TILING_GET(tiling_flags, TILE_SPLIT) > 6)
1123 		return -EINVAL;
1124 
1125 	ubo = to_amdgpu_bo_user(bo);
1126 	ubo->tiling_flags = tiling_flags;
1127 	return 0;
1128 }
1129 
1130 /**
1131  * amdgpu_bo_get_tiling_flags - get tiling flags
1132  * @bo: &amdgpu_bo buffer object
1133  * @tiling_flags: returned flags
1134  *
1135  * Gets buffer object's tiling flags. Used by GEM ioctl or kernel driver to
1136  * set the tiling flags on a buffer.
1137  */
amdgpu_bo_get_tiling_flags(struct amdgpu_bo * bo,u64 * tiling_flags)1138 void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)
1139 {
1140 	struct amdgpu_bo_user *ubo;
1141 
1142 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1143 	dma_resv_assert_held(bo->tbo.base.resv);
1144 	ubo = to_amdgpu_bo_user(bo);
1145 
1146 	if (tiling_flags)
1147 		*tiling_flags = ubo->tiling_flags;
1148 }
1149 
1150 /**
1151  * amdgpu_bo_set_metadata - set metadata
1152  * @bo: &amdgpu_bo buffer object
1153  * @metadata: new metadata
1154  * @metadata_size: size of the new metadata
1155  * @flags: flags of the new metadata
1156  *
1157  * Sets buffer object's metadata, its size and flags.
1158  * Used via GEM ioctl.
1159  *
1160  * Returns:
1161  * 0 for success or a negative error code on failure.
1162  */
amdgpu_bo_set_metadata(struct amdgpu_bo * bo,void * metadata,u32 metadata_size,uint64_t flags)1163 int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata,
1164 			   u32 metadata_size, uint64_t flags)
1165 {
1166 	struct amdgpu_bo_user *ubo;
1167 	void *buffer;
1168 
1169 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1170 	ubo = to_amdgpu_bo_user(bo);
1171 	if (!metadata_size) {
1172 		if (ubo->metadata_size) {
1173 			kfree(ubo->metadata);
1174 			ubo->metadata = NULL;
1175 			ubo->metadata_size = 0;
1176 		}
1177 		return 0;
1178 	}
1179 
1180 	if (metadata == NULL)
1181 		return -EINVAL;
1182 
1183 	buffer = kmemdup(metadata, metadata_size, GFP_KERNEL);
1184 	if (buffer == NULL)
1185 		return -ENOMEM;
1186 
1187 	kfree(ubo->metadata);
1188 	ubo->metadata_flags = flags;
1189 	ubo->metadata = buffer;
1190 	ubo->metadata_size = metadata_size;
1191 
1192 	return 0;
1193 }
1194 
1195 /**
1196  * amdgpu_bo_get_metadata - get metadata
1197  * @bo: &amdgpu_bo buffer object
1198  * @buffer: returned metadata
1199  * @buffer_size: size of the buffer
1200  * @metadata_size: size of the returned metadata
1201  * @flags: flags of the returned metadata
1202  *
1203  * Gets buffer object's metadata, its size and flags. buffer_size shall not be
1204  * less than metadata_size.
1205  * Used via GEM ioctl.
1206  *
1207  * Returns:
1208  * 0 for success or a negative error code on failure.
1209  */
amdgpu_bo_get_metadata(struct amdgpu_bo * bo,void * buffer,size_t buffer_size,uint32_t * metadata_size,uint64_t * flags)1210 int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer,
1211 			   size_t buffer_size, uint32_t *metadata_size,
1212 			   uint64_t *flags)
1213 {
1214 	struct amdgpu_bo_user *ubo;
1215 
1216 	if (!buffer && !metadata_size)
1217 		return -EINVAL;
1218 
1219 	BUG_ON(bo->tbo.type == ttm_bo_type_kernel);
1220 	ubo = to_amdgpu_bo_user(bo);
1221 	if (metadata_size)
1222 		*metadata_size = ubo->metadata_size;
1223 
1224 	if (buffer) {
1225 		if (buffer_size < ubo->metadata_size)
1226 			return -EINVAL;
1227 
1228 		if (ubo->metadata_size)
1229 			memcpy(buffer, ubo->metadata, ubo->metadata_size);
1230 	}
1231 
1232 	if (flags)
1233 		*flags = ubo->metadata_flags;
1234 
1235 	return 0;
1236 }
1237 
1238 /**
1239  * amdgpu_bo_move_notify - notification about a memory move
1240  * @bo: pointer to a buffer object
1241  * @evict: if this move is evicting the buffer from the graphics address space
1242  * @new_mem: new resource for backing the BO
1243  *
1244  * Marks the corresponding &amdgpu_bo buffer object as invalid, also performs
1245  * bookkeeping.
1246  * TTM driver callback which is called when ttm moves a buffer.
1247  */
amdgpu_bo_move_notify(struct ttm_buffer_object * bo,bool evict,struct ttm_resource * new_mem)1248 void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
1249 			   bool evict,
1250 			   struct ttm_resource *new_mem)
1251 {
1252 	struct ttm_resource *old_mem = bo->resource;
1253 	struct amdgpu_bo *abo;
1254 
1255 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1256 		return;
1257 
1258 	abo = ttm_to_amdgpu_bo(bo);
1259 	amdgpu_vm_bo_move(abo, new_mem, evict);
1260 
1261 	amdgpu_bo_kunmap(abo);
1262 
1263 	if (abo->tbo.base.dma_buf && !drm_gem_is_imported(&abo->tbo.base) &&
1264 	    old_mem && old_mem->mem_type != TTM_PL_SYSTEM)
1265 		dma_buf_move_notify(abo->tbo.base.dma_buf);
1266 
1267 	/* move_notify is called before move happens */
1268 	trace_amdgpu_bo_move(abo, new_mem ? new_mem->mem_type : -1,
1269 			     old_mem ? old_mem->mem_type : -1);
1270 }
1271 
1272 /**
1273  * amdgpu_bo_release_notify - notification about a BO being released
1274  * @bo: pointer to a buffer object
1275  *
1276  * Wipes VRAM buffers whose contents should not be leaked before the
1277  * memory is released.
1278  */
amdgpu_bo_release_notify(struct ttm_buffer_object * bo)1279 void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)
1280 {
1281 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1282 	struct dma_fence *fence = NULL;
1283 	struct amdgpu_bo *abo;
1284 	int r;
1285 
1286 	if (!amdgpu_bo_is_amdgpu_bo(bo))
1287 		return;
1288 
1289 	abo = ttm_to_amdgpu_bo(bo);
1290 
1291 	WARN_ON(abo->vm_bo);
1292 
1293 	if (abo->kfd_bo)
1294 		amdgpu_amdkfd_release_notify(abo);
1295 
1296 	/*
1297 	 * We lock the private dma_resv object here and since the BO is about to
1298 	 * be released nobody else should have a pointer to it.
1299 	 * So when this locking here fails something is wrong with the reference
1300 	 * counting.
1301 	 */
1302 	if (WARN_ON_ONCE(!dma_resv_trylock(&bo->base._resv)))
1303 		return;
1304 
1305 	amdgpu_amdkfd_remove_all_eviction_fences(abo);
1306 
1307 	if (!bo->resource || bo->resource->mem_type != TTM_PL_VRAM ||
1308 	    !(abo->flags & AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE) ||
1309 	    adev->in_suspend || drm_dev_is_unplugged(adev_to_drm(adev)))
1310 		goto out;
1311 
1312 	r = dma_resv_reserve_fences(&bo->base._resv, 1);
1313 	if (r)
1314 		goto out;
1315 
1316 	r = amdgpu_fill_buffer(abo, 0, &bo->base._resv, &fence, true);
1317 	if (WARN_ON(r))
1318 		goto out;
1319 
1320 	amdgpu_vram_mgr_set_cleared(bo->resource);
1321 	dma_resv_add_fence(&bo->base._resv, fence, DMA_RESV_USAGE_KERNEL);
1322 	dma_fence_put(fence);
1323 
1324 out:
1325 	dma_resv_unlock(&bo->base._resv);
1326 }
1327 
1328 /**
1329  * amdgpu_bo_fault_reserve_notify - notification about a memory fault
1330  * @bo: pointer to a buffer object
1331  *
1332  * Notifies the driver we are taking a fault on this BO and have reserved it,
1333  * also performs bookkeeping.
1334  * TTM driver callback for dealing with vm faults.
1335  *
1336  * Returns:
1337  * 0 for success or a negative error code on failure.
1338  */
amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object * bo)1339 vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
1340 {
1341 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
1342 	struct ttm_operation_ctx ctx = { false, false };
1343 	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
1344 	int r;
1345 
1346 	/* Remember that this BO was accessed by the CPU */
1347 	abo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
1348 
1349 	if (amdgpu_res_cpu_visible(adev, bo->resource))
1350 		return 0;
1351 
1352 	/* Can't move a pinned BO to visible VRAM */
1353 	if (abo->tbo.pin_count > 0)
1354 		return VM_FAULT_SIGBUS;
1355 
1356 	/* hurrah the memory is not visible ! */
1357 	atomic64_inc(&adev->num_vram_cpu_page_faults);
1358 	amdgpu_bo_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
1359 					AMDGPU_GEM_DOMAIN_GTT);
1360 
1361 	/* Avoid costly evictions; only set GTT as a busy placement */
1362 	abo->placements[0].flags |= TTM_PL_FLAG_DESIRED;
1363 
1364 	r = ttm_bo_validate(bo, &abo->placement, &ctx);
1365 	if (unlikely(r == -EBUSY || r == -ERESTARTSYS))
1366 		return VM_FAULT_NOPAGE;
1367 	else if (unlikely(r))
1368 		return VM_FAULT_SIGBUS;
1369 
1370 	/* this should never happen */
1371 	if (bo->resource->mem_type == TTM_PL_VRAM &&
1372 	    !amdgpu_res_cpu_visible(adev, bo->resource))
1373 		return VM_FAULT_SIGBUS;
1374 
1375 	ttm_bo_move_to_lru_tail_unlocked(bo);
1376 	return 0;
1377 }
1378 
1379 /**
1380  * amdgpu_bo_fence - add fence to buffer object
1381  *
1382  * @bo: buffer object in question
1383  * @fence: fence to add
1384  * @shared: true if fence should be added shared
1385  *
1386  */
amdgpu_bo_fence(struct amdgpu_bo * bo,struct dma_fence * fence,bool shared)1387 void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence,
1388 		     bool shared)
1389 {
1390 	struct dma_resv *resv = bo->tbo.base.resv;
1391 	int r;
1392 
1393 	r = dma_resv_reserve_fences(resv, 1);
1394 	if (r) {
1395 		/* As last resort on OOM we block for the fence */
1396 		dma_fence_wait(fence, false);
1397 		return;
1398 	}
1399 
1400 	dma_resv_add_fence(resv, fence, shared ? DMA_RESV_USAGE_READ :
1401 			   DMA_RESV_USAGE_WRITE);
1402 }
1403 
1404 /**
1405  * amdgpu_bo_sync_wait_resv - Wait for BO reservation fences
1406  *
1407  * @adev: amdgpu device pointer
1408  * @resv: reservation object to sync to
1409  * @sync_mode: synchronization mode
1410  * @owner: fence owner
1411  * @intr: Whether the wait is interruptible
1412  *
1413  * Extract the fences from the reservation object and waits for them to finish.
1414  *
1415  * Returns:
1416  * 0 on success, errno otherwise.
1417  */
amdgpu_bo_sync_wait_resv(struct amdgpu_device * adev,struct dma_resv * resv,enum amdgpu_sync_mode sync_mode,void * owner,bool intr)1418 int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv,
1419 			     enum amdgpu_sync_mode sync_mode, void *owner,
1420 			     bool intr)
1421 {
1422 	struct amdgpu_sync sync;
1423 	int r;
1424 
1425 	amdgpu_sync_create(&sync);
1426 	amdgpu_sync_resv(adev, &sync, resv, sync_mode, owner);
1427 	r = amdgpu_sync_wait(&sync, intr);
1428 	amdgpu_sync_free(&sync);
1429 	return r;
1430 }
1431 
1432 /**
1433  * amdgpu_bo_sync_wait - Wrapper for amdgpu_bo_sync_wait_resv
1434  * @bo: buffer object to wait for
1435  * @owner: fence owner
1436  * @intr: Whether the wait is interruptible
1437  *
1438  * Wrapper to wait for fences in a BO.
1439  * Returns:
1440  * 0 on success, errno otherwise.
1441  */
amdgpu_bo_sync_wait(struct amdgpu_bo * bo,void * owner,bool intr)1442 int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)
1443 {
1444 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1445 
1446 	return amdgpu_bo_sync_wait_resv(adev, bo->tbo.base.resv,
1447 					AMDGPU_SYNC_NE_OWNER, owner, intr);
1448 }
1449 
1450 /**
1451  * amdgpu_bo_gpu_offset - return GPU offset of bo
1452  * @bo:	amdgpu object for which we query the offset
1453  *
1454  * Note: object should either be pinned or reserved when calling this
1455  * function, it might be useful to add check for this for debugging.
1456  *
1457  * Returns:
1458  * current GPU offset of the object.
1459  */
amdgpu_bo_gpu_offset(struct amdgpu_bo * bo)1460 u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
1461 {
1462 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_SYSTEM);
1463 	WARN_ON_ONCE(!dma_resv_is_locked(bo->tbo.base.resv) &&
1464 		     !bo->tbo.pin_count && bo->tbo.type != ttm_bo_type_kernel);
1465 	WARN_ON_ONCE(bo->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET);
1466 	WARN_ON_ONCE(bo->tbo.resource->mem_type == TTM_PL_VRAM &&
1467 		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS));
1468 
1469 	return amdgpu_bo_gpu_offset_no_check(bo);
1470 }
1471 
1472 /**
1473  * amdgpu_bo_fb_aper_addr - return FB aperture GPU offset of the VRAM bo
1474  * @bo:	amdgpu VRAM buffer object for which we query the offset
1475  *
1476  * Returns:
1477  * current FB aperture GPU offset of the object.
1478  */
amdgpu_bo_fb_aper_addr(struct amdgpu_bo * bo)1479 u64 amdgpu_bo_fb_aper_addr(struct amdgpu_bo *bo)
1480 {
1481 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1482 	uint64_t offset, fb_base;
1483 
1484 	WARN_ON_ONCE(bo->tbo.resource->mem_type != TTM_PL_VRAM);
1485 
1486 	fb_base = adev->gmc.fb_start;
1487 	fb_base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1488 	offset = (bo->tbo.resource->start << PAGE_SHIFT) + fb_base;
1489 	return amdgpu_gmc_sign_extend(offset);
1490 }
1491 
1492 /**
1493  * amdgpu_bo_gpu_offset_no_check - return GPU offset of bo
1494  * @bo:	amdgpu object for which we query the offset
1495  *
1496  * Returns:
1497  * current GPU offset of the object without raising warnings.
1498  */
amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo * bo)1499 u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
1500 {
1501 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1502 	uint64_t offset = AMDGPU_BO_INVALID_OFFSET;
1503 
1504 	if (bo->tbo.resource->mem_type == TTM_PL_TT)
1505 		offset = amdgpu_gmc_agp_addr(&bo->tbo);
1506 
1507 	if (offset == AMDGPU_BO_INVALID_OFFSET)
1508 		offset = (bo->tbo.resource->start << PAGE_SHIFT) +
1509 			amdgpu_ttm_domain_start(adev, bo->tbo.resource->mem_type);
1510 
1511 	return amdgpu_gmc_sign_extend(offset);
1512 }
1513 
1514 /**
1515  * amdgpu_bo_mem_stats_placement - bo placement for memory accounting
1516  * @bo:	the buffer object we should look at
1517  *
1518  * BO can have multiple preferred placements, to avoid double counting we want
1519  * to file it under a single placement for memory stats.
1520  * Luckily, if we take the highest set bit in preferred_domains the result is
1521  * quite sensible.
1522  *
1523  * Returns:
1524  * Which of the placements should the BO be accounted under.
1525  */
amdgpu_bo_mem_stats_placement(struct amdgpu_bo * bo)1526 uint32_t amdgpu_bo_mem_stats_placement(struct amdgpu_bo *bo)
1527 {
1528 	uint32_t domain = bo->preferred_domains & AMDGPU_GEM_DOMAIN_MASK;
1529 
1530 	if (!domain)
1531 		return TTM_PL_SYSTEM;
1532 
1533 	switch (rounddown_pow_of_two(domain)) {
1534 	case AMDGPU_GEM_DOMAIN_CPU:
1535 		return TTM_PL_SYSTEM;
1536 	case AMDGPU_GEM_DOMAIN_GTT:
1537 		return TTM_PL_TT;
1538 	case AMDGPU_GEM_DOMAIN_VRAM:
1539 		return TTM_PL_VRAM;
1540 	case AMDGPU_GEM_DOMAIN_GDS:
1541 		return AMDGPU_PL_GDS;
1542 	case AMDGPU_GEM_DOMAIN_GWS:
1543 		return AMDGPU_PL_GWS;
1544 	case AMDGPU_GEM_DOMAIN_OA:
1545 		return AMDGPU_PL_OA;
1546 	case AMDGPU_GEM_DOMAIN_DOORBELL:
1547 		return AMDGPU_PL_DOORBELL;
1548 	default:
1549 		return TTM_PL_SYSTEM;
1550 	}
1551 }
1552 
1553 /**
1554  * amdgpu_bo_get_preferred_domain - get preferred domain
1555  * @adev: amdgpu device object
1556  * @domain: allowed :ref:`memory domains <amdgpu_memory_domains>`
1557  *
1558  * Returns:
1559  * Which of the allowed domains is preferred for allocating the BO.
1560  */
amdgpu_bo_get_preferred_domain(struct amdgpu_device * adev,uint32_t domain)1561 uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
1562 					    uint32_t domain)
1563 {
1564 	if ((domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) &&
1565 	    ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type == CHIP_STONEY))) {
1566 		domain = AMDGPU_GEM_DOMAIN_VRAM;
1567 		if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
1568 			domain = AMDGPU_GEM_DOMAIN_GTT;
1569 	}
1570 	return domain;
1571 }
1572 
1573 #if defined(CONFIG_DEBUG_FS)
1574 #define amdgpu_bo_print_flag(m, bo, flag)		        \
1575 	do {							\
1576 		if (bo->flags & (AMDGPU_GEM_CREATE_ ## flag)) {	\
1577 			seq_printf((m), " " #flag);		\
1578 		}						\
1579 	} while (0)
1580 
1581 /**
1582  * amdgpu_bo_print_info - print BO info in debugfs file
1583  *
1584  * @id: Index or Id of the BO
1585  * @bo: Requested BO for printing info
1586  * @m: debugfs file
1587  *
1588  * Print BO information in debugfs file
1589  *
1590  * Returns:
1591  * Size of the BO in bytes.
1592  */
amdgpu_bo_print_info(int id,struct amdgpu_bo * bo,struct seq_file * m)1593 u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)
1594 {
1595 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
1596 	struct dma_buf_attachment *attachment;
1597 	struct dma_buf *dma_buf;
1598 	const char *placement;
1599 	unsigned int pin_count;
1600 	u64 size;
1601 
1602 	if (dma_resv_trylock(bo->tbo.base.resv)) {
1603 		if (!bo->tbo.resource) {
1604 			placement = "NONE";
1605 		} else {
1606 			switch (bo->tbo.resource->mem_type) {
1607 			case TTM_PL_VRAM:
1608 				if (amdgpu_res_cpu_visible(adev, bo->tbo.resource))
1609 					placement = "VRAM VISIBLE";
1610 				else
1611 					placement = "VRAM";
1612 				break;
1613 			case TTM_PL_TT:
1614 				placement = "GTT";
1615 				break;
1616 			case AMDGPU_PL_GDS:
1617 				placement = "GDS";
1618 				break;
1619 			case AMDGPU_PL_GWS:
1620 				placement = "GWS";
1621 				break;
1622 			case AMDGPU_PL_OA:
1623 				placement = "OA";
1624 				break;
1625 			case AMDGPU_PL_PREEMPT:
1626 				placement = "PREEMPTIBLE";
1627 				break;
1628 			case AMDGPU_PL_DOORBELL:
1629 				placement = "DOORBELL";
1630 				break;
1631 			case TTM_PL_SYSTEM:
1632 			default:
1633 				placement = "CPU";
1634 				break;
1635 			}
1636 		}
1637 		dma_resv_unlock(bo->tbo.base.resv);
1638 	} else {
1639 		placement = "UNKNOWN";
1640 	}
1641 
1642 	size = amdgpu_bo_size(bo);
1643 	seq_printf(m, "\t\t0x%08x: %12lld byte %s",
1644 			id, size, placement);
1645 
1646 	pin_count = READ_ONCE(bo->tbo.pin_count);
1647 	if (pin_count)
1648 		seq_printf(m, " pin count %d", pin_count);
1649 
1650 	dma_buf = READ_ONCE(bo->tbo.base.dma_buf);
1651 	attachment = READ_ONCE(bo->tbo.base.import_attach);
1652 
1653 	if (attachment)
1654 		seq_printf(m, " imported from ino:%lu", file_inode(dma_buf->file)->i_ino);
1655 	else if (dma_buf)
1656 		seq_printf(m, " exported as ino:%lu", file_inode(dma_buf->file)->i_ino);
1657 
1658 	amdgpu_bo_print_flag(m, bo, CPU_ACCESS_REQUIRED);
1659 	amdgpu_bo_print_flag(m, bo, NO_CPU_ACCESS);
1660 	amdgpu_bo_print_flag(m, bo, CPU_GTT_USWC);
1661 	amdgpu_bo_print_flag(m, bo, VRAM_CLEARED);
1662 	amdgpu_bo_print_flag(m, bo, VRAM_CONTIGUOUS);
1663 	amdgpu_bo_print_flag(m, bo, VM_ALWAYS_VALID);
1664 	amdgpu_bo_print_flag(m, bo, EXPLICIT_SYNC);
1665 	/* Add the gem obj resv fence dump*/
1666 	if (dma_resv_trylock(bo->tbo.base.resv)) {
1667 		dma_resv_describe(bo->tbo.base.resv, m);
1668 		dma_resv_unlock(bo->tbo.base.resv);
1669 	}
1670 	seq_puts(m, "\n");
1671 
1672 	return size;
1673 }
1674 #endif
1675