1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2009 VMware, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  * Authors: Michel Dänzer
24  */
25 
26 #include <drm/amdgpu_drm.h>
27 #include "amdgpu.h"
28 #include "amdgpu_uvd.h"
29 #include "amdgpu_vce.h"
30 
31 /* Test BO GTT->VRAM and VRAM->GTT GPU copies across the whole GTT aperture */
amdgpu_do_test_moves(struct amdgpu_device * adev)32 static void amdgpu_do_test_moves(struct amdgpu_device *adev)
33 {
34 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
35 	struct amdgpu_bo *vram_obj = NULL;
36 	struct amdgpu_bo **gtt_obj = NULL;
37 	struct amdgpu_bo_param bp;
38 	uint64_t gart_addr, vram_addr;
39 	unsigned n, size;
40 	int i, r;
41 
42 	size = 1024 * 1024;
43 
44 	/* Number of tests =
45 	 * (Total GTT - gart_pin_size - (2 transfer windows for buffer moves)) / test size
46 	 */
47 	n = adev->gmc.gart_size - atomic64_read(&adev->gart_pin_size);
48 	n -= AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS *
49 		AMDGPU_GPU_PAGE_SIZE;
50 	n /= size;
51 
52 	gtt_obj = kcalloc(n, sizeof(*gtt_obj), GFP_KERNEL);
53 	if (!gtt_obj) {
54 		DRM_ERROR("Failed to allocate %d pointers\n", n);
55 		r = 1;
56 		goto out_cleanup;
57 	}
58 	memset(&bp, 0, sizeof(bp));
59 	bp.size = size;
60 	bp.byte_align = PAGE_SIZE;
61 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
62 	bp.flags = 0;
63 	bp.type = ttm_bo_type_kernel;
64 	bp.resv = NULL;
65 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
66 
67 	r = amdgpu_bo_create(adev, &bp, &vram_obj);
68 	if (r) {
69 		DRM_ERROR("Failed to create VRAM object\n");
70 		goto out_cleanup;
71 	}
72 	r = amdgpu_bo_reserve(vram_obj, false);
73 	if (unlikely(r != 0))
74 		goto out_unref;
75 	r = amdgpu_bo_pin(vram_obj, AMDGPU_GEM_DOMAIN_VRAM);
76 	if (r) {
77 		DRM_ERROR("Failed to pin VRAM object\n");
78 		goto out_unres;
79 	}
80 	vram_addr = amdgpu_bo_gpu_offset(vram_obj);
81 	for (i = 0; i < n; i++) {
82 		void *gtt_map, *vram_map;
83 		void **gart_start, **gart_end;
84 		void **vram_start, **vram_end;
85 		struct dma_fence *fence = NULL;
86 
87 		bp.domain = AMDGPU_GEM_DOMAIN_GTT;
88 		r = amdgpu_bo_create(adev, &bp, gtt_obj + i);
89 		if (r) {
90 			DRM_ERROR("Failed to create GTT object %d\n", i);
91 			goto out_lclean;
92 		}
93 
94 		r = amdgpu_bo_reserve(gtt_obj[i], false);
95 		if (unlikely(r != 0))
96 			goto out_lclean_unref;
97 		r = amdgpu_bo_pin(gtt_obj[i], AMDGPU_GEM_DOMAIN_GTT);
98 		if (r) {
99 			DRM_ERROR("Failed to pin GTT object %d\n", i);
100 			goto out_lclean_unres;
101 		}
102 		r = amdgpu_ttm_alloc_gart(&gtt_obj[i]->tbo);
103 		if (r) {
104 			DRM_ERROR("%p bind failed\n", gtt_obj[i]);
105 			goto out_lclean_unpin;
106 		}
107 		gart_addr = amdgpu_bo_gpu_offset(gtt_obj[i]);
108 
109 		r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
110 		if (r) {
111 			DRM_ERROR("Failed to map GTT object %d\n", i);
112 			goto out_lclean_unpin;
113 		}
114 
115 		for (gart_start = gtt_map, gart_end = gtt_map + size;
116 		     gart_start < gart_end;
117 		     gart_start++)
118 			*gart_start = gart_start;
119 
120 		amdgpu_bo_kunmap(gtt_obj[i]);
121 
122 		r = amdgpu_copy_buffer(ring, gart_addr, vram_addr,
123 				       size, NULL, &fence, false, false, false);
124 
125 		if (r) {
126 			DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
127 			goto out_lclean_unpin;
128 		}
129 
130 		r = dma_fence_wait(fence, false);
131 		if (r) {
132 			DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
133 			goto out_lclean_unpin;
134 		}
135 
136 		dma_fence_put(fence);
137 		fence = NULL;
138 
139 		r = amdgpu_bo_kmap(vram_obj, &vram_map);
140 		if (r) {
141 			DRM_ERROR("Failed to map VRAM object after copy %d\n", i);
142 			goto out_lclean_unpin;
143 		}
144 
145 		for (gart_start = gtt_map, gart_end = gtt_map + size,
146 		     vram_start = vram_map, vram_end = vram_map + size;
147 		     vram_start < vram_end;
148 		     gart_start++, vram_start++) {
149 			if (*vram_start != gart_start) {
150 				DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, "
151 					  "expected 0x%p (GTT/VRAM offset "
152 					  "0x%16llx/0x%16llx)\n",
153 					  i, *vram_start, gart_start,
154 					  (unsigned long long)
155 					  (gart_addr - adev->gmc.gart_start +
156 					   (void *)gart_start - gtt_map),
157 					  (unsigned long long)
158 					  (vram_addr - adev->gmc.vram_start +
159 					   (void *)gart_start - gtt_map));
160 				amdgpu_bo_kunmap(vram_obj);
161 				goto out_lclean_unpin;
162 			}
163 			*vram_start = vram_start;
164 		}
165 
166 		amdgpu_bo_kunmap(vram_obj);
167 
168 		r = amdgpu_copy_buffer(ring, vram_addr, gart_addr,
169 				       size, NULL, &fence, false, false, false);
170 
171 		if (r) {
172 			DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
173 			goto out_lclean_unpin;
174 		}
175 
176 		r = dma_fence_wait(fence, false);
177 		if (r) {
178 			DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
179 			goto out_lclean_unpin;
180 		}
181 
182 		dma_fence_put(fence);
183 		fence = NULL;
184 
185 		r = amdgpu_bo_kmap(gtt_obj[i], &gtt_map);
186 		if (r) {
187 			DRM_ERROR("Failed to map GTT object after copy %d\n", i);
188 			goto out_lclean_unpin;
189 		}
190 
191 		for (gart_start = gtt_map, gart_end = gtt_map + size,
192 		     vram_start = vram_map, vram_end = vram_map + size;
193 		     gart_start < gart_end;
194 		     gart_start++, vram_start++) {
195 			if (*gart_start != vram_start) {
196 				DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, "
197 					  "expected 0x%p (VRAM/GTT offset "
198 					  "0x%16llx/0x%16llx)\n",
199 					  i, *gart_start, vram_start,
200 					  (unsigned long long)
201 					  (vram_addr - adev->gmc.vram_start +
202 					   (void *)vram_start - vram_map),
203 					  (unsigned long long)
204 					  (gart_addr - adev->gmc.gart_start +
205 					   (void *)vram_start - vram_map));
206 				amdgpu_bo_kunmap(gtt_obj[i]);
207 				goto out_lclean_unpin;
208 			}
209 		}
210 
211 		amdgpu_bo_kunmap(gtt_obj[i]);
212 
213 		DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%llx\n",
214 			 gart_addr - adev->gmc.gart_start);
215 		continue;
216 
217 out_lclean_unpin:
218 		amdgpu_bo_unpin(gtt_obj[i]);
219 out_lclean_unres:
220 		amdgpu_bo_unreserve(gtt_obj[i]);
221 out_lclean_unref:
222 		amdgpu_bo_unref(&gtt_obj[i]);
223 out_lclean:
224 		for (--i; i >= 0; --i) {
225 			amdgpu_bo_unpin(gtt_obj[i]);
226 			amdgpu_bo_unreserve(gtt_obj[i]);
227 			amdgpu_bo_unref(&gtt_obj[i]);
228 		}
229 		if (fence)
230 			dma_fence_put(fence);
231 		break;
232 	}
233 
234 	amdgpu_bo_unpin(vram_obj);
235 out_unres:
236 	amdgpu_bo_unreserve(vram_obj);
237 out_unref:
238 	amdgpu_bo_unref(&vram_obj);
239 out_cleanup:
240 	kfree(gtt_obj);
241 	if (r) {
242 		pr_warn("Error while testing BO move\n");
243 	}
244 }
245 
amdgpu_test_moves(struct amdgpu_device * adev)246 void amdgpu_test_moves(struct amdgpu_device *adev)
247 {
248 	if (adev->mman.buffer_funcs)
249 		amdgpu_do_test_moves(adev);
250 }
251