1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24#include <drm/drmP.h>
25#include <drm/amdgpu_drm.h>
26#include "amdgpu.h"
27#include "amdgpu_uvd.h"
28#include "amdgpu_vce.h"
29
30
31static void amdgpu_do_test_moves(struct amdgpu_device *adev)
32{
33 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
34 struct amdgpu_bo *vram_obj = NULL;
35 struct amdgpu_bo **gtt_obj = NULL;
36 uint64_t gart_addr, vram_addr;
37 unsigned n, size;
38 int i, r;
39
40 size = 1024 * 1024;
41
42
43
44
45 n = adev->mc.gart_size - AMDGPU_IB_POOL_SIZE*64*1024;
46 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
47 if (adev->rings[i])
48 n -= adev->rings[i]->ring_size;
49 if (adev->wb.wb_obj)
50 n -= AMDGPU_GPU_PAGE_SIZE;
51 if (adev->irq.ih.ring_obj)
52 n -= adev->irq.ih.ring_size;
53 n /= size;
54
55 gtt_obj = kzalloc(n * sizeof(*gtt_obj), GFP_KERNEL);
56 if (!gtt_obj) {
57 DRM_ERROR("Failed to allocate %d pointers\n", n);
58 r = 1;
59 goto out_cleanup;
60 }
61
62 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
63 AMDGPU_GEM_DOMAIN_VRAM, 0,
64 NULL, NULL, 0, &vram_obj);
65 if (r) {
66 DRM_ERROR("Failed to create VRAM object\n");
67 goto out_cleanup;
68 }
69 r = amdgpu_bo_reserve(vram_obj, false);
70 if (unlikely(r != 0))
71 goto out_unref;
72 r = amdgpu_bo_pin(vram_obj, AMDGPU_GEM_DOMAIN_VRAM, &vram_addr);
73 if (r) {
74 DRM_ERROR("Failed to pin VRAM object\n");
75 goto out_unres;
76 }
77 for (i = 0; i < n; i++) {
78 void *gtt_map, *vram_map;
79 void **gart_start, **gart_end;
80 void **vram_start, **vram_end;
81 struct dma_fence *fence = NULL;
82
83 r = amdgpu_bo_create(adev, size, PAGE_SIZE, true,
84 AMDGPU_GEM_DOMAIN_GTT, 0, NULL,
85 NULL, 0, gtt_obj + i);
86 if (r) {
87 DRM_ERROR("Failed to create GTT object %d\n", i);
88 goto out_lclean;
89 }
90
91 r = amdgpu_bo_reserve(gtt_obj[i], false);
92 if (unlikely(r != 0))
93 goto out_lclean_unref;
94 r = amdgpu_bo_pin(gtt_obj[i], AMDGPU_GEM_DOMAIN_GTT, &gart_addr);
95 if (r) {
96 DRM_ERROR("Failed to pin GTT object %d\n", i);
97 goto out_lclean_unres;
98 }
99
100 r = amdgpu_bo_kmap(gtt_obj[i], >t_map);
101 if (r) {
102 DRM_ERROR("Failed to map GTT object %d\n", i);
103 goto out_lclean_unpin;
104 }
105
106 for (gart_start = gtt_map, gart_end = gtt_map + size;
107 gart_start < gart_end;
108 gart_start++)
109 *gart_start = gart_start;
110
111 amdgpu_bo_kunmap(gtt_obj[i]);
112
113 r = amdgpu_copy_buffer(ring, gart_addr, vram_addr,
114 size, NULL, &fence, false, false);
115
116 if (r) {
117 DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
118 goto out_lclean_unpin;
119 }
120
121 r = dma_fence_wait(fence, false);
122 if (r) {
123 DRM_ERROR("Failed to wait for GTT->VRAM fence %d\n", i);
124 goto out_lclean_unpin;
125 }
126
127 dma_fence_put(fence);
128
129 r = amdgpu_bo_kmap(vram_obj, &vram_map);
130 if (r) {
131 DRM_ERROR("Failed to map VRAM object after copy %d\n", i);
132 goto out_lclean_unpin;
133 }
134
135 for (gart_start = gtt_map, gart_end = gtt_map + size,
136 vram_start = vram_map, vram_end = vram_map + size;
137 vram_start < vram_end;
138 gart_start++, vram_start++) {
139 if (*vram_start != gart_start) {
140 DRM_ERROR("Incorrect GTT->VRAM copy %d: Got 0x%p, "
141 "expected 0x%p (GTT/VRAM offset "
142 "0x%16llx/0x%16llx)\n",
143 i, *vram_start, gart_start,
144 (unsigned long long)
145 (gart_addr - adev->mc.gart_start +
146 (void*)gart_start - gtt_map),
147 (unsigned long long)
148 (vram_addr - adev->mc.vram_start +
149 (void*)gart_start - gtt_map));
150 amdgpu_bo_kunmap(vram_obj);
151 goto out_lclean_unpin;
152 }
153 *vram_start = vram_start;
154 }
155
156 amdgpu_bo_kunmap(vram_obj);
157
158 r = amdgpu_copy_buffer(ring, vram_addr, gart_addr,
159 size, NULL, &fence, false, false);
160
161 if (r) {
162 DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
163 goto out_lclean_unpin;
164 }
165
166 r = dma_fence_wait(fence, false);
167 if (r) {
168 DRM_ERROR("Failed to wait for VRAM->GTT fence %d\n", i);
169 goto out_lclean_unpin;
170 }
171
172 dma_fence_put(fence);
173
174 r = amdgpu_bo_kmap(gtt_obj[i], >t_map);
175 if (r) {
176 DRM_ERROR("Failed to map GTT object after copy %d\n", i);
177 goto out_lclean_unpin;
178 }
179
180 for (gart_start = gtt_map, gart_end = gtt_map + size,
181 vram_start = vram_map, vram_end = vram_map + size;
182 gart_start < gart_end;
183 gart_start++, vram_start++) {
184 if (*gart_start != vram_start) {
185 DRM_ERROR("Incorrect VRAM->GTT copy %d: Got 0x%p, "
186 "expected 0x%p (VRAM/GTT offset "
187 "0x%16llx/0x%16llx)\n",
188 i, *gart_start, vram_start,
189 (unsigned long long)
190 (vram_addr - adev->mc.vram_start +
191 (void*)vram_start - vram_map),
192 (unsigned long long)
193 (gart_addr - adev->mc.gart_start +
194 (void*)vram_start - vram_map));
195 amdgpu_bo_kunmap(gtt_obj[i]);
196 goto out_lclean_unpin;
197 }
198 }
199
200 amdgpu_bo_kunmap(gtt_obj[i]);
201
202 DRM_INFO("Tested GTT->VRAM and VRAM->GTT copy for GTT offset 0x%llx\n",
203 gart_addr - adev->mc.gart_start);
204 continue;
205
206out_lclean_unpin:
207 amdgpu_bo_unpin(gtt_obj[i]);
208out_lclean_unres:
209 amdgpu_bo_unreserve(gtt_obj[i]);
210out_lclean_unref:
211 amdgpu_bo_unref(>t_obj[i]);
212out_lclean:
213 for (--i; i >= 0; --i) {
214 amdgpu_bo_unpin(gtt_obj[i]);
215 amdgpu_bo_unreserve(gtt_obj[i]);
216 amdgpu_bo_unref(>t_obj[i]);
217 }
218 if (fence)
219 dma_fence_put(fence);
220 break;
221 }
222
223 amdgpu_bo_unpin(vram_obj);
224out_unres:
225 amdgpu_bo_unreserve(vram_obj);
226out_unref:
227 amdgpu_bo_unref(&vram_obj);
228out_cleanup:
229 kfree(gtt_obj);
230 if (r) {
231 pr_warn("Error while testing BO move\n");
232 }
233}
234
235void amdgpu_test_moves(struct amdgpu_device *adev)
236{
237 if (adev->mman.buffer_funcs)
238 amdgpu_do_test_moves(adev);
239}
240