drm/amdgpu: fix syncing to VM updates
[firefly-linux-kernel-4.4.55.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_sync.c
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
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  *    Christian König <christian.koenig@amd.com>
29  */
30
31 #include <drm/drmP.h>
32 #include "amdgpu.h"
33 #include "amdgpu_trace.h"
34
35 /**
36  * amdgpu_sync_create - zero init sync object
37  *
38  * @sync: sync object to initialize
39  *
40  * Just clear the sync object for now.
41  */
42 void amdgpu_sync_create(struct amdgpu_sync *sync)
43 {
44         unsigned i;
45
46         for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
47                 sync->semaphores[i] = NULL;
48
49         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
50                 sync->sync_to[i] = NULL;
51
52         sync->last_vm_update = NULL;
53 }
54
55 /**
56  * amdgpu_sync_fence - remember to sync to this fence
57  *
58  * @sync: sync object to add fence to
59  * @fence: fence to sync to
60  *
61  */
62 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
63                       struct fence *f)
64 {
65         struct amdgpu_fence *fence;
66         struct amdgpu_fence *other;
67
68         if (!f)
69                 return 0;
70
71         fence = to_amdgpu_fence(f);
72         if (!fence || fence->ring->adev != adev)
73                 return fence_wait(f, true);
74
75         other = sync->sync_to[fence->ring->idx];
76         sync->sync_to[fence->ring->idx] = amdgpu_fence_ref(
77                 amdgpu_fence_later(fence, other));
78         amdgpu_fence_unref(&other);
79
80         if (fence->owner == AMDGPU_FENCE_OWNER_VM) {
81                 other = sync->last_vm_update;
82                 sync->last_vm_update = amdgpu_fence_ref(
83                         amdgpu_fence_later(fence, other));
84                 amdgpu_fence_unref(&other);
85         }
86
87         return 0;
88 }
89
90 /**
91  * amdgpu_sync_resv - use the semaphores to sync to a reservation object
92  *
93  * @sync: sync object to add fences from reservation object to
94  * @resv: reservation object with embedded fence
95  * @shared: true if we should only sync to the exclusive fence
96  *
97  * Sync to the fence using the semaphore objects
98  */
99 int amdgpu_sync_resv(struct amdgpu_device *adev,
100                      struct amdgpu_sync *sync,
101                      struct reservation_object *resv,
102                      void *owner)
103 {
104         struct reservation_object_list *flist;
105         struct fence *f;
106         struct amdgpu_fence *fence;
107         unsigned i;
108         int r = 0;
109
110         if (resv == NULL)
111                 return -EINVAL;
112
113         /* always sync to the exclusive fence */
114         f = reservation_object_get_excl(resv);
115         r = amdgpu_sync_fence(adev, sync, f);
116
117         flist = reservation_object_get_list(resv);
118         if (!flist || r)
119                 return r;
120
121         for (i = 0; i < flist->shared_count; ++i) {
122                 f = rcu_dereference_protected(flist->shared[i],
123                                               reservation_object_held(resv));
124                 fence = f ? to_amdgpu_fence(f) : NULL;
125                 if (fence && fence->ring->adev == adev) {
126                         /* VM updates are only interesting
127                          * for other VM updates and moves.
128                          */
129                         if ((owner != AMDGPU_FENCE_OWNER_MOVE) &&
130                             (fence->owner != AMDGPU_FENCE_OWNER_MOVE) &&
131                             ((owner == AMDGPU_FENCE_OWNER_VM) !=
132                              (fence->owner == AMDGPU_FENCE_OWNER_VM)))
133                                 continue;
134
135                         /* Ignore fence from the same owner as
136                          * long as it isn't undefined.
137                          */
138                         if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
139                             fence->owner == owner)
140                                 continue;
141                 }
142
143                 r = amdgpu_sync_fence(adev, sync, f);
144                 if (r)
145                         break;
146         }
147         return r;
148 }
149
150 /**
151  * amdgpu_sync_rings - sync ring to all registered fences
152  *
153  * @sync: sync object to use
154  * @ring: ring that needs sync
155  *
156  * Ensure that all registered fences are signaled before letting
157  * the ring continue. The caller must hold the ring lock.
158  */
159 int amdgpu_sync_rings(struct amdgpu_sync *sync,
160                       struct amdgpu_ring *ring)
161 {
162         struct amdgpu_device *adev = ring->adev;
163         unsigned count = 0;
164         int i, r;
165
166         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
167                 struct amdgpu_fence *fence = sync->sync_to[i];
168                 struct amdgpu_semaphore *semaphore;
169                 struct amdgpu_ring *other = adev->rings[i];
170
171                 /* check if we really need to sync */
172                 if (!amdgpu_fence_need_sync(fence, ring))
173                         continue;
174
175                 /* prevent GPU deadlocks */
176                 if (!other->ready) {
177                         dev_err(adev->dev, "Syncing to a disabled ring!");
178                         return -EINVAL;
179                 }
180
181                 if (amdgpu_enable_scheduler || (count >= AMDGPU_NUM_SYNCS)) {
182                         /* not enough room, wait manually */
183                         r = amdgpu_fence_wait(fence, false);
184                         if (r)
185                                 return r;
186                         continue;
187                 }
188                 r = amdgpu_semaphore_create(adev, &semaphore);
189                 if (r)
190                         return r;
191
192                 sync->semaphores[count++] = semaphore;
193
194                 /* allocate enough space for sync command */
195                 r = amdgpu_ring_alloc(other, 16);
196                 if (r)
197                         return r;
198
199                 /* emit the signal semaphore */
200                 if (!amdgpu_semaphore_emit_signal(other, semaphore)) {
201                         /* signaling wasn't successful wait manually */
202                         amdgpu_ring_undo(other);
203                         r = amdgpu_fence_wait(fence, false);
204                         if (r)
205                                 return r;
206                         continue;
207                 }
208
209                 /* we assume caller has already allocated space on waiters ring */
210                 if (!amdgpu_semaphore_emit_wait(ring, semaphore)) {
211                         /* waiting wasn't successful wait manually */
212                         amdgpu_ring_undo(other);
213                         r = amdgpu_fence_wait(fence, false);
214                         if (r)
215                                 return r;
216                         continue;
217                 }
218
219                 amdgpu_ring_commit(other);
220                 amdgpu_fence_note_sync(fence, ring);
221         }
222
223         return 0;
224 }
225
226 /**
227  * amdgpu_sync_free - free the sync object
228  *
229  * @adev: amdgpu_device pointer
230  * @sync: sync object to use
231  * @fence: fence to use for the free
232  *
233  * Free the sync object by freeing all semaphores in it.
234  */
235 void amdgpu_sync_free(struct amdgpu_device *adev,
236                       struct amdgpu_sync *sync,
237                       struct amdgpu_fence *fence)
238 {
239         unsigned i;
240
241         for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
242                 amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
243
244         for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
245                 amdgpu_fence_unref(&sync->sync_to[i]);
246
247         amdgpu_fence_unref(&sync->last_vm_update);
248 }