1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
|
From ce9a1734fbccf5fa7fb28a314f25e1f6f244f988 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig@amd.com>
Date: Mon, 15 Feb 2016 12:33:02 +0100
Subject: [PATCH 0467/1110] drm/amdgpu: use a sync object for VMID fences v2
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
v2: rebase & cleanup
This way we can store more than one fence as user for each VMID.
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com> (v1)
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Kalyan Alle <kalyan.alle@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 7 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c | 105 +++++++++++++++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 54 ++++++++--------
3 files changed, 133 insertions(+), 33 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 4b9fe35..915c2e3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -589,6 +589,9 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
struct amdgpu_sync *sync,
struct reservation_object *resv,
void *owner);
+bool amdgpu_sync_is_idle(struct amdgpu_sync *sync);
+int amdgpu_sync_cycle_fences(struct amdgpu_sync *dst, struct amdgpu_sync *src,
+ struct fence *fence);
struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync);
int amdgpu_sync_wait(struct amdgpu_sync *sync);
void amdgpu_sync_free(struct amdgpu_sync *sync);
@@ -874,8 +877,10 @@ struct amdgpu_vm {
struct amdgpu_vm_id {
struct list_head list;
- struct fence *active;
+ struct fence *first;
+ struct amdgpu_sync active;
atomic_long_t owner;
+
uint64_t pd_gpu_addr;
/* last flushed PD/PT update */
struct fence *flushed_updates;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index c48b4fc..34a9280 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -109,6 +109,29 @@ static void amdgpu_sync_keep_later(struct fence **keep, struct fence *fence)
}
/**
+ * amdgpu_sync_add_later - add the fence to the hash
+ *
+ * @sync: sync object to add the fence to
+ * @f: fence to add
+ *
+ * Tries to add the fence to an existing hash entry. Returns true when an entry
+ * was found, false otherwise.
+ */
+static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct fence *f)
+{
+ struct amdgpu_sync_entry *e;
+
+ hash_for_each_possible(sync->fences, e, node, f->context) {
+ if (unlikely(e->fence->context != f->context))
+ continue;
+
+ amdgpu_sync_keep_later(&e->fence, f);
+ return true;
+ }
+ return false;
+}
+
+/**
* amdgpu_sync_fence - remember to sync to this fence
*
* @sync: sync object to add fence to
@@ -127,13 +150,8 @@ int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM)
amdgpu_sync_keep_later(&sync->last_vm_update, f);
- hash_for_each_possible(sync->fences, e, node, f->context) {
- if (unlikely(e->fence->context != f->context))
- continue;
-
- amdgpu_sync_keep_later(&e->fence, f);
+ if (amdgpu_sync_add_later(sync, f))
return 0;
- }
e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
if (!e)
@@ -204,6 +222,81 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
return r;
}
+/**
+ * amdgpu_sync_is_idle - test if all fences are signaled
+ *
+ * @sync: the sync object
+ *
+ * Returns true if all fences in the sync object are signaled.
+ */
+bool amdgpu_sync_is_idle(struct amdgpu_sync *sync)
+{
+ struct amdgpu_sync_entry *e;
+ struct hlist_node *tmp;
+ int i;
+
+ hash_for_each_safe(sync->fences, i, tmp, e, node) {
+ struct fence *f = e->fence;
+
+ if (fence_is_signaled(f)) {
+ hash_del(&e->node);
+ fence_put(f);
+ kmem_cache_free(amdgpu_sync_slab, e);
+ continue;
+ }
+
+ return false;
+ }
+
+ return true;
+}
+
+/**
+ * amdgpu_sync_cycle_fences - move fences from one sync object into another
+ *
+ * @dst: the destination sync object
+ * @src: the source sync object
+ * @fence: fence to add to source
+ *
+ * Remove all fences from source and put them into destination and add
+ * fence as new one into source.
+ */
+int amdgpu_sync_cycle_fences(struct amdgpu_sync *dst, struct amdgpu_sync *src,
+ struct fence *fence)
+{
+ struct amdgpu_sync_entry *e, *newone;
+ struct hlist_node *tmp;
+ int i;
+
+ /* Allocate the new entry before moving the old ones */
+ newone = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
+ if (!newone)
+ return -ENOMEM;
+
+ hash_for_each_safe(src->fences, i, tmp, e, node) {
+ struct fence *f = e->fence;
+
+ hash_del(&e->node);
+ if (fence_is_signaled(f)) {
+ fence_put(f);
+ kmem_cache_free(amdgpu_sync_slab, e);
+ continue;
+ }
+
+ if (amdgpu_sync_add_later(dst, f)) {
+ kmem_cache_free(amdgpu_sync_slab, e);
+ continue;
+ }
+
+ hash_add(dst->fences, &e->node, f->context);
+ }
+
+ hash_add(src->fences, &newone->node, fence->context);
+ newone->fence = fence_get(fence);
+
+ return 0;
+}
+
struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
{
struct amdgpu_sync_entry *e;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 1c87193..8382b7f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -196,14 +196,13 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
if (usable) {
- r = amdgpu_sync_fence(ring->adev, sync, id->active);
- if (r) {
- mutex_unlock(&adev->vm_manager.lock);
- return r;
- }
+ r = amdgpu_sync_fence(ring->adev, sync, id->first);
+ if (r)
+ goto error;
- fence_put(id->active);
- id->active = fence_get(fence);
+ r = amdgpu_sync_fence(ring->adev, &id->active, fence);
+ if (r)
+ goto error;
list_move_tail(&id->list, &adev->vm_manager.ids_lru);
@@ -221,13 +220,13 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
struct amdgpu_vm_id,
list);
- if (id->active && !fence_is_signaled(id->active)) {
- struct amdgpu_vm_id *tmp;
+ if (!amdgpu_sync_is_idle(&id->active)) {
struct list_head *head = &adev->vm_manager.ids_lru;
+ struct amdgpu_vm_id *tmp;
list_for_each_entry_safe(id, tmp, &adev->vm_manager.ids_lru,
list) {
- if (id->active && fence_is_signaled(id->active)) {
+ if (amdgpu_sync_is_idle(&id->active)) {
list_move(&id->list, head);
head = &id->list;
}
@@ -237,25 +236,27 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
list);
}
- r = amdgpu_sync_fence(ring->adev, sync, id->active);
- if (!r) {
- fence_put(id->active);
- id->active = fence_get(fence);
+ r = amdgpu_sync_cycle_fences(sync, &id->active, fence);
+ if (r)
+ goto error;
- fence_put(id->flushed_updates);
- id->flushed_updates = fence_get(updates);
+ fence_put(id->first);
+ id->first = fence_get(fence);
- id->pd_gpu_addr = pd_addr;
+ fence_put(id->flushed_updates);
+ id->flushed_updates = fence_get(updates);
- list_move_tail(&id->list, &adev->vm_manager.ids_lru);
- atomic_long_set(&id->owner, (long)&vm->ids[ring->idx]);
- vm->ids[ring->idx] = id;
+ id->pd_gpu_addr = pd_addr;
- *vm_id = id - adev->vm_manager.ids;
- *vm_pd_addr = pd_addr;
- trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id, *vm_pd_addr);
- }
+ list_move_tail(&id->list, &adev->vm_manager.ids_lru);
+ atomic_long_set(&id->owner, (long)id);
+ vm->ids[ring->idx] = id;
+ *vm_id = id - adev->vm_manager.ids;
+ *vm_pd_addr = pd_addr;
+ trace_amdgpu_vm_grab_id(vm, ring->idx, *vm_id, *vm_pd_addr);
+
+error:
mutex_unlock(&adev->vm_manager.lock);
return r;
}
@@ -1470,7 +1471,7 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev)
/* skip over VMID 0, since it is the system VM */
for (i = 1; i < adev->vm_manager.num_ids; ++i) {
- amdgpu_vm_reset_id(adev, i);
+ amdgpu_sync_create(&adev->vm_manager.ids[i].active);
list_add_tail(&adev->vm_manager.ids[i].list,
&adev->vm_manager.ids_lru);
}
@@ -1493,7 +1494,8 @@ void amdgpu_vm_manager_fini(struct amdgpu_device *adev)
for (i = 0; i < AMDGPU_NUM_VM; ++i) {
struct amdgpu_vm_id *id = &adev->vm_manager.ids[i];
- fence_put(id->active);
+ fence_put(adev->vm_manager.ids[i].first);
+ amdgpu_sync_free(&adev->vm_manager.ids[i].active);
fence_put(id->flushed_updates);
}
}
--
2.7.4
|
