1 /*
2 * Copyright © 2021 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * 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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include "anv_private.h"
25
26 #include "util/os_time.h"
27
28 static struct anv_bo_sync *
to_anv_bo_sync(struct vk_sync * sync)29 to_anv_bo_sync(struct vk_sync *sync)
30 {
31 assert(sync->type == &anv_bo_sync_type);
32 return container_of(sync, struct anv_bo_sync, sync);
33 }
34
35 static VkResult
anv_bo_sync_init(struct vk_device * vk_device,struct vk_sync * vk_sync,uint64_t initial_value)36 anv_bo_sync_init(struct vk_device *vk_device,
37 struct vk_sync *vk_sync,
38 uint64_t initial_value)
39 {
40 struct anv_device *device = container_of(vk_device, struct anv_device, vk);
41 struct anv_bo_sync *sync = to_anv_bo_sync(vk_sync);
42
43 sync->state = initial_value ? ANV_BO_SYNC_STATE_SIGNALED :
44 ANV_BO_SYNC_STATE_RESET;
45
46 return anv_device_alloc_bo(device, "bo-sync", 4096,
47 ANV_BO_ALLOC_EXTERNAL |
48 ANV_BO_ALLOC_IMPLICIT_SYNC,
49 0 /* explicit_address */,
50 &sync->bo);
51 }
52
53 static void
anv_bo_sync_finish(struct vk_device * vk_device,struct vk_sync * vk_sync)54 anv_bo_sync_finish(struct vk_device *vk_device,
55 struct vk_sync *vk_sync)
56 {
57 struct anv_device *device = container_of(vk_device, struct anv_device, vk);
58 struct anv_bo_sync *sync = to_anv_bo_sync(vk_sync);
59
60 anv_device_release_bo(device, sync->bo);
61 }
62
63 static VkResult
anv_bo_sync_reset(struct vk_device * vk_device,struct vk_sync * vk_sync)64 anv_bo_sync_reset(struct vk_device *vk_device,
65 struct vk_sync *vk_sync)
66 {
67 struct anv_bo_sync *sync = to_anv_bo_sync(vk_sync);
68
69 sync->state = ANV_BO_SYNC_STATE_RESET;
70
71 return VK_SUCCESS;
72 }
73
74 static int64_t
anv_get_relative_timeout(uint64_t abs_timeout)75 anv_get_relative_timeout(uint64_t abs_timeout)
76 {
77 uint64_t now = os_time_get_nano();
78
79 /* We don't want negative timeouts.
80 *
81 * DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is
82 * supposed to block indefinitely timeouts < 0. Unfortunately,
83 * this was broken for a couple of kernel releases. Since there's
84 * no way to know whether or not the kernel we're using is one of
85 * the broken ones, the best we can do is to clamp the timeout to
86 * INT64_MAX. This limits the maximum timeout from 584 years to
87 * 292 years - likely not a big deal.
88 */
89 if (abs_timeout < now)
90 return 0;
91
92 uint64_t rel_timeout = abs_timeout - now;
93 if (rel_timeout > (uint64_t) INT64_MAX)
94 rel_timeout = INT64_MAX;
95
96 return rel_timeout;
97 }
98
99 static VkResult
anv_bo_sync_wait(struct vk_device * vk_device,uint32_t wait_count,const struct vk_sync_wait * waits,enum vk_sync_wait_flags wait_flags,uint64_t abs_timeout_ns)100 anv_bo_sync_wait(struct vk_device *vk_device,
101 uint32_t wait_count,
102 const struct vk_sync_wait *waits,
103 enum vk_sync_wait_flags wait_flags,
104 uint64_t abs_timeout_ns)
105 {
106 struct anv_device *device = container_of(vk_device, struct anv_device, vk);
107 VkResult result;
108
109 uint32_t pending = wait_count;
110 while (pending) {
111 pending = 0;
112 bool signaled = false;
113 for (uint32_t i = 0; i < wait_count; i++) {
114 struct anv_bo_sync *sync = to_anv_bo_sync(waits[i].sync);
115 switch (sync->state) {
116 case ANV_BO_SYNC_STATE_RESET:
117 /* This fence hasn't been submitted yet, we'll catch it the next
118 * time around. Yes, this may mean we dead-loop but, short of
119 * lots of locking and a condition variable, there's not much that
120 * we can do about that.
121 */
122 assert(!(wait_flags & VK_SYNC_WAIT_PENDING));
123 pending++;
124 continue;
125
126 case ANV_BO_SYNC_STATE_SIGNALED:
127 /* This fence is not pending. If waitAll isn't set, we can return
128 * early. Otherwise, we have to keep going.
129 */
130 if (wait_flags & VK_SYNC_WAIT_ANY)
131 return VK_SUCCESS;
132 continue;
133
134 case ANV_BO_SYNC_STATE_SUBMITTED:
135 /* These are the fences we really care about. Go ahead and wait
136 * on it until we hit a timeout.
137 */
138 if (!(wait_flags & VK_SYNC_WAIT_PENDING)) {
139 uint64_t rel_timeout = anv_get_relative_timeout(abs_timeout_ns);
140 result = anv_device_wait(device, sync->bo, rel_timeout);
141 /* This also covers VK_TIMEOUT */
142 if (result != VK_SUCCESS)
143 return result;
144
145 sync->state = ANV_BO_SYNC_STATE_SIGNALED;
146 signaled = true;
147 }
148 if (wait_flags & VK_SYNC_WAIT_ANY)
149 return VK_SUCCESS;
150 break;
151
152 default:
153 unreachable("Invalid BO sync state");
154 }
155 }
156
157 if (pending && !signaled) {
158 /* If we've hit this then someone decided to vkWaitForFences before
159 * they've actually submitted any of them to a queue. This is a
160 * fairly pessimal case, so it's ok to lock here and use a standard
161 * pthreads condition variable.
162 */
163 pthread_mutex_lock(&device->mutex);
164
165 /* It's possible that some of the fences have changed state since the
166 * last time we checked. Now that we have the lock, check for
167 * pending fences again and don't wait if it's changed.
168 */
169 uint32_t now_pending = 0;
170 for (uint32_t i = 0; i < wait_count; i++) {
171 struct anv_bo_sync *sync = to_anv_bo_sync(waits[i].sync);
172 if (sync->state == ANV_BO_SYNC_STATE_RESET)
173 now_pending++;
174 }
175 assert(now_pending <= pending);
176
177 if (now_pending == pending) {
178 struct timespec abstime = {
179 .tv_sec = abs_timeout_ns / NSEC_PER_SEC,
180 .tv_nsec = abs_timeout_ns % NSEC_PER_SEC,
181 };
182
183 ASSERTED int ret;
184 ret = pthread_cond_timedwait(&device->queue_submit,
185 &device->mutex, &abstime);
186 assert(ret != EINVAL);
187 if (os_time_get_nano() >= abs_timeout_ns) {
188 pthread_mutex_unlock(&device->mutex);
189 return VK_TIMEOUT;
190 }
191 }
192
193 pthread_mutex_unlock(&device->mutex);
194 }
195 }
196
197 return VK_SUCCESS;
198 }
199
200 const struct vk_sync_type anv_bo_sync_type = {
201 .size = sizeof(struct anv_bo_sync),
202 .features = VK_SYNC_FEATURE_BINARY |
203 VK_SYNC_FEATURE_GPU_WAIT |
204 VK_SYNC_FEATURE_GPU_MULTI_WAIT |
205 VK_SYNC_FEATURE_CPU_WAIT |
206 VK_SYNC_FEATURE_CPU_RESET |
207 VK_SYNC_FEATURE_WAIT_ANY |
208 VK_SYNC_FEATURE_WAIT_PENDING,
209 .init = anv_bo_sync_init,
210 .finish = anv_bo_sync_finish,
211 .reset = anv_bo_sync_reset,
212 .wait_many = anv_bo_sync_wait,
213 };
214
215 VkResult
anv_create_sync_for_memory(struct vk_device * device,VkDeviceMemory memory,bool signal_memory,struct vk_sync ** sync_out)216 anv_create_sync_for_memory(struct vk_device *device,
217 VkDeviceMemory memory,
218 bool signal_memory,
219 struct vk_sync **sync_out)
220 {
221 ANV_FROM_HANDLE(anv_device_memory, mem, memory);
222 struct anv_bo_sync *bo_sync;
223
224 bo_sync = vk_zalloc(&device->alloc, sizeof(*bo_sync), 8,
225 VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
226 if (bo_sync == NULL)
227 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
228
229 bo_sync->sync.type = &anv_bo_sync_type;
230 bo_sync->state = signal_memory ? ANV_BO_SYNC_STATE_RESET :
231 ANV_BO_SYNC_STATE_SUBMITTED;
232 bo_sync->bo = anv_bo_ref(mem->bo);
233
234 *sync_out = &bo_sync->sync;
235
236 return VK_SUCCESS;
237 }
238