1 /*
2 * Copyright © 2016 Advanced Micro Devices, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a 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,
14 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
15 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
16 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
17 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
21 *
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
24 * of the Software.
25 */
26
27 /* Job queue with execution in a separate thread.
28 *
29 * Jobs can be added from any thread. After that, the wait call can be used
30 * to wait for completion of the job.
31 */
32
33 #ifndef U_QUEUE_H
34 #define U_QUEUE_H
35
36 #include <string.h>
37
38 #include "util/futex.h"
39 #include "util/list.h"
40 #include "util/macros.h"
41 #include "util/os_time.h"
42 #include "util/u_atomic.h"
43 #include "util/u_thread.h"
44
45 #ifdef __cplusplus
46 extern "C" {
47 #endif
48
49 #define UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY (1 << 0)
50 #define UTIL_QUEUE_INIT_RESIZE_IF_FULL (1 << 1)
51
52 #if defined(__GNUC__) && defined(HAVE_LINUX_FUTEX_H)
53 #define UTIL_QUEUE_FENCE_FUTEX
54 #else
55 #define UTIL_QUEUE_FENCE_STANDARD
56 #endif
57
58 #ifdef UTIL_QUEUE_FENCE_FUTEX
59 /* Job completion fence.
60 * Put this into your job structure.
61 */
62 struct util_queue_fence {
63 /* The fence can be in one of three states:
64 * 0 - signaled
65 * 1 - unsignaled
66 * 2 - unsignaled, may have waiters
67 */
68 uint32_t val;
69 };
70
71 static inline void
util_queue_fence_init(struct util_queue_fence * fence)72 util_queue_fence_init(struct util_queue_fence *fence)
73 {
74 fence->val = 0;
75 }
76
77 static inline void
util_queue_fence_destroy(struct util_queue_fence * fence)78 util_queue_fence_destroy(struct util_queue_fence *fence)
79 {
80 assert(fence->val == 0);
81 /* no-op */
82 }
83
84 static inline void
util_queue_fence_signal(struct util_queue_fence * fence)85 util_queue_fence_signal(struct util_queue_fence *fence)
86 {
87 uint32_t val = p_atomic_xchg(&fence->val, 0);
88
89 assert(val != 0);
90
91 if (val == 2)
92 futex_wake(&fence->val, INT_MAX);
93 }
94
95 /**
96 * Move \p fence back into unsignalled state.
97 *
98 * \warning The caller must ensure that no other thread may currently be
99 * waiting (or about to wait) on the fence.
100 */
101 static inline void
util_queue_fence_reset(struct util_queue_fence * fence)102 util_queue_fence_reset(struct util_queue_fence *fence)
103 {
104 #ifdef NDEBUG
105 fence->val = 1;
106 #else
107 uint32_t v = p_atomic_xchg(&fence->val, 1);
108 assert(v == 0);
109 #endif
110 }
111
112 static inline bool
util_queue_fence_is_signalled(struct util_queue_fence * fence)113 util_queue_fence_is_signalled(struct util_queue_fence *fence)
114 {
115 return fence->val == 0;
116 }
117 #endif
118
119 #ifdef UTIL_QUEUE_FENCE_STANDARD
120 /* Job completion fence.
121 * Put this into your job structure.
122 */
123 struct util_queue_fence {
124 mtx_t mutex;
125 cnd_t cond;
126 int signalled;
127 };
128
129 void util_queue_fence_init(struct util_queue_fence *fence);
130 void util_queue_fence_destroy(struct util_queue_fence *fence);
131 void util_queue_fence_signal(struct util_queue_fence *fence);
132
133 /**
134 * Move \p fence back into unsignalled state.
135 *
136 * \warning The caller must ensure that no other thread may currently be
137 * waiting (or about to wait) on the fence.
138 */
139 static inline void
util_queue_fence_reset(struct util_queue_fence * fence)140 util_queue_fence_reset(struct util_queue_fence *fence)
141 {
142 assert(fence->signalled);
143 fence->signalled = 0;
144 }
145
146 static inline bool
util_queue_fence_is_signalled(struct util_queue_fence * fence)147 util_queue_fence_is_signalled(struct util_queue_fence *fence)
148 {
149 return fence->signalled != 0;
150 }
151 #endif
152
153 void
154 _util_queue_fence_wait(struct util_queue_fence *fence);
155
156 static inline void
util_queue_fence_wait(struct util_queue_fence * fence)157 util_queue_fence_wait(struct util_queue_fence *fence)
158 {
159 if (unlikely(!util_queue_fence_is_signalled(fence)))
160 _util_queue_fence_wait(fence);
161 }
162
163 bool
164 _util_queue_fence_wait_timeout(struct util_queue_fence *fence,
165 int64_t abs_timeout);
166
167 /**
168 * Wait for the fence to be signaled with a timeout.
169 *
170 * \param fence the fence
171 * \param abs_timeout the absolute timeout in nanoseconds, relative to the
172 * clock provided by os_time_get_nano.
173 *
174 * \return true if the fence was signaled, false if the timeout occurred.
175 */
176 static inline bool
util_queue_fence_wait_timeout(struct util_queue_fence * fence,int64_t abs_timeout)177 util_queue_fence_wait_timeout(struct util_queue_fence *fence,
178 int64_t abs_timeout)
179 {
180 if (util_queue_fence_is_signalled(fence))
181 return true;
182
183 if (abs_timeout == (int64_t)OS_TIMEOUT_INFINITE) {
184 _util_queue_fence_wait(fence);
185 return true;
186 }
187
188 return _util_queue_fence_wait_timeout(fence, abs_timeout);
189 }
190
191 typedef void (*util_queue_execute_func)(void *job, int thread_index);
192
193 struct util_queue_job {
194 void *job;
195 struct util_queue_fence *fence;
196 util_queue_execute_func execute;
197 util_queue_execute_func cleanup;
198 };
199
200 /* Put this into your context. */
201 struct util_queue {
202 const char *name;
203 mtx_t finish_lock; /* only for util_queue_finish */
204 mtx_t lock;
205 cnd_t has_queued_cond;
206 cnd_t has_space_cond;
207 thrd_t *threads;
208 unsigned flags;
209 int num_queued;
210 unsigned num_threads;
211 int kill_threads;
212 int max_jobs;
213 int write_idx, read_idx; /* ring buffer pointers */
214 struct util_queue_job *jobs;
215
216 /* for cleanup at exit(), protected by exit_mutex */
217 struct list_head head;
218 };
219
220 bool util_queue_init(struct util_queue *queue,
221 const char *name,
222 unsigned max_jobs,
223 unsigned num_threads,
224 unsigned flags);
225 void util_queue_destroy(struct util_queue *queue);
226
227 /* optional cleanup callback is called after fence is signaled: */
228 void util_queue_add_job(struct util_queue *queue,
229 void *job,
230 struct util_queue_fence *fence,
231 util_queue_execute_func execute,
232 util_queue_execute_func cleanup);
233 void util_queue_drop_job(struct util_queue *queue,
234 struct util_queue_fence *fence);
235
236 void util_queue_finish(struct util_queue *queue);
237
238 int64_t util_queue_get_thread_time_nano(struct util_queue *queue,
239 unsigned thread_index);
240
241 /* util_queue needs to be cleared to zeroes for this to work */
242 static inline bool
util_queue_is_initialized(struct util_queue * queue)243 util_queue_is_initialized(struct util_queue *queue)
244 {
245 return queue->threads != NULL;
246 }
247
248 /* Convenient structure for monitoring the queue externally and passing
249 * the structure between Mesa components. The queue doesn't use it directly.
250 */
251 struct util_queue_monitoring
252 {
253 /* For querying the thread busyness. */
254 struct util_queue *queue;
255
256 /* Counters updated by the user of the queue. */
257 unsigned num_offloaded_items;
258 unsigned num_direct_items;
259 unsigned num_syncs;
260 };
261
262 #ifdef __cplusplus
263 }
264 #endif
265
266 #endif
267