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 "simple_mtx.h"
39 #include "util/futex.h"
40 #include "util/list.h"
41 #include "util/macros.h"
42 #include "util/os_time.h"
43 #include "util/u_atomic.h"
44 #include "util/u_thread.h"
45
46 #ifdef __cplusplus
47 extern "C" {
48 #endif
49
50 #define UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY (1 << 0)
51 #define UTIL_QUEUE_INIT_RESIZE_IF_FULL (1 << 1)
52 #define UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY (1 << 2)
53
54 #if UTIL_FUTEX_SUPPORTED
55 #define UTIL_QUEUE_FENCE_FUTEX
56 #else
57 #define UTIL_QUEUE_FENCE_STANDARD
58 #endif
59
60 #ifdef UTIL_QUEUE_FENCE_FUTEX
61 /* Job completion fence.
62 * Put this into your job structure.
63 */
64 struct util_queue_fence {
65 /* The fence can be in one of three states:
66 * 0 - signaled
67 * 1 - unsignaled
68 * 2 - unsignaled, may have waiters
69 */
70 uint32_t val;
71 };
72
73 static inline void
util_queue_fence_init(struct util_queue_fence * fence)74 util_queue_fence_init(struct util_queue_fence *fence)
75 {
76 fence->val = 0;
77 }
78
79 static inline void
util_queue_fence_destroy(struct util_queue_fence * fence)80 util_queue_fence_destroy(struct util_queue_fence *fence)
81 {
82 assert(p_atomic_read_relaxed(&fence->val) == 0);
83 /* no-op */
84 }
85
86 static inline void
util_queue_fence_signal(struct util_queue_fence * fence)87 util_queue_fence_signal(struct util_queue_fence *fence)
88 {
89 uint32_t val = p_atomic_xchg(&fence->val, 0);
90
91 assert(val != 0);
92
93 if (val == 2)
94 futex_wake(&fence->val, INT_MAX);
95 }
96
97 /**
98 * Move \p fence back into unsignalled state.
99 *
100 * \warning The caller must ensure that no other thread may currently be
101 * waiting (or about to wait) on the fence.
102 */
103 static inline void
util_queue_fence_reset(struct util_queue_fence * fence)104 util_queue_fence_reset(struct util_queue_fence *fence)
105 {
106 #ifdef NDEBUG
107 fence->val = 1;
108 #else
109 uint32_t v = p_atomic_xchg(&fence->val, 1);
110 assert(v == 0);
111 #endif
112 }
113
114 static inline bool
util_queue_fence_is_signalled(struct util_queue_fence * fence)115 util_queue_fence_is_signalled(struct util_queue_fence *fence)
116 {
117 return p_atomic_read_relaxed(&fence->val) == 0;
118 }
119 #endif
120
121 #ifdef UTIL_QUEUE_FENCE_STANDARD
122 /* Job completion fence.
123 * Put this into your job structure.
124 */
125 struct util_queue_fence {
126 mtx_t mutex;
127 cnd_t cond;
128 int signalled;
129 };
130
131 void util_queue_fence_init(struct util_queue_fence *fence);
132 void util_queue_fence_destroy(struct util_queue_fence *fence);
133 void util_queue_fence_signal(struct util_queue_fence *fence);
134
135 /**
136 * Move \p fence back into unsignalled state.
137 *
138 * \warning The caller must ensure that no other thread may currently be
139 * waiting (or about to wait) on the fence.
140 */
141 static inline void
util_queue_fence_reset(struct util_queue_fence * fence)142 util_queue_fence_reset(struct util_queue_fence *fence)
143 {
144 assert(fence->signalled);
145 fence->signalled = 0;
146 }
147
148 static inline bool
util_queue_fence_is_signalled(struct util_queue_fence * fence)149 util_queue_fence_is_signalled(struct util_queue_fence *fence)
150 {
151 return fence->signalled != 0;
152 }
153 #endif
154
155 void
156 _util_queue_fence_wait(struct util_queue_fence *fence);
157
158 static inline void
util_queue_fence_wait(struct util_queue_fence * fence)159 util_queue_fence_wait(struct util_queue_fence *fence)
160 {
161 if (unlikely(!util_queue_fence_is_signalled(fence)))
162 _util_queue_fence_wait(fence);
163 }
164
165 bool
166 _util_queue_fence_wait_timeout(struct util_queue_fence *fence,
167 int64_t abs_timeout);
168
169 /**
170 * Wait for the fence to be signaled with a timeout.
171 *
172 * \param fence the fence
173 * \param abs_timeout the absolute timeout in nanoseconds, relative to the
174 * clock provided by os_time_get_nano.
175 *
176 * \return true if the fence was signaled, false if the timeout occurred.
177 */
178 static inline bool
util_queue_fence_wait_timeout(struct util_queue_fence * fence,int64_t abs_timeout)179 util_queue_fence_wait_timeout(struct util_queue_fence *fence,
180 int64_t abs_timeout)
181 {
182 if (util_queue_fence_is_signalled(fence))
183 return true;
184
185 if (abs_timeout == (int64_t)OS_TIMEOUT_INFINITE) {
186 _util_queue_fence_wait(fence);
187 return true;
188 }
189
190 return _util_queue_fence_wait_timeout(fence, abs_timeout);
191 }
192
193 typedef void (*util_queue_execute_func)(void *job, void *gdata, int thread_index);
194
195 struct util_queue_job {
196 void *job;
197 void *global_data;
198 size_t job_size;
199 struct util_queue_fence *fence;
200 util_queue_execute_func execute;
201 util_queue_execute_func cleanup;
202 };
203
204 /* Put this into your context. */
205 struct util_queue {
206 char name[14]; /* 13 characters = the thread name without the index */
207 mtx_t lock;
208 bool create_threads_on_demand;
209 cnd_t has_queued_cond;
210 cnd_t has_space_cond;
211 thrd_t *threads;
212 unsigned flags;
213 int num_queued;
214 unsigned max_threads;
215 unsigned num_threads; /* decreasing this number will terminate threads */
216 int max_jobs;
217 int write_idx, read_idx; /* ring buffer pointers */
218 size_t total_jobs_size; /* memory use of all jobs in the queue */
219 struct util_queue_job *jobs;
220 void *global_data;
221
222 /* for cleanup at exit(), protected by exit_mutex */
223 struct list_head head;
224 };
225
226 bool util_queue_init(struct util_queue *queue,
227 const char *name,
228 unsigned max_jobs,
229 unsigned num_threads,
230 unsigned flags,
231 void *global_data);
232 void util_queue_destroy(struct util_queue *queue);
233
234 /* optional cleanup callback is called after fence is signaled: */
235 void util_queue_add_job(struct util_queue *queue,
236 void *job,
237 struct util_queue_fence *fence,
238 util_queue_execute_func execute,
239 util_queue_execute_func cleanup,
240 const size_t job_size);
241 void util_queue_drop_job(struct util_queue *queue,
242 struct util_queue_fence *fence);
243
244 void util_queue_finish(struct util_queue *queue);
245
246 /* Adjust the number of active threads. The new number of threads can't be
247 * greater than the initial number of threads at the creation of the queue,
248 * and it can't be less than 1.
249 */
250 void
251 util_queue_adjust_num_threads(struct util_queue *queue, unsigned num_threads,
252 bool locked);
253
254 int64_t util_queue_get_thread_time_nano(struct util_queue *queue,
255 unsigned thread_index);
256
257 /* util_queue needs to be cleared to zeroes for this to work */
258 static inline bool
util_queue_is_initialized(struct util_queue * queue)259 util_queue_is_initialized(struct util_queue *queue)
260 {
261 return queue->threads != NULL;
262 }
263
264 /* Convenient structure for monitoring the queue externally and passing
265 * the structure between Mesa components. The queue doesn't use it directly.
266 */
267 struct util_queue_monitoring
268 {
269 /* For querying the thread busyness. */
270 struct util_queue *queue;
271
272 /* Counters updated by the user of the queue. */
273 unsigned num_offloaded_items;
274 unsigned num_direct_items;
275 unsigned num_syncs;
276 unsigned num_batches;
277 };
278
279 #ifdef __cplusplus
280 }
281 #endif
282
283 #endif
284