src/util/u_queue.h

   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/u_atomic.h"
  42 #include "util/u_thread.h"
  43
  44 #ifdef __cplusplus
  45 extern "C" {
  46 #endif
  47
  48 #define UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY      (1 << 0)
  49 #define UTIL_QUEUE_INIT_RESIZE_IF_FULL            (1 << 1)
  50
  51 #if defined(__GNUC__) && defined(HAVE_FUTEX)
  52 #define UTIL_QUEUE_FENCE_FUTEX
  53 #else
  54 #define UTIL_QUEUE_FENCE_STANDARD
  55 #endif
  56
  57 #ifdef UTIL_QUEUE_FENCE_FUTEX
  58 /* Job completion fence.
  59  * Put this into your job structure.
  60  */
  61 struct util_queue_fence {
  62    /* The fence can be in one of three states:
  63     *  0 - signaled
  64     *  1 - unsignaled
  65     *  2 - unsignaled, may have waiters
  66     */
  67    uint32_t val;
  68 };
  69
  70 static inline void
  71 util_queue_fence_init(struct util_queue_fence *fence)
  72 {
  73    fence->val = 0;
  74 }
  75
  76 static inline void
  77 util_queue_fence_destroy(struct util_queue_fence *fence)
  78 {
  79    assert(fence->val == 0);
  80    /* no-op */
  81 }
  82
  83 static inline void
  84 util_queue_fence_signal(struct util_queue_fence *fence)
  85 {
  86    uint32_t val = p_atomic_xchg(&fence->val, 0);
  87
  88    assert(val != 0);
  89
  90    if (val == 2)
  91       futex_wake(&fence->val, INT_MAX);
  92 }
  93
  94 /**
  95  * Move \p fence back into unsignalled state.
  96  *
  97  * \warning The caller must ensure that no other thread may currently be
  98  *          waiting (or about to wait) on the fence.
  99  */
 100 static inline void
 101 util_queue_fence_reset(struct util_queue_fence *fence)
 102 {
 103 #ifdef NDEBUG
 104    fence->val = 1;
 105 #else
 106    uint32_t v = p_atomic_xchg(&fence->val, 1);
 107    assert(v == 0);
 108 #endif
 109 }
 110
 111 static inline bool
 112 util_queue_fence_is_signalled(struct util_queue_fence *fence)
 113 {
 114    return fence->val == 0;
 115 }
 116 #endif
 117
 118 #ifdef UTIL_QUEUE_FENCE_STANDARD
 119 /* Job completion fence.
 120  * Put this into your job structure.
 121  */
 122 struct util_queue_fence {
 123    mtx_t mutex;
 124    cnd_t cond;
 125    int signalled;
 126 };
 127
 128 void util_queue_fence_init(struct util_queue_fence *fence);
 129 void util_queue_fence_destroy(struct util_queue_fence *fence);
 130 void util_queue_fence_signal(struct util_queue_fence *fence);
 131
 132 /**
 133  * Move \p fence back into unsignalled state.
 134  *
 135  * \warning The caller must ensure that no other thread may currently be
 136  *          waiting (or about to wait) on the fence.
 137  */
 138 static inline void
 139 util_queue_fence_reset(struct util_queue_fence *fence)
 140 {
 141    assert(fence->signalled);
 142    fence->signalled = 0;
 143 }
 144
 145 static inline bool
 146 util_queue_fence_is_signalled(struct util_queue_fence *fence)
 147 {
 148    return fence->signalled != 0;
 149 }
 150 #endif
 151
 152 void
 153 _util_queue_fence_wait(struct util_queue_fence *fence);
 154
 155 static inline void
 156 util_queue_fence_wait(struct util_queue_fence *fence)
 157 {
 158    if (unlikely(!util_queue_fence_is_signalled(fence)))
 159       _util_queue_fence_wait(fence);
 160 }
 161
 162 bool
 163 _util_queue_fence_wait_timeout(struct util_queue_fence *fence,
 164                                int64_t abs_timeout);
 165
 166 /**
 167  * Wait for the fence to be signaled with a timeout.
 168  *
 169  * \param fence the fence
 170  * \param abs_timeout the absolute timeout in nanoseconds, relative to the
 171  *                    clock provided by os_time_get_nano.
 172  *
 173  * \return true if the fence was signaled, false if the timeout occurred.
 174  */
 175 static inline bool
 176 util_queue_fence_wait_timeout(struct util_queue_fence *fence,
 177                               int64_t abs_timeout)
 178 {
 179    if (util_queue_fence_is_signalled(fence))
 180       return true;
 181
 182    return _util_queue_fence_wait_timeout(fence, abs_timeout);
 183 }
 184
 185 typedef void (*util_queue_execute_func)(void *job, int thread_index);
 186
 187 struct util_queue_job {
 188    void *job;
 189    struct util_queue_fence *fence;
 190    util_queue_execute_func execute;
 191    util_queue_execute_func cleanup;
 192 };
 193
 194 /* Put this into your context. */
 195 struct util_queue {
 196    const char *name;
 197    mtx_t lock;
 198    cnd_t has_queued_cond;
 199    cnd_t has_space_cond;
 200    thrd_t *threads;
 201    unsigned flags;
 202    int num_queued;
 203    unsigned num_threads;
 204    int kill_threads;
 205    int max_jobs;
 206    int write_idx, read_idx; /* ring buffer pointers */
 207    struct util_queue_job *jobs;
 208
 209    /* for cleanup at exit(), protected by exit_mutex */
 210    struct list_head head;
 211 };
 212
 213 bool util_queue_init(struct util_queue *queue,
 214                      const char *name,
 215                      unsigned max_jobs,
 216                      unsigned num_threads,
 217                      unsigned flags);
 218 void util_queue_destroy(struct util_queue *queue);
 219
 220 /* optional cleanup callback is called after fence is signaled: */
 221 void util_queue_add_job(struct util_queue *queue,
 222                         void *job,
 223                         struct util_queue_fence *fence,
 224                         util_queue_execute_func execute,
 225                         util_queue_execute_func cleanup);
 226 void util_queue_drop_job(struct util_queue *queue,
 227                          struct util_queue_fence *fence);
 228
 229 void util_queue_finish(struct util_queue *queue);
 230
 231 int64_t util_queue_get_thread_time_nano(struct util_queue *queue,
 232                                         unsigned thread_index);
 233
 234 /* util_queue needs to be cleared to zeroes for this to work */
 235 static inline bool
 236 util_queue_is_initialized(struct util_queue *queue)
 237 {
 238    return queue->threads != NULL;
 239 }
 240
 241 /* Convenient structure for monitoring the queue externally and passing
 242  * the structure between Mesa components. The queue doesn't use it directly.
 243  */
 244 struct util_queue_monitoring
 245 {
 246    /* For querying the thread busyness. */
 247    struct util_queue *queue;
 248
 249    /* Counters updated by the user of the queue. */
 250    unsigned num_offloaded_items;
 251    unsigned num_direct_items;
 252    unsigned num_syncs;
 253 };
 254
 255 #ifdef __cplusplus
 256 }
 257 #endif
 258
 259 #endif