src/mesa/main/glthread.c

   1 /*
   2  * Copyright © 2012 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 /** @file glthread.c
  25  *
  26  * Support functions for the glthread feature of Mesa.
  27  *
  28  * In multicore systems, many applications end up CPU-bound with about half
  29  * their time spent inside their rendering thread and half inside Mesa.  To
  30  * alleviate this, we put a shim layer in Mesa at the GL dispatch level that
  31  * quickly logs the GL commands to a buffer to be processed by a worker
  32  * thread.
  33  */
  34
  35 #include "main/mtypes.h"
  36 #include "main/glthread.h"
  37 #include "main/marshal.h"
  38 #include "main/marshal_generated.h"
  39 #include "util/u_thread.h"
  40
  41 #ifdef HAVE_PTHREAD
  42
  43 static void
  44 glthread_allocate_batch(struct gl_context *ctx)
  45 {
  46    struct glthread_state *glthread = ctx->GLThread;
  47
  48    /* TODO: handle memory allocation failure. */
  49    glthread->batch = malloc(sizeof(*glthread->batch));
  50    if (!glthread->batch)
  51       return;
  52    memset(glthread->batch, 0, offsetof(struct glthread_batch, buffer));
  53 }
  54
  55 static void
  56 glthread_unmarshal_batch(struct gl_context *ctx, struct glthread_batch *batch)
  57 {
  58    size_t pos = 0;
  59
  60    _glapi_set_dispatch(ctx->CurrentServerDispatch);
  61
  62    while (pos < batch->used)
  63       pos += _mesa_unmarshal_dispatch_cmd(ctx, &batch->buffer[pos]);
  64
  65    assert(pos == batch->used);
  66
  67    free(batch);
  68 }
  69
  70 static void *
  71 glthread_worker(void *data)
  72 {
  73    struct gl_context *ctx = data;
  74    struct glthread_state *glthread = ctx->GLThread;
  75
  76    ctx->Driver.SetBackgroundContext(ctx);
  77    _glapi_set_context(ctx);
  78
  79    u_thread_setname("mesa_glthread");
  80
  81    pthread_mutex_lock(&glthread->mutex);
  82
  83    while (true) {
  84       struct glthread_batch *batch;
  85
  86       /* Block (dropping the lock) until new work arrives for us. */
  87       while (!glthread->batch_queue && !glthread->shutdown) {
  88          pthread_cond_broadcast(&glthread->work_done);
  89          pthread_cond_wait(&glthread->new_work, &glthread->mutex);
  90       }
  91
  92       batch = glthread->batch_queue;
  93
  94       if (glthread->shutdown && !batch) {
  95          pthread_cond_broadcast(&glthread->work_done);
  96          pthread_mutex_unlock(&glthread->mutex);
  97          return NULL;
  98       }
  99       glthread->batch_queue = batch->next;
 100       if (glthread->batch_queue_tail == &batch->next)
 101          glthread->batch_queue_tail = &glthread->batch_queue;
 102
 103       glthread->busy = true;
 104       pthread_mutex_unlock(&glthread->mutex);
 105
 106       glthread_unmarshal_batch(ctx, batch);
 107
 108       pthread_mutex_lock(&glthread->mutex);
 109       glthread->busy = false;
 110    }
 111
 112    /* UNREACHED */
 113    return NULL;
 114 }
 115
 116 void
 117 _mesa_glthread_init(struct gl_context *ctx)
 118 {
 119    struct glthread_state *glthread = calloc(1, sizeof(*glthread));
 120
 121    if (!glthread)
 122       return;
 123
 124    ctx->MarshalExec = _mesa_create_marshal_table(ctx);
 125    if (!ctx->MarshalExec) {
 126       free(glthread);
 127       return;
 128    }
 129
 130    ctx->CurrentClientDispatch = ctx->MarshalExec;
 131
 132    pthread_mutex_init(&glthread->mutex, NULL);
 133    pthread_cond_init(&glthread->new_work, NULL);
 134    pthread_cond_init(&glthread->work_done, NULL);
 135
 136    glthread->batch_queue_tail = &glthread->batch_queue;
 137    ctx->GLThread = glthread;
 138
 139    glthread_allocate_batch(ctx);
 140
 141    pthread_create(&glthread->thread, NULL, glthread_worker, ctx);
 142 }
 143
 144 void
 145 _mesa_glthread_destroy(struct gl_context *ctx)
 146 {
 147    struct glthread_state *glthread = ctx->GLThread;
 148
 149    if (!glthread)
 150       return;
 151
 152    _mesa_glthread_flush_batch(ctx);
 153
 154    pthread_mutex_lock(&glthread->mutex);
 155    glthread->shutdown = true;
 156    pthread_cond_broadcast(&glthread->new_work);
 157    pthread_mutex_unlock(&glthread->mutex);
 158
 159    /* Since this waits for the thread to exit, it means that all queued work
 160     * will have been completed.
 161     */
 162    pthread_join(glthread->thread, NULL);
 163
 164    pthread_cond_destroy(&glthread->new_work);
 165    pthread_cond_destroy(&glthread->work_done);
 166    pthread_mutex_destroy(&glthread->mutex);
 167
 168    /* Due to the join above, there should be one empty batch allocated at this
 169     * point, and no batches queued.
 170     */
 171    assert(!glthread->batch->used);
 172    assert(!glthread->batch->next);
 173    free(glthread->batch);
 174    assert(!glthread->batch_queue);
 175
 176    free(glthread);
 177    ctx->GLThread = NULL;
 178
 179    _mesa_glthread_restore_dispatch(ctx);
 180 }
 181
 182 void
 183 _mesa_glthread_restore_dispatch(struct gl_context *ctx)
 184 {
 185    /* Remove ourselves from the dispatch table except if another ctx/thread
 186     * already installed a new dispatch table.
 187     *
 188     * Typically glxMakeCurrent will bind a new context (install new table) then
 189     * old context might be deleted.
 190     */
 191    if (_glapi_get_dispatch() == ctx->MarshalExec) {
 192        ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
 193        _glapi_set_dispatch(ctx->CurrentClientDispatch);
 194    }
 195 }
 196
 197 void
 198 _mesa_glthread_flush_batch(struct gl_context *ctx)
 199 {
 200    struct glthread_state *glthread = ctx->GLThread;
 201    struct glthread_batch *batch;
 202
 203    if (!glthread)
 204       return;
 205
 206    batch = glthread->batch;
 207    if (!batch->used)
 208       return;
 209
 210    /* Immediately reallocate a new batch, since the next marshalled call would
 211     * just do it.
 212     */
 213    glthread_allocate_batch(ctx);
 214
 215    /* Debug: execute the batch immediately from this thread.
 216     *
 217     * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
 218     * need to restore it when it returns.
 219     */
 220    if (false) {
 221       glthread_unmarshal_batch(ctx, batch);
 222       _glapi_set_dispatch(ctx->CurrentClientDispatch);
 223       return;
 224    }
 225
 226    pthread_mutex_lock(&glthread->mutex);
 227    *glthread->batch_queue_tail = batch;
 228    glthread->batch_queue_tail = &batch->next;
 229    pthread_cond_broadcast(&glthread->new_work);
 230    pthread_mutex_unlock(&glthread->mutex);
 231 }
 232
 233 /**
 234  * Waits for all pending batches have been unmarshaled.
 235  *
 236  * This can be used by the main thread to synchronize access to the context,
 237  * since the worker thread will be idle after this.
 238  */
 239 void
 240 _mesa_glthread_finish(struct gl_context *ctx)
 241 {
 242    struct glthread_state *glthread = ctx->GLThread;
 243
 244    if (!glthread)
 245       return;
 246
 247    /* If this is called from the worker thread, then we've hit a path that
 248     * might be called from either the main thread or the worker (such as some
 249     * dri interface entrypoints), in which case we don't need to actually
 250     * synchronize against ourself.
 251     */
 252    if (pthread_self() == glthread->thread)
 253       return;
 254
 255    _mesa_glthread_flush_batch(ctx);
 256
 257    pthread_mutex_lock(&glthread->mutex);
 258
 259    while (glthread->batch_queue || glthread->busy)
 260       pthread_cond_wait(&glthread->work_done, &glthread->mutex);
 261
 262    pthread_mutex_unlock(&glthread->mutex);
 263 }
 264
 265 #endif