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