src/apply.c

   1 /*
   2  * Copyright (c) 2008-2013 Apple Inc. All rights reserved.
   3  *
   4  * @APPLE_APACHE_LICENSE_HEADER_START@
   5  *
   6  * Licensed under the Apache License, Version 2.0 (the "License");
   7  * you may not use this file except in compliance with the License.
   8  * You may obtain a copy of the License at
   9  *
  10  *     http://www.apache.org/licenses/LICENSE-2.0
  11  *
  12  * Unless required by applicable law or agreed to in writing, software
  13  * distributed under the License is distributed on an "AS IS" BASIS,
  14  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  15  * See the License for the specific language governing permissions and
  16  * limitations under the License.
  17  *
  18  * @APPLE_APACHE_LICENSE_HEADER_END@
  19  */
  20
  21 #include "internal.h"
  22
  23 typedef void (*dispatch_apply_function_t)(void *, size_t);
  24 static char const * const _dispatch_apply_key = "apply";
  25
  26 #define DISPATCH_APPLY_INVOKE_REDIRECT 0x1
  27 #define DISPATCH_APPLY_INVOKE_WAIT     0x2
  28
  29 DISPATCH_ALWAYS_INLINE
  30 static inline void
  31 _dispatch_apply_invoke2(void *ctxt, long invoke_flags)
  32 {
  33         dispatch_apply_t da = (dispatch_apply_t)ctxt;
  34         size_t const iter = da->da_iterations;
  35         size_t idx, done = 0;
  36
  37         idx = os_atomic_inc_orig2o(da, da_index, acquire);
  38         if (!fastpath(idx < iter)) goto out;
  39
  40         // da_dc is only safe to access once the 'index lock' has been acquired
  41         dispatch_apply_function_t const func = (void *)da->da_dc->dc_func;
  42         void *const da_ctxt = da->da_dc->dc_ctxt;
  43         dispatch_queue_t dq = da->da_dc->dc_data;
  44
  45         _dispatch_perfmon_workitem_dec(); // this unit executes many items
  46
  47         // Handle nested dispatch_apply rdar://problem/9294578
  48         dispatch_thread_context_s apply_ctxt = {
  49                 .dtc_key = _dispatch_apply_key,
  50                 .dtc_apply_nesting = da->da_nested,
  51         };
  52         _dispatch_thread_context_push(&apply_ctxt);
  53
  54         dispatch_thread_frame_s dtf;
  55         pthread_priority_t old_dp;
  56         if (invoke_flags & DISPATCH_APPLY_INVOKE_REDIRECT) {
  57                 _dispatch_thread_frame_push(&dtf, dq);
  58                 old_dp = _dispatch_set_defaultpriority(dq->dq_priority, NULL);
  59         }
  60         dispatch_invoke_flags_t flags = da->da_flags;
  61
  62         // Striding is the responsibility of the caller.
  63         do {
  64                 dispatch_invoke_with_autoreleasepool(flags, {
  65                         _dispatch_client_callout2(da_ctxt, idx, func);
  66                         _dispatch_perfmon_workitem_inc();
  67                         done++;
  68                         idx = os_atomic_inc_orig2o(da, da_index, relaxed);
  69                 });
  70         } while (fastpath(idx < iter));
  71
  72         if (invoke_flags & DISPATCH_APPLY_INVOKE_REDIRECT) {
  73                 _dispatch_reset_defaultpriority(old_dp);
  74                 _dispatch_thread_frame_pop(&dtf);
  75         }
  76
  77         _dispatch_thread_context_pop(&apply_ctxt);
  78
  79         // The thread that finished the last workitem wakes up the possibly waiting
  80         // thread that called dispatch_apply. They could be one and the same.
  81         if (!os_atomic_sub2o(da, da_todo, done, release)) {
  82                 _dispatch_thread_event_signal(&da->da_event);
  83         }
  84 out:
  85         if (invoke_flags & DISPATCH_APPLY_INVOKE_WAIT) {
  86                 _dispatch_thread_event_wait(&da->da_event);
  87                 _dispatch_thread_event_destroy(&da->da_event);
  88         }
  89         if (os_atomic_dec2o(da, da_thr_cnt, release) == 0) {
  90 #if DISPATCH_INTROSPECTION
  91                 _dispatch_continuation_free(da->da_dc);
  92 #endif
  93                 _dispatch_continuation_free((dispatch_continuation_t)da);
  94         }
  95 }
  96
  97 DISPATCH_NOINLINE
  98 void
  99 _dispatch_apply_invoke(void *ctxt)
 100 {
 101         _dispatch_apply_invoke2(ctxt, 0);
 102 }
 103
 104 DISPATCH_NOINLINE
 105 static void
 106 _dispatch_apply_invoke_and_wait(void *ctxt)
 107 {
 108         _dispatch_apply_invoke2(ctxt, DISPATCH_APPLY_INVOKE_WAIT);
 109         _dispatch_perfmon_workitem_inc();
 110 }
 111
 112 DISPATCH_NOINLINE
 113 void
 114 _dispatch_apply_redirect_invoke(void *ctxt)
 115 {
 116         _dispatch_apply_invoke2(ctxt, DISPATCH_APPLY_INVOKE_REDIRECT);
 117 }
 118
 119 DISPATCH_ALWAYS_INLINE
 120 static inline dispatch_invoke_flags_t
 121 _dispatch_apply_autorelease_frequency(dispatch_queue_t dq)
 122 {
 123         dispatch_invoke_flags_t qaf = 0;
 124
 125         while (dq && !qaf) {
 126                 qaf = _dispatch_queue_autorelease_frequency(dq);
 127                 dq = slowpath(dq->do_targetq);
 128         }
 129         return qaf;
 130 }
 131
 132 DISPATCH_NOINLINE
 133 static void
 134 _dispatch_apply_serial(void *ctxt)
 135 {
 136         dispatch_apply_t da = (dispatch_apply_t)ctxt;
 137         dispatch_continuation_t dc = da->da_dc;
 138         size_t const iter = da->da_iterations;
 139         dispatch_invoke_flags_t flags;
 140         size_t idx = 0;
 141
 142         _dispatch_perfmon_workitem_dec(); // this unit executes many items
 143         flags = _dispatch_apply_autorelease_frequency(dc->dc_data);
 144         do {
 145                 dispatch_invoke_with_autoreleasepool(flags, {
 146                         _dispatch_client_callout2(dc->dc_ctxt, idx, (void*)dc->dc_func);
 147                         _dispatch_perfmon_workitem_inc();
 148                 });
 149         } while (++idx < iter);
 150
 151 #if DISPATCH_INTROSPECTION
 152         _dispatch_continuation_free(da->da_dc);
 153 #endif
 154         _dispatch_continuation_free((dispatch_continuation_t)da);
 155 }
 156
 157 DISPATCH_ALWAYS_INLINE
 158 static inline void
 159 _dispatch_apply_f2(dispatch_queue_t dq, dispatch_apply_t da,
 160                 dispatch_function_t func)
 161 {
 162         uint32_t i = 0;
 163         dispatch_continuation_t head = NULL, tail = NULL;
 164
 165         // The current thread does not need a continuation
 166         uint32_t continuation_cnt = da->da_thr_cnt - 1;
 167
 168         dispatch_assert(continuation_cnt);
 169
 170         for (i = 0; i < continuation_cnt; i++) {
 171                 dispatch_continuation_t next = _dispatch_continuation_alloc();
 172                 uintptr_t dc_flags = DISPATCH_OBJ_CONSUME_BIT;
 173
 174                 _dispatch_continuation_init_f(next, dq, da, func, 0, 0, dc_flags);
 175                 next->do_next = head;
 176                 head = next;
 177
 178                 if (!tail) {
 179                         tail = next;
 180                 }
 181         }
 182
 183         _dispatch_thread_event_init(&da->da_event);
 184
 185         _dispatch_queue_push_list(dq, head, tail, head->dc_priority,
 186                         continuation_cnt);
 187         // Call the first element directly
 188         _dispatch_apply_invoke_and_wait(da);
 189 }
 190
 191 DISPATCH_NOINLINE
 192 static void
 193 _dispatch_apply_redirect(void *ctxt)
 194 {
 195         dispatch_apply_t da = (dispatch_apply_t)ctxt;
 196         uint32_t da_width = da->da_thr_cnt - 1;
 197         dispatch_queue_t dq = da->da_dc->dc_data, rq = dq, tq;
 198
 199         do {
 200                 uint32_t width = _dispatch_queue_try_reserve_apply_width(rq, da_width);
 201
 202                 if (slowpath(da_width > width)) {
 203                         uint32_t excess = da_width - width;
 204                         for (tq = dq; tq != rq; tq = tq->do_targetq) {
 205                                 _dispatch_queue_relinquish_width(tq, excess);
 206                         }
 207                         da_width -= excess;
 208                         if (slowpath(!da_width)) {
 209                                 return _dispatch_apply_serial(da);
 210                         }
 211                         da->da_thr_cnt -= excess;
 212                 }
 213                 if (!da->da_flags) {
 214                         // find first queue in descending target queue order that has
 215                         // an autorelease frequency set, and use that as the frequency for
 216                         // this continuation.
 217                         da->da_flags = _dispatch_queue_autorelease_frequency(dq);
 218                 }
 219                 rq = rq->do_targetq;
 220         } while (slowpath(rq->do_targetq));
 221         _dispatch_apply_f2(rq, da, _dispatch_apply_redirect_invoke);
 222         do {
 223                 _dispatch_queue_relinquish_width(dq, da_width);
 224                 dq = dq->do_targetq;
 225         } while (slowpath(dq->do_targetq));
 226 }
 227
 228 #define DISPATCH_APPLY_MAX UINT16_MAX // must be < sqrt(SIZE_MAX)
 229
 230 DISPATCH_NOINLINE
 231 void
 232 dispatch_apply_f(size_t iterations, dispatch_queue_t dq, void *ctxt,
 233                 void (*func)(void *, size_t))
 234 {
 235         if (slowpath(iterations == 0)) {
 236                 return;
 237         }
 238         uint32_t thr_cnt = dispatch_hw_config(active_cpus);
 239         dispatch_thread_context_t dtctxt = _dispatch_thread_context_find(_dispatch_apply_key);
 240         size_t nested = dtctxt ? dtctxt->dtc_apply_nesting : 0;
 241         dispatch_queue_t old_dq = _dispatch_queue_get_current();
 242
 243         if (!slowpath(nested)) {
 244                 nested = iterations;
 245         } else {
 246                 thr_cnt = nested < thr_cnt ? thr_cnt / nested : 1;
 247                 nested = nested < DISPATCH_APPLY_MAX && iterations < DISPATCH_APPLY_MAX
 248                                 ? nested * iterations : DISPATCH_APPLY_MAX;
 249         }
 250         if (iterations < thr_cnt) {
 251                 thr_cnt = (uint32_t)iterations;
 252         }
 253         if (slowpath(dq == DISPATCH_APPLY_CURRENT_ROOT_QUEUE)) {
 254                 dq = old_dq ? old_dq : _dispatch_get_root_queue(
 255                                 _DISPATCH_QOS_CLASS_DEFAULT, false);
 256                 while (slowpath(dq->do_targetq)) {
 257                         dq = dq->do_targetq;
 258                 }
 259         }
 260         struct dispatch_continuation_s dc = {
 261                 .dc_func = (void*)func,
 262                 .dc_ctxt = ctxt,
 263                 .dc_data = dq,
 264         };
 265         dispatch_apply_t da = (typeof(da))_dispatch_continuation_alloc();
 266         da->da_index = 0;
 267         da->da_todo = iterations;
 268         da->da_iterations = iterations;
 269         da->da_nested = nested;
 270         da->da_thr_cnt = thr_cnt;
 271 #if DISPATCH_INTROSPECTION
 272         da->da_dc = _dispatch_continuation_alloc();
 273         *da->da_dc = dc;
 274 #else
 275         da->da_dc = &dc;
 276 #endif
 277         da->da_flags = 0;
 278
 279         if (slowpath(dq->dq_width == 1) || slowpath(thr_cnt <= 1)) {
 280                 return dispatch_sync_f(dq, da, _dispatch_apply_serial);
 281         }
 282         if (slowpath(dq->do_targetq)) {
 283                 if (slowpath(dq == old_dq)) {
 284                         return dispatch_sync_f(dq, da, _dispatch_apply_serial);
 285                 } else {
 286                         return dispatch_sync_f(dq, da, _dispatch_apply_redirect);
 287                 }
 288         }
 289
 290         dispatch_thread_frame_s dtf;
 291         _dispatch_thread_frame_push(&dtf, dq);
 292         _dispatch_apply_f2(dq, da, _dispatch_apply_invoke);
 293         _dispatch_thread_frame_pop(&dtf);
 294 }
 295
 296 #ifdef __BLOCKS__
 297 void
 298 dispatch_apply(size_t iterations, dispatch_queue_t dq, void (^work)(size_t))
 299 {
 300         dispatch_apply_f(iterations, dq, work,
 301                         (dispatch_apply_function_t)_dispatch_Block_invoke(work));
 302 }
 303 #endif
 304
 305 #if 0
 306 #ifdef __BLOCKS__
 307 void
 308 dispatch_stride(size_t offset, size_t stride, size_t iterations,
 309                 dispatch_queue_t dq, void (^work)(size_t))
 310 {
 311         dispatch_stride_f(offset, stride, iterations, dq, work,
 312                         (dispatch_apply_function_t)_dispatch_Block_invoke(work));
 313 }
 314 #endif
 315
 316 DISPATCH_NOINLINE
 317 void
 318 dispatch_stride_f(size_t offset, size_t stride, size_t iterations,
 319                 dispatch_queue_t dq, void *ctxt, void (*func)(void *, size_t))
 320 {
 321         if (stride == 0) {
 322                 stride = 1;
 323         }
 324         dispatch_apply(iterations / stride, queue, ^(size_t idx) {
 325                 size_t i = idx * stride + offset;
 326                 size_t stop = i + stride;
 327                 do {
 328                         func(ctxt, i++);
 329                 } while (i < stop);
 330         });
 331
 332         dispatch_sync(queue, ^{
 333                 size_t i;
 334                 for (i = iterations - (iterations % stride); i < iterations; i++) {
 335                         func(ctxt, i + offset);
 336                 }
 337         });
 338 }
 339 #endif