osfmk/kern/mpsc_queue.c

   1 /*
   2  * Copyright (c) 2018 Apple Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28
  29 #include <machine/machine_cpu.h>
  30 #include <kern/locks.h>
  31 #include <kern/mpsc_queue.h>
  32 #include <kern/thread.h>
  33
  34 #pragma mark Single Consumer calls
  35
  36 __attribute__((noinline))
  37 static mpsc_queue_chain_t
  38 _mpsc_queue_wait_for_enqueuer(struct mpsc_queue_chain *_Atomic *ptr)
  39 {
  40         return hw_wait_while_equals((void **)ptr, NULL);
  41 }
  42
  43 void
  44 mpsc_queue_restore_batch(mpsc_queue_head_t q, mpsc_queue_chain_t first,
  45     mpsc_queue_chain_t last)
  46 {
  47         mpsc_queue_chain_t head = os_atomic_load(&q->mpqh_head.mpqc_next, relaxed);
  48
  49         os_atomic_store(&last->mpqc_next, head, relaxed);
  50
  51         if (head == NULL &&
  52             !os_atomic_cmpxchg(&q->mpqh_tail, &q->mpqh_head, last, release)) {
  53                 head = os_atomic_load(&q->mpqh_head.mpqc_next, relaxed);
  54                 if (__improbable(head == NULL)) {
  55                         head = _mpsc_queue_wait_for_enqueuer(&q->mpqh_head.mpqc_next);
  56                 }
  57                 os_atomic_store(&last->mpqc_next, head, relaxed);
  58         }
  59
  60         os_atomic_store(&q->mpqh_head.mpqc_next, first, relaxed);
  61 }
  62
  63 mpsc_queue_chain_t
  64 mpsc_queue_dequeue_batch(mpsc_queue_head_t q, mpsc_queue_chain_t *tail_out,
  65     os_atomic_dependency_t dependency)
  66 {
  67         mpsc_queue_chain_t head, tail;
  68
  69         q = os_atomic_inject_dependency(q, dependency);
  70
  71         tail = os_atomic_load(&q->mpqh_tail, relaxed);
  72         if (__improbable(tail == &q->mpqh_head)) {
  73                 *tail_out = NULL;
  74                 return NULL;
  75         }
  76
  77         head = os_atomic_load(&q->mpqh_head.mpqc_next, relaxed);
  78         if (__improbable(head == NULL)) {
  79                 head = _mpsc_queue_wait_for_enqueuer(&q->mpqh_head.mpqc_next);
  80         }
  81         os_atomic_store(&q->mpqh_head.mpqc_next, NULL, relaxed);
  82         /*
  83          * 22708742: set tail to &q->mpqh_head with release, so that NULL write
  84          * to head above doesn't clobber the head set by concurrent enqueuer
  85          *
  86          * The other half of the seq_cst is required to pair with any enqueuer that
  87          * contributed to an element in this list (pairs with the release fence in
  88          * __mpsc_queue_append_update_tail().
  89          *
  90          * Making this seq_cst instead of acq_rel makes mpsc_queue_append*()
  91          * visibility transitive (when items hop from one queue to the next)
  92          * which is expected by clients implicitly.
  93          *
  94          * Note that this is the same number of fences that a traditional lock
  95          * would have, but as a once-per-batch cost.
  96          */
  97         *tail_out = os_atomic_xchg(&q->mpqh_tail, &q->mpqh_head, seq_cst);
  98
  99         return head;
 100 }
 101
 102 mpsc_queue_chain_t
 103 mpsc_queue_batch_next(mpsc_queue_chain_t cur, mpsc_queue_chain_t tail)
 104 {
 105         mpsc_queue_chain_t elm = NULL;
 106         if (cur == tail || cur == NULL) {
 107                 return elm;
 108         }
 109
 110         elm = os_atomic_load(&cur->mpqc_next, relaxed);
 111         if (__improbable(elm == NULL)) {
 112                 elm = _mpsc_queue_wait_for_enqueuer(&cur->mpqc_next);
 113         }
 114         return elm;
 115 }
 116
 117 #pragma mark "GCD"-like facilities
 118
 119 static void _mpsc_daemon_queue_drain(mpsc_daemon_queue_t, thread_t);
 120 static void _mpsc_daemon_queue_enqueue(mpsc_daemon_queue_t, mpsc_queue_chain_t);
 121
 122 /* thread based queues */
 123
 124 static void
 125 _mpsc_queue_thread_continue(void *param, wait_result_t wr __unused)
 126 {
 127         mpsc_daemon_queue_t dq = param;
 128
 129         assert(dq->mpd_thread == current_thread());
 130         _mpsc_daemon_queue_drain(dq, dq->mpd_thread);
 131         thread_block_parameter(_mpsc_queue_thread_continue, dq);
 132 }
 133
 134 static void
 135 _mpsc_queue_thread_wakeup(mpsc_daemon_queue_t dq)
 136 {
 137         thread_wakeup_thread((event_t)dq, dq->mpd_thread);
 138 }
 139
 140 static kern_return_t
 141 _mpsc_daemon_queue_init_with_thread(mpsc_daemon_queue_t dq,
 142     mpsc_daemon_invoke_fn_t invoke, int pri, const char *name,
 143     mpsc_daemon_queue_kind_t kind)
 144 {
 145         kern_return_t kr;
 146
 147         *dq = (struct mpsc_daemon_queue){
 148                 .mpd_kind   = kind,
 149                 .mpd_invoke = invoke,
 150                 .mpd_queue  = MPSC_QUEUE_INITIALIZER(dq->mpd_queue),
 151                 .mpd_chain  = { MPSC_QUEUE_NOTQUEUED_MARKER },
 152         };
 153
 154         kr = kernel_thread_create(_mpsc_queue_thread_continue, dq, pri,
 155             &dq->mpd_thread);
 156         if (kr == KERN_SUCCESS) {
 157                 thread_set_thread_name(dq->mpd_thread, name);
 158                 thread_start_in_assert_wait(dq->mpd_thread, (event_t)dq, THREAD_UNINT);
 159                 thread_deallocate(dq->mpd_thread);
 160         }
 161         return kr;
 162 }
 163
 164 kern_return_t
 165 mpsc_daemon_queue_init_with_thread(mpsc_daemon_queue_t dq,
 166     mpsc_daemon_invoke_fn_t invoke, int pri, const char *name)
 167 {
 168         return _mpsc_daemon_queue_init_with_thread(dq, invoke, pri, name,
 169                    MPSC_QUEUE_KIND_THREAD);
 170 }
 171
 172 /* thread-call based queues */
 173
 174 static void
 175 _mpsc_queue_thread_call_drain(thread_call_param_t arg0,
 176     thread_call_param_t arg1 __unused)
 177 {
 178         _mpsc_daemon_queue_drain((mpsc_daemon_queue_t)arg0, NULL);
 179 }
 180
 181 static void
 182 _mpsc_queue_thread_call_wakeup(mpsc_daemon_queue_t dq)
 183 {
 184         thread_call_enter(dq->mpd_call);
 185 }
 186
 187 void
 188 mpsc_daemon_queue_init_with_thread_call(mpsc_daemon_queue_t dq,
 189     mpsc_daemon_invoke_fn_t invoke, thread_call_priority_t pri)
 190 {
 191         *dq = (struct mpsc_daemon_queue){
 192                 .mpd_kind   = MPSC_QUEUE_KIND_THREAD_CALL,
 193                 .mpd_invoke = invoke,
 194                 .mpd_queue  = MPSC_QUEUE_INITIALIZER(dq->mpd_queue),
 195                 .mpd_chain  = { MPSC_QUEUE_NOTQUEUED_MARKER },
 196         };
 197         dq->mpd_call = thread_call_allocate_with_options(
 198                 _mpsc_queue_thread_call_drain, dq, pri, THREAD_CALL_OPTIONS_ONCE);
 199 }
 200
 201 /* nested queues */
 202
 203 void
 204 mpsc_daemon_queue_nested_invoke(mpsc_queue_chain_t elm,
 205     __unused mpsc_daemon_queue_t tq)
 206 {
 207         mpsc_daemon_queue_t dq;
 208         dq = mpsc_queue_element(elm, struct mpsc_daemon_queue, mpd_chain);
 209         _mpsc_daemon_queue_drain(dq, NULL);
 210 }
 211
 212 static void
 213 _mpsc_daemon_queue_nested_wakeup(mpsc_daemon_queue_t dq)
 214 {
 215         _mpsc_daemon_queue_enqueue(dq->mpd_target, &dq->mpd_chain);
 216 }
 217
 218 void
 219 mpsc_daemon_queue_init_with_target(mpsc_daemon_queue_t dq,
 220     mpsc_daemon_invoke_fn_t invoke, mpsc_daemon_queue_t target)
 221 {
 222         *dq = (struct mpsc_daemon_queue){
 223                 .mpd_kind   = MPSC_QUEUE_KIND_NESTED,
 224                 .mpd_invoke = invoke,
 225                 .mpd_target = target,
 226                 .mpd_queue  = MPSC_QUEUE_INITIALIZER(dq->mpd_queue),
 227                 .mpd_chain  = { MPSC_QUEUE_NOTQUEUED_MARKER },
 228         };
 229 }
 230
 231 /* enqueue, drain & cancelation */
 232
 233 static void
 234 _mpsc_daemon_queue_drain(mpsc_daemon_queue_t dq, thread_t self)
 235 {
 236         mpsc_daemon_invoke_fn_t invoke = dq->mpd_invoke;
 237         mpsc_daemon_queue_kind_t kind = dq->mpd_kind;
 238         mpsc_queue_chain_t head, cur, tail;
 239         mpsc_daemon_queue_state_t st;
 240
 241         if (kind == MPSC_QUEUE_KIND_THREAD_CRITICAL) {
 242                 self->options |= TH_OPT_SYSTEM_CRITICAL;
 243         }
 244
 245 again:
 246         /*
 247          * Most of the time we're woken up because we're dirty,
 248          * This atomic xor sets DRAINING and clears WAKEUP in a single atomic
 249          * in that case.
 250          *
 251          * However, if we're woken up for cancelation, the state may be reduced to
 252          * the CANCELED bit set only, and then the xor will actually set WAKEUP.
 253          * We need to correct this and clear it back to avoid looping below.
 254          * This is safe to do as no one is allowed to enqueue more work after
 255          * cancelation has happened.
 256          *
 257          * We use `st` as a dependency token to pair with the release fence in
 258          * _mpsc_daemon_queue_enqueue() which gives us the guarantee that the update
 259          * to the tail of the MPSC queue that made it non empty is visible to us.
 260          */
 261         st = os_atomic_xor(&dq->mpd_state,
 262             MPSC_QUEUE_STATE_DRAINING | MPSC_QUEUE_STATE_WAKEUP, dependency);
 263         assert(st & MPSC_QUEUE_STATE_DRAINING);
 264         if (__improbable(st & MPSC_QUEUE_STATE_WAKEUP)) {
 265                 assert(st & MPSC_QUEUE_STATE_CANCELED);
 266                 os_atomic_andnot(&dq->mpd_state, MPSC_QUEUE_STATE_WAKEUP, relaxed);
 267         }
 268
 269         os_atomic_dependency_t dep = os_atomic_make_dependency((uintptr_t)st);
 270         while ((head = mpsc_queue_dequeue_batch(&dq->mpd_queue, &tail, dep))) {
 271                 mpsc_queue_batch_foreach_safe(cur, head, tail) {
 272                         os_atomic_store(&cur->mpqc_next,
 273                             MPSC_QUEUE_NOTQUEUED_MARKER, relaxed);
 274                         invoke(cur, dq);
 275                 }
 276         }
 277
 278         if (self) {
 279                 assert_wait((event_t)dq, THREAD_UNINT);
 280         }
 281
 282         /*
 283          * Unlike GCD no fence is necessary here: there is no concept similar
 284          * to "dispatch_sync()" that would require changes this thread made to be
 285          * visible to other threads as part of the mpsc_daemon_queue machinery.
 286          *
 287          * Making updates that happened on the daemon queue visible to other threads
 288          * is the responsibility of the client.
 289          */
 290         st = os_atomic_andnot(&dq->mpd_state, MPSC_QUEUE_STATE_DRAINING, relaxed);
 291
 292         /*
 293          * A wakeup has happened while we were draining,
 294          * which means that the queue did an [ empty -> non empty ]
 295          * transition during our drain.
 296          *
 297          * Chances are we already observed and drained everything,
 298          * but we need to be absolutely sure, so start a drain again
 299          * as the enqueuer observed the DRAINING bit and has skipped calling
 300          * _mpsc_daemon_queue_wakeup().
 301          */
 302         if (__improbable(st & MPSC_QUEUE_STATE_WAKEUP)) {
 303                 if (self) {
 304                         clear_wait(self, THREAD_AWAKENED);
 305                 }
 306                 goto again;
 307         }
 308
 309         /* dereferencing `dq` past this point is unsafe */
 310
 311         if (kind == MPSC_QUEUE_KIND_THREAD_CRITICAL) {
 312                 self->options &= ~TH_OPT_SYSTEM_CRITICAL;
 313         }
 314
 315         if (__improbable(st & MPSC_QUEUE_STATE_CANCELED)) {
 316                 thread_wakeup(&dq->mpd_state);
 317                 if (self) {
 318                         clear_wait(self, THREAD_AWAKENED);
 319                         thread_terminate_self();
 320                         __builtin_unreachable();
 321                 }
 322         }
 323 }
 324
 325 static void
 326 _mpsc_daemon_queue_wakeup(mpsc_daemon_queue_t dq)
 327 {
 328         switch (dq->mpd_kind) {
 329         case MPSC_QUEUE_KIND_NESTED:
 330                 _mpsc_daemon_queue_nested_wakeup(dq);
 331                 break;
 332         case MPSC_QUEUE_KIND_THREAD:
 333         case MPSC_QUEUE_KIND_THREAD_CRITICAL:
 334                 _mpsc_queue_thread_wakeup(dq);
 335                 break;
 336         case MPSC_QUEUE_KIND_THREAD_CALL:
 337                 _mpsc_queue_thread_call_wakeup(dq);
 338                 break;
 339         default:
 340                 panic("mpsc_queue[%p]: invalid kind (%d)", dq, dq->mpd_kind);
 341         }
 342 }
 343
 344 static void
 345 _mpsc_daemon_queue_enqueue(mpsc_daemon_queue_t dq, mpsc_queue_chain_t elm)
 346 {
 347         mpsc_daemon_queue_state_t st;
 348
 349         if (mpsc_queue_append(&dq->mpd_queue, elm)) {
 350                 /*
 351                  * Pairs with the acquire fence in _mpsc_daemon_queue_drain().
 352                  */
 353                 st = os_atomic_or_orig(&dq->mpd_state, MPSC_QUEUE_STATE_WAKEUP, release);
 354                 if (__improbable(st & MPSC_QUEUE_STATE_CANCELED)) {
 355                         panic("mpsc_queue[%p]: use after cancelation", dq);
 356                 }
 357
 358                 if ((st & (MPSC_QUEUE_STATE_DRAINING | MPSC_QUEUE_STATE_WAKEUP)) == 0) {
 359                         _mpsc_daemon_queue_wakeup(dq);
 360                 }
 361         }
 362 }
 363
 364 void
 365 mpsc_daemon_enqueue(mpsc_daemon_queue_t dq, mpsc_queue_chain_t elm,
 366     mpsc_queue_options_t options)
 367 {
 368         if (options & MPSC_QUEUE_DISABLE_PREEMPTION) {
 369                 disable_preemption();
 370         }
 371
 372         _mpsc_daemon_queue_enqueue(dq, elm);
 373
 374         if (options & MPSC_QUEUE_DISABLE_PREEMPTION) {
 375                 enable_preemption();
 376         }
 377 }
 378
 379 void
 380 mpsc_daemon_queue_cancel_and_wait(mpsc_daemon_queue_t dq)
 381 {
 382         mpsc_daemon_queue_state_t st;
 383
 384         assert_wait((event_t)&dq->mpd_state, THREAD_UNINT);
 385
 386         st = os_atomic_or_orig(&dq->mpd_state, MPSC_QUEUE_STATE_CANCELED, relaxed);
 387         if (__improbable(st & MPSC_QUEUE_STATE_CANCELED)) {
 388                 panic("mpsc_queue[%p]: cancelled twice (%x)", dq, st);
 389         }
 390
 391         if (dq->mpd_kind == MPSC_QUEUE_KIND_NESTED && st == 0) {
 392                 clear_wait(current_thread(), THREAD_AWAKENED);
 393         } else {
 394                 disable_preemption();
 395                 _mpsc_daemon_queue_wakeup(dq);
 396                 enable_preemption();
 397                 thread_block(THREAD_CONTINUE_NULL);
 398         }
 399
 400         switch (dq->mpd_kind) {
 401         case MPSC_QUEUE_KIND_NESTED:
 402                 dq->mpd_target = NULL;
 403                 break;
 404         case MPSC_QUEUE_KIND_THREAD:
 405         case MPSC_QUEUE_KIND_THREAD_CRITICAL:
 406                 dq->mpd_thread = NULL;
 407                 break;
 408         case MPSC_QUEUE_KIND_THREAD_CALL:
 409                 thread_call_cancel_wait(dq->mpd_call);
 410                 thread_call_free(dq->mpd_call);
 411                 dq->mpd_call = NULL;
 412                 break;
 413         default:
 414                 panic("mpsc_queue[%p]: invalid kind (%d)", dq, dq->mpd_kind);
 415         }
 416         dq->mpd_kind = MPSC_QUEUE_KIND_UNKNOWN;
 417 }
 418
 419 #pragma mark deferred deallocation daemon
 420
 421 static struct mpsc_daemon_queue thread_deferred_deallocation_queue;
 422
 423 void
 424 thread_deallocate_daemon_init(void)
 425 {
 426         kern_return_t kr;
 427
 428         kr = _mpsc_daemon_queue_init_with_thread(&thread_deferred_deallocation_queue,
 429             mpsc_daemon_queue_nested_invoke, MINPRI_KERNEL,
 430             "daemon.deferred-deallocation", MPSC_QUEUE_KIND_THREAD_CRITICAL);
 431         if (kr != KERN_SUCCESS) {
 432                 panic("thread_deallocate_daemon_init: creating daemon failed (%d)", kr);
 433         }
 434 }
 435
 436 void
 437 thread_deallocate_daemon_register_queue(mpsc_daemon_queue_t dq,
 438     mpsc_daemon_invoke_fn_t invoke)
 439 {
 440         mpsc_daemon_queue_init_with_target(dq, invoke,
 441             &thread_deferred_deallocation_queue);
 442 }