lib/libzpool/taskq.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25 /*
  26  * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
  27  */
  28
  29 #include <sys/zfs_context.h>
  30
  31 int taskq_now;
  32 taskq_t *system_taskq;
  33
  34 #define TASKQ_ACTIVE    0x00010000
  35
  36 struct taskq {
  37         kmutex_t        tq_lock;
  38         krwlock_t       tq_threadlock;
  39         kcondvar_t      tq_dispatch_cv;
  40         kcondvar_t      tq_wait_cv;
  41         kthread_t       **tq_threadlist;
  42         int             tq_flags;
  43         int             tq_active;
  44         int             tq_nthreads;
  45         int             tq_nalloc;
  46         int             tq_minalloc;
  47         int             tq_maxalloc;
  48         kcondvar_t      tq_maxalloc_cv;
  49         int             tq_maxalloc_wait;
  50         taskq_ent_t     *tq_freelist;
  51         taskq_ent_t     tq_task;
  52 };
  53
  54 static taskq_ent_t *
  55 task_alloc(taskq_t *tq, int tqflags)
  56 {
  57         taskq_ent_t *t;
  58         int rv;
  59
  60 again:  if ((t = tq->tq_freelist) != NULL && tq->tq_nalloc >= tq->tq_minalloc) {
  61                 ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
  62                 tq->tq_freelist = t->tqent_next;
  63         } else {
  64                 if (tq->tq_nalloc >= tq->tq_maxalloc) {
  65                         if (!(tqflags & KM_SLEEP))
  66                                 return (NULL);
  67
  68                         /*
  69                          * We don't want to exceed tq_maxalloc, but we can't
  70                          * wait for other tasks to complete (and thus free up
  71                          * task structures) without risking deadlock with
  72                          * the caller.  So, we just delay for one second
  73                          * to throttle the allocation rate. If we have tasks
  74                          * complete before one second timeout expires then
  75                          * taskq_ent_free will signal us and we will
  76                          * immediately retry the allocation.
  77                          */
  78                         tq->tq_maxalloc_wait++;
  79                         rv = cv_timedwait(&tq->tq_maxalloc_cv,
  80                             &tq->tq_lock, ddi_get_lbolt() + hz);
  81                         tq->tq_maxalloc_wait--;
  82                         if (rv > 0)
  83                                 goto again;             /* signaled */
  84                 }
  85                 mutex_exit(&tq->tq_lock);
  86
  87                 t = kmem_alloc(sizeof (taskq_ent_t), tqflags);
  88
  89                 mutex_enter(&tq->tq_lock);
  90                 if (t != NULL) {
  91                         /* Make sure we start without any flags */
  92                         t->tqent_flags = 0;
  93                         tq->tq_nalloc++;
  94                 }
  95         }
  96         return (t);
  97 }
  98
  99 static void
 100 task_free(taskq_t *tq, taskq_ent_t *t)
 101 {
 102         if (tq->tq_nalloc <= tq->tq_minalloc) {
 103                 t->tqent_next = tq->tq_freelist;
 104                 tq->tq_freelist = t;
 105         } else {
 106                 tq->tq_nalloc--;
 107                 mutex_exit(&tq->tq_lock);
 108                 kmem_free(t, sizeof (taskq_ent_t));
 109                 mutex_enter(&tq->tq_lock);
 110         }
 111
 112         if (tq->tq_maxalloc_wait)
 113                 cv_signal(&tq->tq_maxalloc_cv);
 114 }
 115
 116 taskqid_t
 117 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t tqflags)
 118 {
 119         taskq_ent_t *t;
 120
 121         if (taskq_now) {
 122                 func(arg);
 123                 return (1);
 124         }
 125
 126         mutex_enter(&tq->tq_lock);
 127         ASSERT(tq->tq_flags & TASKQ_ACTIVE);
 128         if ((t = task_alloc(tq, tqflags)) == NULL) {
 129                 mutex_exit(&tq->tq_lock);
 130                 return (0);
 131         }
 132         if (tqflags & TQ_FRONT) {
 133                 t->tqent_next = tq->tq_task.tqent_next;
 134                 t->tqent_prev = &tq->tq_task;
 135         } else {
 136                 t->tqent_next = &tq->tq_task;
 137                 t->tqent_prev = tq->tq_task.tqent_prev;
 138         }
 139         t->tqent_next->tqent_prev = t;
 140         t->tqent_prev->tqent_next = t;
 141         t->tqent_func = func;
 142         t->tqent_arg = arg;
 143
 144         ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
 145
 146         cv_signal(&tq->tq_dispatch_cv);
 147         mutex_exit(&tq->tq_lock);
 148         return (1);
 149 }
 150
 151 int
 152 taskq_empty_ent(taskq_ent_t *t)
 153 {
 154         return t->tqent_next == NULL;
 155 }
 156
 157 void
 158 taskq_init_ent(taskq_ent_t *t)
 159 {
 160         t->tqent_next = NULL;
 161         t->tqent_prev = NULL;
 162         t->tqent_func = NULL;
 163         t->tqent_arg = NULL;
 164         t->tqent_flags = 0;
 165 }
 166
 167 void
 168 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
 169     taskq_ent_t *t)
 170 {
 171         ASSERT(func != NULL);
 172         ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
 173
 174         /*
 175          * Mark it as a prealloc'd task.  This is important
 176          * to ensure that we don't free it later.
 177          */
 178         t->tqent_flags |= TQENT_FLAG_PREALLOC;
 179         /*
 180          * Enqueue the task to the underlying queue.
 181          */
 182         mutex_enter(&tq->tq_lock);
 183
 184         if (flags & TQ_FRONT) {
 185                 t->tqent_next = tq->tq_task.tqent_next;
 186                 t->tqent_prev = &tq->tq_task;
 187         } else {
 188                 t->tqent_next = &tq->tq_task;
 189                 t->tqent_prev = tq->tq_task.tqent_prev;
 190         }
 191         t->tqent_next->tqent_prev = t;
 192         t->tqent_prev->tqent_next = t;
 193         t->tqent_func = func;
 194         t->tqent_arg = arg;
 195         cv_signal(&tq->tq_dispatch_cv);
 196         mutex_exit(&tq->tq_lock);
 197 }
 198
 199 void
 200 taskq_wait(taskq_t *tq)
 201 {
 202         mutex_enter(&tq->tq_lock);
 203         while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
 204                 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
 205         mutex_exit(&tq->tq_lock);
 206 }
 207
 208 static void
 209 taskq_thread(void *arg)
 210 {
 211         taskq_t *tq = arg;
 212         taskq_ent_t *t;
 213         boolean_t prealloc;
 214
 215         mutex_enter(&tq->tq_lock);
 216         while (tq->tq_flags & TASKQ_ACTIVE) {
 217                 if ((t = tq->tq_task.tqent_next) == &tq->tq_task) {
 218                         if (--tq->tq_active == 0)
 219                                 cv_broadcast(&tq->tq_wait_cv);
 220                         cv_wait(&tq->tq_dispatch_cv, &tq->tq_lock);
 221                         tq->tq_active++;
 222                         continue;
 223                 }
 224                 t->tqent_prev->tqent_next = t->tqent_next;
 225                 t->tqent_next->tqent_prev = t->tqent_prev;
 226                 t->tqent_next = NULL;
 227                 t->tqent_prev = NULL;
 228                 prealloc = t->tqent_flags & TQENT_FLAG_PREALLOC;
 229                 mutex_exit(&tq->tq_lock);
 230
 231                 rw_enter(&tq->tq_threadlock, RW_READER);
 232                 t->tqent_func(t->tqent_arg);
 233                 rw_exit(&tq->tq_threadlock);
 234
 235                 mutex_enter(&tq->tq_lock);
 236                 if (!prealloc)
 237                         task_free(tq, t);
 238         }
 239         tq->tq_nthreads--;
 240         cv_broadcast(&tq->tq_wait_cv);
 241         mutex_exit(&tq->tq_lock);
 242         thread_exit();
 243 }
 244
 245 /*ARGSUSED*/
 246 taskq_t *
 247 taskq_create(const char *name, int nthreads, pri_t pri,
 248         int minalloc, int maxalloc, uint_t flags)
 249 {
 250         taskq_t *tq = kmem_zalloc(sizeof (taskq_t), KM_SLEEP);
 251         int t;
 252
 253         if (flags & TASKQ_THREADS_CPU_PCT) {
 254                 int pct;
 255                 ASSERT3S(nthreads, >=, 0);
 256                 ASSERT3S(nthreads, <=, 100);
 257                 pct = MIN(nthreads, 100);
 258                 pct = MAX(pct, 0);
 259
 260                 nthreads = (sysconf(_SC_NPROCESSORS_ONLN) * pct) / 100;
 261                 nthreads = MAX(nthreads, 1);    /* need at least 1 thread */
 262         } else {
 263                 ASSERT3S(nthreads, >=, 1);
 264         }
 265
 266         rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
 267         mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
 268         cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
 269         cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
 270         cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
 271         tq->tq_flags = flags | TASKQ_ACTIVE;
 272         tq->tq_active = nthreads;
 273         tq->tq_nthreads = nthreads;
 274         tq->tq_minalloc = minalloc;
 275         tq->tq_maxalloc = maxalloc;
 276         tq->tq_task.tqent_next = &tq->tq_task;
 277         tq->tq_task.tqent_prev = &tq->tq_task;
 278         tq->tq_threadlist = kmem_alloc(nthreads*sizeof(kthread_t *), KM_SLEEP);
 279
 280         if (flags & TASKQ_PREPOPULATE) {
 281                 mutex_enter(&tq->tq_lock);
 282                 while (minalloc-- > 0)
 283                         task_free(tq, task_alloc(tq, KM_SLEEP));
 284                 mutex_exit(&tq->tq_lock);
 285         }
 286
 287         for (t = 0; t < nthreads; t++)
 288                 VERIFY((tq->tq_threadlist[t] = thread_create(NULL, 0,
 289                     taskq_thread, tq, TS_RUN, NULL, 0, 0)) != NULL);
 290
 291         return (tq);
 292 }
 293
 294 void
 295 taskq_destroy(taskq_t *tq)
 296 {
 297         int nthreads = tq->tq_nthreads;
 298
 299         taskq_wait(tq);
 300
 301         mutex_enter(&tq->tq_lock);
 302
 303         tq->tq_flags &= ~TASKQ_ACTIVE;
 304         cv_broadcast(&tq->tq_dispatch_cv);
 305
 306         while (tq->tq_nthreads != 0)
 307                 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
 308
 309         tq->tq_minalloc = 0;
 310         while (tq->tq_nalloc != 0) {
 311                 ASSERT(tq->tq_freelist != NULL);
 312                 task_free(tq, task_alloc(tq, KM_SLEEP));
 313         }
 314
 315         mutex_exit(&tq->tq_lock);
 316
 317         kmem_free(tq->tq_threadlist, nthreads * sizeof (kthread_t *));
 318
 319         rw_destroy(&tq->tq_threadlock);
 320         mutex_destroy(&tq->tq_lock);
 321         cv_destroy(&tq->tq_dispatch_cv);
 322         cv_destroy(&tq->tq_wait_cv);
 323         cv_destroy(&tq->tq_maxalloc_cv);
 324
 325         kmem_free(tq, sizeof (taskq_t));
 326 }
 327
 328 int
 329 taskq_member(taskq_t *tq, kthread_t *t)
 330 {
 331         int i;
 332
 333         if (taskq_now)
 334                 return (1);
 335
 336         for (i = 0; i < tq->tq_nthreads; i++)
 337                 if (tq->tq_threadlist[i] == t)
 338                         return (1);
 339
 340         return (0);
 341 }
 342
 343 void
 344 system_taskq_init(void)
 345 {
 346         system_taskq = taskq_create("system_taskq", 64, minclsyspri, 4, 512,
 347             TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
 348 }
 349
 350 void
 351 system_taskq_fini(void)
 352 {
 353         taskq_destroy(system_taskq);
 354         system_taskq = NULL; /* defensive */
 355 }