Home | History | Annotate | Line # | Download | only in kern
kern_threadpool.c revision 1.19
      1 /*	$NetBSD: kern_threadpool.c,v 1.19 2020/09/07 01:08:27 riastradh Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2014, 2018 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Taylor R. Campbell and Jason R. Thorpe.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 /*
     33  * Thread pools.
     34  *
     35  * A thread pool is a collection of worker threads idle or running
     36  * jobs, together with an overseer thread that does not run jobs but
     37  * can be given jobs to assign to a worker thread.  Scheduling a job in
     38  * a thread pool does not allocate or even sleep at all, except perhaps
     39  * on an adaptive lock, unlike kthread_create.  Jobs reuse threads, so
     40  * they do not incur the expense of creating and destroying kthreads
     41  * unless there is not much work to be done.
     42  *
     43  * A per-CPU thread pool (threadpool_percpu) is a collection of thread
     44  * pools, one per CPU bound to that CPU.  For each priority level in
     45  * use, there is one shared unbound thread pool (i.e., pool of threads
     46  * not bound to any CPU) and one shared per-CPU thread pool.
     47  *
     48  * To use the unbound thread pool at priority pri, call
     49  * threadpool_get(&pool, pri).  When you're done, call
     50  * threadpool_put(pool, pri).
     51  *
     52  * To use the per-CPU thread pools at priority pri, call
     53  * threadpool_percpu_get(&pool_percpu, pri), and then use the thread
     54  * pool returned by threadpool_percpu_ref(pool_percpu) for the current
     55  * CPU, or by threadpool_percpu_ref_remote(pool_percpu, ci) for another
     56  * CPU.  When you're done, call threadpool_percpu_put(pool_percpu,
     57  * pri).
     58  *
     59  * +--MACHINE-----------------------------------------------+
     60  * | +--CPU 0-------+ +--CPU 1-------+     +--CPU n-------+ |
     61  * | | <overseer 0> | | <overseer 1> | ... | <overseer n> | |
     62  * | | <idle 0a>    | | <running 1a> | ... | <idle na>    | |
     63  * | | <running 0b> | | <running 1b> | ... | <idle nb>    | |
     64  * | | .            | | .            | ... | .            | |
     65  * | | .            | | .            | ... | .            | |
     66  * | | .            | | .            | ... | .            | |
     67  * | +--------------+ +--------------+     +--------------+ |
     68  * |            +--unbound---------+                        |
     69  * |            | <overseer n+1>   |                        |
     70  * |            | <idle (n+1)a>    |                        |
     71  * |            | <running (n+1)b> |                        |
     72  * |            +------------------+                        |
     73  * +--------------------------------------------------------+
     74  *
     75  * XXX Why one overseer per CPU?  I did that originally to avoid
     76  * touching remote CPUs' memory when scheduling a job, but that still
     77  * requires interprocessor synchronization.  Perhaps we could get by
     78  * with a single overseer thread, at the expense of another pointer in
     79  * struct threadpool_job to identify the CPU on which it must run
     80  * in order for the overseer to schedule it correctly.
     81  */
     82 
     83 #include <sys/cdefs.h>
     84 __KERNEL_RCSID(0, "$NetBSD: kern_threadpool.c,v 1.19 2020/09/07 01:08:27 riastradh Exp $");
     85 
     86 #include <sys/types.h>
     87 #include <sys/param.h>
     88 #include <sys/atomic.h>
     89 #include <sys/condvar.h>
     90 #include <sys/cpu.h>
     91 #include <sys/kernel.h>
     92 #include <sys/kmem.h>
     93 #include <sys/kthread.h>
     94 #include <sys/mutex.h>
     95 #include <sys/once.h>
     96 #include <sys/percpu.h>
     97 #include <sys/pool.h>
     98 #include <sys/proc.h>
     99 #include <sys/queue.h>
    100 #include <sys/sdt.h>
    101 #include <sys/sysctl.h>
    102 #include <sys/systm.h>
    103 #include <sys/threadpool.h>
    104 
    105 /* Probes */
    106 
    107 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, get,
    108     "pri_t"/*pri*/);
    109 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, get__create,
    110     "pri_t"/*pri*/);
    111 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, get__race,
    112     "pri_t"/*pri*/);
    113 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, put,
    114     "struct threadpool *"/*pool*/, "pri_t"/*pri*/);
    115 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, put__destroy,
    116     "struct threadpool *"/*pool*/, "pri_t"/*pri*/);
    117 
    118 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, percpu__get,
    119     "pri_t"/*pri*/);
    120 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, percpu__get__create,
    121     "pri_t"/*pri*/);
    122 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, percpu__get__race,
    123     "pri_t"/*pri*/);
    124 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, percpu__put,
    125     "struct threadpool *"/*pool*/, "pri_t"/*pri*/);
    126 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, percpu__put__destroy,
    127     "struct threadpool *"/*pool*/, "pri_t"/*pri*/);
    128 
    129 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, create,
    130     "struct cpu_info *"/*ci*/, "pri_t"/*pri*/);
    131 SDT_PROBE_DEFINE3(sdt, kernel, threadpool, create__success,
    132     "struct cpu_info *"/*ci*/, "pri_t"/*pri*/, "struct threadpool *"/*pool*/);
    133 SDT_PROBE_DEFINE3(sdt, kernel, threadpool, create__failure,
    134     "struct cpu_info *"/*ci*/, "pri_t"/*pri*/, "int"/*error*/);
    135 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, destroy,
    136     "struct threadpool *"/*pool*/);
    137 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, destroy__wait,
    138     "struct threadpool *"/*pool*/, "uint64_t"/*refcnt*/);
    139 
    140 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, schedule__job,
    141     "struct threadpool *"/*pool*/, "struct threadpool_job *"/*job*/);
    142 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, schedule__job__running,
    143     "struct threadpool *"/*pool*/, "struct threadpool_job *"/*job*/);
    144 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, schedule__job__overseer,
    145     "struct threadpool *"/*pool*/, "struct threadpool_job *"/*job*/);
    146 SDT_PROBE_DEFINE3(sdt, kernel, threadpool, schedule__job__thread,
    147     "struct threadpool *"/*pool*/,
    148     "struct threadpool_job *"/*job*/,
    149     "struct lwp *"/*thread*/);
    150 
    151 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, overseer__start,
    152     "struct threadpool *"/*pool*/);
    153 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, overseer__dying,
    154     "struct threadpool *"/*pool*/);
    155 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, overseer__spawn,
    156     "struct threadpool *"/*pool*/);
    157 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, overseer__race,
    158     "struct threadpool *"/*pool*/,
    159     "struct threadpool_job *"/*job*/);
    160 SDT_PROBE_DEFINE3(sdt, kernel, threadpool, overseer__assign,
    161     "struct threadpool *"/*pool*/,
    162     "struct threadpool_job *"/*job*/,
    163     "struct lwp *"/*thread*/);
    164 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, overseer__exit,
    165     "struct threadpool *"/*pool*/);
    166 
    167 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, thread__start,
    168     "struct threadpool *"/*pool*/);
    169 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, thread__dying,
    170     "struct threadpool *"/*pool*/);
    171 SDT_PROBE_DEFINE2(sdt, kernel, threadpool, thread__job,
    172     "struct threadpool *"/*pool*/, "struct threadpool_job *"/*job*/);
    173 SDT_PROBE_DEFINE1(sdt, kernel, threadpool, thread__exit,
    174     "struct threadpool *"/*pool*/);
    175 
    176 /* Data structures */
    177 
    178 TAILQ_HEAD(job_head, threadpool_job);
    179 TAILQ_HEAD(thread_head, threadpool_thread);
    180 
    181 struct threadpool_thread {
    182 	struct lwp			*tpt_lwp;
    183 	char				*tpt_lwp_savedname;
    184 	struct threadpool		*tpt_pool;
    185 	struct threadpool_job		*tpt_job;
    186 	kcondvar_t			tpt_cv;
    187 	TAILQ_ENTRY(threadpool_thread)	tpt_entry;
    188 };
    189 
    190 struct threadpool {
    191 	kmutex_t			tp_lock;
    192 	struct threadpool_thread	tp_overseer;
    193 	struct job_head			tp_jobs;
    194 	struct thread_head		tp_idle_threads;
    195 	uint64_t			tp_refcnt;
    196 	int				tp_flags;
    197 #define	THREADPOOL_DYING	0x01
    198 	struct cpu_info			*tp_cpu;
    199 	pri_t				tp_pri;
    200 };
    201 
    202 static void	threadpool_hold(struct threadpool *);
    203 static void	threadpool_rele(struct threadpool *);
    204 
    205 static int	threadpool_percpu_create(struct threadpool_percpu **, pri_t);
    206 static void	threadpool_percpu_destroy(struct threadpool_percpu *);
    207 static void	threadpool_percpu_init(void *, void *, struct cpu_info *);
    208 static void	threadpool_percpu_ok(void *, void *, struct cpu_info *);
    209 static void	threadpool_percpu_fini(void *, void *, struct cpu_info *);
    210 
    211 static threadpool_job_fn_t threadpool_job_dead;
    212 
    213 static void	threadpool_job_hold(struct threadpool_job *);
    214 static void	threadpool_job_rele(struct threadpool_job *);
    215 
    216 static void	threadpool_overseer_thread(void *) __dead;
    217 static void	threadpool_thread(void *) __dead;
    218 
    219 static pool_cache_t	threadpool_thread_pc __read_mostly;
    220 
    221 static kmutex_t		threadpools_lock __cacheline_aligned;
    222 
    223 	/* Default to 30 second idle timeout for pool threads. */
    224 static int	threadpool_idle_time_ms = 30 * 1000;
    225 
    226 struct threadpool_unbound {
    227 	struct threadpool		tpu_pool;
    228 
    229 	/* protected by threadpools_lock */
    230 	LIST_ENTRY(threadpool_unbound)	tpu_link;
    231 	uint64_t			tpu_refcnt;
    232 };
    233 
    234 static LIST_HEAD(, threadpool_unbound) unbound_threadpools;
    235 
    236 static struct threadpool_unbound *
    237 threadpool_lookup_unbound(pri_t pri)
    238 {
    239 	struct threadpool_unbound *tpu;
    240 
    241 	LIST_FOREACH(tpu, &unbound_threadpools, tpu_link) {
    242 		if (tpu->tpu_pool.tp_pri == pri)
    243 			return tpu;
    244 	}
    245 	return NULL;
    246 }
    247 
    248 static void
    249 threadpool_insert_unbound(struct threadpool_unbound *tpu)
    250 {
    251 	KASSERT(threadpool_lookup_unbound(tpu->tpu_pool.tp_pri) == NULL);
    252 	LIST_INSERT_HEAD(&unbound_threadpools, tpu, tpu_link);
    253 }
    254 
    255 static void
    256 threadpool_remove_unbound(struct threadpool_unbound *tpu)
    257 {
    258 	KASSERT(threadpool_lookup_unbound(tpu->tpu_pool.tp_pri) == tpu);
    259 	LIST_REMOVE(tpu, tpu_link);
    260 }
    261 
    262 struct threadpool_percpu {
    263 	percpu_t *			tpp_percpu;
    264 	pri_t				tpp_pri;
    265 
    266 	/* protected by threadpools_lock */
    267 	LIST_ENTRY(threadpool_percpu)	tpp_link;
    268 	uint64_t			tpp_refcnt;
    269 };
    270 
    271 static LIST_HEAD(, threadpool_percpu) percpu_threadpools;
    272 
    273 static struct threadpool_percpu *
    274 threadpool_lookup_percpu(pri_t pri)
    275 {
    276 	struct threadpool_percpu *tpp;
    277 
    278 	LIST_FOREACH(tpp, &percpu_threadpools, tpp_link) {
    279 		if (tpp->tpp_pri == pri)
    280 			return tpp;
    281 	}
    282 	return NULL;
    283 }
    284 
    285 static void
    286 threadpool_insert_percpu(struct threadpool_percpu *tpp)
    287 {
    288 	KASSERT(threadpool_lookup_percpu(tpp->tpp_pri) == NULL);
    289 	LIST_INSERT_HEAD(&percpu_threadpools, tpp, tpp_link);
    290 }
    291 
    292 static void
    293 threadpool_remove_percpu(struct threadpool_percpu *tpp)
    294 {
    295 	KASSERT(threadpool_lookup_percpu(tpp->tpp_pri) == tpp);
    296 	LIST_REMOVE(tpp, tpp_link);
    297 }
    298 
    299 static int
    300 sysctl_kern_threadpool_idle_ms(SYSCTLFN_ARGS)
    301 {
    302 	struct sysctlnode node;
    303 	int val, error;
    304 
    305 	node = *rnode;
    306 
    307 	val = threadpool_idle_time_ms;
    308 	node.sysctl_data = &val;
    309 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
    310 	if (error == 0 && newp != NULL) {
    311 		/* Disallow negative values and 0 (forever). */
    312 		if (val < 1)
    313 			error = EINVAL;
    314 		else
    315 			threadpool_idle_time_ms = val;
    316 	}
    317 
    318 	return error;
    319 }
    320 
    321 SYSCTL_SETUP_PROTO(sysctl_threadpool_setup);
    322 
    323 SYSCTL_SETUP(sysctl_threadpool_setup,
    324     "sysctl kern.threadpool subtree setup")
    325 {
    326 	const struct sysctlnode *rnode, *cnode;
    327 	int error __diagused;
    328 
    329 	error = sysctl_createv(clog, 0, NULL, &rnode,
    330 	    CTLFLAG_PERMANENT,
    331 	    CTLTYPE_NODE, "threadpool",
    332 	    SYSCTL_DESCR("threadpool subsystem options"),
    333 	    NULL, 0, NULL, 0,
    334 	    CTL_KERN, CTL_CREATE, CTL_EOL);
    335 	KASSERT(error == 0);
    336 
    337 	error = sysctl_createv(clog, 0, &rnode, &cnode,
    338 	    CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
    339 	    CTLTYPE_INT, "idle_ms",
    340 	    SYSCTL_DESCR("idle thread timeout in ms"),
    341 	    sysctl_kern_threadpool_idle_ms, 0, NULL, 0,
    342 	    CTL_CREATE, CTL_EOL);
    343 	KASSERT(error == 0);
    344 }
    345 
    346 void
    347 threadpools_init(void)
    348 {
    349 
    350 	threadpool_thread_pc =
    351 	    pool_cache_init(sizeof(struct threadpool_thread), 0, 0, 0,
    352 		"thplthrd", NULL, IPL_NONE, NULL, NULL, NULL);
    353 
    354 	LIST_INIT(&unbound_threadpools);
    355 	LIST_INIT(&percpu_threadpools);
    356 	mutex_init(&threadpools_lock, MUTEX_DEFAULT, IPL_NONE);
    357 }
    358 
    359 /* Thread pool creation */
    360 
    361 static bool
    362 threadpool_pri_is_valid(pri_t pri)
    363 {
    364 	return (pri == PRI_NONE || (pri >= PRI_USER && pri < PRI_COUNT));
    365 }
    366 
    367 static int
    368 threadpool_create(struct threadpool *const pool, struct cpu_info *ci,
    369     pri_t pri)
    370 {
    371 	struct lwp *lwp;
    372 	int ktflags;
    373 	int error;
    374 
    375 	KASSERT(threadpool_pri_is_valid(pri));
    376 
    377 	SDT_PROBE2(sdt, kernel, threadpool, create,  ci, pri);
    378 
    379 	mutex_init(&pool->tp_lock, MUTEX_DEFAULT, IPL_VM);
    380 	/* XXX overseer */
    381 	TAILQ_INIT(&pool->tp_jobs);
    382 	TAILQ_INIT(&pool->tp_idle_threads);
    383 	pool->tp_refcnt = 1;		/* overseer's reference */
    384 	pool->tp_flags = 0;
    385 	pool->tp_cpu = ci;
    386 	pool->tp_pri = pri;
    387 
    388 	pool->tp_overseer.tpt_lwp = NULL;
    389 	pool->tp_overseer.tpt_pool = pool;
    390 	pool->tp_overseer.tpt_job = NULL;
    391 	cv_init(&pool->tp_overseer.tpt_cv, "poolover");
    392 
    393 	ktflags = 0;
    394 	ktflags |= KTHREAD_MPSAFE;
    395 	if (pri < PRI_KERNEL)
    396 		ktflags |= KTHREAD_TS;
    397 	error = kthread_create(pri, ktflags, ci, &threadpool_overseer_thread,
    398 	    &pool->tp_overseer, &lwp,
    399 	    "pooloverseer/%d@%d", (ci ? cpu_index(ci) : -1), (int)pri);
    400 	if (error)
    401 		goto fail0;
    402 
    403 	mutex_spin_enter(&pool->tp_lock);
    404 	pool->tp_overseer.tpt_lwp = lwp;
    405 	cv_broadcast(&pool->tp_overseer.tpt_cv);
    406 	mutex_spin_exit(&pool->tp_lock);
    407 
    408 	SDT_PROBE3(sdt, kernel, threadpool, create__success,  ci, pri, pool);
    409 	return 0;
    410 
    411 fail0:	KASSERT(error);
    412 	KASSERT(pool->tp_overseer.tpt_job == NULL);
    413 	KASSERT(pool->tp_overseer.tpt_pool == pool);
    414 	KASSERT(pool->tp_flags == 0);
    415 	KASSERT(pool->tp_refcnt == 0);
    416 	KASSERT(TAILQ_EMPTY(&pool->tp_idle_threads));
    417 	KASSERT(TAILQ_EMPTY(&pool->tp_jobs));
    418 	KASSERT(!cv_has_waiters(&pool->tp_overseer.tpt_cv));
    419 	cv_destroy(&pool->tp_overseer.tpt_cv);
    420 	mutex_destroy(&pool->tp_lock);
    421 	SDT_PROBE3(sdt, kernel, threadpool, create__failure,  ci, pri, error);
    422 	return error;
    423 }
    424 
    425 /* Thread pool destruction */
    426 
    427 static void
    428 threadpool_destroy(struct threadpool *pool)
    429 {
    430 	struct threadpool_thread *thread;
    431 
    432 	SDT_PROBE1(sdt, kernel, threadpool, destroy,  pool);
    433 
    434 	/* Mark the pool dying and wait for threads to commit suicide.  */
    435 	mutex_spin_enter(&pool->tp_lock);
    436 	KASSERT(TAILQ_EMPTY(&pool->tp_jobs));
    437 	pool->tp_flags |= THREADPOOL_DYING;
    438 	cv_broadcast(&pool->tp_overseer.tpt_cv);
    439 	TAILQ_FOREACH(thread, &pool->tp_idle_threads, tpt_entry)
    440 		cv_broadcast(&thread->tpt_cv);
    441 	while (0 < pool->tp_refcnt) {
    442 		SDT_PROBE2(sdt, kernel, threadpool, destroy__wait,
    443 		    pool, pool->tp_refcnt);
    444 		cv_wait(&pool->tp_overseer.tpt_cv, &pool->tp_lock);
    445 	}
    446 	mutex_spin_exit(&pool->tp_lock);
    447 
    448 	KASSERT(pool->tp_overseer.tpt_job == NULL);
    449 	KASSERT(pool->tp_overseer.tpt_pool == pool);
    450 	KASSERT(pool->tp_flags == THREADPOOL_DYING);
    451 	KASSERT(pool->tp_refcnt == 0);
    452 	KASSERT(TAILQ_EMPTY(&pool->tp_idle_threads));
    453 	KASSERT(TAILQ_EMPTY(&pool->tp_jobs));
    454 	KASSERT(!cv_has_waiters(&pool->tp_overseer.tpt_cv));
    455 	cv_destroy(&pool->tp_overseer.tpt_cv);
    456 	mutex_destroy(&pool->tp_lock);
    457 }
    458 
    459 static void
    460 threadpool_hold(struct threadpool *pool)
    461 {
    462 
    463 	KASSERT(mutex_owned(&pool->tp_lock));
    464 	pool->tp_refcnt++;
    465 	KASSERT(pool->tp_refcnt != 0);
    466 }
    467 
    468 static void
    469 threadpool_rele(struct threadpool *pool)
    470 {
    471 
    472 	KASSERT(mutex_owned(&pool->tp_lock));
    473 	KASSERT(0 < pool->tp_refcnt);
    474 	if (--pool->tp_refcnt == 0)
    475 		cv_broadcast(&pool->tp_overseer.tpt_cv);
    476 }
    477 
    478 /* Unbound thread pools */
    479 
    480 int
    481 threadpool_get(struct threadpool **poolp, pri_t pri)
    482 {
    483 	struct threadpool_unbound *tpu, *tmp = NULL;
    484 	int error;
    485 
    486 	ASSERT_SLEEPABLE();
    487 
    488 	SDT_PROBE1(sdt, kernel, threadpool, get,  pri);
    489 
    490 	if (! threadpool_pri_is_valid(pri))
    491 		return EINVAL;
    492 
    493 	mutex_enter(&threadpools_lock);
    494 	tpu = threadpool_lookup_unbound(pri);
    495 	if (tpu == NULL) {
    496 		mutex_exit(&threadpools_lock);
    497 		SDT_PROBE1(sdt, kernel, threadpool, get__create,  pri);
    498 		tmp = kmem_zalloc(sizeof(*tmp), KM_SLEEP);
    499 		error = threadpool_create(&tmp->tpu_pool, NULL, pri);
    500 		if (error) {
    501 			kmem_free(tmp, sizeof(*tmp));
    502 			return error;
    503 		}
    504 		mutex_enter(&threadpools_lock);
    505 		tpu = threadpool_lookup_unbound(pri);
    506 		if (tpu == NULL) {
    507 			tpu = tmp;
    508 			tmp = NULL;
    509 			threadpool_insert_unbound(tpu);
    510 		} else {
    511 			SDT_PROBE1(sdt, kernel, threadpool, get__race,  pri);
    512 		}
    513 	}
    514 	KASSERT(tpu != NULL);
    515 	tpu->tpu_refcnt++;
    516 	KASSERT(tpu->tpu_refcnt != 0);
    517 	mutex_exit(&threadpools_lock);
    518 
    519 	if (tmp != NULL) {
    520 		threadpool_destroy(&tmp->tpu_pool);
    521 		kmem_free(tmp, sizeof(*tmp));
    522 	}
    523 	KASSERT(tpu != NULL);
    524 	*poolp = &tpu->tpu_pool;
    525 	return 0;
    526 }
    527 
    528 void
    529 threadpool_put(struct threadpool *pool, pri_t pri)
    530 {
    531 	struct threadpool_unbound *tpu =
    532 	    container_of(pool, struct threadpool_unbound, tpu_pool);
    533 
    534 	ASSERT_SLEEPABLE();
    535 	KASSERT(threadpool_pri_is_valid(pri));
    536 
    537 	SDT_PROBE2(sdt, kernel, threadpool, put,  pool, pri);
    538 
    539 	mutex_enter(&threadpools_lock);
    540 	KASSERT(tpu == threadpool_lookup_unbound(pri));
    541 	KASSERT(0 < tpu->tpu_refcnt);
    542 	if (--tpu->tpu_refcnt == 0) {
    543 		SDT_PROBE2(sdt, kernel, threadpool, put__destroy,  pool, pri);
    544 		threadpool_remove_unbound(tpu);
    545 	} else {
    546 		tpu = NULL;
    547 	}
    548 	mutex_exit(&threadpools_lock);
    549 
    550 	if (tpu) {
    551 		threadpool_destroy(&tpu->tpu_pool);
    552 		kmem_free(tpu, sizeof(*tpu));
    553 	}
    554 }
    555 
    556 /* Per-CPU thread pools */
    557 
    558 int
    559 threadpool_percpu_get(struct threadpool_percpu **pool_percpup, pri_t pri)
    560 {
    561 	struct threadpool_percpu *pool_percpu, *tmp = NULL;
    562 	int error;
    563 
    564 	ASSERT_SLEEPABLE();
    565 
    566 	SDT_PROBE1(sdt, kernel, threadpool, percpu__get,  pri);
    567 
    568 	if (! threadpool_pri_is_valid(pri))
    569 		return EINVAL;
    570 
    571 	mutex_enter(&threadpools_lock);
    572 	pool_percpu = threadpool_lookup_percpu(pri);
    573 	if (pool_percpu == NULL) {
    574 		mutex_exit(&threadpools_lock);
    575 		SDT_PROBE1(sdt, kernel, threadpool, percpu__get__create,  pri);
    576 		error = threadpool_percpu_create(&tmp, pri);
    577 		if (error)
    578 			return error;
    579 		KASSERT(tmp != NULL);
    580 		mutex_enter(&threadpools_lock);
    581 		pool_percpu = threadpool_lookup_percpu(pri);
    582 		if (pool_percpu == NULL) {
    583 			pool_percpu = tmp;
    584 			tmp = NULL;
    585 			threadpool_insert_percpu(pool_percpu);
    586 		} else {
    587 			SDT_PROBE1(sdt, kernel, threadpool, percpu__get__race,
    588 			    pri);
    589 		}
    590 	}
    591 	KASSERT(pool_percpu != NULL);
    592 	pool_percpu->tpp_refcnt++;
    593 	KASSERT(pool_percpu->tpp_refcnt != 0);
    594 	mutex_exit(&threadpools_lock);
    595 
    596 	if (tmp != NULL)
    597 		threadpool_percpu_destroy(tmp);
    598 	KASSERT(pool_percpu != NULL);
    599 	*pool_percpup = pool_percpu;
    600 	return 0;
    601 }
    602 
    603 void
    604 threadpool_percpu_put(struct threadpool_percpu *pool_percpu, pri_t pri)
    605 {
    606 
    607 	ASSERT_SLEEPABLE();
    608 
    609 	KASSERT(threadpool_pri_is_valid(pri));
    610 
    611 	SDT_PROBE2(sdt, kernel, threadpool, percpu__put,  pool_percpu, pri);
    612 
    613 	mutex_enter(&threadpools_lock);
    614 	KASSERT(pool_percpu == threadpool_lookup_percpu(pri));
    615 	KASSERT(0 < pool_percpu->tpp_refcnt);
    616 	if (--pool_percpu->tpp_refcnt == 0) {
    617 		SDT_PROBE2(sdt, kernel, threadpool, percpu__put__destroy,
    618 		    pool_percpu, pri);
    619 		threadpool_remove_percpu(pool_percpu);
    620 	} else {
    621 		pool_percpu = NULL;
    622 	}
    623 	mutex_exit(&threadpools_lock);
    624 
    625 	if (pool_percpu)
    626 		threadpool_percpu_destroy(pool_percpu);
    627 }
    628 
    629 struct threadpool *
    630 threadpool_percpu_ref(struct threadpool_percpu *pool_percpu)
    631 {
    632 	struct threadpool **poolp, *pool;
    633 
    634 	poolp = percpu_getref(pool_percpu->tpp_percpu);
    635 	pool = *poolp;
    636 	percpu_putref(pool_percpu->tpp_percpu);
    637 
    638 	return pool;
    639 }
    640 
    641 struct threadpool *
    642 threadpool_percpu_ref_remote(struct threadpool_percpu *pool_percpu,
    643     struct cpu_info *ci)
    644 {
    645 	struct threadpool **poolp, *pool;
    646 
    647 	percpu_traverse_enter();
    648 	poolp = percpu_getptr_remote(pool_percpu->tpp_percpu, ci);
    649 	pool = *poolp;
    650 	percpu_traverse_exit();
    651 
    652 	return pool;
    653 }
    654 
    655 static int
    656 threadpool_percpu_create(struct threadpool_percpu **pool_percpup, pri_t pri)
    657 {
    658 	struct threadpool_percpu *pool_percpu;
    659 	bool ok = true;
    660 
    661 	pool_percpu = kmem_zalloc(sizeof(*pool_percpu), KM_SLEEP);
    662 	pool_percpu->tpp_pri = pri;
    663 	pool_percpu->tpp_percpu = percpu_create(sizeof(struct threadpool *),
    664 	    threadpool_percpu_init, threadpool_percpu_fini,
    665 	    (void *)(intptr_t)pri);
    666 
    667 	/*
    668 	 * Verify that all of the CPUs were initialized.
    669 	 *
    670 	 * XXX What to do if we add CPU hotplug?
    671 	 */
    672 	percpu_foreach(pool_percpu->tpp_percpu, &threadpool_percpu_ok, &ok);
    673 	if (!ok)
    674 		goto fail;
    675 
    676 	/* Success!  */
    677 	*pool_percpup = (struct threadpool_percpu *)pool_percpu;
    678 	return 0;
    679 
    680 fail:	percpu_free(pool_percpu->tpp_percpu, sizeof(struct threadpool *));
    681 	kmem_free(pool_percpu, sizeof(*pool_percpu));
    682 	return ENOMEM;
    683 }
    684 
    685 static void
    686 threadpool_percpu_destroy(struct threadpool_percpu *pool_percpu)
    687 {
    688 
    689 	percpu_free(pool_percpu->tpp_percpu, sizeof(struct threadpool *));
    690 	kmem_free(pool_percpu, sizeof(*pool_percpu));
    691 }
    692 
    693 static void
    694 threadpool_percpu_init(void *vpoolp, void *vpri, struct cpu_info *ci)
    695 {
    696 	struct threadpool **const poolp = vpoolp;
    697 	pri_t pri = (intptr_t)(void *)vpri;
    698 	int error;
    699 
    700 	*poolp = kmem_zalloc(sizeof(**poolp), KM_SLEEP);
    701 	error = threadpool_create(*poolp, ci, pri);
    702 	if (error) {
    703 		KASSERT(error == ENOMEM);
    704 		kmem_free(*poolp, sizeof(**poolp));
    705 		*poolp = NULL;
    706 	}
    707 }
    708 
    709 static void
    710 threadpool_percpu_ok(void *vpoolp, void *vokp, struct cpu_info *ci)
    711 {
    712 	struct threadpool **const poolp = vpoolp;
    713 	bool *okp = vokp;
    714 
    715 	if (*poolp == NULL)
    716 		atomic_store_relaxed(okp, false);
    717 }
    718 
    719 static void
    720 threadpool_percpu_fini(void *vpoolp, void *vprip, struct cpu_info *ci)
    721 {
    722 	struct threadpool **const poolp = vpoolp;
    723 
    724 	if (*poolp == NULL)	/* initialization failed */
    725 		return;
    726 	threadpool_destroy(*poolp);
    727 	kmem_free(*poolp, sizeof(**poolp));
    728 }
    729 
    730 /* Thread pool jobs */
    731 
    732 void __printflike(4,5)
    733 threadpool_job_init(struct threadpool_job *job, threadpool_job_fn_t fn,
    734     kmutex_t *lock, const char *fmt, ...)
    735 {
    736 	va_list ap;
    737 
    738 	va_start(ap, fmt);
    739 	(void)vsnprintf(job->job_name, sizeof(job->job_name), fmt, ap);
    740 	va_end(ap);
    741 
    742 	job->job_lock = lock;
    743 	job->job_thread = NULL;
    744 	job->job_refcnt = 0;
    745 	cv_init(&job->job_cv, job->job_name);
    746 	job->job_fn = fn;
    747 }
    748 
    749 static void
    750 threadpool_job_dead(struct threadpool_job *job)
    751 {
    752 
    753 	panic("threadpool job %p ran after destruction", job);
    754 }
    755 
    756 void
    757 threadpool_job_destroy(struct threadpool_job *job)
    758 {
    759 
    760 	ASSERT_SLEEPABLE();
    761 
    762 	KASSERTMSG((job->job_thread == NULL), "job %p still running", job);
    763 
    764 	mutex_enter(job->job_lock);
    765 	while (0 < atomic_load_relaxed(&job->job_refcnt))
    766 		cv_wait(&job->job_cv, job->job_lock);
    767 	mutex_exit(job->job_lock);
    768 
    769 	job->job_lock = NULL;
    770 	KASSERT(job->job_thread == NULL);
    771 	KASSERT(job->job_refcnt == 0);
    772 	KASSERT(!cv_has_waiters(&job->job_cv));
    773 	cv_destroy(&job->job_cv);
    774 	job->job_fn = threadpool_job_dead;
    775 	(void)strlcpy(job->job_name, "deadjob", sizeof(job->job_name));
    776 }
    777 
    778 static void
    779 threadpool_job_hold(struct threadpool_job *job)
    780 {
    781 	unsigned int refcnt __diagused;
    782 
    783 	refcnt = atomic_inc_uint_nv(&job->job_refcnt);
    784 	KASSERT(refcnt != 0);
    785 }
    786 
    787 static void
    788 threadpool_job_rele(struct threadpool_job *job)
    789 {
    790 	unsigned int refcnt;
    791 
    792 	KASSERT(mutex_owned(job->job_lock));
    793 
    794 	refcnt = atomic_dec_uint_nv(&job->job_refcnt);
    795 	KASSERT(refcnt != UINT_MAX);
    796 	if (refcnt == 0)
    797 		cv_broadcast(&job->job_cv);
    798 }
    799 
    800 void
    801 threadpool_job_done(struct threadpool_job *job)
    802 {
    803 
    804 	KASSERT(mutex_owned(job->job_lock));
    805 	KASSERT(job->job_thread != NULL);
    806 	KASSERT(job->job_thread->tpt_lwp == curlwp);
    807 
    808 	/*
    809 	 * We can safely read this field; it's only modified right before
    810 	 * we call the job work function, and we are only preserving it
    811 	 * to use here; no one cares if it contains junk afterward.
    812 	 */
    813 	lwp_lock(curlwp);
    814 	curlwp->l_name = job->job_thread->tpt_lwp_savedname;
    815 	lwp_unlock(curlwp);
    816 
    817 	/*
    818 	 * Inline the work of threadpool_job_rele(); the job is already
    819 	 * locked, the most likely scenario (XXXJRT only scenario?) is
    820 	 * that we're dropping the last reference (the one taken in
    821 	 * threadpool_schedule_job()), and we always do the cv_broadcast()
    822 	 * anyway.
    823 	 */
    824 	KASSERT(0 < atomic_load_relaxed(&job->job_refcnt));
    825 	unsigned int refcnt __diagused = atomic_dec_uint_nv(&job->job_refcnt);
    826 	KASSERT(refcnt != UINT_MAX);
    827 	cv_broadcast(&job->job_cv);
    828 	job->job_thread = NULL;
    829 }
    830 
    831 void
    832 threadpool_schedule_job(struct threadpool *pool, struct threadpool_job *job)
    833 {
    834 
    835 	KASSERT(mutex_owned(job->job_lock));
    836 
    837 	SDT_PROBE2(sdt, kernel, threadpool, schedule__job,  pool, job);
    838 
    839 	/*
    840 	 * If the job's already running, let it keep running.  The job
    841 	 * is guaranteed by the interlock not to end early -- if it had
    842 	 * ended early, threadpool_job_done would have set job_thread
    843 	 * to NULL under the interlock.
    844 	 */
    845 	if (__predict_true(job->job_thread != NULL)) {
    846 		SDT_PROBE2(sdt, kernel, threadpool, schedule__job__running,
    847 		    pool, job);
    848 		return;
    849 	}
    850 
    851 	threadpool_job_hold(job);
    852 
    853 	/* Otherwise, try to assign a thread to the job.  */
    854 	mutex_spin_enter(&pool->tp_lock);
    855 	if (__predict_false(TAILQ_EMPTY(&pool->tp_idle_threads))) {
    856 		/* Nobody's idle.  Give it to the overseer.  */
    857 		SDT_PROBE2(sdt, kernel, threadpool, schedule__job__overseer,
    858 		    pool, job);
    859 		job->job_thread = &pool->tp_overseer;
    860 		TAILQ_INSERT_TAIL(&pool->tp_jobs, job, job_entry);
    861 	} else {
    862 		/* Assign it to the first idle thread.  */
    863 		job->job_thread = TAILQ_FIRST(&pool->tp_idle_threads);
    864 		SDT_PROBE3(sdt, kernel, threadpool, schedule__job__thread,
    865 		    pool, job, job->job_thread->tpt_lwp);
    866 		TAILQ_REMOVE(&pool->tp_idle_threads, job->job_thread,
    867 		    tpt_entry);
    868 		job->job_thread->tpt_job = job;
    869 	}
    870 
    871 	/* Notify whomever we gave it to, overseer or idle thread.  */
    872 	KASSERT(job->job_thread != NULL);
    873 	cv_broadcast(&job->job_thread->tpt_cv);
    874 	mutex_spin_exit(&pool->tp_lock);
    875 }
    876 
    877 bool
    878 threadpool_cancel_job_async(struct threadpool *pool, struct threadpool_job *job)
    879 {
    880 
    881 	KASSERT(mutex_owned(job->job_lock));
    882 
    883 	/*
    884 	 * XXXJRT This fails (albeit safely) when all of the following
    885 	 * are true:
    886 	 *
    887 	 *	=> "pool" is something other than what the job was
    888 	 *	   scheduled on.  This can legitimately occur if,
    889 	 *	   for example, a job is percpu-scheduled on CPU0
    890 	 *	   and then CPU1 attempts to cancel it without taking
    891 	 *	   a remote pool reference.  (this might happen by
    892 	 *	   "luck of the draw").
    893 	 *
    894 	 *	=> "job" is not yet running, but is assigned to the
    895 	 *	   overseer.
    896 	 *
    897 	 * When this happens, this code makes the determination that
    898 	 * the job is already running.  The failure mode is that the
    899 	 * caller is told the job is running, and thus has to wait.
    900 	 * The overseer will eventually get to it and the job will
    901 	 * proceed as if it had been already running.
    902 	 */
    903 
    904 	if (job->job_thread == NULL) {
    905 		/* Nothing to do.  Guaranteed not running.  */
    906 		return true;
    907 	} else if (job->job_thread == &pool->tp_overseer) {
    908 		/* Take it off the list to guarantee it won't run.  */
    909 		job->job_thread = NULL;
    910 		mutex_spin_enter(&pool->tp_lock);
    911 		TAILQ_REMOVE(&pool->tp_jobs, job, job_entry);
    912 		mutex_spin_exit(&pool->tp_lock);
    913 		threadpool_job_rele(job);
    914 		return true;
    915 	} else {
    916 		/* Too late -- already running.  */
    917 		return false;
    918 	}
    919 }
    920 
    921 void
    922 threadpool_cancel_job(struct threadpool *pool, struct threadpool_job *job)
    923 {
    924 
    925 	/*
    926 	 * We may sleep here, but we can't ASSERT_SLEEPABLE() because
    927 	 * the job lock (used to interlock the cv_wait()) may in fact
    928 	 * legitimately be a spin lock, so the assertion would fire
    929 	 * as a false-positive.
    930 	 */
    931 
    932 	KASSERT(mutex_owned(job->job_lock));
    933 
    934 	if (threadpool_cancel_job_async(pool, job))
    935 		return;
    936 
    937 	/* Already running.  Wait for it to complete.  */
    938 	while (job->job_thread != NULL)
    939 		cv_wait(&job->job_cv, job->job_lock);
    940 }
    941 
    942 /* Thread pool overseer thread */
    943 
    944 static void __dead
    945 threadpool_overseer_thread(void *arg)
    946 {
    947 	struct threadpool_thread *const overseer = arg;
    948 	struct threadpool *const pool = overseer->tpt_pool;
    949 	struct lwp *lwp = NULL;
    950 	int ktflags;
    951 	int error;
    952 
    953 	KASSERT((pool->tp_cpu == NULL) || (pool->tp_cpu == curcpu()));
    954 	KASSERT((pool->tp_cpu == NULL) || (curlwp->l_pflag & LP_BOUND));
    955 
    956 	/* Wait until we're initialized.  */
    957 	mutex_spin_enter(&pool->tp_lock);
    958 	while (overseer->tpt_lwp == NULL)
    959 		cv_wait(&overseer->tpt_cv, &pool->tp_lock);
    960 
    961 	SDT_PROBE1(sdt, kernel, threadpool, overseer__start,  pool);
    962 
    963 	for (;;) {
    964 		/* Wait until there's a job.  */
    965 		while (TAILQ_EMPTY(&pool->tp_jobs)) {
    966 			if (ISSET(pool->tp_flags, THREADPOOL_DYING)) {
    967 				SDT_PROBE1(sdt, kernel, threadpool,
    968 				    overseer__dying,  pool);
    969 				break;
    970 			}
    971 			cv_wait(&overseer->tpt_cv, &pool->tp_lock);
    972 		}
    973 		if (__predict_false(TAILQ_EMPTY(&pool->tp_jobs)))
    974 			break;
    975 
    976 		/* If there are no threads, we'll have to try to start one.  */
    977 		if (TAILQ_EMPTY(&pool->tp_idle_threads)) {
    978 			SDT_PROBE1(sdt, kernel, threadpool, overseer__spawn,
    979 			    pool);
    980 			threadpool_hold(pool);
    981 			mutex_spin_exit(&pool->tp_lock);
    982 
    983 			struct threadpool_thread *const thread =
    984 			    pool_cache_get(threadpool_thread_pc, PR_WAITOK);
    985 			thread->tpt_lwp = NULL;
    986 			thread->tpt_pool = pool;
    987 			thread->tpt_job = NULL;
    988 			cv_init(&thread->tpt_cv, "poolthrd");
    989 
    990 			ktflags = 0;
    991 			ktflags |= KTHREAD_MPSAFE;
    992 			if (pool->tp_pri < PRI_KERNEL)
    993 				ktflags |= KTHREAD_TS;
    994 			error = kthread_create(pool->tp_pri, ktflags,
    995 			    pool->tp_cpu, &threadpool_thread, thread, &lwp,
    996 			    "poolthread/%d@%d",
    997 			    (pool->tp_cpu ? cpu_index(pool->tp_cpu) : -1),
    998 			    (int)pool->tp_pri);
    999 
   1000 			mutex_spin_enter(&pool->tp_lock);
   1001 			if (error) {
   1002 				pool_cache_put(threadpool_thread_pc, thread);
   1003 				threadpool_rele(pool);
   1004 				/* XXX What to do to wait for memory?  */
   1005 				(void)kpause("thrdplcr", false, hz,
   1006 				    &pool->tp_lock);
   1007 				continue;
   1008 			}
   1009 			/*
   1010 			 * New kthread now owns the reference to the pool
   1011 			 * taken above.
   1012 			 */
   1013 			KASSERT(lwp != NULL);
   1014 			TAILQ_INSERT_TAIL(&pool->tp_idle_threads, thread,
   1015 			    tpt_entry);
   1016 			thread->tpt_lwp = lwp;
   1017 			lwp = NULL;
   1018 			cv_broadcast(&thread->tpt_cv);
   1019 			continue;
   1020 		}
   1021 
   1022 		/* There are idle threads, so try giving one a job.  */
   1023 		struct threadpool_job *const job = TAILQ_FIRST(&pool->tp_jobs);
   1024 		TAILQ_REMOVE(&pool->tp_jobs, job, job_entry);
   1025 		/*
   1026 		 * Take an extra reference on the job temporarily so that
   1027 		 * it won't disappear on us while we have both locks dropped.
   1028 		 */
   1029 		threadpool_job_hold(job);
   1030 		mutex_spin_exit(&pool->tp_lock);
   1031 
   1032 		mutex_enter(job->job_lock);
   1033 		/* If the job was cancelled, we'll no longer be its thread.  */
   1034 		if (__predict_true(job->job_thread == overseer)) {
   1035 			mutex_spin_enter(&pool->tp_lock);
   1036 			if (__predict_false(
   1037 				    TAILQ_EMPTY(&pool->tp_idle_threads))) {
   1038 				/*
   1039 				 * Someone else snagged the thread
   1040 				 * first.  We'll have to try again.
   1041 				 */
   1042 				SDT_PROBE2(sdt, kernel, threadpool,
   1043 				    overseer__race,  pool, job);
   1044 				TAILQ_INSERT_HEAD(&pool->tp_jobs, job,
   1045 				    job_entry);
   1046 			} else {
   1047 				/*
   1048 				 * Assign the job to the thread and
   1049 				 * wake the thread so it starts work.
   1050 				 */
   1051 				struct threadpool_thread *const thread =
   1052 				    TAILQ_FIRST(&pool->tp_idle_threads);
   1053 
   1054 				SDT_PROBE2(sdt, kernel, threadpool,
   1055 				    overseer__assign,  job, thread->tpt_lwp);
   1056 				KASSERT(thread->tpt_job == NULL);
   1057 				TAILQ_REMOVE(&pool->tp_idle_threads, thread,
   1058 				    tpt_entry);
   1059 				thread->tpt_job = job;
   1060 				job->job_thread = thread;
   1061 				cv_broadcast(&thread->tpt_cv);
   1062 			}
   1063 			mutex_spin_exit(&pool->tp_lock);
   1064 		}
   1065 		threadpool_job_rele(job);
   1066 		mutex_exit(job->job_lock);
   1067 
   1068 		mutex_spin_enter(&pool->tp_lock);
   1069 	}
   1070 	threadpool_rele(pool);
   1071 	mutex_spin_exit(&pool->tp_lock);
   1072 
   1073 	SDT_PROBE1(sdt, kernel, threadpool, overseer__exit,  pool);
   1074 
   1075 	kthread_exit(0);
   1076 }
   1077 
   1078 /* Thread pool thread */
   1079 
   1080 static void __dead
   1081 threadpool_thread(void *arg)
   1082 {
   1083 	struct threadpool_thread *const thread = arg;
   1084 	struct threadpool *const pool = thread->tpt_pool;
   1085 
   1086 	KASSERT((pool->tp_cpu == NULL) || (pool->tp_cpu == curcpu()));
   1087 	KASSERT((pool->tp_cpu == NULL) || (curlwp->l_pflag & LP_BOUND));
   1088 
   1089 	/* Wait until we're initialized and on the queue.  */
   1090 	mutex_spin_enter(&pool->tp_lock);
   1091 	while (thread->tpt_lwp == NULL)
   1092 		cv_wait(&thread->tpt_cv, &pool->tp_lock);
   1093 
   1094 	SDT_PROBE1(sdt, kernel, threadpool, thread__start,  pool);
   1095 
   1096 	KASSERT(thread->tpt_lwp == curlwp);
   1097 	for (;;) {
   1098 		/* Wait until we are assigned a job.  */
   1099 		while (thread->tpt_job == NULL) {
   1100 			if (ISSET(pool->tp_flags, THREADPOOL_DYING)) {
   1101 				SDT_PROBE1(sdt, kernel, threadpool,
   1102 				    thread__dying,  pool);
   1103 				break;
   1104 			}
   1105 			if (cv_timedwait(&thread->tpt_cv, &pool->tp_lock,
   1106 				mstohz(threadpool_idle_time_ms)))
   1107 				break;
   1108 		}
   1109 		if (__predict_false(thread->tpt_job == NULL)) {
   1110 			TAILQ_REMOVE(&pool->tp_idle_threads, thread,
   1111 			    tpt_entry);
   1112 			break;
   1113 		}
   1114 
   1115 		struct threadpool_job *const job = thread->tpt_job;
   1116 		KASSERT(job != NULL);
   1117 
   1118 		/* Set our lwp name to reflect what job we're doing.  */
   1119 		lwp_lock(curlwp);
   1120 		char *const lwp_name __diagused = curlwp->l_name;
   1121 		thread->tpt_lwp_savedname = curlwp->l_name;
   1122 		curlwp->l_name = job->job_name;
   1123 		lwp_unlock(curlwp);
   1124 
   1125 		mutex_spin_exit(&pool->tp_lock);
   1126 
   1127 		SDT_PROBE2(sdt, kernel, threadpool, thread__job,  pool, job);
   1128 
   1129 		/* Run the job.  */
   1130 		(*job->job_fn)(job);
   1131 
   1132 		/* lwp name restored in threadpool_job_done(). */
   1133 		KASSERTMSG((curlwp->l_name == lwp_name),
   1134 		    "someone forgot to call threadpool_job_done()!");
   1135 
   1136 		/*
   1137 		 * We can compare pointers, but we can no longer deference
   1138 		 * job after this because threadpool_job_done() drops the
   1139 		 * last reference on the job while the job is locked.
   1140 		 */
   1141 
   1142 		mutex_spin_enter(&pool->tp_lock);
   1143 		KASSERT(thread->tpt_job == job);
   1144 		thread->tpt_job = NULL;
   1145 		TAILQ_INSERT_TAIL(&pool->tp_idle_threads, thread, tpt_entry);
   1146 	}
   1147 	threadpool_rele(pool);
   1148 	mutex_spin_exit(&pool->tp_lock);
   1149 
   1150 	SDT_PROBE1(sdt, kernel, threadpool, thread__exit,  pool);
   1151 
   1152 	KASSERT(!cv_has_waiters(&thread->tpt_cv));
   1153 	cv_destroy(&thread->tpt_cv);
   1154 	pool_cache_put(threadpool_thread_pc, thread);
   1155 	kthread_exit(0);
   1156 }
   1157