xref: /illumos-gate/usr/src/uts/common/os/taskq.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
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 /*
27  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
28  * Copyright (c) 2017 by Delphix. All rights reserved.
29  * Copyright 2018, Joyent, Inc.
30  * Copyright 2023 RackTop Systems, Inc.
31  */
32 
33 /*
34  * Kernel task queues: general-purpose asynchronous task scheduling.
35  *
36  * A common problem in kernel programming is the need to schedule tasks
37  * to be performed later, by another thread. There are several reasons
38  * you may want or need to do this:
39  *
40  * (1) The task isn't time-critical, but your current code path is.
41  *
42  * (2) The task may require grabbing locks that you already hold.
43  *
44  * (3) The task may need to block (e.g. to wait for memory), but you
45  *     cannot block in your current context.
46  *
47  * (4) Your code path can't complete because of some condition, but you can't
48  *     sleep or fail, so you queue the task for later execution when condition
49  *     disappears.
50  *
51  * (5) You just want a simple way to launch multiple tasks in parallel.
52  *
53  * Task queues provide such a facility. In its simplest form (used when
54  * performance is not a critical consideration) a task queue consists of a
55  * single list of tasks, together with one or more threads to service the
56  * list. There are some cases when this simple queue is not sufficient:
57  *
58  * (1) The task queues are very hot and there is a need to avoid data and lock
59  *	contention over global resources.
60  *
61  * (2) Some tasks may depend on other tasks to complete, so they can't be put in
62  *	the same list managed by the same thread.
63  *
64  * (3) Some tasks may block for a long time, and this should not block other
65  *	tasks in the queue.
66  *
67  * To provide useful service in such cases we define a "dynamic task queue"
68  * which has an individual thread for each of the tasks. These threads are
69  * dynamically created as they are needed and destroyed when they are not in
70  * use. The API for managing task pools is the same as for managing task queues
71  * with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
72  * dynamic task pool behavior is desired.
73  *
74  * Dynamic task queues may also place tasks in the normal queue (called "backing
75  * queue") when task pool runs out of resources. Users of task queues may
76  * disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
77  * flags.
78  *
79  * The backing task queue is also used for scheduling internal tasks needed for
80  * dynamic task queue maintenance.
81  *
82  * INTERFACES ==================================================================
83  *
84  * taskq_t *taskq_create(name, nthreads, pri, minalloc, maxalloc, flags);
85  *
86  *	Create a taskq with specified properties.
87  *	Possible 'flags':
88  *
89  *	  TASKQ_DYNAMIC: Create task pool for task management. If this flag is
90  *		specified, 'nthreads' specifies the maximum number of threads in
91  *		the task queue. Task execution order for dynamic task queues is
92  *		not predictable.
93  *
94  *		If this flag is not specified (default case) a
95  *		single-list task queue is created with 'nthreads' threads
96  *		servicing it. Entries in this queue are managed by
97  *		taskq_ent_alloc() and taskq_ent_free() which try to keep the
98  *		task population between 'minalloc' and 'maxalloc', but the
99  *		latter limit is only advisory for TQ_SLEEP dispatches and the
100  *		former limit is only advisory for TQ_NOALLOC dispatches. If
101  *		TASKQ_PREPOPULATE is set in 'flags', the taskq will be
102  *		prepopulated with 'minalloc' task structures.
103  *
104  *		Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
105  *		executed in the order they are scheduled if nthreads == 1.
106  *		If nthreads > 1, task execution order is not predictable.
107  *
108  *	  TASKQ_PREPOPULATE: Prepopulate task queue with threads.
109  *		Also prepopulate the task queue with 'minalloc' task structures.
110  *
111  *	  TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
112  *		interpreted as a percentage of the # of online CPUs on the
113  *		system.  The taskq subsystem will automatically adjust the
114  *		number of threads in the taskq in response to CPU online
115  *		and offline events, to keep the ratio.  nthreads must be in
116  *		the range [0,100].
117  *
118  *		The calculation used is:
119  *
120  *			MAX((ncpus_online * percentage)/100, 1)
121  *
122  *		This flag is not supported for DYNAMIC task queues.
123  *		This flag is not compatible with TASKQ_CPR_SAFE.
124  *
125  *	  TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
126  *		use their own protocol for handling CPR issues. This flag is not
127  *		supported for DYNAMIC task queues.  This flag is not compatible
128  *		with TASKQ_THREADS_CPU_PCT.
129  *
130  *	The 'pri' field specifies the default priority for the threads that
131  *	service all scheduled tasks.
132  *
133  * taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
134  *    maxalloc, flags);
135  *
136  *	Like taskq_create(), but takes an instance number (or -1 to indicate
137  *	no instance).
138  *
139  * taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxalloc, proc,
140  *    flags);
141  *
142  *	Like taskq_create(), but creates the taskq threads in the specified
143  *	system process.  If proc != &p0, this must be called from a thread
144  *	in that process.
145  *
146  * taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxalloc, proc,
147  *    dc, flags);
148  *
149  *	Like taskq_create_proc(), but the taskq threads will use the
150  *	System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
151  *
152  * void taskq_destroy(tap):
153  *
154  *	Waits for any scheduled tasks to complete, then destroys the taskq.
155  *	Caller should guarantee that no new tasks are scheduled in the closing
156  *	taskq.
157  *
158  * taskqid_t taskq_dispatch(tq, func, arg, flags):
159  *
160  *	Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
161  *	the caller is willing to block for memory.  The function returns an
162  *	opaque value which is zero iff dispatch fails.  If flags is TQ_NOSLEEP
163  *	or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
164  *	and returns TASKQID_INVALID.
165  *
166  *	ASSUMES: func != NULL.
167  *
168  *	Possible flags:
169  *	  TQ_NOSLEEP: Do not wait for resources; may fail.
170  *
171  *	  TQ_NOALLOC: Do not allocate memory; may fail.  May only be used with
172  *		non-dynamic task queues.
173  *
174  *	  TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
175  *		lack of available resources and fail. If this flag is not
176  *		set, and the task pool is exhausted, the task may be scheduled
177  *		in the backing queue. This flag may ONLY be used with dynamic
178  *		task queues.
179  *
180  *		NOTE: This flag should always be used when a task queue is used
181  *		for tasks that may depend on each other for completion.
182  *		Enqueueing dependent tasks may create deadlocks.
183  *
184  *	  TQ_SLEEP:   May block waiting for resources. May still fail for
185  *		dynamic task queues if TQ_NOQUEUE is also specified, otherwise
186  *		always succeed.
187  *
188  *	  TQ_FRONT:   Puts the new task at the front of the queue.  Be careful.
189  *
190  *	NOTE: Dynamic task queues are much more likely to fail in
191  *		taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
192  *		is important to have backup strategies handling such failures.
193  *
194  * void taskq_dispatch_ent(tq, func, arg, flags, tqent)
195  *
196  *	This is a light-weight form of taskq_dispatch(), that uses a
197  *	preallocated taskq_ent_t structure for scheduling.  As a
198  *	result, it does not perform allocations and cannot ever fail.
199  *	Note especially that it cannot be used with TASKQ_DYNAMIC
200  *	taskqs.  The memory for the tqent must not be modified or used
201  *	until the function (func) is called.  (However, func itself
202  *	may safely modify or free this memory, once it is called.)
203  *	Note that the taskq framework will NOT free this memory.
204  *
205  * boolean_t taskq_empty(tq)
206  *
207  *	Queries if there are tasks pending on the queue.
208  *
209  * void taskq_wait(tq):
210  *
211  *	Waits for all previously scheduled tasks to complete.
212  *
213  *	NOTE: It does not stop any new task dispatches.
214  *	      Do NOT call taskq_wait() from a task: it will cause deadlock.
215  *
216  * void taskq_suspend(tq)
217  *
218  *	Suspend all task execution. Tasks already scheduled for a dynamic task
219  *	queue will still be executed, but all new scheduled tasks will be
220  *	suspended until taskq_resume() is called.
221  *
222  * int  taskq_suspended(tq)
223  *
224  *	Returns 1 if taskq is suspended and 0 otherwise. It is intended to
225  *	ASSERT that the task queue is suspended.
226  *
227  * void taskq_resume(tq)
228  *
229  *	Resume task queue execution.
230  *
231  * int  taskq_member(tq, thread)
232  *
233  *	Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
234  *	intended use is to ASSERT that a given function is called in taskq
235  *	context only.
236  *
237  * system_taskq
238  *
239  *	Global system-wide dynamic task queue for common uses. It may be used by
240  *	any subsystem that needs to schedule tasks and does not need to manage
241  *	its own task queues. It is initialized quite early during system boot.
242  *
243  * IMPLEMENTATION ==============================================================
244  *
245  * This is schematic representation of the task queue structures.
246  *
247  *   taskq:
248  *   +-------------+
249  *   | tq_lock     | +---< taskq_ent_free()
250  *   +-------------+ |
251  *   |...          | | tqent:                  tqent:
252  *   +-------------+ | +------------+          +------------+
253  *   | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
254  *   +-------------+   +------------+          +------------+
255  *   |...          |   | ...        |          | ...        |
256  *   +-------------+   +------------+          +------------+
257  *   | tq_task     |    |
258  *   |             |    +-------------->taskq_ent_alloc()
259  * +--------------------------------------------------------------------------+
260  * | |                     |            tqent                   tqent         |
261  * | +---------------------+     +--> +------------+     +--> +------------+  |
262  * | | ...		   |     |    | func, arg  |     |    | func, arg  |  |
263  * +>+---------------------+ <---|-+  +------------+ <---|-+  +------------+  |
264  *   | tq_taskq.tqent_next | ----+ |  | tqent_next | --->+ |  | tqent_next |--+
265  *   +---------------------+	   |  +------------+     ^ |  +------------+
266  * +-| tq_task.tqent_prev  |	   +--| tqent_prev |     | +--| tqent_prev |  ^
267  * | +---------------------+	      +------------+     |    +------------+  |
268  * | |...		   |	      | ...        |     |    | ...        |  |
269  * | +---------------------+	      +------------+     |    +------------+  |
270  * |                                      ^              |                    |
271  * |                                      |              |                    |
272  * +--------------------------------------+--------------+       TQ_APPEND() -+
273  *   |             |                      |
274  *   |...          |   taskq_thread()-----+
275  *   +-------------+
276  *   | tq_buckets  |--+-------> [ NULL ] (for regular task queues)
277  *   +-------------+  |
278  *                    |   DYNAMIC TASK QUEUES:
279  *                    |
280  *                    +-> taskq_bucket[nCPU]		taskq_bucket_dispatch()
281  *                        +-------------------+                    ^
282  *                   +--->| tqbucket_lock     |                    |
283  *                   |    +-------------------+   +--------+      +--------+
284  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  | ^
285  *                   |    +-------------------+<--+--------+<--...+--------+ |
286  *                   |    | ...               |   | thread |      | thread | |
287  *                   |    +-------------------+   +--------+      +--------+ |
288  *                   |    +-------------------+                              |
289  * taskq_dispatch()--+--->| tqbucket_lock     |             TQ_APPEND()------+
290  *      TQ_HASH()    |    +-------------------+   +--------+      +--------+
291  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  |
292  *                   |    +-------------------+<--+--------+<--...+--------+
293  *                   |    | ...               |   | thread |      | thread |
294  *                   |    +-------------------+   +--------+      +--------+
295  *		     +--->	...
296  *
297  *
298  * Task queues use tq_task field to link new entry in the queue. The queue is a
299  * circular doubly-linked list. Entries are put in the end of the list with
300  * TQ_APPEND() and processed from the front of the list by taskq_thread() in
301  * FIFO order. Task queue entries are cached in the free list managed by
302  * taskq_ent_alloc() and taskq_ent_free() functions.
303  *
304  *	All threads used by task queues mark t_taskq field of the thread to
305  *	point to the task queue.
306  *
307  * Taskq Thread Management -----------------------------------------------------
308  *
309  * Taskq's non-dynamic threads are managed with several variables and flags:
310  *
311  *	* tq_nthreads	- The number of threads in taskq_thread() for the
312  *			  taskq.
313  *
314  *	* tq_active	- The number of threads not waiting on a CV in
315  *			  taskq_thread(); includes newly created threads
316  *			  not yet counted in tq_nthreads.
317  *
318  *	* tq_nthreads_target
319  *			- The number of threads desired for the taskq.
320  *
321  *	* tq_flags & TASKQ_CHANGING
322  *			- Indicates that tq_nthreads != tq_nthreads_target.
323  *
324  *	* tq_flags & TASKQ_THREAD_CREATED
325  *			- Indicates that a thread is being created in the taskq.
326  *
327  * During creation, tq_nthreads and tq_active are set to 0, and
328  * tq_nthreads_target is set to the number of threads desired.  The
329  * TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
330  * create the first thread. taskq_thread_create() increments tq_active,
331  * sets TASKQ_THREAD_CREATED, and creates the new thread.
332  *
333  * Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
334  * flag, and increments tq_nthreads.  It stores the new value of
335  * tq_nthreads as its "thread_id", and stores its thread pointer in the
336  * tq_threadlist at the (thread_id - 1).  We keep the thread_id space
337  * densely packed by requiring that only the largest thread_id can exit during
338  * normal adjustment.   The exception is during the destruction of the
339  * taskq; once tq_nthreads_target is set to zero, no new threads will be created
340  * for the taskq queue, so every thread can exit without any ordering being
341  * necessary.
342  *
343  * Threads will only process work if their thread id is <= tq_nthreads_target.
344  *
345  * When TASKQ_CHANGING is set, threads will check the current thread target
346  * whenever they wake up, and do whatever they can to apply its effects.
347  *
348  * TASKQ_THREAD_CPU_PCT --------------------------------------------------------
349  *
350  * When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
351  * percentage in tq_threads_ncpus_pct, start them off with the correct thread
352  * target, and add them to the taskq_cpupct_list for later adjustment.
353  *
354  * We register taskq_cpu_setup() to be called whenever a CPU changes state.  It
355  * walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
356  * if need be, and wakes up all of the threads to process the change.
357  *
358  * Dynamic Task Queues Implementation ------------------------------------------
359  *
360  * For a dynamic task queues there is a 1-to-1 mapping between a thread and
361  * taskq_ent_structure. Each entry is serviced by its own thread and each thread
362  * is controlled by a single entry.
363  *
364  * Entries are distributed over a set of buckets. To avoid using modulo
365  * arithmetics the number of buckets is 2^n and is determined as the nearest
366  * power of two roundown of the number of CPUs in the system. Tunable
367  * variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
368  * is attached to a bucket for its lifetime and can't migrate to other buckets.
369  *
370  * Entries that have scheduled tasks are not placed in any list. The dispatch
371  * function sets their "func" and "arg" fields and signals the corresponding
372  * thread to execute the task. Once the thread executes the task it clears the
373  * "func" field and places an entry on the bucket cache of free entries pointed
374  * by "tqbucket_freelist" field. ALL entries on the free list should have "func"
375  * field equal to NULL. The free list is a circular doubly-linked list identical
376  * in structure to the tq_task list above, but entries are taken from it in LIFO
377  * order - the last freed entry is the first to be allocated. The
378  * taskq_bucket_dispatch() function gets the most recently used entry from the
379  * free list, sets its "func" and "arg" fields and signals a worker thread.
380  *
381  * After executing each task a per-entry thread taskq_d_thread() places its
382  * entry on the bucket free list and goes to a timed sleep. If it wakes up
383  * without getting new task it removes the entry from the free list and destroys
384  * itself. The thread sleep time is controlled by a tunable variable
385  * `taskq_thread_timeout'.
386  *
387  * There are various statistics kept in the bucket which allows for later
388  * analysis of taskq usage patterns. Also, a global copy of taskq creation and
389  * death statistics is kept in the global taskq data structure. Since thread
390  * creation and death happen rarely, updating such global data does not present
391  * a performance problem.
392  *
393  * NOTE: Threads are not bound to any CPU and there is absolutely no association
394  *       between the bucket and actual thread CPU, so buckets are used only to
395  *	 split resources and reduce resource contention. Having threads attached
396  *	 to the CPU denoted by a bucket may reduce number of times the job
397  *	 switches between CPUs.
398  *
399  *	 Current algorithm creates a thread whenever a bucket has no free
400  *	 entries. It would be nice to know how many threads are in the running
401  *	 state and don't create threads if all CPUs are busy with existing
402  *	 tasks, but it is unclear how such strategy can be implemented.
403  *
404  *	 Currently buckets are created statically as an array attached to task
405  *	 queue. On some system with nCPUs < max_ncpus it may waste system
406  *	 memory. One solution may be allocation of buckets when they are first
407  *	 touched, but it is not clear how useful it is.
408  *
409  * SUSPEND/RESUME implementation -----------------------------------------------
410  *
411  *	Before executing a task taskq_thread() (executing non-dynamic task
412  *	queues) obtains taskq's thread lock as a reader. The taskq_suspend()
413  *	function gets the same lock as a writer blocking all non-dynamic task
414  *	execution. The taskq_resume() function releases the lock allowing
415  *	taskq_thread to continue execution.
416  *
417  *	For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
418  *	taskq_suspend() function. After that taskq_bucket_dispatch() always
419  *	fails, so that taskq_dispatch() will either enqueue tasks for a
420  *	suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
421  *	flags.
422  *
423  *	NOTE: taskq_suspend() does not immediately block any tasks already
424  *	      scheduled for dynamic task queues. It only suspends new tasks
425  *	      scheduled after taskq_suspend() was called.
426  *
427  *	taskq_member() function works by comparing a thread t_taskq pointer with
428  *	the passed thread pointer.
429  *
430  * LOCKS and LOCK Hierarchy ----------------------------------------------------
431  *
432  *   There are three locks used in task queues:
433  *
434  *   1) The taskq_t's tq_lock, protecting global task queue state.
435  *
436  *   2) Each per-CPU bucket has a lock for bucket management.
437  *
438  *   3) The global taskq_cpupct_lock, which protects the list of
439  *      TASKQ_THREADS_CPU_PCT taskqs.
440  *
441  *   If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
442  *   lock.
443  *
444  *   If both (1) and (3) are needed, tq_lock should be taken *after*
445  *   taskq_cpupct_lock.
446  *
447  * DEBUG FACILITIES ------------------------------------------------------------
448  *
449  * For DEBUG kernels it is possible to induce random failures to
450  * taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
451  * taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
452  * failures for dynamic and static task queues respectively.
453  *
454  * Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
455  *
456  * TUNABLES --------------------------------------------------------------------
457  *
458  *	system_taskq_size	- Size of the global system_taskq.
459  *				  This value is multiplied by nCPUs to determine
460  *				  actual size.
461  *				  Default value: 64
462  *
463  *	taskq_minimum_nthreads_max
464  *				- Minimum size of the thread list for a taskq.
465  *				  Useful for testing different thread pool
466  *				  sizes by overwriting tq_nthreads_target.
467  *
468  *	taskq_thread_timeout	- Maximum idle time for taskq_d_thread()
469  *				  Default value: 5 minutes
470  *
471  *	taskq_maxbuckets	- Maximum number of buckets in any task queue
472  *				  Default value: 128
473  *
474  *	taskq_search_depth	- Maximum # of buckets searched for a free entry
475  *				  Default value: 4
476  *
477  *	taskq_dmtbf		- Mean time between induced dispatch failures
478  *				  for dynamic task queues.
479  *				  Default value: UINT_MAX (no induced failures)
480  *
481  *	taskq_smtbf		- Mean time between induced dispatch failures
482  *				  for static task queues.
483  *				  Default value: UINT_MAX (no induced failures)
484  *
485  * CONDITIONAL compilation -----------------------------------------------------
486  *
487  *    TASKQ_STATISTIC	- If set will enable bucket statistic (default).
488  *
489  */
490 
491 #include <sys/taskq_impl.h>
492 #include <sys/thread.h>
493 #include <sys/proc.h>
494 #include <sys/kmem.h>
495 #include <sys/vmem.h>
496 #include <sys/callb.h>
497 #include <sys/class.h>
498 #include <sys/systm.h>
499 #include <sys/cmn_err.h>
500 #include <sys/debug.h>
501 #include <sys/vmsystm.h>	/* For throttlefree */
502 #include <sys/sysmacros.h>
503 #include <sys/cpuvar.h>
504 #include <sys/cpupart.h>
505 #include <sys/sdt.h>
506 #include <sys/sysdc.h>
507 #include <sys/note.h>
508 
509 static kmem_cache_t *taskq_ent_cache, *taskq_cache;
510 
511 /*
512  * Pseudo instance numbers for taskqs without explicitly provided instance.
513  */
514 static vmem_t *taskq_id_arena;
515 
516 /* Global system task queue for common use */
517 taskq_t	*system_taskq;
518 
519 /*
520  * Maximum number of entries in global system taskq is
521  *	system_taskq_size * max_ncpus
522  */
523 #define	SYSTEM_TASKQ_SIZE 64
524 int system_taskq_size = SYSTEM_TASKQ_SIZE;
525 
526 /*
527  * Minimum size for tq_nthreads_max; useful for those who want to play around
528  * with increasing a taskq's tq_nthreads_target.
529  */
530 int taskq_minimum_nthreads_max = 1;
531 
532 /*
533  * We want to ensure that when taskq_create() returns, there is at least
534  * one thread ready to handle requests.  To guarantee this, we have to wait
535  * for the second thread, since the first one cannot process requests until
536  * the second thread has been created.
537  */
538 #define	TASKQ_CREATE_ACTIVE_THREADS	2
539 
540 /* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
541 #define	TASKQ_CPUPCT_MAX_PERCENT	1000
542 int taskq_cpupct_max_percent = TASKQ_CPUPCT_MAX_PERCENT;
543 
544 /*
545  * Dynamic task queue threads that don't get any work within
546  * taskq_thread_timeout destroy themselves
547  */
548 #define	TASKQ_THREAD_TIMEOUT (60 * 5)
549 int taskq_thread_timeout = TASKQ_THREAD_TIMEOUT;
550 
551 #define	TASKQ_MAXBUCKETS 128
552 int taskq_maxbuckets = TASKQ_MAXBUCKETS;
553 
554 /*
555  * When a bucket has no available entries another buckets are tried.
556  * taskq_search_depth parameter limits the amount of buckets that we search
557  * before failing. This is mostly useful in systems with many CPUs where we may
558  * spend too much time scanning busy buckets.
559  */
560 #define	TASKQ_SEARCH_DEPTH 4
561 int taskq_search_depth = TASKQ_SEARCH_DEPTH;
562 
563 /*
564  * Hashing function: mix various bits of x. May be pretty much anything.
565  */
566 #define	TQ_HASH(x) ((x) ^ ((x) >> 11) ^ ((x) >> 17) ^ ((x) ^ 27))
567 
568 /*
569  * We do not create any new threads when the system is low on memory and start
570  * throttling memory allocations. The following macro tries to estimate such
571  * condition.
572  */
573 #define	ENOUGH_MEMORY() (freemem > throttlefree)
574 
575 /*
576  * Static functions.
577  */
578 static taskq_t	*taskq_create_common(const char *, int, int, pri_t, int,
579     int, proc_t *, uint_t, uint_t);
580 static void taskq_thread(void *);
581 static void taskq_d_thread(taskq_ent_t *);
582 static void taskq_bucket_extend(void *);
583 static int  taskq_constructor(void *, void *, int);
584 static void taskq_destructor(void *, void *);
585 static int  taskq_ent_constructor(void *, void *, int);
586 static void taskq_ent_destructor(void *, void *);
587 static taskq_ent_t *taskq_ent_alloc(taskq_t *, int);
588 static void taskq_ent_free(taskq_t *, taskq_ent_t *);
589 static int taskq_ent_exists(taskq_t *, task_func_t, void *);
590 static taskq_ent_t *taskq_bucket_dispatch(taskq_bucket_t *, task_func_t,
591     void *);
592 
593 /*
594  * Task queues kstats.
595  */
596 struct taskq_kstat {
597 	kstat_named_t	tq_pid;
598 	kstat_named_t	tq_tasks;
599 	kstat_named_t	tq_executed;
600 	kstat_named_t	tq_maxtasks;
601 	kstat_named_t	tq_totaltime;
602 	kstat_named_t	tq_nalloc;
603 	kstat_named_t	tq_nactive;
604 	kstat_named_t	tq_pri;
605 	kstat_named_t	tq_nthreads;
606 	kstat_named_t	tq_nomem;
607 } taskq_kstat = {
608 	{ "pid",		KSTAT_DATA_UINT64 },
609 	{ "tasks",		KSTAT_DATA_UINT64 },
610 	{ "executed",		KSTAT_DATA_UINT64 },
611 	{ "maxtasks",		KSTAT_DATA_UINT64 },
612 	{ "totaltime",		KSTAT_DATA_UINT64 },
613 	{ "nalloc",		KSTAT_DATA_UINT64 },
614 	{ "nactive",		KSTAT_DATA_UINT64 },
615 	{ "priority",		KSTAT_DATA_UINT64 },
616 	{ "threads",		KSTAT_DATA_UINT64 },
617 	{ "nomem",		KSTAT_DATA_UINT64 },
618 };
619 
620 struct taskq_d_kstat {
621 	kstat_named_t	tqd_pri;
622 	kstat_named_t	tqd_btasks;
623 	kstat_named_t	tqd_bexecuted;
624 	kstat_named_t	tqd_bmaxtasks;
625 	kstat_named_t	tqd_bnalloc;
626 	kstat_named_t	tqd_bnactive;
627 	kstat_named_t	tqd_btotaltime;
628 	kstat_named_t	tqd_hits;
629 	kstat_named_t	tqd_misses;
630 	kstat_named_t	tqd_overflows;
631 	kstat_named_t	tqd_tcreates;
632 	kstat_named_t	tqd_tdeaths;
633 	kstat_named_t	tqd_maxthreads;
634 	kstat_named_t	tqd_nomem;
635 	kstat_named_t	tqd_disptcreates;
636 	kstat_named_t	tqd_totaltime;
637 	kstat_named_t	tqd_nalloc;
638 	kstat_named_t	tqd_nfree;
639 } taskq_d_kstat = {
640 	{ "priority",		KSTAT_DATA_UINT64 },
641 	{ "btasks",		KSTAT_DATA_UINT64 },
642 	{ "bexecuted",		KSTAT_DATA_UINT64 },
643 	{ "bmaxtasks",		KSTAT_DATA_UINT64 },
644 	{ "bnalloc",		KSTAT_DATA_UINT64 },
645 	{ "bnactive",		KSTAT_DATA_UINT64 },
646 	{ "btotaltime",		KSTAT_DATA_UINT64 },
647 	{ "hits",		KSTAT_DATA_UINT64 },
648 	{ "misses",		KSTAT_DATA_UINT64 },
649 	{ "overflows",		KSTAT_DATA_UINT64 },
650 	{ "tcreates",		KSTAT_DATA_UINT64 },
651 	{ "tdeaths",		KSTAT_DATA_UINT64 },
652 	{ "maxthreads",		KSTAT_DATA_UINT64 },
653 	{ "nomem",		KSTAT_DATA_UINT64 },
654 	{ "disptcreates",	KSTAT_DATA_UINT64 },
655 	{ "totaltime",		KSTAT_DATA_UINT64 },
656 	{ "nalloc",		KSTAT_DATA_UINT64 },
657 	{ "nfree",		KSTAT_DATA_UINT64 },
658 };
659 
660 static kmutex_t taskq_kstat_lock;
661 static kmutex_t taskq_d_kstat_lock;
662 static int taskq_kstat_update(kstat_t *, int);
663 static int taskq_d_kstat_update(kstat_t *, int);
664 
665 /*
666  * List of all TASKQ_THREADS_CPU_PCT taskqs.
667  */
668 static list_t taskq_cpupct_list;	/* protected by cpu_lock */
669 
670 /*
671  * Collect per-bucket statistic when TASKQ_STATISTIC is defined.
672  */
673 #define	TASKQ_STATISTIC 1
674 
675 #if TASKQ_STATISTIC
676 #define	TQ_STAT(b, x)	b->tqbucket_stat.x++
677 #else
678 #define	TQ_STAT(b, x)
679 #endif
680 
681 /*
682  * Random fault injection.
683  */
684 uint_t taskq_random;
685 uint_t taskq_dmtbf = UINT_MAX;    /* mean time between injected failures */
686 uint_t taskq_smtbf = UINT_MAX;    /* mean time between injected failures */
687 
688 /*
689  * TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
690  *
691  * TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
692  * they could prepopulate the cache and make sure that they do not use more
693  * then minalloc entries.  So, fault injection in this case insures that
694  * either TASKQ_PREPOPULATE is not set or there are more entries allocated
695  * than is specified by minalloc.  TQ_NOALLOC dispatches are always allowed
696  * to fail, but for simplicity we treat them identically to TQ_NOSLEEP
697  * dispatches.
698  */
699 #ifdef DEBUG
700 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)		\
701 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
702 	if ((flag & TQ_NOSLEEP) &&				\
703 	    taskq_random < 1771875 / taskq_dmtbf) {		\
704 		return (TASKQID_INVALID);			\
705 	}
706 
707 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)		\
708 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
709 	if ((flag & (TQ_NOSLEEP | TQ_NOALLOC)) &&		\
710 	    (!(tq->tq_flags & TASKQ_PREPOPULATE) ||		\
711 	    (tq->tq_nalloc > tq->tq_minalloc)) &&		\
712 	    (taskq_random < (1771875 / taskq_smtbf))) {		\
713 		mutex_exit(&tq->tq_lock);			\
714 		return (TASKQID_INVALID);			\
715 	}
716 #else
717 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)
718 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)
719 #endif
720 
721 #define	IS_EMPTY(l) (((l).tqent_prev == (l).tqent_next) &&	\
722 	((l).tqent_prev == &(l)))
723 
724 /*
725  * Append `tqe' in the end of the doubly-linked list denoted by l.
726  */
727 #define	TQ_APPEND(l, tqe) {					\
728 	tqe->tqent_next = &l;					\
729 	tqe->tqent_prev = l.tqent_prev;				\
730 	tqe->tqent_next->tqent_prev = tqe;			\
731 	tqe->tqent_prev->tqent_next = tqe;			\
732 }
733 /*
734  * Prepend 'tqe' to the beginning of l
735  */
736 #define	TQ_PREPEND(l, tqe) {					\
737 	tqe->tqent_next = l.tqent_next;				\
738 	tqe->tqent_prev = &l;					\
739 	tqe->tqent_next->tqent_prev = tqe;			\
740 	tqe->tqent_prev->tqent_next = tqe;			\
741 }
742 
743 /*
744  * Schedule a task specified by func and arg into the task queue entry tqe.
745  */
746 #define	TQ_DO_ENQUEUE(tq, tqe, func, arg, front) {			\
747 	ASSERT(MUTEX_HELD(&tq->tq_lock));				\
748 	_NOTE(CONSTCOND)						\
749 	if (front) {							\
750 		TQ_PREPEND(tq->tq_task, tqe);				\
751 	} else {							\
752 		TQ_APPEND(tq->tq_task, tqe);				\
753 	}								\
754 	tqe->tqent_func = (func);					\
755 	tqe->tqent_arg = (arg);						\
756 	tq->tq_tasks++;							\
757 	if (tq->tq_tasks - tq->tq_executed > tq->tq_maxtasks)		\
758 		tq->tq_maxtasks = tq->tq_tasks - tq->tq_executed;	\
759 	cv_signal(&tq->tq_dispatch_cv);					\
760 	DTRACE_PROBE2(taskq__enqueue, taskq_t *, tq, taskq_ent_t *, tqe); \
761 }
762 
763 #define	TQ_ENQUEUE(tq, tqe, func, arg)					\
764 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 0)
765 
766 #define	TQ_ENQUEUE_FRONT(tq, tqe, func, arg)				\
767 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 1)
768 
769 /*
770  * Do-nothing task which may be used to prepopulate thread caches.
771  */
772 /*ARGSUSED*/
773 void
774 nulltask(void *unused)
775 {
776 }
777 
778 /*ARGSUSED*/
779 static int
780 taskq_constructor(void *buf, void *cdrarg, int kmflags)
781 {
782 	taskq_t *tq = buf;
783 
784 	bzero(tq, sizeof (taskq_t));
785 
786 	mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
787 	rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
788 	cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
789 	cv_init(&tq->tq_exit_cv, NULL, CV_DEFAULT, NULL);
790 	cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
791 	cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
792 
793 	tq->tq_task.tqent_next = &tq->tq_task;
794 	tq->tq_task.tqent_prev = &tq->tq_task;
795 
796 	return (0);
797 }
798 
799 /*ARGSUSED*/
800 static void
801 taskq_destructor(void *buf, void *cdrarg)
802 {
803 	taskq_t *tq = buf;
804 
805 	ASSERT(tq->tq_nthreads == 0);
806 	ASSERT(tq->tq_buckets == NULL);
807 	ASSERT(tq->tq_tcreates == 0);
808 	ASSERT(tq->tq_tdeaths == 0);
809 
810 	mutex_destroy(&tq->tq_lock);
811 	rw_destroy(&tq->tq_threadlock);
812 	cv_destroy(&tq->tq_dispatch_cv);
813 	cv_destroy(&tq->tq_exit_cv);
814 	cv_destroy(&tq->tq_wait_cv);
815 	cv_destroy(&tq->tq_maxalloc_cv);
816 }
817 
818 /*ARGSUSED*/
819 static int
820 taskq_ent_constructor(void *buf, void *cdrarg, int kmflags)
821 {
822 	taskq_ent_t *tqe = buf;
823 
824 	tqe->tqent_thread = NULL;
825 	cv_init(&tqe->tqent_cv, NULL, CV_DEFAULT, NULL);
826 
827 	return (0);
828 }
829 
830 /*ARGSUSED*/
831 static void
832 taskq_ent_destructor(void *buf, void *cdrarg)
833 {
834 	taskq_ent_t *tqe = buf;
835 
836 	ASSERT(tqe->tqent_thread == NULL);
837 	cv_destroy(&tqe->tqent_cv);
838 }
839 
840 void
841 taskq_init(void)
842 {
843 	taskq_ent_cache = kmem_cache_create("taskq_ent_cache",
844 	    sizeof (taskq_ent_t), 0, taskq_ent_constructor,
845 	    taskq_ent_destructor, NULL, NULL, NULL, 0);
846 	taskq_cache = kmem_cache_create("taskq_cache", sizeof (taskq_t),
847 	    0, taskq_constructor, taskq_destructor, NULL, NULL, NULL, 0);
848 	taskq_id_arena = vmem_create("taskq_id_arena",
849 	    (void *)1, INT32_MAX, 1, NULL, NULL, NULL, 0,
850 	    VM_SLEEP | VMC_IDENTIFIER);
851 
852 	list_create(&taskq_cpupct_list, sizeof (taskq_t),
853 	    offsetof(taskq_t, tq_cpupct_link));
854 }
855 
856 static void
857 taskq_update_nthreads(taskq_t *tq, uint_t ncpus)
858 {
859 	uint_t newtarget = TASKQ_THREADS_PCT(ncpus, tq->tq_threads_ncpus_pct);
860 
861 	ASSERT(MUTEX_HELD(&cpu_lock));
862 	ASSERT(MUTEX_HELD(&tq->tq_lock));
863 
864 	/* We must be going from non-zero to non-zero; no exiting. */
865 	ASSERT3U(tq->tq_nthreads_target, !=, 0);
866 	ASSERT3U(newtarget, !=, 0);
867 
868 	ASSERT3U(newtarget, <=, tq->tq_nthreads_max);
869 	if (newtarget != tq->tq_nthreads_target) {
870 		tq->tq_flags |= TASKQ_CHANGING;
871 		tq->tq_nthreads_target = newtarget;
872 		cv_broadcast(&tq->tq_dispatch_cv);
873 		cv_broadcast(&tq->tq_exit_cv);
874 	}
875 }
876 
877 /* called during task queue creation */
878 static void
879 taskq_cpupct_install(taskq_t *tq, cpupart_t *cpup)
880 {
881 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
882 
883 	mutex_enter(&cpu_lock);
884 	mutex_enter(&tq->tq_lock);
885 	tq->tq_cpupart = cpup->cp_id;
886 	taskq_update_nthreads(tq, cpup->cp_ncpus);
887 	mutex_exit(&tq->tq_lock);
888 
889 	list_insert_tail(&taskq_cpupct_list, tq);
890 	mutex_exit(&cpu_lock);
891 }
892 
893 static void
894 taskq_cpupct_remove(taskq_t *tq)
895 {
896 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
897 
898 	mutex_enter(&cpu_lock);
899 	list_remove(&taskq_cpupct_list, tq);
900 	mutex_exit(&cpu_lock);
901 }
902 
903 /*ARGSUSED*/
904 static int
905 taskq_cpu_setup(cpu_setup_t what, int id, void *arg)
906 {
907 	taskq_t *tq;
908 	cpupart_t *cp = cpu[id]->cpu_part;
909 	uint_t ncpus = cp->cp_ncpus;
910 
911 	ASSERT(MUTEX_HELD(&cpu_lock));
912 	ASSERT(ncpus > 0);
913 
914 	switch (what) {
915 	case CPU_OFF:
916 	case CPU_CPUPART_OUT:
917 		/* offlines are called *before* the cpu is offlined. */
918 		if (ncpus > 1)
919 			ncpus--;
920 		break;
921 
922 	case CPU_ON:
923 	case CPU_CPUPART_IN:
924 		break;
925 
926 	default:
927 		return (0);		/* doesn't affect cpu count */
928 	}
929 
930 	for (tq = list_head(&taskq_cpupct_list); tq != NULL;
931 	    tq = list_next(&taskq_cpupct_list, tq)) {
932 
933 		mutex_enter(&tq->tq_lock);
934 		/*
935 		 * If the taskq is part of the cpuset which is changing,
936 		 * update its nthreads_target.
937 		 */
938 		if (tq->tq_cpupart == cp->cp_id) {
939 			taskq_update_nthreads(tq, ncpus);
940 		}
941 		mutex_exit(&tq->tq_lock);
942 	}
943 	return (0);
944 }
945 
946 void
947 taskq_mp_init(void)
948 {
949 	mutex_enter(&cpu_lock);
950 	register_cpu_setup_func(taskq_cpu_setup, NULL);
951 	/*
952 	 * Make sure we're up to date.  At this point in boot, there is only
953 	 * one processor set, so we only have to update the current CPU.
954 	 */
955 	(void) taskq_cpu_setup(CPU_ON, CPU->cpu_id, NULL);
956 	mutex_exit(&cpu_lock);
957 }
958 
959 /*
960  * Create global system dynamic task queue.
961  */
962 void
963 system_taskq_init(void)
964 {
965 	system_taskq = taskq_create_common("system_taskq", 0,
966 	    system_taskq_size * max_ncpus, minclsyspri, 4, 512, &p0, 0,
967 	    TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
968 }
969 
970 /*
971  * taskq_ent_alloc()
972  *
973  * Allocates a new taskq_ent_t structure either from the free list or from the
974  * cache. Returns NULL if it can't be allocated.
975  *
976  * Assumes: tq->tq_lock is held.
977  */
978 static taskq_ent_t *
979 taskq_ent_alloc(taskq_t *tq, int flags)
980 {
981 	int kmflags = (flags & TQ_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
982 	taskq_ent_t *tqe;
983 	clock_t wait_time;
984 	clock_t	wait_rv;
985 
986 	ASSERT(MUTEX_HELD(&tq->tq_lock));
987 
988 	/*
989 	 * TQ_NOALLOC allocations are allowed to use the freelist, even if
990 	 * we are below tq_minalloc.
991 	 */
992 again:	if ((tqe = tq->tq_freelist) != NULL &&
993 	    ((flags & TQ_NOALLOC) || tq->tq_nalloc >= tq->tq_minalloc)) {
994 		tq->tq_freelist = tqe->tqent_next;
995 	} else {
996 		if (flags & TQ_NOALLOC)
997 			return (NULL);
998 
999 		if (tq->tq_nalloc >= tq->tq_maxalloc) {
1000 			if (kmflags & KM_NOSLEEP)
1001 				return (NULL);
1002 
1003 			/*
1004 			 * We don't want to exceed tq_maxalloc, but we can't
1005 			 * wait for other tasks to complete (and thus free up
1006 			 * task structures) without risking deadlock with
1007 			 * the caller.  So, we just delay for one second
1008 			 * to throttle the allocation rate. If we have tasks
1009 			 * complete before one second timeout expires then
1010 			 * taskq_ent_free will signal us and we will
1011 			 * immediately retry the allocation (reap free).
1012 			 */
1013 			wait_time = ddi_get_lbolt() + hz;
1014 			while (tq->tq_freelist == NULL) {
1015 				tq->tq_maxalloc_wait++;
1016 				wait_rv = cv_timedwait(&tq->tq_maxalloc_cv,
1017 				    &tq->tq_lock, wait_time);
1018 				tq->tq_maxalloc_wait--;
1019 				if (wait_rv == -1)
1020 					break;
1021 			}
1022 			if (tq->tq_freelist)
1023 				goto again;		/* reap freelist */
1024 
1025 		}
1026 		mutex_exit(&tq->tq_lock);
1027 
1028 		tqe = kmem_cache_alloc(taskq_ent_cache, kmflags);
1029 
1030 		mutex_enter(&tq->tq_lock);
1031 		if (tqe != NULL)
1032 			tq->tq_nalloc++;
1033 	}
1034 	return (tqe);
1035 }
1036 
1037 /*
1038  * taskq_ent_free()
1039  *
1040  * Free taskq_ent_t structure by either putting it on the free list or freeing
1041  * it to the cache.
1042  *
1043  * Assumes: tq->tq_lock is held.
1044  */
1045 static void
1046 taskq_ent_free(taskq_t *tq, taskq_ent_t *tqe)
1047 {
1048 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1049 
1050 	if (tq->tq_nalloc <= tq->tq_minalloc) {
1051 		tqe->tqent_next = tq->tq_freelist;
1052 		tq->tq_freelist = tqe;
1053 	} else {
1054 		tq->tq_nalloc--;
1055 		mutex_exit(&tq->tq_lock);
1056 		kmem_cache_free(taskq_ent_cache, tqe);
1057 		mutex_enter(&tq->tq_lock);
1058 	}
1059 
1060 	if (tq->tq_maxalloc_wait)
1061 		cv_signal(&tq->tq_maxalloc_cv);
1062 }
1063 
1064 /*
1065  * taskq_ent_exists()
1066  *
1067  * Return 1 if taskq already has entry for calling 'func(arg)'.
1068  *
1069  * Assumes: tq->tq_lock is held.
1070  */
1071 static int
1072 taskq_ent_exists(taskq_t *tq, task_func_t func, void *arg)
1073 {
1074 	taskq_ent_t	*tqe;
1075 
1076 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1077 
1078 	for (tqe = tq->tq_task.tqent_next; tqe != &tq->tq_task;
1079 	    tqe = tqe->tqent_next)
1080 		if ((tqe->tqent_func == func) && (tqe->tqent_arg == arg))
1081 			return (1);
1082 	return (0);
1083 }
1084 
1085 /*
1086  * Dispatch a task "func(arg)" to a free entry of bucket b.
1087  *
1088  * Assumes: no bucket locks is held.
1089  *
1090  * Returns: a pointer to an entry if dispatch was successful.
1091  *	    NULL if there are no free entries or if the bucket is suspended.
1092  */
1093 static taskq_ent_t *
1094 taskq_bucket_dispatch(taskq_bucket_t *b, task_func_t func, void *arg)
1095 {
1096 	taskq_ent_t *tqe;
1097 
1098 	ASSERT(MUTEX_NOT_HELD(&b->tqbucket_lock));
1099 	ASSERT(func != NULL);
1100 
1101 	mutex_enter(&b->tqbucket_lock);
1102 
1103 	ASSERT(b->tqbucket_nfree != 0 || IS_EMPTY(b->tqbucket_freelist));
1104 	ASSERT(b->tqbucket_nfree == 0 || !IS_EMPTY(b->tqbucket_freelist));
1105 
1106 	/*
1107 	 * Get en entry from the freelist if there is one.
1108 	 * Schedule task into the entry.
1109 	 */
1110 	if ((b->tqbucket_nfree != 0) &&
1111 	    !(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
1112 		tqe = b->tqbucket_freelist.tqent_prev;
1113 
1114 		ASSERT(tqe != &b->tqbucket_freelist);
1115 		ASSERT(tqe->tqent_thread != NULL);
1116 
1117 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1118 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1119 		b->tqbucket_nalloc++;
1120 		b->tqbucket_nfree--;
1121 		tqe->tqent_func = func;
1122 		tqe->tqent_arg = arg;
1123 		TQ_STAT(b, tqs_hits);
1124 		cv_signal(&tqe->tqent_cv);
1125 		DTRACE_PROBE2(taskq__d__enqueue, taskq_bucket_t *, b,
1126 		    taskq_ent_t *, tqe);
1127 	} else {
1128 		tqe = NULL;
1129 		TQ_STAT(b, tqs_misses);
1130 	}
1131 	mutex_exit(&b->tqbucket_lock);
1132 	return (tqe);
1133 }
1134 
1135 /*
1136  * Dispatch a task.
1137  *
1138  * Assumes: func != NULL
1139  *
1140  * Returns: NULL if dispatch failed.
1141  *	    non-NULL if task dispatched successfully.
1142  *	    Actual return value is the pointer to taskq entry that was used to
1143  *	    dispatch a task. This is useful for debugging.
1144  */
1145 taskqid_t
1146 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
1147 {
1148 	taskq_bucket_t *bucket = NULL;	/* Which bucket needs extension */
1149 	taskq_ent_t *tqe = NULL;
1150 	taskq_ent_t *tqe1;
1151 	uint_t bsize;
1152 
1153 	ASSERT(tq != NULL);
1154 	ASSERT(func != NULL);
1155 
1156 	if (!(tq->tq_flags & TASKQ_DYNAMIC)) {
1157 		/*
1158 		 * TQ_NOQUEUE flag can't be used with non-dynamic task queues.
1159 		 */
1160 		ASSERT(!(flags & TQ_NOQUEUE));
1161 		/*
1162 		 * Enqueue the task to the underlying queue.
1163 		 */
1164 		mutex_enter(&tq->tq_lock);
1165 
1166 		TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flags);
1167 
1168 		if ((tqe = taskq_ent_alloc(tq, flags)) == NULL) {
1169 			tq->tq_nomem++;
1170 			mutex_exit(&tq->tq_lock);
1171 			return ((taskqid_t)tqe);
1172 		}
1173 		/* Make sure we start without any flags */
1174 		tqe->tqent_un.tqent_flags = 0;
1175 
1176 		if (flags & TQ_FRONT) {
1177 			TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1178 		} else {
1179 			TQ_ENQUEUE(tq, tqe, func, arg);
1180 		}
1181 		mutex_exit(&tq->tq_lock);
1182 		return ((taskqid_t)tqe);
1183 	}
1184 
1185 	/*
1186 	 * Dynamic taskq dispatching.
1187 	 */
1188 	ASSERT(!(flags & (TQ_NOALLOC | TQ_FRONT)));
1189 	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flags);
1190 
1191 	bsize = tq->tq_nbuckets;
1192 
1193 	if (bsize == 1) {
1194 		/*
1195 		 * In a single-CPU case there is only one bucket, so get
1196 		 * entry directly from there.
1197 		 */
1198 		if ((tqe = taskq_bucket_dispatch(tq->tq_buckets, func, arg))
1199 		    != NULL)
1200 			return ((taskqid_t)tqe);	/* Fastpath */
1201 		bucket = tq->tq_buckets;
1202 	} else {
1203 		int loopcount;
1204 		taskq_bucket_t *b;
1205 		uintptr_t h = ((uintptr_t)CPU + (uintptr_t)arg) >> 3;
1206 
1207 		h = TQ_HASH(h);
1208 
1209 		/*
1210 		 * The 'bucket' points to the original bucket that we hit. If we
1211 		 * can't allocate from it, we search other buckets, but only
1212 		 * extend this one.
1213 		 */
1214 		b = &tq->tq_buckets[h & (bsize - 1)];
1215 		ASSERT(b->tqbucket_taskq == tq);	/* Sanity check */
1216 
1217 		/*
1218 		 * Do a quick check before grabbing the lock. If the bucket does
1219 		 * not have free entries now, chances are very small that it
1220 		 * will after we take the lock, so we just skip it.
1221 		 */
1222 		if (b->tqbucket_nfree != 0) {
1223 			if ((tqe = taskq_bucket_dispatch(b, func, arg)) != NULL)
1224 				return ((taskqid_t)tqe);	/* Fastpath */
1225 		} else {
1226 			TQ_STAT(b, tqs_misses);
1227 		}
1228 
1229 		bucket = b;
1230 		loopcount = MIN(taskq_search_depth, bsize);
1231 		/*
1232 		 * If bucket dispatch failed, search loopcount number of buckets
1233 		 * before we give up and fail.
1234 		 */
1235 		do {
1236 			b = &tq->tq_buckets[++h & (bsize - 1)];
1237 			ASSERT(b->tqbucket_taskq == tq);  /* Sanity check */
1238 			loopcount--;
1239 
1240 			if (b->tqbucket_nfree != 0) {
1241 				tqe = taskq_bucket_dispatch(b, func, arg);
1242 			} else {
1243 				TQ_STAT(b, tqs_misses);
1244 			}
1245 		} while ((tqe == NULL) && (loopcount > 0));
1246 	}
1247 
1248 	/*
1249 	 * At this point we either scheduled a task and (tqe != NULL) or failed
1250 	 * (tqe == NULL). Try to recover from fails.
1251 	 */
1252 
1253 	/*
1254 	 * For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
1255 	 */
1256 	if ((tqe == NULL) && !(flags & TQ_NOSLEEP)) {
1257 		/*
1258 		 * taskq_bucket_extend() may fail to do anything, but this is
1259 		 * fine - we deal with it later. If the bucket was successfully
1260 		 * extended, there is a good chance that taskq_bucket_dispatch()
1261 		 * will get this new entry, unless someone is racing with us and
1262 		 * stealing the new entry from under our nose.
1263 		 * taskq_bucket_extend() may sleep.
1264 		 */
1265 		taskq_bucket_extend(bucket);
1266 		TQ_STAT(bucket, tqs_disptcreates);
1267 		if ((tqe = taskq_bucket_dispatch(bucket, func, arg)) != NULL)
1268 			return ((taskqid_t)tqe);
1269 	}
1270 
1271 	ASSERT(bucket != NULL);
1272 
1273 	/*
1274 	 * Since there are not enough free entries in the bucket, add a
1275 	 * taskq entry to extend it in the background using backing queue
1276 	 * (unless we already have a taskq entry to perform that extension).
1277 	 */
1278 	mutex_enter(&tq->tq_lock);
1279 	if (!taskq_ent_exists(tq, taskq_bucket_extend, bucket)) {
1280 		if ((tqe1 = taskq_ent_alloc(tq, TQ_NOSLEEP)) != NULL) {
1281 			TQ_ENQUEUE_FRONT(tq, tqe1, taskq_bucket_extend, bucket);
1282 		} else {
1283 			tq->tq_nomem++;
1284 		}
1285 	}
1286 
1287 	/*
1288 	 * Dispatch failed and we can't find an entry to schedule a task.
1289 	 * Revert to the backing queue unless TQ_NOQUEUE was asked.
1290 	 */
1291 	if ((tqe == NULL) && !(flags & TQ_NOQUEUE)) {
1292 		if ((tqe = taskq_ent_alloc(tq, flags)) != NULL) {
1293 			TQ_ENQUEUE(tq, tqe, func, arg);
1294 		} else {
1295 			tq->tq_nomem++;
1296 		}
1297 	}
1298 	mutex_exit(&tq->tq_lock);
1299 
1300 	return ((taskqid_t)tqe);
1301 }
1302 
1303 void
1304 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
1305     taskq_ent_t *tqe)
1306 {
1307 	ASSERT(func != NULL);
1308 	ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
1309 
1310 	/*
1311 	 * Mark it as a prealloc'd task.  This is important
1312 	 * to ensure that we don't free it later.
1313 	 */
1314 	tqe->tqent_un.tqent_flags |= TQENT_FLAG_PREALLOC;
1315 	/*
1316 	 * Enqueue the task to the underlying queue.
1317 	 */
1318 	mutex_enter(&tq->tq_lock);
1319 
1320 	if (flags & TQ_FRONT) {
1321 		TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1322 	} else {
1323 		TQ_ENQUEUE(tq, tqe, func, arg);
1324 	}
1325 	mutex_exit(&tq->tq_lock);
1326 }
1327 
1328 /*
1329  * Allow our caller to ask if there are tasks pending on the queue.
1330  */
1331 boolean_t
1332 taskq_empty(taskq_t *tq)
1333 {
1334 	boolean_t rv;
1335 
1336 	ASSERT3P(tq, !=, curthread->t_taskq);
1337 	mutex_enter(&tq->tq_lock);
1338 	rv = (tq->tq_task.tqent_next == &tq->tq_task) && (tq->tq_active == 0);
1339 	mutex_exit(&tq->tq_lock);
1340 
1341 	return (rv);
1342 }
1343 
1344 /*
1345  * Wait for all pending tasks to complete.
1346  * Calling taskq_wait from a task will cause deadlock.
1347  */
1348 void
1349 taskq_wait(taskq_t *tq)
1350 {
1351 	ASSERT(tq != curthread->t_taskq);
1352 
1353 	mutex_enter(&tq->tq_lock);
1354 	while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
1355 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1356 	mutex_exit(&tq->tq_lock);
1357 
1358 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1359 		taskq_bucket_t *b = tq->tq_buckets;
1360 		int bid = 0;
1361 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1362 			mutex_enter(&b->tqbucket_lock);
1363 			while (b->tqbucket_nalloc > 0)
1364 				cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
1365 			mutex_exit(&b->tqbucket_lock);
1366 		}
1367 	}
1368 }
1369 
1370 void
1371 taskq_wait_id(taskq_t *tq, taskqid_t id __unused)
1372 {
1373 	taskq_wait(tq);
1374 }
1375 
1376 /*
1377  * Suspend execution of tasks.
1378  *
1379  * Tasks in the queue part will be suspended immediately upon return from this
1380  * function. Pending tasks in the dynamic part will continue to execute, but all
1381  * new tasks will  be suspended.
1382  */
1383 void
1384 taskq_suspend(taskq_t *tq)
1385 {
1386 	rw_enter(&tq->tq_threadlock, RW_WRITER);
1387 
1388 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1389 		taskq_bucket_t *b = tq->tq_buckets;
1390 		int bid = 0;
1391 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1392 			mutex_enter(&b->tqbucket_lock);
1393 			b->tqbucket_flags |= TQBUCKET_SUSPEND;
1394 			mutex_exit(&b->tqbucket_lock);
1395 		}
1396 	}
1397 	/*
1398 	 * Mark task queue as being suspended. Needed for taskq_suspended().
1399 	 */
1400 	mutex_enter(&tq->tq_lock);
1401 	ASSERT(!(tq->tq_flags & TASKQ_SUSPENDED));
1402 	tq->tq_flags |= TASKQ_SUSPENDED;
1403 	mutex_exit(&tq->tq_lock);
1404 }
1405 
1406 /*
1407  * returns: 1 if tq is suspended, 0 otherwise.
1408  */
1409 int
1410 taskq_suspended(taskq_t *tq)
1411 {
1412 	return ((tq->tq_flags & TASKQ_SUSPENDED) != 0);
1413 }
1414 
1415 /*
1416  * Resume taskq execution.
1417  */
1418 void
1419 taskq_resume(taskq_t *tq)
1420 {
1421 	ASSERT(RW_WRITE_HELD(&tq->tq_threadlock));
1422 
1423 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1424 		taskq_bucket_t *b = tq->tq_buckets;
1425 		int bid = 0;
1426 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1427 			mutex_enter(&b->tqbucket_lock);
1428 			b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
1429 			mutex_exit(&b->tqbucket_lock);
1430 		}
1431 	}
1432 	mutex_enter(&tq->tq_lock);
1433 	ASSERT(tq->tq_flags & TASKQ_SUSPENDED);
1434 	tq->tq_flags &= ~TASKQ_SUSPENDED;
1435 	mutex_exit(&tq->tq_lock);
1436 
1437 	rw_exit(&tq->tq_threadlock);
1438 }
1439 
1440 int
1441 taskq_member(taskq_t *tq, kthread_t *thread)
1442 {
1443 	return (thread->t_taskq == tq);
1444 }
1445 
1446 /*
1447  * Creates a thread in the taskq.  We only allow one outstanding create at
1448  * a time.  We drop and reacquire the tq_lock in order to avoid blocking other
1449  * taskq activity while thread_create() or lwp_kernel_create() run.
1450  *
1451  * The first time we're called, we do some additional setup, and do not
1452  * return until there are enough threads to start servicing requests.
1453  */
1454 static void
1455 taskq_thread_create(taskq_t *tq)
1456 {
1457 	kthread_t	*t;
1458 	const boolean_t	first = (tq->tq_nthreads == 0);
1459 
1460 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1461 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1462 	ASSERT(tq->tq_nthreads < tq->tq_nthreads_target);
1463 	ASSERT(!(tq->tq_flags & TASKQ_THREAD_CREATED));
1464 
1465 
1466 	tq->tq_flags |= TASKQ_THREAD_CREATED;
1467 	tq->tq_active++;
1468 	mutex_exit(&tq->tq_lock);
1469 
1470 	/*
1471 	 * With TASKQ_DUTY_CYCLE the new thread must have an LWP
1472 	 * as explained in ../disp/sysdc.c (for the msacct data).
1473 	 * Normally simple kthreads are preferred, unless the
1474 	 * caller has asked for LWPs for other reasons.
1475 	 */
1476 	if ((tq->tq_flags & (TASKQ_DUTY_CYCLE | TASKQ_THREADS_LWP)) != 0) {
1477 		/* Enforced in taskq_create_common */
1478 		ASSERT3P(tq->tq_proc, !=, &p0);
1479 		t = lwp_kernel_create(tq->tq_proc, taskq_thread, tq, TS_RUN,
1480 		    tq->tq_pri);
1481 	} else {
1482 		t = thread_create(NULL, 0, taskq_thread, tq, 0, tq->tq_proc,
1483 		    TS_RUN, tq->tq_pri);
1484 	}
1485 
1486 	if (!first) {
1487 		mutex_enter(&tq->tq_lock);
1488 		return;
1489 	}
1490 
1491 	/*
1492 	 * We know the thread cannot go away, since tq cannot be
1493 	 * destroyed until creation has completed.  We can therefore
1494 	 * safely dereference t.
1495 	 */
1496 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1497 		taskq_cpupct_install(tq, t->t_cpupart);
1498 	}
1499 	mutex_enter(&tq->tq_lock);
1500 
1501 	/* Wait until we can service requests. */
1502 	while (tq->tq_nthreads != tq->tq_nthreads_target &&
1503 	    tq->tq_nthreads < TASKQ_CREATE_ACTIVE_THREADS) {
1504 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1505 	}
1506 }
1507 
1508 /*
1509  * Common "sleep taskq thread" function, which handles CPR stuff, as well
1510  * as giving a nice common point for debuggers to find inactive threads.
1511  */
1512 static clock_t
1513 taskq_thread_wait(taskq_t *tq, kmutex_t *mx, kcondvar_t *cv,
1514     callb_cpr_t *cprinfo, clock_t timeout)
1515 {
1516 	clock_t ret = 0;
1517 
1518 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1519 		CALLB_CPR_SAFE_BEGIN(cprinfo);
1520 	}
1521 	if (timeout < 0)
1522 		cv_wait(cv, mx);
1523 	else
1524 		ret = cv_reltimedwait(cv, mx, timeout, TR_CLOCK_TICK);
1525 
1526 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1527 		CALLB_CPR_SAFE_END(cprinfo, mx);
1528 	}
1529 
1530 	return (ret);
1531 }
1532 
1533 /*
1534  * Worker thread for processing task queue.
1535  */
1536 static void
1537 taskq_thread(void *arg)
1538 {
1539 	int thread_id;
1540 
1541 	taskq_t *tq = arg;
1542 	taskq_ent_t *tqe;
1543 	callb_cpr_t cprinfo;
1544 	hrtime_t start, end;
1545 	boolean_t freeit;
1546 
1547 	curthread->t_taskq = tq;	/* mark ourselves for taskq_member() */
1548 
1549 	if (curproc != &p0 && (tq->tq_flags & TASKQ_DUTY_CYCLE)) {
1550 		sysdc_thread_enter(curthread, tq->tq_DC,
1551 		    (tq->tq_flags & TASKQ_DC_BATCH) ? SYSDC_THREAD_BATCH : 0);
1552 	}
1553 
1554 	if (tq->tq_flags & TASKQ_CPR_SAFE) {
1555 		CALLB_CPR_INIT_SAFE(curthread, tq->tq_name);
1556 	} else {
1557 		CALLB_CPR_INIT(&cprinfo, &tq->tq_lock, callb_generic_cpr,
1558 		    tq->tq_name);
1559 	}
1560 	mutex_enter(&tq->tq_lock);
1561 	thread_id = ++tq->tq_nthreads;
1562 	ASSERT(tq->tq_flags & TASKQ_THREAD_CREATED);
1563 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1564 	tq->tq_flags &= ~TASKQ_THREAD_CREATED;
1565 
1566 	VERIFY3S(thread_id, <=, tq->tq_nthreads_max);
1567 
1568 	if (tq->tq_nthreads_max == 1)
1569 		tq->tq_thread = curthread;
1570 	else
1571 		tq->tq_threadlist[thread_id - 1] = curthread;
1572 
1573 	/* Allow taskq_create_common()'s taskq_thread_create() to return. */
1574 	if (tq->tq_nthreads == TASKQ_CREATE_ACTIVE_THREADS)
1575 		cv_broadcast(&tq->tq_wait_cv);
1576 
1577 	for (;;) {
1578 		if (tq->tq_flags & TASKQ_CHANGING) {
1579 			/* See if we're no longer needed */
1580 			if (thread_id > tq->tq_nthreads_target) {
1581 				/*
1582 				 * To preserve the one-to-one mapping between
1583 				 * thread_id and thread, we must exit from
1584 				 * highest thread ID to least.
1585 				 *
1586 				 * However, if everyone is exiting, the order
1587 				 * doesn't matter, so just exit immediately.
1588 				 * (this is safe, since you must wait for
1589 				 * nthreads to reach 0 after setting
1590 				 * tq_nthreads_target to 0)
1591 				 */
1592 				if (thread_id == tq->tq_nthreads ||
1593 				    tq->tq_nthreads_target == 0)
1594 					break;
1595 
1596 				/* Wait for higher thread_ids to exit */
1597 				(void) taskq_thread_wait(tq, &tq->tq_lock,
1598 				    &tq->tq_exit_cv, &cprinfo, -1);
1599 				continue;
1600 			}
1601 
1602 			/*
1603 			 * If no thread is starting taskq_thread(), we can
1604 			 * do some bookkeeping.
1605 			 */
1606 			if (!(tq->tq_flags & TASKQ_THREAD_CREATED)) {
1607 				/* Check if we've reached our target */
1608 				if (tq->tq_nthreads == tq->tq_nthreads_target) {
1609 					tq->tq_flags &= ~TASKQ_CHANGING;
1610 					cv_broadcast(&tq->tq_wait_cv);
1611 				}
1612 				/* Check if we need to create a thread */
1613 				if (tq->tq_nthreads < tq->tq_nthreads_target) {
1614 					taskq_thread_create(tq);
1615 					continue; /* tq_lock was dropped */
1616 				}
1617 			}
1618 		}
1619 		if ((tqe = tq->tq_task.tqent_next) == &tq->tq_task) {
1620 			if (--tq->tq_active == 0)
1621 				cv_broadcast(&tq->tq_wait_cv);
1622 			(void) taskq_thread_wait(tq, &tq->tq_lock,
1623 			    &tq->tq_dispatch_cv, &cprinfo, -1);
1624 			tq->tq_active++;
1625 			continue;
1626 		}
1627 
1628 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1629 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1630 		mutex_exit(&tq->tq_lock);
1631 
1632 		/*
1633 		 * For prealloc'd tasks, we don't free anything.  We
1634 		 * have to check this now, because once we call the
1635 		 * function for a prealloc'd taskq, we can't touch the
1636 		 * tqent any longer (calling the function returns the
1637 		 * ownershp of the tqent back to caller of
1638 		 * taskq_dispatch.)
1639 		 */
1640 		if ((!(tq->tq_flags & TASKQ_DYNAMIC)) &&
1641 		    (tqe->tqent_un.tqent_flags & TQENT_FLAG_PREALLOC)) {
1642 			/* clear pointers to assist assertion checks */
1643 			tqe->tqent_next = tqe->tqent_prev = NULL;
1644 			freeit = B_FALSE;
1645 		} else {
1646 			freeit = B_TRUE;
1647 		}
1648 
1649 		rw_enter(&tq->tq_threadlock, RW_READER);
1650 		start = gethrtime();
1651 		DTRACE_PROBE2(taskq__exec__start, taskq_t *, tq,
1652 		    taskq_ent_t *, tqe);
1653 		tqe->tqent_func(tqe->tqent_arg);
1654 		DTRACE_PROBE2(taskq__exec__end, taskq_t *, tq,
1655 		    taskq_ent_t *, tqe);
1656 		end = gethrtime();
1657 		rw_exit(&tq->tq_threadlock);
1658 
1659 		mutex_enter(&tq->tq_lock);
1660 		tq->tq_totaltime += end - start;
1661 		tq->tq_executed++;
1662 
1663 		if (freeit)
1664 			taskq_ent_free(tq, tqe);
1665 	}
1666 
1667 	if (tq->tq_nthreads_max == 1)
1668 		tq->tq_thread = NULL;
1669 	else
1670 		tq->tq_threadlist[thread_id - 1] = NULL;
1671 
1672 	/* We're exiting, and therefore no longer active */
1673 	ASSERT(tq->tq_active > 0);
1674 	tq->tq_active--;
1675 
1676 	ASSERT(tq->tq_nthreads > 0);
1677 	tq->tq_nthreads--;
1678 
1679 	/* Wake up anyone waiting for us to exit */
1680 	cv_broadcast(&tq->tq_exit_cv);
1681 	if (tq->tq_nthreads == tq->tq_nthreads_target) {
1682 		if (!(tq->tq_flags & TASKQ_THREAD_CREATED))
1683 			tq->tq_flags &= ~TASKQ_CHANGING;
1684 
1685 		cv_broadcast(&tq->tq_wait_cv);
1686 	}
1687 
1688 	ASSERT(!(tq->tq_flags & TASKQ_CPR_SAFE));
1689 	CALLB_CPR_EXIT(&cprinfo);		/* drops tq->tq_lock */
1690 	if (curthread->t_lwp != NULL) {
1691 		mutex_enter(&curproc->p_lock);
1692 		lwp_exit();
1693 	} else {
1694 		thread_exit();
1695 	}
1696 }
1697 
1698 /*
1699  * Worker per-entry thread for dynamic dispatches.
1700  */
1701 static void
1702 taskq_d_thread(taskq_ent_t *tqe)
1703 {
1704 	taskq_bucket_t	*bucket = tqe->tqent_un.tqent_bucket;
1705 	taskq_t		*tq = bucket->tqbucket_taskq;
1706 	kmutex_t	*lock = &bucket->tqbucket_lock;
1707 	kcondvar_t	*cv = &tqe->tqent_cv;
1708 	callb_cpr_t	cprinfo;
1709 	clock_t		w = 0;
1710 
1711 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, tq->tq_name);
1712 
1713 	mutex_enter(lock);
1714 
1715 	for (;;) {
1716 		/*
1717 		 * If a task is scheduled (func != NULL), execute it, otherwise
1718 		 * sleep, waiting for a job.
1719 		 */
1720 		if (tqe->tqent_func != NULL) {
1721 			hrtime_t	start;
1722 			hrtime_t	end;
1723 
1724 			ASSERT(bucket->tqbucket_nalloc > 0);
1725 
1726 			/*
1727 			 * It is possible to free the entry right away before
1728 			 * actually executing the task so that subsequent
1729 			 * dispatches may immediately reuse it. But this,
1730 			 * effectively, creates a two-length queue in the entry
1731 			 * and may lead to a deadlock if the execution of the
1732 			 * current task depends on the execution of the next
1733 			 * scheduled task. So, we keep the entry busy until the
1734 			 * task is processed.
1735 			 */
1736 
1737 			mutex_exit(lock);
1738 			start = gethrtime();
1739 			DTRACE_PROBE3(taskq__d__exec__start, taskq_t *, tq,
1740 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1741 			tqe->tqent_func(tqe->tqent_arg);
1742 			DTRACE_PROBE3(taskq__d__exec__end, taskq_t *, tq,
1743 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1744 			end = gethrtime();
1745 			mutex_enter(lock);
1746 			bucket->tqbucket_totaltime += end - start;
1747 
1748 			/*
1749 			 * Return the entry to the bucket free list.
1750 			 */
1751 			tqe->tqent_func = NULL;
1752 			TQ_APPEND(bucket->tqbucket_freelist, tqe);
1753 			bucket->tqbucket_nalloc--;
1754 			bucket->tqbucket_nfree++;
1755 			ASSERT(!IS_EMPTY(bucket->tqbucket_freelist));
1756 			/*
1757 			 * taskq_wait() waits for nalloc to drop to zero on
1758 			 * tqbucket_cv.
1759 			 */
1760 			cv_signal(&bucket->tqbucket_cv);
1761 		}
1762 
1763 		/*
1764 		 * At this point the entry must be in the bucket free list -
1765 		 * either because it was there initially or because it just
1766 		 * finished executing a task and put itself on the free list.
1767 		 */
1768 		ASSERT(bucket->tqbucket_nfree > 0);
1769 		/*
1770 		 * Go to sleep unless we are closing.
1771 		 * If a thread is sleeping too long, it dies.
1772 		 */
1773 		if (! (bucket->tqbucket_flags & TQBUCKET_CLOSE)) {
1774 			w = taskq_thread_wait(tq, lock, cv,
1775 			    &cprinfo, taskq_thread_timeout * hz);
1776 		}
1777 
1778 		/*
1779 		 * At this point we may be in two different states:
1780 		 *
1781 		 * (1) tqent_func is set which means that a new task is
1782 		 *	dispatched and we need to execute it.
1783 		 *
1784 		 * (2) Thread is sleeping for too long or we are closing. In
1785 		 *	both cases destroy the thread and the entry.
1786 		 */
1787 
1788 		/* If func is NULL we should be on the freelist. */
1789 		ASSERT((tqe->tqent_func != NULL) ||
1790 		    (bucket->tqbucket_nfree > 0));
1791 		/* If func is non-NULL we should be allocated */
1792 		ASSERT((tqe->tqent_func == NULL) ||
1793 		    (bucket->tqbucket_nalloc > 0));
1794 
1795 		/* Check freelist consistency */
1796 		ASSERT((bucket->tqbucket_nfree > 0) ||
1797 		    IS_EMPTY(bucket->tqbucket_freelist));
1798 		ASSERT((bucket->tqbucket_nfree == 0) ||
1799 		    !IS_EMPTY(bucket->tqbucket_freelist));
1800 
1801 		if ((tqe->tqent_func == NULL) &&
1802 		    ((w == -1) || (bucket->tqbucket_flags & TQBUCKET_CLOSE))) {
1803 			/*
1804 			 * This thread is sleeping for too long or we are
1805 			 * closing - time to die.
1806 			 */
1807 			break;
1808 		}
1809 	}
1810 
1811 	/*
1812 	 * Thread creation/destruction happens rarely,
1813 	 * so grabbing the lock is not a big performance issue.
1814 	 * The bucket lock is dropped by CALLB_CPR_EXIT().
1815 	 */
1816 
1817 	/* Remove the entry from the free list. */
1818 	tqe->tqent_prev->tqent_next = tqe->tqent_next;
1819 	tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1820 	ASSERT(bucket->tqbucket_nfree > 0);
1821 	bucket->tqbucket_nfree--;
1822 
1823 	TQ_STAT(bucket, tqs_tdeaths);
1824 	cv_signal(&bucket->tqbucket_cv);
1825 	tqe->tqent_thread = NULL;
1826 	mutex_enter(&tq->tq_lock);
1827 	tq->tq_tdeaths++;
1828 	mutex_exit(&tq->tq_lock);
1829 	CALLB_CPR_EXIT(&cprinfo);	/* mutex_exit(lock) */
1830 
1831 	kmem_cache_free(taskq_ent_cache, tqe);
1832 
1833 	if (curthread->t_lwp != NULL) {
1834 		mutex_enter(&curproc->p_lock);
1835 		lwp_exit(); /* noreturn. drops p_lock */
1836 	} else {
1837 		thread_exit();
1838 	}
1839 }
1840 
1841 
1842 /*
1843  * Taskq creation. May sleep for memory.
1844  * Always use automatically generated instances to avoid kstat name space
1845  * collisions.
1846  */
1847 
1848 taskq_t *
1849 taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
1850     int maxalloc, uint_t flags)
1851 {
1852 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1853 
1854 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1855 	    maxalloc, &p0, 0, flags | TASKQ_NOINSTANCE));
1856 }
1857 
1858 /*
1859  * Create an instance of task queue. It is legal to create task queues with the
1860  * same name and different instances.
1861  *
1862  * taskq_create_instance is used by ddi_taskq_create() where it gets the
1863  * instance from ddi_get_instance(). In some cases the instance is not
1864  * initialized and is set to -1. This case is handled as if no instance was
1865  * passed at all.
1866  */
1867 taskq_t *
1868 taskq_create_instance(const char *name, int instance, int nthreads, pri_t pri,
1869     int minalloc, int maxalloc, uint_t flags)
1870 {
1871 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1872 	ASSERT((instance >= 0) || (instance == -1));
1873 
1874 	if (instance < 0) {
1875 		flags |= TASKQ_NOINSTANCE;
1876 	}
1877 
1878 	return (taskq_create_common(name, instance, nthreads,
1879 	    pri, minalloc, maxalloc, &p0, 0, flags));
1880 }
1881 
1882 taskq_t *
1883 taskq_create_proc(const char *name, int nthreads, pri_t pri, int minalloc,
1884     int maxalloc, proc_t *proc, uint_t flags)
1885 {
1886 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1887 	ASSERT(proc->p_flag & SSYS);
1888 
1889 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1890 	    maxalloc, proc, 0, flags | TASKQ_NOINSTANCE));
1891 }
1892 
1893 taskq_t *
1894 taskq_create_sysdc(const char *name, int nthreads, int minalloc,
1895     int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1896 {
1897 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1898 	ASSERT(proc->p_flag & SSYS);
1899 
1900 	return (taskq_create_common(name, 0, nthreads, minclsyspri, minalloc,
1901 	    maxalloc, proc, dc, flags | TASKQ_NOINSTANCE | TASKQ_DUTY_CYCLE));
1902 }
1903 
1904 static taskq_t *
1905 taskq_create_common(const char *name, int instance, int nthreads, pri_t pri,
1906     int minalloc, int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1907 {
1908 	taskq_t *tq = kmem_cache_alloc(taskq_cache, KM_SLEEP);
1909 	uint_t ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
1910 	uint_t bsize;	/* # of buckets - always power of 2 */
1911 	int max_nthreads;
1912 
1913 	/*
1914 	 * TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
1915 	 * mutually incompatible.
1916 	 */
1917 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_CPR_SAFE));
1918 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_THREADS_CPU_PCT));
1919 	IMPLY((flags & TASKQ_CPR_SAFE), !(flags & TASKQ_THREADS_CPU_PCT));
1920 
1921 	/* Cannot have DYNAMIC with DUTY_CYCLE */
1922 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_DUTY_CYCLE));
1923 
1924 	/* Cannot have DUTY_CYCLE with a p0 kernel process */
1925 	IMPLY((flags & TASKQ_DUTY_CYCLE), proc != &p0);
1926 
1927 	/* Cannot have THREADS_LWP with a p0 kernel process */
1928 	IMPLY((flags & TASKQ_THREADS_LWP), proc != &p0);
1929 
1930 	/* Cannot have DC_BATCH without DUTY_CYCLE */
1931 	ASSERT((flags & (TASKQ_DUTY_CYCLE|TASKQ_DC_BATCH)) != TASKQ_DC_BATCH);
1932 
1933 	ASSERT(proc != NULL);
1934 
1935 	bsize = 1 << (highbit(ncpus) - 1);
1936 	ASSERT(bsize >= 1);
1937 	bsize = MIN(bsize, taskq_maxbuckets);
1938 
1939 	if (flags & TASKQ_DYNAMIC) {
1940 		ASSERT3S(nthreads, >=, 1);
1941 		tq->tq_maxsize = nthreads;
1942 
1943 		/* For dynamic task queues use just one backup thread */
1944 		nthreads = max_nthreads = 1;
1945 
1946 	} else if (flags & TASKQ_THREADS_CPU_PCT) {
1947 		uint_t pct;
1948 		ASSERT3S(nthreads, >=, 0);
1949 		pct = nthreads;
1950 
1951 		if (pct > taskq_cpupct_max_percent)
1952 			pct = taskq_cpupct_max_percent;
1953 
1954 		/*
1955 		 * If you're using THREADS_CPU_PCT, the process for the
1956 		 * taskq threads must be curproc.  This allows any pset
1957 		 * binding to be inherited correctly.  If proc is &p0,
1958 		 * we won't be creating LWPs, so new threads will be assigned
1959 		 * to the default processor set.
1960 		 */
1961 		ASSERT(curproc == proc || proc == &p0);
1962 		tq->tq_threads_ncpus_pct = pct;
1963 		nthreads = 1;		/* corrected in taskq_thread_create() */
1964 		max_nthreads = TASKQ_THREADS_PCT(max_ncpus, pct);
1965 
1966 	} else {
1967 		ASSERT3S(nthreads, >=, 1);
1968 		max_nthreads = nthreads;
1969 	}
1970 
1971 	if (max_nthreads < taskq_minimum_nthreads_max)
1972 		max_nthreads = taskq_minimum_nthreads_max;
1973 
1974 	/*
1975 	 * Make sure the name is 0-terminated, and conforms to the rules for
1976 	 * C indentifiers
1977 	 */
1978 	(void) strncpy(tq->tq_name, name, TASKQ_NAMELEN + 1);
1979 	strident_canon(tq->tq_name, TASKQ_NAMELEN + 1);
1980 
1981 	tq->tq_flags = flags | TASKQ_CHANGING;
1982 	tq->tq_active = 0;
1983 	tq->tq_instance = instance;
1984 	tq->tq_nthreads_target = nthreads;
1985 	tq->tq_nthreads_max = max_nthreads;
1986 	tq->tq_minalloc = minalloc;
1987 	tq->tq_maxalloc = maxalloc;
1988 	tq->tq_nbuckets = bsize;
1989 	tq->tq_proc = proc;
1990 	tq->tq_pri = pri;
1991 	tq->tq_DC = dc;
1992 	list_link_init(&tq->tq_cpupct_link);
1993 
1994 	if (max_nthreads > 1)
1995 		tq->tq_threadlist = kmem_alloc(
1996 		    sizeof (kthread_t *) * max_nthreads, KM_SLEEP);
1997 
1998 	mutex_enter(&tq->tq_lock);
1999 	if (flags & TASKQ_PREPOPULATE) {
2000 		while (minalloc-- > 0)
2001 			taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
2002 	}
2003 
2004 	/*
2005 	 * Before we start creating threads for this taskq, take a
2006 	 * zone hold so the zone can't go away before taskq_destroy
2007 	 * makes sure all the taskq threads are gone.  This hold is
2008 	 * similar in purpose to those taken by zthread_create().
2009 	 */
2010 	zone_hold(tq->tq_proc->p_zone);
2011 
2012 	/*
2013 	 * Create the first thread, which will create any other threads
2014 	 * necessary.  taskq_thread_create will not return until we have
2015 	 * enough threads to be able to process requests.
2016 	 */
2017 	taskq_thread_create(tq);
2018 	mutex_exit(&tq->tq_lock);
2019 
2020 	if (flags & TASKQ_DYNAMIC) {
2021 		taskq_bucket_t *bucket = kmem_zalloc(sizeof (taskq_bucket_t) *
2022 		    bsize, KM_SLEEP);
2023 		int b_id;
2024 
2025 		tq->tq_buckets = bucket;
2026 
2027 		/* Initialize each bucket */
2028 		for (b_id = 0; b_id < bsize; b_id++, bucket++) {
2029 			mutex_init(&bucket->tqbucket_lock, NULL, MUTEX_DEFAULT,
2030 			    NULL);
2031 			cv_init(&bucket->tqbucket_cv, NULL, CV_DEFAULT, NULL);
2032 			bucket->tqbucket_taskq = tq;
2033 			bucket->tqbucket_freelist.tqent_next =
2034 			    bucket->tqbucket_freelist.tqent_prev =
2035 			    &bucket->tqbucket_freelist;
2036 			if (flags & TASKQ_PREPOPULATE)
2037 				taskq_bucket_extend(bucket);
2038 		}
2039 	}
2040 
2041 	/*
2042 	 * Install kstats.
2043 	 * We have two cases:
2044 	 *   1) Instance is provided to taskq_create_instance(). In this case it
2045 	 *	should be >= 0 and we use it.
2046 	 *
2047 	 *   2) Instance is not provided and is automatically generated
2048 	 */
2049 	if (flags & TASKQ_NOINSTANCE) {
2050 		instance = tq->tq_instance =
2051 		    (int)(uintptr_t)vmem_alloc(taskq_id_arena, 1, VM_SLEEP);
2052 	}
2053 
2054 	if (flags & TASKQ_DYNAMIC) {
2055 		if ((tq->tq_kstat = kstat_create("unix", instance,
2056 		    tq->tq_name, "taskq_d", KSTAT_TYPE_NAMED,
2057 		    sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
2058 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
2059 			tq->tq_kstat->ks_lock = &taskq_d_kstat_lock;
2060 			tq->tq_kstat->ks_data = &taskq_d_kstat;
2061 			tq->tq_kstat->ks_update = taskq_d_kstat_update;
2062 			tq->tq_kstat->ks_private = tq;
2063 			kstat_install(tq->tq_kstat);
2064 		}
2065 	} else {
2066 		if ((tq->tq_kstat = kstat_create("unix", instance, tq->tq_name,
2067 		    "taskq", KSTAT_TYPE_NAMED,
2068 		    sizeof (taskq_kstat) / sizeof (kstat_named_t),
2069 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
2070 			tq->tq_kstat->ks_lock = &taskq_kstat_lock;
2071 			tq->tq_kstat->ks_data = &taskq_kstat;
2072 			tq->tq_kstat->ks_update = taskq_kstat_update;
2073 			tq->tq_kstat->ks_private = tq;
2074 			kstat_install(tq->tq_kstat);
2075 		}
2076 	}
2077 
2078 	return (tq);
2079 }
2080 
2081 /*
2082  * taskq_destroy().
2083  *
2084  * Assumes: by the time taskq_destroy is called no one will use this task queue
2085  * in any way and no one will try to dispatch entries in it.
2086  */
2087 void
2088 taskq_destroy(taskq_t *tq)
2089 {
2090 	taskq_bucket_t *b = tq->tq_buckets;
2091 	int bid = 0;
2092 
2093 	ASSERT(! (tq->tq_flags & TASKQ_CPR_SAFE));
2094 
2095 	/*
2096 	 * Destroy kstats.
2097 	 */
2098 	if (tq->tq_kstat != NULL) {
2099 		kstat_delete(tq->tq_kstat);
2100 		tq->tq_kstat = NULL;
2101 	}
2102 
2103 	/*
2104 	 * Destroy instance if needed.
2105 	 */
2106 	if (tq->tq_flags & TASKQ_NOINSTANCE) {
2107 		vmem_free(taskq_id_arena, (void *)(uintptr_t)(tq->tq_instance),
2108 		    1);
2109 		tq->tq_instance = 0;
2110 	}
2111 
2112 	/*
2113 	 * Unregister from the cpupct list.
2114 	 */
2115 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
2116 		taskq_cpupct_remove(tq);
2117 	}
2118 
2119 	/*
2120 	 * Wait for any pending entries to complete.
2121 	 */
2122 	taskq_wait(tq);
2123 
2124 	mutex_enter(&tq->tq_lock);
2125 	ASSERT((tq->tq_task.tqent_next == &tq->tq_task) &&
2126 	    (tq->tq_active == 0));
2127 
2128 	/* notify all the threads that they need to exit */
2129 	tq->tq_nthreads_target = 0;
2130 
2131 	tq->tq_flags |= TASKQ_CHANGING;
2132 	cv_broadcast(&tq->tq_dispatch_cv);
2133 	cv_broadcast(&tq->tq_exit_cv);
2134 
2135 	while (tq->tq_nthreads != 0)
2136 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
2137 
2138 	if (tq->tq_nthreads_max != 1)
2139 		kmem_free(tq->tq_threadlist, sizeof (kthread_t *) *
2140 		    tq->tq_nthreads_max);
2141 
2142 	tq->tq_minalloc = 0;
2143 	while (tq->tq_nalloc != 0)
2144 		taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
2145 
2146 	mutex_exit(&tq->tq_lock);
2147 
2148 	/*
2149 	 * Mark each bucket as closing and wakeup all sleeping threads.
2150 	 */
2151 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2152 		taskq_ent_t *tqe;
2153 
2154 		mutex_enter(&b->tqbucket_lock);
2155 
2156 		b->tqbucket_flags |= TQBUCKET_CLOSE;
2157 		/* Wakeup all sleeping threads */
2158 
2159 		for (tqe = b->tqbucket_freelist.tqent_next;
2160 		    tqe != &b->tqbucket_freelist; tqe = tqe->tqent_next)
2161 			cv_signal(&tqe->tqent_cv);
2162 
2163 		ASSERT(b->tqbucket_nalloc == 0);
2164 
2165 		/*
2166 		 * At this point we waited for all pending jobs to complete (in
2167 		 * both the task queue and the bucket and no new jobs should
2168 		 * arrive. Wait for all threads to die.
2169 		 */
2170 		while (b->tqbucket_nfree > 0)
2171 			cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
2172 		mutex_exit(&b->tqbucket_lock);
2173 		mutex_destroy(&b->tqbucket_lock);
2174 		cv_destroy(&b->tqbucket_cv);
2175 	}
2176 
2177 	if (tq->tq_buckets != NULL) {
2178 		ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2179 		kmem_free(tq->tq_buckets,
2180 		    sizeof (taskq_bucket_t) * tq->tq_nbuckets);
2181 
2182 		/* Cleanup fields before returning tq to the cache */
2183 		tq->tq_buckets = NULL;
2184 		tq->tq_tcreates = 0;
2185 		tq->tq_tdeaths = 0;
2186 	} else {
2187 		ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
2188 	}
2189 
2190 	/*
2191 	 * Now that all the taskq threads are gone, we can
2192 	 * drop the zone hold taken in taskq_create_common
2193 	 */
2194 	zone_rele(tq->tq_proc->p_zone);
2195 
2196 	tq->tq_threads_ncpus_pct = 0;
2197 	tq->tq_totaltime = 0;
2198 	tq->tq_tasks = 0;
2199 	tq->tq_maxtasks = 0;
2200 	tq->tq_executed = 0;
2201 	kmem_cache_free(taskq_cache, tq);
2202 }
2203 
2204 /*
2205  * Extend a bucket with a new entry on the free list and attach a worker thread
2206  * to it.
2207  *
2208  * Argument: pointer to the bucket.
2209  *
2210  * This function may quietly fail. It is only used by taskq_dispatch() which
2211  * handles such failures properly.
2212  */
2213 static void
2214 taskq_bucket_extend(void *arg)
2215 {
2216 	taskq_ent_t *tqe;
2217 	taskq_bucket_t *b = (taskq_bucket_t *)arg;
2218 	taskq_t *tq = b->tqbucket_taskq;
2219 	kthread_t *t;
2220 	int nthreads;
2221 
2222 	mutex_enter(&tq->tq_lock);
2223 
2224 	if (! ENOUGH_MEMORY()) {
2225 		tq->tq_nomem++;
2226 		mutex_exit(&tq->tq_lock);
2227 		return;
2228 	}
2229 
2230 	/*
2231 	 * Observe global taskq limits on the number of threads.
2232 	 */
2233 	if (tq->tq_tcreates++ - tq->tq_tdeaths > tq->tq_maxsize) {
2234 		tq->tq_tcreates--;
2235 		mutex_exit(&tq->tq_lock);
2236 		return;
2237 	}
2238 	mutex_exit(&tq->tq_lock);
2239 
2240 	tqe = kmem_cache_alloc(taskq_ent_cache, KM_NOSLEEP);
2241 
2242 	if (tqe == NULL) {
2243 		mutex_enter(&tq->tq_lock);
2244 		tq->tq_nomem++;
2245 		tq->tq_tcreates--;
2246 		mutex_exit(&tq->tq_lock);
2247 		return;
2248 	}
2249 
2250 	ASSERT(tqe->tqent_thread == NULL);
2251 
2252 	tqe->tqent_un.tqent_bucket = b;
2253 
2254 	/*
2255 	 * Create a thread in a TS_STOPPED state first. If it is successfully
2256 	 * created, place the entry on the free list and start the thread.
2257 	 */
2258 	if ((tq->tq_flags & TASKQ_THREADS_LWP) != 0) {
2259 		/* Enforced in taskq_create_common */
2260 		ASSERT3P(tq->tq_proc, !=, &p0);
2261 		t = lwp_kernel_create(tq->tq_proc, taskq_d_thread,
2262 		    tqe, TS_STOPPED, tq->tq_pri);
2263 	} else {
2264 		t = thread_create(NULL, 0, taskq_d_thread, tqe,
2265 		    0, tq->tq_proc, TS_STOPPED, tq->tq_pri);
2266 	}
2267 	tqe->tqent_thread = t;
2268 	t->t_taskq = tq;	/* mark thread as a taskq_member() */
2269 
2270 	/*
2271 	 * Once the entry is ready, link it to the the bucket free list.
2272 	 */
2273 	mutex_enter(&b->tqbucket_lock);
2274 	tqe->tqent_func = NULL;
2275 	TQ_APPEND(b->tqbucket_freelist, tqe);
2276 	b->tqbucket_nfree++;
2277 	TQ_STAT(b, tqs_tcreates);
2278 
2279 #if TASKQ_STATISTIC
2280 	nthreads = b->tqbucket_stat.tqs_tcreates -
2281 	    b->tqbucket_stat.tqs_tdeaths;
2282 	b->tqbucket_stat.tqs_maxthreads = MAX(nthreads,
2283 	    b->tqbucket_stat.tqs_maxthreads);
2284 #endif
2285 
2286 	mutex_exit(&b->tqbucket_lock);
2287 
2288 	/*
2289 	 * Start the stopped thread.
2290 	 */
2291 	if (t->t_lwp != NULL) {
2292 		proc_t *p = tq->tq_proc;
2293 		mutex_enter(&p->p_lock);
2294 		t->t_proc_flag &= ~TP_HOLDLWP;
2295 		lwp_create_done(t);	/* Sets TS_ALLSTART etc. */
2296 		mutex_exit(&p->p_lock);
2297 	} else {
2298 		thread_lock(t);
2299 		t->t_schedflag |= TS_ALLSTART;
2300 		setrun_locked(t);
2301 		thread_unlock(t);
2302 	}
2303 }
2304 
2305 static int
2306 taskq_kstat_update(kstat_t *ksp, int rw)
2307 {
2308 	struct taskq_kstat *tqsp = &taskq_kstat;
2309 	taskq_t *tq = ksp->ks_private;
2310 
2311 	if (rw == KSTAT_WRITE)
2312 		return (EACCES);
2313 
2314 	tqsp->tq_pid.value.ui64 = tq->tq_proc->p_pid;
2315 	tqsp->tq_tasks.value.ui64 = tq->tq_tasks;
2316 	tqsp->tq_executed.value.ui64 = tq->tq_executed;
2317 	tqsp->tq_maxtasks.value.ui64 = tq->tq_maxtasks;
2318 	tqsp->tq_totaltime.value.ui64 = tq->tq_totaltime;
2319 	tqsp->tq_nactive.value.ui64 = tq->tq_active;
2320 	tqsp->tq_nalloc.value.ui64 = tq->tq_nalloc;
2321 	tqsp->tq_pri.value.ui64 = tq->tq_pri;
2322 	tqsp->tq_nthreads.value.ui64 = tq->tq_nthreads;
2323 	tqsp->tq_nomem.value.ui64 = tq->tq_nomem;
2324 	return (0);
2325 }
2326 
2327 static int
2328 taskq_d_kstat_update(kstat_t *ksp, int rw)
2329 {
2330 	struct taskq_d_kstat *tqsp = &taskq_d_kstat;
2331 	taskq_t *tq = ksp->ks_private;
2332 	taskq_bucket_t *b = tq->tq_buckets;
2333 	int bid = 0;
2334 
2335 	if (rw == KSTAT_WRITE)
2336 		return (EACCES);
2337 
2338 	ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2339 
2340 	tqsp->tqd_btasks.value.ui64 = tq->tq_tasks;
2341 	tqsp->tqd_bexecuted.value.ui64 = tq->tq_executed;
2342 	tqsp->tqd_bmaxtasks.value.ui64 = tq->tq_maxtasks;
2343 	tqsp->tqd_bnalloc.value.ui64 = tq->tq_nalloc;
2344 	tqsp->tqd_bnactive.value.ui64 = tq->tq_active;
2345 	tqsp->tqd_btotaltime.value.ui64 = tq->tq_totaltime;
2346 	tqsp->tqd_pri.value.ui64 = tq->tq_pri;
2347 	tqsp->tqd_nomem.value.ui64 = tq->tq_nomem;
2348 
2349 	tqsp->tqd_hits.value.ui64 = 0;
2350 	tqsp->tqd_misses.value.ui64 = 0;
2351 	tqsp->tqd_overflows.value.ui64 = 0;
2352 	tqsp->tqd_tcreates.value.ui64 = 0;
2353 	tqsp->tqd_tdeaths.value.ui64 = 0;
2354 	tqsp->tqd_maxthreads.value.ui64 = 0;
2355 	tqsp->tqd_nomem.value.ui64 = 0;
2356 	tqsp->tqd_disptcreates.value.ui64 = 0;
2357 	tqsp->tqd_totaltime.value.ui64 = 0;
2358 	tqsp->tqd_nalloc.value.ui64 = 0;
2359 	tqsp->tqd_nfree.value.ui64 = 0;
2360 
2361 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2362 		tqsp->tqd_hits.value.ui64 += b->tqbucket_stat.tqs_hits;
2363 		tqsp->tqd_misses.value.ui64 += b->tqbucket_stat.tqs_misses;
2364 		tqsp->tqd_overflows.value.ui64 += b->tqbucket_stat.tqs_overflow;
2365 		tqsp->tqd_tcreates.value.ui64 += b->tqbucket_stat.tqs_tcreates;
2366 		tqsp->tqd_tdeaths.value.ui64 += b->tqbucket_stat.tqs_tdeaths;
2367 		tqsp->tqd_maxthreads.value.ui64 +=
2368 		    b->tqbucket_stat.tqs_maxthreads;
2369 		tqsp->tqd_disptcreates.value.ui64 +=
2370 		    b->tqbucket_stat.tqs_disptcreates;
2371 		tqsp->tqd_totaltime.value.ui64 += b->tqbucket_totaltime;
2372 		tqsp->tqd_nalloc.value.ui64 += b->tqbucket_nalloc;
2373 		tqsp->tqd_nfree.value.ui64 += b->tqbucket_nfree;
2374 	}
2375 	return (0);
2376 }
2377