xref: /illumos-gate/usr/src/uts/common/os/taskq.c (revision c883468ae561a0b211f3633eb2a24aa4306661f2)
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 2011 Nexenta Systems, Inc.  All rights reserved.
28  */
29 
30 /*
31  * Kernel task queues: general-purpose asynchronous task scheduling.
32  *
33  * A common problem in kernel programming is the need to schedule tasks
34  * to be performed later, by another thread. There are several reasons
35  * you may want or need to do this:
36  *
37  * (1) The task isn't time-critical, but your current code path is.
38  *
39  * (2) The task may require grabbing locks that you already hold.
40  *
41  * (3) The task may need to block (e.g. to wait for memory), but you
42  *     cannot block in your current context.
43  *
44  * (4) Your code path can't complete because of some condition, but you can't
45  *     sleep or fail, so you queue the task for later execution when condition
46  *     disappears.
47  *
48  * (5) You just want a simple way to launch multiple tasks in parallel.
49  *
50  * Task queues provide such a facility. In its simplest form (used when
51  * performance is not a critical consideration) a task queue consists of a
52  * single list of tasks, together with one or more threads to service the
53  * list. There are some cases when this simple queue is not sufficient:
54  *
55  * (1) The task queues are very hot and there is a need to avoid data and lock
56  *	contention over global resources.
57  *
58  * (2) Some tasks may depend on other tasks to complete, so they can't be put in
59  *	the same list managed by the same thread.
60  *
61  * (3) Some tasks may block for a long time, and this should not block other
62  *	tasks in the queue.
63  *
64  * To provide useful service in such cases we define a "dynamic task queue"
65  * which has an individual thread for each of the tasks. These threads are
66  * dynamically created as they are needed and destroyed when they are not in
67  * use. The API for managing task pools is the same as for managing task queues
68  * with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
69  * dynamic task pool behavior is desired.
70  *
71  * Dynamic task queues may also place tasks in the normal queue (called "backing
72  * queue") when task pool runs out of resources. Users of task queues may
73  * disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
74  * flags.
75  *
76  * The backing task queue is also used for scheduling internal tasks needed for
77  * dynamic task queue maintenance.
78  *
79  * INTERFACES ==================================================================
80  *
81  * taskq_t *taskq_create(name, nthreads, pri, minalloc, maxall, flags);
82  *
83  *	Create a taskq with specified properties.
84  *	Possible 'flags':
85  *
86  *	  TASKQ_DYNAMIC: Create task pool for task management. If this flag is
87  *		specified, 'nthreads' specifies the maximum number of threads in
88  *		the task queue. Task execution order for dynamic task queues is
89  *		not predictable.
90  *
91  *		If this flag is not specified (default case) a
92  *		single-list task queue is created with 'nthreads' threads
93  *		servicing it. Entries in this queue are managed by
94  *		taskq_ent_alloc() and taskq_ent_free() which try to keep the
95  *		task population between 'minalloc' and 'maxalloc', but the
96  *		latter limit is only advisory for TQ_SLEEP dispatches and the
97  *		former limit is only advisory for TQ_NOALLOC dispatches. If
98  *		TASKQ_PREPOPULATE is set in 'flags', the taskq will be
99  *		prepopulated with 'minalloc' task structures.
100  *
101  *		Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
102  *		executed in the order they are scheduled if nthreads == 1.
103  *		If nthreads > 1, task execution order is not predictable.
104  *
105  *	  TASKQ_PREPOPULATE: Prepopulate task queue with threads.
106  *		Also prepopulate the task queue with 'minalloc' task structures.
107  *
108  *	  TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
109  *		interpreted as a percentage of the # of online CPUs on the
110  *		system.  The taskq subsystem will automatically adjust the
111  *		number of threads in the taskq in response to CPU online
112  *		and offline events, to keep the ratio.  nthreads must be in
113  *		the range [0,100].
114  *
115  *		The calculation used is:
116  *
117  *			MAX((ncpus_online * percentage)/100, 1)
118  *
119  *		This flag is not supported for DYNAMIC task queues.
120  *		This flag is not compatible with TASKQ_CPR_SAFE.
121  *
122  *	  TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
123  *		use their own protocol for handling CPR issues. This flag is not
124  *		supported for DYNAMIC task queues.  This flag is not compatible
125  *		with TASKQ_THREADS_CPU_PCT.
126  *
127  *	The 'pri' field specifies the default priority for the threads that
128  *	service all scheduled tasks.
129  *
130  * taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
131  *    maxall, flags);
132  *
133  *	Like taskq_create(), but takes an instance number (or -1 to indicate
134  *	no instance).
135  *
136  * taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxall, proc,
137  *    flags);
138  *
139  *	Like taskq_create(), but creates the taskq threads in the specified
140  *	system process.  If proc != &p0, this must be called from a thread
141  *	in that process.
142  *
143  * taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxall, proc,
144  *    dc, flags);
145  *
146  *	Like taskq_create_proc(), but the taskq threads will use the
147  *	System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
148  *
149  * void taskq_destroy(tap):
150  *
151  *	Waits for any scheduled tasks to complete, then destroys the taskq.
152  *	Caller should guarantee that no new tasks are scheduled in the closing
153  *	taskq.
154  *
155  * taskqid_t taskq_dispatch(tq, func, arg, flags):
156  *
157  *	Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
158  *	the caller is willing to block for memory.  The function returns an
159  *	opaque value which is zero iff dispatch fails.  If flags is TQ_NOSLEEP
160  *	or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
161  *	and returns (taskqid_t)0.
162  *
163  *	ASSUMES: func != NULL.
164  *
165  *	Possible flags:
166  *	  TQ_NOSLEEP: Do not wait for resources; may fail.
167  *
168  *	  TQ_NOALLOC: Do not allocate memory; may fail.  May only be used with
169  *		non-dynamic task queues.
170  *
171  *	  TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
172  *		lack of available resources and fail. If this flag is not
173  *		set, and the task pool is exhausted, the task may be scheduled
174  *		in the backing queue. This flag may ONLY be used with dynamic
175  *		task queues.
176  *
177  *		NOTE: This flag should always be used when a task queue is used
178  *		for tasks that may depend on each other for completion.
179  *		Enqueueing dependent tasks may create deadlocks.
180  *
181  *	  TQ_SLEEP:   May block waiting for resources. May still fail for
182  *		dynamic task queues if TQ_NOQUEUE is also specified, otherwise
183  *		always succeed.
184  *
185  *	  TQ_FRONT:   Puts the new task at the front of the queue.  Be careful.
186  *
187  *	NOTE: Dynamic task queues are much more likely to fail in
188  *		taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
189  *		is important to have backup strategies handling such failures.
190  *
191  * void taskq_dispatch_ent(tq, func, arg, flags, tqent)
192  *
193  *	This is a light-weight form of taskq_dispatch(), that uses a
194  *	preallocated taskq_ent_t structure for scheduling.  As a
195  *	result, it does not perform allocations and cannot ever fail.
196  *	Note especially that it cannot be used with TASKQ_DYNAMIC
197  *	taskqs.  The memory for the tqent must not be modified or used
198  *	until the function (func) is called.  (However, func itself
199  *	may safely modify or free this memory, once it is called.)
200  *	Note that the taskq framework will NOT free this memory.
201  *
202  * void taskq_wait(tq):
203  *
204  *	Waits for all previously scheduled tasks to complete.
205  *
206  *	NOTE: It does not stop any new task dispatches.
207  *	      Do NOT call taskq_wait() from a task: it will cause deadlock.
208  *
209  * void taskq_suspend(tq)
210  *
211  *	Suspend all task execution. Tasks already scheduled for a dynamic task
212  *	queue will still be executed, but all new scheduled tasks will be
213  *	suspended until taskq_resume() is called.
214  *
215  * int  taskq_suspended(tq)
216  *
217  *	Returns 1 if taskq is suspended and 0 otherwise. It is intended to
218  *	ASSERT that the task queue is suspended.
219  *
220  * void taskq_resume(tq)
221  *
222  *	Resume task queue execution.
223  *
224  * int  taskq_member(tq, thread)
225  *
226  *	Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
227  *	intended use is to ASSERT that a given function is called in taskq
228  *	context only.
229  *
230  * system_taskq
231  *
232  *	Global system-wide dynamic task queue for common uses. It may be used by
233  *	any subsystem that needs to schedule tasks and does not need to manage
234  *	its own task queues. It is initialized quite early during system boot.
235  *
236  * IMPLEMENTATION ==============================================================
237  *
238  * This is schematic representation of the task queue structures.
239  *
240  *   taskq:
241  *   +-------------+
242  *   | tq_lock     | +---< taskq_ent_free()
243  *   +-------------+ |
244  *   |...          | | tqent:                  tqent:
245  *   +-------------+ | +------------+          +------------+
246  *   | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
247  *   +-------------+   +------------+          +------------+
248  *   |...          |   | ...        |          | ...        |
249  *   +-------------+   +------------+          +------------+
250  *   | tq_task     |    |
251  *   |             |    +-------------->taskq_ent_alloc()
252  * +--------------------------------------------------------------------------+
253  * | |                     |            tqent                   tqent         |
254  * | +---------------------+     +--> +------------+     +--> +------------+  |
255  * | | ...		   |     |    | func, arg  |     |    | func, arg  |  |
256  * +>+---------------------+ <---|-+  +------------+ <---|-+  +------------+  |
257  *   | tq_taskq.tqent_next | ----+ |  | tqent_next | --->+ |  | tqent_next |--+
258  *   +---------------------+	   |  +------------+     ^ |  +------------+
259  * +-| tq_task.tqent_prev  |	   +--| tqent_prev |     | +--| tqent_prev |  ^
260  * | +---------------------+	      +------------+     |    +------------+  |
261  * | |...		   |	      | ...        |     |    | ...        |  |
262  * | +---------------------+	      +------------+     |    +------------+  |
263  * |                                      ^              |                    |
264  * |                                      |              |                    |
265  * +--------------------------------------+--------------+       TQ_APPEND() -+
266  *   |             |                      |
267  *   |...          |   taskq_thread()-----+
268  *   +-------------+
269  *   | tq_buckets  |--+-------> [ NULL ] (for regular task queues)
270  *   +-------------+  |
271  *                    |   DYNAMIC TASK QUEUES:
272  *                    |
273  *                    +-> taskq_bucket[nCPU]		taskq_bucket_dispatch()
274  *                        +-------------------+                    ^
275  *                   +--->| tqbucket_lock     |                    |
276  *                   |    +-------------------+   +--------+      +--------+
277  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  | ^
278  *                   |    +-------------------+<--+--------+<--...+--------+ |
279  *                   |    | ...               |   | thread |      | thread | |
280  *                   |    +-------------------+   +--------+      +--------+ |
281  *                   |    +-------------------+                              |
282  * taskq_dispatch()--+--->| tqbucket_lock     |             TQ_APPEND()------+
283  *      TQ_HASH()    |    +-------------------+   +--------+      +--------+
284  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  |
285  *                   |    +-------------------+<--+--------+<--...+--------+
286  *                   |    | ...               |   | thread |      | thread |
287  *                   |    +-------------------+   +--------+      +--------+
288  *		     +--->	...
289  *
290  *
291  * Task queues use tq_task field to link new entry in the queue. The queue is a
292  * circular doubly-linked list. Entries are put in the end of the list with
293  * TQ_APPEND() and processed from the front of the list by taskq_thread() in
294  * FIFO order. Task queue entries are cached in the free list managed by
295  * taskq_ent_alloc() and taskq_ent_free() functions.
296  *
297  *	All threads used by task queues mark t_taskq field of the thread to
298  *	point to the task queue.
299  *
300  * Taskq Thread Management -----------------------------------------------------
301  *
302  * Taskq's non-dynamic threads are managed with several variables and flags:
303  *
304  *	* tq_nthreads	- The number of threads in taskq_thread() for the
305  *			  taskq.
306  *
307  *	* tq_active	- The number of threads not waiting on a CV in
308  *			  taskq_thread(); includes newly created threads
309  *			  not yet counted in tq_nthreads.
310  *
311  *	* tq_nthreads_target
312  *			- The number of threads desired for the taskq.
313  *
314  *	* tq_flags & TASKQ_CHANGING
315  *			- Indicates that tq_nthreads != tq_nthreads_target.
316  *
317  *	* tq_flags & TASKQ_THREAD_CREATED
318  *			- Indicates that a thread is being created in the taskq.
319  *
320  * During creation, tq_nthreads and tq_active are set to 0, and
321  * tq_nthreads_target is set to the number of threads desired.  The
322  * TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
323  * create the first thread. taskq_thread_create() increments tq_active,
324  * sets TASKQ_THREAD_CREATED, and creates the new thread.
325  *
326  * Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
327  * flag, and increments tq_nthreads.  It stores the new value of
328  * tq_nthreads as its "thread_id", and stores its thread pointer in the
329  * tq_threadlist at the (thread_id - 1).  We keep the thread_id space
330  * densely packed by requiring that only the largest thread_id can exit during
331  * normal adjustment.   The exception is during the destruction of the
332  * taskq; once tq_nthreads_target is set to zero, no new threads will be created
333  * for the taskq queue, so every thread can exit without any ordering being
334  * necessary.
335  *
336  * Threads will only process work if their thread id is <= tq_nthreads_target.
337  *
338  * When TASKQ_CHANGING is set, threads will check the current thread target
339  * whenever they wake up, and do whatever they can to apply its effects.
340  *
341  * TASKQ_THREAD_CPU_PCT --------------------------------------------------------
342  *
343  * When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
344  * percentage in tq_threads_ncpus_pct, start them off with the correct thread
345  * target, and add them to the taskq_cpupct_list for later adjustment.
346  *
347  * We register taskq_cpu_setup() to be called whenever a CPU changes state.  It
348  * walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
349  * if need be, and wakes up all of the threads to process the change.
350  *
351  * Dynamic Task Queues Implementation ------------------------------------------
352  *
353  * For a dynamic task queues there is a 1-to-1 mapping between a thread and
354  * taskq_ent_structure. Each entry is serviced by its own thread and each thread
355  * is controlled by a single entry.
356  *
357  * Entries are distributed over a set of buckets. To avoid using modulo
358  * arithmetics the number of buckets is 2^n and is determined as the nearest
359  * power of two roundown of the number of CPUs in the system. Tunable
360  * variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
361  * is attached to a bucket for its lifetime and can't migrate to other buckets.
362  *
363  * Entries that have scheduled tasks are not placed in any list. The dispatch
364  * function sets their "func" and "arg" fields and signals the corresponding
365  * thread to execute the task. Once the thread executes the task it clears the
366  * "func" field and places an entry on the bucket cache of free entries pointed
367  * by "tqbucket_freelist" field. ALL entries on the free list should have "func"
368  * field equal to NULL. The free list is a circular doubly-linked list identical
369  * in structure to the tq_task list above, but entries are taken from it in LIFO
370  * order - the last freed entry is the first to be allocated. The
371  * taskq_bucket_dispatch() function gets the most recently used entry from the
372  * free list, sets its "func" and "arg" fields and signals a worker thread.
373  *
374  * After executing each task a per-entry thread taskq_d_thread() places its
375  * entry on the bucket free list and goes to a timed sleep. If it wakes up
376  * without getting new task it removes the entry from the free list and destroys
377  * itself. The thread sleep time is controlled by a tunable variable
378  * `taskq_thread_timeout'.
379  *
380  * There are various statistics kept in the bucket which allows for later
381  * analysis of taskq usage patterns. Also, a global copy of taskq creation and
382  * death statistics is kept in the global taskq data structure. Since thread
383  * creation and death happen rarely, updating such global data does not present
384  * a performance problem.
385  *
386  * NOTE: Threads are not bound to any CPU and there is absolutely no association
387  *       between the bucket and actual thread CPU, so buckets are used only to
388  *	 split resources and reduce resource contention. Having threads attached
389  *	 to the CPU denoted by a bucket may reduce number of times the job
390  *	 switches between CPUs.
391  *
392  *	 Current algorithm creates a thread whenever a bucket has no free
393  *	 entries. It would be nice to know how many threads are in the running
394  *	 state and don't create threads if all CPUs are busy with existing
395  *	 tasks, but it is unclear how such strategy can be implemented.
396  *
397  *	 Currently buckets are created statically as an array attached to task
398  *	 queue. On some system with nCPUs < max_ncpus it may waste system
399  *	 memory. One solution may be allocation of buckets when they are first
400  *	 touched, but it is not clear how useful it is.
401  *
402  * SUSPEND/RESUME implementation -----------------------------------------------
403  *
404  *	Before executing a task taskq_thread() (executing non-dynamic task
405  *	queues) obtains taskq's thread lock as a reader. The taskq_suspend()
406  *	function gets the same lock as a writer blocking all non-dynamic task
407  *	execution. The taskq_resume() function releases the lock allowing
408  *	taskq_thread to continue execution.
409  *
410  *	For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
411  *	taskq_suspend() function. After that taskq_bucket_dispatch() always
412  *	fails, so that taskq_dispatch() will either enqueue tasks for a
413  *	suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
414  *	flags.
415  *
416  *	NOTE: taskq_suspend() does not immediately block any tasks already
417  *	      scheduled for dynamic task queues. It only suspends new tasks
418  *	      scheduled after taskq_suspend() was called.
419  *
420  *	taskq_member() function works by comparing a thread t_taskq pointer with
421  *	the passed thread pointer.
422  *
423  * LOCKS and LOCK Hierarchy ----------------------------------------------------
424  *
425  *   There are three locks used in task queues:
426  *
427  *   1) The taskq_t's tq_lock, protecting global task queue state.
428  *
429  *   2) Each per-CPU bucket has a lock for bucket management.
430  *
431  *   3) The global taskq_cpupct_lock, which protects the list of
432  *      TASKQ_THREADS_CPU_PCT taskqs.
433  *
434  *   If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
435  *   lock.
436  *
437  *   If both (1) and (3) are needed, tq_lock should be taken *after*
438  *   taskq_cpupct_lock.
439  *
440  * DEBUG FACILITIES ------------------------------------------------------------
441  *
442  * For DEBUG kernels it is possible to induce random failures to
443  * taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
444  * taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
445  * failures for dynamic and static task queues respectively.
446  *
447  * Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
448  *
449  * TUNABLES --------------------------------------------------------------------
450  *
451  *	system_taskq_size	- Size of the global system_taskq.
452  *				  This value is multiplied by nCPUs to determine
453  *				  actual size.
454  *				  Default value: 64
455  *
456  *	taskq_minimum_nthreads_max
457  *				- Minimum size of the thread list for a taskq.
458  *				  Useful for testing different thread pool
459  *				  sizes by overwriting tq_nthreads_target.
460  *
461  *	taskq_thread_timeout	- Maximum idle time for taskq_d_thread()
462  *				  Default value: 5 minutes
463  *
464  *	taskq_maxbuckets	- Maximum number of buckets in any task queue
465  *				  Default value: 128
466  *
467  *	taskq_search_depth	- Maximum # of buckets searched for a free entry
468  *				  Default value: 4
469  *
470  *	taskq_dmtbf		- Mean time between induced dispatch failures
471  *				  for dynamic task queues.
472  *				  Default value: UINT_MAX (no induced failures)
473  *
474  *	taskq_smtbf		- Mean time between induced dispatch failures
475  *				  for static task queues.
476  *				  Default value: UINT_MAX (no induced failures)
477  *
478  * CONDITIONAL compilation -----------------------------------------------------
479  *
480  *    TASKQ_STATISTIC	- If set will enable bucket statistic (default).
481  *
482  */
483 
484 #include <sys/taskq_impl.h>
485 #include <sys/thread.h>
486 #include <sys/proc.h>
487 #include <sys/kmem.h>
488 #include <sys/vmem.h>
489 #include <sys/callb.h>
490 #include <sys/class.h>
491 #include <sys/systm.h>
492 #include <sys/cmn_err.h>
493 #include <sys/debug.h>
494 #include <sys/vmsystm.h>	/* For throttlefree */
495 #include <sys/sysmacros.h>
496 #include <sys/cpuvar.h>
497 #include <sys/cpupart.h>
498 #include <sys/sdt.h>
499 #include <sys/sysdc.h>
500 #include <sys/note.h>
501 
502 static kmem_cache_t *taskq_ent_cache, *taskq_cache;
503 
504 /*
505  * Pseudo instance numbers for taskqs without explicitly provided instance.
506  */
507 static vmem_t *taskq_id_arena;
508 
509 /* Global system task queue for common use */
510 taskq_t	*system_taskq;
511 
512 /*
513  * Maximum number of entries in global system taskq is
514  *	system_taskq_size * max_ncpus
515  */
516 #define	SYSTEM_TASKQ_SIZE 64
517 int system_taskq_size = SYSTEM_TASKQ_SIZE;
518 
519 /*
520  * Minimum size for tq_nthreads_max; useful for those who want to play around
521  * with increasing a taskq's tq_nthreads_target.
522  */
523 int taskq_minimum_nthreads_max = 1;
524 
525 /*
526  * We want to ensure that when taskq_create() returns, there is at least
527  * one thread ready to handle requests.  To guarantee this, we have to wait
528  * for the second thread, since the first one cannot process requests until
529  * the second thread has been created.
530  */
531 #define	TASKQ_CREATE_ACTIVE_THREADS	2
532 
533 /* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
534 #define	TASKQ_CPUPCT_MAX_PERCENT	1000
535 int taskq_cpupct_max_percent = TASKQ_CPUPCT_MAX_PERCENT;
536 
537 /*
538  * Dynamic task queue threads that don't get any work within
539  * taskq_thread_timeout destroy themselves
540  */
541 #define	TASKQ_THREAD_TIMEOUT (60 * 5)
542 int taskq_thread_timeout = TASKQ_THREAD_TIMEOUT;
543 
544 #define	TASKQ_MAXBUCKETS 128
545 int taskq_maxbuckets = TASKQ_MAXBUCKETS;
546 
547 /*
548  * When a bucket has no available entries another buckets are tried.
549  * taskq_search_depth parameter limits the amount of buckets that we search
550  * before failing. This is mostly useful in systems with many CPUs where we may
551  * spend too much time scanning busy buckets.
552  */
553 #define	TASKQ_SEARCH_DEPTH 4
554 int taskq_search_depth = TASKQ_SEARCH_DEPTH;
555 
556 /*
557  * Hashing function: mix various bits of x. May be pretty much anything.
558  */
559 #define	TQ_HASH(x) ((x) ^ ((x) >> 11) ^ ((x) >> 17) ^ ((x) ^ 27))
560 
561 /*
562  * We do not create any new threads when the system is low on memory and start
563  * throttling memory allocations. The following macro tries to estimate such
564  * condition.
565  */
566 #define	ENOUGH_MEMORY() (freemem > throttlefree)
567 
568 /*
569  * Static functions.
570  */
571 static taskq_t	*taskq_create_common(const char *, int, int, pri_t, int,
572     int, proc_t *, uint_t, uint_t);
573 static void taskq_thread(void *);
574 static void taskq_d_thread(taskq_ent_t *);
575 static void taskq_bucket_extend(void *);
576 static int  taskq_constructor(void *, void *, int);
577 static void taskq_destructor(void *, void *);
578 static int  taskq_ent_constructor(void *, void *, int);
579 static void taskq_ent_destructor(void *, void *);
580 static taskq_ent_t *taskq_ent_alloc(taskq_t *, int);
581 static void taskq_ent_free(taskq_t *, taskq_ent_t *);
582 static int taskq_ent_exists(taskq_t *, task_func_t, void *);
583 static taskq_ent_t *taskq_bucket_dispatch(taskq_bucket_t *, task_func_t,
584     void *);
585 
586 /*
587  * Task queues kstats.
588  */
589 struct taskq_kstat {
590 	kstat_named_t	tq_pid;
591 	kstat_named_t	tq_tasks;
592 	kstat_named_t	tq_executed;
593 	kstat_named_t	tq_maxtasks;
594 	kstat_named_t	tq_totaltime;
595 	kstat_named_t	tq_nalloc;
596 	kstat_named_t	tq_nactive;
597 	kstat_named_t	tq_pri;
598 	kstat_named_t	tq_nthreads;
599 } taskq_kstat = {
600 	{ "pid",		KSTAT_DATA_UINT64 },
601 	{ "tasks",		KSTAT_DATA_UINT64 },
602 	{ "executed",		KSTAT_DATA_UINT64 },
603 	{ "maxtasks",		KSTAT_DATA_UINT64 },
604 	{ "totaltime",		KSTAT_DATA_UINT64 },
605 	{ "nactive",		KSTAT_DATA_UINT64 },
606 	{ "nalloc",		KSTAT_DATA_UINT64 },
607 	{ "priority",		KSTAT_DATA_UINT64 },
608 	{ "threads",		KSTAT_DATA_UINT64 },
609 };
610 
611 struct taskq_d_kstat {
612 	kstat_named_t	tqd_pri;
613 	kstat_named_t	tqd_btasks;
614 	kstat_named_t	tqd_bexecuted;
615 	kstat_named_t	tqd_bmaxtasks;
616 	kstat_named_t	tqd_bnalloc;
617 	kstat_named_t	tqd_bnactive;
618 	kstat_named_t	tqd_btotaltime;
619 	kstat_named_t	tqd_hits;
620 	kstat_named_t	tqd_misses;
621 	kstat_named_t	tqd_overflows;
622 	kstat_named_t	tqd_tcreates;
623 	kstat_named_t	tqd_tdeaths;
624 	kstat_named_t	tqd_maxthreads;
625 	kstat_named_t	tqd_nomem;
626 	kstat_named_t	tqd_disptcreates;
627 	kstat_named_t	tqd_totaltime;
628 	kstat_named_t	tqd_nalloc;
629 	kstat_named_t	tqd_nfree;
630 } taskq_d_kstat = {
631 	{ "priority",		KSTAT_DATA_UINT64 },
632 	{ "btasks",		KSTAT_DATA_UINT64 },
633 	{ "bexecuted",		KSTAT_DATA_UINT64 },
634 	{ "bmaxtasks",		KSTAT_DATA_UINT64 },
635 	{ "bnalloc",		KSTAT_DATA_UINT64 },
636 	{ "bnactive",		KSTAT_DATA_UINT64 },
637 	{ "btotaltime",		KSTAT_DATA_UINT64 },
638 	{ "hits",		KSTAT_DATA_UINT64 },
639 	{ "misses",		KSTAT_DATA_UINT64 },
640 	{ "overflows",		KSTAT_DATA_UINT64 },
641 	{ "tcreates",		KSTAT_DATA_UINT64 },
642 	{ "tdeaths",		KSTAT_DATA_UINT64 },
643 	{ "maxthreads",		KSTAT_DATA_UINT64 },
644 	{ "nomem",		KSTAT_DATA_UINT64 },
645 	{ "disptcreates",	KSTAT_DATA_UINT64 },
646 	{ "totaltime",		KSTAT_DATA_UINT64 },
647 	{ "nalloc",		KSTAT_DATA_UINT64 },
648 	{ "nfree",		KSTAT_DATA_UINT64 },
649 };
650 
651 static kmutex_t taskq_kstat_lock;
652 static kmutex_t taskq_d_kstat_lock;
653 static int taskq_kstat_update(kstat_t *, int);
654 static int taskq_d_kstat_update(kstat_t *, int);
655 
656 /*
657  * List of all TASKQ_THREADS_CPU_PCT taskqs.
658  */
659 static list_t taskq_cpupct_list;	/* protected by cpu_lock */
660 
661 /*
662  * Collect per-bucket statistic when TASKQ_STATISTIC is defined.
663  */
664 #define	TASKQ_STATISTIC 1
665 
666 #if TASKQ_STATISTIC
667 #define	TQ_STAT(b, x)	b->tqbucket_stat.x++
668 #else
669 #define	TQ_STAT(b, x)
670 #endif
671 
672 /*
673  * Random fault injection.
674  */
675 uint_t taskq_random;
676 uint_t taskq_dmtbf = UINT_MAX;    /* mean time between injected failures */
677 uint_t taskq_smtbf = UINT_MAX;    /* mean time between injected failures */
678 
679 /*
680  * TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
681  *
682  * TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
683  * they could prepopulate the cache and make sure that they do not use more
684  * then minalloc entries.  So, fault injection in this case insures that
685  * either TASKQ_PREPOPULATE is not set or there are more entries allocated
686  * than is specified by minalloc.  TQ_NOALLOC dispatches are always allowed
687  * to fail, but for simplicity we treat them identically to TQ_NOSLEEP
688  * dispatches.
689  */
690 #ifdef DEBUG
691 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)		\
692 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
693 	if ((flag & TQ_NOSLEEP) &&				\
694 	    taskq_random < 1771875 / taskq_dmtbf) {		\
695 		return (NULL);					\
696 	}
697 
698 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)		\
699 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
700 	if ((flag & (TQ_NOSLEEP | TQ_NOALLOC)) &&		\
701 	    (!(tq->tq_flags & TASKQ_PREPOPULATE) ||		\
702 	    (tq->tq_nalloc > tq->tq_minalloc)) &&		\
703 	    (taskq_random < (1771875 / taskq_smtbf))) {		\
704 		mutex_exit(&tq->tq_lock);			\
705 		return (NULL);					\
706 	}
707 #else
708 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)
709 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)
710 #endif
711 
712 #define	IS_EMPTY(l) (((l).tqent_prev == (l).tqent_next) &&	\
713 	((l).tqent_prev == &(l)))
714 
715 /*
716  * Append `tqe' in the end of the doubly-linked list denoted by l.
717  */
718 #define	TQ_APPEND(l, tqe) {					\
719 	tqe->tqent_next = &l;					\
720 	tqe->tqent_prev = l.tqent_prev;				\
721 	tqe->tqent_next->tqent_prev = tqe;			\
722 	tqe->tqent_prev->tqent_next = tqe;			\
723 }
724 /*
725  * Prepend 'tqe' to the beginning of l
726  */
727 #define	TQ_PREPEND(l, tqe) {					\
728 	tqe->tqent_next = l.tqent_next;				\
729 	tqe->tqent_prev = &l;					\
730 	tqe->tqent_next->tqent_prev = tqe;			\
731 	tqe->tqent_prev->tqent_next = tqe;			\
732 }
733 
734 /*
735  * Schedule a task specified by func and arg into the task queue entry tqe.
736  */
737 #define	TQ_DO_ENQUEUE(tq, tqe, func, arg, front) {			\
738 	ASSERT(MUTEX_HELD(&tq->tq_lock));				\
739 	_NOTE(CONSTCOND)						\
740 	if (front) {							\
741 		TQ_PREPEND(tq->tq_task, tqe);				\
742 	} else {							\
743 		TQ_APPEND(tq->tq_task, tqe);				\
744 	}								\
745 	tqe->tqent_func = (func);					\
746 	tqe->tqent_arg = (arg);						\
747 	tq->tq_tasks++;							\
748 	if (tq->tq_tasks - tq->tq_executed > tq->tq_maxtasks)		\
749 		tq->tq_maxtasks = tq->tq_tasks - tq->tq_executed;	\
750 	cv_signal(&tq->tq_dispatch_cv);					\
751 	DTRACE_PROBE2(taskq__enqueue, taskq_t *, tq, taskq_ent_t *, tqe); \
752 }
753 
754 #define	TQ_ENQUEUE(tq, tqe, func, arg)					\
755 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 0)
756 
757 #define	TQ_ENQUEUE_FRONT(tq, tqe, func, arg)				\
758 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 1)
759 
760 /*
761  * Do-nothing task which may be used to prepopulate thread caches.
762  */
763 /*ARGSUSED*/
764 void
765 nulltask(void *unused)
766 {
767 }
768 
769 /*ARGSUSED*/
770 static int
771 taskq_constructor(void *buf, void *cdrarg, int kmflags)
772 {
773 	taskq_t *tq = buf;
774 
775 	bzero(tq, sizeof (taskq_t));
776 
777 	mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
778 	rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
779 	cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
780 	cv_init(&tq->tq_exit_cv, NULL, CV_DEFAULT, NULL);
781 	cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
782 	cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
783 
784 	tq->tq_task.tqent_next = &tq->tq_task;
785 	tq->tq_task.tqent_prev = &tq->tq_task;
786 
787 	return (0);
788 }
789 
790 /*ARGSUSED*/
791 static void
792 taskq_destructor(void *buf, void *cdrarg)
793 {
794 	taskq_t *tq = buf;
795 
796 	ASSERT(tq->tq_nthreads == 0);
797 	ASSERT(tq->tq_buckets == NULL);
798 	ASSERT(tq->tq_tcreates == 0);
799 	ASSERT(tq->tq_tdeaths == 0);
800 
801 	mutex_destroy(&tq->tq_lock);
802 	rw_destroy(&tq->tq_threadlock);
803 	cv_destroy(&tq->tq_dispatch_cv);
804 	cv_destroy(&tq->tq_exit_cv);
805 	cv_destroy(&tq->tq_wait_cv);
806 	cv_destroy(&tq->tq_maxalloc_cv);
807 }
808 
809 /*ARGSUSED*/
810 static int
811 taskq_ent_constructor(void *buf, void *cdrarg, int kmflags)
812 {
813 	taskq_ent_t *tqe = buf;
814 
815 	tqe->tqent_thread = NULL;
816 	cv_init(&tqe->tqent_cv, NULL, CV_DEFAULT, NULL);
817 
818 	return (0);
819 }
820 
821 /*ARGSUSED*/
822 static void
823 taskq_ent_destructor(void *buf, void *cdrarg)
824 {
825 	taskq_ent_t *tqe = buf;
826 
827 	ASSERT(tqe->tqent_thread == NULL);
828 	cv_destroy(&tqe->tqent_cv);
829 }
830 
831 void
832 taskq_init(void)
833 {
834 	taskq_ent_cache = kmem_cache_create("taskq_ent_cache",
835 	    sizeof (taskq_ent_t), 0, taskq_ent_constructor,
836 	    taskq_ent_destructor, NULL, NULL, NULL, 0);
837 	taskq_cache = kmem_cache_create("taskq_cache", sizeof (taskq_t),
838 	    0, taskq_constructor, taskq_destructor, NULL, NULL, NULL, 0);
839 	taskq_id_arena = vmem_create("taskq_id_arena",
840 	    (void *)1, INT32_MAX, 1, NULL, NULL, NULL, 0,
841 	    VM_SLEEP | VMC_IDENTIFIER);
842 
843 	list_create(&taskq_cpupct_list, sizeof (taskq_t),
844 	    offsetof(taskq_t, tq_cpupct_link));
845 }
846 
847 static void
848 taskq_update_nthreads(taskq_t *tq, uint_t ncpus)
849 {
850 	uint_t newtarget = TASKQ_THREADS_PCT(ncpus, tq->tq_threads_ncpus_pct);
851 
852 	ASSERT(MUTEX_HELD(&cpu_lock));
853 	ASSERT(MUTEX_HELD(&tq->tq_lock));
854 
855 	/* We must be going from non-zero to non-zero; no exiting. */
856 	ASSERT3U(tq->tq_nthreads_target, !=, 0);
857 	ASSERT3U(newtarget, !=, 0);
858 
859 	ASSERT3U(newtarget, <=, tq->tq_nthreads_max);
860 	if (newtarget != tq->tq_nthreads_target) {
861 		tq->tq_flags |= TASKQ_CHANGING;
862 		tq->tq_nthreads_target = newtarget;
863 		cv_broadcast(&tq->tq_dispatch_cv);
864 		cv_broadcast(&tq->tq_exit_cv);
865 	}
866 }
867 
868 /* called during task queue creation */
869 static void
870 taskq_cpupct_install(taskq_t *tq, cpupart_t *cpup)
871 {
872 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
873 
874 	mutex_enter(&cpu_lock);
875 	mutex_enter(&tq->tq_lock);
876 	tq->tq_cpupart = cpup->cp_id;
877 	taskq_update_nthreads(tq, cpup->cp_ncpus);
878 	mutex_exit(&tq->tq_lock);
879 
880 	list_insert_tail(&taskq_cpupct_list, tq);
881 	mutex_exit(&cpu_lock);
882 }
883 
884 static void
885 taskq_cpupct_remove(taskq_t *tq)
886 {
887 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
888 
889 	mutex_enter(&cpu_lock);
890 	list_remove(&taskq_cpupct_list, tq);
891 	mutex_exit(&cpu_lock);
892 }
893 
894 /*ARGSUSED*/
895 static int
896 taskq_cpu_setup(cpu_setup_t what, int id, void *arg)
897 {
898 	taskq_t *tq;
899 	cpupart_t *cp = cpu[id]->cpu_part;
900 	uint_t ncpus = cp->cp_ncpus;
901 
902 	ASSERT(MUTEX_HELD(&cpu_lock));
903 	ASSERT(ncpus > 0);
904 
905 	switch (what) {
906 	case CPU_OFF:
907 	case CPU_CPUPART_OUT:
908 		/* offlines are called *before* the cpu is offlined. */
909 		if (ncpus > 1)
910 			ncpus--;
911 		break;
912 
913 	case CPU_ON:
914 	case CPU_CPUPART_IN:
915 		break;
916 
917 	default:
918 		return (0);		/* doesn't affect cpu count */
919 	}
920 
921 	for (tq = list_head(&taskq_cpupct_list); tq != NULL;
922 	    tq = list_next(&taskq_cpupct_list, tq)) {
923 
924 		mutex_enter(&tq->tq_lock);
925 		/*
926 		 * If the taskq is part of the cpuset which is changing,
927 		 * update its nthreads_target.
928 		 */
929 		if (tq->tq_cpupart == cp->cp_id) {
930 			taskq_update_nthreads(tq, ncpus);
931 		}
932 		mutex_exit(&tq->tq_lock);
933 	}
934 	return (0);
935 }
936 
937 void
938 taskq_mp_init(void)
939 {
940 	mutex_enter(&cpu_lock);
941 	register_cpu_setup_func(taskq_cpu_setup, NULL);
942 	/*
943 	 * Make sure we're up to date.  At this point in boot, there is only
944 	 * one processor set, so we only have to update the current CPU.
945 	 */
946 	(void) taskq_cpu_setup(CPU_ON, CPU->cpu_id, NULL);
947 	mutex_exit(&cpu_lock);
948 }
949 
950 /*
951  * Create global system dynamic task queue.
952  */
953 void
954 system_taskq_init(void)
955 {
956 	system_taskq = taskq_create_common("system_taskq", 0,
957 	    system_taskq_size * max_ncpus, minclsyspri, 4, 512, &p0, 0,
958 	    TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
959 }
960 
961 /*
962  * taskq_ent_alloc()
963  *
964  * Allocates a new taskq_ent_t structure either from the free list or from the
965  * cache. Returns NULL if it can't be allocated.
966  *
967  * Assumes: tq->tq_lock is held.
968  */
969 static taskq_ent_t *
970 taskq_ent_alloc(taskq_t *tq, int flags)
971 {
972 	int kmflags = (flags & TQ_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
973 	taskq_ent_t *tqe;
974 	clock_t wait_time;
975 	clock_t	wait_rv;
976 
977 	ASSERT(MUTEX_HELD(&tq->tq_lock));
978 
979 	/*
980 	 * TQ_NOALLOC allocations are allowed to use the freelist, even if
981 	 * we are below tq_minalloc.
982 	 */
983 again:	if ((tqe = tq->tq_freelist) != NULL &&
984 	    ((flags & TQ_NOALLOC) || tq->tq_nalloc >= tq->tq_minalloc)) {
985 		tq->tq_freelist = tqe->tqent_next;
986 	} else {
987 		if (flags & TQ_NOALLOC)
988 			return (NULL);
989 
990 		if (tq->tq_nalloc >= tq->tq_maxalloc) {
991 			if (kmflags & KM_NOSLEEP)
992 				return (NULL);
993 
994 			/*
995 			 * We don't want to exceed tq_maxalloc, but we can't
996 			 * wait for other tasks to complete (and thus free up
997 			 * task structures) without risking deadlock with
998 			 * the caller.  So, we just delay for one second
999 			 * to throttle the allocation rate. If we have tasks
1000 			 * complete before one second timeout expires then
1001 			 * taskq_ent_free will signal us and we will
1002 			 * immediately retry the allocation (reap free).
1003 			 */
1004 			wait_time = ddi_get_lbolt() + hz;
1005 			while (tq->tq_freelist == NULL) {
1006 				tq->tq_maxalloc_wait++;
1007 				wait_rv = cv_timedwait(&tq->tq_maxalloc_cv,
1008 				    &tq->tq_lock, wait_time);
1009 				tq->tq_maxalloc_wait--;
1010 				if (wait_rv == -1)
1011 					break;
1012 			}
1013 			if (tq->tq_freelist)
1014 				goto again;		/* reap freelist */
1015 
1016 		}
1017 		mutex_exit(&tq->tq_lock);
1018 
1019 		tqe = kmem_cache_alloc(taskq_ent_cache, kmflags);
1020 
1021 		mutex_enter(&tq->tq_lock);
1022 		if (tqe != NULL)
1023 			tq->tq_nalloc++;
1024 	}
1025 	return (tqe);
1026 }
1027 
1028 /*
1029  * taskq_ent_free()
1030  *
1031  * Free taskq_ent_t structure by either putting it on the free list or freeing
1032  * it to the cache.
1033  *
1034  * Assumes: tq->tq_lock is held.
1035  */
1036 static void
1037 taskq_ent_free(taskq_t *tq, taskq_ent_t *tqe)
1038 {
1039 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1040 
1041 	if (tq->tq_nalloc <= tq->tq_minalloc) {
1042 		tqe->tqent_next = tq->tq_freelist;
1043 		tq->tq_freelist = tqe;
1044 	} else {
1045 		tq->tq_nalloc--;
1046 		mutex_exit(&tq->tq_lock);
1047 		kmem_cache_free(taskq_ent_cache, tqe);
1048 		mutex_enter(&tq->tq_lock);
1049 	}
1050 
1051 	if (tq->tq_maxalloc_wait)
1052 		cv_signal(&tq->tq_maxalloc_cv);
1053 }
1054 
1055 /*
1056  * taskq_ent_exists()
1057  *
1058  * Return 1 if taskq already has entry for calling 'func(arg)'.
1059  *
1060  * Assumes: tq->tq_lock is held.
1061  */
1062 static int
1063 taskq_ent_exists(taskq_t *tq, task_func_t func, void *arg)
1064 {
1065 	taskq_ent_t	*tqe;
1066 
1067 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1068 
1069 	for (tqe = tq->tq_task.tqent_next; tqe != &tq->tq_task;
1070 	    tqe = tqe->tqent_next)
1071 		if ((tqe->tqent_func == func) && (tqe->tqent_arg == arg))
1072 			return (1);
1073 	return (0);
1074 }
1075 
1076 /*
1077  * Dispatch a task "func(arg)" to a free entry of bucket b.
1078  *
1079  * Assumes: no bucket locks is held.
1080  *
1081  * Returns: a pointer to an entry if dispatch was successful.
1082  *	    NULL if there are no free entries or if the bucket is suspended.
1083  */
1084 static taskq_ent_t *
1085 taskq_bucket_dispatch(taskq_bucket_t *b, task_func_t func, void *arg)
1086 {
1087 	taskq_ent_t *tqe;
1088 
1089 	ASSERT(MUTEX_NOT_HELD(&b->tqbucket_lock));
1090 	ASSERT(func != NULL);
1091 
1092 	mutex_enter(&b->tqbucket_lock);
1093 
1094 	ASSERT(b->tqbucket_nfree != 0 || IS_EMPTY(b->tqbucket_freelist));
1095 	ASSERT(b->tqbucket_nfree == 0 || !IS_EMPTY(b->tqbucket_freelist));
1096 
1097 	/*
1098 	 * Get en entry from the freelist if there is one.
1099 	 * Schedule task into the entry.
1100 	 */
1101 	if ((b->tqbucket_nfree != 0) &&
1102 	    !(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
1103 		tqe = b->tqbucket_freelist.tqent_prev;
1104 
1105 		ASSERT(tqe != &b->tqbucket_freelist);
1106 		ASSERT(tqe->tqent_thread != NULL);
1107 
1108 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1109 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1110 		b->tqbucket_nalloc++;
1111 		b->tqbucket_nfree--;
1112 		tqe->tqent_func = func;
1113 		tqe->tqent_arg = arg;
1114 		TQ_STAT(b, tqs_hits);
1115 		cv_signal(&tqe->tqent_cv);
1116 		DTRACE_PROBE2(taskq__d__enqueue, taskq_bucket_t *, b,
1117 		    taskq_ent_t *, tqe);
1118 	} else {
1119 		tqe = NULL;
1120 		TQ_STAT(b, tqs_misses);
1121 	}
1122 	mutex_exit(&b->tqbucket_lock);
1123 	return (tqe);
1124 }
1125 
1126 /*
1127  * Dispatch a task.
1128  *
1129  * Assumes: func != NULL
1130  *
1131  * Returns: NULL if dispatch failed.
1132  *	    non-NULL if task dispatched successfully.
1133  *	    Actual return value is the pointer to taskq entry that was used to
1134  *	    dispatch a task. This is useful for debugging.
1135  */
1136 taskqid_t
1137 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
1138 {
1139 	taskq_bucket_t *bucket = NULL;	/* Which bucket needs extension */
1140 	taskq_ent_t *tqe = NULL;
1141 	taskq_ent_t *tqe1;
1142 	uint_t bsize;
1143 
1144 	ASSERT(tq != NULL);
1145 	ASSERT(func != NULL);
1146 
1147 	if (!(tq->tq_flags & TASKQ_DYNAMIC)) {
1148 		/*
1149 		 * TQ_NOQUEUE flag can't be used with non-dynamic task queues.
1150 		 */
1151 		ASSERT(!(flags & TQ_NOQUEUE));
1152 		/*
1153 		 * Enqueue the task to the underlying queue.
1154 		 */
1155 		mutex_enter(&tq->tq_lock);
1156 
1157 		TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flags);
1158 
1159 		if ((tqe = taskq_ent_alloc(tq, flags)) == NULL) {
1160 			mutex_exit(&tq->tq_lock);
1161 			return (NULL);
1162 		}
1163 		/* Make sure we start without any flags */
1164 		tqe->tqent_un.tqent_flags = 0;
1165 
1166 		if (flags & TQ_FRONT) {
1167 			TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1168 		} else {
1169 			TQ_ENQUEUE(tq, tqe, func, arg);
1170 		}
1171 		mutex_exit(&tq->tq_lock);
1172 		return ((taskqid_t)tqe);
1173 	}
1174 
1175 	/*
1176 	 * Dynamic taskq dispatching.
1177 	 */
1178 	ASSERT(!(flags & (TQ_NOALLOC | TQ_FRONT)));
1179 	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flags);
1180 
1181 	bsize = tq->tq_nbuckets;
1182 
1183 	if (bsize == 1) {
1184 		/*
1185 		 * In a single-CPU case there is only one bucket, so get
1186 		 * entry directly from there.
1187 		 */
1188 		if ((tqe = taskq_bucket_dispatch(tq->tq_buckets, func, arg))
1189 		    != NULL)
1190 			return ((taskqid_t)tqe);	/* Fastpath */
1191 		bucket = tq->tq_buckets;
1192 	} else {
1193 		int loopcount;
1194 		taskq_bucket_t *b;
1195 		uintptr_t h = ((uintptr_t)CPU + (uintptr_t)arg) >> 3;
1196 
1197 		h = TQ_HASH(h);
1198 
1199 		/*
1200 		 * The 'bucket' points to the original bucket that we hit. If we
1201 		 * can't allocate from it, we search other buckets, but only
1202 		 * extend this one.
1203 		 */
1204 		b = &tq->tq_buckets[h & (bsize - 1)];
1205 		ASSERT(b->tqbucket_taskq == tq);	/* Sanity check */
1206 
1207 		/*
1208 		 * Do a quick check before grabbing the lock. If the bucket does
1209 		 * not have free entries now, chances are very small that it
1210 		 * will after we take the lock, so we just skip it.
1211 		 */
1212 		if (b->tqbucket_nfree != 0) {
1213 			if ((tqe = taskq_bucket_dispatch(b, func, arg)) != NULL)
1214 				return ((taskqid_t)tqe);	/* Fastpath */
1215 		} else {
1216 			TQ_STAT(b, tqs_misses);
1217 		}
1218 
1219 		bucket = b;
1220 		loopcount = MIN(taskq_search_depth, bsize);
1221 		/*
1222 		 * If bucket dispatch failed, search loopcount number of buckets
1223 		 * before we give up and fail.
1224 		 */
1225 		do {
1226 			b = &tq->tq_buckets[++h & (bsize - 1)];
1227 			ASSERT(b->tqbucket_taskq == tq);  /* Sanity check */
1228 			loopcount--;
1229 
1230 			if (b->tqbucket_nfree != 0) {
1231 				tqe = taskq_bucket_dispatch(b, func, arg);
1232 			} else {
1233 				TQ_STAT(b, tqs_misses);
1234 			}
1235 		} while ((tqe == NULL) && (loopcount > 0));
1236 	}
1237 
1238 	/*
1239 	 * At this point we either scheduled a task and (tqe != NULL) or failed
1240 	 * (tqe == NULL). Try to recover from fails.
1241 	 */
1242 
1243 	/*
1244 	 * For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
1245 	 */
1246 	if ((tqe == NULL) && !(flags & TQ_NOSLEEP)) {
1247 		/*
1248 		 * taskq_bucket_extend() may fail to do anything, but this is
1249 		 * fine - we deal with it later. If the bucket was successfully
1250 		 * extended, there is a good chance that taskq_bucket_dispatch()
1251 		 * will get this new entry, unless someone is racing with us and
1252 		 * stealing the new entry from under our nose.
1253 		 * taskq_bucket_extend() may sleep.
1254 		 */
1255 		taskq_bucket_extend(bucket);
1256 		TQ_STAT(bucket, tqs_disptcreates);
1257 		if ((tqe = taskq_bucket_dispatch(bucket, func, arg)) != NULL)
1258 			return ((taskqid_t)tqe);
1259 	}
1260 
1261 	ASSERT(bucket != NULL);
1262 
1263 	/*
1264 	 * Since there are not enough free entries in the bucket, add a
1265 	 * taskq entry to extend it in the background using backing queue
1266 	 * (unless we already have a taskq entry to perform that extension).
1267 	 */
1268 	mutex_enter(&tq->tq_lock);
1269 	if (!taskq_ent_exists(tq, taskq_bucket_extend, bucket)) {
1270 		if ((tqe1 = taskq_ent_alloc(tq, TQ_NOSLEEP)) != NULL) {
1271 			TQ_ENQUEUE_FRONT(tq, tqe1, taskq_bucket_extend, bucket);
1272 		} else {
1273 			TQ_STAT(bucket, tqs_nomem);
1274 		}
1275 	}
1276 
1277 	/*
1278 	 * Dispatch failed and we can't find an entry to schedule a task.
1279 	 * Revert to the backing queue unless TQ_NOQUEUE was asked.
1280 	 */
1281 	if ((tqe == NULL) && !(flags & TQ_NOQUEUE)) {
1282 		if ((tqe = taskq_ent_alloc(tq, flags)) != NULL) {
1283 			TQ_ENQUEUE(tq, tqe, func, arg);
1284 		} else {
1285 			TQ_STAT(bucket, tqs_nomem);
1286 		}
1287 	}
1288 	mutex_exit(&tq->tq_lock);
1289 
1290 	return ((taskqid_t)tqe);
1291 }
1292 
1293 void
1294 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
1295     taskq_ent_t *tqe)
1296 {
1297 	ASSERT(func != NULL);
1298 	ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
1299 
1300 	/*
1301 	 * Mark it as a prealloc'd task.  This is important
1302 	 * to ensure that we don't free it later.
1303 	 */
1304 	tqe->tqent_un.tqent_flags |= TQENT_FLAG_PREALLOC;
1305 	/*
1306 	 * Enqueue the task to the underlying queue.
1307 	 */
1308 	mutex_enter(&tq->tq_lock);
1309 
1310 	if (flags & TQ_FRONT) {
1311 		TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1312 	} else {
1313 		TQ_ENQUEUE(tq, tqe, func, arg);
1314 	}
1315 	mutex_exit(&tq->tq_lock);
1316 }
1317 
1318 /*
1319  * Wait for all pending tasks to complete.
1320  * Calling taskq_wait from a task will cause deadlock.
1321  */
1322 void
1323 taskq_wait(taskq_t *tq)
1324 {
1325 	ASSERT(tq != curthread->t_taskq);
1326 
1327 	mutex_enter(&tq->tq_lock);
1328 	while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
1329 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1330 	mutex_exit(&tq->tq_lock);
1331 
1332 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1333 		taskq_bucket_t *b = tq->tq_buckets;
1334 		int bid = 0;
1335 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1336 			mutex_enter(&b->tqbucket_lock);
1337 			while (b->tqbucket_nalloc > 0)
1338 				cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
1339 			mutex_exit(&b->tqbucket_lock);
1340 		}
1341 	}
1342 }
1343 
1344 /*
1345  * Suspend execution of tasks.
1346  *
1347  * Tasks in the queue part will be suspended immediately upon return from this
1348  * function. Pending tasks in the dynamic part will continue to execute, but all
1349  * new tasks will  be suspended.
1350  */
1351 void
1352 taskq_suspend(taskq_t *tq)
1353 {
1354 	rw_enter(&tq->tq_threadlock, RW_WRITER);
1355 
1356 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1357 		taskq_bucket_t *b = tq->tq_buckets;
1358 		int bid = 0;
1359 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1360 			mutex_enter(&b->tqbucket_lock);
1361 			b->tqbucket_flags |= TQBUCKET_SUSPEND;
1362 			mutex_exit(&b->tqbucket_lock);
1363 		}
1364 	}
1365 	/*
1366 	 * Mark task queue as being suspended. Needed for taskq_suspended().
1367 	 */
1368 	mutex_enter(&tq->tq_lock);
1369 	ASSERT(!(tq->tq_flags & TASKQ_SUSPENDED));
1370 	tq->tq_flags |= TASKQ_SUSPENDED;
1371 	mutex_exit(&tq->tq_lock);
1372 }
1373 
1374 /*
1375  * returns: 1 if tq is suspended, 0 otherwise.
1376  */
1377 int
1378 taskq_suspended(taskq_t *tq)
1379 {
1380 	return ((tq->tq_flags & TASKQ_SUSPENDED) != 0);
1381 }
1382 
1383 /*
1384  * Resume taskq execution.
1385  */
1386 void
1387 taskq_resume(taskq_t *tq)
1388 {
1389 	ASSERT(RW_WRITE_HELD(&tq->tq_threadlock));
1390 
1391 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1392 		taskq_bucket_t *b = tq->tq_buckets;
1393 		int bid = 0;
1394 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1395 			mutex_enter(&b->tqbucket_lock);
1396 			b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
1397 			mutex_exit(&b->tqbucket_lock);
1398 		}
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 	rw_exit(&tq->tq_threadlock);
1406 }
1407 
1408 int
1409 taskq_member(taskq_t *tq, kthread_t *thread)
1410 {
1411 	return (thread->t_taskq == tq);
1412 }
1413 
1414 /*
1415  * Creates a thread in the taskq.  We only allow one outstanding create at
1416  * a time.  We drop and reacquire the tq_lock in order to avoid blocking other
1417  * taskq activity while thread_create() or lwp_kernel_create() run.
1418  *
1419  * The first time we're called, we do some additional setup, and do not
1420  * return until there are enough threads to start servicing requests.
1421  */
1422 static void
1423 taskq_thread_create(taskq_t *tq)
1424 {
1425 	kthread_t	*t;
1426 	const boolean_t	first = (tq->tq_nthreads == 0);
1427 
1428 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1429 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1430 	ASSERT(tq->tq_nthreads < tq->tq_nthreads_target);
1431 	ASSERT(!(tq->tq_flags & TASKQ_THREAD_CREATED));
1432 
1433 
1434 	tq->tq_flags |= TASKQ_THREAD_CREATED;
1435 	tq->tq_active++;
1436 	mutex_exit(&tq->tq_lock);
1437 
1438 	if (tq->tq_proc != &p0) {
1439 		t = lwp_kernel_create(tq->tq_proc, taskq_thread, tq, TS_RUN,
1440 		    tq->tq_pri);
1441 	} else {
1442 		t = thread_create(NULL, 0, taskq_thread, tq, 0, &p0, TS_RUN,
1443 		    tq->tq_pri);
1444 	}
1445 
1446 	if (!first) {
1447 		mutex_enter(&tq->tq_lock);
1448 		return;
1449 	}
1450 
1451 	/*
1452 	 * We know the thread cannot go away, since tq cannot be
1453 	 * destroyed until creation has completed.  We can therefore
1454 	 * safely dereference t.
1455 	 */
1456 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1457 		taskq_cpupct_install(tq, t->t_cpupart);
1458 	}
1459 	mutex_enter(&tq->tq_lock);
1460 
1461 	/* Wait until we can service requests. */
1462 	while (tq->tq_nthreads != tq->tq_nthreads_target &&
1463 	    tq->tq_nthreads < TASKQ_CREATE_ACTIVE_THREADS) {
1464 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1465 	}
1466 }
1467 
1468 /*
1469  * Common "sleep taskq thread" function, which handles CPR stuff, as well
1470  * as giving a nice common point for debuggers to find inactive threads.
1471  */
1472 static clock_t
1473 taskq_thread_wait(taskq_t *tq, kmutex_t *mx, kcondvar_t *cv,
1474     callb_cpr_t *cprinfo, clock_t timeout)
1475 {
1476 	clock_t ret = 0;
1477 
1478 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1479 		CALLB_CPR_SAFE_BEGIN(cprinfo);
1480 	}
1481 	if (timeout < 0)
1482 		cv_wait(cv, mx);
1483 	else
1484 		ret = cv_reltimedwait(cv, mx, timeout, TR_CLOCK_TICK);
1485 
1486 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1487 		CALLB_CPR_SAFE_END(cprinfo, mx);
1488 	}
1489 
1490 	return (ret);
1491 }
1492 
1493 /*
1494  * Worker thread for processing task queue.
1495  */
1496 static void
1497 taskq_thread(void *arg)
1498 {
1499 	int thread_id;
1500 
1501 	taskq_t *tq = arg;
1502 	taskq_ent_t *tqe;
1503 	callb_cpr_t cprinfo;
1504 	hrtime_t start, end;
1505 	boolean_t freeit;
1506 
1507 	curthread->t_taskq = tq;	/* mark ourselves for taskq_member() */
1508 
1509 	if (curproc != &p0 && (tq->tq_flags & TASKQ_DUTY_CYCLE)) {
1510 		sysdc_thread_enter(curthread, tq->tq_DC,
1511 		    (tq->tq_flags & TASKQ_DC_BATCH) ? SYSDC_THREAD_BATCH : 0);
1512 	}
1513 
1514 	if (tq->tq_flags & TASKQ_CPR_SAFE) {
1515 		CALLB_CPR_INIT_SAFE(curthread, tq->tq_name);
1516 	} else {
1517 		CALLB_CPR_INIT(&cprinfo, &tq->tq_lock, callb_generic_cpr,
1518 		    tq->tq_name);
1519 	}
1520 	mutex_enter(&tq->tq_lock);
1521 	thread_id = ++tq->tq_nthreads;
1522 	ASSERT(tq->tq_flags & TASKQ_THREAD_CREATED);
1523 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1524 	tq->tq_flags &= ~TASKQ_THREAD_CREATED;
1525 
1526 	VERIFY3S(thread_id, <=, tq->tq_nthreads_max);
1527 
1528 	if (tq->tq_nthreads_max == 1)
1529 		tq->tq_thread = curthread;
1530 	else
1531 		tq->tq_threadlist[thread_id - 1] = curthread;
1532 
1533 	/* Allow taskq_create_common()'s taskq_thread_create() to return. */
1534 	if (tq->tq_nthreads == TASKQ_CREATE_ACTIVE_THREADS)
1535 		cv_broadcast(&tq->tq_wait_cv);
1536 
1537 	for (;;) {
1538 		if (tq->tq_flags & TASKQ_CHANGING) {
1539 			/* See if we're no longer needed */
1540 			if (thread_id > tq->tq_nthreads_target) {
1541 				/*
1542 				 * To preserve the one-to-one mapping between
1543 				 * thread_id and thread, we must exit from
1544 				 * highest thread ID to least.
1545 				 *
1546 				 * However, if everyone is exiting, the order
1547 				 * doesn't matter, so just exit immediately.
1548 				 * (this is safe, since you must wait for
1549 				 * nthreads to reach 0 after setting
1550 				 * tq_nthreads_target to 0)
1551 				 */
1552 				if (thread_id == tq->tq_nthreads ||
1553 				    tq->tq_nthreads_target == 0)
1554 					break;
1555 
1556 				/* Wait for higher thread_ids to exit */
1557 				(void) taskq_thread_wait(tq, &tq->tq_lock,
1558 				    &tq->tq_exit_cv, &cprinfo, -1);
1559 				continue;
1560 			}
1561 
1562 			/*
1563 			 * If no thread is starting taskq_thread(), we can
1564 			 * do some bookkeeping.
1565 			 */
1566 			if (!(tq->tq_flags & TASKQ_THREAD_CREATED)) {
1567 				/* Check if we've reached our target */
1568 				if (tq->tq_nthreads == tq->tq_nthreads_target) {
1569 					tq->tq_flags &= ~TASKQ_CHANGING;
1570 					cv_broadcast(&tq->tq_wait_cv);
1571 				}
1572 				/* Check if we need to create a thread */
1573 				if (tq->tq_nthreads < tq->tq_nthreads_target) {
1574 					taskq_thread_create(tq);
1575 					continue; /* tq_lock was dropped */
1576 				}
1577 			}
1578 		}
1579 		if ((tqe = tq->tq_task.tqent_next) == &tq->tq_task) {
1580 			if (--tq->tq_active == 0)
1581 				cv_broadcast(&tq->tq_wait_cv);
1582 			(void) taskq_thread_wait(tq, &tq->tq_lock,
1583 			    &tq->tq_dispatch_cv, &cprinfo, -1);
1584 			tq->tq_active++;
1585 			continue;
1586 		}
1587 
1588 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1589 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1590 		mutex_exit(&tq->tq_lock);
1591 
1592 		/*
1593 		 * For prealloc'd tasks, we don't free anything.  We
1594 		 * have to check this now, because once we call the
1595 		 * function for a prealloc'd taskq, we can't touch the
1596 		 * tqent any longer (calling the function returns the
1597 		 * ownershp of the tqent back to caller of
1598 		 * taskq_dispatch.)
1599 		 */
1600 		if ((!(tq->tq_flags & TASKQ_DYNAMIC)) &&
1601 		    (tqe->tqent_un.tqent_flags & TQENT_FLAG_PREALLOC)) {
1602 			/* clear pointers to assist assertion checks */
1603 			tqe->tqent_next = tqe->tqent_prev = NULL;
1604 			freeit = B_FALSE;
1605 		} else {
1606 			freeit = B_TRUE;
1607 		}
1608 
1609 		rw_enter(&tq->tq_threadlock, RW_READER);
1610 		start = gethrtime();
1611 		DTRACE_PROBE2(taskq__exec__start, taskq_t *, tq,
1612 		    taskq_ent_t *, tqe);
1613 		tqe->tqent_func(tqe->tqent_arg);
1614 		DTRACE_PROBE2(taskq__exec__end, taskq_t *, tq,
1615 		    taskq_ent_t *, tqe);
1616 		end = gethrtime();
1617 		rw_exit(&tq->tq_threadlock);
1618 
1619 		mutex_enter(&tq->tq_lock);
1620 		tq->tq_totaltime += end - start;
1621 		tq->tq_executed++;
1622 
1623 		if (freeit)
1624 			taskq_ent_free(tq, tqe);
1625 	}
1626 
1627 	if (tq->tq_nthreads_max == 1)
1628 		tq->tq_thread = NULL;
1629 	else
1630 		tq->tq_threadlist[thread_id - 1] = NULL;
1631 
1632 	/* We're exiting, and therefore no longer active */
1633 	ASSERT(tq->tq_active > 0);
1634 	tq->tq_active--;
1635 
1636 	ASSERT(tq->tq_nthreads > 0);
1637 	tq->tq_nthreads--;
1638 
1639 	/* Wake up anyone waiting for us to exit */
1640 	cv_broadcast(&tq->tq_exit_cv);
1641 	if (tq->tq_nthreads == tq->tq_nthreads_target) {
1642 		if (!(tq->tq_flags & TASKQ_THREAD_CREATED))
1643 			tq->tq_flags &= ~TASKQ_CHANGING;
1644 
1645 		cv_broadcast(&tq->tq_wait_cv);
1646 	}
1647 
1648 	ASSERT(!(tq->tq_flags & TASKQ_CPR_SAFE));
1649 	CALLB_CPR_EXIT(&cprinfo);		/* drops tq->tq_lock */
1650 	if (curthread->t_lwp != NULL) {
1651 		mutex_enter(&curproc->p_lock);
1652 		lwp_exit();
1653 	} else {
1654 		thread_exit();
1655 	}
1656 }
1657 
1658 /*
1659  * Worker per-entry thread for dynamic dispatches.
1660  */
1661 static void
1662 taskq_d_thread(taskq_ent_t *tqe)
1663 {
1664 	taskq_bucket_t	*bucket = tqe->tqent_un.tqent_bucket;
1665 	taskq_t		*tq = bucket->tqbucket_taskq;
1666 	kmutex_t	*lock = &bucket->tqbucket_lock;
1667 	kcondvar_t	*cv = &tqe->tqent_cv;
1668 	callb_cpr_t	cprinfo;
1669 	clock_t		w;
1670 
1671 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, tq->tq_name);
1672 
1673 	mutex_enter(lock);
1674 
1675 	for (;;) {
1676 		/*
1677 		 * If a task is scheduled (func != NULL), execute it, otherwise
1678 		 * sleep, waiting for a job.
1679 		 */
1680 		if (tqe->tqent_func != NULL) {
1681 			hrtime_t	start;
1682 			hrtime_t	end;
1683 
1684 			ASSERT(bucket->tqbucket_nalloc > 0);
1685 
1686 			/*
1687 			 * It is possible to free the entry right away before
1688 			 * actually executing the task so that subsequent
1689 			 * dispatches may immediately reuse it. But this,
1690 			 * effectively, creates a two-length queue in the entry
1691 			 * and may lead to a deadlock if the execution of the
1692 			 * current task depends on the execution of the next
1693 			 * scheduled task. So, we keep the entry busy until the
1694 			 * task is processed.
1695 			 */
1696 
1697 			mutex_exit(lock);
1698 			start = gethrtime();
1699 			DTRACE_PROBE3(taskq__d__exec__start, taskq_t *, tq,
1700 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1701 			tqe->tqent_func(tqe->tqent_arg);
1702 			DTRACE_PROBE3(taskq__d__exec__end, taskq_t *, tq,
1703 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1704 			end = gethrtime();
1705 			mutex_enter(lock);
1706 			bucket->tqbucket_totaltime += end - start;
1707 
1708 			/*
1709 			 * Return the entry to the bucket free list.
1710 			 */
1711 			tqe->tqent_func = NULL;
1712 			TQ_APPEND(bucket->tqbucket_freelist, tqe);
1713 			bucket->tqbucket_nalloc--;
1714 			bucket->tqbucket_nfree++;
1715 			ASSERT(!IS_EMPTY(bucket->tqbucket_freelist));
1716 			/*
1717 			 * taskq_wait() waits for nalloc to drop to zero on
1718 			 * tqbucket_cv.
1719 			 */
1720 			cv_signal(&bucket->tqbucket_cv);
1721 		}
1722 
1723 		/*
1724 		 * At this point the entry must be in the bucket free list -
1725 		 * either because it was there initially or because it just
1726 		 * finished executing a task and put itself on the free list.
1727 		 */
1728 		ASSERT(bucket->tqbucket_nfree > 0);
1729 		/*
1730 		 * Go to sleep unless we are closing.
1731 		 * If a thread is sleeping too long, it dies.
1732 		 */
1733 		if (! (bucket->tqbucket_flags & TQBUCKET_CLOSE)) {
1734 			w = taskq_thread_wait(tq, lock, cv,
1735 			    &cprinfo, taskq_thread_timeout * hz);
1736 		}
1737 
1738 		/*
1739 		 * At this point we may be in two different states:
1740 		 *
1741 		 * (1) tqent_func is set which means that a new task is
1742 		 *	dispatched and we need to execute it.
1743 		 *
1744 		 * (2) Thread is sleeping for too long or we are closing. In
1745 		 *	both cases destroy the thread and the entry.
1746 		 */
1747 
1748 		/* If func is NULL we should be on the freelist. */
1749 		ASSERT((tqe->tqent_func != NULL) ||
1750 		    (bucket->tqbucket_nfree > 0));
1751 		/* If func is non-NULL we should be allocated */
1752 		ASSERT((tqe->tqent_func == NULL) ||
1753 		    (bucket->tqbucket_nalloc > 0));
1754 
1755 		/* Check freelist consistency */
1756 		ASSERT((bucket->tqbucket_nfree > 0) ||
1757 		    IS_EMPTY(bucket->tqbucket_freelist));
1758 		ASSERT((bucket->tqbucket_nfree == 0) ||
1759 		    !IS_EMPTY(bucket->tqbucket_freelist));
1760 
1761 		if ((tqe->tqent_func == NULL) &&
1762 		    ((w == -1) || (bucket->tqbucket_flags & TQBUCKET_CLOSE))) {
1763 			/*
1764 			 * This thread is sleeping for too long or we are
1765 			 * closing - time to die.
1766 			 * Thread creation/destruction happens rarely,
1767 			 * so grabbing the lock is not a big performance issue.
1768 			 * The bucket lock is dropped by CALLB_CPR_EXIT().
1769 			 */
1770 
1771 			/* Remove the entry from the free list. */
1772 			tqe->tqent_prev->tqent_next = tqe->tqent_next;
1773 			tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1774 			ASSERT(bucket->tqbucket_nfree > 0);
1775 			bucket->tqbucket_nfree--;
1776 
1777 			TQ_STAT(bucket, tqs_tdeaths);
1778 			cv_signal(&bucket->tqbucket_cv);
1779 			tqe->tqent_thread = NULL;
1780 			mutex_enter(&tq->tq_lock);
1781 			tq->tq_tdeaths++;
1782 			mutex_exit(&tq->tq_lock);
1783 			CALLB_CPR_EXIT(&cprinfo);
1784 			kmem_cache_free(taskq_ent_cache, tqe);
1785 			thread_exit();
1786 		}
1787 	}
1788 }
1789 
1790 
1791 /*
1792  * Taskq creation. May sleep for memory.
1793  * Always use automatically generated instances to avoid kstat name space
1794  * collisions.
1795  */
1796 
1797 taskq_t *
1798 taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
1799     int maxalloc, uint_t flags)
1800 {
1801 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1802 
1803 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1804 	    maxalloc, &p0, 0, flags | TASKQ_NOINSTANCE));
1805 }
1806 
1807 /*
1808  * Create an instance of task queue. It is legal to create task queues with the
1809  * same name and different instances.
1810  *
1811  * taskq_create_instance is used by ddi_taskq_create() where it gets the
1812  * instance from ddi_get_instance(). In some cases the instance is not
1813  * initialized and is set to -1. This case is handled as if no instance was
1814  * passed at all.
1815  */
1816 taskq_t *
1817 taskq_create_instance(const char *name, int instance, int nthreads, pri_t pri,
1818     int minalloc, int maxalloc, uint_t flags)
1819 {
1820 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1821 	ASSERT((instance >= 0) || (instance == -1));
1822 
1823 	if (instance < 0) {
1824 		flags |= TASKQ_NOINSTANCE;
1825 	}
1826 
1827 	return (taskq_create_common(name, instance, nthreads,
1828 	    pri, minalloc, maxalloc, &p0, 0, flags));
1829 }
1830 
1831 taskq_t *
1832 taskq_create_proc(const char *name, int nthreads, pri_t pri, int minalloc,
1833     int maxalloc, proc_t *proc, uint_t flags)
1834 {
1835 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1836 	ASSERT(proc->p_flag & SSYS);
1837 
1838 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1839 	    maxalloc, proc, 0, flags | TASKQ_NOINSTANCE));
1840 }
1841 
1842 taskq_t *
1843 taskq_create_sysdc(const char *name, int nthreads, int minalloc,
1844     int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1845 {
1846 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1847 	ASSERT(proc->p_flag & SSYS);
1848 
1849 	return (taskq_create_common(name, 0, nthreads, minclsyspri, minalloc,
1850 	    maxalloc, proc, dc, flags | TASKQ_NOINSTANCE | TASKQ_DUTY_CYCLE));
1851 }
1852 
1853 static taskq_t *
1854 taskq_create_common(const char *name, int instance, int nthreads, pri_t pri,
1855     int minalloc, int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1856 {
1857 	taskq_t *tq = kmem_cache_alloc(taskq_cache, KM_SLEEP);
1858 	uint_t ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
1859 	uint_t bsize;	/* # of buckets - always power of 2 */
1860 	int max_nthreads;
1861 
1862 	/*
1863 	 * TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
1864 	 * mutually incompatible.
1865 	 */
1866 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_CPR_SAFE));
1867 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_THREADS_CPU_PCT));
1868 	IMPLY((flags & TASKQ_CPR_SAFE), !(flags & TASKQ_THREADS_CPU_PCT));
1869 
1870 	/* Cannot have DUTY_CYCLE without a non-p0 kernel process */
1871 	IMPLY((flags & TASKQ_DUTY_CYCLE), proc != &p0);
1872 
1873 	/* Cannot have DC_BATCH without DUTY_CYCLE */
1874 	ASSERT((flags & (TASKQ_DUTY_CYCLE|TASKQ_DC_BATCH)) != TASKQ_DC_BATCH);
1875 
1876 	ASSERT(proc != NULL);
1877 
1878 	bsize = 1 << (highbit(ncpus) - 1);
1879 	ASSERT(bsize >= 1);
1880 	bsize = MIN(bsize, taskq_maxbuckets);
1881 
1882 	if (flags & TASKQ_DYNAMIC) {
1883 		ASSERT3S(nthreads, >=, 1);
1884 		tq->tq_maxsize = nthreads;
1885 
1886 		/* For dynamic task queues use just one backup thread */
1887 		nthreads = max_nthreads = 1;
1888 
1889 	} else if (flags & TASKQ_THREADS_CPU_PCT) {
1890 		uint_t pct;
1891 		ASSERT3S(nthreads, >=, 0);
1892 		pct = nthreads;
1893 
1894 		if (pct > taskq_cpupct_max_percent)
1895 			pct = taskq_cpupct_max_percent;
1896 
1897 		/*
1898 		 * If you're using THREADS_CPU_PCT, the process for the
1899 		 * taskq threads must be curproc.  This allows any pset
1900 		 * binding to be inherited correctly.  If proc is &p0,
1901 		 * we won't be creating LWPs, so new threads will be assigned
1902 		 * to the default processor set.
1903 		 */
1904 		ASSERT(curproc == proc || proc == &p0);
1905 		tq->tq_threads_ncpus_pct = pct;
1906 		nthreads = 1;		/* corrected in taskq_thread_create() */
1907 		max_nthreads = TASKQ_THREADS_PCT(max_ncpus, pct);
1908 
1909 	} else {
1910 		ASSERT3S(nthreads, >=, 1);
1911 		max_nthreads = nthreads;
1912 	}
1913 
1914 	if (max_nthreads < taskq_minimum_nthreads_max)
1915 		max_nthreads = taskq_minimum_nthreads_max;
1916 
1917 	/*
1918 	 * Make sure the name is 0-terminated, and conforms to the rules for
1919 	 * C indentifiers
1920 	 */
1921 	(void) strncpy(tq->tq_name, name, TASKQ_NAMELEN + 1);
1922 	strident_canon(tq->tq_name, TASKQ_NAMELEN + 1);
1923 
1924 	tq->tq_flags = flags | TASKQ_CHANGING;
1925 	tq->tq_active = 0;
1926 	tq->tq_instance = instance;
1927 	tq->tq_nthreads_target = nthreads;
1928 	tq->tq_nthreads_max = max_nthreads;
1929 	tq->tq_minalloc = minalloc;
1930 	tq->tq_maxalloc = maxalloc;
1931 	tq->tq_nbuckets = bsize;
1932 	tq->tq_proc = proc;
1933 	tq->tq_pri = pri;
1934 	tq->tq_DC = dc;
1935 	list_link_init(&tq->tq_cpupct_link);
1936 
1937 	if (max_nthreads > 1)
1938 		tq->tq_threadlist = kmem_alloc(
1939 		    sizeof (kthread_t *) * max_nthreads, KM_SLEEP);
1940 
1941 	mutex_enter(&tq->tq_lock);
1942 	if (flags & TASKQ_PREPOPULATE) {
1943 		while (minalloc-- > 0)
1944 			taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
1945 	}
1946 
1947 	/*
1948 	 * Create the first thread, which will create any other threads
1949 	 * necessary.  taskq_thread_create will not return until we have
1950 	 * enough threads to be able to process requests.
1951 	 */
1952 	taskq_thread_create(tq);
1953 	mutex_exit(&tq->tq_lock);
1954 
1955 	if (flags & TASKQ_DYNAMIC) {
1956 		taskq_bucket_t *bucket = kmem_zalloc(sizeof (taskq_bucket_t) *
1957 		    bsize, KM_SLEEP);
1958 		int b_id;
1959 
1960 		tq->tq_buckets = bucket;
1961 
1962 		/* Initialize each bucket */
1963 		for (b_id = 0; b_id < bsize; b_id++, bucket++) {
1964 			mutex_init(&bucket->tqbucket_lock, NULL, MUTEX_DEFAULT,
1965 			    NULL);
1966 			cv_init(&bucket->tqbucket_cv, NULL, CV_DEFAULT, NULL);
1967 			bucket->tqbucket_taskq = tq;
1968 			bucket->tqbucket_freelist.tqent_next =
1969 			    bucket->tqbucket_freelist.tqent_prev =
1970 			    &bucket->tqbucket_freelist;
1971 			if (flags & TASKQ_PREPOPULATE)
1972 				taskq_bucket_extend(bucket);
1973 		}
1974 	}
1975 
1976 	/*
1977 	 * Install kstats.
1978 	 * We have two cases:
1979 	 *   1) Instance is provided to taskq_create_instance(). In this case it
1980 	 *	should be >= 0 and we use it.
1981 	 *
1982 	 *   2) Instance is not provided and is automatically generated
1983 	 */
1984 	if (flags & TASKQ_NOINSTANCE) {
1985 		instance = tq->tq_instance =
1986 		    (int)(uintptr_t)vmem_alloc(taskq_id_arena, 1, VM_SLEEP);
1987 	}
1988 
1989 	if (flags & TASKQ_DYNAMIC) {
1990 		if ((tq->tq_kstat = kstat_create("unix", instance,
1991 		    tq->tq_name, "taskq_d", KSTAT_TYPE_NAMED,
1992 		    sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
1993 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
1994 			tq->tq_kstat->ks_lock = &taskq_d_kstat_lock;
1995 			tq->tq_kstat->ks_data = &taskq_d_kstat;
1996 			tq->tq_kstat->ks_update = taskq_d_kstat_update;
1997 			tq->tq_kstat->ks_private = tq;
1998 			kstat_install(tq->tq_kstat);
1999 		}
2000 	} else {
2001 		if ((tq->tq_kstat = kstat_create("unix", instance, tq->tq_name,
2002 		    "taskq", KSTAT_TYPE_NAMED,
2003 		    sizeof (taskq_kstat) / sizeof (kstat_named_t),
2004 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
2005 			tq->tq_kstat->ks_lock = &taskq_kstat_lock;
2006 			tq->tq_kstat->ks_data = &taskq_kstat;
2007 			tq->tq_kstat->ks_update = taskq_kstat_update;
2008 			tq->tq_kstat->ks_private = tq;
2009 			kstat_install(tq->tq_kstat);
2010 		}
2011 	}
2012 
2013 	return (tq);
2014 }
2015 
2016 /*
2017  * taskq_destroy().
2018  *
2019  * Assumes: by the time taskq_destroy is called no one will use this task queue
2020  * in any way and no one will try to dispatch entries in it.
2021  */
2022 void
2023 taskq_destroy(taskq_t *tq)
2024 {
2025 	taskq_bucket_t *b = tq->tq_buckets;
2026 	int bid = 0;
2027 
2028 	ASSERT(! (tq->tq_flags & TASKQ_CPR_SAFE));
2029 
2030 	/*
2031 	 * Destroy kstats.
2032 	 */
2033 	if (tq->tq_kstat != NULL) {
2034 		kstat_delete(tq->tq_kstat);
2035 		tq->tq_kstat = NULL;
2036 	}
2037 
2038 	/*
2039 	 * Destroy instance if needed.
2040 	 */
2041 	if (tq->tq_flags & TASKQ_NOINSTANCE) {
2042 		vmem_free(taskq_id_arena, (void *)(uintptr_t)(tq->tq_instance),
2043 		    1);
2044 		tq->tq_instance = 0;
2045 	}
2046 
2047 	/*
2048 	 * Unregister from the cpupct list.
2049 	 */
2050 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
2051 		taskq_cpupct_remove(tq);
2052 	}
2053 
2054 	/*
2055 	 * Wait for any pending entries to complete.
2056 	 */
2057 	taskq_wait(tq);
2058 
2059 	mutex_enter(&tq->tq_lock);
2060 	ASSERT((tq->tq_task.tqent_next == &tq->tq_task) &&
2061 	    (tq->tq_active == 0));
2062 
2063 	/* notify all the threads that they need to exit */
2064 	tq->tq_nthreads_target = 0;
2065 
2066 	tq->tq_flags |= TASKQ_CHANGING;
2067 	cv_broadcast(&tq->tq_dispatch_cv);
2068 	cv_broadcast(&tq->tq_exit_cv);
2069 
2070 	while (tq->tq_nthreads != 0)
2071 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
2072 
2073 	if (tq->tq_nthreads_max != 1)
2074 		kmem_free(tq->tq_threadlist, sizeof (kthread_t *) *
2075 		    tq->tq_nthreads_max);
2076 
2077 	tq->tq_minalloc = 0;
2078 	while (tq->tq_nalloc != 0)
2079 		taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
2080 
2081 	mutex_exit(&tq->tq_lock);
2082 
2083 	/*
2084 	 * Mark each bucket as closing and wakeup all sleeping threads.
2085 	 */
2086 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2087 		taskq_ent_t *tqe;
2088 
2089 		mutex_enter(&b->tqbucket_lock);
2090 
2091 		b->tqbucket_flags |= TQBUCKET_CLOSE;
2092 		/* Wakeup all sleeping threads */
2093 
2094 		for (tqe = b->tqbucket_freelist.tqent_next;
2095 		    tqe != &b->tqbucket_freelist; tqe = tqe->tqent_next)
2096 			cv_signal(&tqe->tqent_cv);
2097 
2098 		ASSERT(b->tqbucket_nalloc == 0);
2099 
2100 		/*
2101 		 * At this point we waited for all pending jobs to complete (in
2102 		 * both the task queue and the bucket and no new jobs should
2103 		 * arrive. Wait for all threads to die.
2104 		 */
2105 		while (b->tqbucket_nfree > 0)
2106 			cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
2107 		mutex_exit(&b->tqbucket_lock);
2108 		mutex_destroy(&b->tqbucket_lock);
2109 		cv_destroy(&b->tqbucket_cv);
2110 	}
2111 
2112 	if (tq->tq_buckets != NULL) {
2113 		ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2114 		kmem_free(tq->tq_buckets,
2115 		    sizeof (taskq_bucket_t) * tq->tq_nbuckets);
2116 
2117 		/* Cleanup fields before returning tq to the cache */
2118 		tq->tq_buckets = NULL;
2119 		tq->tq_tcreates = 0;
2120 		tq->tq_tdeaths = 0;
2121 	} else {
2122 		ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
2123 	}
2124 
2125 	tq->tq_threads_ncpus_pct = 0;
2126 	tq->tq_totaltime = 0;
2127 	tq->tq_tasks = 0;
2128 	tq->tq_maxtasks = 0;
2129 	tq->tq_executed = 0;
2130 	kmem_cache_free(taskq_cache, tq);
2131 }
2132 
2133 /*
2134  * Extend a bucket with a new entry on the free list and attach a worker thread
2135  * to it.
2136  *
2137  * Argument: pointer to the bucket.
2138  *
2139  * This function may quietly fail. It is only used by taskq_dispatch() which
2140  * handles such failures properly.
2141  */
2142 static void
2143 taskq_bucket_extend(void *arg)
2144 {
2145 	taskq_ent_t *tqe;
2146 	taskq_bucket_t *b = (taskq_bucket_t *)arg;
2147 	taskq_t *tq = b->tqbucket_taskq;
2148 	int nthreads;
2149 
2150 	if (! ENOUGH_MEMORY()) {
2151 		TQ_STAT(b, tqs_nomem);
2152 		return;
2153 	}
2154 
2155 	mutex_enter(&tq->tq_lock);
2156 
2157 	/*
2158 	 * Observe global taskq limits on the number of threads.
2159 	 */
2160 	if (tq->tq_tcreates++ - tq->tq_tdeaths > tq->tq_maxsize) {
2161 		tq->tq_tcreates--;
2162 		mutex_exit(&tq->tq_lock);
2163 		return;
2164 	}
2165 	mutex_exit(&tq->tq_lock);
2166 
2167 	tqe = kmem_cache_alloc(taskq_ent_cache, KM_NOSLEEP);
2168 
2169 	if (tqe == NULL) {
2170 		mutex_enter(&tq->tq_lock);
2171 		TQ_STAT(b, tqs_nomem);
2172 		tq->tq_tcreates--;
2173 		mutex_exit(&tq->tq_lock);
2174 		return;
2175 	}
2176 
2177 	ASSERT(tqe->tqent_thread == NULL);
2178 
2179 	tqe->tqent_un.tqent_bucket = b;
2180 
2181 	/*
2182 	 * Create a thread in a TS_STOPPED state first. If it is successfully
2183 	 * created, place the entry on the free list and start the thread.
2184 	 */
2185 	tqe->tqent_thread = thread_create(NULL, 0, taskq_d_thread, tqe,
2186 	    0, &p0, TS_STOPPED, tq->tq_pri);
2187 
2188 	/*
2189 	 * Once the entry is ready, link it to the the bucket free list.
2190 	 */
2191 	mutex_enter(&b->tqbucket_lock);
2192 	tqe->tqent_func = NULL;
2193 	TQ_APPEND(b->tqbucket_freelist, tqe);
2194 	b->tqbucket_nfree++;
2195 	TQ_STAT(b, tqs_tcreates);
2196 
2197 #if TASKQ_STATISTIC
2198 	nthreads = b->tqbucket_stat.tqs_tcreates -
2199 	    b->tqbucket_stat.tqs_tdeaths;
2200 	b->tqbucket_stat.tqs_maxthreads = MAX(nthreads,
2201 	    b->tqbucket_stat.tqs_maxthreads);
2202 #endif
2203 
2204 	mutex_exit(&b->tqbucket_lock);
2205 	/*
2206 	 * Start the stopped thread.
2207 	 */
2208 	thread_lock(tqe->tqent_thread);
2209 	tqe->tqent_thread->t_taskq = tq;
2210 	tqe->tqent_thread->t_schedflag |= TS_ALLSTART;
2211 	setrun_locked(tqe->tqent_thread);
2212 	thread_unlock(tqe->tqent_thread);
2213 }
2214 
2215 static int
2216 taskq_kstat_update(kstat_t *ksp, int rw)
2217 {
2218 	struct taskq_kstat *tqsp = &taskq_kstat;
2219 	taskq_t *tq = ksp->ks_private;
2220 
2221 	if (rw == KSTAT_WRITE)
2222 		return (EACCES);
2223 
2224 	tqsp->tq_pid.value.ui64 = tq->tq_proc->p_pid;
2225 	tqsp->tq_tasks.value.ui64 = tq->tq_tasks;
2226 	tqsp->tq_executed.value.ui64 = tq->tq_executed;
2227 	tqsp->tq_maxtasks.value.ui64 = tq->tq_maxtasks;
2228 	tqsp->tq_totaltime.value.ui64 = tq->tq_totaltime;
2229 	tqsp->tq_nactive.value.ui64 = tq->tq_active;
2230 	tqsp->tq_nalloc.value.ui64 = tq->tq_nalloc;
2231 	tqsp->tq_pri.value.ui64 = tq->tq_pri;
2232 	tqsp->tq_nthreads.value.ui64 = tq->tq_nthreads;
2233 	return (0);
2234 }
2235 
2236 static int
2237 taskq_d_kstat_update(kstat_t *ksp, int rw)
2238 {
2239 	struct taskq_d_kstat *tqsp = &taskq_d_kstat;
2240 	taskq_t *tq = ksp->ks_private;
2241 	taskq_bucket_t *b = tq->tq_buckets;
2242 	int bid = 0;
2243 
2244 	if (rw == KSTAT_WRITE)
2245 		return (EACCES);
2246 
2247 	ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2248 
2249 	tqsp->tqd_btasks.value.ui64 = tq->tq_tasks;
2250 	tqsp->tqd_bexecuted.value.ui64 = tq->tq_executed;
2251 	tqsp->tqd_bmaxtasks.value.ui64 = tq->tq_maxtasks;
2252 	tqsp->tqd_bnalloc.value.ui64 = tq->tq_nalloc;
2253 	tqsp->tqd_bnactive.value.ui64 = tq->tq_active;
2254 	tqsp->tqd_btotaltime.value.ui64 = tq->tq_totaltime;
2255 	tqsp->tqd_pri.value.ui64 = tq->tq_pri;
2256 
2257 	tqsp->tqd_hits.value.ui64 = 0;
2258 	tqsp->tqd_misses.value.ui64 = 0;
2259 	tqsp->tqd_overflows.value.ui64 = 0;
2260 	tqsp->tqd_tcreates.value.ui64 = 0;
2261 	tqsp->tqd_tdeaths.value.ui64 = 0;
2262 	tqsp->tqd_maxthreads.value.ui64 = 0;
2263 	tqsp->tqd_nomem.value.ui64 = 0;
2264 	tqsp->tqd_disptcreates.value.ui64 = 0;
2265 	tqsp->tqd_totaltime.value.ui64 = 0;
2266 	tqsp->tqd_nalloc.value.ui64 = 0;
2267 	tqsp->tqd_nfree.value.ui64 = 0;
2268 
2269 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2270 		tqsp->tqd_hits.value.ui64 += b->tqbucket_stat.tqs_hits;
2271 		tqsp->tqd_misses.value.ui64 += b->tqbucket_stat.tqs_misses;
2272 		tqsp->tqd_overflows.value.ui64 += b->tqbucket_stat.tqs_overflow;
2273 		tqsp->tqd_tcreates.value.ui64 += b->tqbucket_stat.tqs_tcreates;
2274 		tqsp->tqd_tdeaths.value.ui64 += b->tqbucket_stat.tqs_tdeaths;
2275 		tqsp->tqd_maxthreads.value.ui64 +=
2276 		    b->tqbucket_stat.tqs_maxthreads;
2277 		tqsp->tqd_nomem.value.ui64 += b->tqbucket_stat.tqs_nomem;
2278 		tqsp->tqd_disptcreates.value.ui64 +=
2279 		    b->tqbucket_stat.tqs_disptcreates;
2280 		tqsp->tqd_totaltime.value.ui64 += b->tqbucket_totaltime;
2281 		tqsp->tqd_nalloc.value.ui64 += b->tqbucket_nalloc;
2282 		tqsp->tqd_nfree.value.ui64 += b->tqbucket_nfree;
2283 	}
2284 	return (0);
2285 }
2286