xref: /illumos-gate/usr/src/uts/common/os/taskq.c (revision 35a5a3587fd94b666239c157d3722745250ccbd7)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Kernel task queues: general-purpose asynchronous task scheduling.
28  *
29  * A common problem in kernel programming is the need to schedule tasks
30  * to be performed later, by another thread. There are several reasons
31  * you may want or need to do this:
32  *
33  * (1) The task isn't time-critical, but your current code path is.
34  *
35  * (2) The task may require grabbing locks that you already hold.
36  *
37  * (3) The task may need to block (e.g. to wait for memory), but you
38  *     cannot block in your current context.
39  *
40  * (4) Your code path can't complete because of some condition, but you can't
41  *     sleep or fail, so you queue the task for later execution when condition
42  *     disappears.
43  *
44  * (5) You just want a simple way to launch multiple tasks in parallel.
45  *
46  * Task queues provide such a facility. In its simplest form (used when
47  * performance is not a critical consideration) a task queue consists of a
48  * single list of tasks, together with one or more threads to service the
49  * list. There are some cases when this simple queue is not sufficient:
50  *
51  * (1) The task queues are very hot and there is a need to avoid data and lock
52  *	contention over global resources.
53  *
54  * (2) Some tasks may depend on other tasks to complete, so they can't be put in
55  *	the same list managed by the same thread.
56  *
57  * (3) Some tasks may block for a long time, and this should not block other
58  *	tasks in the queue.
59  *
60  * To provide useful service in such cases we define a "dynamic task queue"
61  * which has an individual thread for each of the tasks. These threads are
62  * dynamically created as they are needed and destroyed when they are not in
63  * use. The API for managing task pools is the same as for managing task queues
64  * with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
65  * dynamic task pool behavior is desired.
66  *
67  * Dynamic task queues may also place tasks in the normal queue (called "backing
68  * queue") when task pool runs out of resources. Users of task queues may
69  * disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
70  * flags.
71  *
72  * The backing task queue is also used for scheduling internal tasks needed for
73  * dynamic task queue maintenance.
74  *
75  * INTERFACES ==================================================================
76  *
77  * taskq_t *taskq_create(name, nthreads, pri, minalloc, maxall, flags);
78  *
79  *	Create a taskq with specified properties.
80  *	Possible 'flags':
81  *
82  *	  TASKQ_DYNAMIC: Create task pool for task management. If this flag is
83  *		specified, 'nthreads' specifies the maximum number of threads in
84  *		the task queue. Task execution order for dynamic task queues is
85  *		not predictable.
86  *
87  *		If this flag is not specified (default case) a
88  *		single-list task queue is created with 'nthreads' threads
89  *		servicing it. Entries in this queue are managed by
90  *		taskq_ent_alloc() and taskq_ent_free() which try to keep the
91  *		task population between 'minalloc' and 'maxalloc', but the
92  *		latter limit is only advisory for TQ_SLEEP dispatches and the
93  *		former limit is only advisory for TQ_NOALLOC dispatches. If
94  *		TASKQ_PREPOPULATE is set in 'flags', the taskq will be
95  *		prepopulated with 'minalloc' task structures.
96  *
97  *		Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
98  *		executed in the order they are scheduled if nthreads == 1.
99  *		If nthreads > 1, task execution order is not predictable.
100  *
101  *	  TASKQ_PREPOPULATE: Prepopulate task queue with threads.
102  *		Also prepopulate the task queue with 'minalloc' task structures.
103  *
104  *	  TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
105  *		interpreted as a percentage of the # of online CPUs on the
106  *		system.  The taskq subsystem will automatically adjust the
107  *		number of threads in the taskq in response to CPU online
108  *		and offline events, to keep the ratio.  nthreads must be in
109  *		the range [0,100].
110  *
111  *		The calculation used is:
112  *
113  *			MAX((ncpus_online * percentage)/100, 1)
114  *
115  *		This flag is not supported for DYNAMIC task queues.
116  *		This flag is not compatible with TASKQ_CPR_SAFE.
117  *
118  *	  TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
119  *		use their own protocol for handling CPR issues. This flag is not
120  *		supported for DYNAMIC task queues.  This flag is not compatible
121  *		with TASKQ_THREADS_CPU_PCT.
122  *
123  *	The 'pri' field specifies the default priority for the threads that
124  *	service all scheduled tasks.
125  *
126  * taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
127  *    maxall, flags);
128  *
129  *	Like taskq_create(), but takes an instance number (or -1 to indicate
130  *	no instance).
131  *
132  * taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxall, proc,
133  *    flags);
134  *
135  *	Like taskq_create(), but creates the taskq threads in the specified
136  *	system process.  If proc != &p0, this must be called from a thread
137  *	in that process.
138  *
139  * taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxall, proc,
140  *    dc, flags);
141  *
142  *	Like taskq_create_proc(), but the taskq threads will use the
143  *	System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
144  *
145  * void taskq_destroy(tap):
146  *
147  *	Waits for any scheduled tasks to complete, then destroys the taskq.
148  *	Caller should guarantee that no new tasks are scheduled in the closing
149  *	taskq.
150  *
151  * taskqid_t taskq_dispatch(tq, func, arg, flags):
152  *
153  *	Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
154  *	the caller is willing to block for memory.  The function returns an
155  *	opaque value which is zero iff dispatch fails.  If flags is TQ_NOSLEEP
156  *	or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
157  *	and returns (taskqid_t)0.
158  *
159  *	ASSUMES: func != NULL.
160  *
161  *	Possible flags:
162  *	  TQ_NOSLEEP: Do not wait for resources; may fail.
163  *
164  *	  TQ_NOALLOC: Do not allocate memory; may fail.  May only be used with
165  *		non-dynamic task queues.
166  *
167  *	  TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
168  *		lack of available resources and fail. If this flag is not
169  *		set, and the task pool is exhausted, the task may be scheduled
170  *		in the backing queue. This flag may ONLY be used with dynamic
171  *		task queues.
172  *
173  *		NOTE: This flag should always be used when a task queue is used
174  *		for tasks that may depend on each other for completion.
175  *		Enqueueing dependent tasks may create deadlocks.
176  *
177  *	  TQ_SLEEP:   May block waiting for resources. May still fail for
178  *		dynamic task queues if TQ_NOQUEUE is also specified, otherwise
179  *		always succeed.
180  *
181  *	  TQ_FRONT:   Puts the new task at the front of the queue.  Be careful.
182  *
183  *	NOTE: Dynamic task queues are much more likely to fail in
184  *		taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
185  *		is important to have backup strategies handling such failures.
186  *
187  * void taskq_wait(tq):
188  *
189  *	Waits for all previously scheduled tasks to complete.
190  *
191  *	NOTE: It does not stop any new task dispatches.
192  *	      Do NOT call taskq_wait() from a task: it will cause deadlock.
193  *
194  * void taskq_suspend(tq)
195  *
196  *	Suspend all task execution. Tasks already scheduled for a dynamic task
197  *	queue will still be executed, but all new scheduled tasks will be
198  *	suspended until taskq_resume() is called.
199  *
200  * int  taskq_suspended(tq)
201  *
202  *	Returns 1 if taskq is suspended and 0 otherwise. It is intended to
203  *	ASSERT that the task queue is suspended.
204  *
205  * void taskq_resume(tq)
206  *
207  *	Resume task queue execution.
208  *
209  * int  taskq_member(tq, thread)
210  *
211  *	Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
212  *	intended use is to ASSERT that a given function is called in taskq
213  *	context only.
214  *
215  * system_taskq
216  *
217  *	Global system-wide dynamic task queue for common uses. It may be used by
218  *	any subsystem that needs to schedule tasks and does not need to manage
219  *	its own task queues. It is initialized quite early during system boot.
220  *
221  * IMPLEMENTATION ==============================================================
222  *
223  * This is schematic representation of the task queue structures.
224  *
225  *   taskq:
226  *   +-------------+
227  *   | tq_lock     | +---< taskq_ent_free()
228  *   +-------------+ |
229  *   |...          | | tqent:                  tqent:
230  *   +-------------+ | +------------+          +------------+
231  *   | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
232  *   +-------------+   +------------+          +------------+
233  *   |...          |   | ...        |          | ...        |
234  *   +-------------+   +------------+          +------------+
235  *   | tq_task     |    |
236  *   |             |    +-------------->taskq_ent_alloc()
237  * +--------------------------------------------------------------------------+
238  * | |                     |            tqent                   tqent         |
239  * | +---------------------+     +--> +------------+     +--> +------------+  |
240  * | | ...		   |     |    | func, arg  |     |    | func, arg  |  |
241  * +>+---------------------+ <---|-+  +------------+ <---|-+  +------------+  |
242  *   | tq_taskq.tqent_next | ----+ |  | tqent_next | --->+ |  | tqent_next |--+
243  *   +---------------------+	   |  +------------+     ^ |  +------------+
244  * +-| tq_task.tqent_prev  |	   +--| tqent_prev |     | +--| tqent_prev |  ^
245  * | +---------------------+	      +------------+     |    +------------+  |
246  * | |...		   |	      | ...        |     |    | ...        |  |
247  * | +---------------------+	      +------------+     |    +------------+  |
248  * |                                      ^              |                    |
249  * |                                      |              |                    |
250  * +--------------------------------------+--------------+       TQ_APPEND() -+
251  *   |             |                      |
252  *   |...          |   taskq_thread()-----+
253  *   +-------------+
254  *   | tq_buckets  |--+-------> [ NULL ] (for regular task queues)
255  *   +-------------+  |
256  *                    |   DYNAMIC TASK QUEUES:
257  *                    |
258  *                    +-> taskq_bucket[nCPU]		taskq_bucket_dispatch()
259  *                        +-------------------+                    ^
260  *                   +--->| tqbucket_lock     |                    |
261  *                   |    +-------------------+   +--------+      +--------+
262  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  | ^
263  *                   |    +-------------------+<--+--------+<--...+--------+ |
264  *                   |    | ...               |   | thread |      | thread | |
265  *                   |    +-------------------+   +--------+      +--------+ |
266  *                   |    +-------------------+                              |
267  * taskq_dispatch()--+--->| tqbucket_lock     |             TQ_APPEND()------+
268  *      TQ_HASH()    |    +-------------------+   +--------+      +--------+
269  *                   |    | tqbucket_freelist |-->| tqent  |-->...| tqent  |
270  *                   |    +-------------------+<--+--------+<--...+--------+
271  *                   |    | ...               |   | thread |      | thread |
272  *                   |    +-------------------+   +--------+      +--------+
273  *		     +--->	...
274  *
275  *
276  * Task queues use tq_task field to link new entry in the queue. The queue is a
277  * circular doubly-linked list. Entries are put in the end of the list with
278  * TQ_APPEND() and processed from the front of the list by taskq_thread() in
279  * FIFO order. Task queue entries are cached in the free list managed by
280  * taskq_ent_alloc() and taskq_ent_free() functions.
281  *
282  *	All threads used by task queues mark t_taskq field of the thread to
283  *	point to the task queue.
284  *
285  * Taskq Thread Management -----------------------------------------------------
286  *
287  * Taskq's non-dynamic threads are managed with several variables and flags:
288  *
289  *	* tq_nthreads	- The number of threads in taskq_thread() for the
290  *			  taskq.
291  *
292  *	* tq_active	- The number of threads not waiting on a CV in
293  *			  taskq_thread(); includes newly created threads
294  *			  not yet counted in tq_nthreads.
295  *
296  *	* tq_nthreads_target
297  *			- The number of threads desired for the taskq.
298  *
299  *	* tq_flags & TASKQ_CHANGING
300  *			- Indicates that tq_nthreads != tq_nthreads_target.
301  *
302  *	* tq_flags & TASKQ_THREAD_CREATED
303  *			- Indicates that a thread is being created in the taskq.
304  *
305  * During creation, tq_nthreads and tq_active are set to 0, and
306  * tq_nthreads_target is set to the number of threads desired.  The
307  * TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
308  * create the first thread. taskq_thread_create() increments tq_active,
309  * sets TASKQ_THREAD_CREATED, and creates the new thread.
310  *
311  * Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
312  * flag, and increments tq_nthreads.  It stores the new value of
313  * tq_nthreads as its "thread_id", and stores its thread pointer in the
314  * tq_threadlist at the (thread_id - 1).  We keep the thread_id space
315  * densely packed by requiring that only the largest thread_id can exit during
316  * normal adjustment.   The exception is during the destruction of the
317  * taskq; once tq_nthreads_target is set to zero, no new threads will be created
318  * for the taskq queue, so every thread can exit without any ordering being
319  * necessary.
320  *
321  * Threads will only process work if their thread id is <= tq_nthreads_target.
322  *
323  * When TASKQ_CHANGING is set, threads will check the current thread target
324  * whenever they wake up, and do whatever they can to apply its effects.
325  *
326  * TASKQ_THREAD_CPU_PCT --------------------------------------------------------
327  *
328  * When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
329  * percentage in tq_threads_ncpus_pct, start them off with the correct thread
330  * target, and add them to the taskq_cpupct_list for later adjustment.
331  *
332  * We register taskq_cpu_setup() to be called whenever a CPU changes state.  It
333  * walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
334  * if need be, and wakes up all of the threads to process the change.
335  *
336  * Dynamic Task Queues Implementation ------------------------------------------
337  *
338  * For a dynamic task queues there is a 1-to-1 mapping between a thread and
339  * taskq_ent_structure. Each entry is serviced by its own thread and each thread
340  * is controlled by a single entry.
341  *
342  * Entries are distributed over a set of buckets. To avoid using modulo
343  * arithmetics the number of buckets is 2^n and is determined as the nearest
344  * power of two roundown of the number of CPUs in the system. Tunable
345  * variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
346  * is attached to a bucket for its lifetime and can't migrate to other buckets.
347  *
348  * Entries that have scheduled tasks are not placed in any list. The dispatch
349  * function sets their "func" and "arg" fields and signals the corresponding
350  * thread to execute the task. Once the thread executes the task it clears the
351  * "func" field and places an entry on the bucket cache of free entries pointed
352  * by "tqbucket_freelist" field. ALL entries on the free list should have "func"
353  * field equal to NULL. The free list is a circular doubly-linked list identical
354  * in structure to the tq_task list above, but entries are taken from it in LIFO
355  * order - the last freed entry is the first to be allocated. The
356  * taskq_bucket_dispatch() function gets the most recently used entry from the
357  * free list, sets its "func" and "arg" fields and signals a worker thread.
358  *
359  * After executing each task a per-entry thread taskq_d_thread() places its
360  * entry on the bucket free list and goes to a timed sleep. If it wakes up
361  * without getting new task it removes the entry from the free list and destroys
362  * itself. The thread sleep time is controlled by a tunable variable
363  * `taskq_thread_timeout'.
364  *
365  * There are various statistics kept in the bucket which allows for later
366  * analysis of taskq usage patterns. Also, a global copy of taskq creation and
367  * death statistics is kept in the global taskq data structure. Since thread
368  * creation and death happen rarely, updating such global data does not present
369  * a performance problem.
370  *
371  * NOTE: Threads are not bound to any CPU and there is absolutely no association
372  *       between the bucket and actual thread CPU, so buckets are used only to
373  *	 split resources and reduce resource contention. Having threads attached
374  *	 to the CPU denoted by a bucket may reduce number of times the job
375  *	 switches between CPUs.
376  *
377  *	 Current algorithm creates a thread whenever a bucket has no free
378  *	 entries. It would be nice to know how many threads are in the running
379  *	 state and don't create threads if all CPUs are busy with existing
380  *	 tasks, but it is unclear how such strategy can be implemented.
381  *
382  *	 Currently buckets are created statically as an array attached to task
383  *	 queue. On some system with nCPUs < max_ncpus it may waste system
384  *	 memory. One solution may be allocation of buckets when they are first
385  *	 touched, but it is not clear how useful it is.
386  *
387  * SUSPEND/RESUME implementation -----------------------------------------------
388  *
389  *	Before executing a task taskq_thread() (executing non-dynamic task
390  *	queues) obtains taskq's thread lock as a reader. The taskq_suspend()
391  *	function gets the same lock as a writer blocking all non-dynamic task
392  *	execution. The taskq_resume() function releases the lock allowing
393  *	taskq_thread to continue execution.
394  *
395  *	For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
396  *	taskq_suspend() function. After that taskq_bucket_dispatch() always
397  *	fails, so that taskq_dispatch() will either enqueue tasks for a
398  *	suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
399  *	flags.
400  *
401  *	NOTE: taskq_suspend() does not immediately block any tasks already
402  *	      scheduled for dynamic task queues. It only suspends new tasks
403  *	      scheduled after taskq_suspend() was called.
404  *
405  *	taskq_member() function works by comparing a thread t_taskq pointer with
406  *	the passed thread pointer.
407  *
408  * LOCKS and LOCK Hierarchy ----------------------------------------------------
409  *
410  *   There are three locks used in task queues:
411  *
412  *   1) The taskq_t's tq_lock, protecting global task queue state.
413  *
414  *   2) Each per-CPU bucket has a lock for bucket management.
415  *
416  *   3) The global taskq_cpupct_lock, which protects the list of
417  *      TASKQ_THREADS_CPU_PCT taskqs.
418  *
419  *   If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
420  *   lock.
421  *
422  *   If both (1) and (3) are needed, tq_lock should be taken *after*
423  *   taskq_cpupct_lock.
424  *
425  * DEBUG FACILITIES ------------------------------------------------------------
426  *
427  * For DEBUG kernels it is possible to induce random failures to
428  * taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
429  * taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
430  * failures for dynamic and static task queues respectively.
431  *
432  * Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
433  *
434  * TUNABLES --------------------------------------------------------------------
435  *
436  *	system_taskq_size	- Size of the global system_taskq.
437  *				  This value is multiplied by nCPUs to determine
438  *				  actual size.
439  *				  Default value: 64
440  *
441  *	taskq_minimum_nthreads_max
442  *				- Minimum size of the thread list for a taskq.
443  *				  Useful for testing different thread pool
444  *				  sizes by overwriting tq_nthreads_target.
445  *
446  *	taskq_thread_timeout	- Maximum idle time for taskq_d_thread()
447  *				  Default value: 5 minutes
448  *
449  *	taskq_maxbuckets	- Maximum number of buckets in any task queue
450  *				  Default value: 128
451  *
452  *	taskq_search_depth	- Maximum # of buckets searched for a free entry
453  *				  Default value: 4
454  *
455  *	taskq_dmtbf		- Mean time between induced dispatch failures
456  *				  for dynamic task queues.
457  *				  Default value: UINT_MAX (no induced failures)
458  *
459  *	taskq_smtbf		- Mean time between induced dispatch failures
460  *				  for static task queues.
461  *				  Default value: UINT_MAX (no induced failures)
462  *
463  * CONDITIONAL compilation -----------------------------------------------------
464  *
465  *    TASKQ_STATISTIC	- If set will enable bucket statistic (default).
466  *
467  */
468 
469 #include <sys/taskq_impl.h>
470 #include <sys/thread.h>
471 #include <sys/proc.h>
472 #include <sys/kmem.h>
473 #include <sys/vmem.h>
474 #include <sys/callb.h>
475 #include <sys/class.h>
476 #include <sys/systm.h>
477 #include <sys/cmn_err.h>
478 #include <sys/debug.h>
479 #include <sys/vmsystm.h>	/* For throttlefree */
480 #include <sys/sysmacros.h>
481 #include <sys/cpuvar.h>
482 #include <sys/cpupart.h>
483 #include <sys/sdt.h>
484 #include <sys/sysdc.h>
485 #include <sys/note.h>
486 
487 static kmem_cache_t *taskq_ent_cache, *taskq_cache;
488 
489 /*
490  * Pseudo instance numbers for taskqs without explicitly provided instance.
491  */
492 static vmem_t *taskq_id_arena;
493 
494 /* Global system task queue for common use */
495 taskq_t	*system_taskq;
496 
497 /*
498  * Maximum number of entries in global system taskq is
499  *	system_taskq_size * max_ncpus
500  */
501 #define	SYSTEM_TASKQ_SIZE 64
502 int system_taskq_size = SYSTEM_TASKQ_SIZE;
503 
504 /*
505  * Minimum size for tq_nthreads_max; useful for those who want to play around
506  * with increasing a taskq's tq_nthreads_target.
507  */
508 int taskq_minimum_nthreads_max = 1;
509 
510 /*
511  * We want to ensure that when taskq_create() returns, there is at least
512  * one thread ready to handle requests.  To guarantee this, we have to wait
513  * for the second thread, since the first one cannot process requests until
514  * the second thread has been created.
515  */
516 #define	TASKQ_CREATE_ACTIVE_THREADS	2
517 
518 /* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
519 #define	TASKQ_CPUPCT_MAX_PERCENT	1000
520 int taskq_cpupct_max_percent = TASKQ_CPUPCT_MAX_PERCENT;
521 
522 /*
523  * Dynamic task queue threads that don't get any work within
524  * taskq_thread_timeout destroy themselves
525  */
526 #define	TASKQ_THREAD_TIMEOUT (60 * 5)
527 int taskq_thread_timeout = TASKQ_THREAD_TIMEOUT;
528 
529 #define	TASKQ_MAXBUCKETS 128
530 int taskq_maxbuckets = TASKQ_MAXBUCKETS;
531 
532 /*
533  * When a bucket has no available entries another buckets are tried.
534  * taskq_search_depth parameter limits the amount of buckets that we search
535  * before failing. This is mostly useful in systems with many CPUs where we may
536  * spend too much time scanning busy buckets.
537  */
538 #define	TASKQ_SEARCH_DEPTH 4
539 int taskq_search_depth = TASKQ_SEARCH_DEPTH;
540 
541 /*
542  * Hashing function: mix various bits of x. May be pretty much anything.
543  */
544 #define	TQ_HASH(x) ((x) ^ ((x) >> 11) ^ ((x) >> 17) ^ ((x) ^ 27))
545 
546 /*
547  * We do not create any new threads when the system is low on memory and start
548  * throttling memory allocations. The following macro tries to estimate such
549  * condition.
550  */
551 #define	ENOUGH_MEMORY() (freemem > throttlefree)
552 
553 /*
554  * Static functions.
555  */
556 static taskq_t	*taskq_create_common(const char *, int, int, pri_t, int,
557     int, proc_t *, uint_t, uint_t);
558 static void taskq_thread(void *);
559 static void taskq_d_thread(taskq_ent_t *);
560 static void taskq_bucket_extend(void *);
561 static int  taskq_constructor(void *, void *, int);
562 static void taskq_destructor(void *, void *);
563 static int  taskq_ent_constructor(void *, void *, int);
564 static void taskq_ent_destructor(void *, void *);
565 static taskq_ent_t *taskq_ent_alloc(taskq_t *, int);
566 static void taskq_ent_free(taskq_t *, taskq_ent_t *);
567 static taskq_ent_t *taskq_bucket_dispatch(taskq_bucket_t *, task_func_t,
568     void *);
569 
570 /*
571  * Task queues kstats.
572  */
573 struct taskq_kstat {
574 	kstat_named_t	tq_pid;
575 	kstat_named_t	tq_tasks;
576 	kstat_named_t	tq_executed;
577 	kstat_named_t	tq_maxtasks;
578 	kstat_named_t	tq_totaltime;
579 	kstat_named_t	tq_nalloc;
580 	kstat_named_t	tq_nactive;
581 	kstat_named_t	tq_pri;
582 	kstat_named_t	tq_nthreads;
583 } taskq_kstat = {
584 	{ "pid",		KSTAT_DATA_UINT64 },
585 	{ "tasks",		KSTAT_DATA_UINT64 },
586 	{ "executed",		KSTAT_DATA_UINT64 },
587 	{ "maxtasks",		KSTAT_DATA_UINT64 },
588 	{ "totaltime",		KSTAT_DATA_UINT64 },
589 	{ "nactive",		KSTAT_DATA_UINT64 },
590 	{ "nalloc",		KSTAT_DATA_UINT64 },
591 	{ "priority",		KSTAT_DATA_UINT64 },
592 	{ "threads",		KSTAT_DATA_UINT64 },
593 };
594 
595 struct taskq_d_kstat {
596 	kstat_named_t	tqd_pri;
597 	kstat_named_t	tqd_btasks;
598 	kstat_named_t	tqd_bexecuted;
599 	kstat_named_t	tqd_bmaxtasks;
600 	kstat_named_t	tqd_bnalloc;
601 	kstat_named_t	tqd_bnactive;
602 	kstat_named_t	tqd_btotaltime;
603 	kstat_named_t	tqd_hits;
604 	kstat_named_t	tqd_misses;
605 	kstat_named_t	tqd_overflows;
606 	kstat_named_t	tqd_tcreates;
607 	kstat_named_t	tqd_tdeaths;
608 	kstat_named_t	tqd_maxthreads;
609 	kstat_named_t	tqd_nomem;
610 	kstat_named_t	tqd_disptcreates;
611 	kstat_named_t	tqd_totaltime;
612 	kstat_named_t	tqd_nalloc;
613 	kstat_named_t	tqd_nfree;
614 } taskq_d_kstat = {
615 	{ "priority",		KSTAT_DATA_UINT64 },
616 	{ "btasks",		KSTAT_DATA_UINT64 },
617 	{ "bexecuted",		KSTAT_DATA_UINT64 },
618 	{ "bmaxtasks",		KSTAT_DATA_UINT64 },
619 	{ "bnalloc",		KSTAT_DATA_UINT64 },
620 	{ "bnactive",		KSTAT_DATA_UINT64 },
621 	{ "btotaltime",		KSTAT_DATA_UINT64 },
622 	{ "hits",		KSTAT_DATA_UINT64 },
623 	{ "misses",		KSTAT_DATA_UINT64 },
624 	{ "overflows",		KSTAT_DATA_UINT64 },
625 	{ "tcreates",		KSTAT_DATA_UINT64 },
626 	{ "tdeaths",		KSTAT_DATA_UINT64 },
627 	{ "maxthreads",		KSTAT_DATA_UINT64 },
628 	{ "nomem",		KSTAT_DATA_UINT64 },
629 	{ "disptcreates",	KSTAT_DATA_UINT64 },
630 	{ "totaltime",		KSTAT_DATA_UINT64 },
631 	{ "nalloc",		KSTAT_DATA_UINT64 },
632 	{ "nfree",		KSTAT_DATA_UINT64 },
633 };
634 
635 static kmutex_t taskq_kstat_lock;
636 static kmutex_t taskq_d_kstat_lock;
637 static int taskq_kstat_update(kstat_t *, int);
638 static int taskq_d_kstat_update(kstat_t *, int);
639 
640 /*
641  * List of all TASKQ_THREADS_CPU_PCT taskqs.
642  */
643 static list_t taskq_cpupct_list;	/* protected by cpu_lock */
644 
645 /*
646  * Collect per-bucket statistic when TASKQ_STATISTIC is defined.
647  */
648 #define	TASKQ_STATISTIC 1
649 
650 #if TASKQ_STATISTIC
651 #define	TQ_STAT(b, x)	b->tqbucket_stat.x++
652 #else
653 #define	TQ_STAT(b, x)
654 #endif
655 
656 /*
657  * Random fault injection.
658  */
659 uint_t taskq_random;
660 uint_t taskq_dmtbf = UINT_MAX;    /* mean time between injected failures */
661 uint_t taskq_smtbf = UINT_MAX;    /* mean time between injected failures */
662 
663 /*
664  * TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
665  *
666  * TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
667  * they could prepopulate the cache and make sure that they do not use more
668  * then minalloc entries.  So, fault injection in this case insures that
669  * either TASKQ_PREPOPULATE is not set or there are more entries allocated
670  * than is specified by minalloc.  TQ_NOALLOC dispatches are always allowed
671  * to fail, but for simplicity we treat them identically to TQ_NOSLEEP
672  * dispatches.
673  */
674 #ifdef DEBUG
675 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)		\
676 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
677 	if ((flag & TQ_NOSLEEP) &&				\
678 	    taskq_random < 1771875 / taskq_dmtbf) {		\
679 		return (NULL);					\
680 	}
681 
682 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)		\
683 	taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
684 	if ((flag & (TQ_NOSLEEP | TQ_NOALLOC)) &&		\
685 	    (!(tq->tq_flags & TASKQ_PREPOPULATE) ||		\
686 	    (tq->tq_nalloc > tq->tq_minalloc)) &&		\
687 	    (taskq_random < (1771875 / taskq_smtbf))) {		\
688 		mutex_exit(&tq->tq_lock);			\
689 		return (NULL);					\
690 	}
691 #else
692 #define	TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)
693 #define	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)
694 #endif
695 
696 #define	IS_EMPTY(l) (((l).tqent_prev == (l).tqent_next) &&	\
697 	((l).tqent_prev == &(l)))
698 
699 /*
700  * Append `tqe' in the end of the doubly-linked list denoted by l.
701  */
702 #define	TQ_APPEND(l, tqe) {					\
703 	tqe->tqent_next = &l;					\
704 	tqe->tqent_prev = l.tqent_prev;				\
705 	tqe->tqent_next->tqent_prev = tqe;			\
706 	tqe->tqent_prev->tqent_next = tqe;			\
707 }
708 /*
709  * Prepend 'tqe' to the beginning of l
710  */
711 #define	TQ_PREPEND(l, tqe) {					\
712 	tqe->tqent_next = l.tqent_next;				\
713 	tqe->tqent_prev = &l;					\
714 	tqe->tqent_next->tqent_prev = tqe;			\
715 	tqe->tqent_prev->tqent_next = tqe;			\
716 }
717 
718 /*
719  * Schedule a task specified by func and arg into the task queue entry tqe.
720  */
721 #define	TQ_DO_ENQUEUE(tq, tqe, func, arg, front) {			\
722 	ASSERT(MUTEX_HELD(&tq->tq_lock));				\
723 	_NOTE(CONSTCOND)						\
724 	if (front) {							\
725 		TQ_PREPEND(tq->tq_task, tqe);				\
726 	} else {							\
727 		TQ_APPEND(tq->tq_task, tqe);				\
728 	}								\
729 	tqe->tqent_func = (func);					\
730 	tqe->tqent_arg = (arg);						\
731 	tq->tq_tasks++;							\
732 	if (tq->tq_tasks - tq->tq_executed > tq->tq_maxtasks)		\
733 		tq->tq_maxtasks = tq->tq_tasks - tq->tq_executed;	\
734 	cv_signal(&tq->tq_dispatch_cv);					\
735 	DTRACE_PROBE2(taskq__enqueue, taskq_t *, tq, taskq_ent_t *, tqe); \
736 }
737 
738 #define	TQ_ENQUEUE(tq, tqe, func, arg)					\
739 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 0)
740 
741 #define	TQ_ENQUEUE_FRONT(tq, tqe, func, arg)				\
742 	TQ_DO_ENQUEUE(tq, tqe, func, arg, 1)
743 
744 /*
745  * Do-nothing task which may be used to prepopulate thread caches.
746  */
747 /*ARGSUSED*/
748 void
749 nulltask(void *unused)
750 {
751 }
752 
753 /*ARGSUSED*/
754 static int
755 taskq_constructor(void *buf, void *cdrarg, int kmflags)
756 {
757 	taskq_t *tq = buf;
758 
759 	bzero(tq, sizeof (taskq_t));
760 
761 	mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
762 	rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
763 	cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
764 	cv_init(&tq->tq_exit_cv, NULL, CV_DEFAULT, NULL);
765 	cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
766 
767 	tq->tq_task.tqent_next = &tq->tq_task;
768 	tq->tq_task.tqent_prev = &tq->tq_task;
769 
770 	return (0);
771 }
772 
773 /*ARGSUSED*/
774 static void
775 taskq_destructor(void *buf, void *cdrarg)
776 {
777 	taskq_t *tq = buf;
778 
779 	ASSERT(tq->tq_nthreads == 0);
780 	ASSERT(tq->tq_buckets == NULL);
781 	ASSERT(tq->tq_tcreates == 0);
782 	ASSERT(tq->tq_tdeaths == 0);
783 
784 	mutex_destroy(&tq->tq_lock);
785 	rw_destroy(&tq->tq_threadlock);
786 	cv_destroy(&tq->tq_dispatch_cv);
787 	cv_destroy(&tq->tq_exit_cv);
788 	cv_destroy(&tq->tq_wait_cv);
789 }
790 
791 /*ARGSUSED*/
792 static int
793 taskq_ent_constructor(void *buf, void *cdrarg, int kmflags)
794 {
795 	taskq_ent_t *tqe = buf;
796 
797 	tqe->tqent_thread = NULL;
798 	cv_init(&tqe->tqent_cv, NULL, CV_DEFAULT, NULL);
799 
800 	return (0);
801 }
802 
803 /*ARGSUSED*/
804 static void
805 taskq_ent_destructor(void *buf, void *cdrarg)
806 {
807 	taskq_ent_t *tqe = buf;
808 
809 	ASSERT(tqe->tqent_thread == NULL);
810 	cv_destroy(&tqe->tqent_cv);
811 }
812 
813 void
814 taskq_init(void)
815 {
816 	taskq_ent_cache = kmem_cache_create("taskq_ent_cache",
817 	    sizeof (taskq_ent_t), 0, taskq_ent_constructor,
818 	    taskq_ent_destructor, NULL, NULL, NULL, 0);
819 	taskq_cache = kmem_cache_create("taskq_cache", sizeof (taskq_t),
820 	    0, taskq_constructor, taskq_destructor, NULL, NULL, NULL, 0);
821 	taskq_id_arena = vmem_create("taskq_id_arena",
822 	    (void *)1, INT32_MAX, 1, NULL, NULL, NULL, 0,
823 	    VM_SLEEP | VMC_IDENTIFIER);
824 
825 	list_create(&taskq_cpupct_list, sizeof (taskq_t),
826 	    offsetof(taskq_t, tq_cpupct_link));
827 }
828 
829 static void
830 taskq_update_nthreads(taskq_t *tq, uint_t ncpus)
831 {
832 	uint_t newtarget = TASKQ_THREADS_PCT(ncpus, tq->tq_threads_ncpus_pct);
833 
834 	ASSERT(MUTEX_HELD(&cpu_lock));
835 	ASSERT(MUTEX_HELD(&tq->tq_lock));
836 
837 	/* We must be going from non-zero to non-zero; no exiting. */
838 	ASSERT3U(tq->tq_nthreads_target, !=, 0);
839 	ASSERT3U(newtarget, !=, 0);
840 
841 	ASSERT3U(newtarget, <=, tq->tq_nthreads_max);
842 	if (newtarget != tq->tq_nthreads_target) {
843 		tq->tq_flags |= TASKQ_CHANGING;
844 		tq->tq_nthreads_target = newtarget;
845 		cv_broadcast(&tq->tq_dispatch_cv);
846 		cv_broadcast(&tq->tq_exit_cv);
847 	}
848 }
849 
850 /* called during task queue creation */
851 static void
852 taskq_cpupct_install(taskq_t *tq, cpupart_t *cpup)
853 {
854 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
855 
856 	mutex_enter(&cpu_lock);
857 	mutex_enter(&tq->tq_lock);
858 	tq->tq_cpupart = cpup->cp_id;
859 	taskq_update_nthreads(tq, cpup->cp_ncpus);
860 	mutex_exit(&tq->tq_lock);
861 
862 	list_insert_tail(&taskq_cpupct_list, tq);
863 	mutex_exit(&cpu_lock);
864 }
865 
866 static void
867 taskq_cpupct_remove(taskq_t *tq)
868 {
869 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
870 
871 	mutex_enter(&cpu_lock);
872 	list_remove(&taskq_cpupct_list, tq);
873 	mutex_exit(&cpu_lock);
874 }
875 
876 /*ARGSUSED*/
877 static int
878 taskq_cpu_setup(cpu_setup_t what, int id, void *arg)
879 {
880 	taskq_t *tq;
881 	cpupart_t *cp = cpu[id]->cpu_part;
882 	uint_t ncpus = cp->cp_ncpus;
883 
884 	ASSERT(MUTEX_HELD(&cpu_lock));
885 	ASSERT(ncpus > 0);
886 
887 	switch (what) {
888 	case CPU_OFF:
889 	case CPU_CPUPART_OUT:
890 		/* offlines are called *before* the cpu is offlined. */
891 		if (ncpus > 1)
892 			ncpus--;
893 		break;
894 
895 	case CPU_ON:
896 	case CPU_CPUPART_IN:
897 		break;
898 
899 	default:
900 		return (0);		/* doesn't affect cpu count */
901 	}
902 
903 	for (tq = list_head(&taskq_cpupct_list); tq != NULL;
904 	    tq = list_next(&taskq_cpupct_list, tq)) {
905 
906 		mutex_enter(&tq->tq_lock);
907 		/*
908 		 * If the taskq is part of the cpuset which is changing,
909 		 * update its nthreads_target.
910 		 */
911 		if (tq->tq_cpupart == cp->cp_id) {
912 			taskq_update_nthreads(tq, ncpus);
913 		}
914 		mutex_exit(&tq->tq_lock);
915 	}
916 	return (0);
917 }
918 
919 void
920 taskq_mp_init(void)
921 {
922 	mutex_enter(&cpu_lock);
923 	register_cpu_setup_func(taskq_cpu_setup, NULL);
924 	/*
925 	 * Make sure we're up to date.  At this point in boot, there is only
926 	 * one processor set, so we only have to update the current CPU.
927 	 */
928 	(void) taskq_cpu_setup(CPU_ON, CPU->cpu_id, NULL);
929 	mutex_exit(&cpu_lock);
930 }
931 
932 /*
933  * Create global system dynamic task queue.
934  */
935 void
936 system_taskq_init(void)
937 {
938 	system_taskq = taskq_create_common("system_taskq", 0,
939 	    system_taskq_size * max_ncpus, minclsyspri, 4, 512, &p0, 0,
940 	    TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
941 }
942 
943 /*
944  * taskq_ent_alloc()
945  *
946  * Allocates a new taskq_ent_t structure either from the free list or from the
947  * cache. Returns NULL if it can't be allocated.
948  *
949  * Assumes: tq->tq_lock is held.
950  */
951 static taskq_ent_t *
952 taskq_ent_alloc(taskq_t *tq, int flags)
953 {
954 	int kmflags = (flags & TQ_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
955 
956 	taskq_ent_t *tqe;
957 
958 	ASSERT(MUTEX_HELD(&tq->tq_lock));
959 
960 	/*
961 	 * TQ_NOALLOC allocations are allowed to use the freelist, even if
962 	 * we are below tq_minalloc.
963 	 */
964 	if ((tqe = tq->tq_freelist) != NULL &&
965 	    ((flags & TQ_NOALLOC) || tq->tq_nalloc >= tq->tq_minalloc)) {
966 		tq->tq_freelist = tqe->tqent_next;
967 	} else {
968 		if (flags & TQ_NOALLOC)
969 			return (NULL);
970 
971 		mutex_exit(&tq->tq_lock);
972 		if (tq->tq_nalloc >= tq->tq_maxalloc) {
973 			if (kmflags & KM_NOSLEEP) {
974 				mutex_enter(&tq->tq_lock);
975 				return (NULL);
976 			}
977 			/*
978 			 * We don't want to exceed tq_maxalloc, but we can't
979 			 * wait for other tasks to complete (and thus free up
980 			 * task structures) without risking deadlock with
981 			 * the caller.  So, we just delay for one second
982 			 * to throttle the allocation rate.
983 			 */
984 			delay(hz);
985 		}
986 		tqe = kmem_cache_alloc(taskq_ent_cache, kmflags);
987 		mutex_enter(&tq->tq_lock);
988 		if (tqe != NULL)
989 			tq->tq_nalloc++;
990 	}
991 	return (tqe);
992 }
993 
994 /*
995  * taskq_ent_free()
996  *
997  * Free taskq_ent_t structure by either putting it on the free list or freeing
998  * it to the cache.
999  *
1000  * Assumes: tq->tq_lock is held.
1001  */
1002 static void
1003 taskq_ent_free(taskq_t *tq, taskq_ent_t *tqe)
1004 {
1005 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1006 
1007 	if (tq->tq_nalloc <= tq->tq_minalloc) {
1008 		tqe->tqent_next = tq->tq_freelist;
1009 		tq->tq_freelist = tqe;
1010 	} else {
1011 		tq->tq_nalloc--;
1012 		mutex_exit(&tq->tq_lock);
1013 		kmem_cache_free(taskq_ent_cache, tqe);
1014 		mutex_enter(&tq->tq_lock);
1015 	}
1016 }
1017 
1018 /*
1019  * Dispatch a task "func(arg)" to a free entry of bucket b.
1020  *
1021  * Assumes: no bucket locks is held.
1022  *
1023  * Returns: a pointer to an entry if dispatch was successful.
1024  *	    NULL if there are no free entries or if the bucket is suspended.
1025  */
1026 static taskq_ent_t *
1027 taskq_bucket_dispatch(taskq_bucket_t *b, task_func_t func, void *arg)
1028 {
1029 	taskq_ent_t *tqe;
1030 
1031 	ASSERT(MUTEX_NOT_HELD(&b->tqbucket_lock));
1032 	ASSERT(func != NULL);
1033 
1034 	mutex_enter(&b->tqbucket_lock);
1035 
1036 	ASSERT(b->tqbucket_nfree != 0 || IS_EMPTY(b->tqbucket_freelist));
1037 	ASSERT(b->tqbucket_nfree == 0 || !IS_EMPTY(b->tqbucket_freelist));
1038 
1039 	/*
1040 	 * Get en entry from the freelist if there is one.
1041 	 * Schedule task into the entry.
1042 	 */
1043 	if ((b->tqbucket_nfree != 0) &&
1044 	    !(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
1045 		tqe = b->tqbucket_freelist.tqent_prev;
1046 
1047 		ASSERT(tqe != &b->tqbucket_freelist);
1048 		ASSERT(tqe->tqent_thread != NULL);
1049 
1050 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1051 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1052 		b->tqbucket_nalloc++;
1053 		b->tqbucket_nfree--;
1054 		tqe->tqent_func = func;
1055 		tqe->tqent_arg = arg;
1056 		TQ_STAT(b, tqs_hits);
1057 		cv_signal(&tqe->tqent_cv);
1058 		DTRACE_PROBE2(taskq__d__enqueue, taskq_bucket_t *, b,
1059 		    taskq_ent_t *, tqe);
1060 	} else {
1061 		tqe = NULL;
1062 		TQ_STAT(b, tqs_misses);
1063 	}
1064 	mutex_exit(&b->tqbucket_lock);
1065 	return (tqe);
1066 }
1067 
1068 /*
1069  * Dispatch a task.
1070  *
1071  * Assumes: func != NULL
1072  *
1073  * Returns: NULL if dispatch failed.
1074  *	    non-NULL if task dispatched successfully.
1075  *	    Actual return value is the pointer to taskq entry that was used to
1076  *	    dispatch a task. This is useful for debugging.
1077  */
1078 /* ARGSUSED */
1079 taskqid_t
1080 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
1081 {
1082 	taskq_bucket_t *bucket = NULL;	/* Which bucket needs extension */
1083 	taskq_ent_t *tqe = NULL;
1084 	taskq_ent_t *tqe1;
1085 	uint_t bsize;
1086 
1087 	ASSERT(tq != NULL);
1088 	ASSERT(func != NULL);
1089 
1090 	if (!(tq->tq_flags & TASKQ_DYNAMIC)) {
1091 		/*
1092 		 * TQ_NOQUEUE flag can't be used with non-dynamic task queues.
1093 		 */
1094 		ASSERT(! (flags & TQ_NOQUEUE));
1095 		/*
1096 		 * Enqueue the task to the underlying queue.
1097 		 */
1098 		mutex_enter(&tq->tq_lock);
1099 
1100 		TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flags);
1101 
1102 		if ((tqe = taskq_ent_alloc(tq, flags)) == NULL) {
1103 			mutex_exit(&tq->tq_lock);
1104 			return (NULL);
1105 		}
1106 		if (flags & TQ_FRONT) {
1107 			TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1108 		} else {
1109 			TQ_ENQUEUE(tq, tqe, func, arg);
1110 		}
1111 		mutex_exit(&tq->tq_lock);
1112 		return ((taskqid_t)tqe);
1113 	}
1114 
1115 	/*
1116 	 * Dynamic taskq dispatching.
1117 	 */
1118 	ASSERT(!(flags & (TQ_NOALLOC | TQ_FRONT)));
1119 	TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flags);
1120 
1121 	bsize = tq->tq_nbuckets;
1122 
1123 	if (bsize == 1) {
1124 		/*
1125 		 * In a single-CPU case there is only one bucket, so get
1126 		 * entry directly from there.
1127 		 */
1128 		if ((tqe = taskq_bucket_dispatch(tq->tq_buckets, func, arg))
1129 		    != NULL)
1130 			return ((taskqid_t)tqe);	/* Fastpath */
1131 		bucket = tq->tq_buckets;
1132 	} else {
1133 		int loopcount;
1134 		taskq_bucket_t *b;
1135 		uintptr_t h = ((uintptr_t)CPU + (uintptr_t)arg) >> 3;
1136 
1137 		h = TQ_HASH(h);
1138 
1139 		/*
1140 		 * The 'bucket' points to the original bucket that we hit. If we
1141 		 * can't allocate from it, we search other buckets, but only
1142 		 * extend this one.
1143 		 */
1144 		b = &tq->tq_buckets[h & (bsize - 1)];
1145 		ASSERT(b->tqbucket_taskq == tq);	/* Sanity check */
1146 
1147 		/*
1148 		 * Do a quick check before grabbing the lock. If the bucket does
1149 		 * not have free entries now, chances are very small that it
1150 		 * will after we take the lock, so we just skip it.
1151 		 */
1152 		if (b->tqbucket_nfree != 0) {
1153 			if ((tqe = taskq_bucket_dispatch(b, func, arg)) != NULL)
1154 				return ((taskqid_t)tqe);	/* Fastpath */
1155 		} else {
1156 			TQ_STAT(b, tqs_misses);
1157 		}
1158 
1159 		bucket = b;
1160 		loopcount = MIN(taskq_search_depth, bsize);
1161 		/*
1162 		 * If bucket dispatch failed, search loopcount number of buckets
1163 		 * before we give up and fail.
1164 		 */
1165 		do {
1166 			b = &tq->tq_buckets[++h & (bsize - 1)];
1167 			ASSERT(b->tqbucket_taskq == tq);  /* Sanity check */
1168 			loopcount--;
1169 
1170 			if (b->tqbucket_nfree != 0) {
1171 				tqe = taskq_bucket_dispatch(b, func, arg);
1172 			} else {
1173 				TQ_STAT(b, tqs_misses);
1174 			}
1175 		} while ((tqe == NULL) && (loopcount > 0));
1176 	}
1177 
1178 	/*
1179 	 * At this point we either scheduled a task and (tqe != NULL) or failed
1180 	 * (tqe == NULL). Try to recover from fails.
1181 	 */
1182 
1183 	/*
1184 	 * For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
1185 	 */
1186 	if ((tqe == NULL) && !(flags & TQ_NOSLEEP)) {
1187 		/*
1188 		 * taskq_bucket_extend() may fail to do anything, but this is
1189 		 * fine - we deal with it later. If the bucket was successfully
1190 		 * extended, there is a good chance that taskq_bucket_dispatch()
1191 		 * will get this new entry, unless someone is racing with us and
1192 		 * stealing the new entry from under our nose.
1193 		 * taskq_bucket_extend() may sleep.
1194 		 */
1195 		taskq_bucket_extend(bucket);
1196 		TQ_STAT(bucket, tqs_disptcreates);
1197 		if ((tqe = taskq_bucket_dispatch(bucket, func, arg)) != NULL)
1198 			return ((taskqid_t)tqe);
1199 	}
1200 
1201 	ASSERT(bucket != NULL);
1202 	/*
1203 	 * Since there are not enough free entries in the bucket, extend it
1204 	 * in the background using backing queue.
1205 	 */
1206 	mutex_enter(&tq->tq_lock);
1207 	if ((tqe1 = taskq_ent_alloc(tq, TQ_NOSLEEP)) != NULL) {
1208 		TQ_ENQUEUE(tq, tqe1, taskq_bucket_extend, bucket);
1209 	} else {
1210 		TQ_STAT(bucket, tqs_nomem);
1211 	}
1212 
1213 	/*
1214 	 * Dispatch failed and we can't find an entry to schedule a task.
1215 	 * Revert to the backing queue unless TQ_NOQUEUE was asked.
1216 	 */
1217 	if ((tqe == NULL) && !(flags & TQ_NOQUEUE)) {
1218 		if ((tqe = taskq_ent_alloc(tq, flags)) != NULL) {
1219 			TQ_ENQUEUE(tq, tqe, func, arg);
1220 		} else {
1221 			TQ_STAT(bucket, tqs_nomem);
1222 		}
1223 	}
1224 	mutex_exit(&tq->tq_lock);
1225 
1226 	return ((taskqid_t)tqe);
1227 }
1228 
1229 /*
1230  * Wait for all pending tasks to complete.
1231  * Calling taskq_wait from a task will cause deadlock.
1232  */
1233 void
1234 taskq_wait(taskq_t *tq)
1235 {
1236 	ASSERT(tq != curthread->t_taskq);
1237 
1238 	mutex_enter(&tq->tq_lock);
1239 	while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
1240 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1241 	mutex_exit(&tq->tq_lock);
1242 
1243 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1244 		taskq_bucket_t *b = tq->tq_buckets;
1245 		int bid = 0;
1246 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1247 			mutex_enter(&b->tqbucket_lock);
1248 			while (b->tqbucket_nalloc > 0)
1249 				cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
1250 			mutex_exit(&b->tqbucket_lock);
1251 		}
1252 	}
1253 }
1254 
1255 /*
1256  * Suspend execution of tasks.
1257  *
1258  * Tasks in the queue part will be suspended immediately upon return from this
1259  * function. Pending tasks in the dynamic part will continue to execute, but all
1260  * new tasks will  be suspended.
1261  */
1262 void
1263 taskq_suspend(taskq_t *tq)
1264 {
1265 	rw_enter(&tq->tq_threadlock, RW_WRITER);
1266 
1267 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1268 		taskq_bucket_t *b = tq->tq_buckets;
1269 		int bid = 0;
1270 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1271 			mutex_enter(&b->tqbucket_lock);
1272 			b->tqbucket_flags |= TQBUCKET_SUSPEND;
1273 			mutex_exit(&b->tqbucket_lock);
1274 		}
1275 	}
1276 	/*
1277 	 * Mark task queue as being suspended. Needed for taskq_suspended().
1278 	 */
1279 	mutex_enter(&tq->tq_lock);
1280 	ASSERT(!(tq->tq_flags & TASKQ_SUSPENDED));
1281 	tq->tq_flags |= TASKQ_SUSPENDED;
1282 	mutex_exit(&tq->tq_lock);
1283 }
1284 
1285 /*
1286  * returns: 1 if tq is suspended, 0 otherwise.
1287  */
1288 int
1289 taskq_suspended(taskq_t *tq)
1290 {
1291 	return ((tq->tq_flags & TASKQ_SUSPENDED) != 0);
1292 }
1293 
1294 /*
1295  * Resume taskq execution.
1296  */
1297 void
1298 taskq_resume(taskq_t *tq)
1299 {
1300 	ASSERT(RW_WRITE_HELD(&tq->tq_threadlock));
1301 
1302 	if (tq->tq_flags & TASKQ_DYNAMIC) {
1303 		taskq_bucket_t *b = tq->tq_buckets;
1304 		int bid = 0;
1305 		for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1306 			mutex_enter(&b->tqbucket_lock);
1307 			b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
1308 			mutex_exit(&b->tqbucket_lock);
1309 		}
1310 	}
1311 	mutex_enter(&tq->tq_lock);
1312 	ASSERT(tq->tq_flags & TASKQ_SUSPENDED);
1313 	tq->tq_flags &= ~TASKQ_SUSPENDED;
1314 	mutex_exit(&tq->tq_lock);
1315 
1316 	rw_exit(&tq->tq_threadlock);
1317 }
1318 
1319 int
1320 taskq_member(taskq_t *tq, kthread_t *thread)
1321 {
1322 	return (thread->t_taskq == tq);
1323 }
1324 
1325 /*
1326  * Creates a thread in the taskq.  We only allow one outstanding create at
1327  * a time.  We drop and reacquire the tq_lock in order to avoid blocking other
1328  * taskq activity while thread_create() or lwp_kernel_create() run.
1329  *
1330  * The first time we're called, we do some additional setup, and do not
1331  * return until there are enough threads to start servicing requests.
1332  */
1333 static void
1334 taskq_thread_create(taskq_t *tq)
1335 {
1336 	kthread_t	*t;
1337 	const boolean_t	first = (tq->tq_nthreads == 0);
1338 
1339 	ASSERT(MUTEX_HELD(&tq->tq_lock));
1340 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1341 	ASSERT(tq->tq_nthreads < tq->tq_nthreads_target);
1342 	ASSERT(!(tq->tq_flags & TASKQ_THREAD_CREATED));
1343 
1344 
1345 	tq->tq_flags |= TASKQ_THREAD_CREATED;
1346 	tq->tq_active++;
1347 	mutex_exit(&tq->tq_lock);
1348 
1349 	if (tq->tq_proc != &p0) {
1350 		t = lwp_kernel_create(tq->tq_proc, taskq_thread, tq, TS_RUN,
1351 		    tq->tq_pri);
1352 	} else {
1353 		t = thread_create(NULL, 0, taskq_thread, tq, 0, &p0, TS_RUN,
1354 		    tq->tq_pri);
1355 	}
1356 
1357 	if (!first) {
1358 		mutex_enter(&tq->tq_lock);
1359 		return;
1360 	}
1361 
1362 	/*
1363 	 * We know the thread cannot go away, since tq cannot be
1364 	 * destroyed until creation has completed.  We can therefore
1365 	 * safely dereference t.
1366 	 */
1367 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1368 		taskq_cpupct_install(tq, t->t_cpupart);
1369 	}
1370 	mutex_enter(&tq->tq_lock);
1371 
1372 	/* Wait until we can service requests. */
1373 	while (tq->tq_nthreads != tq->tq_nthreads_target &&
1374 	    tq->tq_nthreads < TASKQ_CREATE_ACTIVE_THREADS) {
1375 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1376 	}
1377 }
1378 
1379 /*
1380  * Common "sleep taskq thread" function, which handles CPR stuff, as well
1381  * as giving a nice common point for debuggers to find inactive threads.
1382  */
1383 static clock_t
1384 taskq_thread_wait(taskq_t *tq, kmutex_t *mx, kcondvar_t *cv,
1385     callb_cpr_t *cprinfo, clock_t timeout)
1386 {
1387 	clock_t ret = 0;
1388 
1389 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1390 		CALLB_CPR_SAFE_BEGIN(cprinfo);
1391 	}
1392 	if (timeout < 0)
1393 		cv_wait(cv, mx);
1394 	else
1395 		ret = cv_reltimedwait(cv, mx, timeout, TR_CLOCK_TICK);
1396 
1397 	if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1398 		CALLB_CPR_SAFE_END(cprinfo, mx);
1399 	}
1400 
1401 	return (ret);
1402 }
1403 
1404 /*
1405  * Worker thread for processing task queue.
1406  */
1407 static void
1408 taskq_thread(void *arg)
1409 {
1410 	int thread_id;
1411 
1412 	taskq_t *tq = arg;
1413 	taskq_ent_t *tqe;
1414 	callb_cpr_t cprinfo;
1415 	hrtime_t start, end;
1416 
1417 	curthread->t_taskq = tq;	/* mark ourselves for taskq_member() */
1418 
1419 	if (curproc != &p0 && (tq->tq_flags & TASKQ_DUTY_CYCLE)) {
1420 		sysdc_thread_enter(curthread, tq->tq_DC,
1421 		    (tq->tq_flags & TASKQ_DC_BATCH) ? SYSDC_THREAD_BATCH : 0);
1422 	}
1423 
1424 	if (tq->tq_flags & TASKQ_CPR_SAFE) {
1425 		CALLB_CPR_INIT_SAFE(curthread, tq->tq_name);
1426 	} else {
1427 		CALLB_CPR_INIT(&cprinfo, &tq->tq_lock, callb_generic_cpr,
1428 		    tq->tq_name);
1429 	}
1430 	mutex_enter(&tq->tq_lock);
1431 	thread_id = ++tq->tq_nthreads;
1432 	ASSERT(tq->tq_flags & TASKQ_THREAD_CREATED);
1433 	ASSERT(tq->tq_flags & TASKQ_CHANGING);
1434 	tq->tq_flags &= ~TASKQ_THREAD_CREATED;
1435 
1436 	VERIFY3S(thread_id, <=, tq->tq_nthreads_max);
1437 
1438 	if (tq->tq_nthreads_max == 1)
1439 		tq->tq_thread = curthread;
1440 	else
1441 		tq->tq_threadlist[thread_id - 1] = curthread;
1442 
1443 	/* Allow taskq_create_common()'s taskq_thread_create() to return. */
1444 	if (tq->tq_nthreads == TASKQ_CREATE_ACTIVE_THREADS)
1445 		cv_broadcast(&tq->tq_wait_cv);
1446 
1447 	for (;;) {
1448 		if (tq->tq_flags & TASKQ_CHANGING) {
1449 			/* See if we're no longer needed */
1450 			if (thread_id > tq->tq_nthreads_target) {
1451 				/*
1452 				 * To preserve the one-to-one mapping between
1453 				 * thread_id and thread, we must exit from
1454 				 * highest thread ID to least.
1455 				 *
1456 				 * However, if everyone is exiting, the order
1457 				 * doesn't matter, so just exit immediately.
1458 				 * (this is safe, since you must wait for
1459 				 * nthreads to reach 0 after setting
1460 				 * tq_nthreads_target to 0)
1461 				 */
1462 				if (thread_id == tq->tq_nthreads ||
1463 				    tq->tq_nthreads_target == 0)
1464 					break;
1465 
1466 				/* Wait for higher thread_ids to exit */
1467 				(void) taskq_thread_wait(tq, &tq->tq_lock,
1468 				    &tq->tq_exit_cv, &cprinfo, -1);
1469 				continue;
1470 			}
1471 
1472 			/*
1473 			 * If no thread is starting taskq_thread(), we can
1474 			 * do some bookkeeping.
1475 			 */
1476 			if (!(tq->tq_flags & TASKQ_THREAD_CREATED)) {
1477 				/* Check if we've reached our target */
1478 				if (tq->tq_nthreads == tq->tq_nthreads_target) {
1479 					tq->tq_flags &= ~TASKQ_CHANGING;
1480 					cv_broadcast(&tq->tq_wait_cv);
1481 				}
1482 				/* Check if we need to create a thread */
1483 				if (tq->tq_nthreads < tq->tq_nthreads_target) {
1484 					taskq_thread_create(tq);
1485 					continue; /* tq_lock was dropped */
1486 				}
1487 			}
1488 		}
1489 		if ((tqe = tq->tq_task.tqent_next) == &tq->tq_task) {
1490 			if (--tq->tq_active == 0)
1491 				cv_broadcast(&tq->tq_wait_cv);
1492 			(void) taskq_thread_wait(tq, &tq->tq_lock,
1493 			    &tq->tq_dispatch_cv, &cprinfo, -1);
1494 			tq->tq_active++;
1495 			continue;
1496 		}
1497 
1498 		tqe->tqent_prev->tqent_next = tqe->tqent_next;
1499 		tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1500 		mutex_exit(&tq->tq_lock);
1501 
1502 		rw_enter(&tq->tq_threadlock, RW_READER);
1503 		start = gethrtime();
1504 		DTRACE_PROBE2(taskq__exec__start, taskq_t *, tq,
1505 		    taskq_ent_t *, tqe);
1506 		tqe->tqent_func(tqe->tqent_arg);
1507 		DTRACE_PROBE2(taskq__exec__end, taskq_t *, tq,
1508 		    taskq_ent_t *, tqe);
1509 		end = gethrtime();
1510 		rw_exit(&tq->tq_threadlock);
1511 
1512 		mutex_enter(&tq->tq_lock);
1513 		tq->tq_totaltime += end - start;
1514 		tq->tq_executed++;
1515 
1516 		taskq_ent_free(tq, tqe);
1517 	}
1518 
1519 	if (tq->tq_nthreads_max == 1)
1520 		tq->tq_thread = NULL;
1521 	else
1522 		tq->tq_threadlist[thread_id - 1] = NULL;
1523 
1524 	/* We're exiting, and therefore no longer active */
1525 	ASSERT(tq->tq_active > 0);
1526 	tq->tq_active--;
1527 
1528 	ASSERT(tq->tq_nthreads > 0);
1529 	tq->tq_nthreads--;
1530 
1531 	/* Wake up anyone waiting for us to exit */
1532 	cv_broadcast(&tq->tq_exit_cv);
1533 	if (tq->tq_nthreads == tq->tq_nthreads_target) {
1534 		if (!(tq->tq_flags & TASKQ_THREAD_CREATED))
1535 			tq->tq_flags &= ~TASKQ_CHANGING;
1536 
1537 		cv_broadcast(&tq->tq_wait_cv);
1538 	}
1539 
1540 	ASSERT(!(tq->tq_flags & TASKQ_CPR_SAFE));
1541 	CALLB_CPR_EXIT(&cprinfo);		/* drops tq->tq_lock */
1542 	if (curthread->t_lwp != NULL) {
1543 		mutex_enter(&curproc->p_lock);
1544 		lwp_exit();
1545 	} else {
1546 		thread_exit();
1547 	}
1548 }
1549 
1550 /*
1551  * Worker per-entry thread for dynamic dispatches.
1552  */
1553 static void
1554 taskq_d_thread(taskq_ent_t *tqe)
1555 {
1556 	taskq_bucket_t	*bucket = tqe->tqent_bucket;
1557 	taskq_t		*tq = bucket->tqbucket_taskq;
1558 	kmutex_t	*lock = &bucket->tqbucket_lock;
1559 	kcondvar_t	*cv = &tqe->tqent_cv;
1560 	callb_cpr_t	cprinfo;
1561 	clock_t		w;
1562 
1563 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, tq->tq_name);
1564 
1565 	mutex_enter(lock);
1566 
1567 	for (;;) {
1568 		/*
1569 		 * If a task is scheduled (func != NULL), execute it, otherwise
1570 		 * sleep, waiting for a job.
1571 		 */
1572 		if (tqe->tqent_func != NULL) {
1573 			hrtime_t	start;
1574 			hrtime_t	end;
1575 
1576 			ASSERT(bucket->tqbucket_nalloc > 0);
1577 
1578 			/*
1579 			 * It is possible to free the entry right away before
1580 			 * actually executing the task so that subsequent
1581 			 * dispatches may immediately reuse it. But this,
1582 			 * effectively, creates a two-length queue in the entry
1583 			 * and may lead to a deadlock if the execution of the
1584 			 * current task depends on the execution of the next
1585 			 * scheduled task. So, we keep the entry busy until the
1586 			 * task is processed.
1587 			 */
1588 
1589 			mutex_exit(lock);
1590 			start = gethrtime();
1591 			DTRACE_PROBE3(taskq__d__exec__start, taskq_t *, tq,
1592 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1593 			tqe->tqent_func(tqe->tqent_arg);
1594 			DTRACE_PROBE3(taskq__d__exec__end, taskq_t *, tq,
1595 			    taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1596 			end = gethrtime();
1597 			mutex_enter(lock);
1598 			bucket->tqbucket_totaltime += end - start;
1599 
1600 			/*
1601 			 * Return the entry to the bucket free list.
1602 			 */
1603 			tqe->tqent_func = NULL;
1604 			TQ_APPEND(bucket->tqbucket_freelist, tqe);
1605 			bucket->tqbucket_nalloc--;
1606 			bucket->tqbucket_nfree++;
1607 			ASSERT(!IS_EMPTY(bucket->tqbucket_freelist));
1608 			/*
1609 			 * taskq_wait() waits for nalloc to drop to zero on
1610 			 * tqbucket_cv.
1611 			 */
1612 			cv_signal(&bucket->tqbucket_cv);
1613 		}
1614 
1615 		/*
1616 		 * At this point the entry must be in the bucket free list -
1617 		 * either because it was there initially or because it just
1618 		 * finished executing a task and put itself on the free list.
1619 		 */
1620 		ASSERT(bucket->tqbucket_nfree > 0);
1621 		/*
1622 		 * Go to sleep unless we are closing.
1623 		 * If a thread is sleeping too long, it dies.
1624 		 */
1625 		if (! (bucket->tqbucket_flags & TQBUCKET_CLOSE)) {
1626 			w = taskq_thread_wait(tq, lock, cv,
1627 			    &cprinfo, taskq_thread_timeout * hz);
1628 		}
1629 
1630 		/*
1631 		 * At this point we may be in two different states:
1632 		 *
1633 		 * (1) tqent_func is set which means that a new task is
1634 		 *	dispatched and we need to execute it.
1635 		 *
1636 		 * (2) Thread is sleeping for too long or we are closing. In
1637 		 *	both cases destroy the thread and the entry.
1638 		 */
1639 
1640 		/* If func is NULL we should be on the freelist. */
1641 		ASSERT((tqe->tqent_func != NULL) ||
1642 		    (bucket->tqbucket_nfree > 0));
1643 		/* If func is non-NULL we should be allocated */
1644 		ASSERT((tqe->tqent_func == NULL) ||
1645 		    (bucket->tqbucket_nalloc > 0));
1646 
1647 		/* Check freelist consistency */
1648 		ASSERT((bucket->tqbucket_nfree > 0) ||
1649 		    IS_EMPTY(bucket->tqbucket_freelist));
1650 		ASSERT((bucket->tqbucket_nfree == 0) ||
1651 		    !IS_EMPTY(bucket->tqbucket_freelist));
1652 
1653 		if ((tqe->tqent_func == NULL) &&
1654 		    ((w == -1) || (bucket->tqbucket_flags & TQBUCKET_CLOSE))) {
1655 			/*
1656 			 * This thread is sleeping for too long or we are
1657 			 * closing - time to die.
1658 			 * Thread creation/destruction happens rarely,
1659 			 * so grabbing the lock is not a big performance issue.
1660 			 * The bucket lock is dropped by CALLB_CPR_EXIT().
1661 			 */
1662 
1663 			/* Remove the entry from the free list. */
1664 			tqe->tqent_prev->tqent_next = tqe->tqent_next;
1665 			tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1666 			ASSERT(bucket->tqbucket_nfree > 0);
1667 			bucket->tqbucket_nfree--;
1668 
1669 			TQ_STAT(bucket, tqs_tdeaths);
1670 			cv_signal(&bucket->tqbucket_cv);
1671 			tqe->tqent_thread = NULL;
1672 			mutex_enter(&tq->tq_lock);
1673 			tq->tq_tdeaths++;
1674 			mutex_exit(&tq->tq_lock);
1675 			CALLB_CPR_EXIT(&cprinfo);
1676 			kmem_cache_free(taskq_ent_cache, tqe);
1677 			thread_exit();
1678 		}
1679 	}
1680 }
1681 
1682 
1683 /*
1684  * Taskq creation. May sleep for memory.
1685  * Always use automatically generated instances to avoid kstat name space
1686  * collisions.
1687  */
1688 
1689 taskq_t *
1690 taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
1691     int maxalloc, uint_t flags)
1692 {
1693 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1694 
1695 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1696 	    maxalloc, &p0, 0, flags | TASKQ_NOINSTANCE));
1697 }
1698 
1699 /*
1700  * Create an instance of task queue. It is legal to create task queues with the
1701  * same name and different instances.
1702  *
1703  * taskq_create_instance is used by ddi_taskq_create() where it gets the
1704  * instance from ddi_get_instance(). In some cases the instance is not
1705  * initialized and is set to -1. This case is handled as if no instance was
1706  * passed at all.
1707  */
1708 taskq_t *
1709 taskq_create_instance(const char *name, int instance, int nthreads, pri_t pri,
1710     int minalloc, int maxalloc, uint_t flags)
1711 {
1712 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1713 	ASSERT((instance >= 0) || (instance == -1));
1714 
1715 	if (instance < 0) {
1716 		flags |= TASKQ_NOINSTANCE;
1717 	}
1718 
1719 	return (taskq_create_common(name, instance, nthreads,
1720 	    pri, minalloc, maxalloc, &p0, 0, flags));
1721 }
1722 
1723 taskq_t *
1724 taskq_create_proc(const char *name, int nthreads, pri_t pri, int minalloc,
1725     int maxalloc, proc_t *proc, uint_t flags)
1726 {
1727 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1728 	ASSERT(proc->p_flag & SSYS);
1729 
1730 	return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1731 	    maxalloc, proc, 0, flags | TASKQ_NOINSTANCE));
1732 }
1733 
1734 taskq_t *
1735 taskq_create_sysdc(const char *name, int nthreads, int minalloc,
1736     int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1737 {
1738 	ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1739 	ASSERT(proc->p_flag & SSYS);
1740 
1741 	return (taskq_create_common(name, 0, nthreads, minclsyspri, minalloc,
1742 	    maxalloc, proc, dc, flags | TASKQ_NOINSTANCE | TASKQ_DUTY_CYCLE));
1743 }
1744 
1745 #define	IMPLY(a, b)	ASSERT((!(a)) || (b)) /* if (a) { ASSERT (b) } */
1746 
1747 static taskq_t *
1748 taskq_create_common(const char *name, int instance, int nthreads, pri_t pri,
1749     int minalloc, int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1750 {
1751 	taskq_t *tq = kmem_cache_alloc(taskq_cache, KM_SLEEP);
1752 	uint_t ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
1753 	uint_t bsize;	/* # of buckets - always power of 2 */
1754 	int max_nthreads;
1755 
1756 	/*
1757 	 * TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
1758 	 * mutually incompatible.
1759 	 */
1760 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_CPR_SAFE));
1761 	IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_THREADS_CPU_PCT));
1762 	IMPLY((flags & TASKQ_CPR_SAFE), !(flags & TASKQ_THREADS_CPU_PCT));
1763 
1764 	/* Cannot have DUTY_CYCLE without a non-p0 kernel process */
1765 	IMPLY((flags & TASKQ_DUTY_CYCLE), proc != &p0);
1766 
1767 	/* Cannot have DC_BATCH without DUTY_CYCLE */
1768 	ASSERT((flags & (TASKQ_DUTY_CYCLE|TASKQ_DC_BATCH)) != TASKQ_DC_BATCH);
1769 
1770 	ASSERT(proc != NULL);
1771 
1772 	bsize = 1 << (highbit(ncpus) - 1);
1773 	ASSERT(bsize >= 1);
1774 	bsize = MIN(bsize, taskq_maxbuckets);
1775 
1776 	if (flags & TASKQ_DYNAMIC) {
1777 		ASSERT3S(nthreads, >=, 1);
1778 		tq->tq_maxsize = nthreads;
1779 
1780 		/* For dynamic task queues use just one backup thread */
1781 		nthreads = max_nthreads = 1;
1782 
1783 	} else if (flags & TASKQ_THREADS_CPU_PCT) {
1784 		uint_t pct;
1785 		ASSERT3S(nthreads, >=, 0);
1786 		pct = nthreads;
1787 
1788 		if (pct > taskq_cpupct_max_percent)
1789 			pct = taskq_cpupct_max_percent;
1790 
1791 		/*
1792 		 * If you're using THREADS_CPU_PCT, the process for the
1793 		 * taskq threads must be curproc.  This allows any pset
1794 		 * binding to be inherited correctly.  If proc is &p0,
1795 		 * we won't be creating LWPs, so new threads will be assigned
1796 		 * to the default processor set.
1797 		 */
1798 		ASSERT(curproc == proc || proc == &p0);
1799 		tq->tq_threads_ncpus_pct = pct;
1800 		nthreads = 1;		/* corrected in taskq_thread_create() */
1801 		max_nthreads = TASKQ_THREADS_PCT(max_ncpus, pct);
1802 
1803 	} else {
1804 		ASSERT3S(nthreads, >=, 1);
1805 		max_nthreads = nthreads;
1806 	}
1807 
1808 	if (max_nthreads < taskq_minimum_nthreads_max)
1809 		max_nthreads = taskq_minimum_nthreads_max;
1810 
1811 	/*
1812 	 * Make sure the name is 0-terminated, and conforms to the rules for
1813 	 * C indentifiers
1814 	 */
1815 	(void) strncpy(tq->tq_name, name, TASKQ_NAMELEN + 1);
1816 	strident_canon(tq->tq_name, TASKQ_NAMELEN + 1);
1817 
1818 	tq->tq_flags = flags | TASKQ_CHANGING;
1819 	tq->tq_active = 0;
1820 	tq->tq_instance = instance;
1821 	tq->tq_nthreads_target = nthreads;
1822 	tq->tq_nthreads_max = max_nthreads;
1823 	tq->tq_minalloc = minalloc;
1824 	tq->tq_maxalloc = maxalloc;
1825 	tq->tq_nbuckets = bsize;
1826 	tq->tq_proc = proc;
1827 	tq->tq_pri = pri;
1828 	tq->tq_DC = dc;
1829 	list_link_init(&tq->tq_cpupct_link);
1830 
1831 	if (max_nthreads > 1)
1832 		tq->tq_threadlist = kmem_alloc(
1833 		    sizeof (kthread_t *) * max_nthreads, KM_SLEEP);
1834 
1835 	mutex_enter(&tq->tq_lock);
1836 	if (flags & TASKQ_PREPOPULATE) {
1837 		while (minalloc-- > 0)
1838 			taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
1839 	}
1840 
1841 	/*
1842 	 * Create the first thread, which will create any other threads
1843 	 * necessary.  taskq_thread_create will not return until we have
1844 	 * enough threads to be able to process requests.
1845 	 */
1846 	taskq_thread_create(tq);
1847 	mutex_exit(&tq->tq_lock);
1848 
1849 	if (flags & TASKQ_DYNAMIC) {
1850 		taskq_bucket_t *bucket = kmem_zalloc(sizeof (taskq_bucket_t) *
1851 		    bsize, KM_SLEEP);
1852 		int b_id;
1853 
1854 		tq->tq_buckets = bucket;
1855 
1856 		/* Initialize each bucket */
1857 		for (b_id = 0; b_id < bsize; b_id++, bucket++) {
1858 			mutex_init(&bucket->tqbucket_lock, NULL, MUTEX_DEFAULT,
1859 			    NULL);
1860 			cv_init(&bucket->tqbucket_cv, NULL, CV_DEFAULT, NULL);
1861 			bucket->tqbucket_taskq = tq;
1862 			bucket->tqbucket_freelist.tqent_next =
1863 			    bucket->tqbucket_freelist.tqent_prev =
1864 			    &bucket->tqbucket_freelist;
1865 			if (flags & TASKQ_PREPOPULATE)
1866 				taskq_bucket_extend(bucket);
1867 		}
1868 	}
1869 
1870 	/*
1871 	 * Install kstats.
1872 	 * We have two cases:
1873 	 *   1) Instance is provided to taskq_create_instance(). In this case it
1874 	 *	should be >= 0 and we use it.
1875 	 *
1876 	 *   2) Instance is not provided and is automatically generated
1877 	 */
1878 	if (flags & TASKQ_NOINSTANCE) {
1879 		instance = tq->tq_instance =
1880 		    (int)(uintptr_t)vmem_alloc(taskq_id_arena, 1, VM_SLEEP);
1881 	}
1882 
1883 	if (flags & TASKQ_DYNAMIC) {
1884 		if ((tq->tq_kstat = kstat_create("unix", instance,
1885 		    tq->tq_name, "taskq_d", KSTAT_TYPE_NAMED,
1886 		    sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
1887 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
1888 			tq->tq_kstat->ks_lock = &taskq_d_kstat_lock;
1889 			tq->tq_kstat->ks_data = &taskq_d_kstat;
1890 			tq->tq_kstat->ks_update = taskq_d_kstat_update;
1891 			tq->tq_kstat->ks_private = tq;
1892 			kstat_install(tq->tq_kstat);
1893 		}
1894 	} else {
1895 		if ((tq->tq_kstat = kstat_create("unix", instance, tq->tq_name,
1896 		    "taskq", KSTAT_TYPE_NAMED,
1897 		    sizeof (taskq_kstat) / sizeof (kstat_named_t),
1898 		    KSTAT_FLAG_VIRTUAL)) != NULL) {
1899 			tq->tq_kstat->ks_lock = &taskq_kstat_lock;
1900 			tq->tq_kstat->ks_data = &taskq_kstat;
1901 			tq->tq_kstat->ks_update = taskq_kstat_update;
1902 			tq->tq_kstat->ks_private = tq;
1903 			kstat_install(tq->tq_kstat);
1904 		}
1905 	}
1906 
1907 	return (tq);
1908 }
1909 
1910 /*
1911  * taskq_destroy().
1912  *
1913  * Assumes: by the time taskq_destroy is called no one will use this task queue
1914  * in any way and no one will try to dispatch entries in it.
1915  */
1916 void
1917 taskq_destroy(taskq_t *tq)
1918 {
1919 	taskq_bucket_t *b = tq->tq_buckets;
1920 	int bid = 0;
1921 
1922 	ASSERT(! (tq->tq_flags & TASKQ_CPR_SAFE));
1923 
1924 	/*
1925 	 * Destroy kstats.
1926 	 */
1927 	if (tq->tq_kstat != NULL) {
1928 		kstat_delete(tq->tq_kstat);
1929 		tq->tq_kstat = NULL;
1930 	}
1931 
1932 	/*
1933 	 * Destroy instance if needed.
1934 	 */
1935 	if (tq->tq_flags & TASKQ_NOINSTANCE) {
1936 		vmem_free(taskq_id_arena, (void *)(uintptr_t)(tq->tq_instance),
1937 		    1);
1938 		tq->tq_instance = 0;
1939 	}
1940 
1941 	/*
1942 	 * Unregister from the cpupct list.
1943 	 */
1944 	if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1945 		taskq_cpupct_remove(tq);
1946 	}
1947 
1948 	/*
1949 	 * Wait for any pending entries to complete.
1950 	 */
1951 	taskq_wait(tq);
1952 
1953 	mutex_enter(&tq->tq_lock);
1954 	ASSERT((tq->tq_task.tqent_next == &tq->tq_task) &&
1955 	    (tq->tq_active == 0));
1956 
1957 	/* notify all the threads that they need to exit */
1958 	tq->tq_nthreads_target = 0;
1959 
1960 	tq->tq_flags |= TASKQ_CHANGING;
1961 	cv_broadcast(&tq->tq_dispatch_cv);
1962 	cv_broadcast(&tq->tq_exit_cv);
1963 
1964 	while (tq->tq_nthreads != 0)
1965 		cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1966 
1967 	if (tq->tq_nthreads_max != 1)
1968 		kmem_free(tq->tq_threadlist, sizeof (kthread_t *) *
1969 		    tq->tq_nthreads_max);
1970 
1971 	tq->tq_minalloc = 0;
1972 	while (tq->tq_nalloc != 0)
1973 		taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
1974 
1975 	mutex_exit(&tq->tq_lock);
1976 
1977 	/*
1978 	 * Mark each bucket as closing and wakeup all sleeping threads.
1979 	 */
1980 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1981 		taskq_ent_t *tqe;
1982 
1983 		mutex_enter(&b->tqbucket_lock);
1984 
1985 		b->tqbucket_flags |= TQBUCKET_CLOSE;
1986 		/* Wakeup all sleeping threads */
1987 
1988 		for (tqe = b->tqbucket_freelist.tqent_next;
1989 		    tqe != &b->tqbucket_freelist; tqe = tqe->tqent_next)
1990 			cv_signal(&tqe->tqent_cv);
1991 
1992 		ASSERT(b->tqbucket_nalloc == 0);
1993 
1994 		/*
1995 		 * At this point we waited for all pending jobs to complete (in
1996 		 * both the task queue and the bucket and no new jobs should
1997 		 * arrive. Wait for all threads to die.
1998 		 */
1999 		while (b->tqbucket_nfree > 0)
2000 			cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
2001 		mutex_exit(&b->tqbucket_lock);
2002 		mutex_destroy(&b->tqbucket_lock);
2003 		cv_destroy(&b->tqbucket_cv);
2004 	}
2005 
2006 	if (tq->tq_buckets != NULL) {
2007 		ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2008 		kmem_free(tq->tq_buckets,
2009 		    sizeof (taskq_bucket_t) * tq->tq_nbuckets);
2010 
2011 		/* Cleanup fields before returning tq to the cache */
2012 		tq->tq_buckets = NULL;
2013 		tq->tq_tcreates = 0;
2014 		tq->tq_tdeaths = 0;
2015 	} else {
2016 		ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
2017 	}
2018 
2019 	tq->tq_threads_ncpus_pct = 0;
2020 	tq->tq_totaltime = 0;
2021 	tq->tq_tasks = 0;
2022 	tq->tq_maxtasks = 0;
2023 	tq->tq_executed = 0;
2024 	kmem_cache_free(taskq_cache, tq);
2025 }
2026 
2027 /*
2028  * Extend a bucket with a new entry on the free list and attach a worker thread
2029  * to it.
2030  *
2031  * Argument: pointer to the bucket.
2032  *
2033  * This function may quietly fail. It is only used by taskq_dispatch() which
2034  * handles such failures properly.
2035  */
2036 static void
2037 taskq_bucket_extend(void *arg)
2038 {
2039 	taskq_ent_t *tqe;
2040 	taskq_bucket_t *b = (taskq_bucket_t *)arg;
2041 	taskq_t *tq = b->tqbucket_taskq;
2042 	int nthreads;
2043 
2044 	if (! ENOUGH_MEMORY()) {
2045 		TQ_STAT(b, tqs_nomem);
2046 		return;
2047 	}
2048 
2049 	mutex_enter(&tq->tq_lock);
2050 
2051 	/*
2052 	 * Observe global taskq limits on the number of threads.
2053 	 */
2054 	if (tq->tq_tcreates++ - tq->tq_tdeaths > tq->tq_maxsize) {
2055 		tq->tq_tcreates--;
2056 		mutex_exit(&tq->tq_lock);
2057 		return;
2058 	}
2059 	mutex_exit(&tq->tq_lock);
2060 
2061 	tqe = kmem_cache_alloc(taskq_ent_cache, KM_NOSLEEP);
2062 
2063 	if (tqe == NULL) {
2064 		mutex_enter(&tq->tq_lock);
2065 		TQ_STAT(b, tqs_nomem);
2066 		tq->tq_tcreates--;
2067 		mutex_exit(&tq->tq_lock);
2068 		return;
2069 	}
2070 
2071 	ASSERT(tqe->tqent_thread == NULL);
2072 
2073 	tqe->tqent_bucket = b;
2074 
2075 	/*
2076 	 * Create a thread in a TS_STOPPED state first. If it is successfully
2077 	 * created, place the entry on the free list and start the thread.
2078 	 */
2079 	tqe->tqent_thread = thread_create(NULL, 0, taskq_d_thread, tqe,
2080 	    0, &p0, TS_STOPPED, tq->tq_pri);
2081 
2082 	/*
2083 	 * Once the entry is ready, link it to the the bucket free list.
2084 	 */
2085 	mutex_enter(&b->tqbucket_lock);
2086 	tqe->tqent_func = NULL;
2087 	TQ_APPEND(b->tqbucket_freelist, tqe);
2088 	b->tqbucket_nfree++;
2089 	TQ_STAT(b, tqs_tcreates);
2090 
2091 #if TASKQ_STATISTIC
2092 	nthreads = b->tqbucket_stat.tqs_tcreates -
2093 	    b->tqbucket_stat.tqs_tdeaths;
2094 	b->tqbucket_stat.tqs_maxthreads = MAX(nthreads,
2095 	    b->tqbucket_stat.tqs_maxthreads);
2096 #endif
2097 
2098 	mutex_exit(&b->tqbucket_lock);
2099 	/*
2100 	 * Start the stopped thread.
2101 	 */
2102 	thread_lock(tqe->tqent_thread);
2103 	tqe->tqent_thread->t_taskq = tq;
2104 	tqe->tqent_thread->t_schedflag |= TS_ALLSTART;
2105 	setrun_locked(tqe->tqent_thread);
2106 	thread_unlock(tqe->tqent_thread);
2107 }
2108 
2109 static int
2110 taskq_kstat_update(kstat_t *ksp, int rw)
2111 {
2112 	struct taskq_kstat *tqsp = &taskq_kstat;
2113 	taskq_t *tq = ksp->ks_private;
2114 
2115 	if (rw == KSTAT_WRITE)
2116 		return (EACCES);
2117 
2118 	tqsp->tq_pid.value.ui64 = tq->tq_proc->p_pid;
2119 	tqsp->tq_tasks.value.ui64 = tq->tq_tasks;
2120 	tqsp->tq_executed.value.ui64 = tq->tq_executed;
2121 	tqsp->tq_maxtasks.value.ui64 = tq->tq_maxtasks;
2122 	tqsp->tq_totaltime.value.ui64 = tq->tq_totaltime;
2123 	tqsp->tq_nactive.value.ui64 = tq->tq_active;
2124 	tqsp->tq_nalloc.value.ui64 = tq->tq_nalloc;
2125 	tqsp->tq_pri.value.ui64 = tq->tq_pri;
2126 	tqsp->tq_nthreads.value.ui64 = tq->tq_nthreads;
2127 	return (0);
2128 }
2129 
2130 static int
2131 taskq_d_kstat_update(kstat_t *ksp, int rw)
2132 {
2133 	struct taskq_d_kstat *tqsp = &taskq_d_kstat;
2134 	taskq_t *tq = ksp->ks_private;
2135 	taskq_bucket_t *b = tq->tq_buckets;
2136 	int bid = 0;
2137 
2138 	if (rw == KSTAT_WRITE)
2139 		return (EACCES);
2140 
2141 	ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2142 
2143 	tqsp->tqd_btasks.value.ui64 = tq->tq_tasks;
2144 	tqsp->tqd_bexecuted.value.ui64 = tq->tq_executed;
2145 	tqsp->tqd_bmaxtasks.value.ui64 = tq->tq_maxtasks;
2146 	tqsp->tqd_bnalloc.value.ui64 = tq->tq_nalloc;
2147 	tqsp->tqd_bnactive.value.ui64 = tq->tq_active;
2148 	tqsp->tqd_btotaltime.value.ui64 = tq->tq_totaltime;
2149 	tqsp->tqd_pri.value.ui64 = tq->tq_pri;
2150 
2151 	tqsp->tqd_hits.value.ui64 = 0;
2152 	tqsp->tqd_misses.value.ui64 = 0;
2153 	tqsp->tqd_overflows.value.ui64 = 0;
2154 	tqsp->tqd_tcreates.value.ui64 = 0;
2155 	tqsp->tqd_tdeaths.value.ui64 = 0;
2156 	tqsp->tqd_maxthreads.value.ui64 = 0;
2157 	tqsp->tqd_nomem.value.ui64 = 0;
2158 	tqsp->tqd_disptcreates.value.ui64 = 0;
2159 	tqsp->tqd_totaltime.value.ui64 = 0;
2160 	tqsp->tqd_nalloc.value.ui64 = 0;
2161 	tqsp->tqd_nfree.value.ui64 = 0;
2162 
2163 	for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2164 		tqsp->tqd_hits.value.ui64 += b->tqbucket_stat.tqs_hits;
2165 		tqsp->tqd_misses.value.ui64 += b->tqbucket_stat.tqs_misses;
2166 		tqsp->tqd_overflows.value.ui64 += b->tqbucket_stat.tqs_overflow;
2167 		tqsp->tqd_tcreates.value.ui64 += b->tqbucket_stat.tqs_tcreates;
2168 		tqsp->tqd_tdeaths.value.ui64 += b->tqbucket_stat.tqs_tdeaths;
2169 		tqsp->tqd_maxthreads.value.ui64 +=
2170 		    b->tqbucket_stat.tqs_maxthreads;
2171 		tqsp->tqd_nomem.value.ui64 += b->tqbucket_stat.tqs_nomem;
2172 		tqsp->tqd_disptcreates.value.ui64 +=
2173 		    b->tqbucket_stat.tqs_disptcreates;
2174 		tqsp->tqd_totaltime.value.ui64 += b->tqbucket_totaltime;
2175 		tqsp->tqd_nalloc.value.ui64 += b->tqbucket_nalloc;
2176 		tqsp->tqd_nfree.value.ui64 += b->tqbucket_nfree;
2177 	}
2178 	return (0);
2179 }
2180