1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 /*
27 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
28 * Copyright (c) 2017 by Delphix. All rights reserved.
29 */
30
31 /*
32 * Kernel task queues: general-purpose asynchronous task scheduling.
33 *
34 * A common problem in kernel programming is the need to schedule tasks
35 * to be performed later, by another thread. There are several reasons
36 * you may want or need to do this:
37 *
38 * (1) The task isn't time-critical, but your current code path is.
39 *
40 * (2) The task may require grabbing locks that you already hold.
41 *
42 * (3) The task may need to block (e.g. to wait for memory), but you
43 * cannot block in your current context.
44 *
45 * (4) Your code path can't complete because of some condition, but you can't
46 * sleep or fail, so you queue the task for later execution when condition
47 * disappears.
48 *
49 * (5) You just want a simple way to launch multiple tasks in parallel.
50 *
51 * Task queues provide such a facility. In its simplest form (used when
52 * performance is not a critical consideration) a task queue consists of a
53 * single list of tasks, together with one or more threads to service the
54 * list. There are some cases when this simple queue is not sufficient:
55 *
56 * (1) The task queues are very hot and there is a need to avoid data and lock
57 * contention over global resources.
58 *
59 * (2) Some tasks may depend on other tasks to complete, so they can't be put in
60 * the same list managed by the same thread.
61 *
62 * (3) Some tasks may block for a long time, and this should not block other
63 * tasks in the queue.
64 *
65 * To provide useful service in such cases we define a "dynamic task queue"
66 * which has an individual thread for each of the tasks. These threads are
67 * dynamically created as they are needed and destroyed when they are not in
68 * use. The API for managing task pools is the same as for managing task queues
69 * with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
70 * dynamic task pool behavior is desired.
71 *
72 * Dynamic task queues may also place tasks in the normal queue (called "backing
73 * queue") when task pool runs out of resources. Users of task queues may
74 * disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
75 * flags.
76 *
77 * The backing task queue is also used for scheduling internal tasks needed for
78 * dynamic task queue maintenance.
79 *
80 * INTERFACES ==================================================================
81 *
82 * taskq_t *taskq_create(name, nthreads, pri, minalloc, maxalloc, flags);
83 *
84 * Create a taskq with specified properties.
85 * Possible 'flags':
86 *
87 * TASKQ_DYNAMIC: Create task pool for task management. If this flag is
88 * specified, 'nthreads' specifies the maximum number of threads in
89 * the task queue. Task execution order for dynamic task queues is
90 * not predictable.
91 *
92 * If this flag is not specified (default case) a
93 * single-list task queue is created with 'nthreads' threads
94 * servicing it. Entries in this queue are managed by
95 * taskq_ent_alloc() and taskq_ent_free() which try to keep the
96 * task population between 'minalloc' and 'maxalloc', but the
97 * latter limit is only advisory for TQ_SLEEP dispatches and the
98 * former limit is only advisory for TQ_NOALLOC dispatches. If
99 * TASKQ_PREPOPULATE is set in 'flags', the taskq will be
100 * prepopulated with 'minalloc' task structures.
101 *
102 * Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
103 * executed in the order they are scheduled if nthreads == 1.
104 * If nthreads > 1, task execution order is not predictable.
105 *
106 * TASKQ_PREPOPULATE: Prepopulate task queue with threads.
107 * Also prepopulate the task queue with 'minalloc' task structures.
108 *
109 * TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
110 * interpreted as a percentage of the # of online CPUs on the
111 * system. The taskq subsystem will automatically adjust the
112 * number of threads in the taskq in response to CPU online
113 * and offline events, to keep the ratio. nthreads must be in
114 * the range [0,100].
115 *
116 * The calculation used is:
117 *
118 * MAX((ncpus_online * percentage)/100, 1)
119 *
120 * This flag is not supported for DYNAMIC task queues.
121 * This flag is not compatible with TASKQ_CPR_SAFE.
122 *
123 * TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
124 * use their own protocol for handling CPR issues. This flag is not
125 * supported for DYNAMIC task queues. This flag is not compatible
126 * with TASKQ_THREADS_CPU_PCT.
127 *
128 * The 'pri' field specifies the default priority for the threads that
129 * service all scheduled tasks.
130 *
131 * taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
132 * maxalloc, flags);
133 *
134 * Like taskq_create(), but takes an instance number (or -1 to indicate
135 * no instance).
136 *
137 * taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxalloc, proc,
138 * flags);
139 *
140 * Like taskq_create(), but creates the taskq threads in the specified
141 * system process. If proc != &p0, this must be called from a thread
142 * in that process.
143 *
144 * taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxalloc, proc,
145 * dc, flags);
146 *
147 * Like taskq_create_proc(), but the taskq threads will use the
148 * System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
149 *
150 * void taskq_destroy(tap):
151 *
152 * Waits for any scheduled tasks to complete, then destroys the taskq.
153 * Caller should guarantee that no new tasks are scheduled in the closing
154 * taskq.
155 *
156 * taskqid_t taskq_dispatch(tq, func, arg, flags):
157 *
158 * Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
159 * the caller is willing to block for memory. The function returns an
160 * opaque value which is zero iff dispatch fails. If flags is TQ_NOSLEEP
161 * or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
162 * and returns (taskqid_t)0.
163 *
164 * ASSUMES: func != NULL.
165 *
166 * Possible flags:
167 * TQ_NOSLEEP: Do not wait for resources; may fail.
168 *
169 * TQ_NOALLOC: Do not allocate memory; may fail. May only be used with
170 * non-dynamic task queues.
171 *
172 * TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
173 * lack of available resources and fail. If this flag is not
174 * set, and the task pool is exhausted, the task may be scheduled
175 * in the backing queue. This flag may ONLY be used with dynamic
176 * task queues.
177 *
178 * NOTE: This flag should always be used when a task queue is used
179 * for tasks that may depend on each other for completion.
180 * Enqueueing dependent tasks may create deadlocks.
181 *
182 * TQ_SLEEP: May block waiting for resources. May still fail for
183 * dynamic task queues if TQ_NOQUEUE is also specified, otherwise
184 * always succeed.
185 *
186 * TQ_FRONT: Puts the new task at the front of the queue. Be careful.
187 *
188 * NOTE: Dynamic task queues are much more likely to fail in
189 * taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
190 * is important to have backup strategies handling such failures.
191 *
192 * void taskq_dispatch_ent(tq, func, arg, flags, tqent)
193 *
194 * This is a light-weight form of taskq_dispatch(), that uses a
195 * preallocated taskq_ent_t structure for scheduling. As a
196 * result, it does not perform allocations and cannot ever fail.
197 * Note especially that it cannot be used with TASKQ_DYNAMIC
198 * taskqs. The memory for the tqent must not be modified or used
199 * until the function (func) is called. (However, func itself
200 * may safely modify or free this memory, once it is called.)
201 * Note that the taskq framework will NOT free this memory.
202 *
203 * void taskq_wait(tq):
204 *
205 * Waits for all previously scheduled tasks to complete.
206 *
207 * NOTE: It does not stop any new task dispatches.
208 * Do NOT call taskq_wait() from a task: it will cause deadlock.
209 *
210 * void taskq_suspend(tq)
211 *
212 * Suspend all task execution. Tasks already scheduled for a dynamic task
213 * queue will still be executed, but all new scheduled tasks will be
214 * suspended until taskq_resume() is called.
215 *
216 * int taskq_suspended(tq)
217 *
218 * Returns 1 if taskq is suspended and 0 otherwise. It is intended to
219 * ASSERT that the task queue is suspended.
220 *
221 * void taskq_resume(tq)
222 *
223 * Resume task queue execution.
224 *
225 * int taskq_member(tq, thread)
226 *
227 * Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
228 * intended use is to ASSERT that a given function is called in taskq
229 * context only.
230 *
231 * system_taskq
232 *
233 * Global system-wide dynamic task queue for common uses. It may be used by
234 * any subsystem that needs to schedule tasks and does not need to manage
235 * its own task queues. It is initialized quite early during system boot.
236 *
237 * IMPLEMENTATION ==============================================================
238 *
239 * This is schematic representation of the task queue structures.
240 *
241 * taskq:
242 * +-------------+
243 * | tq_lock | +---< taskq_ent_free()
244 * +-------------+ |
245 * |... | | tqent: tqent:
246 * +-------------+ | +------------+ +------------+
247 * | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
248 * +-------------+ +------------+ +------------+
249 * |... | | ... | | ... |
250 * +-------------+ +------------+ +------------+
251 * | tq_task | |
252 * | | +-------------->taskq_ent_alloc()
253 * +--------------------------------------------------------------------------+
254 * | | | tqent tqent |
255 * | +---------------------+ +--> +------------+ +--> +------------+ |
256 * | | ... | | | func, arg | | | func, arg | |
257 * +>+---------------------+ <---|-+ +------------+ <---|-+ +------------+ |
258 * | tq_taskq.tqent_next | ----+ | | tqent_next | --->+ | | tqent_next |--+
259 * +---------------------+ | +------------+ ^ | +------------+
260 * +-| tq_task.tqent_prev | +--| tqent_prev | | +--| tqent_prev | ^
261 * | +---------------------+ +------------+ | +------------+ |
262 * | |... | | ... | | | ... | |
263 * | +---------------------+ +------------+ | +------------+ |
264 * | ^ | |
265 * | | | |
266 * +--------------------------------------+--------------+ TQ_APPEND() -+
267 * | | |
268 * |... | taskq_thread()-----+
269 * +-------------+
270 * | tq_buckets |--+-------> [ NULL ] (for regular task queues)
271 * +-------------+ |
272 * | DYNAMIC TASK QUEUES:
273 * |
274 * +-> taskq_bucket[nCPU] taskq_bucket_dispatch()
275 * +-------------------+ ^
276 * +--->| tqbucket_lock | |
277 * | +-------------------+ +--------+ +--------+
278 * | | tqbucket_freelist |-->| tqent |-->...| tqent | ^
279 * | +-------------------+<--+--------+<--...+--------+ |
280 * | | ... | | thread | | thread | |
281 * | +-------------------+ +--------+ +--------+ |
282 * | +-------------------+ |
283 * taskq_dispatch()--+--->| tqbucket_lock | TQ_APPEND()------+
284 * TQ_HASH() | +-------------------+ +--------+ +--------+
285 * | | tqbucket_freelist |-->| tqent |-->...| tqent |
286 * | +-------------------+<--+--------+<--...+--------+
287 * | | ... | | thread | | thread |
288 * | +-------------------+ +--------+ +--------+
289 * +---> ...
290 *
291 *
292 * Task queues use tq_task field to link new entry in the queue. The queue is a
293 * circular doubly-linked list. Entries are put in the end of the list with
294 * TQ_APPEND() and processed from the front of the list by taskq_thread() in
295 * FIFO order. Task queue entries are cached in the free list managed by
296 * taskq_ent_alloc() and taskq_ent_free() functions.
297 *
298 * All threads used by task queues mark t_taskq field of the thread to
299 * point to the task queue.
300 *
301 * Taskq Thread Management -----------------------------------------------------
302 *
303 * Taskq's non-dynamic threads are managed with several variables and flags:
304 *
305 * * tq_nthreads - The number of threads in taskq_thread() for the
306 * taskq.
307 *
308 * * tq_active - The number of threads not waiting on a CV in
309 * taskq_thread(); includes newly created threads
310 * not yet counted in tq_nthreads.
311 *
312 * * tq_nthreads_target
313 * - The number of threads desired for the taskq.
314 *
315 * * tq_flags & TASKQ_CHANGING
316 * - Indicates that tq_nthreads != tq_nthreads_target.
317 *
318 * * tq_flags & TASKQ_THREAD_CREATED
319 * - Indicates that a thread is being created in the taskq.
320 *
321 * During creation, tq_nthreads and tq_active are set to 0, and
322 * tq_nthreads_target is set to the number of threads desired. The
323 * TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
324 * create the first thread. taskq_thread_create() increments tq_active,
325 * sets TASKQ_THREAD_CREATED, and creates the new thread.
326 *
327 * Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
328 * flag, and increments tq_nthreads. It stores the new value of
329 * tq_nthreads as its "thread_id", and stores its thread pointer in the
330 * tq_threadlist at the (thread_id - 1). We keep the thread_id space
331 * densely packed by requiring that only the largest thread_id can exit during
332 * normal adjustment. The exception is during the destruction of the
333 * taskq; once tq_nthreads_target is set to zero, no new threads will be created
334 * for the taskq queue, so every thread can exit without any ordering being
335 * necessary.
336 *
337 * Threads will only process work if their thread id is <= tq_nthreads_target.
338 *
339 * When TASKQ_CHANGING is set, threads will check the current thread target
340 * whenever they wake up, and do whatever they can to apply its effects.
341 *
342 * TASKQ_THREAD_CPU_PCT --------------------------------------------------------
343 *
344 * When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
345 * percentage in tq_threads_ncpus_pct, start them off with the correct thread
346 * target, and add them to the taskq_cpupct_list for later adjustment.
347 *
348 * We register taskq_cpu_setup() to be called whenever a CPU changes state. It
349 * walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
350 * if need be, and wakes up all of the threads to process the change.
351 *
352 * Dynamic Task Queues Implementation ------------------------------------------
353 *
354 * For a dynamic task queues there is a 1-to-1 mapping between a thread and
355 * taskq_ent_structure. Each entry is serviced by its own thread and each thread
356 * is controlled by a single entry.
357 *
358 * Entries are distributed over a set of buckets. To avoid using modulo
359 * arithmetics the number of buckets is 2^n and is determined as the nearest
360 * power of two roundown of the number of CPUs in the system. Tunable
361 * variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
362 * is attached to a bucket for its lifetime and can't migrate to other buckets.
363 *
364 * Entries that have scheduled tasks are not placed in any list. The dispatch
365 * function sets their "func" and "arg" fields and signals the corresponding
366 * thread to execute the task. Once the thread executes the task it clears the
367 * "func" field and places an entry on the bucket cache of free entries pointed
368 * by "tqbucket_freelist" field. ALL entries on the free list should have "func"
369 * field equal to NULL. The free list is a circular doubly-linked list identical
370 * in structure to the tq_task list above, but entries are taken from it in LIFO
371 * order - the last freed entry is the first to be allocated. The
372 * taskq_bucket_dispatch() function gets the most recently used entry from the
373 * free list, sets its "func" and "arg" fields and signals a worker thread.
374 *
375 * After executing each task a per-entry thread taskq_d_thread() places its
376 * entry on the bucket free list and goes to a timed sleep. If it wakes up
377 * without getting new task it removes the entry from the free list and destroys
378 * itself. The thread sleep time is controlled by a tunable variable
379 * `taskq_thread_timeout'.
380 *
381 * There are various statistics kept in the bucket which allows for later
382 * analysis of taskq usage patterns. Also, a global copy of taskq creation and
383 * death statistics is kept in the global taskq data structure. Since thread
384 * creation and death happen rarely, updating such global data does not present
385 * a performance problem.
386 *
387 * NOTE: Threads are not bound to any CPU and there is absolutely no association
388 * between the bucket and actual thread CPU, so buckets are used only to
389 * split resources and reduce resource contention. Having threads attached
390 * to the CPU denoted by a bucket may reduce number of times the job
391 * switches between CPUs.
392 *
393 * Current algorithm creates a thread whenever a bucket has no free
394 * entries. It would be nice to know how many threads are in the running
395 * state and don't create threads if all CPUs are busy with existing
396 * tasks, but it is unclear how such strategy can be implemented.
397 *
398 * Currently buckets are created statically as an array attached to task
399 * queue. On some system with nCPUs < max_ncpus it may waste system
400 * memory. One solution may be allocation of buckets when they are first
401 * touched, but it is not clear how useful it is.
402 *
403 * SUSPEND/RESUME implementation -----------------------------------------------
404 *
405 * Before executing a task taskq_thread() (executing non-dynamic task
406 * queues) obtains taskq's thread lock as a reader. The taskq_suspend()
407 * function gets the same lock as a writer blocking all non-dynamic task
408 * execution. The taskq_resume() function releases the lock allowing
409 * taskq_thread to continue execution.
410 *
411 * For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
412 * taskq_suspend() function. After that taskq_bucket_dispatch() always
413 * fails, so that taskq_dispatch() will either enqueue tasks for a
414 * suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
415 * flags.
416 *
417 * NOTE: taskq_suspend() does not immediately block any tasks already
418 * scheduled for dynamic task queues. It only suspends new tasks
419 * scheduled after taskq_suspend() was called.
420 *
421 * taskq_member() function works by comparing a thread t_taskq pointer with
422 * the passed thread pointer.
423 *
424 * LOCKS and LOCK Hierarchy ----------------------------------------------------
425 *
426 * There are three locks used in task queues:
427 *
428 * 1) The taskq_t's tq_lock, protecting global task queue state.
429 *
430 * 2) Each per-CPU bucket has a lock for bucket management.
431 *
432 * 3) The global taskq_cpupct_lock, which protects the list of
433 * TASKQ_THREADS_CPU_PCT taskqs.
434 *
435 * If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
436 * lock.
437 *
438 * If both (1) and (3) are needed, tq_lock should be taken *after*
439 * taskq_cpupct_lock.
440 *
441 * DEBUG FACILITIES ------------------------------------------------------------
442 *
443 * For DEBUG kernels it is possible to induce random failures to
444 * taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
445 * taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
446 * failures for dynamic and static task queues respectively.
447 *
448 * Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
449 *
450 * TUNABLES --------------------------------------------------------------------
451 *
452 * system_taskq_size - Size of the global system_taskq.
453 * This value is multiplied by nCPUs to determine
454 * actual size.
455 * Default value: 64
456 *
457 * taskq_minimum_nthreads_max
458 * - Minimum size of the thread list for a taskq.
459 * Useful for testing different thread pool
460 * sizes by overwriting tq_nthreads_target.
461 *
462 * taskq_thread_timeout - Maximum idle time for taskq_d_thread()
463 * Default value: 5 minutes
464 *
465 * taskq_maxbuckets - Maximum number of buckets in any task queue
466 * Default value: 128
467 *
468 * taskq_search_depth - Maximum # of buckets searched for a free entry
469 * Default value: 4
470 *
471 * taskq_dmtbf - Mean time between induced dispatch failures
472 * for dynamic task queues.
473 * Default value: UINT_MAX (no induced failures)
474 *
475 * taskq_smtbf - Mean time between induced dispatch failures
476 * for static task queues.
477 * Default value: UINT_MAX (no induced failures)
478 *
479 * CONDITIONAL compilation -----------------------------------------------------
480 *
481 * TASKQ_STATISTIC - If set will enable bucket statistic (default).
482 *
483 */
484
485 #include <sys/taskq_impl.h>
486 #include <sys/thread.h>
487 #include <sys/proc.h>
488 #include <sys/kmem.h>
489 #include <sys/vmem.h>
490 #include <sys/callb.h>
491 #include <sys/class.h>
492 #include <sys/systm.h>
493 #include <sys/cmn_err.h>
494 #include <sys/debug.h>
495 #include <sys/vmsystm.h> /* For throttlefree */
496 #include <sys/sysmacros.h>
497 #include <sys/cpuvar.h>
498 #include <sys/cpupart.h>
499 #include <sys/sdt.h>
500 #include <sys/sysdc.h>
501 #include <sys/note.h>
502
503 static kmem_cache_t *taskq_ent_cache, *taskq_cache;
504
505 /*
506 * Pseudo instance numbers for taskqs without explicitly provided instance.
507 */
508 static vmem_t *taskq_id_arena;
509
510 /* Global system task queue for common use */
511 taskq_t *system_taskq;
512
513 /*
514 * Maximum number of entries in global system taskq is
515 * system_taskq_size * max_ncpus
516 */
517 #define SYSTEM_TASKQ_SIZE 64
518 int system_taskq_size = SYSTEM_TASKQ_SIZE;
519
520 /*
521 * Minimum size for tq_nthreads_max; useful for those who want to play around
522 * with increasing a taskq's tq_nthreads_target.
523 */
524 int taskq_minimum_nthreads_max = 1;
525
526 /*
527 * We want to ensure that when taskq_create() returns, there is at least
528 * one thread ready to handle requests. To guarantee this, we have to wait
529 * for the second thread, since the first one cannot process requests until
530 * the second thread has been created.
531 */
532 #define TASKQ_CREATE_ACTIVE_THREADS 2
533
534 /* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
535 #define TASKQ_CPUPCT_MAX_PERCENT 1000
536 int taskq_cpupct_max_percent = TASKQ_CPUPCT_MAX_PERCENT;
537
538 /*
539 * Dynamic task queue threads that don't get any work within
540 * taskq_thread_timeout destroy themselves
541 */
542 #define TASKQ_THREAD_TIMEOUT (60 * 5)
543 int taskq_thread_timeout = TASKQ_THREAD_TIMEOUT;
544
545 #define TASKQ_MAXBUCKETS 128
546 int taskq_maxbuckets = TASKQ_MAXBUCKETS;
547
548 /*
549 * When a bucket has no available entries another buckets are tried.
550 * taskq_search_depth parameter limits the amount of buckets that we search
551 * before failing. This is mostly useful in systems with many CPUs where we may
552 * spend too much time scanning busy buckets.
553 */
554 #define TASKQ_SEARCH_DEPTH 4
555 int taskq_search_depth = TASKQ_SEARCH_DEPTH;
556
557 /*
558 * Hashing function: mix various bits of x. May be pretty much anything.
559 */
560 #define TQ_HASH(x) ((x) ^ ((x) >> 11) ^ ((x) >> 17) ^ ((x) ^ 27))
561
562 /*
563 * We do not create any new threads when the system is low on memory and start
564 * throttling memory allocations. The following macro tries to estimate such
565 * condition.
566 */
567 #define ENOUGH_MEMORY() (freemem > throttlefree)
568
569 /*
570 * Static functions.
571 */
572 static taskq_t *taskq_create_common(const char *, int, int, pri_t, int,
573 int, proc_t *, uint_t, uint_t);
574 static void taskq_thread(void *);
575 static void taskq_d_thread(taskq_ent_t *);
576 static void taskq_bucket_extend(void *);
577 static int taskq_constructor(void *, void *, int);
578 static void taskq_destructor(void *, void *);
579 static int taskq_ent_constructor(void *, void *, int);
580 static void taskq_ent_destructor(void *, void *);
581 static taskq_ent_t *taskq_ent_alloc(taskq_t *, int);
582 static void taskq_ent_free(taskq_t *, taskq_ent_t *);
583 static int taskq_ent_exists(taskq_t *, task_func_t, void *);
584 static taskq_ent_t *taskq_bucket_dispatch(taskq_bucket_t *, task_func_t,
585 void *);
586
587 /*
588 * Task queues kstats.
589 */
590 struct taskq_kstat {
591 kstat_named_t tq_pid;
592 kstat_named_t tq_tasks;
593 kstat_named_t tq_executed;
594 kstat_named_t tq_maxtasks;
595 kstat_named_t tq_totaltime;
596 kstat_named_t tq_nalloc;
597 kstat_named_t tq_nactive;
598 kstat_named_t tq_pri;
599 kstat_named_t tq_nthreads;
600 kstat_named_t tq_nomem;
601 } taskq_kstat = {
602 { "pid", KSTAT_DATA_UINT64 },
603 { "tasks", KSTAT_DATA_UINT64 },
604 { "executed", KSTAT_DATA_UINT64 },
605 { "maxtasks", KSTAT_DATA_UINT64 },
606 { "totaltime", KSTAT_DATA_UINT64 },
607 { "nalloc", KSTAT_DATA_UINT64 },
608 { "nactive", KSTAT_DATA_UINT64 },
609 { "priority", KSTAT_DATA_UINT64 },
610 { "threads", KSTAT_DATA_UINT64 },
611 { "nomem", KSTAT_DATA_UINT64 },
612 };
613
614 struct taskq_d_kstat {
615 kstat_named_t tqd_pri;
616 kstat_named_t tqd_btasks;
617 kstat_named_t tqd_bexecuted;
618 kstat_named_t tqd_bmaxtasks;
619 kstat_named_t tqd_bnalloc;
620 kstat_named_t tqd_bnactive;
621 kstat_named_t tqd_btotaltime;
622 kstat_named_t tqd_hits;
623 kstat_named_t tqd_misses;
624 kstat_named_t tqd_overflows;
625 kstat_named_t tqd_tcreates;
626 kstat_named_t tqd_tdeaths;
627 kstat_named_t tqd_maxthreads;
628 kstat_named_t tqd_nomem;
629 kstat_named_t tqd_disptcreates;
630 kstat_named_t tqd_totaltime;
631 kstat_named_t tqd_nalloc;
632 kstat_named_t tqd_nfree;
633 } taskq_d_kstat = {
634 { "priority", KSTAT_DATA_UINT64 },
635 { "btasks", KSTAT_DATA_UINT64 },
636 { "bexecuted", KSTAT_DATA_UINT64 },
637 { "bmaxtasks", KSTAT_DATA_UINT64 },
638 { "bnalloc", KSTAT_DATA_UINT64 },
639 { "bnactive", KSTAT_DATA_UINT64 },
640 { "btotaltime", KSTAT_DATA_UINT64 },
641 { "hits", KSTAT_DATA_UINT64 },
642 { "misses", KSTAT_DATA_UINT64 },
643 { "overflows", KSTAT_DATA_UINT64 },
644 { "tcreates", KSTAT_DATA_UINT64 },
645 { "tdeaths", KSTAT_DATA_UINT64 },
646 { "maxthreads", KSTAT_DATA_UINT64 },
647 { "nomem", KSTAT_DATA_UINT64 },
648 { "disptcreates", KSTAT_DATA_UINT64 },
649 { "totaltime", KSTAT_DATA_UINT64 },
650 { "nalloc", KSTAT_DATA_UINT64 },
651 { "nfree", KSTAT_DATA_UINT64 },
652 };
653
654 static kmutex_t taskq_kstat_lock;
655 static kmutex_t taskq_d_kstat_lock;
656 static int taskq_kstat_update(kstat_t *, int);
657 static int taskq_d_kstat_update(kstat_t *, int);
658
659 /*
660 * List of all TASKQ_THREADS_CPU_PCT taskqs.
661 */
662 static list_t taskq_cpupct_list; /* protected by cpu_lock */
663
664 /*
665 * Collect per-bucket statistic when TASKQ_STATISTIC is defined.
666 */
667 #define TASKQ_STATISTIC 1
668
669 #if TASKQ_STATISTIC
670 #define TQ_STAT(b, x) b->tqbucket_stat.x++
671 #else
672 #define TQ_STAT(b, x)
673 #endif
674
675 /*
676 * Random fault injection.
677 */
678 uint_t taskq_random;
679 uint_t taskq_dmtbf = UINT_MAX; /* mean time between injected failures */
680 uint_t taskq_smtbf = UINT_MAX; /* mean time between injected failures */
681
682 /*
683 * TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
684 *
685 * TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
686 * they could prepopulate the cache and make sure that they do not use more
687 * then minalloc entries. So, fault injection in this case insures that
688 * either TASKQ_PREPOPULATE is not set or there are more entries allocated
689 * than is specified by minalloc. TQ_NOALLOC dispatches are always allowed
690 * to fail, but for simplicity we treat them identically to TQ_NOSLEEP
691 * dispatches.
692 */
693 #ifdef DEBUG
694 #define TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag) \
695 taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
696 if ((flag & TQ_NOSLEEP) && \
697 taskq_random < 1771875 / taskq_dmtbf) { \
698 return (NULL); \
699 }
700
701 #define TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag) \
702 taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
703 if ((flag & (TQ_NOSLEEP | TQ_NOALLOC)) && \
704 (!(tq->tq_flags & TASKQ_PREPOPULATE) || \
705 (tq->tq_nalloc > tq->tq_minalloc)) && \
706 (taskq_random < (1771875 / taskq_smtbf))) { \
707 mutex_exit(&tq->tq_lock); \
708 return (NULL); \
709 }
710 #else
711 #define TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)
712 #define TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)
713 #endif
714
715 #define IS_EMPTY(l) (((l).tqent_prev == (l).tqent_next) && \
716 ((l).tqent_prev == &(l)))
717
718 /*
719 * Append `tqe' in the end of the doubly-linked list denoted by l.
720 */
721 #define TQ_APPEND(l, tqe) { \
722 tqe->tqent_next = &l; \
723 tqe->tqent_prev = l.tqent_prev; \
724 tqe->tqent_next->tqent_prev = tqe; \
725 tqe->tqent_prev->tqent_next = tqe; \
726 }
727 /*
728 * Prepend 'tqe' to the beginning of l
729 */
730 #define TQ_PREPEND(l, tqe) { \
731 tqe->tqent_next = l.tqent_next; \
732 tqe->tqent_prev = &l; \
733 tqe->tqent_next->tqent_prev = tqe; \
734 tqe->tqent_prev->tqent_next = tqe; \
735 }
736
737 /*
738 * Schedule a task specified by func and arg into the task queue entry tqe.
739 */
740 #define TQ_DO_ENQUEUE(tq, tqe, func, arg, front) { \
741 ASSERT(MUTEX_HELD(&tq->tq_lock)); \
742 _NOTE(CONSTCOND) \
743 if (front) { \
744 TQ_PREPEND(tq->tq_task, tqe); \
745 } else { \
746 TQ_APPEND(tq->tq_task, tqe); \
747 } \
748 tqe->tqent_func = (func); \
749 tqe->tqent_arg = (arg); \
750 tq->tq_tasks++; \
751 if (tq->tq_tasks - tq->tq_executed > tq->tq_maxtasks) \
752 tq->tq_maxtasks = tq->tq_tasks - tq->tq_executed; \
753 cv_signal(&tq->tq_dispatch_cv); \
754 DTRACE_PROBE2(taskq__enqueue, taskq_t *, tq, taskq_ent_t *, tqe); \
755 }
756
757 #define TQ_ENQUEUE(tq, tqe, func, arg) \
758 TQ_DO_ENQUEUE(tq, tqe, func, arg, 0)
759
760 #define TQ_ENQUEUE_FRONT(tq, tqe, func, arg) \
761 TQ_DO_ENQUEUE(tq, tqe, func, arg, 1)
762
763 /*
764 * Do-nothing task which may be used to prepopulate thread caches.
765 */
766 /*ARGSUSED*/
767 void
768 nulltask(void *unused)
769 {
770 }
771
772 /*ARGSUSED*/
773 static int
774 taskq_constructor(void *buf, void *cdrarg, int kmflags)
775 {
776 taskq_t *tq = buf;
777
778 bzero(tq, sizeof (taskq_t));
779
780 mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
781 rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
782 cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
783 cv_init(&tq->tq_exit_cv, NULL, CV_DEFAULT, NULL);
784 cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
785 cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
786
787 tq->tq_task.tqent_next = &tq->tq_task;
788 tq->tq_task.tqent_prev = &tq->tq_task;
789
790 return (0);
791 }
792
793 /*ARGSUSED*/
794 static void
795 taskq_destructor(void *buf, void *cdrarg)
796 {
797 taskq_t *tq = buf;
798
799 ASSERT(tq->tq_nthreads == 0);
800 ASSERT(tq->tq_buckets == NULL);
801 ASSERT(tq->tq_tcreates == 0);
802 ASSERT(tq->tq_tdeaths == 0);
803
804 mutex_destroy(&tq->tq_lock);
805 rw_destroy(&tq->tq_threadlock);
806 cv_destroy(&tq->tq_dispatch_cv);
807 cv_destroy(&tq->tq_exit_cv);
808 cv_destroy(&tq->tq_wait_cv);
809 cv_destroy(&tq->tq_maxalloc_cv);
810 }
811
812 /*ARGSUSED*/
813 static int
814 taskq_ent_constructor(void *buf, void *cdrarg, int kmflags)
815 {
816 taskq_ent_t *tqe = buf;
817
818 tqe->tqent_thread = NULL;
819 cv_init(&tqe->tqent_cv, NULL, CV_DEFAULT, NULL);
820
821 return (0);
822 }
823
824 /*ARGSUSED*/
825 static void
826 taskq_ent_destructor(void *buf, void *cdrarg)
827 {
828 taskq_ent_t *tqe = buf;
829
830 ASSERT(tqe->tqent_thread == NULL);
831 cv_destroy(&tqe->tqent_cv);
832 }
833
834 void
835 taskq_init(void)
836 {
837 taskq_ent_cache = kmem_cache_create("taskq_ent_cache",
838 sizeof (taskq_ent_t), 0, taskq_ent_constructor,
839 taskq_ent_destructor, NULL, NULL, NULL, 0);
840 taskq_cache = kmem_cache_create("taskq_cache", sizeof (taskq_t),
841 0, taskq_constructor, taskq_destructor, NULL, NULL, NULL, 0);
842 taskq_id_arena = vmem_create("taskq_id_arena",
843 (void *)1, INT32_MAX, 1, NULL, NULL, NULL, 0,
844 VM_SLEEP | VMC_IDENTIFIER);
845
846 list_create(&taskq_cpupct_list, sizeof (taskq_t),
847 offsetof(taskq_t, tq_cpupct_link));
848 }
849
850 static void
851 taskq_update_nthreads(taskq_t *tq, uint_t ncpus)
852 {
853 uint_t newtarget = TASKQ_THREADS_PCT(ncpus, tq->tq_threads_ncpus_pct);
854
855 ASSERT(MUTEX_HELD(&cpu_lock));
856 ASSERT(MUTEX_HELD(&tq->tq_lock));
857
858 /* We must be going from non-zero to non-zero; no exiting. */
859 ASSERT3U(tq->tq_nthreads_target, !=, 0);
860 ASSERT3U(newtarget, !=, 0);
861
862 ASSERT3U(newtarget, <=, tq->tq_nthreads_max);
863 if (newtarget != tq->tq_nthreads_target) {
864 tq->tq_flags |= TASKQ_CHANGING;
865 tq->tq_nthreads_target = newtarget;
866 cv_broadcast(&tq->tq_dispatch_cv);
867 cv_broadcast(&tq->tq_exit_cv);
868 }
869 }
870
871 /* called during task queue creation */
872 static void
873 taskq_cpupct_install(taskq_t *tq, cpupart_t *cpup)
874 {
875 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
876
877 mutex_enter(&cpu_lock);
878 mutex_enter(&tq->tq_lock);
879 tq->tq_cpupart = cpup->cp_id;
880 taskq_update_nthreads(tq, cpup->cp_ncpus);
881 mutex_exit(&tq->tq_lock);
882
883 list_insert_tail(&taskq_cpupct_list, tq);
884 mutex_exit(&cpu_lock);
885 }
886
887 static void
888 taskq_cpupct_remove(taskq_t *tq)
889 {
890 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
891
892 mutex_enter(&cpu_lock);
893 list_remove(&taskq_cpupct_list, tq);
894 mutex_exit(&cpu_lock);
895 }
896
897 /*ARGSUSED*/
898 static int
899 taskq_cpu_setup(cpu_setup_t what, int id, void *arg)
900 {
901 taskq_t *tq;
902 cpupart_t *cp = cpu[id]->cpu_part;
903 uint_t ncpus = cp->cp_ncpus;
904
905 ASSERT(MUTEX_HELD(&cpu_lock));
906 ASSERT(ncpus > 0);
907
908 switch (what) {
909 case CPU_OFF:
910 case CPU_CPUPART_OUT:
911 /* offlines are called *before* the cpu is offlined. */
912 if (ncpus > 1)
913 ncpus--;
914 break;
915
916 case CPU_ON:
917 case CPU_CPUPART_IN:
918 break;
919
920 default:
921 return (0); /* doesn't affect cpu count */
922 }
923
924 for (tq = list_head(&taskq_cpupct_list); tq != NULL;
925 tq = list_next(&taskq_cpupct_list, tq)) {
926
927 mutex_enter(&tq->tq_lock);
928 /*
929 * If the taskq is part of the cpuset which is changing,
930 * update its nthreads_target.
931 */
932 if (tq->tq_cpupart == cp->cp_id) {
933 taskq_update_nthreads(tq, ncpus);
934 }
935 mutex_exit(&tq->tq_lock);
936 }
937 return (0);
938 }
939
940 void
941 taskq_mp_init(void)
942 {
943 mutex_enter(&cpu_lock);
944 register_cpu_setup_func(taskq_cpu_setup, NULL);
945 /*
946 * Make sure we're up to date. At this point in boot, there is only
947 * one processor set, so we only have to update the current CPU.
948 */
949 (void) taskq_cpu_setup(CPU_ON, CPU->cpu_id, NULL);
950 mutex_exit(&cpu_lock);
951 }
952
953 /*
954 * Create global system dynamic task queue.
955 */
956 void
957 system_taskq_init(void)
958 {
959 system_taskq = taskq_create_common("system_taskq", 0,
960 system_taskq_size * max_ncpus, minclsyspri, 4, 512, &p0, 0,
961 TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
962 }
963
964 /*
965 * taskq_ent_alloc()
966 *
967 * Allocates a new taskq_ent_t structure either from the free list or from the
968 * cache. Returns NULL if it can't be allocated.
969 *
970 * Assumes: tq->tq_lock is held.
971 */
972 static taskq_ent_t *
973 taskq_ent_alloc(taskq_t *tq, int flags)
974 {
975 int kmflags = (flags & TQ_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
976 taskq_ent_t *tqe;
977 clock_t wait_time;
978 clock_t wait_rv;
979
980 ASSERT(MUTEX_HELD(&tq->tq_lock));
981
982 /*
983 * TQ_NOALLOC allocations are allowed to use the freelist, even if
984 * we are below tq_minalloc.
985 */
986 again: if ((tqe = tq->tq_freelist) != NULL &&
987 ((flags & TQ_NOALLOC) || tq->tq_nalloc >= tq->tq_minalloc)) {
988 tq->tq_freelist = tqe->tqent_next;
989 } else {
990 if (flags & TQ_NOALLOC)
991 return (NULL);
992
993 if (tq->tq_nalloc >= tq->tq_maxalloc) {
994 if (kmflags & KM_NOSLEEP)
995 return (NULL);
996
997 /*
998 * We don't want to exceed tq_maxalloc, but we can't
999 * wait for other tasks to complete (and thus free up
1000 * task structures) without risking deadlock with
1001 * the caller. So, we just delay for one second
1002 * to throttle the allocation rate. If we have tasks
1003 * complete before one second timeout expires then
1004 * taskq_ent_free will signal us and we will
1005 * immediately retry the allocation (reap free).
1006 */
1007 wait_time = ddi_get_lbolt() + hz;
1008 while (tq->tq_freelist == NULL) {
1009 tq->tq_maxalloc_wait++;
1010 wait_rv = cv_timedwait(&tq->tq_maxalloc_cv,
1011 &tq->tq_lock, wait_time);
1012 tq->tq_maxalloc_wait--;
1013 if (wait_rv == -1)
1014 break;
1015 }
1016 if (tq->tq_freelist)
1017 goto again; /* reap freelist */
1018
1019 }
1020 mutex_exit(&tq->tq_lock);
1021
1022 tqe = kmem_cache_alloc(taskq_ent_cache, kmflags);
1023
1024 mutex_enter(&tq->tq_lock);
1025 if (tqe != NULL)
1026 tq->tq_nalloc++;
1027 }
1028 return (tqe);
1029 }
1030
1031 /*
1032 * taskq_ent_free()
1033 *
1034 * Free taskq_ent_t structure by either putting it on the free list or freeing
1035 * it to the cache.
1036 *
1037 * Assumes: tq->tq_lock is held.
1038 */
1039 static void
1040 taskq_ent_free(taskq_t *tq, taskq_ent_t *tqe)
1041 {
1042 ASSERT(MUTEX_HELD(&tq->tq_lock));
1043
1044 if (tq->tq_nalloc <= tq->tq_minalloc) {
1045 tqe->tqent_next = tq->tq_freelist;
1046 tq->tq_freelist = tqe;
1047 } else {
1048 tq->tq_nalloc--;
1049 mutex_exit(&tq->tq_lock);
1050 kmem_cache_free(taskq_ent_cache, tqe);
1051 mutex_enter(&tq->tq_lock);
1052 }
1053
1054 if (tq->tq_maxalloc_wait)
1055 cv_signal(&tq->tq_maxalloc_cv);
1056 }
1057
1058 /*
1059 * taskq_ent_exists()
1060 *
1061 * Return 1 if taskq already has entry for calling 'func(arg)'.
1062 *
1063 * Assumes: tq->tq_lock is held.
1064 */
1065 static int
1066 taskq_ent_exists(taskq_t *tq, task_func_t func, void *arg)
1067 {
1068 taskq_ent_t *tqe;
1069
1070 ASSERT(MUTEX_HELD(&tq->tq_lock));
1071
1072 for (tqe = tq->tq_task.tqent_next; tqe != &tq->tq_task;
1073 tqe = tqe->tqent_next)
1074 if ((tqe->tqent_func == func) && (tqe->tqent_arg == arg))
1075 return (1);
1076 return (0);
1077 }
1078
1079 /*
1080 * Dispatch a task "func(arg)" to a free entry of bucket b.
1081 *
1082 * Assumes: no bucket locks is held.
1083 *
1084 * Returns: a pointer to an entry if dispatch was successful.
1085 * NULL if there are no free entries or if the bucket is suspended.
1086 */
1087 static taskq_ent_t *
1088 taskq_bucket_dispatch(taskq_bucket_t *b, task_func_t func, void *arg)
1089 {
1090 taskq_ent_t *tqe;
1091
1092 ASSERT(MUTEX_NOT_HELD(&b->tqbucket_lock));
1093 ASSERT(func != NULL);
1094
1095 mutex_enter(&b->tqbucket_lock);
1096
1097 ASSERT(b->tqbucket_nfree != 0 || IS_EMPTY(b->tqbucket_freelist));
1098 ASSERT(b->tqbucket_nfree == 0 || !IS_EMPTY(b->tqbucket_freelist));
1099
1100 /*
1101 * Get en entry from the freelist if there is one.
1102 * Schedule task into the entry.
1103 */
1104 if ((b->tqbucket_nfree != 0) &&
1105 !(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
1106 tqe = b->tqbucket_freelist.tqent_prev;
1107
1108 ASSERT(tqe != &b->tqbucket_freelist);
1109 ASSERT(tqe->tqent_thread != NULL);
1110
1111 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1112 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1113 b->tqbucket_nalloc++;
1114 b->tqbucket_nfree--;
1115 tqe->tqent_func = func;
1116 tqe->tqent_arg = arg;
1117 TQ_STAT(b, tqs_hits);
1118 cv_signal(&tqe->tqent_cv);
1119 DTRACE_PROBE2(taskq__d__enqueue, taskq_bucket_t *, b,
1120 taskq_ent_t *, tqe);
1121 } else {
1122 tqe = NULL;
1123 TQ_STAT(b, tqs_misses);
1124 }
1125 mutex_exit(&b->tqbucket_lock);
1126 return (tqe);
1127 }
1128
1129 /*
1130 * Dispatch a task.
1131 *
1132 * Assumes: func != NULL
1133 *
1134 * Returns: NULL if dispatch failed.
1135 * non-NULL if task dispatched successfully.
1136 * Actual return value is the pointer to taskq entry that was used to
1137 * dispatch a task. This is useful for debugging.
1138 */
1139 taskqid_t
1140 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
1141 {
1142 taskq_bucket_t *bucket = NULL; /* Which bucket needs extension */
1143 taskq_ent_t *tqe = NULL;
1144 taskq_ent_t *tqe1;
1145 uint_t bsize;
1146
1147 ASSERT(tq != NULL);
1148 ASSERT(func != NULL);
1149
1150 if (!(tq->tq_flags & TASKQ_DYNAMIC)) {
1151 /*
1152 * TQ_NOQUEUE flag can't be used with non-dynamic task queues.
1153 */
1154 ASSERT(!(flags & TQ_NOQUEUE));
1155 /*
1156 * Enqueue the task to the underlying queue.
1157 */
1158 mutex_enter(&tq->tq_lock);
1159
1160 TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flags);
1161
1162 if ((tqe = taskq_ent_alloc(tq, flags)) == NULL) {
1163 tq->tq_nomem++;
1164 mutex_exit(&tq->tq_lock);
1165 return (NULL);
1166 }
1167 /* Make sure we start without any flags */
1168 tqe->tqent_un.tqent_flags = 0;
1169
1170 if (flags & TQ_FRONT) {
1171 TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1172 } else {
1173 TQ_ENQUEUE(tq, tqe, func, arg);
1174 }
1175 mutex_exit(&tq->tq_lock);
1176 return ((taskqid_t)tqe);
1177 }
1178
1179 /*
1180 * Dynamic taskq dispatching.
1181 */
1182 ASSERT(!(flags & (TQ_NOALLOC | TQ_FRONT)));
1183 TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flags);
1184
1185 bsize = tq->tq_nbuckets;
1186
1187 if (bsize == 1) {
1188 /*
1189 * In a single-CPU case there is only one bucket, so get
1190 * entry directly from there.
1191 */
1192 if ((tqe = taskq_bucket_dispatch(tq->tq_buckets, func, arg))
1193 != NULL)
1194 return ((taskqid_t)tqe); /* Fastpath */
1195 bucket = tq->tq_buckets;
1196 } else {
1197 int loopcount;
1198 taskq_bucket_t *b;
1199 uintptr_t h = ((uintptr_t)CPU + (uintptr_t)arg) >> 3;
1200
1201 h = TQ_HASH(h);
1202
1203 /*
1204 * The 'bucket' points to the original bucket that we hit. If we
1205 * can't allocate from it, we search other buckets, but only
1206 * extend this one.
1207 */
1208 b = &tq->tq_buckets[h & (bsize - 1)];
1209 ASSERT(b->tqbucket_taskq == tq); /* Sanity check */
1210
1211 /*
1212 * Do a quick check before grabbing the lock. If the bucket does
1213 * not have free entries now, chances are very small that it
1214 * will after we take the lock, so we just skip it.
1215 */
1216 if (b->tqbucket_nfree != 0) {
1217 if ((tqe = taskq_bucket_dispatch(b, func, arg)) != NULL)
1218 return ((taskqid_t)tqe); /* Fastpath */
1219 } else {
1220 TQ_STAT(b, tqs_misses);
1221 }
1222
1223 bucket = b;
1224 loopcount = MIN(taskq_search_depth, bsize);
1225 /*
1226 * If bucket dispatch failed, search loopcount number of buckets
1227 * before we give up and fail.
1228 */
1229 do {
1230 b = &tq->tq_buckets[++h & (bsize - 1)];
1231 ASSERT(b->tqbucket_taskq == tq); /* Sanity check */
1232 loopcount--;
1233
1234 if (b->tqbucket_nfree != 0) {
1235 tqe = taskq_bucket_dispatch(b, func, arg);
1236 } else {
1237 TQ_STAT(b, tqs_misses);
1238 }
1239 } while ((tqe == NULL) && (loopcount > 0));
1240 }
1241
1242 /*
1243 * At this point we either scheduled a task and (tqe != NULL) or failed
1244 * (tqe == NULL). Try to recover from fails.
1245 */
1246
1247 /*
1248 * For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
1249 */
1250 if ((tqe == NULL) && !(flags & TQ_NOSLEEP)) {
1251 /*
1252 * taskq_bucket_extend() may fail to do anything, but this is
1253 * fine - we deal with it later. If the bucket was successfully
1254 * extended, there is a good chance that taskq_bucket_dispatch()
1255 * will get this new entry, unless someone is racing with us and
1256 * stealing the new entry from under our nose.
1257 * taskq_bucket_extend() may sleep.
1258 */
1259 taskq_bucket_extend(bucket);
1260 TQ_STAT(bucket, tqs_disptcreates);
1261 if ((tqe = taskq_bucket_dispatch(bucket, func, arg)) != NULL)
1262 return ((taskqid_t)tqe);
1263 }
1264
1265 ASSERT(bucket != NULL);
1266
1267 /*
1268 * Since there are not enough free entries in the bucket, add a
1269 * taskq entry to extend it in the background using backing queue
1270 * (unless we already have a taskq entry to perform that extension).
1271 */
1272 mutex_enter(&tq->tq_lock);
1273 if (!taskq_ent_exists(tq, taskq_bucket_extend, bucket)) {
1274 if ((tqe1 = taskq_ent_alloc(tq, TQ_NOSLEEP)) != NULL) {
1275 TQ_ENQUEUE_FRONT(tq, tqe1, taskq_bucket_extend, bucket);
1276 } else {
1277 tq->tq_nomem++;
1278 }
1279 }
1280
1281 /*
1282 * Dispatch failed and we can't find an entry to schedule a task.
1283 * Revert to the backing queue unless TQ_NOQUEUE was asked.
1284 */
1285 if ((tqe == NULL) && !(flags & TQ_NOQUEUE)) {
1286 if ((tqe = taskq_ent_alloc(tq, flags)) != NULL) {
1287 TQ_ENQUEUE(tq, tqe, func, arg);
1288 } else {
1289 tq->tq_nomem++;
1290 }
1291 }
1292 mutex_exit(&tq->tq_lock);
1293
1294 return ((taskqid_t)tqe);
1295 }
1296
1297 void
1298 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
1299 taskq_ent_t *tqe)
1300 {
1301 ASSERT(func != NULL);
1302 ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
1303
1304 /*
1305 * Mark it as a prealloc'd task. This is important
1306 * to ensure that we don't free it later.
1307 */
1308 tqe->tqent_un.tqent_flags |= TQENT_FLAG_PREALLOC;
1309 /*
1310 * Enqueue the task to the underlying queue.
1311 */
1312 mutex_enter(&tq->tq_lock);
1313
1314 if (flags & TQ_FRONT) {
1315 TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1316 } else {
1317 TQ_ENQUEUE(tq, tqe, func, arg);
1318 }
1319 mutex_exit(&tq->tq_lock);
1320 }
1321
1322 /*
1323 * Wait for all pending tasks to complete.
1324 * Calling taskq_wait from a task will cause deadlock.
1325 */
1326 void
1327 taskq_wait(taskq_t *tq)
1328 {
1329 ASSERT(tq != curthread->t_taskq);
1330
1331 mutex_enter(&tq->tq_lock);
1332 while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
1333 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1334 mutex_exit(&tq->tq_lock);
1335
1336 if (tq->tq_flags & TASKQ_DYNAMIC) {
1337 taskq_bucket_t *b = tq->tq_buckets;
1338 int bid = 0;
1339 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1340 mutex_enter(&b->tqbucket_lock);
1341 while (b->tqbucket_nalloc > 0)
1342 cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
1343 mutex_exit(&b->tqbucket_lock);
1344 }
1345 }
1346 }
1347
1348 /*
1349 * Suspend execution of tasks.
1350 *
1351 * Tasks in the queue part will be suspended immediately upon return from this
1352 * function. Pending tasks in the dynamic part will continue to execute, but all
1353 * new tasks will be suspended.
1354 */
1355 void
1356 taskq_suspend(taskq_t *tq)
1357 {
1358 rw_enter(&tq->tq_threadlock, RW_WRITER);
1359
1360 if (tq->tq_flags & TASKQ_DYNAMIC) {
1361 taskq_bucket_t *b = tq->tq_buckets;
1362 int bid = 0;
1363 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1364 mutex_enter(&b->tqbucket_lock);
1365 b->tqbucket_flags |= TQBUCKET_SUSPEND;
1366 mutex_exit(&b->tqbucket_lock);
1367 }
1368 }
1369 /*
1370 * Mark task queue as being suspended. Needed for taskq_suspended().
1371 */
1372 mutex_enter(&tq->tq_lock);
1373 ASSERT(!(tq->tq_flags & TASKQ_SUSPENDED));
1374 tq->tq_flags |= TASKQ_SUSPENDED;
1375 mutex_exit(&tq->tq_lock);
1376 }
1377
1378 /*
1379 * returns: 1 if tq is suspended, 0 otherwise.
1380 */
1381 int
1382 taskq_suspended(taskq_t *tq)
1383 {
1384 return ((tq->tq_flags & TASKQ_SUSPENDED) != 0);
1385 }
1386
1387 /*
1388 * Resume taskq execution.
1389 */
1390 void
1391 taskq_resume(taskq_t *tq)
1392 {
1393 ASSERT(RW_WRITE_HELD(&tq->tq_threadlock));
1394
1395 if (tq->tq_flags & TASKQ_DYNAMIC) {
1396 taskq_bucket_t *b = tq->tq_buckets;
1397 int bid = 0;
1398 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1399 mutex_enter(&b->tqbucket_lock);
1400 b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
1401 mutex_exit(&b->tqbucket_lock);
1402 }
1403 }
1404 mutex_enter(&tq->tq_lock);
1405 ASSERT(tq->tq_flags & TASKQ_SUSPENDED);
1406 tq->tq_flags &= ~TASKQ_SUSPENDED;
1407 mutex_exit(&tq->tq_lock);
1408
1409 rw_exit(&tq->tq_threadlock);
1410 }
1411
1412 int
1413 taskq_member(taskq_t *tq, kthread_t *thread)
1414 {
1415 return (thread->t_taskq == tq);
1416 }
1417
1418 /*
1419 * Creates a thread in the taskq. We only allow one outstanding create at
1420 * a time. We drop and reacquire the tq_lock in order to avoid blocking other
1421 * taskq activity while thread_create() or lwp_kernel_create() run.
1422 *
1423 * The first time we're called, we do some additional setup, and do not
1424 * return until there are enough threads to start servicing requests.
1425 */
1426 static void
1427 taskq_thread_create(taskq_t *tq)
1428 {
1429 kthread_t *t;
1430 const boolean_t first = (tq->tq_nthreads == 0);
1431
1432 ASSERT(MUTEX_HELD(&tq->tq_lock));
1433 ASSERT(tq->tq_flags & TASKQ_CHANGING);
1434 ASSERT(tq->tq_nthreads < tq->tq_nthreads_target);
1435 ASSERT(!(tq->tq_flags & TASKQ_THREAD_CREATED));
1436
1437
1438 tq->tq_flags |= TASKQ_THREAD_CREATED;
1439 tq->tq_active++;
1440 mutex_exit(&tq->tq_lock);
1441
1442 /*
1443 * With TASKQ_DUTY_CYCLE the new thread must have an LWP
1444 * as explained in ../disp/sysdc.c (for the msacct data).
1445 * Otherwise simple kthreads are preferred.
1446 */
1447 if ((tq->tq_flags & TASKQ_DUTY_CYCLE) != 0) {
1448 /* Enforced in taskq_create_common */
1449 ASSERT3P(tq->tq_proc, !=, &p0);
1450 t = lwp_kernel_create(tq->tq_proc, taskq_thread, tq, TS_RUN,
1451 tq->tq_pri);
1452 } else {
1453 t = thread_create(NULL, 0, taskq_thread, tq, 0, tq->tq_proc,
1454 TS_RUN, tq->tq_pri);
1455 }
1456
1457 if (!first) {
1458 mutex_enter(&tq->tq_lock);
1459 return;
1460 }
1461
1462 /*
1463 * We know the thread cannot go away, since tq cannot be
1464 * destroyed until creation has completed. We can therefore
1465 * safely dereference t.
1466 */
1467 if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1468 taskq_cpupct_install(tq, t->t_cpupart);
1469 }
1470 mutex_enter(&tq->tq_lock);
1471
1472 /* Wait until we can service requests. */
1473 while (tq->tq_nthreads != tq->tq_nthreads_target &&
1474 tq->tq_nthreads < TASKQ_CREATE_ACTIVE_THREADS) {
1475 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1476 }
1477 }
1478
1479 /*
1480 * Common "sleep taskq thread" function, which handles CPR stuff, as well
1481 * as giving a nice common point for debuggers to find inactive threads.
1482 */
1483 static clock_t
1484 taskq_thread_wait(taskq_t *tq, kmutex_t *mx, kcondvar_t *cv,
1485 callb_cpr_t *cprinfo, clock_t timeout)
1486 {
1487 clock_t ret = 0;
1488
1489 if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1490 CALLB_CPR_SAFE_BEGIN(cprinfo);
1491 }
1492 if (timeout < 0)
1493 cv_wait(cv, mx);
1494 else
1495 ret = cv_reltimedwait(cv, mx, timeout, TR_CLOCK_TICK);
1496
1497 if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1498 CALLB_CPR_SAFE_END(cprinfo, mx);
1499 }
1500
1501 return (ret);
1502 }
1503
1504 /*
1505 * Worker thread for processing task queue.
1506 */
1507 static void
1508 taskq_thread(void *arg)
1509 {
1510 int thread_id;
1511
1512 taskq_t *tq = arg;
1513 taskq_ent_t *tqe;
1514 callb_cpr_t cprinfo;
1515 hrtime_t start, end;
1516 boolean_t freeit;
1517
1518 curthread->t_taskq = tq; /* mark ourselves for taskq_member() */
1519
1520 if (curproc != &p0 && (tq->tq_flags & TASKQ_DUTY_CYCLE)) {
1521 sysdc_thread_enter(curthread, tq->tq_DC,
1522 (tq->tq_flags & TASKQ_DC_BATCH) ? SYSDC_THREAD_BATCH : 0);
1523 }
1524
1525 if (tq->tq_flags & TASKQ_CPR_SAFE) {
1526 CALLB_CPR_INIT_SAFE(curthread, tq->tq_name);
1527 } else {
1528 CALLB_CPR_INIT(&cprinfo, &tq->tq_lock, callb_generic_cpr,
1529 tq->tq_name);
1530 }
1531 mutex_enter(&tq->tq_lock);
1532 thread_id = ++tq->tq_nthreads;
1533 ASSERT(tq->tq_flags & TASKQ_THREAD_CREATED);
1534 ASSERT(tq->tq_flags & TASKQ_CHANGING);
1535 tq->tq_flags &= ~TASKQ_THREAD_CREATED;
1536
1537 VERIFY3S(thread_id, <=, tq->tq_nthreads_max);
1538
1539 if (tq->tq_nthreads_max == 1)
1540 tq->tq_thread = curthread;
1541 else
1542 tq->tq_threadlist[thread_id - 1] = curthread;
1543
1544 /* Allow taskq_create_common()'s taskq_thread_create() to return. */
1545 if (tq->tq_nthreads == TASKQ_CREATE_ACTIVE_THREADS)
1546 cv_broadcast(&tq->tq_wait_cv);
1547
1548 for (;;) {
1549 if (tq->tq_flags & TASKQ_CHANGING) {
1550 /* See if we're no longer needed */
1551 if (thread_id > tq->tq_nthreads_target) {
1552 /*
1553 * To preserve the one-to-one mapping between
1554 * thread_id and thread, we must exit from
1555 * highest thread ID to least.
1556 *
1557 * However, if everyone is exiting, the order
1558 * doesn't matter, so just exit immediately.
1559 * (this is safe, since you must wait for
1560 * nthreads to reach 0 after setting
1561 * tq_nthreads_target to 0)
1562 */
1563 if (thread_id == tq->tq_nthreads ||
1564 tq->tq_nthreads_target == 0)
1565 break;
1566
1567 /* Wait for higher thread_ids to exit */
1568 (void) taskq_thread_wait(tq, &tq->tq_lock,
1569 &tq->tq_exit_cv, &cprinfo, -1);
1570 continue;
1571 }
1572
1573 /*
1574 * If no thread is starting taskq_thread(), we can
1575 * do some bookkeeping.
1576 */
1577 if (!(tq->tq_flags & TASKQ_THREAD_CREATED)) {
1578 /* Check if we've reached our target */
1579 if (tq->tq_nthreads == tq->tq_nthreads_target) {
1580 tq->tq_flags &= ~TASKQ_CHANGING;
1581 cv_broadcast(&tq->tq_wait_cv);
1582 }
1583 /* Check if we need to create a thread */
1584 if (tq->tq_nthreads < tq->tq_nthreads_target) {
1585 taskq_thread_create(tq);
1586 continue; /* tq_lock was dropped */
1587 }
1588 }
1589 }
1590 if ((tqe = tq->tq_task.tqent_next) == &tq->tq_task) {
1591 if (--tq->tq_active == 0)
1592 cv_broadcast(&tq->tq_wait_cv);
1593 (void) taskq_thread_wait(tq, &tq->tq_lock,
1594 &tq->tq_dispatch_cv, &cprinfo, -1);
1595 tq->tq_active++;
1596 continue;
1597 }
1598
1599 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1600 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1601 mutex_exit(&tq->tq_lock);
1602
1603 /*
1604 * For prealloc'd tasks, we don't free anything. We
1605 * have to check this now, because once we call the
1606 * function for a prealloc'd taskq, we can't touch the
1607 * tqent any longer (calling the function returns the
1608 * ownershp of the tqent back to caller of
1609 * taskq_dispatch.)
1610 */
1611 if ((!(tq->tq_flags & TASKQ_DYNAMIC)) &&
1612 (tqe->tqent_un.tqent_flags & TQENT_FLAG_PREALLOC)) {
1613 /* clear pointers to assist assertion checks */
1614 tqe->tqent_next = tqe->tqent_prev = NULL;
1615 freeit = B_FALSE;
1616 } else {
1617 freeit = B_TRUE;
1618 }
1619
1620 rw_enter(&tq->tq_threadlock, RW_READER);
1621 start = gethrtime();
1622 DTRACE_PROBE2(taskq__exec__start, taskq_t *, tq,
1623 taskq_ent_t *, tqe);
1624 tqe->tqent_func(tqe->tqent_arg);
1625 DTRACE_PROBE2(taskq__exec__end, taskq_t *, tq,
1626 taskq_ent_t *, tqe);
1627 end = gethrtime();
1628 rw_exit(&tq->tq_threadlock);
1629
1630 mutex_enter(&tq->tq_lock);
1631 tq->tq_totaltime += end - start;
1632 tq->tq_executed++;
1633
1634 if (freeit)
1635 taskq_ent_free(tq, tqe);
1636 }
1637
1638 if (tq->tq_nthreads_max == 1)
1639 tq->tq_thread = NULL;
1640 else
1641 tq->tq_threadlist[thread_id - 1] = NULL;
1642
1643 /* We're exiting, and therefore no longer active */
1644 ASSERT(tq->tq_active > 0);
1645 tq->tq_active--;
1646
1647 ASSERT(tq->tq_nthreads > 0);
1648 tq->tq_nthreads--;
1649
1650 /* Wake up anyone waiting for us to exit */
1651 cv_broadcast(&tq->tq_exit_cv);
1652 if (tq->tq_nthreads == tq->tq_nthreads_target) {
1653 if (!(tq->tq_flags & TASKQ_THREAD_CREATED))
1654 tq->tq_flags &= ~TASKQ_CHANGING;
1655
1656 cv_broadcast(&tq->tq_wait_cv);
1657 }
1658
1659 ASSERT(!(tq->tq_flags & TASKQ_CPR_SAFE));
1660 CALLB_CPR_EXIT(&cprinfo); /* drops tq->tq_lock */
1661 if (curthread->t_lwp != NULL) {
1662 mutex_enter(&curproc->p_lock);
1663 lwp_exit();
1664 } else {
1665 thread_exit();
1666 }
1667 }
1668
1669 /*
1670 * Worker per-entry thread for dynamic dispatches.
1671 */
1672 static void
1673 taskq_d_thread(taskq_ent_t *tqe)
1674 {
1675 taskq_bucket_t *bucket = tqe->tqent_un.tqent_bucket;
1676 taskq_t *tq = bucket->tqbucket_taskq;
1677 kmutex_t *lock = &bucket->tqbucket_lock;
1678 kcondvar_t *cv = &tqe->tqent_cv;
1679 callb_cpr_t cprinfo;
1680 clock_t w;
1681
1682 CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, tq->tq_name);
1683
1684 mutex_enter(lock);
1685
1686 for (;;) {
1687 /*
1688 * If a task is scheduled (func != NULL), execute it, otherwise
1689 * sleep, waiting for a job.
1690 */
1691 if (tqe->tqent_func != NULL) {
1692 hrtime_t start;
1693 hrtime_t end;
1694
1695 ASSERT(bucket->tqbucket_nalloc > 0);
1696
1697 /*
1698 * It is possible to free the entry right away before
1699 * actually executing the task so that subsequent
1700 * dispatches may immediately reuse it. But this,
1701 * effectively, creates a two-length queue in the entry
1702 * and may lead to a deadlock if the execution of the
1703 * current task depends on the execution of the next
1704 * scheduled task. So, we keep the entry busy until the
1705 * task is processed.
1706 */
1707
1708 mutex_exit(lock);
1709 start = gethrtime();
1710 DTRACE_PROBE3(taskq__d__exec__start, taskq_t *, tq,
1711 taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1712 tqe->tqent_func(tqe->tqent_arg);
1713 DTRACE_PROBE3(taskq__d__exec__end, taskq_t *, tq,
1714 taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1715 end = gethrtime();
1716 mutex_enter(lock);
1717 bucket->tqbucket_totaltime += end - start;
1718
1719 /*
1720 * Return the entry to the bucket free list.
1721 */
1722 tqe->tqent_func = NULL;
1723 TQ_APPEND(bucket->tqbucket_freelist, tqe);
1724 bucket->tqbucket_nalloc--;
1725 bucket->tqbucket_nfree++;
1726 ASSERT(!IS_EMPTY(bucket->tqbucket_freelist));
1727 /*
1728 * taskq_wait() waits for nalloc to drop to zero on
1729 * tqbucket_cv.
1730 */
1731 cv_signal(&bucket->tqbucket_cv);
1732 }
1733
1734 /*
1735 * At this point the entry must be in the bucket free list -
1736 * either because it was there initially or because it just
1737 * finished executing a task and put itself on the free list.
1738 */
1739 ASSERT(bucket->tqbucket_nfree > 0);
1740 /*
1741 * Go to sleep unless we are closing.
1742 * If a thread is sleeping too long, it dies.
1743 */
1744 if (! (bucket->tqbucket_flags & TQBUCKET_CLOSE)) {
1745 w = taskq_thread_wait(tq, lock, cv,
1746 &cprinfo, taskq_thread_timeout * hz);
1747 }
1748
1749 /*
1750 * At this point we may be in two different states:
1751 *
1752 * (1) tqent_func is set which means that a new task is
1753 * dispatched and we need to execute it.
1754 *
1755 * (2) Thread is sleeping for too long or we are closing. In
1756 * both cases destroy the thread and the entry.
1757 */
1758
1759 /* If func is NULL we should be on the freelist. */
1760 ASSERT((tqe->tqent_func != NULL) ||
1761 (bucket->tqbucket_nfree > 0));
1762 /* If func is non-NULL we should be allocated */
1763 ASSERT((tqe->tqent_func == NULL) ||
1764 (bucket->tqbucket_nalloc > 0));
1765
1766 /* Check freelist consistency */
1767 ASSERT((bucket->tqbucket_nfree > 0) ||
1768 IS_EMPTY(bucket->tqbucket_freelist));
1769 ASSERT((bucket->tqbucket_nfree == 0) ||
1770 !IS_EMPTY(bucket->tqbucket_freelist));
1771
1772 if ((tqe->tqent_func == NULL) &&
1773 ((w == -1) || (bucket->tqbucket_flags & TQBUCKET_CLOSE))) {
1774 /*
1775 * This thread is sleeping for too long or we are
1776 * closing - time to die.
1777 * Thread creation/destruction happens rarely,
1778 * so grabbing the lock is not a big performance issue.
1779 * The bucket lock is dropped by CALLB_CPR_EXIT().
1780 */
1781
1782 /* Remove the entry from the free list. */
1783 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1784 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1785 ASSERT(bucket->tqbucket_nfree > 0);
1786 bucket->tqbucket_nfree--;
1787
1788 TQ_STAT(bucket, tqs_tdeaths);
1789 cv_signal(&bucket->tqbucket_cv);
1790 tqe->tqent_thread = NULL;
1791 mutex_enter(&tq->tq_lock);
1792 tq->tq_tdeaths++;
1793 mutex_exit(&tq->tq_lock);
1794 CALLB_CPR_EXIT(&cprinfo);
1795 kmem_cache_free(taskq_ent_cache, tqe);
1796 thread_exit();
1797 }
1798 }
1799 }
1800
1801
1802 /*
1803 * Taskq creation. May sleep for memory.
1804 * Always use automatically generated instances to avoid kstat name space
1805 * collisions.
1806 */
1807
1808 taskq_t *
1809 taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
1810 int maxalloc, uint_t flags)
1811 {
1812 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1813
1814 return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1815 maxalloc, &p0, 0, flags | TASKQ_NOINSTANCE));
1816 }
1817
1818 /*
1819 * Create an instance of task queue. It is legal to create task queues with the
1820 * same name and different instances.
1821 *
1822 * taskq_create_instance is used by ddi_taskq_create() where it gets the
1823 * instance from ddi_get_instance(). In some cases the instance is not
1824 * initialized and is set to -1. This case is handled as if no instance was
1825 * passed at all.
1826 */
1827 taskq_t *
1828 taskq_create_instance(const char *name, int instance, int nthreads, pri_t pri,
1829 int minalloc, int maxalloc, uint_t flags)
1830 {
1831 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1832 ASSERT((instance >= 0) || (instance == -1));
1833
1834 if (instance < 0) {
1835 flags |= TASKQ_NOINSTANCE;
1836 }
1837
1838 return (taskq_create_common(name, instance, nthreads,
1839 pri, minalloc, maxalloc, &p0, 0, flags));
1840 }
1841
1842 taskq_t *
1843 taskq_create_proc(const char *name, int nthreads, pri_t pri, int minalloc,
1844 int maxalloc, proc_t *proc, 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, pri, minalloc,
1850 maxalloc, proc, 0, flags | TASKQ_NOINSTANCE));
1851 }
1852
1853 taskq_t *
1854 taskq_create_sysdc(const char *name, int nthreads, int minalloc,
1855 int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1856 {
1857 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1858 ASSERT(proc->p_flag & SSYS);
1859
1860 return (taskq_create_common(name, 0, nthreads, minclsyspri, minalloc,
1861 maxalloc, proc, dc, flags | TASKQ_NOINSTANCE | TASKQ_DUTY_CYCLE));
1862 }
1863
1864 static taskq_t *
1865 taskq_create_common(const char *name, int instance, int nthreads, pri_t pri,
1866 int minalloc, int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1867 {
1868 taskq_t *tq = kmem_cache_alloc(taskq_cache, KM_SLEEP);
1869 uint_t ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
1870 uint_t bsize; /* # of buckets - always power of 2 */
1871 int max_nthreads;
1872
1873 /*
1874 * TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
1875 * mutually incompatible.
1876 */
1877 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_CPR_SAFE));
1878 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_THREADS_CPU_PCT));
1879 IMPLY((flags & TASKQ_CPR_SAFE), !(flags & TASKQ_THREADS_CPU_PCT));
1880
1881 /* Cannot have DYNAMIC with DUTY_CYCLE */
1882 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_DUTY_CYCLE));
1883
1884 /* Cannot have DUTY_CYCLE with a p0 kernel process */
1885 IMPLY((flags & TASKQ_DUTY_CYCLE), proc != &p0);
1886
1887 /* Cannot have DC_BATCH without DUTY_CYCLE */
1888 ASSERT((flags & (TASKQ_DUTY_CYCLE|TASKQ_DC_BATCH)) != TASKQ_DC_BATCH);
1889
1890 ASSERT(proc != NULL);
1891
1892 bsize = 1 << (highbit(ncpus) - 1);
1893 ASSERT(bsize >= 1);
1894 bsize = MIN(bsize, taskq_maxbuckets);
1895
1896 if (flags & TASKQ_DYNAMIC) {
1897 ASSERT3S(nthreads, >=, 1);
1898 tq->tq_maxsize = nthreads;
1899
1900 /* For dynamic task queues use just one backup thread */
1901 nthreads = max_nthreads = 1;
1902
1903 } else if (flags & TASKQ_THREADS_CPU_PCT) {
1904 uint_t pct;
1905 ASSERT3S(nthreads, >=, 0);
1906 pct = nthreads;
1907
1908 if (pct > taskq_cpupct_max_percent)
1909 pct = taskq_cpupct_max_percent;
1910
1911 /*
1912 * If you're using THREADS_CPU_PCT, the process for the
1913 * taskq threads must be curproc. This allows any pset
1914 * binding to be inherited correctly. If proc is &p0,
1915 * we won't be creating LWPs, so new threads will be assigned
1916 * to the default processor set.
1917 */
1918 ASSERT(curproc == proc || proc == &p0);
1919 tq->tq_threads_ncpus_pct = pct;
1920 nthreads = 1; /* corrected in taskq_thread_create() */
1921 max_nthreads = TASKQ_THREADS_PCT(max_ncpus, pct);
1922
1923 } else {
1924 ASSERT3S(nthreads, >=, 1);
1925 max_nthreads = nthreads;
1926 }
1927
1928 if (max_nthreads < taskq_minimum_nthreads_max)
1929 max_nthreads = taskq_minimum_nthreads_max;
1930
1931 /*
1932 * Make sure the name is 0-terminated, and conforms to the rules for
1933 * C indentifiers
1934 */
1935 (void) strncpy(tq->tq_name, name, TASKQ_NAMELEN + 1);
1936 strident_canon(tq->tq_name, TASKQ_NAMELEN + 1);
1937
1938 tq->tq_flags = flags | TASKQ_CHANGING;
1939 tq->tq_active = 0;
1940 tq->tq_instance = instance;
1941 tq->tq_nthreads_target = nthreads;
1942 tq->tq_nthreads_max = max_nthreads;
1943 tq->tq_minalloc = minalloc;
1944 tq->tq_maxalloc = maxalloc;
1945 tq->tq_nbuckets = bsize;
1946 tq->tq_proc = proc;
1947 tq->tq_pri = pri;
1948 tq->tq_DC = dc;
1949 list_link_init(&tq->tq_cpupct_link);
1950
1951 if (max_nthreads > 1)
1952 tq->tq_threadlist = kmem_alloc(
1953 sizeof (kthread_t *) * max_nthreads, KM_SLEEP);
1954
1955 mutex_enter(&tq->tq_lock);
1956 if (flags & TASKQ_PREPOPULATE) {
1957 while (minalloc-- > 0)
1958 taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
1959 }
1960
1961 /*
1962 * Before we start creating threads for this taskq, take a
1963 * zone hold so the zone can't go away before taskq_destroy
1964 * makes sure all the taskq threads are gone. This hold is
1965 * similar in purpose to those taken by zthread_create().
1966 */
1967 zone_hold(tq->tq_proc->p_zone);
1968
1969 /*
1970 * Create the first thread, which will create any other threads
1971 * necessary. taskq_thread_create will not return until we have
1972 * enough threads to be able to process requests.
1973 */
1974 taskq_thread_create(tq);
1975 mutex_exit(&tq->tq_lock);
1976
1977 if (flags & TASKQ_DYNAMIC) {
1978 taskq_bucket_t *bucket = kmem_zalloc(sizeof (taskq_bucket_t) *
1979 bsize, KM_SLEEP);
1980 int b_id;
1981
1982 tq->tq_buckets = bucket;
1983
1984 /* Initialize each bucket */
1985 for (b_id = 0; b_id < bsize; b_id++, bucket++) {
1986 mutex_init(&bucket->tqbucket_lock, NULL, MUTEX_DEFAULT,
1987 NULL);
1988 cv_init(&bucket->tqbucket_cv, NULL, CV_DEFAULT, NULL);
1989 bucket->tqbucket_taskq = tq;
1990 bucket->tqbucket_freelist.tqent_next =
1991 bucket->tqbucket_freelist.tqent_prev =
1992 &bucket->tqbucket_freelist;
1993 if (flags & TASKQ_PREPOPULATE)
1994 taskq_bucket_extend(bucket);
1995 }
1996 }
1997
1998 /*
1999 * Install kstats.
2000 * We have two cases:
2001 * 1) Instance is provided to taskq_create_instance(). In this case it
2002 * should be >= 0 and we use it.
2003 *
2004 * 2) Instance is not provided and is automatically generated
2005 */
2006 if (flags & TASKQ_NOINSTANCE) {
2007 instance = tq->tq_instance =
2008 (int)(uintptr_t)vmem_alloc(taskq_id_arena, 1, VM_SLEEP);
2009 }
2010
2011 if (flags & TASKQ_DYNAMIC) {
2012 if ((tq->tq_kstat = kstat_create("unix", instance,
2013 tq->tq_name, "taskq_d", KSTAT_TYPE_NAMED,
2014 sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
2015 KSTAT_FLAG_VIRTUAL)) != NULL) {
2016 tq->tq_kstat->ks_lock = &taskq_d_kstat_lock;
2017 tq->tq_kstat->ks_data = &taskq_d_kstat;
2018 tq->tq_kstat->ks_update = taskq_d_kstat_update;
2019 tq->tq_kstat->ks_private = tq;
2020 kstat_install(tq->tq_kstat);
2021 }
2022 } else {
2023 if ((tq->tq_kstat = kstat_create("unix", instance, tq->tq_name,
2024 "taskq", KSTAT_TYPE_NAMED,
2025 sizeof (taskq_kstat) / sizeof (kstat_named_t),
2026 KSTAT_FLAG_VIRTUAL)) != NULL) {
2027 tq->tq_kstat->ks_lock = &taskq_kstat_lock;
2028 tq->tq_kstat->ks_data = &taskq_kstat;
2029 tq->tq_kstat->ks_update = taskq_kstat_update;
2030 tq->tq_kstat->ks_private = tq;
2031 kstat_install(tq->tq_kstat);
2032 }
2033 }
2034
2035 return (tq);
2036 }
2037
2038 /*
2039 * taskq_destroy().
2040 *
2041 * Assumes: by the time taskq_destroy is called no one will use this task queue
2042 * in any way and no one will try to dispatch entries in it.
2043 */
2044 void
2045 taskq_destroy(taskq_t *tq)
2046 {
2047 taskq_bucket_t *b = tq->tq_buckets;
2048 int bid = 0;
2049
2050 ASSERT(! (tq->tq_flags & TASKQ_CPR_SAFE));
2051
2052 /*
2053 * Destroy kstats.
2054 */
2055 if (tq->tq_kstat != NULL) {
2056 kstat_delete(tq->tq_kstat);
2057 tq->tq_kstat = NULL;
2058 }
2059
2060 /*
2061 * Destroy instance if needed.
2062 */
2063 if (tq->tq_flags & TASKQ_NOINSTANCE) {
2064 vmem_free(taskq_id_arena, (void *)(uintptr_t)(tq->tq_instance),
2065 1);
2066 tq->tq_instance = 0;
2067 }
2068
2069 /*
2070 * Unregister from the cpupct list.
2071 */
2072 if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
2073 taskq_cpupct_remove(tq);
2074 }
2075
2076 /*
2077 * Wait for any pending entries to complete.
2078 */
2079 taskq_wait(tq);
2080
2081 mutex_enter(&tq->tq_lock);
2082 ASSERT((tq->tq_task.tqent_next == &tq->tq_task) &&
2083 (tq->tq_active == 0));
2084
2085 /* notify all the threads that they need to exit */
2086 tq->tq_nthreads_target = 0;
2087
2088 tq->tq_flags |= TASKQ_CHANGING;
2089 cv_broadcast(&tq->tq_dispatch_cv);
2090 cv_broadcast(&tq->tq_exit_cv);
2091
2092 while (tq->tq_nthreads != 0)
2093 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
2094
2095 if (tq->tq_nthreads_max != 1)
2096 kmem_free(tq->tq_threadlist, sizeof (kthread_t *) *
2097 tq->tq_nthreads_max);
2098
2099 tq->tq_minalloc = 0;
2100 while (tq->tq_nalloc != 0)
2101 taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
2102
2103 mutex_exit(&tq->tq_lock);
2104
2105 /*
2106 * Mark each bucket as closing and wakeup all sleeping threads.
2107 */
2108 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2109 taskq_ent_t *tqe;
2110
2111 mutex_enter(&b->tqbucket_lock);
2112
2113 b->tqbucket_flags |= TQBUCKET_CLOSE;
2114 /* Wakeup all sleeping threads */
2115
2116 for (tqe = b->tqbucket_freelist.tqent_next;
2117 tqe != &b->tqbucket_freelist; tqe = tqe->tqent_next)
2118 cv_signal(&tqe->tqent_cv);
2119
2120 ASSERT(b->tqbucket_nalloc == 0);
2121
2122 /*
2123 * At this point we waited for all pending jobs to complete (in
2124 * both the task queue and the bucket and no new jobs should
2125 * arrive. Wait for all threads to die.
2126 */
2127 while (b->tqbucket_nfree > 0)
2128 cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
2129 mutex_exit(&b->tqbucket_lock);
2130 mutex_destroy(&b->tqbucket_lock);
2131 cv_destroy(&b->tqbucket_cv);
2132 }
2133
2134 if (tq->tq_buckets != NULL) {
2135 ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2136 kmem_free(tq->tq_buckets,
2137 sizeof (taskq_bucket_t) * tq->tq_nbuckets);
2138
2139 /* Cleanup fields before returning tq to the cache */
2140 tq->tq_buckets = NULL;
2141 tq->tq_tcreates = 0;
2142 tq->tq_tdeaths = 0;
2143 } else {
2144 ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
2145 }
2146
2147 /*
2148 * Now that all the taskq threads are gone, we can
2149 * drop the zone hold taken in taskq_create_common
2150 */
2151 zone_rele(tq->tq_proc->p_zone);
2152
2153 tq->tq_threads_ncpus_pct = 0;
2154 tq->tq_totaltime = 0;
2155 tq->tq_tasks = 0;
2156 tq->tq_maxtasks = 0;
2157 tq->tq_executed = 0;
2158 kmem_cache_free(taskq_cache, tq);
2159 }
2160
2161 /*
2162 * Extend a bucket with a new entry on the free list and attach a worker thread
2163 * to it.
2164 *
2165 * Argument: pointer to the bucket.
2166 *
2167 * This function may quietly fail. It is only used by taskq_dispatch() which
2168 * handles such failures properly.
2169 */
2170 static void
2171 taskq_bucket_extend(void *arg)
2172 {
2173 taskq_ent_t *tqe;
2174 taskq_bucket_t *b = (taskq_bucket_t *)arg;
2175 taskq_t *tq = b->tqbucket_taskq;
2176 int nthreads;
2177
2178 mutex_enter(&tq->tq_lock);
2179
2180 if (! ENOUGH_MEMORY()) {
2181 tq->tq_nomem++;
2182 mutex_exit(&tq->tq_lock);
2183 return;
2184 }
2185
2186 /*
2187 * Observe global taskq limits on the number of threads.
2188 */
2189 if (tq->tq_tcreates++ - tq->tq_tdeaths > tq->tq_maxsize) {
2190 tq->tq_tcreates--;
2191 mutex_exit(&tq->tq_lock);
2192 return;
2193 }
2194 mutex_exit(&tq->tq_lock);
2195
2196 tqe = kmem_cache_alloc(taskq_ent_cache, KM_NOSLEEP);
2197
2198 if (tqe == NULL) {
2199 mutex_enter(&tq->tq_lock);
2200 tq->tq_nomem++;
2201 tq->tq_tcreates--;
2202 mutex_exit(&tq->tq_lock);
2203 return;
2204 }
2205
2206 ASSERT(tqe->tqent_thread == NULL);
2207
2208 tqe->tqent_un.tqent_bucket = b;
2209
2210 /*
2211 * Create a thread in a TS_STOPPED state first. If it is successfully
2212 * created, place the entry on the free list and start the thread.
2213 */
2214 tqe->tqent_thread = thread_create(NULL, 0, taskq_d_thread, tqe,
2215 0, tq->tq_proc, TS_STOPPED, tq->tq_pri);
2216
2217 /*
2218 * Once the entry is ready, link it to the the bucket free list.
2219 */
2220 mutex_enter(&b->tqbucket_lock);
2221 tqe->tqent_func = NULL;
2222 TQ_APPEND(b->tqbucket_freelist, tqe);
2223 b->tqbucket_nfree++;
2224 TQ_STAT(b, tqs_tcreates);
2225
2226 #if TASKQ_STATISTIC
2227 nthreads = b->tqbucket_stat.tqs_tcreates -
2228 b->tqbucket_stat.tqs_tdeaths;
2229 b->tqbucket_stat.tqs_maxthreads = MAX(nthreads,
2230 b->tqbucket_stat.tqs_maxthreads);
2231 #endif
2232
2233 mutex_exit(&b->tqbucket_lock);
2234 /*
2235 * Start the stopped thread.
2236 */
2237 thread_lock(tqe->tqent_thread);
2238 tqe->tqent_thread->t_taskq = tq;
2239 tqe->tqent_thread->t_schedflag |= TS_ALLSTART;
2240 setrun_locked(tqe->tqent_thread);
2241 thread_unlock(tqe->tqent_thread);
2242 }
2243
2244 static int
2245 taskq_kstat_update(kstat_t *ksp, int rw)
2246 {
2247 struct taskq_kstat *tqsp = &taskq_kstat;
2248 taskq_t *tq = ksp->ks_private;
2249
2250 if (rw == KSTAT_WRITE)
2251 return (EACCES);
2252
2253 tqsp->tq_pid.value.ui64 = tq->tq_proc->p_pid;
2254 tqsp->tq_tasks.value.ui64 = tq->tq_tasks;
2255 tqsp->tq_executed.value.ui64 = tq->tq_executed;
2256 tqsp->tq_maxtasks.value.ui64 = tq->tq_maxtasks;
2257 tqsp->tq_totaltime.value.ui64 = tq->tq_totaltime;
2258 tqsp->tq_nactive.value.ui64 = tq->tq_active;
2259 tqsp->tq_nalloc.value.ui64 = tq->tq_nalloc;
2260 tqsp->tq_pri.value.ui64 = tq->tq_pri;
2261 tqsp->tq_nthreads.value.ui64 = tq->tq_nthreads;
2262 tqsp->tq_nomem.value.ui64 = tq->tq_nomem;
2263 return (0);
2264 }
2265
2266 static int
2267 taskq_d_kstat_update(kstat_t *ksp, int rw)
2268 {
2269 struct taskq_d_kstat *tqsp = &taskq_d_kstat;
2270 taskq_t *tq = ksp->ks_private;
2271 taskq_bucket_t *b = tq->tq_buckets;
2272 int bid = 0;
2273
2274 if (rw == KSTAT_WRITE)
2275 return (EACCES);
2276
2277 ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2278
2279 tqsp->tqd_btasks.value.ui64 = tq->tq_tasks;
2280 tqsp->tqd_bexecuted.value.ui64 = tq->tq_executed;
2281 tqsp->tqd_bmaxtasks.value.ui64 = tq->tq_maxtasks;
2282 tqsp->tqd_bnalloc.value.ui64 = tq->tq_nalloc;
2283 tqsp->tqd_bnactive.value.ui64 = tq->tq_active;
2284 tqsp->tqd_btotaltime.value.ui64 = tq->tq_totaltime;
2285 tqsp->tqd_pri.value.ui64 = tq->tq_pri;
2286 tqsp->tqd_nomem.value.ui64 = tq->tq_nomem;
2287
2288 tqsp->tqd_hits.value.ui64 = 0;
2289 tqsp->tqd_misses.value.ui64 = 0;
2290 tqsp->tqd_overflows.value.ui64 = 0;
2291 tqsp->tqd_tcreates.value.ui64 = 0;
2292 tqsp->tqd_tdeaths.value.ui64 = 0;
2293 tqsp->tqd_maxthreads.value.ui64 = 0;
2294 tqsp->tqd_nomem.value.ui64 = 0;
2295 tqsp->tqd_disptcreates.value.ui64 = 0;
2296 tqsp->tqd_totaltime.value.ui64 = 0;
2297 tqsp->tqd_nalloc.value.ui64 = 0;
2298 tqsp->tqd_nfree.value.ui64 = 0;
2299
2300 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2301 tqsp->tqd_hits.value.ui64 += b->tqbucket_stat.tqs_hits;
2302 tqsp->tqd_misses.value.ui64 += b->tqbucket_stat.tqs_misses;
2303 tqsp->tqd_overflows.value.ui64 += b->tqbucket_stat.tqs_overflow;
2304 tqsp->tqd_tcreates.value.ui64 += b->tqbucket_stat.tqs_tcreates;
2305 tqsp->tqd_tdeaths.value.ui64 += b->tqbucket_stat.tqs_tdeaths;
2306 tqsp->tqd_maxthreads.value.ui64 +=
2307 b->tqbucket_stat.tqs_maxthreads;
2308 tqsp->tqd_disptcreates.value.ui64 +=
2309 b->tqbucket_stat.tqs_disptcreates;
2310 tqsp->tqd_totaltime.value.ui64 += b->tqbucket_totaltime;
2311 tqsp->tqd_nalloc.value.ui64 += b->tqbucket_nalloc;
2312 tqsp->tqd_nfree.value.ui64 += b->tqbucket_nfree;
2313 }
2314 return (0);
2315 }