Print this page
OS-6363 system went to dark side of moon for ~467 seconds OS-6404 ARC reclaim should throttle its calls to arc_kmem_reap_now() Reviewed by: Bryan Cantrill <bryan@joyent.com> Reviewed by: Dan McDonald <danmcd@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/os/taskq.c
+++ new/usr/src/uts/common/os/taskq.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
|
↓ open down ↓ |
18 lines elided |
↑ open up ↑ |
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25
26 26 /*
27 27 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
28 28 * Copyright (c) 2017 by Delphix. All rights reserved.
29 + * Copyright (c) 2017, Joyent, Inc.
29 30 */
30 31
31 32 /*
32 33 * Kernel task queues: general-purpose asynchronous task scheduling.
33 34 *
34 35 * A common problem in kernel programming is the need to schedule tasks
35 36 * to be performed later, by another thread. There are several reasons
36 37 * you may want or need to do this:
37 38 *
38 39 * (1) The task isn't time-critical, but your current code path is.
39 40 *
40 41 * (2) The task may require grabbing locks that you already hold.
41 42 *
42 43 * (3) The task may need to block (e.g. to wait for memory), but you
43 44 * cannot block in your current context.
44 45 *
45 46 * (4) Your code path can't complete because of some condition, but you can't
46 47 * sleep or fail, so you queue the task for later execution when condition
47 48 * disappears.
48 49 *
49 50 * (5) You just want a simple way to launch multiple tasks in parallel.
50 51 *
51 52 * Task queues provide such a facility. In its simplest form (used when
52 53 * performance is not a critical consideration) a task queue consists of a
53 54 * single list of tasks, together with one or more threads to service the
54 55 * list. There are some cases when this simple queue is not sufficient:
55 56 *
56 57 * (1) The task queues are very hot and there is a need to avoid data and lock
57 58 * contention over global resources.
58 59 *
59 60 * (2) Some tasks may depend on other tasks to complete, so they can't be put in
60 61 * the same list managed by the same thread.
61 62 *
62 63 * (3) Some tasks may block for a long time, and this should not block other
63 64 * tasks in the queue.
64 65 *
65 66 * To provide useful service in such cases we define a "dynamic task queue"
66 67 * which has an individual thread for each of the tasks. These threads are
67 68 * dynamically created as they are needed and destroyed when they are not in
68 69 * use. The API for managing task pools is the same as for managing task queues
69 70 * with the exception of a taskq creation flag TASKQ_DYNAMIC which tells that
70 71 * dynamic task pool behavior is desired.
71 72 *
72 73 * Dynamic task queues may also place tasks in the normal queue (called "backing
73 74 * queue") when task pool runs out of resources. Users of task queues may
74 75 * disallow such queued scheduling by specifying TQ_NOQUEUE in the dispatch
75 76 * flags.
76 77 *
77 78 * The backing task queue is also used for scheduling internal tasks needed for
78 79 * dynamic task queue maintenance.
79 80 *
80 81 * INTERFACES ==================================================================
81 82 *
82 83 * taskq_t *taskq_create(name, nthreads, pri, minalloc, maxalloc, flags);
83 84 *
84 85 * Create a taskq with specified properties.
85 86 * Possible 'flags':
86 87 *
87 88 * TASKQ_DYNAMIC: Create task pool for task management. If this flag is
88 89 * specified, 'nthreads' specifies the maximum number of threads in
89 90 * the task queue. Task execution order for dynamic task queues is
90 91 * not predictable.
91 92 *
92 93 * If this flag is not specified (default case) a
93 94 * single-list task queue is created with 'nthreads' threads
94 95 * servicing it. Entries in this queue are managed by
95 96 * taskq_ent_alloc() and taskq_ent_free() which try to keep the
96 97 * task population between 'minalloc' and 'maxalloc', but the
97 98 * latter limit is only advisory for TQ_SLEEP dispatches and the
98 99 * former limit is only advisory for TQ_NOALLOC dispatches. If
99 100 * TASKQ_PREPOPULATE is set in 'flags', the taskq will be
100 101 * prepopulated with 'minalloc' task structures.
101 102 *
102 103 * Since non-DYNAMIC taskqs are queues, tasks are guaranteed to be
103 104 * executed in the order they are scheduled if nthreads == 1.
104 105 * If nthreads > 1, task execution order is not predictable.
105 106 *
106 107 * TASKQ_PREPOPULATE: Prepopulate task queue with threads.
107 108 * Also prepopulate the task queue with 'minalloc' task structures.
108 109 *
109 110 * TASKQ_THREADS_CPU_PCT: This flag specifies that 'nthreads' should be
110 111 * interpreted as a percentage of the # of online CPUs on the
111 112 * system. The taskq subsystem will automatically adjust the
112 113 * number of threads in the taskq in response to CPU online
113 114 * and offline events, to keep the ratio. nthreads must be in
114 115 * the range [0,100].
115 116 *
116 117 * The calculation used is:
117 118 *
118 119 * MAX((ncpus_online * percentage)/100, 1)
119 120 *
120 121 * This flag is not supported for DYNAMIC task queues.
121 122 * This flag is not compatible with TASKQ_CPR_SAFE.
122 123 *
123 124 * TASKQ_CPR_SAFE: This flag specifies that users of the task queue will
124 125 * use their own protocol for handling CPR issues. This flag is not
125 126 * supported for DYNAMIC task queues. This flag is not compatible
126 127 * with TASKQ_THREADS_CPU_PCT.
127 128 *
128 129 * The 'pri' field specifies the default priority for the threads that
129 130 * service all scheduled tasks.
130 131 *
131 132 * taskq_t *taskq_create_instance(name, instance, nthreads, pri, minalloc,
132 133 * maxalloc, flags);
133 134 *
134 135 * Like taskq_create(), but takes an instance number (or -1 to indicate
135 136 * no instance).
136 137 *
137 138 * taskq_t *taskq_create_proc(name, nthreads, pri, minalloc, maxalloc, proc,
138 139 * flags);
139 140 *
140 141 * Like taskq_create(), but creates the taskq threads in the specified
141 142 * system process. If proc != &p0, this must be called from a thread
142 143 * in that process.
143 144 *
144 145 * taskq_t *taskq_create_sysdc(name, nthreads, minalloc, maxalloc, proc,
145 146 * dc, flags);
146 147 *
147 148 * Like taskq_create_proc(), but the taskq threads will use the
148 149 * System Duty Cycle (SDC) scheduling class with a duty cycle of dc.
149 150 *
150 151 * void taskq_destroy(tap):
151 152 *
152 153 * Waits for any scheduled tasks to complete, then destroys the taskq.
153 154 * Caller should guarantee that no new tasks are scheduled in the closing
154 155 * taskq.
155 156 *
156 157 * taskqid_t taskq_dispatch(tq, func, arg, flags):
157 158 *
158 159 * Dispatches the task "func(arg)" to taskq. The 'flags' indicates whether
159 160 * the caller is willing to block for memory. The function returns an
160 161 * opaque value which is zero iff dispatch fails. If flags is TQ_NOSLEEP
161 162 * or TQ_NOALLOC and the task can't be dispatched, taskq_dispatch() fails
162 163 * and returns (taskqid_t)0.
163 164 *
164 165 * ASSUMES: func != NULL.
165 166 *
166 167 * Possible flags:
167 168 * TQ_NOSLEEP: Do not wait for resources; may fail.
168 169 *
169 170 * TQ_NOALLOC: Do not allocate memory; may fail. May only be used with
170 171 * non-dynamic task queues.
171 172 *
172 173 * TQ_NOQUEUE: Do not enqueue a task if it can't dispatch it due to
173 174 * lack of available resources and fail. If this flag is not
174 175 * set, and the task pool is exhausted, the task may be scheduled
175 176 * in the backing queue. This flag may ONLY be used with dynamic
176 177 * task queues.
177 178 *
178 179 * NOTE: This flag should always be used when a task queue is used
179 180 * for tasks that may depend on each other for completion.
180 181 * Enqueueing dependent tasks may create deadlocks.
181 182 *
182 183 * TQ_SLEEP: May block waiting for resources. May still fail for
183 184 * dynamic task queues if TQ_NOQUEUE is also specified, otherwise
184 185 * always succeed.
185 186 *
186 187 * TQ_FRONT: Puts the new task at the front of the queue. Be careful.
187 188 *
188 189 * NOTE: Dynamic task queues are much more likely to fail in
189 190 * taskq_dispatch() (especially if TQ_NOQUEUE was specified), so it
190 191 * is important to have backup strategies handling such failures.
191 192 *
192 193 * void taskq_dispatch_ent(tq, func, arg, flags, tqent)
193 194 *
194 195 * This is a light-weight form of taskq_dispatch(), that uses a
195 196 * preallocated taskq_ent_t structure for scheduling. As a
196 197 * result, it does not perform allocations and cannot ever fail.
197 198 * Note especially that it cannot be used with TASKQ_DYNAMIC
198 199 * taskqs. The memory for the tqent must not be modified or used
199 200 * until the function (func) is called. (However, func itself
200 201 * may safely modify or free this memory, once it is called.)
201 202 * Note that the taskq framework will NOT free this memory.
202 203 *
203 204 * void taskq_wait(tq):
204 205 *
205 206 * Waits for all previously scheduled tasks to complete.
206 207 *
207 208 * NOTE: It does not stop any new task dispatches.
208 209 * Do NOT call taskq_wait() from a task: it will cause deadlock.
209 210 *
210 211 * void taskq_suspend(tq)
211 212 *
212 213 * Suspend all task execution. Tasks already scheduled for a dynamic task
213 214 * queue will still be executed, but all new scheduled tasks will be
214 215 * suspended until taskq_resume() is called.
215 216 *
216 217 * int taskq_suspended(tq)
217 218 *
218 219 * Returns 1 if taskq is suspended and 0 otherwise. It is intended to
219 220 * ASSERT that the task queue is suspended.
220 221 *
221 222 * void taskq_resume(tq)
222 223 *
223 224 * Resume task queue execution.
224 225 *
225 226 * int taskq_member(tq, thread)
226 227 *
227 228 * Returns 1 if 'thread' belongs to taskq 'tq' and 0 otherwise. The
228 229 * intended use is to ASSERT that a given function is called in taskq
229 230 * context only.
230 231 *
231 232 * system_taskq
232 233 *
233 234 * Global system-wide dynamic task queue for common uses. It may be used by
234 235 * any subsystem that needs to schedule tasks and does not need to manage
235 236 * its own task queues. It is initialized quite early during system boot.
236 237 *
237 238 * IMPLEMENTATION ==============================================================
238 239 *
239 240 * This is schematic representation of the task queue structures.
240 241 *
241 242 * taskq:
242 243 * +-------------+
243 244 * | tq_lock | +---< taskq_ent_free()
244 245 * +-------------+ |
245 246 * |... | | tqent: tqent:
246 247 * +-------------+ | +------------+ +------------+
247 248 * | tq_freelist |-->| tqent_next |--> ... ->| tqent_next |
248 249 * +-------------+ +------------+ +------------+
249 250 * |... | | ... | | ... |
250 251 * +-------------+ +------------+ +------------+
251 252 * | tq_task | |
252 253 * | | +-------------->taskq_ent_alloc()
253 254 * +--------------------------------------------------------------------------+
254 255 * | | | tqent tqent |
255 256 * | +---------------------+ +--> +------------+ +--> +------------+ |
256 257 * | | ... | | | func, arg | | | func, arg | |
257 258 * +>+---------------------+ <---|-+ +------------+ <---|-+ +------------+ |
258 259 * | tq_taskq.tqent_next | ----+ | | tqent_next | --->+ | | tqent_next |--+
259 260 * +---------------------+ | +------------+ ^ | +------------+
260 261 * +-| tq_task.tqent_prev | +--| tqent_prev | | +--| tqent_prev | ^
261 262 * | +---------------------+ +------------+ | +------------+ |
262 263 * | |... | | ... | | | ... | |
263 264 * | +---------------------+ +------------+ | +------------+ |
264 265 * | ^ | |
265 266 * | | | |
266 267 * +--------------------------------------+--------------+ TQ_APPEND() -+
267 268 * | | |
268 269 * |... | taskq_thread()-----+
269 270 * +-------------+
270 271 * | tq_buckets |--+-------> [ NULL ] (for regular task queues)
271 272 * +-------------+ |
272 273 * | DYNAMIC TASK QUEUES:
273 274 * |
274 275 * +-> taskq_bucket[nCPU] taskq_bucket_dispatch()
275 276 * +-------------------+ ^
276 277 * +--->| tqbucket_lock | |
277 278 * | +-------------------+ +--------+ +--------+
278 279 * | | tqbucket_freelist |-->| tqent |-->...| tqent | ^
279 280 * | +-------------------+<--+--------+<--...+--------+ |
280 281 * | | ... | | thread | | thread | |
281 282 * | +-------------------+ +--------+ +--------+ |
282 283 * | +-------------------+ |
283 284 * taskq_dispatch()--+--->| tqbucket_lock | TQ_APPEND()------+
284 285 * TQ_HASH() | +-------------------+ +--------+ +--------+
285 286 * | | tqbucket_freelist |-->| tqent |-->...| tqent |
286 287 * | +-------------------+<--+--------+<--...+--------+
287 288 * | | ... | | thread | | thread |
288 289 * | +-------------------+ +--------+ +--------+
289 290 * +---> ...
290 291 *
291 292 *
292 293 * Task queues use tq_task field to link new entry in the queue. The queue is a
293 294 * circular doubly-linked list. Entries are put in the end of the list with
294 295 * TQ_APPEND() and processed from the front of the list by taskq_thread() in
295 296 * FIFO order. Task queue entries are cached in the free list managed by
296 297 * taskq_ent_alloc() and taskq_ent_free() functions.
297 298 *
298 299 * All threads used by task queues mark t_taskq field of the thread to
299 300 * point to the task queue.
300 301 *
301 302 * Taskq Thread Management -----------------------------------------------------
302 303 *
303 304 * Taskq's non-dynamic threads are managed with several variables and flags:
304 305 *
305 306 * * tq_nthreads - The number of threads in taskq_thread() for the
306 307 * taskq.
307 308 *
308 309 * * tq_active - The number of threads not waiting on a CV in
309 310 * taskq_thread(); includes newly created threads
310 311 * not yet counted in tq_nthreads.
311 312 *
312 313 * * tq_nthreads_target
313 314 * - The number of threads desired for the taskq.
314 315 *
315 316 * * tq_flags & TASKQ_CHANGING
316 317 * - Indicates that tq_nthreads != tq_nthreads_target.
317 318 *
318 319 * * tq_flags & TASKQ_THREAD_CREATED
319 320 * - Indicates that a thread is being created in the taskq.
320 321 *
321 322 * During creation, tq_nthreads and tq_active are set to 0, and
322 323 * tq_nthreads_target is set to the number of threads desired. The
323 324 * TASKQ_CHANGING flag is set, and taskq_thread_create() is called to
324 325 * create the first thread. taskq_thread_create() increments tq_active,
325 326 * sets TASKQ_THREAD_CREATED, and creates the new thread.
326 327 *
327 328 * Each thread starts in taskq_thread(), clears the TASKQ_THREAD_CREATED
328 329 * flag, and increments tq_nthreads. It stores the new value of
329 330 * tq_nthreads as its "thread_id", and stores its thread pointer in the
330 331 * tq_threadlist at the (thread_id - 1). We keep the thread_id space
331 332 * densely packed by requiring that only the largest thread_id can exit during
332 333 * normal adjustment. The exception is during the destruction of the
333 334 * taskq; once tq_nthreads_target is set to zero, no new threads will be created
334 335 * for the taskq queue, so every thread can exit without any ordering being
335 336 * necessary.
336 337 *
337 338 * Threads will only process work if their thread id is <= tq_nthreads_target.
338 339 *
339 340 * When TASKQ_CHANGING is set, threads will check the current thread target
340 341 * whenever they wake up, and do whatever they can to apply its effects.
341 342 *
342 343 * TASKQ_THREAD_CPU_PCT --------------------------------------------------------
343 344 *
344 345 * When a taskq is created with TASKQ_THREAD_CPU_PCT, we store their requested
345 346 * percentage in tq_threads_ncpus_pct, start them off with the correct thread
346 347 * target, and add them to the taskq_cpupct_list for later adjustment.
347 348 *
348 349 * We register taskq_cpu_setup() to be called whenever a CPU changes state. It
349 350 * walks the list of TASKQ_THREAD_CPU_PCT taskqs, adjusts their nthread_target
350 351 * if need be, and wakes up all of the threads to process the change.
351 352 *
352 353 * Dynamic Task Queues Implementation ------------------------------------------
353 354 *
354 355 * For a dynamic task queues there is a 1-to-1 mapping between a thread and
355 356 * taskq_ent_structure. Each entry is serviced by its own thread and each thread
356 357 * is controlled by a single entry.
357 358 *
358 359 * Entries are distributed over a set of buckets. To avoid using modulo
359 360 * arithmetics the number of buckets is 2^n and is determined as the nearest
360 361 * power of two roundown of the number of CPUs in the system. Tunable
361 362 * variable 'taskq_maxbuckets' limits the maximum number of buckets. Each entry
362 363 * is attached to a bucket for its lifetime and can't migrate to other buckets.
363 364 *
364 365 * Entries that have scheduled tasks are not placed in any list. The dispatch
365 366 * function sets their "func" and "arg" fields and signals the corresponding
366 367 * thread to execute the task. Once the thread executes the task it clears the
367 368 * "func" field and places an entry on the bucket cache of free entries pointed
368 369 * by "tqbucket_freelist" field. ALL entries on the free list should have "func"
369 370 * field equal to NULL. The free list is a circular doubly-linked list identical
370 371 * in structure to the tq_task list above, but entries are taken from it in LIFO
371 372 * order - the last freed entry is the first to be allocated. The
372 373 * taskq_bucket_dispatch() function gets the most recently used entry from the
373 374 * free list, sets its "func" and "arg" fields and signals a worker thread.
374 375 *
375 376 * After executing each task a per-entry thread taskq_d_thread() places its
376 377 * entry on the bucket free list and goes to a timed sleep. If it wakes up
377 378 * without getting new task it removes the entry from the free list and destroys
378 379 * itself. The thread sleep time is controlled by a tunable variable
379 380 * `taskq_thread_timeout'.
380 381 *
381 382 * There are various statistics kept in the bucket which allows for later
382 383 * analysis of taskq usage patterns. Also, a global copy of taskq creation and
383 384 * death statistics is kept in the global taskq data structure. Since thread
384 385 * creation and death happen rarely, updating such global data does not present
385 386 * a performance problem.
386 387 *
387 388 * NOTE: Threads are not bound to any CPU and there is absolutely no association
388 389 * between the bucket and actual thread CPU, so buckets are used only to
389 390 * split resources and reduce resource contention. Having threads attached
390 391 * to the CPU denoted by a bucket may reduce number of times the job
391 392 * switches between CPUs.
392 393 *
393 394 * Current algorithm creates a thread whenever a bucket has no free
394 395 * entries. It would be nice to know how many threads are in the running
395 396 * state and don't create threads if all CPUs are busy with existing
396 397 * tasks, but it is unclear how such strategy can be implemented.
397 398 *
398 399 * Currently buckets are created statically as an array attached to task
399 400 * queue. On some system with nCPUs < max_ncpus it may waste system
400 401 * memory. One solution may be allocation of buckets when they are first
401 402 * touched, but it is not clear how useful it is.
402 403 *
403 404 * SUSPEND/RESUME implementation -----------------------------------------------
404 405 *
405 406 * Before executing a task taskq_thread() (executing non-dynamic task
406 407 * queues) obtains taskq's thread lock as a reader. The taskq_suspend()
407 408 * function gets the same lock as a writer blocking all non-dynamic task
408 409 * execution. The taskq_resume() function releases the lock allowing
409 410 * taskq_thread to continue execution.
410 411 *
411 412 * For dynamic task queues, each bucket is marked as TQBUCKET_SUSPEND by
412 413 * taskq_suspend() function. After that taskq_bucket_dispatch() always
413 414 * fails, so that taskq_dispatch() will either enqueue tasks for a
414 415 * suspended backing queue or fail if TQ_NOQUEUE is specified in dispatch
415 416 * flags.
416 417 *
417 418 * NOTE: taskq_suspend() does not immediately block any tasks already
418 419 * scheduled for dynamic task queues. It only suspends new tasks
419 420 * scheduled after taskq_suspend() was called.
420 421 *
421 422 * taskq_member() function works by comparing a thread t_taskq pointer with
422 423 * the passed thread pointer.
423 424 *
424 425 * LOCKS and LOCK Hierarchy ----------------------------------------------------
425 426 *
426 427 * There are three locks used in task queues:
427 428 *
428 429 * 1) The taskq_t's tq_lock, protecting global task queue state.
429 430 *
430 431 * 2) Each per-CPU bucket has a lock for bucket management.
431 432 *
432 433 * 3) The global taskq_cpupct_lock, which protects the list of
433 434 * TASKQ_THREADS_CPU_PCT taskqs.
434 435 *
435 436 * If both (1) and (2) are needed, tq_lock should be taken *after* the bucket
436 437 * lock.
437 438 *
438 439 * If both (1) and (3) are needed, tq_lock should be taken *after*
439 440 * taskq_cpupct_lock.
440 441 *
441 442 * DEBUG FACILITIES ------------------------------------------------------------
442 443 *
443 444 * For DEBUG kernels it is possible to induce random failures to
444 445 * taskq_dispatch() function when it is given TQ_NOSLEEP argument. The value of
445 446 * taskq_dmtbf and taskq_smtbf tunables control the mean time between induced
446 447 * failures for dynamic and static task queues respectively.
447 448 *
448 449 * Setting TASKQ_STATISTIC to 0 will disable per-bucket statistics.
449 450 *
450 451 * TUNABLES --------------------------------------------------------------------
451 452 *
452 453 * system_taskq_size - Size of the global system_taskq.
453 454 * This value is multiplied by nCPUs to determine
454 455 * actual size.
455 456 * Default value: 64
456 457 *
457 458 * taskq_minimum_nthreads_max
458 459 * - Minimum size of the thread list for a taskq.
459 460 * Useful for testing different thread pool
460 461 * sizes by overwriting tq_nthreads_target.
461 462 *
462 463 * taskq_thread_timeout - Maximum idle time for taskq_d_thread()
463 464 * Default value: 5 minutes
464 465 *
465 466 * taskq_maxbuckets - Maximum number of buckets in any task queue
466 467 * Default value: 128
467 468 *
468 469 * taskq_search_depth - Maximum # of buckets searched for a free entry
469 470 * Default value: 4
470 471 *
471 472 * taskq_dmtbf - Mean time between induced dispatch failures
472 473 * for dynamic task queues.
473 474 * Default value: UINT_MAX (no induced failures)
474 475 *
475 476 * taskq_smtbf - Mean time between induced dispatch failures
476 477 * for static task queues.
477 478 * Default value: UINT_MAX (no induced failures)
478 479 *
479 480 * CONDITIONAL compilation -----------------------------------------------------
480 481 *
481 482 * TASKQ_STATISTIC - If set will enable bucket statistic (default).
482 483 *
483 484 */
484 485
485 486 #include <sys/taskq_impl.h>
486 487 #include <sys/thread.h>
487 488 #include <sys/proc.h>
488 489 #include <sys/kmem.h>
489 490 #include <sys/vmem.h>
490 491 #include <sys/callb.h>
491 492 #include <sys/class.h>
492 493 #include <sys/systm.h>
493 494 #include <sys/cmn_err.h>
494 495 #include <sys/debug.h>
495 496 #include <sys/vmsystm.h> /* For throttlefree */
496 497 #include <sys/sysmacros.h>
497 498 #include <sys/cpuvar.h>
498 499 #include <sys/cpupart.h>
499 500 #include <sys/sdt.h>
500 501 #include <sys/sysdc.h>
501 502 #include <sys/note.h>
502 503
503 504 static kmem_cache_t *taskq_ent_cache, *taskq_cache;
504 505
505 506 /*
506 507 * Pseudo instance numbers for taskqs without explicitly provided instance.
507 508 */
508 509 static vmem_t *taskq_id_arena;
509 510
510 511 /* Global system task queue for common use */
511 512 taskq_t *system_taskq;
512 513
513 514 /*
514 515 * Maximum number of entries in global system taskq is
515 516 * system_taskq_size * max_ncpus
516 517 */
517 518 #define SYSTEM_TASKQ_SIZE 64
518 519 int system_taskq_size = SYSTEM_TASKQ_SIZE;
519 520
520 521 /*
521 522 * Minimum size for tq_nthreads_max; useful for those who want to play around
522 523 * with increasing a taskq's tq_nthreads_target.
523 524 */
524 525 int taskq_minimum_nthreads_max = 1;
525 526
526 527 /*
527 528 * We want to ensure that when taskq_create() returns, there is at least
528 529 * one thread ready to handle requests. To guarantee this, we have to wait
529 530 * for the second thread, since the first one cannot process requests until
530 531 * the second thread has been created.
531 532 */
532 533 #define TASKQ_CREATE_ACTIVE_THREADS 2
533 534
534 535 /* Maximum percentage allowed for TASKQ_THREADS_CPU_PCT */
535 536 #define TASKQ_CPUPCT_MAX_PERCENT 1000
536 537 int taskq_cpupct_max_percent = TASKQ_CPUPCT_MAX_PERCENT;
537 538
538 539 /*
539 540 * Dynamic task queue threads that don't get any work within
540 541 * taskq_thread_timeout destroy themselves
541 542 */
542 543 #define TASKQ_THREAD_TIMEOUT (60 * 5)
543 544 int taskq_thread_timeout = TASKQ_THREAD_TIMEOUT;
544 545
545 546 #define TASKQ_MAXBUCKETS 128
546 547 int taskq_maxbuckets = TASKQ_MAXBUCKETS;
547 548
548 549 /*
549 550 * When a bucket has no available entries another buckets are tried.
550 551 * taskq_search_depth parameter limits the amount of buckets that we search
551 552 * before failing. This is mostly useful in systems with many CPUs where we may
552 553 * spend too much time scanning busy buckets.
553 554 */
554 555 #define TASKQ_SEARCH_DEPTH 4
555 556 int taskq_search_depth = TASKQ_SEARCH_DEPTH;
556 557
557 558 /*
558 559 * Hashing function: mix various bits of x. May be pretty much anything.
559 560 */
560 561 #define TQ_HASH(x) ((x) ^ ((x) >> 11) ^ ((x) >> 17) ^ ((x) ^ 27))
561 562
562 563 /*
563 564 * We do not create any new threads when the system is low on memory and start
564 565 * throttling memory allocations. The following macro tries to estimate such
565 566 * condition.
566 567 */
567 568 #define ENOUGH_MEMORY() (freemem > throttlefree)
568 569
569 570 /*
570 571 * Static functions.
571 572 */
572 573 static taskq_t *taskq_create_common(const char *, int, int, pri_t, int,
573 574 int, proc_t *, uint_t, uint_t);
574 575 static void taskq_thread(void *);
575 576 static void taskq_d_thread(taskq_ent_t *);
576 577 static void taskq_bucket_extend(void *);
577 578 static int taskq_constructor(void *, void *, int);
578 579 static void taskq_destructor(void *, void *);
579 580 static int taskq_ent_constructor(void *, void *, int);
580 581 static void taskq_ent_destructor(void *, void *);
581 582 static taskq_ent_t *taskq_ent_alloc(taskq_t *, int);
582 583 static void taskq_ent_free(taskq_t *, taskq_ent_t *);
583 584 static int taskq_ent_exists(taskq_t *, task_func_t, void *);
584 585 static taskq_ent_t *taskq_bucket_dispatch(taskq_bucket_t *, task_func_t,
585 586 void *);
586 587
587 588 /*
588 589 * Task queues kstats.
589 590 */
590 591 struct taskq_kstat {
591 592 kstat_named_t tq_pid;
592 593 kstat_named_t tq_tasks;
593 594 kstat_named_t tq_executed;
594 595 kstat_named_t tq_maxtasks;
595 596 kstat_named_t tq_totaltime;
596 597 kstat_named_t tq_nalloc;
597 598 kstat_named_t tq_nactive;
598 599 kstat_named_t tq_pri;
599 600 kstat_named_t tq_nthreads;
600 601 kstat_named_t tq_nomem;
601 602 } taskq_kstat = {
602 603 { "pid", KSTAT_DATA_UINT64 },
603 604 { "tasks", KSTAT_DATA_UINT64 },
604 605 { "executed", KSTAT_DATA_UINT64 },
605 606 { "maxtasks", KSTAT_DATA_UINT64 },
606 607 { "totaltime", KSTAT_DATA_UINT64 },
607 608 { "nalloc", KSTAT_DATA_UINT64 },
608 609 { "nactive", KSTAT_DATA_UINT64 },
609 610 { "priority", KSTAT_DATA_UINT64 },
610 611 { "threads", KSTAT_DATA_UINT64 },
611 612 { "nomem", KSTAT_DATA_UINT64 },
612 613 };
613 614
614 615 struct taskq_d_kstat {
615 616 kstat_named_t tqd_pri;
616 617 kstat_named_t tqd_btasks;
617 618 kstat_named_t tqd_bexecuted;
618 619 kstat_named_t tqd_bmaxtasks;
619 620 kstat_named_t tqd_bnalloc;
620 621 kstat_named_t tqd_bnactive;
621 622 kstat_named_t tqd_btotaltime;
622 623 kstat_named_t tqd_hits;
623 624 kstat_named_t tqd_misses;
624 625 kstat_named_t tqd_overflows;
625 626 kstat_named_t tqd_tcreates;
626 627 kstat_named_t tqd_tdeaths;
627 628 kstat_named_t tqd_maxthreads;
628 629 kstat_named_t tqd_nomem;
629 630 kstat_named_t tqd_disptcreates;
630 631 kstat_named_t tqd_totaltime;
631 632 kstat_named_t tqd_nalloc;
632 633 kstat_named_t tqd_nfree;
633 634 } taskq_d_kstat = {
634 635 { "priority", KSTAT_DATA_UINT64 },
635 636 { "btasks", KSTAT_DATA_UINT64 },
636 637 { "bexecuted", KSTAT_DATA_UINT64 },
637 638 { "bmaxtasks", KSTAT_DATA_UINT64 },
638 639 { "bnalloc", KSTAT_DATA_UINT64 },
639 640 { "bnactive", KSTAT_DATA_UINT64 },
640 641 { "btotaltime", KSTAT_DATA_UINT64 },
641 642 { "hits", KSTAT_DATA_UINT64 },
642 643 { "misses", KSTAT_DATA_UINT64 },
643 644 { "overflows", KSTAT_DATA_UINT64 },
644 645 { "tcreates", KSTAT_DATA_UINT64 },
645 646 { "tdeaths", KSTAT_DATA_UINT64 },
646 647 { "maxthreads", KSTAT_DATA_UINT64 },
647 648 { "nomem", KSTAT_DATA_UINT64 },
648 649 { "disptcreates", KSTAT_DATA_UINT64 },
649 650 { "totaltime", KSTAT_DATA_UINT64 },
650 651 { "nalloc", KSTAT_DATA_UINT64 },
651 652 { "nfree", KSTAT_DATA_UINT64 },
652 653 };
653 654
654 655 static kmutex_t taskq_kstat_lock;
655 656 static kmutex_t taskq_d_kstat_lock;
656 657 static int taskq_kstat_update(kstat_t *, int);
657 658 static int taskq_d_kstat_update(kstat_t *, int);
658 659
659 660 /*
660 661 * List of all TASKQ_THREADS_CPU_PCT taskqs.
661 662 */
662 663 static list_t taskq_cpupct_list; /* protected by cpu_lock */
663 664
664 665 /*
665 666 * Collect per-bucket statistic when TASKQ_STATISTIC is defined.
666 667 */
667 668 #define TASKQ_STATISTIC 1
668 669
669 670 #if TASKQ_STATISTIC
670 671 #define TQ_STAT(b, x) b->tqbucket_stat.x++
671 672 #else
672 673 #define TQ_STAT(b, x)
673 674 #endif
674 675
675 676 /*
676 677 * Random fault injection.
677 678 */
678 679 uint_t taskq_random;
679 680 uint_t taskq_dmtbf = UINT_MAX; /* mean time between injected failures */
680 681 uint_t taskq_smtbf = UINT_MAX; /* mean time between injected failures */
681 682
682 683 /*
683 684 * TQ_NOSLEEP dispatches on dynamic task queues are always allowed to fail.
684 685 *
685 686 * TQ_NOSLEEP dispatches on static task queues can't arbitrarily fail because
686 687 * they could prepopulate the cache and make sure that they do not use more
687 688 * then minalloc entries. So, fault injection in this case insures that
688 689 * either TASKQ_PREPOPULATE is not set or there are more entries allocated
689 690 * than is specified by minalloc. TQ_NOALLOC dispatches are always allowed
690 691 * to fail, but for simplicity we treat them identically to TQ_NOSLEEP
691 692 * dispatches.
692 693 */
693 694 #ifdef DEBUG
694 695 #define TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag) \
695 696 taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
696 697 if ((flag & TQ_NOSLEEP) && \
697 698 taskq_random < 1771875 / taskq_dmtbf) { \
698 699 return (NULL); \
699 700 }
700 701
701 702 #define TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag) \
702 703 taskq_random = (taskq_random * 2416 + 374441) % 1771875;\
703 704 if ((flag & (TQ_NOSLEEP | TQ_NOALLOC)) && \
704 705 (!(tq->tq_flags & TASKQ_PREPOPULATE) || \
705 706 (tq->tq_nalloc > tq->tq_minalloc)) && \
706 707 (taskq_random < (1771875 / taskq_smtbf))) { \
707 708 mutex_exit(&tq->tq_lock); \
708 709 return (NULL); \
709 710 }
710 711 #else
711 712 #define TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flag)
712 713 #define TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flag)
713 714 #endif
714 715
715 716 #define IS_EMPTY(l) (((l).tqent_prev == (l).tqent_next) && \
716 717 ((l).tqent_prev == &(l)))
717 718
718 719 /*
719 720 * Append `tqe' in the end of the doubly-linked list denoted by l.
720 721 */
721 722 #define TQ_APPEND(l, tqe) { \
722 723 tqe->tqent_next = &l; \
723 724 tqe->tqent_prev = l.tqent_prev; \
724 725 tqe->tqent_next->tqent_prev = tqe; \
725 726 tqe->tqent_prev->tqent_next = tqe; \
726 727 }
727 728 /*
728 729 * Prepend 'tqe' to the beginning of l
729 730 */
730 731 #define TQ_PREPEND(l, tqe) { \
731 732 tqe->tqent_next = l.tqent_next; \
732 733 tqe->tqent_prev = &l; \
733 734 tqe->tqent_next->tqent_prev = tqe; \
734 735 tqe->tqent_prev->tqent_next = tqe; \
735 736 }
736 737
737 738 /*
738 739 * Schedule a task specified by func and arg into the task queue entry tqe.
739 740 */
740 741 #define TQ_DO_ENQUEUE(tq, tqe, func, arg, front) { \
741 742 ASSERT(MUTEX_HELD(&tq->tq_lock)); \
742 743 _NOTE(CONSTCOND) \
743 744 if (front) { \
744 745 TQ_PREPEND(tq->tq_task, tqe); \
745 746 } else { \
746 747 TQ_APPEND(tq->tq_task, tqe); \
747 748 } \
748 749 tqe->tqent_func = (func); \
749 750 tqe->tqent_arg = (arg); \
750 751 tq->tq_tasks++; \
751 752 if (tq->tq_tasks - tq->tq_executed > tq->tq_maxtasks) \
752 753 tq->tq_maxtasks = tq->tq_tasks - tq->tq_executed; \
753 754 cv_signal(&tq->tq_dispatch_cv); \
754 755 DTRACE_PROBE2(taskq__enqueue, taskq_t *, tq, taskq_ent_t *, tqe); \
755 756 }
756 757
757 758 #define TQ_ENQUEUE(tq, tqe, func, arg) \
758 759 TQ_DO_ENQUEUE(tq, tqe, func, arg, 0)
759 760
760 761 #define TQ_ENQUEUE_FRONT(tq, tqe, func, arg) \
761 762 TQ_DO_ENQUEUE(tq, tqe, func, arg, 1)
762 763
763 764 /*
764 765 * Do-nothing task which may be used to prepopulate thread caches.
765 766 */
766 767 /*ARGSUSED*/
767 768 void
768 769 nulltask(void *unused)
769 770 {
770 771 }
771 772
772 773 /*ARGSUSED*/
773 774 static int
774 775 taskq_constructor(void *buf, void *cdrarg, int kmflags)
775 776 {
776 777 taskq_t *tq = buf;
777 778
778 779 bzero(tq, sizeof (taskq_t));
779 780
780 781 mutex_init(&tq->tq_lock, NULL, MUTEX_DEFAULT, NULL);
781 782 rw_init(&tq->tq_threadlock, NULL, RW_DEFAULT, NULL);
782 783 cv_init(&tq->tq_dispatch_cv, NULL, CV_DEFAULT, NULL);
783 784 cv_init(&tq->tq_exit_cv, NULL, CV_DEFAULT, NULL);
784 785 cv_init(&tq->tq_wait_cv, NULL, CV_DEFAULT, NULL);
785 786 cv_init(&tq->tq_maxalloc_cv, NULL, CV_DEFAULT, NULL);
786 787
787 788 tq->tq_task.tqent_next = &tq->tq_task;
788 789 tq->tq_task.tqent_prev = &tq->tq_task;
789 790
790 791 return (0);
791 792 }
792 793
793 794 /*ARGSUSED*/
794 795 static void
795 796 taskq_destructor(void *buf, void *cdrarg)
796 797 {
797 798 taskq_t *tq = buf;
798 799
799 800 ASSERT(tq->tq_nthreads == 0);
800 801 ASSERT(tq->tq_buckets == NULL);
801 802 ASSERT(tq->tq_tcreates == 0);
802 803 ASSERT(tq->tq_tdeaths == 0);
803 804
804 805 mutex_destroy(&tq->tq_lock);
805 806 rw_destroy(&tq->tq_threadlock);
806 807 cv_destroy(&tq->tq_dispatch_cv);
807 808 cv_destroy(&tq->tq_exit_cv);
808 809 cv_destroy(&tq->tq_wait_cv);
809 810 cv_destroy(&tq->tq_maxalloc_cv);
810 811 }
811 812
812 813 /*ARGSUSED*/
813 814 static int
814 815 taskq_ent_constructor(void *buf, void *cdrarg, int kmflags)
815 816 {
816 817 taskq_ent_t *tqe = buf;
817 818
818 819 tqe->tqent_thread = NULL;
819 820 cv_init(&tqe->tqent_cv, NULL, CV_DEFAULT, NULL);
820 821
821 822 return (0);
822 823 }
823 824
824 825 /*ARGSUSED*/
825 826 static void
826 827 taskq_ent_destructor(void *buf, void *cdrarg)
827 828 {
828 829 taskq_ent_t *tqe = buf;
829 830
830 831 ASSERT(tqe->tqent_thread == NULL);
831 832 cv_destroy(&tqe->tqent_cv);
832 833 }
833 834
834 835 void
835 836 taskq_init(void)
836 837 {
837 838 taskq_ent_cache = kmem_cache_create("taskq_ent_cache",
838 839 sizeof (taskq_ent_t), 0, taskq_ent_constructor,
839 840 taskq_ent_destructor, NULL, NULL, NULL, 0);
840 841 taskq_cache = kmem_cache_create("taskq_cache", sizeof (taskq_t),
841 842 0, taskq_constructor, taskq_destructor, NULL, NULL, NULL, 0);
842 843 taskq_id_arena = vmem_create("taskq_id_arena",
843 844 (void *)1, INT32_MAX, 1, NULL, NULL, NULL, 0,
844 845 VM_SLEEP | VMC_IDENTIFIER);
845 846
846 847 list_create(&taskq_cpupct_list, sizeof (taskq_t),
847 848 offsetof(taskq_t, tq_cpupct_link));
848 849 }
849 850
850 851 static void
851 852 taskq_update_nthreads(taskq_t *tq, uint_t ncpus)
852 853 {
853 854 uint_t newtarget = TASKQ_THREADS_PCT(ncpus, tq->tq_threads_ncpus_pct);
854 855
855 856 ASSERT(MUTEX_HELD(&cpu_lock));
856 857 ASSERT(MUTEX_HELD(&tq->tq_lock));
857 858
858 859 /* We must be going from non-zero to non-zero; no exiting. */
859 860 ASSERT3U(tq->tq_nthreads_target, !=, 0);
860 861 ASSERT3U(newtarget, !=, 0);
861 862
862 863 ASSERT3U(newtarget, <=, tq->tq_nthreads_max);
863 864 if (newtarget != tq->tq_nthreads_target) {
864 865 tq->tq_flags |= TASKQ_CHANGING;
865 866 tq->tq_nthreads_target = newtarget;
866 867 cv_broadcast(&tq->tq_dispatch_cv);
867 868 cv_broadcast(&tq->tq_exit_cv);
868 869 }
869 870 }
870 871
871 872 /* called during task queue creation */
872 873 static void
873 874 taskq_cpupct_install(taskq_t *tq, cpupart_t *cpup)
874 875 {
875 876 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
876 877
877 878 mutex_enter(&cpu_lock);
878 879 mutex_enter(&tq->tq_lock);
879 880 tq->tq_cpupart = cpup->cp_id;
880 881 taskq_update_nthreads(tq, cpup->cp_ncpus);
881 882 mutex_exit(&tq->tq_lock);
882 883
883 884 list_insert_tail(&taskq_cpupct_list, tq);
884 885 mutex_exit(&cpu_lock);
885 886 }
886 887
887 888 static void
888 889 taskq_cpupct_remove(taskq_t *tq)
889 890 {
890 891 ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
891 892
892 893 mutex_enter(&cpu_lock);
893 894 list_remove(&taskq_cpupct_list, tq);
894 895 mutex_exit(&cpu_lock);
895 896 }
896 897
897 898 /*ARGSUSED*/
898 899 static int
899 900 taskq_cpu_setup(cpu_setup_t what, int id, void *arg)
900 901 {
901 902 taskq_t *tq;
902 903 cpupart_t *cp = cpu[id]->cpu_part;
903 904 uint_t ncpus = cp->cp_ncpus;
904 905
905 906 ASSERT(MUTEX_HELD(&cpu_lock));
906 907 ASSERT(ncpus > 0);
907 908
908 909 switch (what) {
909 910 case CPU_OFF:
910 911 case CPU_CPUPART_OUT:
911 912 /* offlines are called *before* the cpu is offlined. */
912 913 if (ncpus > 1)
913 914 ncpus--;
914 915 break;
915 916
916 917 case CPU_ON:
917 918 case CPU_CPUPART_IN:
918 919 break;
919 920
920 921 default:
921 922 return (0); /* doesn't affect cpu count */
922 923 }
923 924
924 925 for (tq = list_head(&taskq_cpupct_list); tq != NULL;
925 926 tq = list_next(&taskq_cpupct_list, tq)) {
926 927
927 928 mutex_enter(&tq->tq_lock);
928 929 /*
929 930 * If the taskq is part of the cpuset which is changing,
930 931 * update its nthreads_target.
931 932 */
932 933 if (tq->tq_cpupart == cp->cp_id) {
933 934 taskq_update_nthreads(tq, ncpus);
934 935 }
935 936 mutex_exit(&tq->tq_lock);
936 937 }
937 938 return (0);
938 939 }
939 940
940 941 void
941 942 taskq_mp_init(void)
942 943 {
943 944 mutex_enter(&cpu_lock);
944 945 register_cpu_setup_func(taskq_cpu_setup, NULL);
945 946 /*
946 947 * Make sure we're up to date. At this point in boot, there is only
947 948 * one processor set, so we only have to update the current CPU.
948 949 */
949 950 (void) taskq_cpu_setup(CPU_ON, CPU->cpu_id, NULL);
950 951 mutex_exit(&cpu_lock);
951 952 }
952 953
953 954 /*
954 955 * Create global system dynamic task queue.
955 956 */
956 957 void
957 958 system_taskq_init(void)
958 959 {
959 960 system_taskq = taskq_create_common("system_taskq", 0,
960 961 system_taskq_size * max_ncpus, minclsyspri, 4, 512, &p0, 0,
961 962 TASKQ_DYNAMIC | TASKQ_PREPOPULATE);
962 963 }
963 964
964 965 /*
965 966 * taskq_ent_alloc()
966 967 *
967 968 * Allocates a new taskq_ent_t structure either from the free list or from the
968 969 * cache. Returns NULL if it can't be allocated.
969 970 *
970 971 * Assumes: tq->tq_lock is held.
971 972 */
972 973 static taskq_ent_t *
973 974 taskq_ent_alloc(taskq_t *tq, int flags)
974 975 {
975 976 int kmflags = (flags & TQ_NOSLEEP) ? KM_NOSLEEP : KM_SLEEP;
976 977 taskq_ent_t *tqe;
977 978 clock_t wait_time;
978 979 clock_t wait_rv;
979 980
980 981 ASSERT(MUTEX_HELD(&tq->tq_lock));
981 982
982 983 /*
983 984 * TQ_NOALLOC allocations are allowed to use the freelist, even if
984 985 * we are below tq_minalloc.
985 986 */
986 987 again: if ((tqe = tq->tq_freelist) != NULL &&
987 988 ((flags & TQ_NOALLOC) || tq->tq_nalloc >= tq->tq_minalloc)) {
988 989 tq->tq_freelist = tqe->tqent_next;
989 990 } else {
990 991 if (flags & TQ_NOALLOC)
991 992 return (NULL);
992 993
993 994 if (tq->tq_nalloc >= tq->tq_maxalloc) {
994 995 if (kmflags & KM_NOSLEEP)
995 996 return (NULL);
996 997
997 998 /*
998 999 * We don't want to exceed tq_maxalloc, but we can't
999 1000 * wait for other tasks to complete (and thus free up
1000 1001 * task structures) without risking deadlock with
1001 1002 * the caller. So, we just delay for one second
1002 1003 * to throttle the allocation rate. If we have tasks
1003 1004 * complete before one second timeout expires then
1004 1005 * taskq_ent_free will signal us and we will
1005 1006 * immediately retry the allocation (reap free).
1006 1007 */
1007 1008 wait_time = ddi_get_lbolt() + hz;
1008 1009 while (tq->tq_freelist == NULL) {
1009 1010 tq->tq_maxalloc_wait++;
1010 1011 wait_rv = cv_timedwait(&tq->tq_maxalloc_cv,
1011 1012 &tq->tq_lock, wait_time);
1012 1013 tq->tq_maxalloc_wait--;
1013 1014 if (wait_rv == -1)
1014 1015 break;
1015 1016 }
1016 1017 if (tq->tq_freelist)
1017 1018 goto again; /* reap freelist */
1018 1019
1019 1020 }
1020 1021 mutex_exit(&tq->tq_lock);
1021 1022
1022 1023 tqe = kmem_cache_alloc(taskq_ent_cache, kmflags);
1023 1024
1024 1025 mutex_enter(&tq->tq_lock);
1025 1026 if (tqe != NULL)
1026 1027 tq->tq_nalloc++;
1027 1028 }
1028 1029 return (tqe);
1029 1030 }
1030 1031
1031 1032 /*
1032 1033 * taskq_ent_free()
1033 1034 *
1034 1035 * Free taskq_ent_t structure by either putting it on the free list or freeing
1035 1036 * it to the cache.
1036 1037 *
1037 1038 * Assumes: tq->tq_lock is held.
1038 1039 */
1039 1040 static void
1040 1041 taskq_ent_free(taskq_t *tq, taskq_ent_t *tqe)
1041 1042 {
1042 1043 ASSERT(MUTEX_HELD(&tq->tq_lock));
1043 1044
1044 1045 if (tq->tq_nalloc <= tq->tq_minalloc) {
1045 1046 tqe->tqent_next = tq->tq_freelist;
1046 1047 tq->tq_freelist = tqe;
1047 1048 } else {
1048 1049 tq->tq_nalloc--;
1049 1050 mutex_exit(&tq->tq_lock);
1050 1051 kmem_cache_free(taskq_ent_cache, tqe);
1051 1052 mutex_enter(&tq->tq_lock);
1052 1053 }
1053 1054
1054 1055 if (tq->tq_maxalloc_wait)
1055 1056 cv_signal(&tq->tq_maxalloc_cv);
1056 1057 }
1057 1058
1058 1059 /*
1059 1060 * taskq_ent_exists()
1060 1061 *
1061 1062 * Return 1 if taskq already has entry for calling 'func(arg)'.
1062 1063 *
1063 1064 * Assumes: tq->tq_lock is held.
1064 1065 */
1065 1066 static int
1066 1067 taskq_ent_exists(taskq_t *tq, task_func_t func, void *arg)
1067 1068 {
1068 1069 taskq_ent_t *tqe;
1069 1070
1070 1071 ASSERT(MUTEX_HELD(&tq->tq_lock));
1071 1072
1072 1073 for (tqe = tq->tq_task.tqent_next; tqe != &tq->tq_task;
1073 1074 tqe = tqe->tqent_next)
1074 1075 if ((tqe->tqent_func == func) && (tqe->tqent_arg == arg))
1075 1076 return (1);
1076 1077 return (0);
1077 1078 }
1078 1079
1079 1080 /*
1080 1081 * Dispatch a task "func(arg)" to a free entry of bucket b.
1081 1082 *
1082 1083 * Assumes: no bucket locks is held.
1083 1084 *
1084 1085 * Returns: a pointer to an entry if dispatch was successful.
1085 1086 * NULL if there are no free entries or if the bucket is suspended.
1086 1087 */
1087 1088 static taskq_ent_t *
1088 1089 taskq_bucket_dispatch(taskq_bucket_t *b, task_func_t func, void *arg)
1089 1090 {
1090 1091 taskq_ent_t *tqe;
1091 1092
1092 1093 ASSERT(MUTEX_NOT_HELD(&b->tqbucket_lock));
1093 1094 ASSERT(func != NULL);
1094 1095
1095 1096 mutex_enter(&b->tqbucket_lock);
1096 1097
1097 1098 ASSERT(b->tqbucket_nfree != 0 || IS_EMPTY(b->tqbucket_freelist));
1098 1099 ASSERT(b->tqbucket_nfree == 0 || !IS_EMPTY(b->tqbucket_freelist));
1099 1100
1100 1101 /*
1101 1102 * Get en entry from the freelist if there is one.
1102 1103 * Schedule task into the entry.
1103 1104 */
1104 1105 if ((b->tqbucket_nfree != 0) &&
1105 1106 !(b->tqbucket_flags & TQBUCKET_SUSPEND)) {
1106 1107 tqe = b->tqbucket_freelist.tqent_prev;
1107 1108
1108 1109 ASSERT(tqe != &b->tqbucket_freelist);
1109 1110 ASSERT(tqe->tqent_thread != NULL);
1110 1111
1111 1112 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1112 1113 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1113 1114 b->tqbucket_nalloc++;
1114 1115 b->tqbucket_nfree--;
1115 1116 tqe->tqent_func = func;
1116 1117 tqe->tqent_arg = arg;
1117 1118 TQ_STAT(b, tqs_hits);
1118 1119 cv_signal(&tqe->tqent_cv);
1119 1120 DTRACE_PROBE2(taskq__d__enqueue, taskq_bucket_t *, b,
1120 1121 taskq_ent_t *, tqe);
1121 1122 } else {
1122 1123 tqe = NULL;
1123 1124 TQ_STAT(b, tqs_misses);
1124 1125 }
1125 1126 mutex_exit(&b->tqbucket_lock);
1126 1127 return (tqe);
1127 1128 }
1128 1129
1129 1130 /*
1130 1131 * Dispatch a task.
1131 1132 *
1132 1133 * Assumes: func != NULL
1133 1134 *
1134 1135 * Returns: NULL if dispatch failed.
1135 1136 * non-NULL if task dispatched successfully.
1136 1137 * Actual return value is the pointer to taskq entry that was used to
1137 1138 * dispatch a task. This is useful for debugging.
1138 1139 */
1139 1140 taskqid_t
1140 1141 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
1141 1142 {
1142 1143 taskq_bucket_t *bucket = NULL; /* Which bucket needs extension */
1143 1144 taskq_ent_t *tqe = NULL;
1144 1145 taskq_ent_t *tqe1;
1145 1146 uint_t bsize;
1146 1147
1147 1148 ASSERT(tq != NULL);
1148 1149 ASSERT(func != NULL);
1149 1150
1150 1151 if (!(tq->tq_flags & TASKQ_DYNAMIC)) {
1151 1152 /*
1152 1153 * TQ_NOQUEUE flag can't be used with non-dynamic task queues.
1153 1154 */
1154 1155 ASSERT(!(flags & TQ_NOQUEUE));
1155 1156 /*
1156 1157 * Enqueue the task to the underlying queue.
1157 1158 */
1158 1159 mutex_enter(&tq->tq_lock);
1159 1160
1160 1161 TASKQ_S_RANDOM_DISPATCH_FAILURE(tq, flags);
1161 1162
1162 1163 if ((tqe = taskq_ent_alloc(tq, flags)) == NULL) {
1163 1164 tq->tq_nomem++;
1164 1165 mutex_exit(&tq->tq_lock);
1165 1166 return (NULL);
1166 1167 }
1167 1168 /* Make sure we start without any flags */
1168 1169 tqe->tqent_un.tqent_flags = 0;
1169 1170
1170 1171 if (flags & TQ_FRONT) {
1171 1172 TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1172 1173 } else {
1173 1174 TQ_ENQUEUE(tq, tqe, func, arg);
1174 1175 }
1175 1176 mutex_exit(&tq->tq_lock);
1176 1177 return ((taskqid_t)tqe);
1177 1178 }
1178 1179
1179 1180 /*
1180 1181 * Dynamic taskq dispatching.
1181 1182 */
1182 1183 ASSERT(!(flags & (TQ_NOALLOC | TQ_FRONT)));
1183 1184 TASKQ_D_RANDOM_DISPATCH_FAILURE(tq, flags);
1184 1185
1185 1186 bsize = tq->tq_nbuckets;
1186 1187
1187 1188 if (bsize == 1) {
1188 1189 /*
1189 1190 * In a single-CPU case there is only one bucket, so get
1190 1191 * entry directly from there.
1191 1192 */
1192 1193 if ((tqe = taskq_bucket_dispatch(tq->tq_buckets, func, arg))
1193 1194 != NULL)
1194 1195 return ((taskqid_t)tqe); /* Fastpath */
1195 1196 bucket = tq->tq_buckets;
1196 1197 } else {
1197 1198 int loopcount;
1198 1199 taskq_bucket_t *b;
1199 1200 uintptr_t h = ((uintptr_t)CPU + (uintptr_t)arg) >> 3;
1200 1201
1201 1202 h = TQ_HASH(h);
1202 1203
1203 1204 /*
1204 1205 * The 'bucket' points to the original bucket that we hit. If we
1205 1206 * can't allocate from it, we search other buckets, but only
1206 1207 * extend this one.
1207 1208 */
1208 1209 b = &tq->tq_buckets[h & (bsize - 1)];
1209 1210 ASSERT(b->tqbucket_taskq == tq); /* Sanity check */
1210 1211
1211 1212 /*
1212 1213 * Do a quick check before grabbing the lock. If the bucket does
1213 1214 * not have free entries now, chances are very small that it
1214 1215 * will after we take the lock, so we just skip it.
1215 1216 */
1216 1217 if (b->tqbucket_nfree != 0) {
1217 1218 if ((tqe = taskq_bucket_dispatch(b, func, arg)) != NULL)
1218 1219 return ((taskqid_t)tqe); /* Fastpath */
1219 1220 } else {
1220 1221 TQ_STAT(b, tqs_misses);
1221 1222 }
1222 1223
1223 1224 bucket = b;
1224 1225 loopcount = MIN(taskq_search_depth, bsize);
1225 1226 /*
1226 1227 * If bucket dispatch failed, search loopcount number of buckets
1227 1228 * before we give up and fail.
1228 1229 */
1229 1230 do {
1230 1231 b = &tq->tq_buckets[++h & (bsize - 1)];
1231 1232 ASSERT(b->tqbucket_taskq == tq); /* Sanity check */
1232 1233 loopcount--;
1233 1234
1234 1235 if (b->tqbucket_nfree != 0) {
1235 1236 tqe = taskq_bucket_dispatch(b, func, arg);
1236 1237 } else {
1237 1238 TQ_STAT(b, tqs_misses);
1238 1239 }
1239 1240 } while ((tqe == NULL) && (loopcount > 0));
1240 1241 }
1241 1242
1242 1243 /*
1243 1244 * At this point we either scheduled a task and (tqe != NULL) or failed
1244 1245 * (tqe == NULL). Try to recover from fails.
1245 1246 */
1246 1247
1247 1248 /*
1248 1249 * For KM_SLEEP dispatches, try to extend the bucket and retry dispatch.
1249 1250 */
1250 1251 if ((tqe == NULL) && !(flags & TQ_NOSLEEP)) {
1251 1252 /*
1252 1253 * taskq_bucket_extend() may fail to do anything, but this is
1253 1254 * fine - we deal with it later. If the bucket was successfully
1254 1255 * extended, there is a good chance that taskq_bucket_dispatch()
1255 1256 * will get this new entry, unless someone is racing with us and
1256 1257 * stealing the new entry from under our nose.
1257 1258 * taskq_bucket_extend() may sleep.
1258 1259 */
1259 1260 taskq_bucket_extend(bucket);
1260 1261 TQ_STAT(bucket, tqs_disptcreates);
1261 1262 if ((tqe = taskq_bucket_dispatch(bucket, func, arg)) != NULL)
1262 1263 return ((taskqid_t)tqe);
1263 1264 }
1264 1265
1265 1266 ASSERT(bucket != NULL);
1266 1267
1267 1268 /*
1268 1269 * Since there are not enough free entries in the bucket, add a
1269 1270 * taskq entry to extend it in the background using backing queue
1270 1271 * (unless we already have a taskq entry to perform that extension).
1271 1272 */
1272 1273 mutex_enter(&tq->tq_lock);
1273 1274 if (!taskq_ent_exists(tq, taskq_bucket_extend, bucket)) {
1274 1275 if ((tqe1 = taskq_ent_alloc(tq, TQ_NOSLEEP)) != NULL) {
1275 1276 TQ_ENQUEUE_FRONT(tq, tqe1, taskq_bucket_extend, bucket);
1276 1277 } else {
1277 1278 tq->tq_nomem++;
1278 1279 }
1279 1280 }
1280 1281
1281 1282 /*
1282 1283 * Dispatch failed and we can't find an entry to schedule a task.
1283 1284 * Revert to the backing queue unless TQ_NOQUEUE was asked.
1284 1285 */
1285 1286 if ((tqe == NULL) && !(flags & TQ_NOQUEUE)) {
1286 1287 if ((tqe = taskq_ent_alloc(tq, flags)) != NULL) {
1287 1288 TQ_ENQUEUE(tq, tqe, func, arg);
1288 1289 } else {
1289 1290 tq->tq_nomem++;
1290 1291 }
1291 1292 }
1292 1293 mutex_exit(&tq->tq_lock);
1293 1294
1294 1295 return ((taskqid_t)tqe);
1295 1296 }
1296 1297
1297 1298 void
1298 1299 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
1299 1300 taskq_ent_t *tqe)
1300 1301 {
1301 1302 ASSERT(func != NULL);
1302 1303 ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
1303 1304
1304 1305 /*
1305 1306 * Mark it as a prealloc'd task. This is important
1306 1307 * to ensure that we don't free it later.
1307 1308 */
1308 1309 tqe->tqent_un.tqent_flags |= TQENT_FLAG_PREALLOC;
1309 1310 /*
1310 1311 * Enqueue the task to the underlying queue.
1311 1312 */
|
↓ open down ↓ |
1273 lines elided |
↑ open up ↑ |
1312 1313 mutex_enter(&tq->tq_lock);
1313 1314
1314 1315 if (flags & TQ_FRONT) {
1315 1316 TQ_ENQUEUE_FRONT(tq, tqe, func, arg);
1316 1317 } else {
1317 1318 TQ_ENQUEUE(tq, tqe, func, arg);
1318 1319 }
1319 1320 mutex_exit(&tq->tq_lock);
1320 1321 }
1321 1322
1323 +/*
1324 + * Allow our caller to ask if there are tasks pending on the queue.
1325 + */
1326 +boolean_t
1327 +taskq_empty(taskq_t *tq)
1328 +{
1329 + boolean_t rv;
1330 +
1331 + ASSERT3P(tq, !=, curthread->t_taskq);
1332 + mutex_enter(&tq->tq_lock);
1333 + rv = (tq->tq_task.tqent_next == &tq->tq_task) && (tq->tq_active == 0);
1334 + mutex_exit(&tq->tq_lock);
1335 +
1336 + return (rv);
1337 +}
1338 +
1322 1339 /*
1323 1340 * Wait for all pending tasks to complete.
1324 1341 * Calling taskq_wait from a task will cause deadlock.
1325 1342 */
1326 1343 void
1327 1344 taskq_wait(taskq_t *tq)
1328 1345 {
1329 1346 ASSERT(tq != curthread->t_taskq);
1330 1347
1331 1348 mutex_enter(&tq->tq_lock);
1332 1349 while (tq->tq_task.tqent_next != &tq->tq_task || tq->tq_active != 0)
1333 1350 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1334 1351 mutex_exit(&tq->tq_lock);
1335 1352
1336 1353 if (tq->tq_flags & TASKQ_DYNAMIC) {
1337 1354 taskq_bucket_t *b = tq->tq_buckets;
1338 1355 int bid = 0;
1339 1356 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1340 1357 mutex_enter(&b->tqbucket_lock);
1341 1358 while (b->tqbucket_nalloc > 0)
1342 1359 cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
1343 1360 mutex_exit(&b->tqbucket_lock);
1344 1361 }
1345 1362 }
1346 1363 }
1347 1364
1348 1365 /*
1349 1366 * Suspend execution of tasks.
1350 1367 *
1351 1368 * Tasks in the queue part will be suspended immediately upon return from this
1352 1369 * function. Pending tasks in the dynamic part will continue to execute, but all
1353 1370 * new tasks will be suspended.
1354 1371 */
1355 1372 void
1356 1373 taskq_suspend(taskq_t *tq)
1357 1374 {
1358 1375 rw_enter(&tq->tq_threadlock, RW_WRITER);
1359 1376
1360 1377 if (tq->tq_flags & TASKQ_DYNAMIC) {
1361 1378 taskq_bucket_t *b = tq->tq_buckets;
1362 1379 int bid = 0;
1363 1380 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1364 1381 mutex_enter(&b->tqbucket_lock);
1365 1382 b->tqbucket_flags |= TQBUCKET_SUSPEND;
1366 1383 mutex_exit(&b->tqbucket_lock);
1367 1384 }
1368 1385 }
1369 1386 /*
1370 1387 * Mark task queue as being suspended. Needed for taskq_suspended().
1371 1388 */
1372 1389 mutex_enter(&tq->tq_lock);
1373 1390 ASSERT(!(tq->tq_flags & TASKQ_SUSPENDED));
1374 1391 tq->tq_flags |= TASKQ_SUSPENDED;
1375 1392 mutex_exit(&tq->tq_lock);
1376 1393 }
1377 1394
1378 1395 /*
1379 1396 * returns: 1 if tq is suspended, 0 otherwise.
1380 1397 */
1381 1398 int
1382 1399 taskq_suspended(taskq_t *tq)
1383 1400 {
1384 1401 return ((tq->tq_flags & TASKQ_SUSPENDED) != 0);
1385 1402 }
1386 1403
1387 1404 /*
1388 1405 * Resume taskq execution.
1389 1406 */
1390 1407 void
1391 1408 taskq_resume(taskq_t *tq)
1392 1409 {
1393 1410 ASSERT(RW_WRITE_HELD(&tq->tq_threadlock));
1394 1411
1395 1412 if (tq->tq_flags & TASKQ_DYNAMIC) {
1396 1413 taskq_bucket_t *b = tq->tq_buckets;
1397 1414 int bid = 0;
1398 1415 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
1399 1416 mutex_enter(&b->tqbucket_lock);
1400 1417 b->tqbucket_flags &= ~TQBUCKET_SUSPEND;
1401 1418 mutex_exit(&b->tqbucket_lock);
1402 1419 }
1403 1420 }
1404 1421 mutex_enter(&tq->tq_lock);
1405 1422 ASSERT(tq->tq_flags & TASKQ_SUSPENDED);
1406 1423 tq->tq_flags &= ~TASKQ_SUSPENDED;
1407 1424 mutex_exit(&tq->tq_lock);
1408 1425
1409 1426 rw_exit(&tq->tq_threadlock);
1410 1427 }
1411 1428
1412 1429 int
1413 1430 taskq_member(taskq_t *tq, kthread_t *thread)
1414 1431 {
1415 1432 return (thread->t_taskq == tq);
1416 1433 }
1417 1434
1418 1435 /*
1419 1436 * Creates a thread in the taskq. We only allow one outstanding create at
1420 1437 * a time. We drop and reacquire the tq_lock in order to avoid blocking other
1421 1438 * taskq activity while thread_create() or lwp_kernel_create() run.
1422 1439 *
1423 1440 * The first time we're called, we do some additional setup, and do not
1424 1441 * return until there are enough threads to start servicing requests.
1425 1442 */
1426 1443 static void
1427 1444 taskq_thread_create(taskq_t *tq)
1428 1445 {
1429 1446 kthread_t *t;
1430 1447 const boolean_t first = (tq->tq_nthreads == 0);
1431 1448
1432 1449 ASSERT(MUTEX_HELD(&tq->tq_lock));
1433 1450 ASSERT(tq->tq_flags & TASKQ_CHANGING);
1434 1451 ASSERT(tq->tq_nthreads < tq->tq_nthreads_target);
1435 1452 ASSERT(!(tq->tq_flags & TASKQ_THREAD_CREATED));
1436 1453
1437 1454
1438 1455 tq->tq_flags |= TASKQ_THREAD_CREATED;
1439 1456 tq->tq_active++;
1440 1457 mutex_exit(&tq->tq_lock);
1441 1458
1442 1459 /*
1443 1460 * With TASKQ_DUTY_CYCLE the new thread must have an LWP
1444 1461 * as explained in ../disp/sysdc.c (for the msacct data).
1445 1462 * Otherwise simple kthreads are preferred.
1446 1463 */
1447 1464 if ((tq->tq_flags & TASKQ_DUTY_CYCLE) != 0) {
1448 1465 /* Enforced in taskq_create_common */
1449 1466 ASSERT3P(tq->tq_proc, !=, &p0);
1450 1467 t = lwp_kernel_create(tq->tq_proc, taskq_thread, tq, TS_RUN,
1451 1468 tq->tq_pri);
1452 1469 } else {
1453 1470 t = thread_create(NULL, 0, taskq_thread, tq, 0, tq->tq_proc,
1454 1471 TS_RUN, tq->tq_pri);
1455 1472 }
1456 1473
1457 1474 if (!first) {
1458 1475 mutex_enter(&tq->tq_lock);
1459 1476 return;
1460 1477 }
1461 1478
1462 1479 /*
1463 1480 * We know the thread cannot go away, since tq cannot be
1464 1481 * destroyed until creation has completed. We can therefore
1465 1482 * safely dereference t.
1466 1483 */
1467 1484 if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
1468 1485 taskq_cpupct_install(tq, t->t_cpupart);
1469 1486 }
1470 1487 mutex_enter(&tq->tq_lock);
1471 1488
1472 1489 /* Wait until we can service requests. */
1473 1490 while (tq->tq_nthreads != tq->tq_nthreads_target &&
1474 1491 tq->tq_nthreads < TASKQ_CREATE_ACTIVE_THREADS) {
1475 1492 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
1476 1493 }
1477 1494 }
1478 1495
1479 1496 /*
1480 1497 * Common "sleep taskq thread" function, which handles CPR stuff, as well
1481 1498 * as giving a nice common point for debuggers to find inactive threads.
1482 1499 */
1483 1500 static clock_t
1484 1501 taskq_thread_wait(taskq_t *tq, kmutex_t *mx, kcondvar_t *cv,
1485 1502 callb_cpr_t *cprinfo, clock_t timeout)
1486 1503 {
1487 1504 clock_t ret = 0;
1488 1505
1489 1506 if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1490 1507 CALLB_CPR_SAFE_BEGIN(cprinfo);
1491 1508 }
1492 1509 if (timeout < 0)
1493 1510 cv_wait(cv, mx);
1494 1511 else
1495 1512 ret = cv_reltimedwait(cv, mx, timeout, TR_CLOCK_TICK);
1496 1513
1497 1514 if (!(tq->tq_flags & TASKQ_CPR_SAFE)) {
1498 1515 CALLB_CPR_SAFE_END(cprinfo, mx);
1499 1516 }
1500 1517
1501 1518 return (ret);
1502 1519 }
1503 1520
1504 1521 /*
1505 1522 * Worker thread for processing task queue.
1506 1523 */
1507 1524 static void
1508 1525 taskq_thread(void *arg)
1509 1526 {
1510 1527 int thread_id;
1511 1528
1512 1529 taskq_t *tq = arg;
1513 1530 taskq_ent_t *tqe;
1514 1531 callb_cpr_t cprinfo;
1515 1532 hrtime_t start, end;
1516 1533 boolean_t freeit;
1517 1534
1518 1535 curthread->t_taskq = tq; /* mark ourselves for taskq_member() */
1519 1536
1520 1537 if (curproc != &p0 && (tq->tq_flags & TASKQ_DUTY_CYCLE)) {
1521 1538 sysdc_thread_enter(curthread, tq->tq_DC,
1522 1539 (tq->tq_flags & TASKQ_DC_BATCH) ? SYSDC_THREAD_BATCH : 0);
1523 1540 }
1524 1541
1525 1542 if (tq->tq_flags & TASKQ_CPR_SAFE) {
1526 1543 CALLB_CPR_INIT_SAFE(curthread, tq->tq_name);
1527 1544 } else {
1528 1545 CALLB_CPR_INIT(&cprinfo, &tq->tq_lock, callb_generic_cpr,
1529 1546 tq->tq_name);
1530 1547 }
1531 1548 mutex_enter(&tq->tq_lock);
1532 1549 thread_id = ++tq->tq_nthreads;
1533 1550 ASSERT(tq->tq_flags & TASKQ_THREAD_CREATED);
1534 1551 ASSERT(tq->tq_flags & TASKQ_CHANGING);
1535 1552 tq->tq_flags &= ~TASKQ_THREAD_CREATED;
1536 1553
1537 1554 VERIFY3S(thread_id, <=, tq->tq_nthreads_max);
1538 1555
1539 1556 if (tq->tq_nthreads_max == 1)
1540 1557 tq->tq_thread = curthread;
1541 1558 else
1542 1559 tq->tq_threadlist[thread_id - 1] = curthread;
1543 1560
1544 1561 /* Allow taskq_create_common()'s taskq_thread_create() to return. */
1545 1562 if (tq->tq_nthreads == TASKQ_CREATE_ACTIVE_THREADS)
1546 1563 cv_broadcast(&tq->tq_wait_cv);
1547 1564
1548 1565 for (;;) {
1549 1566 if (tq->tq_flags & TASKQ_CHANGING) {
1550 1567 /* See if we're no longer needed */
1551 1568 if (thread_id > tq->tq_nthreads_target) {
1552 1569 /*
1553 1570 * To preserve the one-to-one mapping between
1554 1571 * thread_id and thread, we must exit from
1555 1572 * highest thread ID to least.
1556 1573 *
1557 1574 * However, if everyone is exiting, the order
1558 1575 * doesn't matter, so just exit immediately.
1559 1576 * (this is safe, since you must wait for
1560 1577 * nthreads to reach 0 after setting
1561 1578 * tq_nthreads_target to 0)
1562 1579 */
1563 1580 if (thread_id == tq->tq_nthreads ||
1564 1581 tq->tq_nthreads_target == 0)
1565 1582 break;
1566 1583
1567 1584 /* Wait for higher thread_ids to exit */
1568 1585 (void) taskq_thread_wait(tq, &tq->tq_lock,
1569 1586 &tq->tq_exit_cv, &cprinfo, -1);
1570 1587 continue;
1571 1588 }
1572 1589
1573 1590 /*
1574 1591 * If no thread is starting taskq_thread(), we can
1575 1592 * do some bookkeeping.
1576 1593 */
1577 1594 if (!(tq->tq_flags & TASKQ_THREAD_CREATED)) {
1578 1595 /* Check if we've reached our target */
1579 1596 if (tq->tq_nthreads == tq->tq_nthreads_target) {
1580 1597 tq->tq_flags &= ~TASKQ_CHANGING;
1581 1598 cv_broadcast(&tq->tq_wait_cv);
1582 1599 }
1583 1600 /* Check if we need to create a thread */
1584 1601 if (tq->tq_nthreads < tq->tq_nthreads_target) {
1585 1602 taskq_thread_create(tq);
1586 1603 continue; /* tq_lock was dropped */
1587 1604 }
1588 1605 }
1589 1606 }
1590 1607 if ((tqe = tq->tq_task.tqent_next) == &tq->tq_task) {
1591 1608 if (--tq->tq_active == 0)
1592 1609 cv_broadcast(&tq->tq_wait_cv);
1593 1610 (void) taskq_thread_wait(tq, &tq->tq_lock,
1594 1611 &tq->tq_dispatch_cv, &cprinfo, -1);
1595 1612 tq->tq_active++;
1596 1613 continue;
1597 1614 }
1598 1615
1599 1616 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1600 1617 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1601 1618 mutex_exit(&tq->tq_lock);
1602 1619
1603 1620 /*
1604 1621 * For prealloc'd tasks, we don't free anything. We
1605 1622 * have to check this now, because once we call the
1606 1623 * function for a prealloc'd taskq, we can't touch the
1607 1624 * tqent any longer (calling the function returns the
1608 1625 * ownershp of the tqent back to caller of
1609 1626 * taskq_dispatch.)
1610 1627 */
1611 1628 if ((!(tq->tq_flags & TASKQ_DYNAMIC)) &&
1612 1629 (tqe->tqent_un.tqent_flags & TQENT_FLAG_PREALLOC)) {
1613 1630 /* clear pointers to assist assertion checks */
1614 1631 tqe->tqent_next = tqe->tqent_prev = NULL;
1615 1632 freeit = B_FALSE;
1616 1633 } else {
1617 1634 freeit = B_TRUE;
1618 1635 }
1619 1636
1620 1637 rw_enter(&tq->tq_threadlock, RW_READER);
1621 1638 start = gethrtime();
1622 1639 DTRACE_PROBE2(taskq__exec__start, taskq_t *, tq,
1623 1640 taskq_ent_t *, tqe);
1624 1641 tqe->tqent_func(tqe->tqent_arg);
1625 1642 DTRACE_PROBE2(taskq__exec__end, taskq_t *, tq,
1626 1643 taskq_ent_t *, tqe);
1627 1644 end = gethrtime();
1628 1645 rw_exit(&tq->tq_threadlock);
1629 1646
1630 1647 mutex_enter(&tq->tq_lock);
1631 1648 tq->tq_totaltime += end - start;
1632 1649 tq->tq_executed++;
1633 1650
1634 1651 if (freeit)
1635 1652 taskq_ent_free(tq, tqe);
1636 1653 }
1637 1654
1638 1655 if (tq->tq_nthreads_max == 1)
1639 1656 tq->tq_thread = NULL;
1640 1657 else
1641 1658 tq->tq_threadlist[thread_id - 1] = NULL;
1642 1659
1643 1660 /* We're exiting, and therefore no longer active */
1644 1661 ASSERT(tq->tq_active > 0);
1645 1662 tq->tq_active--;
1646 1663
1647 1664 ASSERT(tq->tq_nthreads > 0);
1648 1665 tq->tq_nthreads--;
1649 1666
1650 1667 /* Wake up anyone waiting for us to exit */
1651 1668 cv_broadcast(&tq->tq_exit_cv);
1652 1669 if (tq->tq_nthreads == tq->tq_nthreads_target) {
1653 1670 if (!(tq->tq_flags & TASKQ_THREAD_CREATED))
1654 1671 tq->tq_flags &= ~TASKQ_CHANGING;
1655 1672
1656 1673 cv_broadcast(&tq->tq_wait_cv);
1657 1674 }
1658 1675
1659 1676 ASSERT(!(tq->tq_flags & TASKQ_CPR_SAFE));
1660 1677 CALLB_CPR_EXIT(&cprinfo); /* drops tq->tq_lock */
1661 1678 if (curthread->t_lwp != NULL) {
1662 1679 mutex_enter(&curproc->p_lock);
1663 1680 lwp_exit();
1664 1681 } else {
1665 1682 thread_exit();
1666 1683 }
1667 1684 }
1668 1685
1669 1686 /*
1670 1687 * Worker per-entry thread for dynamic dispatches.
1671 1688 */
1672 1689 static void
1673 1690 taskq_d_thread(taskq_ent_t *tqe)
1674 1691 {
1675 1692 taskq_bucket_t *bucket = tqe->tqent_un.tqent_bucket;
1676 1693 taskq_t *tq = bucket->tqbucket_taskq;
1677 1694 kmutex_t *lock = &bucket->tqbucket_lock;
1678 1695 kcondvar_t *cv = &tqe->tqent_cv;
1679 1696 callb_cpr_t cprinfo;
1680 1697 clock_t w;
1681 1698
1682 1699 CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, tq->tq_name);
1683 1700
1684 1701 mutex_enter(lock);
1685 1702
1686 1703 for (;;) {
1687 1704 /*
1688 1705 * If a task is scheduled (func != NULL), execute it, otherwise
1689 1706 * sleep, waiting for a job.
1690 1707 */
1691 1708 if (tqe->tqent_func != NULL) {
1692 1709 hrtime_t start;
1693 1710 hrtime_t end;
1694 1711
1695 1712 ASSERT(bucket->tqbucket_nalloc > 0);
1696 1713
1697 1714 /*
1698 1715 * It is possible to free the entry right away before
1699 1716 * actually executing the task so that subsequent
1700 1717 * dispatches may immediately reuse it. But this,
1701 1718 * effectively, creates a two-length queue in the entry
1702 1719 * and may lead to a deadlock if the execution of the
1703 1720 * current task depends on the execution of the next
1704 1721 * scheduled task. So, we keep the entry busy until the
1705 1722 * task is processed.
1706 1723 */
1707 1724
1708 1725 mutex_exit(lock);
1709 1726 start = gethrtime();
1710 1727 DTRACE_PROBE3(taskq__d__exec__start, taskq_t *, tq,
1711 1728 taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1712 1729 tqe->tqent_func(tqe->tqent_arg);
1713 1730 DTRACE_PROBE3(taskq__d__exec__end, taskq_t *, tq,
1714 1731 taskq_bucket_t *, bucket, taskq_ent_t *, tqe);
1715 1732 end = gethrtime();
1716 1733 mutex_enter(lock);
1717 1734 bucket->tqbucket_totaltime += end - start;
1718 1735
1719 1736 /*
1720 1737 * Return the entry to the bucket free list.
1721 1738 */
1722 1739 tqe->tqent_func = NULL;
1723 1740 TQ_APPEND(bucket->tqbucket_freelist, tqe);
1724 1741 bucket->tqbucket_nalloc--;
1725 1742 bucket->tqbucket_nfree++;
1726 1743 ASSERT(!IS_EMPTY(bucket->tqbucket_freelist));
1727 1744 /*
1728 1745 * taskq_wait() waits for nalloc to drop to zero on
1729 1746 * tqbucket_cv.
1730 1747 */
1731 1748 cv_signal(&bucket->tqbucket_cv);
1732 1749 }
1733 1750
1734 1751 /*
1735 1752 * At this point the entry must be in the bucket free list -
1736 1753 * either because it was there initially or because it just
1737 1754 * finished executing a task and put itself on the free list.
1738 1755 */
1739 1756 ASSERT(bucket->tqbucket_nfree > 0);
1740 1757 /*
1741 1758 * Go to sleep unless we are closing.
1742 1759 * If a thread is sleeping too long, it dies.
1743 1760 */
1744 1761 if (! (bucket->tqbucket_flags & TQBUCKET_CLOSE)) {
1745 1762 w = taskq_thread_wait(tq, lock, cv,
1746 1763 &cprinfo, taskq_thread_timeout * hz);
1747 1764 }
1748 1765
1749 1766 /*
1750 1767 * At this point we may be in two different states:
1751 1768 *
1752 1769 * (1) tqent_func is set which means that a new task is
1753 1770 * dispatched and we need to execute it.
1754 1771 *
1755 1772 * (2) Thread is sleeping for too long or we are closing. In
1756 1773 * both cases destroy the thread and the entry.
1757 1774 */
1758 1775
1759 1776 /* If func is NULL we should be on the freelist. */
1760 1777 ASSERT((tqe->tqent_func != NULL) ||
1761 1778 (bucket->tqbucket_nfree > 0));
1762 1779 /* If func is non-NULL we should be allocated */
1763 1780 ASSERT((tqe->tqent_func == NULL) ||
1764 1781 (bucket->tqbucket_nalloc > 0));
1765 1782
1766 1783 /* Check freelist consistency */
1767 1784 ASSERT((bucket->tqbucket_nfree > 0) ||
1768 1785 IS_EMPTY(bucket->tqbucket_freelist));
1769 1786 ASSERT((bucket->tqbucket_nfree == 0) ||
1770 1787 !IS_EMPTY(bucket->tqbucket_freelist));
1771 1788
1772 1789 if ((tqe->tqent_func == NULL) &&
1773 1790 ((w == -1) || (bucket->tqbucket_flags & TQBUCKET_CLOSE))) {
1774 1791 /*
1775 1792 * This thread is sleeping for too long or we are
1776 1793 * closing - time to die.
1777 1794 * Thread creation/destruction happens rarely,
1778 1795 * so grabbing the lock is not a big performance issue.
1779 1796 * The bucket lock is dropped by CALLB_CPR_EXIT().
1780 1797 */
1781 1798
1782 1799 /* Remove the entry from the free list. */
1783 1800 tqe->tqent_prev->tqent_next = tqe->tqent_next;
1784 1801 tqe->tqent_next->tqent_prev = tqe->tqent_prev;
1785 1802 ASSERT(bucket->tqbucket_nfree > 0);
1786 1803 bucket->tqbucket_nfree--;
1787 1804
1788 1805 TQ_STAT(bucket, tqs_tdeaths);
1789 1806 cv_signal(&bucket->tqbucket_cv);
1790 1807 tqe->tqent_thread = NULL;
1791 1808 mutex_enter(&tq->tq_lock);
1792 1809 tq->tq_tdeaths++;
1793 1810 mutex_exit(&tq->tq_lock);
1794 1811 CALLB_CPR_EXIT(&cprinfo);
1795 1812 kmem_cache_free(taskq_ent_cache, tqe);
1796 1813 thread_exit();
1797 1814 }
1798 1815 }
1799 1816 }
1800 1817
1801 1818
1802 1819 /*
1803 1820 * Taskq creation. May sleep for memory.
1804 1821 * Always use automatically generated instances to avoid kstat name space
1805 1822 * collisions.
1806 1823 */
1807 1824
1808 1825 taskq_t *
1809 1826 taskq_create(const char *name, int nthreads, pri_t pri, int minalloc,
1810 1827 int maxalloc, uint_t flags)
1811 1828 {
1812 1829 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1813 1830
1814 1831 return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1815 1832 maxalloc, &p0, 0, flags | TASKQ_NOINSTANCE));
1816 1833 }
1817 1834
1818 1835 /*
1819 1836 * Create an instance of task queue. It is legal to create task queues with the
1820 1837 * same name and different instances.
1821 1838 *
1822 1839 * taskq_create_instance is used by ddi_taskq_create() where it gets the
1823 1840 * instance from ddi_get_instance(). In some cases the instance is not
1824 1841 * initialized and is set to -1. This case is handled as if no instance was
1825 1842 * passed at all.
1826 1843 */
1827 1844 taskq_t *
1828 1845 taskq_create_instance(const char *name, int instance, int nthreads, pri_t pri,
1829 1846 int minalloc, int maxalloc, uint_t flags)
1830 1847 {
1831 1848 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1832 1849 ASSERT((instance >= 0) || (instance == -1));
1833 1850
1834 1851 if (instance < 0) {
1835 1852 flags |= TASKQ_NOINSTANCE;
1836 1853 }
1837 1854
1838 1855 return (taskq_create_common(name, instance, nthreads,
1839 1856 pri, minalloc, maxalloc, &p0, 0, flags));
1840 1857 }
1841 1858
1842 1859 taskq_t *
1843 1860 taskq_create_proc(const char *name, int nthreads, pri_t pri, int minalloc,
1844 1861 int maxalloc, proc_t *proc, uint_t flags)
1845 1862 {
1846 1863 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1847 1864 ASSERT(proc->p_flag & SSYS);
1848 1865
1849 1866 return (taskq_create_common(name, 0, nthreads, pri, minalloc,
1850 1867 maxalloc, proc, 0, flags | TASKQ_NOINSTANCE));
1851 1868 }
1852 1869
1853 1870 taskq_t *
1854 1871 taskq_create_sysdc(const char *name, int nthreads, int minalloc,
1855 1872 int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1856 1873 {
1857 1874 ASSERT((flags & ~TASKQ_INTERFACE_FLAGS) == 0);
1858 1875 ASSERT(proc->p_flag & SSYS);
1859 1876
1860 1877 return (taskq_create_common(name, 0, nthreads, minclsyspri, minalloc,
1861 1878 maxalloc, proc, dc, flags | TASKQ_NOINSTANCE | TASKQ_DUTY_CYCLE));
1862 1879 }
1863 1880
1864 1881 static taskq_t *
1865 1882 taskq_create_common(const char *name, int instance, int nthreads, pri_t pri,
1866 1883 int minalloc, int maxalloc, proc_t *proc, uint_t dc, uint_t flags)
1867 1884 {
1868 1885 taskq_t *tq = kmem_cache_alloc(taskq_cache, KM_SLEEP);
1869 1886 uint_t ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
1870 1887 uint_t bsize; /* # of buckets - always power of 2 */
1871 1888 int max_nthreads;
1872 1889
1873 1890 /*
1874 1891 * TASKQ_DYNAMIC, TASKQ_CPR_SAFE and TASKQ_THREADS_CPU_PCT are all
1875 1892 * mutually incompatible.
1876 1893 */
1877 1894 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_CPR_SAFE));
1878 1895 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_THREADS_CPU_PCT));
1879 1896 IMPLY((flags & TASKQ_CPR_SAFE), !(flags & TASKQ_THREADS_CPU_PCT));
1880 1897
1881 1898 /* Cannot have DYNAMIC with DUTY_CYCLE */
1882 1899 IMPLY((flags & TASKQ_DYNAMIC), !(flags & TASKQ_DUTY_CYCLE));
1883 1900
1884 1901 /* Cannot have DUTY_CYCLE with a p0 kernel process */
1885 1902 IMPLY((flags & TASKQ_DUTY_CYCLE), proc != &p0);
1886 1903
1887 1904 /* Cannot have DC_BATCH without DUTY_CYCLE */
1888 1905 ASSERT((flags & (TASKQ_DUTY_CYCLE|TASKQ_DC_BATCH)) != TASKQ_DC_BATCH);
1889 1906
1890 1907 ASSERT(proc != NULL);
1891 1908
1892 1909 bsize = 1 << (highbit(ncpus) - 1);
1893 1910 ASSERT(bsize >= 1);
1894 1911 bsize = MIN(bsize, taskq_maxbuckets);
1895 1912
1896 1913 if (flags & TASKQ_DYNAMIC) {
1897 1914 ASSERT3S(nthreads, >=, 1);
1898 1915 tq->tq_maxsize = nthreads;
1899 1916
1900 1917 /* For dynamic task queues use just one backup thread */
1901 1918 nthreads = max_nthreads = 1;
1902 1919
1903 1920 } else if (flags & TASKQ_THREADS_CPU_PCT) {
1904 1921 uint_t pct;
1905 1922 ASSERT3S(nthreads, >=, 0);
1906 1923 pct = nthreads;
1907 1924
1908 1925 if (pct > taskq_cpupct_max_percent)
1909 1926 pct = taskq_cpupct_max_percent;
1910 1927
1911 1928 /*
1912 1929 * If you're using THREADS_CPU_PCT, the process for the
1913 1930 * taskq threads must be curproc. This allows any pset
1914 1931 * binding to be inherited correctly. If proc is &p0,
1915 1932 * we won't be creating LWPs, so new threads will be assigned
1916 1933 * to the default processor set.
1917 1934 */
1918 1935 ASSERT(curproc == proc || proc == &p0);
1919 1936 tq->tq_threads_ncpus_pct = pct;
1920 1937 nthreads = 1; /* corrected in taskq_thread_create() */
1921 1938 max_nthreads = TASKQ_THREADS_PCT(max_ncpus, pct);
1922 1939
1923 1940 } else {
1924 1941 ASSERT3S(nthreads, >=, 1);
1925 1942 max_nthreads = nthreads;
1926 1943 }
1927 1944
1928 1945 if (max_nthreads < taskq_minimum_nthreads_max)
1929 1946 max_nthreads = taskq_minimum_nthreads_max;
1930 1947
1931 1948 /*
1932 1949 * Make sure the name is 0-terminated, and conforms to the rules for
1933 1950 * C indentifiers
1934 1951 */
1935 1952 (void) strncpy(tq->tq_name, name, TASKQ_NAMELEN + 1);
1936 1953 strident_canon(tq->tq_name, TASKQ_NAMELEN + 1);
1937 1954
1938 1955 tq->tq_flags = flags | TASKQ_CHANGING;
1939 1956 tq->tq_active = 0;
1940 1957 tq->tq_instance = instance;
1941 1958 tq->tq_nthreads_target = nthreads;
1942 1959 tq->tq_nthreads_max = max_nthreads;
1943 1960 tq->tq_minalloc = minalloc;
1944 1961 tq->tq_maxalloc = maxalloc;
1945 1962 tq->tq_nbuckets = bsize;
1946 1963 tq->tq_proc = proc;
1947 1964 tq->tq_pri = pri;
1948 1965 tq->tq_DC = dc;
1949 1966 list_link_init(&tq->tq_cpupct_link);
1950 1967
1951 1968 if (max_nthreads > 1)
1952 1969 tq->tq_threadlist = kmem_alloc(
1953 1970 sizeof (kthread_t *) * max_nthreads, KM_SLEEP);
1954 1971
1955 1972 mutex_enter(&tq->tq_lock);
1956 1973 if (flags & TASKQ_PREPOPULATE) {
1957 1974 while (minalloc-- > 0)
1958 1975 taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
1959 1976 }
1960 1977
1961 1978 /*
1962 1979 * Before we start creating threads for this taskq, take a
1963 1980 * zone hold so the zone can't go away before taskq_destroy
1964 1981 * makes sure all the taskq threads are gone. This hold is
1965 1982 * similar in purpose to those taken by zthread_create().
1966 1983 */
1967 1984 zone_hold(tq->tq_proc->p_zone);
1968 1985
1969 1986 /*
1970 1987 * Create the first thread, which will create any other threads
1971 1988 * necessary. taskq_thread_create will not return until we have
1972 1989 * enough threads to be able to process requests.
1973 1990 */
1974 1991 taskq_thread_create(tq);
1975 1992 mutex_exit(&tq->tq_lock);
1976 1993
1977 1994 if (flags & TASKQ_DYNAMIC) {
1978 1995 taskq_bucket_t *bucket = kmem_zalloc(sizeof (taskq_bucket_t) *
1979 1996 bsize, KM_SLEEP);
1980 1997 int b_id;
1981 1998
1982 1999 tq->tq_buckets = bucket;
1983 2000
1984 2001 /* Initialize each bucket */
1985 2002 for (b_id = 0; b_id < bsize; b_id++, bucket++) {
1986 2003 mutex_init(&bucket->tqbucket_lock, NULL, MUTEX_DEFAULT,
1987 2004 NULL);
1988 2005 cv_init(&bucket->tqbucket_cv, NULL, CV_DEFAULT, NULL);
1989 2006 bucket->tqbucket_taskq = tq;
1990 2007 bucket->tqbucket_freelist.tqent_next =
1991 2008 bucket->tqbucket_freelist.tqent_prev =
1992 2009 &bucket->tqbucket_freelist;
1993 2010 if (flags & TASKQ_PREPOPULATE)
1994 2011 taskq_bucket_extend(bucket);
1995 2012 }
1996 2013 }
1997 2014
1998 2015 /*
1999 2016 * Install kstats.
2000 2017 * We have two cases:
2001 2018 * 1) Instance is provided to taskq_create_instance(). In this case it
2002 2019 * should be >= 0 and we use it.
2003 2020 *
2004 2021 * 2) Instance is not provided and is automatically generated
2005 2022 */
2006 2023 if (flags & TASKQ_NOINSTANCE) {
2007 2024 instance = tq->tq_instance =
2008 2025 (int)(uintptr_t)vmem_alloc(taskq_id_arena, 1, VM_SLEEP);
2009 2026 }
2010 2027
2011 2028 if (flags & TASKQ_DYNAMIC) {
2012 2029 if ((tq->tq_kstat = kstat_create("unix", instance,
2013 2030 tq->tq_name, "taskq_d", KSTAT_TYPE_NAMED,
2014 2031 sizeof (taskq_d_kstat) / sizeof (kstat_named_t),
2015 2032 KSTAT_FLAG_VIRTUAL)) != NULL) {
2016 2033 tq->tq_kstat->ks_lock = &taskq_d_kstat_lock;
2017 2034 tq->tq_kstat->ks_data = &taskq_d_kstat;
2018 2035 tq->tq_kstat->ks_update = taskq_d_kstat_update;
2019 2036 tq->tq_kstat->ks_private = tq;
2020 2037 kstat_install(tq->tq_kstat);
2021 2038 }
2022 2039 } else {
2023 2040 if ((tq->tq_kstat = kstat_create("unix", instance, tq->tq_name,
2024 2041 "taskq", KSTAT_TYPE_NAMED,
2025 2042 sizeof (taskq_kstat) / sizeof (kstat_named_t),
2026 2043 KSTAT_FLAG_VIRTUAL)) != NULL) {
2027 2044 tq->tq_kstat->ks_lock = &taskq_kstat_lock;
2028 2045 tq->tq_kstat->ks_data = &taskq_kstat;
2029 2046 tq->tq_kstat->ks_update = taskq_kstat_update;
2030 2047 tq->tq_kstat->ks_private = tq;
2031 2048 kstat_install(tq->tq_kstat);
2032 2049 }
2033 2050 }
2034 2051
2035 2052 return (tq);
2036 2053 }
2037 2054
2038 2055 /*
2039 2056 * taskq_destroy().
2040 2057 *
2041 2058 * Assumes: by the time taskq_destroy is called no one will use this task queue
2042 2059 * in any way and no one will try to dispatch entries in it.
2043 2060 */
2044 2061 void
2045 2062 taskq_destroy(taskq_t *tq)
2046 2063 {
2047 2064 taskq_bucket_t *b = tq->tq_buckets;
2048 2065 int bid = 0;
2049 2066
2050 2067 ASSERT(! (tq->tq_flags & TASKQ_CPR_SAFE));
2051 2068
2052 2069 /*
2053 2070 * Destroy kstats.
2054 2071 */
2055 2072 if (tq->tq_kstat != NULL) {
2056 2073 kstat_delete(tq->tq_kstat);
2057 2074 tq->tq_kstat = NULL;
2058 2075 }
2059 2076
2060 2077 /*
2061 2078 * Destroy instance if needed.
2062 2079 */
2063 2080 if (tq->tq_flags & TASKQ_NOINSTANCE) {
2064 2081 vmem_free(taskq_id_arena, (void *)(uintptr_t)(tq->tq_instance),
2065 2082 1);
2066 2083 tq->tq_instance = 0;
2067 2084 }
2068 2085
2069 2086 /*
2070 2087 * Unregister from the cpupct list.
2071 2088 */
2072 2089 if (tq->tq_flags & TASKQ_THREADS_CPU_PCT) {
2073 2090 taskq_cpupct_remove(tq);
2074 2091 }
2075 2092
2076 2093 /*
2077 2094 * Wait for any pending entries to complete.
2078 2095 */
2079 2096 taskq_wait(tq);
2080 2097
2081 2098 mutex_enter(&tq->tq_lock);
2082 2099 ASSERT((tq->tq_task.tqent_next == &tq->tq_task) &&
2083 2100 (tq->tq_active == 0));
2084 2101
2085 2102 /* notify all the threads that they need to exit */
2086 2103 tq->tq_nthreads_target = 0;
2087 2104
2088 2105 tq->tq_flags |= TASKQ_CHANGING;
2089 2106 cv_broadcast(&tq->tq_dispatch_cv);
2090 2107 cv_broadcast(&tq->tq_exit_cv);
2091 2108
2092 2109 while (tq->tq_nthreads != 0)
2093 2110 cv_wait(&tq->tq_wait_cv, &tq->tq_lock);
2094 2111
2095 2112 if (tq->tq_nthreads_max != 1)
2096 2113 kmem_free(tq->tq_threadlist, sizeof (kthread_t *) *
2097 2114 tq->tq_nthreads_max);
2098 2115
2099 2116 tq->tq_minalloc = 0;
2100 2117 while (tq->tq_nalloc != 0)
2101 2118 taskq_ent_free(tq, taskq_ent_alloc(tq, TQ_SLEEP));
2102 2119
2103 2120 mutex_exit(&tq->tq_lock);
2104 2121
2105 2122 /*
2106 2123 * Mark each bucket as closing and wakeup all sleeping threads.
2107 2124 */
2108 2125 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2109 2126 taskq_ent_t *tqe;
2110 2127
2111 2128 mutex_enter(&b->tqbucket_lock);
2112 2129
2113 2130 b->tqbucket_flags |= TQBUCKET_CLOSE;
2114 2131 /* Wakeup all sleeping threads */
2115 2132
2116 2133 for (tqe = b->tqbucket_freelist.tqent_next;
2117 2134 tqe != &b->tqbucket_freelist; tqe = tqe->tqent_next)
2118 2135 cv_signal(&tqe->tqent_cv);
2119 2136
2120 2137 ASSERT(b->tqbucket_nalloc == 0);
2121 2138
2122 2139 /*
2123 2140 * At this point we waited for all pending jobs to complete (in
2124 2141 * both the task queue and the bucket and no new jobs should
2125 2142 * arrive. Wait for all threads to die.
2126 2143 */
2127 2144 while (b->tqbucket_nfree > 0)
2128 2145 cv_wait(&b->tqbucket_cv, &b->tqbucket_lock);
2129 2146 mutex_exit(&b->tqbucket_lock);
2130 2147 mutex_destroy(&b->tqbucket_lock);
2131 2148 cv_destroy(&b->tqbucket_cv);
2132 2149 }
2133 2150
2134 2151 if (tq->tq_buckets != NULL) {
2135 2152 ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2136 2153 kmem_free(tq->tq_buckets,
2137 2154 sizeof (taskq_bucket_t) * tq->tq_nbuckets);
2138 2155
2139 2156 /* Cleanup fields before returning tq to the cache */
2140 2157 tq->tq_buckets = NULL;
2141 2158 tq->tq_tcreates = 0;
2142 2159 tq->tq_tdeaths = 0;
2143 2160 } else {
2144 2161 ASSERT(!(tq->tq_flags & TASKQ_DYNAMIC));
2145 2162 }
2146 2163
2147 2164 /*
2148 2165 * Now that all the taskq threads are gone, we can
2149 2166 * drop the zone hold taken in taskq_create_common
2150 2167 */
2151 2168 zone_rele(tq->tq_proc->p_zone);
2152 2169
2153 2170 tq->tq_threads_ncpus_pct = 0;
2154 2171 tq->tq_totaltime = 0;
2155 2172 tq->tq_tasks = 0;
2156 2173 tq->tq_maxtasks = 0;
2157 2174 tq->tq_executed = 0;
2158 2175 kmem_cache_free(taskq_cache, tq);
2159 2176 }
2160 2177
2161 2178 /*
2162 2179 * Extend a bucket with a new entry on the free list and attach a worker thread
2163 2180 * to it.
2164 2181 *
2165 2182 * Argument: pointer to the bucket.
2166 2183 *
2167 2184 * This function may quietly fail. It is only used by taskq_dispatch() which
2168 2185 * handles such failures properly.
2169 2186 */
2170 2187 static void
2171 2188 taskq_bucket_extend(void *arg)
2172 2189 {
2173 2190 taskq_ent_t *tqe;
2174 2191 taskq_bucket_t *b = (taskq_bucket_t *)arg;
2175 2192 taskq_t *tq = b->tqbucket_taskq;
2176 2193 int nthreads;
2177 2194
2178 2195 mutex_enter(&tq->tq_lock);
2179 2196
2180 2197 if (! ENOUGH_MEMORY()) {
2181 2198 tq->tq_nomem++;
2182 2199 mutex_exit(&tq->tq_lock);
2183 2200 return;
2184 2201 }
2185 2202
2186 2203 /*
2187 2204 * Observe global taskq limits on the number of threads.
2188 2205 */
2189 2206 if (tq->tq_tcreates++ - tq->tq_tdeaths > tq->tq_maxsize) {
2190 2207 tq->tq_tcreates--;
2191 2208 mutex_exit(&tq->tq_lock);
2192 2209 return;
2193 2210 }
2194 2211 mutex_exit(&tq->tq_lock);
2195 2212
2196 2213 tqe = kmem_cache_alloc(taskq_ent_cache, KM_NOSLEEP);
2197 2214
2198 2215 if (tqe == NULL) {
2199 2216 mutex_enter(&tq->tq_lock);
2200 2217 tq->tq_nomem++;
2201 2218 tq->tq_tcreates--;
2202 2219 mutex_exit(&tq->tq_lock);
2203 2220 return;
2204 2221 }
2205 2222
2206 2223 ASSERT(tqe->tqent_thread == NULL);
2207 2224
2208 2225 tqe->tqent_un.tqent_bucket = b;
2209 2226
2210 2227 /*
2211 2228 * Create a thread in a TS_STOPPED state first. If it is successfully
2212 2229 * created, place the entry on the free list and start the thread.
2213 2230 */
2214 2231 tqe->tqent_thread = thread_create(NULL, 0, taskq_d_thread, tqe,
2215 2232 0, tq->tq_proc, TS_STOPPED, tq->tq_pri);
2216 2233
2217 2234 /*
2218 2235 * Once the entry is ready, link it to the the bucket free list.
2219 2236 */
2220 2237 mutex_enter(&b->tqbucket_lock);
2221 2238 tqe->tqent_func = NULL;
2222 2239 TQ_APPEND(b->tqbucket_freelist, tqe);
2223 2240 b->tqbucket_nfree++;
2224 2241 TQ_STAT(b, tqs_tcreates);
2225 2242
2226 2243 #if TASKQ_STATISTIC
2227 2244 nthreads = b->tqbucket_stat.tqs_tcreates -
2228 2245 b->tqbucket_stat.tqs_tdeaths;
2229 2246 b->tqbucket_stat.tqs_maxthreads = MAX(nthreads,
2230 2247 b->tqbucket_stat.tqs_maxthreads);
2231 2248 #endif
2232 2249
2233 2250 mutex_exit(&b->tqbucket_lock);
2234 2251 /*
2235 2252 * Start the stopped thread.
2236 2253 */
2237 2254 thread_lock(tqe->tqent_thread);
2238 2255 tqe->tqent_thread->t_taskq = tq;
2239 2256 tqe->tqent_thread->t_schedflag |= TS_ALLSTART;
2240 2257 setrun_locked(tqe->tqent_thread);
2241 2258 thread_unlock(tqe->tqent_thread);
2242 2259 }
2243 2260
2244 2261 static int
2245 2262 taskq_kstat_update(kstat_t *ksp, int rw)
2246 2263 {
2247 2264 struct taskq_kstat *tqsp = &taskq_kstat;
2248 2265 taskq_t *tq = ksp->ks_private;
2249 2266
2250 2267 if (rw == KSTAT_WRITE)
2251 2268 return (EACCES);
2252 2269
2253 2270 tqsp->tq_pid.value.ui64 = tq->tq_proc->p_pid;
2254 2271 tqsp->tq_tasks.value.ui64 = tq->tq_tasks;
2255 2272 tqsp->tq_executed.value.ui64 = tq->tq_executed;
2256 2273 tqsp->tq_maxtasks.value.ui64 = tq->tq_maxtasks;
2257 2274 tqsp->tq_totaltime.value.ui64 = tq->tq_totaltime;
2258 2275 tqsp->tq_nactive.value.ui64 = tq->tq_active;
2259 2276 tqsp->tq_nalloc.value.ui64 = tq->tq_nalloc;
2260 2277 tqsp->tq_pri.value.ui64 = tq->tq_pri;
2261 2278 tqsp->tq_nthreads.value.ui64 = tq->tq_nthreads;
2262 2279 tqsp->tq_nomem.value.ui64 = tq->tq_nomem;
2263 2280 return (0);
2264 2281 }
2265 2282
2266 2283 static int
2267 2284 taskq_d_kstat_update(kstat_t *ksp, int rw)
2268 2285 {
2269 2286 struct taskq_d_kstat *tqsp = &taskq_d_kstat;
2270 2287 taskq_t *tq = ksp->ks_private;
2271 2288 taskq_bucket_t *b = tq->tq_buckets;
2272 2289 int bid = 0;
2273 2290
2274 2291 if (rw == KSTAT_WRITE)
2275 2292 return (EACCES);
2276 2293
2277 2294 ASSERT(tq->tq_flags & TASKQ_DYNAMIC);
2278 2295
2279 2296 tqsp->tqd_btasks.value.ui64 = tq->tq_tasks;
2280 2297 tqsp->tqd_bexecuted.value.ui64 = tq->tq_executed;
2281 2298 tqsp->tqd_bmaxtasks.value.ui64 = tq->tq_maxtasks;
2282 2299 tqsp->tqd_bnalloc.value.ui64 = tq->tq_nalloc;
2283 2300 tqsp->tqd_bnactive.value.ui64 = tq->tq_active;
2284 2301 tqsp->tqd_btotaltime.value.ui64 = tq->tq_totaltime;
2285 2302 tqsp->tqd_pri.value.ui64 = tq->tq_pri;
2286 2303 tqsp->tqd_nomem.value.ui64 = tq->tq_nomem;
2287 2304
2288 2305 tqsp->tqd_hits.value.ui64 = 0;
2289 2306 tqsp->tqd_misses.value.ui64 = 0;
2290 2307 tqsp->tqd_overflows.value.ui64 = 0;
2291 2308 tqsp->tqd_tcreates.value.ui64 = 0;
2292 2309 tqsp->tqd_tdeaths.value.ui64 = 0;
2293 2310 tqsp->tqd_maxthreads.value.ui64 = 0;
2294 2311 tqsp->tqd_nomem.value.ui64 = 0;
2295 2312 tqsp->tqd_disptcreates.value.ui64 = 0;
2296 2313 tqsp->tqd_totaltime.value.ui64 = 0;
2297 2314 tqsp->tqd_nalloc.value.ui64 = 0;
2298 2315 tqsp->tqd_nfree.value.ui64 = 0;
2299 2316
2300 2317 for (; (b != NULL) && (bid < tq->tq_nbuckets); b++, bid++) {
2301 2318 tqsp->tqd_hits.value.ui64 += b->tqbucket_stat.tqs_hits;
2302 2319 tqsp->tqd_misses.value.ui64 += b->tqbucket_stat.tqs_misses;
2303 2320 tqsp->tqd_overflows.value.ui64 += b->tqbucket_stat.tqs_overflow;
2304 2321 tqsp->tqd_tcreates.value.ui64 += b->tqbucket_stat.tqs_tcreates;
2305 2322 tqsp->tqd_tdeaths.value.ui64 += b->tqbucket_stat.tqs_tdeaths;
2306 2323 tqsp->tqd_maxthreads.value.ui64 +=
2307 2324 b->tqbucket_stat.tqs_maxthreads;
2308 2325 tqsp->tqd_disptcreates.value.ui64 +=
2309 2326 b->tqbucket_stat.tqs_disptcreates;
2310 2327 tqsp->tqd_totaltime.value.ui64 += b->tqbucket_totaltime;
2311 2328 tqsp->tqd_nalloc.value.ui64 += b->tqbucket_nalloc;
2312 2329 tqsp->tqd_nfree.value.ui64 += b->tqbucket_nfree;
2313 2330 }
2314 2331 return (0);
2315 2332 }
|
↓ open down ↓ |
984 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX