1 /*
2 * This file and its contents are supplied under the terms of the
3 * Common Development and Distribution License ("CDDL"), version 1.0.
4 * You may only use this file in accordance with the terms of version
5 * 1.0 of the CDDL.
6 *
7 * A full copy of the text of the CDDL should have accompanied this
8 * source. A copy of the CDDL is also available via the Internet at
9 * http://www.illumos.org/license/CDDL.
10 */
11 /*
12 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
13 */
14
15 #include <sys/cmn_err.h>
16 #include <sys/ddi_periodic.h>
17 #include <sys/id_space.h>
18 #include <sys/kobj.h>
19 #include <sys/sysmacros.h>
20 #include <sys/systm.h>
21 #include <sys/taskq.h>
22 #include <sys/taskq_impl.h>
23 #include <sys/time.h>
24 #include <sys/types.h>
25 #include <sys/sdt.h>
26
27 /*
28 * The ddi_periodic_add(9F) Implementation
29 *
30 * This file contains the implementation of the ddi_periodic_add(9F) interface.
31 * It is a thin wrapper around the cyclic subsystem (see documentation in
32 * uts/common/os/cyclic.c), providing a DDI interface for registering
33 * (and unregistering) callbacks for periodic invocation at arbitrary
34 * interrupt levels, or in kernel context.
35 *
36 * Each call to ddi_periodic_add will result in a new opaque handle, as
37 * allocated from an id_space, a new "periodic" object (ddi_periodic_impl_t)
38 * and a registered cyclic.
39 *
40 * Operation
41 *
42 * Whenever the cyclic fires, our cyclic handler checks that the particular
43 * periodic is not dispatched already (we do not support overlapping execution
44 * of the consumer's handler function), and not yet cancelled. If both of
45 * these conditions hold, we mark the periodic as DPF_DISPATCHED and enqueue it
46 * to either the taskq (for DDI_IPL_0) or to one of the soft interrupt queues
47 * (DDI_IPL_1 to DDI_IPL_10).
48 *
49 * While the taskq (or soft interrupt handler) is handling a particular
50 * periodic, we mark it as DPF_EXECUTING. When complete, we reset both
51 * DPF_DISPATCHED and DPF_EXECUTING.
52 *
53 * Cancellation
54 *
55 * ddi_periodic_delete(9F) historically had spectacularly loose semantics with
56 * respect to cancellation concurrent with handler execution. These semantics
57 * are now tighter:
58 *
59 * 1. At most one invocation of ddi_periodic_delete(9F) will actually
60 * perform the deletion, all others will return immediately.
61 * 2. The invocation that performs the deletion will _block_ until
62 * the handler is no longer running, and all resources have been
63 * released.
64 *
65 * We affect this model by removing the cancelling periodic from the
66 * global list and marking it DPF_CANCELLED. This will prevent further
67 * execution of the handler. We then wait on a CV until the DPF_EXECUTING
68 * and DPF_DISPATCHED flags are clear, which means the periodic is removed
69 * from all request queues, is no longer executing, and may be freed. At this
70 * point we return the opaque ID to the id_space and free the memory.
71 *
72 * NOTE:
73 * The ddi_periodic_add(9F) interface is presently limited to a minimum period
74 * of 10ms between firings.
75 */
76
77 /*
78 * Tuneables:
79 */
80 int ddi_periodic_max_id = 1024;
81 int ddi_periodic_taskq_threadcount = 4;
82 hrtime_t ddi_periodic_resolution = 10000000;
83
84 /*
85 * Globals:
86 */
87 static kmem_cache_t *periodic_cache;
88 static id_space_t *periodic_id_space;
89 static taskq_t *periodic_taskq;
90
91 /*
92 * periodics_lock protects the list of all periodics (periodics), and
93 * each of the soft interrupt request queues (periodic_softint_queue).
94 *
95 * Do not hold an individual periodic's lock while obtaining periodics_lock.
96 * While in the periodic_softint_queue list, the periodic will be marked
97 * DPF_DISPATCHED, and thus safe from frees. Only the invocation of
98 * i_untimeout() that removes the periodic from the global list is allowed
99 * to free it.
100 */
101 static kmutex_t periodics_lock;
102 static list_t periodics;
103 static list_t periodic_softint_queue[10]; /* for IPL1 up to IPL10 */
104
105 typedef enum periodic_ipl {
106 PERI_IPL_0 = 0,
107 PERI_IPL_1,
108 PERI_IPL_2,
109 PERI_IPL_3,
110 PERI_IPL_4,
111 PERI_IPL_5,
112 PERI_IPL_6,
113 PERI_IPL_7,
114 PERI_IPL_8,
115 PERI_IPL_9,
116 PERI_IPL_10
117 } periodic_ipl_t;
118
119 static char *
120 periodic_handler_symbol(ddi_periodic_impl_t *dpr)
121 {
122 ulong_t off;
123
124 return (kobj_getsymname((uintptr_t)dpr->dpr_handler, &off));
125 }
126
127 /*
128 * This function may be called either from a soft interrupt handler
129 * (ddi_periodic_softintr), or as a taskq worker function.
130 */
131 static void
132 periodic_execute(void *arg)
133 {
134 ddi_periodic_impl_t *dpr = arg;
135 mutex_enter(&dpr->dpr_lock);
136
137 /*
138 * We must be DISPATCHED, but not yet EXECUTING:
139 */
140 VERIFY((dpr->dpr_flags & (DPF_DISPATCHED | DPF_EXECUTING)) ==
141 DPF_DISPATCHED);
142 VERIFY(dpr->dpr_thread == NULL);
143
144 if (!(dpr->dpr_flags & DPF_CANCELLED)) {
145 int level = dpr->dpr_level;
146 uint64_t count = dpr->dpr_fire_count;
147 /*
148 * If we have not yet been cancelled, then
149 * mark us executing:
150 */
151 dpr->dpr_flags |= DPF_EXECUTING;
152 dpr->dpr_thread = curthread;
153 mutex_exit(&dpr->dpr_lock);
154
155 /*
156 * Execute the handler, without holding locks:
157 */
158 DTRACE_PROBE4(ddi__periodic__execute, void *, dpr->dpr_handler,
159 void *, dpr->dpr_arg, int, level, uint64_t, count);
160 (*dpr->dpr_handler)(dpr->dpr_arg);
161 DTRACE_PROBE4(ddi__periodic__done, void *, dpr->dpr_handler,
162 void *, dpr->dpr_arg, int, level, uint64_t, count);
163
164 mutex_enter(&dpr->dpr_lock);
165 dpr->dpr_thread = NULL;
166 dpr->dpr_fire_count++;
167 }
168
169 /*
170 * We're done with this periodic for now, so release it and
171 * wake anybody that was waiting for us to be finished:
172 */
173 dpr->dpr_flags &= ~(DPF_DISPATCHED | DPF_EXECUTING);
174 cv_broadcast(&dpr->dpr_cv);
175 mutex_exit(&dpr->dpr_lock);
176 }
177
178 void
179 ddi_periodic_softintr(int level)
180 {
181 ddi_periodic_impl_t *dpr;
182 VERIFY(level >= PERI_IPL_1 && level <= PERI_IPL_10);
183
184 mutex_enter(&periodics_lock);
185 /*
186 * Pull the first scheduled periodic off the queue for this priority
187 * level:
188 */
189 while ((dpr = list_remove_head(&periodic_softint_queue[level - 1])) !=
190 NULL) {
191 mutex_exit(&periodics_lock);
192 /*
193 * And execute it:
194 */
195 periodic_execute(dpr);
196 mutex_enter(&periodics_lock);
197 }
198 mutex_exit(&periodics_lock);
199 }
200
201 void
202 ddi_periodic_init(void)
203 {
204 int i;
205
206 /*
207 * Create a kmem_cache for request tracking objects, and a list
208 * to store them in so we can later delete based on opaque handles:
209 */
210 periodic_cache = kmem_cache_create("ddi_periodic",
211 sizeof (ddi_periodic_impl_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
212 list_create(&periodics, sizeof (ddi_periodic_impl_t),
213 offsetof(ddi_periodic_impl_t, dpr_link));
214
215 /*
216 * Initialise the identifier space for ddi_periodic_add(9F):
217 */
218 periodic_id_space = id_space_create("ddi_periodic", 1,
219 ddi_periodic_max_id);
220
221 /*
222 * Initialise the request queue for each soft interrupt level:
223 */
224 for (i = PERI_IPL_1; i <= PERI_IPL_10; i++) {
225 list_create(&periodic_softint_queue[i - 1],
226 sizeof (ddi_periodic_impl_t), offsetof(ddi_periodic_impl_t,
227 dpr_softint_link));
228 }
229
230 /*
231 * Create the taskq for running PERI_IPL_0 handlers. This taskq will
232 * _only_ be used with taskq_dispatch_ent(), and a taskq_ent_t
233 * pre-allocated with the ddi_periodic_impl_t.
234 */
235 periodic_taskq = taskq_create_instance("ddi_periodic_taskq", -1,
236 ddi_periodic_taskq_threadcount, maxclsyspri, 0, 0, 0);
237
238 /*
239 * Initialize the mutex lock used for the soft interrupt request
240 * queues.
241 */
242 mutex_init(&periodics_lock, NULL, MUTEX_ADAPTIVE, NULL);
243 }
244
245 void
246 ddi_periodic_fini(void)
247 {
248 ddi_periodic_impl_t *dpr;
249 /*
250 * Find all periodics that have not yet been unregistered and,
251 * on DEBUG bits, print a warning about this resource leak.
252 */
253 mutex_enter(&periodics_lock);
254 while ((dpr = list_head(&periodics)) != NULL) {
255 #ifdef DEBUG
256 printf("DDI periodic handler not deleted (id=%lx, hdlr=%s)\n",
257 (unsigned long)dpr->dpr_id, periodic_handler_symbol(dpr));
258 #endif
259
260 mutex_exit(&periodics_lock);
261 /*
262 * Delete the periodic ourselves:
263 */
264 i_untimeout((timeout_t)(uintptr_t)dpr->dpr_id);
265 mutex_enter(&periodics_lock);
266 }
267 mutex_exit(&periodics_lock);
268 }
269
270 static void
271 periodic_cyclic_handler(void *arg)
272 {
273 extern void sir_on(int);
274 ddi_periodic_impl_t *dpr = arg;
275
276 mutex_enter(&dpr->dpr_lock);
277 /*
278 * If we've been cancelled, or we're already dispatched, then exit
279 * immediately:
280 */
281 if (dpr->dpr_flags & (DPF_CANCELLED | DPF_DISPATCHED)) {
282 mutex_exit(&dpr->dpr_lock);
283 return;
284 }
285 VERIFY(!(dpr->dpr_flags & DPF_EXECUTING));
286
287 /*
288 * This periodic is not presently dispatched, so dispatch it now:
289 */
290 dpr->dpr_flags |= DPF_DISPATCHED;
291 mutex_exit(&dpr->dpr_lock);
292
293 if (dpr->dpr_level == PERI_IPL_0) {
294 /*
295 * DDI_IPL_0 periodics are dispatched onto the taskq:
296 */
297 taskq_dispatch_ent(periodic_taskq, periodic_execute,
298 dpr, 0, &dpr->dpr_taskq_ent);
299 } else {
300 /*
301 * Higher priority periodics are handled by a soft
302 * interrupt handler. Enqueue us for processing and
303 * fire the soft interrupt:
304 */
305 mutex_enter(&periodics_lock);
306 list_insert_tail(&periodic_softint_queue[dpr->dpr_level - 1],
307 dpr);
308 mutex_exit(&periodics_lock);
309
310 /*
311 * Raise the soft interrupt level for this periodic:
312 */
313 sir_on(dpr->dpr_level);
314 }
315 }
316
317 static void
318 periodic_destroy(ddi_periodic_impl_t *dpr)
319 {
320 if (dpr == NULL)
321 return;
322
323 /*
324 * By now, we should have a periodic that is not busy, and has been
325 * cancelled:
326 */
327 VERIFY(dpr->dpr_flags == DPF_CANCELLED);
328 VERIFY(dpr->dpr_thread == NULL);
329
330 id_free(periodic_id_space, dpr->dpr_id);
331 cv_destroy(&dpr->dpr_cv);
332 mutex_destroy(&dpr->dpr_lock);
333 kmem_cache_free(periodic_cache, dpr);
334 }
335
336 static ddi_periodic_impl_t *
337 periodic_create(void)
338 {
339 ddi_periodic_impl_t *dpr;
340
341 dpr = kmem_cache_alloc(periodic_cache, KM_SLEEP);
342 bzero(dpr, sizeof (*dpr));
343 dpr->dpr_id = id_alloc(periodic_id_space);
344 mutex_init(&dpr->dpr_lock, NULL, MUTEX_ADAPTIVE, NULL);
345 cv_init(&dpr->dpr_cv, NULL, CV_DEFAULT, NULL);
346
347 return (dpr);
348 }
349
350 /*
351 * This function provides the implementation for the ddi_periodic_add(9F)
352 * interface. It registers a periodic handler and returns an opaque identifier
353 * that can be unregistered via ddi_periodic_delete(9F)/i_untimeout().
354 *
355 * It may be called in user or kernel context, provided cpu_lock is not held.
356 */
357 timeout_t
358 i_timeout(void (*func)(void *), void *arg, hrtime_t interval, int level)
359 {
360 cyc_handler_t cyh;
361 cyc_time_t cyt;
362 ddi_periodic_impl_t *dpr;
363
364 VERIFY(func != NULL);
365 VERIFY(level >= 0 && level <= 10);
366
367 /*
368 * Allocate object to track this periodic:
369 */
370 dpr = periodic_create();
371 dpr->dpr_level = level;
372 dpr->dpr_handler = func;
373 dpr->dpr_arg = arg;
374
375 /*
376 * The minimum supported interval between firings of the periodic
377 * handler is 10ms; see ddi_periodic_add(9F) for more details. If a
378 * shorter interval is requested, round up.
379 */
380 if (ddi_periodic_resolution > interval) {
381 cmn_err(CE_WARN,
382 "The periodic timeout (handler=%s, interval=%lld) "
383 "requests a finer interval than the supported resolution. "
384 "It rounds up to %lld\n", periodic_handler_symbol(dpr),
385 interval, ddi_periodic_resolution);
386 interval = ddi_periodic_resolution;
387 }
388
389 /*
390 * Ensure that the interval is an even multiple of the base resolution
391 * that is at least as long as the requested interval.
392 */
393 dpr->dpr_interval = roundup(interval, ddi_periodic_resolution);
394
395 /*
396 * Create the underlying cyclic:
397 */
398 cyh.cyh_func = periodic_cyclic_handler;
399 cyh.cyh_arg = dpr;
400 cyh.cyh_level = CY_LOCK_LEVEL;
401
402 cyt.cyt_when = 0;
403 cyt.cyt_interval = dpr->dpr_interval;
404
405 mutex_enter(&cpu_lock);
406 dpr->dpr_cyclic_id = cyclic_add(&cyh, &cyt);
407 mutex_exit(&cpu_lock);
408
409 /*
410 * Make the id visible to ddi_periodic_delete(9F) before we
411 * return it:
412 */
413 mutex_enter(&periodics_lock);
414 list_insert_tail(&periodics, dpr);
415 mutex_exit(&periodics_lock);
416
417 return ((timeout_t)(uintptr_t)dpr->dpr_id);
418 }
419
420 /*
421 * This function provides the implementation for the ddi_periodic_delete(9F)
422 * interface. It cancels a periodic handler previously registered through
423 * ddi_periodic_add(9F)/i_timeout().
424 *
425 * It may be called in user or kernel context, provided cpu_lock is not held.
426 * It may NOT be called from within a periodic handler.
427 */
428 void
429 i_untimeout(timeout_t id)
430 {
431 ddi_periodic_impl_t *dpr;
432
433 /*
434 * Find the periodic in the list of all periodics and remove it.
435 * If we find in (and remove it from) the global list, we have
436 * license to free it once it is no longer busy.
437 */
438 mutex_enter(&periodics_lock);
439 for (dpr = list_head(&periodics); dpr != NULL; dpr =
440 list_next(&periodics, dpr)) {
441 if (dpr->dpr_id == (id_t)(uintptr_t)id) {
442 list_remove(&periodics, dpr);
443 break;
444 }
445 }
446 mutex_exit(&periodics_lock);
447
448 /*
449 * We could not find a periodic for this id, so bail out:
450 */
451 if (dpr == NULL)
452 return;
453
454 mutex_enter(&dpr->dpr_lock);
455 /*
456 * We should be the only one trying to cancel this periodic:
457 */
458 VERIFY(!(dpr->dpr_flags & DPF_CANCELLED));
459 /*
460 * Removing a periodic from within its own handler function will
461 * cause a deadlock, so panic explicitly.
462 */
463 if (dpr->dpr_thread == curthread) {
464 panic("ddi_periodic_delete(%lx) called from its own handler\n",
465 (unsigned long)dpr->dpr_id);
466 }
467 /*
468 * Mark the periodic as cancelled:
469 */
470 dpr->dpr_flags |= DPF_CANCELLED;
471 mutex_exit(&dpr->dpr_lock);
472
473 /*
474 * Cancel our cyclic. cyclic_remove() guarantees that the cyclic
475 * handler will not run again after it returns. Note that the cyclic
476 * handler merely _dispatches_ the periodic, so this does _not_ mean
477 * the periodic handler is also finished running.
478 */
479 mutex_enter(&cpu_lock);
480 cyclic_remove(dpr->dpr_cyclic_id);
481 mutex_exit(&cpu_lock);
482
483 /*
484 * Wait until the periodic handler is no longer running:
485 */
486 mutex_enter(&dpr->dpr_lock);
487 while (dpr->dpr_flags & (DPF_DISPATCHED | DPF_EXECUTING)) {
488 cv_wait(&dpr->dpr_cv, &dpr->dpr_lock);
489 }
490 mutex_exit(&dpr->dpr_lock);
491
492 periodic_destroy(dpr);
493 }