1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 /*
27 * Copyright 2019 Joyent, Inc.
28 */
29
30 /*
31 * The Cyclic Subsystem
32 * --------------------
33 *
34 * Prehistory
35 *
36 * Historically, most computer architectures have specified interval-based
37 * timer parts (e.g. SPARCstation's counter/timer; Intel's i8254). While
38 * these parts deal in relative (i.e. not absolute) time values, they are
39 * typically used by the operating system to implement the abstraction of
40 * absolute time. As a result, these parts cannot typically be reprogrammed
41 * without introducing error in the system's notion of time.
42 *
43 * Starting in about 1994, chip architectures began specifying high resolution
44 * timestamp registers. As of this writing (1999), all major chip families
45 * (UltraSPARC, PentiumPro, MIPS, PowerPC, Alpha) have high resolution
46 * timestamp registers, and two (UltraSPARC and MIPS) have added the capacity
47 * to interrupt based on timestamp values. These timestamp-compare registers
48 * present a time-based interrupt source which can be reprogrammed arbitrarily
49 * often without introducing error. Given the low cost of implementing such a
50 * timestamp-compare register (and the tangible benefit of eliminating
51 * discrete timer parts), it is reasonable to expect that future chip
52 * architectures will adopt this feature.
53 *
54 * The cyclic subsystem has been designed to take advantage of chip
55 * architectures with the capacity to interrupt based on absolute, high
56 * resolution values of time.
57 *
58 * Subsystem Overview
59 *
60 * The cyclic subsystem is a low-level kernel subsystem designed to provide
61 * arbitrarily high resolution, per-CPU interval timers (to avoid colliding
62 * with existing terms, we dub such an interval timer a "cyclic"). Cyclics
63 * can be specified to fire at high, lock or low interrupt level, and may be
64 * optionally bound to a CPU or a CPU partition. A cyclic's CPU or CPU
65 * partition binding may be changed dynamically; the cyclic will be "juggled"
66 * to a CPU which satisfies the new binding. Alternatively, a cyclic may
67 * be specified to be "omnipresent", denoting firing on all online CPUs.
68 *
69 * Cyclic Subsystem Interface Overview
70 * -----------------------------------
71 *
72 * The cyclic subsystem has interfaces with the kernel at-large, with other
73 * kernel subsystems (e.g. the processor management subsystem, the checkpoint
74 * resume subsystem) and with the platform (the cyclic backend). Each
75 * of these interfaces is given a brief synopsis here, and is described
76 * in full above the interface's implementation.
77 *
78 * The following diagram displays the cyclic subsystem's interfaces to
79 * other kernel components. The arrows denote a "calls" relationship, with
80 * the large arrow indicating the cyclic subsystem's consumer interface.
81 * Each arrow is labeled with the section in which the corresponding
82 * interface is described.
83 *
84 * Kernel at-large consumers
85 * -----------++------------
86 * ||
87 * ||
88 * _||_
89 * \ /
90 * \/
91 * +---------------------+
92 * | |
93 * | Cyclic subsystem |<----------- Other kernel subsystems
94 * | |
95 * +---------------------+
96 * ^ |
97 * | |
98 * | |
99 * | v
100 * +---------------------+
101 * | |
102 * | Cyclic backend |
103 * | (platform specific) |
104 * | |
105 * +---------------------+
106 *
107 *
108 * Kernel At-Large Interfaces
109 *
110 * cyclic_add() <-- Creates a cyclic
111 * cyclic_add_omni() <-- Creates an omnipresent cyclic
112 * cyclic_remove() <-- Removes a cyclic
113 * cyclic_bind() <-- Change a cyclic's CPU or partition binding
114 * cyclic_reprogram() <-- Reprogram a cyclic's expiration
115 * cyclic_move_here() <-- Shuffle cyclic to current CPU
116 *
117 * Inter-subsystem Interfaces
118 *
119 * cyclic_juggle() <-- Juggles cyclics away from a CPU
120 * cyclic_offline() <-- Offlines cyclic operation on a CPU
121 * cyclic_online() <-- Reenables operation on an offlined CPU
122 * cyclic_move_in() <-- Notifies subsystem of change in CPU partition
123 * cyclic_move_out() <-- Notifies subsystem of change in CPU partition
124 * cyclic_suspend() <-- Suspends the cyclic subsystem on all CPUs
125 * cyclic_resume() <-- Resumes the cyclic subsystem on all CPUs
126 *
127 * Backend Interfaces
128 *
129 * cyclic_init() <-- Initializes the cyclic subsystem
130 * cyclic_fire() <-- CY_HIGH_LEVEL interrupt entry point
131 * cyclic_softint() <-- CY_LOCK/LOW_LEVEL soft interrupt entry point
132 *
133 * The backend-supplied interfaces (through the cyc_backend structure) are
134 * documented in detail in <sys/cyclic_impl.h>
135 *
136 *
137 * Cyclic Subsystem Implementation Overview
138 * ----------------------------------------
139 *
140 * The cyclic subsystem is designed to minimize interference between cyclics
141 * on different CPUs. Thus, all of the cyclic subsystem's data structures
142 * hang off of a per-CPU structure, cyc_cpu.
143 *
144 * Each cyc_cpu has a power-of-two sized array of cyclic structures (the
145 * cyp_cyclics member of the cyc_cpu structure). If cyclic_add() is called
146 * and there does not exist a free slot in the cyp_cyclics array, the size of
147 * the array will be doubled. The array will never shrink. Cyclics are
148 * referred to by their index in the cyp_cyclics array, which is of type
149 * cyc_index_t.
150 *
151 * The cyclics are kept sorted by expiration time in the cyc_cpu's heap. The
152 * heap is keyed by cyclic expiration time, with parents expiring earlier
153 * than their children.
154 *
155 * Heap Management
156 *
157 * The heap is managed primarily by cyclic_fire(). Upon entry, cyclic_fire()
158 * compares the root cyclic's expiration time to the current time. If the
159 * expiration time is in the past, cyclic_expire() is called on the root
160 * cyclic. Upon return from cyclic_expire(), the cyclic's new expiration time
161 * is derived by adding its interval to its old expiration time, and a
162 * downheap operation is performed. After the downheap, cyclic_fire()
163 * examines the (potentially changed) root cyclic, repeating the
164 * cyclic_expire()/add interval/cyclic_downheap() sequence until the root
165 * cyclic has an expiration time in the future. This expiration time
166 * (guaranteed to be the earliest in the heap) is then communicated to the
167 * backend via cyb_reprogram. Optimal backends will next call cyclic_fire()
168 * shortly after the root cyclic's expiration time.
169 *
170 * To allow efficient, deterministic downheap operations, we implement the
171 * heap as an array (the cyp_heap member of the cyc_cpu structure), with each
172 * element containing an index into the CPU's cyp_cyclics array.
173 *
174 * The heap is laid out in the array according to the following:
175 *
176 * 1. The root of the heap is always in the 0th element of the heap array
177 * 2. The left and right children of the nth element are element
178 * (((n + 1) << 1) - 1) and element ((n + 1) << 1), respectively.
179 *
180 * This layout is standard (see, e.g., Cormen's "Algorithms"); the proof
181 * that these constraints correctly lay out a heap (or indeed, any binary
182 * tree) is trivial and left to the reader.
183 *
184 * To see the heap by example, assume our cyclics array has the following
185 * members (at time t):
186 *
187 * cy_handler cy_level cy_expire
188 * ---------------------------------------------
189 * [ 0] clock() LOCK t+10000000
190 * [ 1] deadman() HIGH t+1000000000
191 * [ 2] clock_highres_fire() LOW t+100
192 * [ 3] clock_highres_fire() LOW t+1000
193 * [ 4] clock_highres_fire() LOW t+500
194 * [ 5] (free) -- --
195 * [ 6] (free) -- --
196 * [ 7] (free) -- --
197 *
198 * The heap array could be:
199 *
200 * [0] [1] [2] [3] [4] [5] [6] [7]
201 * +-----+-----+-----+-----+-----+-----+-----+-----+
202 * | | | | | | | | |
203 * | 2 | 3 | 4 | 0 | 1 | x | x | x |
204 * | | | | | | | | |
205 * +-----+-----+-----+-----+-----+-----+-----+-----+
206 *
207 * Graphically, this array corresponds to the following (excuse the ASCII art):
208 *
209 * 2
210 * |
211 * +------------------+------------------+
212 * 3 4
213 * |
214 * +---------+--------+
215 * 0 1
216 *
217 * Note that the heap is laid out by layer: all nodes at a given depth are
218 * stored in consecutive elements of the array. Moreover, layers of
219 * consecutive depths are in adjacent element ranges. This property
220 * guarantees high locality of reference during downheap operations.
221 * Specifically, we are guaranteed that we can downheap to a depth of
222 *
223 * lg (cache_line_size / sizeof (cyc_index_t))
224 *
225 * nodes with at most one cache miss. On UltraSPARC (64 byte e-cache line
226 * size), this corresponds to a depth of four nodes. Thus, if there are
227 * fewer than sixteen cyclics in the heap, downheaps on UltraSPARC miss at
228 * most once in the e-cache.
229 *
230 * Downheaps are required to compare siblings as they proceed down the
231 * heap. For downheaps proceeding beyond the one-cache-miss depth, every
232 * access to a left child could potentially miss in the cache. However,
233 * if we assume
234 *
235 * (cache_line_size / sizeof (cyc_index_t)) > 2,
236 *
237 * then all siblings are guaranteed to be on the same cache line. Thus, the
238 * miss on the left child will guarantee a hit on the right child; downheaps
239 * will incur at most one cache miss per layer beyond the one-cache-miss
240 * depth. The total number of cache misses for heap management during a
241 * downheap operation is thus bounded by
242 *
243 * lg (n) - lg (cache_line_size / sizeof (cyc_index_t))
244 *
245 * Traditional pointer-based heaps are implemented without regard to
246 * locality. Downheaps can thus incur two cache misses per layer (one for
247 * each child), but at most one cache miss at the root. This yields a bound
248 * of
249 *
250 * 2 * lg (n) - 1
251 *
252 * on the total cache misses.
253 *
254 * This difference may seem theoretically trivial (the difference is, after
255 * all, constant), but can become substantial in practice -- especially for
256 * caches with very large cache lines and high miss penalties (e.g. TLBs).
257 *
258 * Heaps must always be full, balanced trees. Heap management must therefore
259 * track the next point-of-insertion into the heap. In pointer-based heaps,
260 * recomputing this point takes O(lg (n)). Given the layout of the
261 * array-based implementation, however, the next point-of-insertion is
262 * always:
263 *
264 * heap[number_of_elements]
265 *
266 * We exploit this property by implementing the free-list in the usused
267 * heap elements. Heap insertion, therefore, consists only of filling in
268 * the cyclic at cyp_cyclics[cyp_heap[number_of_elements]], incrementing
269 * the number of elements, and performing an upheap. Heap deletion consists
270 * of decrementing the number of elements, swapping the to-be-deleted element
271 * with the element at cyp_heap[number_of_elements], and downheaping.
272 *
273 * Filling in more details in our earlier example:
274 *
275 * +--- free list head
276 * |
277 * V
278 *
279 * [0] [1] [2] [3] [4] [5] [6] [7]
280 * +-----+-----+-----+-----+-----+-----+-----+-----+
281 * | | | | | | | | |
282 * | 2 | 3 | 4 | 0 | 1 | 5 | 6 | 7 |
283 * | | | | | | | | |
284 * +-----+-----+-----+-----+-----+-----+-----+-----+
285 *
286 * To insert into this heap, we would just need to fill in the cyclic at
287 * cyp_cyclics[5], bump the number of elements (from 5 to 6) and perform
288 * an upheap.
289 *
290 * If we wanted to remove, say, cyp_cyclics[3], we would first scan for it
291 * in the cyp_heap, and discover it at cyp_heap[1]. We would then decrement
292 * the number of elements (from 5 to 4), swap cyp_heap[1] with cyp_heap[4],
293 * and perform a downheap from cyp_heap[1]. The linear scan is required
294 * because the cyclic does not keep a backpointer into the heap. This makes
295 * heap manipulation (e.g. downheaps) faster at the expense of removal
296 * operations.
297 *
298 * Expiry processing
299 *
300 * As alluded to above, cyclic_expire() is called by cyclic_fire() at
301 * CY_HIGH_LEVEL to expire a cyclic. Cyclic subsystem consumers are
302 * guaranteed that for an arbitrary time t in the future, their cyclic
303 * handler will have been called (t - cyt_when) / cyt_interval times. Thus,
304 * there must be a one-to-one mapping between a cyclic's expiration at
305 * CY_HIGH_LEVEL and its execution at the desired level (either CY_HIGH_LEVEL,
306 * CY_LOCK_LEVEL or CY_LOW_LEVEL).
307 *
308 * For CY_HIGH_LEVEL cyclics, this is trivial; cyclic_expire() simply needs
309 * to call the handler.
310 *
311 * For CY_LOCK_LEVEL and CY_LOW_LEVEL cyclics, however, there exists a
312 * potential disconnect: if the CPU is at an interrupt level less than
313 * CY_HIGH_LEVEL but greater than the level of a cyclic for a period of
314 * time longer than twice the cyclic's interval, the cyclic will be expired
315 * twice before it can be handled.
316 *
317 * To maintain the one-to-one mapping, we track the difference between the
318 * number of times a cyclic has been expired and the number of times it's
319 * been handled in a "pending count" (the cy_pend field of the cyclic
320 * structure). cyclic_expire() thus increments the cy_pend count for the
321 * expired cyclic and posts a soft interrupt at the desired level. In the
322 * cyclic subsystem's soft interrupt handler, cyclic_softint(), we repeatedly
323 * call the cyclic handler and decrement cy_pend until we have decremented
324 * cy_pend to zero.
325 *
326 * The Producer/Consumer Buffer
327 *
328 * If we wish to avoid a linear scan of the cyclics array at soft interrupt
329 * level, cyclic_softint() must be able to quickly determine which cyclics
330 * have a non-zero cy_pend count. We thus introduce a per-soft interrupt
331 * level producer/consumer buffer shared with CY_HIGH_LEVEL. These buffers
332 * are encapsulated in the cyc_pcbuffer structure, and, like cyp_heap, are
333 * implemented as cyc_index_t arrays (the cypc_buf member of the cyc_pcbuffer
334 * structure).
335 *
336 * The producer (cyclic_expire() running at CY_HIGH_LEVEL) enqueues a cyclic
337 * by storing the cyclic's index to cypc_buf[cypc_prodndx] and incrementing
338 * cypc_prodndx. The consumer (cyclic_softint() running at either
339 * CY_LOCK_LEVEL or CY_LOW_LEVEL) dequeues a cyclic by loading from
340 * cypc_buf[cypc_consndx] and bumping cypc_consndx. The buffer is empty when
341 * cypc_prodndx == cypc_consndx.
342 *
343 * To bound the size of the producer/consumer buffer, cyclic_expire() only
344 * enqueues a cyclic if its cy_pend was zero (if the cyclic's cy_pend is
345 * non-zero, cyclic_expire() only bumps cy_pend). Symmetrically,
346 * cyclic_softint() only consumes a cyclic after it has decremented the
347 * cy_pend count to zero.
348 *
349 * Returning to our example, here is what the CY_LOW_LEVEL producer/consumer
350 * buffer might look like:
351 *
352 * cypc_consndx ---+ +--- cypc_prodndx
353 * | |
354 * V V
355 *
356 * [0] [1] [2] [3] [4] [5] [6] [7]
357 * +-----+-----+-----+-----+-----+-----+-----+-----+
358 * | | | | | | | | |
359 * | x | x | 3 | 2 | 4 | x | x | x | <== cypc_buf
360 * | | | . | . | . | | | |
361 * +-----+-----+- | -+- | -+- | -+-----+-----+-----+
362 * | | |
363 * | | | cy_pend cy_handler
364 * | | | -------------------------
365 * | | | [ 0] 1 clock()
366 * | | | [ 1] 0 deadman()
367 * | +---- | -------> [ 2] 3 clock_highres_fire()
368 * +---------- | -------> [ 3] 1 clock_highres_fire()
369 * +--------> [ 4] 1 clock_highres_fire()
370 * [ 5] - (free)
371 * [ 6] - (free)
372 * [ 7] - (free)
373 *
374 * In particular, note that clock()'s cy_pend is 1 but that it is _not_ in
375 * this producer/consumer buffer; it would be enqueued in the CY_LOCK_LEVEL
376 * producer/consumer buffer.
377 *
378 * Locking
379 *
380 * Traditionally, access to per-CPU data structures shared between
381 * interrupt levels is serialized by manipulating programmable interrupt
382 * level: readers and writers are required to raise their interrupt level
383 * to that of the highest level writer.
384 *
385 * For the producer/consumer buffers (shared between cyclic_fire()/
386 * cyclic_expire() executing at CY_HIGH_LEVEL and cyclic_softint() executing
387 * at one of CY_LOCK_LEVEL or CY_LOW_LEVEL), forcing cyclic_softint() to raise
388 * programmable interrupt level is undesirable: aside from the additional
389 * latency incurred by manipulating interrupt level in the hot cy_pend
390 * processing path, this would create the potential for soft level cy_pend
391 * processing to delay CY_HIGH_LEVEL firing and expiry processing.
392 * CY_LOCK/LOW_LEVEL cyclics could thereby induce jitter in CY_HIGH_LEVEL
393 * cyclics.
394 *
395 * To minimize jitter, then, we would like the cyclic_fire()/cyclic_expire()
396 * and cyclic_softint() code paths to be lock-free.
397 *
398 * For cyclic_fire()/cyclic_expire(), lock-free execution is straightforward:
399 * because these routines execute at a higher interrupt level than
400 * cyclic_softint(), their actions on the producer/consumer buffer appear
401 * atomic. In particular, the increment of cy_pend appears to occur
402 * atomically with the increment of cypc_prodndx.
403 *
404 * For cyclic_softint(), however, lock-free execution requires more delicacy.
405 * When cyclic_softint() discovers a cyclic in the producer/consumer buffer,
406 * it calls the cyclic's handler and attempts to atomically decrement the
407 * cy_pend count with a compare&swap operation.
408 *
409 * If the compare&swap operation succeeds, cyclic_softint() behaves
410 * conditionally based on the value it atomically wrote to cy_pend:
411 *
412 * - If the cy_pend was decremented to 0, the cyclic has been consumed;
413 * cyclic_softint() increments the cypc_consndx and checks for more
414 * enqueued work.
415 *
416 * - If the count was decremented to a non-zero value, there is more work
417 * to be done on the cyclic; cyclic_softint() calls the cyclic handler
418 * and repeats the atomic decrement process.
419 *
420 * If the compare&swap operation fails, cyclic_softint() knows that
421 * cyclic_expire() has intervened and bumped the cy_pend count (resizes
422 * and removals complicate this, however -- see the sections on their
423 * operation, below). cyclic_softint() thus reloads cy_pend, and re-attempts
424 * the atomic decrement.
425 *
426 * Recall that we bound the size of the producer/consumer buffer by
427 * having cyclic_expire() only enqueue the specified cyclic if its
428 * cy_pend count is zero; this assures that each cyclic is enqueued at
429 * most once. This leads to a critical constraint on cyclic_softint(),
430 * however: after the compare&swap operation which successfully decrements
431 * cy_pend to zero, cyclic_softint() must _not_ re-examine the consumed
432 * cyclic. In part to obey this constraint, cyclic_softint() calls the
433 * cyclic handler before decrementing cy_pend.
434 *
435 * Resizing
436 *
437 * All of the discussion thus far has assumed a static number of cyclics.
438 * Obviously, static limitations are not practical; we need the capacity
439 * to resize our data structures dynamically.
440 *
441 * We resize our data structures lazily, and only on a per-CPU basis.
442 * The size of the data structures always doubles and never shrinks. We
443 * serialize adds (and thus resizes) on cpu_lock; we never need to deal
444 * with concurrent resizes. Resizes should be rare; they may induce jitter
445 * on the CPU being resized, but should not affect cyclic operation on other
446 * CPUs. Pending cyclics may not be dropped during a resize operation.
447 *
448 * Three key cyc_cpu data structures need to be resized: the cyclics array,
449 * the heap array and the producer/consumer buffers. Resizing the first two
450 * is relatively straightforward:
451 *
452 * 1. The new, larger arrays are allocated in cyclic_expand() (called
453 * from cyclic_add()).
454 * 2. cyclic_expand() cross calls cyclic_expand_xcall() on the CPU
455 * undergoing the resize.
456 * 3. cyclic_expand_xcall() raises interrupt level to CY_HIGH_LEVEL
457 * 4. The contents of the old arrays are copied into the new arrays.
458 * 5. The old cyclics array is bzero()'d
459 * 6. The pointers are updated.
460 *
461 * The producer/consumer buffer is dicier: cyclic_expand_xcall() may have
462 * interrupted cyclic_softint() in the middle of consumption. To resize the
463 * producer/consumer buffer, we implement up to two buffers per soft interrupt
464 * level: a hard buffer (the buffer being produced into by cyclic_expire())
465 * and a soft buffer (the buffer from which cyclic_softint() is consuming).
466 * During normal operation, the hard buffer and soft buffer point to the
467 * same underlying producer/consumer buffer.
468 *
469 * During a resize, however, cyclic_expand_xcall() changes the hard buffer
470 * to point to the new, larger producer/consumer buffer; all future
471 * cyclic_expire()'s will produce into the new buffer. cyclic_expand_xcall()
472 * then posts a CY_LOCK_LEVEL soft interrupt, landing in cyclic_softint().
473 *
474 * As under normal operation, cyclic_softint() will consume cyclics from
475 * its soft buffer. After the soft buffer is drained, however,
476 * cyclic_softint() will see that the hard buffer has changed. At that time,
477 * cyclic_softint() will change its soft buffer to point to the hard buffer,
478 * and repeat the producer/consumer buffer draining procedure.
479 *
480 * After the new buffer is drained, cyclic_softint() will determine if both
481 * soft levels have seen their new producer/consumer buffer. If both have,
482 * cyclic_softint() will post on the semaphore cyp_modify_wait. If not, a
483 * soft interrupt will be generated for the remaining level.
484 *
485 * cyclic_expand() blocks on the cyp_modify_wait semaphore (a semaphore is
486 * used instead of a condition variable because of the race between the
487 * sema_p() in cyclic_expand() and the sema_v() in cyclic_softint()). This
488 * allows cyclic_expand() to know when the resize operation is complete;
489 * all of the old buffers (the heap, the cyclics array and the producer/
490 * consumer buffers) can be freed.
491 *
492 * A final caveat on resizing: we described step (5) in the
493 * cyclic_expand_xcall() procedure without providing any motivation. This
494 * step addresses the problem of a cyclic_softint() attempting to decrement
495 * a cy_pend count while interrupted by a cyclic_expand_xcall(). Because
496 * cyclic_softint() has already called the handler by the time cy_pend is
497 * decremented, we want to assure that it doesn't decrement a cy_pend
498 * count in the old cyclics array. By zeroing the old cyclics array in
499 * cyclic_expand_xcall(), we are zeroing out every cy_pend count; when
500 * cyclic_softint() attempts to compare&swap on the cy_pend count, it will
501 * fail and recognize that the count has been zeroed. cyclic_softint() will
502 * update its stale copy of the cyp_cyclics pointer, re-read the cy_pend
503 * count from the new cyclics array, and re-attempt the compare&swap.
504 *
505 * Removals
506 *
507 * Cyclic removals should be rare. To simplify the implementation (and to
508 * allow optimization for the cyclic_fire()/cyclic_expire()/cyclic_softint()
509 * path), we force removals and adds to serialize on cpu_lock.
510 *
511 * Cyclic removal is complicated by a guarantee made to the consumer of
512 * the cyclic subsystem: after cyclic_remove() returns, the cyclic handler
513 * has returned and will never again be called.
514 *
515 * Here is the procedure for cyclic removal:
516 *
517 * 1. cyclic_remove() calls cyclic_remove_xcall() on the CPU undergoing
518 * the removal.
519 * 2. cyclic_remove_xcall() raises interrupt level to CY_HIGH_LEVEL
520 * 3. The current expiration time for the removed cyclic is recorded.
521 * 4. If the cy_pend count on the removed cyclic is non-zero, it
522 * is copied into cyp_rpend and subsequently zeroed.
523 * 5. The cyclic is removed from the heap
524 * 6. If the root of the heap has changed, the backend is reprogrammed.
525 * 7. If the cy_pend count was non-zero cyclic_remove() blocks on the
526 * cyp_modify_wait semaphore.
527 *
528 * The motivation for step (3) is explained in "Juggling", below.
529 *
530 * The cy_pend count is decremented in cyclic_softint() after the cyclic
531 * handler returns. Thus, if we find a cy_pend count of zero in step
532 * (4), we know that cyclic_remove() doesn't need to block.
533 *
534 * If the cy_pend count is non-zero, however, we must block in cyclic_remove()
535 * until cyclic_softint() has finished calling the cyclic handler. To let
536 * cyclic_softint() know that this cyclic has been removed, we zero the
537 * cy_pend count. This will cause cyclic_softint()'s compare&swap to fail.
538 * When cyclic_softint() sees the zero cy_pend count, it knows that it's been
539 * caught during a resize (see "Resizing", above) or that the cyclic has been
540 * removed. In the latter case, it calls cyclic_remove_pend() to call the
541 * cyclic handler cyp_rpend - 1 times, and posts on cyp_modify_wait.
542 *
543 * Juggling
544 *
545 * At first glance, cyclic juggling seems to be a difficult problem. The
546 * subsystem must guarantee that a cyclic doesn't execute simultaneously on
547 * different CPUs, while also assuring that a cyclic fires exactly once
548 * per interval. We solve this problem by leveraging a property of the
549 * platform: gethrtime() is required to increase in lock-step across
550 * multiple CPUs. Therefore, to juggle a cyclic, we remove it from its
551 * CPU, recording its expiration time in the remove cross call (step (3)
552 * in "Removing", above). We then add the cyclic to the new CPU, explicitly
553 * setting its expiration time to the time recorded in the removal. This
554 * leverages the existing cyclic expiry processing, which will compensate
555 * for any time lost while juggling.
556 *
557 * Reprogramming
558 *
559 * Normally, after a cyclic fires, its next expiration is computed from
560 * the current time and the cyclic interval. But there are situations when
561 * the next expiration needs to be reprogrammed by the kernel subsystem that
562 * is using the cyclic. cyclic_reprogram() allows this to be done. This,
563 * unlike the other kernel at-large cyclic API functions, is permitted to
564 * be called from the cyclic handler. This is because it does not use the
565 * cpu_lock to serialize access.
566 *
567 * When cyclic_reprogram() is called for an omni-cyclic, the operation is
568 * applied to the omni-cyclic's component on the current CPU.
569 *
570 * If a high-level cyclic handler reprograms its own cyclic, then
571 * cyclic_fire() detects that and does not recompute the cyclic's next
572 * expiration. However, for a lock-level or a low-level cyclic, the
573 * actual cyclic handler will execute at the lower PIL only after
574 * cyclic_fire() is done with all expired cyclics. To deal with this, such
575 * cyclics can be specified with a special interval of CY_INFINITY (INT64_MAX).
576 * cyclic_fire() recognizes this special value and recomputes the next
577 * expiration to CY_INFINITY. This effectively moves the cyclic to the
578 * bottom of the heap and prevents it from going off until its handler has
579 * had a chance to reprogram it. Infact, this is the way to create and reuse
580 * "one-shot" timers in the context of the cyclic subsystem without using
581 * cyclic_remove().
582 *
583 * Here is the procedure for cyclic reprogramming:
584 *
585 * 1. cyclic_reprogram() calls cyclic_reprogram_xcall() on the CPU
586 * that houses the cyclic.
587 * 2. cyclic_reprogram_xcall() raises interrupt level to CY_HIGH_LEVEL
588 * 3. The cyclic is located in the cyclic heap. The search for this is
589 * done from the bottom of the heap to the top as reprogrammable cyclics
590 * would be located closer to the bottom than the top.
591 * 4. The cyclic expiration is set and the cyclic is moved to its
592 * correct position in the heap (up or down depending on whether the
593 * new expiration is less than or greater than the old one).
594 * 5. If the cyclic move modified the root of the heap, the backend is
595 * reprogrammed.
596 *
597 * Reprogramming can be a frequent event (see the callout subsystem). So,
598 * the serialization used has to be efficient. As with all other cyclic
599 * operations, the interrupt level is raised during reprogramming. Plus,
600 * during reprogramming, the cyclic must not be juggled (regular cyclic)
601 * or stopped (omni-cyclic). The implementation defines a per-cyclic
602 * reader-writer lock to accomplish this. This lock is acquired in the
603 * reader mode by cyclic_reprogram() and writer mode by cyclic_juggle() and
604 * cyclic_omni_stop(). The reader-writer lock makes it efficient if
605 * an omni-cyclic is reprogrammed on different CPUs frequently.
606 *
607 * Note that since the cpu_lock is not used during reprogramming, it is
608 * the responsibility of the user of the reprogrammable cyclic to make sure
609 * that the cyclic is not removed via cyclic_remove() during reprogramming.
610 * This is not an unreasonable requirement as the user will typically have
611 * some sort of synchronization for its cyclic-related activities. This
612 * little caveat exists because the cyclic ID is not really an ID. It is
613 * implemented as a pointer to a structure.
614 *
615 * For cyclics which reprogram themselves during their own handler function,
616 * avoiding the potential race with cyclic_remove() can be a challenge. If a
617 * handler is running and a remote thread issues a cyclic_remove() on its
618 * cyclic (interrupting the handler with the removal xcall), subsequent
619 * attempts to reprogram the cyclics from within the handler will result in a
620 * failure return code from cyclic_reprogram().
621 */
622 #include <sys/cyclic_impl.h>
623 #include <sys/sysmacros.h>
624 #include <sys/systm.h>
625 #include <sys/atomic.h>
626 #include <sys/kmem.h>
627 #include <sys/cmn_err.h>
628 #include <sys/ddi.h>
629 #include <sys/sdt.h>
630
631 #ifdef CYCLIC_TRACE
632
633 /*
634 * cyc_trace_enabled is for the benefit of kernel debuggers.
635 */
636 int cyc_trace_enabled = 1;
637 static cyc_tracebuf_t cyc_ptrace;
638 static cyc_coverage_t cyc_coverage[CY_NCOVERAGE];
639
640 /*
641 * Seen this anywhere?
642 */
643 static uint_t
cyclic_coverage_hash(char * p)644 cyclic_coverage_hash(char *p)
645 {
646 unsigned int g;
647 uint_t hval;
648
649 hval = 0;
650 while (*p) {
651 hval = (hval << 4) + *p++;
652 if ((g = (hval & 0xf0000000)) != 0)
653 hval ^= g >> 24;
654 hval &= ~g;
655 }
656 return (hval);
657 }
658
659 static void
cyclic_coverage(char * why,int level,uint64_t arg0,uint64_t arg1)660 cyclic_coverage(char *why, int level, uint64_t arg0, uint64_t arg1)
661 {
662 uint_t ndx, orig;
663
664 for (ndx = orig = cyclic_coverage_hash(why) % CY_NCOVERAGE; ; ) {
665 if (cyc_coverage[ndx].cyv_why == why)
666 break;
667
668 if (cyc_coverage[ndx].cyv_why != NULL ||
669 atomic_cas_ptr(&cyc_coverage[ndx].cyv_why, NULL, why) !=
670 NULL) {
671
672 if (++ndx == CY_NCOVERAGE)
673 ndx = 0;
674
675 if (ndx == orig)
676 panic("too many cyclic coverage points");
677 continue;
678 }
679
680 /*
681 * If we're here, we have successfully swung our guy into
682 * the position at "ndx".
683 */
684 break;
685 }
686
687 if (level == CY_PASSIVE_LEVEL)
688 cyc_coverage[ndx].cyv_passive_count++;
689 else
690 cyc_coverage[ndx].cyv_count[level]++;
691
692 cyc_coverage[ndx].cyv_arg0 = arg0;
693 cyc_coverage[ndx].cyv_arg1 = arg1;
694 }
695
696 #define CYC_TRACE(cpu, level, why, arg0, arg1) \
697 CYC_TRACE_IMPL(&cpu->cyp_trace[level], level, why, arg0, arg1)
698
699 #define CYC_PTRACE(why, arg0, arg1) \
700 CYC_TRACE_IMPL(&cyc_ptrace, CY_PASSIVE_LEVEL, why, arg0, arg1)
701
702 #define CYC_TRACE_IMPL(buf, level, why, a0, a1) { \
703 if (panicstr == NULL) { \
704 int _ndx = (buf)->cyt_ndx; \
705 cyc_tracerec_t *_rec = &(buf)->cyt_buf[_ndx]; \
706 (buf)->cyt_ndx = (++_ndx == CY_NTRACEREC) ? 0 : _ndx; \
707 _rec->cyt_tstamp = gethrtime_unscaled(); \
708 _rec->cyt_why = (why); \
709 _rec->cyt_arg0 = (uint64_t)(uintptr_t)(a0); \
710 _rec->cyt_arg1 = (uint64_t)(uintptr_t)(a1); \
711 cyclic_coverage(why, level, \
712 (uint64_t)(uintptr_t)(a0), (uint64_t)(uintptr_t)(a1)); \
713 } \
714 }
715
716 #else
717
718 static int cyc_trace_enabled = 0;
719
720 #define CYC_TRACE(cpu, level, why, arg0, arg1)
721 #define CYC_PTRACE(why, arg0, arg1)
722
723 #endif
724
725 #define CYC_TRACE0(cpu, level, why) CYC_TRACE(cpu, level, why, 0, 0)
726 #define CYC_TRACE1(cpu, level, why, arg0) CYC_TRACE(cpu, level, why, arg0, 0)
727
728 #define CYC_PTRACE0(why) CYC_PTRACE(why, 0, 0)
729 #define CYC_PTRACE1(why, arg0) CYC_PTRACE(why, arg0, 0)
730
731 static kmem_cache_t *cyclic_id_cache;
732 static cyc_id_t *cyclic_id_head;
733 static hrtime_t cyclic_resolution;
734 static cyc_backend_t cyclic_backend;
735
736 /*
737 * Returns 1 if the upheap propagated to the root, 0 if it did not. This
738 * allows the caller to reprogram the backend only when the root has been
739 * modified.
740 */
741 static int
cyclic_upheap(cyc_cpu_t * cpu,cyc_index_t ndx)742 cyclic_upheap(cyc_cpu_t *cpu, cyc_index_t ndx)
743 {
744 cyclic_t *cyclics;
745 cyc_index_t *heap;
746 cyc_index_t heap_parent, heap_current = ndx;
747 cyc_index_t parent, current;
748
749 if (heap_current == 0)
750 return (1);
751
752 heap = cpu->cyp_heap;
753 cyclics = cpu->cyp_cyclics;
754 heap_parent = CYC_HEAP_PARENT(heap_current);
755
756 for (;;) {
757 current = heap[heap_current];
758 parent = heap[heap_parent];
759
760 /*
761 * We have an expiration time later than our parent; we're
762 * done.
763 */
764 if (cyclics[current].cy_expire >= cyclics[parent].cy_expire)
765 return (0);
766
767 /*
768 * We need to swap with our parent, and continue up the heap.
769 */
770 heap[heap_parent] = current;
771 heap[heap_current] = parent;
772
773 /*
774 * If we just reached the root, we're done.
775 */
776 if (heap_parent == 0)
777 return (1);
778
779 heap_current = heap_parent;
780 heap_parent = CYC_HEAP_PARENT(heap_current);
781 }
782 }
783
784 static void
cyclic_downheap(cyc_cpu_t * cpu,cyc_index_t ndx)785 cyclic_downheap(cyc_cpu_t *cpu, cyc_index_t ndx)
786 {
787 cyclic_t *cyclics = cpu->cyp_cyclics;
788 cyc_index_t *heap = cpu->cyp_heap;
789
790 cyc_index_t heap_left, heap_right, heap_me = ndx;
791 cyc_index_t left, right, me;
792 cyc_index_t nelems = cpu->cyp_nelems;
793
794 for (;;) {
795 /*
796 * If we don't have a left child (i.e., we're a leaf), we're
797 * done.
798 */
799 if ((heap_left = CYC_HEAP_LEFT(heap_me)) >= nelems)
800 return;
801
802 left = heap[heap_left];
803 me = heap[heap_me];
804
805 heap_right = CYC_HEAP_RIGHT(heap_me);
806
807 /*
808 * Even if we don't have a right child, we still need to compare
809 * our expiration time against that of our left child.
810 */
811 if (heap_right >= nelems)
812 goto comp_left;
813
814 right = heap[heap_right];
815
816 /*
817 * We have both a left and a right child. We need to compare
818 * the expiration times of the children to determine which
819 * expires earlier.
820 */
821 if (cyclics[right].cy_expire < cyclics[left].cy_expire) {
822 /*
823 * Our right child is the earlier of our children.
824 * We'll now compare our expiration time to its; if
825 * ours is the earlier, we're done.
826 */
827 if (cyclics[me].cy_expire <= cyclics[right].cy_expire)
828 return;
829
830 /*
831 * Our right child expires earlier than we do; swap
832 * with our right child, and descend right.
833 */
834 heap[heap_right] = me;
835 heap[heap_me] = right;
836 heap_me = heap_right;
837 continue;
838 }
839
840 comp_left:
841 /*
842 * Our left child is the earlier of our children (or we have
843 * no right child). We'll now compare our expiration time
844 * to its; if ours is the earlier, we're done.
845 */
846 if (cyclics[me].cy_expire <= cyclics[left].cy_expire)
847 return;
848
849 /*
850 * Our left child expires earlier than we do; swap with our
851 * left child, and descend left.
852 */
853 heap[heap_left] = me;
854 heap[heap_me] = left;
855 heap_me = heap_left;
856 }
857 }
858
859 static void
cyclic_expire(cyc_cpu_t * cpu,cyc_index_t ndx,cyclic_t * cyclic)860 cyclic_expire(cyc_cpu_t *cpu, cyc_index_t ndx, cyclic_t *cyclic)
861 {
862 cyc_backend_t *be = cpu->cyp_backend;
863 cyc_level_t level = cyclic->cy_level;
864
865 /*
866 * If this is a CY_HIGH_LEVEL cyclic, just call the handler; we don't
867 * need to worry about the pend count for CY_HIGH_LEVEL cyclics.
868 */
869 if (level == CY_HIGH_LEVEL) {
870 cyc_func_t handler = cyclic->cy_handler;
871 void *arg = cyclic->cy_arg;
872
873 CYC_TRACE(cpu, CY_HIGH_LEVEL, "handler-in", handler, arg);
874 DTRACE_PROBE1(cyclic__start, cyclic_t *, cyclic);
875
876 (*handler)(arg);
877
878 DTRACE_PROBE1(cyclic__end, cyclic_t *, cyclic);
879 CYC_TRACE(cpu, CY_HIGH_LEVEL, "handler-out", handler, arg);
880
881 return;
882 }
883
884 /*
885 * We're at CY_HIGH_LEVEL; this modification to cy_pend need not
886 * be atomic (the high interrupt level assures that it will appear
887 * atomic to any softint currently running).
888 */
889 if (cyclic->cy_pend++ == 0) {
890 cyc_softbuf_t *softbuf = &cpu->cyp_softbuf[level];
891 cyc_pcbuffer_t *pc = &softbuf->cys_buf[softbuf->cys_hard];
892
893 /*
894 * We need to enqueue this cyclic in the soft buffer.
895 */
896 CYC_TRACE(cpu, CY_HIGH_LEVEL, "expire-enq", cyclic,
897 pc->cypc_prodndx);
898 pc->cypc_buf[pc->cypc_prodndx++ & pc->cypc_sizemask] = ndx;
899
900 ASSERT(pc->cypc_prodndx != pc->cypc_consndx);
901 } else {
902 /*
903 * If the pend count is zero after we incremented it, then
904 * we've wrapped (i.e. we had a cy_pend count of over four
905 * billion. In this case, we clamp the pend count at
906 * UINT32_MAX. Yes, cyclics can be lost in this case.
907 */
908 if (cyclic->cy_pend == 0) {
909 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "expire-wrap", cyclic);
910 cyclic->cy_pend = UINT32_MAX;
911 }
912
913 CYC_TRACE(cpu, CY_HIGH_LEVEL, "expire-bump", cyclic, 0);
914 }
915
916 be->cyb_softint(be->cyb_arg, cyclic->cy_level);
917 }
918
919 /*
920 * cyclic_fire(cpu_t *)
921 *
922 * Overview
923 *
924 * cyclic_fire() is the cyclic subsystem's CY_HIGH_LEVEL interrupt handler.
925 * Called by the cyclic backend.
926 *
927 * Arguments and notes
928 *
929 * The only argument is the CPU on which the interrupt is executing;
930 * backends must call into cyclic_fire() on the specified CPU.
931 *
932 * cyclic_fire() may be called spuriously without ill effect. Optimal
933 * backends will call into cyclic_fire() at or shortly after the time
934 * requested via cyb_reprogram(). However, calling cyclic_fire()
935 * arbitrarily late will only manifest latency bubbles; the correctness
936 * of the cyclic subsystem does not rely on the timeliness of the backend.
937 *
938 * cyclic_fire() is wait-free; it will not block or spin.
939 *
940 * Return values
941 *
942 * None.
943 *
944 * Caller's context
945 *
946 * cyclic_fire() must be called from CY_HIGH_LEVEL interrupt context.
947 */
948 void
cyclic_fire(cpu_t * c)949 cyclic_fire(cpu_t *c)
950 {
951 cyc_cpu_t *cpu = c->cpu_cyclic;
952 cyc_backend_t *be = cpu->cyp_backend;
953 cyc_index_t *heap = cpu->cyp_heap;
954 cyclic_t *cyclic, *cyclics = cpu->cyp_cyclics;
955 void *arg = be->cyb_arg;
956 hrtime_t now = gethrtime();
957 hrtime_t exp;
958
959 CYC_TRACE(cpu, CY_HIGH_LEVEL, "fire", now, 0);
960
961 if (cpu->cyp_nelems == 0) {
962 /*
963 * This is a spurious fire. Count it as such, and blow
964 * out of here.
965 */
966 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "fire-spurious");
967 return;
968 }
969
970 for (;;) {
971 cyc_index_t ndx = heap[0];
972
973 cyclic = &cyclics[ndx];
974
975 ASSERT(!(cyclic->cy_flags & CYF_FREE));
976
977 CYC_TRACE(cpu, CY_HIGH_LEVEL, "fire-check", cyclic,
978 cyclic->cy_expire);
979
980 if ((exp = cyclic->cy_expire) > now)
981 break;
982
983 cyclic_expire(cpu, ndx, cyclic);
984
985 /*
986 * If the handler reprogrammed the cyclic, then don't
987 * recompute the expiration. Then, if the interval is
988 * infinity, set the expiration to infinity. This can
989 * be used to create one-shot timers.
990 */
991 if (exp != cyclic->cy_expire) {
992 /*
993 * If a hi level cyclic reprograms itself,
994 * the heap adjustment and reprogramming of the
995 * clock source have already been done at this
996 * point. So, we can continue.
997 */
998 continue;
999 }
1000
1001 if (cyclic->cy_interval == CY_INFINITY)
1002 exp = CY_INFINITY;
1003 else
1004 exp += cyclic->cy_interval;
1005
1006 /*
1007 * If this cyclic will be set to next expire in the distant
1008 * past, we have one of two situations:
1009 *
1010 * a) This is the first firing of a cyclic which had
1011 * cy_expire set to 0.
1012 *
1013 * b) We are tragically late for a cyclic -- most likely
1014 * due to being in the debugger.
1015 *
1016 * In either case, we set the new expiration time to be the
1017 * the next interval boundary. This assures that the
1018 * expiration time modulo the interval is invariant.
1019 *
1020 * We arbitrarily define "distant" to be one second (one second
1021 * is chosen because it's shorter than any foray to the
1022 * debugger while still being longer than any legitimate
1023 * stretch at CY_HIGH_LEVEL).
1024 */
1025
1026 if (now - exp > NANOSEC) {
1027 hrtime_t interval = cyclic->cy_interval;
1028
1029 CYC_TRACE(cpu, CY_HIGH_LEVEL, exp == interval ?
1030 "fire-first" : "fire-swing", now, exp);
1031
1032 exp += ((now - exp) / interval + 1) * interval;
1033 }
1034
1035 cyclic->cy_expire = exp;
1036 cyclic_downheap(cpu, 0);
1037 }
1038
1039 /*
1040 * Now we have a cyclic in the root slot which isn't in the past;
1041 * reprogram the interrupt source.
1042 */
1043 be->cyb_reprogram(arg, exp);
1044 }
1045
1046 static void
cyclic_remove_pend(cyc_cpu_t * cpu,cyc_level_t level,cyclic_t * cyclic)1047 cyclic_remove_pend(cyc_cpu_t *cpu, cyc_level_t level, cyclic_t *cyclic)
1048 {
1049 cyc_func_t handler = cyclic->cy_handler;
1050 void *arg = cyclic->cy_arg;
1051 uint32_t i, rpend = cpu->cyp_rpend - 1;
1052
1053 ASSERT(cyclic->cy_flags & CYF_FREE);
1054 ASSERT(cyclic->cy_pend == 0);
1055 ASSERT(cpu->cyp_state == CYS_REMOVING);
1056 ASSERT(cpu->cyp_rpend > 0);
1057
1058 CYC_TRACE(cpu, level, "remove-rpend", cyclic, cpu->cyp_rpend);
1059
1060 /*
1061 * Note that we only call the handler cyp_rpend - 1 times; this is
1062 * to account for the handler call in cyclic_softint().
1063 */
1064 for (i = 0; i < rpend; i++) {
1065 CYC_TRACE(cpu, level, "rpend-in", handler, arg);
1066 DTRACE_PROBE1(cyclic__start, cyclic_t *, cyclic);
1067
1068 (*handler)(arg);
1069
1070 DTRACE_PROBE1(cyclic__end, cyclic_t *, cyclic);
1071 CYC_TRACE(cpu, level, "rpend-out", handler, arg);
1072 }
1073
1074 /*
1075 * We can now let the remove operation complete.
1076 */
1077 sema_v(&cpu->cyp_modify_wait);
1078 }
1079
1080 /*
1081 * cyclic_softint(cpu_t *cpu, cyc_level_t level)
1082 *
1083 * Overview
1084 *
1085 * cyclic_softint() is the cyclic subsystem's CY_LOCK_LEVEL and CY_LOW_LEVEL
1086 * soft interrupt handler. Called by the cyclic backend.
1087 *
1088 * Arguments and notes
1089 *
1090 * The first argument to cyclic_softint() is the CPU on which the interrupt
1091 * is executing; backends must call into cyclic_softint() on the specified
1092 * CPU. The second argument is the level of the soft interrupt; it must
1093 * be one of CY_LOCK_LEVEL or CY_LOW_LEVEL.
1094 *
1095 * cyclic_softint() will call the handlers for cyclics pending at the
1096 * specified level. cyclic_softint() will not return until all pending
1097 * cyclics at the specified level have been dealt with; intervening
1098 * CY_HIGH_LEVEL interrupts which enqueue cyclics at the specified level
1099 * may therefore prolong cyclic_softint().
1100 *
1101 * cyclic_softint() never disables interrupts, and, if neither a
1102 * cyclic_add() nor a cyclic_remove() is pending on the specified CPU, is
1103 * lock-free. This assures that in the common case, cyclic_softint()
1104 * completes without blocking, and never starves cyclic_fire(). If either
1105 * cyclic_add() or cyclic_remove() is pending, cyclic_softint() may grab
1106 * a dispatcher lock.
1107 *
1108 * While cyclic_softint() is designed for bounded latency, it is obviously
1109 * at the mercy of its cyclic handlers. Because cyclic handlers may block
1110 * arbitrarily, callers of cyclic_softint() should not rely upon
1111 * deterministic completion.
1112 *
1113 * cyclic_softint() may be called spuriously without ill effect.
1114 *
1115 * Return value
1116 *
1117 * None.
1118 *
1119 * Caller's context
1120 *
1121 * The caller must be executing in soft interrupt context at either
1122 * CY_LOCK_LEVEL or CY_LOW_LEVEL. The level passed to cyclic_softint()
1123 * must match the level at which it is executing. On optimal backends,
1124 * the caller will hold no locks. In any case, the caller may not hold
1125 * cpu_lock or any lock acquired by any cyclic handler or held across
1126 * any of cyclic_add(), cyclic_remove(), cyclic_bind() or cyclic_juggle().
1127 */
1128 void
cyclic_softint(cpu_t * c,cyc_level_t level)1129 cyclic_softint(cpu_t *c, cyc_level_t level)
1130 {
1131 cyc_cpu_t *cpu = c->cpu_cyclic;
1132 cyc_softbuf_t *softbuf;
1133 int soft, *buf, consndx, resized = 0, intr_resized = 0;
1134 cyc_pcbuffer_t *pc;
1135 cyclic_t *cyclics = cpu->cyp_cyclics;
1136 int sizemask;
1137
1138 CYC_TRACE(cpu, level, "softint", cyclics, 0);
1139
1140 ASSERT(level < CY_LOW_LEVEL + CY_SOFT_LEVELS);
1141
1142 softbuf = &cpu->cyp_softbuf[level];
1143 top:
1144 soft = softbuf->cys_soft;
1145 ASSERT(soft == 0 || soft == 1);
1146
1147 pc = &softbuf->cys_buf[soft];
1148 buf = pc->cypc_buf;
1149 consndx = pc->cypc_consndx;
1150 sizemask = pc->cypc_sizemask;
1151
1152 CYC_TRACE(cpu, level, "softint-top", cyclics, pc);
1153
1154 while (consndx != pc->cypc_prodndx) {
1155 uint32_t pend, npend, opend;
1156 int consmasked = consndx & sizemask;
1157 cyclic_t *cyclic = &cyclics[buf[consmasked]];
1158 cyc_func_t handler = cyclic->cy_handler;
1159 void *arg = cyclic->cy_arg;
1160
1161 ASSERT(buf[consmasked] < cpu->cyp_size);
1162 CYC_TRACE(cpu, level, "consuming", consndx, cyclic);
1163
1164 /*
1165 * We have found this cyclic in the pcbuffer. We know that
1166 * one of the following is true:
1167 *
1168 * (a) The pend is non-zero. We need to execute the handler
1169 * at least once.
1170 *
1171 * (b) The pend _was_ non-zero, but it's now zero due to a
1172 * resize. We will call the handler once, see that we
1173 * are in this case, and read the new cyclics buffer
1174 * (and hence the old non-zero pend).
1175 *
1176 * (c) The pend _was_ non-zero, but it's now zero due to a
1177 * removal. We will call the handler once, see that we
1178 * are in this case, and call into cyclic_remove_pend()
1179 * to call the cyclic rpend times. We will take into
1180 * account that we have already called the handler once.
1181 *
1182 * Point is: it's safe to call the handler without first
1183 * checking the pend.
1184 */
1185 do {
1186 CYC_TRACE(cpu, level, "handler-in", handler, arg);
1187 DTRACE_PROBE1(cyclic__start, cyclic_t *, cyclic);
1188
1189 (*handler)(arg);
1190
1191 DTRACE_PROBE1(cyclic__end, cyclic_t *, cyclic);
1192 CYC_TRACE(cpu, level, "handler-out", handler, arg);
1193 reread:
1194 pend = cyclic->cy_pend;
1195 npend = pend - 1;
1196
1197 if (pend == 0) {
1198 if (cpu->cyp_state == CYS_REMOVING) {
1199 /*
1200 * This cyclic has been removed while
1201 * it had a non-zero pend count (we
1202 * know it was non-zero because we
1203 * found this cyclic in the pcbuffer).
1204 * There must be a non-zero rpend for
1205 * this CPU, and there must be a remove
1206 * operation blocking; we'll call into
1207 * cyclic_remove_pend() to clean this
1208 * up, and break out of the pend loop.
1209 */
1210 cyclic_remove_pend(cpu, level, cyclic);
1211 break;
1212 }
1213
1214 /*
1215 * We must have had a resize interrupt us.
1216 */
1217 CYC_TRACE(cpu, level, "resize-int", cyclics, 0);
1218 ASSERT(cpu->cyp_state == CYS_EXPANDING);
1219 ASSERT(cyclics != cpu->cyp_cyclics);
1220 ASSERT(resized == 0);
1221 ASSERT(intr_resized == 0);
1222 intr_resized = 1;
1223 cyclics = cpu->cyp_cyclics;
1224 cyclic = &cyclics[buf[consmasked]];
1225 ASSERT(cyclic->cy_handler == handler);
1226 ASSERT(cyclic->cy_arg == arg);
1227 goto reread;
1228 }
1229
1230 if ((opend =
1231 atomic_cas_32(&cyclic->cy_pend, pend, npend)) !=
1232 pend) {
1233 /*
1234 * Our atomic_cas_32 can fail for one of several
1235 * reasons:
1236 *
1237 * (a) An intervening high level bumped up the
1238 * pend count on this cyclic. In this
1239 * case, we will see a higher pend.
1240 *
1241 * (b) The cyclics array has been yanked out
1242 * from underneath us by a resize
1243 * operation. In this case, pend is 0 and
1244 * cyp_state is CYS_EXPANDING.
1245 *
1246 * (c) The cyclic has been removed by an
1247 * intervening remove-xcall. In this case,
1248 * pend will be 0, the cyp_state will be
1249 * CYS_REMOVING, and the cyclic will be
1250 * marked CYF_FREE.
1251 *
1252 * The assertion below checks that we are
1253 * in one of the above situations. The
1254 * action under all three is to return to
1255 * the top of the loop.
1256 */
1257 CYC_TRACE(cpu, level, "cas-fail", opend, pend);
1258 ASSERT(opend > pend || (opend == 0 &&
1259 ((cyclics != cpu->cyp_cyclics &&
1260 cpu->cyp_state == CYS_EXPANDING) ||
1261 (cpu->cyp_state == CYS_REMOVING &&
1262 (cyclic->cy_flags & CYF_FREE)))));
1263 goto reread;
1264 }
1265
1266 /*
1267 * Okay, so we've managed to successfully decrement
1268 * pend. If we just decremented the pend to 0, we're
1269 * done.
1270 */
1271 } while (npend > 0);
1272
1273 pc->cypc_consndx = ++consndx;
1274 }
1275
1276 /*
1277 * If the high level handler is no longer writing to the same
1278 * buffer, then we've had a resize. We need to switch our soft
1279 * index, and goto top.
1280 */
1281 if (soft != softbuf->cys_hard) {
1282 /*
1283 * We can assert that the other buffer has grown by exactly
1284 * one factor of two.
1285 */
1286 CYC_TRACE(cpu, level, "buffer-grow", 0, 0);
1287 ASSERT(cpu->cyp_state == CYS_EXPANDING);
1288 ASSERT(softbuf->cys_buf[softbuf->cys_hard].cypc_sizemask ==
1289 (softbuf->cys_buf[soft].cypc_sizemask << 1) + 1 ||
1290 softbuf->cys_buf[soft].cypc_sizemask == 0);
1291 ASSERT(softbuf->cys_hard == (softbuf->cys_soft ^ 1));
1292
1293 /*
1294 * If our cached cyclics pointer doesn't match cyp_cyclics,
1295 * then we took a resize between our last iteration of the
1296 * pend loop and the check against softbuf->cys_hard.
1297 */
1298 if (cpu->cyp_cyclics != cyclics) {
1299 CYC_TRACE1(cpu, level, "resize-int-int", consndx);
1300 cyclics = cpu->cyp_cyclics;
1301 }
1302
1303 softbuf->cys_soft = softbuf->cys_hard;
1304
1305 ASSERT(resized == 0);
1306 resized = 1;
1307 goto top;
1308 }
1309
1310 /*
1311 * If we were interrupted by a resize operation, then we must have
1312 * seen the hard index change.
1313 */
1314 ASSERT(!(intr_resized == 1 && resized == 0));
1315
1316 if (resized) {
1317 uint32_t lev, nlev;
1318
1319 ASSERT(cpu->cyp_state == CYS_EXPANDING);
1320
1321 do {
1322 lev = cpu->cyp_modify_levels;
1323 nlev = lev + 1;
1324 } while (atomic_cas_32(&cpu->cyp_modify_levels, lev, nlev) !=
1325 lev);
1326
1327 /*
1328 * If we are the last soft level to see the modification,
1329 * post on cyp_modify_wait. Otherwise, (if we're not
1330 * already at low level), post down to the next soft level.
1331 */
1332 if (nlev == CY_SOFT_LEVELS) {
1333 CYC_TRACE0(cpu, level, "resize-kick");
1334 sema_v(&cpu->cyp_modify_wait);
1335 } else {
1336 ASSERT(nlev < CY_SOFT_LEVELS);
1337 if (level != CY_LOW_LEVEL) {
1338 cyc_backend_t *be = cpu->cyp_backend;
1339
1340 CYC_TRACE0(cpu, level, "resize-post");
1341 be->cyb_softint(be->cyb_arg, level - 1);
1342 }
1343 }
1344 }
1345 }
1346
1347 static void
cyclic_expand_xcall(cyc_xcallarg_t * arg)1348 cyclic_expand_xcall(cyc_xcallarg_t *arg)
1349 {
1350 cyc_cpu_t *cpu = arg->cyx_cpu;
1351 cyc_backend_t *be = cpu->cyp_backend;
1352 cyb_arg_t bar = be->cyb_arg;
1353 cyc_cookie_t cookie;
1354 cyc_index_t new_size = arg->cyx_size, size = cpu->cyp_size, i;
1355 cyc_index_t *new_heap = arg->cyx_heap;
1356 cyclic_t *cyclics = cpu->cyp_cyclics, *new_cyclics = arg->cyx_cyclics;
1357
1358 ASSERT(cpu->cyp_state == CYS_EXPANDING);
1359
1360 /*
1361 * This is a little dicey. First, we'll raise our interrupt level
1362 * to CY_HIGH_LEVEL. This CPU already has a new heap, cyclic array,
1363 * etc.; we just need to bcopy them across. As for the softint
1364 * buffers, we'll switch the active buffers. The actual softints will
1365 * take care of consuming any pending cyclics in the old buffer.
1366 */
1367 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
1368
1369 CYC_TRACE(cpu, CY_HIGH_LEVEL, "expand", new_size, 0);
1370
1371 /*
1372 * Assert that the new size is a power of 2.
1373 */
1374 ASSERT((new_size & new_size - 1) == 0);
1375 ASSERT(new_size == (size << 1));
1376 ASSERT(cpu->cyp_heap != NULL && cpu->cyp_cyclics != NULL);
1377
1378 bcopy(cpu->cyp_heap, new_heap, sizeof (cyc_index_t) * size);
1379 bcopy(cyclics, new_cyclics, sizeof (cyclic_t) * size);
1380
1381 /*
1382 * Now run through the old cyclics array, setting pend to 0. To
1383 * softints (which are executing at a lower priority level), the
1384 * pends dropping to 0 will appear atomic with the cyp_cyclics
1385 * pointer changing.
1386 */
1387 for (i = 0; i < size; i++)
1388 cyclics[i].cy_pend = 0;
1389
1390 /*
1391 * Set up the free list, and set all of the new cyclics to be CYF_FREE.
1392 */
1393 for (i = size; i < new_size; i++) {
1394 new_heap[i] = i;
1395 new_cyclics[i].cy_flags = CYF_FREE;
1396 }
1397
1398 /*
1399 * We can go ahead and plow the value of cyp_heap and cyp_cyclics;
1400 * cyclic_expand() has kept a copy.
1401 */
1402 cpu->cyp_heap = new_heap;
1403 cpu->cyp_cyclics = new_cyclics;
1404 cpu->cyp_size = new_size;
1405
1406 /*
1407 * We've switched over the heap and the cyclics array. Now we need
1408 * to switch over our active softint buffer pointers.
1409 */
1410 for (i = CY_LOW_LEVEL; i < CY_LOW_LEVEL + CY_SOFT_LEVELS; i++) {
1411 cyc_softbuf_t *softbuf = &cpu->cyp_softbuf[i];
1412 uchar_t hard = softbuf->cys_hard;
1413
1414 /*
1415 * Assert that we're not in the middle of a resize operation.
1416 */
1417 ASSERT(hard == softbuf->cys_soft);
1418 ASSERT(hard == 0 || hard == 1);
1419 ASSERT(softbuf->cys_buf[hard].cypc_buf != NULL);
1420
1421 softbuf->cys_hard = hard ^ 1;
1422
1423 /*
1424 * The caller (cyclic_expand()) is responsible for setting
1425 * up the new producer-consumer buffer; assert that it's
1426 * been done correctly.
1427 */
1428 ASSERT(softbuf->cys_buf[hard ^ 1].cypc_buf != NULL);
1429 ASSERT(softbuf->cys_buf[hard ^ 1].cypc_prodndx == 0);
1430 ASSERT(softbuf->cys_buf[hard ^ 1].cypc_consndx == 0);
1431 }
1432
1433 /*
1434 * That's all there is to it; now we just need to postdown to
1435 * get the softint chain going.
1436 */
1437 be->cyb_softint(bar, CY_HIGH_LEVEL - 1);
1438 be->cyb_restore_level(bar, cookie);
1439 }
1440
1441 /*
1442 * cyclic_expand() will cross call onto the CPU to perform the actual
1443 * expand operation.
1444 */
1445 static void
cyclic_expand(cyc_cpu_t * cpu)1446 cyclic_expand(cyc_cpu_t *cpu)
1447 {
1448 cyc_index_t new_size, old_size;
1449 cyc_index_t *new_heap, *old_heap;
1450 cyclic_t *new_cyclics, *old_cyclics;
1451 cyc_xcallarg_t arg;
1452 cyc_backend_t *be = cpu->cyp_backend;
1453 char old_hard;
1454 int i;
1455
1456 ASSERT(MUTEX_HELD(&cpu_lock));
1457 ASSERT(cpu->cyp_state == CYS_ONLINE);
1458
1459 cpu->cyp_state = CYS_EXPANDING;
1460
1461 old_heap = cpu->cyp_heap;
1462 old_cyclics = cpu->cyp_cyclics;
1463
1464 if ((new_size = ((old_size = cpu->cyp_size) << 1)) == 0) {
1465 new_size = CY_DEFAULT_PERCPU;
1466 ASSERT(old_heap == NULL && old_cyclics == NULL);
1467 }
1468
1469 /*
1470 * Check that the new_size is a power of 2.
1471 */
1472 ASSERT((new_size - 1 & new_size) == 0);
1473
1474 new_heap = kmem_alloc(sizeof (cyc_index_t) * new_size, KM_SLEEP);
1475 new_cyclics = kmem_zalloc(sizeof (cyclic_t) * new_size, KM_SLEEP);
1476
1477 /*
1478 * We know that no other expansions are in progress (they serialize
1479 * on cpu_lock), so we can safely read the softbuf metadata.
1480 */
1481 old_hard = cpu->cyp_softbuf[0].cys_hard;
1482
1483 for (i = CY_LOW_LEVEL; i < CY_LOW_LEVEL + CY_SOFT_LEVELS; i++) {
1484 cyc_softbuf_t *softbuf = &cpu->cyp_softbuf[i];
1485 char hard = softbuf->cys_hard;
1486 cyc_pcbuffer_t *pc = &softbuf->cys_buf[hard ^ 1];
1487
1488 ASSERT(hard == old_hard);
1489 ASSERT(hard == softbuf->cys_soft);
1490 ASSERT(pc->cypc_buf == NULL);
1491
1492 pc->cypc_buf =
1493 kmem_alloc(sizeof (cyc_index_t) * new_size, KM_SLEEP);
1494 pc->cypc_prodndx = pc->cypc_consndx = 0;
1495 pc->cypc_sizemask = new_size - 1;
1496 }
1497
1498 arg.cyx_cpu = cpu;
1499 arg.cyx_heap = new_heap;
1500 arg.cyx_cyclics = new_cyclics;
1501 arg.cyx_size = new_size;
1502
1503 cpu->cyp_modify_levels = 0;
1504
1505 be->cyb_xcall(be->cyb_arg, cpu->cyp_cpu,
1506 (cyc_func_t)cyclic_expand_xcall, &arg);
1507
1508 /*
1509 * Now block, waiting for the resize operation to complete.
1510 */
1511 sema_p(&cpu->cyp_modify_wait);
1512 ASSERT(cpu->cyp_modify_levels == CY_SOFT_LEVELS);
1513
1514 /*
1515 * The operation is complete; we can now free the old buffers.
1516 */
1517 for (i = CY_LOW_LEVEL; i < CY_LOW_LEVEL + CY_SOFT_LEVELS; i++) {
1518 cyc_softbuf_t *softbuf = &cpu->cyp_softbuf[i];
1519 char hard = softbuf->cys_hard;
1520 cyc_pcbuffer_t *pc = &softbuf->cys_buf[hard ^ 1];
1521
1522 ASSERT(hard == (old_hard ^ 1));
1523 ASSERT(hard == softbuf->cys_soft);
1524
1525 if (pc->cypc_buf == NULL)
1526 continue;
1527
1528 ASSERT(pc->cypc_sizemask == ((new_size - 1) >> 1));
1529
1530 kmem_free(pc->cypc_buf,
1531 sizeof (cyc_index_t) * (pc->cypc_sizemask + 1));
1532 pc->cypc_buf = NULL;
1533 }
1534
1535 if (old_cyclics != NULL) {
1536 ASSERT(old_heap != NULL);
1537 ASSERT(old_size != 0);
1538 kmem_free(old_cyclics, sizeof (cyclic_t) * old_size);
1539 kmem_free(old_heap, sizeof (cyc_index_t) * old_size);
1540 }
1541
1542 ASSERT(cpu->cyp_state == CYS_EXPANDING);
1543 cpu->cyp_state = CYS_ONLINE;
1544 }
1545
1546 /*
1547 * cyclic_pick_cpu will attempt to pick a CPU according to the constraints
1548 * specified by the partition, bound CPU, and flags. Additionally,
1549 * cyclic_pick_cpu() will not pick the avoid CPU; it will return NULL if
1550 * the avoid CPU is the only CPU which satisfies the constraints.
1551 *
1552 * If CYF_CPU_BOUND is set in flags, the specified CPU must be non-NULL.
1553 * If CYF_PART_BOUND is set in flags, the specified partition must be non-NULL.
1554 * If both CYF_CPU_BOUND and CYF_PART_BOUND are set, the specified CPU must
1555 * be in the specified partition.
1556 */
1557 static cyc_cpu_t *
cyclic_pick_cpu(cpupart_t * part,cpu_t * bound,cpu_t * avoid,uint16_t flags)1558 cyclic_pick_cpu(cpupart_t *part, cpu_t *bound, cpu_t *avoid, uint16_t flags)
1559 {
1560 cpu_t *c, *start = (part != NULL) ? part->cp_cpulist : CPU;
1561 cpu_t *online = NULL;
1562 uintptr_t offset;
1563
1564 CYC_PTRACE("pick-cpu", part, bound);
1565
1566 ASSERT(!(flags & CYF_CPU_BOUND) || bound != NULL);
1567 ASSERT(!(flags & CYF_PART_BOUND) || part != NULL);
1568
1569 /*
1570 * If we're bound to our CPU, there isn't much choice involved. We
1571 * need to check that the CPU passed as bound is in the cpupart, and
1572 * that the CPU that we're binding to has been configured.
1573 */
1574 if (flags & CYF_CPU_BOUND) {
1575 CYC_PTRACE("pick-cpu-bound", bound, avoid);
1576
1577 if ((flags & CYF_PART_BOUND) && bound->cpu_part != part)
1578 panic("cyclic_pick_cpu: "
1579 "CPU binding contradicts partition binding");
1580
1581 if (bound == avoid)
1582 return (NULL);
1583
1584 if (bound->cpu_cyclic == NULL)
1585 panic("cyclic_pick_cpu: "
1586 "attempt to bind to non-configured CPU");
1587
1588 return (bound->cpu_cyclic);
1589 }
1590
1591 if (flags & CYF_PART_BOUND) {
1592 CYC_PTRACE("pick-part-bound", bound, avoid);
1593 offset = offsetof(cpu_t, cpu_next_part);
1594 } else {
1595 offset = offsetof(cpu_t, cpu_next_onln);
1596 }
1597
1598 c = start;
1599 do {
1600 if (c->cpu_cyclic == NULL)
1601 continue;
1602
1603 if (c->cpu_cyclic->cyp_state == CYS_OFFLINE)
1604 continue;
1605
1606 if (c == avoid)
1607 continue;
1608
1609 if (c->cpu_flags & CPU_ENABLE)
1610 goto found;
1611
1612 if (online == NULL)
1613 online = c;
1614 } while ((c = *(cpu_t **)((uintptr_t)c + offset)) != start);
1615
1616 /*
1617 * If we're here, we're in one of two situations:
1618 *
1619 * (a) We have a partition-bound cyclic, and there is no CPU in
1620 * our partition which is CPU_ENABLE'd. If we saw another
1621 * non-CYS_OFFLINE CPU in our partition, we'll go with it.
1622 * If not, the avoid CPU must be the only non-CYS_OFFLINE
1623 * CPU in the partition; we're forced to return NULL.
1624 *
1625 * (b) We have a partition-unbound cyclic, in which case there
1626 * must only be one CPU CPU_ENABLE'd, and it must be the one
1627 * we're trying to avoid. If cyclic_juggle()/cyclic_offline()
1628 * are called appropriately, this generally shouldn't happen
1629 * (the offline should fail before getting to this code).
1630 * At any rate: we can't avoid the avoid CPU, so we return
1631 * NULL.
1632 */
1633 if (!(flags & CYF_PART_BOUND)) {
1634 ASSERT(avoid->cpu_flags & CPU_ENABLE);
1635 return (NULL);
1636 }
1637
1638 CYC_PTRACE("pick-no-intr", part, avoid);
1639
1640 if ((c = online) != NULL)
1641 goto found;
1642
1643 CYC_PTRACE("pick-fail", part, avoid);
1644 ASSERT(avoid->cpu_part == start->cpu_part);
1645 return (NULL);
1646
1647 found:
1648 CYC_PTRACE("pick-cpu-found", c, avoid);
1649 ASSERT(c != avoid);
1650 ASSERT(c->cpu_cyclic != NULL);
1651
1652 return (c->cpu_cyclic);
1653 }
1654
1655 static void
cyclic_add_xcall(cyc_xcallarg_t * arg)1656 cyclic_add_xcall(cyc_xcallarg_t *arg)
1657 {
1658 cyc_cpu_t *cpu = arg->cyx_cpu;
1659 cyc_handler_t *hdlr = arg->cyx_hdlr;
1660 cyc_time_t *when = arg->cyx_when;
1661 cyc_backend_t *be = cpu->cyp_backend;
1662 cyc_index_t ndx, nelems;
1663 cyc_cookie_t cookie;
1664 cyb_arg_t bar = be->cyb_arg;
1665 cyclic_t *cyclic;
1666
1667 ASSERT(cpu->cyp_nelems < cpu->cyp_size);
1668
1669 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
1670
1671 CYC_TRACE(cpu, CY_HIGH_LEVEL,
1672 "add-xcall", when->cyt_when, when->cyt_interval);
1673
1674 nelems = cpu->cyp_nelems++;
1675
1676 if (nelems == 0) {
1677 /*
1678 * If this is the first element, we need to enable the
1679 * backend on this CPU.
1680 */
1681 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "enabled");
1682 be->cyb_enable(bar);
1683 }
1684
1685 ndx = cpu->cyp_heap[nelems];
1686 cyclic = &cpu->cyp_cyclics[ndx];
1687
1688 ASSERT(cyclic->cy_flags == CYF_FREE);
1689 cyclic->cy_interval = when->cyt_interval;
1690
1691 if (when->cyt_when == 0) {
1692 /*
1693 * If a start time hasn't been explicitly specified, we'll
1694 * start on the next interval boundary.
1695 */
1696 cyclic->cy_expire = (gethrtime() / cyclic->cy_interval + 1) *
1697 cyclic->cy_interval;
1698 } else {
1699 cyclic->cy_expire = when->cyt_when;
1700 }
1701
1702 cyclic->cy_handler = hdlr->cyh_func;
1703 cyclic->cy_arg = hdlr->cyh_arg;
1704 cyclic->cy_level = hdlr->cyh_level;
1705 cyclic->cy_flags = arg->cyx_flags;
1706
1707 if (cyclic_upheap(cpu, nelems)) {
1708 hrtime_t exp = cyclic->cy_expire;
1709
1710 CYC_TRACE(cpu, CY_HIGH_LEVEL, "add-reprog", cyclic, exp);
1711
1712 /*
1713 * If our upheap propagated to the root, we need to
1714 * reprogram the interrupt source.
1715 */
1716 be->cyb_reprogram(bar, exp);
1717 }
1718 be->cyb_restore_level(bar, cookie);
1719
1720 arg->cyx_ndx = ndx;
1721 }
1722
1723 static cyc_index_t
cyclic_add_here(cyc_cpu_t * cpu,cyc_handler_t * hdlr,cyc_time_t * when,uint16_t flags)1724 cyclic_add_here(cyc_cpu_t *cpu, cyc_handler_t *hdlr,
1725 cyc_time_t *when, uint16_t flags)
1726 {
1727 cyc_backend_t *be = cpu->cyp_backend;
1728 cyb_arg_t bar = be->cyb_arg;
1729 cyc_xcallarg_t arg;
1730
1731 CYC_PTRACE("add-cpu", cpu, hdlr->cyh_func);
1732 ASSERT(MUTEX_HELD(&cpu_lock));
1733 ASSERT(cpu->cyp_state == CYS_ONLINE);
1734 ASSERT(!(cpu->cyp_cpu->cpu_flags & CPU_OFFLINE));
1735 ASSERT(when->cyt_when >= 0 && when->cyt_interval > 0);
1736
1737 if (cpu->cyp_nelems == cpu->cyp_size) {
1738 /*
1739 * This is expensive; it will cross call onto the other
1740 * CPU to perform the expansion.
1741 */
1742 cyclic_expand(cpu);
1743 ASSERT(cpu->cyp_nelems < cpu->cyp_size);
1744 }
1745
1746 /*
1747 * By now, we know that we're going to be able to successfully
1748 * perform the add. Now cross call over to the CPU of interest to
1749 * actually add our cyclic.
1750 */
1751 arg.cyx_cpu = cpu;
1752 arg.cyx_hdlr = hdlr;
1753 arg.cyx_when = when;
1754 arg.cyx_flags = flags;
1755
1756 be->cyb_xcall(bar, cpu->cyp_cpu, (cyc_func_t)cyclic_add_xcall, &arg);
1757
1758 CYC_PTRACE("add-cpu-done", cpu, arg.cyx_ndx);
1759
1760 return (arg.cyx_ndx);
1761 }
1762
1763 static void
cyclic_remove_xcall(cyc_xcallarg_t * arg)1764 cyclic_remove_xcall(cyc_xcallarg_t *arg)
1765 {
1766 cyc_cpu_t *cpu = arg->cyx_cpu;
1767 cyc_backend_t *be = cpu->cyp_backend;
1768 cyb_arg_t bar = be->cyb_arg;
1769 cyc_cookie_t cookie;
1770 cyc_index_t ndx = arg->cyx_ndx, nelems, i;
1771 cyc_index_t *heap, last;
1772 cyclic_t *cyclic;
1773 #ifdef DEBUG
1774 cyc_index_t root;
1775 #endif
1776
1777 ASSERT(cpu->cyp_state == CYS_REMOVING);
1778
1779 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
1780
1781 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "remove-xcall", ndx);
1782
1783 heap = cpu->cyp_heap;
1784 nelems = cpu->cyp_nelems;
1785 ASSERT(nelems > 0);
1786 cyclic = &cpu->cyp_cyclics[ndx];
1787
1788 /*
1789 * Grab the current expiration time. If this cyclic is being
1790 * removed as part of a juggling operation, the expiration time
1791 * will be used when the cyclic is added to the new CPU.
1792 */
1793 if (arg->cyx_when != NULL) {
1794 arg->cyx_when->cyt_when = cyclic->cy_expire;
1795 arg->cyx_when->cyt_interval = cyclic->cy_interval;
1796 }
1797
1798 if (cyclic->cy_pend != 0) {
1799 /*
1800 * The pend is non-zero; this cyclic is currently being
1801 * executed (or will be executed shortly). If the caller
1802 * refuses to wait, we must return (doing nothing). Otherwise,
1803 * we will stash the pend value * in this CPU's rpend, and
1804 * then zero it out. The softint in the pend loop will see
1805 * that we have zeroed out pend, and will call the cyclic
1806 * handler rpend times. The caller will wait until the
1807 * softint has completed calling the cyclic handler.
1808 */
1809 if (arg->cyx_wait == CY_NOWAIT) {
1810 arg->cyx_wait = CY_WAIT;
1811 goto out;
1812 }
1813
1814 ASSERT(cyclic->cy_level != CY_HIGH_LEVEL);
1815 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "remove-pend", cyclic->cy_pend);
1816 cpu->cyp_rpend = cyclic->cy_pend;
1817 cyclic->cy_pend = 0;
1818 }
1819
1820 /*
1821 * Now set the flags to CYF_FREE. We don't need a membar_enter()
1822 * between zeroing pend and setting the flags because we're at
1823 * CY_HIGH_LEVEL (that is, the zeroing of pend and the setting
1824 * of cy_flags appear atomic to softints).
1825 */
1826 cyclic->cy_flags = CYF_FREE;
1827
1828 for (i = 0; i < nelems; i++) {
1829 if (heap[i] == ndx)
1830 break;
1831 }
1832
1833 if (i == nelems)
1834 panic("attempt to remove non-existent cyclic");
1835
1836 cpu->cyp_nelems = --nelems;
1837
1838 if (nelems == 0) {
1839 /*
1840 * If we just removed the last element, then we need to
1841 * disable the backend on this CPU.
1842 */
1843 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "disabled");
1844 be->cyb_disable(bar);
1845 }
1846
1847 if (i == nelems) {
1848 /*
1849 * If we just removed the last element of the heap, then
1850 * we don't have to downheap.
1851 */
1852 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "remove-bottom");
1853 goto out;
1854 }
1855
1856 #ifdef DEBUG
1857 root = heap[0];
1858 #endif
1859
1860 /*
1861 * Swap the last element of the heap with the one we want to
1862 * remove, and downheap (this has the implicit effect of putting
1863 * the newly freed element on the free list).
1864 */
1865 heap[i] = (last = heap[nelems]);
1866 heap[nelems] = ndx;
1867
1868 if (i == 0) {
1869 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "remove-root");
1870 cyclic_downheap(cpu, 0);
1871 } else {
1872 if (cyclic_upheap(cpu, i) == 0) {
1873 /*
1874 * The upheap didn't propagate to the root; if it
1875 * didn't propagate at all, we need to downheap.
1876 */
1877 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "remove-no-root");
1878 if (heap[i] == last) {
1879 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "remove-no-up");
1880 cyclic_downheap(cpu, i);
1881 }
1882 ASSERT(heap[0] == root);
1883 goto out;
1884 }
1885 }
1886
1887 /*
1888 * We're here because we changed the root; we need to reprogram
1889 * the clock source.
1890 */
1891 cyclic = &cpu->cyp_cyclics[heap[0]];
1892
1893 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "remove-reprog");
1894
1895 ASSERT(nelems != 0);
1896 be->cyb_reprogram(bar, cyclic->cy_expire);
1897 out:
1898 be->cyb_restore_level(bar, cookie);
1899 }
1900
1901 static int
cyclic_remove_here(cyc_cpu_t * cpu,cyc_index_t ndx,cyc_time_t * when,int wait)1902 cyclic_remove_here(cyc_cpu_t *cpu, cyc_index_t ndx, cyc_time_t *when, int wait)
1903 {
1904 cyc_backend_t *be = cpu->cyp_backend;
1905 cyc_xcallarg_t arg;
1906 cyclic_t *cyclic = &cpu->cyp_cyclics[ndx];
1907 cyc_level_t level = cyclic->cy_level;
1908
1909 ASSERT(MUTEX_HELD(&cpu_lock));
1910 ASSERT(cpu->cyp_rpend == 0);
1911 ASSERT(wait == CY_WAIT || wait == CY_NOWAIT);
1912
1913 arg.cyx_ndx = ndx;
1914 arg.cyx_cpu = cpu;
1915 arg.cyx_when = when;
1916 arg.cyx_wait = wait;
1917
1918 ASSERT(cpu->cyp_state == CYS_ONLINE);
1919 cpu->cyp_state = CYS_REMOVING;
1920
1921 be->cyb_xcall(be->cyb_arg, cpu->cyp_cpu,
1922 (cyc_func_t)cyclic_remove_xcall, &arg);
1923
1924 /*
1925 * If the cyclic we removed wasn't at CY_HIGH_LEVEL, then we need to
1926 * check the cyp_rpend. If it's non-zero, then we need to wait here
1927 * for all pending cyclic handlers to run.
1928 */
1929 ASSERT(!(level == CY_HIGH_LEVEL && cpu->cyp_rpend != 0));
1930 ASSERT(!(wait == CY_NOWAIT && cpu->cyp_rpend != 0));
1931 ASSERT(!(arg.cyx_wait == CY_NOWAIT && cpu->cyp_rpend != 0));
1932
1933 if (wait != arg.cyx_wait) {
1934 /*
1935 * We are being told that we must wait if we want to
1936 * remove this cyclic; put the CPU back in the CYS_ONLINE
1937 * state and return failure.
1938 */
1939 ASSERT(wait == CY_NOWAIT && arg.cyx_wait == CY_WAIT);
1940 ASSERT(cpu->cyp_state == CYS_REMOVING);
1941 cpu->cyp_state = CYS_ONLINE;
1942
1943 return (0);
1944 }
1945
1946 if (cpu->cyp_rpend != 0)
1947 sema_p(&cpu->cyp_modify_wait);
1948
1949 ASSERT(cpu->cyp_state == CYS_REMOVING);
1950
1951 cpu->cyp_rpend = 0;
1952 cpu->cyp_state = CYS_ONLINE;
1953
1954 return (1);
1955 }
1956
1957 /*
1958 * If cyclic_reprogram() is called on the same CPU as the cyclic's CPU, then
1959 * it calls this function directly. Else, it invokes this function through
1960 * an X-call to the cyclic's CPU.
1961 */
1962 static boolean_t
cyclic_reprogram_cyclic(cyc_cpu_t * cpu,cyc_index_t ndx,hrtime_t expire,boolean_t is_local)1963 cyclic_reprogram_cyclic(cyc_cpu_t *cpu, cyc_index_t ndx, hrtime_t expire,
1964 boolean_t is_local)
1965 {
1966 cyc_backend_t *be = cpu->cyp_backend;
1967 cyb_arg_t bar = be->cyb_arg;
1968 cyc_cookie_t cookie;
1969 cyc_index_t nelems, i;
1970 cyc_index_t *heap;
1971 cyclic_t *cyclic;
1972 hrtime_t oexpire;
1973 int reprog;
1974
1975 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
1976
1977 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "reprog-xcall", ndx);
1978
1979 nelems = cpu->cyp_nelems;
1980 ASSERT(nelems > 0);
1981 heap = cpu->cyp_heap;
1982
1983 /*
1984 * Reprogrammed cyclics are typically one-shot ones that get
1985 * set to infinity on every expiration. We shorten the search by
1986 * searching from the bottom of the heap to the top instead of the
1987 * other way around.
1988 */
1989 for (i = nelems - 1; i >= 0; i--) {
1990 if (heap[i] == ndx)
1991 break;
1992 }
1993 if (i < 0) {
1994 /*
1995 * Report failure (rather than panicking) if and only if the
1996 * cyclic_reprogram() is occurring on the CPU which the cyclic
1997 * resides upon, and there is evidence that a pending cyclic
1998 * was removed from that CPU.
1999 *
2000 * This covers the race where a cyclic is removed out from
2001 * under its running handler, which then attempts a reprogram.
2002 */
2003 if (is_local &&
2004 cpu->cyp_state == CYS_REMOVING && cpu->cyp_rpend > 0) {
2005 return (B_FALSE);
2006 }
2007 panic("attempt to reprogram non-existent cyclic");
2008 }
2009
2010 cyclic = &cpu->cyp_cyclics[ndx];
2011 oexpire = cyclic->cy_expire;
2012 cyclic->cy_expire = expire;
2013
2014 reprog = (i == 0);
2015 if (expire > oexpire) {
2016 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "reprog-down", i);
2017 cyclic_downheap(cpu, i);
2018 } else if (i > 0) {
2019 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "reprog-up", i);
2020 reprog = cyclic_upheap(cpu, i);
2021 }
2022
2023 if (reprog && (cpu->cyp_state != CYS_SUSPENDED)) {
2024 /*
2025 * The root changed. Reprogram the clock source.
2026 */
2027 CYC_TRACE0(cpu, CY_HIGH_LEVEL, "reprog-root");
2028 cyclic = &cpu->cyp_cyclics[heap[0]];
2029 be->cyb_reprogram(bar, cyclic->cy_expire);
2030 }
2031
2032 be->cyb_restore_level(bar, cookie);
2033 return (B_TRUE);
2034 }
2035
2036 static void
cyclic_reprogram_xcall(cyc_xcallarg_t * arg)2037 cyclic_reprogram_xcall(cyc_xcallarg_t *arg)
2038 {
2039 /*
2040 * Cross-call reprogram operations should never fail due to racing
2041 * cyclic removal, as they cannot occur from the handler itself.
2042 */
2043 VERIFY(cyclic_reprogram_cyclic(arg->cyx_cpu, arg->cyx_ndx,
2044 arg->cyx_when->cyt_when, B_FALSE));
2045 }
2046
2047 static void
cyclic_reprogram_here(cyc_cpu_t * cpu,cyc_index_t ndx,hrtime_t expiration)2048 cyclic_reprogram_here(cyc_cpu_t *cpu, cyc_index_t ndx, hrtime_t expiration)
2049 {
2050 cyc_backend_t *be = cpu->cyp_backend;
2051 cyc_xcallarg_t arg;
2052 cyc_time_t when;
2053
2054 ASSERT(expiration > 0);
2055
2056 arg.cyx_ndx = ndx;
2057 arg.cyx_cpu = cpu;
2058 arg.cyx_when = &when;
2059 when.cyt_when = expiration;
2060
2061 be->cyb_xcall(be->cyb_arg, cpu->cyp_cpu,
2062 (cyc_func_t)cyclic_reprogram_xcall, &arg);
2063 }
2064
2065 /*
2066 * cyclic_juggle_one_to() should only be called when the source cyclic
2067 * can be juggled and the destination CPU is known to be able to accept
2068 * it.
2069 */
2070 static void
cyclic_juggle_one_to(cyc_id_t * idp,cyc_cpu_t * dest)2071 cyclic_juggle_one_to(cyc_id_t *idp, cyc_cpu_t *dest)
2072 {
2073 cyc_cpu_t *src = idp->cyi_cpu;
2074 cyc_index_t ndx = idp->cyi_ndx;
2075 cyc_time_t when;
2076 cyc_handler_t hdlr;
2077 cyclic_t *cyclic;
2078 uint16_t flags;
2079 hrtime_t delay;
2080
2081 ASSERT(MUTEX_HELD(&cpu_lock));
2082 ASSERT(src != NULL && idp->cyi_omni_list == NULL);
2083 ASSERT(!(dest->cyp_cpu->cpu_flags & (CPU_QUIESCED | CPU_OFFLINE)));
2084 CYC_PTRACE("juggle-one-to", idp, dest);
2085
2086 cyclic = &src->cyp_cyclics[ndx];
2087
2088 flags = cyclic->cy_flags;
2089 ASSERT(!(flags & CYF_CPU_BOUND) && !(flags & CYF_FREE));
2090
2091 hdlr.cyh_func = cyclic->cy_handler;
2092 hdlr.cyh_level = cyclic->cy_level;
2093 hdlr.cyh_arg = cyclic->cy_arg;
2094
2095 /*
2096 * Before we begin the juggling process, see if the destination
2097 * CPU requires an expansion. If it does, we'll perform the
2098 * expansion before removing the cyclic. This is to prevent us
2099 * from blocking while a system-critical cyclic (notably, the clock
2100 * cyclic) isn't on a CPU.
2101 */
2102 if (dest->cyp_nelems == dest->cyp_size) {
2103 CYC_PTRACE("remove-expand", idp, dest);
2104 cyclic_expand(dest);
2105 ASSERT(dest->cyp_nelems < dest->cyp_size);
2106 }
2107
2108 /*
2109 * Prevent a reprogram of this cyclic while we are relocating it.
2110 * Otherwise, cyclic_reprogram_here() will end up sending an X-call
2111 * to the wrong CPU.
2112 */
2113 rw_enter(&idp->cyi_lock, RW_WRITER);
2114
2115 /*
2116 * Remove the cyclic from the source. As mentioned above, we cannot
2117 * block during this operation; if we cannot remove the cyclic
2118 * without waiting, we spin for a time shorter than the interval, and
2119 * reattempt the (non-blocking) removal. If we continue to fail,
2120 * we will exponentially back off (up to half of the interval).
2121 * Note that the removal will ultimately succeed -- even if the
2122 * cyclic handler is blocked on a resource held by a thread which we
2123 * have preempted, priority inheritance assures that the preempted
2124 * thread will preempt us and continue to progress.
2125 */
2126 for (delay = NANOSEC / MICROSEC; ; delay <<= 1) {
2127 /*
2128 * Before we begin this operation, disable kernel preemption.
2129 */
2130 kpreempt_disable();
2131 if (cyclic_remove_here(src, ndx, &when, CY_NOWAIT))
2132 break;
2133
2134 /*
2135 * The operation failed; enable kernel preemption while
2136 * spinning.
2137 */
2138 kpreempt_enable();
2139
2140 CYC_PTRACE("remove-retry", idp, src);
2141
2142 if (delay > (cyclic->cy_interval >> 1))
2143 delay = cyclic->cy_interval >> 1;
2144
2145 /*
2146 * Drop the RW lock to avoid a deadlock with the cyclic
2147 * handler (because it can potentially call cyclic_reprogram().
2148 */
2149 rw_exit(&idp->cyi_lock);
2150 drv_usecwait((clock_t)(delay / (NANOSEC / MICROSEC)));
2151 rw_enter(&idp->cyi_lock, RW_WRITER);
2152 }
2153
2154 /*
2155 * Now add the cyclic to the destination. This won't block; we
2156 * performed any necessary (blocking) expansion of the destination
2157 * CPU before removing the cyclic from the source CPU.
2158 */
2159 idp->cyi_ndx = cyclic_add_here(dest, &hdlr, &when, flags);
2160 idp->cyi_cpu = dest;
2161 kpreempt_enable();
2162
2163 /*
2164 * Now that we have successfully relocated the cyclic, allow
2165 * it to be reprogrammed.
2166 */
2167 rw_exit(&idp->cyi_lock);
2168 }
2169
2170 static int
cyclic_juggle_one(cyc_id_t * idp)2171 cyclic_juggle_one(cyc_id_t *idp)
2172 {
2173 cyc_index_t ndx = idp->cyi_ndx;
2174 cyc_cpu_t *cpu = idp->cyi_cpu, *dest;
2175 cyclic_t *cyclic = &cpu->cyp_cyclics[ndx];
2176 cpu_t *c = cpu->cyp_cpu;
2177 cpupart_t *part = c->cpu_part;
2178
2179 CYC_PTRACE("juggle-one", idp, cpu);
2180 ASSERT(MUTEX_HELD(&cpu_lock));
2181 ASSERT(!(c->cpu_flags & CPU_OFFLINE));
2182 ASSERT(cpu->cyp_state == CYS_ONLINE);
2183 ASSERT(!(cyclic->cy_flags & CYF_FREE));
2184
2185 if ((dest = cyclic_pick_cpu(part, c, c, cyclic->cy_flags)) == NULL) {
2186 /*
2187 * Bad news: this cyclic can't be juggled.
2188 */
2189 CYC_PTRACE("juggle-fail", idp, cpu)
2190 return (0);
2191 }
2192
2193 cyclic_juggle_one_to(idp, dest);
2194
2195 return (1);
2196 }
2197
2198 static void
cyclic_unbind_cpu(cyclic_id_t id)2199 cyclic_unbind_cpu(cyclic_id_t id)
2200 {
2201 cyc_id_t *idp = (cyc_id_t *)id;
2202 cyc_cpu_t *cpu = idp->cyi_cpu;
2203 cpu_t *c = cpu->cyp_cpu;
2204 cyclic_t *cyclic = &cpu->cyp_cyclics[idp->cyi_ndx];
2205
2206 CYC_PTRACE("unbind-cpu", id, cpu);
2207 ASSERT(MUTEX_HELD(&cpu_lock));
2208 ASSERT(cpu->cyp_state == CYS_ONLINE);
2209 ASSERT(!(cyclic->cy_flags & CYF_FREE));
2210 ASSERT(cyclic->cy_flags & CYF_CPU_BOUND);
2211
2212 cyclic->cy_flags &= ~CYF_CPU_BOUND;
2213
2214 /*
2215 * If we were bound to CPU which has interrupts disabled, we need
2216 * to juggle away. This can only fail if we are bound to a
2217 * processor set, and if every CPU in the processor set has
2218 * interrupts disabled.
2219 */
2220 if (!(c->cpu_flags & CPU_ENABLE)) {
2221 int res = cyclic_juggle_one(idp);
2222
2223 ASSERT((res && idp->cyi_cpu != cpu) ||
2224 (!res && (cyclic->cy_flags & CYF_PART_BOUND)));
2225 }
2226 }
2227
2228 static void
cyclic_bind_cpu(cyclic_id_t id,cpu_t * d)2229 cyclic_bind_cpu(cyclic_id_t id, cpu_t *d)
2230 {
2231 cyc_id_t *idp = (cyc_id_t *)id;
2232 cyc_cpu_t *dest = d->cpu_cyclic, *cpu = idp->cyi_cpu;
2233 cpu_t *c = cpu->cyp_cpu;
2234 cyclic_t *cyclic = &cpu->cyp_cyclics[idp->cyi_ndx];
2235 cpupart_t *part = c->cpu_part;
2236
2237 CYC_PTRACE("bind-cpu", id, dest);
2238 ASSERT(MUTEX_HELD(&cpu_lock));
2239 ASSERT(!(d->cpu_flags & CPU_OFFLINE));
2240 ASSERT(!(c->cpu_flags & CPU_OFFLINE));
2241 ASSERT(cpu->cyp_state == CYS_ONLINE);
2242 ASSERT(dest != NULL);
2243 ASSERT(dest->cyp_state == CYS_ONLINE);
2244 ASSERT(!(cyclic->cy_flags & CYF_FREE));
2245 ASSERT(!(cyclic->cy_flags & CYF_CPU_BOUND));
2246
2247 dest = cyclic_pick_cpu(part, d, NULL, cyclic->cy_flags | CYF_CPU_BOUND);
2248
2249 if (dest != cpu) {
2250 cyclic_juggle_one_to(idp, dest);
2251 cyclic = &dest->cyp_cyclics[idp->cyi_ndx];
2252 }
2253
2254 cyclic->cy_flags |= CYF_CPU_BOUND;
2255 }
2256
2257 static void
cyclic_unbind_cpupart(cyclic_id_t id)2258 cyclic_unbind_cpupart(cyclic_id_t id)
2259 {
2260 cyc_id_t *idp = (cyc_id_t *)id;
2261 cyc_cpu_t *cpu = idp->cyi_cpu;
2262 cpu_t *c = cpu->cyp_cpu;
2263 cyclic_t *cyc = &cpu->cyp_cyclics[idp->cyi_ndx];
2264
2265 CYC_PTRACE("unbind-part", idp, c->cpu_part);
2266 ASSERT(MUTEX_HELD(&cpu_lock));
2267 ASSERT(cpu->cyp_state == CYS_ONLINE);
2268 ASSERT(!(cyc->cy_flags & CYF_FREE));
2269 ASSERT(cyc->cy_flags & CYF_PART_BOUND);
2270
2271 cyc->cy_flags &= ~CYF_PART_BOUND;
2272
2273 /*
2274 * If we're on a CPU which has interrupts disabled (and if this cyclic
2275 * isn't bound to the CPU), we need to juggle away.
2276 */
2277 if (!(c->cpu_flags & CPU_ENABLE) && !(cyc->cy_flags & CYF_CPU_BOUND)) {
2278 int res = cyclic_juggle_one(idp);
2279
2280 ASSERT(res && idp->cyi_cpu != cpu);
2281 }
2282 }
2283
2284 static void
cyclic_bind_cpupart(cyclic_id_t id,cpupart_t * part)2285 cyclic_bind_cpupart(cyclic_id_t id, cpupart_t *part)
2286 {
2287 cyc_id_t *idp = (cyc_id_t *)id;
2288 cyc_cpu_t *cpu = idp->cyi_cpu, *dest;
2289 cpu_t *c = cpu->cyp_cpu;
2290 cyclic_t *cyc = &cpu->cyp_cyclics[idp->cyi_ndx];
2291
2292 CYC_PTRACE("bind-part", idp, part);
2293 ASSERT(MUTEX_HELD(&cpu_lock));
2294 ASSERT(!(c->cpu_flags & CPU_OFFLINE));
2295 ASSERT(cpu->cyp_state == CYS_ONLINE);
2296 ASSERT(!(cyc->cy_flags & CYF_FREE));
2297 ASSERT(!(cyc->cy_flags & CYF_PART_BOUND));
2298 ASSERT(part->cp_ncpus > 0);
2299
2300 dest = cyclic_pick_cpu(part, c, NULL, cyc->cy_flags | CYF_PART_BOUND);
2301
2302 if (dest != cpu) {
2303 cyclic_juggle_one_to(idp, dest);
2304 cyc = &dest->cyp_cyclics[idp->cyi_ndx];
2305 }
2306
2307 cyc->cy_flags |= CYF_PART_BOUND;
2308 }
2309
2310 static void
cyclic_configure(cpu_t * c)2311 cyclic_configure(cpu_t *c)
2312 {
2313 cyc_cpu_t *cpu = kmem_zalloc(sizeof (cyc_cpu_t), KM_SLEEP);
2314 cyc_backend_t *nbe = kmem_zalloc(sizeof (cyc_backend_t), KM_SLEEP);
2315 int i;
2316
2317 CYC_PTRACE1("configure", cpu);
2318 ASSERT(MUTEX_HELD(&cpu_lock));
2319
2320 if (cyclic_id_cache == NULL)
2321 cyclic_id_cache = kmem_cache_create("cyclic_id_cache",
2322 sizeof (cyc_id_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2323
2324 cpu->cyp_cpu = c;
2325
2326 sema_init(&cpu->cyp_modify_wait, 0, NULL, SEMA_DEFAULT, NULL);
2327
2328 cpu->cyp_size = 1;
2329 cpu->cyp_heap = kmem_zalloc(sizeof (cyc_index_t), KM_SLEEP);
2330 cpu->cyp_cyclics = kmem_zalloc(sizeof (cyclic_t), KM_SLEEP);
2331 cpu->cyp_cyclics->cy_flags = CYF_FREE;
2332
2333 for (i = CY_LOW_LEVEL; i < CY_LOW_LEVEL + CY_SOFT_LEVELS; i++) {
2334 /*
2335 * We don't need to set the sizemask; it's already zero
2336 * (which is the appropriate sizemask for a size of 1).
2337 */
2338 cpu->cyp_softbuf[i].cys_buf[0].cypc_buf =
2339 kmem_alloc(sizeof (cyc_index_t), KM_SLEEP);
2340 }
2341
2342 cpu->cyp_state = CYS_OFFLINE;
2343
2344 /*
2345 * Setup the backend for this CPU.
2346 */
2347 bcopy(&cyclic_backend, nbe, sizeof (cyc_backend_t));
2348 nbe->cyb_arg = nbe->cyb_configure(c);
2349 cpu->cyp_backend = nbe;
2350
2351 /*
2352 * On platforms where stray interrupts may be taken during startup,
2353 * the CPU's cpu_cyclic pointer serves as an indicator that the
2354 * cyclic subsystem for this CPU is prepared to field interrupts.
2355 */
2356 membar_producer();
2357
2358 c->cpu_cyclic = cpu;
2359 }
2360
2361 static void
cyclic_unconfigure(cpu_t * c)2362 cyclic_unconfigure(cpu_t *c)
2363 {
2364 cyc_cpu_t *cpu = c->cpu_cyclic;
2365 cyc_backend_t *be = cpu->cyp_backend;
2366 cyb_arg_t bar = be->cyb_arg;
2367 int i;
2368
2369 CYC_PTRACE1("unconfigure", cpu);
2370 ASSERT(MUTEX_HELD(&cpu_lock));
2371 ASSERT(cpu->cyp_state == CYS_OFFLINE);
2372 ASSERT(cpu->cyp_nelems == 0);
2373
2374 /*
2375 * Let the backend know that the CPU is being yanked, and free up
2376 * the backend structure.
2377 */
2378 be->cyb_unconfigure(bar);
2379 kmem_free(be, sizeof (cyc_backend_t));
2380 cpu->cyp_backend = NULL;
2381
2382 /*
2383 * Free up the producer/consumer buffers at each of the soft levels.
2384 */
2385 for (i = CY_LOW_LEVEL; i < CY_LOW_LEVEL + CY_SOFT_LEVELS; i++) {
2386 cyc_softbuf_t *softbuf = &cpu->cyp_softbuf[i];
2387 uchar_t hard = softbuf->cys_hard;
2388 cyc_pcbuffer_t *pc = &softbuf->cys_buf[hard];
2389 size_t bufsize = sizeof (cyc_index_t) * (pc->cypc_sizemask + 1);
2390
2391 /*
2392 * Assert that we're not in the middle of a resize operation.
2393 */
2394 ASSERT(hard == softbuf->cys_soft);
2395 ASSERT(hard == 0 || hard == 1);
2396 ASSERT(pc->cypc_buf != NULL);
2397 ASSERT(softbuf->cys_buf[hard ^ 1].cypc_buf == NULL);
2398
2399 kmem_free(pc->cypc_buf, bufsize);
2400 pc->cypc_buf = NULL;
2401 }
2402
2403 /*
2404 * Finally, clean up our remaining dynamic structures and NULL out
2405 * the cpu_cyclic pointer.
2406 */
2407 kmem_free(cpu->cyp_cyclics, cpu->cyp_size * sizeof (cyclic_t));
2408 kmem_free(cpu->cyp_heap, cpu->cyp_size * sizeof (cyc_index_t));
2409 kmem_free(cpu, sizeof (cyc_cpu_t));
2410
2411 c->cpu_cyclic = NULL;
2412 }
2413
2414 static int
cyclic_cpu_setup(cpu_setup_t what,int id,void * arg __unused)2415 cyclic_cpu_setup(cpu_setup_t what, int id, void *arg __unused)
2416 {
2417 /*
2418 * We are guaranteed that there is still/already an entry in the
2419 * cpu array for this CPU.
2420 */
2421 cpu_t *c = cpu[id];
2422 cyc_cpu_t *cyp = c->cpu_cyclic;
2423
2424 ASSERT(MUTEX_HELD(&cpu_lock));
2425
2426 switch (what) {
2427 case CPU_CONFIG:
2428 ASSERT(cyp == NULL);
2429 cyclic_configure(c);
2430 break;
2431
2432 case CPU_UNCONFIG:
2433 ASSERT(cyp != NULL && cyp->cyp_state == CYS_OFFLINE);
2434 cyclic_unconfigure(c);
2435 break;
2436
2437 default:
2438 break;
2439 }
2440
2441 return (0);
2442 }
2443
2444 static void
cyclic_suspend_xcall(cyc_xcallarg_t * arg)2445 cyclic_suspend_xcall(cyc_xcallarg_t *arg)
2446 {
2447 cyc_cpu_t *cpu = arg->cyx_cpu;
2448 cyc_backend_t *be = cpu->cyp_backend;
2449 cyc_cookie_t cookie;
2450 cyb_arg_t bar = be->cyb_arg;
2451
2452 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
2453
2454 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "suspend-xcall", cpu->cyp_nelems);
2455 ASSERT(cpu->cyp_state == CYS_ONLINE || cpu->cyp_state == CYS_OFFLINE);
2456
2457 /*
2458 * We won't disable this CPU unless it has a non-zero number of
2459 * elements (cpu_lock assures that no one else may be attempting
2460 * to disable this CPU).
2461 */
2462 if (cpu->cyp_nelems > 0) {
2463 ASSERT(cpu->cyp_state == CYS_ONLINE);
2464 be->cyb_disable(bar);
2465 }
2466
2467 if (cpu->cyp_state == CYS_ONLINE)
2468 cpu->cyp_state = CYS_SUSPENDED;
2469
2470 be->cyb_suspend(bar);
2471 be->cyb_restore_level(bar, cookie);
2472 }
2473
2474 static void
cyclic_resume_xcall(cyc_xcallarg_t * arg)2475 cyclic_resume_xcall(cyc_xcallarg_t *arg)
2476 {
2477 cyc_cpu_t *cpu = arg->cyx_cpu;
2478 cyc_backend_t *be = cpu->cyp_backend;
2479 cyc_cookie_t cookie;
2480 cyb_arg_t bar = be->cyb_arg;
2481 cyc_state_t state = cpu->cyp_state;
2482
2483 cookie = be->cyb_set_level(bar, CY_HIGH_LEVEL);
2484
2485 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "resume-xcall", cpu->cyp_nelems);
2486 ASSERT(state == CYS_SUSPENDED || state == CYS_OFFLINE);
2487
2488 be->cyb_resume(bar);
2489
2490 /*
2491 * We won't enable this CPU unless it has a non-zero number of
2492 * elements.
2493 */
2494 if (cpu->cyp_nelems > 0) {
2495 cyclic_t *cyclic = &cpu->cyp_cyclics[cpu->cyp_heap[0]];
2496 hrtime_t exp = cyclic->cy_expire;
2497
2498 CYC_TRACE(cpu, CY_HIGH_LEVEL, "resume-reprog", cyclic, exp);
2499 ASSERT(state == CYS_SUSPENDED);
2500 be->cyb_enable(bar);
2501 be->cyb_reprogram(bar, exp);
2502 }
2503
2504 if (state == CYS_SUSPENDED)
2505 cpu->cyp_state = CYS_ONLINE;
2506
2507 CYC_TRACE1(cpu, CY_HIGH_LEVEL, "resume-done", cpu->cyp_nelems);
2508 be->cyb_restore_level(bar, cookie);
2509 }
2510
2511 static void
cyclic_omni_start(cyc_id_t * idp,cyc_cpu_t * cpu)2512 cyclic_omni_start(cyc_id_t *idp, cyc_cpu_t *cpu)
2513 {
2514 cyc_omni_handler_t *omni = &idp->cyi_omni_hdlr;
2515 cyc_omni_cpu_t *ocpu = kmem_alloc(sizeof (cyc_omni_cpu_t), KM_SLEEP);
2516 cyc_handler_t hdlr;
2517 cyc_time_t when;
2518
2519 CYC_PTRACE("omni-start", cpu, idp);
2520 ASSERT(MUTEX_HELD(&cpu_lock));
2521 ASSERT(cpu->cyp_state == CYS_ONLINE);
2522 ASSERT(idp->cyi_cpu == NULL);
2523
2524 hdlr.cyh_func = NULL;
2525 hdlr.cyh_arg = NULL;
2526 hdlr.cyh_level = CY_LEVELS;
2527
2528 when.cyt_when = 0;
2529 when.cyt_interval = 0;
2530
2531 omni->cyo_online(omni->cyo_arg, cpu->cyp_cpu, &hdlr, &when);
2532
2533 ASSERT(hdlr.cyh_func != NULL);
2534 ASSERT(hdlr.cyh_level < CY_LEVELS);
2535 ASSERT(when.cyt_when >= 0 && when.cyt_interval > 0);
2536
2537 ocpu->cyo_cpu = cpu;
2538 ocpu->cyo_arg = hdlr.cyh_arg;
2539 ocpu->cyo_ndx = cyclic_add_here(cpu, &hdlr, &when, 0);
2540 ocpu->cyo_next = idp->cyi_omni_list;
2541 idp->cyi_omni_list = ocpu;
2542 }
2543
2544 static void
cyclic_omni_stop(cyc_id_t * idp,cyc_cpu_t * cpu)2545 cyclic_omni_stop(cyc_id_t *idp, cyc_cpu_t *cpu)
2546 {
2547 cyc_omni_handler_t *omni = &idp->cyi_omni_hdlr;
2548 cyc_omni_cpu_t *ocpu = idp->cyi_omni_list, *prev = NULL;
2549 clock_t delay;
2550 int ret;
2551
2552 CYC_PTRACE("omni-stop", cpu, idp);
2553 ASSERT(MUTEX_HELD(&cpu_lock));
2554 ASSERT(cpu->cyp_state == CYS_ONLINE);
2555 ASSERT(idp->cyi_cpu == NULL);
2556 ASSERT(ocpu != NULL);
2557
2558 /*
2559 * Prevent a reprogram of this cyclic while we are removing it.
2560 * Otherwise, cyclic_reprogram_here() will end up sending an X-call
2561 * to the offlined CPU.
2562 */
2563 rw_enter(&idp->cyi_lock, RW_WRITER);
2564
2565 while (ocpu != NULL && ocpu->cyo_cpu != cpu) {
2566 prev = ocpu;
2567 ocpu = ocpu->cyo_next;
2568 }
2569
2570 /*
2571 * We _must_ have found an cyc_omni_cpu which corresponds to this
2572 * CPU -- the definition of an omnipresent cyclic is that it runs
2573 * on all online CPUs.
2574 */
2575 ASSERT(ocpu != NULL);
2576
2577 if (prev == NULL) {
2578 idp->cyi_omni_list = ocpu->cyo_next;
2579 } else {
2580 prev->cyo_next = ocpu->cyo_next;
2581 }
2582
2583 /*
2584 * Remove the cyclic from the source. We cannot block during this
2585 * operation because we are holding the cyi_lock which can be held
2586 * by the cyclic handler via cyclic_reprogram().
2587 *
2588 * If we cannot remove the cyclic without waiting, we spin for a time,
2589 * and reattempt the (non-blocking) removal. If the handler is blocked
2590 * on the cyi_lock, then we let go of it in the spin loop to give
2591 * the handler a chance to run. Note that the removal will ultimately
2592 * succeed -- even if the cyclic handler is blocked on a resource
2593 * held by a thread which we have preempted, priority inheritance
2594 * assures that the preempted thread will preempt us and continue
2595 * to progress.
2596 */
2597 for (delay = 1; ; delay <<= 1) {
2598 /*
2599 * Before we begin this operation, disable kernel preemption.
2600 */
2601 kpreempt_disable();
2602 ret = cyclic_remove_here(ocpu->cyo_cpu, ocpu->cyo_ndx, NULL,
2603 CY_NOWAIT);
2604 /*
2605 * Enable kernel preemption while spinning.
2606 */
2607 kpreempt_enable();
2608
2609 if (ret)
2610 break;
2611
2612 CYC_PTRACE("remove-omni-retry", idp, ocpu->cyo_cpu);
2613
2614 /*
2615 * Drop the RW lock to avoid a deadlock with the cyclic
2616 * handler (because it can potentially call cyclic_reprogram().
2617 */
2618 rw_exit(&idp->cyi_lock);
2619 drv_usecwait(delay);
2620 rw_enter(&idp->cyi_lock, RW_WRITER);
2621 }
2622
2623 /*
2624 * Now that we have successfully removed the cyclic, allow the omni
2625 * cyclic to be reprogrammed on other CPUs.
2626 */
2627 rw_exit(&idp->cyi_lock);
2628
2629 /*
2630 * The cyclic has been removed from this CPU; time to call the
2631 * omnipresent offline handler.
2632 */
2633 if (omni->cyo_offline != NULL)
2634 omni->cyo_offline(omni->cyo_arg, cpu->cyp_cpu, ocpu->cyo_arg);
2635
2636 kmem_free(ocpu, sizeof (cyc_omni_cpu_t));
2637 }
2638
2639 static cyc_id_t *
cyclic_new_id()2640 cyclic_new_id()
2641 {
2642 cyc_id_t *idp;
2643
2644 ASSERT(MUTEX_HELD(&cpu_lock));
2645
2646 idp = kmem_cache_alloc(cyclic_id_cache, KM_SLEEP);
2647
2648 /*
2649 * The cyi_cpu field of the cyc_id_t structure tracks the CPU
2650 * associated with the cyclic. If and only if this field is NULL, the
2651 * cyc_id_t is an omnipresent cyclic. Note that cyi_omni_list may be
2652 * NULL for an omnipresent cyclic while the cyclic is being created
2653 * or destroyed.
2654 */
2655 idp->cyi_cpu = NULL;
2656 idp->cyi_ndx = 0;
2657 rw_init(&idp->cyi_lock, NULL, RW_DEFAULT, NULL);
2658
2659 idp->cyi_next = cyclic_id_head;
2660 idp->cyi_prev = NULL;
2661 idp->cyi_omni_list = NULL;
2662
2663 if (cyclic_id_head != NULL) {
2664 ASSERT(cyclic_id_head->cyi_prev == NULL);
2665 cyclic_id_head->cyi_prev = idp;
2666 }
2667
2668 cyclic_id_head = idp;
2669
2670 return (idp);
2671 }
2672
2673 /*
2674 * cyclic_id_t cyclic_add(cyc_handler_t *, cyc_time_t *)
2675 *
2676 * Overview
2677 *
2678 * cyclic_add() will create an unbound cyclic with the specified handler and
2679 * interval. The cyclic will run on a CPU which both has interrupts enabled
2680 * and is in the system CPU partition.
2681 *
2682 * Arguments and notes
2683 *
2684 * As its first argument, cyclic_add() takes a cyc_handler, which has the
2685 * following members:
2686 *
2687 * cyc_func_t cyh_func <-- Cyclic handler
2688 * void *cyh_arg <-- Argument to cyclic handler
2689 * cyc_level_t cyh_level <-- Level at which to fire; must be one of
2690 * CY_LOW_LEVEL, CY_LOCK_LEVEL or CY_HIGH_LEVEL
2691 *
2692 * Note that cyh_level is _not_ an ipl or spl; it must be one the
2693 * CY_*_LEVELs. This layer of abstraction allows the platform to define
2694 * the precise interrupt priority levels, within the following constraints:
2695 *
2696 * CY_LOCK_LEVEL must map to LOCK_LEVEL
2697 * CY_HIGH_LEVEL must map to an ipl greater than LOCK_LEVEL
2698 * CY_LOW_LEVEL must map to an ipl below LOCK_LEVEL
2699 *
2700 * In addition to a cyc_handler, cyclic_add() takes a cyc_time, which
2701 * has the following members:
2702 *
2703 * hrtime_t cyt_when <-- Absolute time, in nanoseconds since boot, at
2704 * which to start firing
2705 * hrtime_t cyt_interval <-- Length of interval, in nanoseconds
2706 *
2707 * gethrtime() is the time source for nanoseconds since boot. If cyt_when
2708 * is set to 0, the cyclic will start to fire when cyt_interval next
2709 * divides the number of nanoseconds since boot.
2710 *
2711 * The cyt_interval field _must_ be filled in by the caller; one-shots are
2712 * _not_ explicitly supported by the cyclic subsystem (cyclic_add() will
2713 * assert that cyt_interval is non-zero). The maximum value for either
2714 * field is INT64_MAX; the caller is responsible for assuring that
2715 * cyt_when + cyt_interval <= INT64_MAX. Neither field may be negative.
2716 *
2717 * For an arbitrary time t in the future, the cyclic handler is guaranteed
2718 * to have been called (t - cyt_when) / cyt_interval times. This will
2719 * be true even if interrupts have been disabled for periods greater than
2720 * cyt_interval nanoseconds. In order to compensate for such periods,
2721 * the cyclic handler may be called a finite number of times with an
2722 * arbitrarily small interval.
2723 *
2724 * The cyclic subsystem will not enforce any lower bound on the interval;
2725 * if the interval is less than the time required to process an interrupt,
2726 * the CPU will wedge. It's the responsibility of the caller to assure that
2727 * either the value of the interval is sane, or that its caller has
2728 * sufficient privilege to deny service (i.e. its caller is root).
2729 *
2730 * The cyclic handler is guaranteed to be single threaded, even while the
2731 * cyclic is being juggled between CPUs (see cyclic_juggle(), below).
2732 * That is, a given cyclic handler will never be executed simultaneously
2733 * on different CPUs.
2734 *
2735 * Return value
2736 *
2737 * cyclic_add() returns a cyclic_id_t, which is guaranteed to be a value
2738 * other than CYCLIC_NONE. cyclic_add() cannot fail.
2739 *
2740 * Caller's context
2741 *
2742 * cpu_lock must be held by the caller, and the caller must not be in
2743 * interrupt context. cyclic_add() will perform a KM_SLEEP kernel
2744 * memory allocation, so the usual rules (e.g. p_lock cannot be held)
2745 * apply. A cyclic may be added even in the presence of CPUs that have
2746 * not been configured with respect to the cyclic subsystem, but only
2747 * configured CPUs will be eligible to run the new cyclic.
2748 *
2749 * Cyclic handler's context
2750 *
2751 * Cyclic handlers will be executed in the interrupt context corresponding
2752 * to the specified level (i.e. either high, lock or low level). The
2753 * usual context rules apply.
2754 *
2755 * A cyclic handler may not grab ANY locks held by the caller of any of
2756 * cyclic_add(), cyclic_remove() or cyclic_bind(); the implementation of
2757 * these functions may require blocking on cyclic handler completion.
2758 * Moreover, cyclic handlers may not make any call back into the cyclic
2759 * subsystem.
2760 */
2761 cyclic_id_t
cyclic_add(cyc_handler_t * hdlr,cyc_time_t * when)2762 cyclic_add(cyc_handler_t *hdlr, cyc_time_t *when)
2763 {
2764 cyc_id_t *idp = cyclic_new_id();
2765
2766 ASSERT(MUTEX_HELD(&cpu_lock));
2767 ASSERT(when->cyt_when >= 0 && when->cyt_interval > 0);
2768
2769 idp->cyi_cpu = cyclic_pick_cpu(NULL, NULL, NULL, 0);
2770 idp->cyi_ndx = cyclic_add_here(idp->cyi_cpu, hdlr, when, 0);
2771
2772 return ((uintptr_t)idp);
2773 }
2774
2775 /*
2776 * cyclic_id_t cyclic_add_omni(cyc_omni_handler_t *)
2777 *
2778 * Overview
2779 *
2780 * cyclic_add_omni() will create an omnipresent cyclic with the specified
2781 * online and offline handlers. Omnipresent cyclics run on all online
2782 * CPUs, including CPUs which have unbound interrupts disabled.
2783 *
2784 * Arguments
2785 *
2786 * As its only argument, cyclic_add_omni() takes a cyc_omni_handler, which
2787 * has the following members:
2788 *
2789 * void (*cyo_online)() <-- Online handler
2790 * void (*cyo_offline)() <-- Offline handler
2791 * void *cyo_arg <-- Argument to be passed to on/offline handlers
2792 *
2793 * Online handler
2794 *
2795 * The cyo_online member is a pointer to a function which has the following
2796 * four arguments:
2797 *
2798 * void * <-- Argument (cyo_arg)
2799 * cpu_t * <-- Pointer to CPU about to be onlined
2800 * cyc_handler_t * <-- Pointer to cyc_handler_t; must be filled in
2801 * by omni online handler
2802 * cyc_time_t * <-- Pointer to cyc_time_t; must be filled in by
2803 * omni online handler
2804 *
2805 * The omni cyclic online handler is always called _before_ the omni
2806 * cyclic begins to fire on the specified CPU. As the above argument
2807 * description implies, the online handler must fill in the two structures
2808 * passed to it: the cyc_handler_t and the cyc_time_t. These are the
2809 * same two structures passed to cyclic_add(), outlined above. This
2810 * allows the omni cyclic to have maximum flexibility; different CPUs may
2811 * optionally
2812 *
2813 * (a) have different intervals
2814 * (b) be explicitly in or out of phase with one another
2815 * (c) have different handlers
2816 * (d) have different handler arguments
2817 * (e) fire at different levels
2818 *
2819 * Of these, (e) seems somewhat dubious, but is nonetheless allowed.
2820 *
2821 * The omni online handler is called in the same context as cyclic_add(),
2822 * and has the same liberties: omni online handlers may perform KM_SLEEP
2823 * kernel memory allocations, and may grab locks which are also acquired
2824 * by cyclic handlers. However, omni cyclic online handlers may _not_
2825 * call back into the cyclic subsystem, and should be generally careful
2826 * about calling into arbitrary kernel subsystems.
2827 *
2828 * Offline handler
2829 *
2830 * The cyo_offline member is a pointer to a function which has the following
2831 * three arguments:
2832 *
2833 * void * <-- Argument (cyo_arg)
2834 * cpu_t * <-- Pointer to CPU about to be offlined
2835 * void * <-- CPU's cyclic argument (that is, value
2836 * to which cyh_arg member of the cyc_handler_t
2837 * was set in the omni online handler)
2838 *
2839 * The omni cyclic offline handler is always called _after_ the omni
2840 * cyclic has ceased firing on the specified CPU. Its purpose is to
2841 * allow cleanup of any resources dynamically allocated in the omni cyclic
2842 * online handler. The context of the offline handler is identical to
2843 * that of the online handler; the same constraints and liberties apply.
2844 *
2845 * The offline handler is optional; it may be NULL.
2846 *
2847 * Return value
2848 *
2849 * cyclic_add_omni() returns a cyclic_id_t, which is guaranteed to be a
2850 * value other than CYCLIC_NONE. cyclic_add_omni() cannot fail.
2851 *
2852 * Caller's context
2853 *
2854 * The caller's context is identical to that of cyclic_add(), specified
2855 * above.
2856 */
2857 cyclic_id_t
cyclic_add_omni(cyc_omni_handler_t * omni)2858 cyclic_add_omni(cyc_omni_handler_t *omni)
2859 {
2860 cyc_id_t *idp = cyclic_new_id();
2861 cyc_cpu_t *cpu;
2862 cpu_t *c;
2863
2864 ASSERT(MUTEX_HELD(&cpu_lock));
2865 ASSERT(omni != NULL && omni->cyo_online != NULL);
2866
2867 idp->cyi_omni_hdlr = *omni;
2868
2869 c = cpu_list;
2870 do {
2871 if ((cpu = c->cpu_cyclic) == NULL)
2872 continue;
2873
2874 if (cpu->cyp_state != CYS_ONLINE) {
2875 ASSERT(cpu->cyp_state == CYS_OFFLINE);
2876 continue;
2877 }
2878
2879 cyclic_omni_start(idp, cpu);
2880 } while ((c = c->cpu_next) != cpu_list);
2881
2882 /*
2883 * We must have found at least one online CPU on which to run
2884 * this cyclic.
2885 */
2886 ASSERT(idp->cyi_omni_list != NULL);
2887 ASSERT(idp->cyi_cpu == NULL);
2888
2889 return ((uintptr_t)idp);
2890 }
2891
2892 /*
2893 * void cyclic_remove(cyclic_id_t)
2894 *
2895 * Overview
2896 *
2897 * cyclic_remove() will remove the specified cyclic from the system.
2898 *
2899 * Arguments and notes
2900 *
2901 * The only argument is a cyclic_id returned from either cyclic_add() or
2902 * cyclic_add_omni().
2903 *
2904 * By the time cyclic_remove() returns, the caller is guaranteed that the
2905 * removed cyclic handler has completed execution (this is the same
2906 * semantic that untimeout() provides). As a result, cyclic_remove() may
2907 * need to block, waiting for the removed cyclic to complete execution.
2908 * This leads to an important constraint on the caller: no lock may be
2909 * held across cyclic_remove() that also may be acquired by a cyclic
2910 * handler.
2911 *
2912 * Return value
2913 *
2914 * None; cyclic_remove() always succeeds.
2915 *
2916 * Caller's context
2917 *
2918 * cpu_lock must be held by the caller, and the caller must not be in
2919 * interrupt context. The caller may not hold any locks which are also
2920 * grabbed by any cyclic handler. See "Arguments and notes", above.
2921 */
2922 void
cyclic_remove(cyclic_id_t id)2923 cyclic_remove(cyclic_id_t id)
2924 {
2925 cyc_id_t *idp = (cyc_id_t *)id;
2926 cyc_id_t *prev = idp->cyi_prev, *next = idp->cyi_next;
2927 cyc_cpu_t *cpu = idp->cyi_cpu;
2928
2929 CYC_PTRACE("remove", idp, idp->cyi_cpu);
2930 ASSERT(MUTEX_HELD(&cpu_lock));
2931
2932 if (cpu != NULL) {
2933 (void) cyclic_remove_here(cpu, idp->cyi_ndx, NULL, CY_WAIT);
2934 } else {
2935 ASSERT(idp->cyi_omni_list != NULL);
2936 while (idp->cyi_omni_list != NULL)
2937 cyclic_omni_stop(idp, idp->cyi_omni_list->cyo_cpu);
2938 }
2939
2940 if (prev != NULL) {
2941 ASSERT(cyclic_id_head != idp);
2942 prev->cyi_next = next;
2943 } else {
2944 ASSERT(cyclic_id_head == idp);
2945 cyclic_id_head = next;
2946 }
2947
2948 if (next != NULL)
2949 next->cyi_prev = prev;
2950
2951 kmem_cache_free(cyclic_id_cache, idp);
2952 }
2953
2954 /*
2955 * void cyclic_bind(cyclic_id_t, cpu_t *, cpupart_t *)
2956 *
2957 * Overview
2958 *
2959 * cyclic_bind() atomically changes the CPU and CPU partition bindings
2960 * of a cyclic.
2961 *
2962 * Arguments and notes
2963 *
2964 * The first argument is a cyclic_id retuned from cyclic_add().
2965 * cyclic_bind() may _not_ be called on a cyclic_id returned from
2966 * cyclic_add_omni().
2967 *
2968 * The second argument specifies the CPU to which to bind the specified
2969 * cyclic. If the specified cyclic is bound to a CPU other than the one
2970 * specified, it will be unbound from its bound CPU. Unbinding the cyclic
2971 * from its CPU may cause it to be juggled to another CPU. If the specified
2972 * CPU is non-NULL, the cyclic will be subsequently rebound to the specified
2973 * CPU.
2974 *
2975 * If a CPU with bound cyclics is transitioned into the P_NOINTR state,
2976 * only cyclics not bound to the CPU can be juggled away; CPU-bound cyclics
2977 * will continue to fire on the P_NOINTR CPU. A CPU with bound cyclics
2978 * cannot be offlined (attempts to offline the CPU will return EBUSY).
2979 * Likewise, cyclics may not be bound to an offline CPU; if the caller
2980 * attempts to bind a cyclic to an offline CPU, the cyclic subsystem will
2981 * panic.
2982 *
2983 * The third argument specifies the CPU partition to which to bind the
2984 * specified cyclic. If the specified cyclic is bound to a CPU partition
2985 * other than the one specified, it will be unbound from its bound
2986 * partition. Unbinding the cyclic from its CPU partition may cause it
2987 * to be juggled to another CPU. If the specified CPU partition is
2988 * non-NULL, the cyclic will be subsequently rebound to the specified CPU
2989 * partition.
2990 *
2991 * It is the caller's responsibility to assure that the specified CPU
2992 * partition contains a CPU. If it does not, the cyclic subsystem will
2993 * panic. A CPU partition with bound cyclics cannot be destroyed (attempts
2994 * to destroy the partition will return EBUSY). If a CPU with
2995 * partition-bound cyclics is transitioned into the P_NOINTR state, cyclics
2996 * bound to the CPU's partition (but not bound to the CPU) will be juggled
2997 * away only if there exists another CPU in the partition in the P_ONLINE
2998 * state.
2999 *
3000 * It is the caller's responsibility to assure that the specified CPU and
3001 * CPU partition are self-consistent. If both parameters are non-NULL,
3002 * and the specified CPU partition does not contain the specified CPU, the
3003 * cyclic subsystem will panic.
3004 *
3005 * It is the caller's responsibility to assure that the specified CPU has
3006 * been configured with respect to the cyclic subsystem. Generally, this
3007 * is always true for valid, on-line CPUs. The only periods of time during
3008 * which this may not be true are during MP boot (i.e. after cyclic_init()
3009 * is called but before cyclic_mp_init() is called) or during dynamic
3010 * reconfiguration; cyclic_bind() should only be called with great care
3011 * from these contexts.
3012 *
3013 * Return value
3014 *
3015 * None; cyclic_bind() always succeeds.
3016 *
3017 * Caller's context
3018 *
3019 * cpu_lock must be held by the caller, and the caller must not be in
3020 * interrupt context. The caller may not hold any locks which are also
3021 * grabbed by any cyclic handler.
3022 */
3023 void
cyclic_bind(cyclic_id_t id,cpu_t * d,cpupart_t * part)3024 cyclic_bind(cyclic_id_t id, cpu_t *d, cpupart_t *part)
3025 {
3026 cyc_id_t *idp = (cyc_id_t *)id;
3027 cyc_cpu_t *cpu = idp->cyi_cpu;
3028 cpu_t *c;
3029 uint16_t flags;
3030
3031 CYC_PTRACE("bind", d, part);
3032 ASSERT(MUTEX_HELD(&cpu_lock));
3033 ASSERT(part == NULL || d == NULL || d->cpu_part == part);
3034
3035 if (cpu == NULL) {
3036 ASSERT(idp->cyi_omni_list != NULL);
3037 panic("attempt to change binding of omnipresent cyclic");
3038 }
3039
3040 c = cpu->cyp_cpu;
3041 flags = cpu->cyp_cyclics[idp->cyi_ndx].cy_flags;
3042
3043 if (c != d && (flags & CYF_CPU_BOUND))
3044 cyclic_unbind_cpu(id);
3045
3046 /*
3047 * Reload our cpu (we may have migrated). We don't have to reload
3048 * the flags field here; if we were CYF_PART_BOUND on entry, we are
3049 * CYF_PART_BOUND now.
3050 */
3051 cpu = idp->cyi_cpu;
3052 c = cpu->cyp_cpu;
3053
3054 if (part != c->cpu_part && (flags & CYF_PART_BOUND))
3055 cyclic_unbind_cpupart(id);
3056
3057 /*
3058 * Now reload the flags field, asserting that if we are CPU bound,
3059 * the CPU was specified (and likewise, if we are partition bound,
3060 * the partition was specified).
3061 */
3062 cpu = idp->cyi_cpu;
3063 c = cpu->cyp_cpu;
3064 flags = cpu->cyp_cyclics[idp->cyi_ndx].cy_flags;
3065 ASSERT(!(flags & CYF_CPU_BOUND) || c == d);
3066 ASSERT(!(flags & CYF_PART_BOUND) || c->cpu_part == part);
3067
3068 if (!(flags & CYF_CPU_BOUND) && d != NULL)
3069 cyclic_bind_cpu(id, d);
3070
3071 if (!(flags & CYF_PART_BOUND) && part != NULL)
3072 cyclic_bind_cpupart(id, part);
3073 }
3074
3075 int
cyclic_reprogram(cyclic_id_t id,hrtime_t expiration)3076 cyclic_reprogram(cyclic_id_t id, hrtime_t expiration)
3077 {
3078 cyc_id_t *idp = (cyc_id_t *)id;
3079 cyc_cpu_t *cpu;
3080 cyc_omni_cpu_t *ocpu;
3081 cyc_index_t ndx;
3082 int res = 1;
3083
3084 ASSERT(expiration > 0);
3085
3086 CYC_PTRACE("reprog", idp, idp->cyi_cpu);
3087
3088 kpreempt_disable();
3089
3090 /*
3091 * Prevent the cyclic from moving or disappearing while we reprogram.
3092 */
3093 rw_enter(&idp->cyi_lock, RW_READER);
3094
3095 if (idp->cyi_cpu == NULL) {
3096 ASSERT(curthread->t_preempt > 0);
3097 cpu = CPU->cpu_cyclic;
3098
3099 /*
3100 * For an omni cyclic, we reprogram the cyclic corresponding
3101 * to the current CPU. Look for it in the list.
3102 */
3103 ocpu = idp->cyi_omni_list;
3104 while (ocpu != NULL) {
3105 if (ocpu->cyo_cpu == cpu)
3106 break;
3107 ocpu = ocpu->cyo_next;
3108 }
3109
3110 if (ocpu == NULL) {
3111 /*
3112 * Didn't find it. This means that CPU offline
3113 * must have removed it racing with us. So,
3114 * nothing to do.
3115 */
3116 rw_exit(&idp->cyi_lock);
3117
3118 kpreempt_enable();
3119
3120 return (0);
3121 }
3122 ndx = ocpu->cyo_ndx;
3123 } else {
3124 cpu = idp->cyi_cpu;
3125 ndx = idp->cyi_ndx;
3126 }
3127
3128 if (cpu->cyp_cpu == CPU) {
3129 /*
3130 * If this reprogram is being done as part of a running cyclic
3131 * handler, it is possible that a racing cyclic_remove() on a
3132 * remote CPU will cause it to fail.
3133 */
3134 if (!cyclic_reprogram_cyclic(cpu, ndx, expiration, B_TRUE)) {
3135 res = 0;
3136 }
3137 } else {
3138 cyclic_reprogram_here(cpu, ndx, expiration);
3139 }
3140
3141 /*
3142 * Allow the cyclic to be moved or removed.
3143 */
3144 rw_exit(&idp->cyi_lock);
3145
3146 kpreempt_enable();
3147
3148 return (res);
3149 }
3150
3151 /*
3152 * void cyclic_move_here(cyclic_id_t)
3153 *
3154 * Overview
3155 *
3156 * cyclic_move_here() attempts to shuffle a cyclic onto the current CPU.
3157 *
3158 * Arguments and notes
3159 *
3160 * The first argument is a cyclic_id returned from cyclic_add().
3161 * cyclic_move_here() may _not_ be called on a cyclic_id returned from
3162 * cyclic_add_omni() or one bound to a CPU or partition via cyclic_bind().
3163 *
3164 * This cyclic shuffling is performed on a best-effort basis. If for some
3165 * reason the current CPU is unsuitable or the thread migrates between CPUs
3166 * during the call, the function may return with the cyclic residing on some
3167 * other CPU.
3168 *
3169 * Return value
3170 *
3171 * None; cyclic_move_here() always reports success.
3172 *
3173 * Caller's context
3174 *
3175 * cpu_lock must be held by the caller, and the caller must not be in
3176 * interrupt context. The caller may not hold any locks which are also
3177 * grabbed by any cyclic handler.
3178 */
3179 void
cyclic_move_here(cyclic_id_t id)3180 cyclic_move_here(cyclic_id_t id)
3181 {
3182 cyc_id_t *idp = (cyc_id_t *)id;
3183 cyc_cpu_t *cc = idp->cyi_cpu;
3184 cpu_t *dest = CPU;
3185
3186 ASSERT(MUTEX_HELD(&cpu_lock));
3187 CYC_PTRACE("move_here", idp, dest);
3188 VERIFY3P(cc, !=, NULL);
3189 VERIFY3U(cc->cyp_cyclics[idp->cyi_ndx].cy_flags &
3190 (CYF_CPU_BOUND|CYF_PART_BOUND), ==, 0);
3191
3192 if (cc->cyp_cpu == dest) {
3193 return;
3194 }
3195
3196 /* Is the destination CPU suitable for a migration target? */
3197 if (dest->cpu_cyclic == NULL ||
3198 dest->cpu_cyclic->cyp_state == CYS_OFFLINE ||
3199 (dest->cpu_flags & CPU_ENABLE) == 0) {
3200 return;
3201 }
3202
3203 cyclic_juggle_one_to(idp, dest->cpu_cyclic);
3204 }
3205
3206 hrtime_t
cyclic_getres()3207 cyclic_getres()
3208 {
3209 return (cyclic_resolution);
3210 }
3211
3212 void
cyclic_init(cyc_backend_t * be,hrtime_t resolution)3213 cyclic_init(cyc_backend_t *be, hrtime_t resolution)
3214 {
3215 ASSERT(MUTEX_HELD(&cpu_lock));
3216
3217 CYC_PTRACE("init", be, resolution);
3218 cyclic_resolution = resolution;
3219
3220 /*
3221 * Copy the passed cyc_backend into the backend template. This must
3222 * be done before the CPU can be configured.
3223 */
3224 bcopy(be, &cyclic_backend, sizeof (cyc_backend_t));
3225
3226 /*
3227 * It's safe to look at the "CPU" pointer without disabling kernel
3228 * preemption; cyclic_init() is called only during startup by the
3229 * cyclic backend.
3230 */
3231 cyclic_configure(CPU);
3232 cyclic_online(CPU);
3233 }
3234
3235 /*
3236 * It is assumed that cyclic_mp_init() is called some time after cyclic
3237 * init (and therefore, after cpu0 has been initialized). We grab cpu_lock,
3238 * find the already initialized CPU, and initialize every other CPU with the
3239 * same backend. Finally, we register a cpu_setup function.
3240 */
3241 void
cyclic_mp_init()3242 cyclic_mp_init()
3243 {
3244 cpu_t *c;
3245
3246 mutex_enter(&cpu_lock);
3247
3248 c = cpu_list;
3249 do {
3250 if (c->cpu_cyclic == NULL) {
3251 cyclic_configure(c);
3252 cyclic_online(c);
3253 }
3254 } while ((c = c->cpu_next) != cpu_list);
3255
3256 register_cpu_setup_func(cyclic_cpu_setup, NULL);
3257 mutex_exit(&cpu_lock);
3258 }
3259
3260 /*
3261 * int cyclic_juggle(cpu_t *)
3262 *
3263 * Overview
3264 *
3265 * cyclic_juggle() juggles as many cyclics as possible away from the
3266 * specified CPU; all remaining cyclics on the CPU will either be CPU-
3267 * or partition-bound.
3268 *
3269 * Arguments and notes
3270 *
3271 * The only argument to cyclic_juggle() is the CPU from which cyclics
3272 * should be juggled. CPU-bound cyclics are never juggled; partition-bound
3273 * cyclics are only juggled if the specified CPU is in the P_NOINTR state
3274 * and there exists a P_ONLINE CPU in the partition. The cyclic subsystem
3275 * assures that a cyclic will never fire late or spuriously, even while
3276 * being juggled.
3277 *
3278 * Return value
3279 *
3280 * cyclic_juggle() returns a non-zero value if all cyclics were able to
3281 * be juggled away from the CPU, and zero if one or more cyclics could
3282 * not be juggled away.
3283 *
3284 * Caller's context
3285 *
3286 * cpu_lock must be held by the caller, and the caller must not be in
3287 * interrupt context. The caller may not hold any locks which are also
3288 * grabbed by any cyclic handler. While cyclic_juggle() _may_ be called
3289 * in any context satisfying these constraints, it _must_ be called
3290 * immediately after clearing CPU_ENABLE (i.e. before dropping cpu_lock).
3291 * Failure to do so could result in an assertion failure in the cyclic
3292 * subsystem.
3293 */
3294 int
cyclic_juggle(cpu_t * c)3295 cyclic_juggle(cpu_t *c)
3296 {
3297 cyc_cpu_t *cpu = c->cpu_cyclic;
3298 cyc_id_t *idp;
3299 int all_juggled = 1;
3300
3301 CYC_PTRACE1("juggle", c);
3302 ASSERT(MUTEX_HELD(&cpu_lock));
3303
3304 /*
3305 * We'll go through each cyclic on the CPU, attempting to juggle
3306 * each one elsewhere.
3307 */
3308 for (idp = cyclic_id_head; idp != NULL; idp = idp->cyi_next) {
3309 if (idp->cyi_cpu != cpu)
3310 continue;
3311
3312 if (cyclic_juggle_one(idp) == 0) {
3313 all_juggled = 0;
3314 continue;
3315 }
3316
3317 ASSERT(idp->cyi_cpu != cpu);
3318 }
3319
3320 return (all_juggled);
3321 }
3322
3323 /*
3324 * int cyclic_offline(cpu_t *)
3325 *
3326 * Overview
3327 *
3328 * cyclic_offline() offlines the cyclic subsystem on the specified CPU.
3329 *
3330 * Arguments and notes
3331 *
3332 * The only argument to cyclic_offline() is a CPU to offline.
3333 * cyclic_offline() will attempt to juggle cyclics away from the specified
3334 * CPU.
3335 *
3336 * Return value
3337 *
3338 * cyclic_offline() returns 1 if all cyclics on the CPU were juggled away
3339 * and the cyclic subsystem on the CPU was successfully offlines.
3340 * cyclic_offline returns 0 if some cyclics remain, blocking the cyclic
3341 * offline operation. All remaining cyclics on the CPU will either be
3342 * CPU- or partition-bound.
3343 *
3344 * See the "Arguments and notes" of cyclic_juggle(), below, for more detail
3345 * on cyclic juggling.
3346 *
3347 * Caller's context
3348 *
3349 * The only caller of cyclic_offline() should be the processor management
3350 * subsystem. It is expected that the caller of cyclic_offline() will
3351 * offline the CPU immediately after cyclic_offline() returns success (i.e.
3352 * before dropping cpu_lock). Moreover, it is expected that the caller will
3353 * fail the CPU offline operation if cyclic_offline() returns failure.
3354 */
3355 int
cyclic_offline(cpu_t * c)3356 cyclic_offline(cpu_t *c)
3357 {
3358 cyc_cpu_t *cpu = c->cpu_cyclic;
3359 cyc_id_t *idp;
3360
3361 CYC_PTRACE1("offline", cpu);
3362 ASSERT(MUTEX_HELD(&cpu_lock));
3363
3364 if (!cyclic_juggle(c))
3365 return (0);
3366
3367 /*
3368 * This CPU is headed offline; we need to now stop omnipresent
3369 * cyclic firing on this CPU.
3370 */
3371 for (idp = cyclic_id_head; idp != NULL; idp = idp->cyi_next) {
3372 if (idp->cyi_cpu != NULL)
3373 continue;
3374
3375 /*
3376 * We cannot possibly be offlining the last CPU; cyi_omni_list
3377 * must be non-NULL.
3378 */
3379 ASSERT(idp->cyi_omni_list != NULL);
3380 cyclic_omni_stop(idp, cpu);
3381 }
3382
3383 ASSERT(cpu->cyp_state == CYS_ONLINE);
3384 cpu->cyp_state = CYS_OFFLINE;
3385
3386 return (1);
3387 }
3388
3389 /*
3390 * void cyclic_online(cpu_t *)
3391 *
3392 * Overview
3393 *
3394 * cyclic_online() onlines a CPU previously offlined with cyclic_offline().
3395 *
3396 * Arguments and notes
3397 *
3398 * cyclic_online()'s only argument is a CPU to online. The specified
3399 * CPU must have been previously offlined with cyclic_offline(). After
3400 * cyclic_online() returns, the specified CPU will be eligible to execute
3401 * cyclics.
3402 *
3403 * Return value
3404 *
3405 * None; cyclic_online() always succeeds.
3406 *
3407 * Caller's context
3408 *
3409 * cyclic_online() should only be called by the processor management
3410 * subsystem; cpu_lock must be held.
3411 */
3412 void
cyclic_online(cpu_t * c)3413 cyclic_online(cpu_t *c)
3414 {
3415 cyc_cpu_t *cpu = c->cpu_cyclic;
3416 cyc_id_t *idp;
3417
3418 CYC_PTRACE1("online", cpu);
3419 ASSERT(c->cpu_flags & CPU_ENABLE);
3420 ASSERT(MUTEX_HELD(&cpu_lock));
3421 ASSERT(cpu->cyp_state == CYS_OFFLINE);
3422
3423 cpu->cyp_state = CYS_ONLINE;
3424
3425 /*
3426 * Now that this CPU is open for business, we need to start firing
3427 * all omnipresent cyclics on it.
3428 */
3429 for (idp = cyclic_id_head; idp != NULL; idp = idp->cyi_next) {
3430 if (idp->cyi_cpu != NULL)
3431 continue;
3432
3433 cyclic_omni_start(idp, cpu);
3434 }
3435 }
3436
3437 /*
3438 * void cyclic_move_in(cpu_t *)
3439 *
3440 * Overview
3441 *
3442 * cyclic_move_in() is called by the CPU partition code immediately after
3443 * the specified CPU has moved into a new partition.
3444 *
3445 * Arguments and notes
3446 *
3447 * The only argument to cyclic_move_in() is a CPU which has moved into a
3448 * new partition. If the specified CPU is P_ONLINE, and every other
3449 * CPU in the specified CPU's new partition is P_NOINTR, cyclic_move_in()
3450 * will juggle all partition-bound, CPU-unbound cyclics to the specified
3451 * CPU.
3452 *
3453 * Return value
3454 *
3455 * None; cyclic_move_in() always succeeds.
3456 *
3457 * Caller's context
3458 *
3459 * cyclic_move_in() should _only_ be called immediately after a CPU has
3460 * moved into a new partition, with cpu_lock held. As with other calls
3461 * into the cyclic subsystem, no lock may be held which is also grabbed
3462 * by any cyclic handler.
3463 */
3464 void
cyclic_move_in(cpu_t * d)3465 cyclic_move_in(cpu_t *d)
3466 {
3467 cyc_id_t *idp;
3468 cyc_cpu_t *dest = d->cpu_cyclic;
3469 cyclic_t *cyclic;
3470 cpupart_t *part = d->cpu_part;
3471
3472 CYC_PTRACE("move-in", dest, part);
3473 ASSERT(MUTEX_HELD(&cpu_lock));
3474
3475 /*
3476 * Look for CYF_PART_BOUND cyclics in the new partition. If
3477 * we find one, check to see if it is currently on a CPU which has
3478 * interrupts disabled. If it is (and if this CPU currently has
3479 * interrupts enabled), we'll juggle those cyclics over here.
3480 */
3481 if (!(d->cpu_flags & CPU_ENABLE)) {
3482 CYC_PTRACE1("move-in-none", dest);
3483 return;
3484 }
3485
3486 for (idp = cyclic_id_head; idp != NULL; idp = idp->cyi_next) {
3487 cyc_cpu_t *cpu = idp->cyi_cpu;
3488 cpu_t *c;
3489
3490 /*
3491 * Omnipresent cyclics are exempt from juggling.
3492 */
3493 if (cpu == NULL)
3494 continue;
3495
3496 c = cpu->cyp_cpu;
3497
3498 if (c->cpu_part != part || (c->cpu_flags & CPU_ENABLE))
3499 continue;
3500
3501 cyclic = &cpu->cyp_cyclics[idp->cyi_ndx];
3502
3503 if (cyclic->cy_flags & CYF_CPU_BOUND)
3504 continue;
3505
3506 /*
3507 * We know that this cyclic is bound to its processor set
3508 * (otherwise, it would not be on a CPU with interrupts
3509 * disabled); juggle it to our CPU.
3510 */
3511 ASSERT(cyclic->cy_flags & CYF_PART_BOUND);
3512 cyclic_juggle_one_to(idp, dest);
3513 }
3514
3515 CYC_PTRACE1("move-in-done", dest);
3516 }
3517
3518 /*
3519 * int cyclic_move_out(cpu_t *)
3520 *
3521 * Overview
3522 *
3523 * cyclic_move_out() is called by the CPU partition code immediately before
3524 * the specified CPU is to move out of its partition.
3525 *
3526 * Arguments and notes
3527 *
3528 * The only argument to cyclic_move_out() is a CPU which is to move out of
3529 * its partition.
3530 *
3531 * cyclic_move_out() will attempt to juggle away all partition-bound
3532 * cyclics. If the specified CPU is the last CPU in a partition with
3533 * partition-bound cyclics, cyclic_move_out() will fail. If there exists
3534 * a partition-bound cyclic which is CPU-bound to the specified CPU,
3535 * cyclic_move_out() will fail.
3536 *
3537 * Note that cyclic_move_out() will _only_ attempt to juggle away
3538 * partition-bound cyclics; CPU-bound cyclics which are not partition-bound
3539 * and unbound cyclics are not affected by changing the partition
3540 * affiliation of the CPU.
3541 *
3542 * Return value
3543 *
3544 * cyclic_move_out() returns 1 if all partition-bound cyclics on the CPU
3545 * were juggled away; 0 if some cyclics remain.
3546 *
3547 * Caller's context
3548 *
3549 * cyclic_move_out() should _only_ be called immediately before a CPU has
3550 * moved out of its partition, with cpu_lock held. It is expected that
3551 * the caller of cyclic_move_out() will change the processor set affiliation
3552 * of the specified CPU immediately after cyclic_move_out() returns
3553 * success (i.e. before dropping cpu_lock). Moreover, it is expected that
3554 * the caller will fail the CPU repartitioning operation if cyclic_move_out()
3555 * returns failure. As with other calls into the cyclic subsystem, no lock
3556 * may be held which is also grabbed by any cyclic handler.
3557 */
3558 int
cyclic_move_out(cpu_t * c)3559 cyclic_move_out(cpu_t *c)
3560 {
3561 cyc_id_t *idp;
3562 cyc_cpu_t *cpu = c->cpu_cyclic, *dest;
3563 cyclic_t *cyclic, *cyclics = cpu->cyp_cyclics;
3564 cpupart_t *part = c->cpu_part;
3565
3566 CYC_PTRACE1("move-out", cpu);
3567 ASSERT(MUTEX_HELD(&cpu_lock));
3568
3569 /*
3570 * If there are any CYF_PART_BOUND cyclics on this CPU, we need
3571 * to try to juggle them away.
3572 */
3573 for (idp = cyclic_id_head; idp != NULL; idp = idp->cyi_next) {
3574
3575 if (idp->cyi_cpu != cpu)
3576 continue;
3577
3578 cyclic = &cyclics[idp->cyi_ndx];
3579
3580 if (!(cyclic->cy_flags & CYF_PART_BOUND))
3581 continue;
3582
3583 dest = cyclic_pick_cpu(part, c, c, cyclic->cy_flags);
3584
3585 if (dest == NULL) {
3586 /*
3587 * We can't juggle this cyclic; we need to return
3588 * failure (we won't bother trying to juggle away
3589 * other cyclics).
3590 */
3591 CYC_PTRACE("move-out-fail", cpu, idp);
3592 return (0);
3593 }
3594 cyclic_juggle_one_to(idp, dest);
3595 }
3596
3597 CYC_PTRACE1("move-out-done", cpu);
3598 return (1);
3599 }
3600
3601 /*
3602 * void cyclic_suspend()
3603 *
3604 * Overview
3605 *
3606 * cyclic_suspend() suspends all cyclic activity throughout the cyclic
3607 * subsystem. It should be called only by subsystems which are attempting
3608 * to suspend the entire system (e.g. checkpoint/resume, dynamic
3609 * reconfiguration).
3610 *
3611 * Arguments and notes
3612 *
3613 * cyclic_suspend() takes no arguments. Each CPU with an active cyclic
3614 * disables its backend (offline CPUs disable their backends as part of
3615 * the cyclic_offline() operation), thereby disabling future CY_HIGH_LEVEL
3616 * interrupts.
3617 *
3618 * Note that disabling CY_HIGH_LEVEL interrupts does not completely preclude
3619 * cyclic handlers from being called after cyclic_suspend() returns: if a
3620 * CY_LOCK_LEVEL or CY_LOW_LEVEL interrupt thread was blocked at the time
3621 * of cyclic_suspend(), cyclic handlers at its level may continue to be
3622 * called after the interrupt thread becomes unblocked. The
3623 * post-cyclic_suspend() activity is bounded by the pend count on all
3624 * cyclics at the time of cyclic_suspend(). Callers concerned with more
3625 * than simply disabling future CY_HIGH_LEVEL interrupts must check for
3626 * this condition.
3627 *
3628 * On most platforms, timestamps from gethrtime() and gethrestime() are not
3629 * guaranteed to monotonically increase between cyclic_suspend() and
3630 * cyclic_resume(). However, timestamps are guaranteed to monotonically
3631 * increase across the entire cyclic_suspend()/cyclic_resume() operation.
3632 * That is, every timestamp obtained before cyclic_suspend() will be less
3633 * than every timestamp obtained after cyclic_resume().
3634 *
3635 * Return value
3636 *
3637 * None; cyclic_suspend() always succeeds.
3638 *
3639 * Caller's context
3640 *
3641 * The cyclic subsystem must be configured on every valid CPU;
3642 * cyclic_suspend() may not be called during boot or during dynamic
3643 * reconfiguration. Additionally, cpu_lock must be held, and the caller
3644 * cannot be in high-level interrupt context. However, unlike most other
3645 * cyclic entry points, cyclic_suspend() may be called with locks held
3646 * which are also acquired by CY_LOCK_LEVEL or CY_LOW_LEVEL cyclic
3647 * handlers.
3648 */
3649 void
cyclic_suspend()3650 cyclic_suspend()
3651 {
3652 cpu_t *c;
3653 cyc_cpu_t *cpu;
3654 cyc_xcallarg_t arg;
3655 cyc_backend_t *be;
3656
3657 CYC_PTRACE0("suspend");
3658 ASSERT(MUTEX_HELD(&cpu_lock));
3659 c = cpu_list;
3660
3661 do {
3662 cpu = c->cpu_cyclic;
3663 be = cpu->cyp_backend;
3664 arg.cyx_cpu = cpu;
3665
3666 be->cyb_xcall(be->cyb_arg, c,
3667 (cyc_func_t)cyclic_suspend_xcall, &arg);
3668 } while ((c = c->cpu_next) != cpu_list);
3669 }
3670
3671 /*
3672 * void cyclic_resume()
3673 *
3674 * cyclic_resume() resumes all cyclic activity throughout the cyclic
3675 * subsystem. It should be called only by system-suspending subsystems.
3676 *
3677 * Arguments and notes
3678 *
3679 * cyclic_resume() takes no arguments. Each CPU with an active cyclic
3680 * reenables and reprograms its backend (offline CPUs are not reenabled).
3681 * On most platforms, timestamps from gethrtime() and gethrestime() are not
3682 * guaranteed to monotonically increase between cyclic_suspend() and
3683 * cyclic_resume(). However, timestamps are guaranteed to monotonically
3684 * increase across the entire cyclic_suspend()/cyclic_resume() operation.
3685 * That is, every timestamp obtained before cyclic_suspend() will be less
3686 * than every timestamp obtained after cyclic_resume().
3687 *
3688 * Return value
3689 *
3690 * None; cyclic_resume() always succeeds.
3691 *
3692 * Caller's context
3693 *
3694 * The cyclic subsystem must be configured on every valid CPU;
3695 * cyclic_resume() may not be called during boot or during dynamic
3696 * reconfiguration. Additionally, cpu_lock must be held, and the caller
3697 * cannot be in high-level interrupt context. However, unlike most other
3698 * cyclic entry points, cyclic_resume() may be called with locks held which
3699 * are also acquired by CY_LOCK_LEVEL or CY_LOW_LEVEL cyclic handlers.
3700 */
3701 void
cyclic_resume()3702 cyclic_resume()
3703 {
3704 cpu_t *c;
3705 cyc_cpu_t *cpu;
3706 cyc_xcallarg_t arg;
3707 cyc_backend_t *be;
3708
3709 CYC_PTRACE0("resume");
3710 ASSERT(MUTEX_HELD(&cpu_lock));
3711
3712 c = cpu_list;
3713
3714 do {
3715 cpu = c->cpu_cyclic;
3716 be = cpu->cyp_backend;
3717 arg.cyx_cpu = cpu;
3718
3719 be->cyb_xcall(be->cyb_arg, c,
3720 (cyc_func_t)cyclic_resume_xcall, &arg);
3721 } while ((c = c->cpu_next) != cpu_list);
3722 }
3723