xref: /titanic_41/usr/src/uts/common/inet/ip/ip_squeue.c (revision d3d50737e566cade9a08d73d2af95105ac7cd960)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * IP interface to squeues.
28  *
29  * IP uses squeues to force serialization of packets, both incoming and
30  * outgoing. Each squeue is associated with a connection instance (conn_t)
31  * above, and a soft ring (if enabled) below. Each CPU will have a default
32  * squeue for outbound connections, and each soft ring of an interface will
33  * have an squeue to which it sends incoming packets. squeues are never
34  * destroyed, and if they become unused they are kept around against future
35  * needs.
36  *
37  * IP organizes its squeues using squeue sets (squeue_set_t). For each CPU
38  * in the system there will be one squeue set, all of whose squeues will be
39  * bound to that CPU, plus one additional set known as the unbound set. Sets
40  * associated with CPUs will have one default squeue, for outbound
41  * connections, and a linked list of squeues used by various NICs for inbound
42  * packets. The unbound set also has a linked list of squeues, but no default
43  * squeue.
44  *
45  * When a CPU goes offline its squeue set is destroyed, and all its squeues
46  * are moved to the unbound set. When a CPU comes online, a new squeue set is
47  * created and the default set is searched for a default squeue formerly bound
48  * to this CPU. If no default squeue is found, a new one is created.
49  *
50  * Two fields of the squeue_t, namely sq_next and sq_set, are owned by IP
51  * and not the squeue code. squeue.c will not touch them, and we can modify
52  * them without holding the squeue lock because of the guarantee that squeues
53  * are never destroyed. ip_squeue locks must be held, however.
54  *
55  * All the squeue sets are protected by a single lock, the sqset_lock. This
56  * is also used to protect the sq_next and sq_set fields of an squeue_t.
57  *
58  * The lock order is: cpu_lock --> ill_lock --> sqset_lock --> sq_lock
59  *
60  * There are two modes of associating connection with squeues. The first mode
61  * associates each connection with the CPU that creates the connection (either
62  * during open time or during accept time). The second mode associates each
63  * connection with a random CPU, effectively distributing load over all CPUs
64  * and all squeues in the system. The mode is controlled by the
65  * ip_squeue_fanout variable.
66  *
67  * NOTE: The fact that there is an association between each connection and
68  * squeue and squeue and CPU does not mean that each connection is always
69  * processed on this CPU and on this CPU only. Any thread calling squeue_enter()
70  * may process the connection on whatever CPU it is scheduled. The squeue to CPU
71  * binding is only relevant for the worker thread.
72  *
73  * INTERFACE:
74  *
75  * squeue_t *ip_squeue_get(ill_rx_ring_t)
76  *
77  * Returns the squeue associated with an ill receive ring. If the ring is
78  * not bound to a CPU, and we're currently servicing the interrupt which
79  * generated the packet, then bind the squeue to CPU.
80  *
81  *
82  * DR Notes
83  * ========
84  *
85  * The ip_squeue_init() registers a call-back function with the CPU DR
86  * subsystem using register_cpu_setup_func(). The call-back function does two
87  * things:
88  *
89  * o When the CPU is going off-line or unconfigured, the worker thread is
90  *	unbound from the CPU. This allows the CPU unconfig code to move it to
91  *	another CPU.
92  *
93  * o When the CPU is going online, it creates a new squeue for this CPU if
94  *	necessary and binds the squeue worker thread to this CPU.
95  *
96  * TUNABLES:
97  *
98  * ip_squeue_fanout: used when TCP calls IP_SQUEUE_GET(). If 1, then
99  * pick the default squeue from a random CPU, otherwise use our CPU's default
100  * squeue.
101  *
102  * ip_squeue_fanout can be accessed and changed using ndd on /dev/tcp or
103  * /dev/ip.
104  *
105  * ip_squeue_worker_wait: global value for the sq_wait field for all squeues *
106  * created. This is the time squeue code waits before waking up the worker
107  * thread after queuing a request.
108  */
109 
110 #include <sys/types.h>
111 #include <sys/debug.h>
112 #include <sys/kmem.h>
113 #include <sys/cpuvar.h>
114 #include <sys/cmn_err.h>
115 
116 #include <inet/common.h>
117 #include <inet/ip.h>
118 #include <netinet/ip6.h>
119 #include <inet/ip_if.h>
120 #include <inet/ip_ire.h>
121 #include <inet/nd.h>
122 #include <inet/ipclassifier.h>
123 #include <sys/types.h>
124 #include <sys/conf.h>
125 #include <sys/sunddi.h>
126 #include <sys/dlpi.h>
127 #include <sys/squeue_impl.h>
128 #include <sys/tihdr.h>
129 #include <inet/udp_impl.h>
130 #include <sys/strsubr.h>
131 #include <sys/zone.h>
132 #include <sys/dld.h>
133 #include <sys/atomic.h>
134 
135 /*
136  * List of all created squeue sets. The list and its size are protected by
137  * sqset_lock.
138  */
139 static squeue_set_t	**sqset_global_list; /* list 0 is the unbound list */
140 static uint_t		sqset_global_size;
141 kmutex_t		sqset_lock;
142 
143 static void (*ip_squeue_create_callback)(squeue_t *) = NULL;
144 
145 /*
146  * ip_squeue_worker_wait: global value for the sq_wait field for all squeues
147  *	created. This is the time squeue code waits before waking up the worker
148  *	thread after queuing a request.
149  */
150 uint_t ip_squeue_worker_wait = 10;
151 
152 static squeue_t *ip_squeue_create(pri_t);
153 static squeue_set_t *ip_squeue_set_create(processorid_t);
154 static int ip_squeue_cpu_setup(cpu_setup_t, int, void *);
155 static void ip_squeue_set_move(squeue_t *, squeue_set_t *);
156 static void ip_squeue_set_destroy(cpu_t *);
157 static void ip_squeue_clean(void *, mblk_t *, void *);
158 
159 #define	CPU_ISON(c) (c != NULL && CPU_ACTIVE(c) && (c->cpu_flags & CPU_EXISTS))
160 
161 static squeue_t *
ip_squeue_create(pri_t pri)162 ip_squeue_create(pri_t pri)
163 {
164 	squeue_t *sqp;
165 
166 	sqp = squeue_create(ip_squeue_worker_wait, pri);
167 	ASSERT(sqp != NULL);
168 	if (ip_squeue_create_callback != NULL)
169 		ip_squeue_create_callback(sqp);
170 	return (sqp);
171 }
172 
173 /*
174  * Create a new squeue_set. If id == -1, then we're creating the unbound set,
175  * which should only happen once when we are first initialized. Otherwise id
176  * is the id of the CPU that needs a set, either because we are initializing
177  * or because the CPU has come online.
178  *
179  * If id != -1, then we need at a minimum to provide a default squeue for the
180  * new set. We search the unbound set for candidates, and if none are found we
181  * create a new one.
182  */
183 static squeue_set_t *
ip_squeue_set_create(processorid_t id)184 ip_squeue_set_create(processorid_t id)
185 {
186 	squeue_set_t	*sqs;
187 	squeue_set_t	*src = sqset_global_list[0];
188 	squeue_t	**lastsqp, *sq;
189 	squeue_t	**defaultq_lastp = NULL;
190 
191 	sqs = kmem_zalloc(sizeof (squeue_set_t), KM_SLEEP);
192 	sqs->sqs_cpuid = id;
193 
194 	if (id == -1) {
195 		ASSERT(sqset_global_size == 0);
196 		sqset_global_list[0] = sqs;
197 		sqset_global_size = 1;
198 		return (sqs);
199 	}
200 
201 	/*
202 	 * When we create an squeue set id != -1, we need to give it a
203 	 * default squeue, in order to support fanout of conns across
204 	 * CPUs. Try to find a former default squeue that matches this
205 	 * cpu id on the unbound squeue set. If no such squeue is found,
206 	 * find some non-default TCP squeue that is free. If still no such
207 	 * candidate is found, create a new squeue.
208 	 */
209 
210 	ASSERT(MUTEX_HELD(&cpu_lock));
211 	mutex_enter(&sqset_lock);
212 	lastsqp = &src->sqs_head;
213 
214 	while (*lastsqp) {
215 		if ((*lastsqp)->sq_bind == id &&
216 		    (*lastsqp)->sq_state & SQS_DEFAULT) {
217 			/*
218 			 * Exact match. Former default squeue of cpu 'id'
219 			 */
220 			ASSERT(!((*lastsqp)->sq_state & SQS_ILL_BOUND));
221 			defaultq_lastp = lastsqp;
222 			break;
223 		}
224 		if (defaultq_lastp == NULL &&
225 		    !((*lastsqp)->sq_state & (SQS_ILL_BOUND | SQS_DEFAULT))) {
226 			/*
227 			 * A free non-default TCP squeue
228 			 */
229 			defaultq_lastp = lastsqp;
230 		}
231 		lastsqp = &(*lastsqp)->sq_next;
232 	}
233 
234 	if (defaultq_lastp != NULL) {
235 		/* Remove from src set and set SQS_DEFAULT */
236 		sq = *defaultq_lastp;
237 		*defaultq_lastp = sq->sq_next;
238 		sq->sq_next = NULL;
239 		if (!(sq->sq_state & SQS_DEFAULT)) {
240 			mutex_enter(&sq->sq_lock);
241 			sq->sq_state |= SQS_DEFAULT;
242 			mutex_exit(&sq->sq_lock);
243 		}
244 	} else {
245 		sq = ip_squeue_create(SQUEUE_DEFAULT_PRIORITY);
246 		sq->sq_state |= SQS_DEFAULT;
247 	}
248 
249 	sq->sq_set = sqs;
250 	sqs->sqs_default = sq;
251 	squeue_bind(sq, id); /* this locks squeue mutex */
252 
253 	ASSERT(sqset_global_size <= NCPU);
254 	sqset_global_list[sqset_global_size++] = sqs;
255 	mutex_exit(&sqset_lock);
256 	return (sqs);
257 }
258 
259 /*
260  * Called by ill_ring_add() to find an squeue to associate with a new ring.
261  */
262 
263 squeue_t *
ip_squeue_getfree(pri_t pri)264 ip_squeue_getfree(pri_t pri)
265 {
266 	squeue_set_t	*sqs = sqset_global_list[0];
267 	squeue_t	*sq;
268 
269 	mutex_enter(&sqset_lock);
270 	for (sq = sqs->sqs_head; sq != NULL; sq = sq->sq_next) {
271 		/*
272 		 * Select a non-default TCP squeue that is free i.e. not
273 		 * bound to any ill.
274 		 */
275 		if (!(sq->sq_state & (SQS_DEFAULT | SQS_ILL_BOUND)))
276 			break;
277 	}
278 
279 	if (sq == NULL) {
280 		sq = ip_squeue_create(pri);
281 		sq->sq_set = sqs;
282 		sq->sq_next = sqs->sqs_head;
283 		sqs->sqs_head = sq;
284 	}
285 
286 	ASSERT(!(sq->sq_state & (SQS_POLL_THR_CONTROL | SQS_WORKER_THR_CONTROL |
287 	    SQS_POLL_CLEANUP_DONE | SQS_POLL_QUIESCE_DONE |
288 	    SQS_POLL_THR_QUIESCED)));
289 
290 	mutex_enter(&sq->sq_lock);
291 	sq->sq_state |= SQS_ILL_BOUND;
292 	mutex_exit(&sq->sq_lock);
293 	mutex_exit(&sqset_lock);
294 
295 	if (sq->sq_priority != pri) {
296 		thread_lock(sq->sq_worker);
297 		(void) thread_change_pri(sq->sq_worker, pri, 0);
298 		thread_unlock(sq->sq_worker);
299 
300 		thread_lock(sq->sq_poll_thr);
301 		(void) thread_change_pri(sq->sq_poll_thr, pri, 0);
302 		thread_unlock(sq->sq_poll_thr);
303 
304 		sq->sq_priority = pri;
305 	}
306 	return (sq);
307 }
308 
309 /*
310  * Initialize IP squeues.
311  */
312 void
ip_squeue_init(void (* callback)(squeue_t *))313 ip_squeue_init(void (*callback)(squeue_t *))
314 {
315 	int i;
316 	squeue_set_t	*sqs;
317 
318 	ASSERT(sqset_global_list == NULL);
319 
320 	ip_squeue_create_callback = callback;
321 	squeue_init();
322 	mutex_init(&sqset_lock, NULL, MUTEX_DEFAULT, NULL);
323 	sqset_global_list =
324 	    kmem_zalloc(sizeof (squeue_set_t *) * (NCPU+1), KM_SLEEP);
325 	sqset_global_size = 0;
326 	/*
327 	 * We are called at system boot time and we don't
328 	 * expect memory allocation failure.
329 	 */
330 	sqs = ip_squeue_set_create(-1);
331 	ASSERT(sqs != NULL);
332 
333 	mutex_enter(&cpu_lock);
334 	/* Create squeue for each active CPU available */
335 	for (i = 0; i < NCPU; i++) {
336 		cpu_t *cp = cpu_get(i);
337 		if (CPU_ISON(cp) && cp->cpu_squeue_set == NULL) {
338 			/*
339 			 * We are called at system boot time and we don't
340 			 * expect memory allocation failure then
341 			 */
342 			cp->cpu_squeue_set = ip_squeue_set_create(cp->cpu_id);
343 			ASSERT(cp->cpu_squeue_set != NULL);
344 		}
345 	}
346 
347 	register_cpu_setup_func(ip_squeue_cpu_setup, NULL);
348 	mutex_exit(&cpu_lock);
349 }
350 
351 /*
352  * Get a default squeue, either from the current CPU or a CPU derived by hash
353  * from the index argument, depending upon the setting of ip_squeue_fanout.
354  */
355 squeue_t *
ip_squeue_random(uint_t index)356 ip_squeue_random(uint_t index)
357 {
358 	squeue_set_t *sqs = NULL;
359 	squeue_t *sq;
360 
361 	/*
362 	 * The minimum value of sqset_global_size is 2, one for the unbound
363 	 * squeue set and another for the squeue set of the zeroth CPU.
364 	 * Even though the value could be changing, it can never go below 2,
365 	 * so the assert does not need the lock protection.
366 	 */
367 	ASSERT(sqset_global_size > 1);
368 
369 	/* Protect against changes to sqset_global_list */
370 	mutex_enter(&sqset_lock);
371 
372 	if (!ip_squeue_fanout)
373 		sqs = CPU->cpu_squeue_set;
374 
375 	/*
376 	 * sqset_global_list[0] corresponds to the unbound squeue set.
377 	 * The computation below picks a set other than the unbound set.
378 	 */
379 	if (sqs == NULL)
380 		sqs = sqset_global_list[(index % (sqset_global_size - 1)) + 1];
381 	sq = sqs->sqs_default;
382 
383 	mutex_exit(&sqset_lock);
384 	ASSERT(sq);
385 	return (sq);
386 }
387 
388 /*
389  * Move squeue from its current set to newset. Not used for default squeues.
390  * Bind or unbind the worker thread as appropriate.
391  */
392 
393 static void
ip_squeue_set_move(squeue_t * sq,squeue_set_t * newset)394 ip_squeue_set_move(squeue_t *sq, squeue_set_t *newset)
395 {
396 	squeue_set_t	*set;
397 	squeue_t	**lastsqp;
398 	processorid_t	cpuid = newset->sqs_cpuid;
399 
400 	ASSERT(!(sq->sq_state & SQS_DEFAULT));
401 	ASSERT(!MUTEX_HELD(&sq->sq_lock));
402 	ASSERT(MUTEX_HELD(&sqset_lock));
403 
404 	set = sq->sq_set;
405 	if (set == newset)
406 		return;
407 
408 	lastsqp = &set->sqs_head;
409 	while (*lastsqp != sq)
410 		lastsqp = &(*lastsqp)->sq_next;
411 
412 	*lastsqp = sq->sq_next;
413 	sq->sq_next = newset->sqs_head;
414 	newset->sqs_head = sq;
415 	sq->sq_set = newset;
416 	if (cpuid == -1)
417 		squeue_unbind(sq);
418 	else
419 		squeue_bind(sq, cpuid);
420 }
421 
422 /*
423  * Move squeue from its current set to cpuid's set and bind to cpuid.
424  */
425 
426 int
ip_squeue_cpu_move(squeue_t * sq,processorid_t cpuid)427 ip_squeue_cpu_move(squeue_t *sq, processorid_t cpuid)
428 {
429 	cpu_t *cpu;
430 	squeue_set_t *set;
431 
432 	if (sq->sq_state & SQS_DEFAULT)
433 		return (-1);
434 
435 	ASSERT(MUTEX_HELD(&cpu_lock));
436 
437 	cpu = cpu_get(cpuid);
438 	if (!CPU_ISON(cpu))
439 		return (-1);
440 
441 	mutex_enter(&sqset_lock);
442 	set = cpu->cpu_squeue_set;
443 	if (set != NULL)
444 		ip_squeue_set_move(sq, set);
445 	mutex_exit(&sqset_lock);
446 	return ((set == NULL) ? -1 : 0);
447 }
448 
449 /*
450  * The mac layer is calling, asking us to move an squeue to a
451  * new CPU. This routine is called with cpu_lock held.
452  */
453 void
ip_squeue_bind_ring(ill_t * ill,ill_rx_ring_t * rx_ring,processorid_t cpuid)454 ip_squeue_bind_ring(ill_t *ill, ill_rx_ring_t *rx_ring, processorid_t cpuid)
455 {
456 	ASSERT(ILL_MAC_PERIM_HELD(ill));
457 	ASSERT(rx_ring->rr_ill == ill);
458 
459 	mutex_enter(&ill->ill_lock);
460 	if (rx_ring->rr_ring_state == RR_FREE ||
461 	    rx_ring->rr_ring_state == RR_FREE_INPROG) {
462 		mutex_exit(&ill->ill_lock);
463 		return;
464 	}
465 
466 	if (ip_squeue_cpu_move(rx_ring->rr_sqp, cpuid) != -1)
467 		rx_ring->rr_ring_state = RR_SQUEUE_BOUND;
468 
469 	mutex_exit(&ill->ill_lock);
470 }
471 
472 void *
ip_squeue_add_ring(ill_t * ill,void * mrp)473 ip_squeue_add_ring(ill_t *ill, void *mrp)
474 {
475 	mac_rx_fifo_t		*mrfp = (mac_rx_fifo_t *)mrp;
476 	ill_rx_ring_t		*rx_ring, *ring_tbl;
477 	int			ip_rx_index;
478 	squeue_t		*sq = NULL;
479 	pri_t			pri;
480 
481 	ASSERT(ILL_MAC_PERIM_HELD(ill));
482 	ASSERT(mrfp->mrf_type == MAC_RX_FIFO);
483 	ASSERT(ill->ill_dld_capab != NULL);
484 
485 	ring_tbl = ill->ill_dld_capab->idc_poll.idp_ring_tbl;
486 
487 	mutex_enter(&ill->ill_lock);
488 	for (ip_rx_index = 0; ip_rx_index < ILL_MAX_RINGS; ip_rx_index++) {
489 		rx_ring = &ring_tbl[ip_rx_index];
490 		if (rx_ring->rr_ring_state == RR_FREE)
491 			break;
492 	}
493 
494 	if (ip_rx_index == ILL_MAX_RINGS) {
495 		/*
496 		 * We ran out of ILL_MAX_RINGS worth rx_ring structures. If
497 		 * we have devices which can overwhelm this limit,
498 		 * ILL_MAX_RING should be made configurable. Meanwhile it
499 		 * cause no panic because driver will pass ip_input a NULL
500 		 * handle which will make IP allocate the default squeue and
501 		 * Polling mode will not be used for this ring.
502 		 */
503 		cmn_err(CE_NOTE,
504 		    "Reached maximum number of receiving rings (%d) for %s\n",
505 		    ILL_MAX_RINGS, ill->ill_name);
506 		mutex_exit(&ill->ill_lock);
507 		return (NULL);
508 	}
509 
510 	bzero(rx_ring, sizeof (ill_rx_ring_t));
511 	rx_ring->rr_rx = (ip_mac_rx_t)mrfp->mrf_receive;
512 	/* XXX: Hard code it to tcp accept for now */
513 	rx_ring->rr_ip_accept = (ip_accept_t)ip_accept_tcp;
514 
515 	rx_ring->rr_intr_handle = mrfp->mrf_intr_handle;
516 	rx_ring->rr_intr_enable = (ip_mac_intr_enable_t)mrfp->mrf_intr_enable;
517 	rx_ring->rr_intr_disable =
518 	    (ip_mac_intr_disable_t)mrfp->mrf_intr_disable;
519 	rx_ring->rr_rx_handle = mrfp->mrf_rx_arg;
520 	rx_ring->rr_ill = ill;
521 
522 	pri = mrfp->mrf_flow_priority;
523 
524 	sq = ip_squeue_getfree(pri);
525 
526 	mutex_enter(&sq->sq_lock);
527 	sq->sq_rx_ring = rx_ring;
528 	rx_ring->rr_sqp = sq;
529 
530 	sq->sq_state |= SQS_POLL_CAPAB;
531 
532 	rx_ring->rr_ring_state = RR_SQUEUE_UNBOUND;
533 	sq->sq_ill = ill;
534 	mutex_exit(&sq->sq_lock);
535 	mutex_exit(&ill->ill_lock);
536 
537 	DTRACE_PROBE4(ill__ring__add, char *, ill->ill_name, ill_t *, ill, int,
538 	    ip_rx_index, void *, mrfp->mrf_rx_arg);
539 
540 	/* Assign the squeue to the specified CPU as well */
541 	mutex_enter(&cpu_lock);
542 	(void) ip_squeue_bind_ring(ill, rx_ring, mrfp->mrf_cpu_id);
543 	mutex_exit(&cpu_lock);
544 
545 	return (rx_ring);
546 }
547 
548 /*
549  * sanitize the squeue etc. Some of the processing
550  * needs to be done from inside the perimeter.
551  */
552 void
ip_squeue_clean_ring(ill_t * ill,ill_rx_ring_t * rx_ring)553 ip_squeue_clean_ring(ill_t *ill, ill_rx_ring_t *rx_ring)
554 {
555 	squeue_t *sqp;
556 
557 	ASSERT(ILL_MAC_PERIM_HELD(ill));
558 	ASSERT(rx_ring != NULL);
559 
560 	/* Just clean one squeue */
561 	mutex_enter(&ill->ill_lock);
562 	if (rx_ring->rr_ring_state == RR_FREE) {
563 		mutex_exit(&ill->ill_lock);
564 		return;
565 	}
566 	rx_ring->rr_ring_state = RR_FREE_INPROG;
567 	sqp = rx_ring->rr_sqp;
568 
569 	mutex_enter(&sqp->sq_lock);
570 	sqp->sq_state |= SQS_POLL_CLEANUP;
571 	cv_signal(&sqp->sq_worker_cv);
572 	mutex_exit(&ill->ill_lock);
573 	while (!(sqp->sq_state & SQS_POLL_CLEANUP_DONE))
574 		cv_wait(&sqp->sq_ctrlop_done_cv, &sqp->sq_lock);
575 	sqp->sq_state &= ~SQS_POLL_CLEANUP_DONE;
576 
577 	ASSERT(!(sqp->sq_state & (SQS_POLL_THR_CONTROL |
578 	    SQS_WORKER_THR_CONTROL | SQS_POLL_QUIESCE_DONE |
579 	    SQS_POLL_THR_QUIESCED)));
580 
581 	cv_signal(&sqp->sq_worker_cv);
582 	mutex_exit(&sqp->sq_lock);
583 
584 	/*
585 	 * Move the squeue to sqset_global_list[0] which holds the set of
586 	 * squeues not bound to any cpu. Note that the squeue is still
587 	 * considered bound to an ill as long as SQS_ILL_BOUND is set.
588 	 */
589 	mutex_enter(&sqset_lock);
590 	ip_squeue_set_move(sqp, sqset_global_list[0]);
591 	mutex_exit(&sqset_lock);
592 
593 	/*
594 	 * CPU going offline can also trigger a move of the squeue to the
595 	 * unbound set sqset_global_list[0]. However the squeue won't be
596 	 * recycled for the next use as long as the SQS_ILL_BOUND flag
597 	 * is set. Hence we clear the SQS_ILL_BOUND flag only towards the
598 	 * end after the move.
599 	 */
600 	mutex_enter(&sqp->sq_lock);
601 	sqp->sq_state &= ~SQS_ILL_BOUND;
602 	mutex_exit(&sqp->sq_lock);
603 
604 	mutex_enter(&ill->ill_lock);
605 	rx_ring->rr_ring_state = RR_FREE;
606 	mutex_exit(&ill->ill_lock);
607 }
608 
609 /*
610  * Stop the squeue from polling. This needs to be done
611  * from inside the perimeter.
612  */
613 void
ip_squeue_quiesce_ring(ill_t * ill,ill_rx_ring_t * rx_ring)614 ip_squeue_quiesce_ring(ill_t *ill, ill_rx_ring_t *rx_ring)
615 {
616 	squeue_t *sqp;
617 
618 	ASSERT(ILL_MAC_PERIM_HELD(ill));
619 	ASSERT(rx_ring != NULL);
620 
621 	sqp = rx_ring->rr_sqp;
622 	mutex_enter(&sqp->sq_lock);
623 	sqp->sq_state |= SQS_POLL_QUIESCE;
624 	cv_signal(&sqp->sq_worker_cv);
625 	while (!(sqp->sq_state & SQS_POLL_QUIESCE_DONE))
626 		cv_wait(&sqp->sq_ctrlop_done_cv, &sqp->sq_lock);
627 
628 	mutex_exit(&sqp->sq_lock);
629 }
630 
631 /*
632  * Restart polling etc. Needs to be inside the perimeter to
633  * prevent races.
634  */
635 void
ip_squeue_restart_ring(ill_t * ill,ill_rx_ring_t * rx_ring)636 ip_squeue_restart_ring(ill_t *ill, ill_rx_ring_t *rx_ring)
637 {
638 	squeue_t *sqp;
639 
640 	ASSERT(ILL_MAC_PERIM_HELD(ill));
641 	ASSERT(rx_ring != NULL);
642 
643 	sqp = rx_ring->rr_sqp;
644 	mutex_enter(&sqp->sq_lock);
645 	/*
646 	 * Handle change in number of rings between the quiesce and
647 	 * restart operations by checking for a previous quiesce before
648 	 * attempting a restart.
649 	 */
650 	if (!(sqp->sq_state & SQS_POLL_QUIESCE_DONE)) {
651 		mutex_exit(&sqp->sq_lock);
652 		return;
653 	}
654 	sqp->sq_state |= SQS_POLL_RESTART;
655 	cv_signal(&sqp->sq_worker_cv);
656 	while (!(sqp->sq_state & SQS_POLL_RESTART_DONE))
657 		cv_wait(&sqp->sq_ctrlop_done_cv, &sqp->sq_lock);
658 	sqp->sq_state &= ~SQS_POLL_RESTART_DONE;
659 	mutex_exit(&sqp->sq_lock);
660 }
661 
662 /*
663  * sanitize all squeues associated with the ill.
664  */
665 void
ip_squeue_clean_all(ill_t * ill)666 ip_squeue_clean_all(ill_t *ill)
667 {
668 	int idx;
669 	ill_rx_ring_t	*rx_ring;
670 
671 	for (idx = 0; idx < ILL_MAX_RINGS; idx++) {
672 		rx_ring = &ill->ill_dld_capab->idc_poll.idp_ring_tbl[idx];
673 		ip_squeue_clean_ring(ill, rx_ring);
674 	}
675 }
676 
677 /*
678  * Used by IP to get the squeue associated with a ring. If the squeue isn't
679  * yet bound to a CPU, and we're being called directly from the NIC's
680  * interrupt, then we know what CPU we want to assign the squeue to, so
681  * dispatch that task to a taskq.
682  */
683 squeue_t *
ip_squeue_get(ill_rx_ring_t * ill_rx_ring)684 ip_squeue_get(ill_rx_ring_t *ill_rx_ring)
685 {
686 	squeue_t 	*sqp;
687 
688 	if ((ill_rx_ring == NULL) || ((sqp = ill_rx_ring->rr_sqp) == NULL))
689 		return (IP_SQUEUE_GET(CPU_PSEUDO_RANDOM()));
690 
691 	return (sqp);
692 }
693 
694 /*
695  * Called when a CPU goes offline. It's squeue_set_t is destroyed, and all
696  * squeues are unboudn and moved to the unbound set.
697  */
698 static void
ip_squeue_set_destroy(cpu_t * cpu)699 ip_squeue_set_destroy(cpu_t *cpu)
700 {
701 	int i;
702 	squeue_t *sqp, *lastsqp = NULL;
703 	squeue_set_t *sqs, *unbound = sqset_global_list[0];
704 
705 	mutex_enter(&sqset_lock);
706 	if ((sqs = cpu->cpu_squeue_set) == NULL) {
707 		mutex_exit(&sqset_lock);
708 		return;
709 	}
710 
711 	/* Move all squeues to unbound set */
712 
713 	for (sqp = sqs->sqs_head; sqp; lastsqp = sqp, sqp = sqp->sq_next) {
714 		squeue_unbind(sqp);
715 		sqp->sq_set = unbound;
716 	}
717 	if (sqs->sqs_head) {
718 		lastsqp->sq_next = unbound->sqs_head;
719 		unbound->sqs_head = sqs->sqs_head;
720 	}
721 
722 	/* Also move default squeue to unbound set */
723 
724 	sqp = sqs->sqs_default;
725 	ASSERT(sqp != NULL);
726 	ASSERT((sqp->sq_state & (SQS_DEFAULT|SQS_ILL_BOUND)) == SQS_DEFAULT);
727 
728 	sqp->sq_next = unbound->sqs_head;
729 	unbound->sqs_head = sqp;
730 	squeue_unbind(sqp);
731 	sqp->sq_set = unbound;
732 
733 	for (i = 1; i < sqset_global_size; i++)
734 		if (sqset_global_list[i] == sqs)
735 			break;
736 
737 	ASSERT(i < sqset_global_size);
738 	sqset_global_list[i] = sqset_global_list[sqset_global_size - 1];
739 	sqset_global_list[sqset_global_size - 1] = NULL;
740 	sqset_global_size--;
741 
742 	mutex_exit(&sqset_lock);
743 	kmem_free(sqs, sizeof (*sqs));
744 }
745 
746 /*
747  * Reconfiguration callback
748  */
749 /* ARGSUSED */
750 static int
ip_squeue_cpu_setup(cpu_setup_t what,int id,void * arg)751 ip_squeue_cpu_setup(cpu_setup_t what, int id, void *arg)
752 {
753 	cpu_t *cp = cpu_get(id);
754 
755 	ASSERT(MUTEX_HELD(&cpu_lock));
756 	switch (what) {
757 	case CPU_CONFIG:
758 	case CPU_ON:
759 	case CPU_INIT:
760 	case CPU_CPUPART_IN:
761 		if (CPU_ISON(cp) && cp->cpu_squeue_set == NULL)
762 			cp->cpu_squeue_set = ip_squeue_set_create(cp->cpu_id);
763 		break;
764 	case CPU_UNCONFIG:
765 	case CPU_OFF:
766 	case CPU_CPUPART_OUT:
767 		if (cp->cpu_squeue_set != NULL) {
768 			ip_squeue_set_destroy(cp);
769 			cp->cpu_squeue_set = NULL;
770 		}
771 		break;
772 	default:
773 		break;
774 	}
775 	return (0);
776 }
777