xref: /illumos-gate/usr/src/uts/common/io/mac/mac_datapath_setup.c (revision 635216b673cf196ac523ff2a7ab715717e553292)
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 #include <sys/types.h>
27 #include <sys/callb.h>
28 #include <sys/sdt.h>
29 #include <sys/strsubr.h>
30 #include <sys/strsun.h>
31 #include <sys/vlan.h>
32 #include <inet/ipsec_impl.h>
33 #include <inet/ip_impl.h>
34 #include <inet/sadb.h>
35 #include <inet/ipsecesp.h>
36 #include <inet/ipsecah.h>
37 
38 #include <sys/mac_impl.h>
39 #include <sys/mac_client_impl.h>
40 #include <sys/mac_client_priv.h>
41 #include <sys/mac_soft_ring.h>
42 #include <sys/mac_flow_impl.h>
43 
44 static void mac_srs_soft_rings_signal(mac_soft_ring_set_t *, uint_t);
45 static void mac_srs_update_fanout_list(mac_soft_ring_set_t *);
46 static void mac_srs_poll_unbind(mac_soft_ring_set_t *);
47 static void mac_srs_worker_unbind(mac_soft_ring_set_t *);
48 static void mac_srs_soft_rings_quiesce(mac_soft_ring_set_t *, uint_t);
49 
50 static int mac_srs_cpu_setup(cpu_setup_t, int, void *);
51 static void mac_srs_worker_bind(mac_soft_ring_set_t *, processorid_t);
52 static void mac_srs_poll_bind(mac_soft_ring_set_t *, processorid_t);
53 static void mac_srs_threads_unbind(mac_soft_ring_set_t *);
54 static void mac_srs_add_glist(mac_soft_ring_set_t *);
55 static void mac_srs_remove_glist(mac_soft_ring_set_t *);
56 static void mac_srs_fanout_list_free(mac_soft_ring_set_t *);
57 static void mac_soft_ring_remove(mac_soft_ring_set_t *, mac_soft_ring_t *);
58 
59 static int mac_compute_soft_ring_count(flow_entry_t *, int);
60 static void mac_walk_srs_and_bind(int);
61 static void mac_walk_srs_and_unbind(int);
62 
63 extern mac_group_t *mac_reserve_rx_group(mac_client_impl_t *, uint8_t *,
64     mac_rx_group_reserve_type_t);
65 extern void mac_release_rx_group(mac_client_impl_t *, mac_group_t *);
66 
67 extern boolean_t mac_latency_optimize;
68 
69 static kmem_cache_t *mac_srs_cache;
70 kmem_cache_t *mac_soft_ring_cache;
71 
72 /*
73  * The duration in msec we wait before signalling the soft ring
74  * worker thread in case packets get queued.
75  */
76 uint32_t mac_soft_ring_worker_wait = 0;
77 
78 /*
79  * A global tunable for turning polling on/off. By default, dynamic
80  * polling is always on and is always very beneficial. It should be
81  * turned off with absolute care and for the rare workload (very
82  * low latency sensitive traffic).
83  */
84 int mac_poll_enable = B_TRUE;
85 
86 /*
87  * Need to set mac_soft_ring_max_q_cnt based on bandwidth and perhaps latency.
88  * Large values could end up in consuming lot of system memory and cause
89  * system hang.
90  */
91 int mac_soft_ring_max_q_cnt = 1024;
92 int mac_soft_ring_min_q_cnt = 256;
93 int mac_soft_ring_poll_thres = 16;
94 
95 /*
96  * Default value of number of TX rings to be assigned to a MAC client.
97  * If less than 'mac_tx_ring_count' worth of Tx rings is available, then
98  * as many as is available will be assigned to the newly created MAC client.
99  * If no TX rings are available, then MAC client(s) will be assigned the
100  * default Tx ring. Default Tx ring can be shared among multiple MAC clients.
101  */
102 uint32_t mac_tx_ring_count = 32;
103 boolean_t mac_tx_serialize = B_FALSE;
104 
105 /*
106  * mac_tx_srs_hiwat is the queue depth threshold at which callers of
107  * mac_tx() will be notified of flow control condition.
108  *
109  * TCP does not honour flow control condition sent up by mac_tx().
110  * Thus provision is made for TCP to allow more packets to be queued
111  * in SRS upto a maximum of mac_tx_srs_max_q_cnt.
112  *
113  * Note that mac_tx_srs_hiwat is always be lesser than
114  * mac_tx_srs_max_q_cnt.
115  */
116 uint32_t mac_tx_srs_max_q_cnt = 100000;
117 uint32_t mac_tx_srs_hiwat = 1000;
118 
119 /*
120  * mac_rx_soft_ring_count, mac_soft_ring_10gig_count:
121  *
122  * Global tunables that determines the number of soft rings to be used for
123  * fanning out incoming traffic on a link. These count will be used only
124  * when no explicit set of CPUs was assigned to the data-links.
125  *
126  * mac_rx_soft_ring_count tunable will come into effect only if
127  * mac_soft_ring_enable is set. mac_soft_ring_enable is turned on by
128  * default only for sun4v platforms.
129  *
130  * mac_rx_soft_ring_10gig_count will come into effect if you are running on a
131  * 10Gbps link and is not dependent upon mac_soft_ring_enable.
132  *
133  * The number of soft rings for fanout for a link or a flow is determined
134  * by mac_compute_soft_ring_count() routine. This routine will take into
135  * account mac_soft_ring_enable, mac_rx_soft_ring_count and
136  * mac_rx_soft_ring_10gig_count to determine the soft ring count for a link.
137  *
138  * If a bandwidth is specified, the determination of the number of soft
139  * rings is based on specified bandwidth, CPU speed and number of CPUs in
140  * the system.
141  */
142 uint_t mac_rx_soft_ring_count = 8;
143 uint_t mac_rx_soft_ring_10gig_count = 8;
144 
145 /*
146  * Every Tx and Rx mac_soft_ring_set_t (mac_srs) created gets added
147  * to mac_srs_g_list and mac_srs_g_lock protects mac_srs_g_list. The
148  * list is used to walk the list of all MAC threads when a CPU is
149  * coming online or going offline.
150  */
151 static mac_soft_ring_set_t *mac_srs_g_list = NULL;
152 static krwlock_t mac_srs_g_lock;
153 
154 /*
155  * Whether the SRS threads should be bound, or not.
156  */
157 boolean_t mac_srs_thread_bind = B_TRUE;
158 
159 /*
160  * CPU to fallback to, used by mac_next_bind_cpu().
161  */
162 processorid_t srs_bind_cpu = 0;
163 
164 /*
165  * If cpu bindings are specified by user, then Tx SRS and its soft
166  * rings should also be bound to the CPUs specified by user. The
167  * CPUs for Tx bindings are at the end of the cpu list provided by
168  * the user. If enough CPUs are not available (for Tx and Rx
169  * SRSes), then the CPUs are shared by both Tx and Rx SRSes.
170  */
171 #define	BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp) {			\
172 	processorid_t cpuid;						\
173 	int i, j;							\
174 	mac_soft_ring_t *softring;					\
175 									\
176 	cpuid = mrp->mrp_cpu[mrp->mrp_ncpus - 1];			\
177 	mac_srs_worker_bind(mac_tx_srs, cpuid);			\
178 	if (TX_MULTI_RING_MODE(mac_tx_srs)) {				\
179 		j =  mrp->mrp_ncpus - 1;				\
180 		for (i = 0;						\
181 		    i < mac_tx_srs->srs_oth_ring_count; i++, j--) {	\
182 			if (j < 0)					\
183 				j =  mrp->mrp_ncpus - 1;		\
184 			cpuid = mrp->mrp_cpu[j];			\
185 			softring = mac_tx_srs->srs_oth_soft_rings[i];	\
186 			(void) mac_soft_ring_bind(softring, cpuid);	\
187 		}							\
188 	}								\
189 }
190 
191 /* INIT and FINI ROUTINES */
192 
193 void
194 mac_soft_ring_init(void)
195 {
196 	mac_soft_ring_cache = kmem_cache_create("mac_soft_ring_cache",
197 	    sizeof (mac_soft_ring_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
198 
199 	mac_srs_cache = kmem_cache_create("mac_srs_cache",
200 	    sizeof (mac_soft_ring_set_t),
201 	    64, NULL, NULL, NULL, NULL, NULL, 0);
202 
203 	rw_init(&mac_srs_g_lock, NULL, RW_DEFAULT, NULL);
204 	mutex_enter(&cpu_lock);
205 	register_cpu_setup_func(mac_srs_cpu_setup, NULL);
206 	mutex_exit(&cpu_lock);
207 }
208 
209 void
210 mac_soft_ring_finish(void)
211 {
212 	mutex_enter(&cpu_lock);
213 	unregister_cpu_setup_func(mac_srs_cpu_setup, NULL);
214 	mutex_exit(&cpu_lock);
215 	rw_destroy(&mac_srs_g_lock);
216 	kmem_cache_destroy(mac_soft_ring_cache);
217 	kmem_cache_destroy(mac_srs_cache);
218 }
219 
220 static void
221 mac_srs_soft_rings_free(mac_soft_ring_set_t *mac_srs, boolean_t release_tx_ring)
222 {
223 	mac_soft_ring_t	*softring, *next, *head;
224 
225 	/*
226 	 * Synchronize with mac_walk_srs_bind/unbind which are callbacks from
227 	 * DR. The callbacks from DR are called with cpu_lock held, and hence
228 	 * can't wait to grab the mac perimeter. The soft ring list is hence
229 	 * protected for read access by srs_lock. Changing the soft ring list
230 	 * needs the mac perimeter and the srs_lock.
231 	 */
232 	mutex_enter(&mac_srs->srs_lock);
233 
234 	head = mac_srs->srs_soft_ring_head;
235 	mac_srs->srs_soft_ring_head = NULL;
236 	mac_srs->srs_soft_ring_tail = NULL;
237 	mac_srs->srs_soft_ring_count = 0;
238 
239 	mutex_exit(&mac_srs->srs_lock);
240 
241 	for (softring = head; softring != NULL; softring = next) {
242 		next = softring->s_ring_next;
243 		mac_soft_ring_free(softring, release_tx_ring);
244 	}
245 }
246 
247 static void
248 mac_srs_add_glist(mac_soft_ring_set_t *mac_srs)
249 {
250 	ASSERT(mac_srs->srs_next == NULL && mac_srs->srs_prev == NULL);
251 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
252 
253 	rw_enter(&mac_srs_g_lock, RW_WRITER);
254 	mutex_enter(&mac_srs->srs_lock);
255 
256 	ASSERT((mac_srs->srs_state & SRS_IN_GLIST) == 0);
257 
258 	if (mac_srs_g_list == NULL) {
259 		mac_srs_g_list = mac_srs;
260 	} else {
261 		mac_srs->srs_next = mac_srs_g_list;
262 		mac_srs_g_list->srs_prev = mac_srs;
263 		mac_srs->srs_prev = NULL;
264 		mac_srs_g_list = mac_srs;
265 	}
266 	mac_srs->srs_state |= SRS_IN_GLIST;
267 
268 	mutex_exit(&mac_srs->srs_lock);
269 	rw_exit(&mac_srs_g_lock);
270 }
271 
272 static void
273 mac_srs_remove_glist(mac_soft_ring_set_t *mac_srs)
274 {
275 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
276 
277 	rw_enter(&mac_srs_g_lock, RW_WRITER);
278 	mutex_enter(&mac_srs->srs_lock);
279 
280 	ASSERT((mac_srs->srs_state & SRS_IN_GLIST) != 0);
281 
282 	if (mac_srs == mac_srs_g_list) {
283 		mac_srs_g_list = mac_srs->srs_next;
284 		if (mac_srs_g_list != NULL)
285 			mac_srs_g_list->srs_prev = NULL;
286 	} else {
287 		mac_srs->srs_prev->srs_next = mac_srs->srs_next;
288 		if (mac_srs->srs_next != NULL)
289 			mac_srs->srs_next->srs_prev = mac_srs->srs_prev;
290 	}
291 	mac_srs->srs_state &= ~SRS_IN_GLIST;
292 
293 	mutex_exit(&mac_srs->srs_lock);
294 	rw_exit(&mac_srs_g_lock);
295 }
296 
297 /* POLLING SETUP AND TEAR DOWN ROUTINES */
298 
299 /*
300  * mac_srs_client_poll_quiesce and mac_srs_client_poll_restart
301  *
302  * These routines are used to call back into the upper layer
303  * (primarily TCP squeue) to stop polling the soft rings or
304  * restart polling.
305  */
306 void
307 mac_srs_client_poll_quiesce(mac_client_impl_t *mcip,
308     mac_soft_ring_set_t *mac_srs)
309 {
310 	mac_soft_ring_t	*softring;
311 
312 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
313 
314 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
315 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
316 		return;
317 	}
318 
319 	for (softring = mac_srs->srs_soft_ring_head;
320 	    softring != NULL; softring = softring->s_ring_next) {
321 		if ((softring->s_ring_type & ST_RING_TCP) &&
322 		    (softring->s_ring_rx_arg2 != NULL)) {
323 			mcip->mci_resource_quiesce(mcip->mci_resource_arg,
324 			    softring->s_ring_rx_arg2);
325 		}
326 	}
327 }
328 
329 void
330 mac_srs_client_poll_restart(mac_client_impl_t *mcip,
331     mac_soft_ring_set_t *mac_srs)
332 {
333 	mac_soft_ring_t	*softring;
334 
335 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
336 
337 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
338 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
339 		return;
340 	}
341 
342 	for (softring = mac_srs->srs_soft_ring_head;
343 	    softring != NULL; softring = softring->s_ring_next) {
344 		if ((softring->s_ring_type & ST_RING_TCP) &&
345 		    (softring->s_ring_rx_arg2 != NULL)) {
346 			mcip->mci_resource_restart(mcip->mci_resource_arg,
347 			    softring->s_ring_rx_arg2);
348 		}
349 	}
350 }
351 
352 /*
353  * Register the given SRS and associated soft rings with the consumer and
354  * enable the polling interface used by the consumer.(i.e IP) over this
355  * SRS and associated soft rings.
356  */
357 void
358 mac_srs_client_poll_enable(mac_client_impl_t *mcip,
359     mac_soft_ring_set_t *mac_srs)
360 {
361 	mac_rx_fifo_t		mrf;
362 	mac_soft_ring_t		*softring;
363 
364 	ASSERT(mac_srs->srs_mcip == mcip);
365 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
366 
367 	if (!(mcip->mci_state_flags & MCIS_CLIENT_POLL_CAPABLE))
368 		return;
369 
370 	bzero(&mrf, sizeof (mac_rx_fifo_t));
371 	mrf.mrf_type = MAC_RX_FIFO;
372 
373 	/*
374 	 * A SRS is capable of acting as a soft ring for cases
375 	 * where no fanout is needed. This is the case for userland
376 	 * flows.
377 	 */
378 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS)
379 		return;
380 
381 	mrf.mrf_receive = (mac_receive_t)mac_soft_ring_poll;
382 	mrf.mrf_intr_enable = (mac_intr_enable_t)mac_soft_ring_intr_enable;
383 	mrf.mrf_intr_disable = (mac_intr_disable_t)mac_soft_ring_intr_disable;
384 	mac_srs->srs_type |= SRST_CLIENT_POLL_ENABLED;
385 
386 	softring = mac_srs->srs_soft_ring_head;
387 	while (softring != NULL) {
388 		if (softring->s_ring_type & (ST_RING_TCP | ST_RING_UDP)) {
389 			/*
390 			 * TCP and UDP support DLS bypass. Squeue polling
391 			 * support implies DLS bypass since the squeue poll
392 			 * path does not have DLS processing.
393 			 */
394 			mac_soft_ring_dls_bypass(softring,
395 			    mcip->mci_direct_rx_fn, mcip->mci_direct_rx_arg);
396 		}
397 		/*
398 		 * Non-TCP protocols don't support squeues. Hence we don't
399 		 * make any ring addition callbacks for non-TCP rings
400 		 */
401 		if (!(softring->s_ring_type & ST_RING_TCP)) {
402 			softring->s_ring_rx_arg2 = NULL;
403 			softring = softring->s_ring_next;
404 			continue;
405 		}
406 		mrf.mrf_rx_arg = softring;
407 		mrf.mrf_intr_handle = (mac_intr_handle_t)softring;
408 		mrf.mrf_cpu_id = softring->s_ring_cpuid;
409 		mrf.mrf_flow_priority = mac_srs->srs_pri;
410 
411 		softring->s_ring_rx_arg2 = mcip->mci_resource_add(
412 		    mcip->mci_resource_arg, (mac_resource_t *)&mrf);
413 
414 		softring = softring->s_ring_next;
415 	}
416 }
417 
418 /*
419  * Unregister the given SRS and associated soft rings with the consumer and
420  * disable the polling interface used by the consumer.(i.e IP) over this
421  * SRS and associated soft rings.
422  */
423 void
424 mac_srs_client_poll_disable(mac_client_impl_t *mcip,
425     mac_soft_ring_set_t *mac_srs)
426 {
427 	mac_soft_ring_t		*softring;
428 
429 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
430 
431 	/*
432 	 * A SRS is capable of acting as a soft ring for cases
433 	 * where no protocol fanout is needed. This is the case
434 	 * for userland flows. Nothing to do here.
435 	 */
436 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS)
437 		return;
438 
439 	mutex_enter(&mac_srs->srs_lock);
440 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
441 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
442 		mutex_exit(&mac_srs->srs_lock);
443 		return;
444 	}
445 	mac_srs->srs_type &= ~(SRST_CLIENT_POLL_ENABLED | SRST_DLS_BYPASS);
446 	mutex_exit(&mac_srs->srs_lock);
447 
448 	/*
449 	 * DLS bypass is now disabled in the case of both TCP and UDP.
450 	 * Reset the soft ring callbacks to the standard 'mac_rx_deliver'
451 	 * callback. In addition, in the case of TCP, invoke IP's callback
452 	 * for ring removal.
453 	 */
454 	for (softring = mac_srs->srs_soft_ring_head;
455 	    softring != NULL; softring = softring->s_ring_next) {
456 		if (!(softring->s_ring_type & (ST_RING_UDP | ST_RING_TCP)))
457 			continue;
458 
459 		if ((softring->s_ring_type & ST_RING_TCP) &&
460 		    softring->s_ring_rx_arg2 != NULL) {
461 			mcip->mci_resource_remove(mcip->mci_resource_arg,
462 			    softring->s_ring_rx_arg2);
463 		}
464 
465 		mutex_enter(&softring->s_ring_lock);
466 		while (softring->s_ring_state & S_RING_PROC) {
467 			softring->s_ring_state |= S_RING_CLIENT_WAIT;
468 			cv_wait(&softring->s_ring_client_cv,
469 			    &softring->s_ring_lock);
470 		}
471 		softring->s_ring_state &= ~S_RING_CLIENT_WAIT;
472 		softring->s_ring_rx_arg2 = NULL;
473 		softring->s_ring_rx_func = mac_rx_deliver;
474 		softring->s_ring_rx_arg1 = mcip;
475 		mutex_exit(&softring->s_ring_lock);
476 	}
477 }
478 
479 /*
480  * Enable or disable poll capability of the SRS on the underlying Rx ring.
481  *
482  * There is a need to enable or disable the poll capability of an SRS over an
483  * Rx ring depending on the number of mac clients sharing the ring and also
484  * whether user flows are configured on it. However the poll state is actively
485  * manipulated by the SRS worker and poll threads and uncoordinated changes by
486  * yet another thread to the underlying capability can surprise them leading
487  * to assert failures. Instead we quiesce the SRS, make the changes and then
488  * restart the SRS.
489  */
490 static void
491 mac_srs_poll_state_change(mac_soft_ring_set_t *mac_srs,
492     boolean_t turn_off_poll_capab, mac_rx_func_t rx_func)
493 {
494 	boolean_t	need_restart = B_FALSE;
495 	mac_srs_rx_t	*srs_rx = &mac_srs->srs_rx;
496 	mac_ring_t	*ring;
497 
498 	if (!SRS_QUIESCED(mac_srs)) {
499 		mac_rx_srs_quiesce(mac_srs, SRS_QUIESCE);
500 		need_restart = B_TRUE;
501 	}
502 
503 	ring = mac_srs->srs_ring;
504 	if ((ring != NULL) &&
505 	    (ring->mr_classify_type == MAC_HW_CLASSIFIER)) {
506 		if (turn_off_poll_capab)
507 			mac_srs->srs_state &= ~SRS_POLLING_CAPAB;
508 		else if (mac_poll_enable)
509 			mac_srs->srs_state |= SRS_POLLING_CAPAB;
510 	}
511 	srs_rx->sr_lower_proc = rx_func;
512 
513 	if (need_restart)
514 		mac_rx_srs_restart(mac_srs);
515 }
516 
517 /* CPU RECONFIGURATION AND FANOUT COMPUTATION ROUTINES */
518 
519 /*
520  * Return the next CPU to be used to bind a MAC kernel thread.
521  */
522 static processorid_t
523 mac_next_bind_cpu(void)
524 {
525 	static processorid_t srs_curr_cpu = -1;
526 	cpu_t *cp;
527 
528 	ASSERT(MUTEX_HELD(&cpu_lock));
529 
530 	srs_curr_cpu++;
531 	cp = cpu_get(srs_curr_cpu);
532 	if (cp == NULL || !cpu_is_online(cp))
533 		srs_curr_cpu = srs_bind_cpu;
534 
535 	return (srs_curr_cpu);
536 }
537 
538 /* ARGSUSED */
539 static int
540 mac_srs_cpu_setup(cpu_setup_t what, int id, void *arg)
541 {
542 	ASSERT(MUTEX_HELD(&cpu_lock));
543 	switch (what) {
544 	case CPU_CONFIG:
545 	case CPU_ON:
546 	case CPU_CPUPART_IN:
547 		mac_walk_srs_and_bind(id);
548 		break;
549 
550 	case CPU_UNCONFIG:
551 	case CPU_OFF:
552 	case CPU_CPUPART_OUT:
553 		mac_walk_srs_and_unbind(id);
554 		break;
555 
556 	default:
557 		break;
558 	}
559 	return (0);
560 }
561 
562 /*
563  * mac_compute_soft_ring_count():
564  *
565  * This routine computes the number of soft rings needed to handle incoming
566  * load given a flow_entry.
567  *
568  * The routine does the following:
569  * 1) soft rings will be created if mac_soft_ring_enable is set.
570  * 2) If the underlying link is a 10Gbps link, then soft rings will be
571  * created even if mac_soft_ring_enable is not set. The number of soft
572  * rings, so created,  will equal mac_rx_soft_ring_10gig_count.
573  * 3) On a sun4v platform (i.e., mac_soft_ring_enable is set), 2 times the
574  * mac_rx_soft_ring_10gig_count number of soft rings will be created for a
575  * 10Gbps link.
576  *
577  * If a bandwidth limit is specified, the number that gets computed is
578  * dependent upon CPU speed, the number of Rx rings configured, and
579  * the bandwidth limit.
580  * If more Rx rings are available, less number of soft rings is needed.
581  *
582  * mac_use_bw_heuristic is another "hidden" variable that can be used to
583  * override the default use of soft ring count computation. Depending upon
584  * the usefulness of it, mac_use_bw_heuristic can later be made into a
585  * data-link property or removed altogether.
586  *
587  * TODO: Cleanup and tighten some of the assumptions.
588  */
589 boolean_t mac_use_bw_heuristic = B_TRUE;
590 static int
591 mac_compute_soft_ring_count(flow_entry_t *flent, int rx_srs_cnt)
592 {
593 	uint64_t cpu_speed, bw = 0;
594 	int srings = 0;
595 	boolean_t bw_enabled = B_FALSE;
596 
597 	ASSERT(!(flent->fe_type & FLOW_USER));
598 	if (flent->fe_resource_props.mrp_mask & MRP_MAXBW &&
599 	    mac_use_bw_heuristic) {
600 		/* bandwidth enabled */
601 		bw_enabled = B_TRUE;
602 		bw = flent->fe_resource_props.mrp_maxbw;
603 	}
604 	if (!bw_enabled) {
605 		/* No bandwidth enabled */
606 		if (mac_soft_ring_enable)
607 			srings = mac_rx_soft_ring_count;
608 
609 		/* Is this a 10Gig link? */
610 		flent->fe_nic_speed = mac_client_stat_get(flent->fe_mcip,
611 		    MAC_STAT_IFSPEED);
612 		/* convert to Mbps */
613 		if (((flent->fe_nic_speed)/1000000) > 1000 &&
614 		    mac_rx_soft_ring_10gig_count > 0) {
615 			/* This is a 10Gig link */
616 			srings = mac_rx_soft_ring_10gig_count;
617 			/*
618 			 * Use 2 times mac_rx_soft_ring_10gig_count for
619 			 * sun4v systems.
620 			 */
621 			if (mac_soft_ring_enable)
622 				srings = srings * 2;
623 		}
624 	} else {
625 		/*
626 		 * Soft ring computation using CPU speed and specified
627 		 * bandwidth limit.
628 		 */
629 		/* Assumption: all CPUs have the same frequency */
630 		cpu_speed = (uint64_t)CPU->cpu_type_info.pi_clock;
631 
632 		/* cpu_speed is in MHz; make bw in units of Mbps.  */
633 		bw = bw/1000000;
634 
635 		if (bw >= 1000) {
636 			/*
637 			 * bw is greater than or equal to 1Gbps.
638 			 * The number of soft rings required is a function
639 			 * of bandwidth and CPU speed. To keep this simple,
640 			 * let's use this rule: 1GHz CPU can handle 1Gbps.
641 			 * If bw is less than 1 Gbps, then there is no need
642 			 * for soft rings. Assumption is that CPU speeds
643 			 * (on modern systems) are at least 1GHz.
644 			 */
645 			srings = bw/cpu_speed;
646 			if (srings <= 1 && mac_soft_ring_enable) {
647 				/*
648 				 * Give at least 2 soft rings
649 				 * for sun4v systems
650 				 */
651 				srings = 2;
652 			}
653 		}
654 	}
655 	/*
656 	 * If the flent has multiple Rx SRSs, then each SRS need not
657 	 * have that many soft rings on top of it. The number of
658 	 * soft rings for each Rx SRS is found by dividing srings by
659 	 * rx_srs_cnt.
660 	 */
661 	if (rx_srs_cnt > 1) {
662 		int remainder;
663 
664 		remainder = srings%rx_srs_cnt;
665 		srings = srings/rx_srs_cnt;
666 		if (remainder != 0)
667 			srings++;
668 		/*
669 		 * Fanning out to 1 soft ring is not very useful.
670 		 * Set it as well to 0 and mac_srs_fanout_init()
671 		 * will take care of creating a single soft ring
672 		 * for proto fanout.
673 		 */
674 		if (srings == 1)
675 			srings = 0;
676 	}
677 	/* Do some more massaging */
678 	srings = min(srings, ncpus);
679 	srings = min(srings, MAX_SR_FANOUT);
680 	return (srings);
681 }
682 
683 /*
684  * Assignment of user specified CPUs to a link.
685  *
686  * Minimum CPUs required to get an optimal assignmet:
687  * For each Rx SRS, atleast two CPUs are needed if mac_latency_optimize
688  * flag is set -- one for polling, one for fanout soft ring.
689  * If mac_latency_optimize is not set, then 3 CPUs are needed -- one
690  * for polling, one for SRS worker thread and one for fanout soft ring.
691  *
692  * The CPUs needed for Tx side is equal to the number of Tx rings
693  * the link is using.
694  *
695  * mac_flow_user_cpu_init() categorizes the CPU assignment depending
696  * upon the number of CPUs in 3 different buckets.
697  *
698  * In the first bucket, the most optimal case is handled. The user has
699  * passed enough number of CPUs and every thread gets its own CPU.
700  *
701  * The second and third are the sub-optimal cases. Enough CPUs are not
702  * available.
703  *
704  * The second bucket handles the case where atleast one distinct CPU is
705  * is available for each of the Rx rings (Rx SRSes) and Tx rings (Tx
706  * SRS or soft rings).
707  *
708  * In the third case (worst case scenario), specified CPU count is less
709  * than the Rx rings configured for the link. In this case, we round
710  * robin the CPUs among the Rx SRSes and Tx SRS/soft rings.
711  */
712 static void
713 mac_flow_user_cpu_init(flow_entry_t *flent, mac_resource_props_t *mrp)
714 {
715 	mac_soft_ring_set_t *rx_srs, *tx_srs;
716 	int i, srs_cnt;
717 	mac_cpus_t *srs_cpu;
718 	int no_of_cpus, cpu_cnt;
719 	int rx_srs_cnt, reqd_rx_cpu_cnt;
720 	int fanout_cpu_cnt, reqd_tx_cpu_cnt;
721 	int reqd_poll_worker_cnt, fanout_cnt_per_srs;
722 
723 	ASSERT(mrp->mrp_fanout_mode == MCM_CPUS);
724 	/*
725 	 * The check for nbc_ncpus to be within limits for
726 	 * the user specified case was done earlier and if
727 	 * not within limits, an error would have been
728 	 * returned to the user.
729 	 */
730 	ASSERT(mrp->mrp_ncpus > 0 && mrp->mrp_ncpus <= MAX_SR_FANOUT);
731 
732 	no_of_cpus = mrp->mrp_ncpus;
733 
734 	if (mrp->mrp_intr_cpu != -1) {
735 		/*
736 		 * interrupt has been re-targetted. Poll
737 		 * thread needs to be bound to interrupt
738 		 * CPU. Presently only fixed interrupts
739 		 * are re-targetted, MSI-x aren't.
740 		 *
741 		 * Find where in the list is the intr
742 		 * CPU and swap it with the first one.
743 		 * We will be using the first CPU in the
744 		 * list for poll.
745 		 */
746 		for (i = 0; i < no_of_cpus; i++) {
747 			if (mrp->mrp_cpu[i] == mrp->mrp_intr_cpu)
748 				break;
749 		}
750 		mrp->mrp_cpu[i] = mrp->mrp_cpu[0];
751 		mrp->mrp_cpu[0] = mrp->mrp_intr_cpu;
752 	}
753 
754 	/*
755 	 * Requirements:
756 	 * The number of CPUs that each Rx ring needs is dependent
757 	 * upon mac_latency_optimize flag.
758 	 * 1) If set, atleast 2 CPUs are needed -- one for
759 	 * polling, one for fanout soft ring.
760 	 * 2) If not set, then atleast 3 CPUs are needed -- one
761 	 * for polling, one for srs worker thread, and one for
762 	 * fanout soft ring.
763 	 */
764 	rx_srs_cnt = (flent->fe_rx_srs_cnt > 1) ?
765 	    (flent->fe_rx_srs_cnt - 1) : flent->fe_rx_srs_cnt;
766 	reqd_rx_cpu_cnt = mac_latency_optimize ?
767 	    (rx_srs_cnt * 2) : (rx_srs_cnt * 3);
768 
769 	/* How many CPUs are needed for Tx side? */
770 	tx_srs = flent->fe_tx_srs;
771 	reqd_tx_cpu_cnt = TX_MULTI_RING_MODE(tx_srs) ?
772 	    tx_srs->srs_oth_ring_count : 1;
773 
774 	/* CPUs needed for Rx SRSes poll and worker threads */
775 	reqd_poll_worker_cnt = mac_latency_optimize ?
776 	    rx_srs_cnt : rx_srs_cnt * 2;
777 
778 	/* Has the user provided enough CPUs? */
779 	if (no_of_cpus >= (reqd_rx_cpu_cnt + reqd_tx_cpu_cnt)) {
780 		/*
781 		 * Best case scenario. There is enough CPUs. All
782 		 * Rx rings will get their own set of CPUs plus
783 		 * Tx soft rings will get their own.
784 		 */
785 		/*
786 		 * fanout_cpu_cnt is the number of CPUs available
787 		 * for Rx side fanout soft rings.
788 		 */
789 		fanout_cpu_cnt = no_of_cpus -
790 		    reqd_poll_worker_cnt - reqd_tx_cpu_cnt;
791 
792 		/*
793 		 * Divide fanout_cpu_cnt by rx_srs_cnt to find
794 		 * out how many fanout soft rings each Rx SRS
795 		 * can have.
796 		 */
797 		fanout_cnt_per_srs = fanout_cpu_cnt/rx_srs_cnt;
798 
799 		/* Do the assignment for the default Rx ring */
800 		cpu_cnt = 0;
801 		rx_srs = flent->fe_rx_srs[0];
802 		ASSERT(rx_srs->srs_ring == NULL);
803 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
804 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
805 		srs_cpu = &rx_srs->srs_cpu;
806 		srs_cpu->mc_ncpus = no_of_cpus;
807 		bcopy(mrp->mrp_cpu,
808 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
809 		srs_cpu->mc_fanout_cnt = fanout_cnt_per_srs;
810 		srs_cpu->mc_pollid = mrp->mrp_cpu[cpu_cnt++];
811 		srs_cpu->mc_intr_cpu = mrp->mrp_intr_cpu;
812 		srs_cpu->mc_workerid = srs_cpu->mc_pollid;
813 		if (!mac_latency_optimize)
814 			srs_cpu->mc_workerid = mrp->mrp_cpu[cpu_cnt++];
815 		for (i = 0; i < fanout_cnt_per_srs; i++)
816 			srs_cpu->mc_fanout_cpus[i] = mrp->mrp_cpu[cpu_cnt++];
817 
818 		/* Do the assignment for h/w Rx SRSes */
819 		if (flent->fe_rx_srs_cnt > 1) {
820 			cpu_cnt = 0;
821 			for (srs_cnt = 1;
822 			    srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
823 				rx_srs = flent->fe_rx_srs[srs_cnt];
824 				ASSERT(rx_srs->srs_ring != NULL);
825 				if (rx_srs->srs_fanout_state ==
826 				    SRS_FANOUT_INIT) {
827 					rx_srs->srs_fanout_state =
828 					    SRS_FANOUT_REINIT;
829 				}
830 				srs_cpu = &rx_srs->srs_cpu;
831 				srs_cpu->mc_ncpus = no_of_cpus;
832 				bcopy(mrp->mrp_cpu, srs_cpu->mc_cpus,
833 				    sizeof (srs_cpu->mc_cpus));
834 				srs_cpu->mc_fanout_cnt = fanout_cnt_per_srs;
835 				/* The first CPU in the list is the intr CPU */
836 				srs_cpu->mc_pollid = mrp->mrp_cpu[cpu_cnt++];
837 				srs_cpu->mc_intr_cpu = mrp->mrp_intr_cpu;
838 				srs_cpu->mc_workerid = srs_cpu->mc_pollid;
839 				if (!mac_latency_optimize) {
840 					srs_cpu->mc_workerid =
841 					    mrp->mrp_cpu[cpu_cnt++];
842 				}
843 				for (i = 0; i < fanout_cnt_per_srs; i++) {
844 					srs_cpu->mc_fanout_cpus[i] =
845 					    mrp->mrp_cpu[cpu_cnt++];
846 				}
847 				ASSERT(cpu_cnt <= no_of_cpus);
848 			}
849 		}
850 		return;
851 	}
852 
853 	/*
854 	 * Sub-optimal case.
855 	 * We have the following information:
856 	 * no_of_cpus - no. of cpus that user passed.
857 	 * rx_srs_cnt - no. of rx rings.
858 	 * reqd_rx_cpu_cnt = mac_latency_optimize?rx_srs_cnt*2:rx_srs_cnt*3
859 	 * reqd_tx_cpu_cnt - no. of cpus reqd. for Tx side.
860 	 * reqd_poll_worker_cnt = mac_latency_optimize?rx_srs_cnt:rx_srs_cnt*2
861 	 */
862 	/*
863 	 * If we bind the Rx fanout soft rings to the same CPUs
864 	 * as poll/worker, would that be enough?
865 	 */
866 	if (no_of_cpus >= (rx_srs_cnt + reqd_tx_cpu_cnt)) {
867 		boolean_t worker_assign = B_FALSE;
868 
869 		/*
870 		 * If mac_latency_optimize is not set, are there
871 		 * enough CPUs to assign a CPU for worker also?
872 		 */
873 		if (no_of_cpus >= (reqd_poll_worker_cnt + reqd_tx_cpu_cnt))
874 			worker_assign = B_TRUE;
875 		/*
876 		 * Zero'th Rx SRS is the default Rx ring. It is not
877 		 * associated with h/w Rx ring.
878 		 */
879 		rx_srs = flent->fe_rx_srs[0];
880 		ASSERT(rx_srs->srs_ring == NULL);
881 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
882 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
883 		cpu_cnt = 0;
884 		srs_cpu = &rx_srs->srs_cpu;
885 		srs_cpu->mc_ncpus = no_of_cpus;
886 		bcopy(mrp->mrp_cpu,
887 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
888 		srs_cpu->mc_fanout_cnt = 1;
889 		srs_cpu->mc_pollid = mrp->mrp_cpu[cpu_cnt++];
890 		srs_cpu->mc_intr_cpu = mrp->mrp_intr_cpu;
891 		srs_cpu->mc_workerid = srs_cpu->mc_pollid;
892 		if (!mac_latency_optimize && worker_assign)
893 			srs_cpu->mc_workerid = mrp->mrp_cpu[cpu_cnt++];
894 		srs_cpu->mc_fanout_cpus[0] = mrp->mrp_cpu[cpu_cnt];
895 
896 		/* Do CPU bindings for SRSes having h/w Rx rings */
897 		if (flent->fe_rx_srs_cnt > 1) {
898 			cpu_cnt = 0;
899 			for (srs_cnt = 1;
900 			    srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
901 				rx_srs = flent->fe_rx_srs[srs_cnt];
902 				ASSERT(rx_srs->srs_ring != NULL);
903 				if (rx_srs->srs_fanout_state ==
904 				    SRS_FANOUT_INIT) {
905 					rx_srs->srs_fanout_state =
906 					    SRS_FANOUT_REINIT;
907 				}
908 				srs_cpu = &rx_srs->srs_cpu;
909 				srs_cpu->mc_ncpus = no_of_cpus;
910 				bcopy(mrp->mrp_cpu, srs_cpu->mc_cpus,
911 				    sizeof (srs_cpu->mc_cpus));
912 				srs_cpu->mc_pollid =
913 				    mrp->mrp_cpu[cpu_cnt];
914 				srs_cpu->mc_intr_cpu = mrp->mrp_intr_cpu;
915 				srs_cpu->mc_workerid = srs_cpu->mc_pollid;
916 				if (!mac_latency_optimize && worker_assign) {
917 					srs_cpu->mc_workerid =
918 					    mrp->mrp_cpu[++cpu_cnt];
919 				}
920 				srs_cpu->mc_fanout_cnt = 1;
921 				srs_cpu->mc_fanout_cpus[0] =
922 				    mrp->mrp_cpu[cpu_cnt];
923 				cpu_cnt++;
924 				ASSERT(cpu_cnt <= no_of_cpus);
925 			}
926 		}
927 		return;
928 	}
929 
930 	/*
931 	 * Real sub-optimal case. Not enough CPUs for poll and
932 	 * Tx soft rings. Do a round robin assignment where
933 	 * each Rx SRS will get the same CPU for poll, worker
934 	 * and fanout soft ring.
935 	 */
936 	cpu_cnt = 0;
937 	for (srs_cnt = 0; srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
938 		rx_srs = flent->fe_rx_srs[srs_cnt];
939 		srs_cpu = &rx_srs->srs_cpu;
940 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
941 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
942 		srs_cpu->mc_ncpus = no_of_cpus;
943 		bcopy(mrp->mrp_cpu,
944 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
945 		srs_cpu->mc_fanout_cnt = 1;
946 		srs_cpu->mc_pollid = mrp->mrp_cpu[cpu_cnt];
947 		srs_cpu->mc_intr_cpu = mrp->mrp_intr_cpu;
948 		srs_cpu->mc_workerid = mrp->mrp_cpu[cpu_cnt];
949 		srs_cpu->mc_fanout_cpus[0] = mrp->mrp_cpu[cpu_cnt];
950 		if (++cpu_cnt >= no_of_cpus)
951 			cpu_cnt = 0;
952 	}
953 }
954 
955 /*
956  * mac_flow_cpu_init():
957  *
958  * Each SRS has a mac_cpu_t structure, srs_cpu. This routine fills in
959  * the CPU binding information in srs_cpu for all Rx SRSes associated
960  * with a flent.
961  */
962 static void
963 mac_flow_cpu_init(flow_entry_t *flent, mac_resource_props_t *mrp)
964 {
965 	mac_soft_ring_set_t *rx_srs;
966 	processorid_t cpuid;
967 	int j, srs_cnt, soft_ring_cnt = 0;
968 	mac_cpus_t *srs_cpu;
969 
970 	if (mrp->mrp_mask & MRP_CPUS_USERSPEC) {
971 		mac_flow_user_cpu_init(flent, mrp);
972 	} else {
973 		/*
974 		 * Compute the number of soft rings needed on top for each Rx
975 		 * SRS. "rx_srs_cnt-1" indicates the number of Rx SRS
976 		 * associated with h/w Rx rings. Soft ring count needed for
977 		 * each h/w Rx SRS is computed and the same is applied to
978 		 * software classified Rx SRS. The first Rx SRS in fe_rx_srs[]
979 		 * is the software classified Rx SRS.
980 		 */
981 		soft_ring_cnt = mac_compute_soft_ring_count(flent,
982 		    flent->fe_rx_srs_cnt - 1);
983 		if (soft_ring_cnt == 0) {
984 			/*
985 			 * Even when soft_ring_cnt is 0, we still need
986 			 * to create a soft ring for TCP, UDP and
987 			 * OTHER. So set it to 1.
988 			 */
989 			soft_ring_cnt = 1;
990 		}
991 		for (srs_cnt = 0; srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
992 			rx_srs = flent->fe_rx_srs[srs_cnt];
993 			srs_cpu = &rx_srs->srs_cpu;
994 			if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT) {
995 				if (soft_ring_cnt == srs_cpu->mc_fanout_cnt)
996 					continue;
997 				rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
998 			}
999 			srs_cpu->mc_ncpus = soft_ring_cnt;
1000 			srs_cpu->mc_fanout_cnt = soft_ring_cnt;
1001 			mutex_enter(&cpu_lock);
1002 			for (j = 0; j < soft_ring_cnt; j++) {
1003 				cpuid = mac_next_bind_cpu();
1004 				srs_cpu->mc_cpus[j] = cpuid;
1005 				srs_cpu->mc_fanout_cpus[j] = cpuid;
1006 			}
1007 			cpuid = mac_next_bind_cpu();
1008 			srs_cpu->mc_pollid = cpuid;
1009 			/* increment ncpus to account for polling cpu */
1010 			srs_cpu->mc_ncpus++;
1011 			srs_cpu->mc_cpus[j++] = cpuid;
1012 			if (!mac_latency_optimize) {
1013 				cpuid = mac_next_bind_cpu();
1014 				srs_cpu->mc_ncpus++;
1015 				srs_cpu->mc_cpus[j++] = cpuid;
1016 			}
1017 			srs_cpu->mc_workerid = cpuid;
1018 			mutex_exit(&cpu_lock);
1019 		}
1020 	}
1021 }
1022 
1023 /*
1024  * DATAPATH SETUP ROUTINES
1025  * (setup SRS and set/update FANOUT, B/W and PRIORITY)
1026  */
1027 
1028 static void
1029 mac_srs_fanout_list_alloc(mac_soft_ring_set_t *mac_srs)
1030 {
1031 	mac_srs->srs_tcp_soft_rings = (mac_soft_ring_t **)
1032 	    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT, KM_SLEEP);
1033 	mac_srs->srs_udp_soft_rings = (mac_soft_ring_t **)
1034 	    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT, KM_SLEEP);
1035 	mac_srs->srs_oth_soft_rings = (mac_soft_ring_t **)
1036 	    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT, KM_SLEEP);
1037 }
1038 
1039 static void
1040 mac_srs_worker_bind(mac_soft_ring_set_t *mac_srs, processorid_t cpuid)
1041 {
1042 	cpu_t *cp;
1043 	boolean_t clear = B_FALSE;
1044 
1045 	ASSERT(MUTEX_HELD(&cpu_lock));
1046 
1047 	if (!mac_srs_thread_bind)
1048 		return;
1049 
1050 	cp = cpu_get(cpuid);
1051 	if (cp == NULL || !cpu_is_online(cp))
1052 		return;
1053 
1054 	mutex_enter(&mac_srs->srs_lock);
1055 	mac_srs->srs_state |= SRS_WORKER_BOUND;
1056 	if (mac_srs->srs_worker_cpuid != -1)
1057 		clear = B_TRUE;
1058 	mac_srs->srs_worker_cpuid = cpuid;
1059 	mutex_exit(&mac_srs->srs_lock);
1060 
1061 	if (clear)
1062 		thread_affinity_clear(mac_srs->srs_worker);
1063 
1064 	thread_affinity_set(mac_srs->srs_worker, cpuid);
1065 	DTRACE_PROBE1(worker__CPU, processorid_t, cpuid);
1066 }
1067 
1068 static void
1069 mac_srs_poll_bind(mac_soft_ring_set_t *mac_srs, processorid_t cpuid)
1070 {
1071 	cpu_t *cp;
1072 	boolean_t clear = B_FALSE;
1073 
1074 	ASSERT(MUTEX_HELD(&cpu_lock));
1075 
1076 	if (!mac_srs_thread_bind || mac_srs->srs_poll_thr == NULL)
1077 		return;
1078 
1079 	cp = cpu_get(cpuid);
1080 	if (cp == NULL || !cpu_is_online(cp))
1081 		return;
1082 
1083 	mutex_enter(&mac_srs->srs_lock);
1084 	mac_srs->srs_state |= SRS_POLL_BOUND;
1085 	if (mac_srs->srs_poll_cpuid != -1)
1086 		clear = B_TRUE;
1087 	mac_srs->srs_poll_cpuid = cpuid;
1088 	mutex_exit(&mac_srs->srs_lock);
1089 
1090 	if (clear)
1091 		thread_affinity_clear(mac_srs->srs_poll_thr);
1092 
1093 	thread_affinity_set(mac_srs->srs_poll_thr, cpuid);
1094 	DTRACE_PROBE1(poll__CPU, processorid_t, cpuid);
1095 }
1096 
1097 /*
1098  * When a CPU comes back online, bind the MAC kernel threads which
1099  * were previously bound to that CPU, and had to be unbound because
1100  * the CPU was going away.
1101  *
1102  * These functions are called with cpu_lock held and hence we can't
1103  * cv_wait to grab the mac perimeter. Since these functions walk the soft
1104  * ring list of an SRS without being in the perimeter, the list itself
1105  * is protected by the SRS lock.
1106  */
1107 static void
1108 mac_walk_srs_and_bind(int cpuid)
1109 {
1110 	mac_soft_ring_set_t *mac_srs;
1111 	mac_soft_ring_t *soft_ring;
1112 
1113 	rw_enter(&mac_srs_g_lock, RW_READER);
1114 
1115 	if ((mac_srs = mac_srs_g_list) == NULL)
1116 		goto done;
1117 
1118 	for (; mac_srs != NULL; mac_srs = mac_srs->srs_next) {
1119 		if (mac_srs->srs_worker_cpuid == -1 &&
1120 		    mac_srs->srs_worker_cpuid_save == cpuid) {
1121 			mac_srs->srs_worker_cpuid_save = -1;
1122 			mac_srs_worker_bind(mac_srs, cpuid);
1123 		}
1124 
1125 		if (!(mac_srs->srs_type & SRST_TX)) {
1126 			if (mac_srs->srs_poll_cpuid == -1 &&
1127 			    mac_srs->srs_poll_cpuid_save == cpuid) {
1128 				mac_srs->srs_poll_cpuid_save = -1;
1129 				mac_srs_poll_bind(mac_srs, cpuid);
1130 			}
1131 		}
1132 
1133 		/* Next tackle the soft rings associated with the srs */
1134 		mutex_enter(&mac_srs->srs_lock);
1135 		for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
1136 		    soft_ring = soft_ring->s_ring_next) {
1137 			if (soft_ring->s_ring_cpuid == -1 &&
1138 			    soft_ring->s_ring_cpuid_save == cpuid) {
1139 				soft_ring->s_ring_cpuid_save = -1;
1140 				(void) mac_soft_ring_bind(soft_ring, cpuid);
1141 			}
1142 		}
1143 		mutex_exit(&mac_srs->srs_lock);
1144 	}
1145 done:
1146 	rw_exit(&mac_srs_g_lock);
1147 }
1148 
1149 /*
1150  * Change the priority of the SRS's poll and worker thread. Additionally,
1151  * update the priority of the worker threads for the SRS's soft rings.
1152  * Need to modify any associated squeue threads.
1153  */
1154 void
1155 mac_update_srs_priority(mac_soft_ring_set_t *mac_srs, pri_t prival)
1156 {
1157 	mac_soft_ring_t		*ringp;
1158 
1159 	mac_srs->srs_pri = prival;
1160 	thread_lock(mac_srs->srs_worker);
1161 	(void) thread_change_pri(mac_srs->srs_worker, mac_srs->srs_pri, 0);
1162 	thread_unlock(mac_srs->srs_worker);
1163 	if (mac_srs->srs_poll_thr != NULL) {
1164 		thread_lock(mac_srs->srs_poll_thr);
1165 		(void) thread_change_pri(mac_srs->srs_poll_thr,
1166 		    mac_srs->srs_pri, 0);
1167 		thread_unlock(mac_srs->srs_poll_thr);
1168 	}
1169 	if ((ringp = mac_srs->srs_soft_ring_head) == NULL)
1170 		return;
1171 	while (ringp != mac_srs->srs_soft_ring_tail) {
1172 		thread_lock(ringp->s_ring_worker);
1173 		(void) thread_change_pri(ringp->s_ring_worker,
1174 		    mac_srs->srs_pri, 0);
1175 		thread_unlock(ringp->s_ring_worker);
1176 		ringp = ringp->s_ring_next;
1177 	}
1178 	ASSERT(ringp == mac_srs->srs_soft_ring_tail);
1179 	thread_lock(ringp->s_ring_worker);
1180 	(void) thread_change_pri(ringp->s_ring_worker, mac_srs->srs_pri, 0);
1181 	thread_unlock(ringp->s_ring_worker);
1182 }
1183 
1184 /*
1185  * Change the receive bandwidth limit.
1186  */
1187 static void
1188 mac_rx_srs_update_bwlimit(mac_soft_ring_set_t *srs, mac_resource_props_t *mrp)
1189 {
1190 	mac_soft_ring_t		*softring;
1191 
1192 	mutex_enter(&srs->srs_lock);
1193 	mutex_enter(&srs->srs_bw->mac_bw_lock);
1194 
1195 	if (mrp->mrp_maxbw == MRP_MAXBW_RESETVAL) {
1196 		/* Reset bandwidth limit */
1197 		if (srs->srs_type & SRST_BW_CONTROL) {
1198 			softring = srs->srs_soft_ring_head;
1199 			while (softring != NULL) {
1200 				softring->s_ring_type &= ~ST_RING_BW_CTL;
1201 				softring = softring->s_ring_next;
1202 			}
1203 			srs->srs_type &= ~SRST_BW_CONTROL;
1204 			srs->srs_drain_func = mac_rx_srs_drain;
1205 		}
1206 	} else {
1207 		/* Set/Modify bandwidth limit */
1208 		srs->srs_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1209 		/*
1210 		 * Give twice the queuing capability before
1211 		 * dropping packets. The unit is bytes/tick.
1212 		 */
1213 		srs->srs_bw->mac_bw_drop_threshold =
1214 		    srs->srs_bw->mac_bw_limit << 1;
1215 		if (!(srs->srs_type & SRST_BW_CONTROL)) {
1216 			softring = srs->srs_soft_ring_head;
1217 			while (softring != NULL) {
1218 				softring->s_ring_type |= ST_RING_BW_CTL;
1219 				softring = softring->s_ring_next;
1220 			}
1221 			srs->srs_type |= SRST_BW_CONTROL;
1222 			srs->srs_drain_func = mac_rx_srs_drain_bw;
1223 		}
1224 	}
1225 done:
1226 	mutex_exit(&srs->srs_bw->mac_bw_lock);
1227 	mutex_exit(&srs->srs_lock);
1228 }
1229 
1230 /* Change the transmit bandwidth limit */
1231 static void
1232 mac_tx_srs_update_bwlimit(mac_soft_ring_set_t *srs, mac_resource_props_t *mrp)
1233 {
1234 	uint32_t		tx_mode;
1235 	mac_srs_tx_t		*srs_tx = &srs->srs_tx;
1236 	mac_client_impl_t	*mcip = srs->srs_mcip;
1237 	mac_impl_t		*mip = mcip->mci_mip;
1238 
1239 	/*
1240 	 * We need to quiesce/restart the client here because mac_tx() and
1241 	 * srs->srs_tx->st_func do not hold srs->srs_lock while accessing
1242 	 * st_mode and related fields, which are modified by the code below.
1243 	 */
1244 	mac_tx_client_quiesce(mcip, SRS_QUIESCE);
1245 
1246 	mutex_enter(&srs->srs_lock);
1247 	mutex_enter(&srs->srs_bw->mac_bw_lock);
1248 
1249 	tx_mode = srs_tx->st_mode;
1250 	if (mrp->mrp_maxbw == MRP_MAXBW_RESETVAL) {
1251 		/* Reset bandwidth limit */
1252 		if (tx_mode == SRS_TX_BW) {
1253 			if (mac_tx_serialize ||
1254 			    (mip->mi_v12n_level & MAC_VIRT_SERIALIZE)) {
1255 				srs_tx->st_mode = SRS_TX_SERIALIZE;
1256 			} else {
1257 				srs_tx->st_mode = SRS_TX_DEFAULT;
1258 			}
1259 		} else if (tx_mode == SRS_TX_BW_FANOUT) {
1260 			srs_tx->st_mode = SRS_TX_FANOUT;
1261 		}
1262 		srs->srs_type &= ~SRST_BW_CONTROL;
1263 	} else {
1264 		/* Set/Modify bandwidth limit */
1265 		srs->srs_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1266 		/*
1267 		 * Give twice the queuing capability before
1268 		 * dropping packets. The unit is bytes/tick.
1269 		 */
1270 		srs->srs_bw->mac_bw_drop_threshold =
1271 		    srs->srs_bw->mac_bw_limit << 1;
1272 		srs->srs_type |= SRST_BW_CONTROL;
1273 		if (tx_mode != SRS_TX_BW &&
1274 		    tx_mode != SRS_TX_BW_FANOUT) {
1275 			if (tx_mode == SRS_TX_SERIALIZE ||
1276 			    tx_mode == SRS_TX_DEFAULT) {
1277 				srs_tx->st_mode = SRS_TX_BW;
1278 			} else if (tx_mode == SRS_TX_FANOUT) {
1279 				srs_tx->st_mode = SRS_TX_BW_FANOUT;
1280 			} else {
1281 				ASSERT(0);
1282 			}
1283 		}
1284 	}
1285 done:
1286 	srs_tx->st_func = mac_tx_get_func(srs_tx->st_mode);
1287 	mutex_exit(&srs->srs_bw->mac_bw_lock);
1288 	mutex_exit(&srs->srs_lock);
1289 
1290 	mac_tx_client_restart(mcip);
1291 }
1292 
1293 /*
1294  * The uber function that deals with any update to bandwidth limits.
1295  */
1296 void
1297 mac_srs_update_bwlimit(flow_entry_t *flent, mac_resource_props_t *mrp)
1298 {
1299 	int			count;
1300 
1301 	for (count = 0; count < flent->fe_rx_srs_cnt; count++)
1302 		mac_rx_srs_update_bwlimit(flent->fe_rx_srs[count], mrp);
1303 	mac_tx_srs_update_bwlimit(flent->fe_tx_srs, mrp);
1304 }
1305 
1306 void
1307 mac_srs_change_upcall(void *arg, mac_direct_rx_t rx_func, void *rx_arg1)
1308 {
1309 	mac_soft_ring_set_t	*mac_srs = arg;
1310 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1311 	mac_soft_ring_t		*softring;
1312 
1313 	mutex_enter(&mac_srs->srs_lock);
1314 	ASSERT((mac_srs->srs_type & SRST_TX) == 0);
1315 	srs_rx->sr_func = rx_func;
1316 	srs_rx->sr_arg1 = rx_arg1;
1317 
1318 	softring = mac_srs->srs_soft_ring_head;
1319 	while (softring != NULL) {
1320 		mutex_enter(&softring->s_ring_lock);
1321 		softring->s_ring_rx_func = rx_func;
1322 		softring->s_ring_rx_arg1 = rx_arg1;
1323 		mutex_exit(&softring->s_ring_lock);
1324 		softring = softring->s_ring_next;
1325 	}
1326 
1327 	mutex_exit(&mac_srs->srs_lock);
1328 }
1329 
1330 /*
1331  * When the first sub-flow is added to a link, we disable polling on the
1332  * link and also modify the entry point to mac_rx_srs_subflow_process.
1333  * (polling is disabled because with the subflow added, accounting
1334  * for polling needs additional logic, it is assumed that when a subflow is
1335  * added, we can take some hit as a result of disabling polling rather than
1336  * adding more complexity - if this becomes a perf. issue we need to
1337  * re-rvaluate this logic).  When the last subflow is removed, we turn back
1338  * polling and also reset the entry point to mac_rx_srs_process.
1339  *
1340  * In the future if there are multiple SRS, we can simply
1341  * take one and give it to the flow rather than disabling polling and
1342  * resetting the entry point.
1343  */
1344 void
1345 mac_client_update_classifier(mac_client_impl_t *mcip, boolean_t enable)
1346 {
1347 	flow_entry_t		*flent = mcip->mci_flent;
1348 	int			i;
1349 	mac_impl_t		*mip = mcip->mci_mip;
1350 	mac_rx_func_t		rx_func;
1351 	uint_t			rx_srs_cnt;
1352 	boolean_t		enable_classifier;
1353 
1354 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
1355 
1356 	enable_classifier = !FLOW_TAB_EMPTY(mcip->mci_subflow_tab) && enable;
1357 
1358 	rx_func = enable_classifier ? mac_rx_srs_subflow_process :
1359 	    mac_rx_srs_process;
1360 
1361 	/* Tell mac_srs_poll_state_change to disable polling if necessary */
1362 	if (mip->mi_state_flags & MIS_POLL_DISABLE)
1363 		enable_classifier = B_TRUE;
1364 
1365 	/*
1366 	 * If receive function has already been configured correctly for
1367 	 * current subflow configuration, do nothing.
1368 	 */
1369 	if (flent->fe_cb_fn == (flow_fn_t)rx_func)
1370 		return;
1371 
1372 	rx_srs_cnt = flent->fe_rx_srs_cnt;
1373 	for (i = 0; i < rx_srs_cnt; i++) {
1374 		ASSERT(flent->fe_rx_srs[i] != NULL);
1375 		mac_srs_poll_state_change(flent->fe_rx_srs[i],
1376 		    enable_classifier, rx_func);
1377 	}
1378 
1379 	/*
1380 	 * Change the S/W classifier so that we can land in the
1381 	 * correct processing function with correct argument.
1382 	 * If all subflows have been removed we can revert to
1383 	 * mac_rx_srsprocess, else we need mac_rx_srs_subflow_process.
1384 	 */
1385 	mutex_enter(&flent->fe_lock);
1386 	flent->fe_cb_fn = (flow_fn_t)rx_func;
1387 	flent->fe_cb_arg1 = (void *)mip;
1388 	flent->fe_cb_arg2 = flent->fe_rx_srs[0];
1389 	mutex_exit(&flent->fe_lock);
1390 }
1391 
1392 static void
1393 mac_srs_update_fanout_list(mac_soft_ring_set_t *mac_srs)
1394 {
1395 	int		tcp_count = 0;
1396 	int		udp_count = 0;
1397 	int		oth_count = 0;
1398 	mac_soft_ring_t *softring;
1399 
1400 	softring = mac_srs->srs_soft_ring_head;
1401 	if (softring == NULL) {
1402 		ASSERT(mac_srs->srs_soft_ring_count == 0);
1403 		mac_srs->srs_tcp_ring_count = 0;
1404 		mac_srs->srs_udp_ring_count = 0;
1405 		mac_srs->srs_oth_ring_count = 0;
1406 		return;
1407 	}
1408 
1409 	softring = mac_srs->srs_soft_ring_head;
1410 	tcp_count = udp_count = oth_count = 0;
1411 
1412 	while (softring != NULL) {
1413 		if (softring->s_ring_type & ST_RING_TCP)
1414 			mac_srs->srs_tcp_soft_rings[tcp_count++] = softring;
1415 		else if (softring->s_ring_type & ST_RING_UDP)
1416 			mac_srs->srs_udp_soft_rings[udp_count++] = softring;
1417 		else
1418 			mac_srs->srs_oth_soft_rings[oth_count++] = softring;
1419 		softring = softring->s_ring_next;
1420 	}
1421 
1422 	ASSERT(mac_srs->srs_soft_ring_count ==
1423 	    (tcp_count + udp_count + oth_count));
1424 
1425 	mac_srs->srs_tcp_ring_count = tcp_count;
1426 	mac_srs->srs_udp_ring_count = udp_count;
1427 	mac_srs->srs_oth_ring_count = oth_count;
1428 }
1429 
1430 void
1431 mac_srs_create_proto_softrings(int id, void *flent, uint16_t type,
1432     pri_t pri, mac_client_impl_t *mcip, mac_soft_ring_set_t *mac_srs,
1433     processorid_t cpuid, mac_direct_rx_t rx_func, void *x_arg1,
1434     mac_resource_handle_t x_arg2, boolean_t set_bypass)
1435 {
1436 	mac_soft_ring_t	*softring;
1437 	mac_rx_fifo_t	mrf;
1438 
1439 	bzero(&mrf, sizeof (mac_rx_fifo_t));
1440 	mrf.mrf_type = MAC_RX_FIFO;
1441 	mrf.mrf_receive = (mac_receive_t)mac_soft_ring_poll;
1442 	mrf.mrf_intr_enable =
1443 	    (mac_intr_enable_t)mac_soft_ring_intr_enable;
1444 	mrf.mrf_intr_disable =
1445 	    (mac_intr_disable_t)mac_soft_ring_intr_disable;
1446 	mrf.mrf_flow_priority = pri;
1447 
1448 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1449 	    (void *)flent, (type|ST_RING_TCP), pri, mcip, mac_srs,
1450 	    cpuid, rx_func, x_arg1, x_arg2);
1451 	softring->s_ring_rx_arg2 = NULL;
1452 
1453 	/*
1454 	 * TCP and UDP support DLS bypass. In addition TCP
1455 	 * squeue can also poll their corresponding soft rings.
1456 	 */
1457 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1458 		mac_soft_ring_dls_bypass(softring,
1459 		    mcip->mci_direct_rx_fn,
1460 		    mcip->mci_direct_rx_arg);
1461 
1462 		mrf.mrf_rx_arg = softring;
1463 		mrf.mrf_intr_handle = (mac_intr_handle_t)softring;
1464 
1465 		/*
1466 		 * Make a call in IP to get a TCP squeue assigned to
1467 		 * this softring to maintain full CPU locality through
1468 		 * the stack and allow the squeue to be able to poll
1469 		 * the softring so the flow control can be pushed
1470 		 * all the way to H/W.
1471 		 */
1472 		softring->s_ring_rx_arg2 =
1473 		    mcip->mci_resource_add((void *)mcip->mci_resource_arg,
1474 		    (mac_resource_t *)&mrf);
1475 	}
1476 
1477 	/*
1478 	 * Non-TCP protocols don't support squeues. Hence we
1479 	 * don't make any ring addition callbacks for non-TCP
1480 	 * rings. Now create the UDP softring and allow it to
1481 	 * bypass the DLS layer.
1482 	 */
1483 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1484 	    (void *)flent, (type|ST_RING_UDP), pri, mcip, mac_srs,
1485 	    cpuid, rx_func, x_arg1, x_arg2);
1486 	softring->s_ring_rx_arg2 = NULL;
1487 
1488 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1489 		mac_soft_ring_dls_bypass(softring,
1490 		    mcip->mci_direct_rx_fn,
1491 		    mcip->mci_direct_rx_arg);
1492 	}
1493 
1494 	/* Create the Oth softrings which has to go through the DLS */
1495 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1496 	    (void *)flent, (type|ST_RING_OTH), pri, mcip, mac_srs,
1497 	    cpuid, rx_func, x_arg1, x_arg2);
1498 	softring->s_ring_rx_arg2 = NULL;
1499 }
1500 
1501 /*
1502  * This routine associates a CPU or a set of CPU to process incoming
1503  * traffic from a mac client. If multiple CPUs are specified, then
1504  * so many soft rings are created with each soft ring worker thread
1505  * bound to a CPU in the set. Each soft ring in turn will be
1506  * associated with an squeue and the squeue will be moved to the
1507  * same CPU as that of the soft ring's.
1508  */
1509 static void
1510 mac_srs_fanout_modify(mac_client_impl_t *mcip, flow_entry_t *flent,
1511     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1512     mac_resource_handle_t x_arg2, mac_soft_ring_set_t *mac_rx_srs,
1513     mac_soft_ring_set_t *mac_tx_srs)
1514 {
1515 	mac_soft_ring_t *softring;
1516 	uint32_t soft_ring_flag = 0;
1517 	processorid_t cpuid = -1;
1518 	boolean_t user_specified;
1519 	int i, srings_present, new_fanout_cnt;
1520 	mac_cpus_t *srs_cpu;
1521 
1522 	user_specified = mrp->mrp_mask & MRP_CPUS_USERSPEC;
1523 	/* fanout state is REINIT. Set it back to INIT */
1524 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_REINIT);
1525 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1526 
1527 	/* how many are present right now */
1528 	srings_present = mac_rx_srs->srs_tcp_ring_count;
1529 	/* new request */
1530 	srs_cpu = &mac_rx_srs->srs_cpu;
1531 	new_fanout_cnt = srs_cpu->mc_fanout_cnt;
1532 
1533 	mutex_enter(&mac_rx_srs->srs_lock);
1534 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1535 		soft_ring_flag |= ST_RING_BW_CTL;
1536 	mutex_exit(&mac_rx_srs->srs_lock);
1537 
1538 	if (new_fanout_cnt > srings_present) {
1539 		/* soft rings increased */
1540 		mutex_enter(&mac_rx_srs->srs_lock);
1541 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1542 		mutex_exit(&mac_rx_srs->srs_lock);
1543 
1544 		for (i = mac_rx_srs->srs_tcp_ring_count;
1545 		    i < new_fanout_cnt; i++) {
1546 			/*
1547 			 * Create the protocol softrings and set the
1548 			 * DLS bypass where possible.
1549 			 */
1550 			mac_srs_create_proto_softrings(i,
1551 			    (void *)flent, soft_ring_flag,
1552 			    mac_rx_srs->srs_pri, mcip, mac_rx_srs, cpuid,
1553 			    rx_func, x_arg1, x_arg2, B_TRUE);
1554 		}
1555 		mac_srs_update_fanout_list(mac_rx_srs);
1556 	} else if (new_fanout_cnt < srings_present) {
1557 		/* soft rings decreased */
1558 		if (new_fanout_cnt == 1) {
1559 			mutex_enter(&mac_rx_srs->srs_lock);
1560 			mac_rx_srs->srs_type &= ~SRST_FANOUT_SRC_IP;
1561 			ASSERT(mac_rx_srs->srs_type & SRST_FANOUT_PROTO);
1562 			mutex_exit(&mac_rx_srs->srs_lock);
1563 		}
1564 		/* Get rid of extra soft rings */
1565 		for (i = new_fanout_cnt;
1566 		    i < mac_rx_srs->srs_tcp_ring_count; i++) {
1567 			softring = mac_rx_srs->srs_tcp_soft_rings[i];
1568 			if (softring->s_ring_rx_arg2 != NULL) {
1569 				mcip->mci_resource_remove(
1570 				    (void *)mcip->mci_resource_arg,
1571 				    softring->s_ring_rx_arg2);
1572 			}
1573 			mac_soft_ring_remove(mac_rx_srs,
1574 			    mac_rx_srs->srs_tcp_soft_rings[i]);
1575 			mac_soft_ring_remove(mac_rx_srs,
1576 			    mac_rx_srs->srs_udp_soft_rings[i]);
1577 			mac_soft_ring_remove(mac_rx_srs,
1578 			    mac_rx_srs->srs_oth_soft_rings[i]);
1579 		}
1580 		mac_srs_update_fanout_list(mac_rx_srs);
1581 	}
1582 
1583 	ASSERT(new_fanout_cnt == mac_rx_srs->srs_tcp_ring_count);
1584 	mutex_enter(&cpu_lock);
1585 	for (i = 0; i < mac_rx_srs->srs_tcp_ring_count; i++) {
1586 		cpuid = srs_cpu->mc_fanout_cpus[i];
1587 		(void) mac_soft_ring_bind(mac_rx_srs->srs_udp_soft_rings[i],
1588 		    cpuid);
1589 		(void) mac_soft_ring_bind(mac_rx_srs->srs_oth_soft_rings[i],
1590 		    cpuid);
1591 		(void) mac_soft_ring_bind(mac_rx_srs->srs_tcp_soft_rings[i],
1592 		    cpuid);
1593 		softring = mac_rx_srs->srs_tcp_soft_rings[i];
1594 		if (softring->s_ring_rx_arg2 != NULL) {
1595 			mcip->mci_resource_bind((void *)mcip->mci_resource_arg,
1596 			    softring->s_ring_rx_arg2, cpuid);
1597 		}
1598 	}
1599 
1600 	mac_srs_worker_bind(mac_rx_srs, srs_cpu->mc_pollid);
1601 	mac_srs_poll_bind(mac_rx_srs, srs_cpu->mc_workerid);
1602 
1603 	/*
1604 	 * Bind Tx srs and soft ring threads too. Let's bind tx
1605 	 * srs to the last cpu in mrp list.
1606 	 */
1607 	if (mac_tx_srs != NULL && user_specified) {
1608 		BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1609 	}
1610 	mutex_exit(&cpu_lock);
1611 }
1612 
1613 /*
1614  * Bind SRS threads and soft rings to CPUs/create fanout list.
1615  */
1616 void
1617 mac_srs_fanout_init(mac_client_impl_t *mcip, flow_entry_t *flent,
1618     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1619     mac_resource_handle_t x_arg2, mac_soft_ring_set_t *mac_rx_srs,
1620     mac_soft_ring_set_t *mac_tx_srs)
1621 {
1622 	int		i;
1623 	processorid_t	cpuid, worker_cpuid, poll_cpuid;
1624 	uint32_t	soft_ring_flag = 0;
1625 	int soft_ring_cnt;
1626 	boolean_t user_specified = B_FALSE;
1627 	mac_cpus_t *srs_cpu = &mac_rx_srs->srs_cpu;
1628 
1629 	/*
1630 	 * Remove the no soft ring flag and we will adjust it
1631 	 * appropriately further down.
1632 	 */
1633 	mutex_enter(&mac_rx_srs->srs_lock);
1634 	mac_rx_srs->srs_type &= ~SRST_NO_SOFT_RINGS;
1635 	mutex_exit(&mac_rx_srs->srs_lock);
1636 
1637 	ASSERT(mac_rx_srs->srs_soft_ring_head == NULL);
1638 
1639 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1640 		soft_ring_flag |= ST_RING_BW_CTL;
1641 
1642 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_UNINIT);
1643 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1644 	user_specified = mrp->mrp_mask & MRP_CPUS_USERSPEC;
1645 	/*
1646 	 * Ring count can be 0 if no fanout is required and no cpu
1647 	 * were specified. Leave the SRS worker and poll thread
1648 	 * unbound
1649 	 */
1650 	ASSERT(mrp != NULL);
1651 	soft_ring_cnt = srs_cpu->mc_fanout_cnt;
1652 
1653 	/* Step 1: bind cpu contains cpu list where threads need to bind */
1654 	if (soft_ring_cnt > 0) {
1655 		mutex_enter(&cpu_lock);
1656 		for (i = 0; i < soft_ring_cnt; i++) {
1657 			cpuid = srs_cpu->mc_fanout_cpus[i];
1658 			/* Create the protocol softrings */
1659 			mac_srs_create_proto_softrings(i, (void *)flent,
1660 			    soft_ring_flag, mac_rx_srs->srs_pri,
1661 			    mcip, mac_rx_srs, cpuid, rx_func,
1662 			    x_arg1, x_arg2, B_FALSE);
1663 		}
1664 		worker_cpuid = srs_cpu->mc_workerid;
1665 		poll_cpuid = srs_cpu->mc_pollid;
1666 		mac_srs_worker_bind(mac_rx_srs, worker_cpuid);
1667 		mac_srs_poll_bind(mac_rx_srs, poll_cpuid);
1668 
1669 		/*
1670 		 * Bind Tx srs and soft ring threads too.
1671 		 * Let's bind tx srs to the last cpu in
1672 		 * mrp list.
1673 		 */
1674 		if (mac_tx_srs == NULL) {
1675 			mutex_exit(&cpu_lock);
1676 			goto alldone;
1677 		}
1678 
1679 		if (user_specified) {
1680 			BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1681 		}
1682 		mutex_exit(&cpu_lock);
1683 	} else {
1684 		mutex_enter(&cpu_lock);
1685 		/*
1686 		 * For a subflow, mrp_workerid and mrp_pollid
1687 		 * is not set.
1688 		 */
1689 		mac_srs_worker_bind(mac_rx_srs, mrp->mrp_workerid);
1690 		mac_srs_poll_bind(mac_rx_srs, mrp->mrp_pollid);
1691 		mutex_exit(&cpu_lock);
1692 		goto no_softrings;
1693 	}
1694 
1695 alldone:
1696 	if (soft_ring_cnt > 1)
1697 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1698 	mac_srs_update_fanout_list(mac_rx_srs);
1699 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1700 	return;
1701 
1702 no_softrings:
1703 	if (mac_rx_srs->srs_type & SRST_FANOUT_PROTO) {
1704 		mutex_enter(&cpu_lock);
1705 		cpuid = mac_next_bind_cpu();
1706 		/* Create the protocol softrings */
1707 		mac_srs_create_proto_softrings(0, (void *)flent,
1708 		    soft_ring_flag, mac_rx_srs->srs_pri,
1709 		    mcip, mac_rx_srs, cpuid, rx_func,
1710 		    x_arg1, x_arg2, B_FALSE);
1711 		mutex_exit(&cpu_lock);
1712 	} else {
1713 		/*
1714 		 * This is the case when there is no fanout which is
1715 		 * true for subflows.
1716 		 */
1717 		mac_rx_srs->srs_type |= SRST_NO_SOFT_RINGS;
1718 	}
1719 	mac_srs_update_fanout_list(mac_rx_srs);
1720 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1721 }
1722 
1723 /*
1724  * mac_fanout_setup:
1725  *
1726  * Calls mac_srs_fanout_init() or modify() depending upon whether
1727  * the SRS is getting initialized or re-initialized.
1728  */
1729 void
1730 mac_fanout_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
1731     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1732     mac_resource_handle_t x_arg2)
1733 {
1734 	mac_soft_ring_set_t *mac_rx_srs, *mac_tx_srs;
1735 	int i, rx_srs_cnt;
1736 
1737 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
1738 	/*
1739 	 * This is an aggregation port. Fanout will be setup
1740 	 * over the aggregation itself.
1741 	 */
1742 	if (mcip->mci_state_flags & MCIS_IS_AGGR_PORT)
1743 		return;
1744 
1745 	mac_rx_srs = flent->fe_rx_srs[0];
1746 	/*
1747 	 * Set up the fanout on the tx side only once, with the
1748 	 * first rx SRS. The CPU binding, fanout, and bandwidth
1749 	 * criteria are common to both RX and TX, so
1750 	 * initializing them along side avoids redundant code.
1751 	 */
1752 	mac_tx_srs = flent->fe_tx_srs;
1753 	rx_srs_cnt = flent->fe_rx_srs_cnt;
1754 
1755 	/* No fanout for subflows */
1756 	if (flent->fe_type & FLOW_USER) {
1757 		mac_srs_fanout_init(mcip, flent, mrp, rx_func,
1758 		    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1759 		return;
1760 	}
1761 
1762 	mac_flow_cpu_init(flent, mrp);
1763 
1764 	/*
1765 	 * Set up fanout for both SW (0th SRS) and HW classified
1766 	 * SRS (the rest of Rx SRSs in flent).
1767 	 */
1768 	for (i = 0; i < rx_srs_cnt; i++) {
1769 		mac_rx_srs = flent->fe_rx_srs[i];
1770 		if (i != 0)
1771 			mac_tx_srs = NULL;
1772 		switch (mac_rx_srs->srs_fanout_state) {
1773 		case SRS_FANOUT_UNINIT:
1774 			mac_srs_fanout_init(mcip, flent, mrp, rx_func,
1775 			    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1776 			break;
1777 		case SRS_FANOUT_INIT:
1778 			break;
1779 		case SRS_FANOUT_REINIT:
1780 			mac_rx_srs_quiesce(mac_rx_srs, SRS_QUIESCE);
1781 			mac_srs_fanout_modify(mcip, flent, mrp, rx_func,
1782 			    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1783 			mac_rx_srs_restart(mac_rx_srs);
1784 			break;
1785 		default:
1786 			VERIFY(mac_rx_srs->srs_fanout_state <=
1787 			    SRS_FANOUT_REINIT);
1788 			break;
1789 		}
1790 	}
1791 }
1792 
1793 /*
1794  * mac_create_soft_ring_set:
1795  *
1796  * Create a mac_soft_ring_set_t (SRS). If soft_ring_fanout_type is
1797  * SRST_TX, an SRS for Tx side is created. Otherwise an SRS for Rx side
1798  * processing is created.
1799  *
1800  * Details on Rx SRS:
1801  * Create a SRS and also add the necessary soft rings for TCP and
1802  * non-TCP based on fanout type and count specified.
1803  *
1804  * mac_soft_ring_fanout, mac_srs_fanout_modify (?),
1805  * mac_soft_ring_stop_workers, mac_soft_ring_set_destroy, etc need
1806  * to be heavily modified.
1807  *
1808  * mi_soft_ring_list_size, mi_soft_ring_size, etc need to disappear.
1809  */
1810 mac_soft_ring_set_t *
1811 mac_srs_create(mac_client_impl_t *mcip, flow_entry_t *flent, uint32_t srs_type,
1812     mac_direct_rx_t rx_func, void *x_arg1, mac_resource_handle_t x_arg2,
1813     mac_ring_t *ring)
1814 {
1815 	mac_soft_ring_set_t 	*mac_srs;
1816 	mac_srs_rx_t		*srs_rx;
1817 	mac_srs_tx_t		*srs_tx;
1818 	mac_bw_ctl_t		*mac_bw;
1819 	mac_resource_props_t	*mrp;
1820 	boolean_t		is_tx_srs = ((srs_type & SRST_TX) != 0);
1821 
1822 	mac_srs = kmem_cache_alloc(mac_srs_cache, KM_SLEEP);
1823 	bzero(mac_srs, sizeof (mac_soft_ring_set_t));
1824 	srs_rx = &mac_srs->srs_rx;
1825 	srs_tx = &mac_srs->srs_tx;
1826 
1827 	mutex_enter(&flent->fe_lock);
1828 
1829 	/*
1830 	 * Get the bandwidth control structure from the flent. Get
1831 	 * rid of any residual values in the control structure for
1832 	 * the tx bw struct and also for the rx, if the rx srs is
1833 	 * the 1st one being brought up (the rx bw ctl struct may
1834 	 * be shared by multiple SRSs)
1835 	 */
1836 	if (is_tx_srs) {
1837 		mac_srs->srs_bw = &flent->fe_tx_bw;
1838 		bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
1839 		flent->fe_tx_srs = mac_srs;
1840 	} else {
1841 		/*
1842 		 * The bw counter (stored in the flent) is shared
1843 		 * by SRS's within an rx group.
1844 		 */
1845 		mac_srs->srs_bw = &flent->fe_rx_bw;
1846 		/* First rx SRS, clear the bw structure */
1847 		if (flent->fe_rx_srs_cnt == 0)
1848 			bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
1849 
1850 		/*
1851 		 * It is better to panic here rather than just assert because
1852 		 * on a non-debug kernel we might end up courrupting memory
1853 		 * and making it difficult to debug.
1854 		 */
1855 		if (flent->fe_rx_srs_cnt >= MAX_RINGS_PER_GROUP) {
1856 			panic("Array Overrun detected due to MAC client %p "
1857 			    " having more rings than %d", (void *)mcip,
1858 			    MAX_RINGS_PER_GROUP);
1859 		}
1860 		flent->fe_rx_srs[flent->fe_rx_srs_cnt] = mac_srs;
1861 		flent->fe_rx_srs_cnt++;
1862 	}
1863 	mac_srs->srs_flent = flent;
1864 	mutex_exit(&flent->fe_lock);
1865 
1866 	mac_srs->srs_state = 0;
1867 	mac_srs->srs_type = (srs_type | SRST_NO_SOFT_RINGS);
1868 	mac_srs->srs_worker_cpuid = mac_srs->srs_worker_cpuid_save = -1;
1869 	mac_srs->srs_poll_cpuid = mac_srs->srs_poll_cpuid_save = -1;
1870 	mac_srs_fanout_list_alloc(mac_srs);
1871 
1872 	/*
1873 	 * For a flow we use the underlying MAC client's priority range with
1874 	 * the priority value to find an absolute priority value. For a MAC
1875 	 * client we use the MAC client's maximum priority as the value.
1876 	 */
1877 	mrp = &flent->fe_effective_props;
1878 	if ((mac_srs->srs_type & SRST_FLOW) != 0) {
1879 		mac_srs->srs_pri = FLOW_PRIORITY(mcip->mci_min_pri,
1880 		    mcip->mci_max_pri, mrp->mrp_priority);
1881 	} else {
1882 		mac_srs->srs_pri = mcip->mci_max_pri;
1883 	}
1884 	mac_srs->srs_mcip = mcip;
1885 	/*
1886 	 * We need to insert the SRS in the global list before
1887 	 * binding the SRS and SR threads. Otherwise there is a
1888 	 * is a small window where the cpu reconfig callbacks
1889 	 * may miss the SRS in the list walk and DR could fail
1890 	 * as there are bound threads.
1891 	 */
1892 	mac_srs_add_glist(mac_srs);
1893 
1894 	/* Initialize bw limit */
1895 	if ((mrp->mrp_mask & MRP_MAXBW) != 0) {
1896 		mac_srs->srs_drain_func = mac_rx_srs_drain_bw;
1897 
1898 		mac_bw = mac_srs->srs_bw;
1899 		mutex_enter(&mac_bw->mac_bw_lock);
1900 		mac_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1901 
1902 		/*
1903 		 * Give twice the queuing capability before
1904 		 * dropping packets. The unit is bytes/tick.
1905 		 */
1906 		mac_bw->mac_bw_drop_threshold = mac_bw->mac_bw_limit << 1;
1907 		mutex_exit(&mac_bw->mac_bw_lock);
1908 		mac_srs->srs_type |= SRST_BW_CONTROL;
1909 	} else {
1910 		mac_srs->srs_drain_func = mac_rx_srs_drain;
1911 	}
1912 
1913 	/*
1914 	 * We use the following policy to control Receive
1915 	 * Side Dynamic Polling:
1916 	 * 1) We switch to poll mode anytime the processing thread causes
1917 	 *    a backlog to build up in SRS and its associated Soft Rings
1918 	 *    (sr_poll_pkt_cnt > 0).
1919 	 * 2) As long as the backlog stays under the low water mark
1920 	 *    (sr_lowat), we poll the H/W for more packets.
1921 	 * 3) If the backlog (sr_poll_pkt_cnt) exceeds low water mark, we
1922 	 *    stay in poll mode but don't poll the H/W for more packets.
1923 	 * 4) Anytime in polling mode, if we poll the H/W for packets and
1924 	 *    find nothing plus we have an existing backlog
1925 	 *    (sr_poll_pkt_cnt > 0), we stay in polling mode but don't poll
1926 	 *    the H/W for packets anymore (let the polling thread go to sleep).
1927 	 * 5) Once the backlog is relived (packets are processed) we reenable
1928 	 *    polling (by signalling the poll thread) only when the backlog
1929 	 *    dips below sr_poll_thres.
1930 	 * 6) sr_hiwat is used exclusively when we are not polling capable
1931 	 *    and is used to decide when to drop packets so the SRS queue
1932 	 *    length doesn't grow infinitely.
1933 	 */
1934 	if (!is_tx_srs) {
1935 		srs_rx->sr_hiwat = mac_soft_ring_max_q_cnt;
1936 		/* Low water mark needs to be less than high water mark */
1937 		srs_rx->sr_lowat = mac_soft_ring_min_q_cnt <=
1938 		    mac_soft_ring_max_q_cnt ? mac_soft_ring_min_q_cnt :
1939 		    (mac_soft_ring_max_q_cnt >> 2);
1940 		/* Poll threshold need to be half of low water mark or less */
1941 		srs_rx->sr_poll_thres = mac_soft_ring_poll_thres <=
1942 		    (srs_rx->sr_lowat >> 1) ? mac_soft_ring_poll_thres :
1943 		    (srs_rx->sr_lowat >> 1);
1944 		if (mac_latency_optimize)
1945 			mac_srs->srs_state |= SRS_LATENCY_OPT;
1946 		else
1947 			mac_srs->srs_state |= SRS_SOFTRING_QUEUE;
1948 	}
1949 
1950 	mac_srs->srs_worker = thread_create(NULL, 0,
1951 	    mac_srs_worker, mac_srs, 0, &p0, TS_RUN, mac_srs->srs_pri);
1952 
1953 	if (is_tx_srs) {
1954 		/* Handle everything about Tx SRS and return */
1955 		mac_srs->srs_drain_func = mac_tx_srs_drain;
1956 		srs_tx->st_max_q_cnt = mac_tx_srs_max_q_cnt;
1957 		srs_tx->st_hiwat =
1958 		    (mac_tx_srs_hiwat > mac_tx_srs_max_q_cnt) ?
1959 		    mac_tx_srs_max_q_cnt : mac_tx_srs_hiwat;
1960 		srs_tx->st_arg1 = x_arg1;
1961 		srs_tx->st_arg2 = x_arg2;
1962 		return (mac_srs);
1963 	}
1964 
1965 	if ((srs_type & SRST_FLOW) != 0 ||
1966 	    FLOW_TAB_EMPTY(mcip->mci_subflow_tab))
1967 		srs_rx->sr_lower_proc = mac_rx_srs_process;
1968 	else
1969 		srs_rx->sr_lower_proc = mac_rx_srs_subflow_process;
1970 
1971 	srs_rx->sr_func = rx_func;
1972 	srs_rx->sr_arg1 = x_arg1;
1973 	srs_rx->sr_arg2 = x_arg2;
1974 
1975 	if (ring != NULL) {
1976 		/* Is the mac_srs created over the RX default group? */
1977 		if (ring->mr_gh == (mac_group_handle_t)
1978 		    (&mcip->mci_mip->mi_rx_groups[0]))
1979 			mac_srs->srs_type |= SRST_DEFAULT_GRP;
1980 
1981 		mac_srs->srs_ring = ring;
1982 		ring->mr_srs = mac_srs;
1983 		ring->mr_classify_type = MAC_HW_CLASSIFIER;
1984 		ring->mr_flag |= MR_INCIPIENT;
1985 
1986 		if (!(mcip->mci_mip->mi_state_flags & MIS_POLL_DISABLE) &&
1987 		    FLOW_TAB_EMPTY(mcip->mci_subflow_tab) && mac_poll_enable)
1988 			mac_srs->srs_state |= SRS_POLLING_CAPAB;
1989 
1990 		mac_srs->srs_poll_thr = thread_create(NULL, 0,
1991 		    mac_rx_srs_poll_ring, mac_srs, 0, &p0, TS_RUN,
1992 		    mac_srs->srs_pri);
1993 		/*
1994 		 * Some drivers require serialization and don't send
1995 		 * packet chains in interrupt context. For such
1996 		 * drivers, we should always queue in soft ring
1997 		 * so that we get a chance to switch into a polling
1998 		 * mode under backlog.
1999 		 */
2000 		if (mcip->mci_mip->mi_v12n_level & MAC_VIRT_SERIALIZE)
2001 			mac_srs->srs_state |= SRS_SOFTRING_QUEUE;
2002 	}
2003 	return (mac_srs);
2004 }
2005 
2006 /*
2007  * Figure out the number of soft rings required. Its dependant on
2008  * if protocol fanout is required (for LINKs), global settings
2009  * require us to do fanout for performance (based on mac_soft_ring_enable),
2010  * or user has specifically requested fanout.
2011  */
2012 static uint32_t
2013 mac_find_fanout(flow_entry_t *flent, uint32_t link_type)
2014 {
2015 	uint32_t			fanout_type;
2016 	mac_resource_props_t		*mrp = &flent->fe_effective_props;
2017 
2018 	/* no fanout for subflows */
2019 	switch (link_type) {
2020 	case SRST_FLOW:
2021 		fanout_type = SRST_NO_SOFT_RINGS;
2022 		break;
2023 	case SRST_LINK:
2024 		fanout_type = SRST_FANOUT_PROTO;
2025 		break;
2026 	}
2027 
2028 	/* A primary NIC/link is being plumbed */
2029 	if (flent->fe_type & FLOW_PRIMARY_MAC) {
2030 		if (mac_soft_ring_enable && mac_rx_soft_ring_count > 1) {
2031 			fanout_type |= SRST_FANOUT_SRC_IP;
2032 		}
2033 	} else if (flent->fe_type & FLOW_VNIC) {
2034 		/* A VNIC is being created */
2035 		if (mrp != NULL && mrp->mrp_ncpus > 0) {
2036 			fanout_type |= SRST_FANOUT_SRC_IP;
2037 		}
2038 	}
2039 
2040 	return (fanout_type);
2041 }
2042 
2043 /*
2044  * Change a group from h/w to s/w classification.
2045  */
2046 static void
2047 mac_rx_switch_grp_to_sw(mac_group_t *group)
2048 {
2049 	mac_ring_t		*ring;
2050 	mac_soft_ring_set_t	*mac_srs;
2051 
2052 	for (ring = group->mrg_rings; ring != NULL; ring = ring->mr_next) {
2053 		if (ring->mr_classify_type == MAC_HW_CLASSIFIER) {
2054 			/*
2055 			 * Remove the SRS associated with the HW ring.
2056 			 * As a result, polling will be disabled.
2057 			 */
2058 			mac_srs = ring->mr_srs;
2059 			ASSERT(mac_srs != NULL);
2060 			mac_rx_srs_remove(mac_srs);
2061 			ring->mr_srs = NULL;
2062 		}
2063 
2064 		if (ring->mr_state != MR_INUSE)
2065 			(void) mac_start_ring(ring);
2066 		/*
2067 		 * We need to perform SW classification
2068 		 * for packets landing in these rings
2069 		 */
2070 		ring->mr_state = MR_INUSE;
2071 		ring->mr_flag = 0;
2072 		ring->mr_classify_type = MAC_SW_CLASSIFIER;
2073 	}
2074 }
2075 
2076 /*
2077  * Create the Rx SRS for S/W classifier and for each ring in the
2078  * group (if exclusive group). Also create the Tx SRS.
2079  */
2080 void
2081 mac_srs_group_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2082     mac_group_t *group, uint32_t link_type)
2083 {
2084 	mac_impl_t		*mip = mcip->mci_mip;
2085 	mac_soft_ring_set_t	*mac_srs;
2086 	mac_soft_ring_set_t	*tx_srs = NULL;
2087 	mac_ring_t 		*ring;
2088 	uint32_t		fanout_type;
2089 	boolean_t		created_srs = B_FALSE;
2090 
2091 	fanout_type = mac_find_fanout(flent, link_type);
2092 
2093 	/* Create the SRS for S/W classification if none exists */
2094 	if (flent->fe_rx_srs[0] == NULL) {
2095 		ASSERT(flent->fe_rx_srs_cnt == 0);
2096 		/* Setup the Rx SRS */
2097 		mac_srs = mac_srs_create(mcip, flent, fanout_type | link_type,
2098 		    mac_rx_deliver, mcip, NULL, NULL);
2099 
2100 		mutex_enter(&flent->fe_lock);
2101 		flent->fe_cb_fn = (flow_fn_t)mac_srs->srs_rx.sr_lower_proc;
2102 		flent->fe_cb_arg1 = (void *)mip;
2103 		flent->fe_cb_arg2 = (void *)mac_srs;
2104 		mutex_exit(&flent->fe_lock);
2105 
2106 		/* Setup the Tx SRS as well */
2107 		ASSERT(flent->fe_tx_srs == NULL);
2108 		tx_srs = mac_srs_create(mcip, flent, SRST_TX | link_type,
2109 		    NULL, mcip, NULL, NULL);
2110 
2111 		if (mcip->mci_share != NULL) {
2112 			mac_srs_tx_t	*tx = &tx_srs->srs_tx;
2113 			ASSERT((mcip->mci_state_flags & MCIS_NO_HWRINGS) == 0);
2114 			/*
2115 			 * A share requires a dedicated TX group.
2116 			 * mac_reserve_tx_group() does the work needed to
2117 			 * allocate a new group and populate that group
2118 			 * with rings according to the driver requirements
2119 			 * and limitations.
2120 			 */
2121 			tx->st_group =
2122 			    mac_reserve_tx_group(mip, mcip->mci_share);
2123 			ASSERT(tx->st_group != NULL);
2124 			tx->st_group->mrg_tx_client = mcip;
2125 		}
2126 		mac_tx_srs_setup(mcip, flent, link_type);
2127 		created_srs = B_TRUE;
2128 	}
2129 
2130 	if (group == NULL) {
2131 		if (created_srs) {
2132 			mac_fanout_setup(mcip, flent,
2133 			    MCIP_RESOURCE_PROPS(mcip), mac_rx_deliver,
2134 			    mcip, NULL);
2135 		}
2136 		return;
2137 	}
2138 
2139 	/*
2140 	 * fanout for default SRS is done when default SRS are created
2141 	 * above. As each ring is added to the group, we setup the
2142 	 * SRS and fanout to it.
2143 	 */
2144 	switch (group->mrg_state) {
2145 	case MAC_GROUP_STATE_RESERVED:
2146 		/*
2147 		 * The group is exclusively ours. Create a SRS
2148 		 * for each ring in the group and allow the
2149 		 * individual SRS to dynamically poll their
2150 		 * Rx ring. Do this only if the  client is not
2151 		 * a VLAN MAC client since for VLAN we do
2152 		 * s/w classification for the VID check.
2153 		 */
2154 		if (i_mac_flow_vid(mcip->mci_flent) != VLAN_ID_NONE)
2155 			break;
2156 		for (ring = group->mrg_rings; ring != NULL;
2157 		    ring = ring->mr_next) {
2158 			switch (ring->mr_state) {
2159 			case MR_INUSE:
2160 			case MR_FREE:
2161 				if (ring->mr_srs != NULL)
2162 					break;
2163 				if (ring->mr_state != MR_INUSE)
2164 					(void) mac_start_ring(ring);
2165 
2166 				ring->mr_state = MR_INUSE;
2167 
2168 				mac_srs = mac_srs_create(mcip, flent,
2169 				    fanout_type | link_type,
2170 				    mac_rx_deliver, mcip, NULL, ring);
2171 				break;
2172 			default:
2173 				cmn_err(CE_PANIC, "srs_setup: mcip = %p "
2174 				    "trying to add UNKNOWN ring = %p\n",
2175 				    (void *)mcip, (void *)ring);
2176 				break;
2177 			}
2178 		}
2179 		break;
2180 	case MAC_GROUP_STATE_SHARED:
2181 		/*
2182 		 * Set all rings of this group to software classified.
2183 		 *
2184 		 * If the group is current RESERVED, the existing mac client
2185 		 * (the only client on this group) is using this group
2186 		 * exclusively.  In that case we need to disable polling on
2187 		 * the rings of the group (if it was enabled), and free the
2188 		 * SRS associated with the rings.
2189 		 */
2190 		mac_rx_switch_grp_to_sw(group);
2191 		break;
2192 	default:
2193 		ASSERT(B_FALSE);
2194 		break;
2195 	}
2196 	mac_fanout_setup(mcip, flent, MCIP_RESOURCE_PROPS(mcip),
2197 	    mac_rx_deliver, mcip, NULL);
2198 }
2199 
2200 void
2201 mac_srs_group_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
2202     uint32_t link_type)
2203 {
2204 	mac_soft_ring_set_t	*mac_srs;
2205 	mac_soft_ring_set_t	*tx_srs;
2206 	mac_srs_tx_t		*tx;
2207 	int			i;
2208 
2209 	for (i = 0; i < flent->fe_rx_srs_cnt; i++) {
2210 		mac_srs = flent->fe_rx_srs[i];
2211 		mac_rx_srs_quiesce(mac_srs, SRS_CONDEMNED);
2212 		/*
2213 		 * Deal with all fanout tear down etc.
2214 		 */
2215 		mac_srs_free(mac_srs);
2216 		flent->fe_rx_srs[i] = NULL;
2217 	}
2218 	flent->fe_rx_srs_cnt = 0;
2219 
2220 	tx_srs = flent->fe_tx_srs;
2221 	tx = &tx_srs->srs_tx;
2222 	switch (link_type) {
2223 	case SRST_FLOW:
2224 		/*
2225 		 * For flows, we need to work with passed
2226 		 * flent to find the Rx/Tx SRS.
2227 		 */
2228 		mac_tx_srs_quiesce(tx_srs, SRS_CONDEMNED);
2229 		break;
2230 	case SRST_LINK:
2231 		mac_tx_client_quiesce(mcip, SRS_CONDEMNED);
2232 		/*
2233 		 * Release the TX resources. First the TX group, if any
2234 		 * was assigned to the MAC client, which will cause the
2235 		 * TX rings to be moved back to the pool. Then free the
2236 		 * rings themselves.
2237 		 */
2238 		if (tx->st_group != NULL) {
2239 			mac_release_tx_group(tx_srs->srs_mcip->mci_mip,
2240 			    tx->st_group);
2241 			tx->st_group = NULL;
2242 		}
2243 		if (tx->st_ring_count != 0) {
2244 			kmem_free(tx->st_rings,
2245 			    sizeof (mac_ring_handle_t) * tx->st_ring_count);
2246 		}
2247 		if (tx->st_arg2 != NULL) {
2248 			ASSERT(tx_srs->srs_type & SRST_TX);
2249 			mac_release_tx_ring(tx->st_arg2);
2250 		}
2251 		break;
2252 	default:
2253 		ASSERT(B_FALSE);
2254 		break;
2255 	}
2256 	mac_srs_free(tx_srs);
2257 	flent->fe_tx_srs = NULL;
2258 }
2259 
2260 /*
2261  * This is the group state machine. The state of an Rx group is given by
2262  * the following table. The default group and its rings are started in
2263  * mac_start itself and the default group stays in SHARED state until
2264  * mac_stop at which time the group and rings are stopped and and it
2265  * reverts to the Registered state.
2266  *
2267  * Typically this function is called on a group after adding or removing a
2268  * client from it, to find out what should be the new state of the group.
2269  * If the new state is RESERVED, then the client that owns this group
2270  * exclusively is also returned. Note that adding or removing a client from
2271  * a group could also impact the default group and the caller needs to
2272  * evaluate the effect on the default group.
2273  *
2274  * Group type		# of clients	mi_nactiveclients	Group State
2275  *			in the group
2276  *
2277  * Non-default		0		N.A.			REGISTERED
2278  * Non-default		1		N.A.			RESERVED
2279  * Non-default		> 1		N.A.			SHARED
2280  *
2281  * Default		0		N.A.			SHARED
2282  * Default		1		1			RESERVED
2283  * Default		1		> 1			SHARED
2284  * Default		> 1		N.A.			SHARED
2285  */
2286 mac_group_state_t
2287 mac_rx_group_next_state(mac_group_t *grp, mac_client_impl_t **group_only_mcip)
2288 {
2289 	mac_impl_t		*mip = (mac_impl_t *)grp->mrg_mh;
2290 
2291 	*group_only_mcip = NULL;
2292 
2293 	/* Non-default group */
2294 
2295 	if (grp != mip->mi_rx_groups) {
2296 		if (MAC_RX_GROUP_NO_CLIENT(grp))
2297 			return (MAC_GROUP_STATE_REGISTERED);
2298 
2299 		*group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(grp);
2300 		if (*group_only_mcip != NULL)
2301 			return (MAC_GROUP_STATE_RESERVED);
2302 
2303 		return (MAC_GROUP_STATE_SHARED);
2304 	}
2305 
2306 	/* Default group */
2307 
2308 	if (MAC_RX_GROUP_NO_CLIENT(grp) || mip->mi_nactiveclients != 1)
2309 		return (MAC_GROUP_STATE_SHARED);
2310 
2311 	*group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(grp);
2312 	ASSERT(*group_only_mcip != NULL);
2313 	return (MAC_GROUP_STATE_RESERVED);
2314 }
2315 
2316 /*
2317  * OVERVIEW NOTES FOR DATAPATH
2318  * ===========================
2319  *
2320  * Create an SRS and setup the corresponding flow function and args.
2321  * Add a classification rule for the flow specified by 'flent' and program
2322  * the hardware classifier when applicable.
2323  *
2324  * Rx ring assignment, SRS, polling and B/W enforcement
2325  * ----------------------------------------------------
2326  *
2327  * We try to use H/W classification on NIC and assign traffic to a
2328  * MAC address to a particular Rx ring. There is a 1-1 mapping
2329  * between a SRS and a Rx ring. The SRS (short for soft ring set)
2330  * dynamically switches the underlying Rx ring between interrupt
2331  * and polling mode and enforces any specified B/W control.
2332  *
2333  * There is always a SRS created and tied to each H/W and S/W rule.
2334  * Whenever we create a H/W rule, we always add the the same rule to
2335  * S/W classifier and tie a SRS to it.
2336  *
2337  * In case a B/W control is specified, its broken into bytes
2338  * per ticks and as soon as the quota for a tick is exhausted,
2339  * the underlying Rx ring is forced into poll mode for remianing
2340  * tick. The SRS poll thread only polls for bytes that are
2341  * allowed to come in the SRS. We typically let 4x the configured
2342  * B/W worth of packets to come in the SRS (to prevent unnecessary
2343  * drops due to bursts) but only process the specified amount.
2344  *
2345  * A Link (primary NIC, VNIC, VLAN or aggr) can have 1 or more
2346  * Rx rings (and corresponding SRSs) assigned to it. The SRS
2347  * in turn can have softrings to do protocol level fanout or
2348  * softrings to do S/W based fanout or both. In case the NIC
2349  * has no Rx rings, we do S/W classification to respective SRS.
2350  * The S/W classification rule is always setup and ready. This
2351  * allows the MAC layer to reassign Rx rings whenever needed
2352  * but packets still continue to flow via the default path and
2353  * getting S/W classified to correct SRS.
2354  *
2355  * In other cases where a NIC or VNIC is plumbed, our goal is use
2356  * H/W classifier and get two Rx ring assigned for the Link. One
2357  * for TCP and one for UDP|SCTP. The respective SRS still do the
2358  * polling on the Rx ring. For Link that is plumbed for IP, there
2359  * is a TCP squeue which also does polling and can control the
2360  * the Rx ring directly (where SRS is just pass through). For
2361  * the following cases, the SRS does the polling underneath.
2362  * 1) non IP based Links (Links which are not plumbed via ifconfig)
2363  *    and paths which have no IP squeues (UDP & SCTP)
2364  * 2) If B/W control is specified on the Link
2365  * 3) If S/W fanout is secified
2366  *
2367  * Note1: As of current implementation, we try to assign only 1 Rx
2368  * ring per Link and more than 1 Rx ring for primary Link for
2369  * H/W based fanout. We always create following softrings per SRS:
2370  * 1) TCP softring which is polled by TCP squeue where possible
2371  *    (and also bypasses DLS)
2372  * 2) UDP/SCTP based which bypasses DLS
2373  * 3) OTH softring which goes via DLS (currently deal with IPv6
2374  *    and non TCP/UDP/SCTP for IPv4 packets).
2375  *
2376  * It is necessary to create 3 softrings since SRS has to poll
2377  * the single Rx ring underneath and enforce any link level B/W
2378  * control (we can't switch the Rx ring in poll mode just based
2379  * on TCP squeue if the same Rx ring is sharing UDP and other
2380  * traffic as well). Once polling is done and any Link level B/W
2381  * control is specified, the packets are assigned to respective
2382  * softring based on protocol. Since TCP has IP based squeue
2383  * which benefits by polling, we separate TCP packets into
2384  * its own softring which can be polled by IP squeue. We need
2385  * to separate out UDP/SCTP to UDP softring since it can bypass
2386  * the DLS layer which has heavy performance advanatges and we
2387  * need a softring (OTH) for rest.
2388  *
2389  * ToDo: The 3 softrings for protocol are needed only till we can
2390  * get rid of DLS from datapath, make IPv4 and IPv6 paths
2391  * symmetric (deal with mac_header_info for v6 and polling for
2392  * IPv4 TCP - ip_accept_tcp is IPv4 specific although squeues
2393  * are generic), and bring SAP based classification to MAC layer
2394  *
2395  * H/W and S/W based fanout and multiple Rx rings per Link
2396  * -------------------------------------------------------
2397  *
2398  * In case, fanout is requested (or determined automatically based
2399  * on Link speed and processor speed), we try to assign multiple
2400  * Rx rings per Link with their respective SRS. In this case
2401  * the NIC should be capable of fanning out incoming packets between
2402  * the assigned Rx rings (H/W based fanout). All the SRS
2403  * individually switch their Rx ring between interrupt and polling
2404  * mode but share a common B/W control counter in case of Link
2405  * level B/W is specified.
2406  *
2407  * If S/W based fanout is specified in lieu of H/W based fanout,
2408  * the Link SRS creates the specified number of softrings for
2409  * each protocol (TCP, UDP, OTH). Incoming packets are fanned
2410  * out to the correct softring based on their protocol and
2411  * protocol specific hash function.
2412  *
2413  * Primary and non primary MAC clients
2414  * -----------------------------------
2415  *
2416  * The NICs, VNICs, Vlans, and Aggrs are typically termed as Links
2417  * and are a Layer 2 construct.
2418  *
2419  * Primary NIC:
2420  *	The Link that owns the primary MAC address and typically
2421  *	is used as the data NIC in non virtualized cases. As such
2422  *	H/W resources are preferntially given to primary NIC. As
2423  *	far as code is concerned, there is no difference in the
2424  *	primary NIC vs VNICs. They are all treated as Links.
2425  *	At the very first call to mac_unicast_add() we program the S/W
2426  *	classifier for the primary MAC address, get a soft ring set
2427  *	(and soft rings based on 'ip_soft_ring_cnt')
2428  *	and a Rx ring assigned for polling to get enabled.
2429  *	When IP get plumbed and negotiates polling, we can
2430  *	let squeue do the polling on TCP softring.
2431  *
2432  * VNICs:
2433  *	Same as any other Link. As long as the H/W resource assignments
2434  *	are equal, the data path and setup for all Links is same.
2435  *
2436  * Flows:
2437  *	Can be configured on Links. They have their own SRS and the
2438  *	S/W classifier is programmed appropriately based on the flow.
2439  *	The flows typically deal with layer 3 and above and
2440  *	creates a soft ring set specific to the flow. The receive
2441  *	side function is switched from mac_rx_srs_process to
2442  *	mac_rx_srs_subflow_process which first tries to assign the
2443  *	packet to appropriate flow SRS and failing which assigns it
2444  *	to link SRS. This allows us to avoid the layered approach
2445  *	which gets complex.
2446  *
2447  * By the time mac_datapath_setup() completes, we already have the
2448  * soft rings set, Rx rings, soft rings, etc figured out and both H/W
2449  * and S/W classifiers programmed. IP is not plumbed yet (and might
2450  * never be for Virtual Machines guest OS path). When IP is plumbed
2451  * (for both NIC and VNIC), we do a capability negotiation for polling
2452  * and upcall functions etc.
2453  *
2454  * Rx ring Assignement NOTES
2455  * -------------------------
2456  *
2457  * For NICs which have only 1 Rx ring (we treat  NICs with no Rx rings
2458  * as NIC with a single default ring), we assign the only ring to
2459  * primary Link as MAC_RX_HW_DEFAULT_RING. The primary Link SRS can do
2460  * polling on it as long as it is the only link in use and we compare
2461  * the MAC address for unicast packets before accepting an incoming
2462  * packet (there is no need for S/W classification in this case). We
2463  * disable polling on the only ring the moment 2nd link gets created
2464  * (the polling remains enabled even though there are broadcast and
2465  * multicast flows created).
2466  *
2467  * If the NIC has more than 1 Rx ring, we assign the default ring (the
2468  * 1st ring) to deal with broadcast, multicast and traffic for other
2469  * NICs which needs S/W classification. We assign the primary mac
2470  * addresses to another ring by specifiying a classification rule for
2471  * primary unicast MAC address to the selected ring. The primary Link
2472  * (and its SRS) can continue to poll the assigned Rx ring at all times
2473  * independantly.
2474  *
2475  * Right now we just assign MAC_RX_HW_DEFAULT_RING to note that it is
2476  * primary NIC and later we will check to see how many Rx rings we
2477  * have and can we get a non default Rx ring for the primary MAC.
2478  *
2479  * Note: In future, if no fanout is specified, we try to assign 2 Rx
2480  * rings for the primary Link with the primary MAC address + TCP going
2481  * to one ring and primary MAC address + UDP|SCTP going to other ring.
2482  * Any remaining traffic for primary MAC address can go to the default
2483  * Rx ring and get S/W classified. This way the respective SRSs don't
2484  * need to do proto fanout and don't need to have softrings at all and
2485  * can poll their respective Rx rings.
2486  *
2487  * As an optimization, when a new NIC or VNIC is created, we can get
2488  * only one Rx ring and make it a TCP specific Rx ring and use the
2489  * H/W default Rx ring for the rest (this Rx ring is never polled).
2490  */
2491 int
2492 mac_datapath_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2493     uint32_t link_type)
2494 {
2495 	mac_impl_t		*mip = mcip->mci_mip;
2496 	mac_group_t		*group = NULL;
2497 	mac_group_t		*default_group;
2498 	int			err;
2499 	uint8_t 		*mac_addr;
2500 	mac_rx_group_reserve_type_t	rtype = MAC_RX_RESERVE_NONDEFAULT;
2501 	mac_group_state_t	next_state;
2502 	mac_client_impl_t	*group_only_mcip;
2503 
2504 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
2505 
2506 	switch (link_type) {
2507 	case SRST_FLOW:
2508 		mac_srs_group_setup(mcip, flent, NULL, link_type);
2509 		return (0);
2510 
2511 	case SRST_LINK:
2512 		mac_addr = flent->fe_flow_desc.fd_dst_mac;
2513 
2514 		/* Check if we need to reserve the default group */
2515 		if (flent->fe_type & FLOW_PRIMARY_MAC)
2516 			rtype = MAC_RX_RESERVE_DEFAULT;
2517 
2518 		if ((mcip->mci_state_flags & MCIS_NO_HWRINGS) == 0) {
2519 			/*
2520 			 * Check to see if we can get an exclusive group for
2521 			 * this mac address or if there already exists a
2522 			 * group that has this mac address (case of VLANs).
2523 			 * If no groups are available, use the default group.
2524 			 */
2525 			group = mac_reserve_rx_group(mcip, mac_addr, rtype);
2526 		}
2527 
2528 		if (group == NULL) {
2529 			if ((mcip->mci_state_flags & MCIS_REQ_HWRINGS) != 0)
2530 				return (ENOSPC);
2531 			group = &mip->mi_rx_groups[0];
2532 		}
2533 
2534 		/*
2535 		 * Some NICs don't support any Rx rings, so there may not
2536 		 * even be a default group.
2537 		 */
2538 		if (group != NULL) {
2539 			flent->fe_rx_ring_group = group;
2540 			/*
2541 			 * Add the client to the group. This could cause
2542 			 * either this group to move to the shared state or
2543 			 * cause the default group to move to the shared state.
2544 			 * The actions on this group are done here, while the
2545 			 * actions on the default group are postponed to
2546 			 * the end of this function.
2547 			 */
2548 			mac_rx_group_add_client(group, mcip);
2549 			next_state = mac_rx_group_next_state(group,
2550 			    &group_only_mcip);
2551 
2552 			ASSERT((next_state == MAC_GROUP_STATE_RESERVED &&
2553 			    mcip == group_only_mcip) ||
2554 			    (next_state == MAC_GROUP_STATE_SHARED &&
2555 			    group_only_mcip == NULL));
2556 
2557 			mac_set_rx_group_state(group, next_state);
2558 		}
2559 
2560 		/*
2561 		 * Setup the Rx and Tx SRSes. If we got a pristine group
2562 		 * exclusively above, mac_srs_group_setup would simply create
2563 		 * the required SRSes. If we ended up sharing a previously
2564 		 * reserved group, mac_srs_group_setup would also dismantle the
2565 		 * SRSes of the previously exclusive group
2566 		 */
2567 		mac_srs_group_setup(mcip, flent, group, link_type);
2568 
2569 		/* Program the S/W Classifer */
2570 		if ((err = mac_flow_add(mip->mi_flow_tab, flent)) != 0)
2571 			goto setup_failed;
2572 
2573 		/* Program the H/W Classifier */
2574 		if ((err = mac_add_macaddr(mip, group, mac_addr,
2575 		    (mcip->mci_state_flags & MCIS_UNICAST_HW) != 0)) != 0)
2576 			goto setup_failed;
2577 		mcip->mci_unicast = mac_find_macaddr(mip, mac_addr);
2578 		ASSERT(mcip->mci_unicast != NULL);
2579 		break;
2580 
2581 	default:
2582 		ASSERT(B_FALSE);
2583 		break;
2584 	}
2585 
2586 	/*
2587 	 * All broadcast and multicast traffic is received only on the default
2588 	 * group. If we have setup the datapath for a non-default group above
2589 	 * then move the default group to shared state to allow distribution of
2590 	 * incoming broadcast traffic to the other groups and dismantle the
2591 	 * SRSes over the default group.
2592 	 */
2593 	if (group != NULL) {
2594 		if (group != mip->mi_rx_groups) {
2595 			default_group = mip->mi_rx_groups;
2596 			if (default_group->mrg_state ==
2597 			    MAC_GROUP_STATE_RESERVED) {
2598 				group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(
2599 				    default_group);
2600 				ASSERT(group_only_mcip != NULL &&
2601 				    mip->mi_nactiveclients > 1);
2602 
2603 				mac_set_rx_group_state(default_group,
2604 				    MAC_GROUP_STATE_SHARED);
2605 				mac_srs_group_setup(group_only_mcip,
2606 				    group_only_mcip->mci_flent,
2607 				    default_group, SRST_LINK);
2608 			}
2609 			ASSERT(default_group->mrg_state ==
2610 			    MAC_GROUP_STATE_SHARED);
2611 		}
2612 		/*
2613 		 * If we get an exclusive group for a VLAN MAC client we
2614 		 * need to take the s/w path to make the additional check for
2615 		 * the vid. Disable polling and set it to s/w classification.
2616 		 */
2617 		if (group->mrg_state == MAC_GROUP_STATE_RESERVED &&
2618 		    i_mac_flow_vid(mcip->mci_flent) != VLAN_ID_NONE) {
2619 			mac_rx_switch_grp_to_sw(group);
2620 		}
2621 	}
2622 	return (0);
2623 
2624 setup_failed:
2625 	mac_datapath_teardown(mcip, flent, link_type);
2626 	return (err);
2627 }
2628 
2629 void
2630 mac_datapath_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
2631     uint32_t link_type)
2632 {
2633 	mac_impl_t		*mip = mcip->mci_mip;
2634 	mac_group_t		*group = NULL;
2635 	mac_client_impl_t	*grp_only_mcip;
2636 	flow_entry_t		*group_only_flent;
2637 	mac_group_t		*default_group;
2638 	boolean_t		check_default_group = B_FALSE;
2639 	mac_group_state_t	next_state;
2640 
2641 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
2642 
2643 	switch (link_type) {
2644 	case SRST_FLOW:
2645 		mac_srs_group_teardown(mcip, flent, SRST_FLOW);
2646 		return;
2647 
2648 	case SRST_LINK:
2649 		/* Stop sending packets */
2650 		mac_tx_client_block(mcip);
2651 
2652 		/* Stop the packets coming from the H/W */
2653 		if (mcip->mci_unicast != NULL) {
2654 			int err;
2655 			err = mac_remove_macaddr(mcip->mci_unicast);
2656 			if (err != 0) {
2657 				cmn_err(CE_WARN, "%s: failed to remove a MAC"
2658 				    " address because of error 0x%x",
2659 				    mip->mi_name, err);
2660 			}
2661 			mcip->mci_unicast = NULL;
2662 		}
2663 
2664 		/* Stop the packets coming from the S/W classifier */
2665 		mac_flow_remove(mip->mi_flow_tab, flent, B_FALSE);
2666 		mac_flow_wait(flent, FLOW_DRIVER_UPCALL);
2667 
2668 		/* Now quiesce and destroy all SRS and soft rings */
2669 		mac_srs_group_teardown(mcip, flent, SRST_LINK);
2670 		ASSERT((mcip->mci_flent == flent) &&
2671 		    (flent->fe_next == NULL));
2672 
2673 		/*
2674 		 * Release our hold on the group as well. We need
2675 		 * to check if the shared group has only one client
2676 		 * left who can use it exclusively. Also, if we
2677 		 * were the last client, release the group.
2678 		 */
2679 		group = flent->fe_rx_ring_group;
2680 		if (group != NULL) {
2681 			mac_rx_group_remove_client(group, mcip);
2682 			next_state = mac_rx_group_next_state(group,
2683 			    &grp_only_mcip);
2684 			if (next_state == MAC_GROUP_STATE_RESERVED) {
2685 				/*
2686 				 * Only one client left on this RX group.
2687 				 */
2688 				ASSERT(grp_only_mcip != NULL);
2689 				mac_set_rx_group_state(group,
2690 				    MAC_GROUP_STATE_RESERVED);
2691 				group_only_flent = grp_only_mcip->mci_flent;
2692 
2693 				/*
2694 				 * The only remaining client has exclusive
2695 				 * access on the group. Allow it to
2696 				 * dynamically poll the H/W rings etc.
2697 				 */
2698 				mac_srs_group_setup(grp_only_mcip,
2699 				    group_only_flent, group, SRST_LINK);
2700 				mac_rx_group_unmark(group, MR_INCIPIENT);
2701 			} else if (next_state == MAC_GROUP_STATE_REGISTERED) {
2702 				/*
2703 				 * This is a non-default group being freed up.
2704 				 * We need to reevaluate the default group
2705 				 * to see if the primary client can get
2706 				 * exclusive access to the default group.
2707 				 */
2708 				ASSERT(group != mip->mi_rx_groups);
2709 				mac_release_rx_group(mcip, group);
2710 				mac_set_rx_group_state(group,
2711 				    MAC_GROUP_STATE_REGISTERED);
2712 				check_default_group = B_TRUE;
2713 			} else {
2714 				ASSERT(next_state == MAC_GROUP_STATE_SHARED);
2715 				mac_set_rx_group_state(group,
2716 				    MAC_GROUP_STATE_SHARED);
2717 				mac_rx_group_unmark(group, MR_CONDEMNED);
2718 			}
2719 			flent->fe_rx_ring_group = NULL;
2720 		}
2721 		break;
2722 	default:
2723 		ASSERT(B_FALSE);
2724 		break;
2725 	}
2726 
2727 	/*
2728 	 * The mac client using the default group gets exclusive access to the
2729 	 * default group if and only if it is the sole client on the entire
2730 	 * mip. If so set the group state to reserved, and set up the SRSes
2731 	 * over the default group.
2732 	 */
2733 	if (check_default_group) {
2734 		default_group = mip->mi_rx_groups;
2735 		ASSERT(default_group->mrg_state == MAC_GROUP_STATE_SHARED);
2736 		next_state = mac_rx_group_next_state(default_group,
2737 		    &grp_only_mcip);
2738 		if (next_state == MAC_GROUP_STATE_RESERVED) {
2739 			ASSERT(grp_only_mcip != NULL &&
2740 			    mip->mi_nactiveclients == 1);
2741 			mac_set_rx_group_state(default_group,
2742 			    MAC_GROUP_STATE_RESERVED);
2743 			mac_srs_group_setup(grp_only_mcip,
2744 			    grp_only_mcip->mci_flent,
2745 			    default_group, SRST_LINK);
2746 			mac_rx_group_unmark(default_group, MR_INCIPIENT);
2747 		}
2748 	}
2749 }
2750 
2751 /* DATAPATH TEAR DOWN ROUTINES (SRS and FANOUT teardown) */
2752 
2753 static void
2754 mac_srs_fanout_list_free(mac_soft_ring_set_t *mac_srs)
2755 {
2756 	ASSERT(mac_srs->srs_tcp_soft_rings != NULL);
2757 	kmem_free(mac_srs->srs_tcp_soft_rings,
2758 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2759 	mac_srs->srs_tcp_soft_rings = NULL;
2760 	ASSERT(mac_srs->srs_udp_soft_rings != NULL);
2761 	kmem_free(mac_srs->srs_udp_soft_rings,
2762 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2763 	mac_srs->srs_udp_soft_rings = NULL;
2764 	ASSERT(mac_srs->srs_oth_soft_rings != NULL);
2765 	kmem_free(mac_srs->srs_oth_soft_rings,
2766 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2767 	mac_srs->srs_oth_soft_rings = NULL;
2768 }
2769 
2770 /*
2771  * An RX SRS is attached to at most one mac_ring.
2772  * A TX SRS  has no  rings.
2773  */
2774 static void
2775 mac_srs_ring_free(mac_soft_ring_set_t *mac_srs)
2776 {
2777 	mac_client_impl_t	*mcip;
2778 	mac_ring_t		*ring;
2779 	flow_entry_t		*flent;
2780 
2781 	ring = mac_srs->srs_ring;
2782 	if (mac_srs->srs_type & SRST_TX) {
2783 		ASSERT(ring == NULL);
2784 		return;
2785 	}
2786 
2787 	if (ring == NULL)
2788 		return;
2789 
2790 	/*
2791 	 * Broadcast flows don't have a client impl association, but they
2792 	 * use only soft rings.
2793 	 */
2794 	flent = mac_srs->srs_flent;
2795 	mcip = flent->fe_mcip;
2796 	ASSERT(mcip != NULL);
2797 
2798 	ring->mr_classify_type = MAC_NO_CLASSIFIER;
2799 	ring->mr_srs = NULL;
2800 }
2801 
2802 /*
2803  * Physical unlink and free of the data structures happen below. This is
2804  * driven from mac_flow_destroy(), on the last refrele of a flow.
2805  *
2806  * Assumes Rx srs is 1-1 mapped with an ring.
2807  */
2808 void
2809 mac_srs_free(mac_soft_ring_set_t *mac_srs)
2810 {
2811 	ASSERT(mac_srs->srs_mcip == NULL ||
2812 	    MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
2813 	ASSERT((mac_srs->srs_state & (SRS_CONDEMNED | SRS_CONDEMNED_DONE |
2814 	    SRS_PROC | SRS_PROC_FAST)) == (SRS_CONDEMNED | SRS_CONDEMNED_DONE));
2815 
2816 	mac_pkt_drop(NULL, NULL, mac_srs->srs_first, B_FALSE);
2817 	mac_srs_ring_free(mac_srs);
2818 	mac_srs_soft_rings_free(mac_srs, B_TRUE);
2819 	mac_srs_fanout_list_free(mac_srs);
2820 
2821 	mac_srs->srs_bw = NULL;
2822 	kmem_cache_free(mac_srs_cache, mac_srs);
2823 }
2824 
2825 static void
2826 mac_srs_soft_rings_quiesce(mac_soft_ring_set_t *mac_srs, uint_t s_ring_flag)
2827 {
2828 	mac_soft_ring_t	*softring;
2829 
2830 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2831 
2832 	mac_srs_soft_rings_signal(mac_srs, s_ring_flag);
2833 	if (s_ring_flag == S_RING_CONDEMNED) {
2834 		while (mac_srs->srs_soft_ring_condemned_count !=
2835 		    mac_srs->srs_soft_ring_count)
2836 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2837 	} else {
2838 		while (mac_srs->srs_soft_ring_quiesced_count !=
2839 		    mac_srs->srs_soft_ring_count)
2840 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2841 	}
2842 	mutex_exit(&mac_srs->srs_lock);
2843 
2844 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2845 	    softring = softring->s_ring_next)
2846 		(void) untimeout(softring->s_ring_tid);
2847 
2848 	(void) untimeout(mac_srs->srs_tid);
2849 
2850 	mutex_enter(&mac_srs->srs_lock);
2851 }
2852 
2853 /*
2854  * The block comment above mac_rx_classify_flow_state_change explains the
2855  * background. At this point upcalls from the driver (both hardware classified
2856  * and software classified) have been cut off. We now need to quiesce the
2857  * SRS worker, poll, and softring threads. The SRS worker thread serves as
2858  * the master controller. The steps involved are described below in the function
2859  */
2860 void
2861 mac_srs_worker_quiesce(mac_soft_ring_set_t *mac_srs)
2862 {
2863 	uint_t			s_ring_flag;
2864 	uint_t			srs_poll_wait_flag;
2865 
2866 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2867 	ASSERT(mac_srs->srs_state & (SRS_CONDEMNED | SRS_QUIESCE));
2868 
2869 	if (mac_srs->srs_state & SRS_CONDEMNED) {
2870 		s_ring_flag = S_RING_CONDEMNED;
2871 		srs_poll_wait_flag = SRS_POLL_THR_EXITED;
2872 	} else {
2873 		s_ring_flag = S_RING_QUIESCE;
2874 		srs_poll_wait_flag = SRS_POLL_THR_QUIESCED;
2875 	}
2876 
2877 	/*
2878 	 * In the case of Rx SRS wait till the poll thread is done.
2879 	 */
2880 	if ((mac_srs->srs_type & SRST_TX) == 0 &&
2881 	    mac_srs->srs_poll_thr != NULL) {
2882 		while (!(mac_srs->srs_state & srs_poll_wait_flag))
2883 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2884 
2885 		/*
2886 		 * Turn off polling as part of the quiesce operation.
2887 		 */
2888 		MAC_SRS_POLLING_OFF(mac_srs);
2889 		mac_srs->srs_state &= ~(SRS_POLLING | SRS_GET_PKTS);
2890 	}
2891 
2892 	/*
2893 	 * Then signal the soft ring worker threads to quiesce or quit
2894 	 * as needed and then wait till that happens.
2895 	 */
2896 	mac_srs_soft_rings_quiesce(mac_srs, s_ring_flag);
2897 
2898 	if (mac_srs->srs_state & SRS_CONDEMNED)
2899 		mac_srs->srs_state |= (SRS_QUIESCE_DONE | SRS_CONDEMNED_DONE);
2900 	else
2901 		mac_srs->srs_state |= SRS_QUIESCE_DONE;
2902 	cv_signal(&mac_srs->srs_quiesce_done_cv);
2903 }
2904 
2905 /*
2906  * Signal an SRS to start a temporary quiesce, or permanent removal, or restart
2907  * a quiesced SRS by setting the appropriate flags and signaling the SRS worker
2908  * or poll thread. This function is internal to the quiescing logic and is
2909  * called internally from the SRS quiesce or flow quiesce or client quiesce
2910  * higher level functions.
2911  */
2912 void
2913 mac_srs_signal(mac_soft_ring_set_t *mac_srs, uint_t srs_flag)
2914 {
2915 	mac_ring_t	*ring;
2916 
2917 	ring = mac_srs->srs_ring;
2918 	ASSERT(ring == NULL || ring->mr_refcnt == 0);
2919 
2920 	if (srs_flag == SRS_CONDEMNED) {
2921 		/*
2922 		 * The SRS is going away. We need to unbind the SRS and SR
2923 		 * threads before removing from the global SRS list. Otherwise
2924 		 * there is a small window where the cpu reconfig callbacks
2925 		 * may miss the SRS in the list walk and DR could fail since
2926 		 * there are still bound threads.
2927 		 */
2928 		mac_srs_threads_unbind(mac_srs);
2929 		mac_srs_remove_glist(mac_srs);
2930 	}
2931 	/*
2932 	 * Wakeup the SRS worker and poll threads.
2933 	 */
2934 	mutex_enter(&mac_srs->srs_lock);
2935 	mac_srs->srs_state |= srs_flag;
2936 	cv_signal(&mac_srs->srs_async);
2937 	cv_signal(&mac_srs->srs_cv);
2938 	mutex_exit(&mac_srs->srs_lock);
2939 }
2940 
2941 /*
2942  * In the Rx side, the quiescing is done bottom up. After the Rx upcalls
2943  * from the driver are done, then the Rx SRS is quiesced and only then can
2944  * we signal the soft rings. Thus this function can't be called arbitrarily
2945  * without satisfying the prerequisites. On the Tx side, the threads from
2946  * top need to quiesced, then the Tx SRS and only then can we signal the
2947  * Tx soft rings.
2948  */
2949 static void
2950 mac_srs_soft_rings_signal(mac_soft_ring_set_t *mac_srs, uint_t sr_flag)
2951 {
2952 	mac_soft_ring_t		*softring;
2953 
2954 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2955 	    softring = softring->s_ring_next)
2956 		mac_soft_ring_signal(softring, sr_flag);
2957 }
2958 
2959 /*
2960  * The block comment above mac_rx_classify_flow_state_change explains the
2961  * background. At this point the SRS is quiesced and we need to restart the
2962  * SRS worker, poll, and softring threads. The SRS worker thread serves as
2963  * the master controller. The steps involved are described below in the function
2964  */
2965 void
2966 mac_srs_worker_restart(mac_soft_ring_set_t *mac_srs)
2967 {
2968 	boolean_t	iam_rx_srs;
2969 	mac_soft_ring_t	*softring;
2970 
2971 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2972 	if ((mac_srs->srs_type & SRST_TX) != 0) {
2973 		iam_rx_srs = B_FALSE;
2974 		ASSERT((mac_srs->srs_state &
2975 		    (SRS_POLL_THR_QUIESCED | SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
2976 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
2977 	} else {
2978 		iam_rx_srs = B_TRUE;
2979 		ASSERT((mac_srs->srs_state &
2980 		    (SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
2981 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
2982 		if (mac_srs->srs_poll_thr != NULL) {
2983 			ASSERT((mac_srs->srs_state & SRS_POLL_THR_QUIESCED) ==
2984 			    SRS_POLL_THR_QUIESCED);
2985 		}
2986 	}
2987 
2988 	/*
2989 	 * Signal any quiesced soft ring workers to restart and wait for the
2990 	 * soft ring down count to come down to zero.
2991 	 */
2992 	if (mac_srs->srs_soft_ring_quiesced_count != 0) {
2993 		for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2994 		    softring = softring->s_ring_next) {
2995 			if (!(softring->s_ring_state & S_RING_QUIESCE))
2996 				continue;
2997 			mac_soft_ring_signal(softring, S_RING_RESTART);
2998 		}
2999 		while (mac_srs->srs_soft_ring_quiesced_count != 0)
3000 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3001 	}
3002 
3003 	mac_srs->srs_state &= ~(SRS_QUIESCE_DONE | SRS_QUIESCE | SRS_RESTART);
3004 	if (iam_rx_srs && mac_srs->srs_poll_thr != NULL) {
3005 		/*
3006 		 * Signal the poll thread and ask it to restart. Wait till it
3007 		 * actually restarts and the SRS_POLL_THR_QUIESCED flag gets
3008 		 * cleared.
3009 		 */
3010 		mac_srs->srs_state |= SRS_POLL_THR_RESTART;
3011 		cv_signal(&mac_srs->srs_cv);
3012 		while (mac_srs->srs_state & SRS_POLL_THR_QUIESCED)
3013 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3014 		ASSERT(!(mac_srs->srs_state & SRS_POLL_THR_RESTART));
3015 	}
3016 	/* Wake up any waiter waiting for the restart to complete */
3017 	mac_srs->srs_state |= SRS_RESTART_DONE;
3018 	cv_signal(&mac_srs->srs_quiesce_done_cv);
3019 }
3020 
3021 static void
3022 mac_srs_worker_unbind(mac_soft_ring_set_t *mac_srs)
3023 {
3024 	mutex_enter(&mac_srs->srs_lock);
3025 	if (!(mac_srs->srs_state & SRS_WORKER_BOUND)) {
3026 		ASSERT(mac_srs->srs_worker_cpuid == -1);
3027 		mutex_exit(&mac_srs->srs_lock);
3028 		return;
3029 	}
3030 
3031 	mac_srs->srs_worker_cpuid = -1;
3032 	mac_srs->srs_state &= ~SRS_WORKER_BOUND;
3033 	thread_affinity_clear(mac_srs->srs_worker);
3034 	mutex_exit(&mac_srs->srs_lock);
3035 }
3036 
3037 static void
3038 mac_srs_poll_unbind(mac_soft_ring_set_t *mac_srs)
3039 {
3040 	mutex_enter(&mac_srs->srs_lock);
3041 	if (mac_srs->srs_poll_thr == NULL ||
3042 	    (mac_srs->srs_state & SRS_POLL_BOUND) == 0) {
3043 		ASSERT(mac_srs->srs_poll_cpuid == -1);
3044 		mutex_exit(&mac_srs->srs_lock);
3045 		return;
3046 	}
3047 
3048 	mac_srs->srs_poll_cpuid = -1;
3049 	mac_srs->srs_state &= ~SRS_POLL_BOUND;
3050 	thread_affinity_clear(mac_srs->srs_poll_thr);
3051 	mutex_exit(&mac_srs->srs_lock);
3052 }
3053 
3054 static void
3055 mac_srs_threads_unbind(mac_soft_ring_set_t *mac_srs)
3056 {
3057 	mac_soft_ring_t	*soft_ring;
3058 
3059 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
3060 
3061 	mutex_enter(&cpu_lock);
3062 	mac_srs_worker_unbind(mac_srs);
3063 	if (!(mac_srs->srs_type & SRST_TX))
3064 		mac_srs_poll_unbind(mac_srs);
3065 
3066 	for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3067 	    soft_ring = soft_ring->s_ring_next) {
3068 		mac_soft_ring_unbind(soft_ring);
3069 	}
3070 	mutex_exit(&cpu_lock);
3071 }
3072 
3073 /*
3074  * When a CPU is going away, unbind all MAC threads which are bound
3075  * to that CPU. The affinity of the thread to the CPU is saved to allow
3076  * the thread to be rebound to the CPU if it comes back online.
3077  */
3078 static void
3079 mac_walk_srs_and_unbind(int cpuid)
3080 {
3081 	mac_soft_ring_set_t *mac_srs;
3082 	mac_soft_ring_t *soft_ring;
3083 
3084 	rw_enter(&mac_srs_g_lock, RW_READER);
3085 
3086 	if ((mac_srs = mac_srs_g_list) == NULL)
3087 		goto done;
3088 
3089 	for (; mac_srs != NULL; mac_srs = mac_srs->srs_next) {
3090 		if (mac_srs->srs_worker_cpuid == cpuid) {
3091 			mac_srs->srs_worker_cpuid_save = cpuid;
3092 			mac_srs_worker_unbind(mac_srs);
3093 		}
3094 
3095 		if (!(mac_srs->srs_type & SRST_TX)) {
3096 			if (mac_srs->srs_poll_cpuid == cpuid) {
3097 				mac_srs->srs_poll_cpuid_save = cpuid;
3098 				mac_srs_poll_unbind(mac_srs);
3099 			}
3100 		}
3101 
3102 		/* Next tackle the soft rings associated with the srs */
3103 		mutex_enter(&mac_srs->srs_lock);
3104 		for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3105 		    soft_ring = soft_ring->s_ring_next) {
3106 			if (soft_ring->s_ring_cpuid == cpuid) {
3107 				soft_ring->s_ring_cpuid_save = cpuid;
3108 				mac_soft_ring_unbind(soft_ring);
3109 			}
3110 		}
3111 		mutex_exit(&mac_srs->srs_lock);
3112 	}
3113 done:
3114 	rw_exit(&mac_srs_g_lock);
3115 }
3116 
3117 /* TX SETUP and TEARDOWN ROUTINES */
3118 
3119 /*
3120  * XXXHIO need to make sure the two mac_tx_srs_{add,del}_ring()
3121  * handle the case where the number of rings is one. I.e. there is
3122  * a ring pointed to by mac_srs->srs_tx_arg2.
3123  */
3124 void
3125 mac_tx_srs_add_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3126 {
3127 	mac_client_impl_t *mcip = mac_srs->srs_mcip;
3128 	mac_soft_ring_t *soft_ring;
3129 	int count = mac_srs->srs_oth_ring_count;
3130 
3131 	ASSERT(mac_srs->srs_state & SRS_QUIESCE);
3132 	soft_ring = mac_soft_ring_create(count, 0, NULL,
3133 	    (ST_RING_OTH | ST_RING_TX), maxclsyspri, mcip, mac_srs, -1,
3134 	    NULL, mcip, (mac_resource_handle_t)tx_ring);
3135 	mac_srs->srs_oth_ring_count++;
3136 	/*
3137 	 * put this soft ring in quiesce mode too so when we restart
3138 	 * all soft rings in the srs are in the same state.
3139 	 */
3140 	mac_soft_ring_signal(soft_ring, S_RING_QUIESCE);
3141 }
3142 
3143 static void
3144 mac_soft_ring_remove(mac_soft_ring_set_t *mac_srs, mac_soft_ring_t *softring)
3145 {
3146 	int sringcnt;
3147 
3148 	mutex_enter(&mac_srs->srs_lock);
3149 	sringcnt = mac_srs->srs_soft_ring_count;
3150 	ASSERT(sringcnt > 0);
3151 	mac_soft_ring_signal(softring, S_RING_CONDEMNED);
3152 
3153 	ASSERT(mac_srs->srs_soft_ring_condemned_count == 0);
3154 	while (mac_srs->srs_soft_ring_condemned_count != 1)
3155 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3156 
3157 	if (softring == mac_srs->srs_soft_ring_head) {
3158 		mac_srs->srs_soft_ring_head = softring->s_ring_next;
3159 		if (mac_srs->srs_soft_ring_head != NULL) {
3160 			mac_srs->srs_soft_ring_head->s_ring_prev = NULL;
3161 		} else {
3162 			mac_srs->srs_soft_ring_tail = NULL;
3163 		}
3164 	} else {
3165 		softring->s_ring_prev->s_ring_next =
3166 		    softring->s_ring_next;
3167 		if (softring->s_ring_next != NULL) {
3168 			softring->s_ring_next->s_ring_prev =
3169 			    softring->s_ring_prev;
3170 		} else {
3171 			mac_srs->srs_soft_ring_tail =
3172 			    softring->s_ring_prev;
3173 		}
3174 	}
3175 	mac_srs->srs_soft_ring_count--;
3176 
3177 	mac_srs->srs_soft_ring_condemned_count--;
3178 	mutex_exit(&mac_srs->srs_lock);
3179 
3180 	mac_soft_ring_free(softring, B_FALSE);
3181 }
3182 
3183 void
3184 mac_tx_srs_del_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3185 {
3186 	int i;
3187 	mac_soft_ring_t *soft_ring, *remove_sring;
3188 
3189 	mutex_enter(&mac_srs->srs_lock);
3190 	for (i = 0; i < mac_srs->srs_oth_ring_count; i++) {
3191 		soft_ring =  mac_srs->srs_oth_soft_rings[i];
3192 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3193 			break;
3194 	}
3195 	mutex_exit(&mac_srs->srs_lock);
3196 	ASSERT(i < mac_srs->srs_oth_ring_count);
3197 	remove_sring = soft_ring;
3198 	mac_soft_ring_remove(mac_srs, remove_sring);
3199 	mac_srs_update_fanout_list(mac_srs);
3200 }
3201 
3202 /*
3203  * mac_tx_srs_setup():
3204  *
3205  * Used to setup Tx rings. If no free Tx ring is available, then default
3206  * Tx ring is used.
3207  */
3208 void
3209 mac_tx_srs_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
3210     uint32_t srs_type)
3211 {
3212 	mac_impl_t *mip = mcip->mci_mip;
3213 	mac_soft_ring_set_t *tx_srs;
3214 	int i, tx_ring_count = 0, tx_rings_reserved = 0;
3215 	mac_ring_handle_t *tx_rings = NULL;
3216 	uint32_t soft_ring_type;
3217 	mac_group_t *grp = NULL;
3218 	mac_ring_t *ring;
3219 	mac_srs_tx_t *tx;
3220 	boolean_t serialize = B_FALSE;
3221 
3222 	tx_srs = flent->fe_tx_srs;
3223 	tx = &tx_srs->srs_tx;
3224 
3225 	if (tx->st_group != NULL) {
3226 		grp = tx->st_group;
3227 		tx_ring_count = grp->mrg_cur_count;
3228 	} else {
3229 		tx_ring_count = mac_tx_ring_count;
3230 	}
3231 
3232 	if (tx_ring_count != 0) {
3233 		tx_rings = kmem_zalloc(sizeof (mac_ring_handle_t) *
3234 		    tx_ring_count, KM_SLEEP);
3235 	}
3236 
3237 	/*
3238 	 * Just use the default ring for now. We need to use
3239 	 * the underlying link's ring set instead of the underlying
3240 	 * NIC's.
3241 	 */
3242 	if (srs_type == SRST_FLOW ||
3243 	    (mcip->mci_state_flags & MCIS_NO_HWRINGS) != 0) {
3244 		/* use default ring */
3245 		tx_rings[0] = (void *)mip->mi_default_tx_ring;
3246 		tx_rings_reserved++;
3247 		goto rings_assigned;
3248 	}
3249 
3250 	if (mcip->mci_share != NULL)
3251 		ring = grp->mrg_rings;
3252 	/*
3253 	 * An attempt is made to reserve 'tx_ring_count' number
3254 	 * of Tx rings. If tx_ring_count is 0, default Tx ring
3255 	 * is used. If it is 1, an attempt is made to reserve one
3256 	 * Tx ring. In both the cases, the ring information is
3257 	 * stored in Tx SRS. If multiple Tx rings are specified,
3258 	 * then each Tx ring will have a Tx-side soft ring. All
3259 	 * these soft rings will be hang off Tx SRS.
3260 	 */
3261 	for (i = 0; i < tx_ring_count; i++) {
3262 		if (mcip->mci_share != NULL) {
3263 			/*
3264 			 * The ring was already chosen and associated
3265 			 * with the TX group. Save it in the new
3266 			 * array to keep as much of the code below common
3267 			 * between the share and non-share cases.
3268 			 */
3269 			ASSERT(ring != NULL);
3270 			tx_rings[i] = (mac_ring_handle_t)ring;
3271 			ring = ring->mr_next;
3272 		} else {
3273 			tx_rings[i] =
3274 			    (mac_ring_handle_t)mac_reserve_tx_ring(mip, NULL);
3275 			if (tx_rings[i] == NULL) {
3276 				/*
3277 				 * We have run out of Tx rings. So
3278 				 * give the default ring too.
3279 				 */
3280 				tx_rings[i] = (void *)mip->mi_default_tx_ring;
3281 				tx_rings_reserved++;
3282 				break;
3283 			}
3284 		}
3285 		tx_rings_reserved++;
3286 	}
3287 
3288 rings_assigned:
3289 	if (mac_tx_serialize || (mip->mi_v12n_level & MAC_VIRT_SERIALIZE))
3290 		serialize = B_TRUE;
3291 	/*
3292 	 * Did we get the requested number of tx rings?
3293 	 * There are 2 actions we can take depending upon the number
3294 	 * of tx_rings we got.
3295 	 * 1) If we got one, then get the tx_ring from the soft ring,
3296 	 * save it in SRS and free up the soft ring.
3297 	 * 2) If we got more than 1, then do the tx fanout among the
3298 	 * rings we obtained.
3299 	 */
3300 	ASSERT(tx_rings_reserved != 0);
3301 	if (tx_rings_reserved == 1) {
3302 		tx->st_arg2 = (void *)tx_rings[0];
3303 		/* For ring_count of 0 or 1, set the tx_mode and return */
3304 		if (tx_srs->srs_type & SRST_BW_CONTROL)
3305 			tx->st_mode = SRS_TX_BW;
3306 		else if (serialize)
3307 			tx->st_mode = SRS_TX_SERIALIZE;
3308 		else
3309 			tx->st_mode = SRS_TX_DEFAULT;
3310 	} else {
3311 		/*
3312 		 * We got multiple Tx rings for Tx fanout.
3313 		 */
3314 		soft_ring_type = ST_RING_OTH | ST_RING_TX;
3315 		if (tx_srs->srs_type & SRST_BW_CONTROL) {
3316 			tx->st_mode = SRS_TX_BW_FANOUT;
3317 		} else {
3318 			tx->st_mode = SRS_TX_FANOUT;
3319 			if (serialize)
3320 				soft_ring_type |= ST_RING_WORKER_ONLY;
3321 		}
3322 		for (i = 0; i < tx_rings_reserved; i++) {
3323 			(void) mac_soft_ring_create(i, 0, NULL, soft_ring_type,
3324 			    maxclsyspri, mcip, tx_srs, -1, NULL, mcip,
3325 			    (mac_resource_handle_t)tx_rings[i]);
3326 		}
3327 		mac_srs_update_fanout_list(tx_srs);
3328 	}
3329 	tx->st_func = mac_tx_get_func(tx->st_mode);
3330 
3331 	DTRACE_PROBE3(tx__srs___setup__return, mac_soft_ring_set_t *, tx_srs,
3332 	    int, tx->st_mode, int, tx_srs->srs_oth_ring_count);
3333 
3334 	if (tx_ring_count != 0) {
3335 		tx->st_ring_count = tx_rings_reserved;
3336 		tx->st_rings = kmem_zalloc(sizeof (mac_ring_handle_t) *
3337 		    tx_rings_reserved, KM_SLEEP);
3338 		for (i = 0; i < tx->st_ring_count; i++)
3339 			tx->st_rings[i] = tx_rings[i];
3340 		kmem_free(tx_rings, sizeof (mac_ring_handle_t) * tx_ring_count);
3341 	}
3342 }
3343 
3344 /*
3345  * Update the fanout of a client if its recorded link speed doesn't match
3346  * its current link speed.
3347  */
3348 void
3349 mac_fanout_recompute_client(mac_client_impl_t *mcip)
3350 {
3351 	uint64_t link_speed;
3352 	mac_resource_props_t *mcip_mrp;
3353 
3354 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
3355 
3356 	link_speed = mac_client_stat_get(mcip->mci_flent->fe_mcip,
3357 	    MAC_STAT_IFSPEED);
3358 
3359 	if ((link_speed != 0) &&
3360 	    (link_speed != mcip->mci_flent->fe_nic_speed)) {
3361 		mcip_mrp = MCIP_RESOURCE_PROPS(mcip);
3362 		mac_fanout_setup(mcip, mcip->mci_flent,
3363 		    mcip_mrp, mac_rx_deliver, mcip, NULL);
3364 	}
3365 }
3366 
3367 /*
3368  * Walk through the list of mac clients for the MAC.
3369  * For each active mac client, recompute the number of soft rings
3370  * associated with every client, only if current speed is different
3371  * from the speed that was previously used for soft ring computation.
3372  * If the cable is disconnected whlie the NIC is started, we would get
3373  * notification with speed set to 0. We do not recompute in that case.
3374  */
3375 void
3376 mac_fanout_recompute(mac_impl_t *mip)
3377 {
3378 	mac_client_impl_t	*mcip;
3379 
3380 	i_mac_perim_enter(mip);
3381 	if ((mip->mi_state_flags & MIS_IS_VNIC) != 0 ||
3382 	    mip->mi_linkstate != LINK_STATE_UP) {
3383 		i_mac_perim_exit(mip);
3384 		return;
3385 	}
3386 
3387 	for (mcip = mip->mi_clients_list; mcip != NULL;
3388 	    mcip = mcip->mci_client_next) {
3389 		if ((mcip->mci_state_flags & MCIS_SHARE_BOUND) != 0 ||
3390 		    !MCIP_DATAPATH_SETUP(mcip))
3391 			continue;
3392 		mac_fanout_recompute_client(mcip);
3393 	}
3394 	i_mac_perim_exit(mip);
3395 }
3396 
3397 /*
3398  * Given a MAC, change the polling state for all its MAC clients.  'enable' is
3399  * B_TRUE to enable polling or B_FALSE to disable.  Polling is enabled by
3400  * default.
3401  */
3402 void
3403 mac_poll_state_change(mac_handle_t mh, boolean_t enable)
3404 {
3405 	mac_impl_t *mip = (mac_impl_t *)mh;
3406 	mac_client_impl_t *mcip;
3407 
3408 	i_mac_perim_enter(mip);
3409 	if (enable)
3410 		mip->mi_state_flags &= ~MIS_POLL_DISABLE;
3411 	else
3412 		mip->mi_state_flags |= MIS_POLL_DISABLE;
3413 	for (mcip = mip->mi_clients_list; mcip != NULL;
3414 	    mcip = mcip->mci_client_next)
3415 		mac_client_update_classifier(mcip, B_TRUE);
3416 	i_mac_perim_exit(mip);
3417 }
3418