xref: /illumos-gate/usr/src/uts/common/io/mac/mac_datapath_setup.c (revision 7a088f03b431bdffa96c3b2175964d4d38420caa)
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 	/*
1362 	 * If receive function has already been configured correctly for
1363 	 * current subflow configuration, do nothing.
1364 	 */
1365 	if (flent->fe_cb_fn == (flow_fn_t)rx_func)
1366 		return;
1367 
1368 	rx_srs_cnt = flent->fe_rx_srs_cnt;
1369 	for (i = 0; i < rx_srs_cnt; i++) {
1370 		ASSERT(flent->fe_rx_srs[i] != NULL);
1371 		mac_srs_poll_state_change(flent->fe_rx_srs[i],
1372 		    enable_classifier, rx_func);
1373 	}
1374 
1375 	/*
1376 	 * Change the S/W classifier so that we can land in the
1377 	 * correct processing function with correct argument.
1378 	 * If all subflows have been removed we can revert to
1379 	 * mac_rx_srsprocess, else we need mac_rx_srs_subflow_process.
1380 	 */
1381 	mutex_enter(&flent->fe_lock);
1382 	flent->fe_cb_fn = (flow_fn_t)rx_func;
1383 	flent->fe_cb_arg1 = (void *)mip;
1384 	flent->fe_cb_arg2 = flent->fe_rx_srs[0];
1385 	mutex_exit(&flent->fe_lock);
1386 }
1387 
1388 static void
1389 mac_srs_update_fanout_list(mac_soft_ring_set_t *mac_srs)
1390 {
1391 	int		tcp_count = 0;
1392 	int		udp_count = 0;
1393 	int		oth_count = 0;
1394 	mac_soft_ring_t *softring;
1395 
1396 	softring = mac_srs->srs_soft_ring_head;
1397 	if (softring == NULL) {
1398 		ASSERT(mac_srs->srs_soft_ring_count == 0);
1399 		mac_srs->srs_tcp_ring_count = 0;
1400 		mac_srs->srs_udp_ring_count = 0;
1401 		mac_srs->srs_oth_ring_count = 0;
1402 		return;
1403 	}
1404 
1405 	softring = mac_srs->srs_soft_ring_head;
1406 	tcp_count = udp_count = oth_count = 0;
1407 
1408 	while (softring != NULL) {
1409 		if (softring->s_ring_type & ST_RING_TCP)
1410 			mac_srs->srs_tcp_soft_rings[tcp_count++] = softring;
1411 		else if (softring->s_ring_type & ST_RING_UDP)
1412 			mac_srs->srs_udp_soft_rings[udp_count++] = softring;
1413 		else
1414 			mac_srs->srs_oth_soft_rings[oth_count++] = softring;
1415 		softring = softring->s_ring_next;
1416 	}
1417 
1418 	ASSERT(mac_srs->srs_soft_ring_count ==
1419 	    (tcp_count + udp_count + oth_count));
1420 
1421 	mac_srs->srs_tcp_ring_count = tcp_count;
1422 	mac_srs->srs_udp_ring_count = udp_count;
1423 	mac_srs->srs_oth_ring_count = oth_count;
1424 }
1425 
1426 void
1427 mac_srs_create_proto_softrings(int id, void *flent, uint16_t type,
1428     pri_t pri, mac_client_impl_t *mcip, mac_soft_ring_set_t *mac_srs,
1429     processorid_t cpuid, mac_direct_rx_t rx_func, void *x_arg1,
1430     mac_resource_handle_t x_arg2, boolean_t set_bypass)
1431 {
1432 	mac_soft_ring_t	*softring;
1433 	mac_rx_fifo_t	mrf;
1434 
1435 	bzero(&mrf, sizeof (mac_rx_fifo_t));
1436 	mrf.mrf_type = MAC_RX_FIFO;
1437 	mrf.mrf_receive = (mac_receive_t)mac_soft_ring_poll;
1438 	mrf.mrf_intr_enable =
1439 	    (mac_intr_enable_t)mac_soft_ring_intr_enable;
1440 	mrf.mrf_intr_disable =
1441 	    (mac_intr_disable_t)mac_soft_ring_intr_disable;
1442 	mrf.mrf_flow_priority = pri;
1443 
1444 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1445 	    (void *)flent, (type|ST_RING_TCP), pri, mcip, mac_srs,
1446 	    cpuid, rx_func, x_arg1, x_arg2);
1447 	softring->s_ring_rx_arg2 = NULL;
1448 
1449 	/*
1450 	 * TCP and UDP support DLS bypass. In addition TCP
1451 	 * squeue can also poll their corresponding soft rings.
1452 	 */
1453 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1454 		mac_soft_ring_dls_bypass(softring,
1455 		    mcip->mci_direct_rx_fn,
1456 		    mcip->mci_direct_rx_arg);
1457 
1458 		mrf.mrf_rx_arg = softring;
1459 		mrf.mrf_intr_handle = (mac_intr_handle_t)softring;
1460 
1461 		/*
1462 		 * Make a call in IP to get a TCP squeue assigned to
1463 		 * this softring to maintain full CPU locality through
1464 		 * the stack and allow the squeue to be able to poll
1465 		 * the softring so the flow control can be pushed
1466 		 * all the way to H/W.
1467 		 */
1468 		softring->s_ring_rx_arg2 =
1469 		    mcip->mci_resource_add((void *)mcip->mci_resource_arg,
1470 		    (mac_resource_t *)&mrf);
1471 	}
1472 
1473 	/*
1474 	 * Non-TCP protocols don't support squeues. Hence we
1475 	 * don't make any ring addition callbacks for non-TCP
1476 	 * rings. Now create the UDP softring and allow it to
1477 	 * bypass the DLS layer.
1478 	 */
1479 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1480 	    (void *)flent, (type|ST_RING_UDP), pri, mcip, mac_srs,
1481 	    cpuid, rx_func, x_arg1, x_arg2);
1482 	softring->s_ring_rx_arg2 = NULL;
1483 
1484 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1485 		mac_soft_ring_dls_bypass(softring,
1486 		    mcip->mci_direct_rx_fn,
1487 		    mcip->mci_direct_rx_arg);
1488 	}
1489 
1490 	/* Create the Oth softrings which has to go through the DLS */
1491 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1492 	    (void *)flent, (type|ST_RING_OTH), pri, mcip, mac_srs,
1493 	    cpuid, rx_func, x_arg1, x_arg2);
1494 	softring->s_ring_rx_arg2 = NULL;
1495 }
1496 
1497 /*
1498  * This routine associates a CPU or a set of CPU to process incoming
1499  * traffic from a mac client. If multiple CPUs are specified, then
1500  * so many soft rings are created with each soft ring worker thread
1501  * bound to a CPU in the set. Each soft ring in turn will be
1502  * associated with an squeue and the squeue will be moved to the
1503  * same CPU as that of the soft ring's.
1504  */
1505 static void
1506 mac_srs_fanout_modify(mac_client_impl_t *mcip, flow_entry_t *flent,
1507     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1508     mac_resource_handle_t x_arg2, mac_soft_ring_set_t *mac_rx_srs,
1509     mac_soft_ring_set_t *mac_tx_srs)
1510 {
1511 	mac_soft_ring_t *softring;
1512 	uint32_t soft_ring_flag = 0;
1513 	processorid_t cpuid = -1;
1514 	boolean_t user_specified;
1515 	int i, srings_present, new_fanout_cnt;
1516 	mac_cpus_t *srs_cpu;
1517 
1518 	user_specified = mrp->mrp_mask & MRP_CPUS_USERSPEC;
1519 	/* fanout state is REINIT. Set it back to INIT */
1520 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_REINIT);
1521 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1522 
1523 	/* how many are present right now */
1524 	srings_present = mac_rx_srs->srs_tcp_ring_count;
1525 	/* new request */
1526 	srs_cpu = &mac_rx_srs->srs_cpu;
1527 	new_fanout_cnt = srs_cpu->mc_fanout_cnt;
1528 
1529 	mutex_enter(&mac_rx_srs->srs_lock);
1530 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1531 		soft_ring_flag |= ST_RING_BW_CTL;
1532 	mutex_exit(&mac_rx_srs->srs_lock);
1533 
1534 	if (new_fanout_cnt > srings_present) {
1535 		/* soft rings increased */
1536 		mutex_enter(&mac_rx_srs->srs_lock);
1537 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1538 		mutex_exit(&mac_rx_srs->srs_lock);
1539 
1540 		for (i = mac_rx_srs->srs_tcp_ring_count;
1541 		    i < new_fanout_cnt; i++) {
1542 			/*
1543 			 * Create the protocol softrings and set the
1544 			 * DLS bypass where possible.
1545 			 */
1546 			mac_srs_create_proto_softrings(i,
1547 			    (void *)flent, soft_ring_flag,
1548 			    mac_rx_srs->srs_pri, mcip, mac_rx_srs, cpuid,
1549 			    rx_func, x_arg1, x_arg2, B_TRUE);
1550 		}
1551 		mac_srs_update_fanout_list(mac_rx_srs);
1552 	} else if (new_fanout_cnt < srings_present) {
1553 		/* soft rings decreased */
1554 		if (new_fanout_cnt == 1) {
1555 			mutex_enter(&mac_rx_srs->srs_lock);
1556 			mac_rx_srs->srs_type &= ~SRST_FANOUT_SRC_IP;
1557 			ASSERT(mac_rx_srs->srs_type & SRST_FANOUT_PROTO);
1558 			mutex_exit(&mac_rx_srs->srs_lock);
1559 		}
1560 		/* Get rid of extra soft rings */
1561 		for (i = new_fanout_cnt;
1562 		    i < mac_rx_srs->srs_tcp_ring_count; i++) {
1563 			softring = mac_rx_srs->srs_tcp_soft_rings[i];
1564 			if (softring->s_ring_rx_arg2 != NULL) {
1565 				mcip->mci_resource_remove(
1566 				    (void *)mcip->mci_resource_arg,
1567 				    softring->s_ring_rx_arg2);
1568 			}
1569 			mac_soft_ring_remove(mac_rx_srs,
1570 			    mac_rx_srs->srs_tcp_soft_rings[i]);
1571 			mac_soft_ring_remove(mac_rx_srs,
1572 			    mac_rx_srs->srs_udp_soft_rings[i]);
1573 			mac_soft_ring_remove(mac_rx_srs,
1574 			    mac_rx_srs->srs_oth_soft_rings[i]);
1575 		}
1576 		mac_srs_update_fanout_list(mac_rx_srs);
1577 	}
1578 
1579 	ASSERT(new_fanout_cnt == mac_rx_srs->srs_tcp_ring_count);
1580 	mutex_enter(&cpu_lock);
1581 	for (i = 0; i < mac_rx_srs->srs_tcp_ring_count; i++) {
1582 		cpuid = srs_cpu->mc_fanout_cpus[i];
1583 		(void) mac_soft_ring_bind(mac_rx_srs->srs_udp_soft_rings[i],
1584 		    cpuid);
1585 		(void) mac_soft_ring_bind(mac_rx_srs->srs_oth_soft_rings[i],
1586 		    cpuid);
1587 		(void) mac_soft_ring_bind(mac_rx_srs->srs_tcp_soft_rings[i],
1588 		    cpuid);
1589 		softring = mac_rx_srs->srs_tcp_soft_rings[i];
1590 		if (softring->s_ring_rx_arg2 != NULL) {
1591 			mcip->mci_resource_bind((void *)mcip->mci_resource_arg,
1592 			    softring->s_ring_rx_arg2, cpuid);
1593 		}
1594 	}
1595 
1596 	mac_srs_worker_bind(mac_rx_srs, srs_cpu->mc_pollid);
1597 	mac_srs_poll_bind(mac_rx_srs, srs_cpu->mc_workerid);
1598 
1599 	/*
1600 	 * Bind Tx srs and soft ring threads too. Let's bind tx
1601 	 * srs to the last cpu in mrp list.
1602 	 */
1603 	if (mac_tx_srs != NULL && user_specified) {
1604 		BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1605 	}
1606 	mutex_exit(&cpu_lock);
1607 }
1608 
1609 /*
1610  * Bind SRS threads and soft rings to CPUs/create fanout list.
1611  */
1612 void
1613 mac_srs_fanout_init(mac_client_impl_t *mcip, flow_entry_t *flent,
1614     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1615     mac_resource_handle_t x_arg2, mac_soft_ring_set_t *mac_rx_srs,
1616     mac_soft_ring_set_t *mac_tx_srs)
1617 {
1618 	int		i;
1619 	processorid_t	cpuid, worker_cpuid, poll_cpuid;
1620 	uint32_t	soft_ring_flag = 0;
1621 	int soft_ring_cnt;
1622 	boolean_t user_specified = B_FALSE;
1623 	mac_cpus_t *srs_cpu = &mac_rx_srs->srs_cpu;
1624 
1625 	/*
1626 	 * Remove the no soft ring flag and we will adjust it
1627 	 * appropriately further down.
1628 	 */
1629 	mutex_enter(&mac_rx_srs->srs_lock);
1630 	mac_rx_srs->srs_type &= ~SRST_NO_SOFT_RINGS;
1631 	mutex_exit(&mac_rx_srs->srs_lock);
1632 
1633 	ASSERT(mac_rx_srs->srs_soft_ring_head == NULL);
1634 
1635 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1636 		soft_ring_flag |= ST_RING_BW_CTL;
1637 
1638 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_UNINIT);
1639 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1640 	user_specified = mrp->mrp_mask & MRP_CPUS_USERSPEC;
1641 	/*
1642 	 * Ring count can be 0 if no fanout is required and no cpu
1643 	 * were specified. Leave the SRS worker and poll thread
1644 	 * unbound
1645 	 */
1646 	ASSERT(mrp != NULL);
1647 	soft_ring_cnt = srs_cpu->mc_fanout_cnt;
1648 
1649 	/* Step 1: bind cpu contains cpu list where threads need to bind */
1650 	if (soft_ring_cnt > 0) {
1651 		mutex_enter(&cpu_lock);
1652 		for (i = 0; i < soft_ring_cnt; i++) {
1653 			cpuid = srs_cpu->mc_fanout_cpus[i];
1654 			/* Create the protocol softrings */
1655 			mac_srs_create_proto_softrings(i, (void *)flent,
1656 			    soft_ring_flag, mac_rx_srs->srs_pri,
1657 			    mcip, mac_rx_srs, cpuid, rx_func,
1658 			    x_arg1, x_arg2, B_FALSE);
1659 		}
1660 		worker_cpuid = srs_cpu->mc_workerid;
1661 		poll_cpuid = srs_cpu->mc_pollid;
1662 		mac_srs_worker_bind(mac_rx_srs, worker_cpuid);
1663 		mac_srs_poll_bind(mac_rx_srs, poll_cpuid);
1664 
1665 		/*
1666 		 * Bind Tx srs and soft ring threads too.
1667 		 * Let's bind tx srs to the last cpu in
1668 		 * mrp list.
1669 		 */
1670 		if (mac_tx_srs == NULL) {
1671 			mutex_exit(&cpu_lock);
1672 			goto alldone;
1673 		}
1674 
1675 		if (user_specified) {
1676 			BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1677 		}
1678 		mutex_exit(&cpu_lock);
1679 	} else {
1680 		mutex_enter(&cpu_lock);
1681 		/*
1682 		 * For a subflow, mrp_workerid and mrp_pollid
1683 		 * is not set.
1684 		 */
1685 		mac_srs_worker_bind(mac_rx_srs, mrp->mrp_workerid);
1686 		mac_srs_poll_bind(mac_rx_srs, mrp->mrp_pollid);
1687 		mutex_exit(&cpu_lock);
1688 		goto no_softrings;
1689 	}
1690 
1691 alldone:
1692 	if (soft_ring_cnt > 1)
1693 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1694 	mac_srs_update_fanout_list(mac_rx_srs);
1695 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1696 	return;
1697 
1698 no_softrings:
1699 	if (mac_rx_srs->srs_type & SRST_FANOUT_PROTO) {
1700 		mutex_enter(&cpu_lock);
1701 		cpuid = mac_next_bind_cpu();
1702 		/* Create the protocol softrings */
1703 		mac_srs_create_proto_softrings(0, (void *)flent,
1704 		    soft_ring_flag, mac_rx_srs->srs_pri,
1705 		    mcip, mac_rx_srs, cpuid, rx_func,
1706 		    x_arg1, x_arg2, B_FALSE);
1707 		mutex_exit(&cpu_lock);
1708 	} else {
1709 		/*
1710 		 * This is the case when there is no fanout which is
1711 		 * true for subflows.
1712 		 */
1713 		mac_rx_srs->srs_type |= SRST_NO_SOFT_RINGS;
1714 	}
1715 	mac_srs_update_fanout_list(mac_rx_srs);
1716 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1717 }
1718 
1719 /*
1720  * mac_fanout_setup:
1721  *
1722  * Calls mac_srs_fanout_init() or modify() depending upon whether
1723  * the SRS is getting initialized or re-initialized.
1724  */
1725 void
1726 mac_fanout_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
1727     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1728     mac_resource_handle_t x_arg2)
1729 {
1730 	mac_soft_ring_set_t *mac_rx_srs, *mac_tx_srs;
1731 	int i, rx_srs_cnt;
1732 
1733 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
1734 	/*
1735 	 * This is an aggregation port. Fanout will be setup
1736 	 * over the aggregation itself.
1737 	 */
1738 	if (mcip->mci_state_flags & MCIS_IS_AGGR_PORT)
1739 		return;
1740 
1741 	mac_rx_srs = flent->fe_rx_srs[0];
1742 	/*
1743 	 * Set up the fanout on the tx side only once, with the
1744 	 * first rx SRS. The CPU binding, fanout, and bandwidth
1745 	 * criteria are common to both RX and TX, so
1746 	 * initializing them along side avoids redundant code.
1747 	 */
1748 	mac_tx_srs = flent->fe_tx_srs;
1749 	rx_srs_cnt = flent->fe_rx_srs_cnt;
1750 
1751 	/* No fanout for subflows */
1752 	if (flent->fe_type & FLOW_USER) {
1753 		mac_srs_fanout_init(mcip, flent, mrp, rx_func,
1754 		    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1755 		return;
1756 	}
1757 
1758 	mac_flow_cpu_init(flent, mrp);
1759 
1760 	/*
1761 	 * Set up fanout for both SW (0th SRS) and HW classified
1762 	 * SRS (the rest of Rx SRSs in flent).
1763 	 */
1764 	for (i = 0; i < rx_srs_cnt; i++) {
1765 		mac_rx_srs = flent->fe_rx_srs[i];
1766 		if (i != 0)
1767 			mac_tx_srs = NULL;
1768 		switch (mac_rx_srs->srs_fanout_state) {
1769 		case SRS_FANOUT_UNINIT:
1770 			mac_srs_fanout_init(mcip, flent, mrp, rx_func,
1771 			    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1772 			break;
1773 		case SRS_FANOUT_INIT:
1774 			break;
1775 		case SRS_FANOUT_REINIT:
1776 			mac_rx_srs_quiesce(mac_rx_srs, SRS_QUIESCE);
1777 			mac_srs_fanout_modify(mcip, flent, mrp, rx_func,
1778 			    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs);
1779 			mac_rx_srs_restart(mac_rx_srs);
1780 			break;
1781 		default:
1782 			VERIFY(mac_rx_srs->srs_fanout_state <=
1783 			    SRS_FANOUT_REINIT);
1784 			break;
1785 		}
1786 	}
1787 }
1788 
1789 /*
1790  * mac_create_soft_ring_set:
1791  *
1792  * Create a mac_soft_ring_set_t (SRS). If soft_ring_fanout_type is
1793  * SRST_TX, an SRS for Tx side is created. Otherwise an SRS for Rx side
1794  * processing is created.
1795  *
1796  * Details on Rx SRS:
1797  * Create a SRS and also add the necessary soft rings for TCP and
1798  * non-TCP based on fanout type and count specified.
1799  *
1800  * mac_soft_ring_fanout, mac_srs_fanout_modify (?),
1801  * mac_soft_ring_stop_workers, mac_soft_ring_set_destroy, etc need
1802  * to be heavily modified.
1803  *
1804  * mi_soft_ring_list_size, mi_soft_ring_size, etc need to disappear.
1805  */
1806 mac_soft_ring_set_t *
1807 mac_srs_create(mac_client_impl_t *mcip, flow_entry_t *flent, uint32_t srs_type,
1808     mac_direct_rx_t rx_func, void *x_arg1, mac_resource_handle_t x_arg2,
1809     mac_ring_t *ring)
1810 {
1811 	mac_soft_ring_set_t 	*mac_srs;
1812 	mac_srs_rx_t		*srs_rx;
1813 	mac_srs_tx_t		*srs_tx;
1814 	mac_bw_ctl_t		*mac_bw;
1815 	mac_resource_props_t	*mrp;
1816 	boolean_t		is_tx_srs = ((srs_type & SRST_TX) != 0);
1817 
1818 	mac_srs = kmem_cache_alloc(mac_srs_cache, KM_SLEEP);
1819 	bzero(mac_srs, sizeof (mac_soft_ring_set_t));
1820 	srs_rx = &mac_srs->srs_rx;
1821 	srs_tx = &mac_srs->srs_tx;
1822 
1823 	mutex_enter(&flent->fe_lock);
1824 
1825 	/*
1826 	 * Get the bandwidth control structure from the flent. Get
1827 	 * rid of any residual values in the control structure for
1828 	 * the tx bw struct and also for the rx, if the rx srs is
1829 	 * the 1st one being brought up (the rx bw ctl struct may
1830 	 * be shared by multiple SRSs)
1831 	 */
1832 	if (is_tx_srs) {
1833 		mac_srs->srs_bw = &flent->fe_tx_bw;
1834 		bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
1835 		flent->fe_tx_srs = mac_srs;
1836 	} else {
1837 		/*
1838 		 * The bw counter (stored in the flent) is shared
1839 		 * by SRS's within an rx group.
1840 		 */
1841 		mac_srs->srs_bw = &flent->fe_rx_bw;
1842 		/* First rx SRS, clear the bw structure */
1843 		if (flent->fe_rx_srs_cnt == 0)
1844 			bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
1845 		ASSERT(flent->fe_rx_srs_cnt < MAX_RINGS_PER_GROUP);
1846 		flent->fe_rx_srs[flent->fe_rx_srs_cnt] = mac_srs;
1847 		flent->fe_rx_srs_cnt++;
1848 	}
1849 	mac_srs->srs_flent = flent;
1850 	mutex_exit(&flent->fe_lock);
1851 
1852 	mac_srs->srs_state = 0;
1853 	mac_srs->srs_type = (srs_type | SRST_NO_SOFT_RINGS);
1854 	mac_srs->srs_worker_cpuid = mac_srs->srs_worker_cpuid_save = -1;
1855 	mac_srs->srs_poll_cpuid = mac_srs->srs_poll_cpuid_save = -1;
1856 	mac_srs_fanout_list_alloc(mac_srs);
1857 
1858 	/*
1859 	 * For a flow we use the underlying MAC client's priority range with
1860 	 * the priority value to find an absolute priority value. For a MAC
1861 	 * client we use the MAC client's maximum priority as the value.
1862 	 */
1863 	mrp = &flent->fe_effective_props;
1864 	if ((mac_srs->srs_type & SRST_FLOW) != 0) {
1865 		mac_srs->srs_pri = FLOW_PRIORITY(mcip->mci_min_pri,
1866 		    mcip->mci_max_pri, mrp->mrp_priority);
1867 	} else {
1868 		mac_srs->srs_pri = mcip->mci_max_pri;
1869 	}
1870 	mac_srs->srs_mcip = mcip;
1871 	/*
1872 	 * We need to insert the SRS in the global list before
1873 	 * binding the SRS and SR threads. Otherwise there is a
1874 	 * is a small window where the cpu reconfig callbacks
1875 	 * may miss the SRS in the list walk and DR could fail
1876 	 * as there are bound threads.
1877 	 */
1878 	mac_srs_add_glist(mac_srs);
1879 
1880 	/* Initialize bw limit */
1881 	if ((mrp->mrp_mask & MRP_MAXBW) != 0) {
1882 		mac_srs->srs_drain_func = mac_rx_srs_drain_bw;
1883 
1884 		mac_bw = mac_srs->srs_bw;
1885 		mutex_enter(&mac_bw->mac_bw_lock);
1886 		mac_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1887 
1888 		/*
1889 		 * Give twice the queuing capability before
1890 		 * dropping packets. The unit is bytes/tick.
1891 		 */
1892 		mac_bw->mac_bw_drop_threshold = mac_bw->mac_bw_limit << 1;
1893 		mutex_exit(&mac_bw->mac_bw_lock);
1894 		mac_srs->srs_type |= SRST_BW_CONTROL;
1895 	} else {
1896 		mac_srs->srs_drain_func = mac_rx_srs_drain;
1897 	}
1898 
1899 	/*
1900 	 * We use the following policy to control Receive
1901 	 * Side Dynamic Polling:
1902 	 * 1) We switch to poll mode anytime the processing thread causes
1903 	 *    a backlog to build up in SRS and its associated Soft Rings
1904 	 *    (sr_poll_pkt_cnt > 0).
1905 	 * 2) As long as the backlog stays under the low water mark
1906 	 *    (sr_lowat), we poll the H/W for more packets.
1907 	 * 3) If the backlog (sr_poll_pkt_cnt) exceeds low water mark, we
1908 	 *    stay in poll mode but don't poll the H/W for more packets.
1909 	 * 4) Anytime in polling mode, if we poll the H/W for packets and
1910 	 *    find nothing plus we have an existing backlog
1911 	 *    (sr_poll_pkt_cnt > 0), we stay in polling mode but don't poll
1912 	 *    the H/W for packets anymore (let the polling thread go to sleep).
1913 	 * 5) Once the backlog is relived (packets are processed) we reenable
1914 	 *    polling (by signalling the poll thread) only when the backlog
1915 	 *    dips below sr_poll_thres.
1916 	 * 6) sr_hiwat is used exclusively when we are not polling capable
1917 	 *    and is used to decide when to drop packets so the SRS queue
1918 	 *    length doesn't grow infinitely.
1919 	 */
1920 	if (!is_tx_srs) {
1921 		srs_rx->sr_hiwat = mac_soft_ring_max_q_cnt;
1922 		/* Low water mark needs to be less than high water mark */
1923 		srs_rx->sr_lowat = mac_soft_ring_min_q_cnt <=
1924 		    mac_soft_ring_max_q_cnt ? mac_soft_ring_min_q_cnt :
1925 		    (mac_soft_ring_max_q_cnt >> 2);
1926 		/* Poll threshold need to be half of low water mark or less */
1927 		srs_rx->sr_poll_thres = mac_soft_ring_poll_thres <=
1928 		    (srs_rx->sr_lowat >> 1) ? mac_soft_ring_poll_thres :
1929 		    (srs_rx->sr_lowat >> 1);
1930 		if (mac_latency_optimize)
1931 			mac_srs->srs_state |=
1932 			    (SRS_LATENCY_OPT|SRS_SOFTRING_QUEUE);
1933 	}
1934 
1935 	mac_srs->srs_worker = thread_create(NULL, 0,
1936 	    mac_srs_worker, mac_srs, 0, &p0, TS_RUN, mac_srs->srs_pri);
1937 
1938 	if (is_tx_srs) {
1939 		/* Handle everything about Tx SRS and return */
1940 		mac_srs->srs_drain_func = mac_tx_srs_drain;
1941 		srs_tx->st_max_q_cnt = mac_tx_srs_max_q_cnt;
1942 		srs_tx->st_hiwat =
1943 		    (mac_tx_srs_hiwat > mac_tx_srs_max_q_cnt) ?
1944 		    mac_tx_srs_max_q_cnt : mac_tx_srs_hiwat;
1945 		srs_tx->st_arg1 = x_arg1;
1946 		srs_tx->st_arg2 = x_arg2;
1947 		return (mac_srs);
1948 	}
1949 
1950 	if ((srs_type & SRST_FLOW) != 0 ||
1951 	    FLOW_TAB_EMPTY(mcip->mci_subflow_tab))
1952 		srs_rx->sr_lower_proc = mac_rx_srs_process;
1953 	else
1954 		srs_rx->sr_lower_proc = mac_rx_srs_subflow_process;
1955 
1956 	srs_rx->sr_func = rx_func;
1957 	srs_rx->sr_arg1 = x_arg1;
1958 	srs_rx->sr_arg2 = x_arg2;
1959 
1960 	if (ring != NULL) {
1961 		/* Is the mac_srs created over the RX default group? */
1962 		if (ring->mr_gh == (mac_group_handle_t)
1963 		    (&mcip->mci_mip->mi_rx_groups[0]))
1964 			mac_srs->srs_type |= SRST_DEFAULT_GRP;
1965 
1966 		mac_srs->srs_ring = ring;
1967 		ring->mr_srs = mac_srs;
1968 		ring->mr_classify_type = MAC_HW_CLASSIFIER;
1969 		ring->mr_flag |= MR_INCIPIENT;
1970 
1971 		if (FLOW_TAB_EMPTY(mcip->mci_subflow_tab) && mac_poll_enable)
1972 			mac_srs->srs_state |= SRS_POLLING_CAPAB;
1973 
1974 		mac_srs->srs_poll_thr = thread_create(NULL, 0,
1975 		    mac_rx_srs_poll_ring, mac_srs, 0, &p0, TS_RUN,
1976 		    mac_srs->srs_pri);
1977 		/*
1978 		 * Some drivers require serialization and don't send
1979 		 * packet chains in interrupt context. For such
1980 		 * drivers, we should always queue in soft ring
1981 		 * so that we get a chance to switch into a polling
1982 		 * mode under backlog.
1983 		 */
1984 		if (mcip->mci_mip->mi_v12n_level & MAC_VIRT_SERIALIZE)
1985 			mac_srs->srs_state |= SRS_SOFTRING_QUEUE;
1986 	}
1987 	return (mac_srs);
1988 }
1989 
1990 /*
1991  * Figure out the number of soft rings required. Its dependant on
1992  * if protocol fanout is required (for LINKs), global settings
1993  * require us to do fanout for performance (based on mac_soft_ring_enable),
1994  * or user has specifically requested fanout.
1995  */
1996 static uint32_t
1997 mac_find_fanout(flow_entry_t *flent, uint32_t link_type)
1998 {
1999 	uint32_t			fanout_type;
2000 	mac_resource_props_t		*mrp = &flent->fe_effective_props;
2001 
2002 	/* no fanout for subflows */
2003 	switch (link_type) {
2004 	case SRST_FLOW:
2005 		fanout_type = SRST_NO_SOFT_RINGS;
2006 		break;
2007 	case SRST_LINK:
2008 		fanout_type = SRST_FANOUT_PROTO;
2009 		break;
2010 	}
2011 
2012 	/* A primary NIC/link is being plumbed */
2013 	if (flent->fe_type & FLOW_PRIMARY_MAC) {
2014 		if (mac_soft_ring_enable && mac_rx_soft_ring_count > 1) {
2015 			fanout_type |= SRST_FANOUT_SRC_IP;
2016 		}
2017 	} else if (flent->fe_type & FLOW_VNIC) {
2018 		/* A VNIC is being created */
2019 		if (mrp != NULL && mrp->mrp_ncpus > 0) {
2020 			fanout_type |= SRST_FANOUT_SRC_IP;
2021 		}
2022 	}
2023 
2024 	return (fanout_type);
2025 }
2026 
2027 /*
2028  * Change a group from h/w to s/w classification.
2029  */
2030 static void
2031 mac_rx_switch_grp_to_sw(mac_group_t *group)
2032 {
2033 	mac_ring_t		*ring;
2034 	mac_soft_ring_set_t	*mac_srs;
2035 
2036 	for (ring = group->mrg_rings; ring != NULL; ring = ring->mr_next) {
2037 		if (ring->mr_classify_type == MAC_HW_CLASSIFIER) {
2038 			/*
2039 			 * Remove the SRS associated with the HW ring.
2040 			 * As a result, polling will be disabled.
2041 			 */
2042 			mac_srs = ring->mr_srs;
2043 			ASSERT(mac_srs != NULL);
2044 			mac_rx_srs_remove(mac_srs);
2045 			ring->mr_srs = NULL;
2046 		}
2047 
2048 		if (ring->mr_state != MR_INUSE)
2049 			(void) mac_start_ring(ring);
2050 		/*
2051 		 * We need to perform SW classification
2052 		 * for packets landing in these rings
2053 		 */
2054 		ring->mr_state = MR_INUSE;
2055 		ring->mr_flag = 0;
2056 		ring->mr_classify_type = MAC_SW_CLASSIFIER;
2057 	}
2058 }
2059 
2060 /*
2061  * Create the Rx SRS for S/W classifier and for each ring in the
2062  * group (if exclusive group). Also create the Tx SRS.
2063  */
2064 void
2065 mac_srs_group_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2066     mac_group_t *group, uint32_t link_type)
2067 {
2068 	mac_impl_t		*mip = mcip->mci_mip;
2069 	mac_soft_ring_set_t	*mac_srs;
2070 	mac_soft_ring_set_t	*tx_srs = NULL;
2071 	mac_ring_t 		*ring;
2072 	uint32_t		fanout_type;
2073 	boolean_t		created_srs = B_FALSE;
2074 
2075 	fanout_type = mac_find_fanout(flent, link_type);
2076 
2077 	/* Create the SRS for S/W classification if none exists */
2078 	if (flent->fe_rx_srs[0] == NULL) {
2079 		ASSERT(flent->fe_rx_srs_cnt == 0);
2080 		/* Setup the Rx SRS */
2081 		mac_srs = mac_srs_create(mcip, flent, fanout_type | link_type,
2082 		    mac_rx_deliver, mcip, NULL, NULL);
2083 
2084 		mutex_enter(&flent->fe_lock);
2085 		flent->fe_cb_fn = (flow_fn_t)mac_srs->srs_rx.sr_lower_proc;
2086 		flent->fe_cb_arg1 = (void *)mip;
2087 		flent->fe_cb_arg2 = (void *)mac_srs;
2088 		mutex_exit(&flent->fe_lock);
2089 
2090 		/* Setup the Tx SRS as well */
2091 		ASSERT(flent->fe_tx_srs == NULL);
2092 		tx_srs = mac_srs_create(mcip, flent, SRST_TX | link_type,
2093 		    NULL, mcip, NULL, NULL);
2094 
2095 		if (mcip->mci_share != NULL) {
2096 			mac_srs_tx_t	*tx = &tx_srs->srs_tx;
2097 			ASSERT((mcip->mci_state_flags & MCIS_NO_HWRINGS) == 0);
2098 			/*
2099 			 * A share requires a dedicated TX group.
2100 			 * mac_reserve_tx_group() does the work needed to
2101 			 * allocate a new group and populate that group
2102 			 * with rings according to the driver requirements
2103 			 * and limitations.
2104 			 */
2105 			tx->st_group =
2106 			    mac_reserve_tx_group(mip, mcip->mci_share);
2107 			ASSERT(tx->st_group != NULL);
2108 			tx->st_group->mrg_tx_client = mcip;
2109 		}
2110 		mac_tx_srs_setup(mcip, flent, link_type);
2111 		created_srs = B_TRUE;
2112 	}
2113 
2114 	if (group == NULL) {
2115 		if (created_srs) {
2116 			mac_fanout_setup(mcip, flent,
2117 			    MCIP_RESOURCE_PROPS(mcip), mac_rx_deliver,
2118 			    mcip, NULL);
2119 		}
2120 		return;
2121 	}
2122 
2123 	/*
2124 	 * fanout for default SRS is done when default SRS are created
2125 	 * above. As each ring is added to the group, we setup the
2126 	 * SRS and fanout to it.
2127 	 */
2128 	switch (group->mrg_state) {
2129 	case MAC_GROUP_STATE_RESERVED:
2130 		/*
2131 		 * The group is exclusively ours. Create a SRS
2132 		 * for each ring in the group and allow the
2133 		 * individual SRS to dynamically poll their
2134 		 * Rx ring. Do this only if the  client is not
2135 		 * a VLAN MAC client since for VLAN we do
2136 		 * s/w classification for the VID check.
2137 		 */
2138 		if (i_mac_flow_vid(mcip->mci_flent) != VLAN_ID_NONE)
2139 			break;
2140 		for (ring = group->mrg_rings; ring != NULL;
2141 		    ring = ring->mr_next) {
2142 			switch (ring->mr_state) {
2143 			case MR_INUSE:
2144 			case MR_FREE:
2145 				if (ring->mr_srs != NULL)
2146 					break;
2147 				if (ring->mr_state != MR_INUSE)
2148 					(void) mac_start_ring(ring);
2149 
2150 				ring->mr_state = MR_INUSE;
2151 
2152 				mac_srs = mac_srs_create(mcip, flent,
2153 				    fanout_type | link_type,
2154 				    mac_rx_deliver, mcip, NULL, ring);
2155 				break;
2156 			default:
2157 				cmn_err(CE_PANIC, "srs_setup: mcip = %p "
2158 				    "trying to add UNKNOWN ring = %p\n",
2159 				    (void *)mcip, (void *)ring);
2160 				break;
2161 			}
2162 		}
2163 		break;
2164 	case MAC_GROUP_STATE_SHARED:
2165 		/*
2166 		 * Set all rings of this group to software classified.
2167 		 *
2168 		 * If the group is current RESERVED, the existing mac client
2169 		 * (the only client on this group) is using this group
2170 		 * exclusively.  In that case we need to disable polling on
2171 		 * the rings of the group (if it was enabled), and free the
2172 		 * SRS associated with the rings.
2173 		 */
2174 		mac_rx_switch_grp_to_sw(group);
2175 		break;
2176 	default:
2177 		ASSERT(B_FALSE);
2178 		break;
2179 	}
2180 	mac_fanout_setup(mcip, flent, MCIP_RESOURCE_PROPS(mcip),
2181 	    mac_rx_deliver, mcip, NULL);
2182 }
2183 
2184 void
2185 mac_srs_group_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
2186     uint32_t link_type)
2187 {
2188 	mac_soft_ring_set_t	*mac_srs;
2189 	mac_soft_ring_set_t	*tx_srs;
2190 	mac_srs_tx_t		*tx;
2191 	int			i;
2192 
2193 	for (i = 0; i < flent->fe_rx_srs_cnt; i++) {
2194 		mac_srs = flent->fe_rx_srs[i];
2195 		mac_rx_srs_quiesce(mac_srs, SRS_CONDEMNED);
2196 		/*
2197 		 * Deal with all fanout tear down etc.
2198 		 */
2199 		mac_srs_free(mac_srs);
2200 		flent->fe_rx_srs[i] = NULL;
2201 	}
2202 	flent->fe_rx_srs_cnt = 0;
2203 
2204 	tx_srs = flent->fe_tx_srs;
2205 	tx = &tx_srs->srs_tx;
2206 	switch (link_type) {
2207 	case SRST_FLOW:
2208 		/*
2209 		 * For flows, we need to work with passed
2210 		 * flent to find the Rx/Tx SRS.
2211 		 */
2212 		mac_tx_srs_quiesce(tx_srs, SRS_CONDEMNED);
2213 		break;
2214 	case SRST_LINK:
2215 		mac_tx_client_quiesce(mcip, SRS_CONDEMNED);
2216 		/*
2217 		 * Release the TX resources. First the TX group, if any
2218 		 * was assigned to the MAC client, which will cause the
2219 		 * TX rings to be moved back to the pool. Then free the
2220 		 * rings themselves.
2221 		 */
2222 		if (tx->st_group != NULL) {
2223 			mac_release_tx_group(tx_srs->srs_mcip->mci_mip,
2224 			    tx->st_group);
2225 			tx->st_group = NULL;
2226 		}
2227 		if (tx->st_arg2 != NULL) {
2228 			ASSERT(tx_srs->srs_type & SRST_TX);
2229 			mac_release_tx_ring(tx->st_arg2);
2230 		}
2231 		break;
2232 	default:
2233 		ASSERT(B_FALSE);
2234 		break;
2235 	}
2236 	mac_srs_free(tx_srs);
2237 	flent->fe_tx_srs = NULL;
2238 }
2239 
2240 /*
2241  * This is the group state machine. The state of an Rx group is given by
2242  * the following table. The default group and its rings are started in
2243  * mac_start itself and the default group stays in SHARED state until
2244  * mac_stop at which time the group and rings are stopped and and it
2245  * reverts to the Registered state.
2246  *
2247  * Typically this function is called on a group after adding or removing a
2248  * client from it, to find out what should be the new state of the group.
2249  * If the new state is RESERVED, then the client that owns this group
2250  * exclusively is also returned. Note that adding or removing a client from
2251  * a group could also impact the default group and the caller needs to
2252  * evaluate the effect on the default group.
2253  *
2254  * Group type		# of clients	mi_nactiveclients	Group State
2255  *			in the group
2256  *
2257  * Non-default		0		N.A.			REGISTERED
2258  * Non-default		1		N.A.			RESERVED
2259  * Non-default		> 1		N.A.			SHARED
2260  *
2261  * Default		0		N.A.			SHARED
2262  * Default		1		1			RESERVED
2263  * Default		1		> 1			SHARED
2264  * Default		> 1		N.A.			SHARED
2265  */
2266 mac_group_state_t
2267 mac_rx_group_next_state(mac_group_t *grp, mac_client_impl_t **group_only_mcip)
2268 {
2269 	mac_impl_t		*mip = (mac_impl_t *)grp->mrg_mh;
2270 
2271 	*group_only_mcip = NULL;
2272 
2273 	/* Non-default group */
2274 
2275 	if (grp != mip->mi_rx_groups) {
2276 		if (MAC_RX_GROUP_NO_CLIENT(grp))
2277 			return (MAC_GROUP_STATE_REGISTERED);
2278 
2279 		*group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(grp);
2280 		if (*group_only_mcip != NULL)
2281 			return (MAC_GROUP_STATE_RESERVED);
2282 
2283 		return (MAC_GROUP_STATE_SHARED);
2284 	}
2285 
2286 	/* Default group */
2287 
2288 	if (MAC_RX_GROUP_NO_CLIENT(grp) || mip->mi_nactiveclients != 1)
2289 		return (MAC_GROUP_STATE_SHARED);
2290 
2291 	*group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(grp);
2292 	ASSERT(*group_only_mcip != NULL);
2293 	return (MAC_GROUP_STATE_RESERVED);
2294 }
2295 
2296 /*
2297  * OVERVIEW NOTES FOR DATAPATH
2298  * ===========================
2299  *
2300  * Create an SRS and setup the corresponding flow function and args.
2301  * Add a classification rule for the flow specified by 'flent' and program
2302  * the hardware classifier when applicable.
2303  *
2304  * Rx ring assignment, SRS, polling and B/W enforcement
2305  * ----------------------------------------------------
2306  *
2307  * We try to use H/W classification on NIC and assign traffic to a
2308  * MAC address to a particular Rx ring. There is a 1-1 mapping
2309  * between a SRS and a Rx ring. The SRS (short for soft ring set)
2310  * dynamically switches the underlying Rx ring between interrupt
2311  * and polling mode and enforces any specified B/W control.
2312  *
2313  * There is always a SRS created and tied to each H/W and S/W rule.
2314  * Whenever we create a H/W rule, we always add the the same rule to
2315  * S/W classifier and tie a SRS to it.
2316  *
2317  * In case a B/W control is specified, its broken into bytes
2318  * per ticks and as soon as the quota for a tick is exhausted,
2319  * the underlying Rx ring is forced into poll mode for remianing
2320  * tick. The SRS poll thread only polls for bytes that are
2321  * allowed to come in the SRS. We typically let 4x the configured
2322  * B/W worth of packets to come in the SRS (to prevent unnecessary
2323  * drops due to bursts) but only process the specified amount.
2324  *
2325  * A Link (primary NIC, VNIC, VLAN or aggr) can have 1 or more
2326  * Rx rings (and corresponding SRSs) assigned to it. The SRS
2327  * in turn can have softrings to do protocol level fanout or
2328  * softrings to do S/W based fanout or both. In case the NIC
2329  * has no Rx rings, we do S/W classification to respective SRS.
2330  * The S/W classification rule is always setup and ready. This
2331  * allows the MAC layer to reassign Rx rings whenever needed
2332  * but packets still continue to flow via the default path and
2333  * getting S/W classified to correct SRS.
2334  *
2335  * In other cases where a NIC or VNIC is plumbed, our goal is use
2336  * H/W classifier and get two Rx ring assigned for the Link. One
2337  * for TCP and one for UDP|SCTP. The respective SRS still do the
2338  * polling on the Rx ring. For Link that is plumbed for IP, there
2339  * is a TCP squeue which also does polling and can control the
2340  * the Rx ring directly (where SRS is just pass through). For
2341  * the following cases, the SRS does the polling underneath.
2342  * 1) non IP based Links (Links which are not plumbed via ifconfig)
2343  *    and paths which have no IP squeues (UDP & SCTP)
2344  * 2) If B/W control is specified on the Link
2345  * 3) If S/W fanout is secified
2346  *
2347  * Note1: As of current implementation, we try to assign only 1 Rx
2348  * ring per Link and more than 1 Rx ring for primary Link for
2349  * H/W based fanout. We always create following softrings per SRS:
2350  * 1) TCP softring which is polled by TCP squeue where possible
2351  *    (and also bypasses DLS)
2352  * 2) UDP/SCTP based which bypasses DLS
2353  * 3) OTH softring which goes via DLS (currently deal with IPv6
2354  *    and non TCP/UDP/SCTP for IPv4 packets).
2355  *
2356  * It is necessary to create 3 softrings since SRS has to poll
2357  * the single Rx ring underneath and enforce any link level B/W
2358  * control (we can't switch the Rx ring in poll mode just based
2359  * on TCP squeue if the same Rx ring is sharing UDP and other
2360  * traffic as well). Once polling is done and any Link level B/W
2361  * control is specified, the packets are assigned to respective
2362  * softring based on protocol. Since TCP has IP based squeue
2363  * which benefits by polling, we separate TCP packets into
2364  * its own softring which can be polled by IP squeue. We need
2365  * to separate out UDP/SCTP to UDP softring since it can bypass
2366  * the DLS layer which has heavy performance advanatges and we
2367  * need a softring (OTH) for rest.
2368  *
2369  * ToDo: The 3 softrings for protocol are needed only till we can
2370  * get rid of DLS from datapath, make IPv4 and IPv6 paths
2371  * symmetric (deal with mac_header_info for v6 and polling for
2372  * IPv4 TCP - ip_accept_tcp is IPv4 specific although squeues
2373  * are generic), and bring SAP based classification to MAC layer
2374  *
2375  * H/W and S/W based fanout and multiple Rx rings per Link
2376  * -------------------------------------------------------
2377  *
2378  * In case, fanout is requested (or determined automatically based
2379  * on Link speed and processor speed), we try to assign multiple
2380  * Rx rings per Link with their respective SRS. In this case
2381  * the NIC should be capable of fanning out incoming packets between
2382  * the assigned Rx rings (H/W based fanout). All the SRS
2383  * individually switch their Rx ring between interrupt and polling
2384  * mode but share a common B/W control counter in case of Link
2385  * level B/W is specified.
2386  *
2387  * If S/W based fanout is specified in lieu of H/W based fanout,
2388  * the Link SRS creates the specified number of softrings for
2389  * each protocol (TCP, UDP, OTH). Incoming packets are fanned
2390  * out to the correct softring based on their protocol and
2391  * protocol specific hash function.
2392  *
2393  * Primary and non primary MAC clients
2394  * -----------------------------------
2395  *
2396  * The NICs, VNICs, Vlans, and Aggrs are typically termed as Links
2397  * and are a Layer 2 construct.
2398  *
2399  * Primary NIC:
2400  *	The Link that owns the primary MAC address and typically
2401  *	is used as the data NIC in non virtualized cases. As such
2402  *	H/W resources are preferntially given to primary NIC. As
2403  *	far as code is concerned, there is no difference in the
2404  *	primary NIC vs VNICs. They are all treated as Links.
2405  *	At the very first call to mac_unicast_add() we program the S/W
2406  *	classifier for the primary MAC address, get a soft ring set
2407  *	(and soft rings based on 'ip_soft_ring_cnt')
2408  *	and a Rx ring assigned for polling to get enabled.
2409  *	When IP get plumbed and negotiates polling, we can
2410  *	let squeue do the polling on TCP softring.
2411  *
2412  * VNICs:
2413  *	Same as any other Link. As long as the H/W resource assignments
2414  *	are equal, the data path and setup for all Links is same.
2415  *
2416  * Flows:
2417  *	Can be configured on Links. They have their own SRS and the
2418  *	S/W classifier is programmed appropriately based on the flow.
2419  *	The flows typically deal with layer 3 and above and
2420  *	creates a soft ring set specific to the flow. The receive
2421  *	side function is switched from mac_rx_srs_process to
2422  *	mac_rx_srs_subflow_process which first tries to assign the
2423  *	packet to appropriate flow SRS and failing which assigns it
2424  *	to link SRS. This allows us to avoid the layered approach
2425  *	which gets complex.
2426  *
2427  * By the time mac_datapath_setup() completes, we already have the
2428  * soft rings set, Rx rings, soft rings, etc figured out and both H/W
2429  * and S/W classifiers programmed. IP is not plumbed yet (and might
2430  * never be for Virtual Machines guest OS path). When IP is plumbed
2431  * (for both NIC and VNIC), we do a capability negotiation for polling
2432  * and upcall functions etc.
2433  *
2434  * Rx ring Assignement NOTES
2435  * -------------------------
2436  *
2437  * For NICs which have only 1 Rx ring (we treat  NICs with no Rx rings
2438  * as NIC with a single default ring), we assign the only ring to
2439  * primary Link as MAC_RX_HW_DEFAULT_RING. The primary Link SRS can do
2440  * polling on it as long as it is the only link in use and we compare
2441  * the MAC address for unicast packets before accepting an incoming
2442  * packet (there is no need for S/W classification in this case). We
2443  * disable polling on the only ring the moment 2nd link gets created
2444  * (the polling remains enabled even though there are broadcast and
2445  * multicast flows created).
2446  *
2447  * If the NIC has more than 1 Rx ring, we assign the default ring (the
2448  * 1st ring) to deal with broadcast, multicast and traffic for other
2449  * NICs which needs S/W classification. We assign the primary mac
2450  * addresses to another ring by specifiying a classification rule for
2451  * primary unicast MAC address to the selected ring. The primary Link
2452  * (and its SRS) can continue to poll the assigned Rx ring at all times
2453  * independantly.
2454  *
2455  * Right now we just assign MAC_RX_HW_DEFAULT_RING to note that it is
2456  * primary NIC and later we will check to see how many Rx rings we
2457  * have and can we get a non default Rx ring for the primary MAC.
2458  *
2459  * Note: In future, if no fanout is specified, we try to assign 2 Rx
2460  * rings for the primary Link with the primary MAC address + TCP going
2461  * to one ring and primary MAC address + UDP|SCTP going to other ring.
2462  * Any remaining traffic for primary MAC address can go to the default
2463  * Rx ring and get S/W classified. This way the respective SRSs don't
2464  * need to do proto fanout and don't need to have softrings at all and
2465  * can poll their respective Rx rings.
2466  *
2467  * As an optimization, when a new NIC or VNIC is created, we can get
2468  * only one Rx ring and make it a TCP specific Rx ring and use the
2469  * H/W default Rx ring for the rest (this Rx ring is never polled).
2470  */
2471 int
2472 mac_datapath_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2473     uint32_t link_type)
2474 {
2475 	mac_impl_t		*mip = mcip->mci_mip;
2476 	mac_group_t		*group = NULL;
2477 	mac_group_t		*default_group;
2478 	int			err;
2479 	uint8_t 		*mac_addr;
2480 	mac_rx_group_reserve_type_t	rtype = MAC_RX_RESERVE_NONDEFAULT;
2481 	mac_group_state_t	next_state;
2482 	mac_client_impl_t	*group_only_mcip;
2483 
2484 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
2485 
2486 	switch (link_type) {
2487 	case SRST_FLOW:
2488 		mac_srs_group_setup(mcip, flent, NULL, link_type);
2489 		return (0);
2490 
2491 	case SRST_LINK:
2492 		mac_addr = flent->fe_flow_desc.fd_dst_mac;
2493 
2494 		/* Check if we need to reserve the default group */
2495 		if (flent->fe_type & FLOW_PRIMARY_MAC)
2496 			rtype = MAC_RX_RESERVE_DEFAULT;
2497 
2498 		if ((mcip->mci_state_flags & MCIS_NO_HWRINGS) == 0) {
2499 			/*
2500 			 * Check to see if we can get an exclusive group for
2501 			 * this mac address or if there already exists a
2502 			 * group that has this mac address (case of VLANs).
2503 			 * If no groups are available, use the default group.
2504 			 */
2505 			group = mac_reserve_rx_group(mcip, mac_addr, rtype);
2506 		}
2507 
2508 		if (group == NULL) {
2509 			if ((mcip->mci_state_flags & MCIS_REQ_HWRINGS) != 0)
2510 				return (ENOSPC);
2511 			group = &mip->mi_rx_groups[0];
2512 		}
2513 
2514 		/*
2515 		 * Some NICs don't support any Rx rings, so there may not
2516 		 * even be a default group.
2517 		 */
2518 		if (group != NULL) {
2519 			flent->fe_rx_ring_group = group;
2520 			/*
2521 			 * Add the client to the group. This could cause
2522 			 * either this group to move to the shared state or
2523 			 * cause the default group to move to the shared state.
2524 			 * The actions on this group are done here, while the
2525 			 * actions on the default group are postponed to
2526 			 * the end of this function.
2527 			 */
2528 			mac_rx_group_add_client(group, mcip);
2529 			next_state = mac_rx_group_next_state(group,
2530 			    &group_only_mcip);
2531 
2532 			ASSERT((next_state == MAC_GROUP_STATE_RESERVED &&
2533 			    mcip == group_only_mcip) ||
2534 			    (next_state == MAC_GROUP_STATE_SHARED &&
2535 			    group_only_mcip == NULL));
2536 
2537 			mac_set_rx_group_state(group, next_state);
2538 		}
2539 
2540 		/*
2541 		 * Setup the Rx and Tx SRSes. If we got a pristine group
2542 		 * exclusively above, mac_srs_group_setup would simply create
2543 		 * the required SRSes. If we ended up sharing a previously
2544 		 * reserved group, mac_srs_group_setup would also dismantle the
2545 		 * SRSes of the previously exclusive group
2546 		 */
2547 		mac_srs_group_setup(mcip, flent, group, link_type);
2548 
2549 		/* Program the S/W Classifer */
2550 		if ((err = mac_flow_add(mip->mi_flow_tab, flent)) != 0)
2551 			goto setup_failed;
2552 
2553 		/* Program the H/W Classifier */
2554 		if ((err = mac_add_macaddr(mip, group, mac_addr,
2555 		    (mcip->mci_state_flags & MCIS_UNICAST_HW) != 0)) != 0)
2556 			goto setup_failed;
2557 		mcip->mci_unicast = mac_find_macaddr(mip, mac_addr);
2558 		ASSERT(mcip->mci_unicast != NULL);
2559 		break;
2560 
2561 	default:
2562 		ASSERT(B_FALSE);
2563 		break;
2564 	}
2565 
2566 	/*
2567 	 * All broadcast and multicast traffic is received only on the default
2568 	 * group. If we have setup the datapath for a non-default group above
2569 	 * then move the default group to shared state to allow distribution of
2570 	 * incoming broadcast traffic to the other groups and dismantle the
2571 	 * SRSes over the default group.
2572 	 */
2573 	if (group != NULL) {
2574 		if (group != mip->mi_rx_groups) {
2575 			default_group = mip->mi_rx_groups;
2576 			if (default_group->mrg_state ==
2577 			    MAC_GROUP_STATE_RESERVED) {
2578 				group_only_mcip = MAC_RX_GROUP_ONLY_CLIENT(
2579 				    default_group);
2580 				ASSERT(group_only_mcip != NULL &&
2581 				    mip->mi_nactiveclients > 1);
2582 
2583 				mac_set_rx_group_state(default_group,
2584 				    MAC_GROUP_STATE_SHARED);
2585 				mac_srs_group_setup(group_only_mcip,
2586 				    group_only_mcip->mci_flent,
2587 				    default_group, SRST_LINK);
2588 			}
2589 			ASSERT(default_group->mrg_state ==
2590 			    MAC_GROUP_STATE_SHARED);
2591 		}
2592 		/*
2593 		 * If we get an exclusive group for a VLAN MAC client we
2594 		 * need to take the s/w path to make the additional check for
2595 		 * the vid. Disable polling and set it to s/w classification.
2596 		 */
2597 		if (group->mrg_state == MAC_GROUP_STATE_RESERVED &&
2598 		    i_mac_flow_vid(mcip->mci_flent) != VLAN_ID_NONE) {
2599 			mac_rx_switch_grp_to_sw(group);
2600 		}
2601 	}
2602 	return (0);
2603 
2604 setup_failed:
2605 	mac_datapath_teardown(mcip, flent, link_type);
2606 	return (err);
2607 }
2608 
2609 void
2610 mac_datapath_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
2611     uint32_t link_type)
2612 {
2613 	mac_impl_t		*mip = mcip->mci_mip;
2614 	mac_group_t		*group = NULL;
2615 	mac_client_impl_t	*grp_only_mcip;
2616 	flow_entry_t		*group_only_flent;
2617 	mac_group_t		*default_group;
2618 	boolean_t		check_default_group = B_FALSE;
2619 	mac_group_state_t	next_state;
2620 
2621 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
2622 
2623 	switch (link_type) {
2624 	case SRST_FLOW:
2625 		mac_srs_group_teardown(mcip, flent, SRST_FLOW);
2626 		return;
2627 
2628 	case SRST_LINK:
2629 		/* Stop sending packets */
2630 		mac_tx_client_block(mcip);
2631 
2632 		/* Stop the packets coming from the H/W */
2633 		if (mcip->mci_unicast != NULL) {
2634 			int err;
2635 			err = mac_remove_macaddr(mcip->mci_unicast);
2636 			if (err != 0) {
2637 				cmn_err(CE_WARN, "%s: failed to remove a MAC"
2638 				    " address because of error 0x%x",
2639 				    mip->mi_name, err);
2640 			}
2641 			mcip->mci_unicast = NULL;
2642 		}
2643 
2644 		/* Stop the packets coming from the S/W classifier */
2645 		mac_flow_remove(mip->mi_flow_tab, flent, B_FALSE);
2646 		mac_flow_wait(flent, FLOW_DRIVER_UPCALL);
2647 
2648 		/* Now quiesce and destroy all SRS and soft rings */
2649 		mac_srs_group_teardown(mcip, flent, SRST_LINK);
2650 		ASSERT((mcip->mci_flent == flent) &&
2651 		    (flent->fe_next == NULL));
2652 
2653 		/*
2654 		 * Release our hold on the group as well. We need
2655 		 * to check if the shared group has only one client
2656 		 * left who can use it exclusively. Also, if we
2657 		 * were the last client, release the group.
2658 		 */
2659 		group = flent->fe_rx_ring_group;
2660 		if (group != NULL) {
2661 			mac_rx_group_remove_client(group, mcip);
2662 			next_state = mac_rx_group_next_state(group,
2663 			    &grp_only_mcip);
2664 			if (next_state == MAC_GROUP_STATE_RESERVED) {
2665 				/*
2666 				 * Only one client left on this RX group.
2667 				 */
2668 				ASSERT(grp_only_mcip != NULL);
2669 				mac_set_rx_group_state(group,
2670 				    MAC_GROUP_STATE_RESERVED);
2671 				group_only_flent = grp_only_mcip->mci_flent;
2672 
2673 				/*
2674 				 * The only remaining client has exclusive
2675 				 * access on the group. Allow it to
2676 				 * dynamically poll the H/W rings etc.
2677 				 */
2678 				mac_srs_group_setup(grp_only_mcip,
2679 				    group_only_flent, group, SRST_LINK);
2680 				mac_rx_group_unmark(group, MR_INCIPIENT);
2681 			} else if (next_state == MAC_GROUP_STATE_REGISTERED) {
2682 				/*
2683 				 * This is a non-default group being freed up.
2684 				 * We need to reevaluate the default group
2685 				 * to see if the primary client can get
2686 				 * exclusive access to the default group.
2687 				 */
2688 				ASSERT(group != mip->mi_rx_groups);
2689 				mac_release_rx_group(mcip, group);
2690 				mac_set_rx_group_state(group,
2691 				    MAC_GROUP_STATE_REGISTERED);
2692 				check_default_group = B_TRUE;
2693 			} else {
2694 				ASSERT(next_state == MAC_GROUP_STATE_SHARED);
2695 				mac_set_rx_group_state(group,
2696 				    MAC_GROUP_STATE_SHARED);
2697 				mac_rx_group_unmark(group, MR_CONDEMNED);
2698 			}
2699 			flent->fe_rx_ring_group = NULL;
2700 		}
2701 		break;
2702 	default:
2703 		ASSERT(B_FALSE);
2704 		break;
2705 	}
2706 
2707 	/*
2708 	 * The mac client using the default group gets exclusive access to the
2709 	 * default group if and only if it is the sole client on the entire
2710 	 * mip. If so set the group state to reserved, and set up the SRSes
2711 	 * over the default group.
2712 	 */
2713 	if (check_default_group) {
2714 		default_group = mip->mi_rx_groups;
2715 		ASSERT(default_group->mrg_state == MAC_GROUP_STATE_SHARED);
2716 		next_state = mac_rx_group_next_state(default_group,
2717 		    &grp_only_mcip);
2718 		if (next_state == MAC_GROUP_STATE_RESERVED) {
2719 			ASSERT(grp_only_mcip != NULL &&
2720 			    mip->mi_nactiveclients == 1);
2721 			mac_set_rx_group_state(default_group,
2722 			    MAC_GROUP_STATE_RESERVED);
2723 			mac_srs_group_setup(grp_only_mcip,
2724 			    grp_only_mcip->mci_flent,
2725 			    default_group, SRST_LINK);
2726 			mac_rx_group_unmark(default_group, MR_INCIPIENT);
2727 		}
2728 	}
2729 }
2730 
2731 /* DATAPATH TEAR DOWN ROUTINES (SRS and FANOUT teardown) */
2732 
2733 static void
2734 mac_srs_fanout_list_free(mac_soft_ring_set_t *mac_srs)
2735 {
2736 	ASSERT(mac_srs->srs_tcp_soft_rings != NULL);
2737 	kmem_free(mac_srs->srs_tcp_soft_rings,
2738 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2739 	mac_srs->srs_tcp_soft_rings = NULL;
2740 	ASSERT(mac_srs->srs_udp_soft_rings != NULL);
2741 	kmem_free(mac_srs->srs_udp_soft_rings,
2742 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2743 	mac_srs->srs_udp_soft_rings = NULL;
2744 	ASSERT(mac_srs->srs_oth_soft_rings != NULL);
2745 	kmem_free(mac_srs->srs_oth_soft_rings,
2746 	    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
2747 	mac_srs->srs_oth_soft_rings = NULL;
2748 }
2749 
2750 /*
2751  * An RX SRS is attached to at most one mac_ring.
2752  * A TX SRS  has no  rings.
2753  */
2754 static void
2755 mac_srs_ring_free(mac_soft_ring_set_t *mac_srs)
2756 {
2757 	mac_client_impl_t	*mcip;
2758 	mac_ring_t		*ring;
2759 	flow_entry_t		*flent;
2760 
2761 	ring = mac_srs->srs_ring;
2762 	if (mac_srs->srs_type & SRST_TX) {
2763 		ASSERT(ring == NULL);
2764 		return;
2765 	}
2766 
2767 	if (ring == NULL)
2768 		return;
2769 
2770 	/*
2771 	 * Broadcast flows don't have a client impl association, but they
2772 	 * use only soft rings.
2773 	 */
2774 	flent = mac_srs->srs_flent;
2775 	mcip = flent->fe_mcip;
2776 	ASSERT(mcip != NULL);
2777 
2778 	ring->mr_classify_type = MAC_NO_CLASSIFIER;
2779 	ring->mr_srs = NULL;
2780 }
2781 
2782 /*
2783  * Physical unlink and free of the data structures happen below. This is
2784  * driven from mac_flow_destroy(), on the last refrele of a flow.
2785  *
2786  * Assumes Rx srs is 1-1 mapped with an ring.
2787  */
2788 void
2789 mac_srs_free(mac_soft_ring_set_t *mac_srs)
2790 {
2791 	ASSERT(mac_srs->srs_mcip == NULL ||
2792 	    MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
2793 	ASSERT((mac_srs->srs_state & (SRS_CONDEMNED | SRS_CONDEMNED_DONE |
2794 	    SRS_PROC | SRS_PROC_FAST)) == (SRS_CONDEMNED | SRS_CONDEMNED_DONE));
2795 
2796 	mac_pkt_drop(NULL, NULL, mac_srs->srs_first, B_FALSE);
2797 	mac_srs_ring_free(mac_srs);
2798 	mac_srs_soft_rings_free(mac_srs, B_TRUE);
2799 	mac_srs_fanout_list_free(mac_srs);
2800 
2801 	mac_srs->srs_bw = NULL;
2802 	kmem_cache_free(mac_srs_cache, mac_srs);
2803 }
2804 
2805 static void
2806 mac_srs_soft_rings_quiesce(mac_soft_ring_set_t *mac_srs, uint_t s_ring_flag)
2807 {
2808 	mac_soft_ring_t	*softring;
2809 
2810 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2811 
2812 	mac_srs_soft_rings_signal(mac_srs, s_ring_flag);
2813 	if (s_ring_flag == S_RING_CONDEMNED) {
2814 		while (mac_srs->srs_soft_ring_condemned_count !=
2815 		    mac_srs->srs_soft_ring_count)
2816 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2817 	} else {
2818 		while (mac_srs->srs_soft_ring_quiesced_count !=
2819 		    mac_srs->srs_soft_ring_count)
2820 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2821 	}
2822 	mutex_exit(&mac_srs->srs_lock);
2823 
2824 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2825 	    softring = softring->s_ring_next)
2826 		(void) untimeout(softring->s_ring_tid);
2827 
2828 	(void) untimeout(mac_srs->srs_tid);
2829 
2830 	mutex_enter(&mac_srs->srs_lock);
2831 }
2832 
2833 /*
2834  * The block comment above mac_rx_classify_flow_state_change explains the
2835  * background. At this point upcalls from the driver (both hardware classified
2836  * and software classified) have been cut off. We now need to quiesce the
2837  * SRS worker, poll, and softring threads. The SRS worker thread serves as
2838  * the master controller. The steps involved are described below in the function
2839  */
2840 void
2841 mac_srs_worker_quiesce(mac_soft_ring_set_t *mac_srs)
2842 {
2843 	uint_t			s_ring_flag;
2844 	uint_t			srs_poll_wait_flag;
2845 
2846 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2847 	ASSERT(mac_srs->srs_state & (SRS_CONDEMNED | SRS_QUIESCE));
2848 
2849 	if (mac_srs->srs_state & SRS_CONDEMNED) {
2850 		s_ring_flag = S_RING_CONDEMNED;
2851 		srs_poll_wait_flag = SRS_POLL_THR_EXITED;
2852 	} else {
2853 		s_ring_flag = S_RING_QUIESCE;
2854 		srs_poll_wait_flag = SRS_POLL_THR_QUIESCED;
2855 	}
2856 
2857 	/*
2858 	 * In the case of Rx SRS wait till the poll thread is done.
2859 	 */
2860 	if ((mac_srs->srs_type & SRST_TX) == 0 &&
2861 	    mac_srs->srs_poll_thr != NULL) {
2862 		while (!(mac_srs->srs_state & srs_poll_wait_flag))
2863 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2864 
2865 		/*
2866 		 * Turn off polling as part of the quiesce operation.
2867 		 */
2868 		MAC_SRS_POLLING_OFF(mac_srs);
2869 		mac_srs->srs_state &= ~(SRS_POLLING | SRS_GET_PKTS);
2870 	}
2871 
2872 	/*
2873 	 * Then signal the soft ring worker threads to quiesce or quit
2874 	 * as needed and then wait till that happens.
2875 	 */
2876 	mac_srs_soft_rings_quiesce(mac_srs, s_ring_flag);
2877 
2878 	if (mac_srs->srs_state & SRS_CONDEMNED)
2879 		mac_srs->srs_state |= (SRS_QUIESCE_DONE | SRS_CONDEMNED_DONE);
2880 	else
2881 		mac_srs->srs_state |= SRS_QUIESCE_DONE;
2882 	cv_signal(&mac_srs->srs_quiesce_done_cv);
2883 }
2884 
2885 /*
2886  * Signal an SRS to start a temporary quiesce, or permanent removal, or restart
2887  * a quiesced SRS by setting the appropriate flags and signaling the SRS worker
2888  * or poll thread. This function is internal to the quiescing logic and is
2889  * called internally from the SRS quiesce or flow quiesce or client quiesce
2890  * higher level functions.
2891  */
2892 void
2893 mac_srs_signal(mac_soft_ring_set_t *mac_srs, uint_t srs_flag)
2894 {
2895 	mac_ring_t	*ring;
2896 
2897 	ring = mac_srs->srs_ring;
2898 	ASSERT(ring == NULL || ring->mr_refcnt == 0);
2899 
2900 	if (srs_flag == SRS_CONDEMNED) {
2901 		/*
2902 		 * The SRS is going away. We need to unbind the SRS and SR
2903 		 * threads before removing from the global SRS list. Otherwise
2904 		 * there is a small window where the cpu reconfig callbacks
2905 		 * may miss the SRS in the list walk and DR could fail since
2906 		 * there are still bound threads.
2907 		 */
2908 		mac_srs_threads_unbind(mac_srs);
2909 		mac_srs_remove_glist(mac_srs);
2910 	}
2911 	/*
2912 	 * Wakeup the SRS worker and poll threads.
2913 	 */
2914 	mutex_enter(&mac_srs->srs_lock);
2915 	mac_srs->srs_state |= srs_flag;
2916 	cv_signal(&mac_srs->srs_async);
2917 	cv_signal(&mac_srs->srs_cv);
2918 	mutex_exit(&mac_srs->srs_lock);
2919 }
2920 
2921 /*
2922  * In the Rx side, the quiescing is done bottom up. After the Rx upcalls
2923  * from the driver are done, then the Rx SRS is quiesced and only then can
2924  * we signal the soft rings. Thus this function can't be called arbitrarily
2925  * without satisfying the prerequisites. On the Tx side, the threads from
2926  * top need to quiesced, then the Tx SRS and only then can we signal the
2927  * Tx soft rings.
2928  */
2929 static void
2930 mac_srs_soft_rings_signal(mac_soft_ring_set_t *mac_srs, uint_t sr_flag)
2931 {
2932 	mac_soft_ring_t		*softring;
2933 
2934 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2935 	    softring = softring->s_ring_next)
2936 		mac_soft_ring_signal(softring, sr_flag);
2937 }
2938 
2939 /*
2940  * The block comment above mac_rx_classify_flow_state_change explains the
2941  * background. At this point the SRS is quiesced and we need to restart the
2942  * SRS worker, poll, and softring threads. The SRS worker thread serves as
2943  * the master controller. The steps involved are described below in the function
2944  */
2945 void
2946 mac_srs_worker_restart(mac_soft_ring_set_t *mac_srs)
2947 {
2948 	boolean_t	iam_rx_srs;
2949 	mac_soft_ring_t	*softring;
2950 
2951 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
2952 	if ((mac_srs->srs_type & SRST_TX) != 0) {
2953 		iam_rx_srs = B_FALSE;
2954 		ASSERT((mac_srs->srs_state &
2955 		    (SRS_POLL_THR_QUIESCED | SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
2956 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
2957 	} else {
2958 		iam_rx_srs = B_TRUE;
2959 		ASSERT((mac_srs->srs_state &
2960 		    (SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
2961 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
2962 		if (mac_srs->srs_poll_thr != NULL) {
2963 			ASSERT((mac_srs->srs_state & SRS_POLL_THR_QUIESCED) ==
2964 			    SRS_POLL_THR_QUIESCED);
2965 		}
2966 	}
2967 
2968 	/*
2969 	 * Signal any quiesced soft ring workers to restart and wait for the
2970 	 * soft ring down count to come down to zero.
2971 	 */
2972 	if (mac_srs->srs_soft_ring_quiesced_count != 0) {
2973 		for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
2974 		    softring = softring->s_ring_next) {
2975 			if (!(softring->s_ring_state & S_RING_QUIESCE))
2976 				continue;
2977 			mac_soft_ring_signal(softring, S_RING_RESTART);
2978 		}
2979 		while (mac_srs->srs_soft_ring_quiesced_count != 0)
2980 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2981 	}
2982 
2983 	mac_srs->srs_state &= ~(SRS_QUIESCE_DONE | SRS_QUIESCE | SRS_RESTART);
2984 	if (iam_rx_srs && mac_srs->srs_poll_thr != NULL) {
2985 		/*
2986 		 * Signal the poll thread and ask it to restart. Wait till it
2987 		 * actually restarts and the SRS_POLL_THR_QUIESCED flag gets
2988 		 * cleared.
2989 		 */
2990 		mac_srs->srs_state |= SRS_POLL_THR_RESTART;
2991 		cv_signal(&mac_srs->srs_cv);
2992 		while (mac_srs->srs_state & SRS_POLL_THR_QUIESCED)
2993 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2994 		ASSERT(!(mac_srs->srs_state & SRS_POLL_THR_RESTART));
2995 	}
2996 	/* Wake up any waiter waiting for the restart to complete */
2997 	mac_srs->srs_state |= SRS_RESTART_DONE;
2998 	cv_signal(&mac_srs->srs_quiesce_done_cv);
2999 }
3000 
3001 static void
3002 mac_srs_worker_unbind(mac_soft_ring_set_t *mac_srs)
3003 {
3004 	mutex_enter(&mac_srs->srs_lock);
3005 	if (!(mac_srs->srs_state & SRS_WORKER_BOUND)) {
3006 		ASSERT(mac_srs->srs_worker_cpuid == -1);
3007 		mutex_exit(&mac_srs->srs_lock);
3008 		return;
3009 	}
3010 
3011 	mac_srs->srs_worker_cpuid = -1;
3012 	mac_srs->srs_state &= ~SRS_WORKER_BOUND;
3013 	thread_affinity_clear(mac_srs->srs_worker);
3014 	mutex_exit(&mac_srs->srs_lock);
3015 }
3016 
3017 static void
3018 mac_srs_poll_unbind(mac_soft_ring_set_t *mac_srs)
3019 {
3020 	mutex_enter(&mac_srs->srs_lock);
3021 	if (mac_srs->srs_poll_thr == NULL ||
3022 	    (mac_srs->srs_state & SRS_POLL_BOUND) == 0) {
3023 		ASSERT(mac_srs->srs_poll_cpuid == -1);
3024 		mutex_exit(&mac_srs->srs_lock);
3025 		return;
3026 	}
3027 
3028 	mac_srs->srs_poll_cpuid = -1;
3029 	mac_srs->srs_state &= ~SRS_POLL_BOUND;
3030 	thread_affinity_clear(mac_srs->srs_poll_thr);
3031 	mutex_exit(&mac_srs->srs_lock);
3032 }
3033 
3034 static void
3035 mac_srs_threads_unbind(mac_soft_ring_set_t *mac_srs)
3036 {
3037 	mac_soft_ring_t	*soft_ring;
3038 
3039 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
3040 
3041 	mutex_enter(&cpu_lock);
3042 	mac_srs_worker_unbind(mac_srs);
3043 	if (!(mac_srs->srs_type & SRST_TX))
3044 		mac_srs_poll_unbind(mac_srs);
3045 
3046 	for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3047 	    soft_ring = soft_ring->s_ring_next) {
3048 		mac_soft_ring_unbind(soft_ring);
3049 	}
3050 	mutex_exit(&cpu_lock);
3051 }
3052 
3053 /*
3054  * When a CPU is going away, unbind all MAC threads which are bound
3055  * to that CPU. The affinity of the thread to the CPU is saved to allow
3056  * the thread to be rebound to the CPU if it comes back online.
3057  */
3058 static void
3059 mac_walk_srs_and_unbind(int cpuid)
3060 {
3061 	mac_soft_ring_set_t *mac_srs;
3062 	mac_soft_ring_t *soft_ring;
3063 
3064 	rw_enter(&mac_srs_g_lock, RW_READER);
3065 
3066 	if ((mac_srs = mac_srs_g_list) == NULL)
3067 		goto done;
3068 
3069 	for (; mac_srs != NULL; mac_srs = mac_srs->srs_next) {
3070 		if (mac_srs->srs_worker_cpuid == cpuid) {
3071 			mac_srs->srs_worker_cpuid_save = cpuid;
3072 			mac_srs_worker_unbind(mac_srs);
3073 		}
3074 
3075 		if (!(mac_srs->srs_type & SRST_TX)) {
3076 			if (mac_srs->srs_poll_cpuid == cpuid) {
3077 				mac_srs->srs_poll_cpuid_save = cpuid;
3078 				mac_srs_poll_unbind(mac_srs);
3079 			}
3080 		}
3081 
3082 		/* Next tackle the soft rings associated with the srs */
3083 		mutex_enter(&mac_srs->srs_lock);
3084 		for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3085 		    soft_ring = soft_ring->s_ring_next) {
3086 			if (soft_ring->s_ring_cpuid == cpuid) {
3087 				soft_ring->s_ring_cpuid_save = cpuid;
3088 				mac_soft_ring_unbind(soft_ring);
3089 			}
3090 		}
3091 		mutex_exit(&mac_srs->srs_lock);
3092 	}
3093 done:
3094 	rw_exit(&mac_srs_g_lock);
3095 }
3096 
3097 /* TX SETUP and TEARDOWN ROUTINES */
3098 
3099 /*
3100  * XXXHIO need to make sure the two mac_tx_srs_{add,del}_ring()
3101  * handle the case where the number of rings is one. I.e. there is
3102  * a ring pointed to by mac_srs->srs_tx_arg2.
3103  */
3104 void
3105 mac_tx_srs_add_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3106 {
3107 	mac_client_impl_t *mcip = mac_srs->srs_mcip;
3108 	mac_soft_ring_t *soft_ring;
3109 	int count = mac_srs->srs_oth_ring_count;
3110 
3111 	ASSERT(mac_srs->srs_state & SRS_QUIESCE);
3112 	soft_ring = mac_soft_ring_create(count, 0, NULL,
3113 	    (ST_RING_OTH | ST_RING_TX), maxclsyspri, mcip, mac_srs, -1,
3114 	    NULL, mcip, (mac_resource_handle_t)tx_ring);
3115 	mac_srs->srs_oth_ring_count++;
3116 	/*
3117 	 * put this soft ring in quiesce mode too so when we restart
3118 	 * all soft rings in the srs are in the same state.
3119 	 */
3120 	mac_soft_ring_signal(soft_ring, S_RING_QUIESCE);
3121 }
3122 
3123 static void
3124 mac_soft_ring_remove(mac_soft_ring_set_t *mac_srs, mac_soft_ring_t *softring)
3125 {
3126 	int sringcnt;
3127 
3128 	mutex_enter(&mac_srs->srs_lock);
3129 	sringcnt = mac_srs->srs_soft_ring_count;
3130 	ASSERT(sringcnt > 0);
3131 	mac_soft_ring_signal(softring, S_RING_CONDEMNED);
3132 
3133 	ASSERT(mac_srs->srs_soft_ring_condemned_count == 0);
3134 	while (mac_srs->srs_soft_ring_condemned_count != 1)
3135 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3136 
3137 	if (softring == mac_srs->srs_soft_ring_head) {
3138 		mac_srs->srs_soft_ring_head = softring->s_ring_next;
3139 		if (mac_srs->srs_soft_ring_head != NULL) {
3140 			mac_srs->srs_soft_ring_head->s_ring_prev = NULL;
3141 		} else {
3142 			mac_srs->srs_soft_ring_tail = NULL;
3143 		}
3144 	} else {
3145 		softring->s_ring_prev->s_ring_next =
3146 		    softring->s_ring_next;
3147 		if (softring->s_ring_next != NULL) {
3148 			softring->s_ring_next->s_ring_prev =
3149 			    softring->s_ring_prev;
3150 		} else {
3151 			mac_srs->srs_soft_ring_tail =
3152 			    softring->s_ring_prev;
3153 		}
3154 	}
3155 	mac_srs->srs_soft_ring_count--;
3156 
3157 	mac_srs->srs_soft_ring_condemned_count--;
3158 	mutex_exit(&mac_srs->srs_lock);
3159 
3160 	mac_soft_ring_free(softring, B_FALSE);
3161 }
3162 
3163 void
3164 mac_tx_srs_del_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3165 {
3166 	int i;
3167 	mac_soft_ring_t *soft_ring, *remove_sring;
3168 
3169 	mutex_enter(&mac_srs->srs_lock);
3170 	for (i = 0; i < mac_srs->srs_oth_ring_count; i++) {
3171 		soft_ring =  mac_srs->srs_oth_soft_rings[i];
3172 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3173 			break;
3174 	}
3175 	mutex_exit(&mac_srs->srs_lock);
3176 	ASSERT(i < mac_srs->srs_oth_ring_count);
3177 	remove_sring = soft_ring;
3178 	mac_soft_ring_remove(mac_srs, remove_sring);
3179 	mac_srs_update_fanout_list(mac_srs);
3180 }
3181 
3182 /*
3183  * mac_tx_srs_setup():
3184  *
3185  * Used to setup Tx rings. If no free Tx ring is available, then default
3186  * Tx ring is used.
3187  */
3188 void
3189 mac_tx_srs_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
3190     uint32_t srs_type)
3191 {
3192 	mac_impl_t *mip = mcip->mci_mip;
3193 	mac_soft_ring_set_t *tx_srs;
3194 	int i, tx_ring_count = 0, tx_rings_reserved = 0;
3195 	mac_ring_handle_t *tx_ring = NULL;
3196 	uint32_t soft_ring_type;
3197 	mac_group_t *grp = NULL;
3198 	mac_ring_t *ring;
3199 	mac_srs_tx_t *tx;
3200 	boolean_t serialize = B_FALSE;
3201 
3202 	tx_srs = flent->fe_tx_srs;
3203 	tx = &tx_srs->srs_tx;
3204 
3205 	if (tx->st_group != NULL) {
3206 		grp = tx->st_group;
3207 		tx_ring_count = grp->mrg_cur_count;
3208 	} else {
3209 		tx_ring_count = mac_tx_ring_count;
3210 	}
3211 
3212 	if (tx_ring_count != 0) {
3213 		tx_ring = kmem_zalloc(sizeof (mac_ring_handle_t) *
3214 		    tx_ring_count, KM_SLEEP);
3215 	}
3216 
3217 	/*
3218 	 * Just use the default ring for now. We need to use
3219 	 * the underlying link's ring set instead of the underlying
3220 	 * NIC's.
3221 	 */
3222 	if (srs_type == SRST_FLOW ||
3223 	    (mcip->mci_state_flags & MCIS_NO_HWRINGS) != 0)
3224 		goto use_default_ring;
3225 
3226 	if (mcip->mci_share != NULL)
3227 		ring = grp->mrg_rings;
3228 	/*
3229 	 * An attempt is made to reserve 'tx_ring_count' number
3230 	 * of Tx rings. If tx_ring_count is 0, default Tx ring
3231 	 * is used. If it is 1, an attempt is made to reserve one
3232 	 * Tx ring. In both the cases, the ring information is
3233 	 * stored in Tx SRS. If multiple Tx rings are specified,
3234 	 * then each Tx ring will have a Tx-side soft ring. All
3235 	 * these soft rings will be hang off Tx SRS.
3236 	 */
3237 	for (i = 0, tx_rings_reserved = 0;
3238 	    i < tx_ring_count; i++, tx_rings_reserved++) {
3239 		if (mcip->mci_share != NULL) {
3240 			/*
3241 			 * The ring was already chosen and associated
3242 			 * with the TX group. Save it in the new
3243 			 * array to keep as much of the code below common
3244 			 * between the share and non-share cases.
3245 			 */
3246 			ASSERT(ring != NULL);
3247 			tx_ring[i] = (mac_ring_handle_t)ring;
3248 			ring = ring->mr_next;
3249 		} else {
3250 			tx_ring[i] =
3251 			    (mac_ring_handle_t)mac_reserve_tx_ring(mip, NULL);
3252 			if (tx_ring[i] == NULL)
3253 				break;
3254 		}
3255 	}
3256 	if (mac_tx_serialize || (mip->mi_v12n_level & MAC_VIRT_SERIALIZE))
3257 		serialize = B_TRUE;
3258 	/*
3259 	 * Did we get the requested number of tx rings?
3260 	 * There are 3 actions we can take depending upon the number
3261 	 * of tx_rings we got.
3262 	 * 1) If we got none, then hook up the tx_srs with the
3263 	 * default ring.
3264 	 * 2) If we got one, then get the tx_ring from the soft ring,
3265 	 * save it in SRS and free up the soft ring.
3266 	 * 3) If we got more than 1, then do the tx fanout among the
3267 	 * rings we obtained.
3268 	 */
3269 	switch (tx_rings_reserved) {
3270 	case 1:
3271 		/*
3272 		 * No need to allocate Tx soft rings. Tx-side soft
3273 		 * rings are for Tx fanout case. Just use Tx SRS.
3274 		 */
3275 		/* FALLTHRU */
3276 
3277 	case 0:
3278 use_default_ring:
3279 		if (tx_rings_reserved == 0)
3280 			tx->st_arg2 = (void *)mip->mi_default_tx_ring;
3281 		else
3282 			tx->st_arg2 = (void *)tx_ring[0];
3283 		/* For ring_count of 0 or 1, set the tx_mode and return */
3284 		if (tx_srs->srs_type & SRST_BW_CONTROL)
3285 			tx->st_mode = SRS_TX_BW;
3286 		else if (serialize)
3287 			tx->st_mode = SRS_TX_SERIALIZE;
3288 		else
3289 			tx->st_mode = SRS_TX_DEFAULT;
3290 		break;
3291 
3292 	default:
3293 		/*
3294 		 * We got multiple Tx rings for Tx fanout.
3295 		 *
3296 		 * cpuid of -1 is passed. This creates an unbound
3297 		 * worker thread. Instead the code should get CPU
3298 		 * binding information and pass that to
3299 		 * mac_soft_ring_create(). This needs to be done
3300 		 * in conjunction with Rx-side soft ring
3301 		 * bindings.
3302 		 */
3303 		soft_ring_type = ST_RING_OTH | ST_RING_TX;
3304 		if (tx_srs->srs_type & SRST_BW_CONTROL) {
3305 			tx->st_mode = SRS_TX_BW_FANOUT;
3306 		} else {
3307 			tx->st_mode = SRS_TX_FANOUT;
3308 			if (serialize)
3309 				soft_ring_type |= ST_RING_WORKER_ONLY;
3310 		}
3311 		for (i = 0; i < tx_rings_reserved; i++) {
3312 			(void) mac_soft_ring_create(i, 0, NULL, soft_ring_type,
3313 			    maxclsyspri, mcip, tx_srs, -1, NULL, mcip,
3314 			    (mac_resource_handle_t)tx_ring[i]);
3315 		}
3316 		mac_srs_update_fanout_list(tx_srs);
3317 	}
3318 	tx->st_func = mac_tx_get_func(tx->st_mode);
3319 
3320 	DTRACE_PROBE3(tx__srs___setup__return, mac_soft_ring_set_t *, tx_srs,
3321 	    int, tx->st_mode, int, tx_srs->srs_oth_ring_count);
3322 
3323 	if (tx_ring_count != 0) {
3324 		kmem_free(tx_ring,
3325 		    sizeof (mac_ring_handle_t) * tx_ring_count);
3326 	}
3327 }
3328 
3329 /*
3330  * Update the fanout of a client if its recorded link speed doesn't match
3331  * its current link speed.
3332  */
3333 void
3334 mac_fanout_recompute_client(mac_client_impl_t *mcip)
3335 {
3336 	uint64_t link_speed;
3337 	mac_resource_props_t *mcip_mrp;
3338 
3339 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
3340 
3341 	link_speed = mac_client_stat_get(mcip->mci_flent->fe_mcip,
3342 	    MAC_STAT_IFSPEED);
3343 
3344 	if ((link_speed != 0) &&
3345 	    (link_speed != mcip->mci_flent->fe_nic_speed)) {
3346 		mcip_mrp = MCIP_RESOURCE_PROPS(mcip);
3347 		mac_fanout_setup(mcip, mcip->mci_flent,
3348 		    mcip_mrp, mac_rx_deliver, mcip, NULL);
3349 	}
3350 }
3351 
3352 /*
3353  * Walk through the list of mac clients for the MAC.
3354  * For each active mac client, recompute the number of soft rings
3355  * associated with every client, only if current speed is different
3356  * from the speed that was previously used for soft ring computation.
3357  * If the cable is disconnected whlie the NIC is started, we would get
3358  * notification with speed set to 0. We do not recompute in that case.
3359  */
3360 void
3361 mac_fanout_recompute(mac_impl_t *mip)
3362 {
3363 	mac_client_impl_t	*mcip;
3364 
3365 	i_mac_perim_enter(mip);
3366 	if ((mip->mi_state_flags & MIS_IS_VNIC) != 0 ||
3367 	    mip->mi_linkstate != LINK_STATE_UP) {
3368 		i_mac_perim_exit(mip);
3369 		return;
3370 	}
3371 
3372 	for (mcip = mip->mi_clients_list; mcip != NULL;
3373 	    mcip = mcip->mci_client_next) {
3374 		if ((mcip->mci_state_flags & MCIS_SHARE_BOUND) != 0 ||
3375 		    !MCIP_DATAPATH_SETUP(mcip))
3376 			continue;
3377 		mac_fanout_recompute_client(mcip);
3378 	}
3379 	i_mac_perim_exit(mip);
3380 }
3381