xref: /illumos-gate/usr/src/uts/common/io/mac/mac_datapath_setup.c (revision 5328fc53d11d7151861fa272e4fb0248b8f0e145)
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 (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2017, Joyent, Inc.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/callb.h>
28 #include <sys/cpupart.h>
29 #include <sys/pool.h>
30 #include <sys/pool_pset.h>
31 #include <sys/sdt.h>
32 #include <sys/strsubr.h>
33 #include <sys/strsun.h>
34 #include <sys/vlan.h>
35 #include <inet/ipsec_impl.h>
36 #include <inet/ip_impl.h>
37 #include <inet/sadb.h>
38 #include <inet/ipsecesp.h>
39 #include <inet/ipsecah.h>
40 
41 #include <sys/mac_impl.h>
42 #include <sys/mac_client_impl.h>
43 #include <sys/mac_client_priv.h>
44 #include <sys/mac_soft_ring.h>
45 #include <sys/mac_flow_impl.h>
46 #include <sys/mac_stat.h>
47 
48 static void mac_srs_soft_rings_signal(mac_soft_ring_set_t *, uint_t);
49 static void mac_srs_update_fanout_list(mac_soft_ring_set_t *);
50 static void mac_srs_poll_unbind(mac_soft_ring_set_t *);
51 static void mac_srs_worker_unbind(mac_soft_ring_set_t *);
52 static void mac_srs_soft_rings_quiesce(mac_soft_ring_set_t *, uint_t);
53 
54 static int mac_srs_cpu_setup(cpu_setup_t, int, void *);
55 static void mac_srs_worker_bind(mac_soft_ring_set_t *, processorid_t);
56 static void mac_srs_poll_bind(mac_soft_ring_set_t *, processorid_t);
57 static void mac_srs_threads_unbind(mac_soft_ring_set_t *);
58 static void mac_srs_add_glist(mac_soft_ring_set_t *);
59 static void mac_srs_remove_glist(mac_soft_ring_set_t *);
60 static void mac_srs_fanout_list_free(mac_soft_ring_set_t *);
61 static void mac_soft_ring_remove(mac_soft_ring_set_t *, mac_soft_ring_t *);
62 
63 static int mac_compute_soft_ring_count(flow_entry_t *, int, int);
64 static void mac_walk_srs_and_bind(int);
65 static void mac_walk_srs_and_unbind(int);
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 boolean_t mac_tx_serialize = B_FALSE;
96 
97 /*
98  * mac_tx_srs_hiwat is the queue depth threshold at which callers of
99  * mac_tx() will be notified of flow control condition.
100  *
101  * TCP does not honour flow control condition sent up by mac_tx().
102  * Thus provision is made for TCP to allow more packets to be queued
103  * in SRS upto a maximum of mac_tx_srs_max_q_cnt.
104  *
105  * Note that mac_tx_srs_hiwat is always be lesser than
106  * mac_tx_srs_max_q_cnt.
107  */
108 uint32_t mac_tx_srs_max_q_cnt = 100000;
109 uint32_t mac_tx_srs_hiwat = 1000;
110 
111 /*
112  * mac_rx_soft_ring_count, mac_soft_ring_10gig_count:
113  *
114  * Global tunables that determines the number of soft rings to be used for
115  * fanning out incoming traffic on a link. These count will be used only
116  * when no explicit set of CPUs was assigned to the data-links.
117  *
118  * mac_rx_soft_ring_count tunable will come into effect only if
119  * mac_soft_ring_enable is set. mac_soft_ring_enable is turned on by
120  * default only for sun4v platforms.
121  *
122  * mac_rx_soft_ring_10gig_count will come into effect if you are running on a
123  * 10Gbps link and is not dependent upon mac_soft_ring_enable.
124  *
125  * The number of soft rings for fanout for a link or a flow is determined
126  * by mac_compute_soft_ring_count() routine. This routine will take into
127  * account mac_soft_ring_enable, mac_rx_soft_ring_count and
128  * mac_rx_soft_ring_10gig_count to determine the soft ring count for a link.
129  *
130  * If a bandwidth is specified, the determination of the number of soft
131  * rings is based on specified bandwidth, CPU speed and number of CPUs in
132  * the system.
133  */
134 uint_t mac_rx_soft_ring_count = 8;
135 uint_t mac_rx_soft_ring_10gig_count = 8;
136 
137 /*
138  * Every Tx and Rx mac_soft_ring_set_t (mac_srs) created gets added
139  * to mac_srs_g_list and mac_srs_g_lock protects mac_srs_g_list. The
140  * list is used to walk the list of all MAC threads when a CPU is
141  * coming online or going offline.
142  */
143 static mac_soft_ring_set_t *mac_srs_g_list = NULL;
144 static krwlock_t mac_srs_g_lock;
145 
146 /*
147  * Whether the SRS threads should be bound, or not.
148  */
149 boolean_t mac_srs_thread_bind = B_TRUE;
150 
151 /*
152  * Whether Rx/Tx interrupts should be re-targeted. Disabled by default.
153  * dladm command would override this.
154  */
155 boolean_t mac_tx_intr_retarget = B_FALSE;
156 boolean_t mac_rx_intr_retarget = B_FALSE;
157 
158 /*
159  * If cpu bindings are specified by user, then Tx SRS and its soft
160  * rings should also be bound to the CPUs specified by user. The
161  * CPUs for Tx bindings are at the end of the cpu list provided by
162  * the user. If enough CPUs are not available (for Tx and Rx
163  * SRSes), then the CPUs are shared by both Tx and Rx SRSes.
164  */
165 #define	BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp) {			\
166 	processorid_t cpuid;						\
167 	int i;								\
168 	mac_soft_ring_t *softring;					\
169 	mac_cpus_t *srs_cpu;						\
170 									\
171 	srs_cpu = &mac_tx_srs->srs_cpu;					\
172 	cpuid = srs_cpu->mc_tx_fanout_cpus[0];				\
173 	mac_srs_worker_bind(mac_tx_srs, cpuid);				\
174 	if (MAC_TX_SOFT_RINGS(mac_tx_srs)) {				\
175 		for (i = 0; i < mac_tx_srs->srs_tx_ring_count; i++) {	\
176 			cpuid = srs_cpu->mc_tx_fanout_cpus[i];		\
177 			softring = mac_tx_srs->srs_tx_soft_rings[i];	\
178 			if (cpuid != -1) {				\
179 				(void) mac_soft_ring_bind(softring,	\
180 				    cpuid);				\
181 			}						\
182 		}							\
183 	}								\
184 }
185 
186 /*
187  * Re-targeting is allowed only for exclusive group or for primary.
188  */
189 #define	RETARGETABLE_CLIENT(group, mcip)				\
190 	((((group) != NULL) &&						\
191 	    ((group)->mrg_state == MAC_GROUP_STATE_RESERVED)) ||	\
192 	    mac_is_primary_client(mcip))
193 
194 #define	MAC_RING_RETARGETABLE(ring)					\
195 	(((ring) != NULL) &&						\
196 	    ((ring)->mr_info.mri_intr.mi_ddi_handle != NULL) &&		\
197 	    !((ring)->mr_info.mri_intr.mi_ddi_shared))
198 
199 
200 /* INIT and FINI ROUTINES */
201 
202 void
203 mac_soft_ring_init(void)
204 {
205 	mac_soft_ring_cache = kmem_cache_create("mac_soft_ring_cache",
206 	    sizeof (mac_soft_ring_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
207 
208 	mac_srs_cache = kmem_cache_create("mac_srs_cache",
209 	    sizeof (mac_soft_ring_set_t),
210 	    64, NULL, NULL, NULL, NULL, NULL, 0);
211 
212 	rw_init(&mac_srs_g_lock, NULL, RW_DEFAULT, NULL);
213 	mutex_enter(&cpu_lock);
214 	register_cpu_setup_func(mac_srs_cpu_setup, NULL);
215 	mutex_exit(&cpu_lock);
216 }
217 
218 void
219 mac_soft_ring_finish(void)
220 {
221 	mutex_enter(&cpu_lock);
222 	unregister_cpu_setup_func(mac_srs_cpu_setup, NULL);
223 	mutex_exit(&cpu_lock);
224 	rw_destroy(&mac_srs_g_lock);
225 	kmem_cache_destroy(mac_soft_ring_cache);
226 	kmem_cache_destroy(mac_srs_cache);
227 }
228 
229 static void
230 mac_srs_soft_rings_free(mac_soft_ring_set_t *mac_srs)
231 {
232 	mac_soft_ring_t	*softring, *next, *head;
233 
234 	/*
235 	 * Synchronize with mac_walk_srs_bind/unbind which are callbacks from
236 	 * DR. The callbacks from DR are called with cpu_lock held, and hence
237 	 * can't wait to grab the mac perimeter. The soft ring list is hence
238 	 * protected for read access by srs_lock. Changing the soft ring list
239 	 * needs the mac perimeter and the srs_lock.
240 	 */
241 	mutex_enter(&mac_srs->srs_lock);
242 
243 	head = mac_srs->srs_soft_ring_head;
244 	mac_srs->srs_soft_ring_head = NULL;
245 	mac_srs->srs_soft_ring_tail = NULL;
246 	mac_srs->srs_soft_ring_count = 0;
247 
248 	mutex_exit(&mac_srs->srs_lock);
249 
250 	for (softring = head; softring != NULL; softring = next) {
251 		next = softring->s_ring_next;
252 		mac_soft_ring_free(softring);
253 	}
254 }
255 
256 static void
257 mac_srs_add_glist(mac_soft_ring_set_t *mac_srs)
258 {
259 	ASSERT(mac_srs->srs_next == NULL && mac_srs->srs_prev == NULL);
260 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
261 
262 	rw_enter(&mac_srs_g_lock, RW_WRITER);
263 	mutex_enter(&mac_srs->srs_lock);
264 
265 	ASSERT((mac_srs->srs_state & SRS_IN_GLIST) == 0);
266 
267 	if (mac_srs_g_list == NULL) {
268 		mac_srs_g_list = mac_srs;
269 	} else {
270 		mac_srs->srs_next = mac_srs_g_list;
271 		mac_srs_g_list->srs_prev = mac_srs;
272 		mac_srs->srs_prev = NULL;
273 		mac_srs_g_list = mac_srs;
274 	}
275 	mac_srs->srs_state |= SRS_IN_GLIST;
276 
277 	mutex_exit(&mac_srs->srs_lock);
278 	rw_exit(&mac_srs_g_lock);
279 }
280 
281 static void
282 mac_srs_remove_glist(mac_soft_ring_set_t *mac_srs)
283 {
284 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
285 
286 	rw_enter(&mac_srs_g_lock, RW_WRITER);
287 	mutex_enter(&mac_srs->srs_lock);
288 
289 	ASSERT((mac_srs->srs_state & SRS_IN_GLIST) != 0);
290 
291 	if (mac_srs == mac_srs_g_list) {
292 		mac_srs_g_list = mac_srs->srs_next;
293 		if (mac_srs_g_list != NULL)
294 			mac_srs_g_list->srs_prev = NULL;
295 	} else {
296 		mac_srs->srs_prev->srs_next = mac_srs->srs_next;
297 		if (mac_srs->srs_next != NULL)
298 			mac_srs->srs_next->srs_prev = mac_srs->srs_prev;
299 	}
300 	mac_srs->srs_state &= ~SRS_IN_GLIST;
301 
302 	mutex_exit(&mac_srs->srs_lock);
303 	rw_exit(&mac_srs_g_lock);
304 }
305 
306 /* POLLING SETUP AND TEAR DOWN ROUTINES */
307 
308 /*
309  * mac_srs_client_poll_quiesce and mac_srs_client_poll_restart
310  *
311  * These routines are used to call back into the upper layer
312  * (primarily TCP squeue) to stop polling the soft rings or
313  * restart polling.
314  */
315 void
316 mac_srs_client_poll_quiesce(mac_client_impl_t *mcip,
317     mac_soft_ring_set_t *mac_srs)
318 {
319 	mac_soft_ring_t	*softring;
320 
321 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
322 
323 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
324 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
325 		return;
326 	}
327 
328 	for (softring = mac_srs->srs_soft_ring_head;
329 	    softring != NULL; softring = softring->s_ring_next) {
330 		if ((softring->s_ring_type & ST_RING_TCP) &&
331 		    (softring->s_ring_rx_arg2 != NULL)) {
332 			mcip->mci_resource_quiesce(mcip->mci_resource_arg,
333 			    softring->s_ring_rx_arg2);
334 		}
335 	}
336 }
337 
338 void
339 mac_srs_client_poll_restart(mac_client_impl_t *mcip,
340     mac_soft_ring_set_t *mac_srs)
341 {
342 	mac_soft_ring_t	*softring;
343 
344 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
345 
346 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
347 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
348 		return;
349 	}
350 
351 	for (softring = mac_srs->srs_soft_ring_head;
352 	    softring != NULL; softring = softring->s_ring_next) {
353 		if ((softring->s_ring_type & ST_RING_TCP) &&
354 		    (softring->s_ring_rx_arg2 != NULL)) {
355 			mcip->mci_resource_restart(mcip->mci_resource_arg,
356 			    softring->s_ring_rx_arg2);
357 		}
358 	}
359 }
360 
361 /*
362  * Register the given SRS and associated soft rings with the consumer and
363  * enable the polling interface used by the consumer.(i.e IP) over this
364  * SRS and associated soft rings.
365  */
366 void
367 mac_srs_client_poll_enable(mac_client_impl_t *mcip,
368     mac_soft_ring_set_t *mac_srs)
369 {
370 	mac_rx_fifo_t		mrf;
371 	mac_soft_ring_t		*softring;
372 
373 	ASSERT(mac_srs->srs_mcip == mcip);
374 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
375 
376 	if (!(mcip->mci_state_flags & MCIS_CLIENT_POLL_CAPABLE))
377 		return;
378 
379 	bzero(&mrf, sizeof (mac_rx_fifo_t));
380 	mrf.mrf_type = MAC_RX_FIFO;
381 
382 	/*
383 	 * A SRS is capable of acting as a soft ring for cases
384 	 * where no fanout is needed. This is the case for userland
385 	 * flows.
386 	 */
387 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS)
388 		return;
389 
390 	mrf.mrf_receive = (mac_receive_t)mac_soft_ring_poll;
391 	mrf.mrf_intr_enable = (mac_intr_enable_t)mac_soft_ring_intr_enable;
392 	mrf.mrf_intr_disable = (mac_intr_disable_t)mac_soft_ring_intr_disable;
393 	mac_srs->srs_type |= SRST_CLIENT_POLL_ENABLED;
394 
395 	softring = mac_srs->srs_soft_ring_head;
396 	while (softring != NULL) {
397 		if (softring->s_ring_type & (ST_RING_TCP | ST_RING_UDP)) {
398 			/*
399 			 * TCP and UDP support DLS bypass. Squeue polling
400 			 * support implies DLS bypass since the squeue poll
401 			 * path does not have DLS processing.
402 			 */
403 			mac_soft_ring_dls_bypass(softring,
404 			    mcip->mci_direct_rx_fn, mcip->mci_direct_rx_arg);
405 		}
406 		/*
407 		 * Non-TCP protocols don't support squeues. Hence we don't
408 		 * make any ring addition callbacks for non-TCP rings
409 		 */
410 		if (!(softring->s_ring_type & ST_RING_TCP)) {
411 			softring->s_ring_rx_arg2 = NULL;
412 			softring = softring->s_ring_next;
413 			continue;
414 		}
415 		mrf.mrf_rx_arg = softring;
416 		mrf.mrf_intr_handle = (mac_intr_handle_t)softring;
417 		mrf.mrf_cpu_id = softring->s_ring_cpuid;
418 		mrf.mrf_flow_priority = mac_srs->srs_pri;
419 
420 		softring->s_ring_rx_arg2 = mcip->mci_resource_add(
421 		    mcip->mci_resource_arg, (mac_resource_t *)&mrf);
422 
423 		softring = softring->s_ring_next;
424 	}
425 }
426 
427 /*
428  * Unregister the given SRS and associated soft rings with the consumer and
429  * disable the polling interface used by the consumer.(i.e IP) over this
430  * SRS and associated soft rings.
431  */
432 void
433 mac_srs_client_poll_disable(mac_client_impl_t *mcip,
434     mac_soft_ring_set_t *mac_srs)
435 {
436 	mac_soft_ring_t		*softring;
437 
438 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
439 
440 	/*
441 	 * A SRS is capable of acting as a soft ring for cases
442 	 * where no protocol fanout is needed. This is the case
443 	 * for userland flows. Nothing to do here.
444 	 */
445 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS)
446 		return;
447 
448 	mutex_enter(&mac_srs->srs_lock);
449 	if (!(mac_srs->srs_type & SRST_CLIENT_POLL_ENABLED)) {
450 		ASSERT(!(mac_srs->srs_type & SRST_DLS_BYPASS));
451 		mutex_exit(&mac_srs->srs_lock);
452 		return;
453 	}
454 	mac_srs->srs_type &= ~(SRST_CLIENT_POLL_ENABLED | SRST_DLS_BYPASS);
455 	mutex_exit(&mac_srs->srs_lock);
456 
457 	/*
458 	 * DLS bypass is now disabled in the case of both TCP and UDP.
459 	 * Reset the soft ring callbacks to the standard 'mac_rx_deliver'
460 	 * callback. In addition, in the case of TCP, invoke IP's callback
461 	 * for ring removal.
462 	 */
463 	for (softring = mac_srs->srs_soft_ring_head;
464 	    softring != NULL; softring = softring->s_ring_next) {
465 		if (!(softring->s_ring_type & (ST_RING_UDP | ST_RING_TCP)))
466 			continue;
467 
468 		if ((softring->s_ring_type & ST_RING_TCP) &&
469 		    softring->s_ring_rx_arg2 != NULL) {
470 			mcip->mci_resource_remove(mcip->mci_resource_arg,
471 			    softring->s_ring_rx_arg2);
472 		}
473 
474 		mutex_enter(&softring->s_ring_lock);
475 		while (softring->s_ring_state & S_RING_PROC) {
476 			softring->s_ring_state |= S_RING_CLIENT_WAIT;
477 			cv_wait(&softring->s_ring_client_cv,
478 			    &softring->s_ring_lock);
479 		}
480 		softring->s_ring_state &= ~S_RING_CLIENT_WAIT;
481 		softring->s_ring_rx_arg2 = NULL;
482 		softring->s_ring_rx_func = mac_rx_deliver;
483 		softring->s_ring_rx_arg1 = mcip;
484 		mutex_exit(&softring->s_ring_lock);
485 	}
486 }
487 
488 /*
489  * Enable or disable poll capability of the SRS on the underlying Rx ring.
490  *
491  * There is a need to enable or disable the poll capability of an SRS over an
492  * Rx ring depending on the number of mac clients sharing the ring and also
493  * whether user flows are configured on it. However the poll state is actively
494  * manipulated by the SRS worker and poll threads and uncoordinated changes by
495  * yet another thread to the underlying capability can surprise them leading
496  * to assert failures. Instead we quiesce the SRS, make the changes and then
497  * restart the SRS.
498  */
499 static void
500 mac_srs_poll_state_change(mac_soft_ring_set_t *mac_srs,
501     boolean_t turn_off_poll_capab, mac_rx_func_t rx_func)
502 {
503 	boolean_t	need_restart = B_FALSE;
504 	mac_srs_rx_t	*srs_rx = &mac_srs->srs_rx;
505 	mac_ring_t	*ring;
506 
507 	if (!SRS_QUIESCED(mac_srs)) {
508 		mac_rx_srs_quiesce(mac_srs, SRS_QUIESCE);
509 		need_restart = B_TRUE;
510 	}
511 
512 	ring = mac_srs->srs_ring;
513 	if ((ring != NULL) &&
514 	    (ring->mr_classify_type == MAC_HW_CLASSIFIER)) {
515 		if (turn_off_poll_capab)
516 			mac_srs->srs_state &= ~SRS_POLLING_CAPAB;
517 		else if (mac_poll_enable)
518 			mac_srs->srs_state |= SRS_POLLING_CAPAB;
519 	}
520 	srs_rx->sr_lower_proc = rx_func;
521 
522 	if (need_restart)
523 		mac_rx_srs_restart(mac_srs);
524 }
525 
526 /* CPU RECONFIGURATION AND FANOUT COMPUTATION ROUTINES */
527 
528 /*
529  * Return the next CPU to be used to bind a MAC kernel thread.
530  * If a cpupart is specified, the cpu chosen must be from that
531  * cpu partition.
532  */
533 static processorid_t
534 mac_next_bind_cpu(cpupart_t *cpupart)
535 {
536 	static cpu_t		*cp = NULL;
537 	cpu_t			*cp_start;
538 
539 	ASSERT(MUTEX_HELD(&cpu_lock));
540 
541 	if (cp == NULL)
542 		cp = cpu_list;
543 
544 	cp = cp->cpu_next_onln;
545 	cp_start = cp;
546 
547 	do {
548 		if ((cpupart == NULL) || (cp->cpu_part == cpupart))
549 			return (cp->cpu_id);
550 
551 	} while ((cp = cp->cpu_next_onln) != cp_start);
552 
553 	return (-1);	/* No matching CPU found online */
554 }
555 
556 /* ARGSUSED */
557 static int
558 mac_srs_cpu_setup(cpu_setup_t what, int id, void *arg)
559 {
560 	ASSERT(MUTEX_HELD(&cpu_lock));
561 	switch (what) {
562 	case CPU_CONFIG:
563 	case CPU_ON:
564 	case CPU_CPUPART_IN:
565 		mac_walk_srs_and_bind(id);
566 		break;
567 
568 	case CPU_UNCONFIG:
569 	case CPU_OFF:
570 	case CPU_CPUPART_OUT:
571 		mac_walk_srs_and_unbind(id);
572 		break;
573 
574 	default:
575 		break;
576 	}
577 	return (0);
578 }
579 
580 /*
581  * mac_compute_soft_ring_count():
582  *
583  * This routine computes the number of soft rings needed to handle incoming
584  * load given a flow_entry.
585  *
586  * The routine does the following:
587  * 1) soft rings will be created if mac_soft_ring_enable is set.
588  * 2) If the underlying link is a 10Gbps link, then soft rings will be
589  * created even if mac_soft_ring_enable is not set. The number of soft
590  * rings, so created,  will equal mac_rx_soft_ring_10gig_count.
591  * 3) On a sun4v platform (i.e., mac_soft_ring_enable is set), 2 times the
592  * mac_rx_soft_ring_10gig_count number of soft rings will be created for a
593  * 10Gbps link.
594  *
595  * If a bandwidth limit is specified, the number that gets computed is
596  * dependent upon CPU speed, the number of Rx rings configured, and
597  * the bandwidth limit.
598  * If more Rx rings are available, less number of soft rings is needed.
599  *
600  * mac_use_bw_heuristic is another "hidden" variable that can be used to
601  * override the default use of soft ring count computation. Depending upon
602  * the usefulness of it, mac_use_bw_heuristic can later be made into a
603  * data-link property or removed altogether.
604  *
605  * TODO: Cleanup and tighten some of the assumptions.
606  */
607 boolean_t mac_use_bw_heuristic = B_TRUE;
608 static int
609 mac_compute_soft_ring_count(flow_entry_t *flent, int rx_srs_cnt, int maxcpus)
610 {
611 	uint64_t cpu_speed, bw = 0;
612 	int srings = 0;
613 	boolean_t bw_enabled = B_FALSE;
614 
615 	ASSERT(!(flent->fe_type & FLOW_USER));
616 	if (flent->fe_resource_props.mrp_mask & MRP_MAXBW &&
617 	    mac_use_bw_heuristic) {
618 		/* bandwidth enabled */
619 		bw_enabled = B_TRUE;
620 		bw = flent->fe_resource_props.mrp_maxbw;
621 	}
622 	if (!bw_enabled) {
623 		/* No bandwidth enabled */
624 		if (mac_soft_ring_enable)
625 			srings = mac_rx_soft_ring_count;
626 
627 		/* Is this a 10Gig link? */
628 		flent->fe_nic_speed = mac_client_stat_get(flent->fe_mcip,
629 		    MAC_STAT_IFSPEED);
630 		/* convert to Mbps */
631 		if (((flent->fe_nic_speed)/1000000) > 1000 &&
632 		    mac_rx_soft_ring_10gig_count > 0) {
633 			/* This is a 10Gig link */
634 			srings = mac_rx_soft_ring_10gig_count;
635 			/*
636 			 * Use 2 times mac_rx_soft_ring_10gig_count for
637 			 * sun4v systems.
638 			 */
639 			if (mac_soft_ring_enable)
640 				srings = srings * 2;
641 		}
642 	} else {
643 		/*
644 		 * Soft ring computation using CPU speed and specified
645 		 * bandwidth limit.
646 		 */
647 		/* Assumption: all CPUs have the same frequency */
648 		cpu_speed = (uint64_t)CPU->cpu_type_info.pi_clock;
649 
650 		/* cpu_speed is in MHz; make bw in units of Mbps.  */
651 		bw = bw/1000000;
652 
653 		if (bw >= 1000) {
654 			/*
655 			 * bw is greater than or equal to 1Gbps.
656 			 * The number of soft rings required is a function
657 			 * of bandwidth and CPU speed. To keep this simple,
658 			 * let's use this rule: 1GHz CPU can handle 1Gbps.
659 			 * If bw is less than 1 Gbps, then there is no need
660 			 * for soft rings. Assumption is that CPU speeds
661 			 * (on modern systems) are at least 1GHz.
662 			 */
663 			srings = bw/cpu_speed;
664 			if (srings <= 1 && mac_soft_ring_enable) {
665 				/*
666 				 * Give at least 2 soft rings
667 				 * for sun4v systems
668 				 */
669 				srings = 2;
670 			}
671 		}
672 	}
673 	/*
674 	 * If the flent has multiple Rx SRSs, then each SRS need not
675 	 * have that many soft rings on top of it. The number of
676 	 * soft rings for each Rx SRS is found by dividing srings by
677 	 * rx_srs_cnt.
678 	 */
679 	if (rx_srs_cnt > 1) {
680 		int remainder;
681 
682 		remainder = srings%rx_srs_cnt;
683 		srings = srings/rx_srs_cnt;
684 		if (remainder != 0)
685 			srings++;
686 		/*
687 		 * Fanning out to 1 soft ring is not very useful.
688 		 * Set it as well to 0 and mac_srs_fanout_init()
689 		 * will take care of creating a single soft ring
690 		 * for proto fanout.
691 		 */
692 		if (srings == 1)
693 			srings = 0;
694 	}
695 	/* Do some more massaging */
696 	srings = min(srings, maxcpus);
697 	srings = min(srings, MAX_SR_FANOUT);
698 	return (srings);
699 }
700 
701 /*
702  * mac_tx_cpu_init:
703  * set up CPUs for Tx interrupt re-targeting and Tx worker
704  * thread binding
705  */
706 static void
707 mac_tx_cpu_init(flow_entry_t *flent, mac_resource_props_t *mrp,
708     cpupart_t *cpupart)
709 {
710 	mac_soft_ring_set_t *tx_srs = flent->fe_tx_srs;
711 	mac_srs_tx_t *srs_tx = &tx_srs->srs_tx;
712 	mac_cpus_t *srs_cpu = &tx_srs->srs_cpu;
713 	mac_soft_ring_t *sringp;
714 	mac_ring_t *ring;
715 	processorid_t worker_cpuid;
716 	boolean_t retargetable_client = B_FALSE;
717 	int i, j;
718 
719 	if (RETARGETABLE_CLIENT((mac_group_t *)flent->fe_tx_ring_group,
720 	    flent->fe_mcip)) {
721 		retargetable_client = B_TRUE;
722 	}
723 
724 	if (MAC_TX_SOFT_RINGS(tx_srs)) {
725 		if (mrp != NULL)
726 			j = mrp->mrp_ncpus - 1;
727 		for (i = 0; i < tx_srs->srs_tx_ring_count; i++) {
728 			if (mrp != NULL) {
729 				if (j < 0)
730 					j = mrp->mrp_ncpus - 1;
731 				worker_cpuid = mrp->mrp_cpu[j];
732 			} else {
733 				/*
734 				 * Bind interrupt to the next CPU available
735 				 * and leave the worker unbound.
736 				 */
737 				worker_cpuid = -1;
738 			}
739 			sringp = tx_srs->srs_tx_soft_rings[i];
740 			ring = (mac_ring_t *)sringp->s_ring_tx_arg2;
741 			srs_cpu->mc_tx_fanout_cpus[i] = worker_cpuid;
742 			if (MAC_RING_RETARGETABLE(ring) &&
743 			    retargetable_client) {
744 				mutex_enter(&cpu_lock);
745 				srs_cpu->mc_tx_intr_cpu[i] =
746 				    (mrp != NULL) ? mrp->mrp_cpu[j] :
747 				    (mac_tx_intr_retarget ?
748 				    mac_next_bind_cpu(cpupart) : -1);
749 				mutex_exit(&cpu_lock);
750 			} else {
751 				srs_cpu->mc_tx_intr_cpu[i] = -1;
752 			}
753 			if (mrp != NULL)
754 				j--;
755 		}
756 	} else {
757 		/* Tx mac_ring_handle_t is stored in st_arg2 */
758 		srs_cpu->mc_tx_fanout_cpus[0] =
759 		    (mrp != NULL) ? mrp->mrp_cpu[mrp->mrp_ncpus - 1] : -1;
760 		ring = (mac_ring_t *)srs_tx->st_arg2;
761 		if (MAC_RING_RETARGETABLE(ring) && retargetable_client) {
762 			mutex_enter(&cpu_lock);
763 			srs_cpu->mc_tx_intr_cpu[0] = (mrp != NULL) ?
764 			    mrp->mrp_cpu[mrp->mrp_ncpus - 1] :
765 			    (mac_tx_intr_retarget ?
766 			    mac_next_bind_cpu(cpupart) : -1);
767 			mutex_exit(&cpu_lock);
768 		} else {
769 			srs_cpu->mc_tx_intr_cpu[0] = -1;
770 		}
771 	}
772 }
773 
774 /*
775  * Assignment of user specified CPUs to a link.
776  *
777  * Minimum CPUs required to get an optimal assignmet:
778  * For each Rx SRS, atleast two CPUs are needed if mac_latency_optimize
779  * flag is set -- one for polling, one for fanout soft ring.
780  * If mac_latency_optimize is not set, then 3 CPUs are needed -- one
781  * for polling, one for SRS worker thread and one for fanout soft ring.
782  *
783  * The CPUs needed for Tx side is equal to the number of Tx rings
784  * the link is using.
785  *
786  * mac_flow_user_cpu_init() categorizes the CPU assignment depending
787  * upon the number of CPUs in 3 different buckets.
788  *
789  * In the first bucket, the most optimal case is handled. The user has
790  * passed enough number of CPUs and every thread gets its own CPU.
791  *
792  * The second and third are the sub-optimal cases. Enough CPUs are not
793  * available.
794  *
795  * The second bucket handles the case where atleast one distinct CPU is
796  * is available for each of the Rx rings (Rx SRSes) and Tx rings (Tx
797  * SRS or soft rings).
798  *
799  * In the third case (worst case scenario), specified CPU count is less
800  * than the Rx rings configured for the link. In this case, we round
801  * robin the CPUs among the Rx SRSes and Tx SRS/soft rings.
802  */
803 static void
804 mac_flow_user_cpu_init(flow_entry_t *flent, mac_resource_props_t *mrp)
805 {
806 	mac_soft_ring_set_t *rx_srs, *tx_srs;
807 	int i, srs_cnt;
808 	mac_cpus_t *srs_cpu;
809 	int no_of_cpus, cpu_cnt;
810 	int rx_srs_cnt, reqd_rx_cpu_cnt;
811 	int fanout_cpu_cnt, reqd_tx_cpu_cnt;
812 	int reqd_poll_worker_cnt, fanout_cnt_per_srs;
813 	mac_resource_props_t *emrp = &flent->fe_effective_props;
814 
815 	ASSERT(mrp->mrp_fanout_mode == MCM_CPUS);
816 	/*
817 	 * The check for nbc_ncpus to be within limits for
818 	 * the user specified case was done earlier and if
819 	 * not within limits, an error would have been
820 	 * returned to the user.
821 	 */
822 	ASSERT(mrp->mrp_ncpus > 0);
823 
824 	no_of_cpus = mrp->mrp_ncpus;
825 
826 	if (mrp->mrp_rx_intr_cpu != -1) {
827 		/*
828 		 * interrupt has been re-targetted. Poll
829 		 * thread needs to be bound to interrupt
830 		 * CPU.
831 		 *
832 		 * Find where in the list is the intr
833 		 * CPU and swap it with the first one.
834 		 * We will be using the first CPU in the
835 		 * list for poll.
836 		 */
837 		for (i = 0; i < no_of_cpus; i++) {
838 			if (mrp->mrp_cpu[i] == mrp->mrp_rx_intr_cpu)
839 				break;
840 		}
841 		mrp->mrp_cpu[i] = mrp->mrp_cpu[0];
842 		mrp->mrp_cpu[0] = mrp->mrp_rx_intr_cpu;
843 	}
844 
845 	/*
846 	 * Requirements:
847 	 * The number of CPUs that each Rx ring needs is dependent
848 	 * upon mac_latency_optimize flag.
849 	 * 1) If set, atleast 2 CPUs are needed -- one for
850 	 * polling, one for fanout soft ring.
851 	 * 2) If not set, then atleast 3 CPUs are needed -- one
852 	 * for polling, one for srs worker thread, and one for
853 	 * fanout soft ring.
854 	 */
855 	rx_srs_cnt = (flent->fe_rx_srs_cnt > 1) ?
856 	    (flent->fe_rx_srs_cnt - 1) : flent->fe_rx_srs_cnt;
857 	reqd_rx_cpu_cnt = mac_latency_optimize ?
858 	    (rx_srs_cnt * 2) : (rx_srs_cnt * 3);
859 
860 	/* How many CPUs are needed for Tx side? */
861 	tx_srs = flent->fe_tx_srs;
862 	reqd_tx_cpu_cnt = MAC_TX_SOFT_RINGS(tx_srs) ?
863 	    tx_srs->srs_tx_ring_count : 1;
864 
865 	/* CPUs needed for Rx SRSes poll and worker threads */
866 	reqd_poll_worker_cnt = mac_latency_optimize ?
867 	    rx_srs_cnt : rx_srs_cnt * 2;
868 
869 	/* Has the user provided enough CPUs? */
870 	if (no_of_cpus >= (reqd_rx_cpu_cnt + reqd_tx_cpu_cnt)) {
871 		/*
872 		 * Best case scenario. There is enough CPUs. All
873 		 * Rx rings will get their own set of CPUs plus
874 		 * Tx soft rings will get their own.
875 		 */
876 		/*
877 		 * fanout_cpu_cnt is the number of CPUs available
878 		 * for Rx side fanout soft rings.
879 		 */
880 		fanout_cpu_cnt = no_of_cpus -
881 		    reqd_poll_worker_cnt - reqd_tx_cpu_cnt;
882 
883 		/*
884 		 * Divide fanout_cpu_cnt by rx_srs_cnt to find
885 		 * out how many fanout soft rings each Rx SRS
886 		 * can have.
887 		 */
888 		fanout_cnt_per_srs = fanout_cpu_cnt/rx_srs_cnt;
889 
890 		/* fanout_cnt_per_srs should not be >  MAX_SR_FANOUT */
891 		fanout_cnt_per_srs = min(fanout_cnt_per_srs, MAX_SR_FANOUT);
892 
893 		/* Do the assignment for the default Rx ring */
894 		cpu_cnt = 0;
895 		rx_srs = flent->fe_rx_srs[0];
896 		ASSERT(rx_srs->srs_ring == NULL);
897 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
898 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
899 		srs_cpu = &rx_srs->srs_cpu;
900 		srs_cpu->mc_ncpus = no_of_cpus;
901 		bcopy(mrp->mrp_cpu,
902 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
903 		srs_cpu->mc_rx_fanout_cnt = fanout_cnt_per_srs;
904 		srs_cpu->mc_rx_pollid = mrp->mrp_cpu[cpu_cnt++];
905 		/* Retarget the interrupt to the same CPU as the poll */
906 		srs_cpu->mc_rx_intr_cpu = srs_cpu->mc_rx_pollid;
907 		srs_cpu->mc_rx_workerid = (mac_latency_optimize ?
908 		    srs_cpu->mc_rx_pollid : mrp->mrp_cpu[cpu_cnt++]);
909 		for (i = 0; i < fanout_cnt_per_srs; i++)
910 			srs_cpu->mc_rx_fanout_cpus[i] = mrp->mrp_cpu[cpu_cnt++];
911 
912 		/* Do the assignment for h/w Rx SRSes */
913 		if (flent->fe_rx_srs_cnt > 1) {
914 			cpu_cnt = 0;
915 			for (srs_cnt = 1;
916 			    srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
917 				rx_srs = flent->fe_rx_srs[srs_cnt];
918 				ASSERT(rx_srs->srs_ring != NULL);
919 				if (rx_srs->srs_fanout_state ==
920 				    SRS_FANOUT_INIT) {
921 					rx_srs->srs_fanout_state =
922 					    SRS_FANOUT_REINIT;
923 				}
924 				srs_cpu = &rx_srs->srs_cpu;
925 				srs_cpu->mc_ncpus = no_of_cpus;
926 				bcopy(mrp->mrp_cpu, srs_cpu->mc_cpus,
927 				    sizeof (srs_cpu->mc_cpus));
928 				srs_cpu->mc_rx_fanout_cnt = fanout_cnt_per_srs;
929 				/* The first CPU in the list is the intr CPU */
930 				srs_cpu->mc_rx_pollid = mrp->mrp_cpu[cpu_cnt++];
931 				srs_cpu->mc_rx_intr_cpu = srs_cpu->mc_rx_pollid;
932 				srs_cpu->mc_rx_workerid =
933 				    (mac_latency_optimize ?
934 				    srs_cpu->mc_rx_pollid :
935 				    mrp->mrp_cpu[cpu_cnt++]);
936 				for (i = 0; i < fanout_cnt_per_srs; i++) {
937 					srs_cpu->mc_rx_fanout_cpus[i] =
938 					    mrp->mrp_cpu[cpu_cnt++];
939 				}
940 				ASSERT(cpu_cnt <= no_of_cpus);
941 			}
942 		}
943 		goto tx_cpu_init;
944 	}
945 
946 	/*
947 	 * Sub-optimal case.
948 	 * We have the following information:
949 	 * no_of_cpus - no. of cpus that user passed.
950 	 * rx_srs_cnt - no. of rx rings.
951 	 * reqd_rx_cpu_cnt = mac_latency_optimize?rx_srs_cnt*2:rx_srs_cnt*3
952 	 * reqd_tx_cpu_cnt - no. of cpus reqd. for Tx side.
953 	 * reqd_poll_worker_cnt = mac_latency_optimize?rx_srs_cnt:rx_srs_cnt*2
954 	 */
955 	/*
956 	 * If we bind the Rx fanout soft rings to the same CPUs
957 	 * as poll/worker, would that be enough?
958 	 */
959 	if (no_of_cpus >= (rx_srs_cnt + reqd_tx_cpu_cnt)) {
960 		boolean_t worker_assign = B_FALSE;
961 
962 		/*
963 		 * If mac_latency_optimize is not set, are there
964 		 * enough CPUs to assign a CPU for worker also?
965 		 */
966 		if (no_of_cpus >= (reqd_poll_worker_cnt + reqd_tx_cpu_cnt))
967 			worker_assign = B_TRUE;
968 		/*
969 		 * Zero'th Rx SRS is the default Rx ring. It is not
970 		 * associated with h/w Rx ring.
971 		 */
972 		rx_srs = flent->fe_rx_srs[0];
973 		ASSERT(rx_srs->srs_ring == NULL);
974 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
975 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
976 		cpu_cnt = 0;
977 		srs_cpu = &rx_srs->srs_cpu;
978 		srs_cpu->mc_ncpus = no_of_cpus;
979 		bcopy(mrp->mrp_cpu,
980 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
981 		srs_cpu->mc_rx_fanout_cnt = 1;
982 		srs_cpu->mc_rx_pollid = mrp->mrp_cpu[cpu_cnt++];
983 		/* Retarget the interrupt to the same CPU as the poll */
984 		srs_cpu->mc_rx_intr_cpu = srs_cpu->mc_rx_pollid;
985 		srs_cpu->mc_rx_workerid =
986 		    ((!mac_latency_optimize && worker_assign) ?
987 		    mrp->mrp_cpu[cpu_cnt++] : srs_cpu->mc_rx_pollid);
988 
989 		srs_cpu->mc_rx_fanout_cpus[0] = mrp->mrp_cpu[cpu_cnt];
990 
991 		/* Do CPU bindings for SRSes having h/w Rx rings */
992 		if (flent->fe_rx_srs_cnt > 1) {
993 			cpu_cnt = 0;
994 			for (srs_cnt = 1;
995 			    srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
996 				rx_srs = flent->fe_rx_srs[srs_cnt];
997 				ASSERT(rx_srs->srs_ring != NULL);
998 				if (rx_srs->srs_fanout_state ==
999 				    SRS_FANOUT_INIT) {
1000 					rx_srs->srs_fanout_state =
1001 					    SRS_FANOUT_REINIT;
1002 				}
1003 				srs_cpu = &rx_srs->srs_cpu;
1004 				srs_cpu->mc_ncpus = no_of_cpus;
1005 				bcopy(mrp->mrp_cpu, srs_cpu->mc_cpus,
1006 				    sizeof (srs_cpu->mc_cpus));
1007 				srs_cpu->mc_rx_pollid =
1008 				    mrp->mrp_cpu[cpu_cnt];
1009 				srs_cpu->mc_rx_intr_cpu = srs_cpu->mc_rx_pollid;
1010 				srs_cpu->mc_rx_workerid =
1011 				    ((!mac_latency_optimize && worker_assign) ?
1012 				    mrp->mrp_cpu[++cpu_cnt] :
1013 				    srs_cpu->mc_rx_pollid);
1014 				srs_cpu->mc_rx_fanout_cnt = 1;
1015 				srs_cpu->mc_rx_fanout_cpus[0] =
1016 				    mrp->mrp_cpu[cpu_cnt];
1017 				cpu_cnt++;
1018 				ASSERT(cpu_cnt <= no_of_cpus);
1019 			}
1020 		}
1021 		goto tx_cpu_init;
1022 	}
1023 
1024 	/*
1025 	 * Real sub-optimal case. Not enough CPUs for poll and
1026 	 * Tx soft rings. Do a round robin assignment where
1027 	 * each Rx SRS will get the same CPU for poll, worker
1028 	 * and fanout soft ring.
1029 	 */
1030 	cpu_cnt = 0;
1031 	for (srs_cnt = 0; srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
1032 		rx_srs = flent->fe_rx_srs[srs_cnt];
1033 		srs_cpu = &rx_srs->srs_cpu;
1034 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
1035 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
1036 		srs_cpu->mc_ncpus = no_of_cpus;
1037 		bcopy(mrp->mrp_cpu,
1038 		    srs_cpu->mc_cpus, sizeof (srs_cpu->mc_cpus));
1039 		srs_cpu->mc_rx_fanout_cnt = 1;
1040 		srs_cpu->mc_rx_pollid = mrp->mrp_cpu[cpu_cnt];
1041 		/* Retarget the interrupt to the same CPU as the poll */
1042 		srs_cpu->mc_rx_intr_cpu = srs_cpu->mc_rx_pollid;
1043 		srs_cpu->mc_rx_workerid = mrp->mrp_cpu[cpu_cnt];
1044 		srs_cpu->mc_rx_fanout_cpus[0] = mrp->mrp_cpu[cpu_cnt];
1045 		if (++cpu_cnt >= no_of_cpus)
1046 			cpu_cnt = 0;
1047 	}
1048 
1049 tx_cpu_init:
1050 	mac_tx_cpu_init(flent, mrp, NULL);
1051 
1052 	/*
1053 	 * Copy the user specified CPUs to the effective CPUs
1054 	 */
1055 	for (i = 0; i < mrp->mrp_ncpus; i++) {
1056 		emrp->mrp_cpu[i] = mrp->mrp_cpu[i];
1057 	}
1058 	emrp->mrp_ncpus = mrp->mrp_ncpus;
1059 	emrp->mrp_mask = mrp->mrp_mask;
1060 	bzero(emrp->mrp_pool, MAXPATHLEN);
1061 }
1062 
1063 /*
1064  * mac_flow_cpu_init():
1065  *
1066  * Each SRS has a mac_cpu_t structure, srs_cpu. This routine fills in
1067  * the CPU binding information in srs_cpu for all Rx SRSes associated
1068  * with a flent.
1069  */
1070 static void
1071 mac_flow_cpu_init(flow_entry_t *flent, cpupart_t *cpupart)
1072 {
1073 	mac_soft_ring_set_t *rx_srs;
1074 	processorid_t cpuid;
1075 	int i, j, k, srs_cnt, nscpus, maxcpus, soft_ring_cnt = 0;
1076 	mac_cpus_t *srs_cpu;
1077 	mac_resource_props_t *emrp = &flent->fe_effective_props;
1078 	uint32_t cpus[MRP_NCPUS];
1079 
1080 	/*
1081 	 * The maximum number of CPUs available can either be
1082 	 * the number of CPUs in the pool or the number of CPUs
1083 	 * in the system.
1084 	 */
1085 	maxcpus = (cpupart != NULL) ? cpupart->cp_ncpus : ncpus;
1086 
1087 	/*
1088 	 * Compute the number of soft rings needed on top for each Rx
1089 	 * SRS. "rx_srs_cnt-1" indicates the number of Rx SRS
1090 	 * associated with h/w Rx rings. Soft ring count needed for
1091 	 * each h/w Rx SRS is computed and the same is applied to
1092 	 * software classified Rx SRS. The first Rx SRS in fe_rx_srs[]
1093 	 * is the software classified Rx SRS.
1094 	 */
1095 	soft_ring_cnt = mac_compute_soft_ring_count(flent,
1096 	    flent->fe_rx_srs_cnt - 1, maxcpus);
1097 	if (soft_ring_cnt == 0) {
1098 		/*
1099 		 * Even when soft_ring_cnt is 0, we still need
1100 		 * to create a soft ring for TCP, UDP and
1101 		 * OTHER. So set it to 1.
1102 		 */
1103 		soft_ring_cnt = 1;
1104 	}
1105 	for (srs_cnt = 0; srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
1106 		rx_srs = flent->fe_rx_srs[srs_cnt];
1107 		srs_cpu = &rx_srs->srs_cpu;
1108 		if (rx_srs->srs_fanout_state == SRS_FANOUT_INIT)
1109 			rx_srs->srs_fanout_state = SRS_FANOUT_REINIT;
1110 		srs_cpu->mc_ncpus = soft_ring_cnt;
1111 		srs_cpu->mc_rx_fanout_cnt = soft_ring_cnt;
1112 		mutex_enter(&cpu_lock);
1113 		for (j = 0; j < soft_ring_cnt; j++) {
1114 			cpuid = mac_next_bind_cpu(cpupart);
1115 			srs_cpu->mc_cpus[j] = cpuid;
1116 			srs_cpu->mc_rx_fanout_cpus[j] = cpuid;
1117 		}
1118 		cpuid = mac_next_bind_cpu(cpupart);
1119 		srs_cpu->mc_rx_pollid = cpuid;
1120 		srs_cpu->mc_rx_intr_cpu = (mac_rx_intr_retarget ?
1121 		    srs_cpu->mc_rx_pollid : -1);
1122 		/* increment ncpus to account for polling cpu */
1123 		srs_cpu->mc_ncpus++;
1124 		srs_cpu->mc_cpus[j++] = cpuid;
1125 		if (!mac_latency_optimize) {
1126 			cpuid = mac_next_bind_cpu(cpupart);
1127 			srs_cpu->mc_ncpus++;
1128 			srs_cpu->mc_cpus[j++] = cpuid;
1129 		}
1130 		srs_cpu->mc_rx_workerid = cpuid;
1131 		mutex_exit(&cpu_lock);
1132 	}
1133 
1134 	nscpus = 0;
1135 	for (srs_cnt = 0; srs_cnt < flent->fe_rx_srs_cnt; srs_cnt++) {
1136 		rx_srs = flent->fe_rx_srs[srs_cnt];
1137 		srs_cpu = &rx_srs->srs_cpu;
1138 		for (j = 0; j < srs_cpu->mc_ncpus; j++) {
1139 			cpus[nscpus++] = srs_cpu->mc_cpus[j];
1140 		}
1141 	}
1142 
1143 
1144 	/*
1145 	 * Copy cpu list to fe_effective_props
1146 	 * without duplicates.
1147 	 */
1148 	k = 0;
1149 	for (i = 0; i < nscpus; i++) {
1150 		for (j = 0; j < k; j++) {
1151 			if (emrp->mrp_cpu[j] == cpus[i])
1152 				break;
1153 		}
1154 		if (j == k)
1155 			emrp->mrp_cpu[k++] = cpus[i];
1156 	}
1157 	emrp->mrp_ncpus = k;
1158 
1159 	mac_tx_cpu_init(flent, NULL, cpupart);
1160 }
1161 
1162 /*
1163  * DATAPATH SETUP ROUTINES
1164  * (setup SRS and set/update FANOUT, B/W and PRIORITY)
1165  */
1166 
1167 /*
1168  * mac_srs_fanout_list_alloc:
1169  *
1170  * The underlying device can expose upto MAX_RINGS_PER_GROUP worth of
1171  * rings to a client. In such a case, MAX_RINGS_PER_GROUP worth of
1172  * array space is needed to store Tx soft rings. Thus we allocate so
1173  * much array space for srs_tx_soft_rings.
1174  *
1175  * And when it is an aggr, again we allocate MAX_RINGS_PER_GROUP worth
1176  * of space to st_soft_rings. This array is used for quick access to
1177  * soft ring associated with a pseudo Tx ring based on the pseudo
1178  * ring's index (mr_index).
1179  */
1180 static void
1181 mac_srs_fanout_list_alloc(mac_soft_ring_set_t *mac_srs)
1182 {
1183 	mac_client_impl_t *mcip = mac_srs->srs_mcip;
1184 
1185 	if (mac_srs->srs_type & SRST_TX) {
1186 		mac_srs->srs_tx_soft_rings = (mac_soft_ring_t **)
1187 		    kmem_zalloc(sizeof (mac_soft_ring_t *) *
1188 		    MAX_RINGS_PER_GROUP, KM_SLEEP);
1189 		if (mcip->mci_state_flags & MCIS_IS_AGGR) {
1190 			mac_srs_tx_t *tx = &mac_srs->srs_tx;
1191 
1192 			tx->st_soft_rings = (mac_soft_ring_t **)
1193 			    kmem_zalloc(sizeof (mac_soft_ring_t *) *
1194 			    MAX_RINGS_PER_GROUP, KM_SLEEP);
1195 		}
1196 	} else {
1197 		mac_srs->srs_tcp_soft_rings = (mac_soft_ring_t **)
1198 		    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT,
1199 		    KM_SLEEP);
1200 		mac_srs->srs_udp_soft_rings = (mac_soft_ring_t **)
1201 		    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT,
1202 		    KM_SLEEP);
1203 		mac_srs->srs_oth_soft_rings = (mac_soft_ring_t **)
1204 		    kmem_zalloc(sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT,
1205 		    KM_SLEEP);
1206 	}
1207 }
1208 
1209 static void
1210 mac_srs_worker_bind(mac_soft_ring_set_t *mac_srs, processorid_t cpuid)
1211 {
1212 	cpu_t *cp;
1213 	boolean_t clear = B_FALSE;
1214 
1215 	ASSERT(MUTEX_HELD(&cpu_lock));
1216 
1217 	if (!mac_srs_thread_bind)
1218 		return;
1219 
1220 	cp = cpu_get(cpuid);
1221 	if (cp == NULL || !cpu_is_online(cp))
1222 		return;
1223 
1224 	mutex_enter(&mac_srs->srs_lock);
1225 	mac_srs->srs_state |= SRS_WORKER_BOUND;
1226 	if (mac_srs->srs_worker_cpuid != -1)
1227 		clear = B_TRUE;
1228 	mac_srs->srs_worker_cpuid = cpuid;
1229 	mutex_exit(&mac_srs->srs_lock);
1230 
1231 	if (clear)
1232 		thread_affinity_clear(mac_srs->srs_worker);
1233 
1234 	thread_affinity_set(mac_srs->srs_worker, cpuid);
1235 	DTRACE_PROBE1(worker__CPU, processorid_t, cpuid);
1236 }
1237 
1238 static void
1239 mac_srs_poll_bind(mac_soft_ring_set_t *mac_srs, processorid_t cpuid)
1240 {
1241 	cpu_t *cp;
1242 	boolean_t clear = B_FALSE;
1243 
1244 	ASSERT(MUTEX_HELD(&cpu_lock));
1245 
1246 	if (!mac_srs_thread_bind || mac_srs->srs_poll_thr == NULL)
1247 		return;
1248 
1249 	cp = cpu_get(cpuid);
1250 	if (cp == NULL || !cpu_is_online(cp))
1251 		return;
1252 
1253 	mutex_enter(&mac_srs->srs_lock);
1254 	mac_srs->srs_state |= SRS_POLL_BOUND;
1255 	if (mac_srs->srs_poll_cpuid != -1)
1256 		clear = B_TRUE;
1257 	mac_srs->srs_poll_cpuid = cpuid;
1258 	mutex_exit(&mac_srs->srs_lock);
1259 
1260 	if (clear)
1261 		thread_affinity_clear(mac_srs->srs_poll_thr);
1262 
1263 	thread_affinity_set(mac_srs->srs_poll_thr, cpuid);
1264 	DTRACE_PROBE1(poll__CPU, processorid_t, cpuid);
1265 }
1266 
1267 /*
1268  * Re-target interrupt to the passed CPU. If re-target is successful,
1269  * set mc_rx_intr_cpu to the re-targeted CPU. Otherwise set it to -1.
1270  */
1271 void
1272 mac_rx_srs_retarget_intr(mac_soft_ring_set_t *mac_srs, processorid_t cpuid)
1273 {
1274 	cpu_t *cp;
1275 	mac_ring_t *ring = mac_srs->srs_ring;
1276 	mac_intr_t *mintr = &ring->mr_info.mri_intr;
1277 	flow_entry_t *flent = mac_srs->srs_flent;
1278 	boolean_t primary = mac_is_primary_client(mac_srs->srs_mcip);
1279 
1280 	ASSERT(MUTEX_HELD(&cpu_lock));
1281 
1282 	/*
1283 	 * Don't re-target the interrupt for these cases:
1284 	 * 1) ring is NULL
1285 	 * 2) the interrupt is shared (mi_ddi_shared)
1286 	 * 3) ddi_handle is NULL and !primary
1287 	 * 4) primary, ddi_handle is NULL but fe_rx_srs_cnt > 2
1288 	 * Case 3 & 4 are because of mac_client_intr_cpu() routine.
1289 	 * This routine will re-target fixed interrupt for primary
1290 	 * mac client if the client has only one ring. In that
1291 	 * case, mc_rx_intr_cpu will already have the correct value.
1292 	 */
1293 	if (ring == NULL || mintr->mi_ddi_shared || cpuid == -1 ||
1294 	    (mintr->mi_ddi_handle == NULL && !primary) || (primary &&
1295 	    mintr->mi_ddi_handle == NULL && flent->fe_rx_srs_cnt > 2)) {
1296 		mac_srs->srs_cpu.mc_rx_intr_cpu = -1;
1297 		return;
1298 	}
1299 
1300 	if (mintr->mi_ddi_handle == NULL)
1301 		return;
1302 
1303 	cp = cpu_get(cpuid);
1304 	if (cp == NULL || !cpu_is_online(cp))
1305 		return;
1306 
1307 	/* Drop the cpu_lock as set_intr_affinity() holds it */
1308 	mutex_exit(&cpu_lock);
1309 	if (set_intr_affinity(mintr->mi_ddi_handle, cpuid) == DDI_SUCCESS)
1310 		mac_srs->srs_cpu.mc_rx_intr_cpu = cpuid;
1311 	else
1312 		mac_srs->srs_cpu.mc_rx_intr_cpu = -1;
1313 	mutex_enter(&cpu_lock);
1314 }
1315 
1316 /*
1317  * Re-target Tx interrupts
1318  */
1319 void
1320 mac_tx_srs_retarget_intr(mac_soft_ring_set_t *mac_srs)
1321 {
1322 	cpu_t *cp;
1323 	mac_ring_t *ring;
1324 	mac_intr_t *mintr;
1325 	mac_soft_ring_t *sringp;
1326 	mac_srs_tx_t *srs_tx;
1327 	mac_cpus_t *srs_cpu;
1328 	processorid_t cpuid;
1329 	int i;
1330 
1331 	ASSERT(MUTEX_HELD(&cpu_lock));
1332 
1333 	srs_cpu = &mac_srs->srs_cpu;
1334 	if (MAC_TX_SOFT_RINGS(mac_srs)) {
1335 		for (i = 0; i < mac_srs->srs_tx_ring_count; i++) {
1336 			sringp = mac_srs->srs_tx_soft_rings[i];
1337 			ring = (mac_ring_t *)sringp->s_ring_tx_arg2;
1338 			cpuid = srs_cpu->mc_tx_intr_cpu[i];
1339 			cp = cpu_get(cpuid);
1340 			if (cp == NULL || !cpu_is_online(cp) ||
1341 			    !MAC_RING_RETARGETABLE(ring)) {
1342 				srs_cpu->mc_tx_retargeted_cpu[i] = -1;
1343 				continue;
1344 			}
1345 			mintr = &ring->mr_info.mri_intr;
1346 			/*
1347 			 * Drop the cpu_lock as set_intr_affinity()
1348 			 * holds it
1349 			 */
1350 			mutex_exit(&cpu_lock);
1351 			if (set_intr_affinity(mintr->mi_ddi_handle,
1352 			    cpuid) == DDI_SUCCESS) {
1353 				srs_cpu->mc_tx_retargeted_cpu[i] = cpuid;
1354 			} else {
1355 				srs_cpu->mc_tx_retargeted_cpu[i] = -1;
1356 			}
1357 			mutex_enter(&cpu_lock);
1358 		}
1359 	} else {
1360 		cpuid = srs_cpu->mc_tx_intr_cpu[0];
1361 		cp = cpu_get(cpuid);
1362 		if (cp == NULL || !cpu_is_online(cp)) {
1363 			srs_cpu->mc_tx_retargeted_cpu[0] = -1;
1364 			return;
1365 		}
1366 		srs_tx = &mac_srs->srs_tx;
1367 		ring = (mac_ring_t *)srs_tx->st_arg2;
1368 		if (MAC_RING_RETARGETABLE(ring)) {
1369 			mintr = &ring->mr_info.mri_intr;
1370 			mutex_exit(&cpu_lock);
1371 			if ((set_intr_affinity(mintr->mi_ddi_handle,
1372 			    cpuid) == DDI_SUCCESS)) {
1373 				srs_cpu->mc_tx_retargeted_cpu[0] = cpuid;
1374 			} else {
1375 				srs_cpu->mc_tx_retargeted_cpu[0] = -1;
1376 			}
1377 			mutex_enter(&cpu_lock);
1378 		}
1379 	}
1380 }
1381 
1382 /*
1383  * When a CPU comes back online, bind the MAC kernel threads which
1384  * were previously bound to that CPU, and had to be unbound because
1385  * the CPU was going away.
1386  *
1387  * These functions are called with cpu_lock held and hence we can't
1388  * cv_wait to grab the mac perimeter. Since these functions walk the soft
1389  * ring list of an SRS without being in the perimeter, the list itself
1390  * is protected by the SRS lock.
1391  */
1392 static void
1393 mac_walk_srs_and_bind(int cpuid)
1394 {
1395 	mac_soft_ring_set_t *mac_srs;
1396 	mac_soft_ring_t *soft_ring;
1397 
1398 	rw_enter(&mac_srs_g_lock, RW_READER);
1399 
1400 	if ((mac_srs = mac_srs_g_list) == NULL)
1401 		goto done;
1402 
1403 	for (; mac_srs != NULL; mac_srs = mac_srs->srs_next) {
1404 		if (mac_srs->srs_worker_cpuid == -1 &&
1405 		    mac_srs->srs_worker_cpuid_save == cpuid) {
1406 			mac_srs->srs_worker_cpuid_save = -1;
1407 			mac_srs_worker_bind(mac_srs, cpuid);
1408 		}
1409 
1410 		if (!(mac_srs->srs_type & SRST_TX)) {
1411 			if (mac_srs->srs_poll_cpuid == -1 &&
1412 			    mac_srs->srs_poll_cpuid_save == cpuid) {
1413 				mac_srs->srs_poll_cpuid_save = -1;
1414 				mac_srs_poll_bind(mac_srs, cpuid);
1415 			}
1416 		}
1417 
1418 		/* Next tackle the soft rings associated with the srs */
1419 		mutex_enter(&mac_srs->srs_lock);
1420 		for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
1421 		    soft_ring = soft_ring->s_ring_next) {
1422 			if (soft_ring->s_ring_cpuid == -1 &&
1423 			    soft_ring->s_ring_cpuid_save == cpuid) {
1424 				soft_ring->s_ring_cpuid_save = -1;
1425 				(void) mac_soft_ring_bind(soft_ring, cpuid);
1426 			}
1427 		}
1428 		mutex_exit(&mac_srs->srs_lock);
1429 	}
1430 done:
1431 	rw_exit(&mac_srs_g_lock);
1432 }
1433 
1434 /*
1435  * Change the priority of the SRS's poll and worker thread. Additionally,
1436  * update the priority of the worker threads for the SRS's soft rings.
1437  * Need to modify any associated squeue threads.
1438  */
1439 void
1440 mac_update_srs_priority(mac_soft_ring_set_t *mac_srs, pri_t prival)
1441 {
1442 	mac_soft_ring_t		*ringp;
1443 
1444 	mac_srs->srs_pri = prival;
1445 	thread_lock(mac_srs->srs_worker);
1446 	(void) thread_change_pri(mac_srs->srs_worker, mac_srs->srs_pri, 0);
1447 	thread_unlock(mac_srs->srs_worker);
1448 	if (mac_srs->srs_poll_thr != NULL) {
1449 		thread_lock(mac_srs->srs_poll_thr);
1450 		(void) thread_change_pri(mac_srs->srs_poll_thr,
1451 		    mac_srs->srs_pri, 0);
1452 		thread_unlock(mac_srs->srs_poll_thr);
1453 	}
1454 	if ((ringp = mac_srs->srs_soft_ring_head) == NULL)
1455 		return;
1456 	while (ringp != mac_srs->srs_soft_ring_tail) {
1457 		thread_lock(ringp->s_ring_worker);
1458 		(void) thread_change_pri(ringp->s_ring_worker,
1459 		    mac_srs->srs_pri, 0);
1460 		thread_unlock(ringp->s_ring_worker);
1461 		ringp = ringp->s_ring_next;
1462 	}
1463 	ASSERT(ringp == mac_srs->srs_soft_ring_tail);
1464 	thread_lock(ringp->s_ring_worker);
1465 	(void) thread_change_pri(ringp->s_ring_worker, mac_srs->srs_pri, 0);
1466 	thread_unlock(ringp->s_ring_worker);
1467 }
1468 
1469 /*
1470  * Change the receive bandwidth limit.
1471  */
1472 static void
1473 mac_rx_srs_update_bwlimit(mac_soft_ring_set_t *srs, mac_resource_props_t *mrp)
1474 {
1475 	mac_soft_ring_t		*softring;
1476 
1477 	mutex_enter(&srs->srs_lock);
1478 	mutex_enter(&srs->srs_bw->mac_bw_lock);
1479 
1480 	if (mrp->mrp_maxbw == MRP_MAXBW_RESETVAL) {
1481 		/* Reset bandwidth limit */
1482 		if (srs->srs_type & SRST_BW_CONTROL) {
1483 			softring = srs->srs_soft_ring_head;
1484 			while (softring != NULL) {
1485 				softring->s_ring_type &= ~ST_RING_BW_CTL;
1486 				softring = softring->s_ring_next;
1487 			}
1488 			srs->srs_type &= ~SRST_BW_CONTROL;
1489 			srs->srs_drain_func = mac_rx_srs_drain;
1490 		}
1491 	} else {
1492 		/* Set/Modify bandwidth limit */
1493 		srs->srs_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1494 		/*
1495 		 * Give twice the queuing capability before
1496 		 * dropping packets. The unit is bytes/tick.
1497 		 */
1498 		srs->srs_bw->mac_bw_drop_threshold =
1499 		    srs->srs_bw->mac_bw_limit << 1;
1500 		if (!(srs->srs_type & SRST_BW_CONTROL)) {
1501 			softring = srs->srs_soft_ring_head;
1502 			while (softring != NULL) {
1503 				softring->s_ring_type |= ST_RING_BW_CTL;
1504 				softring = softring->s_ring_next;
1505 			}
1506 			srs->srs_type |= SRST_BW_CONTROL;
1507 			srs->srs_drain_func = mac_rx_srs_drain_bw;
1508 		}
1509 	}
1510 done:
1511 	mutex_exit(&srs->srs_bw->mac_bw_lock);
1512 	mutex_exit(&srs->srs_lock);
1513 }
1514 
1515 /* Change the transmit bandwidth limit */
1516 static void
1517 mac_tx_srs_update_bwlimit(mac_soft_ring_set_t *srs, mac_resource_props_t *mrp)
1518 {
1519 	uint32_t		tx_mode, ring_info = 0;
1520 	mac_srs_tx_t		*srs_tx = &srs->srs_tx;
1521 	mac_client_impl_t	*mcip = srs->srs_mcip;
1522 
1523 	/*
1524 	 * We need to quiesce/restart the client here because mac_tx() and
1525 	 * srs->srs_tx->st_func do not hold srs->srs_lock while accessing
1526 	 * st_mode and related fields, which are modified by the code below.
1527 	 */
1528 	mac_tx_client_quiesce((mac_client_handle_t)mcip);
1529 
1530 	mutex_enter(&srs->srs_lock);
1531 	mutex_enter(&srs->srs_bw->mac_bw_lock);
1532 
1533 	tx_mode = srs_tx->st_mode;
1534 	if (mrp->mrp_maxbw == MRP_MAXBW_RESETVAL) {
1535 		/* Reset bandwidth limit */
1536 		if (tx_mode == SRS_TX_BW) {
1537 			if (srs_tx->st_arg2 != NULL)
1538 				ring_info = mac_hwring_getinfo(srs_tx->st_arg2);
1539 			if (mac_tx_serialize ||
1540 			    (ring_info & MAC_RING_TX_SERIALIZE)) {
1541 				srs_tx->st_mode = SRS_TX_SERIALIZE;
1542 			} else {
1543 				srs_tx->st_mode = SRS_TX_DEFAULT;
1544 			}
1545 		} else if (tx_mode == SRS_TX_BW_FANOUT) {
1546 			srs_tx->st_mode = SRS_TX_FANOUT;
1547 		} else if (tx_mode == SRS_TX_BW_AGGR) {
1548 			srs_tx->st_mode = SRS_TX_AGGR;
1549 		}
1550 		srs->srs_type &= ~SRST_BW_CONTROL;
1551 	} else {
1552 		/* Set/Modify bandwidth limit */
1553 		srs->srs_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
1554 		/*
1555 		 * Give twice the queuing capability before
1556 		 * dropping packets. The unit is bytes/tick.
1557 		 */
1558 		srs->srs_bw->mac_bw_drop_threshold =
1559 		    srs->srs_bw->mac_bw_limit << 1;
1560 		srs->srs_type |= SRST_BW_CONTROL;
1561 		if (tx_mode != SRS_TX_BW && tx_mode != SRS_TX_BW_FANOUT &&
1562 		    tx_mode != SRS_TX_BW_AGGR) {
1563 			if (tx_mode == SRS_TX_SERIALIZE ||
1564 			    tx_mode == SRS_TX_DEFAULT) {
1565 				srs_tx->st_mode = SRS_TX_BW;
1566 			} else if (tx_mode == SRS_TX_FANOUT) {
1567 				srs_tx->st_mode = SRS_TX_BW_FANOUT;
1568 			} else if (tx_mode == SRS_TX_AGGR) {
1569 				srs_tx->st_mode = SRS_TX_BW_AGGR;
1570 			} else {
1571 				ASSERT(0);
1572 			}
1573 		}
1574 	}
1575 done:
1576 	srs_tx->st_func = mac_tx_get_func(srs_tx->st_mode);
1577 	mutex_exit(&srs->srs_bw->mac_bw_lock);
1578 	mutex_exit(&srs->srs_lock);
1579 
1580 	mac_tx_client_restart((mac_client_handle_t)mcip);
1581 }
1582 
1583 /*
1584  * The uber function that deals with any update to bandwidth limits.
1585  */
1586 void
1587 mac_srs_update_bwlimit(flow_entry_t *flent, mac_resource_props_t *mrp)
1588 {
1589 	int			count;
1590 
1591 	for (count = 0; count < flent->fe_rx_srs_cnt; count++)
1592 		mac_rx_srs_update_bwlimit(flent->fe_rx_srs[count], mrp);
1593 	mac_tx_srs_update_bwlimit(flent->fe_tx_srs, mrp);
1594 }
1595 
1596 /*
1597  * When the first sub-flow is added to a link, we disable polling on the
1598  * link and also modify the entry point to mac_rx_srs_subflow_process.
1599  * (polling is disabled because with the subflow added, accounting
1600  * for polling needs additional logic, it is assumed that when a subflow is
1601  * added, we can take some hit as a result of disabling polling rather than
1602  * adding more complexity - if this becomes a perf. issue we need to
1603  * re-rvaluate this logic).  When the last subflow is removed, we turn back
1604  * polling and also reset the entry point to mac_rx_srs_process.
1605  *
1606  * In the future if there are multiple SRS, we can simply
1607  * take one and give it to the flow rather than disabling polling and
1608  * resetting the entry point.
1609  */
1610 void
1611 mac_client_update_classifier(mac_client_impl_t *mcip, boolean_t enable)
1612 {
1613 	flow_entry_t		*flent = mcip->mci_flent;
1614 	int			i;
1615 	mac_impl_t		*mip = mcip->mci_mip;
1616 	mac_rx_func_t		rx_func;
1617 	uint_t			rx_srs_cnt;
1618 	boolean_t		enable_classifier;
1619 
1620 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
1621 
1622 	enable_classifier = !FLOW_TAB_EMPTY(mcip->mci_subflow_tab) && enable;
1623 
1624 	rx_func = enable_classifier ? mac_rx_srs_subflow_process :
1625 	    mac_rx_srs_process;
1626 
1627 	/* Tell mac_srs_poll_state_change to disable polling if necessary */
1628 	if (mip->mi_state_flags & MIS_POLL_DISABLE)
1629 		enable_classifier = B_TRUE;
1630 
1631 	/*
1632 	 * If receive function has already been configured correctly for
1633 	 * current subflow configuration, do nothing.
1634 	 */
1635 	if (flent->fe_cb_fn == (flow_fn_t)rx_func)
1636 		return;
1637 
1638 	rx_srs_cnt = flent->fe_rx_srs_cnt;
1639 	for (i = 0; i < rx_srs_cnt; i++) {
1640 		ASSERT(flent->fe_rx_srs[i] != NULL);
1641 		mac_srs_poll_state_change(flent->fe_rx_srs[i],
1642 		    enable_classifier, rx_func);
1643 	}
1644 
1645 	/*
1646 	 * Change the S/W classifier so that we can land in the
1647 	 * correct processing function with correct argument.
1648 	 * If all subflows have been removed we can revert to
1649 	 * mac_rx_srsprocess, else we need mac_rx_srs_subflow_process.
1650 	 */
1651 	mutex_enter(&flent->fe_lock);
1652 	flent->fe_cb_fn = (flow_fn_t)rx_func;
1653 	flent->fe_cb_arg1 = (void *)mip;
1654 	flent->fe_cb_arg2 = flent->fe_rx_srs[0];
1655 	mutex_exit(&flent->fe_lock);
1656 }
1657 
1658 static void
1659 mac_srs_update_fanout_list(mac_soft_ring_set_t *mac_srs)
1660 {
1661 	int tcp_count = 0, udp_count = 0, oth_count = 0, tx_count = 0;
1662 	mac_soft_ring_t *softring;
1663 
1664 	softring = mac_srs->srs_soft_ring_head;
1665 	if (softring == NULL) {
1666 		ASSERT(mac_srs->srs_soft_ring_count == 0);
1667 		mac_srs->srs_tcp_ring_count = 0;
1668 		mac_srs->srs_udp_ring_count = 0;
1669 		mac_srs->srs_oth_ring_count = 0;
1670 		mac_srs->srs_tx_ring_count = 0;
1671 		return;
1672 	}
1673 
1674 	while (softring != NULL) {
1675 		if (softring->s_ring_type & ST_RING_TCP) {
1676 			mac_srs->srs_tcp_soft_rings[tcp_count++] = softring;
1677 		} else if (softring->s_ring_type & ST_RING_UDP) {
1678 			mac_srs->srs_udp_soft_rings[udp_count++] = softring;
1679 		} else if (softring->s_ring_type & ST_RING_OTH) {
1680 			mac_srs->srs_oth_soft_rings[oth_count++] = softring;
1681 		} else {
1682 			ASSERT(softring->s_ring_type & ST_RING_TX);
1683 			mac_srs->srs_tx_soft_rings[tx_count++] = softring;
1684 		}
1685 		softring = softring->s_ring_next;
1686 	}
1687 
1688 	ASSERT(mac_srs->srs_soft_ring_count ==
1689 	    (tcp_count + udp_count + oth_count + tx_count));
1690 	mac_srs->srs_tcp_ring_count = tcp_count;
1691 	mac_srs->srs_udp_ring_count = udp_count;
1692 	mac_srs->srs_oth_ring_count = oth_count;
1693 	mac_srs->srs_tx_ring_count = tx_count;
1694 }
1695 
1696 void
1697 mac_srs_create_proto_softrings(int id, uint16_t type, pri_t pri,
1698     mac_client_impl_t *mcip, mac_soft_ring_set_t *mac_srs,
1699     processorid_t cpuid, mac_direct_rx_t rx_func, void *x_arg1,
1700     mac_resource_handle_t x_arg2, boolean_t set_bypass)
1701 {
1702 	mac_soft_ring_t	*softring;
1703 	mac_rx_fifo_t	mrf;
1704 
1705 	bzero(&mrf, sizeof (mac_rx_fifo_t));
1706 	mrf.mrf_type = MAC_RX_FIFO;
1707 	mrf.mrf_receive = (mac_receive_t)mac_soft_ring_poll;
1708 	mrf.mrf_intr_enable = (mac_intr_enable_t)mac_soft_ring_intr_enable;
1709 	mrf.mrf_intr_disable = (mac_intr_disable_t)mac_soft_ring_intr_disable;
1710 	mrf.mrf_flow_priority = pri;
1711 
1712 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1713 	    (type|ST_RING_TCP), pri, mcip, mac_srs,
1714 	    cpuid, rx_func, x_arg1, x_arg2);
1715 	softring->s_ring_rx_arg2 = NULL;
1716 
1717 	/*
1718 	 * TCP and UDP support DLS bypass. In addition TCP
1719 	 * squeue can also poll their corresponding soft rings.
1720 	 */
1721 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1722 		mac_soft_ring_dls_bypass(softring,
1723 		    mcip->mci_direct_rx_fn,
1724 		    mcip->mci_direct_rx_arg);
1725 
1726 		mrf.mrf_rx_arg = softring;
1727 		mrf.mrf_intr_handle = (mac_intr_handle_t)softring;
1728 
1729 		/*
1730 		 * Make a call in IP to get a TCP squeue assigned to
1731 		 * this softring to maintain full CPU locality through
1732 		 * the stack and allow the squeue to be able to poll
1733 		 * the softring so the flow control can be pushed
1734 		 * all the way to H/W.
1735 		 */
1736 		softring->s_ring_rx_arg2 =
1737 		    mcip->mci_resource_add((void *)mcip->mci_resource_arg,
1738 		    (mac_resource_t *)&mrf);
1739 	}
1740 
1741 	/*
1742 	 * Non-TCP protocols don't support squeues. Hence we
1743 	 * don't make any ring addition callbacks for non-TCP
1744 	 * rings. Now create the UDP softring and allow it to
1745 	 * bypass the DLS layer.
1746 	 */
1747 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1748 	    (type|ST_RING_UDP), pri, mcip, mac_srs,
1749 	    cpuid, rx_func, x_arg1, x_arg2);
1750 	softring->s_ring_rx_arg2 = NULL;
1751 
1752 	if (set_bypass && (mcip->mci_resource_arg != NULL)) {
1753 		mac_soft_ring_dls_bypass(softring,
1754 		    mcip->mci_direct_rx_fn,
1755 		    mcip->mci_direct_rx_arg);
1756 	}
1757 
1758 	/* Create the Oth softrings which has to go through the DLS */
1759 	softring = mac_soft_ring_create(id, mac_soft_ring_worker_wait,
1760 	    (type|ST_RING_OTH), pri, mcip, mac_srs,
1761 	    cpuid, rx_func, x_arg1, x_arg2);
1762 	softring->s_ring_rx_arg2 = NULL;
1763 }
1764 
1765 /*
1766  * This routine associates a CPU or a set of CPU to process incoming
1767  * traffic from a mac client. If multiple CPUs are specified, then
1768  * so many soft rings are created with each soft ring worker thread
1769  * bound to a CPU in the set. Each soft ring in turn will be
1770  * associated with an squeue and the squeue will be moved to the
1771  * same CPU as that of the soft ring's.
1772  */
1773 static void
1774 mac_srs_fanout_modify(mac_client_impl_t *mcip, mac_direct_rx_t rx_func,
1775     void *x_arg1, mac_resource_handle_t x_arg2,
1776     mac_soft_ring_set_t *mac_rx_srs, mac_soft_ring_set_t *mac_tx_srs)
1777 {
1778 	mac_soft_ring_t *softring;
1779 	uint32_t soft_ring_flag = 0;
1780 	processorid_t cpuid = -1;
1781 	int i, srings_present, new_fanout_cnt;
1782 	mac_cpus_t *srs_cpu;
1783 
1784 	/* fanout state is REINIT. Set it back to INIT */
1785 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_REINIT);
1786 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1787 
1788 	/* how many are present right now */
1789 	srings_present = mac_rx_srs->srs_tcp_ring_count;
1790 	/* new request */
1791 	srs_cpu = &mac_rx_srs->srs_cpu;
1792 	new_fanout_cnt = srs_cpu->mc_rx_fanout_cnt;
1793 
1794 	mutex_enter(&mac_rx_srs->srs_lock);
1795 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1796 		soft_ring_flag |= ST_RING_BW_CTL;
1797 	mutex_exit(&mac_rx_srs->srs_lock);
1798 
1799 	if (new_fanout_cnt > srings_present) {
1800 		/* soft rings increased */
1801 		mutex_enter(&mac_rx_srs->srs_lock);
1802 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1803 		mutex_exit(&mac_rx_srs->srs_lock);
1804 
1805 		for (i = mac_rx_srs->srs_tcp_ring_count;
1806 		    i < new_fanout_cnt; i++) {
1807 			/*
1808 			 * Create the protocol softrings and set the
1809 			 * DLS bypass where possible.
1810 			 */
1811 			mac_srs_create_proto_softrings(i, soft_ring_flag,
1812 			    mac_rx_srs->srs_pri, mcip, mac_rx_srs, cpuid,
1813 			    rx_func, x_arg1, x_arg2, B_TRUE);
1814 		}
1815 		mac_srs_update_fanout_list(mac_rx_srs);
1816 	} else if (new_fanout_cnt < srings_present) {
1817 		/* soft rings decreased */
1818 		if (new_fanout_cnt == 1) {
1819 			mutex_enter(&mac_rx_srs->srs_lock);
1820 			mac_rx_srs->srs_type &= ~SRST_FANOUT_SRC_IP;
1821 			ASSERT(mac_rx_srs->srs_type & SRST_FANOUT_PROTO);
1822 			mutex_exit(&mac_rx_srs->srs_lock);
1823 		}
1824 		/* Get rid of extra soft rings */
1825 		for (i = new_fanout_cnt;
1826 		    i < mac_rx_srs->srs_tcp_ring_count; i++) {
1827 			softring = mac_rx_srs->srs_tcp_soft_rings[i];
1828 			if (softring->s_ring_rx_arg2 != NULL) {
1829 				mcip->mci_resource_remove(
1830 				    (void *)mcip->mci_resource_arg,
1831 				    softring->s_ring_rx_arg2);
1832 			}
1833 			mac_soft_ring_remove(mac_rx_srs,
1834 			    mac_rx_srs->srs_tcp_soft_rings[i]);
1835 			mac_soft_ring_remove(mac_rx_srs,
1836 			    mac_rx_srs->srs_udp_soft_rings[i]);
1837 			mac_soft_ring_remove(mac_rx_srs,
1838 			    mac_rx_srs->srs_oth_soft_rings[i]);
1839 		}
1840 		mac_srs_update_fanout_list(mac_rx_srs);
1841 	}
1842 
1843 	ASSERT(new_fanout_cnt == mac_rx_srs->srs_tcp_ring_count);
1844 	mutex_enter(&cpu_lock);
1845 	for (i = 0; i < mac_rx_srs->srs_tcp_ring_count; i++) {
1846 		cpuid = srs_cpu->mc_rx_fanout_cpus[i];
1847 		(void) mac_soft_ring_bind(mac_rx_srs->srs_udp_soft_rings[i],
1848 		    cpuid);
1849 		(void) mac_soft_ring_bind(mac_rx_srs->srs_oth_soft_rings[i],
1850 		    cpuid);
1851 		(void) mac_soft_ring_bind(mac_rx_srs->srs_tcp_soft_rings[i],
1852 		    cpuid);
1853 		softring = mac_rx_srs->srs_tcp_soft_rings[i];
1854 		if (softring->s_ring_rx_arg2 != NULL) {
1855 			mcip->mci_resource_bind((void *)mcip->mci_resource_arg,
1856 			    softring->s_ring_rx_arg2, cpuid);
1857 		}
1858 	}
1859 
1860 	mac_srs_worker_bind(mac_rx_srs, srs_cpu->mc_rx_workerid);
1861 	mac_srs_poll_bind(mac_rx_srs, srs_cpu->mc_rx_pollid);
1862 	mac_rx_srs_retarget_intr(mac_rx_srs, srs_cpu->mc_rx_intr_cpu);
1863 	/*
1864 	 * Bind Tx srs and soft ring threads too. Let's bind tx
1865 	 * srs to the last cpu in mrp list.
1866 	 */
1867 	if (mac_tx_srs != NULL) {
1868 		BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1869 		mac_tx_srs_retarget_intr(mac_tx_srs);
1870 	}
1871 	mutex_exit(&cpu_lock);
1872 }
1873 
1874 /*
1875  * Bind SRS threads and soft rings to CPUs/create fanout list.
1876  */
1877 void
1878 mac_srs_fanout_init(mac_client_impl_t *mcip, mac_resource_props_t *mrp,
1879     mac_direct_rx_t rx_func, void *x_arg1, mac_resource_handle_t x_arg2,
1880     mac_soft_ring_set_t *mac_rx_srs, mac_soft_ring_set_t *mac_tx_srs,
1881     cpupart_t *cpupart)
1882 {
1883 	int		i;
1884 	processorid_t	cpuid;
1885 	uint32_t	soft_ring_flag = 0;
1886 	int soft_ring_cnt;
1887 	mac_cpus_t *srs_cpu = &mac_rx_srs->srs_cpu;
1888 
1889 	/*
1890 	 * Remove the no soft ring flag and we will adjust it
1891 	 * appropriately further down.
1892 	 */
1893 	mutex_enter(&mac_rx_srs->srs_lock);
1894 	mac_rx_srs->srs_type &= ~SRST_NO_SOFT_RINGS;
1895 	mutex_exit(&mac_rx_srs->srs_lock);
1896 
1897 	ASSERT(mac_rx_srs->srs_soft_ring_head == NULL);
1898 
1899 	if (mac_rx_srs->srs_type & SRST_BW_CONTROL)
1900 		soft_ring_flag |= ST_RING_BW_CTL;
1901 
1902 	ASSERT(mac_rx_srs->srs_fanout_state == SRS_FANOUT_UNINIT);
1903 	mac_rx_srs->srs_fanout_state = SRS_FANOUT_INIT;
1904 	/*
1905 	 * Ring count can be 0 if no fanout is required and no cpu
1906 	 * were specified. Leave the SRS worker and poll thread
1907 	 * unbound
1908 	 */
1909 	ASSERT(mrp != NULL);
1910 	soft_ring_cnt = srs_cpu->mc_rx_fanout_cnt;
1911 
1912 	/* Step 1: bind cpu contains cpu list where threads need to bind */
1913 	if (soft_ring_cnt > 0) {
1914 		mutex_enter(&cpu_lock);
1915 		for (i = 0; i < soft_ring_cnt; i++) {
1916 			cpuid = srs_cpu->mc_rx_fanout_cpus[i];
1917 			/* Create the protocol softrings */
1918 			mac_srs_create_proto_softrings(i, soft_ring_flag,
1919 			    mac_rx_srs->srs_pri, mcip, mac_rx_srs, cpuid,
1920 			    rx_func, x_arg1, x_arg2, B_FALSE);
1921 		}
1922 		mac_srs_worker_bind(mac_rx_srs, srs_cpu->mc_rx_workerid);
1923 		mac_srs_poll_bind(mac_rx_srs, srs_cpu->mc_rx_pollid);
1924 		mac_rx_srs_retarget_intr(mac_rx_srs, srs_cpu->mc_rx_intr_cpu);
1925 		/*
1926 		 * Bind Tx srs and soft ring threads too.
1927 		 * Let's bind tx srs to the last cpu in
1928 		 * mrp list.
1929 		 */
1930 		if (mac_tx_srs == NULL) {
1931 			mutex_exit(&cpu_lock);
1932 			goto alldone;
1933 		}
1934 
1935 		BIND_TX_SRS_AND_SOFT_RINGS(mac_tx_srs, mrp);
1936 		mac_tx_srs_retarget_intr(mac_tx_srs);
1937 		mutex_exit(&cpu_lock);
1938 	} else {
1939 		mutex_enter(&cpu_lock);
1940 		/*
1941 		 * For a subflow, mrp_workerid and mrp_pollid
1942 		 * is not set.
1943 		 */
1944 		mac_srs_worker_bind(mac_rx_srs, mrp->mrp_rx_workerid);
1945 		mac_srs_poll_bind(mac_rx_srs, mrp->mrp_rx_pollid);
1946 		mutex_exit(&cpu_lock);
1947 		goto no_softrings;
1948 	}
1949 
1950 alldone:
1951 	if (soft_ring_cnt > 1)
1952 		mac_rx_srs->srs_type |= SRST_FANOUT_SRC_IP;
1953 	mac_srs_update_fanout_list(mac_rx_srs);
1954 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1955 	return;
1956 
1957 no_softrings:
1958 	if (mac_rx_srs->srs_type & SRST_FANOUT_PROTO) {
1959 		mutex_enter(&cpu_lock);
1960 		cpuid = mac_next_bind_cpu(cpupart);
1961 		/* Create the protocol softrings */
1962 		mac_srs_create_proto_softrings(0, soft_ring_flag,
1963 		    mac_rx_srs->srs_pri, mcip, mac_rx_srs, cpuid,
1964 		    rx_func, x_arg1, x_arg2, B_FALSE);
1965 		mutex_exit(&cpu_lock);
1966 	} else {
1967 		/*
1968 		 * This is the case when there is no fanout which is
1969 		 * true for subflows.
1970 		 */
1971 		mac_rx_srs->srs_type |= SRST_NO_SOFT_RINGS;
1972 	}
1973 	mac_srs_update_fanout_list(mac_rx_srs);
1974 	mac_srs_client_poll_enable(mcip, mac_rx_srs);
1975 }
1976 
1977 /*
1978  * mac_fanout_setup:
1979  *
1980  * Calls mac_srs_fanout_init() or modify() depending upon whether
1981  * the SRS is getting initialized or re-initialized.
1982  */
1983 void
1984 mac_fanout_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
1985     mac_resource_props_t *mrp, mac_direct_rx_t rx_func, void *x_arg1,
1986     mac_resource_handle_t x_arg2, cpupart_t *cpupart)
1987 {
1988 	mac_soft_ring_set_t *mac_rx_srs, *mac_tx_srs;
1989 	int i, rx_srs_cnt;
1990 
1991 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
1992 	/*
1993 	 * This is an aggregation port. Fanout will be setup
1994 	 * over the aggregation itself.
1995 	 */
1996 	if (mcip->mci_state_flags & MCIS_EXCLUSIVE)
1997 		return;
1998 
1999 	mac_rx_srs = flent->fe_rx_srs[0];
2000 	/*
2001 	 * Set up the fanout on the tx side only once, with the
2002 	 * first rx SRS. The CPU binding, fanout, and bandwidth
2003 	 * criteria are common to both RX and TX, so
2004 	 * initializing them along side avoids redundant code.
2005 	 */
2006 	mac_tx_srs = flent->fe_tx_srs;
2007 	rx_srs_cnt = flent->fe_rx_srs_cnt;
2008 
2009 	/* No fanout for subflows */
2010 	if (flent->fe_type & FLOW_USER) {
2011 		mac_srs_fanout_init(mcip, mrp, rx_func,
2012 		    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs,
2013 		    cpupart);
2014 		return;
2015 	}
2016 
2017 	if (mrp->mrp_mask & MRP_CPUS_USERSPEC)
2018 		mac_flow_user_cpu_init(flent, mrp);
2019 	else
2020 		mac_flow_cpu_init(flent, cpupart);
2021 
2022 	mrp->mrp_rx_fanout_cnt = mac_rx_srs->srs_cpu.mc_rx_fanout_cnt;
2023 
2024 	/*
2025 	 * Set up fanout for both SW (0th SRS) and HW classified
2026 	 * SRS (the rest of Rx SRSs in flent).
2027 	 */
2028 	for (i = 0; i < rx_srs_cnt; i++) {
2029 		mac_rx_srs = flent->fe_rx_srs[i];
2030 		if (i != 0)
2031 			mac_tx_srs = NULL;
2032 		switch (mac_rx_srs->srs_fanout_state) {
2033 		case SRS_FANOUT_UNINIT:
2034 			mac_srs_fanout_init(mcip, mrp, rx_func,
2035 			    x_arg1, x_arg2, mac_rx_srs, mac_tx_srs,
2036 			    cpupart);
2037 			break;
2038 		case SRS_FANOUT_INIT:
2039 			break;
2040 		case SRS_FANOUT_REINIT:
2041 			mac_rx_srs_quiesce(mac_rx_srs, SRS_QUIESCE);
2042 			mac_srs_fanout_modify(mcip, rx_func, x_arg1,
2043 			    x_arg2, mac_rx_srs, mac_tx_srs);
2044 			mac_rx_srs_restart(mac_rx_srs);
2045 			break;
2046 		default:
2047 			VERIFY(mac_rx_srs->srs_fanout_state <=
2048 			    SRS_FANOUT_REINIT);
2049 			break;
2050 		}
2051 	}
2052 }
2053 
2054 /*
2055  * mac_srs_create:
2056  *
2057  * Create a mac_soft_ring_set_t (SRS). If soft_ring_fanout_type is
2058  * SRST_TX, an SRS for Tx side is created. Otherwise an SRS for Rx side
2059  * processing is created.
2060  *
2061  * Details on Rx SRS:
2062  * Create a SRS and also add the necessary soft rings for TCP and
2063  * non-TCP based on fanout type and count specified.
2064  *
2065  * mac_soft_ring_fanout, mac_srs_fanout_modify (?),
2066  * mac_soft_ring_stop_workers, mac_soft_ring_set_destroy, etc need
2067  * to be heavily modified.
2068  *
2069  * mi_soft_ring_list_size, mi_soft_ring_size, etc need to disappear.
2070  */
2071 mac_soft_ring_set_t *
2072 mac_srs_create(mac_client_impl_t *mcip, flow_entry_t *flent, uint32_t srs_type,
2073     mac_direct_rx_t rx_func, void *x_arg1, mac_resource_handle_t x_arg2,
2074     mac_ring_t *ring)
2075 {
2076 	mac_soft_ring_set_t	*mac_srs;
2077 	mac_srs_rx_t		*srs_rx;
2078 	mac_srs_tx_t		*srs_tx;
2079 	mac_bw_ctl_t		*mac_bw;
2080 	mac_resource_props_t	*mrp;
2081 	boolean_t		is_tx_srs = ((srs_type & SRST_TX) != 0);
2082 
2083 	mac_srs = kmem_cache_alloc(mac_srs_cache, KM_SLEEP);
2084 	bzero(mac_srs, sizeof (mac_soft_ring_set_t));
2085 	srs_rx = &mac_srs->srs_rx;
2086 	srs_tx = &mac_srs->srs_tx;
2087 
2088 	mutex_enter(&flent->fe_lock);
2089 
2090 	/*
2091 	 * Get the bandwidth control structure from the flent. Get
2092 	 * rid of any residual values in the control structure for
2093 	 * the tx bw struct and also for the rx, if the rx srs is
2094 	 * the 1st one being brought up (the rx bw ctl struct may
2095 	 * be shared by multiple SRSs)
2096 	 */
2097 	if (is_tx_srs) {
2098 		mac_srs->srs_bw = &flent->fe_tx_bw;
2099 		bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
2100 		flent->fe_tx_srs = mac_srs;
2101 	} else {
2102 		/*
2103 		 * The bw counter (stored in the flent) is shared
2104 		 * by SRS's within an rx group.
2105 		 */
2106 		mac_srs->srs_bw = &flent->fe_rx_bw;
2107 		/* First rx SRS, clear the bw structure */
2108 		if (flent->fe_rx_srs_cnt == 0)
2109 			bzero(mac_srs->srs_bw, sizeof (mac_bw_ctl_t));
2110 
2111 		/*
2112 		 * It is better to panic here rather than just assert because
2113 		 * on a non-debug kernel we might end up courrupting memory
2114 		 * and making it difficult to debug.
2115 		 */
2116 		if (flent->fe_rx_srs_cnt >= MAX_RINGS_PER_GROUP) {
2117 			panic("Array Overrun detected due to MAC client %p "
2118 			    " having more rings than %d", (void *)mcip,
2119 			    MAX_RINGS_PER_GROUP);
2120 		}
2121 		flent->fe_rx_srs[flent->fe_rx_srs_cnt] = mac_srs;
2122 		flent->fe_rx_srs_cnt++;
2123 	}
2124 	mac_srs->srs_flent = flent;
2125 	mutex_exit(&flent->fe_lock);
2126 
2127 	mac_srs->srs_state = 0;
2128 	mac_srs->srs_type = (srs_type | SRST_NO_SOFT_RINGS);
2129 	mac_srs->srs_worker_cpuid = mac_srs->srs_worker_cpuid_save = -1;
2130 	mac_srs->srs_poll_cpuid = mac_srs->srs_poll_cpuid_save = -1;
2131 	mac_srs->srs_mcip = mcip;
2132 	mac_srs_fanout_list_alloc(mac_srs);
2133 
2134 	/*
2135 	 * For a flow we use the underlying MAC client's priority range with
2136 	 * the priority value to find an absolute priority value. For a MAC
2137 	 * client we use the MAC client's maximum priority as the value.
2138 	 */
2139 	mrp = &flent->fe_effective_props;
2140 	if ((mac_srs->srs_type & SRST_FLOW) != 0) {
2141 		mac_srs->srs_pri = FLOW_PRIORITY(mcip->mci_min_pri,
2142 		    mcip->mci_max_pri, mrp->mrp_priority);
2143 	} else {
2144 		mac_srs->srs_pri = mcip->mci_max_pri;
2145 	}
2146 	/*
2147 	 * We need to insert the SRS in the global list before
2148 	 * binding the SRS and SR threads. Otherwise there is a
2149 	 * is a small window where the cpu reconfig callbacks
2150 	 * may miss the SRS in the list walk and DR could fail
2151 	 * as there are bound threads.
2152 	 */
2153 	mac_srs_add_glist(mac_srs);
2154 
2155 	/* Initialize bw limit */
2156 	if ((mrp->mrp_mask & MRP_MAXBW) != 0) {
2157 		mac_srs->srs_drain_func = mac_rx_srs_drain_bw;
2158 
2159 		mac_bw = mac_srs->srs_bw;
2160 		mutex_enter(&mac_bw->mac_bw_lock);
2161 		mac_bw->mac_bw_limit = FLOW_BYTES_PER_TICK(mrp->mrp_maxbw);
2162 
2163 		/*
2164 		 * Give twice the queuing capability before
2165 		 * dropping packets. The unit is bytes/tick.
2166 		 */
2167 		mac_bw->mac_bw_drop_threshold = mac_bw->mac_bw_limit << 1;
2168 		mutex_exit(&mac_bw->mac_bw_lock);
2169 		mac_srs->srs_type |= SRST_BW_CONTROL;
2170 	} else {
2171 		mac_srs->srs_drain_func = mac_rx_srs_drain;
2172 	}
2173 
2174 	/*
2175 	 * We use the following policy to control Receive
2176 	 * Side Dynamic Polling:
2177 	 * 1) We switch to poll mode anytime the processing thread causes
2178 	 *    a backlog to build up in SRS and its associated Soft Rings
2179 	 *    (sr_poll_pkt_cnt > 0).
2180 	 * 2) As long as the backlog stays under the low water mark
2181 	 *    (sr_lowat), we poll the H/W for more packets.
2182 	 * 3) If the backlog (sr_poll_pkt_cnt) exceeds low water mark, we
2183 	 *    stay in poll mode but don't poll the H/W for more packets.
2184 	 * 4) Anytime in polling mode, if we poll the H/W for packets and
2185 	 *    find nothing plus we have an existing backlog
2186 	 *    (sr_poll_pkt_cnt > 0), we stay in polling mode but don't poll
2187 	 *    the H/W for packets anymore (let the polling thread go to sleep).
2188 	 * 5) Once the backlog is relived (packets are processed) we reenable
2189 	 *    polling (by signalling the poll thread) only when the backlog
2190 	 *    dips below sr_poll_thres.
2191 	 * 6) sr_hiwat is used exclusively when we are not polling capable
2192 	 *    and is used to decide when to drop packets so the SRS queue
2193 	 *    length doesn't grow infinitely.
2194 	 */
2195 	if (!is_tx_srs) {
2196 		srs_rx->sr_hiwat = mac_soft_ring_max_q_cnt;
2197 		/* Low water mark needs to be less than high water mark */
2198 		srs_rx->sr_lowat = mac_soft_ring_min_q_cnt <=
2199 		    mac_soft_ring_max_q_cnt ? mac_soft_ring_min_q_cnt :
2200 		    (mac_soft_ring_max_q_cnt >> 2);
2201 		/* Poll threshold need to be half of low water mark or less */
2202 		srs_rx->sr_poll_thres = mac_soft_ring_poll_thres <=
2203 		    (srs_rx->sr_lowat >> 1) ? mac_soft_ring_poll_thres :
2204 		    (srs_rx->sr_lowat >> 1);
2205 		if (mac_latency_optimize)
2206 			mac_srs->srs_state |= SRS_LATENCY_OPT;
2207 		else
2208 			mac_srs->srs_state |= SRS_SOFTRING_QUEUE;
2209 	}
2210 
2211 	mac_srs->srs_worker = thread_create(NULL, 0,
2212 	    mac_srs_worker, mac_srs, 0, &p0, TS_RUN, mac_srs->srs_pri);
2213 
2214 	if (is_tx_srs) {
2215 		/* Handle everything about Tx SRS and return */
2216 		mac_srs->srs_drain_func = mac_tx_srs_drain;
2217 		srs_tx->st_max_q_cnt = mac_tx_srs_max_q_cnt;
2218 		srs_tx->st_hiwat =
2219 		    (mac_tx_srs_hiwat > mac_tx_srs_max_q_cnt) ?
2220 		    mac_tx_srs_max_q_cnt : mac_tx_srs_hiwat;
2221 		srs_tx->st_arg1 = x_arg1;
2222 		srs_tx->st_arg2 = x_arg2;
2223 		goto done;
2224 	}
2225 
2226 	if ((srs_type & SRST_FLOW) != 0 ||
2227 	    FLOW_TAB_EMPTY(mcip->mci_subflow_tab))
2228 		srs_rx->sr_lower_proc = mac_rx_srs_process;
2229 	else
2230 		srs_rx->sr_lower_proc = mac_rx_srs_subflow_process;
2231 
2232 	srs_rx->sr_func = rx_func;
2233 	srs_rx->sr_arg1 = x_arg1;
2234 	srs_rx->sr_arg2 = x_arg2;
2235 
2236 	if (ring != NULL) {
2237 		uint_t ring_info;
2238 
2239 		/* Is the mac_srs created over the RX default group? */
2240 		if (ring->mr_gh == (mac_group_handle_t)
2241 		    MAC_DEFAULT_RX_GROUP(mcip->mci_mip)) {
2242 			mac_srs->srs_type |= SRST_DEFAULT_GRP;
2243 		}
2244 		mac_srs->srs_ring = ring;
2245 		ring->mr_srs = mac_srs;
2246 		ring->mr_classify_type = MAC_HW_CLASSIFIER;
2247 		ring->mr_flag |= MR_INCIPIENT;
2248 
2249 		if (!(mcip->mci_mip->mi_state_flags & MIS_POLL_DISABLE) &&
2250 		    FLOW_TAB_EMPTY(mcip->mci_subflow_tab) && mac_poll_enable)
2251 			mac_srs->srs_state |= SRS_POLLING_CAPAB;
2252 
2253 		mac_srs->srs_poll_thr = thread_create(NULL, 0,
2254 		    mac_rx_srs_poll_ring, mac_srs, 0, &p0, TS_RUN,
2255 		    mac_srs->srs_pri);
2256 		/*
2257 		 * Some drivers require serialization and don't send
2258 		 * packet chains in interrupt context. For such
2259 		 * drivers, we should always queue in soft ring
2260 		 * so that we get a chance to switch into a polling
2261 		 * mode under backlog.
2262 		 */
2263 		ring_info = mac_hwring_getinfo((mac_ring_handle_t)ring);
2264 		if (ring_info & MAC_RING_RX_ENQUEUE)
2265 			mac_srs->srs_state |= SRS_SOFTRING_QUEUE;
2266 	}
2267 done:
2268 	mac_srs_stat_create(mac_srs);
2269 	return (mac_srs);
2270 }
2271 
2272 /*
2273  * Figure out the number of soft rings required. Its dependant on
2274  * if protocol fanout is required (for LINKs), global settings
2275  * require us to do fanout for performance (based on mac_soft_ring_enable),
2276  * or user has specifically requested fanout.
2277  */
2278 static uint32_t
2279 mac_find_fanout(flow_entry_t *flent, uint32_t link_type)
2280 {
2281 	uint32_t			fanout_type;
2282 	mac_resource_props_t		*mrp = &flent->fe_effective_props;
2283 
2284 	/* no fanout for subflows */
2285 	switch (link_type) {
2286 	case SRST_FLOW:
2287 		fanout_type = SRST_NO_SOFT_RINGS;
2288 		break;
2289 	case SRST_LINK:
2290 		fanout_type = SRST_FANOUT_PROTO;
2291 		break;
2292 	}
2293 
2294 	/* A primary NIC/link is being plumbed */
2295 	if (flent->fe_type & FLOW_PRIMARY_MAC) {
2296 		if (mac_soft_ring_enable && mac_rx_soft_ring_count > 1) {
2297 			fanout_type |= SRST_FANOUT_SRC_IP;
2298 		}
2299 	} else if (flent->fe_type & FLOW_VNIC) {
2300 		/* A VNIC is being created */
2301 		if (mrp != NULL && mrp->mrp_ncpus > 0) {
2302 			fanout_type |= SRST_FANOUT_SRC_IP;
2303 		}
2304 	}
2305 
2306 	return (fanout_type);
2307 }
2308 
2309 /*
2310  * Change a group from h/w to s/w classification.
2311  */
2312 void
2313 mac_rx_switch_grp_to_sw(mac_group_t *group)
2314 {
2315 	mac_ring_t		*ring;
2316 	mac_soft_ring_set_t	*mac_srs;
2317 
2318 	for (ring = group->mrg_rings; ring != NULL; ring = ring->mr_next) {
2319 		if (ring->mr_classify_type == MAC_HW_CLASSIFIER) {
2320 			/*
2321 			 * Remove the SRS associated with the HW ring.
2322 			 * As a result, polling will be disabled.
2323 			 */
2324 			mac_srs = ring->mr_srs;
2325 			ASSERT(mac_srs != NULL);
2326 			mac_rx_srs_remove(mac_srs);
2327 			ring->mr_srs = NULL;
2328 		}
2329 
2330 		if (ring->mr_state != MR_INUSE)
2331 			(void) mac_start_ring(ring);
2332 
2333 		/*
2334 		 * We need to perform SW classification
2335 		 * for packets landing in these rings
2336 		 */
2337 		ring->mr_flag = 0;
2338 		ring->mr_classify_type = MAC_SW_CLASSIFIER;
2339 	}
2340 }
2341 
2342 /*
2343  * Create the Rx SRS for S/W classifier and for each ring in the
2344  * group (if exclusive group). Also create the Tx SRS.
2345  */
2346 void
2347 mac_srs_group_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2348     uint32_t link_type)
2349 {
2350 	cpupart_t		*cpupart;
2351 	mac_resource_props_t	*mrp = MCIP_RESOURCE_PROPS(mcip);
2352 	mac_resource_props_t	*emrp = MCIP_EFFECTIVE_PROPS(mcip);
2353 	boolean_t		use_default = B_FALSE;
2354 
2355 	mac_rx_srs_group_setup(mcip, flent, link_type);
2356 	mac_tx_srs_group_setup(mcip, flent, link_type);
2357 
2358 	pool_lock();
2359 	cpupart = mac_pset_find(mrp, &use_default);
2360 	mac_fanout_setup(mcip, flent, MCIP_RESOURCE_PROPS(mcip),
2361 	    mac_rx_deliver, mcip, NULL, cpupart);
2362 	mac_set_pool_effective(use_default, cpupart, mrp, emrp);
2363 	pool_unlock();
2364 }
2365 
2366 /*
2367  * Set up the RX SRSs. If the S/W SRS is not set, set  it up, if there
2368  * is a group associated with this MAC client, set up SRSs for individual
2369  * h/w rings.
2370  */
2371 void
2372 mac_rx_srs_group_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2373     uint32_t link_type)
2374 {
2375 	mac_impl_t		*mip = mcip->mci_mip;
2376 	mac_soft_ring_set_t	*mac_srs;
2377 	mac_ring_t		*ring;
2378 	uint32_t		fanout_type;
2379 	mac_group_t		*rx_group = flent->fe_rx_ring_group;
2380 
2381 	fanout_type = mac_find_fanout(flent, link_type);
2382 
2383 	/* Create the SRS for S/W classification if none exists */
2384 	if (flent->fe_rx_srs[0] == NULL) {
2385 		ASSERT(flent->fe_rx_srs_cnt == 0);
2386 		/* Setup the Rx SRS */
2387 		mac_srs = mac_srs_create(mcip, flent, fanout_type | link_type,
2388 		    mac_rx_deliver, mcip, NULL, NULL);
2389 		mutex_enter(&flent->fe_lock);
2390 		flent->fe_cb_fn = (flow_fn_t)mac_srs->srs_rx.sr_lower_proc;
2391 		flent->fe_cb_arg1 = (void *)mip;
2392 		flent->fe_cb_arg2 = (void *)mac_srs;
2393 		mutex_exit(&flent->fe_lock);
2394 	}
2395 
2396 	if (rx_group == NULL)
2397 		return;
2398 	/*
2399 	 * fanout for default SRS is done when default SRS are created
2400 	 * above. As each ring is added to the group, we setup the
2401 	 * SRS and fanout to it.
2402 	 */
2403 	switch (rx_group->mrg_state) {
2404 	case MAC_GROUP_STATE_RESERVED:
2405 		for (ring = rx_group->mrg_rings; ring != NULL;
2406 		    ring = ring->mr_next) {
2407 			switch (ring->mr_state) {
2408 			case MR_INUSE:
2409 			case MR_FREE:
2410 				if (ring->mr_srs != NULL)
2411 					break;
2412 				if (ring->mr_state != MR_INUSE)
2413 					(void) mac_start_ring(ring);
2414 
2415 				/*
2416 				 * Since the group is exclusively ours create
2417 				 * an SRS for this ring to allow the
2418 				 * individual SRS to dynamically poll the
2419 				 * ring. Do this only if the  client is not
2420 				 * a VLAN MAC client, since for VLAN we do
2421 				 * s/w classification for the VID check, and
2422 				 * if it has a unicast address.
2423 				 */
2424 				if ((mcip->mci_state_flags &
2425 				    MCIS_NO_UNICAST_ADDR) ||
2426 				    i_mac_flow_vid(mcip->mci_flent) !=
2427 				    VLAN_ID_NONE) {
2428 					break;
2429 				}
2430 				mac_srs = mac_srs_create(mcip, flent,
2431 				    fanout_type | link_type,
2432 				    mac_rx_deliver, mcip, NULL, ring);
2433 				break;
2434 			default:
2435 				cmn_err(CE_PANIC,
2436 				    "srs_setup: mcip = %p "
2437 				    "trying to add UNKNOWN ring = %p\n",
2438 				    (void *)mcip, (void *)ring);
2439 				break;
2440 			}
2441 		}
2442 		break;
2443 	case MAC_GROUP_STATE_SHARED:
2444 		/*
2445 		 * Set all rings of this group to software classified.
2446 		 *
2447 		 * If the group is current RESERVED, the existing mac
2448 		 * client (the only client on this group) is using
2449 		 * this group exclusively.  In that case we need to
2450 		 * disable polling on the rings of the group (if it
2451 		 * was enabled), and free the SRS associated with the
2452 		 * rings.
2453 		 */
2454 		mac_rx_switch_grp_to_sw(rx_group);
2455 		break;
2456 	default:
2457 		ASSERT(B_FALSE);
2458 		break;
2459 	}
2460 }
2461 
2462 /*
2463  * Set up the TX SRS.
2464  */
2465 void
2466 mac_tx_srs_group_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2467     uint32_t link_type)
2468 {
2469 	int			cnt;
2470 	int			ringcnt;
2471 	mac_ring_t		*ring;
2472 	mac_group_t		*grp;
2473 
2474 	/*
2475 	 * If we are opened exclusively (like aggr does for aggr_ports),
2476 	 * don't set up Tx SRS and Tx soft rings as they won't be used.
2477 	 * The same thing has to be done for Rx side also. See bug:
2478 	 * 6880080
2479 	 */
2480 	if (mcip->mci_state_flags & MCIS_EXCLUSIVE) {
2481 		/*
2482 		 * If we have rings, start them here.
2483 		 */
2484 		if (flent->fe_tx_ring_group == NULL)
2485 			return;
2486 		grp = (mac_group_t *)flent->fe_tx_ring_group;
2487 		ringcnt = grp->mrg_cur_count;
2488 		ring = grp->mrg_rings;
2489 		for (cnt = 0; cnt < ringcnt; cnt++) {
2490 			if (ring->mr_state != MR_INUSE) {
2491 				(void) mac_start_ring(ring);
2492 			}
2493 			ring = ring->mr_next;
2494 		}
2495 		return;
2496 	}
2497 	if (flent->fe_tx_srs == NULL) {
2498 		(void) mac_srs_create(mcip, flent, SRST_TX | link_type,
2499 		    NULL, mcip, NULL, NULL);
2500 	}
2501 	mac_tx_srs_setup(mcip, flent);
2502 }
2503 
2504 /*
2505  * Remove all the RX SRSs. If we want to remove only the SRSs associated
2506  * with h/w rings, leave the S/W SRS alone. This is used when we want to
2507  * move the MAC client from one group to another, so we need to teardown
2508  * on the h/w SRSs.
2509  */
2510 void
2511 mac_rx_srs_group_teardown(flow_entry_t *flent, boolean_t hwonly)
2512 {
2513 	mac_soft_ring_set_t	*mac_srs;
2514 	int			i;
2515 	int			count = flent->fe_rx_srs_cnt;
2516 
2517 	for (i = 0; i < count; i++) {
2518 		if (i == 0 && hwonly)
2519 			continue;
2520 		mac_srs = flent->fe_rx_srs[i];
2521 		mac_rx_srs_quiesce(mac_srs, SRS_CONDEMNED);
2522 		mac_srs_free(mac_srs);
2523 		flent->fe_rx_srs[i] = NULL;
2524 		flent->fe_rx_srs_cnt--;
2525 	}
2526 	ASSERT(!hwonly || flent->fe_rx_srs_cnt == 1);
2527 	ASSERT(hwonly || flent->fe_rx_srs_cnt == 0);
2528 }
2529 
2530 /*
2531  * Remove the TX SRS.
2532  */
2533 void
2534 mac_tx_srs_group_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
2535     uint32_t link_type)
2536 {
2537 	mac_soft_ring_set_t	*tx_srs;
2538 	mac_srs_tx_t		*tx;
2539 
2540 	if ((tx_srs = flent->fe_tx_srs) == NULL)
2541 		return;
2542 
2543 	tx = &tx_srs->srs_tx;
2544 	switch (link_type) {
2545 	case SRST_FLOW:
2546 		/*
2547 		 * For flows, we need to work with passed
2548 		 * flent to find the Rx/Tx SRS.
2549 		 */
2550 		mac_tx_srs_quiesce(tx_srs, SRS_CONDEMNED);
2551 		break;
2552 	case SRST_LINK:
2553 		mac_tx_client_condemn((mac_client_handle_t)mcip);
2554 		if (tx->st_arg2 != NULL) {
2555 			ASSERT(tx_srs->srs_type & SRST_TX);
2556 			/*
2557 			 * The ring itself will be stopped when
2558 			 * we release the group or in the
2559 			 * mac_datapath_teardown (for the default
2560 			 * group)
2561 			 */
2562 			tx->st_arg2 = NULL;
2563 		}
2564 		break;
2565 	default:
2566 		ASSERT(B_FALSE);
2567 		break;
2568 	}
2569 	mac_srs_free(tx_srs);
2570 	flent->fe_tx_srs = NULL;
2571 }
2572 
2573 /*
2574  * This is the group state machine.
2575  *
2576  * The state of an Rx group is given by
2577  * the following table. The default group and its rings are started in
2578  * mac_start itself and the default group stays in SHARED state until
2579  * mac_stop at which time the group and rings are stopped and and it
2580  * reverts to the Registered state.
2581  *
2582  * Typically this function is called on a group after adding or removing a
2583  * client from it, to find out what should be the new state of the group.
2584  * If the new state is RESERVED, then the client that owns this group
2585  * exclusively is also returned. Note that adding or removing a client from
2586  * a group could also impact the default group and the caller needs to
2587  * evaluate the effect on the default group.
2588  *
2589  * Group type		# of clients	mi_nactiveclients	Group State
2590  *			in the group
2591  *
2592  * Non-default		0		N.A.			REGISTERED
2593  * Non-default		1		N.A.			RESERVED
2594  *
2595  * Default		0		N.A.			SHARED
2596  * Default		1		1			RESERVED
2597  * Default		1		> 1			SHARED
2598  * Default		> 1		N.A.			SHARED
2599  *
2600  * For a TX group, the following is the state table.
2601  *
2602  * Group type		# of clients	Group State
2603  *			in the group
2604  *
2605  * Non-default		0		REGISTERED
2606  * Non-default		1		RESERVED
2607  *
2608  * Default		0		REGISTERED
2609  * Default		1		RESERVED
2610  * Default		> 1		SHARED
2611  */
2612 mac_group_state_t
2613 mac_group_next_state(mac_group_t *grp, mac_client_impl_t **group_only_mcip,
2614     mac_group_t *defgrp, boolean_t rx_group)
2615 {
2616 	mac_impl_t		*mip = (mac_impl_t *)grp->mrg_mh;
2617 
2618 	*group_only_mcip = NULL;
2619 
2620 	/* Non-default group */
2621 
2622 	if (grp != defgrp) {
2623 		if (MAC_GROUP_NO_CLIENT(grp))
2624 			return (MAC_GROUP_STATE_REGISTERED);
2625 
2626 		*group_only_mcip = MAC_GROUP_ONLY_CLIENT(grp);
2627 		if (*group_only_mcip != NULL)
2628 			return (MAC_GROUP_STATE_RESERVED);
2629 
2630 		return (MAC_GROUP_STATE_SHARED);
2631 	}
2632 
2633 	/* Default group */
2634 
2635 	if (MAC_GROUP_NO_CLIENT(grp)) {
2636 		if (rx_group)
2637 			return (MAC_GROUP_STATE_SHARED);
2638 		else
2639 			return (MAC_GROUP_STATE_REGISTERED);
2640 	}
2641 	*group_only_mcip = MAC_GROUP_ONLY_CLIENT(grp);
2642 	if (*group_only_mcip == NULL)
2643 		return (MAC_GROUP_STATE_SHARED);
2644 
2645 	if (rx_group && mip->mi_nactiveclients != 1)
2646 		return (MAC_GROUP_STATE_SHARED);
2647 
2648 	ASSERT(*group_only_mcip != NULL);
2649 	return (MAC_GROUP_STATE_RESERVED);
2650 }
2651 
2652 /*
2653  * OVERVIEW NOTES FOR DATAPATH
2654  * ===========================
2655  *
2656  * Create an SRS and setup the corresponding flow function and args.
2657  * Add a classification rule for the flow specified by 'flent' and program
2658  * the hardware classifier when applicable.
2659  *
2660  * Rx ring assignment, SRS, polling and B/W enforcement
2661  * ----------------------------------------------------
2662  *
2663  * We try to use H/W classification on NIC and assign traffic to a
2664  * MAC address to a particular Rx ring. There is a 1-1 mapping
2665  * between a SRS and a Rx ring. The SRS (short for soft ring set)
2666  * dynamically switches the underlying Rx ring between interrupt
2667  * and polling mode and enforces any specified B/W control.
2668  *
2669  * There is always a SRS created and tied to each H/W and S/W rule.
2670  * Whenever we create a H/W rule, we always add the the same rule to
2671  * S/W classifier and tie a SRS to it.
2672  *
2673  * In case a B/W control is specified, its broken into bytes
2674  * per ticks and as soon as the quota for a tick is exhausted,
2675  * the underlying Rx ring is forced into poll mode for remianing
2676  * tick. The SRS poll thread only polls for bytes that are
2677  * allowed to come in the SRS. We typically let 4x the configured
2678  * B/W worth of packets to come in the SRS (to prevent unnecessary
2679  * drops due to bursts) but only process the specified amount.
2680  *
2681  * A Link (primary NIC, VNIC, VLAN or aggr) can have 1 or more
2682  * Rx rings (and corresponding SRSs) assigned to it. The SRS
2683  * in turn can have softrings to do protocol level fanout or
2684  * softrings to do S/W based fanout or both. In case the NIC
2685  * has no Rx rings, we do S/W classification to respective SRS.
2686  * The S/W classification rule is always setup and ready. This
2687  * allows the MAC layer to reassign Rx rings whenever needed
2688  * but packets still continue to flow via the default path and
2689  * getting S/W classified to correct SRS.
2690  *
2691  * In other cases where a NIC or VNIC is plumbed, our goal is use
2692  * H/W classifier and get two Rx ring assigned for the Link. One
2693  * for TCP and one for UDP|SCTP. The respective SRS still do the
2694  * polling on the Rx ring. For Link that is plumbed for IP, there
2695  * is a TCP squeue which also does polling and can control the
2696  * the Rx ring directly (where SRS is just pass through). For
2697  * the following cases, the SRS does the polling underneath.
2698  * 1) non IP based Links (Links which are not plumbed via ifconfig)
2699  *    and paths which have no IP squeues (UDP & SCTP)
2700  * 2) If B/W control is specified on the Link
2701  * 3) If S/W fanout is secified
2702  *
2703  * Note1: As of current implementation, we try to assign only 1 Rx
2704  * ring per Link and more than 1 Rx ring for primary Link for
2705  * H/W based fanout. We always create following softrings per SRS:
2706  * 1) TCP softring which is polled by TCP squeue where possible
2707  *    (and also bypasses DLS)
2708  * 2) UDP/SCTP based which bypasses DLS
2709  * 3) OTH softring which goes via DLS (currently deal with IPv6
2710  *    and non TCP/UDP/SCTP for IPv4 packets).
2711  *
2712  * It is necessary to create 3 softrings since SRS has to poll
2713  * the single Rx ring underneath and enforce any link level B/W
2714  * control (we can't switch the Rx ring in poll mode just based
2715  * on TCP squeue if the same Rx ring is sharing UDP and other
2716  * traffic as well). Once polling is done and any Link level B/W
2717  * control is specified, the packets are assigned to respective
2718  * softring based on protocol. Since TCP has IP based squeue
2719  * which benefits by polling, we separate TCP packets into
2720  * its own softring which can be polled by IP squeue. We need
2721  * to separate out UDP/SCTP to UDP softring since it can bypass
2722  * the DLS layer which has heavy performance advanatges and we
2723  * need a softring (OTH) for rest.
2724  *
2725  * ToDo: The 3 softrings for protocol are needed only till we can
2726  * get rid of DLS from datapath, make IPv4 and IPv6 paths
2727  * symmetric (deal with mac_header_info for v6 and polling for
2728  * IPv4 TCP - ip_accept_tcp is IPv4 specific although squeues
2729  * are generic), and bring SAP based classification to MAC layer
2730  *
2731  * H/W and S/W based fanout and multiple Rx rings per Link
2732  * -------------------------------------------------------
2733  *
2734  * In case, fanout is requested (or determined automatically based
2735  * on Link speed and processor speed), we try to assign multiple
2736  * Rx rings per Link with their respective SRS. In this case
2737  * the NIC should be capable of fanning out incoming packets between
2738  * the assigned Rx rings (H/W based fanout). All the SRS
2739  * individually switch their Rx ring between interrupt and polling
2740  * mode but share a common B/W control counter in case of Link
2741  * level B/W is specified.
2742  *
2743  * If S/W based fanout is specified in lieu of H/W based fanout,
2744  * the Link SRS creates the specified number of softrings for
2745  * each protocol (TCP, UDP, OTH). Incoming packets are fanned
2746  * out to the correct softring based on their protocol and
2747  * protocol specific hash function.
2748  *
2749  * Primary and non primary MAC clients
2750  * -----------------------------------
2751  *
2752  * The NICs, VNICs, Vlans, and Aggrs are typically termed as Links
2753  * and are a Layer 2 construct.
2754  *
2755  * Primary NIC:
2756  *	The Link that owns the primary MAC address and typically
2757  *	is used as the data NIC in non virtualized cases. As such
2758  *	H/W resources are preferntially given to primary NIC. As
2759  *	far as code is concerned, there is no difference in the
2760  *	primary NIC vs VNICs. They are all treated as Links.
2761  *	At the very first call to mac_unicast_add() we program the S/W
2762  *	classifier for the primary MAC address, get a soft ring set
2763  *	(and soft rings based on 'ip_soft_ring_cnt')
2764  *	and a Rx ring assigned for polling to get enabled.
2765  *	When IP get plumbed and negotiates polling, we can
2766  *	let squeue do the polling on TCP softring.
2767  *
2768  * VNICs:
2769  *	Same as any other Link. As long as the H/W resource assignments
2770  *	are equal, the data path and setup for all Links is same.
2771  *
2772  * Flows:
2773  *	Can be configured on Links. They have their own SRS and the
2774  *	S/W classifier is programmed appropriately based on the flow.
2775  *	The flows typically deal with layer 3 and above and
2776  *	creates a soft ring set specific to the flow. The receive
2777  *	side function is switched from mac_rx_srs_process to
2778  *	mac_rx_srs_subflow_process which first tries to assign the
2779  *	packet to appropriate flow SRS and failing which assigns it
2780  *	to link SRS. This allows us to avoid the layered approach
2781  *	which gets complex.
2782  *
2783  * By the time mac_datapath_setup() completes, we already have the
2784  * soft rings set, Rx rings, soft rings, etc figured out and both H/W
2785  * and S/W classifiers programmed. IP is not plumbed yet (and might
2786  * never be for Virtual Machines guest OS path). When IP is plumbed
2787  * (for both NIC and VNIC), we do a capability negotiation for polling
2788  * and upcall functions etc.
2789  *
2790  * Rx ring Assignement NOTES
2791  * -------------------------
2792  *
2793  * For NICs which have only 1 Rx ring (we treat  NICs with no Rx rings
2794  * as NIC with a single default ring), we assign the only ring to
2795  * primary Link. The primary Link SRS can do polling on it as long as
2796  * it is the only link in use and we compare the MAC address for unicast
2797  * packets before accepting an incoming packet (there is no need for S/W
2798  * classification in this case). We disable polling on the only ring the
2799  * moment 2nd link gets created (the polling remains enabled even though
2800  * there are broadcast and * multicast flows created).
2801  *
2802  * If the NIC has more than 1 Rx ring, we assign the default ring (the
2803  * 1st ring) to deal with broadcast, multicast and traffic for other
2804  * NICs which needs S/W classification. We assign the primary mac
2805  * addresses to another ring by specifiying a classification rule for
2806  * primary unicast MAC address to the selected ring. The primary Link
2807  * (and its SRS) can continue to poll the assigned Rx ring at all times
2808  * independantly.
2809  *
2810  * Note: In future, if no fanout is specified, we try to assign 2 Rx
2811  * rings for the primary Link with the primary MAC address + TCP going
2812  * to one ring and primary MAC address + UDP|SCTP going to other ring.
2813  * Any remaining traffic for primary MAC address can go to the default
2814  * Rx ring and get S/W classified. This way the respective SRSs don't
2815  * need to do proto fanout and don't need to have softrings at all and
2816  * can poll their respective Rx rings.
2817  *
2818  * As an optimization, when a new NIC or VNIC is created, we can get
2819  * only one Rx ring and make it a TCP specific Rx ring and use the
2820  * H/W default Rx ring for the rest (this Rx ring is never polled).
2821  *
2822  * For clients that don't have MAC address, but want to receive and
2823  * transmit packets (e.g, bpf, gvrp etc.), we need to setup the datapath.
2824  * For such clients (identified by the MCIS_NO_UNICAST_ADDR flag) we
2825  * always give the default group and use software classification (i.e.
2826  * even if this is the only client in the default group, we will
2827  * leave group as shared).
2828  */
2829 int
2830 mac_datapath_setup(mac_client_impl_t *mcip, flow_entry_t *flent,
2831     uint32_t link_type)
2832 {
2833 	mac_impl_t		*mip = mcip->mci_mip;
2834 	mac_group_t		*rgroup = NULL;
2835 	mac_group_t		*tgroup = NULL;
2836 	mac_group_t		*default_rgroup;
2837 	mac_group_t		*default_tgroup;
2838 	int			err;
2839 	uint8_t			*mac_addr;
2840 	mac_group_state_t	next_state;
2841 	mac_client_impl_t	*group_only_mcip;
2842 	mac_resource_props_t	*mrp = MCIP_RESOURCE_PROPS(mcip);
2843 	mac_resource_props_t	*emrp = MCIP_EFFECTIVE_PROPS(mcip);
2844 	boolean_t		rxhw;
2845 	boolean_t		txhw;
2846 	boolean_t		use_default = B_FALSE;
2847 	cpupart_t		*cpupart;
2848 	boolean_t		no_unicast;
2849 	boolean_t		isprimary = flent->fe_type & FLOW_PRIMARY_MAC;
2850 	mac_client_impl_t	*reloc_pmcip = NULL;
2851 
2852 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
2853 
2854 	switch (link_type) {
2855 	case SRST_FLOW:
2856 		mac_srs_group_setup(mcip, flent, link_type);
2857 		return (0);
2858 
2859 	case SRST_LINK:
2860 		no_unicast = mcip->mci_state_flags & MCIS_NO_UNICAST_ADDR;
2861 		mac_addr = flent->fe_flow_desc.fd_dst_mac;
2862 
2863 		/* Default RX group */
2864 		default_rgroup = MAC_DEFAULT_RX_GROUP(mip);
2865 
2866 		/* Default TX group */
2867 		default_tgroup = MAC_DEFAULT_TX_GROUP(mip);
2868 
2869 		if (no_unicast) {
2870 			rgroup = default_rgroup;
2871 			tgroup = default_tgroup;
2872 			goto grp_found;
2873 		}
2874 		rxhw = (mrp->mrp_mask & MRP_RX_RINGS) &&
2875 		    (mrp->mrp_nrxrings > 0 ||
2876 		    (mrp->mrp_mask & MRP_RXRINGS_UNSPEC));
2877 		txhw = (mrp->mrp_mask & MRP_TX_RINGS) &&
2878 		    (mrp->mrp_ntxrings > 0 ||
2879 		    (mrp->mrp_mask & MRP_TXRINGS_UNSPEC));
2880 
2881 		/*
2882 		 * By default we have given the primary all the rings
2883 		 * i.e. the default group. Let's see if the primary
2884 		 * needs to be relocated so that the addition of this
2885 		 * client doesn't impact the primary's performance,
2886 		 * i.e. if the primary is in the default group and
2887 		 * we add this client, the primary will lose polling.
2888 		 * We do this only for NICs supporting dynamic ring
2889 		 * grouping and only when this is the first client
2890 		 * after the primary (i.e. nactiveclients is 2)
2891 		 */
2892 		if (!isprimary && mip->mi_nactiveclients == 2 &&
2893 		    (group_only_mcip = mac_primary_client_handle(mip)) !=
2894 		    NULL && mip->mi_rx_group_type == MAC_GROUP_TYPE_DYNAMIC) {
2895 			reloc_pmcip = mac_check_primary_relocation(
2896 			    group_only_mcip, rxhw);
2897 		}
2898 		/*
2899 		 * Check to see if we can get an exclusive group for
2900 		 * this mac address or if there already exists a
2901 		 * group that has this mac address (case of VLANs).
2902 		 * If no groups are available, use the default group.
2903 		 */
2904 		rgroup = mac_reserve_rx_group(mcip, mac_addr, B_FALSE);
2905 		if (rgroup == NULL && rxhw) {
2906 			err = ENOSPC;
2907 			goto setup_failed;
2908 		} else if (rgroup == NULL) {
2909 			rgroup = default_rgroup;
2910 		}
2911 		/*
2912 		 * Check to see if we can get an exclusive group for
2913 		 * this mac client. If no groups are available, use
2914 		 * the default group.
2915 		 */
2916 		tgroup = mac_reserve_tx_group(mcip, B_FALSE);
2917 		if (tgroup == NULL && txhw) {
2918 			if (rgroup != NULL && rgroup != default_rgroup)
2919 				mac_release_rx_group(mcip, rgroup);
2920 			err = ENOSPC;
2921 			goto setup_failed;
2922 		} else if (tgroup == NULL) {
2923 			tgroup = default_tgroup;
2924 		}
2925 
2926 		/*
2927 		 * Some NICs don't support any Rx rings, so there may not
2928 		 * even be a default group.
2929 		 */
2930 	grp_found:
2931 		if (rgroup != NULL) {
2932 			if (rgroup != default_rgroup &&
2933 			    MAC_GROUP_NO_CLIENT(rgroup) &&
2934 			    (rxhw || mcip->mci_share != 0)) {
2935 				MAC_RX_GRP_RESERVED(mip);
2936 				if (mip->mi_rx_group_type ==
2937 				    MAC_GROUP_TYPE_DYNAMIC) {
2938 					MAC_RX_RING_RESERVED(mip,
2939 					    rgroup->mrg_cur_count);
2940 				}
2941 			}
2942 			flent->fe_rx_ring_group = rgroup;
2943 			/*
2944 			 * Add the client to the group. This could cause
2945 			 * either this group to move to the shared state or
2946 			 * cause the default group to move to the shared state.
2947 			 * The actions on this group are done here, while the
2948 			 * actions on the default group are postponed to
2949 			 * the end of this function.
2950 			 */
2951 			mac_group_add_client(rgroup, mcip);
2952 			next_state = mac_group_next_state(rgroup,
2953 			    &group_only_mcip, default_rgroup, B_TRUE);
2954 			mac_set_group_state(rgroup, next_state);
2955 		}
2956 
2957 		if (tgroup != NULL) {
2958 			if (tgroup != default_tgroup &&
2959 			    MAC_GROUP_NO_CLIENT(tgroup) &&
2960 			    (txhw || mcip->mci_share != 0)) {
2961 				MAC_TX_GRP_RESERVED(mip);
2962 				if (mip->mi_tx_group_type ==
2963 				    MAC_GROUP_TYPE_DYNAMIC) {
2964 					MAC_TX_RING_RESERVED(mip,
2965 					    tgroup->mrg_cur_count);
2966 				}
2967 			}
2968 			flent->fe_tx_ring_group = tgroup;
2969 			mac_group_add_client(tgroup, mcip);
2970 			next_state = mac_group_next_state(tgroup,
2971 			    &group_only_mcip, default_tgroup, B_FALSE);
2972 			tgroup->mrg_state = next_state;
2973 		}
2974 		/*
2975 		 * Setup the Rx and Tx SRSes. If we got a pristine group
2976 		 * exclusively above, mac_srs_group_setup would simply create
2977 		 * the required SRSes. If we ended up sharing a previously
2978 		 * reserved group, mac_srs_group_setup would also dismantle the
2979 		 * SRSes of the previously exclusive group
2980 		 */
2981 		mac_srs_group_setup(mcip, flent, link_type);
2982 
2983 		/* We are setting up minimal datapath only */
2984 		if (no_unicast)
2985 			break;
2986 		/* Program the S/W Classifer */
2987 		if ((err = mac_flow_add(mip->mi_flow_tab, flent)) != 0)
2988 			goto setup_failed;
2989 
2990 		/* Program the H/W Classifier */
2991 		if ((err = mac_add_macaddr(mip, rgroup, mac_addr,
2992 		    (mcip->mci_state_flags & MCIS_UNICAST_HW) != 0)) != 0)
2993 			goto setup_failed;
2994 		mcip->mci_unicast = mac_find_macaddr(mip, mac_addr);
2995 		ASSERT(mcip->mci_unicast != NULL);
2996 		/* (Re)init the v6 token & local addr used by link protection */
2997 		mac_protect_update_mac_token(mcip);
2998 		break;
2999 
3000 	default:
3001 		ASSERT(B_FALSE);
3002 		break;
3003 	}
3004 
3005 	/*
3006 	 * All broadcast and multicast traffic is received only on the default
3007 	 * group. If we have setup the datapath for a non-default group above
3008 	 * then move the default group to shared state to allow distribution of
3009 	 * incoming broadcast traffic to the other groups and dismantle the
3010 	 * SRSes over the default group.
3011 	 */
3012 	if (rgroup != NULL) {
3013 		if (rgroup != default_rgroup) {
3014 			if (default_rgroup->mrg_state ==
3015 			    MAC_GROUP_STATE_RESERVED) {
3016 				group_only_mcip = MAC_GROUP_ONLY_CLIENT(
3017 				    default_rgroup);
3018 				ASSERT(group_only_mcip != NULL &&
3019 				    mip->mi_nactiveclients > 1);
3020 
3021 				mac_set_group_state(default_rgroup,
3022 				    MAC_GROUP_STATE_SHARED);
3023 				mac_rx_srs_group_setup(group_only_mcip,
3024 				    group_only_mcip->mci_flent, SRST_LINK);
3025 				pool_lock();
3026 				cpupart = mac_pset_find(mrp, &use_default);
3027 				mac_fanout_setup(group_only_mcip,
3028 				    group_only_mcip->mci_flent,
3029 				    MCIP_RESOURCE_PROPS(group_only_mcip),
3030 				    mac_rx_deliver, group_only_mcip, NULL,
3031 				    cpupart);
3032 				mac_set_pool_effective(use_default, cpupart,
3033 				    mrp, emrp);
3034 				pool_unlock();
3035 			}
3036 			ASSERT(default_rgroup->mrg_state ==
3037 			    MAC_GROUP_STATE_SHARED);
3038 		}
3039 		/*
3040 		 * If we get an exclusive group for a VLAN MAC client we
3041 		 * need to take the s/w path to make the additional check for
3042 		 * the vid. Disable polling and set it to s/w classification.
3043 		 * Similarly for clients that don't have a unicast address.
3044 		 */
3045 		if (rgroup->mrg_state == MAC_GROUP_STATE_RESERVED &&
3046 		    (i_mac_flow_vid(flent) != VLAN_ID_NONE || no_unicast)) {
3047 			mac_rx_switch_grp_to_sw(rgroup);
3048 		}
3049 	}
3050 	mac_set_rings_effective(mcip);
3051 	return (0);
3052 
3053 setup_failed:
3054 	/* Switch the primary back to default group */
3055 	if (reloc_pmcip != NULL) {
3056 		(void) mac_rx_switch_group(reloc_pmcip,
3057 		    reloc_pmcip->mci_flent->fe_rx_ring_group, default_rgroup);
3058 	}
3059 	mac_datapath_teardown(mcip, flent, link_type);
3060 	return (err);
3061 }
3062 
3063 void
3064 mac_datapath_teardown(mac_client_impl_t *mcip, flow_entry_t *flent,
3065     uint32_t link_type)
3066 {
3067 	mac_impl_t		*mip = mcip->mci_mip;
3068 	mac_group_t		*group = NULL;
3069 	mac_client_impl_t	*grp_only_mcip;
3070 	flow_entry_t		*group_only_flent;
3071 	mac_group_t		*default_group;
3072 	boolean_t		check_default_group = B_FALSE;
3073 	mac_group_state_t	next_state;
3074 	mac_resource_props_t	*mrp = MCIP_RESOURCE_PROPS(mcip);
3075 
3076 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mip));
3077 
3078 	switch (link_type) {
3079 	case SRST_FLOW:
3080 		mac_rx_srs_group_teardown(flent, B_FALSE);
3081 		mac_tx_srs_group_teardown(mcip, flent, SRST_FLOW);
3082 		return;
3083 
3084 	case SRST_LINK:
3085 		/* Stop sending packets */
3086 		mac_tx_client_block(mcip);
3087 
3088 		/* Stop the packets coming from the H/W */
3089 		if (mcip->mci_unicast != NULL) {
3090 			int err;
3091 			err = mac_remove_macaddr(mcip->mci_unicast);
3092 			if (err != 0) {
3093 				cmn_err(CE_WARN, "%s: failed to remove a MAC"
3094 				    " address because of error 0x%x",
3095 				    mip->mi_name, err);
3096 			}
3097 			mcip->mci_unicast = NULL;
3098 		}
3099 
3100 		/* Stop the packets coming from the S/W classifier */
3101 		mac_flow_remove(mip->mi_flow_tab, flent, B_FALSE);
3102 		mac_flow_wait(flent, FLOW_DRIVER_UPCALL);
3103 
3104 		/* Now quiesce and destroy all SRS and soft rings */
3105 		mac_rx_srs_group_teardown(flent, B_FALSE);
3106 		mac_tx_srs_group_teardown(mcip, flent, SRST_LINK);
3107 
3108 		ASSERT((mcip->mci_flent == flent) &&
3109 		    (flent->fe_next == NULL));
3110 
3111 		/*
3112 		 * Release our hold on the group as well. We need
3113 		 * to check if the shared group has only one client
3114 		 * left who can use it exclusively. Also, if we
3115 		 * were the last client, release the group.
3116 		 */
3117 		group = flent->fe_rx_ring_group;
3118 		default_group = MAC_DEFAULT_RX_GROUP(mip);
3119 		if (group != NULL) {
3120 			mac_group_remove_client(group, mcip);
3121 			next_state = mac_group_next_state(group,
3122 			    &grp_only_mcip, default_group, B_TRUE);
3123 			if (next_state == MAC_GROUP_STATE_RESERVED) {
3124 				/*
3125 				 * Only one client left on this RX group.
3126 				 */
3127 				ASSERT(grp_only_mcip != NULL);
3128 				mac_set_group_state(group,
3129 				    MAC_GROUP_STATE_RESERVED);
3130 				group_only_flent = grp_only_mcip->mci_flent;
3131 
3132 				/*
3133 				 * The only remaining client has exclusive
3134 				 * access on the group. Allow it to
3135 				 * dynamically poll the H/W rings etc.
3136 				 */
3137 				mac_rx_srs_group_setup(grp_only_mcip,
3138 				    group_only_flent, SRST_LINK);
3139 				mac_fanout_setup(grp_only_mcip,
3140 				    group_only_flent,
3141 				    MCIP_RESOURCE_PROPS(grp_only_mcip),
3142 				    mac_rx_deliver, grp_only_mcip, NULL, NULL);
3143 				mac_rx_group_unmark(group, MR_INCIPIENT);
3144 				mac_set_rings_effective(grp_only_mcip);
3145 			} else if (next_state == MAC_GROUP_STATE_REGISTERED) {
3146 				/*
3147 				 * This is a non-default group being freed up.
3148 				 * We need to reevaluate the default group
3149 				 * to see if the primary client can get
3150 				 * exclusive access to the default group.
3151 				 */
3152 				ASSERT(group != MAC_DEFAULT_RX_GROUP(mip));
3153 				if (mrp->mrp_mask & MRP_RX_RINGS) {
3154 					MAC_RX_GRP_RELEASED(mip);
3155 					if (mip->mi_rx_group_type ==
3156 					    MAC_GROUP_TYPE_DYNAMIC) {
3157 						MAC_RX_RING_RELEASED(mip,
3158 						    group->mrg_cur_count);
3159 					}
3160 				}
3161 				mac_release_rx_group(mcip, group);
3162 				mac_set_group_state(group,
3163 				    MAC_GROUP_STATE_REGISTERED);
3164 				check_default_group = B_TRUE;
3165 			} else {
3166 				ASSERT(next_state == MAC_GROUP_STATE_SHARED);
3167 				mac_set_group_state(group,
3168 				    MAC_GROUP_STATE_SHARED);
3169 				mac_rx_group_unmark(group, MR_CONDEMNED);
3170 			}
3171 			flent->fe_rx_ring_group = NULL;
3172 		}
3173 		/*
3174 		 * Remove the client from the TX group. Additionally, if
3175 		 * this a non-default group, then we also need to release
3176 		 * the group.
3177 		 */
3178 		group = flent->fe_tx_ring_group;
3179 		default_group = MAC_DEFAULT_TX_GROUP(mip);
3180 		if (group != NULL) {
3181 			mac_group_remove_client(group, mcip);
3182 			next_state = mac_group_next_state(group,
3183 			    &grp_only_mcip, default_group, B_FALSE);
3184 			if (next_state == MAC_GROUP_STATE_REGISTERED) {
3185 				if (group != default_group) {
3186 					if (mrp->mrp_mask & MRP_TX_RINGS) {
3187 						MAC_TX_GRP_RELEASED(mip);
3188 						if (mip->mi_tx_group_type ==
3189 						    MAC_GROUP_TYPE_DYNAMIC) {
3190 							MAC_TX_RING_RELEASED(
3191 							    mip, group->
3192 							    mrg_cur_count);
3193 						}
3194 					}
3195 					mac_release_tx_group(mcip, group);
3196 					/*
3197 					 * If the default group is reserved,
3198 					 * then we need to set the effective
3199 					 * rings as we would have given
3200 					 * back some rings when the group
3201 					 * was released
3202 					 */
3203 					if (mip->mi_tx_group_type ==
3204 					    MAC_GROUP_TYPE_DYNAMIC &&
3205 					    default_group->mrg_state ==
3206 					    MAC_GROUP_STATE_RESERVED) {
3207 						grp_only_mcip =
3208 						    MAC_GROUP_ONLY_CLIENT
3209 						    (default_group);
3210 						mac_set_rings_effective(
3211 						    grp_only_mcip);
3212 					}
3213 				} else {
3214 					mac_ring_t	*ring;
3215 					int		cnt;
3216 					int		ringcnt;
3217 
3218 					/*
3219 					 * Stop all the rings except the
3220 					 * default ring.
3221 					 */
3222 					ringcnt = group->mrg_cur_count;
3223 					ring = group->mrg_rings;
3224 					for (cnt = 0; cnt < ringcnt; cnt++) {
3225 						if (ring->mr_state ==
3226 						    MR_INUSE && ring !=
3227 						    (mac_ring_t *)
3228 						    mip->mi_default_tx_ring) {
3229 							mac_stop_ring(ring);
3230 							ring->mr_flag = 0;
3231 						}
3232 						ring = ring->mr_next;
3233 					}
3234 				}
3235 			} else if (next_state == MAC_GROUP_STATE_RESERVED) {
3236 				mac_set_rings_effective(grp_only_mcip);
3237 			}
3238 			flent->fe_tx_ring_group = NULL;
3239 			group->mrg_state = next_state;
3240 		}
3241 		break;
3242 	default:
3243 		ASSERT(B_FALSE);
3244 		break;
3245 	}
3246 
3247 	/*
3248 	 * The mac client using the default group gets exclusive access to the
3249 	 * default group if and only if it is the sole client on the entire
3250 	 * mip. If so set the group state to reserved, and set up the SRSes
3251 	 * over the default group.
3252 	 */
3253 	if (check_default_group) {
3254 		default_group = MAC_DEFAULT_RX_GROUP(mip);
3255 		ASSERT(default_group->mrg_state == MAC_GROUP_STATE_SHARED);
3256 		next_state = mac_group_next_state(default_group,
3257 		    &grp_only_mcip, default_group, B_TRUE);
3258 		if (next_state == MAC_GROUP_STATE_RESERVED) {
3259 			ASSERT(grp_only_mcip != NULL &&
3260 			    mip->mi_nactiveclients == 1);
3261 			mac_set_group_state(default_group,
3262 			    MAC_GROUP_STATE_RESERVED);
3263 			mac_rx_srs_group_setup(grp_only_mcip,
3264 			    grp_only_mcip->mci_flent, SRST_LINK);
3265 			mac_fanout_setup(grp_only_mcip,
3266 			    grp_only_mcip->mci_flent,
3267 			    MCIP_RESOURCE_PROPS(grp_only_mcip), mac_rx_deliver,
3268 			    grp_only_mcip, NULL, NULL);
3269 			mac_rx_group_unmark(default_group, MR_INCIPIENT);
3270 			mac_set_rings_effective(grp_only_mcip);
3271 		}
3272 	}
3273 
3274 	/*
3275 	 * If the primary is the only one left and the MAC supports
3276 	 * dynamic grouping, we need to see if the primary needs to
3277 	 * be moved to the default group so that it can use all the
3278 	 * H/W rings.
3279 	 */
3280 	if (!(flent->fe_type & FLOW_PRIMARY_MAC) &&
3281 	    mip->mi_nactiveclients == 1 &&
3282 	    mip->mi_rx_group_type == MAC_GROUP_TYPE_DYNAMIC) {
3283 		default_group = MAC_DEFAULT_RX_GROUP(mip);
3284 		grp_only_mcip = mac_primary_client_handle(mip);
3285 		if (grp_only_mcip == NULL)
3286 			return;
3287 		group_only_flent = grp_only_mcip->mci_flent;
3288 		mrp = MCIP_RESOURCE_PROPS(grp_only_mcip);
3289 		/*
3290 		 * If the primary has an explicit property set, leave it
3291 		 * alone.
3292 		 */
3293 		if (mrp->mrp_mask & MRP_RX_RINGS)
3294 			return;
3295 		/*
3296 		 * Switch the primary to the default group.
3297 		 */
3298 		(void) mac_rx_switch_group(grp_only_mcip,
3299 		    group_only_flent->fe_rx_ring_group, default_group);
3300 	}
3301 }
3302 
3303 /* DATAPATH TEAR DOWN ROUTINES (SRS and FANOUT teardown) */
3304 
3305 static void
3306 mac_srs_fanout_list_free(mac_soft_ring_set_t *mac_srs)
3307 {
3308 	if (mac_srs->srs_type & SRST_TX) {
3309 		mac_srs_tx_t *tx;
3310 
3311 		ASSERT(mac_srs->srs_tcp_soft_rings == NULL);
3312 		ASSERT(mac_srs->srs_udp_soft_rings == NULL);
3313 		ASSERT(mac_srs->srs_oth_soft_rings == NULL);
3314 		ASSERT(mac_srs->srs_tx_soft_rings != NULL);
3315 		kmem_free(mac_srs->srs_tx_soft_rings,
3316 		    sizeof (mac_soft_ring_t *) * MAX_RINGS_PER_GROUP);
3317 		mac_srs->srs_tx_soft_rings = NULL;
3318 		tx = &mac_srs->srs_tx;
3319 		if (tx->st_soft_rings != NULL) {
3320 			kmem_free(tx->st_soft_rings,
3321 			    sizeof (mac_soft_ring_t *) * MAX_RINGS_PER_GROUP);
3322 		}
3323 	} else {
3324 		ASSERT(mac_srs->srs_tx_soft_rings == NULL);
3325 		ASSERT(mac_srs->srs_tcp_soft_rings != NULL);
3326 		kmem_free(mac_srs->srs_tcp_soft_rings,
3327 		    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
3328 		mac_srs->srs_tcp_soft_rings = NULL;
3329 		ASSERT(mac_srs->srs_udp_soft_rings != NULL);
3330 		kmem_free(mac_srs->srs_udp_soft_rings,
3331 		    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
3332 		mac_srs->srs_udp_soft_rings = NULL;
3333 		ASSERT(mac_srs->srs_oth_soft_rings != NULL);
3334 		kmem_free(mac_srs->srs_oth_soft_rings,
3335 		    sizeof (mac_soft_ring_t *) * MAX_SR_FANOUT);
3336 		mac_srs->srs_oth_soft_rings = NULL;
3337 	}
3338 }
3339 
3340 /*
3341  * An RX SRS is attached to at most one mac_ring.
3342  * A TX SRS  has no  rings.
3343  */
3344 static void
3345 mac_srs_ring_free(mac_soft_ring_set_t *mac_srs)
3346 {
3347 	mac_client_impl_t	*mcip;
3348 	mac_ring_t		*ring;
3349 	flow_entry_t		*flent;
3350 
3351 	ring = mac_srs->srs_ring;
3352 	if (mac_srs->srs_type & SRST_TX) {
3353 		ASSERT(ring == NULL);
3354 		return;
3355 	}
3356 
3357 	if (ring == NULL)
3358 		return;
3359 
3360 	/*
3361 	 * Broadcast flows don't have a client impl association, but they
3362 	 * use only soft rings.
3363 	 */
3364 	flent = mac_srs->srs_flent;
3365 	mcip = flent->fe_mcip;
3366 	ASSERT(mcip != NULL);
3367 
3368 	ring->mr_classify_type = MAC_NO_CLASSIFIER;
3369 	ring->mr_srs = NULL;
3370 }
3371 
3372 /*
3373  * Physical unlink and free of the data structures happen below. This is
3374  * driven from mac_flow_destroy(), on the last refrele of a flow.
3375  *
3376  * Assumes Rx srs is 1-1 mapped with an ring.
3377  */
3378 void
3379 mac_srs_free(mac_soft_ring_set_t *mac_srs)
3380 {
3381 	ASSERT(mac_srs->srs_mcip == NULL ||
3382 	    MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
3383 	ASSERT((mac_srs->srs_state & (SRS_CONDEMNED | SRS_CONDEMNED_DONE |
3384 	    SRS_PROC | SRS_PROC_FAST)) == (SRS_CONDEMNED | SRS_CONDEMNED_DONE));
3385 
3386 	mac_pkt_drop(NULL, NULL, mac_srs->srs_first, B_FALSE);
3387 	mac_srs_ring_free(mac_srs);
3388 	mac_srs_soft_rings_free(mac_srs);
3389 	mac_srs_fanout_list_free(mac_srs);
3390 
3391 	mac_srs->srs_bw = NULL;
3392 	mac_srs_stat_delete(mac_srs);
3393 	kmem_cache_free(mac_srs_cache, mac_srs);
3394 }
3395 
3396 static void
3397 mac_srs_soft_rings_quiesce(mac_soft_ring_set_t *mac_srs, uint_t s_ring_flag)
3398 {
3399 	mac_soft_ring_t	*softring;
3400 
3401 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
3402 
3403 	mac_srs_soft_rings_signal(mac_srs, s_ring_flag);
3404 	if (s_ring_flag == S_RING_CONDEMNED) {
3405 		while (mac_srs->srs_soft_ring_condemned_count !=
3406 		    mac_srs->srs_soft_ring_count)
3407 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3408 	} else {
3409 		while (mac_srs->srs_soft_ring_quiesced_count !=
3410 		    mac_srs->srs_soft_ring_count)
3411 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3412 	}
3413 	mutex_exit(&mac_srs->srs_lock);
3414 
3415 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
3416 	    softring = softring->s_ring_next) {
3417 		(void) untimeout(softring->s_ring_tid);
3418 		softring->s_ring_tid = NULL;
3419 	}
3420 
3421 	(void) untimeout(mac_srs->srs_tid);
3422 	mac_srs->srs_tid = NULL;
3423 
3424 	mutex_enter(&mac_srs->srs_lock);
3425 }
3426 
3427 /*
3428  * The block comment above mac_rx_classify_flow_state_change explains the
3429  * background. At this point upcalls from the driver (both hardware classified
3430  * and software classified) have been cut off. We now need to quiesce the
3431  * SRS worker, poll, and softring threads. The SRS worker thread serves as
3432  * the master controller. The steps involved are described below in the function
3433  */
3434 void
3435 mac_srs_worker_quiesce(mac_soft_ring_set_t *mac_srs)
3436 {
3437 	uint_t			s_ring_flag;
3438 	uint_t			srs_poll_wait_flag;
3439 
3440 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
3441 	ASSERT(mac_srs->srs_state & (SRS_CONDEMNED | SRS_QUIESCE));
3442 
3443 	if (mac_srs->srs_state & SRS_CONDEMNED) {
3444 		s_ring_flag = S_RING_CONDEMNED;
3445 		srs_poll_wait_flag = SRS_POLL_THR_EXITED;
3446 	} else {
3447 		s_ring_flag = S_RING_QUIESCE;
3448 		srs_poll_wait_flag = SRS_POLL_THR_QUIESCED;
3449 	}
3450 
3451 	/*
3452 	 * In the case of Rx SRS wait till the poll thread is done.
3453 	 */
3454 	if ((mac_srs->srs_type & SRST_TX) == 0 &&
3455 	    mac_srs->srs_poll_thr != NULL) {
3456 		while (!(mac_srs->srs_state & srs_poll_wait_flag))
3457 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3458 
3459 		/*
3460 		 * Turn off polling as part of the quiesce operation.
3461 		 */
3462 		MAC_SRS_POLLING_OFF(mac_srs);
3463 		mac_srs->srs_state &= ~(SRS_POLLING | SRS_GET_PKTS);
3464 	}
3465 
3466 	/*
3467 	 * Then signal the soft ring worker threads to quiesce or quit
3468 	 * as needed and then wait till that happens.
3469 	 */
3470 	mac_srs_soft_rings_quiesce(mac_srs, s_ring_flag);
3471 
3472 	if (mac_srs->srs_state & SRS_CONDEMNED)
3473 		mac_srs->srs_state |= (SRS_QUIESCE_DONE | SRS_CONDEMNED_DONE);
3474 	else
3475 		mac_srs->srs_state |= SRS_QUIESCE_DONE;
3476 	cv_signal(&mac_srs->srs_quiesce_done_cv);
3477 }
3478 
3479 /*
3480  * Signal an SRS to start a temporary quiesce, or permanent removal, or restart
3481  * a quiesced SRS by setting the appropriate flags and signaling the SRS worker
3482  * or poll thread. This function is internal to the quiescing logic and is
3483  * called internally from the SRS quiesce or flow quiesce or client quiesce
3484  * higher level functions.
3485  */
3486 void
3487 mac_srs_signal(mac_soft_ring_set_t *mac_srs, uint_t srs_flag)
3488 {
3489 	mac_ring_t	*ring;
3490 
3491 	ring = mac_srs->srs_ring;
3492 	ASSERT(ring == NULL || ring->mr_refcnt == 0);
3493 
3494 	if (srs_flag == SRS_CONDEMNED) {
3495 		/*
3496 		 * The SRS is going away. We need to unbind the SRS and SR
3497 		 * threads before removing from the global SRS list. Otherwise
3498 		 * there is a small window where the cpu reconfig callbacks
3499 		 * may miss the SRS in the list walk and DR could fail since
3500 		 * there are still bound threads.
3501 		 */
3502 		mac_srs_threads_unbind(mac_srs);
3503 		mac_srs_remove_glist(mac_srs);
3504 	}
3505 	/*
3506 	 * Wakeup the SRS worker and poll threads.
3507 	 */
3508 	mutex_enter(&mac_srs->srs_lock);
3509 	mac_srs->srs_state |= srs_flag;
3510 	cv_signal(&mac_srs->srs_async);
3511 	cv_signal(&mac_srs->srs_cv);
3512 	mutex_exit(&mac_srs->srs_lock);
3513 }
3514 
3515 /*
3516  * In the Rx side, the quiescing is done bottom up. After the Rx upcalls
3517  * from the driver are done, then the Rx SRS is quiesced and only then can
3518  * we signal the soft rings. Thus this function can't be called arbitrarily
3519  * without satisfying the prerequisites. On the Tx side, the threads from
3520  * top need to quiesced, then the Tx SRS and only then can we signal the
3521  * Tx soft rings.
3522  */
3523 static void
3524 mac_srs_soft_rings_signal(mac_soft_ring_set_t *mac_srs, uint_t sr_flag)
3525 {
3526 	mac_soft_ring_t		*softring;
3527 
3528 	for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
3529 	    softring = softring->s_ring_next)
3530 		mac_soft_ring_signal(softring, sr_flag);
3531 }
3532 
3533 /*
3534  * The block comment above mac_rx_classify_flow_state_change explains the
3535  * background. At this point the SRS is quiesced and we need to restart the
3536  * SRS worker, poll, and softring threads. The SRS worker thread serves as
3537  * the master controller. The steps involved are described below in the function
3538  */
3539 void
3540 mac_srs_worker_restart(mac_soft_ring_set_t *mac_srs)
3541 {
3542 	boolean_t	iam_rx_srs;
3543 	mac_soft_ring_t	*softring;
3544 
3545 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
3546 	if ((mac_srs->srs_type & SRST_TX) != 0) {
3547 		iam_rx_srs = B_FALSE;
3548 		ASSERT((mac_srs->srs_state &
3549 		    (SRS_POLL_THR_QUIESCED | SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
3550 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
3551 	} else {
3552 		iam_rx_srs = B_TRUE;
3553 		ASSERT((mac_srs->srs_state &
3554 		    (SRS_QUIESCE_DONE | SRS_QUIESCE)) ==
3555 		    (SRS_QUIESCE_DONE | SRS_QUIESCE));
3556 		if (mac_srs->srs_poll_thr != NULL) {
3557 			ASSERT((mac_srs->srs_state & SRS_POLL_THR_QUIESCED) ==
3558 			    SRS_POLL_THR_QUIESCED);
3559 		}
3560 	}
3561 
3562 	/*
3563 	 * Signal any quiesced soft ring workers to restart and wait for the
3564 	 * soft ring down count to come down to zero.
3565 	 */
3566 	if (mac_srs->srs_soft_ring_quiesced_count != 0) {
3567 		for (softring = mac_srs->srs_soft_ring_head; softring != NULL;
3568 		    softring = softring->s_ring_next) {
3569 			if (!(softring->s_ring_state & S_RING_QUIESCE))
3570 				continue;
3571 			mac_soft_ring_signal(softring, S_RING_RESTART);
3572 		}
3573 		while (mac_srs->srs_soft_ring_quiesced_count != 0)
3574 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3575 	}
3576 
3577 	mac_srs->srs_state &= ~(SRS_QUIESCE_DONE | SRS_QUIESCE | SRS_RESTART);
3578 	if (iam_rx_srs && mac_srs->srs_poll_thr != NULL) {
3579 		/*
3580 		 * Signal the poll thread and ask it to restart. Wait till it
3581 		 * actually restarts and the SRS_POLL_THR_QUIESCED flag gets
3582 		 * cleared.
3583 		 */
3584 		mac_srs->srs_state |= SRS_POLL_THR_RESTART;
3585 		cv_signal(&mac_srs->srs_cv);
3586 		while (mac_srs->srs_state & SRS_POLL_THR_QUIESCED)
3587 			cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3588 		ASSERT(!(mac_srs->srs_state & SRS_POLL_THR_RESTART));
3589 	}
3590 	/* Wake up any waiter waiting for the restart to complete */
3591 	mac_srs->srs_state |= SRS_RESTART_DONE;
3592 	cv_signal(&mac_srs->srs_quiesce_done_cv);
3593 }
3594 
3595 static void
3596 mac_srs_worker_unbind(mac_soft_ring_set_t *mac_srs)
3597 {
3598 	mutex_enter(&mac_srs->srs_lock);
3599 	if (!(mac_srs->srs_state & SRS_WORKER_BOUND)) {
3600 		ASSERT(mac_srs->srs_worker_cpuid == -1);
3601 		mutex_exit(&mac_srs->srs_lock);
3602 		return;
3603 	}
3604 
3605 	mac_srs->srs_worker_cpuid = -1;
3606 	mac_srs->srs_state &= ~SRS_WORKER_BOUND;
3607 	thread_affinity_clear(mac_srs->srs_worker);
3608 	mutex_exit(&mac_srs->srs_lock);
3609 }
3610 
3611 static void
3612 mac_srs_poll_unbind(mac_soft_ring_set_t *mac_srs)
3613 {
3614 	mutex_enter(&mac_srs->srs_lock);
3615 	if (mac_srs->srs_poll_thr == NULL ||
3616 	    (mac_srs->srs_state & SRS_POLL_BOUND) == 0) {
3617 		ASSERT(mac_srs->srs_poll_cpuid == -1);
3618 		mutex_exit(&mac_srs->srs_lock);
3619 		return;
3620 	}
3621 
3622 	mac_srs->srs_poll_cpuid = -1;
3623 	mac_srs->srs_state &= ~SRS_POLL_BOUND;
3624 	thread_affinity_clear(mac_srs->srs_poll_thr);
3625 	mutex_exit(&mac_srs->srs_lock);
3626 }
3627 
3628 static void
3629 mac_srs_threads_unbind(mac_soft_ring_set_t *mac_srs)
3630 {
3631 	mac_soft_ring_t	*soft_ring;
3632 
3633 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mac_srs->srs_mcip->mci_mip));
3634 
3635 	mutex_enter(&cpu_lock);
3636 	mac_srs_worker_unbind(mac_srs);
3637 	if (!(mac_srs->srs_type & SRST_TX))
3638 		mac_srs_poll_unbind(mac_srs);
3639 
3640 	for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3641 	    soft_ring = soft_ring->s_ring_next) {
3642 		mac_soft_ring_unbind(soft_ring);
3643 	}
3644 	mutex_exit(&cpu_lock);
3645 }
3646 
3647 /*
3648  * When a CPU is going away, unbind all MAC threads which are bound
3649  * to that CPU. The affinity of the thread to the CPU is saved to allow
3650  * the thread to be rebound to the CPU if it comes back online.
3651  */
3652 static void
3653 mac_walk_srs_and_unbind(int cpuid)
3654 {
3655 	mac_soft_ring_set_t *mac_srs;
3656 	mac_soft_ring_t *soft_ring;
3657 
3658 	rw_enter(&mac_srs_g_lock, RW_READER);
3659 
3660 	if ((mac_srs = mac_srs_g_list) == NULL)
3661 		goto done;
3662 
3663 	for (; mac_srs != NULL; mac_srs = mac_srs->srs_next) {
3664 		if (mac_srs->srs_worker_cpuid == cpuid) {
3665 			mac_srs->srs_worker_cpuid_save = cpuid;
3666 			mac_srs_worker_unbind(mac_srs);
3667 		}
3668 
3669 		if (!(mac_srs->srs_type & SRST_TX)) {
3670 			if (mac_srs->srs_poll_cpuid == cpuid) {
3671 				mac_srs->srs_poll_cpuid_save = cpuid;
3672 				mac_srs_poll_unbind(mac_srs);
3673 			}
3674 		}
3675 
3676 		/* Next tackle the soft rings associated with the srs */
3677 		mutex_enter(&mac_srs->srs_lock);
3678 		for (soft_ring = mac_srs->srs_soft_ring_head; soft_ring != NULL;
3679 		    soft_ring = soft_ring->s_ring_next) {
3680 			if (soft_ring->s_ring_cpuid == cpuid) {
3681 				soft_ring->s_ring_cpuid_save = cpuid;
3682 				mac_soft_ring_unbind(soft_ring);
3683 			}
3684 		}
3685 		mutex_exit(&mac_srs->srs_lock);
3686 	}
3687 done:
3688 	rw_exit(&mac_srs_g_lock);
3689 }
3690 
3691 /* TX SETUP and TEARDOWN ROUTINES */
3692 
3693 /*
3694  * XXXHIO need to make sure the two mac_tx_srs_{add,del}_ring()
3695  * handle the case where the number of rings is one. I.e. there is
3696  * a ring pointed to by mac_srs->srs_tx_arg2.
3697  */
3698 void
3699 mac_tx_srs_add_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3700 {
3701 	mac_client_impl_t *mcip = mac_srs->srs_mcip;
3702 	mac_soft_ring_t *soft_ring;
3703 	int count = mac_srs->srs_tx_ring_count;
3704 	uint32_t soft_ring_type = ST_RING_TX;
3705 	uint_t ring_info;
3706 
3707 	ASSERT(mac_srs->srs_state & SRS_QUIESCE);
3708 	ring_info = mac_hwring_getinfo((mac_ring_handle_t)tx_ring);
3709 	if (mac_tx_serialize || (ring_info & MAC_RING_TX_SERIALIZE))
3710 		soft_ring_type |= ST_RING_WORKER_ONLY;
3711 	soft_ring = mac_soft_ring_create(count, 0,
3712 	    soft_ring_type, maxclsyspri, mcip, mac_srs, -1,
3713 	    NULL, mcip, (mac_resource_handle_t)tx_ring);
3714 	mac_srs->srs_tx_ring_count++;
3715 	mac_srs_update_fanout_list(mac_srs);
3716 	/*
3717 	 * put this soft ring in quiesce mode too so when we restart
3718 	 * all soft rings in the srs are in the same state.
3719 	 */
3720 	mac_soft_ring_signal(soft_ring, S_RING_QUIESCE);
3721 }
3722 
3723 static void
3724 mac_soft_ring_remove(mac_soft_ring_set_t *mac_srs, mac_soft_ring_t *softring)
3725 {
3726 	int sringcnt;
3727 
3728 	mutex_enter(&mac_srs->srs_lock);
3729 	sringcnt = mac_srs->srs_soft_ring_count;
3730 	ASSERT(sringcnt > 0);
3731 	mac_soft_ring_signal(softring, S_RING_CONDEMNED);
3732 
3733 	ASSERT(mac_srs->srs_soft_ring_condemned_count == 0);
3734 	while (mac_srs->srs_soft_ring_condemned_count != 1)
3735 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
3736 
3737 	if (softring == mac_srs->srs_soft_ring_head) {
3738 		mac_srs->srs_soft_ring_head = softring->s_ring_next;
3739 		if (mac_srs->srs_soft_ring_head != NULL) {
3740 			mac_srs->srs_soft_ring_head->s_ring_prev = NULL;
3741 		} else {
3742 			mac_srs->srs_soft_ring_tail = NULL;
3743 		}
3744 	} else {
3745 		softring->s_ring_prev->s_ring_next =
3746 		    softring->s_ring_next;
3747 		if (softring->s_ring_next != NULL) {
3748 			softring->s_ring_next->s_ring_prev =
3749 			    softring->s_ring_prev;
3750 		} else {
3751 			mac_srs->srs_soft_ring_tail =
3752 			    softring->s_ring_prev;
3753 		}
3754 	}
3755 	mac_srs->srs_soft_ring_count--;
3756 
3757 	mac_srs->srs_soft_ring_condemned_count--;
3758 	mutex_exit(&mac_srs->srs_lock);
3759 
3760 	mac_soft_ring_free(softring);
3761 }
3762 
3763 void
3764 mac_tx_srs_del_ring(mac_soft_ring_set_t *mac_srs, mac_ring_t *tx_ring)
3765 {
3766 	int i;
3767 	mac_soft_ring_t *soft_ring, *remove_sring;
3768 	mac_client_impl_t *mcip = mac_srs->srs_mcip;
3769 
3770 	mutex_enter(&mac_srs->srs_lock);
3771 	for (i = 0; i < mac_srs->srs_tx_ring_count; i++) {
3772 		soft_ring =  mac_srs->srs_tx_soft_rings[i];
3773 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3774 			break;
3775 	}
3776 	mutex_exit(&mac_srs->srs_lock);
3777 	ASSERT(i < mac_srs->srs_tx_ring_count);
3778 	remove_sring = soft_ring;
3779 	/*
3780 	 * In the case of aggr, the soft ring associated with a Tx ring
3781 	 * is also stored in st_soft_rings[] array. That entry should
3782 	 * be removed.
3783 	 */
3784 	if (mcip->mci_state_flags & MCIS_IS_AGGR) {
3785 		mac_srs_tx_t *tx = &mac_srs->srs_tx;
3786 
3787 		ASSERT(tx->st_soft_rings[tx_ring->mr_index] == remove_sring);
3788 		tx->st_soft_rings[tx_ring->mr_index] = NULL;
3789 	}
3790 	mac_soft_ring_remove(mac_srs, remove_sring);
3791 	mac_srs_update_fanout_list(mac_srs);
3792 }
3793 
3794 /*
3795  * mac_tx_srs_setup():
3796  * Used to setup Tx rings. If no free Tx ring is available, then default
3797  * Tx ring is used.
3798  */
3799 void
3800 mac_tx_srs_setup(mac_client_impl_t *mcip, flow_entry_t *flent)
3801 {
3802 	mac_impl_t		*mip = mcip->mci_mip;
3803 	mac_soft_ring_set_t	*tx_srs = flent->fe_tx_srs;
3804 	int			i;
3805 	int			tx_ring_count = 0;
3806 	uint32_t		soft_ring_type;
3807 	mac_group_t		*grp = NULL;
3808 	mac_ring_t		*ring;
3809 	mac_srs_tx_t		*tx = &tx_srs->srs_tx;
3810 	boolean_t		is_aggr;
3811 	uint_t			ring_info = 0;
3812 
3813 	is_aggr = (mcip->mci_state_flags & MCIS_IS_AGGR) != 0;
3814 	grp = flent->fe_tx_ring_group;
3815 	if (grp == NULL) {
3816 		ring = (mac_ring_t *)mip->mi_default_tx_ring;
3817 		goto no_group;
3818 	}
3819 	tx_ring_count = grp->mrg_cur_count;
3820 	ring = grp->mrg_rings;
3821 	/*
3822 	 * An attempt is made to reserve 'tx_ring_count' number
3823 	 * of Tx rings. If tx_ring_count is 0, default Tx ring
3824 	 * is used. If it is 1, an attempt is made to reserve one
3825 	 * Tx ring. In both the cases, the ring information is
3826 	 * stored in Tx SRS. If multiple Tx rings are specified,
3827 	 * then each Tx ring will have a Tx-side soft ring. All
3828 	 * these soft rings will be hang off Tx SRS.
3829 	 */
3830 	switch (grp->mrg_state) {
3831 		case MAC_GROUP_STATE_SHARED:
3832 		case MAC_GROUP_STATE_RESERVED:
3833 			if (tx_ring_count <= 1 && !is_aggr) {
3834 no_group:
3835 				if (ring != NULL &&
3836 				    ring->mr_state != MR_INUSE) {
3837 					(void) mac_start_ring(ring);
3838 					ring_info = mac_hwring_getinfo(
3839 					    (mac_ring_handle_t)ring);
3840 				}
3841 				tx->st_arg2 = (void *)ring;
3842 				mac_tx_srs_stat_recreate(tx_srs, B_FALSE);
3843 				if (tx_srs->srs_type & SRST_BW_CONTROL) {
3844 					tx->st_mode = SRS_TX_BW;
3845 				} else if (mac_tx_serialize ||
3846 				    (ring_info & MAC_RING_TX_SERIALIZE)) {
3847 					tx->st_mode = SRS_TX_SERIALIZE;
3848 				} else {
3849 					tx->st_mode = SRS_TX_DEFAULT;
3850 				}
3851 				break;
3852 			}
3853 			soft_ring_type = ST_RING_TX;
3854 			if (tx_srs->srs_type & SRST_BW_CONTROL) {
3855 				tx->st_mode = is_aggr ?
3856 				    SRS_TX_BW_AGGR : SRS_TX_BW_FANOUT;
3857 			} else {
3858 				tx->st_mode = is_aggr ? SRS_TX_AGGR :
3859 				    SRS_TX_FANOUT;
3860 			}
3861 			for (i = 0; i < tx_ring_count; i++) {
3862 				ASSERT(ring != NULL);
3863 				switch (ring->mr_state) {
3864 				case MR_INUSE:
3865 				case MR_FREE:
3866 					ASSERT(ring->mr_srs == NULL);
3867 
3868 					if (ring->mr_state != MR_INUSE)
3869 						(void) mac_start_ring(ring);
3870 					ring_info = mac_hwring_getinfo(
3871 					    (mac_ring_handle_t)ring);
3872 					if (mac_tx_serialize || (ring_info &
3873 					    MAC_RING_TX_SERIALIZE)) {
3874 						soft_ring_type |=
3875 						    ST_RING_WORKER_ONLY;
3876 					}
3877 					(void) mac_soft_ring_create(i, 0,
3878 					    soft_ring_type, maxclsyspri,
3879 					    mcip, tx_srs, -1, NULL, mcip,
3880 					    (mac_resource_handle_t)ring);
3881 					break;
3882 				default:
3883 					cmn_err(CE_PANIC,
3884 					    "srs_setup: mcip = %p "
3885 					    "trying to add UNKNOWN ring = %p\n",
3886 					    (void *)mcip, (void *)ring);
3887 					break;
3888 				}
3889 				ring = ring->mr_next;
3890 			}
3891 			mac_srs_update_fanout_list(tx_srs);
3892 			break;
3893 		default:
3894 			ASSERT(B_FALSE);
3895 			break;
3896 	}
3897 	tx->st_func = mac_tx_get_func(tx->st_mode);
3898 	if (is_aggr) {
3899 		VERIFY(i_mac_capab_get((mac_handle_t)mip,
3900 		    MAC_CAPAB_AGGR, &tx->st_capab_aggr));
3901 	}
3902 	DTRACE_PROBE3(tx__srs___setup__return, mac_soft_ring_set_t *, tx_srs,
3903 	    int, tx->st_mode, int, tx_srs->srs_tx_ring_count);
3904 }
3905 
3906 /*
3907  * Update the fanout of a client if its recorded link speed doesn't match
3908  * its current link speed.
3909  */
3910 void
3911 mac_fanout_recompute_client(mac_client_impl_t *mcip, cpupart_t *cpupart)
3912 {
3913 	uint64_t link_speed;
3914 	mac_resource_props_t *mcip_mrp;
3915 	flow_entry_t *flent = mcip->mci_flent;
3916 	mac_soft_ring_set_t *rx_srs;
3917 	mac_cpus_t *srs_cpu;
3918 	int soft_ring_count, maxcpus;
3919 
3920 	ASSERT(MAC_PERIM_HELD((mac_handle_t)mcip->mci_mip));
3921 
3922 	link_speed = mac_client_stat_get(mcip->mci_flent->fe_mcip,
3923 	    MAC_STAT_IFSPEED);
3924 
3925 	if ((link_speed != 0) &&
3926 	    (link_speed != mcip->mci_flent->fe_nic_speed)) {
3927 		mcip_mrp = MCIP_RESOURCE_PROPS(mcip);
3928 		/*
3929 		 * Before calling mac_fanout_setup(), check to see if
3930 		 * the SRSes already have the right number of soft
3931 		 * rings. mac_fanout_setup() is a heavy duty operation
3932 		 * where new cpu bindings are done for SRS and soft
3933 		 * ring threads and interrupts re-targeted.
3934 		 */
3935 		maxcpus = (cpupart != NULL) ? cpupart->cp_ncpus : ncpus;
3936 		soft_ring_count = mac_compute_soft_ring_count(flent,
3937 		    flent->fe_rx_srs_cnt - 1, maxcpus);
3938 		/*
3939 		 * If soft_ring_count returned by
3940 		 * mac_compute_soft_ring_count() is 0, bump it
3941 		 * up by 1 because we always have atleast one
3942 		 * TCP, UDP, and OTH soft ring associated with
3943 		 * an SRS.
3944 		 */
3945 		soft_ring_count = (soft_ring_count == 0) ?
3946 		    1 : soft_ring_count;
3947 		rx_srs = flent->fe_rx_srs[0];
3948 		srs_cpu = &rx_srs->srs_cpu;
3949 		if (soft_ring_count != srs_cpu->mc_rx_fanout_cnt) {
3950 			mac_fanout_setup(mcip, flent, mcip_mrp,
3951 			    mac_rx_deliver, mcip, NULL, cpupart);
3952 		}
3953 	}
3954 }
3955 
3956 /*
3957  * Walk through the list of mac clients for the MAC.
3958  * For each active mac client, recompute the number of soft rings
3959  * associated with every client, only if current speed is different
3960  * from the speed that was previously used for soft ring computation.
3961  * If the cable is disconnected whlie the NIC is started, we would get
3962  * notification with speed set to 0. We do not recompute in that case.
3963  */
3964 void
3965 mac_fanout_recompute(mac_impl_t *mip)
3966 {
3967 	mac_client_impl_t	*mcip;
3968 	cpupart_t		*cpupart;
3969 	boolean_t		use_default;
3970 	mac_resource_props_t	*mrp, *emrp;
3971 
3972 	i_mac_perim_enter(mip);
3973 	if ((mip->mi_state_flags & MIS_IS_VNIC) != 0 ||
3974 	    mip->mi_linkstate != LINK_STATE_UP) {
3975 		i_mac_perim_exit(mip);
3976 		return;
3977 	}
3978 
3979 	for (mcip = mip->mi_clients_list; mcip != NULL;
3980 	    mcip = mcip->mci_client_next) {
3981 		if ((mcip->mci_state_flags & MCIS_SHARE_BOUND) != 0 ||
3982 		    !MCIP_DATAPATH_SETUP(mcip))
3983 			continue;
3984 		mrp = MCIP_RESOURCE_PROPS(mcip);
3985 		emrp = MCIP_EFFECTIVE_PROPS(mcip);
3986 		use_default = B_FALSE;
3987 		pool_lock();
3988 		cpupart = mac_pset_find(mrp, &use_default);
3989 		mac_fanout_recompute_client(mcip, cpupart);
3990 		mac_set_pool_effective(use_default, cpupart, mrp, emrp);
3991 		pool_unlock();
3992 	}
3993 	i_mac_perim_exit(mip);
3994 }
3995 
3996 /*
3997  * Given a MAC, change the polling state for all its MAC clients.  'enable' is
3998  * B_TRUE to enable polling or B_FALSE to disable.  Polling is enabled by
3999  * default.
4000  */
4001 void
4002 mac_poll_state_change(mac_handle_t mh, boolean_t enable)
4003 {
4004 	mac_impl_t *mip = (mac_impl_t *)mh;
4005 	mac_client_impl_t *mcip;
4006 
4007 	i_mac_perim_enter(mip);
4008 	if (enable)
4009 		mip->mi_state_flags &= ~MIS_POLL_DISABLE;
4010 	else
4011 		mip->mi_state_flags |= MIS_POLL_DISABLE;
4012 	for (mcip = mip->mi_clients_list; mcip != NULL;
4013 	    mcip = mcip->mci_client_next)
4014 		mac_client_update_classifier(mcip, B_TRUE);
4015 	i_mac_perim_exit(mip);
4016 }
4017