xref: /illumos-gate/usr/src/uts/common/io/aggr/aggr_grp.c (revision 09fbbb7d1b5a956a9b81304070dcff318a4a158e)
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) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2020 Joyent, Inc.
24  * Copyright 2020 RackTop Systems, Inc.
25  */
26 
27 /*
28  * IEEE 802.3ad Link Aggregation -- Link Aggregation Groups.
29  *
30  * An instance of the structure aggr_grp_t is allocated for each
31  * link aggregation group. When created, aggr_grp_t objects are
32  * entered into the aggr_grp_hash hash table maintained by the modhash
33  * module. The hash key is the linkid associated with the link
34  * aggregation group.
35  *
36  * Each aggregation contains a set of ports. The port is represented
37  * by the aggr_port_t structure. A port consists of a single MAC
38  * client which has exclusive (MCIS_EXCLUSIVE) use of the underlying
39  * MAC. This client is used by the aggr to send and receive LACP
40  * traffic. Each port client takes on the same MAC unicast address --
41  * the address of the aggregation itself (taken from the first port by
42  * default).
43  *
44  * The MAC client that hangs off each aggr port is not your typical
45  * MAC client. Not only does it have exclusive control of the MAC, but
46  * it also has no Tx or Rx SRSes. An SRS is designed to queue and
47  * fanout traffic among L4 protocols; but the aggr is an intermediary,
48  * not a consumer. Instead of using SRSes, the aggr puts the
49  * underlying hardware rings into passthru mode and ships packets up
50  * via a direct call to aggr_recv_cb(). This allows aggr to enforce
51  * LACP while passing all other traffic up to clients of the aggr.
52  *
53  * Pseudo Rx Groups and Rings
54  * --------------------------
55  *
56  * It is imperative for client performance that the aggr provide as
57  * many MAC groups as possible. In order to use the underlying HW
58  * resources, aggr creates pseudo groups to aggregate the underlying
59  * HW groups. Every HW group gets mapped to a pseudo group; and every
60  * HW ring in that group gets mapped to a pseudo ring. The pseudo
61  * group at index 0 combines all the HW groups at index 0 from each
62  * port, etc. The aggr's MAC then creates normal MAC groups and rings
63  * out of these pseudo groups and rings to present to the aggr's
64  * clients. To the clients, the aggr's groups and rings are absolutely
65  * no different than a NIC's groups or rings.
66  *
67  * Pseudo Tx Rings
68  * ---------------
69  *
70  * The underlying ports (NICs) in an aggregation can have Tx rings. To
71  * enhance aggr's performance, these Tx rings are made available to
72  * the aggr layer as pseudo Tx rings. The concept of pseudo rings are
73  * not new. They are already present and implemented on the Rx side.
74  * The same concept is extended to the Tx side where each Tx ring of
75  * an underlying port is reflected in aggr as a pseudo Tx ring. Thus
76  * each pseudo Tx ring will map to a specific hardware Tx ring. Even
77  * in the case of a NIC that does not have a Tx ring, a pseudo Tx ring
78  * is given to the aggregation layer.
79  *
80  * With this change, the outgoing stack depth looks much better:
81  *
82  * mac_tx() -> mac_tx_aggr_mode() -> mac_tx_soft_ring_process() ->
83  * mac_tx_send() -> aggr_ring_rx() -> <driver>_ring_tx()
84  *
85  * Two new modes are introduced to mac_tx() to handle aggr pseudo Tx rings:
86  * SRS_TX_AGGR and SRS_TX_BW_AGGR.
87  *
88  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
89  * invokes an aggr function, aggr_find_tx_ring(), to find a (pseudo) Tx
90  * ring belonging to a port on which the packet has to be sent.
91  * aggr_find_tx_ring() first finds the outgoing port based on L2/L3/L4
92  * policy and then uses the fanout_hint passed to it to pick a Tx ring from
93  * the selected port.
94  *
95  * In SRS_TX_BW_AGGR mode, mac_tx_bw_mode() function is called where
96  * bandwidth limit is applied first on the outgoing packet and the packets
97  * allowed to go out would call mac_tx_aggr_mode() to send the packet on a
98  * particular Tx ring.
99  */
100 
101 #include <sys/types.h>
102 #include <sys/sysmacros.h>
103 #include <sys/conf.h>
104 #include <sys/cmn_err.h>
105 #include <sys/disp.h>
106 #include <sys/list.h>
107 #include <sys/ksynch.h>
108 #include <sys/kmem.h>
109 #include <sys/stream.h>
110 #include <sys/modctl.h>
111 #include <sys/ddi.h>
112 #include <sys/sunddi.h>
113 #include <sys/atomic.h>
114 #include <sys/stat.h>
115 #include <sys/modhash.h>
116 #include <sys/id_space.h>
117 #include <sys/strsun.h>
118 #include <sys/cred.h>
119 #include <sys/dlpi.h>
120 #include <sys/zone.h>
121 #include <sys/mac_provider.h>
122 #include <sys/dls.h>
123 #include <sys/vlan.h>
124 #include <sys/aggr.h>
125 #include <sys/aggr_impl.h>
126 
127 static int aggr_m_start(void *);
128 static void aggr_m_stop(void *);
129 static int aggr_m_promisc(void *, boolean_t);
130 static int aggr_m_multicst(void *, boolean_t, const uint8_t *);
131 static int aggr_m_unicst(void *, const uint8_t *);
132 static int aggr_m_stat(void *, uint_t, uint64_t *);
133 static void aggr_m_ioctl(void *, queue_t *, mblk_t *);
134 static boolean_t aggr_m_capab_get(void *, mac_capab_t, void *);
135 static int aggr_m_setprop(void *, const char *, mac_prop_id_t, uint_t,
136     const void *);
137 static void aggr_m_propinfo(void *, const char *, mac_prop_id_t,
138     mac_prop_info_handle_t);
139 
140 static aggr_port_t *aggr_grp_port_lookup(aggr_grp_t *, datalink_id_t);
141 static int aggr_grp_rem_port(aggr_grp_t *, aggr_port_t *, boolean_t *,
142     boolean_t *);
143 
144 static void aggr_grp_capab_set(aggr_grp_t *);
145 static boolean_t aggr_grp_capab_check(aggr_grp_t *, aggr_port_t *);
146 static uint_t aggr_grp_max_sdu(aggr_grp_t *);
147 static uint32_t aggr_grp_max_margin(aggr_grp_t *);
148 static boolean_t aggr_grp_sdu_check(aggr_grp_t *, aggr_port_t *);
149 static boolean_t aggr_grp_margin_check(aggr_grp_t *, aggr_port_t *);
150 
151 static int aggr_add_pseudo_rx_group(aggr_port_t *, aggr_pseudo_rx_group_t *);
152 static void aggr_rem_pseudo_rx_group(aggr_port_t *, aggr_pseudo_rx_group_t *);
153 static int aggr_pseudo_disable_intr(mac_intr_handle_t);
154 static int aggr_pseudo_enable_intr(mac_intr_handle_t);
155 static int aggr_pseudo_start_rx_ring(mac_ring_driver_t, uint64_t);
156 static void aggr_pseudo_stop_rx_ring(mac_ring_driver_t);
157 static int aggr_addmac(void *, const uint8_t *);
158 static int aggr_remmac(void *, const uint8_t *);
159 static int aggr_addvlan(mac_group_driver_t, uint16_t);
160 static int aggr_remvlan(mac_group_driver_t, uint16_t);
161 static mblk_t *aggr_rx_poll(void *, int);
162 static void aggr_fill_ring(void *, mac_ring_type_t, const int,
163     const int, mac_ring_info_t *, mac_ring_handle_t);
164 static void aggr_fill_group(void *, mac_ring_type_t, const int,
165     mac_group_info_t *, mac_group_handle_t);
166 
167 static kmem_cache_t	*aggr_grp_cache;
168 static mod_hash_t	*aggr_grp_hash;
169 static krwlock_t	aggr_grp_lock;
170 static uint_t		aggr_grp_cnt;
171 static id_space_t	*key_ids;
172 
173 #define	GRP_HASHSZ		64
174 #define	GRP_HASH_KEY(linkid)	((mod_hash_key_t)(uintptr_t)linkid)
175 #define	AGGR_PORT_NAME_DELIMIT '-'
176 
177 static uchar_t aggr_zero_mac[] = {0, 0, 0, 0, 0, 0};
178 
179 #define	AGGR_M_CALLBACK_FLAGS	\
180 	(MC_IOCTL | MC_GETCAPAB | MC_SETPROP | MC_PROPINFO)
181 
182 static mac_callbacks_t aggr_m_callbacks = {
183 	AGGR_M_CALLBACK_FLAGS,
184 	aggr_m_stat,
185 	aggr_m_start,
186 	aggr_m_stop,
187 	aggr_m_promisc,
188 	aggr_m_multicst,
189 	NULL,
190 	NULL,
191 	NULL,
192 	aggr_m_ioctl,
193 	aggr_m_capab_get,
194 	NULL,
195 	NULL,
196 	aggr_m_setprop,
197 	NULL,
198 	aggr_m_propinfo
199 };
200 
201 /*ARGSUSED*/
202 static int
203 aggr_grp_constructor(void *buf, void *arg, int kmflag)
204 {
205 	aggr_grp_t *grp = buf;
206 
207 	bzero(grp, sizeof (*grp));
208 	mutex_init(&grp->lg_lacp_lock, NULL, MUTEX_DEFAULT, NULL);
209 	cv_init(&grp->lg_lacp_cv, NULL, CV_DEFAULT, NULL);
210 	rw_init(&grp->lg_tx_lock, NULL, RW_DRIVER, NULL);
211 	mutex_init(&grp->lg_port_lock, NULL, MUTEX_DEFAULT, NULL);
212 	cv_init(&grp->lg_port_cv, NULL, CV_DEFAULT, NULL);
213 	mutex_init(&grp->lg_tx_flowctl_lock, NULL, MUTEX_DEFAULT, NULL);
214 	cv_init(&grp->lg_tx_flowctl_cv, NULL, CV_DEFAULT, NULL);
215 	grp->lg_link_state = LINK_STATE_UNKNOWN;
216 	return (0);
217 }
218 
219 /*ARGSUSED*/
220 static void
221 aggr_grp_destructor(void *buf, void *arg)
222 {
223 	aggr_grp_t *grp = buf;
224 
225 	if (grp->lg_tx_ports != NULL) {
226 		kmem_free(grp->lg_tx_ports,
227 		    grp->lg_tx_ports_size * sizeof (aggr_port_t *));
228 	}
229 
230 	mutex_destroy(&grp->lg_lacp_lock);
231 	cv_destroy(&grp->lg_lacp_cv);
232 	mutex_destroy(&grp->lg_port_lock);
233 	cv_destroy(&grp->lg_port_cv);
234 	rw_destroy(&grp->lg_tx_lock);
235 	mutex_destroy(&grp->lg_tx_flowctl_lock);
236 	cv_destroy(&grp->lg_tx_flowctl_cv);
237 }
238 
239 void
240 aggr_grp_init(void)
241 {
242 	aggr_grp_cache = kmem_cache_create("aggr_grp_cache",
243 	    sizeof (aggr_grp_t), 0, aggr_grp_constructor,
244 	    aggr_grp_destructor, NULL, NULL, NULL, 0);
245 
246 	aggr_grp_hash = mod_hash_create_idhash("aggr_grp_hash",
247 	    GRP_HASHSZ, mod_hash_null_valdtor);
248 	rw_init(&aggr_grp_lock, NULL, RW_DEFAULT, NULL);
249 	aggr_grp_cnt = 0;
250 
251 	/*
252 	 * Allocate an id space to manage key values (when key is not
253 	 * specified). The range of the id space will be from
254 	 * (AGGR_MAX_KEY + 1) to UINT16_MAX, because the LACP protocol
255 	 * uses a 16-bit key.
256 	 */
257 	key_ids = id_space_create("aggr_key_ids", AGGR_MAX_KEY + 1, UINT16_MAX);
258 	ASSERT(key_ids != NULL);
259 }
260 
261 void
262 aggr_grp_fini(void)
263 {
264 	id_space_destroy(key_ids);
265 	rw_destroy(&aggr_grp_lock);
266 	mod_hash_destroy_idhash(aggr_grp_hash);
267 	kmem_cache_destroy(aggr_grp_cache);
268 }
269 
270 uint_t
271 aggr_grp_count(void)
272 {
273 	uint_t	count;
274 
275 	rw_enter(&aggr_grp_lock, RW_READER);
276 	count = aggr_grp_cnt;
277 	rw_exit(&aggr_grp_lock);
278 	return (count);
279 }
280 
281 /*
282  * Since both aggr_port_notify_cb() and aggr_port_timer_thread() functions
283  * requires the mac perimeter, this function holds a reference of the aggr
284  * and aggr won't call mac_unregister() until this reference drops to 0.
285  */
286 void
287 aggr_grp_port_hold(aggr_port_t *port)
288 {
289 	aggr_grp_t	*grp = port->lp_grp;
290 
291 	AGGR_PORT_REFHOLD(port);
292 	mutex_enter(&grp->lg_port_lock);
293 	grp->lg_port_ref++;
294 	mutex_exit(&grp->lg_port_lock);
295 }
296 
297 /*
298  * Release the reference of the grp and inform aggr_grp_delete() calling
299  * mac_unregister() is now safe.
300  */
301 void
302 aggr_grp_port_rele(aggr_port_t *port)
303 {
304 	aggr_grp_t	*grp = port->lp_grp;
305 
306 	mutex_enter(&grp->lg_port_lock);
307 	if (--grp->lg_port_ref == 0)
308 		cv_signal(&grp->lg_port_cv);
309 	mutex_exit(&grp->lg_port_lock);
310 	AGGR_PORT_REFRELE(port);
311 }
312 
313 /*
314  * Wait for the port's lacp timer thread and the port's notification callback
315  * to exit.
316  */
317 void
318 aggr_grp_port_wait(aggr_grp_t *grp)
319 {
320 	mutex_enter(&grp->lg_port_lock);
321 	if (grp->lg_port_ref != 0)
322 		cv_wait(&grp->lg_port_cv, &grp->lg_port_lock);
323 	mutex_exit(&grp->lg_port_lock);
324 }
325 
326 /*
327  * Attach a port to a link aggregation group.
328  *
329  * A port is attached to a link aggregation group once its speed
330  * and link state have been verified.
331  *
332  * Returns B_TRUE if the group link state or speed has changed. If
333  * it's the case, the caller must notify the MAC layer via a call
334  * to mac_link().
335  */
336 boolean_t
337 aggr_grp_attach_port(aggr_grp_t *grp, aggr_port_t *port)
338 {
339 	boolean_t link_state_changed = B_FALSE;
340 
341 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
342 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
343 
344 	if (port->lp_state == AGGR_PORT_STATE_ATTACHED)
345 		return (B_FALSE);
346 
347 	/*
348 	 * Validate the MAC port link speed and update the group
349 	 * link speed if needed.
350 	 */
351 	if (port->lp_ifspeed == 0 ||
352 	    port->lp_link_state != LINK_STATE_UP ||
353 	    port->lp_link_duplex != LINK_DUPLEX_FULL) {
354 		/*
355 		 * Can't attach a MAC port with unknown link speed,
356 		 * down link, or not in full duplex mode.
357 		 */
358 		return (B_FALSE);
359 	}
360 
361 	mutex_enter(&grp->lg_stat_lock);
362 	if (grp->lg_ifspeed == 0) {
363 		/*
364 		 * The group inherits the speed of the first link being
365 		 * attached.
366 		 */
367 		grp->lg_ifspeed = port->lp_ifspeed;
368 		link_state_changed = B_TRUE;
369 	} else if (grp->lg_ifspeed != port->lp_ifspeed) {
370 		/*
371 		 * The link speed of the MAC port must be the same as
372 		 * the group link speed, as per 802.3ad. Since it is
373 		 * not, the attach is cancelled.
374 		 */
375 		mutex_exit(&grp->lg_stat_lock);
376 		return (B_FALSE);
377 	}
378 	mutex_exit(&grp->lg_stat_lock);
379 
380 	grp->lg_nattached_ports++;
381 
382 	/*
383 	 * Update the group link state.
384 	 */
385 	if (grp->lg_link_state != LINK_STATE_UP) {
386 		grp->lg_link_state = LINK_STATE_UP;
387 		mutex_enter(&grp->lg_stat_lock);
388 		grp->lg_link_duplex = LINK_DUPLEX_FULL;
389 		mutex_exit(&grp->lg_stat_lock);
390 		link_state_changed = B_TRUE;
391 	}
392 
393 	/*
394 	 * Update port's state.
395 	 */
396 	port->lp_state = AGGR_PORT_STATE_ATTACHED;
397 
398 	aggr_grp_multicst_port(port, B_TRUE);
399 
400 	/*
401 	 * The port client doesn't have an Rx SRS; instead of calling
402 	 * mac_rx_set() we set the client's flow callback directly.
403 	 * This datapath is used only when the port's driver doesn't
404 	 * support MAC_CAPAB_RINGS. Drivers with ring support will
405 	 * deliver traffic to the aggr via ring passthru.
406 	 */
407 	mac_client_set_flow_cb(port->lp_mch, aggr_recv_cb, port);
408 
409 	/*
410 	 * If LACP is OFF, the port can be used to send data as soon
411 	 * as its link is up and verified to be compatible with the
412 	 * aggregation.
413 	 *
414 	 * If LACP is active or passive, notify the LACP subsystem, which
415 	 * will enable sending on the port following the LACP protocol.
416 	 */
417 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
418 		aggr_send_port_enable(port);
419 	else
420 		aggr_lacp_port_attached(port);
421 
422 	return (link_state_changed);
423 }
424 
425 boolean_t
426 aggr_grp_detach_port(aggr_grp_t *grp, aggr_port_t *port)
427 {
428 	boolean_t link_state_changed = B_FALSE;
429 
430 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
431 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
432 
433 	/* update state */
434 	if (port->lp_state != AGGR_PORT_STATE_ATTACHED)
435 		return (B_FALSE);
436 
437 	mac_client_clear_flow_cb(port->lp_mch);
438 
439 	aggr_grp_multicst_port(port, B_FALSE);
440 
441 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
442 		aggr_send_port_disable(port);
443 	else
444 		aggr_lacp_port_detached(port);
445 
446 	port->lp_state = AGGR_PORT_STATE_STANDBY;
447 
448 	grp->lg_nattached_ports--;
449 	if (grp->lg_nattached_ports == 0) {
450 		/* the last attached MAC port of the group is being detached */
451 		grp->lg_link_state = LINK_STATE_DOWN;
452 		mutex_enter(&grp->lg_stat_lock);
453 		grp->lg_ifspeed = 0;
454 		grp->lg_link_duplex = LINK_DUPLEX_UNKNOWN;
455 		mutex_exit(&grp->lg_stat_lock);
456 		link_state_changed = B_TRUE;
457 	}
458 
459 	return (link_state_changed);
460 }
461 
462 /*
463  * Update the MAC addresses of the constituent ports of the specified
464  * group. This function is invoked:
465  * - after creating a new aggregation group.
466  * - after adding new ports to an aggregation group.
467  * - after removing a port from a group when the MAC address of
468  *   that port was used for the MAC address of the group.
469  * - after the MAC address of a port changed when the MAC address
470  *   of that port was used for the MAC address of the group.
471  *
472  * Return true if the link state of the aggregation changed, for example
473  * as a result of a failure changing the MAC address of one of the
474  * constituent ports.
475  */
476 boolean_t
477 aggr_grp_update_ports_mac(aggr_grp_t *grp)
478 {
479 	aggr_port_t *cport;
480 	boolean_t link_state_changed = B_FALSE;
481 	mac_perim_handle_t mph;
482 
483 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
484 
485 	for (cport = grp->lg_ports; cport != NULL;
486 	    cport = cport->lp_next) {
487 		mac_perim_enter_by_mh(cport->lp_mh, &mph);
488 		if (aggr_port_unicst(cport) != 0) {
489 			if (aggr_grp_detach_port(grp, cport))
490 				link_state_changed = B_TRUE;
491 		} else {
492 			/*
493 			 * If a port was detached because of a previous
494 			 * failure changing the MAC address, the port is
495 			 * reattached when it successfully changes the MAC
496 			 * address now, and this might cause the link state
497 			 * of the aggregation to change.
498 			 */
499 			if (aggr_grp_attach_port(grp, cport))
500 				link_state_changed = B_TRUE;
501 		}
502 		mac_perim_exit(mph);
503 	}
504 	return (link_state_changed);
505 }
506 
507 /*
508  * Invoked when the MAC address of a port has changed. If the port's
509  * MAC address was used for the group MAC address, set mac_addr_changedp
510  * to B_TRUE to indicate to the caller that it should send a MAC_NOTE_UNICST
511  * notification. If the link state changes due to detach/attach of
512  * the constituent port, set link_state_changedp to B_TRUE to indicate
513  * to the caller that it should send a MAC_NOTE_LINK notification. In both
514  * cases, it is the responsibility of the caller to invoke notification
515  * functions after releasing the the port lock.
516  */
517 void
518 aggr_grp_port_mac_changed(aggr_grp_t *grp, aggr_port_t *port,
519     boolean_t *mac_addr_changedp, boolean_t *link_state_changedp)
520 {
521 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
522 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
523 	ASSERT(mac_addr_changedp != NULL);
524 	ASSERT(link_state_changedp != NULL);
525 
526 	*mac_addr_changedp = B_FALSE;
527 	*link_state_changedp = B_FALSE;
528 
529 	if (grp->lg_addr_fixed) {
530 		/*
531 		 * The group is using a fixed MAC address or an automatic
532 		 * MAC address has not been set.
533 		 */
534 		return;
535 	}
536 
537 	if (grp->lg_mac_addr_port == port) {
538 		/*
539 		 * The MAC address of the port was assigned to the group
540 		 * MAC address. Update the group MAC address.
541 		 */
542 		bcopy(port->lp_addr, grp->lg_addr, ETHERADDRL);
543 		*mac_addr_changedp = B_TRUE;
544 	} else {
545 		/*
546 		 * Update the actual port MAC address to the MAC address
547 		 * of the group.
548 		 */
549 		if (aggr_port_unicst(port) != 0) {
550 			*link_state_changedp = aggr_grp_detach_port(grp, port);
551 		} else {
552 			/*
553 			 * If a port was detached because of a previous
554 			 * failure changing the MAC address, the port is
555 			 * reattached when it successfully changes the MAC
556 			 * address now, and this might cause the link state
557 			 * of the aggregation to change.
558 			 */
559 			*link_state_changedp = aggr_grp_attach_port(grp, port);
560 		}
561 	}
562 }
563 
564 /*
565  * Add a port to a link aggregation group.
566  */
567 static int
568 aggr_grp_add_port(aggr_grp_t *grp, datalink_id_t port_linkid, boolean_t force,
569     aggr_port_t **pp)
570 {
571 	aggr_port_t *port, **cport;
572 	mac_perim_handle_t mph;
573 	zoneid_t port_zoneid = ALL_ZONES;
574 	int err;
575 
576 	/* The port must be in the same zone as the aggregation. */
577 	if (zone_check_datalink(&port_zoneid, port_linkid) != 0)
578 		port_zoneid = GLOBAL_ZONEID;
579 	if (grp->lg_zoneid != port_zoneid)
580 		return (EBUSY);
581 
582 	/*
583 	 * If we are creating the aggr, then there is no MAC handle
584 	 * and thus no perimeter to hold. If we are adding a port to
585 	 * an existing aggr, then the perimiter of the aggr's MAC must
586 	 * be held.
587 	 */
588 	ASSERT(grp->lg_mh == NULL || MAC_PERIM_HELD(grp->lg_mh));
589 
590 	err = aggr_port_create(grp, port_linkid, force, &port);
591 	if (err != 0)
592 		return (err);
593 
594 	mac_perim_enter_by_mh(port->lp_mh, &mph);
595 
596 	/* Add the new port to the end of the list. */
597 	cport = &grp->lg_ports;
598 	while (*cport != NULL)
599 		cport = &((*cport)->lp_next);
600 	*cport = port;
601 
602 	/*
603 	 * Back reference to the group it is member of. A port always
604 	 * holds a reference to its group to ensure that the back
605 	 * reference is always valid.
606 	 */
607 	port->lp_grp = grp;
608 	AGGR_GRP_REFHOLD(grp);
609 	grp->lg_nports++;
610 	if (grp->lg_nports > grp->lg_nports_high)
611 		grp->lg_nports_high = grp->lg_nports;
612 
613 	aggr_lacp_init_port(port);
614 	mac_perim_exit(mph);
615 
616 	if (pp != NULL)
617 		*pp = port;
618 
619 	return (0);
620 }
621 
622 /*
623  * This is called when the 'lg_tx_ports' arrangement has changed and
624  * we need to update the corresponding 'mi_default_tx_ring'. This
625  * happens for several reasons.
626  *
627  *     - A pseudo TX mac group was added or removed.
628  *     - An LACP message has changed the port's state.
629  *     - A link event has changed the port's state.
630  *
631  * In any case, we see if there is at least one port enabled (see
632  * 'aggr_send_port_enable()'), and if so we use its first ring as the
633  * mac's default TX ring.
634  *
635  * Note, because we only have a single TX group, we don't have to
636  * worry about the rings moving between groups and the chance that mac
637  * will reassign it unless someone removes a port, at which point, we
638  * play it safe and call this again.
639  */
640 void
641 aggr_grp_update_default(aggr_grp_t *grp)
642 {
643 	aggr_port_t *port;
644 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
645 
646 	rw_enter(&grp->lg_tx_lock, RW_WRITER);
647 
648 	if (grp->lg_ntx_ports == 0) {
649 		rw_exit(&grp->lg_tx_lock);
650 		return;
651 	}
652 
653 	port = grp->lg_tx_ports[0];
654 	ASSERT(port->lp_tx_ring_cnt > 0);
655 	mac_hwring_set_default(grp->lg_mh, port->lp_pseudo_tx_rings[0]);
656 	rw_exit(&grp->lg_tx_lock);
657 }
658 
659 /*
660  * Add a pseudo RX ring for the given HW ring handle.
661  */
662 static int
663 aggr_add_pseudo_rx_ring(aggr_port_t *port,
664     aggr_pseudo_rx_group_t *rx_grp, mac_ring_handle_t hw_rh)
665 {
666 	aggr_pseudo_rx_ring_t	*ring;
667 	int			err;
668 	int			j;
669 
670 	for (j = 0; j < MAX_RINGS_PER_GROUP; j++) {
671 		ring = rx_grp->arg_rings + j;
672 		if (!(ring->arr_flags & MAC_PSEUDO_RING_INUSE))
673 			break;
674 	}
675 
676 	/*
677 	 * No slot for this new RX ring.
678 	 */
679 	if (j == MAX_RINGS_PER_GROUP)
680 		return (ENOSPC);
681 
682 	ring->arr_flags |= MAC_PSEUDO_RING_INUSE;
683 	ring->arr_hw_rh = hw_rh;
684 	ring->arr_port = port;
685 	ring->arr_grp = rx_grp;
686 	rx_grp->arg_ring_cnt++;
687 
688 	/*
689 	 * The group is already registered, dynamically add a new ring to the
690 	 * mac group.
691 	 */
692 	if ((err = mac_group_add_ring(rx_grp->arg_gh, j)) != 0) {
693 		ring->arr_flags &= ~MAC_PSEUDO_RING_INUSE;
694 		ring->arr_hw_rh = NULL;
695 		ring->arr_port = NULL;
696 		ring->arr_grp = NULL;
697 		rx_grp->arg_ring_cnt--;
698 	} else {
699 		/*
700 		 * This must run after the MAC is registered.
701 		 */
702 		ASSERT3P(ring->arr_rh, !=, NULL);
703 		mac_hwring_set_passthru(hw_rh, (mac_rx_t)aggr_recv_cb,
704 		    (void *)port, (mac_resource_handle_t)ring);
705 	}
706 	return (err);
707 }
708 
709 /*
710  * Remove the pseudo RX ring of the given HW ring handle.
711  */
712 static void
713 aggr_rem_pseudo_rx_ring(aggr_pseudo_rx_group_t *rx_grp, mac_ring_handle_t hw_rh)
714 {
715 	for (uint_t j = 0; j < MAX_RINGS_PER_GROUP; j++) {
716 		aggr_pseudo_rx_ring_t *ring = rx_grp->arg_rings + j;
717 
718 		if (!(ring->arr_flags & MAC_PSEUDO_RING_INUSE) ||
719 		    ring->arr_hw_rh != hw_rh) {
720 			continue;
721 		}
722 
723 		mac_group_rem_ring(rx_grp->arg_gh, ring->arr_rh);
724 
725 		ring->arr_flags &= ~MAC_PSEUDO_RING_INUSE;
726 		ring->arr_hw_rh = NULL;
727 		ring->arr_port = NULL;
728 		ring->arr_grp = NULL;
729 		rx_grp->arg_ring_cnt--;
730 		mac_hwring_clear_passthru(hw_rh);
731 		break;
732 	}
733 }
734 
735 /*
736  * Create pseudo rings over the HW rings of the port.
737  *
738  * o Create a pseudo ring in rx_grp per HW ring in the port's HW group.
739  *
740  * o Program existing unicast filters on the pseudo group into the HW group.
741  *
742  * o Program existing VLAN filters on the pseudo group into the HW group.
743  */
744 static int
745 aggr_add_pseudo_rx_group(aggr_port_t *port, aggr_pseudo_rx_group_t *rx_grp)
746 {
747 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP];
748 	aggr_unicst_addr_t	*addr, *a;
749 	mac_perim_handle_t	pmph;
750 	aggr_vlan_t		*avp;
751 	uint_t			hw_rh_cnt, i;
752 	int			err = 0;
753 	uint_t			g_idx = rx_grp->arg_index;
754 
755 	ASSERT(MAC_PERIM_HELD(port->lp_grp->lg_mh));
756 	ASSERT3U(g_idx, <, MAX_GROUPS_PER_PORT);
757 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
758 
759 	i = 0;
760 	addr = NULL;
761 	/*
762 	 * This function must be called after the aggr registers its
763 	 * MAC and its Rx groups have been initialized.
764 	 */
765 	ASSERT(rx_grp->arg_gh != NULL);
766 
767 	/*
768 	 * Get the list of the underlying HW rings.
769 	 */
770 	hw_rh_cnt = mac_hwrings_idx_get(port->lp_mh, g_idx,
771 	    &port->lp_hwghs[g_idx], hw_rh, MAC_RING_TYPE_RX);
772 
773 	/*
774 	 * Add existing VLAN and unicast address filters to the port.
775 	 */
776 	for (avp = list_head(&rx_grp->arg_vlans); avp != NULL;
777 	    avp = list_next(&rx_grp->arg_vlans, avp)) {
778 		if ((err = aggr_port_addvlan(port, g_idx, avp->av_vid)) != 0)
779 			goto err;
780 	}
781 
782 	for (addr = rx_grp->arg_macaddr; addr != NULL; addr = addr->aua_next) {
783 		if ((err = aggr_port_addmac(port, g_idx, addr->aua_addr)) != 0)
784 			goto err;
785 	}
786 
787 	for (i = 0; i < hw_rh_cnt; i++) {
788 		err = aggr_add_pseudo_rx_ring(port, rx_grp, hw_rh[i]);
789 		if (err != 0)
790 			goto err;
791 	}
792 
793 	mac_perim_exit(pmph);
794 	return (0);
795 
796 err:
797 	ASSERT(err != 0);
798 
799 	for (uint_t j = 0; j < i; j++)
800 		aggr_rem_pseudo_rx_ring(rx_grp, hw_rh[j]);
801 
802 	for (a = rx_grp->arg_macaddr; a != addr; a = a->aua_next)
803 		aggr_port_remmac(port, g_idx, a->aua_addr);
804 
805 	if (avp != NULL)
806 		avp = list_prev(&rx_grp->arg_vlans, avp);
807 
808 	for (; avp != NULL; avp = list_prev(&rx_grp->arg_vlans, avp)) {
809 		int err2;
810 
811 		if ((err2 = aggr_port_remvlan(port, g_idx, avp->av_vid)) != 0) {
812 			cmn_err(CE_WARN, "Failed to remove VLAN %u from port %s"
813 			    ": errno %d.", avp->av_vid,
814 			    mac_client_name(port->lp_mch), err2);
815 		}
816 	}
817 
818 	port->lp_hwghs[g_idx] = NULL;
819 	mac_perim_exit(pmph);
820 	return (err);
821 }
822 
823 /*
824  * Destroy the pseudo rings mapping to this port and remove all VLAN
825  * and unicast filters from this port. Even if there are no underlying
826  * HW rings we must still remove the unicast filters to take the port
827  * out of promisc mode.
828  */
829 static void
830 aggr_rem_pseudo_rx_group(aggr_port_t *port, aggr_pseudo_rx_group_t *rx_grp)
831 {
832 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP];
833 	aggr_unicst_addr_t	*addr;
834 	mac_perim_handle_t	pmph;
835 	uint_t			hw_rh_cnt;
836 	uint_t			g_idx = rx_grp->arg_index;
837 
838 	ASSERT(MAC_PERIM_HELD(port->lp_grp->lg_mh));
839 	ASSERT3U(g_idx, <, MAX_GROUPS_PER_PORT);
840 	ASSERT3P(rx_grp->arg_gh, !=, NULL);
841 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
842 
843 	hw_rh_cnt = mac_hwrings_idx_get(port->lp_mh, g_idx, NULL, hw_rh,
844 	    MAC_RING_TYPE_RX);
845 
846 	for (uint_t i = 0; i < hw_rh_cnt; i++)
847 		aggr_rem_pseudo_rx_ring(rx_grp, hw_rh[i]);
848 
849 	for (addr = rx_grp->arg_macaddr; addr != NULL; addr = addr->aua_next)
850 		aggr_port_remmac(port, g_idx, addr->aua_addr);
851 
852 	for (aggr_vlan_t *avp = list_head(&rx_grp->arg_vlans); avp != NULL;
853 	    avp = list_next(&rx_grp->arg_vlans, avp)) {
854 		int err;
855 
856 		if ((err = aggr_port_remvlan(port, g_idx, avp->av_vid)) != 0) {
857 			cmn_err(CE_WARN, "Failed to remove VLAN %u from port %s"
858 			    ": errno %d.", avp->av_vid,
859 			    mac_client_name(port->lp_mch), err);
860 		}
861 	}
862 
863 	port->lp_hwghs[g_idx] = NULL;
864 	mac_perim_exit(pmph);
865 }
866 
867 /*
868  * Add a pseudo TX ring for the given HW ring handle.
869  */
870 static int
871 aggr_add_pseudo_tx_ring(aggr_port_t *port,
872     aggr_pseudo_tx_group_t *tx_grp, mac_ring_handle_t hw_rh,
873     mac_ring_handle_t *pseudo_rh)
874 {
875 	aggr_pseudo_tx_ring_t	*ring;
876 	int			err;
877 	int			i;
878 
879 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
880 	for (i = 0; i < MAX_RINGS_PER_GROUP; i++) {
881 		ring = tx_grp->atg_rings + i;
882 		if (!(ring->atr_flags & MAC_PSEUDO_RING_INUSE))
883 			break;
884 	}
885 	/*
886 	 * No slot for this new TX ring.
887 	 */
888 	if (i == MAX_RINGS_PER_GROUP)
889 		return (ENOSPC);
890 	/*
891 	 * The following 4 statements needs to be done before
892 	 * calling mac_group_add_ring(). Otherwise it will
893 	 * result in an assertion failure in mac_init_ring().
894 	 */
895 	ring->atr_flags |= MAC_PSEUDO_RING_INUSE;
896 	ring->atr_hw_rh = hw_rh;
897 	ring->atr_port = port;
898 	tx_grp->atg_ring_cnt++;
899 
900 	/*
901 	 * The TX side has no concept of ring groups unlike RX groups.
902 	 * There is just a single group which stores all the TX rings.
903 	 * This group will be used to store aggr's pseudo TX rings.
904 	 */
905 	if ((err = mac_group_add_ring(tx_grp->atg_gh, i)) != 0) {
906 		ring->atr_flags &= ~MAC_PSEUDO_RING_INUSE;
907 		ring->atr_hw_rh = NULL;
908 		ring->atr_port = NULL;
909 		tx_grp->atg_ring_cnt--;
910 	} else {
911 		*pseudo_rh = mac_find_ring(tx_grp->atg_gh, i);
912 		if (hw_rh != NULL) {
913 			mac_hwring_setup(hw_rh, (mac_resource_handle_t)ring,
914 			    mac_find_ring(tx_grp->atg_gh, i));
915 		}
916 	}
917 
918 	return (err);
919 }
920 
921 /*
922  * Remove the pseudo TX ring of the given HW ring handle.
923  */
924 static void
925 aggr_rem_pseudo_tx_ring(aggr_pseudo_tx_group_t *tx_grp,
926     mac_ring_handle_t pseudo_hw_rh)
927 {
928 	aggr_pseudo_tx_ring_t	*ring;
929 	int			i;
930 
931 	for (i = 0; i < MAX_RINGS_PER_GROUP; i++) {
932 		ring = tx_grp->atg_rings + i;
933 		if (ring->atr_rh != pseudo_hw_rh)
934 			continue;
935 
936 		ASSERT(ring->atr_flags & MAC_PSEUDO_RING_INUSE);
937 		mac_group_rem_ring(tx_grp->atg_gh, pseudo_hw_rh);
938 		ring->atr_flags &= ~MAC_PSEUDO_RING_INUSE;
939 		mac_hwring_teardown(ring->atr_hw_rh);
940 		ring->atr_hw_rh = NULL;
941 		ring->atr_port = NULL;
942 		tx_grp->atg_ring_cnt--;
943 		break;
944 	}
945 }
946 
947 /*
948  * This function is called to create pseudo rings over hardware rings of
949  * the underlying device. There is a 1:1 mapping between the pseudo TX
950  * rings of the aggr and the hardware rings of the underlying port.
951  */
952 static int
953 aggr_add_pseudo_tx_group(aggr_port_t *port, aggr_pseudo_tx_group_t *tx_grp,
954     uint_t limit)
955 {
956 	aggr_grp_t		*grp = port->lp_grp;
957 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP], pseudo_rh;
958 	mac_perim_handle_t	pmph;
959 	int			hw_rh_cnt, i = 0, j;
960 	int			err = 0;
961 
962 	if (limit == 0)
963 		return (ENOSPC);
964 
965 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
966 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
967 
968 	/*
969 	 * Get the list the the underlying HW rings.
970 	 */
971 	hw_rh_cnt = mac_hwrings_get(port->lp_mch, NULL, hw_rh,
972 	    MAC_RING_TYPE_TX);
973 
974 	/*
975 	 * Even if the underlying NIC does not have TX rings, we
976 	 * still make a psuedo TX ring for that NIC with NULL as
977 	 * the ring handle.
978 	 */
979 	if (hw_rh_cnt == 0)
980 		port->lp_tx_ring_cnt = 1;
981 	else
982 		port->lp_tx_ring_cnt = MIN(hw_rh_cnt, limit);
983 
984 	port->lp_tx_ring_alloc = port->lp_tx_ring_cnt;
985 	port->lp_tx_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
986 	    port->lp_tx_ring_alloc), KM_SLEEP);
987 	port->lp_pseudo_tx_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
988 	    port->lp_tx_ring_alloc), KM_SLEEP);
989 
990 	if (hw_rh_cnt == 0) {
991 		if ((err = aggr_add_pseudo_tx_ring(port, tx_grp,
992 		    NULL, &pseudo_rh)) == 0) {
993 			port->lp_tx_rings[0] = NULL;
994 			port->lp_pseudo_tx_rings[0] = pseudo_rh;
995 		}
996 	} else {
997 		for (i = 0; err == 0 && i < port->lp_tx_ring_cnt; i++) {
998 			err = aggr_add_pseudo_tx_ring(port,
999 			    tx_grp, hw_rh[i], &pseudo_rh);
1000 			if (err != 0)
1001 				break;
1002 			port->lp_tx_rings[i] = hw_rh[i];
1003 			port->lp_pseudo_tx_rings[i] = pseudo_rh;
1004 		}
1005 	}
1006 
1007 	if (err != 0) {
1008 		if (hw_rh_cnt != 0) {
1009 			for (j = 0; j < i; j++) {
1010 				aggr_rem_pseudo_tx_ring(tx_grp,
1011 				    port->lp_pseudo_tx_rings[j]);
1012 			}
1013 		}
1014 		kmem_free(port->lp_tx_rings,
1015 		    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_alloc));
1016 		kmem_free(port->lp_pseudo_tx_rings,
1017 		    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_alloc));
1018 		port->lp_tx_ring_cnt = 0;
1019 		port->lp_tx_ring_alloc = 0;
1020 	} else {
1021 		port->lp_tx_grp_added = B_TRUE;
1022 		port->lp_tx_notify_mh = mac_client_tx_notify(port->lp_mch,
1023 		    aggr_tx_ring_update, port);
1024 	}
1025 	mac_perim_exit(pmph);
1026 	aggr_grp_update_default(grp);
1027 	return (err);
1028 }
1029 
1030 /*
1031  * This function is called by aggr to remove pseudo TX rings over the
1032  * HW rings of the underlying port.
1033  */
1034 static void
1035 aggr_rem_pseudo_tx_group(aggr_port_t *port, aggr_pseudo_tx_group_t *tx_grp)
1036 {
1037 	aggr_grp_t		*grp = port->lp_grp;
1038 	mac_perim_handle_t	pmph;
1039 	int			i;
1040 
1041 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1042 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
1043 
1044 	if (!port->lp_tx_grp_added)
1045 		goto done;
1046 
1047 	ASSERT(tx_grp->atg_gh != NULL);
1048 
1049 	for (i = 0; i < port->lp_tx_ring_cnt; i++)
1050 		aggr_rem_pseudo_tx_ring(tx_grp, port->lp_pseudo_tx_rings[i]);
1051 
1052 	kmem_free(port->lp_tx_rings,
1053 	    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_alloc));
1054 	kmem_free(port->lp_pseudo_tx_rings,
1055 	    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_alloc));
1056 
1057 	port->lp_tx_ring_cnt = 0;
1058 	(void) mac_client_tx_notify(port->lp_mch, NULL, port->lp_tx_notify_mh);
1059 	port->lp_tx_grp_added = B_FALSE;
1060 	aggr_grp_update_default(grp);
1061 done:
1062 	mac_perim_exit(pmph);
1063 }
1064 
1065 static int
1066 aggr_pseudo_disable_intr(mac_intr_handle_t ih)
1067 {
1068 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)ih;
1069 	return (mac_hwring_disable_intr(rr_ring->arr_hw_rh));
1070 }
1071 
1072 static int
1073 aggr_pseudo_enable_intr(mac_intr_handle_t ih)
1074 {
1075 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)ih;
1076 	return (mac_hwring_enable_intr(rr_ring->arr_hw_rh));
1077 }
1078 
1079 /*
1080  * Start the pseudo ring. Since the pseudo ring is just an abstraction
1081  * over an actual HW ring, the real task is to start the underlying HW
1082  * ring.
1083  */
1084 static int
1085 aggr_pseudo_start_rx_ring(mac_ring_driver_t arg, uint64_t mr_gen)
1086 {
1087 	int err;
1088 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)arg;
1089 
1090 	err = mac_hwring_start(rr_ring->arr_hw_rh);
1091 
1092 	if (err != 0)
1093 		return (err);
1094 
1095 	rr_ring->arr_gen = mr_gen;
1096 	return (err);
1097 }
1098 
1099 /*
1100  * Stop the pseudo ring. Since the pseudo ring is just an abstraction
1101  * over an actual HW ring, the real task is to stop the underlying HW
1102  * ring.
1103  */
1104 static void
1105 aggr_pseudo_stop_rx_ring(mac_ring_driver_t arg)
1106 {
1107 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)arg;
1108 
1109 	/*
1110 	 * The rings underlying the default group must stay up to
1111 	 * continue receiving LACP traffic. We would normally never
1112 	 * stop the default Rx rings because of the primary MAC
1113 	 * client; but aggr's primary MAC client doesn't call
1114 	 * mac_unicast_add() and thus mi_active is 0 when the last
1115 	 * non-primary client is deleted.
1116 	 */
1117 	if (rr_ring->arr_grp->arg_index != 0)
1118 		mac_hwring_stop(rr_ring->arr_hw_rh);
1119 }
1120 
1121 /*
1122  * Trim each port in a group to ensure it uses no more than tx_ring_limit
1123  * rings.
1124  */
1125 static void
1126 aggr_grp_balance_tx(aggr_grp_t *grp, uint_t tx_ring_limit)
1127 {
1128 	aggr_port_t *port;
1129 	mac_perim_handle_t mph;
1130 	uint_t i, tx_ring_cnt;
1131 
1132 	ASSERT(tx_ring_limit > 0);
1133 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1134 
1135 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1136 		mac_perim_enter_by_mh(port->lp_mh, &mph);
1137 
1138 		/*
1139 		 * Reduce the Tx ring count first to prevent rings being
1140 		 * used as they are removed.
1141 		 */
1142 		rw_enter(&grp->lg_tx_lock, RW_WRITER);
1143 		if (port->lp_tx_ring_cnt <= tx_ring_limit) {
1144 			rw_exit(&grp->lg_tx_lock);
1145 			mac_perim_exit(mph);
1146 			continue;
1147 		}
1148 
1149 		tx_ring_cnt = port->lp_tx_ring_cnt;
1150 		port->lp_tx_ring_cnt = tx_ring_limit;
1151 		rw_exit(&grp->lg_tx_lock);
1152 
1153 		for (i = tx_ring_cnt - 1; i >= tx_ring_limit; i--) {
1154 			aggr_rem_pseudo_tx_ring(&grp->lg_tx_group,
1155 			    port->lp_pseudo_tx_rings[i]);
1156 
1157 		}
1158 
1159 		mac_perim_exit(mph);
1160 	}
1161 }
1162 
1163 /*
1164  * Add one or more ports to an existing link aggregation group.
1165  */
1166 int
1167 aggr_grp_add_ports(datalink_id_t linkid, uint_t nports, boolean_t force,
1168     laioc_port_t *ports)
1169 {
1170 	int rc;
1171 	uint_t port_added = 0;
1172 	uint_t grp_added;
1173 	uint_t nports_high, tx_ring_limit;
1174 	aggr_grp_t *grp = NULL;
1175 	aggr_port_t *port;
1176 	boolean_t link_state_changed = B_FALSE;
1177 	mac_perim_handle_t mph, pmph;
1178 
1179 	/* Get the aggr corresponding to linkid. */
1180 	rw_enter(&aggr_grp_lock, RW_READER);
1181 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1182 	    (mod_hash_val_t *)&grp) != 0) {
1183 		rw_exit(&aggr_grp_lock);
1184 		return (ENOENT);
1185 	}
1186 	AGGR_GRP_REFHOLD(grp);
1187 
1188 	/*
1189 	 * Hold the perimeter so that the aggregation can't be destroyed.
1190 	 */
1191 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1192 	rw_exit(&aggr_grp_lock);
1193 
1194 	/*
1195 	 * Limit the number of Tx rings per port. When determining the
1196 	 * number of ports take into consideration the existing high
1197 	 * value, and what the new high value may be after this request.
1198 	 */
1199 	nports_high = MAX(grp->lg_nports_high, grp->lg_nports + nports);
1200 	tx_ring_limit = MAX_RINGS_PER_GROUP / nports_high;
1201 
1202 	if (tx_ring_limit == 0) {
1203 		rc = ENOSPC;
1204 		goto bail;
1205 	}
1206 
1207 	/*
1208 	 * Balance the Tx rings so each port has a fair share of rings.
1209 	 */
1210 	aggr_grp_balance_tx(grp, tx_ring_limit);
1211 
1212 	/* Add the specified ports to the aggr. */
1213 	for (uint_t i = 0; i < nports; i++) {
1214 		grp_added = 0;
1215 
1216 		if ((rc = aggr_grp_add_port(grp, ports[i].lp_linkid,
1217 		    force, &port)) != 0) {
1218 			goto bail;
1219 		}
1220 
1221 		ASSERT(port != NULL);
1222 		port_added++;
1223 
1224 		/* check capabilities */
1225 		if (!aggr_grp_capab_check(grp, port) ||
1226 		    !aggr_grp_sdu_check(grp, port) ||
1227 		    !aggr_grp_margin_check(grp, port)) {
1228 			rc = ENOTSUP;
1229 			goto bail;
1230 		}
1231 
1232 		/*
1233 		 * Create the pseudo ring for each HW ring of the underlying
1234 		 * port.
1235 		 */
1236 		rc = aggr_add_pseudo_tx_group(port, &grp->lg_tx_group,
1237 		    tx_ring_limit);
1238 		if (rc != 0)
1239 			goto bail;
1240 
1241 		for (uint_t j = 0; j < grp->lg_rx_group_count; j++) {
1242 			rc = aggr_add_pseudo_rx_group(port,
1243 			    &grp->lg_rx_groups[j]);
1244 
1245 			if (rc != 0)
1246 				goto bail;
1247 
1248 			grp_added++;
1249 		}
1250 
1251 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
1252 
1253 		/* set LACP mode */
1254 		aggr_port_lacp_set_mode(grp, port);
1255 
1256 		/* start port if group has already been started */
1257 		if (grp->lg_started) {
1258 			rc = aggr_port_start(port);
1259 			if (rc != 0) {
1260 				mac_perim_exit(pmph);
1261 				goto bail;
1262 			}
1263 
1264 			/*
1265 			 * Turn on the promiscuous mode over the port when it
1266 			 * is requested to be turned on to receive the
1267 			 * non-primary address over a port, or the promiscuous
1268 			 * mode is enabled over the aggr.
1269 			 */
1270 			if (grp->lg_promisc || port->lp_prom_addr != NULL) {
1271 				rc = aggr_port_promisc(port, B_TRUE);
1272 				if (rc != 0) {
1273 					mac_perim_exit(pmph);
1274 					goto bail;
1275 				}
1276 			}
1277 		}
1278 		mac_perim_exit(pmph);
1279 
1280 		/*
1281 		 * Attach each port if necessary.
1282 		 */
1283 		if (aggr_port_notify_link(grp, port))
1284 			link_state_changed = B_TRUE;
1285 
1286 		/*
1287 		 * Initialize the callback functions for this port.
1288 		 */
1289 		aggr_port_init_callbacks(port);
1290 	}
1291 
1292 	/* update the MAC address of the constituent ports */
1293 	if (aggr_grp_update_ports_mac(grp))
1294 		link_state_changed = B_TRUE;
1295 
1296 	if (link_state_changed)
1297 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1298 
1299 bail:
1300 	if (rc != 0) {
1301 		/* stop and remove ports that have been added */
1302 		for (uint_t i = 0; i < port_added; i++) {
1303 			uint_t grp_remove;
1304 
1305 			port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1306 			ASSERT(port != NULL);
1307 
1308 			if (grp->lg_started) {
1309 				mac_perim_enter_by_mh(port->lp_mh, &pmph);
1310 				(void) aggr_port_promisc(port, B_FALSE);
1311 				aggr_port_stop(port);
1312 				mac_perim_exit(pmph);
1313 			}
1314 
1315 			aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
1316 
1317 			/*
1318 			 * Only the last port could have a partial set
1319 			 * of groups added.
1320 			 */
1321 			grp_remove = (i + 1 == port_added) ? grp_added :
1322 			    grp->lg_rx_group_count;
1323 
1324 			for (uint_t j = 0; j < grp_remove; j++) {
1325 				aggr_rem_pseudo_rx_group(port,
1326 				    &grp->lg_rx_groups[j]);
1327 			}
1328 
1329 			(void) aggr_grp_rem_port(grp, port, NULL, NULL);
1330 		}
1331 	}
1332 
1333 	mac_perim_exit(mph);
1334 	AGGR_GRP_REFRELE(grp);
1335 	return (rc);
1336 }
1337 
1338 static int
1339 aggr_grp_modify_common(aggr_grp_t *grp, uint8_t update_mask, uint32_t policy,
1340     boolean_t mac_fixed, const uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode,
1341     aggr_lacp_timer_t lacp_timer)
1342 {
1343 	boolean_t mac_addr_changed = B_FALSE;
1344 	boolean_t link_state_changed = B_FALSE;
1345 	mac_perim_handle_t pmph;
1346 
1347 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1348 
1349 	/* validate fixed address if specified */
1350 	if ((update_mask & AGGR_MODIFY_MAC) && mac_fixed &&
1351 	    ((bcmp(aggr_zero_mac, mac_addr, ETHERADDRL) == 0) ||
1352 	    (mac_addr[0] & 0x01))) {
1353 		return (EINVAL);
1354 	}
1355 
1356 	/* update policy if requested */
1357 	if (update_mask & AGGR_MODIFY_POLICY)
1358 		aggr_send_update_policy(grp, policy);
1359 
1360 	/* update unicast MAC address if requested */
1361 	if (update_mask & AGGR_MODIFY_MAC) {
1362 		if (mac_fixed) {
1363 			/* user-supplied MAC address */
1364 			grp->lg_mac_addr_port = NULL;
1365 			if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) != 0) {
1366 				bcopy(mac_addr, grp->lg_addr, ETHERADDRL);
1367 				mac_addr_changed = B_TRUE;
1368 			}
1369 		} else if (grp->lg_addr_fixed) {
1370 			/* switch from user-supplied to automatic */
1371 			aggr_port_t *port = grp->lg_ports;
1372 
1373 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1374 			bcopy(port->lp_addr, grp->lg_addr, ETHERADDRL);
1375 			grp->lg_mac_addr_port = port;
1376 			mac_addr_changed = B_TRUE;
1377 			mac_perim_exit(pmph);
1378 		}
1379 		grp->lg_addr_fixed = mac_fixed;
1380 	}
1381 
1382 	if (mac_addr_changed)
1383 		link_state_changed = aggr_grp_update_ports_mac(grp);
1384 
1385 	if (update_mask & AGGR_MODIFY_LACP_MODE)
1386 		aggr_lacp_update_mode(grp, lacp_mode);
1387 
1388 	if (update_mask & AGGR_MODIFY_LACP_TIMER)
1389 		aggr_lacp_update_timer(grp, lacp_timer);
1390 
1391 	if (link_state_changed)
1392 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1393 
1394 	if (mac_addr_changed)
1395 		mac_unicst_update(grp->lg_mh, grp->lg_addr);
1396 
1397 	return (0);
1398 }
1399 
1400 /*
1401  * Update properties of an existing link aggregation group.
1402  */
1403 int
1404 aggr_grp_modify(datalink_id_t linkid, uint8_t update_mask, uint32_t policy,
1405     boolean_t mac_fixed, const uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode,
1406     aggr_lacp_timer_t lacp_timer)
1407 {
1408 	aggr_grp_t *grp = NULL;
1409 	mac_perim_handle_t mph;
1410 	int err;
1411 
1412 	/* get group corresponding to linkid */
1413 	rw_enter(&aggr_grp_lock, RW_READER);
1414 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1415 	    (mod_hash_val_t *)&grp) != 0) {
1416 		rw_exit(&aggr_grp_lock);
1417 		return (ENOENT);
1418 	}
1419 	AGGR_GRP_REFHOLD(grp);
1420 
1421 	/*
1422 	 * Hold the perimeter so that the aggregation won't be destroyed.
1423 	 */
1424 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1425 	rw_exit(&aggr_grp_lock);
1426 
1427 	err = aggr_grp_modify_common(grp, update_mask, policy, mac_fixed,
1428 	    mac_addr, lacp_mode, lacp_timer);
1429 
1430 	mac_perim_exit(mph);
1431 	AGGR_GRP_REFRELE(grp);
1432 	return (err);
1433 }
1434 
1435 /*
1436  * Create a new link aggregation group upon request from administrator.
1437  * Returns 0 on success, an errno on failure.
1438  */
1439 int
1440 aggr_grp_create(datalink_id_t linkid, uint32_t key, uint_t nports,
1441     laioc_port_t *ports, uint32_t policy, boolean_t mac_fixed, boolean_t force,
1442     uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode, aggr_lacp_timer_t lacp_timer,
1443     cred_t *credp)
1444 {
1445 	aggr_grp_t *grp = NULL;
1446 	aggr_port_t *port;
1447 	aggr_port_t *last_attached = NULL;
1448 	mac_register_t *mac;
1449 	boolean_t link_state_changed;
1450 	mac_perim_handle_t mph, pmph;
1451 	datalink_id_t tempid;
1452 	boolean_t mac_registered = B_FALSE;
1453 	uint_t tx_ring_limit;
1454 	int err;
1455 	int i, j;
1456 	kt_did_t tid = 0;
1457 
1458 	/* need at least one port */
1459 	if (nports == 0)
1460 		return (EINVAL);
1461 
1462 	rw_enter(&aggr_grp_lock, RW_WRITER);
1463 
1464 	/* does a group with the same linkid already exist? */
1465 	err = mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1466 	    (mod_hash_val_t *)&grp);
1467 	if (err == 0) {
1468 		rw_exit(&aggr_grp_lock);
1469 		return (EEXIST);
1470 	}
1471 
1472 	grp = kmem_cache_alloc(aggr_grp_cache, KM_SLEEP);
1473 
1474 	grp->lg_refs = 1;
1475 	grp->lg_closing = B_FALSE;
1476 	grp->lg_force = force;
1477 	grp->lg_linkid = linkid;
1478 	grp->lg_zoneid = crgetzoneid(credp);
1479 	grp->lg_ifspeed = 0;
1480 	grp->lg_link_state = LINK_STATE_UNKNOWN;
1481 	grp->lg_link_duplex = LINK_DUPLEX_UNKNOWN;
1482 	grp->lg_started = B_FALSE;
1483 	grp->lg_promisc = B_FALSE;
1484 	grp->lg_lacp_done = B_FALSE;
1485 	grp->lg_tx_notify_done = B_FALSE;
1486 	grp->lg_lacp_head = grp->lg_lacp_tail = NULL;
1487 	grp->lg_lacp_rx_thread = thread_create(NULL, 0,
1488 	    aggr_lacp_rx_thread, grp, 0, &p0, TS_RUN, minclsyspri);
1489 	grp->lg_tx_notify_thread = thread_create(NULL, 0,
1490 	    aggr_tx_notify_thread, grp, 0, &p0, TS_RUN, minclsyspri);
1491 	grp->lg_tx_blocked_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
1492 	    MAX_RINGS_PER_GROUP), KM_SLEEP);
1493 	grp->lg_tx_blocked_cnt = 0;
1494 	bzero(&grp->lg_rx_groups,
1495 	    sizeof (aggr_pseudo_rx_group_t) * MAX_GROUPS_PER_PORT);
1496 	bzero(&grp->lg_tx_group, sizeof (aggr_pseudo_tx_group_t));
1497 	aggr_lacp_init_grp(grp);
1498 
1499 	/* add MAC ports to group */
1500 	grp->lg_ports = NULL;
1501 	grp->lg_nports = 0;
1502 	grp->lg_nattached_ports = 0;
1503 	grp->lg_ntx_ports = 0;
1504 
1505 	/*
1506 	 * If key is not specified by the user, allocate the key.
1507 	 */
1508 	if ((key == 0) && ((key = (uint32_t)id_alloc(key_ids)) == 0)) {
1509 		err = ENOMEM;
1510 		goto bail;
1511 	}
1512 	grp->lg_key = key;
1513 
1514 	for (i = 0; i < nports; i++) {
1515 		err = aggr_grp_add_port(grp, ports[i].lp_linkid, force, &port);
1516 		if (err != 0)
1517 			goto bail;
1518 	}
1519 
1520 	grp->lg_rx_group_count = 1;
1521 
1522 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1523 		uint_t num_rgroups;
1524 
1525 		mac_perim_enter_by_mh(port->lp_mh, &mph);
1526 		num_rgroups = mac_get_num_rx_groups(port->lp_mh);
1527 		mac_perim_exit(mph);
1528 
1529 		/*
1530 		 * Utilize all the groups in a port. If some ports
1531 		 * have less groups than others, then traffic destined
1532 		 * for the same unicast address may be HW classified
1533 		 * on some ports but SW classified by aggr when
1534 		 * arriving on other ports.
1535 		 */
1536 		grp->lg_rx_group_count = MAX(grp->lg_rx_group_count,
1537 		    num_rgroups);
1538 	}
1539 
1540 	/*
1541 	 * There could be cases where the hardware provides more
1542 	 * groups than aggr can support. Make sure we never go above
1543 	 * the max aggr can support.
1544 	 */
1545 	grp->lg_rx_group_count = MIN(grp->lg_rx_group_count,
1546 	    MAX_GROUPS_PER_PORT);
1547 
1548 	ASSERT3U(grp->lg_rx_group_count, >, 0);
1549 	for (i = 0; i < MAX_GROUPS_PER_PORT; i++) {
1550 		grp->lg_rx_groups[i].arg_index = i;
1551 		grp->lg_rx_groups[i].arg_untagged = 0;
1552 		list_create(&(grp->lg_rx_groups[i].arg_vlans),
1553 		    sizeof (aggr_vlan_t), offsetof(aggr_vlan_t, av_link));
1554 	}
1555 
1556 	/*
1557 	 * If no explicit MAC address was specified by the administrator,
1558 	 * set it to the MAC address of the first port.
1559 	 */
1560 	grp->lg_addr_fixed = mac_fixed;
1561 	if (grp->lg_addr_fixed) {
1562 		/* validate specified address */
1563 		if (bcmp(aggr_zero_mac, mac_addr, ETHERADDRL) == 0) {
1564 			err = EINVAL;
1565 			goto bail;
1566 		}
1567 		bcopy(mac_addr, grp->lg_addr, ETHERADDRL);
1568 	} else {
1569 		bcopy(grp->lg_ports->lp_addr, grp->lg_addr, ETHERADDRL);
1570 		grp->lg_mac_addr_port = grp->lg_ports;
1571 	}
1572 
1573 	/* Set the initial group capabilities. */
1574 	aggr_grp_capab_set(grp);
1575 
1576 	if ((mac = mac_alloc(MAC_VERSION)) == NULL) {
1577 		err = ENOMEM;
1578 		goto bail;
1579 	}
1580 	mac->m_type_ident = MAC_PLUGIN_IDENT_ETHER;
1581 	mac->m_driver = grp;
1582 	mac->m_dip = aggr_dip;
1583 	mac->m_instance = grp->lg_key > AGGR_MAX_KEY ? (uint_t)-1 : grp->lg_key;
1584 	mac->m_src_addr = grp->lg_addr;
1585 	mac->m_callbacks = &aggr_m_callbacks;
1586 	mac->m_min_sdu = 0;
1587 	mac->m_max_sdu = grp->lg_max_sdu = aggr_grp_max_sdu(grp);
1588 	mac->m_margin = aggr_grp_max_margin(grp);
1589 	mac->m_v12n = MAC_VIRT_LEVEL1;
1590 	err = mac_register(mac, &grp->lg_mh);
1591 	mac_free(mac);
1592 	if (err != 0)
1593 		goto bail;
1594 
1595 	err = dls_devnet_create(grp->lg_mh, grp->lg_linkid, crgetzoneid(credp));
1596 	if (err != 0) {
1597 		(void) mac_unregister(grp->lg_mh);
1598 		grp->lg_mh = NULL;
1599 		goto bail;
1600 	}
1601 
1602 	mac_registered = B_TRUE;
1603 
1604 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1605 
1606 	/*
1607 	 * Update the MAC address of the constituent ports.
1608 	 * None of the port is attached at this time, the link state of the
1609 	 * aggregation will not change.
1610 	 *
1611 	 * All ports take on the primary MAC address of the aggr
1612 	 * (lg_aggr). At this point, none of the ports are attached;
1613 	 * thus the link state of the aggregation will not change.
1614 	 */
1615 	link_state_changed = aggr_grp_update_ports_mac(grp);
1616 	ASSERT(!link_state_changed);
1617 
1618 	/* Update outbound load balancing policy. */
1619 	aggr_send_update_policy(grp, policy);
1620 
1621 	/* Set LACP mode. */
1622 	aggr_lacp_set_mode(grp, lacp_mode, lacp_timer);
1623 
1624 	/*
1625 	 * The pseudo Tx group holds a maximum of MAX_RINGS_PER_GROUP
1626 	 * rings, when all the Tx rings of all the ports are accumulated
1627 	 * it is conceivable this limit is exceeded. We try and prevent
1628 	 * this by limiting the number of rings an individual port will use.
1629 	 *
1630 	 * - When an aggr is first created, we will not let an
1631 	 *   individual port use more than MAX_RINGS_PER_GROUP/nports
1632 	 *   rings.
1633 	 * - As ports are added to an existing aggr, each of the
1634 	 *   ports will not use more than MAX_RINGS_PER_GROUP/nports_high.
1635 	 *   Where nports_high is the highest number of ports the aggr has
1636 	 *   held (including any ports being added). This may involve
1637 	 *   trimming rings from existing ports.
1638 	 */
1639 
1640 	/* Leave room for 4 ports */
1641 	tx_ring_limit = MAX_RINGS_PER_GROUP / MAX(4, nports);
1642 
1643 	/*
1644 	 * Attach each port if necessary.
1645 	 */
1646 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1647 		/*
1648 		 * Create the pseudo ring for each HW ring of the
1649 		 * underlying port. Note that this is done after the
1650 		 * aggr registers its MAC.
1651 		 */
1652 		err = aggr_add_pseudo_tx_group(port, &grp->lg_tx_group,
1653 		    tx_ring_limit);
1654 
1655 		if (err != 0) {
1656 			mac_perim_exit(mph);
1657 			goto bail;
1658 		}
1659 
1660 		for (i = 0; i < grp->lg_rx_group_count; i++) {
1661 			err = aggr_add_pseudo_rx_group(port,
1662 			    &grp->lg_rx_groups[i]);
1663 
1664 			if (err != 0) {
1665 				/*
1666 				 * Undo what we have added for the current
1667 				 * port.
1668 				 */
1669 				aggr_rem_pseudo_tx_group(port,
1670 				    &grp->lg_tx_group);
1671 
1672 				for (j = 0; j < i; j++) {
1673 					aggr_rem_pseudo_rx_group(port,
1674 					    &grp->lg_rx_groups[j]);
1675 				}
1676 
1677 				mac_perim_exit(mph);
1678 				goto bail;
1679 			}
1680 		}
1681 
1682 		if (aggr_port_notify_link(grp, port))
1683 			link_state_changed = B_TRUE;
1684 
1685 		/*
1686 		 * Initialize the callback functions for this port.
1687 		 */
1688 		aggr_port_init_callbacks(port);
1689 
1690 		last_attached = port;
1691 	}
1692 
1693 	if (link_state_changed)
1694 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1695 
1696 	/* add new group to hash table */
1697 	err = mod_hash_insert(aggr_grp_hash, GRP_HASH_KEY(linkid),
1698 	    (mod_hash_val_t)grp);
1699 	ASSERT(err == 0);
1700 	aggr_grp_cnt++;
1701 
1702 	mac_perim_exit(mph);
1703 	rw_exit(&aggr_grp_lock);
1704 	return (0);
1705 
1706 bail:
1707 	grp->lg_closing = B_TRUE;
1708 
1709 	/*
1710 	 * Inform the lacp_rx thread to exit.
1711 	 */
1712 	mutex_enter(&grp->lg_lacp_lock);
1713 	grp->lg_lacp_done = B_TRUE;
1714 	cv_signal(&grp->lg_lacp_cv);
1715 	while (grp->lg_lacp_rx_thread != NULL)
1716 		cv_wait(&grp->lg_lacp_cv, &grp->lg_lacp_lock);
1717 	mutex_exit(&grp->lg_lacp_lock);
1718 	/*
1719 	 * Inform the tx_notify thread to exit.
1720 	 */
1721 	mutex_enter(&grp->lg_tx_flowctl_lock);
1722 	if (grp->lg_tx_notify_thread != NULL) {
1723 		tid = grp->lg_tx_notify_thread->t_did;
1724 		grp->lg_tx_notify_done = B_TRUE;
1725 		cv_signal(&grp->lg_tx_flowctl_cv);
1726 	}
1727 	mutex_exit(&grp->lg_tx_flowctl_lock);
1728 	if (tid != 0)
1729 		thread_join(tid);
1730 
1731 	if (mac_registered) {
1732 		(void) dls_devnet_destroy(grp->lg_mh, &tempid, B_TRUE);
1733 		(void) mac_disable(grp->lg_mh);
1734 
1735 		if (last_attached != NULL) {
1736 			/*
1737 			 * Detach and clean up ports added.
1738 			 */
1739 			mac_perim_enter_by_mh(grp->lg_mh, &mph);
1740 
1741 			for (port = grp->lg_ports; ; port = port->lp_next) {
1742 				mac_perim_enter_by_mh(port->lp_mh, &pmph);
1743 				(void) aggr_grp_detach_port(grp, port);
1744 				mac_perim_exit(pmph);
1745 
1746 				aggr_rem_pseudo_tx_group(port,
1747 				    &grp->lg_tx_group);
1748 
1749 				for (i = 0; i < grp->lg_rx_group_count; i++) {
1750 					aggr_rem_pseudo_rx_group(port,
1751 					    &grp->lg_rx_groups[i]);
1752 				}
1753 				if (port == last_attached)
1754 					break;
1755 			}
1756 
1757 			mac_perim_exit(mph);
1758 		}
1759 
1760 		(void) mac_unregister(grp->lg_mh);
1761 	}
1762 
1763 	port = grp->lg_ports;
1764 	while (port != NULL) {
1765 		aggr_port_t *cport;
1766 
1767 		cport = port->lp_next;
1768 		aggr_port_delete(port);
1769 		port = cport;
1770 	}
1771 
1772 	kmem_free(grp->lg_tx_blocked_rings,
1773 	    (sizeof (mac_ring_handle_t *) * MAX_RINGS_PER_GROUP));
1774 	rw_exit(&aggr_grp_lock);
1775 	AGGR_GRP_REFRELE(grp);
1776 	return (err);
1777 }
1778 
1779 /*
1780  * Return a pointer to the member of a group with specified linkid.
1781  */
1782 static aggr_port_t *
1783 aggr_grp_port_lookup(aggr_grp_t *grp, datalink_id_t linkid)
1784 {
1785 	aggr_port_t *port;
1786 
1787 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1788 
1789 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1790 		if (port->lp_linkid == linkid)
1791 			break;
1792 	}
1793 
1794 	return (port);
1795 }
1796 
1797 /*
1798  * Stop, detach and remove a port from a link aggregation group.
1799  */
1800 static int
1801 aggr_grp_rem_port(aggr_grp_t *grp, aggr_port_t *port,
1802     boolean_t *mac_addr_changedp, boolean_t *link_state_changedp)
1803 {
1804 	int rc = 0;
1805 	aggr_port_t **pport;
1806 	boolean_t mac_addr_changed = B_FALSE;
1807 	boolean_t link_state_changed = B_FALSE;
1808 	mac_perim_handle_t mph;
1809 	uint64_t val;
1810 	uint_t i;
1811 	uint_t stat;
1812 
1813 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1814 	ASSERT(grp->lg_nports > 1);
1815 	ASSERT(!grp->lg_closing);
1816 
1817 	/* unlink port */
1818 	for (pport = &grp->lg_ports; *pport != port;
1819 	    pport = &(*pport)->lp_next) {
1820 		if (*pport == NULL) {
1821 			rc = ENOENT;
1822 			goto done;
1823 		}
1824 	}
1825 	*pport = port->lp_next;
1826 
1827 	mac_perim_enter_by_mh(port->lp_mh, &mph);
1828 
1829 	/*
1830 	 * If the MAC address of the port being removed was assigned
1831 	 * to the group, update the group MAC address
1832 	 * using the MAC address of a different port.
1833 	 */
1834 	if (!grp->lg_addr_fixed && grp->lg_mac_addr_port == port) {
1835 		/*
1836 		 * Set the MAC address of the group to the
1837 		 * MAC address of its first port.
1838 		 */
1839 		bcopy(grp->lg_ports->lp_addr, grp->lg_addr, ETHERADDRL);
1840 		grp->lg_mac_addr_port = grp->lg_ports;
1841 		mac_addr_changed = B_TRUE;
1842 	}
1843 
1844 	link_state_changed = aggr_grp_detach_port(grp, port);
1845 
1846 	/*
1847 	 * Add the counter statistics of the ports while it was aggregated
1848 	 * to the group's residual statistics.  This is done by obtaining
1849 	 * the current counter from the underlying MAC then subtracting the
1850 	 * value of the counter at the moment it was added to the
1851 	 * aggregation.
1852 	 */
1853 	for (i = 0; i < MAC_NSTAT; i++) {
1854 		stat = i + MAC_STAT_MIN;
1855 		if (!MAC_STAT_ISACOUNTER(stat))
1856 			continue;
1857 		val = aggr_port_stat(port, stat);
1858 		val -= port->lp_stat[i];
1859 		mutex_enter(&grp->lg_stat_lock);
1860 		grp->lg_stat[i] += val;
1861 		mutex_exit(&grp->lg_stat_lock);
1862 	}
1863 	for (i = 0; i < ETHER_NSTAT; i++) {
1864 		stat = i + MACTYPE_STAT_MIN;
1865 		if (!ETHER_STAT_ISACOUNTER(stat))
1866 			continue;
1867 		val = aggr_port_stat(port, stat);
1868 		val -= port->lp_ether_stat[i];
1869 		mutex_enter(&grp->lg_stat_lock);
1870 		grp->lg_ether_stat[i] += val;
1871 		mutex_exit(&grp->lg_stat_lock);
1872 	}
1873 
1874 	grp->lg_nports--;
1875 	mac_perim_exit(mph);
1876 
1877 	aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
1878 	aggr_port_delete(port);
1879 
1880 	/*
1881 	 * If the group MAC address has changed, update the MAC address of
1882 	 * the remaining constituent ports according to the new MAC
1883 	 * address of the group.
1884 	 */
1885 	if (mac_addr_changed && aggr_grp_update_ports_mac(grp))
1886 		link_state_changed = B_TRUE;
1887 
1888 done:
1889 	if (mac_addr_changedp != NULL)
1890 		*mac_addr_changedp = mac_addr_changed;
1891 	if (link_state_changedp != NULL)
1892 		*link_state_changedp = link_state_changed;
1893 
1894 	return (rc);
1895 }
1896 
1897 /*
1898  * Remove one or more ports from an existing link aggregation group.
1899  */
1900 int
1901 aggr_grp_rem_ports(datalink_id_t linkid, uint_t nports, laioc_port_t *ports)
1902 {
1903 	int rc = 0;
1904 	uint_t i;
1905 	aggr_grp_t *grp = NULL;
1906 	aggr_port_t *port;
1907 	boolean_t mac_addr_update = B_FALSE, mac_addr_changed;
1908 	boolean_t link_state_update = B_FALSE, link_state_changed;
1909 	mac_perim_handle_t mph, pmph;
1910 
1911 	/* get group corresponding to linkid */
1912 	rw_enter(&aggr_grp_lock, RW_READER);
1913 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1914 	    (mod_hash_val_t *)&grp) != 0) {
1915 		rw_exit(&aggr_grp_lock);
1916 		return (ENOENT);
1917 	}
1918 	AGGR_GRP_REFHOLD(grp);
1919 
1920 	/*
1921 	 * Hold the perimeter so that the aggregation won't be destroyed.
1922 	 */
1923 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1924 	rw_exit(&aggr_grp_lock);
1925 
1926 	/* we need to keep at least one port per group */
1927 	if (nports >= grp->lg_nports) {
1928 		rc = EINVAL;
1929 		goto bail;
1930 	}
1931 
1932 	/* first verify that all the groups are valid */
1933 	for (i = 0; i < nports; i++) {
1934 		if (aggr_grp_port_lookup(grp, ports[i].lp_linkid) == NULL) {
1935 			/* port not found */
1936 			rc = ENOENT;
1937 			goto bail;
1938 		}
1939 	}
1940 
1941 	/* clear the promiscous mode for the specified ports */
1942 	for (i = 0; i < nports && rc == 0; i++) {
1943 		/* lookup port */
1944 		port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1945 		ASSERT(port != NULL);
1946 
1947 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
1948 		rc = aggr_port_promisc(port, B_FALSE);
1949 		mac_perim_exit(pmph);
1950 	}
1951 	if (rc != 0) {
1952 		for (i = 0; i < nports; i++) {
1953 			port = aggr_grp_port_lookup(grp,
1954 			    ports[i].lp_linkid);
1955 			ASSERT(port != NULL);
1956 
1957 			/*
1958 			 * Turn the promiscuous mode back on if it is required
1959 			 * to receive the non-primary address over a port, or
1960 			 * the promiscous mode is enabled over the aggr.
1961 			 */
1962 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1963 			if (port->lp_started && (grp->lg_promisc ||
1964 			    port->lp_prom_addr != NULL)) {
1965 				(void) aggr_port_promisc(port, B_TRUE);
1966 			}
1967 			mac_perim_exit(pmph);
1968 		}
1969 		goto bail;
1970 	}
1971 
1972 	/* remove the specified ports from group */
1973 	for (i = 0; i < nports; i++) {
1974 		/* lookup port */
1975 		port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1976 		ASSERT(port != NULL);
1977 
1978 		/* stop port if group has already been started */
1979 		if (grp->lg_started) {
1980 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1981 			aggr_port_stop(port);
1982 			mac_perim_exit(pmph);
1983 		}
1984 
1985 		/*
1986 		 * aggr_rem_pseudo_tx_group() is not called here. Instead
1987 		 * it is called from inside aggr_grp_rem_port() after the
1988 		 * port has been detached. The reason is that
1989 		 * aggr_rem_pseudo_tx_group() removes one ring at a time
1990 		 * and if there is still traffic going on, then there
1991 		 * is the possibility of aggr_find_tx_ring() returning a
1992 		 * removed ring for transmission. Once the port has been
1993 		 * detached, that port will not be used and
1994 		 * aggr_find_tx_ring() will not return any rings
1995 		 * belonging to it.
1996 		 */
1997 		for (uint_t j = 0; j < grp->lg_rx_group_count; j++)
1998 			aggr_rem_pseudo_rx_group(port, &grp->lg_rx_groups[j]);
1999 
2000 		/* remove port from group */
2001 		rc = aggr_grp_rem_port(grp, port, &mac_addr_changed,
2002 		    &link_state_changed);
2003 		ASSERT(rc == 0);
2004 		mac_addr_update = mac_addr_update || mac_addr_changed;
2005 		link_state_update = link_state_update || link_state_changed;
2006 	}
2007 
2008 bail:
2009 	if (mac_addr_update)
2010 		mac_unicst_update(grp->lg_mh, grp->lg_addr);
2011 	if (link_state_update)
2012 		mac_link_update(grp->lg_mh, grp->lg_link_state);
2013 
2014 	mac_perim_exit(mph);
2015 	AGGR_GRP_REFRELE(grp);
2016 
2017 	return (rc);
2018 }
2019 
2020 int
2021 aggr_grp_delete(datalink_id_t linkid, cred_t *cred)
2022 {
2023 	aggr_grp_t *grp = NULL;
2024 	aggr_port_t *port, *cport;
2025 	datalink_id_t tmpid;
2026 	mod_hash_val_t val;
2027 	mac_perim_handle_t mph, pmph;
2028 	int err;
2029 	kt_did_t tid = 0;
2030 
2031 	rw_enter(&aggr_grp_lock, RW_WRITER);
2032 
2033 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
2034 	    (mod_hash_val_t *)&grp) != 0) {
2035 		rw_exit(&aggr_grp_lock);
2036 		return (ENOENT);
2037 	}
2038 
2039 	/*
2040 	 * Note that dls_devnet_destroy() must be called before lg_lock is
2041 	 * held. Otherwise, it will deadlock if another thread is in
2042 	 * aggr_m_stat() and thus has a kstat_hold() on the kstats that
2043 	 * dls_devnet_destroy() needs to delete.
2044 	 */
2045 	if ((err = dls_devnet_destroy(grp->lg_mh, &tmpid, B_TRUE)) != 0) {
2046 		rw_exit(&aggr_grp_lock);
2047 		return (err);
2048 	}
2049 	ASSERT(linkid == tmpid);
2050 
2051 	/*
2052 	 * Unregister from the MAC service module. Since this can
2053 	 * fail if a client hasn't closed the MAC port, we gracefully
2054 	 * fail the operation.
2055 	 */
2056 	if ((err = mac_disable(grp->lg_mh)) != 0) {
2057 		(void) dls_devnet_create(grp->lg_mh, linkid, crgetzoneid(cred));
2058 		rw_exit(&aggr_grp_lock);
2059 		return (err);
2060 	}
2061 	(void) mod_hash_remove(aggr_grp_hash, GRP_HASH_KEY(linkid), &val);
2062 	ASSERT(grp == (aggr_grp_t *)val);
2063 
2064 	ASSERT(aggr_grp_cnt > 0);
2065 	aggr_grp_cnt--;
2066 	rw_exit(&aggr_grp_lock);
2067 
2068 	/*
2069 	 * Inform the lacp_rx thread to exit.
2070 	 */
2071 	mutex_enter(&grp->lg_lacp_lock);
2072 	grp->lg_lacp_done = B_TRUE;
2073 	cv_signal(&grp->lg_lacp_cv);
2074 	while (grp->lg_lacp_rx_thread != NULL)
2075 		cv_wait(&grp->lg_lacp_cv, &grp->lg_lacp_lock);
2076 	mutex_exit(&grp->lg_lacp_lock);
2077 	/*
2078 	 * Inform the tx_notify_thread to exit.
2079 	 */
2080 	mutex_enter(&grp->lg_tx_flowctl_lock);
2081 	if (grp->lg_tx_notify_thread != NULL) {
2082 		tid = grp->lg_tx_notify_thread->t_did;
2083 		grp->lg_tx_notify_done = B_TRUE;
2084 		cv_signal(&grp->lg_tx_flowctl_cv);
2085 	}
2086 	mutex_exit(&grp->lg_tx_flowctl_lock);
2087 	if (tid != 0)
2088 		thread_join(tid);
2089 
2090 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2091 
2092 	grp->lg_closing = B_TRUE;
2093 	/* detach and free MAC ports associated with group */
2094 	port = grp->lg_ports;
2095 	while (port != NULL) {
2096 		cport = port->lp_next;
2097 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2098 		if (grp->lg_started)
2099 			aggr_port_stop(port);
2100 		(void) aggr_grp_detach_port(grp, port);
2101 		mac_perim_exit(pmph);
2102 		aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
2103 		for (uint_t i = 0; i < grp->lg_rx_group_count; i++)
2104 			aggr_rem_pseudo_rx_group(port, &grp->lg_rx_groups[i]);
2105 		aggr_port_delete(port);
2106 		port = cport;
2107 	}
2108 
2109 	mac_perim_exit(mph);
2110 
2111 	kmem_free(grp->lg_tx_blocked_rings,
2112 	    (sizeof (mac_ring_handle_t *) * MAX_RINGS_PER_GROUP));
2113 	/*
2114 	 * Wait for the port's lacp timer thread and its notification callback
2115 	 * to exit before calling mac_unregister() since both needs to access
2116 	 * the mac perimeter of the grp.
2117 	 */
2118 	aggr_grp_port_wait(grp);
2119 
2120 	VERIFY(mac_unregister(grp->lg_mh) == 0);
2121 	grp->lg_mh = NULL;
2122 
2123 	for (uint_t i = 0; i < MAX_GROUPS_PER_PORT; i++) {
2124 		list_destroy(&(grp->lg_rx_groups[i].arg_vlans));
2125 	}
2126 
2127 	AGGR_GRP_REFRELE(grp);
2128 	return (0);
2129 }
2130 
2131 void
2132 aggr_grp_free(aggr_grp_t *grp)
2133 {
2134 	ASSERT(grp->lg_refs == 0);
2135 	ASSERT(grp->lg_port_ref == 0);
2136 	if (grp->lg_key > AGGR_MAX_KEY) {
2137 		id_free(key_ids, grp->lg_key);
2138 		grp->lg_key = 0;
2139 	}
2140 	kmem_cache_free(aggr_grp_cache, grp);
2141 }
2142 
2143 int
2144 aggr_grp_info(datalink_id_t linkid, void *fn_arg,
2145     aggr_grp_info_new_grp_fn_t new_grp_fn,
2146     aggr_grp_info_new_port_fn_t new_port_fn, cred_t *cred)
2147 {
2148 	aggr_grp_t	*grp;
2149 	aggr_port_t	*port;
2150 	mac_perim_handle_t mph, pmph;
2151 	int		rc = 0;
2152 
2153 	/*
2154 	 * Make sure that the aggregation link is visible from the caller's
2155 	 * zone.
2156 	 */
2157 	if (!dls_devnet_islinkvisible(linkid, crgetzoneid(cred)))
2158 		return (ENOENT);
2159 
2160 	rw_enter(&aggr_grp_lock, RW_READER);
2161 
2162 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
2163 	    (mod_hash_val_t *)&grp) != 0) {
2164 		rw_exit(&aggr_grp_lock);
2165 		return (ENOENT);
2166 	}
2167 	AGGR_GRP_REFHOLD(grp);
2168 
2169 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2170 	rw_exit(&aggr_grp_lock);
2171 
2172 	rc = new_grp_fn(fn_arg, grp->lg_linkid,
2173 	    (grp->lg_key > AGGR_MAX_KEY) ? 0 : grp->lg_key, grp->lg_addr,
2174 	    grp->lg_addr_fixed, grp->lg_force, grp->lg_tx_policy,
2175 	    grp->lg_nports, grp->lg_lacp_mode, grp->aggr.PeriodicTimer);
2176 
2177 	if (rc != 0)
2178 		goto bail;
2179 
2180 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2181 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2182 		rc = new_port_fn(fn_arg, port->lp_linkid, port->lp_addr,
2183 		    port->lp_state, &port->lp_lacp.ActorOperPortState);
2184 		mac_perim_exit(pmph);
2185 
2186 		if (rc != 0)
2187 			goto bail;
2188 	}
2189 
2190 bail:
2191 	mac_perim_exit(mph);
2192 	AGGR_GRP_REFRELE(grp);
2193 	return (rc);
2194 }
2195 
2196 /*ARGSUSED*/
2197 static void
2198 aggr_m_ioctl(void *arg, queue_t *q, mblk_t *mp)
2199 {
2200 	miocnak(q, mp, 0, ENOTSUP);
2201 }
2202 
2203 static int
2204 aggr_grp_stat(aggr_grp_t *grp, uint_t stat, uint64_t *val)
2205 {
2206 	aggr_port_t	*port;
2207 	uint_t		stat_index;
2208 
2209 	ASSERT(MUTEX_HELD(&grp->lg_stat_lock));
2210 
2211 	/* We only aggregate counter statistics. */
2212 	if (IS_MAC_STAT(stat) && !MAC_STAT_ISACOUNTER(stat) ||
2213 	    IS_MACTYPE_STAT(stat) && !ETHER_STAT_ISACOUNTER(stat)) {
2214 		return (ENOTSUP);
2215 	}
2216 
2217 	/*
2218 	 * Counter statistics for a group are computed by aggregating the
2219 	 * counters of the members MACs while they were aggregated, plus
2220 	 * the residual counter of the group itself, which is updated each
2221 	 * time a MAC is removed from the group.
2222 	 */
2223 	*val = 0;
2224 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2225 		/* actual port statistic */
2226 		*val += aggr_port_stat(port, stat);
2227 		/*
2228 		 * minus the port stat when it was added, plus any residual
2229 		 * amount for the group.
2230 		 */
2231 		if (IS_MAC_STAT(stat)) {
2232 			stat_index = stat - MAC_STAT_MIN;
2233 			*val -= port->lp_stat[stat_index];
2234 			*val += grp->lg_stat[stat_index];
2235 		} else if (IS_MACTYPE_STAT(stat)) {
2236 			stat_index = stat - MACTYPE_STAT_MIN;
2237 			*val -= port->lp_ether_stat[stat_index];
2238 			*val += grp->lg_ether_stat[stat_index];
2239 		}
2240 	}
2241 	return (0);
2242 }
2243 
2244 int
2245 aggr_rx_ring_stat(mac_ring_driver_t rdriver, uint_t stat, uint64_t *val)
2246 {
2247 	aggr_pseudo_rx_ring_t   *rx_ring = (aggr_pseudo_rx_ring_t *)rdriver;
2248 
2249 	if (rx_ring->arr_hw_rh != NULL) {
2250 		*val = mac_pseudo_rx_ring_stat_get(rx_ring->arr_hw_rh, stat);
2251 	} else {
2252 		aggr_port_t	*port = rx_ring->arr_port;
2253 
2254 		*val = mac_stat_get(port->lp_mh, stat);
2255 
2256 	}
2257 	return (0);
2258 }
2259 
2260 int
2261 aggr_tx_ring_stat(mac_ring_driver_t rdriver, uint_t stat, uint64_t *val)
2262 {
2263 	aggr_pseudo_tx_ring_t   *tx_ring = (aggr_pseudo_tx_ring_t *)rdriver;
2264 
2265 	if (tx_ring->atr_hw_rh != NULL) {
2266 		*val = mac_pseudo_tx_ring_stat_get(tx_ring->atr_hw_rh, stat);
2267 	} else {
2268 		aggr_port_t	*port = tx_ring->atr_port;
2269 
2270 		*val = mac_stat_get(port->lp_mh, stat);
2271 	}
2272 	return (0);
2273 }
2274 
2275 static int
2276 aggr_m_stat(void *arg, uint_t stat, uint64_t *val)
2277 {
2278 	aggr_grp_t		*grp = arg;
2279 	int			rval = 0;
2280 
2281 	mutex_enter(&grp->lg_stat_lock);
2282 
2283 	switch (stat) {
2284 	case MAC_STAT_IFSPEED:
2285 		*val = grp->lg_ifspeed;
2286 		break;
2287 
2288 	case ETHER_STAT_LINK_DUPLEX:
2289 		*val = grp->lg_link_duplex;
2290 		break;
2291 
2292 	default:
2293 		/*
2294 		 * For all other statistics, we return the aggregated stat
2295 		 * from the underlying ports.  aggr_grp_stat() will set
2296 		 * rval appropriately if the statistic isn't a counter.
2297 		 */
2298 		rval = aggr_grp_stat(grp, stat, val);
2299 	}
2300 
2301 	mutex_exit(&grp->lg_stat_lock);
2302 	return (rval);
2303 }
2304 
2305 static int
2306 aggr_m_start(void *arg)
2307 {
2308 	aggr_grp_t *grp = arg;
2309 	aggr_port_t *port;
2310 	mac_perim_handle_t mph, pmph;
2311 
2312 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2313 
2314 	/*
2315 	 * Attempts to start all configured members of the group.
2316 	 * Group members will be attached when their link-up notification
2317 	 * is received.
2318 	 */
2319 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2320 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2321 		if (aggr_port_start(port) != 0) {
2322 			mac_perim_exit(pmph);
2323 			continue;
2324 		}
2325 
2326 		/*
2327 		 * Turn on the promiscuous mode if it is required to receive
2328 		 * the non-primary address over a port, or the promiscous
2329 		 * mode is enabled over the aggr.
2330 		 */
2331 		if (grp->lg_promisc || port->lp_prom_addr != NULL) {
2332 			if (aggr_port_promisc(port, B_TRUE) != 0)
2333 				aggr_port_stop(port);
2334 		}
2335 		mac_perim_exit(pmph);
2336 	}
2337 
2338 	grp->lg_started = B_TRUE;
2339 
2340 	mac_perim_exit(mph);
2341 	return (0);
2342 }
2343 
2344 static void
2345 aggr_m_stop(void *arg)
2346 {
2347 	aggr_grp_t *grp = arg;
2348 	aggr_port_t *port;
2349 	mac_perim_handle_t mph, pmph;
2350 
2351 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2352 
2353 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2354 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2355 
2356 		/* reset port promiscuous mode */
2357 		(void) aggr_port_promisc(port, B_FALSE);
2358 
2359 		aggr_port_stop(port);
2360 		mac_perim_exit(pmph);
2361 	}
2362 
2363 	grp->lg_started = B_FALSE;
2364 	mac_perim_exit(mph);
2365 }
2366 
2367 static int
2368 aggr_m_promisc(void *arg, boolean_t on)
2369 {
2370 	aggr_grp_t *grp = arg;
2371 	aggr_port_t *port;
2372 	boolean_t link_state_changed = B_FALSE;
2373 	mac_perim_handle_t mph, pmph;
2374 
2375 	AGGR_GRP_REFHOLD(grp);
2376 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2377 
2378 	ASSERT(!grp->lg_closing);
2379 
2380 	if (on == grp->lg_promisc)
2381 		goto bail;
2382 
2383 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2384 		int	err = 0;
2385 
2386 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2387 		AGGR_PORT_REFHOLD(port);
2388 		if (!on && (port->lp_prom_addr == NULL))
2389 			err = aggr_port_promisc(port, B_FALSE);
2390 		else if (on && port->lp_started)
2391 			err = aggr_port_promisc(port, B_TRUE);
2392 
2393 		if (err != 0) {
2394 			if (aggr_grp_detach_port(grp, port))
2395 				link_state_changed = B_TRUE;
2396 		} else {
2397 			/*
2398 			 * If a port was detached because of a previous
2399 			 * failure changing the promiscuity, the port
2400 			 * is reattached when it successfully changes
2401 			 * the promiscuity now, and this might cause
2402 			 * the link state of the aggregation to change.
2403 			 */
2404 			if (aggr_grp_attach_port(grp, port))
2405 				link_state_changed = B_TRUE;
2406 		}
2407 		mac_perim_exit(pmph);
2408 		AGGR_PORT_REFRELE(port);
2409 	}
2410 
2411 	grp->lg_promisc = on;
2412 
2413 	if (link_state_changed)
2414 		mac_link_update(grp->lg_mh, grp->lg_link_state);
2415 
2416 bail:
2417 	mac_perim_exit(mph);
2418 	AGGR_GRP_REFRELE(grp);
2419 
2420 	return (0);
2421 }
2422 
2423 static void
2424 aggr_grp_port_rename(const char *new_name, void *arg)
2425 {
2426 	/*
2427 	 * aggr port's mac client name is the format of "aggr link name" plus
2428 	 * AGGR_PORT_NAME_DELIMIT plus "underneath link name".
2429 	 */
2430 	int aggr_len, link_len, clnt_name_len, i;
2431 	char *str_end, *str_st, *str_del;
2432 	char aggr_name[MAXNAMELEN];
2433 	char link_name[MAXNAMELEN];
2434 	char *clnt_name;
2435 	aggr_grp_t *aggr_grp = arg;
2436 	aggr_port_t *aggr_port = aggr_grp->lg_ports;
2437 
2438 	for (i = 0; i < aggr_grp->lg_nports; i++) {
2439 		clnt_name = mac_client_name(aggr_port->lp_mch);
2440 		clnt_name_len = strlen(clnt_name);
2441 		str_st = clnt_name;
2442 		str_end = &(clnt_name[clnt_name_len]);
2443 		str_del = strchr(str_st, AGGR_PORT_NAME_DELIMIT);
2444 		ASSERT(str_del != NULL);
2445 		aggr_len = (intptr_t)((uintptr_t)str_del - (uintptr_t)str_st);
2446 		link_len = (intptr_t)((uintptr_t)str_end - (uintptr_t)str_del);
2447 		bzero(aggr_name, MAXNAMELEN);
2448 		bzero(link_name, MAXNAMELEN);
2449 		bcopy(clnt_name, aggr_name, aggr_len);
2450 		bcopy(str_del, link_name, link_len + 1);
2451 		bzero(clnt_name, MAXNAMELEN);
2452 		(void) snprintf(clnt_name, MAXNAMELEN, "%s%s", new_name,
2453 		    link_name);
2454 
2455 		(void) mac_rename_primary(aggr_port->lp_mh, NULL);
2456 		aggr_port = aggr_port->lp_next;
2457 	}
2458 }
2459 
2460 /*
2461  * Initialize the capabilities that are advertised for the group
2462  * according to the capabilities of the constituent ports.
2463  */
2464 static boolean_t
2465 aggr_m_capab_get(void *arg, mac_capab_t cap, void *cap_data)
2466 {
2467 	aggr_grp_t *grp = arg;
2468 
2469 	switch (cap) {
2470 	case MAC_CAPAB_HCKSUM: {
2471 		uint32_t *hcksum_txflags = cap_data;
2472 		*hcksum_txflags = grp->lg_hcksum_txflags;
2473 		break;
2474 	}
2475 	case MAC_CAPAB_LSO: {
2476 		mac_capab_lso_t *cap_lso = cap_data;
2477 
2478 		if (grp->lg_lso) {
2479 			*cap_lso = grp->lg_cap_lso;
2480 			break;
2481 		} else {
2482 			return (B_FALSE);
2483 		}
2484 	}
2485 	case MAC_CAPAB_NO_NATIVEVLAN:
2486 		return (!grp->lg_vlan);
2487 	case MAC_CAPAB_NO_ZCOPY:
2488 		return (!grp->lg_zcopy);
2489 	case MAC_CAPAB_RINGS: {
2490 		mac_capab_rings_t *cap_rings = cap_data;
2491 		uint_t ring_cnt = 0;
2492 
2493 		for (uint_t i = 0; i < grp->lg_rx_group_count; i++)
2494 			ring_cnt += grp->lg_rx_groups[i].arg_ring_cnt;
2495 
2496 		if (cap_rings->mr_type == MAC_RING_TYPE_RX) {
2497 			cap_rings->mr_group_type = MAC_GROUP_TYPE_STATIC;
2498 			cap_rings->mr_rnum = ring_cnt;
2499 			cap_rings->mr_gnum = grp->lg_rx_group_count;
2500 			cap_rings->mr_gaddring = NULL;
2501 			cap_rings->mr_gremring = NULL;
2502 		} else {
2503 			cap_rings->mr_group_type = MAC_GROUP_TYPE_STATIC;
2504 			cap_rings->mr_rnum = grp->lg_tx_group.atg_ring_cnt;
2505 			cap_rings->mr_gnum = 0;
2506 		}
2507 		cap_rings->mr_rget = aggr_fill_ring;
2508 		cap_rings->mr_gget = aggr_fill_group;
2509 		break;
2510 	}
2511 	case MAC_CAPAB_AGGR:
2512 	{
2513 		mac_capab_aggr_t *aggr_cap;
2514 
2515 		if (cap_data != NULL) {
2516 			aggr_cap = cap_data;
2517 			aggr_cap->mca_rename_fn = aggr_grp_port_rename;
2518 			aggr_cap->mca_unicst = aggr_m_unicst;
2519 			aggr_cap->mca_find_tx_ring_fn = aggr_find_tx_ring;
2520 			aggr_cap->mca_arg = arg;
2521 		}
2522 		return (B_TRUE);
2523 	}
2524 	default:
2525 		return (B_FALSE);
2526 	}
2527 	return (B_TRUE);
2528 }
2529 
2530 /*
2531  * Callback function for MAC layer to register groups.
2532  */
2533 static void
2534 aggr_fill_group(void *arg, mac_ring_type_t rtype, const int index,
2535     mac_group_info_t *infop, mac_group_handle_t gh)
2536 {
2537 	aggr_grp_t *grp = arg;
2538 
2539 	if (rtype == MAC_RING_TYPE_RX) {
2540 		aggr_pseudo_rx_group_t *rx_group = &grp->lg_rx_groups[index];
2541 
2542 		rx_group->arg_gh = gh;
2543 		rx_group->arg_grp = grp;
2544 
2545 		infop->mgi_driver = (mac_group_driver_t)rx_group;
2546 		infop->mgi_start = NULL;
2547 		infop->mgi_stop = NULL;
2548 		infop->mgi_addmac = aggr_addmac;
2549 		infop->mgi_remmac = aggr_remmac;
2550 		infop->mgi_count = rx_group->arg_ring_cnt;
2551 
2552 		/*
2553 		 * Always set the HW VLAN callbacks. They are smart
2554 		 * enough to know when a port has HW VLAN filters to
2555 		 * program and when it doesn't.
2556 		 */
2557 		infop->mgi_addvlan = aggr_addvlan;
2558 		infop->mgi_remvlan = aggr_remvlan;
2559 	} else {
2560 		aggr_pseudo_tx_group_t *tx_group = &grp->lg_tx_group;
2561 
2562 		ASSERT3S(index, ==, 0);
2563 		tx_group->atg_gh = gh;
2564 	}
2565 }
2566 
2567 /*
2568  * Callback funtion for MAC layer to register all rings.
2569  */
2570 static void
2571 aggr_fill_ring(void *arg, mac_ring_type_t rtype, const int rg_index,
2572     const int index, mac_ring_info_t *infop, mac_ring_handle_t rh)
2573 {
2574 	aggr_grp_t	*grp = arg;
2575 
2576 	switch (rtype) {
2577 	case MAC_RING_TYPE_RX: {
2578 		aggr_pseudo_rx_group_t	*rx_group;
2579 		aggr_pseudo_rx_ring_t	*rx_ring;
2580 		mac_intr_t		aggr_mac_intr;
2581 
2582 		rx_group = &grp->lg_rx_groups[rg_index];
2583 		ASSERT3S(index, >=, 0);
2584 		ASSERT3S(index, <, rx_group->arg_ring_cnt);
2585 		rx_ring = rx_group->arg_rings + index;
2586 		rx_ring->arr_rh = rh;
2587 
2588 		/*
2589 		 * Entrypoint to enable interrupt (disable poll) and
2590 		 * disable interrupt (enable poll).
2591 		 */
2592 		aggr_mac_intr.mi_handle = (mac_intr_handle_t)rx_ring;
2593 		aggr_mac_intr.mi_enable = aggr_pseudo_enable_intr;
2594 		aggr_mac_intr.mi_disable = aggr_pseudo_disable_intr;
2595 		aggr_mac_intr.mi_ddi_handle = NULL;
2596 
2597 		infop->mri_driver = (mac_ring_driver_t)rx_ring;
2598 		infop->mri_start = aggr_pseudo_start_rx_ring;
2599 		infop->mri_stop = aggr_pseudo_stop_rx_ring;
2600 
2601 		infop->mri_intr = aggr_mac_intr;
2602 		infop->mri_poll = aggr_rx_poll;
2603 
2604 		infop->mri_stat = aggr_rx_ring_stat;
2605 		break;
2606 	}
2607 	case MAC_RING_TYPE_TX: {
2608 		aggr_pseudo_tx_group_t	*tx_group = &grp->lg_tx_group;
2609 		aggr_pseudo_tx_ring_t	*tx_ring;
2610 
2611 		ASSERT(rg_index == -1);
2612 		ASSERT(index < tx_group->atg_ring_cnt);
2613 
2614 		tx_ring = &tx_group->atg_rings[index];
2615 		tx_ring->atr_rh = rh;
2616 
2617 		infop->mri_driver = (mac_ring_driver_t)tx_ring;
2618 		infop->mri_start = NULL;
2619 		infop->mri_stop = NULL;
2620 		infop->mri_tx = aggr_ring_tx;
2621 		infop->mri_stat = aggr_tx_ring_stat;
2622 		/*
2623 		 * Use the hw TX ring handle to find if the ring needs
2624 		 * serialization or not. For NICs that do not expose
2625 		 * Tx rings, atr_hw_rh will be NULL.
2626 		 */
2627 		if (tx_ring->atr_hw_rh != NULL) {
2628 			infop->mri_flags =
2629 			    mac_hwring_getinfo(tx_ring->atr_hw_rh);
2630 		}
2631 		break;
2632 	}
2633 	default:
2634 		break;
2635 	}
2636 }
2637 
2638 static mblk_t *
2639 aggr_rx_poll(void *arg, int bytes_to_pickup)
2640 {
2641 	aggr_pseudo_rx_ring_t *rr_ring = arg;
2642 	aggr_port_t *port = rr_ring->arr_port;
2643 	aggr_grp_t *grp = port->lp_grp;
2644 	mblk_t *mp_chain, *mp, **mpp;
2645 
2646 	mp_chain = mac_hwring_poll(rr_ring->arr_hw_rh, bytes_to_pickup);
2647 
2648 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
2649 		return (mp_chain);
2650 
2651 	mpp = &mp_chain;
2652 	while ((mp = *mpp) != NULL) {
2653 		if (MBLKL(mp) >= sizeof (struct ether_header)) {
2654 			struct ether_header *ehp;
2655 
2656 			ehp = (struct ether_header *)mp->b_rptr;
2657 			if (ntohs(ehp->ether_type) == ETHERTYPE_SLOW) {
2658 				*mpp = mp->b_next;
2659 				mp->b_next = NULL;
2660 				aggr_recv_lacp(port,
2661 				    (mac_resource_handle_t)rr_ring, mp);
2662 				continue;
2663 			}
2664 		}
2665 
2666 		if (!port->lp_collector_enabled) {
2667 			*mpp = mp->b_next;
2668 			mp->b_next = NULL;
2669 			freemsg(mp);
2670 			continue;
2671 		}
2672 		mpp = &mp->b_next;
2673 	}
2674 	return (mp_chain);
2675 }
2676 
2677 static int
2678 aggr_addmac(void *arg, const uint8_t *mac_addr)
2679 {
2680 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)arg;
2681 	aggr_unicst_addr_t	*addr, **pprev;
2682 	aggr_grp_t		*grp = rx_group->arg_grp;
2683 	aggr_port_t		*port, *p;
2684 	mac_perim_handle_t	mph;
2685 	int			err = 0;
2686 	uint_t			idx = rx_group->arg_index;
2687 
2688 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2689 
2690 	if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) == 0) {
2691 		mac_perim_exit(mph);
2692 		return (0);
2693 	}
2694 
2695 	/*
2696 	 * Insert this mac address into the list of mac addresses owned by
2697 	 * the aggregation pseudo group.
2698 	 */
2699 	pprev = &rx_group->arg_macaddr;
2700 	while ((addr = *pprev) != NULL) {
2701 		if (bcmp(mac_addr, addr->aua_addr, ETHERADDRL) == 0) {
2702 			mac_perim_exit(mph);
2703 			return (EEXIST);
2704 		}
2705 		pprev = &addr->aua_next;
2706 	}
2707 	addr = kmem_alloc(sizeof (aggr_unicst_addr_t), KM_SLEEP);
2708 	bcopy(mac_addr, addr->aua_addr, ETHERADDRL);
2709 	addr->aua_next = NULL;
2710 	*pprev = addr;
2711 
2712 	for (port = grp->lg_ports; port != NULL; port = port->lp_next)
2713 		if ((err = aggr_port_addmac(port, idx, mac_addr)) != 0)
2714 			break;
2715 
2716 	if (err != 0) {
2717 		for (p = grp->lg_ports; p != port; p = p->lp_next)
2718 			aggr_port_remmac(p, idx, mac_addr);
2719 
2720 		*pprev = NULL;
2721 		kmem_free(addr, sizeof (aggr_unicst_addr_t));
2722 	}
2723 
2724 	mac_perim_exit(mph);
2725 	return (err);
2726 }
2727 
2728 static int
2729 aggr_remmac(void *arg, const uint8_t *mac_addr)
2730 {
2731 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)arg;
2732 	aggr_unicst_addr_t	*addr, **pprev;
2733 	aggr_grp_t		*grp = rx_group->arg_grp;
2734 	aggr_port_t		*port;
2735 	mac_perim_handle_t	mph;
2736 	int			err = 0;
2737 
2738 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2739 
2740 	if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) == 0) {
2741 		mac_perim_exit(mph);
2742 		return (0);
2743 	}
2744 
2745 	/*
2746 	 * Insert this mac address into the list of mac addresses owned by
2747 	 * the aggregation pseudo group.
2748 	 */
2749 	pprev = &rx_group->arg_macaddr;
2750 	while ((addr = *pprev) != NULL) {
2751 		if (bcmp(mac_addr, addr->aua_addr, ETHERADDRL) != 0) {
2752 			pprev = &addr->aua_next;
2753 			continue;
2754 		}
2755 		break;
2756 	}
2757 	if (addr == NULL) {
2758 		mac_perim_exit(mph);
2759 		return (EINVAL);
2760 	}
2761 
2762 	for (port = grp->lg_ports; port != NULL; port = port->lp_next)
2763 		aggr_port_remmac(port, rx_group->arg_index, mac_addr);
2764 
2765 	*pprev = addr->aua_next;
2766 	kmem_free(addr, sizeof (aggr_unicst_addr_t));
2767 
2768 	mac_perim_exit(mph);
2769 	return (err);
2770 }
2771 
2772 /*
2773  * Search for VID in the Rx group's list and return a pointer if
2774  * found. Otherwise return NULL.
2775  */
2776 static aggr_vlan_t *
2777 aggr_find_vlan(aggr_pseudo_rx_group_t *rx_group, uint16_t vid)
2778 {
2779 	ASSERT(MAC_PERIM_HELD(rx_group->arg_grp->lg_mh));
2780 	for (aggr_vlan_t *avp = list_head(&rx_group->arg_vlans); avp != NULL;
2781 	    avp = list_next(&rx_group->arg_vlans, avp)) {
2782 		if (avp->av_vid == vid)
2783 			return (avp);
2784 	}
2785 
2786 	return (NULL);
2787 }
2788 
2789 /*
2790  * Accept traffic on the specified VID.
2791  *
2792  * Persist VLAN state in the aggr so that ports added later will
2793  * receive the correct filters. In the future it would be nice to
2794  * allow aggr to iterate its clients instead of duplicating state.
2795  */
2796 static int
2797 aggr_addvlan(mac_group_driver_t gdriver, uint16_t vid)
2798 {
2799 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)gdriver;
2800 	aggr_grp_t		*aggr = rx_group->arg_grp;
2801 	aggr_port_t		*port, *p;
2802 	mac_perim_handle_t	mph;
2803 	int			err = 0;
2804 	aggr_vlan_t		*avp = NULL;
2805 	uint_t			idx = rx_group->arg_index;
2806 
2807 	mac_perim_enter_by_mh(aggr->lg_mh, &mph);
2808 
2809 	if (vid == MAC_VLAN_UNTAGGED) {
2810 		/*
2811 		 * Aggr is both a MAC provider and MAC client. As a
2812 		 * MAC provider it is passed MAC_VLAN_UNTAGGED by its
2813 		 * client. As a client itself, it should pass
2814 		 * VLAN_ID_NONE to its ports.
2815 		 */
2816 		vid = VLAN_ID_NONE;
2817 		rx_group->arg_untagged++;
2818 		goto update_ports;
2819 	}
2820 
2821 	avp = aggr_find_vlan(rx_group, vid);
2822 
2823 	if (avp != NULL) {
2824 		avp->av_refs++;
2825 		mac_perim_exit(mph);
2826 		return (0);
2827 	}
2828 
2829 	avp = kmem_zalloc(sizeof (aggr_vlan_t), KM_SLEEP);
2830 	avp->av_vid = vid;
2831 	avp->av_refs = 1;
2832 
2833 update_ports:
2834 	for (port = aggr->lg_ports; port != NULL; port = port->lp_next)
2835 		if ((err = aggr_port_addvlan(port, idx, vid)) != 0)
2836 			break;
2837 
2838 	if (err != 0) {
2839 		/*
2840 		 * If any of these calls fail then we are in a
2841 		 * situation where the ports have different HW state.
2842 		 * There's no reasonable action the MAC client can
2843 		 * take in this scenario to rectify the situation.
2844 		 */
2845 		for (p = aggr->lg_ports; p != port; p = p->lp_next) {
2846 			int err2;
2847 
2848 			if ((err2 = aggr_port_remvlan(p, idx, vid)) != 0) {
2849 				cmn_err(CE_WARN, "Failed to remove VLAN %u"
2850 				    " from port %s: errno %d.", vid,
2851 				    mac_client_name(p->lp_mch), err2);
2852 			}
2853 
2854 		}
2855 
2856 		if (vid == VLAN_ID_NONE)
2857 			rx_group->arg_untagged--;
2858 
2859 		if (avp != NULL) {
2860 			kmem_free(avp, sizeof (aggr_vlan_t));
2861 			avp = NULL;
2862 		}
2863 	}
2864 
2865 	if (avp != NULL)
2866 		list_insert_tail(&rx_group->arg_vlans, avp);
2867 
2868 done:
2869 	mac_perim_exit(mph);
2870 	return (err);
2871 }
2872 
2873 /*
2874  * Stop accepting traffic on this VLAN if it's the last use of this VLAN.
2875  */
2876 static int
2877 aggr_remvlan(mac_group_driver_t gdriver, uint16_t vid)
2878 {
2879 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)gdriver;
2880 	aggr_grp_t		*aggr = rx_group->arg_grp;
2881 	aggr_port_t		*port, *p;
2882 	mac_perim_handle_t	mph;
2883 	int			err = 0;
2884 	aggr_vlan_t		*avp = NULL;
2885 	uint_t			idx = rx_group->arg_index;
2886 
2887 	mac_perim_enter_by_mh(aggr->lg_mh, &mph);
2888 
2889 	/*
2890 	 * See the comment in aggr_addvlan().
2891 	 */
2892 	if (vid == MAC_VLAN_UNTAGGED) {
2893 		vid = VLAN_ID_NONE;
2894 		rx_group->arg_untagged--;
2895 
2896 		if (rx_group->arg_untagged > 0)
2897 			goto done;
2898 
2899 		goto update_ports;
2900 	}
2901 
2902 	avp = aggr_find_vlan(rx_group, vid);
2903 
2904 	if (avp == NULL) {
2905 		err = ENOENT;
2906 		goto done;
2907 	}
2908 
2909 	avp->av_refs--;
2910 
2911 	if (avp->av_refs > 0)
2912 		goto done;
2913 
2914 update_ports:
2915 	for (port = aggr->lg_ports; port != NULL; port = port->lp_next)
2916 		if ((err = aggr_port_remvlan(port, idx, vid)) != 0)
2917 			break;
2918 
2919 	/*
2920 	 * See the comment in aggr_addvlan() for justification of the
2921 	 * use of VERIFY here.
2922 	 */
2923 	if (err != 0) {
2924 		for (p = aggr->lg_ports; p != port; p = p->lp_next) {
2925 			int err2;
2926 
2927 			if ((err2 = aggr_port_addvlan(p, idx, vid)) != 0) {
2928 				cmn_err(CE_WARN, "Failed to add VLAN %u"
2929 				    " to port %s: errno %d.", vid,
2930 				    mac_client_name(p->lp_mch), err2);
2931 			}
2932 		}
2933 
2934 		if (avp != NULL)
2935 			avp->av_refs++;
2936 
2937 		if (vid == VLAN_ID_NONE)
2938 			rx_group->arg_untagged++;
2939 
2940 		goto done;
2941 	}
2942 
2943 	if (err == 0 && avp != NULL) {
2944 		VERIFY3U(avp->av_refs, ==, 0);
2945 		list_remove(&rx_group->arg_vlans, avp);
2946 		kmem_free(avp, sizeof (aggr_vlan_t));
2947 	}
2948 
2949 done:
2950 	mac_perim_exit(mph);
2951 	return (err);
2952 }
2953 
2954 /*
2955  * Add or remove the multicast addresses that are defined for the group
2956  * to or from the specified port.
2957  *
2958  * Note that aggr_grp_multicst_port(..., B_TRUE) is called when the port
2959  * is started and attached, and aggr_grp_multicst_port(..., B_FALSE) is
2960  * called when the port is either stopped or detached.
2961  */
2962 void
2963 aggr_grp_multicst_port(aggr_port_t *port, boolean_t add)
2964 {
2965 	aggr_grp_t *grp = port->lp_grp;
2966 
2967 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
2968 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
2969 
2970 	if (!port->lp_started || port->lp_state != AGGR_PORT_STATE_ATTACHED)
2971 		return;
2972 
2973 	mac_multicast_refresh(grp->lg_mh, aggr_port_multicst, port, add);
2974 }
2975 
2976 static int
2977 aggr_m_multicst(void *arg, boolean_t add, const uint8_t *addrp)
2978 {
2979 	aggr_grp_t *grp = arg;
2980 	aggr_port_t *port = NULL, *errport = NULL;
2981 	mac_perim_handle_t mph;
2982 	int err = 0;
2983 
2984 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2985 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2986 		if (port->lp_state != AGGR_PORT_STATE_ATTACHED ||
2987 		    !port->lp_started) {
2988 			continue;
2989 		}
2990 		err = aggr_port_multicst(port, add, addrp);
2991 		if (err != 0) {
2992 			errport = port;
2993 			break;
2994 		}
2995 	}
2996 
2997 	/*
2998 	 * At least one port caused error return and this error is returned to
2999 	 * mac, eventually a NAK would be sent upwards.
3000 	 * Some ports have this multicast address listed now, and some don't.
3001 	 * Treat this error as a whole aggr failure not individual port failure.
3002 	 * Therefore remove this multicast address from other ports.
3003 	 */
3004 	if ((err != 0) && add) {
3005 		for (port = grp->lg_ports; port != errport;
3006 		    port = port->lp_next) {
3007 			if (port->lp_state != AGGR_PORT_STATE_ATTACHED ||
3008 			    !port->lp_started) {
3009 				continue;
3010 			}
3011 			(void) aggr_port_multicst(port, B_FALSE, addrp);
3012 		}
3013 	}
3014 	mac_perim_exit(mph);
3015 	return (err);
3016 }
3017 
3018 static int
3019 aggr_m_unicst(void *arg, const uint8_t *macaddr)
3020 {
3021 	aggr_grp_t *grp = arg;
3022 	mac_perim_handle_t mph;
3023 	int err;
3024 
3025 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
3026 	err = aggr_grp_modify_common(grp, AGGR_MODIFY_MAC, 0, B_TRUE, macaddr,
3027 	    0, 0);
3028 	mac_perim_exit(mph);
3029 	return (err);
3030 }
3031 
3032 /*
3033  * Initialize the capabilities that are advertised for the group
3034  * according to the capabilities of the constituent ports.
3035  */
3036 static void
3037 aggr_grp_capab_set(aggr_grp_t *grp)
3038 {
3039 	uint32_t cksum;
3040 	aggr_port_t *port;
3041 	mac_capab_lso_t cap_lso;
3042 
3043 	ASSERT(grp->lg_mh == NULL);
3044 	ASSERT(grp->lg_ports != NULL);
3045 
3046 	grp->lg_hcksum_txflags = (uint32_t)-1;
3047 	grp->lg_zcopy = B_TRUE;
3048 	grp->lg_vlan = B_TRUE;
3049 
3050 	grp->lg_lso = B_TRUE;
3051 	grp->lg_cap_lso.lso_flags = (t_uscalar_t)-1;
3052 	grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max = (t_uscalar_t)-1;
3053 
3054 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
3055 		if (!mac_capab_get(port->lp_mh, MAC_CAPAB_HCKSUM, &cksum))
3056 			cksum = 0;
3057 		grp->lg_hcksum_txflags &= cksum;
3058 
3059 		grp->lg_vlan &=
3060 		    !mac_capab_get(port->lp_mh, MAC_CAPAB_NO_NATIVEVLAN, NULL);
3061 
3062 		grp->lg_zcopy &=
3063 		    !mac_capab_get(port->lp_mh, MAC_CAPAB_NO_ZCOPY, NULL);
3064 
3065 		grp->lg_lso &=
3066 		    mac_capab_get(port->lp_mh, MAC_CAPAB_LSO, &cap_lso);
3067 		if (grp->lg_lso) {
3068 			grp->lg_cap_lso.lso_flags &= cap_lso.lso_flags;
3069 			if (grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max >
3070 			    cap_lso.lso_basic_tcp_ipv4.lso_max)
3071 				grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max =
3072 				    cap_lso.lso_basic_tcp_ipv4.lso_max;
3073 		}
3074 	}
3075 }
3076 
3077 /*
3078  * Checks whether the capabilities of the port being added are compatible
3079  * with the current capabilities of the aggregation.
3080  */
3081 static boolean_t
3082 aggr_grp_capab_check(aggr_grp_t *grp, aggr_port_t *port)
3083 {
3084 	uint32_t hcksum_txflags;
3085 
3086 	ASSERT(grp->lg_ports != NULL);
3087 
3088 	if (((!mac_capab_get(port->lp_mh, MAC_CAPAB_NO_NATIVEVLAN, NULL)) &
3089 	    grp->lg_vlan) != grp->lg_vlan) {
3090 		return (B_FALSE);
3091 	}
3092 
3093 	if (((!mac_capab_get(port->lp_mh, MAC_CAPAB_NO_ZCOPY, NULL)) &
3094 	    grp->lg_zcopy) != grp->lg_zcopy) {
3095 		return (B_FALSE);
3096 	}
3097 
3098 	if (!mac_capab_get(port->lp_mh, MAC_CAPAB_HCKSUM, &hcksum_txflags)) {
3099 		if (grp->lg_hcksum_txflags != 0)
3100 			return (B_FALSE);
3101 	} else if ((hcksum_txflags & grp->lg_hcksum_txflags) !=
3102 	    grp->lg_hcksum_txflags) {
3103 		return (B_FALSE);
3104 	}
3105 
3106 	if (grp->lg_lso) {
3107 		mac_capab_lso_t cap_lso;
3108 
3109 		if (mac_capab_get(port->lp_mh, MAC_CAPAB_LSO, &cap_lso)) {
3110 			if ((grp->lg_cap_lso.lso_flags & cap_lso.lso_flags) !=
3111 			    grp->lg_cap_lso.lso_flags)
3112 				return (B_FALSE);
3113 			if (grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max >
3114 			    cap_lso.lso_basic_tcp_ipv4.lso_max)
3115 				return (B_FALSE);
3116 		} else {
3117 			return (B_FALSE);
3118 		}
3119 	}
3120 
3121 	return (B_TRUE);
3122 }
3123 
3124 /*
3125  * Returns the maximum SDU according to the SDU of the constituent ports.
3126  */
3127 static uint_t
3128 aggr_grp_max_sdu(aggr_grp_t *grp)
3129 {
3130 	uint_t max_sdu = (uint_t)-1;
3131 	aggr_port_t *port;
3132 
3133 	ASSERT(grp->lg_ports != NULL);
3134 
3135 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
3136 		uint_t port_sdu_max;
3137 
3138 		mac_sdu_get(port->lp_mh, NULL, &port_sdu_max);
3139 		if (max_sdu > port_sdu_max)
3140 			max_sdu = port_sdu_max;
3141 	}
3142 
3143 	return (max_sdu);
3144 }
3145 
3146 /*
3147  * Checks if the maximum SDU of the specified port is compatible
3148  * with the maximum SDU of the specified aggregation group, returns
3149  * B_TRUE if it is, B_FALSE otherwise.
3150  */
3151 static boolean_t
3152 aggr_grp_sdu_check(aggr_grp_t *grp, aggr_port_t *port)
3153 {
3154 	uint_t port_sdu_max;
3155 
3156 	mac_sdu_get(port->lp_mh, NULL, &port_sdu_max);
3157 	return (port_sdu_max >= grp->lg_max_sdu);
3158 }
3159 
3160 /*
3161  * Returns the maximum margin according to the margin of the constituent ports.
3162  */
3163 static uint32_t
3164 aggr_grp_max_margin(aggr_grp_t *grp)
3165 {
3166 	uint32_t margin = UINT32_MAX;
3167 	aggr_port_t *port;
3168 
3169 	ASSERT(grp->lg_mh == NULL);
3170 	ASSERT(grp->lg_ports != NULL);
3171 
3172 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
3173 		if (margin > port->lp_margin)
3174 			margin = port->lp_margin;
3175 	}
3176 
3177 	grp->lg_margin = margin;
3178 	return (margin);
3179 }
3180 
3181 /*
3182  * Checks if the maximum margin of the specified port is compatible
3183  * with the maximum margin of the specified aggregation group, returns
3184  * B_TRUE if it is, B_FALSE otherwise.
3185  */
3186 static boolean_t
3187 aggr_grp_margin_check(aggr_grp_t *grp, aggr_port_t *port)
3188 {
3189 	if (port->lp_margin >= grp->lg_margin)
3190 		return (B_TRUE);
3191 
3192 	/*
3193 	 * See whether the current margin value is allowed to be changed to
3194 	 * the new value.
3195 	 */
3196 	if (!mac_margin_update(grp->lg_mh, port->lp_margin))
3197 		return (B_FALSE);
3198 
3199 	grp->lg_margin = port->lp_margin;
3200 	return (B_TRUE);
3201 }
3202 
3203 /*
3204  * Set MTU on individual ports of an aggregation group
3205  */
3206 static int
3207 aggr_set_port_sdu(aggr_grp_t *grp, aggr_port_t *port, uint32_t sdu,
3208     uint32_t *old_mtu)
3209 {
3210 	boolean_t		removed = B_FALSE;
3211 	mac_perim_handle_t	mph;
3212 	mac_diag_t		diag;
3213 	int			err, rv, retry = 0;
3214 
3215 	if (port->lp_mah != NULL) {
3216 		(void) mac_unicast_remove(port->lp_mch, port->lp_mah);
3217 		port->lp_mah = NULL;
3218 		removed = B_TRUE;
3219 	}
3220 	err = mac_set_mtu(port->lp_mh, sdu, old_mtu);
3221 try_again:
3222 	if (removed && (rv = mac_unicast_add(port->lp_mch, NULL,
3223 	    MAC_UNICAST_PRIMARY | MAC_UNICAST_DISABLE_TX_VID_CHECK,
3224 	    &port->lp_mah, 0, &diag)) != 0) {
3225 		/*
3226 		 * following is a workaround for a bug in 'bge' driver.
3227 		 * See CR 6794654 for more information and this work around
3228 		 * will be removed once the CR is fixed.
3229 		 */
3230 		if (rv == EIO && retry++ < 3) {
3231 			delay(2 * hz);
3232 			goto try_again;
3233 		}
3234 		/*
3235 		 * if mac_unicast_add() failed while setting the MTU,
3236 		 * detach the port from the group.
3237 		 */
3238 		mac_perim_enter_by_mh(port->lp_mh, &mph);
3239 		(void) aggr_grp_detach_port(grp, port);
3240 		mac_perim_exit(mph);
3241 		cmn_err(CE_WARN, "Unable to restart the port %s while "
3242 		    "setting MTU. Detaching the port from the aggregation.",
3243 		    mac_client_name(port->lp_mch));
3244 	}
3245 	return (err);
3246 }
3247 
3248 static int
3249 aggr_sdu_update(aggr_grp_t *grp, uint32_t sdu)
3250 {
3251 	int			err = 0, i, rv;
3252 	aggr_port_t		*port;
3253 	uint32_t		*mtu;
3254 
3255 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
3256 
3257 	/*
3258 	 * If the MTU being set is equal to aggr group's maximum
3259 	 * allowable value, then there is nothing to change
3260 	 */
3261 	if (sdu == grp->lg_max_sdu)
3262 		return (0);
3263 
3264 	/* 0 is aggr group's min sdu */
3265 	if (sdu == 0)
3266 		return (EINVAL);
3267 
3268 	mtu = kmem_alloc(sizeof (uint32_t) * grp->lg_nports, KM_SLEEP);
3269 	for (port = grp->lg_ports, i = 0; port != NULL && err == 0;
3270 	    port = port->lp_next, i++) {
3271 		err = aggr_set_port_sdu(grp, port, sdu, mtu + i);
3272 	}
3273 	if (err != 0) {
3274 		/* recover from error: reset the mtus of the ports */
3275 		aggr_port_t *tmp;
3276 
3277 		for (tmp = grp->lg_ports, i = 0; tmp != port;
3278 		    tmp = tmp->lp_next, i++) {
3279 			(void) aggr_set_port_sdu(grp, tmp, *(mtu + i), NULL);
3280 		}
3281 		goto bail;
3282 	}
3283 	grp->lg_max_sdu = aggr_grp_max_sdu(grp);
3284 	rv = mac_maxsdu_update(grp->lg_mh, grp->lg_max_sdu);
3285 	ASSERT(rv == 0);
3286 bail:
3287 	kmem_free(mtu, sizeof (uint32_t) * grp->lg_nports);
3288 	return (err);
3289 }
3290 
3291 /*
3292  * Callback functions for set/get of properties
3293  */
3294 /*ARGSUSED*/
3295 static int
3296 aggr_m_setprop(void *m_driver, const char *pr_name, mac_prop_id_t pr_num,
3297     uint_t pr_valsize, const void *pr_val)
3298 {
3299 	int		err = ENOTSUP;
3300 	aggr_grp_t	*grp = m_driver;
3301 
3302 	switch (pr_num) {
3303 	case MAC_PROP_MTU: {
3304 		uint32_t	mtu;
3305 
3306 		if (pr_valsize < sizeof (mtu)) {
3307 			err = EINVAL;
3308 			break;
3309 		}
3310 		bcopy(pr_val, &mtu, sizeof (mtu));
3311 		err = aggr_sdu_update(grp, mtu);
3312 		break;
3313 	}
3314 	default:
3315 		break;
3316 	}
3317 	return (err);
3318 }
3319 
3320 typedef struct rboundary {
3321 	uint32_t	bval;
3322 	int		btype;
3323 } rboundary_t;
3324 
3325 /*
3326  * This function finds the intersection of mtu ranges stored in arrays -
3327  * mrange[0] ... mrange[mcount -1]. It returns the intersection in rval.
3328  * Individual arrays are assumed to contain non-overlapping ranges.
3329  * Algorithm:
3330  *   A range has two boundaries - min and max. We scan all arrays and store
3331  * each boundary as a separate element in a temporary array. We also store
3332  * the boundary types, min or max, as +1 or -1 respectively in the temporary
3333  * array. Then we sort the temporary array in ascending order. We scan the
3334  * sorted array from lower to higher values and keep a cumulative sum of
3335  * boundary types. Element in the temporary array for which the sum reaches
3336  * mcount is a min boundary of a range in the result and next element will be
3337  * max boundary.
3338  *
3339  * Example for mcount = 3,
3340  *
3341  *  ----|_________|-------|_______|----|__|------ mrange[0]
3342  *
3343  *  -------|________|--|____________|-----|___|-- mrange[1]
3344  *
3345  *  --------|________________|-------|____|------ mrange[2]
3346  *
3347  *                                      3 2 1
3348  *                                       \|/
3349  *      1  23     2 1  2  3  2    1 01 2  V   0  <- the sum
3350  *  ----|--||-----|-|--|--|--|----|-||-|--|---|-- sorted array
3351  *
3352  *                                 same min and max
3353  *                                        V
3354  *  --------|_____|-------|__|------------|------ intersecting ranges
3355  */
3356 void
3357 aggr_mtu_range_intersection(mac_propval_range_t **mrange, int mcount,
3358     mac_propval_uint32_range_t **prval, int *prmaxcnt, int *prcount)
3359 {
3360 	mac_propval_uint32_range_t	*rval, *ur;
3361 	int				rmaxcnt, rcount;
3362 	size_t				sz_range32;
3363 	rboundary_t			*ta; /* temporary array */
3364 	rboundary_t			temp;
3365 	boolean_t			range_started = B_FALSE;
3366 	int				i, j, m, sum;
3367 
3368 	sz_range32 = sizeof (mac_propval_uint32_range_t);
3369 
3370 	for (i = 0, rmaxcnt = 0; i < mcount; i++)
3371 		rmaxcnt += mrange[i]->mpr_count;
3372 
3373 	/* Allocate enough space to store the results */
3374 	rval = kmem_alloc(rmaxcnt * sz_range32, KM_SLEEP);
3375 
3376 	/* Number of boundaries are twice as many as ranges */
3377 	ta = kmem_alloc(2 * rmaxcnt * sizeof (rboundary_t), KM_SLEEP);
3378 
3379 	for (i = 0, m = 0; i < mcount; i++) {
3380 		ur = &(mrange[i]->mpr_range_uint32[0]);
3381 		for (j = 0; j < mrange[i]->mpr_count; j++) {
3382 			ta[m].bval = ur[j].mpur_min;
3383 			ta[m++].btype = 1;
3384 			ta[m].bval = ur[j].mpur_max;
3385 			ta[m++].btype = -1;
3386 		}
3387 	}
3388 
3389 	/*
3390 	 * Sort the temporary array in ascending order of bval;
3391 	 * if boundary values are same then sort on btype.
3392 	 */
3393 	for (i = 0; i < m-1; i++) {
3394 		for (j = i+1; j < m; j++) {
3395 			if ((ta[i].bval > ta[j].bval) ||
3396 			    ((ta[i].bval == ta[j].bval) &&
3397 			    (ta[i].btype < ta[j].btype))) {
3398 				temp = ta[i];
3399 				ta[i] = ta[j];
3400 				ta[j] = temp;
3401 			}
3402 		}
3403 	}
3404 
3405 	/* Walk through temporary array to find all ranges in the results */
3406 	for (i = 0, sum = 0, rcount = 0; i < m; i++) {
3407 		sum += ta[i].btype;
3408 		if (sum == mcount) {
3409 			rval[rcount].mpur_min = ta[i].bval;
3410 			range_started = B_TRUE;
3411 		} else if (sum < mcount && range_started) {
3412 			rval[rcount++].mpur_max = ta[i].bval;
3413 			range_started = B_FALSE;
3414 		}
3415 	}
3416 
3417 	*prval = rval;
3418 	*prmaxcnt = rmaxcnt;
3419 	*prcount = rcount;
3420 
3421 	kmem_free(ta, 2 * rmaxcnt * sizeof (rboundary_t));
3422 }
3423 
3424 /*
3425  * Returns the mtu ranges which could be supported by aggr group.
3426  * prmaxcnt returns the size of the buffer prval, prcount returns
3427  * the number of valid entries in prval. Caller is responsible
3428  * for freeing up prval.
3429  */
3430 int
3431 aggr_grp_possible_mtu_range(aggr_grp_t *grp, mac_propval_uint32_range_t **prval,
3432     int *prmaxcnt, int *prcount)
3433 {
3434 	mac_propval_range_t		**vals;
3435 	aggr_port_t			*port;
3436 	mac_perim_handle_t		mph;
3437 	uint_t				i, numr;
3438 	int				err = 0;
3439 	size_t				sz_propval, sz_range32;
3440 	size_t				size;
3441 
3442 	sz_propval = sizeof (mac_propval_range_t);
3443 	sz_range32 = sizeof (mac_propval_uint32_range_t);
3444 
3445 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
3446 
3447 	vals = kmem_zalloc(sizeof (mac_propval_range_t *) * grp->lg_nports,
3448 	    KM_SLEEP);
3449 
3450 	for (port = grp->lg_ports, i = 0; port != NULL;
3451 	    port = port->lp_next, i++) {
3452 
3453 		size = sz_propval;
3454 		vals[i] = kmem_alloc(size, KM_SLEEP);
3455 		vals[i]->mpr_count = 1;
3456 
3457 		mac_perim_enter_by_mh(port->lp_mh, &mph);
3458 
3459 		err = mac_prop_info(port->lp_mh, MAC_PROP_MTU, NULL,
3460 		    NULL, 0, vals[i], NULL);
3461 		if (err == ENOSPC) {
3462 			/*
3463 			 * Not enough space to hold all ranges.
3464 			 * Allocate extra space as indicated and retry.
3465 			 */
3466 			numr = vals[i]->mpr_count;
3467 			kmem_free(vals[i], sz_propval);
3468 			size = sz_propval + (numr - 1) * sz_range32;
3469 			vals[i] = kmem_alloc(size, KM_SLEEP);
3470 			vals[i]->mpr_count = numr;
3471 			err = mac_prop_info(port->lp_mh, MAC_PROP_MTU, NULL,
3472 			    NULL, 0, vals[i], NULL);
3473 			ASSERT(err != ENOSPC);
3474 		}
3475 		mac_perim_exit(mph);
3476 		if (err != 0) {
3477 			kmem_free(vals[i], size);
3478 			vals[i] = NULL;
3479 			break;
3480 		}
3481 	}
3482 
3483 	/*
3484 	 * if any of the underlying ports does not support changing MTU then
3485 	 * just return ENOTSUP
3486 	 */
3487 	if (port != NULL) {
3488 		ASSERT(err != 0);
3489 		goto done;
3490 	}
3491 
3492 	aggr_mtu_range_intersection(vals, grp->lg_nports, prval, prmaxcnt,
3493 	    prcount);
3494 
3495 done:
3496 	for (i = 0; i < grp->lg_nports; i++) {
3497 		if (vals[i] != NULL) {
3498 			numr = vals[i]->mpr_count;
3499 			size = sz_propval + (numr - 1) * sz_range32;
3500 			kmem_free(vals[i], size);
3501 		}
3502 	}
3503 
3504 	kmem_free(vals, sizeof (mac_propval_range_t *) * grp->lg_nports);
3505 	return (err);
3506 }
3507 
3508 static void
3509 aggr_m_propinfo(void *m_driver, const char *pr_name, mac_prop_id_t pr_num,
3510     mac_prop_info_handle_t prh)
3511 {
3512 	aggr_grp_t			*grp = m_driver;
3513 	mac_propval_uint32_range_t	*rval = NULL;
3514 	int				i, rcount, rmaxcnt;
3515 	int				err = 0;
3516 
3517 	_NOTE(ARGUNUSED(pr_name));
3518 
3519 	switch (pr_num) {
3520 	case MAC_PROP_MTU:
3521 
3522 		err = aggr_grp_possible_mtu_range(grp, &rval, &rmaxcnt,
3523 		    &rcount);
3524 		if (err != 0) {
3525 			ASSERT(rval == NULL);
3526 			return;
3527 		}
3528 		for (i = 0; i < rcount; i++) {
3529 			mac_prop_info_set_range_uint32(prh,
3530 			    rval[i].mpur_min, rval[i].mpur_max);
3531 		}
3532 		kmem_free(rval, sizeof (mac_propval_uint32_range_t) * rmaxcnt);
3533 		break;
3534 	}
3535 }
3536