xref: /illumos-gate/usr/src/uts/common/io/aggr/aggr_grp.c (revision 2aeafac3612e19716bf8164f89c3c9196342979c)
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  */
25 
26 /*
27  * IEEE 802.3ad Link Aggregation -- Link Aggregation Groups.
28  *
29  * An instance of the structure aggr_grp_t is allocated for each
30  * link aggregation group. When created, aggr_grp_t objects are
31  * entered into the aggr_grp_hash hash table maintained by the modhash
32  * module. The hash key is the linkid associated with the link
33  * aggregation group.
34  *
35  * Each aggregation contains a set of ports. The port is represented
36  * by the aggr_port_t structure. A port consists of a single MAC
37  * client which has exclusive (MCIS_EXCLUSIVE) use of the underlying
38  * MAC. This client is used by the aggr to send and receive LACP
39  * traffic. Each port client takes on the same MAC unicast address --
40  * the address of the aggregation itself (taken from the first port by
41  * default).
42  *
43  * The MAC client that hangs off each aggr port is not your typical
44  * MAC client. Not only does it have exclusive control of the MAC, but
45  * it also has no Tx or Rx SRSes. An SRS is designed to queue and
46  * fanout traffic among L4 protocols; but the aggr is an intermediary,
47  * not a consumer. Instead of using SRSes, the aggr puts the
48  * underlying hardware rings into passthru mode and ships packets up
49  * via a direct call to aggr_recv_cb(). This allows aggr to enforce
50  * LACP while passing all other traffic up to clients of the aggr.
51  *
52  * Pseudo Rx Groups and Rings
53  * --------------------------
54  *
55  * It is imperative for client performance that the aggr provide as
56  * many MAC groups as possible. In order to use the underlying HW
57  * resources, aggr creates pseudo groups to aggregate the underlying
58  * HW groups. Every HW group gets mapped to a pseudo group; and every
59  * HW ring in that group gets mapped to a pseudo ring. The pseudo
60  * group at index 0 combines all the HW groups at index 0 from each
61  * port, etc. The aggr's MAC then creates normal MAC groups and rings
62  * out of these pseudo groups and rings to present to the aggr's
63  * clients. To the clients, the aggr's groups and rings are absolutely
64  * no different than a NIC's groups or rings.
65  *
66  * Pseudo Tx Rings
67  * ---------------
68  *
69  * The underlying ports (NICs) in an aggregation can have Tx rings. To
70  * enhance aggr's performance, these Tx rings are made available to
71  * the aggr layer as pseudo Tx rings. The concept of pseudo rings are
72  * not new. They are already present and implemented on the Rx side.
73  * The same concept is extended to the Tx side where each Tx ring of
74  * an underlying port is reflected in aggr as a pseudo Tx ring. Thus
75  * each pseudo Tx ring will map to a specific hardware Tx ring. Even
76  * in the case of a NIC that does not have a Tx ring, a pseudo Tx ring
77  * is given to the aggregation layer.
78  *
79  * With this change, the outgoing stack depth looks much better:
80  *
81  * mac_tx() -> mac_tx_aggr_mode() -> mac_tx_soft_ring_process() ->
82  * mac_tx_send() -> aggr_ring_rx() -> <driver>_ring_tx()
83  *
84  * Two new modes are introduced to mac_tx() to handle aggr pseudo Tx rings:
85  * SRS_TX_AGGR and SRS_TX_BW_AGGR.
86  *
87  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
88  * invokes an aggr function, aggr_find_tx_ring(), to find a (pseudo) Tx
89  * ring belonging to a port on which the packet has to be sent.
90  * aggr_find_tx_ring() first finds the outgoing port based on L2/L3/L4
91  * policy and then uses the fanout_hint passed to it to pick a Tx ring from
92  * the selected port.
93  *
94  * In SRS_TX_BW_AGGR mode, mac_tx_bw_mode() function is called where
95  * bandwidth limit is applied first on the outgoing packet and the packets
96  * allowed to go out would call mac_tx_aggr_mode() to send the packet on a
97  * particular Tx ring.
98  */
99 
100 #include <sys/types.h>
101 #include <sys/sysmacros.h>
102 #include <sys/conf.h>
103 #include <sys/cmn_err.h>
104 #include <sys/disp.h>
105 #include <sys/list.h>
106 #include <sys/ksynch.h>
107 #include <sys/kmem.h>
108 #include <sys/stream.h>
109 #include <sys/modctl.h>
110 #include <sys/ddi.h>
111 #include <sys/sunddi.h>
112 #include <sys/atomic.h>
113 #include <sys/stat.h>
114 #include <sys/modhash.h>
115 #include <sys/id_space.h>
116 #include <sys/strsun.h>
117 #include <sys/cred.h>
118 #include <sys/dlpi.h>
119 #include <sys/zone.h>
120 #include <sys/mac_provider.h>
121 #include <sys/dls.h>
122 #include <sys/vlan.h>
123 #include <sys/aggr.h>
124 #include <sys/aggr_impl.h>
125 
126 static int aggr_m_start(void *);
127 static void aggr_m_stop(void *);
128 static int aggr_m_promisc(void *, boolean_t);
129 static int aggr_m_multicst(void *, boolean_t, const uint8_t *);
130 static int aggr_m_unicst(void *, const uint8_t *);
131 static int aggr_m_stat(void *, uint_t, uint64_t *);
132 static void aggr_m_ioctl(void *, queue_t *, mblk_t *);
133 static boolean_t aggr_m_capab_get(void *, mac_capab_t, void *);
134 static int aggr_m_setprop(void *, const char *, mac_prop_id_t, uint_t,
135     const void *);
136 static void aggr_m_propinfo(void *, const char *, mac_prop_id_t,
137     mac_prop_info_handle_t);
138 
139 static aggr_port_t *aggr_grp_port_lookup(aggr_grp_t *, datalink_id_t);
140 static int aggr_grp_rem_port(aggr_grp_t *, aggr_port_t *, boolean_t *,
141     boolean_t *);
142 
143 static void aggr_grp_capab_set(aggr_grp_t *);
144 static boolean_t aggr_grp_capab_check(aggr_grp_t *, aggr_port_t *);
145 static uint_t aggr_grp_max_sdu(aggr_grp_t *);
146 static uint32_t aggr_grp_max_margin(aggr_grp_t *);
147 static boolean_t aggr_grp_sdu_check(aggr_grp_t *, aggr_port_t *);
148 static boolean_t aggr_grp_margin_check(aggr_grp_t *, aggr_port_t *);
149 
150 static int aggr_add_pseudo_rx_group(aggr_port_t *, aggr_pseudo_rx_group_t *);
151 static void aggr_rem_pseudo_rx_group(aggr_port_t *, aggr_pseudo_rx_group_t *);
152 static int aggr_pseudo_disable_intr(mac_intr_handle_t);
153 static int aggr_pseudo_enable_intr(mac_intr_handle_t);
154 static int aggr_pseudo_start_rx_ring(mac_ring_driver_t, uint64_t);
155 static void aggr_pseudo_stop_rx_ring(mac_ring_driver_t);
156 static int aggr_addmac(void *, const uint8_t *);
157 static int aggr_remmac(void *, const uint8_t *);
158 static int aggr_addvlan(mac_group_driver_t, uint16_t);
159 static int aggr_remvlan(mac_group_driver_t, uint16_t);
160 static mblk_t *aggr_rx_poll(void *, int);
161 static void aggr_fill_ring(void *, mac_ring_type_t, const int,
162     const int, mac_ring_info_t *, mac_ring_handle_t);
163 static void aggr_fill_group(void *, mac_ring_type_t, const int,
164     mac_group_info_t *, mac_group_handle_t);
165 
166 static kmem_cache_t	*aggr_grp_cache;
167 static mod_hash_t	*aggr_grp_hash;
168 static krwlock_t	aggr_grp_lock;
169 static uint_t		aggr_grp_cnt;
170 static id_space_t	*key_ids;
171 
172 #define	GRP_HASHSZ		64
173 #define	GRP_HASH_KEY(linkid)	((mod_hash_key_t)(uintptr_t)linkid)
174 #define	AGGR_PORT_NAME_DELIMIT '-'
175 
176 static uchar_t aggr_zero_mac[] = {0, 0, 0, 0, 0, 0};
177 
178 #define	AGGR_M_CALLBACK_FLAGS	\
179 	(MC_IOCTL | MC_GETCAPAB | MC_SETPROP | MC_PROPINFO)
180 
181 static mac_callbacks_t aggr_m_callbacks = {
182 	AGGR_M_CALLBACK_FLAGS,
183 	aggr_m_stat,
184 	aggr_m_start,
185 	aggr_m_stop,
186 	aggr_m_promisc,
187 	aggr_m_multicst,
188 	NULL,
189 	NULL,
190 	NULL,
191 	aggr_m_ioctl,
192 	aggr_m_capab_get,
193 	NULL,
194 	NULL,
195 	aggr_m_setprop,
196 	NULL,
197 	aggr_m_propinfo
198 };
199 
200 /*ARGSUSED*/
201 static int
202 aggr_grp_constructor(void *buf, void *arg, int kmflag)
203 {
204 	aggr_grp_t *grp = buf;
205 
206 	bzero(grp, sizeof (*grp));
207 	mutex_init(&grp->lg_lacp_lock, NULL, MUTEX_DEFAULT, NULL);
208 	cv_init(&grp->lg_lacp_cv, NULL, CV_DEFAULT, NULL);
209 	rw_init(&grp->lg_tx_lock, NULL, RW_DRIVER, NULL);
210 	mutex_init(&grp->lg_port_lock, NULL, MUTEX_DEFAULT, NULL);
211 	cv_init(&grp->lg_port_cv, NULL, CV_DEFAULT, NULL);
212 	mutex_init(&grp->lg_tx_flowctl_lock, NULL, MUTEX_DEFAULT, NULL);
213 	cv_init(&grp->lg_tx_flowctl_cv, NULL, CV_DEFAULT, NULL);
214 	grp->lg_link_state = LINK_STATE_UNKNOWN;
215 	return (0);
216 }
217 
218 /*ARGSUSED*/
219 static void
220 aggr_grp_destructor(void *buf, void *arg)
221 {
222 	aggr_grp_t *grp = buf;
223 
224 	if (grp->lg_tx_ports != NULL) {
225 		kmem_free(grp->lg_tx_ports,
226 		    grp->lg_tx_ports_size * sizeof (aggr_port_t *));
227 	}
228 
229 	mutex_destroy(&grp->lg_lacp_lock);
230 	cv_destroy(&grp->lg_lacp_cv);
231 	mutex_destroy(&grp->lg_port_lock);
232 	cv_destroy(&grp->lg_port_cv);
233 	rw_destroy(&grp->lg_tx_lock);
234 	mutex_destroy(&grp->lg_tx_flowctl_lock);
235 	cv_destroy(&grp->lg_tx_flowctl_cv);
236 }
237 
238 void
239 aggr_grp_init(void)
240 {
241 	aggr_grp_cache = kmem_cache_create("aggr_grp_cache",
242 	    sizeof (aggr_grp_t), 0, aggr_grp_constructor,
243 	    aggr_grp_destructor, NULL, NULL, NULL, 0);
244 
245 	aggr_grp_hash = mod_hash_create_idhash("aggr_grp_hash",
246 	    GRP_HASHSZ, mod_hash_null_valdtor);
247 	rw_init(&aggr_grp_lock, NULL, RW_DEFAULT, NULL);
248 	aggr_grp_cnt = 0;
249 
250 	/*
251 	 * Allocate an id space to manage key values (when key is not
252 	 * specified). The range of the id space will be from
253 	 * (AGGR_MAX_KEY + 1) to UINT16_MAX, because the LACP protocol
254 	 * uses a 16-bit key.
255 	 */
256 	key_ids = id_space_create("aggr_key_ids", AGGR_MAX_KEY + 1, UINT16_MAX);
257 	ASSERT(key_ids != NULL);
258 }
259 
260 void
261 aggr_grp_fini(void)
262 {
263 	id_space_destroy(key_ids);
264 	rw_destroy(&aggr_grp_lock);
265 	mod_hash_destroy_idhash(aggr_grp_hash);
266 	kmem_cache_destroy(aggr_grp_cache);
267 }
268 
269 uint_t
270 aggr_grp_count(void)
271 {
272 	uint_t	count;
273 
274 	rw_enter(&aggr_grp_lock, RW_READER);
275 	count = aggr_grp_cnt;
276 	rw_exit(&aggr_grp_lock);
277 	return (count);
278 }
279 
280 /*
281  * Since both aggr_port_notify_cb() and aggr_port_timer_thread() functions
282  * requires the mac perimeter, this function holds a reference of the aggr
283  * and aggr won't call mac_unregister() until this reference drops to 0.
284  */
285 void
286 aggr_grp_port_hold(aggr_port_t *port)
287 {
288 	aggr_grp_t	*grp = port->lp_grp;
289 
290 	AGGR_PORT_REFHOLD(port);
291 	mutex_enter(&grp->lg_port_lock);
292 	grp->lg_port_ref++;
293 	mutex_exit(&grp->lg_port_lock);
294 }
295 
296 /*
297  * Release the reference of the grp and inform aggr_grp_delete() calling
298  * mac_unregister() is now safe.
299  */
300 void
301 aggr_grp_port_rele(aggr_port_t *port)
302 {
303 	aggr_grp_t	*grp = port->lp_grp;
304 
305 	mutex_enter(&grp->lg_port_lock);
306 	if (--grp->lg_port_ref == 0)
307 		cv_signal(&grp->lg_port_cv);
308 	mutex_exit(&grp->lg_port_lock);
309 	AGGR_PORT_REFRELE(port);
310 }
311 
312 /*
313  * Wait for the port's lacp timer thread and the port's notification callback
314  * to exit.
315  */
316 void
317 aggr_grp_port_wait(aggr_grp_t *grp)
318 {
319 	mutex_enter(&grp->lg_port_lock);
320 	if (grp->lg_port_ref != 0)
321 		cv_wait(&grp->lg_port_cv, &grp->lg_port_lock);
322 	mutex_exit(&grp->lg_port_lock);
323 }
324 
325 /*
326  * Attach a port to a link aggregation group.
327  *
328  * A port is attached to a link aggregation group once its speed
329  * and link state have been verified.
330  *
331  * Returns B_TRUE if the group link state or speed has changed. If
332  * it's the case, the caller must notify the MAC layer via a call
333  * to mac_link().
334  */
335 boolean_t
336 aggr_grp_attach_port(aggr_grp_t *grp, aggr_port_t *port)
337 {
338 	boolean_t link_state_changed = B_FALSE;
339 
340 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
341 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
342 
343 	if (port->lp_state == AGGR_PORT_STATE_ATTACHED)
344 		return (B_FALSE);
345 
346 	/*
347 	 * Validate the MAC port link speed and update the group
348 	 * link speed if needed.
349 	 */
350 	if (port->lp_ifspeed == 0 ||
351 	    port->lp_link_state != LINK_STATE_UP ||
352 	    port->lp_link_duplex != LINK_DUPLEX_FULL) {
353 		/*
354 		 * Can't attach a MAC port with unknown link speed,
355 		 * down link, or not in full duplex mode.
356 		 */
357 		return (B_FALSE);
358 	}
359 
360 	mutex_enter(&grp->lg_stat_lock);
361 	if (grp->lg_ifspeed == 0) {
362 		/*
363 		 * The group inherits the speed of the first link being
364 		 * attached.
365 		 */
366 		grp->lg_ifspeed = port->lp_ifspeed;
367 		link_state_changed = B_TRUE;
368 	} else if (grp->lg_ifspeed != port->lp_ifspeed) {
369 		/*
370 		 * The link speed of the MAC port must be the same as
371 		 * the group link speed, as per 802.3ad. Since it is
372 		 * not, the attach is cancelled.
373 		 */
374 		mutex_exit(&grp->lg_stat_lock);
375 		return (B_FALSE);
376 	}
377 	mutex_exit(&grp->lg_stat_lock);
378 
379 	grp->lg_nattached_ports++;
380 
381 	/*
382 	 * Update the group link state.
383 	 */
384 	if (grp->lg_link_state != LINK_STATE_UP) {
385 		grp->lg_link_state = LINK_STATE_UP;
386 		mutex_enter(&grp->lg_stat_lock);
387 		grp->lg_link_duplex = LINK_DUPLEX_FULL;
388 		mutex_exit(&grp->lg_stat_lock);
389 		link_state_changed = B_TRUE;
390 	}
391 
392 	/*
393 	 * Update port's state.
394 	 */
395 	port->lp_state = AGGR_PORT_STATE_ATTACHED;
396 
397 	aggr_grp_multicst_port(port, B_TRUE);
398 
399 	/*
400 	 * The port client doesn't have an Rx SRS; instead of calling
401 	 * mac_rx_set() we set the client's flow callback directly.
402 	 * This datapath is used only when the port's driver doesn't
403 	 * support MAC_CAPAB_RINGS. Drivers with ring support will
404 	 * deliver traffic to the aggr via ring passthru.
405 	 */
406 	mac_client_set_flow_cb(port->lp_mch, aggr_recv_cb, port);
407 
408 	/*
409 	 * If LACP is OFF, the port can be used to send data as soon
410 	 * as its link is up and verified to be compatible with the
411 	 * aggregation.
412 	 *
413 	 * If LACP is active or passive, notify the LACP subsystem, which
414 	 * will enable sending on the port following the LACP protocol.
415 	 */
416 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
417 		aggr_send_port_enable(port);
418 	else
419 		aggr_lacp_port_attached(port);
420 
421 	return (link_state_changed);
422 }
423 
424 boolean_t
425 aggr_grp_detach_port(aggr_grp_t *grp, aggr_port_t *port)
426 {
427 	boolean_t link_state_changed = B_FALSE;
428 
429 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
430 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
431 
432 	/* update state */
433 	if (port->lp_state != AGGR_PORT_STATE_ATTACHED)
434 		return (B_FALSE);
435 
436 	mac_client_clear_flow_cb(port->lp_mch);
437 
438 	aggr_grp_multicst_port(port, B_FALSE);
439 
440 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
441 		aggr_send_port_disable(port);
442 	else
443 		aggr_lacp_port_detached(port);
444 
445 	port->lp_state = AGGR_PORT_STATE_STANDBY;
446 
447 	grp->lg_nattached_ports--;
448 	if (grp->lg_nattached_ports == 0) {
449 		/* the last attached MAC port of the group is being detached */
450 		grp->lg_link_state = LINK_STATE_DOWN;
451 		mutex_enter(&grp->lg_stat_lock);
452 		grp->lg_ifspeed = 0;
453 		grp->lg_link_duplex = LINK_DUPLEX_UNKNOWN;
454 		mutex_exit(&grp->lg_stat_lock);
455 		link_state_changed = B_TRUE;
456 	}
457 
458 	return (link_state_changed);
459 }
460 
461 /*
462  * Update the MAC addresses of the constituent ports of the specified
463  * group. This function is invoked:
464  * - after creating a new aggregation group.
465  * - after adding new ports to an aggregation group.
466  * - after removing a port from a group when the MAC address of
467  *   that port was used for the MAC address of the group.
468  * - after the MAC address of a port changed when the MAC address
469  *   of that port was used for the MAC address of the group.
470  *
471  * Return true if the link state of the aggregation changed, for example
472  * as a result of a failure changing the MAC address of one of the
473  * constituent ports.
474  */
475 boolean_t
476 aggr_grp_update_ports_mac(aggr_grp_t *grp)
477 {
478 	aggr_port_t *cport;
479 	boolean_t link_state_changed = B_FALSE;
480 	mac_perim_handle_t mph;
481 
482 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
483 
484 	for (cport = grp->lg_ports; cport != NULL;
485 	    cport = cport->lp_next) {
486 		mac_perim_enter_by_mh(cport->lp_mh, &mph);
487 		if (aggr_port_unicst(cport) != 0) {
488 			if (aggr_grp_detach_port(grp, cport))
489 				link_state_changed = B_TRUE;
490 		} else {
491 			/*
492 			 * If a port was detached because of a previous
493 			 * failure changing the MAC address, the port is
494 			 * reattached when it successfully changes the MAC
495 			 * address now, and this might cause the link state
496 			 * of the aggregation to change.
497 			 */
498 			if (aggr_grp_attach_port(grp, cport))
499 				link_state_changed = B_TRUE;
500 		}
501 		mac_perim_exit(mph);
502 	}
503 	return (link_state_changed);
504 }
505 
506 /*
507  * Invoked when the MAC address of a port has changed. If the port's
508  * MAC address was used for the group MAC address, set mac_addr_changedp
509  * to B_TRUE to indicate to the caller that it should send a MAC_NOTE_UNICST
510  * notification. If the link state changes due to detach/attach of
511  * the constituent port, set link_state_changedp to B_TRUE to indicate
512  * to the caller that it should send a MAC_NOTE_LINK notification. In both
513  * cases, it is the responsibility of the caller to invoke notification
514  * functions after releasing the the port lock.
515  */
516 void
517 aggr_grp_port_mac_changed(aggr_grp_t *grp, aggr_port_t *port,
518     boolean_t *mac_addr_changedp, boolean_t *link_state_changedp)
519 {
520 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
521 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
522 	ASSERT(mac_addr_changedp != NULL);
523 	ASSERT(link_state_changedp != NULL);
524 
525 	*mac_addr_changedp = B_FALSE;
526 	*link_state_changedp = B_FALSE;
527 
528 	if (grp->lg_addr_fixed) {
529 		/*
530 		 * The group is using a fixed MAC address or an automatic
531 		 * MAC address has not been set.
532 		 */
533 		return;
534 	}
535 
536 	if (grp->lg_mac_addr_port == port) {
537 		/*
538 		 * The MAC address of the port was assigned to the group
539 		 * MAC address. Update the group MAC address.
540 		 */
541 		bcopy(port->lp_addr, grp->lg_addr, ETHERADDRL);
542 		*mac_addr_changedp = B_TRUE;
543 	} else {
544 		/*
545 		 * Update the actual port MAC address to the MAC address
546 		 * of the group.
547 		 */
548 		if (aggr_port_unicst(port) != 0) {
549 			*link_state_changedp = aggr_grp_detach_port(grp, port);
550 		} else {
551 			/*
552 			 * If a port was detached because of a previous
553 			 * failure changing the MAC address, the port is
554 			 * reattached when it successfully changes the MAC
555 			 * address now, and this might cause the link state
556 			 * of the aggregation to change.
557 			 */
558 			*link_state_changedp = aggr_grp_attach_port(grp, port);
559 		}
560 	}
561 }
562 
563 /*
564  * Add a port to a link aggregation group.
565  */
566 static int
567 aggr_grp_add_port(aggr_grp_t *grp, datalink_id_t port_linkid, boolean_t force,
568     aggr_port_t **pp)
569 {
570 	aggr_port_t *port, **cport;
571 	mac_perim_handle_t mph;
572 	zoneid_t port_zoneid = ALL_ZONES;
573 	int err;
574 
575 	/* The port must be in the same zone as the aggregation. */
576 	if (zone_check_datalink(&port_zoneid, port_linkid) != 0)
577 		port_zoneid = GLOBAL_ZONEID;
578 	if (grp->lg_zoneid != port_zoneid)
579 		return (EBUSY);
580 
581 	/*
582 	 * If we are creating the aggr, then there is no MAC handle
583 	 * and thus no perimeter to hold. If we are adding a port to
584 	 * an existing aggr, then the perimiter of the aggr's MAC must
585 	 * be held.
586 	 */
587 	ASSERT(grp->lg_mh == NULL || MAC_PERIM_HELD(grp->lg_mh));
588 
589 	err = aggr_port_create(grp, port_linkid, force, &port);
590 	if (err != 0)
591 		return (err);
592 
593 	mac_perim_enter_by_mh(port->lp_mh, &mph);
594 
595 	/* Add the new port to the end of the list. */
596 	cport = &grp->lg_ports;
597 	while (*cport != NULL)
598 		cport = &((*cport)->lp_next);
599 	*cport = port;
600 
601 	/*
602 	 * Back reference to the group it is member of. A port always
603 	 * holds a reference to its group to ensure that the back
604 	 * reference is always valid.
605 	 */
606 	port->lp_grp = grp;
607 	AGGR_GRP_REFHOLD(grp);
608 	grp->lg_nports++;
609 
610 	aggr_lacp_init_port(port);
611 	mac_perim_exit(mph);
612 
613 	if (pp != NULL)
614 		*pp = port;
615 
616 	return (0);
617 }
618 
619 /*
620  * This is called in response to either our LACP state machine or a MAC
621  * notification that the link has gone down via aggr_send_port_disable(). At
622  * this point, we may need to update our default ring. To that end, we go
623  * through the set of ports (underlying datalinks in an aggregation) that are
624  * currently enabled to transmit data. If all our links have been disabled for
625  * transmit, then we don't do anything.
626  *
627  * Note, because we only have a single TX group, we don't have to worry about
628  * the rings moving between groups and the chance that mac will reassign it
629  * unless someone removes a port, at which point, we play it safe and call this
630  * again.
631  */
632 void
633 aggr_grp_update_default(aggr_grp_t *grp)
634 {
635 	aggr_port_t *port;
636 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
637 
638 	rw_enter(&grp->lg_tx_lock, RW_WRITER);
639 
640 	if (grp->lg_ntx_ports == 0) {
641 		rw_exit(&grp->lg_tx_lock);
642 		return;
643 	}
644 
645 	port = grp->lg_tx_ports[0];
646 	ASSERT(port->lp_tx_ring_cnt > 0);
647 	mac_hwring_set_default(grp->lg_mh, port->lp_pseudo_tx_rings[0]);
648 	rw_exit(&grp->lg_tx_lock);
649 }
650 
651 /*
652  * Add a pseudo RX ring for the given HW ring handle.
653  */
654 static int
655 aggr_add_pseudo_rx_ring(aggr_port_t *port,
656     aggr_pseudo_rx_group_t *rx_grp, mac_ring_handle_t hw_rh)
657 {
658 	aggr_pseudo_rx_ring_t	*ring;
659 	int			err;
660 	int			j;
661 
662 	for (j = 0; j < MAX_RINGS_PER_GROUP; j++) {
663 		ring = rx_grp->arg_rings + j;
664 		if (!(ring->arr_flags & MAC_PSEUDO_RING_INUSE))
665 			break;
666 	}
667 
668 	/*
669 	 * No slot for this new RX ring.
670 	 */
671 	if (j == MAX_RINGS_PER_GROUP)
672 		return (EIO);
673 
674 	ring->arr_flags |= MAC_PSEUDO_RING_INUSE;
675 	ring->arr_hw_rh = hw_rh;
676 	ring->arr_port = port;
677 	ring->arr_grp = rx_grp;
678 	rx_grp->arg_ring_cnt++;
679 
680 	/*
681 	 * The group is already registered, dynamically add a new ring to the
682 	 * mac group.
683 	 */
684 	if ((err = mac_group_add_ring(rx_grp->arg_gh, j)) != 0) {
685 		ring->arr_flags &= ~MAC_PSEUDO_RING_INUSE;
686 		ring->arr_hw_rh = NULL;
687 		ring->arr_port = NULL;
688 		ring->arr_grp = NULL;
689 		rx_grp->arg_ring_cnt--;
690 	} else {
691 		/*
692 		 * This must run after the MAC is registered.
693 		 */
694 		ASSERT3P(ring->arr_rh, !=, NULL);
695 		mac_hwring_set_passthru(hw_rh, (mac_rx_t)aggr_recv_cb,
696 		    (void *)port, (mac_resource_handle_t)ring);
697 	}
698 	return (err);
699 }
700 
701 /*
702  * Remove the pseudo RX ring of the given HW ring handle.
703  */
704 static void
705 aggr_rem_pseudo_rx_ring(aggr_pseudo_rx_group_t *rx_grp, mac_ring_handle_t hw_rh)
706 {
707 	for (uint_t j = 0; j < MAX_RINGS_PER_GROUP; j++) {
708 		aggr_pseudo_rx_ring_t *ring = rx_grp->arg_rings + j;
709 
710 		if (!(ring->arr_flags & MAC_PSEUDO_RING_INUSE) ||
711 		    ring->arr_hw_rh != hw_rh) {
712 			continue;
713 		}
714 
715 		mac_group_rem_ring(rx_grp->arg_gh, ring->arr_rh);
716 
717 		ring->arr_flags &= ~MAC_PSEUDO_RING_INUSE;
718 		ring->arr_hw_rh = NULL;
719 		ring->arr_port = NULL;
720 		ring->arr_grp = NULL;
721 		rx_grp->arg_ring_cnt--;
722 		mac_hwring_clear_passthru(hw_rh);
723 		break;
724 	}
725 }
726 
727 /*
728  * Create pseudo rings over the HW rings of the port.
729  *
730  * o Create a pseudo ring in rx_grp per HW ring in the port's HW group.
731  *
732  * o Program existing unicast filters on the pseudo group into the HW group.
733  *
734  * o Program existing VLAN filters on the pseudo group into the HW group.
735  */
736 static int
737 aggr_add_pseudo_rx_group(aggr_port_t *port, aggr_pseudo_rx_group_t *rx_grp)
738 {
739 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP];
740 	aggr_unicst_addr_t	*addr, *a;
741 	mac_perim_handle_t	pmph;
742 	aggr_vlan_t		*avp;
743 	uint_t			hw_rh_cnt, i;
744 	int			err = 0;
745 	uint_t			g_idx = rx_grp->arg_index;
746 
747 	ASSERT(MAC_PERIM_HELD(port->lp_grp->lg_mh));
748 	ASSERT3U(g_idx, <, MAX_GROUPS_PER_PORT);
749 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
750 
751 	i = 0;
752 	addr = NULL;
753 	/*
754 	 * This function must be called after the aggr registers its
755 	 * MAC and its Rx groups have been initialized.
756 	 */
757 	ASSERT(rx_grp->arg_gh != NULL);
758 
759 	/*
760 	 * Get the list of the underlying HW rings.
761 	 */
762 	hw_rh_cnt = mac_hwrings_idx_get(port->lp_mh, g_idx,
763 	    &port->lp_hwghs[g_idx], hw_rh, MAC_RING_TYPE_RX);
764 
765 	/*
766 	 * Add existing VLAN and unicast address filters to the port.
767 	 */
768 	for (avp = list_head(&rx_grp->arg_vlans); avp != NULL;
769 	    avp = list_next(&rx_grp->arg_vlans, avp)) {
770 		if ((err = aggr_port_addvlan(port, g_idx, avp->av_vid)) != 0)
771 			goto err;
772 	}
773 
774 	for (addr = rx_grp->arg_macaddr; addr != NULL; addr = addr->aua_next) {
775 		if ((err = aggr_port_addmac(port, g_idx, addr->aua_addr)) != 0)
776 			goto err;
777 	}
778 
779 	for (i = 0; i < hw_rh_cnt; i++) {
780 		err = aggr_add_pseudo_rx_ring(port, rx_grp, hw_rh[i]);
781 		if (err != 0)
782 			goto err;
783 	}
784 
785 	mac_perim_exit(pmph);
786 	return (0);
787 
788 err:
789 	ASSERT(err != 0);
790 
791 	for (uint_t j = 0; j < i; j++)
792 		aggr_rem_pseudo_rx_ring(rx_grp, hw_rh[j]);
793 
794 	for (a = rx_grp->arg_macaddr; a != addr; a = a->aua_next)
795 		aggr_port_remmac(port, g_idx, a->aua_addr);
796 
797 	if (avp != NULL)
798 		avp = list_prev(&rx_grp->arg_vlans, avp);
799 
800 	for (; avp != NULL; avp = list_prev(&rx_grp->arg_vlans, avp)) {
801 		int err2;
802 
803 		if ((err2 = aggr_port_remvlan(port, g_idx, avp->av_vid)) != 0) {
804 			cmn_err(CE_WARN, "Failed to remove VLAN %u from port %s"
805 			    ": errno %d.", avp->av_vid,
806 			    mac_client_name(port->lp_mch), err2);
807 		}
808 	}
809 
810 	port->lp_hwghs[g_idx] = NULL;
811 	mac_perim_exit(pmph);
812 	return (err);
813 }
814 
815 /*
816  * Destroy the pseudo rings mapping to this port and remove all VLAN
817  * and unicast filters from this port. Even if there are no underlying
818  * HW rings we must still remove the unicast filters to take the port
819  * out of promisc mode.
820  */
821 static void
822 aggr_rem_pseudo_rx_group(aggr_port_t *port, aggr_pseudo_rx_group_t *rx_grp)
823 {
824 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP];
825 	aggr_unicst_addr_t	*addr;
826 	mac_perim_handle_t	pmph;
827 	uint_t			hw_rh_cnt;
828 	uint_t			g_idx = rx_grp->arg_index;
829 
830 	ASSERT(MAC_PERIM_HELD(port->lp_grp->lg_mh));
831 	ASSERT3U(g_idx, <, MAX_GROUPS_PER_PORT);
832 	ASSERT3P(rx_grp->arg_gh, !=, NULL);
833 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
834 
835 	hw_rh_cnt = mac_hwrings_idx_get(port->lp_mh, g_idx, NULL, hw_rh,
836 	    MAC_RING_TYPE_RX);
837 
838 	for (uint_t i = 0; i < hw_rh_cnt; i++)
839 		aggr_rem_pseudo_rx_ring(rx_grp, hw_rh[i]);
840 
841 	for (addr = rx_grp->arg_macaddr; addr != NULL; addr = addr->aua_next)
842 		aggr_port_remmac(port, g_idx, addr->aua_addr);
843 
844 	for (aggr_vlan_t *avp = list_head(&rx_grp->arg_vlans); avp != NULL;
845 	    avp = list_next(&rx_grp->arg_vlans, avp)) {
846 		int err;
847 
848 		if ((err = aggr_port_remvlan(port, g_idx, avp->av_vid)) != 0) {
849 			cmn_err(CE_WARN, "Failed to remove VLAN %u from port %s"
850 			    ": errno %d.", avp->av_vid,
851 			    mac_client_name(port->lp_mch), err);
852 		}
853 	}
854 
855 	port->lp_hwghs[g_idx] = NULL;
856 	mac_perim_exit(pmph);
857 }
858 
859 /*
860  * Add a pseudo TX ring for the given HW ring handle.
861  */
862 static int
863 aggr_add_pseudo_tx_ring(aggr_port_t *port,
864     aggr_pseudo_tx_group_t *tx_grp, mac_ring_handle_t hw_rh,
865     mac_ring_handle_t *pseudo_rh)
866 {
867 	aggr_pseudo_tx_ring_t	*ring;
868 	int			err;
869 	int			i;
870 
871 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
872 	for (i = 0; i < MAX_RINGS_PER_GROUP; i++) {
873 		ring = tx_grp->atg_rings + i;
874 		if (!(ring->atr_flags & MAC_PSEUDO_RING_INUSE))
875 			break;
876 	}
877 	/*
878 	 * No slot for this new TX ring.
879 	 */
880 	if (i == MAX_RINGS_PER_GROUP)
881 		return (EIO);
882 	/*
883 	 * The following 4 statements needs to be done before
884 	 * calling mac_group_add_ring(). Otherwise it will
885 	 * result in an assertion failure in mac_init_ring().
886 	 */
887 	ring->atr_flags |= MAC_PSEUDO_RING_INUSE;
888 	ring->atr_hw_rh = hw_rh;
889 	ring->atr_port = port;
890 	tx_grp->atg_ring_cnt++;
891 
892 	/*
893 	 * The TX side has no concept of ring groups unlike RX groups.
894 	 * There is just a single group which stores all the TX rings.
895 	 * This group will be used to store aggr's pseudo TX rings.
896 	 */
897 	if ((err = mac_group_add_ring(tx_grp->atg_gh, i)) != 0) {
898 		ring->atr_flags &= ~MAC_PSEUDO_RING_INUSE;
899 		ring->atr_hw_rh = NULL;
900 		ring->atr_port = NULL;
901 		tx_grp->atg_ring_cnt--;
902 	} else {
903 		*pseudo_rh = mac_find_ring(tx_grp->atg_gh, i);
904 		if (hw_rh != NULL) {
905 			mac_hwring_setup(hw_rh, (mac_resource_handle_t)ring,
906 			    mac_find_ring(tx_grp->atg_gh, i));
907 		}
908 	}
909 
910 	return (err);
911 }
912 
913 /*
914  * Remove the pseudo TX ring of the given HW ring handle.
915  */
916 static void
917 aggr_rem_pseudo_tx_ring(aggr_pseudo_tx_group_t *tx_grp,
918     mac_ring_handle_t pseudo_hw_rh)
919 {
920 	aggr_pseudo_tx_ring_t	*ring;
921 	int			i;
922 
923 	for (i = 0; i < MAX_RINGS_PER_GROUP; i++) {
924 		ring = tx_grp->atg_rings + i;
925 		if (ring->atr_rh != pseudo_hw_rh)
926 			continue;
927 
928 		ASSERT(ring->atr_flags & MAC_PSEUDO_RING_INUSE);
929 		mac_group_rem_ring(tx_grp->atg_gh, pseudo_hw_rh);
930 		ring->atr_flags &= ~MAC_PSEUDO_RING_INUSE;
931 		mac_hwring_teardown(ring->atr_hw_rh);
932 		ring->atr_hw_rh = NULL;
933 		ring->atr_port = NULL;
934 		tx_grp->atg_ring_cnt--;
935 		break;
936 	}
937 }
938 
939 /*
940  * This function is called to create pseudo rings over hardware rings of
941  * the underlying device. There is a 1:1 mapping between the pseudo TX
942  * rings of the aggr and the hardware rings of the underlying port.
943  */
944 static int
945 aggr_add_pseudo_tx_group(aggr_port_t *port, aggr_pseudo_tx_group_t *tx_grp)
946 {
947 	aggr_grp_t		*grp = port->lp_grp;
948 	mac_ring_handle_t	hw_rh[MAX_RINGS_PER_GROUP], pseudo_rh;
949 	mac_perim_handle_t	pmph;
950 	int			hw_rh_cnt, i = 0, j;
951 	int			err = 0;
952 
953 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
954 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
955 
956 	/*
957 	 * Get the list the the underlying HW rings.
958 	 */
959 	hw_rh_cnt = mac_hwrings_get(port->lp_mch, NULL, hw_rh,
960 	    MAC_RING_TYPE_TX);
961 
962 	/*
963 	 * Even if the underlying NIC does not have TX rings, we
964 	 * still make a psuedo TX ring for that NIC with NULL as
965 	 * the ring handle.
966 	 */
967 	if (hw_rh_cnt == 0)
968 		port->lp_tx_ring_cnt = 1;
969 	else
970 		port->lp_tx_ring_cnt = hw_rh_cnt;
971 
972 	port->lp_tx_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
973 	    port->lp_tx_ring_cnt), KM_SLEEP);
974 	port->lp_pseudo_tx_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
975 	    port->lp_tx_ring_cnt), KM_SLEEP);
976 
977 	if (hw_rh_cnt == 0) {
978 		if ((err = aggr_add_pseudo_tx_ring(port, tx_grp,
979 		    NULL, &pseudo_rh)) == 0) {
980 			port->lp_tx_rings[0] = NULL;
981 			port->lp_pseudo_tx_rings[0] = pseudo_rh;
982 		}
983 	} else {
984 		for (i = 0; err == 0 && i < hw_rh_cnt; i++) {
985 			err = aggr_add_pseudo_tx_ring(port,
986 			    tx_grp, hw_rh[i], &pseudo_rh);
987 			if (err != 0)
988 				break;
989 			port->lp_tx_rings[i] = hw_rh[i];
990 			port->lp_pseudo_tx_rings[i] = pseudo_rh;
991 		}
992 	}
993 
994 	if (err != 0) {
995 		if (hw_rh_cnt != 0) {
996 			for (j = 0; j < i; j++) {
997 				aggr_rem_pseudo_tx_ring(tx_grp,
998 				    port->lp_pseudo_tx_rings[j]);
999 			}
1000 		}
1001 		kmem_free(port->lp_tx_rings,
1002 		    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_cnt));
1003 		kmem_free(port->lp_pseudo_tx_rings,
1004 		    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_cnt));
1005 		port->lp_tx_ring_cnt = 0;
1006 	} else {
1007 		port->lp_tx_grp_added = B_TRUE;
1008 		port->lp_tx_notify_mh = mac_client_tx_notify(port->lp_mch,
1009 		    aggr_tx_ring_update, port);
1010 	}
1011 	mac_perim_exit(pmph);
1012 	aggr_grp_update_default(grp);
1013 	return (err);
1014 }
1015 
1016 /*
1017  * This function is called by aggr to remove pseudo TX rings over the
1018  * HW rings of the underlying port.
1019  */
1020 static void
1021 aggr_rem_pseudo_tx_group(aggr_port_t *port, aggr_pseudo_tx_group_t *tx_grp)
1022 {
1023 	aggr_grp_t		*grp = port->lp_grp;
1024 	mac_perim_handle_t	pmph;
1025 	int			i;
1026 
1027 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1028 	mac_perim_enter_by_mh(port->lp_mh, &pmph);
1029 
1030 	if (!port->lp_tx_grp_added)
1031 		goto done;
1032 
1033 	ASSERT(tx_grp->atg_gh != NULL);
1034 
1035 	for (i = 0; i < port->lp_tx_ring_cnt; i++)
1036 		aggr_rem_pseudo_tx_ring(tx_grp, port->lp_pseudo_tx_rings[i]);
1037 
1038 	kmem_free(port->lp_tx_rings,
1039 	    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_cnt));
1040 	kmem_free(port->lp_pseudo_tx_rings,
1041 	    (sizeof (mac_ring_handle_t *) * port->lp_tx_ring_cnt));
1042 
1043 	port->lp_tx_ring_cnt = 0;
1044 	(void) mac_client_tx_notify(port->lp_mch, NULL, port->lp_tx_notify_mh);
1045 	port->lp_tx_grp_added = B_FALSE;
1046 	aggr_grp_update_default(grp);
1047 done:
1048 	mac_perim_exit(pmph);
1049 }
1050 
1051 static int
1052 aggr_pseudo_disable_intr(mac_intr_handle_t ih)
1053 {
1054 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)ih;
1055 	return (mac_hwring_disable_intr(rr_ring->arr_hw_rh));
1056 }
1057 
1058 static int
1059 aggr_pseudo_enable_intr(mac_intr_handle_t ih)
1060 {
1061 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)ih;
1062 	return (mac_hwring_enable_intr(rr_ring->arr_hw_rh));
1063 }
1064 
1065 /*
1066  * Start the pseudo ring. Since the pseudo ring is just an abstraction
1067  * over an actual HW ring, the real task is to start the underlying HW
1068  * ring.
1069  */
1070 static int
1071 aggr_pseudo_start_rx_ring(mac_ring_driver_t arg, uint64_t mr_gen)
1072 {
1073 	int err;
1074 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)arg;
1075 
1076 	err = mac_hwring_start(rr_ring->arr_hw_rh);
1077 
1078 	if (err != 0)
1079 		return (err);
1080 
1081 	rr_ring->arr_gen = mr_gen;
1082 	return (err);
1083 }
1084 
1085 /*
1086  * Stop the pseudo ring. Since the pseudo ring is just an abstraction
1087  * over an actual HW ring, the real task is to stop the underlying HW
1088  * ring.
1089  */
1090 static void
1091 aggr_pseudo_stop_rx_ring(mac_ring_driver_t arg)
1092 {
1093 	aggr_pseudo_rx_ring_t *rr_ring = (aggr_pseudo_rx_ring_t *)arg;
1094 
1095 	/*
1096 	 * The rings underlying the default group must stay up to
1097 	 * continue receiving LACP traffic. We would normally never
1098 	 * stop the default Rx rings because of the primary MAC
1099 	 * client; but aggr's primary MAC client doesn't call
1100 	 * mac_unicast_add() and thus mi_active is 0 when the last
1101 	 * non-primary client is deleted.
1102 	 */
1103 	if (rr_ring->arr_grp->arg_index != 0)
1104 		mac_hwring_stop(rr_ring->arr_hw_rh);
1105 }
1106 
1107 /*
1108  * Add one or more ports to an existing link aggregation group.
1109  */
1110 int
1111 aggr_grp_add_ports(datalink_id_t linkid, uint_t nports, boolean_t force,
1112     laioc_port_t *ports)
1113 {
1114 	int rc;
1115 	uint_t port_added = 0;
1116 	uint_t grp_added;
1117 	aggr_grp_t *grp = NULL;
1118 	aggr_port_t *port;
1119 	boolean_t link_state_changed = B_FALSE;
1120 	mac_perim_handle_t mph, pmph;
1121 
1122 	/* Get the aggr corresponding to linkid. */
1123 	rw_enter(&aggr_grp_lock, RW_READER);
1124 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1125 	    (mod_hash_val_t *)&grp) != 0) {
1126 		rw_exit(&aggr_grp_lock);
1127 		return (ENOENT);
1128 	}
1129 	AGGR_GRP_REFHOLD(grp);
1130 
1131 	/*
1132 	 * Hold the perimeter so that the aggregation can't be destroyed.
1133 	 */
1134 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1135 	rw_exit(&aggr_grp_lock);
1136 
1137 	/* Add the specified ports to the aggr. */
1138 	for (uint_t i = 0; i < nports; i++) {
1139 		grp_added = 0;
1140 
1141 		if ((rc = aggr_grp_add_port(grp, ports[i].lp_linkid,
1142 		    force, &port)) != 0) {
1143 			goto bail;
1144 		}
1145 
1146 		ASSERT(port != NULL);
1147 		port_added++;
1148 
1149 		/* check capabilities */
1150 		if (!aggr_grp_capab_check(grp, port) ||
1151 		    !aggr_grp_sdu_check(grp, port) ||
1152 		    !aggr_grp_margin_check(grp, port)) {
1153 			rc = ENOTSUP;
1154 			goto bail;
1155 		}
1156 
1157 		/*
1158 		 * Create the pseudo ring for each HW ring of the underlying
1159 		 * port.
1160 		 */
1161 		rc = aggr_add_pseudo_tx_group(port, &grp->lg_tx_group);
1162 		if (rc != 0)
1163 			goto bail;
1164 
1165 		for (uint_t j = 0; j < grp->lg_rx_group_count; j++) {
1166 			rc = aggr_add_pseudo_rx_group(port,
1167 			    &grp->lg_rx_groups[j]);
1168 
1169 			if (rc != 0)
1170 				goto bail;
1171 
1172 			grp_added++;
1173 		}
1174 
1175 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
1176 
1177 		/* set LACP mode */
1178 		aggr_port_lacp_set_mode(grp, port);
1179 
1180 		/* start port if group has already been started */
1181 		if (grp->lg_started) {
1182 			rc = aggr_port_start(port);
1183 			if (rc != 0) {
1184 				mac_perim_exit(pmph);
1185 				goto bail;
1186 			}
1187 
1188 			/*
1189 			 * Turn on the promiscuous mode over the port when it
1190 			 * is requested to be turned on to receive the
1191 			 * non-primary address over a port, or the promiscuous
1192 			 * mode is enabled over the aggr.
1193 			 */
1194 			if (grp->lg_promisc || port->lp_prom_addr != NULL) {
1195 				rc = aggr_port_promisc(port, B_TRUE);
1196 				if (rc != 0) {
1197 					mac_perim_exit(pmph);
1198 					goto bail;
1199 				}
1200 			}
1201 		}
1202 		mac_perim_exit(pmph);
1203 
1204 		/*
1205 		 * Attach each port if necessary.
1206 		 */
1207 		if (aggr_port_notify_link(grp, port))
1208 			link_state_changed = B_TRUE;
1209 
1210 		/*
1211 		 * Initialize the callback functions for this port.
1212 		 */
1213 		aggr_port_init_callbacks(port);
1214 	}
1215 
1216 	/* update the MAC address of the constituent ports */
1217 	if (aggr_grp_update_ports_mac(grp))
1218 		link_state_changed = B_TRUE;
1219 
1220 	if (link_state_changed)
1221 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1222 
1223 bail:
1224 	if (rc != 0) {
1225 		/* stop and remove ports that have been added */
1226 		for (uint_t i = 0; i < port_added; i++) {
1227 			uint_t grp_remove;
1228 
1229 			port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1230 			ASSERT(port != NULL);
1231 
1232 			if (grp->lg_started) {
1233 				mac_perim_enter_by_mh(port->lp_mh, &pmph);
1234 				(void) aggr_port_promisc(port, B_FALSE);
1235 				aggr_port_stop(port);
1236 				mac_perim_exit(pmph);
1237 			}
1238 
1239 			aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
1240 
1241 			/*
1242 			 * Only the last port could have a partial set
1243 			 * of groups added.
1244 			 */
1245 			grp_remove = (i + 1 == port_added) ? grp_added :
1246 			    grp->lg_rx_group_count;
1247 
1248 			for (uint_t j = 0; j < grp_remove; j++) {
1249 				aggr_rem_pseudo_rx_group(port,
1250 				    &grp->lg_rx_groups[j]);
1251 			}
1252 
1253 			(void) aggr_grp_rem_port(grp, port, NULL, NULL);
1254 		}
1255 	}
1256 
1257 	mac_perim_exit(mph);
1258 	AGGR_GRP_REFRELE(grp);
1259 	return (rc);
1260 }
1261 
1262 static int
1263 aggr_grp_modify_common(aggr_grp_t *grp, uint8_t update_mask, uint32_t policy,
1264     boolean_t mac_fixed, const uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode,
1265     aggr_lacp_timer_t lacp_timer)
1266 {
1267 	boolean_t mac_addr_changed = B_FALSE;
1268 	boolean_t link_state_changed = B_FALSE;
1269 	mac_perim_handle_t pmph;
1270 
1271 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1272 
1273 	/* validate fixed address if specified */
1274 	if ((update_mask & AGGR_MODIFY_MAC) && mac_fixed &&
1275 	    ((bcmp(aggr_zero_mac, mac_addr, ETHERADDRL) == 0) ||
1276 	    (mac_addr[0] & 0x01))) {
1277 		return (EINVAL);
1278 	}
1279 
1280 	/* update policy if requested */
1281 	if (update_mask & AGGR_MODIFY_POLICY)
1282 		aggr_send_update_policy(grp, policy);
1283 
1284 	/* update unicast MAC address if requested */
1285 	if (update_mask & AGGR_MODIFY_MAC) {
1286 		if (mac_fixed) {
1287 			/* user-supplied MAC address */
1288 			grp->lg_mac_addr_port = NULL;
1289 			if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) != 0) {
1290 				bcopy(mac_addr, grp->lg_addr, ETHERADDRL);
1291 				mac_addr_changed = B_TRUE;
1292 			}
1293 		} else if (grp->lg_addr_fixed) {
1294 			/* switch from user-supplied to automatic */
1295 			aggr_port_t *port = grp->lg_ports;
1296 
1297 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1298 			bcopy(port->lp_addr, grp->lg_addr, ETHERADDRL);
1299 			grp->lg_mac_addr_port = port;
1300 			mac_addr_changed = B_TRUE;
1301 			mac_perim_exit(pmph);
1302 		}
1303 		grp->lg_addr_fixed = mac_fixed;
1304 	}
1305 
1306 	if (mac_addr_changed)
1307 		link_state_changed = aggr_grp_update_ports_mac(grp);
1308 
1309 	if (update_mask & AGGR_MODIFY_LACP_MODE)
1310 		aggr_lacp_update_mode(grp, lacp_mode);
1311 
1312 	if (update_mask & AGGR_MODIFY_LACP_TIMER)
1313 		aggr_lacp_update_timer(grp, lacp_timer);
1314 
1315 	if (link_state_changed)
1316 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1317 
1318 	if (mac_addr_changed)
1319 		mac_unicst_update(grp->lg_mh, grp->lg_addr);
1320 
1321 	return (0);
1322 }
1323 
1324 /*
1325  * Update properties of an existing link aggregation group.
1326  */
1327 int
1328 aggr_grp_modify(datalink_id_t linkid, uint8_t update_mask, uint32_t policy,
1329     boolean_t mac_fixed, const uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode,
1330     aggr_lacp_timer_t lacp_timer)
1331 {
1332 	aggr_grp_t *grp = NULL;
1333 	mac_perim_handle_t mph;
1334 	int err;
1335 
1336 	/* get group corresponding to linkid */
1337 	rw_enter(&aggr_grp_lock, RW_READER);
1338 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1339 	    (mod_hash_val_t *)&grp) != 0) {
1340 		rw_exit(&aggr_grp_lock);
1341 		return (ENOENT);
1342 	}
1343 	AGGR_GRP_REFHOLD(grp);
1344 
1345 	/*
1346 	 * Hold the perimeter so that the aggregation won't be destroyed.
1347 	 */
1348 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1349 	rw_exit(&aggr_grp_lock);
1350 
1351 	err = aggr_grp_modify_common(grp, update_mask, policy, mac_fixed,
1352 	    mac_addr, lacp_mode, lacp_timer);
1353 
1354 	mac_perim_exit(mph);
1355 	AGGR_GRP_REFRELE(grp);
1356 	return (err);
1357 }
1358 
1359 /*
1360  * Create a new link aggregation group upon request from administrator.
1361  * Returns 0 on success, an errno on failure.
1362  */
1363 int
1364 aggr_grp_create(datalink_id_t linkid, uint32_t key, uint_t nports,
1365     laioc_port_t *ports, uint32_t policy, boolean_t mac_fixed, boolean_t force,
1366     uchar_t *mac_addr, aggr_lacp_mode_t lacp_mode, aggr_lacp_timer_t lacp_timer,
1367     cred_t *credp)
1368 {
1369 	aggr_grp_t *grp = NULL;
1370 	aggr_port_t *port;
1371 	mac_register_t *mac;
1372 	boolean_t link_state_changed;
1373 	mac_perim_handle_t mph;
1374 	int err;
1375 	int i;
1376 	kt_did_t tid = 0;
1377 
1378 	/* need at least one port */
1379 	if (nports == 0)
1380 		return (EINVAL);
1381 
1382 	rw_enter(&aggr_grp_lock, RW_WRITER);
1383 
1384 	/* does a group with the same linkid already exist? */
1385 	err = mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1386 	    (mod_hash_val_t *)&grp);
1387 	if (err == 0) {
1388 		rw_exit(&aggr_grp_lock);
1389 		return (EEXIST);
1390 	}
1391 
1392 	grp = kmem_cache_alloc(aggr_grp_cache, KM_SLEEP);
1393 
1394 	grp->lg_refs = 1;
1395 	grp->lg_closing = B_FALSE;
1396 	grp->lg_force = force;
1397 	grp->lg_linkid = linkid;
1398 	grp->lg_zoneid = crgetzoneid(credp);
1399 	grp->lg_ifspeed = 0;
1400 	grp->lg_link_state = LINK_STATE_UNKNOWN;
1401 	grp->lg_link_duplex = LINK_DUPLEX_UNKNOWN;
1402 	grp->lg_started = B_FALSE;
1403 	grp->lg_promisc = B_FALSE;
1404 	grp->lg_lacp_done = B_FALSE;
1405 	grp->lg_tx_notify_done = B_FALSE;
1406 	grp->lg_lacp_head = grp->lg_lacp_tail = NULL;
1407 	grp->lg_lacp_rx_thread = thread_create(NULL, 0,
1408 	    aggr_lacp_rx_thread, grp, 0, &p0, TS_RUN, minclsyspri);
1409 	grp->lg_tx_notify_thread = thread_create(NULL, 0,
1410 	    aggr_tx_notify_thread, grp, 0, &p0, TS_RUN, minclsyspri);
1411 	grp->lg_tx_blocked_rings = kmem_zalloc((sizeof (mac_ring_handle_t *) *
1412 	    MAX_RINGS_PER_GROUP), KM_SLEEP);
1413 	grp->lg_tx_blocked_cnt = 0;
1414 	bzero(&grp->lg_rx_groups,
1415 	    sizeof (aggr_pseudo_rx_group_t) * MAX_GROUPS_PER_PORT);
1416 	bzero(&grp->lg_tx_group, sizeof (aggr_pseudo_tx_group_t));
1417 	aggr_lacp_init_grp(grp);
1418 
1419 	/* add MAC ports to group */
1420 	grp->lg_ports = NULL;
1421 	grp->lg_nports = 0;
1422 	grp->lg_nattached_ports = 0;
1423 	grp->lg_ntx_ports = 0;
1424 
1425 	/*
1426 	 * If key is not specified by the user, allocate the key.
1427 	 */
1428 	if ((key == 0) && ((key = (uint32_t)id_alloc(key_ids)) == 0)) {
1429 		err = ENOMEM;
1430 		goto bail;
1431 	}
1432 	grp->lg_key = key;
1433 
1434 	for (i = 0; i < nports; i++) {
1435 		err = aggr_grp_add_port(grp, ports[i].lp_linkid, force, &port);
1436 		if (err != 0)
1437 			goto bail;
1438 	}
1439 
1440 	grp->lg_rx_group_count = 1;
1441 
1442 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1443 		uint_t num_rgroups;
1444 
1445 		mac_perim_enter_by_mh(port->lp_mh, &mph);
1446 		num_rgroups = mac_get_num_rx_groups(port->lp_mh);
1447 		mac_perim_exit(mph);
1448 
1449 		/*
1450 		 * Utilize all the groups in a port. If some ports
1451 		 * have less groups than others, then traffic destined
1452 		 * for the same unicast address may be HW classified
1453 		 * on some ports but SW classified by aggr when
1454 		 * arriving on other ports.
1455 		 */
1456 		grp->lg_rx_group_count = MAX(grp->lg_rx_group_count,
1457 		    num_rgroups);
1458 	}
1459 
1460 	/*
1461 	 * There could be cases where the hardware provides more
1462 	 * groups than aggr can support. Make sure we never go above
1463 	 * the max aggr can support.
1464 	 */
1465 	grp->lg_rx_group_count = MIN(grp->lg_rx_group_count,
1466 	    MAX_GROUPS_PER_PORT);
1467 
1468 	ASSERT3U(grp->lg_rx_group_count, >, 0);
1469 	for (i = 0; i < MAX_GROUPS_PER_PORT; i++) {
1470 		grp->lg_rx_groups[i].arg_index = i;
1471 		grp->lg_rx_groups[i].arg_untagged = 0;
1472 		list_create(&(grp->lg_rx_groups[i].arg_vlans),
1473 		    sizeof (aggr_vlan_t), offsetof(aggr_vlan_t, av_link));
1474 	}
1475 
1476 	/*
1477 	 * If no explicit MAC address was specified by the administrator,
1478 	 * set it to the MAC address of the first port.
1479 	 */
1480 	grp->lg_addr_fixed = mac_fixed;
1481 	if (grp->lg_addr_fixed) {
1482 		/* validate specified address */
1483 		if (bcmp(aggr_zero_mac, mac_addr, ETHERADDRL) == 0) {
1484 			err = EINVAL;
1485 			goto bail;
1486 		}
1487 		bcopy(mac_addr, grp->lg_addr, ETHERADDRL);
1488 	} else {
1489 		bcopy(grp->lg_ports->lp_addr, grp->lg_addr, ETHERADDRL);
1490 		grp->lg_mac_addr_port = grp->lg_ports;
1491 	}
1492 
1493 	/* Set the initial group capabilities. */
1494 	aggr_grp_capab_set(grp);
1495 
1496 	if ((mac = mac_alloc(MAC_VERSION)) == NULL) {
1497 		err = ENOMEM;
1498 		goto bail;
1499 	}
1500 	mac->m_type_ident = MAC_PLUGIN_IDENT_ETHER;
1501 	mac->m_driver = grp;
1502 	mac->m_dip = aggr_dip;
1503 	mac->m_instance = grp->lg_key > AGGR_MAX_KEY ? (uint_t)-1 : grp->lg_key;
1504 	mac->m_src_addr = grp->lg_addr;
1505 	mac->m_callbacks = &aggr_m_callbacks;
1506 	mac->m_min_sdu = 0;
1507 	mac->m_max_sdu = grp->lg_max_sdu = aggr_grp_max_sdu(grp);
1508 	mac->m_margin = aggr_grp_max_margin(grp);
1509 	mac->m_v12n = MAC_VIRT_LEVEL1;
1510 	err = mac_register(mac, &grp->lg_mh);
1511 	mac_free(mac);
1512 	if (err != 0)
1513 		goto bail;
1514 
1515 	err = dls_devnet_create(grp->lg_mh, grp->lg_linkid, crgetzoneid(credp));
1516 	if (err != 0) {
1517 		(void) mac_unregister(grp->lg_mh);
1518 		grp->lg_mh = NULL;
1519 		goto bail;
1520 	}
1521 
1522 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1523 
1524 	/*
1525 	 * Update the MAC address of the constituent ports.
1526 	 * None of the port is attached at this time, the link state of the
1527 	 * aggregation will not change.
1528 	 *
1529 	 * All ports take on the primary MAC address of the aggr
1530 	 * (lg_aggr). At this point, none of the ports are attached;
1531 	 * thus the link state of the aggregation will not change.
1532 	 */
1533 	link_state_changed = aggr_grp_update_ports_mac(grp);
1534 	ASSERT(!link_state_changed);
1535 
1536 	/* Update outbound load balancing policy. */
1537 	aggr_send_update_policy(grp, policy);
1538 
1539 	/* Set LACP mode. */
1540 	aggr_lacp_set_mode(grp, lacp_mode, lacp_timer);
1541 
1542 	/*
1543 	 * Attach each port if necessary.
1544 	 */
1545 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1546 		/*
1547 		 * Create the pseudo ring for each HW ring of the
1548 		 * underlying port. Note that this is done after the
1549 		 * aggr registers its MAC.
1550 		 */
1551 		VERIFY3S(aggr_add_pseudo_tx_group(port, &grp->lg_tx_group),
1552 		    ==, 0);
1553 
1554 		for (i = 0; i < grp->lg_rx_group_count; i++) {
1555 			VERIFY3S(aggr_add_pseudo_rx_group(port,
1556 			    &grp->lg_rx_groups[i]), ==, 0);
1557 		}
1558 
1559 		if (aggr_port_notify_link(grp, port))
1560 			link_state_changed = B_TRUE;
1561 
1562 		/*
1563 		 * Initialize the callback functions for this port.
1564 		 */
1565 		aggr_port_init_callbacks(port);
1566 	}
1567 
1568 	if (link_state_changed)
1569 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1570 
1571 	/* add new group to hash table */
1572 	err = mod_hash_insert(aggr_grp_hash, GRP_HASH_KEY(linkid),
1573 	    (mod_hash_val_t)grp);
1574 	ASSERT(err == 0);
1575 	aggr_grp_cnt++;
1576 
1577 	mac_perim_exit(mph);
1578 	rw_exit(&aggr_grp_lock);
1579 	return (0);
1580 
1581 bail:
1582 
1583 	grp->lg_closing = B_TRUE;
1584 
1585 	port = grp->lg_ports;
1586 	while (port != NULL) {
1587 		aggr_port_t *cport;
1588 
1589 		cport = port->lp_next;
1590 		aggr_port_delete(port);
1591 		port = cport;
1592 	}
1593 
1594 	/*
1595 	 * Inform the lacp_rx thread to exit.
1596 	 */
1597 	mutex_enter(&grp->lg_lacp_lock);
1598 	grp->lg_lacp_done = B_TRUE;
1599 	cv_signal(&grp->lg_lacp_cv);
1600 	while (grp->lg_lacp_rx_thread != NULL)
1601 		cv_wait(&grp->lg_lacp_cv, &grp->lg_lacp_lock);
1602 	mutex_exit(&grp->lg_lacp_lock);
1603 	/*
1604 	 * Inform the tx_notify thread to exit.
1605 	 */
1606 	mutex_enter(&grp->lg_tx_flowctl_lock);
1607 	if (grp->lg_tx_notify_thread != NULL) {
1608 		tid = grp->lg_tx_notify_thread->t_did;
1609 		grp->lg_tx_notify_done = B_TRUE;
1610 		cv_signal(&grp->lg_tx_flowctl_cv);
1611 	}
1612 	mutex_exit(&grp->lg_tx_flowctl_lock);
1613 	if (tid != 0)
1614 		thread_join(tid);
1615 
1616 	kmem_free(grp->lg_tx_blocked_rings,
1617 	    (sizeof (mac_ring_handle_t *) * MAX_RINGS_PER_GROUP));
1618 	rw_exit(&aggr_grp_lock);
1619 	AGGR_GRP_REFRELE(grp);
1620 	return (err);
1621 }
1622 
1623 /*
1624  * Return a pointer to the member of a group with specified linkid.
1625  */
1626 static aggr_port_t *
1627 aggr_grp_port_lookup(aggr_grp_t *grp, datalink_id_t linkid)
1628 {
1629 	aggr_port_t *port;
1630 
1631 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1632 
1633 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
1634 		if (port->lp_linkid == linkid)
1635 			break;
1636 	}
1637 
1638 	return (port);
1639 }
1640 
1641 /*
1642  * Stop, detach and remove a port from a link aggregation group.
1643  */
1644 static int
1645 aggr_grp_rem_port(aggr_grp_t *grp, aggr_port_t *port,
1646     boolean_t *mac_addr_changedp, boolean_t *link_state_changedp)
1647 {
1648 	int rc = 0;
1649 	aggr_port_t **pport;
1650 	boolean_t mac_addr_changed = B_FALSE;
1651 	boolean_t link_state_changed = B_FALSE;
1652 	mac_perim_handle_t mph;
1653 	uint64_t val;
1654 	uint_t i;
1655 	uint_t stat;
1656 
1657 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
1658 	ASSERT(grp->lg_nports > 1);
1659 	ASSERT(!grp->lg_closing);
1660 
1661 	/* unlink port */
1662 	for (pport = &grp->lg_ports; *pport != port;
1663 	    pport = &(*pport)->lp_next) {
1664 		if (*pport == NULL) {
1665 			rc = ENOENT;
1666 			goto done;
1667 		}
1668 	}
1669 	*pport = port->lp_next;
1670 
1671 	mac_perim_enter_by_mh(port->lp_mh, &mph);
1672 
1673 	/*
1674 	 * If the MAC address of the port being removed was assigned
1675 	 * to the group, update the group MAC address
1676 	 * using the MAC address of a different port.
1677 	 */
1678 	if (!grp->lg_addr_fixed && grp->lg_mac_addr_port == port) {
1679 		/*
1680 		 * Set the MAC address of the group to the
1681 		 * MAC address of its first port.
1682 		 */
1683 		bcopy(grp->lg_ports->lp_addr, grp->lg_addr, ETHERADDRL);
1684 		grp->lg_mac_addr_port = grp->lg_ports;
1685 		mac_addr_changed = B_TRUE;
1686 	}
1687 
1688 	link_state_changed = aggr_grp_detach_port(grp, port);
1689 
1690 	/*
1691 	 * Add the counter statistics of the ports while it was aggregated
1692 	 * to the group's residual statistics.  This is done by obtaining
1693 	 * the current counter from the underlying MAC then subtracting the
1694 	 * value of the counter at the moment it was added to the
1695 	 * aggregation.
1696 	 */
1697 	for (i = 0; i < MAC_NSTAT; i++) {
1698 		stat = i + MAC_STAT_MIN;
1699 		if (!MAC_STAT_ISACOUNTER(stat))
1700 			continue;
1701 		val = aggr_port_stat(port, stat);
1702 		val -= port->lp_stat[i];
1703 		mutex_enter(&grp->lg_stat_lock);
1704 		grp->lg_stat[i] += val;
1705 		mutex_exit(&grp->lg_stat_lock);
1706 	}
1707 	for (i = 0; i < ETHER_NSTAT; i++) {
1708 		stat = i + MACTYPE_STAT_MIN;
1709 		if (!ETHER_STAT_ISACOUNTER(stat))
1710 			continue;
1711 		val = aggr_port_stat(port, stat);
1712 		val -= port->lp_ether_stat[i];
1713 		mutex_enter(&grp->lg_stat_lock);
1714 		grp->lg_ether_stat[i] += val;
1715 		mutex_exit(&grp->lg_stat_lock);
1716 	}
1717 
1718 	grp->lg_nports--;
1719 	mac_perim_exit(mph);
1720 
1721 	aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
1722 	aggr_port_delete(port);
1723 
1724 	/*
1725 	 * If the group MAC address has changed, update the MAC address of
1726 	 * the remaining constituent ports according to the new MAC
1727 	 * address of the group.
1728 	 */
1729 	if (mac_addr_changed && aggr_grp_update_ports_mac(grp))
1730 		link_state_changed = B_TRUE;
1731 
1732 done:
1733 	if (mac_addr_changedp != NULL)
1734 		*mac_addr_changedp = mac_addr_changed;
1735 	if (link_state_changedp != NULL)
1736 		*link_state_changedp = link_state_changed;
1737 
1738 	return (rc);
1739 }
1740 
1741 /*
1742  * Remove one or more ports from an existing link aggregation group.
1743  */
1744 int
1745 aggr_grp_rem_ports(datalink_id_t linkid, uint_t nports, laioc_port_t *ports)
1746 {
1747 	int rc = 0, i;
1748 	aggr_grp_t *grp = NULL;
1749 	aggr_port_t *port;
1750 	boolean_t mac_addr_update = B_FALSE, mac_addr_changed;
1751 	boolean_t link_state_update = B_FALSE, link_state_changed;
1752 	mac_perim_handle_t mph, pmph;
1753 
1754 	/* get group corresponding to linkid */
1755 	rw_enter(&aggr_grp_lock, RW_READER);
1756 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1757 	    (mod_hash_val_t *)&grp) != 0) {
1758 		rw_exit(&aggr_grp_lock);
1759 		return (ENOENT);
1760 	}
1761 	AGGR_GRP_REFHOLD(grp);
1762 
1763 	/*
1764 	 * Hold the perimeter so that the aggregation won't be destroyed.
1765 	 */
1766 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1767 	rw_exit(&aggr_grp_lock);
1768 
1769 	/* we need to keep at least one port per group */
1770 	if (nports >= grp->lg_nports) {
1771 		rc = EINVAL;
1772 		goto bail;
1773 	}
1774 
1775 	/* first verify that all the groups are valid */
1776 	for (i = 0; i < nports; i++) {
1777 		if (aggr_grp_port_lookup(grp, ports[i].lp_linkid) == NULL) {
1778 			/* port not found */
1779 			rc = ENOENT;
1780 			goto bail;
1781 		}
1782 	}
1783 
1784 	/* clear the promiscous mode for the specified ports */
1785 	for (i = 0; i < nports && rc == 0; i++) {
1786 		/* lookup port */
1787 		port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1788 		ASSERT(port != NULL);
1789 
1790 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
1791 		rc = aggr_port_promisc(port, B_FALSE);
1792 		mac_perim_exit(pmph);
1793 	}
1794 	if (rc != 0) {
1795 		for (i = 0; i < nports; i++) {
1796 			port = aggr_grp_port_lookup(grp,
1797 			    ports[i].lp_linkid);
1798 			ASSERT(port != NULL);
1799 
1800 			/*
1801 			 * Turn the promiscuous mode back on if it is required
1802 			 * to receive the non-primary address over a port, or
1803 			 * the promiscous mode is enabled over the aggr.
1804 			 */
1805 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1806 			if (port->lp_started && (grp->lg_promisc ||
1807 			    port->lp_prom_addr != NULL)) {
1808 				(void) aggr_port_promisc(port, B_TRUE);
1809 			}
1810 			mac_perim_exit(pmph);
1811 		}
1812 		goto bail;
1813 	}
1814 
1815 	/* remove the specified ports from group */
1816 	for (i = 0; i < nports; i++) {
1817 		/* lookup port */
1818 		port = aggr_grp_port_lookup(grp, ports[i].lp_linkid);
1819 		ASSERT(port != NULL);
1820 
1821 		/* stop port if group has already been started */
1822 		if (grp->lg_started) {
1823 			mac_perim_enter_by_mh(port->lp_mh, &pmph);
1824 			aggr_port_stop(port);
1825 			mac_perim_exit(pmph);
1826 		}
1827 
1828 		/*
1829 		 * aggr_rem_pseudo_tx_group() is not called here. Instead
1830 		 * it is called from inside aggr_grp_rem_port() after the
1831 		 * port has been detached. The reason is that
1832 		 * aggr_rem_pseudo_tx_group() removes one ring at a time
1833 		 * and if there is still traffic going on, then there
1834 		 * is the possibility of aggr_find_tx_ring() returning a
1835 		 * removed ring for transmission. Once the port has been
1836 		 * detached, that port will not be used and
1837 		 * aggr_find_tx_ring() will not return any rings
1838 		 * belonging to it.
1839 		 */
1840 		for (i = 0; i < grp->lg_rx_group_count; i++)
1841 			aggr_rem_pseudo_rx_group(port, &grp->lg_rx_groups[i]);
1842 
1843 		/* remove port from group */
1844 		rc = aggr_grp_rem_port(grp, port, &mac_addr_changed,
1845 		    &link_state_changed);
1846 		ASSERT(rc == 0);
1847 		mac_addr_update = mac_addr_update || mac_addr_changed;
1848 		link_state_update = link_state_update || link_state_changed;
1849 	}
1850 
1851 bail:
1852 	if (mac_addr_update)
1853 		mac_unicst_update(grp->lg_mh, grp->lg_addr);
1854 	if (link_state_update)
1855 		mac_link_update(grp->lg_mh, grp->lg_link_state);
1856 
1857 	mac_perim_exit(mph);
1858 	AGGR_GRP_REFRELE(grp);
1859 
1860 	return (rc);
1861 }
1862 
1863 int
1864 aggr_grp_delete(datalink_id_t linkid, cred_t *cred)
1865 {
1866 	aggr_grp_t *grp = NULL;
1867 	aggr_port_t *port, *cport;
1868 	datalink_id_t tmpid;
1869 	mod_hash_val_t val;
1870 	mac_perim_handle_t mph, pmph;
1871 	int err;
1872 	kt_did_t tid = 0;
1873 
1874 	rw_enter(&aggr_grp_lock, RW_WRITER);
1875 
1876 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
1877 	    (mod_hash_val_t *)&grp) != 0) {
1878 		rw_exit(&aggr_grp_lock);
1879 		return (ENOENT);
1880 	}
1881 
1882 	/*
1883 	 * Note that dls_devnet_destroy() must be called before lg_lock is
1884 	 * held. Otherwise, it will deadlock if another thread is in
1885 	 * aggr_m_stat() and thus has a kstat_hold() on the kstats that
1886 	 * dls_devnet_destroy() needs to delete.
1887 	 */
1888 	if ((err = dls_devnet_destroy(grp->lg_mh, &tmpid, B_TRUE)) != 0) {
1889 		rw_exit(&aggr_grp_lock);
1890 		return (err);
1891 	}
1892 	ASSERT(linkid == tmpid);
1893 
1894 	/*
1895 	 * Unregister from the MAC service module. Since this can
1896 	 * fail if a client hasn't closed the MAC port, we gracefully
1897 	 * fail the operation.
1898 	 */
1899 	if ((err = mac_disable(grp->lg_mh)) != 0) {
1900 		(void) dls_devnet_create(grp->lg_mh, linkid, crgetzoneid(cred));
1901 		rw_exit(&aggr_grp_lock);
1902 		return (err);
1903 	}
1904 	(void) mod_hash_remove(aggr_grp_hash, GRP_HASH_KEY(linkid), &val);
1905 	ASSERT(grp == (aggr_grp_t *)val);
1906 
1907 	ASSERT(aggr_grp_cnt > 0);
1908 	aggr_grp_cnt--;
1909 	rw_exit(&aggr_grp_lock);
1910 
1911 	/*
1912 	 * Inform the lacp_rx thread to exit.
1913 	 */
1914 	mutex_enter(&grp->lg_lacp_lock);
1915 	grp->lg_lacp_done = B_TRUE;
1916 	cv_signal(&grp->lg_lacp_cv);
1917 	while (grp->lg_lacp_rx_thread != NULL)
1918 		cv_wait(&grp->lg_lacp_cv, &grp->lg_lacp_lock);
1919 	mutex_exit(&grp->lg_lacp_lock);
1920 	/*
1921 	 * Inform the tx_notify_thread to exit.
1922 	 */
1923 	mutex_enter(&grp->lg_tx_flowctl_lock);
1924 	if (grp->lg_tx_notify_thread != NULL) {
1925 		tid = grp->lg_tx_notify_thread->t_did;
1926 		grp->lg_tx_notify_done = B_TRUE;
1927 		cv_signal(&grp->lg_tx_flowctl_cv);
1928 	}
1929 	mutex_exit(&grp->lg_tx_flowctl_lock);
1930 	if (tid != 0)
1931 		thread_join(tid);
1932 
1933 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
1934 
1935 	grp->lg_closing = B_TRUE;
1936 	/* detach and free MAC ports associated with group */
1937 	port = grp->lg_ports;
1938 	while (port != NULL) {
1939 		cport = port->lp_next;
1940 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
1941 		if (grp->lg_started)
1942 			aggr_port_stop(port);
1943 		(void) aggr_grp_detach_port(grp, port);
1944 		mac_perim_exit(pmph);
1945 		aggr_rem_pseudo_tx_group(port, &grp->lg_tx_group);
1946 		for (uint_t i = 0; i < grp->lg_rx_group_count; i++)
1947 			aggr_rem_pseudo_rx_group(port, &grp->lg_rx_groups[i]);
1948 		aggr_port_delete(port);
1949 		port = cport;
1950 	}
1951 
1952 	mac_perim_exit(mph);
1953 
1954 	kmem_free(grp->lg_tx_blocked_rings,
1955 	    (sizeof (mac_ring_handle_t *) * MAX_RINGS_PER_GROUP));
1956 	/*
1957 	 * Wait for the port's lacp timer thread and its notification callback
1958 	 * to exit before calling mac_unregister() since both needs to access
1959 	 * the mac perimeter of the grp.
1960 	 */
1961 	aggr_grp_port_wait(grp);
1962 
1963 	VERIFY(mac_unregister(grp->lg_mh) == 0);
1964 	grp->lg_mh = NULL;
1965 
1966 	for (uint_t i = 0; i < MAX_GROUPS_PER_PORT; i++) {
1967 		list_destroy(&(grp->lg_rx_groups[i].arg_vlans));
1968 	}
1969 
1970 	AGGR_GRP_REFRELE(grp);
1971 	return (0);
1972 }
1973 
1974 void
1975 aggr_grp_free(aggr_grp_t *grp)
1976 {
1977 	ASSERT(grp->lg_refs == 0);
1978 	ASSERT(grp->lg_port_ref == 0);
1979 	if (grp->lg_key > AGGR_MAX_KEY) {
1980 		id_free(key_ids, grp->lg_key);
1981 		grp->lg_key = 0;
1982 	}
1983 	kmem_cache_free(aggr_grp_cache, grp);
1984 }
1985 
1986 int
1987 aggr_grp_info(datalink_id_t linkid, void *fn_arg,
1988     aggr_grp_info_new_grp_fn_t new_grp_fn,
1989     aggr_grp_info_new_port_fn_t new_port_fn, cred_t *cred)
1990 {
1991 	aggr_grp_t	*grp;
1992 	aggr_port_t	*port;
1993 	mac_perim_handle_t mph, pmph;
1994 	int		rc = 0;
1995 
1996 	/*
1997 	 * Make sure that the aggregation link is visible from the caller's
1998 	 * zone.
1999 	 */
2000 	if (!dls_devnet_islinkvisible(linkid, crgetzoneid(cred)))
2001 		return (ENOENT);
2002 
2003 	rw_enter(&aggr_grp_lock, RW_READER);
2004 
2005 	if (mod_hash_find(aggr_grp_hash, GRP_HASH_KEY(linkid),
2006 	    (mod_hash_val_t *)&grp) != 0) {
2007 		rw_exit(&aggr_grp_lock);
2008 		return (ENOENT);
2009 	}
2010 	AGGR_GRP_REFHOLD(grp);
2011 
2012 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2013 	rw_exit(&aggr_grp_lock);
2014 
2015 	rc = new_grp_fn(fn_arg, grp->lg_linkid,
2016 	    (grp->lg_key > AGGR_MAX_KEY) ? 0 : grp->lg_key, grp->lg_addr,
2017 	    grp->lg_addr_fixed, grp->lg_force, grp->lg_tx_policy,
2018 	    grp->lg_nports, grp->lg_lacp_mode, grp->aggr.PeriodicTimer);
2019 
2020 	if (rc != 0)
2021 		goto bail;
2022 
2023 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2024 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2025 		rc = new_port_fn(fn_arg, port->lp_linkid, port->lp_addr,
2026 		    port->lp_state, &port->lp_lacp.ActorOperPortState);
2027 		mac_perim_exit(pmph);
2028 
2029 		if (rc != 0)
2030 			goto bail;
2031 	}
2032 
2033 bail:
2034 	mac_perim_exit(mph);
2035 	AGGR_GRP_REFRELE(grp);
2036 	return (rc);
2037 }
2038 
2039 /*ARGSUSED*/
2040 static void
2041 aggr_m_ioctl(void *arg, queue_t *q, mblk_t *mp)
2042 {
2043 	miocnak(q, mp, 0, ENOTSUP);
2044 }
2045 
2046 static int
2047 aggr_grp_stat(aggr_grp_t *grp, uint_t stat, uint64_t *val)
2048 {
2049 	aggr_port_t	*port;
2050 	uint_t		stat_index;
2051 
2052 	ASSERT(MUTEX_HELD(&grp->lg_stat_lock));
2053 
2054 	/* We only aggregate counter statistics. */
2055 	if (IS_MAC_STAT(stat) && !MAC_STAT_ISACOUNTER(stat) ||
2056 	    IS_MACTYPE_STAT(stat) && !ETHER_STAT_ISACOUNTER(stat)) {
2057 		return (ENOTSUP);
2058 	}
2059 
2060 	/*
2061 	 * Counter statistics for a group are computed by aggregating the
2062 	 * counters of the members MACs while they were aggregated, plus
2063 	 * the residual counter of the group itself, which is updated each
2064 	 * time a MAC is removed from the group.
2065 	 */
2066 	*val = 0;
2067 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2068 		/* actual port statistic */
2069 		*val += aggr_port_stat(port, stat);
2070 		/*
2071 		 * minus the port stat when it was added, plus any residual
2072 		 * amount for the group.
2073 		 */
2074 		if (IS_MAC_STAT(stat)) {
2075 			stat_index = stat - MAC_STAT_MIN;
2076 			*val -= port->lp_stat[stat_index];
2077 			*val += grp->lg_stat[stat_index];
2078 		} else if (IS_MACTYPE_STAT(stat)) {
2079 			stat_index = stat - MACTYPE_STAT_MIN;
2080 			*val -= port->lp_ether_stat[stat_index];
2081 			*val += grp->lg_ether_stat[stat_index];
2082 		}
2083 	}
2084 	return (0);
2085 }
2086 
2087 int
2088 aggr_rx_ring_stat(mac_ring_driver_t rdriver, uint_t stat, uint64_t *val)
2089 {
2090 	aggr_pseudo_rx_ring_t   *rx_ring = (aggr_pseudo_rx_ring_t *)rdriver;
2091 
2092 	if (rx_ring->arr_hw_rh != NULL) {
2093 		*val = mac_pseudo_rx_ring_stat_get(rx_ring->arr_hw_rh, stat);
2094 	} else {
2095 		aggr_port_t	*port = rx_ring->arr_port;
2096 
2097 		*val = mac_stat_get(port->lp_mh, stat);
2098 
2099 	}
2100 	return (0);
2101 }
2102 
2103 int
2104 aggr_tx_ring_stat(mac_ring_driver_t rdriver, uint_t stat, uint64_t *val)
2105 {
2106 	aggr_pseudo_tx_ring_t   *tx_ring = (aggr_pseudo_tx_ring_t *)rdriver;
2107 
2108 	if (tx_ring->atr_hw_rh != NULL) {
2109 		*val = mac_pseudo_tx_ring_stat_get(tx_ring->atr_hw_rh, stat);
2110 	} else {
2111 		aggr_port_t	*port = tx_ring->atr_port;
2112 
2113 		*val = mac_stat_get(port->lp_mh, stat);
2114 	}
2115 	return (0);
2116 }
2117 
2118 static int
2119 aggr_m_stat(void *arg, uint_t stat, uint64_t *val)
2120 {
2121 	aggr_grp_t		*grp = arg;
2122 	int			rval = 0;
2123 
2124 	mutex_enter(&grp->lg_stat_lock);
2125 
2126 	switch (stat) {
2127 	case MAC_STAT_IFSPEED:
2128 		*val = grp->lg_ifspeed;
2129 		break;
2130 
2131 	case ETHER_STAT_LINK_DUPLEX:
2132 		*val = grp->lg_link_duplex;
2133 		break;
2134 
2135 	default:
2136 		/*
2137 		 * For all other statistics, we return the aggregated stat
2138 		 * from the underlying ports.  aggr_grp_stat() will set
2139 		 * rval appropriately if the statistic isn't a counter.
2140 		 */
2141 		rval = aggr_grp_stat(grp, stat, val);
2142 	}
2143 
2144 	mutex_exit(&grp->lg_stat_lock);
2145 	return (rval);
2146 }
2147 
2148 static int
2149 aggr_m_start(void *arg)
2150 {
2151 	aggr_grp_t *grp = arg;
2152 	aggr_port_t *port;
2153 	mac_perim_handle_t mph, pmph;
2154 
2155 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2156 
2157 	/*
2158 	 * Attempts to start all configured members of the group.
2159 	 * Group members will be attached when their link-up notification
2160 	 * is received.
2161 	 */
2162 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2163 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2164 		if (aggr_port_start(port) != 0) {
2165 			mac_perim_exit(pmph);
2166 			continue;
2167 		}
2168 
2169 		/*
2170 		 * Turn on the promiscuous mode if it is required to receive
2171 		 * the non-primary address over a port, or the promiscous
2172 		 * mode is enabled over the aggr.
2173 		 */
2174 		if (grp->lg_promisc || port->lp_prom_addr != NULL) {
2175 			if (aggr_port_promisc(port, B_TRUE) != 0)
2176 				aggr_port_stop(port);
2177 		}
2178 		mac_perim_exit(pmph);
2179 	}
2180 
2181 	grp->lg_started = B_TRUE;
2182 
2183 	mac_perim_exit(mph);
2184 	return (0);
2185 }
2186 
2187 static void
2188 aggr_m_stop(void *arg)
2189 {
2190 	aggr_grp_t *grp = arg;
2191 	aggr_port_t *port;
2192 	mac_perim_handle_t mph, pmph;
2193 
2194 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2195 
2196 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2197 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2198 
2199 		/* reset port promiscuous mode */
2200 		(void) aggr_port_promisc(port, B_FALSE);
2201 
2202 		aggr_port_stop(port);
2203 		mac_perim_exit(pmph);
2204 	}
2205 
2206 	grp->lg_started = B_FALSE;
2207 	mac_perim_exit(mph);
2208 }
2209 
2210 static int
2211 aggr_m_promisc(void *arg, boolean_t on)
2212 {
2213 	aggr_grp_t *grp = arg;
2214 	aggr_port_t *port;
2215 	boolean_t link_state_changed = B_FALSE;
2216 	mac_perim_handle_t mph, pmph;
2217 
2218 	AGGR_GRP_REFHOLD(grp);
2219 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2220 
2221 	ASSERT(!grp->lg_closing);
2222 
2223 	if (on == grp->lg_promisc)
2224 		goto bail;
2225 
2226 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2227 		int	err = 0;
2228 
2229 		mac_perim_enter_by_mh(port->lp_mh, &pmph);
2230 		AGGR_PORT_REFHOLD(port);
2231 		if (!on && (port->lp_prom_addr == NULL))
2232 			err = aggr_port_promisc(port, B_FALSE);
2233 		else if (on && port->lp_started)
2234 			err = aggr_port_promisc(port, B_TRUE);
2235 
2236 		if (err != 0) {
2237 			if (aggr_grp_detach_port(grp, port))
2238 				link_state_changed = B_TRUE;
2239 		} else {
2240 			/*
2241 			 * If a port was detached because of a previous
2242 			 * failure changing the promiscuity, the port
2243 			 * is reattached when it successfully changes
2244 			 * the promiscuity now, and this might cause
2245 			 * the link state of the aggregation to change.
2246 			 */
2247 			if (aggr_grp_attach_port(grp, port))
2248 				link_state_changed = B_TRUE;
2249 		}
2250 		mac_perim_exit(pmph);
2251 		AGGR_PORT_REFRELE(port);
2252 	}
2253 
2254 	grp->lg_promisc = on;
2255 
2256 	if (link_state_changed)
2257 		mac_link_update(grp->lg_mh, grp->lg_link_state);
2258 
2259 bail:
2260 	mac_perim_exit(mph);
2261 	AGGR_GRP_REFRELE(grp);
2262 
2263 	return (0);
2264 }
2265 
2266 static void
2267 aggr_grp_port_rename(const char *new_name, void *arg)
2268 {
2269 	/*
2270 	 * aggr port's mac client name is the format of "aggr link name" plus
2271 	 * AGGR_PORT_NAME_DELIMIT plus "underneath link name".
2272 	 */
2273 	int aggr_len, link_len, clnt_name_len, i;
2274 	char *str_end, *str_st, *str_del;
2275 	char aggr_name[MAXNAMELEN];
2276 	char link_name[MAXNAMELEN];
2277 	char *clnt_name;
2278 	aggr_grp_t *aggr_grp = arg;
2279 	aggr_port_t *aggr_port = aggr_grp->lg_ports;
2280 
2281 	for (i = 0; i < aggr_grp->lg_nports; i++) {
2282 		clnt_name = mac_client_name(aggr_port->lp_mch);
2283 		clnt_name_len = strlen(clnt_name);
2284 		str_st = clnt_name;
2285 		str_end = &(clnt_name[clnt_name_len]);
2286 		str_del = strchr(str_st, AGGR_PORT_NAME_DELIMIT);
2287 		ASSERT(str_del != NULL);
2288 		aggr_len = (intptr_t)((uintptr_t)str_del - (uintptr_t)str_st);
2289 		link_len = (intptr_t)((uintptr_t)str_end - (uintptr_t)str_del);
2290 		bzero(aggr_name, MAXNAMELEN);
2291 		bzero(link_name, MAXNAMELEN);
2292 		bcopy(clnt_name, aggr_name, aggr_len);
2293 		bcopy(str_del, link_name, link_len + 1);
2294 		bzero(clnt_name, MAXNAMELEN);
2295 		(void) snprintf(clnt_name, MAXNAMELEN, "%s%s", new_name,
2296 		    link_name);
2297 
2298 		(void) mac_rename_primary(aggr_port->lp_mh, NULL);
2299 		aggr_port = aggr_port->lp_next;
2300 	}
2301 }
2302 
2303 /*
2304  * Initialize the capabilities that are advertised for the group
2305  * according to the capabilities of the constituent ports.
2306  */
2307 static boolean_t
2308 aggr_m_capab_get(void *arg, mac_capab_t cap, void *cap_data)
2309 {
2310 	aggr_grp_t *grp = arg;
2311 
2312 	switch (cap) {
2313 	case MAC_CAPAB_HCKSUM: {
2314 		uint32_t *hcksum_txflags = cap_data;
2315 		*hcksum_txflags = grp->lg_hcksum_txflags;
2316 		break;
2317 	}
2318 	case MAC_CAPAB_LSO: {
2319 		mac_capab_lso_t *cap_lso = cap_data;
2320 
2321 		if (grp->lg_lso) {
2322 			*cap_lso = grp->lg_cap_lso;
2323 			break;
2324 		} else {
2325 			return (B_FALSE);
2326 		}
2327 	}
2328 	case MAC_CAPAB_NO_NATIVEVLAN:
2329 		return (!grp->lg_vlan);
2330 	case MAC_CAPAB_NO_ZCOPY:
2331 		return (!grp->lg_zcopy);
2332 	case MAC_CAPAB_RINGS: {
2333 		mac_capab_rings_t *cap_rings = cap_data;
2334 		uint_t ring_cnt = 0;
2335 
2336 		for (uint_t i = 0; i < grp->lg_rx_group_count; i++)
2337 			ring_cnt += grp->lg_rx_groups[i].arg_ring_cnt;
2338 
2339 		if (cap_rings->mr_type == MAC_RING_TYPE_RX) {
2340 			cap_rings->mr_group_type = MAC_GROUP_TYPE_STATIC;
2341 			cap_rings->mr_rnum = ring_cnt;
2342 			cap_rings->mr_gnum = grp->lg_rx_group_count;
2343 			cap_rings->mr_gaddring = NULL;
2344 			cap_rings->mr_gremring = NULL;
2345 		} else {
2346 			cap_rings->mr_group_type = MAC_GROUP_TYPE_STATIC;
2347 			cap_rings->mr_rnum = grp->lg_tx_group.atg_ring_cnt;
2348 			cap_rings->mr_gnum = 0;
2349 		}
2350 		cap_rings->mr_rget = aggr_fill_ring;
2351 		cap_rings->mr_gget = aggr_fill_group;
2352 		break;
2353 	}
2354 	case MAC_CAPAB_AGGR:
2355 	{
2356 		mac_capab_aggr_t *aggr_cap;
2357 
2358 		if (cap_data != NULL) {
2359 			aggr_cap = cap_data;
2360 			aggr_cap->mca_rename_fn = aggr_grp_port_rename;
2361 			aggr_cap->mca_unicst = aggr_m_unicst;
2362 			aggr_cap->mca_find_tx_ring_fn = aggr_find_tx_ring;
2363 			aggr_cap->mca_arg = arg;
2364 		}
2365 		return (B_TRUE);
2366 	}
2367 	default:
2368 		return (B_FALSE);
2369 	}
2370 	return (B_TRUE);
2371 }
2372 
2373 /*
2374  * Callback function for MAC layer to register groups.
2375  */
2376 static void
2377 aggr_fill_group(void *arg, mac_ring_type_t rtype, const int index,
2378     mac_group_info_t *infop, mac_group_handle_t gh)
2379 {
2380 	aggr_grp_t *grp = arg;
2381 
2382 	if (rtype == MAC_RING_TYPE_RX) {
2383 		aggr_pseudo_rx_group_t *rx_group = &grp->lg_rx_groups[index];
2384 
2385 		rx_group->arg_gh = gh;
2386 		rx_group->arg_grp = grp;
2387 
2388 		infop->mgi_driver = (mac_group_driver_t)rx_group;
2389 		infop->mgi_start = NULL;
2390 		infop->mgi_stop = NULL;
2391 		infop->mgi_addmac = aggr_addmac;
2392 		infop->mgi_remmac = aggr_remmac;
2393 		infop->mgi_count = rx_group->arg_ring_cnt;
2394 
2395 		/*
2396 		 * Always set the HW VLAN callbacks. They are smart
2397 		 * enough to know when a port has HW VLAN filters to
2398 		 * program and when it doesn't.
2399 		 */
2400 		infop->mgi_addvlan = aggr_addvlan;
2401 		infop->mgi_remvlan = aggr_remvlan;
2402 	} else {
2403 		aggr_pseudo_tx_group_t *tx_group = &grp->lg_tx_group;
2404 
2405 		ASSERT3S(index, ==, 0);
2406 		tx_group->atg_gh = gh;
2407 	}
2408 }
2409 
2410 /*
2411  * Callback funtion for MAC layer to register all rings.
2412  */
2413 static void
2414 aggr_fill_ring(void *arg, mac_ring_type_t rtype, const int rg_index,
2415     const int index, mac_ring_info_t *infop, mac_ring_handle_t rh)
2416 {
2417 	aggr_grp_t	*grp = arg;
2418 
2419 	switch (rtype) {
2420 	case MAC_RING_TYPE_RX: {
2421 		aggr_pseudo_rx_group_t	*rx_group;
2422 		aggr_pseudo_rx_ring_t	*rx_ring;
2423 		mac_intr_t		aggr_mac_intr;
2424 
2425 		rx_group = &grp->lg_rx_groups[rg_index];
2426 		ASSERT3S(index, >=, 0);
2427 		ASSERT3S(index, <, rx_group->arg_ring_cnt);
2428 		rx_ring = rx_group->arg_rings + index;
2429 		rx_ring->arr_rh = rh;
2430 
2431 		/*
2432 		 * Entrypoint to enable interrupt (disable poll) and
2433 		 * disable interrupt (enable poll).
2434 		 */
2435 		aggr_mac_intr.mi_handle = (mac_intr_handle_t)rx_ring;
2436 		aggr_mac_intr.mi_enable = aggr_pseudo_enable_intr;
2437 		aggr_mac_intr.mi_disable = aggr_pseudo_disable_intr;
2438 		aggr_mac_intr.mi_ddi_handle = NULL;
2439 
2440 		infop->mri_driver = (mac_ring_driver_t)rx_ring;
2441 		infop->mri_start = aggr_pseudo_start_rx_ring;
2442 		infop->mri_stop = aggr_pseudo_stop_rx_ring;
2443 
2444 		infop->mri_intr = aggr_mac_intr;
2445 		infop->mri_poll = aggr_rx_poll;
2446 
2447 		infop->mri_stat = aggr_rx_ring_stat;
2448 		break;
2449 	}
2450 	case MAC_RING_TYPE_TX: {
2451 		aggr_pseudo_tx_group_t	*tx_group = &grp->lg_tx_group;
2452 		aggr_pseudo_tx_ring_t	*tx_ring;
2453 
2454 		ASSERT(rg_index == -1);
2455 		ASSERT(index < tx_group->atg_ring_cnt);
2456 
2457 		tx_ring = &tx_group->atg_rings[index];
2458 		tx_ring->atr_rh = rh;
2459 
2460 		infop->mri_driver = (mac_ring_driver_t)tx_ring;
2461 		infop->mri_start = NULL;
2462 		infop->mri_stop = NULL;
2463 		infop->mri_tx = aggr_ring_tx;
2464 		infop->mri_stat = aggr_tx_ring_stat;
2465 		/*
2466 		 * Use the hw TX ring handle to find if the ring needs
2467 		 * serialization or not. For NICs that do not expose
2468 		 * Tx rings, atr_hw_rh will be NULL.
2469 		 */
2470 		if (tx_ring->atr_hw_rh != NULL) {
2471 			infop->mri_flags =
2472 			    mac_hwring_getinfo(tx_ring->atr_hw_rh);
2473 		}
2474 		break;
2475 	}
2476 	default:
2477 		break;
2478 	}
2479 }
2480 
2481 static mblk_t *
2482 aggr_rx_poll(void *arg, int bytes_to_pickup)
2483 {
2484 	aggr_pseudo_rx_ring_t *rr_ring = arg;
2485 	aggr_port_t *port = rr_ring->arr_port;
2486 	aggr_grp_t *grp = port->lp_grp;
2487 	mblk_t *mp_chain, *mp, **mpp;
2488 
2489 	mp_chain = mac_hwring_poll(rr_ring->arr_hw_rh, bytes_to_pickup);
2490 
2491 	if (grp->lg_lacp_mode == AGGR_LACP_OFF)
2492 		return (mp_chain);
2493 
2494 	mpp = &mp_chain;
2495 	while ((mp = *mpp) != NULL) {
2496 		if (MBLKL(mp) >= sizeof (struct ether_header)) {
2497 			struct ether_header *ehp;
2498 
2499 			ehp = (struct ether_header *)mp->b_rptr;
2500 			if (ntohs(ehp->ether_type) == ETHERTYPE_SLOW) {
2501 				*mpp = mp->b_next;
2502 				mp->b_next = NULL;
2503 				aggr_recv_lacp(port,
2504 				    (mac_resource_handle_t)rr_ring, mp);
2505 				continue;
2506 			}
2507 		}
2508 
2509 		if (!port->lp_collector_enabled) {
2510 			*mpp = mp->b_next;
2511 			mp->b_next = NULL;
2512 			freemsg(mp);
2513 			continue;
2514 		}
2515 		mpp = &mp->b_next;
2516 	}
2517 	return (mp_chain);
2518 }
2519 
2520 static int
2521 aggr_addmac(void *arg, const uint8_t *mac_addr)
2522 {
2523 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)arg;
2524 	aggr_unicst_addr_t	*addr, **pprev;
2525 	aggr_grp_t		*grp = rx_group->arg_grp;
2526 	aggr_port_t		*port, *p;
2527 	mac_perim_handle_t	mph;
2528 	int			err = 0;
2529 	uint_t			idx = rx_group->arg_index;
2530 
2531 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2532 
2533 	if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) == 0) {
2534 		mac_perim_exit(mph);
2535 		return (0);
2536 	}
2537 
2538 	/*
2539 	 * Insert this mac address into the list of mac addresses owned by
2540 	 * the aggregation pseudo group.
2541 	 */
2542 	pprev = &rx_group->arg_macaddr;
2543 	while ((addr = *pprev) != NULL) {
2544 		if (bcmp(mac_addr, addr->aua_addr, ETHERADDRL) == 0) {
2545 			mac_perim_exit(mph);
2546 			return (EEXIST);
2547 		}
2548 		pprev = &addr->aua_next;
2549 	}
2550 	addr = kmem_alloc(sizeof (aggr_unicst_addr_t), KM_SLEEP);
2551 	bcopy(mac_addr, addr->aua_addr, ETHERADDRL);
2552 	addr->aua_next = NULL;
2553 	*pprev = addr;
2554 
2555 	for (port = grp->lg_ports; port != NULL; port = port->lp_next)
2556 		if ((err = aggr_port_addmac(port, idx, mac_addr)) != 0)
2557 			break;
2558 
2559 	if (err != 0) {
2560 		for (p = grp->lg_ports; p != port; p = p->lp_next)
2561 			aggr_port_remmac(p, idx, mac_addr);
2562 
2563 		*pprev = NULL;
2564 		kmem_free(addr, sizeof (aggr_unicst_addr_t));
2565 	}
2566 
2567 	mac_perim_exit(mph);
2568 	return (err);
2569 }
2570 
2571 static int
2572 aggr_remmac(void *arg, const uint8_t *mac_addr)
2573 {
2574 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)arg;
2575 	aggr_unicst_addr_t	*addr, **pprev;
2576 	aggr_grp_t		*grp = rx_group->arg_grp;
2577 	aggr_port_t		*port;
2578 	mac_perim_handle_t	mph;
2579 	int			err = 0;
2580 
2581 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2582 
2583 	if (bcmp(mac_addr, grp->lg_addr, ETHERADDRL) == 0) {
2584 		mac_perim_exit(mph);
2585 		return (0);
2586 	}
2587 
2588 	/*
2589 	 * Insert this mac address into the list of mac addresses owned by
2590 	 * the aggregation pseudo group.
2591 	 */
2592 	pprev = &rx_group->arg_macaddr;
2593 	while ((addr = *pprev) != NULL) {
2594 		if (bcmp(mac_addr, addr->aua_addr, ETHERADDRL) != 0) {
2595 			pprev = &addr->aua_next;
2596 			continue;
2597 		}
2598 		break;
2599 	}
2600 	if (addr == NULL) {
2601 		mac_perim_exit(mph);
2602 		return (EINVAL);
2603 	}
2604 
2605 	for (port = grp->lg_ports; port != NULL; port = port->lp_next)
2606 		aggr_port_remmac(port, rx_group->arg_index, mac_addr);
2607 
2608 	*pprev = addr->aua_next;
2609 	kmem_free(addr, sizeof (aggr_unicst_addr_t));
2610 
2611 	mac_perim_exit(mph);
2612 	return (err);
2613 }
2614 
2615 /*
2616  * Search for VID in the Rx group's list and return a pointer if
2617  * found. Otherwise return NULL.
2618  */
2619 static aggr_vlan_t *
2620 aggr_find_vlan(aggr_pseudo_rx_group_t *rx_group, uint16_t vid)
2621 {
2622 	ASSERT(MAC_PERIM_HELD(rx_group->arg_grp->lg_mh));
2623 	for (aggr_vlan_t *avp = list_head(&rx_group->arg_vlans); avp != NULL;
2624 	    avp = list_next(&rx_group->arg_vlans, avp)) {
2625 		if (avp->av_vid == vid)
2626 			return (avp);
2627 	}
2628 
2629 	return (NULL);
2630 }
2631 
2632 /*
2633  * Accept traffic on the specified VID.
2634  *
2635  * Persist VLAN state in the aggr so that ports added later will
2636  * receive the correct filters. In the future it would be nice to
2637  * allow aggr to iterate its clients instead of duplicating state.
2638  */
2639 static int
2640 aggr_addvlan(mac_group_driver_t gdriver, uint16_t vid)
2641 {
2642 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)gdriver;
2643 	aggr_grp_t		*aggr = rx_group->arg_grp;
2644 	aggr_port_t		*port, *p;
2645 	mac_perim_handle_t	mph;
2646 	int			err = 0;
2647 	aggr_vlan_t		*avp = NULL;
2648 	uint_t			idx = rx_group->arg_index;
2649 
2650 	mac_perim_enter_by_mh(aggr->lg_mh, &mph);
2651 
2652 	if (vid == MAC_VLAN_UNTAGGED) {
2653 		/*
2654 		 * Aggr is both a MAC provider and MAC client. As a
2655 		 * MAC provider it is passed MAC_VLAN_UNTAGGED by its
2656 		 * client. As a client itself, it should pass
2657 		 * VLAN_ID_NONE to its ports.
2658 		 */
2659 		vid = VLAN_ID_NONE;
2660 		rx_group->arg_untagged++;
2661 		goto update_ports;
2662 	}
2663 
2664 	avp = aggr_find_vlan(rx_group, vid);
2665 
2666 	if (avp != NULL) {
2667 		avp->av_refs++;
2668 		mac_perim_exit(mph);
2669 		return (0);
2670 	}
2671 
2672 	avp = kmem_zalloc(sizeof (aggr_vlan_t), KM_SLEEP);
2673 	avp->av_vid = vid;
2674 	avp->av_refs = 1;
2675 
2676 update_ports:
2677 	for (port = aggr->lg_ports; port != NULL; port = port->lp_next)
2678 		if ((err = aggr_port_addvlan(port, idx, vid)) != 0)
2679 			break;
2680 
2681 	if (err != 0) {
2682 		/*
2683 		 * If any of these calls fail then we are in a
2684 		 * situation where the ports have different HW state.
2685 		 * There's no reasonable action the MAC client can
2686 		 * take in this scenario to rectify the situation.
2687 		 */
2688 		for (p = aggr->lg_ports; p != port; p = p->lp_next) {
2689 			int err2;
2690 
2691 			if ((err2 = aggr_port_remvlan(p, idx, vid)) != 0) {
2692 				cmn_err(CE_WARN, "Failed to remove VLAN %u"
2693 				    " from port %s: errno %d.", vid,
2694 				    mac_client_name(p->lp_mch), err2);
2695 			}
2696 
2697 		}
2698 
2699 		if (vid == VLAN_ID_NONE)
2700 			rx_group->arg_untagged--;
2701 
2702 		if (avp != NULL) {
2703 			kmem_free(avp, sizeof (aggr_vlan_t));
2704 			avp = NULL;
2705 		}
2706 	}
2707 
2708 	if (avp != NULL)
2709 		list_insert_tail(&rx_group->arg_vlans, avp);
2710 
2711 done:
2712 	mac_perim_exit(mph);
2713 	return (err);
2714 }
2715 
2716 /*
2717  * Stop accepting traffic on this VLAN if it's the last use of this VLAN.
2718  */
2719 static int
2720 aggr_remvlan(mac_group_driver_t gdriver, uint16_t vid)
2721 {
2722 	aggr_pseudo_rx_group_t	*rx_group = (aggr_pseudo_rx_group_t *)gdriver;
2723 	aggr_grp_t		*aggr = rx_group->arg_grp;
2724 	aggr_port_t		*port, *p;
2725 	mac_perim_handle_t	mph;
2726 	int			err = 0;
2727 	aggr_vlan_t		*avp = NULL;
2728 	uint_t			idx = rx_group->arg_index;
2729 
2730 	mac_perim_enter_by_mh(aggr->lg_mh, &mph);
2731 
2732 	/*
2733 	 * See the comment in aggr_addvlan().
2734 	 */
2735 	if (vid == MAC_VLAN_UNTAGGED) {
2736 		vid = VLAN_ID_NONE;
2737 		rx_group->arg_untagged--;
2738 
2739 		if (rx_group->arg_untagged > 0)
2740 			goto done;
2741 
2742 		goto update_ports;
2743 	}
2744 
2745 	avp = aggr_find_vlan(rx_group, vid);
2746 
2747 	if (avp == NULL) {
2748 		err = ENOENT;
2749 		goto done;
2750 	}
2751 
2752 	avp->av_refs--;
2753 
2754 	if (avp->av_refs > 0)
2755 		goto done;
2756 
2757 update_ports:
2758 	for (port = aggr->lg_ports; port != NULL; port = port->lp_next)
2759 		if ((err = aggr_port_remvlan(port, idx, vid)) != 0)
2760 			break;
2761 
2762 	/*
2763 	 * See the comment in aggr_addvlan() for justification of the
2764 	 * use of VERIFY here.
2765 	 */
2766 	if (err != 0) {
2767 		for (p = aggr->lg_ports; p != port; p = p->lp_next) {
2768 			int err2;
2769 
2770 			if ((err2 = aggr_port_addvlan(p, idx, vid)) != 0) {
2771 				cmn_err(CE_WARN, "Failed to add VLAN %u"
2772 				    " to port %s: errno %d.", vid,
2773 				    mac_client_name(p->lp_mch), err2);
2774 			}
2775 		}
2776 
2777 		if (avp != NULL)
2778 			avp->av_refs++;
2779 
2780 		if (vid == VLAN_ID_NONE)
2781 			rx_group->arg_untagged++;
2782 
2783 		goto done;
2784 	}
2785 
2786 	if (err == 0 && avp != NULL) {
2787 		VERIFY3U(avp->av_refs, ==, 0);
2788 		list_remove(&rx_group->arg_vlans, avp);
2789 		kmem_free(avp, sizeof (aggr_vlan_t));
2790 	}
2791 
2792 done:
2793 	mac_perim_exit(mph);
2794 	return (err);
2795 }
2796 
2797 /*
2798  * Add or remove the multicast addresses that are defined for the group
2799  * to or from the specified port.
2800  *
2801  * Note that aggr_grp_multicst_port(..., B_TRUE) is called when the port
2802  * is started and attached, and aggr_grp_multicst_port(..., B_FALSE) is
2803  * called when the port is either stopped or detached.
2804  */
2805 void
2806 aggr_grp_multicst_port(aggr_port_t *port, boolean_t add)
2807 {
2808 	aggr_grp_t *grp = port->lp_grp;
2809 
2810 	ASSERT(MAC_PERIM_HELD(port->lp_mh));
2811 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
2812 
2813 	if (!port->lp_started || port->lp_state != AGGR_PORT_STATE_ATTACHED)
2814 		return;
2815 
2816 	mac_multicast_refresh(grp->lg_mh, aggr_port_multicst, port, add);
2817 }
2818 
2819 static int
2820 aggr_m_multicst(void *arg, boolean_t add, const uint8_t *addrp)
2821 {
2822 	aggr_grp_t *grp = arg;
2823 	aggr_port_t *port = NULL, *errport = NULL;
2824 	mac_perim_handle_t mph;
2825 	int err = 0;
2826 
2827 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2828 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2829 		if (port->lp_state != AGGR_PORT_STATE_ATTACHED ||
2830 		    !port->lp_started) {
2831 			continue;
2832 		}
2833 		err = aggr_port_multicst(port, add, addrp);
2834 		if (err != 0) {
2835 			errport = port;
2836 			break;
2837 		}
2838 	}
2839 
2840 	/*
2841 	 * At least one port caused error return and this error is returned to
2842 	 * mac, eventually a NAK would be sent upwards.
2843 	 * Some ports have this multicast address listed now, and some don't.
2844 	 * Treat this error as a whole aggr failure not individual port failure.
2845 	 * Therefore remove this multicast address from other ports.
2846 	 */
2847 	if ((err != 0) && add) {
2848 		for (port = grp->lg_ports; port != errport;
2849 		    port = port->lp_next) {
2850 			if (port->lp_state != AGGR_PORT_STATE_ATTACHED ||
2851 			    !port->lp_started) {
2852 				continue;
2853 			}
2854 			(void) aggr_port_multicst(port, B_FALSE, addrp);
2855 		}
2856 	}
2857 	mac_perim_exit(mph);
2858 	return (err);
2859 }
2860 
2861 static int
2862 aggr_m_unicst(void *arg, const uint8_t *macaddr)
2863 {
2864 	aggr_grp_t *grp = arg;
2865 	mac_perim_handle_t mph;
2866 	int err;
2867 
2868 	mac_perim_enter_by_mh(grp->lg_mh, &mph);
2869 	err = aggr_grp_modify_common(grp, AGGR_MODIFY_MAC, 0, B_TRUE, macaddr,
2870 	    0, 0);
2871 	mac_perim_exit(mph);
2872 	return (err);
2873 }
2874 
2875 /*
2876  * Initialize the capabilities that are advertised for the group
2877  * according to the capabilities of the constituent ports.
2878  */
2879 static void
2880 aggr_grp_capab_set(aggr_grp_t *grp)
2881 {
2882 	uint32_t cksum;
2883 	aggr_port_t *port;
2884 	mac_capab_lso_t cap_lso;
2885 
2886 	ASSERT(grp->lg_mh == NULL);
2887 	ASSERT(grp->lg_ports != NULL);
2888 
2889 	grp->lg_hcksum_txflags = (uint32_t)-1;
2890 	grp->lg_zcopy = B_TRUE;
2891 	grp->lg_vlan = B_TRUE;
2892 
2893 	grp->lg_lso = B_TRUE;
2894 	grp->lg_cap_lso.lso_flags = (t_uscalar_t)-1;
2895 	grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max = (t_uscalar_t)-1;
2896 
2897 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2898 		if (!mac_capab_get(port->lp_mh, MAC_CAPAB_HCKSUM, &cksum))
2899 			cksum = 0;
2900 		grp->lg_hcksum_txflags &= cksum;
2901 
2902 		grp->lg_vlan &=
2903 		    !mac_capab_get(port->lp_mh, MAC_CAPAB_NO_NATIVEVLAN, NULL);
2904 
2905 		grp->lg_zcopy &=
2906 		    !mac_capab_get(port->lp_mh, MAC_CAPAB_NO_ZCOPY, NULL);
2907 
2908 		grp->lg_lso &=
2909 		    mac_capab_get(port->lp_mh, MAC_CAPAB_LSO, &cap_lso);
2910 		if (grp->lg_lso) {
2911 			grp->lg_cap_lso.lso_flags &= cap_lso.lso_flags;
2912 			if (grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max >
2913 			    cap_lso.lso_basic_tcp_ipv4.lso_max)
2914 				grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max =
2915 				    cap_lso.lso_basic_tcp_ipv4.lso_max;
2916 		}
2917 	}
2918 }
2919 
2920 /*
2921  * Checks whether the capabilities of the port being added are compatible
2922  * with the current capabilities of the aggregation.
2923  */
2924 static boolean_t
2925 aggr_grp_capab_check(aggr_grp_t *grp, aggr_port_t *port)
2926 {
2927 	uint32_t hcksum_txflags;
2928 
2929 	ASSERT(grp->lg_ports != NULL);
2930 
2931 	if (((!mac_capab_get(port->lp_mh, MAC_CAPAB_NO_NATIVEVLAN, NULL)) &
2932 	    grp->lg_vlan) != grp->lg_vlan) {
2933 		return (B_FALSE);
2934 	}
2935 
2936 	if (((!mac_capab_get(port->lp_mh, MAC_CAPAB_NO_ZCOPY, NULL)) &
2937 	    grp->lg_zcopy) != grp->lg_zcopy) {
2938 		return (B_FALSE);
2939 	}
2940 
2941 	if (!mac_capab_get(port->lp_mh, MAC_CAPAB_HCKSUM, &hcksum_txflags)) {
2942 		if (grp->lg_hcksum_txflags != 0)
2943 			return (B_FALSE);
2944 	} else if ((hcksum_txflags & grp->lg_hcksum_txflags) !=
2945 	    grp->lg_hcksum_txflags) {
2946 		return (B_FALSE);
2947 	}
2948 
2949 	if (grp->lg_lso) {
2950 		mac_capab_lso_t cap_lso;
2951 
2952 		if (mac_capab_get(port->lp_mh, MAC_CAPAB_LSO, &cap_lso)) {
2953 			if ((grp->lg_cap_lso.lso_flags & cap_lso.lso_flags) !=
2954 			    grp->lg_cap_lso.lso_flags)
2955 				return (B_FALSE);
2956 			if (grp->lg_cap_lso.lso_basic_tcp_ipv4.lso_max >
2957 			    cap_lso.lso_basic_tcp_ipv4.lso_max)
2958 				return (B_FALSE);
2959 		} else {
2960 			return (B_FALSE);
2961 		}
2962 	}
2963 
2964 	return (B_TRUE);
2965 }
2966 
2967 /*
2968  * Returns the maximum SDU according to the SDU of the constituent ports.
2969  */
2970 static uint_t
2971 aggr_grp_max_sdu(aggr_grp_t *grp)
2972 {
2973 	uint_t max_sdu = (uint_t)-1;
2974 	aggr_port_t *port;
2975 
2976 	ASSERT(grp->lg_ports != NULL);
2977 
2978 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
2979 		uint_t port_sdu_max;
2980 
2981 		mac_sdu_get(port->lp_mh, NULL, &port_sdu_max);
2982 		if (max_sdu > port_sdu_max)
2983 			max_sdu = port_sdu_max;
2984 	}
2985 
2986 	return (max_sdu);
2987 }
2988 
2989 /*
2990  * Checks if the maximum SDU of the specified port is compatible
2991  * with the maximum SDU of the specified aggregation group, returns
2992  * B_TRUE if it is, B_FALSE otherwise.
2993  */
2994 static boolean_t
2995 aggr_grp_sdu_check(aggr_grp_t *grp, aggr_port_t *port)
2996 {
2997 	uint_t port_sdu_max;
2998 
2999 	mac_sdu_get(port->lp_mh, NULL, &port_sdu_max);
3000 	return (port_sdu_max >= grp->lg_max_sdu);
3001 }
3002 
3003 /*
3004  * Returns the maximum margin according to the margin of the constituent ports.
3005  */
3006 static uint32_t
3007 aggr_grp_max_margin(aggr_grp_t *grp)
3008 {
3009 	uint32_t margin = UINT32_MAX;
3010 	aggr_port_t *port;
3011 
3012 	ASSERT(grp->lg_mh == NULL);
3013 	ASSERT(grp->lg_ports != NULL);
3014 
3015 	for (port = grp->lg_ports; port != NULL; port = port->lp_next) {
3016 		if (margin > port->lp_margin)
3017 			margin = port->lp_margin;
3018 	}
3019 
3020 	grp->lg_margin = margin;
3021 	return (margin);
3022 }
3023 
3024 /*
3025  * Checks if the maximum margin of the specified port is compatible
3026  * with the maximum margin of the specified aggregation group, returns
3027  * B_TRUE if it is, B_FALSE otherwise.
3028  */
3029 static boolean_t
3030 aggr_grp_margin_check(aggr_grp_t *grp, aggr_port_t *port)
3031 {
3032 	if (port->lp_margin >= grp->lg_margin)
3033 		return (B_TRUE);
3034 
3035 	/*
3036 	 * See whether the current margin value is allowed to be changed to
3037 	 * the new value.
3038 	 */
3039 	if (!mac_margin_update(grp->lg_mh, port->lp_margin))
3040 		return (B_FALSE);
3041 
3042 	grp->lg_margin = port->lp_margin;
3043 	return (B_TRUE);
3044 }
3045 
3046 /*
3047  * Set MTU on individual ports of an aggregation group
3048  */
3049 static int
3050 aggr_set_port_sdu(aggr_grp_t *grp, aggr_port_t *port, uint32_t sdu,
3051     uint32_t *old_mtu)
3052 {
3053 	boolean_t		removed = B_FALSE;
3054 	mac_perim_handle_t	mph;
3055 	mac_diag_t		diag;
3056 	int			err, rv, retry = 0;
3057 
3058 	if (port->lp_mah != NULL) {
3059 		(void) mac_unicast_remove(port->lp_mch, port->lp_mah);
3060 		port->lp_mah = NULL;
3061 		removed = B_TRUE;
3062 	}
3063 	err = mac_set_mtu(port->lp_mh, sdu, old_mtu);
3064 try_again:
3065 	if (removed && (rv = mac_unicast_add(port->lp_mch, NULL,
3066 	    MAC_UNICAST_PRIMARY | MAC_UNICAST_DISABLE_TX_VID_CHECK,
3067 	    &port->lp_mah, 0, &diag)) != 0) {
3068 		/*
3069 		 * following is a workaround for a bug in 'bge' driver.
3070 		 * See CR 6794654 for more information and this work around
3071 		 * will be removed once the CR is fixed.
3072 		 */
3073 		if (rv == EIO && retry++ < 3) {
3074 			delay(2 * hz);
3075 			goto try_again;
3076 		}
3077 		/*
3078 		 * if mac_unicast_add() failed while setting the MTU,
3079 		 * detach the port from the group.
3080 		 */
3081 		mac_perim_enter_by_mh(port->lp_mh, &mph);
3082 		(void) aggr_grp_detach_port(grp, port);
3083 		mac_perim_exit(mph);
3084 		cmn_err(CE_WARN, "Unable to restart the port %s while "
3085 		    "setting MTU. Detaching the port from the aggregation.",
3086 		    mac_client_name(port->lp_mch));
3087 	}
3088 	return (err);
3089 }
3090 
3091 static int
3092 aggr_sdu_update(aggr_grp_t *grp, uint32_t sdu)
3093 {
3094 	int			err = 0, i, rv;
3095 	aggr_port_t		*port;
3096 	uint32_t		*mtu;
3097 
3098 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
3099 
3100 	/*
3101 	 * If the MTU being set is equal to aggr group's maximum
3102 	 * allowable value, then there is nothing to change
3103 	 */
3104 	if (sdu == grp->lg_max_sdu)
3105 		return (0);
3106 
3107 	/* 0 is aggr group's min sdu */
3108 	if (sdu == 0)
3109 		return (EINVAL);
3110 
3111 	mtu = kmem_alloc(sizeof (uint32_t) * grp->lg_nports, KM_SLEEP);
3112 	for (port = grp->lg_ports, i = 0; port != NULL && err == 0;
3113 	    port = port->lp_next, i++) {
3114 		err = aggr_set_port_sdu(grp, port, sdu, mtu + i);
3115 	}
3116 	if (err != 0) {
3117 		/* recover from error: reset the mtus of the ports */
3118 		aggr_port_t *tmp;
3119 
3120 		for (tmp = grp->lg_ports, i = 0; tmp != port;
3121 		    tmp = tmp->lp_next, i++) {
3122 			(void) aggr_set_port_sdu(grp, tmp, *(mtu + i), NULL);
3123 		}
3124 		goto bail;
3125 	}
3126 	grp->lg_max_sdu = aggr_grp_max_sdu(grp);
3127 	rv = mac_maxsdu_update(grp->lg_mh, grp->lg_max_sdu);
3128 	ASSERT(rv == 0);
3129 bail:
3130 	kmem_free(mtu, sizeof (uint32_t) * grp->lg_nports);
3131 	return (err);
3132 }
3133 
3134 /*
3135  * Callback functions for set/get of properties
3136  */
3137 /*ARGSUSED*/
3138 static int
3139 aggr_m_setprop(void *m_driver, const char *pr_name, mac_prop_id_t pr_num,
3140     uint_t pr_valsize, const void *pr_val)
3141 {
3142 	int		err = ENOTSUP;
3143 	aggr_grp_t	*grp = m_driver;
3144 
3145 	switch (pr_num) {
3146 	case MAC_PROP_MTU: {
3147 		uint32_t	mtu;
3148 
3149 		if (pr_valsize < sizeof (mtu)) {
3150 			err = EINVAL;
3151 			break;
3152 		}
3153 		bcopy(pr_val, &mtu, sizeof (mtu));
3154 		err = aggr_sdu_update(grp, mtu);
3155 		break;
3156 	}
3157 	default:
3158 		break;
3159 	}
3160 	return (err);
3161 }
3162 
3163 typedef struct rboundary {
3164 	uint32_t	bval;
3165 	int		btype;
3166 } rboundary_t;
3167 
3168 /*
3169  * This function finds the intersection of mtu ranges stored in arrays -
3170  * mrange[0] ... mrange[mcount -1]. It returns the intersection in rval.
3171  * Individual arrays are assumed to contain non-overlapping ranges.
3172  * Algorithm:
3173  *   A range has two boundaries - min and max. We scan all arrays and store
3174  * each boundary as a separate element in a temporary array. We also store
3175  * the boundary types, min or max, as +1 or -1 respectively in the temporary
3176  * array. Then we sort the temporary array in ascending order. We scan the
3177  * sorted array from lower to higher values and keep a cumulative sum of
3178  * boundary types. Element in the temporary array for which the sum reaches
3179  * mcount is a min boundary of a range in the result and next element will be
3180  * max boundary.
3181  *
3182  * Example for mcount = 3,
3183  *
3184  *  ----|_________|-------|_______|----|__|------ mrange[0]
3185  *
3186  *  -------|________|--|____________|-----|___|-- mrange[1]
3187  *
3188  *  --------|________________|-------|____|------ mrange[2]
3189  *
3190  *                                      3 2 1
3191  *                                       \|/
3192  *      1  23     2 1  2  3  2    1 01 2  V   0  <- the sum
3193  *  ----|--||-----|-|--|--|--|----|-||-|--|---|-- sorted array
3194  *
3195  *                                 same min and max
3196  *                                        V
3197  *  --------|_____|-------|__|------------|------ intersecting ranges
3198  */
3199 void
3200 aggr_mtu_range_intersection(mac_propval_range_t **mrange, int mcount,
3201     mac_propval_uint32_range_t **prval, int *prmaxcnt, int *prcount)
3202 {
3203 	mac_propval_uint32_range_t	*rval, *ur;
3204 	int				rmaxcnt, rcount;
3205 	size_t				sz_range32;
3206 	rboundary_t			*ta; /* temporary array */
3207 	rboundary_t			temp;
3208 	boolean_t			range_started = B_FALSE;
3209 	int				i, j, m, sum;
3210 
3211 	sz_range32 = sizeof (mac_propval_uint32_range_t);
3212 
3213 	for (i = 0, rmaxcnt = 0; i < mcount; i++)
3214 		rmaxcnt += mrange[i]->mpr_count;
3215 
3216 	/* Allocate enough space to store the results */
3217 	rval = kmem_alloc(rmaxcnt * sz_range32, KM_SLEEP);
3218 
3219 	/* Number of boundaries are twice as many as ranges */
3220 	ta = kmem_alloc(2 * rmaxcnt * sizeof (rboundary_t), KM_SLEEP);
3221 
3222 	for (i = 0, m = 0; i < mcount; i++) {
3223 		ur = &(mrange[i]->mpr_range_uint32[0]);
3224 		for (j = 0; j < mrange[i]->mpr_count; j++) {
3225 			ta[m].bval = ur[j].mpur_min;
3226 			ta[m++].btype = 1;
3227 			ta[m].bval = ur[j].mpur_max;
3228 			ta[m++].btype = -1;
3229 		}
3230 	}
3231 
3232 	/*
3233 	 * Sort the temporary array in ascending order of bval;
3234 	 * if boundary values are same then sort on btype.
3235 	 */
3236 	for (i = 0; i < m-1; i++) {
3237 		for (j = i+1; j < m; j++) {
3238 			if ((ta[i].bval > ta[j].bval) ||
3239 			    ((ta[i].bval == ta[j].bval) &&
3240 			    (ta[i].btype < ta[j].btype))) {
3241 				temp = ta[i];
3242 				ta[i] = ta[j];
3243 				ta[j] = temp;
3244 			}
3245 		}
3246 	}
3247 
3248 	/* Walk through temporary array to find all ranges in the results */
3249 	for (i = 0, sum = 0, rcount = 0; i < m; i++) {
3250 		sum += ta[i].btype;
3251 		if (sum == mcount) {
3252 			rval[rcount].mpur_min = ta[i].bval;
3253 			range_started = B_TRUE;
3254 		} else if (sum < mcount && range_started) {
3255 			rval[rcount++].mpur_max = ta[i].bval;
3256 			range_started = B_FALSE;
3257 		}
3258 	}
3259 
3260 	*prval = rval;
3261 	*prmaxcnt = rmaxcnt;
3262 	*prcount = rcount;
3263 
3264 	kmem_free(ta, 2 * rmaxcnt * sizeof (rboundary_t));
3265 }
3266 
3267 /*
3268  * Returns the mtu ranges which could be supported by aggr group.
3269  * prmaxcnt returns the size of the buffer prval, prcount returns
3270  * the number of valid entries in prval. Caller is responsible
3271  * for freeing up prval.
3272  */
3273 int
3274 aggr_grp_possible_mtu_range(aggr_grp_t *grp, mac_propval_uint32_range_t **prval,
3275     int *prmaxcnt, int *prcount)
3276 {
3277 	mac_propval_range_t		**vals;
3278 	aggr_port_t			*port;
3279 	mac_perim_handle_t		mph;
3280 	uint_t				i, numr;
3281 	int				err = 0;
3282 	size_t				sz_propval, sz_range32;
3283 	size_t				size;
3284 
3285 	sz_propval = sizeof (mac_propval_range_t);
3286 	sz_range32 = sizeof (mac_propval_uint32_range_t);
3287 
3288 	ASSERT(MAC_PERIM_HELD(grp->lg_mh));
3289 
3290 	vals = kmem_zalloc(sizeof (mac_propval_range_t *) * grp->lg_nports,
3291 	    KM_SLEEP);
3292 
3293 	for (port = grp->lg_ports, i = 0; port != NULL;
3294 	    port = port->lp_next, i++) {
3295 
3296 		size = sz_propval;
3297 		vals[i] = kmem_alloc(size, KM_SLEEP);
3298 		vals[i]->mpr_count = 1;
3299 
3300 		mac_perim_enter_by_mh(port->lp_mh, &mph);
3301 
3302 		err = mac_prop_info(port->lp_mh, MAC_PROP_MTU, NULL,
3303 		    NULL, 0, vals[i], NULL);
3304 		if (err == ENOSPC) {
3305 			/*
3306 			 * Not enough space to hold all ranges.
3307 			 * Allocate extra space as indicated and retry.
3308 			 */
3309 			numr = vals[i]->mpr_count;
3310 			kmem_free(vals[i], sz_propval);
3311 			size = sz_propval + (numr - 1) * sz_range32;
3312 			vals[i] = kmem_alloc(size, KM_SLEEP);
3313 			vals[i]->mpr_count = numr;
3314 			err = mac_prop_info(port->lp_mh, MAC_PROP_MTU, NULL,
3315 			    NULL, 0, vals[i], NULL);
3316 			ASSERT(err != ENOSPC);
3317 		}
3318 		mac_perim_exit(mph);
3319 		if (err != 0) {
3320 			kmem_free(vals[i], size);
3321 			vals[i] = NULL;
3322 			break;
3323 		}
3324 	}
3325 
3326 	/*
3327 	 * if any of the underlying ports does not support changing MTU then
3328 	 * just return ENOTSUP
3329 	 */
3330 	if (port != NULL) {
3331 		ASSERT(err != 0);
3332 		goto done;
3333 	}
3334 
3335 	aggr_mtu_range_intersection(vals, grp->lg_nports, prval, prmaxcnt,
3336 	    prcount);
3337 
3338 done:
3339 	for (i = 0; i < grp->lg_nports; i++) {
3340 		if (vals[i] != NULL) {
3341 			numr = vals[i]->mpr_count;
3342 			size = sz_propval + (numr - 1) * sz_range32;
3343 			kmem_free(vals[i], size);
3344 		}
3345 	}
3346 
3347 	kmem_free(vals, sizeof (mac_propval_range_t *) * grp->lg_nports);
3348 	return (err);
3349 }
3350 
3351 static void
3352 aggr_m_propinfo(void *m_driver, const char *pr_name, mac_prop_id_t pr_num,
3353     mac_prop_info_handle_t prh)
3354 {
3355 	aggr_grp_t			*grp = m_driver;
3356 	mac_propval_uint32_range_t	*rval = NULL;
3357 	int				i, rcount, rmaxcnt;
3358 	int				err = 0;
3359 
3360 	_NOTE(ARGUNUSED(pr_name));
3361 
3362 	switch (pr_num) {
3363 	case MAC_PROP_MTU:
3364 
3365 		err = aggr_grp_possible_mtu_range(grp, &rval, &rmaxcnt,
3366 		    &rcount);
3367 		if (err != 0) {
3368 			ASSERT(rval == NULL);
3369 			return;
3370 		}
3371 		for (i = 0; i < rcount; i++) {
3372 			mac_prop_info_set_range_uint32(prh,
3373 			    rval[i].mpur_min, rval[i].mpur_max);
3374 		}
3375 		kmem_free(rval, sizeof (mac_propval_uint32_range_t) * rmaxcnt);
3376 		break;
3377 	}
3378 }
3379