xref: /titanic_41/usr/src/uts/common/io/mac/mac_sched.c (revision 80e2ca8596e3435bc3b76f3c597833ea0a87f85e)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/callb.h>
28 #include <sys/sdt.h>
29 #include <sys/strsubr.h>
30 #include <sys/strsun.h>
31 #include <sys/vlan.h>
32 #include <inet/ipsec_impl.h>
33 #include <inet/ip_impl.h>
34 #include <inet/sadb.h>
35 #include <inet/ipsecesp.h>
36 #include <inet/ipsecah.h>
37 #include <inet/ip6.h>
38 
39 #include <sys/mac_impl.h>
40 #include <sys/mac_client_impl.h>
41 #include <sys/mac_client_priv.h>
42 #include <sys/mac_soft_ring.h>
43 #include <sys/mac_flow_impl.h>
44 
45 static mac_tx_cookie_t mac_tx_single_ring_mode(mac_soft_ring_set_t *, mblk_t *,
46     uintptr_t, uint16_t, mblk_t **);
47 static mac_tx_cookie_t mac_tx_serializer_mode(mac_soft_ring_set_t *, mblk_t *,
48     uintptr_t, uint16_t, mblk_t **);
49 static mac_tx_cookie_t mac_tx_fanout_mode(mac_soft_ring_set_t *, mblk_t *,
50     uintptr_t, uint16_t, mblk_t **);
51 static mac_tx_cookie_t mac_tx_bw_mode(mac_soft_ring_set_t *, mblk_t *,
52     uintptr_t, uint16_t, mblk_t **);
53 static mac_tx_cookie_t mac_tx_aggr_mode(mac_soft_ring_set_t *, mblk_t *,
54     uintptr_t, uint16_t, mblk_t **);
55 
56 typedef struct mac_tx_mode_s {
57 	mac_tx_srs_mode_t	mac_tx_mode;
58 	mac_tx_func_t		mac_tx_func;
59 } mac_tx_mode_t;
60 
61 /*
62  * There are seven modes of operation on the Tx side. These modes get set
63  * in mac_tx_srs_setup(). Except for the experimental TX_SERIALIZE mode,
64  * none of the other modes are user configurable. They get selected by
65  * the system depending upon whether the link (or flow) has multiple Tx
66  * rings or a bandwidth configured, or if the link is an aggr, etc.
67  *
68  * When the Tx SRS is operating in aggr mode (st_mode) or if there are
69  * multiple Tx rings owned by Tx SRS, then each Tx ring (pseudo or
70  * otherwise) will have a soft ring associated with it. These soft rings
71  * are stored in srs_tx_soft_rings[] array.
72  *
73  * Additionally in the case of aggr, there is the st_soft_rings[] array
74  * in the mac_srs_tx_t structure. This array is used to store the same
75  * set of soft rings that are present in srs_tx_soft_rings[] array but
76  * in a different manner. The soft ring associated with the pseudo Tx
77  * ring is saved at mr_index (of the pseudo ring) in st_soft_rings[]
78  * array. This helps in quickly getting the soft ring associated with the
79  * Tx ring when aggr_find_tx_ring() returns the pseudo Tx ring that is to
80  * be used for transmit.
81  */
82 mac_tx_mode_t mac_tx_mode_list[] = {
83 	{SRS_TX_DEFAULT,	mac_tx_single_ring_mode},
84 	{SRS_TX_SERIALIZE,	mac_tx_serializer_mode},
85 	{SRS_TX_FANOUT,		mac_tx_fanout_mode},
86 	{SRS_TX_BW,		mac_tx_bw_mode},
87 	{SRS_TX_BW_FANOUT,	mac_tx_bw_mode},
88 	{SRS_TX_AGGR,		mac_tx_aggr_mode},
89 	{SRS_TX_BW_AGGR,	mac_tx_bw_mode}
90 };
91 
92 /*
93  * Soft Ring Set (SRS) - The Run time code that deals with
94  * dynamic polling from the hardware, bandwidth enforcement,
95  * fanout etc.
96  *
97  * We try to use H/W classification on NIC and assign traffic for
98  * a MAC address to a particular Rx ring or ring group. There is a
99  * 1-1 mapping between a SRS and a Rx ring. The SRS dynamically
100  * switches the underlying Rx ring between interrupt and
101  * polling mode and enforces any specified B/W control.
102  *
103  * There is always a SRS created and tied to each H/W and S/W rule.
104  * Whenever we create a H/W rule, we always add the the same rule to
105  * S/W classifier and tie a SRS to it.
106  *
107  * In case a B/W control is specified, it is broken into bytes
108  * per ticks and as soon as the quota for a tick is exhausted,
109  * the underlying Rx ring is forced into poll mode for remainder of
110  * the tick. The SRS poll thread only polls for bytes that are
111  * allowed to come in the SRS. We typically let 4x the configured
112  * B/W worth of packets to come in the SRS (to prevent unnecessary
113  * drops due to bursts) but only process the specified amount.
114  *
115  * A MAC client (e.g. a VNIC or aggr) can have 1 or more
116  * Rx rings (and corresponding SRSs) assigned to it. The SRS
117  * in turn can have softrings to do protocol level fanout or
118  * softrings to do S/W based fanout or both. In case the NIC
119  * has no Rx rings, we do S/W classification to respective SRS.
120  * The S/W classification rule is always setup and ready. This
121  * allows the MAC layer to reassign Rx rings whenever needed
122  * but packets still continue to flow via the default path and
123  * getting S/W classified to correct SRS.
124  *
125  * The SRS's are used on both Tx and Rx side. They use the same
126  * data structure but the processing routines have slightly different
127  * semantics due to the fact that Rx side needs to do dynamic
128  * polling etc.
129  *
130  * Dynamic Polling Notes
131  * =====================
132  *
133  * Each Soft ring set is capable of switching its Rx ring between
134  * interrupt and poll mode and actively 'polls' for packets in
135  * poll mode. If the SRS is implementing a B/W limit, it makes
136  * sure that only Max allowed packets are pulled in poll mode
137  * and goes to poll mode as soon as B/W limit is exceeded. As
138  * such, there are no overheads to implement B/W limits.
139  *
140  * In poll mode, its better to keep the pipeline going where the
141  * SRS worker thread keeps processing packets and poll thread
142  * keeps bringing more packets (specially if they get to run
143  * on different CPUs). This also prevents the overheads associated
144  * by excessive signalling (on NUMA machines, this can be
145  * pretty devastating). The exception is latency optimized case
146  * where worker thread does no work and interrupt and poll thread
147  * are allowed to do their own drain.
148  *
149  * We use the following policy to control Dynamic Polling:
150  * 1) We switch to poll mode anytime the processing
151  *    thread causes a backlog to build up in SRS and
152  *    its associated Soft Rings (sr_poll_pkt_cnt > 0).
153  * 2) As long as the backlog stays under the low water
154  *    mark (sr_lowat), we poll the H/W for more packets.
155  * 3) If the backlog (sr_poll_pkt_cnt) exceeds low
156  *    water mark, we stay in poll mode but don't poll
157  *    the H/W for more packets.
158  * 4) Anytime in polling mode, if we poll the H/W for
159  *    packets and find nothing plus we have an existing
160  *    backlog (sr_poll_pkt_cnt > 0), we stay in polling
161  *    mode but don't poll the H/W for packets anymore
162  *    (let the polling thread go to sleep).
163  * 5) Once the backlog is relived (packets are processed)
164  *    we reenable polling (by signalling the poll thread)
165  *    only when the backlog dips below sr_poll_thres.
166  * 6) sr_hiwat is used exclusively when we are not
167  *    polling capable and is used to decide when to
168  *    drop packets so the SRS queue length doesn't grow
169  *    infinitely.
170  *
171  * NOTE: Also see the block level comment on top of mac_soft_ring.c
172  */
173 
174 /*
175  * mac_latency_optimize
176  *
177  * Controls whether the poll thread can process the packets inline
178  * or let the SRS worker thread do the processing. This applies if
179  * the SRS was not being processed. For latency sensitive traffic,
180  * this needs to be true to allow inline processing. For throughput
181  * under load, this should be false.
182  *
183  * This (and other similar) tunable should be rolled into a link
184  * or flow specific workload hint that can be set using dladm
185  * linkprop (instead of multiple such tunables).
186  */
187 boolean_t mac_latency_optimize = B_TRUE;
188 
189 /*
190  * MAC_RX_SRS_ENQUEUE_CHAIN and MAC_TX_SRS_ENQUEUE_CHAIN
191  *
192  * queue a mp or chain in soft ring set and increment the
193  * local count (srs_count) for the SRS and the shared counter
194  * (srs_poll_pkt_cnt - shared between SRS and its soft rings
195  * to track the total unprocessed packets for polling to work
196  * correctly).
197  *
198  * The size (total bytes queued) counters are incremented only
199  * if we are doing B/W control.
200  */
201 #define	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {		\
202 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
203 	if ((mac_srs)->srs_last != NULL)				\
204 		(mac_srs)->srs_last->b_next = (head);			\
205 	else								\
206 		(mac_srs)->srs_first = (head);				\
207 	(mac_srs)->srs_last = (tail);					\
208 	(mac_srs)->srs_count += count;					\
209 }
210 
211 #define	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
212 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
213 									\
214 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
215 	srs_rx->sr_poll_pkt_cnt += count;				\
216 	ASSERT(srs_rx->sr_poll_pkt_cnt > 0);				\
217 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
218 		(mac_srs)->srs_size += (sz);				\
219 		mutex_enter(&(mac_srs)->srs_bw->mac_bw_lock);		\
220 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
221 		mutex_exit(&(mac_srs)->srs_bw->mac_bw_lock);		\
222 	}								\
223 }
224 
225 #define	MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
226 	mac_srs->srs_state |= SRS_ENQUEUED;				\
227 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
228 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
229 		(mac_srs)->srs_size += (sz);				\
230 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
231 	}								\
232 }
233 
234 /*
235  * Turn polling on routines
236  */
237 #define	MAC_SRS_POLLING_ON(mac_srs) {					\
238 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
239 	if (((mac_srs)->srs_state &					\
240 	    (SRS_POLLING_CAPAB|SRS_POLLING)) == SRS_POLLING_CAPAB) {	\
241 		(mac_srs)->srs_state |= SRS_POLLING;			\
242 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
243 		    (mac_srs)->srs_ring);				\
244 		(mac_srs)->srs_rx.sr_poll_on++;				\
245 	}								\
246 }
247 
248 #define	MAC_SRS_WORKER_POLLING_ON(mac_srs) {				\
249 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
250 	if (((mac_srs)->srs_state &					\
251 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_POLLING)) == 		\
252 	    (SRS_POLLING_CAPAB|SRS_WORKER)) {				\
253 		(mac_srs)->srs_state |= SRS_POLLING;			\
254 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
255 		    (mac_srs)->srs_ring);				\
256 		(mac_srs)->srs_rx.sr_worker_poll_on++;			\
257 	}								\
258 }
259 
260 /*
261  * MAC_SRS_POLL_RING
262  *
263  * Signal the SRS poll thread to poll the underlying H/W ring
264  * provided it wasn't already polling (SRS_GET_PKTS was set).
265  *
266  * Poll thread gets to run only from mac_rx_srs_drain() and only
267  * if the drain was being done by the worker thread.
268  */
269 #define	MAC_SRS_POLL_RING(mac_srs) {					\
270 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
271 									\
272 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
273 	srs_rx->sr_poll_thr_sig++;					\
274 	if (((mac_srs)->srs_state & 					\
275 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_GET_PKTS)) ==		\
276 		(SRS_WORKER|SRS_POLLING_CAPAB)) {			\
277 		(mac_srs)->srs_state |= SRS_GET_PKTS;			\
278 		cv_signal(&(mac_srs)->srs_cv);   			\
279 	} else {							\
280 		srs_rx->sr_poll_thr_busy++;				\
281 	}								\
282 }
283 
284 /*
285  * MAC_SRS_CHECK_BW_CONTROL
286  *
287  * Check to see if next tick has started so we can reset the
288  * SRS_BW_ENFORCED flag and allow more packets to come in the
289  * system.
290  */
291 #define	MAC_SRS_CHECK_BW_CONTROL(mac_srs) {				\
292 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
293 	ASSERT(((mac_srs)->srs_type & SRST_TX) ||			\
294 	    MUTEX_HELD(&(mac_srs)->srs_bw->mac_bw_lock));		\
295 	clock_t now = ddi_get_lbolt();					\
296 	if ((mac_srs)->srs_bw->mac_bw_curr_time != now) {		\
297 		(mac_srs)->srs_bw->mac_bw_curr_time = now;		\
298 		(mac_srs)->srs_bw->mac_bw_used = 0;	       		\
299 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
300 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
301 	}								\
302 }
303 
304 /*
305  * MAC_SRS_WORKER_WAKEUP
306  *
307  * Wake up the SRS worker thread to process the queue as long as
308  * no one else is processing the queue. If we are optimizing for
309  * latency, we wake up the worker thread immediately or else we
310  * wait mac_srs_worker_wakeup_ticks before worker thread gets
311  * woken up.
312  */
313 int mac_srs_worker_wakeup_ticks = 0;
314 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
315 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
316 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
317 		(mac_srs)->srs_tid == NULL) {				\
318 		if (((mac_srs)->srs_state & SRS_LATENCY_OPT) ||		\
319 			(mac_srs_worker_wakeup_ticks == 0))		\
320 			cv_signal(&(mac_srs)->srs_async);		\
321 		else							\
322 			(mac_srs)->srs_tid =				\
323 				timeout(mac_srs_fire, (mac_srs),	\
324 					mac_srs_worker_wakeup_ticks);	\
325 	}								\
326 }
327 
328 #define	TX_BANDWIDTH_MODE(mac_srs)				\
329 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
330 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT ||	\
331 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_AGGR)
332 
333 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
334 	if (tx_mode == SRS_TX_BW_FANOUT)				\
335 		(void) mac_tx_fanout_mode(mac_srs, head, hint, 0, NULL);\
336 	else								\
337 		(void) mac_tx_aggr_mode(mac_srs, head, hint, 0, NULL);	\
338 }
339 
340 /*
341  * MAC_TX_SRS_BLOCK
342  *
343  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
344  * will be set only if srs_tx_woken_up is FALSE. If
345  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
346  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
347  * attempt to transmit again and not setting SRS_TX_BLOCKED does
348  * that.
349  */
350 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
351 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
352 	if ((srs)->srs_tx.st_woken_up) {			\
353 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
354 	} else {						\
355 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
356 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
357 		(srs)->srs_tx.st_stat.mts_blockcnt++;		\
358 	}							\
359 }
360 
361 /*
362  * MAC_TX_SRS_TEST_HIWAT
363  *
364  * Called before queueing a packet onto Tx SRS to test and set
365  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
366  */
367 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
368 	boolean_t enqueue = 1;						\
369 									\
370 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
371 		/*							\
372 		 * flow-controlled. Store srs in cookie so that it	\
373 		 * can be returned as mac_tx_cookie_t to client		\
374 		 */							\
375 		(srs)->srs_state |= SRS_TX_HIWAT;			\
376 		cookie = (mac_tx_cookie_t)srs;				\
377 		(srs)->srs_tx.st_hiwat_cnt++;				\
378 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
379 			/* increment freed stats */			\
380 			(srs)->srs_tx.st_stat.mts_sdrops += cnt;	\
381 			/*						\
382 			 * b_prev may be set to the fanout hint		\
383 			 * hence can't use freemsg directly		\
384 			 */						\
385 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);	\
386 			DTRACE_PROBE1(tx_queued_hiwat,			\
387 			    mac_soft_ring_set_t *, srs);		\
388 			enqueue = 0;					\
389 		}							\
390 	}								\
391 	if (enqueue)							\
392 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
393 }
394 
395 /* Some utility macros */
396 #define	MAC_SRS_BW_LOCK(srs)						\
397 	if (!(srs->srs_type & SRST_TX))					\
398 		mutex_enter(&srs->srs_bw->mac_bw_lock);
399 
400 #define	MAC_SRS_BW_UNLOCK(srs)						\
401 	if (!(srs->srs_type & SRST_TX))					\
402 		mutex_exit(&srs->srs_bw->mac_bw_lock);
403 
404 #define	MAC_TX_SRS_DROP_MESSAGE(srs, mp, cookie) {		\
405 	mac_pkt_drop(NULL, NULL, mp, B_FALSE);			\
406 	/* increment freed stats */				\
407 	mac_srs->srs_tx.st_stat.mts_sdrops++;			\
408 	cookie = (mac_tx_cookie_t)srs;				\
409 }
410 
411 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
412 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
413 	cookie = (mac_tx_cookie_t)srs;					\
414 	*ret_mp = mp_chain;						\
415 }
416 
417 /*
418  * Drop the rx packet and advance to the next one in the chain.
419  */
420 static void
421 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
422 {
423 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
424 
425 	ASSERT(mp->b_next == NULL);
426 	mutex_enter(&srs->srs_lock);
427 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
428 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
429 	mutex_exit(&srs->srs_lock);
430 
431 	srs_rx->sr_stat.mrs_sdrops++;
432 	freemsg(mp);
433 }
434 
435 /* DATAPATH RUNTIME ROUTINES */
436 
437 /*
438  * mac_srs_fire
439  *
440  * Timer callback routine for waking up the SRS worker thread.
441  */
442 static void
443 mac_srs_fire(void *arg)
444 {
445 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
446 
447 	mutex_enter(&mac_srs->srs_lock);
448 	if (mac_srs->srs_tid == 0) {
449 		mutex_exit(&mac_srs->srs_lock);
450 		return;
451 	}
452 
453 	mac_srs->srs_tid = 0;
454 	if (!(mac_srs->srs_state & SRS_PROC))
455 		cv_signal(&mac_srs->srs_async);
456 
457 	mutex_exit(&mac_srs->srs_lock);
458 }
459 
460 /*
461  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
462  * and it is used on the TX path.
463  */
464 #define	HASH_HINT(hint)	\
465 	((hint) ^ ((hint) >> 24) ^ ((hint) >> 16) ^ ((hint) >> 8))
466 
467 
468 /*
469  * hash based on the src address and the port information.
470  */
471 #define	HASH_ADDR(src, ports)					\
472 	(ntohl((src)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
473 	((ports) >> 8) ^ (ports))
474 
475 #define	COMPUTE_INDEX(key, sz)	(key % sz)
476 
477 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
478 	if ((tail) != NULL) {						\
479 		ASSERT((tail)->b_next == NULL);				\
480 		(tail)->b_next = (mp);					\
481 	} else {							\
482 		ASSERT((head) == NULL);					\
483 		(head) = (mp);						\
484 	}								\
485 	(tail) = (mp);							\
486 	(cnt)++;							\
487 	if ((bw_ctl))							\
488 		(sz) += (sz0);						\
489 }
490 
491 #define	MAC_FANOUT_DEFAULT	0
492 #define	MAC_FANOUT_RND_ROBIN	1
493 int mac_fanout_type = MAC_FANOUT_DEFAULT;
494 
495 #define	MAX_SR_TYPES	3
496 /* fanout types for port based hashing */
497 enum pkt_type {
498 	V4_TCP = 0,
499 	V4_UDP,
500 	OTH,
501 	UNDEF
502 };
503 
504 /*
505  * In general we do port based hashing to spread traffic over different
506  * softrings. The below tunable allows to override that behavior. Setting it
507  * to B_TRUE allows to do a fanout based on src ipv6 address. This behavior
508  * is also the applicable to ipv6 packets carrying multiple optional headers
509  * and other uncommon packet types.
510  */
511 boolean_t mac_src_ipv6_fanout = B_FALSE;
512 
513 /*
514  * Pair of local and remote ports in the transport header
515  */
516 #define	PORTS_SIZE 4
517 
518 /*
519  * mac_rx_srs_proto_fanout
520  *
521  * This routine delivers packets destined to an SRS into one of the
522  * protocol soft rings.
523  *
524  * Given a chain of packets we need to split it up into multiple sub chains
525  * destined into TCP, UDP or OTH soft ring. Instead of entering
526  * the soft ring one packet at a time, we want to enter it in the form of a
527  * chain otherwise we get this start/stop behaviour where the worker thread
528  * goes to sleep and then next packets comes in forcing it to wake up etc.
529  */
530 static void
531 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
532 {
533 	struct ether_header		*ehp;
534 	struct ether_vlan_header	*evhp;
535 	uint32_t			sap;
536 	ipha_t				*ipha;
537 	uint8_t				*dstaddr;
538 	size_t				hdrsize;
539 	mblk_t				*mp;
540 	mblk_t				*headmp[MAX_SR_TYPES];
541 	mblk_t				*tailmp[MAX_SR_TYPES];
542 	int				cnt[MAX_SR_TYPES];
543 	size_t				sz[MAX_SR_TYPES];
544 	size_t				sz1;
545 	boolean_t			bw_ctl;
546 	boolean_t			hw_classified;
547 	boolean_t			dls_bypass;
548 	boolean_t			is_ether;
549 	boolean_t			is_unicast;
550 	enum pkt_type			type;
551 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
552 
553 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
554 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
555 
556 	/*
557 	 * If we don't have a Rx ring, S/W classification would have done
558 	 * its job and its a packet meant for us. If we were polling on
559 	 * the default ring (i.e. there was a ring assigned to this SRS),
560 	 * then we need to make sure that the mac address really belongs
561 	 * to us.
562 	 */
563 	hw_classified = mac_srs->srs_ring != NULL &&
564 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
565 
566 	/*
567 	 * Special clients (eg. VLAN, non ether, etc) need DLS
568 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
569 	 * such SRSs. Another way of disabling bypass is to set the
570 	 * MCIS_RX_BYPASS_DISABLE flag.
571 	 */
572 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
573 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
574 
575 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
576 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
577 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
578 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
579 
580 	/*
581 	 * We got a chain from SRS that we need to send to the soft rings.
582 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
583 	 * performance reasons), we need to separate out v4_tcp, v4_udp
584 	 * and the rest goes in other.
585 	 */
586 	while (head != NULL) {
587 		mp = head;
588 		head = head->b_next;
589 		mp->b_next = NULL;
590 
591 		type = OTH;
592 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
593 
594 		if (is_ether) {
595 			/*
596 			 * At this point we can be sure the packet at least
597 			 * has an ether header.
598 			 */
599 			if (sz1 < sizeof (struct ether_header)) {
600 				mac_rx_drop_pkt(mac_srs, mp);
601 				continue;
602 			}
603 			ehp = (struct ether_header *)mp->b_rptr;
604 
605 			/*
606 			 * Determine if this is a VLAN or non-VLAN packet.
607 			 */
608 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
609 				evhp = (struct ether_vlan_header *)mp->b_rptr;
610 				sap = ntohs(evhp->ether_type);
611 				hdrsize = sizeof (struct ether_vlan_header);
612 				/*
613 				 * Check if the VID of the packet, if any,
614 				 * belongs to this client.
615 				 */
616 				if (!mac_client_check_flow_vid(mcip,
617 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
618 					mac_rx_drop_pkt(mac_srs, mp);
619 					continue;
620 				}
621 			} else {
622 				hdrsize = sizeof (struct ether_header);
623 			}
624 			is_unicast =
625 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
626 			dstaddr = (uint8_t *)&ehp->ether_dhost;
627 		} else {
628 			mac_header_info_t		mhi;
629 
630 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
631 			    mp, &mhi) != 0) {
632 				mac_rx_drop_pkt(mac_srs, mp);
633 				continue;
634 			}
635 			hdrsize = mhi.mhi_hdrsize;
636 			sap = mhi.mhi_bindsap;
637 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
638 			dstaddr = (uint8_t *)mhi.mhi_daddr;
639 		}
640 
641 		if (!dls_bypass) {
642 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
643 			    cnt[type], bw_ctl, sz[type], sz1, mp);
644 			continue;
645 		}
646 
647 		if (sap == ETHERTYPE_IP) {
648 			/*
649 			 * If we are H/W classified, but we have promisc
650 			 * on, then we need to check for the unicast address.
651 			 */
652 			if (hw_classified && mcip->mci_promisc_list != NULL) {
653 				mac_address_t		*map;
654 
655 				rw_enter(&mcip->mci_rw_lock, RW_READER);
656 				map = mcip->mci_unicast;
657 				if (bcmp(dstaddr, map->ma_addr,
658 				    map->ma_len) == 0)
659 					type = UNDEF;
660 				rw_exit(&mcip->mci_rw_lock);
661 			} else if (is_unicast) {
662 				type = UNDEF;
663 			}
664 		}
665 
666 		/*
667 		 * This needs to become a contract with the driver for
668 		 * the fast path.
669 		 *
670 		 * In the normal case the packet will have at least the L2
671 		 * header and the IP + Transport header in the same mblk.
672 		 * This is usually the case when the NIC driver sends up
673 		 * the packet. This is also true when the stack generates
674 		 * a packet that is looped back and when the stack uses the
675 		 * fastpath mechanism. The normal case is optimized for
676 		 * performance and may bypass DLS. All other cases go through
677 		 * the 'OTH' type path without DLS bypass.
678 		 */
679 
680 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
681 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
682 			type = OTH;
683 
684 		if (type == OTH) {
685 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
686 			    cnt[type], bw_ctl, sz[type], sz1, mp);
687 			continue;
688 		}
689 
690 		ASSERT(type == UNDEF);
691 		/*
692 		 * We look for at least 4 bytes past the IP header to get
693 		 * the port information. If we get an IP fragment, we don't
694 		 * have the port information, and we use just the protocol
695 		 * information.
696 		 */
697 		switch (ipha->ipha_protocol) {
698 		case IPPROTO_TCP:
699 			type = V4_TCP;
700 			mp->b_rptr += hdrsize;
701 			break;
702 		case IPPROTO_UDP:
703 			type = V4_UDP;
704 			mp->b_rptr += hdrsize;
705 			break;
706 		default:
707 			type = OTH;
708 			break;
709 		}
710 
711 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
712 		    bw_ctl, sz[type], sz1, mp);
713 	}
714 
715 	for (type = V4_TCP; type < UNDEF; type++) {
716 		if (headmp[type] != NULL) {
717 			mac_soft_ring_t			*softring;
718 
719 			ASSERT(tailmp[type]->b_next == NULL);
720 			switch (type) {
721 			case V4_TCP:
722 				softring = mac_srs->srs_tcp_soft_rings[0];
723 				break;
724 			case V4_UDP:
725 				softring = mac_srs->srs_udp_soft_rings[0];
726 				break;
727 			case OTH:
728 				softring = mac_srs->srs_oth_soft_rings[0];
729 			}
730 			mac_rx_soft_ring_process(mcip, softring,
731 			    headmp[type], tailmp[type], cnt[type], sz[type]);
732 		}
733 	}
734 }
735 
736 int	fanout_unalligned = 0;
737 
738 /*
739  * mac_rx_srs_long_fanout
740  *
741  * The fanout routine for IPv6
742  */
743 static int
744 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
745     uint32_t sap, size_t hdrsize, enum pkt_type *type, uint_t *indx)
746 {
747 	ip6_t		*ip6h;
748 	uint8_t		*whereptr;
749 	uint_t		hash;
750 	uint16_t	remlen;
751 	uint8_t		nexthdr;
752 	uint16_t	hdr_len;
753 
754 	if (sap == ETHERTYPE_IPV6) {
755 		boolean_t	modifiable = B_TRUE;
756 
757 		ASSERT(MBLKL(mp) >= hdrsize);
758 
759 		ip6h = (ip6_t *)(mp->b_rptr + hdrsize);
760 		if ((unsigned char *)ip6h == mp->b_wptr) {
761 			/*
762 			 * The first mblk_t only includes the mac header.
763 			 * Note that it is safe to change the mp pointer here,
764 			 * as the subsequent operation does not assume mp
765 			 * points to the start of the mac header.
766 			 */
767 			mp = mp->b_cont;
768 
769 			/*
770 			 * Make sure ip6h holds the full ip6_t structure.
771 			 */
772 			if (mp == NULL)
773 				return (-1);
774 
775 			if (MBLKL(mp) < IPV6_HDR_LEN) {
776 				modifiable = (DB_REF(mp) == 1);
777 
778 				if (modifiable &&
779 				    !pullupmsg(mp, IPV6_HDR_LEN)) {
780 					return (-1);
781 				}
782 			}
783 
784 			ip6h = (ip6_t *)mp->b_rptr;
785 		}
786 
787 		if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
788 		    ((unsigned char *)ip6h + IPV6_HDR_LEN > mp->b_wptr)) {
789 			/*
790 			 * If either ip6h is not alligned, or ip6h does not
791 			 * hold the complete ip6_t structure (a pullupmsg()
792 			 * is not an option since it would result in an
793 			 * unalligned ip6h), fanout to the default ring. Note
794 			 * that this may cause packets reordering.
795 			 */
796 			*indx = 0;
797 			*type = OTH;
798 			fanout_unalligned++;
799 			return (0);
800 		}
801 
802 		remlen = ntohs(ip6h->ip6_plen);
803 		nexthdr = ip6h->ip6_nxt;
804 
805 		if (remlen < MIN_EHDR_LEN)
806 			return (-1);
807 		/*
808 		 * Do src based fanout if below tunable is set to B_TRUE or
809 		 * when mac_ip_hdr_length_v6() fails because of malformed
810 		 * packets or because mblk's need to be concatenated using
811 		 * pullupmsg().
812 		 */
813 		if (mac_src_ipv6_fanout || !mac_ip_hdr_length_v6(ip6h,
814 		    mp->b_wptr, &hdr_len, &nexthdr, NULL)) {
815 			goto src_based_fanout;
816 		}
817 		whereptr = (uint8_t *)ip6h + hdr_len;
818 
819 		/* If the transport is one of below, we do port based fanout */
820 		switch (nexthdr) {
821 		case IPPROTO_TCP:
822 		case IPPROTO_UDP:
823 		case IPPROTO_SCTP:
824 		case IPPROTO_ESP:
825 			/*
826 			 * If the ports in the transport header is not part of
827 			 * the mblk, do src_based_fanout, instead of calling
828 			 * pullupmsg().
829 			 */
830 			if (mp->b_cont != NULL &&
831 			    whereptr + PORTS_SIZE > mp->b_wptr) {
832 				goto src_based_fanout;
833 			}
834 			break;
835 		default:
836 			break;
837 		}
838 
839 		switch (nexthdr) {
840 		case IPPROTO_TCP:
841 			hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
842 			    *(uint32_t *)whereptr);
843 			*indx = COMPUTE_INDEX(hash,
844 			    mac_srs->srs_tcp_ring_count);
845 			*type = OTH;
846 			break;
847 
848 		case IPPROTO_UDP:
849 		case IPPROTO_SCTP:
850 		case IPPROTO_ESP:
851 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
852 				hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
853 				    *(uint32_t *)whereptr);
854 				*indx = COMPUTE_INDEX(hash,
855 				    mac_srs->srs_udp_ring_count);
856 			} else {
857 				*indx = mac_srs->srs_ind %
858 				    mac_srs->srs_udp_ring_count;
859 				mac_srs->srs_ind++;
860 			}
861 			*type = OTH;
862 			break;
863 
864 			/* For all other protocol, do source based fanout */
865 		default:
866 			goto src_based_fanout;
867 		}
868 	} else {
869 		*indx = 0;
870 		*type = OTH;
871 	}
872 	return (0);
873 
874 src_based_fanout:
875 	hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src), (uint32_t)0);
876 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
877 	*type = OTH;
878 	return (0);
879 }
880 
881 /*
882  * mac_rx_srs_fanout
883  *
884  * This routine delivers packets destined to an SRS into a soft ring member
885  * of the set.
886  *
887  * Given a chain of packets we need to split it up into multiple sub chains
888  * destined for one of the TCP, UDP or OTH soft rings. Instead of entering
889  * the soft ring one packet at a time, we want to enter it in the form of a
890  * chain otherwise we get this start/stop behaviour where the worker thread
891  * goes to sleep and then next packets comes in forcing it to wake up etc.
892  *
893  * Note:
894  * Since we know what is the maximum fanout possible, we create a 2D array
895  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
896  * variables so that we can enter the softrings with chain. We need the
897  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
898  * for each packet would be expensive). If we ever want to have the
899  * ability to have unlimited fanout, we should probably declare a head,
900  * tail, cnt, sz with each soft ring (a data struct which contains a softring
901  * along with these members) and create an array of this uber struct so we
902  * don't have to do kmem_alloc.
903  */
904 int	fanout_oth1 = 0;
905 int	fanout_oth2 = 0;
906 int	fanout_oth3 = 0;
907 int	fanout_oth4 = 0;
908 int	fanout_oth5 = 0;
909 
910 static void
911 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
912 {
913 	struct ether_header		*ehp;
914 	struct ether_vlan_header	*evhp;
915 	uint32_t			sap;
916 	ipha_t				*ipha;
917 	uint8_t				*dstaddr;
918 	uint_t				indx;
919 	size_t				ports_offset;
920 	size_t				ipha_len;
921 	size_t				hdrsize;
922 	uint_t				hash;
923 	mblk_t				*mp;
924 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
925 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
926 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
927 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
928 	size_t				sz1;
929 	boolean_t			bw_ctl;
930 	boolean_t			hw_classified;
931 	boolean_t			dls_bypass;
932 	boolean_t			is_ether;
933 	boolean_t			is_unicast;
934 	int				fanout_cnt;
935 	enum pkt_type			type;
936 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
937 
938 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
939 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
940 
941 	/*
942 	 * If we don't have a Rx ring, S/W classification would have done
943 	 * its job and its a packet meant for us. If we were polling on
944 	 * the default ring (i.e. there was a ring assigned to this SRS),
945 	 * then we need to make sure that the mac address really belongs
946 	 * to us.
947 	 */
948 	hw_classified = mac_srs->srs_ring != NULL &&
949 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
950 
951 	/*
952 	 * Special clients (eg. VLAN, non ether, etc) need DLS
953 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
954 	 * such SRSs. Another way of disabling bypass is to set the
955 	 * MCIS_RX_BYPASS_DISABLE flag.
956 	 */
957 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
958 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
959 
960 	/*
961 	 * Since the softrings are never destroyed and we always
962 	 * create equal number of softrings for TCP, UDP and rest,
963 	 * its OK to check one of them for count and use it without
964 	 * any lock. In future, if soft rings get destroyed because
965 	 * of reduction in fanout, we will need to ensure that happens
966 	 * behind the SRS_PROC.
967 	 */
968 	fanout_cnt = mac_srs->srs_tcp_ring_count;
969 
970 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
971 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
972 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
973 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
974 
975 	/*
976 	 * We got a chain from SRS that we need to send to the soft rings.
977 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
978 	 * performance reasons), we need to separate out v4_tcp, v4_udp
979 	 * and the rest goes in other.
980 	 */
981 	while (head != NULL) {
982 		mp = head;
983 		head = head->b_next;
984 		mp->b_next = NULL;
985 
986 		type = OTH;
987 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
988 
989 		if (is_ether) {
990 			/*
991 			 * At this point we can be sure the packet at least
992 			 * has an ether header.
993 			 */
994 			if (sz1 < sizeof (struct ether_header)) {
995 				mac_rx_drop_pkt(mac_srs, mp);
996 				continue;
997 			}
998 			ehp = (struct ether_header *)mp->b_rptr;
999 
1000 			/*
1001 			 * Determine if this is a VLAN or non-VLAN packet.
1002 			 */
1003 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1004 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1005 				sap = ntohs(evhp->ether_type);
1006 				hdrsize = sizeof (struct ether_vlan_header);
1007 				/*
1008 				 * Check if the VID of the packet, if any,
1009 				 * belongs to this client.
1010 				 */
1011 				if (!mac_client_check_flow_vid(mcip,
1012 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1013 					mac_rx_drop_pkt(mac_srs, mp);
1014 					continue;
1015 				}
1016 			} else {
1017 				hdrsize = sizeof (struct ether_header);
1018 			}
1019 			is_unicast =
1020 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
1021 			dstaddr = (uint8_t *)&ehp->ether_dhost;
1022 		} else {
1023 			mac_header_info_t		mhi;
1024 
1025 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
1026 			    mp, &mhi) != 0) {
1027 				mac_rx_drop_pkt(mac_srs, mp);
1028 				continue;
1029 			}
1030 			hdrsize = mhi.mhi_hdrsize;
1031 			sap = mhi.mhi_bindsap;
1032 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
1033 			dstaddr = (uint8_t *)mhi.mhi_daddr;
1034 		}
1035 
1036 		if (!dls_bypass) {
1037 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1038 			    hdrsize, &type, &indx) == -1) {
1039 				mac_rx_drop_pkt(mac_srs, mp);
1040 				continue;
1041 			}
1042 
1043 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1044 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1045 			    sz[type][indx], sz1, mp);
1046 			continue;
1047 		}
1048 
1049 
1050 		/*
1051 		 * If we are using the default Rx ring where H/W or S/W
1052 		 * classification has not happened, we need to verify if
1053 		 * this unicast packet really belongs to us.
1054 		 */
1055 		if (sap == ETHERTYPE_IP) {
1056 			/*
1057 			 * If we are H/W classified, but we have promisc
1058 			 * on, then we need to check for the unicast address.
1059 			 */
1060 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1061 				mac_address_t		*map;
1062 
1063 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1064 				map = mcip->mci_unicast;
1065 				if (bcmp(dstaddr, map->ma_addr,
1066 				    map->ma_len) == 0)
1067 					type = UNDEF;
1068 				rw_exit(&mcip->mci_rw_lock);
1069 			} else if (is_unicast) {
1070 				type = UNDEF;
1071 			}
1072 		}
1073 
1074 		/*
1075 		 * This needs to become a contract with the driver for
1076 		 * the fast path.
1077 		 */
1078 
1079 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
1080 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
1081 			type = OTH;
1082 			fanout_oth1++;
1083 		}
1084 
1085 		if (type != OTH) {
1086 			uint16_t	frag_offset_flags;
1087 
1088 			switch (ipha->ipha_protocol) {
1089 			case IPPROTO_TCP:
1090 			case IPPROTO_UDP:
1091 			case IPPROTO_SCTP:
1092 			case IPPROTO_ESP:
1093 				ipha_len = IPH_HDR_LENGTH(ipha);
1094 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
1095 				    mp->b_wptr) {
1096 					type = OTH;
1097 					break;
1098 				}
1099 				frag_offset_flags =
1100 				    ntohs(ipha->ipha_fragment_offset_and_flags);
1101 				if ((frag_offset_flags &
1102 				    (IPH_MF | IPH_OFFSET)) != 0) {
1103 					type = OTH;
1104 					fanout_oth3++;
1105 					break;
1106 				}
1107 				ports_offset = hdrsize + ipha_len;
1108 				break;
1109 			default:
1110 				type = OTH;
1111 				fanout_oth4++;
1112 				break;
1113 			}
1114 		}
1115 
1116 		if (type == OTH) {
1117 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1118 			    hdrsize, &type, &indx) == -1) {
1119 				mac_rx_drop_pkt(mac_srs, mp);
1120 				continue;
1121 			}
1122 
1123 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1124 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1125 			    sz[type][indx], sz1, mp);
1126 			continue;
1127 		}
1128 
1129 		ASSERT(type == UNDEF);
1130 
1131 		/*
1132 		 * XXX-Sunay: We should hold srs_lock since ring_count
1133 		 * below can change. But if we are always called from
1134 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
1135 		 * enforce that ring_count can't be changed i.e.
1136 		 * to change fanout type or ring count, the calling
1137 		 * thread needs to be behind SRS_PROC.
1138 		 */
1139 		switch (ipha->ipha_protocol) {
1140 		case IPPROTO_TCP:
1141 			/*
1142 			 * Note that for ESP, we fanout on SPI and it is at the
1143 			 * same offset as the 2x16-bit ports. So it is clumped
1144 			 * along with TCP, UDP and SCTP.
1145 			 */
1146 			hash = HASH_ADDR(ipha->ipha_src,
1147 			    *(uint32_t *)(mp->b_rptr + ports_offset));
1148 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1149 			type = V4_TCP;
1150 			mp->b_rptr += hdrsize;
1151 			break;
1152 		case IPPROTO_UDP:
1153 		case IPPROTO_SCTP:
1154 		case IPPROTO_ESP:
1155 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1156 				hash = HASH_ADDR(ipha->ipha_src,
1157 				    *(uint32_t *)(mp->b_rptr + ports_offset));
1158 				indx = COMPUTE_INDEX(hash,
1159 				    mac_srs->srs_udp_ring_count);
1160 			} else {
1161 				indx = mac_srs->srs_ind %
1162 				    mac_srs->srs_udp_ring_count;
1163 				mac_srs->srs_ind++;
1164 			}
1165 			type = V4_UDP;
1166 			mp->b_rptr += hdrsize;
1167 			break;
1168 		default:
1169 			indx = 0;
1170 			type = OTH;
1171 		}
1172 
1173 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
1174 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
1175 	}
1176 
1177 	for (type = V4_TCP; type < UNDEF; type++) {
1178 		int	i;
1179 
1180 		for (i = 0; i < fanout_cnt; i++) {
1181 			if (headmp[type][i] != NULL) {
1182 				mac_soft_ring_t	*softring;
1183 
1184 				ASSERT(tailmp[type][i]->b_next == NULL);
1185 				switch (type) {
1186 				case V4_TCP:
1187 					softring =
1188 					    mac_srs->srs_tcp_soft_rings[i];
1189 					break;
1190 				case V4_UDP:
1191 					softring =
1192 					    mac_srs->srs_udp_soft_rings[i];
1193 					break;
1194 				case OTH:
1195 					softring =
1196 					    mac_srs->srs_oth_soft_rings[i];
1197 					break;
1198 				}
1199 				mac_rx_soft_ring_process(mcip,
1200 				    softring, headmp[type][i], tailmp[type][i],
1201 				    cnt[type][i], sz[type][i]);
1202 			}
1203 		}
1204 	}
1205 }
1206 
1207 #define	SRS_BYTES_TO_PICKUP	150000
1208 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
1209 
1210 /*
1211  * mac_rx_srs_poll_ring
1212  *
1213  * This SRS Poll thread uses this routine to poll the underlying hardware
1214  * Rx ring to get a chain of packets. It can inline process that chain
1215  * if mac_latency_optimize is set (default) or signal the SRS worker thread
1216  * to do the remaining processing.
1217  *
1218  * Since packets come in the system via interrupt or poll path, we also
1219  * update the stats and deal with promiscous clients here.
1220  */
1221 void
1222 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
1223 {
1224 	kmutex_t 		*lock = &mac_srs->srs_lock;
1225 	kcondvar_t 		*async = &mac_srs->srs_cv;
1226 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1227 	mblk_t 			*head, *tail, *mp;
1228 	callb_cpr_t 		cprinfo;
1229 	ssize_t 		bytes_to_pickup;
1230 	size_t 			sz;
1231 	int			count;
1232 	mac_client_impl_t	*smcip;
1233 
1234 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
1235 	mutex_enter(lock);
1236 
1237 start:
1238 	for (;;) {
1239 		if (mac_srs->srs_state & SRS_PAUSE)
1240 			goto done;
1241 
1242 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
1243 		cv_wait(async, lock);
1244 		CALLB_CPR_SAFE_END(&cprinfo, lock);
1245 
1246 		if (mac_srs->srs_state & SRS_PAUSE)
1247 			goto done;
1248 
1249 check_again:
1250 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
1251 			/*
1252 			 * We pick as many bytes as we are allowed to queue.
1253 			 * Its possible that we will exceed the total
1254 			 * packets queued in case this SRS is part of the
1255 			 * Rx ring group since > 1 poll thread can be pulling
1256 			 * upto the max allowed packets at the same time
1257 			 * but that should be OK.
1258 			 */
1259 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1260 			bytes_to_pickup =
1261 			    mac_srs->srs_bw->mac_bw_drop_threshold -
1262 			    mac_srs->srs_bw->mac_bw_sz;
1263 			/*
1264 			 * We shouldn't have been signalled if we
1265 			 * have 0 or less bytes to pick but since
1266 			 * some of the bytes accounting is driver
1267 			 * dependant, we do the safety check.
1268 			 */
1269 			if (bytes_to_pickup < 0)
1270 				bytes_to_pickup = 0;
1271 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1272 		} else {
1273 			/*
1274 			 * ToDO: Need to change the polling API
1275 			 * to add a packet count and a flag which
1276 			 * tells the driver whether we want packets
1277 			 * based on a count, or bytes, or all the
1278 			 * packets queued in the driver/HW. This
1279 			 * way, we never have to check the limits
1280 			 * on poll path. We truly let only as many
1281 			 * packets enter the system as we are willing
1282 			 * to process or queue.
1283 			 *
1284 			 * Something along the lines of
1285 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
1286 			 *	mac_srs->srs_poll_pkt_cnt
1287 			 */
1288 
1289 			/*
1290 			 * Since we are not doing B/W control, pick
1291 			 * as many packets as allowed.
1292 			 */
1293 			bytes_to_pickup = max_bytes_to_pickup;
1294 		}
1295 
1296 		/* Poll the underlying Hardware */
1297 		mutex_exit(lock);
1298 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
1299 		mutex_enter(lock);
1300 
1301 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
1302 		    SRS_POLL_THR_OWNER);
1303 
1304 		mp = tail = head;
1305 		count = 0;
1306 		sz = 0;
1307 		while (mp != NULL) {
1308 			tail = mp;
1309 			sz += msgdsize(mp);
1310 			mp = mp->b_next;
1311 			count++;
1312 		}
1313 
1314 		if (head != NULL) {
1315 			tail->b_next = NULL;
1316 			smcip = mac_srs->srs_mcip;
1317 
1318 			SRS_RX_STAT_UPDATE(mac_srs, pollbytes, sz);
1319 			SRS_RX_STAT_UPDATE(mac_srs, pollcnt, count);
1320 
1321 			/*
1322 			 * If there are any promiscuous mode callbacks
1323 			 * defined for this MAC client, pass them a copy
1324 			 * if appropriate and also update the counters.
1325 			 */
1326 			if (smcip != NULL) {
1327 				if (smcip->mci_mip->mi_promisc_list != NULL) {
1328 					mutex_exit(lock);
1329 					mac_promisc_dispatch(smcip->mci_mip,
1330 					    head, NULL);
1331 					mutex_enter(lock);
1332 				}
1333 			}
1334 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
1335 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1336 				mac_srs->srs_bw->mac_bw_polled += sz;
1337 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1338 			}
1339 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
1340 			    count, sz);
1341 			if (count <= 10)
1342 				srs_rx->sr_stat.mrs_chaincntundr10++;
1343 			else if (count > 10 && count <= 50)
1344 				srs_rx->sr_stat.mrs_chaincnt10to50++;
1345 			else
1346 				srs_rx->sr_stat.mrs_chaincntover50++;
1347 		}
1348 
1349 		/*
1350 		 * We are guaranteed that SRS_PROC will be set if we
1351 		 * are here. Also, poll thread gets to run only if
1352 		 * the drain was being done by a worker thread although
1353 		 * its possible that worker thread is still running
1354 		 * and poll thread was sent down to keep the pipeline
1355 		 * going instead of doing a complete drain and then
1356 		 * trying to poll the NIC.
1357 		 *
1358 		 * So we need to check SRS_WORKER flag to make sure
1359 		 * that the worker thread is not processing the queue
1360 		 * in parallel to us. The flags and conditions are
1361 		 * protected by the srs_lock to prevent any race. We
1362 		 * ensure that we don't drop the srs_lock from now
1363 		 * till the end and similarly we don't drop the srs_lock
1364 		 * in mac_rx_srs_drain() till similar condition check
1365 		 * are complete. The mac_rx_srs_drain() needs to ensure
1366 		 * that SRS_WORKER flag remains set as long as its
1367 		 * processing the queue.
1368 		 */
1369 		if (!(mac_srs->srs_state & SRS_WORKER) &&
1370 		    (mac_srs->srs_first != NULL)) {
1371 			/*
1372 			 * We have packets to process and worker thread
1373 			 * is not running. Check to see if poll thread is
1374 			 * allowed to process.
1375 			 */
1376 			if (mac_srs->srs_state & SRS_LATENCY_OPT) {
1377 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
1378 				if (!(mac_srs->srs_state & SRS_PAUSE) &&
1379 				    srs_rx->sr_poll_pkt_cnt <=
1380 				    srs_rx->sr_lowat) {
1381 					srs_rx->sr_poll_again++;
1382 					goto check_again;
1383 				}
1384 				/*
1385 				 * We are already above low water mark
1386 				 * so stay in the polling mode but no
1387 				 * need to poll. Once we dip below
1388 				 * the polling threshold, the processing
1389 				 * thread (soft ring) will signal us
1390 				 * to poll again (MAC_UPDATE_SRS_COUNT)
1391 				 */
1392 				srs_rx->sr_poll_drain_no_poll++;
1393 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1394 				/*
1395 				 * In B/W control case, its possible
1396 				 * that the backlog built up due to
1397 				 * B/W limit being reached and packets
1398 				 * are queued only in SRS. In this case,
1399 				 * we should schedule worker thread
1400 				 * since no one else will wake us up.
1401 				 */
1402 				if ((mac_srs->srs_type & SRST_BW_CONTROL) &&
1403 				    (mac_srs->srs_tid == NULL)) {
1404 					mac_srs->srs_tid =
1405 					    timeout(mac_srs_fire, mac_srs, 1);
1406 					srs_rx->sr_poll_worker_wakeup++;
1407 				}
1408 			} else {
1409 				/*
1410 				 * Wakeup the worker thread for more processing.
1411 				 * We optimize for throughput in this case.
1412 				 */
1413 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1414 				MAC_SRS_WORKER_WAKEUP(mac_srs);
1415 				srs_rx->sr_poll_sig_worker++;
1416 			}
1417 		} else if ((mac_srs->srs_first == NULL) &&
1418 		    !(mac_srs->srs_state & SRS_WORKER)) {
1419 			/*
1420 			 * There is nothing queued in SRS and
1421 			 * no worker thread running. Plus we
1422 			 * didn't get anything from the H/W
1423 			 * as well (head == NULL);
1424 			 */
1425 			ASSERT(head == NULL);
1426 			mac_srs->srs_state &=
1427 			    ~(SRS_PROC|SRS_GET_PKTS);
1428 
1429 			/*
1430 			 * If we have a packets in soft ring, don't allow
1431 			 * more packets to come into this SRS by keeping the
1432 			 * interrupts off but not polling the H/W. The
1433 			 * poll thread will get signaled as soon as
1434 			 * srs_poll_pkt_cnt dips below poll threshold.
1435 			 */
1436 			if (srs_rx->sr_poll_pkt_cnt == 0) {
1437 				srs_rx->sr_poll_intr_enable++;
1438 				MAC_SRS_POLLING_OFF(mac_srs);
1439 			} else {
1440 				/*
1441 				 * We know nothing is queued in SRS
1442 				 * since we are here after checking
1443 				 * srs_first is NULL. The backlog
1444 				 * is entirely due to packets queued
1445 				 * in Soft ring which will wake us up
1446 				 * and get the interface out of polling
1447 				 * mode once the backlog dips below
1448 				 * sr_poll_thres.
1449 				 */
1450 				srs_rx->sr_poll_no_poll++;
1451 			}
1452 		} else {
1453 			/*
1454 			 * Worker thread is already running.
1455 			 * Nothing much to do. If the polling
1456 			 * was enabled, worker thread will deal
1457 			 * with that.
1458 			 */
1459 			mac_srs->srs_state &= ~SRS_GET_PKTS;
1460 			srs_rx->sr_poll_goto_sleep++;
1461 		}
1462 	}
1463 done:
1464 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
1465 	cv_signal(&mac_srs->srs_async);
1466 	/*
1467 	 * If this is a temporary quiesce then wait for the restart signal
1468 	 * from the srs worker. Then clear the flags and signal the srs worker
1469 	 * to ensure a positive handshake and go back to start.
1470 	 */
1471 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
1472 		cv_wait(async, lock);
1473 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
1474 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
1475 		mac_srs->srs_state &=
1476 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
1477 		cv_signal(&mac_srs->srs_async);
1478 		goto start;
1479 	} else {
1480 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
1481 		cv_signal(&mac_srs->srs_async);
1482 		CALLB_CPR_EXIT(&cprinfo);
1483 		thread_exit();
1484 	}
1485 }
1486 
1487 /*
1488  * mac_srs_pick_chain
1489  *
1490  * In Bandwidth control case, checks how many packets can be processed
1491  * and return them in a sub chain.
1492  */
1493 static mblk_t *
1494 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
1495     size_t *chain_sz, int *chain_cnt)
1496 {
1497 	mblk_t 			*head = NULL;
1498 	mblk_t 			*tail = NULL;
1499 	size_t			sz;
1500 	size_t 			tsz = 0;
1501 	int			cnt = 0;
1502 	mblk_t 			*mp;
1503 
1504 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1505 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1506 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
1507 	    mac_srs->srs_bw->mac_bw_limit) ||
1508 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
1509 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1510 		head = mac_srs->srs_first;
1511 		mac_srs->srs_first = NULL;
1512 		*chain_tail = mac_srs->srs_last;
1513 		mac_srs->srs_last = NULL;
1514 		*chain_sz = mac_srs->srs_size;
1515 		*chain_cnt = mac_srs->srs_count;
1516 		mac_srs->srs_count = 0;
1517 		mac_srs->srs_size = 0;
1518 		return (head);
1519 	}
1520 
1521 	/*
1522 	 * Can't clear the entire backlog.
1523 	 * Need to find how many packets to pick
1524 	 */
1525 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
1526 	while ((mp = mac_srs->srs_first) != NULL) {
1527 		sz = msgdsize(mp);
1528 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
1529 		    mac_srs->srs_bw->mac_bw_limit) {
1530 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
1531 				mac_srs->srs_bw->mac_bw_state |=
1532 				    SRS_BW_ENFORCED;
1533 			break;
1534 		}
1535 
1536 		/*
1537 		 * The _size & cnt is  decremented from the softrings
1538 		 * when they send up the packet for polling to work
1539 		 * properly.
1540 		 */
1541 		tsz += sz;
1542 		cnt++;
1543 		mac_srs->srs_count--;
1544 		mac_srs->srs_size -= sz;
1545 		if (tail != NULL)
1546 			tail->b_next = mp;
1547 		else
1548 			head = mp;
1549 		tail = mp;
1550 		mac_srs->srs_first = mac_srs->srs_first->b_next;
1551 	}
1552 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1553 	if (mac_srs->srs_first == NULL)
1554 		mac_srs->srs_last = NULL;
1555 
1556 	if (tail != NULL)
1557 		tail->b_next = NULL;
1558 	*chain_tail = tail;
1559 	*chain_cnt = cnt;
1560 	*chain_sz = tsz;
1561 
1562 	return (head);
1563 }
1564 
1565 /*
1566  * mac_rx_srs_drain
1567  *
1568  * The SRS drain routine. Gets to run to clear the queue. Any thread
1569  * (worker, interrupt, poll) can call this based on processing model.
1570  * The first thing we do is disable interrupts if possible and then
1571  * drain the queue. we also try to poll the underlying hardware if
1572  * there is a dedicated hardware Rx ring assigned to this SRS.
1573  *
1574  * There is a equivalent drain routine in bandwidth control mode
1575  * mac_rx_srs_drain_bw. There is some code duplication between the two
1576  * routines but they are highly performance sensitive and are easier
1577  * to read/debug if they stay separate. Any code changes here might
1578  * also apply to mac_rx_srs_drain_bw as well.
1579  */
1580 void
1581 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1582 {
1583 	mblk_t 			*head;
1584 	mblk_t			*tail;
1585 	timeout_id_t 		tid;
1586 	int			cnt = 0;
1587 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1588 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1589 
1590 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1591 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
1592 
1593 	/* If we are blanked i.e. can't do upcalls, then we are done */
1594 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1595 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1596 		    (mac_srs->srs_state & SRS_PAUSE));
1597 		goto out;
1598 	}
1599 
1600 	if (mac_srs->srs_first == NULL)
1601 		goto out;
1602 
1603 	if (!(mac_srs->srs_state & SRS_LATENCY_OPT) &&
1604 	    (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)) {
1605 		/*
1606 		 * In the normal case, the SRS worker thread does no
1607 		 * work and we wait for a backlog to build up before
1608 		 * we switch into polling mode. In case we are
1609 		 * optimizing for throughput, we use the worker thread
1610 		 * as well. The goal is to let worker thread process
1611 		 * the queue and poll thread to feed packets into
1612 		 * the queue. As such, we should signal the poll
1613 		 * thread to try and get more packets.
1614 		 *
1615 		 * We could have pulled this check in the POLL_RING
1616 		 * macro itself but keeping it explicit here makes
1617 		 * the architecture more human understandable.
1618 		 */
1619 		MAC_SRS_POLL_RING(mac_srs);
1620 	}
1621 
1622 again:
1623 	head = mac_srs->srs_first;
1624 	mac_srs->srs_first = NULL;
1625 	tail = mac_srs->srs_last;
1626 	mac_srs->srs_last = NULL;
1627 	cnt = mac_srs->srs_count;
1628 	mac_srs->srs_count = 0;
1629 
1630 	ASSERT(head != NULL);
1631 	ASSERT(tail != NULL);
1632 
1633 	if ((tid = mac_srs->srs_tid) != 0)
1634 		mac_srs->srs_tid = 0;
1635 
1636 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1637 
1638 
1639 	/*
1640 	 * mcip is NULL for broadcast and multicast flows. The promisc
1641 	 * callbacks for broadcast and multicast packets are delivered from
1642 	 * mac_rx() and we don't need to worry about that case in this path
1643 	 */
1644 	if (mcip != NULL) {
1645 		if (mcip->mci_promisc_list != NULL) {
1646 			mutex_exit(&mac_srs->srs_lock);
1647 			mac_promisc_client_dispatch(mcip, head);
1648 			mutex_enter(&mac_srs->srs_lock);
1649 		}
1650 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
1651 			mutex_exit(&mac_srs->srs_lock);
1652 			mac_protect_intercept_dhcp(mcip, head);
1653 			mutex_enter(&mac_srs->srs_lock);
1654 		}
1655 	}
1656 
1657 	/*
1658 	 * Check if SRS itself is doing the processing
1659 	 * This direct path does not apply when subflows are present. In this
1660 	 * case, packets need to be dispatched to a soft ring according to the
1661 	 * flow's bandwidth and other resources contraints.
1662 	 */
1663 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1664 		mac_direct_rx_t		proc;
1665 		void			*arg1;
1666 		mac_resource_handle_t	arg2;
1667 
1668 		/*
1669 		 * This is the case when a Rx is directly
1670 		 * assigned and we have a fully classified
1671 		 * protocol chain. We can deal with it in
1672 		 * one shot.
1673 		 */
1674 		proc = srs_rx->sr_func;
1675 		arg1 = srs_rx->sr_arg1;
1676 		arg2 = srs_rx->sr_arg2;
1677 
1678 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1679 		mutex_exit(&mac_srs->srs_lock);
1680 		if (tid != 0) {
1681 			(void) untimeout(tid);
1682 			tid = 0;
1683 		}
1684 
1685 		proc(arg1, arg2, head, NULL);
1686 		/*
1687 		 * Decrement the size and count here itelf
1688 		 * since the packet has been processed.
1689 		 */
1690 		mutex_enter(&mac_srs->srs_lock);
1691 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1692 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1693 			cv_signal(&mac_srs->srs_client_cv);
1694 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1695 	} else {
1696 		/* Some kind of softrings based fanout is required */
1697 		mutex_exit(&mac_srs->srs_lock);
1698 		if (tid != 0) {
1699 			(void) untimeout(tid);
1700 			tid = 0;
1701 		}
1702 
1703 		/*
1704 		 * Since the fanout routines can deal with chains,
1705 		 * shoot the entire chain up.
1706 		 */
1707 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1708 			mac_rx_srs_fanout(mac_srs, head);
1709 		else
1710 			mac_rx_srs_proto_fanout(mac_srs, head);
1711 		mutex_enter(&mac_srs->srs_lock);
1712 	}
1713 
1714 	if (!(mac_srs->srs_state & (SRS_BLANK|SRS_PAUSE)) &&
1715 	    (mac_srs->srs_first != NULL)) {
1716 		/*
1717 		 * More packets arrived while we were clearing the
1718 		 * SRS. This can be possible because of one of
1719 		 * three conditions below:
1720 		 * 1) The driver is using multiple worker threads
1721 		 *    to send the packets to us.
1722 		 * 2) The driver has a race in switching
1723 		 *    between interrupt and polling mode or
1724 		 * 3) Packets are arriving in this SRS via the
1725 		 *    S/W classification as well.
1726 		 *
1727 		 * We should switch to polling mode and see if we
1728 		 * need to send the poll thread down. Also, signal
1729 		 * the worker thread to process whats just arrived.
1730 		 */
1731 		MAC_SRS_POLLING_ON(mac_srs);
1732 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat) {
1733 			srs_rx->sr_drain_poll_sig++;
1734 			MAC_SRS_POLL_RING(mac_srs);
1735 		}
1736 
1737 		/*
1738 		 * If we didn't signal the poll thread, we need
1739 		 * to deal with the pending packets ourselves.
1740 		 */
1741 		if (proc_type == SRS_WORKER) {
1742 			srs_rx->sr_drain_again++;
1743 			goto again;
1744 		} else {
1745 			srs_rx->sr_drain_worker_sig++;
1746 			cv_signal(&mac_srs->srs_async);
1747 		}
1748 	}
1749 
1750 out:
1751 	if (mac_srs->srs_state & SRS_GET_PKTS) {
1752 		/*
1753 		 * Poll thread is already running. Leave the
1754 		 * SRS_RPOC set and hand over the control to
1755 		 * poll thread.
1756 		 */
1757 		mac_srs->srs_state &= ~proc_type;
1758 		srs_rx->sr_drain_poll_running++;
1759 		return;
1760 	}
1761 
1762 	/*
1763 	 * Even if there are no packets queued in SRS, we
1764 	 * need to make sure that the shared counter is
1765 	 * clear and any associated softrings have cleared
1766 	 * all the backlog. Otherwise, leave the interface
1767 	 * in polling mode and the poll thread will get
1768 	 * signalled once the count goes down to zero.
1769 	 *
1770 	 * If someone is already draining the queue (SRS_PROC is
1771 	 * set) when the srs_poll_pkt_cnt goes down to zero,
1772 	 * then it means that drain is already running and we
1773 	 * will turn off polling at that time if there is
1774 	 * no backlog.
1775 	 *
1776 	 * As long as there are packets queued either
1777 	 * in soft ring set or its soft rings, we will leave
1778 	 * the interface in polling mode (even if the drain
1779 	 * was done being the interrupt thread). We signal
1780 	 * the poll thread as well if we have dipped below
1781 	 * low water mark.
1782 	 *
1783 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
1784 	 * since that turn polling on only for worker thread.
1785 	 * Its not worth turning polling on for interrupt
1786 	 * thread (since NIC will not issue another interrupt)
1787 	 * unless a backlog builds up.
1788 	 */
1789 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
1790 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
1791 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1792 		srs_rx->sr_drain_keep_polling++;
1793 		MAC_SRS_POLLING_ON(mac_srs);
1794 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1795 			MAC_SRS_POLL_RING(mac_srs);
1796 		return;
1797 	}
1798 
1799 	/* Nothing else to do. Get out of poll mode */
1800 	MAC_SRS_POLLING_OFF(mac_srs);
1801 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1802 	srs_rx->sr_drain_finish_intr++;
1803 }
1804 
1805 /*
1806  * mac_rx_srs_drain_bw
1807  *
1808  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
1809  * (worker, interrupt, poll) can call this based on processing model.
1810  * The first thing we do is disable interrupts if possible and then
1811  * drain the queue. we also try to poll the underlying hardware if
1812  * there is a dedicated hardware Rx ring assigned to this SRS.
1813  *
1814  * There is a equivalent drain routine in non bandwidth control mode
1815  * mac_rx_srs_drain. There is some code duplication between the two
1816  * routines but they are highly performance sensitive and are easier
1817  * to read/debug if they stay separate. Any code changes here might
1818  * also apply to mac_rx_srs_drain as well.
1819  */
1820 void
1821 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1822 {
1823 	mblk_t 			*head;
1824 	mblk_t			*tail;
1825 	timeout_id_t 		tid;
1826 	size_t			sz = 0;
1827 	int			cnt = 0;
1828 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1829 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1830 	clock_t			now;
1831 
1832 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1833 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
1834 again:
1835 	/* Check if we are doing B/W control */
1836 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1837 	now = ddi_get_lbolt();
1838 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
1839 		mac_srs->srs_bw->mac_bw_curr_time = now;
1840 		mac_srs->srs_bw->mac_bw_used = 0;
1841 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
1842 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
1843 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
1844 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1845 		goto done;
1846 	} else if (mac_srs->srs_bw->mac_bw_used >
1847 	    mac_srs->srs_bw->mac_bw_limit) {
1848 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
1849 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1850 		goto done;
1851 	}
1852 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1853 
1854 	/* If we are blanked i.e. can't do upcalls, then we are done */
1855 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1856 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1857 		    (mac_srs->srs_state & SRS_PAUSE));
1858 		goto done;
1859 	}
1860 
1861 	sz = 0;
1862 	cnt = 0;
1863 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
1864 		/*
1865 		 * We couldn't pick up a single packet.
1866 		 */
1867 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1868 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
1869 		    (mac_srs->srs_size != 0) &&
1870 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1871 			/*
1872 			 * Seems like configured B/W doesn't
1873 			 * even allow processing of 1 packet
1874 			 * per tick.
1875 			 *
1876 			 * XXX: raise the limit to processing
1877 			 * at least 1 packet per tick.
1878 			 */
1879 			mac_srs->srs_bw->mac_bw_limit +=
1880 			    mac_srs->srs_bw->mac_bw_limit;
1881 			mac_srs->srs_bw->mac_bw_drop_threshold +=
1882 			    mac_srs->srs_bw->mac_bw_drop_threshold;
1883 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
1884 			    "raised B/W limit to %d since not even a "
1885 			    "single packet can be processed per "
1886 			    "tick %d\n", (void *)mac_srs,
1887 			    (int)mac_srs->srs_bw->mac_bw_limit,
1888 			    (int)msgdsize(mac_srs->srs_first));
1889 		}
1890 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1891 		goto done;
1892 	}
1893 
1894 	ASSERT(head != NULL);
1895 	ASSERT(tail != NULL);
1896 
1897 	/* zero bandwidth: drop all and return to interrupt mode */
1898 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1899 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
1900 		srs_rx->sr_stat.mrs_sdrops += cnt;
1901 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
1902 		mac_srs->srs_bw->mac_bw_sz -= sz;
1903 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
1904 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1905 		mac_pkt_drop(NULL, NULL, head, B_FALSE);
1906 		goto leave_poll;
1907 	} else {
1908 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1909 	}
1910 
1911 	if ((tid = mac_srs->srs_tid) != 0)
1912 		mac_srs->srs_tid = 0;
1913 
1914 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1915 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
1916 
1917 	/*
1918 	 * mcip is NULL for broadcast and multicast flows. The promisc
1919 	 * callbacks for broadcast and multicast packets are delivered from
1920 	 * mac_rx() and we don't need to worry about that case in this path
1921 	 */
1922 	if (mcip != NULL) {
1923 		if (mcip->mci_promisc_list != NULL) {
1924 			mutex_exit(&mac_srs->srs_lock);
1925 			mac_promisc_client_dispatch(mcip, head);
1926 			mutex_enter(&mac_srs->srs_lock);
1927 		}
1928 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
1929 			mutex_exit(&mac_srs->srs_lock);
1930 			mac_protect_intercept_dhcp(mcip, head);
1931 			mutex_enter(&mac_srs->srs_lock);
1932 		}
1933 	}
1934 
1935 	/*
1936 	 * Check if SRS itself is doing the processing
1937 	 * This direct path does not apply when subflows are present. In this
1938 	 * case, packets need to be dispatched to a soft ring according to the
1939 	 * flow's bandwidth and other resources contraints.
1940 	 */
1941 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1942 		mac_direct_rx_t		proc;
1943 		void			*arg1;
1944 		mac_resource_handle_t	arg2;
1945 
1946 		/*
1947 		 * This is the case when a Rx is directly
1948 		 * assigned and we have a fully classified
1949 		 * protocol chain. We can deal with it in
1950 		 * one shot.
1951 		 */
1952 		proc = srs_rx->sr_func;
1953 		arg1 = srs_rx->sr_arg1;
1954 		arg2 = srs_rx->sr_arg2;
1955 
1956 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1957 		mutex_exit(&mac_srs->srs_lock);
1958 		if (tid != 0) {
1959 			(void) untimeout(tid);
1960 			tid = 0;
1961 		}
1962 
1963 		proc(arg1, arg2, head, NULL);
1964 		/*
1965 		 * Decrement the size and count here itelf
1966 		 * since the packet has been processed.
1967 		 */
1968 		mutex_enter(&mac_srs->srs_lock);
1969 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1970 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
1971 
1972 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1973 			cv_signal(&mac_srs->srs_client_cv);
1974 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1975 	} else {
1976 		/* Some kind of softrings based fanout is required */
1977 		mutex_exit(&mac_srs->srs_lock);
1978 		if (tid != 0) {
1979 			(void) untimeout(tid);
1980 			tid = 0;
1981 		}
1982 
1983 		/*
1984 		 * Since the fanout routines can deal with chains,
1985 		 * shoot the entire chain up.
1986 		 */
1987 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1988 			mac_rx_srs_fanout(mac_srs, head);
1989 		else
1990 			mac_rx_srs_proto_fanout(mac_srs, head);
1991 		mutex_enter(&mac_srs->srs_lock);
1992 	}
1993 
1994 	/*
1995 	 * Send the poll thread to pick up any packets arrived
1996 	 * so far. This also serves as the last check in case
1997 	 * nothing else is queued in the SRS. The poll thread
1998 	 * is signalled only in the case the drain was done
1999 	 * by the worker thread and SRS_WORKER is set. The
2000 	 * worker thread can run in parallel as long as the
2001 	 * SRS_WORKER flag is set. We we have nothing else to
2002 	 * process, we can exit while leaving SRS_PROC set
2003 	 * which gives the poll thread control to process and
2004 	 * cleanup once it returns from the NIC.
2005 	 *
2006 	 * If we have nothing else to process, we need to
2007 	 * ensure that we keep holding the srs_lock till
2008 	 * all the checks below are done and control is
2009 	 * handed to the poll thread if it was running.
2010 	 */
2011 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2012 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2013 		if (mac_srs->srs_first != NULL) {
2014 			if (proc_type == SRS_WORKER) {
2015 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2016 				if (srs_rx->sr_poll_pkt_cnt <=
2017 				    srs_rx->sr_lowat)
2018 					MAC_SRS_POLL_RING(mac_srs);
2019 				goto again;
2020 			} else {
2021 				cv_signal(&mac_srs->srs_async);
2022 			}
2023 		}
2024 	}
2025 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2026 
2027 done:
2028 
2029 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2030 		/*
2031 		 * Poll thread is already running. Leave the
2032 		 * SRS_RPOC set and hand over the control to
2033 		 * poll thread.
2034 		 */
2035 		mac_srs->srs_state &= ~proc_type;
2036 		return;
2037 	}
2038 
2039 	/*
2040 	 * If we can't process packets because we have exceeded
2041 	 * B/W limit for this tick, just set the timeout
2042 	 * and leave.
2043 	 *
2044 	 * Even if there are no packets queued in SRS, we
2045 	 * need to make sure that the shared counter is
2046 	 * clear and any associated softrings have cleared
2047 	 * all the backlog. Otherwise, leave the interface
2048 	 * in polling mode and the poll thread will get
2049 	 * signalled once the count goes down to zero.
2050 	 *
2051 	 * If someone is already draining the queue (SRS_PROC is
2052 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2053 	 * then it means that drain is already running and we
2054 	 * will turn off polling at that time if there is
2055 	 * no backlog. As long as there are packets queued either
2056 	 * is soft ring set or its soft rings, we will leave
2057 	 * the interface in polling mode.
2058 	 */
2059 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2060 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
2061 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
2062 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
2063 		MAC_SRS_POLLING_ON(mac_srs);
2064 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2065 		if ((mac_srs->srs_first != NULL) &&
2066 		    (mac_srs->srs_tid == NULL))
2067 			mac_srs->srs_tid = timeout(mac_srs_fire,
2068 			    mac_srs, 1);
2069 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2070 		return;
2071 	}
2072 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2073 
2074 leave_poll:
2075 
2076 	/* Nothing else to do. Get out of poll mode */
2077 	MAC_SRS_POLLING_OFF(mac_srs);
2078 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2079 }
2080 
2081 /*
2082  * mac_srs_worker
2083  *
2084  * The SRS worker routine. Drains the queue when no one else is
2085  * processing it.
2086  */
2087 void
2088 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
2089 {
2090 	kmutex_t 		*lock = &mac_srs->srs_lock;
2091 	kcondvar_t 		*async = &mac_srs->srs_async;
2092 	callb_cpr_t		cprinfo;
2093 	boolean_t		bw_ctl_flag;
2094 
2095 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
2096 	mutex_enter(lock);
2097 
2098 start:
2099 	for (;;) {
2100 		bw_ctl_flag = B_FALSE;
2101 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2102 			MAC_SRS_BW_LOCK(mac_srs);
2103 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2104 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2105 				bw_ctl_flag = B_TRUE;
2106 			MAC_SRS_BW_UNLOCK(mac_srs);
2107 		}
2108 		/*
2109 		 * The SRS_BW_ENFORCED flag may change since we have dropped
2110 		 * the mac_bw_lock. However the drain function can handle both
2111 		 * a drainable SRS or a bandwidth controlled SRS, and the
2112 		 * effect of scheduling a timeout is to wakeup the worker
2113 		 * thread which in turn will call the drain function. Since
2114 		 * we release the srs_lock atomically only in the cv_wait there
2115 		 * isn't a fear of waiting for ever.
2116 		 */
2117 		while (((mac_srs->srs_state & SRS_PROC) ||
2118 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
2119 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
2120 		    !(mac_srs->srs_state & SRS_PAUSE)) {
2121 			/*
2122 			 * If we have packets queued and we are here
2123 			 * because B/W control is in place, we better
2124 			 * schedule the worker wakeup after 1 tick
2125 			 * to see if bandwidth control can be relaxed.
2126 			 */
2127 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
2128 				/*
2129 				 * We need to ensure that a timer  is already
2130 				 * scheduled or we force  schedule one for
2131 				 * later so that we can continue processing
2132 				 * after this  quanta is over.
2133 				 */
2134 				mac_srs->srs_tid = timeout(mac_srs_fire,
2135 				    mac_srs, 1);
2136 			}
2137 wait:
2138 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
2139 			cv_wait(async, lock);
2140 			CALLB_CPR_SAFE_END(&cprinfo, lock);
2141 
2142 			if (mac_srs->srs_state & SRS_PAUSE)
2143 				goto done;
2144 			if (mac_srs->srs_state & SRS_PROC)
2145 				goto wait;
2146 
2147 			if (mac_srs->srs_first != NULL &&
2148 			    mac_srs->srs_type & SRST_BW_CONTROL) {
2149 				MAC_SRS_BW_LOCK(mac_srs);
2150 				if (mac_srs->srs_bw->mac_bw_state &
2151 				    SRS_BW_ENFORCED) {
2152 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2153 				}
2154 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
2155 				    SRS_BW_ENFORCED;
2156 				MAC_SRS_BW_UNLOCK(mac_srs);
2157 			}
2158 		}
2159 
2160 		if (mac_srs->srs_state & SRS_PAUSE)
2161 			goto done;
2162 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
2163 	}
2164 done:
2165 	/*
2166 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
2167 	 * from both hard and soft classifications and waits for such threads
2168 	 * to finish before signaling the worker. So at this point the only
2169 	 * thread left that could be competing with the worker is the poll
2170 	 * thread. In the case of Tx, there shouldn't be any thread holding
2171 	 * SRS_PROC at this point.
2172 	 */
2173 	if (!(mac_srs->srs_state & SRS_PROC)) {
2174 		mac_srs->srs_state |= SRS_PROC;
2175 	} else {
2176 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
2177 		/*
2178 		 * Poll thread still owns the SRS and is still running
2179 		 */
2180 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
2181 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2182 		    SRS_POLL_THR_OWNER));
2183 	}
2184 	mac_srs_worker_quiesce(mac_srs);
2185 	/*
2186 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
2187 	 * of the quiesce operation
2188 	 */
2189 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
2190 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2191 
2192 	if (mac_srs->srs_state & SRS_RESTART) {
2193 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2194 		mac_srs_worker_restart(mac_srs);
2195 		mac_srs->srs_state &= ~SRS_PROC;
2196 		goto start;
2197 	}
2198 
2199 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
2200 		mac_srs_worker_quiesce(mac_srs);
2201 
2202 	mac_srs->srs_state &= ~SRS_PROC;
2203 	/* The macro drops the srs_lock */
2204 	CALLB_CPR_EXIT(&cprinfo);
2205 	thread_exit();
2206 }
2207 
2208 /*
2209  * mac_rx_srs_subflow_process
2210  *
2211  * Receive side routine called from interrupt path when there are
2212  * sub flows present on this SRS.
2213  */
2214 /* ARGSUSED */
2215 void
2216 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
2217     mblk_t *mp_chain, boolean_t loopback)
2218 {
2219 	flow_entry_t		*flent = NULL;
2220 	flow_entry_t		*prev_flent = NULL;
2221 	mblk_t			*mp = NULL;
2222 	mblk_t			*tail = NULL;
2223 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2224 	mac_client_impl_t	*mcip;
2225 
2226 	mcip = mac_srs->srs_mcip;
2227 	ASSERT(mcip != NULL);
2228 
2229 	/*
2230 	 * We need to determine the SRS for every packet
2231 	 * by walking the flow table, if we don't get any,
2232 	 * then we proceed using the SRS we came with.
2233 	 */
2234 	mp = tail = mp_chain;
2235 	while (mp != NULL) {
2236 
2237 		/*
2238 		 * We will increment the stats for the mactching subflow.
2239 		 * when we get the bytes/pkt count for the classified packets
2240 		 * later in mac_rx_srs_process.
2241 		 */
2242 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
2243 		    FLOW_INBOUND, &flent);
2244 
2245 		if (mp == mp_chain || flent == prev_flent) {
2246 			if (prev_flent != NULL)
2247 				FLOW_REFRELE(prev_flent);
2248 			prev_flent = flent;
2249 			flent = NULL;
2250 			tail = mp;
2251 			mp = mp->b_next;
2252 			continue;
2253 		}
2254 		tail->b_next = NULL;
2255 		/*
2256 		 * A null indicates, this is for the mac_srs itself.
2257 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
2258 		 */
2259 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2260 			mac_rx_srs_process(arg,
2261 			    (mac_resource_handle_t)mac_srs, mp_chain,
2262 			    loopback);
2263 		} else {
2264 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2265 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
2266 			FLOW_REFRELE(prev_flent);
2267 		}
2268 		prev_flent = flent;
2269 		flent = NULL;
2270 		mp_chain = mp;
2271 		tail = mp;
2272 		mp = mp->b_next;
2273 	}
2274 	/* Last chain */
2275 	ASSERT(mp_chain != NULL);
2276 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2277 		mac_rx_srs_process(arg,
2278 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
2279 	} else {
2280 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2281 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
2282 		FLOW_REFRELE(prev_flent);
2283 	}
2284 }
2285 
2286 /*
2287  * mac_rx_srs_process
2288  *
2289  * Receive side routine called from the interrupt path.
2290  *
2291  * loopback is set to force a context switch on the loopback
2292  * path between MAC clients.
2293  */
2294 /* ARGSUSED */
2295 void
2296 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
2297     boolean_t loopback)
2298 {
2299 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2300 	mblk_t			*mp, *tail, *head;
2301 	int			count = 0;
2302 	int			count1;
2303 	size_t			sz = 0;
2304 	size_t			chain_sz, sz1;
2305 	mac_bw_ctl_t		*mac_bw;
2306 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2307 
2308 	/*
2309 	 * Set the tail, count and sz. We set the sz irrespective
2310 	 * of whether we are doing B/W control or not for the
2311 	 * purpose of updating the stats.
2312 	 */
2313 	mp = tail = mp_chain;
2314 	while (mp != NULL) {
2315 		tail = mp;
2316 		count++;
2317 		sz += msgdsize(mp);
2318 		mp = mp->b_next;
2319 	}
2320 
2321 	mutex_enter(&mac_srs->srs_lock);
2322 
2323 	if (loopback) {
2324 		SRS_RX_STAT_UPDATE(mac_srs, lclbytes, sz);
2325 		SRS_RX_STAT_UPDATE(mac_srs, lclcnt, count);
2326 
2327 	} else {
2328 		SRS_RX_STAT_UPDATE(mac_srs, intrbytes, sz);
2329 		SRS_RX_STAT_UPDATE(mac_srs, intrcnt, count);
2330 	}
2331 
2332 	/*
2333 	 * If the SRS in already being processed; has been blanked;
2334 	 * can be processed by worker thread only; or the B/W limit
2335 	 * has been reached, then queue the chain and check if
2336 	 * worker thread needs to be awakend.
2337 	 */
2338 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
2339 		mac_bw = mac_srs->srs_bw;
2340 		ASSERT(mac_bw != NULL);
2341 		mutex_enter(&mac_bw->mac_bw_lock);
2342 		mac_bw->mac_bw_intr += sz;
2343 		if (mac_bw->mac_bw_limit == 0) {
2344 			/* zero bandwidth: drop all */
2345 			srs_rx->sr_stat.mrs_sdrops += count;
2346 			mac_bw->mac_bw_drop_bytes += sz;
2347 			mutex_exit(&mac_bw->mac_bw_lock);
2348 			mutex_exit(&mac_srs->srs_lock);
2349 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
2350 			return;
2351 		} else {
2352 			if ((mac_bw->mac_bw_sz + sz) <=
2353 			    mac_bw->mac_bw_drop_threshold) {
2354 				mutex_exit(&mac_bw->mac_bw_lock);
2355 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
2356 				    tail, count, sz);
2357 			} else {
2358 				mp = mp_chain;
2359 				chain_sz = 0;
2360 				count1 = 0;
2361 				tail = NULL;
2362 				head = NULL;
2363 				while (mp != NULL) {
2364 					sz1 = msgdsize(mp);
2365 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
2366 					    mac_bw->mac_bw_drop_threshold)
2367 						break;
2368 					chain_sz += sz1;
2369 					count1++;
2370 					tail = mp;
2371 					mp = mp->b_next;
2372 				}
2373 				mutex_exit(&mac_bw->mac_bw_lock);
2374 				if (tail != NULL) {
2375 					head = tail->b_next;
2376 					tail->b_next = NULL;
2377 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
2378 					    mp_chain, tail, count1, chain_sz);
2379 					sz -= chain_sz;
2380 					count -= count1;
2381 				} else {
2382 					/* Can't pick up any */
2383 					head = mp_chain;
2384 				}
2385 				if (head != NULL) {
2386 					/* Drop any packet over the threshold */
2387 					srs_rx->sr_stat.mrs_sdrops += count;
2388 					mutex_enter(&mac_bw->mac_bw_lock);
2389 					mac_bw->mac_bw_drop_bytes += sz;
2390 					mutex_exit(&mac_bw->mac_bw_lock);
2391 					freemsgchain(head);
2392 				}
2393 			}
2394 			MAC_SRS_WORKER_WAKEUP(mac_srs);
2395 			mutex_exit(&mac_srs->srs_lock);
2396 			return;
2397 		}
2398 	}
2399 
2400 	/*
2401 	 * If the total number of packets queued in the SRS and
2402 	 * its associated soft rings exceeds the max allowed,
2403 	 * then drop the chain. If we are polling capable, this
2404 	 * shouldn't be happening.
2405 	 */
2406 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
2407 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
2408 		mac_bw = mac_srs->srs_bw;
2409 		srs_rx->sr_stat.mrs_sdrops += count;
2410 		mutex_enter(&mac_bw->mac_bw_lock);
2411 		mac_bw->mac_bw_drop_bytes += sz;
2412 		mutex_exit(&mac_bw->mac_bw_lock);
2413 		freemsgchain(mp_chain);
2414 		mutex_exit(&mac_srs->srs_lock);
2415 		return;
2416 	}
2417 
2418 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
2419 
2420 	if (!(mac_srs->srs_state & SRS_PROC)) {
2421 		/*
2422 		 * If we are coming via loopback or if we are not
2423 		 * optimizing for latency, we should signal the
2424 		 * worker thread.
2425 		 */
2426 		if (loopback || !(mac_srs->srs_state & SRS_LATENCY_OPT)) {
2427 			/*
2428 			 * For loopback, We need to let the worker take
2429 			 * over as we don't want to continue in the same
2430 			 * thread even if we can. This could lead to stack
2431 			 * overflows and may also end up using
2432 			 * resources (cpu) incorrectly.
2433 			 */
2434 			cv_signal(&mac_srs->srs_async);
2435 		} else {
2436 			/*
2437 			 * Seems like no one is processing the SRS and
2438 			 * there is no backlog. We also inline process
2439 			 * our packet if its a single packet in non
2440 			 * latency optimized case (in latency optimized
2441 			 * case, we inline process chains of any size).
2442 			 */
2443 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
2444 		}
2445 	}
2446 	mutex_exit(&mac_srs->srs_lock);
2447 }
2448 
2449 /* TX SIDE ROUTINES (RUNTIME) */
2450 
2451 /*
2452  * mac_tx_srs_no_desc
2453  *
2454  * This routine is called by Tx single ring default mode
2455  * when Tx ring runs out of descs.
2456  */
2457 mac_tx_cookie_t
2458 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2459     uint16_t flag, mblk_t **ret_mp)
2460 {
2461 	mac_tx_cookie_t cookie = NULL;
2462 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
2463 	boolean_t wakeup_worker = B_TRUE;
2464 	uint32_t tx_mode = srs_tx->st_mode;
2465 	int cnt, sz;
2466 	mblk_t *tail;
2467 
2468 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
2469 	if (flag & MAC_DROP_ON_NO_DESC) {
2470 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2471 	} else {
2472 		if (mac_srs->srs_first != NULL)
2473 			wakeup_worker = B_FALSE;
2474 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2475 		if (flag & MAC_TX_NO_ENQUEUE) {
2476 			/*
2477 			 * If TX_QUEUED is not set, queue the
2478 			 * packet and let mac_tx_srs_drain()
2479 			 * set the TX_BLOCKED bit for the
2480 			 * reasons explained above. Otherwise,
2481 			 * return the mblks.
2482 			 */
2483 			if (wakeup_worker) {
2484 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2485 				    mp_chain, tail, cnt, sz);
2486 			} else {
2487 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
2488 				    mp_chain, ret_mp, cookie);
2489 			}
2490 		} else {
2491 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2492 			    tail, cnt, sz, cookie);
2493 		}
2494 		if (wakeup_worker)
2495 			cv_signal(&mac_srs->srs_async);
2496 	}
2497 	return (cookie);
2498 }
2499 
2500 /*
2501  * mac_tx_srs_enqueue
2502  *
2503  * This routine is called when Tx SRS is operating in either serializer
2504  * or bandwidth mode. In serializer mode, a packet will get enqueued
2505  * when a thread cannot enter SRS exclusively. In bandwidth mode,
2506  * packets gets queued if allowed byte-count limit for a tick is
2507  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
2508  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
2509  * the default mode or fanout mode. Here packets get dropped or
2510  * returned back to the caller only after hi-watermark worth of data
2511  * is queued.
2512  */
2513 static mac_tx_cookie_t
2514 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2515     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
2516 {
2517 	mac_tx_cookie_t cookie = NULL;
2518 	int cnt, sz;
2519 	mblk_t *tail;
2520 	boolean_t wakeup_worker = B_TRUE;
2521 
2522 	/*
2523 	 * Ignore fanout hint if we don't have multiple tx rings.
2524 	 */
2525 	if (!MAC_TX_SOFT_RINGS(mac_srs))
2526 		fanout_hint = 0;
2527 
2528 	if (mac_srs->srs_first != NULL)
2529 		wakeup_worker = B_FALSE;
2530 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2531 	if (flag & MAC_DROP_ON_NO_DESC) {
2532 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
2533 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2534 		} else {
2535 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2536 			    mp_chain, tail, cnt, sz);
2537 		}
2538 	} else if (flag & MAC_TX_NO_ENQUEUE) {
2539 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
2540 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
2541 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
2542 			    ret_mp, cookie);
2543 		} else {
2544 			mp_chain->b_prev = (mblk_t *)fanout_hint;
2545 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2546 			    mp_chain, tail, cnt, sz);
2547 		}
2548 	} else {
2549 		/*
2550 		 * If you are BW_ENFORCED, just enqueue the
2551 		 * packet. srs_worker will drain it at the
2552 		 * prescribed rate. Before enqueueing, save
2553 		 * the fanout hint.
2554 		 */
2555 		mp_chain->b_prev = (mblk_t *)fanout_hint;
2556 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2557 		    tail, cnt, sz, cookie);
2558 	}
2559 	if (wakeup_worker)
2560 		cv_signal(&mac_srs->srs_async);
2561 	return (cookie);
2562 }
2563 
2564 /*
2565  * There are seven tx modes:
2566  *
2567  * 1) Default mode (SRS_TX_DEFAULT)
2568  * 2) Serialization mode (SRS_TX_SERIALIZE)
2569  * 3) Fanout mode (SRS_TX_FANOUT)
2570  * 4) Bandwdith mode (SRS_TX_BW)
2571  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
2572  * 6) aggr Tx mode (SRS_TX_AGGR)
2573  * 7) aggr Tx bw mode (SRS_TX_BW_AGGR)
2574  *
2575  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
2576  * based on the number of Tx rings requested for an SRS and whether
2577  * bandwidth control is requested or not.
2578  *
2579  * The default mode (i.e., no fanout/no bandwidth) is used when the
2580  * underlying NIC does not have Tx rings or just one Tx ring. In this mode,
2581  * the SRS acts as a pass-thru. Packets will go directly to mac_tx_send().
2582  * When the underlying Tx ring runs out of Tx descs, it starts queueing up
2583  * packets in SRS. When flow-control is relieved, the srs_worker drains
2584  * the queued packets and informs blocked clients to restart sending
2585  * packets.
2586  *
2587  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized. This
2588  * mode is used when the link has no Tx rings or only one Tx ring.
2589  *
2590  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
2591  * Tx rings. Each Tx ring will have a soft ring associated with it.
2592  * These soft rings will be hung off the Tx SRS. Queueing if it happens
2593  * due to lack of Tx desc will be in individual soft ring (and not srs)
2594  * associated with Tx ring.
2595  *
2596  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
2597  * only if bw is available. Otherwise the packets will be queued in
2598  * SRS. If fanout to multiple Tx rings is configured, the packets will
2599  * be fanned out among the soft rings associated with the Tx rings.
2600  *
2601  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
2602  * invokes an aggr function, aggr_find_tx_ring(), to find a pseudo Tx ring
2603  * belonging to a port on which the packet has to be sent. Aggr will
2604  * always have a pseudo Tx ring associated with it even when it is an
2605  * aggregation over a single NIC that has no Tx rings. Even in such a
2606  * case, the single pseudo Tx ring will have a soft ring associated with
2607  * it and the soft ring will hang off the SRS.
2608  *
2609  * If a bandwidth is specified for an aggr, SRS_TX_BW_AGGR mode is used.
2610  * In this mode, the bandwidth is first applied on the outgoing packets
2611  * and later mac_tx_addr_mode() function is called to send the packet out
2612  * of one of the pseudo Tx rings.
2613  *
2614  * Four flags are used in srs_state for indicating flow control
2615  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
2616  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
2617  * driver below.
2618  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
2619  * and flow-control pressure is applied back to clients. The clients expect
2620  * wakeup when flow-control is relieved.
2621  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
2622  * got returned back to client either due to lack of Tx descs or due to bw
2623  * control reasons. The clients expect a wakeup when condition is relieved.
2624  *
2625  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
2626  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
2627  * MAC_TX_NO_ENQUEUE
2628  * Mac clients that do not want packets to be enqueued in the mac layer set
2629  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
2630  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
2631  * behaviour of this flag is different when the Tx is running in serializer
2632  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
2633  * get dropped when Tx high watermark is reached.
2634  * There are some mac clients like vsw, aggr that want the mblks to be
2635  * returned back to clients instead of being queued in Tx SRS (or Tx soft
2636  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
2637  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
2638  * In the default and Tx fanout mode, the un-transmitted mblks will be
2639  * returned back to the clients when the driver runs out of Tx descs.
2640  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
2641  * soft ring) so that the clients can be woken up when Tx desc become
2642  * available. When running in serializer or bandwidth mode mode,
2643  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
2644  */
2645 
2646 mac_tx_func_t
2647 mac_tx_get_func(uint32_t mode)
2648 {
2649 	return (mac_tx_mode_list[mode].mac_tx_func);
2650 }
2651 
2652 /* ARGSUSED */
2653 static mac_tx_cookie_t
2654 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2655     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2656 {
2657 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2658 	mac_tx_stats_t		stats;
2659 	mac_tx_cookie_t		cookie = NULL;
2660 
2661 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
2662 
2663 	/* Regular case with a single Tx ring */
2664 	/*
2665 	 * SRS_TX_BLOCKED is set when underlying NIC runs
2666 	 * out of Tx descs and messages start getting
2667 	 * queued. It won't get reset until
2668 	 * tx_srs_drain() completely drains out the
2669 	 * messages.
2670 	 */
2671 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2672 		/* Tx descs/resources not available */
2673 		mutex_enter(&mac_srs->srs_lock);
2674 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2675 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
2676 			    flag, ret_mp);
2677 			mutex_exit(&mac_srs->srs_lock);
2678 			return (cookie);
2679 		}
2680 		/*
2681 		 * While we were computing mblk count, the
2682 		 * flow control condition got relieved.
2683 		 * Continue with the transmission.
2684 		 */
2685 		mutex_exit(&mac_srs->srs_lock);
2686 	}
2687 
2688 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2689 	    mp_chain, &stats);
2690 
2691 	/*
2692 	 * Multiple threads could be here sending packets.
2693 	 * Under such conditions, it is not possible to
2694 	 * automically set SRS_TX_BLOCKED bit to indicate
2695 	 * out of tx desc condition. To atomically set
2696 	 * this, we queue the returned packet and do
2697 	 * the setting of SRS_TX_BLOCKED in
2698 	 * mac_tx_srs_drain().
2699 	 */
2700 	if (mp_chain != NULL) {
2701 		mutex_enter(&mac_srs->srs_lock);
2702 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
2703 		mutex_exit(&mac_srs->srs_lock);
2704 		return (cookie);
2705 	}
2706 	SRS_TX_STATS_UPDATE(mac_srs, &stats);
2707 
2708 	return (NULL);
2709 }
2710 
2711 /*
2712  * mac_tx_serialize_mode
2713  *
2714  * This is an experimental mode implemented as per the request of PAE.
2715  * In this mode, all callers attempting to send a packet to the NIC
2716  * will get serialized. Only one thread at any time will access the
2717  * NIC to send the packet out.
2718  */
2719 /* ARGSUSED */
2720 static mac_tx_cookie_t
2721 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2722     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2723 {
2724 	mac_tx_stats_t		stats;
2725 	mac_tx_cookie_t		cookie = NULL;
2726 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2727 
2728 	/* Single ring, serialize below */
2729 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
2730 	mutex_enter(&mac_srs->srs_lock);
2731 	if ((mac_srs->srs_first != NULL) ||
2732 	    (mac_srs->srs_state & SRS_PROC)) {
2733 		/*
2734 		 * In serialization mode, queue all packets until
2735 		 * TX_HIWAT is set.
2736 		 * If drop bit is set, drop if TX_HIWAT is set.
2737 		 * If no_enqueue is set, still enqueue until hiwat
2738 		 * is set and return mblks after TX_HIWAT is set.
2739 		 */
2740 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
2741 		    flag, NULL, ret_mp);
2742 		mutex_exit(&mac_srs->srs_lock);
2743 		return (cookie);
2744 	}
2745 	/*
2746 	 * No packets queued, nothing on proc and no flow
2747 	 * control condition. Fast-path, ok. Do inline
2748 	 * processing.
2749 	 */
2750 	mac_srs->srs_state |= SRS_PROC;
2751 	mutex_exit(&mac_srs->srs_lock);
2752 
2753 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2754 	    mp_chain, &stats);
2755 
2756 	mutex_enter(&mac_srs->srs_lock);
2757 	mac_srs->srs_state &= ~SRS_PROC;
2758 	if (mp_chain != NULL) {
2759 		cookie = mac_tx_srs_enqueue(mac_srs,
2760 		    mp_chain, flag, NULL, ret_mp);
2761 	}
2762 	if (mac_srs->srs_first != NULL) {
2763 		/*
2764 		 * We processed inline our packet and a new
2765 		 * packet/s got queued while we were
2766 		 * processing. Wakeup srs worker
2767 		 */
2768 		cv_signal(&mac_srs->srs_async);
2769 	}
2770 	mutex_exit(&mac_srs->srs_lock);
2771 
2772 	if (cookie == NULL)
2773 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
2774 
2775 	return (cookie);
2776 }
2777 
2778 /*
2779  * mac_tx_fanout_mode
2780  *
2781  * In this mode, the SRS will have access to multiple Tx rings to send
2782  * the packet out. The fanout hint that is passed as an argument is
2783  * used to find an appropriate ring to fanout the traffic. Each Tx
2784  * ring, in turn,  will have a soft ring associated with it. If a Tx
2785  * ring runs out of Tx desc's the returned packet will be queued in
2786  * the soft ring associated with that Tx ring. The srs itself will not
2787  * queue any packets.
2788  */
2789 
2790 #define	MAC_TX_SOFT_RING_PROCESS(chain) {		       		\
2791 	index = COMPUTE_INDEX(hash, mac_srs->srs_tx_ring_count),	\
2792 	softring = mac_srs->srs_tx_soft_rings[index];			\
2793 	cookie = mac_tx_soft_ring_process(softring, chain, flag, ret_mp); \
2794 	DTRACE_PROBE2(tx__fanout, uint64_t, hash, uint_t, index);	\
2795 }
2796 
2797 static mac_tx_cookie_t
2798 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2799     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2800 {
2801 	mac_soft_ring_t		*softring;
2802 	uint64_t		hash;
2803 	uint_t			index;
2804 	mac_tx_cookie_t		cookie = NULL;
2805 
2806 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
2807 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
2808 	if (fanout_hint != 0) {
2809 		/*
2810 		 * The hint is specified by the caller, simply pass the
2811 		 * whole chain to the soft ring.
2812 		 */
2813 		hash = HASH_HINT(fanout_hint);
2814 		MAC_TX_SOFT_RING_PROCESS(mp_chain);
2815 	} else {
2816 		mblk_t *last_mp, *cur_mp, *sub_chain;
2817 		uint64_t last_hash = 0;
2818 		uint_t media = mac_srs->srs_mcip->mci_mip->mi_info.mi_media;
2819 
2820 		/*
2821 		 * Compute the hash from the contents (headers) of the
2822 		 * packets of the mblk chain. Split the chains into
2823 		 * subchains of the same conversation.
2824 		 *
2825 		 * Since there may be more than one ring used for
2826 		 * sub-chains of the same call, and since the caller
2827 		 * does not maintain per conversation state since it
2828 		 * passed a zero hint, unsent subchains will be
2829 		 * dropped.
2830 		 */
2831 
2832 		flag |= MAC_DROP_ON_NO_DESC;
2833 		ret_mp = NULL;
2834 
2835 		ASSERT(ret_mp == NULL);
2836 
2837 		sub_chain = NULL;
2838 		last_mp = NULL;
2839 
2840 		for (cur_mp = mp_chain; cur_mp != NULL;
2841 		    cur_mp = cur_mp->b_next) {
2842 			hash = mac_pkt_hash(media, cur_mp, MAC_PKT_HASH_L4,
2843 			    B_TRUE);
2844 			if (last_hash != 0 && hash != last_hash) {
2845 				/*
2846 				 * Starting a different subchain, send current
2847 				 * chain out.
2848 				 */
2849 				ASSERT(last_mp != NULL);
2850 				last_mp->b_next = NULL;
2851 				MAC_TX_SOFT_RING_PROCESS(sub_chain);
2852 				sub_chain = NULL;
2853 			}
2854 
2855 			/* add packet to subchain */
2856 			if (sub_chain == NULL)
2857 				sub_chain = cur_mp;
2858 			last_mp = cur_mp;
2859 			last_hash = hash;
2860 		}
2861 
2862 		if (sub_chain != NULL) {
2863 			/* send last subchain */
2864 			ASSERT(last_mp != NULL);
2865 			last_mp->b_next = NULL;
2866 			MAC_TX_SOFT_RING_PROCESS(sub_chain);
2867 		}
2868 
2869 		cookie = NULL;
2870 	}
2871 
2872 	return (cookie);
2873 }
2874 
2875 /*
2876  * mac_tx_bw_mode
2877  *
2878  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
2879  * only if bw is available. Otherwise the packets will be queued in
2880  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
2881  * out to a Tx rings.
2882  */
2883 static mac_tx_cookie_t
2884 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2885     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2886 {
2887 	int			cnt, sz;
2888 	mblk_t			*tail;
2889 	mac_tx_cookie_t		cookie = NULL;
2890 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2891 	clock_t			now;
2892 
2893 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
2894 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2895 	mutex_enter(&mac_srs->srs_lock);
2896 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2897 		/*
2898 		 * zero bandwidth, no traffic is sent: drop the packets,
2899 		 * or return the whole chain if the caller requests all
2900 		 * unsent packets back.
2901 		 */
2902 		if (flag & MAC_TX_NO_ENQUEUE) {
2903 			cookie = (mac_tx_cookie_t)mac_srs;
2904 			*ret_mp = mp_chain;
2905 		} else {
2906 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2907 		}
2908 		mutex_exit(&mac_srs->srs_lock);
2909 		return (cookie);
2910 	} else if ((mac_srs->srs_first != NULL) ||
2911 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2912 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2913 		    fanout_hint, ret_mp);
2914 		mutex_exit(&mac_srs->srs_lock);
2915 		return (cookie);
2916 	}
2917 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2918 	now = ddi_get_lbolt();
2919 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
2920 		mac_srs->srs_bw->mac_bw_curr_time = now;
2921 		mac_srs->srs_bw->mac_bw_used = 0;
2922 	} else if (mac_srs->srs_bw->mac_bw_used >
2923 	    mac_srs->srs_bw->mac_bw_limit) {
2924 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2925 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2926 		    mp_chain, tail, cnt, sz);
2927 		/*
2928 		 * Wakeup worker thread. Note that worker
2929 		 * thread has to be woken up so that it
2930 		 * can fire up the timer to be woken up
2931 		 * on the next tick. Also once
2932 		 * BW_ENFORCED is set, it can only be
2933 		 * reset by srs_worker thread. Until then
2934 		 * all packets will get queued up in SRS
2935 		 * and hence this this code path won't be
2936 		 * entered until BW_ENFORCED is reset.
2937 		 */
2938 		cv_signal(&mac_srs->srs_async);
2939 		mutex_exit(&mac_srs->srs_lock);
2940 		return (cookie);
2941 	}
2942 
2943 	mac_srs->srs_bw->mac_bw_used += sz;
2944 	mutex_exit(&mac_srs->srs_lock);
2945 
2946 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
2947 		mac_soft_ring_t *softring;
2948 		uint_t indx, hash;
2949 
2950 		hash = HASH_HINT(fanout_hint);
2951 		indx = COMPUTE_INDEX(hash,
2952 		    mac_srs->srs_tx_ring_count);
2953 		softring = mac_srs->srs_tx_soft_rings[indx];
2954 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
2955 		    ret_mp));
2956 	} else if (srs_tx->st_mode == SRS_TX_BW_AGGR) {
2957 		return (mac_tx_aggr_mode(mac_srs, mp_chain,
2958 		    fanout_hint, flag, ret_mp));
2959 	} else {
2960 		mac_tx_stats_t		stats;
2961 
2962 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2963 		    mp_chain, &stats);
2964 
2965 		if (mp_chain != NULL) {
2966 			mutex_enter(&mac_srs->srs_lock);
2967 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2968 			if (mac_srs->srs_bw->mac_bw_used > sz)
2969 				mac_srs->srs_bw->mac_bw_used -= sz;
2970 			else
2971 				mac_srs->srs_bw->mac_bw_used = 0;
2972 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2973 			    fanout_hint, ret_mp);
2974 			mutex_exit(&mac_srs->srs_lock);
2975 			return (cookie);
2976 		}
2977 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
2978 
2979 		return (NULL);
2980 	}
2981 }
2982 
2983 /*
2984  * mac_tx_aggr_mode
2985  *
2986  * This routine invokes an aggr function, aggr_find_tx_ring(), to find
2987  * a (pseudo) Tx ring belonging to a port on which the packet has to
2988  * be sent. aggr_find_tx_ring() first finds the outgoing port based on
2989  * L2/L3/L4 policy and then uses the fanout_hint passed to it to pick
2990  * a Tx ring from the selected port.
2991  *
2992  * Note that a port can be deleted from the aggregation. In such a case,
2993  * the aggregation layer first separates the port from the rest of the
2994  * ports making sure that port (and thus any Tx rings associated with
2995  * it) won't get selected in the call to aggr_find_tx_ring() function.
2996  * Later calls are made to mac_group_rem_ring() passing pseudo Tx ring
2997  * handles one by one which in turn will quiesce the Tx SRS and remove
2998  * the soft ring associated with the pseudo Tx ring. Unlike Rx side
2999  * where a cookie is used to protect against mac_rx_ring() calls on
3000  * rings that have been removed, no such cookie is needed on the Tx
3001  * side as the pseudo Tx ring won't be available anymore to
3002  * aggr_find_tx_ring() once the port has been removed.
3003  */
3004 static mac_tx_cookie_t
3005 mac_tx_aggr_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3006     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3007 {
3008 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3009 	mac_tx_ring_fn_t	find_tx_ring_fn;
3010 	mac_ring_handle_t	ring = NULL;
3011 	void			*arg;
3012 	mac_soft_ring_t		*sringp;
3013 
3014 	find_tx_ring_fn = srs_tx->st_capab_aggr.mca_find_tx_ring_fn;
3015 	arg = srs_tx->st_capab_aggr.mca_arg;
3016 	if (find_tx_ring_fn(arg, mp_chain, fanout_hint, &ring) == NULL)
3017 		return (NULL);
3018 	sringp = srs_tx->st_soft_rings[((mac_ring_t *)ring)->mr_index];
3019 	return (mac_tx_soft_ring_process(sringp, mp_chain, flag, ret_mp));
3020 }
3021 
3022 void
3023 mac_tx_invoke_callbacks(mac_client_impl_t *mcip, mac_tx_cookie_t cookie)
3024 {
3025 	mac_cb_t *mcb;
3026 	mac_tx_notify_cb_t *mtnfp;
3027 
3028 	/* Wakeup callback registered clients */
3029 	MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
3030 	for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
3031 	    mcb = mcb->mcb_nextp) {
3032 		mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
3033 		mtnfp->mtnf_fn(mtnfp->mtnf_arg, cookie);
3034 	}
3035 	MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
3036 	    &mcip->mci_tx_notify_cb_list);
3037 }
3038 
3039 /* ARGSUSED */
3040 void
3041 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
3042 {
3043 	mblk_t			*head, *tail;
3044 	size_t			sz;
3045 	uint32_t		tx_mode;
3046 	uint_t			saved_pkt_count;
3047 	mac_tx_stats_t		stats;
3048 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3049 	clock_t			now;
3050 
3051 	saved_pkt_count = 0;
3052 	ASSERT(mutex_owned(&mac_srs->srs_lock));
3053 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
3054 
3055 	mac_srs->srs_state |= SRS_PROC;
3056 
3057 	tx_mode = srs_tx->st_mode;
3058 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
3059 		if (mac_srs->srs_first != NULL) {
3060 			head = mac_srs->srs_first;
3061 			tail = mac_srs->srs_last;
3062 			saved_pkt_count = mac_srs->srs_count;
3063 			mac_srs->srs_first = NULL;
3064 			mac_srs->srs_last = NULL;
3065 			mac_srs->srs_count = 0;
3066 			mutex_exit(&mac_srs->srs_lock);
3067 
3068 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3069 			    head, &stats);
3070 
3071 			mutex_enter(&mac_srs->srs_lock);
3072 			if (head != NULL) {
3073 				/* Device out of tx desc, set block */
3074 				if (head->b_next == NULL)
3075 					VERIFY(head == tail);
3076 				tail->b_next = mac_srs->srs_first;
3077 				mac_srs->srs_first = head;
3078 				mac_srs->srs_count +=
3079 				    (saved_pkt_count - stats.mts_opackets);
3080 				if (mac_srs->srs_last == NULL)
3081 					mac_srs->srs_last = tail;
3082 				MAC_TX_SRS_BLOCK(mac_srs, head);
3083 			} else {
3084 				srs_tx->st_woken_up = B_FALSE;
3085 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
3086 			}
3087 		}
3088 	} else if (tx_mode == SRS_TX_BW) {
3089 		/*
3090 		 * We are here because the timer fired and we have some data
3091 		 * to tranmit. Also mac_tx_srs_worker should have reset
3092 		 * SRS_BW_ENFORCED flag
3093 		 */
3094 		ASSERT(!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED));
3095 		head = tail = mac_srs->srs_first;
3096 		while (mac_srs->srs_first != NULL) {
3097 			tail = mac_srs->srs_first;
3098 			tail->b_prev = NULL;
3099 			mac_srs->srs_first = tail->b_next;
3100 			if (mac_srs->srs_first == NULL)
3101 				mac_srs->srs_last = NULL;
3102 			mac_srs->srs_count--;
3103 			sz = msgdsize(tail);
3104 			mac_srs->srs_size -= sz;
3105 			saved_pkt_count++;
3106 			MAC_TX_UPDATE_BW_INFO(mac_srs, sz);
3107 
3108 			if (mac_srs->srs_bw->mac_bw_used <
3109 			    mac_srs->srs_bw->mac_bw_limit)
3110 				continue;
3111 
3112 			now = ddi_get_lbolt();
3113 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3114 				mac_srs->srs_bw->mac_bw_curr_time = now;
3115 				mac_srs->srs_bw->mac_bw_used = sz;
3116 				continue;
3117 			}
3118 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3119 			break;
3120 		}
3121 
3122 		ASSERT((head == NULL && tail == NULL) ||
3123 		    (head != NULL && tail != NULL));
3124 		if (tail != NULL) {
3125 			tail->b_next = NULL;
3126 			mutex_exit(&mac_srs->srs_lock);
3127 
3128 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3129 			    head, &stats);
3130 
3131 			mutex_enter(&mac_srs->srs_lock);
3132 			if (head != NULL) {
3133 				uint_t size_sent;
3134 
3135 				/* Device out of tx desc, set block */
3136 				if (head->b_next == NULL)
3137 					VERIFY(head == tail);
3138 				tail->b_next = mac_srs->srs_first;
3139 				mac_srs->srs_first = head;
3140 				mac_srs->srs_count +=
3141 				    (saved_pkt_count - stats.mts_opackets);
3142 				if (mac_srs->srs_last == NULL)
3143 					mac_srs->srs_last = tail;
3144 				size_sent = sz - stats.mts_obytes;
3145 				mac_srs->srs_size += size_sent;
3146 				mac_srs->srs_bw->mac_bw_sz += size_sent;
3147 				if (mac_srs->srs_bw->mac_bw_used > size_sent) {
3148 					mac_srs->srs_bw->mac_bw_used -=
3149 					    size_sent;
3150 				} else {
3151 					mac_srs->srs_bw->mac_bw_used = 0;
3152 				}
3153 				MAC_TX_SRS_BLOCK(mac_srs, head);
3154 			} else {
3155 				srs_tx->st_woken_up = B_FALSE;
3156 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
3157 			}
3158 		}
3159 	} else if (tx_mode == SRS_TX_BW_FANOUT || tx_mode == SRS_TX_BW_AGGR) {
3160 		mblk_t *prev;
3161 		uint64_t hint;
3162 
3163 		/*
3164 		 * We are here because the timer fired and we
3165 		 * have some quota to tranmit.
3166 		 */
3167 		prev = NULL;
3168 		head = tail = mac_srs->srs_first;
3169 		while (mac_srs->srs_first != NULL) {
3170 			tail = mac_srs->srs_first;
3171 			mac_srs->srs_first = tail->b_next;
3172 			if (mac_srs->srs_first == NULL)
3173 				mac_srs->srs_last = NULL;
3174 			mac_srs->srs_count--;
3175 			sz = msgdsize(tail);
3176 			mac_srs->srs_size -= sz;
3177 			mac_srs->srs_bw->mac_bw_used += sz;
3178 			if (prev == NULL)
3179 				hint = (ulong_t)tail->b_prev;
3180 			if (hint != (ulong_t)tail->b_prev) {
3181 				prev->b_next = NULL;
3182 				mutex_exit(&mac_srs->srs_lock);
3183 				TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3184 				head = tail;
3185 				hint = (ulong_t)tail->b_prev;
3186 				mutex_enter(&mac_srs->srs_lock);
3187 			}
3188 
3189 			prev = tail;
3190 			tail->b_prev = NULL;
3191 			if (mac_srs->srs_bw->mac_bw_used <
3192 			    mac_srs->srs_bw->mac_bw_limit)
3193 				continue;
3194 
3195 			now = ddi_get_lbolt();
3196 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3197 				mac_srs->srs_bw->mac_bw_curr_time = now;
3198 				mac_srs->srs_bw->mac_bw_used = 0;
3199 				continue;
3200 			}
3201 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3202 			break;
3203 		}
3204 		ASSERT((head == NULL && tail == NULL) ||
3205 		    (head != NULL && tail != NULL));
3206 		if (tail != NULL) {
3207 			tail->b_next = NULL;
3208 			mutex_exit(&mac_srs->srs_lock);
3209 			TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3210 			mutex_enter(&mac_srs->srs_lock);
3211 		}
3212 	}
3213 	/*
3214 	 * SRS_TX_FANOUT case not considered here because packets
3215 	 * won't be queued in the SRS for this case. Packets will
3216 	 * be sent directly to soft rings underneath and if there
3217 	 * is any queueing at all, it would be in Tx side soft
3218 	 * rings.
3219 	 */
3220 
3221 	/*
3222 	 * When srs_count becomes 0, reset SRS_TX_HIWAT and
3223 	 * SRS_TX_WAKEUP_CLIENT and wakeup registered clients.
3224 	 */
3225 	if (mac_srs->srs_count == 0 && (mac_srs->srs_state &
3226 	    (SRS_TX_HIWAT | SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED))) {
3227 		mac_client_impl_t *mcip = mac_srs->srs_mcip;
3228 		boolean_t wakeup_required = B_FALSE;
3229 
3230 		if (mac_srs->srs_state &
3231 		    (SRS_TX_HIWAT|SRS_TX_WAKEUP_CLIENT)) {
3232 			wakeup_required = B_TRUE;
3233 		}
3234 		mac_srs->srs_state &= ~(SRS_TX_HIWAT |
3235 		    SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED);
3236 		mutex_exit(&mac_srs->srs_lock);
3237 		if (wakeup_required) {
3238 			mac_tx_invoke_callbacks(mcip, (mac_tx_cookie_t)mac_srs);
3239 			/*
3240 			 * If the client is not the primary MAC client, then we
3241 			 * need to send the notification to the clients upper
3242 			 * MAC, i.e. mci_upper_mip.
3243 			 */
3244 			mac_tx_notify(mcip->mci_upper_mip != NULL ?
3245 			    mcip->mci_upper_mip : mcip->mci_mip);
3246 		}
3247 		mutex_enter(&mac_srs->srs_lock);
3248 	}
3249 	mac_srs->srs_state &= ~SRS_PROC;
3250 }
3251 
3252 /*
3253  * Given a packet, get the flow_entry that identifies the flow
3254  * to which that packet belongs. The flow_entry will contain
3255  * the transmit function to be used to send the packet. If the
3256  * function returns NULL, the packet should be sent using the
3257  * underlying NIC.
3258  */
3259 static flow_entry_t *
3260 mac_tx_classify(mac_impl_t *mip, mblk_t *mp)
3261 {
3262 	flow_entry_t		*flent = NULL;
3263 	mac_client_impl_t	*mcip;
3264 	int	err;
3265 
3266 	/*
3267 	 * Do classification on the packet.
3268 	 */
3269 	err = mac_flow_lookup(mip->mi_flow_tab, mp, FLOW_OUTBOUND, &flent);
3270 	if (err != 0)
3271 		return (NULL);
3272 
3273 	/*
3274 	 * This flent might just be an additional one on the MAC client,
3275 	 * i.e. for classification purposes (different fdesc), however
3276 	 * the resources, SRS et. al., are in the mci_flent, so if
3277 	 * this isn't the mci_flent, we need to get it.
3278 	 */
3279 	if ((mcip = flent->fe_mcip) != NULL && mcip->mci_flent != flent) {
3280 		FLOW_REFRELE(flent);
3281 		flent = mcip->mci_flent;
3282 		FLOW_TRY_REFHOLD(flent, err);
3283 		if (err != 0)
3284 			return (NULL);
3285 	}
3286 
3287 	return (flent);
3288 }
3289 
3290 /*
3291  * This macro is only meant to be used by mac_tx_send().
3292  */
3293 #define	CHECK_VID_AND_ADD_TAG(mp) {			\
3294 	if (vid_check) {				\
3295 		int err = 0;				\
3296 							\
3297 		MAC_VID_CHECK(src_mcip, (mp), err);	\
3298 		if (err != 0) {				\
3299 			freemsg((mp));			\
3300 			(mp) = next;			\
3301 			oerrors++;			\
3302 			continue;			\
3303 		}					\
3304 	}						\
3305 	if (add_tag) {					\
3306 		(mp) = mac_add_vlan_tag((mp), 0, vid);	\
3307 		if ((mp) == NULL) {			\
3308 			(mp) = next;			\
3309 			oerrors++;			\
3310 			continue;			\
3311 		}					\
3312 	}						\
3313 }
3314 
3315 mblk_t *
3316 mac_tx_send(mac_client_handle_t mch, mac_ring_handle_t ring, mblk_t *mp_chain,
3317     mac_tx_stats_t *stats)
3318 {
3319 	mac_client_impl_t *src_mcip = (mac_client_impl_t *)mch;
3320 	mac_impl_t *mip = src_mcip->mci_mip;
3321 	uint_t obytes = 0, opackets = 0, oerrors = 0;
3322 	mblk_t *mp = NULL, *next;
3323 	boolean_t vid_check, add_tag;
3324 	uint16_t vid = 0;
3325 
3326 	if (mip->mi_nclients > 1) {
3327 		vid_check = MAC_VID_CHECK_NEEDED(src_mcip);
3328 		add_tag = MAC_TAG_NEEDED(src_mcip);
3329 		if (add_tag)
3330 			vid = mac_client_vid(mch);
3331 	} else {
3332 		ASSERT(mip->mi_nclients == 1);
3333 		vid_check = add_tag = B_FALSE;
3334 	}
3335 
3336 	/*
3337 	 * Fastpath: if there's only one client, we simply send
3338 	 * the packet down to the underlying NIC.
3339 	 */
3340 	if (mip->mi_nactiveclients == 1) {
3341 		DTRACE_PROBE2(fastpath,
3342 		    mac_client_impl_t *, src_mcip, mblk_t *, mp_chain);
3343 
3344 		mp = mp_chain;
3345 		while (mp != NULL) {
3346 			next = mp->b_next;
3347 			mp->b_next = NULL;
3348 			opackets++;
3349 			obytes += (mp->b_cont == NULL ? MBLKL(mp) :
3350 			    msgdsize(mp));
3351 
3352 			CHECK_VID_AND_ADD_TAG(mp);
3353 			MAC_TX(mip, ring, mp, src_mcip);
3354 
3355 			/*
3356 			 * If the driver is out of descriptors and does a
3357 			 * partial send it will return a chain of unsent
3358 			 * mblks. Adjust the accounting stats.
3359 			 */
3360 			if (mp != NULL) {
3361 				opackets--;
3362 				obytes -= msgdsize(mp);
3363 				mp->b_next = next;
3364 				break;
3365 			}
3366 			mp = next;
3367 		}
3368 		goto done;
3369 	}
3370 
3371 	/*
3372 	 * No fastpath, we either have more than one MAC client
3373 	 * defined on top of the same MAC, or one or more MAC
3374 	 * client promiscuous callbacks.
3375 	 */
3376 	DTRACE_PROBE3(slowpath, mac_client_impl_t *,
3377 	    src_mcip, int, mip->mi_nclients, mblk_t *, mp_chain);
3378 
3379 	mp = mp_chain;
3380 	while (mp != NULL) {
3381 		flow_entry_t *dst_flow_ent;
3382 		void *flow_cookie;
3383 		size_t	pkt_size;
3384 		mblk_t *mp1;
3385 
3386 		next = mp->b_next;
3387 		mp->b_next = NULL;
3388 		opackets++;
3389 		pkt_size = (mp->b_cont == NULL ? MBLKL(mp) : msgdsize(mp));
3390 		obytes += pkt_size;
3391 		CHECK_VID_AND_ADD_TAG(mp);
3392 
3393 		/*
3394 		 * Find the destination.
3395 		 */
3396 		dst_flow_ent = mac_tx_classify(mip, mp);
3397 
3398 		if (dst_flow_ent != NULL) {
3399 			size_t	hdrsize;
3400 			int	err = 0;
3401 
3402 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
3403 				struct ether_vlan_header *evhp =
3404 				    (struct ether_vlan_header *)mp->b_rptr;
3405 
3406 				if (ntohs(evhp->ether_tpid) == ETHERTYPE_VLAN)
3407 					hdrsize = sizeof (*evhp);
3408 				else
3409 					hdrsize = sizeof (struct ether_header);
3410 			} else {
3411 				mac_header_info_t	mhi;
3412 
3413 				err = mac_header_info((mac_handle_t)mip,
3414 				    mp, &mhi);
3415 				if (err == 0)
3416 					hdrsize = mhi.mhi_hdrsize;
3417 			}
3418 
3419 			/*
3420 			 * Got a matching flow. It's either another
3421 			 * MAC client, or a broadcast/multicast flow.
3422 			 * Make sure the packet size is within the
3423 			 * allowed size. If not drop the packet and
3424 			 * move to next packet.
3425 			 */
3426 			if (err != 0 ||
3427 			    (pkt_size - hdrsize) > mip->mi_sdu_max) {
3428 				oerrors++;
3429 				DTRACE_PROBE2(loopback__drop, size_t, pkt_size,
3430 				    mblk_t *, mp);
3431 				freemsg(mp);
3432 				mp = next;
3433 				FLOW_REFRELE(dst_flow_ent);
3434 				continue;
3435 			}
3436 			flow_cookie = mac_flow_get_client_cookie(dst_flow_ent);
3437 			if (flow_cookie != NULL) {
3438 				/*
3439 				 * The vnic_bcast_send function expects
3440 				 * to receive the sender MAC client
3441 				 * as value for arg2.
3442 				 */
3443 				mac_bcast_send(flow_cookie, src_mcip, mp,
3444 				    B_TRUE);
3445 			} else {
3446 				/*
3447 				 * loopback the packet to a local MAC
3448 				 * client. We force a context switch
3449 				 * if both source and destination MAC
3450 				 * clients are used by IP, i.e.
3451 				 * bypass is set.
3452 				 */
3453 				boolean_t do_switch;
3454 				mac_client_impl_t *dst_mcip =
3455 				    dst_flow_ent->fe_mcip;
3456 
3457 				/*
3458 				 * Check if there are promiscuous mode
3459 				 * callbacks defined. This check is
3460 				 * done here in the 'else' case and
3461 				 * not in other cases because this
3462 				 * path is for local loopback
3463 				 * communication which does not go
3464 				 * through MAC_TX(). For paths that go
3465 				 * through MAC_TX(), the promisc_list
3466 				 * check is done inside the MAC_TX()
3467 				 * macro.
3468 				 */
3469 				if (mip->mi_promisc_list != NULL)
3470 					mac_promisc_dispatch(mip, mp, src_mcip);
3471 
3472 				do_switch = ((src_mcip->mci_state_flags &
3473 				    dst_mcip->mci_state_flags &
3474 				    MCIS_CLIENT_POLL_CAPABLE) != 0);
3475 
3476 				if ((mp1 = mac_fix_cksum(mp)) != NULL) {
3477 					(dst_flow_ent->fe_cb_fn)(
3478 					    dst_flow_ent->fe_cb_arg1,
3479 					    dst_flow_ent->fe_cb_arg2,
3480 					    mp1, do_switch);
3481 				}
3482 			}
3483 			FLOW_REFRELE(dst_flow_ent);
3484 		} else {
3485 			/*
3486 			 * Unknown destination, send via the underlying
3487 			 * NIC.
3488 			 */
3489 			MAC_TX(mip, ring, mp, src_mcip);
3490 			if (mp != NULL) {
3491 				/*
3492 				 * Adjust for the last packet that
3493 				 * could not be transmitted
3494 				 */
3495 				opackets--;
3496 				obytes -= pkt_size;
3497 				mp->b_next = next;
3498 				break;
3499 			}
3500 		}
3501 		mp = next;
3502 	}
3503 
3504 done:
3505 	stats->mts_obytes = obytes;
3506 	stats->mts_opackets = opackets;
3507 	stats->mts_oerrors = oerrors;
3508 	return (mp);
3509 }
3510 
3511 /*
3512  * mac_tx_srs_ring_present
3513  *
3514  * Returns whether the specified ring is part of the specified SRS.
3515  */
3516 boolean_t
3517 mac_tx_srs_ring_present(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3518 {
3519 	int i;
3520 	mac_soft_ring_t *soft_ring;
3521 
3522 	if (srs->srs_tx.st_arg2 == tx_ring)
3523 		return (B_TRUE);
3524 
3525 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
3526 		soft_ring =  srs->srs_tx_soft_rings[i];
3527 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3528 			return (B_TRUE);
3529 	}
3530 
3531 	return (B_FALSE);
3532 }
3533 
3534 /*
3535  * mac_tx_srs_get_soft_ring
3536  *
3537  * Returns the TX soft ring associated with the given ring, if present.
3538  */
3539 mac_soft_ring_t *
3540 mac_tx_srs_get_soft_ring(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3541 {
3542 	int		i;
3543 	mac_soft_ring_t	*soft_ring;
3544 
3545 	if (srs->srs_tx.st_arg2 == tx_ring)
3546 		return (NULL);
3547 
3548 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
3549 		soft_ring =  srs->srs_tx_soft_rings[i];
3550 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3551 			return (soft_ring);
3552 	}
3553 
3554 	return (NULL);
3555 }
3556 
3557 /*
3558  * mac_tx_srs_wakeup
3559  *
3560  * Called when Tx desc become available. Wakeup the appropriate worker
3561  * thread after resetting the SRS_TX_BLOCKED/S_RING_BLOCK bit in the
3562  * state field.
3563  */
3564 void
3565 mac_tx_srs_wakeup(mac_soft_ring_set_t *mac_srs, mac_ring_handle_t ring)
3566 {
3567 	int i;
3568 	mac_soft_ring_t *sringp;
3569 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3570 
3571 	mutex_enter(&mac_srs->srs_lock);
3572 	/*
3573 	 * srs_tx_ring_count == 0 is the single ring mode case. In
3574 	 * this mode, there will not be Tx soft rings associated
3575 	 * with the SRS.
3576 	 */
3577 	if (!MAC_TX_SOFT_RINGS(mac_srs)) {
3578 		if (srs_tx->st_arg2 == ring &&
3579 		    mac_srs->srs_state & SRS_TX_BLOCKED) {
3580 			mac_srs->srs_state &= ~SRS_TX_BLOCKED;
3581 			srs_tx->st_stat.mts_unblockcnt++;
3582 			cv_signal(&mac_srs->srs_async);
3583 		}
3584 		/*
3585 		 * A wakeup can come before tx_srs_drain() could
3586 		 * grab srs lock and set SRS_TX_BLOCKED. So
3587 		 * always set woken_up flag when we come here.
3588 		 */
3589 		srs_tx->st_woken_up = B_TRUE;
3590 		mutex_exit(&mac_srs->srs_lock);
3591 		return;
3592 	}
3593 
3594 	/*
3595 	 * If you are here, it is for FANOUT, BW_FANOUT,
3596 	 * AGGR_MODE or AGGR_BW_MODE case
3597 	 */
3598 	for (i = 0; i < mac_srs->srs_tx_ring_count; i++) {
3599 		sringp = mac_srs->srs_tx_soft_rings[i];
3600 		mutex_enter(&sringp->s_ring_lock);
3601 		if (sringp->s_ring_tx_arg2 == ring) {
3602 			if (sringp->s_ring_state & S_RING_BLOCK) {
3603 				sringp->s_ring_state &= ~S_RING_BLOCK;
3604 				sringp->s_st_stat.mts_unblockcnt++;
3605 				cv_signal(&sringp->s_ring_async);
3606 			}
3607 			sringp->s_ring_tx_woken_up = B_TRUE;
3608 		}
3609 		mutex_exit(&sringp->s_ring_lock);
3610 	}
3611 	mutex_exit(&mac_srs->srs_lock);
3612 }
3613 
3614 /*
3615  * Once the driver is done draining, send a MAC_NOTE_TX notification to unleash
3616  * the blocked clients again.
3617  */
3618 void
3619 mac_tx_notify(mac_impl_t *mip)
3620 {
3621 	i_mac_notify(mip, MAC_NOTE_TX);
3622 }
3623 
3624 /*
3625  * RX SOFTRING RELATED FUNCTIONS
3626  *
3627  * These functions really belong in mac_soft_ring.c and here for
3628  * a short period.
3629  */
3630 
3631 #define	SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3632 	/*								\
3633 	 * Enqueue our mblk chain.					\
3634 	 */								\
3635 	ASSERT(MUTEX_HELD(&(ringp)->s_ring_lock));			\
3636 									\
3637 	if ((ringp)->s_ring_last != NULL)				\
3638 		(ringp)->s_ring_last->b_next = (mp);			\
3639 	else								\
3640 		(ringp)->s_ring_first = (mp);				\
3641 	(ringp)->s_ring_last = (tail);					\
3642 	(ringp)->s_ring_count += (cnt);					\
3643 	ASSERT((ringp)->s_ring_count > 0);				\
3644 	if ((ringp)->s_ring_type & ST_RING_BW_CTL) {			\
3645 		(ringp)->s_ring_size += sz;				\
3646 	}								\
3647 }
3648 
3649 /*
3650  * Default entry point to deliver a packet chain to a MAC client.
3651  * If the MAC client has flows, do the classification with these
3652  * flows as well.
3653  */
3654 /* ARGSUSED */
3655 void
3656 mac_rx_deliver(void *arg1, mac_resource_handle_t mrh, mblk_t *mp_chain,
3657     mac_header_info_t *arg3)
3658 {
3659 	mac_client_impl_t *mcip = arg1;
3660 
3661 	if (mcip->mci_nvids == 1 &&
3662 	    !(mcip->mci_state_flags & MCIS_STRIP_DISABLE)) {
3663 		/*
3664 		 * If the client has exactly one VID associated with it
3665 		 * and striping of VLAN header is not disabled,
3666 		 * remove the VLAN tag from the packet before
3667 		 * passing it on to the client's receive callback.
3668 		 * Note that this needs to be done after we dispatch
3669 		 * the packet to the promiscuous listeners of the
3670 		 * client, since they expect to see the whole
3671 		 * frame including the VLAN headers.
3672 		 */
3673 		mp_chain = mac_strip_vlan_tag_chain(mp_chain);
3674 	}
3675 
3676 	mcip->mci_rx_fn(mcip->mci_rx_arg, mrh, mp_chain, B_FALSE);
3677 }
3678 
3679 /*
3680  * mac_rx_soft_ring_process
3681  *
3682  * process a chain for a given soft ring. The number of packets queued
3683  * in the SRS and its associated soft rings (including this one) is
3684  * very small (tracked by srs_poll_pkt_cnt), then allow the entering
3685  * thread (interrupt or poll thread) to do inline processing. This
3686  * helps keep the latency down under low load.
3687  *
3688  * The proc and arg for each mblk is already stored in the mblk in
3689  * appropriate places.
3690  */
3691 /* ARGSUSED */
3692 void
3693 mac_rx_soft_ring_process(mac_client_impl_t *mcip, mac_soft_ring_t *ringp,
3694     mblk_t *mp_chain, mblk_t *tail, int cnt, size_t sz)
3695 {
3696 	mac_direct_rx_t		proc;
3697 	void			*arg1;
3698 	mac_resource_handle_t	arg2;
3699 	mac_soft_ring_set_t	*mac_srs = ringp->s_ring_set;
3700 
3701 	ASSERT(ringp != NULL);
3702 	ASSERT(mp_chain != NULL);
3703 	ASSERT(tail != NULL);
3704 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3705 
3706 	mutex_enter(&ringp->s_ring_lock);
3707 	ringp->s_ring_total_inpkt += cnt;
3708 	ringp->s_ring_total_rbytes += sz;
3709 	if ((mac_srs->srs_rx.sr_poll_pkt_cnt <= 1) &&
3710 	    !(ringp->s_ring_type & ST_RING_WORKER_ONLY)) {
3711 		/* If on processor or blanking on, then enqueue and return */
3712 		if (ringp->s_ring_state & S_RING_BLANK ||
3713 		    ringp->s_ring_state & S_RING_PROC) {
3714 			SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3715 			mutex_exit(&ringp->s_ring_lock);
3716 			return;
3717 		}
3718 		proc = ringp->s_ring_rx_func;
3719 		arg1 = ringp->s_ring_rx_arg1;
3720 		arg2 = ringp->s_ring_rx_arg2;
3721 		/*
3722 		 * See if anything is already queued. If we are the
3723 		 * first packet, do inline processing else queue the
3724 		 * packet and do the drain.
3725 		 */
3726 		if (ringp->s_ring_first == NULL) {
3727 			/*
3728 			 * Fast-path, ok to process and nothing queued.
3729 			 */
3730 			ringp->s_ring_run = curthread;
3731 			ringp->s_ring_state |= (S_RING_PROC);
3732 
3733 			mutex_exit(&ringp->s_ring_lock);
3734 
3735 			/*
3736 			 * We are the chain of 1 packet so
3737 			 * go through this fast path.
3738 			 */
3739 			ASSERT(mp_chain->b_next == NULL);
3740 
3741 			(*proc)(arg1, arg2, mp_chain, NULL);
3742 
3743 			ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3744 			/*
3745 			 * If we have a soft ring set which is doing
3746 			 * bandwidth control, we need to decrement
3747 			 * srs_size and count so it the SRS can have a
3748 			 * accurate idea of what is the real data
3749 			 * queued between SRS and its soft rings. We
3750 			 * decrement the counters only when the packet
3751 			 * gets processed by both SRS and the soft ring.
3752 			 */
3753 			mutex_enter(&mac_srs->srs_lock);
3754 			MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
3755 			MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
3756 			mutex_exit(&mac_srs->srs_lock);
3757 
3758 			mutex_enter(&ringp->s_ring_lock);
3759 			ringp->s_ring_run = NULL;
3760 			ringp->s_ring_state &= ~S_RING_PROC;
3761 			if (ringp->s_ring_state & S_RING_CLIENT_WAIT)
3762 				cv_signal(&ringp->s_ring_client_cv);
3763 
3764 			if ((ringp->s_ring_first == NULL) ||
3765 			    (ringp->s_ring_state & S_RING_BLANK)) {
3766 				/*
3767 				 * We processed inline our packet and
3768 				 * nothing new has arrived or our
3769 				 * receiver doesn't want to receive
3770 				 * any packets. We are done.
3771 				 */
3772 				mutex_exit(&ringp->s_ring_lock);
3773 				return;
3774 			}
3775 		} else {
3776 			SOFT_RING_ENQUEUE_CHAIN(ringp,
3777 			    mp_chain, tail, cnt, sz);
3778 		}
3779 
3780 		/*
3781 		 * We are here because either we couldn't do inline
3782 		 * processing (because something was already
3783 		 * queued), or we had a chain of more than one
3784 		 * packet, or something else arrived after we were
3785 		 * done with inline processing.
3786 		 */
3787 		ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3788 		ASSERT(ringp->s_ring_first != NULL);
3789 
3790 		ringp->s_ring_drain_func(ringp);
3791 		mutex_exit(&ringp->s_ring_lock);
3792 		return;
3793 	} else {
3794 		/* ST_RING_WORKER_ONLY case */
3795 		SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3796 		mac_soft_ring_worker_wakeup(ringp);
3797 		mutex_exit(&ringp->s_ring_lock);
3798 	}
3799 }
3800 
3801 /*
3802  * TX SOFTRING RELATED FUNCTIONS
3803  *
3804  * These functions really belong in mac_soft_ring.c and here for
3805  * a short period.
3806  */
3807 
3808 #define	TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3809 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));			\
3810 	ringp->s_ring_state |= S_RING_ENQUEUED;				\
3811 	SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);	\
3812 }
3813 
3814 /*
3815  * mac_tx_sring_queued
3816  *
3817  * When we are out of transmit descriptors and we already have a
3818  * queue that exceeds hiwat (or the client called us with
3819  * MAC_TX_NO_ENQUEUE or MAC_DROP_ON_NO_DESC flag), return the
3820  * soft ring pointer as the opaque cookie for the client enable
3821  * flow control.
3822  */
3823 static mac_tx_cookie_t
3824 mac_tx_sring_enqueue(mac_soft_ring_t *ringp, mblk_t *mp_chain, uint16_t flag,
3825     mblk_t **ret_mp)
3826 {
3827 	int cnt;
3828 	size_t sz;
3829 	mblk_t *tail;
3830 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3831 	mac_tx_cookie_t cookie = NULL;
3832 	boolean_t wakeup_worker = B_TRUE;
3833 
3834 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3835 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3836 	if (flag & MAC_DROP_ON_NO_DESC) {
3837 		mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
3838 		/* increment freed stats */
3839 		ringp->s_ring_drops += cnt;
3840 		cookie = (mac_tx_cookie_t)ringp;
3841 	} else {
3842 		if (ringp->s_ring_first != NULL)
3843 			wakeup_worker = B_FALSE;
3844 
3845 		if (flag & MAC_TX_NO_ENQUEUE) {
3846 			/*
3847 			 * If QUEUED is not set, queue the packet
3848 			 * and let mac_tx_soft_ring_drain() set
3849 			 * the TX_BLOCKED bit for the reasons
3850 			 * explained above. Otherwise, return the
3851 			 * mblks.
3852 			 */
3853 			if (wakeup_worker) {
3854 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp,
3855 				    mp_chain, tail, cnt, sz);
3856 			} else {
3857 				ringp->s_ring_state |= S_RING_WAKEUP_CLIENT;
3858 				cookie = (mac_tx_cookie_t)ringp;
3859 				*ret_mp = mp_chain;
3860 			}
3861 		} else {
3862 			boolean_t enqueue = B_TRUE;
3863 
3864 			if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3865 				/*
3866 				 * flow-controlled. Store ringp in cookie
3867 				 * so that it can be returned as
3868 				 * mac_tx_cookie_t to client
3869 				 */
3870 				ringp->s_ring_state |= S_RING_TX_HIWAT;
3871 				cookie = (mac_tx_cookie_t)ringp;
3872 				ringp->s_ring_hiwat_cnt++;
3873 				if (ringp->s_ring_count >
3874 				    ringp->s_ring_tx_max_q_cnt) {
3875 					/* increment freed stats */
3876 					ringp->s_ring_drops += cnt;
3877 					/*
3878 					 * b_prev may be set to the fanout hint
3879 					 * hence can't use freemsg directly
3880 					 */
3881 					mac_pkt_drop(NULL, NULL,
3882 					    mp_chain, B_FALSE);
3883 					DTRACE_PROBE1(tx_queued_hiwat,
3884 					    mac_soft_ring_t *, ringp);
3885 					enqueue = B_FALSE;
3886 				}
3887 			}
3888 			if (enqueue) {
3889 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain,
3890 				    tail, cnt, sz);
3891 			}
3892 		}
3893 		if (wakeup_worker)
3894 			cv_signal(&ringp->s_ring_async);
3895 	}
3896 	return (cookie);
3897 }
3898 
3899 
3900 /*
3901  * mac_tx_soft_ring_process
3902  *
3903  * This routine is called when fanning out outgoing traffic among
3904  * multipe Tx rings.
3905  * Note that a soft ring is associated with a h/w Tx ring.
3906  */
3907 mac_tx_cookie_t
3908 mac_tx_soft_ring_process(mac_soft_ring_t *ringp, mblk_t *mp_chain,
3909     uint16_t flag, mblk_t **ret_mp)
3910 {
3911 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3912 	int	cnt;
3913 	size_t	sz;
3914 	mblk_t	*tail;
3915 	mac_tx_cookie_t cookie = NULL;
3916 
3917 	ASSERT(ringp != NULL);
3918 	ASSERT(mp_chain != NULL);
3919 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3920 	/*
3921 	 * The following modes can come here: SRS_TX_BW_FANOUT,
3922 	 * SRS_TX_FANOUT, SRS_TX_AGGR, SRS_TX_BW_AGGR.
3923 	 */
3924 	ASSERT(MAC_TX_SOFT_RINGS(mac_srs));
3925 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3926 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT ||
3927 	    mac_srs->srs_tx.st_mode == SRS_TX_AGGR ||
3928 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_AGGR);
3929 
3930 	if (ringp->s_ring_type & ST_RING_WORKER_ONLY) {
3931 		/* Serialization mode */
3932 
3933 		mutex_enter(&ringp->s_ring_lock);
3934 		if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3935 			cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3936 			    flag, ret_mp);
3937 			mutex_exit(&ringp->s_ring_lock);
3938 			return (cookie);
3939 		}
3940 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3941 		TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3942 		if (ringp->s_ring_state & (S_RING_BLOCK | S_RING_PROC)) {
3943 			/*
3944 			 * If ring is blocked due to lack of Tx
3945 			 * descs, just return. Worker thread
3946 			 * will get scheduled when Tx desc's
3947 			 * become available.
3948 			 */
3949 			mutex_exit(&ringp->s_ring_lock);
3950 			return (cookie);
3951 		}
3952 		mac_soft_ring_worker_wakeup(ringp);
3953 		mutex_exit(&ringp->s_ring_lock);
3954 		return (cookie);
3955 	} else {
3956 		/* Default fanout mode */
3957 		/*
3958 		 * S_RING_BLOCKED is set when underlying NIC runs
3959 		 * out of Tx descs and messages start getting
3960 		 * queued. It won't get reset until
3961 		 * tx_srs_drain() completely drains out the
3962 		 * messages.
3963 		 */
3964 		mac_tx_stats_t		stats;
3965 
3966 		if (ringp->s_ring_state & S_RING_ENQUEUED) {
3967 			/* Tx descs/resources not available */
3968 			mutex_enter(&ringp->s_ring_lock);
3969 			if (ringp->s_ring_state & S_RING_ENQUEUED) {
3970 				cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3971 				    flag, ret_mp);
3972 				mutex_exit(&ringp->s_ring_lock);
3973 				return (cookie);
3974 			}
3975 			/*
3976 			 * While we were computing mblk count, the
3977 			 * flow control condition got relieved.
3978 			 * Continue with the transmission.
3979 			 */
3980 			mutex_exit(&ringp->s_ring_lock);
3981 		}
3982 
3983 		mp_chain = mac_tx_send(ringp->s_ring_tx_arg1,
3984 		    ringp->s_ring_tx_arg2, mp_chain, &stats);
3985 
3986 		/*
3987 		 * Multiple threads could be here sending packets.
3988 		 * Under such conditions, it is not possible to
3989 		 * automically set S_RING_BLOCKED bit to indicate
3990 		 * out of tx desc condition. To atomically set
3991 		 * this, we queue the returned packet and do
3992 		 * the setting of S_RING_BLOCKED in
3993 		 * mac_tx_soft_ring_drain().
3994 		 */
3995 		if (mp_chain != NULL) {
3996 			mutex_enter(&ringp->s_ring_lock);
3997 			cookie =
3998 			    mac_tx_sring_enqueue(ringp, mp_chain, flag, ret_mp);
3999 			mutex_exit(&ringp->s_ring_lock);
4000 			return (cookie);
4001 		}
4002 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
4003 		SOFTRING_TX_STATS_UPDATE(ringp, &stats);
4004 
4005 		return (NULL);
4006 	}
4007 }
4008