xref: /titanic_41/usr/src/uts/common/io/mac/mac_sched.c (revision 96a62ada8aa6cb19b04270da282e7e21ba74b808)
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 
54 typedef struct mac_tx_mode_s {
55 	mac_tx_srs_mode_t	mac_tx_mode;
56 	mac_tx_func_t		mac_tx_func;
57 } mac_tx_mode_t;
58 
59 /*
60  * There are five modes of operation on the Tx side. These modes get set
61  * in mac_tx_srs_setup(). Except for the experimental TX_SERIALIZE mode,
62  * none of the other modes are user configurable. They get selected by
63  * the system depending upon whether the link (or flow) has multiple Tx
64  * rings or a bandwidth configured, etc.
65  */
66 mac_tx_mode_t mac_tx_mode_list[] = {
67 	{SRS_TX_DEFAULT,	mac_tx_single_ring_mode},
68 	{SRS_TX_SERIALIZE,	mac_tx_serializer_mode},
69 	{SRS_TX_FANOUT,		mac_tx_fanout_mode},
70 	{SRS_TX_BW,		mac_tx_bw_mode},
71 	{SRS_TX_BW_FANOUT,	mac_tx_bw_mode}
72 };
73 
74 /*
75  * Soft Ring Set (SRS) - The Run time code that deals with
76  * dynamic polling from the hardware, bandwidth enforcement,
77  * fanout etc.
78  *
79  * We try to use H/W classification on NIC and assign traffic for
80  * a MAC address to a particular Rx ring or ring group. There is a
81  * 1-1 mapping between a SRS and a Rx ring. The SRS dynamically
82  * switches the underlying Rx ring between interrupt and
83  * polling mode and enforces any specified B/W control.
84  *
85  * There is always a SRS created and tied to each H/W and S/W rule.
86  * Whenever we create a H/W rule, we always add the the same rule to
87  * S/W classifier and tie a SRS to it.
88  *
89  * In case a B/W control is specified, it is broken into bytes
90  * per ticks and as soon as the quota for a tick is exhausted,
91  * the underlying Rx ring is forced into poll mode for remainder of
92  * the tick. The SRS poll thread only polls for bytes that are
93  * allowed to come in the SRS. We typically let 4x the configured
94  * B/W worth of packets to come in the SRS (to prevent unnecessary
95  * drops due to bursts) but only process the specified amount.
96  *
97  * A MAC client (e.g. a VNIC or aggr) can have 1 or more
98  * Rx rings (and corresponding SRSs) assigned to it. The SRS
99  * in turn can have softrings to do protocol level fanout or
100  * softrings to do S/W based fanout or both. In case the NIC
101  * has no Rx rings, we do S/W classification to respective SRS.
102  * The S/W classification rule is always setup and ready. This
103  * allows the MAC layer to reassign Rx rings whenever needed
104  * but packets still continue to flow via the default path and
105  * getting S/W classified to correct SRS.
106  *
107  * The SRS's are used on both Tx and Rx side. They use the same
108  * data structure but the processing routines have slightly different
109  * semantics due to the fact that Rx side needs to do dynamic
110  * polling etc.
111  *
112  * Dynamic Polling Notes
113  * =====================
114  *
115  * Each Soft ring set is capable of switching its Rx ring between
116  * interrupt and poll mode and actively 'polls' for packets in
117  * poll mode. If the SRS is implementing a B/W limit, it makes
118  * sure that only Max allowed packets are pulled in poll mode
119  * and goes to poll mode as soon as B/W limit is exceeded. As
120  * such, there are no overheads to implement B/W limits.
121  *
122  * In poll mode, its better to keep the pipeline going where the
123  * SRS worker thread keeps processing packets and poll thread
124  * keeps bringing more packets (specially if they get to run
125  * on different CPUs). This also prevents the overheads associated
126  * by excessive signalling (on NUMA machines, this can be
127  * pretty devastating). The exception is latency optimized case
128  * where worker thread does no work and interrupt and poll thread
129  * are allowed to do their own drain.
130  *
131  * We use the following policy to control Dynamic Polling:
132  * 1) We switch to poll mode anytime the processing
133  *    thread causes a backlog to build up in SRS and
134  *    its associated Soft Rings (sr_poll_pkt_cnt > 0).
135  * 2) As long as the backlog stays under the low water
136  *    mark (sr_lowat), we poll the H/W for more packets.
137  * 3) If the backlog (sr_poll_pkt_cnt) exceeds low
138  *    water mark, we stay in poll mode but don't poll
139  *    the H/W for more packets.
140  * 4) Anytime in polling mode, if we poll the H/W for
141  *    packets and find nothing plus we have an existing
142  *    backlog (sr_poll_pkt_cnt > 0), we stay in polling
143  *    mode but don't poll the H/W for packets anymore
144  *    (let the polling thread go to sleep).
145  * 5) Once the backlog is relived (packets are processed)
146  *    we reenable polling (by signalling the poll thread)
147  *    only when the backlog dips below sr_poll_thres.
148  * 6) sr_hiwat is used exclusively when we are not
149  *    polling capable and is used to decide when to
150  *    drop packets so the SRS queue length doesn't grow
151  *    infinitely.
152  *
153  * NOTE: Also see the block level comment on top of mac_soft_ring.c
154  */
155 
156 /*
157  * mac_latency_optimize
158  *
159  * Controls whether the poll thread can process the packets inline
160  * or let the SRS worker thread do the processing. This applies if
161  * the SRS was not being processed. For latency sensitive traffic,
162  * this needs to be true to allow inline processing. For throughput
163  * under load, this should be false.
164  *
165  * This (and other similar) tunable should be rolled into a link
166  * or flow specific workload hint that can be set using dladm
167  * linkprop (instead of multiple such tunables).
168  */
169 boolean_t mac_latency_optimize = B_TRUE;
170 
171 /*
172  * MAC_RX_SRS_ENQUEUE_CHAIN and MAC_TX_SRS_ENQUEUE_CHAIN
173  *
174  * queue a mp or chain in soft ring set and increment the
175  * local count (srs_count) for the SRS and the shared counter
176  * (srs_poll_pkt_cnt - shared between SRS and its soft rings
177  * to track the total unprocessed packets for polling to work
178  * correctly).
179  *
180  * The size (total bytes queued) counters are incremented only
181  * if we are doing B/W control.
182  */
183 #define	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {		\
184 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
185 	if ((mac_srs)->srs_last != NULL)				\
186 		(mac_srs)->srs_last->b_next = (head);			\
187 	else								\
188 		(mac_srs)->srs_first = (head);				\
189 	(mac_srs)->srs_last = (tail);					\
190 	(mac_srs)->srs_count += count;					\
191 }
192 
193 #define	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
194 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
195 									\
196 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
197 	srs_rx->sr_poll_pkt_cnt += count;				\
198 	ASSERT(srs_rx->sr_poll_pkt_cnt > 0);				\
199 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
200 		(mac_srs)->srs_size += (sz);				\
201 		mutex_enter(&(mac_srs)->srs_bw->mac_bw_lock);		\
202 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
203 		mutex_exit(&(mac_srs)->srs_bw->mac_bw_lock);		\
204 	}								\
205 }
206 
207 #define	MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
208 	mac_srs->srs_state |= SRS_ENQUEUED;				\
209 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
210 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
211 		(mac_srs)->srs_size += (sz);				\
212 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
213 	}								\
214 }
215 
216 /*
217  * Turn polling on routines
218  */
219 #define	MAC_SRS_POLLING_ON(mac_srs) {					\
220 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
221 	if (((mac_srs)->srs_state &					\
222 	    (SRS_POLLING_CAPAB|SRS_POLLING)) == SRS_POLLING_CAPAB) {	\
223 		(mac_srs)->srs_state |= SRS_POLLING;			\
224 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
225 		    (mac_srs)->srs_ring);				\
226 		(mac_srs)->srs_rx.sr_poll_on++;				\
227 	}								\
228 }
229 
230 #define	MAC_SRS_WORKER_POLLING_ON(mac_srs) {				\
231 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
232 	if (((mac_srs)->srs_state &					\
233 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_POLLING)) == 		\
234 	    (SRS_POLLING_CAPAB|SRS_WORKER)) {				\
235 		(mac_srs)->srs_state |= SRS_POLLING;			\
236 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
237 		    (mac_srs)->srs_ring);				\
238 		(mac_srs)->srs_rx.sr_worker_poll_on++;			\
239 	}								\
240 }
241 
242 /*
243  * MAC_SRS_POLL_RING
244  *
245  * Signal the SRS poll thread to poll the underlying H/W ring
246  * provided it wasn't already polling (SRS_GET_PKTS was set).
247  *
248  * Poll thread gets to run only from mac_rx_srs_drain() and only
249  * if the drain was being done by the worker thread.
250  */
251 #define	MAC_SRS_POLL_RING(mac_srs) {					\
252 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
253 									\
254 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
255 	srs_rx->sr_poll_thr_sig++;					\
256 	if (((mac_srs)->srs_state & 					\
257 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_GET_PKTS)) ==		\
258 		(SRS_WORKER|SRS_POLLING_CAPAB)) {			\
259 		(mac_srs)->srs_state |= SRS_GET_PKTS;			\
260 		cv_signal(&(mac_srs)->srs_cv);   			\
261 	} else {							\
262 		srs_rx->sr_poll_thr_busy++;				\
263 	}								\
264 }
265 
266 /*
267  * MAC_SRS_CHECK_BW_CONTROL
268  *
269  * Check to see if next tick has started so we can reset the
270  * SRS_BW_ENFORCED flag and allow more packets to come in the
271  * system.
272  */
273 #define	MAC_SRS_CHECK_BW_CONTROL(mac_srs) {				\
274 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
275 	ASSERT(((mac_srs)->srs_type & SRST_TX) ||			\
276 	    MUTEX_HELD(&(mac_srs)->srs_bw->mac_bw_lock));		\
277 	clock_t now = ddi_get_lbolt();					\
278 	if ((mac_srs)->srs_bw->mac_bw_curr_time != now) {		\
279 		(mac_srs)->srs_bw->mac_bw_curr_time = now;		\
280 		(mac_srs)->srs_bw->mac_bw_used = 0;	       		\
281 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
282 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
283 	}								\
284 }
285 
286 /*
287  * MAC_SRS_WORKER_WAKEUP
288  *
289  * Wake up the SRS worker thread to process the queue as long as
290  * no one else is processing the queue. If we are optimizing for
291  * latency, we wake up the worker thread immediately or else we
292  * wait mac_srs_worker_wakeup_ticks before worker thread gets
293  * woken up.
294  */
295 int mac_srs_worker_wakeup_ticks = 0;
296 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
297 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
298 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
299 		(mac_srs)->srs_tid == NULL) {				\
300 		if (((mac_srs)->srs_state & SRS_LATENCY_OPT) ||		\
301 			(mac_srs_worker_wakeup_ticks == 0))		\
302 			cv_signal(&(mac_srs)->srs_async);		\
303 		else							\
304 			(mac_srs)->srs_tid =				\
305 				timeout(mac_srs_fire, (mac_srs),	\
306 					mac_srs_worker_wakeup_ticks);	\
307 	}								\
308 }
309 
310 #define	TX_SINGLE_RING_MODE(mac_srs)				\
311 	((mac_srs)->srs_tx.st_mode == SRS_TX_DEFAULT || 	\
312 	    (mac_srs)->srs_tx.st_mode == SRS_TX_SERIALIZE ||	\
313 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW)
314 
315 #define	TX_BANDWIDTH_MODE(mac_srs)				\
316 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
317 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT)
318 
319 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
320 	uint_t hash, indx;						\
321 	hash = HASH_HINT(hint);					\
322 	indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);	\
323 	softring = mac_srs->srs_oth_soft_rings[indx];			\
324 	(void) (mac_tx_soft_ring_process(softring, head, 0, NULL));	\
325 }
326 
327 /*
328  * MAC_TX_SRS_BLOCK
329  *
330  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
331  * will be set only if srs_tx_woken_up is FALSE. If
332  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
333  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
334  * attempt to transmit again and not setting SRS_TX_BLOCKED does
335  * that.
336  */
337 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
338 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
339 	if ((srs)->srs_tx.st_woken_up) {			\
340 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
341 	} else {						\
342 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
343 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
344 		(srs)->srs_tx.st_blocked_cnt++;			\
345 	}							\
346 }
347 
348 /*
349  * MAC_TX_SRS_TEST_HIWAT
350  *
351  * Called before queueing a packet onto Tx SRS to test and set
352  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
353  */
354 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
355 	boolean_t enqueue = 1;						\
356 									\
357 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
358 		/*							\
359 		 * flow-controlled. Store srs in cookie so that it	\
360 		 * can be returned as mac_tx_cookie_t to client		\
361 		 */							\
362 		(srs)->srs_state |= SRS_TX_HIWAT;			\
363 		cookie = (mac_tx_cookie_t)srs;				\
364 		(srs)->srs_tx.st_hiwat_cnt++;				\
365 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
366 			/* increment freed stats */			\
367 			(srs)->srs_tx.st_drop_count += cnt;		\
368 			/*						\
369 			 * b_prev may be set to the fanout hint		\
370 			 * hence can't use freemsg directly		\
371 			 */						\
372 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);	\
373 			DTRACE_PROBE1(tx_queued_hiwat,			\
374 			    mac_soft_ring_set_t *, srs);		\
375 			enqueue = 0;					\
376 		}							\
377 	}								\
378 	if (enqueue)							\
379 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
380 }
381 
382 /* Some utility macros */
383 #define	MAC_SRS_BW_LOCK(srs)						\
384 	if (!(srs->srs_type & SRST_TX))					\
385 		mutex_enter(&srs->srs_bw->mac_bw_lock);
386 
387 #define	MAC_SRS_BW_UNLOCK(srs)						\
388 	if (!(srs->srs_type & SRST_TX))					\
389 		mutex_exit(&srs->srs_bw->mac_bw_lock);
390 
391 #define	MAC_TX_SRS_DROP_MESSAGE(srs, mp, cookie) {		\
392 	mac_pkt_drop(NULL, NULL, mp, B_FALSE);			\
393 	/* increment freed stats */				\
394 	mac_srs->srs_tx.st_drop_count++;			\
395 	cookie = (mac_tx_cookie_t)srs;				\
396 }
397 
398 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
399 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
400 	cookie = (mac_tx_cookie_t)srs;					\
401 	*ret_mp = mp_chain;						\
402 }
403 
404 /*
405  * Drop the rx packet and advance to the next one in the chain.
406  */
407 static void
408 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
409 {
410 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
411 
412 	ASSERT(mp->b_next == NULL);
413 	mutex_enter(&srs->srs_lock);
414 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
415 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
416 	mutex_exit(&srs->srs_lock);
417 
418 	srs_rx->sr_drop_count++;
419 	freemsg(mp);
420 }
421 
422 /* DATAPATH RUNTIME ROUTINES */
423 
424 /*
425  * mac_srs_fire
426  *
427  * Timer callback routine for waking up the SRS worker thread.
428  */
429 static void
430 mac_srs_fire(void *arg)
431 {
432 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
433 
434 	mutex_enter(&mac_srs->srs_lock);
435 	if (mac_srs->srs_tid == 0) {
436 		mutex_exit(&mac_srs->srs_lock);
437 		return;
438 	}
439 
440 	mac_srs->srs_tid = 0;
441 	if (!(mac_srs->srs_state & SRS_PROC))
442 		cv_signal(&mac_srs->srs_async);
443 
444 	mutex_exit(&mac_srs->srs_lock);
445 }
446 
447 /*
448  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
449  * and it is used on the TX path.
450  */
451 #define	HASH_HINT(hint) \
452 	((hint) ^ ((hint) >> 24) ^ ((hint) >> 16) ^ ((hint) >> 8))
453 
454 
455 /*
456  * hash based on the src address and the port information.
457  */
458 #define	HASH_ADDR(src, ports)					\
459 	(ntohl((src)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
460 	((ports) >> 8) ^ (ports))
461 
462 #define	COMPUTE_INDEX(key, sz)	(key % sz)
463 
464 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
465 	if ((tail) != NULL) {						\
466 		ASSERT((tail)->b_next == NULL);				\
467 		(tail)->b_next = (mp);					\
468 	} else {							\
469 		ASSERT((head) == NULL);					\
470 		(head) = (mp);						\
471 	}								\
472 	(tail) = (mp);							\
473 	(cnt)++;							\
474 	if ((bw_ctl))							\
475 		(sz) += (sz0);						\
476 }
477 
478 #define	MAC_FANOUT_DEFAULT	0
479 #define	MAC_FANOUT_RND_ROBIN	1
480 int mac_fanout_type = MAC_FANOUT_DEFAULT;
481 
482 #define	MAX_SR_TYPES	3
483 /* fanout types for port based hashing */
484 enum pkt_type {
485 	V4_TCP = 0,
486 	V4_UDP,
487 	OTH,
488 	UNDEF
489 };
490 
491 /*
492  * In general we do port based hashing to spread traffic over different
493  * softrings. The below tunable allows to override that behavior. Setting it
494  * to B_TRUE allows to do a fanout based on src ipv6 address. This behavior
495  * is also the applicable to ipv6 packets carrying multiple optional headers
496  * and other uncommon packet types.
497  */
498 boolean_t mac_src_ipv6_fanout = B_FALSE;
499 
500 /*
501  * Pair of local and remote ports in the transport header
502  */
503 #define	PORTS_SIZE 4
504 
505 /*
506  * mac_rx_srs_proto_fanout
507  *
508  * This routine delivers packets destined to an SRS into one of the
509  * protocol soft rings.
510  *
511  * Given a chain of packets we need to split it up into multiple sub chains
512  * destined into TCP, UDP or OTH soft ring. Instead of entering
513  * the soft ring one packet at a time, we want to enter it in the form of a
514  * chain otherwise we get this start/stop behaviour where the worker thread
515  * goes to sleep and then next packets comes in forcing it to wake up etc.
516  */
517 static void
518 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
519 {
520 	struct ether_header		*ehp;
521 	struct ether_vlan_header	*evhp;
522 	uint32_t			sap;
523 	ipha_t				*ipha;
524 	uint8_t				*dstaddr;
525 	size_t				hdrsize;
526 	mblk_t				*mp;
527 	mblk_t				*headmp[MAX_SR_TYPES];
528 	mblk_t				*tailmp[MAX_SR_TYPES];
529 	int				cnt[MAX_SR_TYPES];
530 	size_t				sz[MAX_SR_TYPES];
531 	size_t				sz1;
532 	boolean_t			bw_ctl;
533 	boolean_t			hw_classified;
534 	boolean_t			dls_bypass;
535 	boolean_t			is_ether;
536 	boolean_t			is_unicast;
537 	enum pkt_type			type;
538 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
539 
540 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
541 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
542 
543 	/*
544 	 * If we don't have a Rx ring, S/W classification would have done
545 	 * its job and its a packet meant for us. If we were polling on
546 	 * the default ring (i.e. there was a ring assigned to this SRS),
547 	 * then we need to make sure that the mac address really belongs
548 	 * to us.
549 	 */
550 	hw_classified = mac_srs->srs_ring != NULL &&
551 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
552 
553 	/*
554 	 * Special clients (eg. VLAN, non ether, etc) need DLS
555 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
556 	 * such SRSs. Another way of disabling bypass is to set the
557 	 * MCIS_RX_BYPASS_DISABLE flag.
558 	 */
559 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
560 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
561 
562 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
563 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
564 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
565 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
566 
567 	/*
568 	 * We got a chain from SRS that we need to send to the soft rings.
569 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
570 	 * performance reasons), we need to separate out v4_tcp, v4_udp
571 	 * and the rest goes in other.
572 	 */
573 	while (head != NULL) {
574 		mp = head;
575 		head = head->b_next;
576 		mp->b_next = NULL;
577 
578 		type = OTH;
579 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
580 
581 		if (is_ether) {
582 			/*
583 			 * At this point we can be sure the packet at least
584 			 * has an ether header.
585 			 */
586 			if (sz1 < sizeof (struct ether_header)) {
587 				mac_rx_drop_pkt(mac_srs, mp);
588 				continue;
589 			}
590 			ehp = (struct ether_header *)mp->b_rptr;
591 
592 			/*
593 			 * Determine if this is a VLAN or non-VLAN packet.
594 			 */
595 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
596 				evhp = (struct ether_vlan_header *)mp->b_rptr;
597 				sap = ntohs(evhp->ether_type);
598 				hdrsize = sizeof (struct ether_vlan_header);
599 				/*
600 				 * Check if the VID of the packet, if any,
601 				 * belongs to this client.
602 				 */
603 				if (!mac_client_check_flow_vid(mcip,
604 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
605 					mac_rx_drop_pkt(mac_srs, mp);
606 					continue;
607 				}
608 			} else {
609 				hdrsize = sizeof (struct ether_header);
610 			}
611 			is_unicast =
612 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
613 			dstaddr = (uint8_t *)&ehp->ether_dhost;
614 		} else {
615 			mac_header_info_t		mhi;
616 
617 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
618 			    mp, &mhi) != 0) {
619 				mac_rx_drop_pkt(mac_srs, mp);
620 				continue;
621 			}
622 			hdrsize = mhi.mhi_hdrsize;
623 			sap = mhi.mhi_bindsap;
624 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
625 			dstaddr = (uint8_t *)mhi.mhi_daddr;
626 		}
627 
628 		if (!dls_bypass) {
629 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
630 			    cnt[type], bw_ctl, sz[type], sz1, mp);
631 			continue;
632 		}
633 
634 		if (sap == ETHERTYPE_IP) {
635 			/*
636 			 * If we are H/W classified, but we have promisc
637 			 * on, then we need to check for the unicast address.
638 			 */
639 			if (hw_classified && mcip->mci_promisc_list != NULL) {
640 				mac_address_t		*map;
641 
642 				rw_enter(&mcip->mci_rw_lock, RW_READER);
643 				map = mcip->mci_unicast;
644 				if (bcmp(dstaddr, map->ma_addr,
645 				    map->ma_len) == 0)
646 					type = UNDEF;
647 				rw_exit(&mcip->mci_rw_lock);
648 			} else if (is_unicast) {
649 				type = UNDEF;
650 			}
651 		}
652 
653 		/*
654 		 * This needs to become a contract with the driver for
655 		 * the fast path.
656 		 *
657 		 * In the normal case the packet will have at least the L2
658 		 * header and the IP + Transport header in the same mblk.
659 		 * This is usually the case when the NIC driver sends up
660 		 * the packet. This is also true when the stack generates
661 		 * a packet that is looped back and when the stack uses the
662 		 * fastpath mechanism. The normal case is optimized for
663 		 * performance and may bypass DLS. All other cases go through
664 		 * the 'OTH' type path without DLS bypass.
665 		 */
666 
667 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
668 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
669 			type = OTH;
670 
671 		if (type == OTH) {
672 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
673 			    cnt[type], bw_ctl, sz[type], sz1, mp);
674 			continue;
675 		}
676 
677 		ASSERT(type == UNDEF);
678 		/*
679 		 * We look for at least 4 bytes past the IP header to get
680 		 * the port information. If we get an IP fragment, we don't
681 		 * have the port information, and we use just the protocol
682 		 * information.
683 		 */
684 		switch (ipha->ipha_protocol) {
685 		case IPPROTO_TCP:
686 			type = V4_TCP;
687 			mp->b_rptr += hdrsize;
688 			break;
689 		case IPPROTO_UDP:
690 			type = V4_UDP;
691 			mp->b_rptr += hdrsize;
692 			break;
693 		default:
694 			type = OTH;
695 			break;
696 		}
697 
698 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
699 		    bw_ctl, sz[type], sz1, mp);
700 	}
701 
702 	for (type = V4_TCP; type < UNDEF; type++) {
703 		if (headmp[type] != NULL) {
704 			mac_soft_ring_t			*softring;
705 
706 			ASSERT(tailmp[type]->b_next == NULL);
707 			switch (type) {
708 			case V4_TCP:
709 				softring = mac_srs->srs_tcp_soft_rings[0];
710 				break;
711 			case V4_UDP:
712 				softring = mac_srs->srs_udp_soft_rings[0];
713 				break;
714 			case OTH:
715 				softring = mac_srs->srs_oth_soft_rings[0];
716 			}
717 			mac_rx_soft_ring_process(mcip, softring,
718 			    headmp[type], tailmp[type], cnt[type], sz[type]);
719 		}
720 	}
721 }
722 
723 int	fanout_unalligned = 0;
724 
725 /*
726  * mac_rx_srs_long_fanout
727  *
728  * The fanout routine for IPv6
729  */
730 static int
731 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
732     uint32_t sap, size_t hdrsize, enum pkt_type *type, uint_t *indx)
733 {
734 	ip6_t		*ip6h;
735 	uint8_t		*whereptr;
736 	uint_t		hash;
737 	uint16_t	remlen;
738 	uint8_t		nexthdr;
739 	uint16_t	hdr_len;
740 
741 	if (sap == ETHERTYPE_IPV6) {
742 		boolean_t	modifiable = B_TRUE;
743 
744 		ASSERT(MBLKL(mp) >= hdrsize);
745 
746 		ip6h = (ip6_t *)(mp->b_rptr + hdrsize);
747 		if ((unsigned char *)ip6h == mp->b_wptr) {
748 			/*
749 			 * The first mblk_t only includes the mac header.
750 			 * Note that it is safe to change the mp pointer here,
751 			 * as the subsequent operation does not assume mp
752 			 * points to the start of the mac header.
753 			 */
754 			mp = mp->b_cont;
755 
756 			/*
757 			 * Make sure ip6h holds the full ip6_t structure.
758 			 */
759 			if (mp == NULL)
760 				return (-1);
761 
762 			if (MBLKL(mp) < IPV6_HDR_LEN) {
763 				modifiable = (DB_REF(mp) == 1);
764 
765 				if (modifiable &&
766 				    !pullupmsg(mp, IPV6_HDR_LEN)) {
767 					return (-1);
768 				}
769 			}
770 
771 			ip6h = (ip6_t *)mp->b_rptr;
772 		}
773 
774 		if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
775 		    ((unsigned char *)ip6h + IPV6_HDR_LEN > mp->b_wptr)) {
776 			/*
777 			 * If either ip6h is not alligned, or ip6h does not
778 			 * hold the complete ip6_t structure (a pullupmsg()
779 			 * is not an option since it would result in an
780 			 * unalligned ip6h), fanout to the default ring. Note
781 			 * that this may cause packets reordering.
782 			 */
783 			*indx = 0;
784 			*type = OTH;
785 			fanout_unalligned++;
786 			return (0);
787 		}
788 
789 		remlen = ntohs(ip6h->ip6_plen);
790 		nexthdr = ip6h->ip6_nxt;
791 
792 		if (remlen < MIN_EHDR_LEN)
793 			return (-1);
794 		/*
795 		 * Do src based fanout if below tunable is set to B_TRUE or
796 		 * when mac_ip_hdr_length_v6() fails because of malformed
797 		 * packets or because mblk's need to be concatenated using
798 		 * pullupmsg().
799 		 */
800 		if (mac_src_ipv6_fanout || !mac_ip_hdr_length_v6(mp, ip6h,
801 		    &hdr_len, &nexthdr, NULL, NULL)) {
802 			goto src_based_fanout;
803 		}
804 		whereptr = (uint8_t *)ip6h + hdr_len;
805 
806 		/* If the transport is one of below, we do port based fanout */
807 		switch (nexthdr) {
808 		case IPPROTO_TCP:
809 		case IPPROTO_UDP:
810 		case IPPROTO_SCTP:
811 		case IPPROTO_ESP:
812 			/*
813 			 * If the ports in the transport header is not part of
814 			 * the mblk, do src_based_fanout, instead of calling
815 			 * pullupmsg().
816 			 */
817 			if (mp->b_cont != NULL &&
818 			    whereptr + PORTS_SIZE > mp->b_wptr) {
819 				goto src_based_fanout;
820 			}
821 			break;
822 		default:
823 			break;
824 		}
825 
826 		switch (nexthdr) {
827 		case IPPROTO_TCP:
828 			hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
829 			    *(uint32_t *)whereptr);
830 			*indx = COMPUTE_INDEX(hash,
831 			    mac_srs->srs_tcp_ring_count);
832 			*type = OTH;
833 			break;
834 
835 		case IPPROTO_UDP:
836 		case IPPROTO_SCTP:
837 		case IPPROTO_ESP:
838 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
839 				hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
840 				    *(uint32_t *)whereptr);
841 				*indx = COMPUTE_INDEX(hash,
842 				    mac_srs->srs_udp_ring_count);
843 			} else {
844 				*indx = mac_srs->srs_ind %
845 				    mac_srs->srs_udp_ring_count;
846 				mac_srs->srs_ind++;
847 			}
848 			*type = OTH;
849 			break;
850 
851 			/* For all other protocol, do source based fanout */
852 		default:
853 			goto src_based_fanout;
854 		}
855 	} else {
856 		*indx = 0;
857 		*type = OTH;
858 	}
859 	return (0);
860 
861 src_based_fanout:
862 	hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src), (uint32_t)0);
863 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
864 	*type = OTH;
865 	return (0);
866 }
867 
868 /*
869  * mac_rx_srs_fanout
870  *
871  * This routine delivers packets destined to an SRS into a soft ring member
872  * of the set.
873  *
874  * Given a chain of packets we need to split it up into multiple sub chains
875  * destined for one of the TCP, UDP or OTH soft rings. Instead of entering
876  * the soft ring one packet at a time, we want to enter it in the form of a
877  * chain otherwise we get this start/stop behaviour where the worker thread
878  * goes to sleep and then next packets comes in forcing it to wake up etc.
879  *
880  * Note:
881  * Since we know what is the maximum fanout possible, we create a 2D array
882  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
883  * variables so that we can enter the softrings with chain. We need the
884  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
885  * for each packet would be expensive). If we ever want to have the
886  * ability to have unlimited fanout, we should probably declare a head,
887  * tail, cnt, sz with each soft ring (a data struct which contains a softring
888  * along with these members) and create an array of this uber struct so we
889  * don't have to do kmem_alloc.
890  */
891 int	fanout_oth1 = 0;
892 int	fanout_oth2 = 0;
893 int	fanout_oth3 = 0;
894 int	fanout_oth4 = 0;
895 int	fanout_oth5 = 0;
896 
897 static void
898 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
899 {
900 	struct ether_header		*ehp;
901 	struct ether_vlan_header	*evhp;
902 	uint32_t			sap;
903 	ipha_t				*ipha;
904 	uint8_t				*dstaddr;
905 	uint_t				indx;
906 	size_t				ports_offset;
907 	size_t				ipha_len;
908 	size_t				hdrsize;
909 	uint_t				hash;
910 	mblk_t				*mp;
911 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
912 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
913 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
914 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
915 	size_t				sz1;
916 	boolean_t			bw_ctl;
917 	boolean_t			hw_classified;
918 	boolean_t			dls_bypass;
919 	boolean_t			is_ether;
920 	boolean_t			is_unicast;
921 	int				fanout_cnt;
922 	enum pkt_type			type;
923 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
924 
925 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
926 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
927 
928 	/*
929 	 * If we don't have a Rx ring, S/W classification would have done
930 	 * its job and its a packet meant for us. If we were polling on
931 	 * the default ring (i.e. there was a ring assigned to this SRS),
932 	 * then we need to make sure that the mac address really belongs
933 	 * to us.
934 	 */
935 	hw_classified = mac_srs->srs_ring != NULL &&
936 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
937 
938 	/*
939 	 * Special clients (eg. VLAN, non ether, etc) need DLS
940 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
941 	 * such SRSs. Another way of disabling bypass is to set the
942 	 * MCIS_RX_BYPASS_DISABLE flag.
943 	 */
944 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
945 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
946 
947 	/*
948 	 * Since the softrings are never destroyed and we always
949 	 * create equal number of softrings for TCP, UDP and rest,
950 	 * its OK to check one of them for count and use it without
951 	 * any lock. In future, if soft rings get destroyed because
952 	 * of reduction in fanout, we will need to ensure that happens
953 	 * behind the SRS_PROC.
954 	 */
955 	fanout_cnt = mac_srs->srs_tcp_ring_count;
956 
957 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
958 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
959 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
960 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
961 
962 	/*
963 	 * We got a chain from SRS that we need to send to the soft rings.
964 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
965 	 * performance reasons), we need to separate out v4_tcp, v4_udp
966 	 * and the rest goes in other.
967 	 */
968 	while (head != NULL) {
969 		mp = head;
970 		head = head->b_next;
971 		mp->b_next = NULL;
972 
973 		type = OTH;
974 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
975 
976 		if (is_ether) {
977 			/*
978 			 * At this point we can be sure the packet at least
979 			 * has an ether header.
980 			 */
981 			if (sz1 < sizeof (struct ether_header)) {
982 				mac_rx_drop_pkt(mac_srs, mp);
983 				continue;
984 			}
985 			ehp = (struct ether_header *)mp->b_rptr;
986 
987 			/*
988 			 * Determine if this is a VLAN or non-VLAN packet.
989 			 */
990 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
991 				evhp = (struct ether_vlan_header *)mp->b_rptr;
992 				sap = ntohs(evhp->ether_type);
993 				hdrsize = sizeof (struct ether_vlan_header);
994 				/*
995 				 * Check if the VID of the packet, if any,
996 				 * belongs to this client.
997 				 */
998 				if (!mac_client_check_flow_vid(mcip,
999 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1000 					mac_rx_drop_pkt(mac_srs, mp);
1001 					continue;
1002 				}
1003 			} else {
1004 				hdrsize = sizeof (struct ether_header);
1005 			}
1006 			is_unicast =
1007 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
1008 			dstaddr = (uint8_t *)&ehp->ether_dhost;
1009 		} else {
1010 			mac_header_info_t		mhi;
1011 
1012 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
1013 			    mp, &mhi) != 0) {
1014 				mac_rx_drop_pkt(mac_srs, mp);
1015 				continue;
1016 			}
1017 			hdrsize = mhi.mhi_hdrsize;
1018 			sap = mhi.mhi_bindsap;
1019 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
1020 			dstaddr = (uint8_t *)mhi.mhi_daddr;
1021 		}
1022 
1023 		if (!dls_bypass) {
1024 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1025 			    hdrsize, &type, &indx) == -1) {
1026 				mac_rx_drop_pkt(mac_srs, mp);
1027 				continue;
1028 			}
1029 
1030 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1031 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1032 			    sz[type][indx], sz1, mp);
1033 			continue;
1034 		}
1035 
1036 
1037 		/*
1038 		 * If we are using the default Rx ring where H/W or S/W
1039 		 * classification has not happened, we need to verify if
1040 		 * this unicast packet really belongs to us.
1041 		 */
1042 		if (sap == ETHERTYPE_IP) {
1043 			/*
1044 			 * If we are H/W classified, but we have promisc
1045 			 * on, then we need to check for the unicast address.
1046 			 */
1047 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1048 				mac_address_t		*map;
1049 
1050 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1051 				map = mcip->mci_unicast;
1052 				if (bcmp(dstaddr, map->ma_addr,
1053 				    map->ma_len) == 0)
1054 					type = UNDEF;
1055 				rw_exit(&mcip->mci_rw_lock);
1056 			} else if (is_unicast) {
1057 				type = UNDEF;
1058 			}
1059 		}
1060 
1061 		/*
1062 		 * This needs to become a contract with the driver for
1063 		 * the fast path.
1064 		 */
1065 
1066 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
1067 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
1068 			type = OTH;
1069 			fanout_oth1++;
1070 		}
1071 
1072 		if (type != OTH) {
1073 			uint16_t	frag_offset_flags;
1074 
1075 			switch (ipha->ipha_protocol) {
1076 			case IPPROTO_TCP:
1077 			case IPPROTO_UDP:
1078 			case IPPROTO_SCTP:
1079 			case IPPROTO_ESP:
1080 				ipha_len = IPH_HDR_LENGTH(ipha);
1081 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
1082 				    mp->b_wptr) {
1083 					type = OTH;
1084 					break;
1085 				}
1086 				frag_offset_flags =
1087 				    ntohs(ipha->ipha_fragment_offset_and_flags);
1088 				if ((frag_offset_flags &
1089 				    (IPH_MF | IPH_OFFSET)) != 0) {
1090 					type = OTH;
1091 					fanout_oth3++;
1092 					break;
1093 				}
1094 				ports_offset = hdrsize + ipha_len;
1095 				break;
1096 			default:
1097 				type = OTH;
1098 				fanout_oth4++;
1099 				break;
1100 			}
1101 		}
1102 
1103 		if (type == OTH) {
1104 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1105 			    hdrsize, &type, &indx) == -1) {
1106 				mac_rx_drop_pkt(mac_srs, mp);
1107 				continue;
1108 			}
1109 
1110 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1111 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1112 			    sz[type][indx], sz1, mp);
1113 			continue;
1114 		}
1115 
1116 		ASSERT(type == UNDEF);
1117 
1118 		/*
1119 		 * XXX-Sunay: We should hold srs_lock since ring_count
1120 		 * below can change. But if we are always called from
1121 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
1122 		 * enforce that ring_count can't be changed i.e.
1123 		 * to change fanout type or ring count, the calling
1124 		 * thread needs to be behind SRS_PROC.
1125 		 */
1126 		switch (ipha->ipha_protocol) {
1127 		case IPPROTO_TCP:
1128 			/*
1129 			 * Note that for ESP, we fanout on SPI and it is at the
1130 			 * same offset as the 2x16-bit ports. So it is clumped
1131 			 * along with TCP, UDP and SCTP.
1132 			 */
1133 			hash = HASH_ADDR(ipha->ipha_src,
1134 			    *(uint32_t *)(mp->b_rptr + ports_offset));
1135 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1136 			type = V4_TCP;
1137 			mp->b_rptr += hdrsize;
1138 			break;
1139 		case IPPROTO_UDP:
1140 		case IPPROTO_SCTP:
1141 		case IPPROTO_ESP:
1142 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1143 				hash = HASH_ADDR(ipha->ipha_src,
1144 				    *(uint32_t *)(mp->b_rptr + ports_offset));
1145 				indx = COMPUTE_INDEX(hash,
1146 				    mac_srs->srs_udp_ring_count);
1147 			} else {
1148 				indx = mac_srs->srs_ind %
1149 				    mac_srs->srs_udp_ring_count;
1150 				mac_srs->srs_ind++;
1151 			}
1152 			type = V4_UDP;
1153 			mp->b_rptr += hdrsize;
1154 			break;
1155 		default:
1156 			indx = 0;
1157 			type = OTH;
1158 		}
1159 
1160 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
1161 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
1162 	}
1163 
1164 	for (type = V4_TCP; type < UNDEF; type++) {
1165 		int	i;
1166 
1167 		for (i = 0; i < fanout_cnt; i++) {
1168 			if (headmp[type][i] != NULL) {
1169 				mac_soft_ring_t	*softring;
1170 
1171 				ASSERT(tailmp[type][i]->b_next == NULL);
1172 				switch (type) {
1173 				case V4_TCP:
1174 					softring =
1175 					    mac_srs->srs_tcp_soft_rings[i];
1176 					break;
1177 				case V4_UDP:
1178 					softring =
1179 					    mac_srs->srs_udp_soft_rings[i];
1180 					break;
1181 				case OTH:
1182 					softring =
1183 					    mac_srs->srs_oth_soft_rings[i];
1184 					break;
1185 				}
1186 				mac_rx_soft_ring_process(mcip,
1187 				    softring, headmp[type][i], tailmp[type][i],
1188 				    cnt[type][i], sz[type][i]);
1189 			}
1190 		}
1191 	}
1192 }
1193 
1194 #define	SRS_BYTES_TO_PICKUP	150000
1195 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
1196 
1197 /*
1198  * mac_rx_srs_poll_ring
1199  *
1200  * This SRS Poll thread uses this routine to poll the underlying hardware
1201  * Rx ring to get a chain of packets. It can inline process that chain
1202  * if mac_latency_optimize is set (default) or signal the SRS worker thread
1203  * to do the remaining processing.
1204  *
1205  * Since packets come in the system via interrupt or poll path, we also
1206  * update the stats and deal with promiscous clients here.
1207  */
1208 void
1209 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
1210 {
1211 	kmutex_t 		*lock = &mac_srs->srs_lock;
1212 	kcondvar_t 		*async = &mac_srs->srs_cv;
1213 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1214 	mblk_t 			*head, *tail, *mp;
1215 	callb_cpr_t 		cprinfo;
1216 	ssize_t 		bytes_to_pickup;
1217 	size_t 			sz;
1218 	int			count;
1219 	mac_client_impl_t	*smcip;
1220 
1221 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
1222 	mutex_enter(lock);
1223 
1224 start:
1225 	for (;;) {
1226 		if (mac_srs->srs_state & SRS_PAUSE)
1227 			goto done;
1228 
1229 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
1230 		cv_wait(async, lock);
1231 		CALLB_CPR_SAFE_END(&cprinfo, lock);
1232 
1233 		if (mac_srs->srs_state & SRS_PAUSE)
1234 			goto done;
1235 
1236 check_again:
1237 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
1238 			/*
1239 			 * We pick as many bytes as we are allowed to queue.
1240 			 * Its possible that we will exceed the total
1241 			 * packets queued in case this SRS is part of the
1242 			 * Rx ring group since > 1 poll thread can be pulling
1243 			 * upto the max allowed packets at the same time
1244 			 * but that should be OK.
1245 			 */
1246 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1247 			bytes_to_pickup =
1248 			    mac_srs->srs_bw->mac_bw_drop_threshold -
1249 			    mac_srs->srs_bw->mac_bw_sz;
1250 			/*
1251 			 * We shouldn't have been signalled if we
1252 			 * have 0 or less bytes to pick but since
1253 			 * some of the bytes accounting is driver
1254 			 * dependant, we do the safety check.
1255 			 */
1256 			if (bytes_to_pickup < 0)
1257 				bytes_to_pickup = 0;
1258 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1259 		} else {
1260 			/*
1261 			 * ToDO: Need to change the polling API
1262 			 * to add a packet count and a flag which
1263 			 * tells the driver whether we want packets
1264 			 * based on a count, or bytes, or all the
1265 			 * packets queued in the driver/HW. This
1266 			 * way, we never have to check the limits
1267 			 * on poll path. We truly let only as many
1268 			 * packets enter the system as we are willing
1269 			 * to process or queue.
1270 			 *
1271 			 * Something along the lines of
1272 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
1273 			 *	mac_srs->srs_poll_pkt_cnt
1274 			 */
1275 
1276 			/*
1277 			 * Since we are not doing B/W control, pick
1278 			 * as many packets as allowed.
1279 			 */
1280 			bytes_to_pickup = max_bytes_to_pickup;
1281 		}
1282 
1283 		/* Poll the underlying Hardware */
1284 		mutex_exit(lock);
1285 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
1286 		mutex_enter(lock);
1287 
1288 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
1289 		    SRS_POLL_THR_OWNER);
1290 
1291 		mp = tail = head;
1292 		count = 0;
1293 		sz = 0;
1294 		while (mp != NULL) {
1295 			tail = mp;
1296 			sz += msgdsize(mp);
1297 			mp = mp->b_next;
1298 			count++;
1299 		}
1300 
1301 		if (head != NULL) {
1302 			tail->b_next = NULL;
1303 			smcip = mac_srs->srs_mcip;
1304 
1305 			if ((mac_srs->srs_type & SRST_FLOW) ||
1306 			    (smcip == NULL)) {
1307 				FLOW_STAT_UPDATE(mac_srs->srs_flent,
1308 				    rbytes, sz);
1309 				FLOW_STAT_UPDATE(mac_srs->srs_flent,
1310 				    ipackets, count);
1311 			}
1312 
1313 			/*
1314 			 * If there are any promiscuous mode callbacks
1315 			 * defined for this MAC client, pass them a copy
1316 			 * if appropriate and also update the counters.
1317 			 */
1318 			if (smcip != NULL) {
1319 				smcip->mci_stat_ibytes += sz;
1320 				smcip->mci_stat_ipackets += count;
1321 
1322 				if (smcip->mci_mip->mi_promisc_list != NULL) {
1323 					mutex_exit(lock);
1324 					mac_promisc_dispatch(smcip->mci_mip,
1325 					    head, NULL);
1326 					mutex_enter(lock);
1327 				}
1328 			}
1329 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
1330 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1331 				mac_srs->srs_bw->mac_bw_polled += sz;
1332 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1333 			}
1334 			srs_rx->sr_poll_count += count;
1335 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
1336 			    count, sz);
1337 			if (count <= 10)
1338 				srs_rx->sr_chain_cnt_undr10++;
1339 			else if (count > 10 && count <= 50)
1340 				srs_rx->sr_chain_cnt_10to50++;
1341 			else
1342 				srs_rx->sr_chain_cnt_over50++;
1343 		}
1344 
1345 		/*
1346 		 * We are guaranteed that SRS_PROC will be set if we
1347 		 * are here. Also, poll thread gets to run only if
1348 		 * the drain was being done by a worker thread although
1349 		 * its possible that worker thread is still running
1350 		 * and poll thread was sent down to keep the pipeline
1351 		 * going instead of doing a complete drain and then
1352 		 * trying to poll the NIC.
1353 		 *
1354 		 * So we need to check SRS_WORKER flag to make sure
1355 		 * that the worker thread is not processing the queue
1356 		 * in parallel to us. The flags and conditions are
1357 		 * protected by the srs_lock to prevent any race. We
1358 		 * ensure that we don't drop the srs_lock from now
1359 		 * till the end and similarly we don't drop the srs_lock
1360 		 * in mac_rx_srs_drain() till similar condition check
1361 		 * are complete. The mac_rx_srs_drain() needs to ensure
1362 		 * that SRS_WORKER flag remains set as long as its
1363 		 * processing the queue.
1364 		 */
1365 		if (!(mac_srs->srs_state & SRS_WORKER) &&
1366 		    (mac_srs->srs_first != NULL)) {
1367 			/*
1368 			 * We have packets to process and worker thread
1369 			 * is not running. Check to see if poll thread is
1370 			 * allowed to process.
1371 			 */
1372 			if (mac_srs->srs_state & SRS_LATENCY_OPT) {
1373 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
1374 				if (!(mac_srs->srs_state & SRS_PAUSE) &&
1375 				    srs_rx->sr_poll_pkt_cnt <=
1376 				    srs_rx->sr_lowat) {
1377 					srs_rx->sr_poll_again++;
1378 					goto check_again;
1379 				}
1380 				/*
1381 				 * We are already above low water mark
1382 				 * so stay in the polling mode but no
1383 				 * need to poll. Once we dip below
1384 				 * the polling threshold, the processing
1385 				 * thread (soft ring) will signal us
1386 				 * to poll again (MAC_UPDATE_SRS_COUNT)
1387 				 */
1388 				srs_rx->sr_poll_drain_no_poll++;
1389 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1390 				/*
1391 				 * In B/W control case, its possible
1392 				 * that the backlog built up due to
1393 				 * B/W limit being reached and packets
1394 				 * are queued only in SRS. In this case,
1395 				 * we should schedule worker thread
1396 				 * since no one else will wake us up.
1397 				 */
1398 				if ((mac_srs->srs_type & SRST_BW_CONTROL) &&
1399 				    (mac_srs->srs_tid == NULL)) {
1400 					mac_srs->srs_tid =
1401 					    timeout(mac_srs_fire, mac_srs, 1);
1402 					srs_rx->sr_poll_worker_wakeup++;
1403 				}
1404 			} else {
1405 				/*
1406 				 * Wakeup the worker thread for more processing.
1407 				 * We optimize for throughput in this case.
1408 				 */
1409 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1410 				MAC_SRS_WORKER_WAKEUP(mac_srs);
1411 				srs_rx->sr_poll_sig_worker++;
1412 			}
1413 		} else if ((mac_srs->srs_first == NULL) &&
1414 		    !(mac_srs->srs_state & SRS_WORKER)) {
1415 			/*
1416 			 * There is nothing queued in SRS and
1417 			 * no worker thread running. Plus we
1418 			 * didn't get anything from the H/W
1419 			 * as well (head == NULL);
1420 			 */
1421 			ASSERT(head == NULL);
1422 			mac_srs->srs_state &=
1423 			    ~(SRS_PROC|SRS_GET_PKTS);
1424 
1425 			/*
1426 			 * If we have a packets in soft ring, don't allow
1427 			 * more packets to come into this SRS by keeping the
1428 			 * interrupts off but not polling the H/W. The
1429 			 * poll thread will get signaled as soon as
1430 			 * srs_poll_pkt_cnt dips below poll threshold.
1431 			 */
1432 			if (srs_rx->sr_poll_pkt_cnt == 0) {
1433 				srs_rx->sr_poll_intr_enable++;
1434 				MAC_SRS_POLLING_OFF(mac_srs);
1435 			} else {
1436 				/*
1437 				 * We know nothing is queued in SRS
1438 				 * since we are here after checking
1439 				 * srs_first is NULL. The backlog
1440 				 * is entirely due to packets queued
1441 				 * in Soft ring which will wake us up
1442 				 * and get the interface out of polling
1443 				 * mode once the backlog dips below
1444 				 * sr_poll_thres.
1445 				 */
1446 				srs_rx->sr_poll_no_poll++;
1447 			}
1448 		} else {
1449 			/*
1450 			 * Worker thread is already running.
1451 			 * Nothing much to do. If the polling
1452 			 * was enabled, worker thread will deal
1453 			 * with that.
1454 			 */
1455 			mac_srs->srs_state &= ~SRS_GET_PKTS;
1456 			srs_rx->sr_poll_goto_sleep++;
1457 		}
1458 	}
1459 done:
1460 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
1461 	cv_signal(&mac_srs->srs_async);
1462 	/*
1463 	 * If this is a temporary quiesce then wait for the restart signal
1464 	 * from the srs worker. Then clear the flags and signal the srs worker
1465 	 * to ensure a positive handshake and go back to start.
1466 	 */
1467 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
1468 		cv_wait(async, lock);
1469 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
1470 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
1471 		mac_srs->srs_state &=
1472 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
1473 		cv_signal(&mac_srs->srs_async);
1474 		goto start;
1475 	} else {
1476 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
1477 		cv_signal(&mac_srs->srs_async);
1478 		CALLB_CPR_EXIT(&cprinfo);
1479 		thread_exit();
1480 	}
1481 }
1482 
1483 /*
1484  * mac_srs_pick_chain
1485  *
1486  * In Bandwidth control case, checks how many packets can be processed
1487  * and return them in a sub chain.
1488  */
1489 static mblk_t *
1490 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
1491     size_t *chain_sz, int *chain_cnt)
1492 {
1493 	mblk_t 			*head = NULL;
1494 	mblk_t 			*tail = NULL;
1495 	size_t			sz;
1496 	size_t 			tsz = 0;
1497 	int			cnt = 0;
1498 	mblk_t 			*mp;
1499 
1500 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1501 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1502 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
1503 	    mac_srs->srs_bw->mac_bw_limit) ||
1504 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
1505 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1506 		head = mac_srs->srs_first;
1507 		mac_srs->srs_first = NULL;
1508 		*chain_tail = mac_srs->srs_last;
1509 		mac_srs->srs_last = NULL;
1510 		*chain_sz = mac_srs->srs_size;
1511 		*chain_cnt = mac_srs->srs_count;
1512 		mac_srs->srs_count = 0;
1513 		mac_srs->srs_size = 0;
1514 		return (head);
1515 	}
1516 
1517 	/*
1518 	 * Can't clear the entire backlog.
1519 	 * Need to find how many packets to pick
1520 	 */
1521 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
1522 	while ((mp = mac_srs->srs_first) != NULL) {
1523 		sz = msgdsize(mp);
1524 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
1525 		    mac_srs->srs_bw->mac_bw_limit) {
1526 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
1527 				mac_srs->srs_bw->mac_bw_state |=
1528 				    SRS_BW_ENFORCED;
1529 			break;
1530 		}
1531 
1532 		/*
1533 		 * The _size & cnt is  decremented from the softrings
1534 		 * when they send up the packet for polling to work
1535 		 * properly.
1536 		 */
1537 		tsz += sz;
1538 		cnt++;
1539 		mac_srs->srs_count--;
1540 		mac_srs->srs_size -= sz;
1541 		if (tail != NULL)
1542 			tail->b_next = mp;
1543 		else
1544 			head = mp;
1545 		tail = mp;
1546 		mac_srs->srs_first = mac_srs->srs_first->b_next;
1547 	}
1548 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1549 	if (mac_srs->srs_first == NULL)
1550 		mac_srs->srs_last = NULL;
1551 
1552 	if (tail != NULL)
1553 		tail->b_next = NULL;
1554 	*chain_tail = tail;
1555 	*chain_cnt = cnt;
1556 	*chain_sz = tsz;
1557 
1558 	return (head);
1559 }
1560 
1561 /*
1562  * mac_rx_srs_drain
1563  *
1564  * The SRS drain routine. Gets to run to clear the queue. Any thread
1565  * (worker, interrupt, poll) can call this based on processing model.
1566  * The first thing we do is disable interrupts if possible and then
1567  * drain the queue. we also try to poll the underlying hardware if
1568  * there is a dedicated hardware Rx ring assigned to this SRS.
1569  *
1570  * There is a equivalent drain routine in bandwidth control mode
1571  * mac_rx_srs_drain_bw. There is some code duplication between the two
1572  * routines but they are highly performance sensitive and are easier
1573  * to read/debug if they stay separate. Any code changes here might
1574  * also apply to mac_rx_srs_drain_bw as well.
1575  */
1576 void
1577 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1578 {
1579 	mblk_t 			*head;
1580 	mblk_t			*tail;
1581 	timeout_id_t 		tid;
1582 	int			cnt = 0;
1583 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1584 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1585 
1586 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1587 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
1588 
1589 	/* If we are blanked i.e. can't do upcalls, then we are done */
1590 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1591 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1592 		    (mac_srs->srs_state & SRS_PAUSE));
1593 		goto out;
1594 	}
1595 
1596 	if (mac_srs->srs_first == NULL)
1597 		goto out;
1598 
1599 	if (!(mac_srs->srs_state & SRS_LATENCY_OPT) &&
1600 	    (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)) {
1601 		/*
1602 		 * In the normal case, the SRS worker thread does no
1603 		 * work and we wait for a backlog to build up before
1604 		 * we switch into polling mode. In case we are
1605 		 * optimizing for throughput, we use the worker thread
1606 		 * as well. The goal is to let worker thread process
1607 		 * the queue and poll thread to feed packets into
1608 		 * the queue. As such, we should signal the poll
1609 		 * thread to try and get more packets.
1610 		 *
1611 		 * We could have pulled this check in the POLL_RING
1612 		 * macro itself but keeping it explicit here makes
1613 		 * the architecture more human understandable.
1614 		 */
1615 		MAC_SRS_POLL_RING(mac_srs);
1616 	}
1617 
1618 again:
1619 	head = mac_srs->srs_first;
1620 	mac_srs->srs_first = NULL;
1621 	tail = mac_srs->srs_last;
1622 	mac_srs->srs_last = NULL;
1623 	cnt = mac_srs->srs_count;
1624 	mac_srs->srs_count = 0;
1625 
1626 	ASSERT(head != NULL);
1627 	ASSERT(tail != NULL);
1628 
1629 	if ((tid = mac_srs->srs_tid) != 0)
1630 		mac_srs->srs_tid = 0;
1631 
1632 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1633 
1634 
1635 	/*
1636 	 * mcip is NULL for broadcast and multicast flows. The promisc
1637 	 * callbacks for broadcast and multicast packets are delivered from
1638 	 * mac_rx() and we don't need to worry about that case in this path
1639 	 */
1640 	if (mcip != NULL && mcip->mci_promisc_list != NULL) {
1641 		mutex_exit(&mac_srs->srs_lock);
1642 		mac_promisc_client_dispatch(mcip, head);
1643 		mutex_enter(&mac_srs->srs_lock);
1644 	}
1645 
1646 	/*
1647 	 * Check if SRS itself is doing the processing
1648 	 * This direct path does not apply when subflows are present. In this
1649 	 * case, packets need to be dispatched to a soft ring according to the
1650 	 * flow's bandwidth and other resources contraints.
1651 	 */
1652 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1653 		mac_direct_rx_t		proc;
1654 		void			*arg1;
1655 		mac_resource_handle_t	arg2;
1656 
1657 		/*
1658 		 * This is the case when a Rx is directly
1659 		 * assigned and we have a fully classified
1660 		 * protocol chain. We can deal with it in
1661 		 * one shot.
1662 		 */
1663 		proc = srs_rx->sr_func;
1664 		arg1 = srs_rx->sr_arg1;
1665 		arg2 = srs_rx->sr_arg2;
1666 
1667 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1668 		mutex_exit(&mac_srs->srs_lock);
1669 		if (tid != 0) {
1670 			(void) untimeout(tid);
1671 			tid = 0;
1672 		}
1673 
1674 		proc(arg1, arg2, head, NULL);
1675 		/*
1676 		 * Decrement the size and count here itelf
1677 		 * since the packet has been processed.
1678 		 */
1679 		mutex_enter(&mac_srs->srs_lock);
1680 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1681 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1682 			cv_signal(&mac_srs->srs_client_cv);
1683 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1684 	} else {
1685 		/* Some kind of softrings based fanout is required */
1686 		mutex_exit(&mac_srs->srs_lock);
1687 		if (tid != 0) {
1688 			(void) untimeout(tid);
1689 			tid = 0;
1690 		}
1691 
1692 		/*
1693 		 * Since the fanout routines can deal with chains,
1694 		 * shoot the entire chain up.
1695 		 */
1696 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1697 			mac_rx_srs_fanout(mac_srs, head);
1698 		else
1699 			mac_rx_srs_proto_fanout(mac_srs, head);
1700 		mutex_enter(&mac_srs->srs_lock);
1701 	}
1702 
1703 	if (!(mac_srs->srs_state & (SRS_BLANK|SRS_PAUSE)) &&
1704 	    (mac_srs->srs_first != NULL)) {
1705 		/*
1706 		 * More packets arrived while we were clearing the
1707 		 * SRS. This can be possible because of one of
1708 		 * three conditions below:
1709 		 * 1) The driver is using multiple worker threads
1710 		 *    to send the packets to us.
1711 		 * 2) The driver has a race in switching
1712 		 *    between interrupt and polling mode or
1713 		 * 3) Packets are arriving in this SRS via the
1714 		 *    S/W classification as well.
1715 		 *
1716 		 * We should switch to polling mode and see if we
1717 		 * need to send the poll thread down. Also, signal
1718 		 * the worker thread to process whats just arrived.
1719 		 */
1720 		MAC_SRS_POLLING_ON(mac_srs);
1721 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat) {
1722 			srs_rx->sr_drain_poll_sig++;
1723 			MAC_SRS_POLL_RING(mac_srs);
1724 		}
1725 
1726 		/*
1727 		 * If we didn't signal the poll thread, we need
1728 		 * to deal with the pending packets ourselves.
1729 		 */
1730 		if (proc_type == SRS_WORKER) {
1731 			srs_rx->sr_drain_again++;
1732 			goto again;
1733 		} else {
1734 			srs_rx->sr_drain_worker_sig++;
1735 			cv_signal(&mac_srs->srs_async);
1736 		}
1737 	}
1738 
1739 out:
1740 	if (mac_srs->srs_state & SRS_GET_PKTS) {
1741 		/*
1742 		 * Poll thread is already running. Leave the
1743 		 * SRS_RPOC set and hand over the control to
1744 		 * poll thread.
1745 		 */
1746 		mac_srs->srs_state &= ~proc_type;
1747 		srs_rx->sr_drain_poll_running++;
1748 		return;
1749 	}
1750 
1751 	/*
1752 	 * Even if there are no packets queued in SRS, we
1753 	 * need to make sure that the shared counter is
1754 	 * clear and any associated softrings have cleared
1755 	 * all the backlog. Otherwise, leave the interface
1756 	 * in polling mode and the poll thread will get
1757 	 * signalled once the count goes down to zero.
1758 	 *
1759 	 * If someone is already draining the queue (SRS_PROC is
1760 	 * set) when the srs_poll_pkt_cnt goes down to zero,
1761 	 * then it means that drain is already running and we
1762 	 * will turn off polling at that time if there is
1763 	 * no backlog.
1764 	 *
1765 	 * As long as there are packets queued either
1766 	 * in soft ring set or its soft rings, we will leave
1767 	 * the interface in polling mode (even if the drain
1768 	 * was done being the interrupt thread). We signal
1769 	 * the poll thread as well if we have dipped below
1770 	 * low water mark.
1771 	 *
1772 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
1773 	 * since that turn polling on only for worker thread.
1774 	 * Its not worth turning polling on for interrupt
1775 	 * thread (since NIC will not issue another interrupt)
1776 	 * unless a backlog builds up.
1777 	 */
1778 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
1779 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
1780 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1781 		srs_rx->sr_drain_keep_polling++;
1782 		MAC_SRS_POLLING_ON(mac_srs);
1783 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1784 			MAC_SRS_POLL_RING(mac_srs);
1785 		return;
1786 	}
1787 
1788 	/* Nothing else to do. Get out of poll mode */
1789 	MAC_SRS_POLLING_OFF(mac_srs);
1790 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1791 	srs_rx->sr_drain_finish_intr++;
1792 }
1793 
1794 /*
1795  * mac_rx_srs_drain_bw
1796  *
1797  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
1798  * (worker, interrupt, poll) can call this based on processing model.
1799  * The first thing we do is disable interrupts if possible and then
1800  * drain the queue. we also try to poll the underlying hardware if
1801  * there is a dedicated hardware Rx ring assigned to this SRS.
1802  *
1803  * There is a equivalent drain routine in non bandwidth control mode
1804  * mac_rx_srs_drain. There is some code duplication between the two
1805  * routines but they are highly performance sensitive and are easier
1806  * to read/debug if they stay separate. Any code changes here might
1807  * also apply to mac_rx_srs_drain as well.
1808  */
1809 void
1810 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1811 {
1812 	mblk_t 			*head;
1813 	mblk_t			*tail;
1814 	timeout_id_t 		tid;
1815 	size_t			sz = 0;
1816 	int			cnt = 0;
1817 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1818 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1819 	clock_t			now;
1820 
1821 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1822 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
1823 again:
1824 	/* Check if we are doing B/W control */
1825 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1826 	now = ddi_get_lbolt();
1827 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
1828 		mac_srs->srs_bw->mac_bw_curr_time = now;
1829 		mac_srs->srs_bw->mac_bw_used = 0;
1830 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
1831 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
1832 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
1833 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1834 		goto done;
1835 	} else if (mac_srs->srs_bw->mac_bw_used >
1836 	    mac_srs->srs_bw->mac_bw_limit) {
1837 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
1838 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1839 		goto done;
1840 	}
1841 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1842 
1843 	/* If we are blanked i.e. can't do upcalls, then we are done */
1844 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1845 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1846 		    (mac_srs->srs_state & SRS_PAUSE));
1847 		goto done;
1848 	}
1849 
1850 	sz = 0;
1851 	cnt = 0;
1852 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
1853 		/*
1854 		 * We couldn't pick up a single packet.
1855 		 */
1856 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1857 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
1858 		    (mac_srs->srs_size != 0) &&
1859 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1860 			/*
1861 			 * Seems like configured B/W doesn't
1862 			 * even allow processing of 1 packet
1863 			 * per tick.
1864 			 *
1865 			 * XXX: raise the limit to processing
1866 			 * at least 1 packet per tick.
1867 			 */
1868 			mac_srs->srs_bw->mac_bw_limit +=
1869 			    mac_srs->srs_bw->mac_bw_limit;
1870 			mac_srs->srs_bw->mac_bw_drop_threshold +=
1871 			    mac_srs->srs_bw->mac_bw_drop_threshold;
1872 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
1873 			    "raised B/W limit to %d since not even a "
1874 			    "single packet can be processed per "
1875 			    "tick %d\n", (void *)mac_srs,
1876 			    (int)mac_srs->srs_bw->mac_bw_limit,
1877 			    (int)msgdsize(mac_srs->srs_first));
1878 		}
1879 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1880 		goto done;
1881 	}
1882 
1883 	ASSERT(head != NULL);
1884 	ASSERT(tail != NULL);
1885 
1886 	/* zero bandwidth: drop all and return to interrupt mode */
1887 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1888 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
1889 		srs_rx->sr_drop_count += cnt;
1890 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
1891 		mac_srs->srs_bw->mac_bw_sz -= sz;
1892 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
1893 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1894 		mac_pkt_drop(NULL, NULL, head, B_FALSE);
1895 		goto leave_poll;
1896 	} else {
1897 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1898 	}
1899 
1900 	if ((tid = mac_srs->srs_tid) != 0)
1901 		mac_srs->srs_tid = 0;
1902 
1903 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1904 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
1905 
1906 	/*
1907 	 * mcip is NULL for broadcast and multicast flows. The promisc
1908 	 * callbacks for broadcast and multicast packets are delivered from
1909 	 * mac_rx() and we don't need to worry about that case in this path
1910 	 */
1911 	if (mcip != NULL && mcip->mci_promisc_list != NULL) {
1912 		mutex_exit(&mac_srs->srs_lock);
1913 		mac_promisc_client_dispatch(mcip, head);
1914 		mutex_enter(&mac_srs->srs_lock);
1915 	}
1916 
1917 	/*
1918 	 * Check if SRS itself is doing the processing
1919 	 * This direct path does not apply when subflows are present. In this
1920 	 * case, packets need to be dispatched to a soft ring according to the
1921 	 * flow's bandwidth and other resources contraints.
1922 	 */
1923 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1924 		mac_direct_rx_t		proc;
1925 		void			*arg1;
1926 		mac_resource_handle_t	arg2;
1927 
1928 		/*
1929 		 * This is the case when a Rx is directly
1930 		 * assigned and we have a fully classified
1931 		 * protocol chain. We can deal with it in
1932 		 * one shot.
1933 		 */
1934 		proc = srs_rx->sr_func;
1935 		arg1 = srs_rx->sr_arg1;
1936 		arg2 = srs_rx->sr_arg2;
1937 
1938 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1939 		mutex_exit(&mac_srs->srs_lock);
1940 		if (tid != 0) {
1941 			(void) untimeout(tid);
1942 			tid = 0;
1943 		}
1944 
1945 		proc(arg1, arg2, head, NULL);
1946 		/*
1947 		 * Decrement the size and count here itelf
1948 		 * since the packet has been processed.
1949 		 */
1950 		mutex_enter(&mac_srs->srs_lock);
1951 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1952 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
1953 
1954 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1955 			cv_signal(&mac_srs->srs_client_cv);
1956 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1957 	} else {
1958 		/* Some kind of softrings based fanout is required */
1959 		mutex_exit(&mac_srs->srs_lock);
1960 		if (tid != 0) {
1961 			(void) untimeout(tid);
1962 			tid = 0;
1963 		}
1964 
1965 		/*
1966 		 * Since the fanout routines can deal with chains,
1967 		 * shoot the entire chain up.
1968 		 */
1969 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1970 			mac_rx_srs_fanout(mac_srs, head);
1971 		else
1972 			mac_rx_srs_proto_fanout(mac_srs, head);
1973 		mutex_enter(&mac_srs->srs_lock);
1974 	}
1975 
1976 	/*
1977 	 * Send the poll thread to pick up any packets arrived
1978 	 * so far. This also serves as the last check in case
1979 	 * nothing else is queued in the SRS. The poll thread
1980 	 * is signalled only in the case the drain was done
1981 	 * by the worker thread and SRS_WORKER is set. The
1982 	 * worker thread can run in parallel as long as the
1983 	 * SRS_WORKER flag is set. We we have nothing else to
1984 	 * process, we can exit while leaving SRS_PROC set
1985 	 * which gives the poll thread control to process and
1986 	 * cleanup once it returns from the NIC.
1987 	 *
1988 	 * If we have nothing else to process, we need to
1989 	 * ensure that we keep holding the srs_lock till
1990 	 * all the checks below are done and control is
1991 	 * handed to the poll thread if it was running.
1992 	 */
1993 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1994 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1995 		if (mac_srs->srs_first != NULL) {
1996 			if (proc_type == SRS_WORKER) {
1997 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1998 				if (srs_rx->sr_poll_pkt_cnt <=
1999 				    srs_rx->sr_lowat)
2000 					MAC_SRS_POLL_RING(mac_srs);
2001 				goto again;
2002 			} else {
2003 				cv_signal(&mac_srs->srs_async);
2004 			}
2005 		}
2006 	}
2007 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2008 
2009 done:
2010 
2011 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2012 		/*
2013 		 * Poll thread is already running. Leave the
2014 		 * SRS_RPOC set and hand over the control to
2015 		 * poll thread.
2016 		 */
2017 		mac_srs->srs_state &= ~proc_type;
2018 		return;
2019 	}
2020 
2021 	/*
2022 	 * If we can't process packets because we have exceeded
2023 	 * B/W limit for this tick, just set the timeout
2024 	 * and leave.
2025 	 *
2026 	 * Even if there are no packets queued in SRS, we
2027 	 * need to make sure that the shared counter is
2028 	 * clear and any associated softrings have cleared
2029 	 * all the backlog. Otherwise, leave the interface
2030 	 * in polling mode and the poll thread will get
2031 	 * signalled once the count goes down to zero.
2032 	 *
2033 	 * If someone is already draining the queue (SRS_PROC is
2034 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2035 	 * then it means that drain is already running and we
2036 	 * will turn off polling at that time if there is
2037 	 * no backlog. As long as there are packets queued either
2038 	 * is soft ring set or its soft rings, we will leave
2039 	 * the interface in polling mode.
2040 	 */
2041 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2042 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
2043 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
2044 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
2045 		MAC_SRS_POLLING_ON(mac_srs);
2046 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2047 		if ((mac_srs->srs_first != NULL) &&
2048 		    (mac_srs->srs_tid == NULL))
2049 			mac_srs->srs_tid = timeout(mac_srs_fire,
2050 			    mac_srs, 1);
2051 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2052 		return;
2053 	}
2054 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2055 
2056 leave_poll:
2057 
2058 	/* Nothing else to do. Get out of poll mode */
2059 	MAC_SRS_POLLING_OFF(mac_srs);
2060 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2061 }
2062 
2063 /*
2064  * mac_srs_worker
2065  *
2066  * The SRS worker routine. Drains the queue when no one else is
2067  * processing it.
2068  */
2069 void
2070 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
2071 {
2072 	kmutex_t 		*lock = &mac_srs->srs_lock;
2073 	kcondvar_t 		*async = &mac_srs->srs_async;
2074 	callb_cpr_t		cprinfo;
2075 	boolean_t		bw_ctl_flag;
2076 
2077 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
2078 	mutex_enter(lock);
2079 
2080 start:
2081 	for (;;) {
2082 		bw_ctl_flag = B_FALSE;
2083 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2084 			MAC_SRS_BW_LOCK(mac_srs);
2085 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2086 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2087 				bw_ctl_flag = B_TRUE;
2088 			MAC_SRS_BW_UNLOCK(mac_srs);
2089 		}
2090 		/*
2091 		 * The SRS_BW_ENFORCED flag may change since we have dropped
2092 		 * the mac_bw_lock. However the drain function can handle both
2093 		 * a drainable SRS or a bandwidth controlled SRS, and the
2094 		 * effect of scheduling a timeout is to wakeup the worker
2095 		 * thread which in turn will call the drain function. Since
2096 		 * we release the srs_lock atomically only in the cv_wait there
2097 		 * isn't a fear of waiting for ever.
2098 		 */
2099 		while (((mac_srs->srs_state & SRS_PROC) ||
2100 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
2101 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
2102 		    !(mac_srs->srs_state & SRS_PAUSE)) {
2103 			/*
2104 			 * If we have packets queued and we are here
2105 			 * because B/W control is in place, we better
2106 			 * schedule the worker wakeup after 1 tick
2107 			 * to see if bandwidth control can be relaxed.
2108 			 */
2109 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
2110 				/*
2111 				 * We need to ensure that a timer  is already
2112 				 * scheduled or we force  schedule one for
2113 				 * later so that we can continue processing
2114 				 * after this  quanta is over.
2115 				 */
2116 				mac_srs->srs_tid = timeout(mac_srs_fire,
2117 				    mac_srs, 1);
2118 			}
2119 wait:
2120 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
2121 			cv_wait(async, lock);
2122 			CALLB_CPR_SAFE_END(&cprinfo, lock);
2123 
2124 			if (mac_srs->srs_state & SRS_PAUSE)
2125 				goto done;
2126 			if (mac_srs->srs_state & SRS_PROC)
2127 				goto wait;
2128 
2129 			if (mac_srs->srs_first != NULL &&
2130 			    mac_srs->srs_type & SRST_BW_CONTROL) {
2131 				MAC_SRS_BW_LOCK(mac_srs);
2132 				if (mac_srs->srs_bw->mac_bw_state &
2133 				    SRS_BW_ENFORCED) {
2134 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2135 				}
2136 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
2137 				    SRS_BW_ENFORCED;
2138 				MAC_SRS_BW_UNLOCK(mac_srs);
2139 			}
2140 		}
2141 
2142 		if (mac_srs->srs_state & SRS_PAUSE)
2143 			goto done;
2144 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
2145 	}
2146 done:
2147 	/*
2148 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
2149 	 * from both hard and soft classifications and waits for such threads
2150 	 * to finish before signaling the worker. So at this point the only
2151 	 * thread left that could be competing with the worker is the poll
2152 	 * thread. In the case of Tx, there shouldn't be any thread holding
2153 	 * SRS_PROC at this point.
2154 	 */
2155 	if (!(mac_srs->srs_state & SRS_PROC)) {
2156 		mac_srs->srs_state |= SRS_PROC;
2157 	} else {
2158 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
2159 		/*
2160 		 * Poll thread still owns the SRS and is still running
2161 		 */
2162 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
2163 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2164 		    SRS_POLL_THR_OWNER));
2165 	}
2166 	mac_srs_worker_quiesce(mac_srs);
2167 	/*
2168 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
2169 	 * of the quiesce operation
2170 	 */
2171 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
2172 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2173 
2174 	if (mac_srs->srs_state & SRS_RESTART) {
2175 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2176 		mac_srs_worker_restart(mac_srs);
2177 		mac_srs->srs_state &= ~SRS_PROC;
2178 		goto start;
2179 	}
2180 
2181 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
2182 		mac_srs_worker_quiesce(mac_srs);
2183 
2184 	mac_srs->srs_state &= ~SRS_PROC;
2185 	/* The macro drops the srs_lock */
2186 	CALLB_CPR_EXIT(&cprinfo);
2187 	thread_exit();
2188 }
2189 
2190 /*
2191  * mac_rx_srs_subflow_process
2192  *
2193  * Receive side routine called from interrupt path when there are
2194  * sub flows present on this SRS.
2195  */
2196 /* ARGSUSED */
2197 void
2198 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
2199     mblk_t *mp_chain, boolean_t loopback)
2200 {
2201 	flow_entry_t		*flent = NULL;
2202 	flow_entry_t		*prev_flent = NULL;
2203 	mblk_t			*mp = NULL;
2204 	mblk_t			*tail = NULL;
2205 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2206 	mac_client_impl_t	*mcip;
2207 
2208 	mcip = mac_srs->srs_mcip;
2209 	ASSERT(mcip != NULL);
2210 
2211 	/*
2212 	 * We need to determine the SRS for every packet
2213 	 * by walking the flow table, if we don't get any,
2214 	 * then we proceed using the SRS we came with.
2215 	 */
2216 	mp = tail = mp_chain;
2217 	while (mp != NULL) {
2218 
2219 		/*
2220 		 * We will increment the stats for the mactching subflow.
2221 		 * when we get the bytes/pkt count for the classified packets
2222 		 * later in mac_rx_srs_process.
2223 		 */
2224 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
2225 		    FLOW_INBOUND, &flent);
2226 
2227 		if (mp == mp_chain || flent == prev_flent) {
2228 			if (prev_flent != NULL)
2229 				FLOW_REFRELE(prev_flent);
2230 			prev_flent = flent;
2231 			flent = NULL;
2232 			tail = mp;
2233 			mp = mp->b_next;
2234 			continue;
2235 		}
2236 		tail->b_next = NULL;
2237 		/*
2238 		 * A null indicates, this is for the mac_srs itself.
2239 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
2240 		 */
2241 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2242 			mac_rx_srs_process(arg,
2243 			    (mac_resource_handle_t)mac_srs, mp_chain,
2244 			    loopback);
2245 		} else {
2246 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2247 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
2248 			FLOW_REFRELE(prev_flent);
2249 		}
2250 		prev_flent = flent;
2251 		flent = NULL;
2252 		mp_chain = mp;
2253 		tail = mp;
2254 		mp = mp->b_next;
2255 	}
2256 	/* Last chain */
2257 	ASSERT(mp_chain != NULL);
2258 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2259 		mac_rx_srs_process(arg,
2260 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
2261 	} else {
2262 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2263 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
2264 		FLOW_REFRELE(prev_flent);
2265 	}
2266 }
2267 
2268 /*
2269  * mac_rx_srs_process
2270  *
2271  * Receive side routine called from the interrupt path.
2272  *
2273  * loopback is set to force a context switch on the loopback
2274  * path between MAC clients.
2275  */
2276 /* ARGSUSED */
2277 void
2278 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
2279     boolean_t loopback)
2280 {
2281 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2282 	mblk_t			*mp, *tail, *head;
2283 	int			count = 0;
2284 	int			count1;
2285 	size_t			sz = 0;
2286 	size_t			chain_sz, sz1;
2287 	mac_bw_ctl_t		*mac_bw;
2288 	mac_client_impl_t	*smcip;
2289 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2290 
2291 	/*
2292 	 * Set the tail, count and sz. We set the sz irrespective
2293 	 * of whether we are doing B/W control or not for the
2294 	 * purpose of updating the stats.
2295 	 */
2296 	mp = tail = mp_chain;
2297 	while (mp != NULL) {
2298 		tail = mp;
2299 		count++;
2300 		sz += msgdsize(mp);
2301 		mp = mp->b_next;
2302 	}
2303 
2304 	mutex_enter(&mac_srs->srs_lock);
2305 	smcip = mac_srs->srs_mcip;
2306 
2307 	if (mac_srs->srs_type & SRST_FLOW || smcip == NULL) {
2308 		FLOW_STAT_UPDATE(mac_srs->srs_flent, rbytes, sz);
2309 		FLOW_STAT_UPDATE(mac_srs->srs_flent, ipackets, count);
2310 	}
2311 	if (smcip != NULL) {
2312 		smcip->mci_stat_ibytes += sz;
2313 		smcip->mci_stat_ipackets += count;
2314 	}
2315 
2316 	/*
2317 	 * If the SRS in already being processed; has been blanked;
2318 	 * can be processed by worker thread only; or the B/W limit
2319 	 * has been reached, then queue the chain and check if
2320 	 * worker thread needs to be awakend.
2321 	 */
2322 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
2323 		mac_bw = mac_srs->srs_bw;
2324 		ASSERT(mac_bw != NULL);
2325 		mutex_enter(&mac_bw->mac_bw_lock);
2326 		/* Count the packets and bytes via interrupt */
2327 		srs_rx->sr_intr_count += count;
2328 		mac_bw->mac_bw_intr += sz;
2329 		if (mac_bw->mac_bw_limit == 0) {
2330 			/* zero bandwidth: drop all */
2331 			srs_rx->sr_drop_count += count;
2332 			mac_bw->mac_bw_drop_bytes += sz;
2333 			mutex_exit(&mac_bw->mac_bw_lock);
2334 			mutex_exit(&mac_srs->srs_lock);
2335 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
2336 			return;
2337 		} else {
2338 			if ((mac_bw->mac_bw_sz + sz) <=
2339 			    mac_bw->mac_bw_drop_threshold) {
2340 				mutex_exit(&mac_bw->mac_bw_lock);
2341 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
2342 				    tail, count, sz);
2343 			} else {
2344 				mp = mp_chain;
2345 				chain_sz = 0;
2346 				count1 = 0;
2347 				tail = NULL;
2348 				head = NULL;
2349 				while (mp != NULL) {
2350 					sz1 = msgdsize(mp);
2351 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
2352 					    mac_bw->mac_bw_drop_threshold)
2353 						break;
2354 					chain_sz += sz1;
2355 					count1++;
2356 					tail = mp;
2357 					mp = mp->b_next;
2358 				}
2359 				mutex_exit(&mac_bw->mac_bw_lock);
2360 				if (tail != NULL) {
2361 					head = tail->b_next;
2362 					tail->b_next = NULL;
2363 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
2364 					    mp_chain, tail, count1, chain_sz);
2365 					sz -= chain_sz;
2366 					count -= count1;
2367 				} else {
2368 					/* Can't pick up any */
2369 					head = mp_chain;
2370 				}
2371 				if (head != NULL) {
2372 					/* Drop any packet over the threshold */
2373 					srs_rx->sr_drop_count += count;
2374 					mutex_enter(&mac_bw->mac_bw_lock);
2375 					mac_bw->mac_bw_drop_bytes += sz;
2376 					mutex_exit(&mac_bw->mac_bw_lock);
2377 					freemsgchain(head);
2378 				}
2379 			}
2380 			MAC_SRS_WORKER_WAKEUP(mac_srs);
2381 			mutex_exit(&mac_srs->srs_lock);
2382 			return;
2383 		}
2384 	}
2385 
2386 	/*
2387 	 * If the total number of packets queued in the SRS and
2388 	 * its associated soft rings exceeds the max allowed,
2389 	 * then drop the chain. If we are polling capable, this
2390 	 * shouldn't be happening.
2391 	 */
2392 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
2393 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
2394 		mac_bw = mac_srs->srs_bw;
2395 		srs_rx->sr_drop_count += count;
2396 		mutex_enter(&mac_bw->mac_bw_lock);
2397 		mac_bw->mac_bw_drop_bytes += sz;
2398 		mutex_exit(&mac_bw->mac_bw_lock);
2399 		freemsgchain(mp_chain);
2400 		mutex_exit(&mac_srs->srs_lock);
2401 		return;
2402 	}
2403 
2404 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
2405 	/* Count the packets entering via interrupt path */
2406 	srs_rx->sr_intr_count += count;
2407 
2408 	if (!(mac_srs->srs_state & SRS_PROC)) {
2409 		/*
2410 		 * If we are coming via loopback or if we are not
2411 		 * optimizing for latency, we should signal the
2412 		 * worker thread.
2413 		 */
2414 		if (loopback || !(mac_srs->srs_state & SRS_LATENCY_OPT)) {
2415 			/*
2416 			 * For loopback, We need to let the worker take
2417 			 * over as we don't want to continue in the same
2418 			 * thread even if we can. This could lead to stack
2419 			 * overflows and may also end up using
2420 			 * resources (cpu) incorrectly.
2421 			 */
2422 			cv_signal(&mac_srs->srs_async);
2423 		} else {
2424 			/*
2425 			 * Seems like no one is processing the SRS and
2426 			 * there is no backlog. We also inline process
2427 			 * our packet if its a single packet in non
2428 			 * latency optimized case (in latency optimized
2429 			 * case, we inline process chains of any size).
2430 			 */
2431 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
2432 		}
2433 	}
2434 	mutex_exit(&mac_srs->srs_lock);
2435 }
2436 
2437 /* TX SIDE ROUTINES (RUNTIME) */
2438 
2439 /*
2440  * mac_tx_srs_no_desc
2441  *
2442  * This routine is called by Tx single ring default mode
2443  * when Tx ring runs out of descs.
2444  */
2445 mac_tx_cookie_t
2446 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2447     uint16_t flag, mblk_t **ret_mp)
2448 {
2449 	mac_tx_cookie_t cookie = NULL;
2450 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
2451 	boolean_t wakeup_worker = B_TRUE;
2452 	uint32_t tx_mode = srs_tx->st_mode;
2453 	int cnt, sz;
2454 	mblk_t *tail;
2455 
2456 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
2457 	if (flag & MAC_DROP_ON_NO_DESC) {
2458 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2459 	} else {
2460 		if (mac_srs->srs_first != NULL)
2461 			wakeup_worker = B_FALSE;
2462 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2463 		if (flag & MAC_TX_NO_ENQUEUE) {
2464 			/*
2465 			 * If TX_QUEUED is not set, queue the
2466 			 * packet and let mac_tx_srs_drain()
2467 			 * set the TX_BLOCKED bit for the
2468 			 * reasons explained above. Otherwise,
2469 			 * return the mblks.
2470 			 */
2471 			if (wakeup_worker) {
2472 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2473 				    mp_chain, tail, cnt, sz);
2474 			} else {
2475 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
2476 				    mp_chain, ret_mp, cookie);
2477 			}
2478 		} else {
2479 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2480 			    tail, cnt, sz, cookie);
2481 		}
2482 		if (wakeup_worker)
2483 			cv_signal(&mac_srs->srs_async);
2484 	}
2485 	return (cookie);
2486 }
2487 
2488 /*
2489  * mac_tx_srs_enqueue
2490  *
2491  * This routine is called when Tx SRS is operating in either serializer
2492  * or bandwidth mode. In serializer mode, a packet will get enqueued
2493  * when a thread cannot enter SRS exclusively. In bandwidth mode,
2494  * packets gets queued if allowed byte-count limit for a tick is
2495  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
2496  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
2497  * the default mode or fanout mode. Here packets get dropped or
2498  * returned back to the caller only after hi-watermark worth of data
2499  * is queued.
2500  */
2501 static mac_tx_cookie_t
2502 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2503     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
2504 {
2505 	mac_tx_cookie_t cookie = NULL;
2506 	int cnt, sz;
2507 	mblk_t *tail;
2508 	boolean_t wakeup_worker = B_TRUE;
2509 
2510 	/*
2511 	 * Ignore fanout hint if we don't have multiple tx rings.
2512 	 */
2513 	if (!TX_MULTI_RING_MODE(mac_srs))
2514 		fanout_hint = 0;
2515 
2516 	if (mac_srs->srs_first != NULL)
2517 		wakeup_worker = B_FALSE;
2518 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2519 	if (flag & MAC_DROP_ON_NO_DESC) {
2520 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
2521 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2522 		} else {
2523 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2524 			    mp_chain, tail, cnt, sz);
2525 		}
2526 	} else if (flag & MAC_TX_NO_ENQUEUE) {
2527 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
2528 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
2529 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
2530 			    ret_mp, cookie);
2531 		} else {
2532 			mp_chain->b_prev = (mblk_t *)fanout_hint;
2533 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2534 			    mp_chain, tail, cnt, sz);
2535 		}
2536 	} else {
2537 		/*
2538 		 * If you are BW_ENFORCED, just enqueue the
2539 		 * packet. srs_worker will drain it at the
2540 		 * prescribed rate. Before enqueueing, save
2541 		 * the fanout hint.
2542 		 */
2543 		mp_chain->b_prev = (mblk_t *)fanout_hint;
2544 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2545 		    tail, cnt, sz, cookie);
2546 	}
2547 	if (wakeup_worker)
2548 		cv_signal(&mac_srs->srs_async);
2549 	return (cookie);
2550 }
2551 
2552 /*
2553  * There are five tx modes:
2554  *
2555  * 1) Default mode (SRS_TX_DEFAULT)
2556  * 2) Serialization mode (SRS_TX_SERIALIZE)
2557  * 3) Fanout mode (SRS_TX_FANOUT)
2558  * 4) Bandwdith mode (SRS_TX_BW)
2559  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
2560  *
2561  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
2562  * based on the number of Tx rings requested for an SRS and whether
2563  * bandwidth control is requested or not.
2564  *
2565  * In the default mode (i.e., no fanout/no bandwidth), the SRS acts as a
2566  * pass-thru. Packets will go directly to mac_tx_send(). When the underlying
2567  * Tx ring runs out of Tx descs, it starts queueing up packets in SRS.
2568  * When flow-control is relieved, the srs_worker drains the queued
2569  * packets and informs blocked clients to restart sending packets.
2570  *
2571  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized.
2572  *
2573  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
2574  * Tx rings. Each Tx ring will have a soft ring associated with it.
2575  * These soft rings will be hung off the Tx SRS. Queueing if it happens
2576  * due to lack of Tx desc will be in individual soft ring (and not srs)
2577  * associated with Tx ring.
2578  *
2579  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
2580  * only if bw is available. Otherwise the packets will be queued in
2581  * SRS. If fanout to multiple Tx rings is configured, the packets will
2582  * be fanned out among the soft rings associated with the Tx rings.
2583  *
2584  * Four flags are used in srs_state for indicating flow control
2585  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
2586  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
2587  * driver below.
2588  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
2589  * and flow-control pressure is applied back to clients. The clients expect
2590  * wakeup when flow-control is relieved.
2591  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
2592  * got returned back to client either due to lack of Tx descs or due to bw
2593  * control reasons. The clients expect a wakeup when condition is relieved.
2594  *
2595  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
2596  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
2597  * MAC_TX_NO_ENQUEUE
2598  * Mac clients that do not want packets to be enqueued in the mac layer set
2599  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
2600  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
2601  * behaviour of this flag is different when the Tx is running in serializer
2602  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
2603  * get dropped when Tx high watermark is reached.
2604  * There are some mac clients like vsw, aggr that want the mblks to be
2605  * returned back to clients instead of being queued in Tx SRS (or Tx soft
2606  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
2607  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
2608  * In the default and Tx fanout mode, the un-transmitted mblks will be
2609  * returned back to the clients when the driver runs out of Tx descs.
2610  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
2611  * soft ring) so that the clients can be woken up when Tx desc become
2612  * available. When running in serializer or bandwidth mode mode,
2613  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
2614  */
2615 
2616 mac_tx_func_t
2617 mac_tx_get_func(uint32_t mode)
2618 {
2619 	return (mac_tx_mode_list[mode].mac_tx_func);
2620 }
2621 
2622 /* ARGSUSED */
2623 static mac_tx_cookie_t
2624 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2625     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2626 {
2627 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2628 	boolean_t		is_subflow;
2629 	mac_tx_stats_t		stats;
2630 	mac_tx_cookie_t		cookie = NULL;
2631 
2632 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
2633 
2634 	/* Regular case with a single Tx ring */
2635 	/*
2636 	 * SRS_TX_BLOCKED is set when underlying NIC runs
2637 	 * out of Tx descs and messages start getting
2638 	 * queued. It won't get reset until
2639 	 * tx_srs_drain() completely drains out the
2640 	 * messages.
2641 	 */
2642 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2643 		/* Tx descs/resources not available */
2644 		mutex_enter(&mac_srs->srs_lock);
2645 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2646 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
2647 			    flag, ret_mp);
2648 			mutex_exit(&mac_srs->srs_lock);
2649 			return (cookie);
2650 		}
2651 		/*
2652 		 * While we were computing mblk count, the
2653 		 * flow control condition got relieved.
2654 		 * Continue with the transmission.
2655 		 */
2656 		mutex_exit(&mac_srs->srs_lock);
2657 	}
2658 
2659 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2660 
2661 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2662 	    mp_chain, (is_subflow ? &stats : NULL));
2663 
2664 	/*
2665 	 * Multiple threads could be here sending packets.
2666 	 * Under such conditions, it is not possible to
2667 	 * automically set SRS_TX_BLOCKED bit to indicate
2668 	 * out of tx desc condition. To atomically set
2669 	 * this, we queue the returned packet and do
2670 	 * the setting of SRS_TX_BLOCKED in
2671 	 * mac_tx_srs_drain().
2672 	 */
2673 	if (mp_chain != NULL) {
2674 		mutex_enter(&mac_srs->srs_lock);
2675 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
2676 		mutex_exit(&mac_srs->srs_lock);
2677 		return (cookie);
2678 	}
2679 
2680 	if (is_subflow)
2681 		FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2682 
2683 	return (NULL);
2684 }
2685 
2686 /*
2687  * mac_tx_serialize_mode
2688  *
2689  * This is an experimental mode implemented as per the request of PAE.
2690  * In this mode, all callers attempting to send a packet to the NIC
2691  * will get serialized. Only one thread at any time will access the
2692  * NIC to send the packet out.
2693  */
2694 /* ARGSUSED */
2695 static mac_tx_cookie_t
2696 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2697     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2698 {
2699 	boolean_t		is_subflow;
2700 	mac_tx_stats_t		stats;
2701 	mac_tx_cookie_t		cookie = NULL;
2702 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2703 
2704 	/* Single ring, serialize below */
2705 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
2706 	mutex_enter(&mac_srs->srs_lock);
2707 	if ((mac_srs->srs_first != NULL) ||
2708 	    (mac_srs->srs_state & SRS_PROC)) {
2709 		/*
2710 		 * In serialization mode, queue all packets until
2711 		 * TX_HIWAT is set.
2712 		 * If drop bit is set, drop if TX_HIWAT is set.
2713 		 * If no_enqueue is set, still enqueue until hiwat
2714 		 * is set and return mblks after TX_HIWAT is set.
2715 		 */
2716 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
2717 		    flag, NULL, ret_mp);
2718 		mutex_exit(&mac_srs->srs_lock);
2719 		return (cookie);
2720 	}
2721 	/*
2722 	 * No packets queued, nothing on proc and no flow
2723 	 * control condition. Fast-path, ok. Do inline
2724 	 * processing.
2725 	 */
2726 	mac_srs->srs_state |= SRS_PROC;
2727 	mutex_exit(&mac_srs->srs_lock);
2728 
2729 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2730 
2731 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2732 	    mp_chain, (is_subflow ? &stats : NULL));
2733 
2734 	mutex_enter(&mac_srs->srs_lock);
2735 	mac_srs->srs_state &= ~SRS_PROC;
2736 	if (mp_chain != NULL) {
2737 		cookie = mac_tx_srs_enqueue(mac_srs,
2738 		    mp_chain, flag, NULL, ret_mp);
2739 	}
2740 	if (mac_srs->srs_first != NULL) {
2741 		/*
2742 		 * We processed inline our packet and a new
2743 		 * packet/s got queued while we were
2744 		 * processing. Wakeup srs worker
2745 		 */
2746 		cv_signal(&mac_srs->srs_async);
2747 	}
2748 	mutex_exit(&mac_srs->srs_lock);
2749 
2750 	if (is_subflow && cookie == NULL)
2751 		FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2752 
2753 	return (cookie);
2754 }
2755 
2756 /*
2757  * mac_tx_fanout_mode
2758  *
2759  * In this mode, the SRS will have access to multiple Tx rings to send
2760  * the packet out. The fanout hint that is passed as an argument is
2761  * used to find an appropriate ring to fanout the traffic. Each Tx
2762  * ring, in turn,  will have a soft ring associated with it. If a Tx
2763  * ring runs out of Tx desc's the returned packet will be queued in
2764  * the soft ring associated with that Tx ring. The srs itself will not
2765  * queue any packets.
2766  */
2767 
2768 #define	MAC_TX_SOFT_RING_PROCESS(chain) {		       		\
2769 	index = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count),	\
2770 	softring = mac_srs->srs_oth_soft_rings[index];			\
2771 	cookie = mac_tx_soft_ring_process(softring, chain, flag, ret_mp); \
2772 	DTRACE_PROBE2(tx__fanout, uint64_t, hash, uint_t, index);	\
2773 }
2774 
2775 static mac_tx_cookie_t
2776 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2777     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2778 {
2779 	mac_soft_ring_t		*softring;
2780 	uint64_t		hash;
2781 	uint_t			index;
2782 	mac_tx_cookie_t		cookie = NULL;
2783 
2784 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT);
2785 	if (fanout_hint != 0) {
2786 		/*
2787 		 * The hint is specified by the caller, simply pass the
2788 		 * whole chain to the soft ring.
2789 		 */
2790 		hash = HASH_HINT(fanout_hint);
2791 		MAC_TX_SOFT_RING_PROCESS(mp_chain);
2792 	} else {
2793 		mblk_t *last_mp, *cur_mp, *sub_chain;
2794 		uint64_t last_hash = 0;
2795 		uint_t media = mac_srs->srs_mcip->mci_mip->mi_info.mi_media;
2796 
2797 		/*
2798 		 * Compute the hash from the contents (headers) of the
2799 		 * packets of the mblk chain. Split the chains into
2800 		 * subchains of the same conversation.
2801 		 *
2802 		 * Since there may be more than one ring used for
2803 		 * sub-chains of the same call, and since the caller
2804 		 * does not maintain per conversation state since it
2805 		 * passed a zero hint, unsent subchains will be
2806 		 * dropped.
2807 		 */
2808 
2809 		flag |= MAC_DROP_ON_NO_DESC;
2810 		ret_mp = NULL;
2811 
2812 		ASSERT(ret_mp == NULL);
2813 
2814 		sub_chain = NULL;
2815 		last_mp = NULL;
2816 
2817 		for (cur_mp = mp_chain; cur_mp != NULL;
2818 		    cur_mp = cur_mp->b_next) {
2819 			hash = mac_pkt_hash(media, cur_mp, MAC_PKT_HASH_L4,
2820 			    B_TRUE);
2821 			if (last_hash != 0 && hash != last_hash) {
2822 				/*
2823 				 * Starting a different subchain, send current
2824 				 * chain out.
2825 				 */
2826 				ASSERT(last_mp != NULL);
2827 				last_mp->b_next = NULL;
2828 				MAC_TX_SOFT_RING_PROCESS(sub_chain);
2829 				sub_chain = NULL;
2830 			}
2831 
2832 			/* add packet to subchain */
2833 			if (sub_chain == NULL)
2834 				sub_chain = cur_mp;
2835 			last_mp = cur_mp;
2836 			last_hash = hash;
2837 		}
2838 
2839 		if (sub_chain != NULL) {
2840 			/* send last subchain */
2841 			ASSERT(last_mp != NULL);
2842 			last_mp->b_next = NULL;
2843 			MAC_TX_SOFT_RING_PROCESS(sub_chain);
2844 		}
2845 
2846 		cookie = NULL;
2847 	}
2848 
2849 	return (cookie);
2850 }
2851 
2852 /*
2853  * mac_tx_bw_mode
2854  *
2855  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
2856  * only if bw is available. Otherwise the packets will be queued in
2857  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
2858  * out to a Tx rings.
2859  */
2860 static mac_tx_cookie_t
2861 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2862     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2863 {
2864 	int			cnt, sz;
2865 	mblk_t			*tail;
2866 	mac_tx_cookie_t		cookie = NULL;
2867 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2868 	clock_t			now;
2869 
2870 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
2871 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2872 	mutex_enter(&mac_srs->srs_lock);
2873 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2874 		/*
2875 		 * zero bandwidth, no traffic is sent: drop the packets,
2876 		 * or return the whole chain if the caller requests all
2877 		 * unsent packets back.
2878 		 */
2879 		if (flag & MAC_TX_NO_ENQUEUE) {
2880 			cookie = (mac_tx_cookie_t)mac_srs;
2881 			*ret_mp = mp_chain;
2882 		} else {
2883 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2884 		}
2885 		mutex_exit(&mac_srs->srs_lock);
2886 		return (cookie);
2887 	} else if ((mac_srs->srs_first != NULL) ||
2888 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2889 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2890 		    fanout_hint, ret_mp);
2891 		mutex_exit(&mac_srs->srs_lock);
2892 		return (cookie);
2893 	}
2894 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2895 	now = ddi_get_lbolt();
2896 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
2897 		mac_srs->srs_bw->mac_bw_curr_time = now;
2898 		mac_srs->srs_bw->mac_bw_used = 0;
2899 	} else if (mac_srs->srs_bw->mac_bw_used >
2900 	    mac_srs->srs_bw->mac_bw_limit) {
2901 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2902 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2903 		    mp_chain, tail, cnt, sz);
2904 		/*
2905 		 * Wakeup worker thread. Note that worker
2906 		 * thread has to be woken up so that it
2907 		 * can fire up the timer to be woken up
2908 		 * on the next tick. Also once
2909 		 * BW_ENFORCED is set, it can only be
2910 		 * reset by srs_worker thread. Until then
2911 		 * all packets will get queued up in SRS
2912 		 * and hence this this code path won't be
2913 		 * entered until BW_ENFORCED is reset.
2914 		 */
2915 		cv_signal(&mac_srs->srs_async);
2916 		mutex_exit(&mac_srs->srs_lock);
2917 		return (cookie);
2918 	}
2919 
2920 	mac_srs->srs_bw->mac_bw_used += sz;
2921 	mutex_exit(&mac_srs->srs_lock);
2922 
2923 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
2924 		mac_soft_ring_t *softring;
2925 		uint_t indx, hash;
2926 
2927 		hash = HASH_HINT(fanout_hint);
2928 		indx = COMPUTE_INDEX(hash,
2929 		    mac_srs->srs_oth_ring_count);
2930 		softring = mac_srs->srs_oth_soft_rings[indx];
2931 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
2932 		    ret_mp));
2933 	} else {
2934 		boolean_t		is_subflow;
2935 		mac_tx_stats_t		stats;
2936 
2937 		is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2938 
2939 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2940 		    mp_chain, (is_subflow ? &stats : NULL));
2941 
2942 		if (mp_chain != NULL) {
2943 			mutex_enter(&mac_srs->srs_lock);
2944 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2945 			if (mac_srs->srs_bw->mac_bw_used > sz)
2946 				mac_srs->srs_bw->mac_bw_used -= sz;
2947 			else
2948 				mac_srs->srs_bw->mac_bw_used = 0;
2949 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2950 			    fanout_hint, ret_mp);
2951 			mutex_exit(&mac_srs->srs_lock);
2952 			return (cookie);
2953 		}
2954 		if (is_subflow)
2955 			FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2956 
2957 		return (NULL);
2958 	}
2959 }
2960 
2961 /* ARGSUSED */
2962 void
2963 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2964 {
2965 	mblk_t			*head, *tail;
2966 	size_t			sz;
2967 	uint32_t		tx_mode;
2968 	uint_t			saved_pkt_count;
2969 	boolean_t		is_subflow;
2970 	mac_tx_stats_t		stats;
2971 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2972 	clock_t			now;
2973 
2974 	saved_pkt_count = 0;
2975 	ASSERT(mutex_owned(&mac_srs->srs_lock));
2976 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
2977 
2978 	mac_srs->srs_state |= SRS_PROC;
2979 
2980 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2981 	tx_mode = srs_tx->st_mode;
2982 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
2983 		if (mac_srs->srs_first != NULL) {
2984 			head = mac_srs->srs_first;
2985 			tail = mac_srs->srs_last;
2986 			saved_pkt_count = mac_srs->srs_count;
2987 			mac_srs->srs_first = NULL;
2988 			mac_srs->srs_last = NULL;
2989 			mac_srs->srs_count = 0;
2990 			mutex_exit(&mac_srs->srs_lock);
2991 
2992 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2993 			    head, &stats);
2994 
2995 			mutex_enter(&mac_srs->srs_lock);
2996 			if (head != NULL) {
2997 				/* Device out of tx desc, set block */
2998 				if (head->b_next == NULL)
2999 					VERIFY(head == tail);
3000 				tail->b_next = mac_srs->srs_first;
3001 				mac_srs->srs_first = head;
3002 				mac_srs->srs_count +=
3003 				    (saved_pkt_count - stats.ts_opackets);
3004 				if (mac_srs->srs_last == NULL)
3005 					mac_srs->srs_last = tail;
3006 				MAC_TX_SRS_BLOCK(mac_srs, head);
3007 			} else {
3008 				srs_tx->st_woken_up = B_FALSE;
3009 				if (is_subflow) {
3010 					FLOW_TX_STATS_UPDATE(
3011 					    mac_srs->srs_flent, &stats);
3012 				}
3013 			}
3014 		}
3015 	} else if (tx_mode == SRS_TX_BW) {
3016 		/*
3017 		 * We are here because the timer fired and we have some data
3018 		 * to tranmit. Also mac_tx_srs_worker should have reset
3019 		 * SRS_BW_ENFORCED flag
3020 		 */
3021 		ASSERT(!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED));
3022 		head = tail = mac_srs->srs_first;
3023 		while (mac_srs->srs_first != NULL) {
3024 			tail = mac_srs->srs_first;
3025 			tail->b_prev = NULL;
3026 			mac_srs->srs_first = tail->b_next;
3027 			if (mac_srs->srs_first == NULL)
3028 				mac_srs->srs_last = NULL;
3029 			mac_srs->srs_count--;
3030 			sz = msgdsize(tail);
3031 			mac_srs->srs_size -= sz;
3032 			saved_pkt_count++;
3033 			MAC_TX_UPDATE_BW_INFO(mac_srs, sz);
3034 
3035 			if (mac_srs->srs_bw->mac_bw_used <
3036 			    mac_srs->srs_bw->mac_bw_limit)
3037 				continue;
3038 
3039 			now = ddi_get_lbolt();
3040 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3041 				mac_srs->srs_bw->mac_bw_curr_time = now;
3042 				mac_srs->srs_bw->mac_bw_used = sz;
3043 				continue;
3044 			}
3045 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3046 			break;
3047 		}
3048 
3049 		ASSERT((head == NULL && tail == NULL) ||
3050 		    (head != NULL && tail != NULL));
3051 		if (tail != NULL) {
3052 			tail->b_next = NULL;
3053 			mutex_exit(&mac_srs->srs_lock);
3054 
3055 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3056 			    head, &stats);
3057 
3058 			mutex_enter(&mac_srs->srs_lock);
3059 			if (head != NULL) {
3060 				uint_t size_sent;
3061 
3062 				/* Device out of tx desc, set block */
3063 				if (head->b_next == NULL)
3064 					VERIFY(head == tail);
3065 				tail->b_next = mac_srs->srs_first;
3066 				mac_srs->srs_first = head;
3067 				mac_srs->srs_count +=
3068 				    (saved_pkt_count - stats.ts_opackets);
3069 				if (mac_srs->srs_last == NULL)
3070 					mac_srs->srs_last = tail;
3071 				size_sent = sz - stats.ts_obytes;
3072 				mac_srs->srs_size += size_sent;
3073 				mac_srs->srs_bw->mac_bw_sz += size_sent;
3074 				if (mac_srs->srs_bw->mac_bw_used > size_sent) {
3075 					mac_srs->srs_bw->mac_bw_used -=
3076 					    size_sent;
3077 				} else {
3078 					mac_srs->srs_bw->mac_bw_used = 0;
3079 				}
3080 				MAC_TX_SRS_BLOCK(mac_srs, head);
3081 			} else {
3082 				srs_tx->st_woken_up = B_FALSE;
3083 				if (is_subflow) {
3084 					FLOW_TX_STATS_UPDATE(
3085 					    mac_srs->srs_flent, &stats);
3086 				}
3087 			}
3088 		}
3089 	} else if (tx_mode == SRS_TX_BW_FANOUT) {
3090 		mblk_t *prev;
3091 		mac_soft_ring_t *softring;
3092 		uint64_t hint;
3093 
3094 		/*
3095 		 * We are here because the timer fired and we
3096 		 * have some quota to tranmit.
3097 		 */
3098 		prev = NULL;
3099 		head = tail = mac_srs->srs_first;
3100 		while (mac_srs->srs_first != NULL) {
3101 			tail = mac_srs->srs_first;
3102 			mac_srs->srs_first = tail->b_next;
3103 			if (mac_srs->srs_first == NULL)
3104 				mac_srs->srs_last = NULL;
3105 			mac_srs->srs_count--;
3106 			sz = msgdsize(tail);
3107 			mac_srs->srs_size -= sz;
3108 			mac_srs->srs_bw->mac_bw_used += sz;
3109 			if (prev == NULL)
3110 				hint = (ulong_t)tail->b_prev;
3111 			if (hint != (ulong_t)tail->b_prev) {
3112 				prev->b_next = NULL;
3113 				mutex_exit(&mac_srs->srs_lock);
3114 				TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3115 				head = tail;
3116 				hint = (ulong_t)tail->b_prev;
3117 				mutex_enter(&mac_srs->srs_lock);
3118 			}
3119 
3120 			prev = tail;
3121 			tail->b_prev = NULL;
3122 			if (mac_srs->srs_bw->mac_bw_used <
3123 			    mac_srs->srs_bw->mac_bw_limit)
3124 				continue;
3125 
3126 			now = ddi_get_lbolt();
3127 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3128 				mac_srs->srs_bw->mac_bw_curr_time = now;
3129 				mac_srs->srs_bw->mac_bw_used = 0;
3130 				continue;
3131 			}
3132 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3133 			break;
3134 		}
3135 		ASSERT((head == NULL && tail == NULL) ||
3136 		    (head != NULL && tail != NULL));
3137 		if (tail != NULL) {
3138 			tail->b_next = NULL;
3139 			mutex_exit(&mac_srs->srs_lock);
3140 			TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3141 			mutex_enter(&mac_srs->srs_lock);
3142 		}
3143 	}
3144 	/*
3145 	 * SRS_TX_FANOUT case not considered here because packets
3146 	 * won't be queued in the SRS for this case. Packets will
3147 	 * be sent directly to soft rings underneath and if there
3148 	 * is any queueing at all, it would be in Tx side soft
3149 	 * rings.
3150 	 */
3151 
3152 	/*
3153 	 * When srs_count becomes 0, reset SRS_TX_HIWAT and
3154 	 * SRS_TX_WAKEUP_CLIENT and wakeup registered clients.
3155 	 */
3156 	if (mac_srs->srs_count == 0 && (mac_srs->srs_state &
3157 	    (SRS_TX_HIWAT | SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED))) {
3158 		mac_tx_notify_cb_t *mtnfp;
3159 		mac_cb_t *mcb;
3160 		mac_client_impl_t *mcip = mac_srs->srs_mcip;
3161 		boolean_t wakeup_required = B_FALSE;
3162 
3163 		if (mac_srs->srs_state &
3164 		    (SRS_TX_HIWAT|SRS_TX_WAKEUP_CLIENT)) {
3165 			wakeup_required = B_TRUE;
3166 		}
3167 		mac_srs->srs_state &= ~(SRS_TX_HIWAT |
3168 		    SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED);
3169 		mutex_exit(&mac_srs->srs_lock);
3170 		if (wakeup_required) {
3171 			/* Wakeup callback registered clients */
3172 			MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
3173 			for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
3174 			    mcb = mcb->mcb_nextp) {
3175 				mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
3176 				mtnfp->mtnf_fn(mtnfp->mtnf_arg,
3177 				    (mac_tx_cookie_t)mac_srs);
3178 			}
3179 			MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
3180 			    &mcip->mci_tx_notify_cb_list);
3181 			/*
3182 			 * If the client is not the primary MAC client, then we
3183 			 * need to send the notification to the clients upper
3184 			 * MAC, i.e. mci_upper_mip.
3185 			 */
3186 			mac_tx_notify(mcip->mci_upper_mip != NULL ?
3187 			    mcip->mci_upper_mip : mcip->mci_mip);
3188 		}
3189 		mutex_enter(&mac_srs->srs_lock);
3190 	}
3191 	mac_srs->srs_state &= ~SRS_PROC;
3192 }
3193 
3194 /*
3195  * Given a packet, get the flow_entry that identifies the flow
3196  * to which that packet belongs. The flow_entry will contain
3197  * the transmit function to be used to send the packet. If the
3198  * function returns NULL, the packet should be sent using the
3199  * underlying NIC.
3200  */
3201 static flow_entry_t *
3202 mac_tx_classify(mac_impl_t *mip, mblk_t *mp)
3203 {
3204 	flow_entry_t		*flent = NULL;
3205 	mac_client_impl_t	*mcip;
3206 	int	err;
3207 
3208 	/*
3209 	 * Do classification on the packet.
3210 	 */
3211 	err = mac_flow_lookup(mip->mi_flow_tab, mp, FLOW_OUTBOUND, &flent);
3212 	if (err != 0)
3213 		return (NULL);
3214 
3215 	/*
3216 	 * This flent might just be an additional one on the MAC client,
3217 	 * i.e. for classification purposes (different fdesc), however
3218 	 * the resources, SRS et. al., are in the mci_flent, so if
3219 	 * this isn't the mci_flent, we need to get it.
3220 	 */
3221 	if ((mcip = flent->fe_mcip) != NULL && mcip->mci_flent != flent) {
3222 		FLOW_REFRELE(flent);
3223 		flent = mcip->mci_flent;
3224 		FLOW_TRY_REFHOLD(flent, err);
3225 		if (err != 0)
3226 			return (NULL);
3227 	}
3228 
3229 	return (flent);
3230 }
3231 
3232 /*
3233  * This macro is only meant to be used by mac_tx_send().
3234  */
3235 #define	CHECK_VID_AND_ADD_TAG(mp) {			\
3236 	if (vid_check) {				\
3237 		int err = 0;				\
3238 							\
3239 		MAC_VID_CHECK(src_mcip, (mp), err);	\
3240 		if (err != 0) {				\
3241 			freemsg((mp));			\
3242 			(mp) = next;			\
3243 			oerrors++;			\
3244 			continue;			\
3245 		}					\
3246 	}						\
3247 	if (add_tag) {					\
3248 		(mp) = mac_add_vlan_tag((mp), 0, vid);	\
3249 		if ((mp) == NULL) {			\
3250 			(mp) = next;			\
3251 			oerrors++;			\
3252 			continue;			\
3253 		}					\
3254 	}						\
3255 }
3256 
3257 mblk_t *
3258 mac_tx_send(mac_client_handle_t mch, mac_ring_handle_t ring, mblk_t *mp_chain,
3259     mac_tx_stats_t *stats)
3260 {
3261 	mac_client_impl_t *src_mcip = (mac_client_impl_t *)mch;
3262 	mac_impl_t *mip = src_mcip->mci_mip;
3263 	uint_t obytes = 0, opackets = 0, oerrors = 0;
3264 	mblk_t *mp = NULL, *next;
3265 	boolean_t vid_check, add_tag;
3266 	uint16_t vid = 0;
3267 
3268 	if (mip->mi_nclients > 1) {
3269 		vid_check = MAC_VID_CHECK_NEEDED(src_mcip);
3270 		add_tag = MAC_TAG_NEEDED(src_mcip);
3271 		if (add_tag)
3272 			vid = mac_client_vid(mch);
3273 	} else {
3274 		ASSERT(mip->mi_nclients == 1);
3275 		vid_check = add_tag = B_FALSE;
3276 	}
3277 
3278 	/*
3279 	 * Fastpath: if there's only one client, and there's no
3280 	 * multicast listeners, we simply send the packet down to the
3281 	 * underlying NIC.
3282 	 */
3283 	if (mip->mi_nactiveclients == 1 && mip->mi_promisc_list == NULL)  {
3284 		DTRACE_PROBE2(fastpath,
3285 		    mac_client_impl_t *, src_mcip, mblk_t *, mp_chain);
3286 
3287 		mp = mp_chain;
3288 		while (mp != NULL) {
3289 			next = mp->b_next;
3290 			mp->b_next = NULL;
3291 			opackets++;
3292 			obytes += (mp->b_cont == NULL ? MBLKL(mp) :
3293 			    msgdsize(mp));
3294 
3295 			CHECK_VID_AND_ADD_TAG(mp);
3296 			MAC_TX(mip, ring, mp,
3297 			    ((src_mcip->mci_state_flags & MCIS_SHARE_BOUND) !=
3298 			    0));
3299 
3300 			/*
3301 			 * If the driver is out of descriptors and does a
3302 			 * partial send it will return a chain of unsent
3303 			 * mblks. Adjust the accounting stats.
3304 			 */
3305 			if (mp != NULL) {
3306 				opackets--;
3307 				obytes -= msgdsize(mp);
3308 				mp->b_next = next;
3309 				break;
3310 			}
3311 			mp = next;
3312 		}
3313 		goto done;
3314 	}
3315 
3316 	/*
3317 	 * No fastpath, we either have more than one MAC client
3318 	 * defined on top of the same MAC, or one or more MAC
3319 	 * client promiscuous callbacks.
3320 	 */
3321 	DTRACE_PROBE3(slowpath, mac_client_impl_t *,
3322 	    src_mcip, int, mip->mi_nclients, mblk_t *, mp_chain);
3323 
3324 	mp = mp_chain;
3325 	while (mp != NULL) {
3326 		flow_entry_t *dst_flow_ent;
3327 		void *flow_cookie;
3328 		size_t	pkt_size;
3329 		mblk_t *mp1;
3330 
3331 		next = mp->b_next;
3332 		mp->b_next = NULL;
3333 		opackets++;
3334 		pkt_size = (mp->b_cont == NULL ? MBLKL(mp) : msgdsize(mp));
3335 		obytes += pkt_size;
3336 		CHECK_VID_AND_ADD_TAG(mp);
3337 
3338 		/*
3339 		 * Check if there are promiscuous mode callbacks defined.
3340 		 */
3341 		if (mip->mi_promisc_list != NULL)
3342 			mac_promisc_dispatch(mip, mp, src_mcip);
3343 
3344 		/*
3345 		 * Find the destination.
3346 		 */
3347 		dst_flow_ent = mac_tx_classify(mip, mp);
3348 
3349 		if (dst_flow_ent != NULL) {
3350 			size_t	hdrsize;
3351 			int	err = 0;
3352 
3353 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
3354 				struct ether_vlan_header *evhp =
3355 				    (struct ether_vlan_header *)mp->b_rptr;
3356 
3357 				if (ntohs(evhp->ether_tpid) == ETHERTYPE_VLAN)
3358 					hdrsize = sizeof (*evhp);
3359 				else
3360 					hdrsize = sizeof (struct ether_header);
3361 			} else {
3362 				mac_header_info_t	mhi;
3363 
3364 				err = mac_header_info((mac_handle_t)mip,
3365 				    mp, &mhi);
3366 				if (err == 0)
3367 					hdrsize = mhi.mhi_hdrsize;
3368 			}
3369 
3370 			/*
3371 			 * Got a matching flow. It's either another
3372 			 * MAC client, or a broadcast/multicast flow.
3373 			 * Make sure the packet size is within the
3374 			 * allowed size. If not drop the packet and
3375 			 * move to next packet.
3376 			 */
3377 			if (err != 0 ||
3378 			    (pkt_size - hdrsize) > mip->mi_sdu_max) {
3379 				oerrors++;
3380 				DTRACE_PROBE2(loopback__drop, size_t, pkt_size,
3381 				    mblk_t *, mp);
3382 				freemsg(mp);
3383 				mp = next;
3384 				FLOW_REFRELE(dst_flow_ent);
3385 				continue;
3386 			}
3387 			flow_cookie = mac_flow_get_client_cookie(dst_flow_ent);
3388 			if (flow_cookie != NULL) {
3389 				/*
3390 				 * The vnic_bcast_send function expects
3391 				 * to receive the sender MAC client
3392 				 * as value for arg2.
3393 				 */
3394 				mac_bcast_send(flow_cookie, src_mcip, mp,
3395 				    B_TRUE);
3396 			} else {
3397 				/*
3398 				 * loopback the packet to a
3399 				 * local MAC client. We force a context
3400 				 * switch if both source and destination
3401 				 * MAC clients are used by IP, i.e. bypass
3402 				 * is set.
3403 				 */
3404 				boolean_t do_switch;
3405 				mac_client_impl_t *dst_mcip =
3406 				    dst_flow_ent->fe_mcip;
3407 
3408 				do_switch = ((src_mcip->mci_state_flags &
3409 				    dst_mcip->mci_state_flags &
3410 				    MCIS_CLIENT_POLL_CAPABLE) != 0);
3411 
3412 				if ((mp1 = mac_fix_cksum(mp)) != NULL) {
3413 					(dst_flow_ent->fe_cb_fn)(
3414 					    dst_flow_ent->fe_cb_arg1,
3415 					    dst_flow_ent->fe_cb_arg2,
3416 					    mp1, do_switch);
3417 				}
3418 			}
3419 			FLOW_REFRELE(dst_flow_ent);
3420 		} else {
3421 			/*
3422 			 * Unknown destination, send via the underlying
3423 			 * NIC.
3424 			 */
3425 			MAC_TX(mip, ring, mp,
3426 			    ((src_mcip->mci_state_flags & MCIS_SHARE_BOUND) !=
3427 			    0));
3428 			if (mp != NULL) {
3429 				/*
3430 				 * Adjust for the last packet that
3431 				 * could not be transmitted
3432 				 */
3433 				opackets--;
3434 				obytes -= pkt_size;
3435 				mp->b_next = next;
3436 				break;
3437 			}
3438 		}
3439 		mp = next;
3440 	}
3441 
3442 done:
3443 	src_mcip->mci_stat_obytes += obytes;
3444 	src_mcip->mci_stat_opackets += opackets;
3445 	src_mcip->mci_stat_oerrors += oerrors;
3446 
3447 	if (stats != NULL) {
3448 		stats->ts_opackets = opackets;
3449 		stats->ts_obytes = obytes;
3450 		stats->ts_oerrors = oerrors;
3451 	}
3452 	return (mp);
3453 }
3454 
3455 /*
3456  * mac_tx_srs_ring_present
3457  *
3458  * Returns whether the specified ring is part of the specified SRS.
3459  */
3460 boolean_t
3461 mac_tx_srs_ring_present(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3462 {
3463 	int i;
3464 	mac_soft_ring_t *soft_ring;
3465 
3466 	if (srs->srs_tx.st_arg2 == tx_ring)
3467 		return (B_TRUE);
3468 
3469 	for (i = 0; i < srs->srs_oth_ring_count; i++) {
3470 		soft_ring =  srs->srs_oth_soft_rings[i];
3471 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3472 			return (B_TRUE);
3473 	}
3474 
3475 	return (B_FALSE);
3476 }
3477 
3478 /*
3479  * mac_tx_srs_wakeup
3480  *
3481  * Called when Tx desc become available. Wakeup the appropriate worker
3482  * thread after resetting the SRS_TX_BLOCKED/S_RING_BLOCK bit in the
3483  * state field.
3484  */
3485 void
3486 mac_tx_srs_wakeup(mac_soft_ring_set_t *mac_srs, mac_ring_handle_t ring)
3487 {
3488 	int i;
3489 	mac_soft_ring_t *sringp;
3490 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3491 
3492 	mutex_enter(&mac_srs->srs_lock);
3493 	if (TX_SINGLE_RING_MODE(mac_srs)) {
3494 		if (srs_tx->st_arg2 == ring &&
3495 		    mac_srs->srs_state & SRS_TX_BLOCKED) {
3496 			mac_srs->srs_state &= ~SRS_TX_BLOCKED;
3497 			srs_tx->st_unblocked_cnt++;
3498 			cv_signal(&mac_srs->srs_async);
3499 		}
3500 		/*
3501 		 * A wakeup can come before tx_srs_drain() could
3502 		 * grab srs lock and set SRS_TX_BLOCKED. So
3503 		 * always set woken_up flag when we come here.
3504 		 */
3505 		srs_tx->st_woken_up = B_TRUE;
3506 		mutex_exit(&mac_srs->srs_lock);
3507 		return;
3508 	}
3509 
3510 	/* If you are here, it is for FANOUT or BW_FANOUT case */
3511 	ASSERT(TX_MULTI_RING_MODE(mac_srs));
3512 	for (i = 0; i < mac_srs->srs_oth_ring_count; i++) {
3513 		sringp = mac_srs->srs_oth_soft_rings[i];
3514 		mutex_enter(&sringp->s_ring_lock);
3515 		if (sringp->s_ring_tx_arg2 == ring) {
3516 			if (sringp->s_ring_state & S_RING_BLOCK) {
3517 				sringp->s_ring_state &= ~S_RING_BLOCK;
3518 				sringp->s_ring_unblocked_cnt++;
3519 				cv_signal(&sringp->s_ring_async);
3520 			}
3521 			sringp->s_ring_tx_woken_up = B_TRUE;
3522 		}
3523 		mutex_exit(&sringp->s_ring_lock);
3524 	}
3525 	mutex_exit(&mac_srs->srs_lock);
3526 }
3527 
3528 /*
3529  * Once the driver is done draining, send a MAC_NOTE_TX notification to unleash
3530  * the blocked clients again.
3531  */
3532 void
3533 mac_tx_notify(mac_impl_t *mip)
3534 {
3535 	i_mac_notify(mip, MAC_NOTE_TX);
3536 }
3537 
3538 /*
3539  * RX SOFTRING RELATED FUNCTIONS
3540  *
3541  * These functions really belong in mac_soft_ring.c and here for
3542  * a short period.
3543  */
3544 
3545 #define	SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3546 	/*								\
3547 	 * Enqueue our mblk chain.					\
3548 	 */								\
3549 	ASSERT(MUTEX_HELD(&(ringp)->s_ring_lock));			\
3550 									\
3551 	if ((ringp)->s_ring_last != NULL)				\
3552 		(ringp)->s_ring_last->b_next = (mp);			\
3553 	else								\
3554 		(ringp)->s_ring_first = (mp);				\
3555 	(ringp)->s_ring_last = (tail);					\
3556 	(ringp)->s_ring_count += (cnt);					\
3557 	ASSERT((ringp)->s_ring_count > 0);				\
3558 	if ((ringp)->s_ring_type & ST_RING_BW_CTL) {			\
3559 		(ringp)->s_ring_size += sz;				\
3560 	}								\
3561 }
3562 
3563 /*
3564  * Default entry point to deliver a packet chain to a MAC client.
3565  * If the MAC client has flows, do the classification with these
3566  * flows as well.
3567  */
3568 /* ARGSUSED */
3569 void
3570 mac_rx_deliver(void *arg1, mac_resource_handle_t mrh, mblk_t *mp_chain,
3571     mac_header_info_t *arg3)
3572 {
3573 	mac_client_impl_t *mcip = arg1;
3574 
3575 	if (mcip->mci_nvids == 1 &&
3576 	    !(mcip->mci_state_flags & MCIS_STRIP_DISABLE)) {
3577 		/*
3578 		 * If the client has exactly one VID associated with it
3579 		 * and striping of VLAN header is not disabled,
3580 		 * remove the VLAN tag from the packet before
3581 		 * passing it on to the client's receive callback.
3582 		 * Note that this needs to be done after we dispatch
3583 		 * the packet to the promiscuous listeners of the
3584 		 * client, since they expect to see the whole
3585 		 * frame including the VLAN headers.
3586 		 */
3587 		mp_chain = mac_strip_vlan_tag_chain(mp_chain);
3588 	}
3589 
3590 	mcip->mci_rx_fn(mcip->mci_rx_arg, mrh, mp_chain, B_FALSE);
3591 }
3592 
3593 /*
3594  * mac_rx_soft_ring_process
3595  *
3596  * process a chain for a given soft ring. The number of packets queued
3597  * in the SRS and its associated soft rings (including this one) is
3598  * very small (tracked by srs_poll_pkt_cnt), then allow the entering
3599  * thread (interrupt or poll thread) to do inline processing. This
3600  * helps keep the latency down under low load.
3601  *
3602  * The proc and arg for each mblk is already stored in the mblk in
3603  * appropriate places.
3604  */
3605 /* ARGSUSED */
3606 void
3607 mac_rx_soft_ring_process(mac_client_impl_t *mcip, mac_soft_ring_t *ringp,
3608     mblk_t *mp_chain, mblk_t *tail, int cnt, size_t sz)
3609 {
3610 	mac_direct_rx_t		proc;
3611 	void			*arg1;
3612 	mac_resource_handle_t	arg2;
3613 	mac_soft_ring_set_t	*mac_srs = ringp->s_ring_set;
3614 
3615 	ASSERT(ringp != NULL);
3616 	ASSERT(mp_chain != NULL);
3617 	ASSERT(tail != NULL);
3618 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3619 
3620 	mutex_enter(&ringp->s_ring_lock);
3621 	ringp->s_ring_total_inpkt += cnt;
3622 	if ((mac_srs->srs_rx.sr_poll_pkt_cnt <= 1) &&
3623 	    !(ringp->s_ring_type & ST_RING_WORKER_ONLY)) {
3624 		/* If on processor or blanking on, then enqueue and return */
3625 		if (ringp->s_ring_state & S_RING_BLANK ||
3626 		    ringp->s_ring_state & S_RING_PROC) {
3627 			SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3628 			mutex_exit(&ringp->s_ring_lock);
3629 			return;
3630 		}
3631 		proc = ringp->s_ring_rx_func;
3632 		arg1 = ringp->s_ring_rx_arg1;
3633 		arg2 = ringp->s_ring_rx_arg2;
3634 		/*
3635 		 * See if anything is already queued. If we are the
3636 		 * first packet, do inline processing else queue the
3637 		 * packet and do the drain.
3638 		 */
3639 		if (ringp->s_ring_first == NULL) {
3640 			/*
3641 			 * Fast-path, ok to process and nothing queued.
3642 			 */
3643 			ringp->s_ring_run = curthread;
3644 			ringp->s_ring_state |= (S_RING_PROC);
3645 
3646 			mutex_exit(&ringp->s_ring_lock);
3647 
3648 			/*
3649 			 * We are the chain of 1 packet so
3650 			 * go through this fast path.
3651 			 */
3652 			ASSERT(mp_chain->b_next == NULL);
3653 
3654 			(*proc)(arg1, arg2, mp_chain, NULL);
3655 
3656 			ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3657 			/*
3658 			 * If we have a soft ring set which is doing
3659 			 * bandwidth control, we need to decrement
3660 			 * srs_size and count so it the SRS can have a
3661 			 * accurate idea of what is the real data
3662 			 * queued between SRS and its soft rings. We
3663 			 * decrement the counters only when the packet
3664 			 * gets processed by both SRS and the soft ring.
3665 			 */
3666 			mutex_enter(&mac_srs->srs_lock);
3667 			MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
3668 			MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
3669 			mutex_exit(&mac_srs->srs_lock);
3670 
3671 			mutex_enter(&ringp->s_ring_lock);
3672 			ringp->s_ring_run = NULL;
3673 			ringp->s_ring_state &= ~S_RING_PROC;
3674 			if (ringp->s_ring_state & S_RING_CLIENT_WAIT)
3675 				cv_signal(&ringp->s_ring_client_cv);
3676 
3677 			if ((ringp->s_ring_first == NULL) ||
3678 			    (ringp->s_ring_state & S_RING_BLANK)) {
3679 				/*
3680 				 * We processed inline our packet and
3681 				 * nothing new has arrived or our
3682 				 * receiver doesn't want to receive
3683 				 * any packets. We are done.
3684 				 */
3685 				mutex_exit(&ringp->s_ring_lock);
3686 				return;
3687 			}
3688 		} else {
3689 			SOFT_RING_ENQUEUE_CHAIN(ringp,
3690 			    mp_chain, tail, cnt, sz);
3691 		}
3692 
3693 		/*
3694 		 * We are here because either we couldn't do inline
3695 		 * processing (because something was already
3696 		 * queued), or we had a chain of more than one
3697 		 * packet, or something else arrived after we were
3698 		 * done with inline processing.
3699 		 */
3700 		ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3701 		ASSERT(ringp->s_ring_first != NULL);
3702 
3703 		ringp->s_ring_drain_func(ringp);
3704 		mutex_exit(&ringp->s_ring_lock);
3705 		return;
3706 	} else {
3707 		/* ST_RING_WORKER_ONLY case */
3708 		SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3709 		mac_soft_ring_worker_wakeup(ringp);
3710 		mutex_exit(&ringp->s_ring_lock);
3711 	}
3712 }
3713 
3714 /*
3715  * TX SOFTRING RELATED FUNCTIONS
3716  *
3717  * These functions really belong in mac_soft_ring.c and here for
3718  * a short period.
3719  */
3720 
3721 #define	TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3722 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));			\
3723 	ringp->s_ring_state |= S_RING_ENQUEUED;				\
3724 	SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);	\
3725 }
3726 
3727 /*
3728  * mac_tx_sring_queued
3729  *
3730  * When we are out of transmit descriptors and we already have a
3731  * queue that exceeds hiwat (or the client called us with
3732  * MAC_TX_NO_ENQUEUE or MAC_DROP_ON_NO_DESC flag), return the
3733  * soft ring pointer as the opaque cookie for the client enable
3734  * flow control.
3735  */
3736 static mac_tx_cookie_t
3737 mac_tx_sring_enqueue(mac_soft_ring_t *ringp, mblk_t *mp_chain, uint16_t flag,
3738     mblk_t **ret_mp)
3739 {
3740 	int cnt;
3741 	size_t sz;
3742 	mblk_t *tail;
3743 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3744 	mac_tx_cookie_t cookie = NULL;
3745 	boolean_t wakeup_worker = B_TRUE;
3746 
3747 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3748 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3749 	if (flag & MAC_DROP_ON_NO_DESC) {
3750 		mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
3751 		/* increment freed stats */
3752 		ringp->s_ring_drops += cnt;
3753 		cookie = (mac_tx_cookie_t)ringp;
3754 	} else {
3755 		if (ringp->s_ring_first != NULL)
3756 			wakeup_worker = B_FALSE;
3757 
3758 		if (flag & MAC_TX_NO_ENQUEUE) {
3759 			/*
3760 			 * If QUEUED is not set, queue the packet
3761 			 * and let mac_tx_soft_ring_drain() set
3762 			 * the TX_BLOCKED bit for the reasons
3763 			 * explained above. Otherwise, return the
3764 			 * mblks.
3765 			 */
3766 			if (wakeup_worker) {
3767 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp,
3768 				    mp_chain, tail, cnt, sz);
3769 			} else {
3770 				ringp->s_ring_state |= S_RING_WAKEUP_CLIENT;
3771 				cookie = (mac_tx_cookie_t)ringp;
3772 				*ret_mp = mp_chain;
3773 			}
3774 		} else {
3775 			boolean_t enqueue = B_TRUE;
3776 
3777 			if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3778 				/*
3779 				 * flow-controlled. Store ringp in cookie
3780 				 * so that it can be returned as
3781 				 * mac_tx_cookie_t to client
3782 				 */
3783 				ringp->s_ring_state |= S_RING_TX_HIWAT;
3784 				cookie = (mac_tx_cookie_t)ringp;
3785 				ringp->s_ring_hiwat_cnt++;
3786 				if (ringp->s_ring_count >
3787 				    ringp->s_ring_tx_max_q_cnt) {
3788 					/* increment freed stats */
3789 					ringp->s_ring_drops += cnt;
3790 					/*
3791 					 * b_prev may be set to the fanout hint
3792 					 * hence can't use freemsg directly
3793 					 */
3794 					mac_pkt_drop(NULL, NULL,
3795 					    mp_chain, B_FALSE);
3796 					DTRACE_PROBE1(tx_queued_hiwat,
3797 					    mac_soft_ring_t *, ringp);
3798 					enqueue = B_FALSE;
3799 				}
3800 			}
3801 			if (enqueue) {
3802 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain,
3803 				    tail, cnt, sz);
3804 			}
3805 		}
3806 		if (wakeup_worker)
3807 			cv_signal(&ringp->s_ring_async);
3808 	}
3809 	return (cookie);
3810 }
3811 
3812 
3813 /*
3814  * mac_tx_soft_ring_process
3815  *
3816  * This routine is called when fanning out outgoing traffic among
3817  * multipe Tx rings.
3818  * Note that a soft ring is associated with a h/w Tx ring.
3819  */
3820 mac_tx_cookie_t
3821 mac_tx_soft_ring_process(mac_soft_ring_t *ringp, mblk_t *mp_chain,
3822     uint16_t flag, mblk_t **ret_mp)
3823 {
3824 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3825 	int	cnt;
3826 	size_t	sz;
3827 	mblk_t	*tail;
3828 	mac_tx_cookie_t cookie = NULL;
3829 
3830 	ASSERT(ringp != NULL);
3831 	ASSERT(mp_chain != NULL);
3832 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3833 	/*
3834 	 * Only two modes can come here; either it can be
3835 	 * SRS_TX_BW_FANOUT or SRS_TX_FANOUT
3836 	 */
3837 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3838 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
3839 
3840 	if (ringp->s_ring_type & ST_RING_WORKER_ONLY) {
3841 		/* Serialization mode */
3842 
3843 		mutex_enter(&ringp->s_ring_lock);
3844 		if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3845 			cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3846 			    flag, ret_mp);
3847 			mutex_exit(&ringp->s_ring_lock);
3848 			return (cookie);
3849 		}
3850 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3851 		TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3852 		if (ringp->s_ring_state & (S_RING_BLOCK | S_RING_PROC)) {
3853 			/*
3854 			 * If ring is blocked due to lack of Tx
3855 			 * descs, just return. Worker thread
3856 			 * will get scheduled when Tx desc's
3857 			 * become available.
3858 			 */
3859 			mutex_exit(&ringp->s_ring_lock);
3860 			return (cookie);
3861 		}
3862 		mac_soft_ring_worker_wakeup(ringp);
3863 		mutex_exit(&ringp->s_ring_lock);
3864 		return (cookie);
3865 	} else {
3866 		/* Default fanout mode */
3867 		/*
3868 		 * S_RING_BLOCKED is set when underlying NIC runs
3869 		 * out of Tx descs and messages start getting
3870 		 * queued. It won't get reset until
3871 		 * tx_srs_drain() completely drains out the
3872 		 * messages.
3873 		 */
3874 		boolean_t		is_subflow;
3875 		mac_tx_stats_t		stats;
3876 
3877 		if (ringp->s_ring_state & S_RING_ENQUEUED) {
3878 			/* Tx descs/resources not available */
3879 			mutex_enter(&ringp->s_ring_lock);
3880 			if (ringp->s_ring_state & S_RING_ENQUEUED) {
3881 				cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3882 				    flag, ret_mp);
3883 				mutex_exit(&ringp->s_ring_lock);
3884 				return (cookie);
3885 			}
3886 			/*
3887 			 * While we were computing mblk count, the
3888 			 * flow control condition got relieved.
3889 			 * Continue with the transmission.
3890 			 */
3891 			mutex_exit(&ringp->s_ring_lock);
3892 		}
3893 		is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
3894 
3895 		mp_chain = mac_tx_send(ringp->s_ring_tx_arg1,
3896 		    ringp->s_ring_tx_arg2, mp_chain,
3897 		    (is_subflow ? &stats : NULL));
3898 
3899 		/*
3900 		 * Multiple threads could be here sending packets.
3901 		 * Under such conditions, it is not possible to
3902 		 * automically set S_RING_BLOCKED bit to indicate
3903 		 * out of tx desc condition. To atomically set
3904 		 * this, we queue the returned packet and do
3905 		 * the setting of S_RING_BLOCKED in
3906 		 * mac_tx_soft_ring_drain().
3907 		 */
3908 		if (mp_chain != NULL) {
3909 			mutex_enter(&ringp->s_ring_lock);
3910 			cookie =
3911 			    mac_tx_sring_enqueue(ringp, mp_chain, flag, ret_mp);
3912 			mutex_exit(&ringp->s_ring_lock);
3913 			return (cookie);
3914 		}
3915 		if (is_subflow) {
3916 			FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
3917 		}
3918 		return (NULL);
3919 	}
3920 }
3921