xref: /titanic_41/usr/src/uts/common/io/mac/mac_sched.c (revision 0f1702c5201310f0529cd5abb77652e5e9b241b6)
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 2008 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 	if ((mac_srs)->srs_bw->mac_bw_curr_time != lbolt) {    		\
278 		(mac_srs)->srs_bw->mac_bw_curr_time = lbolt;   		\
279 		(mac_srs)->srs_bw->mac_bw_used = 0;	       		\
280 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
281 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
282 	}								\
283 }
284 
285 /*
286  * MAC_SRS_WORKER_WAKEUP
287  *
288  * Wake up the SRS worker thread to process the queue as long as
289  * no one else is processing the queue. If we are optimizing for
290  * latency, we wake up the worker thread immediately or else we
291  * wait mac_srs_worker_wakeup_ticks before worker thread gets
292  * woken up.
293  */
294 int mac_srs_worker_wakeup_ticks = 0;
295 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
296 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
297 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
298 		(mac_srs)->srs_tid == NULL) {				\
299 		if (mac_latency_optimize ||				\
300 			(mac_srs_worker_wakeup_ticks == 0))		\
301 			cv_signal(&(mac_srs)->srs_async);		\
302 		else							\
303 			(mac_srs)->srs_tid =				\
304 				timeout(mac_srs_fire, (mac_srs),	\
305 					mac_srs_worker_wakeup_ticks);	\
306 	}								\
307 }
308 
309 #define	TX_SINGLE_RING_MODE(mac_srs)				\
310 	((mac_srs)->srs_tx.st_mode == SRS_TX_DEFAULT || 	\
311 	    (mac_srs)->srs_tx.st_mode == SRS_TX_SERIALIZE ||	\
312 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW)
313 
314 #define	TX_BANDWIDTH_MODE(mac_srs)				\
315 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
316 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT)
317 
318 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
319 	uint_t hash, indx;						\
320 	hash = HASH_HINT(hint);					\
321 	indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);	\
322 	softring = mac_srs->srs_oth_soft_rings[indx];			\
323 	(void) (mac_tx_soft_ring_process(softring, head, 0, NULL));	\
324 }
325 
326 /*
327  * MAC_TX_SRS_BLOCK
328  *
329  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
330  * will be set only if srs_tx_woken_up is FALSE. If
331  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
332  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
333  * attempt to transmit again and not setting SRS_TX_BLOCKED does
334  * that.
335  */
336 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
337 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
338 	if ((srs)->srs_tx.st_woken_up) {			\
339 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
340 	} else {						\
341 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
342 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
343 		(srs)->srs_tx.st_blocked_cnt++;			\
344 	}							\
345 }
346 
347 /*
348  * MAC_TX_SRS_TEST_HIWAT
349  *
350  * Called before queueing a packet onto Tx SRS to test and set
351  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
352  */
353 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
354 	boolean_t enqueue = 1;						\
355 									\
356 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
357 		/*							\
358 		 * flow-controlled. Store srs in cookie so that it	\
359 		 * can be returned as mac_tx_cookie_t to client		\
360 		 */							\
361 		(srs)->srs_state |= SRS_TX_HIWAT;			\
362 		cookie = (mac_tx_cookie_t)srs;				\
363 		(srs)->srs_tx.st_hiwat_cnt++;				\
364 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
365 			/* increment freed stats */			\
366 			(srs)->srs_tx.st_drop_count += cnt;		\
367 			/*						\
368 			 * b_prev may be set to the fanout hint		\
369 			 * hence can't use freemsg directly		\
370 			 */						\
371 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);	\
372 			DTRACE_PROBE1(tx_queued_hiwat,			\
373 			    mac_soft_ring_set_t *, srs);		\
374 			enqueue = 0;					\
375 		}							\
376 	}								\
377 	if (enqueue)							\
378 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
379 }
380 
381 /* Some utility macros */
382 #define	MAC_SRS_BW_LOCK(srs)						\
383 	if (!(srs->srs_type & SRST_TX))					\
384 		mutex_enter(&srs->srs_bw->mac_bw_lock);
385 
386 #define	MAC_SRS_BW_UNLOCK(srs)						\
387 	if (!(srs->srs_type & SRST_TX))					\
388 		mutex_exit(&srs->srs_bw->mac_bw_lock);
389 
390 #define	MAC_TX_SRS_DROP_MESSAGE(srs, mp, cookie) {		\
391 	mac_pkt_drop(NULL, NULL, mp, B_FALSE);			\
392 	/* increment freed stats */				\
393 	mac_srs->srs_tx.st_drop_count++;			\
394 	cookie = (mac_tx_cookie_t)srs;				\
395 }
396 
397 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
398 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
399 	cookie = (mac_tx_cookie_t)srs;					\
400 	*ret_mp = mp_chain;						\
401 }
402 
403 /*
404  * Drop the rx packet and advance to the next one in the chain.
405  */
406 static void
407 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
408 {
409 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
410 
411 	ASSERT(mp->b_next == NULL);
412 	mutex_enter(&srs->srs_lock);
413 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
414 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
415 	mutex_exit(&srs->srs_lock);
416 
417 	srs_rx->sr_drop_count++;
418 	freemsg(mp);
419 }
420 
421 /* DATAPATH RUNTIME ROUTINES */
422 
423 /*
424  * mac_srs_fire
425  *
426  * Timer callback routine for waking up the SRS worker thread.
427  */
428 static void
429 mac_srs_fire(void *arg)
430 {
431 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
432 
433 	mutex_enter(&mac_srs->srs_lock);
434 	if (mac_srs->srs_tid == 0) {
435 		mutex_exit(&mac_srs->srs_lock);
436 		return;
437 	}
438 
439 	mac_srs->srs_tid = 0;
440 	if (!(mac_srs->srs_state & SRS_PROC))
441 		cv_signal(&mac_srs->srs_async);
442 
443 	mutex_exit(&mac_srs->srs_lock);
444 }
445 
446 /*
447  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
448  * and it is used on the TX path.
449  */
450 #define	HASH_HINT(hint)	(((hint) << 17) | ((hint) >> 16))
451 
452 /*
453  * hash based on the src address and the port information.
454  */
455 #define	HASH_ADDR(src, ports)					\
456 	(ntohl((src)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
457 	((ports) >> 8) ^ (ports))
458 
459 #define	COMPUTE_INDEX(key, sz)	(key % sz)
460 
461 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
462 	if ((tail) != NULL) {						\
463 		ASSERT((tail)->b_next == NULL);				\
464 		(tail)->b_next = (mp);					\
465 	} else {							\
466 		ASSERT((head) == NULL);					\
467 		(head) = (mp);						\
468 	}								\
469 	(tail) = (mp);							\
470 	(cnt)++;							\
471 	if ((bw_ctl))							\
472 		(sz) += (sz0);						\
473 }
474 
475 #define	MAC_FANOUT_DEFAULT	0
476 #define	MAC_FANOUT_RND_ROBIN	1
477 int mac_fanout_type = MAC_FANOUT_DEFAULT;
478 
479 #define	MAX_SR_TYPES	3
480 /* fanout types for port based hashing */
481 enum pkt_type {
482 	V4_TCP = 0,
483 	V4_UDP,
484 	OTH,
485 	UNDEF
486 };
487 
488 /*
489  * In general we do port based hashing to spread traffic over different
490  * softrings. The below tunable allows to override that behavior. Setting it
491  * to B_TRUE allows to do a fanout based on src ipv6 address. This behavior
492  * is also the applicable to ipv6 packets carrying multiple optional headers
493  * and other uncommon packet types.
494  */
495 boolean_t mac_src_ipv6_fanout = B_FALSE;
496 
497 /*
498  * Pair of local and remote ports in the transport header
499  */
500 #define	PORTS_SIZE 4
501 
502 /*
503  * mac_rx_srs_proto_fanout
504  *
505  * This routine delivers packets destined to an SRS into one of the
506  * protocol soft rings.
507  *
508  * Given a chain of packets we need to split it up into multiple sub chains
509  * destined into TCP, UDP or OTH soft ring. Instead of entering
510  * the soft ring one packet at a time, we want to enter it in the form of a
511  * chain otherwise we get this start/stop behaviour where the worker thread
512  * goes to sleep and then next packets comes in forcing it to wake up etc.
513  */
514 static void
515 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
516 {
517 	struct ether_header		*ehp;
518 	uint16_t			etype;
519 	ipha_t				*ipha;
520 	mac_soft_ring_t			*softring;
521 	size_t				ether_hlen;
522 	mblk_t				*mp;
523 	mblk_t				*headmp[MAX_SR_TYPES];
524 	mblk_t				*tailmp[MAX_SR_TYPES];
525 	int				cnt[MAX_SR_TYPES];
526 	size_t				sz[MAX_SR_TYPES];
527 	size_t				sz1;
528 	boolean_t			bw_ctl = B_FALSE;
529 	boolean_t			hw_classified;
530 	boolean_t			dls_bypass = B_TRUE;
531 	enum				pkt_type type;
532 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
533 	struct ether_vlan_header	*evhp;
534 
535 	if (mac_srs->srs_type & SRST_BW_CONTROL)
536 		bw_ctl = B_TRUE;
537 
538 	/*
539 	 * If we don't have a Rx ring, S/W classification would have done
540 	 * its job and its a packet meant for us. If we were polling on
541 	 * the default ring (i.e. there was a ring assigned to this SRS),
542 	 * then we need to make sure that the mac address really belongs
543 	 * to us.
544 	 */
545 	hw_classified = mac_srs->srs_ring != NULL &&
546 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
547 
548 	/*
549 	 * Special clients (eg. VLAN, non ether, etc) need DLS
550 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
551 	 * such SRSs.
552 	 */
553 	if (!(mac_srs->srs_type & SRST_DLS_BYPASS))
554 		dls_bypass = B_FALSE;
555 
556 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
557 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
558 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
559 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
560 
561 	/*
562 	 * We got a chain from SRS that we need to send to the soft rings.
563 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
564 	 * performance reasons), we need to separate out v4_tcp, v4_udp
565 	 * and the rest goes in other.
566 	 */
567 	while (head != NULL) {
568 		mp = head;
569 		head = head->b_next;
570 		mp->b_next = NULL;
571 
572 		type = OTH;
573 		sz1 = msgdsize(mp);
574 
575 		if (!dls_bypass) {
576 			mac_impl_t	*mip = mcip->mci_mip;
577 
578 			ehp = (struct ether_header *)mp->b_rptr;
579 
580 			/*
581 			 * For VLAN packets, if the VLAN id doesn't belong
582 			 * to this client, we drop the packet.
583 			 */
584 			if (mip->mi_info.mi_nativemedia == DL_ETHER &&
585 			    ntohs(ehp->ether_type) == VLAN_TPID) {
586 				/*
587 				 * LINTED: cast may result in improper
588 				 * alignment
589 				 */
590 				evhp = (struct ether_vlan_header *)ehp;
591 				if (!mac_client_check_flow_vid(mcip,
592 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
593 					mac_rx_drop_pkt(mac_srs, mp);
594 					continue;
595 				}
596 			}
597 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
598 			    cnt[type], bw_ctl, sz[type], sz1, mp);
599 			continue;
600 		}
601 
602 		/*
603 		 * At this point we can be sure the packet at least
604 		 * has an ether header.
605 		 */
606 		if (sz1 < sizeof (struct ether_header)) {
607 			mac_rx_drop_pkt(mac_srs, mp);
608 			continue;
609 		}
610 		/* LINTED: cast may result in improper alignment */
611 		ehp = (struct ether_header *)mp->b_rptr;
612 
613 		/*
614 		 * Determine if this is a VLAN or non-VLAN packet.
615 		 */
616 		if ((etype = ntohs(ehp->ether_type)) == VLAN_TPID) {
617 			/* LINTED: cast may result in improper alignment */
618 			evhp = (struct ether_vlan_header *)mp->b_rptr;
619 			etype = ntohs(evhp->ether_type);
620 			ether_hlen = sizeof (struct ether_vlan_header);
621 			/*
622 			 * Check if the VID of the packet, if any, belongs
623 			 * to this client.
624 			 */
625 			if (!mac_client_check_flow_vid(mcip,
626 			    VLAN_ID(ntohs(evhp->ether_tci)))) {
627 				mac_rx_drop_pkt(mac_srs, mp);
628 				continue;
629 			}
630 		} else {
631 			ether_hlen = sizeof (struct ether_header);
632 		}
633 
634 		if (etype == 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(&ehp->ether_dhost, map->ma_addr,
645 				    map->ma_len) == 0)
646 					type = UNDEF;
647 				rw_exit(&mcip->mci_rw_lock);
648 			} else if (((((uint8_t *)&ehp->ether_dhost)[0] &
649 			    0x01) == 0)) {
650 				type = UNDEF;
651 			}
652 		}
653 
654 		/*
655 		 * This needs to become a contract with the driver for
656 		 * the fast path.
657 		 *
658 		 * In the normal case the packet will have at least the L2
659 		 * header and the IP + Transport header in the same mblk.
660 		 * This is usually the case when the NIC driver sends up
661 		 * the packet. This is also true when the stack generates
662 		 * a packet that is looped back and when the stack uses the
663 		 * fastpath mechanism. The normal case is optimized for
664 		 * performance and may bypass DLS. All other cases go through
665 		 * the 'OTH' type path without DLS bypass.
666 		 */
667 
668 		/* LINTED: cast may result in improper alignment */
669 		ipha = (ipha_t *)(mp->b_rptr + ether_hlen);
670 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
671 			type = OTH;
672 
673 		if (type == OTH) {
674 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
675 			    cnt[type], bw_ctl, sz[type], sz1, mp);
676 			continue;
677 		}
678 
679 		ASSERT(type == UNDEF);
680 		/*
681 		 * We look for at least 4 bytes past the IP header to get
682 		 * the port information. If we get an IP fragment, we don't
683 		 * have the port information, and we use just the protocol
684 		 * information.
685 		 */
686 		switch (ipha->ipha_protocol) {
687 		case IPPROTO_TCP:
688 			type = V4_TCP;
689 			mp->b_rptr += ether_hlen;
690 			break;
691 		case IPPROTO_UDP:
692 			type = V4_UDP;
693 			mp->b_rptr += ether_hlen;
694 			break;
695 		default:
696 			type = OTH;
697 			break;
698 		}
699 
700 		ASSERT(type != UNDEF);
701 
702 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
703 		    bw_ctl, sz[type], sz1, mp);
704 	}
705 
706 	for (type = V4_TCP; type < UNDEF; type++) {
707 		if (headmp[type] != NULL) {
708 			ASSERT(tailmp[type]->b_next == NULL);
709 			switch (type) {
710 			case V4_TCP:
711 				softring = mac_srs->srs_tcp_soft_rings[0];
712 				break;
713 			case V4_UDP:
714 				softring = mac_srs->srs_udp_soft_rings[0];
715 				break;
716 			case OTH:
717 				softring = mac_srs->srs_oth_soft_rings[0];
718 			}
719 			mac_rx_soft_ring_process(mac_srs->srs_mcip, softring,
720 			    headmp[type], tailmp[type], cnt[type], sz[type]);
721 		}
722 	}
723 }
724 
725 int	fanout_unalligned = 0;
726 
727 /*
728  * mac_rx_srs_long_fanout
729  *
730  * The fanout routine for IPv6
731  */
732 static int
733 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
734     uint16_t etype, enum pkt_type *type, uint_t *indx)
735 {
736 	ip6_t		*ip6h;
737 	uint8_t		*whereptr;
738 	uint_t		hash;
739 	uint16_t	remlen;
740 	uint8_t		nexthdr;
741 	uint16_t	hdr_len;
742 
743 	if (etype == ETHERTYPE_IPV6) {
744 		boolean_t	modifiable = B_TRUE;
745 
746 		ASSERT(MBLKL(mp) >= sizeof (struct ether_header));
747 
748 		ip6h = (ip6_t *)(mp->b_rptr + sizeof (struct ether_header));
749 		if ((unsigned char *)ip6h == mp->b_wptr) {
750 			/*
751 			 * The first mblk_t only includes the ethernet header.
752 			 * Note that it is safe to change the mp pointer here,
753 			 * as the subsequent operation does not assume mp
754 			 * points to the start of the ethernet header.
755 			 */
756 			mp = mp->b_cont;
757 
758 			/*
759 			 * Make sure ip6h holds the full ip6_t structure.
760 			 */
761 			if (mp == NULL)
762 				return (-1);
763 
764 			if (MBLKL(mp) < IPV6_HDR_LEN) {
765 				modifiable = (DB_REF(mp) == 1);
766 
767 				if (modifiable &&
768 				    !pullupmsg(mp, IPV6_HDR_LEN)) {
769 					return (-1);
770 				}
771 			}
772 
773 			ip6h = (ip6_t *)mp->b_rptr;
774 		}
775 
776 		if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
777 		    ((unsigned char *)ip6h + IPV6_HDR_LEN > mp->b_wptr)) {
778 			/*
779 			 * If either ip6h is not alligned, or ip6h does not
780 			 * hold the complete ip6_t structure (a pullupmsg()
781 			 * is not an option since it would result in an
782 			 * unalligned ip6h), fanout to the default ring. Note
783 			 * that this may cause packets reordering.
784 			 */
785 			*indx = 0;
786 			*type = OTH;
787 			fanout_unalligned++;
788 			return (0);
789 		}
790 
791 		remlen = ntohs(ip6h->ip6_plen);
792 		nexthdr = ip6h->ip6_nxt;
793 
794 		if (remlen < MIN_EHDR_LEN)
795 			return (-1);
796 		/*
797 		 * Do src based fanout if below tunable is set to B_TRUE or
798 		 * when mac_ip_hdr_length_v6() fails because of malformed
799 		 * packets or because mblk's need to be concatenated using
800 		 * pullupmsg().
801 		 */
802 		if (mac_src_ipv6_fanout || !mac_ip_hdr_length_v6(mp, ip6h,
803 		    &hdr_len, &nexthdr)) {
804 			goto src_based_fanout;
805 		}
806 		whereptr = (uint8_t *)ip6h + hdr_len;
807 
808 		/* If the transport is one of below, we do port based fanout */
809 		switch (nexthdr) {
810 		case IPPROTO_TCP:
811 		case IPPROTO_UDP:
812 		case IPPROTO_SCTP:
813 		case IPPROTO_ESP:
814 			/*
815 			 * If the ports in the transport header is not part of
816 			 * the mblk, do src_based_fanout, instead of calling
817 			 * pullupmsg().
818 			 */
819 			if (mp->b_cont != NULL &&
820 			    whereptr + PORTS_SIZE > mp->b_wptr) {
821 				goto src_based_fanout;
822 			}
823 			break;
824 		default:
825 			break;
826 		}
827 
828 		switch (nexthdr) {
829 		case IPPROTO_TCP:
830 			hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
831 			    *(uint32_t *)whereptr);
832 			*indx = COMPUTE_INDEX(hash,
833 			    mac_srs->srs_tcp_ring_count);
834 			*type = OTH;
835 			break;
836 
837 		case IPPROTO_UDP:
838 		case IPPROTO_SCTP:
839 		case IPPROTO_ESP:
840 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
841 				hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
842 				    *(uint32_t *)whereptr);
843 				*indx = COMPUTE_INDEX(hash,
844 				    mac_srs->srs_udp_ring_count);
845 			} else {
846 				*indx = mac_srs->srs_ind %
847 				    mac_srs->srs_udp_ring_count;
848 				mac_srs->srs_ind++;
849 			}
850 			*type = OTH;
851 			break;
852 
853 			/* For all other protocol, do source based fanout */
854 		default:
855 			goto src_based_fanout;
856 		}
857 	} else {
858 		*indx = 0;
859 		*type = OTH;
860 	}
861 	return (0);
862 
863 src_based_fanout:
864 	hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src), (uint32_t)0);
865 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
866 	*type = OTH;
867 	return (0);
868 }
869 
870 /*
871  * mac_rx_srs_fanout
872  *
873  * This routine delivers packets destined to an SRS into a soft ring member
874  * of the set.
875  *
876  * Given a chain of packets we need to split it up into multiple sub chains
877  * destined for one of the TCP, UDP or OTH soft rings. Instead of entering
878  * the soft ring one packet at a time, we want to enter it in the form of a
879  * chain otherwise we get this start/stop behaviour where the worker thread
880  * goes to sleep and then next packets comes in forcing it to wake up etc.
881  *
882  * Note:
883  * Since we know what is the maximum fanout possible, we create a 2D array
884  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
885  * variables so that we can enter the softrings with chain. We need the
886  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
887  * for each packet would be expensive). If we ever want to have the
888  * ability to have unlimited fanout, we should probably declare a head,
889  * tail, cnt, sz with each soft ring (a data struct which contains a softring
890  * along with these members) and create an array of this uber struct so we
891  * don't have to do kmem_alloc.
892  */
893 int	fanout_oth1 = 0;
894 int	fanout_oth2 = 0;
895 int	fanout_oth3 = 0;
896 int	fanout_oth4 = 0;
897 int	fanout_oth5 = 0;
898 
899 static void
900 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
901 {
902 	struct ether_header		*ehp;
903 	uint16_t			etype;
904 	ipha_t				*ipha;
905 	uint_t				indx;
906 	int				ports_offset = -1;
907 	int				ipha_len;
908 	uint_t				hash;
909 	mac_soft_ring_t			*softring;
910 	size_t				ether_hlen;
911 	uint16_t			frag_offset_flags;
912 	mblk_t				*mp;
913 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
914 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
915 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
916 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
917 	size_t				sz1;
918 	boolean_t			bw_ctl = B_FALSE;
919 	boolean_t			hw_classified;
920 	boolean_t			dls_bypass = B_TRUE;
921 	int				i;
922 	int				fanout_cnt;
923 	enum 				pkt_type type;
924 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
925 	struct ether_vlan_header	*evhp;
926 
927 	if (mac_srs->srs_type & SRST_BW_CONTROL)
928 		bw_ctl = B_TRUE;
929 
930 	/*
931 	 * If we don't have a Rx ring, S/W classification would have done
932 	 * its job and its a packet meant for us. If we were polling on
933 	 * the default ring (i.e. there was a ring assigned to this SRS),
934 	 * then we need to make sure that the mac address really belongs
935 	 * to us.
936 	 */
937 	hw_classified = mac_srs->srs_ring != NULL &&
938 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
939 
940 	/*
941 	 * Special clients (eg. VLAN, non ether, etc) need DLS
942 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
943 	 * such SRSs.
944 	 */
945 	if (!(mac_srs->srs_type & SRST_DLS_BYPASS))
946 		dls_bypass = B_FALSE;
947 
948 	/*
949 	 * Since the softrings are never destroyed and we always
950 	 * create equal number of softrings for TCP, UDP and rest,
951 	 * its OK to check one of them for count and use it without
952 	 * any lock. In future, if soft rings get destroyed because
953 	 * of reduction in fanout, we will need to ensure that happens
954 	 * behind the SRS_PROC.
955 	 */
956 	fanout_cnt = mac_srs->srs_tcp_ring_count;
957 
958 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
959 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
960 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
961 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
962 
963 	/*
964 	 * We got a chain from SRS that we need to send to the soft rings.
965 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
966 	 * performance reasons), we need to separate out v4_tcp, v4_udp
967 	 * and the rest goes in other.
968 	 */
969 	while (head != NULL) {
970 		mp = head;
971 		head = head->b_next;
972 		mp->b_next = NULL;
973 
974 		type = OTH;
975 		sz1 = msgdsize(mp);
976 
977 		if (!dls_bypass) {
978 			mac_impl_t	*mip = mcip->mci_mip;
979 
980 			indx = 0;
981 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
982 				ehp = (struct ether_header *)mp->b_rptr;
983 				etype = ntohs(ehp->ether_type);
984 				/*
985 				 * For VLAN packets, if the VLAN id doesn't
986 				 * belong to this client, we drop the packet.
987 				 */
988 				if (etype == VLAN_TPID) {
989 					/*
990 					 * LINTED: cast may result in improper
991 					 * alignment
992 					 */
993 					evhp = (struct ether_vlan_header *)
994 					    mp->b_rptr;
995 					if (!mac_client_check_flow_vid(mcip,
996 					    VLAN_ID(ntohs(evhp->ether_tci)))) {
997 						mac_rx_drop_pkt(mac_srs, mp);
998 						continue;
999 					}
1000 				}
1001 				if (mac_rx_srs_long_fanout(mac_srs, mp, etype,
1002 				    &type, &indx) == -1) {
1003 					mac_rx_drop_pkt(mac_srs, mp);
1004 					continue;
1005 				}
1006 			}
1007 
1008 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1009 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1010 			    sz[type][indx], sz1, mp);
1011 			continue;
1012 		}
1013 
1014 		/*
1015 		 * At this point we can be sure the packet at least
1016 		 * has an ether header. On the outbound side, GLD/stack
1017 		 * ensure this. On the inbound side, the driver needs
1018 		 * to ensure this.
1019 		 */
1020 		if (sz1 < sizeof (struct ether_header)) {
1021 			mac_rx_drop_pkt(mac_srs, mp);
1022 			continue;
1023 		}
1024 		/* LINTED: cast may result in improper alignment */
1025 		ehp = (struct ether_header *)mp->b_rptr;
1026 
1027 		/*
1028 		 * Determine if this is a VLAN or non-VLAN packet.
1029 		 */
1030 		if ((etype = ntohs(ehp->ether_type)) == VLAN_TPID) {
1031 			/* LINTED: cast may result in improper alignment */
1032 			evhp = (struct ether_vlan_header *)mp->b_rptr;
1033 			etype = ntohs(evhp->ether_type);
1034 			ether_hlen = sizeof (struct ether_vlan_header);
1035 			/*
1036 			 * Check if the VID of the packet, if any, belongs
1037 			 * to this client.
1038 			 */
1039 			if (!mac_client_check_flow_vid(mcip,
1040 			    VLAN_ID(ntohs(evhp->ether_tci)))) {
1041 				mac_rx_drop_pkt(mac_srs, mp);
1042 				continue;
1043 			}
1044 		} else {
1045 			ether_hlen = sizeof (struct ether_header);
1046 		}
1047 
1048 
1049 		/*
1050 		 * If we are using the default Rx ring where H/W or S/W
1051 		 * classification has not happened, we need to verify if
1052 		 * this unicast packet really belongs to us.
1053 		 */
1054 		if (etype == ETHERTYPE_IP) {
1055 			/*
1056 			 * If we are H/W classified, but we have promisc
1057 			 * on, then we need to check for the unicast address.
1058 			 */
1059 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1060 				mac_address_t		*map;
1061 
1062 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1063 				map = mcip->mci_unicast;
1064 				if (bcmp(&ehp->ether_dhost, map->ma_addr,
1065 				    map->ma_len) == 0)
1066 					type = UNDEF;
1067 				rw_exit(&mcip->mci_rw_lock);
1068 			} else if (((((uint8_t *)&ehp->ether_dhost)[0] &
1069 			    0x01) == 0)) {
1070 				type = UNDEF;
1071 			}
1072 		}
1073 
1074 		/*
1075 		 * This needs to become a contract with the driver for
1076 		 * the fast path.
1077 		 */
1078 
1079 		/* LINTED: cast may result in improper alignment */
1080 		ipha = (ipha_t *)(mp->b_rptr + ether_hlen);
1081 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
1082 			type = OTH;
1083 			fanout_oth1++;
1084 		}
1085 
1086 		if (type != OTH) {
1087 			switch (ipha->ipha_protocol) {
1088 			case IPPROTO_TCP:
1089 			case IPPROTO_UDP:
1090 			case IPPROTO_SCTP:
1091 			case IPPROTO_ESP:
1092 				ipha_len = IPH_HDR_LENGTH(ipha);
1093 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
1094 				    mp->b_wptr) {
1095 					type = OTH;
1096 					break;
1097 				}
1098 				frag_offset_flags =
1099 				    ntohs(ipha->ipha_fragment_offset_and_flags);
1100 				if ((frag_offset_flags &
1101 				    (IPH_MF | IPH_OFFSET)) != 0) {
1102 					type = OTH;
1103 					fanout_oth3++;
1104 					break;
1105 				}
1106 				ports_offset = ether_hlen + ipha_len;
1107 				break;
1108 			default:
1109 				type = OTH;
1110 				fanout_oth4++;
1111 				break;
1112 			}
1113 		}
1114 
1115 		if (type == OTH) {
1116 			if (mac_rx_srs_long_fanout(mac_srs, mp, etype,
1117 			    &type, &indx) == -1) {
1118 				mac_rx_drop_pkt(mac_srs, mp);
1119 				continue;
1120 			}
1121 
1122 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1123 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1124 			    sz[type][indx], sz1, mp);
1125 			continue;
1126 		}
1127 
1128 		ASSERT(type == UNDEF);
1129 
1130 		/*
1131 		 * XXX-Sunay: We should hold srs_lock since ring_count
1132 		 * below can change. But if we are always called from
1133 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
1134 		 * enforce that ring_count can't be changed i.e.
1135 		 * to change fanout type or ring count, the calling
1136 		 * thread needs to be behind SRS_PROC.
1137 		 */
1138 		switch (ipha->ipha_protocol) {
1139 		case IPPROTO_TCP:
1140 			/*
1141 			 * Note that for ESP, we fanout on SPI and it is at the
1142 			 * same offset as the 2x16-bit ports. So it is clumped
1143 			 * along with TCP, UDP and SCTP.
1144 			 */
1145 			hash = HASH_ADDR(ipha->ipha_src,
1146 			    *(uint32_t *)(mp->b_rptr + ports_offset));
1147 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1148 			type = V4_TCP;
1149 			mp->b_rptr += ether_hlen;
1150 			break;
1151 		case IPPROTO_UDP:
1152 		case IPPROTO_SCTP:
1153 		case IPPROTO_ESP:
1154 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1155 				hash = HASH_ADDR(ipha->ipha_src,
1156 				    *(uint32_t *)(mp->b_rptr + ports_offset));
1157 				indx = COMPUTE_INDEX(hash,
1158 				    mac_srs->srs_udp_ring_count);
1159 			} else {
1160 				indx = mac_srs->srs_ind %
1161 				    mac_srs->srs_udp_ring_count;
1162 				mac_srs->srs_ind++;
1163 			}
1164 			type = V4_UDP;
1165 			mp->b_rptr += ether_hlen;
1166 			break;
1167 		}
1168 
1169 		ASSERT(type != UNDEF);
1170 
1171 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
1172 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
1173 	}
1174 
1175 	for (type = V4_TCP; type < UNDEF; type++) {
1176 		for (i = 0; i < fanout_cnt; i++) {
1177 			if (headmp[type][i] != NULL) {
1178 				ASSERT(tailmp[type][i]->b_next == NULL);
1179 				switch (type) {
1180 				case V4_TCP:
1181 					softring =
1182 					    mac_srs->srs_tcp_soft_rings[i];
1183 					break;
1184 				case V4_UDP:
1185 					softring =
1186 					    mac_srs->srs_udp_soft_rings[i];
1187 					break;
1188 				case OTH:
1189 					softring =
1190 					    mac_srs->srs_oth_soft_rings[i];
1191 					break;
1192 				}
1193 				mac_rx_soft_ring_process(mac_srs->srs_mcip,
1194 				    softring, headmp[type][i], tailmp[type][i],
1195 				    cnt[type][i], sz[type][i]);
1196 			}
1197 		}
1198 	}
1199 }
1200 
1201 #define	SRS_BYTES_TO_PICKUP	150000
1202 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
1203 
1204 /*
1205  * mac_rx_srs_poll_ring
1206  *
1207  * This SRS Poll thread uses this routine to poll the underlying hardware
1208  * Rx ring to get a chain of packets. It can inline process that chain
1209  * if mac_latency_optimize is set (default) or signal the SRS worker thread
1210  * to do the remaining processing.
1211  *
1212  * Since packets come in the system via interrupt or poll path, we also
1213  * update the stats and deal with promiscous clients here.
1214  */
1215 void
1216 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
1217 {
1218 	kmutex_t 		*lock = &mac_srs->srs_lock;
1219 	kcondvar_t 		*async = &mac_srs->srs_cv;
1220 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1221 	mblk_t 			*head, *tail, *mp;
1222 	callb_cpr_t 		cprinfo;
1223 	ssize_t 		bytes_to_pickup;
1224 	size_t 			sz;
1225 	int			count;
1226 	mac_client_impl_t	*smcip;
1227 
1228 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
1229 	mutex_enter(lock);
1230 
1231 start:
1232 	for (;;) {
1233 		if (mac_srs->srs_state & SRS_PAUSE)
1234 			goto done;
1235 
1236 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
1237 		cv_wait(async, lock);
1238 		CALLB_CPR_SAFE_END(&cprinfo, lock);
1239 
1240 		if (mac_srs->srs_state & SRS_PAUSE)
1241 			goto done;
1242 
1243 check_again:
1244 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
1245 			/*
1246 			 * We pick as many bytes as we are allowed to queue.
1247 			 * Its possible that we will exceed the total
1248 			 * packets queued in case this SRS is part of the
1249 			 * Rx ring group since > 1 poll thread can be pulling
1250 			 * upto the max allowed packets at the same time
1251 			 * but that should be OK.
1252 			 */
1253 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1254 			bytes_to_pickup =
1255 			    mac_srs->srs_bw->mac_bw_drop_threshold -
1256 			    mac_srs->srs_bw->mac_bw_sz;
1257 			/*
1258 			 * We shouldn't have been signalled if we
1259 			 * have 0 or less bytes to pick but since
1260 			 * some of the bytes accounting is driver
1261 			 * dependant, we do the safety check.
1262 			 */
1263 			if (bytes_to_pickup < 0)
1264 				bytes_to_pickup = 0;
1265 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1266 		} else {
1267 			/*
1268 			 * ToDO: Need to change the polling API
1269 			 * to add a packet count and a flag which
1270 			 * tells the driver whether we want packets
1271 			 * based on a count, or bytes, or all the
1272 			 * packets queued in the driver/HW. This
1273 			 * way, we never have to check the limits
1274 			 * on poll path. We truly let only as many
1275 			 * packets enter the system as we are willing
1276 			 * to process or queue.
1277 			 *
1278 			 * Something along the lines of
1279 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
1280 			 *	mac_srs->srs_poll_pkt_cnt
1281 			 */
1282 
1283 			/*
1284 			 * Since we are not doing B/W control, pick
1285 			 * as many packets as allowed.
1286 			 */
1287 			bytes_to_pickup = max_bytes_to_pickup;
1288 		}
1289 
1290 		/* Poll the underlying Hardware */
1291 		mutex_exit(lock);
1292 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
1293 		mutex_enter(lock);
1294 
1295 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
1296 		    SRS_POLL_THR_OWNER);
1297 
1298 		mp = tail = head;
1299 		count = 0;
1300 		sz = 0;
1301 		while (mp != NULL) {
1302 			tail = mp;
1303 			sz += msgdsize(mp);
1304 			mp = mp->b_next;
1305 			count++;
1306 		}
1307 
1308 		if (head != NULL) {
1309 			tail->b_next = NULL;
1310 			smcip = mac_srs->srs_mcip;
1311 
1312 			if ((mac_srs->srs_type & SRST_FLOW) ||
1313 			    (smcip == NULL)) {
1314 				FLOW_STAT_UPDATE(mac_srs->srs_flent,
1315 				    rbytes, sz);
1316 				FLOW_STAT_UPDATE(mac_srs->srs_flent,
1317 				    ipackets, count);
1318 			}
1319 
1320 			/*
1321 			 * If there are any promiscuous mode callbacks
1322 			 * defined for this MAC client, pass them a copy
1323 			 * if appropriate and also update the counters.
1324 			 */
1325 			if (smcip != NULL) {
1326 				smcip->mci_stat_ibytes += sz;
1327 				smcip->mci_stat_ipackets += count;
1328 
1329 				if (smcip->mci_mip->mi_promisc_list != NULL) {
1330 					mutex_exit(lock);
1331 					mac_promisc_dispatch(smcip->mci_mip,
1332 					    head, NULL);
1333 					mutex_enter(lock);
1334 				}
1335 			}
1336 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
1337 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1338 				mac_srs->srs_bw->mac_bw_polled += sz;
1339 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1340 			}
1341 			srs_rx->sr_poll_count += count;
1342 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
1343 			    count, sz);
1344 			if (count <= 10)
1345 				srs_rx->sr_chain_cnt_undr10++;
1346 			else if (count > 10 && count <= 50)
1347 				srs_rx->sr_chain_cnt_10to50++;
1348 			else
1349 				srs_rx->sr_chain_cnt_over50++;
1350 		}
1351 
1352 		/*
1353 		 * We are guaranteed that SRS_PROC will be set if we
1354 		 * are here. Also, poll thread gets to run only if
1355 		 * the drain was being done by a worker thread although
1356 		 * its possible that worker thread is still running
1357 		 * and poll thread was sent down to keep the pipeline
1358 		 * going instead of doing a complete drain and then
1359 		 * trying to poll the NIC.
1360 		 *
1361 		 * So we need to check SRS_WORKER flag to make sure
1362 		 * that the worker thread is not processing the queue
1363 		 * in parallel to us. The flags and conditions are
1364 		 * protected by the srs_lock to prevent any race. We
1365 		 * ensure that we don't drop the srs_lock from now
1366 		 * till the end and similarly we don't drop the srs_lock
1367 		 * in mac_rx_srs_drain() till similar condition check
1368 		 * are complete. The mac_rx_srs_drain() needs to ensure
1369 		 * that SRS_WORKER flag remains set as long as its
1370 		 * processing the queue.
1371 		 */
1372 		if (!(mac_srs->srs_state & SRS_WORKER) &&
1373 		    (mac_srs->srs_first != NULL)) {
1374 			/*
1375 			 * We have packets to process and worker thread
1376 			 * is not running.  Check to see if poll thread is
1377 			 * allowed to process. Let it do processing only if it
1378 			 * picked up some packets from the NIC otherwise
1379 			 * wakeup the worker thread.
1380 			 */
1381 			if ((mac_srs->srs_state & SRS_LATENCY_OPT) &&
1382 			    (head != NULL)) {
1383 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
1384 				if (srs_rx->sr_poll_pkt_cnt <=
1385 				    srs_rx->sr_lowat) {
1386 					srs_rx->sr_poll_again++;
1387 					goto check_again;
1388 				} else {
1389 					/*
1390 					 * We are already above low water mark
1391 					 * so stay in the polling mode but no
1392 					 * need to poll. Once we dip below
1393 					 * the polling threshold, the processing
1394 					 * thread (soft ring) will signal us
1395 					 * to poll again (MAC_UPDATE_SRS_COUNT)
1396 					 */
1397 					srs_rx->sr_poll_drain_no_poll++;
1398 					mac_srs->srs_state &=
1399 					    ~(SRS_PROC|SRS_GET_PKTS);
1400 					/*
1401 					 * In B/W control case, its possible
1402 					 * that the backlog built up due to
1403 					 * B/W limit being reached and packets
1404 					 * are queued only in SRS. In this case,
1405 					 * we should schedule worker thread
1406 					 * since no one else will wake us up.
1407 					 */
1408 					if ((mac_srs->srs_type &
1409 					    SRST_BW_CONTROL) &&
1410 					    (mac_srs->srs_tid == NULL)) {
1411 						mac_srs->srs_tid =
1412 						    timeout(mac_srs_fire,
1413 						    mac_srs, 1);
1414 						srs_rx->sr_poll_worker_wakeup++;
1415 					}
1416 				}
1417 			} else {
1418 				/*
1419 				 * Wakeup the worker thread for more processing.
1420 				 * We optimize for throughput in this case.
1421 				 */
1422 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1423 				MAC_SRS_WORKER_WAKEUP(mac_srs);
1424 				srs_rx->sr_poll_sig_worker++;
1425 			}
1426 		} else if ((mac_srs->srs_first == NULL) &&
1427 		    !(mac_srs->srs_state & SRS_WORKER)) {
1428 			/*
1429 			 * There is nothing queued in SRS and
1430 			 * no worker thread running. Plus we
1431 			 * didn't get anything from the H/W
1432 			 * as well (head == NULL);
1433 			 */
1434 			ASSERT(head == NULL);
1435 			mac_srs->srs_state &=
1436 			    ~(SRS_PROC|SRS_GET_PKTS);
1437 
1438 			/*
1439 			 * If we have a packets in soft ring, don't allow
1440 			 * more packets to come into this SRS by keeping the
1441 			 * interrupts off but not polling the H/W. The
1442 			 * poll thread will get signaled as soon as
1443 			 * srs_poll_pkt_cnt dips below poll threshold.
1444 			 */
1445 			if (srs_rx->sr_poll_pkt_cnt == 0) {
1446 				srs_rx->sr_poll_intr_enable++;
1447 				MAC_SRS_POLLING_OFF(mac_srs);
1448 			} else {
1449 				/*
1450 				 * We know nothing is queued in SRS
1451 				 * since we are here after checking
1452 				 * srs_first is NULL. The backlog
1453 				 * is entirely due to packets queued
1454 				 * in Soft ring which will wake us up
1455 				 * and get the interface out of polling
1456 				 * mode once the backlog dips below
1457 				 * sr_poll_thres.
1458 				 */
1459 				srs_rx->sr_poll_no_poll++;
1460 			}
1461 		} else {
1462 			/*
1463 			 * Worker thread is already running.
1464 			 * Nothing much to do. If the polling
1465 			 * was enabled, worker thread will deal
1466 			 * with that.
1467 			 */
1468 			mac_srs->srs_state &= ~SRS_GET_PKTS;
1469 			srs_rx->sr_poll_goto_sleep++;
1470 		}
1471 	}
1472 done:
1473 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
1474 	cv_signal(&mac_srs->srs_async);
1475 	/*
1476 	 * If this is a temporary quiesce then wait for the restart signal
1477 	 * from the srs worker. Then clear the flags and signal the srs worker
1478 	 * to ensure a positive handshake and go back to start.
1479 	 */
1480 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
1481 		cv_wait(async, lock);
1482 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
1483 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
1484 		mac_srs->srs_state &=
1485 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
1486 		cv_signal(&mac_srs->srs_async);
1487 		goto start;
1488 	} else {
1489 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
1490 		cv_signal(&mac_srs->srs_async);
1491 		CALLB_CPR_EXIT(&cprinfo);
1492 		thread_exit();
1493 	}
1494 }
1495 
1496 /*
1497  * mac_srs_pick_chain
1498  *
1499  * In Bandwidth control case, checks how many packets can be processed
1500  * and return them in a sub chain.
1501  */
1502 static mblk_t *
1503 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
1504     size_t *chain_sz, int *chain_cnt)
1505 {
1506 	mblk_t 			*head = NULL;
1507 	mblk_t 			*tail = NULL;
1508 	size_t			sz;
1509 	size_t 			tsz = 0;
1510 	int			cnt = 0;
1511 	mblk_t 			*mp;
1512 
1513 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1514 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1515 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
1516 	    mac_srs->srs_bw->mac_bw_limit) ||
1517 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
1518 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1519 		head = mac_srs->srs_first;
1520 		mac_srs->srs_first = NULL;
1521 		*chain_tail = mac_srs->srs_last;
1522 		mac_srs->srs_last = NULL;
1523 		*chain_sz = mac_srs->srs_size;
1524 		*chain_cnt = mac_srs->srs_count;
1525 		mac_srs->srs_count = 0;
1526 		mac_srs->srs_size = 0;
1527 		return (head);
1528 	}
1529 
1530 	/*
1531 	 * Can't clear the entire backlog.
1532 	 * Need to find how many packets to pick
1533 	 */
1534 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
1535 	while ((mp = mac_srs->srs_first) != NULL) {
1536 		sz = msgdsize(mp);
1537 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
1538 		    mac_srs->srs_bw->mac_bw_limit) {
1539 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
1540 				mac_srs->srs_bw->mac_bw_state |=
1541 				    SRS_BW_ENFORCED;
1542 			break;
1543 		}
1544 
1545 		/*
1546 		 * The _size & cnt is  decremented from the softrings
1547 		 * when they send up the packet for polling to work
1548 		 * properly.
1549 		 */
1550 		tsz += sz;
1551 		cnt++;
1552 		mac_srs->srs_count--;
1553 		mac_srs->srs_size -= sz;
1554 		if (tail != NULL)
1555 			tail->b_next = mp;
1556 		else
1557 			head = mp;
1558 		tail = mp;
1559 		mac_srs->srs_first = mac_srs->srs_first->b_next;
1560 	}
1561 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1562 	if (mac_srs->srs_first == NULL)
1563 		mac_srs->srs_last = NULL;
1564 
1565 	if (tail != NULL)
1566 		tail->b_next = NULL;
1567 	*chain_tail = tail;
1568 	*chain_cnt = cnt;
1569 	*chain_sz = tsz;
1570 
1571 	return (head);
1572 }
1573 
1574 /*
1575  * mac_rx_srs_drain
1576  *
1577  * The SRS drain routine. Gets to run to clear the queue. Any thread
1578  * (worker, interrupt, poll) can call this based on processing model.
1579  * The first thing we do is disable interrupts if possible and then
1580  * drain the queue. we also try to poll the underlying hardware if
1581  * there is a dedicated hardware Rx ring assigned to this SRS.
1582  *
1583  * There is a equivalent drain routine in bandwidth control mode
1584  * mac_rx_srs_drain_bw. There is some code duplication between the two
1585  * routines but they are highly performance sensitive and are easier
1586  * to read/debug if they stay separate. Any code changes here might
1587  * also apply to mac_rx_srs_drain_bw as well.
1588  */
1589 void
1590 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1591 {
1592 	mblk_t 			*head;
1593 	mblk_t			*tail;
1594 	timeout_id_t 		tid;
1595 	int			cnt = 0;
1596 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1597 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1598 
1599 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1600 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
1601 again:
1602 	/* If we are blanked i.e. can't do upcalls, then we are done */
1603 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1604 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1605 		    (mac_srs->srs_state & SRS_PAUSE));
1606 		goto out;
1607 	}
1608 
1609 	if (mac_srs->srs_first == NULL)
1610 		goto out;
1611 
1612 	head = mac_srs->srs_first;
1613 	mac_srs->srs_first = NULL;
1614 	tail = mac_srs->srs_last;
1615 	mac_srs->srs_last = NULL;
1616 	cnt = mac_srs->srs_count;
1617 	mac_srs->srs_count = 0;
1618 
1619 	ASSERT(head != NULL);
1620 	ASSERT(tail != NULL);
1621 
1622 	if ((tid = mac_srs->srs_tid) != 0)
1623 		mac_srs->srs_tid = 0;
1624 
1625 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1626 
1627 	/* Switch to polling mode */
1628 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
1629 	if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1630 		MAC_SRS_POLL_RING(mac_srs);
1631 	/*
1632 	 * mcip is NULL for broadcast and multicast flows. The promisc
1633 	 * callbacks for broadcast and multicast packets are delivered from
1634 	 * mac_rx() and we don't need to worry about that case in this path
1635 	 */
1636 	if (mcip != NULL && mcip->mci_promisc_list != NULL) {
1637 		mutex_exit(&mac_srs->srs_lock);
1638 		mac_promisc_client_dispatch(mcip, head);
1639 		mutex_enter(&mac_srs->srs_lock);
1640 	}
1641 
1642 	/*
1643 	 * Check if SRS itself is doing the processing
1644 	 * This direct path does not apply when subflows are present. In this
1645 	 * case, packets need to be dispatched to a soft ring according to the
1646 	 * flow's bandwidth and other resources contraints.
1647 	 */
1648 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1649 		mac_direct_rx_t		proc;
1650 		void			*arg1;
1651 		mac_resource_handle_t	arg2;
1652 
1653 		/*
1654 		 * This is the case when a Rx is directly
1655 		 * assigned and we have a fully classified
1656 		 * protocol chain. We can deal with it in
1657 		 * one shot.
1658 		 */
1659 		proc = srs_rx->sr_func;
1660 		arg1 = srs_rx->sr_arg1;
1661 		arg2 = srs_rx->sr_arg2;
1662 
1663 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1664 		mutex_exit(&mac_srs->srs_lock);
1665 		if (tid != 0) {
1666 			(void) untimeout(tid);
1667 			tid = 0;
1668 		}
1669 
1670 		proc(arg1, arg2, head, NULL);
1671 		/*
1672 		 * Decrement the size and count here itelf
1673 		 * since the packet has been processed.
1674 		 */
1675 		mutex_enter(&mac_srs->srs_lock);
1676 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1677 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1678 			cv_signal(&mac_srs->srs_client_cv);
1679 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1680 	} else {
1681 		/* Some kind of softrings based fanout is required */
1682 		mutex_exit(&mac_srs->srs_lock);
1683 		if (tid != 0) {
1684 			(void) untimeout(tid);
1685 			tid = 0;
1686 		}
1687 
1688 		/*
1689 		 * Since the fanout routines can deal with chains,
1690 		 * shoot the entire chain up.
1691 		 */
1692 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1693 			mac_rx_srs_fanout(mac_srs, head);
1694 		else
1695 			mac_rx_srs_proto_fanout(mac_srs, head);
1696 		mutex_enter(&mac_srs->srs_lock);
1697 	}
1698 
1699 	/*
1700 	 * Send the poll thread to pick up any packets arrived
1701 	 * so far. This also serves as the last check in case
1702 	 * nothing else is queued in the SRS. The poll thread
1703 	 * is signalled only in the case the drain was done
1704 	 * by the worker thread and SRS_WORKER is set. The
1705 	 * worker thread can run in parallel as long as the
1706 	 * SRS_WORKER flag is set. We we have nothing else to
1707 	 * process, we can exit while leaving SRS_PROC set
1708 	 * which gives the poll thread control to process and
1709 	 * cleanup once it returns from the NIC.
1710 	 *
1711 	 * If we have nothing else to process, we need to
1712 	 * ensure that we keep holding the srs_lock till
1713 	 * all the checks below are done and control is
1714 	 * handed to the poll thread if it was running.
1715 	 */
1716 	if (mac_srs->srs_first != NULL) {
1717 		if (proc_type == SRS_WORKER) {
1718 			if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1719 				MAC_SRS_POLL_RING(mac_srs);
1720 			srs_rx->sr_drain_again++;
1721 			goto again;
1722 		} else {
1723 			srs_rx->sr_drain_worker_sig++;
1724 			cv_signal(&mac_srs->srs_async);
1725 		}
1726 	}
1727 
1728 out:
1729 
1730 	if (mac_srs->srs_state & SRS_GET_PKTS) {
1731 		/*
1732 		 * Poll thread is already running. Leave the
1733 		 * SRS_RPOC set and hand over the control to
1734 		 * poll thread.
1735 		 */
1736 		mac_srs->srs_state &= ~proc_type;
1737 		srs_rx->sr_drain_poll_running++;
1738 		return;
1739 	}
1740 
1741 	/*
1742 	 * Even if there are no packets queued in SRS, we
1743 	 * need to make sure that the shared counter is
1744 	 * clear and any associated softrings have cleared
1745 	 * all the backlog. Otherwise, leave the interface
1746 	 * in polling mode and the poll thread will get
1747 	 * signalled once the count goes down to zero.
1748 	 *
1749 	 * If someone is already draining the queue (SRS_PROC is
1750 	 * set) when the srs_poll_pkt_cnt goes down to zero,
1751 	 * then it means that drain is already running and we
1752 	 * will turn off polling at that time if there is
1753 	 * no backlog.
1754 	 *
1755 	 * As long as there are packets queued either
1756 	 * in soft ring set or its soft rings, we will leave
1757 	 * the interface in polling mode (even if the drain
1758 	 * was done being the interrupt thread). We signal
1759 	 * the poll thread as well if we have dipped below
1760 	 * low water mark.
1761 	 *
1762 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
1763 	 * since that turn polling on only for worker thread.
1764 	 * Its not worth turning polling on for interrupt
1765 	 * thread (since NIC will not issue another interrupt)
1766 	 * unless a backlog builds up.
1767 	 */
1768 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
1769 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
1770 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1771 		srs_rx->sr_drain_keep_polling++;
1772 		MAC_SRS_POLLING_ON(mac_srs);
1773 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1774 			MAC_SRS_POLL_RING(mac_srs);
1775 		return;
1776 	}
1777 
1778 	/* Nothing else to do. Get out of poll mode */
1779 	MAC_SRS_POLLING_OFF(mac_srs);
1780 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1781 	srs_rx->sr_drain_finish_intr++;
1782 }
1783 
1784 /*
1785  * mac_rx_srs_drain_bw
1786  *
1787  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
1788  * (worker, interrupt, poll) can call this based on processing model.
1789  * The first thing we do is disable interrupts if possible and then
1790  * drain the queue. we also try to poll the underlying hardware if
1791  * there is a dedicated hardware Rx ring assigned to this SRS.
1792  *
1793  * There is a equivalent drain routine in non bandwidth control mode
1794  * mac_rx_srs_drain. There is some code duplication between the two
1795  * routines but they are highly performance sensitive and are easier
1796  * to read/debug if they stay separate. Any code changes here might
1797  * also apply to mac_rx_srs_drain as well.
1798  */
1799 void
1800 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1801 {
1802 	mblk_t 			*head;
1803 	mblk_t			*tail;
1804 	timeout_id_t 		tid;
1805 	size_t			sz = 0;
1806 	int			cnt = 0;
1807 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1808 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1809 
1810 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1811 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
1812 again:
1813 	/* Check if we are doing B/W control */
1814 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1815 	if (mac_srs->srs_bw->mac_bw_curr_time != lbolt) {
1816 		mac_srs->srs_bw->mac_bw_curr_time = lbolt;
1817 		mac_srs->srs_bw->mac_bw_used = 0;
1818 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
1819 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
1820 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
1821 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1822 		goto done;
1823 	} else if (mac_srs->srs_bw->mac_bw_used >
1824 	    mac_srs->srs_bw->mac_bw_limit) {
1825 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
1826 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1827 		goto done;
1828 	}
1829 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1830 
1831 	/* If we are blanked i.e. can't do upcalls, then we are done */
1832 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1833 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1834 		    (mac_srs->srs_state & SRS_PAUSE));
1835 		goto done;
1836 	}
1837 
1838 	sz = 0;
1839 	cnt = 0;
1840 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
1841 		/*
1842 		 * We couldn't pick up a single packet.
1843 		 */
1844 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1845 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
1846 		    (mac_srs->srs_size != 0) &&
1847 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1848 			/*
1849 			 * Seems like configured B/W doesn't
1850 			 * even allow processing of 1 packet
1851 			 * per tick.
1852 			 *
1853 			 * XXX: raise the limit to processing
1854 			 * at least 1 packet per tick.
1855 			 */
1856 			mac_srs->srs_bw->mac_bw_limit +=
1857 			    mac_srs->srs_bw->mac_bw_limit;
1858 			mac_srs->srs_bw->mac_bw_drop_threshold +=
1859 			    mac_srs->srs_bw->mac_bw_drop_threshold;
1860 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
1861 			    "raised B/W limit to %d since not even a "
1862 			    "single packet can be processed per "
1863 			    "tick %d\n", (void *)mac_srs,
1864 			    (int)mac_srs->srs_bw->mac_bw_limit,
1865 			    (int)msgdsize(mac_srs->srs_first));
1866 		}
1867 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1868 		goto done;
1869 	}
1870 
1871 	ASSERT(head != NULL);
1872 	ASSERT(tail != NULL);
1873 
1874 	/* zero bandwidth: drop all and return to interrupt mode */
1875 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1876 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
1877 		srs_rx->sr_drop_count += cnt;
1878 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
1879 		mac_srs->srs_bw->mac_bw_sz -= sz;
1880 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
1881 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1882 		mac_pkt_drop(NULL, NULL, head, B_FALSE);
1883 		goto leave_poll;
1884 	} else {
1885 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1886 	}
1887 
1888 	/*
1889 	 * We can continue processing the queue.
1890 	 * We need to figure out if there is a fanout needed or
1891 	 * we can just process this here.
1892 	 */
1893 
1894 	if ((tid = mac_srs->srs_tid) != 0)
1895 		mac_srs->srs_tid = 0;
1896 
1897 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1898 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
1899 
1900 	/*
1901 	 * mcip is NULL for broadcast and multicast flows. The promisc
1902 	 * callbacks for broadcast and multicast packets are delivered from
1903 	 * mac_rx() and we don't need to worry about that case in this path
1904 	 */
1905 	if (mcip != NULL && mcip->mci_promisc_list != NULL) {
1906 		mutex_exit(&mac_srs->srs_lock);
1907 		mac_promisc_client_dispatch(mcip, head);
1908 		mutex_enter(&mac_srs->srs_lock);
1909 	}
1910 
1911 	/*
1912 	 * Check if SRS itself is doing the processing
1913 	 * This direct path does not apply when subflows are present. In this
1914 	 * case, packets need to be dispatched to a soft ring according to the
1915 	 * flow's bandwidth and other resources contraints.
1916 	 */
1917 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1918 		mac_direct_rx_t		proc;
1919 		void			*arg1;
1920 		mac_resource_handle_t	arg2;
1921 
1922 		/*
1923 		 * This is the case when a Rx is directly
1924 		 * assigned and we have a fully classified
1925 		 * protocol chain. We can deal with it in
1926 		 * one shot.
1927 		 */
1928 		proc = srs_rx->sr_func;
1929 		arg1 = srs_rx->sr_arg1;
1930 		arg2 = srs_rx->sr_arg2;
1931 
1932 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1933 		mutex_exit(&mac_srs->srs_lock);
1934 		if (tid != 0) {
1935 			(void) untimeout(tid);
1936 			tid = 0;
1937 		}
1938 
1939 		proc(arg1, arg2, head, NULL);
1940 		/*
1941 		 * Decrement the size and count here itelf
1942 		 * since the packet has been processed.
1943 		 */
1944 		mutex_enter(&mac_srs->srs_lock);
1945 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1946 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
1947 
1948 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1949 			cv_signal(&mac_srs->srs_client_cv);
1950 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1951 	} else {
1952 		/* Some kind of softrings based fanout is required */
1953 		mutex_exit(&mac_srs->srs_lock);
1954 		if (tid != 0) {
1955 			(void) untimeout(tid);
1956 			tid = 0;
1957 		}
1958 
1959 		/*
1960 		 * Since the fanout routines can deal with chains,
1961 		 * shoot the entire chain up.
1962 		 */
1963 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1964 			mac_rx_srs_fanout(mac_srs, head);
1965 		else
1966 			mac_rx_srs_proto_fanout(mac_srs, head);
1967 		mutex_enter(&mac_srs->srs_lock);
1968 	}
1969 
1970 	/*
1971 	 * Send the poll thread to pick up any packets arrived
1972 	 * so far. This also serves as the last check in case
1973 	 * nothing else is queued in the SRS. The poll thread
1974 	 * is signalled only in the case the drain was done
1975 	 * by the worker thread and SRS_WORKER is set. The
1976 	 * worker thread can run in parallel as long as the
1977 	 * SRS_WORKER flag is set. We we have nothing else to
1978 	 * process, we can exit while leaving SRS_PROC set
1979 	 * which gives the poll thread control to process and
1980 	 * cleanup once it returns from the NIC.
1981 	 *
1982 	 * If we have nothing else to process, we need to
1983 	 * ensure that we keep holding the srs_lock till
1984 	 * all the checks below are done and control is
1985 	 * handed to the poll thread if it was running.
1986 	 */
1987 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1988 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1989 		if (mac_srs->srs_first != NULL) {
1990 			if (proc_type == SRS_WORKER) {
1991 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1992 				if (srs_rx->sr_poll_pkt_cnt <=
1993 				    srs_rx->sr_lowat)
1994 					MAC_SRS_POLL_RING(mac_srs);
1995 				goto again;
1996 			} else {
1997 				cv_signal(&mac_srs->srs_async);
1998 			}
1999 		}
2000 	}
2001 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2002 
2003 done:
2004 
2005 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2006 		/*
2007 		 * Poll thread is already running. Leave the
2008 		 * SRS_RPOC set and hand over the control to
2009 		 * poll thread.
2010 		 */
2011 		mac_srs->srs_state &= ~proc_type;
2012 		return;
2013 	}
2014 
2015 	/*
2016 	 * If we can't process packets because we have exceeded
2017 	 * B/W limit for this tick, just set the timeout
2018 	 * and leave.
2019 	 *
2020 	 * Even if there are no packets queued in SRS, we
2021 	 * need to make sure that the shared counter is
2022 	 * clear and any associated softrings have cleared
2023 	 * all the backlog. Otherwise, leave the interface
2024 	 * in polling mode and the poll thread will get
2025 	 * signalled once the count goes down to zero.
2026 	 *
2027 	 * If someone is already draining the queue (SRS_PROC is
2028 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2029 	 * then it means that drain is already running and we
2030 	 * will turn off polling at that time if there is
2031 	 * no backlog. As long as there are packets queued either
2032 	 * is soft ring set or its soft rings, we will leave
2033 	 * the interface in polling mode.
2034 	 */
2035 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2036 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
2037 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
2038 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
2039 		MAC_SRS_POLLING_ON(mac_srs);
2040 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2041 		if ((mac_srs->srs_first != NULL) &&
2042 		    (mac_srs->srs_tid == NULL))
2043 			mac_srs->srs_tid = timeout(mac_srs_fire,
2044 			    mac_srs, 1);
2045 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2046 		return;
2047 	}
2048 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2049 
2050 leave_poll:
2051 
2052 	/* Nothing else to do. Get out of poll mode */
2053 	MAC_SRS_POLLING_OFF(mac_srs);
2054 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2055 }
2056 
2057 /*
2058  * mac_srs_worker
2059  *
2060  * The SRS worker routine. Drains the queue when no one else is
2061  * processing it.
2062  */
2063 void
2064 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
2065 {
2066 	kmutex_t 		*lock = &mac_srs->srs_lock;
2067 	kcondvar_t 		*async = &mac_srs->srs_async;
2068 	callb_cpr_t		cprinfo;
2069 	boolean_t		bw_ctl_flag;
2070 
2071 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
2072 	mutex_enter(lock);
2073 
2074 start:
2075 	for (;;) {
2076 		bw_ctl_flag = B_FALSE;
2077 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2078 			MAC_SRS_BW_LOCK(mac_srs);
2079 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2080 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2081 				bw_ctl_flag = B_TRUE;
2082 			MAC_SRS_BW_UNLOCK(mac_srs);
2083 		}
2084 		/*
2085 		 * The SRS_BW_ENFORCED flag may change since we have dropped
2086 		 * the mac_bw_lock. However the drain function can handle both
2087 		 * a drainable SRS or a bandwidth controlled SRS, and the
2088 		 * effect of scheduling a timeout is to wakeup the worker
2089 		 * thread which in turn will call the drain function. Since
2090 		 * we release the srs_lock atomically only in the cv_wait there
2091 		 * isn't a fear of waiting for ever.
2092 		 */
2093 		while (((mac_srs->srs_state & SRS_PROC) ||
2094 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
2095 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
2096 		    !(mac_srs->srs_state & SRS_PAUSE)) {
2097 			/*
2098 			 * If we have packets queued and we are here
2099 			 * because B/W control is in place, we better
2100 			 * schedule the worker wakeup after 1 tick
2101 			 * to see if bandwidth control can be relaxed.
2102 			 */
2103 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
2104 				/*
2105 				 * We need to ensure that a timer  is already
2106 				 * scheduled or we force  schedule one for
2107 				 * later so that we can continue processing
2108 				 * after this  quanta is over.
2109 				 */
2110 				mac_srs->srs_tid = timeout(mac_srs_fire,
2111 				    mac_srs, 1);
2112 			}
2113 wait:
2114 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
2115 			cv_wait(async, lock);
2116 			CALLB_CPR_SAFE_END(&cprinfo, lock);
2117 
2118 			if (mac_srs->srs_state & SRS_PAUSE)
2119 				goto done;
2120 			if (mac_srs->srs_state & SRS_PROC)
2121 				goto wait;
2122 
2123 			if (mac_srs->srs_first != NULL &&
2124 			    mac_srs->srs_type & SRST_BW_CONTROL) {
2125 				MAC_SRS_BW_LOCK(mac_srs);
2126 				if (mac_srs->srs_bw->mac_bw_state &
2127 				    SRS_BW_ENFORCED) {
2128 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2129 				}
2130 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
2131 				    SRS_BW_ENFORCED;
2132 				MAC_SRS_BW_UNLOCK(mac_srs);
2133 			}
2134 		}
2135 
2136 		if (mac_srs->srs_state & SRS_PAUSE)
2137 			goto done;
2138 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
2139 	}
2140 done:
2141 	/*
2142 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
2143 	 * from both hard and soft classifications and waits for such threads
2144 	 * to finish before signaling the worker. So at this point the only
2145 	 * thread left that could be competing with the worker is the poll
2146 	 * thread. In the case of Tx, there shouldn't be any thread holding
2147 	 * SRS_PROC at this point.
2148 	 */
2149 	if (!(mac_srs->srs_state & SRS_PROC)) {
2150 		mac_srs->srs_state |= SRS_PROC;
2151 	} else {
2152 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
2153 		/*
2154 		 * Poll thread still owns the SRS and is still running
2155 		 */
2156 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
2157 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2158 		    SRS_POLL_THR_OWNER));
2159 	}
2160 	mac_srs_worker_quiesce(mac_srs);
2161 	/*
2162 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
2163 	 * of the quiesce operation
2164 	 */
2165 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
2166 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2167 
2168 	if (mac_srs->srs_state & SRS_RESTART) {
2169 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2170 		mac_srs_worker_restart(mac_srs);
2171 		mac_srs->srs_state &= ~SRS_PROC;
2172 		goto start;
2173 	}
2174 
2175 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
2176 		mac_srs_worker_quiesce(mac_srs);
2177 
2178 	mac_srs->srs_state &= ~SRS_PROC;
2179 	/* The macro drops the srs_lock */
2180 	CALLB_CPR_EXIT(&cprinfo);
2181 	thread_exit();
2182 }
2183 
2184 /*
2185  * mac_rx_srs_subflow_process
2186  *
2187  * Receive side routine called from interrupt path when there are
2188  * sub flows present on this SRS.
2189  */
2190 /* ARGSUSED */
2191 void
2192 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
2193     mblk_t *mp_chain, boolean_t loopback)
2194 {
2195 	flow_entry_t		*flent = NULL;
2196 	flow_entry_t		*prev_flent = NULL;
2197 	mblk_t			*mp = NULL;
2198 	mblk_t			*tail = NULL;
2199 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2200 	mac_client_impl_t	*mcip;
2201 
2202 	mcip = mac_srs->srs_mcip;
2203 	ASSERT(mcip != NULL);
2204 
2205 	/*
2206 	 * We need to determine the SRS for every packet
2207 	 * by walking the flow table, if we don't get any,
2208 	 * then we proceed using the SRS we came with.
2209 	 */
2210 	mp = tail = mp_chain;
2211 	while (mp != NULL) {
2212 
2213 		/*
2214 		 * We will increment the stats for the mactching subflow.
2215 		 * when we get the bytes/pkt count for the classified packets
2216 		 * later in mac_rx_srs_process.
2217 		 */
2218 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
2219 		    FLOW_INBOUND, &flent);
2220 
2221 		if (mp == mp_chain || flent == prev_flent) {
2222 			if (prev_flent != NULL)
2223 				FLOW_REFRELE(prev_flent);
2224 			prev_flent = flent;
2225 			flent = NULL;
2226 			tail = mp;
2227 			mp = mp->b_next;
2228 			continue;
2229 		}
2230 		tail->b_next = NULL;
2231 		/*
2232 		 * A null indicates, this is for the mac_srs itself.
2233 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
2234 		 */
2235 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2236 			mac_rx_srs_process(arg,
2237 			    (mac_resource_handle_t)mac_srs, mp_chain,
2238 			    loopback);
2239 		} else {
2240 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2241 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
2242 			FLOW_REFRELE(prev_flent);
2243 		}
2244 		prev_flent = flent;
2245 		flent = NULL;
2246 		mp_chain = mp;
2247 		tail = mp;
2248 		mp = mp->b_next;
2249 	}
2250 	/* Last chain */
2251 	ASSERT(mp_chain != NULL);
2252 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2253 		mac_rx_srs_process(arg,
2254 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
2255 	} else {
2256 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2257 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
2258 		FLOW_REFRELE(prev_flent);
2259 	}
2260 }
2261 
2262 /*
2263  * mac_rx_srs_process
2264  *
2265  * Receive side routine called from the interrupt path.
2266  *
2267  * loopback is set to force a context switch on the loopback
2268  * path between MAC clients.
2269  */
2270 /* ARGSUSED */
2271 void
2272 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
2273     boolean_t loopback)
2274 {
2275 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2276 	mblk_t			*mp, *tail, *head;
2277 	int			count = 0;
2278 	int			count1;
2279 	size_t			sz = 0;
2280 	size_t			chain_sz, sz1;
2281 	mac_bw_ctl_t		*mac_bw;
2282 	mac_client_impl_t	*smcip;
2283 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2284 
2285 	/*
2286 	 * Set the tail, count and sz. We set the sz irrespective
2287 	 * of whether we are doing B/W control or not for the
2288 	 * purpose of updating the stats.
2289 	 */
2290 	mp = tail = mp_chain;
2291 	while (mp != NULL) {
2292 		tail = mp;
2293 		count++;
2294 		sz += msgdsize(mp);
2295 		mp = mp->b_next;
2296 	}
2297 
2298 	mutex_enter(&mac_srs->srs_lock);
2299 	smcip = mac_srs->srs_mcip;
2300 
2301 	if (mac_srs->srs_type & SRST_FLOW || smcip == NULL) {
2302 		FLOW_STAT_UPDATE(mac_srs->srs_flent, rbytes, sz);
2303 		FLOW_STAT_UPDATE(mac_srs->srs_flent, ipackets, count);
2304 	}
2305 	if (smcip != NULL) {
2306 		smcip->mci_stat_ibytes += sz;
2307 		smcip->mci_stat_ipackets += count;
2308 	}
2309 
2310 	/*
2311 	 * If the SRS in already being processed; has been blanked;
2312 	 * can be processed by worker thread only; or the B/W limit
2313 	 * has been reached, then queue the chain and check if
2314 	 * worker thread needs to be awakend.
2315 	 */
2316 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
2317 		mac_bw = mac_srs->srs_bw;
2318 		ASSERT(mac_bw != NULL);
2319 		mutex_enter(&mac_bw->mac_bw_lock);
2320 		/* Count the packets and bytes via interrupt */
2321 		srs_rx->sr_intr_count += count;
2322 		mac_bw->mac_bw_intr += sz;
2323 		if (mac_bw->mac_bw_limit == 0) {
2324 			/* zero bandwidth: drop all */
2325 			srs_rx->sr_drop_count += count;
2326 			mac_bw->mac_bw_drop_bytes += sz;
2327 			mutex_exit(&mac_bw->mac_bw_lock);
2328 			mutex_exit(&mac_srs->srs_lock);
2329 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
2330 			return;
2331 		} else {
2332 			if ((mac_bw->mac_bw_sz + sz) <=
2333 			    mac_bw->mac_bw_drop_threshold) {
2334 				mutex_exit(&mac_bw->mac_bw_lock);
2335 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
2336 				    tail, count, sz);
2337 			} else {
2338 				mp = mp_chain;
2339 				chain_sz = 0;
2340 				count1 = 0;
2341 				tail = NULL;
2342 				head = NULL;
2343 				while (mp != NULL) {
2344 					sz1 = msgdsize(mp);
2345 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
2346 					    mac_bw->mac_bw_drop_threshold)
2347 						break;
2348 					chain_sz += sz1;
2349 					count1++;
2350 					tail = mp;
2351 					mp = mp->b_next;
2352 				}
2353 				mutex_exit(&mac_bw->mac_bw_lock);
2354 				if (tail != NULL) {
2355 					head = tail->b_next;
2356 					tail->b_next = NULL;
2357 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
2358 					    mp_chain, tail, count1, chain_sz);
2359 					sz -= chain_sz;
2360 					count -= count1;
2361 				} else {
2362 					/* Can't pick up any */
2363 					head = mp_chain;
2364 				}
2365 				if (head != NULL) {
2366 					/* Drop any packet over the threshold */
2367 					srs_rx->sr_drop_count += count;
2368 					mutex_enter(&mac_bw->mac_bw_lock);
2369 					mac_bw->mac_bw_drop_bytes += sz;
2370 					mutex_exit(&mac_bw->mac_bw_lock);
2371 					freemsgchain(head);
2372 				}
2373 			}
2374 			MAC_SRS_WORKER_WAKEUP(mac_srs);
2375 			mutex_exit(&mac_srs->srs_lock);
2376 			return;
2377 		}
2378 	}
2379 
2380 	/*
2381 	 * If the total number of packets queued in the SRS and
2382 	 * its associated soft rings exceeds the max allowed,
2383 	 * then drop the chain. If we are polling capable, this
2384 	 * shouldn't be happening.
2385 	 */
2386 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
2387 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
2388 		mac_bw = mac_srs->srs_bw;
2389 		srs_rx->sr_drop_count += count;
2390 		mutex_enter(&mac_bw->mac_bw_lock);
2391 		mac_bw->mac_bw_drop_bytes += sz;
2392 		mutex_exit(&mac_bw->mac_bw_lock);
2393 		freemsgchain(mp_chain);
2394 		mutex_exit(&mac_srs->srs_lock);
2395 		return;
2396 	}
2397 
2398 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
2399 	/* Count the packets entering via interrupt path */
2400 	srs_rx->sr_intr_count += count;
2401 
2402 	if (!(mac_srs->srs_state & SRS_PROC)) {
2403 		/*
2404 		 * If we are coming via loopback or if we are not
2405 		 * optimizing for latency, we should signal the
2406 		 * worker thread.
2407 		 */
2408 		if (loopback || ((count > 1) &&
2409 		    !(mac_srs->srs_state & SRS_LATENCY_OPT))) {
2410 			/*
2411 			 * For loopback, We need to let the worker take
2412 			 * over as we don't want to continue in the same
2413 			 * thread even if we can. This could lead to stack
2414 			 * overflows and may also end up using
2415 			 * resources (cpu) incorrectly.
2416 			 */
2417 			cv_signal(&mac_srs->srs_async);
2418 		} else {
2419 			/*
2420 			 * Seems like no one is processing the SRS and
2421 			 * there is no backlog. We also inline process
2422 			 * our packet if its a single packet in non
2423 			 * latency optimized case (in latency optimized
2424 			 * case, we inline process chains of any size).
2425 			 */
2426 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
2427 		}
2428 	}
2429 	mutex_exit(&mac_srs->srs_lock);
2430 }
2431 
2432 /* TX SIDE ROUTINES (RUNTIME) */
2433 
2434 /*
2435  * mac_tx_srs_no_desc
2436  *
2437  * This routine is called by Tx single ring default mode
2438  * when Tx ring runs out of descs.
2439  */
2440 mac_tx_cookie_t
2441 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2442     uint16_t flag, mblk_t **ret_mp)
2443 {
2444 	mac_tx_cookie_t cookie = NULL;
2445 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
2446 	boolean_t wakeup_worker = B_TRUE;
2447 	uint32_t tx_mode = srs_tx->st_mode;
2448 	int cnt, sz;
2449 	mblk_t *tail;
2450 
2451 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
2452 	if (flag & MAC_DROP_ON_NO_DESC) {
2453 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2454 	} else {
2455 		if (mac_srs->srs_first != NULL)
2456 			wakeup_worker = B_FALSE;
2457 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2458 		if (flag & MAC_TX_NO_ENQUEUE) {
2459 			/*
2460 			 * If TX_QUEUED is not set, queue the
2461 			 * packet and let mac_tx_srs_drain()
2462 			 * set the TX_BLOCKED bit for the
2463 			 * reasons explained above. Otherwise,
2464 			 * return the mblks.
2465 			 */
2466 			if (wakeup_worker) {
2467 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2468 				    mp_chain, tail, cnt, sz);
2469 			} else {
2470 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
2471 				    mp_chain, ret_mp, cookie);
2472 			}
2473 		} else {
2474 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2475 			    tail, cnt, sz, cookie);
2476 		}
2477 		if (wakeup_worker)
2478 			cv_signal(&mac_srs->srs_async);
2479 	}
2480 	return (cookie);
2481 }
2482 
2483 /*
2484  * mac_tx_srs_enqueue
2485  *
2486  * This routine is called when Tx SRS is operating in either serializer
2487  * or bandwidth mode. In serializer mode, a packet will get enqueued
2488  * when a thread cannot enter SRS exclusively. In bandwidth mode,
2489  * packets gets queued if allowed byte-count limit for a tick is
2490  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
2491  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
2492  * the default mode or fanout mode. Here packets get dropped or
2493  * returned back to the caller only after hi-watermark worth of data
2494  * is queued.
2495  */
2496 static mac_tx_cookie_t
2497 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2498     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
2499 {
2500 	mac_tx_cookie_t cookie = NULL;
2501 	int cnt, sz;
2502 	mblk_t *tail;
2503 	boolean_t wakeup_worker = B_TRUE;
2504 
2505 	if (mac_srs->srs_first != NULL)
2506 		wakeup_worker = B_FALSE;
2507 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2508 	if (flag & MAC_DROP_ON_NO_DESC) {
2509 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
2510 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2511 		} else {
2512 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2513 			    mp_chain, tail, cnt, sz);
2514 		}
2515 	} else if (flag & MAC_TX_NO_ENQUEUE) {
2516 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
2517 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
2518 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
2519 			    ret_mp, cookie);
2520 		} else {
2521 			mp_chain->b_prev = (mblk_t *)fanout_hint;
2522 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2523 			    mp_chain, tail, cnt, sz);
2524 		}
2525 	} else {
2526 		/*
2527 		 * If you are BW_ENFORCED, just enqueue the
2528 		 * packet. srs_worker will drain it at the
2529 		 * prescribed rate. Before enqueueing, save
2530 		 * the fanout hint.
2531 		 */
2532 		mp_chain->b_prev = (mblk_t *)fanout_hint;
2533 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2534 		    tail, cnt, sz, cookie);
2535 	}
2536 	if (wakeup_worker)
2537 		cv_signal(&mac_srs->srs_async);
2538 	return (cookie);
2539 }
2540 
2541 /*
2542  * There are five tx modes:
2543  *
2544  * 1) Default mode (SRS_TX_DEFAULT)
2545  * 2) Serialization mode (SRS_TX_SERIALIZE)
2546  * 3) Fanout mode (SRS_TX_FANOUT)
2547  * 4) Bandwdith mode (SRS_TX_BW)
2548  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
2549  *
2550  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
2551  * based on the number of Tx rings requested for an SRS and whether
2552  * bandwidth control is requested or not.
2553  *
2554  * In the default mode (i.e., no fanout/no bandwidth), the SRS acts as a
2555  * pass-thru. Packets will go directly to mac_tx_send(). When the underlying
2556  * Tx ring runs out of Tx descs, it starts queueing up packets in SRS.
2557  * When flow-control is relieved, the srs_worker drains the queued
2558  * packets and informs blocked clients to restart sending packets.
2559  *
2560  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized.
2561  *
2562  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
2563  * Tx rings. Each Tx ring will have a soft ring associated with it.
2564  * These soft rings will be hung off the Tx SRS. Queueing if it happens
2565  * due to lack of Tx desc will be in individual soft ring (and not srs)
2566  * associated with Tx ring.
2567  *
2568  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
2569  * only if bw is available. Otherwise the packets will be queued in
2570  * SRS. If fanout to multiple Tx rings is configured, the packets will
2571  * be fanned out among the soft rings associated with the Tx rings.
2572  *
2573  * Four flags are used in srs_state for indicating flow control
2574  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
2575  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
2576  * driver below.
2577  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
2578  * and flow-control pressure is applied back to clients. The clients expect
2579  * wakeup when flow-control is relieved.
2580  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
2581  * got returned back to client either due to lack of Tx descs or due to bw
2582  * control reasons. The clients expect a wakeup when condition is relieved.
2583  *
2584  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
2585  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
2586  * MAC_TX_NO_ENQUEUE
2587  * Mac clients that do not want packets to be enqueued in the mac layer set
2588  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
2589  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
2590  * behaviour of this flag is different when the Tx is running in serializer
2591  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
2592  * get dropped when Tx high watermark is reached.
2593  * There are some mac clients like vsw, aggr that want the mblks to be
2594  * returned back to clients instead of being queued in Tx SRS (or Tx soft
2595  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
2596  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
2597  * In the default and Tx fanout mode, the un-transmitted mblks will be
2598  * returned back to the clients when the driver runs out of Tx descs.
2599  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
2600  * soft ring) so that the clients can be woken up when Tx desc become
2601  * available. When running in serializer or bandwidth mode mode,
2602  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
2603  */
2604 
2605 mac_tx_func_t
2606 mac_tx_get_func(uint32_t mode)
2607 {
2608 	return (mac_tx_mode_list[mode].mac_tx_func);
2609 }
2610 
2611 /* ARGSUSED */
2612 static mac_tx_cookie_t
2613 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2614     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2615 {
2616 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2617 	boolean_t		is_subflow;
2618 	mac_tx_stats_t		stats;
2619 	mac_tx_cookie_t		cookie = NULL;
2620 
2621 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
2622 
2623 	/* Regular case with a single Tx ring */
2624 	/*
2625 	 * SRS_TX_BLOCKED is set when underlying NIC runs
2626 	 * out of Tx descs and messages start getting
2627 	 * queued. It won't get reset until
2628 	 * tx_srs_drain() completely drains out the
2629 	 * messages.
2630 	 */
2631 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2632 		/* Tx descs/resources not available */
2633 		mutex_enter(&mac_srs->srs_lock);
2634 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2635 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
2636 			    flag, ret_mp);
2637 			mutex_exit(&mac_srs->srs_lock);
2638 			return (cookie);
2639 		}
2640 		/*
2641 		 * While we were computing mblk count, the
2642 		 * flow control condition got relieved.
2643 		 * Continue with the transmission.
2644 		 */
2645 		mutex_exit(&mac_srs->srs_lock);
2646 	}
2647 
2648 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2649 
2650 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2651 	    mp_chain, (is_subflow ? &stats : NULL));
2652 
2653 	/*
2654 	 * Multiple threads could be here sending packets.
2655 	 * Under such conditions, it is not possible to
2656 	 * automically set SRS_TX_BLOCKED bit to indicate
2657 	 * out of tx desc condition. To atomically set
2658 	 * this, we queue the returned packet and do
2659 	 * the setting of SRS_TX_BLOCKED in
2660 	 * mac_tx_srs_drain().
2661 	 */
2662 	if (mp_chain != NULL) {
2663 		mutex_enter(&mac_srs->srs_lock);
2664 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
2665 		mutex_exit(&mac_srs->srs_lock);
2666 		return (cookie);
2667 	}
2668 
2669 	if (is_subflow)
2670 		FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2671 
2672 	return (NULL);
2673 }
2674 
2675 /*
2676  * mac_tx_serialize_mode
2677  *
2678  * This is an experimental mode implemented as per the request of PAE.
2679  * In this mode, all callers attempting to send a packet to the NIC
2680  * will get serialized. Only one thread at any time will access the
2681  * NIC to send the packet out.
2682  */
2683 /* ARGSUSED */
2684 static mac_tx_cookie_t
2685 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2686     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2687 {
2688 	boolean_t		is_subflow;
2689 	mac_tx_stats_t		stats;
2690 	mac_tx_cookie_t		cookie = NULL;
2691 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2692 
2693 	/* Single ring, serialize below */
2694 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
2695 	mutex_enter(&mac_srs->srs_lock);
2696 	if ((mac_srs->srs_first != NULL) ||
2697 	    (mac_srs->srs_state & SRS_PROC)) {
2698 		/*
2699 		 * In serialization mode, queue all packets until
2700 		 * TX_HIWAT is set.
2701 		 * If drop bit is set, drop if TX_HIWAT is set.
2702 		 * If no_enqueue is set, still enqueue until hiwat
2703 		 * is set and return mblks after TX_HIWAT is set.
2704 		 */
2705 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
2706 		    flag, NULL, ret_mp);
2707 		mutex_exit(&mac_srs->srs_lock);
2708 		return (cookie);
2709 	}
2710 	/*
2711 	 * No packets queued, nothing on proc and no flow
2712 	 * control condition. Fast-path, ok. Do inline
2713 	 * processing.
2714 	 */
2715 	mac_srs->srs_state |= SRS_PROC;
2716 	mutex_exit(&mac_srs->srs_lock);
2717 
2718 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2719 
2720 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2721 	    mp_chain, (is_subflow ? &stats : NULL));
2722 
2723 	mutex_enter(&mac_srs->srs_lock);
2724 	mac_srs->srs_state &= ~SRS_PROC;
2725 	if (mp_chain != NULL) {
2726 		cookie = mac_tx_srs_enqueue(mac_srs,
2727 		    mp_chain, flag, NULL, ret_mp);
2728 	}
2729 	if (mac_srs->srs_first != NULL) {
2730 		/*
2731 		 * We processed inline our packet and a new
2732 		 * packet/s got queued while we were
2733 		 * processing. Wakeup srs worker
2734 		 */
2735 		cv_signal(&mac_srs->srs_async);
2736 	}
2737 	mutex_exit(&mac_srs->srs_lock);
2738 
2739 	if (is_subflow && cookie == NULL)
2740 		FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2741 
2742 	return (cookie);
2743 }
2744 
2745 /*
2746  * mac_tx_fanout_mode
2747  *
2748  * In this mode, the SRS will have access to multiple Tx rings to send
2749  * the packet out. The fanout hint that is passed as an argument is
2750  * used to find an appropriate ring to fanout the traffic. Each Tx
2751  * ring, in turn,  will have a soft ring associated with it. If a Tx
2752  * ring runs out of Tx desc's the returned packet will be queued in
2753  * the soft ring associated with that Tx ring. The srs itself will not
2754  * queue any packets.
2755  */
2756 static mac_tx_cookie_t
2757 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2758     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2759 {
2760 	mac_soft_ring_t		*softring;
2761 	uint_t			indx, hash;
2762 
2763 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT);
2764 	hash = HASH_HINT(fanout_hint);
2765 	indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
2766 	softring = mac_srs->srs_oth_soft_rings[indx];
2767 	return (mac_tx_soft_ring_process(softring, mp_chain, flag, ret_mp));
2768 }
2769 
2770 /*
2771  * mac_tx_bw_mode
2772  *
2773  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
2774  * only if bw is available. Otherwise the packets will be queued in
2775  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
2776  * out to a Tx rings.
2777  */
2778 static mac_tx_cookie_t
2779 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2780     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2781 {
2782 	int			cnt, sz;
2783 	mblk_t			*tail;
2784 	mac_tx_cookie_t		cookie = NULL;
2785 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2786 
2787 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
2788 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2789 	mutex_enter(&mac_srs->srs_lock);
2790 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2791 		/* zero bandwidth: drop all */
2792 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2793 		mutex_exit(&mac_srs->srs_lock);
2794 		return (cookie);
2795 	} else if ((mac_srs->srs_first != NULL) ||
2796 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2797 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2798 		    fanout_hint, ret_mp);
2799 		mutex_exit(&mac_srs->srs_lock);
2800 		return (cookie);
2801 	}
2802 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2803 	if (mac_srs->srs_bw->mac_bw_curr_time != lbolt) {
2804 		mac_srs->srs_bw->mac_bw_curr_time = lbolt;
2805 		mac_srs->srs_bw->mac_bw_used = 0;
2806 	} else if (mac_srs->srs_bw->mac_bw_used >
2807 	    mac_srs->srs_bw->mac_bw_limit) {
2808 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2809 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2810 		    mp_chain, tail, cnt, sz);
2811 		/*
2812 		 * Wakeup worker thread. Note that worker
2813 		 * thread has to be woken up so that it
2814 		 * can fire up the timer to be woken up
2815 		 * on the next tick. Also once
2816 		 * BW_ENFORCED is set, it can only be
2817 		 * reset by srs_worker thread. Until then
2818 		 * all packets will get queued up in SRS
2819 		 * and hence this this code path won't be
2820 		 * entered until BW_ENFORCED is reset.
2821 		 */
2822 		cv_signal(&mac_srs->srs_async);
2823 		mutex_exit(&mac_srs->srs_lock);
2824 		return (cookie);
2825 	}
2826 
2827 	mac_srs->srs_bw->mac_bw_used += sz;
2828 	mutex_exit(&mac_srs->srs_lock);
2829 
2830 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
2831 		mac_soft_ring_t *softring;
2832 		uint_t indx, hash;
2833 
2834 		hash = HASH_HINT(fanout_hint);
2835 		indx = COMPUTE_INDEX(hash,
2836 		    mac_srs->srs_oth_ring_count);
2837 		softring = mac_srs->srs_oth_soft_rings[indx];
2838 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
2839 		    ret_mp));
2840 	} else {
2841 		boolean_t		is_subflow;
2842 		mac_tx_stats_t		stats;
2843 
2844 		is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2845 
2846 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2847 		    mp_chain, (is_subflow ? &stats : NULL));
2848 
2849 		if (mp_chain != NULL) {
2850 			mutex_enter(&mac_srs->srs_lock);
2851 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2852 			if (mac_srs->srs_bw->mac_bw_used > sz)
2853 				mac_srs->srs_bw->mac_bw_used -= sz;
2854 			else
2855 				mac_srs->srs_bw->mac_bw_used = 0;
2856 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2857 			    fanout_hint, ret_mp);
2858 			mutex_exit(&mac_srs->srs_lock);
2859 			return (cookie);
2860 		}
2861 		if (is_subflow)
2862 			FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
2863 
2864 		return (NULL);
2865 	}
2866 }
2867 
2868 /* ARGSUSED */
2869 void
2870 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
2871 {
2872 	mblk_t			*head, *tail;
2873 	size_t			sz;
2874 	uint32_t		tx_mode;
2875 	uint_t			saved_pkt_count;
2876 	boolean_t		is_subflow;
2877 	mac_tx_stats_t		stats;
2878 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2879 
2880 	saved_pkt_count = 0;
2881 	ASSERT(mutex_owned(&mac_srs->srs_lock));
2882 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
2883 
2884 	mac_srs->srs_state |= SRS_PROC;
2885 
2886 	is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
2887 	tx_mode = srs_tx->st_mode;
2888 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
2889 		if (mac_srs->srs_first != NULL) {
2890 			head = mac_srs->srs_first;
2891 			tail = mac_srs->srs_last;
2892 			saved_pkt_count = mac_srs->srs_count;
2893 			mac_srs->srs_first = NULL;
2894 			mac_srs->srs_last = NULL;
2895 			mac_srs->srs_count = 0;
2896 			mutex_exit(&mac_srs->srs_lock);
2897 
2898 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2899 			    head, &stats);
2900 
2901 			mutex_enter(&mac_srs->srs_lock);
2902 			if (head != NULL) {
2903 				/* Device out of tx desc, set block */
2904 				if (head->b_next == NULL)
2905 					VERIFY(head == tail);
2906 				tail->b_next = mac_srs->srs_first;
2907 				mac_srs->srs_first = head;
2908 				mac_srs->srs_count +=
2909 				    (saved_pkt_count - stats.ts_opackets);
2910 				if (mac_srs->srs_last == NULL)
2911 					mac_srs->srs_last = tail;
2912 				MAC_TX_SRS_BLOCK(mac_srs, head);
2913 			} else {
2914 				srs_tx->st_woken_up = B_FALSE;
2915 				if (is_subflow) {
2916 					FLOW_TX_STATS_UPDATE(
2917 					    mac_srs->srs_flent, &stats);
2918 				}
2919 			}
2920 		}
2921 	} else if (tx_mode == SRS_TX_BW) {
2922 		/*
2923 		 * We are here because the timer fired and we have some data
2924 		 * to tranmit. Also mac_tx_srs_worker should have reset
2925 		 * SRS_BW_ENFORCED flag
2926 		 */
2927 		ASSERT(!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED));
2928 		head = tail = mac_srs->srs_first;
2929 		while (mac_srs->srs_first != NULL) {
2930 			tail = mac_srs->srs_first;
2931 			tail->b_prev = NULL;
2932 			mac_srs->srs_first = tail->b_next;
2933 			if (mac_srs->srs_first == NULL)
2934 				mac_srs->srs_last = NULL;
2935 			mac_srs->srs_count--;
2936 			sz = msgdsize(tail);
2937 			mac_srs->srs_size -= sz;
2938 			saved_pkt_count++;
2939 			MAC_TX_UPDATE_BW_INFO(mac_srs, sz);
2940 
2941 			if (mac_srs->srs_bw->mac_bw_used <
2942 			    mac_srs->srs_bw->mac_bw_limit)
2943 				continue;
2944 
2945 			if (mac_srs->srs_bw->mac_bw_curr_time != lbolt) {
2946 				mac_srs->srs_bw->mac_bw_curr_time = lbolt;
2947 				mac_srs->srs_bw->mac_bw_used = sz;
2948 				continue;
2949 			}
2950 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2951 			break;
2952 		}
2953 
2954 		ASSERT((head == NULL && tail == NULL) ||
2955 		    (head != NULL && tail != NULL));
2956 		if (tail != NULL) {
2957 			tail->b_next = NULL;
2958 			mutex_exit(&mac_srs->srs_lock);
2959 
2960 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2961 			    head, &stats);
2962 
2963 			mutex_enter(&mac_srs->srs_lock);
2964 			if (head != NULL) {
2965 				uint_t size_sent;
2966 
2967 				/* Device out of tx desc, set block */
2968 				if (head->b_next == NULL)
2969 					VERIFY(head == tail);
2970 				tail->b_next = mac_srs->srs_first;
2971 				mac_srs->srs_first = head;
2972 				mac_srs->srs_count +=
2973 				    (saved_pkt_count - stats.ts_opackets);
2974 				if (mac_srs->srs_last == NULL)
2975 					mac_srs->srs_last = tail;
2976 				size_sent = sz - stats.ts_obytes;
2977 				mac_srs->srs_size += size_sent;
2978 				mac_srs->srs_bw->mac_bw_sz += size_sent;
2979 				if (mac_srs->srs_bw->mac_bw_used > size_sent) {
2980 					mac_srs->srs_bw->mac_bw_used -=
2981 					    size_sent;
2982 				} else {
2983 					mac_srs->srs_bw->mac_bw_used = 0;
2984 				}
2985 				MAC_TX_SRS_BLOCK(mac_srs, head);
2986 			} else {
2987 				srs_tx->st_woken_up = B_FALSE;
2988 				if (is_subflow) {
2989 					FLOW_TX_STATS_UPDATE(
2990 					    mac_srs->srs_flent, &stats);
2991 				}
2992 			}
2993 		}
2994 	} else if (tx_mode == SRS_TX_BW_FANOUT) {
2995 		mblk_t *prev;
2996 		mac_soft_ring_t *softring;
2997 		uint64_t hint;
2998 
2999 		/*
3000 		 * We are here because the timer fired and we
3001 		 * have some quota to tranmit.
3002 		 */
3003 		prev = NULL;
3004 		head = tail = mac_srs->srs_first;
3005 		while (mac_srs->srs_first != NULL) {
3006 			tail = mac_srs->srs_first;
3007 			mac_srs->srs_first = tail->b_next;
3008 			if (mac_srs->srs_first == NULL)
3009 				mac_srs->srs_last = NULL;
3010 			mac_srs->srs_count--;
3011 			sz = msgdsize(tail);
3012 			mac_srs->srs_size -= sz;
3013 			mac_srs->srs_bw->mac_bw_used += sz;
3014 			if (prev == NULL)
3015 				hint = (ulong_t)tail->b_prev;
3016 			if (hint != (ulong_t)tail->b_prev) {
3017 				prev->b_next = NULL;
3018 				mutex_exit(&mac_srs->srs_lock);
3019 				TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3020 				head = tail;
3021 				hint = (ulong_t)tail->b_prev;
3022 				mutex_enter(&mac_srs->srs_lock);
3023 			}
3024 
3025 			prev = tail;
3026 			tail->b_prev = NULL;
3027 			if (mac_srs->srs_bw->mac_bw_used <
3028 			    mac_srs->srs_bw->mac_bw_limit)
3029 				continue;
3030 
3031 			if (mac_srs->srs_bw->mac_bw_curr_time != lbolt) {
3032 				mac_srs->srs_bw->mac_bw_curr_time = lbolt;
3033 				mac_srs->srs_bw->mac_bw_used = 0;
3034 				continue;
3035 			}
3036 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3037 			break;
3038 		}
3039 		ASSERT((head == NULL && tail == NULL) ||
3040 		    (head != NULL && tail != NULL));
3041 		if (tail != NULL) {
3042 			tail->b_next = NULL;
3043 			mutex_exit(&mac_srs->srs_lock);
3044 			TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3045 			mutex_enter(&mac_srs->srs_lock);
3046 		}
3047 	}
3048 	/*
3049 	 * SRS_TX_FANOUT case not considered here because packets
3050 	 * won't be queued in the SRS for this case. Packets will
3051 	 * be sent directly to soft rings underneath and if there
3052 	 * is any queueing at all, it would be in Tx side soft
3053 	 * rings.
3054 	 */
3055 
3056 	/*
3057 	 * When srs_count becomes 0, reset SRS_TX_HIWAT and
3058 	 * SRS_TX_WAKEUP_CLIENT and wakeup registered clients.
3059 	 */
3060 	if (mac_srs->srs_count == 0 && (mac_srs->srs_state &
3061 	    (SRS_TX_HIWAT | SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED))) {
3062 		mac_tx_notify_cb_t *mtnfp;
3063 		mac_cb_t *mcb;
3064 		mac_client_impl_t *mcip = mac_srs->srs_mcip;
3065 		boolean_t wakeup_required = B_FALSE;
3066 
3067 		if (mac_srs->srs_state &
3068 		    (SRS_TX_HIWAT|SRS_TX_WAKEUP_CLIENT)) {
3069 			wakeup_required = B_TRUE;
3070 		}
3071 		mac_srs->srs_state &= ~(SRS_TX_HIWAT |
3072 		    SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED);
3073 		mutex_exit(&mac_srs->srs_lock);
3074 		if (wakeup_required) {
3075 			/* Wakeup callback registered clients */
3076 			MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
3077 			for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
3078 			    mcb = mcb->mcb_nextp) {
3079 				mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
3080 				mtnfp->mtnf_fn(mtnfp->mtnf_arg,
3081 				    (mac_tx_cookie_t)mac_srs);
3082 			}
3083 			MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
3084 			    &mcip->mci_tx_notify_cb_list);
3085 			/*
3086 			 * If the client is not the primary MAC client, then we
3087 			 * need to send the notification to the clients upper
3088 			 * MAC, i.e. mci_upper_mip.
3089 			 */
3090 			mac_tx_notify(mcip->mci_upper_mip != NULL ?
3091 			    mcip->mci_upper_mip : mcip->mci_mip);
3092 		}
3093 		mutex_enter(&mac_srs->srs_lock);
3094 	}
3095 	mac_srs->srs_state &= ~SRS_PROC;
3096 }
3097 
3098 /*
3099  * Given a packet, get the flow_entry that identifies the flow
3100  * to which that packet belongs. The flow_entry will contain
3101  * the transmit function to be used to send the packet. If the
3102  * function returns NULL, the packet should be sent using the
3103  * underlying NIC.
3104  */
3105 static flow_entry_t *
3106 mac_tx_classify(mac_impl_t *mip, mblk_t *mp)
3107 {
3108 	flow_entry_t		*flent = NULL;
3109 	mac_client_impl_t	*mcip;
3110 	int	err;
3111 
3112 	/*
3113 	 * Do classification on the packet.
3114 	 */
3115 	err = mac_flow_lookup(mip->mi_flow_tab, mp, FLOW_OUTBOUND, &flent);
3116 	if (err != 0)
3117 		return (NULL);
3118 
3119 	/*
3120 	 * This flent might just be an additional one on the MAC client,
3121 	 * i.e. for classification purposes (different fdesc), however
3122 	 * the resources, SRS et. al., are in the mci_flent, so if
3123 	 * this isn't the mci_flent, we need to get it.
3124 	 */
3125 	if ((mcip = flent->fe_mcip) != NULL && mcip->mci_flent != flent) {
3126 		FLOW_REFRELE(flent);
3127 		flent = mcip->mci_flent;
3128 		FLOW_TRY_REFHOLD(flent, err);
3129 		if (err != 0)
3130 			return (NULL);
3131 	}
3132 
3133 	return (flent);
3134 }
3135 
3136 /*
3137  * This macro is only meant to be used by mac_tx_send().
3138  */
3139 #define	CHECK_VID_AND_ADD_TAG(mp) {			\
3140 	if (vid_check) {				\
3141 		int err = 0;				\
3142 							\
3143 		MAC_VID_CHECK(src_mcip, (mp), err);	\
3144 		if (err != 0) {				\
3145 			freemsg((mp));			\
3146 			(mp) = next;			\
3147 			oerrors++;			\
3148 			continue;			\
3149 		}					\
3150 	}						\
3151 	if (add_tag) {					\
3152 		(mp) = mac_add_vlan_tag((mp), 0, vid);	\
3153 		if ((mp) == NULL) {			\
3154 			(mp) = next;			\
3155 			oerrors++;			\
3156 			continue;			\
3157 		}					\
3158 	}						\
3159 }
3160 
3161 mblk_t *
3162 mac_tx_send(mac_client_handle_t mch, mac_ring_handle_t ring, mblk_t *mp_chain,
3163     mac_tx_stats_t *stats)
3164 {
3165 	mac_client_impl_t *src_mcip = (mac_client_impl_t *)mch;
3166 	mac_impl_t *mip = src_mcip->mci_mip;
3167 	uint_t obytes = 0, opackets = 0, oerrors = 0;
3168 	mblk_t *mp = NULL, *next;
3169 	boolean_t vid_check, add_tag;
3170 	uint16_t vid = 0;
3171 
3172 	if (mip->mi_nclients > 1) {
3173 		vid_check = MAC_VID_CHECK_NEEDED(src_mcip);
3174 		add_tag = MAC_TAG_NEEDED(src_mcip);
3175 		if (add_tag)
3176 			vid = mac_client_vid(mch);
3177 	} else {
3178 		ASSERT(mip->mi_nclients == 1);
3179 		vid_check = add_tag = B_FALSE;
3180 	}
3181 
3182 	/*
3183 	 * Fastpath: if there's only one client, and there's no
3184 	 * multicast listeners, we simply send the packet down to the
3185 	 * underlying NIC.
3186 	 */
3187 	if (mip->mi_nactiveclients == 1 && mip->mi_promisc_list == NULL)  {
3188 		DTRACE_PROBE2(fastpath,
3189 		    mac_client_impl_t *, src_mcip, mblk_t *, mp_chain);
3190 
3191 		mp = mp_chain;
3192 		while (mp != NULL) {
3193 			next = mp->b_next;
3194 			mp->b_next = NULL;
3195 			opackets++;
3196 			obytes += (mp->b_cont == NULL ? MBLKL(mp) :
3197 			    msgdsize(mp));
3198 
3199 			CHECK_VID_AND_ADD_TAG(mp);
3200 			MAC_TX(mip, ring, mp, src_mcip);
3201 
3202 			/*
3203 			 * If the driver is out of descriptors and does a
3204 			 * partial send it will return a chain of unsent
3205 			 * mblks. Adjust the accounting stats.
3206 			 */
3207 			if (mp != NULL) {
3208 				opackets--;
3209 				obytes -= msgdsize(mp);
3210 				mp->b_next = next;
3211 				break;
3212 			}
3213 			mp = next;
3214 		}
3215 		goto done;
3216 	}
3217 
3218 	/*
3219 	 * No fastpath, we either have more than one MAC client
3220 	 * defined on top of the same MAC, or one or more MAC
3221 	 * client promiscuous callbacks.
3222 	 */
3223 	DTRACE_PROBE3(slowpath, mac_client_impl_t *,
3224 	    src_mcip, int, mip->mi_nclients, mblk_t *, mp_chain);
3225 
3226 	if (mip->mi_promisc_list != NULL)
3227 		mac_promisc_dispatch(mip, mp_chain, src_mcip);
3228 
3229 	mp = mp_chain;
3230 	while (mp != NULL) {
3231 		flow_entry_t *dst_flow_ent;
3232 		void *flow_cookie;
3233 		size_t	pkt_size;
3234 		mblk_t *mp1;
3235 
3236 		next = mp->b_next;
3237 		mp->b_next = NULL;
3238 		opackets++;
3239 		pkt_size = (mp->b_cont == NULL ? MBLKL(mp) : msgdsize(mp));
3240 		obytes += pkt_size;
3241 		CHECK_VID_AND_ADD_TAG(mp);
3242 
3243 		/*
3244 		 * Find the destination.
3245 		 */
3246 		dst_flow_ent = mac_tx_classify(mip, mp);
3247 
3248 		if (dst_flow_ent != NULL) {
3249 			size_t	hdrsize;
3250 			int	err = 0;
3251 
3252 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
3253 				struct ether_vlan_header *evhp =
3254 				    (struct ether_vlan_header *)mp->b_rptr;
3255 
3256 				if (ntohs(evhp->ether_tpid) == ETHERTYPE_VLAN)
3257 					hdrsize = sizeof (*evhp);
3258 				else
3259 					hdrsize = sizeof (struct ether_header);
3260 			} else {
3261 				mac_header_info_t	mhi;
3262 
3263 				err = mac_header_info((mac_handle_t)mip,
3264 				    mp, &mhi);
3265 				if (err == 0)
3266 					hdrsize = mhi.mhi_hdrsize;
3267 			}
3268 
3269 			/*
3270 			 * Got a matching flow. It's either another
3271 			 * MAC client, or a broadcast/multicast flow.
3272 			 * Make sure the packet size is within the
3273 			 * allowed size. If not drop the packet and
3274 			 * move to next packet.
3275 			 */
3276 			if (err != 0 ||
3277 			    (pkt_size - hdrsize) > mip->mi_sdu_max) {
3278 				oerrors++;
3279 				DTRACE_PROBE2(loopback__drop, size_t, pkt_size,
3280 				    mblk_t *, mp);
3281 				freemsg(mp);
3282 				mp = next;
3283 				FLOW_REFRELE(dst_flow_ent);
3284 				continue;
3285 			}
3286 			flow_cookie = mac_flow_get_client_cookie(dst_flow_ent);
3287 			if (flow_cookie != NULL) {
3288 				/*
3289 				 * The vnic_bcast_send function expects
3290 				 * to receive the sender MAC client
3291 				 * as value for arg2.
3292 				 */
3293 				mac_bcast_send(flow_cookie, src_mcip, mp,
3294 				    B_TRUE);
3295 			} else {
3296 				/*
3297 				 * loopback the packet to a
3298 				 * local MAC client. We force a context
3299 				 * switch if both source and destination
3300 				 * MAC clients are used by IP, i.e. bypass
3301 				 * is set.
3302 				 */
3303 				boolean_t do_switch;
3304 				mac_client_impl_t *dst_mcip =
3305 				    dst_flow_ent->fe_mcip;
3306 
3307 				do_switch = ((src_mcip->mci_state_flags &
3308 				    dst_mcip->mci_state_flags &
3309 				    MCIS_CLIENT_POLL_CAPABLE) != 0);
3310 
3311 				if ((mp1 = mac_fix_cksum(mp)) != NULL) {
3312 					(dst_flow_ent->fe_cb_fn)(
3313 					    dst_flow_ent->fe_cb_arg1,
3314 					    dst_flow_ent->fe_cb_arg2,
3315 					    mp1, do_switch);
3316 				}
3317 			}
3318 			FLOW_REFRELE(dst_flow_ent);
3319 		} else {
3320 			/*
3321 			 * Unknown destination, send via the underlying
3322 			 * NIC.
3323 			 */
3324 			MAC_TX(mip, ring, mp, src_mcip);
3325 			if (mp != NULL) {
3326 				/*
3327 				 * Adjust for the last packet that
3328 				 * could not be transmitted
3329 				 */
3330 				opackets--;
3331 				obytes -= pkt_size;
3332 				mp->b_next = next;
3333 				break;
3334 			}
3335 		}
3336 		mp = next;
3337 	}
3338 
3339 done:
3340 	src_mcip->mci_stat_obytes += obytes;
3341 	src_mcip->mci_stat_opackets += opackets;
3342 	src_mcip->mci_stat_oerrors += oerrors;
3343 
3344 	if (stats != NULL) {
3345 		stats->ts_opackets = opackets;
3346 		stats->ts_obytes = obytes;
3347 		stats->ts_oerrors = oerrors;
3348 	}
3349 	return (mp);
3350 }
3351 
3352 /*
3353  * mac_tx_srs_ring_present
3354  *
3355  * Returns whether the specified ring is part of the specified SRS.
3356  */
3357 boolean_t
3358 mac_tx_srs_ring_present(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3359 {
3360 	int i;
3361 	mac_soft_ring_t *soft_ring;
3362 
3363 	if (srs->srs_tx.st_arg2 == tx_ring)
3364 		return (B_TRUE);
3365 
3366 	for (i = 0; i < srs->srs_oth_ring_count; i++) {
3367 		soft_ring =  srs->srs_oth_soft_rings[i];
3368 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3369 			return (B_TRUE);
3370 	}
3371 
3372 	return (B_FALSE);
3373 }
3374 
3375 /*
3376  * mac_tx_srs_wakeup
3377  *
3378  * Called when Tx desc become available. Wakeup the appropriate worker
3379  * thread after resetting the SRS_TX_BLOCKED/S_RING_BLOCK bit in the
3380  * state field.
3381  */
3382 void
3383 mac_tx_srs_wakeup(mac_soft_ring_set_t *mac_srs, mac_ring_handle_t ring)
3384 {
3385 	int i;
3386 	mac_soft_ring_t *sringp;
3387 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3388 
3389 	mutex_enter(&mac_srs->srs_lock);
3390 	if (TX_SINGLE_RING_MODE(mac_srs)) {
3391 		if (srs_tx->st_arg2 == ring &&
3392 		    mac_srs->srs_state & SRS_TX_BLOCKED) {
3393 			mac_srs->srs_state &= ~SRS_TX_BLOCKED;
3394 			srs_tx->st_unblocked_cnt++;
3395 			cv_signal(&mac_srs->srs_async);
3396 		}
3397 		/*
3398 		 * A wakeup can come before tx_srs_drain() could
3399 		 * grab srs lock and set SRS_TX_BLOCKED. So
3400 		 * always set woken_up flag when we come here.
3401 		 */
3402 		srs_tx->st_woken_up = B_TRUE;
3403 		mutex_exit(&mac_srs->srs_lock);
3404 		return;
3405 	}
3406 
3407 	/* If you are here, it is for FANOUT or BW_FANOUT case */
3408 	ASSERT(TX_MULTI_RING_MODE(mac_srs));
3409 	for (i = 0; i < mac_srs->srs_oth_ring_count; i++) {
3410 		sringp = mac_srs->srs_oth_soft_rings[i];
3411 		mutex_enter(&sringp->s_ring_lock);
3412 		if (sringp->s_ring_tx_arg2 == ring) {
3413 			if (sringp->s_ring_state & S_RING_BLOCK) {
3414 				sringp->s_ring_state &= ~S_RING_BLOCK;
3415 				sringp->s_ring_unblocked_cnt++;
3416 				cv_signal(&sringp->s_ring_async);
3417 			}
3418 			sringp->s_ring_tx_woken_up = B_TRUE;
3419 		}
3420 		mutex_exit(&sringp->s_ring_lock);
3421 	}
3422 	mutex_exit(&mac_srs->srs_lock);
3423 }
3424 
3425 /*
3426  * Once the driver is done draining, send a MAC_NOTE_TX notification to unleash
3427  * the blocked clients again.
3428  */
3429 void
3430 mac_tx_notify(mac_impl_t *mip)
3431 {
3432 	i_mac_notify(mip, MAC_NOTE_TX);
3433 }
3434 
3435 /*
3436  * RX SOFTRING RELATED FUNCTIONS
3437  *
3438  * These functions really belong in mac_soft_ring.c and here for
3439  * a short period.
3440  */
3441 
3442 #define	SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3443 	/*								\
3444 	 * Enqueue our mblk chain.					\
3445 	 */								\
3446 	ASSERT(MUTEX_HELD(&(ringp)->s_ring_lock));			\
3447 									\
3448 	if ((ringp)->s_ring_last != NULL)				\
3449 		(ringp)->s_ring_last->b_next = (mp);			\
3450 	else								\
3451 		(ringp)->s_ring_first = (mp);				\
3452 	(ringp)->s_ring_last = (tail);					\
3453 	(ringp)->s_ring_count += (cnt);					\
3454 	ASSERT((ringp)->s_ring_count > 0);				\
3455 	if ((ringp)->s_ring_type & ST_RING_BW_CTL) {			\
3456 		(ringp)->s_ring_size += sz;				\
3457 	}								\
3458 }
3459 
3460 /*
3461  * Default entry point to deliver a packet chain to a MAC client.
3462  * If the MAC client has flows, do the classification with these
3463  * flows as well.
3464  */
3465 /* ARGSUSED */
3466 void
3467 mac_rx_deliver(void *arg1, mac_resource_handle_t mrh, mblk_t *mp_chain,
3468     mac_header_info_t *arg3)
3469 {
3470 	mac_client_impl_t *mcip = arg1;
3471 
3472 	if (mcip->mci_nvids == 1 &&
3473 	    !(mcip->mci_state_flags & MCIS_TAG_DISABLE)) {
3474 		/*
3475 		 * If the client has exactly one VID associated with it
3476 		 * and striping of VLAN header is not disabled,
3477 		 * remove the VLAN tag from the packet before
3478 		 * passing it on to the client's receive callback.
3479 		 * Note that this needs to be done after we dispatch
3480 		 * the packet to the promiscuous listeners of the
3481 		 * client, since they expect to see the whole
3482 		 * frame including the VLAN headers.
3483 		 */
3484 		mp_chain = mac_strip_vlan_tag_chain(mp_chain);
3485 	}
3486 
3487 	mcip->mci_rx_fn(mcip->mci_rx_arg, mrh, mp_chain, B_FALSE);
3488 }
3489 
3490 /*
3491  * mac_rx_soft_ring_process
3492  *
3493  * process a chain for a given soft ring. The number of packets queued
3494  * in the SRS and its associated soft rings (including this one) is
3495  * very small (tracked by srs_poll_pkt_cnt), then allow the entering
3496  * thread (interrupt or poll thread) to do inline processing. This
3497  * helps keep the latency down under low load.
3498  *
3499  * The proc and arg for each mblk is already stored in the mblk in
3500  * appropriate places.
3501  */
3502 /* ARGSUSED */
3503 void
3504 mac_rx_soft_ring_process(mac_client_impl_t *mcip, mac_soft_ring_t *ringp,
3505     mblk_t *mp_chain, mblk_t *tail, int cnt, size_t sz)
3506 {
3507 	mac_direct_rx_t		proc;
3508 	void			*arg1;
3509 	mac_resource_handle_t	arg2;
3510 	mac_soft_ring_set_t	*mac_srs = ringp->s_ring_set;
3511 
3512 	ASSERT(ringp != NULL);
3513 	ASSERT(mp_chain != NULL);
3514 	ASSERT(tail != NULL);
3515 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3516 
3517 	mutex_enter(&ringp->s_ring_lock);
3518 	ringp->s_ring_total_inpkt += cnt;
3519 	if ((ringp->s_ring_type & ST_RING_ANY) ||
3520 	    ((mac_srs->srs_rx.sr_poll_pkt_cnt <= 1) &&
3521 	    !mac_srs->srs_rx.sr_enqueue_always)) {
3522 		/* If on processor or blanking on, then enqueue and return */
3523 		if (ringp->s_ring_state & S_RING_BLANK ||
3524 		    ringp->s_ring_state & S_RING_PROC) {
3525 			SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3526 			mutex_exit(&ringp->s_ring_lock);
3527 			return;
3528 		}
3529 
3530 		proc = ringp->s_ring_rx_func;
3531 		arg1 = ringp->s_ring_rx_arg1;
3532 		arg2 = ringp->s_ring_rx_arg2;
3533 		/*
3534 		 * See if anything is already queued. If we are the
3535 		 * first packet, do inline processing else queue the
3536 		 * packet and do the drain.
3537 		 */
3538 		if (ringp->s_ring_first == NULL) {
3539 			/*
3540 			 * Fast-path, ok to process and nothing queued.
3541 			 */
3542 			ringp->s_ring_run = curthread;
3543 			ringp->s_ring_state |= (S_RING_PROC);
3544 
3545 			mutex_exit(&ringp->s_ring_lock);
3546 
3547 			/*
3548 			 * We are the chain of 1 packet so
3549 			 * go through this fast path.
3550 			 */
3551 			ASSERT(mp_chain->b_next == NULL);
3552 
3553 			(*proc)(arg1, arg2, mp_chain, NULL);
3554 
3555 			ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3556 			/*
3557 			 * If we have a soft ring set which is doing
3558 			 * bandwidth control, we need to decrement
3559 			 * srs_size and count so it the SRS can have a
3560 			 * accurate idea of what is the real data
3561 			 * queued between SRS and its soft rings. We
3562 			 * decrement the counters only when the packet
3563 			 * gets processed by both SRS and the soft ring.
3564 			 */
3565 			mutex_enter(&mac_srs->srs_lock);
3566 			MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
3567 			MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
3568 			mutex_exit(&mac_srs->srs_lock);
3569 
3570 			mutex_enter(&ringp->s_ring_lock);
3571 			ringp->s_ring_run = NULL;
3572 			ringp->s_ring_state &= ~S_RING_PROC;
3573 			if (ringp->s_ring_state & S_RING_CLIENT_WAIT)
3574 				cv_signal(&ringp->s_ring_client_cv);
3575 
3576 			if ((ringp->s_ring_first == NULL) ||
3577 			    (ringp->s_ring_state & S_RING_BLANK)) {
3578 				/*
3579 				 * We processed inline our packet and
3580 				 * nothing new has arrived or our
3581 				 * receiver doesn't want to receive
3582 				 * any packets. We are done.
3583 				 */
3584 				mutex_exit(&ringp->s_ring_lock);
3585 				return;
3586 			}
3587 		} else {
3588 			SOFT_RING_ENQUEUE_CHAIN(ringp,
3589 			    mp_chain, tail, cnt, sz);
3590 		}
3591 
3592 		/*
3593 		 * We are here because either we couldn't do inline
3594 		 * processing (because something was already
3595 		 * queued), or we had a chain of more than one
3596 		 * packet, or something else arrived after we were
3597 		 * done with inline processing.
3598 		 */
3599 		ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3600 		ASSERT(ringp->s_ring_first != NULL);
3601 
3602 		ringp->s_ring_drain_func(ringp);
3603 		mutex_exit(&ringp->s_ring_lock);
3604 		return;
3605 	} else {
3606 		/* ST_RING_WORKER_ONLY case */
3607 		SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3608 		mac_soft_ring_worker_wakeup(ringp);
3609 		mutex_exit(&ringp->s_ring_lock);
3610 	}
3611 }
3612 
3613 /*
3614  * TX SOFTRING RELATED FUNCTIONS
3615  *
3616  * These functions really belong in mac_soft_ring.c and here for
3617  * a short period.
3618  */
3619 
3620 #define	TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3621 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));			\
3622 	ringp->s_ring_state |= S_RING_ENQUEUED;				\
3623 	SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);	\
3624 }
3625 
3626 /*
3627  * mac_tx_sring_queued
3628  *
3629  * When we are out of transmit descriptors and we already have a
3630  * queue that exceeds hiwat (or the client called us with
3631  * MAC_TX_NO_ENQUEUE or MAC_DROP_ON_NO_DESC flag), return the
3632  * soft ring pointer as the opaque cookie for the client enable
3633  * flow control.
3634  */
3635 static mac_tx_cookie_t
3636 mac_tx_sring_enqueue(mac_soft_ring_t *ringp, mblk_t *mp_chain, uint16_t flag,
3637     mblk_t **ret_mp)
3638 {
3639 	int cnt;
3640 	size_t sz;
3641 	mblk_t *tail;
3642 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3643 	mac_tx_cookie_t cookie = NULL;
3644 	boolean_t wakeup_worker = B_TRUE;
3645 
3646 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3647 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3648 	if (flag & MAC_DROP_ON_NO_DESC) {
3649 		mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
3650 		/* increment freed stats */
3651 		ringp->s_ring_drops += cnt;
3652 		cookie = (mac_tx_cookie_t)ringp;
3653 	} else {
3654 		if (ringp->s_ring_first != NULL)
3655 			wakeup_worker = B_FALSE;
3656 
3657 		if (flag & MAC_TX_NO_ENQUEUE) {
3658 			/*
3659 			 * If QUEUED is not set, queue the packet
3660 			 * and let mac_tx_soft_ring_drain() set
3661 			 * the TX_BLOCKED bit for the reasons
3662 			 * explained above. Otherwise, return the
3663 			 * mblks.
3664 			 */
3665 			if (wakeup_worker) {
3666 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp,
3667 				    mp_chain, tail, cnt, sz);
3668 			} else {
3669 				ringp->s_ring_state |= S_RING_WAKEUP_CLIENT;
3670 				cookie = (mac_tx_cookie_t)ringp;
3671 				*ret_mp = mp_chain;
3672 			}
3673 		} else {
3674 			boolean_t enqueue = B_TRUE;
3675 
3676 			if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3677 				/*
3678 				 * flow-controlled. Store ringp in cookie
3679 				 * so that it can be returned as
3680 				 * mac_tx_cookie_t to client
3681 				 */
3682 				ringp->s_ring_state |= S_RING_TX_HIWAT;
3683 				cookie = (mac_tx_cookie_t)ringp;
3684 				ringp->s_ring_hiwat_cnt++;
3685 				if (ringp->s_ring_count >
3686 				    ringp->s_ring_tx_max_q_cnt) {
3687 					/* increment freed stats */
3688 					ringp->s_ring_drops += cnt;
3689 					/*
3690 					 * b_prev may be set to the fanout hint
3691 					 * hence can't use freemsg directly
3692 					 */
3693 					mac_pkt_drop(NULL, NULL,
3694 					    mp_chain, B_FALSE);
3695 					DTRACE_PROBE1(tx_queued_hiwat,
3696 					    mac_soft_ring_t *, ringp);
3697 					enqueue = B_FALSE;
3698 				}
3699 			}
3700 			if (enqueue) {
3701 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain,
3702 				    tail, cnt, sz);
3703 			}
3704 		}
3705 		if (wakeup_worker)
3706 			cv_signal(&ringp->s_ring_async);
3707 	}
3708 	return (cookie);
3709 }
3710 
3711 
3712 /*
3713  * mac_tx_soft_ring_process
3714  *
3715  * This routine is called when fanning out outgoing traffic among
3716  * multipe Tx rings.
3717  * Note that a soft ring is associated with a h/w Tx ring.
3718  */
3719 mac_tx_cookie_t
3720 mac_tx_soft_ring_process(mac_soft_ring_t *ringp, mblk_t *mp_chain,
3721     uint16_t flag, mblk_t **ret_mp)
3722 {
3723 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3724 	int	cnt;
3725 	size_t	sz;
3726 	mblk_t	*tail;
3727 	mac_tx_cookie_t cookie = NULL;
3728 
3729 	ASSERT(ringp != NULL);
3730 	ASSERT(mp_chain != NULL);
3731 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3732 	/*
3733 	 * Only two modes can come here; either it can be
3734 	 * SRS_TX_BW_FANOUT or SRS_TX_FANOUT
3735 	 */
3736 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3737 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
3738 
3739 	if (ringp->s_ring_type & ST_RING_WORKER_ONLY) {
3740 		/* Serialization mode */
3741 
3742 		mutex_enter(&ringp->s_ring_lock);
3743 		if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3744 			cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3745 			    flag, ret_mp);
3746 			mutex_exit(&ringp->s_ring_lock);
3747 			return (cookie);
3748 		}
3749 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3750 		TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3751 		if (ringp->s_ring_state & (S_RING_BLOCK | S_RING_PROC)) {
3752 			/*
3753 			 * If ring is blocked due to lack of Tx
3754 			 * descs, just return. Worker thread
3755 			 * will get scheduled when Tx desc's
3756 			 * become available.
3757 			 */
3758 			mutex_exit(&ringp->s_ring_lock);
3759 			return (cookie);
3760 		}
3761 		mac_soft_ring_worker_wakeup(ringp);
3762 		mutex_exit(&ringp->s_ring_lock);
3763 		return (cookie);
3764 	} else {
3765 		/* Default fanout mode */
3766 		/*
3767 		 * S_RING_BLOCKED is set when underlying NIC runs
3768 		 * out of Tx descs and messages start getting
3769 		 * queued. It won't get reset until
3770 		 * tx_srs_drain() completely drains out the
3771 		 * messages.
3772 		 */
3773 		boolean_t		is_subflow;
3774 		mac_tx_stats_t		stats;
3775 
3776 		if (ringp->s_ring_state & S_RING_ENQUEUED) {
3777 			/* Tx descs/resources not available */
3778 			mutex_enter(&ringp->s_ring_lock);
3779 			if (ringp->s_ring_state & S_RING_ENQUEUED) {
3780 				cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3781 				    flag, ret_mp);
3782 				mutex_exit(&ringp->s_ring_lock);
3783 				return (cookie);
3784 			}
3785 			/*
3786 			 * While we were computing mblk count, the
3787 			 * flow control condition got relieved.
3788 			 * Continue with the transmission.
3789 			 */
3790 			mutex_exit(&ringp->s_ring_lock);
3791 		}
3792 		is_subflow = ((mac_srs->srs_type & SRST_FLOW) != 0);
3793 
3794 		mp_chain = mac_tx_send(ringp->s_ring_tx_arg1,
3795 		    ringp->s_ring_tx_arg2, mp_chain,
3796 		    (is_subflow ? &stats : NULL));
3797 
3798 		/*
3799 		 * Multiple threads could be here sending packets.
3800 		 * Under such conditions, it is not possible to
3801 		 * automically set S_RING_BLOCKED bit to indicate
3802 		 * out of tx desc condition. To atomically set
3803 		 * this, we queue the returned packet and do
3804 		 * the setting of S_RING_BLOCKED in
3805 		 * mac_tx_soft_ring_drain().
3806 		 */
3807 		if (mp_chain != NULL) {
3808 			mutex_enter(&ringp->s_ring_lock);
3809 			cookie =
3810 			    mac_tx_sring_enqueue(ringp, mp_chain, flag, ret_mp);
3811 			mutex_exit(&ringp->s_ring_lock);
3812 			return (cookie);
3813 		}
3814 		if (is_subflow) {
3815 			FLOW_TX_STATS_UPDATE(mac_srs->srs_flent, &stats);
3816 		}
3817 		return (NULL);
3818 	}
3819 }
3820