xref: /titanic_44/usr/src/uts/common/io/mac/mac_sched.c (revision 70f9559bd0c02885d84a425eaafc8c280df10efb)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright 2011 Joyent, Inc.  All rights reserved.
27  */
28 
29 #include <sys/types.h>
30 #include <sys/callb.h>
31 #include <sys/sdt.h>
32 #include <sys/strsubr.h>
33 #include <sys/strsun.h>
34 #include <sys/vlan.h>
35 #include <sys/stack.h>
36 #include <sys/archsystm.h>
37 #include <inet/ipsec_impl.h>
38 #include <inet/ip_impl.h>
39 #include <inet/sadb.h>
40 #include <inet/ipsecesp.h>
41 #include <inet/ipsecah.h>
42 #include <inet/ip6.h>
43 
44 #include <sys/mac_impl.h>
45 #include <sys/mac_client_impl.h>
46 #include <sys/mac_client_priv.h>
47 #include <sys/mac_soft_ring.h>
48 #include <sys/mac_flow_impl.h>
49 
50 static mac_tx_cookie_t mac_tx_single_ring_mode(mac_soft_ring_set_t *, mblk_t *,
51     uintptr_t, uint16_t, mblk_t **);
52 static mac_tx_cookie_t mac_tx_serializer_mode(mac_soft_ring_set_t *, mblk_t *,
53     uintptr_t, uint16_t, mblk_t **);
54 static mac_tx_cookie_t mac_tx_fanout_mode(mac_soft_ring_set_t *, mblk_t *,
55     uintptr_t, uint16_t, mblk_t **);
56 static mac_tx_cookie_t mac_tx_bw_mode(mac_soft_ring_set_t *, mblk_t *,
57     uintptr_t, uint16_t, mblk_t **);
58 static mac_tx_cookie_t mac_tx_aggr_mode(mac_soft_ring_set_t *, mblk_t *,
59     uintptr_t, uint16_t, mblk_t **);
60 
61 typedef struct mac_tx_mode_s {
62 	mac_tx_srs_mode_t	mac_tx_mode;
63 	mac_tx_func_t		mac_tx_func;
64 } mac_tx_mode_t;
65 
66 /*
67  * There are seven modes of operation on the Tx side. These modes get set
68  * in mac_tx_srs_setup(). Except for the experimental TX_SERIALIZE mode,
69  * none of the other modes are user configurable. They get selected by
70  * the system depending upon whether the link (or flow) has multiple Tx
71  * rings or a bandwidth configured, or if the link is an aggr, etc.
72  *
73  * When the Tx SRS is operating in aggr mode (st_mode) or if there are
74  * multiple Tx rings owned by Tx SRS, then each Tx ring (pseudo or
75  * otherwise) will have a soft ring associated with it. These soft rings
76  * are stored in srs_tx_soft_rings[] array.
77  *
78  * Additionally in the case of aggr, there is the st_soft_rings[] array
79  * in the mac_srs_tx_t structure. This array is used to store the same
80  * set of soft rings that are present in srs_tx_soft_rings[] array but
81  * in a different manner. The soft ring associated with the pseudo Tx
82  * ring is saved at mr_index (of the pseudo ring) in st_soft_rings[]
83  * array. This helps in quickly getting the soft ring associated with the
84  * Tx ring when aggr_find_tx_ring() returns the pseudo Tx ring that is to
85  * be used for transmit.
86  */
87 mac_tx_mode_t mac_tx_mode_list[] = {
88 	{SRS_TX_DEFAULT,	mac_tx_single_ring_mode},
89 	{SRS_TX_SERIALIZE,	mac_tx_serializer_mode},
90 	{SRS_TX_FANOUT,		mac_tx_fanout_mode},
91 	{SRS_TX_BW,		mac_tx_bw_mode},
92 	{SRS_TX_BW_FANOUT,	mac_tx_bw_mode},
93 	{SRS_TX_AGGR,		mac_tx_aggr_mode},
94 	{SRS_TX_BW_AGGR,	mac_tx_bw_mode}
95 };
96 
97 /*
98  * Soft Ring Set (SRS) - The Run time code that deals with
99  * dynamic polling from the hardware, bandwidth enforcement,
100  * fanout etc.
101  *
102  * We try to use H/W classification on NIC and assign traffic for
103  * a MAC address to a particular Rx ring or ring group. There is a
104  * 1-1 mapping between a SRS and a Rx ring. The SRS dynamically
105  * switches the underlying Rx ring between interrupt and
106  * polling mode and enforces any specified B/W control.
107  *
108  * There is always a SRS created and tied to each H/W and S/W rule.
109  * Whenever we create a H/W rule, we always add the the same rule to
110  * S/W classifier and tie a SRS to it.
111  *
112  * In case a B/W control is specified, it is broken into bytes
113  * per ticks and as soon as the quota for a tick is exhausted,
114  * the underlying Rx ring is forced into poll mode for remainder of
115  * the tick. The SRS poll thread only polls for bytes that are
116  * allowed to come in the SRS. We typically let 4x the configured
117  * B/W worth of packets to come in the SRS (to prevent unnecessary
118  * drops due to bursts) but only process the specified amount.
119  *
120  * A MAC client (e.g. a VNIC or aggr) can have 1 or more
121  * Rx rings (and corresponding SRSs) assigned to it. The SRS
122  * in turn can have softrings to do protocol level fanout or
123  * softrings to do S/W based fanout or both. In case the NIC
124  * has no Rx rings, we do S/W classification to respective SRS.
125  * The S/W classification rule is always setup and ready. This
126  * allows the MAC layer to reassign Rx rings whenever needed
127  * but packets still continue to flow via the default path and
128  * getting S/W classified to correct SRS.
129  *
130  * The SRS's are used on both Tx and Rx side. They use the same
131  * data structure but the processing routines have slightly different
132  * semantics due to the fact that Rx side needs to do dynamic
133  * polling etc.
134  *
135  * Dynamic Polling Notes
136  * =====================
137  *
138  * Each Soft ring set is capable of switching its Rx ring between
139  * interrupt and poll mode and actively 'polls' for packets in
140  * poll mode. If the SRS is implementing a B/W limit, it makes
141  * sure that only Max allowed packets are pulled in poll mode
142  * and goes to poll mode as soon as B/W limit is exceeded. As
143  * such, there are no overheads to implement B/W limits.
144  *
145  * In poll mode, its better to keep the pipeline going where the
146  * SRS worker thread keeps processing packets and poll thread
147  * keeps bringing more packets (specially if they get to run
148  * on different CPUs). This also prevents the overheads associated
149  * by excessive signalling (on NUMA machines, this can be
150  * pretty devastating). The exception is latency optimized case
151  * where worker thread does no work and interrupt and poll thread
152  * are allowed to do their own drain.
153  *
154  * We use the following policy to control Dynamic Polling:
155  * 1) We switch to poll mode anytime the processing
156  *    thread causes a backlog to build up in SRS and
157  *    its associated Soft Rings (sr_poll_pkt_cnt > 0).
158  * 2) As long as the backlog stays under the low water
159  *    mark (sr_lowat), we poll the H/W for more packets.
160  * 3) If the backlog (sr_poll_pkt_cnt) exceeds low
161  *    water mark, we stay in poll mode but don't poll
162  *    the H/W for more packets.
163  * 4) Anytime in polling mode, if we poll the H/W for
164  *    packets and find nothing plus we have an existing
165  *    backlog (sr_poll_pkt_cnt > 0), we stay in polling
166  *    mode but don't poll the H/W for packets anymore
167  *    (let the polling thread go to sleep).
168  * 5) Once the backlog is relived (packets are processed)
169  *    we reenable polling (by signalling the poll thread)
170  *    only when the backlog dips below sr_poll_thres.
171  * 6) sr_hiwat is used exclusively when we are not
172  *    polling capable and is used to decide when to
173  *    drop packets so the SRS queue length doesn't grow
174  *    infinitely.
175  *
176  * NOTE: Also see the block level comment on top of mac_soft_ring.c
177  */
178 
179 /*
180  * mac_latency_optimize
181  *
182  * Controls whether the poll thread can process the packets inline
183  * or let the SRS worker thread do the processing. This applies if
184  * the SRS was not being processed. For latency sensitive traffic,
185  * this needs to be true to allow inline processing. For throughput
186  * under load, this should be false.
187  *
188  * This (and other similar) tunable should be rolled into a link
189  * or flow specific workload hint that can be set using dladm
190  * linkprop (instead of multiple such tunables).
191  */
192 boolean_t mac_latency_optimize = B_TRUE;
193 
194 /*
195  * MAC_RX_SRS_ENQUEUE_CHAIN and MAC_TX_SRS_ENQUEUE_CHAIN
196  *
197  * queue a mp or chain in soft ring set and increment the
198  * local count (srs_count) for the SRS and the shared counter
199  * (srs_poll_pkt_cnt - shared between SRS and its soft rings
200  * to track the total unprocessed packets for polling to work
201  * correctly).
202  *
203  * The size (total bytes queued) counters are incremented only
204  * if we are doing B/W control.
205  */
206 #define	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {		\
207 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
208 	if ((mac_srs)->srs_last != NULL)				\
209 		(mac_srs)->srs_last->b_next = (head);			\
210 	else								\
211 		(mac_srs)->srs_first = (head);				\
212 	(mac_srs)->srs_last = (tail);					\
213 	(mac_srs)->srs_count += count;					\
214 }
215 
216 #define	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
217 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
218 									\
219 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
220 	srs_rx->sr_poll_pkt_cnt += count;				\
221 	ASSERT(srs_rx->sr_poll_pkt_cnt > 0);				\
222 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
223 		(mac_srs)->srs_size += (sz);				\
224 		mutex_enter(&(mac_srs)->srs_bw->mac_bw_lock);		\
225 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
226 		mutex_exit(&(mac_srs)->srs_bw->mac_bw_lock);		\
227 	}								\
228 }
229 
230 #define	MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz) {	\
231 	mac_srs->srs_state |= SRS_ENQUEUED;				\
232 	MAC_SRS_ENQUEUE_CHAIN(mac_srs, head, tail, count, sz);		\
233 	if ((mac_srs)->srs_type & SRST_BW_CONTROL) {			\
234 		(mac_srs)->srs_size += (sz);				\
235 		(mac_srs)->srs_bw->mac_bw_sz += (sz);			\
236 	}								\
237 }
238 
239 /*
240  * Turn polling on routines
241  */
242 #define	MAC_SRS_POLLING_ON(mac_srs) {					\
243 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
244 	if (((mac_srs)->srs_state &					\
245 	    (SRS_POLLING_CAPAB|SRS_POLLING)) == SRS_POLLING_CAPAB) {	\
246 		(mac_srs)->srs_state |= SRS_POLLING;			\
247 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
248 		    (mac_srs)->srs_ring);				\
249 		(mac_srs)->srs_rx.sr_poll_on++;				\
250 	}								\
251 }
252 
253 #define	MAC_SRS_WORKER_POLLING_ON(mac_srs) {				\
254 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
255 	if (((mac_srs)->srs_state &					\
256 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_POLLING)) == 		\
257 	    (SRS_POLLING_CAPAB|SRS_WORKER)) {				\
258 		(mac_srs)->srs_state |= SRS_POLLING;			\
259 		(void) mac_hwring_disable_intr((mac_ring_handle_t)	\
260 		    (mac_srs)->srs_ring);				\
261 		(mac_srs)->srs_rx.sr_worker_poll_on++;			\
262 	}								\
263 }
264 
265 /*
266  * MAC_SRS_POLL_RING
267  *
268  * Signal the SRS poll thread to poll the underlying H/W ring
269  * provided it wasn't already polling (SRS_GET_PKTS was set).
270  *
271  * Poll thread gets to run only from mac_rx_srs_drain() and only
272  * if the drain was being done by the worker thread.
273  */
274 #define	MAC_SRS_POLL_RING(mac_srs) {					\
275 	mac_srs_rx_t	*srs_rx = &(mac_srs)->srs_rx;			\
276 									\
277 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
278 	srs_rx->sr_poll_thr_sig++;					\
279 	if (((mac_srs)->srs_state & 					\
280 	    (SRS_POLLING_CAPAB|SRS_WORKER|SRS_GET_PKTS)) ==		\
281 		(SRS_WORKER|SRS_POLLING_CAPAB)) {			\
282 		(mac_srs)->srs_state |= SRS_GET_PKTS;			\
283 		cv_signal(&(mac_srs)->srs_cv);   			\
284 	} else {							\
285 		srs_rx->sr_poll_thr_busy++;				\
286 	}								\
287 }
288 
289 /*
290  * MAC_SRS_CHECK_BW_CONTROL
291  *
292  * Check to see if next tick has started so we can reset the
293  * SRS_BW_ENFORCED flag and allow more packets to come in the
294  * system.
295  */
296 #define	MAC_SRS_CHECK_BW_CONTROL(mac_srs) {				\
297 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
298 	ASSERT(((mac_srs)->srs_type & SRST_TX) ||			\
299 	    MUTEX_HELD(&(mac_srs)->srs_bw->mac_bw_lock));		\
300 	clock_t now = ddi_get_lbolt();					\
301 	if ((mac_srs)->srs_bw->mac_bw_curr_time != now) {		\
302 		(mac_srs)->srs_bw->mac_bw_curr_time = now;		\
303 		(mac_srs)->srs_bw->mac_bw_used = 0;	       		\
304 		if ((mac_srs)->srs_bw->mac_bw_state & SRS_BW_ENFORCED)	\
305 			(mac_srs)->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED; \
306 	}								\
307 }
308 
309 /*
310  * MAC_SRS_WORKER_WAKEUP
311  *
312  * Wake up the SRS worker thread to process the queue as long as
313  * no one else is processing the queue. If we are optimizing for
314  * latency, we wake up the worker thread immediately or else we
315  * wait mac_srs_worker_wakeup_ticks before worker thread gets
316  * woken up.
317  */
318 int mac_srs_worker_wakeup_ticks = 0;
319 #define	MAC_SRS_WORKER_WAKEUP(mac_srs) {				\
320 	ASSERT(MUTEX_HELD(&(mac_srs)->srs_lock));			\
321 	if (!((mac_srs)->srs_state & SRS_PROC) &&			\
322 		(mac_srs)->srs_tid == NULL) {				\
323 		if (((mac_srs)->srs_state & SRS_LATENCY_OPT) ||		\
324 			(mac_srs_worker_wakeup_ticks == 0))		\
325 			cv_signal(&(mac_srs)->srs_async);		\
326 		else							\
327 			(mac_srs)->srs_tid =				\
328 				timeout(mac_srs_fire, (mac_srs),	\
329 					mac_srs_worker_wakeup_ticks);	\
330 	}								\
331 }
332 
333 #define	TX_BANDWIDTH_MODE(mac_srs)				\
334 	((mac_srs)->srs_tx.st_mode == SRS_TX_BW ||		\
335 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_FANOUT ||	\
336 	    (mac_srs)->srs_tx.st_mode == SRS_TX_BW_AGGR)
337 
338 #define	TX_SRS_TO_SOFT_RING(mac_srs, head, hint) {			\
339 	if (tx_mode == SRS_TX_BW_FANOUT)				\
340 		(void) mac_tx_fanout_mode(mac_srs, head, hint, 0, NULL);\
341 	else								\
342 		(void) mac_tx_aggr_mode(mac_srs, head, hint, 0, NULL);	\
343 }
344 
345 /*
346  * MAC_TX_SRS_BLOCK
347  *
348  * Always called from mac_tx_srs_drain() function. SRS_TX_BLOCKED
349  * will be set only if srs_tx_woken_up is FALSE. If
350  * srs_tx_woken_up is TRUE, it indicates that the wakeup arrived
351  * before we grabbed srs_lock to set SRS_TX_BLOCKED. We need to
352  * attempt to transmit again and not setting SRS_TX_BLOCKED does
353  * that.
354  */
355 #define	MAC_TX_SRS_BLOCK(srs, mp)	{			\
356 	ASSERT(MUTEX_HELD(&(srs)->srs_lock));			\
357 	if ((srs)->srs_tx.st_woken_up) {			\
358 		(srs)->srs_tx.st_woken_up = B_FALSE;		\
359 	} else {						\
360 		ASSERT(!((srs)->srs_state & SRS_TX_BLOCKED));	\
361 		(srs)->srs_state |= SRS_TX_BLOCKED;		\
362 		(srs)->srs_tx.st_stat.mts_blockcnt++;		\
363 	}							\
364 }
365 
366 /*
367  * MAC_TX_SRS_TEST_HIWAT
368  *
369  * Called before queueing a packet onto Tx SRS to test and set
370  * SRS_TX_HIWAT if srs_count exceeds srs_tx_hiwat.
371  */
372 #define	MAC_TX_SRS_TEST_HIWAT(srs, mp, tail, cnt, sz, cookie) {		\
373 	boolean_t enqueue = 1;						\
374 									\
375 	if ((srs)->srs_count > (srs)->srs_tx.st_hiwat) {		\
376 		/*							\
377 		 * flow-controlled. Store srs in cookie so that it	\
378 		 * can be returned as mac_tx_cookie_t to client		\
379 		 */							\
380 		(srs)->srs_state |= SRS_TX_HIWAT;			\
381 		cookie = (mac_tx_cookie_t)srs;				\
382 		(srs)->srs_tx.st_hiwat_cnt++;				\
383 		if ((srs)->srs_count > (srs)->srs_tx.st_max_q_cnt) {	\
384 			/* increment freed stats */			\
385 			(srs)->srs_tx.st_stat.mts_sdrops += cnt;	\
386 			/*						\
387 			 * b_prev may be set to the fanout hint		\
388 			 * hence can't use freemsg directly		\
389 			 */						\
390 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);	\
391 			DTRACE_PROBE1(tx_queued_hiwat,			\
392 			    mac_soft_ring_set_t *, srs);		\
393 			enqueue = 0;					\
394 		}							\
395 	}								\
396 	if (enqueue)							\
397 		MAC_TX_SRS_ENQUEUE_CHAIN(srs, mp, tail, cnt, sz);	\
398 }
399 
400 /* Some utility macros */
401 #define	MAC_SRS_BW_LOCK(srs)						\
402 	if (!(srs->srs_type & SRST_TX))					\
403 		mutex_enter(&srs->srs_bw->mac_bw_lock);
404 
405 #define	MAC_SRS_BW_UNLOCK(srs)						\
406 	if (!(srs->srs_type & SRST_TX))					\
407 		mutex_exit(&srs->srs_bw->mac_bw_lock);
408 
409 #define	MAC_TX_SRS_DROP_MESSAGE(srs, mp, cookie) {		\
410 	mac_pkt_drop(NULL, NULL, mp, B_FALSE);			\
411 	/* increment freed stats */				\
412 	mac_srs->srs_tx.st_stat.mts_sdrops++;			\
413 	cookie = (mac_tx_cookie_t)srs;				\
414 }
415 
416 #define	MAC_TX_SET_NO_ENQUEUE(srs, mp_chain, ret_mp, cookie) {		\
417 	mac_srs->srs_state |= SRS_TX_WAKEUP_CLIENT;			\
418 	cookie = (mac_tx_cookie_t)srs;					\
419 	*ret_mp = mp_chain;						\
420 }
421 
422 /*
423  * MAC_RX_SRS_TOODEEP
424  *
425  * Macro called as part of receive-side processing to determine if handling
426  * can occur in situ (in the interrupt thread) or if it should be left to a
427  * worker thread.  Note that the constant used to make this determination is
428  * not entirely made-up, and is a result of some emprical validation. That
429  * said, the constant is left as a static variable to allow it to be
430  * dynamically tuned in the field if and as needed.
431  */
432 static uintptr_t mac_rx_srs_stack_needed = 10240;
433 static uint_t mac_rx_srs_stack_toodeep;
434 
435 #ifndef STACK_GROWTH_DOWN
436 #error Downward stack growth assumed.
437 #endif
438 
439 #define	MAC_RX_SRS_TOODEEP() (STACK_BIAS + (uintptr_t)getfp() - \
440 	(uintptr_t)curthread->t_stkbase < mac_rx_srs_stack_needed && \
441 	++mac_rx_srs_stack_toodeep)
442 
443 
444 /*
445  * Drop the rx packet and advance to the next one in the chain.
446  */
447 static void
448 mac_rx_drop_pkt(mac_soft_ring_set_t *srs, mblk_t *mp)
449 {
450 	mac_srs_rx_t	*srs_rx = &srs->srs_rx;
451 
452 	ASSERT(mp->b_next == NULL);
453 	mutex_enter(&srs->srs_lock);
454 	MAC_UPDATE_SRS_COUNT_LOCKED(srs, 1);
455 	MAC_UPDATE_SRS_SIZE_LOCKED(srs, msgdsize(mp));
456 	mutex_exit(&srs->srs_lock);
457 
458 	srs_rx->sr_stat.mrs_sdrops++;
459 	freemsg(mp);
460 }
461 
462 /* DATAPATH RUNTIME ROUTINES */
463 
464 /*
465  * mac_srs_fire
466  *
467  * Timer callback routine for waking up the SRS worker thread.
468  */
469 static void
470 mac_srs_fire(void *arg)
471 {
472 	mac_soft_ring_set_t *mac_srs = (mac_soft_ring_set_t *)arg;
473 
474 	mutex_enter(&mac_srs->srs_lock);
475 	if (mac_srs->srs_tid == 0) {
476 		mutex_exit(&mac_srs->srs_lock);
477 		return;
478 	}
479 
480 	mac_srs->srs_tid = 0;
481 	if (!(mac_srs->srs_state & SRS_PROC))
482 		cv_signal(&mac_srs->srs_async);
483 
484 	mutex_exit(&mac_srs->srs_lock);
485 }
486 
487 /*
488  * 'hint' is fanout_hint (type of uint64_t) which is given by the TCP/IP stack,
489  * and it is used on the TX path.
490  */
491 #define	HASH_HINT(hint)	\
492 	((hint) ^ ((hint) >> 24) ^ ((hint) >> 16) ^ ((hint) >> 8))
493 
494 
495 /*
496  * hash based on the src address and the port information.
497  */
498 #define	HASH_ADDR(src, ports)					\
499 	(ntohl((src)) ^ ((ports) >> 24) ^ ((ports) >> 16) ^	\
500 	((ports) >> 8) ^ (ports))
501 
502 #define	COMPUTE_INDEX(key, sz)	(key % sz)
503 
504 #define	FANOUT_ENQUEUE_MP(head, tail, cnt, bw_ctl, sz, sz0, mp) {	\
505 	if ((tail) != NULL) {						\
506 		ASSERT((tail)->b_next == NULL);				\
507 		(tail)->b_next = (mp);					\
508 	} else {							\
509 		ASSERT((head) == NULL);					\
510 		(head) = (mp);						\
511 	}								\
512 	(tail) = (mp);							\
513 	(cnt)++;							\
514 	if ((bw_ctl))							\
515 		(sz) += (sz0);						\
516 }
517 
518 #define	MAC_FANOUT_DEFAULT	0
519 #define	MAC_FANOUT_RND_ROBIN	1
520 int mac_fanout_type = MAC_FANOUT_DEFAULT;
521 
522 #define	MAX_SR_TYPES	3
523 /* fanout types for port based hashing */
524 enum pkt_type {
525 	V4_TCP = 0,
526 	V4_UDP,
527 	OTH,
528 	UNDEF
529 };
530 
531 /*
532  * In general we do port based hashing to spread traffic over different
533  * softrings. The below tunable allows to override that behavior. Setting it
534  * to B_TRUE allows to do a fanout based on src ipv6 address. This behavior
535  * is also the applicable to ipv6 packets carrying multiple optional headers
536  * and other uncommon packet types.
537  */
538 boolean_t mac_src_ipv6_fanout = B_FALSE;
539 
540 /*
541  * Pair of local and remote ports in the transport header
542  */
543 #define	PORTS_SIZE 4
544 
545 /*
546  * mac_rx_srs_proto_fanout
547  *
548  * This routine delivers packets destined to an SRS into one of the
549  * protocol soft rings.
550  *
551  * Given a chain of packets we need to split it up into multiple sub chains
552  * destined into TCP, UDP or OTH soft ring. Instead of entering
553  * the soft ring one packet at a time, we want to enter it in the form of a
554  * chain otherwise we get this start/stop behaviour where the worker thread
555  * goes to sleep and then next packets comes in forcing it to wake up etc.
556  */
557 static void
558 mac_rx_srs_proto_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
559 {
560 	struct ether_header		*ehp;
561 	struct ether_vlan_header	*evhp;
562 	uint32_t			sap;
563 	ipha_t				*ipha;
564 	uint8_t				*dstaddr;
565 	size_t				hdrsize;
566 	mblk_t				*mp;
567 	mblk_t				*headmp[MAX_SR_TYPES];
568 	mblk_t				*tailmp[MAX_SR_TYPES];
569 	int				cnt[MAX_SR_TYPES];
570 	size_t				sz[MAX_SR_TYPES];
571 	size_t				sz1;
572 	boolean_t			bw_ctl;
573 	boolean_t			hw_classified;
574 	boolean_t			dls_bypass;
575 	boolean_t			is_ether;
576 	boolean_t			is_unicast;
577 	enum pkt_type			type;
578 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
579 
580 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
581 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
582 
583 	/*
584 	 * If we don't have a Rx ring, S/W classification would have done
585 	 * its job and its a packet meant for us. If we were polling on
586 	 * the default ring (i.e. there was a ring assigned to this SRS),
587 	 * then we need to make sure that the mac address really belongs
588 	 * to us.
589 	 */
590 	hw_classified = mac_srs->srs_ring != NULL &&
591 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
592 
593 	/*
594 	 * Special clients (eg. VLAN, non ether, etc) need DLS
595 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
596 	 * such SRSs. Another way of disabling bypass is to set the
597 	 * MCIS_RX_BYPASS_DISABLE flag.
598 	 */
599 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
600 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
601 
602 	bzero(headmp, MAX_SR_TYPES * sizeof (mblk_t *));
603 	bzero(tailmp, MAX_SR_TYPES * sizeof (mblk_t *));
604 	bzero(cnt, MAX_SR_TYPES * sizeof (int));
605 	bzero(sz, MAX_SR_TYPES * sizeof (size_t));
606 
607 	/*
608 	 * We got a chain from SRS that we need to send to the soft rings.
609 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
610 	 * performance reasons), we need to separate out v4_tcp, v4_udp
611 	 * and the rest goes in other.
612 	 */
613 	while (head != NULL) {
614 		mp = head;
615 		head = head->b_next;
616 		mp->b_next = NULL;
617 
618 		type = OTH;
619 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
620 
621 		if (is_ether) {
622 			/*
623 			 * At this point we can be sure the packet at least
624 			 * has an ether header.
625 			 */
626 			if (sz1 < sizeof (struct ether_header)) {
627 				mac_rx_drop_pkt(mac_srs, mp);
628 				continue;
629 			}
630 			ehp = (struct ether_header *)mp->b_rptr;
631 
632 			/*
633 			 * Determine if this is a VLAN or non-VLAN packet.
634 			 */
635 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
636 				evhp = (struct ether_vlan_header *)mp->b_rptr;
637 				sap = ntohs(evhp->ether_type);
638 				hdrsize = sizeof (struct ether_vlan_header);
639 				/*
640 				 * Check if the VID of the packet, if any,
641 				 * belongs to this client.
642 				 */
643 				if (!mac_client_check_flow_vid(mcip,
644 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
645 					mac_rx_drop_pkt(mac_srs, mp);
646 					continue;
647 				}
648 			} else {
649 				hdrsize = sizeof (struct ether_header);
650 			}
651 			is_unicast =
652 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
653 			dstaddr = (uint8_t *)&ehp->ether_dhost;
654 		} else {
655 			mac_header_info_t		mhi;
656 
657 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
658 			    mp, &mhi) != 0) {
659 				mac_rx_drop_pkt(mac_srs, mp);
660 				continue;
661 			}
662 			hdrsize = mhi.mhi_hdrsize;
663 			sap = mhi.mhi_bindsap;
664 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
665 			dstaddr = (uint8_t *)mhi.mhi_daddr;
666 		}
667 
668 		if (!dls_bypass) {
669 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
670 			    cnt[type], bw_ctl, sz[type], sz1, mp);
671 			continue;
672 		}
673 
674 		if (sap == ETHERTYPE_IP) {
675 			/*
676 			 * If we are H/W classified, but we have promisc
677 			 * on, then we need to check for the unicast address.
678 			 */
679 			if (hw_classified && mcip->mci_promisc_list != NULL) {
680 				mac_address_t		*map;
681 
682 				rw_enter(&mcip->mci_rw_lock, RW_READER);
683 				map = mcip->mci_unicast;
684 				if (bcmp(dstaddr, map->ma_addr,
685 				    map->ma_len) == 0)
686 					type = UNDEF;
687 				rw_exit(&mcip->mci_rw_lock);
688 			} else if (is_unicast) {
689 				type = UNDEF;
690 			}
691 		}
692 
693 		/*
694 		 * This needs to become a contract with the driver for
695 		 * the fast path.
696 		 *
697 		 * In the normal case the packet will have at least the L2
698 		 * header and the IP + Transport header in the same mblk.
699 		 * This is usually the case when the NIC driver sends up
700 		 * the packet. This is also true when the stack generates
701 		 * a packet that is looped back and when the stack uses the
702 		 * fastpath mechanism. The normal case is optimized for
703 		 * performance and may bypass DLS. All other cases go through
704 		 * the 'OTH' type path without DLS bypass.
705 		 */
706 
707 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
708 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha))
709 			type = OTH;
710 
711 		if (type == OTH) {
712 			FANOUT_ENQUEUE_MP(headmp[type], tailmp[type],
713 			    cnt[type], bw_ctl, sz[type], sz1, mp);
714 			continue;
715 		}
716 
717 		ASSERT(type == UNDEF);
718 		/*
719 		 * We look for at least 4 bytes past the IP header to get
720 		 * the port information. If we get an IP fragment, we don't
721 		 * have the port information, and we use just the protocol
722 		 * information.
723 		 */
724 		switch (ipha->ipha_protocol) {
725 		case IPPROTO_TCP:
726 			type = V4_TCP;
727 			mp->b_rptr += hdrsize;
728 			break;
729 		case IPPROTO_UDP:
730 			type = V4_UDP;
731 			mp->b_rptr += hdrsize;
732 			break;
733 		default:
734 			type = OTH;
735 			break;
736 		}
737 
738 		FANOUT_ENQUEUE_MP(headmp[type], tailmp[type], cnt[type],
739 		    bw_ctl, sz[type], sz1, mp);
740 	}
741 
742 	for (type = V4_TCP; type < UNDEF; type++) {
743 		if (headmp[type] != NULL) {
744 			mac_soft_ring_t			*softring;
745 
746 			ASSERT(tailmp[type]->b_next == NULL);
747 			switch (type) {
748 			case V4_TCP:
749 				softring = mac_srs->srs_tcp_soft_rings[0];
750 				break;
751 			case V4_UDP:
752 				softring = mac_srs->srs_udp_soft_rings[0];
753 				break;
754 			case OTH:
755 				softring = mac_srs->srs_oth_soft_rings[0];
756 			}
757 			mac_rx_soft_ring_process(mcip, softring,
758 			    headmp[type], tailmp[type], cnt[type], sz[type]);
759 		}
760 	}
761 }
762 
763 int	fanout_unalligned = 0;
764 
765 /*
766  * mac_rx_srs_long_fanout
767  *
768  * The fanout routine for IPv6
769  */
770 static int
771 mac_rx_srs_long_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *mp,
772     uint32_t sap, size_t hdrsize, enum pkt_type *type, uint_t *indx)
773 {
774 	ip6_t		*ip6h;
775 	uint8_t		*whereptr;
776 	uint_t		hash;
777 	uint16_t	remlen;
778 	uint8_t		nexthdr;
779 	uint16_t	hdr_len;
780 
781 	if (sap == ETHERTYPE_IPV6) {
782 		boolean_t	modifiable = B_TRUE;
783 
784 		ASSERT(MBLKL(mp) >= hdrsize);
785 
786 		ip6h = (ip6_t *)(mp->b_rptr + hdrsize);
787 		if ((unsigned char *)ip6h == mp->b_wptr) {
788 			/*
789 			 * The first mblk_t only includes the mac header.
790 			 * Note that it is safe to change the mp pointer here,
791 			 * as the subsequent operation does not assume mp
792 			 * points to the start of the mac header.
793 			 */
794 			mp = mp->b_cont;
795 
796 			/*
797 			 * Make sure ip6h holds the full ip6_t structure.
798 			 */
799 			if (mp == NULL)
800 				return (-1);
801 
802 			if (MBLKL(mp) < IPV6_HDR_LEN) {
803 				modifiable = (DB_REF(mp) == 1);
804 
805 				if (modifiable &&
806 				    !pullupmsg(mp, IPV6_HDR_LEN)) {
807 					return (-1);
808 				}
809 			}
810 
811 			ip6h = (ip6_t *)mp->b_rptr;
812 		}
813 
814 		if (!modifiable || !(OK_32PTR((char *)ip6h)) ||
815 		    ((unsigned char *)ip6h + IPV6_HDR_LEN > mp->b_wptr)) {
816 			/*
817 			 * If either ip6h is not alligned, or ip6h does not
818 			 * hold the complete ip6_t structure (a pullupmsg()
819 			 * is not an option since it would result in an
820 			 * unalligned ip6h), fanout to the default ring. Note
821 			 * that this may cause packets reordering.
822 			 */
823 			*indx = 0;
824 			*type = OTH;
825 			fanout_unalligned++;
826 			return (0);
827 		}
828 
829 		remlen = ntohs(ip6h->ip6_plen);
830 		nexthdr = ip6h->ip6_nxt;
831 
832 		if (remlen < MIN_EHDR_LEN)
833 			return (-1);
834 		/*
835 		 * Do src based fanout if below tunable is set to B_TRUE or
836 		 * when mac_ip_hdr_length_v6() fails because of malformed
837 		 * packets or because mblk's need to be concatenated using
838 		 * pullupmsg().
839 		 */
840 		if (mac_src_ipv6_fanout || !mac_ip_hdr_length_v6(ip6h,
841 		    mp->b_wptr, &hdr_len, &nexthdr, NULL)) {
842 			goto src_based_fanout;
843 		}
844 		whereptr = (uint8_t *)ip6h + hdr_len;
845 
846 		/* If the transport is one of below, we do port based fanout */
847 		switch (nexthdr) {
848 		case IPPROTO_TCP:
849 		case IPPROTO_UDP:
850 		case IPPROTO_SCTP:
851 		case IPPROTO_ESP:
852 			/*
853 			 * If the ports in the transport header is not part of
854 			 * the mblk, do src_based_fanout, instead of calling
855 			 * pullupmsg().
856 			 */
857 			if (mp->b_cont != NULL &&
858 			    whereptr + PORTS_SIZE > mp->b_wptr) {
859 				goto src_based_fanout;
860 			}
861 			break;
862 		default:
863 			break;
864 		}
865 
866 		switch (nexthdr) {
867 		case IPPROTO_TCP:
868 			hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
869 			    *(uint32_t *)whereptr);
870 			*indx = COMPUTE_INDEX(hash,
871 			    mac_srs->srs_tcp_ring_count);
872 			*type = OTH;
873 			break;
874 
875 		case IPPROTO_UDP:
876 		case IPPROTO_SCTP:
877 		case IPPROTO_ESP:
878 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
879 				hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src),
880 				    *(uint32_t *)whereptr);
881 				*indx = COMPUTE_INDEX(hash,
882 				    mac_srs->srs_udp_ring_count);
883 			} else {
884 				*indx = mac_srs->srs_ind %
885 				    mac_srs->srs_udp_ring_count;
886 				mac_srs->srs_ind++;
887 			}
888 			*type = OTH;
889 			break;
890 
891 			/* For all other protocol, do source based fanout */
892 		default:
893 			goto src_based_fanout;
894 		}
895 	} else {
896 		*indx = 0;
897 		*type = OTH;
898 	}
899 	return (0);
900 
901 src_based_fanout:
902 	hash = HASH_ADDR(V4_PART_OF_V6(ip6h->ip6_src), (uint32_t)0);
903 	*indx = COMPUTE_INDEX(hash, mac_srs->srs_oth_ring_count);
904 	*type = OTH;
905 	return (0);
906 }
907 
908 /*
909  * mac_rx_srs_fanout
910  *
911  * This routine delivers packets destined to an SRS into a soft ring member
912  * of the set.
913  *
914  * Given a chain of packets we need to split it up into multiple sub chains
915  * destined for one of the TCP, UDP or OTH soft rings. Instead of entering
916  * the soft ring one packet at a time, we want to enter it in the form of a
917  * chain otherwise we get this start/stop behaviour where the worker thread
918  * goes to sleep and then next packets comes in forcing it to wake up etc.
919  *
920  * Note:
921  * Since we know what is the maximum fanout possible, we create a 2D array
922  * of 'softring types * MAX_SR_FANOUT' for the head, tail, cnt and sz
923  * variables so that we can enter the softrings with chain. We need the
924  * MAX_SR_FANOUT so we can allocate the arrays on the stack (a kmem_alloc
925  * for each packet would be expensive). If we ever want to have the
926  * ability to have unlimited fanout, we should probably declare a head,
927  * tail, cnt, sz with each soft ring (a data struct which contains a softring
928  * along with these members) and create an array of this uber struct so we
929  * don't have to do kmem_alloc.
930  */
931 int	fanout_oth1 = 0;
932 int	fanout_oth2 = 0;
933 int	fanout_oth3 = 0;
934 int	fanout_oth4 = 0;
935 int	fanout_oth5 = 0;
936 
937 static void
938 mac_rx_srs_fanout(mac_soft_ring_set_t *mac_srs, mblk_t *head)
939 {
940 	struct ether_header		*ehp;
941 	struct ether_vlan_header	*evhp;
942 	uint32_t			sap;
943 	ipha_t				*ipha;
944 	uint8_t				*dstaddr;
945 	uint_t				indx;
946 	size_t				ports_offset;
947 	size_t				ipha_len;
948 	size_t				hdrsize;
949 	uint_t				hash;
950 	mblk_t				*mp;
951 	mblk_t				*headmp[MAX_SR_TYPES][MAX_SR_FANOUT];
952 	mblk_t				*tailmp[MAX_SR_TYPES][MAX_SR_FANOUT];
953 	int				cnt[MAX_SR_TYPES][MAX_SR_FANOUT];
954 	size_t				sz[MAX_SR_TYPES][MAX_SR_FANOUT];
955 	size_t				sz1;
956 	boolean_t			bw_ctl;
957 	boolean_t			hw_classified;
958 	boolean_t			dls_bypass;
959 	boolean_t			is_ether;
960 	boolean_t			is_unicast;
961 	int				fanout_cnt;
962 	enum pkt_type			type;
963 	mac_client_impl_t		*mcip = mac_srs->srs_mcip;
964 
965 	is_ether = (mcip->mci_mip->mi_info.mi_nativemedia == DL_ETHER);
966 	bw_ctl = ((mac_srs->srs_type & SRST_BW_CONTROL) != 0);
967 
968 	/*
969 	 * If we don't have a Rx ring, S/W classification would have done
970 	 * its job and its a packet meant for us. If we were polling on
971 	 * the default ring (i.e. there was a ring assigned to this SRS),
972 	 * then we need to make sure that the mac address really belongs
973 	 * to us.
974 	 */
975 	hw_classified = mac_srs->srs_ring != NULL &&
976 	    mac_srs->srs_ring->mr_classify_type == MAC_HW_CLASSIFIER;
977 
978 	/*
979 	 * Special clients (eg. VLAN, non ether, etc) need DLS
980 	 * processing in the Rx path. SRST_DLS_BYPASS will be clear for
981 	 * such SRSs. Another way of disabling bypass is to set the
982 	 * MCIS_RX_BYPASS_DISABLE flag.
983 	 */
984 	dls_bypass = ((mac_srs->srs_type & SRST_DLS_BYPASS) != 0) &&
985 	    ((mcip->mci_state_flags & MCIS_RX_BYPASS_DISABLE) == 0);
986 
987 	/*
988 	 * Since the softrings are never destroyed and we always
989 	 * create equal number of softrings for TCP, UDP and rest,
990 	 * its OK to check one of them for count and use it without
991 	 * any lock. In future, if soft rings get destroyed because
992 	 * of reduction in fanout, we will need to ensure that happens
993 	 * behind the SRS_PROC.
994 	 */
995 	fanout_cnt = mac_srs->srs_tcp_ring_count;
996 
997 	bzero(headmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
998 	bzero(tailmp, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (mblk_t *));
999 	bzero(cnt, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (int));
1000 	bzero(sz, MAX_SR_TYPES * MAX_SR_FANOUT * sizeof (size_t));
1001 
1002 	/*
1003 	 * We got a chain from SRS that we need to send to the soft rings.
1004 	 * Since squeues for TCP & IPv4 sap poll their soft rings (for
1005 	 * performance reasons), we need to separate out v4_tcp, v4_udp
1006 	 * and the rest goes in other.
1007 	 */
1008 	while (head != NULL) {
1009 		mp = head;
1010 		head = head->b_next;
1011 		mp->b_next = NULL;
1012 
1013 		type = OTH;
1014 		sz1 = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1015 
1016 		if (is_ether) {
1017 			/*
1018 			 * At this point we can be sure the packet at least
1019 			 * has an ether header.
1020 			 */
1021 			if (sz1 < sizeof (struct ether_header)) {
1022 				mac_rx_drop_pkt(mac_srs, mp);
1023 				continue;
1024 			}
1025 			ehp = (struct ether_header *)mp->b_rptr;
1026 
1027 			/*
1028 			 * Determine if this is a VLAN or non-VLAN packet.
1029 			 */
1030 			if ((sap = ntohs(ehp->ether_type)) == VLAN_TPID) {
1031 				evhp = (struct ether_vlan_header *)mp->b_rptr;
1032 				sap = ntohs(evhp->ether_type);
1033 				hdrsize = sizeof (struct ether_vlan_header);
1034 				/*
1035 				 * Check if the VID of the packet, if any,
1036 				 * belongs to this client.
1037 				 */
1038 				if (!mac_client_check_flow_vid(mcip,
1039 				    VLAN_ID(ntohs(evhp->ether_tci)))) {
1040 					mac_rx_drop_pkt(mac_srs, mp);
1041 					continue;
1042 				}
1043 			} else {
1044 				hdrsize = sizeof (struct ether_header);
1045 			}
1046 			is_unicast =
1047 			    ((((uint8_t *)&ehp->ether_dhost)[0] & 0x01) == 0);
1048 			dstaddr = (uint8_t *)&ehp->ether_dhost;
1049 		} else {
1050 			mac_header_info_t		mhi;
1051 
1052 			if (mac_header_info((mac_handle_t)mcip->mci_mip,
1053 			    mp, &mhi) != 0) {
1054 				mac_rx_drop_pkt(mac_srs, mp);
1055 				continue;
1056 			}
1057 			hdrsize = mhi.mhi_hdrsize;
1058 			sap = mhi.mhi_bindsap;
1059 			is_unicast = (mhi.mhi_dsttype == MAC_ADDRTYPE_UNICAST);
1060 			dstaddr = (uint8_t *)mhi.mhi_daddr;
1061 		}
1062 
1063 		if (!dls_bypass) {
1064 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1065 			    hdrsize, &type, &indx) == -1) {
1066 				mac_rx_drop_pkt(mac_srs, mp);
1067 				continue;
1068 			}
1069 
1070 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1071 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1072 			    sz[type][indx], sz1, mp);
1073 			continue;
1074 		}
1075 
1076 
1077 		/*
1078 		 * If we are using the default Rx ring where H/W or S/W
1079 		 * classification has not happened, we need to verify if
1080 		 * this unicast packet really belongs to us.
1081 		 */
1082 		if (sap == ETHERTYPE_IP) {
1083 			/*
1084 			 * If we are H/W classified, but we have promisc
1085 			 * on, then we need to check for the unicast address.
1086 			 */
1087 			if (hw_classified && mcip->mci_promisc_list != NULL) {
1088 				mac_address_t		*map;
1089 
1090 				rw_enter(&mcip->mci_rw_lock, RW_READER);
1091 				map = mcip->mci_unicast;
1092 				if (bcmp(dstaddr, map->ma_addr,
1093 				    map->ma_len) == 0)
1094 					type = UNDEF;
1095 				rw_exit(&mcip->mci_rw_lock);
1096 			} else if (is_unicast) {
1097 				type = UNDEF;
1098 			}
1099 		}
1100 
1101 		/*
1102 		 * This needs to become a contract with the driver for
1103 		 * the fast path.
1104 		 */
1105 
1106 		ipha = (ipha_t *)(mp->b_rptr + hdrsize);
1107 		if ((type != OTH) && MBLK_RX_FANOUT_SLOWPATH(mp, ipha)) {
1108 			type = OTH;
1109 			fanout_oth1++;
1110 		}
1111 
1112 		if (type != OTH) {
1113 			uint16_t	frag_offset_flags;
1114 
1115 			switch (ipha->ipha_protocol) {
1116 			case IPPROTO_TCP:
1117 			case IPPROTO_UDP:
1118 			case IPPROTO_SCTP:
1119 			case IPPROTO_ESP:
1120 				ipha_len = IPH_HDR_LENGTH(ipha);
1121 				if ((uchar_t *)ipha + ipha_len + PORTS_SIZE >
1122 				    mp->b_wptr) {
1123 					type = OTH;
1124 					break;
1125 				}
1126 				frag_offset_flags =
1127 				    ntohs(ipha->ipha_fragment_offset_and_flags);
1128 				if ((frag_offset_flags &
1129 				    (IPH_MF | IPH_OFFSET)) != 0) {
1130 					type = OTH;
1131 					fanout_oth3++;
1132 					break;
1133 				}
1134 				ports_offset = hdrsize + ipha_len;
1135 				break;
1136 			default:
1137 				type = OTH;
1138 				fanout_oth4++;
1139 				break;
1140 			}
1141 		}
1142 
1143 		if (type == OTH) {
1144 			if (mac_rx_srs_long_fanout(mac_srs, mp, sap,
1145 			    hdrsize, &type, &indx) == -1) {
1146 				mac_rx_drop_pkt(mac_srs, mp);
1147 				continue;
1148 			}
1149 
1150 			FANOUT_ENQUEUE_MP(headmp[type][indx],
1151 			    tailmp[type][indx], cnt[type][indx], bw_ctl,
1152 			    sz[type][indx], sz1, mp);
1153 			continue;
1154 		}
1155 
1156 		ASSERT(type == UNDEF);
1157 
1158 		/*
1159 		 * XXX-Sunay: We should hold srs_lock since ring_count
1160 		 * below can change. But if we are always called from
1161 		 * mac_rx_srs_drain and SRS_PROC is set, then we can
1162 		 * enforce that ring_count can't be changed i.e.
1163 		 * to change fanout type or ring count, the calling
1164 		 * thread needs to be behind SRS_PROC.
1165 		 */
1166 		switch (ipha->ipha_protocol) {
1167 		case IPPROTO_TCP:
1168 			/*
1169 			 * Note that for ESP, we fanout on SPI and it is at the
1170 			 * same offset as the 2x16-bit ports. So it is clumped
1171 			 * along with TCP, UDP and SCTP.
1172 			 */
1173 			hash = HASH_ADDR(ipha->ipha_src,
1174 			    *(uint32_t *)(mp->b_rptr + ports_offset));
1175 			indx = COMPUTE_INDEX(hash, mac_srs->srs_tcp_ring_count);
1176 			type = V4_TCP;
1177 			mp->b_rptr += hdrsize;
1178 			break;
1179 		case IPPROTO_UDP:
1180 		case IPPROTO_SCTP:
1181 		case IPPROTO_ESP:
1182 			if (mac_fanout_type == MAC_FANOUT_DEFAULT) {
1183 				hash = HASH_ADDR(ipha->ipha_src,
1184 				    *(uint32_t *)(mp->b_rptr + ports_offset));
1185 				indx = COMPUTE_INDEX(hash,
1186 				    mac_srs->srs_udp_ring_count);
1187 			} else {
1188 				indx = mac_srs->srs_ind %
1189 				    mac_srs->srs_udp_ring_count;
1190 				mac_srs->srs_ind++;
1191 			}
1192 			type = V4_UDP;
1193 			mp->b_rptr += hdrsize;
1194 			break;
1195 		default:
1196 			indx = 0;
1197 			type = OTH;
1198 		}
1199 
1200 		FANOUT_ENQUEUE_MP(headmp[type][indx], tailmp[type][indx],
1201 		    cnt[type][indx], bw_ctl, sz[type][indx], sz1, mp);
1202 	}
1203 
1204 	for (type = V4_TCP; type < UNDEF; type++) {
1205 		int	i;
1206 
1207 		for (i = 0; i < fanout_cnt; i++) {
1208 			if (headmp[type][i] != NULL) {
1209 				mac_soft_ring_t	*softring;
1210 
1211 				ASSERT(tailmp[type][i]->b_next == NULL);
1212 				switch (type) {
1213 				case V4_TCP:
1214 					softring =
1215 					    mac_srs->srs_tcp_soft_rings[i];
1216 					break;
1217 				case V4_UDP:
1218 					softring =
1219 					    mac_srs->srs_udp_soft_rings[i];
1220 					break;
1221 				case OTH:
1222 					softring =
1223 					    mac_srs->srs_oth_soft_rings[i];
1224 					break;
1225 				}
1226 				mac_rx_soft_ring_process(mcip,
1227 				    softring, headmp[type][i], tailmp[type][i],
1228 				    cnt[type][i], sz[type][i]);
1229 			}
1230 		}
1231 	}
1232 }
1233 
1234 #define	SRS_BYTES_TO_PICKUP	150000
1235 ssize_t	max_bytes_to_pickup = SRS_BYTES_TO_PICKUP;
1236 
1237 /*
1238  * mac_rx_srs_poll_ring
1239  *
1240  * This SRS Poll thread uses this routine to poll the underlying hardware
1241  * Rx ring to get a chain of packets. It can inline process that chain
1242  * if mac_latency_optimize is set (default) or signal the SRS worker thread
1243  * to do the remaining processing.
1244  *
1245  * Since packets come in the system via interrupt or poll path, we also
1246  * update the stats and deal with promiscous clients here.
1247  */
1248 void
1249 mac_rx_srs_poll_ring(mac_soft_ring_set_t *mac_srs)
1250 {
1251 	kmutex_t 		*lock = &mac_srs->srs_lock;
1252 	kcondvar_t 		*async = &mac_srs->srs_cv;
1253 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1254 	mblk_t 			*head, *tail, *mp;
1255 	callb_cpr_t 		cprinfo;
1256 	ssize_t 		bytes_to_pickup;
1257 	size_t 			sz;
1258 	int			count;
1259 	mac_client_impl_t	*smcip;
1260 
1261 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "mac_srs_poll");
1262 	mutex_enter(lock);
1263 
1264 start:
1265 	for (;;) {
1266 		if (mac_srs->srs_state & SRS_PAUSE)
1267 			goto done;
1268 
1269 		CALLB_CPR_SAFE_BEGIN(&cprinfo);
1270 		cv_wait(async, lock);
1271 		CALLB_CPR_SAFE_END(&cprinfo, lock);
1272 
1273 		if (mac_srs->srs_state & SRS_PAUSE)
1274 			goto done;
1275 
1276 check_again:
1277 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
1278 			/*
1279 			 * We pick as many bytes as we are allowed to queue.
1280 			 * Its possible that we will exceed the total
1281 			 * packets queued in case this SRS is part of the
1282 			 * Rx ring group since > 1 poll thread can be pulling
1283 			 * upto the max allowed packets at the same time
1284 			 * but that should be OK.
1285 			 */
1286 			mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1287 			bytes_to_pickup =
1288 			    mac_srs->srs_bw->mac_bw_drop_threshold -
1289 			    mac_srs->srs_bw->mac_bw_sz;
1290 			/*
1291 			 * We shouldn't have been signalled if we
1292 			 * have 0 or less bytes to pick but since
1293 			 * some of the bytes accounting is driver
1294 			 * dependant, we do the safety check.
1295 			 */
1296 			if (bytes_to_pickup < 0)
1297 				bytes_to_pickup = 0;
1298 			mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1299 		} else {
1300 			/*
1301 			 * ToDO: Need to change the polling API
1302 			 * to add a packet count and a flag which
1303 			 * tells the driver whether we want packets
1304 			 * based on a count, or bytes, or all the
1305 			 * packets queued in the driver/HW. This
1306 			 * way, we never have to check the limits
1307 			 * on poll path. We truly let only as many
1308 			 * packets enter the system as we are willing
1309 			 * to process or queue.
1310 			 *
1311 			 * Something along the lines of
1312 			 * pkts_to_pickup = mac_soft_ring_max_q_cnt -
1313 			 *	mac_srs->srs_poll_pkt_cnt
1314 			 */
1315 
1316 			/*
1317 			 * Since we are not doing B/W control, pick
1318 			 * as many packets as allowed.
1319 			 */
1320 			bytes_to_pickup = max_bytes_to_pickup;
1321 		}
1322 
1323 		/* Poll the underlying Hardware */
1324 		mutex_exit(lock);
1325 		head = MAC_HWRING_POLL(mac_srs->srs_ring, (int)bytes_to_pickup);
1326 		mutex_enter(lock);
1327 
1328 		ASSERT((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
1329 		    SRS_POLL_THR_OWNER);
1330 
1331 		mp = tail = head;
1332 		count = 0;
1333 		sz = 0;
1334 		while (mp != NULL) {
1335 			tail = mp;
1336 			sz += msgdsize(mp);
1337 			mp = mp->b_next;
1338 			count++;
1339 		}
1340 
1341 		if (head != NULL) {
1342 			tail->b_next = NULL;
1343 			smcip = mac_srs->srs_mcip;
1344 
1345 			SRS_RX_STAT_UPDATE(mac_srs, pollbytes, sz);
1346 			SRS_RX_STAT_UPDATE(mac_srs, pollcnt, count);
1347 
1348 			/*
1349 			 * If there are any promiscuous mode callbacks
1350 			 * defined for this MAC client, pass them a copy
1351 			 * if appropriate and also update the counters.
1352 			 */
1353 			if (smcip != NULL) {
1354 				if (smcip->mci_mip->mi_promisc_list != NULL) {
1355 					mutex_exit(lock);
1356 					mac_promisc_dispatch(smcip->mci_mip,
1357 					    head, NULL);
1358 					mutex_enter(lock);
1359 				}
1360 			}
1361 			if (mac_srs->srs_type & SRST_BW_CONTROL) {
1362 				mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1363 				mac_srs->srs_bw->mac_bw_polled += sz;
1364 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1365 			}
1366 			MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, head, tail,
1367 			    count, sz);
1368 			if (count <= 10)
1369 				srs_rx->sr_stat.mrs_chaincntundr10++;
1370 			else if (count > 10 && count <= 50)
1371 				srs_rx->sr_stat.mrs_chaincnt10to50++;
1372 			else
1373 				srs_rx->sr_stat.mrs_chaincntover50++;
1374 		}
1375 
1376 		/*
1377 		 * We are guaranteed that SRS_PROC will be set if we
1378 		 * are here. Also, poll thread gets to run only if
1379 		 * the drain was being done by a worker thread although
1380 		 * its possible that worker thread is still running
1381 		 * and poll thread was sent down to keep the pipeline
1382 		 * going instead of doing a complete drain and then
1383 		 * trying to poll the NIC.
1384 		 *
1385 		 * So we need to check SRS_WORKER flag to make sure
1386 		 * that the worker thread is not processing the queue
1387 		 * in parallel to us. The flags and conditions are
1388 		 * protected by the srs_lock to prevent any race. We
1389 		 * ensure that we don't drop the srs_lock from now
1390 		 * till the end and similarly we don't drop the srs_lock
1391 		 * in mac_rx_srs_drain() till similar condition check
1392 		 * are complete. The mac_rx_srs_drain() needs to ensure
1393 		 * that SRS_WORKER flag remains set as long as its
1394 		 * processing the queue.
1395 		 */
1396 		if (!(mac_srs->srs_state & SRS_WORKER) &&
1397 		    (mac_srs->srs_first != NULL)) {
1398 			/*
1399 			 * We have packets to process and worker thread
1400 			 * is not running. Check to see if poll thread is
1401 			 * allowed to process.
1402 			 */
1403 			if (mac_srs->srs_state & SRS_LATENCY_OPT) {
1404 				mac_srs->srs_drain_func(mac_srs, SRS_POLL_PROC);
1405 				if (!(mac_srs->srs_state & SRS_PAUSE) &&
1406 				    srs_rx->sr_poll_pkt_cnt <=
1407 				    srs_rx->sr_lowat) {
1408 					srs_rx->sr_poll_again++;
1409 					goto check_again;
1410 				}
1411 				/*
1412 				 * We are already above low water mark
1413 				 * so stay in the polling mode but no
1414 				 * need to poll. Once we dip below
1415 				 * the polling threshold, the processing
1416 				 * thread (soft ring) will signal us
1417 				 * to poll again (MAC_UPDATE_SRS_COUNT)
1418 				 */
1419 				srs_rx->sr_poll_drain_no_poll++;
1420 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1421 				/*
1422 				 * In B/W control case, its possible
1423 				 * that the backlog built up due to
1424 				 * B/W limit being reached and packets
1425 				 * are queued only in SRS. In this case,
1426 				 * we should schedule worker thread
1427 				 * since no one else will wake us up.
1428 				 */
1429 				if ((mac_srs->srs_type & SRST_BW_CONTROL) &&
1430 				    (mac_srs->srs_tid == NULL)) {
1431 					mac_srs->srs_tid =
1432 					    timeout(mac_srs_fire, mac_srs, 1);
1433 					srs_rx->sr_poll_worker_wakeup++;
1434 				}
1435 			} else {
1436 				/*
1437 				 * Wakeup the worker thread for more processing.
1438 				 * We optimize for throughput in this case.
1439 				 */
1440 				mac_srs->srs_state &= ~(SRS_PROC|SRS_GET_PKTS);
1441 				MAC_SRS_WORKER_WAKEUP(mac_srs);
1442 				srs_rx->sr_poll_sig_worker++;
1443 			}
1444 		} else if ((mac_srs->srs_first == NULL) &&
1445 		    !(mac_srs->srs_state & SRS_WORKER)) {
1446 			/*
1447 			 * There is nothing queued in SRS and
1448 			 * no worker thread running. Plus we
1449 			 * didn't get anything from the H/W
1450 			 * as well (head == NULL);
1451 			 */
1452 			ASSERT(head == NULL);
1453 			mac_srs->srs_state &=
1454 			    ~(SRS_PROC|SRS_GET_PKTS);
1455 
1456 			/*
1457 			 * If we have a packets in soft ring, don't allow
1458 			 * more packets to come into this SRS by keeping the
1459 			 * interrupts off but not polling the H/W. The
1460 			 * poll thread will get signaled as soon as
1461 			 * srs_poll_pkt_cnt dips below poll threshold.
1462 			 */
1463 			if (srs_rx->sr_poll_pkt_cnt == 0) {
1464 				srs_rx->sr_poll_intr_enable++;
1465 				MAC_SRS_POLLING_OFF(mac_srs);
1466 			} else {
1467 				/*
1468 				 * We know nothing is queued in SRS
1469 				 * since we are here after checking
1470 				 * srs_first is NULL. The backlog
1471 				 * is entirely due to packets queued
1472 				 * in Soft ring which will wake us up
1473 				 * and get the interface out of polling
1474 				 * mode once the backlog dips below
1475 				 * sr_poll_thres.
1476 				 */
1477 				srs_rx->sr_poll_no_poll++;
1478 			}
1479 		} else {
1480 			/*
1481 			 * Worker thread is already running.
1482 			 * Nothing much to do. If the polling
1483 			 * was enabled, worker thread will deal
1484 			 * with that.
1485 			 */
1486 			mac_srs->srs_state &= ~SRS_GET_PKTS;
1487 			srs_rx->sr_poll_goto_sleep++;
1488 		}
1489 	}
1490 done:
1491 	mac_srs->srs_state |= SRS_POLL_THR_QUIESCED;
1492 	cv_signal(&mac_srs->srs_async);
1493 	/*
1494 	 * If this is a temporary quiesce then wait for the restart signal
1495 	 * from the srs worker. Then clear the flags and signal the srs worker
1496 	 * to ensure a positive handshake and go back to start.
1497 	 */
1498 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_POLL_THR_RESTART)))
1499 		cv_wait(async, lock);
1500 	if (mac_srs->srs_state & SRS_POLL_THR_RESTART) {
1501 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
1502 		mac_srs->srs_state &=
1503 		    ~(SRS_POLL_THR_QUIESCED | SRS_POLL_THR_RESTART);
1504 		cv_signal(&mac_srs->srs_async);
1505 		goto start;
1506 	} else {
1507 		mac_srs->srs_state |= SRS_POLL_THR_EXITED;
1508 		cv_signal(&mac_srs->srs_async);
1509 		CALLB_CPR_EXIT(&cprinfo);
1510 		thread_exit();
1511 	}
1512 }
1513 
1514 /*
1515  * mac_srs_pick_chain
1516  *
1517  * In Bandwidth control case, checks how many packets can be processed
1518  * and return them in a sub chain.
1519  */
1520 static mblk_t *
1521 mac_srs_pick_chain(mac_soft_ring_set_t *mac_srs, mblk_t **chain_tail,
1522     size_t *chain_sz, int *chain_cnt)
1523 {
1524 	mblk_t 			*head = NULL;
1525 	mblk_t 			*tail = NULL;
1526 	size_t			sz;
1527 	size_t 			tsz = 0;
1528 	int			cnt = 0;
1529 	mblk_t 			*mp;
1530 
1531 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1532 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1533 	if (((mac_srs->srs_bw->mac_bw_used + mac_srs->srs_size) <=
1534 	    mac_srs->srs_bw->mac_bw_limit) ||
1535 	    (mac_srs->srs_bw->mac_bw_limit == 0)) {
1536 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1537 		head = mac_srs->srs_first;
1538 		mac_srs->srs_first = NULL;
1539 		*chain_tail = mac_srs->srs_last;
1540 		mac_srs->srs_last = NULL;
1541 		*chain_sz = mac_srs->srs_size;
1542 		*chain_cnt = mac_srs->srs_count;
1543 		mac_srs->srs_count = 0;
1544 		mac_srs->srs_size = 0;
1545 		return (head);
1546 	}
1547 
1548 	/*
1549 	 * Can't clear the entire backlog.
1550 	 * Need to find how many packets to pick
1551 	 */
1552 	ASSERT(MUTEX_HELD(&mac_srs->srs_bw->mac_bw_lock));
1553 	while ((mp = mac_srs->srs_first) != NULL) {
1554 		sz = msgdsize(mp);
1555 		if ((tsz + sz + mac_srs->srs_bw->mac_bw_used) >
1556 		    mac_srs->srs_bw->mac_bw_limit) {
1557 			if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED))
1558 				mac_srs->srs_bw->mac_bw_state |=
1559 				    SRS_BW_ENFORCED;
1560 			break;
1561 		}
1562 
1563 		/*
1564 		 * The _size & cnt is  decremented from the softrings
1565 		 * when they send up the packet for polling to work
1566 		 * properly.
1567 		 */
1568 		tsz += sz;
1569 		cnt++;
1570 		mac_srs->srs_count--;
1571 		mac_srs->srs_size -= sz;
1572 		if (tail != NULL)
1573 			tail->b_next = mp;
1574 		else
1575 			head = mp;
1576 		tail = mp;
1577 		mac_srs->srs_first = mac_srs->srs_first->b_next;
1578 	}
1579 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1580 	if (mac_srs->srs_first == NULL)
1581 		mac_srs->srs_last = NULL;
1582 
1583 	if (tail != NULL)
1584 		tail->b_next = NULL;
1585 	*chain_tail = tail;
1586 	*chain_cnt = cnt;
1587 	*chain_sz = tsz;
1588 
1589 	return (head);
1590 }
1591 
1592 /*
1593  * mac_rx_srs_drain
1594  *
1595  * The SRS drain routine. Gets to run to clear the queue. Any thread
1596  * (worker, interrupt, poll) can call this based on processing model.
1597  * The first thing we do is disable interrupts if possible and then
1598  * drain the queue. we also try to poll the underlying hardware if
1599  * there is a dedicated hardware Rx ring assigned to this SRS.
1600  *
1601  * There is a equivalent drain routine in bandwidth control mode
1602  * mac_rx_srs_drain_bw. There is some code duplication between the two
1603  * routines but they are highly performance sensitive and are easier
1604  * to read/debug if they stay separate. Any code changes here might
1605  * also apply to mac_rx_srs_drain_bw as well.
1606  */
1607 void
1608 mac_rx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1609 {
1610 	mblk_t 			*head;
1611 	mblk_t			*tail;
1612 	timeout_id_t 		tid;
1613 	int			cnt = 0;
1614 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1615 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1616 
1617 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1618 	ASSERT(!(mac_srs->srs_type & SRST_BW_CONTROL));
1619 
1620 	/* If we are blanked i.e. can't do upcalls, then we are done */
1621 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1622 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1623 		    (mac_srs->srs_state & SRS_PAUSE));
1624 		goto out;
1625 	}
1626 
1627 	if (mac_srs->srs_first == NULL)
1628 		goto out;
1629 
1630 	if (!(mac_srs->srs_state & SRS_LATENCY_OPT) &&
1631 	    (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)) {
1632 		/*
1633 		 * In the normal case, the SRS worker thread does no
1634 		 * work and we wait for a backlog to build up before
1635 		 * we switch into polling mode. In case we are
1636 		 * optimizing for throughput, we use the worker thread
1637 		 * as well. The goal is to let worker thread process
1638 		 * the queue and poll thread to feed packets into
1639 		 * the queue. As such, we should signal the poll
1640 		 * thread to try and get more packets.
1641 		 *
1642 		 * We could have pulled this check in the POLL_RING
1643 		 * macro itself but keeping it explicit here makes
1644 		 * the architecture more human understandable.
1645 		 */
1646 		MAC_SRS_POLL_RING(mac_srs);
1647 	}
1648 
1649 again:
1650 	head = mac_srs->srs_first;
1651 	mac_srs->srs_first = NULL;
1652 	tail = mac_srs->srs_last;
1653 	mac_srs->srs_last = NULL;
1654 	cnt = mac_srs->srs_count;
1655 	mac_srs->srs_count = 0;
1656 
1657 	ASSERT(head != NULL);
1658 	ASSERT(tail != NULL);
1659 
1660 	if ((tid = mac_srs->srs_tid) != 0)
1661 		mac_srs->srs_tid = 0;
1662 
1663 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1664 
1665 
1666 	/*
1667 	 * mcip is NULL for broadcast and multicast flows. The promisc
1668 	 * callbacks for broadcast and multicast packets are delivered from
1669 	 * mac_rx() and we don't need to worry about that case in this path
1670 	 */
1671 	if (mcip != NULL) {
1672 		if (mcip->mci_promisc_list != NULL) {
1673 			mutex_exit(&mac_srs->srs_lock);
1674 			mac_promisc_client_dispatch(mcip, head);
1675 			mutex_enter(&mac_srs->srs_lock);
1676 		}
1677 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
1678 			mutex_exit(&mac_srs->srs_lock);
1679 			mac_protect_intercept_dhcp(mcip, head);
1680 			mutex_enter(&mac_srs->srs_lock);
1681 		}
1682 	}
1683 
1684 	/*
1685 	 * Check if SRS itself is doing the processing
1686 	 * This direct path does not apply when subflows are present. In this
1687 	 * case, packets need to be dispatched to a soft ring according to the
1688 	 * flow's bandwidth and other resources contraints.
1689 	 */
1690 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1691 		mac_direct_rx_t		proc;
1692 		void			*arg1;
1693 		mac_resource_handle_t	arg2;
1694 
1695 		/*
1696 		 * This is the case when a Rx is directly
1697 		 * assigned and we have a fully classified
1698 		 * protocol chain. We can deal with it in
1699 		 * one shot.
1700 		 */
1701 		proc = srs_rx->sr_func;
1702 		arg1 = srs_rx->sr_arg1;
1703 		arg2 = srs_rx->sr_arg2;
1704 
1705 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1706 		mutex_exit(&mac_srs->srs_lock);
1707 		if (tid != 0) {
1708 			(void) untimeout(tid);
1709 			tid = 0;
1710 		}
1711 
1712 		proc(arg1, arg2, head, NULL);
1713 		/*
1714 		 * Decrement the size and count here itelf
1715 		 * since the packet has been processed.
1716 		 */
1717 		mutex_enter(&mac_srs->srs_lock);
1718 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1719 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
1720 			cv_signal(&mac_srs->srs_client_cv);
1721 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
1722 	} else {
1723 		/* Some kind of softrings based fanout is required */
1724 		mutex_exit(&mac_srs->srs_lock);
1725 		if (tid != 0) {
1726 			(void) untimeout(tid);
1727 			tid = 0;
1728 		}
1729 
1730 		/*
1731 		 * Since the fanout routines can deal with chains,
1732 		 * shoot the entire chain up.
1733 		 */
1734 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
1735 			mac_rx_srs_fanout(mac_srs, head);
1736 		else
1737 			mac_rx_srs_proto_fanout(mac_srs, head);
1738 		mutex_enter(&mac_srs->srs_lock);
1739 	}
1740 
1741 	if (!(mac_srs->srs_state & (SRS_BLANK|SRS_PAUSE)) &&
1742 	    (mac_srs->srs_first != NULL)) {
1743 		/*
1744 		 * More packets arrived while we were clearing the
1745 		 * SRS. This can be possible because of one of
1746 		 * three conditions below:
1747 		 * 1) The driver is using multiple worker threads
1748 		 *    to send the packets to us.
1749 		 * 2) The driver has a race in switching
1750 		 *    between interrupt and polling mode or
1751 		 * 3) Packets are arriving in this SRS via the
1752 		 *    S/W classification as well.
1753 		 *
1754 		 * We should switch to polling mode and see if we
1755 		 * need to send the poll thread down. Also, signal
1756 		 * the worker thread to process whats just arrived.
1757 		 */
1758 		MAC_SRS_POLLING_ON(mac_srs);
1759 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat) {
1760 			srs_rx->sr_drain_poll_sig++;
1761 			MAC_SRS_POLL_RING(mac_srs);
1762 		}
1763 
1764 		/*
1765 		 * If we didn't signal the poll thread, we need
1766 		 * to deal with the pending packets ourselves.
1767 		 */
1768 		if (proc_type == SRS_WORKER) {
1769 			srs_rx->sr_drain_again++;
1770 			goto again;
1771 		} else {
1772 			srs_rx->sr_drain_worker_sig++;
1773 			cv_signal(&mac_srs->srs_async);
1774 		}
1775 	}
1776 
1777 out:
1778 	if (mac_srs->srs_state & SRS_GET_PKTS) {
1779 		/*
1780 		 * Poll thread is already running. Leave the
1781 		 * SRS_RPOC set and hand over the control to
1782 		 * poll thread.
1783 		 */
1784 		mac_srs->srs_state &= ~proc_type;
1785 		srs_rx->sr_drain_poll_running++;
1786 		return;
1787 	}
1788 
1789 	/*
1790 	 * Even if there are no packets queued in SRS, we
1791 	 * need to make sure that the shared counter is
1792 	 * clear and any associated softrings have cleared
1793 	 * all the backlog. Otherwise, leave the interface
1794 	 * in polling mode and the poll thread will get
1795 	 * signalled once the count goes down to zero.
1796 	 *
1797 	 * If someone is already draining the queue (SRS_PROC is
1798 	 * set) when the srs_poll_pkt_cnt goes down to zero,
1799 	 * then it means that drain is already running and we
1800 	 * will turn off polling at that time if there is
1801 	 * no backlog.
1802 	 *
1803 	 * As long as there are packets queued either
1804 	 * in soft ring set or its soft rings, we will leave
1805 	 * the interface in polling mode (even if the drain
1806 	 * was done being the interrupt thread). We signal
1807 	 * the poll thread as well if we have dipped below
1808 	 * low water mark.
1809 	 *
1810 	 * NOTE: We can't use the MAC_SRS_POLLING_ON macro
1811 	 * since that turn polling on only for worker thread.
1812 	 * Its not worth turning polling on for interrupt
1813 	 * thread (since NIC will not issue another interrupt)
1814 	 * unless a backlog builds up.
1815 	 */
1816 	if ((srs_rx->sr_poll_pkt_cnt > 0) &&
1817 	    (mac_srs->srs_state & SRS_POLLING_CAPAB)) {
1818 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1819 		srs_rx->sr_drain_keep_polling++;
1820 		MAC_SRS_POLLING_ON(mac_srs);
1821 		if (srs_rx->sr_poll_pkt_cnt <= srs_rx->sr_lowat)
1822 			MAC_SRS_POLL_RING(mac_srs);
1823 		return;
1824 	}
1825 
1826 	/* Nothing else to do. Get out of poll mode */
1827 	MAC_SRS_POLLING_OFF(mac_srs);
1828 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
1829 	srs_rx->sr_drain_finish_intr++;
1830 }
1831 
1832 /*
1833  * mac_rx_srs_drain_bw
1834  *
1835  * The SRS BW drain routine. Gets to run to clear the queue. Any thread
1836  * (worker, interrupt, poll) can call this based on processing model.
1837  * The first thing we do is disable interrupts if possible and then
1838  * drain the queue. we also try to poll the underlying hardware if
1839  * there is a dedicated hardware Rx ring assigned to this SRS.
1840  *
1841  * There is a equivalent drain routine in non bandwidth control mode
1842  * mac_rx_srs_drain. There is some code duplication between the two
1843  * routines but they are highly performance sensitive and are easier
1844  * to read/debug if they stay separate. Any code changes here might
1845  * also apply to mac_rx_srs_drain as well.
1846  */
1847 void
1848 mac_rx_srs_drain_bw(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
1849 {
1850 	mblk_t 			*head;
1851 	mblk_t			*tail;
1852 	timeout_id_t 		tid;
1853 	size_t			sz = 0;
1854 	int			cnt = 0;
1855 	mac_client_impl_t	*mcip = mac_srs->srs_mcip;
1856 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
1857 	clock_t			now;
1858 
1859 	ASSERT(MUTEX_HELD(&mac_srs->srs_lock));
1860 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
1861 again:
1862 	/* Check if we are doing B/W control */
1863 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1864 	now = ddi_get_lbolt();
1865 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
1866 		mac_srs->srs_bw->mac_bw_curr_time = now;
1867 		mac_srs->srs_bw->mac_bw_used = 0;
1868 		if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
1869 			mac_srs->srs_bw->mac_bw_state &= ~SRS_BW_ENFORCED;
1870 	} else if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) {
1871 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1872 		goto done;
1873 	} else if (mac_srs->srs_bw->mac_bw_used >
1874 	    mac_srs->srs_bw->mac_bw_limit) {
1875 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
1876 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1877 		goto done;
1878 	}
1879 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1880 
1881 	/* If we are blanked i.e. can't do upcalls, then we are done */
1882 	if (mac_srs->srs_state & (SRS_BLANK | SRS_PAUSE)) {
1883 		ASSERT((mac_srs->srs_type & SRST_NO_SOFT_RINGS) ||
1884 		    (mac_srs->srs_state & SRS_PAUSE));
1885 		goto done;
1886 	}
1887 
1888 	sz = 0;
1889 	cnt = 0;
1890 	if ((head = mac_srs_pick_chain(mac_srs, &tail, &sz, &cnt)) == NULL) {
1891 		/*
1892 		 * We couldn't pick up a single packet.
1893 		 */
1894 		mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1895 		if ((mac_srs->srs_bw->mac_bw_used == 0) &&
1896 		    (mac_srs->srs_size != 0) &&
1897 		    !(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
1898 			/*
1899 			 * Seems like configured B/W doesn't
1900 			 * even allow processing of 1 packet
1901 			 * per tick.
1902 			 *
1903 			 * XXX: raise the limit to processing
1904 			 * at least 1 packet per tick.
1905 			 */
1906 			mac_srs->srs_bw->mac_bw_limit +=
1907 			    mac_srs->srs_bw->mac_bw_limit;
1908 			mac_srs->srs_bw->mac_bw_drop_threshold +=
1909 			    mac_srs->srs_bw->mac_bw_drop_threshold;
1910 			cmn_err(CE_NOTE, "mac_rx_srs_drain: srs(%p) "
1911 			    "raised B/W limit to %d since not even a "
1912 			    "single packet can be processed per "
1913 			    "tick %d\n", (void *)mac_srs,
1914 			    (int)mac_srs->srs_bw->mac_bw_limit,
1915 			    (int)msgdsize(mac_srs->srs_first));
1916 		}
1917 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1918 		goto done;
1919 	}
1920 
1921 	ASSERT(head != NULL);
1922 	ASSERT(tail != NULL);
1923 
1924 	/* zero bandwidth: drop all and return to interrupt mode */
1925 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
1926 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
1927 		srs_rx->sr_stat.mrs_sdrops += cnt;
1928 		ASSERT(mac_srs->srs_bw->mac_bw_sz >= sz);
1929 		mac_srs->srs_bw->mac_bw_sz -= sz;
1930 		mac_srs->srs_bw->mac_bw_drop_bytes += sz;
1931 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1932 		mac_pkt_drop(NULL, NULL, head, B_FALSE);
1933 		goto leave_poll;
1934 	} else {
1935 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
1936 	}
1937 
1938 	if ((tid = mac_srs->srs_tid) != 0)
1939 		mac_srs->srs_tid = 0;
1940 
1941 	mac_srs->srs_state |= (SRS_PROC|proc_type);
1942 	MAC_SRS_WORKER_POLLING_ON(mac_srs);
1943 
1944 	/*
1945 	 * mcip is NULL for broadcast and multicast flows. The promisc
1946 	 * callbacks for broadcast and multicast packets are delivered from
1947 	 * mac_rx() and we don't need to worry about that case in this path
1948 	 */
1949 	if (mcip != NULL) {
1950 		if (mcip->mci_promisc_list != NULL) {
1951 			mutex_exit(&mac_srs->srs_lock);
1952 			mac_promisc_client_dispatch(mcip, head);
1953 			mutex_enter(&mac_srs->srs_lock);
1954 		}
1955 		if (MAC_PROTECT_ENABLED(mcip, MPT_IPNOSPOOF)) {
1956 			mutex_exit(&mac_srs->srs_lock);
1957 			mac_protect_intercept_dhcp(mcip, head);
1958 			mutex_enter(&mac_srs->srs_lock);
1959 		}
1960 	}
1961 
1962 	/*
1963 	 * Check if SRS itself is doing the processing
1964 	 * This direct path does not apply when subflows are present. In this
1965 	 * case, packets need to be dispatched to a soft ring according to the
1966 	 * flow's bandwidth and other resources contraints.
1967 	 */
1968 	if (mac_srs->srs_type & SRST_NO_SOFT_RINGS) {
1969 		mac_direct_rx_t		proc;
1970 		void			*arg1;
1971 		mac_resource_handle_t	arg2;
1972 
1973 		/*
1974 		 * This is the case when a Rx is directly
1975 		 * assigned and we have a fully classified
1976 		 * protocol chain. We can deal with it in
1977 		 * one shot.
1978 		 */
1979 		proc = srs_rx->sr_func;
1980 		arg1 = srs_rx->sr_arg1;
1981 		arg2 = srs_rx->sr_arg2;
1982 
1983 		mac_srs->srs_state |= SRS_CLIENT_PROC;
1984 		mutex_exit(&mac_srs->srs_lock);
1985 		if (tid != 0) {
1986 			(void) untimeout(tid);
1987 			tid = 0;
1988 		}
1989 
1990 		proc(arg1, arg2, head, NULL);
1991 		/*
1992 		 * Decrement the size and count here itelf
1993 		 * since the packet has been processed.
1994 		 */
1995 		mutex_enter(&mac_srs->srs_lock);
1996 		MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
1997 		MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
1998 
1999 		if (mac_srs->srs_state & SRS_CLIENT_WAIT)
2000 			cv_signal(&mac_srs->srs_client_cv);
2001 		mac_srs->srs_state &= ~SRS_CLIENT_PROC;
2002 	} else {
2003 		/* Some kind of softrings based fanout is required */
2004 		mutex_exit(&mac_srs->srs_lock);
2005 		if (tid != 0) {
2006 			(void) untimeout(tid);
2007 			tid = 0;
2008 		}
2009 
2010 		/*
2011 		 * Since the fanout routines can deal with chains,
2012 		 * shoot the entire chain up.
2013 		 */
2014 		if (mac_srs->srs_type & SRST_FANOUT_SRC_IP)
2015 			mac_rx_srs_fanout(mac_srs, head);
2016 		else
2017 			mac_rx_srs_proto_fanout(mac_srs, head);
2018 		mutex_enter(&mac_srs->srs_lock);
2019 	}
2020 
2021 	/*
2022 	 * Send the poll thread to pick up any packets arrived
2023 	 * so far. This also serves as the last check in case
2024 	 * nothing else is queued in the SRS. The poll thread
2025 	 * is signalled only in the case the drain was done
2026 	 * by the worker thread and SRS_WORKER is set. The
2027 	 * worker thread can run in parallel as long as the
2028 	 * SRS_WORKER flag is set. We we have nothing else to
2029 	 * process, we can exit while leaving SRS_PROC set
2030 	 * which gives the poll thread control to process and
2031 	 * cleanup once it returns from the NIC.
2032 	 *
2033 	 * If we have nothing else to process, we need to
2034 	 * ensure that we keep holding the srs_lock till
2035 	 * all the checks below are done and control is
2036 	 * handed to the poll thread if it was running.
2037 	 */
2038 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2039 	if (!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2040 		if (mac_srs->srs_first != NULL) {
2041 			if (proc_type == SRS_WORKER) {
2042 				mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2043 				if (srs_rx->sr_poll_pkt_cnt <=
2044 				    srs_rx->sr_lowat)
2045 					MAC_SRS_POLL_RING(mac_srs);
2046 				goto again;
2047 			} else {
2048 				cv_signal(&mac_srs->srs_async);
2049 			}
2050 		}
2051 	}
2052 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2053 
2054 done:
2055 
2056 	if (mac_srs->srs_state & SRS_GET_PKTS) {
2057 		/*
2058 		 * Poll thread is already running. Leave the
2059 		 * SRS_RPOC set and hand over the control to
2060 		 * poll thread.
2061 		 */
2062 		mac_srs->srs_state &= ~proc_type;
2063 		return;
2064 	}
2065 
2066 	/*
2067 	 * If we can't process packets because we have exceeded
2068 	 * B/W limit for this tick, just set the timeout
2069 	 * and leave.
2070 	 *
2071 	 * Even if there are no packets queued in SRS, we
2072 	 * need to make sure that the shared counter is
2073 	 * clear and any associated softrings have cleared
2074 	 * all the backlog. Otherwise, leave the interface
2075 	 * in polling mode and the poll thread will get
2076 	 * signalled once the count goes down to zero.
2077 	 *
2078 	 * If someone is already draining the queue (SRS_PROC is
2079 	 * set) when the srs_poll_pkt_cnt goes down to zero,
2080 	 * then it means that drain is already running and we
2081 	 * will turn off polling at that time if there is
2082 	 * no backlog. As long as there are packets queued either
2083 	 * is soft ring set or its soft rings, we will leave
2084 	 * the interface in polling mode.
2085 	 */
2086 	mutex_enter(&mac_srs->srs_bw->mac_bw_lock);
2087 	if ((mac_srs->srs_state & SRS_POLLING_CAPAB) &&
2088 	    ((mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED) ||
2089 	    (srs_rx->sr_poll_pkt_cnt > 0))) {
2090 		MAC_SRS_POLLING_ON(mac_srs);
2091 		mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2092 		if ((mac_srs->srs_first != NULL) &&
2093 		    (mac_srs->srs_tid == NULL))
2094 			mac_srs->srs_tid = timeout(mac_srs_fire,
2095 			    mac_srs, 1);
2096 		mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2097 		return;
2098 	}
2099 	mutex_exit(&mac_srs->srs_bw->mac_bw_lock);
2100 
2101 leave_poll:
2102 
2103 	/* Nothing else to do. Get out of poll mode */
2104 	MAC_SRS_POLLING_OFF(mac_srs);
2105 	mac_srs->srs_state &= ~(SRS_PROC|proc_type);
2106 }
2107 
2108 /*
2109  * mac_srs_worker
2110  *
2111  * The SRS worker routine. Drains the queue when no one else is
2112  * processing it.
2113  */
2114 void
2115 mac_srs_worker(mac_soft_ring_set_t *mac_srs)
2116 {
2117 	kmutex_t 		*lock = &mac_srs->srs_lock;
2118 	kcondvar_t 		*async = &mac_srs->srs_async;
2119 	callb_cpr_t		cprinfo;
2120 	boolean_t		bw_ctl_flag;
2121 
2122 	CALLB_CPR_INIT(&cprinfo, lock, callb_generic_cpr, "srs_worker");
2123 	mutex_enter(lock);
2124 
2125 start:
2126 	for (;;) {
2127 		bw_ctl_flag = B_FALSE;
2128 		if (mac_srs->srs_type & SRST_BW_CONTROL) {
2129 			MAC_SRS_BW_LOCK(mac_srs);
2130 			MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2131 			if (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)
2132 				bw_ctl_flag = B_TRUE;
2133 			MAC_SRS_BW_UNLOCK(mac_srs);
2134 		}
2135 		/*
2136 		 * The SRS_BW_ENFORCED flag may change since we have dropped
2137 		 * the mac_bw_lock. However the drain function can handle both
2138 		 * a drainable SRS or a bandwidth controlled SRS, and the
2139 		 * effect of scheduling a timeout is to wakeup the worker
2140 		 * thread which in turn will call the drain function. Since
2141 		 * we release the srs_lock atomically only in the cv_wait there
2142 		 * isn't a fear of waiting for ever.
2143 		 */
2144 		while (((mac_srs->srs_state & SRS_PROC) ||
2145 		    (mac_srs->srs_first == NULL) || bw_ctl_flag ||
2146 		    (mac_srs->srs_state & SRS_TX_BLOCKED)) &&
2147 		    !(mac_srs->srs_state & SRS_PAUSE)) {
2148 			/*
2149 			 * If we have packets queued and we are here
2150 			 * because B/W control is in place, we better
2151 			 * schedule the worker wakeup after 1 tick
2152 			 * to see if bandwidth control can be relaxed.
2153 			 */
2154 			if (bw_ctl_flag && mac_srs->srs_tid == NULL) {
2155 				/*
2156 				 * We need to ensure that a timer  is already
2157 				 * scheduled or we force  schedule one for
2158 				 * later so that we can continue processing
2159 				 * after this  quanta is over.
2160 				 */
2161 				mac_srs->srs_tid = timeout(mac_srs_fire,
2162 				    mac_srs, 1);
2163 			}
2164 wait:
2165 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
2166 			cv_wait(async, lock);
2167 			CALLB_CPR_SAFE_END(&cprinfo, lock);
2168 
2169 			if (mac_srs->srs_state & SRS_PAUSE)
2170 				goto done;
2171 			if (mac_srs->srs_state & SRS_PROC)
2172 				goto wait;
2173 
2174 			if (mac_srs->srs_first != NULL &&
2175 			    mac_srs->srs_type & SRST_BW_CONTROL) {
2176 				MAC_SRS_BW_LOCK(mac_srs);
2177 				if (mac_srs->srs_bw->mac_bw_state &
2178 				    SRS_BW_ENFORCED) {
2179 					MAC_SRS_CHECK_BW_CONTROL(mac_srs);
2180 				}
2181 				bw_ctl_flag = mac_srs->srs_bw->mac_bw_state &
2182 				    SRS_BW_ENFORCED;
2183 				MAC_SRS_BW_UNLOCK(mac_srs);
2184 			}
2185 		}
2186 
2187 		if (mac_srs->srs_state & SRS_PAUSE)
2188 			goto done;
2189 		mac_srs->srs_drain_func(mac_srs, SRS_WORKER);
2190 	}
2191 done:
2192 	/*
2193 	 * The Rx SRS quiesce logic first cuts off packet supply to the SRS
2194 	 * from both hard and soft classifications and waits for such threads
2195 	 * to finish before signaling the worker. So at this point the only
2196 	 * thread left that could be competing with the worker is the poll
2197 	 * thread. In the case of Tx, there shouldn't be any thread holding
2198 	 * SRS_PROC at this point.
2199 	 */
2200 	if (!(mac_srs->srs_state & SRS_PROC)) {
2201 		mac_srs->srs_state |= SRS_PROC;
2202 	} else {
2203 		ASSERT((mac_srs->srs_type & SRST_TX) == 0);
2204 		/*
2205 		 * Poll thread still owns the SRS and is still running
2206 		 */
2207 		ASSERT((mac_srs->srs_poll_thr == NULL) ||
2208 		    ((mac_srs->srs_state & SRS_POLL_THR_OWNER) ==
2209 		    SRS_POLL_THR_OWNER));
2210 	}
2211 	mac_srs_worker_quiesce(mac_srs);
2212 	/*
2213 	 * Wait for the SRS_RESTART or SRS_CONDEMNED signal from the initiator
2214 	 * of the quiesce operation
2215 	 */
2216 	while (!(mac_srs->srs_state & (SRS_CONDEMNED | SRS_RESTART)))
2217 		cv_wait(&mac_srs->srs_async, &mac_srs->srs_lock);
2218 
2219 	if (mac_srs->srs_state & SRS_RESTART) {
2220 		ASSERT(!(mac_srs->srs_state & SRS_CONDEMNED));
2221 		mac_srs_worker_restart(mac_srs);
2222 		mac_srs->srs_state &= ~SRS_PROC;
2223 		goto start;
2224 	}
2225 
2226 	if (!(mac_srs->srs_state & SRS_CONDEMNED_DONE))
2227 		mac_srs_worker_quiesce(mac_srs);
2228 
2229 	mac_srs->srs_state &= ~SRS_PROC;
2230 	/* The macro drops the srs_lock */
2231 	CALLB_CPR_EXIT(&cprinfo);
2232 	thread_exit();
2233 }
2234 
2235 /*
2236  * mac_rx_srs_subflow_process
2237  *
2238  * Receive side routine called from interrupt path when there are
2239  * sub flows present on this SRS.
2240  */
2241 /* ARGSUSED */
2242 void
2243 mac_rx_srs_subflow_process(void *arg, mac_resource_handle_t srs,
2244     mblk_t *mp_chain, boolean_t loopback)
2245 {
2246 	flow_entry_t		*flent = NULL;
2247 	flow_entry_t		*prev_flent = NULL;
2248 	mblk_t			*mp = NULL;
2249 	mblk_t			*tail = NULL;
2250 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2251 	mac_client_impl_t	*mcip;
2252 
2253 	mcip = mac_srs->srs_mcip;
2254 	ASSERT(mcip != NULL);
2255 
2256 	/*
2257 	 * We need to determine the SRS for every packet
2258 	 * by walking the flow table, if we don't get any,
2259 	 * then we proceed using the SRS we came with.
2260 	 */
2261 	mp = tail = mp_chain;
2262 	while (mp != NULL) {
2263 
2264 		/*
2265 		 * We will increment the stats for the mactching subflow.
2266 		 * when we get the bytes/pkt count for the classified packets
2267 		 * later in mac_rx_srs_process.
2268 		 */
2269 		(void) mac_flow_lookup(mcip->mci_subflow_tab, mp,
2270 		    FLOW_INBOUND, &flent);
2271 
2272 		if (mp == mp_chain || flent == prev_flent) {
2273 			if (prev_flent != NULL)
2274 				FLOW_REFRELE(prev_flent);
2275 			prev_flent = flent;
2276 			flent = NULL;
2277 			tail = mp;
2278 			mp = mp->b_next;
2279 			continue;
2280 		}
2281 		tail->b_next = NULL;
2282 		/*
2283 		 * A null indicates, this is for the mac_srs itself.
2284 		 * XXX-venu : probably assert for fe_rx_srs_cnt == 0.
2285 		 */
2286 		if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2287 			mac_rx_srs_process(arg,
2288 			    (mac_resource_handle_t)mac_srs, mp_chain,
2289 			    loopback);
2290 		} else {
2291 			(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2292 			    prev_flent->fe_cb_arg2, mp_chain, loopback);
2293 			FLOW_REFRELE(prev_flent);
2294 		}
2295 		prev_flent = flent;
2296 		flent = NULL;
2297 		mp_chain = mp;
2298 		tail = mp;
2299 		mp = mp->b_next;
2300 	}
2301 	/* Last chain */
2302 	ASSERT(mp_chain != NULL);
2303 	if (prev_flent == NULL || prev_flent->fe_rx_srs_cnt == 0) {
2304 		mac_rx_srs_process(arg,
2305 		    (mac_resource_handle_t)mac_srs, mp_chain, loopback);
2306 	} else {
2307 		(prev_flent->fe_cb_fn)(prev_flent->fe_cb_arg1,
2308 		    prev_flent->fe_cb_arg2, mp_chain, loopback);
2309 		FLOW_REFRELE(prev_flent);
2310 	}
2311 }
2312 
2313 /*
2314  * mac_rx_srs_process
2315  *
2316  * Receive side routine called from the interrupt path.
2317  *
2318  * loopback is set to force a context switch on the loopback
2319  * path between MAC clients.
2320  */
2321 /* ARGSUSED */
2322 void
2323 mac_rx_srs_process(void *arg, mac_resource_handle_t srs, mblk_t *mp_chain,
2324     boolean_t loopback)
2325 {
2326 	mac_soft_ring_set_t	*mac_srs = (mac_soft_ring_set_t *)srs;
2327 	mblk_t			*mp, *tail, *head;
2328 	int			count = 0;
2329 	int			count1;
2330 	size_t			sz = 0;
2331 	size_t			chain_sz, sz1;
2332 	mac_bw_ctl_t		*mac_bw;
2333 	mac_srs_rx_t		*srs_rx = &mac_srs->srs_rx;
2334 
2335 	/*
2336 	 * Set the tail, count and sz. We set the sz irrespective
2337 	 * of whether we are doing B/W control or not for the
2338 	 * purpose of updating the stats.
2339 	 */
2340 	mp = tail = mp_chain;
2341 	while (mp != NULL) {
2342 		tail = mp;
2343 		count++;
2344 		sz += msgdsize(mp);
2345 		mp = mp->b_next;
2346 	}
2347 
2348 	mutex_enter(&mac_srs->srs_lock);
2349 
2350 	if (loopback) {
2351 		SRS_RX_STAT_UPDATE(mac_srs, lclbytes, sz);
2352 		SRS_RX_STAT_UPDATE(mac_srs, lclcnt, count);
2353 
2354 	} else {
2355 		SRS_RX_STAT_UPDATE(mac_srs, intrbytes, sz);
2356 		SRS_RX_STAT_UPDATE(mac_srs, intrcnt, count);
2357 	}
2358 
2359 	/*
2360 	 * If the SRS in already being processed; has been blanked;
2361 	 * can be processed by worker thread only; or the B/W limit
2362 	 * has been reached, then queue the chain and check if
2363 	 * worker thread needs to be awakend.
2364 	 */
2365 	if (mac_srs->srs_type & SRST_BW_CONTROL) {
2366 		mac_bw = mac_srs->srs_bw;
2367 		ASSERT(mac_bw != NULL);
2368 		mutex_enter(&mac_bw->mac_bw_lock);
2369 		mac_bw->mac_bw_intr += sz;
2370 		if (mac_bw->mac_bw_limit == 0) {
2371 			/* zero bandwidth: drop all */
2372 			srs_rx->sr_stat.mrs_sdrops += count;
2373 			mac_bw->mac_bw_drop_bytes += sz;
2374 			mutex_exit(&mac_bw->mac_bw_lock);
2375 			mutex_exit(&mac_srs->srs_lock);
2376 			mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
2377 			return;
2378 		} else {
2379 			if ((mac_bw->mac_bw_sz + sz) <=
2380 			    mac_bw->mac_bw_drop_threshold) {
2381 				mutex_exit(&mac_bw->mac_bw_lock);
2382 				MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain,
2383 				    tail, count, sz);
2384 			} else {
2385 				mp = mp_chain;
2386 				chain_sz = 0;
2387 				count1 = 0;
2388 				tail = NULL;
2389 				head = NULL;
2390 				while (mp != NULL) {
2391 					sz1 = msgdsize(mp);
2392 					if (mac_bw->mac_bw_sz + chain_sz + sz1 >
2393 					    mac_bw->mac_bw_drop_threshold)
2394 						break;
2395 					chain_sz += sz1;
2396 					count1++;
2397 					tail = mp;
2398 					mp = mp->b_next;
2399 				}
2400 				mutex_exit(&mac_bw->mac_bw_lock);
2401 				if (tail != NULL) {
2402 					head = tail->b_next;
2403 					tail->b_next = NULL;
2404 					MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs,
2405 					    mp_chain, tail, count1, chain_sz);
2406 					sz -= chain_sz;
2407 					count -= count1;
2408 				} else {
2409 					/* Can't pick up any */
2410 					head = mp_chain;
2411 				}
2412 				if (head != NULL) {
2413 					/* Drop any packet over the threshold */
2414 					srs_rx->sr_stat.mrs_sdrops += count;
2415 					mutex_enter(&mac_bw->mac_bw_lock);
2416 					mac_bw->mac_bw_drop_bytes += sz;
2417 					mutex_exit(&mac_bw->mac_bw_lock);
2418 					freemsgchain(head);
2419 				}
2420 			}
2421 			MAC_SRS_WORKER_WAKEUP(mac_srs);
2422 			mutex_exit(&mac_srs->srs_lock);
2423 			return;
2424 		}
2425 	}
2426 
2427 	/*
2428 	 * If the total number of packets queued in the SRS and
2429 	 * its associated soft rings exceeds the max allowed,
2430 	 * then drop the chain. If we are polling capable, this
2431 	 * shouldn't be happening.
2432 	 */
2433 	if (!(mac_srs->srs_type & SRST_BW_CONTROL) &&
2434 	    (srs_rx->sr_poll_pkt_cnt > srs_rx->sr_hiwat)) {
2435 		mac_bw = mac_srs->srs_bw;
2436 		srs_rx->sr_stat.mrs_sdrops += count;
2437 		mutex_enter(&mac_bw->mac_bw_lock);
2438 		mac_bw->mac_bw_drop_bytes += sz;
2439 		mutex_exit(&mac_bw->mac_bw_lock);
2440 		freemsgchain(mp_chain);
2441 		mutex_exit(&mac_srs->srs_lock);
2442 		return;
2443 	}
2444 
2445 	MAC_RX_SRS_ENQUEUE_CHAIN(mac_srs, mp_chain, tail, count, sz);
2446 
2447 	if (!(mac_srs->srs_state & SRS_PROC)) {
2448 		/*
2449 		 * If we are coming via loopback, if we are not optimizing for
2450 		 * latency, or if our stack is running deep, we should signal
2451 		 * the worker thread.
2452 		 */
2453 		if (loopback || !(mac_srs->srs_state & SRS_LATENCY_OPT) ||
2454 		    MAC_RX_SRS_TOODEEP()) {
2455 			/*
2456 			 * For loopback, We need to let the worker take
2457 			 * over as we don't want to continue in the same
2458 			 * thread even if we can. This could lead to stack
2459 			 * overflows and may also end up using
2460 			 * resources (cpu) incorrectly.
2461 			 */
2462 			cv_signal(&mac_srs->srs_async);
2463 		} else {
2464 			/*
2465 			 * Seems like no one is processing the SRS and
2466 			 * there is no backlog. We also inline process
2467 			 * our packet if its a single packet in non
2468 			 * latency optimized case (in latency optimized
2469 			 * case, we inline process chains of any size).
2470 			 */
2471 			mac_srs->srs_drain_func(mac_srs, SRS_PROC_FAST);
2472 		}
2473 	}
2474 	mutex_exit(&mac_srs->srs_lock);
2475 }
2476 
2477 /* TX SIDE ROUTINES (RUNTIME) */
2478 
2479 /*
2480  * mac_tx_srs_no_desc
2481  *
2482  * This routine is called by Tx single ring default mode
2483  * when Tx ring runs out of descs.
2484  */
2485 mac_tx_cookie_t
2486 mac_tx_srs_no_desc(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2487     uint16_t flag, mblk_t **ret_mp)
2488 {
2489 	mac_tx_cookie_t cookie = NULL;
2490 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
2491 	boolean_t wakeup_worker = B_TRUE;
2492 	uint32_t tx_mode = srs_tx->st_mode;
2493 	int cnt, sz;
2494 	mblk_t *tail;
2495 
2496 	ASSERT(tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_BW);
2497 	if (flag & MAC_DROP_ON_NO_DESC) {
2498 		MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2499 	} else {
2500 		if (mac_srs->srs_first != NULL)
2501 			wakeup_worker = B_FALSE;
2502 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2503 		if (flag & MAC_TX_NO_ENQUEUE) {
2504 			/*
2505 			 * If TX_QUEUED is not set, queue the
2506 			 * packet and let mac_tx_srs_drain()
2507 			 * set the TX_BLOCKED bit for the
2508 			 * reasons explained above. Otherwise,
2509 			 * return the mblks.
2510 			 */
2511 			if (wakeup_worker) {
2512 				MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2513 				    mp_chain, tail, cnt, sz);
2514 			} else {
2515 				MAC_TX_SET_NO_ENQUEUE(mac_srs,
2516 				    mp_chain, ret_mp, cookie);
2517 			}
2518 		} else {
2519 			MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2520 			    tail, cnt, sz, cookie);
2521 		}
2522 		if (wakeup_worker)
2523 			cv_signal(&mac_srs->srs_async);
2524 	}
2525 	return (cookie);
2526 }
2527 
2528 /*
2529  * mac_tx_srs_enqueue
2530  *
2531  * This routine is called when Tx SRS is operating in either serializer
2532  * or bandwidth mode. In serializer mode, a packet will get enqueued
2533  * when a thread cannot enter SRS exclusively. In bandwidth mode,
2534  * packets gets queued if allowed byte-count limit for a tick is
2535  * exceeded. The action that gets taken when MAC_DROP_ON_NO_DESC and
2536  * MAC_TX_NO_ENQUEUE is set is different than when operaing in either
2537  * the default mode or fanout mode. Here packets get dropped or
2538  * returned back to the caller only after hi-watermark worth of data
2539  * is queued.
2540  */
2541 static mac_tx_cookie_t
2542 mac_tx_srs_enqueue(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2543     uint16_t flag, uintptr_t fanout_hint, mblk_t **ret_mp)
2544 {
2545 	mac_tx_cookie_t cookie = NULL;
2546 	int cnt, sz;
2547 	mblk_t *tail;
2548 	boolean_t wakeup_worker = B_TRUE;
2549 
2550 	/*
2551 	 * Ignore fanout hint if we don't have multiple tx rings.
2552 	 */
2553 	if (!MAC_TX_SOFT_RINGS(mac_srs))
2554 		fanout_hint = 0;
2555 
2556 	if (mac_srs->srs_first != NULL)
2557 		wakeup_worker = B_FALSE;
2558 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2559 	if (flag & MAC_DROP_ON_NO_DESC) {
2560 		if (mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) {
2561 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2562 		} else {
2563 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2564 			    mp_chain, tail, cnt, sz);
2565 		}
2566 	} else if (flag & MAC_TX_NO_ENQUEUE) {
2567 		if ((mac_srs->srs_count > mac_srs->srs_tx.st_hiwat) ||
2568 		    (mac_srs->srs_state & SRS_TX_WAKEUP_CLIENT)) {
2569 			MAC_TX_SET_NO_ENQUEUE(mac_srs, mp_chain,
2570 			    ret_mp, cookie);
2571 		} else {
2572 			mp_chain->b_prev = (mblk_t *)fanout_hint;
2573 			MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2574 			    mp_chain, tail, cnt, sz);
2575 		}
2576 	} else {
2577 		/*
2578 		 * If you are BW_ENFORCED, just enqueue the
2579 		 * packet. srs_worker will drain it at the
2580 		 * prescribed rate. Before enqueueing, save
2581 		 * the fanout hint.
2582 		 */
2583 		mp_chain->b_prev = (mblk_t *)fanout_hint;
2584 		MAC_TX_SRS_TEST_HIWAT(mac_srs, mp_chain,
2585 		    tail, cnt, sz, cookie);
2586 	}
2587 	if (wakeup_worker)
2588 		cv_signal(&mac_srs->srs_async);
2589 	return (cookie);
2590 }
2591 
2592 /*
2593  * There are seven tx modes:
2594  *
2595  * 1) Default mode (SRS_TX_DEFAULT)
2596  * 2) Serialization mode (SRS_TX_SERIALIZE)
2597  * 3) Fanout mode (SRS_TX_FANOUT)
2598  * 4) Bandwdith mode (SRS_TX_BW)
2599  * 5) Fanout and Bandwidth mode (SRS_TX_BW_FANOUT)
2600  * 6) aggr Tx mode (SRS_TX_AGGR)
2601  * 7) aggr Tx bw mode (SRS_TX_BW_AGGR)
2602  *
2603  * The tx mode in which an SRS operates is decided in mac_tx_srs_setup()
2604  * based on the number of Tx rings requested for an SRS and whether
2605  * bandwidth control is requested or not.
2606  *
2607  * The default mode (i.e., no fanout/no bandwidth) is used when the
2608  * underlying NIC does not have Tx rings or just one Tx ring. In this mode,
2609  * the SRS acts as a pass-thru. Packets will go directly to mac_tx_send().
2610  * When the underlying Tx ring runs out of Tx descs, it starts queueing up
2611  * packets in SRS. When flow-control is relieved, the srs_worker drains
2612  * the queued packets and informs blocked clients to restart sending
2613  * packets.
2614  *
2615  * In the SRS_TX_SERIALIZE mode, all calls to mac_tx() are serialized. This
2616  * mode is used when the link has no Tx rings or only one Tx ring.
2617  *
2618  * In the SRS_TX_FANOUT mode, packets will be fanned out to multiple
2619  * Tx rings. Each Tx ring will have a soft ring associated with it.
2620  * These soft rings will be hung off the Tx SRS. Queueing if it happens
2621  * due to lack of Tx desc will be in individual soft ring (and not srs)
2622  * associated with Tx ring.
2623  *
2624  * In the TX_BW mode, tx srs will allow packets to go down to Tx ring
2625  * only if bw is available. Otherwise the packets will be queued in
2626  * SRS. If fanout to multiple Tx rings is configured, the packets will
2627  * be fanned out among the soft rings associated with the Tx rings.
2628  *
2629  * In SRS_TX_AGGR mode, mac_tx_aggr_mode() routine is called. This routine
2630  * invokes an aggr function, aggr_find_tx_ring(), to find a pseudo Tx ring
2631  * belonging to a port on which the packet has to be sent. Aggr will
2632  * always have a pseudo Tx ring associated with it even when it is an
2633  * aggregation over a single NIC that has no Tx rings. Even in such a
2634  * case, the single pseudo Tx ring will have a soft ring associated with
2635  * it and the soft ring will hang off the SRS.
2636  *
2637  * If a bandwidth is specified for an aggr, SRS_TX_BW_AGGR mode is used.
2638  * In this mode, the bandwidth is first applied on the outgoing packets
2639  * and later mac_tx_addr_mode() function is called to send the packet out
2640  * of one of the pseudo Tx rings.
2641  *
2642  * Four flags are used in srs_state for indicating flow control
2643  * conditions : SRS_TX_BLOCKED, SRS_TX_HIWAT, SRS_TX_WAKEUP_CLIENT.
2644  * SRS_TX_BLOCKED indicates out of Tx descs. SRS expects a wakeup from the
2645  * driver below.
2646  * SRS_TX_HIWAT indicates packet count enqueued in Tx SRS exceeded Tx hiwat
2647  * and flow-control pressure is applied back to clients. The clients expect
2648  * wakeup when flow-control is relieved.
2649  * SRS_TX_WAKEUP_CLIENT get set when (flag == MAC_TX_NO_ENQUEUE) and mblk
2650  * got returned back to client either due to lack of Tx descs or due to bw
2651  * control reasons. The clients expect a wakeup when condition is relieved.
2652  *
2653  * The fourth argument to mac_tx() is the flag. Normally it will be 0 but
2654  * some clients set the following values too: MAC_DROP_ON_NO_DESC,
2655  * MAC_TX_NO_ENQUEUE
2656  * Mac clients that do not want packets to be enqueued in the mac layer set
2657  * MAC_DROP_ON_NO_DESC value. The packets won't be queued in the Tx SRS or
2658  * Tx soft rings but instead get dropped when the NIC runs out of desc. The
2659  * behaviour of this flag is different when the Tx is running in serializer
2660  * or bandwidth mode. Under these (Serializer, bandwidth) modes, the packet
2661  * get dropped when Tx high watermark is reached.
2662  * There are some mac clients like vsw, aggr that want the mblks to be
2663  * returned back to clients instead of being queued in Tx SRS (or Tx soft
2664  * rings) under flow-control (i.e., out of desc or exceeding bw limits)
2665  * conditions. These clients call mac_tx() with MAC_TX_NO_ENQUEUE flag set.
2666  * In the default and Tx fanout mode, the un-transmitted mblks will be
2667  * returned back to the clients when the driver runs out of Tx descs.
2668  * SRS_TX_WAKEUP_CLIENT (or S_RING_WAKEUP_CLIENT) will be set in SRS (or
2669  * soft ring) so that the clients can be woken up when Tx desc become
2670  * available. When running in serializer or bandwidth mode mode,
2671  * SRS_TX_WAKEUP_CLIENT will be set when tx hi-watermark is reached.
2672  */
2673 
2674 mac_tx_func_t
2675 mac_tx_get_func(uint32_t mode)
2676 {
2677 	return (mac_tx_mode_list[mode].mac_tx_func);
2678 }
2679 
2680 /* ARGSUSED */
2681 static mac_tx_cookie_t
2682 mac_tx_single_ring_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2683     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2684 {
2685 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2686 	mac_tx_stats_t		stats;
2687 	mac_tx_cookie_t		cookie = NULL;
2688 
2689 	ASSERT(srs_tx->st_mode == SRS_TX_DEFAULT);
2690 
2691 	/* Regular case with a single Tx ring */
2692 	/*
2693 	 * SRS_TX_BLOCKED is set when underlying NIC runs
2694 	 * out of Tx descs and messages start getting
2695 	 * queued. It won't get reset until
2696 	 * tx_srs_drain() completely drains out the
2697 	 * messages.
2698 	 */
2699 	if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2700 		/* Tx descs/resources not available */
2701 		mutex_enter(&mac_srs->srs_lock);
2702 		if ((mac_srs->srs_state & SRS_ENQUEUED) != 0) {
2703 			cookie = mac_tx_srs_no_desc(mac_srs, mp_chain,
2704 			    flag, ret_mp);
2705 			mutex_exit(&mac_srs->srs_lock);
2706 			return (cookie);
2707 		}
2708 		/*
2709 		 * While we were computing mblk count, the
2710 		 * flow control condition got relieved.
2711 		 * Continue with the transmission.
2712 		 */
2713 		mutex_exit(&mac_srs->srs_lock);
2714 	}
2715 
2716 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2717 	    mp_chain, &stats);
2718 
2719 	/*
2720 	 * Multiple threads could be here sending packets.
2721 	 * Under such conditions, it is not possible to
2722 	 * automically set SRS_TX_BLOCKED bit to indicate
2723 	 * out of tx desc condition. To atomically set
2724 	 * this, we queue the returned packet and do
2725 	 * the setting of SRS_TX_BLOCKED in
2726 	 * mac_tx_srs_drain().
2727 	 */
2728 	if (mp_chain != NULL) {
2729 		mutex_enter(&mac_srs->srs_lock);
2730 		cookie = mac_tx_srs_no_desc(mac_srs, mp_chain, flag, ret_mp);
2731 		mutex_exit(&mac_srs->srs_lock);
2732 		return (cookie);
2733 	}
2734 	SRS_TX_STATS_UPDATE(mac_srs, &stats);
2735 
2736 	return (NULL);
2737 }
2738 
2739 /*
2740  * mac_tx_serialize_mode
2741  *
2742  * This is an experimental mode implemented as per the request of PAE.
2743  * In this mode, all callers attempting to send a packet to the NIC
2744  * will get serialized. Only one thread at any time will access the
2745  * NIC to send the packet out.
2746  */
2747 /* ARGSUSED */
2748 static mac_tx_cookie_t
2749 mac_tx_serializer_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2750     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2751 {
2752 	mac_tx_stats_t		stats;
2753 	mac_tx_cookie_t		cookie = NULL;
2754 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2755 
2756 	/* Single ring, serialize below */
2757 	ASSERT(srs_tx->st_mode == SRS_TX_SERIALIZE);
2758 	mutex_enter(&mac_srs->srs_lock);
2759 	if ((mac_srs->srs_first != NULL) ||
2760 	    (mac_srs->srs_state & SRS_PROC)) {
2761 		/*
2762 		 * In serialization mode, queue all packets until
2763 		 * TX_HIWAT is set.
2764 		 * If drop bit is set, drop if TX_HIWAT is set.
2765 		 * If no_enqueue is set, still enqueue until hiwat
2766 		 * is set and return mblks after TX_HIWAT is set.
2767 		 */
2768 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain,
2769 		    flag, NULL, ret_mp);
2770 		mutex_exit(&mac_srs->srs_lock);
2771 		return (cookie);
2772 	}
2773 	/*
2774 	 * No packets queued, nothing on proc and no flow
2775 	 * control condition. Fast-path, ok. Do inline
2776 	 * processing.
2777 	 */
2778 	mac_srs->srs_state |= SRS_PROC;
2779 	mutex_exit(&mac_srs->srs_lock);
2780 
2781 	mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2782 	    mp_chain, &stats);
2783 
2784 	mutex_enter(&mac_srs->srs_lock);
2785 	mac_srs->srs_state &= ~SRS_PROC;
2786 	if (mp_chain != NULL) {
2787 		cookie = mac_tx_srs_enqueue(mac_srs,
2788 		    mp_chain, flag, NULL, ret_mp);
2789 	}
2790 	if (mac_srs->srs_first != NULL) {
2791 		/*
2792 		 * We processed inline our packet and a new
2793 		 * packet/s got queued while we were
2794 		 * processing. Wakeup srs worker
2795 		 */
2796 		cv_signal(&mac_srs->srs_async);
2797 	}
2798 	mutex_exit(&mac_srs->srs_lock);
2799 
2800 	if (cookie == NULL)
2801 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
2802 
2803 	return (cookie);
2804 }
2805 
2806 /*
2807  * mac_tx_fanout_mode
2808  *
2809  * In this mode, the SRS will have access to multiple Tx rings to send
2810  * the packet out. The fanout hint that is passed as an argument is
2811  * used to find an appropriate ring to fanout the traffic. Each Tx
2812  * ring, in turn,  will have a soft ring associated with it. If a Tx
2813  * ring runs out of Tx desc's the returned packet will be queued in
2814  * the soft ring associated with that Tx ring. The srs itself will not
2815  * queue any packets.
2816  */
2817 
2818 #define	MAC_TX_SOFT_RING_PROCESS(chain) {		       		\
2819 	index = COMPUTE_INDEX(hash, mac_srs->srs_tx_ring_count),	\
2820 	softring = mac_srs->srs_tx_soft_rings[index];			\
2821 	cookie = mac_tx_soft_ring_process(softring, chain, flag, ret_mp); \
2822 	DTRACE_PROBE2(tx__fanout, uint64_t, hash, uint_t, index);	\
2823 }
2824 
2825 static mac_tx_cookie_t
2826 mac_tx_fanout_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2827     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2828 {
2829 	mac_soft_ring_t		*softring;
2830 	uint64_t		hash;
2831 	uint_t			index;
2832 	mac_tx_cookie_t		cookie = NULL;
2833 
2834 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
2835 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT);
2836 	if (fanout_hint != 0) {
2837 		/*
2838 		 * The hint is specified by the caller, simply pass the
2839 		 * whole chain to the soft ring.
2840 		 */
2841 		hash = HASH_HINT(fanout_hint);
2842 		MAC_TX_SOFT_RING_PROCESS(mp_chain);
2843 	} else {
2844 		mblk_t *last_mp, *cur_mp, *sub_chain;
2845 		uint64_t last_hash = 0;
2846 		uint_t media = mac_srs->srs_mcip->mci_mip->mi_info.mi_media;
2847 
2848 		/*
2849 		 * Compute the hash from the contents (headers) of the
2850 		 * packets of the mblk chain. Split the chains into
2851 		 * subchains of the same conversation.
2852 		 *
2853 		 * Since there may be more than one ring used for
2854 		 * sub-chains of the same call, and since the caller
2855 		 * does not maintain per conversation state since it
2856 		 * passed a zero hint, unsent subchains will be
2857 		 * dropped.
2858 		 */
2859 
2860 		flag |= MAC_DROP_ON_NO_DESC;
2861 		ret_mp = NULL;
2862 
2863 		ASSERT(ret_mp == NULL);
2864 
2865 		sub_chain = NULL;
2866 		last_mp = NULL;
2867 
2868 		for (cur_mp = mp_chain; cur_mp != NULL;
2869 		    cur_mp = cur_mp->b_next) {
2870 			hash = mac_pkt_hash(media, cur_mp, MAC_PKT_HASH_L4,
2871 			    B_TRUE);
2872 			if (last_hash != 0 && hash != last_hash) {
2873 				/*
2874 				 * Starting a different subchain, send current
2875 				 * chain out.
2876 				 */
2877 				ASSERT(last_mp != NULL);
2878 				last_mp->b_next = NULL;
2879 				MAC_TX_SOFT_RING_PROCESS(sub_chain);
2880 				sub_chain = NULL;
2881 			}
2882 
2883 			/* add packet to subchain */
2884 			if (sub_chain == NULL)
2885 				sub_chain = cur_mp;
2886 			last_mp = cur_mp;
2887 			last_hash = hash;
2888 		}
2889 
2890 		if (sub_chain != NULL) {
2891 			/* send last subchain */
2892 			ASSERT(last_mp != NULL);
2893 			last_mp->b_next = NULL;
2894 			MAC_TX_SOFT_RING_PROCESS(sub_chain);
2895 		}
2896 
2897 		cookie = NULL;
2898 	}
2899 
2900 	return (cookie);
2901 }
2902 
2903 /*
2904  * mac_tx_bw_mode
2905  *
2906  * In the bandwidth mode, Tx srs will allow packets to go down to Tx ring
2907  * only if bw is available. Otherwise the packets will be queued in
2908  * SRS. If the SRS has multiple Tx rings, then packets will get fanned
2909  * out to a Tx rings.
2910  */
2911 static mac_tx_cookie_t
2912 mac_tx_bw_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
2913     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
2914 {
2915 	int			cnt, sz;
2916 	mblk_t			*tail;
2917 	mac_tx_cookie_t		cookie = NULL;
2918 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
2919 	clock_t			now;
2920 
2921 	ASSERT(TX_BANDWIDTH_MODE(mac_srs));
2922 	ASSERT(mac_srs->srs_type & SRST_BW_CONTROL);
2923 	mutex_enter(&mac_srs->srs_lock);
2924 	if (mac_srs->srs_bw->mac_bw_limit == 0) {
2925 		/*
2926 		 * zero bandwidth, no traffic is sent: drop the packets,
2927 		 * or return the whole chain if the caller requests all
2928 		 * unsent packets back.
2929 		 */
2930 		if (flag & MAC_TX_NO_ENQUEUE) {
2931 			cookie = (mac_tx_cookie_t)mac_srs;
2932 			*ret_mp = mp_chain;
2933 		} else {
2934 			MAC_TX_SRS_DROP_MESSAGE(mac_srs, mp_chain, cookie);
2935 		}
2936 		mutex_exit(&mac_srs->srs_lock);
2937 		return (cookie);
2938 	} else if ((mac_srs->srs_first != NULL) ||
2939 	    (mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED)) {
2940 		cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
2941 		    fanout_hint, ret_mp);
2942 		mutex_exit(&mac_srs->srs_lock);
2943 		return (cookie);
2944 	}
2945 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2946 	now = ddi_get_lbolt();
2947 	if (mac_srs->srs_bw->mac_bw_curr_time != now) {
2948 		mac_srs->srs_bw->mac_bw_curr_time = now;
2949 		mac_srs->srs_bw->mac_bw_used = 0;
2950 	} else if (mac_srs->srs_bw->mac_bw_used >
2951 	    mac_srs->srs_bw->mac_bw_limit) {
2952 		mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
2953 		MAC_TX_SRS_ENQUEUE_CHAIN(mac_srs,
2954 		    mp_chain, tail, cnt, sz);
2955 		/*
2956 		 * Wakeup worker thread. Note that worker
2957 		 * thread has to be woken up so that it
2958 		 * can fire up the timer to be woken up
2959 		 * on the next tick. Also once
2960 		 * BW_ENFORCED is set, it can only be
2961 		 * reset by srs_worker thread. Until then
2962 		 * all packets will get queued up in SRS
2963 		 * and hence this this code path won't be
2964 		 * entered until BW_ENFORCED is reset.
2965 		 */
2966 		cv_signal(&mac_srs->srs_async);
2967 		mutex_exit(&mac_srs->srs_lock);
2968 		return (cookie);
2969 	}
2970 
2971 	mac_srs->srs_bw->mac_bw_used += sz;
2972 	mutex_exit(&mac_srs->srs_lock);
2973 
2974 	if (srs_tx->st_mode == SRS_TX_BW_FANOUT) {
2975 		mac_soft_ring_t *softring;
2976 		uint_t indx, hash;
2977 
2978 		hash = HASH_HINT(fanout_hint);
2979 		indx = COMPUTE_INDEX(hash,
2980 		    mac_srs->srs_tx_ring_count);
2981 		softring = mac_srs->srs_tx_soft_rings[indx];
2982 		return (mac_tx_soft_ring_process(softring, mp_chain, flag,
2983 		    ret_mp));
2984 	} else if (srs_tx->st_mode == SRS_TX_BW_AGGR) {
2985 		return (mac_tx_aggr_mode(mac_srs, mp_chain,
2986 		    fanout_hint, flag, ret_mp));
2987 	} else {
2988 		mac_tx_stats_t		stats;
2989 
2990 		mp_chain = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
2991 		    mp_chain, &stats);
2992 
2993 		if (mp_chain != NULL) {
2994 			mutex_enter(&mac_srs->srs_lock);
2995 			MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
2996 			if (mac_srs->srs_bw->mac_bw_used > sz)
2997 				mac_srs->srs_bw->mac_bw_used -= sz;
2998 			else
2999 				mac_srs->srs_bw->mac_bw_used = 0;
3000 			cookie = mac_tx_srs_enqueue(mac_srs, mp_chain, flag,
3001 			    fanout_hint, ret_mp);
3002 			mutex_exit(&mac_srs->srs_lock);
3003 			return (cookie);
3004 		}
3005 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
3006 
3007 		return (NULL);
3008 	}
3009 }
3010 
3011 /*
3012  * mac_tx_aggr_mode
3013  *
3014  * This routine invokes an aggr function, aggr_find_tx_ring(), to find
3015  * a (pseudo) Tx ring belonging to a port on which the packet has to
3016  * be sent. aggr_find_tx_ring() first finds the outgoing port based on
3017  * L2/L3/L4 policy and then uses the fanout_hint passed to it to pick
3018  * a Tx ring from the selected port.
3019  *
3020  * Note that a port can be deleted from the aggregation. In such a case,
3021  * the aggregation layer first separates the port from the rest of the
3022  * ports making sure that port (and thus any Tx rings associated with
3023  * it) won't get selected in the call to aggr_find_tx_ring() function.
3024  * Later calls are made to mac_group_rem_ring() passing pseudo Tx ring
3025  * handles one by one which in turn will quiesce the Tx SRS and remove
3026  * the soft ring associated with the pseudo Tx ring. Unlike Rx side
3027  * where a cookie is used to protect against mac_rx_ring() calls on
3028  * rings that have been removed, no such cookie is needed on the Tx
3029  * side as the pseudo Tx ring won't be available anymore to
3030  * aggr_find_tx_ring() once the port has been removed.
3031  */
3032 static mac_tx_cookie_t
3033 mac_tx_aggr_mode(mac_soft_ring_set_t *mac_srs, mblk_t *mp_chain,
3034     uintptr_t fanout_hint, uint16_t flag, mblk_t **ret_mp)
3035 {
3036 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3037 	mac_tx_ring_fn_t	find_tx_ring_fn;
3038 	mac_ring_handle_t	ring = NULL;
3039 	void			*arg;
3040 	mac_soft_ring_t		*sringp;
3041 
3042 	find_tx_ring_fn = srs_tx->st_capab_aggr.mca_find_tx_ring_fn;
3043 	arg = srs_tx->st_capab_aggr.mca_arg;
3044 	if (find_tx_ring_fn(arg, mp_chain, fanout_hint, &ring) == NULL)
3045 		return (NULL);
3046 	sringp = srs_tx->st_soft_rings[((mac_ring_t *)ring)->mr_index];
3047 	return (mac_tx_soft_ring_process(sringp, mp_chain, flag, ret_mp));
3048 }
3049 
3050 void
3051 mac_tx_invoke_callbacks(mac_client_impl_t *mcip, mac_tx_cookie_t cookie)
3052 {
3053 	mac_cb_t *mcb;
3054 	mac_tx_notify_cb_t *mtnfp;
3055 
3056 	/* Wakeup callback registered clients */
3057 	MAC_CALLBACK_WALKER_INC(&mcip->mci_tx_notify_cb_info);
3058 	for (mcb = mcip->mci_tx_notify_cb_list; mcb != NULL;
3059 	    mcb = mcb->mcb_nextp) {
3060 		mtnfp = (mac_tx_notify_cb_t *)mcb->mcb_objp;
3061 		mtnfp->mtnf_fn(mtnfp->mtnf_arg, cookie);
3062 	}
3063 	MAC_CALLBACK_WALKER_DCR(&mcip->mci_tx_notify_cb_info,
3064 	    &mcip->mci_tx_notify_cb_list);
3065 }
3066 
3067 /* ARGSUSED */
3068 void
3069 mac_tx_srs_drain(mac_soft_ring_set_t *mac_srs, uint_t proc_type)
3070 {
3071 	mblk_t			*head, *tail;
3072 	size_t			sz;
3073 	uint32_t		tx_mode;
3074 	uint_t			saved_pkt_count;
3075 	mac_tx_stats_t		stats;
3076 	mac_srs_tx_t		*srs_tx = &mac_srs->srs_tx;
3077 	clock_t			now;
3078 
3079 	saved_pkt_count = 0;
3080 	ASSERT(mutex_owned(&mac_srs->srs_lock));
3081 	ASSERT(!(mac_srs->srs_state & SRS_PROC));
3082 
3083 	mac_srs->srs_state |= SRS_PROC;
3084 
3085 	tx_mode = srs_tx->st_mode;
3086 	if (tx_mode == SRS_TX_DEFAULT || tx_mode == SRS_TX_SERIALIZE) {
3087 		if (mac_srs->srs_first != NULL) {
3088 			head = mac_srs->srs_first;
3089 			tail = mac_srs->srs_last;
3090 			saved_pkt_count = mac_srs->srs_count;
3091 			mac_srs->srs_first = NULL;
3092 			mac_srs->srs_last = NULL;
3093 			mac_srs->srs_count = 0;
3094 			mutex_exit(&mac_srs->srs_lock);
3095 
3096 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3097 			    head, &stats);
3098 
3099 			mutex_enter(&mac_srs->srs_lock);
3100 			if (head != NULL) {
3101 				/* Device out of tx desc, set block */
3102 				if (head->b_next == NULL)
3103 					VERIFY(head == tail);
3104 				tail->b_next = mac_srs->srs_first;
3105 				mac_srs->srs_first = head;
3106 				mac_srs->srs_count +=
3107 				    (saved_pkt_count - stats.mts_opackets);
3108 				if (mac_srs->srs_last == NULL)
3109 					mac_srs->srs_last = tail;
3110 				MAC_TX_SRS_BLOCK(mac_srs, head);
3111 			} else {
3112 				srs_tx->st_woken_up = B_FALSE;
3113 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
3114 			}
3115 		}
3116 	} else if (tx_mode == SRS_TX_BW) {
3117 		/*
3118 		 * We are here because the timer fired and we have some data
3119 		 * to tranmit. Also mac_tx_srs_worker should have reset
3120 		 * SRS_BW_ENFORCED flag
3121 		 */
3122 		ASSERT(!(mac_srs->srs_bw->mac_bw_state & SRS_BW_ENFORCED));
3123 		head = tail = mac_srs->srs_first;
3124 		while (mac_srs->srs_first != NULL) {
3125 			tail = mac_srs->srs_first;
3126 			tail->b_prev = NULL;
3127 			mac_srs->srs_first = tail->b_next;
3128 			if (mac_srs->srs_first == NULL)
3129 				mac_srs->srs_last = NULL;
3130 			mac_srs->srs_count--;
3131 			sz = msgdsize(tail);
3132 			mac_srs->srs_size -= sz;
3133 			saved_pkt_count++;
3134 			MAC_TX_UPDATE_BW_INFO(mac_srs, sz);
3135 
3136 			if (mac_srs->srs_bw->mac_bw_used <
3137 			    mac_srs->srs_bw->mac_bw_limit)
3138 				continue;
3139 
3140 			now = ddi_get_lbolt();
3141 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3142 				mac_srs->srs_bw->mac_bw_curr_time = now;
3143 				mac_srs->srs_bw->mac_bw_used = sz;
3144 				continue;
3145 			}
3146 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3147 			break;
3148 		}
3149 
3150 		ASSERT((head == NULL && tail == NULL) ||
3151 		    (head != NULL && tail != NULL));
3152 		if (tail != NULL) {
3153 			tail->b_next = NULL;
3154 			mutex_exit(&mac_srs->srs_lock);
3155 
3156 			head = mac_tx_send(srs_tx->st_arg1, srs_tx->st_arg2,
3157 			    head, &stats);
3158 
3159 			mutex_enter(&mac_srs->srs_lock);
3160 			if (head != NULL) {
3161 				uint_t size_sent;
3162 
3163 				/* Device out of tx desc, set block */
3164 				if (head->b_next == NULL)
3165 					VERIFY(head == tail);
3166 				tail->b_next = mac_srs->srs_first;
3167 				mac_srs->srs_first = head;
3168 				mac_srs->srs_count +=
3169 				    (saved_pkt_count - stats.mts_opackets);
3170 				if (mac_srs->srs_last == NULL)
3171 					mac_srs->srs_last = tail;
3172 				size_sent = sz - stats.mts_obytes;
3173 				mac_srs->srs_size += size_sent;
3174 				mac_srs->srs_bw->mac_bw_sz += size_sent;
3175 				if (mac_srs->srs_bw->mac_bw_used > size_sent) {
3176 					mac_srs->srs_bw->mac_bw_used -=
3177 					    size_sent;
3178 				} else {
3179 					mac_srs->srs_bw->mac_bw_used = 0;
3180 				}
3181 				MAC_TX_SRS_BLOCK(mac_srs, head);
3182 			} else {
3183 				srs_tx->st_woken_up = B_FALSE;
3184 				SRS_TX_STATS_UPDATE(mac_srs, &stats);
3185 			}
3186 		}
3187 	} else if (tx_mode == SRS_TX_BW_FANOUT || tx_mode == SRS_TX_BW_AGGR) {
3188 		mblk_t *prev;
3189 		uint64_t hint;
3190 
3191 		/*
3192 		 * We are here because the timer fired and we
3193 		 * have some quota to tranmit.
3194 		 */
3195 		prev = NULL;
3196 		head = tail = mac_srs->srs_first;
3197 		while (mac_srs->srs_first != NULL) {
3198 			tail = mac_srs->srs_first;
3199 			mac_srs->srs_first = tail->b_next;
3200 			if (mac_srs->srs_first == NULL)
3201 				mac_srs->srs_last = NULL;
3202 			mac_srs->srs_count--;
3203 			sz = msgdsize(tail);
3204 			mac_srs->srs_size -= sz;
3205 			mac_srs->srs_bw->mac_bw_used += sz;
3206 			if (prev == NULL)
3207 				hint = (ulong_t)tail->b_prev;
3208 			if (hint != (ulong_t)tail->b_prev) {
3209 				prev->b_next = NULL;
3210 				mutex_exit(&mac_srs->srs_lock);
3211 				TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3212 				head = tail;
3213 				hint = (ulong_t)tail->b_prev;
3214 				mutex_enter(&mac_srs->srs_lock);
3215 			}
3216 
3217 			prev = tail;
3218 			tail->b_prev = NULL;
3219 			if (mac_srs->srs_bw->mac_bw_used <
3220 			    mac_srs->srs_bw->mac_bw_limit)
3221 				continue;
3222 
3223 			now = ddi_get_lbolt();
3224 			if (mac_srs->srs_bw->mac_bw_curr_time != now) {
3225 				mac_srs->srs_bw->mac_bw_curr_time = now;
3226 				mac_srs->srs_bw->mac_bw_used = 0;
3227 				continue;
3228 			}
3229 			mac_srs->srs_bw->mac_bw_state |= SRS_BW_ENFORCED;
3230 			break;
3231 		}
3232 		ASSERT((head == NULL && tail == NULL) ||
3233 		    (head != NULL && tail != NULL));
3234 		if (tail != NULL) {
3235 			tail->b_next = NULL;
3236 			mutex_exit(&mac_srs->srs_lock);
3237 			TX_SRS_TO_SOFT_RING(mac_srs, head, hint);
3238 			mutex_enter(&mac_srs->srs_lock);
3239 		}
3240 	}
3241 	/*
3242 	 * SRS_TX_FANOUT case not considered here because packets
3243 	 * won't be queued in the SRS for this case. Packets will
3244 	 * be sent directly to soft rings underneath and if there
3245 	 * is any queueing at all, it would be in Tx side soft
3246 	 * rings.
3247 	 */
3248 
3249 	/*
3250 	 * When srs_count becomes 0, reset SRS_TX_HIWAT and
3251 	 * SRS_TX_WAKEUP_CLIENT and wakeup registered clients.
3252 	 */
3253 	if (mac_srs->srs_count == 0 && (mac_srs->srs_state &
3254 	    (SRS_TX_HIWAT | SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED))) {
3255 		mac_client_impl_t *mcip = mac_srs->srs_mcip;
3256 		boolean_t wakeup_required = B_FALSE;
3257 
3258 		if (mac_srs->srs_state &
3259 		    (SRS_TX_HIWAT|SRS_TX_WAKEUP_CLIENT)) {
3260 			wakeup_required = B_TRUE;
3261 		}
3262 		mac_srs->srs_state &= ~(SRS_TX_HIWAT |
3263 		    SRS_TX_WAKEUP_CLIENT | SRS_ENQUEUED);
3264 		mutex_exit(&mac_srs->srs_lock);
3265 		if (wakeup_required) {
3266 			mac_tx_invoke_callbacks(mcip, (mac_tx_cookie_t)mac_srs);
3267 			/*
3268 			 * If the client is not the primary MAC client, then we
3269 			 * need to send the notification to the clients upper
3270 			 * MAC, i.e. mci_upper_mip.
3271 			 */
3272 			mac_tx_notify(mcip->mci_upper_mip != NULL ?
3273 			    mcip->mci_upper_mip : mcip->mci_mip);
3274 		}
3275 		mutex_enter(&mac_srs->srs_lock);
3276 	}
3277 	mac_srs->srs_state &= ~SRS_PROC;
3278 }
3279 
3280 /*
3281  * Given a packet, get the flow_entry that identifies the flow
3282  * to which that packet belongs. The flow_entry will contain
3283  * the transmit function to be used to send the packet. If the
3284  * function returns NULL, the packet should be sent using the
3285  * underlying NIC.
3286  */
3287 static flow_entry_t *
3288 mac_tx_classify(mac_impl_t *mip, mblk_t *mp)
3289 {
3290 	flow_entry_t		*flent = NULL;
3291 	mac_client_impl_t	*mcip;
3292 	int	err;
3293 
3294 	/*
3295 	 * Do classification on the packet.
3296 	 */
3297 	err = mac_flow_lookup(mip->mi_flow_tab, mp, FLOW_OUTBOUND, &flent);
3298 	if (err != 0)
3299 		return (NULL);
3300 
3301 	/*
3302 	 * This flent might just be an additional one on the MAC client,
3303 	 * i.e. for classification purposes (different fdesc), however
3304 	 * the resources, SRS et. al., are in the mci_flent, so if
3305 	 * this isn't the mci_flent, we need to get it.
3306 	 */
3307 	if ((mcip = flent->fe_mcip) != NULL && mcip->mci_flent != flent) {
3308 		FLOW_REFRELE(flent);
3309 		flent = mcip->mci_flent;
3310 		FLOW_TRY_REFHOLD(flent, err);
3311 		if (err != 0)
3312 			return (NULL);
3313 	}
3314 
3315 	return (flent);
3316 }
3317 
3318 /*
3319  * This macro is only meant to be used by mac_tx_send().
3320  */
3321 #define	CHECK_VID_AND_ADD_TAG(mp) {			\
3322 	if (vid_check) {				\
3323 		int err = 0;				\
3324 							\
3325 		MAC_VID_CHECK(src_mcip, (mp), err);	\
3326 		if (err != 0) {				\
3327 			freemsg((mp));			\
3328 			(mp) = next;			\
3329 			oerrors++;			\
3330 			continue;			\
3331 		}					\
3332 	}						\
3333 	if (add_tag) {					\
3334 		(mp) = mac_add_vlan_tag((mp), 0, vid);	\
3335 		if ((mp) == NULL) {			\
3336 			(mp) = next;			\
3337 			oerrors++;			\
3338 			continue;			\
3339 		}					\
3340 	}						\
3341 }
3342 
3343 mblk_t *
3344 mac_tx_send(mac_client_handle_t mch, mac_ring_handle_t ring, mblk_t *mp_chain,
3345     mac_tx_stats_t *stats)
3346 {
3347 	mac_client_impl_t *src_mcip = (mac_client_impl_t *)mch;
3348 	mac_impl_t *mip = src_mcip->mci_mip;
3349 	uint_t obytes = 0, opackets = 0, oerrors = 0;
3350 	mblk_t *mp = NULL, *next;
3351 	boolean_t vid_check, add_tag;
3352 	uint16_t vid = 0;
3353 
3354 	if (mip->mi_nclients > 1) {
3355 		vid_check = MAC_VID_CHECK_NEEDED(src_mcip);
3356 		add_tag = MAC_TAG_NEEDED(src_mcip);
3357 		if (add_tag)
3358 			vid = mac_client_vid(mch);
3359 	} else {
3360 		ASSERT(mip->mi_nclients == 1);
3361 		vid_check = add_tag = B_FALSE;
3362 	}
3363 
3364 	/*
3365 	 * Fastpath: if there's only one client, we simply send
3366 	 * the packet down to the underlying NIC.
3367 	 */
3368 	if (mip->mi_nactiveclients == 1) {
3369 		DTRACE_PROBE2(fastpath,
3370 		    mac_client_impl_t *, src_mcip, mblk_t *, mp_chain);
3371 
3372 		mp = mp_chain;
3373 		while (mp != NULL) {
3374 			next = mp->b_next;
3375 			mp->b_next = NULL;
3376 			opackets++;
3377 			obytes += (mp->b_cont == NULL ? MBLKL(mp) :
3378 			    msgdsize(mp));
3379 
3380 			CHECK_VID_AND_ADD_TAG(mp);
3381 			MAC_TX(mip, ring, mp, src_mcip);
3382 
3383 			/*
3384 			 * If the driver is out of descriptors and does a
3385 			 * partial send it will return a chain of unsent
3386 			 * mblks. Adjust the accounting stats.
3387 			 */
3388 			if (mp != NULL) {
3389 				opackets--;
3390 				obytes -= msgdsize(mp);
3391 				mp->b_next = next;
3392 				break;
3393 			}
3394 			mp = next;
3395 		}
3396 		goto done;
3397 	}
3398 
3399 	/*
3400 	 * No fastpath, we either have more than one MAC client
3401 	 * defined on top of the same MAC, or one or more MAC
3402 	 * client promiscuous callbacks.
3403 	 */
3404 	DTRACE_PROBE3(slowpath, mac_client_impl_t *,
3405 	    src_mcip, int, mip->mi_nclients, mblk_t *, mp_chain);
3406 
3407 	mp = mp_chain;
3408 	while (mp != NULL) {
3409 		flow_entry_t *dst_flow_ent;
3410 		void *flow_cookie;
3411 		size_t	pkt_size;
3412 		mblk_t *mp1;
3413 
3414 		next = mp->b_next;
3415 		mp->b_next = NULL;
3416 		opackets++;
3417 		pkt_size = (mp->b_cont == NULL ? MBLKL(mp) : msgdsize(mp));
3418 		obytes += pkt_size;
3419 		CHECK_VID_AND_ADD_TAG(mp);
3420 
3421 		/*
3422 		 * Find the destination.
3423 		 */
3424 		dst_flow_ent = mac_tx_classify(mip, mp);
3425 
3426 		if (dst_flow_ent != NULL) {
3427 			size_t	hdrsize;
3428 			int	err = 0;
3429 
3430 			if (mip->mi_info.mi_nativemedia == DL_ETHER) {
3431 				struct ether_vlan_header *evhp =
3432 				    (struct ether_vlan_header *)mp->b_rptr;
3433 
3434 				if (ntohs(evhp->ether_tpid) == ETHERTYPE_VLAN)
3435 					hdrsize = sizeof (*evhp);
3436 				else
3437 					hdrsize = sizeof (struct ether_header);
3438 			} else {
3439 				mac_header_info_t	mhi;
3440 
3441 				err = mac_header_info((mac_handle_t)mip,
3442 				    mp, &mhi);
3443 				if (err == 0)
3444 					hdrsize = mhi.mhi_hdrsize;
3445 			}
3446 
3447 			/*
3448 			 * Got a matching flow. It's either another
3449 			 * MAC client, or a broadcast/multicast flow.
3450 			 * Make sure the packet size is within the
3451 			 * allowed size. If not drop the packet and
3452 			 * move to next packet.
3453 			 */
3454 			if (err != 0 ||
3455 			    (pkt_size - hdrsize) > mip->mi_sdu_max) {
3456 				oerrors++;
3457 				DTRACE_PROBE2(loopback__drop, size_t, pkt_size,
3458 				    mblk_t *, mp);
3459 				freemsg(mp);
3460 				mp = next;
3461 				FLOW_REFRELE(dst_flow_ent);
3462 				continue;
3463 			}
3464 			flow_cookie = mac_flow_get_client_cookie(dst_flow_ent);
3465 			if (flow_cookie != NULL) {
3466 				/*
3467 				 * The vnic_bcast_send function expects
3468 				 * to receive the sender MAC client
3469 				 * as value for arg2.
3470 				 */
3471 				mac_bcast_send(flow_cookie, src_mcip, mp,
3472 				    B_TRUE);
3473 			} else {
3474 				/*
3475 				 * loopback the packet to a local MAC
3476 				 * client. We force a context switch
3477 				 * if both source and destination MAC
3478 				 * clients are used by IP, i.e.
3479 				 * bypass is set.
3480 				 */
3481 				boolean_t do_switch;
3482 				mac_client_impl_t *dst_mcip =
3483 				    dst_flow_ent->fe_mcip;
3484 
3485 				/*
3486 				 * Check if there are promiscuous mode
3487 				 * callbacks defined. This check is
3488 				 * done here in the 'else' case and
3489 				 * not in other cases because this
3490 				 * path is for local loopback
3491 				 * communication which does not go
3492 				 * through MAC_TX(). For paths that go
3493 				 * through MAC_TX(), the promisc_list
3494 				 * check is done inside the MAC_TX()
3495 				 * macro.
3496 				 */
3497 				if (mip->mi_promisc_list != NULL)
3498 					mac_promisc_dispatch(mip, mp, src_mcip);
3499 
3500 				do_switch = ((src_mcip->mci_state_flags &
3501 				    dst_mcip->mci_state_flags &
3502 				    MCIS_CLIENT_POLL_CAPABLE) != 0);
3503 
3504 				if ((mp1 = mac_fix_cksum(mp)) != NULL) {
3505 					(dst_flow_ent->fe_cb_fn)(
3506 					    dst_flow_ent->fe_cb_arg1,
3507 					    dst_flow_ent->fe_cb_arg2,
3508 					    mp1, do_switch);
3509 				}
3510 			}
3511 			FLOW_REFRELE(dst_flow_ent);
3512 		} else {
3513 			/*
3514 			 * Unknown destination, send via the underlying
3515 			 * NIC.
3516 			 */
3517 			MAC_TX(mip, ring, mp, src_mcip);
3518 			if (mp != NULL) {
3519 				/*
3520 				 * Adjust for the last packet that
3521 				 * could not be transmitted
3522 				 */
3523 				opackets--;
3524 				obytes -= pkt_size;
3525 				mp->b_next = next;
3526 				break;
3527 			}
3528 		}
3529 		mp = next;
3530 	}
3531 
3532 done:
3533 	stats->mts_obytes = obytes;
3534 	stats->mts_opackets = opackets;
3535 	stats->mts_oerrors = oerrors;
3536 	return (mp);
3537 }
3538 
3539 /*
3540  * mac_tx_srs_ring_present
3541  *
3542  * Returns whether the specified ring is part of the specified SRS.
3543  */
3544 boolean_t
3545 mac_tx_srs_ring_present(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3546 {
3547 	int i;
3548 	mac_soft_ring_t *soft_ring;
3549 
3550 	if (srs->srs_tx.st_arg2 == tx_ring)
3551 		return (B_TRUE);
3552 
3553 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
3554 		soft_ring =  srs->srs_tx_soft_rings[i];
3555 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3556 			return (B_TRUE);
3557 	}
3558 
3559 	return (B_FALSE);
3560 }
3561 
3562 /*
3563  * mac_tx_srs_get_soft_ring
3564  *
3565  * Returns the TX soft ring associated with the given ring, if present.
3566  */
3567 mac_soft_ring_t *
3568 mac_tx_srs_get_soft_ring(mac_soft_ring_set_t *srs, mac_ring_t *tx_ring)
3569 {
3570 	int		i;
3571 	mac_soft_ring_t	*soft_ring;
3572 
3573 	if (srs->srs_tx.st_arg2 == tx_ring)
3574 		return (NULL);
3575 
3576 	for (i = 0; i < srs->srs_tx_ring_count; i++) {
3577 		soft_ring =  srs->srs_tx_soft_rings[i];
3578 		if (soft_ring->s_ring_tx_arg2 == tx_ring)
3579 			return (soft_ring);
3580 	}
3581 
3582 	return (NULL);
3583 }
3584 
3585 /*
3586  * mac_tx_srs_wakeup
3587  *
3588  * Called when Tx desc become available. Wakeup the appropriate worker
3589  * thread after resetting the SRS_TX_BLOCKED/S_RING_BLOCK bit in the
3590  * state field.
3591  */
3592 void
3593 mac_tx_srs_wakeup(mac_soft_ring_set_t *mac_srs, mac_ring_handle_t ring)
3594 {
3595 	int i;
3596 	mac_soft_ring_t *sringp;
3597 	mac_srs_tx_t *srs_tx = &mac_srs->srs_tx;
3598 
3599 	mutex_enter(&mac_srs->srs_lock);
3600 	/*
3601 	 * srs_tx_ring_count == 0 is the single ring mode case. In
3602 	 * this mode, there will not be Tx soft rings associated
3603 	 * with the SRS.
3604 	 */
3605 	if (!MAC_TX_SOFT_RINGS(mac_srs)) {
3606 		if (srs_tx->st_arg2 == ring &&
3607 		    mac_srs->srs_state & SRS_TX_BLOCKED) {
3608 			mac_srs->srs_state &= ~SRS_TX_BLOCKED;
3609 			srs_tx->st_stat.mts_unblockcnt++;
3610 			cv_signal(&mac_srs->srs_async);
3611 		}
3612 		/*
3613 		 * A wakeup can come before tx_srs_drain() could
3614 		 * grab srs lock and set SRS_TX_BLOCKED. So
3615 		 * always set woken_up flag when we come here.
3616 		 */
3617 		srs_tx->st_woken_up = B_TRUE;
3618 		mutex_exit(&mac_srs->srs_lock);
3619 		return;
3620 	}
3621 
3622 	/*
3623 	 * If you are here, it is for FANOUT, BW_FANOUT,
3624 	 * AGGR_MODE or AGGR_BW_MODE case
3625 	 */
3626 	for (i = 0; i < mac_srs->srs_tx_ring_count; i++) {
3627 		sringp = mac_srs->srs_tx_soft_rings[i];
3628 		mutex_enter(&sringp->s_ring_lock);
3629 		if (sringp->s_ring_tx_arg2 == ring) {
3630 			if (sringp->s_ring_state & S_RING_BLOCK) {
3631 				sringp->s_ring_state &= ~S_RING_BLOCK;
3632 				sringp->s_st_stat.mts_unblockcnt++;
3633 				cv_signal(&sringp->s_ring_async);
3634 			}
3635 			sringp->s_ring_tx_woken_up = B_TRUE;
3636 		}
3637 		mutex_exit(&sringp->s_ring_lock);
3638 	}
3639 	mutex_exit(&mac_srs->srs_lock);
3640 }
3641 
3642 /*
3643  * Once the driver is done draining, send a MAC_NOTE_TX notification to unleash
3644  * the blocked clients again.
3645  */
3646 void
3647 mac_tx_notify(mac_impl_t *mip)
3648 {
3649 	i_mac_notify(mip, MAC_NOTE_TX);
3650 }
3651 
3652 /*
3653  * RX SOFTRING RELATED FUNCTIONS
3654  *
3655  * These functions really belong in mac_soft_ring.c and here for
3656  * a short period.
3657  */
3658 
3659 #define	SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3660 	/*								\
3661 	 * Enqueue our mblk chain.					\
3662 	 */								\
3663 	ASSERT(MUTEX_HELD(&(ringp)->s_ring_lock));			\
3664 									\
3665 	if ((ringp)->s_ring_last != NULL)				\
3666 		(ringp)->s_ring_last->b_next = (mp);			\
3667 	else								\
3668 		(ringp)->s_ring_first = (mp);				\
3669 	(ringp)->s_ring_last = (tail);					\
3670 	(ringp)->s_ring_count += (cnt);					\
3671 	ASSERT((ringp)->s_ring_count > 0);				\
3672 	if ((ringp)->s_ring_type & ST_RING_BW_CTL) {			\
3673 		(ringp)->s_ring_size += sz;				\
3674 	}								\
3675 }
3676 
3677 /*
3678  * Default entry point to deliver a packet chain to a MAC client.
3679  * If the MAC client has flows, do the classification with these
3680  * flows as well.
3681  */
3682 /* ARGSUSED */
3683 void
3684 mac_rx_deliver(void *arg1, mac_resource_handle_t mrh, mblk_t *mp_chain,
3685     mac_header_info_t *arg3)
3686 {
3687 	mac_client_impl_t *mcip = arg1;
3688 
3689 	if (mcip->mci_nvids == 1 &&
3690 	    !(mcip->mci_state_flags & MCIS_STRIP_DISABLE)) {
3691 		/*
3692 		 * If the client has exactly one VID associated with it
3693 		 * and striping of VLAN header is not disabled,
3694 		 * remove the VLAN tag from the packet before
3695 		 * passing it on to the client's receive callback.
3696 		 * Note that this needs to be done after we dispatch
3697 		 * the packet to the promiscuous listeners of the
3698 		 * client, since they expect to see the whole
3699 		 * frame including the VLAN headers.
3700 		 */
3701 		mp_chain = mac_strip_vlan_tag_chain(mp_chain);
3702 	}
3703 
3704 	mcip->mci_rx_fn(mcip->mci_rx_arg, mrh, mp_chain, B_FALSE);
3705 }
3706 
3707 /*
3708  * mac_rx_soft_ring_process
3709  *
3710  * process a chain for a given soft ring. The number of packets queued
3711  * in the SRS and its associated soft rings (including this one) is
3712  * very small (tracked by srs_poll_pkt_cnt), then allow the entering
3713  * thread (interrupt or poll thread) to do inline processing. This
3714  * helps keep the latency down under low load.
3715  *
3716  * The proc and arg for each mblk is already stored in the mblk in
3717  * appropriate places.
3718  */
3719 /* ARGSUSED */
3720 void
3721 mac_rx_soft_ring_process(mac_client_impl_t *mcip, mac_soft_ring_t *ringp,
3722     mblk_t *mp_chain, mblk_t *tail, int cnt, size_t sz)
3723 {
3724 	mac_direct_rx_t		proc;
3725 	void			*arg1;
3726 	mac_resource_handle_t	arg2;
3727 	mac_soft_ring_set_t	*mac_srs = ringp->s_ring_set;
3728 
3729 	ASSERT(ringp != NULL);
3730 	ASSERT(mp_chain != NULL);
3731 	ASSERT(tail != NULL);
3732 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3733 
3734 	mutex_enter(&ringp->s_ring_lock);
3735 	ringp->s_ring_total_inpkt += cnt;
3736 	ringp->s_ring_total_rbytes += sz;
3737 	if ((mac_srs->srs_rx.sr_poll_pkt_cnt <= 1) &&
3738 	    !(ringp->s_ring_type & ST_RING_WORKER_ONLY)) {
3739 		/* If on processor or blanking on, then enqueue and return */
3740 		if (ringp->s_ring_state & S_RING_BLANK ||
3741 		    ringp->s_ring_state & S_RING_PROC) {
3742 			SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3743 			mutex_exit(&ringp->s_ring_lock);
3744 			return;
3745 		}
3746 		proc = ringp->s_ring_rx_func;
3747 		arg1 = ringp->s_ring_rx_arg1;
3748 		arg2 = ringp->s_ring_rx_arg2;
3749 		/*
3750 		 * See if anything is already queued. If we are the
3751 		 * first packet, do inline processing else queue the
3752 		 * packet and do the drain.
3753 		 */
3754 		if (ringp->s_ring_first == NULL) {
3755 			/*
3756 			 * Fast-path, ok to process and nothing queued.
3757 			 */
3758 			ringp->s_ring_run = curthread;
3759 			ringp->s_ring_state |= (S_RING_PROC);
3760 
3761 			mutex_exit(&ringp->s_ring_lock);
3762 
3763 			/*
3764 			 * We are the chain of 1 packet so
3765 			 * go through this fast path.
3766 			 */
3767 			ASSERT(mp_chain->b_next == NULL);
3768 
3769 			(*proc)(arg1, arg2, mp_chain, NULL);
3770 
3771 			ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3772 			/*
3773 			 * If we have a soft ring set which is doing
3774 			 * bandwidth control, we need to decrement
3775 			 * srs_size and count so it the SRS can have a
3776 			 * accurate idea of what is the real data
3777 			 * queued between SRS and its soft rings. We
3778 			 * decrement the counters only when the packet
3779 			 * gets processed by both SRS and the soft ring.
3780 			 */
3781 			mutex_enter(&mac_srs->srs_lock);
3782 			MAC_UPDATE_SRS_COUNT_LOCKED(mac_srs, cnt);
3783 			MAC_UPDATE_SRS_SIZE_LOCKED(mac_srs, sz);
3784 			mutex_exit(&mac_srs->srs_lock);
3785 
3786 			mutex_enter(&ringp->s_ring_lock);
3787 			ringp->s_ring_run = NULL;
3788 			ringp->s_ring_state &= ~S_RING_PROC;
3789 			if (ringp->s_ring_state & S_RING_CLIENT_WAIT)
3790 				cv_signal(&ringp->s_ring_client_cv);
3791 
3792 			if ((ringp->s_ring_first == NULL) ||
3793 			    (ringp->s_ring_state & S_RING_BLANK)) {
3794 				/*
3795 				 * We processed inline our packet and
3796 				 * nothing new has arrived or our
3797 				 * receiver doesn't want to receive
3798 				 * any packets. We are done.
3799 				 */
3800 				mutex_exit(&ringp->s_ring_lock);
3801 				return;
3802 			}
3803 		} else {
3804 			SOFT_RING_ENQUEUE_CHAIN(ringp,
3805 			    mp_chain, tail, cnt, sz);
3806 		}
3807 
3808 		/*
3809 		 * We are here because either we couldn't do inline
3810 		 * processing (because something was already
3811 		 * queued), or we had a chain of more than one
3812 		 * packet, or something else arrived after we were
3813 		 * done with inline processing.
3814 		 */
3815 		ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3816 		ASSERT(ringp->s_ring_first != NULL);
3817 
3818 		ringp->s_ring_drain_func(ringp);
3819 		mutex_exit(&ringp->s_ring_lock);
3820 		return;
3821 	} else {
3822 		/* ST_RING_WORKER_ONLY case */
3823 		SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3824 		mac_soft_ring_worker_wakeup(ringp);
3825 		mutex_exit(&ringp->s_ring_lock);
3826 	}
3827 }
3828 
3829 /*
3830  * TX SOFTRING RELATED FUNCTIONS
3831  *
3832  * These functions really belong in mac_soft_ring.c and here for
3833  * a short period.
3834  */
3835 
3836 #define	TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp, tail, cnt, sz) {	       	\
3837 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));			\
3838 	ringp->s_ring_state |= S_RING_ENQUEUED;				\
3839 	SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);	\
3840 }
3841 
3842 /*
3843  * mac_tx_sring_queued
3844  *
3845  * When we are out of transmit descriptors and we already have a
3846  * queue that exceeds hiwat (or the client called us with
3847  * MAC_TX_NO_ENQUEUE or MAC_DROP_ON_NO_DESC flag), return the
3848  * soft ring pointer as the opaque cookie for the client enable
3849  * flow control.
3850  */
3851 static mac_tx_cookie_t
3852 mac_tx_sring_enqueue(mac_soft_ring_t *ringp, mblk_t *mp_chain, uint16_t flag,
3853     mblk_t **ret_mp)
3854 {
3855 	int cnt;
3856 	size_t sz;
3857 	mblk_t *tail;
3858 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3859 	mac_tx_cookie_t cookie = NULL;
3860 	boolean_t wakeup_worker = B_TRUE;
3861 
3862 	ASSERT(MUTEX_HELD(&ringp->s_ring_lock));
3863 	MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3864 	if (flag & MAC_DROP_ON_NO_DESC) {
3865 		mac_pkt_drop(NULL, NULL, mp_chain, B_FALSE);
3866 		/* increment freed stats */
3867 		ringp->s_ring_drops += cnt;
3868 		cookie = (mac_tx_cookie_t)ringp;
3869 	} else {
3870 		if (ringp->s_ring_first != NULL)
3871 			wakeup_worker = B_FALSE;
3872 
3873 		if (flag & MAC_TX_NO_ENQUEUE) {
3874 			/*
3875 			 * If QUEUED is not set, queue the packet
3876 			 * and let mac_tx_soft_ring_drain() set
3877 			 * the TX_BLOCKED bit for the reasons
3878 			 * explained above. Otherwise, return the
3879 			 * mblks.
3880 			 */
3881 			if (wakeup_worker) {
3882 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp,
3883 				    mp_chain, tail, cnt, sz);
3884 			} else {
3885 				ringp->s_ring_state |= S_RING_WAKEUP_CLIENT;
3886 				cookie = (mac_tx_cookie_t)ringp;
3887 				*ret_mp = mp_chain;
3888 			}
3889 		} else {
3890 			boolean_t enqueue = B_TRUE;
3891 
3892 			if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3893 				/*
3894 				 * flow-controlled. Store ringp in cookie
3895 				 * so that it can be returned as
3896 				 * mac_tx_cookie_t to client
3897 				 */
3898 				ringp->s_ring_state |= S_RING_TX_HIWAT;
3899 				cookie = (mac_tx_cookie_t)ringp;
3900 				ringp->s_ring_hiwat_cnt++;
3901 				if (ringp->s_ring_count >
3902 				    ringp->s_ring_tx_max_q_cnt) {
3903 					/* increment freed stats */
3904 					ringp->s_ring_drops += cnt;
3905 					/*
3906 					 * b_prev may be set to the fanout hint
3907 					 * hence can't use freemsg directly
3908 					 */
3909 					mac_pkt_drop(NULL, NULL,
3910 					    mp_chain, B_FALSE);
3911 					DTRACE_PROBE1(tx_queued_hiwat,
3912 					    mac_soft_ring_t *, ringp);
3913 					enqueue = B_FALSE;
3914 				}
3915 			}
3916 			if (enqueue) {
3917 				TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain,
3918 				    tail, cnt, sz);
3919 			}
3920 		}
3921 		if (wakeup_worker)
3922 			cv_signal(&ringp->s_ring_async);
3923 	}
3924 	return (cookie);
3925 }
3926 
3927 
3928 /*
3929  * mac_tx_soft_ring_process
3930  *
3931  * This routine is called when fanning out outgoing traffic among
3932  * multipe Tx rings.
3933  * Note that a soft ring is associated with a h/w Tx ring.
3934  */
3935 mac_tx_cookie_t
3936 mac_tx_soft_ring_process(mac_soft_ring_t *ringp, mblk_t *mp_chain,
3937     uint16_t flag, mblk_t **ret_mp)
3938 {
3939 	mac_soft_ring_set_t *mac_srs = ringp->s_ring_set;
3940 	int	cnt;
3941 	size_t	sz;
3942 	mblk_t	*tail;
3943 	mac_tx_cookie_t cookie = NULL;
3944 
3945 	ASSERT(ringp != NULL);
3946 	ASSERT(mp_chain != NULL);
3947 	ASSERT(MUTEX_NOT_HELD(&ringp->s_ring_lock));
3948 	/*
3949 	 * The following modes can come here: SRS_TX_BW_FANOUT,
3950 	 * SRS_TX_FANOUT, SRS_TX_AGGR, SRS_TX_BW_AGGR.
3951 	 */
3952 	ASSERT(MAC_TX_SOFT_RINGS(mac_srs));
3953 	ASSERT(mac_srs->srs_tx.st_mode == SRS_TX_FANOUT ||
3954 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_FANOUT ||
3955 	    mac_srs->srs_tx.st_mode == SRS_TX_AGGR ||
3956 	    mac_srs->srs_tx.st_mode == SRS_TX_BW_AGGR);
3957 
3958 	if (ringp->s_ring_type & ST_RING_WORKER_ONLY) {
3959 		/* Serialization mode */
3960 
3961 		mutex_enter(&ringp->s_ring_lock);
3962 		if (ringp->s_ring_count > ringp->s_ring_tx_hiwat) {
3963 			cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3964 			    flag, ret_mp);
3965 			mutex_exit(&ringp->s_ring_lock);
3966 			return (cookie);
3967 		}
3968 		MAC_COUNT_CHAIN(mac_srs, mp_chain, tail, cnt, sz);
3969 		TX_SOFT_RING_ENQUEUE_CHAIN(ringp, mp_chain, tail, cnt, sz);
3970 		if (ringp->s_ring_state & (S_RING_BLOCK | S_RING_PROC)) {
3971 			/*
3972 			 * If ring is blocked due to lack of Tx
3973 			 * descs, just return. Worker thread
3974 			 * will get scheduled when Tx desc's
3975 			 * become available.
3976 			 */
3977 			mutex_exit(&ringp->s_ring_lock);
3978 			return (cookie);
3979 		}
3980 		mac_soft_ring_worker_wakeup(ringp);
3981 		mutex_exit(&ringp->s_ring_lock);
3982 		return (cookie);
3983 	} else {
3984 		/* Default fanout mode */
3985 		/*
3986 		 * S_RING_BLOCKED is set when underlying NIC runs
3987 		 * out of Tx descs and messages start getting
3988 		 * queued. It won't get reset until
3989 		 * tx_srs_drain() completely drains out the
3990 		 * messages.
3991 		 */
3992 		mac_tx_stats_t		stats;
3993 
3994 		if (ringp->s_ring_state & S_RING_ENQUEUED) {
3995 			/* Tx descs/resources not available */
3996 			mutex_enter(&ringp->s_ring_lock);
3997 			if (ringp->s_ring_state & S_RING_ENQUEUED) {
3998 				cookie = mac_tx_sring_enqueue(ringp, mp_chain,
3999 				    flag, ret_mp);
4000 				mutex_exit(&ringp->s_ring_lock);
4001 				return (cookie);
4002 			}
4003 			/*
4004 			 * While we were computing mblk count, the
4005 			 * flow control condition got relieved.
4006 			 * Continue with the transmission.
4007 			 */
4008 			mutex_exit(&ringp->s_ring_lock);
4009 		}
4010 
4011 		mp_chain = mac_tx_send(ringp->s_ring_tx_arg1,
4012 		    ringp->s_ring_tx_arg2, mp_chain, &stats);
4013 
4014 		/*
4015 		 * Multiple threads could be here sending packets.
4016 		 * Under such conditions, it is not possible to
4017 		 * automically set S_RING_BLOCKED bit to indicate
4018 		 * out of tx desc condition. To atomically set
4019 		 * this, we queue the returned packet and do
4020 		 * the setting of S_RING_BLOCKED in
4021 		 * mac_tx_soft_ring_drain().
4022 		 */
4023 		if (mp_chain != NULL) {
4024 			mutex_enter(&ringp->s_ring_lock);
4025 			cookie =
4026 			    mac_tx_sring_enqueue(ringp, mp_chain, flag, ret_mp);
4027 			mutex_exit(&ringp->s_ring_lock);
4028 			return (cookie);
4029 		}
4030 		SRS_TX_STATS_UPDATE(mac_srs, &stats);
4031 		SOFTRING_TX_STATS_UPDATE(ringp, &stats);
4032 
4033 		return (NULL);
4034 	}
4035 }
4036