xref: /illumos-gate/usr/src/uts/common/os/strsubr.c (revision 56ffda179176d14e2fb160bdc85f6da601f1eac7)
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 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
22 /*	  All Rights Reserved  	*/
23 
24 
25 /*
26  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  * Copyright (c) 2016 by Delphix. All rights reserved.
29  * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
30  * Copyright 2018 Joyent, Inc.
31  */
32 
33 #include <sys/types.h>
34 #include <sys/sysmacros.h>
35 #include <sys/param.h>
36 #include <sys/errno.h>
37 #include <sys/signal.h>
38 #include <sys/proc.h>
39 #include <sys/conf.h>
40 #include <sys/cred.h>
41 #include <sys/user.h>
42 #include <sys/vnode.h>
43 #include <sys/file.h>
44 #include <sys/session.h>
45 #include <sys/stream.h>
46 #include <sys/strsubr.h>
47 #include <sys/stropts.h>
48 #include <sys/poll.h>
49 #include <sys/systm.h>
50 #include <sys/cpuvar.h>
51 #include <sys/uio.h>
52 #include <sys/cmn_err.h>
53 #include <sys/priocntl.h>
54 #include <sys/procset.h>
55 #include <sys/vmem.h>
56 #include <sys/bitmap.h>
57 #include <sys/kmem.h>
58 #include <sys/siginfo.h>
59 #include <sys/vtrace.h>
60 #include <sys/callb.h>
61 #include <sys/debug.h>
62 #include <sys/modctl.h>
63 #include <sys/vmsystm.h>
64 #include <vm/page.h>
65 #include <sys/atomic.h>
66 #include <sys/suntpi.h>
67 #include <sys/strlog.h>
68 #include <sys/promif.h>
69 #include <sys/project.h>
70 #include <sys/vm.h>
71 #include <sys/taskq.h>
72 #include <sys/sunddi.h>
73 #include <sys/sunldi_impl.h>
74 #include <sys/strsun.h>
75 #include <sys/isa_defs.h>
76 #include <sys/multidata.h>
77 #include <sys/pattr.h>
78 #include <sys/strft.h>
79 #include <sys/fs/snode.h>
80 #include <sys/zone.h>
81 #include <sys/open.h>
82 #include <sys/sunldi.h>
83 #include <sys/sad.h>
84 #include <sys/netstack.h>
85 
86 #define	O_SAMESTR(q)	(((q)->q_next) && \
87 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
88 
89 /*
90  * WARNING:
91  * The variables and routines in this file are private, belonging
92  * to the STREAMS subsystem. These should not be used by modules
93  * or drivers. Compatibility will not be guaranteed.
94  */
95 
96 /*
97  * Id value used to distinguish between different multiplexor links.
98  */
99 static int32_t lnk_id = 0;
100 
101 #define	STREAMS_LOPRI MINCLSYSPRI
102 static pri_t streams_lopri = STREAMS_LOPRI;
103 
104 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
105 typedef struct str_stat {
106 	kstat_named_t	sqenables;
107 	kstat_named_t	stenables;
108 	kstat_named_t	syncqservice;
109 	kstat_named_t	freebs;
110 	kstat_named_t	qwr_outer;
111 	kstat_named_t	rservice;
112 	kstat_named_t	strwaits;
113 	kstat_named_t	taskqfails;
114 	kstat_named_t	bufcalls;
115 	kstat_named_t	qhelps;
116 	kstat_named_t	qremoved;
117 	kstat_named_t	sqremoved;
118 	kstat_named_t	bcwaits;
119 	kstat_named_t	sqtoomany;
120 } str_stat_t;
121 
122 static str_stat_t str_statistics = {
123 	{ "sqenables",		KSTAT_DATA_UINT64 },
124 	{ "stenables",		KSTAT_DATA_UINT64 },
125 	{ "syncqservice",	KSTAT_DATA_UINT64 },
126 	{ "freebs",		KSTAT_DATA_UINT64 },
127 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
128 	{ "rservice",		KSTAT_DATA_UINT64 },
129 	{ "strwaits",		KSTAT_DATA_UINT64 },
130 	{ "taskqfails",		KSTAT_DATA_UINT64 },
131 	{ "bufcalls",		KSTAT_DATA_UINT64 },
132 	{ "qhelps",		KSTAT_DATA_UINT64 },
133 	{ "qremoved",		KSTAT_DATA_UINT64 },
134 	{ "sqremoved",		KSTAT_DATA_UINT64 },
135 	{ "bcwaits",		KSTAT_DATA_UINT64 },
136 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
137 };
138 
139 static kstat_t *str_kstat;
140 
141 /*
142  * qrunflag was used previously to control background scheduling of queues. It
143  * is not used anymore, but kept here in case some module still wants to access
144  * it via qready() and setqsched macros.
145  */
146 char qrunflag;			/*  Unused */
147 
148 /*
149  * Most of the streams scheduling is done via task queues. Task queues may fail
150  * for non-sleep dispatches, so there are two backup threads servicing failed
151  * requests for queues and syncqs. Both of these threads also service failed
152  * dispatches freebs requests. Queues are put in the list specified by `qhead'
153  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
154  * requests are put into `freebs_list' which has no tail pointer. All three
155  * lists are protected by a single `service_queue' lock and use
156  * `services_to_run' condition variable for signaling background threads. Use of
157  * a single lock should not be a problem because it is only used under heavy
158  * loads when task queues start to fail and at that time it may be a good idea
159  * to throttle scheduling requests.
160  *
161  * NOTE: queues and syncqs should be scheduled by two separate threads because
162  * queue servicing may be blocked waiting for a syncq which may be also
163  * scheduled for background execution. This may create a deadlock when only one
164  * thread is used for both.
165  */
166 
167 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
168 
169 static kmutex_t service_queue;		/* protects all of servicing vars */
170 static kcondvar_t services_to_run;	/* wake up background service thread */
171 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
172 
173 /*
174  * List of queues scheduled for background processing due to lack of resources
175  * in the task queues. Protected by service_queue lock;
176  */
177 static struct queue *qhead;
178 static struct queue *qtail;
179 
180 /*
181  * Same list for syncqs
182  */
183 static syncq_t *sqhead;
184 static syncq_t *sqtail;
185 
186 static mblk_t *freebs_list;	/* list of buffers to free */
187 
188 /*
189  * Backup threads for servicing queues and syncqs
190  */
191 kthread_t *streams_qbkgrnd_thread;
192 kthread_t *streams_sqbkgrnd_thread;
193 
194 /*
195  * Bufcalls related variables.
196  */
197 struct bclist	strbcalls;	/* list of waiting bufcalls */
198 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
199 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
200 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
201 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
202 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
203 
204 kmutex_t	strresources;	/* protects global resources */
205 kmutex_t	muxifier;	/* single-threads multiplexor creation */
206 
207 static void	*str_stack_init(netstackid_t stackid, netstack_t *ns);
208 static void	str_stack_shutdown(netstackid_t stackid, void *arg);
209 static void	str_stack_fini(netstackid_t stackid, void *arg);
210 
211 /*
212  * run_queues is no longer used, but is kept in case some 3rd party
213  * module/driver decides to use it.
214  */
215 int run_queues = 0;
216 
217 /*
218  * sq_max_size is the depth of the syncq (in number of messages) before
219  * qfill_syncq() starts QFULL'ing destination queues. As its primary
220  * consumer - IP is no longer D_MTPERMOD, but there may be other
221  * modules/drivers depend on this syncq flow control, we prefer to
222  * choose a large number as the default value. For potential
223  * performance gain, this value is tunable in /etc/system.
224  */
225 int sq_max_size = 10000;
226 
227 /*
228  * The number of ciputctrl structures per syncq and stream we create when
229  * needed.
230  */
231 int n_ciputctrl;
232 int max_n_ciputctrl = 16;
233 /*
234  * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
235  */
236 int min_n_ciputctrl = 2;
237 
238 /*
239  * Per-driver/module syncqs
240  * ========================
241  *
242  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
243  * perdm structures, new entries being added (and new syncqs allocated) when
244  * setq() encounters a module/driver with a streamtab that it hasn't seen
245  * before.
246  * The reason for this mechanism is that some modules and drivers share a
247  * common streamtab and it is necessary for those modules and drivers to also
248  * share a common PERMOD syncq.
249  *
250  * perdm_list --> dm_str == streamtab_1
251  *                dm_sq == syncq_1
252  *                dm_ref
253  *                dm_next --> dm_str == streamtab_2
254  *                            dm_sq == syncq_2
255  *                            dm_ref
256  *                            dm_next --> ... NULL
257  *
258  * The dm_ref field is incremented for each new driver/module that takes
259  * a reference to the perdm structure and hence shares the syncq.
260  * References are held in the fmodsw_impl_t structure for each STREAMS module
261  * or the dev_impl array (indexed by device major number) for each driver.
262  *
263  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
264  *		     ^                 ^ ^               ^
265  *                   |  ______________/  |               |
266  *                   | /                 |               |
267  * dev_impl:     ...|x|y|...          module A	      module B
268  *
269  * When a module/driver is unloaded the reference count is decremented and,
270  * when it falls to zero, the perdm structure is removed from the list and
271  * the syncq is freed (see rele_dm()).
272  */
273 perdm_t *perdm_list = NULL;
274 static krwlock_t perdm_rwlock;
275 cdevsw_impl_t *devimpl;
276 
277 extern struct qinit strdata;
278 extern struct qinit stwdata;
279 
280 static void runservice(queue_t *);
281 static void streams_bufcall_service(void);
282 static void streams_qbkgrnd_service(void);
283 static void streams_sqbkgrnd_service(void);
284 static syncq_t *new_syncq(void);
285 static void free_syncq(syncq_t *);
286 static void outer_insert(syncq_t *, syncq_t *);
287 static void outer_remove(syncq_t *, syncq_t *);
288 static void write_now(syncq_t *);
289 static void clr_qfull(queue_t *);
290 static void runbufcalls(void);
291 static void sqenable(syncq_t *);
292 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
293 static void wait_q_syncq(queue_t *);
294 static void backenable_insertedq(queue_t *);
295 
296 static void queue_service(queue_t *);
297 static void stream_service(stdata_t *);
298 static void syncq_service(syncq_t *);
299 static void qwriter_outer_service(syncq_t *);
300 static void mblk_free(mblk_t *);
301 #ifdef DEBUG
302 static int qprocsareon(queue_t *);
303 #endif
304 
305 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
306 static void reset_nfsrv_ptr(queue_t *, queue_t *);
307 void set_qfull(queue_t *);
308 
309 static void sq_run_events(syncq_t *);
310 static int propagate_syncq(queue_t *);
311 
312 static void	blocksq(syncq_t *, ushort_t, int);
313 static void	unblocksq(syncq_t *, ushort_t, int);
314 static int	dropsq(syncq_t *, uint16_t);
315 static void	emptysq(syncq_t *);
316 static sqlist_t *sqlist_alloc(struct stdata *, int);
317 static void	sqlist_free(sqlist_t *);
318 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
319 static void	sqlist_insert(sqlist_t *, syncq_t *);
320 static void	sqlist_insertall(sqlist_t *, queue_t *);
321 
322 static void	strsetuio(stdata_t *);
323 
324 struct kmem_cache *stream_head_cache;
325 struct kmem_cache *queue_cache;
326 struct kmem_cache *syncq_cache;
327 struct kmem_cache *qband_cache;
328 struct kmem_cache *linkinfo_cache;
329 struct kmem_cache *ciputctrl_cache = NULL;
330 
331 static linkinfo_t *linkinfo_list;
332 
333 /* Global esballoc throttling queue */
334 static esb_queue_t system_esbq;
335 
336 /* Array of esballoc throttling queues, of length esbq_nelem */
337 static esb_queue_t *volatile system_esbq_array;
338 static int esbq_nelem;
339 static kmutex_t esbq_lock;
340 static int esbq_log2_cpus_per_q = 0;
341 
342 /* Scale the system_esbq length by setting number of CPUs per queue. */
343 uint_t esbq_cpus_per_q = 1;
344 
345 /*
346  * esballoc tunable parameters.
347  */
348 int		esbq_max_qlen = 0x16;	/* throttled queue length */
349 clock_t		esbq_timeout = 0x8;	/* timeout to process esb queue */
350 
351 /*
352  * Routines to handle esballoc queueing.
353  */
354 static void esballoc_process_queue(esb_queue_t *);
355 static void esballoc_enqueue_mblk(mblk_t *);
356 static void esballoc_timer(void *);
357 static void esballoc_set_timer(esb_queue_t *, clock_t);
358 static void esballoc_mblk_free(mblk_t *);
359 
360 /*
361  *  Qinit structure and Module_info structures
362  *	for passthru read and write queues
363  */
364 
365 static int pass_rput(queue_t *, mblk_t *);
366 static int pass_wput(queue_t *, mblk_t *);
367 static queue_t *link_addpassthru(stdata_t *);
368 static void link_rempassthru(queue_t *);
369 
370 struct  module_info passthru_info = {
371 	0,
372 	"passthru",
373 	0,
374 	INFPSZ,
375 	STRHIGH,
376 	STRLOW
377 };
378 
379 struct  qinit passthru_rinit = {
380 	pass_rput,
381 	NULL,
382 	NULL,
383 	NULL,
384 	NULL,
385 	&passthru_info,
386 	NULL
387 };
388 
389 struct  qinit passthru_winit = {
390 	pass_wput,
391 	NULL,
392 	NULL,
393 	NULL,
394 	NULL,
395 	&passthru_info,
396 	NULL
397 };
398 
399 /*
400  * Verify correctness of list head/tail pointers.
401  */
402 #define	LISTCHECK(head, tail, link) {				\
403 	EQUIV(head, tail);					\
404 	IMPLY(tail != NULL, tail->link == NULL);		\
405 }
406 
407 /*
408  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
409  * using a `link' field.
410  */
411 #define	ENQUEUE(el, head, tail, link) {				\
412 	ASSERT(el->link == NULL);				\
413 	LISTCHECK(head, tail, link);				\
414 	if (head == NULL)					\
415 		head = el;					\
416 	else							\
417 		tail->link = el;				\
418 	tail = el;						\
419 }
420 
421 /*
422  * Dequeue the first element of the list denoted by `head' and `tail' pointers
423  * using a `link' field and put result into `el'.
424  */
425 #define	DQ(el, head, tail, link) {				\
426 	LISTCHECK(head, tail, link);				\
427 	el = head;						\
428 	if (head != NULL) {					\
429 		head = head->link;				\
430 		if (head == NULL)				\
431 			tail = NULL;				\
432 		el->link = NULL;				\
433 	}							\
434 }
435 
436 /*
437  * Remove `el' from the list using `chase' and `curr' pointers and return result
438  * in `succeed'.
439  */
440 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
441 	LISTCHECK(head, tail, link);				\
442 	chase = NULL;						\
443 	succeed = 0;						\
444 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
445 		chase = curr;					\
446 	if (curr != NULL) {					\
447 		succeed = 1;					\
448 		ASSERT(curr == el);				\
449 		if (chase != NULL)				\
450 			chase->link = curr->link;		\
451 		else						\
452 			head = curr->link;			\
453 		curr->link = NULL;				\
454 		if (curr == tail)				\
455 			tail = chase;				\
456 	}							\
457 	LISTCHECK(head, tail, link);				\
458 }
459 
460 /* Handling of delayed messages on the inner syncq. */
461 
462 /*
463  * DEBUG versions should use function versions (to simplify tracing) and
464  * non-DEBUG kernels should use macro versions.
465  */
466 
467 /*
468  * Put a queue on the syncq list of queues.
469  * Assumes SQLOCK held.
470  */
471 #define	SQPUT_Q(sq, qp)							\
472 {									\
473 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
474 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
475 		/* The queue should not be linked anywhere */		\
476 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
477 		/* Head and tail may only be NULL simultaneously */	\
478 		EQUIV(sq->sq_head, sq->sq_tail);			\
479 		/* Queue may be only enqueued on its syncq */		\
480 		ASSERT(sq == qp->q_syncq);				\
481 		/* Check the correctness of SQ_MESSAGES flag */		\
482 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
483 		/* Sanity check first/last elements of the list */	\
484 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
485 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
486 		/*							\
487 		 * Sanity check of priority field: empty queue should	\
488 		 * have zero priority					\
489 		 * and nqueues equal to zero.				\
490 		 */							\
491 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
492 		/* Sanity check of sq_nqueues field */			\
493 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
494 		if (sq->sq_head == NULL) {				\
495 			sq->sq_head = sq->sq_tail = qp;			\
496 			sq->sq_flags |= SQ_MESSAGES;			\
497 		} else if (qp->q_spri == 0) {				\
498 			qp->q_sqprev = sq->sq_tail;			\
499 			sq->sq_tail->q_sqnext = qp;			\
500 			sq->sq_tail = qp;				\
501 		} else {						\
502 			/*						\
503 			 * Put this queue in priority order: higher	\
504 			 * priority gets closer to the head.		\
505 			 */						\
506 			queue_t **qpp = &sq->sq_tail;			\
507 			queue_t *qnext = NULL;				\
508 									\
509 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
510 				qnext = *qpp;				\
511 				qpp = &(*qpp)->q_sqprev;		\
512 			}						\
513 			qp->q_sqnext = qnext;				\
514 			qp->q_sqprev = *qpp;				\
515 			if (*qpp != NULL) {				\
516 				(*qpp)->q_sqnext = qp;			\
517 			} else {					\
518 				sq->sq_head = qp;			\
519 				sq->sq_pri = sq->sq_head->q_spri;	\
520 			}						\
521 			*qpp = qp;					\
522 		}							\
523 		qp->q_sqflags |= Q_SQQUEUED;				\
524 		qp->q_sqtstamp = ddi_get_lbolt();			\
525 		sq->sq_nqueues++;					\
526 	}								\
527 }
528 
529 /*
530  * Remove a queue from the syncq list
531  * Assumes SQLOCK held.
532  */
533 #define	SQRM_Q(sq, qp)							\
534 	{								\
535 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
536 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
537 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
538 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
539 		/* Check that the queue is actually in the list */	\
540 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
541 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
542 		ASSERT(sq->sq_nqueues != 0);				\
543 		if (qp->q_sqprev == NULL) {				\
544 			/* First queue on list, make head q_sqnext */	\
545 			sq->sq_head = qp->q_sqnext;			\
546 		} else {						\
547 			/* Make prev->next == next */			\
548 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
549 		}							\
550 		if (qp->q_sqnext == NULL) {				\
551 			/* Last queue on list, make tail sqprev */	\
552 			sq->sq_tail = qp->q_sqprev;			\
553 		} else {						\
554 			/* Make next->prev == prev */			\
555 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
556 		}							\
557 		/* clear out references on this queue */		\
558 		qp->q_sqprev = qp->q_sqnext = NULL;			\
559 		qp->q_sqflags &= ~Q_SQQUEUED;				\
560 		/* If there is nothing queued, clear SQ_MESSAGES */	\
561 		if (sq->sq_head != NULL) {				\
562 			sq->sq_pri = sq->sq_head->q_spri;		\
563 		} else	{						\
564 			sq->sq_flags &= ~SQ_MESSAGES;			\
565 			sq->sq_pri = 0;					\
566 		}							\
567 		sq->sq_nqueues--;					\
568 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
569 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
570 	}
571 
572 /* Hide the definition from the header file. */
573 #ifdef SQPUT_MP
574 #undef SQPUT_MP
575 #endif
576 
577 /*
578  * Put a message on the queue syncq.
579  * Assumes QLOCK held.
580  */
581 #define	SQPUT_MP(qp, mp)						\
582 	{								\
583 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
584 		ASSERT(qp->q_sqhead == NULL ||				\
585 		    (qp->q_sqtail != NULL &&				\
586 		    qp->q_sqtail->b_next == NULL));			\
587 		qp->q_syncqmsgs++;					\
588 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
589 		if (qp->q_sqhead == NULL) {				\
590 			qp->q_sqhead = qp->q_sqtail = mp;		\
591 		} else {						\
592 			qp->q_sqtail->b_next = mp;			\
593 			qp->q_sqtail = mp;				\
594 		}							\
595 		ASSERT(qp->q_syncqmsgs > 0);				\
596 		set_qfull(qp);						\
597 	}
598 
599 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
600 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
601 		if ((sq)->sq_ciputctrl != NULL) {			\
602 			int i;						\
603 			int nlocks = (sq)->sq_nciputctrl;		\
604 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
605 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
606 			for (i = 0; i <= nlocks; i++) {			\
607 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
608 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
609 			}						\
610 		}							\
611 	}
612 
613 
614 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
615 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
616 		if ((sq)->sq_ciputctrl != NULL) {			\
617 			int i;						\
618 			int nlocks = (sq)->sq_nciputctrl;		\
619 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
620 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
621 			for (i = 0; i <= nlocks; i++) {			\
622 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
623 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
624 			}						\
625 		}							\
626 	}
627 
628 /*
629  * Run service procedures for all queues in the stream head.
630  */
631 #define	STR_SERVICE(stp, q) {						\
632 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
633 	while (stp->sd_qhead != NULL) {					\
634 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
635 		ASSERT(stp->sd_nqueues > 0);				\
636 		stp->sd_nqueues--;					\
637 		ASSERT(!(q->q_flag & QINSERVICE));			\
638 		mutex_exit(&stp->sd_qlock);				\
639 		queue_service(q);					\
640 		mutex_enter(&stp->sd_qlock);				\
641 	}								\
642 	ASSERT(stp->sd_nqueues == 0);					\
643 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
644 }
645 
646 /*
647  * Constructor/destructor routines for the stream head cache
648  */
649 /* ARGSUSED */
650 static int
651 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
652 {
653 	stdata_t *stp = buf;
654 
655 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
656 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
657 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
658 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
659 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
660 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
661 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
662 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
663 	stp->sd_wrq = NULL;
664 
665 	return (0);
666 }
667 
668 /* ARGSUSED */
669 static void
670 stream_head_destructor(void *buf, void *cdrarg)
671 {
672 	stdata_t *stp = buf;
673 
674 	mutex_destroy(&stp->sd_lock);
675 	mutex_destroy(&stp->sd_reflock);
676 	mutex_destroy(&stp->sd_qlock);
677 	cv_destroy(&stp->sd_monitor);
678 	cv_destroy(&stp->sd_iocmonitor);
679 	cv_destroy(&stp->sd_refmonitor);
680 	cv_destroy(&stp->sd_qcv);
681 	cv_destroy(&stp->sd_zcopy_wait);
682 }
683 
684 /*
685  * Constructor/destructor routines for the queue cache
686  */
687 /* ARGSUSED */
688 static int
689 queue_constructor(void *buf, void *cdrarg, int kmflags)
690 {
691 	queinfo_t *qip = buf;
692 	queue_t *qp = &qip->qu_rqueue;
693 	queue_t *wqp = &qip->qu_wqueue;
694 	syncq_t	*sq = &qip->qu_syncq;
695 
696 	qp->q_first = NULL;
697 	qp->q_link = NULL;
698 	qp->q_count = 0;
699 	qp->q_mblkcnt = 0;
700 	qp->q_sqhead = NULL;
701 	qp->q_sqtail = NULL;
702 	qp->q_sqnext = NULL;
703 	qp->q_sqprev = NULL;
704 	qp->q_sqflags = 0;
705 	qp->q_rwcnt = 0;
706 	qp->q_spri = 0;
707 
708 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
709 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
710 
711 	wqp->q_first = NULL;
712 	wqp->q_link = NULL;
713 	wqp->q_count = 0;
714 	wqp->q_mblkcnt = 0;
715 	wqp->q_sqhead = NULL;
716 	wqp->q_sqtail = NULL;
717 	wqp->q_sqnext = NULL;
718 	wqp->q_sqprev = NULL;
719 	wqp->q_sqflags = 0;
720 	wqp->q_rwcnt = 0;
721 	wqp->q_spri = 0;
722 
723 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
724 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
725 
726 	sq->sq_head = NULL;
727 	sq->sq_tail = NULL;
728 	sq->sq_evhead = NULL;
729 	sq->sq_evtail = NULL;
730 	sq->sq_callbpend = NULL;
731 	sq->sq_outer = NULL;
732 	sq->sq_onext = NULL;
733 	sq->sq_oprev = NULL;
734 	sq->sq_next = NULL;
735 	sq->sq_svcflags = 0;
736 	sq->sq_servcount = 0;
737 	sq->sq_needexcl = 0;
738 	sq->sq_nqueues = 0;
739 	sq->sq_pri = 0;
740 
741 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
742 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
743 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
744 
745 	return (0);
746 }
747 
748 /* ARGSUSED */
749 static void
750 queue_destructor(void *buf, void *cdrarg)
751 {
752 	queinfo_t *qip = buf;
753 	queue_t *qp = &qip->qu_rqueue;
754 	queue_t *wqp = &qip->qu_wqueue;
755 	syncq_t	*sq = &qip->qu_syncq;
756 
757 	ASSERT(qp->q_sqhead == NULL);
758 	ASSERT(wqp->q_sqhead == NULL);
759 	ASSERT(qp->q_sqnext == NULL);
760 	ASSERT(wqp->q_sqnext == NULL);
761 	ASSERT(qp->q_rwcnt == 0);
762 	ASSERT(wqp->q_rwcnt == 0);
763 
764 	mutex_destroy(&qp->q_lock);
765 	cv_destroy(&qp->q_wait);
766 
767 	mutex_destroy(&wqp->q_lock);
768 	cv_destroy(&wqp->q_wait);
769 
770 	mutex_destroy(&sq->sq_lock);
771 	cv_destroy(&sq->sq_wait);
772 	cv_destroy(&sq->sq_exitwait);
773 }
774 
775 /*
776  * Constructor/destructor routines for the syncq cache
777  */
778 /* ARGSUSED */
779 static int
780 syncq_constructor(void *buf, void *cdrarg, int kmflags)
781 {
782 	syncq_t	*sq = buf;
783 
784 	bzero(buf, sizeof (syncq_t));
785 
786 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
787 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
788 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
789 
790 	return (0);
791 }
792 
793 /* ARGSUSED */
794 static void
795 syncq_destructor(void *buf, void *cdrarg)
796 {
797 	syncq_t	*sq = buf;
798 
799 	ASSERT(sq->sq_head == NULL);
800 	ASSERT(sq->sq_tail == NULL);
801 	ASSERT(sq->sq_evhead == NULL);
802 	ASSERT(sq->sq_evtail == NULL);
803 	ASSERT(sq->sq_callbpend == NULL);
804 	ASSERT(sq->sq_callbflags == 0);
805 	ASSERT(sq->sq_outer == NULL);
806 	ASSERT(sq->sq_onext == NULL);
807 	ASSERT(sq->sq_oprev == NULL);
808 	ASSERT(sq->sq_next == NULL);
809 	ASSERT(sq->sq_needexcl == 0);
810 	ASSERT(sq->sq_svcflags == 0);
811 	ASSERT(sq->sq_servcount == 0);
812 	ASSERT(sq->sq_nqueues == 0);
813 	ASSERT(sq->sq_pri == 0);
814 	ASSERT(sq->sq_count == 0);
815 	ASSERT(sq->sq_rmqcount == 0);
816 	ASSERT(sq->sq_cancelid == 0);
817 	ASSERT(sq->sq_ciputctrl == NULL);
818 	ASSERT(sq->sq_nciputctrl == 0);
819 	ASSERT(sq->sq_type == 0);
820 	ASSERT(sq->sq_flags == 0);
821 
822 	mutex_destroy(&sq->sq_lock);
823 	cv_destroy(&sq->sq_wait);
824 	cv_destroy(&sq->sq_exitwait);
825 }
826 
827 /* ARGSUSED */
828 static int
829 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
830 {
831 	ciputctrl_t *cip = buf;
832 	int i;
833 
834 	for (i = 0; i < n_ciputctrl; i++) {
835 		cip[i].ciputctrl_count = SQ_FASTPUT;
836 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
837 	}
838 
839 	return (0);
840 }
841 
842 /* ARGSUSED */
843 static void
844 ciputctrl_destructor(void *buf, void *cdrarg)
845 {
846 	ciputctrl_t *cip = buf;
847 	int i;
848 
849 	for (i = 0; i < n_ciputctrl; i++) {
850 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
851 		mutex_destroy(&cip[i].ciputctrl_lock);
852 	}
853 }
854 
855 /*
856  * Init routine run from main at boot time.
857  */
858 void
859 strinit(void)
860 {
861 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
862 
863 	stream_head_cache = kmem_cache_create("stream_head_cache",
864 	    sizeof (stdata_t), 0,
865 	    stream_head_constructor, stream_head_destructor, NULL,
866 	    NULL, NULL, 0);
867 
868 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
869 	    queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
870 
871 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
872 	    syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
873 
874 	qband_cache = kmem_cache_create("qband_cache",
875 	    sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
876 
877 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
878 	    sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
879 
880 	n_ciputctrl = ncpus;
881 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
882 	ASSERT(n_ciputctrl >= 1);
883 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
884 	if (n_ciputctrl >= min_n_ciputctrl) {
885 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
886 		    sizeof (ciputctrl_t) * n_ciputctrl,
887 		    sizeof (ciputctrl_t), ciputctrl_constructor,
888 		    ciputctrl_destructor, NULL, NULL, NULL, 0);
889 	}
890 
891 	streams_taskq = system_taskq;
892 
893 	if (streams_taskq == NULL)
894 		panic("strinit: no memory for streams taskq!");
895 
896 	bc_bkgrnd_thread = thread_create(NULL, 0,
897 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
898 
899 	streams_qbkgrnd_thread = thread_create(NULL, 0,
900 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
901 
902 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
903 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
904 
905 	/*
906 	 * Create STREAMS kstats.
907 	 */
908 	str_kstat = kstat_create("streams", 0, "strstat",
909 	    "net", KSTAT_TYPE_NAMED,
910 	    sizeof (str_statistics) / sizeof (kstat_named_t),
911 	    KSTAT_FLAG_VIRTUAL);
912 
913 	if (str_kstat != NULL) {
914 		str_kstat->ks_data = &str_statistics;
915 		kstat_install(str_kstat);
916 	}
917 
918 	/*
919 	 * TPI support routine initialisation.
920 	 */
921 	tpi_init();
922 
923 	/*
924 	 * Handle to have autopush and persistent link information per
925 	 * zone.
926 	 * Note: uses shutdown hook instead of destroy hook so that the
927 	 * persistent links can be torn down before the destroy hooks
928 	 * in the TCP/IP stack are called.
929 	 */
930 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
931 	    str_stack_fini);
932 }
933 
934 void
935 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
936 {
937 	struct stdata *stp;
938 
939 	ASSERT(vp->v_stream);
940 	stp = vp->v_stream;
941 	/* Have to hold sd_lock to prevent siglist from changing */
942 	mutex_enter(&stp->sd_lock);
943 	if (stp->sd_sigflags & event)
944 		strsendsig(stp->sd_siglist, event, band, error);
945 	mutex_exit(&stp->sd_lock);
946 }
947 
948 /*
949  * Send the "sevent" set of signals to a process.
950  * This might send more than one signal if the process is registered
951  * for multiple events. The caller should pass in an sevent that only
952  * includes the events for which the process has registered.
953  */
954 static void
955 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
956     uchar_t band, int error)
957 {
958 	ASSERT(MUTEX_HELD(&proc->p_lock));
959 
960 	info->si_band = 0;
961 	info->si_errno = 0;
962 
963 	if (sevent & S_ERROR) {
964 		sevent &= ~S_ERROR;
965 		info->si_code = POLL_ERR;
966 		info->si_errno = error;
967 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
968 		    "strsendsig:proc %p info %p", proc, info);
969 		sigaddq(proc, NULL, info, KM_NOSLEEP);
970 		info->si_errno = 0;
971 	}
972 	if (sevent & S_HANGUP) {
973 		sevent &= ~S_HANGUP;
974 		info->si_code = POLL_HUP;
975 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
976 		    "strsendsig:proc %p info %p", proc, info);
977 		sigaddq(proc, NULL, info, KM_NOSLEEP);
978 	}
979 	if (sevent & S_HIPRI) {
980 		sevent &= ~S_HIPRI;
981 		info->si_code = POLL_PRI;
982 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
983 		    "strsendsig:proc %p info %p", proc, info);
984 		sigaddq(proc, NULL, info, KM_NOSLEEP);
985 	}
986 	if (sevent & S_RDBAND) {
987 		sevent &= ~S_RDBAND;
988 		if (events & S_BANDURG)
989 			sigtoproc(proc, NULL, SIGURG);
990 		else
991 			sigtoproc(proc, NULL, SIGPOLL);
992 	}
993 	if (sevent & S_WRBAND) {
994 		sevent &= ~S_WRBAND;
995 		sigtoproc(proc, NULL, SIGPOLL);
996 	}
997 	if (sevent & S_INPUT) {
998 		sevent &= ~S_INPUT;
999 		info->si_code = POLL_IN;
1000 		info->si_band = band;
1001 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1002 		    "strsendsig:proc %p info %p", proc, info);
1003 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1004 		info->si_band = 0;
1005 	}
1006 	if (sevent & S_OUTPUT) {
1007 		sevent &= ~S_OUTPUT;
1008 		info->si_code = POLL_OUT;
1009 		info->si_band = band;
1010 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1011 		    "strsendsig:proc %p info %p", proc, info);
1012 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1013 		info->si_band = 0;
1014 	}
1015 	if (sevent & S_MSG) {
1016 		sevent &= ~S_MSG;
1017 		info->si_code = POLL_MSG;
1018 		info->si_band = band;
1019 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1020 		    "strsendsig:proc %p info %p", proc, info);
1021 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1022 		info->si_band = 0;
1023 	}
1024 	if (sevent & S_RDNORM) {
1025 		sevent &= ~S_RDNORM;
1026 		sigtoproc(proc, NULL, SIGPOLL);
1027 	}
1028 	if (sevent != 0) {
1029 		panic("strsendsig: unknown event(s) %x", sevent);
1030 	}
1031 }
1032 
1033 /*
1034  * Send SIGPOLL/SIGURG signal to all processes and process groups
1035  * registered on the given signal list that want a signal for at
1036  * least one of the specified events.
1037  *
1038  * Must be called with exclusive access to siglist (caller holding sd_lock).
1039  *
1040  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1041  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1042  * while it is in the siglist.
1043  *
1044  * For performance reasons (MP scalability) the code drops pidlock
1045  * when sending signals to a single process.
1046  * When sending to a process group the code holds
1047  * pidlock to prevent the membership in the process group from changing
1048  * while walking the p_pglink list.
1049  */
1050 void
1051 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1052 {
1053 	strsig_t *ssp;
1054 	k_siginfo_t info;
1055 	struct pid *pidp;
1056 	proc_t  *proc;
1057 
1058 	info.si_signo = SIGPOLL;
1059 	info.si_errno = 0;
1060 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1061 		int sevent;
1062 
1063 		sevent = ssp->ss_events & event;
1064 		if (sevent == 0)
1065 			continue;
1066 
1067 		if ((pidp = ssp->ss_pidp) == NULL) {
1068 			/* pid was released but still on event list */
1069 			continue;
1070 		}
1071 
1072 
1073 		if (ssp->ss_pid > 0) {
1074 			/*
1075 			 * XXX This unfortunately still generates
1076 			 * a signal when a fd is closed but
1077 			 * the proc is active.
1078 			 */
1079 			ASSERT(ssp->ss_pid == pidp->pid_id);
1080 
1081 			mutex_enter(&pidlock);
1082 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1083 			if (proc == NULL) {
1084 				mutex_exit(&pidlock);
1085 				continue;
1086 			}
1087 			mutex_enter(&proc->p_lock);
1088 			mutex_exit(&pidlock);
1089 			dosendsig(proc, ssp->ss_events, sevent, &info,
1090 			    band, error);
1091 			mutex_exit(&proc->p_lock);
1092 		} else {
1093 			/*
1094 			 * Send to process group. Hold pidlock across
1095 			 * calls to dosendsig().
1096 			 */
1097 			pid_t pgrp = -ssp->ss_pid;
1098 
1099 			mutex_enter(&pidlock);
1100 			proc = pgfind_zone(pgrp, ALL_ZONES);
1101 			while (proc != NULL) {
1102 				mutex_enter(&proc->p_lock);
1103 				dosendsig(proc, ssp->ss_events, sevent,
1104 				    &info, band, error);
1105 				mutex_exit(&proc->p_lock);
1106 				proc = proc->p_pglink;
1107 			}
1108 			mutex_exit(&pidlock);
1109 		}
1110 	}
1111 }
1112 
1113 /*
1114  * Attach a stream device or module.
1115  * qp is a read queue; the new queue goes in so its next
1116  * read ptr is the argument, and the write queue corresponding
1117  * to the argument points to this queue. Return 0 on success,
1118  * or a non-zero errno on failure.
1119  */
1120 int
1121 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1122     boolean_t is_insert)
1123 {
1124 	major_t			major;
1125 	cdevsw_impl_t		*dp;
1126 	struct streamtab	*str;
1127 	queue_t			*rq;
1128 	queue_t			*wrq;
1129 	uint32_t		qflag;
1130 	uint32_t		sqtype;
1131 	perdm_t			*dmp;
1132 	int			error;
1133 	int			sflag;
1134 
1135 	rq = allocq();
1136 	wrq = _WR(rq);
1137 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1138 
1139 	if (fp != NULL) {
1140 		str = fp->f_str;
1141 		qflag = fp->f_qflag;
1142 		sqtype = fp->f_sqtype;
1143 		dmp = fp->f_dmp;
1144 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1145 		sflag = MODOPEN;
1146 
1147 		/*
1148 		 * stash away a pointer to the module structure so we can
1149 		 * unref it in qdetach.
1150 		 */
1151 		rq->q_fp = fp;
1152 	} else {
1153 		ASSERT(!is_insert);
1154 
1155 		major = getmajor(*devp);
1156 		dp = &devimpl[major];
1157 
1158 		str = dp->d_str;
1159 		ASSERT(str == STREAMSTAB(major));
1160 
1161 		qflag = dp->d_qflag;
1162 		ASSERT(qflag & QISDRV);
1163 		sqtype = dp->d_sqtype;
1164 
1165 		/* create perdm_t if needed */
1166 		if (NEED_DM(dp->d_dmp, qflag))
1167 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1168 
1169 		dmp = dp->d_dmp;
1170 		sflag = 0;
1171 	}
1172 
1173 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1174 	    "qattach:qflag == %X(%X)", qflag, *devp);
1175 
1176 	/* setq might sleep in allocator - avoid holding locks. */
1177 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1178 
1179 	/*
1180 	 * Before calling the module's open routine, set up the q_next
1181 	 * pointer for inserting a module in the middle of a stream.
1182 	 *
1183 	 * Note that we can always set _QINSERTING and set up q_next
1184 	 * pointer for both inserting and pushing a module.  Then there
1185 	 * is no need for the is_insert parameter.  In insertq(), called
1186 	 * by qprocson(), assume that q_next of the new module always points
1187 	 * to the correct queue and use it for insertion.  Everything should
1188 	 * work out fine.  But in the first release of _I_INSERT, we
1189 	 * distinguish between inserting and pushing to make sure that
1190 	 * pushing a module follows the same code path as before.
1191 	 */
1192 	if (is_insert) {
1193 		rq->q_flag |= _QINSERTING;
1194 		rq->q_next = qp;
1195 	}
1196 
1197 	/*
1198 	 * If there is an outer perimeter get exclusive access during
1199 	 * the open procedure.  Bump up the reference count on the queue.
1200 	 */
1201 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1202 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1203 	if (error != 0)
1204 		goto failed;
1205 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1206 	ASSERT(qprocsareon(rq));
1207 	return (0);
1208 
1209 failed:
1210 	rq->q_flag &= ~_QINSERTING;
1211 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1212 		qprocsoff(rq);
1213 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1214 	rq->q_next = wrq->q_next = NULL;
1215 	qdetach(rq, 0, 0, crp, B_FALSE);
1216 	return (error);
1217 }
1218 
1219 /*
1220  * Handle second open of stream. For modules, set the
1221  * last argument to MODOPEN and do not pass any open flags.
1222  * Ignore dummydev since this is not the first open.
1223  */
1224 int
1225 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1226 {
1227 	int	error;
1228 	dev_t dummydev;
1229 	queue_t *wqp = _WR(qp);
1230 
1231 	ASSERT(qp->q_flag & QREADR);
1232 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1233 
1234 	dummydev = *devp;
1235 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1236 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1237 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1238 		mutex_enter(&STREAM(qp)->sd_lock);
1239 		qp->q_stream->sd_flag |= STREOPENFAIL;
1240 		mutex_exit(&STREAM(qp)->sd_lock);
1241 		return (error);
1242 	}
1243 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1244 
1245 	/*
1246 	 * successful open should have done qprocson()
1247 	 */
1248 	ASSERT(qprocsareon(_RD(qp)));
1249 	return (0);
1250 }
1251 
1252 /*
1253  * Detach a stream module or device.
1254  * If clmode == 1 then the module or driver was opened and its
1255  * close routine must be called. If clmode == 0, the module
1256  * or driver was never opened or the open failed, and so its close
1257  * should not be called.
1258  */
1259 void
1260 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1261 {
1262 	queue_t *wqp = _WR(qp);
1263 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1264 
1265 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1266 		stream_runservice(STREAM(qp));
1267 
1268 	if (clmode) {
1269 		/*
1270 		 * Make sure that all the messages on the write side syncq are
1271 		 * processed and nothing is left. Since we are closing, no new
1272 		 * messages may appear there.
1273 		 */
1274 		wait_q_syncq(wqp);
1275 
1276 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1277 		if (is_remove) {
1278 			mutex_enter(QLOCK(qp));
1279 			qp->q_flag |= _QREMOVING;
1280 			mutex_exit(QLOCK(qp));
1281 		}
1282 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1283 		/*
1284 		 * Check that qprocsoff() was actually called.
1285 		 */
1286 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1287 
1288 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1289 	} else {
1290 		disable_svc(qp);
1291 	}
1292 
1293 	/*
1294 	 * Allow any threads blocked in entersq to proceed and discover
1295 	 * the QWCLOSE is set.
1296 	 * Note: This assumes that all users of entersq check QWCLOSE.
1297 	 * Currently runservice is the only entersq that can happen
1298 	 * after removeq has finished.
1299 	 * Removeq will have discarded all messages destined to the closing
1300 	 * pair of queues from the syncq.
1301 	 * NOTE: Calling a function inside an assert is unconventional.
1302 	 * However, it does not cause any problem since flush_syncq() does
1303 	 * not change any state except when it returns non-zero i.e.
1304 	 * when the assert will trigger.
1305 	 */
1306 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1307 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1308 	ASSERT((qp->q_flag & QPERMOD) ||
1309 	    ((qp->q_syncq->sq_head == NULL) &&
1310 	    (wqp->q_syncq->sq_head == NULL)));
1311 
1312 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1313 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1314 	if (qp->q_fp != NULL)
1315 		fmodsw_rele(qp->q_fp);
1316 
1317 	/* freeq removes us from the outer perimeter if any */
1318 	freeq(qp);
1319 }
1320 
1321 /* Prevent service procedures from being called */
1322 void
1323 disable_svc(queue_t *qp)
1324 {
1325 	queue_t *wqp = _WR(qp);
1326 
1327 	ASSERT(qp->q_flag & QREADR);
1328 	mutex_enter(QLOCK(qp));
1329 	qp->q_flag |= QWCLOSE;
1330 	mutex_exit(QLOCK(qp));
1331 	mutex_enter(QLOCK(wqp));
1332 	wqp->q_flag |= QWCLOSE;
1333 	mutex_exit(QLOCK(wqp));
1334 }
1335 
1336 /* Allow service procedures to be called again */
1337 void
1338 enable_svc(queue_t *qp)
1339 {
1340 	queue_t *wqp = _WR(qp);
1341 
1342 	ASSERT(qp->q_flag & QREADR);
1343 	mutex_enter(QLOCK(qp));
1344 	qp->q_flag &= ~QWCLOSE;
1345 	mutex_exit(QLOCK(qp));
1346 	mutex_enter(QLOCK(wqp));
1347 	wqp->q_flag &= ~QWCLOSE;
1348 	mutex_exit(QLOCK(wqp));
1349 }
1350 
1351 /*
1352  * Remove queue from qhead/qtail if it is enabled.
1353  * Only reset QENAB if the queue was removed from the runlist.
1354  * A queue goes through 3 stages:
1355  *	It is on the service list and QENAB is set.
1356  *	It is removed from the service list but QENAB is still set.
1357  *	QENAB gets changed to QINSERVICE.
1358  *	QINSERVICE is reset (when the service procedure is done)
1359  * Thus we can not reset QENAB unless we actually removed it from the service
1360  * queue.
1361  */
1362 void
1363 remove_runlist(queue_t *qp)
1364 {
1365 	if (qp->q_flag & QENAB && qhead != NULL) {
1366 		queue_t *q_chase;
1367 		queue_t *q_curr;
1368 		int removed;
1369 
1370 		mutex_enter(&service_queue);
1371 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1372 		mutex_exit(&service_queue);
1373 		if (removed) {
1374 			STRSTAT(qremoved);
1375 			qp->q_flag &= ~QENAB;
1376 		}
1377 	}
1378 }
1379 
1380 
1381 /*
1382  * Wait for any pending service processing to complete.
1383  * The removal of queues from the runlist is not atomic with the
1384  * clearing of the QENABLED flag and setting the INSERVICE flag.
1385  * consequently it is possible for remove_runlist in strclose
1386  * to not find the queue on the runlist but for it to be QENABLED
1387  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1388  * as well as INSERVICE.
1389  */
1390 void
1391 wait_svc(queue_t *qp)
1392 {
1393 	queue_t *wqp = _WR(qp);
1394 
1395 	ASSERT(qp->q_flag & QREADR);
1396 
1397 	/*
1398 	 * Try to remove queues from qhead/qtail list.
1399 	 */
1400 	if (qhead != NULL) {
1401 		remove_runlist(qp);
1402 		remove_runlist(wqp);
1403 	}
1404 	/*
1405 	 * Wait till the syncqs associated with the queue disappear from the
1406 	 * background processing list.
1407 	 * This only needs to be done for non-PERMOD perimeters since
1408 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1409 	 * when the last module/driver is unloaded.
1410 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1411 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1412 	 * functions remove the queue from its syncq list, so sqthread will not
1413 	 * try to access the queue.
1414 	 */
1415 	if (!(qp->q_flag & QPERMOD)) {
1416 		syncq_t *rsq = qp->q_syncq;
1417 		syncq_t *wsq = wqp->q_syncq;
1418 
1419 		/*
1420 		 * Disable rsq and wsq and wait for any background processing of
1421 		 * syncq to complete.
1422 		 */
1423 		wait_sq_svc(rsq);
1424 		if (wsq != rsq)
1425 			wait_sq_svc(wsq);
1426 	}
1427 
1428 	mutex_enter(QLOCK(qp));
1429 	while (qp->q_flag & (QINSERVICE|QENAB))
1430 		cv_wait(&qp->q_wait, QLOCK(qp));
1431 	mutex_exit(QLOCK(qp));
1432 	mutex_enter(QLOCK(wqp));
1433 	while (wqp->q_flag & (QINSERVICE|QENAB))
1434 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1435 	mutex_exit(QLOCK(wqp));
1436 }
1437 
1438 /*
1439  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1440  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1441  * also be set, and is passed through to allocb_cred_wait().
1442  *
1443  * Returns errno on failure, zero on success.
1444  */
1445 int
1446 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1447 {
1448 	mblk_t *tmp;
1449 	ssize_t  count;
1450 	int error = 0;
1451 
1452 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1453 	    (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1454 
1455 	if (bp->b_datap->db_type == M_IOCTL) {
1456 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1457 	} else {
1458 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1459 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1460 	}
1461 	/*
1462 	 * strdoioctl validates ioc_count, so if this assert fails it
1463 	 * cannot be due to user error.
1464 	 */
1465 	ASSERT(count >= 0);
1466 
1467 	if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1468 	    curproc->p_pid)) == NULL) {
1469 		return (error);
1470 	}
1471 	error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1472 	if (error != 0) {
1473 		freeb(tmp);
1474 		return (error);
1475 	}
1476 	DB_CPID(tmp) = curproc->p_pid;
1477 	tmp->b_wptr += count;
1478 	bp->b_cont = tmp;
1479 
1480 	return (0);
1481 }
1482 
1483 /*
1484  * Copy ioctl data to user-land. Return non-zero errno on failure,
1485  * 0 for success.
1486  */
1487 int
1488 getiocd(mblk_t *bp, char *arg, int copymode)
1489 {
1490 	ssize_t count;
1491 	size_t  n;
1492 	int	error;
1493 
1494 	if (bp->b_datap->db_type == M_IOCACK)
1495 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1496 	else {
1497 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1498 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1499 	}
1500 	ASSERT(count >= 0);
1501 
1502 	for (bp = bp->b_cont; bp && count;
1503 	    count -= n, bp = bp->b_cont, arg += n) {
1504 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1505 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1506 		if (error)
1507 			return (error);
1508 	}
1509 	ASSERT(count == 0);
1510 	return (0);
1511 }
1512 
1513 /*
1514  * Allocate a linkinfo entry given the write queue of the
1515  * bottom module of the top stream and the write queue of the
1516  * stream head of the bottom stream.
1517  */
1518 linkinfo_t *
1519 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1520 {
1521 	linkinfo_t *linkp;
1522 
1523 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1524 
1525 	linkp->li_lblk.l_qtop = qup;
1526 	linkp->li_lblk.l_qbot = qdown;
1527 	linkp->li_fpdown = fpdown;
1528 
1529 	mutex_enter(&strresources);
1530 	linkp->li_next = linkinfo_list;
1531 	linkp->li_prev = NULL;
1532 	if (linkp->li_next)
1533 		linkp->li_next->li_prev = linkp;
1534 	linkinfo_list = linkp;
1535 	linkp->li_lblk.l_index = ++lnk_id;
1536 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1537 	mutex_exit(&strresources);
1538 
1539 	return (linkp);
1540 }
1541 
1542 /*
1543  * Free a linkinfo entry.
1544  */
1545 void
1546 lbfree(linkinfo_t *linkp)
1547 {
1548 	mutex_enter(&strresources);
1549 	if (linkp->li_next)
1550 		linkp->li_next->li_prev = linkp->li_prev;
1551 	if (linkp->li_prev)
1552 		linkp->li_prev->li_next = linkp->li_next;
1553 	else
1554 		linkinfo_list = linkp->li_next;
1555 	mutex_exit(&strresources);
1556 
1557 	kmem_cache_free(linkinfo_cache, linkp);
1558 }
1559 
1560 /*
1561  * Check for a potential linking cycle.
1562  * Return 1 if a link will result in a cycle,
1563  * and 0 otherwise.
1564  */
1565 int
1566 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1567 {
1568 	struct mux_node *np;
1569 	struct mux_edge *ep;
1570 	int i;
1571 	major_t lomaj;
1572 	major_t upmaj;
1573 	/*
1574 	 * if the lower stream is a pipe/FIFO, return, since link
1575 	 * cycles can not happen on pipes/FIFOs
1576 	 */
1577 	if (lostp->sd_vnode->v_type == VFIFO)
1578 		return (0);
1579 
1580 	for (i = 0; i < ss->ss_devcnt; i++) {
1581 		np = &ss->ss_mux_nodes[i];
1582 		MUX_CLEAR(np);
1583 	}
1584 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1585 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1586 	np = &ss->ss_mux_nodes[lomaj];
1587 	for (;;) {
1588 		if (!MUX_DIDVISIT(np)) {
1589 			if (np->mn_imaj == upmaj)
1590 				return (1);
1591 			if (np->mn_outp == NULL) {
1592 				MUX_VISIT(np);
1593 				if (np->mn_originp == NULL)
1594 					return (0);
1595 				np = np->mn_originp;
1596 				continue;
1597 			}
1598 			MUX_VISIT(np);
1599 			np->mn_startp = np->mn_outp;
1600 		} else {
1601 			if (np->mn_startp == NULL) {
1602 				if (np->mn_originp == NULL)
1603 					return (0);
1604 				else {
1605 					np = np->mn_originp;
1606 					continue;
1607 				}
1608 			}
1609 			/*
1610 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1611 			 * ignore the edge and move on. ep->me_nodep gets
1612 			 * set to NULL in mux_addedge() if it is a FIFO.
1613 			 *
1614 			 */
1615 			ep = np->mn_startp;
1616 			np->mn_startp = ep->me_nextp;
1617 			if (ep->me_nodep == NULL)
1618 				continue;
1619 			ep->me_nodep->mn_originp = np;
1620 			np = ep->me_nodep;
1621 		}
1622 	}
1623 }
1624 
1625 /*
1626  * Find linkinfo entry corresponding to the parameters.
1627  */
1628 linkinfo_t *
1629 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1630 {
1631 	linkinfo_t *linkp;
1632 	struct mux_edge *mep;
1633 	struct mux_node *mnp;
1634 	queue_t *qup;
1635 
1636 	mutex_enter(&strresources);
1637 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1638 		qup = getendq(stp->sd_wrq);
1639 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1640 			if ((qup == linkp->li_lblk.l_qtop) &&
1641 			    (!index || (index == linkp->li_lblk.l_index))) {
1642 				mutex_exit(&strresources);
1643 				return (linkp);
1644 			}
1645 		}
1646 	} else {
1647 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1648 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1649 		mep = mnp->mn_outp;
1650 		while (mep) {
1651 			if ((index == 0) || (index == mep->me_muxid))
1652 				break;
1653 			mep = mep->me_nextp;
1654 		}
1655 		if (!mep) {
1656 			mutex_exit(&strresources);
1657 			return (NULL);
1658 		}
1659 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1660 			if ((!linkp->li_lblk.l_qtop) &&
1661 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1662 				mutex_exit(&strresources);
1663 				return (linkp);
1664 			}
1665 		}
1666 	}
1667 	mutex_exit(&strresources);
1668 	return (NULL);
1669 }
1670 
1671 /*
1672  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1673  * last queue on the chain and return it.
1674  */
1675 queue_t *
1676 getendq(queue_t *q)
1677 {
1678 	ASSERT(q != NULL);
1679 	while (_SAMESTR(q))
1680 		q = q->q_next;
1681 	return (q);
1682 }
1683 
1684 /*
1685  * Wait for the syncq count to drop to zero.
1686  * sq could be either outer or inner.
1687  */
1688 
1689 static void
1690 wait_syncq(syncq_t *sq)
1691 {
1692 	uint16_t count;
1693 
1694 	mutex_enter(SQLOCK(sq));
1695 	count = sq->sq_count;
1696 	SQ_PUTLOCKS_ENTER(sq);
1697 	SUM_SQ_PUTCOUNTS(sq, count);
1698 	while (count != 0) {
1699 		sq->sq_flags |= SQ_WANTWAKEUP;
1700 		SQ_PUTLOCKS_EXIT(sq);
1701 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1702 		count = sq->sq_count;
1703 		SQ_PUTLOCKS_ENTER(sq);
1704 		SUM_SQ_PUTCOUNTS(sq, count);
1705 	}
1706 	SQ_PUTLOCKS_EXIT(sq);
1707 	mutex_exit(SQLOCK(sq));
1708 }
1709 
1710 /*
1711  * Wait while there are any messages for the queue in its syncq.
1712  */
1713 static void
1714 wait_q_syncq(queue_t *q)
1715 {
1716 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1717 		syncq_t *sq = q->q_syncq;
1718 
1719 		mutex_enter(SQLOCK(sq));
1720 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1721 			sq->sq_flags |= SQ_WANTWAKEUP;
1722 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1723 		}
1724 		mutex_exit(SQLOCK(sq));
1725 	}
1726 }
1727 
1728 
1729 int
1730 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1731     int lhlink)
1732 {
1733 	struct stdata *stp;
1734 	struct strioctl strioc;
1735 	struct linkinfo *linkp;
1736 	struct stdata *stpdown;
1737 	struct streamtab *str;
1738 	queue_t *passq;
1739 	syncq_t *passyncq;
1740 	queue_t *rq;
1741 	cdevsw_impl_t *dp;
1742 	uint32_t qflag;
1743 	uint32_t sqtype;
1744 	perdm_t *dmp;
1745 	int error = 0;
1746 	netstack_t *ns;
1747 	str_stack_t *ss;
1748 
1749 	stp = vp->v_stream;
1750 	TRACE_1(TR_FAC_STREAMS_FR,
1751 	    TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1752 	/*
1753 	 * Test for invalid upper stream
1754 	 */
1755 	if (stp->sd_flag & STRHUP) {
1756 		return (ENXIO);
1757 	}
1758 	if (vp->v_type == VFIFO) {
1759 		return (EINVAL);
1760 	}
1761 	if (stp->sd_strtab == NULL) {
1762 		return (EINVAL);
1763 	}
1764 	if (!stp->sd_strtab->st_muxwinit) {
1765 		return (EINVAL);
1766 	}
1767 	if (fpdown == NULL) {
1768 		return (EBADF);
1769 	}
1770 	ns = netstack_find_by_cred(crp);
1771 	ASSERT(ns != NULL);
1772 	ss = ns->netstack_str;
1773 	ASSERT(ss != NULL);
1774 
1775 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1776 		netstack_rele(ss->ss_netstack);
1777 		return (EINVAL);
1778 	}
1779 	mutex_enter(&muxifier);
1780 	if (stp->sd_flag & STPLEX) {
1781 		mutex_exit(&muxifier);
1782 		netstack_rele(ss->ss_netstack);
1783 		return (ENXIO);
1784 	}
1785 
1786 	/*
1787 	 * Test for invalid lower stream.
1788 	 * The check for the v_type != VFIFO and having a major
1789 	 * number not >= devcnt is done to avoid problems with
1790 	 * adding mux_node entry past the end of mux_nodes[].
1791 	 * For FIFO's we don't add an entry so this isn't a
1792 	 * problem.
1793 	 */
1794 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1795 	    (stpdown == stp) || (stpdown->sd_flag &
1796 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1797 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1798 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1799 	    linkcycle(stp, stpdown, ss)) {
1800 		mutex_exit(&muxifier);
1801 		netstack_rele(ss->ss_netstack);
1802 		return (EINVAL);
1803 	}
1804 	TRACE_1(TR_FAC_STREAMS_FR,
1805 	    TR_STPDOWN, "stpdown:%p", stpdown);
1806 	rq = getendq(stp->sd_wrq);
1807 	if (cmd == I_PLINK)
1808 		rq = NULL;
1809 
1810 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1811 
1812 	strioc.ic_cmd = cmd;
1813 	strioc.ic_timout = INFTIM;
1814 	strioc.ic_len = sizeof (struct linkblk);
1815 	strioc.ic_dp = (char *)&linkp->li_lblk;
1816 
1817 	/*
1818 	 * STRPLUMB protects plumbing changes and should be set before
1819 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1820 	 * and cleared in the end of mlink when passthru queue is removed.
1821 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1822 	 * queue is in-place (it is not a proper module and doesn't have open
1823 	 * entry point).
1824 	 *
1825 	 * STPLEX prevents any threads from entering the stream from above. It
1826 	 * can't be set before the call to link_addpassthru() because putnext
1827 	 * from below may cause stream head I/O routines to be called and these
1828 	 * routines assert that STPLEX is not set. After link_addpassthru()
1829 	 * nothing may come from below since the pass queue syncq is blocked.
1830 	 * Note also that STPLEX should be cleared before the call to
1831 	 * link_rempassthru() since when messages start flowing to the stream
1832 	 * head (e.g. because of message propagation from the pass queue) stream
1833 	 * head I/O routines may be called with STPLEX flag set.
1834 	 *
1835 	 * When STPLEX is set, nothing may come into the stream from above and
1836 	 * it is safe to do a setq which will change stream head. So, the
1837 	 * correct sequence of actions is:
1838 	 *
1839 	 * 1) Set STRPLUMB
1840 	 * 2) Call link_addpassthru()
1841 	 * 3) Set STPLEX
1842 	 * 4) Call setq and update the stream state
1843 	 * 5) Clear STPLEX
1844 	 * 6) Call link_rempassthru()
1845 	 * 7) Clear STRPLUMB
1846 	 *
1847 	 * The same sequence applies to munlink() code.
1848 	 */
1849 	mutex_enter(&stpdown->sd_lock);
1850 	stpdown->sd_flag |= STRPLUMB;
1851 	mutex_exit(&stpdown->sd_lock);
1852 	/*
1853 	 * Add passthru queue below lower mux. This will block
1854 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1855 	 */
1856 	passq = link_addpassthru(stpdown);
1857 
1858 	mutex_enter(&stpdown->sd_lock);
1859 	stpdown->sd_flag |= STPLEX;
1860 	mutex_exit(&stpdown->sd_lock);
1861 
1862 	rq = _RD(stpdown->sd_wrq);
1863 	/*
1864 	 * There may be messages in the streamhead's syncq due to messages
1865 	 * that arrived before link_addpassthru() was done. To avoid
1866 	 * background processing of the syncq happening simultaneous with
1867 	 * setq processing, we disable the streamhead syncq and wait until
1868 	 * existing background thread finishes working on it.
1869 	 */
1870 	wait_sq_svc(rq->q_syncq);
1871 	passyncq = passq->q_syncq;
1872 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1873 		blocksq(passyncq, SQ_BLOCKED, 0);
1874 
1875 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1876 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1877 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1878 
1879 	/* setq might sleep in allocator - avoid holding locks. */
1880 	/* Note: we are holding muxifier here. */
1881 
1882 	str = stp->sd_strtab;
1883 	dp = &devimpl[getmajor(vp->v_rdev)];
1884 	ASSERT(dp->d_str == str);
1885 
1886 	qflag = dp->d_qflag;
1887 	sqtype = dp->d_sqtype;
1888 
1889 	/* create perdm_t if needed */
1890 	if (NEED_DM(dp->d_dmp, qflag))
1891 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1892 
1893 	dmp = dp->d_dmp;
1894 
1895 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1896 	    B_TRUE);
1897 
1898 	/*
1899 	 * XXX Remove any "odd" messages from the queue.
1900 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1901 	 */
1902 	error = strdoioctl(stp, &strioc, FNATIVE,
1903 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1904 	if (error != 0) {
1905 		lbfree(linkp);
1906 
1907 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1908 			blocksq(passyncq, SQ_BLOCKED, 0);
1909 		/*
1910 		 * Restore the stream head queue and then remove
1911 		 * the passq. Turn off STPLEX before we turn on
1912 		 * the stream by removing the passq.
1913 		 */
1914 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1915 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1916 		    B_TRUE);
1917 
1918 		mutex_enter(&stpdown->sd_lock);
1919 		stpdown->sd_flag &= ~STPLEX;
1920 		mutex_exit(&stpdown->sd_lock);
1921 
1922 		link_rempassthru(passq);
1923 
1924 		mutex_enter(&stpdown->sd_lock);
1925 		stpdown->sd_flag &= ~STRPLUMB;
1926 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1927 		cv_broadcast(&stpdown->sd_monitor);
1928 		mutex_exit(&stpdown->sd_lock);
1929 
1930 		mutex_exit(&muxifier);
1931 		netstack_rele(ss->ss_netstack);
1932 		return (error);
1933 	}
1934 	mutex_enter(&fpdown->f_tlock);
1935 	fpdown->f_count++;
1936 	mutex_exit(&fpdown->f_tlock);
1937 
1938 	/*
1939 	 * if we've made it here the linkage is all set up so we should also
1940 	 * set up the layered driver linkages
1941 	 */
1942 
1943 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1944 	if (cmd == I_LINK) {
1945 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1946 	} else {
1947 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1948 	}
1949 
1950 	link_rempassthru(passq);
1951 
1952 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1953 
1954 	/*
1955 	 * Mark the upper stream as having dependent links
1956 	 * so that strclose can clean it up.
1957 	 */
1958 	if (cmd == I_LINK) {
1959 		mutex_enter(&stp->sd_lock);
1960 		stp->sd_flag |= STRHASLINKS;
1961 		mutex_exit(&stp->sd_lock);
1962 	}
1963 	/*
1964 	 * Wake up any other processes that may have been
1965 	 * waiting on the lower stream. These will all
1966 	 * error out.
1967 	 */
1968 	mutex_enter(&stpdown->sd_lock);
1969 	/* The passthru module is removed so we may release STRPLUMB */
1970 	stpdown->sd_flag &= ~STRPLUMB;
1971 	cv_broadcast(&rq->q_wait);
1972 	cv_broadcast(&_WR(rq)->q_wait);
1973 	cv_broadcast(&stpdown->sd_monitor);
1974 	mutex_exit(&stpdown->sd_lock);
1975 	mutex_exit(&muxifier);
1976 	*rvalp = linkp->li_lblk.l_index;
1977 	netstack_rele(ss->ss_netstack);
1978 	return (0);
1979 }
1980 
1981 int
1982 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1983 {
1984 	int		ret;
1985 	struct file	*fpdown;
1986 
1987 	fpdown = getf(arg);
1988 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1989 	if (fpdown != NULL)
1990 		releasef(arg);
1991 	return (ret);
1992 }
1993 
1994 /*
1995  * Unlink a multiplexor link. Stp is the controlling stream for the
1996  * link, and linkp points to the link's entry in the linkinfo list.
1997  * The muxifier lock must be held on entry and is dropped on exit.
1998  *
1999  * NOTE : Currently it is assumed that mux would process all the messages
2000  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
2001  * of the mux to handle all the messages that arrive before UNLINK.
2002  * If the mux has to send down messages on its lower stream before
2003  * ACKing I_UNLINK, then it *should* know to handle messages even
2004  * after the UNLINK is acked (actually it should be able to handle till we
2005  * re-block the read side of the pass queue here). If the mux does not
2006  * open up the lower stream, any messages that arrive during UNLINK
2007  * will be put in the stream head. In the case of lower stream opening
2008  * up, some messages might land in the stream head depending on when
2009  * the message arrived and when the read side of the pass queue was
2010  * re-blocked.
2011  */
2012 int
2013 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2014     str_stack_t *ss)
2015 {
2016 	struct strioctl strioc;
2017 	struct stdata *stpdown;
2018 	queue_t *rq, *wrq;
2019 	queue_t	*passq;
2020 	syncq_t *passyncq;
2021 	int error = 0;
2022 	file_t *fpdown;
2023 
2024 	ASSERT(MUTEX_HELD(&muxifier));
2025 
2026 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2027 
2028 	/*
2029 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2030 	 */
2031 	mutex_enter(&stpdown->sd_lock);
2032 	stpdown->sd_flag |= STRPLUMB;
2033 	mutex_exit(&stpdown->sd_lock);
2034 
2035 	/*
2036 	 * Add passthru queue below lower mux. This will block
2037 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2038 	 */
2039 	passq = link_addpassthru(stpdown);
2040 
2041 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2042 		strioc.ic_cmd = I_UNLINK;
2043 	else
2044 		strioc.ic_cmd = I_PUNLINK;
2045 	strioc.ic_timout = INFTIM;
2046 	strioc.ic_len = sizeof (struct linkblk);
2047 	strioc.ic_dp = (char *)&linkp->li_lblk;
2048 
2049 	error = strdoioctl(stp, &strioc, FNATIVE,
2050 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2051 
2052 	/*
2053 	 * If there was an error and this is not called via strclose,
2054 	 * return to the user. Otherwise, pretend there was no error
2055 	 * and close the link.
2056 	 */
2057 	if (error) {
2058 		if (flag & LINKCLOSE) {
2059 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2060 			    "unlink ioctl, closing anyway (%d)\n", error);
2061 		} else {
2062 			link_rempassthru(passq);
2063 			mutex_enter(&stpdown->sd_lock);
2064 			stpdown->sd_flag &= ~STRPLUMB;
2065 			cv_broadcast(&stpdown->sd_monitor);
2066 			mutex_exit(&stpdown->sd_lock);
2067 			mutex_exit(&muxifier);
2068 			return (error);
2069 		}
2070 	}
2071 
2072 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2073 	fpdown = linkp->li_fpdown;
2074 	lbfree(linkp);
2075 
2076 	/*
2077 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2078 	 * can slow others down. It's okay to since attempts to mlink() this
2079 	 * stream will be stopped because STPLEX is still set in the stdata
2080 	 * structure, and munlink() is stopped because mux_rmvedge() and
2081 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2082 	 * respectively.  Note that we defer the closef() of fpdown until
2083 	 * after we drop muxifier since strclose() can call munlinkall().
2084 	 */
2085 	mutex_exit(&muxifier);
2086 
2087 	wrq = stpdown->sd_wrq;
2088 	rq = _RD(wrq);
2089 
2090 	/*
2091 	 * Get rid of outstanding service procedure runs, before we make
2092 	 * it a stream head, since a stream head doesn't have any service
2093 	 * procedure.
2094 	 */
2095 	disable_svc(rq);
2096 	wait_svc(rq);
2097 
2098 	/*
2099 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2100 	 * is queued up to be finished. mux should take care that nothing is
2101 	 * send down to this queue. We should do it now as we're going to block
2102 	 * passyncq if it was unblocked.
2103 	 */
2104 	if (wrq->q_flag & QPERMOD) {
2105 		syncq_t	*sq = wrq->q_syncq;
2106 
2107 		mutex_enter(SQLOCK(sq));
2108 		while (wrq->q_sqflags & Q_SQQUEUED) {
2109 			sq->sq_flags |= SQ_WANTWAKEUP;
2110 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2111 		}
2112 		mutex_exit(SQLOCK(sq));
2113 	}
2114 	passyncq = passq->q_syncq;
2115 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2116 
2117 		syncq_t *sq, *outer;
2118 
2119 		/*
2120 		 * Messages could be flowing from underneath. We will
2121 		 * block the read side of the passq. This would be
2122 		 * sufficient for QPAIR and QPERQ muxes to ensure
2123 		 * that no data is flowing up into this queue
2124 		 * and hence no thread active in this instance of
2125 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2126 		 * could be messages on the inner and outer/inner
2127 		 * syncqs respectively. We will wait for them to drain.
2128 		 * Because passq is blocked messages end up in the syncq
2129 		 * And qfill_syncq could possibly end up setting QFULL
2130 		 * which will access the rq->q_flag. Hence, we have to
2131 		 * acquire the QLOCK in setq.
2132 		 *
2133 		 * XXX Messages can also flow from top into this
2134 		 * queue though the unlink is over (Ex. some instance
2135 		 * in putnext() called from top that has still not
2136 		 * accessed this queue. And also putq(lowerq) ?).
2137 		 * Solution : How about blocking the l_qtop queue ?
2138 		 * Do we really care about such pure D_MP muxes ?
2139 		 */
2140 
2141 		blocksq(passyncq, SQ_BLOCKED, 0);
2142 
2143 		sq = rq->q_syncq;
2144 		if ((outer = sq->sq_outer) != NULL) {
2145 
2146 			/*
2147 			 * We have to just wait for the outer sq_count
2148 			 * drop to zero. As this does not prevent new
2149 			 * messages to enter the outer perimeter, this
2150 			 * is subject to starvation.
2151 			 *
2152 			 * NOTE :Because of blocksq above, messages could
2153 			 * be in the inner syncq only because of some
2154 			 * thread holding the outer perimeter exclusively.
2155 			 * Hence it would be sufficient to wait for the
2156 			 * exclusive holder of the outer perimeter to drain
2157 			 * the inner and outer syncqs. But we will not depend
2158 			 * on this feature and hence check the inner syncqs
2159 			 * separately.
2160 			 */
2161 			wait_syncq(outer);
2162 		}
2163 
2164 
2165 		/*
2166 		 * There could be messages destined for
2167 		 * this queue. Let the exclusive holder
2168 		 * drain it.
2169 		 */
2170 
2171 		wait_syncq(sq);
2172 		ASSERT((rq->q_flag & QPERMOD) ||
2173 		    ((rq->q_syncq->sq_head == NULL) &&
2174 		    (_WR(rq)->q_syncq->sq_head == NULL)));
2175 	}
2176 
2177 	/*
2178 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2179 	 * case as we don't disable its syncq or remove it off the syncq
2180 	 * service list.
2181 	 */
2182 	if (rq->q_flag & QPERMOD) {
2183 		syncq_t	*sq = rq->q_syncq;
2184 
2185 		mutex_enter(SQLOCK(sq));
2186 		while (rq->q_sqflags & Q_SQQUEUED) {
2187 			sq->sq_flags |= SQ_WANTWAKEUP;
2188 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2189 		}
2190 		mutex_exit(SQLOCK(sq));
2191 	}
2192 
2193 	/*
2194 	 * flush_syncq changes states only when there are some messages to
2195 	 * free, i.e. when it returns non-zero value to return.
2196 	 */
2197 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2198 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2199 
2200 	/*
2201 	 * Nobody else should know about this queue now.
2202 	 * If the mux did not process the messages before
2203 	 * acking the I_UNLINK, free them now.
2204 	 */
2205 
2206 	flushq(rq, FLUSHALL);
2207 	flushq(_WR(rq), FLUSHALL);
2208 
2209 	/*
2210 	 * Convert the mux lower queue into a stream head queue.
2211 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2212 	 */
2213 	rq->q_ptr = wrq->q_ptr = stpdown;
2214 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2215 
2216 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2217 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2218 
2219 	enable_svc(rq);
2220 
2221 	/*
2222 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2223 	 * needs to be set to prevent reopen() of the stream - such reopen may
2224 	 * try to call non-existent pass queue open routine and panic.
2225 	 */
2226 	mutex_enter(&stpdown->sd_lock);
2227 	stpdown->sd_flag &= ~STPLEX;
2228 	mutex_exit(&stpdown->sd_lock);
2229 
2230 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2231 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2232 
2233 	/* clean up the layered driver linkages */
2234 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2235 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2236 	} else {
2237 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2238 	}
2239 
2240 	link_rempassthru(passq);
2241 
2242 	/*
2243 	 * Now all plumbing changes are finished and STRPLUMB is no
2244 	 * longer needed.
2245 	 */
2246 	mutex_enter(&stpdown->sd_lock);
2247 	stpdown->sd_flag &= ~STRPLUMB;
2248 	cv_broadcast(&stpdown->sd_monitor);
2249 	mutex_exit(&stpdown->sd_lock);
2250 
2251 	(void) closef(fpdown);
2252 	return (0);
2253 }
2254 
2255 /*
2256  * Unlink all multiplexor links for which stp is the controlling stream.
2257  * Return 0, or a non-zero errno on failure.
2258  */
2259 int
2260 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2261 {
2262 	linkinfo_t *linkp;
2263 	int error = 0;
2264 
2265 	mutex_enter(&muxifier);
2266 	while (linkp = findlinks(stp, 0, flag, ss)) {
2267 		/*
2268 		 * munlink() releases the muxifier lock.
2269 		 */
2270 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2271 			return (error);
2272 		mutex_enter(&muxifier);
2273 	}
2274 	mutex_exit(&muxifier);
2275 	return (0);
2276 }
2277 
2278 /*
2279  * A multiplexor link has been made. Add an
2280  * edge to the directed graph.
2281  */
2282 void
2283 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2284 {
2285 	struct mux_node *np;
2286 	struct mux_edge *ep;
2287 	major_t upmaj;
2288 	major_t lomaj;
2289 
2290 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2291 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2292 	np = &ss->ss_mux_nodes[upmaj];
2293 	if (np->mn_outp) {
2294 		ep = np->mn_outp;
2295 		while (ep->me_nextp)
2296 			ep = ep->me_nextp;
2297 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2298 		ep = ep->me_nextp;
2299 	} else {
2300 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2301 		ep = np->mn_outp;
2302 	}
2303 	ep->me_nextp = NULL;
2304 	ep->me_muxid = muxid;
2305 	/*
2306 	 * Save the dev_t for the purposes of str_stack_shutdown.
2307 	 * str_stack_shutdown assumes that the device allows reopen, since
2308 	 * this dev_t is the one after any cloning by xx_open().
2309 	 * Would prefer finding the dev_t from before any cloning,
2310 	 * but specfs doesn't retain that.
2311 	 */
2312 	ep->me_dev = upstp->sd_vnode->v_rdev;
2313 	if (lostp->sd_vnode->v_type == VFIFO)
2314 		ep->me_nodep = NULL;
2315 	else
2316 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2317 }
2318 
2319 /*
2320  * A multiplexor link has been removed. Remove the
2321  * edge in the directed graph.
2322  */
2323 void
2324 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2325 {
2326 	struct mux_node *np;
2327 	struct mux_edge *ep;
2328 	struct mux_edge *pep = NULL;
2329 	major_t upmaj;
2330 
2331 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2332 	np = &ss->ss_mux_nodes[upmaj];
2333 	ASSERT(np->mn_outp != NULL);
2334 	ep = np->mn_outp;
2335 	while (ep) {
2336 		if (ep->me_muxid == muxid) {
2337 			if (pep)
2338 				pep->me_nextp = ep->me_nextp;
2339 			else
2340 				np->mn_outp = ep->me_nextp;
2341 			kmem_free(ep, sizeof (struct mux_edge));
2342 			return;
2343 		}
2344 		pep = ep;
2345 		ep = ep->me_nextp;
2346 	}
2347 	ASSERT(0);	/* should not reach here */
2348 }
2349 
2350 /*
2351  * Translate the device flags (from conf.h) to the corresponding
2352  * qflag and sq_flag (type) values.
2353  */
2354 int
2355 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2356     uint32_t *sqtypep)
2357 {
2358 	uint32_t qflag = 0;
2359 	uint32_t sqtype = 0;
2360 
2361 	if (devflag & _D_OLD)
2362 		goto bad;
2363 
2364 	/* Inner perimeter presence and scope */
2365 	switch (devflag & D_MTINNER_MASK) {
2366 	case D_MP:
2367 		qflag |= QMTSAFE;
2368 		sqtype |= SQ_CI;
2369 		break;
2370 	case D_MTPERQ|D_MP:
2371 		qflag |= QPERQ;
2372 		break;
2373 	case D_MTQPAIR|D_MP:
2374 		qflag |= QPAIR;
2375 		break;
2376 	case D_MTPERMOD|D_MP:
2377 		qflag |= QPERMOD;
2378 		break;
2379 	default:
2380 		goto bad;
2381 	}
2382 
2383 	/* Outer perimeter */
2384 	if (devflag & D_MTOUTPERIM) {
2385 		switch (devflag & D_MTINNER_MASK) {
2386 		case D_MP:
2387 		case D_MTPERQ|D_MP:
2388 		case D_MTQPAIR|D_MP:
2389 			break;
2390 		default:
2391 			goto bad;
2392 		}
2393 		qflag |= QMTOUTPERIM;
2394 	}
2395 
2396 	/* Inner perimeter modifiers */
2397 	if (devflag & D_MTINNER_MOD) {
2398 		switch (devflag & D_MTINNER_MASK) {
2399 		case D_MP:
2400 			goto bad;
2401 		default:
2402 			break;
2403 		}
2404 		if (devflag & D_MTPUTSHARED)
2405 			sqtype |= SQ_CIPUT;
2406 		if (devflag & _D_MTOCSHARED) {
2407 			/*
2408 			 * The code in putnext assumes that it has the
2409 			 * highest concurrency by not checking sq_count.
2410 			 * Thus _D_MTOCSHARED can only be supported when
2411 			 * D_MTPUTSHARED is set.
2412 			 */
2413 			if (!(devflag & D_MTPUTSHARED))
2414 				goto bad;
2415 			sqtype |= SQ_CIOC;
2416 		}
2417 		if (devflag & _D_MTCBSHARED) {
2418 			/*
2419 			 * The code in putnext assumes that it has the
2420 			 * highest concurrency by not checking sq_count.
2421 			 * Thus _D_MTCBSHARED can only be supported when
2422 			 * D_MTPUTSHARED is set.
2423 			 */
2424 			if (!(devflag & D_MTPUTSHARED))
2425 				goto bad;
2426 			sqtype |= SQ_CICB;
2427 		}
2428 		if (devflag & _D_MTSVCSHARED) {
2429 			/*
2430 			 * The code in putnext assumes that it has the
2431 			 * highest concurrency by not checking sq_count.
2432 			 * Thus _D_MTSVCSHARED can only be supported when
2433 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2434 			 * supported only for QPERMOD.
2435 			 */
2436 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2437 				goto bad;
2438 			sqtype |= SQ_CISVC;
2439 		}
2440 	}
2441 
2442 	/* Default outer perimeter concurrency */
2443 	sqtype |= SQ_CO;
2444 
2445 	/* Outer perimeter modifiers */
2446 	if (devflag & D_MTOCEXCL) {
2447 		if (!(devflag & D_MTOUTPERIM)) {
2448 			/* No outer perimeter */
2449 			goto bad;
2450 		}
2451 		sqtype &= ~SQ_COOC;
2452 	}
2453 
2454 	/* Synchronous Streams extended qinit structure */
2455 	if (devflag & D_SYNCSTR)
2456 		qflag |= QSYNCSTR;
2457 
2458 	/*
2459 	 * Private flag used by a transport module to indicate
2460 	 * to sockfs that it supports direct-access mode without
2461 	 * having to go through STREAMS.
2462 	 */
2463 	if (devflag & _D_DIRECT) {
2464 		/* Reject unless the module is fully-MT (no perimeter) */
2465 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2466 			goto bad;
2467 		qflag |= _QDIRECT;
2468 	}
2469 
2470 	/*
2471 	 * Private flag used to indicate that a streams module should only
2472 	 * be pushed once. The TTY streams modules have this flag since if
2473 	 * libc believes itself to be an xpg4 process then it will
2474 	 * automatically and unconditionally push them when a PTS device is
2475 	 * opened. If an application is not aware of this then without this
2476 	 * flag we would end up with duplicate modules.
2477 	 */
2478 	if (devflag & _D_SINGLE_INSTANCE)
2479 		qflag |= _QSINGLE_INSTANCE;
2480 
2481 	*qflagp = qflag;
2482 	*sqtypep = sqtype;
2483 	return (0);
2484 
2485 bad:
2486 	cmn_err(CE_WARN,
2487 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2488 	    (int)(qflag & D_MTSAFETY_MASK),
2489 	    stp->st_rdinit->qi_minfo->mi_idname);
2490 
2491 	return (EINVAL);
2492 }
2493 
2494 /*
2495  * Set the interface values for a pair of queues (qinit structure,
2496  * packet sizes, water marks).
2497  * setq assumes that the caller does not have a claim (entersq or claimq)
2498  * on the queue.
2499  */
2500 void
2501 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2502     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2503 {
2504 	queue_t *wq;
2505 	syncq_t	*sq, *outer;
2506 
2507 	ASSERT(rq->q_flag & QREADR);
2508 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2509 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2510 
2511 	wq = _WR(rq);
2512 	rq->q_qinfo = rinit;
2513 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2514 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2515 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2516 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2517 	wq->q_qinfo = winit;
2518 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2519 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2520 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2521 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2522 
2523 	/* Remove old syncqs */
2524 	sq = rq->q_syncq;
2525 	outer = sq->sq_outer;
2526 	if (outer != NULL) {
2527 		ASSERT(wq->q_syncq->sq_outer == outer);
2528 		outer_remove(outer, rq->q_syncq);
2529 		if (wq->q_syncq != rq->q_syncq)
2530 			outer_remove(outer, wq->q_syncq);
2531 	}
2532 	ASSERT(sq->sq_outer == NULL);
2533 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2534 
2535 	if (sq != SQ(rq)) {
2536 		if (!(rq->q_flag & QPERMOD))
2537 			free_syncq(sq);
2538 		if (wq->q_syncq == rq->q_syncq)
2539 			wq->q_syncq = NULL;
2540 		rq->q_syncq = NULL;
2541 	}
2542 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2543 	    wq->q_syncq != SQ(rq)) {
2544 		free_syncq(wq->q_syncq);
2545 		wq->q_syncq = NULL;
2546 	}
2547 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2548 	    rq->q_syncq->sq_tail == NULL));
2549 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2550 	    wq->q_syncq->sq_tail == NULL));
2551 
2552 	if (!(rq->q_flag & QPERMOD) &&
2553 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2554 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2555 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2556 		    rq->q_syncq->sq_nciputctrl, 0);
2557 		ASSERT(ciputctrl_cache != NULL);
2558 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2559 		rq->q_syncq->sq_ciputctrl = NULL;
2560 		rq->q_syncq->sq_nciputctrl = 0;
2561 	}
2562 
2563 	if (!(wq->q_flag & QPERMOD) &&
2564 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2565 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2566 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2567 		    wq->q_syncq->sq_nciputctrl, 0);
2568 		ASSERT(ciputctrl_cache != NULL);
2569 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2570 		wq->q_syncq->sq_ciputctrl = NULL;
2571 		wq->q_syncq->sq_nciputctrl = 0;
2572 	}
2573 
2574 	sq = SQ(rq);
2575 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2576 	ASSERT(sq->sq_outer == NULL);
2577 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2578 
2579 	/*
2580 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2581 	 * bits in sq_flag based on the sqtype.
2582 	 */
2583 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2584 
2585 	rq->q_syncq = wq->q_syncq = sq;
2586 	sq->sq_type = sqtype;
2587 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2588 
2589 	/*
2590 	 *  We are making sq_svcflags zero,
2591 	 *  resetting SQ_DISABLED in case it was set by
2592 	 *  wait_svc() in the munlink path.
2593 	 *
2594 	 */
2595 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2596 	sq->sq_svcflags = 0;
2597 
2598 	/*
2599 	 * We need to acquire the lock here for the mlink and munlink case,
2600 	 * where canputnext, backenable, etc can access the q_flag.
2601 	 */
2602 	if (lock_needed) {
2603 		mutex_enter(QLOCK(rq));
2604 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2605 		mutex_exit(QLOCK(rq));
2606 		mutex_enter(QLOCK(wq));
2607 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2608 		mutex_exit(QLOCK(wq));
2609 	} else {
2610 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2611 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2612 	}
2613 
2614 	if (qflag & QPERQ) {
2615 		/* Allocate a separate syncq for the write side */
2616 		sq = new_syncq();
2617 		sq->sq_type = rq->q_syncq->sq_type;
2618 		sq->sq_flags = rq->q_syncq->sq_flags;
2619 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2620 		    sq->sq_oprev == NULL);
2621 		wq->q_syncq = sq;
2622 	}
2623 	if (qflag & QPERMOD) {
2624 		sq = dmp->dm_sq;
2625 
2626 		/*
2627 		 * Assert that we do have an inner perimeter syncq and that it
2628 		 * does not have an outer perimeter associated with it.
2629 		 */
2630 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2631 		    sq->sq_oprev == NULL);
2632 		rq->q_syncq = wq->q_syncq = sq;
2633 	}
2634 	if (qflag & QMTOUTPERIM) {
2635 		outer = dmp->dm_sq;
2636 
2637 		ASSERT(outer->sq_outer == NULL);
2638 		outer_insert(outer, rq->q_syncq);
2639 		if (wq->q_syncq != rq->q_syncq)
2640 			outer_insert(outer, wq->q_syncq);
2641 	}
2642 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2643 	    (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2644 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2645 	    (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2646 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2647 
2648 	/*
2649 	 * Initialize struio() types.
2650 	 */
2651 	rq->q_struiot =
2652 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2653 	wq->q_struiot =
2654 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2655 }
2656 
2657 perdm_t *
2658 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2659 {
2660 	syncq_t	*sq;
2661 	perdm_t	**pp;
2662 	perdm_t	*p;
2663 	perdm_t	*dmp;
2664 
2665 	ASSERT(str != NULL);
2666 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2667 
2668 	rw_enter(&perdm_rwlock, RW_READER);
2669 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2670 		if (p->dm_str == str) {	/* found one */
2671 			atomic_inc_32(&(p->dm_ref));
2672 			rw_exit(&perdm_rwlock);
2673 			return (p);
2674 		}
2675 	}
2676 	rw_exit(&perdm_rwlock);
2677 
2678 	sq = new_syncq();
2679 	if (qflag & QPERMOD) {
2680 		sq->sq_type = sqtype | SQ_PERMOD;
2681 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2682 	} else {
2683 		ASSERT(qflag & QMTOUTPERIM);
2684 		sq->sq_onext = sq->sq_oprev = sq;
2685 	}
2686 
2687 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2688 	dmp->dm_sq = sq;
2689 	dmp->dm_str = str;
2690 	dmp->dm_ref = 1;
2691 	dmp->dm_next = NULL;
2692 
2693 	rw_enter(&perdm_rwlock, RW_WRITER);
2694 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2695 		if (p->dm_str == str) {	/* already present */
2696 			p->dm_ref++;
2697 			rw_exit(&perdm_rwlock);
2698 			free_syncq(sq);
2699 			kmem_free(dmp, sizeof (perdm_t));
2700 			return (p);
2701 		}
2702 	}
2703 
2704 	*pp = dmp;
2705 	rw_exit(&perdm_rwlock);
2706 	return (dmp);
2707 }
2708 
2709 void
2710 rele_dm(perdm_t *dmp)
2711 {
2712 	perdm_t **pp;
2713 	perdm_t *p;
2714 
2715 	rw_enter(&perdm_rwlock, RW_WRITER);
2716 	ASSERT(dmp->dm_ref > 0);
2717 
2718 	if (--dmp->dm_ref > 0) {
2719 		rw_exit(&perdm_rwlock);
2720 		return;
2721 	}
2722 
2723 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2724 		if (p == dmp)
2725 			break;
2726 	ASSERT(p == dmp);
2727 	*pp = p->dm_next;
2728 	rw_exit(&perdm_rwlock);
2729 
2730 	/*
2731 	 * Wait for any background processing that relies on the
2732 	 * syncq to complete before it is freed.
2733 	 */
2734 	wait_sq_svc(p->dm_sq);
2735 	free_syncq(p->dm_sq);
2736 	kmem_free(p, sizeof (perdm_t));
2737 }
2738 
2739 /*
2740  * Make a protocol message given control and data buffers.
2741  * n.b., this can block; be careful of what locks you hold when calling it.
2742  *
2743  * If sd_maxblk is less than *iosize this routine can fail part way through
2744  * (due to an allocation failure). In this case on return *iosize will contain
2745  * the amount that was consumed. Otherwise *iosize will not be modified
2746  * i.e. it will contain the amount that was consumed.
2747  */
2748 int
2749 strmakemsg(
2750 	struct strbuf *mctl,
2751 	ssize_t *iosize,
2752 	struct uio *uiop,
2753 	stdata_t *stp,
2754 	int32_t flag,
2755 	mblk_t **mpp)
2756 {
2757 	mblk_t *mpctl = NULL;
2758 	mblk_t *mpdata = NULL;
2759 	int error;
2760 
2761 	ASSERT(uiop != NULL);
2762 
2763 	*mpp = NULL;
2764 	/* Create control part, if any */
2765 	if ((mctl != NULL) && (mctl->len >= 0)) {
2766 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2767 		if (error)
2768 			return (error);
2769 	}
2770 	/* Create data part, if any */
2771 	if (*iosize >= 0) {
2772 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2773 		if (error) {
2774 			freemsg(mpctl);
2775 			return (error);
2776 		}
2777 	}
2778 	if (mpctl != NULL) {
2779 		if (mpdata != NULL)
2780 			linkb(mpctl, mpdata);
2781 		*mpp = mpctl;
2782 	} else {
2783 		*mpp = mpdata;
2784 	}
2785 	return (0);
2786 }
2787 
2788 /*
2789  * Make the control part of a protocol message given a control buffer.
2790  * n.b., this can block; be careful of what locks you hold when calling it.
2791  */
2792 int
2793 strmakectl(
2794 	struct strbuf *mctl,
2795 	int32_t flag,
2796 	int32_t fflag,
2797 	mblk_t **mpp)
2798 {
2799 	mblk_t *bp = NULL;
2800 	unsigned char msgtype;
2801 	int error = 0;
2802 	cred_t *cr = CRED();
2803 
2804 	/* We do not support interrupt threads using the stream head to send */
2805 	ASSERT(cr != NULL);
2806 
2807 	*mpp = NULL;
2808 	/*
2809 	 * Create control part of message, if any.
2810 	 */
2811 	if ((mctl != NULL) && (mctl->len >= 0)) {
2812 		caddr_t base;
2813 		int ctlcount;
2814 		int allocsz;
2815 
2816 		if (flag & RS_HIPRI)
2817 			msgtype = M_PCPROTO;
2818 		else
2819 			msgtype = M_PROTO;
2820 
2821 		ctlcount = mctl->len;
2822 		base = mctl->buf;
2823 
2824 		/*
2825 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2826 		 * blocks by increasing the size to something more usable.
2827 		 */
2828 		allocsz = MAX(ctlcount, 64);
2829 
2830 		/*
2831 		 * Range checking has already been done; simply try
2832 		 * to allocate a message block for the ctl part.
2833 		 */
2834 		while ((bp = allocb_cred(allocsz, cr,
2835 		    curproc->p_pid)) == NULL) {
2836 			if (fflag & (FNDELAY|FNONBLOCK))
2837 				return (EAGAIN);
2838 			if (error = strwaitbuf(allocsz, BPRI_MED))
2839 				return (error);
2840 		}
2841 
2842 		bp->b_datap->db_type = msgtype;
2843 		if (copyin(base, bp->b_wptr, ctlcount)) {
2844 			freeb(bp);
2845 			return (EFAULT);
2846 		}
2847 		bp->b_wptr += ctlcount;
2848 	}
2849 	*mpp = bp;
2850 	return (0);
2851 }
2852 
2853 /*
2854  * Make a protocol message given data buffers.
2855  * n.b., this can block; be careful of what locks you hold when calling it.
2856  *
2857  * If sd_maxblk is less than *iosize this routine can fail part way through
2858  * (due to an allocation failure). In this case on return *iosize will contain
2859  * the amount that was consumed. Otherwise *iosize will not be modified
2860  * i.e. it will contain the amount that was consumed.
2861  */
2862 int
2863 strmakedata(
2864 	ssize_t   *iosize,
2865 	struct uio *uiop,
2866 	stdata_t *stp,
2867 	int32_t flag,
2868 	mblk_t **mpp)
2869 {
2870 	mblk_t *mp = NULL;
2871 	mblk_t *bp;
2872 	int wroff = (int)stp->sd_wroff;
2873 	int tail_len = (int)stp->sd_tail;
2874 	int extra = wroff + tail_len;
2875 	int error = 0;
2876 	ssize_t maxblk;
2877 	ssize_t count = *iosize;
2878 	cred_t *cr;
2879 
2880 	*mpp = NULL;
2881 	if (count < 0)
2882 		return (0);
2883 
2884 	/* We do not support interrupt threads using the stream head to send */
2885 	cr = CRED();
2886 	ASSERT(cr != NULL);
2887 
2888 	maxblk = stp->sd_maxblk;
2889 	if (maxblk == INFPSZ)
2890 		maxblk = count;
2891 
2892 	/*
2893 	 * Create data part of message, if any.
2894 	 */
2895 	do {
2896 		ssize_t size;
2897 		dblk_t  *dp;
2898 
2899 		ASSERT(uiop);
2900 
2901 		size = MIN(count, maxblk);
2902 
2903 		while ((bp = allocb_cred(size + extra, cr,
2904 		    curproc->p_pid)) == NULL) {
2905 			error = EAGAIN;
2906 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2907 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2908 				if (count == *iosize) {
2909 					freemsg(mp);
2910 					return (error);
2911 				} else {
2912 					*iosize -= count;
2913 					*mpp = mp;
2914 					return (0);
2915 				}
2916 			}
2917 		}
2918 		dp = bp->b_datap;
2919 		dp->db_cpid = curproc->p_pid;
2920 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2921 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2922 
2923 		if (flag & STRUIO_POSTPONE) {
2924 			/*
2925 			 * Setup the stream uio portion of the
2926 			 * dblk for subsequent use by struioget().
2927 			 */
2928 			dp->db_struioflag = STRUIO_SPEC;
2929 			dp->db_cksumstart = 0;
2930 			dp->db_cksumstuff = 0;
2931 			dp->db_cksumend = size;
2932 			*(long long *)dp->db_struioun.data = 0ll;
2933 			bp->b_wptr += size;
2934 		} else {
2935 			if (stp->sd_copyflag & STRCOPYCACHED)
2936 				uiop->uio_extflg |= UIO_COPY_CACHED;
2937 
2938 			if (size != 0) {
2939 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2940 				    uiop);
2941 				if (error != 0) {
2942 					freeb(bp);
2943 					freemsg(mp);
2944 					return (error);
2945 				}
2946 			}
2947 			bp->b_wptr += size;
2948 
2949 			if (stp->sd_wputdatafunc != NULL) {
2950 				mblk_t *newbp;
2951 
2952 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2953 				    bp, NULL, NULL, NULL, NULL);
2954 				if (newbp == NULL) {
2955 					freeb(bp);
2956 					freemsg(mp);
2957 					return (ECOMM);
2958 				}
2959 				bp = newbp;
2960 			}
2961 		}
2962 
2963 		count -= size;
2964 
2965 		if (mp == NULL)
2966 			mp = bp;
2967 		else
2968 			linkb(mp, bp);
2969 	} while (count > 0);
2970 
2971 	*mpp = mp;
2972 	return (0);
2973 }
2974 
2975 /*
2976  * Wait for a buffer to become available. Return non-zero errno
2977  * if not able to wait, 0 if buffer is probably there.
2978  */
2979 int
2980 strwaitbuf(size_t size, int pri)
2981 {
2982 	bufcall_id_t id;
2983 
2984 	mutex_enter(&bcall_monitor);
2985 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2986 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2987 		mutex_exit(&bcall_monitor);
2988 		return (ENOSR);
2989 	}
2990 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2991 		unbufcall(id);
2992 		mutex_exit(&bcall_monitor);
2993 		return (EINTR);
2994 	}
2995 	unbufcall(id);
2996 	mutex_exit(&bcall_monitor);
2997 	return (0);
2998 }
2999 
3000 /*
3001  * This function waits for a read or write event to happen on a stream.
3002  * fmode can specify FNDELAY and/or FNONBLOCK.
3003  * The timeout is in ms with -1 meaning infinite.
3004  * The flag values work as follows:
3005  *	READWAIT	Check for read side errors, send M_READ
3006  *	GETWAIT		Check for read side errors, no M_READ
3007  *	WRITEWAIT	Check for write side errors.
3008  *	NOINTR		Do not return error if nonblocking or timeout.
3009  * 	STR_NOERROR	Ignore all errors except STPLEX.
3010  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
3011  *	STR_PEEK	Pass through the strgeterr().
3012  */
3013 int
3014 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
3015     int *done)
3016 {
3017 	int slpflg, errs;
3018 	int error;
3019 	kcondvar_t *sleepon;
3020 	mblk_t *mp;
3021 	ssize_t *rd_count;
3022 	clock_t rval;
3023 
3024 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3025 	if ((flag & READWAIT) || (flag & GETWAIT)) {
3026 		slpflg = RSLEEP;
3027 		sleepon = &_RD(stp->sd_wrq)->q_wait;
3028 		errs = STRDERR|STPLEX;
3029 	} else {
3030 		slpflg = WSLEEP;
3031 		sleepon = &stp->sd_wrq->q_wait;
3032 		errs = STWRERR|STRHUP|STPLEX;
3033 	}
3034 	if (flag & STR_NOERROR)
3035 		errs = STPLEX;
3036 
3037 	if (stp->sd_wakeq & slpflg) {
3038 		/*
3039 		 * A strwakeq() is pending, no need to sleep.
3040 		 */
3041 		stp->sd_wakeq &= ~slpflg;
3042 		*done = 0;
3043 		return (0);
3044 	}
3045 
3046 	if (stp->sd_flag & errs) {
3047 		/*
3048 		 * Check for errors before going to sleep since the
3049 		 * caller might not have checked this while holding
3050 		 * sd_lock.
3051 		 */
3052 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3053 		if (error != 0) {
3054 			*done = 1;
3055 			return (error);
3056 		}
3057 	}
3058 
3059 	/*
3060 	 * If any module downstream has requested read notification
3061 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3062 	 * down stream.
3063 	 */
3064 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3065 		mutex_exit(&stp->sd_lock);
3066 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3067 		    (flag & STR_NOSIG), &error))) {
3068 			mutex_enter(&stp->sd_lock);
3069 			*done = 1;
3070 			return (error);
3071 		}
3072 		mp->b_datap->db_type = M_READ;
3073 		rd_count = (ssize_t *)mp->b_wptr;
3074 		*rd_count = count;
3075 		mp->b_wptr += sizeof (ssize_t);
3076 		/*
3077 		 * Send the number of bytes requested by the
3078 		 * read as the argument to M_READ.
3079 		 */
3080 		stream_willservice(stp);
3081 		putnext(stp->sd_wrq, mp);
3082 		stream_runservice(stp);
3083 		mutex_enter(&stp->sd_lock);
3084 
3085 		/*
3086 		 * If any data arrived due to inline processing
3087 		 * of putnext(), don't sleep.
3088 		 */
3089 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3090 			*done = 0;
3091 			return (0);
3092 		}
3093 	}
3094 
3095 	if (fmode & (FNDELAY|FNONBLOCK)) {
3096 		if (!(flag & NOINTR))
3097 			error = EAGAIN;
3098 		else
3099 			error = 0;
3100 		*done = 1;
3101 		return (error);
3102 	}
3103 
3104 	stp->sd_flag |= slpflg;
3105 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3106 	    "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3107 	    stp, flag, count, fmode, done);
3108 
3109 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3110 	if (rval > 0) {
3111 		/* EMPTY */
3112 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3113 		    "strwaitq awakes(2):%X, %X, %X, %X, %X",
3114 		    stp, flag, count, fmode, done);
3115 	} else if (rval == 0) {
3116 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3117 		    "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3118 		    stp, flag, count, fmode, done);
3119 		stp->sd_flag &= ~slpflg;
3120 		cv_broadcast(sleepon);
3121 		if (!(flag & NOINTR))
3122 			error = EINTR;
3123 		else
3124 			error = 0;
3125 		*done = 1;
3126 		return (error);
3127 	} else {
3128 		/* timeout */
3129 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3130 		    "strwaitq timeout:%p, %X, %lX, %X, %p",
3131 		    stp, flag, count, fmode, done);
3132 		*done = 1;
3133 		if (!(flag & NOINTR))
3134 			return (ETIME);
3135 		else
3136 			return (0);
3137 	}
3138 	/*
3139 	 * If the caller implements delayed errors (i.e. queued after data)
3140 	 * we can not check for errors here since data as well as an
3141 	 * error might have arrived at the stream head. We return to
3142 	 * have the caller check the read queue before checking for errors.
3143 	 */
3144 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3145 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3146 		if (error != 0) {
3147 			*done = 1;
3148 			return (error);
3149 		}
3150 	}
3151 	*done = 0;
3152 	return (0);
3153 }
3154 
3155 /*
3156  * Perform job control discipline access checks.
3157  * Return 0 for success and the errno for failure.
3158  */
3159 
3160 #define	cantsend(p, t, sig) \
3161 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3162 
3163 int
3164 straccess(struct stdata *stp, enum jcaccess mode)
3165 {
3166 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3167 	kthread_t *t = curthread;
3168 	proc_t *p = ttoproc(t);
3169 	sess_t *sp;
3170 
3171 	ASSERT(mutex_owned(&stp->sd_lock));
3172 
3173 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3174 		return (0);
3175 
3176 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3177 
3178 	for (;;) {
3179 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3180 		sp = p->p_sessp;
3181 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3182 
3183 		/*
3184 		 * If this is not the calling process's controlling terminal
3185 		 * or if the calling process is already in the foreground
3186 		 * then allow access.
3187 		 */
3188 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3189 		    p->p_pgidp == stp->sd_pgidp) {
3190 			mutex_exit(&sp->s_lock);
3191 			mutex_exit(&p->p_splock);
3192 			mutex_exit(&p->p_lock);
3193 			return (0);
3194 		}
3195 
3196 		/*
3197 		 * Check to see if controlling terminal has been deallocated.
3198 		 */
3199 		if (sp->s_vp == NULL) {
3200 			if (!cantsend(p, t, SIGHUP))
3201 				sigtoproc(p, t, SIGHUP);
3202 			mutex_exit(&sp->s_lock);
3203 			mutex_exit(&p->p_splock);
3204 			mutex_exit(&p->p_lock);
3205 			return (EIO);
3206 		}
3207 
3208 		mutex_exit(&sp->s_lock);
3209 		mutex_exit(&p->p_splock);
3210 
3211 		if (mode == JCGETP) {
3212 			mutex_exit(&p->p_lock);
3213 			return (0);
3214 		}
3215 
3216 		if (mode == JCREAD) {
3217 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3218 				mutex_exit(&p->p_lock);
3219 				return (EIO);
3220 			}
3221 			mutex_exit(&p->p_lock);
3222 			mutex_exit(&stp->sd_lock);
3223 			pgsignal(p->p_pgidp, SIGTTIN);
3224 			mutex_enter(&stp->sd_lock);
3225 			mutex_enter(&p->p_lock);
3226 		} else {  /* mode == JCWRITE or JCSETP */
3227 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3228 			    cantsend(p, t, SIGTTOU)) {
3229 				mutex_exit(&p->p_lock);
3230 				return (0);
3231 			}
3232 			if (p->p_detached) {
3233 				mutex_exit(&p->p_lock);
3234 				return (EIO);
3235 			}
3236 			mutex_exit(&p->p_lock);
3237 			mutex_exit(&stp->sd_lock);
3238 			pgsignal(p->p_pgidp, SIGTTOU);
3239 			mutex_enter(&stp->sd_lock);
3240 			mutex_enter(&p->p_lock);
3241 		}
3242 
3243 		/*
3244 		 * We call cv_wait_sig_swap() to cause the appropriate
3245 		 * action for the jobcontrol signal to take place.
3246 		 * If the signal is being caught, we will take the
3247 		 * EINTR error return.  Otherwise, the default action
3248 		 * of causing the process to stop will take place.
3249 		 * In this case, we rely on the periodic cv_broadcast() on
3250 		 * &lbolt_cv to wake us up to loop around and test again.
3251 		 * We can't get here if the signal is ignored or
3252 		 * if the current thread is blocking the signal.
3253 		 */
3254 		mutex_exit(&stp->sd_lock);
3255 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3256 			mutex_exit(&p->p_lock);
3257 			mutex_enter(&stp->sd_lock);
3258 			return (EINTR);
3259 		}
3260 		mutex_exit(&p->p_lock);
3261 		mutex_enter(&stp->sd_lock);
3262 		mutex_enter(&p->p_lock);
3263 	}
3264 }
3265 
3266 /*
3267  * Return size of message of block type (bp->b_datap->db_type)
3268  */
3269 size_t
3270 xmsgsize(mblk_t *bp)
3271 {
3272 	unsigned char type;
3273 	size_t count = 0;
3274 
3275 	type = bp->b_datap->db_type;
3276 
3277 	for (; bp; bp = bp->b_cont) {
3278 		if (type != bp->b_datap->db_type)
3279 			break;
3280 		ASSERT(bp->b_wptr >= bp->b_rptr);
3281 		count += bp->b_wptr - bp->b_rptr;
3282 	}
3283 	return (count);
3284 }
3285 
3286 /*
3287  * Allocate a stream head.
3288  */
3289 struct stdata *
3290 shalloc(queue_t *qp)
3291 {
3292 	stdata_t *stp;
3293 
3294 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3295 
3296 	stp->sd_wrq = _WR(qp);
3297 	stp->sd_strtab = NULL;
3298 	stp->sd_iocid = 0;
3299 	stp->sd_mate = NULL;
3300 	stp->sd_freezer = NULL;
3301 	stp->sd_refcnt = 0;
3302 	stp->sd_wakeq = 0;
3303 	stp->sd_anchor = 0;
3304 	stp->sd_struiowrq = NULL;
3305 	stp->sd_struiordq = NULL;
3306 	stp->sd_struiodnak = 0;
3307 	stp->sd_struionak = NULL;
3308 	stp->sd_t_audit_data = NULL;
3309 	stp->sd_rput_opt = 0;
3310 	stp->sd_wput_opt = 0;
3311 	stp->sd_read_opt = 0;
3312 	stp->sd_rprotofunc = strrput_proto;
3313 	stp->sd_rmiscfunc = strrput_misc;
3314 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3315 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3316 	stp->sd_ciputctrl = NULL;
3317 	stp->sd_nciputctrl = 0;
3318 	stp->sd_qhead = NULL;
3319 	stp->sd_qtail = NULL;
3320 	stp->sd_servid = NULL;
3321 	stp->sd_nqueues = 0;
3322 	stp->sd_svcflags = 0;
3323 	stp->sd_copyflag = 0;
3324 
3325 	return (stp);
3326 }
3327 
3328 /*
3329  * Free a stream head.
3330  */
3331 void
3332 shfree(stdata_t *stp)
3333 {
3334 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3335 
3336 	stp->sd_wrq = NULL;
3337 
3338 	mutex_enter(&stp->sd_qlock);
3339 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3340 		STRSTAT(strwaits);
3341 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3342 	}
3343 	mutex_exit(&stp->sd_qlock);
3344 
3345 	if (stp->sd_ciputctrl != NULL) {
3346 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3347 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3348 		    stp->sd_nciputctrl, 0);
3349 		ASSERT(ciputctrl_cache != NULL);
3350 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3351 		stp->sd_ciputctrl = NULL;
3352 		stp->sd_nciputctrl = 0;
3353 	}
3354 	ASSERT(stp->sd_qhead == NULL);
3355 	ASSERT(stp->sd_qtail == NULL);
3356 	ASSERT(stp->sd_nqueues == 0);
3357 	kmem_cache_free(stream_head_cache, stp);
3358 }
3359 
3360 /*
3361  * Allocate a pair of queues and a syncq for the pair
3362  */
3363 queue_t *
3364 allocq(void)
3365 {
3366 	queinfo_t *qip;
3367 	queue_t *qp, *wqp;
3368 	syncq_t	*sq;
3369 
3370 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3371 
3372 	qp = &qip->qu_rqueue;
3373 	wqp = &qip->qu_wqueue;
3374 	sq = &qip->qu_syncq;
3375 
3376 	qp->q_last	= NULL;
3377 	qp->q_next	= NULL;
3378 	qp->q_ptr	= NULL;
3379 	qp->q_flag	= QUSE | QREADR;
3380 	qp->q_bandp	= NULL;
3381 	qp->q_stream	= NULL;
3382 	qp->q_syncq	= sq;
3383 	qp->q_nband	= 0;
3384 	qp->q_nfsrv	= NULL;
3385 	qp->q_draining	= 0;
3386 	qp->q_syncqmsgs	= 0;
3387 	qp->q_spri	= 0;
3388 	qp->q_qtstamp	= 0;
3389 	qp->q_sqtstamp	= 0;
3390 	qp->q_fp	= NULL;
3391 
3392 	wqp->q_last	= NULL;
3393 	wqp->q_next	= NULL;
3394 	wqp->q_ptr	= NULL;
3395 	wqp->q_flag	= QUSE;
3396 	wqp->q_bandp	= NULL;
3397 	wqp->q_stream	= NULL;
3398 	wqp->q_syncq	= sq;
3399 	wqp->q_nband	= 0;
3400 	wqp->q_nfsrv	= NULL;
3401 	wqp->q_draining	= 0;
3402 	wqp->q_syncqmsgs = 0;
3403 	wqp->q_qtstamp	= 0;
3404 	wqp->q_sqtstamp	= 0;
3405 	wqp->q_spri	= 0;
3406 
3407 	sq->sq_count	= 0;
3408 	sq->sq_rmqcount	= 0;
3409 	sq->sq_flags	= 0;
3410 	sq->sq_type	= 0;
3411 	sq->sq_callbflags = 0;
3412 	sq->sq_cancelid	= 0;
3413 	sq->sq_ciputctrl = NULL;
3414 	sq->sq_nciputctrl = 0;
3415 	sq->sq_needexcl = 0;
3416 	sq->sq_svcflags = 0;
3417 
3418 	return (qp);
3419 }
3420 
3421 /*
3422  * Free a pair of queues and the "attached" syncq.
3423  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3424  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3425  */
3426 void
3427 freeq(queue_t *qp)
3428 {
3429 	qband_t *qbp, *nqbp;
3430 	syncq_t *sq, *outer;
3431 	queue_t *wqp = _WR(qp);
3432 
3433 	ASSERT(qp->q_flag & QREADR);
3434 
3435 	/*
3436 	 * If a previously dispatched taskq job is scheduled to run
3437 	 * sync_service() or a service routine is scheduled for the
3438 	 * queues about to be freed, wait here until all service is
3439 	 * done on the queue and all associated queues and syncqs.
3440 	 */
3441 	wait_svc(qp);
3442 
3443 	(void) flush_syncq(qp->q_syncq, qp);
3444 	(void) flush_syncq(wqp->q_syncq, wqp);
3445 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3446 
3447 	/*
3448 	 * Flush the queues before q_next is set to NULL This is needed
3449 	 * in order to backenable any downstream queue before we go away.
3450 	 * Note: we are already removed from the stream so that the
3451 	 * backenabling will not cause any messages to be delivered to our
3452 	 * put procedures.
3453 	 */
3454 	flushq(qp, FLUSHALL);
3455 	flushq(wqp, FLUSHALL);
3456 
3457 	/* Tidy up - removeq only does a half-remove from stream */
3458 	qp->q_next = wqp->q_next = NULL;
3459 	ASSERT(!(qp->q_flag & QENAB));
3460 	ASSERT(!(wqp->q_flag & QENAB));
3461 
3462 	outer = qp->q_syncq->sq_outer;
3463 	if (outer != NULL) {
3464 		outer_remove(outer, qp->q_syncq);
3465 		if (wqp->q_syncq != qp->q_syncq)
3466 			outer_remove(outer, wqp->q_syncq);
3467 	}
3468 	/*
3469 	 * Free any syncqs that are outside what allocq returned.
3470 	 */
3471 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3472 		free_syncq(qp->q_syncq);
3473 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3474 		free_syncq(wqp->q_syncq);
3475 
3476 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3477 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3478 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3479 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3480 	sq = SQ(qp);
3481 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3482 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3483 	ASSERT(sq->sq_outer == NULL);
3484 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3485 	ASSERT(sq->sq_callbpend == NULL);
3486 	ASSERT(sq->sq_needexcl == 0);
3487 
3488 	if (sq->sq_ciputctrl != NULL) {
3489 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3490 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3491 		    sq->sq_nciputctrl, 0);
3492 		ASSERT(ciputctrl_cache != NULL);
3493 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3494 		sq->sq_ciputctrl = NULL;
3495 		sq->sq_nciputctrl = 0;
3496 	}
3497 
3498 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3499 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3500 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3501 
3502 	qp->q_flag &= ~QUSE;
3503 	wqp->q_flag &= ~QUSE;
3504 
3505 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3506 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3507 
3508 	qbp = qp->q_bandp;
3509 	while (qbp) {
3510 		nqbp = qbp->qb_next;
3511 		freeband(qbp);
3512 		qbp = nqbp;
3513 	}
3514 	qbp = wqp->q_bandp;
3515 	while (qbp) {
3516 		nqbp = qbp->qb_next;
3517 		freeband(qbp);
3518 		qbp = nqbp;
3519 	}
3520 	kmem_cache_free(queue_cache, qp);
3521 }
3522 
3523 /*
3524  * Allocate a qband structure.
3525  */
3526 qband_t *
3527 allocband(void)
3528 {
3529 	qband_t *qbp;
3530 
3531 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3532 	if (qbp == NULL)
3533 		return (NULL);
3534 
3535 	qbp->qb_next	= NULL;
3536 	qbp->qb_count	= 0;
3537 	qbp->qb_mblkcnt	= 0;
3538 	qbp->qb_first	= NULL;
3539 	qbp->qb_last	= NULL;
3540 	qbp->qb_flag	= 0;
3541 
3542 	return (qbp);
3543 }
3544 
3545 /*
3546  * Free a qband structure.
3547  */
3548 void
3549 freeband(qband_t *qbp)
3550 {
3551 	kmem_cache_free(qband_cache, qbp);
3552 }
3553 
3554 /*
3555  * Just like putnextctl(9F), except that allocb_wait() is used.
3556  *
3557  * Consolidation Private, and of course only callable from the stream head or
3558  * routines that may block.
3559  */
3560 int
3561 putnextctl_wait(queue_t *q, int type)
3562 {
3563 	mblk_t *bp;
3564 	int error;
3565 
3566 	if ((datamsg(type) && (type != M_DELAY)) ||
3567 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3568 		return (0);
3569 
3570 	bp->b_datap->db_type = (unsigned char)type;
3571 	putnext(q, bp);
3572 	return (1);
3573 }
3574 
3575 /*
3576  * Run any possible bufcalls.
3577  */
3578 void
3579 runbufcalls(void)
3580 {
3581 	strbufcall_t *bcp;
3582 
3583 	mutex_enter(&bcall_monitor);
3584 	mutex_enter(&strbcall_lock);
3585 
3586 	if (strbcalls.bc_head) {
3587 		size_t count;
3588 		int nevent;
3589 
3590 		/*
3591 		 * count how many events are on the list
3592 		 * now so we can check to avoid looping
3593 		 * in low memory situations
3594 		 */
3595 		nevent = 0;
3596 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3597 			nevent++;
3598 
3599 		/*
3600 		 * get estimate of available memory from kmem_avail().
3601 		 * awake all bufcall functions waiting for
3602 		 * memory whose request could be satisfied
3603 		 * by 'count' memory and let 'em fight for it.
3604 		 */
3605 		count = kmem_avail();
3606 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3607 			STRSTAT(bufcalls);
3608 			--nevent;
3609 			if (bcp->bc_size <= count) {
3610 				bcp->bc_executor = curthread;
3611 				mutex_exit(&strbcall_lock);
3612 				(*bcp->bc_func)(bcp->bc_arg);
3613 				mutex_enter(&strbcall_lock);
3614 				bcp->bc_executor = NULL;
3615 				cv_broadcast(&bcall_cv);
3616 				strbcalls.bc_head = bcp->bc_next;
3617 				kmem_free(bcp, sizeof (strbufcall_t));
3618 			} else {
3619 				/*
3620 				 * too big, try again later - note
3621 				 * that nevent was decremented above
3622 				 * so we won't retry this one on this
3623 				 * iteration of the loop
3624 				 */
3625 				if (bcp->bc_next != NULL) {
3626 					strbcalls.bc_head = bcp->bc_next;
3627 					bcp->bc_next = NULL;
3628 					strbcalls.bc_tail->bc_next = bcp;
3629 					strbcalls.bc_tail = bcp;
3630 				}
3631 			}
3632 		}
3633 		if (strbcalls.bc_head == NULL)
3634 			strbcalls.bc_tail = NULL;
3635 	}
3636 
3637 	mutex_exit(&strbcall_lock);
3638 	mutex_exit(&bcall_monitor);
3639 }
3640 
3641 
3642 /*
3643  * Actually run queue's service routine.
3644  */
3645 static void
3646 runservice(queue_t *q)
3647 {
3648 	qband_t *qbp;
3649 
3650 	ASSERT(q->q_qinfo->qi_srvp);
3651 again:
3652 	entersq(q->q_syncq, SQ_SVC);
3653 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3654 	    "runservice starts:%p", q);
3655 
3656 	if (!(q->q_flag & QWCLOSE))
3657 		(*q->q_qinfo->qi_srvp)(q);
3658 
3659 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3660 	    "runservice ends:(%p)", q);
3661 
3662 	leavesq(q->q_syncq, SQ_SVC);
3663 
3664 	mutex_enter(QLOCK(q));
3665 	if (q->q_flag & QENAB) {
3666 		q->q_flag &= ~QENAB;
3667 		mutex_exit(QLOCK(q));
3668 		goto again;
3669 	}
3670 	q->q_flag &= ~QINSERVICE;
3671 	q->q_flag &= ~QBACK;
3672 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3673 		qbp->qb_flag &= ~QB_BACK;
3674 	/*
3675 	 * Wakeup thread waiting for the service procedure
3676 	 * to be run (strclose and qdetach).
3677 	 */
3678 	cv_broadcast(&q->q_wait);
3679 
3680 	mutex_exit(QLOCK(q));
3681 }
3682 
3683 /*
3684  * Background processing of bufcalls.
3685  */
3686 void
3687 streams_bufcall_service(void)
3688 {
3689 	callb_cpr_t	cprinfo;
3690 
3691 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3692 	    "streams_bufcall_service");
3693 
3694 	mutex_enter(&strbcall_lock);
3695 
3696 	for (;;) {
3697 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3698 			mutex_exit(&strbcall_lock);
3699 			runbufcalls();
3700 			mutex_enter(&strbcall_lock);
3701 		}
3702 		if (strbcalls.bc_head != NULL) {
3703 			STRSTAT(bcwaits);
3704 			/* Wait for memory to become available */
3705 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3706 			(void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3707 			    SEC_TO_TICK(60), TR_CLOCK_TICK);
3708 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3709 		}
3710 
3711 		/* Wait for new work to arrive */
3712 		if (strbcalls.bc_head == NULL) {
3713 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3714 			cv_wait(&strbcall_cv, &strbcall_lock);
3715 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3716 		}
3717 	}
3718 }
3719 
3720 /*
3721  * Background processing of streams background tasks which failed
3722  * taskq_dispatch.
3723  */
3724 static void
3725 streams_qbkgrnd_service(void)
3726 {
3727 	callb_cpr_t cprinfo;
3728 	queue_t *q;
3729 
3730 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3731 	    "streams_bkgrnd_service");
3732 
3733 	mutex_enter(&service_queue);
3734 
3735 	for (;;) {
3736 		/*
3737 		 * Wait for work to arrive.
3738 		 */
3739 		while ((freebs_list == NULL) && (qhead == NULL)) {
3740 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3741 			cv_wait(&services_to_run, &service_queue);
3742 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3743 		}
3744 		/*
3745 		 * Handle all pending freebs requests to free memory.
3746 		 */
3747 		while (freebs_list != NULL) {
3748 			mblk_t *mp = freebs_list;
3749 			freebs_list = mp->b_next;
3750 			mutex_exit(&service_queue);
3751 			mblk_free(mp);
3752 			mutex_enter(&service_queue);
3753 		}
3754 		/*
3755 		 * Run pending queues.
3756 		 */
3757 		while (qhead != NULL) {
3758 			DQ(q, qhead, qtail, q_link);
3759 			ASSERT(q != NULL);
3760 			mutex_exit(&service_queue);
3761 			queue_service(q);
3762 			mutex_enter(&service_queue);
3763 		}
3764 		ASSERT(qhead == NULL && qtail == NULL);
3765 	}
3766 }
3767 
3768 /*
3769  * Background processing of streams background tasks which failed
3770  * taskq_dispatch.
3771  */
3772 static void
3773 streams_sqbkgrnd_service(void)
3774 {
3775 	callb_cpr_t cprinfo;
3776 	syncq_t *sq;
3777 
3778 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3779 	    "streams_sqbkgrnd_service");
3780 
3781 	mutex_enter(&service_queue);
3782 
3783 	for (;;) {
3784 		/*
3785 		 * Wait for work to arrive.
3786 		 */
3787 		while (sqhead == NULL) {
3788 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3789 			cv_wait(&syncqs_to_run, &service_queue);
3790 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3791 		}
3792 
3793 		/*
3794 		 * Run pending syncqs.
3795 		 */
3796 		while (sqhead != NULL) {
3797 			DQ(sq, sqhead, sqtail, sq_next);
3798 			ASSERT(sq != NULL);
3799 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3800 			mutex_exit(&service_queue);
3801 			syncq_service(sq);
3802 			mutex_enter(&service_queue);
3803 		}
3804 	}
3805 }
3806 
3807 /*
3808  * Disable the syncq and wait for background syncq processing to complete.
3809  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3810  * list.
3811  */
3812 void
3813 wait_sq_svc(syncq_t *sq)
3814 {
3815 	mutex_enter(SQLOCK(sq));
3816 	sq->sq_svcflags |= SQ_DISABLED;
3817 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3818 		syncq_t *sq_chase;
3819 		syncq_t *sq_curr;
3820 		int removed;
3821 
3822 		ASSERT(sq->sq_servcount == 1);
3823 		mutex_enter(&service_queue);
3824 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3825 		mutex_exit(&service_queue);
3826 		if (removed) {
3827 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3828 			sq->sq_servcount = 0;
3829 			STRSTAT(sqremoved);
3830 			goto done;
3831 		}
3832 	}
3833 	while (sq->sq_servcount != 0) {
3834 		sq->sq_flags |= SQ_WANTWAKEUP;
3835 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3836 	}
3837 done:
3838 	mutex_exit(SQLOCK(sq));
3839 }
3840 
3841 /*
3842  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3843  * Add the argument to the end of the sqhead list and set the flag
3844  * indicating this syncq has been enabled.  If it has already been
3845  * enabled, don't do anything.
3846  * This routine assumes that SQLOCK is held.
3847  * NOTE that the lock order is to have the SQLOCK first,
3848  * so if the service_syncq lock is held, we need to release it
3849  * before acquiring the SQLOCK (mostly relevant for the background
3850  * thread, and this seems to be common among the STREAMS global locks).
3851  * Note that the sq_svcflags are protected by the SQLOCK.
3852  */
3853 void
3854 sqenable(syncq_t *sq)
3855 {
3856 	/*
3857 	 * This is probably not important except for where I believe it
3858 	 * is being called.  At that point, it should be held (and it
3859 	 * is a pain to release it just for this routine, so don't do
3860 	 * it).
3861 	 */
3862 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3863 
3864 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3865 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3866 
3867 	/*
3868 	 * Do not put on list if background thread is scheduled or
3869 	 * syncq is disabled.
3870 	 */
3871 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3872 		return;
3873 
3874 	/*
3875 	 * Check whether we should enable sq at all.
3876 	 * Non PERMOD syncqs may be drained by at most one thread.
3877 	 * PERMOD syncqs may be drained by several threads but we limit the
3878 	 * total amount to the lesser of
3879 	 *	Number of queues on the squeue and
3880 	 *	Number of CPUs.
3881 	 */
3882 	if (sq->sq_servcount != 0) {
3883 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3884 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3885 			STRSTAT(sqtoomany);
3886 			return;
3887 		}
3888 	}
3889 
3890 	sq->sq_tstamp = ddi_get_lbolt();
3891 	STRSTAT(sqenables);
3892 
3893 	/* Attempt a taskq dispatch */
3894 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3895 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3896 	if (sq->sq_servid != NULL) {
3897 		sq->sq_servcount++;
3898 		return;
3899 	}
3900 
3901 	/*
3902 	 * This taskq dispatch failed, but a previous one may have succeeded.
3903 	 * Don't try to schedule on the background thread whilst there is
3904 	 * outstanding taskq processing.
3905 	 */
3906 	if (sq->sq_servcount != 0)
3907 		return;
3908 
3909 	/*
3910 	 * System is low on resources and can't perform a non-sleeping
3911 	 * dispatch. Schedule the syncq for a background thread and mark the
3912 	 * syncq to avoid any further taskq dispatch attempts.
3913 	 */
3914 	mutex_enter(&service_queue);
3915 	STRSTAT(taskqfails);
3916 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3917 	sq->sq_svcflags |= SQ_BGTHREAD;
3918 	sq->sq_servcount = 1;
3919 	cv_signal(&syncqs_to_run);
3920 	mutex_exit(&service_queue);
3921 }
3922 
3923 /*
3924  * Note: fifo_close() depends on the mblk_t on the queue being freed
3925  * asynchronously. The asynchronous freeing of messages breaks the
3926  * recursive call chain of fifo_close() while there are I_SENDFD type of
3927  * messages referring to other file pointers on the queue. Then when
3928  * closing pipes it can avoid stack overflow in case of daisy-chained
3929  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3930  * share the same fifolock_t).
3931  *
3932  * No need to kpreempt_disable to access cpu_seqid.  If we migrate and
3933  * the esb queue does not match the new CPU, that is OK.
3934  */
3935 void
3936 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3937 {
3938 	int qindex = CPU->cpu_seqid >> esbq_log2_cpus_per_q;
3939 	esb_queue_t *eqp;
3940 
3941 	ASSERT(dbp->db_mblk == mp);
3942 	ASSERT(qindex < esbq_nelem);
3943 
3944 	eqp = system_esbq_array;
3945 	if (eqp != NULL) {
3946 		eqp += qindex;
3947 	} else {
3948 		mutex_enter(&esbq_lock);
3949 		if (kmem_ready && system_esbq_array == NULL)
3950 			system_esbq_array = (esb_queue_t *)kmem_zalloc(
3951 			    esbq_nelem * sizeof (esb_queue_t), KM_NOSLEEP);
3952 		mutex_exit(&esbq_lock);
3953 		eqp = system_esbq_array;
3954 		if (eqp != NULL)
3955 			eqp += qindex;
3956 		else
3957 			eqp = &system_esbq;
3958 	}
3959 
3960 	/*
3961 	 * Check data sanity. The dblock should have non-empty free function.
3962 	 * It is better to panic here then later when the dblock is freed
3963 	 * asynchronously when the context is lost.
3964 	 */
3965 	if (dbp->db_frtnp->free_func == NULL) {
3966 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3967 		    (void *)dbp);
3968 	}
3969 
3970 	mutex_enter(&eqp->eq_lock);
3971 	/* queue the new mblk on the esballoc queue */
3972 	if (eqp->eq_head == NULL) {
3973 		eqp->eq_head = eqp->eq_tail = mp;
3974 	} else {
3975 		eqp->eq_tail->b_next = mp;
3976 		eqp->eq_tail = mp;
3977 	}
3978 	eqp->eq_len++;
3979 
3980 	/* If we're the first thread to reach the threshold, process */
3981 	if (eqp->eq_len >= esbq_max_qlen &&
3982 	    !(eqp->eq_flags & ESBQ_PROCESSING))
3983 		esballoc_process_queue(eqp);
3984 
3985 	esballoc_set_timer(eqp, esbq_timeout);
3986 	mutex_exit(&eqp->eq_lock);
3987 }
3988 
3989 static void
3990 esballoc_process_queue(esb_queue_t *eqp)
3991 {
3992 	mblk_t	*mp;
3993 
3994 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
3995 
3996 	eqp->eq_flags |= ESBQ_PROCESSING;
3997 
3998 	do {
3999 		/*
4000 		 * Detach the message chain for processing.
4001 		 */
4002 		mp = eqp->eq_head;
4003 		eqp->eq_tail->b_next = NULL;
4004 		eqp->eq_head = eqp->eq_tail = NULL;
4005 		eqp->eq_len = 0;
4006 		mutex_exit(&eqp->eq_lock);
4007 
4008 		/*
4009 		 * Process the message chain.
4010 		 */
4011 		esballoc_enqueue_mblk(mp);
4012 		mutex_enter(&eqp->eq_lock);
4013 	} while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
4014 
4015 	eqp->eq_flags &= ~ESBQ_PROCESSING;
4016 }
4017 
4018 /*
4019  * taskq callback routine to free esballoced mblk's
4020  */
4021 static void
4022 esballoc_mblk_free(mblk_t *mp)
4023 {
4024 	mblk_t	*nextmp;
4025 
4026 	for (; mp != NULL; mp = nextmp) {
4027 		nextmp = mp->b_next;
4028 		mp->b_next = NULL;
4029 		mblk_free(mp);
4030 	}
4031 }
4032 
4033 static void
4034 esballoc_enqueue_mblk(mblk_t *mp)
4035 {
4036 
4037 	if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
4038 	    TQ_NOSLEEP) == TASKQID_INVALID) {
4039 		mblk_t *first_mp = mp;
4040 		/*
4041 		 * System is low on resources and can't perform a non-sleeping
4042 		 * dispatch. Schedule for a background thread.
4043 		 */
4044 		mutex_enter(&service_queue);
4045 		STRSTAT(taskqfails);
4046 
4047 		while (mp->b_next != NULL)
4048 			mp = mp->b_next;
4049 
4050 		mp->b_next = freebs_list;
4051 		freebs_list = first_mp;
4052 		cv_signal(&services_to_run);
4053 		mutex_exit(&service_queue);
4054 	}
4055 }
4056 
4057 static void
4058 esballoc_timer(void *arg)
4059 {
4060 	esb_queue_t *eqp = arg;
4061 
4062 	mutex_enter(&eqp->eq_lock);
4063 	eqp->eq_flags &= ~ESBQ_TIMER;
4064 
4065 	if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4066 	    eqp->eq_len > 0)
4067 		esballoc_process_queue(eqp);
4068 
4069 	esballoc_set_timer(eqp, esbq_timeout);
4070 	mutex_exit(&eqp->eq_lock);
4071 }
4072 
4073 static void
4074 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4075 {
4076 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4077 
4078 	if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4079 		(void) timeout(esballoc_timer, eqp, eq_timeout);
4080 		eqp->eq_flags |= ESBQ_TIMER;
4081 	}
4082 }
4083 
4084 /*
4085  * Setup esbq array length based upon NCPU scaled by CPUs per
4086  * queue. Use static system_esbq until kmem_ready and we can
4087  * create an array in freebs_enqueue().
4088  */
4089 void
4090 esballoc_queue_init(void)
4091 {
4092 	esbq_log2_cpus_per_q = highbit(esbq_cpus_per_q - 1);
4093 	esbq_cpus_per_q = 1 << esbq_log2_cpus_per_q;
4094 	esbq_nelem = howmany(NCPU, esbq_cpus_per_q);
4095 	system_esbq.eq_len = 0;
4096 	system_esbq.eq_head = system_esbq.eq_tail = NULL;
4097 	system_esbq.eq_flags = 0;
4098 }
4099 
4100 /*
4101  * Set the QBACK or QB_BACK flag in the given queue for
4102  * the given priority band.
4103  */
4104 void
4105 setqback(queue_t *q, unsigned char pri)
4106 {
4107 	int i;
4108 	qband_t *qbp;
4109 	qband_t **qbpp;
4110 
4111 	ASSERT(MUTEX_HELD(QLOCK(q)));
4112 	if (pri != 0) {
4113 		if (pri > q->q_nband) {
4114 			qbpp = &q->q_bandp;
4115 			while (*qbpp)
4116 				qbpp = &(*qbpp)->qb_next;
4117 			while (pri > q->q_nband) {
4118 				if ((*qbpp = allocband()) == NULL) {
4119 					cmn_err(CE_WARN,
4120 					    "setqback: can't allocate qband\n");
4121 					return;
4122 				}
4123 				(*qbpp)->qb_hiwat = q->q_hiwat;
4124 				(*qbpp)->qb_lowat = q->q_lowat;
4125 				q->q_nband++;
4126 				qbpp = &(*qbpp)->qb_next;
4127 			}
4128 		}
4129 		qbp = q->q_bandp;
4130 		i = pri;
4131 		while (--i)
4132 			qbp = qbp->qb_next;
4133 		qbp->qb_flag |= QB_BACK;
4134 	} else {
4135 		q->q_flag |= QBACK;
4136 	}
4137 }
4138 
4139 int
4140 strcopyin(void *from, void *to, size_t len, int copyflag)
4141 {
4142 	if (copyflag & U_TO_K) {
4143 		ASSERT((copyflag & K_TO_K) == 0);
4144 		if (copyin(from, to, len))
4145 			return (EFAULT);
4146 	} else {
4147 		ASSERT(copyflag & K_TO_K);
4148 		bcopy(from, to, len);
4149 	}
4150 	return (0);
4151 }
4152 
4153 int
4154 strcopyout(void *from, void *to, size_t len, int copyflag)
4155 {
4156 	if (copyflag & U_TO_K) {
4157 		if (copyout(from, to, len))
4158 			return (EFAULT);
4159 	} else {
4160 		ASSERT(copyflag & K_TO_K);
4161 		bcopy(from, to, len);
4162 	}
4163 	return (0);
4164 }
4165 
4166 /*
4167  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4168  * It assumes that the stream head lock is already held, whereas strsignal()
4169  * acquires the lock first.  This routine was created because a few callers
4170  * release the stream head lock before calling only to re-acquire it after
4171  * it returns.
4172  */
4173 void
4174 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4175 {
4176 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4177 	switch (sig) {
4178 	case SIGPOLL:
4179 		if (stp->sd_sigflags & S_MSG)
4180 			strsendsig(stp->sd_siglist, S_MSG, band, 0);
4181 		break;
4182 	default:
4183 		if (stp->sd_pgidp)
4184 			pgsignal(stp->sd_pgidp, sig);
4185 		break;
4186 	}
4187 }
4188 
4189 void
4190 strsignal(stdata_t *stp, int sig, int32_t band)
4191 {
4192 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4193 	    "strsignal:%p, %X, %X", stp, sig, band);
4194 
4195 	mutex_enter(&stp->sd_lock);
4196 	switch (sig) {
4197 	case SIGPOLL:
4198 		if (stp->sd_sigflags & S_MSG)
4199 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4200 		break;
4201 
4202 	default:
4203 		if (stp->sd_pgidp) {
4204 			pgsignal(stp->sd_pgidp, sig);
4205 		}
4206 		break;
4207 	}
4208 	mutex_exit(&stp->sd_lock);
4209 }
4210 
4211 void
4212 strhup(stdata_t *stp)
4213 {
4214 	ASSERT(mutex_owned(&stp->sd_lock));
4215 	pollwakeup(&stp->sd_pollist, POLLHUP);
4216 	if (stp->sd_sigflags & S_HANGUP)
4217 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4218 }
4219 
4220 /*
4221  * Backenable the first queue upstream from `q' with a service procedure.
4222  */
4223 void
4224 backenable(queue_t *q, uchar_t pri)
4225 {
4226 	queue_t	*nq;
4227 
4228 	/*
4229 	 * Our presence might not prevent other modules in our own
4230 	 * stream from popping/pushing since the caller of getq might not
4231 	 * have a claim on the queue (some drivers do a getq on somebody
4232 	 * else's queue - they know that the queue itself is not going away
4233 	 * but the framework has to guarantee q_next in that stream).
4234 	 */
4235 	claimstr(q);
4236 
4237 	/* Find nearest back queue with service proc */
4238 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4239 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4240 	}
4241 
4242 	if (nq) {
4243 		kthread_t *freezer;
4244 		/*
4245 		 * backenable can be called either with no locks held
4246 		 * or with the stream frozen (the latter occurs when a module
4247 		 * calls rmvq with the stream frozen). If the stream is frozen
4248 		 * by the caller the caller will hold all qlocks in the stream.
4249 		 * Note that a frozen stream doesn't freeze a mated stream,
4250 		 * so we explicitly check for that.
4251 		 */
4252 		freezer = STREAM(q)->sd_freezer;
4253 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4254 			mutex_enter(QLOCK(nq));
4255 		}
4256 #ifdef DEBUG
4257 		else {
4258 			ASSERT(frozenstr(q));
4259 			ASSERT(MUTEX_HELD(QLOCK(q)));
4260 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4261 		}
4262 #endif
4263 		setqback(nq, pri);
4264 		qenable_locked(nq);
4265 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4266 			mutex_exit(QLOCK(nq));
4267 	}
4268 	releasestr(q);
4269 }
4270 
4271 /*
4272  * Return the appropriate errno when one of flags_to_check is set
4273  * in sd_flags. Uses the exported error routines if they are set.
4274  * Will return 0 if non error is set (or if the exported error routines
4275  * do not return an error).
4276  *
4277  * If there is both a read and write error to check, we prefer the read error.
4278  * Also, give preference to recorded errno's over the error functions.
4279  * The flags that are handled are:
4280  *	STPLEX		return EINVAL
4281  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4282  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4283  *	STRHUP		return sd_werror
4284  *
4285  * If the caller indicates that the operation is a peek, a nonpersistent error
4286  * is not cleared.
4287  */
4288 int
4289 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4290 {
4291 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4292 	int error = 0;
4293 
4294 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4295 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4296 	if (sd_flag & STPLEX)
4297 		error = EINVAL;
4298 	else if (sd_flag & STRDERR) {
4299 		error = stp->sd_rerror;
4300 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4301 			/*
4302 			 * Read errors are non-persistent i.e. discarded once
4303 			 * returned to a non-peeking caller,
4304 			 */
4305 			stp->sd_rerror = 0;
4306 			stp->sd_flag &= ~STRDERR;
4307 		}
4308 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4309 			int clearerr = 0;
4310 
4311 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4312 			    &clearerr);
4313 			if (clearerr) {
4314 				stp->sd_flag &= ~STRDERR;
4315 				stp->sd_rderrfunc = NULL;
4316 			}
4317 		}
4318 	} else if (sd_flag & STWRERR) {
4319 		error = stp->sd_werror;
4320 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4321 			/*
4322 			 * Write errors are non-persistent i.e. discarded once
4323 			 * returned to a non-peeking caller,
4324 			 */
4325 			stp->sd_werror = 0;
4326 			stp->sd_flag &= ~STWRERR;
4327 		}
4328 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4329 			int clearerr = 0;
4330 
4331 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4332 			    &clearerr);
4333 			if (clearerr) {
4334 				stp->sd_flag &= ~STWRERR;
4335 				stp->sd_wrerrfunc = NULL;
4336 			}
4337 		}
4338 	} else if (sd_flag & STRHUP) {
4339 		/* sd_werror set when STRHUP */
4340 		error = stp->sd_werror;
4341 	}
4342 	return (error);
4343 }
4344 
4345 
4346 /*
4347  * Single-thread open/close/push/pop
4348  * for twisted streams also
4349  */
4350 int
4351 strstartplumb(stdata_t *stp, int flag, int cmd)
4352 {
4353 	int waited = 1;
4354 	int error = 0;
4355 
4356 	if (STRMATED(stp)) {
4357 		struct stdata *stmatep = stp->sd_mate;
4358 
4359 		STRLOCKMATES(stp);
4360 		while (waited) {
4361 			waited = 0;
4362 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4363 				if ((cmd == I_POP) &&
4364 				    (flag & (FNDELAY|FNONBLOCK))) {
4365 					STRUNLOCKMATES(stp);
4366 					return (EAGAIN);
4367 				}
4368 				waited = 1;
4369 				mutex_exit(&stp->sd_lock);
4370 				if (!cv_wait_sig(&stmatep->sd_monitor,
4371 				    &stmatep->sd_lock)) {
4372 					mutex_exit(&stmatep->sd_lock);
4373 					return (EINTR);
4374 				}
4375 				mutex_exit(&stmatep->sd_lock);
4376 				STRLOCKMATES(stp);
4377 			}
4378 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4379 				if ((cmd == I_POP) &&
4380 				    (flag & (FNDELAY|FNONBLOCK))) {
4381 					STRUNLOCKMATES(stp);
4382 					return (EAGAIN);
4383 				}
4384 				waited = 1;
4385 				mutex_exit(&stmatep->sd_lock);
4386 				if (!cv_wait_sig(&stp->sd_monitor,
4387 				    &stp->sd_lock)) {
4388 					mutex_exit(&stp->sd_lock);
4389 					return (EINTR);
4390 				}
4391 				mutex_exit(&stp->sd_lock);
4392 				STRLOCKMATES(stp);
4393 			}
4394 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4395 				error = strgeterr(stp,
4396 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4397 				if (error != 0) {
4398 					STRUNLOCKMATES(stp);
4399 					return (error);
4400 				}
4401 			}
4402 		}
4403 		stp->sd_flag |= STRPLUMB;
4404 		STRUNLOCKMATES(stp);
4405 	} else {
4406 		mutex_enter(&stp->sd_lock);
4407 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4408 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4409 			    (flag & (FNDELAY|FNONBLOCK))) {
4410 				mutex_exit(&stp->sd_lock);
4411 				return (EAGAIN);
4412 			}
4413 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4414 				mutex_exit(&stp->sd_lock);
4415 				return (EINTR);
4416 			}
4417 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4418 				error = strgeterr(stp,
4419 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4420 				if (error != 0) {
4421 					mutex_exit(&stp->sd_lock);
4422 					return (error);
4423 				}
4424 			}
4425 		}
4426 		stp->sd_flag |= STRPLUMB;
4427 		mutex_exit(&stp->sd_lock);
4428 	}
4429 	return (0);
4430 }
4431 
4432 /*
4433  * Complete the plumbing operation associated with stream `stp'.
4434  */
4435 void
4436 strendplumb(stdata_t *stp)
4437 {
4438 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4439 	ASSERT(stp->sd_flag & STRPLUMB);
4440 	stp->sd_flag &= ~STRPLUMB;
4441 	cv_broadcast(&stp->sd_monitor);
4442 }
4443 
4444 /*
4445  * This describes how the STREAMS framework handles synchronization
4446  * during open/push and close/pop.
4447  * The key interfaces for open and close are qprocson and qprocsoff,
4448  * respectively. While the close case in general is harder both open
4449  * have close have significant similarities.
4450  *
4451  * During close the STREAMS framework has to both ensure that there
4452  * are no stale references to the queue pair (and syncq) that
4453  * are being closed and also provide the guarantees that are documented
4454  * in qprocsoff(9F).
4455  * If there are stale references to the queue that is closing it can
4456  * result in kernel memory corruption or kernel panics.
4457  *
4458  * Note that is it up to the module/driver to ensure that it itself
4459  * does not have any stale references to the closing queues once its close
4460  * routine returns. This includes:
4461  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4462  *    associated with the queues. For timeout and bufcall callbacks the
4463  *    module/driver also has to ensure (or wait for) any callbacks that
4464  *    are in progress.
4465  *  - If the module/driver is using esballoc it has to ensure that any
4466  *    esballoc free functions do not refer to a queue that has closed.
4467  *    (Note that in general the close routine can not wait for the esballoc'ed
4468  *    messages to be freed since that can cause a deadlock.)
4469  *  - Cancelling any interrupts that refer to the closing queues and
4470  *    also ensuring that there are no interrupts in progress that will
4471  *    refer to the closing queues once the close routine returns.
4472  *  - For multiplexors removing any driver global state that refers to
4473  *    the closing queue and also ensuring that there are no threads in
4474  *    the multiplexor that has picked up a queue pointer but not yet
4475  *    finished using it.
4476  *
4477  * In addition, a driver/module can only reference the q_next pointer
4478  * in its open, close, put, or service procedures or in a
4479  * qtimeout/qbufcall callback procedure executing "on" the correct
4480  * stream. Thus it can not reference the q_next pointer in an interrupt
4481  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4482  * it can not reference q_next of a different queue e.g. in a mux that
4483  * passes messages from one queues put/service procedure to another queue.
4484  * In all the cases when the driver/module can not access the q_next
4485  * field it must use the *next* versions e.g. canputnext instead of
4486  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4487  *
4488  *
4489  * Assuming that the driver/module conforms to the above constraints
4490  * the STREAMS framework has to avoid stale references to q_next for all
4491  * the framework internal cases which include (but are not limited to):
4492  *  - Threads in canput/canputnext/backenable and elsewhere that are
4493  *    walking q_next.
4494  *  - Messages on a syncq that have a reference to the queue through b_queue.
4495  *  - Messages on an outer perimeter (syncq) that have a reference to the
4496  *    queue through b_queue.
4497  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4498  *    Note that only canput and bcanput use q_nfsrv without any locking.
4499  *
4500  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4501  * after qprocsoff returns, the framework has to ensure that no threads can
4502  * enter the put or service routines for the closing read or write-side queue.
4503  * In addition to preventing "direct" entry into the put procedures
4504  * the framework also has to prevent messages being drained from
4505  * the syncq or the outer perimeter.
4506  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4507  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4508  * qprocsoff has returned.
4509  * Note that if a module/driver uses put(9F) on one of its own queues
4510  * it is up to the module/driver to ensure that the put() doesn't
4511  * get called when the queue is closing.
4512  *
4513  *
4514  * The framework aspects of the above "contract" is implemented by
4515  * qprocsoff, removeq, and strlock:
4516  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4517  *    entering the service procedures.
4518  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4519  *    canputnext, backenable etc from dereferencing the q_next that will
4520  *    soon change.
4521  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4522  *    or other q_next walker that uses claimstr/releasestr to finish.
4523  *  - optionally for every syncq in the stream strlock acquires all the
4524  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4525  *    that no thread executes in the put or service procedures and that no
4526  *    thread is draining into the module/driver. This ensures that no
4527  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4528  *    currently executing hence no such thread can end up with the old stale
4529  *    q_next value and no canput/backenable can have the old stale
4530  *    q_nfsrv/q_next.
4531  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4532  *    have either finished or observed the QWCLOSE flag and gone away.
4533  */
4534 
4535 
4536 /*
4537  * Get all the locks necessary to change q_next.
4538  *
4539  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4540  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4541  * the only threads inside the syncq are threads currently calling removeq().
4542  * Since threads calling removeq() are in the process of removing their queues
4543  * from the stream, we do not need to worry about them accessing a stale q_next
4544  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4545  * for them can cause deadlock).
4546  *
4547  * This routine is subject to starvation since it does not set any flag to
4548  * prevent threads from entering a module in the stream (i.e. sq_count can
4549  * increase on some syncq while it is waiting on some other syncq).
4550  *
4551  * Assumes that only one thread attempts to call strlock for a given
4552  * stream. If this is not the case the two threads would deadlock.
4553  * This assumption is guaranteed since strlock is only called by insertq
4554  * and removeq and streams plumbing changes are single-threaded for
4555  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4556  *
4557  * For pipes, it is not difficult to atomically designate a pair of streams
4558  * to be mated. Once mated atomically by the framework the twisted pair remain
4559  * configured that way until dismantled atomically by the framework.
4560  * When plumbing takes place on a twisted stream it is necessary to ensure that
4561  * this operation is done exclusively on the twisted stream since two such
4562  * operations, each initiated on different ends of the pipe will deadlock
4563  * waiting for each other to complete.
4564  *
4565  * On entry, no locks should be held.
4566  * The locks acquired and held by strlock depends on a few factors.
4567  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4568  *   and held on exit and all sq_count are at an acceptable level.
4569  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4570  *   sd_refcnt being zero.
4571  */
4572 
4573 static void
4574 strlock(struct stdata *stp, sqlist_t *sqlist)
4575 {
4576 	syncql_t *sql, *sql2;
4577 retry:
4578 	/*
4579 	 * Wait for any claimstr to go away.
4580 	 */
4581 	if (STRMATED(stp)) {
4582 		struct stdata *stp1, *stp2;
4583 
4584 		STRLOCKMATES(stp);
4585 		/*
4586 		 * Note that the selection of locking order is not
4587 		 * important, just that they are always acquired in
4588 		 * the same order.  To assure this, we choose this
4589 		 * order based on the value of the pointer, and since
4590 		 * the pointer will not change for the life of this
4591 		 * pair, we will always grab the locks in the same
4592 		 * order (and hence, prevent deadlocks).
4593 		 */
4594 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4595 			stp1 = stp;
4596 			stp2 = stp->sd_mate;
4597 		} else {
4598 			stp2 = stp;
4599 			stp1 = stp->sd_mate;
4600 		}
4601 		mutex_enter(&stp1->sd_reflock);
4602 		if (stp1->sd_refcnt > 0) {
4603 			STRUNLOCKMATES(stp);
4604 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4605 			mutex_exit(&stp1->sd_reflock);
4606 			goto retry;
4607 		}
4608 		mutex_enter(&stp2->sd_reflock);
4609 		if (stp2->sd_refcnt > 0) {
4610 			STRUNLOCKMATES(stp);
4611 			mutex_exit(&stp1->sd_reflock);
4612 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4613 			mutex_exit(&stp2->sd_reflock);
4614 			goto retry;
4615 		}
4616 		STREAM_PUTLOCKS_ENTER(stp1);
4617 		STREAM_PUTLOCKS_ENTER(stp2);
4618 	} else {
4619 		mutex_enter(&stp->sd_lock);
4620 		mutex_enter(&stp->sd_reflock);
4621 		while (stp->sd_refcnt > 0) {
4622 			mutex_exit(&stp->sd_lock);
4623 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4624 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4625 				mutex_exit(&stp->sd_reflock);
4626 				mutex_enter(&stp->sd_lock);
4627 				mutex_enter(&stp->sd_reflock);
4628 			}
4629 		}
4630 		STREAM_PUTLOCKS_ENTER(stp);
4631 	}
4632 
4633 	if (sqlist == NULL)
4634 		return;
4635 
4636 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4637 		syncq_t *sq = sql->sql_sq;
4638 		uint16_t count;
4639 
4640 		mutex_enter(SQLOCK(sq));
4641 		count = sq->sq_count;
4642 		ASSERT(sq->sq_rmqcount <= count);
4643 		SQ_PUTLOCKS_ENTER(sq);
4644 		SUM_SQ_PUTCOUNTS(sq, count);
4645 		if (count == sq->sq_rmqcount)
4646 			continue;
4647 
4648 		/* Failed - drop all locks that we have acquired so far */
4649 		if (STRMATED(stp)) {
4650 			STREAM_PUTLOCKS_EXIT(stp);
4651 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4652 			STRUNLOCKMATES(stp);
4653 			mutex_exit(&stp->sd_reflock);
4654 			mutex_exit(&stp->sd_mate->sd_reflock);
4655 		} else {
4656 			STREAM_PUTLOCKS_EXIT(stp);
4657 			mutex_exit(&stp->sd_lock);
4658 			mutex_exit(&stp->sd_reflock);
4659 		}
4660 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4661 		    sql2 = sql2->sql_next) {
4662 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4663 			mutex_exit(SQLOCK(sql2->sql_sq));
4664 		}
4665 
4666 		/*
4667 		 * The wait loop below may starve when there are many threads
4668 		 * claiming the syncq. This is especially a problem with permod
4669 		 * syncqs (IP). To lessen the impact of the problem we increment
4670 		 * sq_needexcl and clear fastbits so that putnexts will slow
4671 		 * down and call sqenable instead of draining right away.
4672 		 */
4673 		sq->sq_needexcl++;
4674 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4675 		while (count > sq->sq_rmqcount) {
4676 			sq->sq_flags |= SQ_WANTWAKEUP;
4677 			SQ_PUTLOCKS_EXIT(sq);
4678 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4679 			count = sq->sq_count;
4680 			SQ_PUTLOCKS_ENTER(sq);
4681 			SUM_SQ_PUTCOUNTS(sq, count);
4682 		}
4683 		sq->sq_needexcl--;
4684 		if (sq->sq_needexcl == 0)
4685 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4686 		SQ_PUTLOCKS_EXIT(sq);
4687 		ASSERT(count == sq->sq_rmqcount);
4688 		mutex_exit(SQLOCK(sq));
4689 		goto retry;
4690 	}
4691 }
4692 
4693 /*
4694  * Drop all the locks that strlock acquired.
4695  */
4696 static void
4697 strunlock(struct stdata *stp, sqlist_t *sqlist)
4698 {
4699 	syncql_t *sql;
4700 
4701 	if (STRMATED(stp)) {
4702 		STREAM_PUTLOCKS_EXIT(stp);
4703 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4704 		STRUNLOCKMATES(stp);
4705 		mutex_exit(&stp->sd_reflock);
4706 		mutex_exit(&stp->sd_mate->sd_reflock);
4707 	} else {
4708 		STREAM_PUTLOCKS_EXIT(stp);
4709 		mutex_exit(&stp->sd_lock);
4710 		mutex_exit(&stp->sd_reflock);
4711 	}
4712 
4713 	if (sqlist == NULL)
4714 		return;
4715 
4716 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4717 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4718 		mutex_exit(SQLOCK(sql->sql_sq));
4719 	}
4720 }
4721 
4722 /*
4723  * When the module has service procedure, we need check if the next
4724  * module which has service procedure is in flow control to trigger
4725  * the backenable.
4726  */
4727 static void
4728 backenable_insertedq(queue_t *q)
4729 {
4730 	qband_t	*qbp;
4731 
4732 	claimstr(q);
4733 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4734 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4735 			backenable(q, 0);
4736 
4737 		qbp = q->q_next->q_nfsrv->q_bandp;
4738 		for (; qbp != NULL; qbp = qbp->qb_next)
4739 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4740 				backenable(q, qbp->qb_first->b_band);
4741 	}
4742 	releasestr(q);
4743 }
4744 
4745 /*
4746  * Given two read queues, insert a new single one after another.
4747  *
4748  * This routine acquires all the necessary locks in order to change
4749  * q_next and related pointer using strlock().
4750  * It depends on the stream head ensuring that there are no concurrent
4751  * insertq or removeq on the same stream. The stream head ensures this
4752  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4753  *
4754  * Note that no syncq locks are held during the q_next change. This is
4755  * applied to all streams since, unlike removeq, there is no problem of stale
4756  * pointers when adding a module to the stream. Thus drivers/modules that do a
4757  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4758  * applied this optimization to all streams.
4759  */
4760 void
4761 insertq(struct stdata *stp, queue_t *new)
4762 {
4763 	queue_t	*after;
4764 	queue_t *wafter;
4765 	queue_t *wnew = _WR(new);
4766 	boolean_t have_fifo = B_FALSE;
4767 
4768 	if (new->q_flag & _QINSERTING) {
4769 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4770 		after = new->q_next;
4771 		wafter = _WR(new->q_next);
4772 	} else {
4773 		after = _RD(stp->sd_wrq);
4774 		wafter = stp->sd_wrq;
4775 	}
4776 
4777 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4778 	    "insertq:%p, %p", after, new);
4779 	ASSERT(after->q_flag & QREADR);
4780 	ASSERT(new->q_flag & QREADR);
4781 
4782 	strlock(stp, NULL);
4783 
4784 	/* Do we have a FIFO? */
4785 	if (wafter->q_next == after) {
4786 		have_fifo = B_TRUE;
4787 		wnew->q_next = new;
4788 	} else {
4789 		wnew->q_next = wafter->q_next;
4790 	}
4791 	new->q_next = after;
4792 
4793 	set_nfsrv_ptr(new, wnew, after, wafter);
4794 	/*
4795 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4796 	 * so only reset this flag after calling it.
4797 	 */
4798 	new->q_flag &= ~_QINSERTING;
4799 
4800 	if (have_fifo) {
4801 		wafter->q_next = wnew;
4802 	} else {
4803 		if (wafter->q_next)
4804 			_OTHERQ(wafter->q_next)->q_next = new;
4805 		wafter->q_next = wnew;
4806 	}
4807 
4808 	set_qend(new);
4809 	/* The QEND flag might have to be updated for the upstream guy */
4810 	set_qend(after);
4811 
4812 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4813 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4814 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4815 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4816 	strsetuio(stp);
4817 
4818 	/*
4819 	 * If this was a module insertion, bump the push count.
4820 	 */
4821 	if (!(new->q_flag & QISDRV))
4822 		stp->sd_pushcnt++;
4823 
4824 	strunlock(stp, NULL);
4825 
4826 	/* check if the write Q needs backenable */
4827 	backenable_insertedq(wnew);
4828 
4829 	/* check if the read Q needs backenable */
4830 	backenable_insertedq(new);
4831 }
4832 
4833 /*
4834  * Given a read queue, unlink it from any neighbors.
4835  *
4836  * This routine acquires all the necessary locks in order to
4837  * change q_next and related pointers and also guard against
4838  * stale references (e.g. through q_next) to the queue that
4839  * is being removed. It also plays part of the role in ensuring
4840  * that the module's/driver's put procedure doesn't get called
4841  * after qprocsoff returns.
4842  *
4843  * Removeq depends on the stream head ensuring that there are
4844  * no concurrent insertq or removeq on the same stream. The
4845  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4846  * STRPLUMB.
4847  *
4848  * The set of locks needed to remove the queue is different in
4849  * different cases:
4850  *
4851  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4852  * waiting for the syncq reference count to drop to 0 indicating that no
4853  * non-close threads are present anywhere in the stream. This ensures that any
4854  * module/driver can reference q_next in its open, close, put, or service
4855  * procedures.
4856  *
4857  * The sq_rmqcount counter tracks the number of threads inside removeq().
4858  * strlock() ensures that there is either no threads executing inside perimeter
4859  * or there is only a thread calling qprocsoff().
4860  *
4861  * strlock() compares the value of sq_count with the number of threads inside
4862  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4863  * any threads waiting in strlock() when the sq_rmqcount increases.
4864  */
4865 
4866 void
4867 removeq(queue_t *qp)
4868 {
4869 	queue_t *wqp = _WR(qp);
4870 	struct stdata *stp = STREAM(qp);
4871 	sqlist_t *sqlist = NULL;
4872 	boolean_t isdriver;
4873 	int moved;
4874 	syncq_t *sq = qp->q_syncq;
4875 	syncq_t *wsq = wqp->q_syncq;
4876 
4877 	ASSERT(stp);
4878 
4879 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4880 	    "removeq:%p %p", qp, wqp);
4881 	ASSERT(qp->q_flag&QREADR);
4882 
4883 	/*
4884 	 * For queues using Synchronous streams, we must wait for all threads in
4885 	 * rwnext() to drain out before proceeding.
4886 	 */
4887 	if (qp->q_flag & QSYNCSTR) {
4888 		/* First, we need wakeup any threads blocked in rwnext() */
4889 		mutex_enter(SQLOCK(sq));
4890 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4891 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4892 			cv_broadcast(&sq->sq_wait);
4893 		}
4894 		mutex_exit(SQLOCK(sq));
4895 
4896 		if (wsq != sq) {
4897 			mutex_enter(SQLOCK(wsq));
4898 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4899 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4900 				cv_broadcast(&wsq->sq_wait);
4901 			}
4902 			mutex_exit(SQLOCK(wsq));
4903 		}
4904 
4905 		mutex_enter(QLOCK(qp));
4906 		while (qp->q_rwcnt > 0) {
4907 			qp->q_flag |= QWANTRMQSYNC;
4908 			cv_wait(&qp->q_wait, QLOCK(qp));
4909 		}
4910 		mutex_exit(QLOCK(qp));
4911 
4912 		mutex_enter(QLOCK(wqp));
4913 		while (wqp->q_rwcnt > 0) {
4914 			wqp->q_flag |= QWANTRMQSYNC;
4915 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4916 		}
4917 		mutex_exit(QLOCK(wqp));
4918 	}
4919 
4920 	mutex_enter(SQLOCK(sq));
4921 	sq->sq_rmqcount++;
4922 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4923 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4924 		cv_broadcast(&sq->sq_wait);
4925 	}
4926 	mutex_exit(SQLOCK(sq));
4927 
4928 	isdriver = (qp->q_flag & QISDRV);
4929 
4930 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4931 	strlock(stp, sqlist);
4932 
4933 	reset_nfsrv_ptr(qp, wqp);
4934 
4935 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4936 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4937 	/* Do we have a FIFO? */
4938 	if (wqp->q_next == qp) {
4939 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4940 	} else {
4941 		if (wqp->q_next)
4942 			backq(qp)->q_next = qp->q_next;
4943 		if (qp->q_next)
4944 			backq(wqp)->q_next = wqp->q_next;
4945 	}
4946 
4947 	/* The QEND flag might have to be updated for the upstream guy */
4948 	if (qp->q_next)
4949 		set_qend(qp->q_next);
4950 
4951 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4952 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4953 
4954 	/*
4955 	 * Move any messages destined for the put procedures to the next
4956 	 * syncq in line. Otherwise free them.
4957 	 */
4958 	moved = 0;
4959 	/*
4960 	 * Quick check to see whether there are any messages or events.
4961 	 */
4962 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4963 		moved += propagate_syncq(qp);
4964 	if (wqp->q_syncqmsgs != 0 ||
4965 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4966 		moved += propagate_syncq(wqp);
4967 
4968 	strsetuio(stp);
4969 
4970 	/*
4971 	 * If this was a module removal, decrement the push count.
4972 	 */
4973 	if (!isdriver)
4974 		stp->sd_pushcnt--;
4975 
4976 	strunlock(stp, sqlist);
4977 	sqlist_free(sqlist);
4978 
4979 	/*
4980 	 * Make sure any messages that were propagated are drained.
4981 	 * Also clear any QFULL bit caused by messages that were propagated.
4982 	 */
4983 
4984 	if (qp->q_next != NULL) {
4985 		clr_qfull(qp);
4986 		/*
4987 		 * For the driver calling qprocsoff, propagate_syncq
4988 		 * frees all the messages instead of putting it in
4989 		 * the stream head
4990 		 */
4991 		if (!isdriver && (moved > 0))
4992 			emptysq(qp->q_next->q_syncq);
4993 	}
4994 	if (wqp->q_next != NULL) {
4995 		clr_qfull(wqp);
4996 		/*
4997 		 * We come here for any pop of a module except for the
4998 		 * case of driver being removed. We don't call emptysq
4999 		 * if we did not move any messages. This will avoid holding
5000 		 * PERMOD syncq locks in emptysq
5001 		 */
5002 		if (moved > 0)
5003 			emptysq(wqp->q_next->q_syncq);
5004 	}
5005 
5006 	mutex_enter(SQLOCK(sq));
5007 	sq->sq_rmqcount--;
5008 	mutex_exit(SQLOCK(sq));
5009 }
5010 
5011 /*
5012  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
5013  * SQ_WRITER) on a syncq.
5014  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
5015  * sync queue and waits until sq_count reaches maxcnt.
5016  *
5017  * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
5018  * does not care about putnext threads that are in the middle of calling put
5019  * entry points.
5020  *
5021  * This routine is used for both inner and outer syncqs.
5022  */
5023 static void
5024 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
5025 {
5026 	uint16_t count = 0;
5027 
5028 	mutex_enter(SQLOCK(sq));
5029 	/*
5030 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
5031 	 * SQ_FROZEN will be set if there is a frozen stream that has a
5032 	 * queue which also refers to this "shared" syncq.
5033 	 * SQ_BLOCKED will be set if there is "off" queue which also
5034 	 * refers to this "shared" syncq.
5035 	 */
5036 	if (maxcnt != -1) {
5037 		count = sq->sq_count;
5038 		SQ_PUTLOCKS_ENTER(sq);
5039 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5040 		SUM_SQ_PUTCOUNTS(sq, count);
5041 	}
5042 	sq->sq_needexcl++;
5043 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5044 
5045 	while ((sq->sq_flags & flag) ||
5046 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
5047 		sq->sq_flags |= SQ_WANTWAKEUP;
5048 		if (maxcnt != -1) {
5049 			SQ_PUTLOCKS_EXIT(sq);
5050 		}
5051 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5052 		if (maxcnt != -1) {
5053 			count = sq->sq_count;
5054 			SQ_PUTLOCKS_ENTER(sq);
5055 			SUM_SQ_PUTCOUNTS(sq, count);
5056 		}
5057 	}
5058 	sq->sq_needexcl--;
5059 	sq->sq_flags |= flag;
5060 	ASSERT(maxcnt == -1 || count == maxcnt);
5061 	if (maxcnt != -1) {
5062 		if (sq->sq_needexcl == 0) {
5063 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5064 		}
5065 		SQ_PUTLOCKS_EXIT(sq);
5066 	} else if (sq->sq_needexcl == 0) {
5067 		SQ_PUTCOUNT_SETFAST(sq);
5068 	}
5069 
5070 	mutex_exit(SQLOCK(sq));
5071 }
5072 
5073 /*
5074  * Reset a flag that was set with blocksq.
5075  *
5076  * Can not use this routine to reset SQ_WRITER.
5077  *
5078  * If "isouter" is set then the syncq is assumed to be an outer perimeter
5079  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5080  * to handle the queued qwriter operations.
5081  *
5082  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5083  * sq_putlocks are used.
5084  */
5085 static void
5086 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5087 {
5088 	uint16_t flags;
5089 
5090 	mutex_enter(SQLOCK(sq));
5091 	ASSERT(resetflag != SQ_WRITER);
5092 	ASSERT(sq->sq_flags & resetflag);
5093 	flags = sq->sq_flags & ~resetflag;
5094 	sq->sq_flags = flags;
5095 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5096 		if (flags & SQ_WANTWAKEUP) {
5097 			flags &= ~SQ_WANTWAKEUP;
5098 			cv_broadcast(&sq->sq_wait);
5099 		}
5100 		sq->sq_flags = flags;
5101 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5102 			if (!isouter) {
5103 				/* drain_syncq drops SQLOCK */
5104 				drain_syncq(sq);
5105 				return;
5106 			}
5107 		}
5108 	}
5109 	mutex_exit(SQLOCK(sq));
5110 }
5111 
5112 /*
5113  * Reset a flag that was set with blocksq.
5114  * Does not drain the syncq. Use emptysq() for that.
5115  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5116  *
5117  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5118  * sq_putlocks are used.
5119  */
5120 static int
5121 dropsq(syncq_t *sq, uint16_t resetflag)
5122 {
5123 	uint16_t flags;
5124 
5125 	mutex_enter(SQLOCK(sq));
5126 	ASSERT(sq->sq_flags & resetflag);
5127 	flags = sq->sq_flags & ~resetflag;
5128 	if (flags & SQ_WANTWAKEUP) {
5129 		flags &= ~SQ_WANTWAKEUP;
5130 		cv_broadcast(&sq->sq_wait);
5131 	}
5132 	sq->sq_flags = flags;
5133 	mutex_exit(SQLOCK(sq));
5134 	if (flags & SQ_QUEUED)
5135 		return (1);
5136 	return (0);
5137 }
5138 
5139 /*
5140  * Empty all the messages on a syncq.
5141  *
5142  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5143  * sq_putlocks are used.
5144  */
5145 static void
5146 emptysq(syncq_t *sq)
5147 {
5148 	uint16_t flags;
5149 
5150 	mutex_enter(SQLOCK(sq));
5151 	flags = sq->sq_flags;
5152 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5153 		/*
5154 		 * To prevent potential recursive invocation of drain_syncq we
5155 		 * do not call drain_syncq if count is non-zero.
5156 		 */
5157 		if (sq->sq_count == 0) {
5158 			/* drain_syncq() drops SQLOCK */
5159 			drain_syncq(sq);
5160 			return;
5161 		} else
5162 			sqenable(sq);
5163 	}
5164 	mutex_exit(SQLOCK(sq));
5165 }
5166 
5167 /*
5168  * Ordered insert while removing duplicates.
5169  */
5170 static void
5171 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5172 {
5173 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5174 
5175 	prev_sqlpp = &sqlist->sqlist_head;
5176 	while ((sqlp = *prev_sqlpp) != NULL) {
5177 		if (sqlp->sql_sq >= sqp) {
5178 			if (sqlp->sql_sq == sqp)	/* duplicate */
5179 				return;
5180 			break;
5181 		}
5182 		prev_sqlpp = &sqlp->sql_next;
5183 	}
5184 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5185 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5186 	new_sqlp->sql_next = sqlp;
5187 	new_sqlp->sql_sq = sqp;
5188 	*prev_sqlpp = new_sqlp;
5189 }
5190 
5191 /*
5192  * Walk the write side queues until we hit either the driver
5193  * or a twist in the stream (_SAMESTR will return false in both
5194  * these cases) then turn around and walk the read side queues
5195  * back up to the stream head.
5196  */
5197 static void
5198 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5199 {
5200 	while (q != NULL) {
5201 		sqlist_insert(sqlist, q->q_syncq);
5202 
5203 		if (_SAMESTR(q))
5204 			q = q->q_next;
5205 		else if (!(q->q_flag & QREADR))
5206 			q = _RD(q);
5207 		else
5208 			q = NULL;
5209 	}
5210 }
5211 
5212 /*
5213  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5214  * associated with the "q" parameter. The resulting list is sorted in a
5215  * canonical order and is free of duplicates.
5216  * Assumes the passed queue is a _RD(q).
5217  */
5218 static sqlist_t *
5219 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5220 {
5221 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5222 
5223 	/*
5224 	 * start with the current queue/qpair
5225 	 */
5226 	ASSERT(q->q_flag & QREADR);
5227 
5228 	sqlist_insert(sqlist, q->q_syncq);
5229 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5230 
5231 	sqlist_insertall(sqlist, stp->sd_wrq);
5232 	if (do_twist)
5233 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5234 
5235 	return (sqlist);
5236 }
5237 
5238 static sqlist_t *
5239 sqlist_alloc(struct stdata *stp, int kmflag)
5240 {
5241 	size_t sqlist_size;
5242 	sqlist_t *sqlist;
5243 
5244 	/*
5245 	 * Allocate 2 syncql_t's for each pushed module. Note that
5246 	 * the sqlist_t structure already has 4 syncql_t's built in:
5247 	 * 2 for the stream head, and 2 for the driver/other stream head.
5248 	 */
5249 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5250 	    sizeof (sqlist_t);
5251 	if (STRMATED(stp))
5252 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5253 	sqlist = kmem_alloc(sqlist_size, kmflag);
5254 
5255 	sqlist->sqlist_head = NULL;
5256 	sqlist->sqlist_size = sqlist_size;
5257 	sqlist->sqlist_index = 0;
5258 
5259 	return (sqlist);
5260 }
5261 
5262 /*
5263  * Free the list created by sqlist_alloc()
5264  */
5265 static void
5266 sqlist_free(sqlist_t *sqlist)
5267 {
5268 	kmem_free(sqlist, sqlist->sqlist_size);
5269 }
5270 
5271 /*
5272  * Prevent any new entries into any syncq in this stream.
5273  * Used by freezestr.
5274  */
5275 void
5276 strblock(queue_t *q)
5277 {
5278 	struct stdata	*stp;
5279 	syncql_t	*sql;
5280 	sqlist_t	*sqlist;
5281 
5282 	q = _RD(q);
5283 
5284 	stp = STREAM(q);
5285 	ASSERT(stp != NULL);
5286 
5287 	/*
5288 	 * Get a sorted list with all the duplicates removed containing
5289 	 * all the syncqs referenced by this stream.
5290 	 */
5291 	sqlist = sqlist_build(q, stp, B_FALSE);
5292 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5293 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5294 	sqlist_free(sqlist);
5295 }
5296 
5297 /*
5298  * Release the block on new entries into this stream
5299  */
5300 void
5301 strunblock(queue_t *q)
5302 {
5303 	struct stdata	*stp;
5304 	syncql_t	*sql;
5305 	sqlist_t	*sqlist;
5306 	int		drain_needed;
5307 
5308 	q = _RD(q);
5309 
5310 	/*
5311 	 * Get a sorted list with all the duplicates removed containing
5312 	 * all the syncqs referenced by this stream.
5313 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5314 	 * starting to drain them; otherwise the draining might
5315 	 * cause a freezestr in some module on the stream (which
5316 	 * would deadlock).
5317 	 */
5318 	stp = STREAM(q);
5319 	ASSERT(stp != NULL);
5320 	sqlist = sqlist_build(q, stp, B_FALSE);
5321 	drain_needed = 0;
5322 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5323 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5324 	if (drain_needed) {
5325 		for (sql = sqlist->sqlist_head; sql != NULL;
5326 		    sql = sql->sql_next)
5327 			emptysq(sql->sql_sq);
5328 	}
5329 	sqlist_free(sqlist);
5330 }
5331 
5332 #ifdef DEBUG
5333 static int
5334 qprocsareon(queue_t *rq)
5335 {
5336 	if (rq->q_next == NULL)
5337 		return (0);
5338 	return (_WR(rq->q_next)->q_next == _WR(rq));
5339 }
5340 
5341 int
5342 qclaimed(queue_t *q)
5343 {
5344 	uint_t count;
5345 
5346 	count = q->q_syncq->sq_count;
5347 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5348 	return (count != 0);
5349 }
5350 
5351 /*
5352  * Check if anyone has frozen this stream with freezestr
5353  */
5354 int
5355 frozenstr(queue_t *q)
5356 {
5357 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5358 }
5359 #endif /* DEBUG */
5360 
5361 /*
5362  * Enter a queue.
5363  * Obsoleted interface. Should not be used.
5364  */
5365 void
5366 enterq(queue_t *q)
5367 {
5368 	entersq(q->q_syncq, SQ_CALLBACK);
5369 }
5370 
5371 void
5372 leaveq(queue_t *q)
5373 {
5374 	leavesq(q->q_syncq, SQ_CALLBACK);
5375 }
5376 
5377 /*
5378  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5379  * to check.
5380  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5381  * calls and the running of open, close and service procedures.
5382  *
5383  * If c_inner bit is set no need to grab sq_putlocks since we don't care
5384  * if other threads have entered or are entering put entry point.
5385  *
5386  * If c_inner bit is set it might have been possible to use
5387  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5388  * open/close path for IP) but since the count may need to be decremented in
5389  * qwait() we wouldn't know which counter to decrement. Currently counter is
5390  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5391  * in the future we might use curthread id bits to select the counter and this
5392  * would stay constant across routine calls.
5393  */
5394 void
5395 entersq(syncq_t *sq, int entrypoint)
5396 {
5397 	uint16_t	count = 0;
5398 	uint16_t	flags;
5399 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5400 	uint16_t	type;
5401 	uint_t		c_inner = entrypoint & SQ_CI;
5402 	uint_t		c_outer = entrypoint & SQ_CO;
5403 
5404 	/*
5405 	 * Increment ref count to keep closes out of this queue.
5406 	 */
5407 	ASSERT(sq);
5408 	ASSERT(c_inner && c_outer);
5409 	mutex_enter(SQLOCK(sq));
5410 	flags = sq->sq_flags;
5411 	type = sq->sq_type;
5412 	if (!(type & c_inner)) {
5413 		/* Make sure all putcounts now use slowlock. */
5414 		count = sq->sq_count;
5415 		SQ_PUTLOCKS_ENTER(sq);
5416 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5417 		SUM_SQ_PUTCOUNTS(sq, count);
5418 		sq->sq_needexcl++;
5419 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5420 		waitflags |= SQ_MESSAGES;
5421 	}
5422 	/*
5423 	 * Wait until we can enter the inner perimeter.
5424 	 * If we want exclusive access we wait until sq_count is 0.
5425 	 * We have to do this before entering the outer perimeter in order
5426 	 * to preserve put/close message ordering.
5427 	 */
5428 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5429 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5430 		if (!(type & c_inner)) {
5431 			SQ_PUTLOCKS_EXIT(sq);
5432 		}
5433 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5434 		if (!(type & c_inner)) {
5435 			count = sq->sq_count;
5436 			SQ_PUTLOCKS_ENTER(sq);
5437 			SUM_SQ_PUTCOUNTS(sq, count);
5438 		}
5439 		flags = sq->sq_flags;
5440 	}
5441 
5442 	if (!(type & c_inner)) {
5443 		ASSERT(sq->sq_needexcl > 0);
5444 		sq->sq_needexcl--;
5445 		if (sq->sq_needexcl == 0) {
5446 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5447 		}
5448 	}
5449 
5450 	/* Check if we need to enter the outer perimeter */
5451 	if (!(type & c_outer)) {
5452 		/*
5453 		 * We have to enter the outer perimeter exclusively before
5454 		 * we can increment sq_count to avoid deadlock. This implies
5455 		 * that we have to re-check sq_flags and sq_count.
5456 		 *
5457 		 * is it possible to have c_inner set when c_outer is not set?
5458 		 */
5459 		if (!(type & c_inner)) {
5460 			SQ_PUTLOCKS_EXIT(sq);
5461 		}
5462 		mutex_exit(SQLOCK(sq));
5463 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5464 		mutex_enter(SQLOCK(sq));
5465 		flags = sq->sq_flags;
5466 		/*
5467 		 * there should be no need to recheck sq_putcounts
5468 		 * because outer_enter() has already waited for them to clear
5469 		 * after setting SQ_WRITER.
5470 		 */
5471 		count = sq->sq_count;
5472 #ifdef DEBUG
5473 		/*
5474 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5475 		 * of doing an ASSERT internally. Others should do
5476 		 * something like
5477 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5478 		 * without the need to #ifdef DEBUG it.
5479 		 */
5480 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5481 #endif
5482 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5483 		    (!(type & c_inner) && count != 0)) {
5484 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5485 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5486 			count = sq->sq_count;
5487 			flags = sq->sq_flags;
5488 		}
5489 	}
5490 
5491 	sq->sq_count++;
5492 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5493 	if (!(type & c_inner)) {
5494 		/* Exclusive entry */
5495 		ASSERT(sq->sq_count == 1);
5496 		sq->sq_flags |= SQ_EXCL;
5497 		if (type & c_outer) {
5498 			SQ_PUTLOCKS_EXIT(sq);
5499 		}
5500 	}
5501 	mutex_exit(SQLOCK(sq));
5502 }
5503 
5504 /*
5505  * Leave a syncq. Announce to framework that closes may proceed.
5506  * c_inner and c_outer specify which concurrency bits to check.
5507  *
5508  * Must never be called from driver or module put entry point.
5509  *
5510  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5511  * sq_putlocks are used.
5512  */
5513 void
5514 leavesq(syncq_t *sq, int entrypoint)
5515 {
5516 	uint16_t	flags;
5517 	uint16_t	type;
5518 	uint_t		c_outer = entrypoint & SQ_CO;
5519 #ifdef DEBUG
5520 	uint_t		c_inner = entrypoint & SQ_CI;
5521 #endif
5522 
5523 	/*
5524 	 * Decrement ref count, drain the syncq if possible, and wake up
5525 	 * any waiting close.
5526 	 */
5527 	ASSERT(sq);
5528 	ASSERT(c_inner && c_outer);
5529 	mutex_enter(SQLOCK(sq));
5530 	flags = sq->sq_flags;
5531 	type = sq->sq_type;
5532 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5533 
5534 		if (flags & SQ_WANTWAKEUP) {
5535 			flags &= ~SQ_WANTWAKEUP;
5536 			cv_broadcast(&sq->sq_wait);
5537 		}
5538 		if (flags & SQ_WANTEXWAKEUP) {
5539 			flags &= ~SQ_WANTEXWAKEUP;
5540 			cv_broadcast(&sq->sq_exitwait);
5541 		}
5542 
5543 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5544 			/*
5545 			 * The syncq needs to be drained. "Exit" the syncq
5546 			 * before calling drain_syncq.
5547 			 */
5548 			ASSERT(sq->sq_count != 0);
5549 			sq->sq_count--;
5550 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5551 			sq->sq_flags = flags & ~SQ_EXCL;
5552 			drain_syncq(sq);
5553 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5554 			/* Check if we need to exit the outer perimeter */
5555 			/* XXX will this ever be true? */
5556 			if (!(type & c_outer))
5557 				outer_exit(sq->sq_outer);
5558 			return;
5559 		}
5560 	}
5561 	ASSERT(sq->sq_count != 0);
5562 	sq->sq_count--;
5563 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5564 	sq->sq_flags = flags & ~SQ_EXCL;
5565 	mutex_exit(SQLOCK(sq));
5566 
5567 	/* Check if we need to exit the outer perimeter */
5568 	if (!(sq->sq_type & c_outer))
5569 		outer_exit(sq->sq_outer);
5570 }
5571 
5572 /*
5573  * Prevent q_next from changing in this stream by incrementing sq_count.
5574  *
5575  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5576  * sq_putlocks are used.
5577  */
5578 void
5579 claimq(queue_t *qp)
5580 {
5581 	syncq_t	*sq = qp->q_syncq;
5582 
5583 	mutex_enter(SQLOCK(sq));
5584 	sq->sq_count++;
5585 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5586 	mutex_exit(SQLOCK(sq));
5587 }
5588 
5589 /*
5590  * Undo claimq.
5591  *
5592  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5593  * sq_putlocks are used.
5594  */
5595 void
5596 releaseq(queue_t *qp)
5597 {
5598 	syncq_t	*sq = qp->q_syncq;
5599 	uint16_t flags;
5600 
5601 	mutex_enter(SQLOCK(sq));
5602 	ASSERT(sq->sq_count > 0);
5603 	sq->sq_count--;
5604 
5605 	flags = sq->sq_flags;
5606 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5607 		if (flags & SQ_WANTWAKEUP) {
5608 			flags &= ~SQ_WANTWAKEUP;
5609 			cv_broadcast(&sq->sq_wait);
5610 		}
5611 		sq->sq_flags = flags;
5612 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5613 			/*
5614 			 * To prevent potential recursive invocation of
5615 			 * drain_syncq we do not call drain_syncq if count is
5616 			 * non-zero.
5617 			 */
5618 			if (sq->sq_count == 0) {
5619 				drain_syncq(sq);
5620 				return;
5621 			} else
5622 				sqenable(sq);
5623 		}
5624 	}
5625 	mutex_exit(SQLOCK(sq));
5626 }
5627 
5628 /*
5629  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5630  */
5631 void
5632 claimstr(queue_t *qp)
5633 {
5634 	struct stdata *stp = STREAM(qp);
5635 
5636 	mutex_enter(&stp->sd_reflock);
5637 	stp->sd_refcnt++;
5638 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5639 	mutex_exit(&stp->sd_reflock);
5640 }
5641 
5642 /*
5643  * Undo claimstr.
5644  */
5645 void
5646 releasestr(queue_t *qp)
5647 {
5648 	struct stdata *stp = STREAM(qp);
5649 
5650 	mutex_enter(&stp->sd_reflock);
5651 	ASSERT(stp->sd_refcnt != 0);
5652 	if (--stp->sd_refcnt == 0)
5653 		cv_broadcast(&stp->sd_refmonitor);
5654 	mutex_exit(&stp->sd_reflock);
5655 }
5656 
5657 static syncq_t *
5658 new_syncq(void)
5659 {
5660 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5661 }
5662 
5663 static void
5664 free_syncq(syncq_t *sq)
5665 {
5666 	ASSERT(sq->sq_head == NULL);
5667 	ASSERT(sq->sq_outer == NULL);
5668 	ASSERT(sq->sq_callbpend == NULL);
5669 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5670 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5671 
5672 	if (sq->sq_ciputctrl != NULL) {
5673 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5674 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5675 		    sq->sq_nciputctrl, 0);
5676 		ASSERT(ciputctrl_cache != NULL);
5677 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5678 	}
5679 
5680 	sq->sq_tail = NULL;
5681 	sq->sq_evhead = NULL;
5682 	sq->sq_evtail = NULL;
5683 	sq->sq_ciputctrl = NULL;
5684 	sq->sq_nciputctrl = 0;
5685 	sq->sq_count = 0;
5686 	sq->sq_rmqcount = 0;
5687 	sq->sq_callbflags = 0;
5688 	sq->sq_cancelid = 0;
5689 	sq->sq_next = NULL;
5690 	sq->sq_needexcl = 0;
5691 	sq->sq_svcflags = 0;
5692 	sq->sq_nqueues = 0;
5693 	sq->sq_pri = 0;
5694 	sq->sq_onext = NULL;
5695 	sq->sq_oprev = NULL;
5696 	sq->sq_flags = 0;
5697 	sq->sq_type = 0;
5698 	sq->sq_servcount = 0;
5699 
5700 	kmem_cache_free(syncq_cache, sq);
5701 }
5702 
5703 /* Outer perimeter code */
5704 
5705 /*
5706  * The outer syncq uses the fields and flags in the syncq slightly
5707  * differently from the inner syncqs.
5708  *	sq_count	Incremented when there are pending or running
5709  *			writers at the outer perimeter to prevent the set of
5710  *			inner syncqs that belong to the outer perimeter from
5711  *			changing.
5712  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5713  *
5714  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5715  *			inner syncqs are added to or removed from the
5716  *			outer perimeter.
5717  *	SQ_QUEUED	sq_head/tail has messages or events queued.
5718  *
5719  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5720  *			setting the SQ_WRITER flag.
5721  */
5722 
5723 /*
5724  * Get write access at the outer perimeter.
5725  * Note that read access is done by entersq, putnext, and put by simply
5726  * incrementing sq_count in the inner syncq.
5727  *
5728  * Waits until "flags" is no longer set in the outer to prevent multiple
5729  * threads from having write access at the same time. SQ_WRITER has to be part
5730  * of "flags".
5731  *
5732  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5733  * until the outer_exit is finished.
5734  *
5735  * outer_enter is vulnerable to starvation since it does not prevent new
5736  * threads from entering the inner syncqs while it is waiting for sq_count to
5737  * go to zero.
5738  */
5739 void
5740 outer_enter(syncq_t *outer, uint16_t flags)
5741 {
5742 	syncq_t	*sq;
5743 	int	wait_needed;
5744 	uint16_t	count;
5745 
5746 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5747 	    outer->sq_oprev != NULL);
5748 	ASSERT(flags & SQ_WRITER);
5749 
5750 retry:
5751 	mutex_enter(SQLOCK(outer));
5752 	while (outer->sq_flags & flags) {
5753 		outer->sq_flags |= SQ_WANTWAKEUP;
5754 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5755 	}
5756 
5757 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5758 	outer->sq_flags |= SQ_WRITER;
5759 	outer->sq_count++;
5760 	ASSERT(outer->sq_count != 0);	/* wraparound */
5761 	wait_needed = 0;
5762 	/*
5763 	 * Set SQ_WRITER on all the inner syncqs while holding
5764 	 * the SQLOCK on the outer syncq. This ensures that the changing
5765 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5766 	 */
5767 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5768 		mutex_enter(SQLOCK(sq));
5769 		count = sq->sq_count;
5770 		SQ_PUTLOCKS_ENTER(sq);
5771 		sq->sq_flags |= SQ_WRITER;
5772 		SUM_SQ_PUTCOUNTS(sq, count);
5773 		if (count != 0)
5774 			wait_needed = 1;
5775 		SQ_PUTLOCKS_EXIT(sq);
5776 		mutex_exit(SQLOCK(sq));
5777 	}
5778 	mutex_exit(SQLOCK(outer));
5779 
5780 	/*
5781 	 * Get everybody out of the syncqs sequentially.
5782 	 * Note that we don't actually need to acquire the PUTLOCKS, since
5783 	 * we have already cleared the fastbit, and set QWRITER.  By
5784 	 * definition, the count can not increase since putnext will
5785 	 * take the slowlock path (and the purpose of acquiring the
5786 	 * putlocks was to make sure it didn't increase while we were
5787 	 * waiting).
5788 	 *
5789 	 * Note that we still acquire the PUTLOCKS to be safe.
5790 	 */
5791 	if (wait_needed) {
5792 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5793 			mutex_enter(SQLOCK(sq));
5794 			count = sq->sq_count;
5795 			SQ_PUTLOCKS_ENTER(sq);
5796 			SUM_SQ_PUTCOUNTS(sq, count);
5797 			while (count != 0) {
5798 				sq->sq_flags |= SQ_WANTWAKEUP;
5799 				SQ_PUTLOCKS_EXIT(sq);
5800 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5801 				count = sq->sq_count;
5802 				SQ_PUTLOCKS_ENTER(sq);
5803 				SUM_SQ_PUTCOUNTS(sq, count);
5804 			}
5805 			SQ_PUTLOCKS_EXIT(sq);
5806 			mutex_exit(SQLOCK(sq));
5807 		}
5808 		/*
5809 		 * Verify that none of the flags got set while we
5810 		 * were waiting for the sq_counts to drop.
5811 		 * If this happens we exit and retry entering the
5812 		 * outer perimeter.
5813 		 */
5814 		mutex_enter(SQLOCK(outer));
5815 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5816 			mutex_exit(SQLOCK(outer));
5817 			outer_exit(outer);
5818 			goto retry;
5819 		}
5820 		mutex_exit(SQLOCK(outer));
5821 	}
5822 }
5823 
5824 /*
5825  * Drop the write access at the outer perimeter.
5826  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5827  * decrementing sq_count.
5828  */
5829 void
5830 outer_exit(syncq_t *outer)
5831 {
5832 	syncq_t	*sq;
5833 	int	 drain_needed;
5834 	uint16_t flags;
5835 
5836 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5837 	    outer->sq_oprev != NULL);
5838 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5839 
5840 	/*
5841 	 * Atomically (from the perspective of threads calling become_writer)
5842 	 * drop the write access at the outer perimeter by holding
5843 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5844 	 * SQ_WRITER.
5845 	 * This defines a locking order between the outer perimeter
5846 	 * SQLOCK and the inner perimeter SQLOCKs.
5847 	 */
5848 	mutex_enter(SQLOCK(outer));
5849 	flags = outer->sq_flags;
5850 	ASSERT(outer->sq_flags & SQ_WRITER);
5851 	if (flags & SQ_QUEUED) {
5852 		write_now(outer);
5853 		flags = outer->sq_flags;
5854 	}
5855 
5856 	/*
5857 	 * sq_onext is stable since sq_count has not yet been decreased.
5858 	 * Reset the SQ_WRITER flags in all syncqs.
5859 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5860 	 * inner syncqs.
5861 	 */
5862 	drain_needed = 0;
5863 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5864 		drain_needed += dropsq(sq, SQ_WRITER);
5865 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5866 	flags &= ~SQ_WRITER;
5867 	if (drain_needed) {
5868 		outer->sq_flags = flags;
5869 		mutex_exit(SQLOCK(outer));
5870 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5871 			emptysq(sq);
5872 		mutex_enter(SQLOCK(outer));
5873 		flags = outer->sq_flags;
5874 	}
5875 	if (flags & SQ_WANTWAKEUP) {
5876 		flags &= ~SQ_WANTWAKEUP;
5877 		cv_broadcast(&outer->sq_wait);
5878 	}
5879 	outer->sq_flags = flags;
5880 	ASSERT(outer->sq_count > 0);
5881 	outer->sq_count--;
5882 	mutex_exit(SQLOCK(outer));
5883 }
5884 
5885 /*
5886  * Add another syncq to an outer perimeter.
5887  * Block out all other access to the outer perimeter while it is being
5888  * changed using blocksq.
5889  * Assumes that the caller has *not* done an outer_enter.
5890  *
5891  * Vulnerable to starvation in blocksq.
5892  */
5893 static void
5894 outer_insert(syncq_t *outer, syncq_t *sq)
5895 {
5896 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5897 	    outer->sq_oprev != NULL);
5898 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5899 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5900 
5901 	/* Get exclusive access to the outer perimeter list */
5902 	blocksq(outer, SQ_BLOCKED, 0);
5903 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5904 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5905 
5906 	mutex_enter(SQLOCK(sq));
5907 	sq->sq_outer = outer;
5908 	outer->sq_onext->sq_oprev = sq;
5909 	sq->sq_onext = outer->sq_onext;
5910 	outer->sq_onext = sq;
5911 	sq->sq_oprev = outer;
5912 	mutex_exit(SQLOCK(sq));
5913 	unblocksq(outer, SQ_BLOCKED, 1);
5914 }
5915 
5916 /*
5917  * Remove a syncq from an outer perimeter.
5918  * Block out all other access to the outer perimeter while it is being
5919  * changed using blocksq.
5920  * Assumes that the caller has *not* done an outer_enter.
5921  *
5922  * Vulnerable to starvation in blocksq.
5923  */
5924 static void
5925 outer_remove(syncq_t *outer, syncq_t *sq)
5926 {
5927 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5928 	    outer->sq_oprev != NULL);
5929 	ASSERT(sq->sq_outer == outer);
5930 
5931 	/* Get exclusive access to the outer perimeter list */
5932 	blocksq(outer, SQ_BLOCKED, 0);
5933 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5934 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5935 
5936 	mutex_enter(SQLOCK(sq));
5937 	sq->sq_outer = NULL;
5938 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5939 	sq->sq_oprev->sq_onext = sq->sq_onext;
5940 	sq->sq_oprev = sq->sq_onext = NULL;
5941 	mutex_exit(SQLOCK(sq));
5942 	unblocksq(outer, SQ_BLOCKED, 1);
5943 }
5944 
5945 /*
5946  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5947  * If this is the first callback for this outer perimeter then add
5948  * this outer perimeter to the list of outer perimeters that
5949  * the qwriter_outer_thread will process.
5950  *
5951  * Increments sq_count in the outer syncq to prevent the membership
5952  * of the outer perimeter (in terms of inner syncqs) to change while
5953  * the callback is pending.
5954  */
5955 static void
5956 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5957 {
5958 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5959 
5960 	mp->b_prev = (mblk_t *)func;
5961 	mp->b_queue = q;
5962 	mp->b_next = NULL;
5963 	outer->sq_count++;	/* Decremented when dequeued */
5964 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5965 	if (outer->sq_evhead == NULL) {
5966 		/* First message. */
5967 		outer->sq_evhead = outer->sq_evtail = mp;
5968 		outer->sq_flags |= SQ_EVENTS;
5969 		mutex_exit(SQLOCK(outer));
5970 		STRSTAT(qwr_outer);
5971 		(void) taskq_dispatch(streams_taskq,
5972 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5973 	} else {
5974 		ASSERT(outer->sq_flags & SQ_EVENTS);
5975 		outer->sq_evtail->b_next = mp;
5976 		outer->sq_evtail = mp;
5977 		mutex_exit(SQLOCK(outer));
5978 	}
5979 }
5980 
5981 /*
5982  * Try and upgrade to write access at the outer perimeter. If this can
5983  * not be done without blocking then queue the callback to be done
5984  * by the qwriter_outer_thread.
5985  *
5986  * This routine can only be called from put or service procedures plus
5987  * asynchronous callback routines that have properly entered the queue (with
5988  * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
5989  * associated with q.
5990  */
5991 void
5992 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5993 {
5994 	syncq_t	*osq, *sq, *outer;
5995 	int	failed;
5996 	uint16_t flags;
5997 
5998 	osq = q->q_syncq;
5999 	outer = osq->sq_outer;
6000 	if (outer == NULL)
6001 		panic("qwriter(PERIM_OUTER): no outer perimeter");
6002 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6003 	    outer->sq_oprev != NULL);
6004 
6005 	mutex_enter(SQLOCK(outer));
6006 	flags = outer->sq_flags;
6007 	/*
6008 	 * If some thread is traversing sq_next, or if we are blocked by
6009 	 * outer_insert or outer_remove, or if the we already have queued
6010 	 * callbacks, then queue this callback for later processing.
6011 	 *
6012 	 * Also queue the qwriter for an interrupt thread in order
6013 	 * to reduce the time spent running at high IPL.
6014 	 * to identify there are events.
6015 	 */
6016 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
6017 		/*
6018 		 * Queue the become_writer request.
6019 		 * The queueing is atomic under SQLOCK(outer) in order
6020 		 * to synchronize with outer_exit.
6021 		 * queue_writer will drop the outer SQLOCK
6022 		 */
6023 		if (flags & SQ_BLOCKED) {
6024 			/* Must set SQ_WRITER on inner perimeter */
6025 			mutex_enter(SQLOCK(osq));
6026 			osq->sq_flags |= SQ_WRITER;
6027 			mutex_exit(SQLOCK(osq));
6028 		} else {
6029 			if (!(flags & SQ_WRITER)) {
6030 				/*
6031 				 * The outer could have been SQ_BLOCKED thus
6032 				 * SQ_WRITER might not be set on the inner.
6033 				 */
6034 				mutex_enter(SQLOCK(osq));
6035 				osq->sq_flags |= SQ_WRITER;
6036 				mutex_exit(SQLOCK(osq));
6037 			}
6038 			ASSERT(osq->sq_flags & SQ_WRITER);
6039 		}
6040 		queue_writer(outer, func, q, mp);
6041 		return;
6042 	}
6043 	/*
6044 	 * We are half-way to exclusive access to the outer perimeter.
6045 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
6046 	 * while the inner syncqs are traversed.
6047 	 */
6048 	outer->sq_count++;
6049 	ASSERT(outer->sq_count != 0);	/* wraparound */
6050 	flags |= SQ_WRITER;
6051 	/*
6052 	 * Check if we can run the function immediately. Mark all
6053 	 * syncqs with the writer flag to prevent new entries into
6054 	 * put and service procedures.
6055 	 *
6056 	 * Set SQ_WRITER on all the inner syncqs while holding
6057 	 * the SQLOCK on the outer syncq. This ensures that the changing
6058 	 * of SQ_WRITER is atomic under the outer SQLOCK.
6059 	 */
6060 	failed = 0;
6061 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6062 		uint16_t count;
6063 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
6064 
6065 		mutex_enter(SQLOCK(sq));
6066 		count = sq->sq_count;
6067 		SQ_PUTLOCKS_ENTER(sq);
6068 		SUM_SQ_PUTCOUNTS(sq, count);
6069 		if (sq->sq_count > maxcnt)
6070 			failed = 1;
6071 		sq->sq_flags |= SQ_WRITER;
6072 		SQ_PUTLOCKS_EXIT(sq);
6073 		mutex_exit(SQLOCK(sq));
6074 	}
6075 	if (failed) {
6076 		/*
6077 		 * Some other thread has a read claim on the outer perimeter.
6078 		 * Queue the callback for deferred processing.
6079 		 *
6080 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6081 		 * so that other qwriter(OUTER) calls will queue their
6082 		 * callbacks as well. queue_writer increments sq_count so we
6083 		 * decrement to compensate for the our increment.
6084 		 *
6085 		 * Dropping SQ_WRITER enables the writer thread to work
6086 		 * on this outer perimeter.
6087 		 */
6088 		outer->sq_flags = flags;
6089 		queue_writer(outer, func, q, mp);
6090 		/* queue_writer dropper the lock */
6091 		mutex_enter(SQLOCK(outer));
6092 		ASSERT(outer->sq_count > 0);
6093 		outer->sq_count--;
6094 		ASSERT(outer->sq_flags & SQ_WRITER);
6095 		flags = outer->sq_flags;
6096 		flags &= ~SQ_WRITER;
6097 		if (flags & SQ_WANTWAKEUP) {
6098 			flags &= ~SQ_WANTWAKEUP;
6099 			cv_broadcast(&outer->sq_wait);
6100 		}
6101 		outer->sq_flags = flags;
6102 		mutex_exit(SQLOCK(outer));
6103 		return;
6104 	} else {
6105 		outer->sq_flags = flags;
6106 		mutex_exit(SQLOCK(outer));
6107 	}
6108 
6109 	/* Can run it immediately */
6110 	(*func)(q, mp);
6111 
6112 	outer_exit(outer);
6113 }
6114 
6115 /*
6116  * Dequeue all writer callbacks from the outer perimeter and run them.
6117  */
6118 static void
6119 write_now(syncq_t *outer)
6120 {
6121 	mblk_t		*mp;
6122 	queue_t		*q;
6123 	void	(*func)();
6124 
6125 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6126 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6127 	    outer->sq_oprev != NULL);
6128 	while ((mp = outer->sq_evhead) != NULL) {
6129 		/*
6130 		 * queues cannot be placed on the queuelist on the outer
6131 		 * perimeter.
6132 		 */
6133 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6134 		ASSERT((outer->sq_flags & SQ_EVENTS));
6135 
6136 		outer->sq_evhead = mp->b_next;
6137 		if (outer->sq_evhead == NULL) {
6138 			outer->sq_evtail = NULL;
6139 			outer->sq_flags &= ~SQ_EVENTS;
6140 		}
6141 		ASSERT(outer->sq_count != 0);
6142 		outer->sq_count--;	/* Incremented when enqueued. */
6143 		mutex_exit(SQLOCK(outer));
6144 		/*
6145 		 * Drop the message if the queue is closing.
6146 		 * Make sure that the queue is "claimed" when the callback
6147 		 * is run in order to satisfy various ASSERTs.
6148 		 */
6149 		q = mp->b_queue;
6150 		func = (void (*)())mp->b_prev;
6151 		ASSERT(func != NULL);
6152 		mp->b_next = mp->b_prev = NULL;
6153 		if (q->q_flag & QWCLOSE) {
6154 			freemsg(mp);
6155 		} else {
6156 			claimq(q);
6157 			(*func)(q, mp);
6158 			releaseq(q);
6159 		}
6160 		mutex_enter(SQLOCK(outer));
6161 	}
6162 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6163 }
6164 
6165 /*
6166  * The list of messages on the inner syncq is effectively hashed
6167  * by destination queue.  These destination queues are doubly
6168  * linked lists (hopefully) in priority order.  Messages are then
6169  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6170  * Additional messages are linked together by the b_next/b_prev
6171  * elements in the mblk, with (similar to putq()) the first message
6172  * having a NULL b_prev and the last message having a NULL b_next.
6173  *
6174  * Events, such as qwriter callbacks, are put onto a list in FIFO
6175  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6176  * linked list, and messages here MUST be processed in the order queued.
6177  */
6178 
6179 /*
6180  * Run the events on the syncq event list (sq_evhead).
6181  * Assumes there is only one claim on the syncq, it is
6182  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6183  * Messages here are processed in order, with the SQ_EXCL bit
6184  * held all the way through till the last message is processed.
6185  */
6186 void
6187 sq_run_events(syncq_t *sq)
6188 {
6189 	mblk_t		*bp;
6190 	queue_t		*qp;
6191 	uint16_t	flags = sq->sq_flags;
6192 	void		(*func)();
6193 
6194 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6195 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6196 	    sq->sq_oprev == NULL) ||
6197 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6198 	    sq->sq_oprev != NULL));
6199 
6200 	ASSERT(flags & SQ_EXCL);
6201 	ASSERT(sq->sq_count == 1);
6202 
6203 	/*
6204 	 * We need to process all of the events on this list.  It
6205 	 * is possible that new events will be added while we are
6206 	 * away processing a callback, so on every loop, we start
6207 	 * back at the beginning of the list.
6208 	 */
6209 	/*
6210 	 * We have to reaccess sq_evhead since there is a
6211 	 * possibility of a new entry while we were running
6212 	 * the callback.
6213 	 */
6214 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6215 		ASSERT(bp->b_queue->q_syncq == sq);
6216 		ASSERT(sq->sq_flags & SQ_EVENTS);
6217 
6218 		qp = bp->b_queue;
6219 		func = (void (*)())bp->b_prev;
6220 		ASSERT(func != NULL);
6221 
6222 		/*
6223 		 * Messages from the event queue must be taken off in
6224 		 * FIFO order.
6225 		 */
6226 		ASSERT(sq->sq_evhead == bp);
6227 		sq->sq_evhead = bp->b_next;
6228 
6229 		if (bp->b_next == NULL) {
6230 			/* Deleting last */
6231 			ASSERT(sq->sq_evtail == bp);
6232 			sq->sq_evtail = NULL;
6233 			sq->sq_flags &= ~SQ_EVENTS;
6234 		}
6235 		bp->b_prev = bp->b_next = NULL;
6236 		ASSERT(bp->b_datap->db_ref != 0);
6237 
6238 		mutex_exit(SQLOCK(sq));
6239 
6240 		(*func)(qp, bp);
6241 
6242 		mutex_enter(SQLOCK(sq));
6243 		/*
6244 		 * re-read the flags, since they could have changed.
6245 		 */
6246 		flags = sq->sq_flags;
6247 		ASSERT(flags & SQ_EXCL);
6248 	}
6249 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6250 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6251 
6252 	if (flags & SQ_WANTWAKEUP) {
6253 		flags &= ~SQ_WANTWAKEUP;
6254 		cv_broadcast(&sq->sq_wait);
6255 	}
6256 	if (flags & SQ_WANTEXWAKEUP) {
6257 		flags &= ~SQ_WANTEXWAKEUP;
6258 		cv_broadcast(&sq->sq_exitwait);
6259 	}
6260 	sq->sq_flags = flags;
6261 }
6262 
6263 /*
6264  * Put messages on the event list.
6265  * If we can go exclusive now, do so and process the event list, otherwise
6266  * let the last claim service this list (or wake the sqthread).
6267  * This procedure assumes SQLOCK is held.  To run the event list, it
6268  * must be called with no claims.
6269  */
6270 static void
6271 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6272 {
6273 	uint16_t count;
6274 
6275 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6276 	ASSERT(func != NULL);
6277 
6278 	/*
6279 	 * This is a callback.  Add it to the list of callbacks
6280 	 * and see about upgrading.
6281 	 */
6282 	mp->b_prev = (mblk_t *)func;
6283 	mp->b_queue = q;
6284 	mp->b_next = NULL;
6285 	if (sq->sq_evhead == NULL) {
6286 		sq->sq_evhead = sq->sq_evtail = mp;
6287 		sq->sq_flags |= SQ_EVENTS;
6288 	} else {
6289 		ASSERT(sq->sq_evtail != NULL);
6290 		ASSERT(sq->sq_evtail->b_next == NULL);
6291 		ASSERT(sq->sq_flags & SQ_EVENTS);
6292 		sq->sq_evtail->b_next = mp;
6293 		sq->sq_evtail = mp;
6294 	}
6295 	/*
6296 	 * We have set SQ_EVENTS, so threads will have to
6297 	 * unwind out of the perimeter, and new entries will
6298 	 * not grab a putlock.  But we still need to know
6299 	 * how many threads have already made a claim to the
6300 	 * syncq, so grab the putlocks, and sum the counts.
6301 	 * If there are no claims on the syncq, we can upgrade
6302 	 * to exclusive, and run the event list.
6303 	 * NOTE: We hold the SQLOCK, so we can just grab the
6304 	 * putlocks.
6305 	 */
6306 	count = sq->sq_count;
6307 	SQ_PUTLOCKS_ENTER(sq);
6308 	SUM_SQ_PUTCOUNTS(sq, count);
6309 	/*
6310 	 * We have no claim, so we need to check if there
6311 	 * are no others, then we can upgrade.
6312 	 */
6313 	/*
6314 	 * There are currently no claims on
6315 	 * the syncq by this thread (at least on this entry). The thread who has
6316 	 * the claim should drain syncq.
6317 	 */
6318 	if (count > 0) {
6319 		/*
6320 		 * Can't upgrade - other threads inside.
6321 		 */
6322 		SQ_PUTLOCKS_EXIT(sq);
6323 		mutex_exit(SQLOCK(sq));
6324 		return;
6325 	}
6326 	/*
6327 	 * Need to set SQ_EXCL and make a claim on the syncq.
6328 	 */
6329 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6330 	sq->sq_flags |= SQ_EXCL;
6331 	ASSERT(sq->sq_count == 0);
6332 	sq->sq_count++;
6333 	SQ_PUTLOCKS_EXIT(sq);
6334 
6335 	/* Process the events list */
6336 	sq_run_events(sq);
6337 
6338 	/*
6339 	 * Release our claim...
6340 	 */
6341 	sq->sq_count--;
6342 
6343 	/*
6344 	 * And release SQ_EXCL.
6345 	 * We don't need to acquire the putlocks to release
6346 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6347 	 */
6348 	sq->sq_flags &= ~SQ_EXCL;
6349 
6350 	/*
6351 	 * sq_run_events should have released SQ_EXCL
6352 	 */
6353 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6354 
6355 	/*
6356 	 * If anything happened while we were running the
6357 	 * events (or was there before), we need to process
6358 	 * them now.  We shouldn't be exclusive sine we
6359 	 * released the perimeter above (plus, we asserted
6360 	 * for it).
6361 	 */
6362 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6363 		drain_syncq(sq);
6364 	else
6365 		mutex_exit(SQLOCK(sq));
6366 }
6367 
6368 /*
6369  * Perform delayed processing. The caller has to make sure that it is safe
6370  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6371  * set).
6372  *
6373  * Assume that the caller has NO claims on the syncq.  However, a claim
6374  * on the syncq does not indicate that a thread is draining the syncq.
6375  * There may be more claims on the syncq than there are threads draining
6376  * (i.e.  #_threads_draining <= sq_count)
6377  *
6378  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6379  * in order to preserve qwriter(OUTER) ordering constraints.
6380  *
6381  * sq_putcount only needs to be checked when dispatching the queued
6382  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6383  */
6384 void
6385 drain_syncq(syncq_t *sq)
6386 {
6387 	queue_t		*qp;
6388 	uint16_t	count;
6389 	uint16_t	type = sq->sq_type;
6390 	uint16_t	flags = sq->sq_flags;
6391 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6392 
6393 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6394 	    "drain_syncq start:%p", sq);
6395 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6396 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6397 	    sq->sq_oprev == NULL) ||
6398 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6399 	    sq->sq_oprev != NULL));
6400 
6401 	/*
6402 	 * Drop SQ_SERVICE flag.
6403 	 */
6404 	if (bg_service)
6405 		sq->sq_svcflags &= ~SQ_SERVICE;
6406 
6407 	/*
6408 	 * If SQ_EXCL is set, someone else is processing this syncq - let them
6409 	 * finish the job.
6410 	 */
6411 	if (flags & SQ_EXCL) {
6412 		if (bg_service) {
6413 			ASSERT(sq->sq_servcount != 0);
6414 			sq->sq_servcount--;
6415 		}
6416 		mutex_exit(SQLOCK(sq));
6417 		return;
6418 	}
6419 
6420 	/*
6421 	 * This routine can be called by a background thread if
6422 	 * it was scheduled by a hi-priority thread.  SO, if there are
6423 	 * NOT messages queued, return (remember, we have the SQLOCK,
6424 	 * and it cannot change until we release it). Wakeup any waiters also.
6425 	 */
6426 	if (!(flags & SQ_QUEUED)) {
6427 		if (flags & SQ_WANTWAKEUP) {
6428 			flags &= ~SQ_WANTWAKEUP;
6429 			cv_broadcast(&sq->sq_wait);
6430 		}
6431 		if (flags & SQ_WANTEXWAKEUP) {
6432 			flags &= ~SQ_WANTEXWAKEUP;
6433 			cv_broadcast(&sq->sq_exitwait);
6434 		}
6435 		sq->sq_flags = flags;
6436 		if (bg_service) {
6437 			ASSERT(sq->sq_servcount != 0);
6438 			sq->sq_servcount--;
6439 		}
6440 		mutex_exit(SQLOCK(sq));
6441 		return;
6442 	}
6443 
6444 	/*
6445 	 * If this is not a concurrent put perimeter, we need to
6446 	 * become exclusive to drain.  Also, if not CIPUT, we would
6447 	 * not have acquired a putlock, so we don't need to check
6448 	 * the putcounts.  If not entering with a claim, we test
6449 	 * for sq_count == 0.
6450 	 */
6451 	type = sq->sq_type;
6452 	if (!(type & SQ_CIPUT)) {
6453 		if (sq->sq_count > 1) {
6454 			if (bg_service) {
6455 				ASSERT(sq->sq_servcount != 0);
6456 				sq->sq_servcount--;
6457 			}
6458 			mutex_exit(SQLOCK(sq));
6459 			return;
6460 		}
6461 		sq->sq_flags |= SQ_EXCL;
6462 	}
6463 
6464 	/*
6465 	 * This is where we make a claim to the syncq.
6466 	 * This can either be done by incrementing a putlock, or
6467 	 * the sq_count.  But since we already have the SQLOCK
6468 	 * here, we just bump the sq_count.
6469 	 *
6470 	 * Note that after we make a claim, we need to let the code
6471 	 * fall through to the end of this routine to clean itself
6472 	 * up.  A return in the while loop will put the syncq in a
6473 	 * very bad state.
6474 	 */
6475 	sq->sq_count++;
6476 	ASSERT(sq->sq_count != 0);	/* wraparound */
6477 
6478 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6479 		/*
6480 		 * If we are told to stayaway or went exclusive,
6481 		 * we are done.
6482 		 */
6483 		if (flags & (SQ_STAYAWAY)) {
6484 			break;
6485 		}
6486 
6487 		/*
6488 		 * If there are events to run, do so.
6489 		 * We have one claim to the syncq, so if there are
6490 		 * more than one, other threads are running.
6491 		 */
6492 		if (sq->sq_evhead != NULL) {
6493 			ASSERT(sq->sq_flags & SQ_EVENTS);
6494 
6495 			count = sq->sq_count;
6496 			SQ_PUTLOCKS_ENTER(sq);
6497 			SUM_SQ_PUTCOUNTS(sq, count);
6498 			if (count > 1) {
6499 				SQ_PUTLOCKS_EXIT(sq);
6500 				/* Can't upgrade - other threads inside */
6501 				break;
6502 			}
6503 			ASSERT((flags & SQ_EXCL) == 0);
6504 			sq->sq_flags = flags | SQ_EXCL;
6505 			SQ_PUTLOCKS_EXIT(sq);
6506 			/*
6507 			 * we have the only claim, run the events,
6508 			 * sq_run_events will clear the SQ_EXCL flag.
6509 			 */
6510 			sq_run_events(sq);
6511 
6512 			/*
6513 			 * If this is a CIPUT perimeter, we need
6514 			 * to drop the SQ_EXCL flag so we can properly
6515 			 * continue draining the syncq.
6516 			 */
6517 			if (type & SQ_CIPUT) {
6518 				ASSERT(sq->sq_flags & SQ_EXCL);
6519 				sq->sq_flags &= ~SQ_EXCL;
6520 			}
6521 
6522 			/*
6523 			 * And go back to the beginning just in case
6524 			 * anything changed while we were away.
6525 			 */
6526 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6527 			continue;
6528 		}
6529 
6530 		ASSERT(sq->sq_evhead == NULL);
6531 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6532 
6533 		/*
6534 		 * Find the queue that is not draining.
6535 		 *
6536 		 * q_draining is protected by QLOCK which we do not hold.
6537 		 * But if it was set, then a thread was draining, and if it gets
6538 		 * cleared, then it was because the thread has successfully
6539 		 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6540 		 * state to happen, a thread needs the SQLOCK which we hold, and
6541 		 * if there was such a flag, we would have already seen it.
6542 		 */
6543 
6544 		for (qp = sq->sq_head;
6545 		    qp != NULL && (qp->q_draining ||
6546 		    (qp->q_sqflags & Q_SQDRAINING));
6547 		    qp = qp->q_sqnext)
6548 			;
6549 
6550 		if (qp == NULL)
6551 			break;
6552 
6553 		/*
6554 		 * We have a queue to work on, and we hold the
6555 		 * SQLOCK and one claim, call qdrain_syncq.
6556 		 * This means we need to release the SQLOCK and
6557 		 * acquire the QLOCK (OK since we have a claim).
6558 		 * Note that qdrain_syncq will actually dequeue
6559 		 * this queue from the sq_head list when it is
6560 		 * convinced all the work is done and release
6561 		 * the QLOCK before returning.
6562 		 */
6563 		qp->q_sqflags |= Q_SQDRAINING;
6564 		mutex_exit(SQLOCK(sq));
6565 		mutex_enter(QLOCK(qp));
6566 		qdrain_syncq(sq, qp);
6567 		mutex_enter(SQLOCK(sq));
6568 
6569 		/* The queue is drained */
6570 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6571 		qp->q_sqflags &= ~Q_SQDRAINING;
6572 		/*
6573 		 * NOTE: After this point qp should not be used since it may be
6574 		 * closed.
6575 		 */
6576 	}
6577 
6578 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6579 	flags = sq->sq_flags;
6580 
6581 	/*
6582 	 * sq->sq_head cannot change because we hold the
6583 	 * sqlock. However, a thread CAN decide that it is no longer
6584 	 * going to drain that queue.  However, this should be due to
6585 	 * a GOAWAY state, and we should see that here.
6586 	 *
6587 	 * This loop is not very efficient. One solution may be adding a second
6588 	 * pointer to the "draining" queue, but it is difficult to do when
6589 	 * queues are inserted in the middle due to priority ordering. Another
6590 	 * possibility is to yank the queue out of the sq list and put it onto
6591 	 * the "draining list" and then put it back if it can't be drained.
6592 	 */
6593 
6594 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6595 	    (type & SQ_CI) || sq->sq_head->q_draining);
6596 
6597 	/* Drop SQ_EXCL for non-CIPUT perimeters */
6598 	if (!(type & SQ_CIPUT))
6599 		flags &= ~SQ_EXCL;
6600 	ASSERT((flags & SQ_EXCL) == 0);
6601 
6602 	/* Wake up any waiters. */
6603 	if (flags & SQ_WANTWAKEUP) {
6604 		flags &= ~SQ_WANTWAKEUP;
6605 		cv_broadcast(&sq->sq_wait);
6606 	}
6607 	if (flags & SQ_WANTEXWAKEUP) {
6608 		flags &= ~SQ_WANTEXWAKEUP;
6609 		cv_broadcast(&sq->sq_exitwait);
6610 	}
6611 	sq->sq_flags = flags;
6612 
6613 	ASSERT(sq->sq_count != 0);
6614 	/* Release our claim. */
6615 	sq->sq_count--;
6616 
6617 	if (bg_service) {
6618 		ASSERT(sq->sq_servcount != 0);
6619 		sq->sq_servcount--;
6620 	}
6621 
6622 	mutex_exit(SQLOCK(sq));
6623 
6624 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6625 	    "drain_syncq end:%p", sq);
6626 }
6627 
6628 
6629 /*
6630  *
6631  * qdrain_syncq can be called (currently) from only one of two places:
6632  *	drain_syncq
6633  * 	putnext  (or some variation of it).
6634  * and eventually
6635  * 	qwait(_sig)
6636  *
6637  * If called from drain_syncq, we found it in the list of queues needing
6638  * service, so there is work to be done (or it wouldn't be in the list).
6639  *
6640  * If called from some putnext variation, it was because the
6641  * perimeter is open, but messages are blocking a putnext and
6642  * there is not a thread working on it.  Now a thread could start
6643  * working on it while we are getting ready to do so ourself, but
6644  * the thread would set the q_draining flag, and we can spin out.
6645  *
6646  * As for qwait(_sig), I think I shall let it continue to call
6647  * drain_syncq directly (after all, it will get here eventually).
6648  *
6649  * qdrain_syncq has to terminate when:
6650  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6651  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6652  *
6653  * ASSUMES:
6654  *	One claim
6655  * 	QLOCK held
6656  * 	SQLOCK not held
6657  *	Will release QLOCK before returning
6658  */
6659 void
6660 qdrain_syncq(syncq_t *sq, queue_t *q)
6661 {
6662 	mblk_t		*bp;
6663 #ifdef DEBUG
6664 	uint16_t	count;
6665 #endif
6666 
6667 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6668 	    "drain_syncq start:%p", sq);
6669 	ASSERT(q->q_syncq == sq);
6670 	ASSERT(MUTEX_HELD(QLOCK(q)));
6671 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6672 	/*
6673 	 * For non-CIPUT perimeters, we should be called with the exclusive bit
6674 	 * set already. For CIPUT perimeters, we will be doing a concurrent
6675 	 * drain, so it better not be set.
6676 	 */
6677 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6678 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6679 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6680 	/*
6681 	 * All outer pointers are set, or none of them are
6682 	 */
6683 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6684 	    sq->sq_oprev == NULL) ||
6685 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6686 	    sq->sq_oprev != NULL));
6687 #ifdef DEBUG
6688 	count = sq->sq_count;
6689 	/*
6690 	 * This is OK without the putlocks, because we have one
6691 	 * claim either from the sq_count, or a putcount.  We could
6692 	 * get an erroneous value from other counts, but ours won't
6693 	 * change, so one way or another, we will have at least a
6694 	 * value of one.
6695 	 */
6696 	SUM_SQ_PUTCOUNTS(sq, count);
6697 	ASSERT(count >= 1);
6698 #endif /* DEBUG */
6699 
6700 	/*
6701 	 * The first thing to do is find out if a thread is already draining
6702 	 * this queue. If so, we are done, just return.
6703 	 */
6704 	if (q->q_draining) {
6705 		mutex_exit(QLOCK(q));
6706 		return;
6707 	}
6708 
6709 	/*
6710 	 * If the perimeter is exclusive, there is nothing we can do right now,
6711 	 * go away. Note that there is nothing to prevent this case from
6712 	 * changing right after this check, but the spin-out will catch it.
6713 	 */
6714 
6715 	/* Tell other threads that we are draining this queue */
6716 	q->q_draining = 1;	/* Protected by QLOCK */
6717 
6718 	/*
6719 	 * If there is nothing to do, clear QFULL as necessary. This caters for
6720 	 * the case where an empty queue was enqueued onto the syncq.
6721 	 */
6722 	if (q->q_sqhead == NULL) {
6723 		ASSERT(q->q_syncqmsgs == 0);
6724 		mutex_exit(QLOCK(q));
6725 		clr_qfull(q);
6726 		mutex_enter(QLOCK(q));
6727 	}
6728 
6729 	/*
6730 	 * Note that q_sqhead must be re-checked here in case another message
6731 	 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6732 	 */
6733 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6734 		/*
6735 		 * Because we can enter this routine just because a putnext is
6736 		 * blocked, we need to spin out if the perimeter wants to go
6737 		 * exclusive as well as just blocked. We need to spin out also
6738 		 * if events are queued on the syncq.
6739 		 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6740 		 * set it, and it can't become exclusive while we hold a claim.
6741 		 */
6742 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6743 			break;
6744 		}
6745 
6746 #ifdef DEBUG
6747 		/*
6748 		 * Since we are in qdrain_syncq, we already know the queue,
6749 		 * but for sanity, we want to check this against the qp that
6750 		 * was passed in by bp->b_queue.
6751 		 */
6752 
6753 		ASSERT(bp->b_queue == q);
6754 		ASSERT(bp->b_queue->q_syncq == sq);
6755 		bp->b_queue = NULL;
6756 
6757 		/*
6758 		 * We would have the following check in the DEBUG code:
6759 		 *
6760 		 * if (bp->b_prev != NULL)  {
6761 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6762 		 * }
6763 		 *
6764 		 * This can't be done, however, since IP modifies qinfo
6765 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6766 		 * qinfo), invalidating the check.
6767 		 * So the assignment to func is left here, but the ASSERT itself
6768 		 * is removed until the whole issue is resolved.
6769 		 */
6770 #endif
6771 		ASSERT(q->q_sqhead == bp);
6772 		q->q_sqhead = bp->b_next;
6773 		bp->b_prev = bp->b_next = NULL;
6774 		ASSERT(q->q_syncqmsgs > 0);
6775 		mutex_exit(QLOCK(q));
6776 
6777 		ASSERT(bp->b_datap->db_ref != 0);
6778 
6779 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6780 
6781 		mutex_enter(QLOCK(q));
6782 
6783 		/*
6784 		 * q_syncqmsgs should only be decremented after executing the
6785 		 * put procedure to avoid message re-ordering. This is due to an
6786 		 * optimisation in putnext() which can call the put procedure
6787 		 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6788 		 * being set).
6789 		 *
6790 		 * We also need to clear QFULL in the next service procedure
6791 		 * queue if this is the last message destined for that queue.
6792 		 *
6793 		 * It would make better sense to have some sort of tunable for
6794 		 * the low water mark, but these semantics are not yet defined.
6795 		 * So, alas, we use a constant.
6796 		 */
6797 		if (--q->q_syncqmsgs == 0) {
6798 			mutex_exit(QLOCK(q));
6799 			clr_qfull(q);
6800 			mutex_enter(QLOCK(q));
6801 		}
6802 
6803 		/*
6804 		 * Always clear SQ_EXCL when CIPUT in order to handle
6805 		 * qwriter(INNER). The putp() can call qwriter and get exclusive
6806 		 * access IFF this is the only claim. So, we need to test for
6807 		 * this possibility, acquire the mutex and clear the bit.
6808 		 */
6809 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6810 			mutex_enter(SQLOCK(sq));
6811 			sq->sq_flags &= ~SQ_EXCL;
6812 			mutex_exit(SQLOCK(sq));
6813 		}
6814 	}
6815 
6816 	/*
6817 	 * We should either have no messages on this queue, or we were told to
6818 	 * goaway by a waiter (which we will wake up at the end of this
6819 	 * function).
6820 	 */
6821 	ASSERT((q->q_sqhead == NULL) ||
6822 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6823 
6824 	ASSERT(MUTEX_HELD(QLOCK(q)));
6825 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6826 
6827 	/* Remove the q from the syncq list if all the messages are drained. */
6828 	if (q->q_sqhead == NULL) {
6829 		ASSERT(q->q_syncqmsgs == 0);
6830 		mutex_enter(SQLOCK(sq));
6831 		if (q->q_sqflags & Q_SQQUEUED)
6832 			SQRM_Q(sq, q);
6833 		mutex_exit(SQLOCK(sq));
6834 		/*
6835 		 * Since the queue is removed from the list, reset its priority.
6836 		 */
6837 		q->q_spri = 0;
6838 	}
6839 
6840 	/*
6841 	 * Remember, the q_draining flag is used to let another thread know
6842 	 * that there is a thread currently draining the messages for a queue.
6843 	 * Since we are now done with this queue (even if there may be messages
6844 	 * still there), we need to clear this flag so some thread will work on
6845 	 * it if needed.
6846 	 */
6847 	ASSERT(q->q_draining);
6848 	q->q_draining = 0;
6849 
6850 	/* Called with a claim, so OK to drop all locks. */
6851 	mutex_exit(QLOCK(q));
6852 
6853 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6854 	    "drain_syncq end:%p", sq);
6855 }
6856 /* END OF QDRAIN_SYNCQ  */
6857 
6858 
6859 /*
6860  * This is the mate to qdrain_syncq, except that it is putting the message onto
6861  * the queue instead of draining. Since the message is destined for the queue
6862  * that is selected, there is no need to identify the function because the
6863  * message is intended for the put routine for the queue. For debug kernels,
6864  * this routine will do it anyway just in case.
6865  *
6866  * After the message is enqueued on the syncq, it calls putnext_tail()
6867  * which will schedule a background thread to actually process the message.
6868  *
6869  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6870  * SQLOCK(sq) and QLOCK(q) are not held.
6871  */
6872 void
6873 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6874 {
6875 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6876 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6877 	ASSERT(sq->sq_count > 0);
6878 	ASSERT(q->q_syncq == sq);
6879 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6880 	    sq->sq_oprev == NULL) ||
6881 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6882 	    sq->sq_oprev != NULL));
6883 
6884 	mutex_enter(QLOCK(q));
6885 
6886 #ifdef DEBUG
6887 	/*
6888 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6889 	 * to trace the queue that the message is intended for.  Note
6890 	 * that the original use was to identify the queue and function
6891 	 * to call on the drain.  In the new syncq, we have the context
6892 	 * of the queue that we are draining, so call it's putproc and
6893 	 * don't rely on the saved values.  But for debug this is still
6894 	 * useful information.
6895 	 */
6896 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6897 	mp->b_queue = q;
6898 	mp->b_next = NULL;
6899 #endif
6900 	ASSERT(q->q_syncq == sq);
6901 	/*
6902 	 * Enqueue the message on the list.
6903 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6904 	 * protect it.  So it's ok to acquire SQLOCK after SQPUT_MP().
6905 	 */
6906 	SQPUT_MP(q, mp);
6907 	mutex_enter(SQLOCK(sq));
6908 
6909 	/*
6910 	 * And queue on syncq for scheduling, if not already queued.
6911 	 * Note that we need the SQLOCK for this, and for testing flags
6912 	 * at the end to see if we will drain.  So grab it now, and
6913 	 * release it before we call qdrain_syncq or return.
6914 	 */
6915 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6916 		q->q_spri = curthread->t_pri;
6917 		SQPUT_Q(sq, q);
6918 	}
6919 #ifdef DEBUG
6920 	else {
6921 		/*
6922 		 * All of these conditions MUST be true!
6923 		 */
6924 		ASSERT(sq->sq_tail != NULL);
6925 		if (sq->sq_tail == sq->sq_head) {
6926 			ASSERT((q->q_sqprev == NULL) &&
6927 			    (q->q_sqnext == NULL));
6928 		} else {
6929 			ASSERT((q->q_sqprev != NULL) ||
6930 			    (q->q_sqnext != NULL));
6931 		}
6932 		ASSERT(sq->sq_flags & SQ_QUEUED);
6933 		ASSERT(q->q_syncqmsgs != 0);
6934 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6935 	}
6936 #endif
6937 	mutex_exit(QLOCK(q));
6938 	/*
6939 	 * SQLOCK is still held, so sq_count can be safely decremented.
6940 	 */
6941 	sq->sq_count--;
6942 
6943 	putnext_tail(sq, q, 0);
6944 	/* Should not reference sq or q after this point. */
6945 }
6946 
6947 /*  End of qfill_syncq  */
6948 
6949 /*
6950  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6951  * that would be put into qp by drain_syncq.
6952  * Used when deleting the syncq (qp == NULL) or when detaching
6953  * a queue (qp != NULL).
6954  * Return non-zero if one or more messages were freed.
6955  *
6956  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
6957  * sq_putlocks are used.
6958  *
6959  * NOTE: This function assumes that it is called from the close() context and
6960  * that all the queues in the syncq are going away. For this reason it doesn't
6961  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6962  * currently valid, but it is useful to rethink this function to behave properly
6963  * in other cases.
6964  */
6965 int
6966 flush_syncq(syncq_t *sq, queue_t *qp)
6967 {
6968 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6969 	queue_t		*q;
6970 	int		ret = 0;
6971 
6972 	mutex_enter(SQLOCK(sq));
6973 
6974 	/*
6975 	 * Before we leave, we need to make sure there are no
6976 	 * events listed for this queue.  All events for this queue
6977 	 * will just be freed.
6978 	 */
6979 	if (qp != NULL && sq->sq_evhead != NULL) {
6980 		ASSERT(sq->sq_flags & SQ_EVENTS);
6981 
6982 		mp_prev = NULL;
6983 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6984 			mp_next = bp->b_next;
6985 			if (bp->b_queue == qp) {
6986 				/* Delete this message */
6987 				if (mp_prev != NULL) {
6988 					mp_prev->b_next = mp_next;
6989 					/*
6990 					 * Update sq_evtail if the last element
6991 					 * is removed.
6992 					 */
6993 					if (bp == sq->sq_evtail) {
6994 						ASSERT(mp_next == NULL);
6995 						sq->sq_evtail = mp_prev;
6996 					}
6997 				} else
6998 					sq->sq_evhead = mp_next;
6999 				if (sq->sq_evhead == NULL)
7000 					sq->sq_flags &= ~SQ_EVENTS;
7001 				bp->b_prev = bp->b_next = NULL;
7002 				freemsg(bp);
7003 				ret++;
7004 			} else {
7005 				mp_prev = bp;
7006 			}
7007 		}
7008 	}
7009 
7010 	/*
7011 	 * Walk sq_head and:
7012 	 *	- match qp if qp is set, remove it's messages
7013 	 *	- all if qp is not set
7014 	 */
7015 	q = sq->sq_head;
7016 	while (q != NULL) {
7017 		ASSERT(q->q_syncq == sq);
7018 		if ((qp == NULL) || (qp == q)) {
7019 			/*
7020 			 * Yank the messages as a list off the queue
7021 			 */
7022 			mp_head = q->q_sqhead;
7023 			/*
7024 			 * We do not have QLOCK(q) here (which is safe due to
7025 			 * assumptions mentioned above). To obtain the lock we
7026 			 * need to release SQLOCK which may allow lots of things
7027 			 * to change upon us. This place requires more analysis.
7028 			 */
7029 			q->q_sqhead = q->q_sqtail = NULL;
7030 			ASSERT(mp_head->b_queue &&
7031 			    mp_head->b_queue->q_syncq == sq);
7032 
7033 			/*
7034 			 * Free each of the messages.
7035 			 */
7036 			for (bp = mp_head; bp != NULL; bp = mp_next) {
7037 				mp_next = bp->b_next;
7038 				bp->b_prev = bp->b_next = NULL;
7039 				freemsg(bp);
7040 				ret++;
7041 			}
7042 			/*
7043 			 * Now remove the queue from the syncq.
7044 			 */
7045 			ASSERT(q->q_sqflags & Q_SQQUEUED);
7046 			SQRM_Q(sq, q);
7047 			q->q_spri = 0;
7048 			q->q_syncqmsgs = 0;
7049 
7050 			/*
7051 			 * If qp was specified, we are done with it and are
7052 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
7053 			 * waiters while we still have the SQLOCK.
7054 			 */
7055 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7056 				sq->sq_flags &= ~SQ_WANTWAKEUP;
7057 				cv_broadcast(&sq->sq_wait);
7058 			}
7059 			/* Drop SQLOCK across clr_qfull */
7060 			mutex_exit(SQLOCK(sq));
7061 
7062 			/*
7063 			 * We avoid doing the test that drain_syncq does and
7064 			 * unconditionally clear qfull for every flushed
7065 			 * message. Since flush_syncq is only called during
7066 			 * close this should not be a problem.
7067 			 */
7068 			clr_qfull(q);
7069 			if (qp != NULL) {
7070 				return (ret);
7071 			} else {
7072 				mutex_enter(SQLOCK(sq));
7073 				/*
7074 				 * The head was removed by SQRM_Q above.
7075 				 * reread the new head and flush it.
7076 				 */
7077 				q = sq->sq_head;
7078 			}
7079 		} else {
7080 			q = q->q_sqnext;
7081 		}
7082 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
7083 	}
7084 
7085 	if (sq->sq_flags & SQ_WANTWAKEUP) {
7086 		sq->sq_flags &= ~SQ_WANTWAKEUP;
7087 		cv_broadcast(&sq->sq_wait);
7088 	}
7089 
7090 	mutex_exit(SQLOCK(sq));
7091 	return (ret);
7092 }
7093 
7094 /*
7095  * Propagate all messages from a syncq to the next syncq that are associated
7096  * with the specified queue. If the queue is attached to a driver or if the
7097  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7098  *
7099  * Assumes that the stream is strlock()'ed. We don't come here if there
7100  * are no messages to propagate.
7101  *
7102  * NOTE : If the queue is attached to a driver, all the messages are freed
7103  * as there is no point in propagating the messages from the driver syncq
7104  * to the closing stream head which will in turn get freed later.
7105  */
7106 static int
7107 propagate_syncq(queue_t *qp)
7108 {
7109 	mblk_t		*bp, *head, *tail, *prev, *next;
7110 	syncq_t 	*sq;
7111 	queue_t		*nqp;
7112 	syncq_t		*nsq;
7113 	boolean_t	isdriver;
7114 	int 		moved = 0;
7115 	uint16_t	flags;
7116 	pri_t		priority = curthread->t_pri;
7117 #ifdef DEBUG
7118 	void		(*func)();
7119 #endif
7120 
7121 	sq = qp->q_syncq;
7122 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7123 	/* debug macro */
7124 	SQ_PUTLOCKS_HELD(sq);
7125 	/*
7126 	 * As entersq() does not increment the sq_count for
7127 	 * the write side, check sq_count for non-QPERQ
7128 	 * perimeters alone.
7129 	 */
7130 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7131 
7132 	/*
7133 	 * propagate_syncq() can be called because of either messages on the
7134 	 * queue syncq or because on events on the queue syncq. Do actual
7135 	 * message propagations if there are any messages.
7136 	 */
7137 	if (qp->q_syncqmsgs) {
7138 		isdriver = (qp->q_flag & QISDRV);
7139 
7140 		if (!isdriver) {
7141 			nqp = qp->q_next;
7142 			nsq = nqp->q_syncq;
7143 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7144 			/* debug macro */
7145 			SQ_PUTLOCKS_HELD(nsq);
7146 #ifdef DEBUG
7147 			func = (void (*)())nqp->q_qinfo->qi_putp;
7148 #endif
7149 		}
7150 
7151 		SQRM_Q(sq, qp);
7152 		priority = MAX(qp->q_spri, priority);
7153 		qp->q_spri = 0;
7154 		head = qp->q_sqhead;
7155 		tail = qp->q_sqtail;
7156 		qp->q_sqhead = qp->q_sqtail = NULL;
7157 		qp->q_syncqmsgs = 0;
7158 
7159 		/*
7160 		 * Walk the list of messages, and free them if this is a driver,
7161 		 * otherwise reset the b_prev and b_queue value to the new putp.
7162 		 * Afterward, we will just add the head to the end of the next
7163 		 * syncq, and point the tail to the end of this one.
7164 		 */
7165 
7166 		for (bp = head; bp != NULL; bp = next) {
7167 			next = bp->b_next;
7168 			if (isdriver) {
7169 				bp->b_prev = bp->b_next = NULL;
7170 				freemsg(bp);
7171 				continue;
7172 			}
7173 			/* Change the q values for this message */
7174 			bp->b_queue = nqp;
7175 #ifdef DEBUG
7176 			bp->b_prev = (mblk_t *)func;
7177 #endif
7178 			moved++;
7179 		}
7180 		/*
7181 		 * Attach list of messages to the end of the new queue (if there
7182 		 * is a list of messages).
7183 		 */
7184 
7185 		if (!isdriver && head != NULL) {
7186 			ASSERT(tail != NULL);
7187 			if (nqp->q_sqhead == NULL) {
7188 				nqp->q_sqhead = head;
7189 			} else {
7190 				ASSERT(nqp->q_sqtail != NULL);
7191 				nqp->q_sqtail->b_next = head;
7192 			}
7193 			nqp->q_sqtail = tail;
7194 			/*
7195 			 * When messages are moved from high priority queue to
7196 			 * another queue, the destination queue priority is
7197 			 * upgraded.
7198 			 */
7199 
7200 			if (priority > nqp->q_spri)
7201 				nqp->q_spri = priority;
7202 
7203 			SQPUT_Q(nsq, nqp);
7204 
7205 			nqp->q_syncqmsgs += moved;
7206 			ASSERT(nqp->q_syncqmsgs != 0);
7207 		}
7208 	}
7209 
7210 	/*
7211 	 * Before we leave, we need to make sure there are no
7212 	 * events listed for this queue.  All events for this queue
7213 	 * will just be freed.
7214 	 */
7215 	if (sq->sq_evhead != NULL) {
7216 		ASSERT(sq->sq_flags & SQ_EVENTS);
7217 		prev = NULL;
7218 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7219 			next = bp->b_next;
7220 			if (bp->b_queue == qp) {
7221 				/* Delete this message */
7222 				if (prev != NULL) {
7223 					prev->b_next = next;
7224 					/*
7225 					 * Update sq_evtail if the last element
7226 					 * is removed.
7227 					 */
7228 					if (bp == sq->sq_evtail) {
7229 						ASSERT(next == NULL);
7230 						sq->sq_evtail = prev;
7231 					}
7232 				} else
7233 					sq->sq_evhead = next;
7234 				if (sq->sq_evhead == NULL)
7235 					sq->sq_flags &= ~SQ_EVENTS;
7236 				bp->b_prev = bp->b_next = NULL;
7237 				freemsg(bp);
7238 			} else {
7239 				prev = bp;
7240 			}
7241 		}
7242 	}
7243 
7244 	flags = sq->sq_flags;
7245 
7246 	/* Wake up any waiter before leaving. */
7247 	if (flags & SQ_WANTWAKEUP) {
7248 		flags &= ~SQ_WANTWAKEUP;
7249 		cv_broadcast(&sq->sq_wait);
7250 	}
7251 	sq->sq_flags = flags;
7252 
7253 	return (moved);
7254 }
7255 
7256 /*
7257  * Try and upgrade to exclusive access at the inner perimeter. If this can
7258  * not be done without blocking then request will be queued on the syncq
7259  * and drain_syncq will run it later.
7260  *
7261  * This routine can only be called from put or service procedures plus
7262  * asynchronous callback routines that have properly entered the queue (with
7263  * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7264  * associated with q.
7265  */
7266 void
7267 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7268 {
7269 	syncq_t	*sq = q->q_syncq;
7270 	uint16_t count;
7271 
7272 	mutex_enter(SQLOCK(sq));
7273 	count = sq->sq_count;
7274 	SQ_PUTLOCKS_ENTER(sq);
7275 	SUM_SQ_PUTCOUNTS(sq, count);
7276 	ASSERT(count >= 1);
7277 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7278 
7279 	if (count == 1) {
7280 		/*
7281 		 * Can upgrade. This case also handles nested qwriter calls
7282 		 * (when the qwriter callback function calls qwriter). In that
7283 		 * case SQ_EXCL is already set.
7284 		 */
7285 		sq->sq_flags |= SQ_EXCL;
7286 		SQ_PUTLOCKS_EXIT(sq);
7287 		mutex_exit(SQLOCK(sq));
7288 		(*func)(q, mp);
7289 		/*
7290 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7291 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7292 		 * until putnext, leavesq, or drain_syncq drops it.
7293 		 * That way we handle nested qwriter(INNER) without dropping
7294 		 * SQ_EXCL until the outermost qwriter callback routine is
7295 		 * done.
7296 		 */
7297 		return;
7298 	}
7299 	SQ_PUTLOCKS_EXIT(sq);
7300 	sqfill_events(sq, q, mp, func);
7301 }
7302 
7303 /*
7304  * Synchronous callback support functions
7305  */
7306 
7307 /*
7308  * Allocate a callback parameter structure.
7309  * Assumes that caller initializes the flags and the id.
7310  * Acquires SQLOCK(sq) if non-NULL is returned.
7311  */
7312 callbparams_t *
7313 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7314 {
7315 	callbparams_t *cbp;
7316 	size_t size = sizeof (callbparams_t);
7317 
7318 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7319 
7320 	/*
7321 	 * Only try tryhard allocation if the caller is ready to panic.
7322 	 * Otherwise just fail.
7323 	 */
7324 	if (cbp == NULL) {
7325 		if (kmflags & KM_PANIC)
7326 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7327 			    &size, kmflags);
7328 		else
7329 			return (NULL);
7330 	}
7331 
7332 	ASSERT(size >= sizeof (callbparams_t));
7333 	cbp->cbp_size = size;
7334 	cbp->cbp_sq = sq;
7335 	cbp->cbp_func = func;
7336 	cbp->cbp_arg = arg;
7337 	mutex_enter(SQLOCK(sq));
7338 	cbp->cbp_next = sq->sq_callbpend;
7339 	sq->sq_callbpend = cbp;
7340 	return (cbp);
7341 }
7342 
7343 void
7344 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7345 {
7346 	callbparams_t **pp, *p;
7347 
7348 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7349 
7350 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7351 		if (p == cbp) {
7352 			*pp = p->cbp_next;
7353 			kmem_free(p, p->cbp_size);
7354 			return;
7355 		}
7356 	}
7357 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7358 	    "callbparams_free: not found\n"));
7359 }
7360 
7361 void
7362 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7363 {
7364 	callbparams_t **pp, *p;
7365 
7366 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7367 
7368 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7369 		if (p->cbp_id == id && p->cbp_flags == flag) {
7370 			*pp = p->cbp_next;
7371 			kmem_free(p, p->cbp_size);
7372 			return;
7373 		}
7374 	}
7375 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7376 	    "callbparams_free_id: not found\n"));
7377 }
7378 
7379 /*
7380  * Callback wrapper function used by once-only callbacks that can be
7381  * cancelled (qtimeout and qbufcall)
7382  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7383  * cancelled by the qun* functions.
7384  */
7385 void
7386 qcallbwrapper(void *arg)
7387 {
7388 	callbparams_t *cbp = arg;
7389 	syncq_t	*sq;
7390 	uint16_t count = 0;
7391 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7392 	uint16_t type;
7393 
7394 	sq = cbp->cbp_sq;
7395 	mutex_enter(SQLOCK(sq));
7396 	type = sq->sq_type;
7397 	if (!(type & SQ_CICB)) {
7398 		count = sq->sq_count;
7399 		SQ_PUTLOCKS_ENTER(sq);
7400 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7401 		SUM_SQ_PUTCOUNTS(sq, count);
7402 		sq->sq_needexcl++;
7403 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7404 		waitflags |= SQ_MESSAGES;
7405 	}
7406 	/* Can not handle exclusive entry at outer perimeter */
7407 	ASSERT(type & SQ_COCB);
7408 
7409 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7410 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7411 		    (sq->sq_cancelid == cbp->cbp_id)) {
7412 			/* timeout has been cancelled */
7413 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7414 			callbparams_free(sq, cbp);
7415 			if (!(type & SQ_CICB)) {
7416 				ASSERT(sq->sq_needexcl > 0);
7417 				sq->sq_needexcl--;
7418 				if (sq->sq_needexcl == 0) {
7419 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7420 				}
7421 				SQ_PUTLOCKS_EXIT(sq);
7422 			}
7423 			mutex_exit(SQLOCK(sq));
7424 			return;
7425 		}
7426 		sq->sq_flags |= SQ_WANTWAKEUP;
7427 		if (!(type & SQ_CICB)) {
7428 			SQ_PUTLOCKS_EXIT(sq);
7429 		}
7430 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7431 		if (!(type & SQ_CICB)) {
7432 			count = sq->sq_count;
7433 			SQ_PUTLOCKS_ENTER(sq);
7434 			SUM_SQ_PUTCOUNTS(sq, count);
7435 		}
7436 	}
7437 
7438 	sq->sq_count++;
7439 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7440 	if (!(type & SQ_CICB)) {
7441 		ASSERT(count == 0);
7442 		sq->sq_flags |= SQ_EXCL;
7443 		ASSERT(sq->sq_needexcl > 0);
7444 		sq->sq_needexcl--;
7445 		if (sq->sq_needexcl == 0) {
7446 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7447 		}
7448 		SQ_PUTLOCKS_EXIT(sq);
7449 	}
7450 
7451 	mutex_exit(SQLOCK(sq));
7452 
7453 	cbp->cbp_func(cbp->cbp_arg);
7454 
7455 	/*
7456 	 * We drop the lock only for leavesq to re-acquire it.
7457 	 * Possible optimization is inline of leavesq.
7458 	 */
7459 	mutex_enter(SQLOCK(sq));
7460 	callbparams_free(sq, cbp);
7461 	mutex_exit(SQLOCK(sq));
7462 	leavesq(sq, SQ_CALLBACK);
7463 }
7464 
7465 /*
7466  * No need to grab sq_putlocks here. See comment in strsubr.h that
7467  * explains when sq_putlocks are used.
7468  *
7469  * sq_count (or one of the sq_putcounts) has already been
7470  * decremented by the caller, and if SQ_QUEUED, we need to call
7471  * drain_syncq (the global syncq drain).
7472  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7473  * one of two states, non-CIPUT perimeter, and we need to clear
7474  * it, or we went exclusive in the put procedure.  In any case,
7475  * we want to clear the bit now, and it is probably easier to do
7476  * this at the beginning of this function (remember, we hold
7477  * the SQLOCK).  Lastly, if there are other messages queued
7478  * on the syncq (and not for our destination), enable the syncq
7479  * for background work.
7480  */
7481 
7482 /* ARGSUSED */
7483 void
7484 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7485 {
7486 	uint16_t	flags = sq->sq_flags;
7487 
7488 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7489 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7490 
7491 	/* Clear SQ_EXCL if set in passflags */
7492 	if (passflags & SQ_EXCL) {
7493 		flags &= ~SQ_EXCL;
7494 	}
7495 	if (flags & SQ_WANTWAKEUP) {
7496 		flags &= ~SQ_WANTWAKEUP;
7497 		cv_broadcast(&sq->sq_wait);
7498 	}
7499 	if (flags & SQ_WANTEXWAKEUP) {
7500 		flags &= ~SQ_WANTEXWAKEUP;
7501 		cv_broadcast(&sq->sq_exitwait);
7502 	}
7503 	sq->sq_flags = flags;
7504 
7505 	/*
7506 	 * We have cleared SQ_EXCL if we were asked to, and started
7507 	 * the wakeup process for waiters.  If there are no writers
7508 	 * then we need to drain the syncq if we were told to, or
7509 	 * enable the background thread to do it.
7510 	 */
7511 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7512 		if ((passflags & SQ_QUEUED) ||
7513 		    (sq->sq_svcflags & SQ_DISABLED)) {
7514 			/* drain_syncq will take care of events in the list */
7515 			drain_syncq(sq);
7516 			return;
7517 		} else if (flags & SQ_QUEUED) {
7518 			sqenable(sq);
7519 		}
7520 	}
7521 	/* Drop the SQLOCK on exit */
7522 	mutex_exit(SQLOCK(sq));
7523 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7524 	    "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7525 }
7526 
7527 void
7528 set_qend(queue_t *q)
7529 {
7530 	mutex_enter(QLOCK(q));
7531 	if (!O_SAMESTR(q))
7532 		q->q_flag |= QEND;
7533 	else
7534 		q->q_flag &= ~QEND;
7535 	mutex_exit(QLOCK(q));
7536 	q = _OTHERQ(q);
7537 	mutex_enter(QLOCK(q));
7538 	if (!O_SAMESTR(q))
7539 		q->q_flag |= QEND;
7540 	else
7541 		q->q_flag &= ~QEND;
7542 	mutex_exit(QLOCK(q));
7543 }
7544 
7545 /*
7546  * Set QFULL in next service procedure queue (that cares) if not already
7547  * set and if there are already more messages on the syncq than
7548  * sq_max_size.  If sq_max_size is 0, no flow control will be asserted on
7549  * any syncq.
7550  *
7551  * The fq here is the next queue with a service procedure.  This is where
7552  * we would fail canputnext, so this is where we need to set QFULL.
7553  * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7554  *
7555  * We already have QLOCK at this point. To avoid cross-locks with
7556  * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7557  * SQLOCK and sd_reflock, we need to drop respective locks first.
7558  */
7559 void
7560 set_qfull(queue_t *q)
7561 {
7562 	queue_t		*fq = NULL;
7563 
7564 	ASSERT(MUTEX_HELD(QLOCK(q)));
7565 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7566 	    (q->q_syncqmsgs > sq_max_size)) {
7567 		if ((fq = q->q_nfsrv) == q) {
7568 			fq->q_flag |= QFULL;
7569 		} else {
7570 			mutex_exit(QLOCK(q));
7571 			mutex_enter(QLOCK(fq));
7572 			fq->q_flag |= QFULL;
7573 			mutex_exit(QLOCK(fq));
7574 			mutex_enter(QLOCK(q));
7575 		}
7576 	}
7577 }
7578 
7579 void
7580 clr_qfull(queue_t *q)
7581 {
7582 	queue_t	*oq = q;
7583 
7584 	q = q->q_nfsrv;
7585 	/* Fast check if there is any work to do before getting the lock. */
7586 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7587 		return;
7588 	}
7589 
7590 	/*
7591 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7592 	 * for QFULL being set.
7593 	 */
7594 	mutex_enter(QLOCK(q));
7595 	/*
7596 	 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7597 	 * Hence clear the QFULL.
7598 	 * If both q_count and q_mblkcnt are less than the hiwat mark,
7599 	 * clear the QFULL.
7600 	 */
7601 	if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7602 	    (q->q_mblkcnt < q->q_hiwat))) {
7603 		q->q_flag &= ~QFULL;
7604 		/*
7605 		 * A little more confusing, how about this way:
7606 		 * if someone wants to write,
7607 		 * AND
7608 		 *    both counts are less than the lowat mark
7609 		 *    OR
7610 		 *    the lowat mark is zero
7611 		 * THEN
7612 		 * backenable
7613 		 */
7614 		if ((q->q_flag & QWANTW) &&
7615 		    (((q->q_count < q->q_lowat) &&
7616 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7617 			q->q_flag &= ~QWANTW;
7618 			mutex_exit(QLOCK(q));
7619 			backenable(oq, 0);
7620 		} else
7621 			mutex_exit(QLOCK(q));
7622 	} else
7623 		mutex_exit(QLOCK(q));
7624 }
7625 
7626 /*
7627  * Set the forward service procedure pointer.
7628  *
7629  * Called at insert-time to cache a queue's next forward service procedure in
7630  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7631  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7632  * inserted does not have a service procedure, then q_nfsrv points to the next
7633  * queue forward that has a service procedure.  If the queue is at the logical
7634  * end of the stream (driver for write side, stream head for the read side)
7635  * and does not have a service procedure, then q_nfsrv also points to itself.
7636  */
7637 void
7638 set_nfsrv_ptr(
7639 	queue_t  *rnew,		/* read queue pointer to new module */
7640 	queue_t  *wnew,		/* write queue pointer to new module */
7641 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7642 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7643 {
7644 	queue_t *qp;
7645 
7646 	if (prev_wq->q_next == NULL) {
7647 		/*
7648 		 * Insert the driver, initialize the driver and stream head.
7649 		 * In this case, prev_rq/prev_wq should be the stream head.
7650 		 * _I_INSERT does not allow inserting a driver.  Make sure
7651 		 * that it is not an insertion.
7652 		 */
7653 		ASSERT(!(rnew->q_flag & _QINSERTING));
7654 		wnew->q_nfsrv = wnew;
7655 		if (rnew->q_qinfo->qi_srvp)
7656 			rnew->q_nfsrv = rnew;
7657 		else
7658 			rnew->q_nfsrv = prev_rq;
7659 		prev_rq->q_nfsrv = prev_rq;
7660 		prev_wq->q_nfsrv = prev_wq;
7661 	} else {
7662 		/*
7663 		 * set up read side q_nfsrv pointer.  This MUST be done
7664 		 * before setting the write side, because the setting of
7665 		 * the write side for a fifo may depend on it.
7666 		 *
7667 		 * Suppose we have a fifo that only has pipemod pushed.
7668 		 * pipemod has no read or write service procedures, so
7669 		 * nfsrv for both pipemod queues points to prev_rq (the
7670 		 * stream read head).  Now push bufmod (which has only a
7671 		 * read service procedure).  Doing the write side first,
7672 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7673 		 * is WRONG; the next queue forward from wnew with a
7674 		 * service procedure will be rnew, not the stream read head.
7675 		 * Since the downstream queue (which in the case of a fifo
7676 		 * is the read queue rnew) can affect upstream queues, it
7677 		 * needs to be done first.  Setting up the read side first
7678 		 * sets nfsrv for both pipemod queues to rnew and then
7679 		 * when the write side is set up, wnew-q_nfsrv will also
7680 		 * point to rnew.
7681 		 */
7682 		if (rnew->q_qinfo->qi_srvp) {
7683 			/*
7684 			 * use _OTHERQ() because, if this is a pipe, next
7685 			 * module may have been pushed from other end and
7686 			 * q_next could be a read queue.
7687 			 */
7688 			qp = _OTHERQ(prev_wq->q_next);
7689 			while (qp && qp->q_nfsrv != qp) {
7690 				qp->q_nfsrv = rnew;
7691 				qp = backq(qp);
7692 			}
7693 			rnew->q_nfsrv = rnew;
7694 		} else
7695 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7696 
7697 		/* set up write side q_nfsrv pointer */
7698 		if (wnew->q_qinfo->qi_srvp) {
7699 			wnew->q_nfsrv = wnew;
7700 
7701 			/*
7702 			 * For insertion, need to update nfsrv of the modules
7703 			 * above which do not have a service routine.
7704 			 */
7705 			if (rnew->q_flag & _QINSERTING) {
7706 				for (qp = prev_wq;
7707 				    qp != NULL && qp->q_nfsrv != qp;
7708 				    qp = backq(qp)) {
7709 					qp->q_nfsrv = wnew->q_nfsrv;
7710 				}
7711 			}
7712 		} else {
7713 			if (prev_wq->q_next == prev_rq)
7714 				/*
7715 				 * Since prev_wq/prev_rq are the middle of a
7716 				 * fifo, wnew/rnew will also be the middle of
7717 				 * a fifo and wnew's nfsrv is same as rnew's.
7718 				 */
7719 				wnew->q_nfsrv = rnew->q_nfsrv;
7720 			else
7721 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7722 		}
7723 	}
7724 }
7725 
7726 /*
7727  * Reset the forward service procedure pointer; called at remove-time.
7728  */
7729 void
7730 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7731 {
7732 	queue_t *tmp_qp;
7733 
7734 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7735 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7736 		for (tmp_qp = backq(wqp);
7737 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7738 		    tmp_qp = backq(tmp_qp)) {
7739 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7740 		}
7741 	}
7742 
7743 	/* reset the read side q_nfsrv pointer */
7744 	if (rqp->q_qinfo->qi_srvp) {
7745 		if (wqp->q_next) {	/* non-driver case */
7746 			tmp_qp = _OTHERQ(wqp->q_next);
7747 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7748 				/* Note that rqp->q_next cannot be NULL */
7749 				ASSERT(rqp->q_next != NULL);
7750 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7751 				tmp_qp = backq(tmp_qp);
7752 			}
7753 		}
7754 	}
7755 }
7756 
7757 /*
7758  * This routine should be called after all stream geometry changes to update
7759  * the stream head cached struio() rd/wr queue pointers. Note must be called
7760  * with the streamlock()ed.
7761  *
7762  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7763  *	 an explicit synchronous barrier module queue. That is, a queue that
7764  *	 has specified a struio() type.
7765  */
7766 static void
7767 strsetuio(stdata_t *stp)
7768 {
7769 	queue_t *wrq;
7770 
7771 	if (stp->sd_flag & STPLEX) {
7772 		/*
7773 		 * Not streamhead, but a mux, so no Synchronous STREAMS.
7774 		 */
7775 		stp->sd_struiowrq = NULL;
7776 		stp->sd_struiordq = NULL;
7777 		return;
7778 	}
7779 	/*
7780 	 * Scan the write queue(s) while synchronous
7781 	 * until we find a qinfo uio type specified.
7782 	 */
7783 	wrq = stp->sd_wrq->q_next;
7784 	while (wrq) {
7785 		if (wrq->q_struiot == STRUIOT_NONE) {
7786 			wrq = 0;
7787 			break;
7788 		}
7789 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7790 			break;
7791 		if (! _SAMESTR(wrq)) {
7792 			wrq = 0;
7793 			break;
7794 		}
7795 		wrq = wrq->q_next;
7796 	}
7797 	stp->sd_struiowrq = wrq;
7798 	/*
7799 	 * Scan the read queue(s) while synchronous
7800 	 * until we find a qinfo uio type specified.
7801 	 */
7802 	wrq = stp->sd_wrq->q_next;
7803 	while (wrq) {
7804 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7805 			wrq = 0;
7806 			break;
7807 		}
7808 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7809 			break;
7810 		if (! _SAMESTR(wrq)) {
7811 			wrq = 0;
7812 			break;
7813 		}
7814 		wrq = wrq->q_next;
7815 	}
7816 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7817 }
7818 
7819 static int
7820 pass_rput(queue_t *q, mblk_t *mp)
7821 {
7822 	putnext(q, mp);
7823 	return (0);
7824 }
7825 
7826 /*
7827  * pass_wput, unblocks the passthru queues, so that
7828  * messages can arrive at muxs lower read queue, before
7829  * I_LINK/I_UNLINK is acked/nacked.
7830  */
7831 static int
7832 pass_wput(queue_t *q, mblk_t *mp)
7833 {
7834 	syncq_t *sq;
7835 
7836 	sq = _RD(q)->q_syncq;
7837 	if (sq->sq_flags & SQ_BLOCKED)
7838 		unblocksq(sq, SQ_BLOCKED, 0);
7839 	putnext(q, mp);
7840 	return (0);
7841 }
7842 
7843 /*
7844  * Set up queues for the link/unlink.
7845  * Create a new queue and block it and then insert it
7846  * below the stream head on the lower stream.
7847  * This prevents any messages from arriving during the setq
7848  * as well as while the mux is processing the LINK/I_UNLINK.
7849  * The blocked passq is unblocked once the LINK/I_UNLINK has
7850  * been acked or nacked or if a message is generated and sent
7851  * down muxs write put procedure.
7852  * See pass_wput().
7853  *
7854  * After the new queue is inserted, all messages coming from below are
7855  * blocked. The call to strlock will ensure that all activity in the stream head
7856  * read queue syncq is stopped (sq_count drops to zero).
7857  */
7858 static queue_t *
7859 link_addpassthru(stdata_t *stpdown)
7860 {
7861 	queue_t *passq;
7862 	sqlist_t sqlist;
7863 
7864 	passq = allocq();
7865 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7866 	/* setq might sleep in allocator - avoid holding locks. */
7867 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7868 	    SQ_CI|SQ_CO, B_FALSE);
7869 	claimq(passq);
7870 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7871 	insertq(STREAM(passq), passq);
7872 
7873 	/*
7874 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7875 	 * since we are going to change q_ptr in the stream head.  Note that
7876 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7877 	 */
7878 	sqlist.sqlist_head = NULL;
7879 	sqlist.sqlist_index = 0;
7880 	sqlist.sqlist_size = sizeof (sqlist_t);
7881 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7882 	strlock(stpdown, &sqlist);
7883 	strunlock(stpdown, &sqlist);
7884 
7885 	releaseq(passq);
7886 	return (passq);
7887 }
7888 
7889 /*
7890  * Let messages flow up into the mux by removing
7891  * the passq.
7892  */
7893 static void
7894 link_rempassthru(queue_t *passq)
7895 {
7896 	claimq(passq);
7897 	removeq(passq);
7898 	releaseq(passq);
7899 	freeq(passq);
7900 }
7901 
7902 /*
7903  * Wait for the condition variable pointed to by `cvp' to be signaled,
7904  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7905  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7906  * then the wait will be non-interruptible.
7907  *
7908  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7909  */
7910 clock_t
7911 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7912 {
7913 	clock_t ret;
7914 
7915 	if (tim < 0) {
7916 		if (nosigs) {
7917 			cv_wait(cvp, mp);
7918 			ret = 1;
7919 		} else {
7920 			ret = cv_wait_sig(cvp, mp);
7921 		}
7922 	} else if (tim > 0) {
7923 		/*
7924 		 * convert milliseconds to clock ticks
7925 		 */
7926 		if (nosigs) {
7927 			ret = cv_reltimedwait(cvp, mp,
7928 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7929 		} else {
7930 			ret = cv_reltimedwait_sig(cvp, mp,
7931 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7932 		}
7933 	} else {
7934 		ret = -1;
7935 	}
7936 	return (ret);
7937 }
7938 
7939 /*
7940  * Wait until the stream head can determine if it is at the mark but
7941  * don't wait forever to prevent a race condition between the "mark" state
7942  * in the stream head and any mark state in the caller/user of this routine.
7943  *
7944  * This is used by sockets and for a socket it would be incorrect
7945  * to return a failure for SIOCATMARK when there is no data in the receive
7946  * queue and the marked urgent data is traveling up the stream.
7947  *
7948  * This routine waits until the mark is known by waiting for one of these
7949  * three events:
7950  *	The stream head read queue becoming non-empty (including an EOF).
7951  *	The STRATMARK flag being set (due to a MSGMARKNEXT message).
7952  *	The STRNOTATMARK flag being set (which indicates that the transport
7953  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7954  *	the mark).
7955  *
7956  * The routine returns 1 if the stream is at the mark; 0 if it can
7957  * be determined that the stream is not at the mark.
7958  * If the wait times out and it can't determine
7959  * whether or not the stream might be at the mark the routine will return -1.
7960  *
7961  * Note: This routine should only be used when a mark is pending i.e.,
7962  * in the socket case the SIGURG has been posted.
7963  * Note2: This can not wakeup just because synchronous streams indicate
7964  * that data is available since it is not possible to use the synchronous
7965  * streams interfaces to determine the b_flag value for the data queued below
7966  * the stream head.
7967  */
7968 int
7969 strwaitmark(vnode_t *vp)
7970 {
7971 	struct stdata *stp = vp->v_stream;
7972 	queue_t *rq = _RD(stp->sd_wrq);
7973 	int mark;
7974 
7975 	mutex_enter(&stp->sd_lock);
7976 	while (rq->q_first == NULL &&
7977 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7978 		stp->sd_flag |= RSLEEP;
7979 
7980 		/* Wait for 100 milliseconds for any state change. */
7981 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7982 			mutex_exit(&stp->sd_lock);
7983 			return (-1);
7984 		}
7985 	}
7986 	if (stp->sd_flag & STRATMARK)
7987 		mark = 1;
7988 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7989 		mark = 1;
7990 	else
7991 		mark = 0;
7992 
7993 	mutex_exit(&stp->sd_lock);
7994 	return (mark);
7995 }
7996 
7997 /*
7998  * Set a read side error. If persist is set change the socket error
7999  * to persistent. If errfunc is set install the function as the exported
8000  * error handler.
8001  */
8002 void
8003 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8004 {
8005 	struct stdata *stp = vp->v_stream;
8006 
8007 	mutex_enter(&stp->sd_lock);
8008 	stp->sd_rerror = error;
8009 	if (error == 0 && errfunc == NULL)
8010 		stp->sd_flag &= ~STRDERR;
8011 	else
8012 		stp->sd_flag |= STRDERR;
8013 	if (persist) {
8014 		stp->sd_flag &= ~STRDERRNONPERSIST;
8015 	} else {
8016 		stp->sd_flag |= STRDERRNONPERSIST;
8017 	}
8018 	stp->sd_rderrfunc = errfunc;
8019 	if (error != 0 || errfunc != NULL) {
8020 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8021 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8022 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8023 
8024 		mutex_exit(&stp->sd_lock);
8025 		pollwakeup(&stp->sd_pollist, POLLERR);
8026 		mutex_enter(&stp->sd_lock);
8027 
8028 		if (stp->sd_sigflags & S_ERROR)
8029 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8030 	}
8031 	mutex_exit(&stp->sd_lock);
8032 }
8033 
8034 /*
8035  * Set a write side error. If persist is set change the socket error
8036  * to persistent.
8037  */
8038 void
8039 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8040 {
8041 	struct stdata *stp = vp->v_stream;
8042 
8043 	mutex_enter(&stp->sd_lock);
8044 	stp->sd_werror = error;
8045 	if (error == 0 && errfunc == NULL)
8046 		stp->sd_flag &= ~STWRERR;
8047 	else
8048 		stp->sd_flag |= STWRERR;
8049 	if (persist) {
8050 		stp->sd_flag &= ~STWRERRNONPERSIST;
8051 	} else {
8052 		stp->sd_flag |= STWRERRNONPERSIST;
8053 	}
8054 	stp->sd_wrerrfunc = errfunc;
8055 	if (error != 0 || errfunc != NULL) {
8056 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8057 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8058 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8059 
8060 		mutex_exit(&stp->sd_lock);
8061 		pollwakeup(&stp->sd_pollist, POLLERR);
8062 		mutex_enter(&stp->sd_lock);
8063 
8064 		if (stp->sd_sigflags & S_ERROR)
8065 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8066 	}
8067 	mutex_exit(&stp->sd_lock);
8068 }
8069 
8070 /*
8071  * Make the stream return 0 (EOF) when all data has been read.
8072  * No effect on write side.
8073  */
8074 void
8075 strseteof(vnode_t *vp, int eof)
8076 {
8077 	struct stdata *stp = vp->v_stream;
8078 
8079 	mutex_enter(&stp->sd_lock);
8080 	if (!eof) {
8081 		stp->sd_flag &= ~STREOF;
8082 		mutex_exit(&stp->sd_lock);
8083 		return;
8084 	}
8085 	stp->sd_flag |= STREOF;
8086 	if (stp->sd_flag & RSLEEP) {
8087 		stp->sd_flag &= ~RSLEEP;
8088 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8089 	}
8090 
8091 	mutex_exit(&stp->sd_lock);
8092 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8093 	mutex_enter(&stp->sd_lock);
8094 
8095 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8096 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8097 	mutex_exit(&stp->sd_lock);
8098 }
8099 
8100 void
8101 strflushrq(vnode_t *vp, int flag)
8102 {
8103 	struct stdata *stp = vp->v_stream;
8104 
8105 	mutex_enter(&stp->sd_lock);
8106 	flushq(_RD(stp->sd_wrq), flag);
8107 	mutex_exit(&stp->sd_lock);
8108 }
8109 
8110 void
8111 strsetrputhooks(vnode_t *vp, uint_t flags,
8112     msgfunc_t protofunc, msgfunc_t miscfunc)
8113 {
8114 	struct stdata *stp = vp->v_stream;
8115 
8116 	mutex_enter(&stp->sd_lock);
8117 
8118 	if (protofunc == NULL)
8119 		stp->sd_rprotofunc = strrput_proto;
8120 	else
8121 		stp->sd_rprotofunc = protofunc;
8122 
8123 	if (miscfunc == NULL)
8124 		stp->sd_rmiscfunc = strrput_misc;
8125 	else
8126 		stp->sd_rmiscfunc = miscfunc;
8127 
8128 	if (flags & SH_CONSOL_DATA)
8129 		stp->sd_rput_opt |= SR_CONSOL_DATA;
8130 	else
8131 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8132 
8133 	if (flags & SH_SIGALLDATA)
8134 		stp->sd_rput_opt |= SR_SIGALLDATA;
8135 	else
8136 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
8137 
8138 	if (flags & SH_IGN_ZEROLEN)
8139 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8140 	else
8141 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8142 
8143 	mutex_exit(&stp->sd_lock);
8144 }
8145 
8146 void
8147 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8148 {
8149 	struct stdata *stp = vp->v_stream;
8150 
8151 	mutex_enter(&stp->sd_lock);
8152 	stp->sd_closetime = closetime;
8153 
8154 	if (flags & SH_SIGPIPE)
8155 		stp->sd_wput_opt |= SW_SIGPIPE;
8156 	else
8157 		stp->sd_wput_opt &= ~SW_SIGPIPE;
8158 	if (flags & SH_RECHECK_ERR)
8159 		stp->sd_wput_opt |= SW_RECHECK_ERR;
8160 	else
8161 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8162 
8163 	mutex_exit(&stp->sd_lock);
8164 }
8165 
8166 void
8167 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8168 {
8169 	struct stdata *stp = vp->v_stream;
8170 
8171 	mutex_enter(&stp->sd_lock);
8172 
8173 	stp->sd_rputdatafunc = rdatafunc;
8174 	stp->sd_wputdatafunc = wdatafunc;
8175 
8176 	mutex_exit(&stp->sd_lock);
8177 }
8178 
8179 /* Used within framework when the queue is already locked */
8180 void
8181 qenable_locked(queue_t *q)
8182 {
8183 	stdata_t *stp = STREAM(q);
8184 
8185 	ASSERT(MUTEX_HELD(QLOCK(q)));
8186 
8187 	if (!q->q_qinfo->qi_srvp)
8188 		return;
8189 
8190 	/*
8191 	 * Do not place on run queue if already enabled or closing.
8192 	 */
8193 	if (q->q_flag & (QWCLOSE|QENAB))
8194 		return;
8195 
8196 	/*
8197 	 * mark queue enabled and place on run list if it is not already being
8198 	 * serviced. If it is serviced, the runservice() function will detect
8199 	 * that QENAB is set and call service procedure before clearing
8200 	 * QINSERVICE flag.
8201 	 */
8202 	q->q_flag |= QENAB;
8203 	if (q->q_flag & QINSERVICE)
8204 		return;
8205 
8206 	/* Record the time of qenable */
8207 	q->q_qtstamp = ddi_get_lbolt();
8208 
8209 	/*
8210 	 * Put the queue in the stp list and schedule it for background
8211 	 * processing if it is not already scheduled or if stream head does not
8212 	 * intent to process it in the foreground later by setting
8213 	 * STRS_WILLSERVICE flag.
8214 	 */
8215 	mutex_enter(&stp->sd_qlock);
8216 	/*
8217 	 * If there are already something on the list, stp flags should show
8218 	 * intention to drain it.
8219 	 */
8220 	IMPLY(STREAM_NEEDSERVICE(stp),
8221 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8222 
8223 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8224 	stp->sd_nqueues++;
8225 
8226 	/*
8227 	 * If no one will drain this stream we are the first producer and
8228 	 * need to schedule it for background thread.
8229 	 */
8230 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8231 		/*
8232 		 * No one will service this stream later, so we have to
8233 		 * schedule it now.
8234 		 */
8235 		STRSTAT(stenables);
8236 		stp->sd_svcflags |= STRS_SCHEDULED;
8237 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8238 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8239 
8240 		if (stp->sd_servid == NULL) {
8241 			/*
8242 			 * Task queue failed so fail over to the backup
8243 			 * servicing thread.
8244 			 */
8245 			STRSTAT(taskqfails);
8246 			/*
8247 			 * It is safe to clear STRS_SCHEDULED flag because it
8248 			 * was set by this thread above.
8249 			 */
8250 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8251 
8252 			/*
8253 			 * Failover scheduling is protected by service_queue
8254 			 * lock.
8255 			 */
8256 			mutex_enter(&service_queue);
8257 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8258 			ASSERT(q->q_link == NULL);
8259 			/*
8260 			 * Append the queue to qhead/qtail list.
8261 			 */
8262 			if (qhead == NULL)
8263 				qhead = q;
8264 			else
8265 				qtail->q_link = q;
8266 			qtail = q;
8267 			/*
8268 			 * Clear stp queue list.
8269 			 */
8270 			stp->sd_qhead = stp->sd_qtail = NULL;
8271 			stp->sd_nqueues = 0;
8272 			/*
8273 			 * Wakeup background queue processing thread.
8274 			 */
8275 			cv_signal(&services_to_run);
8276 			mutex_exit(&service_queue);
8277 		}
8278 	}
8279 	mutex_exit(&stp->sd_qlock);
8280 }
8281 
8282 static void
8283 queue_service(queue_t *q)
8284 {
8285 	/*
8286 	 * The queue in the list should have
8287 	 * QENAB flag set and should not have
8288 	 * QINSERVICE flag set. QINSERVICE is
8289 	 * set when the queue is dequeued and
8290 	 * qenable_locked doesn't enqueue a
8291 	 * queue with QINSERVICE set.
8292 	 */
8293 
8294 	ASSERT(!(q->q_flag & QINSERVICE));
8295 	ASSERT((q->q_flag & QENAB));
8296 	mutex_enter(QLOCK(q));
8297 	q->q_flag &= ~QENAB;
8298 	q->q_flag |= QINSERVICE;
8299 	mutex_exit(QLOCK(q));
8300 	runservice(q);
8301 }
8302 
8303 static void
8304 syncq_service(syncq_t *sq)
8305 {
8306 	STRSTAT(syncqservice);
8307 	mutex_enter(SQLOCK(sq));
8308 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8309 	ASSERT(sq->sq_servcount != 0);
8310 	ASSERT(sq->sq_next == NULL);
8311 
8312 	/* if we came here from the background thread, clear the flag */
8313 	if (sq->sq_svcflags & SQ_BGTHREAD)
8314 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8315 
8316 	/* let drain_syncq know that it's being called in the background */
8317 	sq->sq_svcflags |= SQ_SERVICE;
8318 	drain_syncq(sq);
8319 }
8320 
8321 static void
8322 qwriter_outer_service(syncq_t *outer)
8323 {
8324 	/*
8325 	 * Note that SQ_WRITER is used on the outer perimeter
8326 	 * to signal that a qwriter(OUTER) is either investigating
8327 	 * running or that it is actually running a function.
8328 	 */
8329 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8330 
8331 	/*
8332 	 * All inner syncq are empty and have SQ_WRITER set
8333 	 * to block entering the outer perimeter.
8334 	 *
8335 	 * We do not need to explicitly call write_now since
8336 	 * outer_exit does it for us.
8337 	 */
8338 	outer_exit(outer);
8339 }
8340 
8341 static void
8342 mblk_free(mblk_t *mp)
8343 {
8344 	dblk_t *dbp = mp->b_datap;
8345 	frtn_t *frp = dbp->db_frtnp;
8346 
8347 	mp->b_next = NULL;
8348 	if (dbp->db_fthdr != NULL)
8349 		str_ftfree(dbp);
8350 
8351 	ASSERT(dbp->db_fthdr == NULL);
8352 	frp->free_func(frp->free_arg);
8353 	ASSERT(dbp->db_mblk == mp);
8354 
8355 	if (dbp->db_credp != NULL) {
8356 		crfree(dbp->db_credp);
8357 		dbp->db_credp = NULL;
8358 	}
8359 	dbp->db_cpid = -1;
8360 	dbp->db_struioflag = 0;
8361 	dbp->db_struioun.cksum.flags = 0;
8362 
8363 	kmem_cache_free(dbp->db_cache, dbp);
8364 }
8365 
8366 /*
8367  * Background processing of the stream queue list.
8368  */
8369 static void
8370 stream_service(stdata_t *stp)
8371 {
8372 	queue_t *q;
8373 
8374 	mutex_enter(&stp->sd_qlock);
8375 
8376 	STR_SERVICE(stp, q);
8377 
8378 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8379 	stp->sd_servid = NULL;
8380 	cv_signal(&stp->sd_qcv);
8381 	mutex_exit(&stp->sd_qlock);
8382 }
8383 
8384 /*
8385  * Foreground processing of the stream queue list.
8386  */
8387 void
8388 stream_runservice(stdata_t *stp)
8389 {
8390 	queue_t *q;
8391 
8392 	mutex_enter(&stp->sd_qlock);
8393 	STRSTAT(rservice);
8394 	/*
8395 	 * We are going to drain this stream queue list, so qenable_locked will
8396 	 * not schedule it until we finish.
8397 	 */
8398 	stp->sd_svcflags |= STRS_WILLSERVICE;
8399 
8400 	STR_SERVICE(stp, q);
8401 
8402 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8403 	mutex_exit(&stp->sd_qlock);
8404 	/*
8405 	 * Help backup background thread to drain the qhead/qtail list.
8406 	 */
8407 	while (qhead != NULL) {
8408 		STRSTAT(qhelps);
8409 		mutex_enter(&service_queue);
8410 		DQ(q, qhead, qtail, q_link);
8411 		mutex_exit(&service_queue);
8412 		if (q != NULL)
8413 			queue_service(q);
8414 	}
8415 }
8416 
8417 void
8418 stream_willservice(stdata_t *stp)
8419 {
8420 	mutex_enter(&stp->sd_qlock);
8421 	stp->sd_svcflags |= STRS_WILLSERVICE;
8422 	mutex_exit(&stp->sd_qlock);
8423 }
8424 
8425 /*
8426  * Replace the cred currently in the mblk with a different one.
8427  * Also update db_cpid.
8428  */
8429 void
8430 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8431 {
8432 	dblk_t *dbp = mp->b_datap;
8433 	cred_t *ocr = dbp->db_credp;
8434 
8435 	ASSERT(cr != NULL);
8436 
8437 	if (cr != ocr) {
8438 		crhold(dbp->db_credp = cr);
8439 		if (ocr != NULL)
8440 			crfree(ocr);
8441 	}
8442 	/* Don't overwrite with NOPID */
8443 	if (cpid != NOPID)
8444 		dbp->db_cpid = cpid;
8445 }
8446 
8447 /*
8448  * If the src message has a cred, then replace the cred currently in the mblk
8449  * with it.
8450  * Also update db_cpid.
8451  */
8452 void
8453 mblk_copycred(mblk_t *mp, const mblk_t *src)
8454 {
8455 	dblk_t *dbp = mp->b_datap;
8456 	cred_t *cr, *ocr;
8457 	pid_t cpid;
8458 
8459 	cr = msg_getcred(src, &cpid);
8460 	if (cr == NULL)
8461 		return;
8462 
8463 	ocr = dbp->db_credp;
8464 	if (cr != ocr) {
8465 		crhold(dbp->db_credp = cr);
8466 		if (ocr != NULL)
8467 			crfree(ocr);
8468 	}
8469 	/* Don't overwrite with NOPID */
8470 	if (cpid != NOPID)
8471 		dbp->db_cpid = cpid;
8472 }
8473 
8474 
8475 /*
8476  * Now that NIC drivers are expected to deal only with M_DATA mblks, the
8477  * hcksum_assoc and hcksum_retrieve functions are deprecated in favor of their
8478  * respective mac_hcksum_set and mac_hcksum_get counterparts.
8479  */
8480 int
8481 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8482     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8483     uint32_t flags, int km_flags)
8484 {
8485 	int rc = 0;
8486 
8487 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8488 	if (mp->b_datap->db_type == M_DATA) {
8489 		/* Associate values for M_DATA type */
8490 		DB_CKSUMSTART(mp) = (intptr_t)start;
8491 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8492 		DB_CKSUMEND(mp) = (intptr_t)end;
8493 		DB_CKSUMFLAGS(mp) = flags;
8494 		DB_CKSUM16(mp) = (uint16_t)value;
8495 
8496 	} else {
8497 		pattrinfo_t pa_info;
8498 
8499 		ASSERT(mmd != NULL);
8500 
8501 		pa_info.type = PATTR_HCKSUM;
8502 		pa_info.len = sizeof (pattr_hcksum_t);
8503 
8504 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8505 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8506 
8507 			hck->hcksum_start_offset = start;
8508 			hck->hcksum_stuff_offset = stuff;
8509 			hck->hcksum_end_offset = end;
8510 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8511 			hck->hcksum_flags = flags;
8512 		} else {
8513 			rc = -1;
8514 		}
8515 	}
8516 	return (rc);
8517 }
8518 
8519 void
8520 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8521     uint32_t *start, uint32_t *stuff, uint32_t *end,
8522     uint32_t *value, uint32_t *flags)
8523 {
8524 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8525 	if (mp->b_datap->db_type == M_DATA) {
8526 		if (flags != NULL) {
8527 			*flags = DB_CKSUMFLAGS(mp) & HCK_FLAGS;
8528 			if ((*flags & (HCK_PARTIALCKSUM |
8529 			    HCK_FULLCKSUM)) != 0) {
8530 				if (value != NULL)
8531 					*value = (uint32_t)DB_CKSUM16(mp);
8532 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
8533 					if (start != NULL)
8534 						*start =
8535 						    (uint32_t)DB_CKSUMSTART(mp);
8536 					if (stuff != NULL)
8537 						*stuff =
8538 						    (uint32_t)DB_CKSUMSTUFF(mp);
8539 					if (end != NULL)
8540 						*end =
8541 						    (uint32_t)DB_CKSUMEND(mp);
8542 				}
8543 			}
8544 		}
8545 	} else {
8546 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8547 
8548 		ASSERT(mmd != NULL);
8549 
8550 		/* get hardware checksum attribute */
8551 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8552 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8553 
8554 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8555 			if (flags != NULL)
8556 				*flags = hck->hcksum_flags;
8557 			if (start != NULL)
8558 				*start = hck->hcksum_start_offset;
8559 			if (stuff != NULL)
8560 				*stuff = hck->hcksum_stuff_offset;
8561 			if (end != NULL)
8562 				*end = hck->hcksum_end_offset;
8563 			if (value != NULL)
8564 				*value = (uint32_t)
8565 				    hck->hcksum_cksum_val.inet_cksum;
8566 		}
8567 	}
8568 }
8569 
8570 void
8571 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8572 {
8573 	ASSERT(DB_TYPE(mp) == M_DATA);
8574 	ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8575 
8576 	/* Set the flags */
8577 	DB_LSOFLAGS(mp) |= flags;
8578 	DB_LSOMSS(mp) = mss;
8579 }
8580 
8581 void
8582 lso_info_cleanup(mblk_t *mp)
8583 {
8584 	ASSERT(DB_TYPE(mp) == M_DATA);
8585 
8586 	/* Clear the flags */
8587 	DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8588 	DB_LSOMSS(mp) = 0;
8589 }
8590 
8591 /*
8592  * Checksum buffer *bp for len bytes with psum partial checksum,
8593  * or 0 if none, and return the 16 bit partial checksum.
8594  */
8595 unsigned
8596 bcksum(uchar_t *bp, int len, unsigned int psum)
8597 {
8598 	int odd = len & 1;
8599 	extern unsigned int ip_ocsum();
8600 
8601 	if (((intptr_t)bp & 1) == 0 && !odd) {
8602 		/*
8603 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8604 		 */
8605 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8606 	}
8607 	if (((intptr_t)bp & 1) != 0) {
8608 		/*
8609 		 * Bp isn't 16 bit aligned.
8610 		 */
8611 		unsigned int tsum;
8612 
8613 #ifdef _LITTLE_ENDIAN
8614 		psum += *bp;
8615 #else
8616 		psum += *bp << 8;
8617 #endif
8618 		len--;
8619 		bp++;
8620 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8621 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8622 		if (len & 1) {
8623 			bp += len - 1;
8624 #ifdef _LITTLE_ENDIAN
8625 			psum += *bp << 8;
8626 #else
8627 			psum += *bp;
8628 #endif
8629 		}
8630 	} else {
8631 		/*
8632 		 * Bp is 16 bit aligned.
8633 		 */
8634 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8635 		if (odd) {
8636 			bp += len - 1;
8637 #ifdef _LITTLE_ENDIAN
8638 			psum += *bp;
8639 #else
8640 			psum += *bp << 8;
8641 #endif
8642 		}
8643 	}
8644 	/*
8645 	 * Normalize psum to 16 bits before returning the new partial
8646 	 * checksum. The max psum value before normalization is 0x3FDFE.
8647 	 */
8648 	return ((psum >> 16) + (psum & 0xFFFF));
8649 }
8650 
8651 boolean_t
8652 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8653 {
8654 	boolean_t rc;
8655 
8656 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8657 	if (DB_TYPE(mp) == M_DATA) {
8658 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8659 	} else {
8660 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8661 
8662 		ASSERT(mmd != NULL);
8663 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8664 	}
8665 	return (rc);
8666 }
8667 
8668 void
8669 freemsgchain(mblk_t *mp)
8670 {
8671 	mblk_t	*next;
8672 
8673 	while (mp != NULL) {
8674 		next = mp->b_next;
8675 		mp->b_next = NULL;
8676 
8677 		freemsg(mp);
8678 		mp = next;
8679 	}
8680 }
8681 
8682 mblk_t *
8683 copymsgchain(mblk_t *mp)
8684 {
8685 	mblk_t	*nmp = NULL;
8686 	mblk_t	**nmpp = &nmp;
8687 
8688 	for (; mp != NULL; mp = mp->b_next) {
8689 		if ((*nmpp = copymsg(mp)) == NULL) {
8690 			freemsgchain(nmp);
8691 			return (NULL);
8692 		}
8693 
8694 		nmpp = &((*nmpp)->b_next);
8695 	}
8696 
8697 	return (nmp);
8698 }
8699 
8700 /* NOTE: Do not add code after this point. */
8701 #undef QLOCK
8702 
8703 /*
8704  * Replacement for QLOCK macro for those that can't use it.
8705  */
8706 kmutex_t *
8707 QLOCK(queue_t *q)
8708 {
8709 	return (&(q)->q_lock);
8710 }
8711 
8712 /*
8713  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8714  */
8715 #undef runqueues
8716 void
8717 runqueues(void)
8718 {
8719 }
8720 
8721 #undef queuerun
8722 void
8723 queuerun(void)
8724 {
8725 }
8726 
8727 /*
8728  * Initialize the STR stack instance, which tracks autopush and persistent
8729  * links.
8730  */
8731 /* ARGSUSED */
8732 static void *
8733 str_stack_init(netstackid_t stackid, netstack_t *ns)
8734 {
8735 	str_stack_t	*ss;
8736 	int i;
8737 
8738 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8739 	ss->ss_netstack = ns;
8740 
8741 	/*
8742 	 * set up autopush
8743 	 */
8744 	sad_initspace(ss);
8745 
8746 	/*
8747 	 * set up mux_node structures.
8748 	 */
8749 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8750 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8751 	    ss->ss_devcnt), KM_SLEEP);
8752 	for (i = 0; i < ss->ss_devcnt; i++)
8753 		ss->ss_mux_nodes[i].mn_imaj = i;
8754 	return (ss);
8755 }
8756 
8757 /*
8758  * Note: run at zone shutdown and not destroy so that the PLINKs are
8759  * gone by the time other cleanup happens from the destroy callbacks.
8760  */
8761 static void
8762 str_stack_shutdown(netstackid_t stackid, void *arg)
8763 {
8764 	str_stack_t *ss = (str_stack_t *)arg;
8765 	int i;
8766 	cred_t *cr;
8767 
8768 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8769 	ASSERT(cr != NULL);
8770 
8771 	/* Undo all the I_PLINKs for this zone */
8772 	for (i = 0; i < ss->ss_devcnt; i++) {
8773 		struct mux_edge		*ep;
8774 		ldi_handle_t		lh;
8775 		ldi_ident_t		li;
8776 		int			ret;
8777 		int			rval;
8778 		dev_t			rdev;
8779 
8780 		ep = ss->ss_mux_nodes[i].mn_outp;
8781 		if (ep == NULL)
8782 			continue;
8783 		ret = ldi_ident_from_major((major_t)i, &li);
8784 		if (ret != 0) {
8785 			continue;
8786 		}
8787 		rdev = ep->me_dev;
8788 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8789 		    cr, &lh, li);
8790 		if (ret != 0) {
8791 			ldi_ident_release(li);
8792 			continue;
8793 		}
8794 
8795 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8796 		    cr, &rval);
8797 		if (ret) {
8798 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8799 			ldi_ident_release(li);
8800 			continue;
8801 		}
8802 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8803 
8804 		/* Close layered handles */
8805 		ldi_ident_release(li);
8806 	}
8807 	crfree(cr);
8808 
8809 	sad_freespace(ss);
8810 
8811 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8812 	ss->ss_mux_nodes = NULL;
8813 }
8814 
8815 /*
8816  * Free the structure; str_stack_shutdown did the other cleanup work.
8817  */
8818 /* ARGSUSED */
8819 static void
8820 str_stack_fini(netstackid_t stackid, void *arg)
8821 {
8822 	str_stack_t	*ss = (str_stack_t *)arg;
8823 
8824 	kmem_free(ss, sizeof (*ss));
8825 }
8826