xref: /titanic_41/usr/src/uts/common/os/strsubr.c (revision d73ae94e59c019f5cc3221ee0a0012d02091b40e)
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 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 #include <sys/types.h>
33 #include <sys/sysmacros.h>
34 #include <sys/param.h>
35 #include <sys/errno.h>
36 #include <sys/signal.h>
37 #include <sys/proc.h>
38 #include <sys/conf.h>
39 #include <sys/cred.h>
40 #include <sys/user.h>
41 #include <sys/vnode.h>
42 #include <sys/file.h>
43 #include <sys/session.h>
44 #include <sys/stream.h>
45 #include <sys/strsubr.h>
46 #include <sys/stropts.h>
47 #include <sys/poll.h>
48 #include <sys/systm.h>
49 #include <sys/cpuvar.h>
50 #include <sys/uio.h>
51 #include <sys/cmn_err.h>
52 #include <sys/priocntl.h>
53 #include <sys/procset.h>
54 #include <sys/vmem.h>
55 #include <sys/bitmap.h>
56 #include <sys/kmem.h>
57 #include <sys/siginfo.h>
58 #include <sys/vtrace.h>
59 #include <sys/callb.h>
60 #include <sys/debug.h>
61 #include <sys/modctl.h>
62 #include <sys/vmsystm.h>
63 #include <vm/page.h>
64 #include <sys/atomic.h>
65 #include <sys/suntpi.h>
66 #include <sys/strlog.h>
67 #include <sys/promif.h>
68 #include <sys/project.h>
69 #include <sys/vm.h>
70 #include <sys/taskq.h>
71 #include <sys/sunddi.h>
72 #include <sys/sunldi_impl.h>
73 #include <sys/strsun.h>
74 #include <sys/isa_defs.h>
75 #include <sys/multidata.h>
76 #include <sys/pattr.h>
77 #include <sys/strft.h>
78 #include <sys/fs/snode.h>
79 #include <sys/zone.h>
80 
81 #define	O_SAMESTR(q)	(((q)->q_next) && \
82 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
83 
84 /*
85  * WARNING:
86  * The variables and routines in this file are private, belonging
87  * to the STREAMS subsystem. These should not be used by modules
88  * or drivers. Compatibility will not be guaranteed.
89  */
90 
91 /*
92  * Id value used to distinguish between different multiplexor links.
93  */
94 static int32_t lnk_id = 0;
95 
96 #define	STREAMS_LOPRI MINCLSYSPRI
97 static pri_t streams_lopri = STREAMS_LOPRI;
98 
99 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
100 typedef struct str_stat {
101 	kstat_named_t	sqenables;
102 	kstat_named_t	stenables;
103 	kstat_named_t	syncqservice;
104 	kstat_named_t	freebs;
105 	kstat_named_t	qwr_outer;
106 	kstat_named_t	rservice;
107 	kstat_named_t	strwaits;
108 	kstat_named_t	taskqfails;
109 	kstat_named_t	bufcalls;
110 	kstat_named_t	qhelps;
111 	kstat_named_t	qremoved;
112 	kstat_named_t	sqremoved;
113 	kstat_named_t	bcwaits;
114 	kstat_named_t	sqtoomany;
115 } str_stat_t;
116 
117 static str_stat_t str_statistics = {
118 	{ "sqenables",		KSTAT_DATA_UINT64 },
119 	{ "stenables",		KSTAT_DATA_UINT64 },
120 	{ "syncqservice",	KSTAT_DATA_UINT64 },
121 	{ "freebs",		KSTAT_DATA_UINT64 },
122 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
123 	{ "rservice",		KSTAT_DATA_UINT64 },
124 	{ "strwaits",		KSTAT_DATA_UINT64 },
125 	{ "taskqfails",		KSTAT_DATA_UINT64 },
126 	{ "bufcalls",		KSTAT_DATA_UINT64 },
127 	{ "qhelps",		KSTAT_DATA_UINT64 },
128 	{ "qremoved",		KSTAT_DATA_UINT64 },
129 	{ "sqremoved",		KSTAT_DATA_UINT64 },
130 	{ "bcwaits",		KSTAT_DATA_UINT64 },
131 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
132 };
133 
134 static kstat_t *str_kstat;
135 
136 /*
137  * qrunflag was used previously to control background scheduling of queues. It
138  * is not used anymore, but kept here in case some module still wants to access
139  * it via qready() and setqsched macros.
140  */
141 char qrunflag;			/*  Unused */
142 
143 /*
144  * Most of the streams scheduling is done via task queues. Task queues may fail
145  * for non-sleep dispatches, so there are two backup threads servicing failed
146  * requests for queues and syncqs. Both of these threads also service failed
147  * dispatches freebs requests. Queues are put in the list specified by `qhead'
148  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
149  * requests are put into `freebs_list' which has no tail pointer. All three
150  * lists are protected by a single `service_queue' lock and use
151  * `services_to_run' condition variable for signaling background threads. Use of
152  * a single lock should not be a problem because it is only used under heavy
153  * loads when task queues start to fail and at that time it may be a good idea
154  * to throttle scheduling requests.
155  *
156  * NOTE: queues and syncqs should be scheduled by two separate threads because
157  * queue servicing may be blocked waiting for a syncq which may be also
158  * scheduled for background execution. This may create a deadlock when only one
159  * thread is used for both.
160  */
161 
162 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
163 
164 static kmutex_t service_queue;		/* protects all of servicing vars */
165 static kcondvar_t services_to_run;	/* wake up background service thread */
166 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
167 
168 /*
169  * List of queues scheduled for background processing dueue to lack of resources
170  * in the task queues. Protected by service_queue lock;
171  */
172 static struct queue *qhead;
173 static struct queue *qtail;
174 
175 /*
176  * Same list for syncqs
177  */
178 static syncq_t *sqhead;
179 static syncq_t *sqtail;
180 
181 static mblk_t *freebs_list;	/* list of buffers to free */
182 
183 /*
184  * Backup threads for servicing queues and syncqs
185  */
186 kthread_t *streams_qbkgrnd_thread;
187 kthread_t *streams_sqbkgrnd_thread;
188 
189 /*
190  * Bufcalls related variables.
191  */
192 struct bclist	strbcalls;	/* list of waiting bufcalls */
193 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
194 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
195 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
196 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
197 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
198 
199 kmutex_t	strresources;	/* protects global resources */
200 kmutex_t	muxifier;	/* single-threads multiplexor creation */
201 
202 extern void	time_to_wait(clock_t *, clock_t);
203 
204 /*
205  * run_queues is no longer used, but is kept in case some 3-d party
206  * module/driver decides to use it.
207  */
208 int run_queues = 0;
209 
210 /*
211  * sq_max_size is the depth of the syncq (in number of messages) before
212  * qfill_syncq() starts QFULL'ing destination queues. As its primary
213  * consumer - IP is no longer D_MTPERMOD, but there may be other
214  * modules/drivers depend on this syncq flow control, we prefer to
215  * choose a large number as the default value. For potential
216  * performance gain, this value is tunable in /etc/system.
217  */
218 int sq_max_size = 10000;
219 
220 /*
221  * the number of ciputctrl structures per syncq and stream we create when
222  * needed.
223  */
224 int n_ciputctrl;
225 int max_n_ciputctrl = 16;
226 /*
227  * if n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
228  */
229 int min_n_ciputctrl = 2;
230 
231 static struct mux_node *mux_nodes;	/* mux info for cycle checking */
232 
233 /*
234  * Per-driver/module syncqs
235  * ========================
236  *
237  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
238  * perdm structures, new entries being added (and new syncqs allocated) when
239  * setq() encounters a module/driver with a streamtab that it hasn't seen
240  * before.
241  * The reason for this mechanism is that some modules and drivers share a
242  * common streamtab and it is necessary for those modules and drivers to also
243  * share a common PERMOD syncq.
244  *
245  * perdm_list --> dm_str == streamtab_1
246  *                dm_sq == syncq_1
247  *                dm_ref
248  *                dm_next --> dm_str == streamtab_2
249  *                            dm_sq == syncq_2
250  *                            dm_ref
251  *                            dm_next --> ... NULL
252  *
253  * The dm_ref field is incremented for each new driver/module that takes
254  * a reference to the perdm structure and hence shares the syncq.
255  * References are held in the fmodsw_impl_t structure for each STREAMS module
256  * or the dev_impl array (indexed by device major number) for each driver.
257  *
258  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
259  *		     ^                 ^ ^               ^
260  *                   |  ______________/  |               |
261  *                   | /                 |               |
262  * dev_impl:     ...|x|y|...          module A	      module B
263  *
264  * When a module/driver is unloaded the reference count is decremented and,
265  * when it falls to zero, the perdm structure is removed from the list and
266  * the syncq is freed (see rele_dm()).
267  */
268 perdm_t *perdm_list = NULL;
269 static krwlock_t perdm_rwlock;
270 cdevsw_impl_t *devimpl;
271 
272 extern struct qinit strdata;
273 extern struct qinit stwdata;
274 
275 static void runservice(queue_t *);
276 static void streams_bufcall_service(void);
277 static void streams_qbkgrnd_service(void);
278 static void streams_sqbkgrnd_service(void);
279 static syncq_t *new_syncq(void);
280 static void free_syncq(syncq_t *);
281 static void outer_insert(syncq_t *, syncq_t *);
282 static void outer_remove(syncq_t *, syncq_t *);
283 static void write_now(syncq_t *);
284 static void clr_qfull(queue_t *);
285 static void enable_svc(queue_t *);
286 static void runbufcalls(void);
287 static void sqenable(syncq_t *);
288 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
289 static void wait_q_syncq(queue_t *);
290 static void backenable_insertedq(queue_t *);
291 
292 static void queue_service(queue_t *);
293 static void stream_service(stdata_t *);
294 static void syncq_service(syncq_t *);
295 static void qwriter_outer_service(syncq_t *);
296 static void mblk_free(mblk_t *);
297 #ifdef DEBUG
298 static int qprocsareon(queue_t *);
299 #endif
300 
301 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
302 static void reset_nfsrv_ptr(queue_t *, queue_t *);
303 
304 static void sq_run_events(syncq_t *);
305 static int propagate_syncq(queue_t *);
306 
307 static void	blocksq(syncq_t *, ushort_t, int);
308 static void	unblocksq(syncq_t *, ushort_t, int);
309 static int	dropsq(syncq_t *, uint16_t);
310 static void	emptysq(syncq_t *);
311 static sqlist_t *sqlist_alloc(struct stdata *, int);
312 static void	sqlist_free(sqlist_t *);
313 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
314 static void	sqlist_insert(sqlist_t *, syncq_t *);
315 static void	sqlist_insertall(sqlist_t *, queue_t *);
316 
317 static void	strsetuio(stdata_t *);
318 
319 struct kmem_cache *stream_head_cache;
320 struct kmem_cache *queue_cache;
321 struct kmem_cache *syncq_cache;
322 struct kmem_cache *qband_cache;
323 struct kmem_cache *linkinfo_cache;
324 struct kmem_cache *ciputctrl_cache = NULL;
325 
326 static linkinfo_t *linkinfo_list;
327 
328 /*
329  *  Qinit structure and Module_info structures
330  *	for passthru read and write queues
331  */
332 
333 static void pass_wput(queue_t *, mblk_t *);
334 static queue_t *link_addpassthru(stdata_t *);
335 static void link_rempassthru(queue_t *);
336 
337 struct  module_info passthru_info = {
338 	0,
339 	"passthru",
340 	0,
341 	INFPSZ,
342 	STRHIGH,
343 	STRLOW
344 };
345 
346 struct  qinit passthru_rinit = {
347 	(int (*)())putnext,
348 	NULL,
349 	NULL,
350 	NULL,
351 	NULL,
352 	&passthru_info,
353 	NULL
354 };
355 
356 struct  qinit passthru_winit = {
357 	(int (*)()) pass_wput,
358 	NULL,
359 	NULL,
360 	NULL,
361 	NULL,
362 	&passthru_info,
363 	NULL
364 };
365 
366 /*
367  * Special form of assertion: verify that X implies Y i.e. when X is true Y
368  * should also be true.
369  */
370 #define	IMPLY(X, Y)	ASSERT(!(X) || (Y))
371 
372 /*
373  * Logical equivalence. Verify that both X and Y are either TRUE or FALSE.
374  */
375 #define	EQUIV(X, Y)	{ IMPLY(X, Y); IMPLY(Y, X); }
376 
377 /*
378  * Verify correctness of list head/tail pointers.
379  */
380 #define	LISTCHECK(head, tail, link) {				\
381 	EQUIV(head, tail);					\
382 	IMPLY(tail != NULL, tail->link == NULL);		\
383 }
384 
385 /*
386  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
387  * using a `link' field.
388  */
389 #define	ENQUEUE(el, head, tail, link) {				\
390 	ASSERT(el->link == NULL);				\
391 	LISTCHECK(head, tail, link);				\
392 	if (head == NULL)					\
393 		head = el;					\
394 	else							\
395 		tail->link = el;				\
396 	tail = el;						\
397 }
398 
399 /*
400  * Dequeue the first element of the list denoted by `head' and `tail' pointers
401  * using a `link' field and put result into `el'.
402  */
403 #define	DQ(el, head, tail, link) {				\
404 	LISTCHECK(head, tail, link);				\
405 	el = head;						\
406 	if (head != NULL) {					\
407 		head = head->link;				\
408 		if (head == NULL)				\
409 			tail = NULL;				\
410 		el->link = NULL;				\
411 	}							\
412 }
413 
414 /*
415  * Remove `el' from the list using `chase' and `curr' pointers and return result
416  * in `succeed'.
417  */
418 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
419 	LISTCHECK(head, tail, link);				\
420 	chase = NULL;						\
421 	succeed = 0;						\
422 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
423 		chase = curr;					\
424 	if (curr != NULL) {					\
425 		succeed = 1;					\
426 		ASSERT(curr == el);				\
427 		if (chase != NULL)				\
428 			chase->link = curr->link;		\
429 		else						\
430 			head = curr->link;			\
431 		curr->link = NULL;				\
432 		if (curr == tail)				\
433 			tail = chase;				\
434 	}							\
435 	LISTCHECK(head, tail, link);				\
436 }
437 
438 /* Handling of delayed messages on the inner syncq. */
439 
440 /*
441  * DEBUG versions should use function versions (to simplify tracing) and
442  * non-DEBUG kernels should use macro versions.
443  */
444 
445 /*
446  * Put a queue on the syncq list of queues.
447  * Assumes SQLOCK held.
448  */
449 #define	SQPUT_Q(sq, qp)							\
450 {									\
451 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
452 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
453 		/* The queue should not be linked anywhere */		\
454 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
455 		/* Head and tail may only be NULL simultaneously */	\
456 		EQUIV(sq->sq_head, sq->sq_tail);			\
457 		/* Queue may be only enqueyed on its syncq */		\
458 		ASSERT(sq == qp->q_syncq);				\
459 		/* Check the correctness of SQ_MESSAGES flag */		\
460 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
461 		/* Sanity check first/last elements of the list */	\
462 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
463 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
464 		/*							\
465 		 * Sanity check of priority field: empty queue should	\
466 		 * have zero priority					\
467 		 * and nqueues equal to zero.				\
468 		 */							\
469 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
470 		/* Sanity check of sq_nqueues field */			\
471 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
472 		if (sq->sq_head == NULL) {				\
473 			sq->sq_head = sq->sq_tail = qp;			\
474 			sq->sq_flags |= SQ_MESSAGES;			\
475 		} else if (qp->q_spri == 0) {				\
476 			qp->q_sqprev = sq->sq_tail;			\
477 			sq->sq_tail->q_sqnext = qp;			\
478 			sq->sq_tail = qp;				\
479 		} else {						\
480 			/*						\
481 			 * Put this queue in priority order: higher	\
482 			 * priority gets closer to the head.		\
483 			 */						\
484 			queue_t **qpp = &sq->sq_tail;			\
485 			queue_t *qnext = NULL;				\
486 									\
487 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
488 				qnext = *qpp;				\
489 				qpp = &(*qpp)->q_sqprev;		\
490 			}						\
491 			qp->q_sqnext = qnext;				\
492 			qp->q_sqprev = *qpp;				\
493 			if (*qpp != NULL) {				\
494 				(*qpp)->q_sqnext = qp;			\
495 			} else {					\
496 				sq->sq_head = qp;			\
497 				sq->sq_pri = sq->sq_head->q_spri;	\
498 			}						\
499 			*qpp = qp;					\
500 		}							\
501 		qp->q_sqflags |= Q_SQQUEUED;				\
502 		qp->q_sqtstamp = lbolt;					\
503 		sq->sq_nqueues++;					\
504 	}								\
505 }
506 
507 /*
508  * Remove a queue from the syncq list
509  * Assumes SQLOCK held.
510  */
511 #define	SQRM_Q(sq, qp)							\
512 	{								\
513 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
514 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
515 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
516 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
517 		/* Check that the queue is actually in the list */	\
518 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
519 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
520 		ASSERT(sq->sq_nqueues != 0);				\
521 		if (qp->q_sqprev == NULL) {				\
522 			/* First queue on list, make head q_sqnext */	\
523 			sq->sq_head = qp->q_sqnext;			\
524 		} else {						\
525 			/* Make prev->next == next */			\
526 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
527 		}							\
528 		if (qp->q_sqnext == NULL) {				\
529 			/* Last queue on list, make tail sqprev */	\
530 			sq->sq_tail = qp->q_sqprev;			\
531 		} else {						\
532 			/* Make next->prev == prev */			\
533 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
534 		}							\
535 		/* clear out references on this queue */		\
536 		qp->q_sqprev = qp->q_sqnext = NULL;			\
537 		qp->q_sqflags &= ~Q_SQQUEUED;				\
538 		/* If there is nothing queued, clear SQ_MESSAGES */	\
539 		if (sq->sq_head != NULL) {				\
540 			sq->sq_pri = sq->sq_head->q_spri;		\
541 		} else	{						\
542 			sq->sq_flags &= ~SQ_MESSAGES;			\
543 			sq->sq_pri = 0;					\
544 		}							\
545 		sq->sq_nqueues--;					\
546 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
547 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
548 	}
549 
550 /* Hide the definition from the header file. */
551 #ifdef SQPUT_MP
552 #undef SQPUT_MP
553 #endif
554 
555 /*
556  * Put a message on the queue syncq.
557  * Assumes QLOCK held.
558  */
559 #define	SQPUT_MP(qp, mp)						\
560 	{								\
561 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
562 		ASSERT(qp->q_sqhead == NULL ||				\
563 		    (qp->q_sqtail != NULL &&				\
564 		    qp->q_sqtail->b_next == NULL));			\
565 		qp->q_syncqmsgs++;					\
566 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
567 		if (qp->q_sqhead == NULL) {				\
568 			qp->q_sqhead = qp->q_sqtail = mp;		\
569 		} else {						\
570 			qp->q_sqtail->b_next = mp;			\
571 			qp->q_sqtail = mp;				\
572 		}							\
573 		ASSERT(qp->q_syncqmsgs > 0);				\
574 	}
575 
576 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
577 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
578 		if ((sq)->sq_ciputctrl != NULL) {			\
579 			int i;						\
580 			int nlocks = (sq)->sq_nciputctrl;		\
581 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
582 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
583 			for (i = 0; i <= nlocks; i++) {			\
584 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
585 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
586 			}						\
587 		}							\
588 	}
589 
590 
591 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
592 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
593 		if ((sq)->sq_ciputctrl != NULL) {			\
594 			int i;						\
595 			int nlocks = (sq)->sq_nciputctrl;		\
596 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
597 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
598 			for (i = 0; i <= nlocks; i++) {			\
599 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
600 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
601 			}						\
602 		}							\
603 	}
604 
605 /*
606  * Run service procedures for all queues in the stream head.
607  */
608 #define	STR_SERVICE(stp, q) {						\
609 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
610 	while (stp->sd_qhead != NULL) {					\
611 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
612 		ASSERT(stp->sd_nqueues > 0);				\
613 		stp->sd_nqueues--;					\
614 		ASSERT(!(q->q_flag & QINSERVICE));			\
615 		mutex_exit(&stp->sd_qlock);				\
616 		queue_service(q);					\
617 		mutex_enter(&stp->sd_qlock);				\
618 	}								\
619 	ASSERT(stp->sd_nqueues == 0);					\
620 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
621 }
622 
623 /*
624  * constructor/destructor routines for the stream head cache
625  */
626 /* ARGSUSED */
627 static int
628 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
629 {
630 	stdata_t *stp = buf;
631 
632 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
633 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
634 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
635 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
636 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
637 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
638 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
639 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
640 	stp->sd_wrq = NULL;
641 
642 	return (0);
643 }
644 
645 /* ARGSUSED */
646 static void
647 stream_head_destructor(void *buf, void *cdrarg)
648 {
649 	stdata_t *stp = buf;
650 
651 	mutex_destroy(&stp->sd_lock);
652 	mutex_destroy(&stp->sd_reflock);
653 	mutex_destroy(&stp->sd_qlock);
654 	cv_destroy(&stp->sd_monitor);
655 	cv_destroy(&stp->sd_iocmonitor);
656 	cv_destroy(&stp->sd_refmonitor);
657 	cv_destroy(&stp->sd_qcv);
658 	cv_destroy(&stp->sd_zcopy_wait);
659 }
660 
661 /*
662  * constructor/destructor routines for the queue cache
663  */
664 /* ARGSUSED */
665 static int
666 queue_constructor(void *buf, void *cdrarg, int kmflags)
667 {
668 	queinfo_t *qip = buf;
669 	queue_t *qp = &qip->qu_rqueue;
670 	queue_t *wqp = &qip->qu_wqueue;
671 	syncq_t	*sq = &qip->qu_syncq;
672 
673 	qp->q_first = NULL;
674 	qp->q_link = NULL;
675 	qp->q_count = 0;
676 	qp->q_mblkcnt = 0;
677 	qp->q_sqhead = NULL;
678 	qp->q_sqtail = NULL;
679 	qp->q_sqnext = NULL;
680 	qp->q_sqprev = NULL;
681 	qp->q_sqflags = 0;
682 	qp->q_rwcnt = 0;
683 	qp->q_spri = 0;
684 
685 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
686 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
687 
688 	wqp->q_first = NULL;
689 	wqp->q_link = NULL;
690 	wqp->q_count = 0;
691 	wqp->q_mblkcnt = 0;
692 	wqp->q_sqhead = NULL;
693 	wqp->q_sqtail = NULL;
694 	wqp->q_sqnext = NULL;
695 	wqp->q_sqprev = NULL;
696 	wqp->q_sqflags = 0;
697 	wqp->q_rwcnt = 0;
698 	wqp->q_spri = 0;
699 
700 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
701 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
702 
703 	sq->sq_head = NULL;
704 	sq->sq_tail = NULL;
705 	sq->sq_evhead = NULL;
706 	sq->sq_evtail = NULL;
707 	sq->sq_callbpend = NULL;
708 	sq->sq_outer = NULL;
709 	sq->sq_onext = NULL;
710 	sq->sq_oprev = NULL;
711 	sq->sq_next = NULL;
712 	sq->sq_svcflags = 0;
713 	sq->sq_servcount = 0;
714 	sq->sq_needexcl = 0;
715 	sq->sq_nqueues = 0;
716 	sq->sq_pri = 0;
717 
718 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
719 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
720 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
721 
722 	return (0);
723 }
724 
725 /* ARGSUSED */
726 static void
727 queue_destructor(void *buf, void *cdrarg)
728 {
729 	queinfo_t *qip = buf;
730 	queue_t *qp = &qip->qu_rqueue;
731 	queue_t *wqp = &qip->qu_wqueue;
732 	syncq_t	*sq = &qip->qu_syncq;
733 
734 	ASSERT(qp->q_sqhead == NULL);
735 	ASSERT(wqp->q_sqhead == NULL);
736 	ASSERT(qp->q_sqnext == NULL);
737 	ASSERT(wqp->q_sqnext == NULL);
738 	ASSERT(qp->q_rwcnt == 0);
739 	ASSERT(wqp->q_rwcnt == 0);
740 
741 	mutex_destroy(&qp->q_lock);
742 	cv_destroy(&qp->q_wait);
743 
744 	mutex_destroy(&wqp->q_lock);
745 	cv_destroy(&wqp->q_wait);
746 
747 	mutex_destroy(&sq->sq_lock);
748 	cv_destroy(&sq->sq_wait);
749 	cv_destroy(&sq->sq_exitwait);
750 }
751 
752 /*
753  * constructor/destructor routines for the syncq cache
754  */
755 /* ARGSUSED */
756 static int
757 syncq_constructor(void *buf, void *cdrarg, int kmflags)
758 {
759 	syncq_t	*sq = buf;
760 
761 	bzero(buf, sizeof (syncq_t));
762 
763 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
764 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
765 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
766 
767 	return (0);
768 }
769 
770 /* ARGSUSED */
771 static void
772 syncq_destructor(void *buf, void *cdrarg)
773 {
774 	syncq_t	*sq = buf;
775 
776 	ASSERT(sq->sq_head == NULL);
777 	ASSERT(sq->sq_tail == NULL);
778 	ASSERT(sq->sq_evhead == NULL);
779 	ASSERT(sq->sq_evtail == NULL);
780 	ASSERT(sq->sq_callbpend == NULL);
781 	ASSERT(sq->sq_callbflags == 0);
782 	ASSERT(sq->sq_outer == NULL);
783 	ASSERT(sq->sq_onext == NULL);
784 	ASSERT(sq->sq_oprev == NULL);
785 	ASSERT(sq->sq_next == NULL);
786 	ASSERT(sq->sq_needexcl == 0);
787 	ASSERT(sq->sq_svcflags == 0);
788 	ASSERT(sq->sq_servcount == 0);
789 	ASSERT(sq->sq_nqueues == 0);
790 	ASSERT(sq->sq_pri == 0);
791 	ASSERT(sq->sq_count == 0);
792 	ASSERT(sq->sq_rmqcount == 0);
793 	ASSERT(sq->sq_cancelid == 0);
794 	ASSERT(sq->sq_ciputctrl == NULL);
795 	ASSERT(sq->sq_nciputctrl == 0);
796 	ASSERT(sq->sq_type == 0);
797 	ASSERT(sq->sq_flags == 0);
798 
799 	mutex_destroy(&sq->sq_lock);
800 	cv_destroy(&sq->sq_wait);
801 	cv_destroy(&sq->sq_exitwait);
802 }
803 
804 /* ARGSUSED */
805 static int
806 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
807 {
808 	ciputctrl_t *cip = buf;
809 	int i;
810 
811 	for (i = 0; i < n_ciputctrl; i++) {
812 		cip[i].ciputctrl_count = SQ_FASTPUT;
813 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
814 	}
815 
816 	return (0);
817 }
818 
819 /* ARGSUSED */
820 static void
821 ciputctrl_destructor(void *buf, void *cdrarg)
822 {
823 	ciputctrl_t *cip = buf;
824 	int i;
825 
826 	for (i = 0; i < n_ciputctrl; i++) {
827 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
828 		mutex_destroy(&cip[i].ciputctrl_lock);
829 	}
830 }
831 
832 /*
833  * Init routine run from main at boot time.
834  */
835 void
836 strinit(void)
837 {
838 	int i;
839 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
840 
841 	/*
842 	 * Set up mux_node structures.
843 	 */
844 	mux_nodes = kmem_zalloc((sizeof (struct mux_node) * devcnt), KM_SLEEP);
845 	for (i = 0; i < devcnt; i++)
846 		mux_nodes[i].mn_imaj = i;
847 
848 	stream_head_cache = kmem_cache_create("stream_head_cache",
849 		sizeof (stdata_t), 0,
850 		stream_head_constructor, stream_head_destructor, NULL,
851 		NULL, NULL, 0);
852 
853 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
854 		queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
855 
856 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
857 		syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
858 
859 	qband_cache = kmem_cache_create("qband_cache",
860 		sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
861 
862 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
863 		sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
864 
865 	n_ciputctrl = ncpus;
866 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
867 	ASSERT(n_ciputctrl >= 1);
868 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
869 	if (n_ciputctrl >= min_n_ciputctrl) {
870 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
871 			sizeof (ciputctrl_t) * n_ciputctrl,
872 			sizeof (ciputctrl_t), ciputctrl_constructor,
873 			ciputctrl_destructor, NULL, NULL, NULL, 0);
874 	}
875 
876 	streams_taskq = system_taskq;
877 
878 	if (streams_taskq == NULL)
879 		panic("strinit: no memory for streams taskq!");
880 
881 	bc_bkgrnd_thread = thread_create(NULL, 0,
882 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
883 
884 	streams_qbkgrnd_thread = thread_create(NULL, 0,
885 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
886 
887 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
888 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
889 
890 	/*
891 	 * Create STREAMS kstats.
892 	 */
893 	str_kstat = kstat_create("streams", 0, "strstat",
894 	    "net", KSTAT_TYPE_NAMED,
895 	    sizeof (str_statistics) / sizeof (kstat_named_t),
896 	    KSTAT_FLAG_VIRTUAL);
897 
898 	if (str_kstat != NULL) {
899 		str_kstat->ks_data = &str_statistics;
900 		kstat_install(str_kstat);
901 	}
902 
903 	/*
904 	 * TPI support routine initialisation.
905 	 */
906 	tpi_init();
907 }
908 
909 void
910 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
911 {
912 	struct stdata *stp;
913 
914 	ASSERT(vp->v_stream);
915 	stp = vp->v_stream;
916 	/* Have to hold sd_lock to prevent siglist from changing */
917 	mutex_enter(&stp->sd_lock);
918 	if (stp->sd_sigflags & event)
919 		strsendsig(stp->sd_siglist, event, band, error);
920 	mutex_exit(&stp->sd_lock);
921 }
922 
923 /*
924  * Send the "sevent" set of signals to a process.
925  * This might send more than one signal if the process is registered
926  * for multiple events. The caller should pass in an sevent that only
927  * includes the events for which the process has registered.
928  */
929 static void
930 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
931 	uchar_t band, int error)
932 {
933 	ASSERT(MUTEX_HELD(&proc->p_lock));
934 
935 	info->si_band = 0;
936 	info->si_errno = 0;
937 
938 	if (sevent & S_ERROR) {
939 		sevent &= ~S_ERROR;
940 		info->si_code = POLL_ERR;
941 		info->si_errno = error;
942 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
943 			"strsendsig:proc %p info %p", proc, info);
944 		sigaddq(proc, NULL, info, KM_NOSLEEP);
945 		info->si_errno = 0;
946 	}
947 	if (sevent & S_HANGUP) {
948 		sevent &= ~S_HANGUP;
949 		info->si_code = POLL_HUP;
950 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
951 			"strsendsig:proc %p info %p", proc, info);
952 		sigaddq(proc, NULL, info, KM_NOSLEEP);
953 	}
954 	if (sevent & S_HIPRI) {
955 		sevent &= ~S_HIPRI;
956 		info->si_code = POLL_PRI;
957 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
958 			"strsendsig:proc %p info %p", proc, info);
959 		sigaddq(proc, NULL, info, KM_NOSLEEP);
960 	}
961 	if (sevent & S_RDBAND) {
962 		sevent &= ~S_RDBAND;
963 		if (events & S_BANDURG)
964 			sigtoproc(proc, NULL, SIGURG);
965 		else
966 			sigtoproc(proc, NULL, SIGPOLL);
967 	}
968 	if (sevent & S_WRBAND) {
969 		sevent &= ~S_WRBAND;
970 		sigtoproc(proc, NULL, SIGPOLL);
971 	}
972 	if (sevent & S_INPUT) {
973 		sevent &= ~S_INPUT;
974 		info->si_code = POLL_IN;
975 		info->si_band = band;
976 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
977 			"strsendsig:proc %p info %p", proc, info);
978 		sigaddq(proc, NULL, info, KM_NOSLEEP);
979 		info->si_band = 0;
980 	}
981 	if (sevent & S_OUTPUT) {
982 		sevent &= ~S_OUTPUT;
983 		info->si_code = POLL_OUT;
984 		info->si_band = band;
985 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
986 			"strsendsig:proc %p info %p", proc, info);
987 		sigaddq(proc, NULL, info, KM_NOSLEEP);
988 		info->si_band = 0;
989 	}
990 	if (sevent & S_MSG) {
991 		sevent &= ~S_MSG;
992 		info->si_code = POLL_MSG;
993 		info->si_band = band;
994 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
995 			"strsendsig:proc %p info %p", proc, info);
996 		sigaddq(proc, NULL, info, KM_NOSLEEP);
997 		info->si_band = 0;
998 	}
999 	if (sevent & S_RDNORM) {
1000 		sevent &= ~S_RDNORM;
1001 		sigtoproc(proc, NULL, SIGPOLL);
1002 	}
1003 	if (sevent != 0) {
1004 		panic("strsendsig: unknown event(s) %x", sevent);
1005 	}
1006 }
1007 
1008 /*
1009  * Send SIGPOLL/SIGURG signal to all processes and process groups
1010  * registered on the given signal list that want a signal for at
1011  * least one of the specified events.
1012  *
1013  * Must be called with exclusive access to siglist (caller holding sd_lock).
1014  *
1015  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1016  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1017  * while it is in the siglist.
1018  *
1019  * For performance reasons (MP scalability) the code drops pidlock
1020  * when sending signals to a single process.
1021  * When sending to a process group the code holds
1022  * pidlock to prevent the membership in the process group from changing
1023  * while walking the p_pglink list.
1024  */
1025 void
1026 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1027 {
1028 	strsig_t *ssp;
1029 	k_siginfo_t info;
1030 	struct pid *pidp;
1031 	proc_t  *proc;
1032 
1033 	info.si_signo = SIGPOLL;
1034 	info.si_errno = 0;
1035 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1036 		int sevent;
1037 
1038 		sevent = ssp->ss_events & event;
1039 		if (sevent == 0)
1040 			continue;
1041 
1042 		if ((pidp = ssp->ss_pidp) == NULL) {
1043 			/* pid was released but still on event list */
1044 			continue;
1045 		}
1046 
1047 
1048 		if (ssp->ss_pid > 0) {
1049 			/*
1050 			 * XXX This unfortunately still generates
1051 			 * a signal when a fd is closed but
1052 			 * the proc is active.
1053 			 */
1054 			ASSERT(ssp->ss_pid == pidp->pid_id);
1055 
1056 			mutex_enter(&pidlock);
1057 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1058 			if (proc == NULL) {
1059 				mutex_exit(&pidlock);
1060 				continue;
1061 			}
1062 			mutex_enter(&proc->p_lock);
1063 			mutex_exit(&pidlock);
1064 			dosendsig(proc, ssp->ss_events, sevent, &info,
1065 				band, error);
1066 			mutex_exit(&proc->p_lock);
1067 		} else {
1068 			/*
1069 			 * Send to process group. Hold pidlock across
1070 			 * calls to dosendsig().
1071 			 */
1072 			pid_t pgrp = -ssp->ss_pid;
1073 
1074 			mutex_enter(&pidlock);
1075 			proc = pgfind_zone(pgrp, ALL_ZONES);
1076 			while (proc != NULL) {
1077 				mutex_enter(&proc->p_lock);
1078 				dosendsig(proc, ssp->ss_events, sevent,
1079 					&info, band, error);
1080 				mutex_exit(&proc->p_lock);
1081 				proc = proc->p_pglink;
1082 			}
1083 			mutex_exit(&pidlock);
1084 		}
1085 	}
1086 }
1087 
1088 /*
1089  * Attach a stream device or module.
1090  * qp is a read queue; the new queue goes in so its next
1091  * read ptr is the argument, and the write queue corresponding
1092  * to the argument points to this queue. Return 0 on success,
1093  * or a non-zero errno on failure.
1094  */
1095 int
1096 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1097     boolean_t is_insert)
1098 {
1099 	major_t			major;
1100 	cdevsw_impl_t		*dp;
1101 	struct streamtab	*str;
1102 	queue_t			*rq;
1103 	queue_t			*wrq;
1104 	uint32_t		qflag;
1105 	uint32_t		sqtype;
1106 	perdm_t			*dmp;
1107 	int			error;
1108 	int			sflag;
1109 
1110 	rq = allocq();
1111 	wrq = _WR(rq);
1112 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1113 
1114 	if (fp != NULL) {
1115 		str = fp->f_str;
1116 		qflag = fp->f_qflag;
1117 		sqtype = fp->f_sqtype;
1118 		dmp = fp->f_dmp;
1119 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1120 		sflag = MODOPEN;
1121 
1122 		/*
1123 		 * stash away a pointer to the module structure so we can
1124 		 * unref it in qdetach.
1125 		 */
1126 		rq->q_fp = fp;
1127 	} else {
1128 		ASSERT(!is_insert);
1129 
1130 		major = getmajor(*devp);
1131 		dp = &devimpl[major];
1132 
1133 		str = dp->d_str;
1134 		ASSERT(str == STREAMSTAB(major));
1135 
1136 		qflag = dp->d_qflag;
1137 		ASSERT(qflag & QISDRV);
1138 		sqtype = dp->d_sqtype;
1139 
1140 		/* create perdm_t if needed */
1141 		if (NEED_DM(dp->d_dmp, qflag))
1142 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1143 
1144 		dmp = dp->d_dmp;
1145 		sflag = 0;
1146 	}
1147 
1148 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1149 	    "qattach:qflag == %X(%X)", qflag, *devp);
1150 
1151 	/* setq might sleep in allocator - avoid holding locks. */
1152 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1153 
1154 	/*
1155 	 * Before calling the module's open routine, set up the q_next
1156 	 * pointer for inserting a module in the middle of a stream.
1157 	 *
1158 	 * Note that we can always set _QINSERTING and set up q_next
1159 	 * pointer for both inserting and pushing a module.  Then there
1160 	 * is no need for the is_insert parameter.  In insertq(), called
1161 	 * by qprocson(), assume that q_next of the new module always points
1162 	 * to the correct queue and use it for insertion.  Everything should
1163 	 * work out fine.  But in the first release of _I_INSERT, we
1164 	 * distinguish between inserting and pushing to make sure that
1165 	 * pushing a module follows the same code path as before.
1166 	 */
1167 	if (is_insert) {
1168 		rq->q_flag |= _QINSERTING;
1169 		rq->q_next = qp;
1170 	}
1171 
1172 	/*
1173 	 * If there is an outer perimeter get exclusive access during
1174 	 * the open procedure.  Bump up the reference count on the queue.
1175 	 */
1176 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1177 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1178 	if (error != 0)
1179 		goto failed;
1180 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1181 	ASSERT(qprocsareon(rq));
1182 	return (0);
1183 
1184 failed:
1185 	rq->q_flag &= ~_QINSERTING;
1186 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1187 		qprocsoff(rq);
1188 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1189 	rq->q_next = wrq->q_next = NULL;
1190 	qdetach(rq, 0, 0, crp, B_FALSE);
1191 	return (error);
1192 }
1193 
1194 /*
1195  * Handle second open of stream. For modules, set the
1196  * last argument to MODOPEN and do not pass any open flags.
1197  * Ignore dummydev since this is not the first open.
1198  */
1199 int
1200 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1201 {
1202 	int	error;
1203 	dev_t dummydev;
1204 	queue_t *wqp = _WR(qp);
1205 
1206 	ASSERT(qp->q_flag & QREADR);
1207 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1208 
1209 	dummydev = *devp;
1210 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1211 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1212 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1213 		mutex_enter(&STREAM(qp)->sd_lock);
1214 		qp->q_stream->sd_flag |= STREOPENFAIL;
1215 		mutex_exit(&STREAM(qp)->sd_lock);
1216 		return (error);
1217 	}
1218 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1219 
1220 	/*
1221 	 * successful open should have done qprocson()
1222 	 */
1223 	ASSERT(qprocsareon(_RD(qp)));
1224 	return (0);
1225 }
1226 
1227 /*
1228  * Detach a stream module or device.
1229  * If clmode == 1 then the module or driver was opened and its
1230  * close routine must be called. If clmode == 0, the module
1231  * or driver was never opened or the open failed, and so its close
1232  * should not be called.
1233  */
1234 void
1235 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1236 {
1237 	queue_t *wqp = _WR(qp);
1238 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1239 
1240 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1241 		stream_runservice(STREAM(qp));
1242 
1243 	if (clmode) {
1244 		/*
1245 		 * Make sure that all the messages on the write side syncq are
1246 		 * processed and nothing is left. Since we are closing, no new
1247 		 * messages may appear there.
1248 		 */
1249 		wait_q_syncq(wqp);
1250 
1251 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1252 		if (is_remove) {
1253 			mutex_enter(QLOCK(qp));
1254 			qp->q_flag |= _QREMOVING;
1255 			mutex_exit(QLOCK(qp));
1256 		}
1257 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1258 		/*
1259 		 * Check that qprocsoff() was actually called.
1260 		 */
1261 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1262 
1263 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1264 	} else {
1265 		disable_svc(qp);
1266 	}
1267 
1268 	/*
1269 	 * Allow any threads blocked in entersq to proceed and discover
1270 	 * the QWCLOSE is set.
1271 	 * Note: This assumes that all users of entersq check QWCLOSE.
1272 	 * Currently runservice is the only entersq that can happen
1273 	 * after removeq has finished.
1274 	 * Removeq will have discarded all messages destined to the closing
1275 	 * pair of queues from the syncq.
1276 	 * NOTE: Calling a function inside an assert is unconventional.
1277 	 * However, it does not cause any problem since flush_syncq() does
1278 	 * not change any state except when it returns non-zero i.e.
1279 	 * when the assert will trigger.
1280 	 */
1281 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1282 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1283 	ASSERT((qp->q_flag & QPERMOD) ||
1284 		((qp->q_syncq->sq_head == NULL) &&
1285 		(wqp->q_syncq->sq_head == NULL)));
1286 
1287 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1288 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1289 	if (qp->q_fp != NULL)
1290 		fmodsw_rele(qp->q_fp);
1291 
1292 	/* freeq removes us from the outer perimeter if any */
1293 	freeq(qp);
1294 }
1295 
1296 /* Prevent service procedures from being called */
1297 void
1298 disable_svc(queue_t *qp)
1299 {
1300 	queue_t *wqp = _WR(qp);
1301 
1302 	ASSERT(qp->q_flag & QREADR);
1303 	mutex_enter(QLOCK(qp));
1304 	qp->q_flag |= QWCLOSE;
1305 	mutex_exit(QLOCK(qp));
1306 	mutex_enter(QLOCK(wqp));
1307 	wqp->q_flag |= QWCLOSE;
1308 	mutex_exit(QLOCK(wqp));
1309 }
1310 
1311 /* allow service procedures to be called again */
1312 void
1313 enable_svc(queue_t *qp)
1314 {
1315 	queue_t *wqp = _WR(qp);
1316 
1317 	ASSERT(qp->q_flag & QREADR);
1318 	mutex_enter(QLOCK(qp));
1319 	qp->q_flag &= ~QWCLOSE;
1320 	mutex_exit(QLOCK(qp));
1321 	mutex_enter(QLOCK(wqp));
1322 	wqp->q_flag &= ~QWCLOSE;
1323 	mutex_exit(QLOCK(wqp));
1324 }
1325 
1326 /*
1327  * Remove queue from qhead/qtail if it is enabled.
1328  * Only reset QENAB if the queue was removed from the runlist.
1329  * A queue goes through 3 stages:
1330  *	It is on the service list and QENAB is set.
1331  *	It is removed from the service list but QENAB is still set.
1332  *	QENAB gets changed to QINSERVICE.
1333  *	QINSERVICE is reset (when the service procedure is done)
1334  * Thus we can not reset QENAB unless we actually removed it from the service
1335  * queue.
1336  */
1337 void
1338 remove_runlist(queue_t *qp)
1339 {
1340 	if (qp->q_flag & QENAB && qhead != NULL) {
1341 		queue_t *q_chase;
1342 		queue_t *q_curr;
1343 		int removed;
1344 
1345 		mutex_enter(&service_queue);
1346 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1347 		mutex_exit(&service_queue);
1348 		if (removed) {
1349 			STRSTAT(qremoved);
1350 			qp->q_flag &= ~QENAB;
1351 		}
1352 	}
1353 }
1354 
1355 
1356 /*
1357  * wait for any pending service processing to complete.
1358  * The removal of queues from the runlist is not atomic with the
1359  * clearing of the QENABLED flag and setting the INSERVICE flag.
1360  * consequently it is possible for remove_runlist in strclose
1361  * to not find the queue on the runlist but for it to be QENABLED
1362  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1363  * as well as INSERVICE.
1364  */
1365 void
1366 wait_svc(queue_t *qp)
1367 {
1368 	queue_t *wqp = _WR(qp);
1369 
1370 	ASSERT(qp->q_flag & QREADR);
1371 
1372 	/*
1373 	 * Try to remove queues from qhead/qtail list.
1374 	 */
1375 	if (qhead != NULL) {
1376 		remove_runlist(qp);
1377 		remove_runlist(wqp);
1378 	}
1379 	/*
1380 	 * Wait till the syncqs associated with the queue
1381 	 * will dissapear from background processing list.
1382 	 * This only needs to be done for non-PERMOD perimeters since
1383 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1384 	 * when the last module/driver is unloaded.
1385 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1386 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1387 	 * function remove the queue from its syncq list, so sqthread will not
1388 	 * try to access the queue.
1389 	 */
1390 	if (!(qp->q_flag & QPERMOD)) {
1391 		syncq_t *rsq = qp->q_syncq;
1392 		syncq_t *wsq = wqp->q_syncq;
1393 
1394 		/*
1395 		 * Disable rsq and wsq and wait for any background processing of
1396 		 * syncq to complete.
1397 		 */
1398 		wait_sq_svc(rsq);
1399 		if (wsq != rsq)
1400 			wait_sq_svc(wsq);
1401 	}
1402 
1403 	mutex_enter(QLOCK(qp));
1404 	while (qp->q_flag & (QINSERVICE|QENAB))
1405 		cv_wait(&qp->q_wait, QLOCK(qp));
1406 	mutex_exit(QLOCK(qp));
1407 	mutex_enter(QLOCK(wqp));
1408 	while (wqp->q_flag & (QINSERVICE|QENAB))
1409 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1410 	mutex_exit(QLOCK(wqp));
1411 }
1412 
1413 /*
1414  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1415  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1416  * also be set, and is passed through to allocb_cred_wait().
1417  *
1418  * Returns errno on failure, zero on success.
1419  */
1420 int
1421 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1422 {
1423 	mblk_t *tmp;
1424 	ssize_t  count;
1425 	size_t n;
1426 	int error = 0;
1427 
1428 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1429 		(flag & (U_TO_K | K_TO_K)) == K_TO_K);
1430 
1431 	if (bp->b_datap->db_type == M_IOCTL) {
1432 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1433 	} else {
1434 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1435 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1436 	}
1437 	/*
1438 	 * strdoioctl validates ioc_count, so if this assert fails it
1439 	 * cannot be due to user error.
1440 	 */
1441 	ASSERT(count >= 0);
1442 
1443 	while (count > 0) {
1444 		n = MIN(MAXIOCBSZ, count);
1445 		if ((tmp = allocb_cred_wait(n, (flag & STR_NOSIG), &error,
1446 		    cr)) == NULL) {
1447 			return (error);
1448 		}
1449 		error = strcopyin(arg, tmp->b_wptr, n, flag & (U_TO_K|K_TO_K));
1450 		if (error != 0) {
1451 			freeb(tmp);
1452 			return (error);
1453 		}
1454 		arg += n;
1455 		DB_CPID(tmp) = curproc->p_pid;
1456 		tmp->b_wptr += n;
1457 		count -= n;
1458 		bp = (bp->b_cont = tmp);
1459 	}
1460 
1461 	return (0);
1462 }
1463 
1464 /*
1465  * Copy ioctl data to user-land. Return non-zero errno on failure,
1466  * 0 for success.
1467  */
1468 int
1469 getiocd(mblk_t *bp, char *arg, int copymode)
1470 {
1471 	ssize_t count;
1472 	size_t  n;
1473 	int	error;
1474 
1475 	if (bp->b_datap->db_type == M_IOCACK)
1476 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1477 	else {
1478 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1479 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1480 	}
1481 	ASSERT(count >= 0);
1482 
1483 	for (bp = bp->b_cont; bp && count;
1484 	    count -= n, bp = bp->b_cont, arg += n) {
1485 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1486 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1487 		if (error)
1488 			return (error);
1489 	}
1490 	ASSERT(count == 0);
1491 	return (0);
1492 }
1493 
1494 /*
1495  * Allocate a linkinfo entry given the write queue of the
1496  * bottom module of the top stream and the write queue of the
1497  * stream head of the bottom stream.
1498  */
1499 linkinfo_t *
1500 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1501 {
1502 	linkinfo_t *linkp;
1503 
1504 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1505 
1506 	linkp->li_lblk.l_qtop = qup;
1507 	linkp->li_lblk.l_qbot = qdown;
1508 	linkp->li_fpdown = fpdown;
1509 
1510 	mutex_enter(&strresources);
1511 	linkp->li_next = linkinfo_list;
1512 	linkp->li_prev = NULL;
1513 	if (linkp->li_next)
1514 		linkp->li_next->li_prev = linkp;
1515 	linkinfo_list = linkp;
1516 	linkp->li_lblk.l_index = ++lnk_id;
1517 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1518 	mutex_exit(&strresources);
1519 
1520 	return (linkp);
1521 }
1522 
1523 /*
1524  * Free a linkinfo entry.
1525  */
1526 void
1527 lbfree(linkinfo_t *linkp)
1528 {
1529 	mutex_enter(&strresources);
1530 	if (linkp->li_next)
1531 		linkp->li_next->li_prev = linkp->li_prev;
1532 	if (linkp->li_prev)
1533 		linkp->li_prev->li_next = linkp->li_next;
1534 	else
1535 		linkinfo_list = linkp->li_next;
1536 	mutex_exit(&strresources);
1537 
1538 	kmem_cache_free(linkinfo_cache, linkp);
1539 }
1540 
1541 /*
1542  * Check for a potential linking cycle.
1543  * Return 1 if a link will result in a cycle,
1544  * and 0 otherwise.
1545  */
1546 int
1547 linkcycle(stdata_t *upstp, stdata_t *lostp)
1548 {
1549 	struct mux_node *np;
1550 	struct mux_edge *ep;
1551 	int i;
1552 	major_t lomaj;
1553 	major_t upmaj;
1554 	/*
1555 	 * if the lower stream is a pipe/FIFO, return, since link
1556 	 * cycles can not happen on pipes/FIFOs
1557 	 */
1558 	if (lostp->sd_vnode->v_type == VFIFO)
1559 		return (0);
1560 
1561 	for (i = 0; i < devcnt; i++) {
1562 		np = &mux_nodes[i];
1563 		MUX_CLEAR(np);
1564 	}
1565 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1566 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1567 	np = &mux_nodes[lomaj];
1568 	for (;;) {
1569 		if (!MUX_DIDVISIT(np)) {
1570 			if (np->mn_imaj == upmaj)
1571 				return (1);
1572 			if (np->mn_outp == NULL) {
1573 				MUX_VISIT(np);
1574 				if (np->mn_originp == NULL)
1575 					return (0);
1576 				np = np->mn_originp;
1577 				continue;
1578 			}
1579 			MUX_VISIT(np);
1580 			np->mn_startp = np->mn_outp;
1581 		} else {
1582 			if (np->mn_startp == NULL) {
1583 				if (np->mn_originp == NULL)
1584 					return (0);
1585 				else {
1586 					np = np->mn_originp;
1587 					continue;
1588 				}
1589 			}
1590 			/*
1591 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1592 			 * ignore the edge and move on. ep->me_nodep gets
1593 			 * set to NULL in mux_addedge() if it is a FIFO.
1594 			 *
1595 			 */
1596 			ep = np->mn_startp;
1597 			np->mn_startp = ep->me_nextp;
1598 			if (ep->me_nodep == NULL)
1599 				continue;
1600 			ep->me_nodep->mn_originp = np;
1601 			np = ep->me_nodep;
1602 		}
1603 	}
1604 }
1605 
1606 /*
1607  * Find linkinfo entry corresponding to the parameters.
1608  */
1609 linkinfo_t *
1610 findlinks(stdata_t *stp, int index, int type)
1611 {
1612 	linkinfo_t *linkp;
1613 	struct mux_edge *mep;
1614 	struct mux_node *mnp;
1615 	queue_t *qup;
1616 
1617 	mutex_enter(&strresources);
1618 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1619 		qup = getendq(stp->sd_wrq);
1620 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1621 			if ((qup == linkp->li_lblk.l_qtop) &&
1622 			    (!index || (index == linkp->li_lblk.l_index))) {
1623 				mutex_exit(&strresources);
1624 				return (linkp);
1625 			}
1626 		}
1627 	} else {
1628 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1629 		mnp = &mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1630 		mep = mnp->mn_outp;
1631 		while (mep) {
1632 			if ((index == 0) || (index == mep->me_muxid))
1633 				break;
1634 			mep = mep->me_nextp;
1635 		}
1636 		if (!mep) {
1637 			mutex_exit(&strresources);
1638 			return (NULL);
1639 		}
1640 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1641 			if ((!linkp->li_lblk.l_qtop) &&
1642 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1643 				mutex_exit(&strresources);
1644 				return (linkp);
1645 			}
1646 		}
1647 	}
1648 	mutex_exit(&strresources);
1649 	return (NULL);
1650 }
1651 
1652 /*
1653  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1654  * last queue on the chain and return it.
1655  */
1656 queue_t *
1657 getendq(queue_t *q)
1658 {
1659 	ASSERT(q != NULL);
1660 	while (_SAMESTR(q))
1661 		q = q->q_next;
1662 	return (q);
1663 }
1664 
1665 /*
1666  * wait for the syncq count to drop to zero.
1667  * sq could be either outer or inner.
1668  */
1669 
1670 static void
1671 wait_syncq(syncq_t *sq)
1672 {
1673 	uint16_t count;
1674 
1675 	mutex_enter(SQLOCK(sq));
1676 	count = sq->sq_count;
1677 	SQ_PUTLOCKS_ENTER(sq);
1678 	SUM_SQ_PUTCOUNTS(sq, count);
1679 	while (count != 0) {
1680 		sq->sq_flags |= SQ_WANTWAKEUP;
1681 		SQ_PUTLOCKS_EXIT(sq);
1682 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1683 		count = sq->sq_count;
1684 		SQ_PUTLOCKS_ENTER(sq);
1685 		SUM_SQ_PUTCOUNTS(sq, count);
1686 	}
1687 	SQ_PUTLOCKS_EXIT(sq);
1688 	mutex_exit(SQLOCK(sq));
1689 }
1690 
1691 /*
1692  * Wait while there are any messages for the queue in its syncq.
1693  */
1694 static void
1695 wait_q_syncq(queue_t *q)
1696 {
1697 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1698 		syncq_t *sq = q->q_syncq;
1699 
1700 		mutex_enter(SQLOCK(sq));
1701 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1702 			sq->sq_flags |= SQ_WANTWAKEUP;
1703 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1704 		}
1705 		mutex_exit(SQLOCK(sq));
1706 	}
1707 }
1708 
1709 
1710 int
1711 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1712     int lhlink)
1713 {
1714 	struct stdata *stp;
1715 	struct strioctl strioc;
1716 	struct linkinfo *linkp;
1717 	struct stdata *stpdown;
1718 	struct streamtab *str;
1719 	queue_t *passq;
1720 	syncq_t *passyncq;
1721 	queue_t *rq;
1722 	cdevsw_impl_t *dp;
1723 	uint32_t qflag;
1724 	uint32_t sqtype;
1725 	perdm_t *dmp;
1726 	int error = 0;
1727 
1728 	stp = vp->v_stream;
1729 	TRACE_1(TR_FAC_STREAMS_FR,
1730 		TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1731 	/*
1732 	 * Test for invalid upper stream
1733 	 */
1734 	if (stp->sd_flag & STRHUP) {
1735 		return (ENXIO);
1736 	}
1737 	if (vp->v_type == VFIFO) {
1738 		return (EINVAL);
1739 	}
1740 	if (stp->sd_strtab == NULL) {
1741 		return (EINVAL);
1742 	}
1743 	if (!stp->sd_strtab->st_muxwinit) {
1744 		return (EINVAL);
1745 	}
1746 	if (fpdown == NULL) {
1747 		return (EBADF);
1748 	}
1749 	if (getmajor(stp->sd_vnode->v_rdev) >= devcnt) {
1750 		return (EINVAL);
1751 	}
1752 	mutex_enter(&muxifier);
1753 	if (stp->sd_flag & STPLEX) {
1754 		mutex_exit(&muxifier);
1755 		return (ENXIO);
1756 	}
1757 
1758 	/*
1759 	 * Test for invalid lower stream.
1760 	 * The check for the v_type != VFIFO and having a major
1761 	 * number not >= devcnt is done to avoid problems with
1762 	 * adding mux_node entry past the end of mux_nodes[].
1763 	 * For FIFO's we don't add an entry so this isn't a
1764 	 * problem.
1765 	 */
1766 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1767 	    (stpdown == stp) || (stpdown->sd_flag &
1768 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1769 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1770 	    (getmajor(stpdown->sd_vnode->v_rdev) >= devcnt)) ||
1771 	    linkcycle(stp, stpdown)) {
1772 		mutex_exit(&muxifier);
1773 		return (EINVAL);
1774 	}
1775 	TRACE_1(TR_FAC_STREAMS_FR,
1776 		TR_STPDOWN, "stpdown:%p", stpdown);
1777 	rq = getendq(stp->sd_wrq);
1778 	if (cmd == I_PLINK)
1779 		rq = NULL;
1780 
1781 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1782 
1783 	strioc.ic_cmd = cmd;
1784 	strioc.ic_timout = INFTIM;
1785 	strioc.ic_len = sizeof (struct linkblk);
1786 	strioc.ic_dp = (char *)&linkp->li_lblk;
1787 
1788 	/*
1789 	 * STRPLUMB protects plumbing changes and should be set before
1790 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1791 	 * and cleared in the end of mlink when passthru queue is removed.
1792 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1793 	 * queue is in-place (it is not a proper module and doesn't have open
1794 	 * entry point).
1795 	 *
1796 	 * STPLEX prevents any threads from entering the stream from above. It
1797 	 * can't be set before the call to link_addpassthru() because putnext
1798 	 * from below may cause stream head I/O routines to be called and these
1799 	 * routines assert that STPLEX is not set. After link_addpassthru()
1800 	 * nothing may come from below since the pass queue syncq is blocked.
1801 	 * Note also that STPLEX should be cleared before the call to
1802 	 * link_remmpassthru() since when messages start flowing to the stream
1803 	 * head (e.g. because of message propagation from the pass queue) stream
1804 	 * head I/O routines may be called with STPLEX flag set.
1805 	 *
1806 	 * When STPLEX is set, nothing may come into the stream from above and
1807 	 * it is safe to do a setq which will change stream head. So, the
1808 	 * correct sequence of actions is:
1809 	 *
1810 	 * 1) Set STRPLUMB
1811 	 * 2) Call link_addpassthru()
1812 	 * 3) Set STPLEX
1813 	 * 4) Call setq and update the stream state
1814 	 * 5) Clear STPLEX
1815 	 * 6) Call link_rempassthru()
1816 	 * 7) Clear STRPLUMB
1817 	 *
1818 	 * The same sequence applies to munlink() code.
1819 	 */
1820 	mutex_enter(&stpdown->sd_lock);
1821 	stpdown->sd_flag |= STRPLUMB;
1822 	mutex_exit(&stpdown->sd_lock);
1823 	/*
1824 	 * Add passthru queue below lower mux. This will block
1825 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1826 	 */
1827 	passq = link_addpassthru(stpdown);
1828 
1829 	mutex_enter(&stpdown->sd_lock);
1830 	stpdown->sd_flag |= STPLEX;
1831 	mutex_exit(&stpdown->sd_lock);
1832 
1833 	rq = _RD(stpdown->sd_wrq);
1834 	/*
1835 	 * There may be messages in the streamhead's syncq due to messages
1836 	 * that arrived before link_addpassthru() was done. To avoid
1837 	 * background processing of the syncq happening simultaneous with
1838 	 * setq processing, we disable the streamhead syncq and wait until
1839 	 * existing background thread finishes working on it.
1840 	 */
1841 	wait_sq_svc(rq->q_syncq);
1842 	passyncq = passq->q_syncq;
1843 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1844 		blocksq(passyncq, SQ_BLOCKED, 0);
1845 
1846 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1847 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1848 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1849 
1850 	/* setq might sleep in allocator - avoid holding locks. */
1851 	/* Note: we are holding muxifier here. */
1852 
1853 	str = stp->sd_strtab;
1854 	dp = &devimpl[getmajor(vp->v_rdev)];
1855 	ASSERT(dp->d_str == str);
1856 
1857 	qflag = dp->d_qflag;
1858 	sqtype = dp->d_sqtype;
1859 
1860 	/* create perdm_t if needed */
1861 	if (NEED_DM(dp->d_dmp, qflag))
1862 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1863 
1864 	dmp = dp->d_dmp;
1865 
1866 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1867 	    B_TRUE);
1868 
1869 	/*
1870 	 * XXX Remove any "odd" messages from the queue.
1871 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1872 	 */
1873 	error = strdoioctl(stp, &strioc, FNATIVE,
1874 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1875 	if (error != 0) {
1876 		lbfree(linkp);
1877 
1878 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1879 			blocksq(passyncq, SQ_BLOCKED, 0);
1880 		/*
1881 		 * Restore the stream head queue and then remove
1882 		 * the passq. Turn off STPLEX before we turn on
1883 		 * the stream by removing the passq.
1884 		 */
1885 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1886 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1887 		    B_TRUE);
1888 
1889 		mutex_enter(&stpdown->sd_lock);
1890 		stpdown->sd_flag &= ~STPLEX;
1891 		mutex_exit(&stpdown->sd_lock);
1892 
1893 		link_rempassthru(passq);
1894 
1895 		mutex_enter(&stpdown->sd_lock);
1896 		stpdown->sd_flag &= ~STRPLUMB;
1897 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1898 		cv_broadcast(&stpdown->sd_monitor);
1899 		mutex_exit(&stpdown->sd_lock);
1900 
1901 		mutex_exit(&muxifier);
1902 		return (error);
1903 	}
1904 	mutex_enter(&fpdown->f_tlock);
1905 	fpdown->f_count++;
1906 	mutex_exit(&fpdown->f_tlock);
1907 
1908 	/*
1909 	 * if we've made it here the linkage is all set up so we should also
1910 	 * set up the layered driver linkages
1911 	 */
1912 
1913 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1914 	if (cmd == I_LINK) {
1915 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1916 	} else {
1917 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1918 	}
1919 
1920 	link_rempassthru(passq);
1921 
1922 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index);
1923 
1924 	/*
1925 	 * Mark the upper stream as having dependent links
1926 	 * so that strclose can clean it up.
1927 	 */
1928 	if (cmd == I_LINK) {
1929 		mutex_enter(&stp->sd_lock);
1930 		stp->sd_flag |= STRHASLINKS;
1931 		mutex_exit(&stp->sd_lock);
1932 	}
1933 	/*
1934 	 * Wake up any other processes that may have been
1935 	 * waiting on the lower stream. These will all
1936 	 * error out.
1937 	 */
1938 	mutex_enter(&stpdown->sd_lock);
1939 	/* The passthru module is removed so we may release STRPLUMB */
1940 	stpdown->sd_flag &= ~STRPLUMB;
1941 	cv_broadcast(&rq->q_wait);
1942 	cv_broadcast(&_WR(rq)->q_wait);
1943 	cv_broadcast(&stpdown->sd_monitor);
1944 	mutex_exit(&stpdown->sd_lock);
1945 	mutex_exit(&muxifier);
1946 	*rvalp = linkp->li_lblk.l_index;
1947 	return (0);
1948 }
1949 
1950 int
1951 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1952 {
1953 	int		ret;
1954 	struct file	*fpdown;
1955 
1956 	fpdown = getf(arg);
1957 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1958 	if (fpdown != NULL)
1959 		releasef(arg);
1960 	return (ret);
1961 }
1962 
1963 /*
1964  * Unlink a multiplexor link. Stp is the controlling stream for the
1965  * link, and linkp points to the link's entry in the linkinfo list.
1966  * The muxifier lock must be held on entry and is dropped on exit.
1967  *
1968  * NOTE : Currently it is assumed that mux would process all the messages
1969  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
1970  * of the mux to handle all the messages that arrive before UNLINK.
1971  * If the mux has to send down messages on its lower stream before
1972  * ACKing I_UNLINK, then it *should* know to handle messages even
1973  * after the UNLINK is acked (actually it should be able to handle till we
1974  * re-block the read side of the pass queue here). If the mux does not
1975  * open up the lower stream, any messages that arrive during UNLINK
1976  * will be put in the stream head. In the case of lower stream opening
1977  * up, some messages might land in the stream head depending on when
1978  * the message arrived and when the read side of the pass queue was
1979  * re-blocked.
1980  */
1981 int
1982 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp)
1983 {
1984 	struct strioctl strioc;
1985 	struct stdata *stpdown;
1986 	queue_t *rq, *wrq;
1987 	queue_t	*passq;
1988 	syncq_t *passyncq;
1989 	int error = 0;
1990 	file_t *fpdown;
1991 
1992 	ASSERT(MUTEX_HELD(&muxifier));
1993 
1994 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
1995 
1996 	/*
1997 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
1998 	 */
1999 	mutex_enter(&stpdown->sd_lock);
2000 	stpdown->sd_flag |= STRPLUMB;
2001 	mutex_exit(&stpdown->sd_lock);
2002 
2003 	/*
2004 	 * Add passthru queue below lower mux. This will block
2005 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2006 	 */
2007 	passq = link_addpassthru(stpdown);
2008 
2009 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2010 		strioc.ic_cmd = I_UNLINK;
2011 	else
2012 		strioc.ic_cmd = I_PUNLINK;
2013 	strioc.ic_timout = INFTIM;
2014 	strioc.ic_len = sizeof (struct linkblk);
2015 	strioc.ic_dp = (char *)&linkp->li_lblk;
2016 
2017 	error = strdoioctl(stp, &strioc, FNATIVE,
2018 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2019 
2020 	/*
2021 	 * If there was an error and this is not called via strclose,
2022 	 * return to the user. Otherwise, pretend there was no error
2023 	 * and close the link.
2024 	 */
2025 	if (error) {
2026 		if (flag & LINKCLOSE) {
2027 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2028 			    "unlink ioctl, closing anyway (%d)\n", error);
2029 		} else {
2030 			link_rempassthru(passq);
2031 			mutex_enter(&stpdown->sd_lock);
2032 			stpdown->sd_flag &= ~STRPLUMB;
2033 			cv_broadcast(&stpdown->sd_monitor);
2034 			mutex_exit(&stpdown->sd_lock);
2035 			mutex_exit(&muxifier);
2036 			return (error);
2037 		}
2038 	}
2039 
2040 	mux_rmvedge(stp, linkp->li_lblk.l_index);
2041 	fpdown = linkp->li_fpdown;
2042 	lbfree(linkp);
2043 
2044 	/*
2045 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2046 	 * can slow others down. It's okay to since attempts to mlink() this
2047 	 * stream will be stopped because STPLEX is still set in the stdata
2048 	 * structure, and munlink() is stopped because mux_rmvedge() and
2049 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2050 	 * respectively.  Note that we defer the closef() of fpdown until
2051 	 * after we drop muxifier since strclose() can call munlinkall().
2052 	 */
2053 	mutex_exit(&muxifier);
2054 
2055 	wrq = stpdown->sd_wrq;
2056 	rq = _RD(wrq);
2057 
2058 	/*
2059 	 * Get rid of outstanding service procedure runs, before we make
2060 	 * it a stream head, since a stream head doesn't have any service
2061 	 * procedure.
2062 	 */
2063 	disable_svc(rq);
2064 	wait_svc(rq);
2065 
2066 	/*
2067 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2068 	 * is queued up to be finished. mux should take care that nothing is
2069 	 * send down to this queue. We should do it now as we're going to block
2070 	 * passyncq if it was unblocked.
2071 	 */
2072 	if (wrq->q_flag & QPERMOD) {
2073 		syncq_t	*sq = wrq->q_syncq;
2074 
2075 		mutex_enter(SQLOCK(sq));
2076 		while (wrq->q_sqflags & Q_SQQUEUED) {
2077 			sq->sq_flags |= SQ_WANTWAKEUP;
2078 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2079 		}
2080 		mutex_exit(SQLOCK(sq));
2081 	}
2082 	passyncq = passq->q_syncq;
2083 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2084 
2085 		syncq_t *sq, *outer;
2086 
2087 		/*
2088 		 * Messages could be flowing from underneath. We will
2089 		 * block the read side of the passq. This would be
2090 		 * sufficient for QPAIR and QPERQ muxes to ensure
2091 		 * that no data is flowing up into this queue
2092 		 * and hence no thread active in this instance of
2093 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2094 		 * could be messages on the inner and outer/inner
2095 		 * syncqs respectively. We will wait for them to drain.
2096 		 * Because passq is blocked messages end up in the syncq
2097 		 * And qfill_syncq could possibly end up setting QFULL
2098 		 * which will access the rq->q_flag. Hence, we have to
2099 		 * acquire the QLOCK in setq.
2100 		 *
2101 		 * XXX Messages can also flow from top into this
2102 		 * queue though the unlink is over (Ex. some instance
2103 		 * in putnext() called from top that has still not
2104 		 * accessed this queue. And also putq(lowerq) ?).
2105 		 * Solution : How about blocking the l_qtop queue ?
2106 		 * Do we really care about such pure D_MP muxes ?
2107 		 */
2108 
2109 		blocksq(passyncq, SQ_BLOCKED, 0);
2110 
2111 		sq = rq->q_syncq;
2112 		if ((outer = sq->sq_outer) != NULL) {
2113 
2114 			/*
2115 			 * We have to just wait for the outer sq_count
2116 			 * drop to zero. As this does not prevent new
2117 			 * messages to enter the outer perimeter, this
2118 			 * is subject to starvation.
2119 			 *
2120 			 * NOTE :Because of blocksq above, messages could
2121 			 * be in the inner syncq only because of some
2122 			 * thread holding the outer perimeter exclusively.
2123 			 * Hence it would be sufficient to wait for the
2124 			 * exclusive holder of the outer perimeter to drain
2125 			 * the inner and outer syncqs. But we will not depend
2126 			 * on this feature and hence check the inner syncqs
2127 			 * separately.
2128 			 */
2129 			wait_syncq(outer);
2130 		}
2131 
2132 
2133 		/*
2134 		 * There could be messages destined for
2135 		 * this queue. Let the exclusive holder
2136 		 * drain it.
2137 		 */
2138 
2139 		wait_syncq(sq);
2140 		ASSERT((rq->q_flag & QPERMOD) ||
2141 			((rq->q_syncq->sq_head == NULL) &&
2142 			(_WR(rq)->q_syncq->sq_head == NULL)));
2143 	}
2144 
2145 	/*
2146 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2147 	 * case as we don't disable its syncq or remove it off the syncq
2148 	 * service list.
2149 	 */
2150 	if (rq->q_flag & QPERMOD) {
2151 		syncq_t	*sq = rq->q_syncq;
2152 
2153 		mutex_enter(SQLOCK(sq));
2154 		while (rq->q_sqflags & Q_SQQUEUED) {
2155 			sq->sq_flags |= SQ_WANTWAKEUP;
2156 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2157 		}
2158 		mutex_exit(SQLOCK(sq));
2159 	}
2160 
2161 	/*
2162 	 * flush_syncq changes states only when there is some messages to
2163 	 * free. ie when it returns non-zero value to return.
2164 	 */
2165 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2166 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2167 
2168 	/*
2169 	 * No body else should know about this queue now.
2170 	 * If the mux did not process the messages before
2171 	 * acking the I_UNLINK, free them now.
2172 	 */
2173 
2174 	flushq(rq, FLUSHALL);
2175 	flushq(_WR(rq), FLUSHALL);
2176 
2177 	/*
2178 	 * Convert the mux lower queue into a stream head queue.
2179 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2180 	 */
2181 	rq->q_ptr = wrq->q_ptr = stpdown;
2182 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2183 
2184 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2185 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2186 
2187 	enable_svc(rq);
2188 
2189 	/*
2190 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2191 	 * needs to be set to prevent reopen() of the stream - such reopen may
2192 	 * try to call non-existent pass queue open routine and panic.
2193 	 */
2194 	mutex_enter(&stpdown->sd_lock);
2195 	stpdown->sd_flag &= ~STPLEX;
2196 	mutex_exit(&stpdown->sd_lock);
2197 
2198 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2199 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2200 
2201 	/* clean up the layered driver linkages */
2202 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2203 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2204 	} else {
2205 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2206 	}
2207 
2208 	link_rempassthru(passq);
2209 
2210 	/*
2211 	 * Now all plumbing changes are finished and STRPLUMB is no
2212 	 * longer needed.
2213 	 */
2214 	mutex_enter(&stpdown->sd_lock);
2215 	stpdown->sd_flag &= ~STRPLUMB;
2216 	cv_broadcast(&stpdown->sd_monitor);
2217 	mutex_exit(&stpdown->sd_lock);
2218 
2219 	(void) closef(fpdown);
2220 	return (0);
2221 }
2222 
2223 /*
2224  * Unlink all multiplexor links for which stp is the controlling stream.
2225  * Return 0, or a non-zero errno on failure.
2226  */
2227 int
2228 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp)
2229 {
2230 	linkinfo_t *linkp;
2231 	int error = 0;
2232 
2233 	mutex_enter(&muxifier);
2234 	while (linkp = findlinks(stp, 0, flag)) {
2235 		/*
2236 		 * munlink() releases the muxifier lock.
2237 		 */
2238 		if (error = munlink(stp, linkp, flag, crp, rvalp))
2239 			return (error);
2240 		mutex_enter(&muxifier);
2241 	}
2242 	mutex_exit(&muxifier);
2243 	return (0);
2244 }
2245 
2246 /*
2247  * A multiplexor link has been made. Add an
2248  * edge to the directed graph.
2249  */
2250 void
2251 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid)
2252 {
2253 	struct mux_node *np;
2254 	struct mux_edge *ep;
2255 	major_t upmaj;
2256 	major_t lomaj;
2257 
2258 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2259 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2260 	np = &mux_nodes[upmaj];
2261 	if (np->mn_outp) {
2262 		ep = np->mn_outp;
2263 		while (ep->me_nextp)
2264 			ep = ep->me_nextp;
2265 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2266 		ep = ep->me_nextp;
2267 	} else {
2268 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2269 		ep = np->mn_outp;
2270 	}
2271 	ep->me_nextp = NULL;
2272 	ep->me_muxid = muxid;
2273 	if (lostp->sd_vnode->v_type == VFIFO)
2274 		ep->me_nodep = NULL;
2275 	else
2276 		ep->me_nodep = &mux_nodes[lomaj];
2277 }
2278 
2279 /*
2280  * A multiplexor link has been removed. Remove the
2281  * edge in the directed graph.
2282  */
2283 void
2284 mux_rmvedge(stdata_t *upstp, int muxid)
2285 {
2286 	struct mux_node *np;
2287 	struct mux_edge *ep;
2288 	struct mux_edge *pep = NULL;
2289 	major_t upmaj;
2290 
2291 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2292 	np = &mux_nodes[upmaj];
2293 	ASSERT(np->mn_outp != NULL);
2294 	ep = np->mn_outp;
2295 	while (ep) {
2296 		if (ep->me_muxid == muxid) {
2297 			if (pep)
2298 				pep->me_nextp = ep->me_nextp;
2299 			else
2300 				np->mn_outp = ep->me_nextp;
2301 			kmem_free(ep, sizeof (struct mux_edge));
2302 			return;
2303 		}
2304 		pep = ep;
2305 		ep = ep->me_nextp;
2306 	}
2307 	ASSERT(0);	/* should not reach here */
2308 }
2309 
2310 /*
2311  * Translate the device flags (from conf.h) to the corresponding
2312  * qflag and sq_flag (type) values.
2313  */
2314 int
2315 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2316 	uint32_t *sqtypep)
2317 {
2318 	uint32_t qflag = 0;
2319 	uint32_t sqtype = 0;
2320 
2321 	if (devflag & _D_OLD)
2322 		goto bad;
2323 
2324 	/* Inner perimeter presence and scope */
2325 	switch (devflag & D_MTINNER_MASK) {
2326 	case D_MP:
2327 		qflag |= QMTSAFE;
2328 		sqtype |= SQ_CI;
2329 		break;
2330 	case D_MTPERQ|D_MP:
2331 		qflag |= QPERQ;
2332 		break;
2333 	case D_MTQPAIR|D_MP:
2334 		qflag |= QPAIR;
2335 		break;
2336 	case D_MTPERMOD|D_MP:
2337 		qflag |= QPERMOD;
2338 		break;
2339 	default:
2340 		goto bad;
2341 	}
2342 
2343 	/* Outer perimeter */
2344 	if (devflag & D_MTOUTPERIM) {
2345 		switch (devflag & D_MTINNER_MASK) {
2346 		case D_MP:
2347 		case D_MTPERQ|D_MP:
2348 		case D_MTQPAIR|D_MP:
2349 			break;
2350 		default:
2351 			goto bad;
2352 		}
2353 		qflag |= QMTOUTPERIM;
2354 	}
2355 
2356 	/* Inner perimeter modifiers */
2357 	if (devflag & D_MTINNER_MOD) {
2358 		switch (devflag & D_MTINNER_MASK) {
2359 		case D_MP:
2360 			goto bad;
2361 		default:
2362 			break;
2363 		}
2364 		if (devflag & D_MTPUTSHARED)
2365 			sqtype |= SQ_CIPUT;
2366 		if (devflag & _D_MTOCSHARED) {
2367 			/*
2368 			 * The code in putnext assumes that it has the
2369 			 * highest concurrency by not checking sq_count.
2370 			 * Thus _D_MTOCSHARED can only be supported when
2371 			 * D_MTPUTSHARED is set.
2372 			 */
2373 			if (!(devflag & D_MTPUTSHARED))
2374 				goto bad;
2375 			sqtype |= SQ_CIOC;
2376 		}
2377 		if (devflag & _D_MTCBSHARED) {
2378 			/*
2379 			 * The code in putnext assumes that it has the
2380 			 * highest concurrency by not checking sq_count.
2381 			 * Thus _D_MTCBSHARED can only be supported when
2382 			 * D_MTPUTSHARED is set.
2383 			 */
2384 			if (!(devflag & D_MTPUTSHARED))
2385 				goto bad;
2386 			sqtype |= SQ_CICB;
2387 		}
2388 		if (devflag & _D_MTSVCSHARED) {
2389 			/*
2390 			 * The code in putnext assumes that it has the
2391 			 * highest concurrency by not checking sq_count.
2392 			 * Thus _D_MTSVCSHARED can only be supported when
2393 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2394 			 * supported only for QPERMOD.
2395 			 */
2396 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2397 				goto bad;
2398 			sqtype |= SQ_CISVC;
2399 		}
2400 	}
2401 
2402 	/* Default outer perimeter concurrency */
2403 	sqtype |= SQ_CO;
2404 
2405 	/* Outer perimeter modifiers */
2406 	if (devflag & D_MTOCEXCL) {
2407 		if (!(devflag & D_MTOUTPERIM)) {
2408 			/* No outer perimeter */
2409 			goto bad;
2410 		}
2411 		sqtype &= ~SQ_COOC;
2412 	}
2413 
2414 	/* Synchronous Streams extended qinit structure */
2415 	if (devflag & D_SYNCSTR)
2416 		qflag |= QSYNCSTR;
2417 
2418 	/*
2419 	 * Private flag used by a transport module to indicate
2420 	 * to sockfs that it supports direct-access mode without
2421 	 * having to go through STREAMS.
2422 	 */
2423 	if (devflag & _D_DIRECT) {
2424 		/* Reject unless the module is fully-MT (no perimeter) */
2425 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2426 			goto bad;
2427 		qflag |= _QDIRECT;
2428 	}
2429 
2430 	*qflagp = qflag;
2431 	*sqtypep = sqtype;
2432 	return (0);
2433 
2434 bad:
2435 	cmn_err(CE_WARN,
2436 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2437 	    (int)(qflag & D_MTSAFETY_MASK),
2438 	    stp->st_rdinit->qi_minfo->mi_idname);
2439 
2440 	return (EINVAL);
2441 }
2442 
2443 /*
2444  * Set the interface values for a pair of queues (qinit structure,
2445  * packet sizes, water marks).
2446  * setq assumes that the caller does not have a claim (entersq or claimq)
2447  * on the queue.
2448  */
2449 void
2450 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2451     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2452 {
2453 	queue_t *wq;
2454 	syncq_t	*sq, *outer;
2455 
2456 	ASSERT(rq->q_flag & QREADR);
2457 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2458 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2459 
2460 	wq = _WR(rq);
2461 	rq->q_qinfo = rinit;
2462 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2463 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2464 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2465 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2466 	wq->q_qinfo = winit;
2467 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2468 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2469 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2470 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2471 
2472 	/* Remove old syncqs */
2473 	sq = rq->q_syncq;
2474 	outer = sq->sq_outer;
2475 	if (outer != NULL) {
2476 		ASSERT(wq->q_syncq->sq_outer == outer);
2477 		outer_remove(outer, rq->q_syncq);
2478 		if (wq->q_syncq != rq->q_syncq)
2479 			outer_remove(outer, wq->q_syncq);
2480 	}
2481 	ASSERT(sq->sq_outer == NULL);
2482 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2483 
2484 	if (sq != SQ(rq)) {
2485 		if (!(rq->q_flag & QPERMOD))
2486 			free_syncq(sq);
2487 		if (wq->q_syncq == rq->q_syncq)
2488 			wq->q_syncq = NULL;
2489 		rq->q_syncq = NULL;
2490 	}
2491 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2492 	    wq->q_syncq != SQ(rq)) {
2493 		free_syncq(wq->q_syncq);
2494 		wq->q_syncq = NULL;
2495 	}
2496 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2497 				rq->q_syncq->sq_tail == NULL));
2498 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2499 				wq->q_syncq->sq_tail == NULL));
2500 
2501 	if (!(rq->q_flag & QPERMOD) &&
2502 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2503 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2504 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2505 		    rq->q_syncq->sq_nciputctrl, 0);
2506 		ASSERT(ciputctrl_cache != NULL);
2507 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2508 		rq->q_syncq->sq_ciputctrl = NULL;
2509 		rq->q_syncq->sq_nciputctrl = 0;
2510 	}
2511 
2512 	if (!(wq->q_flag & QPERMOD) &&
2513 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2514 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2515 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2516 		    wq->q_syncq->sq_nciputctrl, 0);
2517 		ASSERT(ciputctrl_cache != NULL);
2518 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2519 		wq->q_syncq->sq_ciputctrl = NULL;
2520 		wq->q_syncq->sq_nciputctrl = 0;
2521 	}
2522 
2523 	sq = SQ(rq);
2524 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2525 	ASSERT(sq->sq_outer == NULL);
2526 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2527 
2528 	/*
2529 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2530 	 * bits in sq_flag based on the sqtype.
2531 	 */
2532 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2533 
2534 	rq->q_syncq = wq->q_syncq = sq;
2535 	sq->sq_type = sqtype;
2536 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2537 
2538 	/*
2539 	 *  We are making sq_svcflags zero,
2540 	 *  resetting SQ_DISABLED in case it was set by
2541 	 *  wait_svc() in the munlink path.
2542 	 *
2543 	 */
2544 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2545 	sq->sq_svcflags = 0;
2546 
2547 	/*
2548 	 * We need to acquire the lock here for the mlink and munlink case,
2549 	 * where canputnext, backenable, etc can access the q_flag.
2550 	 */
2551 	if (lock_needed) {
2552 		mutex_enter(QLOCK(rq));
2553 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2554 		mutex_exit(QLOCK(rq));
2555 		mutex_enter(QLOCK(wq));
2556 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2557 		mutex_exit(QLOCK(wq));
2558 	} else {
2559 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2560 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2561 	}
2562 
2563 	if (qflag & QPERQ) {
2564 		/* Allocate a separate syncq for the write side */
2565 		sq = new_syncq();
2566 		sq->sq_type = rq->q_syncq->sq_type;
2567 		sq->sq_flags = rq->q_syncq->sq_flags;
2568 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2569 		    sq->sq_oprev == NULL);
2570 		wq->q_syncq = sq;
2571 	}
2572 	if (qflag & QPERMOD) {
2573 		sq = dmp->dm_sq;
2574 
2575 		/*
2576 		 * Assert that we do have an inner perimeter syncq and that it
2577 		 * does not have an outer perimeter associated with it.
2578 		 */
2579 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2580 		    sq->sq_oprev == NULL);
2581 		rq->q_syncq = wq->q_syncq = sq;
2582 	}
2583 	if (qflag & QMTOUTPERIM) {
2584 		outer = dmp->dm_sq;
2585 
2586 		ASSERT(outer->sq_outer == NULL);
2587 		outer_insert(outer, rq->q_syncq);
2588 		if (wq->q_syncq != rq->q_syncq)
2589 			outer_insert(outer, wq->q_syncq);
2590 	}
2591 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2592 		(rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2593 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2594 		(wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2595 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2596 
2597 	/*
2598 	 * Initialize struio() types.
2599 	 */
2600 	rq->q_struiot =
2601 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2602 	wq->q_struiot =
2603 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2604 }
2605 
2606 perdm_t *
2607 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2608 {
2609 	syncq_t	*sq;
2610 	perdm_t	**pp;
2611 	perdm_t	*p;
2612 	perdm_t	*dmp;
2613 
2614 	ASSERT(str != NULL);
2615 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2616 
2617 	rw_enter(&perdm_rwlock, RW_READER);
2618 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2619 		if (p->dm_str == str) {	/* found one */
2620 			atomic_add_32(&(p->dm_ref), 1);
2621 			rw_exit(&perdm_rwlock);
2622 			return (p);
2623 		}
2624 	}
2625 	rw_exit(&perdm_rwlock);
2626 
2627 	sq = new_syncq();
2628 	if (qflag & QPERMOD) {
2629 		sq->sq_type = sqtype | SQ_PERMOD;
2630 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2631 	} else {
2632 		ASSERT(qflag & QMTOUTPERIM);
2633 		sq->sq_onext = sq->sq_oprev = sq;
2634 	}
2635 
2636 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2637 	dmp->dm_sq = sq;
2638 	dmp->dm_str = str;
2639 	dmp->dm_ref = 1;
2640 	dmp->dm_next = NULL;
2641 
2642 	rw_enter(&perdm_rwlock, RW_WRITER);
2643 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2644 		if (p->dm_str == str) {	/* already present */
2645 			p->dm_ref++;
2646 			rw_exit(&perdm_rwlock);
2647 			free_syncq(sq);
2648 			kmem_free(dmp, sizeof (perdm_t));
2649 			return (p);
2650 		}
2651 	}
2652 
2653 	*pp = dmp;
2654 	rw_exit(&perdm_rwlock);
2655 	return (dmp);
2656 }
2657 
2658 void
2659 rele_dm(perdm_t *dmp)
2660 {
2661 	perdm_t **pp;
2662 	perdm_t *p;
2663 
2664 	rw_enter(&perdm_rwlock, RW_WRITER);
2665 	ASSERT(dmp->dm_ref > 0);
2666 
2667 	if (--dmp->dm_ref > 0) {
2668 		rw_exit(&perdm_rwlock);
2669 		return;
2670 	}
2671 
2672 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2673 		if (p == dmp)
2674 			break;
2675 	ASSERT(p == dmp);
2676 	*pp = p->dm_next;
2677 	rw_exit(&perdm_rwlock);
2678 
2679 	/*
2680 	 * Wait for any background processing that relies on the
2681 	 * syncq to complete before it is freed.
2682 	 */
2683 	wait_sq_svc(p->dm_sq);
2684 	free_syncq(p->dm_sq);
2685 	kmem_free(p, sizeof (perdm_t));
2686 }
2687 
2688 /*
2689  * Make a protocol message given control and data buffers.
2690  * n.b., this can block; be careful of what locks you hold when calling it.
2691  *
2692  * If sd_maxblk is less than *iosize this routine can fail part way through
2693  * (due to an allocation failure). In this case on return *iosize will contain
2694  * the amount that was consumed. Otherwise *iosize will not be modified
2695  * i.e. it will contain the amount that was consumed.
2696  */
2697 int
2698 strmakemsg(
2699 	struct strbuf *mctl,
2700 	ssize_t *iosize,
2701 	struct uio *uiop,
2702 	stdata_t *stp,
2703 	int32_t flag,
2704 	mblk_t **mpp)
2705 {
2706 	mblk_t *mpctl = NULL;
2707 	mblk_t *mpdata = NULL;
2708 	int error;
2709 
2710 	ASSERT(uiop != NULL);
2711 
2712 	*mpp = NULL;
2713 	/* Create control part, if any */
2714 	if ((mctl != NULL) && (mctl->len >= 0)) {
2715 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2716 		if (error)
2717 			return (error);
2718 	}
2719 	/* Create data part, if any */
2720 	if (*iosize >= 0) {
2721 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2722 		if (error) {
2723 			freemsg(mpctl);
2724 			return (error);
2725 		}
2726 	}
2727 	if (mpctl != NULL) {
2728 		if (mpdata != NULL)
2729 			linkb(mpctl, mpdata);
2730 		*mpp = mpctl;
2731 	} else {
2732 		*mpp = mpdata;
2733 	}
2734 	return (0);
2735 }
2736 
2737 /*
2738  * Make the control part of a protocol message given a control buffer.
2739  * n.b., this can block; be careful of what locks you hold when calling it.
2740  */
2741 int
2742 strmakectl(
2743 	struct strbuf *mctl,
2744 	int32_t flag,
2745 	int32_t fflag,
2746 	mblk_t **mpp)
2747 {
2748 	mblk_t *bp = NULL;
2749 	unsigned char msgtype;
2750 	int error = 0;
2751 
2752 	*mpp = NULL;
2753 	/*
2754 	 * Create control part of message, if any.
2755 	 */
2756 	if ((mctl != NULL) && (mctl->len >= 0)) {
2757 		caddr_t base;
2758 		int ctlcount;
2759 		int allocsz;
2760 
2761 		if (flag & RS_HIPRI)
2762 			msgtype = M_PCPROTO;
2763 		else
2764 			msgtype = M_PROTO;
2765 
2766 		ctlcount = mctl->len;
2767 		base = mctl->buf;
2768 
2769 		/*
2770 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2771 		 * blocks by increasing the size to something more usable.
2772 		 */
2773 		allocsz = MAX(ctlcount, 64);
2774 
2775 		/*
2776 		 * Range checking has already been done; simply try
2777 		 * to allocate a message block for the ctl part.
2778 		 */
2779 		while (!(bp = allocb(allocsz, BPRI_MED))) {
2780 			if (fflag & (FNDELAY|FNONBLOCK))
2781 				return (EAGAIN);
2782 			if (error = strwaitbuf(allocsz, BPRI_MED))
2783 				return (error);
2784 		}
2785 
2786 		bp->b_datap->db_type = msgtype;
2787 		if (copyin(base, bp->b_wptr, ctlcount)) {
2788 			freeb(bp);
2789 			return (EFAULT);
2790 		}
2791 		bp->b_wptr += ctlcount;
2792 	}
2793 	*mpp = bp;
2794 	return (0);
2795 }
2796 
2797 /*
2798  * Make a protocol message given data buffers.
2799  * n.b., this can block; be careful of what locks you hold when calling it.
2800  *
2801  * If sd_maxblk is less than *iosize this routine can fail part way through
2802  * (due to an allocation failure). In this case on return *iosize will contain
2803  * the amount that was consumed. Otherwise *iosize will not be modified
2804  * i.e. it will contain the amount that was consumed.
2805  */
2806 int
2807 strmakedata(
2808 	ssize_t   *iosize,
2809 	struct uio *uiop,
2810 	stdata_t *stp,
2811 	int32_t flag,
2812 	mblk_t **mpp)
2813 {
2814 	mblk_t *mp = NULL;
2815 	mblk_t *bp;
2816 	int wroff = (int)stp->sd_wroff;
2817 	int tail_len = (int)stp->sd_tail;
2818 	int extra = wroff + tail_len;
2819 	int error = 0;
2820 	ssize_t maxblk;
2821 	ssize_t count = *iosize;
2822 	cred_t *cr = CRED();
2823 
2824 	*mpp = NULL;
2825 	if (count < 0)
2826 		return (0);
2827 
2828 	maxblk = stp->sd_maxblk;
2829 	if (maxblk == INFPSZ)
2830 		maxblk = count;
2831 
2832 	/*
2833 	 * Create data part of message, if any.
2834 	 */
2835 	do {
2836 		ssize_t size;
2837 		dblk_t  *dp;
2838 
2839 		ASSERT(uiop);
2840 
2841 		size = MIN(count, maxblk);
2842 
2843 		while ((bp = allocb_cred(size + extra, cr)) == NULL) {
2844 			error = EAGAIN;
2845 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2846 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2847 				if (count == *iosize) {
2848 					freemsg(mp);
2849 					return (error);
2850 				} else {
2851 					*iosize -= count;
2852 					*mpp = mp;
2853 					return (0);
2854 				}
2855 			}
2856 		}
2857 		dp = bp->b_datap;
2858 		dp->db_cpid = curproc->p_pid;
2859 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2860 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2861 
2862 		if (flag & STRUIO_POSTPONE) {
2863 			/*
2864 			 * Setup the stream uio portion of the
2865 			 * dblk for subsequent use by struioget().
2866 			 */
2867 			dp->db_struioflag = STRUIO_SPEC;
2868 			dp->db_cksumstart = 0;
2869 			dp->db_cksumstuff = 0;
2870 			dp->db_cksumend = size;
2871 			*(long long *)dp->db_struioun.data = 0ll;
2872 			bp->b_wptr += size;
2873 		} else {
2874 			if (stp->sd_copyflag & STRCOPYCACHED)
2875 				uiop->uio_extflg |= UIO_COPY_CACHED;
2876 
2877 			if (size != 0) {
2878 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2879 				    uiop);
2880 				if (error != 0) {
2881 					freeb(bp);
2882 					freemsg(mp);
2883 					return (error);
2884 				}
2885 			}
2886 			bp->b_wptr += size;
2887 
2888 			if (stp->sd_wputdatafunc != NULL) {
2889 				mblk_t *newbp;
2890 
2891 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2892 				    bp, NULL, NULL, NULL, NULL);
2893 				if (newbp == NULL) {
2894 					freeb(bp);
2895 					freemsg(mp);
2896 					return (ECOMM);
2897 				}
2898 				bp = newbp;
2899 			}
2900 		}
2901 
2902 		count -= size;
2903 
2904 		if (mp == NULL)
2905 			mp = bp;
2906 		else
2907 			linkb(mp, bp);
2908 	} while (count > 0);
2909 
2910 	*mpp = mp;
2911 	return (0);
2912 }
2913 
2914 /*
2915  * Wait for a buffer to become available. Return non-zero errno
2916  * if not able to wait, 0 if buffer is probably there.
2917  */
2918 int
2919 strwaitbuf(size_t size, int pri)
2920 {
2921 	bufcall_id_t id;
2922 
2923 	mutex_enter(&bcall_monitor);
2924 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2925 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2926 		mutex_exit(&bcall_monitor);
2927 		return (ENOSR);
2928 	}
2929 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2930 		unbufcall(id);
2931 		mutex_exit(&bcall_monitor);
2932 		return (EINTR);
2933 	}
2934 	unbufcall(id);
2935 	mutex_exit(&bcall_monitor);
2936 	return (0);
2937 }
2938 
2939 /*
2940  * This function waits for a read or write event to happen on a stream.
2941  * fmode can specify FNDELAY and/or FNONBLOCK.
2942  * The timeout is in ms with -1 meaning infinite.
2943  * The flag values work as follows:
2944  *	READWAIT	Check for read side errors, send M_READ
2945  *	GETWAIT		Check for read side errors, no M_READ
2946  *	WRITEWAIT	Check for write side errors.
2947  *	NOINTR		Do not return error if nonblocking or timeout.
2948  * 	STR_NOERROR	Ignore all errors except STPLEX.
2949  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
2950  *	STR_PEEK	Pass through the strgeterr().
2951  */
2952 int
2953 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
2954     int *done)
2955 {
2956 	int slpflg, errs;
2957 	int error;
2958 	kcondvar_t *sleepon;
2959 	mblk_t *mp;
2960 	ssize_t *rd_count;
2961 	clock_t rval;
2962 
2963 	ASSERT(MUTEX_HELD(&stp->sd_lock));
2964 	if ((flag & READWAIT) || (flag & GETWAIT)) {
2965 		slpflg = RSLEEP;
2966 		sleepon = &_RD(stp->sd_wrq)->q_wait;
2967 		errs = STRDERR|STPLEX;
2968 	} else {
2969 		slpflg = WSLEEP;
2970 		sleepon = &stp->sd_wrq->q_wait;
2971 		errs = STWRERR|STRHUP|STPLEX;
2972 	}
2973 	if (flag & STR_NOERROR)
2974 		errs = STPLEX;
2975 
2976 	if (stp->sd_wakeq & slpflg) {
2977 		/*
2978 		 * A strwakeq() is pending, no need to sleep.
2979 		 */
2980 		stp->sd_wakeq &= ~slpflg;
2981 		*done = 0;
2982 		return (0);
2983 	}
2984 
2985 	if (fmode & (FNDELAY|FNONBLOCK)) {
2986 		if (!(flag & NOINTR))
2987 			error = EAGAIN;
2988 		else
2989 			error = 0;
2990 		*done = 1;
2991 		return (error);
2992 	}
2993 
2994 	if (stp->sd_flag & errs) {
2995 		/*
2996 		 * Check for errors before going to sleep since the
2997 		 * caller might not have checked this while holding
2998 		 * sd_lock.
2999 		 */
3000 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3001 		if (error != 0) {
3002 			*done = 1;
3003 			return (error);
3004 		}
3005 	}
3006 
3007 	/*
3008 	 * If any module downstream has requested read notification
3009 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3010 	 * down stream.
3011 	 */
3012 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3013 		mutex_exit(&stp->sd_lock);
3014 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3015 		    (flag & STR_NOSIG), &error))) {
3016 			mutex_enter(&stp->sd_lock);
3017 			*done = 1;
3018 			return (error);
3019 		}
3020 		mp->b_datap->db_type = M_READ;
3021 		rd_count = (ssize_t *)mp->b_wptr;
3022 		*rd_count = count;
3023 		mp->b_wptr += sizeof (ssize_t);
3024 		/*
3025 		 * Send the number of bytes requested by the
3026 		 * read as the argument to M_READ.
3027 		 */
3028 		stream_willservice(stp);
3029 		putnext(stp->sd_wrq, mp);
3030 		stream_runservice(stp);
3031 		mutex_enter(&stp->sd_lock);
3032 
3033 		/*
3034 		 * If any data arrived due to inline processing
3035 		 * of putnext(), don't sleep.
3036 		 */
3037 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3038 			*done = 0;
3039 			return (0);
3040 		}
3041 	}
3042 
3043 	stp->sd_flag |= slpflg;
3044 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3045 		"strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3046 		stp, flag, count, fmode, done);
3047 
3048 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3049 	if (rval > 0) {
3050 		/* EMPTY */
3051 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3052 			"strwaitq awakes(2):%X, %X, %X, %X, %X",
3053 			stp, flag, count, fmode, done);
3054 	} else if (rval == 0) {
3055 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3056 			"strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3057 			stp, flag, count, fmode, done);
3058 		stp->sd_flag &= ~slpflg;
3059 		cv_broadcast(sleepon);
3060 		if (!(flag & NOINTR))
3061 			error = EINTR;
3062 		else
3063 			error = 0;
3064 		*done = 1;
3065 		return (error);
3066 	} else {
3067 		/* timeout */
3068 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3069 			"strwaitq timeout:%p, %X, %lX, %X, %p",
3070 			stp, flag, count, fmode, done);
3071 		*done = 1;
3072 		if (!(flag & NOINTR))
3073 			return (ETIME);
3074 		else
3075 			return (0);
3076 	}
3077 	/*
3078 	 * If the caller implements delayed errors (i.e. queued after data)
3079 	 * we can not check for errors here since data as well as an
3080 	 * error might have arrived at the stream head. We return to
3081 	 * have the caller check the read queue before checking for errors.
3082 	 */
3083 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3084 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3085 		if (error != 0) {
3086 			*done = 1;
3087 			return (error);
3088 		}
3089 	}
3090 	*done = 0;
3091 	return (0);
3092 }
3093 
3094 /*
3095  * Perform job control discipline access checks.
3096  * Return 0 for success and the errno for failure.
3097  */
3098 
3099 #define	cantsend(p, t, sig) \
3100 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3101 
3102 int
3103 straccess(struct stdata *stp, enum jcaccess mode)
3104 {
3105 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3106 	kthread_t *t = curthread;
3107 	proc_t *p = ttoproc(t);
3108 	sess_t *sp;
3109 
3110 	ASSERT(mutex_owned(&stp->sd_lock));
3111 
3112 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3113 		return (0);
3114 
3115 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3116 
3117 	for (;;) {
3118 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3119 		sp = p->p_sessp;
3120 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3121 
3122 		/*
3123 		 * If this is not the calling process's controlling terminal
3124 		 * or if the calling process is already in the foreground
3125 		 * then allow access.
3126 		 */
3127 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3128 		    p->p_pgidp == stp->sd_pgidp) {
3129 			mutex_exit(&sp->s_lock);
3130 			mutex_exit(&p->p_splock);
3131 			mutex_exit(&p->p_lock);
3132 			return (0);
3133 		}
3134 
3135 		/*
3136 		 * Check to see if controlling terminal has been deallocated.
3137 		 */
3138 		if (sp->s_vp == NULL) {
3139 			if (!cantsend(p, t, SIGHUP))
3140 				sigtoproc(p, t, SIGHUP);
3141 			mutex_exit(&sp->s_lock);
3142 			mutex_exit(&p->p_splock);
3143 			mutex_exit(&p->p_lock);
3144 			return (EIO);
3145 		}
3146 
3147 		mutex_exit(&sp->s_lock);
3148 		mutex_exit(&p->p_splock);
3149 
3150 		if (mode == JCGETP) {
3151 			mutex_exit(&p->p_lock);
3152 			return (0);
3153 		}
3154 
3155 		if (mode == JCREAD) {
3156 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3157 				mutex_exit(&p->p_lock);
3158 				return (EIO);
3159 			}
3160 			mutex_exit(&p->p_lock);
3161 			mutex_exit(&stp->sd_lock);
3162 			pgsignal(p->p_pgidp, SIGTTIN);
3163 			mutex_enter(&stp->sd_lock);
3164 			mutex_enter(&p->p_lock);
3165 		} else {  /* mode == JCWRITE or JCSETP */
3166 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3167 			    cantsend(p, t, SIGTTOU)) {
3168 				mutex_exit(&p->p_lock);
3169 				return (0);
3170 			}
3171 			if (p->p_detached) {
3172 				mutex_exit(&p->p_lock);
3173 				return (EIO);
3174 			}
3175 			mutex_exit(&p->p_lock);
3176 			mutex_exit(&stp->sd_lock);
3177 			pgsignal(p->p_pgidp, SIGTTOU);
3178 			mutex_enter(&stp->sd_lock);
3179 			mutex_enter(&p->p_lock);
3180 		}
3181 
3182 		/*
3183 		 * We call cv_wait_sig_swap() to cause the appropriate
3184 		 * action for the jobcontrol signal to take place.
3185 		 * If the signal is being caught, we will take the
3186 		 * EINTR error return.  Otherwise, the default action
3187 		 * of causing the process to stop will take place.
3188 		 * In this case, we rely on the periodic cv_broadcast() on
3189 		 * &lbolt_cv to wake us up to loop around and test again.
3190 		 * We can't get here if the signal is ignored or
3191 		 * if the current thread is blocking the signal.
3192 		 */
3193 		mutex_exit(&stp->sd_lock);
3194 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3195 			mutex_exit(&p->p_lock);
3196 			mutex_enter(&stp->sd_lock);
3197 			return (EINTR);
3198 		}
3199 		mutex_exit(&p->p_lock);
3200 		mutex_enter(&stp->sd_lock);
3201 		mutex_enter(&p->p_lock);
3202 	}
3203 }
3204 
3205 /*
3206  * Return size of message of block type (bp->b_datap->db_type)
3207  */
3208 size_t
3209 xmsgsize(mblk_t *bp)
3210 {
3211 	unsigned char type;
3212 	size_t count = 0;
3213 
3214 	type = bp->b_datap->db_type;
3215 
3216 	for (; bp; bp = bp->b_cont) {
3217 		if (type != bp->b_datap->db_type)
3218 			break;
3219 		ASSERT(bp->b_wptr >= bp->b_rptr);
3220 		count += bp->b_wptr - bp->b_rptr;
3221 	}
3222 	return (count);
3223 }
3224 
3225 /*
3226  * Allocate a stream head.
3227  */
3228 struct stdata *
3229 shalloc(queue_t *qp)
3230 {
3231 	stdata_t *stp;
3232 
3233 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3234 
3235 	stp->sd_wrq = _WR(qp);
3236 	stp->sd_strtab = NULL;
3237 	stp->sd_iocid = 0;
3238 	stp->sd_mate = NULL;
3239 	stp->sd_freezer = NULL;
3240 	stp->sd_refcnt = 0;
3241 	stp->sd_wakeq = 0;
3242 	stp->sd_anchor = 0;
3243 	stp->sd_struiowrq = NULL;
3244 	stp->sd_struiordq = NULL;
3245 	stp->sd_struiodnak = 0;
3246 	stp->sd_struionak = NULL;
3247 #ifdef C2_AUDIT
3248 	stp->sd_t_audit_data = NULL;
3249 #endif
3250 	stp->sd_rput_opt = 0;
3251 	stp->sd_wput_opt = 0;
3252 	stp->sd_read_opt = 0;
3253 	stp->sd_rprotofunc = strrput_proto;
3254 	stp->sd_rmiscfunc = strrput_misc;
3255 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3256 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3257 	stp->sd_ciputctrl = NULL;
3258 	stp->sd_nciputctrl = 0;
3259 	stp->sd_qhead = NULL;
3260 	stp->sd_qtail = NULL;
3261 	stp->sd_servid = NULL;
3262 	stp->sd_nqueues = 0;
3263 	stp->sd_svcflags = 0;
3264 	stp->sd_copyflag = 0;
3265 
3266 	return (stp);
3267 }
3268 
3269 /*
3270  * Free a stream head.
3271  */
3272 void
3273 shfree(stdata_t *stp)
3274 {
3275 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3276 
3277 	stp->sd_wrq = NULL;
3278 
3279 	mutex_enter(&stp->sd_qlock);
3280 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3281 		STRSTAT(strwaits);
3282 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3283 	}
3284 	mutex_exit(&stp->sd_qlock);
3285 
3286 	if (stp->sd_ciputctrl != NULL) {
3287 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3288 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3289 		    stp->sd_nciputctrl, 0);
3290 		ASSERT(ciputctrl_cache != NULL);
3291 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3292 		stp->sd_ciputctrl = NULL;
3293 		stp->sd_nciputctrl = 0;
3294 	}
3295 	ASSERT(stp->sd_qhead == NULL);
3296 	ASSERT(stp->sd_qtail == NULL);
3297 	ASSERT(stp->sd_nqueues == 0);
3298 	kmem_cache_free(stream_head_cache, stp);
3299 }
3300 
3301 /*
3302  * Allocate a pair of queues and a syncq for the pair
3303  */
3304 queue_t *
3305 allocq(void)
3306 {
3307 	queinfo_t *qip;
3308 	queue_t *qp, *wqp;
3309 	syncq_t	*sq;
3310 
3311 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3312 
3313 	qp = &qip->qu_rqueue;
3314 	wqp = &qip->qu_wqueue;
3315 	sq = &qip->qu_syncq;
3316 
3317 	qp->q_last	= NULL;
3318 	qp->q_next	= NULL;
3319 	qp->q_ptr	= NULL;
3320 	qp->q_flag	= QUSE | QREADR;
3321 	qp->q_bandp	= NULL;
3322 	qp->q_stream	= NULL;
3323 	qp->q_syncq	= sq;
3324 	qp->q_nband	= 0;
3325 	qp->q_nfsrv	= NULL;
3326 	qp->q_draining	= 0;
3327 	qp->q_syncqmsgs	= 0;
3328 	qp->q_spri	= 0;
3329 	qp->q_qtstamp	= 0;
3330 	qp->q_sqtstamp	= 0;
3331 	qp->q_fp	= NULL;
3332 
3333 	wqp->q_last	= NULL;
3334 	wqp->q_next	= NULL;
3335 	wqp->q_ptr	= NULL;
3336 	wqp->q_flag	= QUSE;
3337 	wqp->q_bandp	= NULL;
3338 	wqp->q_stream	= NULL;
3339 	wqp->q_syncq	= sq;
3340 	wqp->q_nband	= 0;
3341 	wqp->q_nfsrv	= NULL;
3342 	wqp->q_draining	= 0;
3343 	wqp->q_syncqmsgs = 0;
3344 	wqp->q_qtstamp	= 0;
3345 	wqp->q_sqtstamp	= 0;
3346 	wqp->q_spri	= 0;
3347 
3348 	sq->sq_count	= 0;
3349 	sq->sq_rmqcount	= 0;
3350 	sq->sq_flags	= 0;
3351 	sq->sq_type	= 0;
3352 	sq->sq_callbflags = 0;
3353 	sq->sq_cancelid	= 0;
3354 	sq->sq_ciputctrl = NULL;
3355 	sq->sq_nciputctrl = 0;
3356 	sq->sq_needexcl = 0;
3357 	sq->sq_svcflags = 0;
3358 
3359 	return (qp);
3360 }
3361 
3362 /*
3363  * Free a pair of queues and the "attached" syncq.
3364  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3365  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3366  */
3367 void
3368 freeq(queue_t *qp)
3369 {
3370 	qband_t *qbp, *nqbp;
3371 	syncq_t *sq, *outer;
3372 	queue_t *wqp = _WR(qp);
3373 
3374 	ASSERT(qp->q_flag & QREADR);
3375 
3376 	/*
3377 	 * If a previously dispatched taskq job is scheduled to run
3378 	 * sync_service() or a service routine is scheduled for the
3379 	 * queues about to be freed, wait here until all service is
3380 	 * done on the queue and all associated queues and syncqs.
3381 	 */
3382 	wait_svc(qp);
3383 
3384 	(void) flush_syncq(qp->q_syncq, qp);
3385 	(void) flush_syncq(wqp->q_syncq, wqp);
3386 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3387 
3388 	/*
3389 	 * Flush the queues before q_next is set to NULL This is needed
3390 	 * in order to backenable any downstream queue before we go away.
3391 	 * Note: we are already removed from the stream so that the
3392 	 * backenabling will not cause any messages to be delivered to our
3393 	 * put procedures.
3394 	 */
3395 	flushq(qp, FLUSHALL);
3396 	flushq(wqp, FLUSHALL);
3397 
3398 	/* Tidy up - removeq only does a half-remove from stream */
3399 	qp->q_next = wqp->q_next = NULL;
3400 	ASSERT(!(qp->q_flag & QENAB));
3401 	ASSERT(!(wqp->q_flag & QENAB));
3402 
3403 	outer = qp->q_syncq->sq_outer;
3404 	if (outer != NULL) {
3405 		outer_remove(outer, qp->q_syncq);
3406 		if (wqp->q_syncq != qp->q_syncq)
3407 			outer_remove(outer, wqp->q_syncq);
3408 	}
3409 	/*
3410 	 * Free any syncqs that are outside what allocq returned.
3411 	 */
3412 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3413 		free_syncq(qp->q_syncq);
3414 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3415 		free_syncq(wqp->q_syncq);
3416 
3417 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3418 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3419 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3420 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3421 	sq = SQ(qp);
3422 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3423 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3424 	ASSERT(sq->sq_outer == NULL);
3425 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3426 	ASSERT(sq->sq_callbpend == NULL);
3427 	ASSERT(sq->sq_needexcl == 0);
3428 
3429 	if (sq->sq_ciputctrl != NULL) {
3430 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3431 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3432 		    sq->sq_nciputctrl, 0);
3433 		ASSERT(ciputctrl_cache != NULL);
3434 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3435 		sq->sq_ciputctrl = NULL;
3436 		sq->sq_nciputctrl = 0;
3437 	}
3438 
3439 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3440 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3441 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3442 
3443 	qp->q_flag &= ~QUSE;
3444 	wqp->q_flag &= ~QUSE;
3445 
3446 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3447 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3448 
3449 	qbp = qp->q_bandp;
3450 	while (qbp) {
3451 		nqbp = qbp->qb_next;
3452 		freeband(qbp);
3453 		qbp = nqbp;
3454 	}
3455 	qbp = wqp->q_bandp;
3456 	while (qbp) {
3457 		nqbp = qbp->qb_next;
3458 		freeband(qbp);
3459 		qbp = nqbp;
3460 	}
3461 	kmem_cache_free(queue_cache, qp);
3462 }
3463 
3464 /*
3465  * Allocate a qband structure.
3466  */
3467 qband_t *
3468 allocband(void)
3469 {
3470 	qband_t *qbp;
3471 
3472 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3473 	if (qbp == NULL)
3474 		return (NULL);
3475 
3476 	qbp->qb_next	= NULL;
3477 	qbp->qb_count	= 0;
3478 	qbp->qb_mblkcnt	= 0;
3479 	qbp->qb_first	= NULL;
3480 	qbp->qb_last	= NULL;
3481 	qbp->qb_flag	= 0;
3482 
3483 	return (qbp);
3484 }
3485 
3486 /*
3487  * Free a qband structure.
3488  */
3489 void
3490 freeband(qband_t *qbp)
3491 {
3492 	kmem_cache_free(qband_cache, qbp);
3493 }
3494 
3495 /*
3496  * Just like putnextctl(9F), except that allocb_wait() is used.
3497  *
3498  * Consolidation Private, and of course only callable from the stream head or
3499  * routines that may block.
3500  */
3501 int
3502 putnextctl_wait(queue_t *q, int type)
3503 {
3504 	mblk_t *bp;
3505 	int error;
3506 
3507 	if ((datamsg(type) && (type != M_DELAY)) ||
3508 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3509 		return (0);
3510 
3511 	bp->b_datap->db_type = (unsigned char)type;
3512 	putnext(q, bp);
3513 	return (1);
3514 }
3515 
3516 /*
3517  * run any possible bufcalls.
3518  */
3519 void
3520 runbufcalls(void)
3521 {
3522 	strbufcall_t *bcp;
3523 
3524 	mutex_enter(&bcall_monitor);
3525 	mutex_enter(&strbcall_lock);
3526 
3527 	if (strbcalls.bc_head) {
3528 		size_t count;
3529 		int nevent;
3530 
3531 		/*
3532 		 * count how many events are on the list
3533 		 * now so we can check to avoid looping
3534 		 * in low memory situations
3535 		 */
3536 		nevent = 0;
3537 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3538 			nevent++;
3539 
3540 		/*
3541 		 * get estimate of available memory from kmem_avail().
3542 		 * awake all bufcall functions waiting for
3543 		 * memory whose request could be satisfied
3544 		 * by 'count' memory and let 'em fight for it.
3545 		 */
3546 		count = kmem_avail();
3547 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3548 			STRSTAT(bufcalls);
3549 			--nevent;
3550 			if (bcp->bc_size <= count) {
3551 				bcp->bc_executor = curthread;
3552 				mutex_exit(&strbcall_lock);
3553 				(*bcp->bc_func)(bcp->bc_arg);
3554 				mutex_enter(&strbcall_lock);
3555 				bcp->bc_executor = NULL;
3556 				cv_broadcast(&bcall_cv);
3557 				strbcalls.bc_head = bcp->bc_next;
3558 				kmem_free(bcp, sizeof (strbufcall_t));
3559 			} else {
3560 				/*
3561 				 * too big, try again later - note
3562 				 * that nevent was decremented above
3563 				 * so we won't retry this one on this
3564 				 * iteration of the loop
3565 				 */
3566 				if (bcp->bc_next != NULL) {
3567 					strbcalls.bc_head = bcp->bc_next;
3568 					bcp->bc_next = NULL;
3569 					strbcalls.bc_tail->bc_next = bcp;
3570 					strbcalls.bc_tail = bcp;
3571 				}
3572 			}
3573 		}
3574 		if (strbcalls.bc_head == NULL)
3575 			strbcalls.bc_tail = NULL;
3576 	}
3577 
3578 	mutex_exit(&strbcall_lock);
3579 	mutex_exit(&bcall_monitor);
3580 }
3581 
3582 
3583 /*
3584  * actually run queue's service routine.
3585  */
3586 static void
3587 runservice(queue_t *q)
3588 {
3589 	qband_t *qbp;
3590 
3591 	ASSERT(q->q_qinfo->qi_srvp);
3592 again:
3593 	entersq(q->q_syncq, SQ_SVC);
3594 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3595 		"runservice starts:%p", q);
3596 
3597 	if (!(q->q_flag & QWCLOSE))
3598 		(*q->q_qinfo->qi_srvp)(q);
3599 
3600 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3601 		"runservice ends:(%p)", q);
3602 
3603 	leavesq(q->q_syncq, SQ_SVC);
3604 
3605 	mutex_enter(QLOCK(q));
3606 	if (q->q_flag & QENAB) {
3607 		q->q_flag &= ~QENAB;
3608 		mutex_exit(QLOCK(q));
3609 		goto again;
3610 	}
3611 	q->q_flag &= ~QINSERVICE;
3612 	q->q_flag &= ~QBACK;
3613 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3614 		qbp->qb_flag &= ~QB_BACK;
3615 	/*
3616 	 * Wakeup thread waiting for the service procedure
3617 	 * to be run (strclose and qdetach).
3618 	 */
3619 	cv_broadcast(&q->q_wait);
3620 
3621 	mutex_exit(QLOCK(q));
3622 }
3623 
3624 /*
3625  * Background processing of bufcalls.
3626  */
3627 void
3628 streams_bufcall_service(void)
3629 {
3630 	callb_cpr_t	cprinfo;
3631 
3632 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3633 	    "streams_bufcall_service");
3634 
3635 	mutex_enter(&strbcall_lock);
3636 
3637 	for (;;) {
3638 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3639 			mutex_exit(&strbcall_lock);
3640 			runbufcalls();
3641 			mutex_enter(&strbcall_lock);
3642 		}
3643 		if (strbcalls.bc_head != NULL) {
3644 			clock_t wt, tick;
3645 
3646 			STRSTAT(bcwaits);
3647 			/* Wait for memory to become available */
3648 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3649 			tick = SEC_TO_TICK(60);
3650 			time_to_wait(&wt, tick);
3651 			(void) cv_timedwait(&memavail_cv, &strbcall_lock, wt);
3652 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3653 		}
3654 
3655 		/* Wait for new work to arrive */
3656 		if (strbcalls.bc_head == NULL) {
3657 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3658 			cv_wait(&strbcall_cv, &strbcall_lock);
3659 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3660 		}
3661 	}
3662 }
3663 
3664 /*
3665  * Background processing of streams background tasks which failed
3666  * taskq_dispatch.
3667  */
3668 static void
3669 streams_qbkgrnd_service(void)
3670 {
3671 	callb_cpr_t cprinfo;
3672 	queue_t *q;
3673 
3674 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3675 	    "streams_bkgrnd_service");
3676 
3677 	mutex_enter(&service_queue);
3678 
3679 	for (;;) {
3680 		/*
3681 		 * Wait for work to arrive.
3682 		 */
3683 		while ((freebs_list == NULL) && (qhead == NULL)) {
3684 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3685 			cv_wait(&services_to_run, &service_queue);
3686 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3687 		}
3688 		/*
3689 		 * Handle all pending freebs requests to free memory.
3690 		 */
3691 		while (freebs_list != NULL) {
3692 			mblk_t *mp = freebs_list;
3693 			freebs_list = mp->b_next;
3694 			mutex_exit(&service_queue);
3695 			mblk_free(mp);
3696 			mutex_enter(&service_queue);
3697 		}
3698 		/*
3699 		 * Run pending queues.
3700 		 */
3701 		while (qhead != NULL) {
3702 			DQ(q, qhead, qtail, q_link);
3703 			ASSERT(q != NULL);
3704 			mutex_exit(&service_queue);
3705 			queue_service(q);
3706 			mutex_enter(&service_queue);
3707 		}
3708 		ASSERT(qhead == NULL && qtail == NULL);
3709 	}
3710 }
3711 
3712 /*
3713  * Background processing of streams background tasks which failed
3714  * taskq_dispatch.
3715  */
3716 static void
3717 streams_sqbkgrnd_service(void)
3718 {
3719 	callb_cpr_t cprinfo;
3720 	syncq_t *sq;
3721 
3722 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3723 	    "streams_sqbkgrnd_service");
3724 
3725 	mutex_enter(&service_queue);
3726 
3727 	for (;;) {
3728 		/*
3729 		 * Wait for work to arrive.
3730 		 */
3731 		while (sqhead == NULL) {
3732 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3733 			cv_wait(&syncqs_to_run, &service_queue);
3734 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3735 		}
3736 
3737 		/*
3738 		 * Run pending syncqs.
3739 		 */
3740 		while (sqhead != NULL) {
3741 			DQ(sq, sqhead, sqtail, sq_next);
3742 			ASSERT(sq != NULL);
3743 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3744 			mutex_exit(&service_queue);
3745 			syncq_service(sq);
3746 			mutex_enter(&service_queue);
3747 		}
3748 	}
3749 }
3750 
3751 /*
3752  * Disable the syncq and wait for background syncq processing to complete.
3753  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3754  * list.
3755  */
3756 void
3757 wait_sq_svc(syncq_t *sq)
3758 {
3759 	mutex_enter(SQLOCK(sq));
3760 	sq->sq_svcflags |= SQ_DISABLED;
3761 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3762 		syncq_t *sq_chase;
3763 		syncq_t *sq_curr;
3764 		int removed;
3765 
3766 		ASSERT(sq->sq_servcount == 1);
3767 		mutex_enter(&service_queue);
3768 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3769 		mutex_exit(&service_queue);
3770 		if (removed) {
3771 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3772 			sq->sq_servcount = 0;
3773 			STRSTAT(sqremoved);
3774 			goto done;
3775 		}
3776 	}
3777 	while (sq->sq_servcount != 0) {
3778 		sq->sq_flags |= SQ_WANTWAKEUP;
3779 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3780 	}
3781 done:
3782 	mutex_exit(SQLOCK(sq));
3783 }
3784 
3785 /*
3786  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3787  * Add the argument to the end of the sqhead list and set the flag
3788  * indicating this syncq has been enabled.  If it has already been
3789  * enabled, don't do anything.
3790  * This routine assumes that SQLOCK is held.
3791  * NOTE that the lock order is to have the SQLOCK first,
3792  * so if the service_syncq lock is held, we need to release it
3793  * before aquiring the SQLOCK (mostly relevant for the background
3794  * thread, and this seems to be common among the STREAMS global locks).
3795  * Note the the sq_svcflags are protected by the SQLOCK.
3796  */
3797 void
3798 sqenable(syncq_t *sq)
3799 {
3800 	/*
3801 	 * This is probably not important except for where I believe it
3802 	 * is being called.  At that point, it should be held (and it
3803 	 * is a pain to release it just for this routine, so don't do
3804 	 * it).
3805 	 */
3806 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3807 
3808 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3809 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3810 
3811 	/*
3812 	 * Do not put on list if background thread is scheduled or
3813 	 * syncq is disabled.
3814 	 */
3815 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3816 		return;
3817 
3818 	/*
3819 	 * Check whether we should enable sq at all.
3820 	 * Non PERMOD syncqs may be drained by at most one thread.
3821 	 * PERMOD syncqs may be drained by several threads but we limit the
3822 	 * total amount to the lesser of
3823 	 *	Number of queues on the squeue and
3824 	 *	Number of CPUs.
3825 	 */
3826 	if (sq->sq_servcount != 0) {
3827 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3828 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3829 			STRSTAT(sqtoomany);
3830 			return;
3831 		}
3832 	}
3833 
3834 	sq->sq_tstamp = lbolt;
3835 	STRSTAT(sqenables);
3836 
3837 	/* Attempt a taskq dispatch */
3838 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3839 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3840 	if (sq->sq_servid != NULL) {
3841 		sq->sq_servcount++;
3842 		return;
3843 	}
3844 
3845 	/*
3846 	 * This taskq dispatch failed, but a previous one may have succeeded.
3847 	 * Don't try to schedule on the background thread whilst there is
3848 	 * outstanding taskq processing.
3849 	 */
3850 	if (sq->sq_servcount != 0)
3851 		return;
3852 
3853 	/*
3854 	 * System is low on resources and can't perform a non-sleeping
3855 	 * dispatch. Schedule the syncq for a background thread and mark the
3856 	 * syncq to avoid any further taskq dispatch attempts.
3857 	 */
3858 	mutex_enter(&service_queue);
3859 	STRSTAT(taskqfails);
3860 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3861 	sq->sq_svcflags |= SQ_BGTHREAD;
3862 	sq->sq_servcount = 1;
3863 	cv_signal(&syncqs_to_run);
3864 	mutex_exit(&service_queue);
3865 }
3866 
3867 /*
3868  * Note: fifo_close() depends on the mblk_t on the queue being freed
3869  * asynchronously. The asynchronous freeing of messages breaks the
3870  * recursive call chain of fifo_close() while there are I_SENDFD type of
3871  * messages refering other file pointers on the queue. Then when
3872  * closing pipes it can avoid stack overflow in case of daisy-chained
3873  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3874  * share the same fifolock_t).
3875  */
3876 
3877 /* ARGSUSED */
3878 void
3879 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3880 {
3881 	ASSERT(dbp->db_mblk == mp);
3882 
3883 	/*
3884 	 * Check data sanity. The dblock should have non-empty free function.
3885 	 * It is better to panic here then later when the dblock is freed
3886 	 * asynchronously when the context is lost.
3887 	 */
3888 	if (dbp->db_frtnp->free_func == NULL) {
3889 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3890 		    (void *) dbp);
3891 	}
3892 
3893 	STRSTAT(freebs);
3894 	if (taskq_dispatch(streams_taskq, (task_func_t *)mblk_free, mp,
3895 	    TQ_NOSLEEP) == NULL) {
3896 		/*
3897 		 * System is low on resources and can't perform a non-sleeping
3898 		 * dispatch. Schedule for a background thread.
3899 		 */
3900 		mutex_enter(&service_queue);
3901 		STRSTAT(taskqfails);
3902 		mp->b_next = freebs_list;
3903 		freebs_list = mp;
3904 		cv_signal(&services_to_run);
3905 		mutex_exit(&service_queue);
3906 	}
3907 }
3908 
3909 /*
3910  * Set the QBACK or QB_BACK flag in the given queue for
3911  * the given priority band.
3912  */
3913 void
3914 setqback(queue_t *q, unsigned char pri)
3915 {
3916 	int i;
3917 	qband_t *qbp;
3918 	qband_t **qbpp;
3919 
3920 	ASSERT(MUTEX_HELD(QLOCK(q)));
3921 	if (pri != 0) {
3922 		if (pri > q->q_nband) {
3923 			qbpp = &q->q_bandp;
3924 			while (*qbpp)
3925 				qbpp = &(*qbpp)->qb_next;
3926 			while (pri > q->q_nband) {
3927 				if ((*qbpp = allocband()) == NULL) {
3928 					cmn_err(CE_WARN,
3929 					    "setqback: can't allocate qband\n");
3930 					return;
3931 				}
3932 				(*qbpp)->qb_hiwat = q->q_hiwat;
3933 				(*qbpp)->qb_lowat = q->q_lowat;
3934 				q->q_nband++;
3935 				qbpp = &(*qbpp)->qb_next;
3936 			}
3937 		}
3938 		qbp = q->q_bandp;
3939 		i = pri;
3940 		while (--i)
3941 			qbp = qbp->qb_next;
3942 		qbp->qb_flag |= QB_BACK;
3943 	} else {
3944 		q->q_flag |= QBACK;
3945 	}
3946 }
3947 
3948 int
3949 strcopyin(void *from, void *to, size_t len, int copyflag)
3950 {
3951 	if (copyflag & U_TO_K) {
3952 		ASSERT((copyflag & K_TO_K) == 0);
3953 		if (copyin(from, to, len))
3954 			return (EFAULT);
3955 	} else {
3956 		ASSERT(copyflag & K_TO_K);
3957 		bcopy(from, to, len);
3958 	}
3959 	return (0);
3960 }
3961 
3962 int
3963 strcopyout(void *from, void *to, size_t len, int copyflag)
3964 {
3965 	if (copyflag & U_TO_K) {
3966 		if (copyout(from, to, len))
3967 			return (EFAULT);
3968 	} else {
3969 		ASSERT(copyflag & K_TO_K);
3970 		bcopy(from, to, len);
3971 	}
3972 	return (0);
3973 }
3974 
3975 /*
3976  * strsignal_nolock() posts a signal to the process(es) at the stream head.
3977  * It assumes that the stream head lock is already held, whereas strsignal()
3978  * acquires the lock first.  This routine was created because a few callers
3979  * release the stream head lock before calling only to re-acquire it after
3980  * it returns.
3981  */
3982 void
3983 strsignal_nolock(stdata_t *stp, int sig, int32_t band)
3984 {
3985 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3986 	switch (sig) {
3987 	case SIGPOLL:
3988 		if (stp->sd_sigflags & S_MSG)
3989 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
3990 		break;
3991 
3992 	default:
3993 		if (stp->sd_pgidp) {
3994 			pgsignal(stp->sd_pgidp, sig);
3995 		}
3996 		break;
3997 	}
3998 }
3999 
4000 void
4001 strsignal(stdata_t *stp, int sig, int32_t band)
4002 {
4003 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4004 		"strsignal:%p, %X, %X", stp, sig, band);
4005 
4006 	mutex_enter(&stp->sd_lock);
4007 	switch (sig) {
4008 	case SIGPOLL:
4009 		if (stp->sd_sigflags & S_MSG)
4010 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4011 		break;
4012 
4013 	default:
4014 		if (stp->sd_pgidp) {
4015 			pgsignal(stp->sd_pgidp, sig);
4016 		}
4017 		break;
4018 	}
4019 	mutex_exit(&stp->sd_lock);
4020 }
4021 
4022 void
4023 strhup(stdata_t *stp)
4024 {
4025 	ASSERT(mutex_owned(&stp->sd_lock));
4026 	pollwakeup(&stp->sd_pollist, POLLHUP);
4027 	if (stp->sd_sigflags & S_HANGUP)
4028 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4029 }
4030 
4031 /*
4032  * Backenable the first queue upstream from `q' with a service procedure.
4033  */
4034 void
4035 backenable(queue_t *q, uchar_t pri)
4036 {
4037 	queue_t	*nq;
4038 
4039 	/*
4040 	 * our presence might not prevent other modules in our own
4041 	 * stream from popping/pushing since the caller of getq might not
4042 	 * have a claim on the queue (some drivers do a getq on somebody
4043 	 * else's queue - they know that the queue itself is not going away
4044 	 * but the framework has to guarantee q_next in that stream.)
4045 	 */
4046 	claimstr(q);
4047 
4048 	/* find nearest back queue with service proc */
4049 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4050 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4051 	}
4052 
4053 	if (nq) {
4054 		kthread_t *freezer;
4055 		/*
4056 		 * backenable can be called either with no locks held
4057 		 * or with the stream frozen (the latter occurs when a module
4058 		 * calls rmvq with the stream frozen.) If the stream is frozen
4059 		 * by the caller the caller will hold all qlocks in the stream.
4060 		 */
4061 		freezer = STREAM(q)->sd_freezer;
4062 		if (freezer != curthread) {
4063 			mutex_enter(QLOCK(nq));
4064 		}
4065 #ifdef DEBUG
4066 		else {
4067 			ASSERT(frozenstr(q));
4068 			ASSERT(MUTEX_HELD(QLOCK(q)));
4069 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4070 		}
4071 #endif
4072 		setqback(nq, pri);
4073 		qenable_locked(nq);
4074 		if (freezer != curthread)
4075 			mutex_exit(QLOCK(nq));
4076 	}
4077 	releasestr(q);
4078 }
4079 
4080 /*
4081  * Return the appropriate errno when one of flags_to_check is set
4082  * in sd_flags. Uses the exported error routines if they are set.
4083  * Will return 0 if non error is set (or if the exported error routines
4084  * do not return an error).
4085  *
4086  * If there is both a read and write error to check we prefer the read error.
4087  * Also, give preference to recorded errno's over the error functions.
4088  * The flags that are handled are:
4089  *	STPLEX		return EINVAL
4090  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4091  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4092  *	STRHUP		return sd_werror
4093  *
4094  * If the caller indicates that the operation is a peek a nonpersistent error
4095  * is not cleared.
4096  */
4097 int
4098 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4099 {
4100 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4101 	int error = 0;
4102 
4103 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4104 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4105 	if (sd_flag & STPLEX)
4106 		error = EINVAL;
4107 	else if (sd_flag & STRDERR) {
4108 		error = stp->sd_rerror;
4109 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4110 			/*
4111 			 * Read errors are non-persistent i.e. discarded once
4112 			 * returned to a non-peeking caller,
4113 			 */
4114 			stp->sd_rerror = 0;
4115 			stp->sd_flag &= ~STRDERR;
4116 		}
4117 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4118 			int clearerr = 0;
4119 
4120 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4121 						&clearerr);
4122 			if (clearerr) {
4123 				stp->sd_flag &= ~STRDERR;
4124 				stp->sd_rderrfunc = NULL;
4125 			}
4126 		}
4127 	} else if (sd_flag & STWRERR) {
4128 		error = stp->sd_werror;
4129 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4130 			/*
4131 			 * Write errors are non-persistent i.e. discarded once
4132 			 * returned to a non-peeking caller,
4133 			 */
4134 			stp->sd_werror = 0;
4135 			stp->sd_flag &= ~STWRERR;
4136 		}
4137 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4138 			int clearerr = 0;
4139 
4140 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4141 						&clearerr);
4142 			if (clearerr) {
4143 				stp->sd_flag &= ~STWRERR;
4144 				stp->sd_wrerrfunc = NULL;
4145 			}
4146 		}
4147 	} else if (sd_flag & STRHUP) {
4148 		/* sd_werror set when STRHUP */
4149 		error = stp->sd_werror;
4150 	}
4151 	return (error);
4152 }
4153 
4154 
4155 /*
4156  * single-thread open/close/push/pop
4157  * for twisted streams also
4158  */
4159 int
4160 strstartplumb(stdata_t *stp, int flag, int cmd)
4161 {
4162 	int waited = 1;
4163 	int error = 0;
4164 
4165 	if (STRMATED(stp)) {
4166 		struct stdata *stmatep = stp->sd_mate;
4167 
4168 		STRLOCKMATES(stp);
4169 		while (waited) {
4170 			waited = 0;
4171 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4172 				if ((cmd == I_POP) &&
4173 				    (flag & (FNDELAY|FNONBLOCK))) {
4174 					STRUNLOCKMATES(stp);
4175 					return (EAGAIN);
4176 				}
4177 				waited = 1;
4178 				mutex_exit(&stp->sd_lock);
4179 				if (!cv_wait_sig(&stmatep->sd_monitor,
4180 				    &stmatep->sd_lock)) {
4181 					mutex_exit(&stmatep->sd_lock);
4182 					return (EINTR);
4183 				}
4184 				mutex_exit(&stmatep->sd_lock);
4185 				STRLOCKMATES(stp);
4186 			}
4187 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4188 				if ((cmd == I_POP) &&
4189 					(flag & (FNDELAY|FNONBLOCK))) {
4190 					STRUNLOCKMATES(stp);
4191 					return (EAGAIN);
4192 				}
4193 				waited = 1;
4194 				mutex_exit(&stmatep->sd_lock);
4195 				if (!cv_wait_sig(&stp->sd_monitor,
4196 				    &stp->sd_lock)) {
4197 					mutex_exit(&stp->sd_lock);
4198 					return (EINTR);
4199 				}
4200 				mutex_exit(&stp->sd_lock);
4201 				STRLOCKMATES(stp);
4202 			}
4203 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4204 				error = strgeterr(stp,
4205 					STRDERR|STWRERR|STRHUP|STPLEX, 0);
4206 				if (error != 0) {
4207 					STRUNLOCKMATES(stp);
4208 					return (error);
4209 				}
4210 			}
4211 		}
4212 		stp->sd_flag |= STRPLUMB;
4213 		STRUNLOCKMATES(stp);
4214 	} else {
4215 		mutex_enter(&stp->sd_lock);
4216 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4217 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4218 			    (flag & (FNDELAY|FNONBLOCK))) {
4219 				mutex_exit(&stp->sd_lock);
4220 				return (EAGAIN);
4221 			}
4222 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4223 				mutex_exit(&stp->sd_lock);
4224 				return (EINTR);
4225 			}
4226 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4227 				error = strgeterr(stp,
4228 					STRDERR|STWRERR|STRHUP|STPLEX, 0);
4229 				if (error != 0) {
4230 					mutex_exit(&stp->sd_lock);
4231 					return (error);
4232 				}
4233 			}
4234 		}
4235 		stp->sd_flag |= STRPLUMB;
4236 		mutex_exit(&stp->sd_lock);
4237 	}
4238 	return (0);
4239 }
4240 
4241 /*
4242  * Complete the plumbing operation associated with stream `stp'.
4243  */
4244 void
4245 strendplumb(stdata_t *stp)
4246 {
4247 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4248 	ASSERT(stp->sd_flag & STRPLUMB);
4249 	stp->sd_flag &= ~STRPLUMB;
4250 	cv_broadcast(&stp->sd_monitor);
4251 }
4252 
4253 /*
4254  * This describes how the STREAMS framework handles synchronization
4255  * during open/push and close/pop.
4256  * The key interfaces for open and close are qprocson and qprocsoff,
4257  * respectively. While the close case in general is harder both open
4258  * have close have significant similarities.
4259  *
4260  * During close the STREAMS framework has to both ensure that there
4261  * are no stale references to the queue pair (and syncq) that
4262  * are being closed and also provide the guarantees that are documented
4263  * in qprocsoff(9F).
4264  * If there are stale references to the queue that is closing it can
4265  * result in kernel memory corruption or kernel panics.
4266  *
4267  * Note that is it up to the module/driver to ensure that it itself
4268  * does not have any stale references to the closing queues once its close
4269  * routine returns. This includes:
4270  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4271  *    associated with the queues. For timeout and bufcall callbacks the
4272  *    module/driver also has to ensure (or wait for) any callbacks that
4273  *    are in progress.
4274  *  - If the module/driver is using esballoc it has to ensure that any
4275  *    esballoc free functions do not refer to a queue that has closed.
4276  *    (Note that in general the close routine can not wait for the esballoc'ed
4277  *    messages to be freed since that can cause a deadlock.)
4278  *  - Cancelling any interrupts that refer to the closing queues and
4279  *    also ensuring that there are no interrupts in progress that will
4280  *    refer to the closing queues once the close routine returns.
4281  *  - For multiplexors removing any driver global state that refers to
4282  *    the closing queue and also ensuring that there are no threads in
4283  *    the multiplexor that has picked up a queue pointer but not yet
4284  *    finished using it.
4285  *
4286  * In addition, a driver/module can only reference the q_next pointer
4287  * in its open, close, put, or service procedures or in a
4288  * qtimeout/qbufcall callback procedure executing "on" the correct
4289  * stream. Thus it can not reference the q_next pointer in an interrupt
4290  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4291  * it can not reference q_next of a different queue e.g. in a mux that
4292  * passes messages from one queues put/service procedure to another queue.
4293  * In all the cases when the driver/module can not access the q_next
4294  * field it must use the *next* versions e.g. canputnext instead of
4295  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4296  *
4297  *
4298  * Assuming that the driver/module conforms to the above constraints
4299  * the STREAMS framework has to avoid stale references to q_next for all
4300  * the framework internal cases which include (but are not limited to):
4301  *  - Threads in canput/canputnext/backenable and elsewhere that are
4302  *    walking q_next.
4303  *  - Messages on a syncq that have a reference to the queue through b_queue.
4304  *  - Messages on an outer perimeter (syncq) that have a reference to the
4305  *    queue through b_queue.
4306  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4307  *    Note that only canput and bcanput use q_nfsrv without any locking.
4308  *
4309  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4310  * after qprocsoff returns, the framework has to ensure that no threads can
4311  * enter the put or service routines for the closing read or write-side queue.
4312  * In addition to preventing "direct" entry into the put procedures
4313  * the framework also has to prevent messages being drained from
4314  * the syncq or the outer perimeter.
4315  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4316  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4317  * qprocsoff has returned.
4318  * Note that if a module/driver uses put(9F) on one of its own queues
4319  * it is up to the module/driver to ensure that the put() doesn't
4320  * get called when the queue is closing.
4321  *
4322  *
4323  * The framework aspects of the above "contract" is implemented by
4324  * qprocsoff, removeq, and strlock:
4325  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4326  *    entering the service procedures.
4327  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4328  *    canputnext, backenable etc from dereferencing the q_next that will
4329  *    soon change.
4330  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4331  *    or other q_next walker that uses claimstr/releasestr to finish.
4332  *  - optionally for every syncq in the stream strlock acquires all the
4333  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4334  *    that no thread executes in the put or service procedures and that no
4335  *    thread is draining into the module/driver. This ensures that no
4336  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4337  *    currently executing hence no such thread can end up with the old stale
4338  *    q_next value and no canput/backenable can have the old stale
4339  *    q_nfsrv/q_next.
4340  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4341  *    have either finished or observed the QWCLOSE flag and gone away.
4342  */
4343 
4344 
4345 /*
4346  * Get all the locks necessary to change q_next.
4347  *
4348  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for  the
4349  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4350  * the only threads inside the sqncq are threads currently calling removeq().
4351  * Since threads calling removeq() are in the process of removing their queues
4352  * from the stream, we do not need to worry about them accessing a stale q_next
4353  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4354  * for them can cause deadlock).
4355  *
4356  * This routine is subject to starvation since it does not set any flag to
4357  * prevent threads from entering a module in the stream(i.e. sq_count can
4358  * increase on some syncq while it is waiting on some other syncq.)
4359  *
4360  * Assumes that only one thread attempts to call strlock for a given
4361  * stream. If this is not the case the two threads would deadlock.
4362  * This assumption is guaranteed since strlock is only called by insertq
4363  * and removeq and streams plumbing changes are single-threaded for
4364  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4365  *
4366  * For pipes, it is not difficult to atomically designate a pair of streams
4367  * to be mated. Once mated atomically by the framework the twisted pair remain
4368  * configured that way until dismantled atomically by the framework.
4369  * When plumbing takes place on a twisted stream it is necessary to ensure that
4370  * this operation is done exclusively on the twisted stream since two such
4371  * operations, each initiated on different ends of the pipe will deadlock
4372  * waiting for each other to complete.
4373  *
4374  * On entry, no locks should be held.
4375  * The locks acquired and held by strlock depends on a few factors.
4376  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4377  *   and held on exit and all sq_count are at an acceptable level.
4378  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4379  *   sd_refcnt being zero.
4380  */
4381 
4382 static void
4383 strlock(struct stdata *stp, sqlist_t *sqlist)
4384 {
4385 	syncql_t *sql, *sql2;
4386 retry:
4387 	/*
4388 	 * Wait for any claimstr to go away.
4389 	 */
4390 	if (STRMATED(stp)) {
4391 		struct stdata *stp1, *stp2;
4392 
4393 		STRLOCKMATES(stp);
4394 		/*
4395 		 * Note that the selection of locking order is not
4396 		 * important, just that they are always aquired in
4397 		 * the same order.  To assure this, we choose this
4398 		 * order based on the value of the pointer, and since
4399 		 * the pointer will not change for the life of this
4400 		 * pair, we will always grab the locks in the same
4401 		 * order (and hence, prevent deadlocks).
4402 		 */
4403 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4404 			stp1 = stp;
4405 			stp2 = stp->sd_mate;
4406 		} else {
4407 			stp2 = stp;
4408 			stp1 = stp->sd_mate;
4409 		}
4410 		mutex_enter(&stp1->sd_reflock);
4411 		if (stp1->sd_refcnt > 0) {
4412 			STRUNLOCKMATES(stp);
4413 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4414 			mutex_exit(&stp1->sd_reflock);
4415 			goto retry;
4416 		}
4417 		mutex_enter(&stp2->sd_reflock);
4418 		if (stp2->sd_refcnt > 0) {
4419 			STRUNLOCKMATES(stp);
4420 			mutex_exit(&stp1->sd_reflock);
4421 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4422 			mutex_exit(&stp2->sd_reflock);
4423 			goto retry;
4424 		}
4425 		STREAM_PUTLOCKS_ENTER(stp1);
4426 		STREAM_PUTLOCKS_ENTER(stp2);
4427 	} else {
4428 		mutex_enter(&stp->sd_lock);
4429 		mutex_enter(&stp->sd_reflock);
4430 		while (stp->sd_refcnt > 0) {
4431 			mutex_exit(&stp->sd_lock);
4432 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4433 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4434 				mutex_exit(&stp->sd_reflock);
4435 				mutex_enter(&stp->sd_lock);
4436 				mutex_enter(&stp->sd_reflock);
4437 			}
4438 		}
4439 		STREAM_PUTLOCKS_ENTER(stp);
4440 	}
4441 
4442 	if (sqlist == NULL)
4443 		return;
4444 
4445 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4446 		syncq_t *sq = sql->sql_sq;
4447 		uint16_t count;
4448 
4449 		mutex_enter(SQLOCK(sq));
4450 		count = sq->sq_count;
4451 		ASSERT(sq->sq_rmqcount <= count);
4452 		SQ_PUTLOCKS_ENTER(sq);
4453 		SUM_SQ_PUTCOUNTS(sq, count);
4454 		if (count == sq->sq_rmqcount)
4455 			continue;
4456 
4457 		/* Failed - drop all locks that we have acquired so far */
4458 		if (STRMATED(stp)) {
4459 			STREAM_PUTLOCKS_EXIT(stp);
4460 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4461 			STRUNLOCKMATES(stp);
4462 			mutex_exit(&stp->sd_reflock);
4463 			mutex_exit(&stp->sd_mate->sd_reflock);
4464 		} else {
4465 			STREAM_PUTLOCKS_EXIT(stp);
4466 			mutex_exit(&stp->sd_lock);
4467 			mutex_exit(&stp->sd_reflock);
4468 		}
4469 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4470 		    sql2 = sql2->sql_next) {
4471 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4472 			mutex_exit(SQLOCK(sql2->sql_sq));
4473 		}
4474 
4475 		/*
4476 		 * The wait loop below may starve when there are many threads
4477 		 * claiming the syncq. This is especially a problem with permod
4478 		 * syncqs (IP). To lessen the impact of the problem we increment
4479 		 * sq_needexcl and clear fastbits so that putnexts will slow
4480 		 * down and call sqenable instead of draining right away.
4481 		 */
4482 		sq->sq_needexcl++;
4483 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4484 		while (count > sq->sq_rmqcount) {
4485 			sq->sq_flags |= SQ_WANTWAKEUP;
4486 			SQ_PUTLOCKS_EXIT(sq);
4487 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4488 			count = sq->sq_count;
4489 			SQ_PUTLOCKS_ENTER(sq);
4490 			SUM_SQ_PUTCOUNTS(sq, count);
4491 		}
4492 		sq->sq_needexcl--;
4493 		if (sq->sq_needexcl == 0)
4494 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4495 		SQ_PUTLOCKS_EXIT(sq);
4496 		ASSERT(count == sq->sq_rmqcount);
4497 		mutex_exit(SQLOCK(sq));
4498 		goto retry;
4499 	}
4500 }
4501 
4502 /*
4503  * Drop all the locks that strlock acquired.
4504  */
4505 static void
4506 strunlock(struct stdata *stp, sqlist_t *sqlist)
4507 {
4508 	syncql_t *sql;
4509 
4510 	if (STRMATED(stp)) {
4511 		STREAM_PUTLOCKS_EXIT(stp);
4512 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4513 		STRUNLOCKMATES(stp);
4514 		mutex_exit(&stp->sd_reflock);
4515 		mutex_exit(&stp->sd_mate->sd_reflock);
4516 	} else {
4517 		STREAM_PUTLOCKS_EXIT(stp);
4518 		mutex_exit(&stp->sd_lock);
4519 		mutex_exit(&stp->sd_reflock);
4520 	}
4521 
4522 	if (sqlist == NULL)
4523 		return;
4524 
4525 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4526 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4527 		mutex_exit(SQLOCK(sql->sql_sq));
4528 	}
4529 }
4530 
4531 /*
4532  * When the module has service procedure, we need check if the next
4533  * module which has service procedure is in flow control to trigger
4534  * the backenable.
4535  */
4536 static void
4537 backenable_insertedq(queue_t *q)
4538 {
4539 	qband_t	*qbp;
4540 
4541 	claimstr(q);
4542 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4543 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4544 			backenable(q, 0);
4545 
4546 		qbp = q->q_next->q_nfsrv->q_bandp;
4547 		for (; qbp != NULL; qbp = qbp->qb_next)
4548 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4549 				backenable(q, qbp->qb_first->b_band);
4550 	}
4551 	releasestr(q);
4552 }
4553 
4554 /*
4555  * Given two read queues, insert a new single one after another.
4556  *
4557  * This routine acquires all the necessary locks in order to change
4558  * q_next and related pointer using strlock().
4559  * It depends on the stream head ensuring that there are no concurrent
4560  * insertq or removeq on the same stream. The stream head ensures this
4561  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4562  *
4563  * Note that no syncq locks are held during the q_next change. This is
4564  * applied to all streams since, unlike removeq, there is no problem of stale
4565  * pointers when adding a module to the stream. Thus drivers/modules that do a
4566  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4567  * applied this optimization to all streams.
4568  */
4569 void
4570 insertq(struct stdata *stp, queue_t *new)
4571 {
4572 	queue_t	*after;
4573 	queue_t *wafter;
4574 	queue_t *wnew = _WR(new);
4575 	boolean_t have_fifo = B_FALSE;
4576 
4577 	if (new->q_flag & _QINSERTING) {
4578 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4579 		after = new->q_next;
4580 		wafter = _WR(new->q_next);
4581 	} else {
4582 		after = _RD(stp->sd_wrq);
4583 		wafter = stp->sd_wrq;
4584 	}
4585 
4586 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4587 		"insertq:%p, %p", after, new);
4588 	ASSERT(after->q_flag & QREADR);
4589 	ASSERT(new->q_flag & QREADR);
4590 
4591 	strlock(stp, NULL);
4592 
4593 	/* Do we have a FIFO? */
4594 	if (wafter->q_next == after) {
4595 		have_fifo = B_TRUE;
4596 		wnew->q_next = new;
4597 	} else {
4598 		wnew->q_next = wafter->q_next;
4599 	}
4600 	new->q_next = after;
4601 
4602 	set_nfsrv_ptr(new, wnew, after, wafter);
4603 	/*
4604 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4605 	 * so only reset this flag after calling it.
4606 	 */
4607 	new->q_flag &= ~_QINSERTING;
4608 
4609 	if (have_fifo) {
4610 		wafter->q_next = wnew;
4611 	} else {
4612 		if (wafter->q_next)
4613 			_OTHERQ(wafter->q_next)->q_next = new;
4614 		wafter->q_next = wnew;
4615 	}
4616 
4617 	set_qend(new);
4618 	/* The QEND flag might have to be updated for the upstream guy */
4619 	set_qend(after);
4620 
4621 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4622 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4623 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4624 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4625 	strsetuio(stp);
4626 
4627 	/*
4628 	 * If this was a module insertion, bump the push count.
4629 	 */
4630 	if (!(new->q_flag & QISDRV))
4631 		stp->sd_pushcnt++;
4632 
4633 	strunlock(stp, NULL);
4634 
4635 	/* check if the write Q needs backenable */
4636 	backenable_insertedq(wnew);
4637 
4638 	/* check if the read Q needs backenable */
4639 	backenable_insertedq(new);
4640 }
4641 
4642 /*
4643  * Given a read queue, unlink it from any neighbors.
4644  *
4645  * This routine acquires all the necessary locks in order to
4646  * change q_next and related pointers and also guard against
4647  * stale references (e.g. through q_next) to the queue that
4648  * is being removed. It also plays part of the role in ensuring
4649  * that the module's/driver's put procedure doesn't get called
4650  * after qprocsoff returns.
4651  *
4652  * Removeq depends on the stream head ensuring that there are
4653  * no concurrent insertq or removeq on the same stream. The
4654  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4655  * STRPLUMB.
4656  *
4657  * The set of locks needed to remove the queue is different in
4658  * different cases:
4659  *
4660  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4661  * waiting for the syncq reference count to drop to 0 indicating that no
4662  * non-close threads are present anywhere in the stream. This ensures that any
4663  * module/driver can reference q_next in its open, close, put, or service
4664  * procedures.
4665  *
4666  * The sq_rmqcount counter tracks the number of threads inside removeq().
4667  * strlock() ensures that there is either no threads executing inside perimeter
4668  * or there is only a thread calling qprocsoff().
4669  *
4670  * strlock() compares the value of sq_count with the number of threads inside
4671  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4672  * any threads waiting in strlock() when the sq_rmqcount increases.
4673  */
4674 
4675 void
4676 removeq(queue_t *qp)
4677 {
4678 	queue_t *wqp = _WR(qp);
4679 	struct stdata *stp = STREAM(qp);
4680 	sqlist_t *sqlist = NULL;
4681 	boolean_t isdriver;
4682 	int moved;
4683 	syncq_t *sq = qp->q_syncq;
4684 	syncq_t *wsq = wqp->q_syncq;
4685 
4686 	ASSERT(stp);
4687 
4688 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4689 		"removeq:%p %p", qp, wqp);
4690 	ASSERT(qp->q_flag&QREADR);
4691 
4692 	/*
4693 	 * For queues using Synchronous streams, we must wait for all threads in
4694 	 * rwnext() to drain out before proceeding.
4695 	 */
4696 	if (qp->q_flag & QSYNCSTR) {
4697 		/* First, we need wakeup any threads blocked in rwnext() */
4698 		mutex_enter(SQLOCK(sq));
4699 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4700 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4701 			cv_broadcast(&sq->sq_wait);
4702 		}
4703 		mutex_exit(SQLOCK(sq));
4704 
4705 		if (wsq != sq) {
4706 			mutex_enter(SQLOCK(wsq));
4707 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4708 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4709 				cv_broadcast(&wsq->sq_wait);
4710 			}
4711 			mutex_exit(SQLOCK(wsq));
4712 		}
4713 
4714 		mutex_enter(QLOCK(qp));
4715 		while (qp->q_rwcnt > 0) {
4716 			qp->q_flag |= QWANTRMQSYNC;
4717 			cv_wait(&qp->q_wait, QLOCK(qp));
4718 		}
4719 		mutex_exit(QLOCK(qp));
4720 
4721 		mutex_enter(QLOCK(wqp));
4722 		while (wqp->q_rwcnt > 0) {
4723 			wqp->q_flag |= QWANTRMQSYNC;
4724 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4725 		}
4726 		mutex_exit(QLOCK(wqp));
4727 	}
4728 
4729 	mutex_enter(SQLOCK(sq));
4730 	sq->sq_rmqcount++;
4731 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4732 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4733 		cv_broadcast(&sq->sq_wait);
4734 	}
4735 	mutex_exit(SQLOCK(sq));
4736 
4737 	isdriver = (qp->q_flag & QISDRV);
4738 
4739 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4740 	strlock(stp, sqlist);
4741 
4742 	reset_nfsrv_ptr(qp, wqp);
4743 
4744 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4745 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4746 	/* Do we have a FIFO? */
4747 	if (wqp->q_next == qp) {
4748 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4749 	} else {
4750 		if (wqp->q_next)
4751 			backq(qp)->q_next = qp->q_next;
4752 		if (qp->q_next)
4753 			backq(wqp)->q_next = wqp->q_next;
4754 	}
4755 
4756 	/* The QEND flag might have to be updated for the upstream guy */
4757 	if (qp->q_next)
4758 		set_qend(qp->q_next);
4759 
4760 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4761 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4762 
4763 	/*
4764 	 * Move any messages destined for the put procedures to the next
4765 	 * syncq in line. Otherwise free them.
4766 	 */
4767 	moved = 0;
4768 	/*
4769 	 * Quick check to see whether there are any messages or events.
4770 	 */
4771 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4772 		moved += propagate_syncq(qp);
4773 	if (wqp->q_syncqmsgs != 0 ||
4774 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4775 		moved += propagate_syncq(wqp);
4776 
4777 	strsetuio(stp);
4778 
4779 	/*
4780 	 * If this was a module removal, decrement the push count.
4781 	 */
4782 	if (!isdriver)
4783 		stp->sd_pushcnt--;
4784 
4785 	strunlock(stp, sqlist);
4786 	sqlist_free(sqlist);
4787 
4788 	/*
4789 	 * Make sure any messages that were propagated are drained.
4790 	 * Also clear any QFULL bit caused by messages that were propagated.
4791 	 */
4792 
4793 	if (qp->q_next != NULL) {
4794 		clr_qfull(qp);
4795 		/*
4796 		 * For the driver calling qprocsoff, propagate_syncq
4797 		 * frees all the messages instead of putting it in
4798 		 * the stream head
4799 		 */
4800 		if (!isdriver && (moved > 0))
4801 			emptysq(qp->q_next->q_syncq);
4802 	}
4803 	if (wqp->q_next != NULL) {
4804 		clr_qfull(wqp);
4805 		/*
4806 		 * We come here for any pop of a module except for the
4807 		 * case of driver being removed. We don't call emptysq
4808 		 * if we did not move any messages. This will avoid holding
4809 		 * PERMOD syncq locks in emptysq
4810 		 */
4811 		if (moved > 0)
4812 			emptysq(wqp->q_next->q_syncq);
4813 	}
4814 
4815 	mutex_enter(SQLOCK(sq));
4816 	sq->sq_rmqcount--;
4817 	mutex_exit(SQLOCK(sq));
4818 }
4819 
4820 /*
4821  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
4822  * SQ_WRITER) on a syncq.
4823  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
4824  * sync queue and waits until sq_count reaches maxcnt.
4825  *
4826  * if maxcnt is -1 there's no need to grab sq_putlocks since the caller
4827  * does not care about putnext threads that are in the middle of calling put
4828  * entry points.
4829  *
4830  * This routine is used for both inner and outer syncqs.
4831  */
4832 static void
4833 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
4834 {
4835 	uint16_t count = 0;
4836 
4837 	mutex_enter(SQLOCK(sq));
4838 	/*
4839 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
4840 	 * SQ_FROZEN will be set if there is a frozen stream that has a
4841 	 * queue which also refers to this "shared" syncq.
4842 	 * SQ_BLOCKED will be set if there is "off" queue which also
4843 	 * refers to this "shared" syncq.
4844 	 */
4845 	if (maxcnt != -1) {
4846 		count = sq->sq_count;
4847 		SQ_PUTLOCKS_ENTER(sq);
4848 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4849 		SUM_SQ_PUTCOUNTS(sq, count);
4850 	}
4851 	sq->sq_needexcl++;
4852 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
4853 
4854 	while ((sq->sq_flags & flag) ||
4855 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
4856 		sq->sq_flags |= SQ_WANTWAKEUP;
4857 		if (maxcnt != -1) {
4858 			SQ_PUTLOCKS_EXIT(sq);
4859 		}
4860 		cv_wait(&sq->sq_wait, SQLOCK(sq));
4861 		if (maxcnt != -1) {
4862 			count = sq->sq_count;
4863 			SQ_PUTLOCKS_ENTER(sq);
4864 			SUM_SQ_PUTCOUNTS(sq, count);
4865 		}
4866 	}
4867 	sq->sq_needexcl--;
4868 	sq->sq_flags |= flag;
4869 	ASSERT(maxcnt == -1 || count == maxcnt);
4870 	if (maxcnt != -1) {
4871 		if (sq->sq_needexcl == 0) {
4872 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4873 		}
4874 		SQ_PUTLOCKS_EXIT(sq);
4875 	} else if (sq->sq_needexcl == 0) {
4876 		SQ_PUTCOUNT_SETFAST(sq);
4877 	}
4878 
4879 	mutex_exit(SQLOCK(sq));
4880 }
4881 
4882 /*
4883  * Reset a flag that was set with blocksq.
4884  *
4885  * Can not use this routine to reset SQ_WRITER.
4886  *
4887  * If "isouter" is set then the syncq is assumed to be an outer perimeter
4888  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
4889  * to handle the queued qwriter operations.
4890  *
4891  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4892  * sq_putlocks are used.
4893  */
4894 static void
4895 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
4896 {
4897 	uint16_t flags;
4898 
4899 	mutex_enter(SQLOCK(sq));
4900 	ASSERT(resetflag != SQ_WRITER);
4901 	ASSERT(sq->sq_flags & resetflag);
4902 	flags = sq->sq_flags & ~resetflag;
4903 	sq->sq_flags = flags;
4904 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
4905 		if (flags & SQ_WANTWAKEUP) {
4906 			flags &= ~SQ_WANTWAKEUP;
4907 			cv_broadcast(&sq->sq_wait);
4908 		}
4909 		sq->sq_flags = flags;
4910 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
4911 			if (!isouter) {
4912 				/* drain_syncq drops SQLOCK */
4913 				drain_syncq(sq);
4914 				return;
4915 			}
4916 		}
4917 	}
4918 	mutex_exit(SQLOCK(sq));
4919 }
4920 
4921 /*
4922  * Reset a flag that was set with blocksq.
4923  * Does not drain the syncq. Use emptysq() for that.
4924  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
4925  *
4926  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4927  * sq_putlocks are used.
4928  */
4929 static int
4930 dropsq(syncq_t *sq, uint16_t resetflag)
4931 {
4932 	uint16_t flags;
4933 
4934 	mutex_enter(SQLOCK(sq));
4935 	ASSERT(sq->sq_flags & resetflag);
4936 	flags = sq->sq_flags & ~resetflag;
4937 	if (flags & SQ_WANTWAKEUP) {
4938 		flags &= ~SQ_WANTWAKEUP;
4939 		cv_broadcast(&sq->sq_wait);
4940 	}
4941 	sq->sq_flags = flags;
4942 	mutex_exit(SQLOCK(sq));
4943 	if (flags & SQ_QUEUED)
4944 		return (1);
4945 	return (0);
4946 }
4947 
4948 /*
4949  * Empty all the messages on a syncq.
4950  *
4951  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4952  * sq_putlocks are used.
4953  */
4954 static void
4955 emptysq(syncq_t *sq)
4956 {
4957 	uint16_t flags;
4958 
4959 	mutex_enter(SQLOCK(sq));
4960 	flags = sq->sq_flags;
4961 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
4962 		/*
4963 		 * To prevent potential recursive invocation of drain_syncq we
4964 		 * do not call drain_syncq if count is non-zero.
4965 		 */
4966 		if (sq->sq_count == 0) {
4967 			/* drain_syncq() drops SQLOCK */
4968 			drain_syncq(sq);
4969 			return;
4970 		} else
4971 			sqenable(sq);
4972 	}
4973 	mutex_exit(SQLOCK(sq));
4974 }
4975 
4976 /*
4977  * Ordered insert while removing duplicates.
4978  */
4979 static void
4980 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
4981 {
4982 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
4983 
4984 	prev_sqlpp = &sqlist->sqlist_head;
4985 	while ((sqlp = *prev_sqlpp) != NULL) {
4986 		if (sqlp->sql_sq >= sqp) {
4987 			if (sqlp->sql_sq == sqp)	/* duplicate */
4988 				return;
4989 			break;
4990 		}
4991 		prev_sqlpp = &sqlp->sql_next;
4992 	}
4993 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
4994 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
4995 	new_sqlp->sql_next = sqlp;
4996 	new_sqlp->sql_sq = sqp;
4997 	*prev_sqlpp = new_sqlp;
4998 }
4999 
5000 /*
5001  * Walk the write side queues until we hit either the driver
5002  * or a twist in the stream (_SAMESTR will return false in both
5003  * these cases) then turn around and walk the read side queues
5004  * back up to the stream head.
5005  */
5006 static void
5007 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5008 {
5009 	while (q != NULL) {
5010 		sqlist_insert(sqlist, q->q_syncq);
5011 
5012 		if (_SAMESTR(q))
5013 			q = q->q_next;
5014 		else if (!(q->q_flag & QREADR))
5015 			q = _RD(q);
5016 		else
5017 			q = NULL;
5018 	}
5019 }
5020 
5021 /*
5022  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5023  * associated with the "q" parameter. The resulting list is sorted in a
5024  * canonical order and is free of duplicates.
5025  * Assumes the passed queue is a _RD(q).
5026  */
5027 static sqlist_t *
5028 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5029 {
5030 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5031 
5032 	/*
5033 	 * start with the current queue/qpair
5034 	 */
5035 	ASSERT(q->q_flag & QREADR);
5036 
5037 	sqlist_insert(sqlist, q->q_syncq);
5038 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5039 
5040 	sqlist_insertall(sqlist, stp->sd_wrq);
5041 	if (do_twist)
5042 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5043 
5044 	return (sqlist);
5045 }
5046 
5047 static sqlist_t *
5048 sqlist_alloc(struct stdata *stp, int kmflag)
5049 {
5050 	size_t sqlist_size;
5051 	sqlist_t *sqlist;
5052 
5053 	/*
5054 	 * Allocate 2 syncql_t's for each pushed module. Note that
5055 	 * the sqlist_t structure already has 4 syncql_t's built in:
5056 	 * 2 for the stream head, and 2 for the driver/other stream head.
5057 	 */
5058 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5059 		sizeof (sqlist_t);
5060 	if (STRMATED(stp))
5061 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5062 	sqlist = kmem_alloc(sqlist_size, kmflag);
5063 
5064 	sqlist->sqlist_head = NULL;
5065 	sqlist->sqlist_size = sqlist_size;
5066 	sqlist->sqlist_index = 0;
5067 
5068 	return (sqlist);
5069 }
5070 
5071 /*
5072  * Free the list created by sqlist_alloc()
5073  */
5074 static void
5075 sqlist_free(sqlist_t *sqlist)
5076 {
5077 	kmem_free(sqlist, sqlist->sqlist_size);
5078 }
5079 
5080 /*
5081  * Prevent any new entries into any syncq in this stream.
5082  * Used by freezestr.
5083  */
5084 void
5085 strblock(queue_t *q)
5086 {
5087 	struct stdata	*stp;
5088 	syncql_t	*sql;
5089 	sqlist_t	*sqlist;
5090 
5091 	q = _RD(q);
5092 
5093 	stp = STREAM(q);
5094 	ASSERT(stp != NULL);
5095 
5096 	/*
5097 	 * Get a sorted list with all the duplicates removed containing
5098 	 * all the syncqs referenced by this stream.
5099 	 */
5100 	sqlist = sqlist_build(q, stp, B_FALSE);
5101 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5102 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5103 	sqlist_free(sqlist);
5104 }
5105 
5106 /*
5107  * Release the block on new entries into this stream
5108  */
5109 void
5110 strunblock(queue_t *q)
5111 {
5112 	struct stdata	*stp;
5113 	syncql_t	*sql;
5114 	sqlist_t	*sqlist;
5115 	int		drain_needed;
5116 
5117 	q = _RD(q);
5118 
5119 	/*
5120 	 * Get a sorted list with all the duplicates removed containing
5121 	 * all the syncqs referenced by this stream.
5122 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5123 	 * starting to drain them; otherwise the draining might
5124 	 * cause a freezestr in some module on the stream (which
5125 	 * would deadlock.)
5126 	 */
5127 	stp = STREAM(q);
5128 	ASSERT(stp != NULL);
5129 	sqlist = sqlist_build(q, stp, B_FALSE);
5130 	drain_needed = 0;
5131 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5132 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5133 	if (drain_needed) {
5134 		for (sql = sqlist->sqlist_head; sql != NULL;
5135 		    sql = sql->sql_next)
5136 			emptysq(sql->sql_sq);
5137 	}
5138 	sqlist_free(sqlist);
5139 }
5140 
5141 #ifdef DEBUG
5142 static int
5143 qprocsareon(queue_t *rq)
5144 {
5145 	if (rq->q_next == NULL)
5146 		return (0);
5147 	return (_WR(rq->q_next)->q_next == _WR(rq));
5148 }
5149 
5150 int
5151 qclaimed(queue_t *q)
5152 {
5153 	uint_t count;
5154 
5155 	count = q->q_syncq->sq_count;
5156 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5157 	return (count != 0);
5158 }
5159 
5160 /*
5161  * Check if anyone has frozen this stream with freezestr
5162  */
5163 int
5164 frozenstr(queue_t *q)
5165 {
5166 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5167 }
5168 #endif /* DEBUG */
5169 
5170 /*
5171  * Enter a queue.
5172  * Obsoleted interface. Should not be used.
5173  */
5174 void
5175 enterq(queue_t *q)
5176 {
5177 	entersq(q->q_syncq, SQ_CALLBACK);
5178 }
5179 
5180 void
5181 leaveq(queue_t *q)
5182 {
5183 	leavesq(q->q_syncq, SQ_CALLBACK);
5184 }
5185 
5186 /*
5187  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5188  * to check.
5189  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5190  * calls and the running of open, close and service procedures.
5191  *
5192  * if c_inner bit is set no need to grab sq_putlocks since we don't care
5193  * if other threads have entered or are entering put entry point.
5194  *
5195  * if c_inner bit is set it might have been posible to use
5196  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5197  * open/close path for IP) but since the count may need to be decremented in
5198  * qwait() we wouldn't know which counter to decrement. Currently counter is
5199  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5200  * in the future we might use curthread id bits to select the counter and this
5201  * would stay constant across routine calls.
5202  */
5203 void
5204 entersq(syncq_t *sq, int entrypoint)
5205 {
5206 	uint16_t	count = 0;
5207 	uint16_t	flags;
5208 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5209 	uint16_t	type;
5210 	uint_t		c_inner = entrypoint & SQ_CI;
5211 	uint_t		c_outer = entrypoint & SQ_CO;
5212 
5213 	/*
5214 	 * Increment ref count to keep closes out of this queue.
5215 	 */
5216 	ASSERT(sq);
5217 	ASSERT(c_inner && c_outer);
5218 	mutex_enter(SQLOCK(sq));
5219 	flags = sq->sq_flags;
5220 	type = sq->sq_type;
5221 	if (!(type & c_inner)) {
5222 		/* Make sure all putcounts now use slowlock. */
5223 		count = sq->sq_count;
5224 		SQ_PUTLOCKS_ENTER(sq);
5225 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5226 		SUM_SQ_PUTCOUNTS(sq, count);
5227 		sq->sq_needexcl++;
5228 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5229 		waitflags |= SQ_MESSAGES;
5230 	}
5231 	/*
5232 	 * Wait until we can enter the inner perimeter.
5233 	 * If we want exclusive access we wait until sq_count is 0.
5234 	 * We have to do this before entering the outer perimeter in order
5235 	 * to preserve put/close message ordering.
5236 	 */
5237 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5238 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5239 		if (!(type & c_inner)) {
5240 			SQ_PUTLOCKS_EXIT(sq);
5241 		}
5242 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5243 		if (!(type & c_inner)) {
5244 			count = sq->sq_count;
5245 			SQ_PUTLOCKS_ENTER(sq);
5246 			SUM_SQ_PUTCOUNTS(sq, count);
5247 		}
5248 		flags = sq->sq_flags;
5249 	}
5250 
5251 	if (!(type & c_inner)) {
5252 		ASSERT(sq->sq_needexcl > 0);
5253 		sq->sq_needexcl--;
5254 		if (sq->sq_needexcl == 0) {
5255 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5256 		}
5257 	}
5258 
5259 	/* Check if we need to enter the outer perimeter */
5260 	if (!(type & c_outer)) {
5261 		/*
5262 		 * We have to enter the outer perimeter exclusively before
5263 		 * we can increment sq_count to avoid deadlock. This implies
5264 		 * that we have to re-check sq_flags and sq_count.
5265 		 *
5266 		 * is it possible to have c_inner set when c_outer is not set?
5267 		 */
5268 		if (!(type & c_inner)) {
5269 			SQ_PUTLOCKS_EXIT(sq);
5270 		}
5271 		mutex_exit(SQLOCK(sq));
5272 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5273 		mutex_enter(SQLOCK(sq));
5274 		flags = sq->sq_flags;
5275 		/*
5276 		 * there should be no need to recheck sq_putcounts
5277 		 * because outer_enter() has already waited for them to clear
5278 		 * after setting SQ_WRITER.
5279 		 */
5280 		count = sq->sq_count;
5281 #ifdef DEBUG
5282 		/*
5283 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5284 		 * of doing an ASSERT internally. Others should do
5285 		 * something like
5286 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5287 		 * without the need to #ifdef DEBUG it.
5288 		 */
5289 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5290 #endif
5291 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5292 		    (!(type & c_inner) && count != 0)) {
5293 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5294 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5295 			count = sq->sq_count;
5296 			flags = sq->sq_flags;
5297 		}
5298 	}
5299 
5300 	sq->sq_count++;
5301 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5302 	if (!(type & c_inner)) {
5303 		/* Exclusive entry */
5304 		ASSERT(sq->sq_count == 1);
5305 		sq->sq_flags |= SQ_EXCL;
5306 		if (type & c_outer) {
5307 			SQ_PUTLOCKS_EXIT(sq);
5308 		}
5309 	}
5310 	mutex_exit(SQLOCK(sq));
5311 }
5312 
5313 /*
5314  * leave a syncq. announce to framework that closes may proceed.
5315  * c_inner and c_outer specifies which concurrency bits
5316  * to check.
5317  *
5318  * must never be called from driver or module put entry point.
5319  *
5320  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5321  * sq_putlocks are used.
5322  */
5323 void
5324 leavesq(syncq_t *sq, int entrypoint)
5325 {
5326 	uint16_t	flags;
5327 	uint16_t	type;
5328 	uint_t		c_outer = entrypoint & SQ_CO;
5329 #ifdef DEBUG
5330 	uint_t		c_inner = entrypoint & SQ_CI;
5331 #endif
5332 
5333 	/*
5334 	 * decrement ref count, drain the syncq if possible, and wake up
5335 	 * any waiting close.
5336 	 */
5337 	ASSERT(sq);
5338 	ASSERT(c_inner && c_outer);
5339 	mutex_enter(SQLOCK(sq));
5340 	flags = sq->sq_flags;
5341 	type = sq->sq_type;
5342 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5343 
5344 		if (flags & SQ_WANTWAKEUP) {
5345 			flags &= ~SQ_WANTWAKEUP;
5346 			cv_broadcast(&sq->sq_wait);
5347 		}
5348 		if (flags & SQ_WANTEXWAKEUP) {
5349 			flags &= ~SQ_WANTEXWAKEUP;
5350 			cv_broadcast(&sq->sq_exitwait);
5351 		}
5352 
5353 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5354 			/*
5355 			 * The syncq needs to be drained. "Exit" the syncq
5356 			 * before calling drain_syncq.
5357 			 */
5358 			ASSERT(sq->sq_count != 0);
5359 			sq->sq_count--;
5360 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5361 			sq->sq_flags = flags & ~SQ_EXCL;
5362 			drain_syncq(sq);
5363 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5364 			/* Check if we need to exit the outer perimeter */
5365 			/* XXX will this ever be true? */
5366 			if (!(type & c_outer))
5367 				outer_exit(sq->sq_outer);
5368 			return;
5369 		}
5370 	}
5371 	ASSERT(sq->sq_count != 0);
5372 	sq->sq_count--;
5373 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5374 	sq->sq_flags = flags & ~SQ_EXCL;
5375 	mutex_exit(SQLOCK(sq));
5376 
5377 	/* Check if we need to exit the outer perimeter */
5378 	if (!(sq->sq_type & c_outer))
5379 		outer_exit(sq->sq_outer);
5380 }
5381 
5382 /*
5383  * Prevent q_next from changing in this stream by incrementing sq_count.
5384  *
5385  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5386  * sq_putlocks are used.
5387  */
5388 void
5389 claimq(queue_t *qp)
5390 {
5391 	syncq_t	*sq = qp->q_syncq;
5392 
5393 	mutex_enter(SQLOCK(sq));
5394 	sq->sq_count++;
5395 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5396 	mutex_exit(SQLOCK(sq));
5397 }
5398 
5399 /*
5400  * Undo claimq.
5401  *
5402  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5403  * sq_putlocks are used.
5404  */
5405 void
5406 releaseq(queue_t *qp)
5407 {
5408 	syncq_t	*sq = qp->q_syncq;
5409 	uint16_t flags;
5410 
5411 	mutex_enter(SQLOCK(sq));
5412 	ASSERT(sq->sq_count > 0);
5413 	sq->sq_count--;
5414 
5415 	flags = sq->sq_flags;
5416 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5417 		if (flags & SQ_WANTWAKEUP) {
5418 			flags &= ~SQ_WANTWAKEUP;
5419 			cv_broadcast(&sq->sq_wait);
5420 		}
5421 		sq->sq_flags = flags;
5422 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5423 			/*
5424 			 * To prevent potential recursive invocation of
5425 			 * drain_syncq we do not call drain_syncq if count is
5426 			 * non-zero.
5427 			 */
5428 			if (sq->sq_count == 0) {
5429 				drain_syncq(sq);
5430 				return;
5431 			} else
5432 				sqenable(sq);
5433 		}
5434 	}
5435 	mutex_exit(SQLOCK(sq));
5436 }
5437 
5438 /*
5439  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5440  */
5441 void
5442 claimstr(queue_t *qp)
5443 {
5444 	struct stdata *stp = STREAM(qp);
5445 
5446 	mutex_enter(&stp->sd_reflock);
5447 	stp->sd_refcnt++;
5448 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5449 	mutex_exit(&stp->sd_reflock);
5450 }
5451 
5452 /*
5453  * Undo claimstr.
5454  */
5455 void
5456 releasestr(queue_t *qp)
5457 {
5458 	struct stdata *stp = STREAM(qp);
5459 
5460 	mutex_enter(&stp->sd_reflock);
5461 	ASSERT(stp->sd_refcnt != 0);
5462 	if (--stp->sd_refcnt == 0)
5463 		cv_broadcast(&stp->sd_refmonitor);
5464 	mutex_exit(&stp->sd_reflock);
5465 }
5466 
5467 static syncq_t *
5468 new_syncq(void)
5469 {
5470 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5471 }
5472 
5473 static void
5474 free_syncq(syncq_t *sq)
5475 {
5476 	ASSERT(sq->sq_head == NULL);
5477 	ASSERT(sq->sq_outer == NULL);
5478 	ASSERT(sq->sq_callbpend == NULL);
5479 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5480 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5481 
5482 	if (sq->sq_ciputctrl != NULL) {
5483 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5484 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5485 		    sq->sq_nciputctrl, 0);
5486 		ASSERT(ciputctrl_cache != NULL);
5487 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5488 	}
5489 
5490 	sq->sq_tail = NULL;
5491 	sq->sq_evhead = NULL;
5492 	sq->sq_evtail = NULL;
5493 	sq->sq_ciputctrl = NULL;
5494 	sq->sq_nciputctrl = 0;
5495 	sq->sq_count = 0;
5496 	sq->sq_rmqcount = 0;
5497 	sq->sq_callbflags = 0;
5498 	sq->sq_cancelid = 0;
5499 	sq->sq_next = NULL;
5500 	sq->sq_needexcl = 0;
5501 	sq->sq_svcflags = 0;
5502 	sq->sq_nqueues = 0;
5503 	sq->sq_pri = 0;
5504 	sq->sq_onext = NULL;
5505 	sq->sq_oprev = NULL;
5506 	sq->sq_flags = 0;
5507 	sq->sq_type = 0;
5508 	sq->sq_servcount = 0;
5509 
5510 	kmem_cache_free(syncq_cache, sq);
5511 }
5512 
5513 /* Outer perimeter code */
5514 
5515 /*
5516  * The outer syncq uses the fields and flags in the syncq slightly
5517  * differently from the inner syncqs.
5518  *	sq_count	Incremented when there are pending or running
5519  *			writers at the outer perimeter to prevent the set of
5520  *			inner syncqs that belong to the outer perimeter from
5521  *			changing.
5522  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5523  *
5524  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5525  *			inner syncqs are added to or removed from the
5526  *			outer perimeter.
5527  *	SQ_QUEUED	sq_head/tail has messages or eventsqueued.
5528  *
5529  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5530  *			setting the SQ_WRITER flag.
5531  */
5532 
5533 /*
5534  * Get write access at the outer perimeter.
5535  * Note that read access is done by entersq, putnext, and put by simply
5536  * incrementing sq_count in the inner syncq.
5537  *
5538  * Waits until "flags" is no longer set in the outer to prevent multiple
5539  * threads from having write access at the same time. SQ_WRITER has to be part
5540  * of "flags".
5541  *
5542  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5543  * until the outer_exit is finished.
5544  *
5545  * outer_enter is vulnerable to starvation since it does not prevent new
5546  * threads from entering the inner syncqs while it is waiting for sq_count to
5547  * go to zero.
5548  */
5549 void
5550 outer_enter(syncq_t *outer, uint16_t flags)
5551 {
5552 	syncq_t	*sq;
5553 	int	wait_needed;
5554 	uint16_t	count;
5555 
5556 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5557 	    outer->sq_oprev != NULL);
5558 	ASSERT(flags & SQ_WRITER);
5559 
5560 retry:
5561 	mutex_enter(SQLOCK(outer));
5562 	while (outer->sq_flags & flags) {
5563 		outer->sq_flags |= SQ_WANTWAKEUP;
5564 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5565 	}
5566 
5567 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5568 	outer->sq_flags |= SQ_WRITER;
5569 	outer->sq_count++;
5570 	ASSERT(outer->sq_count != 0);	/* wraparound */
5571 	wait_needed = 0;
5572 	/*
5573 	 * Set SQ_WRITER on all the inner syncqs while holding
5574 	 * the SQLOCK on the outer syncq. This ensures that the changing
5575 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5576 	 */
5577 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5578 		mutex_enter(SQLOCK(sq));
5579 		count = sq->sq_count;
5580 		SQ_PUTLOCKS_ENTER(sq);
5581 		sq->sq_flags |= SQ_WRITER;
5582 		SUM_SQ_PUTCOUNTS(sq, count);
5583 		if (count != 0)
5584 			wait_needed = 1;
5585 		SQ_PUTLOCKS_EXIT(sq);
5586 		mutex_exit(SQLOCK(sq));
5587 	}
5588 	mutex_exit(SQLOCK(outer));
5589 
5590 	/*
5591 	 * Get everybody out of the syncqs sequentially.
5592 	 * Note that we don't actually need to aqiure the PUTLOCKS, since
5593 	 * we have already cleared the fastbit, and set QWRITER.  By
5594 	 * definition, the count can not increase since putnext will
5595 	 * take the slowlock path (and the purpose of aquiring the
5596 	 * putlocks was to make sure it didn't increase while we were
5597 	 * waiting).
5598 	 *
5599 	 * Note that we still aquire the PUTLOCKS to be safe.
5600 	 */
5601 	if (wait_needed) {
5602 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5603 			mutex_enter(SQLOCK(sq));
5604 			count = sq->sq_count;
5605 			SQ_PUTLOCKS_ENTER(sq);
5606 			SUM_SQ_PUTCOUNTS(sq, count);
5607 			while (count != 0) {
5608 				sq->sq_flags |= SQ_WANTWAKEUP;
5609 				SQ_PUTLOCKS_EXIT(sq);
5610 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5611 				count = sq->sq_count;
5612 				SQ_PUTLOCKS_ENTER(sq);
5613 				SUM_SQ_PUTCOUNTS(sq, count);
5614 			}
5615 			SQ_PUTLOCKS_EXIT(sq);
5616 			mutex_exit(SQLOCK(sq));
5617 		}
5618 		/*
5619 		 * Verify that none of the flags got set while we
5620 		 * were waiting for the sq_counts to drop.
5621 		 * If this happens we exit and retry entering the
5622 		 * outer perimeter.
5623 		 */
5624 		mutex_enter(SQLOCK(outer));
5625 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5626 			mutex_exit(SQLOCK(outer));
5627 			outer_exit(outer);
5628 			goto retry;
5629 		}
5630 		mutex_exit(SQLOCK(outer));
5631 	}
5632 }
5633 
5634 /*
5635  * Drop the write access at the outer perimeter.
5636  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5637  * decrementing sq_count.
5638  */
5639 void
5640 outer_exit(syncq_t *outer)
5641 {
5642 	syncq_t	*sq;
5643 	int	 drain_needed;
5644 	uint16_t flags;
5645 
5646 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5647 	    outer->sq_oprev != NULL);
5648 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5649 
5650 	/*
5651 	 * Atomically (from the perspective of threads calling become_writer)
5652 	 * drop the write access at the outer perimeter by holding
5653 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5654 	 * SQ_WRITER.
5655 	 * This defines a locking order between the outer perimeter
5656 	 * SQLOCK and the inner perimeter SQLOCKs.
5657 	 */
5658 	mutex_enter(SQLOCK(outer));
5659 	flags = outer->sq_flags;
5660 	ASSERT(outer->sq_flags & SQ_WRITER);
5661 	if (flags & SQ_QUEUED) {
5662 		write_now(outer);
5663 		flags = outer->sq_flags;
5664 	}
5665 
5666 	/*
5667 	 * sq_onext is stable since sq_count has not yet been decreased.
5668 	 * Reset the SQ_WRITER flags in all syncqs.
5669 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5670 	 * inner syncqs.
5671 	 */
5672 	drain_needed = 0;
5673 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5674 		drain_needed += dropsq(sq, SQ_WRITER);
5675 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5676 	flags &= ~SQ_WRITER;
5677 	if (drain_needed) {
5678 		outer->sq_flags = flags;
5679 		mutex_exit(SQLOCK(outer));
5680 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5681 			emptysq(sq);
5682 		mutex_enter(SQLOCK(outer));
5683 		flags = outer->sq_flags;
5684 	}
5685 	if (flags & SQ_WANTWAKEUP) {
5686 		flags &= ~SQ_WANTWAKEUP;
5687 		cv_broadcast(&outer->sq_wait);
5688 	}
5689 	outer->sq_flags = flags;
5690 	ASSERT(outer->sq_count > 0);
5691 	outer->sq_count--;
5692 	mutex_exit(SQLOCK(outer));
5693 }
5694 
5695 /*
5696  * Add another syncq to an outer perimeter.
5697  * Block out all other access to the outer perimeter while it is being
5698  * changed using blocksq.
5699  * Assumes that the caller has *not* done an outer_enter.
5700  *
5701  * Vulnerable to starvation in blocksq.
5702  */
5703 static void
5704 outer_insert(syncq_t *outer, syncq_t *sq)
5705 {
5706 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5707 	    outer->sq_oprev != NULL);
5708 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5709 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5710 
5711 	/* Get exclusive access to the outer perimeter list */
5712 	blocksq(outer, SQ_BLOCKED, 0);
5713 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5714 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5715 
5716 	mutex_enter(SQLOCK(sq));
5717 	sq->sq_outer = outer;
5718 	outer->sq_onext->sq_oprev = sq;
5719 	sq->sq_onext = outer->sq_onext;
5720 	outer->sq_onext = sq;
5721 	sq->sq_oprev = outer;
5722 	mutex_exit(SQLOCK(sq));
5723 	unblocksq(outer, SQ_BLOCKED, 1);
5724 }
5725 
5726 /*
5727  * Remove a syncq from an outer perimeter.
5728  * Block out all other access to the outer perimeter while it is being
5729  * changed using blocksq.
5730  * Assumes that the caller has *not* done an outer_enter.
5731  *
5732  * Vulnerable to starvation in blocksq.
5733  */
5734 static void
5735 outer_remove(syncq_t *outer, syncq_t *sq)
5736 {
5737 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5738 	    outer->sq_oprev != NULL);
5739 	ASSERT(sq->sq_outer == outer);
5740 
5741 	/* Get exclusive access to the outer perimeter list */
5742 	blocksq(outer, SQ_BLOCKED, 0);
5743 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5744 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5745 
5746 	mutex_enter(SQLOCK(sq));
5747 	sq->sq_outer = NULL;
5748 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5749 	sq->sq_oprev->sq_onext = sq->sq_onext;
5750 	sq->sq_oprev = sq->sq_onext = NULL;
5751 	mutex_exit(SQLOCK(sq));
5752 	unblocksq(outer, SQ_BLOCKED, 1);
5753 }
5754 
5755 /*
5756  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5757  * If this is the first callback for this outer perimeter then add
5758  * this outer perimeter to the list of outer perimeters that
5759  * the qwriter_outer_thread will process.
5760  *
5761  * Increments sq_count in the outer syncq to prevent the membership
5762  * of the outer perimeter (in terms of inner syncqs) to change while
5763  * the callback is pending.
5764  */
5765 static void
5766 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5767 {
5768 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5769 
5770 	mp->b_prev = (mblk_t *)func;
5771 	mp->b_queue = q;
5772 	mp->b_next = NULL;
5773 	outer->sq_count++;	/* Decremented when dequeued */
5774 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5775 	if (outer->sq_evhead == NULL) {
5776 		/* First message. */
5777 		outer->sq_evhead = outer->sq_evtail = mp;
5778 		outer->sq_flags |= SQ_EVENTS;
5779 		mutex_exit(SQLOCK(outer));
5780 		STRSTAT(qwr_outer);
5781 		(void) taskq_dispatch(streams_taskq,
5782 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5783 	} else {
5784 		ASSERT(outer->sq_flags & SQ_EVENTS);
5785 		outer->sq_evtail->b_next = mp;
5786 		outer->sq_evtail = mp;
5787 		mutex_exit(SQLOCK(outer));
5788 	}
5789 }
5790 
5791 /*
5792  * Try and upgrade to write access at the outer perimeter. If this can
5793  * not be done without blocking then queue the callback to be done
5794  * by the qwriter_outer_thread.
5795  *
5796  * This routine can only be called from put or service procedures plus
5797  * asynchronous callback routines that have properly entered to
5798  * queue (with entersq.) Thus qwriter(OUTER) assumes the caller has one claim
5799  * on the syncq associated with q.
5800  */
5801 void
5802 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5803 {
5804 	syncq_t	*osq, *sq, *outer;
5805 	int	failed;
5806 	uint16_t flags;
5807 
5808 	osq = q->q_syncq;
5809 	outer = osq->sq_outer;
5810 	if (outer == NULL)
5811 		panic("qwriter(PERIM_OUTER): no outer perimeter");
5812 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5813 	    outer->sq_oprev != NULL);
5814 
5815 	mutex_enter(SQLOCK(outer));
5816 	flags = outer->sq_flags;
5817 	/*
5818 	 * If some thread is traversing sq_next, or if we are blocked by
5819 	 * outer_insert or outer_remove, or if the we already have queued
5820 	 * callbacks, then queue this callback for later processing.
5821 	 *
5822 	 * Also queue the qwriter for an interrupt thread in order
5823 	 * to reduce the time spent running at high IPL.
5824 	 * to identify there are events.
5825 	 */
5826 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
5827 		/*
5828 		 * Queue the become_writer request.
5829 		 * The queueing is atomic under SQLOCK(outer) in order
5830 		 * to synchronize with outer_exit.
5831 		 * queue_writer will drop the outer SQLOCK
5832 		 */
5833 		if (flags & SQ_BLOCKED) {
5834 			/* Must set SQ_WRITER on inner perimeter */
5835 			mutex_enter(SQLOCK(osq));
5836 			osq->sq_flags |= SQ_WRITER;
5837 			mutex_exit(SQLOCK(osq));
5838 		} else {
5839 			if (!(flags & SQ_WRITER)) {
5840 				/*
5841 				 * The outer could have been SQ_BLOCKED thus
5842 				 * SQ_WRITER might not be set on the inner.
5843 				 */
5844 				mutex_enter(SQLOCK(osq));
5845 				osq->sq_flags |= SQ_WRITER;
5846 				mutex_exit(SQLOCK(osq));
5847 			}
5848 			ASSERT(osq->sq_flags & SQ_WRITER);
5849 		}
5850 		queue_writer(outer, func, q, mp);
5851 		return;
5852 	}
5853 	/*
5854 	 * We are half-way to exclusive access to the outer perimeter.
5855 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
5856 	 * while the inner syncqs are traversed.
5857 	 */
5858 	outer->sq_count++;
5859 	ASSERT(outer->sq_count != 0);	/* wraparound */
5860 	flags |= SQ_WRITER;
5861 	/*
5862 	 * Check if we can run the function immediately. Mark all
5863 	 * syncqs with the writer flag to prevent new entries into
5864 	 * put and service procedures.
5865 	 *
5866 	 * Set SQ_WRITER on all the inner syncqs while holding
5867 	 * the SQLOCK on the outer syncq. This ensures that the changing
5868 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5869 	 */
5870 	failed = 0;
5871 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5872 		uint16_t count;
5873 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
5874 
5875 		mutex_enter(SQLOCK(sq));
5876 		count = sq->sq_count;
5877 		SQ_PUTLOCKS_ENTER(sq);
5878 		SUM_SQ_PUTCOUNTS(sq, count);
5879 		if (sq->sq_count > maxcnt)
5880 			failed = 1;
5881 		sq->sq_flags |= SQ_WRITER;
5882 		SQ_PUTLOCKS_EXIT(sq);
5883 		mutex_exit(SQLOCK(sq));
5884 	}
5885 	if (failed) {
5886 		/*
5887 		 * Some other thread has a read claim on the outer perimeter.
5888 		 * Queue the callback for deferred processing.
5889 		 *
5890 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
5891 		 * so that other qwriter(OUTER) calls will queue their
5892 		 * callbacks as well. queue_writer increments sq_count so we
5893 		 * decrement to compensate for the our increment.
5894 		 *
5895 		 * Dropping SQ_WRITER enables the writer thread to work
5896 		 * on this outer perimeter.
5897 		 */
5898 		outer->sq_flags = flags;
5899 		queue_writer(outer, func, q, mp);
5900 		/* queue_writer dropper the lock */
5901 		mutex_enter(SQLOCK(outer));
5902 		ASSERT(outer->sq_count > 0);
5903 		outer->sq_count--;
5904 		ASSERT(outer->sq_flags & SQ_WRITER);
5905 		flags = outer->sq_flags;
5906 		flags &= ~SQ_WRITER;
5907 		if (flags & SQ_WANTWAKEUP) {
5908 			flags &= ~SQ_WANTWAKEUP;
5909 			cv_broadcast(&outer->sq_wait);
5910 		}
5911 		outer->sq_flags = flags;
5912 		mutex_exit(SQLOCK(outer));
5913 		return;
5914 	} else {
5915 		outer->sq_flags = flags;
5916 		mutex_exit(SQLOCK(outer));
5917 	}
5918 
5919 	/* Can run it immediately */
5920 	(*func)(q, mp);
5921 
5922 	outer_exit(outer);
5923 }
5924 
5925 /*
5926  * Dequeue all writer callbacks from the outer perimeter and run them.
5927  */
5928 static void
5929 write_now(syncq_t *outer)
5930 {
5931 	mblk_t		*mp;
5932 	queue_t		*q;
5933 	void	(*func)();
5934 
5935 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5936 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5937 	    outer->sq_oprev != NULL);
5938 	while ((mp = outer->sq_evhead) != NULL) {
5939 		/*
5940 		 * queues cannot be placed on the queuelist on the outer
5941 		 * perimiter.
5942 		 */
5943 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
5944 		ASSERT((outer->sq_flags & SQ_EVENTS));
5945 
5946 		outer->sq_evhead = mp->b_next;
5947 		if (outer->sq_evhead == NULL) {
5948 			outer->sq_evtail = NULL;
5949 			outer->sq_flags &= ~SQ_EVENTS;
5950 		}
5951 		ASSERT(outer->sq_count != 0);
5952 		outer->sq_count--;	/* Incremented when enqueued. */
5953 		mutex_exit(SQLOCK(outer));
5954 		/*
5955 		 * Drop the message if the queue is closing.
5956 		 * Make sure that the queue is "claimed" when the callback
5957 		 * is run in order to satisfy various ASSERTs.
5958 		 */
5959 		q = mp->b_queue;
5960 		func = (void (*)())mp->b_prev;
5961 		ASSERT(func != NULL);
5962 		mp->b_next = mp->b_prev = NULL;
5963 		if (q->q_flag & QWCLOSE) {
5964 			freemsg(mp);
5965 		} else {
5966 			claimq(q);
5967 			(*func)(q, mp);
5968 			releaseq(q);
5969 		}
5970 		mutex_enter(SQLOCK(outer));
5971 	}
5972 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5973 }
5974 
5975 /*
5976  * The list of messages on the inner syncq is effectively hashed
5977  * by destination queue.  These destination queues are doubly
5978  * linked lists (hopefully) in priority order.  Messages are then
5979  * put on the queue referenced by the q_sqhead/q_sqtail elements.
5980  * Additional messages are linked together by the b_next/b_prev
5981  * elements in the mblk, with (similar to putq()) the first message
5982  * having a NULL b_prev and the last message having a NULL b_next.
5983  *
5984  * Events, such as qwriter callbacks, are put onto a list in FIFO
5985  * order referenced by sq_evhead, and sq_evtail.  This is a singly
5986  * linked list, and messages here MUST be processed in the order queued.
5987  */
5988 
5989 /*
5990  * Run the events on the syncq event list (sq_evhead).
5991  * Assumes there is only one claim on the syncq, it is
5992  * already exclusive (SQ_EXCL set), and the SQLOCK held.
5993  * Messages here are processed in order, with the SQ_EXCL bit
5994  * held all the way through till the last message is processed.
5995  */
5996 void
5997 sq_run_events(syncq_t *sq)
5998 {
5999 	mblk_t		*bp;
6000 	queue_t		*qp;
6001 	uint16_t	flags = sq->sq_flags;
6002 	void		(*func)();
6003 
6004 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6005 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6006 		sq->sq_oprev == NULL) ||
6007 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6008 		sq->sq_oprev != NULL));
6009 
6010 	ASSERT(flags & SQ_EXCL);
6011 	ASSERT(sq->sq_count == 1);
6012 
6013 	/*
6014 	 * We need to process all of the events on this list.  It
6015 	 * is possible that new events will be added while we are
6016 	 * away processing a callback, so on every loop, we start
6017 	 * back at the beginning of the list.
6018 	 */
6019 	/*
6020 	 * We have to reaccess sq_evhead since there is a
6021 	 * possibility of a new entry while we were running
6022 	 * the callback.
6023 	 */
6024 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6025 		ASSERT(bp->b_queue->q_syncq == sq);
6026 		ASSERT(sq->sq_flags & SQ_EVENTS);
6027 
6028 		qp = bp->b_queue;
6029 		func = (void (*)())bp->b_prev;
6030 		ASSERT(func != NULL);
6031 
6032 		/*
6033 		 * Messages from the event queue must be taken off in
6034 		 * FIFO order.
6035 		 */
6036 		ASSERT(sq->sq_evhead == bp);
6037 		sq->sq_evhead = bp->b_next;
6038 
6039 		if (bp->b_next == NULL) {
6040 			/* Deleting last */
6041 			ASSERT(sq->sq_evtail == bp);
6042 			sq->sq_evtail = NULL;
6043 			sq->sq_flags &= ~SQ_EVENTS;
6044 		}
6045 		bp->b_prev = bp->b_next = NULL;
6046 		ASSERT(bp->b_datap->db_ref != 0);
6047 
6048 		mutex_exit(SQLOCK(sq));
6049 
6050 		(*func)(qp, bp);
6051 
6052 		mutex_enter(SQLOCK(sq));
6053 		/*
6054 		 * re-read the flags, since they could have changed.
6055 		 */
6056 		flags = sq->sq_flags;
6057 		ASSERT(flags & SQ_EXCL);
6058 	}
6059 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6060 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6061 
6062 	if (flags & SQ_WANTWAKEUP) {
6063 		flags &= ~SQ_WANTWAKEUP;
6064 		cv_broadcast(&sq->sq_wait);
6065 	}
6066 	if (flags & SQ_WANTEXWAKEUP) {
6067 		flags &= ~SQ_WANTEXWAKEUP;
6068 		cv_broadcast(&sq->sq_exitwait);
6069 	}
6070 	sq->sq_flags = flags;
6071 }
6072 
6073 /*
6074  * Put messages on the event list.
6075  * If we can go exclusive now, do so and process the event list, otherwise
6076  * let the last claim service this list (or wake the sqthread).
6077  * This procedure assumes SQLOCK is held.  To run the event list, it
6078  * must be called with no claims.
6079  */
6080 static void
6081 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6082 {
6083 	uint16_t count;
6084 
6085 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6086 	ASSERT(func != NULL);
6087 
6088 	/*
6089 	 * This is a callback.  Add it to the list of callbacks
6090 	 * and see about upgrading.
6091 	 */
6092 	mp->b_prev = (mblk_t *)func;
6093 	mp->b_queue = q;
6094 	mp->b_next = NULL;
6095 	if (sq->sq_evhead == NULL) {
6096 		sq->sq_evhead = sq->sq_evtail = mp;
6097 		sq->sq_flags |= SQ_EVENTS;
6098 	} else {
6099 		ASSERT(sq->sq_evtail != NULL);
6100 		ASSERT(sq->sq_evtail->b_next == NULL);
6101 		ASSERT(sq->sq_flags & SQ_EVENTS);
6102 		sq->sq_evtail->b_next = mp;
6103 		sq->sq_evtail = mp;
6104 	}
6105 	/*
6106 	 * We have set SQ_EVENTS, so threads will have to
6107 	 * unwind out of the perimiter, and new entries will
6108 	 * not grab a putlock.  But we still need to know
6109 	 * how many threads have already made a claim to the
6110 	 * syncq, so grab the putlocks, and sum the counts.
6111 	 * If there are no claims on the syncq, we can upgrade
6112 	 * to exclusive, and run the event list.
6113 	 * NOTE: We hold the SQLOCK, so we can just grab the
6114 	 * putlocks.
6115 	 */
6116 	count = sq->sq_count;
6117 	SQ_PUTLOCKS_ENTER(sq);
6118 	SUM_SQ_PUTCOUNTS(sq, count);
6119 	/*
6120 	 * We have no claim, so we need to check if there
6121 	 * are no others, then we can upgrade.
6122 	 */
6123 	/*
6124 	 * There are currently no claims on
6125 	 * the syncq by this thread (at least on this entry). The thread who has
6126 	 * the claim should drain syncq.
6127 	 */
6128 	if (count > 0) {
6129 		/*
6130 		 * Can't upgrade - other threads inside.
6131 		 */
6132 		SQ_PUTLOCKS_EXIT(sq);
6133 		mutex_exit(SQLOCK(sq));
6134 		return;
6135 	}
6136 	/*
6137 	 * Need to set SQ_EXCL and make a claim on the syncq.
6138 	 */
6139 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6140 	sq->sq_flags |= SQ_EXCL;
6141 	ASSERT(sq->sq_count == 0);
6142 	sq->sq_count++;
6143 	SQ_PUTLOCKS_EXIT(sq);
6144 
6145 	/* Process the events list */
6146 	sq_run_events(sq);
6147 
6148 	/*
6149 	 * Release our claim...
6150 	 */
6151 	sq->sq_count--;
6152 
6153 	/*
6154 	 * And release SQ_EXCL.
6155 	 * We don't need to acquire the putlocks to release
6156 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6157 	 */
6158 	sq->sq_flags &= ~SQ_EXCL;
6159 
6160 	/*
6161 	 * sq_run_events should have released SQ_EXCL
6162 	 */
6163 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6164 
6165 	/*
6166 	 * If anything happened while we were running the
6167 	 * events (or was there before), we need to process
6168 	 * them now.  We shouldn't be exclusive sine we
6169 	 * released the perimiter above (plus, we asserted
6170 	 * for it).
6171 	 */
6172 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6173 		drain_syncq(sq);
6174 	else
6175 		mutex_exit(SQLOCK(sq));
6176 }
6177 
6178 /*
6179  * Perform delayed processing. The caller has to make sure that it is safe
6180  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6181  * set.)
6182  *
6183  * Assume that the caller has NO claims on the syncq.  However, a claim
6184  * on the syncq does not indicate that a thread is draining the syncq.
6185  * There may be more claims on the syncq than there are threads draining
6186  * (i.e.  #_threads_draining <= sq_count)
6187  *
6188  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6189  * in order to preserve qwriter(OUTER) ordering constraints.
6190  *
6191  * sq_putcount only needs to be checked when dispatching the queued
6192  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6193  */
6194 void
6195 drain_syncq(syncq_t *sq)
6196 {
6197 	queue_t		*qp;
6198 	uint16_t	count;
6199 	uint16_t	type = sq->sq_type;
6200 	uint16_t	flags = sq->sq_flags;
6201 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6202 
6203 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6204 		"drain_syncq start:%p", sq);
6205 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6206 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6207 		sq->sq_oprev == NULL) ||
6208 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6209 		sq->sq_oprev != NULL));
6210 
6211 	/*
6212 	 * Drop SQ_SERVICE flag.
6213 	 */
6214 	if (bg_service)
6215 		sq->sq_svcflags &= ~SQ_SERVICE;
6216 
6217 	/*
6218 	 * If SQ_EXCL is set, someone else is processing this syncq - let him
6219 	 * finish the job.
6220 	 */
6221 	if (flags & SQ_EXCL) {
6222 		if (bg_service) {
6223 			ASSERT(sq->sq_servcount != 0);
6224 			sq->sq_servcount--;
6225 		}
6226 		mutex_exit(SQLOCK(sq));
6227 		return;
6228 	}
6229 
6230 	/*
6231 	 * This routine can be called by a background thread if
6232 	 * it was scheduled by a hi-priority thread.  SO, if there are
6233 	 * NOT messages queued, return (remember, we have the SQLOCK,
6234 	 * and it cannot change until we release it). Wakeup any waiters also.
6235 	 */
6236 	if (!(flags & SQ_QUEUED)) {
6237 		if (flags & SQ_WANTWAKEUP) {
6238 			flags &= ~SQ_WANTWAKEUP;
6239 			cv_broadcast(&sq->sq_wait);
6240 		}
6241 		if (flags & SQ_WANTEXWAKEUP) {
6242 			flags &= ~SQ_WANTEXWAKEUP;
6243 			cv_broadcast(&sq->sq_exitwait);
6244 		}
6245 		sq->sq_flags = flags;
6246 		if (bg_service) {
6247 			ASSERT(sq->sq_servcount != 0);
6248 			sq->sq_servcount--;
6249 		}
6250 		mutex_exit(SQLOCK(sq));
6251 		return;
6252 	}
6253 
6254 	/*
6255 	 * If this is not a concurrent put perimiter, we need to
6256 	 * become exclusive to drain.  Also, if not CIPUT, we would
6257 	 * not have acquired a putlock, so we don't need to check
6258 	 * the putcounts.  If not entering with a claim, we test
6259 	 * for sq_count == 0.
6260 	 */
6261 	type = sq->sq_type;
6262 	if (!(type & SQ_CIPUT)) {
6263 		if (sq->sq_count > 1) {
6264 			if (bg_service) {
6265 				ASSERT(sq->sq_servcount != 0);
6266 				sq->sq_servcount--;
6267 			}
6268 			mutex_exit(SQLOCK(sq));
6269 			return;
6270 		}
6271 		sq->sq_flags |= SQ_EXCL;
6272 	}
6273 
6274 	/*
6275 	 * This is where we make a claim to the syncq.
6276 	 * This can either be done by incrementing a putlock, or
6277 	 * the sq_count.  But since we already have the SQLOCK
6278 	 * here, we just bump the sq_count.
6279 	 *
6280 	 * Note that after we make a claim, we need to let the code
6281 	 * fall through to the end of this routine to clean itself
6282 	 * up.  A return in the while loop will put the syncq in a
6283 	 * very bad state.
6284 	 */
6285 	sq->sq_count++;
6286 	ASSERT(sq->sq_count != 0);	/* wraparound */
6287 
6288 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6289 		/*
6290 		 * If we are told to stayaway or went exclusive,
6291 		 * we are done.
6292 		 */
6293 		if (flags & (SQ_STAYAWAY)) {
6294 			break;
6295 		}
6296 
6297 		/*
6298 		 * If there are events to run, do so.
6299 		 * We have one claim to the syncq, so if there are
6300 		 * more than one, other threads are running.
6301 		 */
6302 		if (sq->sq_evhead != NULL) {
6303 			ASSERT(sq->sq_flags & SQ_EVENTS);
6304 
6305 			count = sq->sq_count;
6306 			SQ_PUTLOCKS_ENTER(sq);
6307 			SUM_SQ_PUTCOUNTS(sq, count);
6308 			if (count > 1) {
6309 				SQ_PUTLOCKS_EXIT(sq);
6310 				/* Can't upgrade - other threads inside */
6311 				break;
6312 			}
6313 			ASSERT((flags & SQ_EXCL) == 0);
6314 			sq->sq_flags = flags | SQ_EXCL;
6315 			SQ_PUTLOCKS_EXIT(sq);
6316 			/*
6317 			 * we have the only claim, run the events,
6318 			 * sq_run_events will clear the SQ_EXCL flag.
6319 			 */
6320 			sq_run_events(sq);
6321 
6322 			/*
6323 			 * If this is a CIPUT perimiter, we need
6324 			 * to drop the SQ_EXCL flag so we can properly
6325 			 * continue draining the syncq.
6326 			 */
6327 			if (type & SQ_CIPUT) {
6328 				ASSERT(sq->sq_flags & SQ_EXCL);
6329 				sq->sq_flags &= ~SQ_EXCL;
6330 			}
6331 
6332 			/*
6333 			 * And go back to the beginning just in case
6334 			 * anything changed while we were away.
6335 			 */
6336 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6337 			continue;
6338 		}
6339 
6340 		ASSERT(sq->sq_evhead == NULL);
6341 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6342 
6343 		/*
6344 		 * Find the queue that is not draining.
6345 		 *
6346 		 * q_draining is protected by QLOCK which we do not hold.
6347 		 * But if it was set, then a thread was draining, and if it gets
6348 		 * cleared, then it was because the thread has successfully
6349 		 * drained the syncq, or a GOAWAY state occured. For the GOAWAY
6350 		 * state to happen, a thread needs the SQLOCK which we hold, and
6351 		 * if there was such a flag, we whould have already seen it.
6352 		 */
6353 
6354 		for (qp = sq->sq_head;
6355 		    qp != NULL && (qp->q_draining ||
6356 			(qp->q_sqflags & Q_SQDRAINING));
6357 		    qp = qp->q_sqnext)
6358 			;
6359 
6360 		if (qp == NULL)
6361 			break;
6362 
6363 		/*
6364 		 * We have a queue to work on, and we hold the
6365 		 * SQLOCK and one claim, call qdrain_syncq.
6366 		 * This means we need to release the SQLOCK and
6367 		 * aquire the QLOCK (OK since we have a claim).
6368 		 * Note that qdrain_syncq will actually dequeue
6369 		 * this queue from the sq_head list when it is
6370 		 * convinced all the work is done and release
6371 		 * the QLOCK before returning.
6372 		 */
6373 		qp->q_sqflags |= Q_SQDRAINING;
6374 		mutex_exit(SQLOCK(sq));
6375 		mutex_enter(QLOCK(qp));
6376 		qdrain_syncq(sq, qp);
6377 		mutex_enter(SQLOCK(sq));
6378 
6379 		/* The queue is drained */
6380 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6381 		qp->q_sqflags &= ~Q_SQDRAINING;
6382 		/*
6383 		 * NOTE: After this point qp should not be used since it may be
6384 		 * closed.
6385 		 */
6386 	}
6387 
6388 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6389 	flags = sq->sq_flags;
6390 
6391 	/*
6392 	 * sq->sq_head cannot change because we hold the
6393 	 * sqlock. However, a thread CAN decide that it is no longer
6394 	 * going to drain that queue.  However, this should be due to
6395 	 * a GOAWAY state, and we should see that here.
6396 	 *
6397 	 * This loop is not very efficient. One solution may be adding a second
6398 	 * pointer to the "draining" queue, but it is difficult to do when
6399 	 * queues are inserted in the middle due to priority ordering. Another
6400 	 * possibility is to yank the queue out of the sq list and put it onto
6401 	 * the "draining list" and then put it back if it can't be drained.
6402 	 */
6403 
6404 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6405 		(type & SQ_CI) || sq->sq_head->q_draining);
6406 
6407 	/* Drop SQ_EXCL for non-CIPUT perimiters */
6408 	if (!(type & SQ_CIPUT))
6409 		flags &= ~SQ_EXCL;
6410 	ASSERT((flags & SQ_EXCL) == 0);
6411 
6412 	/* Wake up any waiters. */
6413 	if (flags & SQ_WANTWAKEUP) {
6414 		flags &= ~SQ_WANTWAKEUP;
6415 		cv_broadcast(&sq->sq_wait);
6416 	}
6417 	if (flags & SQ_WANTEXWAKEUP) {
6418 		flags &= ~SQ_WANTEXWAKEUP;
6419 		cv_broadcast(&sq->sq_exitwait);
6420 	}
6421 	sq->sq_flags = flags;
6422 
6423 	ASSERT(sq->sq_count != 0);
6424 	/* Release our claim. */
6425 	sq->sq_count--;
6426 
6427 	if (bg_service) {
6428 		ASSERT(sq->sq_servcount != 0);
6429 		sq->sq_servcount--;
6430 	}
6431 
6432 	mutex_exit(SQLOCK(sq));
6433 
6434 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6435 		"drain_syncq end:%p", sq);
6436 }
6437 
6438 
6439 /*
6440  *
6441  * qdrain_syncq can be called (currently) from only one of two places:
6442  *	drain_syncq
6443  * 	putnext  (or some variation of it).
6444  * and eventually
6445  * 	qwait(_sig)
6446  *
6447  * If called from drain_syncq, we found it in the list
6448  * of queue's needing service, so there is work to be done (or it
6449  * wouldn't be on the list).
6450  *
6451  * If called from some putnext variation, it was because the
6452  * perimiter is open, but messages are blocking a putnext and
6453  * there is not a thread working on it.  Now a thread could start
6454  * working on it while we are getting ready to do so ourself, but
6455  * the thread would set the q_draining flag, and we can spin out.
6456  *
6457  * As for qwait(_sig), I think I shall let it continue to call
6458  * drain_syncq directly (after all, it will get here eventually).
6459  *
6460  * qdrain_syncq has to terminate when:
6461  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6462  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6463  *
6464  * ASSUMES:
6465  *	One claim
6466  * 	QLOCK held
6467  * 	SQLOCK not held
6468  *	Will release QLOCK before returning
6469  */
6470 void
6471 qdrain_syncq(syncq_t *sq, queue_t *q)
6472 {
6473 	mblk_t		*bp;
6474 	boolean_t	do_clr;
6475 #ifdef DEBUG
6476 	uint16_t	count;
6477 #endif
6478 
6479 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6480 		"drain_syncq start:%p", sq);
6481 	ASSERT(q->q_syncq == sq);
6482 	ASSERT(MUTEX_HELD(QLOCK(q)));
6483 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6484 	/*
6485 	 * For non-CIPUT perimiters, we should be called with the
6486 	 * exclusive bit set already.  For non-CIPUT perimiters we
6487 	 * will be doing a concurrent drain, so it better not be set.
6488 	 */
6489 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6490 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6491 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6492 	/*
6493 	 * All outer pointers are set, or none of them are
6494 	 */
6495 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6496 		sq->sq_oprev == NULL) ||
6497 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6498 		sq->sq_oprev != NULL));
6499 #ifdef DEBUG
6500 	count = sq->sq_count;
6501 	/*
6502 	 * This is OK without the putlocks, because we have one
6503 	 * claim either from the sq_count, or a putcount.  We could
6504 	 * get an erroneous value from other counts, but ours won't
6505 	 * change, so one way or another, we will have at least a
6506 	 * value of one.
6507 	 */
6508 	SUM_SQ_PUTCOUNTS(sq, count);
6509 	ASSERT(count >= 1);
6510 #endif /* DEBUG */
6511 
6512 	/*
6513 	 * The first thing to do here, is find out if a thread is already
6514 	 * draining this queue or the queue is closing. If so, we are done,
6515 	 * just return. Also, if there are no messages, we are done as well.
6516 	 * Note that we check the q_sqhead since there is s window of
6517 	 * opportunity for us to enter here because Q_SQQUEUED was set, but is
6518 	 * not anymore.
6519 	 */
6520 	if (q->q_draining || (q->q_sqhead == NULL)) {
6521 		mutex_exit(QLOCK(q));
6522 		return;
6523 	}
6524 
6525 	/*
6526 	 * If the perimiter is exclusive, there is nothing we can
6527 	 * do right now, go away.
6528 	 * Note that there is nothing to prevent this case from changing
6529 	 * right after this check, but the spin-out will catch it.
6530 	 */
6531 
6532 	/* Tell other threads that we are draining this queue */
6533 	q->q_draining = 1;	/* Protected by QLOCK */
6534 
6535 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6536 
6537 		/*
6538 		 * Because we can enter this routine just because
6539 		 * a putnext is blocked, we need to spin out if
6540 		 * the perimiter wants to go exclusive as well
6541 		 * as just blocked. We need to spin out also if
6542 		 * events are queued on the syncq.
6543 		 * Don't check for SQ_EXCL, because non-CIPUT
6544 		 * perimiters would set it, and it can't become
6545 		 * exclusive while we hold a claim.
6546 		 */
6547 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6548 			break;
6549 		}
6550 
6551 #ifdef DEBUG
6552 		/*
6553 		 * Since we are in qdrain_syncq, we already know the queue,
6554 		 * but for sanity, we want to check this against the qp that
6555 		 * was passed in by bp->b_queue.
6556 		 */
6557 
6558 		ASSERT(bp->b_queue == q);
6559 		ASSERT(bp->b_queue->q_syncq == sq);
6560 		bp->b_queue = NULL;
6561 
6562 		/*
6563 		 * We would have the following check in the DEBUG code:
6564 		 *
6565 		 * if (bp->b_prev != NULL)  {
6566 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6567 		 * }
6568 		 *
6569 		 * This can't be done, however, since IP modifies qinfo
6570 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6571 		 * qinfo), invalidating the check.
6572 		 * So the assignment to func is left here, but the ASSERT itself
6573 		 * is removed until the whole issue is resolved.
6574 		 */
6575 #endif
6576 		ASSERT(q->q_sqhead == bp);
6577 		q->q_sqhead = bp->b_next;
6578 		bp->b_prev = bp->b_next = NULL;
6579 		ASSERT(q->q_syncqmsgs > 0);
6580 		mutex_exit(QLOCK(q));
6581 
6582 		ASSERT(bp->b_datap->db_ref != 0);
6583 
6584 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6585 
6586 		mutex_enter(QLOCK(q));
6587 		/*
6588 		 * We should decrement q_syncqmsgs only after executing the
6589 		 * put procedure to avoid a possible race with putnext().
6590 		 * In putnext() though it sees Q_SQQUEUED is set, there is
6591 		 * an optimization which allows putnext to call the put
6592 		 * procedure directly if (q_syncqmsgs == 0) and thus
6593 		 * a message reodering could otherwise occur.
6594 		 */
6595 		q->q_syncqmsgs--;
6596 
6597 		/*
6598 		 * Clear QFULL in the next service procedure queue if
6599 		 * this is the last message destined to that queue.
6600 		 *
6601 		 * It would make better sense to have some sort of
6602 		 * tunable for the low water mark, but these symantics
6603 		 * are not yet defined.  So, alas, we use a constant.
6604 		 */
6605 		do_clr = (q->q_syncqmsgs == 0);
6606 		mutex_exit(QLOCK(q));
6607 
6608 		if (do_clr)
6609 			clr_qfull(q);
6610 
6611 		mutex_enter(QLOCK(q));
6612 		/*
6613 		 * Always clear SQ_EXCL when CIPUT in order to handle
6614 		 * qwriter(INNER).
6615 		 */
6616 		/*
6617 		 * The putp() can call qwriter and get exclusive access
6618 		 * IFF this is the only claim.  So, we need to test for
6619 		 * this possibility so we can aquire the mutex and clear
6620 		 * the bit.
6621 		 */
6622 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6623 			mutex_enter(SQLOCK(sq));
6624 			sq->sq_flags &= ~SQ_EXCL;
6625 			mutex_exit(SQLOCK(sq));
6626 		}
6627 	}
6628 
6629 	/*
6630 	 * We should either have no queues on the syncq, or we were
6631 	 * told to goaway by a waiter (which we will wake up at the
6632 	 * end of this function).
6633 	 */
6634 	ASSERT((q->q_sqhead == NULL) ||
6635 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6636 
6637 	ASSERT(MUTEX_HELD(QLOCK(q)));
6638 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6639 
6640 	/*
6641 	 * Remove the q from the syncq list if all the messages are
6642 	 * drained.
6643 	 */
6644 	if (q->q_sqhead == NULL) {
6645 		mutex_enter(SQLOCK(sq));
6646 		if (q->q_sqflags & Q_SQQUEUED)
6647 			SQRM_Q(sq, q);
6648 		mutex_exit(SQLOCK(sq));
6649 		/*
6650 		 * Since the queue is removed from the list, reset its priority.
6651 		 */
6652 		q->q_spri = 0;
6653 	}
6654 
6655 	/*
6656 	 * Remember, the q_draining flag is used to let another
6657 	 * thread know that there is a thread currently draining
6658 	 * the messages for a queue.  Since we are now done with
6659 	 * this queue (even if there may be messages still there),
6660 	 * we need to clear this flag so some thread will work
6661 	 * on it if needed.
6662 	 */
6663 	ASSERT(q->q_draining);
6664 	q->q_draining = 0;
6665 
6666 	/* called with a claim, so OK to drop all locks. */
6667 	mutex_exit(QLOCK(q));
6668 
6669 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6670 		"drain_syncq end:%p", sq);
6671 }
6672 /* END OF QDRAIN_SYNCQ  */
6673 
6674 
6675 /*
6676  * This is the mate to qdrain_syncq, except that it is putting the
6677  * message onto the the queue instead draining.  Since the
6678  * message is destined for the queue that is selected, there is
6679  * no need to identify the function because the message is
6680  * intended for the put routine for the queue.  But this
6681  * routine will do it anyway just in case (but only for debug kernels).
6682  *
6683  * After the message is enqueued on the syncq, it calls putnext_tail()
6684  * which will schedule a background thread to actually process the message.
6685  *
6686  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6687  * SQLOCK(sq) and QLOCK(q) are not held.
6688  */
6689 void
6690 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6691 {
6692 	queue_t		*fq = NULL;
6693 
6694 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6695 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6696 	ASSERT(sq->sq_count > 0);
6697 	ASSERT(q->q_syncq == sq);
6698 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6699 		sq->sq_oprev == NULL) ||
6700 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6701 		sq->sq_oprev != NULL));
6702 
6703 	mutex_enter(QLOCK(q));
6704 
6705 	/*
6706 	 * Set QFULL in next service procedure queue (that cares) if not
6707 	 * already set and if there are already more messages on the syncq
6708 	 * than sq_max_size.  If sq_max_size is 0, no flow control will be
6709 	 * asserted on any syncq.
6710 	 *
6711 	 * The fq here is the next queue with a service procedure.
6712 	 * This is where we would fail canputnext, so this is where we
6713 	 * need to set QFULL.
6714 	 *
6715 	 * LOCKING HIERARCHY: In the case when fq != q we need to
6716 	 *  a) Take QLOCK(fq) to set QFULL flag and
6717 	 *  b) Take sd_reflock in the case of the hot stream to update
6718 	 *  	sd_refcnt.
6719 	 * We already have QLOCK at this point. To avoid cross-locks with
6720 	 * freezestr() which grabs all QLOCKs and with strlock() which grabs
6721 	 * both SQLOCK and sd_reflock, we need to drop respective locks first.
6722 	 */
6723 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
6724 	    (q->q_syncqmsgs > sq_max_size)) {
6725 		if ((fq = q->q_nfsrv) == q) {
6726 			fq->q_flag |= QFULL;
6727 		} else {
6728 			mutex_exit(QLOCK(q));
6729 			mutex_enter(QLOCK(fq));
6730 			fq->q_flag |= QFULL;
6731 			mutex_exit(QLOCK(fq));
6732 			mutex_enter(QLOCK(q));
6733 		}
6734 	}
6735 
6736 #ifdef DEBUG
6737 	/*
6738 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6739 	 * to trace the queue that the message is intended for.  Note
6740 	 * that the original use was to identify the queue and function
6741 	 * to call on the drain.  In the new syncq, we have the context
6742 	 * of the queue that we are draining, so call it's putproc and
6743 	 * don't rely on the saved values.  But for debug this is still
6744 	 * usefull information.
6745 	 */
6746 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6747 	mp->b_queue = q;
6748 	mp->b_next = NULL;
6749 #endif
6750 	ASSERT(q->q_syncq == sq);
6751 	/*
6752 	 * Enqueue the message on the list.
6753 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6754 	 * protect it.  So its ok to acquire SQLOCK after SQPUT_MP().
6755 	 */
6756 	SQPUT_MP(q, mp);
6757 	mutex_enter(SQLOCK(sq));
6758 
6759 	/*
6760 	 * And queue on syncq for scheduling, if not already queued.
6761 	 * Note that we need the SQLOCK for this, and for testing flags
6762 	 * at the end to see if we will drain.  So grab it now, and
6763 	 * release it before we call qdrain_syncq or return.
6764 	 */
6765 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6766 		q->q_spri = curthread->t_pri;
6767 		SQPUT_Q(sq, q);
6768 	}
6769 #ifdef DEBUG
6770 	else {
6771 		/*
6772 		 * All of these conditions MUST be true!
6773 		 */
6774 		ASSERT(sq->sq_tail != NULL);
6775 		if (sq->sq_tail == sq->sq_head) {
6776 			ASSERT((q->q_sqprev == NULL) &&
6777 			    (q->q_sqnext == NULL));
6778 		} else {
6779 			ASSERT((q->q_sqprev != NULL) ||
6780 			    (q->q_sqnext != NULL));
6781 		}
6782 		ASSERT(sq->sq_flags & SQ_QUEUED);
6783 		ASSERT(q->q_syncqmsgs != 0);
6784 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6785 	}
6786 #endif
6787 	mutex_exit(QLOCK(q));
6788 	/*
6789 	 * SQLOCK is still held, so sq_count can be safely decremented.
6790 	 */
6791 	sq->sq_count--;
6792 
6793 	putnext_tail(sq, q, 0);
6794 	/* Should not reference sq or q after this point. */
6795 }
6796 
6797 /*  End of qfill_syncq  */
6798 
6799 /*
6800  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6801  * that would be put into qp by drain_syncq.
6802  * Used when deleting the syncq (qp == NULL) or when detaching
6803  * a queue (qp != NULL).
6804  * Return non-zero if one or more messages were freed.
6805  *
6806  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
6807  * sq_putlocks are used.
6808  *
6809  * NOTE: This function assumes that it is called from the close() context and
6810  * that all the queues in the syncq are going aay. For this reason it doesn't
6811  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6812  * currently valid, but it is useful to rethink this function to behave properly
6813  * in other cases.
6814  */
6815 int
6816 flush_syncq(syncq_t *sq, queue_t *qp)
6817 {
6818 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6819 	queue_t		*q;
6820 	int		ret = 0;
6821 
6822 	mutex_enter(SQLOCK(sq));
6823 
6824 	/*
6825 	 * Before we leave, we need to make sure there are no
6826 	 * events listed for this queue.  All events for this queue
6827 	 * will just be freed.
6828 	 */
6829 	if (qp != NULL && sq->sq_evhead != NULL) {
6830 		ASSERT(sq->sq_flags & SQ_EVENTS);
6831 
6832 		mp_prev = NULL;
6833 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6834 			mp_next = bp->b_next;
6835 			if (bp->b_queue == qp) {
6836 				/* Delete this message */
6837 				if (mp_prev != NULL) {
6838 					mp_prev->b_next = mp_next;
6839 					/*
6840 					 * Update sq_evtail if the last element
6841 					 * is removed.
6842 					 */
6843 					if (bp == sq->sq_evtail) {
6844 						ASSERT(mp_next == NULL);
6845 						sq->sq_evtail = mp_prev;
6846 					}
6847 				} else
6848 					sq->sq_evhead = mp_next;
6849 				if (sq->sq_evhead == NULL)
6850 					sq->sq_flags &= ~SQ_EVENTS;
6851 				bp->b_prev = bp->b_next = NULL;
6852 				freemsg(bp);
6853 				ret++;
6854 			} else {
6855 				mp_prev = bp;
6856 			}
6857 		}
6858 	}
6859 
6860 	/*
6861 	 * Walk sq_head and:
6862 	 *	- match qp if qp is set, remove it's messages
6863 	 *	- all if qp is not set
6864 	 */
6865 	q = sq->sq_head;
6866 	while (q != NULL) {
6867 		ASSERT(q->q_syncq == sq);
6868 		if ((qp == NULL) || (qp == q)) {
6869 			/*
6870 			 * Yank the messages as a list off the queue
6871 			 */
6872 			mp_head = q->q_sqhead;
6873 			/*
6874 			 * We do not have QLOCK(q) here (which is safe due to
6875 			 * assumptions mentioned above). To obtain the lock we
6876 			 * need to release SQLOCK which may allow lots of things
6877 			 * to change upon us. This place requires more analysis.
6878 			 */
6879 			q->q_sqhead = q->q_sqtail = NULL;
6880 			ASSERT(mp_head->b_queue &&
6881 			    mp_head->b_queue->q_syncq == sq);
6882 
6883 			/*
6884 			 * Free each of the messages.
6885 			 */
6886 			for (bp = mp_head; bp != NULL; bp = mp_next) {
6887 				mp_next = bp->b_next;
6888 				bp->b_prev = bp->b_next = NULL;
6889 				freemsg(bp);
6890 				ret++;
6891 			}
6892 			/*
6893 			 * Now remove the queue from the syncq.
6894 			 */
6895 			ASSERT(q->q_sqflags & Q_SQQUEUED);
6896 			SQRM_Q(sq, q);
6897 			q->q_spri = 0;
6898 			q->q_syncqmsgs = 0;
6899 
6900 			/*
6901 			 * If qp was specified, we are done with it and are
6902 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
6903 			 * waiters while we still have the SQLOCK.
6904 			 */
6905 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
6906 				sq->sq_flags &= ~SQ_WANTWAKEUP;
6907 				cv_broadcast(&sq->sq_wait);
6908 			}
6909 			/* Drop SQLOCK across clr_qfull */
6910 			mutex_exit(SQLOCK(sq));
6911 
6912 			/*
6913 			 * We avoid doing the test that drain_syncq does and
6914 			 * unconditionally clear qfull for every flushed
6915 			 * message. Since flush_syncq is only called during
6916 			 * close this should not be a problem.
6917 			 */
6918 			clr_qfull(q);
6919 			if (qp != NULL) {
6920 				return (ret);
6921 			} else {
6922 				mutex_enter(SQLOCK(sq));
6923 				/*
6924 				 * The head was removed by SQRM_Q above.
6925 				 * reread the new head and flush it.
6926 				 */
6927 				q = sq->sq_head;
6928 			}
6929 		} else {
6930 			q = q->q_sqnext;
6931 		}
6932 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
6933 	}
6934 
6935 	if (sq->sq_flags & SQ_WANTWAKEUP) {
6936 		sq->sq_flags &= ~SQ_WANTWAKEUP;
6937 		cv_broadcast(&sq->sq_wait);
6938 	}
6939 
6940 	mutex_exit(SQLOCK(sq));
6941 	return (ret);
6942 }
6943 
6944 /*
6945  * Propagate all messages from a syncq to the next syncq that are associated
6946  * with the specified queue. If the queue is attached to a driver or if the
6947  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
6948  *
6949  * Assumes that the stream is strlock()'ed. We don't come here if there
6950  * are no messages to propagate.
6951  *
6952  * NOTE : If the queue is attached to a driver, all the messages are freed
6953  * as there is no point in propagating the messages from the driver syncq
6954  * to the closing stream head which will in turn get freed later.
6955  */
6956 static int
6957 propagate_syncq(queue_t *qp)
6958 {
6959 	mblk_t		*bp, *head, *tail, *prev, *next;
6960 	syncq_t 	*sq;
6961 	queue_t		*nqp;
6962 	syncq_t		*nsq;
6963 	boolean_t	isdriver;
6964 	int 		moved = 0;
6965 	uint16_t	flags;
6966 	pri_t		priority = curthread->t_pri;
6967 #ifdef DEBUG
6968 	void		(*func)();
6969 #endif
6970 
6971 	sq = qp->q_syncq;
6972 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6973 	/* debug macro */
6974 	SQ_PUTLOCKS_HELD(sq);
6975 	/*
6976 	 * As entersq() does not increment the sq_count for
6977 	 * the write side, check sq_count for non-QPERQ
6978 	 * perimeters alone.
6979 	 */
6980 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
6981 
6982 	/*
6983 	 * propagate_syncq() can be called because of either messages on the
6984 	 * queue syncq or because on events on the queue syncq. Do actual
6985 	 * message propagations if there are any messages.
6986 	 */
6987 	if (qp->q_syncqmsgs) {
6988 		isdriver = (qp->q_flag & QISDRV);
6989 
6990 		if (!isdriver) {
6991 			nqp = qp->q_next;
6992 			nsq = nqp->q_syncq;
6993 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
6994 			/* debug macro */
6995 			SQ_PUTLOCKS_HELD(nsq);
6996 #ifdef DEBUG
6997 			func = (void (*)())nqp->q_qinfo->qi_putp;
6998 #endif
6999 		}
7000 
7001 		SQRM_Q(sq, qp);
7002 		priority = MAX(qp->q_spri, priority);
7003 		qp->q_spri = 0;
7004 		head = qp->q_sqhead;
7005 		tail = qp->q_sqtail;
7006 		qp->q_sqhead = qp->q_sqtail = NULL;
7007 		qp->q_syncqmsgs = 0;
7008 
7009 		/*
7010 		 * Walk the list of messages, and free them if this is a driver,
7011 		 * otherwise reset the b_prev and b_queue value to the new putp.
7012 		 * Afterward, we will just add the head to the end of the next
7013 		 * syncq, and point the tail to the end of this one.
7014 		 */
7015 
7016 		for (bp = head; bp != NULL; bp = next) {
7017 			next = bp->b_next;
7018 			if (isdriver) {
7019 				bp->b_prev = bp->b_next = NULL;
7020 				freemsg(bp);
7021 				continue;
7022 			}
7023 			/* Change the q values for this message */
7024 			bp->b_queue = nqp;
7025 #ifdef DEBUG
7026 			bp->b_prev = (mblk_t *)func;
7027 #endif
7028 			moved++;
7029 		}
7030 		/*
7031 		 * Attach list of messages to the end of the new queue (if there
7032 		 * is a list of messages).
7033 		 */
7034 
7035 		if (!isdriver && head != NULL) {
7036 			ASSERT(tail != NULL);
7037 			if (nqp->q_sqhead == NULL) {
7038 				nqp->q_sqhead = head;
7039 			} else {
7040 				ASSERT(nqp->q_sqtail != NULL);
7041 				nqp->q_sqtail->b_next = head;
7042 			}
7043 			nqp->q_sqtail = tail;
7044 			/*
7045 			 * When messages are moved from high priority queue to
7046 			 * another queue, the destination queue priority is
7047 			 * upgraded.
7048 			 */
7049 
7050 			if (priority > nqp->q_spri)
7051 				nqp->q_spri = priority;
7052 
7053 			SQPUT_Q(nsq, nqp);
7054 
7055 			nqp->q_syncqmsgs += moved;
7056 			ASSERT(nqp->q_syncqmsgs != 0);
7057 		}
7058 	}
7059 
7060 	/*
7061 	 * Before we leave, we need to make sure there are no
7062 	 * events listed for this queue.  All events for this queue
7063 	 * will just be freed.
7064 	 */
7065 	if (sq->sq_evhead != NULL) {
7066 		ASSERT(sq->sq_flags & SQ_EVENTS);
7067 		prev = NULL;
7068 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7069 			next = bp->b_next;
7070 			if (bp->b_queue == qp) {
7071 				/* Delete this message */
7072 				if (prev != NULL) {
7073 					prev->b_next = next;
7074 					/*
7075 					 * Update sq_evtail if the last element
7076 					 * is removed.
7077 					 */
7078 					if (bp == sq->sq_evtail) {
7079 						ASSERT(next == NULL);
7080 						sq->sq_evtail = prev;
7081 					}
7082 				} else
7083 					sq->sq_evhead = next;
7084 				if (sq->sq_evhead == NULL)
7085 					sq->sq_flags &= ~SQ_EVENTS;
7086 				bp->b_prev = bp->b_next = NULL;
7087 				freemsg(bp);
7088 			} else {
7089 				prev = bp;
7090 			}
7091 		}
7092 	}
7093 
7094 	flags = sq->sq_flags;
7095 
7096 	/* Wake up any waiter before leaving. */
7097 	if (flags & SQ_WANTWAKEUP) {
7098 		flags &= ~SQ_WANTWAKEUP;
7099 		cv_broadcast(&sq->sq_wait);
7100 	}
7101 	sq->sq_flags = flags;
7102 
7103 	return (moved);
7104 }
7105 
7106 /*
7107  * Try and upgrade to exclusive access at the inner perimeter. If this can
7108  * not be done without blocking then request will be queued on the syncq
7109  * and drain_syncq will run it later.
7110  *
7111  * This routine can only be called from put or service procedures plus
7112  * asynchronous callback routines that have properly entered to
7113  * queue (with entersq.) Thus qwriter_inner assumes the caller has one claim
7114  * on the syncq associated with q.
7115  */
7116 void
7117 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7118 {
7119 	syncq_t	*sq = q->q_syncq;
7120 	uint16_t count;
7121 
7122 	mutex_enter(SQLOCK(sq));
7123 	count = sq->sq_count;
7124 	SQ_PUTLOCKS_ENTER(sq);
7125 	SUM_SQ_PUTCOUNTS(sq, count);
7126 	ASSERT(count >= 1);
7127 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7128 
7129 	if (count == 1) {
7130 		/*
7131 		 * Can upgrade. This case also handles nested qwriter calls
7132 		 * (when the qwriter callback function calls qwriter). In that
7133 		 * case SQ_EXCL is already set.
7134 		 */
7135 		sq->sq_flags |= SQ_EXCL;
7136 		SQ_PUTLOCKS_EXIT(sq);
7137 		mutex_exit(SQLOCK(sq));
7138 		(*func)(q, mp);
7139 		/*
7140 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7141 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7142 		 * until putnext, leavesq, or drain_syncq drops it.
7143 		 * That way we handle nested qwriter(INNER) without dropping
7144 		 * SQ_EXCL until the outermost qwriter callback routine is
7145 		 * done.
7146 		 */
7147 		return;
7148 	}
7149 	SQ_PUTLOCKS_EXIT(sq);
7150 	sqfill_events(sq, q, mp, func);
7151 }
7152 
7153 /*
7154  * Synchronous callback support functions
7155  */
7156 
7157 /*
7158  * Allocate a callback parameter structure.
7159  * Assumes that caller initializes the flags and the id.
7160  * Acquires SQLOCK(sq) if non-NULL is returned.
7161  */
7162 callbparams_t *
7163 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7164 {
7165 	callbparams_t *cbp;
7166 	size_t size = sizeof (callbparams_t);
7167 
7168 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7169 
7170 	/*
7171 	 * Only try tryhard allocation if the caller is ready to panic.
7172 	 * Otherwise just fail.
7173 	 */
7174 	if (cbp == NULL) {
7175 		if (kmflags & KM_PANIC)
7176 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7177 			    &size, kmflags);
7178 		else
7179 			return (NULL);
7180 	}
7181 
7182 	ASSERT(size >= sizeof (callbparams_t));
7183 	cbp->cbp_size = size;
7184 	cbp->cbp_sq = sq;
7185 	cbp->cbp_func = func;
7186 	cbp->cbp_arg = arg;
7187 	mutex_enter(SQLOCK(sq));
7188 	cbp->cbp_next = sq->sq_callbpend;
7189 	sq->sq_callbpend = cbp;
7190 	return (cbp);
7191 }
7192 
7193 void
7194 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7195 {
7196 	callbparams_t **pp, *p;
7197 
7198 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7199 
7200 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7201 		if (p == cbp) {
7202 			*pp = p->cbp_next;
7203 			kmem_free(p, p->cbp_size);
7204 			return;
7205 		}
7206 	}
7207 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7208 	    "callbparams_free: not found\n"));
7209 }
7210 
7211 void
7212 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7213 {
7214 	callbparams_t **pp, *p;
7215 
7216 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7217 
7218 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7219 		if (p->cbp_id == id && p->cbp_flags == flag) {
7220 			*pp = p->cbp_next;
7221 			kmem_free(p, p->cbp_size);
7222 			return;
7223 		}
7224 	}
7225 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7226 	    "callbparams_free_id: not found\n"));
7227 }
7228 
7229 /*
7230  * Callback wrapper function used by once-only callbacks that can be
7231  * cancelled (qtimeout and qbufcall)
7232  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7233  * cancelled by the qun* functions.
7234  */
7235 void
7236 qcallbwrapper(void *arg)
7237 {
7238 	callbparams_t *cbp = arg;
7239 	syncq_t	*sq;
7240 	uint16_t count = 0;
7241 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7242 	uint16_t type;
7243 
7244 	sq = cbp->cbp_sq;
7245 	mutex_enter(SQLOCK(sq));
7246 	type = sq->sq_type;
7247 	if (!(type & SQ_CICB)) {
7248 		count = sq->sq_count;
7249 		SQ_PUTLOCKS_ENTER(sq);
7250 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7251 		SUM_SQ_PUTCOUNTS(sq, count);
7252 		sq->sq_needexcl++;
7253 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7254 		waitflags |= SQ_MESSAGES;
7255 	}
7256 	/* Can not handle exlusive entry at outer perimeter */
7257 	ASSERT(type & SQ_COCB);
7258 
7259 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7260 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7261 		    (sq->sq_cancelid == cbp->cbp_id)) {
7262 			/* timeout has been cancelled */
7263 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7264 			callbparams_free(sq, cbp);
7265 			if (!(type & SQ_CICB)) {
7266 				ASSERT(sq->sq_needexcl > 0);
7267 				sq->sq_needexcl--;
7268 				if (sq->sq_needexcl == 0) {
7269 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7270 				}
7271 				SQ_PUTLOCKS_EXIT(sq);
7272 			}
7273 			mutex_exit(SQLOCK(sq));
7274 			return;
7275 		}
7276 		sq->sq_flags |= SQ_WANTWAKEUP;
7277 		if (!(type & SQ_CICB)) {
7278 			SQ_PUTLOCKS_EXIT(sq);
7279 		}
7280 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7281 		if (!(type & SQ_CICB)) {
7282 			count = sq->sq_count;
7283 			SQ_PUTLOCKS_ENTER(sq);
7284 			SUM_SQ_PUTCOUNTS(sq, count);
7285 		}
7286 	}
7287 
7288 	sq->sq_count++;
7289 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7290 	if (!(type & SQ_CICB)) {
7291 		ASSERT(count == 0);
7292 		sq->sq_flags |= SQ_EXCL;
7293 		ASSERT(sq->sq_needexcl > 0);
7294 		sq->sq_needexcl--;
7295 		if (sq->sq_needexcl == 0) {
7296 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7297 		}
7298 		SQ_PUTLOCKS_EXIT(sq);
7299 	}
7300 
7301 	mutex_exit(SQLOCK(sq));
7302 
7303 	cbp->cbp_func(cbp->cbp_arg);
7304 
7305 	/*
7306 	 * We drop the lock only for leavesq to re-acquire it.
7307 	 * Possible optimization is inline of leavesq.
7308 	 */
7309 	mutex_enter(SQLOCK(sq));
7310 	callbparams_free(sq, cbp);
7311 	mutex_exit(SQLOCK(sq));
7312 	leavesq(sq, SQ_CALLBACK);
7313 }
7314 
7315 /*
7316  * no need to grab sq_putlocks here. See comment in strsubr.h that
7317  * explains when sq_putlocks are used.
7318  *
7319  * sq_count (or one of the sq_putcounts) has already been
7320  * decremented by the caller, and if SQ_QUEUED, we need to call
7321  * drain_syncq (the global syncq drain).
7322  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7323  * one of two states, non-CIPUT perimiter, and we need to clear
7324  * it, or we went exclusive in the put procedure.  In any case,
7325  * we want to clear the bit now, and it is probably easier to do
7326  * this at the beginning of this function (remember, we hold
7327  * the SQLOCK).  Lastly, if there are other messages queued
7328  * on the syncq (and not for our destination), enable the syncq
7329  * for background work.
7330  */
7331 
7332 /* ARGSUSED */
7333 void
7334 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7335 {
7336 	uint16_t	flags = sq->sq_flags;
7337 
7338 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7339 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7340 
7341 	/* Clear SQ_EXCL if set in passflags */
7342 	if (passflags & SQ_EXCL) {
7343 		flags &= ~SQ_EXCL;
7344 	}
7345 	if (flags & SQ_WANTWAKEUP) {
7346 		flags &= ~SQ_WANTWAKEUP;
7347 		cv_broadcast(&sq->sq_wait);
7348 	}
7349 	if (flags & SQ_WANTEXWAKEUP) {
7350 		flags &= ~SQ_WANTEXWAKEUP;
7351 		cv_broadcast(&sq->sq_exitwait);
7352 	}
7353 	sq->sq_flags = flags;
7354 
7355 	/*
7356 	 * We have cleared SQ_EXCL if we were asked to, and started
7357 	 * the wakeup process for waiters.  If there are no writers
7358 	 * then we need to drain the syncq if we were told to, or
7359 	 * enable the background thread to do it.
7360 	 */
7361 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7362 		if ((passflags & SQ_QUEUED) ||
7363 		    (sq->sq_svcflags & SQ_DISABLED)) {
7364 			/* drain_syncq will take care of events in the list */
7365 			drain_syncq(sq);
7366 			return;
7367 		} else if (flags & SQ_QUEUED) {
7368 			sqenable(sq);
7369 		}
7370 	}
7371 	/* Drop the SQLOCK on exit */
7372 	mutex_exit(SQLOCK(sq));
7373 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7374 		"putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7375 }
7376 
7377 void
7378 set_qend(queue_t *q)
7379 {
7380 	mutex_enter(QLOCK(q));
7381 	if (!O_SAMESTR(q))
7382 		q->q_flag |= QEND;
7383 	else
7384 		q->q_flag &= ~QEND;
7385 	mutex_exit(QLOCK(q));
7386 	q = _OTHERQ(q);
7387 	mutex_enter(QLOCK(q));
7388 	if (!O_SAMESTR(q))
7389 		q->q_flag |= QEND;
7390 	else
7391 		q->q_flag &= ~QEND;
7392 	mutex_exit(QLOCK(q));
7393 }
7394 
7395 
7396 void
7397 clr_qfull(queue_t *q)
7398 {
7399 	queue_t	*oq = q;
7400 
7401 	q = q->q_nfsrv;
7402 	/* Fast check if there is any work to do before getting the lock. */
7403 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7404 		return;
7405 	}
7406 
7407 	/*
7408 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7409 	 * for QFULL being set.
7410 	 */
7411 	mutex_enter(QLOCK(q));
7412 	/*
7413 	 * If both q_count and q_mblkcnt are less than the hiwat mark
7414 	 */
7415 	if ((q->q_count < q->q_hiwat) && (q->q_mblkcnt < q->q_hiwat)) {
7416 		q->q_flag &= ~QFULL;
7417 		/*
7418 		 * A little more confusing, how about this way:
7419 		 * if someone wants to write,
7420 		 * AND
7421 		 *    both counts are less than the lowat mark
7422 		 *    OR
7423 		 *    the lowat mark is zero
7424 		 * THEN
7425 		 * backenable
7426 		 */
7427 		if ((q->q_flag & QWANTW) &&
7428 		    (((q->q_count < q->q_lowat) &&
7429 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7430 			q->q_flag &= ~QWANTW;
7431 			mutex_exit(QLOCK(q));
7432 			backenable(oq, 0);
7433 		} else
7434 			mutex_exit(QLOCK(q));
7435 	} else
7436 		mutex_exit(QLOCK(q));
7437 }
7438 
7439 /*
7440  * Set the forward service procedure pointer.
7441  *
7442  * Called at insert-time to cache a queue's next forward service procedure in
7443  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7444  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7445  * inserted does not have a service procedure, then q_nfsrv points to the next
7446  * queue forward that has a service procedure.  If the queue is at the logical
7447  * end of the stream (driver for write side, stream head for the read side)
7448  * and does not have a service procedure, then q_nfsrv also points to itself.
7449  */
7450 void
7451 set_nfsrv_ptr(
7452 	queue_t  *rnew,		/* read queue pointer to new module */
7453 	queue_t  *wnew,		/* write queue pointer to new module */
7454 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7455 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7456 {
7457 	queue_t *qp;
7458 
7459 	if (prev_wq->q_next == NULL) {
7460 		/*
7461 		 * Insert the driver, initialize the driver and stream head.
7462 		 * In this case, prev_rq/prev_wq should be the stream head.
7463 		 * _I_INSERT does not allow inserting a driver.  Make sure
7464 		 * that it is not an insertion.
7465 		 */
7466 		ASSERT(!(rnew->q_flag & _QINSERTING));
7467 		wnew->q_nfsrv = wnew;
7468 		if (rnew->q_qinfo->qi_srvp)
7469 			rnew->q_nfsrv = rnew;
7470 		else
7471 			rnew->q_nfsrv = prev_rq;
7472 		prev_rq->q_nfsrv = prev_rq;
7473 		prev_wq->q_nfsrv = prev_wq;
7474 	} else {
7475 		/*
7476 		 * set up read side q_nfsrv pointer.  This MUST be done
7477 		 * before setting the write side, because the setting of
7478 		 * the write side for a fifo may depend on it.
7479 		 *
7480 		 * Suppose we have a fifo that only has pipemod pushed.
7481 		 * pipemod has no read or write service procedures, so
7482 		 * nfsrv for both pipemod queues points to prev_rq (the
7483 		 * stream read head).  Now push bufmod (which has only a
7484 		 * read service procedure).  Doing the write side first,
7485 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7486 		 * is WRONG; the next queue forward from wnew with a
7487 		 * service procedure will be rnew, not the stream read head.
7488 		 * Since the downstream queue (which in the case of a fifo
7489 		 * is the read queue rnew) can affect upstream queues, it
7490 		 * needs to be done first.  Setting up the read side first
7491 		 * sets nfsrv for both pipemod queues to rnew and then
7492 		 * when the write side is set up, wnew-q_nfsrv will also
7493 		 * point to rnew.
7494 		 */
7495 		if (rnew->q_qinfo->qi_srvp) {
7496 			/*
7497 			 * use _OTHERQ() because, if this is a pipe, next
7498 			 * module may have been pushed from other end and
7499 			 * q_next could be a read queue.
7500 			 */
7501 			qp = _OTHERQ(prev_wq->q_next);
7502 			while (qp && qp->q_nfsrv != qp) {
7503 				qp->q_nfsrv = rnew;
7504 				qp = backq(qp);
7505 			}
7506 			rnew->q_nfsrv = rnew;
7507 		} else
7508 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7509 
7510 		/* set up write side q_nfsrv pointer */
7511 		if (wnew->q_qinfo->qi_srvp) {
7512 			wnew->q_nfsrv = wnew;
7513 
7514 			/*
7515 			 * For insertion, need to update nfsrv of the modules
7516 			 * above which do not have a service routine.
7517 			 */
7518 			if (rnew->q_flag & _QINSERTING) {
7519 				for (qp = prev_wq;
7520 				    qp != NULL && qp->q_nfsrv != qp;
7521 				    qp = backq(qp)) {
7522 					qp->q_nfsrv = wnew->q_nfsrv;
7523 				}
7524 			}
7525 		} else {
7526 			if (prev_wq->q_next == prev_rq)
7527 				/*
7528 				 * Since prev_wq/prev_rq are the middle of a
7529 				 * fifo, wnew/rnew will also be the middle of
7530 				 * a fifo and wnew's nfsrv is same as rnew's.
7531 				 */
7532 				wnew->q_nfsrv = rnew->q_nfsrv;
7533 			else
7534 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7535 		}
7536 	}
7537 }
7538 
7539 /*
7540  * Reset the forward service procedure pointer; called at remove-time.
7541  */
7542 void
7543 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7544 {
7545 	queue_t *tmp_qp;
7546 
7547 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7548 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7549 		for (tmp_qp = backq(wqp);
7550 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7551 		    tmp_qp = backq(tmp_qp)) {
7552 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7553 		}
7554 	}
7555 
7556 	/* reset the read side q_nfsrv pointer */
7557 	if (rqp->q_qinfo->qi_srvp) {
7558 		if (wqp->q_next) {	/* non-driver case */
7559 			tmp_qp = _OTHERQ(wqp->q_next);
7560 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7561 				/* Note that rqp->q_next cannot be NULL */
7562 				ASSERT(rqp->q_next != NULL);
7563 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7564 				tmp_qp = backq(tmp_qp);
7565 			}
7566 		}
7567 	}
7568 }
7569 
7570 /*
7571  * This routine should be called after all stream geometry changes to update
7572  * the stream head cached struio() rd/wr queue pointers. Note must be called
7573  * with the streamlock()ed.
7574  *
7575  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7576  *	 an explicit synchronous barrier module queue. That is, a queue that
7577  *	 has specified a struio() type.
7578  */
7579 static void
7580 strsetuio(stdata_t *stp)
7581 {
7582 	queue_t *wrq;
7583 
7584 	if (stp->sd_flag & STPLEX) {
7585 		/*
7586 		 * Not stremahead, but a mux, so no Synchronous STREAMS.
7587 		 */
7588 		stp->sd_struiowrq = NULL;
7589 		stp->sd_struiordq = NULL;
7590 		return;
7591 	}
7592 	/*
7593 	 * Scan the write queue(s) while synchronous
7594 	 * until we find a qinfo uio type specified.
7595 	 */
7596 	wrq = stp->sd_wrq->q_next;
7597 	while (wrq) {
7598 		if (wrq->q_struiot == STRUIOT_NONE) {
7599 			wrq = 0;
7600 			break;
7601 		}
7602 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7603 			break;
7604 		if (! _SAMESTR(wrq)) {
7605 			wrq = 0;
7606 			break;
7607 		}
7608 		wrq = wrq->q_next;
7609 	}
7610 	stp->sd_struiowrq = wrq;
7611 	/*
7612 	 * Scan the read queue(s) while synchronous
7613 	 * until we find a qinfo uio type specified.
7614 	 */
7615 	wrq = stp->sd_wrq->q_next;
7616 	while (wrq) {
7617 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7618 			wrq = 0;
7619 			break;
7620 		}
7621 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7622 			break;
7623 		if (! _SAMESTR(wrq)) {
7624 			wrq = 0;
7625 			break;
7626 		}
7627 		wrq = wrq->q_next;
7628 	}
7629 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7630 }
7631 
7632 /*
7633  * pass_wput, unblocks the passthru queues, so that
7634  * messages can arrive at muxs lower read queue, before
7635  * I_LINK/I_UNLINK is acked/nacked.
7636  */
7637 static void
7638 pass_wput(queue_t *q, mblk_t *mp)
7639 {
7640 	syncq_t *sq;
7641 
7642 	sq = _RD(q)->q_syncq;
7643 	if (sq->sq_flags & SQ_BLOCKED)
7644 		unblocksq(sq, SQ_BLOCKED, 0);
7645 	putnext(q, mp);
7646 }
7647 
7648 /*
7649  * Set up queues for the link/unlink.
7650  * Create a new queue and block it and then insert it
7651  * below the stream head on the lower stream.
7652  * This prevents any messages from arriving during the setq
7653  * as well as while the mux is processing the LINK/I_UNLINK.
7654  * The blocked passq is unblocked once the LINK/I_UNLINK has
7655  * been acked or nacked or if a message is generated and sent
7656  * down muxs write put procedure.
7657  * see pass_wput().
7658  *
7659  * After the new queue is inserted, all messages coming from below are
7660  * blocked. The call to strlock will ensure that all activity in the stream head
7661  * read queue syncq is stopped (sq_count drops to zero).
7662  */
7663 static queue_t *
7664 link_addpassthru(stdata_t *stpdown)
7665 {
7666 	queue_t *passq;
7667 	sqlist_t sqlist;
7668 
7669 	passq = allocq();
7670 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7671 	/* setq might sleep in allocator - avoid holding locks. */
7672 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7673 	    SQ_CI|SQ_CO, B_FALSE);
7674 	claimq(passq);
7675 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7676 	insertq(STREAM(passq), passq);
7677 
7678 	/*
7679 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7680 	 * since we are going to change q_ptr in the stream head.  Note that
7681 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7682 	 */
7683 	sqlist.sqlist_head = NULL;
7684 	sqlist.sqlist_index = 0;
7685 	sqlist.sqlist_size = sizeof (sqlist_t);
7686 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7687 	strlock(stpdown, &sqlist);
7688 	strunlock(stpdown, &sqlist);
7689 
7690 	releaseq(passq);
7691 	return (passq);
7692 }
7693 
7694 /*
7695  * Let messages flow up into the mux by removing
7696  * the passq.
7697  */
7698 static void
7699 link_rempassthru(queue_t *passq)
7700 {
7701 	claimq(passq);
7702 	removeq(passq);
7703 	releaseq(passq);
7704 	freeq(passq);
7705 }
7706 
7707 /*
7708  * Wait for the condition variable pointed to by `cvp' to be signaled,
7709  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7710  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7711  * then the wait will be non-interruptible.
7712  *
7713  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7714  */
7715 clock_t
7716 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7717 {
7718 	clock_t ret, now, tick;
7719 
7720 	if (tim < 0) {
7721 		if (nosigs) {
7722 			cv_wait(cvp, mp);
7723 			ret = 1;
7724 		} else {
7725 			ret = cv_wait_sig(cvp, mp);
7726 		}
7727 	} else if (tim > 0) {
7728 		/*
7729 		 * convert milliseconds to clock ticks
7730 		 */
7731 		tick = MSEC_TO_TICK_ROUNDUP(tim);
7732 		time_to_wait(&now, tick);
7733 		if (nosigs) {
7734 			ret = cv_timedwait(cvp, mp, now);
7735 		} else {
7736 			ret = cv_timedwait_sig(cvp, mp, now);
7737 		}
7738 	} else {
7739 		ret = -1;
7740 	}
7741 	return (ret);
7742 }
7743 
7744 /*
7745  * Wait until the stream head can determine if it is at the mark but
7746  * don't wait forever to prevent a race condition between the "mark" state
7747  * in the stream head and any mark state in the caller/user of this routine.
7748  *
7749  * This is used by sockets and for a socket it would be incorrect
7750  * to return a failure for SIOCATMARK when there is no data in the receive
7751  * queue and the marked urgent data is traveling up the stream.
7752  *
7753  * This routine waits until the mark is known by waiting for one of these
7754  * three events:
7755  *	The stream head read queue becoming non-empty (including an EOF)
7756  *	The STRATMARK flag being set. (Due to a MSGMARKNEXT message.)
7757  *	The STRNOTATMARK flag being set (which indicates that the transport
7758  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7759  *	the mark).
7760  *
7761  * The routine returns 1 if the stream is at the mark; 0 if it can
7762  * be determined that the stream is not at the mark.
7763  * If the wait times out and it can't determine
7764  * whether or not the stream might be at the mark the routine will return -1.
7765  *
7766  * Note: This routine should only be used when a mark is pending i.e.,
7767  * in the socket case the SIGURG has been posted.
7768  * Note2: This can not wakeup just because synchronous streams indicate
7769  * that data is available since it is not possible to use the synchronous
7770  * streams interfaces to determine the b_flag value for the data queued below
7771  * the stream head.
7772  */
7773 int
7774 strwaitmark(vnode_t *vp)
7775 {
7776 	struct stdata *stp = vp->v_stream;
7777 	queue_t *rq = _RD(stp->sd_wrq);
7778 	int mark;
7779 
7780 	mutex_enter(&stp->sd_lock);
7781 	while (rq->q_first == NULL &&
7782 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7783 		stp->sd_flag |= RSLEEP;
7784 
7785 		/* Wait for 100 milliseconds for any state change. */
7786 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7787 			mutex_exit(&stp->sd_lock);
7788 			return (-1);
7789 		}
7790 	}
7791 	if (stp->sd_flag & STRATMARK)
7792 		mark = 1;
7793 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7794 		mark = 1;
7795 	else
7796 		mark = 0;
7797 
7798 	mutex_exit(&stp->sd_lock);
7799 	return (mark);
7800 }
7801 
7802 /*
7803  * Set a read side error. If persist is set change the socket error
7804  * to persistent. If errfunc is set install the function as the exported
7805  * error handler.
7806  */
7807 void
7808 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7809 {
7810 	struct stdata *stp = vp->v_stream;
7811 
7812 	mutex_enter(&stp->sd_lock);
7813 	stp->sd_rerror = error;
7814 	if (error == 0 && errfunc == NULL)
7815 		stp->sd_flag &= ~STRDERR;
7816 	else
7817 		stp->sd_flag |= STRDERR;
7818 	if (persist) {
7819 		stp->sd_flag &= ~STRDERRNONPERSIST;
7820 	} else {
7821 		stp->sd_flag |= STRDERRNONPERSIST;
7822 	}
7823 	stp->sd_rderrfunc = errfunc;
7824 	if (error != 0 || errfunc != NULL) {
7825 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7826 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7827 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7828 
7829 		mutex_exit(&stp->sd_lock);
7830 		pollwakeup(&stp->sd_pollist, POLLERR);
7831 		mutex_enter(&stp->sd_lock);
7832 
7833 		if (stp->sd_sigflags & S_ERROR)
7834 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
7835 	}
7836 	mutex_exit(&stp->sd_lock);
7837 }
7838 
7839 /*
7840  * Set a write side error. If persist is set change the socket error
7841  * to persistent.
7842  */
7843 void
7844 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7845 {
7846 	struct stdata *stp = vp->v_stream;
7847 
7848 	mutex_enter(&stp->sd_lock);
7849 	stp->sd_werror = error;
7850 	if (error == 0 && errfunc == NULL)
7851 		stp->sd_flag &= ~STWRERR;
7852 	else
7853 		stp->sd_flag |= STWRERR;
7854 	if (persist) {
7855 		stp->sd_flag &= ~STWRERRNONPERSIST;
7856 	} else {
7857 		stp->sd_flag |= STWRERRNONPERSIST;
7858 	}
7859 	stp->sd_wrerrfunc = errfunc;
7860 	if (error != 0 || errfunc != NULL) {
7861 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7862 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7863 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7864 
7865 		mutex_exit(&stp->sd_lock);
7866 		pollwakeup(&stp->sd_pollist, POLLERR);
7867 		mutex_enter(&stp->sd_lock);
7868 
7869 		if (stp->sd_sigflags & S_ERROR)
7870 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
7871 	}
7872 	mutex_exit(&stp->sd_lock);
7873 }
7874 
7875 /*
7876  * Make the stream return 0 (EOF) when all data has been read.
7877  * No effect on write side.
7878  */
7879 void
7880 strseteof(vnode_t *vp, int eof)
7881 {
7882 	struct stdata *stp = vp->v_stream;
7883 
7884 	mutex_enter(&stp->sd_lock);
7885 	if (!eof) {
7886 		stp->sd_flag &= ~STREOF;
7887 		mutex_exit(&stp->sd_lock);
7888 		return;
7889 	}
7890 	stp->sd_flag |= STREOF;
7891 	if (stp->sd_flag & RSLEEP) {
7892 		stp->sd_flag &= ~RSLEEP;
7893 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
7894 	}
7895 
7896 	mutex_exit(&stp->sd_lock);
7897 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
7898 	mutex_enter(&stp->sd_lock);
7899 
7900 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
7901 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
7902 	mutex_exit(&stp->sd_lock);
7903 }
7904 
7905 void
7906 strflushrq(vnode_t *vp, int flag)
7907 {
7908 	struct stdata *stp = vp->v_stream;
7909 
7910 	mutex_enter(&stp->sd_lock);
7911 	flushq(_RD(stp->sd_wrq), flag);
7912 	mutex_exit(&stp->sd_lock);
7913 }
7914 
7915 void
7916 strsetrputhooks(vnode_t *vp, uint_t flags,
7917 		msgfunc_t protofunc, msgfunc_t miscfunc)
7918 {
7919 	struct stdata *stp = vp->v_stream;
7920 
7921 	mutex_enter(&stp->sd_lock);
7922 
7923 	if (protofunc == NULL)
7924 		stp->sd_rprotofunc = strrput_proto;
7925 	else
7926 		stp->sd_rprotofunc = protofunc;
7927 
7928 	if (miscfunc == NULL)
7929 		stp->sd_rmiscfunc = strrput_misc;
7930 	else
7931 		stp->sd_rmiscfunc = miscfunc;
7932 
7933 	if (flags & SH_CONSOL_DATA)
7934 		stp->sd_rput_opt |= SR_CONSOL_DATA;
7935 	else
7936 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
7937 
7938 	if (flags & SH_SIGALLDATA)
7939 		stp->sd_rput_opt |= SR_SIGALLDATA;
7940 	else
7941 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
7942 
7943 	if (flags & SH_IGN_ZEROLEN)
7944 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
7945 	else
7946 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
7947 
7948 	mutex_exit(&stp->sd_lock);
7949 }
7950 
7951 void
7952 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
7953 {
7954 	struct stdata *stp = vp->v_stream;
7955 
7956 	mutex_enter(&stp->sd_lock);
7957 	stp->sd_closetime = closetime;
7958 
7959 	if (flags & SH_SIGPIPE)
7960 		stp->sd_wput_opt |= SW_SIGPIPE;
7961 	else
7962 		stp->sd_wput_opt &= ~SW_SIGPIPE;
7963 	if (flags & SH_RECHECK_ERR)
7964 		stp->sd_wput_opt |= SW_RECHECK_ERR;
7965 	else
7966 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
7967 
7968 	mutex_exit(&stp->sd_lock);
7969 }
7970 
7971 void
7972 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
7973 {
7974 	struct stdata *stp = vp->v_stream;
7975 
7976 	mutex_enter(&stp->sd_lock);
7977 
7978 	stp->sd_rputdatafunc = rdatafunc;
7979 	stp->sd_wputdatafunc = wdatafunc;
7980 
7981 	mutex_exit(&stp->sd_lock);
7982 }
7983 
7984 /* Used within framework when the queue is already locked */
7985 void
7986 qenable_locked(queue_t *q)
7987 {
7988 	stdata_t *stp = STREAM(q);
7989 
7990 	ASSERT(MUTEX_HELD(QLOCK(q)));
7991 
7992 	if (!q->q_qinfo->qi_srvp)
7993 		return;
7994 
7995 	/*
7996 	 * Do not place on run queue if already enabled or closing.
7997 	 */
7998 	if (q->q_flag & (QWCLOSE|QENAB))
7999 		return;
8000 
8001 	/*
8002 	 * mark queue enabled and place on run list if it is not already being
8003 	 * serviced. If it is serviced, the runservice() function will detect
8004 	 * that QENAB is set and call service procedure before clearing
8005 	 * QINSERVICE flag.
8006 	 */
8007 	q->q_flag |= QENAB;
8008 	if (q->q_flag & QINSERVICE)
8009 		return;
8010 
8011 	/* Record the time of qenable */
8012 	q->q_qtstamp = lbolt;
8013 
8014 	/*
8015 	 * Put the queue in the stp list and schedule it for background
8016 	 * processing if it is not already scheduled or if stream head does not
8017 	 * intent to process it in the foreground later by setting
8018 	 * STRS_WILLSERVICE flag.
8019 	 */
8020 	mutex_enter(&stp->sd_qlock);
8021 	/*
8022 	 * If there are already something on the list, stp flags should show
8023 	 * intention to drain it.
8024 	 */
8025 	IMPLY(STREAM_NEEDSERVICE(stp),
8026 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8027 
8028 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8029 	stp->sd_nqueues++;
8030 
8031 	/*
8032 	 * If no one will drain this stream we are the first producer and
8033 	 * need to schedule it for background thread.
8034 	 */
8035 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8036 		/*
8037 		 * No one will service this stream later, so we have to
8038 		 * schedule it now.
8039 		 */
8040 		STRSTAT(stenables);
8041 		stp->sd_svcflags |= STRS_SCHEDULED;
8042 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8043 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8044 
8045 		if (stp->sd_servid == NULL) {
8046 			/*
8047 			 * Task queue failed so fail over to the backup
8048 			 * servicing thread.
8049 			 */
8050 			STRSTAT(taskqfails);
8051 			/*
8052 			 * It is safe to clear STRS_SCHEDULED flag because it
8053 			 * was set by this thread above.
8054 			 */
8055 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8056 
8057 			/*
8058 			 * Failover scheduling is protected by service_queue
8059 			 * lock.
8060 			 */
8061 			mutex_enter(&service_queue);
8062 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8063 			ASSERT(q->q_link == NULL);
8064 			/*
8065 			 * Append the queue to qhead/qtail list.
8066 			 */
8067 			if (qhead == NULL)
8068 				qhead = q;
8069 			else
8070 				qtail->q_link = q;
8071 			qtail = q;
8072 			/*
8073 			 * Clear stp queue list.
8074 			 */
8075 			stp->sd_qhead = stp->sd_qtail = NULL;
8076 			stp->sd_nqueues = 0;
8077 			/*
8078 			 * Wakeup background queue processing thread.
8079 			 */
8080 			cv_signal(&services_to_run);
8081 			mutex_exit(&service_queue);
8082 		}
8083 	}
8084 	mutex_exit(&stp->sd_qlock);
8085 }
8086 
8087 static void
8088 queue_service(queue_t *q)
8089 {
8090 	/*
8091 	 * The queue in the list should have
8092 	 * QENAB flag set and should not have
8093 	 * QINSERVICE flag set. QINSERVICE is
8094 	 * set when the queue is dequeued and
8095 	 * qenable_locked doesn't enqueue a
8096 	 * queue with QINSERVICE set.
8097 	 */
8098 
8099 	ASSERT(!(q->q_flag & QINSERVICE));
8100 	ASSERT((q->q_flag & QENAB));
8101 	mutex_enter(QLOCK(q));
8102 	q->q_flag &= ~QENAB;
8103 	q->q_flag |= QINSERVICE;
8104 	mutex_exit(QLOCK(q));
8105 	runservice(q);
8106 }
8107 
8108 static void
8109 syncq_service(syncq_t *sq)
8110 {
8111 	STRSTAT(syncqservice);
8112 	mutex_enter(SQLOCK(sq));
8113 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8114 	ASSERT(sq->sq_servcount != 0);
8115 	ASSERT(sq->sq_next == NULL);
8116 
8117 	/* if we came here from the background thread, clear the flag */
8118 	if (sq->sq_svcflags & SQ_BGTHREAD)
8119 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8120 
8121 	/* let drain_syncq know that it's being called in the background */
8122 	sq->sq_svcflags |= SQ_SERVICE;
8123 	drain_syncq(sq);
8124 }
8125 
8126 static void
8127 qwriter_outer_service(syncq_t *outer)
8128 {
8129 	/*
8130 	 * Note that SQ_WRITER is used on the outer perimeter
8131 	 * to signal that a qwriter(OUTER) is either investigating
8132 	 * running or that it is actually running a function.
8133 	 */
8134 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8135 
8136 	/*
8137 	 * All inner syncq are empty and have SQ_WRITER set
8138 	 * to block entering the outer perimeter.
8139 	 *
8140 	 * We do not need to explicitly call write_now since
8141 	 * outer_exit does it for us.
8142 	 */
8143 	outer_exit(outer);
8144 }
8145 
8146 static void
8147 mblk_free(mblk_t *mp)
8148 {
8149 	dblk_t *dbp = mp->b_datap;
8150 	frtn_t *frp = dbp->db_frtnp;
8151 
8152 	mp->b_next = NULL;
8153 	if (dbp->db_fthdr != NULL)
8154 		str_ftfree(dbp);
8155 
8156 	ASSERT(dbp->db_fthdr == NULL);
8157 	frp->free_func(frp->free_arg);
8158 	ASSERT(dbp->db_mblk == mp);
8159 
8160 	if (dbp->db_credp != NULL) {
8161 		crfree(dbp->db_credp);
8162 		dbp->db_credp = NULL;
8163 	}
8164 	dbp->db_cpid = -1;
8165 	dbp->db_struioflag = 0;
8166 	dbp->db_struioun.cksum.flags = 0;
8167 
8168 	kmem_cache_free(dbp->db_cache, dbp);
8169 }
8170 
8171 /*
8172  * Background processing of the stream queue list.
8173  */
8174 static void
8175 stream_service(stdata_t *stp)
8176 {
8177 	queue_t *q;
8178 
8179 	mutex_enter(&stp->sd_qlock);
8180 
8181 	STR_SERVICE(stp, q);
8182 
8183 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8184 	stp->sd_servid = NULL;
8185 	cv_signal(&stp->sd_qcv);
8186 	mutex_exit(&stp->sd_qlock);
8187 }
8188 
8189 /*
8190  * Foreground processing of the stream queue list.
8191  */
8192 void
8193 stream_runservice(stdata_t *stp)
8194 {
8195 	queue_t *q;
8196 
8197 	mutex_enter(&stp->sd_qlock);
8198 	STRSTAT(rservice);
8199 	/*
8200 	 * We are going to drain this stream queue list, so qenable_locked will
8201 	 * not schedule it until we finish.
8202 	 */
8203 	stp->sd_svcflags |= STRS_WILLSERVICE;
8204 
8205 	STR_SERVICE(stp, q);
8206 
8207 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8208 	mutex_exit(&stp->sd_qlock);
8209 	/*
8210 	 * Help backup background thread to drain the qhead/qtail list.
8211 	 */
8212 	while (qhead != NULL) {
8213 		STRSTAT(qhelps);
8214 		mutex_enter(&service_queue);
8215 		DQ(q, qhead, qtail, q_link);
8216 		mutex_exit(&service_queue);
8217 		if (q != NULL)
8218 			queue_service(q);
8219 	}
8220 }
8221 
8222 void
8223 stream_willservice(stdata_t *stp)
8224 {
8225 	mutex_enter(&stp->sd_qlock);
8226 	stp->sd_svcflags |= STRS_WILLSERVICE;
8227 	mutex_exit(&stp->sd_qlock);
8228 }
8229 
8230 /*
8231  * Replace the cred currently in the mblk with a different one.
8232  */
8233 void
8234 mblk_setcred(mblk_t *mp, cred_t *cr)
8235 {
8236 	cred_t *ocr = DB_CRED(mp);
8237 
8238 	ASSERT(cr != NULL);
8239 
8240 	if (cr != ocr) {
8241 		crhold(mp->b_datap->db_credp = cr);
8242 		if (ocr != NULL)
8243 			crfree(ocr);
8244 	}
8245 }
8246 
8247 int
8248 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8249     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8250     uint32_t flags, int km_flags)
8251 {
8252 	int rc = 0;
8253 
8254 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8255 	if (mp->b_datap->db_type == M_DATA) {
8256 		/* Associate values for M_DATA type */
8257 		DB_CKSUMSTART(mp) = (intptr_t)start;
8258 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8259 		DB_CKSUMEND(mp) = (intptr_t)end;
8260 		DB_CKSUMFLAGS(mp) = flags;
8261 		DB_CKSUM16(mp) = (uint16_t)value;
8262 
8263 	} else {
8264 		pattrinfo_t pa_info;
8265 
8266 		ASSERT(mmd != NULL);
8267 
8268 		pa_info.type = PATTR_HCKSUM;
8269 		pa_info.len = sizeof (pattr_hcksum_t);
8270 
8271 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8272 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8273 
8274 			hck->hcksum_start_offset = start;
8275 			hck->hcksum_stuff_offset = stuff;
8276 			hck->hcksum_end_offset = end;
8277 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8278 			hck->hcksum_flags = flags;
8279 		} else {
8280 			rc = -1;
8281 		}
8282 	}
8283 	return (rc);
8284 }
8285 
8286 void
8287 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8288     uint32_t *start, uint32_t *stuff, uint32_t *end,
8289     uint32_t *value, uint32_t *flags)
8290 {
8291 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8292 	if (mp->b_datap->db_type == M_DATA) {
8293 		if (flags != NULL) {
8294 			*flags = DB_CKSUMFLAGS(mp);
8295 			if (*flags & HCK_PARTIALCKSUM) {
8296 				if (start != NULL)
8297 					*start = (uint32_t)DB_CKSUMSTART(mp);
8298 				if (stuff != NULL)
8299 					*stuff = (uint32_t)DB_CKSUMSTUFF(mp);
8300 				if (end != NULL)
8301 					*end = (uint32_t)DB_CKSUMEND(mp);
8302 				if (value != NULL)
8303 					*value = (uint32_t)DB_CKSUM16(mp);
8304 			} else if ((*flags & HW_LSO) && (value != NULL))
8305 				*value = (uint32_t)DB_LSOMSS(mp);
8306 		}
8307 	} else {
8308 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8309 
8310 		ASSERT(mmd != NULL);
8311 
8312 		/* get hardware checksum attribute */
8313 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8314 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8315 
8316 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8317 			if (flags != NULL)
8318 				*flags = hck->hcksum_flags;
8319 			if (start != NULL)
8320 				*start = hck->hcksum_start_offset;
8321 			if (stuff != NULL)
8322 				*stuff = hck->hcksum_stuff_offset;
8323 			if (end != NULL)
8324 				*end = hck->hcksum_end_offset;
8325 			if (value != NULL)
8326 				*value = (uint32_t)
8327 				    hck->hcksum_cksum_val.inet_cksum;
8328 		}
8329 	}
8330 }
8331 
8332 /*
8333  * Checksum buffer *bp for len bytes with psum partial checksum,
8334  * or 0 if none, and return the 16 bit partial checksum.
8335  */
8336 unsigned
8337 bcksum(uchar_t *bp, int len, unsigned int psum)
8338 {
8339 	int odd = len & 1;
8340 	extern unsigned int ip_ocsum();
8341 
8342 	if (((intptr_t)bp & 1) == 0 && !odd) {
8343 		/*
8344 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8345 		 */
8346 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8347 	}
8348 	if (((intptr_t)bp & 1) != 0) {
8349 		/*
8350 		 * Bp isn't 16 bit aligned.
8351 		 */
8352 		unsigned int tsum;
8353 
8354 #ifdef _LITTLE_ENDIAN
8355 		psum += *bp;
8356 #else
8357 		psum += *bp << 8;
8358 #endif
8359 		len--;
8360 		bp++;
8361 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8362 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8363 		if (len & 1) {
8364 			bp += len - 1;
8365 #ifdef _LITTLE_ENDIAN
8366 			psum += *bp << 8;
8367 #else
8368 			psum += *bp;
8369 #endif
8370 		}
8371 	} else {
8372 		/*
8373 		 * Bp is 16 bit aligned.
8374 		 */
8375 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8376 		if (odd) {
8377 			bp += len - 1;
8378 #ifdef _LITTLE_ENDIAN
8379 			psum += *bp;
8380 #else
8381 			psum += *bp << 8;
8382 #endif
8383 		}
8384 	}
8385 	/*
8386 	 * Normalize psum to 16 bits before returning the new partial
8387 	 * checksum. The max psum value before normalization is 0x3FDFE.
8388 	 */
8389 	return ((psum >> 16) + (psum & 0xFFFF));
8390 }
8391 
8392 boolean_t
8393 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8394 {
8395 	boolean_t rc;
8396 
8397 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8398 	if (DB_TYPE(mp) == M_DATA) {
8399 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8400 	} else {
8401 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8402 
8403 		ASSERT(mmd != NULL);
8404 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8405 	}
8406 	return (rc);
8407 }
8408 
8409 void
8410 freemsgchain(mblk_t *mp)
8411 {
8412 	mblk_t	*next;
8413 
8414 	while (mp != NULL) {
8415 		next = mp->b_next;
8416 		mp->b_next = NULL;
8417 
8418 		freemsg(mp);
8419 		mp = next;
8420 	}
8421 }
8422 
8423 mblk_t *
8424 copymsgchain(mblk_t *mp)
8425 {
8426 	mblk_t	*nmp = NULL;
8427 	mblk_t	**nmpp = &nmp;
8428 
8429 	for (; mp != NULL; mp = mp->b_next) {
8430 		if ((*nmpp = copymsg(mp)) == NULL) {
8431 			freemsgchain(nmp);
8432 			return (NULL);
8433 		}
8434 
8435 		nmpp = &((*nmpp)->b_next);
8436 	}
8437 
8438 	return (nmp);
8439 }
8440 
8441 /* NOTE: Do not add code after this point. */
8442 #undef QLOCK
8443 
8444 /*
8445  * replacement for QLOCK macro for those that can't use it.
8446  */
8447 kmutex_t *
8448 QLOCK(queue_t *q)
8449 {
8450 	return (&(q)->q_lock);
8451 }
8452 
8453 /*
8454  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8455  */
8456 #undef runqueues
8457 void
8458 runqueues(void)
8459 {
8460 }
8461 
8462 #undef queuerun
8463 void
8464 queuerun(void)
8465 {
8466 }
8467