xref: /titanic_41/usr/src/uts/common/os/strsubr.c (revision 63360950109af2ce85a962ca61f40b8782f11100)
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 2007 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 #include <sys/open.h>
81 #include <sys/sunldi.h>
82 #include <sys/sad.h>
83 #include <sys/netstack.h>
84 
85 #define	O_SAMESTR(q)	(((q)->q_next) && \
86 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
87 
88 /*
89  * WARNING:
90  * The variables and routines in this file are private, belonging
91  * to the STREAMS subsystem. These should not be used by modules
92  * or drivers. Compatibility will not be guaranteed.
93  */
94 
95 /*
96  * Id value used to distinguish between different multiplexor links.
97  */
98 static int32_t lnk_id = 0;
99 
100 #define	STREAMS_LOPRI MINCLSYSPRI
101 static pri_t streams_lopri = STREAMS_LOPRI;
102 
103 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
104 typedef struct str_stat {
105 	kstat_named_t	sqenables;
106 	kstat_named_t	stenables;
107 	kstat_named_t	syncqservice;
108 	kstat_named_t	freebs;
109 	kstat_named_t	qwr_outer;
110 	kstat_named_t	rservice;
111 	kstat_named_t	strwaits;
112 	kstat_named_t	taskqfails;
113 	kstat_named_t	bufcalls;
114 	kstat_named_t	qhelps;
115 	kstat_named_t	qremoved;
116 	kstat_named_t	sqremoved;
117 	kstat_named_t	bcwaits;
118 	kstat_named_t	sqtoomany;
119 } str_stat_t;
120 
121 static str_stat_t str_statistics = {
122 	{ "sqenables",		KSTAT_DATA_UINT64 },
123 	{ "stenables",		KSTAT_DATA_UINT64 },
124 	{ "syncqservice",	KSTAT_DATA_UINT64 },
125 	{ "freebs",		KSTAT_DATA_UINT64 },
126 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
127 	{ "rservice",		KSTAT_DATA_UINT64 },
128 	{ "strwaits",		KSTAT_DATA_UINT64 },
129 	{ "taskqfails",		KSTAT_DATA_UINT64 },
130 	{ "bufcalls",		KSTAT_DATA_UINT64 },
131 	{ "qhelps",		KSTAT_DATA_UINT64 },
132 	{ "qremoved",		KSTAT_DATA_UINT64 },
133 	{ "sqremoved",		KSTAT_DATA_UINT64 },
134 	{ "bcwaits",		KSTAT_DATA_UINT64 },
135 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
136 };
137 
138 static kstat_t *str_kstat;
139 
140 /*
141  * qrunflag was used previously to control background scheduling of queues. It
142  * is not used anymore, but kept here in case some module still wants to access
143  * it via qready() and setqsched macros.
144  */
145 char qrunflag;			/*  Unused */
146 
147 /*
148  * Most of the streams scheduling is done via task queues. Task queues may fail
149  * for non-sleep dispatches, so there are two backup threads servicing failed
150  * requests for queues and syncqs. Both of these threads also service failed
151  * dispatches freebs requests. Queues are put in the list specified by `qhead'
152  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
153  * requests are put into `freebs_list' which has no tail pointer. All three
154  * lists are protected by a single `service_queue' lock and use
155  * `services_to_run' condition variable for signaling background threads. Use of
156  * a single lock should not be a problem because it is only used under heavy
157  * loads when task queues start to fail and at that time it may be a good idea
158  * to throttle scheduling requests.
159  *
160  * NOTE: queues and syncqs should be scheduled by two separate threads because
161  * queue servicing may be blocked waiting for a syncq which may be also
162  * scheduled for background execution. This may create a deadlock when only one
163  * thread is used for both.
164  */
165 
166 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
167 
168 static kmutex_t service_queue;		/* protects all of servicing vars */
169 static kcondvar_t services_to_run;	/* wake up background service thread */
170 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
171 
172 /*
173  * List of queues scheduled for background processing dueue to lack of resources
174  * in the task queues. Protected by service_queue lock;
175  */
176 static struct queue *qhead;
177 static struct queue *qtail;
178 
179 /*
180  * Same list for syncqs
181  */
182 static syncq_t *sqhead;
183 static syncq_t *sqtail;
184 
185 static mblk_t *freebs_list;	/* list of buffers to free */
186 
187 /*
188  * Backup threads for servicing queues and syncqs
189  */
190 kthread_t *streams_qbkgrnd_thread;
191 kthread_t *streams_sqbkgrnd_thread;
192 
193 /*
194  * Bufcalls related variables.
195  */
196 struct bclist	strbcalls;	/* list of waiting bufcalls */
197 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
198 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
199 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
200 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
201 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
202 
203 kmutex_t	strresources;	/* protects global resources */
204 kmutex_t	muxifier;	/* single-threads multiplexor creation */
205 
206 static void	*str_stack_init(netstackid_t stackid, netstack_t *ns);
207 static void	str_stack_shutdown(netstackid_t stackid, void *arg);
208 static void	str_stack_fini(netstackid_t stackid, void *arg);
209 
210 extern void	time_to_wait(clock_t *, clock_t);
211 
212 /*
213  * run_queues is no longer used, but is kept in case some 3-d party
214  * module/driver decides to use it.
215  */
216 int run_queues = 0;
217 
218 /*
219  * sq_max_size is the depth of the syncq (in number of messages) before
220  * qfill_syncq() starts QFULL'ing destination queues. As its primary
221  * consumer - IP is no longer D_MTPERMOD, but there may be other
222  * modules/drivers depend on this syncq flow control, we prefer to
223  * choose a large number as the default value. For potential
224  * performance gain, this value is tunable in /etc/system.
225  */
226 int sq_max_size = 10000;
227 
228 /*
229  * the number of ciputctrl structures per syncq and stream we create when
230  * needed.
231  */
232 int n_ciputctrl;
233 int max_n_ciputctrl = 16;
234 /*
235  * if n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
236  */
237 int min_n_ciputctrl = 2;
238 
239 /*
240  * Per-driver/module syncqs
241  * ========================
242  *
243  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
244  * perdm structures, new entries being added (and new syncqs allocated) when
245  * setq() encounters a module/driver with a streamtab that it hasn't seen
246  * before.
247  * The reason for this mechanism is that some modules and drivers share a
248  * common streamtab and it is necessary for those modules and drivers to also
249  * share a common PERMOD syncq.
250  *
251  * perdm_list --> dm_str == streamtab_1
252  *                dm_sq == syncq_1
253  *                dm_ref
254  *                dm_next --> dm_str == streamtab_2
255  *                            dm_sq == syncq_2
256  *                            dm_ref
257  *                            dm_next --> ... NULL
258  *
259  * The dm_ref field is incremented for each new driver/module that takes
260  * a reference to the perdm structure and hence shares the syncq.
261  * References are held in the fmodsw_impl_t structure for each STREAMS module
262  * or the dev_impl array (indexed by device major number) for each driver.
263  *
264  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
265  *		     ^                 ^ ^               ^
266  *                   |  ______________/  |               |
267  *                   | /                 |               |
268  * dev_impl:     ...|x|y|...          module A	      module B
269  *
270  * When a module/driver is unloaded the reference count is decremented and,
271  * when it falls to zero, the perdm structure is removed from the list and
272  * the syncq is freed (see rele_dm()).
273  */
274 perdm_t *perdm_list = NULL;
275 static krwlock_t perdm_rwlock;
276 cdevsw_impl_t *devimpl;
277 
278 extern struct qinit strdata;
279 extern struct qinit stwdata;
280 
281 static void runservice(queue_t *);
282 static void streams_bufcall_service(void);
283 static void streams_qbkgrnd_service(void);
284 static void streams_sqbkgrnd_service(void);
285 static syncq_t *new_syncq(void);
286 static void free_syncq(syncq_t *);
287 static void outer_insert(syncq_t *, syncq_t *);
288 static void outer_remove(syncq_t *, syncq_t *);
289 static void write_now(syncq_t *);
290 static void clr_qfull(queue_t *);
291 static void enable_svc(queue_t *);
292 static void runbufcalls(void);
293 static void sqenable(syncq_t *);
294 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
295 static void wait_q_syncq(queue_t *);
296 static void backenable_insertedq(queue_t *);
297 
298 static void queue_service(queue_t *);
299 static void stream_service(stdata_t *);
300 static void syncq_service(syncq_t *);
301 static void qwriter_outer_service(syncq_t *);
302 static void mblk_free(mblk_t *);
303 #ifdef DEBUG
304 static int qprocsareon(queue_t *);
305 #endif
306 
307 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
308 static void reset_nfsrv_ptr(queue_t *, queue_t *);
309 
310 static void sq_run_events(syncq_t *);
311 static int propagate_syncq(queue_t *);
312 
313 static void	blocksq(syncq_t *, ushort_t, int);
314 static void	unblocksq(syncq_t *, ushort_t, int);
315 static int	dropsq(syncq_t *, uint16_t);
316 static void	emptysq(syncq_t *);
317 static sqlist_t *sqlist_alloc(struct stdata *, int);
318 static void	sqlist_free(sqlist_t *);
319 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
320 static void	sqlist_insert(sqlist_t *, syncq_t *);
321 static void	sqlist_insertall(sqlist_t *, queue_t *);
322 
323 static void	strsetuio(stdata_t *);
324 
325 struct kmem_cache *stream_head_cache;
326 struct kmem_cache *queue_cache;
327 struct kmem_cache *syncq_cache;
328 struct kmem_cache *qband_cache;
329 struct kmem_cache *linkinfo_cache;
330 struct kmem_cache *ciputctrl_cache = NULL;
331 
332 static linkinfo_t *linkinfo_list;
333 
334 /*
335  *  Qinit structure and Module_info structures
336  *	for passthru read and write queues
337  */
338 
339 static void pass_wput(queue_t *, mblk_t *);
340 static queue_t *link_addpassthru(stdata_t *);
341 static void link_rempassthru(queue_t *);
342 
343 struct  module_info passthru_info = {
344 	0,
345 	"passthru",
346 	0,
347 	INFPSZ,
348 	STRHIGH,
349 	STRLOW
350 };
351 
352 struct  qinit passthru_rinit = {
353 	(int (*)())putnext,
354 	NULL,
355 	NULL,
356 	NULL,
357 	NULL,
358 	&passthru_info,
359 	NULL
360 };
361 
362 struct  qinit passthru_winit = {
363 	(int (*)()) pass_wput,
364 	NULL,
365 	NULL,
366 	NULL,
367 	NULL,
368 	&passthru_info,
369 	NULL
370 };
371 
372 /*
373  * Special form of assertion: verify that X implies Y i.e. when X is true Y
374  * should also be true.
375  */
376 #define	IMPLY(X, Y)	ASSERT(!(X) || (Y))
377 
378 /*
379  * Logical equivalence. Verify that both X and Y are either TRUE or FALSE.
380  */
381 #define	EQUIV(X, Y)	{ IMPLY(X, Y); IMPLY(Y, X); }
382 
383 /*
384  * Verify correctness of list head/tail pointers.
385  */
386 #define	LISTCHECK(head, tail, link) {				\
387 	EQUIV(head, tail);					\
388 	IMPLY(tail != NULL, tail->link == NULL);		\
389 }
390 
391 /*
392  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
393  * using a `link' field.
394  */
395 #define	ENQUEUE(el, head, tail, link) {				\
396 	ASSERT(el->link == NULL);				\
397 	LISTCHECK(head, tail, link);				\
398 	if (head == NULL)					\
399 		head = el;					\
400 	else							\
401 		tail->link = el;				\
402 	tail = el;						\
403 }
404 
405 /*
406  * Dequeue the first element of the list denoted by `head' and `tail' pointers
407  * using a `link' field and put result into `el'.
408  */
409 #define	DQ(el, head, tail, link) {				\
410 	LISTCHECK(head, tail, link);				\
411 	el = head;						\
412 	if (head != NULL) {					\
413 		head = head->link;				\
414 		if (head == NULL)				\
415 			tail = NULL;				\
416 		el->link = NULL;				\
417 	}							\
418 }
419 
420 /*
421  * Remove `el' from the list using `chase' and `curr' pointers and return result
422  * in `succeed'.
423  */
424 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
425 	LISTCHECK(head, tail, link);				\
426 	chase = NULL;						\
427 	succeed = 0;						\
428 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
429 		chase = curr;					\
430 	if (curr != NULL) {					\
431 		succeed = 1;					\
432 		ASSERT(curr == el);				\
433 		if (chase != NULL)				\
434 			chase->link = curr->link;		\
435 		else						\
436 			head = curr->link;			\
437 		curr->link = NULL;				\
438 		if (curr == tail)				\
439 			tail = chase;				\
440 	}							\
441 	LISTCHECK(head, tail, link);				\
442 }
443 
444 /* Handling of delayed messages on the inner syncq. */
445 
446 /*
447  * DEBUG versions should use function versions (to simplify tracing) and
448  * non-DEBUG kernels should use macro versions.
449  */
450 
451 /*
452  * Put a queue on the syncq list of queues.
453  * Assumes SQLOCK held.
454  */
455 #define	SQPUT_Q(sq, qp)							\
456 {									\
457 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
458 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
459 		/* The queue should not be linked anywhere */		\
460 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
461 		/* Head and tail may only be NULL simultaneously */	\
462 		EQUIV(sq->sq_head, sq->sq_tail);			\
463 		/* Queue may be only enqueyed on its syncq */		\
464 		ASSERT(sq == qp->q_syncq);				\
465 		/* Check the correctness of SQ_MESSAGES flag */		\
466 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
467 		/* Sanity check first/last elements of the list */	\
468 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
469 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
470 		/*							\
471 		 * Sanity check of priority field: empty queue should	\
472 		 * have zero priority					\
473 		 * and nqueues equal to zero.				\
474 		 */							\
475 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
476 		/* Sanity check of sq_nqueues field */			\
477 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
478 		if (sq->sq_head == NULL) {				\
479 			sq->sq_head = sq->sq_tail = qp;			\
480 			sq->sq_flags |= SQ_MESSAGES;			\
481 		} else if (qp->q_spri == 0) {				\
482 			qp->q_sqprev = sq->sq_tail;			\
483 			sq->sq_tail->q_sqnext = qp;			\
484 			sq->sq_tail = qp;				\
485 		} else {						\
486 			/*						\
487 			 * Put this queue in priority order: higher	\
488 			 * priority gets closer to the head.		\
489 			 */						\
490 			queue_t **qpp = &sq->sq_tail;			\
491 			queue_t *qnext = NULL;				\
492 									\
493 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
494 				qnext = *qpp;				\
495 				qpp = &(*qpp)->q_sqprev;		\
496 			}						\
497 			qp->q_sqnext = qnext;				\
498 			qp->q_sqprev = *qpp;				\
499 			if (*qpp != NULL) {				\
500 				(*qpp)->q_sqnext = qp;			\
501 			} else {					\
502 				sq->sq_head = qp;			\
503 				sq->sq_pri = sq->sq_head->q_spri;	\
504 			}						\
505 			*qpp = qp;					\
506 		}							\
507 		qp->q_sqflags |= Q_SQQUEUED;				\
508 		qp->q_sqtstamp = lbolt;					\
509 		sq->sq_nqueues++;					\
510 	}								\
511 }
512 
513 /*
514  * Remove a queue from the syncq list
515  * Assumes SQLOCK held.
516  */
517 #define	SQRM_Q(sq, qp)							\
518 	{								\
519 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
520 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
521 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
522 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
523 		/* Check that the queue is actually in the list */	\
524 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
525 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
526 		ASSERT(sq->sq_nqueues != 0);				\
527 		if (qp->q_sqprev == NULL) {				\
528 			/* First queue on list, make head q_sqnext */	\
529 			sq->sq_head = qp->q_sqnext;			\
530 		} else {						\
531 			/* Make prev->next == next */			\
532 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
533 		}							\
534 		if (qp->q_sqnext == NULL) {				\
535 			/* Last queue on list, make tail sqprev */	\
536 			sq->sq_tail = qp->q_sqprev;			\
537 		} else {						\
538 			/* Make next->prev == prev */			\
539 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
540 		}							\
541 		/* clear out references on this queue */		\
542 		qp->q_sqprev = qp->q_sqnext = NULL;			\
543 		qp->q_sqflags &= ~Q_SQQUEUED;				\
544 		/* If there is nothing queued, clear SQ_MESSAGES */	\
545 		if (sq->sq_head != NULL) {				\
546 			sq->sq_pri = sq->sq_head->q_spri;		\
547 		} else	{						\
548 			sq->sq_flags &= ~SQ_MESSAGES;			\
549 			sq->sq_pri = 0;					\
550 		}							\
551 		sq->sq_nqueues--;					\
552 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
553 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
554 	}
555 
556 /* Hide the definition from the header file. */
557 #ifdef SQPUT_MP
558 #undef SQPUT_MP
559 #endif
560 
561 /*
562  * Put a message on the queue syncq.
563  * Assumes QLOCK held.
564  */
565 #define	SQPUT_MP(qp, mp)						\
566 	{								\
567 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
568 		ASSERT(qp->q_sqhead == NULL ||				\
569 		    (qp->q_sqtail != NULL &&				\
570 		    qp->q_sqtail->b_next == NULL));			\
571 		qp->q_syncqmsgs++;					\
572 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
573 		if (qp->q_sqhead == NULL) {				\
574 			qp->q_sqhead = qp->q_sqtail = mp;		\
575 		} else {						\
576 			qp->q_sqtail->b_next = mp;			\
577 			qp->q_sqtail = mp;				\
578 		}							\
579 		ASSERT(qp->q_syncqmsgs > 0);				\
580 	}
581 
582 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
583 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
584 		if ((sq)->sq_ciputctrl != NULL) {			\
585 			int i;						\
586 			int nlocks = (sq)->sq_nciputctrl;		\
587 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
588 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
589 			for (i = 0; i <= nlocks; i++) {			\
590 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
591 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
592 			}						\
593 		}							\
594 	}
595 
596 
597 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
598 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
599 		if ((sq)->sq_ciputctrl != NULL) {			\
600 			int i;						\
601 			int nlocks = (sq)->sq_nciputctrl;		\
602 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
603 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
604 			for (i = 0; i <= nlocks; i++) {			\
605 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
606 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
607 			}						\
608 		}							\
609 	}
610 
611 /*
612  * Run service procedures for all queues in the stream head.
613  */
614 #define	STR_SERVICE(stp, q) {						\
615 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
616 	while (stp->sd_qhead != NULL) {					\
617 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
618 		ASSERT(stp->sd_nqueues > 0);				\
619 		stp->sd_nqueues--;					\
620 		ASSERT(!(q->q_flag & QINSERVICE));			\
621 		mutex_exit(&stp->sd_qlock);				\
622 		queue_service(q);					\
623 		mutex_enter(&stp->sd_qlock);				\
624 	}								\
625 	ASSERT(stp->sd_nqueues == 0);					\
626 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
627 }
628 
629 /*
630  * constructor/destructor routines for the stream head cache
631  */
632 /* ARGSUSED */
633 static int
634 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
635 {
636 	stdata_t *stp = buf;
637 
638 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
639 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
640 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
641 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
642 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
643 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
644 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
645 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
646 	stp->sd_wrq = NULL;
647 
648 	return (0);
649 }
650 
651 /* ARGSUSED */
652 static void
653 stream_head_destructor(void *buf, void *cdrarg)
654 {
655 	stdata_t *stp = buf;
656 
657 	mutex_destroy(&stp->sd_lock);
658 	mutex_destroy(&stp->sd_reflock);
659 	mutex_destroy(&stp->sd_qlock);
660 	cv_destroy(&stp->sd_monitor);
661 	cv_destroy(&stp->sd_iocmonitor);
662 	cv_destroy(&stp->sd_refmonitor);
663 	cv_destroy(&stp->sd_qcv);
664 	cv_destroy(&stp->sd_zcopy_wait);
665 }
666 
667 /*
668  * constructor/destructor routines for the queue cache
669  */
670 /* ARGSUSED */
671 static int
672 queue_constructor(void *buf, void *cdrarg, int kmflags)
673 {
674 	queinfo_t *qip = buf;
675 	queue_t *qp = &qip->qu_rqueue;
676 	queue_t *wqp = &qip->qu_wqueue;
677 	syncq_t	*sq = &qip->qu_syncq;
678 
679 	qp->q_first = NULL;
680 	qp->q_link = NULL;
681 	qp->q_count = 0;
682 	qp->q_mblkcnt = 0;
683 	qp->q_sqhead = NULL;
684 	qp->q_sqtail = NULL;
685 	qp->q_sqnext = NULL;
686 	qp->q_sqprev = NULL;
687 	qp->q_sqflags = 0;
688 	qp->q_rwcnt = 0;
689 	qp->q_spri = 0;
690 
691 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
692 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
693 
694 	wqp->q_first = NULL;
695 	wqp->q_link = NULL;
696 	wqp->q_count = 0;
697 	wqp->q_mblkcnt = 0;
698 	wqp->q_sqhead = NULL;
699 	wqp->q_sqtail = NULL;
700 	wqp->q_sqnext = NULL;
701 	wqp->q_sqprev = NULL;
702 	wqp->q_sqflags = 0;
703 	wqp->q_rwcnt = 0;
704 	wqp->q_spri = 0;
705 
706 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
707 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
708 
709 	sq->sq_head = NULL;
710 	sq->sq_tail = NULL;
711 	sq->sq_evhead = NULL;
712 	sq->sq_evtail = NULL;
713 	sq->sq_callbpend = NULL;
714 	sq->sq_outer = NULL;
715 	sq->sq_onext = NULL;
716 	sq->sq_oprev = NULL;
717 	sq->sq_next = NULL;
718 	sq->sq_svcflags = 0;
719 	sq->sq_servcount = 0;
720 	sq->sq_needexcl = 0;
721 	sq->sq_nqueues = 0;
722 	sq->sq_pri = 0;
723 
724 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
725 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
726 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
727 
728 	return (0);
729 }
730 
731 /* ARGSUSED */
732 static void
733 queue_destructor(void *buf, void *cdrarg)
734 {
735 	queinfo_t *qip = buf;
736 	queue_t *qp = &qip->qu_rqueue;
737 	queue_t *wqp = &qip->qu_wqueue;
738 	syncq_t	*sq = &qip->qu_syncq;
739 
740 	ASSERT(qp->q_sqhead == NULL);
741 	ASSERT(wqp->q_sqhead == NULL);
742 	ASSERT(qp->q_sqnext == NULL);
743 	ASSERT(wqp->q_sqnext == NULL);
744 	ASSERT(qp->q_rwcnt == 0);
745 	ASSERT(wqp->q_rwcnt == 0);
746 
747 	mutex_destroy(&qp->q_lock);
748 	cv_destroy(&qp->q_wait);
749 
750 	mutex_destroy(&wqp->q_lock);
751 	cv_destroy(&wqp->q_wait);
752 
753 	mutex_destroy(&sq->sq_lock);
754 	cv_destroy(&sq->sq_wait);
755 	cv_destroy(&sq->sq_exitwait);
756 }
757 
758 /*
759  * constructor/destructor routines for the syncq cache
760  */
761 /* ARGSUSED */
762 static int
763 syncq_constructor(void *buf, void *cdrarg, int kmflags)
764 {
765 	syncq_t	*sq = buf;
766 
767 	bzero(buf, sizeof (syncq_t));
768 
769 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
770 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
771 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
772 
773 	return (0);
774 }
775 
776 /* ARGSUSED */
777 static void
778 syncq_destructor(void *buf, void *cdrarg)
779 {
780 	syncq_t	*sq = buf;
781 
782 	ASSERT(sq->sq_head == NULL);
783 	ASSERT(sq->sq_tail == NULL);
784 	ASSERT(sq->sq_evhead == NULL);
785 	ASSERT(sq->sq_evtail == NULL);
786 	ASSERT(sq->sq_callbpend == NULL);
787 	ASSERT(sq->sq_callbflags == 0);
788 	ASSERT(sq->sq_outer == NULL);
789 	ASSERT(sq->sq_onext == NULL);
790 	ASSERT(sq->sq_oprev == NULL);
791 	ASSERT(sq->sq_next == NULL);
792 	ASSERT(sq->sq_needexcl == 0);
793 	ASSERT(sq->sq_svcflags == 0);
794 	ASSERT(sq->sq_servcount == 0);
795 	ASSERT(sq->sq_nqueues == 0);
796 	ASSERT(sq->sq_pri == 0);
797 	ASSERT(sq->sq_count == 0);
798 	ASSERT(sq->sq_rmqcount == 0);
799 	ASSERT(sq->sq_cancelid == 0);
800 	ASSERT(sq->sq_ciputctrl == NULL);
801 	ASSERT(sq->sq_nciputctrl == 0);
802 	ASSERT(sq->sq_type == 0);
803 	ASSERT(sq->sq_flags == 0);
804 
805 	mutex_destroy(&sq->sq_lock);
806 	cv_destroy(&sq->sq_wait);
807 	cv_destroy(&sq->sq_exitwait);
808 }
809 
810 /* ARGSUSED */
811 static int
812 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
813 {
814 	ciputctrl_t *cip = buf;
815 	int i;
816 
817 	for (i = 0; i < n_ciputctrl; i++) {
818 		cip[i].ciputctrl_count = SQ_FASTPUT;
819 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
820 	}
821 
822 	return (0);
823 }
824 
825 /* ARGSUSED */
826 static void
827 ciputctrl_destructor(void *buf, void *cdrarg)
828 {
829 	ciputctrl_t *cip = buf;
830 	int i;
831 
832 	for (i = 0; i < n_ciputctrl; i++) {
833 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
834 		mutex_destroy(&cip[i].ciputctrl_lock);
835 	}
836 }
837 
838 /*
839  * Init routine run from main at boot time.
840  */
841 void
842 strinit(void)
843 {
844 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
845 
846 	stream_head_cache = kmem_cache_create("stream_head_cache",
847 		sizeof (stdata_t), 0,
848 		stream_head_constructor, stream_head_destructor, NULL,
849 		NULL, NULL, 0);
850 
851 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
852 		queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
853 
854 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
855 		syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
856 
857 	qband_cache = kmem_cache_create("qband_cache",
858 		sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
859 
860 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
861 		sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
862 
863 	n_ciputctrl = ncpus;
864 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
865 	ASSERT(n_ciputctrl >= 1);
866 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
867 	if (n_ciputctrl >= min_n_ciputctrl) {
868 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
869 			sizeof (ciputctrl_t) * n_ciputctrl,
870 			sizeof (ciputctrl_t), ciputctrl_constructor,
871 			ciputctrl_destructor, NULL, NULL, NULL, 0);
872 	}
873 
874 	streams_taskq = system_taskq;
875 
876 	if (streams_taskq == NULL)
877 		panic("strinit: no memory for streams taskq!");
878 
879 	bc_bkgrnd_thread = thread_create(NULL, 0,
880 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
881 
882 	streams_qbkgrnd_thread = thread_create(NULL, 0,
883 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
884 
885 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
886 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
887 
888 	/*
889 	 * Create STREAMS kstats.
890 	 */
891 	str_kstat = kstat_create("streams", 0, "strstat",
892 	    "net", KSTAT_TYPE_NAMED,
893 	    sizeof (str_statistics) / sizeof (kstat_named_t),
894 	    KSTAT_FLAG_VIRTUAL);
895 
896 	if (str_kstat != NULL) {
897 		str_kstat->ks_data = &str_statistics;
898 		kstat_install(str_kstat);
899 	}
900 
901 	/*
902 	 * TPI support routine initialisation.
903 	 */
904 	tpi_init();
905 
906 	/*
907 	 * Handle to have autopush and persistent link information per
908 	 * zone.
909 	 * Note: uses shutdown hook instead of destroy hook so that the
910 	 * persistent links can be torn down before the destroy hooks
911 	 * in the TCP/IP stack are called.
912 	 */
913 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
914 	    str_stack_fini);
915 }
916 
917 void
918 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
919 {
920 	struct stdata *stp;
921 
922 	ASSERT(vp->v_stream);
923 	stp = vp->v_stream;
924 	/* Have to hold sd_lock to prevent siglist from changing */
925 	mutex_enter(&stp->sd_lock);
926 	if (stp->sd_sigflags & event)
927 		strsendsig(stp->sd_siglist, event, band, error);
928 	mutex_exit(&stp->sd_lock);
929 }
930 
931 /*
932  * Send the "sevent" set of signals to a process.
933  * This might send more than one signal if the process is registered
934  * for multiple events. The caller should pass in an sevent that only
935  * includes the events for which the process has registered.
936  */
937 static void
938 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
939 	uchar_t band, int error)
940 {
941 	ASSERT(MUTEX_HELD(&proc->p_lock));
942 
943 	info->si_band = 0;
944 	info->si_errno = 0;
945 
946 	if (sevent & S_ERROR) {
947 		sevent &= ~S_ERROR;
948 		info->si_code = POLL_ERR;
949 		info->si_errno = error;
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 		info->si_errno = 0;
954 	}
955 	if (sevent & S_HANGUP) {
956 		sevent &= ~S_HANGUP;
957 		info->si_code = POLL_HUP;
958 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
959 			"strsendsig:proc %p info %p", proc, info);
960 		sigaddq(proc, NULL, info, KM_NOSLEEP);
961 	}
962 	if (sevent & S_HIPRI) {
963 		sevent &= ~S_HIPRI;
964 		info->si_code = POLL_PRI;
965 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
966 			"strsendsig:proc %p info %p", proc, info);
967 		sigaddq(proc, NULL, info, KM_NOSLEEP);
968 	}
969 	if (sevent & S_RDBAND) {
970 		sevent &= ~S_RDBAND;
971 		if (events & S_BANDURG)
972 			sigtoproc(proc, NULL, SIGURG);
973 		else
974 			sigtoproc(proc, NULL, SIGPOLL);
975 	}
976 	if (sevent & S_WRBAND) {
977 		sevent &= ~S_WRBAND;
978 		sigtoproc(proc, NULL, SIGPOLL);
979 	}
980 	if (sevent & S_INPUT) {
981 		sevent &= ~S_INPUT;
982 		info->si_code = POLL_IN;
983 		info->si_band = band;
984 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
985 			"strsendsig:proc %p info %p", proc, info);
986 		sigaddq(proc, NULL, info, KM_NOSLEEP);
987 		info->si_band = 0;
988 	}
989 	if (sevent & S_OUTPUT) {
990 		sevent &= ~S_OUTPUT;
991 		info->si_code = POLL_OUT;
992 		info->si_band = band;
993 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
994 			"strsendsig:proc %p info %p", proc, info);
995 		sigaddq(proc, NULL, info, KM_NOSLEEP);
996 		info->si_band = 0;
997 	}
998 	if (sevent & S_MSG) {
999 		sevent &= ~S_MSG;
1000 		info->si_code = POLL_MSG;
1001 		info->si_band = band;
1002 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1003 			"strsendsig:proc %p info %p", proc, info);
1004 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1005 		info->si_band = 0;
1006 	}
1007 	if (sevent & S_RDNORM) {
1008 		sevent &= ~S_RDNORM;
1009 		sigtoproc(proc, NULL, SIGPOLL);
1010 	}
1011 	if (sevent != 0) {
1012 		panic("strsendsig: unknown event(s) %x", sevent);
1013 	}
1014 }
1015 
1016 /*
1017  * Send SIGPOLL/SIGURG signal to all processes and process groups
1018  * registered on the given signal list that want a signal for at
1019  * least one of the specified events.
1020  *
1021  * Must be called with exclusive access to siglist (caller holding sd_lock).
1022  *
1023  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1024  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1025  * while it is in the siglist.
1026  *
1027  * For performance reasons (MP scalability) the code drops pidlock
1028  * when sending signals to a single process.
1029  * When sending to a process group the code holds
1030  * pidlock to prevent the membership in the process group from changing
1031  * while walking the p_pglink list.
1032  */
1033 void
1034 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1035 {
1036 	strsig_t *ssp;
1037 	k_siginfo_t info;
1038 	struct pid *pidp;
1039 	proc_t  *proc;
1040 
1041 	info.si_signo = SIGPOLL;
1042 	info.si_errno = 0;
1043 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1044 		int sevent;
1045 
1046 		sevent = ssp->ss_events & event;
1047 		if (sevent == 0)
1048 			continue;
1049 
1050 		if ((pidp = ssp->ss_pidp) == NULL) {
1051 			/* pid was released but still on event list */
1052 			continue;
1053 		}
1054 
1055 
1056 		if (ssp->ss_pid > 0) {
1057 			/*
1058 			 * XXX This unfortunately still generates
1059 			 * a signal when a fd is closed but
1060 			 * the proc is active.
1061 			 */
1062 			ASSERT(ssp->ss_pid == pidp->pid_id);
1063 
1064 			mutex_enter(&pidlock);
1065 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1066 			if (proc == NULL) {
1067 				mutex_exit(&pidlock);
1068 				continue;
1069 			}
1070 			mutex_enter(&proc->p_lock);
1071 			mutex_exit(&pidlock);
1072 			dosendsig(proc, ssp->ss_events, sevent, &info,
1073 				band, error);
1074 			mutex_exit(&proc->p_lock);
1075 		} else {
1076 			/*
1077 			 * Send to process group. Hold pidlock across
1078 			 * calls to dosendsig().
1079 			 */
1080 			pid_t pgrp = -ssp->ss_pid;
1081 
1082 			mutex_enter(&pidlock);
1083 			proc = pgfind_zone(pgrp, ALL_ZONES);
1084 			while (proc != NULL) {
1085 				mutex_enter(&proc->p_lock);
1086 				dosendsig(proc, ssp->ss_events, sevent,
1087 					&info, band, error);
1088 				mutex_exit(&proc->p_lock);
1089 				proc = proc->p_pglink;
1090 			}
1091 			mutex_exit(&pidlock);
1092 		}
1093 	}
1094 }
1095 
1096 /*
1097  * Attach a stream device or module.
1098  * qp is a read queue; the new queue goes in so its next
1099  * read ptr is the argument, and the write queue corresponding
1100  * to the argument points to this queue. Return 0 on success,
1101  * or a non-zero errno on failure.
1102  */
1103 int
1104 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1105     boolean_t is_insert)
1106 {
1107 	major_t			major;
1108 	cdevsw_impl_t		*dp;
1109 	struct streamtab	*str;
1110 	queue_t			*rq;
1111 	queue_t			*wrq;
1112 	uint32_t		qflag;
1113 	uint32_t		sqtype;
1114 	perdm_t			*dmp;
1115 	int			error;
1116 	int			sflag;
1117 
1118 	rq = allocq();
1119 	wrq = _WR(rq);
1120 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1121 
1122 	if (fp != NULL) {
1123 		str = fp->f_str;
1124 		qflag = fp->f_qflag;
1125 		sqtype = fp->f_sqtype;
1126 		dmp = fp->f_dmp;
1127 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1128 		sflag = MODOPEN;
1129 
1130 		/*
1131 		 * stash away a pointer to the module structure so we can
1132 		 * unref it in qdetach.
1133 		 */
1134 		rq->q_fp = fp;
1135 	} else {
1136 		ASSERT(!is_insert);
1137 
1138 		major = getmajor(*devp);
1139 		dp = &devimpl[major];
1140 
1141 		str = dp->d_str;
1142 		ASSERT(str == STREAMSTAB(major));
1143 
1144 		qflag = dp->d_qflag;
1145 		ASSERT(qflag & QISDRV);
1146 		sqtype = dp->d_sqtype;
1147 
1148 		/* create perdm_t if needed */
1149 		if (NEED_DM(dp->d_dmp, qflag))
1150 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1151 
1152 		dmp = dp->d_dmp;
1153 		sflag = 0;
1154 	}
1155 
1156 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1157 	    "qattach:qflag == %X(%X)", qflag, *devp);
1158 
1159 	/* setq might sleep in allocator - avoid holding locks. */
1160 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1161 
1162 	/*
1163 	 * Before calling the module's open routine, set up the q_next
1164 	 * pointer for inserting a module in the middle of a stream.
1165 	 *
1166 	 * Note that we can always set _QINSERTING and set up q_next
1167 	 * pointer for both inserting and pushing a module.  Then there
1168 	 * is no need for the is_insert parameter.  In insertq(), called
1169 	 * by qprocson(), assume that q_next of the new module always points
1170 	 * to the correct queue and use it for insertion.  Everything should
1171 	 * work out fine.  But in the first release of _I_INSERT, we
1172 	 * distinguish between inserting and pushing to make sure that
1173 	 * pushing a module follows the same code path as before.
1174 	 */
1175 	if (is_insert) {
1176 		rq->q_flag |= _QINSERTING;
1177 		rq->q_next = qp;
1178 	}
1179 
1180 	/*
1181 	 * If there is an outer perimeter get exclusive access during
1182 	 * the open procedure.  Bump up the reference count on the queue.
1183 	 */
1184 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1185 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1186 	if (error != 0)
1187 		goto failed;
1188 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1189 	ASSERT(qprocsareon(rq));
1190 	return (0);
1191 
1192 failed:
1193 	rq->q_flag &= ~_QINSERTING;
1194 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1195 		qprocsoff(rq);
1196 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1197 	rq->q_next = wrq->q_next = NULL;
1198 	qdetach(rq, 0, 0, crp, B_FALSE);
1199 	return (error);
1200 }
1201 
1202 /*
1203  * Handle second open of stream. For modules, set the
1204  * last argument to MODOPEN and do not pass any open flags.
1205  * Ignore dummydev since this is not the first open.
1206  */
1207 int
1208 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1209 {
1210 	int	error;
1211 	dev_t dummydev;
1212 	queue_t *wqp = _WR(qp);
1213 
1214 	ASSERT(qp->q_flag & QREADR);
1215 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1216 
1217 	dummydev = *devp;
1218 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1219 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1220 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1221 		mutex_enter(&STREAM(qp)->sd_lock);
1222 		qp->q_stream->sd_flag |= STREOPENFAIL;
1223 		mutex_exit(&STREAM(qp)->sd_lock);
1224 		return (error);
1225 	}
1226 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1227 
1228 	/*
1229 	 * successful open should have done qprocson()
1230 	 */
1231 	ASSERT(qprocsareon(_RD(qp)));
1232 	return (0);
1233 }
1234 
1235 /*
1236  * Detach a stream module or device.
1237  * If clmode == 1 then the module or driver was opened and its
1238  * close routine must be called. If clmode == 0, the module
1239  * or driver was never opened or the open failed, and so its close
1240  * should not be called.
1241  */
1242 void
1243 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1244 {
1245 	queue_t *wqp = _WR(qp);
1246 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1247 
1248 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1249 		stream_runservice(STREAM(qp));
1250 
1251 	if (clmode) {
1252 		/*
1253 		 * Make sure that all the messages on the write side syncq are
1254 		 * processed and nothing is left. Since we are closing, no new
1255 		 * messages may appear there.
1256 		 */
1257 		wait_q_syncq(wqp);
1258 
1259 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1260 		if (is_remove) {
1261 			mutex_enter(QLOCK(qp));
1262 			qp->q_flag |= _QREMOVING;
1263 			mutex_exit(QLOCK(qp));
1264 		}
1265 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1266 		/*
1267 		 * Check that qprocsoff() was actually called.
1268 		 */
1269 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1270 
1271 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1272 	} else {
1273 		disable_svc(qp);
1274 	}
1275 
1276 	/*
1277 	 * Allow any threads blocked in entersq to proceed and discover
1278 	 * the QWCLOSE is set.
1279 	 * Note: This assumes that all users of entersq check QWCLOSE.
1280 	 * Currently runservice is the only entersq that can happen
1281 	 * after removeq has finished.
1282 	 * Removeq will have discarded all messages destined to the closing
1283 	 * pair of queues from the syncq.
1284 	 * NOTE: Calling a function inside an assert is unconventional.
1285 	 * However, it does not cause any problem since flush_syncq() does
1286 	 * not change any state except when it returns non-zero i.e.
1287 	 * when the assert will trigger.
1288 	 */
1289 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1290 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1291 	ASSERT((qp->q_flag & QPERMOD) ||
1292 		((qp->q_syncq->sq_head == NULL) &&
1293 		(wqp->q_syncq->sq_head == NULL)));
1294 
1295 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1296 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1297 	if (qp->q_fp != NULL)
1298 		fmodsw_rele(qp->q_fp);
1299 
1300 	/* freeq removes us from the outer perimeter if any */
1301 	freeq(qp);
1302 }
1303 
1304 /* Prevent service procedures from being called */
1305 void
1306 disable_svc(queue_t *qp)
1307 {
1308 	queue_t *wqp = _WR(qp);
1309 
1310 	ASSERT(qp->q_flag & QREADR);
1311 	mutex_enter(QLOCK(qp));
1312 	qp->q_flag |= QWCLOSE;
1313 	mutex_exit(QLOCK(qp));
1314 	mutex_enter(QLOCK(wqp));
1315 	wqp->q_flag |= QWCLOSE;
1316 	mutex_exit(QLOCK(wqp));
1317 }
1318 
1319 /* allow service procedures to be called again */
1320 void
1321 enable_svc(queue_t *qp)
1322 {
1323 	queue_t *wqp = _WR(qp);
1324 
1325 	ASSERT(qp->q_flag & QREADR);
1326 	mutex_enter(QLOCK(qp));
1327 	qp->q_flag &= ~QWCLOSE;
1328 	mutex_exit(QLOCK(qp));
1329 	mutex_enter(QLOCK(wqp));
1330 	wqp->q_flag &= ~QWCLOSE;
1331 	mutex_exit(QLOCK(wqp));
1332 }
1333 
1334 /*
1335  * Remove queue from qhead/qtail if it is enabled.
1336  * Only reset QENAB if the queue was removed from the runlist.
1337  * A queue goes through 3 stages:
1338  *	It is on the service list and QENAB is set.
1339  *	It is removed from the service list but QENAB is still set.
1340  *	QENAB gets changed to QINSERVICE.
1341  *	QINSERVICE is reset (when the service procedure is done)
1342  * Thus we can not reset QENAB unless we actually removed it from the service
1343  * queue.
1344  */
1345 void
1346 remove_runlist(queue_t *qp)
1347 {
1348 	if (qp->q_flag & QENAB && qhead != NULL) {
1349 		queue_t *q_chase;
1350 		queue_t *q_curr;
1351 		int removed;
1352 
1353 		mutex_enter(&service_queue);
1354 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1355 		mutex_exit(&service_queue);
1356 		if (removed) {
1357 			STRSTAT(qremoved);
1358 			qp->q_flag &= ~QENAB;
1359 		}
1360 	}
1361 }
1362 
1363 
1364 /*
1365  * wait for any pending service processing to complete.
1366  * The removal of queues from the runlist is not atomic with the
1367  * clearing of the QENABLED flag and setting the INSERVICE flag.
1368  * consequently it is possible for remove_runlist in strclose
1369  * to not find the queue on the runlist but for it to be QENABLED
1370  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1371  * as well as INSERVICE.
1372  */
1373 void
1374 wait_svc(queue_t *qp)
1375 {
1376 	queue_t *wqp = _WR(qp);
1377 
1378 	ASSERT(qp->q_flag & QREADR);
1379 
1380 	/*
1381 	 * Try to remove queues from qhead/qtail list.
1382 	 */
1383 	if (qhead != NULL) {
1384 		remove_runlist(qp);
1385 		remove_runlist(wqp);
1386 	}
1387 	/*
1388 	 * Wait till the syncqs associated with the queue
1389 	 * will dissapear from background processing list.
1390 	 * This only needs to be done for non-PERMOD perimeters since
1391 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1392 	 * when the last module/driver is unloaded.
1393 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1394 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1395 	 * function remove the queue from its syncq list, so sqthread will not
1396 	 * try to access the queue.
1397 	 */
1398 	if (!(qp->q_flag & QPERMOD)) {
1399 		syncq_t *rsq = qp->q_syncq;
1400 		syncq_t *wsq = wqp->q_syncq;
1401 
1402 		/*
1403 		 * Disable rsq and wsq and wait for any background processing of
1404 		 * syncq to complete.
1405 		 */
1406 		wait_sq_svc(rsq);
1407 		if (wsq != rsq)
1408 			wait_sq_svc(wsq);
1409 	}
1410 
1411 	mutex_enter(QLOCK(qp));
1412 	while (qp->q_flag & (QINSERVICE|QENAB))
1413 		cv_wait(&qp->q_wait, QLOCK(qp));
1414 	mutex_exit(QLOCK(qp));
1415 	mutex_enter(QLOCK(wqp));
1416 	while (wqp->q_flag & (QINSERVICE|QENAB))
1417 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1418 	mutex_exit(QLOCK(wqp));
1419 }
1420 
1421 /*
1422  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1423  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1424  * also be set, and is passed through to allocb_cred_wait().
1425  *
1426  * Returns errno on failure, zero on success.
1427  */
1428 int
1429 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1430 {
1431 	mblk_t *tmp;
1432 	ssize_t  count;
1433 	size_t n;
1434 	int error = 0;
1435 
1436 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1437 		(flag & (U_TO_K | K_TO_K)) == K_TO_K);
1438 
1439 	if (bp->b_datap->db_type == M_IOCTL) {
1440 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1441 	} else {
1442 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1443 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1444 	}
1445 	/*
1446 	 * strdoioctl validates ioc_count, so if this assert fails it
1447 	 * cannot be due to user error.
1448 	 */
1449 	ASSERT(count >= 0);
1450 
1451 	while (count > 0) {
1452 		n = MIN(MAXIOCBSZ, count);
1453 		if ((tmp = allocb_cred_wait(n, (flag & STR_NOSIG), &error,
1454 		    cr)) == NULL) {
1455 			return (error);
1456 		}
1457 		error = strcopyin(arg, tmp->b_wptr, n, flag & (U_TO_K|K_TO_K));
1458 		if (error != 0) {
1459 			freeb(tmp);
1460 			return (error);
1461 		}
1462 		arg += n;
1463 		DB_CPID(tmp) = curproc->p_pid;
1464 		tmp->b_wptr += n;
1465 		count -= n;
1466 		bp = (bp->b_cont = tmp);
1467 	}
1468 
1469 	return (0);
1470 }
1471 
1472 /*
1473  * Copy ioctl data to user-land. Return non-zero errno on failure,
1474  * 0 for success.
1475  */
1476 int
1477 getiocd(mblk_t *bp, char *arg, int copymode)
1478 {
1479 	ssize_t count;
1480 	size_t  n;
1481 	int	error;
1482 
1483 	if (bp->b_datap->db_type == M_IOCACK)
1484 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1485 	else {
1486 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1487 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1488 	}
1489 	ASSERT(count >= 0);
1490 
1491 	for (bp = bp->b_cont; bp && count;
1492 	    count -= n, bp = bp->b_cont, arg += n) {
1493 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1494 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1495 		if (error)
1496 			return (error);
1497 	}
1498 	ASSERT(count == 0);
1499 	return (0);
1500 }
1501 
1502 /*
1503  * Allocate a linkinfo entry given the write queue of the
1504  * bottom module of the top stream and the write queue of the
1505  * stream head of the bottom stream.
1506  */
1507 linkinfo_t *
1508 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1509 {
1510 	linkinfo_t *linkp;
1511 
1512 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1513 
1514 	linkp->li_lblk.l_qtop = qup;
1515 	linkp->li_lblk.l_qbot = qdown;
1516 	linkp->li_fpdown = fpdown;
1517 
1518 	mutex_enter(&strresources);
1519 	linkp->li_next = linkinfo_list;
1520 	linkp->li_prev = NULL;
1521 	if (linkp->li_next)
1522 		linkp->li_next->li_prev = linkp;
1523 	linkinfo_list = linkp;
1524 	linkp->li_lblk.l_index = ++lnk_id;
1525 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1526 	mutex_exit(&strresources);
1527 
1528 	return (linkp);
1529 }
1530 
1531 /*
1532  * Free a linkinfo entry.
1533  */
1534 void
1535 lbfree(linkinfo_t *linkp)
1536 {
1537 	mutex_enter(&strresources);
1538 	if (linkp->li_next)
1539 		linkp->li_next->li_prev = linkp->li_prev;
1540 	if (linkp->li_prev)
1541 		linkp->li_prev->li_next = linkp->li_next;
1542 	else
1543 		linkinfo_list = linkp->li_next;
1544 	mutex_exit(&strresources);
1545 
1546 	kmem_cache_free(linkinfo_cache, linkp);
1547 }
1548 
1549 /*
1550  * Check for a potential linking cycle.
1551  * Return 1 if a link will result in a cycle,
1552  * and 0 otherwise.
1553  */
1554 int
1555 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1556 {
1557 	struct mux_node *np;
1558 	struct mux_edge *ep;
1559 	int i;
1560 	major_t lomaj;
1561 	major_t upmaj;
1562 	/*
1563 	 * if the lower stream is a pipe/FIFO, return, since link
1564 	 * cycles can not happen on pipes/FIFOs
1565 	 */
1566 	if (lostp->sd_vnode->v_type == VFIFO)
1567 		return (0);
1568 
1569 	for (i = 0; i < ss->ss_devcnt; i++) {
1570 		np = &ss->ss_mux_nodes[i];
1571 		MUX_CLEAR(np);
1572 	}
1573 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1574 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1575 	np = &ss->ss_mux_nodes[lomaj];
1576 	for (;;) {
1577 		if (!MUX_DIDVISIT(np)) {
1578 			if (np->mn_imaj == upmaj)
1579 				return (1);
1580 			if (np->mn_outp == NULL) {
1581 				MUX_VISIT(np);
1582 				if (np->mn_originp == NULL)
1583 					return (0);
1584 				np = np->mn_originp;
1585 				continue;
1586 			}
1587 			MUX_VISIT(np);
1588 			np->mn_startp = np->mn_outp;
1589 		} else {
1590 			if (np->mn_startp == NULL) {
1591 				if (np->mn_originp == NULL)
1592 					return (0);
1593 				else {
1594 					np = np->mn_originp;
1595 					continue;
1596 				}
1597 			}
1598 			/*
1599 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1600 			 * ignore the edge and move on. ep->me_nodep gets
1601 			 * set to NULL in mux_addedge() if it is a FIFO.
1602 			 *
1603 			 */
1604 			ep = np->mn_startp;
1605 			np->mn_startp = ep->me_nextp;
1606 			if (ep->me_nodep == NULL)
1607 				continue;
1608 			ep->me_nodep->mn_originp = np;
1609 			np = ep->me_nodep;
1610 		}
1611 	}
1612 }
1613 
1614 /*
1615  * Find linkinfo entry corresponding to the parameters.
1616  */
1617 linkinfo_t *
1618 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1619 {
1620 	linkinfo_t *linkp;
1621 	struct mux_edge *mep;
1622 	struct mux_node *mnp;
1623 	queue_t *qup;
1624 
1625 	mutex_enter(&strresources);
1626 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1627 		qup = getendq(stp->sd_wrq);
1628 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1629 			if ((qup == linkp->li_lblk.l_qtop) &&
1630 			    (!index || (index == linkp->li_lblk.l_index))) {
1631 				mutex_exit(&strresources);
1632 				return (linkp);
1633 			}
1634 		}
1635 	} else {
1636 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1637 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1638 		mep = mnp->mn_outp;
1639 		while (mep) {
1640 			if ((index == 0) || (index == mep->me_muxid))
1641 				break;
1642 			mep = mep->me_nextp;
1643 		}
1644 		if (!mep) {
1645 			mutex_exit(&strresources);
1646 			return (NULL);
1647 		}
1648 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1649 			if ((!linkp->li_lblk.l_qtop) &&
1650 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1651 				mutex_exit(&strresources);
1652 				return (linkp);
1653 			}
1654 		}
1655 	}
1656 	mutex_exit(&strresources);
1657 	return (NULL);
1658 }
1659 
1660 /*
1661  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1662  * last queue on the chain and return it.
1663  */
1664 queue_t *
1665 getendq(queue_t *q)
1666 {
1667 	ASSERT(q != NULL);
1668 	while (_SAMESTR(q))
1669 		q = q->q_next;
1670 	return (q);
1671 }
1672 
1673 /*
1674  * wait for the syncq count to drop to zero.
1675  * sq could be either outer or inner.
1676  */
1677 
1678 static void
1679 wait_syncq(syncq_t *sq)
1680 {
1681 	uint16_t count;
1682 
1683 	mutex_enter(SQLOCK(sq));
1684 	count = sq->sq_count;
1685 	SQ_PUTLOCKS_ENTER(sq);
1686 	SUM_SQ_PUTCOUNTS(sq, count);
1687 	while (count != 0) {
1688 		sq->sq_flags |= SQ_WANTWAKEUP;
1689 		SQ_PUTLOCKS_EXIT(sq);
1690 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1691 		count = sq->sq_count;
1692 		SQ_PUTLOCKS_ENTER(sq);
1693 		SUM_SQ_PUTCOUNTS(sq, count);
1694 	}
1695 	SQ_PUTLOCKS_EXIT(sq);
1696 	mutex_exit(SQLOCK(sq));
1697 }
1698 
1699 /*
1700  * Wait while there are any messages for the queue in its syncq.
1701  */
1702 static void
1703 wait_q_syncq(queue_t *q)
1704 {
1705 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1706 		syncq_t *sq = q->q_syncq;
1707 
1708 		mutex_enter(SQLOCK(sq));
1709 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1710 			sq->sq_flags |= SQ_WANTWAKEUP;
1711 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1712 		}
1713 		mutex_exit(SQLOCK(sq));
1714 	}
1715 }
1716 
1717 
1718 int
1719 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1720     int lhlink)
1721 {
1722 	struct stdata *stp;
1723 	struct strioctl strioc;
1724 	struct linkinfo *linkp;
1725 	struct stdata *stpdown;
1726 	struct streamtab *str;
1727 	queue_t *passq;
1728 	syncq_t *passyncq;
1729 	queue_t *rq;
1730 	cdevsw_impl_t *dp;
1731 	uint32_t qflag;
1732 	uint32_t sqtype;
1733 	perdm_t *dmp;
1734 	int error = 0;
1735 	netstack_t *ns;
1736 	str_stack_t *ss;
1737 
1738 	stp = vp->v_stream;
1739 	TRACE_1(TR_FAC_STREAMS_FR,
1740 		TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1741 	/*
1742 	 * Test for invalid upper stream
1743 	 */
1744 	if (stp->sd_flag & STRHUP) {
1745 		return (ENXIO);
1746 	}
1747 	if (vp->v_type == VFIFO) {
1748 		return (EINVAL);
1749 	}
1750 	if (stp->sd_strtab == NULL) {
1751 		return (EINVAL);
1752 	}
1753 	if (!stp->sd_strtab->st_muxwinit) {
1754 		return (EINVAL);
1755 	}
1756 	if (fpdown == NULL) {
1757 		return (EBADF);
1758 	}
1759 	ns = netstack_find_by_cred(crp);
1760 	ASSERT(ns != NULL);
1761 	ss = ns->netstack_str;
1762 	ASSERT(ss != NULL);
1763 
1764 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1765 		netstack_rele(ss->ss_netstack);
1766 		return (EINVAL);
1767 	}
1768 	mutex_enter(&muxifier);
1769 	if (stp->sd_flag & STPLEX) {
1770 		mutex_exit(&muxifier);
1771 		netstack_rele(ss->ss_netstack);
1772 		return (ENXIO);
1773 	}
1774 
1775 	/*
1776 	 * Test for invalid lower stream.
1777 	 * The check for the v_type != VFIFO and having a major
1778 	 * number not >= devcnt is done to avoid problems with
1779 	 * adding mux_node entry past the end of mux_nodes[].
1780 	 * For FIFO's we don't add an entry so this isn't a
1781 	 * problem.
1782 	 */
1783 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1784 	    (stpdown == stp) || (stpdown->sd_flag &
1785 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1786 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1787 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1788 	    linkcycle(stp, stpdown, ss)) {
1789 		mutex_exit(&muxifier);
1790 		netstack_rele(ss->ss_netstack);
1791 		return (EINVAL);
1792 	}
1793 	TRACE_1(TR_FAC_STREAMS_FR,
1794 		TR_STPDOWN, "stpdown:%p", stpdown);
1795 	rq = getendq(stp->sd_wrq);
1796 	if (cmd == I_PLINK)
1797 		rq = NULL;
1798 
1799 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1800 
1801 	strioc.ic_cmd = cmd;
1802 	strioc.ic_timout = INFTIM;
1803 	strioc.ic_len = sizeof (struct linkblk);
1804 	strioc.ic_dp = (char *)&linkp->li_lblk;
1805 
1806 	/*
1807 	 * STRPLUMB protects plumbing changes and should be set before
1808 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1809 	 * and cleared in the end of mlink when passthru queue is removed.
1810 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1811 	 * queue is in-place (it is not a proper module and doesn't have open
1812 	 * entry point).
1813 	 *
1814 	 * STPLEX prevents any threads from entering the stream from above. It
1815 	 * can't be set before the call to link_addpassthru() because putnext
1816 	 * from below may cause stream head I/O routines to be called and these
1817 	 * routines assert that STPLEX is not set. After link_addpassthru()
1818 	 * nothing may come from below since the pass queue syncq is blocked.
1819 	 * Note also that STPLEX should be cleared before the call to
1820 	 * link_remmpassthru() since when messages start flowing to the stream
1821 	 * head (e.g. because of message propagation from the pass queue) stream
1822 	 * head I/O routines may be called with STPLEX flag set.
1823 	 *
1824 	 * When STPLEX is set, nothing may come into the stream from above and
1825 	 * it is safe to do a setq which will change stream head. So, the
1826 	 * correct sequence of actions is:
1827 	 *
1828 	 * 1) Set STRPLUMB
1829 	 * 2) Call link_addpassthru()
1830 	 * 3) Set STPLEX
1831 	 * 4) Call setq and update the stream state
1832 	 * 5) Clear STPLEX
1833 	 * 6) Call link_rempassthru()
1834 	 * 7) Clear STRPLUMB
1835 	 *
1836 	 * The same sequence applies to munlink() code.
1837 	 */
1838 	mutex_enter(&stpdown->sd_lock);
1839 	stpdown->sd_flag |= STRPLUMB;
1840 	mutex_exit(&stpdown->sd_lock);
1841 	/*
1842 	 * Add passthru queue below lower mux. This will block
1843 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1844 	 */
1845 	passq = link_addpassthru(stpdown);
1846 
1847 	mutex_enter(&stpdown->sd_lock);
1848 	stpdown->sd_flag |= STPLEX;
1849 	mutex_exit(&stpdown->sd_lock);
1850 
1851 	rq = _RD(stpdown->sd_wrq);
1852 	/*
1853 	 * There may be messages in the streamhead's syncq due to messages
1854 	 * that arrived before link_addpassthru() was done. To avoid
1855 	 * background processing of the syncq happening simultaneous with
1856 	 * setq processing, we disable the streamhead syncq and wait until
1857 	 * existing background thread finishes working on it.
1858 	 */
1859 	wait_sq_svc(rq->q_syncq);
1860 	passyncq = passq->q_syncq;
1861 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1862 		blocksq(passyncq, SQ_BLOCKED, 0);
1863 
1864 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1865 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1866 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1867 
1868 	/* setq might sleep in allocator - avoid holding locks. */
1869 	/* Note: we are holding muxifier here. */
1870 
1871 	str = stp->sd_strtab;
1872 	dp = &devimpl[getmajor(vp->v_rdev)];
1873 	ASSERT(dp->d_str == str);
1874 
1875 	qflag = dp->d_qflag;
1876 	sqtype = dp->d_sqtype;
1877 
1878 	/* create perdm_t if needed */
1879 	if (NEED_DM(dp->d_dmp, qflag))
1880 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1881 
1882 	dmp = dp->d_dmp;
1883 
1884 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1885 	    B_TRUE);
1886 
1887 	/*
1888 	 * XXX Remove any "odd" messages from the queue.
1889 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1890 	 */
1891 	error = strdoioctl(stp, &strioc, FNATIVE,
1892 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1893 	if (error != 0) {
1894 		lbfree(linkp);
1895 
1896 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1897 			blocksq(passyncq, SQ_BLOCKED, 0);
1898 		/*
1899 		 * Restore the stream head queue and then remove
1900 		 * the passq. Turn off STPLEX before we turn on
1901 		 * the stream by removing the passq.
1902 		 */
1903 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1904 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1905 		    B_TRUE);
1906 
1907 		mutex_enter(&stpdown->sd_lock);
1908 		stpdown->sd_flag &= ~STPLEX;
1909 		mutex_exit(&stpdown->sd_lock);
1910 
1911 		link_rempassthru(passq);
1912 
1913 		mutex_enter(&stpdown->sd_lock);
1914 		stpdown->sd_flag &= ~STRPLUMB;
1915 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1916 		cv_broadcast(&stpdown->sd_monitor);
1917 		mutex_exit(&stpdown->sd_lock);
1918 
1919 		mutex_exit(&muxifier);
1920 		netstack_rele(ss->ss_netstack);
1921 		return (error);
1922 	}
1923 	mutex_enter(&fpdown->f_tlock);
1924 	fpdown->f_count++;
1925 	mutex_exit(&fpdown->f_tlock);
1926 
1927 	/*
1928 	 * if we've made it here the linkage is all set up so we should also
1929 	 * set up the layered driver linkages
1930 	 */
1931 
1932 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1933 	if (cmd == I_LINK) {
1934 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1935 	} else {
1936 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1937 	}
1938 
1939 	link_rempassthru(passq);
1940 
1941 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1942 
1943 	/*
1944 	 * Mark the upper stream as having dependent links
1945 	 * so that strclose can clean it up.
1946 	 */
1947 	if (cmd == I_LINK) {
1948 		mutex_enter(&stp->sd_lock);
1949 		stp->sd_flag |= STRHASLINKS;
1950 		mutex_exit(&stp->sd_lock);
1951 	}
1952 	/*
1953 	 * Wake up any other processes that may have been
1954 	 * waiting on the lower stream. These will all
1955 	 * error out.
1956 	 */
1957 	mutex_enter(&stpdown->sd_lock);
1958 	/* The passthru module is removed so we may release STRPLUMB */
1959 	stpdown->sd_flag &= ~STRPLUMB;
1960 	cv_broadcast(&rq->q_wait);
1961 	cv_broadcast(&_WR(rq)->q_wait);
1962 	cv_broadcast(&stpdown->sd_monitor);
1963 	mutex_exit(&stpdown->sd_lock);
1964 	mutex_exit(&muxifier);
1965 	*rvalp = linkp->li_lblk.l_index;
1966 	netstack_rele(ss->ss_netstack);
1967 	return (0);
1968 }
1969 
1970 int
1971 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1972 {
1973 	int		ret;
1974 	struct file	*fpdown;
1975 
1976 	fpdown = getf(arg);
1977 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1978 	if (fpdown != NULL)
1979 		releasef(arg);
1980 	return (ret);
1981 }
1982 
1983 /*
1984  * Unlink a multiplexor link. Stp is the controlling stream for the
1985  * link, and linkp points to the link's entry in the linkinfo list.
1986  * The muxifier lock must be held on entry and is dropped on exit.
1987  *
1988  * NOTE : Currently it is assumed that mux would process all the messages
1989  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
1990  * of the mux to handle all the messages that arrive before UNLINK.
1991  * If the mux has to send down messages on its lower stream before
1992  * ACKing I_UNLINK, then it *should* know to handle messages even
1993  * after the UNLINK is acked (actually it should be able to handle till we
1994  * re-block the read side of the pass queue here). If the mux does not
1995  * open up the lower stream, any messages that arrive during UNLINK
1996  * will be put in the stream head. In the case of lower stream opening
1997  * up, some messages might land in the stream head depending on when
1998  * the message arrived and when the read side of the pass queue was
1999  * re-blocked.
2000  */
2001 int
2002 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2003     str_stack_t *ss)
2004 {
2005 	struct strioctl strioc;
2006 	struct stdata *stpdown;
2007 	queue_t *rq, *wrq;
2008 	queue_t	*passq;
2009 	syncq_t *passyncq;
2010 	int error = 0;
2011 	file_t *fpdown;
2012 
2013 	ASSERT(MUTEX_HELD(&muxifier));
2014 
2015 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2016 
2017 	/*
2018 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2019 	 */
2020 	mutex_enter(&stpdown->sd_lock);
2021 	stpdown->sd_flag |= STRPLUMB;
2022 	mutex_exit(&stpdown->sd_lock);
2023 
2024 	/*
2025 	 * Add passthru queue below lower mux. This will block
2026 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2027 	 */
2028 	passq = link_addpassthru(stpdown);
2029 
2030 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2031 		strioc.ic_cmd = I_UNLINK;
2032 	else
2033 		strioc.ic_cmd = I_PUNLINK;
2034 	strioc.ic_timout = INFTIM;
2035 	strioc.ic_len = sizeof (struct linkblk);
2036 	strioc.ic_dp = (char *)&linkp->li_lblk;
2037 
2038 	error = strdoioctl(stp, &strioc, FNATIVE,
2039 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2040 
2041 	/*
2042 	 * If there was an error and this is not called via strclose,
2043 	 * return to the user. Otherwise, pretend there was no error
2044 	 * and close the link.
2045 	 */
2046 	if (error) {
2047 		if (flag & LINKCLOSE) {
2048 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2049 			    "unlink ioctl, closing anyway (%d)\n", error);
2050 		} else {
2051 			link_rempassthru(passq);
2052 			mutex_enter(&stpdown->sd_lock);
2053 			stpdown->sd_flag &= ~STRPLUMB;
2054 			cv_broadcast(&stpdown->sd_monitor);
2055 			mutex_exit(&stpdown->sd_lock);
2056 			mutex_exit(&muxifier);
2057 			return (error);
2058 		}
2059 	}
2060 
2061 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2062 	fpdown = linkp->li_fpdown;
2063 	lbfree(linkp);
2064 
2065 	/*
2066 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2067 	 * can slow others down. It's okay to since attempts to mlink() this
2068 	 * stream will be stopped because STPLEX is still set in the stdata
2069 	 * structure, and munlink() is stopped because mux_rmvedge() and
2070 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2071 	 * respectively.  Note that we defer the closef() of fpdown until
2072 	 * after we drop muxifier since strclose() can call munlinkall().
2073 	 */
2074 	mutex_exit(&muxifier);
2075 
2076 	wrq = stpdown->sd_wrq;
2077 	rq = _RD(wrq);
2078 
2079 	/*
2080 	 * Get rid of outstanding service procedure runs, before we make
2081 	 * it a stream head, since a stream head doesn't have any service
2082 	 * procedure.
2083 	 */
2084 	disable_svc(rq);
2085 	wait_svc(rq);
2086 
2087 	/*
2088 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2089 	 * is queued up to be finished. mux should take care that nothing is
2090 	 * send down to this queue. We should do it now as we're going to block
2091 	 * passyncq if it was unblocked.
2092 	 */
2093 	if (wrq->q_flag & QPERMOD) {
2094 		syncq_t	*sq = wrq->q_syncq;
2095 
2096 		mutex_enter(SQLOCK(sq));
2097 		while (wrq->q_sqflags & Q_SQQUEUED) {
2098 			sq->sq_flags |= SQ_WANTWAKEUP;
2099 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2100 		}
2101 		mutex_exit(SQLOCK(sq));
2102 	}
2103 	passyncq = passq->q_syncq;
2104 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2105 
2106 		syncq_t *sq, *outer;
2107 
2108 		/*
2109 		 * Messages could be flowing from underneath. We will
2110 		 * block the read side of the passq. This would be
2111 		 * sufficient for QPAIR and QPERQ muxes to ensure
2112 		 * that no data is flowing up into this queue
2113 		 * and hence no thread active in this instance of
2114 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2115 		 * could be messages on the inner and outer/inner
2116 		 * syncqs respectively. We will wait for them to drain.
2117 		 * Because passq is blocked messages end up in the syncq
2118 		 * And qfill_syncq could possibly end up setting QFULL
2119 		 * which will access the rq->q_flag. Hence, we have to
2120 		 * acquire the QLOCK in setq.
2121 		 *
2122 		 * XXX Messages can also flow from top into this
2123 		 * queue though the unlink is over (Ex. some instance
2124 		 * in putnext() called from top that has still not
2125 		 * accessed this queue. And also putq(lowerq) ?).
2126 		 * Solution : How about blocking the l_qtop queue ?
2127 		 * Do we really care about such pure D_MP muxes ?
2128 		 */
2129 
2130 		blocksq(passyncq, SQ_BLOCKED, 0);
2131 
2132 		sq = rq->q_syncq;
2133 		if ((outer = sq->sq_outer) != NULL) {
2134 
2135 			/*
2136 			 * We have to just wait for the outer sq_count
2137 			 * drop to zero. As this does not prevent new
2138 			 * messages to enter the outer perimeter, this
2139 			 * is subject to starvation.
2140 			 *
2141 			 * NOTE :Because of blocksq above, messages could
2142 			 * be in the inner syncq only because of some
2143 			 * thread holding the outer perimeter exclusively.
2144 			 * Hence it would be sufficient to wait for the
2145 			 * exclusive holder of the outer perimeter to drain
2146 			 * the inner and outer syncqs. But we will not depend
2147 			 * on this feature and hence check the inner syncqs
2148 			 * separately.
2149 			 */
2150 			wait_syncq(outer);
2151 		}
2152 
2153 
2154 		/*
2155 		 * There could be messages destined for
2156 		 * this queue. Let the exclusive holder
2157 		 * drain it.
2158 		 */
2159 
2160 		wait_syncq(sq);
2161 		ASSERT((rq->q_flag & QPERMOD) ||
2162 			((rq->q_syncq->sq_head == NULL) &&
2163 			(_WR(rq)->q_syncq->sq_head == NULL)));
2164 	}
2165 
2166 	/*
2167 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2168 	 * case as we don't disable its syncq or remove it off the syncq
2169 	 * service list.
2170 	 */
2171 	if (rq->q_flag & QPERMOD) {
2172 		syncq_t	*sq = rq->q_syncq;
2173 
2174 		mutex_enter(SQLOCK(sq));
2175 		while (rq->q_sqflags & Q_SQQUEUED) {
2176 			sq->sq_flags |= SQ_WANTWAKEUP;
2177 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2178 		}
2179 		mutex_exit(SQLOCK(sq));
2180 	}
2181 
2182 	/*
2183 	 * flush_syncq changes states only when there is some messages to
2184 	 * free. ie when it returns non-zero value to return.
2185 	 */
2186 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2187 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2188 
2189 	/*
2190 	 * No body else should know about this queue now.
2191 	 * If the mux did not process the messages before
2192 	 * acking the I_UNLINK, free them now.
2193 	 */
2194 
2195 	flushq(rq, FLUSHALL);
2196 	flushq(_WR(rq), FLUSHALL);
2197 
2198 	/*
2199 	 * Convert the mux lower queue into a stream head queue.
2200 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2201 	 */
2202 	rq->q_ptr = wrq->q_ptr = stpdown;
2203 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2204 
2205 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2206 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2207 
2208 	enable_svc(rq);
2209 
2210 	/*
2211 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2212 	 * needs to be set to prevent reopen() of the stream - such reopen may
2213 	 * try to call non-existent pass queue open routine and panic.
2214 	 */
2215 	mutex_enter(&stpdown->sd_lock);
2216 	stpdown->sd_flag &= ~STPLEX;
2217 	mutex_exit(&stpdown->sd_lock);
2218 
2219 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2220 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2221 
2222 	/* clean up the layered driver linkages */
2223 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2224 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2225 	} else {
2226 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2227 	}
2228 
2229 	link_rempassthru(passq);
2230 
2231 	/*
2232 	 * Now all plumbing changes are finished and STRPLUMB is no
2233 	 * longer needed.
2234 	 */
2235 	mutex_enter(&stpdown->sd_lock);
2236 	stpdown->sd_flag &= ~STRPLUMB;
2237 	cv_broadcast(&stpdown->sd_monitor);
2238 	mutex_exit(&stpdown->sd_lock);
2239 
2240 	(void) closef(fpdown);
2241 	return (0);
2242 }
2243 
2244 /*
2245  * Unlink all multiplexor links for which stp is the controlling stream.
2246  * Return 0, or a non-zero errno on failure.
2247  */
2248 int
2249 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2250 {
2251 	linkinfo_t *linkp;
2252 	int error = 0;
2253 
2254 	mutex_enter(&muxifier);
2255 	while (linkp = findlinks(stp, 0, flag, ss)) {
2256 		/*
2257 		 * munlink() releases the muxifier lock.
2258 		 */
2259 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2260 			return (error);
2261 		mutex_enter(&muxifier);
2262 	}
2263 	mutex_exit(&muxifier);
2264 	return (0);
2265 }
2266 
2267 /*
2268  * A multiplexor link has been made. Add an
2269  * edge to the directed graph.
2270  */
2271 void
2272 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2273 {
2274 	struct mux_node *np;
2275 	struct mux_edge *ep;
2276 	major_t upmaj;
2277 	major_t lomaj;
2278 
2279 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2280 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2281 	np = &ss->ss_mux_nodes[upmaj];
2282 	if (np->mn_outp) {
2283 		ep = np->mn_outp;
2284 		while (ep->me_nextp)
2285 			ep = ep->me_nextp;
2286 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2287 		ep = ep->me_nextp;
2288 	} else {
2289 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2290 		ep = np->mn_outp;
2291 	}
2292 	ep->me_nextp = NULL;
2293 	ep->me_muxid = muxid;
2294 	/*
2295 	 * Save the dev_t for the purposes of str_stack_shutdown.
2296 	 * str_stack_shutdown assumes that the device allows reopen, since
2297 	 * this dev_t is the one after any cloning by xx_open().
2298 	 * Would prefer finding the dev_t from before any cloning,
2299 	 * but specfs doesn't retain that.
2300 	 */
2301 	ep->me_dev = upstp->sd_vnode->v_rdev;
2302 	if (lostp->sd_vnode->v_type == VFIFO)
2303 		ep->me_nodep = NULL;
2304 	else
2305 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2306 }
2307 
2308 /*
2309  * A multiplexor link has been removed. Remove the
2310  * edge in the directed graph.
2311  */
2312 void
2313 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2314 {
2315 	struct mux_node *np;
2316 	struct mux_edge *ep;
2317 	struct mux_edge *pep = NULL;
2318 	major_t upmaj;
2319 
2320 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2321 	np = &ss->ss_mux_nodes[upmaj];
2322 	ASSERT(np->mn_outp != NULL);
2323 	ep = np->mn_outp;
2324 	while (ep) {
2325 		if (ep->me_muxid == muxid) {
2326 			if (pep)
2327 				pep->me_nextp = ep->me_nextp;
2328 			else
2329 				np->mn_outp = ep->me_nextp;
2330 			kmem_free(ep, sizeof (struct mux_edge));
2331 			return;
2332 		}
2333 		pep = ep;
2334 		ep = ep->me_nextp;
2335 	}
2336 	ASSERT(0);	/* should not reach here */
2337 }
2338 
2339 /*
2340  * Translate the device flags (from conf.h) to the corresponding
2341  * qflag and sq_flag (type) values.
2342  */
2343 int
2344 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2345 	uint32_t *sqtypep)
2346 {
2347 	uint32_t qflag = 0;
2348 	uint32_t sqtype = 0;
2349 
2350 	if (devflag & _D_OLD)
2351 		goto bad;
2352 
2353 	/* Inner perimeter presence and scope */
2354 	switch (devflag & D_MTINNER_MASK) {
2355 	case D_MP:
2356 		qflag |= QMTSAFE;
2357 		sqtype |= SQ_CI;
2358 		break;
2359 	case D_MTPERQ|D_MP:
2360 		qflag |= QPERQ;
2361 		break;
2362 	case D_MTQPAIR|D_MP:
2363 		qflag |= QPAIR;
2364 		break;
2365 	case D_MTPERMOD|D_MP:
2366 		qflag |= QPERMOD;
2367 		break;
2368 	default:
2369 		goto bad;
2370 	}
2371 
2372 	/* Outer perimeter */
2373 	if (devflag & D_MTOUTPERIM) {
2374 		switch (devflag & D_MTINNER_MASK) {
2375 		case D_MP:
2376 		case D_MTPERQ|D_MP:
2377 		case D_MTQPAIR|D_MP:
2378 			break;
2379 		default:
2380 			goto bad;
2381 		}
2382 		qflag |= QMTOUTPERIM;
2383 	}
2384 
2385 	/* Inner perimeter modifiers */
2386 	if (devflag & D_MTINNER_MOD) {
2387 		switch (devflag & D_MTINNER_MASK) {
2388 		case D_MP:
2389 			goto bad;
2390 		default:
2391 			break;
2392 		}
2393 		if (devflag & D_MTPUTSHARED)
2394 			sqtype |= SQ_CIPUT;
2395 		if (devflag & _D_MTOCSHARED) {
2396 			/*
2397 			 * The code in putnext assumes that it has the
2398 			 * highest concurrency by not checking sq_count.
2399 			 * Thus _D_MTOCSHARED can only be supported when
2400 			 * D_MTPUTSHARED is set.
2401 			 */
2402 			if (!(devflag & D_MTPUTSHARED))
2403 				goto bad;
2404 			sqtype |= SQ_CIOC;
2405 		}
2406 		if (devflag & _D_MTCBSHARED) {
2407 			/*
2408 			 * The code in putnext assumes that it has the
2409 			 * highest concurrency by not checking sq_count.
2410 			 * Thus _D_MTCBSHARED can only be supported when
2411 			 * D_MTPUTSHARED is set.
2412 			 */
2413 			if (!(devflag & D_MTPUTSHARED))
2414 				goto bad;
2415 			sqtype |= SQ_CICB;
2416 		}
2417 		if (devflag & _D_MTSVCSHARED) {
2418 			/*
2419 			 * The code in putnext assumes that it has the
2420 			 * highest concurrency by not checking sq_count.
2421 			 * Thus _D_MTSVCSHARED can only be supported when
2422 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2423 			 * supported only for QPERMOD.
2424 			 */
2425 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2426 				goto bad;
2427 			sqtype |= SQ_CISVC;
2428 		}
2429 	}
2430 
2431 	/* Default outer perimeter concurrency */
2432 	sqtype |= SQ_CO;
2433 
2434 	/* Outer perimeter modifiers */
2435 	if (devflag & D_MTOCEXCL) {
2436 		if (!(devflag & D_MTOUTPERIM)) {
2437 			/* No outer perimeter */
2438 			goto bad;
2439 		}
2440 		sqtype &= ~SQ_COOC;
2441 	}
2442 
2443 	/* Synchronous Streams extended qinit structure */
2444 	if (devflag & D_SYNCSTR)
2445 		qflag |= QSYNCSTR;
2446 
2447 	/*
2448 	 * Private flag used by a transport module to indicate
2449 	 * to sockfs that it supports direct-access mode without
2450 	 * having to go through STREAMS.
2451 	 */
2452 	if (devflag & _D_DIRECT) {
2453 		/* Reject unless the module is fully-MT (no perimeter) */
2454 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2455 			goto bad;
2456 		qflag |= _QDIRECT;
2457 	}
2458 
2459 	*qflagp = qflag;
2460 	*sqtypep = sqtype;
2461 	return (0);
2462 
2463 bad:
2464 	cmn_err(CE_WARN,
2465 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2466 	    (int)(qflag & D_MTSAFETY_MASK),
2467 	    stp->st_rdinit->qi_minfo->mi_idname);
2468 
2469 	return (EINVAL);
2470 }
2471 
2472 /*
2473  * Set the interface values for a pair of queues (qinit structure,
2474  * packet sizes, water marks).
2475  * setq assumes that the caller does not have a claim (entersq or claimq)
2476  * on the queue.
2477  */
2478 void
2479 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2480     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2481 {
2482 	queue_t *wq;
2483 	syncq_t	*sq, *outer;
2484 
2485 	ASSERT(rq->q_flag & QREADR);
2486 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2487 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2488 
2489 	wq = _WR(rq);
2490 	rq->q_qinfo = rinit;
2491 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2492 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2493 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2494 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2495 	wq->q_qinfo = winit;
2496 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2497 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2498 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2499 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2500 
2501 	/* Remove old syncqs */
2502 	sq = rq->q_syncq;
2503 	outer = sq->sq_outer;
2504 	if (outer != NULL) {
2505 		ASSERT(wq->q_syncq->sq_outer == outer);
2506 		outer_remove(outer, rq->q_syncq);
2507 		if (wq->q_syncq != rq->q_syncq)
2508 			outer_remove(outer, wq->q_syncq);
2509 	}
2510 	ASSERT(sq->sq_outer == NULL);
2511 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2512 
2513 	if (sq != SQ(rq)) {
2514 		if (!(rq->q_flag & QPERMOD))
2515 			free_syncq(sq);
2516 		if (wq->q_syncq == rq->q_syncq)
2517 			wq->q_syncq = NULL;
2518 		rq->q_syncq = NULL;
2519 	}
2520 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2521 	    wq->q_syncq != SQ(rq)) {
2522 		free_syncq(wq->q_syncq);
2523 		wq->q_syncq = NULL;
2524 	}
2525 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2526 				rq->q_syncq->sq_tail == NULL));
2527 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2528 				wq->q_syncq->sq_tail == NULL));
2529 
2530 	if (!(rq->q_flag & QPERMOD) &&
2531 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2532 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2533 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2534 		    rq->q_syncq->sq_nciputctrl, 0);
2535 		ASSERT(ciputctrl_cache != NULL);
2536 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2537 		rq->q_syncq->sq_ciputctrl = NULL;
2538 		rq->q_syncq->sq_nciputctrl = 0;
2539 	}
2540 
2541 	if (!(wq->q_flag & QPERMOD) &&
2542 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2543 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2544 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2545 		    wq->q_syncq->sq_nciputctrl, 0);
2546 		ASSERT(ciputctrl_cache != NULL);
2547 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2548 		wq->q_syncq->sq_ciputctrl = NULL;
2549 		wq->q_syncq->sq_nciputctrl = 0;
2550 	}
2551 
2552 	sq = SQ(rq);
2553 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2554 	ASSERT(sq->sq_outer == NULL);
2555 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2556 
2557 	/*
2558 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2559 	 * bits in sq_flag based on the sqtype.
2560 	 */
2561 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2562 
2563 	rq->q_syncq = wq->q_syncq = sq;
2564 	sq->sq_type = sqtype;
2565 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2566 
2567 	/*
2568 	 *  We are making sq_svcflags zero,
2569 	 *  resetting SQ_DISABLED in case it was set by
2570 	 *  wait_svc() in the munlink path.
2571 	 *
2572 	 */
2573 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2574 	sq->sq_svcflags = 0;
2575 
2576 	/*
2577 	 * We need to acquire the lock here for the mlink and munlink case,
2578 	 * where canputnext, backenable, etc can access the q_flag.
2579 	 */
2580 	if (lock_needed) {
2581 		mutex_enter(QLOCK(rq));
2582 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2583 		mutex_exit(QLOCK(rq));
2584 		mutex_enter(QLOCK(wq));
2585 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2586 		mutex_exit(QLOCK(wq));
2587 	} else {
2588 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2589 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2590 	}
2591 
2592 	if (qflag & QPERQ) {
2593 		/* Allocate a separate syncq for the write side */
2594 		sq = new_syncq();
2595 		sq->sq_type = rq->q_syncq->sq_type;
2596 		sq->sq_flags = rq->q_syncq->sq_flags;
2597 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2598 		    sq->sq_oprev == NULL);
2599 		wq->q_syncq = sq;
2600 	}
2601 	if (qflag & QPERMOD) {
2602 		sq = dmp->dm_sq;
2603 
2604 		/*
2605 		 * Assert that we do have an inner perimeter syncq and that it
2606 		 * does not have an outer perimeter associated with it.
2607 		 */
2608 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2609 		    sq->sq_oprev == NULL);
2610 		rq->q_syncq = wq->q_syncq = sq;
2611 	}
2612 	if (qflag & QMTOUTPERIM) {
2613 		outer = dmp->dm_sq;
2614 
2615 		ASSERT(outer->sq_outer == NULL);
2616 		outer_insert(outer, rq->q_syncq);
2617 		if (wq->q_syncq != rq->q_syncq)
2618 			outer_insert(outer, wq->q_syncq);
2619 	}
2620 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2621 		(rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2622 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2623 		(wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2624 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2625 
2626 	/*
2627 	 * Initialize struio() types.
2628 	 */
2629 	rq->q_struiot =
2630 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2631 	wq->q_struiot =
2632 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2633 }
2634 
2635 perdm_t *
2636 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2637 {
2638 	syncq_t	*sq;
2639 	perdm_t	**pp;
2640 	perdm_t	*p;
2641 	perdm_t	*dmp;
2642 
2643 	ASSERT(str != NULL);
2644 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2645 
2646 	rw_enter(&perdm_rwlock, RW_READER);
2647 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2648 		if (p->dm_str == str) {	/* found one */
2649 			atomic_add_32(&(p->dm_ref), 1);
2650 			rw_exit(&perdm_rwlock);
2651 			return (p);
2652 		}
2653 	}
2654 	rw_exit(&perdm_rwlock);
2655 
2656 	sq = new_syncq();
2657 	if (qflag & QPERMOD) {
2658 		sq->sq_type = sqtype | SQ_PERMOD;
2659 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2660 	} else {
2661 		ASSERT(qflag & QMTOUTPERIM);
2662 		sq->sq_onext = sq->sq_oprev = sq;
2663 	}
2664 
2665 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2666 	dmp->dm_sq = sq;
2667 	dmp->dm_str = str;
2668 	dmp->dm_ref = 1;
2669 	dmp->dm_next = NULL;
2670 
2671 	rw_enter(&perdm_rwlock, RW_WRITER);
2672 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2673 		if (p->dm_str == str) {	/* already present */
2674 			p->dm_ref++;
2675 			rw_exit(&perdm_rwlock);
2676 			free_syncq(sq);
2677 			kmem_free(dmp, sizeof (perdm_t));
2678 			return (p);
2679 		}
2680 	}
2681 
2682 	*pp = dmp;
2683 	rw_exit(&perdm_rwlock);
2684 	return (dmp);
2685 }
2686 
2687 void
2688 rele_dm(perdm_t *dmp)
2689 {
2690 	perdm_t **pp;
2691 	perdm_t *p;
2692 
2693 	rw_enter(&perdm_rwlock, RW_WRITER);
2694 	ASSERT(dmp->dm_ref > 0);
2695 
2696 	if (--dmp->dm_ref > 0) {
2697 		rw_exit(&perdm_rwlock);
2698 		return;
2699 	}
2700 
2701 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2702 		if (p == dmp)
2703 			break;
2704 	ASSERT(p == dmp);
2705 	*pp = p->dm_next;
2706 	rw_exit(&perdm_rwlock);
2707 
2708 	/*
2709 	 * Wait for any background processing that relies on the
2710 	 * syncq to complete before it is freed.
2711 	 */
2712 	wait_sq_svc(p->dm_sq);
2713 	free_syncq(p->dm_sq);
2714 	kmem_free(p, sizeof (perdm_t));
2715 }
2716 
2717 /*
2718  * Make a protocol message given control and data buffers.
2719  * n.b., this can block; be careful of what locks you hold when calling it.
2720  *
2721  * If sd_maxblk is less than *iosize this routine can fail part way through
2722  * (due to an allocation failure). In this case on return *iosize will contain
2723  * the amount that was consumed. Otherwise *iosize will not be modified
2724  * i.e. it will contain the amount that was consumed.
2725  */
2726 int
2727 strmakemsg(
2728 	struct strbuf *mctl,
2729 	ssize_t *iosize,
2730 	struct uio *uiop,
2731 	stdata_t *stp,
2732 	int32_t flag,
2733 	mblk_t **mpp)
2734 {
2735 	mblk_t *mpctl = NULL;
2736 	mblk_t *mpdata = NULL;
2737 	int error;
2738 
2739 	ASSERT(uiop != NULL);
2740 
2741 	*mpp = NULL;
2742 	/* Create control part, if any */
2743 	if ((mctl != NULL) && (mctl->len >= 0)) {
2744 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2745 		if (error)
2746 			return (error);
2747 	}
2748 	/* Create data part, if any */
2749 	if (*iosize >= 0) {
2750 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2751 		if (error) {
2752 			freemsg(mpctl);
2753 			return (error);
2754 		}
2755 	}
2756 	if (mpctl != NULL) {
2757 		if (mpdata != NULL)
2758 			linkb(mpctl, mpdata);
2759 		*mpp = mpctl;
2760 	} else {
2761 		*mpp = mpdata;
2762 	}
2763 	return (0);
2764 }
2765 
2766 /*
2767  * Make the control part of a protocol message given a control buffer.
2768  * n.b., this can block; be careful of what locks you hold when calling it.
2769  */
2770 int
2771 strmakectl(
2772 	struct strbuf *mctl,
2773 	int32_t flag,
2774 	int32_t fflag,
2775 	mblk_t **mpp)
2776 {
2777 	mblk_t *bp = NULL;
2778 	unsigned char msgtype;
2779 	int error = 0;
2780 
2781 	*mpp = NULL;
2782 	/*
2783 	 * Create control part of message, if any.
2784 	 */
2785 	if ((mctl != NULL) && (mctl->len >= 0)) {
2786 		caddr_t base;
2787 		int ctlcount;
2788 		int allocsz;
2789 
2790 		if (flag & RS_HIPRI)
2791 			msgtype = M_PCPROTO;
2792 		else
2793 			msgtype = M_PROTO;
2794 
2795 		ctlcount = mctl->len;
2796 		base = mctl->buf;
2797 
2798 		/*
2799 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2800 		 * blocks by increasing the size to something more usable.
2801 		 */
2802 		allocsz = MAX(ctlcount, 64);
2803 
2804 		/*
2805 		 * Range checking has already been done; simply try
2806 		 * to allocate a message block for the ctl part.
2807 		 */
2808 		while (!(bp = allocb(allocsz, BPRI_MED))) {
2809 			if (fflag & (FNDELAY|FNONBLOCK))
2810 				return (EAGAIN);
2811 			if (error = strwaitbuf(allocsz, BPRI_MED))
2812 				return (error);
2813 		}
2814 
2815 		bp->b_datap->db_type = msgtype;
2816 		if (copyin(base, bp->b_wptr, ctlcount)) {
2817 			freeb(bp);
2818 			return (EFAULT);
2819 		}
2820 		bp->b_wptr += ctlcount;
2821 	}
2822 	*mpp = bp;
2823 	return (0);
2824 }
2825 
2826 /*
2827  * Make a protocol message given data buffers.
2828  * n.b., this can block; be careful of what locks you hold when calling it.
2829  *
2830  * If sd_maxblk is less than *iosize this routine can fail part way through
2831  * (due to an allocation failure). In this case on return *iosize will contain
2832  * the amount that was consumed. Otherwise *iosize will not be modified
2833  * i.e. it will contain the amount that was consumed.
2834  */
2835 int
2836 strmakedata(
2837 	ssize_t   *iosize,
2838 	struct uio *uiop,
2839 	stdata_t *stp,
2840 	int32_t flag,
2841 	mblk_t **mpp)
2842 {
2843 	mblk_t *mp = NULL;
2844 	mblk_t *bp;
2845 	int wroff = (int)stp->sd_wroff;
2846 	int tail_len = (int)stp->sd_tail;
2847 	int extra = wroff + tail_len;
2848 	int error = 0;
2849 	ssize_t maxblk;
2850 	ssize_t count = *iosize;
2851 	cred_t *cr = CRED();
2852 
2853 	*mpp = NULL;
2854 	if (count < 0)
2855 		return (0);
2856 
2857 	maxblk = stp->sd_maxblk;
2858 	if (maxblk == INFPSZ)
2859 		maxblk = count;
2860 
2861 	/*
2862 	 * Create data part of message, if any.
2863 	 */
2864 	do {
2865 		ssize_t size;
2866 		dblk_t  *dp;
2867 
2868 		ASSERT(uiop);
2869 
2870 		size = MIN(count, maxblk);
2871 
2872 		while ((bp = allocb_cred(size + extra, cr)) == NULL) {
2873 			error = EAGAIN;
2874 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2875 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2876 				if (count == *iosize) {
2877 					freemsg(mp);
2878 					return (error);
2879 				} else {
2880 					*iosize -= count;
2881 					*mpp = mp;
2882 					return (0);
2883 				}
2884 			}
2885 		}
2886 		dp = bp->b_datap;
2887 		dp->db_cpid = curproc->p_pid;
2888 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2889 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2890 
2891 		if (flag & STRUIO_POSTPONE) {
2892 			/*
2893 			 * Setup the stream uio portion of the
2894 			 * dblk for subsequent use by struioget().
2895 			 */
2896 			dp->db_struioflag = STRUIO_SPEC;
2897 			dp->db_cksumstart = 0;
2898 			dp->db_cksumstuff = 0;
2899 			dp->db_cksumend = size;
2900 			*(long long *)dp->db_struioun.data = 0ll;
2901 			bp->b_wptr += size;
2902 		} else {
2903 			if (stp->sd_copyflag & STRCOPYCACHED)
2904 				uiop->uio_extflg |= UIO_COPY_CACHED;
2905 
2906 			if (size != 0) {
2907 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2908 				    uiop);
2909 				if (error != 0) {
2910 					freeb(bp);
2911 					freemsg(mp);
2912 					return (error);
2913 				}
2914 			}
2915 			bp->b_wptr += size;
2916 
2917 			if (stp->sd_wputdatafunc != NULL) {
2918 				mblk_t *newbp;
2919 
2920 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2921 				    bp, NULL, NULL, NULL, NULL);
2922 				if (newbp == NULL) {
2923 					freeb(bp);
2924 					freemsg(mp);
2925 					return (ECOMM);
2926 				}
2927 				bp = newbp;
2928 			}
2929 		}
2930 
2931 		count -= size;
2932 
2933 		if (mp == NULL)
2934 			mp = bp;
2935 		else
2936 			linkb(mp, bp);
2937 	} while (count > 0);
2938 
2939 	*mpp = mp;
2940 	return (0);
2941 }
2942 
2943 /*
2944  * Wait for a buffer to become available. Return non-zero errno
2945  * if not able to wait, 0 if buffer is probably there.
2946  */
2947 int
2948 strwaitbuf(size_t size, int pri)
2949 {
2950 	bufcall_id_t id;
2951 
2952 	mutex_enter(&bcall_monitor);
2953 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2954 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2955 		mutex_exit(&bcall_monitor);
2956 		return (ENOSR);
2957 	}
2958 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2959 		unbufcall(id);
2960 		mutex_exit(&bcall_monitor);
2961 		return (EINTR);
2962 	}
2963 	unbufcall(id);
2964 	mutex_exit(&bcall_monitor);
2965 	return (0);
2966 }
2967 
2968 /*
2969  * This function waits for a read or write event to happen on a stream.
2970  * fmode can specify FNDELAY and/or FNONBLOCK.
2971  * The timeout is in ms with -1 meaning infinite.
2972  * The flag values work as follows:
2973  *	READWAIT	Check for read side errors, send M_READ
2974  *	GETWAIT		Check for read side errors, no M_READ
2975  *	WRITEWAIT	Check for write side errors.
2976  *	NOINTR		Do not return error if nonblocking or timeout.
2977  * 	STR_NOERROR	Ignore all errors except STPLEX.
2978  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
2979  *	STR_PEEK	Pass through the strgeterr().
2980  */
2981 int
2982 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
2983     int *done)
2984 {
2985 	int slpflg, errs;
2986 	int error;
2987 	kcondvar_t *sleepon;
2988 	mblk_t *mp;
2989 	ssize_t *rd_count;
2990 	clock_t rval;
2991 
2992 	ASSERT(MUTEX_HELD(&stp->sd_lock));
2993 	if ((flag & READWAIT) || (flag & GETWAIT)) {
2994 		slpflg = RSLEEP;
2995 		sleepon = &_RD(stp->sd_wrq)->q_wait;
2996 		errs = STRDERR|STPLEX;
2997 	} else {
2998 		slpflg = WSLEEP;
2999 		sleepon = &stp->sd_wrq->q_wait;
3000 		errs = STWRERR|STRHUP|STPLEX;
3001 	}
3002 	if (flag & STR_NOERROR)
3003 		errs = STPLEX;
3004 
3005 	if (stp->sd_wakeq & slpflg) {
3006 		/*
3007 		 * A strwakeq() is pending, no need to sleep.
3008 		 */
3009 		stp->sd_wakeq &= ~slpflg;
3010 		*done = 0;
3011 		return (0);
3012 	}
3013 
3014 	if (fmode & (FNDELAY|FNONBLOCK)) {
3015 		if (!(flag & NOINTR))
3016 			error = EAGAIN;
3017 		else
3018 			error = 0;
3019 		*done = 1;
3020 		return (error);
3021 	}
3022 
3023 	if (stp->sd_flag & errs) {
3024 		/*
3025 		 * Check for errors before going to sleep since the
3026 		 * caller might not have checked this while holding
3027 		 * sd_lock.
3028 		 */
3029 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3030 		if (error != 0) {
3031 			*done = 1;
3032 			return (error);
3033 		}
3034 	}
3035 
3036 	/*
3037 	 * If any module downstream has requested read notification
3038 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3039 	 * down stream.
3040 	 */
3041 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3042 		mutex_exit(&stp->sd_lock);
3043 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3044 		    (flag & STR_NOSIG), &error))) {
3045 			mutex_enter(&stp->sd_lock);
3046 			*done = 1;
3047 			return (error);
3048 		}
3049 		mp->b_datap->db_type = M_READ;
3050 		rd_count = (ssize_t *)mp->b_wptr;
3051 		*rd_count = count;
3052 		mp->b_wptr += sizeof (ssize_t);
3053 		/*
3054 		 * Send the number of bytes requested by the
3055 		 * read as the argument to M_READ.
3056 		 */
3057 		stream_willservice(stp);
3058 		putnext(stp->sd_wrq, mp);
3059 		stream_runservice(stp);
3060 		mutex_enter(&stp->sd_lock);
3061 
3062 		/*
3063 		 * If any data arrived due to inline processing
3064 		 * of putnext(), don't sleep.
3065 		 */
3066 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3067 			*done = 0;
3068 			return (0);
3069 		}
3070 	}
3071 
3072 	stp->sd_flag |= slpflg;
3073 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3074 		"strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3075 		stp, flag, count, fmode, done);
3076 
3077 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3078 	if (rval > 0) {
3079 		/* EMPTY */
3080 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3081 			"strwaitq awakes(2):%X, %X, %X, %X, %X",
3082 			stp, flag, count, fmode, done);
3083 	} else if (rval == 0) {
3084 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3085 			"strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3086 			stp, flag, count, fmode, done);
3087 		stp->sd_flag &= ~slpflg;
3088 		cv_broadcast(sleepon);
3089 		if (!(flag & NOINTR))
3090 			error = EINTR;
3091 		else
3092 			error = 0;
3093 		*done = 1;
3094 		return (error);
3095 	} else {
3096 		/* timeout */
3097 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3098 			"strwaitq timeout:%p, %X, %lX, %X, %p",
3099 			stp, flag, count, fmode, done);
3100 		*done = 1;
3101 		if (!(flag & NOINTR))
3102 			return (ETIME);
3103 		else
3104 			return (0);
3105 	}
3106 	/*
3107 	 * If the caller implements delayed errors (i.e. queued after data)
3108 	 * we can not check for errors here since data as well as an
3109 	 * error might have arrived at the stream head. We return to
3110 	 * have the caller check the read queue before checking for errors.
3111 	 */
3112 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3113 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3114 		if (error != 0) {
3115 			*done = 1;
3116 			return (error);
3117 		}
3118 	}
3119 	*done = 0;
3120 	return (0);
3121 }
3122 
3123 /*
3124  * Perform job control discipline access checks.
3125  * Return 0 for success and the errno for failure.
3126  */
3127 
3128 #define	cantsend(p, t, sig) \
3129 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3130 
3131 int
3132 straccess(struct stdata *stp, enum jcaccess mode)
3133 {
3134 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3135 	kthread_t *t = curthread;
3136 	proc_t *p = ttoproc(t);
3137 	sess_t *sp;
3138 
3139 	ASSERT(mutex_owned(&stp->sd_lock));
3140 
3141 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3142 		return (0);
3143 
3144 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3145 
3146 	for (;;) {
3147 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3148 		sp = p->p_sessp;
3149 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3150 
3151 		/*
3152 		 * If this is not the calling process's controlling terminal
3153 		 * or if the calling process is already in the foreground
3154 		 * then allow access.
3155 		 */
3156 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3157 		    p->p_pgidp == stp->sd_pgidp) {
3158 			mutex_exit(&sp->s_lock);
3159 			mutex_exit(&p->p_splock);
3160 			mutex_exit(&p->p_lock);
3161 			return (0);
3162 		}
3163 
3164 		/*
3165 		 * Check to see if controlling terminal has been deallocated.
3166 		 */
3167 		if (sp->s_vp == NULL) {
3168 			if (!cantsend(p, t, SIGHUP))
3169 				sigtoproc(p, t, SIGHUP);
3170 			mutex_exit(&sp->s_lock);
3171 			mutex_exit(&p->p_splock);
3172 			mutex_exit(&p->p_lock);
3173 			return (EIO);
3174 		}
3175 
3176 		mutex_exit(&sp->s_lock);
3177 		mutex_exit(&p->p_splock);
3178 
3179 		if (mode == JCGETP) {
3180 			mutex_exit(&p->p_lock);
3181 			return (0);
3182 		}
3183 
3184 		if (mode == JCREAD) {
3185 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3186 				mutex_exit(&p->p_lock);
3187 				return (EIO);
3188 			}
3189 			mutex_exit(&p->p_lock);
3190 			mutex_exit(&stp->sd_lock);
3191 			pgsignal(p->p_pgidp, SIGTTIN);
3192 			mutex_enter(&stp->sd_lock);
3193 			mutex_enter(&p->p_lock);
3194 		} else {  /* mode == JCWRITE or JCSETP */
3195 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3196 			    cantsend(p, t, SIGTTOU)) {
3197 				mutex_exit(&p->p_lock);
3198 				return (0);
3199 			}
3200 			if (p->p_detached) {
3201 				mutex_exit(&p->p_lock);
3202 				return (EIO);
3203 			}
3204 			mutex_exit(&p->p_lock);
3205 			mutex_exit(&stp->sd_lock);
3206 			pgsignal(p->p_pgidp, SIGTTOU);
3207 			mutex_enter(&stp->sd_lock);
3208 			mutex_enter(&p->p_lock);
3209 		}
3210 
3211 		/*
3212 		 * We call cv_wait_sig_swap() to cause the appropriate
3213 		 * action for the jobcontrol signal to take place.
3214 		 * If the signal is being caught, we will take the
3215 		 * EINTR error return.  Otherwise, the default action
3216 		 * of causing the process to stop will take place.
3217 		 * In this case, we rely on the periodic cv_broadcast() on
3218 		 * &lbolt_cv to wake us up to loop around and test again.
3219 		 * We can't get here if the signal is ignored or
3220 		 * if the current thread is blocking the signal.
3221 		 */
3222 		mutex_exit(&stp->sd_lock);
3223 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3224 			mutex_exit(&p->p_lock);
3225 			mutex_enter(&stp->sd_lock);
3226 			return (EINTR);
3227 		}
3228 		mutex_exit(&p->p_lock);
3229 		mutex_enter(&stp->sd_lock);
3230 		mutex_enter(&p->p_lock);
3231 	}
3232 }
3233 
3234 /*
3235  * Return size of message of block type (bp->b_datap->db_type)
3236  */
3237 size_t
3238 xmsgsize(mblk_t *bp)
3239 {
3240 	unsigned char type;
3241 	size_t count = 0;
3242 
3243 	type = bp->b_datap->db_type;
3244 
3245 	for (; bp; bp = bp->b_cont) {
3246 		if (type != bp->b_datap->db_type)
3247 			break;
3248 		ASSERT(bp->b_wptr >= bp->b_rptr);
3249 		count += bp->b_wptr - bp->b_rptr;
3250 	}
3251 	return (count);
3252 }
3253 
3254 /*
3255  * Allocate a stream head.
3256  */
3257 struct stdata *
3258 shalloc(queue_t *qp)
3259 {
3260 	stdata_t *stp;
3261 
3262 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3263 
3264 	stp->sd_wrq = _WR(qp);
3265 	stp->sd_strtab = NULL;
3266 	stp->sd_iocid = 0;
3267 	stp->sd_mate = NULL;
3268 	stp->sd_freezer = NULL;
3269 	stp->sd_refcnt = 0;
3270 	stp->sd_wakeq = 0;
3271 	stp->sd_anchor = 0;
3272 	stp->sd_struiowrq = NULL;
3273 	stp->sd_struiordq = NULL;
3274 	stp->sd_struiodnak = 0;
3275 	stp->sd_struionak = NULL;
3276 #ifdef C2_AUDIT
3277 	stp->sd_t_audit_data = NULL;
3278 #endif
3279 	stp->sd_rput_opt = 0;
3280 	stp->sd_wput_opt = 0;
3281 	stp->sd_read_opt = 0;
3282 	stp->sd_rprotofunc = strrput_proto;
3283 	stp->sd_rmiscfunc = strrput_misc;
3284 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3285 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3286 	stp->sd_ciputctrl = NULL;
3287 	stp->sd_nciputctrl = 0;
3288 	stp->sd_qhead = NULL;
3289 	stp->sd_qtail = NULL;
3290 	stp->sd_servid = NULL;
3291 	stp->sd_nqueues = 0;
3292 	stp->sd_svcflags = 0;
3293 	stp->sd_copyflag = 0;
3294 
3295 	return (stp);
3296 }
3297 
3298 /*
3299  * Free a stream head.
3300  */
3301 void
3302 shfree(stdata_t *stp)
3303 {
3304 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3305 
3306 	stp->sd_wrq = NULL;
3307 
3308 	mutex_enter(&stp->sd_qlock);
3309 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3310 		STRSTAT(strwaits);
3311 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3312 	}
3313 	mutex_exit(&stp->sd_qlock);
3314 
3315 	if (stp->sd_ciputctrl != NULL) {
3316 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3317 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3318 		    stp->sd_nciputctrl, 0);
3319 		ASSERT(ciputctrl_cache != NULL);
3320 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3321 		stp->sd_ciputctrl = NULL;
3322 		stp->sd_nciputctrl = 0;
3323 	}
3324 	ASSERT(stp->sd_qhead == NULL);
3325 	ASSERT(stp->sd_qtail == NULL);
3326 	ASSERT(stp->sd_nqueues == 0);
3327 	kmem_cache_free(stream_head_cache, stp);
3328 }
3329 
3330 /*
3331  * Allocate a pair of queues and a syncq for the pair
3332  */
3333 queue_t *
3334 allocq(void)
3335 {
3336 	queinfo_t *qip;
3337 	queue_t *qp, *wqp;
3338 	syncq_t	*sq;
3339 
3340 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3341 
3342 	qp = &qip->qu_rqueue;
3343 	wqp = &qip->qu_wqueue;
3344 	sq = &qip->qu_syncq;
3345 
3346 	qp->q_last	= NULL;
3347 	qp->q_next	= NULL;
3348 	qp->q_ptr	= NULL;
3349 	qp->q_flag	= QUSE | QREADR;
3350 	qp->q_bandp	= NULL;
3351 	qp->q_stream	= NULL;
3352 	qp->q_syncq	= sq;
3353 	qp->q_nband	= 0;
3354 	qp->q_nfsrv	= NULL;
3355 	qp->q_draining	= 0;
3356 	qp->q_syncqmsgs	= 0;
3357 	qp->q_spri	= 0;
3358 	qp->q_qtstamp	= 0;
3359 	qp->q_sqtstamp	= 0;
3360 	qp->q_fp	= NULL;
3361 
3362 	wqp->q_last	= NULL;
3363 	wqp->q_next	= NULL;
3364 	wqp->q_ptr	= NULL;
3365 	wqp->q_flag	= QUSE;
3366 	wqp->q_bandp	= NULL;
3367 	wqp->q_stream	= NULL;
3368 	wqp->q_syncq	= sq;
3369 	wqp->q_nband	= 0;
3370 	wqp->q_nfsrv	= NULL;
3371 	wqp->q_draining	= 0;
3372 	wqp->q_syncqmsgs = 0;
3373 	wqp->q_qtstamp	= 0;
3374 	wqp->q_sqtstamp	= 0;
3375 	wqp->q_spri	= 0;
3376 
3377 	sq->sq_count	= 0;
3378 	sq->sq_rmqcount	= 0;
3379 	sq->sq_flags	= 0;
3380 	sq->sq_type	= 0;
3381 	sq->sq_callbflags = 0;
3382 	sq->sq_cancelid	= 0;
3383 	sq->sq_ciputctrl = NULL;
3384 	sq->sq_nciputctrl = 0;
3385 	sq->sq_needexcl = 0;
3386 	sq->sq_svcflags = 0;
3387 
3388 	return (qp);
3389 }
3390 
3391 /*
3392  * Free a pair of queues and the "attached" syncq.
3393  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3394  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3395  */
3396 void
3397 freeq(queue_t *qp)
3398 {
3399 	qband_t *qbp, *nqbp;
3400 	syncq_t *sq, *outer;
3401 	queue_t *wqp = _WR(qp);
3402 
3403 	ASSERT(qp->q_flag & QREADR);
3404 
3405 	/*
3406 	 * If a previously dispatched taskq job is scheduled to run
3407 	 * sync_service() or a service routine is scheduled for the
3408 	 * queues about to be freed, wait here until all service is
3409 	 * done on the queue and all associated queues and syncqs.
3410 	 */
3411 	wait_svc(qp);
3412 
3413 	(void) flush_syncq(qp->q_syncq, qp);
3414 	(void) flush_syncq(wqp->q_syncq, wqp);
3415 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3416 
3417 	/*
3418 	 * Flush the queues before q_next is set to NULL This is needed
3419 	 * in order to backenable any downstream queue before we go away.
3420 	 * Note: we are already removed from the stream so that the
3421 	 * backenabling will not cause any messages to be delivered to our
3422 	 * put procedures.
3423 	 */
3424 	flushq(qp, FLUSHALL);
3425 	flushq(wqp, FLUSHALL);
3426 
3427 	/* Tidy up - removeq only does a half-remove from stream */
3428 	qp->q_next = wqp->q_next = NULL;
3429 	ASSERT(!(qp->q_flag & QENAB));
3430 	ASSERT(!(wqp->q_flag & QENAB));
3431 
3432 	outer = qp->q_syncq->sq_outer;
3433 	if (outer != NULL) {
3434 		outer_remove(outer, qp->q_syncq);
3435 		if (wqp->q_syncq != qp->q_syncq)
3436 			outer_remove(outer, wqp->q_syncq);
3437 	}
3438 	/*
3439 	 * Free any syncqs that are outside what allocq returned.
3440 	 */
3441 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3442 		free_syncq(qp->q_syncq);
3443 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3444 		free_syncq(wqp->q_syncq);
3445 
3446 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3447 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3448 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3449 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3450 	sq = SQ(qp);
3451 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3452 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3453 	ASSERT(sq->sq_outer == NULL);
3454 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3455 	ASSERT(sq->sq_callbpend == NULL);
3456 	ASSERT(sq->sq_needexcl == 0);
3457 
3458 	if (sq->sq_ciputctrl != NULL) {
3459 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3460 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3461 		    sq->sq_nciputctrl, 0);
3462 		ASSERT(ciputctrl_cache != NULL);
3463 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3464 		sq->sq_ciputctrl = NULL;
3465 		sq->sq_nciputctrl = 0;
3466 	}
3467 
3468 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3469 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3470 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3471 
3472 	qp->q_flag &= ~QUSE;
3473 	wqp->q_flag &= ~QUSE;
3474 
3475 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3476 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3477 
3478 	qbp = qp->q_bandp;
3479 	while (qbp) {
3480 		nqbp = qbp->qb_next;
3481 		freeband(qbp);
3482 		qbp = nqbp;
3483 	}
3484 	qbp = wqp->q_bandp;
3485 	while (qbp) {
3486 		nqbp = qbp->qb_next;
3487 		freeband(qbp);
3488 		qbp = nqbp;
3489 	}
3490 	kmem_cache_free(queue_cache, qp);
3491 }
3492 
3493 /*
3494  * Allocate a qband structure.
3495  */
3496 qband_t *
3497 allocband(void)
3498 {
3499 	qband_t *qbp;
3500 
3501 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3502 	if (qbp == NULL)
3503 		return (NULL);
3504 
3505 	qbp->qb_next	= NULL;
3506 	qbp->qb_count	= 0;
3507 	qbp->qb_mblkcnt	= 0;
3508 	qbp->qb_first	= NULL;
3509 	qbp->qb_last	= NULL;
3510 	qbp->qb_flag	= 0;
3511 
3512 	return (qbp);
3513 }
3514 
3515 /*
3516  * Free a qband structure.
3517  */
3518 void
3519 freeband(qband_t *qbp)
3520 {
3521 	kmem_cache_free(qband_cache, qbp);
3522 }
3523 
3524 /*
3525  * Just like putnextctl(9F), except that allocb_wait() is used.
3526  *
3527  * Consolidation Private, and of course only callable from the stream head or
3528  * routines that may block.
3529  */
3530 int
3531 putnextctl_wait(queue_t *q, int type)
3532 {
3533 	mblk_t *bp;
3534 	int error;
3535 
3536 	if ((datamsg(type) && (type != M_DELAY)) ||
3537 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3538 		return (0);
3539 
3540 	bp->b_datap->db_type = (unsigned char)type;
3541 	putnext(q, bp);
3542 	return (1);
3543 }
3544 
3545 /*
3546  * run any possible bufcalls.
3547  */
3548 void
3549 runbufcalls(void)
3550 {
3551 	strbufcall_t *bcp;
3552 
3553 	mutex_enter(&bcall_monitor);
3554 	mutex_enter(&strbcall_lock);
3555 
3556 	if (strbcalls.bc_head) {
3557 		size_t count;
3558 		int nevent;
3559 
3560 		/*
3561 		 * count how many events are on the list
3562 		 * now so we can check to avoid looping
3563 		 * in low memory situations
3564 		 */
3565 		nevent = 0;
3566 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3567 			nevent++;
3568 
3569 		/*
3570 		 * get estimate of available memory from kmem_avail().
3571 		 * awake all bufcall functions waiting for
3572 		 * memory whose request could be satisfied
3573 		 * by 'count' memory and let 'em fight for it.
3574 		 */
3575 		count = kmem_avail();
3576 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3577 			STRSTAT(bufcalls);
3578 			--nevent;
3579 			if (bcp->bc_size <= count) {
3580 				bcp->bc_executor = curthread;
3581 				mutex_exit(&strbcall_lock);
3582 				(*bcp->bc_func)(bcp->bc_arg);
3583 				mutex_enter(&strbcall_lock);
3584 				bcp->bc_executor = NULL;
3585 				cv_broadcast(&bcall_cv);
3586 				strbcalls.bc_head = bcp->bc_next;
3587 				kmem_free(bcp, sizeof (strbufcall_t));
3588 			} else {
3589 				/*
3590 				 * too big, try again later - note
3591 				 * that nevent was decremented above
3592 				 * so we won't retry this one on this
3593 				 * iteration of the loop
3594 				 */
3595 				if (bcp->bc_next != NULL) {
3596 					strbcalls.bc_head = bcp->bc_next;
3597 					bcp->bc_next = NULL;
3598 					strbcalls.bc_tail->bc_next = bcp;
3599 					strbcalls.bc_tail = bcp;
3600 				}
3601 			}
3602 		}
3603 		if (strbcalls.bc_head == NULL)
3604 			strbcalls.bc_tail = NULL;
3605 	}
3606 
3607 	mutex_exit(&strbcall_lock);
3608 	mutex_exit(&bcall_monitor);
3609 }
3610 
3611 
3612 /*
3613  * actually run queue's service routine.
3614  */
3615 static void
3616 runservice(queue_t *q)
3617 {
3618 	qband_t *qbp;
3619 
3620 	ASSERT(q->q_qinfo->qi_srvp);
3621 again:
3622 	entersq(q->q_syncq, SQ_SVC);
3623 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3624 		"runservice starts:%p", q);
3625 
3626 	if (!(q->q_flag & QWCLOSE))
3627 		(*q->q_qinfo->qi_srvp)(q);
3628 
3629 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3630 		"runservice ends:(%p)", q);
3631 
3632 	leavesq(q->q_syncq, SQ_SVC);
3633 
3634 	mutex_enter(QLOCK(q));
3635 	if (q->q_flag & QENAB) {
3636 		q->q_flag &= ~QENAB;
3637 		mutex_exit(QLOCK(q));
3638 		goto again;
3639 	}
3640 	q->q_flag &= ~QINSERVICE;
3641 	q->q_flag &= ~QBACK;
3642 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3643 		qbp->qb_flag &= ~QB_BACK;
3644 	/*
3645 	 * Wakeup thread waiting for the service procedure
3646 	 * to be run (strclose and qdetach).
3647 	 */
3648 	cv_broadcast(&q->q_wait);
3649 
3650 	mutex_exit(QLOCK(q));
3651 }
3652 
3653 /*
3654  * Background processing of bufcalls.
3655  */
3656 void
3657 streams_bufcall_service(void)
3658 {
3659 	callb_cpr_t	cprinfo;
3660 
3661 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3662 	    "streams_bufcall_service");
3663 
3664 	mutex_enter(&strbcall_lock);
3665 
3666 	for (;;) {
3667 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3668 			mutex_exit(&strbcall_lock);
3669 			runbufcalls();
3670 			mutex_enter(&strbcall_lock);
3671 		}
3672 		if (strbcalls.bc_head != NULL) {
3673 			clock_t wt, tick;
3674 
3675 			STRSTAT(bcwaits);
3676 			/* Wait for memory to become available */
3677 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3678 			tick = SEC_TO_TICK(60);
3679 			time_to_wait(&wt, tick);
3680 			(void) cv_timedwait(&memavail_cv, &strbcall_lock, wt);
3681 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3682 		}
3683 
3684 		/* Wait for new work to arrive */
3685 		if (strbcalls.bc_head == NULL) {
3686 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3687 			cv_wait(&strbcall_cv, &strbcall_lock);
3688 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3689 		}
3690 	}
3691 }
3692 
3693 /*
3694  * Background processing of streams background tasks which failed
3695  * taskq_dispatch.
3696  */
3697 static void
3698 streams_qbkgrnd_service(void)
3699 {
3700 	callb_cpr_t cprinfo;
3701 	queue_t *q;
3702 
3703 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3704 	    "streams_bkgrnd_service");
3705 
3706 	mutex_enter(&service_queue);
3707 
3708 	for (;;) {
3709 		/*
3710 		 * Wait for work to arrive.
3711 		 */
3712 		while ((freebs_list == NULL) && (qhead == NULL)) {
3713 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3714 			cv_wait(&services_to_run, &service_queue);
3715 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3716 		}
3717 		/*
3718 		 * Handle all pending freebs requests to free memory.
3719 		 */
3720 		while (freebs_list != NULL) {
3721 			mblk_t *mp = freebs_list;
3722 			freebs_list = mp->b_next;
3723 			mutex_exit(&service_queue);
3724 			mblk_free(mp);
3725 			mutex_enter(&service_queue);
3726 		}
3727 		/*
3728 		 * Run pending queues.
3729 		 */
3730 		while (qhead != NULL) {
3731 			DQ(q, qhead, qtail, q_link);
3732 			ASSERT(q != NULL);
3733 			mutex_exit(&service_queue);
3734 			queue_service(q);
3735 			mutex_enter(&service_queue);
3736 		}
3737 		ASSERT(qhead == NULL && qtail == NULL);
3738 	}
3739 }
3740 
3741 /*
3742  * Background processing of streams background tasks which failed
3743  * taskq_dispatch.
3744  */
3745 static void
3746 streams_sqbkgrnd_service(void)
3747 {
3748 	callb_cpr_t cprinfo;
3749 	syncq_t *sq;
3750 
3751 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3752 	    "streams_sqbkgrnd_service");
3753 
3754 	mutex_enter(&service_queue);
3755 
3756 	for (;;) {
3757 		/*
3758 		 * Wait for work to arrive.
3759 		 */
3760 		while (sqhead == NULL) {
3761 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3762 			cv_wait(&syncqs_to_run, &service_queue);
3763 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3764 		}
3765 
3766 		/*
3767 		 * Run pending syncqs.
3768 		 */
3769 		while (sqhead != NULL) {
3770 			DQ(sq, sqhead, sqtail, sq_next);
3771 			ASSERT(sq != NULL);
3772 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3773 			mutex_exit(&service_queue);
3774 			syncq_service(sq);
3775 			mutex_enter(&service_queue);
3776 		}
3777 	}
3778 }
3779 
3780 /*
3781  * Disable the syncq and wait for background syncq processing to complete.
3782  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3783  * list.
3784  */
3785 void
3786 wait_sq_svc(syncq_t *sq)
3787 {
3788 	mutex_enter(SQLOCK(sq));
3789 	sq->sq_svcflags |= SQ_DISABLED;
3790 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3791 		syncq_t *sq_chase;
3792 		syncq_t *sq_curr;
3793 		int removed;
3794 
3795 		ASSERT(sq->sq_servcount == 1);
3796 		mutex_enter(&service_queue);
3797 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3798 		mutex_exit(&service_queue);
3799 		if (removed) {
3800 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3801 			sq->sq_servcount = 0;
3802 			STRSTAT(sqremoved);
3803 			goto done;
3804 		}
3805 	}
3806 	while (sq->sq_servcount != 0) {
3807 		sq->sq_flags |= SQ_WANTWAKEUP;
3808 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3809 	}
3810 done:
3811 	mutex_exit(SQLOCK(sq));
3812 }
3813 
3814 /*
3815  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3816  * Add the argument to the end of the sqhead list and set the flag
3817  * indicating this syncq has been enabled.  If it has already been
3818  * enabled, don't do anything.
3819  * This routine assumes that SQLOCK is held.
3820  * NOTE that the lock order is to have the SQLOCK first,
3821  * so if the service_syncq lock is held, we need to release it
3822  * before aquiring the SQLOCK (mostly relevant for the background
3823  * thread, and this seems to be common among the STREAMS global locks).
3824  * Note the the sq_svcflags are protected by the SQLOCK.
3825  */
3826 void
3827 sqenable(syncq_t *sq)
3828 {
3829 	/*
3830 	 * This is probably not important except for where I believe it
3831 	 * is being called.  At that point, it should be held (and it
3832 	 * is a pain to release it just for this routine, so don't do
3833 	 * it).
3834 	 */
3835 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3836 
3837 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3838 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3839 
3840 	/*
3841 	 * Do not put on list if background thread is scheduled or
3842 	 * syncq is disabled.
3843 	 */
3844 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3845 		return;
3846 
3847 	/*
3848 	 * Check whether we should enable sq at all.
3849 	 * Non PERMOD syncqs may be drained by at most one thread.
3850 	 * PERMOD syncqs may be drained by several threads but we limit the
3851 	 * total amount to the lesser of
3852 	 *	Number of queues on the squeue and
3853 	 *	Number of CPUs.
3854 	 */
3855 	if (sq->sq_servcount != 0) {
3856 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3857 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3858 			STRSTAT(sqtoomany);
3859 			return;
3860 		}
3861 	}
3862 
3863 	sq->sq_tstamp = lbolt;
3864 	STRSTAT(sqenables);
3865 
3866 	/* Attempt a taskq dispatch */
3867 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3868 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3869 	if (sq->sq_servid != NULL) {
3870 		sq->sq_servcount++;
3871 		return;
3872 	}
3873 
3874 	/*
3875 	 * This taskq dispatch failed, but a previous one may have succeeded.
3876 	 * Don't try to schedule on the background thread whilst there is
3877 	 * outstanding taskq processing.
3878 	 */
3879 	if (sq->sq_servcount != 0)
3880 		return;
3881 
3882 	/*
3883 	 * System is low on resources and can't perform a non-sleeping
3884 	 * dispatch. Schedule the syncq for a background thread and mark the
3885 	 * syncq to avoid any further taskq dispatch attempts.
3886 	 */
3887 	mutex_enter(&service_queue);
3888 	STRSTAT(taskqfails);
3889 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3890 	sq->sq_svcflags |= SQ_BGTHREAD;
3891 	sq->sq_servcount = 1;
3892 	cv_signal(&syncqs_to_run);
3893 	mutex_exit(&service_queue);
3894 }
3895 
3896 /*
3897  * Note: fifo_close() depends on the mblk_t on the queue being freed
3898  * asynchronously. The asynchronous freeing of messages breaks the
3899  * recursive call chain of fifo_close() while there are I_SENDFD type of
3900  * messages refering other file pointers on the queue. Then when
3901  * closing pipes it can avoid stack overflow in case of daisy-chained
3902  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3903  * share the same fifolock_t).
3904  */
3905 
3906 /* ARGSUSED */
3907 void
3908 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3909 {
3910 	ASSERT(dbp->db_mblk == mp);
3911 
3912 	/*
3913 	 * Check data sanity. The dblock should have non-empty free function.
3914 	 * It is better to panic here then later when the dblock is freed
3915 	 * asynchronously when the context is lost.
3916 	 */
3917 	if (dbp->db_frtnp->free_func == NULL) {
3918 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3919 		    (void *) dbp);
3920 	}
3921 
3922 	STRSTAT(freebs);
3923 	if (taskq_dispatch(streams_taskq, (task_func_t *)mblk_free, mp,
3924 	    TQ_NOSLEEP) == NULL) {
3925 		/*
3926 		 * System is low on resources and can't perform a non-sleeping
3927 		 * dispatch. Schedule for a background thread.
3928 		 */
3929 		mutex_enter(&service_queue);
3930 		STRSTAT(taskqfails);
3931 		mp->b_next = freebs_list;
3932 		freebs_list = mp;
3933 		cv_signal(&services_to_run);
3934 		mutex_exit(&service_queue);
3935 	}
3936 }
3937 
3938 /*
3939  * Set the QBACK or QB_BACK flag in the given queue for
3940  * the given priority band.
3941  */
3942 void
3943 setqback(queue_t *q, unsigned char pri)
3944 {
3945 	int i;
3946 	qband_t *qbp;
3947 	qband_t **qbpp;
3948 
3949 	ASSERT(MUTEX_HELD(QLOCK(q)));
3950 	if (pri != 0) {
3951 		if (pri > q->q_nband) {
3952 			qbpp = &q->q_bandp;
3953 			while (*qbpp)
3954 				qbpp = &(*qbpp)->qb_next;
3955 			while (pri > q->q_nband) {
3956 				if ((*qbpp = allocband()) == NULL) {
3957 					cmn_err(CE_WARN,
3958 					    "setqback: can't allocate qband\n");
3959 					return;
3960 				}
3961 				(*qbpp)->qb_hiwat = q->q_hiwat;
3962 				(*qbpp)->qb_lowat = q->q_lowat;
3963 				q->q_nband++;
3964 				qbpp = &(*qbpp)->qb_next;
3965 			}
3966 		}
3967 		qbp = q->q_bandp;
3968 		i = pri;
3969 		while (--i)
3970 			qbp = qbp->qb_next;
3971 		qbp->qb_flag |= QB_BACK;
3972 	} else {
3973 		q->q_flag |= QBACK;
3974 	}
3975 }
3976 
3977 int
3978 strcopyin(void *from, void *to, size_t len, int copyflag)
3979 {
3980 	if (copyflag & U_TO_K) {
3981 		ASSERT((copyflag & K_TO_K) == 0);
3982 		if (copyin(from, to, len))
3983 			return (EFAULT);
3984 	} else {
3985 		ASSERT(copyflag & K_TO_K);
3986 		bcopy(from, to, len);
3987 	}
3988 	return (0);
3989 }
3990 
3991 int
3992 strcopyout(void *from, void *to, size_t len, int copyflag)
3993 {
3994 	if (copyflag & U_TO_K) {
3995 		if (copyout(from, to, len))
3996 			return (EFAULT);
3997 	} else {
3998 		ASSERT(copyflag & K_TO_K);
3999 		bcopy(from, to, len);
4000 	}
4001 	return (0);
4002 }
4003 
4004 /*
4005  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4006  * It assumes that the stream head lock is already held, whereas strsignal()
4007  * acquires the lock first.  This routine was created because a few callers
4008  * release the stream head lock before calling only to re-acquire it after
4009  * it returns.
4010  */
4011 void
4012 strsignal_nolock(stdata_t *stp, int sig, int32_t band)
4013 {
4014 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4015 	switch (sig) {
4016 	case SIGPOLL:
4017 		if (stp->sd_sigflags & S_MSG)
4018 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4019 		break;
4020 
4021 	default:
4022 		if (stp->sd_pgidp) {
4023 			pgsignal(stp->sd_pgidp, sig);
4024 		}
4025 		break;
4026 	}
4027 }
4028 
4029 void
4030 strsignal(stdata_t *stp, int sig, int32_t band)
4031 {
4032 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4033 		"strsignal:%p, %X, %X", stp, sig, band);
4034 
4035 	mutex_enter(&stp->sd_lock);
4036 	switch (sig) {
4037 	case SIGPOLL:
4038 		if (stp->sd_sigflags & S_MSG)
4039 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4040 		break;
4041 
4042 	default:
4043 		if (stp->sd_pgidp) {
4044 			pgsignal(stp->sd_pgidp, sig);
4045 		}
4046 		break;
4047 	}
4048 	mutex_exit(&stp->sd_lock);
4049 }
4050 
4051 void
4052 strhup(stdata_t *stp)
4053 {
4054 	ASSERT(mutex_owned(&stp->sd_lock));
4055 	pollwakeup(&stp->sd_pollist, POLLHUP);
4056 	if (stp->sd_sigflags & S_HANGUP)
4057 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4058 }
4059 
4060 /*
4061  * Backenable the first queue upstream from `q' with a service procedure.
4062  */
4063 void
4064 backenable(queue_t *q, uchar_t pri)
4065 {
4066 	queue_t	*nq;
4067 
4068 	/*
4069 	 * our presence might not prevent other modules in our own
4070 	 * stream from popping/pushing since the caller of getq might not
4071 	 * have a claim on the queue (some drivers do a getq on somebody
4072 	 * else's queue - they know that the queue itself is not going away
4073 	 * but the framework has to guarantee q_next in that stream.)
4074 	 */
4075 	claimstr(q);
4076 
4077 	/* find nearest back queue with service proc */
4078 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4079 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4080 	}
4081 
4082 	if (nq) {
4083 		kthread_t *freezer;
4084 		/*
4085 		 * backenable can be called either with no locks held
4086 		 * or with the stream frozen (the latter occurs when a module
4087 		 * calls rmvq with the stream frozen.) If the stream is frozen
4088 		 * by the caller the caller will hold all qlocks in the stream.
4089 		 * Note that a frozen stream doesn't freeze a mated stream,
4090 		 * so we explicitly check for that.
4091 		 */
4092 		freezer = STREAM(q)->sd_freezer;
4093 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4094 			mutex_enter(QLOCK(nq));
4095 		}
4096 #ifdef DEBUG
4097 		else {
4098 			ASSERT(frozenstr(q));
4099 			ASSERT(MUTEX_HELD(QLOCK(q)));
4100 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4101 		}
4102 #endif
4103 		setqback(nq, pri);
4104 		qenable_locked(nq);
4105 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4106 			mutex_exit(QLOCK(nq));
4107 	}
4108 	releasestr(q);
4109 }
4110 
4111 /*
4112  * Return the appropriate errno when one of flags_to_check is set
4113  * in sd_flags. Uses the exported error routines if they are set.
4114  * Will return 0 if non error is set (or if the exported error routines
4115  * do not return an error).
4116  *
4117  * If there is both a read and write error to check we prefer the read error.
4118  * Also, give preference to recorded errno's over the error functions.
4119  * The flags that are handled are:
4120  *	STPLEX		return EINVAL
4121  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4122  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4123  *	STRHUP		return sd_werror
4124  *
4125  * If the caller indicates that the operation is a peek a nonpersistent error
4126  * is not cleared.
4127  */
4128 int
4129 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4130 {
4131 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4132 	int error = 0;
4133 
4134 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4135 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4136 	if (sd_flag & STPLEX)
4137 		error = EINVAL;
4138 	else if (sd_flag & STRDERR) {
4139 		error = stp->sd_rerror;
4140 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4141 			/*
4142 			 * Read errors are non-persistent i.e. discarded once
4143 			 * returned to a non-peeking caller,
4144 			 */
4145 			stp->sd_rerror = 0;
4146 			stp->sd_flag &= ~STRDERR;
4147 		}
4148 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4149 			int clearerr = 0;
4150 
4151 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4152 						&clearerr);
4153 			if (clearerr) {
4154 				stp->sd_flag &= ~STRDERR;
4155 				stp->sd_rderrfunc = NULL;
4156 			}
4157 		}
4158 	} else if (sd_flag & STWRERR) {
4159 		error = stp->sd_werror;
4160 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4161 			/*
4162 			 * Write errors are non-persistent i.e. discarded once
4163 			 * returned to a non-peeking caller,
4164 			 */
4165 			stp->sd_werror = 0;
4166 			stp->sd_flag &= ~STWRERR;
4167 		}
4168 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4169 			int clearerr = 0;
4170 
4171 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4172 						&clearerr);
4173 			if (clearerr) {
4174 				stp->sd_flag &= ~STWRERR;
4175 				stp->sd_wrerrfunc = NULL;
4176 			}
4177 		}
4178 	} else if (sd_flag & STRHUP) {
4179 		/* sd_werror set when STRHUP */
4180 		error = stp->sd_werror;
4181 	}
4182 	return (error);
4183 }
4184 
4185 
4186 /*
4187  * single-thread open/close/push/pop
4188  * for twisted streams also
4189  */
4190 int
4191 strstartplumb(stdata_t *stp, int flag, int cmd)
4192 {
4193 	int waited = 1;
4194 	int error = 0;
4195 
4196 	if (STRMATED(stp)) {
4197 		struct stdata *stmatep = stp->sd_mate;
4198 
4199 		STRLOCKMATES(stp);
4200 		while (waited) {
4201 			waited = 0;
4202 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4203 				if ((cmd == I_POP) &&
4204 				    (flag & (FNDELAY|FNONBLOCK))) {
4205 					STRUNLOCKMATES(stp);
4206 					return (EAGAIN);
4207 				}
4208 				waited = 1;
4209 				mutex_exit(&stp->sd_lock);
4210 				if (!cv_wait_sig(&stmatep->sd_monitor,
4211 				    &stmatep->sd_lock)) {
4212 					mutex_exit(&stmatep->sd_lock);
4213 					return (EINTR);
4214 				}
4215 				mutex_exit(&stmatep->sd_lock);
4216 				STRLOCKMATES(stp);
4217 			}
4218 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4219 				if ((cmd == I_POP) &&
4220 					(flag & (FNDELAY|FNONBLOCK))) {
4221 					STRUNLOCKMATES(stp);
4222 					return (EAGAIN);
4223 				}
4224 				waited = 1;
4225 				mutex_exit(&stmatep->sd_lock);
4226 				if (!cv_wait_sig(&stp->sd_monitor,
4227 				    &stp->sd_lock)) {
4228 					mutex_exit(&stp->sd_lock);
4229 					return (EINTR);
4230 				}
4231 				mutex_exit(&stp->sd_lock);
4232 				STRLOCKMATES(stp);
4233 			}
4234 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4235 				error = strgeterr(stp,
4236 					STRDERR|STWRERR|STRHUP|STPLEX, 0);
4237 				if (error != 0) {
4238 					STRUNLOCKMATES(stp);
4239 					return (error);
4240 				}
4241 			}
4242 		}
4243 		stp->sd_flag |= STRPLUMB;
4244 		STRUNLOCKMATES(stp);
4245 	} else {
4246 		mutex_enter(&stp->sd_lock);
4247 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4248 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4249 			    (flag & (FNDELAY|FNONBLOCK))) {
4250 				mutex_exit(&stp->sd_lock);
4251 				return (EAGAIN);
4252 			}
4253 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4254 				mutex_exit(&stp->sd_lock);
4255 				return (EINTR);
4256 			}
4257 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4258 				error = strgeterr(stp,
4259 					STRDERR|STWRERR|STRHUP|STPLEX, 0);
4260 				if (error != 0) {
4261 					mutex_exit(&stp->sd_lock);
4262 					return (error);
4263 				}
4264 			}
4265 		}
4266 		stp->sd_flag |= STRPLUMB;
4267 		mutex_exit(&stp->sd_lock);
4268 	}
4269 	return (0);
4270 }
4271 
4272 /*
4273  * Complete the plumbing operation associated with stream `stp'.
4274  */
4275 void
4276 strendplumb(stdata_t *stp)
4277 {
4278 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4279 	ASSERT(stp->sd_flag & STRPLUMB);
4280 	stp->sd_flag &= ~STRPLUMB;
4281 	cv_broadcast(&stp->sd_monitor);
4282 }
4283 
4284 /*
4285  * This describes how the STREAMS framework handles synchronization
4286  * during open/push and close/pop.
4287  * The key interfaces for open and close are qprocson and qprocsoff,
4288  * respectively. While the close case in general is harder both open
4289  * have close have significant similarities.
4290  *
4291  * During close the STREAMS framework has to both ensure that there
4292  * are no stale references to the queue pair (and syncq) that
4293  * are being closed and also provide the guarantees that are documented
4294  * in qprocsoff(9F).
4295  * If there are stale references to the queue that is closing it can
4296  * result in kernel memory corruption or kernel panics.
4297  *
4298  * Note that is it up to the module/driver to ensure that it itself
4299  * does not have any stale references to the closing queues once its close
4300  * routine returns. This includes:
4301  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4302  *    associated with the queues. For timeout and bufcall callbacks the
4303  *    module/driver also has to ensure (or wait for) any callbacks that
4304  *    are in progress.
4305  *  - If the module/driver is using esballoc it has to ensure that any
4306  *    esballoc free functions do not refer to a queue that has closed.
4307  *    (Note that in general the close routine can not wait for the esballoc'ed
4308  *    messages to be freed since that can cause a deadlock.)
4309  *  - Cancelling any interrupts that refer to the closing queues and
4310  *    also ensuring that there are no interrupts in progress that will
4311  *    refer to the closing queues once the close routine returns.
4312  *  - For multiplexors removing any driver global state that refers to
4313  *    the closing queue and also ensuring that there are no threads in
4314  *    the multiplexor that has picked up a queue pointer but not yet
4315  *    finished using it.
4316  *
4317  * In addition, a driver/module can only reference the q_next pointer
4318  * in its open, close, put, or service procedures or in a
4319  * qtimeout/qbufcall callback procedure executing "on" the correct
4320  * stream. Thus it can not reference the q_next pointer in an interrupt
4321  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4322  * it can not reference q_next of a different queue e.g. in a mux that
4323  * passes messages from one queues put/service procedure to another queue.
4324  * In all the cases when the driver/module can not access the q_next
4325  * field it must use the *next* versions e.g. canputnext instead of
4326  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4327  *
4328  *
4329  * Assuming that the driver/module conforms to the above constraints
4330  * the STREAMS framework has to avoid stale references to q_next for all
4331  * the framework internal cases which include (but are not limited to):
4332  *  - Threads in canput/canputnext/backenable and elsewhere that are
4333  *    walking q_next.
4334  *  - Messages on a syncq that have a reference to the queue through b_queue.
4335  *  - Messages on an outer perimeter (syncq) that have a reference to the
4336  *    queue through b_queue.
4337  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4338  *    Note that only canput and bcanput use q_nfsrv without any locking.
4339  *
4340  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4341  * after qprocsoff returns, the framework has to ensure that no threads can
4342  * enter the put or service routines for the closing read or write-side queue.
4343  * In addition to preventing "direct" entry into the put procedures
4344  * the framework also has to prevent messages being drained from
4345  * the syncq or the outer perimeter.
4346  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4347  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4348  * qprocsoff has returned.
4349  * Note that if a module/driver uses put(9F) on one of its own queues
4350  * it is up to the module/driver to ensure that the put() doesn't
4351  * get called when the queue is closing.
4352  *
4353  *
4354  * The framework aspects of the above "contract" is implemented by
4355  * qprocsoff, removeq, and strlock:
4356  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4357  *    entering the service procedures.
4358  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4359  *    canputnext, backenable etc from dereferencing the q_next that will
4360  *    soon change.
4361  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4362  *    or other q_next walker that uses claimstr/releasestr to finish.
4363  *  - optionally for every syncq in the stream strlock acquires all the
4364  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4365  *    that no thread executes in the put or service procedures and that no
4366  *    thread is draining into the module/driver. This ensures that no
4367  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4368  *    currently executing hence no such thread can end up with the old stale
4369  *    q_next value and no canput/backenable can have the old stale
4370  *    q_nfsrv/q_next.
4371  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4372  *    have either finished or observed the QWCLOSE flag and gone away.
4373  */
4374 
4375 
4376 /*
4377  * Get all the locks necessary to change q_next.
4378  *
4379  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for  the
4380  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4381  * the only threads inside the sqncq are threads currently calling removeq().
4382  * Since threads calling removeq() are in the process of removing their queues
4383  * from the stream, we do not need to worry about them accessing a stale q_next
4384  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4385  * for them can cause deadlock).
4386  *
4387  * This routine is subject to starvation since it does not set any flag to
4388  * prevent threads from entering a module in the stream(i.e. sq_count can
4389  * increase on some syncq while it is waiting on some other syncq.)
4390  *
4391  * Assumes that only one thread attempts to call strlock for a given
4392  * stream. If this is not the case the two threads would deadlock.
4393  * This assumption is guaranteed since strlock is only called by insertq
4394  * and removeq and streams plumbing changes are single-threaded for
4395  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4396  *
4397  * For pipes, it is not difficult to atomically designate a pair of streams
4398  * to be mated. Once mated atomically by the framework the twisted pair remain
4399  * configured that way until dismantled atomically by the framework.
4400  * When plumbing takes place on a twisted stream it is necessary to ensure that
4401  * this operation is done exclusively on the twisted stream since two such
4402  * operations, each initiated on different ends of the pipe will deadlock
4403  * waiting for each other to complete.
4404  *
4405  * On entry, no locks should be held.
4406  * The locks acquired and held by strlock depends on a few factors.
4407  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4408  *   and held on exit and all sq_count are at an acceptable level.
4409  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4410  *   sd_refcnt being zero.
4411  */
4412 
4413 static void
4414 strlock(struct stdata *stp, sqlist_t *sqlist)
4415 {
4416 	syncql_t *sql, *sql2;
4417 retry:
4418 	/*
4419 	 * Wait for any claimstr to go away.
4420 	 */
4421 	if (STRMATED(stp)) {
4422 		struct stdata *stp1, *stp2;
4423 
4424 		STRLOCKMATES(stp);
4425 		/*
4426 		 * Note that the selection of locking order is not
4427 		 * important, just that they are always aquired in
4428 		 * the same order.  To assure this, we choose this
4429 		 * order based on the value of the pointer, and since
4430 		 * the pointer will not change for the life of this
4431 		 * pair, we will always grab the locks in the same
4432 		 * order (and hence, prevent deadlocks).
4433 		 */
4434 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4435 			stp1 = stp;
4436 			stp2 = stp->sd_mate;
4437 		} else {
4438 			stp2 = stp;
4439 			stp1 = stp->sd_mate;
4440 		}
4441 		mutex_enter(&stp1->sd_reflock);
4442 		if (stp1->sd_refcnt > 0) {
4443 			STRUNLOCKMATES(stp);
4444 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4445 			mutex_exit(&stp1->sd_reflock);
4446 			goto retry;
4447 		}
4448 		mutex_enter(&stp2->sd_reflock);
4449 		if (stp2->sd_refcnt > 0) {
4450 			STRUNLOCKMATES(stp);
4451 			mutex_exit(&stp1->sd_reflock);
4452 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4453 			mutex_exit(&stp2->sd_reflock);
4454 			goto retry;
4455 		}
4456 		STREAM_PUTLOCKS_ENTER(stp1);
4457 		STREAM_PUTLOCKS_ENTER(stp2);
4458 	} else {
4459 		mutex_enter(&stp->sd_lock);
4460 		mutex_enter(&stp->sd_reflock);
4461 		while (stp->sd_refcnt > 0) {
4462 			mutex_exit(&stp->sd_lock);
4463 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4464 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4465 				mutex_exit(&stp->sd_reflock);
4466 				mutex_enter(&stp->sd_lock);
4467 				mutex_enter(&stp->sd_reflock);
4468 			}
4469 		}
4470 		STREAM_PUTLOCKS_ENTER(stp);
4471 	}
4472 
4473 	if (sqlist == NULL)
4474 		return;
4475 
4476 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4477 		syncq_t *sq = sql->sql_sq;
4478 		uint16_t count;
4479 
4480 		mutex_enter(SQLOCK(sq));
4481 		count = sq->sq_count;
4482 		ASSERT(sq->sq_rmqcount <= count);
4483 		SQ_PUTLOCKS_ENTER(sq);
4484 		SUM_SQ_PUTCOUNTS(sq, count);
4485 		if (count == sq->sq_rmqcount)
4486 			continue;
4487 
4488 		/* Failed - drop all locks that we have acquired so far */
4489 		if (STRMATED(stp)) {
4490 			STREAM_PUTLOCKS_EXIT(stp);
4491 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4492 			STRUNLOCKMATES(stp);
4493 			mutex_exit(&stp->sd_reflock);
4494 			mutex_exit(&stp->sd_mate->sd_reflock);
4495 		} else {
4496 			STREAM_PUTLOCKS_EXIT(stp);
4497 			mutex_exit(&stp->sd_lock);
4498 			mutex_exit(&stp->sd_reflock);
4499 		}
4500 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4501 		    sql2 = sql2->sql_next) {
4502 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4503 			mutex_exit(SQLOCK(sql2->sql_sq));
4504 		}
4505 
4506 		/*
4507 		 * The wait loop below may starve when there are many threads
4508 		 * claiming the syncq. This is especially a problem with permod
4509 		 * syncqs (IP). To lessen the impact of the problem we increment
4510 		 * sq_needexcl and clear fastbits so that putnexts will slow
4511 		 * down and call sqenable instead of draining right away.
4512 		 */
4513 		sq->sq_needexcl++;
4514 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4515 		while (count > sq->sq_rmqcount) {
4516 			sq->sq_flags |= SQ_WANTWAKEUP;
4517 			SQ_PUTLOCKS_EXIT(sq);
4518 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4519 			count = sq->sq_count;
4520 			SQ_PUTLOCKS_ENTER(sq);
4521 			SUM_SQ_PUTCOUNTS(sq, count);
4522 		}
4523 		sq->sq_needexcl--;
4524 		if (sq->sq_needexcl == 0)
4525 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4526 		SQ_PUTLOCKS_EXIT(sq);
4527 		ASSERT(count == sq->sq_rmqcount);
4528 		mutex_exit(SQLOCK(sq));
4529 		goto retry;
4530 	}
4531 }
4532 
4533 /*
4534  * Drop all the locks that strlock acquired.
4535  */
4536 static void
4537 strunlock(struct stdata *stp, sqlist_t *sqlist)
4538 {
4539 	syncql_t *sql;
4540 
4541 	if (STRMATED(stp)) {
4542 		STREAM_PUTLOCKS_EXIT(stp);
4543 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4544 		STRUNLOCKMATES(stp);
4545 		mutex_exit(&stp->sd_reflock);
4546 		mutex_exit(&stp->sd_mate->sd_reflock);
4547 	} else {
4548 		STREAM_PUTLOCKS_EXIT(stp);
4549 		mutex_exit(&stp->sd_lock);
4550 		mutex_exit(&stp->sd_reflock);
4551 	}
4552 
4553 	if (sqlist == NULL)
4554 		return;
4555 
4556 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4557 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4558 		mutex_exit(SQLOCK(sql->sql_sq));
4559 	}
4560 }
4561 
4562 /*
4563  * When the module has service procedure, we need check if the next
4564  * module which has service procedure is in flow control to trigger
4565  * the backenable.
4566  */
4567 static void
4568 backenable_insertedq(queue_t *q)
4569 {
4570 	qband_t	*qbp;
4571 
4572 	claimstr(q);
4573 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4574 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4575 			backenable(q, 0);
4576 
4577 		qbp = q->q_next->q_nfsrv->q_bandp;
4578 		for (; qbp != NULL; qbp = qbp->qb_next)
4579 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4580 				backenable(q, qbp->qb_first->b_band);
4581 	}
4582 	releasestr(q);
4583 }
4584 
4585 /*
4586  * Given two read queues, insert a new single one after another.
4587  *
4588  * This routine acquires all the necessary locks in order to change
4589  * q_next and related pointer using strlock().
4590  * It depends on the stream head ensuring that there are no concurrent
4591  * insertq or removeq on the same stream. The stream head ensures this
4592  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4593  *
4594  * Note that no syncq locks are held during the q_next change. This is
4595  * applied to all streams since, unlike removeq, there is no problem of stale
4596  * pointers when adding a module to the stream. Thus drivers/modules that do a
4597  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4598  * applied this optimization to all streams.
4599  */
4600 void
4601 insertq(struct stdata *stp, queue_t *new)
4602 {
4603 	queue_t	*after;
4604 	queue_t *wafter;
4605 	queue_t *wnew = _WR(new);
4606 	boolean_t have_fifo = B_FALSE;
4607 
4608 	if (new->q_flag & _QINSERTING) {
4609 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4610 		after = new->q_next;
4611 		wafter = _WR(new->q_next);
4612 	} else {
4613 		after = _RD(stp->sd_wrq);
4614 		wafter = stp->sd_wrq;
4615 	}
4616 
4617 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4618 		"insertq:%p, %p", after, new);
4619 	ASSERT(after->q_flag & QREADR);
4620 	ASSERT(new->q_flag & QREADR);
4621 
4622 	strlock(stp, NULL);
4623 
4624 	/* Do we have a FIFO? */
4625 	if (wafter->q_next == after) {
4626 		have_fifo = B_TRUE;
4627 		wnew->q_next = new;
4628 	} else {
4629 		wnew->q_next = wafter->q_next;
4630 	}
4631 	new->q_next = after;
4632 
4633 	set_nfsrv_ptr(new, wnew, after, wafter);
4634 	/*
4635 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4636 	 * so only reset this flag after calling it.
4637 	 */
4638 	new->q_flag &= ~_QINSERTING;
4639 
4640 	if (have_fifo) {
4641 		wafter->q_next = wnew;
4642 	} else {
4643 		if (wafter->q_next)
4644 			_OTHERQ(wafter->q_next)->q_next = new;
4645 		wafter->q_next = wnew;
4646 	}
4647 
4648 	set_qend(new);
4649 	/* The QEND flag might have to be updated for the upstream guy */
4650 	set_qend(after);
4651 
4652 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4653 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4654 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4655 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4656 	strsetuio(stp);
4657 
4658 	/*
4659 	 * If this was a module insertion, bump the push count.
4660 	 */
4661 	if (!(new->q_flag & QISDRV))
4662 		stp->sd_pushcnt++;
4663 
4664 	strunlock(stp, NULL);
4665 
4666 	/* check if the write Q needs backenable */
4667 	backenable_insertedq(wnew);
4668 
4669 	/* check if the read Q needs backenable */
4670 	backenable_insertedq(new);
4671 }
4672 
4673 /*
4674  * Given a read queue, unlink it from any neighbors.
4675  *
4676  * This routine acquires all the necessary locks in order to
4677  * change q_next and related pointers and also guard against
4678  * stale references (e.g. through q_next) to the queue that
4679  * is being removed. It also plays part of the role in ensuring
4680  * that the module's/driver's put procedure doesn't get called
4681  * after qprocsoff returns.
4682  *
4683  * Removeq depends on the stream head ensuring that there are
4684  * no concurrent insertq or removeq on the same stream. The
4685  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4686  * STRPLUMB.
4687  *
4688  * The set of locks needed to remove the queue is different in
4689  * different cases:
4690  *
4691  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4692  * waiting for the syncq reference count to drop to 0 indicating that no
4693  * non-close threads are present anywhere in the stream. This ensures that any
4694  * module/driver can reference q_next in its open, close, put, or service
4695  * procedures.
4696  *
4697  * The sq_rmqcount counter tracks the number of threads inside removeq().
4698  * strlock() ensures that there is either no threads executing inside perimeter
4699  * or there is only a thread calling qprocsoff().
4700  *
4701  * strlock() compares the value of sq_count with the number of threads inside
4702  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4703  * any threads waiting in strlock() when the sq_rmqcount increases.
4704  */
4705 
4706 void
4707 removeq(queue_t *qp)
4708 {
4709 	queue_t *wqp = _WR(qp);
4710 	struct stdata *stp = STREAM(qp);
4711 	sqlist_t *sqlist = NULL;
4712 	boolean_t isdriver;
4713 	int moved;
4714 	syncq_t *sq = qp->q_syncq;
4715 	syncq_t *wsq = wqp->q_syncq;
4716 
4717 	ASSERT(stp);
4718 
4719 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4720 		"removeq:%p %p", qp, wqp);
4721 	ASSERT(qp->q_flag&QREADR);
4722 
4723 	/*
4724 	 * For queues using Synchronous streams, we must wait for all threads in
4725 	 * rwnext() to drain out before proceeding.
4726 	 */
4727 	if (qp->q_flag & QSYNCSTR) {
4728 		/* First, we need wakeup any threads blocked in rwnext() */
4729 		mutex_enter(SQLOCK(sq));
4730 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4731 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4732 			cv_broadcast(&sq->sq_wait);
4733 		}
4734 		mutex_exit(SQLOCK(sq));
4735 
4736 		if (wsq != sq) {
4737 			mutex_enter(SQLOCK(wsq));
4738 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4739 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4740 				cv_broadcast(&wsq->sq_wait);
4741 			}
4742 			mutex_exit(SQLOCK(wsq));
4743 		}
4744 
4745 		mutex_enter(QLOCK(qp));
4746 		while (qp->q_rwcnt > 0) {
4747 			qp->q_flag |= QWANTRMQSYNC;
4748 			cv_wait(&qp->q_wait, QLOCK(qp));
4749 		}
4750 		mutex_exit(QLOCK(qp));
4751 
4752 		mutex_enter(QLOCK(wqp));
4753 		while (wqp->q_rwcnt > 0) {
4754 			wqp->q_flag |= QWANTRMQSYNC;
4755 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4756 		}
4757 		mutex_exit(QLOCK(wqp));
4758 	}
4759 
4760 	mutex_enter(SQLOCK(sq));
4761 	sq->sq_rmqcount++;
4762 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4763 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4764 		cv_broadcast(&sq->sq_wait);
4765 	}
4766 	mutex_exit(SQLOCK(sq));
4767 
4768 	isdriver = (qp->q_flag & QISDRV);
4769 
4770 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4771 	strlock(stp, sqlist);
4772 
4773 	reset_nfsrv_ptr(qp, wqp);
4774 
4775 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4776 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4777 	/* Do we have a FIFO? */
4778 	if (wqp->q_next == qp) {
4779 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4780 	} else {
4781 		if (wqp->q_next)
4782 			backq(qp)->q_next = qp->q_next;
4783 		if (qp->q_next)
4784 			backq(wqp)->q_next = wqp->q_next;
4785 	}
4786 
4787 	/* The QEND flag might have to be updated for the upstream guy */
4788 	if (qp->q_next)
4789 		set_qend(qp->q_next);
4790 
4791 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4792 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4793 
4794 	/*
4795 	 * Move any messages destined for the put procedures to the next
4796 	 * syncq in line. Otherwise free them.
4797 	 */
4798 	moved = 0;
4799 	/*
4800 	 * Quick check to see whether there are any messages or events.
4801 	 */
4802 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4803 		moved += propagate_syncq(qp);
4804 	if (wqp->q_syncqmsgs != 0 ||
4805 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4806 		moved += propagate_syncq(wqp);
4807 
4808 	strsetuio(stp);
4809 
4810 	/*
4811 	 * If this was a module removal, decrement the push count.
4812 	 */
4813 	if (!isdriver)
4814 		stp->sd_pushcnt--;
4815 
4816 	strunlock(stp, sqlist);
4817 	sqlist_free(sqlist);
4818 
4819 	/*
4820 	 * Make sure any messages that were propagated are drained.
4821 	 * Also clear any QFULL bit caused by messages that were propagated.
4822 	 */
4823 
4824 	if (qp->q_next != NULL) {
4825 		clr_qfull(qp);
4826 		/*
4827 		 * For the driver calling qprocsoff, propagate_syncq
4828 		 * frees all the messages instead of putting it in
4829 		 * the stream head
4830 		 */
4831 		if (!isdriver && (moved > 0))
4832 			emptysq(qp->q_next->q_syncq);
4833 	}
4834 	if (wqp->q_next != NULL) {
4835 		clr_qfull(wqp);
4836 		/*
4837 		 * We come here for any pop of a module except for the
4838 		 * case of driver being removed. We don't call emptysq
4839 		 * if we did not move any messages. This will avoid holding
4840 		 * PERMOD syncq locks in emptysq
4841 		 */
4842 		if (moved > 0)
4843 			emptysq(wqp->q_next->q_syncq);
4844 	}
4845 
4846 	mutex_enter(SQLOCK(sq));
4847 	sq->sq_rmqcount--;
4848 	mutex_exit(SQLOCK(sq));
4849 }
4850 
4851 /*
4852  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
4853  * SQ_WRITER) on a syncq.
4854  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
4855  * sync queue and waits until sq_count reaches maxcnt.
4856  *
4857  * if maxcnt is -1 there's no need to grab sq_putlocks since the caller
4858  * does not care about putnext threads that are in the middle of calling put
4859  * entry points.
4860  *
4861  * This routine is used for both inner and outer syncqs.
4862  */
4863 static void
4864 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
4865 {
4866 	uint16_t count = 0;
4867 
4868 	mutex_enter(SQLOCK(sq));
4869 	/*
4870 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
4871 	 * SQ_FROZEN will be set if there is a frozen stream that has a
4872 	 * queue which also refers to this "shared" syncq.
4873 	 * SQ_BLOCKED will be set if there is "off" queue which also
4874 	 * refers to this "shared" syncq.
4875 	 */
4876 	if (maxcnt != -1) {
4877 		count = sq->sq_count;
4878 		SQ_PUTLOCKS_ENTER(sq);
4879 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4880 		SUM_SQ_PUTCOUNTS(sq, count);
4881 	}
4882 	sq->sq_needexcl++;
4883 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
4884 
4885 	while ((sq->sq_flags & flag) ||
4886 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
4887 		sq->sq_flags |= SQ_WANTWAKEUP;
4888 		if (maxcnt != -1) {
4889 			SQ_PUTLOCKS_EXIT(sq);
4890 		}
4891 		cv_wait(&sq->sq_wait, SQLOCK(sq));
4892 		if (maxcnt != -1) {
4893 			count = sq->sq_count;
4894 			SQ_PUTLOCKS_ENTER(sq);
4895 			SUM_SQ_PUTCOUNTS(sq, count);
4896 		}
4897 	}
4898 	sq->sq_needexcl--;
4899 	sq->sq_flags |= flag;
4900 	ASSERT(maxcnt == -1 || count == maxcnt);
4901 	if (maxcnt != -1) {
4902 		if (sq->sq_needexcl == 0) {
4903 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4904 		}
4905 		SQ_PUTLOCKS_EXIT(sq);
4906 	} else if (sq->sq_needexcl == 0) {
4907 		SQ_PUTCOUNT_SETFAST(sq);
4908 	}
4909 
4910 	mutex_exit(SQLOCK(sq));
4911 }
4912 
4913 /*
4914  * Reset a flag that was set with blocksq.
4915  *
4916  * Can not use this routine to reset SQ_WRITER.
4917  *
4918  * If "isouter" is set then the syncq is assumed to be an outer perimeter
4919  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
4920  * to handle the queued qwriter operations.
4921  *
4922  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4923  * sq_putlocks are used.
4924  */
4925 static void
4926 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
4927 {
4928 	uint16_t flags;
4929 
4930 	mutex_enter(SQLOCK(sq));
4931 	ASSERT(resetflag != SQ_WRITER);
4932 	ASSERT(sq->sq_flags & resetflag);
4933 	flags = sq->sq_flags & ~resetflag;
4934 	sq->sq_flags = flags;
4935 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
4936 		if (flags & SQ_WANTWAKEUP) {
4937 			flags &= ~SQ_WANTWAKEUP;
4938 			cv_broadcast(&sq->sq_wait);
4939 		}
4940 		sq->sq_flags = flags;
4941 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
4942 			if (!isouter) {
4943 				/* drain_syncq drops SQLOCK */
4944 				drain_syncq(sq);
4945 				return;
4946 			}
4947 		}
4948 	}
4949 	mutex_exit(SQLOCK(sq));
4950 }
4951 
4952 /*
4953  * Reset a flag that was set with blocksq.
4954  * Does not drain the syncq. Use emptysq() for that.
4955  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
4956  *
4957  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4958  * sq_putlocks are used.
4959  */
4960 static int
4961 dropsq(syncq_t *sq, uint16_t resetflag)
4962 {
4963 	uint16_t flags;
4964 
4965 	mutex_enter(SQLOCK(sq));
4966 	ASSERT(sq->sq_flags & resetflag);
4967 	flags = sq->sq_flags & ~resetflag;
4968 	if (flags & SQ_WANTWAKEUP) {
4969 		flags &= ~SQ_WANTWAKEUP;
4970 		cv_broadcast(&sq->sq_wait);
4971 	}
4972 	sq->sq_flags = flags;
4973 	mutex_exit(SQLOCK(sq));
4974 	if (flags & SQ_QUEUED)
4975 		return (1);
4976 	return (0);
4977 }
4978 
4979 /*
4980  * Empty all the messages on a syncq.
4981  *
4982  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
4983  * sq_putlocks are used.
4984  */
4985 static void
4986 emptysq(syncq_t *sq)
4987 {
4988 	uint16_t flags;
4989 
4990 	mutex_enter(SQLOCK(sq));
4991 	flags = sq->sq_flags;
4992 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
4993 		/*
4994 		 * To prevent potential recursive invocation of drain_syncq we
4995 		 * do not call drain_syncq if count is non-zero.
4996 		 */
4997 		if (sq->sq_count == 0) {
4998 			/* drain_syncq() drops SQLOCK */
4999 			drain_syncq(sq);
5000 			return;
5001 		} else
5002 			sqenable(sq);
5003 	}
5004 	mutex_exit(SQLOCK(sq));
5005 }
5006 
5007 /*
5008  * Ordered insert while removing duplicates.
5009  */
5010 static void
5011 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5012 {
5013 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5014 
5015 	prev_sqlpp = &sqlist->sqlist_head;
5016 	while ((sqlp = *prev_sqlpp) != NULL) {
5017 		if (sqlp->sql_sq >= sqp) {
5018 			if (sqlp->sql_sq == sqp)	/* duplicate */
5019 				return;
5020 			break;
5021 		}
5022 		prev_sqlpp = &sqlp->sql_next;
5023 	}
5024 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5025 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5026 	new_sqlp->sql_next = sqlp;
5027 	new_sqlp->sql_sq = sqp;
5028 	*prev_sqlpp = new_sqlp;
5029 }
5030 
5031 /*
5032  * Walk the write side queues until we hit either the driver
5033  * or a twist in the stream (_SAMESTR will return false in both
5034  * these cases) then turn around and walk the read side queues
5035  * back up to the stream head.
5036  */
5037 static void
5038 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5039 {
5040 	while (q != NULL) {
5041 		sqlist_insert(sqlist, q->q_syncq);
5042 
5043 		if (_SAMESTR(q))
5044 			q = q->q_next;
5045 		else if (!(q->q_flag & QREADR))
5046 			q = _RD(q);
5047 		else
5048 			q = NULL;
5049 	}
5050 }
5051 
5052 /*
5053  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5054  * associated with the "q" parameter. The resulting list is sorted in a
5055  * canonical order and is free of duplicates.
5056  * Assumes the passed queue is a _RD(q).
5057  */
5058 static sqlist_t *
5059 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5060 {
5061 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5062 
5063 	/*
5064 	 * start with the current queue/qpair
5065 	 */
5066 	ASSERT(q->q_flag & QREADR);
5067 
5068 	sqlist_insert(sqlist, q->q_syncq);
5069 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5070 
5071 	sqlist_insertall(sqlist, stp->sd_wrq);
5072 	if (do_twist)
5073 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5074 
5075 	return (sqlist);
5076 }
5077 
5078 static sqlist_t *
5079 sqlist_alloc(struct stdata *stp, int kmflag)
5080 {
5081 	size_t sqlist_size;
5082 	sqlist_t *sqlist;
5083 
5084 	/*
5085 	 * Allocate 2 syncql_t's for each pushed module. Note that
5086 	 * the sqlist_t structure already has 4 syncql_t's built in:
5087 	 * 2 for the stream head, and 2 for the driver/other stream head.
5088 	 */
5089 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5090 		sizeof (sqlist_t);
5091 	if (STRMATED(stp))
5092 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5093 	sqlist = kmem_alloc(sqlist_size, kmflag);
5094 
5095 	sqlist->sqlist_head = NULL;
5096 	sqlist->sqlist_size = sqlist_size;
5097 	sqlist->sqlist_index = 0;
5098 
5099 	return (sqlist);
5100 }
5101 
5102 /*
5103  * Free the list created by sqlist_alloc()
5104  */
5105 static void
5106 sqlist_free(sqlist_t *sqlist)
5107 {
5108 	kmem_free(sqlist, sqlist->sqlist_size);
5109 }
5110 
5111 /*
5112  * Prevent any new entries into any syncq in this stream.
5113  * Used by freezestr.
5114  */
5115 void
5116 strblock(queue_t *q)
5117 {
5118 	struct stdata	*stp;
5119 	syncql_t	*sql;
5120 	sqlist_t	*sqlist;
5121 
5122 	q = _RD(q);
5123 
5124 	stp = STREAM(q);
5125 	ASSERT(stp != NULL);
5126 
5127 	/*
5128 	 * Get a sorted list with all the duplicates removed containing
5129 	 * all the syncqs referenced by this stream.
5130 	 */
5131 	sqlist = sqlist_build(q, stp, B_FALSE);
5132 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5133 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5134 	sqlist_free(sqlist);
5135 }
5136 
5137 /*
5138  * Release the block on new entries into this stream
5139  */
5140 void
5141 strunblock(queue_t *q)
5142 {
5143 	struct stdata	*stp;
5144 	syncql_t	*sql;
5145 	sqlist_t	*sqlist;
5146 	int		drain_needed;
5147 
5148 	q = _RD(q);
5149 
5150 	/*
5151 	 * Get a sorted list with all the duplicates removed containing
5152 	 * all the syncqs referenced by this stream.
5153 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5154 	 * starting to drain them; otherwise the draining might
5155 	 * cause a freezestr in some module on the stream (which
5156 	 * would deadlock.)
5157 	 */
5158 	stp = STREAM(q);
5159 	ASSERT(stp != NULL);
5160 	sqlist = sqlist_build(q, stp, B_FALSE);
5161 	drain_needed = 0;
5162 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5163 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5164 	if (drain_needed) {
5165 		for (sql = sqlist->sqlist_head; sql != NULL;
5166 		    sql = sql->sql_next)
5167 			emptysq(sql->sql_sq);
5168 	}
5169 	sqlist_free(sqlist);
5170 }
5171 
5172 #ifdef DEBUG
5173 static int
5174 qprocsareon(queue_t *rq)
5175 {
5176 	if (rq->q_next == NULL)
5177 		return (0);
5178 	return (_WR(rq->q_next)->q_next == _WR(rq));
5179 }
5180 
5181 int
5182 qclaimed(queue_t *q)
5183 {
5184 	uint_t count;
5185 
5186 	count = q->q_syncq->sq_count;
5187 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5188 	return (count != 0);
5189 }
5190 
5191 /*
5192  * Check if anyone has frozen this stream with freezestr
5193  */
5194 int
5195 frozenstr(queue_t *q)
5196 {
5197 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5198 }
5199 #endif /* DEBUG */
5200 
5201 /*
5202  * Enter a queue.
5203  * Obsoleted interface. Should not be used.
5204  */
5205 void
5206 enterq(queue_t *q)
5207 {
5208 	entersq(q->q_syncq, SQ_CALLBACK);
5209 }
5210 
5211 void
5212 leaveq(queue_t *q)
5213 {
5214 	leavesq(q->q_syncq, SQ_CALLBACK);
5215 }
5216 
5217 /*
5218  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5219  * to check.
5220  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5221  * calls and the running of open, close and service procedures.
5222  *
5223  * if c_inner bit is set no need to grab sq_putlocks since we don't care
5224  * if other threads have entered or are entering put entry point.
5225  *
5226  * if c_inner bit is set it might have been posible to use
5227  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5228  * open/close path for IP) but since the count may need to be decremented in
5229  * qwait() we wouldn't know which counter to decrement. Currently counter is
5230  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5231  * in the future we might use curthread id bits to select the counter and this
5232  * would stay constant across routine calls.
5233  */
5234 void
5235 entersq(syncq_t *sq, int entrypoint)
5236 {
5237 	uint16_t	count = 0;
5238 	uint16_t	flags;
5239 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5240 	uint16_t	type;
5241 	uint_t		c_inner = entrypoint & SQ_CI;
5242 	uint_t		c_outer = entrypoint & SQ_CO;
5243 
5244 	/*
5245 	 * Increment ref count to keep closes out of this queue.
5246 	 */
5247 	ASSERT(sq);
5248 	ASSERT(c_inner && c_outer);
5249 	mutex_enter(SQLOCK(sq));
5250 	flags = sq->sq_flags;
5251 	type = sq->sq_type;
5252 	if (!(type & c_inner)) {
5253 		/* Make sure all putcounts now use slowlock. */
5254 		count = sq->sq_count;
5255 		SQ_PUTLOCKS_ENTER(sq);
5256 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5257 		SUM_SQ_PUTCOUNTS(sq, count);
5258 		sq->sq_needexcl++;
5259 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5260 		waitflags |= SQ_MESSAGES;
5261 	}
5262 	/*
5263 	 * Wait until we can enter the inner perimeter.
5264 	 * If we want exclusive access we wait until sq_count is 0.
5265 	 * We have to do this before entering the outer perimeter in order
5266 	 * to preserve put/close message ordering.
5267 	 */
5268 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5269 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5270 		if (!(type & c_inner)) {
5271 			SQ_PUTLOCKS_EXIT(sq);
5272 		}
5273 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5274 		if (!(type & c_inner)) {
5275 			count = sq->sq_count;
5276 			SQ_PUTLOCKS_ENTER(sq);
5277 			SUM_SQ_PUTCOUNTS(sq, count);
5278 		}
5279 		flags = sq->sq_flags;
5280 	}
5281 
5282 	if (!(type & c_inner)) {
5283 		ASSERT(sq->sq_needexcl > 0);
5284 		sq->sq_needexcl--;
5285 		if (sq->sq_needexcl == 0) {
5286 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5287 		}
5288 	}
5289 
5290 	/* Check if we need to enter the outer perimeter */
5291 	if (!(type & c_outer)) {
5292 		/*
5293 		 * We have to enter the outer perimeter exclusively before
5294 		 * we can increment sq_count to avoid deadlock. This implies
5295 		 * that we have to re-check sq_flags and sq_count.
5296 		 *
5297 		 * is it possible to have c_inner set when c_outer is not set?
5298 		 */
5299 		if (!(type & c_inner)) {
5300 			SQ_PUTLOCKS_EXIT(sq);
5301 		}
5302 		mutex_exit(SQLOCK(sq));
5303 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5304 		mutex_enter(SQLOCK(sq));
5305 		flags = sq->sq_flags;
5306 		/*
5307 		 * there should be no need to recheck sq_putcounts
5308 		 * because outer_enter() has already waited for them to clear
5309 		 * after setting SQ_WRITER.
5310 		 */
5311 		count = sq->sq_count;
5312 #ifdef DEBUG
5313 		/*
5314 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5315 		 * of doing an ASSERT internally. Others should do
5316 		 * something like
5317 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5318 		 * without the need to #ifdef DEBUG it.
5319 		 */
5320 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5321 #endif
5322 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5323 		    (!(type & c_inner) && count != 0)) {
5324 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5325 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5326 			count = sq->sq_count;
5327 			flags = sq->sq_flags;
5328 		}
5329 	}
5330 
5331 	sq->sq_count++;
5332 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5333 	if (!(type & c_inner)) {
5334 		/* Exclusive entry */
5335 		ASSERT(sq->sq_count == 1);
5336 		sq->sq_flags |= SQ_EXCL;
5337 		if (type & c_outer) {
5338 			SQ_PUTLOCKS_EXIT(sq);
5339 		}
5340 	}
5341 	mutex_exit(SQLOCK(sq));
5342 }
5343 
5344 /*
5345  * leave a syncq. announce to framework that closes may proceed.
5346  * c_inner and c_outer specifies which concurrency bits
5347  * to check.
5348  *
5349  * must never be called from driver or module put entry point.
5350  *
5351  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5352  * sq_putlocks are used.
5353  */
5354 void
5355 leavesq(syncq_t *sq, int entrypoint)
5356 {
5357 	uint16_t	flags;
5358 	uint16_t	type;
5359 	uint_t		c_outer = entrypoint & SQ_CO;
5360 #ifdef DEBUG
5361 	uint_t		c_inner = entrypoint & SQ_CI;
5362 #endif
5363 
5364 	/*
5365 	 * decrement ref count, drain the syncq if possible, and wake up
5366 	 * any waiting close.
5367 	 */
5368 	ASSERT(sq);
5369 	ASSERT(c_inner && c_outer);
5370 	mutex_enter(SQLOCK(sq));
5371 	flags = sq->sq_flags;
5372 	type = sq->sq_type;
5373 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5374 
5375 		if (flags & SQ_WANTWAKEUP) {
5376 			flags &= ~SQ_WANTWAKEUP;
5377 			cv_broadcast(&sq->sq_wait);
5378 		}
5379 		if (flags & SQ_WANTEXWAKEUP) {
5380 			flags &= ~SQ_WANTEXWAKEUP;
5381 			cv_broadcast(&sq->sq_exitwait);
5382 		}
5383 
5384 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5385 			/*
5386 			 * The syncq needs to be drained. "Exit" the syncq
5387 			 * before calling drain_syncq.
5388 			 */
5389 			ASSERT(sq->sq_count != 0);
5390 			sq->sq_count--;
5391 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5392 			sq->sq_flags = flags & ~SQ_EXCL;
5393 			drain_syncq(sq);
5394 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5395 			/* Check if we need to exit the outer perimeter */
5396 			/* XXX will this ever be true? */
5397 			if (!(type & c_outer))
5398 				outer_exit(sq->sq_outer);
5399 			return;
5400 		}
5401 	}
5402 	ASSERT(sq->sq_count != 0);
5403 	sq->sq_count--;
5404 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5405 	sq->sq_flags = flags & ~SQ_EXCL;
5406 	mutex_exit(SQLOCK(sq));
5407 
5408 	/* Check if we need to exit the outer perimeter */
5409 	if (!(sq->sq_type & c_outer))
5410 		outer_exit(sq->sq_outer);
5411 }
5412 
5413 /*
5414  * Prevent q_next from changing in this stream by incrementing sq_count.
5415  *
5416  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5417  * sq_putlocks are used.
5418  */
5419 void
5420 claimq(queue_t *qp)
5421 {
5422 	syncq_t	*sq = qp->q_syncq;
5423 
5424 	mutex_enter(SQLOCK(sq));
5425 	sq->sq_count++;
5426 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5427 	mutex_exit(SQLOCK(sq));
5428 }
5429 
5430 /*
5431  * Undo claimq.
5432  *
5433  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
5434  * sq_putlocks are used.
5435  */
5436 void
5437 releaseq(queue_t *qp)
5438 {
5439 	syncq_t	*sq = qp->q_syncq;
5440 	uint16_t flags;
5441 
5442 	mutex_enter(SQLOCK(sq));
5443 	ASSERT(sq->sq_count > 0);
5444 	sq->sq_count--;
5445 
5446 	flags = sq->sq_flags;
5447 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5448 		if (flags & SQ_WANTWAKEUP) {
5449 			flags &= ~SQ_WANTWAKEUP;
5450 			cv_broadcast(&sq->sq_wait);
5451 		}
5452 		sq->sq_flags = flags;
5453 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5454 			/*
5455 			 * To prevent potential recursive invocation of
5456 			 * drain_syncq we do not call drain_syncq if count is
5457 			 * non-zero.
5458 			 */
5459 			if (sq->sq_count == 0) {
5460 				drain_syncq(sq);
5461 				return;
5462 			} else
5463 				sqenable(sq);
5464 		}
5465 	}
5466 	mutex_exit(SQLOCK(sq));
5467 }
5468 
5469 /*
5470  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5471  */
5472 void
5473 claimstr(queue_t *qp)
5474 {
5475 	struct stdata *stp = STREAM(qp);
5476 
5477 	mutex_enter(&stp->sd_reflock);
5478 	stp->sd_refcnt++;
5479 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5480 	mutex_exit(&stp->sd_reflock);
5481 }
5482 
5483 /*
5484  * Undo claimstr.
5485  */
5486 void
5487 releasestr(queue_t *qp)
5488 {
5489 	struct stdata *stp = STREAM(qp);
5490 
5491 	mutex_enter(&stp->sd_reflock);
5492 	ASSERT(stp->sd_refcnt != 0);
5493 	if (--stp->sd_refcnt == 0)
5494 		cv_broadcast(&stp->sd_refmonitor);
5495 	mutex_exit(&stp->sd_reflock);
5496 }
5497 
5498 static syncq_t *
5499 new_syncq(void)
5500 {
5501 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5502 }
5503 
5504 static void
5505 free_syncq(syncq_t *sq)
5506 {
5507 	ASSERT(sq->sq_head == NULL);
5508 	ASSERT(sq->sq_outer == NULL);
5509 	ASSERT(sq->sq_callbpend == NULL);
5510 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5511 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5512 
5513 	if (sq->sq_ciputctrl != NULL) {
5514 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5515 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5516 		    sq->sq_nciputctrl, 0);
5517 		ASSERT(ciputctrl_cache != NULL);
5518 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5519 	}
5520 
5521 	sq->sq_tail = NULL;
5522 	sq->sq_evhead = NULL;
5523 	sq->sq_evtail = NULL;
5524 	sq->sq_ciputctrl = NULL;
5525 	sq->sq_nciputctrl = 0;
5526 	sq->sq_count = 0;
5527 	sq->sq_rmqcount = 0;
5528 	sq->sq_callbflags = 0;
5529 	sq->sq_cancelid = 0;
5530 	sq->sq_next = NULL;
5531 	sq->sq_needexcl = 0;
5532 	sq->sq_svcflags = 0;
5533 	sq->sq_nqueues = 0;
5534 	sq->sq_pri = 0;
5535 	sq->sq_onext = NULL;
5536 	sq->sq_oprev = NULL;
5537 	sq->sq_flags = 0;
5538 	sq->sq_type = 0;
5539 	sq->sq_servcount = 0;
5540 
5541 	kmem_cache_free(syncq_cache, sq);
5542 }
5543 
5544 /* Outer perimeter code */
5545 
5546 /*
5547  * The outer syncq uses the fields and flags in the syncq slightly
5548  * differently from the inner syncqs.
5549  *	sq_count	Incremented when there are pending or running
5550  *			writers at the outer perimeter to prevent the set of
5551  *			inner syncqs that belong to the outer perimeter from
5552  *			changing.
5553  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5554  *
5555  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5556  *			inner syncqs are added to or removed from the
5557  *			outer perimeter.
5558  *	SQ_QUEUED	sq_head/tail has messages or eventsqueued.
5559  *
5560  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5561  *			setting the SQ_WRITER flag.
5562  */
5563 
5564 /*
5565  * Get write access at the outer perimeter.
5566  * Note that read access is done by entersq, putnext, and put by simply
5567  * incrementing sq_count in the inner syncq.
5568  *
5569  * Waits until "flags" is no longer set in the outer to prevent multiple
5570  * threads from having write access at the same time. SQ_WRITER has to be part
5571  * of "flags".
5572  *
5573  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5574  * until the outer_exit is finished.
5575  *
5576  * outer_enter is vulnerable to starvation since it does not prevent new
5577  * threads from entering the inner syncqs while it is waiting for sq_count to
5578  * go to zero.
5579  */
5580 void
5581 outer_enter(syncq_t *outer, uint16_t flags)
5582 {
5583 	syncq_t	*sq;
5584 	int	wait_needed;
5585 	uint16_t	count;
5586 
5587 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5588 	    outer->sq_oprev != NULL);
5589 	ASSERT(flags & SQ_WRITER);
5590 
5591 retry:
5592 	mutex_enter(SQLOCK(outer));
5593 	while (outer->sq_flags & flags) {
5594 		outer->sq_flags |= SQ_WANTWAKEUP;
5595 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5596 	}
5597 
5598 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5599 	outer->sq_flags |= SQ_WRITER;
5600 	outer->sq_count++;
5601 	ASSERT(outer->sq_count != 0);	/* wraparound */
5602 	wait_needed = 0;
5603 	/*
5604 	 * Set SQ_WRITER on all the inner syncqs while holding
5605 	 * the SQLOCK on the outer syncq. This ensures that the changing
5606 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5607 	 */
5608 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5609 		mutex_enter(SQLOCK(sq));
5610 		count = sq->sq_count;
5611 		SQ_PUTLOCKS_ENTER(sq);
5612 		sq->sq_flags |= SQ_WRITER;
5613 		SUM_SQ_PUTCOUNTS(sq, count);
5614 		if (count != 0)
5615 			wait_needed = 1;
5616 		SQ_PUTLOCKS_EXIT(sq);
5617 		mutex_exit(SQLOCK(sq));
5618 	}
5619 	mutex_exit(SQLOCK(outer));
5620 
5621 	/*
5622 	 * Get everybody out of the syncqs sequentially.
5623 	 * Note that we don't actually need to aqiure the PUTLOCKS, since
5624 	 * we have already cleared the fastbit, and set QWRITER.  By
5625 	 * definition, the count can not increase since putnext will
5626 	 * take the slowlock path (and the purpose of aquiring the
5627 	 * putlocks was to make sure it didn't increase while we were
5628 	 * waiting).
5629 	 *
5630 	 * Note that we still aquire the PUTLOCKS to be safe.
5631 	 */
5632 	if (wait_needed) {
5633 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5634 			mutex_enter(SQLOCK(sq));
5635 			count = sq->sq_count;
5636 			SQ_PUTLOCKS_ENTER(sq);
5637 			SUM_SQ_PUTCOUNTS(sq, count);
5638 			while (count != 0) {
5639 				sq->sq_flags |= SQ_WANTWAKEUP;
5640 				SQ_PUTLOCKS_EXIT(sq);
5641 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5642 				count = sq->sq_count;
5643 				SQ_PUTLOCKS_ENTER(sq);
5644 				SUM_SQ_PUTCOUNTS(sq, count);
5645 			}
5646 			SQ_PUTLOCKS_EXIT(sq);
5647 			mutex_exit(SQLOCK(sq));
5648 		}
5649 		/*
5650 		 * Verify that none of the flags got set while we
5651 		 * were waiting for the sq_counts to drop.
5652 		 * If this happens we exit and retry entering the
5653 		 * outer perimeter.
5654 		 */
5655 		mutex_enter(SQLOCK(outer));
5656 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5657 			mutex_exit(SQLOCK(outer));
5658 			outer_exit(outer);
5659 			goto retry;
5660 		}
5661 		mutex_exit(SQLOCK(outer));
5662 	}
5663 }
5664 
5665 /*
5666  * Drop the write access at the outer perimeter.
5667  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5668  * decrementing sq_count.
5669  */
5670 void
5671 outer_exit(syncq_t *outer)
5672 {
5673 	syncq_t	*sq;
5674 	int	 drain_needed;
5675 	uint16_t flags;
5676 
5677 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5678 	    outer->sq_oprev != NULL);
5679 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5680 
5681 	/*
5682 	 * Atomically (from the perspective of threads calling become_writer)
5683 	 * drop the write access at the outer perimeter by holding
5684 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5685 	 * SQ_WRITER.
5686 	 * This defines a locking order between the outer perimeter
5687 	 * SQLOCK and the inner perimeter SQLOCKs.
5688 	 */
5689 	mutex_enter(SQLOCK(outer));
5690 	flags = outer->sq_flags;
5691 	ASSERT(outer->sq_flags & SQ_WRITER);
5692 	if (flags & SQ_QUEUED) {
5693 		write_now(outer);
5694 		flags = outer->sq_flags;
5695 	}
5696 
5697 	/*
5698 	 * sq_onext is stable since sq_count has not yet been decreased.
5699 	 * Reset the SQ_WRITER flags in all syncqs.
5700 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5701 	 * inner syncqs.
5702 	 */
5703 	drain_needed = 0;
5704 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5705 		drain_needed += dropsq(sq, SQ_WRITER);
5706 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5707 	flags &= ~SQ_WRITER;
5708 	if (drain_needed) {
5709 		outer->sq_flags = flags;
5710 		mutex_exit(SQLOCK(outer));
5711 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5712 			emptysq(sq);
5713 		mutex_enter(SQLOCK(outer));
5714 		flags = outer->sq_flags;
5715 	}
5716 	if (flags & SQ_WANTWAKEUP) {
5717 		flags &= ~SQ_WANTWAKEUP;
5718 		cv_broadcast(&outer->sq_wait);
5719 	}
5720 	outer->sq_flags = flags;
5721 	ASSERT(outer->sq_count > 0);
5722 	outer->sq_count--;
5723 	mutex_exit(SQLOCK(outer));
5724 }
5725 
5726 /*
5727  * Add another syncq to 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_insert(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 == NULL && sq->sq_onext == NULL &&
5740 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5741 
5742 	/* Get exclusive access to the outer perimeter list */
5743 	blocksq(outer, SQ_BLOCKED, 0);
5744 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5745 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5746 
5747 	mutex_enter(SQLOCK(sq));
5748 	sq->sq_outer = outer;
5749 	outer->sq_onext->sq_oprev = sq;
5750 	sq->sq_onext = outer->sq_onext;
5751 	outer->sq_onext = sq;
5752 	sq->sq_oprev = outer;
5753 	mutex_exit(SQLOCK(sq));
5754 	unblocksq(outer, SQ_BLOCKED, 1);
5755 }
5756 
5757 /*
5758  * Remove a syncq from an outer perimeter.
5759  * Block out all other access to the outer perimeter while it is being
5760  * changed using blocksq.
5761  * Assumes that the caller has *not* done an outer_enter.
5762  *
5763  * Vulnerable to starvation in blocksq.
5764  */
5765 static void
5766 outer_remove(syncq_t *outer, syncq_t *sq)
5767 {
5768 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5769 	    outer->sq_oprev != NULL);
5770 	ASSERT(sq->sq_outer == outer);
5771 
5772 	/* Get exclusive access to the outer perimeter list */
5773 	blocksq(outer, SQ_BLOCKED, 0);
5774 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5775 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5776 
5777 	mutex_enter(SQLOCK(sq));
5778 	sq->sq_outer = NULL;
5779 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5780 	sq->sq_oprev->sq_onext = sq->sq_onext;
5781 	sq->sq_oprev = sq->sq_onext = NULL;
5782 	mutex_exit(SQLOCK(sq));
5783 	unblocksq(outer, SQ_BLOCKED, 1);
5784 }
5785 
5786 /*
5787  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5788  * If this is the first callback for this outer perimeter then add
5789  * this outer perimeter to the list of outer perimeters that
5790  * the qwriter_outer_thread will process.
5791  *
5792  * Increments sq_count in the outer syncq to prevent the membership
5793  * of the outer perimeter (in terms of inner syncqs) to change while
5794  * the callback is pending.
5795  */
5796 static void
5797 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5798 {
5799 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5800 
5801 	mp->b_prev = (mblk_t *)func;
5802 	mp->b_queue = q;
5803 	mp->b_next = NULL;
5804 	outer->sq_count++;	/* Decremented when dequeued */
5805 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5806 	if (outer->sq_evhead == NULL) {
5807 		/* First message. */
5808 		outer->sq_evhead = outer->sq_evtail = mp;
5809 		outer->sq_flags |= SQ_EVENTS;
5810 		mutex_exit(SQLOCK(outer));
5811 		STRSTAT(qwr_outer);
5812 		(void) taskq_dispatch(streams_taskq,
5813 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5814 	} else {
5815 		ASSERT(outer->sq_flags & SQ_EVENTS);
5816 		outer->sq_evtail->b_next = mp;
5817 		outer->sq_evtail = mp;
5818 		mutex_exit(SQLOCK(outer));
5819 	}
5820 }
5821 
5822 /*
5823  * Try and upgrade to write access at the outer perimeter. If this can
5824  * not be done without blocking then queue the callback to be done
5825  * by the qwriter_outer_thread.
5826  *
5827  * This routine can only be called from put or service procedures plus
5828  * asynchronous callback routines that have properly entered to
5829  * queue (with entersq.) Thus qwriter(OUTER) assumes the caller has one claim
5830  * on the syncq associated with q.
5831  */
5832 void
5833 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5834 {
5835 	syncq_t	*osq, *sq, *outer;
5836 	int	failed;
5837 	uint16_t flags;
5838 
5839 	osq = q->q_syncq;
5840 	outer = osq->sq_outer;
5841 	if (outer == NULL)
5842 		panic("qwriter(PERIM_OUTER): no outer perimeter");
5843 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5844 	    outer->sq_oprev != NULL);
5845 
5846 	mutex_enter(SQLOCK(outer));
5847 	flags = outer->sq_flags;
5848 	/*
5849 	 * If some thread is traversing sq_next, or if we are blocked by
5850 	 * outer_insert or outer_remove, or if the we already have queued
5851 	 * callbacks, then queue this callback for later processing.
5852 	 *
5853 	 * Also queue the qwriter for an interrupt thread in order
5854 	 * to reduce the time spent running at high IPL.
5855 	 * to identify there are events.
5856 	 */
5857 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
5858 		/*
5859 		 * Queue the become_writer request.
5860 		 * The queueing is atomic under SQLOCK(outer) in order
5861 		 * to synchronize with outer_exit.
5862 		 * queue_writer will drop the outer SQLOCK
5863 		 */
5864 		if (flags & SQ_BLOCKED) {
5865 			/* Must set SQ_WRITER on inner perimeter */
5866 			mutex_enter(SQLOCK(osq));
5867 			osq->sq_flags |= SQ_WRITER;
5868 			mutex_exit(SQLOCK(osq));
5869 		} else {
5870 			if (!(flags & SQ_WRITER)) {
5871 				/*
5872 				 * The outer could have been SQ_BLOCKED thus
5873 				 * SQ_WRITER might not be set on the inner.
5874 				 */
5875 				mutex_enter(SQLOCK(osq));
5876 				osq->sq_flags |= SQ_WRITER;
5877 				mutex_exit(SQLOCK(osq));
5878 			}
5879 			ASSERT(osq->sq_flags & SQ_WRITER);
5880 		}
5881 		queue_writer(outer, func, q, mp);
5882 		return;
5883 	}
5884 	/*
5885 	 * We are half-way to exclusive access to the outer perimeter.
5886 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
5887 	 * while the inner syncqs are traversed.
5888 	 */
5889 	outer->sq_count++;
5890 	ASSERT(outer->sq_count != 0);	/* wraparound */
5891 	flags |= SQ_WRITER;
5892 	/*
5893 	 * Check if we can run the function immediately. Mark all
5894 	 * syncqs with the writer flag to prevent new entries into
5895 	 * put and service procedures.
5896 	 *
5897 	 * Set SQ_WRITER on all the inner syncqs while holding
5898 	 * the SQLOCK on the outer syncq. This ensures that the changing
5899 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5900 	 */
5901 	failed = 0;
5902 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5903 		uint16_t count;
5904 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
5905 
5906 		mutex_enter(SQLOCK(sq));
5907 		count = sq->sq_count;
5908 		SQ_PUTLOCKS_ENTER(sq);
5909 		SUM_SQ_PUTCOUNTS(sq, count);
5910 		if (sq->sq_count > maxcnt)
5911 			failed = 1;
5912 		sq->sq_flags |= SQ_WRITER;
5913 		SQ_PUTLOCKS_EXIT(sq);
5914 		mutex_exit(SQLOCK(sq));
5915 	}
5916 	if (failed) {
5917 		/*
5918 		 * Some other thread has a read claim on the outer perimeter.
5919 		 * Queue the callback for deferred processing.
5920 		 *
5921 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
5922 		 * so that other qwriter(OUTER) calls will queue their
5923 		 * callbacks as well. queue_writer increments sq_count so we
5924 		 * decrement to compensate for the our increment.
5925 		 *
5926 		 * Dropping SQ_WRITER enables the writer thread to work
5927 		 * on this outer perimeter.
5928 		 */
5929 		outer->sq_flags = flags;
5930 		queue_writer(outer, func, q, mp);
5931 		/* queue_writer dropper the lock */
5932 		mutex_enter(SQLOCK(outer));
5933 		ASSERT(outer->sq_count > 0);
5934 		outer->sq_count--;
5935 		ASSERT(outer->sq_flags & SQ_WRITER);
5936 		flags = outer->sq_flags;
5937 		flags &= ~SQ_WRITER;
5938 		if (flags & SQ_WANTWAKEUP) {
5939 			flags &= ~SQ_WANTWAKEUP;
5940 			cv_broadcast(&outer->sq_wait);
5941 		}
5942 		outer->sq_flags = flags;
5943 		mutex_exit(SQLOCK(outer));
5944 		return;
5945 	} else {
5946 		outer->sq_flags = flags;
5947 		mutex_exit(SQLOCK(outer));
5948 	}
5949 
5950 	/* Can run it immediately */
5951 	(*func)(q, mp);
5952 
5953 	outer_exit(outer);
5954 }
5955 
5956 /*
5957  * Dequeue all writer callbacks from the outer perimeter and run them.
5958  */
5959 static void
5960 write_now(syncq_t *outer)
5961 {
5962 	mblk_t		*mp;
5963 	queue_t		*q;
5964 	void	(*func)();
5965 
5966 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5967 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5968 	    outer->sq_oprev != NULL);
5969 	while ((mp = outer->sq_evhead) != NULL) {
5970 		/*
5971 		 * queues cannot be placed on the queuelist on the outer
5972 		 * perimiter.
5973 		 */
5974 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
5975 		ASSERT((outer->sq_flags & SQ_EVENTS));
5976 
5977 		outer->sq_evhead = mp->b_next;
5978 		if (outer->sq_evhead == NULL) {
5979 			outer->sq_evtail = NULL;
5980 			outer->sq_flags &= ~SQ_EVENTS;
5981 		}
5982 		ASSERT(outer->sq_count != 0);
5983 		outer->sq_count--;	/* Incremented when enqueued. */
5984 		mutex_exit(SQLOCK(outer));
5985 		/*
5986 		 * Drop the message if the queue is closing.
5987 		 * Make sure that the queue is "claimed" when the callback
5988 		 * is run in order to satisfy various ASSERTs.
5989 		 */
5990 		q = mp->b_queue;
5991 		func = (void (*)())mp->b_prev;
5992 		ASSERT(func != NULL);
5993 		mp->b_next = mp->b_prev = NULL;
5994 		if (q->q_flag & QWCLOSE) {
5995 			freemsg(mp);
5996 		} else {
5997 			claimq(q);
5998 			(*func)(q, mp);
5999 			releaseq(q);
6000 		}
6001 		mutex_enter(SQLOCK(outer));
6002 	}
6003 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6004 }
6005 
6006 /*
6007  * The list of messages on the inner syncq is effectively hashed
6008  * by destination queue.  These destination queues are doubly
6009  * linked lists (hopefully) in priority order.  Messages are then
6010  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6011  * Additional messages are linked together by the b_next/b_prev
6012  * elements in the mblk, with (similar to putq()) the first message
6013  * having a NULL b_prev and the last message having a NULL b_next.
6014  *
6015  * Events, such as qwriter callbacks, are put onto a list in FIFO
6016  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6017  * linked list, and messages here MUST be processed in the order queued.
6018  */
6019 
6020 /*
6021  * Run the events on the syncq event list (sq_evhead).
6022  * Assumes there is only one claim on the syncq, it is
6023  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6024  * Messages here are processed in order, with the SQ_EXCL bit
6025  * held all the way through till the last message is processed.
6026  */
6027 void
6028 sq_run_events(syncq_t *sq)
6029 {
6030 	mblk_t		*bp;
6031 	queue_t		*qp;
6032 	uint16_t	flags = sq->sq_flags;
6033 	void		(*func)();
6034 
6035 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6036 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6037 		sq->sq_oprev == NULL) ||
6038 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6039 		sq->sq_oprev != NULL));
6040 
6041 	ASSERT(flags & SQ_EXCL);
6042 	ASSERT(sq->sq_count == 1);
6043 
6044 	/*
6045 	 * We need to process all of the events on this list.  It
6046 	 * is possible that new events will be added while we are
6047 	 * away processing a callback, so on every loop, we start
6048 	 * back at the beginning of the list.
6049 	 */
6050 	/*
6051 	 * We have to reaccess sq_evhead since there is a
6052 	 * possibility of a new entry while we were running
6053 	 * the callback.
6054 	 */
6055 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6056 		ASSERT(bp->b_queue->q_syncq == sq);
6057 		ASSERT(sq->sq_flags & SQ_EVENTS);
6058 
6059 		qp = bp->b_queue;
6060 		func = (void (*)())bp->b_prev;
6061 		ASSERT(func != NULL);
6062 
6063 		/*
6064 		 * Messages from the event queue must be taken off in
6065 		 * FIFO order.
6066 		 */
6067 		ASSERT(sq->sq_evhead == bp);
6068 		sq->sq_evhead = bp->b_next;
6069 
6070 		if (bp->b_next == NULL) {
6071 			/* Deleting last */
6072 			ASSERT(sq->sq_evtail == bp);
6073 			sq->sq_evtail = NULL;
6074 			sq->sq_flags &= ~SQ_EVENTS;
6075 		}
6076 		bp->b_prev = bp->b_next = NULL;
6077 		ASSERT(bp->b_datap->db_ref != 0);
6078 
6079 		mutex_exit(SQLOCK(sq));
6080 
6081 		(*func)(qp, bp);
6082 
6083 		mutex_enter(SQLOCK(sq));
6084 		/*
6085 		 * re-read the flags, since they could have changed.
6086 		 */
6087 		flags = sq->sq_flags;
6088 		ASSERT(flags & SQ_EXCL);
6089 	}
6090 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6091 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6092 
6093 	if (flags & SQ_WANTWAKEUP) {
6094 		flags &= ~SQ_WANTWAKEUP;
6095 		cv_broadcast(&sq->sq_wait);
6096 	}
6097 	if (flags & SQ_WANTEXWAKEUP) {
6098 		flags &= ~SQ_WANTEXWAKEUP;
6099 		cv_broadcast(&sq->sq_exitwait);
6100 	}
6101 	sq->sq_flags = flags;
6102 }
6103 
6104 /*
6105  * Put messages on the event list.
6106  * If we can go exclusive now, do so and process the event list, otherwise
6107  * let the last claim service this list (or wake the sqthread).
6108  * This procedure assumes SQLOCK is held.  To run the event list, it
6109  * must be called with no claims.
6110  */
6111 static void
6112 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6113 {
6114 	uint16_t count;
6115 
6116 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6117 	ASSERT(func != NULL);
6118 
6119 	/*
6120 	 * This is a callback.  Add it to the list of callbacks
6121 	 * and see about upgrading.
6122 	 */
6123 	mp->b_prev = (mblk_t *)func;
6124 	mp->b_queue = q;
6125 	mp->b_next = NULL;
6126 	if (sq->sq_evhead == NULL) {
6127 		sq->sq_evhead = sq->sq_evtail = mp;
6128 		sq->sq_flags |= SQ_EVENTS;
6129 	} else {
6130 		ASSERT(sq->sq_evtail != NULL);
6131 		ASSERT(sq->sq_evtail->b_next == NULL);
6132 		ASSERT(sq->sq_flags & SQ_EVENTS);
6133 		sq->sq_evtail->b_next = mp;
6134 		sq->sq_evtail = mp;
6135 	}
6136 	/*
6137 	 * We have set SQ_EVENTS, so threads will have to
6138 	 * unwind out of the perimiter, and new entries will
6139 	 * not grab a putlock.  But we still need to know
6140 	 * how many threads have already made a claim to the
6141 	 * syncq, so grab the putlocks, and sum the counts.
6142 	 * If there are no claims on the syncq, we can upgrade
6143 	 * to exclusive, and run the event list.
6144 	 * NOTE: We hold the SQLOCK, so we can just grab the
6145 	 * putlocks.
6146 	 */
6147 	count = sq->sq_count;
6148 	SQ_PUTLOCKS_ENTER(sq);
6149 	SUM_SQ_PUTCOUNTS(sq, count);
6150 	/*
6151 	 * We have no claim, so we need to check if there
6152 	 * are no others, then we can upgrade.
6153 	 */
6154 	/*
6155 	 * There are currently no claims on
6156 	 * the syncq by this thread (at least on this entry). The thread who has
6157 	 * the claim should drain syncq.
6158 	 */
6159 	if (count > 0) {
6160 		/*
6161 		 * Can't upgrade - other threads inside.
6162 		 */
6163 		SQ_PUTLOCKS_EXIT(sq);
6164 		mutex_exit(SQLOCK(sq));
6165 		return;
6166 	}
6167 	/*
6168 	 * Need to set SQ_EXCL and make a claim on the syncq.
6169 	 */
6170 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6171 	sq->sq_flags |= SQ_EXCL;
6172 	ASSERT(sq->sq_count == 0);
6173 	sq->sq_count++;
6174 	SQ_PUTLOCKS_EXIT(sq);
6175 
6176 	/* Process the events list */
6177 	sq_run_events(sq);
6178 
6179 	/*
6180 	 * Release our claim...
6181 	 */
6182 	sq->sq_count--;
6183 
6184 	/*
6185 	 * And release SQ_EXCL.
6186 	 * We don't need to acquire the putlocks to release
6187 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6188 	 */
6189 	sq->sq_flags &= ~SQ_EXCL;
6190 
6191 	/*
6192 	 * sq_run_events should have released SQ_EXCL
6193 	 */
6194 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6195 
6196 	/*
6197 	 * If anything happened while we were running the
6198 	 * events (or was there before), we need to process
6199 	 * them now.  We shouldn't be exclusive sine we
6200 	 * released the perimiter above (plus, we asserted
6201 	 * for it).
6202 	 */
6203 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6204 		drain_syncq(sq);
6205 	else
6206 		mutex_exit(SQLOCK(sq));
6207 }
6208 
6209 /*
6210  * Perform delayed processing. The caller has to make sure that it is safe
6211  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6212  * set.)
6213  *
6214  * Assume that the caller has NO claims on the syncq.  However, a claim
6215  * on the syncq does not indicate that a thread is draining the syncq.
6216  * There may be more claims on the syncq than there are threads draining
6217  * (i.e.  #_threads_draining <= sq_count)
6218  *
6219  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6220  * in order to preserve qwriter(OUTER) ordering constraints.
6221  *
6222  * sq_putcount only needs to be checked when dispatching the queued
6223  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6224  */
6225 void
6226 drain_syncq(syncq_t *sq)
6227 {
6228 	queue_t		*qp;
6229 	uint16_t	count;
6230 	uint16_t	type = sq->sq_type;
6231 	uint16_t	flags = sq->sq_flags;
6232 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6233 
6234 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6235 		"drain_syncq start:%p", sq);
6236 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6237 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6238 		sq->sq_oprev == NULL) ||
6239 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6240 		sq->sq_oprev != NULL));
6241 
6242 	/*
6243 	 * Drop SQ_SERVICE flag.
6244 	 */
6245 	if (bg_service)
6246 		sq->sq_svcflags &= ~SQ_SERVICE;
6247 
6248 	/*
6249 	 * If SQ_EXCL is set, someone else is processing this syncq - let him
6250 	 * finish the job.
6251 	 */
6252 	if (flags & SQ_EXCL) {
6253 		if (bg_service) {
6254 			ASSERT(sq->sq_servcount != 0);
6255 			sq->sq_servcount--;
6256 		}
6257 		mutex_exit(SQLOCK(sq));
6258 		return;
6259 	}
6260 
6261 	/*
6262 	 * This routine can be called by a background thread if
6263 	 * it was scheduled by a hi-priority thread.  SO, if there are
6264 	 * NOT messages queued, return (remember, we have the SQLOCK,
6265 	 * and it cannot change until we release it). Wakeup any waiters also.
6266 	 */
6267 	if (!(flags & SQ_QUEUED)) {
6268 		if (flags & SQ_WANTWAKEUP) {
6269 			flags &= ~SQ_WANTWAKEUP;
6270 			cv_broadcast(&sq->sq_wait);
6271 		}
6272 		if (flags & SQ_WANTEXWAKEUP) {
6273 			flags &= ~SQ_WANTEXWAKEUP;
6274 			cv_broadcast(&sq->sq_exitwait);
6275 		}
6276 		sq->sq_flags = flags;
6277 		if (bg_service) {
6278 			ASSERT(sq->sq_servcount != 0);
6279 			sq->sq_servcount--;
6280 		}
6281 		mutex_exit(SQLOCK(sq));
6282 		return;
6283 	}
6284 
6285 	/*
6286 	 * If this is not a concurrent put perimiter, we need to
6287 	 * become exclusive to drain.  Also, if not CIPUT, we would
6288 	 * not have acquired a putlock, so we don't need to check
6289 	 * the putcounts.  If not entering with a claim, we test
6290 	 * for sq_count == 0.
6291 	 */
6292 	type = sq->sq_type;
6293 	if (!(type & SQ_CIPUT)) {
6294 		if (sq->sq_count > 1) {
6295 			if (bg_service) {
6296 				ASSERT(sq->sq_servcount != 0);
6297 				sq->sq_servcount--;
6298 			}
6299 			mutex_exit(SQLOCK(sq));
6300 			return;
6301 		}
6302 		sq->sq_flags |= SQ_EXCL;
6303 	}
6304 
6305 	/*
6306 	 * This is where we make a claim to the syncq.
6307 	 * This can either be done by incrementing a putlock, or
6308 	 * the sq_count.  But since we already have the SQLOCK
6309 	 * here, we just bump the sq_count.
6310 	 *
6311 	 * Note that after we make a claim, we need to let the code
6312 	 * fall through to the end of this routine to clean itself
6313 	 * up.  A return in the while loop will put the syncq in a
6314 	 * very bad state.
6315 	 */
6316 	sq->sq_count++;
6317 	ASSERT(sq->sq_count != 0);	/* wraparound */
6318 
6319 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6320 		/*
6321 		 * If we are told to stayaway or went exclusive,
6322 		 * we are done.
6323 		 */
6324 		if (flags & (SQ_STAYAWAY)) {
6325 			break;
6326 		}
6327 
6328 		/*
6329 		 * If there are events to run, do so.
6330 		 * We have one claim to the syncq, so if there are
6331 		 * more than one, other threads are running.
6332 		 */
6333 		if (sq->sq_evhead != NULL) {
6334 			ASSERT(sq->sq_flags & SQ_EVENTS);
6335 
6336 			count = sq->sq_count;
6337 			SQ_PUTLOCKS_ENTER(sq);
6338 			SUM_SQ_PUTCOUNTS(sq, count);
6339 			if (count > 1) {
6340 				SQ_PUTLOCKS_EXIT(sq);
6341 				/* Can't upgrade - other threads inside */
6342 				break;
6343 			}
6344 			ASSERT((flags & SQ_EXCL) == 0);
6345 			sq->sq_flags = flags | SQ_EXCL;
6346 			SQ_PUTLOCKS_EXIT(sq);
6347 			/*
6348 			 * we have the only claim, run the events,
6349 			 * sq_run_events will clear the SQ_EXCL flag.
6350 			 */
6351 			sq_run_events(sq);
6352 
6353 			/*
6354 			 * If this is a CIPUT perimiter, we need
6355 			 * to drop the SQ_EXCL flag so we can properly
6356 			 * continue draining the syncq.
6357 			 */
6358 			if (type & SQ_CIPUT) {
6359 				ASSERT(sq->sq_flags & SQ_EXCL);
6360 				sq->sq_flags &= ~SQ_EXCL;
6361 			}
6362 
6363 			/*
6364 			 * And go back to the beginning just in case
6365 			 * anything changed while we were away.
6366 			 */
6367 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6368 			continue;
6369 		}
6370 
6371 		ASSERT(sq->sq_evhead == NULL);
6372 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6373 
6374 		/*
6375 		 * Find the queue that is not draining.
6376 		 *
6377 		 * q_draining is protected by QLOCK which we do not hold.
6378 		 * But if it was set, then a thread was draining, and if it gets
6379 		 * cleared, then it was because the thread has successfully
6380 		 * drained the syncq, or a GOAWAY state occured. For the GOAWAY
6381 		 * state to happen, a thread needs the SQLOCK which we hold, and
6382 		 * if there was such a flag, we whould have already seen it.
6383 		 */
6384 
6385 		for (qp = sq->sq_head;
6386 		    qp != NULL && (qp->q_draining ||
6387 			(qp->q_sqflags & Q_SQDRAINING));
6388 		    qp = qp->q_sqnext)
6389 			;
6390 
6391 		if (qp == NULL)
6392 			break;
6393 
6394 		/*
6395 		 * We have a queue to work on, and we hold the
6396 		 * SQLOCK and one claim, call qdrain_syncq.
6397 		 * This means we need to release the SQLOCK and
6398 		 * aquire the QLOCK (OK since we have a claim).
6399 		 * Note that qdrain_syncq will actually dequeue
6400 		 * this queue from the sq_head list when it is
6401 		 * convinced all the work is done and release
6402 		 * the QLOCK before returning.
6403 		 */
6404 		qp->q_sqflags |= Q_SQDRAINING;
6405 		mutex_exit(SQLOCK(sq));
6406 		mutex_enter(QLOCK(qp));
6407 		qdrain_syncq(sq, qp);
6408 		mutex_enter(SQLOCK(sq));
6409 
6410 		/* The queue is drained */
6411 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6412 		qp->q_sqflags &= ~Q_SQDRAINING;
6413 		/*
6414 		 * NOTE: After this point qp should not be used since it may be
6415 		 * closed.
6416 		 */
6417 	}
6418 
6419 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6420 	flags = sq->sq_flags;
6421 
6422 	/*
6423 	 * sq->sq_head cannot change because we hold the
6424 	 * sqlock. However, a thread CAN decide that it is no longer
6425 	 * going to drain that queue.  However, this should be due to
6426 	 * a GOAWAY state, and we should see that here.
6427 	 *
6428 	 * This loop is not very efficient. One solution may be adding a second
6429 	 * pointer to the "draining" queue, but it is difficult to do when
6430 	 * queues are inserted in the middle due to priority ordering. Another
6431 	 * possibility is to yank the queue out of the sq list and put it onto
6432 	 * the "draining list" and then put it back if it can't be drained.
6433 	 */
6434 
6435 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6436 		(type & SQ_CI) || sq->sq_head->q_draining);
6437 
6438 	/* Drop SQ_EXCL for non-CIPUT perimiters */
6439 	if (!(type & SQ_CIPUT))
6440 		flags &= ~SQ_EXCL;
6441 	ASSERT((flags & SQ_EXCL) == 0);
6442 
6443 	/* Wake up any waiters. */
6444 	if (flags & SQ_WANTWAKEUP) {
6445 		flags &= ~SQ_WANTWAKEUP;
6446 		cv_broadcast(&sq->sq_wait);
6447 	}
6448 	if (flags & SQ_WANTEXWAKEUP) {
6449 		flags &= ~SQ_WANTEXWAKEUP;
6450 		cv_broadcast(&sq->sq_exitwait);
6451 	}
6452 	sq->sq_flags = flags;
6453 
6454 	ASSERT(sq->sq_count != 0);
6455 	/* Release our claim. */
6456 	sq->sq_count--;
6457 
6458 	if (bg_service) {
6459 		ASSERT(sq->sq_servcount != 0);
6460 		sq->sq_servcount--;
6461 	}
6462 
6463 	mutex_exit(SQLOCK(sq));
6464 
6465 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6466 		"drain_syncq end:%p", sq);
6467 }
6468 
6469 
6470 /*
6471  *
6472  * qdrain_syncq can be called (currently) from only one of two places:
6473  *	drain_syncq
6474  * 	putnext  (or some variation of it).
6475  * and eventually
6476  * 	qwait(_sig)
6477  *
6478  * If called from drain_syncq, we found it in the list
6479  * of queue's needing service, so there is work to be done (or it
6480  * wouldn't be on the list).
6481  *
6482  * If called from some putnext variation, it was because the
6483  * perimiter is open, but messages are blocking a putnext and
6484  * there is not a thread working on it.  Now a thread could start
6485  * working on it while we are getting ready to do so ourself, but
6486  * the thread would set the q_draining flag, and we can spin out.
6487  *
6488  * As for qwait(_sig), I think I shall let it continue to call
6489  * drain_syncq directly (after all, it will get here eventually).
6490  *
6491  * qdrain_syncq has to terminate when:
6492  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6493  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6494  *
6495  * ASSUMES:
6496  *	One claim
6497  * 	QLOCK held
6498  * 	SQLOCK not held
6499  *	Will release QLOCK before returning
6500  */
6501 void
6502 qdrain_syncq(syncq_t *sq, queue_t *q)
6503 {
6504 	mblk_t		*bp;
6505 	boolean_t	do_clr;
6506 #ifdef DEBUG
6507 	uint16_t	count;
6508 #endif
6509 
6510 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6511 		"drain_syncq start:%p", sq);
6512 	ASSERT(q->q_syncq == sq);
6513 	ASSERT(MUTEX_HELD(QLOCK(q)));
6514 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6515 	/*
6516 	 * For non-CIPUT perimiters, we should be called with the
6517 	 * exclusive bit set already.  For non-CIPUT perimiters we
6518 	 * will be doing a concurrent drain, so it better not be set.
6519 	 */
6520 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6521 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6522 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6523 	/*
6524 	 * All outer pointers are set, or none of them are
6525 	 */
6526 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6527 		sq->sq_oprev == NULL) ||
6528 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6529 		sq->sq_oprev != NULL));
6530 #ifdef DEBUG
6531 	count = sq->sq_count;
6532 	/*
6533 	 * This is OK without the putlocks, because we have one
6534 	 * claim either from the sq_count, or a putcount.  We could
6535 	 * get an erroneous value from other counts, but ours won't
6536 	 * change, so one way or another, we will have at least a
6537 	 * value of one.
6538 	 */
6539 	SUM_SQ_PUTCOUNTS(sq, count);
6540 	ASSERT(count >= 1);
6541 #endif /* DEBUG */
6542 
6543 	/*
6544 	 * The first thing to do here, is find out if a thread is already
6545 	 * draining this queue or the queue is closing. If so, we are done,
6546 	 * just return. Also, if there are no messages, we are done as well.
6547 	 * Note that we check the q_sqhead since there is s window of
6548 	 * opportunity for us to enter here because Q_SQQUEUED was set, but is
6549 	 * not anymore.
6550 	 */
6551 	if (q->q_draining || (q->q_sqhead == NULL)) {
6552 		mutex_exit(QLOCK(q));
6553 		return;
6554 	}
6555 
6556 	/*
6557 	 * If the perimiter is exclusive, there is nothing we can
6558 	 * do right now, go away.
6559 	 * Note that there is nothing to prevent this case from changing
6560 	 * right after this check, but the spin-out will catch it.
6561 	 */
6562 
6563 	/* Tell other threads that we are draining this queue */
6564 	q->q_draining = 1;	/* Protected by QLOCK */
6565 
6566 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6567 
6568 		/*
6569 		 * Because we can enter this routine just because
6570 		 * a putnext is blocked, we need to spin out if
6571 		 * the perimiter wants to go exclusive as well
6572 		 * as just blocked. We need to spin out also if
6573 		 * events are queued on the syncq.
6574 		 * Don't check for SQ_EXCL, because non-CIPUT
6575 		 * perimiters would set it, and it can't become
6576 		 * exclusive while we hold a claim.
6577 		 */
6578 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6579 			break;
6580 		}
6581 
6582 #ifdef DEBUG
6583 		/*
6584 		 * Since we are in qdrain_syncq, we already know the queue,
6585 		 * but for sanity, we want to check this against the qp that
6586 		 * was passed in by bp->b_queue.
6587 		 */
6588 
6589 		ASSERT(bp->b_queue == q);
6590 		ASSERT(bp->b_queue->q_syncq == sq);
6591 		bp->b_queue = NULL;
6592 
6593 		/*
6594 		 * We would have the following check in the DEBUG code:
6595 		 *
6596 		 * if (bp->b_prev != NULL)  {
6597 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6598 		 * }
6599 		 *
6600 		 * This can't be done, however, since IP modifies qinfo
6601 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6602 		 * qinfo), invalidating the check.
6603 		 * So the assignment to func is left here, but the ASSERT itself
6604 		 * is removed until the whole issue is resolved.
6605 		 */
6606 #endif
6607 		ASSERT(q->q_sqhead == bp);
6608 		q->q_sqhead = bp->b_next;
6609 		bp->b_prev = bp->b_next = NULL;
6610 		ASSERT(q->q_syncqmsgs > 0);
6611 		mutex_exit(QLOCK(q));
6612 
6613 		ASSERT(bp->b_datap->db_ref != 0);
6614 
6615 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6616 
6617 		mutex_enter(QLOCK(q));
6618 		/*
6619 		 * We should decrement q_syncqmsgs only after executing the
6620 		 * put procedure to avoid a possible race with putnext().
6621 		 * In putnext() though it sees Q_SQQUEUED is set, there is
6622 		 * an optimization which allows putnext to call the put
6623 		 * procedure directly if (q_syncqmsgs == 0) and thus
6624 		 * a message reodering could otherwise occur.
6625 		 */
6626 		q->q_syncqmsgs--;
6627 
6628 		/*
6629 		 * Clear QFULL in the next service procedure queue if
6630 		 * this is the last message destined to that queue.
6631 		 *
6632 		 * It would make better sense to have some sort of
6633 		 * tunable for the low water mark, but these symantics
6634 		 * are not yet defined.  So, alas, we use a constant.
6635 		 */
6636 		do_clr = (q->q_syncqmsgs == 0);
6637 		mutex_exit(QLOCK(q));
6638 
6639 		if (do_clr)
6640 			clr_qfull(q);
6641 
6642 		mutex_enter(QLOCK(q));
6643 		/*
6644 		 * Always clear SQ_EXCL when CIPUT in order to handle
6645 		 * qwriter(INNER).
6646 		 */
6647 		/*
6648 		 * The putp() can call qwriter and get exclusive access
6649 		 * IFF this is the only claim.  So, we need to test for
6650 		 * this possibility so we can aquire the mutex and clear
6651 		 * the bit.
6652 		 */
6653 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6654 			mutex_enter(SQLOCK(sq));
6655 			sq->sq_flags &= ~SQ_EXCL;
6656 			mutex_exit(SQLOCK(sq));
6657 		}
6658 	}
6659 
6660 	/*
6661 	 * We should either have no queues on the syncq, or we were
6662 	 * told to goaway by a waiter (which we will wake up at the
6663 	 * end of this function).
6664 	 */
6665 	ASSERT((q->q_sqhead == NULL) ||
6666 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6667 
6668 	ASSERT(MUTEX_HELD(QLOCK(q)));
6669 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6670 
6671 	/*
6672 	 * Remove the q from the syncq list if all the messages are
6673 	 * drained.
6674 	 */
6675 	if (q->q_sqhead == NULL) {
6676 		mutex_enter(SQLOCK(sq));
6677 		if (q->q_sqflags & Q_SQQUEUED)
6678 			SQRM_Q(sq, q);
6679 		mutex_exit(SQLOCK(sq));
6680 		/*
6681 		 * Since the queue is removed from the list, reset its priority.
6682 		 */
6683 		q->q_spri = 0;
6684 	}
6685 
6686 	/*
6687 	 * Remember, the q_draining flag is used to let another
6688 	 * thread know that there is a thread currently draining
6689 	 * the messages for a queue.  Since we are now done with
6690 	 * this queue (even if there may be messages still there),
6691 	 * we need to clear this flag so some thread will work
6692 	 * on it if needed.
6693 	 */
6694 	ASSERT(q->q_draining);
6695 	q->q_draining = 0;
6696 
6697 	/* called with a claim, so OK to drop all locks. */
6698 	mutex_exit(QLOCK(q));
6699 
6700 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6701 		"drain_syncq end:%p", sq);
6702 }
6703 /* END OF QDRAIN_SYNCQ  */
6704 
6705 
6706 /*
6707  * This is the mate to qdrain_syncq, except that it is putting the
6708  * message onto the the queue instead draining.  Since the
6709  * message is destined for the queue that is selected, there is
6710  * no need to identify the function because the message is
6711  * intended for the put routine for the queue.  But this
6712  * routine will do it anyway just in case (but only for debug kernels).
6713  *
6714  * After the message is enqueued on the syncq, it calls putnext_tail()
6715  * which will schedule a background thread to actually process the message.
6716  *
6717  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6718  * SQLOCK(sq) and QLOCK(q) are not held.
6719  */
6720 void
6721 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6722 {
6723 	queue_t		*fq = NULL;
6724 
6725 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6726 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6727 	ASSERT(sq->sq_count > 0);
6728 	ASSERT(q->q_syncq == sq);
6729 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6730 		sq->sq_oprev == NULL) ||
6731 		(sq->sq_outer != NULL && sq->sq_onext != NULL &&
6732 		sq->sq_oprev != NULL));
6733 
6734 	mutex_enter(QLOCK(q));
6735 
6736 	/*
6737 	 * Set QFULL in next service procedure queue (that cares) if not
6738 	 * already set and if there are already more messages on the syncq
6739 	 * than sq_max_size.  If sq_max_size is 0, no flow control will be
6740 	 * asserted on any syncq.
6741 	 *
6742 	 * The fq here is the next queue with a service procedure.
6743 	 * This is where we would fail canputnext, so this is where we
6744 	 * need to set QFULL.
6745 	 *
6746 	 * LOCKING HIERARCHY: In the case when fq != q we need to
6747 	 *  a) Take QLOCK(fq) to set QFULL flag and
6748 	 *  b) Take sd_reflock in the case of the hot stream to update
6749 	 *  	sd_refcnt.
6750 	 * We already have QLOCK at this point. To avoid cross-locks with
6751 	 * freezestr() which grabs all QLOCKs and with strlock() which grabs
6752 	 * both SQLOCK and sd_reflock, we need to drop respective locks first.
6753 	 */
6754 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
6755 	    (q->q_syncqmsgs > sq_max_size)) {
6756 		if ((fq = q->q_nfsrv) == q) {
6757 			fq->q_flag |= QFULL;
6758 		} else {
6759 			mutex_exit(QLOCK(q));
6760 			mutex_enter(QLOCK(fq));
6761 			fq->q_flag |= QFULL;
6762 			mutex_exit(QLOCK(fq));
6763 			mutex_enter(QLOCK(q));
6764 		}
6765 	}
6766 
6767 #ifdef DEBUG
6768 	/*
6769 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6770 	 * to trace the queue that the message is intended for.  Note
6771 	 * that the original use was to identify the queue and function
6772 	 * to call on the drain.  In the new syncq, we have the context
6773 	 * of the queue that we are draining, so call it's putproc and
6774 	 * don't rely on the saved values.  But for debug this is still
6775 	 * usefull information.
6776 	 */
6777 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6778 	mp->b_queue = q;
6779 	mp->b_next = NULL;
6780 #endif
6781 	ASSERT(q->q_syncq == sq);
6782 	/*
6783 	 * Enqueue the message on the list.
6784 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6785 	 * protect it.  So its ok to acquire SQLOCK after SQPUT_MP().
6786 	 */
6787 	SQPUT_MP(q, mp);
6788 	mutex_enter(SQLOCK(sq));
6789 
6790 	/*
6791 	 * And queue on syncq for scheduling, if not already queued.
6792 	 * Note that we need the SQLOCK for this, and for testing flags
6793 	 * at the end to see if we will drain.  So grab it now, and
6794 	 * release it before we call qdrain_syncq or return.
6795 	 */
6796 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6797 		q->q_spri = curthread->t_pri;
6798 		SQPUT_Q(sq, q);
6799 	}
6800 #ifdef DEBUG
6801 	else {
6802 		/*
6803 		 * All of these conditions MUST be true!
6804 		 */
6805 		ASSERT(sq->sq_tail != NULL);
6806 		if (sq->sq_tail == sq->sq_head) {
6807 			ASSERT((q->q_sqprev == NULL) &&
6808 			    (q->q_sqnext == NULL));
6809 		} else {
6810 			ASSERT((q->q_sqprev != NULL) ||
6811 			    (q->q_sqnext != NULL));
6812 		}
6813 		ASSERT(sq->sq_flags & SQ_QUEUED);
6814 		ASSERT(q->q_syncqmsgs != 0);
6815 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6816 	}
6817 #endif
6818 	mutex_exit(QLOCK(q));
6819 	/*
6820 	 * SQLOCK is still held, so sq_count can be safely decremented.
6821 	 */
6822 	sq->sq_count--;
6823 
6824 	putnext_tail(sq, q, 0);
6825 	/* Should not reference sq or q after this point. */
6826 }
6827 
6828 /*  End of qfill_syncq  */
6829 
6830 /*
6831  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6832  * that would be put into qp by drain_syncq.
6833  * Used when deleting the syncq (qp == NULL) or when detaching
6834  * a queue (qp != NULL).
6835  * Return non-zero if one or more messages were freed.
6836  *
6837  * no need to grab sq_putlocks here. See comment in strsubr.h that explains when
6838  * sq_putlocks are used.
6839  *
6840  * NOTE: This function assumes that it is called from the close() context and
6841  * that all the queues in the syncq are going aay. For this reason it doesn't
6842  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6843  * currently valid, but it is useful to rethink this function to behave properly
6844  * in other cases.
6845  */
6846 int
6847 flush_syncq(syncq_t *sq, queue_t *qp)
6848 {
6849 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6850 	queue_t		*q;
6851 	int		ret = 0;
6852 
6853 	mutex_enter(SQLOCK(sq));
6854 
6855 	/*
6856 	 * Before we leave, we need to make sure there are no
6857 	 * events listed for this queue.  All events for this queue
6858 	 * will just be freed.
6859 	 */
6860 	if (qp != NULL && sq->sq_evhead != NULL) {
6861 		ASSERT(sq->sq_flags & SQ_EVENTS);
6862 
6863 		mp_prev = NULL;
6864 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6865 			mp_next = bp->b_next;
6866 			if (bp->b_queue == qp) {
6867 				/* Delete this message */
6868 				if (mp_prev != NULL) {
6869 					mp_prev->b_next = mp_next;
6870 					/*
6871 					 * Update sq_evtail if the last element
6872 					 * is removed.
6873 					 */
6874 					if (bp == sq->sq_evtail) {
6875 						ASSERT(mp_next == NULL);
6876 						sq->sq_evtail = mp_prev;
6877 					}
6878 				} else
6879 					sq->sq_evhead = mp_next;
6880 				if (sq->sq_evhead == NULL)
6881 					sq->sq_flags &= ~SQ_EVENTS;
6882 				bp->b_prev = bp->b_next = NULL;
6883 				freemsg(bp);
6884 				ret++;
6885 			} else {
6886 				mp_prev = bp;
6887 			}
6888 		}
6889 	}
6890 
6891 	/*
6892 	 * Walk sq_head and:
6893 	 *	- match qp if qp is set, remove it's messages
6894 	 *	- all if qp is not set
6895 	 */
6896 	q = sq->sq_head;
6897 	while (q != NULL) {
6898 		ASSERT(q->q_syncq == sq);
6899 		if ((qp == NULL) || (qp == q)) {
6900 			/*
6901 			 * Yank the messages as a list off the queue
6902 			 */
6903 			mp_head = q->q_sqhead;
6904 			/*
6905 			 * We do not have QLOCK(q) here (which is safe due to
6906 			 * assumptions mentioned above). To obtain the lock we
6907 			 * need to release SQLOCK which may allow lots of things
6908 			 * to change upon us. This place requires more analysis.
6909 			 */
6910 			q->q_sqhead = q->q_sqtail = NULL;
6911 			ASSERT(mp_head->b_queue &&
6912 			    mp_head->b_queue->q_syncq == sq);
6913 
6914 			/*
6915 			 * Free each of the messages.
6916 			 */
6917 			for (bp = mp_head; bp != NULL; bp = mp_next) {
6918 				mp_next = bp->b_next;
6919 				bp->b_prev = bp->b_next = NULL;
6920 				freemsg(bp);
6921 				ret++;
6922 			}
6923 			/*
6924 			 * Now remove the queue from the syncq.
6925 			 */
6926 			ASSERT(q->q_sqflags & Q_SQQUEUED);
6927 			SQRM_Q(sq, q);
6928 			q->q_spri = 0;
6929 			q->q_syncqmsgs = 0;
6930 
6931 			/*
6932 			 * If qp was specified, we are done with it and are
6933 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
6934 			 * waiters while we still have the SQLOCK.
6935 			 */
6936 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
6937 				sq->sq_flags &= ~SQ_WANTWAKEUP;
6938 				cv_broadcast(&sq->sq_wait);
6939 			}
6940 			/* Drop SQLOCK across clr_qfull */
6941 			mutex_exit(SQLOCK(sq));
6942 
6943 			/*
6944 			 * We avoid doing the test that drain_syncq does and
6945 			 * unconditionally clear qfull for every flushed
6946 			 * message. Since flush_syncq is only called during
6947 			 * close this should not be a problem.
6948 			 */
6949 			clr_qfull(q);
6950 			if (qp != NULL) {
6951 				return (ret);
6952 			} else {
6953 				mutex_enter(SQLOCK(sq));
6954 				/*
6955 				 * The head was removed by SQRM_Q above.
6956 				 * reread the new head and flush it.
6957 				 */
6958 				q = sq->sq_head;
6959 			}
6960 		} else {
6961 			q = q->q_sqnext;
6962 		}
6963 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
6964 	}
6965 
6966 	if (sq->sq_flags & SQ_WANTWAKEUP) {
6967 		sq->sq_flags &= ~SQ_WANTWAKEUP;
6968 		cv_broadcast(&sq->sq_wait);
6969 	}
6970 
6971 	mutex_exit(SQLOCK(sq));
6972 	return (ret);
6973 }
6974 
6975 /*
6976  * Propagate all messages from a syncq to the next syncq that are associated
6977  * with the specified queue. If the queue is attached to a driver or if the
6978  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
6979  *
6980  * Assumes that the stream is strlock()'ed. We don't come here if there
6981  * are no messages to propagate.
6982  *
6983  * NOTE : If the queue is attached to a driver, all the messages are freed
6984  * as there is no point in propagating the messages from the driver syncq
6985  * to the closing stream head which will in turn get freed later.
6986  */
6987 static int
6988 propagate_syncq(queue_t *qp)
6989 {
6990 	mblk_t		*bp, *head, *tail, *prev, *next;
6991 	syncq_t 	*sq;
6992 	queue_t		*nqp;
6993 	syncq_t		*nsq;
6994 	boolean_t	isdriver;
6995 	int 		moved = 0;
6996 	uint16_t	flags;
6997 	pri_t		priority = curthread->t_pri;
6998 #ifdef DEBUG
6999 	void		(*func)();
7000 #endif
7001 
7002 	sq = qp->q_syncq;
7003 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7004 	/* debug macro */
7005 	SQ_PUTLOCKS_HELD(sq);
7006 	/*
7007 	 * As entersq() does not increment the sq_count for
7008 	 * the write side, check sq_count for non-QPERQ
7009 	 * perimeters alone.
7010 	 */
7011 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7012 
7013 	/*
7014 	 * propagate_syncq() can be called because of either messages on the
7015 	 * queue syncq or because on events on the queue syncq. Do actual
7016 	 * message propagations if there are any messages.
7017 	 */
7018 	if (qp->q_syncqmsgs) {
7019 		isdriver = (qp->q_flag & QISDRV);
7020 
7021 		if (!isdriver) {
7022 			nqp = qp->q_next;
7023 			nsq = nqp->q_syncq;
7024 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7025 			/* debug macro */
7026 			SQ_PUTLOCKS_HELD(nsq);
7027 #ifdef DEBUG
7028 			func = (void (*)())nqp->q_qinfo->qi_putp;
7029 #endif
7030 		}
7031 
7032 		SQRM_Q(sq, qp);
7033 		priority = MAX(qp->q_spri, priority);
7034 		qp->q_spri = 0;
7035 		head = qp->q_sqhead;
7036 		tail = qp->q_sqtail;
7037 		qp->q_sqhead = qp->q_sqtail = NULL;
7038 		qp->q_syncqmsgs = 0;
7039 
7040 		/*
7041 		 * Walk the list of messages, and free them if this is a driver,
7042 		 * otherwise reset the b_prev and b_queue value to the new putp.
7043 		 * Afterward, we will just add the head to the end of the next
7044 		 * syncq, and point the tail to the end of this one.
7045 		 */
7046 
7047 		for (bp = head; bp != NULL; bp = next) {
7048 			next = bp->b_next;
7049 			if (isdriver) {
7050 				bp->b_prev = bp->b_next = NULL;
7051 				freemsg(bp);
7052 				continue;
7053 			}
7054 			/* Change the q values for this message */
7055 			bp->b_queue = nqp;
7056 #ifdef DEBUG
7057 			bp->b_prev = (mblk_t *)func;
7058 #endif
7059 			moved++;
7060 		}
7061 		/*
7062 		 * Attach list of messages to the end of the new queue (if there
7063 		 * is a list of messages).
7064 		 */
7065 
7066 		if (!isdriver && head != NULL) {
7067 			ASSERT(tail != NULL);
7068 			if (nqp->q_sqhead == NULL) {
7069 				nqp->q_sqhead = head;
7070 			} else {
7071 				ASSERT(nqp->q_sqtail != NULL);
7072 				nqp->q_sqtail->b_next = head;
7073 			}
7074 			nqp->q_sqtail = tail;
7075 			/*
7076 			 * When messages are moved from high priority queue to
7077 			 * another queue, the destination queue priority is
7078 			 * upgraded.
7079 			 */
7080 
7081 			if (priority > nqp->q_spri)
7082 				nqp->q_spri = priority;
7083 
7084 			SQPUT_Q(nsq, nqp);
7085 
7086 			nqp->q_syncqmsgs += moved;
7087 			ASSERT(nqp->q_syncqmsgs != 0);
7088 		}
7089 	}
7090 
7091 	/*
7092 	 * Before we leave, we need to make sure there are no
7093 	 * events listed for this queue.  All events for this queue
7094 	 * will just be freed.
7095 	 */
7096 	if (sq->sq_evhead != NULL) {
7097 		ASSERT(sq->sq_flags & SQ_EVENTS);
7098 		prev = NULL;
7099 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7100 			next = bp->b_next;
7101 			if (bp->b_queue == qp) {
7102 				/* Delete this message */
7103 				if (prev != NULL) {
7104 					prev->b_next = next;
7105 					/*
7106 					 * Update sq_evtail if the last element
7107 					 * is removed.
7108 					 */
7109 					if (bp == sq->sq_evtail) {
7110 						ASSERT(next == NULL);
7111 						sq->sq_evtail = prev;
7112 					}
7113 				} else
7114 					sq->sq_evhead = next;
7115 				if (sq->sq_evhead == NULL)
7116 					sq->sq_flags &= ~SQ_EVENTS;
7117 				bp->b_prev = bp->b_next = NULL;
7118 				freemsg(bp);
7119 			} else {
7120 				prev = bp;
7121 			}
7122 		}
7123 	}
7124 
7125 	flags = sq->sq_flags;
7126 
7127 	/* Wake up any waiter before leaving. */
7128 	if (flags & SQ_WANTWAKEUP) {
7129 		flags &= ~SQ_WANTWAKEUP;
7130 		cv_broadcast(&sq->sq_wait);
7131 	}
7132 	sq->sq_flags = flags;
7133 
7134 	return (moved);
7135 }
7136 
7137 /*
7138  * Try and upgrade to exclusive access at the inner perimeter. If this can
7139  * not be done without blocking then request will be queued on the syncq
7140  * and drain_syncq will run it later.
7141  *
7142  * This routine can only be called from put or service procedures plus
7143  * asynchronous callback routines that have properly entered to
7144  * queue (with entersq.) Thus qwriter_inner assumes the caller has one claim
7145  * on the syncq associated with q.
7146  */
7147 void
7148 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7149 {
7150 	syncq_t	*sq = q->q_syncq;
7151 	uint16_t count;
7152 
7153 	mutex_enter(SQLOCK(sq));
7154 	count = sq->sq_count;
7155 	SQ_PUTLOCKS_ENTER(sq);
7156 	SUM_SQ_PUTCOUNTS(sq, count);
7157 	ASSERT(count >= 1);
7158 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7159 
7160 	if (count == 1) {
7161 		/*
7162 		 * Can upgrade. This case also handles nested qwriter calls
7163 		 * (when the qwriter callback function calls qwriter). In that
7164 		 * case SQ_EXCL is already set.
7165 		 */
7166 		sq->sq_flags |= SQ_EXCL;
7167 		SQ_PUTLOCKS_EXIT(sq);
7168 		mutex_exit(SQLOCK(sq));
7169 		(*func)(q, mp);
7170 		/*
7171 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7172 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7173 		 * until putnext, leavesq, or drain_syncq drops it.
7174 		 * That way we handle nested qwriter(INNER) without dropping
7175 		 * SQ_EXCL until the outermost qwriter callback routine is
7176 		 * done.
7177 		 */
7178 		return;
7179 	}
7180 	SQ_PUTLOCKS_EXIT(sq);
7181 	sqfill_events(sq, q, mp, func);
7182 }
7183 
7184 /*
7185  * Synchronous callback support functions
7186  */
7187 
7188 /*
7189  * Allocate a callback parameter structure.
7190  * Assumes that caller initializes the flags and the id.
7191  * Acquires SQLOCK(sq) if non-NULL is returned.
7192  */
7193 callbparams_t *
7194 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7195 {
7196 	callbparams_t *cbp;
7197 	size_t size = sizeof (callbparams_t);
7198 
7199 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7200 
7201 	/*
7202 	 * Only try tryhard allocation if the caller is ready to panic.
7203 	 * Otherwise just fail.
7204 	 */
7205 	if (cbp == NULL) {
7206 		if (kmflags & KM_PANIC)
7207 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7208 			    &size, kmflags);
7209 		else
7210 			return (NULL);
7211 	}
7212 
7213 	ASSERT(size >= sizeof (callbparams_t));
7214 	cbp->cbp_size = size;
7215 	cbp->cbp_sq = sq;
7216 	cbp->cbp_func = func;
7217 	cbp->cbp_arg = arg;
7218 	mutex_enter(SQLOCK(sq));
7219 	cbp->cbp_next = sq->sq_callbpend;
7220 	sq->sq_callbpend = cbp;
7221 	return (cbp);
7222 }
7223 
7224 void
7225 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7226 {
7227 	callbparams_t **pp, *p;
7228 
7229 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7230 
7231 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7232 		if (p == cbp) {
7233 			*pp = p->cbp_next;
7234 			kmem_free(p, p->cbp_size);
7235 			return;
7236 		}
7237 	}
7238 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7239 	    "callbparams_free: not found\n"));
7240 }
7241 
7242 void
7243 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7244 {
7245 	callbparams_t **pp, *p;
7246 
7247 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7248 
7249 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7250 		if (p->cbp_id == id && p->cbp_flags == flag) {
7251 			*pp = p->cbp_next;
7252 			kmem_free(p, p->cbp_size);
7253 			return;
7254 		}
7255 	}
7256 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7257 	    "callbparams_free_id: not found\n"));
7258 }
7259 
7260 /*
7261  * Callback wrapper function used by once-only callbacks that can be
7262  * cancelled (qtimeout and qbufcall)
7263  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7264  * cancelled by the qun* functions.
7265  */
7266 void
7267 qcallbwrapper(void *arg)
7268 {
7269 	callbparams_t *cbp = arg;
7270 	syncq_t	*sq;
7271 	uint16_t count = 0;
7272 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7273 	uint16_t type;
7274 
7275 	sq = cbp->cbp_sq;
7276 	mutex_enter(SQLOCK(sq));
7277 	type = sq->sq_type;
7278 	if (!(type & SQ_CICB)) {
7279 		count = sq->sq_count;
7280 		SQ_PUTLOCKS_ENTER(sq);
7281 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7282 		SUM_SQ_PUTCOUNTS(sq, count);
7283 		sq->sq_needexcl++;
7284 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7285 		waitflags |= SQ_MESSAGES;
7286 	}
7287 	/* Can not handle exlusive entry at outer perimeter */
7288 	ASSERT(type & SQ_COCB);
7289 
7290 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7291 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7292 		    (sq->sq_cancelid == cbp->cbp_id)) {
7293 			/* timeout has been cancelled */
7294 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7295 			callbparams_free(sq, cbp);
7296 			if (!(type & SQ_CICB)) {
7297 				ASSERT(sq->sq_needexcl > 0);
7298 				sq->sq_needexcl--;
7299 				if (sq->sq_needexcl == 0) {
7300 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7301 				}
7302 				SQ_PUTLOCKS_EXIT(sq);
7303 			}
7304 			mutex_exit(SQLOCK(sq));
7305 			return;
7306 		}
7307 		sq->sq_flags |= SQ_WANTWAKEUP;
7308 		if (!(type & SQ_CICB)) {
7309 			SQ_PUTLOCKS_EXIT(sq);
7310 		}
7311 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7312 		if (!(type & SQ_CICB)) {
7313 			count = sq->sq_count;
7314 			SQ_PUTLOCKS_ENTER(sq);
7315 			SUM_SQ_PUTCOUNTS(sq, count);
7316 		}
7317 	}
7318 
7319 	sq->sq_count++;
7320 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7321 	if (!(type & SQ_CICB)) {
7322 		ASSERT(count == 0);
7323 		sq->sq_flags |= SQ_EXCL;
7324 		ASSERT(sq->sq_needexcl > 0);
7325 		sq->sq_needexcl--;
7326 		if (sq->sq_needexcl == 0) {
7327 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7328 		}
7329 		SQ_PUTLOCKS_EXIT(sq);
7330 	}
7331 
7332 	mutex_exit(SQLOCK(sq));
7333 
7334 	cbp->cbp_func(cbp->cbp_arg);
7335 
7336 	/*
7337 	 * We drop the lock only for leavesq to re-acquire it.
7338 	 * Possible optimization is inline of leavesq.
7339 	 */
7340 	mutex_enter(SQLOCK(sq));
7341 	callbparams_free(sq, cbp);
7342 	mutex_exit(SQLOCK(sq));
7343 	leavesq(sq, SQ_CALLBACK);
7344 }
7345 
7346 /*
7347  * no need to grab sq_putlocks here. See comment in strsubr.h that
7348  * explains when sq_putlocks are used.
7349  *
7350  * sq_count (or one of the sq_putcounts) has already been
7351  * decremented by the caller, and if SQ_QUEUED, we need to call
7352  * drain_syncq (the global syncq drain).
7353  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7354  * one of two states, non-CIPUT perimiter, and we need to clear
7355  * it, or we went exclusive in the put procedure.  In any case,
7356  * we want to clear the bit now, and it is probably easier to do
7357  * this at the beginning of this function (remember, we hold
7358  * the SQLOCK).  Lastly, if there are other messages queued
7359  * on the syncq (and not for our destination), enable the syncq
7360  * for background work.
7361  */
7362 
7363 /* ARGSUSED */
7364 void
7365 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7366 {
7367 	uint16_t	flags = sq->sq_flags;
7368 
7369 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7370 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7371 
7372 	/* Clear SQ_EXCL if set in passflags */
7373 	if (passflags & SQ_EXCL) {
7374 		flags &= ~SQ_EXCL;
7375 	}
7376 	if (flags & SQ_WANTWAKEUP) {
7377 		flags &= ~SQ_WANTWAKEUP;
7378 		cv_broadcast(&sq->sq_wait);
7379 	}
7380 	if (flags & SQ_WANTEXWAKEUP) {
7381 		flags &= ~SQ_WANTEXWAKEUP;
7382 		cv_broadcast(&sq->sq_exitwait);
7383 	}
7384 	sq->sq_flags = flags;
7385 
7386 	/*
7387 	 * We have cleared SQ_EXCL if we were asked to, and started
7388 	 * the wakeup process for waiters.  If there are no writers
7389 	 * then we need to drain the syncq if we were told to, or
7390 	 * enable the background thread to do it.
7391 	 */
7392 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7393 		if ((passflags & SQ_QUEUED) ||
7394 		    (sq->sq_svcflags & SQ_DISABLED)) {
7395 			/* drain_syncq will take care of events in the list */
7396 			drain_syncq(sq);
7397 			return;
7398 		} else if (flags & SQ_QUEUED) {
7399 			sqenable(sq);
7400 		}
7401 	}
7402 	/* Drop the SQLOCK on exit */
7403 	mutex_exit(SQLOCK(sq));
7404 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7405 		"putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7406 }
7407 
7408 void
7409 set_qend(queue_t *q)
7410 {
7411 	mutex_enter(QLOCK(q));
7412 	if (!O_SAMESTR(q))
7413 		q->q_flag |= QEND;
7414 	else
7415 		q->q_flag &= ~QEND;
7416 	mutex_exit(QLOCK(q));
7417 	q = _OTHERQ(q);
7418 	mutex_enter(QLOCK(q));
7419 	if (!O_SAMESTR(q))
7420 		q->q_flag |= QEND;
7421 	else
7422 		q->q_flag &= ~QEND;
7423 	mutex_exit(QLOCK(q));
7424 }
7425 
7426 
7427 void
7428 clr_qfull(queue_t *q)
7429 {
7430 	queue_t	*oq = q;
7431 
7432 	q = q->q_nfsrv;
7433 	/* Fast check if there is any work to do before getting the lock. */
7434 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7435 		return;
7436 	}
7437 
7438 	/*
7439 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7440 	 * for QFULL being set.
7441 	 */
7442 	mutex_enter(QLOCK(q));
7443 	/*
7444 	 * If both q_count and q_mblkcnt are less than the hiwat mark
7445 	 */
7446 	if ((q->q_count < q->q_hiwat) && (q->q_mblkcnt < q->q_hiwat)) {
7447 		q->q_flag &= ~QFULL;
7448 		/*
7449 		 * A little more confusing, how about this way:
7450 		 * if someone wants to write,
7451 		 * AND
7452 		 *    both counts are less than the lowat mark
7453 		 *    OR
7454 		 *    the lowat mark is zero
7455 		 * THEN
7456 		 * backenable
7457 		 */
7458 		if ((q->q_flag & QWANTW) &&
7459 		    (((q->q_count < q->q_lowat) &&
7460 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7461 			q->q_flag &= ~QWANTW;
7462 			mutex_exit(QLOCK(q));
7463 			backenable(oq, 0);
7464 		} else
7465 			mutex_exit(QLOCK(q));
7466 	} else
7467 		mutex_exit(QLOCK(q));
7468 }
7469 
7470 /*
7471  * Set the forward service procedure pointer.
7472  *
7473  * Called at insert-time to cache a queue's next forward service procedure in
7474  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7475  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7476  * inserted does not have a service procedure, then q_nfsrv points to the next
7477  * queue forward that has a service procedure.  If the queue is at the logical
7478  * end of the stream (driver for write side, stream head for the read side)
7479  * and does not have a service procedure, then q_nfsrv also points to itself.
7480  */
7481 void
7482 set_nfsrv_ptr(
7483 	queue_t  *rnew,		/* read queue pointer to new module */
7484 	queue_t  *wnew,		/* write queue pointer to new module */
7485 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7486 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7487 {
7488 	queue_t *qp;
7489 
7490 	if (prev_wq->q_next == NULL) {
7491 		/*
7492 		 * Insert the driver, initialize the driver and stream head.
7493 		 * In this case, prev_rq/prev_wq should be the stream head.
7494 		 * _I_INSERT does not allow inserting a driver.  Make sure
7495 		 * that it is not an insertion.
7496 		 */
7497 		ASSERT(!(rnew->q_flag & _QINSERTING));
7498 		wnew->q_nfsrv = wnew;
7499 		if (rnew->q_qinfo->qi_srvp)
7500 			rnew->q_nfsrv = rnew;
7501 		else
7502 			rnew->q_nfsrv = prev_rq;
7503 		prev_rq->q_nfsrv = prev_rq;
7504 		prev_wq->q_nfsrv = prev_wq;
7505 	} else {
7506 		/*
7507 		 * set up read side q_nfsrv pointer.  This MUST be done
7508 		 * before setting the write side, because the setting of
7509 		 * the write side for a fifo may depend on it.
7510 		 *
7511 		 * Suppose we have a fifo that only has pipemod pushed.
7512 		 * pipemod has no read or write service procedures, so
7513 		 * nfsrv for both pipemod queues points to prev_rq (the
7514 		 * stream read head).  Now push bufmod (which has only a
7515 		 * read service procedure).  Doing the write side first,
7516 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7517 		 * is WRONG; the next queue forward from wnew with a
7518 		 * service procedure will be rnew, not the stream read head.
7519 		 * Since the downstream queue (which in the case of a fifo
7520 		 * is the read queue rnew) can affect upstream queues, it
7521 		 * needs to be done first.  Setting up the read side first
7522 		 * sets nfsrv for both pipemod queues to rnew and then
7523 		 * when the write side is set up, wnew-q_nfsrv will also
7524 		 * point to rnew.
7525 		 */
7526 		if (rnew->q_qinfo->qi_srvp) {
7527 			/*
7528 			 * use _OTHERQ() because, if this is a pipe, next
7529 			 * module may have been pushed from other end and
7530 			 * q_next could be a read queue.
7531 			 */
7532 			qp = _OTHERQ(prev_wq->q_next);
7533 			while (qp && qp->q_nfsrv != qp) {
7534 				qp->q_nfsrv = rnew;
7535 				qp = backq(qp);
7536 			}
7537 			rnew->q_nfsrv = rnew;
7538 		} else
7539 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7540 
7541 		/* set up write side q_nfsrv pointer */
7542 		if (wnew->q_qinfo->qi_srvp) {
7543 			wnew->q_nfsrv = wnew;
7544 
7545 			/*
7546 			 * For insertion, need to update nfsrv of the modules
7547 			 * above which do not have a service routine.
7548 			 */
7549 			if (rnew->q_flag & _QINSERTING) {
7550 				for (qp = prev_wq;
7551 				    qp != NULL && qp->q_nfsrv != qp;
7552 				    qp = backq(qp)) {
7553 					qp->q_nfsrv = wnew->q_nfsrv;
7554 				}
7555 			}
7556 		} else {
7557 			if (prev_wq->q_next == prev_rq)
7558 				/*
7559 				 * Since prev_wq/prev_rq are the middle of a
7560 				 * fifo, wnew/rnew will also be the middle of
7561 				 * a fifo and wnew's nfsrv is same as rnew's.
7562 				 */
7563 				wnew->q_nfsrv = rnew->q_nfsrv;
7564 			else
7565 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7566 		}
7567 	}
7568 }
7569 
7570 /*
7571  * Reset the forward service procedure pointer; called at remove-time.
7572  */
7573 void
7574 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7575 {
7576 	queue_t *tmp_qp;
7577 
7578 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7579 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7580 		for (tmp_qp = backq(wqp);
7581 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7582 		    tmp_qp = backq(tmp_qp)) {
7583 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7584 		}
7585 	}
7586 
7587 	/* reset the read side q_nfsrv pointer */
7588 	if (rqp->q_qinfo->qi_srvp) {
7589 		if (wqp->q_next) {	/* non-driver case */
7590 			tmp_qp = _OTHERQ(wqp->q_next);
7591 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7592 				/* Note that rqp->q_next cannot be NULL */
7593 				ASSERT(rqp->q_next != NULL);
7594 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7595 				tmp_qp = backq(tmp_qp);
7596 			}
7597 		}
7598 	}
7599 }
7600 
7601 /*
7602  * This routine should be called after all stream geometry changes to update
7603  * the stream head cached struio() rd/wr queue pointers. Note must be called
7604  * with the streamlock()ed.
7605  *
7606  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7607  *	 an explicit synchronous barrier module queue. That is, a queue that
7608  *	 has specified a struio() type.
7609  */
7610 static void
7611 strsetuio(stdata_t *stp)
7612 {
7613 	queue_t *wrq;
7614 
7615 	if (stp->sd_flag & STPLEX) {
7616 		/*
7617 		 * Not stremahead, but a mux, so no Synchronous STREAMS.
7618 		 */
7619 		stp->sd_struiowrq = NULL;
7620 		stp->sd_struiordq = NULL;
7621 		return;
7622 	}
7623 	/*
7624 	 * Scan the write queue(s) while synchronous
7625 	 * until we find a qinfo uio type specified.
7626 	 */
7627 	wrq = stp->sd_wrq->q_next;
7628 	while (wrq) {
7629 		if (wrq->q_struiot == STRUIOT_NONE) {
7630 			wrq = 0;
7631 			break;
7632 		}
7633 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7634 			break;
7635 		if (! _SAMESTR(wrq)) {
7636 			wrq = 0;
7637 			break;
7638 		}
7639 		wrq = wrq->q_next;
7640 	}
7641 	stp->sd_struiowrq = wrq;
7642 	/*
7643 	 * Scan the read queue(s) while synchronous
7644 	 * until we find a qinfo uio type specified.
7645 	 */
7646 	wrq = stp->sd_wrq->q_next;
7647 	while (wrq) {
7648 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7649 			wrq = 0;
7650 			break;
7651 		}
7652 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7653 			break;
7654 		if (! _SAMESTR(wrq)) {
7655 			wrq = 0;
7656 			break;
7657 		}
7658 		wrq = wrq->q_next;
7659 	}
7660 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7661 }
7662 
7663 /*
7664  * pass_wput, unblocks the passthru queues, so that
7665  * messages can arrive at muxs lower read queue, before
7666  * I_LINK/I_UNLINK is acked/nacked.
7667  */
7668 static void
7669 pass_wput(queue_t *q, mblk_t *mp)
7670 {
7671 	syncq_t *sq;
7672 
7673 	sq = _RD(q)->q_syncq;
7674 	if (sq->sq_flags & SQ_BLOCKED)
7675 		unblocksq(sq, SQ_BLOCKED, 0);
7676 	putnext(q, mp);
7677 }
7678 
7679 /*
7680  * Set up queues for the link/unlink.
7681  * Create a new queue and block it and then insert it
7682  * below the stream head on the lower stream.
7683  * This prevents any messages from arriving during the setq
7684  * as well as while the mux is processing the LINK/I_UNLINK.
7685  * The blocked passq is unblocked once the LINK/I_UNLINK has
7686  * been acked or nacked or if a message is generated and sent
7687  * down muxs write put procedure.
7688  * see pass_wput().
7689  *
7690  * After the new queue is inserted, all messages coming from below are
7691  * blocked. The call to strlock will ensure that all activity in the stream head
7692  * read queue syncq is stopped (sq_count drops to zero).
7693  */
7694 static queue_t *
7695 link_addpassthru(stdata_t *stpdown)
7696 {
7697 	queue_t *passq;
7698 	sqlist_t sqlist;
7699 
7700 	passq = allocq();
7701 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7702 	/* setq might sleep in allocator - avoid holding locks. */
7703 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7704 	    SQ_CI|SQ_CO, B_FALSE);
7705 	claimq(passq);
7706 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7707 	insertq(STREAM(passq), passq);
7708 
7709 	/*
7710 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7711 	 * since we are going to change q_ptr in the stream head.  Note that
7712 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7713 	 */
7714 	sqlist.sqlist_head = NULL;
7715 	sqlist.sqlist_index = 0;
7716 	sqlist.sqlist_size = sizeof (sqlist_t);
7717 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7718 	strlock(stpdown, &sqlist);
7719 	strunlock(stpdown, &sqlist);
7720 
7721 	releaseq(passq);
7722 	return (passq);
7723 }
7724 
7725 /*
7726  * Let messages flow up into the mux by removing
7727  * the passq.
7728  */
7729 static void
7730 link_rempassthru(queue_t *passq)
7731 {
7732 	claimq(passq);
7733 	removeq(passq);
7734 	releaseq(passq);
7735 	freeq(passq);
7736 }
7737 
7738 /*
7739  * Wait for the condition variable pointed to by `cvp' to be signaled,
7740  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7741  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7742  * then the wait will be non-interruptible.
7743  *
7744  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7745  */
7746 clock_t
7747 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7748 {
7749 	clock_t ret, now, tick;
7750 
7751 	if (tim < 0) {
7752 		if (nosigs) {
7753 			cv_wait(cvp, mp);
7754 			ret = 1;
7755 		} else {
7756 			ret = cv_wait_sig(cvp, mp);
7757 		}
7758 	} else if (tim > 0) {
7759 		/*
7760 		 * convert milliseconds to clock ticks
7761 		 */
7762 		tick = MSEC_TO_TICK_ROUNDUP(tim);
7763 		time_to_wait(&now, tick);
7764 		if (nosigs) {
7765 			ret = cv_timedwait(cvp, mp, now);
7766 		} else {
7767 			ret = cv_timedwait_sig(cvp, mp, now);
7768 		}
7769 	} else {
7770 		ret = -1;
7771 	}
7772 	return (ret);
7773 }
7774 
7775 /*
7776  * Wait until the stream head can determine if it is at the mark but
7777  * don't wait forever to prevent a race condition between the "mark" state
7778  * in the stream head and any mark state in the caller/user of this routine.
7779  *
7780  * This is used by sockets and for a socket it would be incorrect
7781  * to return a failure for SIOCATMARK when there is no data in the receive
7782  * queue and the marked urgent data is traveling up the stream.
7783  *
7784  * This routine waits until the mark is known by waiting for one of these
7785  * three events:
7786  *	The stream head read queue becoming non-empty (including an EOF)
7787  *	The STRATMARK flag being set. (Due to a MSGMARKNEXT message.)
7788  *	The STRNOTATMARK flag being set (which indicates that the transport
7789  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7790  *	the mark).
7791  *
7792  * The routine returns 1 if the stream is at the mark; 0 if it can
7793  * be determined that the stream is not at the mark.
7794  * If the wait times out and it can't determine
7795  * whether or not the stream might be at the mark the routine will return -1.
7796  *
7797  * Note: This routine should only be used when a mark is pending i.e.,
7798  * in the socket case the SIGURG has been posted.
7799  * Note2: This can not wakeup just because synchronous streams indicate
7800  * that data is available since it is not possible to use the synchronous
7801  * streams interfaces to determine the b_flag value for the data queued below
7802  * the stream head.
7803  */
7804 int
7805 strwaitmark(vnode_t *vp)
7806 {
7807 	struct stdata *stp = vp->v_stream;
7808 	queue_t *rq = _RD(stp->sd_wrq);
7809 	int mark;
7810 
7811 	mutex_enter(&stp->sd_lock);
7812 	while (rq->q_first == NULL &&
7813 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7814 		stp->sd_flag |= RSLEEP;
7815 
7816 		/* Wait for 100 milliseconds for any state change. */
7817 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7818 			mutex_exit(&stp->sd_lock);
7819 			return (-1);
7820 		}
7821 	}
7822 	if (stp->sd_flag & STRATMARK)
7823 		mark = 1;
7824 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7825 		mark = 1;
7826 	else
7827 		mark = 0;
7828 
7829 	mutex_exit(&stp->sd_lock);
7830 	return (mark);
7831 }
7832 
7833 /*
7834  * Set a read side error. If persist is set change the socket error
7835  * to persistent. If errfunc is set install the function as the exported
7836  * error handler.
7837  */
7838 void
7839 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7840 {
7841 	struct stdata *stp = vp->v_stream;
7842 
7843 	mutex_enter(&stp->sd_lock);
7844 	stp->sd_rerror = error;
7845 	if (error == 0 && errfunc == NULL)
7846 		stp->sd_flag &= ~STRDERR;
7847 	else
7848 		stp->sd_flag |= STRDERR;
7849 	if (persist) {
7850 		stp->sd_flag &= ~STRDERRNONPERSIST;
7851 	} else {
7852 		stp->sd_flag |= STRDERRNONPERSIST;
7853 	}
7854 	stp->sd_rderrfunc = errfunc;
7855 	if (error != 0 || errfunc != NULL) {
7856 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7857 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7858 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7859 
7860 		mutex_exit(&stp->sd_lock);
7861 		pollwakeup(&stp->sd_pollist, POLLERR);
7862 		mutex_enter(&stp->sd_lock);
7863 
7864 		if (stp->sd_sigflags & S_ERROR)
7865 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
7866 	}
7867 	mutex_exit(&stp->sd_lock);
7868 }
7869 
7870 /*
7871  * Set a write side error. If persist is set change the socket error
7872  * to persistent.
7873  */
7874 void
7875 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7876 {
7877 	struct stdata *stp = vp->v_stream;
7878 
7879 	mutex_enter(&stp->sd_lock);
7880 	stp->sd_werror = error;
7881 	if (error == 0 && errfunc == NULL)
7882 		stp->sd_flag &= ~STWRERR;
7883 	else
7884 		stp->sd_flag |= STWRERR;
7885 	if (persist) {
7886 		stp->sd_flag &= ~STWRERRNONPERSIST;
7887 	} else {
7888 		stp->sd_flag |= STWRERRNONPERSIST;
7889 	}
7890 	stp->sd_wrerrfunc = errfunc;
7891 	if (error != 0 || errfunc != NULL) {
7892 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7893 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7894 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7895 
7896 		mutex_exit(&stp->sd_lock);
7897 		pollwakeup(&stp->sd_pollist, POLLERR);
7898 		mutex_enter(&stp->sd_lock);
7899 
7900 		if (stp->sd_sigflags & S_ERROR)
7901 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
7902 	}
7903 	mutex_exit(&stp->sd_lock);
7904 }
7905 
7906 /*
7907  * Make the stream return 0 (EOF) when all data has been read.
7908  * No effect on write side.
7909  */
7910 void
7911 strseteof(vnode_t *vp, int eof)
7912 {
7913 	struct stdata *stp = vp->v_stream;
7914 
7915 	mutex_enter(&stp->sd_lock);
7916 	if (!eof) {
7917 		stp->sd_flag &= ~STREOF;
7918 		mutex_exit(&stp->sd_lock);
7919 		return;
7920 	}
7921 	stp->sd_flag |= STREOF;
7922 	if (stp->sd_flag & RSLEEP) {
7923 		stp->sd_flag &= ~RSLEEP;
7924 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
7925 	}
7926 
7927 	mutex_exit(&stp->sd_lock);
7928 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
7929 	mutex_enter(&stp->sd_lock);
7930 
7931 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
7932 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
7933 	mutex_exit(&stp->sd_lock);
7934 }
7935 
7936 void
7937 strflushrq(vnode_t *vp, int flag)
7938 {
7939 	struct stdata *stp = vp->v_stream;
7940 
7941 	mutex_enter(&stp->sd_lock);
7942 	flushq(_RD(stp->sd_wrq), flag);
7943 	mutex_exit(&stp->sd_lock);
7944 }
7945 
7946 void
7947 strsetrputhooks(vnode_t *vp, uint_t flags,
7948 		msgfunc_t protofunc, msgfunc_t miscfunc)
7949 {
7950 	struct stdata *stp = vp->v_stream;
7951 
7952 	mutex_enter(&stp->sd_lock);
7953 
7954 	if (protofunc == NULL)
7955 		stp->sd_rprotofunc = strrput_proto;
7956 	else
7957 		stp->sd_rprotofunc = protofunc;
7958 
7959 	if (miscfunc == NULL)
7960 		stp->sd_rmiscfunc = strrput_misc;
7961 	else
7962 		stp->sd_rmiscfunc = miscfunc;
7963 
7964 	if (flags & SH_CONSOL_DATA)
7965 		stp->sd_rput_opt |= SR_CONSOL_DATA;
7966 	else
7967 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
7968 
7969 	if (flags & SH_SIGALLDATA)
7970 		stp->sd_rput_opt |= SR_SIGALLDATA;
7971 	else
7972 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
7973 
7974 	if (flags & SH_IGN_ZEROLEN)
7975 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
7976 	else
7977 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
7978 
7979 	mutex_exit(&stp->sd_lock);
7980 }
7981 
7982 void
7983 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
7984 {
7985 	struct stdata *stp = vp->v_stream;
7986 
7987 	mutex_enter(&stp->sd_lock);
7988 	stp->sd_closetime = closetime;
7989 
7990 	if (flags & SH_SIGPIPE)
7991 		stp->sd_wput_opt |= SW_SIGPIPE;
7992 	else
7993 		stp->sd_wput_opt &= ~SW_SIGPIPE;
7994 	if (flags & SH_RECHECK_ERR)
7995 		stp->sd_wput_opt |= SW_RECHECK_ERR;
7996 	else
7997 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
7998 
7999 	mutex_exit(&stp->sd_lock);
8000 }
8001 
8002 void
8003 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8004 {
8005 	struct stdata *stp = vp->v_stream;
8006 
8007 	mutex_enter(&stp->sd_lock);
8008 
8009 	stp->sd_rputdatafunc = rdatafunc;
8010 	stp->sd_wputdatafunc = wdatafunc;
8011 
8012 	mutex_exit(&stp->sd_lock);
8013 }
8014 
8015 /* Used within framework when the queue is already locked */
8016 void
8017 qenable_locked(queue_t *q)
8018 {
8019 	stdata_t *stp = STREAM(q);
8020 
8021 	ASSERT(MUTEX_HELD(QLOCK(q)));
8022 
8023 	if (!q->q_qinfo->qi_srvp)
8024 		return;
8025 
8026 	/*
8027 	 * Do not place on run queue if already enabled or closing.
8028 	 */
8029 	if (q->q_flag & (QWCLOSE|QENAB))
8030 		return;
8031 
8032 	/*
8033 	 * mark queue enabled and place on run list if it is not already being
8034 	 * serviced. If it is serviced, the runservice() function will detect
8035 	 * that QENAB is set and call service procedure before clearing
8036 	 * QINSERVICE flag.
8037 	 */
8038 	q->q_flag |= QENAB;
8039 	if (q->q_flag & QINSERVICE)
8040 		return;
8041 
8042 	/* Record the time of qenable */
8043 	q->q_qtstamp = lbolt;
8044 
8045 	/*
8046 	 * Put the queue in the stp list and schedule it for background
8047 	 * processing if it is not already scheduled or if stream head does not
8048 	 * intent to process it in the foreground later by setting
8049 	 * STRS_WILLSERVICE flag.
8050 	 */
8051 	mutex_enter(&stp->sd_qlock);
8052 	/*
8053 	 * If there are already something on the list, stp flags should show
8054 	 * intention to drain it.
8055 	 */
8056 	IMPLY(STREAM_NEEDSERVICE(stp),
8057 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8058 
8059 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8060 	stp->sd_nqueues++;
8061 
8062 	/*
8063 	 * If no one will drain this stream we are the first producer and
8064 	 * need to schedule it for background thread.
8065 	 */
8066 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8067 		/*
8068 		 * No one will service this stream later, so we have to
8069 		 * schedule it now.
8070 		 */
8071 		STRSTAT(stenables);
8072 		stp->sd_svcflags |= STRS_SCHEDULED;
8073 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8074 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8075 
8076 		if (stp->sd_servid == NULL) {
8077 			/*
8078 			 * Task queue failed so fail over to the backup
8079 			 * servicing thread.
8080 			 */
8081 			STRSTAT(taskqfails);
8082 			/*
8083 			 * It is safe to clear STRS_SCHEDULED flag because it
8084 			 * was set by this thread above.
8085 			 */
8086 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8087 
8088 			/*
8089 			 * Failover scheduling is protected by service_queue
8090 			 * lock.
8091 			 */
8092 			mutex_enter(&service_queue);
8093 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8094 			ASSERT(q->q_link == NULL);
8095 			/*
8096 			 * Append the queue to qhead/qtail list.
8097 			 */
8098 			if (qhead == NULL)
8099 				qhead = q;
8100 			else
8101 				qtail->q_link = q;
8102 			qtail = q;
8103 			/*
8104 			 * Clear stp queue list.
8105 			 */
8106 			stp->sd_qhead = stp->sd_qtail = NULL;
8107 			stp->sd_nqueues = 0;
8108 			/*
8109 			 * Wakeup background queue processing thread.
8110 			 */
8111 			cv_signal(&services_to_run);
8112 			mutex_exit(&service_queue);
8113 		}
8114 	}
8115 	mutex_exit(&stp->sd_qlock);
8116 }
8117 
8118 static void
8119 queue_service(queue_t *q)
8120 {
8121 	/*
8122 	 * The queue in the list should have
8123 	 * QENAB flag set and should not have
8124 	 * QINSERVICE flag set. QINSERVICE is
8125 	 * set when the queue is dequeued and
8126 	 * qenable_locked doesn't enqueue a
8127 	 * queue with QINSERVICE set.
8128 	 */
8129 
8130 	ASSERT(!(q->q_flag & QINSERVICE));
8131 	ASSERT((q->q_flag & QENAB));
8132 	mutex_enter(QLOCK(q));
8133 	q->q_flag &= ~QENAB;
8134 	q->q_flag |= QINSERVICE;
8135 	mutex_exit(QLOCK(q));
8136 	runservice(q);
8137 }
8138 
8139 static void
8140 syncq_service(syncq_t *sq)
8141 {
8142 	STRSTAT(syncqservice);
8143 	mutex_enter(SQLOCK(sq));
8144 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8145 	ASSERT(sq->sq_servcount != 0);
8146 	ASSERT(sq->sq_next == NULL);
8147 
8148 	/* if we came here from the background thread, clear the flag */
8149 	if (sq->sq_svcflags & SQ_BGTHREAD)
8150 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8151 
8152 	/* let drain_syncq know that it's being called in the background */
8153 	sq->sq_svcflags |= SQ_SERVICE;
8154 	drain_syncq(sq);
8155 }
8156 
8157 static void
8158 qwriter_outer_service(syncq_t *outer)
8159 {
8160 	/*
8161 	 * Note that SQ_WRITER is used on the outer perimeter
8162 	 * to signal that a qwriter(OUTER) is either investigating
8163 	 * running or that it is actually running a function.
8164 	 */
8165 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8166 
8167 	/*
8168 	 * All inner syncq are empty and have SQ_WRITER set
8169 	 * to block entering the outer perimeter.
8170 	 *
8171 	 * We do not need to explicitly call write_now since
8172 	 * outer_exit does it for us.
8173 	 */
8174 	outer_exit(outer);
8175 }
8176 
8177 static void
8178 mblk_free(mblk_t *mp)
8179 {
8180 	dblk_t *dbp = mp->b_datap;
8181 	frtn_t *frp = dbp->db_frtnp;
8182 
8183 	mp->b_next = NULL;
8184 	if (dbp->db_fthdr != NULL)
8185 		str_ftfree(dbp);
8186 
8187 	ASSERT(dbp->db_fthdr == NULL);
8188 	frp->free_func(frp->free_arg);
8189 	ASSERT(dbp->db_mblk == mp);
8190 
8191 	if (dbp->db_credp != NULL) {
8192 		crfree(dbp->db_credp);
8193 		dbp->db_credp = NULL;
8194 	}
8195 	dbp->db_cpid = -1;
8196 	dbp->db_struioflag = 0;
8197 	dbp->db_struioun.cksum.flags = 0;
8198 
8199 	kmem_cache_free(dbp->db_cache, dbp);
8200 }
8201 
8202 /*
8203  * Background processing of the stream queue list.
8204  */
8205 static void
8206 stream_service(stdata_t *stp)
8207 {
8208 	queue_t *q;
8209 
8210 	mutex_enter(&stp->sd_qlock);
8211 
8212 	STR_SERVICE(stp, q);
8213 
8214 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8215 	stp->sd_servid = NULL;
8216 	cv_signal(&stp->sd_qcv);
8217 	mutex_exit(&stp->sd_qlock);
8218 }
8219 
8220 /*
8221  * Foreground processing of the stream queue list.
8222  */
8223 void
8224 stream_runservice(stdata_t *stp)
8225 {
8226 	queue_t *q;
8227 
8228 	mutex_enter(&stp->sd_qlock);
8229 	STRSTAT(rservice);
8230 	/*
8231 	 * We are going to drain this stream queue list, so qenable_locked will
8232 	 * not schedule it until we finish.
8233 	 */
8234 	stp->sd_svcflags |= STRS_WILLSERVICE;
8235 
8236 	STR_SERVICE(stp, q);
8237 
8238 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8239 	mutex_exit(&stp->sd_qlock);
8240 	/*
8241 	 * Help backup background thread to drain the qhead/qtail list.
8242 	 */
8243 	while (qhead != NULL) {
8244 		STRSTAT(qhelps);
8245 		mutex_enter(&service_queue);
8246 		DQ(q, qhead, qtail, q_link);
8247 		mutex_exit(&service_queue);
8248 		if (q != NULL)
8249 			queue_service(q);
8250 	}
8251 }
8252 
8253 void
8254 stream_willservice(stdata_t *stp)
8255 {
8256 	mutex_enter(&stp->sd_qlock);
8257 	stp->sd_svcflags |= STRS_WILLSERVICE;
8258 	mutex_exit(&stp->sd_qlock);
8259 }
8260 
8261 /*
8262  * Replace the cred currently in the mblk with a different one.
8263  */
8264 void
8265 mblk_setcred(mblk_t *mp, cred_t *cr)
8266 {
8267 	cred_t *ocr = DB_CRED(mp);
8268 
8269 	ASSERT(cr != NULL);
8270 
8271 	if (cr != ocr) {
8272 		crhold(mp->b_datap->db_credp = cr);
8273 		if (ocr != NULL)
8274 			crfree(ocr);
8275 	}
8276 }
8277 
8278 int
8279 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8280     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8281     uint32_t flags, int km_flags)
8282 {
8283 	int rc = 0;
8284 
8285 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8286 	if (mp->b_datap->db_type == M_DATA) {
8287 		/* Associate values for M_DATA type */
8288 		DB_CKSUMSTART(mp) = (intptr_t)start;
8289 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8290 		DB_CKSUMEND(mp) = (intptr_t)end;
8291 		DB_CKSUMFLAGS(mp) = flags;
8292 		DB_CKSUM16(mp) = (uint16_t)value;
8293 
8294 	} else {
8295 		pattrinfo_t pa_info;
8296 
8297 		ASSERT(mmd != NULL);
8298 
8299 		pa_info.type = PATTR_HCKSUM;
8300 		pa_info.len = sizeof (pattr_hcksum_t);
8301 
8302 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8303 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8304 
8305 			hck->hcksum_start_offset = start;
8306 			hck->hcksum_stuff_offset = stuff;
8307 			hck->hcksum_end_offset = end;
8308 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8309 			hck->hcksum_flags = flags;
8310 		} else {
8311 			rc = -1;
8312 		}
8313 	}
8314 	return (rc);
8315 }
8316 
8317 void
8318 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8319     uint32_t *start, uint32_t *stuff, uint32_t *end,
8320     uint32_t *value, uint32_t *flags)
8321 {
8322 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8323 	if (mp->b_datap->db_type == M_DATA) {
8324 		if (flags != NULL) {
8325 			*flags = DB_CKSUMFLAGS(mp);
8326 			if (*flags & HCK_PARTIALCKSUM) {
8327 				if (start != NULL)
8328 					*start = (uint32_t)DB_CKSUMSTART(mp);
8329 				if (stuff != NULL)
8330 					*stuff = (uint32_t)DB_CKSUMSTUFF(mp);
8331 				if (end != NULL)
8332 					*end = (uint32_t)DB_CKSUMEND(mp);
8333 				if (value != NULL)
8334 					*value = (uint32_t)DB_CKSUM16(mp);
8335 			} else if ((*flags & HW_LSO) && (value != NULL))
8336 				*value = (uint32_t)DB_LSOMSS(mp);
8337 		}
8338 	} else {
8339 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8340 
8341 		ASSERT(mmd != NULL);
8342 
8343 		/* get hardware checksum attribute */
8344 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8345 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8346 
8347 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8348 			if (flags != NULL)
8349 				*flags = hck->hcksum_flags;
8350 			if (start != NULL)
8351 				*start = hck->hcksum_start_offset;
8352 			if (stuff != NULL)
8353 				*stuff = hck->hcksum_stuff_offset;
8354 			if (end != NULL)
8355 				*end = hck->hcksum_end_offset;
8356 			if (value != NULL)
8357 				*value = (uint32_t)
8358 				    hck->hcksum_cksum_val.inet_cksum;
8359 		}
8360 	}
8361 }
8362 
8363 /*
8364  * Checksum buffer *bp for len bytes with psum partial checksum,
8365  * or 0 if none, and return the 16 bit partial checksum.
8366  */
8367 unsigned
8368 bcksum(uchar_t *bp, int len, unsigned int psum)
8369 {
8370 	int odd = len & 1;
8371 	extern unsigned int ip_ocsum();
8372 
8373 	if (((intptr_t)bp & 1) == 0 && !odd) {
8374 		/*
8375 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8376 		 */
8377 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8378 	}
8379 	if (((intptr_t)bp & 1) != 0) {
8380 		/*
8381 		 * Bp isn't 16 bit aligned.
8382 		 */
8383 		unsigned int tsum;
8384 
8385 #ifdef _LITTLE_ENDIAN
8386 		psum += *bp;
8387 #else
8388 		psum += *bp << 8;
8389 #endif
8390 		len--;
8391 		bp++;
8392 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8393 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8394 		if (len & 1) {
8395 			bp += len - 1;
8396 #ifdef _LITTLE_ENDIAN
8397 			psum += *bp << 8;
8398 #else
8399 			psum += *bp;
8400 #endif
8401 		}
8402 	} else {
8403 		/*
8404 		 * Bp is 16 bit aligned.
8405 		 */
8406 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8407 		if (odd) {
8408 			bp += len - 1;
8409 #ifdef _LITTLE_ENDIAN
8410 			psum += *bp;
8411 #else
8412 			psum += *bp << 8;
8413 #endif
8414 		}
8415 	}
8416 	/*
8417 	 * Normalize psum to 16 bits before returning the new partial
8418 	 * checksum. The max psum value before normalization is 0x3FDFE.
8419 	 */
8420 	return ((psum >> 16) + (psum & 0xFFFF));
8421 }
8422 
8423 boolean_t
8424 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8425 {
8426 	boolean_t rc;
8427 
8428 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8429 	if (DB_TYPE(mp) == M_DATA) {
8430 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8431 	} else {
8432 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8433 
8434 		ASSERT(mmd != NULL);
8435 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8436 	}
8437 	return (rc);
8438 }
8439 
8440 void
8441 freemsgchain(mblk_t *mp)
8442 {
8443 	mblk_t	*next;
8444 
8445 	while (mp != NULL) {
8446 		next = mp->b_next;
8447 		mp->b_next = NULL;
8448 
8449 		freemsg(mp);
8450 		mp = next;
8451 	}
8452 }
8453 
8454 mblk_t *
8455 copymsgchain(mblk_t *mp)
8456 {
8457 	mblk_t	*nmp = NULL;
8458 	mblk_t	**nmpp = &nmp;
8459 
8460 	for (; mp != NULL; mp = mp->b_next) {
8461 		if ((*nmpp = copymsg(mp)) == NULL) {
8462 			freemsgchain(nmp);
8463 			return (NULL);
8464 		}
8465 
8466 		nmpp = &((*nmpp)->b_next);
8467 	}
8468 
8469 	return (nmp);
8470 }
8471 
8472 /* NOTE: Do not add code after this point. */
8473 #undef QLOCK
8474 
8475 /*
8476  * replacement for QLOCK macro for those that can't use it.
8477  */
8478 kmutex_t *
8479 QLOCK(queue_t *q)
8480 {
8481 	return (&(q)->q_lock);
8482 }
8483 
8484 /*
8485  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8486  */
8487 #undef runqueues
8488 void
8489 runqueues(void)
8490 {
8491 }
8492 
8493 #undef queuerun
8494 void
8495 queuerun(void)
8496 {
8497 }
8498 
8499 /*
8500  * Initialize the STR stack instance, which tracks autopush and persistent
8501  * links.
8502  */
8503 /* ARGSUSED */
8504 static void *
8505 str_stack_init(netstackid_t stackid, netstack_t *ns)
8506 {
8507 	str_stack_t	*ss;
8508 	int i;
8509 
8510 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8511 	ss->ss_netstack = ns;
8512 
8513 	/*
8514 	 * set up autopush
8515 	 */
8516 	sad_initspace(ss);
8517 
8518 	/*
8519 	 * set up mux_node structures.
8520 	 */
8521 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8522 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8523 					    ss->ss_devcnt), KM_SLEEP);
8524 	for (i = 0; i < ss->ss_devcnt; i++)
8525 		ss->ss_mux_nodes[i].mn_imaj = i;
8526 	return (ss);
8527 }
8528 
8529 /*
8530  * Note: run at zone shutdown and not destroy so that the PLINKs are
8531  * gone by the time other cleanup happens from the destroy callbacks.
8532  */
8533 static void
8534 str_stack_shutdown(netstackid_t stackid, void *arg)
8535 {
8536 	str_stack_t *ss = (str_stack_t *)arg;
8537 	int i;
8538 	cred_t *cr;
8539 
8540 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8541 	ASSERT(cr != NULL);
8542 
8543 	/* Undo all the I_PLINKs for this zone */
8544 	for (i = 0; i < ss->ss_devcnt; i++) {
8545 		struct mux_edge		*ep;
8546 		ldi_handle_t		lh;
8547 		ldi_ident_t		li;
8548 		int			ret;
8549 		int			rval;
8550 		dev_t			rdev;
8551 
8552 		ep = ss->ss_mux_nodes[i].mn_outp;
8553 		if (ep == NULL)
8554 			continue;
8555 		ret = ldi_ident_from_major((major_t)i, &li);
8556 		if (ret != 0) {
8557 			continue;
8558 		}
8559 		rdev = ep->me_dev;
8560 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8561 		    cr, &lh, li);
8562 		if (ret != 0) {
8563 			ldi_ident_release(li);
8564 			continue;
8565 		}
8566 
8567 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8568 		    cr, &rval);
8569 		if (ret) {
8570 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8571 			ldi_ident_release(li);
8572 			continue;
8573 		}
8574 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8575 
8576 		/* Close layered handles */
8577 		ldi_ident_release(li);
8578 	}
8579 	crfree(cr);
8580 
8581 	sad_freespace(ss);
8582 
8583 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8584 	ss->ss_mux_nodes = NULL;
8585 }
8586 
8587 /*
8588  * Free the structure; str_stack_shutdown did the other cleanup work.
8589  */
8590 /* ARGSUSED */
8591 static void
8592 str_stack_fini(netstackid_t stackid, void *arg)
8593 {
8594 	str_stack_t	*ss = (str_stack_t *)arg;
8595 
8596 	kmem_free(ss, sizeof (*ss));
8597 }
8598