xref: /illumos-gate/usr/src/uts/common/os/strsubr.c (revision bc1f688b4872ace323eaddbb1a6365d054e7bf56)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
22 /*	  All Rights Reserved  	*/
23 
24 
25 /*
26  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  * Copyright (c) 2016 by Delphix. All rights reserved.
29  * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
30  */
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 due 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 /*
211  * run_queues is no longer used, but is kept in case some 3rd party
212  * module/driver decides to use it.
213  */
214 int run_queues = 0;
215 
216 /*
217  * sq_max_size is the depth of the syncq (in number of messages) before
218  * qfill_syncq() starts QFULL'ing destination queues. As its primary
219  * consumer - IP is no longer D_MTPERMOD, but there may be other
220  * modules/drivers depend on this syncq flow control, we prefer to
221  * choose a large number as the default value. For potential
222  * performance gain, this value is tunable in /etc/system.
223  */
224 int sq_max_size = 10000;
225 
226 /*
227  * The number of ciputctrl structures per syncq and stream we create when
228  * needed.
229  */
230 int n_ciputctrl;
231 int max_n_ciputctrl = 16;
232 /*
233  * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
234  */
235 int min_n_ciputctrl = 2;
236 
237 /*
238  * Per-driver/module syncqs
239  * ========================
240  *
241  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
242  * perdm structures, new entries being added (and new syncqs allocated) when
243  * setq() encounters a module/driver with a streamtab that it hasn't seen
244  * before.
245  * The reason for this mechanism is that some modules and drivers share a
246  * common streamtab and it is necessary for those modules and drivers to also
247  * share a common PERMOD syncq.
248  *
249  * perdm_list --> dm_str == streamtab_1
250  *                dm_sq == syncq_1
251  *                dm_ref
252  *                dm_next --> dm_str == streamtab_2
253  *                            dm_sq == syncq_2
254  *                            dm_ref
255  *                            dm_next --> ... NULL
256  *
257  * The dm_ref field is incremented for each new driver/module that takes
258  * a reference to the perdm structure and hence shares the syncq.
259  * References are held in the fmodsw_impl_t structure for each STREAMS module
260  * or the dev_impl array (indexed by device major number) for each driver.
261  *
262  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
263  *		     ^                 ^ ^               ^
264  *                   |  ______________/  |               |
265  *                   | /                 |               |
266  * dev_impl:     ...|x|y|...          module A	      module B
267  *
268  * When a module/driver is unloaded the reference count is decremented and,
269  * when it falls to zero, the perdm structure is removed from the list and
270  * the syncq is freed (see rele_dm()).
271  */
272 perdm_t *perdm_list = NULL;
273 static krwlock_t perdm_rwlock;
274 cdevsw_impl_t *devimpl;
275 
276 extern struct qinit strdata;
277 extern struct qinit stwdata;
278 
279 static void runservice(queue_t *);
280 static void streams_bufcall_service(void);
281 static void streams_qbkgrnd_service(void);
282 static void streams_sqbkgrnd_service(void);
283 static syncq_t *new_syncq(void);
284 static void free_syncq(syncq_t *);
285 static void outer_insert(syncq_t *, syncq_t *);
286 static void outer_remove(syncq_t *, syncq_t *);
287 static void write_now(syncq_t *);
288 static void clr_qfull(queue_t *);
289 static void runbufcalls(void);
290 static void sqenable(syncq_t *);
291 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
292 static void wait_q_syncq(queue_t *);
293 static void backenable_insertedq(queue_t *);
294 
295 static void queue_service(queue_t *);
296 static void stream_service(stdata_t *);
297 static void syncq_service(syncq_t *);
298 static void qwriter_outer_service(syncq_t *);
299 static void mblk_free(mblk_t *);
300 #ifdef DEBUG
301 static int qprocsareon(queue_t *);
302 #endif
303 
304 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
305 static void reset_nfsrv_ptr(queue_t *, queue_t *);
306 void set_qfull(queue_t *);
307 
308 static void sq_run_events(syncq_t *);
309 static int propagate_syncq(queue_t *);
310 
311 static void	blocksq(syncq_t *, ushort_t, int);
312 static void	unblocksq(syncq_t *, ushort_t, int);
313 static int	dropsq(syncq_t *, uint16_t);
314 static void	emptysq(syncq_t *);
315 static sqlist_t *sqlist_alloc(struct stdata *, int);
316 static void	sqlist_free(sqlist_t *);
317 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
318 static void	sqlist_insert(sqlist_t *, syncq_t *);
319 static void	sqlist_insertall(sqlist_t *, queue_t *);
320 
321 static void	strsetuio(stdata_t *);
322 
323 struct kmem_cache *stream_head_cache;
324 struct kmem_cache *queue_cache;
325 struct kmem_cache *syncq_cache;
326 struct kmem_cache *qband_cache;
327 struct kmem_cache *linkinfo_cache;
328 struct kmem_cache *ciputctrl_cache = NULL;
329 
330 static linkinfo_t *linkinfo_list;
331 
332 /* Global esballoc throttling queue */
333 static esb_queue_t system_esbq;
334 
335 /* Array of esballoc throttling queues, of length esbq_nelem */
336 static esb_queue_t *volatile system_esbq_array;
337 static int esbq_nelem;
338 static kmutex_t esbq_lock;
339 static int esbq_log2_cpus_per_q = 0;
340 
341 /* Scale the system_esbq length by setting number of CPUs per queue. */
342 uint_t esbq_cpus_per_q = 1;
343 
344 /*
345  * esballoc tunable parameters.
346  */
347 int		esbq_max_qlen = 0x16;	/* throttled queue length */
348 clock_t		esbq_timeout = 0x8;	/* timeout to process esb queue */
349 
350 /*
351  * Routines to handle esballoc queueing.
352  */
353 static void esballoc_process_queue(esb_queue_t *);
354 static void esballoc_enqueue_mblk(mblk_t *);
355 static void esballoc_timer(void *);
356 static void esballoc_set_timer(esb_queue_t *, clock_t);
357 static void esballoc_mblk_free(mblk_t *);
358 
359 /*
360  *  Qinit structure and Module_info structures
361  *	for passthru read and write queues
362  */
363 
364 static int pass_rput(queue_t *, mblk_t *);
365 static int pass_wput(queue_t *, mblk_t *);
366 static queue_t *link_addpassthru(stdata_t *);
367 static void link_rempassthru(queue_t *);
368 
369 struct  module_info passthru_info = {
370 	0,
371 	"passthru",
372 	0,
373 	INFPSZ,
374 	STRHIGH,
375 	STRLOW
376 };
377 
378 struct  qinit passthru_rinit = {
379 	pass_rput,
380 	NULL,
381 	NULL,
382 	NULL,
383 	NULL,
384 	&passthru_info,
385 	NULL
386 };
387 
388 struct  qinit passthru_winit = {
389 	pass_wput,
390 	NULL,
391 	NULL,
392 	NULL,
393 	NULL,
394 	&passthru_info,
395 	NULL
396 };
397 
398 /*
399  * Verify correctness of list head/tail pointers.
400  */
401 #define	LISTCHECK(head, tail, link) {				\
402 	EQUIV(head, tail);					\
403 	IMPLY(tail != NULL, tail->link == NULL);		\
404 }
405 
406 /*
407  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
408  * using a `link' field.
409  */
410 #define	ENQUEUE(el, head, tail, link) {				\
411 	ASSERT(el->link == NULL);				\
412 	LISTCHECK(head, tail, link);				\
413 	if (head == NULL)					\
414 		head = el;					\
415 	else							\
416 		tail->link = el;				\
417 	tail = el;						\
418 }
419 
420 /*
421  * Dequeue the first element of the list denoted by `head' and `tail' pointers
422  * using a `link' field and put result into `el'.
423  */
424 #define	DQ(el, head, tail, link) {				\
425 	LISTCHECK(head, tail, link);				\
426 	el = head;						\
427 	if (head != NULL) {					\
428 		head = head->link;				\
429 		if (head == NULL)				\
430 			tail = NULL;				\
431 		el->link = NULL;				\
432 	}							\
433 }
434 
435 /*
436  * Remove `el' from the list using `chase' and `curr' pointers and return result
437  * in `succeed'.
438  */
439 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
440 	LISTCHECK(head, tail, link);				\
441 	chase = NULL;						\
442 	succeed = 0;						\
443 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
444 		chase = curr;					\
445 	if (curr != NULL) {					\
446 		succeed = 1;					\
447 		ASSERT(curr == el);				\
448 		if (chase != NULL)				\
449 			chase->link = curr->link;		\
450 		else						\
451 			head = curr->link;			\
452 		curr->link = NULL;				\
453 		if (curr == tail)				\
454 			tail = chase;				\
455 	}							\
456 	LISTCHECK(head, tail, link);				\
457 }
458 
459 /* Handling of delayed messages on the inner syncq. */
460 
461 /*
462  * DEBUG versions should use function versions (to simplify tracing) and
463  * non-DEBUG kernels should use macro versions.
464  */
465 
466 /*
467  * Put a queue on the syncq list of queues.
468  * Assumes SQLOCK held.
469  */
470 #define	SQPUT_Q(sq, qp)							\
471 {									\
472 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
473 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
474 		/* The queue should not be linked anywhere */		\
475 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
476 		/* Head and tail may only be NULL simultaneously */	\
477 		EQUIV(sq->sq_head, sq->sq_tail);			\
478 		/* Queue may be only enqueued on its syncq */		\
479 		ASSERT(sq == qp->q_syncq);				\
480 		/* Check the correctness of SQ_MESSAGES flag */		\
481 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
482 		/* Sanity check first/last elements of the list */	\
483 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
484 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
485 		/*							\
486 		 * Sanity check of priority field: empty queue should	\
487 		 * have zero priority					\
488 		 * and nqueues equal to zero.				\
489 		 */							\
490 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
491 		/* Sanity check of sq_nqueues field */			\
492 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
493 		if (sq->sq_head == NULL) {				\
494 			sq->sq_head = sq->sq_tail = qp;			\
495 			sq->sq_flags |= SQ_MESSAGES;			\
496 		} else if (qp->q_spri == 0) {				\
497 			qp->q_sqprev = sq->sq_tail;			\
498 			sq->sq_tail->q_sqnext = qp;			\
499 			sq->sq_tail = qp;				\
500 		} else {						\
501 			/*						\
502 			 * Put this queue in priority order: higher	\
503 			 * priority gets closer to the head.		\
504 			 */						\
505 			queue_t **qpp = &sq->sq_tail;			\
506 			queue_t *qnext = NULL;				\
507 									\
508 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
509 				qnext = *qpp;				\
510 				qpp = &(*qpp)->q_sqprev;		\
511 			}						\
512 			qp->q_sqnext = qnext;				\
513 			qp->q_sqprev = *qpp;				\
514 			if (*qpp != NULL) {				\
515 				(*qpp)->q_sqnext = qp;			\
516 			} else {					\
517 				sq->sq_head = qp;			\
518 				sq->sq_pri = sq->sq_head->q_spri;	\
519 			}						\
520 			*qpp = qp;					\
521 		}							\
522 		qp->q_sqflags |= Q_SQQUEUED;				\
523 		qp->q_sqtstamp = ddi_get_lbolt();			\
524 		sq->sq_nqueues++;					\
525 	}								\
526 }
527 
528 /*
529  * Remove a queue from the syncq list
530  * Assumes SQLOCK held.
531  */
532 #define	SQRM_Q(sq, qp)							\
533 	{								\
534 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
535 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
536 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
537 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
538 		/* Check that the queue is actually in the list */	\
539 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
540 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
541 		ASSERT(sq->sq_nqueues != 0);				\
542 		if (qp->q_sqprev == NULL) {				\
543 			/* First queue on list, make head q_sqnext */	\
544 			sq->sq_head = qp->q_sqnext;			\
545 		} else {						\
546 			/* Make prev->next == next */			\
547 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
548 		}							\
549 		if (qp->q_sqnext == NULL) {				\
550 			/* Last queue on list, make tail sqprev */	\
551 			sq->sq_tail = qp->q_sqprev;			\
552 		} else {						\
553 			/* Make next->prev == prev */			\
554 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
555 		}							\
556 		/* clear out references on this queue */		\
557 		qp->q_sqprev = qp->q_sqnext = NULL;			\
558 		qp->q_sqflags &= ~Q_SQQUEUED;				\
559 		/* If there is nothing queued, clear SQ_MESSAGES */	\
560 		if (sq->sq_head != NULL) {				\
561 			sq->sq_pri = sq->sq_head->q_spri;		\
562 		} else	{						\
563 			sq->sq_flags &= ~SQ_MESSAGES;			\
564 			sq->sq_pri = 0;					\
565 		}							\
566 		sq->sq_nqueues--;					\
567 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
568 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
569 	}
570 
571 /* Hide the definition from the header file. */
572 #ifdef SQPUT_MP
573 #undef SQPUT_MP
574 #endif
575 
576 /*
577  * Put a message on the queue syncq.
578  * Assumes QLOCK held.
579  */
580 #define	SQPUT_MP(qp, mp)						\
581 	{								\
582 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
583 		ASSERT(qp->q_sqhead == NULL ||				\
584 		    (qp->q_sqtail != NULL &&				\
585 		    qp->q_sqtail->b_next == NULL));			\
586 		qp->q_syncqmsgs++;					\
587 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
588 		if (qp->q_sqhead == NULL) {				\
589 			qp->q_sqhead = qp->q_sqtail = mp;		\
590 		} else {						\
591 			qp->q_sqtail->b_next = mp;			\
592 			qp->q_sqtail = mp;				\
593 		}							\
594 		ASSERT(qp->q_syncqmsgs > 0);				\
595 		set_qfull(qp);						\
596 	}
597 
598 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
599 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
600 		if ((sq)->sq_ciputctrl != NULL) {			\
601 			int i;						\
602 			int nlocks = (sq)->sq_nciputctrl;		\
603 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
604 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
605 			for (i = 0; i <= nlocks; i++) {			\
606 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
607 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
608 			}						\
609 		}							\
610 	}
611 
612 
613 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
614 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
615 		if ((sq)->sq_ciputctrl != NULL) {			\
616 			int i;						\
617 			int nlocks = (sq)->sq_nciputctrl;		\
618 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
619 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
620 			for (i = 0; i <= nlocks; i++) {			\
621 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
622 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
623 			}						\
624 		}							\
625 	}
626 
627 /*
628  * Run service procedures for all queues in the stream head.
629  */
630 #define	STR_SERVICE(stp, q) {						\
631 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
632 	while (stp->sd_qhead != NULL) {					\
633 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
634 		ASSERT(stp->sd_nqueues > 0);				\
635 		stp->sd_nqueues--;					\
636 		ASSERT(!(q->q_flag & QINSERVICE));			\
637 		mutex_exit(&stp->sd_qlock);				\
638 		queue_service(q);					\
639 		mutex_enter(&stp->sd_qlock);				\
640 	}								\
641 	ASSERT(stp->sd_nqueues == 0);					\
642 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
643 }
644 
645 /*
646  * Constructor/destructor routines for the stream head cache
647  */
648 /* ARGSUSED */
649 static int
650 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
651 {
652 	stdata_t *stp = buf;
653 
654 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
655 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
656 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
657 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
658 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
659 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
660 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
661 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
662 	stp->sd_wrq = NULL;
663 
664 	return (0);
665 }
666 
667 /* ARGSUSED */
668 static void
669 stream_head_destructor(void *buf, void *cdrarg)
670 {
671 	stdata_t *stp = buf;
672 
673 	mutex_destroy(&stp->sd_lock);
674 	mutex_destroy(&stp->sd_reflock);
675 	mutex_destroy(&stp->sd_qlock);
676 	cv_destroy(&stp->sd_monitor);
677 	cv_destroy(&stp->sd_iocmonitor);
678 	cv_destroy(&stp->sd_refmonitor);
679 	cv_destroy(&stp->sd_qcv);
680 	cv_destroy(&stp->sd_zcopy_wait);
681 }
682 
683 /*
684  * Constructor/destructor routines for the queue cache
685  */
686 /* ARGSUSED */
687 static int
688 queue_constructor(void *buf, void *cdrarg, int kmflags)
689 {
690 	queinfo_t *qip = buf;
691 	queue_t *qp = &qip->qu_rqueue;
692 	queue_t *wqp = &qip->qu_wqueue;
693 	syncq_t	*sq = &qip->qu_syncq;
694 
695 	qp->q_first = NULL;
696 	qp->q_link = NULL;
697 	qp->q_count = 0;
698 	qp->q_mblkcnt = 0;
699 	qp->q_sqhead = NULL;
700 	qp->q_sqtail = NULL;
701 	qp->q_sqnext = NULL;
702 	qp->q_sqprev = NULL;
703 	qp->q_sqflags = 0;
704 	qp->q_rwcnt = 0;
705 	qp->q_spri = 0;
706 
707 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
708 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
709 
710 	wqp->q_first = NULL;
711 	wqp->q_link = NULL;
712 	wqp->q_count = 0;
713 	wqp->q_mblkcnt = 0;
714 	wqp->q_sqhead = NULL;
715 	wqp->q_sqtail = NULL;
716 	wqp->q_sqnext = NULL;
717 	wqp->q_sqprev = NULL;
718 	wqp->q_sqflags = 0;
719 	wqp->q_rwcnt = 0;
720 	wqp->q_spri = 0;
721 
722 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
723 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
724 
725 	sq->sq_head = NULL;
726 	sq->sq_tail = NULL;
727 	sq->sq_evhead = NULL;
728 	sq->sq_evtail = NULL;
729 	sq->sq_callbpend = NULL;
730 	sq->sq_outer = NULL;
731 	sq->sq_onext = NULL;
732 	sq->sq_oprev = NULL;
733 	sq->sq_next = NULL;
734 	sq->sq_svcflags = 0;
735 	sq->sq_servcount = 0;
736 	sq->sq_needexcl = 0;
737 	sq->sq_nqueues = 0;
738 	sq->sq_pri = 0;
739 
740 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
741 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
742 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
743 
744 	return (0);
745 }
746 
747 /* ARGSUSED */
748 static void
749 queue_destructor(void *buf, void *cdrarg)
750 {
751 	queinfo_t *qip = buf;
752 	queue_t *qp = &qip->qu_rqueue;
753 	queue_t *wqp = &qip->qu_wqueue;
754 	syncq_t	*sq = &qip->qu_syncq;
755 
756 	ASSERT(qp->q_sqhead == NULL);
757 	ASSERT(wqp->q_sqhead == NULL);
758 	ASSERT(qp->q_sqnext == NULL);
759 	ASSERT(wqp->q_sqnext == NULL);
760 	ASSERT(qp->q_rwcnt == 0);
761 	ASSERT(wqp->q_rwcnt == 0);
762 
763 	mutex_destroy(&qp->q_lock);
764 	cv_destroy(&qp->q_wait);
765 
766 	mutex_destroy(&wqp->q_lock);
767 	cv_destroy(&wqp->q_wait);
768 
769 	mutex_destroy(&sq->sq_lock);
770 	cv_destroy(&sq->sq_wait);
771 	cv_destroy(&sq->sq_exitwait);
772 }
773 
774 /*
775  * Constructor/destructor routines for the syncq cache
776  */
777 /* ARGSUSED */
778 static int
779 syncq_constructor(void *buf, void *cdrarg, int kmflags)
780 {
781 	syncq_t	*sq = buf;
782 
783 	bzero(buf, sizeof (syncq_t));
784 
785 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
786 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
787 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
788 
789 	return (0);
790 }
791 
792 /* ARGSUSED */
793 static void
794 syncq_destructor(void *buf, void *cdrarg)
795 {
796 	syncq_t	*sq = buf;
797 
798 	ASSERT(sq->sq_head == NULL);
799 	ASSERT(sq->sq_tail == NULL);
800 	ASSERT(sq->sq_evhead == NULL);
801 	ASSERT(sq->sq_evtail == NULL);
802 	ASSERT(sq->sq_callbpend == NULL);
803 	ASSERT(sq->sq_callbflags == 0);
804 	ASSERT(sq->sq_outer == NULL);
805 	ASSERT(sq->sq_onext == NULL);
806 	ASSERT(sq->sq_oprev == NULL);
807 	ASSERT(sq->sq_next == NULL);
808 	ASSERT(sq->sq_needexcl == 0);
809 	ASSERT(sq->sq_svcflags == 0);
810 	ASSERT(sq->sq_servcount == 0);
811 	ASSERT(sq->sq_nqueues == 0);
812 	ASSERT(sq->sq_pri == 0);
813 	ASSERT(sq->sq_count == 0);
814 	ASSERT(sq->sq_rmqcount == 0);
815 	ASSERT(sq->sq_cancelid == 0);
816 	ASSERT(sq->sq_ciputctrl == NULL);
817 	ASSERT(sq->sq_nciputctrl == 0);
818 	ASSERT(sq->sq_type == 0);
819 	ASSERT(sq->sq_flags == 0);
820 
821 	mutex_destroy(&sq->sq_lock);
822 	cv_destroy(&sq->sq_wait);
823 	cv_destroy(&sq->sq_exitwait);
824 }
825 
826 /* ARGSUSED */
827 static int
828 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
829 {
830 	ciputctrl_t *cip = buf;
831 	int i;
832 
833 	for (i = 0; i < n_ciputctrl; i++) {
834 		cip[i].ciputctrl_count = SQ_FASTPUT;
835 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
836 	}
837 
838 	return (0);
839 }
840 
841 /* ARGSUSED */
842 static void
843 ciputctrl_destructor(void *buf, void *cdrarg)
844 {
845 	ciputctrl_t *cip = buf;
846 	int i;
847 
848 	for (i = 0; i < n_ciputctrl; i++) {
849 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
850 		mutex_destroy(&cip[i].ciputctrl_lock);
851 	}
852 }
853 
854 /*
855  * Init routine run from main at boot time.
856  */
857 void
858 strinit(void)
859 {
860 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
861 
862 	stream_head_cache = kmem_cache_create("stream_head_cache",
863 	    sizeof (stdata_t), 0,
864 	    stream_head_constructor, stream_head_destructor, NULL,
865 	    NULL, NULL, 0);
866 
867 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
868 	    queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
869 
870 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
871 	    syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
872 
873 	qband_cache = kmem_cache_create("qband_cache",
874 	    sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
875 
876 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
877 	    sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
878 
879 	n_ciputctrl = ncpus;
880 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
881 	ASSERT(n_ciputctrl >= 1);
882 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
883 	if (n_ciputctrl >= min_n_ciputctrl) {
884 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
885 		    sizeof (ciputctrl_t) * n_ciputctrl,
886 		    sizeof (ciputctrl_t), ciputctrl_constructor,
887 		    ciputctrl_destructor, NULL, NULL, NULL, 0);
888 	}
889 
890 	streams_taskq = system_taskq;
891 
892 	if (streams_taskq == NULL)
893 		panic("strinit: no memory for streams taskq!");
894 
895 	bc_bkgrnd_thread = thread_create(NULL, 0,
896 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
897 
898 	streams_qbkgrnd_thread = thread_create(NULL, 0,
899 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
900 
901 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
902 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
903 
904 	/*
905 	 * Create STREAMS kstats.
906 	 */
907 	str_kstat = kstat_create("streams", 0, "strstat",
908 	    "net", KSTAT_TYPE_NAMED,
909 	    sizeof (str_statistics) / sizeof (kstat_named_t),
910 	    KSTAT_FLAG_VIRTUAL);
911 
912 	if (str_kstat != NULL) {
913 		str_kstat->ks_data = &str_statistics;
914 		kstat_install(str_kstat);
915 	}
916 
917 	/*
918 	 * TPI support routine initialisation.
919 	 */
920 	tpi_init();
921 
922 	/*
923 	 * Handle to have autopush and persistent link information per
924 	 * zone.
925 	 * Note: uses shutdown hook instead of destroy hook so that the
926 	 * persistent links can be torn down before the destroy hooks
927 	 * in the TCP/IP stack are called.
928 	 */
929 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
930 	    str_stack_fini);
931 }
932 
933 void
934 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
935 {
936 	struct stdata *stp;
937 
938 	ASSERT(vp->v_stream);
939 	stp = vp->v_stream;
940 	/* Have to hold sd_lock to prevent siglist from changing */
941 	mutex_enter(&stp->sd_lock);
942 	if (stp->sd_sigflags & event)
943 		strsendsig(stp->sd_siglist, event, band, error);
944 	mutex_exit(&stp->sd_lock);
945 }
946 
947 /*
948  * Send the "sevent" set of signals to a process.
949  * This might send more than one signal if the process is registered
950  * for multiple events. The caller should pass in an sevent that only
951  * includes the events for which the process has registered.
952  */
953 static void
954 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
955     uchar_t band, int error)
956 {
957 	ASSERT(MUTEX_HELD(&proc->p_lock));
958 
959 	info->si_band = 0;
960 	info->si_errno = 0;
961 
962 	if (sevent & S_ERROR) {
963 		sevent &= ~S_ERROR;
964 		info->si_code = POLL_ERR;
965 		info->si_errno = error;
966 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
967 		    "strsendsig:proc %p info %p", proc, info);
968 		sigaddq(proc, NULL, info, KM_NOSLEEP);
969 		info->si_errno = 0;
970 	}
971 	if (sevent & S_HANGUP) {
972 		sevent &= ~S_HANGUP;
973 		info->si_code = POLL_HUP;
974 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
975 		    "strsendsig:proc %p info %p", proc, info);
976 		sigaddq(proc, NULL, info, KM_NOSLEEP);
977 	}
978 	if (sevent & S_HIPRI) {
979 		sevent &= ~S_HIPRI;
980 		info->si_code = POLL_PRI;
981 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
982 		    "strsendsig:proc %p info %p", proc, info);
983 		sigaddq(proc, NULL, info, KM_NOSLEEP);
984 	}
985 	if (sevent & S_RDBAND) {
986 		sevent &= ~S_RDBAND;
987 		if (events & S_BANDURG)
988 			sigtoproc(proc, NULL, SIGURG);
989 		else
990 			sigtoproc(proc, NULL, SIGPOLL);
991 	}
992 	if (sevent & S_WRBAND) {
993 		sevent &= ~S_WRBAND;
994 		sigtoproc(proc, NULL, SIGPOLL);
995 	}
996 	if (sevent & S_INPUT) {
997 		sevent &= ~S_INPUT;
998 		info->si_code = POLL_IN;
999 		info->si_band = band;
1000 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1001 		    "strsendsig:proc %p info %p", proc, info);
1002 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1003 		info->si_band = 0;
1004 	}
1005 	if (sevent & S_OUTPUT) {
1006 		sevent &= ~S_OUTPUT;
1007 		info->si_code = POLL_OUT;
1008 		info->si_band = band;
1009 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1010 		    "strsendsig:proc %p info %p", proc, info);
1011 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1012 		info->si_band = 0;
1013 	}
1014 	if (sevent & S_MSG) {
1015 		sevent &= ~S_MSG;
1016 		info->si_code = POLL_MSG;
1017 		info->si_band = band;
1018 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1019 		    "strsendsig:proc %p info %p", proc, info);
1020 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1021 		info->si_band = 0;
1022 	}
1023 	if (sevent & S_RDNORM) {
1024 		sevent &= ~S_RDNORM;
1025 		sigtoproc(proc, NULL, SIGPOLL);
1026 	}
1027 	if (sevent != 0) {
1028 		panic("strsendsig: unknown event(s) %x", sevent);
1029 	}
1030 }
1031 
1032 /*
1033  * Send SIGPOLL/SIGURG signal to all processes and process groups
1034  * registered on the given signal list that want a signal for at
1035  * least one of the specified events.
1036  *
1037  * Must be called with exclusive access to siglist (caller holding sd_lock).
1038  *
1039  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1040  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1041  * while it is in the siglist.
1042  *
1043  * For performance reasons (MP scalability) the code drops pidlock
1044  * when sending signals to a single process.
1045  * When sending to a process group the code holds
1046  * pidlock to prevent the membership in the process group from changing
1047  * while walking the p_pglink list.
1048  */
1049 void
1050 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1051 {
1052 	strsig_t *ssp;
1053 	k_siginfo_t info;
1054 	struct pid *pidp;
1055 	proc_t  *proc;
1056 
1057 	info.si_signo = SIGPOLL;
1058 	info.si_errno = 0;
1059 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1060 		int sevent;
1061 
1062 		sevent = ssp->ss_events & event;
1063 		if (sevent == 0)
1064 			continue;
1065 
1066 		if ((pidp = ssp->ss_pidp) == NULL) {
1067 			/* pid was released but still on event list */
1068 			continue;
1069 		}
1070 
1071 
1072 		if (ssp->ss_pid > 0) {
1073 			/*
1074 			 * XXX This unfortunately still generates
1075 			 * a signal when a fd is closed but
1076 			 * the proc is active.
1077 			 */
1078 			ASSERT(ssp->ss_pid == pidp->pid_id);
1079 
1080 			mutex_enter(&pidlock);
1081 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1082 			if (proc == NULL) {
1083 				mutex_exit(&pidlock);
1084 				continue;
1085 			}
1086 			mutex_enter(&proc->p_lock);
1087 			mutex_exit(&pidlock);
1088 			dosendsig(proc, ssp->ss_events, sevent, &info,
1089 			    band, error);
1090 			mutex_exit(&proc->p_lock);
1091 		} else {
1092 			/*
1093 			 * Send to process group. Hold pidlock across
1094 			 * calls to dosendsig().
1095 			 */
1096 			pid_t pgrp = -ssp->ss_pid;
1097 
1098 			mutex_enter(&pidlock);
1099 			proc = pgfind_zone(pgrp, ALL_ZONES);
1100 			while (proc != NULL) {
1101 				mutex_enter(&proc->p_lock);
1102 				dosendsig(proc, ssp->ss_events, sevent,
1103 				    &info, band, error);
1104 				mutex_exit(&proc->p_lock);
1105 				proc = proc->p_pglink;
1106 			}
1107 			mutex_exit(&pidlock);
1108 		}
1109 	}
1110 }
1111 
1112 /*
1113  * Attach a stream device or module.
1114  * qp is a read queue; the new queue goes in so its next
1115  * read ptr is the argument, and the write queue corresponding
1116  * to the argument points to this queue. Return 0 on success,
1117  * or a non-zero errno on failure.
1118  */
1119 int
1120 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1121     boolean_t is_insert)
1122 {
1123 	major_t			major;
1124 	cdevsw_impl_t		*dp;
1125 	struct streamtab	*str;
1126 	queue_t			*rq;
1127 	queue_t			*wrq;
1128 	uint32_t		qflag;
1129 	uint32_t		sqtype;
1130 	perdm_t			*dmp;
1131 	int			error;
1132 	int			sflag;
1133 
1134 	rq = allocq();
1135 	wrq = _WR(rq);
1136 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1137 
1138 	if (fp != NULL) {
1139 		str = fp->f_str;
1140 		qflag = fp->f_qflag;
1141 		sqtype = fp->f_sqtype;
1142 		dmp = fp->f_dmp;
1143 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1144 		sflag = MODOPEN;
1145 
1146 		/*
1147 		 * stash away a pointer to the module structure so we can
1148 		 * unref it in qdetach.
1149 		 */
1150 		rq->q_fp = fp;
1151 	} else {
1152 		ASSERT(!is_insert);
1153 
1154 		major = getmajor(*devp);
1155 		dp = &devimpl[major];
1156 
1157 		str = dp->d_str;
1158 		ASSERT(str == STREAMSTAB(major));
1159 
1160 		qflag = dp->d_qflag;
1161 		ASSERT(qflag & QISDRV);
1162 		sqtype = dp->d_sqtype;
1163 
1164 		/* create perdm_t if needed */
1165 		if (NEED_DM(dp->d_dmp, qflag))
1166 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1167 
1168 		dmp = dp->d_dmp;
1169 		sflag = 0;
1170 	}
1171 
1172 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1173 	    "qattach:qflag == %X(%X)", qflag, *devp);
1174 
1175 	/* setq might sleep in allocator - avoid holding locks. */
1176 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1177 
1178 	/*
1179 	 * Before calling the module's open routine, set up the q_next
1180 	 * pointer for inserting a module in the middle of a stream.
1181 	 *
1182 	 * Note that we can always set _QINSERTING and set up q_next
1183 	 * pointer for both inserting and pushing a module.  Then there
1184 	 * is no need for the is_insert parameter.  In insertq(), called
1185 	 * by qprocson(), assume that q_next of the new module always points
1186 	 * to the correct queue and use it for insertion.  Everything should
1187 	 * work out fine.  But in the first release of _I_INSERT, we
1188 	 * distinguish between inserting and pushing to make sure that
1189 	 * pushing a module follows the same code path as before.
1190 	 */
1191 	if (is_insert) {
1192 		rq->q_flag |= _QINSERTING;
1193 		rq->q_next = qp;
1194 	}
1195 
1196 	/*
1197 	 * If there is an outer perimeter get exclusive access during
1198 	 * the open procedure.  Bump up the reference count on the queue.
1199 	 */
1200 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1201 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1202 	if (error != 0)
1203 		goto failed;
1204 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1205 	ASSERT(qprocsareon(rq));
1206 	return (0);
1207 
1208 failed:
1209 	rq->q_flag &= ~_QINSERTING;
1210 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1211 		qprocsoff(rq);
1212 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1213 	rq->q_next = wrq->q_next = NULL;
1214 	qdetach(rq, 0, 0, crp, B_FALSE);
1215 	return (error);
1216 }
1217 
1218 /*
1219  * Handle second open of stream. For modules, set the
1220  * last argument to MODOPEN and do not pass any open flags.
1221  * Ignore dummydev since this is not the first open.
1222  */
1223 int
1224 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1225 {
1226 	int	error;
1227 	dev_t dummydev;
1228 	queue_t *wqp = _WR(qp);
1229 
1230 	ASSERT(qp->q_flag & QREADR);
1231 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1232 
1233 	dummydev = *devp;
1234 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1235 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1236 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1237 		mutex_enter(&STREAM(qp)->sd_lock);
1238 		qp->q_stream->sd_flag |= STREOPENFAIL;
1239 		mutex_exit(&STREAM(qp)->sd_lock);
1240 		return (error);
1241 	}
1242 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1243 
1244 	/*
1245 	 * successful open should have done qprocson()
1246 	 */
1247 	ASSERT(qprocsareon(_RD(qp)));
1248 	return (0);
1249 }
1250 
1251 /*
1252  * Detach a stream module or device.
1253  * If clmode == 1 then the module or driver was opened and its
1254  * close routine must be called. If clmode == 0, the module
1255  * or driver was never opened or the open failed, and so its close
1256  * should not be called.
1257  */
1258 void
1259 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1260 {
1261 	queue_t *wqp = _WR(qp);
1262 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1263 
1264 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1265 		stream_runservice(STREAM(qp));
1266 
1267 	if (clmode) {
1268 		/*
1269 		 * Make sure that all the messages on the write side syncq are
1270 		 * processed and nothing is left. Since we are closing, no new
1271 		 * messages may appear there.
1272 		 */
1273 		wait_q_syncq(wqp);
1274 
1275 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1276 		if (is_remove) {
1277 			mutex_enter(QLOCK(qp));
1278 			qp->q_flag |= _QREMOVING;
1279 			mutex_exit(QLOCK(qp));
1280 		}
1281 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1282 		/*
1283 		 * Check that qprocsoff() was actually called.
1284 		 */
1285 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1286 
1287 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1288 	} else {
1289 		disable_svc(qp);
1290 	}
1291 
1292 	/*
1293 	 * Allow any threads blocked in entersq to proceed and discover
1294 	 * the QWCLOSE is set.
1295 	 * Note: This assumes that all users of entersq check QWCLOSE.
1296 	 * Currently runservice is the only entersq that can happen
1297 	 * after removeq has finished.
1298 	 * Removeq will have discarded all messages destined to the closing
1299 	 * pair of queues from the syncq.
1300 	 * NOTE: Calling a function inside an assert is unconventional.
1301 	 * However, it does not cause any problem since flush_syncq() does
1302 	 * not change any state except when it returns non-zero i.e.
1303 	 * when the assert will trigger.
1304 	 */
1305 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1306 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1307 	ASSERT((qp->q_flag & QPERMOD) ||
1308 	    ((qp->q_syncq->sq_head == NULL) &&
1309 	    (wqp->q_syncq->sq_head == NULL)));
1310 
1311 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1312 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1313 	if (qp->q_fp != NULL)
1314 		fmodsw_rele(qp->q_fp);
1315 
1316 	/* freeq removes us from the outer perimeter if any */
1317 	freeq(qp);
1318 }
1319 
1320 /* Prevent service procedures from being called */
1321 void
1322 disable_svc(queue_t *qp)
1323 {
1324 	queue_t *wqp = _WR(qp);
1325 
1326 	ASSERT(qp->q_flag & QREADR);
1327 	mutex_enter(QLOCK(qp));
1328 	qp->q_flag |= QWCLOSE;
1329 	mutex_exit(QLOCK(qp));
1330 	mutex_enter(QLOCK(wqp));
1331 	wqp->q_flag |= QWCLOSE;
1332 	mutex_exit(QLOCK(wqp));
1333 }
1334 
1335 /* Allow service procedures to be called again */
1336 void
1337 enable_svc(queue_t *qp)
1338 {
1339 	queue_t *wqp = _WR(qp);
1340 
1341 	ASSERT(qp->q_flag & QREADR);
1342 	mutex_enter(QLOCK(qp));
1343 	qp->q_flag &= ~QWCLOSE;
1344 	mutex_exit(QLOCK(qp));
1345 	mutex_enter(QLOCK(wqp));
1346 	wqp->q_flag &= ~QWCLOSE;
1347 	mutex_exit(QLOCK(wqp));
1348 }
1349 
1350 /*
1351  * Remove queue from qhead/qtail if it is enabled.
1352  * Only reset QENAB if the queue was removed from the runlist.
1353  * A queue goes through 3 stages:
1354  *	It is on the service list and QENAB is set.
1355  *	It is removed from the service list but QENAB is still set.
1356  *	QENAB gets changed to QINSERVICE.
1357  *	QINSERVICE is reset (when the service procedure is done)
1358  * Thus we can not reset QENAB unless we actually removed it from the service
1359  * queue.
1360  */
1361 void
1362 remove_runlist(queue_t *qp)
1363 {
1364 	if (qp->q_flag & QENAB && qhead != NULL) {
1365 		queue_t *q_chase;
1366 		queue_t *q_curr;
1367 		int removed;
1368 
1369 		mutex_enter(&service_queue);
1370 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1371 		mutex_exit(&service_queue);
1372 		if (removed) {
1373 			STRSTAT(qremoved);
1374 			qp->q_flag &= ~QENAB;
1375 		}
1376 	}
1377 }
1378 
1379 
1380 /*
1381  * Wait for any pending service processing to complete.
1382  * The removal of queues from the runlist is not atomic with the
1383  * clearing of the QENABLED flag and setting the INSERVICE flag.
1384  * consequently it is possible for remove_runlist in strclose
1385  * to not find the queue on the runlist but for it to be QENABLED
1386  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1387  * as well as INSERVICE.
1388  */
1389 void
1390 wait_svc(queue_t *qp)
1391 {
1392 	queue_t *wqp = _WR(qp);
1393 
1394 	ASSERT(qp->q_flag & QREADR);
1395 
1396 	/*
1397 	 * Try to remove queues from qhead/qtail list.
1398 	 */
1399 	if (qhead != NULL) {
1400 		remove_runlist(qp);
1401 		remove_runlist(wqp);
1402 	}
1403 	/*
1404 	 * Wait till the syncqs associated with the queue disappear from the
1405 	 * background processing list.
1406 	 * This only needs to be done for non-PERMOD perimeters since
1407 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1408 	 * when the last module/driver is unloaded.
1409 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1410 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1411 	 * functions remove the queue from its syncq list, so sqthread will not
1412 	 * try to access the queue.
1413 	 */
1414 	if (!(qp->q_flag & QPERMOD)) {
1415 		syncq_t *rsq = qp->q_syncq;
1416 		syncq_t *wsq = wqp->q_syncq;
1417 
1418 		/*
1419 		 * Disable rsq and wsq and wait for any background processing of
1420 		 * syncq to complete.
1421 		 */
1422 		wait_sq_svc(rsq);
1423 		if (wsq != rsq)
1424 			wait_sq_svc(wsq);
1425 	}
1426 
1427 	mutex_enter(QLOCK(qp));
1428 	while (qp->q_flag & (QINSERVICE|QENAB))
1429 		cv_wait(&qp->q_wait, QLOCK(qp));
1430 	mutex_exit(QLOCK(qp));
1431 	mutex_enter(QLOCK(wqp));
1432 	while (wqp->q_flag & (QINSERVICE|QENAB))
1433 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1434 	mutex_exit(QLOCK(wqp));
1435 }
1436 
1437 /*
1438  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1439  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1440  * also be set, and is passed through to allocb_cred_wait().
1441  *
1442  * Returns errno on failure, zero on success.
1443  */
1444 int
1445 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1446 {
1447 	mblk_t *tmp;
1448 	ssize_t  count;
1449 	int error = 0;
1450 
1451 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1452 	    (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1453 
1454 	if (bp->b_datap->db_type == M_IOCTL) {
1455 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1456 	} else {
1457 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1458 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1459 	}
1460 	/*
1461 	 * strdoioctl validates ioc_count, so if this assert fails it
1462 	 * cannot be due to user error.
1463 	 */
1464 	ASSERT(count >= 0);
1465 
1466 	if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1467 	    curproc->p_pid)) == NULL) {
1468 		return (error);
1469 	}
1470 	error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1471 	if (error != 0) {
1472 		freeb(tmp);
1473 		return (error);
1474 	}
1475 	DB_CPID(tmp) = curproc->p_pid;
1476 	tmp->b_wptr += count;
1477 	bp->b_cont = tmp;
1478 
1479 	return (0);
1480 }
1481 
1482 /*
1483  * Copy ioctl data to user-land. Return non-zero errno on failure,
1484  * 0 for success.
1485  */
1486 int
1487 getiocd(mblk_t *bp, char *arg, int copymode)
1488 {
1489 	ssize_t count;
1490 	size_t  n;
1491 	int	error;
1492 
1493 	if (bp->b_datap->db_type == M_IOCACK)
1494 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1495 	else {
1496 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1497 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1498 	}
1499 	ASSERT(count >= 0);
1500 
1501 	for (bp = bp->b_cont; bp && count;
1502 	    count -= n, bp = bp->b_cont, arg += n) {
1503 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1504 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1505 		if (error)
1506 			return (error);
1507 	}
1508 	ASSERT(count == 0);
1509 	return (0);
1510 }
1511 
1512 /*
1513  * Allocate a linkinfo entry given the write queue of the
1514  * bottom module of the top stream and the write queue of the
1515  * stream head of the bottom stream.
1516  */
1517 linkinfo_t *
1518 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1519 {
1520 	linkinfo_t *linkp;
1521 
1522 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1523 
1524 	linkp->li_lblk.l_qtop = qup;
1525 	linkp->li_lblk.l_qbot = qdown;
1526 	linkp->li_fpdown = fpdown;
1527 
1528 	mutex_enter(&strresources);
1529 	linkp->li_next = linkinfo_list;
1530 	linkp->li_prev = NULL;
1531 	if (linkp->li_next)
1532 		linkp->li_next->li_prev = linkp;
1533 	linkinfo_list = linkp;
1534 	linkp->li_lblk.l_index = ++lnk_id;
1535 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1536 	mutex_exit(&strresources);
1537 
1538 	return (linkp);
1539 }
1540 
1541 /*
1542  * Free a linkinfo entry.
1543  */
1544 void
1545 lbfree(linkinfo_t *linkp)
1546 {
1547 	mutex_enter(&strresources);
1548 	if (linkp->li_next)
1549 		linkp->li_next->li_prev = linkp->li_prev;
1550 	if (linkp->li_prev)
1551 		linkp->li_prev->li_next = linkp->li_next;
1552 	else
1553 		linkinfo_list = linkp->li_next;
1554 	mutex_exit(&strresources);
1555 
1556 	kmem_cache_free(linkinfo_cache, linkp);
1557 }
1558 
1559 /*
1560  * Check for a potential linking cycle.
1561  * Return 1 if a link will result in a cycle,
1562  * and 0 otherwise.
1563  */
1564 int
1565 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1566 {
1567 	struct mux_node *np;
1568 	struct mux_edge *ep;
1569 	int i;
1570 	major_t lomaj;
1571 	major_t upmaj;
1572 	/*
1573 	 * if the lower stream is a pipe/FIFO, return, since link
1574 	 * cycles can not happen on pipes/FIFOs
1575 	 */
1576 	if (lostp->sd_vnode->v_type == VFIFO)
1577 		return (0);
1578 
1579 	for (i = 0; i < ss->ss_devcnt; i++) {
1580 		np = &ss->ss_mux_nodes[i];
1581 		MUX_CLEAR(np);
1582 	}
1583 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1584 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1585 	np = &ss->ss_mux_nodes[lomaj];
1586 	for (;;) {
1587 		if (!MUX_DIDVISIT(np)) {
1588 			if (np->mn_imaj == upmaj)
1589 				return (1);
1590 			if (np->mn_outp == NULL) {
1591 				MUX_VISIT(np);
1592 				if (np->mn_originp == NULL)
1593 					return (0);
1594 				np = np->mn_originp;
1595 				continue;
1596 			}
1597 			MUX_VISIT(np);
1598 			np->mn_startp = np->mn_outp;
1599 		} else {
1600 			if (np->mn_startp == NULL) {
1601 				if (np->mn_originp == NULL)
1602 					return (0);
1603 				else {
1604 					np = np->mn_originp;
1605 					continue;
1606 				}
1607 			}
1608 			/*
1609 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1610 			 * ignore the edge and move on. ep->me_nodep gets
1611 			 * set to NULL in mux_addedge() if it is a FIFO.
1612 			 *
1613 			 */
1614 			ep = np->mn_startp;
1615 			np->mn_startp = ep->me_nextp;
1616 			if (ep->me_nodep == NULL)
1617 				continue;
1618 			ep->me_nodep->mn_originp = np;
1619 			np = ep->me_nodep;
1620 		}
1621 	}
1622 }
1623 
1624 /*
1625  * Find linkinfo entry corresponding to the parameters.
1626  */
1627 linkinfo_t *
1628 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1629 {
1630 	linkinfo_t *linkp;
1631 	struct mux_edge *mep;
1632 	struct mux_node *mnp;
1633 	queue_t *qup;
1634 
1635 	mutex_enter(&strresources);
1636 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1637 		qup = getendq(stp->sd_wrq);
1638 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1639 			if ((qup == linkp->li_lblk.l_qtop) &&
1640 			    (!index || (index == linkp->li_lblk.l_index))) {
1641 				mutex_exit(&strresources);
1642 				return (linkp);
1643 			}
1644 		}
1645 	} else {
1646 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1647 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1648 		mep = mnp->mn_outp;
1649 		while (mep) {
1650 			if ((index == 0) || (index == mep->me_muxid))
1651 				break;
1652 			mep = mep->me_nextp;
1653 		}
1654 		if (!mep) {
1655 			mutex_exit(&strresources);
1656 			return (NULL);
1657 		}
1658 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1659 			if ((!linkp->li_lblk.l_qtop) &&
1660 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1661 				mutex_exit(&strresources);
1662 				return (linkp);
1663 			}
1664 		}
1665 	}
1666 	mutex_exit(&strresources);
1667 	return (NULL);
1668 }
1669 
1670 /*
1671  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1672  * last queue on the chain and return it.
1673  */
1674 queue_t *
1675 getendq(queue_t *q)
1676 {
1677 	ASSERT(q != NULL);
1678 	while (_SAMESTR(q))
1679 		q = q->q_next;
1680 	return (q);
1681 }
1682 
1683 /*
1684  * Wait for the syncq count to drop to zero.
1685  * sq could be either outer or inner.
1686  */
1687 
1688 static void
1689 wait_syncq(syncq_t *sq)
1690 {
1691 	uint16_t count;
1692 
1693 	mutex_enter(SQLOCK(sq));
1694 	count = sq->sq_count;
1695 	SQ_PUTLOCKS_ENTER(sq);
1696 	SUM_SQ_PUTCOUNTS(sq, count);
1697 	while (count != 0) {
1698 		sq->sq_flags |= SQ_WANTWAKEUP;
1699 		SQ_PUTLOCKS_EXIT(sq);
1700 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1701 		count = sq->sq_count;
1702 		SQ_PUTLOCKS_ENTER(sq);
1703 		SUM_SQ_PUTCOUNTS(sq, count);
1704 	}
1705 	SQ_PUTLOCKS_EXIT(sq);
1706 	mutex_exit(SQLOCK(sq));
1707 }
1708 
1709 /*
1710  * Wait while there are any messages for the queue in its syncq.
1711  */
1712 static void
1713 wait_q_syncq(queue_t *q)
1714 {
1715 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1716 		syncq_t *sq = q->q_syncq;
1717 
1718 		mutex_enter(SQLOCK(sq));
1719 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1720 			sq->sq_flags |= SQ_WANTWAKEUP;
1721 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1722 		}
1723 		mutex_exit(SQLOCK(sq));
1724 	}
1725 }
1726 
1727 
1728 int
1729 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1730     int lhlink)
1731 {
1732 	struct stdata *stp;
1733 	struct strioctl strioc;
1734 	struct linkinfo *linkp;
1735 	struct stdata *stpdown;
1736 	struct streamtab *str;
1737 	queue_t *passq;
1738 	syncq_t *passyncq;
1739 	queue_t *rq;
1740 	cdevsw_impl_t *dp;
1741 	uint32_t qflag;
1742 	uint32_t sqtype;
1743 	perdm_t *dmp;
1744 	int error = 0;
1745 	netstack_t *ns;
1746 	str_stack_t *ss;
1747 
1748 	stp = vp->v_stream;
1749 	TRACE_1(TR_FAC_STREAMS_FR,
1750 	    TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1751 	/*
1752 	 * Test for invalid upper stream
1753 	 */
1754 	if (stp->sd_flag & STRHUP) {
1755 		return (ENXIO);
1756 	}
1757 	if (vp->v_type == VFIFO) {
1758 		return (EINVAL);
1759 	}
1760 	if (stp->sd_strtab == NULL) {
1761 		return (EINVAL);
1762 	}
1763 	if (!stp->sd_strtab->st_muxwinit) {
1764 		return (EINVAL);
1765 	}
1766 	if (fpdown == NULL) {
1767 		return (EBADF);
1768 	}
1769 	ns = netstack_find_by_cred(crp);
1770 	ASSERT(ns != NULL);
1771 	ss = ns->netstack_str;
1772 	ASSERT(ss != NULL);
1773 
1774 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1775 		netstack_rele(ss->ss_netstack);
1776 		return (EINVAL);
1777 	}
1778 	mutex_enter(&muxifier);
1779 	if (stp->sd_flag & STPLEX) {
1780 		mutex_exit(&muxifier);
1781 		netstack_rele(ss->ss_netstack);
1782 		return (ENXIO);
1783 	}
1784 
1785 	/*
1786 	 * Test for invalid lower stream.
1787 	 * The check for the v_type != VFIFO and having a major
1788 	 * number not >= devcnt is done to avoid problems with
1789 	 * adding mux_node entry past the end of mux_nodes[].
1790 	 * For FIFO's we don't add an entry so this isn't a
1791 	 * problem.
1792 	 */
1793 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1794 	    (stpdown == stp) || (stpdown->sd_flag &
1795 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1796 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1797 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1798 	    linkcycle(stp, stpdown, ss)) {
1799 		mutex_exit(&muxifier);
1800 		netstack_rele(ss->ss_netstack);
1801 		return (EINVAL);
1802 	}
1803 	TRACE_1(TR_FAC_STREAMS_FR,
1804 	    TR_STPDOWN, "stpdown:%p", stpdown);
1805 	rq = getendq(stp->sd_wrq);
1806 	if (cmd == I_PLINK)
1807 		rq = NULL;
1808 
1809 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1810 
1811 	strioc.ic_cmd = cmd;
1812 	strioc.ic_timout = INFTIM;
1813 	strioc.ic_len = sizeof (struct linkblk);
1814 	strioc.ic_dp = (char *)&linkp->li_lblk;
1815 
1816 	/*
1817 	 * STRPLUMB protects plumbing changes and should be set before
1818 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1819 	 * and cleared in the end of mlink when passthru queue is removed.
1820 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1821 	 * queue is in-place (it is not a proper module and doesn't have open
1822 	 * entry point).
1823 	 *
1824 	 * STPLEX prevents any threads from entering the stream from above. It
1825 	 * can't be set before the call to link_addpassthru() because putnext
1826 	 * from below may cause stream head I/O routines to be called and these
1827 	 * routines assert that STPLEX is not set. After link_addpassthru()
1828 	 * nothing may come from below since the pass queue syncq is blocked.
1829 	 * Note also that STPLEX should be cleared before the call to
1830 	 * link_rempassthru() since when messages start flowing to the stream
1831 	 * head (e.g. because of message propagation from the pass queue) stream
1832 	 * head I/O routines may be called with STPLEX flag set.
1833 	 *
1834 	 * When STPLEX is set, nothing may come into the stream from above and
1835 	 * it is safe to do a setq which will change stream head. So, the
1836 	 * correct sequence of actions is:
1837 	 *
1838 	 * 1) Set STRPLUMB
1839 	 * 2) Call link_addpassthru()
1840 	 * 3) Set STPLEX
1841 	 * 4) Call setq and update the stream state
1842 	 * 5) Clear STPLEX
1843 	 * 6) Call link_rempassthru()
1844 	 * 7) Clear STRPLUMB
1845 	 *
1846 	 * The same sequence applies to munlink() code.
1847 	 */
1848 	mutex_enter(&stpdown->sd_lock);
1849 	stpdown->sd_flag |= STRPLUMB;
1850 	mutex_exit(&stpdown->sd_lock);
1851 	/*
1852 	 * Add passthru queue below lower mux. This will block
1853 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1854 	 */
1855 	passq = link_addpassthru(stpdown);
1856 
1857 	mutex_enter(&stpdown->sd_lock);
1858 	stpdown->sd_flag |= STPLEX;
1859 	mutex_exit(&stpdown->sd_lock);
1860 
1861 	rq = _RD(stpdown->sd_wrq);
1862 	/*
1863 	 * There may be messages in the streamhead's syncq due to messages
1864 	 * that arrived before link_addpassthru() was done. To avoid
1865 	 * background processing of the syncq happening simultaneous with
1866 	 * setq processing, we disable the streamhead syncq and wait until
1867 	 * existing background thread finishes working on it.
1868 	 */
1869 	wait_sq_svc(rq->q_syncq);
1870 	passyncq = passq->q_syncq;
1871 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1872 		blocksq(passyncq, SQ_BLOCKED, 0);
1873 
1874 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1875 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1876 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1877 
1878 	/* setq might sleep in allocator - avoid holding locks. */
1879 	/* Note: we are holding muxifier here. */
1880 
1881 	str = stp->sd_strtab;
1882 	dp = &devimpl[getmajor(vp->v_rdev)];
1883 	ASSERT(dp->d_str == str);
1884 
1885 	qflag = dp->d_qflag;
1886 	sqtype = dp->d_sqtype;
1887 
1888 	/* create perdm_t if needed */
1889 	if (NEED_DM(dp->d_dmp, qflag))
1890 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1891 
1892 	dmp = dp->d_dmp;
1893 
1894 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1895 	    B_TRUE);
1896 
1897 	/*
1898 	 * XXX Remove any "odd" messages from the queue.
1899 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1900 	 */
1901 	error = strdoioctl(stp, &strioc, FNATIVE,
1902 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1903 	if (error != 0) {
1904 		lbfree(linkp);
1905 
1906 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1907 			blocksq(passyncq, SQ_BLOCKED, 0);
1908 		/*
1909 		 * Restore the stream head queue and then remove
1910 		 * the passq. Turn off STPLEX before we turn on
1911 		 * the stream by removing the passq.
1912 		 */
1913 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1914 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1915 		    B_TRUE);
1916 
1917 		mutex_enter(&stpdown->sd_lock);
1918 		stpdown->sd_flag &= ~STPLEX;
1919 		mutex_exit(&stpdown->sd_lock);
1920 
1921 		link_rempassthru(passq);
1922 
1923 		mutex_enter(&stpdown->sd_lock);
1924 		stpdown->sd_flag &= ~STRPLUMB;
1925 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1926 		cv_broadcast(&stpdown->sd_monitor);
1927 		mutex_exit(&stpdown->sd_lock);
1928 
1929 		mutex_exit(&muxifier);
1930 		netstack_rele(ss->ss_netstack);
1931 		return (error);
1932 	}
1933 	mutex_enter(&fpdown->f_tlock);
1934 	fpdown->f_count++;
1935 	mutex_exit(&fpdown->f_tlock);
1936 
1937 	/*
1938 	 * if we've made it here the linkage is all set up so we should also
1939 	 * set up the layered driver linkages
1940 	 */
1941 
1942 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1943 	if (cmd == I_LINK) {
1944 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1945 	} else {
1946 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1947 	}
1948 
1949 	link_rempassthru(passq);
1950 
1951 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1952 
1953 	/*
1954 	 * Mark the upper stream as having dependent links
1955 	 * so that strclose can clean it up.
1956 	 */
1957 	if (cmd == I_LINK) {
1958 		mutex_enter(&stp->sd_lock);
1959 		stp->sd_flag |= STRHASLINKS;
1960 		mutex_exit(&stp->sd_lock);
1961 	}
1962 	/*
1963 	 * Wake up any other processes that may have been
1964 	 * waiting on the lower stream. These will all
1965 	 * error out.
1966 	 */
1967 	mutex_enter(&stpdown->sd_lock);
1968 	/* The passthru module is removed so we may release STRPLUMB */
1969 	stpdown->sd_flag &= ~STRPLUMB;
1970 	cv_broadcast(&rq->q_wait);
1971 	cv_broadcast(&_WR(rq)->q_wait);
1972 	cv_broadcast(&stpdown->sd_monitor);
1973 	mutex_exit(&stpdown->sd_lock);
1974 	mutex_exit(&muxifier);
1975 	*rvalp = linkp->li_lblk.l_index;
1976 	netstack_rele(ss->ss_netstack);
1977 	return (0);
1978 }
1979 
1980 int
1981 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1982 {
1983 	int		ret;
1984 	struct file	*fpdown;
1985 
1986 	fpdown = getf(arg);
1987 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1988 	if (fpdown != NULL)
1989 		releasef(arg);
1990 	return (ret);
1991 }
1992 
1993 /*
1994  * Unlink a multiplexor link. Stp is the controlling stream for the
1995  * link, and linkp points to the link's entry in the linkinfo list.
1996  * The muxifier lock must be held on entry and is dropped on exit.
1997  *
1998  * NOTE : Currently it is assumed that mux would process all the messages
1999  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
2000  * of the mux to handle all the messages that arrive before UNLINK.
2001  * If the mux has to send down messages on its lower stream before
2002  * ACKing I_UNLINK, then it *should* know to handle messages even
2003  * after the UNLINK is acked (actually it should be able to handle till we
2004  * re-block the read side of the pass queue here). If the mux does not
2005  * open up the lower stream, any messages that arrive during UNLINK
2006  * will be put in the stream head. In the case of lower stream opening
2007  * up, some messages might land in the stream head depending on when
2008  * the message arrived and when the read side of the pass queue was
2009  * re-blocked.
2010  */
2011 int
2012 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2013     str_stack_t *ss)
2014 {
2015 	struct strioctl strioc;
2016 	struct stdata *stpdown;
2017 	queue_t *rq, *wrq;
2018 	queue_t	*passq;
2019 	syncq_t *passyncq;
2020 	int error = 0;
2021 	file_t *fpdown;
2022 
2023 	ASSERT(MUTEX_HELD(&muxifier));
2024 
2025 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2026 
2027 	/*
2028 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2029 	 */
2030 	mutex_enter(&stpdown->sd_lock);
2031 	stpdown->sd_flag |= STRPLUMB;
2032 	mutex_exit(&stpdown->sd_lock);
2033 
2034 	/*
2035 	 * Add passthru queue below lower mux. This will block
2036 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2037 	 */
2038 	passq = link_addpassthru(stpdown);
2039 
2040 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2041 		strioc.ic_cmd = I_UNLINK;
2042 	else
2043 		strioc.ic_cmd = I_PUNLINK;
2044 	strioc.ic_timout = INFTIM;
2045 	strioc.ic_len = sizeof (struct linkblk);
2046 	strioc.ic_dp = (char *)&linkp->li_lblk;
2047 
2048 	error = strdoioctl(stp, &strioc, FNATIVE,
2049 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2050 
2051 	/*
2052 	 * If there was an error and this is not called via strclose,
2053 	 * return to the user. Otherwise, pretend there was no error
2054 	 * and close the link.
2055 	 */
2056 	if (error) {
2057 		if (flag & LINKCLOSE) {
2058 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2059 			    "unlink ioctl, closing anyway (%d)\n", error);
2060 		} else {
2061 			link_rempassthru(passq);
2062 			mutex_enter(&stpdown->sd_lock);
2063 			stpdown->sd_flag &= ~STRPLUMB;
2064 			cv_broadcast(&stpdown->sd_monitor);
2065 			mutex_exit(&stpdown->sd_lock);
2066 			mutex_exit(&muxifier);
2067 			return (error);
2068 		}
2069 	}
2070 
2071 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2072 	fpdown = linkp->li_fpdown;
2073 	lbfree(linkp);
2074 
2075 	/*
2076 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2077 	 * can slow others down. It's okay to since attempts to mlink() this
2078 	 * stream will be stopped because STPLEX is still set in the stdata
2079 	 * structure, and munlink() is stopped because mux_rmvedge() and
2080 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2081 	 * respectively.  Note that we defer the closef() of fpdown until
2082 	 * after we drop muxifier since strclose() can call munlinkall().
2083 	 */
2084 	mutex_exit(&muxifier);
2085 
2086 	wrq = stpdown->sd_wrq;
2087 	rq = _RD(wrq);
2088 
2089 	/*
2090 	 * Get rid of outstanding service procedure runs, before we make
2091 	 * it a stream head, since a stream head doesn't have any service
2092 	 * procedure.
2093 	 */
2094 	disable_svc(rq);
2095 	wait_svc(rq);
2096 
2097 	/*
2098 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2099 	 * is queued up to be finished. mux should take care that nothing is
2100 	 * send down to this queue. We should do it now as we're going to block
2101 	 * passyncq if it was unblocked.
2102 	 */
2103 	if (wrq->q_flag & QPERMOD) {
2104 		syncq_t	*sq = wrq->q_syncq;
2105 
2106 		mutex_enter(SQLOCK(sq));
2107 		while (wrq->q_sqflags & Q_SQQUEUED) {
2108 			sq->sq_flags |= SQ_WANTWAKEUP;
2109 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2110 		}
2111 		mutex_exit(SQLOCK(sq));
2112 	}
2113 	passyncq = passq->q_syncq;
2114 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2115 
2116 		syncq_t *sq, *outer;
2117 
2118 		/*
2119 		 * Messages could be flowing from underneath. We will
2120 		 * block the read side of the passq. This would be
2121 		 * sufficient for QPAIR and QPERQ muxes to ensure
2122 		 * that no data is flowing up into this queue
2123 		 * and hence no thread active in this instance of
2124 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2125 		 * could be messages on the inner and outer/inner
2126 		 * syncqs respectively. We will wait for them to drain.
2127 		 * Because passq is blocked messages end up in the syncq
2128 		 * And qfill_syncq could possibly end up setting QFULL
2129 		 * which will access the rq->q_flag. Hence, we have to
2130 		 * acquire the QLOCK in setq.
2131 		 *
2132 		 * XXX Messages can also flow from top into this
2133 		 * queue though the unlink is over (Ex. some instance
2134 		 * in putnext() called from top that has still not
2135 		 * accessed this queue. And also putq(lowerq) ?).
2136 		 * Solution : How about blocking the l_qtop queue ?
2137 		 * Do we really care about such pure D_MP muxes ?
2138 		 */
2139 
2140 		blocksq(passyncq, SQ_BLOCKED, 0);
2141 
2142 		sq = rq->q_syncq;
2143 		if ((outer = sq->sq_outer) != NULL) {
2144 
2145 			/*
2146 			 * We have to just wait for the outer sq_count
2147 			 * drop to zero. As this does not prevent new
2148 			 * messages to enter the outer perimeter, this
2149 			 * is subject to starvation.
2150 			 *
2151 			 * NOTE :Because of blocksq above, messages could
2152 			 * be in the inner syncq only because of some
2153 			 * thread holding the outer perimeter exclusively.
2154 			 * Hence it would be sufficient to wait for the
2155 			 * exclusive holder of the outer perimeter to drain
2156 			 * the inner and outer syncqs. But we will not depend
2157 			 * on this feature and hence check the inner syncqs
2158 			 * separately.
2159 			 */
2160 			wait_syncq(outer);
2161 		}
2162 
2163 
2164 		/*
2165 		 * There could be messages destined for
2166 		 * this queue. Let the exclusive holder
2167 		 * drain it.
2168 		 */
2169 
2170 		wait_syncq(sq);
2171 		ASSERT((rq->q_flag & QPERMOD) ||
2172 		    ((rq->q_syncq->sq_head == NULL) &&
2173 		    (_WR(rq)->q_syncq->sq_head == NULL)));
2174 	}
2175 
2176 	/*
2177 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2178 	 * case as we don't disable its syncq or remove it off the syncq
2179 	 * service list.
2180 	 */
2181 	if (rq->q_flag & QPERMOD) {
2182 		syncq_t	*sq = rq->q_syncq;
2183 
2184 		mutex_enter(SQLOCK(sq));
2185 		while (rq->q_sqflags & Q_SQQUEUED) {
2186 			sq->sq_flags |= SQ_WANTWAKEUP;
2187 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2188 		}
2189 		mutex_exit(SQLOCK(sq));
2190 	}
2191 
2192 	/*
2193 	 * flush_syncq changes states only when there are some messages to
2194 	 * free, i.e. when it returns non-zero value to return.
2195 	 */
2196 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2197 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2198 
2199 	/*
2200 	 * Nobody else should know about this queue now.
2201 	 * If the mux did not process the messages before
2202 	 * acking the I_UNLINK, free them now.
2203 	 */
2204 
2205 	flushq(rq, FLUSHALL);
2206 	flushq(_WR(rq), FLUSHALL);
2207 
2208 	/*
2209 	 * Convert the mux lower queue into a stream head queue.
2210 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2211 	 */
2212 	rq->q_ptr = wrq->q_ptr = stpdown;
2213 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2214 
2215 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2216 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2217 
2218 	enable_svc(rq);
2219 
2220 	/*
2221 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2222 	 * needs to be set to prevent reopen() of the stream - such reopen may
2223 	 * try to call non-existent pass queue open routine and panic.
2224 	 */
2225 	mutex_enter(&stpdown->sd_lock);
2226 	stpdown->sd_flag &= ~STPLEX;
2227 	mutex_exit(&stpdown->sd_lock);
2228 
2229 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2230 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2231 
2232 	/* clean up the layered driver linkages */
2233 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2234 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2235 	} else {
2236 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2237 	}
2238 
2239 	link_rempassthru(passq);
2240 
2241 	/*
2242 	 * Now all plumbing changes are finished and STRPLUMB is no
2243 	 * longer needed.
2244 	 */
2245 	mutex_enter(&stpdown->sd_lock);
2246 	stpdown->sd_flag &= ~STRPLUMB;
2247 	cv_broadcast(&stpdown->sd_monitor);
2248 	mutex_exit(&stpdown->sd_lock);
2249 
2250 	(void) closef(fpdown);
2251 	return (0);
2252 }
2253 
2254 /*
2255  * Unlink all multiplexor links for which stp is the controlling stream.
2256  * Return 0, or a non-zero errno on failure.
2257  */
2258 int
2259 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2260 {
2261 	linkinfo_t *linkp;
2262 	int error = 0;
2263 
2264 	mutex_enter(&muxifier);
2265 	while (linkp = findlinks(stp, 0, flag, ss)) {
2266 		/*
2267 		 * munlink() releases the muxifier lock.
2268 		 */
2269 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2270 			return (error);
2271 		mutex_enter(&muxifier);
2272 	}
2273 	mutex_exit(&muxifier);
2274 	return (0);
2275 }
2276 
2277 /*
2278  * A multiplexor link has been made. Add an
2279  * edge to the directed graph.
2280  */
2281 void
2282 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2283 {
2284 	struct mux_node *np;
2285 	struct mux_edge *ep;
2286 	major_t upmaj;
2287 	major_t lomaj;
2288 
2289 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2290 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2291 	np = &ss->ss_mux_nodes[upmaj];
2292 	if (np->mn_outp) {
2293 		ep = np->mn_outp;
2294 		while (ep->me_nextp)
2295 			ep = ep->me_nextp;
2296 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2297 		ep = ep->me_nextp;
2298 	} else {
2299 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2300 		ep = np->mn_outp;
2301 	}
2302 	ep->me_nextp = NULL;
2303 	ep->me_muxid = muxid;
2304 	/*
2305 	 * Save the dev_t for the purposes of str_stack_shutdown.
2306 	 * str_stack_shutdown assumes that the device allows reopen, since
2307 	 * this dev_t is the one after any cloning by xx_open().
2308 	 * Would prefer finding the dev_t from before any cloning,
2309 	 * but specfs doesn't retain that.
2310 	 */
2311 	ep->me_dev = upstp->sd_vnode->v_rdev;
2312 	if (lostp->sd_vnode->v_type == VFIFO)
2313 		ep->me_nodep = NULL;
2314 	else
2315 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2316 }
2317 
2318 /*
2319  * A multiplexor link has been removed. Remove the
2320  * edge in the directed graph.
2321  */
2322 void
2323 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2324 {
2325 	struct mux_node *np;
2326 	struct mux_edge *ep;
2327 	struct mux_edge *pep = NULL;
2328 	major_t upmaj;
2329 
2330 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2331 	np = &ss->ss_mux_nodes[upmaj];
2332 	ASSERT(np->mn_outp != NULL);
2333 	ep = np->mn_outp;
2334 	while (ep) {
2335 		if (ep->me_muxid == muxid) {
2336 			if (pep)
2337 				pep->me_nextp = ep->me_nextp;
2338 			else
2339 				np->mn_outp = ep->me_nextp;
2340 			kmem_free(ep, sizeof (struct mux_edge));
2341 			return;
2342 		}
2343 		pep = ep;
2344 		ep = ep->me_nextp;
2345 	}
2346 	ASSERT(0);	/* should not reach here */
2347 }
2348 
2349 /*
2350  * Translate the device flags (from conf.h) to the corresponding
2351  * qflag and sq_flag (type) values.
2352  */
2353 int
2354 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2355     uint32_t *sqtypep)
2356 {
2357 	uint32_t qflag = 0;
2358 	uint32_t sqtype = 0;
2359 
2360 	if (devflag & _D_OLD)
2361 		goto bad;
2362 
2363 	/* Inner perimeter presence and scope */
2364 	switch (devflag & D_MTINNER_MASK) {
2365 	case D_MP:
2366 		qflag |= QMTSAFE;
2367 		sqtype |= SQ_CI;
2368 		break;
2369 	case D_MTPERQ|D_MP:
2370 		qflag |= QPERQ;
2371 		break;
2372 	case D_MTQPAIR|D_MP:
2373 		qflag |= QPAIR;
2374 		break;
2375 	case D_MTPERMOD|D_MP:
2376 		qflag |= QPERMOD;
2377 		break;
2378 	default:
2379 		goto bad;
2380 	}
2381 
2382 	/* Outer perimeter */
2383 	if (devflag & D_MTOUTPERIM) {
2384 		switch (devflag & D_MTINNER_MASK) {
2385 		case D_MP:
2386 		case D_MTPERQ|D_MP:
2387 		case D_MTQPAIR|D_MP:
2388 			break;
2389 		default:
2390 			goto bad;
2391 		}
2392 		qflag |= QMTOUTPERIM;
2393 	}
2394 
2395 	/* Inner perimeter modifiers */
2396 	if (devflag & D_MTINNER_MOD) {
2397 		switch (devflag & D_MTINNER_MASK) {
2398 		case D_MP:
2399 			goto bad;
2400 		default:
2401 			break;
2402 		}
2403 		if (devflag & D_MTPUTSHARED)
2404 			sqtype |= SQ_CIPUT;
2405 		if (devflag & _D_MTOCSHARED) {
2406 			/*
2407 			 * The code in putnext assumes that it has the
2408 			 * highest concurrency by not checking sq_count.
2409 			 * Thus _D_MTOCSHARED can only be supported when
2410 			 * D_MTPUTSHARED is set.
2411 			 */
2412 			if (!(devflag & D_MTPUTSHARED))
2413 				goto bad;
2414 			sqtype |= SQ_CIOC;
2415 		}
2416 		if (devflag & _D_MTCBSHARED) {
2417 			/*
2418 			 * The code in putnext assumes that it has the
2419 			 * highest concurrency by not checking sq_count.
2420 			 * Thus _D_MTCBSHARED can only be supported when
2421 			 * D_MTPUTSHARED is set.
2422 			 */
2423 			if (!(devflag & D_MTPUTSHARED))
2424 				goto bad;
2425 			sqtype |= SQ_CICB;
2426 		}
2427 		if (devflag & _D_MTSVCSHARED) {
2428 			/*
2429 			 * The code in putnext assumes that it has the
2430 			 * highest concurrency by not checking sq_count.
2431 			 * Thus _D_MTSVCSHARED can only be supported when
2432 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2433 			 * supported only for QPERMOD.
2434 			 */
2435 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2436 				goto bad;
2437 			sqtype |= SQ_CISVC;
2438 		}
2439 	}
2440 
2441 	/* Default outer perimeter concurrency */
2442 	sqtype |= SQ_CO;
2443 
2444 	/* Outer perimeter modifiers */
2445 	if (devflag & D_MTOCEXCL) {
2446 		if (!(devflag & D_MTOUTPERIM)) {
2447 			/* No outer perimeter */
2448 			goto bad;
2449 		}
2450 		sqtype &= ~SQ_COOC;
2451 	}
2452 
2453 	/* Synchronous Streams extended qinit structure */
2454 	if (devflag & D_SYNCSTR)
2455 		qflag |= QSYNCSTR;
2456 
2457 	/*
2458 	 * Private flag used by a transport module to indicate
2459 	 * to sockfs that it supports direct-access mode without
2460 	 * having to go through STREAMS.
2461 	 */
2462 	if (devflag & _D_DIRECT) {
2463 		/* Reject unless the module is fully-MT (no perimeter) */
2464 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2465 			goto bad;
2466 		qflag |= _QDIRECT;
2467 	}
2468 
2469 	/*
2470 	 * Private flag used to indicate that a streams module should only
2471 	 * be pushed once. The TTY streams modules have this flag since if
2472 	 * libc believes itself to be an xpg4 process then it will
2473 	 * automatically and unconditionally push them when a PTS device is
2474 	 * opened. If an application is not aware of this then without this
2475 	 * flag we would end up with duplicate modules.
2476 	 */
2477 	if (devflag & _D_SINGLE_INSTANCE)
2478 		qflag |= _QSINGLE_INSTANCE;
2479 
2480 	*qflagp = qflag;
2481 	*sqtypep = sqtype;
2482 	return (0);
2483 
2484 bad:
2485 	cmn_err(CE_WARN,
2486 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2487 	    (int)(qflag & D_MTSAFETY_MASK),
2488 	    stp->st_rdinit->qi_minfo->mi_idname);
2489 
2490 	return (EINVAL);
2491 }
2492 
2493 /*
2494  * Set the interface values for a pair of queues (qinit structure,
2495  * packet sizes, water marks).
2496  * setq assumes that the caller does not have a claim (entersq or claimq)
2497  * on the queue.
2498  */
2499 void
2500 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2501     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2502 {
2503 	queue_t *wq;
2504 	syncq_t	*sq, *outer;
2505 
2506 	ASSERT(rq->q_flag & QREADR);
2507 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2508 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2509 
2510 	wq = _WR(rq);
2511 	rq->q_qinfo = rinit;
2512 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2513 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2514 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2515 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2516 	wq->q_qinfo = winit;
2517 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2518 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2519 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2520 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2521 
2522 	/* Remove old syncqs */
2523 	sq = rq->q_syncq;
2524 	outer = sq->sq_outer;
2525 	if (outer != NULL) {
2526 		ASSERT(wq->q_syncq->sq_outer == outer);
2527 		outer_remove(outer, rq->q_syncq);
2528 		if (wq->q_syncq != rq->q_syncq)
2529 			outer_remove(outer, wq->q_syncq);
2530 	}
2531 	ASSERT(sq->sq_outer == NULL);
2532 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2533 
2534 	if (sq != SQ(rq)) {
2535 		if (!(rq->q_flag & QPERMOD))
2536 			free_syncq(sq);
2537 		if (wq->q_syncq == rq->q_syncq)
2538 			wq->q_syncq = NULL;
2539 		rq->q_syncq = NULL;
2540 	}
2541 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2542 	    wq->q_syncq != SQ(rq)) {
2543 		free_syncq(wq->q_syncq);
2544 		wq->q_syncq = NULL;
2545 	}
2546 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2547 	    rq->q_syncq->sq_tail == NULL));
2548 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2549 	    wq->q_syncq->sq_tail == NULL));
2550 
2551 	if (!(rq->q_flag & QPERMOD) &&
2552 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2553 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2554 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2555 		    rq->q_syncq->sq_nciputctrl, 0);
2556 		ASSERT(ciputctrl_cache != NULL);
2557 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2558 		rq->q_syncq->sq_ciputctrl = NULL;
2559 		rq->q_syncq->sq_nciputctrl = 0;
2560 	}
2561 
2562 	if (!(wq->q_flag & QPERMOD) &&
2563 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2564 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2565 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2566 		    wq->q_syncq->sq_nciputctrl, 0);
2567 		ASSERT(ciputctrl_cache != NULL);
2568 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2569 		wq->q_syncq->sq_ciputctrl = NULL;
2570 		wq->q_syncq->sq_nciputctrl = 0;
2571 	}
2572 
2573 	sq = SQ(rq);
2574 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2575 	ASSERT(sq->sq_outer == NULL);
2576 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2577 
2578 	/*
2579 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2580 	 * bits in sq_flag based on the sqtype.
2581 	 */
2582 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2583 
2584 	rq->q_syncq = wq->q_syncq = sq;
2585 	sq->sq_type = sqtype;
2586 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2587 
2588 	/*
2589 	 *  We are making sq_svcflags zero,
2590 	 *  resetting SQ_DISABLED in case it was set by
2591 	 *  wait_svc() in the munlink path.
2592 	 *
2593 	 */
2594 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2595 	sq->sq_svcflags = 0;
2596 
2597 	/*
2598 	 * We need to acquire the lock here for the mlink and munlink case,
2599 	 * where canputnext, backenable, etc can access the q_flag.
2600 	 */
2601 	if (lock_needed) {
2602 		mutex_enter(QLOCK(rq));
2603 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2604 		mutex_exit(QLOCK(rq));
2605 		mutex_enter(QLOCK(wq));
2606 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2607 		mutex_exit(QLOCK(wq));
2608 	} else {
2609 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2610 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2611 	}
2612 
2613 	if (qflag & QPERQ) {
2614 		/* Allocate a separate syncq for the write side */
2615 		sq = new_syncq();
2616 		sq->sq_type = rq->q_syncq->sq_type;
2617 		sq->sq_flags = rq->q_syncq->sq_flags;
2618 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2619 		    sq->sq_oprev == NULL);
2620 		wq->q_syncq = sq;
2621 	}
2622 	if (qflag & QPERMOD) {
2623 		sq = dmp->dm_sq;
2624 
2625 		/*
2626 		 * Assert that we do have an inner perimeter syncq and that it
2627 		 * does not have an outer perimeter associated with it.
2628 		 */
2629 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2630 		    sq->sq_oprev == NULL);
2631 		rq->q_syncq = wq->q_syncq = sq;
2632 	}
2633 	if (qflag & QMTOUTPERIM) {
2634 		outer = dmp->dm_sq;
2635 
2636 		ASSERT(outer->sq_outer == NULL);
2637 		outer_insert(outer, rq->q_syncq);
2638 		if (wq->q_syncq != rq->q_syncq)
2639 			outer_insert(outer, wq->q_syncq);
2640 	}
2641 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2642 	    (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2643 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2644 	    (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2645 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2646 
2647 	/*
2648 	 * Initialize struio() types.
2649 	 */
2650 	rq->q_struiot =
2651 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2652 	wq->q_struiot =
2653 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2654 }
2655 
2656 perdm_t *
2657 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2658 {
2659 	syncq_t	*sq;
2660 	perdm_t	**pp;
2661 	perdm_t	*p;
2662 	perdm_t	*dmp;
2663 
2664 	ASSERT(str != NULL);
2665 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2666 
2667 	rw_enter(&perdm_rwlock, RW_READER);
2668 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2669 		if (p->dm_str == str) {	/* found one */
2670 			atomic_inc_32(&(p->dm_ref));
2671 			rw_exit(&perdm_rwlock);
2672 			return (p);
2673 		}
2674 	}
2675 	rw_exit(&perdm_rwlock);
2676 
2677 	sq = new_syncq();
2678 	if (qflag & QPERMOD) {
2679 		sq->sq_type = sqtype | SQ_PERMOD;
2680 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2681 	} else {
2682 		ASSERT(qflag & QMTOUTPERIM);
2683 		sq->sq_onext = sq->sq_oprev = sq;
2684 	}
2685 
2686 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2687 	dmp->dm_sq = sq;
2688 	dmp->dm_str = str;
2689 	dmp->dm_ref = 1;
2690 	dmp->dm_next = NULL;
2691 
2692 	rw_enter(&perdm_rwlock, RW_WRITER);
2693 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2694 		if (p->dm_str == str) {	/* already present */
2695 			p->dm_ref++;
2696 			rw_exit(&perdm_rwlock);
2697 			free_syncq(sq);
2698 			kmem_free(dmp, sizeof (perdm_t));
2699 			return (p);
2700 		}
2701 	}
2702 
2703 	*pp = dmp;
2704 	rw_exit(&perdm_rwlock);
2705 	return (dmp);
2706 }
2707 
2708 void
2709 rele_dm(perdm_t *dmp)
2710 {
2711 	perdm_t **pp;
2712 	perdm_t *p;
2713 
2714 	rw_enter(&perdm_rwlock, RW_WRITER);
2715 	ASSERT(dmp->dm_ref > 0);
2716 
2717 	if (--dmp->dm_ref > 0) {
2718 		rw_exit(&perdm_rwlock);
2719 		return;
2720 	}
2721 
2722 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2723 		if (p == dmp)
2724 			break;
2725 	ASSERT(p == dmp);
2726 	*pp = p->dm_next;
2727 	rw_exit(&perdm_rwlock);
2728 
2729 	/*
2730 	 * Wait for any background processing that relies on the
2731 	 * syncq to complete before it is freed.
2732 	 */
2733 	wait_sq_svc(p->dm_sq);
2734 	free_syncq(p->dm_sq);
2735 	kmem_free(p, sizeof (perdm_t));
2736 }
2737 
2738 /*
2739  * Make a protocol message given control and data buffers.
2740  * n.b., this can block; be careful of what locks you hold when calling it.
2741  *
2742  * If sd_maxblk is less than *iosize this routine can fail part way through
2743  * (due to an allocation failure). In this case on return *iosize will contain
2744  * the amount that was consumed. Otherwise *iosize will not be modified
2745  * i.e. it will contain the amount that was consumed.
2746  */
2747 int
2748 strmakemsg(
2749 	struct strbuf *mctl,
2750 	ssize_t *iosize,
2751 	struct uio *uiop,
2752 	stdata_t *stp,
2753 	int32_t flag,
2754 	mblk_t **mpp)
2755 {
2756 	mblk_t *mpctl = NULL;
2757 	mblk_t *mpdata = NULL;
2758 	int error;
2759 
2760 	ASSERT(uiop != NULL);
2761 
2762 	*mpp = NULL;
2763 	/* Create control part, if any */
2764 	if ((mctl != NULL) && (mctl->len >= 0)) {
2765 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2766 		if (error)
2767 			return (error);
2768 	}
2769 	/* Create data part, if any */
2770 	if (*iosize >= 0) {
2771 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2772 		if (error) {
2773 			freemsg(mpctl);
2774 			return (error);
2775 		}
2776 	}
2777 	if (mpctl != NULL) {
2778 		if (mpdata != NULL)
2779 			linkb(mpctl, mpdata);
2780 		*mpp = mpctl;
2781 	} else {
2782 		*mpp = mpdata;
2783 	}
2784 	return (0);
2785 }
2786 
2787 /*
2788  * Make the control part of a protocol message given a control buffer.
2789  * n.b., this can block; be careful of what locks you hold when calling it.
2790  */
2791 int
2792 strmakectl(
2793 	struct strbuf *mctl,
2794 	int32_t flag,
2795 	int32_t fflag,
2796 	mblk_t **mpp)
2797 {
2798 	mblk_t *bp = NULL;
2799 	unsigned char msgtype;
2800 	int error = 0;
2801 	cred_t *cr = CRED();
2802 
2803 	/* We do not support interrupt threads using the stream head to send */
2804 	ASSERT(cr != NULL);
2805 
2806 	*mpp = NULL;
2807 	/*
2808 	 * Create control part of message, if any.
2809 	 */
2810 	if ((mctl != NULL) && (mctl->len >= 0)) {
2811 		caddr_t base;
2812 		int ctlcount;
2813 		int allocsz;
2814 
2815 		if (flag & RS_HIPRI)
2816 			msgtype = M_PCPROTO;
2817 		else
2818 			msgtype = M_PROTO;
2819 
2820 		ctlcount = mctl->len;
2821 		base = mctl->buf;
2822 
2823 		/*
2824 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2825 		 * blocks by increasing the size to something more usable.
2826 		 */
2827 		allocsz = MAX(ctlcount, 64);
2828 
2829 		/*
2830 		 * Range checking has already been done; simply try
2831 		 * to allocate a message block for the ctl part.
2832 		 */
2833 		while ((bp = allocb_cred(allocsz, cr,
2834 		    curproc->p_pid)) == NULL) {
2835 			if (fflag & (FNDELAY|FNONBLOCK))
2836 				return (EAGAIN);
2837 			if (error = strwaitbuf(allocsz, BPRI_MED))
2838 				return (error);
2839 		}
2840 
2841 		bp->b_datap->db_type = msgtype;
2842 		if (copyin(base, bp->b_wptr, ctlcount)) {
2843 			freeb(bp);
2844 			return (EFAULT);
2845 		}
2846 		bp->b_wptr += ctlcount;
2847 	}
2848 	*mpp = bp;
2849 	return (0);
2850 }
2851 
2852 /*
2853  * Make a protocol message given data buffers.
2854  * n.b., this can block; be careful of what locks you hold when calling it.
2855  *
2856  * If sd_maxblk is less than *iosize this routine can fail part way through
2857  * (due to an allocation failure). In this case on return *iosize will contain
2858  * the amount that was consumed. Otherwise *iosize will not be modified
2859  * i.e. it will contain the amount that was consumed.
2860  */
2861 int
2862 strmakedata(
2863 	ssize_t   *iosize,
2864 	struct uio *uiop,
2865 	stdata_t *stp,
2866 	int32_t flag,
2867 	mblk_t **mpp)
2868 {
2869 	mblk_t *mp = NULL;
2870 	mblk_t *bp;
2871 	int wroff = (int)stp->sd_wroff;
2872 	int tail_len = (int)stp->sd_tail;
2873 	int extra = wroff + tail_len;
2874 	int error = 0;
2875 	ssize_t maxblk;
2876 	ssize_t count = *iosize;
2877 	cred_t *cr;
2878 
2879 	*mpp = NULL;
2880 	if (count < 0)
2881 		return (0);
2882 
2883 	/* We do not support interrupt threads using the stream head to send */
2884 	cr = CRED();
2885 	ASSERT(cr != NULL);
2886 
2887 	maxblk = stp->sd_maxblk;
2888 	if (maxblk == INFPSZ)
2889 		maxblk = count;
2890 
2891 	/*
2892 	 * Create data part of message, if any.
2893 	 */
2894 	do {
2895 		ssize_t size;
2896 		dblk_t  *dp;
2897 
2898 		ASSERT(uiop);
2899 
2900 		size = MIN(count, maxblk);
2901 
2902 		while ((bp = allocb_cred(size + extra, cr,
2903 		    curproc->p_pid)) == NULL) {
2904 			error = EAGAIN;
2905 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2906 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2907 				if (count == *iosize) {
2908 					freemsg(mp);
2909 					return (error);
2910 				} else {
2911 					*iosize -= count;
2912 					*mpp = mp;
2913 					return (0);
2914 				}
2915 			}
2916 		}
2917 		dp = bp->b_datap;
2918 		dp->db_cpid = curproc->p_pid;
2919 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2920 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2921 
2922 		if (flag & STRUIO_POSTPONE) {
2923 			/*
2924 			 * Setup the stream uio portion of the
2925 			 * dblk for subsequent use by struioget().
2926 			 */
2927 			dp->db_struioflag = STRUIO_SPEC;
2928 			dp->db_cksumstart = 0;
2929 			dp->db_cksumstuff = 0;
2930 			dp->db_cksumend = size;
2931 			*(long long *)dp->db_struioun.data = 0ll;
2932 			bp->b_wptr += size;
2933 		} else {
2934 			if (stp->sd_copyflag & STRCOPYCACHED)
2935 				uiop->uio_extflg |= UIO_COPY_CACHED;
2936 
2937 			if (size != 0) {
2938 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2939 				    uiop);
2940 				if (error != 0) {
2941 					freeb(bp);
2942 					freemsg(mp);
2943 					return (error);
2944 				}
2945 			}
2946 			bp->b_wptr += size;
2947 
2948 			if (stp->sd_wputdatafunc != NULL) {
2949 				mblk_t *newbp;
2950 
2951 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2952 				    bp, NULL, NULL, NULL, NULL);
2953 				if (newbp == NULL) {
2954 					freeb(bp);
2955 					freemsg(mp);
2956 					return (ECOMM);
2957 				}
2958 				bp = newbp;
2959 			}
2960 		}
2961 
2962 		count -= size;
2963 
2964 		if (mp == NULL)
2965 			mp = bp;
2966 		else
2967 			linkb(mp, bp);
2968 	} while (count > 0);
2969 
2970 	*mpp = mp;
2971 	return (0);
2972 }
2973 
2974 /*
2975  * Wait for a buffer to become available. Return non-zero errno
2976  * if not able to wait, 0 if buffer is probably there.
2977  */
2978 int
2979 strwaitbuf(size_t size, int pri)
2980 {
2981 	bufcall_id_t id;
2982 
2983 	mutex_enter(&bcall_monitor);
2984 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2985 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2986 		mutex_exit(&bcall_monitor);
2987 		return (ENOSR);
2988 	}
2989 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2990 		unbufcall(id);
2991 		mutex_exit(&bcall_monitor);
2992 		return (EINTR);
2993 	}
2994 	unbufcall(id);
2995 	mutex_exit(&bcall_monitor);
2996 	return (0);
2997 }
2998 
2999 /*
3000  * This function waits for a read or write event to happen on a stream.
3001  * fmode can specify FNDELAY and/or FNONBLOCK.
3002  * The timeout is in ms with -1 meaning infinite.
3003  * The flag values work as follows:
3004  *	READWAIT	Check for read side errors, send M_READ
3005  *	GETWAIT		Check for read side errors, no M_READ
3006  *	WRITEWAIT	Check for write side errors.
3007  *	NOINTR		Do not return error if nonblocking or timeout.
3008  * 	STR_NOERROR	Ignore all errors except STPLEX.
3009  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
3010  *	STR_PEEK	Pass through the strgeterr().
3011  */
3012 int
3013 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
3014     int *done)
3015 {
3016 	int slpflg, errs;
3017 	int error;
3018 	kcondvar_t *sleepon;
3019 	mblk_t *mp;
3020 	ssize_t *rd_count;
3021 	clock_t rval;
3022 
3023 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3024 	if ((flag & READWAIT) || (flag & GETWAIT)) {
3025 		slpflg = RSLEEP;
3026 		sleepon = &_RD(stp->sd_wrq)->q_wait;
3027 		errs = STRDERR|STPLEX;
3028 	} else {
3029 		slpflg = WSLEEP;
3030 		sleepon = &stp->sd_wrq->q_wait;
3031 		errs = STWRERR|STRHUP|STPLEX;
3032 	}
3033 	if (flag & STR_NOERROR)
3034 		errs = STPLEX;
3035 
3036 	if (stp->sd_wakeq & slpflg) {
3037 		/*
3038 		 * A strwakeq() is pending, no need to sleep.
3039 		 */
3040 		stp->sd_wakeq &= ~slpflg;
3041 		*done = 0;
3042 		return (0);
3043 	}
3044 
3045 	if (stp->sd_flag & errs) {
3046 		/*
3047 		 * Check for errors before going to sleep since the
3048 		 * caller might not have checked this while holding
3049 		 * sd_lock.
3050 		 */
3051 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3052 		if (error != 0) {
3053 			*done = 1;
3054 			return (error);
3055 		}
3056 	}
3057 
3058 	/*
3059 	 * If any module downstream has requested read notification
3060 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3061 	 * down stream.
3062 	 */
3063 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3064 		mutex_exit(&stp->sd_lock);
3065 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3066 		    (flag & STR_NOSIG), &error))) {
3067 			mutex_enter(&stp->sd_lock);
3068 			*done = 1;
3069 			return (error);
3070 		}
3071 		mp->b_datap->db_type = M_READ;
3072 		rd_count = (ssize_t *)mp->b_wptr;
3073 		*rd_count = count;
3074 		mp->b_wptr += sizeof (ssize_t);
3075 		/*
3076 		 * Send the number of bytes requested by the
3077 		 * read as the argument to M_READ.
3078 		 */
3079 		stream_willservice(stp);
3080 		putnext(stp->sd_wrq, mp);
3081 		stream_runservice(stp);
3082 		mutex_enter(&stp->sd_lock);
3083 
3084 		/*
3085 		 * If any data arrived due to inline processing
3086 		 * of putnext(), don't sleep.
3087 		 */
3088 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3089 			*done = 0;
3090 			return (0);
3091 		}
3092 	}
3093 
3094 	if (fmode & (FNDELAY|FNONBLOCK)) {
3095 		if (!(flag & NOINTR))
3096 			error = EAGAIN;
3097 		else
3098 			error = 0;
3099 		*done = 1;
3100 		return (error);
3101 	}
3102 
3103 	stp->sd_flag |= slpflg;
3104 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3105 	    "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3106 	    stp, flag, count, fmode, done);
3107 
3108 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3109 	if (rval > 0) {
3110 		/* EMPTY */
3111 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3112 		    "strwaitq awakes(2):%X, %X, %X, %X, %X",
3113 		    stp, flag, count, fmode, done);
3114 	} else if (rval == 0) {
3115 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3116 		    "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3117 		    stp, flag, count, fmode, done);
3118 		stp->sd_flag &= ~slpflg;
3119 		cv_broadcast(sleepon);
3120 		if (!(flag & NOINTR))
3121 			error = EINTR;
3122 		else
3123 			error = 0;
3124 		*done = 1;
3125 		return (error);
3126 	} else {
3127 		/* timeout */
3128 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3129 		    "strwaitq timeout:%p, %X, %lX, %X, %p",
3130 		    stp, flag, count, fmode, done);
3131 		*done = 1;
3132 		if (!(flag & NOINTR))
3133 			return (ETIME);
3134 		else
3135 			return (0);
3136 	}
3137 	/*
3138 	 * If the caller implements delayed errors (i.e. queued after data)
3139 	 * we can not check for errors here since data as well as an
3140 	 * error might have arrived at the stream head. We return to
3141 	 * have the caller check the read queue before checking for errors.
3142 	 */
3143 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3144 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3145 		if (error != 0) {
3146 			*done = 1;
3147 			return (error);
3148 		}
3149 	}
3150 	*done = 0;
3151 	return (0);
3152 }
3153 
3154 /*
3155  * Perform job control discipline access checks.
3156  * Return 0 for success and the errno for failure.
3157  */
3158 
3159 #define	cantsend(p, t, sig) \
3160 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3161 
3162 int
3163 straccess(struct stdata *stp, enum jcaccess mode)
3164 {
3165 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3166 	kthread_t *t = curthread;
3167 	proc_t *p = ttoproc(t);
3168 	sess_t *sp;
3169 
3170 	ASSERT(mutex_owned(&stp->sd_lock));
3171 
3172 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3173 		return (0);
3174 
3175 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3176 
3177 	for (;;) {
3178 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3179 		sp = p->p_sessp;
3180 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3181 
3182 		/*
3183 		 * If this is not the calling process's controlling terminal
3184 		 * or if the calling process is already in the foreground
3185 		 * then allow access.
3186 		 */
3187 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3188 		    p->p_pgidp == stp->sd_pgidp) {
3189 			mutex_exit(&sp->s_lock);
3190 			mutex_exit(&p->p_splock);
3191 			mutex_exit(&p->p_lock);
3192 			return (0);
3193 		}
3194 
3195 		/*
3196 		 * Check to see if controlling terminal has been deallocated.
3197 		 */
3198 		if (sp->s_vp == NULL) {
3199 			if (!cantsend(p, t, SIGHUP))
3200 				sigtoproc(p, t, SIGHUP);
3201 			mutex_exit(&sp->s_lock);
3202 			mutex_exit(&p->p_splock);
3203 			mutex_exit(&p->p_lock);
3204 			return (EIO);
3205 		}
3206 
3207 		mutex_exit(&sp->s_lock);
3208 		mutex_exit(&p->p_splock);
3209 
3210 		if (mode == JCGETP) {
3211 			mutex_exit(&p->p_lock);
3212 			return (0);
3213 		}
3214 
3215 		if (mode == JCREAD) {
3216 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3217 				mutex_exit(&p->p_lock);
3218 				return (EIO);
3219 			}
3220 			mutex_exit(&p->p_lock);
3221 			mutex_exit(&stp->sd_lock);
3222 			pgsignal(p->p_pgidp, SIGTTIN);
3223 			mutex_enter(&stp->sd_lock);
3224 			mutex_enter(&p->p_lock);
3225 		} else {  /* mode == JCWRITE or JCSETP */
3226 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3227 			    cantsend(p, t, SIGTTOU)) {
3228 				mutex_exit(&p->p_lock);
3229 				return (0);
3230 			}
3231 			if (p->p_detached) {
3232 				mutex_exit(&p->p_lock);
3233 				return (EIO);
3234 			}
3235 			mutex_exit(&p->p_lock);
3236 			mutex_exit(&stp->sd_lock);
3237 			pgsignal(p->p_pgidp, SIGTTOU);
3238 			mutex_enter(&stp->sd_lock);
3239 			mutex_enter(&p->p_lock);
3240 		}
3241 
3242 		/*
3243 		 * We call cv_wait_sig_swap() to cause the appropriate
3244 		 * action for the jobcontrol signal to take place.
3245 		 * If the signal is being caught, we will take the
3246 		 * EINTR error return.  Otherwise, the default action
3247 		 * of causing the process to stop will take place.
3248 		 * In this case, we rely on the periodic cv_broadcast() on
3249 		 * &lbolt_cv to wake us up to loop around and test again.
3250 		 * We can't get here if the signal is ignored or
3251 		 * if the current thread is blocking the signal.
3252 		 */
3253 		mutex_exit(&stp->sd_lock);
3254 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3255 			mutex_exit(&p->p_lock);
3256 			mutex_enter(&stp->sd_lock);
3257 			return (EINTR);
3258 		}
3259 		mutex_exit(&p->p_lock);
3260 		mutex_enter(&stp->sd_lock);
3261 		mutex_enter(&p->p_lock);
3262 	}
3263 }
3264 
3265 /*
3266  * Return size of message of block type (bp->b_datap->db_type)
3267  */
3268 size_t
3269 xmsgsize(mblk_t *bp)
3270 {
3271 	unsigned char type;
3272 	size_t count = 0;
3273 
3274 	type = bp->b_datap->db_type;
3275 
3276 	for (; bp; bp = bp->b_cont) {
3277 		if (type != bp->b_datap->db_type)
3278 			break;
3279 		ASSERT(bp->b_wptr >= bp->b_rptr);
3280 		count += bp->b_wptr - bp->b_rptr;
3281 	}
3282 	return (count);
3283 }
3284 
3285 /*
3286  * Allocate a stream head.
3287  */
3288 struct stdata *
3289 shalloc(queue_t *qp)
3290 {
3291 	stdata_t *stp;
3292 
3293 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3294 
3295 	stp->sd_wrq = _WR(qp);
3296 	stp->sd_strtab = NULL;
3297 	stp->sd_iocid = 0;
3298 	stp->sd_mate = NULL;
3299 	stp->sd_freezer = NULL;
3300 	stp->sd_refcnt = 0;
3301 	stp->sd_wakeq = 0;
3302 	stp->sd_anchor = 0;
3303 	stp->sd_struiowrq = NULL;
3304 	stp->sd_struiordq = NULL;
3305 	stp->sd_struiodnak = 0;
3306 	stp->sd_struionak = NULL;
3307 	stp->sd_t_audit_data = NULL;
3308 	stp->sd_rput_opt = 0;
3309 	stp->sd_wput_opt = 0;
3310 	stp->sd_read_opt = 0;
3311 	stp->sd_rprotofunc = strrput_proto;
3312 	stp->sd_rmiscfunc = strrput_misc;
3313 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3314 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3315 	stp->sd_ciputctrl = NULL;
3316 	stp->sd_nciputctrl = 0;
3317 	stp->sd_qhead = NULL;
3318 	stp->sd_qtail = NULL;
3319 	stp->sd_servid = NULL;
3320 	stp->sd_nqueues = 0;
3321 	stp->sd_svcflags = 0;
3322 	stp->sd_copyflag = 0;
3323 
3324 	return (stp);
3325 }
3326 
3327 /*
3328  * Free a stream head.
3329  */
3330 void
3331 shfree(stdata_t *stp)
3332 {
3333 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3334 
3335 	stp->sd_wrq = NULL;
3336 
3337 	mutex_enter(&stp->sd_qlock);
3338 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3339 		STRSTAT(strwaits);
3340 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3341 	}
3342 	mutex_exit(&stp->sd_qlock);
3343 
3344 	if (stp->sd_ciputctrl != NULL) {
3345 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3346 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3347 		    stp->sd_nciputctrl, 0);
3348 		ASSERT(ciputctrl_cache != NULL);
3349 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3350 		stp->sd_ciputctrl = NULL;
3351 		stp->sd_nciputctrl = 0;
3352 	}
3353 	ASSERT(stp->sd_qhead == NULL);
3354 	ASSERT(stp->sd_qtail == NULL);
3355 	ASSERT(stp->sd_nqueues == 0);
3356 	kmem_cache_free(stream_head_cache, stp);
3357 }
3358 
3359 /*
3360  * Allocate a pair of queues and a syncq for the pair
3361  */
3362 queue_t *
3363 allocq(void)
3364 {
3365 	queinfo_t *qip;
3366 	queue_t *qp, *wqp;
3367 	syncq_t	*sq;
3368 
3369 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3370 
3371 	qp = &qip->qu_rqueue;
3372 	wqp = &qip->qu_wqueue;
3373 	sq = &qip->qu_syncq;
3374 
3375 	qp->q_last	= NULL;
3376 	qp->q_next	= NULL;
3377 	qp->q_ptr	= NULL;
3378 	qp->q_flag	= QUSE | QREADR;
3379 	qp->q_bandp	= NULL;
3380 	qp->q_stream	= NULL;
3381 	qp->q_syncq	= sq;
3382 	qp->q_nband	= 0;
3383 	qp->q_nfsrv	= NULL;
3384 	qp->q_draining	= 0;
3385 	qp->q_syncqmsgs	= 0;
3386 	qp->q_spri	= 0;
3387 	qp->q_qtstamp	= 0;
3388 	qp->q_sqtstamp	= 0;
3389 	qp->q_fp	= NULL;
3390 
3391 	wqp->q_last	= NULL;
3392 	wqp->q_next	= NULL;
3393 	wqp->q_ptr	= NULL;
3394 	wqp->q_flag	= QUSE;
3395 	wqp->q_bandp	= NULL;
3396 	wqp->q_stream	= NULL;
3397 	wqp->q_syncq	= sq;
3398 	wqp->q_nband	= 0;
3399 	wqp->q_nfsrv	= NULL;
3400 	wqp->q_draining	= 0;
3401 	wqp->q_syncqmsgs = 0;
3402 	wqp->q_qtstamp	= 0;
3403 	wqp->q_sqtstamp	= 0;
3404 	wqp->q_spri	= 0;
3405 
3406 	sq->sq_count	= 0;
3407 	sq->sq_rmqcount	= 0;
3408 	sq->sq_flags	= 0;
3409 	sq->sq_type	= 0;
3410 	sq->sq_callbflags = 0;
3411 	sq->sq_cancelid	= 0;
3412 	sq->sq_ciputctrl = NULL;
3413 	sq->sq_nciputctrl = 0;
3414 	sq->sq_needexcl = 0;
3415 	sq->sq_svcflags = 0;
3416 
3417 	return (qp);
3418 }
3419 
3420 /*
3421  * Free a pair of queues and the "attached" syncq.
3422  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3423  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3424  */
3425 void
3426 freeq(queue_t *qp)
3427 {
3428 	qband_t *qbp, *nqbp;
3429 	syncq_t *sq, *outer;
3430 	queue_t *wqp = _WR(qp);
3431 
3432 	ASSERT(qp->q_flag & QREADR);
3433 
3434 	/*
3435 	 * If a previously dispatched taskq job is scheduled to run
3436 	 * sync_service() or a service routine is scheduled for the
3437 	 * queues about to be freed, wait here until all service is
3438 	 * done on the queue and all associated queues and syncqs.
3439 	 */
3440 	wait_svc(qp);
3441 
3442 	(void) flush_syncq(qp->q_syncq, qp);
3443 	(void) flush_syncq(wqp->q_syncq, wqp);
3444 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3445 
3446 	/*
3447 	 * Flush the queues before q_next is set to NULL This is needed
3448 	 * in order to backenable any downstream queue before we go away.
3449 	 * Note: we are already removed from the stream so that the
3450 	 * backenabling will not cause any messages to be delivered to our
3451 	 * put procedures.
3452 	 */
3453 	flushq(qp, FLUSHALL);
3454 	flushq(wqp, FLUSHALL);
3455 
3456 	/* Tidy up - removeq only does a half-remove from stream */
3457 	qp->q_next = wqp->q_next = NULL;
3458 	ASSERT(!(qp->q_flag & QENAB));
3459 	ASSERT(!(wqp->q_flag & QENAB));
3460 
3461 	outer = qp->q_syncq->sq_outer;
3462 	if (outer != NULL) {
3463 		outer_remove(outer, qp->q_syncq);
3464 		if (wqp->q_syncq != qp->q_syncq)
3465 			outer_remove(outer, wqp->q_syncq);
3466 	}
3467 	/*
3468 	 * Free any syncqs that are outside what allocq returned.
3469 	 */
3470 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3471 		free_syncq(qp->q_syncq);
3472 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3473 		free_syncq(wqp->q_syncq);
3474 
3475 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3476 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3477 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3478 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3479 	sq = SQ(qp);
3480 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3481 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3482 	ASSERT(sq->sq_outer == NULL);
3483 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3484 	ASSERT(sq->sq_callbpend == NULL);
3485 	ASSERT(sq->sq_needexcl == 0);
3486 
3487 	if (sq->sq_ciputctrl != NULL) {
3488 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3489 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3490 		    sq->sq_nciputctrl, 0);
3491 		ASSERT(ciputctrl_cache != NULL);
3492 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3493 		sq->sq_ciputctrl = NULL;
3494 		sq->sq_nciputctrl = 0;
3495 	}
3496 
3497 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3498 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3499 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3500 
3501 	qp->q_flag &= ~QUSE;
3502 	wqp->q_flag &= ~QUSE;
3503 
3504 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3505 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3506 
3507 	qbp = qp->q_bandp;
3508 	while (qbp) {
3509 		nqbp = qbp->qb_next;
3510 		freeband(qbp);
3511 		qbp = nqbp;
3512 	}
3513 	qbp = wqp->q_bandp;
3514 	while (qbp) {
3515 		nqbp = qbp->qb_next;
3516 		freeband(qbp);
3517 		qbp = nqbp;
3518 	}
3519 	kmem_cache_free(queue_cache, qp);
3520 }
3521 
3522 /*
3523  * Allocate a qband structure.
3524  */
3525 qband_t *
3526 allocband(void)
3527 {
3528 	qband_t *qbp;
3529 
3530 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3531 	if (qbp == NULL)
3532 		return (NULL);
3533 
3534 	qbp->qb_next	= NULL;
3535 	qbp->qb_count	= 0;
3536 	qbp->qb_mblkcnt	= 0;
3537 	qbp->qb_first	= NULL;
3538 	qbp->qb_last	= NULL;
3539 	qbp->qb_flag	= 0;
3540 
3541 	return (qbp);
3542 }
3543 
3544 /*
3545  * Free a qband structure.
3546  */
3547 void
3548 freeband(qband_t *qbp)
3549 {
3550 	kmem_cache_free(qband_cache, qbp);
3551 }
3552 
3553 /*
3554  * Just like putnextctl(9F), except that allocb_wait() is used.
3555  *
3556  * Consolidation Private, and of course only callable from the stream head or
3557  * routines that may block.
3558  */
3559 int
3560 putnextctl_wait(queue_t *q, int type)
3561 {
3562 	mblk_t *bp;
3563 	int error;
3564 
3565 	if ((datamsg(type) && (type != M_DELAY)) ||
3566 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3567 		return (0);
3568 
3569 	bp->b_datap->db_type = (unsigned char)type;
3570 	putnext(q, bp);
3571 	return (1);
3572 }
3573 
3574 /*
3575  * Run any possible bufcalls.
3576  */
3577 void
3578 runbufcalls(void)
3579 {
3580 	strbufcall_t *bcp;
3581 
3582 	mutex_enter(&bcall_monitor);
3583 	mutex_enter(&strbcall_lock);
3584 
3585 	if (strbcalls.bc_head) {
3586 		size_t count;
3587 		int nevent;
3588 
3589 		/*
3590 		 * count how many events are on the list
3591 		 * now so we can check to avoid looping
3592 		 * in low memory situations
3593 		 */
3594 		nevent = 0;
3595 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3596 			nevent++;
3597 
3598 		/*
3599 		 * get estimate of available memory from kmem_avail().
3600 		 * awake all bufcall functions waiting for
3601 		 * memory whose request could be satisfied
3602 		 * by 'count' memory and let 'em fight for it.
3603 		 */
3604 		count = kmem_avail();
3605 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3606 			STRSTAT(bufcalls);
3607 			--nevent;
3608 			if (bcp->bc_size <= count) {
3609 				bcp->bc_executor = curthread;
3610 				mutex_exit(&strbcall_lock);
3611 				(*bcp->bc_func)(bcp->bc_arg);
3612 				mutex_enter(&strbcall_lock);
3613 				bcp->bc_executor = NULL;
3614 				cv_broadcast(&bcall_cv);
3615 				strbcalls.bc_head = bcp->bc_next;
3616 				kmem_free(bcp, sizeof (strbufcall_t));
3617 			} else {
3618 				/*
3619 				 * too big, try again later - note
3620 				 * that nevent was decremented above
3621 				 * so we won't retry this one on this
3622 				 * iteration of the loop
3623 				 */
3624 				if (bcp->bc_next != NULL) {
3625 					strbcalls.bc_head = bcp->bc_next;
3626 					bcp->bc_next = NULL;
3627 					strbcalls.bc_tail->bc_next = bcp;
3628 					strbcalls.bc_tail = bcp;
3629 				}
3630 			}
3631 		}
3632 		if (strbcalls.bc_head == NULL)
3633 			strbcalls.bc_tail = NULL;
3634 	}
3635 
3636 	mutex_exit(&strbcall_lock);
3637 	mutex_exit(&bcall_monitor);
3638 }
3639 
3640 
3641 /*
3642  * Actually run queue's service routine.
3643  */
3644 static void
3645 runservice(queue_t *q)
3646 {
3647 	qband_t *qbp;
3648 
3649 	ASSERT(q->q_qinfo->qi_srvp);
3650 again:
3651 	entersq(q->q_syncq, SQ_SVC);
3652 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3653 	    "runservice starts:%p", q);
3654 
3655 	if (!(q->q_flag & QWCLOSE))
3656 		(*q->q_qinfo->qi_srvp)(q);
3657 
3658 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3659 	    "runservice ends:(%p)", q);
3660 
3661 	leavesq(q->q_syncq, SQ_SVC);
3662 
3663 	mutex_enter(QLOCK(q));
3664 	if (q->q_flag & QENAB) {
3665 		q->q_flag &= ~QENAB;
3666 		mutex_exit(QLOCK(q));
3667 		goto again;
3668 	}
3669 	q->q_flag &= ~QINSERVICE;
3670 	q->q_flag &= ~QBACK;
3671 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3672 		qbp->qb_flag &= ~QB_BACK;
3673 	/*
3674 	 * Wakeup thread waiting for the service procedure
3675 	 * to be run (strclose and qdetach).
3676 	 */
3677 	cv_broadcast(&q->q_wait);
3678 
3679 	mutex_exit(QLOCK(q));
3680 }
3681 
3682 /*
3683  * Background processing of bufcalls.
3684  */
3685 void
3686 streams_bufcall_service(void)
3687 {
3688 	callb_cpr_t	cprinfo;
3689 
3690 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3691 	    "streams_bufcall_service");
3692 
3693 	mutex_enter(&strbcall_lock);
3694 
3695 	for (;;) {
3696 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3697 			mutex_exit(&strbcall_lock);
3698 			runbufcalls();
3699 			mutex_enter(&strbcall_lock);
3700 		}
3701 		if (strbcalls.bc_head != NULL) {
3702 			STRSTAT(bcwaits);
3703 			/* Wait for memory to become available */
3704 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3705 			(void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3706 			    SEC_TO_TICK(60), TR_CLOCK_TICK);
3707 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3708 		}
3709 
3710 		/* Wait for new work to arrive */
3711 		if (strbcalls.bc_head == NULL) {
3712 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3713 			cv_wait(&strbcall_cv, &strbcall_lock);
3714 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3715 		}
3716 	}
3717 }
3718 
3719 /*
3720  * Background processing of streams background tasks which failed
3721  * taskq_dispatch.
3722  */
3723 static void
3724 streams_qbkgrnd_service(void)
3725 {
3726 	callb_cpr_t cprinfo;
3727 	queue_t *q;
3728 
3729 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3730 	    "streams_bkgrnd_service");
3731 
3732 	mutex_enter(&service_queue);
3733 
3734 	for (;;) {
3735 		/*
3736 		 * Wait for work to arrive.
3737 		 */
3738 		while ((freebs_list == NULL) && (qhead == NULL)) {
3739 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3740 			cv_wait(&services_to_run, &service_queue);
3741 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3742 		}
3743 		/*
3744 		 * Handle all pending freebs requests to free memory.
3745 		 */
3746 		while (freebs_list != NULL) {
3747 			mblk_t *mp = freebs_list;
3748 			freebs_list = mp->b_next;
3749 			mutex_exit(&service_queue);
3750 			mblk_free(mp);
3751 			mutex_enter(&service_queue);
3752 		}
3753 		/*
3754 		 * Run pending queues.
3755 		 */
3756 		while (qhead != NULL) {
3757 			DQ(q, qhead, qtail, q_link);
3758 			ASSERT(q != NULL);
3759 			mutex_exit(&service_queue);
3760 			queue_service(q);
3761 			mutex_enter(&service_queue);
3762 		}
3763 		ASSERT(qhead == NULL && qtail == NULL);
3764 	}
3765 }
3766 
3767 /*
3768  * Background processing of streams background tasks which failed
3769  * taskq_dispatch.
3770  */
3771 static void
3772 streams_sqbkgrnd_service(void)
3773 {
3774 	callb_cpr_t cprinfo;
3775 	syncq_t *sq;
3776 
3777 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3778 	    "streams_sqbkgrnd_service");
3779 
3780 	mutex_enter(&service_queue);
3781 
3782 	for (;;) {
3783 		/*
3784 		 * Wait for work to arrive.
3785 		 */
3786 		while (sqhead == NULL) {
3787 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3788 			cv_wait(&syncqs_to_run, &service_queue);
3789 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3790 		}
3791 
3792 		/*
3793 		 * Run pending syncqs.
3794 		 */
3795 		while (sqhead != NULL) {
3796 			DQ(sq, sqhead, sqtail, sq_next);
3797 			ASSERT(sq != NULL);
3798 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3799 			mutex_exit(&service_queue);
3800 			syncq_service(sq);
3801 			mutex_enter(&service_queue);
3802 		}
3803 	}
3804 }
3805 
3806 /*
3807  * Disable the syncq and wait for background syncq processing to complete.
3808  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3809  * list.
3810  */
3811 void
3812 wait_sq_svc(syncq_t *sq)
3813 {
3814 	mutex_enter(SQLOCK(sq));
3815 	sq->sq_svcflags |= SQ_DISABLED;
3816 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3817 		syncq_t *sq_chase;
3818 		syncq_t *sq_curr;
3819 		int removed;
3820 
3821 		ASSERT(sq->sq_servcount == 1);
3822 		mutex_enter(&service_queue);
3823 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3824 		mutex_exit(&service_queue);
3825 		if (removed) {
3826 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3827 			sq->sq_servcount = 0;
3828 			STRSTAT(sqremoved);
3829 			goto done;
3830 		}
3831 	}
3832 	while (sq->sq_servcount != 0) {
3833 		sq->sq_flags |= SQ_WANTWAKEUP;
3834 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3835 	}
3836 done:
3837 	mutex_exit(SQLOCK(sq));
3838 }
3839 
3840 /*
3841  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3842  * Add the argument to the end of the sqhead list and set the flag
3843  * indicating this syncq has been enabled.  If it has already been
3844  * enabled, don't do anything.
3845  * This routine assumes that SQLOCK is held.
3846  * NOTE that the lock order is to have the SQLOCK first,
3847  * so if the service_syncq lock is held, we need to release it
3848  * before acquiring the SQLOCK (mostly relevant for the background
3849  * thread, and this seems to be common among the STREAMS global locks).
3850  * Note that the sq_svcflags are protected by the SQLOCK.
3851  */
3852 void
3853 sqenable(syncq_t *sq)
3854 {
3855 	/*
3856 	 * This is probably not important except for where I believe it
3857 	 * is being called.  At that point, it should be held (and it
3858 	 * is a pain to release it just for this routine, so don't do
3859 	 * it).
3860 	 */
3861 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3862 
3863 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3864 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3865 
3866 	/*
3867 	 * Do not put on list if background thread is scheduled or
3868 	 * syncq is disabled.
3869 	 */
3870 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3871 		return;
3872 
3873 	/*
3874 	 * Check whether we should enable sq at all.
3875 	 * Non PERMOD syncqs may be drained by at most one thread.
3876 	 * PERMOD syncqs may be drained by several threads but we limit the
3877 	 * total amount to the lesser of
3878 	 *	Number of queues on the squeue and
3879 	 *	Number of CPUs.
3880 	 */
3881 	if (sq->sq_servcount != 0) {
3882 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3883 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3884 			STRSTAT(sqtoomany);
3885 			return;
3886 		}
3887 	}
3888 
3889 	sq->sq_tstamp = ddi_get_lbolt();
3890 	STRSTAT(sqenables);
3891 
3892 	/* Attempt a taskq dispatch */
3893 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3894 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3895 	if (sq->sq_servid != NULL) {
3896 		sq->sq_servcount++;
3897 		return;
3898 	}
3899 
3900 	/*
3901 	 * This taskq dispatch failed, but a previous one may have succeeded.
3902 	 * Don't try to schedule on the background thread whilst there is
3903 	 * outstanding taskq processing.
3904 	 */
3905 	if (sq->sq_servcount != 0)
3906 		return;
3907 
3908 	/*
3909 	 * System is low on resources and can't perform a non-sleeping
3910 	 * dispatch. Schedule the syncq for a background thread and mark the
3911 	 * syncq to avoid any further taskq dispatch attempts.
3912 	 */
3913 	mutex_enter(&service_queue);
3914 	STRSTAT(taskqfails);
3915 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3916 	sq->sq_svcflags |= SQ_BGTHREAD;
3917 	sq->sq_servcount = 1;
3918 	cv_signal(&syncqs_to_run);
3919 	mutex_exit(&service_queue);
3920 }
3921 
3922 /*
3923  * Note: fifo_close() depends on the mblk_t on the queue being freed
3924  * asynchronously. The asynchronous freeing of messages breaks the
3925  * recursive call chain of fifo_close() while there are I_SENDFD type of
3926  * messages referring to other file pointers on the queue. Then when
3927  * closing pipes it can avoid stack overflow in case of daisy-chained
3928  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3929  * share the same fifolock_t).
3930  *
3931  * No need to kpreempt_disable to access cpu_seqid.  If we migrate and
3932  * the esb queue does not match the new CPU, that is OK.
3933  */
3934 void
3935 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3936 {
3937 	int qindex = CPU->cpu_seqid >> esbq_log2_cpus_per_q;
3938 	esb_queue_t *eqp;
3939 
3940 	ASSERT(dbp->db_mblk == mp);
3941 	ASSERT(qindex < esbq_nelem);
3942 
3943 	eqp = system_esbq_array;
3944 	if (eqp != NULL) {
3945 		eqp += qindex;
3946 	} else {
3947 		mutex_enter(&esbq_lock);
3948 		if (kmem_ready && system_esbq_array == NULL)
3949 			system_esbq_array = (esb_queue_t *)kmem_zalloc(
3950 			    esbq_nelem * sizeof (esb_queue_t), KM_NOSLEEP);
3951 		mutex_exit(&esbq_lock);
3952 		eqp = system_esbq_array;
3953 		if (eqp != NULL)
3954 			eqp += qindex;
3955 		else
3956 			eqp = &system_esbq;
3957 	}
3958 
3959 	/*
3960 	 * Check data sanity. The dblock should have non-empty free function.
3961 	 * It is better to panic here then later when the dblock is freed
3962 	 * asynchronously when the context is lost.
3963 	 */
3964 	if (dbp->db_frtnp->free_func == NULL) {
3965 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3966 		    (void *)dbp);
3967 	}
3968 
3969 	mutex_enter(&eqp->eq_lock);
3970 	/* queue the new mblk on the esballoc queue */
3971 	if (eqp->eq_head == NULL) {
3972 		eqp->eq_head = eqp->eq_tail = mp;
3973 	} else {
3974 		eqp->eq_tail->b_next = mp;
3975 		eqp->eq_tail = mp;
3976 	}
3977 	eqp->eq_len++;
3978 
3979 	/* If we're the first thread to reach the threshold, process */
3980 	if (eqp->eq_len >= esbq_max_qlen &&
3981 	    !(eqp->eq_flags & ESBQ_PROCESSING))
3982 		esballoc_process_queue(eqp);
3983 
3984 	esballoc_set_timer(eqp, esbq_timeout);
3985 	mutex_exit(&eqp->eq_lock);
3986 }
3987 
3988 static void
3989 esballoc_process_queue(esb_queue_t *eqp)
3990 {
3991 	mblk_t	*mp;
3992 
3993 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
3994 
3995 	eqp->eq_flags |= ESBQ_PROCESSING;
3996 
3997 	do {
3998 		/*
3999 		 * Detach the message chain for processing.
4000 		 */
4001 		mp = eqp->eq_head;
4002 		eqp->eq_tail->b_next = NULL;
4003 		eqp->eq_head = eqp->eq_tail = NULL;
4004 		eqp->eq_len = 0;
4005 		mutex_exit(&eqp->eq_lock);
4006 
4007 		/*
4008 		 * Process the message chain.
4009 		 */
4010 		esballoc_enqueue_mblk(mp);
4011 		mutex_enter(&eqp->eq_lock);
4012 	} while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
4013 
4014 	eqp->eq_flags &= ~ESBQ_PROCESSING;
4015 }
4016 
4017 /*
4018  * taskq callback routine to free esballoced mblk's
4019  */
4020 static void
4021 esballoc_mblk_free(mblk_t *mp)
4022 {
4023 	mblk_t	*nextmp;
4024 
4025 	for (; mp != NULL; mp = nextmp) {
4026 		nextmp = mp->b_next;
4027 		mp->b_next = NULL;
4028 		mblk_free(mp);
4029 	}
4030 }
4031 
4032 static void
4033 esballoc_enqueue_mblk(mblk_t *mp)
4034 {
4035 
4036 	if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
4037 	    TQ_NOSLEEP) == NULL) {
4038 		mblk_t *first_mp = mp;
4039 		/*
4040 		 * System is low on resources and can't perform a non-sleeping
4041 		 * dispatch. Schedule for a background thread.
4042 		 */
4043 		mutex_enter(&service_queue);
4044 		STRSTAT(taskqfails);
4045 
4046 		while (mp->b_next != NULL)
4047 			mp = mp->b_next;
4048 
4049 		mp->b_next = freebs_list;
4050 		freebs_list = first_mp;
4051 		cv_signal(&services_to_run);
4052 		mutex_exit(&service_queue);
4053 	}
4054 }
4055 
4056 static void
4057 esballoc_timer(void *arg)
4058 {
4059 	esb_queue_t *eqp = arg;
4060 
4061 	mutex_enter(&eqp->eq_lock);
4062 	eqp->eq_flags &= ~ESBQ_TIMER;
4063 
4064 	if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4065 	    eqp->eq_len > 0)
4066 		esballoc_process_queue(eqp);
4067 
4068 	esballoc_set_timer(eqp, esbq_timeout);
4069 	mutex_exit(&eqp->eq_lock);
4070 }
4071 
4072 static void
4073 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4074 {
4075 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4076 
4077 	if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4078 		(void) timeout(esballoc_timer, eqp, eq_timeout);
4079 		eqp->eq_flags |= ESBQ_TIMER;
4080 	}
4081 }
4082 
4083 /*
4084  * Setup esbq array length based upon NCPU scaled by CPUs per
4085  * queue. Use static system_esbq until kmem_ready and we can
4086  * create an array in freebs_enqueue().
4087  */
4088 void
4089 esballoc_queue_init(void)
4090 {
4091 	esbq_log2_cpus_per_q = highbit(esbq_cpus_per_q - 1);
4092 	esbq_cpus_per_q = 1 << esbq_log2_cpus_per_q;
4093 	esbq_nelem = howmany(NCPU, esbq_cpus_per_q);
4094 	system_esbq.eq_len = 0;
4095 	system_esbq.eq_head = system_esbq.eq_tail = NULL;
4096 	system_esbq.eq_flags = 0;
4097 }
4098 
4099 /*
4100  * Set the QBACK or QB_BACK flag in the given queue for
4101  * the given priority band.
4102  */
4103 void
4104 setqback(queue_t *q, unsigned char pri)
4105 {
4106 	int i;
4107 	qband_t *qbp;
4108 	qband_t **qbpp;
4109 
4110 	ASSERT(MUTEX_HELD(QLOCK(q)));
4111 	if (pri != 0) {
4112 		if (pri > q->q_nband) {
4113 			qbpp = &q->q_bandp;
4114 			while (*qbpp)
4115 				qbpp = &(*qbpp)->qb_next;
4116 			while (pri > q->q_nband) {
4117 				if ((*qbpp = allocband()) == NULL) {
4118 					cmn_err(CE_WARN,
4119 					    "setqback: can't allocate qband\n");
4120 					return;
4121 				}
4122 				(*qbpp)->qb_hiwat = q->q_hiwat;
4123 				(*qbpp)->qb_lowat = q->q_lowat;
4124 				q->q_nband++;
4125 				qbpp = &(*qbpp)->qb_next;
4126 			}
4127 		}
4128 		qbp = q->q_bandp;
4129 		i = pri;
4130 		while (--i)
4131 			qbp = qbp->qb_next;
4132 		qbp->qb_flag |= QB_BACK;
4133 	} else {
4134 		q->q_flag |= QBACK;
4135 	}
4136 }
4137 
4138 int
4139 strcopyin(void *from, void *to, size_t len, int copyflag)
4140 {
4141 	if (copyflag & U_TO_K) {
4142 		ASSERT((copyflag & K_TO_K) == 0);
4143 		if (copyin(from, to, len))
4144 			return (EFAULT);
4145 	} else {
4146 		ASSERT(copyflag & K_TO_K);
4147 		bcopy(from, to, len);
4148 	}
4149 	return (0);
4150 }
4151 
4152 int
4153 strcopyout(void *from, void *to, size_t len, int copyflag)
4154 {
4155 	if (copyflag & U_TO_K) {
4156 		if (copyout(from, to, len))
4157 			return (EFAULT);
4158 	} else {
4159 		ASSERT(copyflag & K_TO_K);
4160 		bcopy(from, to, len);
4161 	}
4162 	return (0);
4163 }
4164 
4165 /*
4166  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4167  * It assumes that the stream head lock is already held, whereas strsignal()
4168  * acquires the lock first.  This routine was created because a few callers
4169  * release the stream head lock before calling only to re-acquire it after
4170  * it returns.
4171  */
4172 void
4173 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4174 {
4175 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4176 	switch (sig) {
4177 	case SIGPOLL:
4178 		if (stp->sd_sigflags & S_MSG)
4179 			strsendsig(stp->sd_siglist, S_MSG, band, 0);
4180 		break;
4181 	default:
4182 		if (stp->sd_pgidp)
4183 			pgsignal(stp->sd_pgidp, sig);
4184 		break;
4185 	}
4186 }
4187 
4188 void
4189 strsignal(stdata_t *stp, int sig, int32_t band)
4190 {
4191 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4192 	    "strsignal:%p, %X, %X", stp, sig, band);
4193 
4194 	mutex_enter(&stp->sd_lock);
4195 	switch (sig) {
4196 	case SIGPOLL:
4197 		if (stp->sd_sigflags & S_MSG)
4198 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4199 		break;
4200 
4201 	default:
4202 		if (stp->sd_pgidp) {
4203 			pgsignal(stp->sd_pgidp, sig);
4204 		}
4205 		break;
4206 	}
4207 	mutex_exit(&stp->sd_lock);
4208 }
4209 
4210 void
4211 strhup(stdata_t *stp)
4212 {
4213 	ASSERT(mutex_owned(&stp->sd_lock));
4214 	pollwakeup(&stp->sd_pollist, POLLHUP);
4215 	if (stp->sd_sigflags & S_HANGUP)
4216 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4217 }
4218 
4219 /*
4220  * Backenable the first queue upstream from `q' with a service procedure.
4221  */
4222 void
4223 backenable(queue_t *q, uchar_t pri)
4224 {
4225 	queue_t	*nq;
4226 
4227 	/*
4228 	 * Our presence might not prevent other modules in our own
4229 	 * stream from popping/pushing since the caller of getq might not
4230 	 * have a claim on the queue (some drivers do a getq on somebody
4231 	 * else's queue - they know that the queue itself is not going away
4232 	 * but the framework has to guarantee q_next in that stream).
4233 	 */
4234 	claimstr(q);
4235 
4236 	/* Find nearest back queue with service proc */
4237 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4238 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4239 	}
4240 
4241 	if (nq) {
4242 		kthread_t *freezer;
4243 		/*
4244 		 * backenable can be called either with no locks held
4245 		 * or with the stream frozen (the latter occurs when a module
4246 		 * calls rmvq with the stream frozen). If the stream is frozen
4247 		 * by the caller the caller will hold all qlocks in the stream.
4248 		 * Note that a frozen stream doesn't freeze a mated stream,
4249 		 * so we explicitly check for that.
4250 		 */
4251 		freezer = STREAM(q)->sd_freezer;
4252 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4253 			mutex_enter(QLOCK(nq));
4254 		}
4255 #ifdef DEBUG
4256 		else {
4257 			ASSERT(frozenstr(q));
4258 			ASSERT(MUTEX_HELD(QLOCK(q)));
4259 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4260 		}
4261 #endif
4262 		setqback(nq, pri);
4263 		qenable_locked(nq);
4264 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4265 			mutex_exit(QLOCK(nq));
4266 	}
4267 	releasestr(q);
4268 }
4269 
4270 /*
4271  * Return the appropriate errno when one of flags_to_check is set
4272  * in sd_flags. Uses the exported error routines if they are set.
4273  * Will return 0 if non error is set (or if the exported error routines
4274  * do not return an error).
4275  *
4276  * If there is both a read and write error to check, we prefer the read error.
4277  * Also, give preference to recorded errno's over the error functions.
4278  * The flags that are handled are:
4279  *	STPLEX		return EINVAL
4280  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4281  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4282  *	STRHUP		return sd_werror
4283  *
4284  * If the caller indicates that the operation is a peek, a nonpersistent error
4285  * is not cleared.
4286  */
4287 int
4288 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4289 {
4290 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4291 	int error = 0;
4292 
4293 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4294 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4295 	if (sd_flag & STPLEX)
4296 		error = EINVAL;
4297 	else if (sd_flag & STRDERR) {
4298 		error = stp->sd_rerror;
4299 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4300 			/*
4301 			 * Read errors are non-persistent i.e. discarded once
4302 			 * returned to a non-peeking caller,
4303 			 */
4304 			stp->sd_rerror = 0;
4305 			stp->sd_flag &= ~STRDERR;
4306 		}
4307 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4308 			int clearerr = 0;
4309 
4310 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4311 			    &clearerr);
4312 			if (clearerr) {
4313 				stp->sd_flag &= ~STRDERR;
4314 				stp->sd_rderrfunc = NULL;
4315 			}
4316 		}
4317 	} else if (sd_flag & STWRERR) {
4318 		error = stp->sd_werror;
4319 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4320 			/*
4321 			 * Write errors are non-persistent i.e. discarded once
4322 			 * returned to a non-peeking caller,
4323 			 */
4324 			stp->sd_werror = 0;
4325 			stp->sd_flag &= ~STWRERR;
4326 		}
4327 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4328 			int clearerr = 0;
4329 
4330 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4331 			    &clearerr);
4332 			if (clearerr) {
4333 				stp->sd_flag &= ~STWRERR;
4334 				stp->sd_wrerrfunc = NULL;
4335 			}
4336 		}
4337 	} else if (sd_flag & STRHUP) {
4338 		/* sd_werror set when STRHUP */
4339 		error = stp->sd_werror;
4340 	}
4341 	return (error);
4342 }
4343 
4344 
4345 /*
4346  * Single-thread open/close/push/pop
4347  * for twisted streams also
4348  */
4349 int
4350 strstartplumb(stdata_t *stp, int flag, int cmd)
4351 {
4352 	int waited = 1;
4353 	int error = 0;
4354 
4355 	if (STRMATED(stp)) {
4356 		struct stdata *stmatep = stp->sd_mate;
4357 
4358 		STRLOCKMATES(stp);
4359 		while (waited) {
4360 			waited = 0;
4361 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4362 				if ((cmd == I_POP) &&
4363 				    (flag & (FNDELAY|FNONBLOCK))) {
4364 					STRUNLOCKMATES(stp);
4365 					return (EAGAIN);
4366 				}
4367 				waited = 1;
4368 				mutex_exit(&stp->sd_lock);
4369 				if (!cv_wait_sig(&stmatep->sd_monitor,
4370 				    &stmatep->sd_lock)) {
4371 					mutex_exit(&stmatep->sd_lock);
4372 					return (EINTR);
4373 				}
4374 				mutex_exit(&stmatep->sd_lock);
4375 				STRLOCKMATES(stp);
4376 			}
4377 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4378 				if ((cmd == I_POP) &&
4379 				    (flag & (FNDELAY|FNONBLOCK))) {
4380 					STRUNLOCKMATES(stp);
4381 					return (EAGAIN);
4382 				}
4383 				waited = 1;
4384 				mutex_exit(&stmatep->sd_lock);
4385 				if (!cv_wait_sig(&stp->sd_monitor,
4386 				    &stp->sd_lock)) {
4387 					mutex_exit(&stp->sd_lock);
4388 					return (EINTR);
4389 				}
4390 				mutex_exit(&stp->sd_lock);
4391 				STRLOCKMATES(stp);
4392 			}
4393 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4394 				error = strgeterr(stp,
4395 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4396 				if (error != 0) {
4397 					STRUNLOCKMATES(stp);
4398 					return (error);
4399 				}
4400 			}
4401 		}
4402 		stp->sd_flag |= STRPLUMB;
4403 		STRUNLOCKMATES(stp);
4404 	} else {
4405 		mutex_enter(&stp->sd_lock);
4406 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4407 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4408 			    (flag & (FNDELAY|FNONBLOCK))) {
4409 				mutex_exit(&stp->sd_lock);
4410 				return (EAGAIN);
4411 			}
4412 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4413 				mutex_exit(&stp->sd_lock);
4414 				return (EINTR);
4415 			}
4416 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4417 				error = strgeterr(stp,
4418 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4419 				if (error != 0) {
4420 					mutex_exit(&stp->sd_lock);
4421 					return (error);
4422 				}
4423 			}
4424 		}
4425 		stp->sd_flag |= STRPLUMB;
4426 		mutex_exit(&stp->sd_lock);
4427 	}
4428 	return (0);
4429 }
4430 
4431 /*
4432  * Complete the plumbing operation associated with stream `stp'.
4433  */
4434 void
4435 strendplumb(stdata_t *stp)
4436 {
4437 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4438 	ASSERT(stp->sd_flag & STRPLUMB);
4439 	stp->sd_flag &= ~STRPLUMB;
4440 	cv_broadcast(&stp->sd_monitor);
4441 }
4442 
4443 /*
4444  * This describes how the STREAMS framework handles synchronization
4445  * during open/push and close/pop.
4446  * The key interfaces for open and close are qprocson and qprocsoff,
4447  * respectively. While the close case in general is harder both open
4448  * have close have significant similarities.
4449  *
4450  * During close the STREAMS framework has to both ensure that there
4451  * are no stale references to the queue pair (and syncq) that
4452  * are being closed and also provide the guarantees that are documented
4453  * in qprocsoff(9F).
4454  * If there are stale references to the queue that is closing it can
4455  * result in kernel memory corruption or kernel panics.
4456  *
4457  * Note that is it up to the module/driver to ensure that it itself
4458  * does not have any stale references to the closing queues once its close
4459  * routine returns. This includes:
4460  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4461  *    associated with the queues. For timeout and bufcall callbacks the
4462  *    module/driver also has to ensure (or wait for) any callbacks that
4463  *    are in progress.
4464  *  - If the module/driver is using esballoc it has to ensure that any
4465  *    esballoc free functions do not refer to a queue that has closed.
4466  *    (Note that in general the close routine can not wait for the esballoc'ed
4467  *    messages to be freed since that can cause a deadlock.)
4468  *  - Cancelling any interrupts that refer to the closing queues and
4469  *    also ensuring that there are no interrupts in progress that will
4470  *    refer to the closing queues once the close routine returns.
4471  *  - For multiplexors removing any driver global state that refers to
4472  *    the closing queue and also ensuring that there are no threads in
4473  *    the multiplexor that has picked up a queue pointer but not yet
4474  *    finished using it.
4475  *
4476  * In addition, a driver/module can only reference the q_next pointer
4477  * in its open, close, put, or service procedures or in a
4478  * qtimeout/qbufcall callback procedure executing "on" the correct
4479  * stream. Thus it can not reference the q_next pointer in an interrupt
4480  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4481  * it can not reference q_next of a different queue e.g. in a mux that
4482  * passes messages from one queues put/service procedure to another queue.
4483  * In all the cases when the driver/module can not access the q_next
4484  * field it must use the *next* versions e.g. canputnext instead of
4485  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4486  *
4487  *
4488  * Assuming that the driver/module conforms to the above constraints
4489  * the STREAMS framework has to avoid stale references to q_next for all
4490  * the framework internal cases which include (but are not limited to):
4491  *  - Threads in canput/canputnext/backenable and elsewhere that are
4492  *    walking q_next.
4493  *  - Messages on a syncq that have a reference to the queue through b_queue.
4494  *  - Messages on an outer perimeter (syncq) that have a reference to the
4495  *    queue through b_queue.
4496  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4497  *    Note that only canput and bcanput use q_nfsrv without any locking.
4498  *
4499  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4500  * after qprocsoff returns, the framework has to ensure that no threads can
4501  * enter the put or service routines for the closing read or write-side queue.
4502  * In addition to preventing "direct" entry into the put procedures
4503  * the framework also has to prevent messages being drained from
4504  * the syncq or the outer perimeter.
4505  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4506  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4507  * qprocsoff has returned.
4508  * Note that if a module/driver uses put(9F) on one of its own queues
4509  * it is up to the module/driver to ensure that the put() doesn't
4510  * get called when the queue is closing.
4511  *
4512  *
4513  * The framework aspects of the above "contract" is implemented by
4514  * qprocsoff, removeq, and strlock:
4515  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4516  *    entering the service procedures.
4517  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4518  *    canputnext, backenable etc from dereferencing the q_next that will
4519  *    soon change.
4520  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4521  *    or other q_next walker that uses claimstr/releasestr to finish.
4522  *  - optionally for every syncq in the stream strlock acquires all the
4523  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4524  *    that no thread executes in the put or service procedures and that no
4525  *    thread is draining into the module/driver. This ensures that no
4526  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4527  *    currently executing hence no such thread can end up with the old stale
4528  *    q_next value and no canput/backenable can have the old stale
4529  *    q_nfsrv/q_next.
4530  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4531  *    have either finished or observed the QWCLOSE flag and gone away.
4532  */
4533 
4534 
4535 /*
4536  * Get all the locks necessary to change q_next.
4537  *
4538  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4539  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4540  * the only threads inside the syncq are threads currently calling removeq().
4541  * Since threads calling removeq() are in the process of removing their queues
4542  * from the stream, we do not need to worry about them accessing a stale q_next
4543  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4544  * for them can cause deadlock).
4545  *
4546  * This routine is subject to starvation since it does not set any flag to
4547  * prevent threads from entering a module in the stream (i.e. sq_count can
4548  * increase on some syncq while it is waiting on some other syncq).
4549  *
4550  * Assumes that only one thread attempts to call strlock for a given
4551  * stream. If this is not the case the two threads would deadlock.
4552  * This assumption is guaranteed since strlock is only called by insertq
4553  * and removeq and streams plumbing changes are single-threaded for
4554  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4555  *
4556  * For pipes, it is not difficult to atomically designate a pair of streams
4557  * to be mated. Once mated atomically by the framework the twisted pair remain
4558  * configured that way until dismantled atomically by the framework.
4559  * When plumbing takes place on a twisted stream it is necessary to ensure that
4560  * this operation is done exclusively on the twisted stream since two such
4561  * operations, each initiated on different ends of the pipe will deadlock
4562  * waiting for each other to complete.
4563  *
4564  * On entry, no locks should be held.
4565  * The locks acquired and held by strlock depends on a few factors.
4566  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4567  *   and held on exit and all sq_count are at an acceptable level.
4568  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4569  *   sd_refcnt being zero.
4570  */
4571 
4572 static void
4573 strlock(struct stdata *stp, sqlist_t *sqlist)
4574 {
4575 	syncql_t *sql, *sql2;
4576 retry:
4577 	/*
4578 	 * Wait for any claimstr to go away.
4579 	 */
4580 	if (STRMATED(stp)) {
4581 		struct stdata *stp1, *stp2;
4582 
4583 		STRLOCKMATES(stp);
4584 		/*
4585 		 * Note that the selection of locking order is not
4586 		 * important, just that they are always acquired in
4587 		 * the same order.  To assure this, we choose this
4588 		 * order based on the value of the pointer, and since
4589 		 * the pointer will not change for the life of this
4590 		 * pair, we will always grab the locks in the same
4591 		 * order (and hence, prevent deadlocks).
4592 		 */
4593 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4594 			stp1 = stp;
4595 			stp2 = stp->sd_mate;
4596 		} else {
4597 			stp2 = stp;
4598 			stp1 = stp->sd_mate;
4599 		}
4600 		mutex_enter(&stp1->sd_reflock);
4601 		if (stp1->sd_refcnt > 0) {
4602 			STRUNLOCKMATES(stp);
4603 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4604 			mutex_exit(&stp1->sd_reflock);
4605 			goto retry;
4606 		}
4607 		mutex_enter(&stp2->sd_reflock);
4608 		if (stp2->sd_refcnt > 0) {
4609 			STRUNLOCKMATES(stp);
4610 			mutex_exit(&stp1->sd_reflock);
4611 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4612 			mutex_exit(&stp2->sd_reflock);
4613 			goto retry;
4614 		}
4615 		STREAM_PUTLOCKS_ENTER(stp1);
4616 		STREAM_PUTLOCKS_ENTER(stp2);
4617 	} else {
4618 		mutex_enter(&stp->sd_lock);
4619 		mutex_enter(&stp->sd_reflock);
4620 		while (stp->sd_refcnt > 0) {
4621 			mutex_exit(&stp->sd_lock);
4622 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4623 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4624 				mutex_exit(&stp->sd_reflock);
4625 				mutex_enter(&stp->sd_lock);
4626 				mutex_enter(&stp->sd_reflock);
4627 			}
4628 		}
4629 		STREAM_PUTLOCKS_ENTER(stp);
4630 	}
4631 
4632 	if (sqlist == NULL)
4633 		return;
4634 
4635 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4636 		syncq_t *sq = sql->sql_sq;
4637 		uint16_t count;
4638 
4639 		mutex_enter(SQLOCK(sq));
4640 		count = sq->sq_count;
4641 		ASSERT(sq->sq_rmqcount <= count);
4642 		SQ_PUTLOCKS_ENTER(sq);
4643 		SUM_SQ_PUTCOUNTS(sq, count);
4644 		if (count == sq->sq_rmqcount)
4645 			continue;
4646 
4647 		/* Failed - drop all locks that we have acquired so far */
4648 		if (STRMATED(stp)) {
4649 			STREAM_PUTLOCKS_EXIT(stp);
4650 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4651 			STRUNLOCKMATES(stp);
4652 			mutex_exit(&stp->sd_reflock);
4653 			mutex_exit(&stp->sd_mate->sd_reflock);
4654 		} else {
4655 			STREAM_PUTLOCKS_EXIT(stp);
4656 			mutex_exit(&stp->sd_lock);
4657 			mutex_exit(&stp->sd_reflock);
4658 		}
4659 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4660 		    sql2 = sql2->sql_next) {
4661 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4662 			mutex_exit(SQLOCK(sql2->sql_sq));
4663 		}
4664 
4665 		/*
4666 		 * The wait loop below may starve when there are many threads
4667 		 * claiming the syncq. This is especially a problem with permod
4668 		 * syncqs (IP). To lessen the impact of the problem we increment
4669 		 * sq_needexcl and clear fastbits so that putnexts will slow
4670 		 * down and call sqenable instead of draining right away.
4671 		 */
4672 		sq->sq_needexcl++;
4673 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4674 		while (count > sq->sq_rmqcount) {
4675 			sq->sq_flags |= SQ_WANTWAKEUP;
4676 			SQ_PUTLOCKS_EXIT(sq);
4677 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4678 			count = sq->sq_count;
4679 			SQ_PUTLOCKS_ENTER(sq);
4680 			SUM_SQ_PUTCOUNTS(sq, count);
4681 		}
4682 		sq->sq_needexcl--;
4683 		if (sq->sq_needexcl == 0)
4684 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4685 		SQ_PUTLOCKS_EXIT(sq);
4686 		ASSERT(count == sq->sq_rmqcount);
4687 		mutex_exit(SQLOCK(sq));
4688 		goto retry;
4689 	}
4690 }
4691 
4692 /*
4693  * Drop all the locks that strlock acquired.
4694  */
4695 static void
4696 strunlock(struct stdata *stp, sqlist_t *sqlist)
4697 {
4698 	syncql_t *sql;
4699 
4700 	if (STRMATED(stp)) {
4701 		STREAM_PUTLOCKS_EXIT(stp);
4702 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4703 		STRUNLOCKMATES(stp);
4704 		mutex_exit(&stp->sd_reflock);
4705 		mutex_exit(&stp->sd_mate->sd_reflock);
4706 	} else {
4707 		STREAM_PUTLOCKS_EXIT(stp);
4708 		mutex_exit(&stp->sd_lock);
4709 		mutex_exit(&stp->sd_reflock);
4710 	}
4711 
4712 	if (sqlist == NULL)
4713 		return;
4714 
4715 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4716 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4717 		mutex_exit(SQLOCK(sql->sql_sq));
4718 	}
4719 }
4720 
4721 /*
4722  * When the module has service procedure, we need check if the next
4723  * module which has service procedure is in flow control to trigger
4724  * the backenable.
4725  */
4726 static void
4727 backenable_insertedq(queue_t *q)
4728 {
4729 	qband_t	*qbp;
4730 
4731 	claimstr(q);
4732 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4733 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4734 			backenable(q, 0);
4735 
4736 		qbp = q->q_next->q_nfsrv->q_bandp;
4737 		for (; qbp != NULL; qbp = qbp->qb_next)
4738 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4739 				backenable(q, qbp->qb_first->b_band);
4740 	}
4741 	releasestr(q);
4742 }
4743 
4744 /*
4745  * Given two read queues, insert a new single one after another.
4746  *
4747  * This routine acquires all the necessary locks in order to change
4748  * q_next and related pointer using strlock().
4749  * It depends on the stream head ensuring that there are no concurrent
4750  * insertq or removeq on the same stream. The stream head ensures this
4751  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4752  *
4753  * Note that no syncq locks are held during the q_next change. This is
4754  * applied to all streams since, unlike removeq, there is no problem of stale
4755  * pointers when adding a module to the stream. Thus drivers/modules that do a
4756  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4757  * applied this optimization to all streams.
4758  */
4759 void
4760 insertq(struct stdata *stp, queue_t *new)
4761 {
4762 	queue_t	*after;
4763 	queue_t *wafter;
4764 	queue_t *wnew = _WR(new);
4765 	boolean_t have_fifo = B_FALSE;
4766 
4767 	if (new->q_flag & _QINSERTING) {
4768 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4769 		after = new->q_next;
4770 		wafter = _WR(new->q_next);
4771 	} else {
4772 		after = _RD(stp->sd_wrq);
4773 		wafter = stp->sd_wrq;
4774 	}
4775 
4776 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4777 	    "insertq:%p, %p", after, new);
4778 	ASSERT(after->q_flag & QREADR);
4779 	ASSERT(new->q_flag & QREADR);
4780 
4781 	strlock(stp, NULL);
4782 
4783 	/* Do we have a FIFO? */
4784 	if (wafter->q_next == after) {
4785 		have_fifo = B_TRUE;
4786 		wnew->q_next = new;
4787 	} else {
4788 		wnew->q_next = wafter->q_next;
4789 	}
4790 	new->q_next = after;
4791 
4792 	set_nfsrv_ptr(new, wnew, after, wafter);
4793 	/*
4794 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4795 	 * so only reset this flag after calling it.
4796 	 */
4797 	new->q_flag &= ~_QINSERTING;
4798 
4799 	if (have_fifo) {
4800 		wafter->q_next = wnew;
4801 	} else {
4802 		if (wafter->q_next)
4803 			_OTHERQ(wafter->q_next)->q_next = new;
4804 		wafter->q_next = wnew;
4805 	}
4806 
4807 	set_qend(new);
4808 	/* The QEND flag might have to be updated for the upstream guy */
4809 	set_qend(after);
4810 
4811 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4812 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4813 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4814 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4815 	strsetuio(stp);
4816 
4817 	/*
4818 	 * If this was a module insertion, bump the push count.
4819 	 */
4820 	if (!(new->q_flag & QISDRV))
4821 		stp->sd_pushcnt++;
4822 
4823 	strunlock(stp, NULL);
4824 
4825 	/* check if the write Q needs backenable */
4826 	backenable_insertedq(wnew);
4827 
4828 	/* check if the read Q needs backenable */
4829 	backenable_insertedq(new);
4830 }
4831 
4832 /*
4833  * Given a read queue, unlink it from any neighbors.
4834  *
4835  * This routine acquires all the necessary locks in order to
4836  * change q_next and related pointers and also guard against
4837  * stale references (e.g. through q_next) to the queue that
4838  * is being removed. It also plays part of the role in ensuring
4839  * that the module's/driver's put procedure doesn't get called
4840  * after qprocsoff returns.
4841  *
4842  * Removeq depends on the stream head ensuring that there are
4843  * no concurrent insertq or removeq on the same stream. The
4844  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4845  * STRPLUMB.
4846  *
4847  * The set of locks needed to remove the queue is different in
4848  * different cases:
4849  *
4850  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4851  * waiting for the syncq reference count to drop to 0 indicating that no
4852  * non-close threads are present anywhere in the stream. This ensures that any
4853  * module/driver can reference q_next in its open, close, put, or service
4854  * procedures.
4855  *
4856  * The sq_rmqcount counter tracks the number of threads inside removeq().
4857  * strlock() ensures that there is either no threads executing inside perimeter
4858  * or there is only a thread calling qprocsoff().
4859  *
4860  * strlock() compares the value of sq_count with the number of threads inside
4861  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4862  * any threads waiting in strlock() when the sq_rmqcount increases.
4863  */
4864 
4865 void
4866 removeq(queue_t *qp)
4867 {
4868 	queue_t *wqp = _WR(qp);
4869 	struct stdata *stp = STREAM(qp);
4870 	sqlist_t *sqlist = NULL;
4871 	boolean_t isdriver;
4872 	int moved;
4873 	syncq_t *sq = qp->q_syncq;
4874 	syncq_t *wsq = wqp->q_syncq;
4875 
4876 	ASSERT(stp);
4877 
4878 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4879 	    "removeq:%p %p", qp, wqp);
4880 	ASSERT(qp->q_flag&QREADR);
4881 
4882 	/*
4883 	 * For queues using Synchronous streams, we must wait for all threads in
4884 	 * rwnext() to drain out before proceeding.
4885 	 */
4886 	if (qp->q_flag & QSYNCSTR) {
4887 		/* First, we need wakeup any threads blocked in rwnext() */
4888 		mutex_enter(SQLOCK(sq));
4889 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4890 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4891 			cv_broadcast(&sq->sq_wait);
4892 		}
4893 		mutex_exit(SQLOCK(sq));
4894 
4895 		if (wsq != sq) {
4896 			mutex_enter(SQLOCK(wsq));
4897 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4898 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4899 				cv_broadcast(&wsq->sq_wait);
4900 			}
4901 			mutex_exit(SQLOCK(wsq));
4902 		}
4903 
4904 		mutex_enter(QLOCK(qp));
4905 		while (qp->q_rwcnt > 0) {
4906 			qp->q_flag |= QWANTRMQSYNC;
4907 			cv_wait(&qp->q_wait, QLOCK(qp));
4908 		}
4909 		mutex_exit(QLOCK(qp));
4910 
4911 		mutex_enter(QLOCK(wqp));
4912 		while (wqp->q_rwcnt > 0) {
4913 			wqp->q_flag |= QWANTRMQSYNC;
4914 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4915 		}
4916 		mutex_exit(QLOCK(wqp));
4917 	}
4918 
4919 	mutex_enter(SQLOCK(sq));
4920 	sq->sq_rmqcount++;
4921 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4922 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4923 		cv_broadcast(&sq->sq_wait);
4924 	}
4925 	mutex_exit(SQLOCK(sq));
4926 
4927 	isdriver = (qp->q_flag & QISDRV);
4928 
4929 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4930 	strlock(stp, sqlist);
4931 
4932 	reset_nfsrv_ptr(qp, wqp);
4933 
4934 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4935 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4936 	/* Do we have a FIFO? */
4937 	if (wqp->q_next == qp) {
4938 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4939 	} else {
4940 		if (wqp->q_next)
4941 			backq(qp)->q_next = qp->q_next;
4942 		if (qp->q_next)
4943 			backq(wqp)->q_next = wqp->q_next;
4944 	}
4945 
4946 	/* The QEND flag might have to be updated for the upstream guy */
4947 	if (qp->q_next)
4948 		set_qend(qp->q_next);
4949 
4950 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4951 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4952 
4953 	/*
4954 	 * Move any messages destined for the put procedures to the next
4955 	 * syncq in line. Otherwise free them.
4956 	 */
4957 	moved = 0;
4958 	/*
4959 	 * Quick check to see whether there are any messages or events.
4960 	 */
4961 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4962 		moved += propagate_syncq(qp);
4963 	if (wqp->q_syncqmsgs != 0 ||
4964 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4965 		moved += propagate_syncq(wqp);
4966 
4967 	strsetuio(stp);
4968 
4969 	/*
4970 	 * If this was a module removal, decrement the push count.
4971 	 */
4972 	if (!isdriver)
4973 		stp->sd_pushcnt--;
4974 
4975 	strunlock(stp, sqlist);
4976 	sqlist_free(sqlist);
4977 
4978 	/*
4979 	 * Make sure any messages that were propagated are drained.
4980 	 * Also clear any QFULL bit caused by messages that were propagated.
4981 	 */
4982 
4983 	if (qp->q_next != NULL) {
4984 		clr_qfull(qp);
4985 		/*
4986 		 * For the driver calling qprocsoff, propagate_syncq
4987 		 * frees all the messages instead of putting it in
4988 		 * the stream head
4989 		 */
4990 		if (!isdriver && (moved > 0))
4991 			emptysq(qp->q_next->q_syncq);
4992 	}
4993 	if (wqp->q_next != NULL) {
4994 		clr_qfull(wqp);
4995 		/*
4996 		 * We come here for any pop of a module except for the
4997 		 * case of driver being removed. We don't call emptysq
4998 		 * if we did not move any messages. This will avoid holding
4999 		 * PERMOD syncq locks in emptysq
5000 		 */
5001 		if (moved > 0)
5002 			emptysq(wqp->q_next->q_syncq);
5003 	}
5004 
5005 	mutex_enter(SQLOCK(sq));
5006 	sq->sq_rmqcount--;
5007 	mutex_exit(SQLOCK(sq));
5008 }
5009 
5010 /*
5011  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
5012  * SQ_WRITER) on a syncq.
5013  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
5014  * sync queue and waits until sq_count reaches maxcnt.
5015  *
5016  * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
5017  * does not care about putnext threads that are in the middle of calling put
5018  * entry points.
5019  *
5020  * This routine is used for both inner and outer syncqs.
5021  */
5022 static void
5023 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
5024 {
5025 	uint16_t count = 0;
5026 
5027 	mutex_enter(SQLOCK(sq));
5028 	/*
5029 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
5030 	 * SQ_FROZEN will be set if there is a frozen stream that has a
5031 	 * queue which also refers to this "shared" syncq.
5032 	 * SQ_BLOCKED will be set if there is "off" queue which also
5033 	 * refers to this "shared" syncq.
5034 	 */
5035 	if (maxcnt != -1) {
5036 		count = sq->sq_count;
5037 		SQ_PUTLOCKS_ENTER(sq);
5038 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5039 		SUM_SQ_PUTCOUNTS(sq, count);
5040 	}
5041 	sq->sq_needexcl++;
5042 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5043 
5044 	while ((sq->sq_flags & flag) ||
5045 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
5046 		sq->sq_flags |= SQ_WANTWAKEUP;
5047 		if (maxcnt != -1) {
5048 			SQ_PUTLOCKS_EXIT(sq);
5049 		}
5050 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5051 		if (maxcnt != -1) {
5052 			count = sq->sq_count;
5053 			SQ_PUTLOCKS_ENTER(sq);
5054 			SUM_SQ_PUTCOUNTS(sq, count);
5055 		}
5056 	}
5057 	sq->sq_needexcl--;
5058 	sq->sq_flags |= flag;
5059 	ASSERT(maxcnt == -1 || count == maxcnt);
5060 	if (maxcnt != -1) {
5061 		if (sq->sq_needexcl == 0) {
5062 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5063 		}
5064 		SQ_PUTLOCKS_EXIT(sq);
5065 	} else if (sq->sq_needexcl == 0) {
5066 		SQ_PUTCOUNT_SETFAST(sq);
5067 	}
5068 
5069 	mutex_exit(SQLOCK(sq));
5070 }
5071 
5072 /*
5073  * Reset a flag that was set with blocksq.
5074  *
5075  * Can not use this routine to reset SQ_WRITER.
5076  *
5077  * If "isouter" is set then the syncq is assumed to be an outer perimeter
5078  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5079  * to handle the queued qwriter operations.
5080  *
5081  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5082  * sq_putlocks are used.
5083  */
5084 static void
5085 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5086 {
5087 	uint16_t flags;
5088 
5089 	mutex_enter(SQLOCK(sq));
5090 	ASSERT(resetflag != SQ_WRITER);
5091 	ASSERT(sq->sq_flags & resetflag);
5092 	flags = sq->sq_flags & ~resetflag;
5093 	sq->sq_flags = flags;
5094 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5095 		if (flags & SQ_WANTWAKEUP) {
5096 			flags &= ~SQ_WANTWAKEUP;
5097 			cv_broadcast(&sq->sq_wait);
5098 		}
5099 		sq->sq_flags = flags;
5100 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5101 			if (!isouter) {
5102 				/* drain_syncq drops SQLOCK */
5103 				drain_syncq(sq);
5104 				return;
5105 			}
5106 		}
5107 	}
5108 	mutex_exit(SQLOCK(sq));
5109 }
5110 
5111 /*
5112  * Reset a flag that was set with blocksq.
5113  * Does not drain the syncq. Use emptysq() for that.
5114  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5115  *
5116  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5117  * sq_putlocks are used.
5118  */
5119 static int
5120 dropsq(syncq_t *sq, uint16_t resetflag)
5121 {
5122 	uint16_t flags;
5123 
5124 	mutex_enter(SQLOCK(sq));
5125 	ASSERT(sq->sq_flags & resetflag);
5126 	flags = sq->sq_flags & ~resetflag;
5127 	if (flags & SQ_WANTWAKEUP) {
5128 		flags &= ~SQ_WANTWAKEUP;
5129 		cv_broadcast(&sq->sq_wait);
5130 	}
5131 	sq->sq_flags = flags;
5132 	mutex_exit(SQLOCK(sq));
5133 	if (flags & SQ_QUEUED)
5134 		return (1);
5135 	return (0);
5136 }
5137 
5138 /*
5139  * Empty all the messages on a syncq.
5140  *
5141  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5142  * sq_putlocks are used.
5143  */
5144 static void
5145 emptysq(syncq_t *sq)
5146 {
5147 	uint16_t flags;
5148 
5149 	mutex_enter(SQLOCK(sq));
5150 	flags = sq->sq_flags;
5151 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5152 		/*
5153 		 * To prevent potential recursive invocation of drain_syncq we
5154 		 * do not call drain_syncq if count is non-zero.
5155 		 */
5156 		if (sq->sq_count == 0) {
5157 			/* drain_syncq() drops SQLOCK */
5158 			drain_syncq(sq);
5159 			return;
5160 		} else
5161 			sqenable(sq);
5162 	}
5163 	mutex_exit(SQLOCK(sq));
5164 }
5165 
5166 /*
5167  * Ordered insert while removing duplicates.
5168  */
5169 static void
5170 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5171 {
5172 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5173 
5174 	prev_sqlpp = &sqlist->sqlist_head;
5175 	while ((sqlp = *prev_sqlpp) != NULL) {
5176 		if (sqlp->sql_sq >= sqp) {
5177 			if (sqlp->sql_sq == sqp)	/* duplicate */
5178 				return;
5179 			break;
5180 		}
5181 		prev_sqlpp = &sqlp->sql_next;
5182 	}
5183 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5184 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5185 	new_sqlp->sql_next = sqlp;
5186 	new_sqlp->sql_sq = sqp;
5187 	*prev_sqlpp = new_sqlp;
5188 }
5189 
5190 /*
5191  * Walk the write side queues until we hit either the driver
5192  * or a twist in the stream (_SAMESTR will return false in both
5193  * these cases) then turn around and walk the read side queues
5194  * back up to the stream head.
5195  */
5196 static void
5197 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5198 {
5199 	while (q != NULL) {
5200 		sqlist_insert(sqlist, q->q_syncq);
5201 
5202 		if (_SAMESTR(q))
5203 			q = q->q_next;
5204 		else if (!(q->q_flag & QREADR))
5205 			q = _RD(q);
5206 		else
5207 			q = NULL;
5208 	}
5209 }
5210 
5211 /*
5212  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5213  * associated with the "q" parameter. The resulting list is sorted in a
5214  * canonical order and is free of duplicates.
5215  * Assumes the passed queue is a _RD(q).
5216  */
5217 static sqlist_t *
5218 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5219 {
5220 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5221 
5222 	/*
5223 	 * start with the current queue/qpair
5224 	 */
5225 	ASSERT(q->q_flag & QREADR);
5226 
5227 	sqlist_insert(sqlist, q->q_syncq);
5228 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5229 
5230 	sqlist_insertall(sqlist, stp->sd_wrq);
5231 	if (do_twist)
5232 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5233 
5234 	return (sqlist);
5235 }
5236 
5237 static sqlist_t *
5238 sqlist_alloc(struct stdata *stp, int kmflag)
5239 {
5240 	size_t sqlist_size;
5241 	sqlist_t *sqlist;
5242 
5243 	/*
5244 	 * Allocate 2 syncql_t's for each pushed module. Note that
5245 	 * the sqlist_t structure already has 4 syncql_t's built in:
5246 	 * 2 for the stream head, and 2 for the driver/other stream head.
5247 	 */
5248 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5249 	    sizeof (sqlist_t);
5250 	if (STRMATED(stp))
5251 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5252 	sqlist = kmem_alloc(sqlist_size, kmflag);
5253 
5254 	sqlist->sqlist_head = NULL;
5255 	sqlist->sqlist_size = sqlist_size;
5256 	sqlist->sqlist_index = 0;
5257 
5258 	return (sqlist);
5259 }
5260 
5261 /*
5262  * Free the list created by sqlist_alloc()
5263  */
5264 static void
5265 sqlist_free(sqlist_t *sqlist)
5266 {
5267 	kmem_free(sqlist, sqlist->sqlist_size);
5268 }
5269 
5270 /*
5271  * Prevent any new entries into any syncq in this stream.
5272  * Used by freezestr.
5273  */
5274 void
5275 strblock(queue_t *q)
5276 {
5277 	struct stdata	*stp;
5278 	syncql_t	*sql;
5279 	sqlist_t	*sqlist;
5280 
5281 	q = _RD(q);
5282 
5283 	stp = STREAM(q);
5284 	ASSERT(stp != NULL);
5285 
5286 	/*
5287 	 * Get a sorted list with all the duplicates removed containing
5288 	 * all the syncqs referenced by this stream.
5289 	 */
5290 	sqlist = sqlist_build(q, stp, B_FALSE);
5291 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5292 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5293 	sqlist_free(sqlist);
5294 }
5295 
5296 /*
5297  * Release the block on new entries into this stream
5298  */
5299 void
5300 strunblock(queue_t *q)
5301 {
5302 	struct stdata	*stp;
5303 	syncql_t	*sql;
5304 	sqlist_t	*sqlist;
5305 	int		drain_needed;
5306 
5307 	q = _RD(q);
5308 
5309 	/*
5310 	 * Get a sorted list with all the duplicates removed containing
5311 	 * all the syncqs referenced by this stream.
5312 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5313 	 * starting to drain them; otherwise the draining might
5314 	 * cause a freezestr in some module on the stream (which
5315 	 * would deadlock).
5316 	 */
5317 	stp = STREAM(q);
5318 	ASSERT(stp != NULL);
5319 	sqlist = sqlist_build(q, stp, B_FALSE);
5320 	drain_needed = 0;
5321 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5322 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5323 	if (drain_needed) {
5324 		for (sql = sqlist->sqlist_head; sql != NULL;
5325 		    sql = sql->sql_next)
5326 			emptysq(sql->sql_sq);
5327 	}
5328 	sqlist_free(sqlist);
5329 }
5330 
5331 #ifdef DEBUG
5332 static int
5333 qprocsareon(queue_t *rq)
5334 {
5335 	if (rq->q_next == NULL)
5336 		return (0);
5337 	return (_WR(rq->q_next)->q_next == _WR(rq));
5338 }
5339 
5340 int
5341 qclaimed(queue_t *q)
5342 {
5343 	uint_t count;
5344 
5345 	count = q->q_syncq->sq_count;
5346 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5347 	return (count != 0);
5348 }
5349 
5350 /*
5351  * Check if anyone has frozen this stream with freezestr
5352  */
5353 int
5354 frozenstr(queue_t *q)
5355 {
5356 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5357 }
5358 #endif /* DEBUG */
5359 
5360 /*
5361  * Enter a queue.
5362  * Obsoleted interface. Should not be used.
5363  */
5364 void
5365 enterq(queue_t *q)
5366 {
5367 	entersq(q->q_syncq, SQ_CALLBACK);
5368 }
5369 
5370 void
5371 leaveq(queue_t *q)
5372 {
5373 	leavesq(q->q_syncq, SQ_CALLBACK);
5374 }
5375 
5376 /*
5377  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5378  * to check.
5379  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5380  * calls and the running of open, close and service procedures.
5381  *
5382  * If c_inner bit is set no need to grab sq_putlocks since we don't care
5383  * if other threads have entered or are entering put entry point.
5384  *
5385  * If c_inner bit is set it might have been possible to use
5386  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5387  * open/close path for IP) but since the count may need to be decremented in
5388  * qwait() we wouldn't know which counter to decrement. Currently counter is
5389  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5390  * in the future we might use curthread id bits to select the counter and this
5391  * would stay constant across routine calls.
5392  */
5393 void
5394 entersq(syncq_t *sq, int entrypoint)
5395 {
5396 	uint16_t	count = 0;
5397 	uint16_t	flags;
5398 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5399 	uint16_t	type;
5400 	uint_t		c_inner = entrypoint & SQ_CI;
5401 	uint_t		c_outer = entrypoint & SQ_CO;
5402 
5403 	/*
5404 	 * Increment ref count to keep closes out of this queue.
5405 	 */
5406 	ASSERT(sq);
5407 	ASSERT(c_inner && c_outer);
5408 	mutex_enter(SQLOCK(sq));
5409 	flags = sq->sq_flags;
5410 	type = sq->sq_type;
5411 	if (!(type & c_inner)) {
5412 		/* Make sure all putcounts now use slowlock. */
5413 		count = sq->sq_count;
5414 		SQ_PUTLOCKS_ENTER(sq);
5415 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5416 		SUM_SQ_PUTCOUNTS(sq, count);
5417 		sq->sq_needexcl++;
5418 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5419 		waitflags |= SQ_MESSAGES;
5420 	}
5421 	/*
5422 	 * Wait until we can enter the inner perimeter.
5423 	 * If we want exclusive access we wait until sq_count is 0.
5424 	 * We have to do this before entering the outer perimeter in order
5425 	 * to preserve put/close message ordering.
5426 	 */
5427 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5428 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5429 		if (!(type & c_inner)) {
5430 			SQ_PUTLOCKS_EXIT(sq);
5431 		}
5432 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5433 		if (!(type & c_inner)) {
5434 			count = sq->sq_count;
5435 			SQ_PUTLOCKS_ENTER(sq);
5436 			SUM_SQ_PUTCOUNTS(sq, count);
5437 		}
5438 		flags = sq->sq_flags;
5439 	}
5440 
5441 	if (!(type & c_inner)) {
5442 		ASSERT(sq->sq_needexcl > 0);
5443 		sq->sq_needexcl--;
5444 		if (sq->sq_needexcl == 0) {
5445 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5446 		}
5447 	}
5448 
5449 	/* Check if we need to enter the outer perimeter */
5450 	if (!(type & c_outer)) {
5451 		/*
5452 		 * We have to enter the outer perimeter exclusively before
5453 		 * we can increment sq_count to avoid deadlock. This implies
5454 		 * that we have to re-check sq_flags and sq_count.
5455 		 *
5456 		 * is it possible to have c_inner set when c_outer is not set?
5457 		 */
5458 		if (!(type & c_inner)) {
5459 			SQ_PUTLOCKS_EXIT(sq);
5460 		}
5461 		mutex_exit(SQLOCK(sq));
5462 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5463 		mutex_enter(SQLOCK(sq));
5464 		flags = sq->sq_flags;
5465 		/*
5466 		 * there should be no need to recheck sq_putcounts
5467 		 * because outer_enter() has already waited for them to clear
5468 		 * after setting SQ_WRITER.
5469 		 */
5470 		count = sq->sq_count;
5471 #ifdef DEBUG
5472 		/*
5473 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5474 		 * of doing an ASSERT internally. Others should do
5475 		 * something like
5476 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5477 		 * without the need to #ifdef DEBUG it.
5478 		 */
5479 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5480 #endif
5481 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5482 		    (!(type & c_inner) && count != 0)) {
5483 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5484 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5485 			count = sq->sq_count;
5486 			flags = sq->sq_flags;
5487 		}
5488 	}
5489 
5490 	sq->sq_count++;
5491 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5492 	if (!(type & c_inner)) {
5493 		/* Exclusive entry */
5494 		ASSERT(sq->sq_count == 1);
5495 		sq->sq_flags |= SQ_EXCL;
5496 		if (type & c_outer) {
5497 			SQ_PUTLOCKS_EXIT(sq);
5498 		}
5499 	}
5500 	mutex_exit(SQLOCK(sq));
5501 }
5502 
5503 /*
5504  * Leave a syncq. Announce to framework that closes may proceed.
5505  * c_inner and c_outer specify which concurrency bits to check.
5506  *
5507  * Must never be called from driver or module put entry point.
5508  *
5509  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5510  * sq_putlocks are used.
5511  */
5512 void
5513 leavesq(syncq_t *sq, int entrypoint)
5514 {
5515 	uint16_t	flags;
5516 	uint16_t	type;
5517 	uint_t		c_outer = entrypoint & SQ_CO;
5518 #ifdef DEBUG
5519 	uint_t		c_inner = entrypoint & SQ_CI;
5520 #endif
5521 
5522 	/*
5523 	 * Decrement ref count, drain the syncq if possible, and wake up
5524 	 * any waiting close.
5525 	 */
5526 	ASSERT(sq);
5527 	ASSERT(c_inner && c_outer);
5528 	mutex_enter(SQLOCK(sq));
5529 	flags = sq->sq_flags;
5530 	type = sq->sq_type;
5531 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5532 
5533 		if (flags & SQ_WANTWAKEUP) {
5534 			flags &= ~SQ_WANTWAKEUP;
5535 			cv_broadcast(&sq->sq_wait);
5536 		}
5537 		if (flags & SQ_WANTEXWAKEUP) {
5538 			flags &= ~SQ_WANTEXWAKEUP;
5539 			cv_broadcast(&sq->sq_exitwait);
5540 		}
5541 
5542 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5543 			/*
5544 			 * The syncq needs to be drained. "Exit" the syncq
5545 			 * before calling drain_syncq.
5546 			 */
5547 			ASSERT(sq->sq_count != 0);
5548 			sq->sq_count--;
5549 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5550 			sq->sq_flags = flags & ~SQ_EXCL;
5551 			drain_syncq(sq);
5552 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5553 			/* Check if we need to exit the outer perimeter */
5554 			/* XXX will this ever be true? */
5555 			if (!(type & c_outer))
5556 				outer_exit(sq->sq_outer);
5557 			return;
5558 		}
5559 	}
5560 	ASSERT(sq->sq_count != 0);
5561 	sq->sq_count--;
5562 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5563 	sq->sq_flags = flags & ~SQ_EXCL;
5564 	mutex_exit(SQLOCK(sq));
5565 
5566 	/* Check if we need to exit the outer perimeter */
5567 	if (!(sq->sq_type & c_outer))
5568 		outer_exit(sq->sq_outer);
5569 }
5570 
5571 /*
5572  * Prevent q_next from changing in this stream by incrementing sq_count.
5573  *
5574  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5575  * sq_putlocks are used.
5576  */
5577 void
5578 claimq(queue_t *qp)
5579 {
5580 	syncq_t	*sq = qp->q_syncq;
5581 
5582 	mutex_enter(SQLOCK(sq));
5583 	sq->sq_count++;
5584 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5585 	mutex_exit(SQLOCK(sq));
5586 }
5587 
5588 /*
5589  * Undo claimq.
5590  *
5591  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5592  * sq_putlocks are used.
5593  */
5594 void
5595 releaseq(queue_t *qp)
5596 {
5597 	syncq_t	*sq = qp->q_syncq;
5598 	uint16_t flags;
5599 
5600 	mutex_enter(SQLOCK(sq));
5601 	ASSERT(sq->sq_count > 0);
5602 	sq->sq_count--;
5603 
5604 	flags = sq->sq_flags;
5605 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5606 		if (flags & SQ_WANTWAKEUP) {
5607 			flags &= ~SQ_WANTWAKEUP;
5608 			cv_broadcast(&sq->sq_wait);
5609 		}
5610 		sq->sq_flags = flags;
5611 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5612 			/*
5613 			 * To prevent potential recursive invocation of
5614 			 * drain_syncq we do not call drain_syncq if count is
5615 			 * non-zero.
5616 			 */
5617 			if (sq->sq_count == 0) {
5618 				drain_syncq(sq);
5619 				return;
5620 			} else
5621 				sqenable(sq);
5622 		}
5623 	}
5624 	mutex_exit(SQLOCK(sq));
5625 }
5626 
5627 /*
5628  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5629  */
5630 void
5631 claimstr(queue_t *qp)
5632 {
5633 	struct stdata *stp = STREAM(qp);
5634 
5635 	mutex_enter(&stp->sd_reflock);
5636 	stp->sd_refcnt++;
5637 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5638 	mutex_exit(&stp->sd_reflock);
5639 }
5640 
5641 /*
5642  * Undo claimstr.
5643  */
5644 void
5645 releasestr(queue_t *qp)
5646 {
5647 	struct stdata *stp = STREAM(qp);
5648 
5649 	mutex_enter(&stp->sd_reflock);
5650 	ASSERT(stp->sd_refcnt != 0);
5651 	if (--stp->sd_refcnt == 0)
5652 		cv_broadcast(&stp->sd_refmonitor);
5653 	mutex_exit(&stp->sd_reflock);
5654 }
5655 
5656 static syncq_t *
5657 new_syncq(void)
5658 {
5659 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5660 }
5661 
5662 static void
5663 free_syncq(syncq_t *sq)
5664 {
5665 	ASSERT(sq->sq_head == NULL);
5666 	ASSERT(sq->sq_outer == NULL);
5667 	ASSERT(sq->sq_callbpend == NULL);
5668 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5669 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5670 
5671 	if (sq->sq_ciputctrl != NULL) {
5672 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5673 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5674 		    sq->sq_nciputctrl, 0);
5675 		ASSERT(ciputctrl_cache != NULL);
5676 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5677 	}
5678 
5679 	sq->sq_tail = NULL;
5680 	sq->sq_evhead = NULL;
5681 	sq->sq_evtail = NULL;
5682 	sq->sq_ciputctrl = NULL;
5683 	sq->sq_nciputctrl = 0;
5684 	sq->sq_count = 0;
5685 	sq->sq_rmqcount = 0;
5686 	sq->sq_callbflags = 0;
5687 	sq->sq_cancelid = 0;
5688 	sq->sq_next = NULL;
5689 	sq->sq_needexcl = 0;
5690 	sq->sq_svcflags = 0;
5691 	sq->sq_nqueues = 0;
5692 	sq->sq_pri = 0;
5693 	sq->sq_onext = NULL;
5694 	sq->sq_oprev = NULL;
5695 	sq->sq_flags = 0;
5696 	sq->sq_type = 0;
5697 	sq->sq_servcount = 0;
5698 
5699 	kmem_cache_free(syncq_cache, sq);
5700 }
5701 
5702 /* Outer perimeter code */
5703 
5704 /*
5705  * The outer syncq uses the fields and flags in the syncq slightly
5706  * differently from the inner syncqs.
5707  *	sq_count	Incremented when there are pending or running
5708  *			writers at the outer perimeter to prevent the set of
5709  *			inner syncqs that belong to the outer perimeter from
5710  *			changing.
5711  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5712  *
5713  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5714  *			inner syncqs are added to or removed from the
5715  *			outer perimeter.
5716  *	SQ_QUEUED	sq_head/tail has messages or events queued.
5717  *
5718  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5719  *			setting the SQ_WRITER flag.
5720  */
5721 
5722 /*
5723  * Get write access at the outer perimeter.
5724  * Note that read access is done by entersq, putnext, and put by simply
5725  * incrementing sq_count in the inner syncq.
5726  *
5727  * Waits until "flags" is no longer set in the outer to prevent multiple
5728  * threads from having write access at the same time. SQ_WRITER has to be part
5729  * of "flags".
5730  *
5731  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5732  * until the outer_exit is finished.
5733  *
5734  * outer_enter is vulnerable to starvation since it does not prevent new
5735  * threads from entering the inner syncqs while it is waiting for sq_count to
5736  * go to zero.
5737  */
5738 void
5739 outer_enter(syncq_t *outer, uint16_t flags)
5740 {
5741 	syncq_t	*sq;
5742 	int	wait_needed;
5743 	uint16_t	count;
5744 
5745 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5746 	    outer->sq_oprev != NULL);
5747 	ASSERT(flags & SQ_WRITER);
5748 
5749 retry:
5750 	mutex_enter(SQLOCK(outer));
5751 	while (outer->sq_flags & flags) {
5752 		outer->sq_flags |= SQ_WANTWAKEUP;
5753 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5754 	}
5755 
5756 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5757 	outer->sq_flags |= SQ_WRITER;
5758 	outer->sq_count++;
5759 	ASSERT(outer->sq_count != 0);	/* wraparound */
5760 	wait_needed = 0;
5761 	/*
5762 	 * Set SQ_WRITER on all the inner syncqs while holding
5763 	 * the SQLOCK on the outer syncq. This ensures that the changing
5764 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5765 	 */
5766 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5767 		mutex_enter(SQLOCK(sq));
5768 		count = sq->sq_count;
5769 		SQ_PUTLOCKS_ENTER(sq);
5770 		sq->sq_flags |= SQ_WRITER;
5771 		SUM_SQ_PUTCOUNTS(sq, count);
5772 		if (count != 0)
5773 			wait_needed = 1;
5774 		SQ_PUTLOCKS_EXIT(sq);
5775 		mutex_exit(SQLOCK(sq));
5776 	}
5777 	mutex_exit(SQLOCK(outer));
5778 
5779 	/*
5780 	 * Get everybody out of the syncqs sequentially.
5781 	 * Note that we don't actually need to acquire the PUTLOCKS, since
5782 	 * we have already cleared the fastbit, and set QWRITER.  By
5783 	 * definition, the count can not increase since putnext will
5784 	 * take the slowlock path (and the purpose of acquiring the
5785 	 * putlocks was to make sure it didn't increase while we were
5786 	 * waiting).
5787 	 *
5788 	 * Note that we still acquire the PUTLOCKS to be safe.
5789 	 */
5790 	if (wait_needed) {
5791 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5792 			mutex_enter(SQLOCK(sq));
5793 			count = sq->sq_count;
5794 			SQ_PUTLOCKS_ENTER(sq);
5795 			SUM_SQ_PUTCOUNTS(sq, count);
5796 			while (count != 0) {
5797 				sq->sq_flags |= SQ_WANTWAKEUP;
5798 				SQ_PUTLOCKS_EXIT(sq);
5799 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5800 				count = sq->sq_count;
5801 				SQ_PUTLOCKS_ENTER(sq);
5802 				SUM_SQ_PUTCOUNTS(sq, count);
5803 			}
5804 			SQ_PUTLOCKS_EXIT(sq);
5805 			mutex_exit(SQLOCK(sq));
5806 		}
5807 		/*
5808 		 * Verify that none of the flags got set while we
5809 		 * were waiting for the sq_counts to drop.
5810 		 * If this happens we exit and retry entering the
5811 		 * outer perimeter.
5812 		 */
5813 		mutex_enter(SQLOCK(outer));
5814 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5815 			mutex_exit(SQLOCK(outer));
5816 			outer_exit(outer);
5817 			goto retry;
5818 		}
5819 		mutex_exit(SQLOCK(outer));
5820 	}
5821 }
5822 
5823 /*
5824  * Drop the write access at the outer perimeter.
5825  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5826  * decrementing sq_count.
5827  */
5828 void
5829 outer_exit(syncq_t *outer)
5830 {
5831 	syncq_t	*sq;
5832 	int	 drain_needed;
5833 	uint16_t flags;
5834 
5835 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5836 	    outer->sq_oprev != NULL);
5837 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5838 
5839 	/*
5840 	 * Atomically (from the perspective of threads calling become_writer)
5841 	 * drop the write access at the outer perimeter by holding
5842 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5843 	 * SQ_WRITER.
5844 	 * This defines a locking order between the outer perimeter
5845 	 * SQLOCK and the inner perimeter SQLOCKs.
5846 	 */
5847 	mutex_enter(SQLOCK(outer));
5848 	flags = outer->sq_flags;
5849 	ASSERT(outer->sq_flags & SQ_WRITER);
5850 	if (flags & SQ_QUEUED) {
5851 		write_now(outer);
5852 		flags = outer->sq_flags;
5853 	}
5854 
5855 	/*
5856 	 * sq_onext is stable since sq_count has not yet been decreased.
5857 	 * Reset the SQ_WRITER flags in all syncqs.
5858 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5859 	 * inner syncqs.
5860 	 */
5861 	drain_needed = 0;
5862 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5863 		drain_needed += dropsq(sq, SQ_WRITER);
5864 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5865 	flags &= ~SQ_WRITER;
5866 	if (drain_needed) {
5867 		outer->sq_flags = flags;
5868 		mutex_exit(SQLOCK(outer));
5869 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5870 			emptysq(sq);
5871 		mutex_enter(SQLOCK(outer));
5872 		flags = outer->sq_flags;
5873 	}
5874 	if (flags & SQ_WANTWAKEUP) {
5875 		flags &= ~SQ_WANTWAKEUP;
5876 		cv_broadcast(&outer->sq_wait);
5877 	}
5878 	outer->sq_flags = flags;
5879 	ASSERT(outer->sq_count > 0);
5880 	outer->sq_count--;
5881 	mutex_exit(SQLOCK(outer));
5882 }
5883 
5884 /*
5885  * Add another syncq to an outer perimeter.
5886  * Block out all other access to the outer perimeter while it is being
5887  * changed using blocksq.
5888  * Assumes that the caller has *not* done an outer_enter.
5889  *
5890  * Vulnerable to starvation in blocksq.
5891  */
5892 static void
5893 outer_insert(syncq_t *outer, syncq_t *sq)
5894 {
5895 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5896 	    outer->sq_oprev != NULL);
5897 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5898 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5899 
5900 	/* Get exclusive access to the outer perimeter list */
5901 	blocksq(outer, SQ_BLOCKED, 0);
5902 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5903 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5904 
5905 	mutex_enter(SQLOCK(sq));
5906 	sq->sq_outer = outer;
5907 	outer->sq_onext->sq_oprev = sq;
5908 	sq->sq_onext = outer->sq_onext;
5909 	outer->sq_onext = sq;
5910 	sq->sq_oprev = outer;
5911 	mutex_exit(SQLOCK(sq));
5912 	unblocksq(outer, SQ_BLOCKED, 1);
5913 }
5914 
5915 /*
5916  * Remove a syncq from an outer perimeter.
5917  * Block out all other access to the outer perimeter while it is being
5918  * changed using blocksq.
5919  * Assumes that the caller has *not* done an outer_enter.
5920  *
5921  * Vulnerable to starvation in blocksq.
5922  */
5923 static void
5924 outer_remove(syncq_t *outer, syncq_t *sq)
5925 {
5926 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5927 	    outer->sq_oprev != NULL);
5928 	ASSERT(sq->sq_outer == outer);
5929 
5930 	/* Get exclusive access to the outer perimeter list */
5931 	blocksq(outer, SQ_BLOCKED, 0);
5932 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5933 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5934 
5935 	mutex_enter(SQLOCK(sq));
5936 	sq->sq_outer = NULL;
5937 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5938 	sq->sq_oprev->sq_onext = sq->sq_onext;
5939 	sq->sq_oprev = sq->sq_onext = NULL;
5940 	mutex_exit(SQLOCK(sq));
5941 	unblocksq(outer, SQ_BLOCKED, 1);
5942 }
5943 
5944 /*
5945  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5946  * If this is the first callback for this outer perimeter then add
5947  * this outer perimeter to the list of outer perimeters that
5948  * the qwriter_outer_thread will process.
5949  *
5950  * Increments sq_count in the outer syncq to prevent the membership
5951  * of the outer perimeter (in terms of inner syncqs) to change while
5952  * the callback is pending.
5953  */
5954 static void
5955 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5956 {
5957 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5958 
5959 	mp->b_prev = (mblk_t *)func;
5960 	mp->b_queue = q;
5961 	mp->b_next = NULL;
5962 	outer->sq_count++;	/* Decremented when dequeued */
5963 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5964 	if (outer->sq_evhead == NULL) {
5965 		/* First message. */
5966 		outer->sq_evhead = outer->sq_evtail = mp;
5967 		outer->sq_flags |= SQ_EVENTS;
5968 		mutex_exit(SQLOCK(outer));
5969 		STRSTAT(qwr_outer);
5970 		(void) taskq_dispatch(streams_taskq,
5971 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5972 	} else {
5973 		ASSERT(outer->sq_flags & SQ_EVENTS);
5974 		outer->sq_evtail->b_next = mp;
5975 		outer->sq_evtail = mp;
5976 		mutex_exit(SQLOCK(outer));
5977 	}
5978 }
5979 
5980 /*
5981  * Try and upgrade to write access at the outer perimeter. If this can
5982  * not be done without blocking then queue the callback to be done
5983  * by the qwriter_outer_thread.
5984  *
5985  * This routine can only be called from put or service procedures plus
5986  * asynchronous callback routines that have properly entered the queue (with
5987  * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
5988  * associated with q.
5989  */
5990 void
5991 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5992 {
5993 	syncq_t	*osq, *sq, *outer;
5994 	int	failed;
5995 	uint16_t flags;
5996 
5997 	osq = q->q_syncq;
5998 	outer = osq->sq_outer;
5999 	if (outer == NULL)
6000 		panic("qwriter(PERIM_OUTER): no outer perimeter");
6001 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6002 	    outer->sq_oprev != NULL);
6003 
6004 	mutex_enter(SQLOCK(outer));
6005 	flags = outer->sq_flags;
6006 	/*
6007 	 * If some thread is traversing sq_next, or if we are blocked by
6008 	 * outer_insert or outer_remove, or if the we already have queued
6009 	 * callbacks, then queue this callback for later processing.
6010 	 *
6011 	 * Also queue the qwriter for an interrupt thread in order
6012 	 * to reduce the time spent running at high IPL.
6013 	 * to identify there are events.
6014 	 */
6015 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
6016 		/*
6017 		 * Queue the become_writer request.
6018 		 * The queueing is atomic under SQLOCK(outer) in order
6019 		 * to synchronize with outer_exit.
6020 		 * queue_writer will drop the outer SQLOCK
6021 		 */
6022 		if (flags & SQ_BLOCKED) {
6023 			/* Must set SQ_WRITER on inner perimeter */
6024 			mutex_enter(SQLOCK(osq));
6025 			osq->sq_flags |= SQ_WRITER;
6026 			mutex_exit(SQLOCK(osq));
6027 		} else {
6028 			if (!(flags & SQ_WRITER)) {
6029 				/*
6030 				 * The outer could have been SQ_BLOCKED thus
6031 				 * SQ_WRITER might not be set on the inner.
6032 				 */
6033 				mutex_enter(SQLOCK(osq));
6034 				osq->sq_flags |= SQ_WRITER;
6035 				mutex_exit(SQLOCK(osq));
6036 			}
6037 			ASSERT(osq->sq_flags & SQ_WRITER);
6038 		}
6039 		queue_writer(outer, func, q, mp);
6040 		return;
6041 	}
6042 	/*
6043 	 * We are half-way to exclusive access to the outer perimeter.
6044 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
6045 	 * while the inner syncqs are traversed.
6046 	 */
6047 	outer->sq_count++;
6048 	ASSERT(outer->sq_count != 0);	/* wraparound */
6049 	flags |= SQ_WRITER;
6050 	/*
6051 	 * Check if we can run the function immediately. Mark all
6052 	 * syncqs with the writer flag to prevent new entries into
6053 	 * put and service procedures.
6054 	 *
6055 	 * Set SQ_WRITER on all the inner syncqs while holding
6056 	 * the SQLOCK on the outer syncq. This ensures that the changing
6057 	 * of SQ_WRITER is atomic under the outer SQLOCK.
6058 	 */
6059 	failed = 0;
6060 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6061 		uint16_t count;
6062 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
6063 
6064 		mutex_enter(SQLOCK(sq));
6065 		count = sq->sq_count;
6066 		SQ_PUTLOCKS_ENTER(sq);
6067 		SUM_SQ_PUTCOUNTS(sq, count);
6068 		if (sq->sq_count > maxcnt)
6069 			failed = 1;
6070 		sq->sq_flags |= SQ_WRITER;
6071 		SQ_PUTLOCKS_EXIT(sq);
6072 		mutex_exit(SQLOCK(sq));
6073 	}
6074 	if (failed) {
6075 		/*
6076 		 * Some other thread has a read claim on the outer perimeter.
6077 		 * Queue the callback for deferred processing.
6078 		 *
6079 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6080 		 * so that other qwriter(OUTER) calls will queue their
6081 		 * callbacks as well. queue_writer increments sq_count so we
6082 		 * decrement to compensate for the our increment.
6083 		 *
6084 		 * Dropping SQ_WRITER enables the writer thread to work
6085 		 * on this outer perimeter.
6086 		 */
6087 		outer->sq_flags = flags;
6088 		queue_writer(outer, func, q, mp);
6089 		/* queue_writer dropper the lock */
6090 		mutex_enter(SQLOCK(outer));
6091 		ASSERT(outer->sq_count > 0);
6092 		outer->sq_count--;
6093 		ASSERT(outer->sq_flags & SQ_WRITER);
6094 		flags = outer->sq_flags;
6095 		flags &= ~SQ_WRITER;
6096 		if (flags & SQ_WANTWAKEUP) {
6097 			flags &= ~SQ_WANTWAKEUP;
6098 			cv_broadcast(&outer->sq_wait);
6099 		}
6100 		outer->sq_flags = flags;
6101 		mutex_exit(SQLOCK(outer));
6102 		return;
6103 	} else {
6104 		outer->sq_flags = flags;
6105 		mutex_exit(SQLOCK(outer));
6106 	}
6107 
6108 	/* Can run it immediately */
6109 	(*func)(q, mp);
6110 
6111 	outer_exit(outer);
6112 }
6113 
6114 /*
6115  * Dequeue all writer callbacks from the outer perimeter and run them.
6116  */
6117 static void
6118 write_now(syncq_t *outer)
6119 {
6120 	mblk_t		*mp;
6121 	queue_t		*q;
6122 	void	(*func)();
6123 
6124 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6125 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6126 	    outer->sq_oprev != NULL);
6127 	while ((mp = outer->sq_evhead) != NULL) {
6128 		/*
6129 		 * queues cannot be placed on the queuelist on the outer
6130 		 * perimeter.
6131 		 */
6132 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6133 		ASSERT((outer->sq_flags & SQ_EVENTS));
6134 
6135 		outer->sq_evhead = mp->b_next;
6136 		if (outer->sq_evhead == NULL) {
6137 			outer->sq_evtail = NULL;
6138 			outer->sq_flags &= ~SQ_EVENTS;
6139 		}
6140 		ASSERT(outer->sq_count != 0);
6141 		outer->sq_count--;	/* Incremented when enqueued. */
6142 		mutex_exit(SQLOCK(outer));
6143 		/*
6144 		 * Drop the message if the queue is closing.
6145 		 * Make sure that the queue is "claimed" when the callback
6146 		 * is run in order to satisfy various ASSERTs.
6147 		 */
6148 		q = mp->b_queue;
6149 		func = (void (*)())mp->b_prev;
6150 		ASSERT(func != NULL);
6151 		mp->b_next = mp->b_prev = NULL;
6152 		if (q->q_flag & QWCLOSE) {
6153 			freemsg(mp);
6154 		} else {
6155 			claimq(q);
6156 			(*func)(q, mp);
6157 			releaseq(q);
6158 		}
6159 		mutex_enter(SQLOCK(outer));
6160 	}
6161 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6162 }
6163 
6164 /*
6165  * The list of messages on the inner syncq is effectively hashed
6166  * by destination queue.  These destination queues are doubly
6167  * linked lists (hopefully) in priority order.  Messages are then
6168  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6169  * Additional messages are linked together by the b_next/b_prev
6170  * elements in the mblk, with (similar to putq()) the first message
6171  * having a NULL b_prev and the last message having a NULL b_next.
6172  *
6173  * Events, such as qwriter callbacks, are put onto a list in FIFO
6174  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6175  * linked list, and messages here MUST be processed in the order queued.
6176  */
6177 
6178 /*
6179  * Run the events on the syncq event list (sq_evhead).
6180  * Assumes there is only one claim on the syncq, it is
6181  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6182  * Messages here are processed in order, with the SQ_EXCL bit
6183  * held all the way through till the last message is processed.
6184  */
6185 void
6186 sq_run_events(syncq_t *sq)
6187 {
6188 	mblk_t		*bp;
6189 	queue_t		*qp;
6190 	uint16_t	flags = sq->sq_flags;
6191 	void		(*func)();
6192 
6193 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6194 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6195 	    sq->sq_oprev == NULL) ||
6196 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6197 	    sq->sq_oprev != NULL));
6198 
6199 	ASSERT(flags & SQ_EXCL);
6200 	ASSERT(sq->sq_count == 1);
6201 
6202 	/*
6203 	 * We need to process all of the events on this list.  It
6204 	 * is possible that new events will be added while we are
6205 	 * away processing a callback, so on every loop, we start
6206 	 * back at the beginning of the list.
6207 	 */
6208 	/*
6209 	 * We have to reaccess sq_evhead since there is a
6210 	 * possibility of a new entry while we were running
6211 	 * the callback.
6212 	 */
6213 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6214 		ASSERT(bp->b_queue->q_syncq == sq);
6215 		ASSERT(sq->sq_flags & SQ_EVENTS);
6216 
6217 		qp = bp->b_queue;
6218 		func = (void (*)())bp->b_prev;
6219 		ASSERT(func != NULL);
6220 
6221 		/*
6222 		 * Messages from the event queue must be taken off in
6223 		 * FIFO order.
6224 		 */
6225 		ASSERT(sq->sq_evhead == bp);
6226 		sq->sq_evhead = bp->b_next;
6227 
6228 		if (bp->b_next == NULL) {
6229 			/* Deleting last */
6230 			ASSERT(sq->sq_evtail == bp);
6231 			sq->sq_evtail = NULL;
6232 			sq->sq_flags &= ~SQ_EVENTS;
6233 		}
6234 		bp->b_prev = bp->b_next = NULL;
6235 		ASSERT(bp->b_datap->db_ref != 0);
6236 
6237 		mutex_exit(SQLOCK(sq));
6238 
6239 		(*func)(qp, bp);
6240 
6241 		mutex_enter(SQLOCK(sq));
6242 		/*
6243 		 * re-read the flags, since they could have changed.
6244 		 */
6245 		flags = sq->sq_flags;
6246 		ASSERT(flags & SQ_EXCL);
6247 	}
6248 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6249 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6250 
6251 	if (flags & SQ_WANTWAKEUP) {
6252 		flags &= ~SQ_WANTWAKEUP;
6253 		cv_broadcast(&sq->sq_wait);
6254 	}
6255 	if (flags & SQ_WANTEXWAKEUP) {
6256 		flags &= ~SQ_WANTEXWAKEUP;
6257 		cv_broadcast(&sq->sq_exitwait);
6258 	}
6259 	sq->sq_flags = flags;
6260 }
6261 
6262 /*
6263  * Put messages on the event list.
6264  * If we can go exclusive now, do so and process the event list, otherwise
6265  * let the last claim service this list (or wake the sqthread).
6266  * This procedure assumes SQLOCK is held.  To run the event list, it
6267  * must be called with no claims.
6268  */
6269 static void
6270 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6271 {
6272 	uint16_t count;
6273 
6274 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6275 	ASSERT(func != NULL);
6276 
6277 	/*
6278 	 * This is a callback.  Add it to the list of callbacks
6279 	 * and see about upgrading.
6280 	 */
6281 	mp->b_prev = (mblk_t *)func;
6282 	mp->b_queue = q;
6283 	mp->b_next = NULL;
6284 	if (sq->sq_evhead == NULL) {
6285 		sq->sq_evhead = sq->sq_evtail = mp;
6286 		sq->sq_flags |= SQ_EVENTS;
6287 	} else {
6288 		ASSERT(sq->sq_evtail != NULL);
6289 		ASSERT(sq->sq_evtail->b_next == NULL);
6290 		ASSERT(sq->sq_flags & SQ_EVENTS);
6291 		sq->sq_evtail->b_next = mp;
6292 		sq->sq_evtail = mp;
6293 	}
6294 	/*
6295 	 * We have set SQ_EVENTS, so threads will have to
6296 	 * unwind out of the perimeter, and new entries will
6297 	 * not grab a putlock.  But we still need to know
6298 	 * how many threads have already made a claim to the
6299 	 * syncq, so grab the putlocks, and sum the counts.
6300 	 * If there are no claims on the syncq, we can upgrade
6301 	 * to exclusive, and run the event list.
6302 	 * NOTE: We hold the SQLOCK, so we can just grab the
6303 	 * putlocks.
6304 	 */
6305 	count = sq->sq_count;
6306 	SQ_PUTLOCKS_ENTER(sq);
6307 	SUM_SQ_PUTCOUNTS(sq, count);
6308 	/*
6309 	 * We have no claim, so we need to check if there
6310 	 * are no others, then we can upgrade.
6311 	 */
6312 	/*
6313 	 * There are currently no claims on
6314 	 * the syncq by this thread (at least on this entry). The thread who has
6315 	 * the claim should drain syncq.
6316 	 */
6317 	if (count > 0) {
6318 		/*
6319 		 * Can't upgrade - other threads inside.
6320 		 */
6321 		SQ_PUTLOCKS_EXIT(sq);
6322 		mutex_exit(SQLOCK(sq));
6323 		return;
6324 	}
6325 	/*
6326 	 * Need to set SQ_EXCL and make a claim on the syncq.
6327 	 */
6328 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6329 	sq->sq_flags |= SQ_EXCL;
6330 	ASSERT(sq->sq_count == 0);
6331 	sq->sq_count++;
6332 	SQ_PUTLOCKS_EXIT(sq);
6333 
6334 	/* Process the events list */
6335 	sq_run_events(sq);
6336 
6337 	/*
6338 	 * Release our claim...
6339 	 */
6340 	sq->sq_count--;
6341 
6342 	/*
6343 	 * And release SQ_EXCL.
6344 	 * We don't need to acquire the putlocks to release
6345 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6346 	 */
6347 	sq->sq_flags &= ~SQ_EXCL;
6348 
6349 	/*
6350 	 * sq_run_events should have released SQ_EXCL
6351 	 */
6352 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6353 
6354 	/*
6355 	 * If anything happened while we were running the
6356 	 * events (or was there before), we need to process
6357 	 * them now.  We shouldn't be exclusive sine we
6358 	 * released the perimeter above (plus, we asserted
6359 	 * for it).
6360 	 */
6361 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6362 		drain_syncq(sq);
6363 	else
6364 		mutex_exit(SQLOCK(sq));
6365 }
6366 
6367 /*
6368  * Perform delayed processing. The caller has to make sure that it is safe
6369  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6370  * set).
6371  *
6372  * Assume that the caller has NO claims on the syncq.  However, a claim
6373  * on the syncq does not indicate that a thread is draining the syncq.
6374  * There may be more claims on the syncq than there are threads draining
6375  * (i.e.  #_threads_draining <= sq_count)
6376  *
6377  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6378  * in order to preserve qwriter(OUTER) ordering constraints.
6379  *
6380  * sq_putcount only needs to be checked when dispatching the queued
6381  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6382  */
6383 void
6384 drain_syncq(syncq_t *sq)
6385 {
6386 	queue_t		*qp;
6387 	uint16_t	count;
6388 	uint16_t	type = sq->sq_type;
6389 	uint16_t	flags = sq->sq_flags;
6390 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6391 
6392 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6393 	    "drain_syncq start:%p", sq);
6394 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6395 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6396 	    sq->sq_oprev == NULL) ||
6397 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6398 	    sq->sq_oprev != NULL));
6399 
6400 	/*
6401 	 * Drop SQ_SERVICE flag.
6402 	 */
6403 	if (bg_service)
6404 		sq->sq_svcflags &= ~SQ_SERVICE;
6405 
6406 	/*
6407 	 * If SQ_EXCL is set, someone else is processing this syncq - let them
6408 	 * finish the job.
6409 	 */
6410 	if (flags & SQ_EXCL) {
6411 		if (bg_service) {
6412 			ASSERT(sq->sq_servcount != 0);
6413 			sq->sq_servcount--;
6414 		}
6415 		mutex_exit(SQLOCK(sq));
6416 		return;
6417 	}
6418 
6419 	/*
6420 	 * This routine can be called by a background thread if
6421 	 * it was scheduled by a hi-priority thread.  SO, if there are
6422 	 * NOT messages queued, return (remember, we have the SQLOCK,
6423 	 * and it cannot change until we release it). Wakeup any waiters also.
6424 	 */
6425 	if (!(flags & SQ_QUEUED)) {
6426 		if (flags & SQ_WANTWAKEUP) {
6427 			flags &= ~SQ_WANTWAKEUP;
6428 			cv_broadcast(&sq->sq_wait);
6429 		}
6430 		if (flags & SQ_WANTEXWAKEUP) {
6431 			flags &= ~SQ_WANTEXWAKEUP;
6432 			cv_broadcast(&sq->sq_exitwait);
6433 		}
6434 		sq->sq_flags = flags;
6435 		if (bg_service) {
6436 			ASSERT(sq->sq_servcount != 0);
6437 			sq->sq_servcount--;
6438 		}
6439 		mutex_exit(SQLOCK(sq));
6440 		return;
6441 	}
6442 
6443 	/*
6444 	 * If this is not a concurrent put perimeter, we need to
6445 	 * become exclusive to drain.  Also, if not CIPUT, we would
6446 	 * not have acquired a putlock, so we don't need to check
6447 	 * the putcounts.  If not entering with a claim, we test
6448 	 * for sq_count == 0.
6449 	 */
6450 	type = sq->sq_type;
6451 	if (!(type & SQ_CIPUT)) {
6452 		if (sq->sq_count > 1) {
6453 			if (bg_service) {
6454 				ASSERT(sq->sq_servcount != 0);
6455 				sq->sq_servcount--;
6456 			}
6457 			mutex_exit(SQLOCK(sq));
6458 			return;
6459 		}
6460 		sq->sq_flags |= SQ_EXCL;
6461 	}
6462 
6463 	/*
6464 	 * This is where we make a claim to the syncq.
6465 	 * This can either be done by incrementing a putlock, or
6466 	 * the sq_count.  But since we already have the SQLOCK
6467 	 * here, we just bump the sq_count.
6468 	 *
6469 	 * Note that after we make a claim, we need to let the code
6470 	 * fall through to the end of this routine to clean itself
6471 	 * up.  A return in the while loop will put the syncq in a
6472 	 * very bad state.
6473 	 */
6474 	sq->sq_count++;
6475 	ASSERT(sq->sq_count != 0);	/* wraparound */
6476 
6477 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6478 		/*
6479 		 * If we are told to stayaway or went exclusive,
6480 		 * we are done.
6481 		 */
6482 		if (flags & (SQ_STAYAWAY)) {
6483 			break;
6484 		}
6485 
6486 		/*
6487 		 * If there are events to run, do so.
6488 		 * We have one claim to the syncq, so if there are
6489 		 * more than one, other threads are running.
6490 		 */
6491 		if (sq->sq_evhead != NULL) {
6492 			ASSERT(sq->sq_flags & SQ_EVENTS);
6493 
6494 			count = sq->sq_count;
6495 			SQ_PUTLOCKS_ENTER(sq);
6496 			SUM_SQ_PUTCOUNTS(sq, count);
6497 			if (count > 1) {
6498 				SQ_PUTLOCKS_EXIT(sq);
6499 				/* Can't upgrade - other threads inside */
6500 				break;
6501 			}
6502 			ASSERT((flags & SQ_EXCL) == 0);
6503 			sq->sq_flags = flags | SQ_EXCL;
6504 			SQ_PUTLOCKS_EXIT(sq);
6505 			/*
6506 			 * we have the only claim, run the events,
6507 			 * sq_run_events will clear the SQ_EXCL flag.
6508 			 */
6509 			sq_run_events(sq);
6510 
6511 			/*
6512 			 * If this is a CIPUT perimeter, we need
6513 			 * to drop the SQ_EXCL flag so we can properly
6514 			 * continue draining the syncq.
6515 			 */
6516 			if (type & SQ_CIPUT) {
6517 				ASSERT(sq->sq_flags & SQ_EXCL);
6518 				sq->sq_flags &= ~SQ_EXCL;
6519 			}
6520 
6521 			/*
6522 			 * And go back to the beginning just in case
6523 			 * anything changed while we were away.
6524 			 */
6525 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6526 			continue;
6527 		}
6528 
6529 		ASSERT(sq->sq_evhead == NULL);
6530 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6531 
6532 		/*
6533 		 * Find the queue that is not draining.
6534 		 *
6535 		 * q_draining is protected by QLOCK which we do not hold.
6536 		 * But if it was set, then a thread was draining, and if it gets
6537 		 * cleared, then it was because the thread has successfully
6538 		 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6539 		 * state to happen, a thread needs the SQLOCK which we hold, and
6540 		 * if there was such a flag, we would have already seen it.
6541 		 */
6542 
6543 		for (qp = sq->sq_head;
6544 		    qp != NULL && (qp->q_draining ||
6545 		    (qp->q_sqflags & Q_SQDRAINING));
6546 		    qp = qp->q_sqnext)
6547 			;
6548 
6549 		if (qp == NULL)
6550 			break;
6551 
6552 		/*
6553 		 * We have a queue to work on, and we hold the
6554 		 * SQLOCK and one claim, call qdrain_syncq.
6555 		 * This means we need to release the SQLOCK and
6556 		 * acquire the QLOCK (OK since we have a claim).
6557 		 * Note that qdrain_syncq will actually dequeue
6558 		 * this queue from the sq_head list when it is
6559 		 * convinced all the work is done and release
6560 		 * the QLOCK before returning.
6561 		 */
6562 		qp->q_sqflags |= Q_SQDRAINING;
6563 		mutex_exit(SQLOCK(sq));
6564 		mutex_enter(QLOCK(qp));
6565 		qdrain_syncq(sq, qp);
6566 		mutex_enter(SQLOCK(sq));
6567 
6568 		/* The queue is drained */
6569 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6570 		qp->q_sqflags &= ~Q_SQDRAINING;
6571 		/*
6572 		 * NOTE: After this point qp should not be used since it may be
6573 		 * closed.
6574 		 */
6575 	}
6576 
6577 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6578 	flags = sq->sq_flags;
6579 
6580 	/*
6581 	 * sq->sq_head cannot change because we hold the
6582 	 * sqlock. However, a thread CAN decide that it is no longer
6583 	 * going to drain that queue.  However, this should be due to
6584 	 * a GOAWAY state, and we should see that here.
6585 	 *
6586 	 * This loop is not very efficient. One solution may be adding a second
6587 	 * pointer to the "draining" queue, but it is difficult to do when
6588 	 * queues are inserted in the middle due to priority ordering. Another
6589 	 * possibility is to yank the queue out of the sq list and put it onto
6590 	 * the "draining list" and then put it back if it can't be drained.
6591 	 */
6592 
6593 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6594 	    (type & SQ_CI) || sq->sq_head->q_draining);
6595 
6596 	/* Drop SQ_EXCL for non-CIPUT perimeters */
6597 	if (!(type & SQ_CIPUT))
6598 		flags &= ~SQ_EXCL;
6599 	ASSERT((flags & SQ_EXCL) == 0);
6600 
6601 	/* Wake up any waiters. */
6602 	if (flags & SQ_WANTWAKEUP) {
6603 		flags &= ~SQ_WANTWAKEUP;
6604 		cv_broadcast(&sq->sq_wait);
6605 	}
6606 	if (flags & SQ_WANTEXWAKEUP) {
6607 		flags &= ~SQ_WANTEXWAKEUP;
6608 		cv_broadcast(&sq->sq_exitwait);
6609 	}
6610 	sq->sq_flags = flags;
6611 
6612 	ASSERT(sq->sq_count != 0);
6613 	/* Release our claim. */
6614 	sq->sq_count--;
6615 
6616 	if (bg_service) {
6617 		ASSERT(sq->sq_servcount != 0);
6618 		sq->sq_servcount--;
6619 	}
6620 
6621 	mutex_exit(SQLOCK(sq));
6622 
6623 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6624 	    "drain_syncq end:%p", sq);
6625 }
6626 
6627 
6628 /*
6629  *
6630  * qdrain_syncq can be called (currently) from only one of two places:
6631  *	drain_syncq
6632  * 	putnext  (or some variation of it).
6633  * and eventually
6634  * 	qwait(_sig)
6635  *
6636  * If called from drain_syncq, we found it in the list of queues needing
6637  * service, so there is work to be done (or it wouldn't be in the list).
6638  *
6639  * If called from some putnext variation, it was because the
6640  * perimeter is open, but messages are blocking a putnext and
6641  * there is not a thread working on it.  Now a thread could start
6642  * working on it while we are getting ready to do so ourself, but
6643  * the thread would set the q_draining flag, and we can spin out.
6644  *
6645  * As for qwait(_sig), I think I shall let it continue to call
6646  * drain_syncq directly (after all, it will get here eventually).
6647  *
6648  * qdrain_syncq has to terminate when:
6649  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6650  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6651  *
6652  * ASSUMES:
6653  *	One claim
6654  * 	QLOCK held
6655  * 	SQLOCK not held
6656  *	Will release QLOCK before returning
6657  */
6658 void
6659 qdrain_syncq(syncq_t *sq, queue_t *q)
6660 {
6661 	mblk_t		*bp;
6662 #ifdef DEBUG
6663 	uint16_t	count;
6664 #endif
6665 
6666 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6667 	    "drain_syncq start:%p", sq);
6668 	ASSERT(q->q_syncq == sq);
6669 	ASSERT(MUTEX_HELD(QLOCK(q)));
6670 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6671 	/*
6672 	 * For non-CIPUT perimeters, we should be called with the exclusive bit
6673 	 * set already. For CIPUT perimeters, we will be doing a concurrent
6674 	 * drain, so it better not be set.
6675 	 */
6676 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6677 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6678 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6679 	/*
6680 	 * All outer pointers are set, or none of them are
6681 	 */
6682 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6683 	    sq->sq_oprev == NULL) ||
6684 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6685 	    sq->sq_oprev != NULL));
6686 #ifdef DEBUG
6687 	count = sq->sq_count;
6688 	/*
6689 	 * This is OK without the putlocks, because we have one
6690 	 * claim either from the sq_count, or a putcount.  We could
6691 	 * get an erroneous value from other counts, but ours won't
6692 	 * change, so one way or another, we will have at least a
6693 	 * value of one.
6694 	 */
6695 	SUM_SQ_PUTCOUNTS(sq, count);
6696 	ASSERT(count >= 1);
6697 #endif /* DEBUG */
6698 
6699 	/*
6700 	 * The first thing to do is find out if a thread is already draining
6701 	 * this queue. If so, we are done, just return.
6702 	 */
6703 	if (q->q_draining) {
6704 		mutex_exit(QLOCK(q));
6705 		return;
6706 	}
6707 
6708 	/*
6709 	 * If the perimeter is exclusive, there is nothing we can do right now,
6710 	 * go away. Note that there is nothing to prevent this case from
6711 	 * changing right after this check, but the spin-out will catch it.
6712 	 */
6713 
6714 	/* Tell other threads that we are draining this queue */
6715 	q->q_draining = 1;	/* Protected by QLOCK */
6716 
6717 	/*
6718 	 * If there is nothing to do, clear QFULL as necessary. This caters for
6719 	 * the case where an empty queue was enqueued onto the syncq.
6720 	 */
6721 	if (q->q_sqhead == NULL) {
6722 		ASSERT(q->q_syncqmsgs == 0);
6723 		mutex_exit(QLOCK(q));
6724 		clr_qfull(q);
6725 		mutex_enter(QLOCK(q));
6726 	}
6727 
6728 	/*
6729 	 * Note that q_sqhead must be re-checked here in case another message
6730 	 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6731 	 */
6732 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6733 		/*
6734 		 * Because we can enter this routine just because a putnext is
6735 		 * blocked, we need to spin out if the perimeter wants to go
6736 		 * exclusive as well as just blocked. We need to spin out also
6737 		 * if events are queued on the syncq.
6738 		 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6739 		 * set it, and it can't become exclusive while we hold a claim.
6740 		 */
6741 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6742 			break;
6743 		}
6744 
6745 #ifdef DEBUG
6746 		/*
6747 		 * Since we are in qdrain_syncq, we already know the queue,
6748 		 * but for sanity, we want to check this against the qp that
6749 		 * was passed in by bp->b_queue.
6750 		 */
6751 
6752 		ASSERT(bp->b_queue == q);
6753 		ASSERT(bp->b_queue->q_syncq == sq);
6754 		bp->b_queue = NULL;
6755 
6756 		/*
6757 		 * We would have the following check in the DEBUG code:
6758 		 *
6759 		 * if (bp->b_prev != NULL)  {
6760 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6761 		 * }
6762 		 *
6763 		 * This can't be done, however, since IP modifies qinfo
6764 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6765 		 * qinfo), invalidating the check.
6766 		 * So the assignment to func is left here, but the ASSERT itself
6767 		 * is removed until the whole issue is resolved.
6768 		 */
6769 #endif
6770 		ASSERT(q->q_sqhead == bp);
6771 		q->q_sqhead = bp->b_next;
6772 		bp->b_prev = bp->b_next = NULL;
6773 		ASSERT(q->q_syncqmsgs > 0);
6774 		mutex_exit(QLOCK(q));
6775 
6776 		ASSERT(bp->b_datap->db_ref != 0);
6777 
6778 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6779 
6780 		mutex_enter(QLOCK(q));
6781 
6782 		/*
6783 		 * q_syncqmsgs should only be decremented after executing the
6784 		 * put procedure to avoid message re-ordering. This is due to an
6785 		 * optimisation in putnext() which can call the put procedure
6786 		 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6787 		 * being set).
6788 		 *
6789 		 * We also need to clear QFULL in the next service procedure
6790 		 * queue if this is the last message destined for that queue.
6791 		 *
6792 		 * It would make better sense to have some sort of tunable for
6793 		 * the low water mark, but these semantics are not yet defined.
6794 		 * So, alas, we use a constant.
6795 		 */
6796 		if (--q->q_syncqmsgs == 0) {
6797 			mutex_exit(QLOCK(q));
6798 			clr_qfull(q);
6799 			mutex_enter(QLOCK(q));
6800 		}
6801 
6802 		/*
6803 		 * Always clear SQ_EXCL when CIPUT in order to handle
6804 		 * qwriter(INNER). The putp() can call qwriter and get exclusive
6805 		 * access IFF this is the only claim. So, we need to test for
6806 		 * this possibility, acquire the mutex and clear the bit.
6807 		 */
6808 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6809 			mutex_enter(SQLOCK(sq));
6810 			sq->sq_flags &= ~SQ_EXCL;
6811 			mutex_exit(SQLOCK(sq));
6812 		}
6813 	}
6814 
6815 	/*
6816 	 * We should either have no messages on this queue, or we were told to
6817 	 * goaway by a waiter (which we will wake up at the end of this
6818 	 * function).
6819 	 */
6820 	ASSERT((q->q_sqhead == NULL) ||
6821 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6822 
6823 	ASSERT(MUTEX_HELD(QLOCK(q)));
6824 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6825 
6826 	/* Remove the q from the syncq list if all the messages are drained. */
6827 	if (q->q_sqhead == NULL) {
6828 		ASSERT(q->q_syncqmsgs == 0);
6829 		mutex_enter(SQLOCK(sq));
6830 		if (q->q_sqflags & Q_SQQUEUED)
6831 			SQRM_Q(sq, q);
6832 		mutex_exit(SQLOCK(sq));
6833 		/*
6834 		 * Since the queue is removed from the list, reset its priority.
6835 		 */
6836 		q->q_spri = 0;
6837 	}
6838 
6839 	/*
6840 	 * Remember, the q_draining flag is used to let another thread know
6841 	 * that there is a thread currently draining the messages for a queue.
6842 	 * Since we are now done with this queue (even if there may be messages
6843 	 * still there), we need to clear this flag so some thread will work on
6844 	 * it if needed.
6845 	 */
6846 	ASSERT(q->q_draining);
6847 	q->q_draining = 0;
6848 
6849 	/* Called with a claim, so OK to drop all locks. */
6850 	mutex_exit(QLOCK(q));
6851 
6852 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6853 	    "drain_syncq end:%p", sq);
6854 }
6855 /* END OF QDRAIN_SYNCQ  */
6856 
6857 
6858 /*
6859  * This is the mate to qdrain_syncq, except that it is putting the message onto
6860  * the queue instead of draining. Since the message is destined for the queue
6861  * that is selected, there is no need to identify the function because the
6862  * message is intended for the put routine for the queue. For debug kernels,
6863  * this routine will do it anyway just in case.
6864  *
6865  * After the message is enqueued on the syncq, it calls putnext_tail()
6866  * which will schedule a background thread to actually process the message.
6867  *
6868  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6869  * SQLOCK(sq) and QLOCK(q) are not held.
6870  */
6871 void
6872 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6873 {
6874 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6875 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6876 	ASSERT(sq->sq_count > 0);
6877 	ASSERT(q->q_syncq == sq);
6878 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6879 	    sq->sq_oprev == NULL) ||
6880 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6881 	    sq->sq_oprev != NULL));
6882 
6883 	mutex_enter(QLOCK(q));
6884 
6885 #ifdef DEBUG
6886 	/*
6887 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6888 	 * to trace the queue that the message is intended for.  Note
6889 	 * that the original use was to identify the queue and function
6890 	 * to call on the drain.  In the new syncq, we have the context
6891 	 * of the queue that we are draining, so call it's putproc and
6892 	 * don't rely on the saved values.  But for debug this is still
6893 	 * useful information.
6894 	 */
6895 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6896 	mp->b_queue = q;
6897 	mp->b_next = NULL;
6898 #endif
6899 	ASSERT(q->q_syncq == sq);
6900 	/*
6901 	 * Enqueue the message on the list.
6902 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6903 	 * protect it.  So it's ok to acquire SQLOCK after SQPUT_MP().
6904 	 */
6905 	SQPUT_MP(q, mp);
6906 	mutex_enter(SQLOCK(sq));
6907 
6908 	/*
6909 	 * And queue on syncq for scheduling, if not already queued.
6910 	 * Note that we need the SQLOCK for this, and for testing flags
6911 	 * at the end to see if we will drain.  So grab it now, and
6912 	 * release it before we call qdrain_syncq or return.
6913 	 */
6914 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6915 		q->q_spri = curthread->t_pri;
6916 		SQPUT_Q(sq, q);
6917 	}
6918 #ifdef DEBUG
6919 	else {
6920 		/*
6921 		 * All of these conditions MUST be true!
6922 		 */
6923 		ASSERT(sq->sq_tail != NULL);
6924 		if (sq->sq_tail == sq->sq_head) {
6925 			ASSERT((q->q_sqprev == NULL) &&
6926 			    (q->q_sqnext == NULL));
6927 		} else {
6928 			ASSERT((q->q_sqprev != NULL) ||
6929 			    (q->q_sqnext != NULL));
6930 		}
6931 		ASSERT(sq->sq_flags & SQ_QUEUED);
6932 		ASSERT(q->q_syncqmsgs != 0);
6933 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6934 	}
6935 #endif
6936 	mutex_exit(QLOCK(q));
6937 	/*
6938 	 * SQLOCK is still held, so sq_count can be safely decremented.
6939 	 */
6940 	sq->sq_count--;
6941 
6942 	putnext_tail(sq, q, 0);
6943 	/* Should not reference sq or q after this point. */
6944 }
6945 
6946 /*  End of qfill_syncq  */
6947 
6948 /*
6949  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6950  * that would be put into qp by drain_syncq.
6951  * Used when deleting the syncq (qp == NULL) or when detaching
6952  * a queue (qp != NULL).
6953  * Return non-zero if one or more messages were freed.
6954  *
6955  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
6956  * sq_putlocks are used.
6957  *
6958  * NOTE: This function assumes that it is called from the close() context and
6959  * that all the queues in the syncq are going away. For this reason it doesn't
6960  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6961  * currently valid, but it is useful to rethink this function to behave properly
6962  * in other cases.
6963  */
6964 int
6965 flush_syncq(syncq_t *sq, queue_t *qp)
6966 {
6967 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6968 	queue_t		*q;
6969 	int		ret = 0;
6970 
6971 	mutex_enter(SQLOCK(sq));
6972 
6973 	/*
6974 	 * Before we leave, we need to make sure there are no
6975 	 * events listed for this queue.  All events for this queue
6976 	 * will just be freed.
6977 	 */
6978 	if (qp != NULL && sq->sq_evhead != NULL) {
6979 		ASSERT(sq->sq_flags & SQ_EVENTS);
6980 
6981 		mp_prev = NULL;
6982 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6983 			mp_next = bp->b_next;
6984 			if (bp->b_queue == qp) {
6985 				/* Delete this message */
6986 				if (mp_prev != NULL) {
6987 					mp_prev->b_next = mp_next;
6988 					/*
6989 					 * Update sq_evtail if the last element
6990 					 * is removed.
6991 					 */
6992 					if (bp == sq->sq_evtail) {
6993 						ASSERT(mp_next == NULL);
6994 						sq->sq_evtail = mp_prev;
6995 					}
6996 				} else
6997 					sq->sq_evhead = mp_next;
6998 				if (sq->sq_evhead == NULL)
6999 					sq->sq_flags &= ~SQ_EVENTS;
7000 				bp->b_prev = bp->b_next = NULL;
7001 				freemsg(bp);
7002 				ret++;
7003 			} else {
7004 				mp_prev = bp;
7005 			}
7006 		}
7007 	}
7008 
7009 	/*
7010 	 * Walk sq_head and:
7011 	 *	- match qp if qp is set, remove it's messages
7012 	 *	- all if qp is not set
7013 	 */
7014 	q = sq->sq_head;
7015 	while (q != NULL) {
7016 		ASSERT(q->q_syncq == sq);
7017 		if ((qp == NULL) || (qp == q)) {
7018 			/*
7019 			 * Yank the messages as a list off the queue
7020 			 */
7021 			mp_head = q->q_sqhead;
7022 			/*
7023 			 * We do not have QLOCK(q) here (which is safe due to
7024 			 * assumptions mentioned above). To obtain the lock we
7025 			 * need to release SQLOCK which may allow lots of things
7026 			 * to change upon us. This place requires more analysis.
7027 			 */
7028 			q->q_sqhead = q->q_sqtail = NULL;
7029 			ASSERT(mp_head->b_queue &&
7030 			    mp_head->b_queue->q_syncq == sq);
7031 
7032 			/*
7033 			 * Free each of the messages.
7034 			 */
7035 			for (bp = mp_head; bp != NULL; bp = mp_next) {
7036 				mp_next = bp->b_next;
7037 				bp->b_prev = bp->b_next = NULL;
7038 				freemsg(bp);
7039 				ret++;
7040 			}
7041 			/*
7042 			 * Now remove the queue from the syncq.
7043 			 */
7044 			ASSERT(q->q_sqflags & Q_SQQUEUED);
7045 			SQRM_Q(sq, q);
7046 			q->q_spri = 0;
7047 			q->q_syncqmsgs = 0;
7048 
7049 			/*
7050 			 * If qp was specified, we are done with it and are
7051 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
7052 			 * waiters while we still have the SQLOCK.
7053 			 */
7054 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7055 				sq->sq_flags &= ~SQ_WANTWAKEUP;
7056 				cv_broadcast(&sq->sq_wait);
7057 			}
7058 			/* Drop SQLOCK across clr_qfull */
7059 			mutex_exit(SQLOCK(sq));
7060 
7061 			/*
7062 			 * We avoid doing the test that drain_syncq does and
7063 			 * unconditionally clear qfull for every flushed
7064 			 * message. Since flush_syncq is only called during
7065 			 * close this should not be a problem.
7066 			 */
7067 			clr_qfull(q);
7068 			if (qp != NULL) {
7069 				return (ret);
7070 			} else {
7071 				mutex_enter(SQLOCK(sq));
7072 				/*
7073 				 * The head was removed by SQRM_Q above.
7074 				 * reread the new head and flush it.
7075 				 */
7076 				q = sq->sq_head;
7077 			}
7078 		} else {
7079 			q = q->q_sqnext;
7080 		}
7081 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
7082 	}
7083 
7084 	if (sq->sq_flags & SQ_WANTWAKEUP) {
7085 		sq->sq_flags &= ~SQ_WANTWAKEUP;
7086 		cv_broadcast(&sq->sq_wait);
7087 	}
7088 
7089 	mutex_exit(SQLOCK(sq));
7090 	return (ret);
7091 }
7092 
7093 /*
7094  * Propagate all messages from a syncq to the next syncq that are associated
7095  * with the specified queue. If the queue is attached to a driver or if the
7096  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7097  *
7098  * Assumes that the stream is strlock()'ed. We don't come here if there
7099  * are no messages to propagate.
7100  *
7101  * NOTE : If the queue is attached to a driver, all the messages are freed
7102  * as there is no point in propagating the messages from the driver syncq
7103  * to the closing stream head which will in turn get freed later.
7104  */
7105 static int
7106 propagate_syncq(queue_t *qp)
7107 {
7108 	mblk_t		*bp, *head, *tail, *prev, *next;
7109 	syncq_t 	*sq;
7110 	queue_t		*nqp;
7111 	syncq_t		*nsq;
7112 	boolean_t	isdriver;
7113 	int 		moved = 0;
7114 	uint16_t	flags;
7115 	pri_t		priority = curthread->t_pri;
7116 #ifdef DEBUG
7117 	void		(*func)();
7118 #endif
7119 
7120 	sq = qp->q_syncq;
7121 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7122 	/* debug macro */
7123 	SQ_PUTLOCKS_HELD(sq);
7124 	/*
7125 	 * As entersq() does not increment the sq_count for
7126 	 * the write side, check sq_count for non-QPERQ
7127 	 * perimeters alone.
7128 	 */
7129 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7130 
7131 	/*
7132 	 * propagate_syncq() can be called because of either messages on the
7133 	 * queue syncq or because on events on the queue syncq. Do actual
7134 	 * message propagations if there are any messages.
7135 	 */
7136 	if (qp->q_syncqmsgs) {
7137 		isdriver = (qp->q_flag & QISDRV);
7138 
7139 		if (!isdriver) {
7140 			nqp = qp->q_next;
7141 			nsq = nqp->q_syncq;
7142 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7143 			/* debug macro */
7144 			SQ_PUTLOCKS_HELD(nsq);
7145 #ifdef DEBUG
7146 			func = (void (*)())nqp->q_qinfo->qi_putp;
7147 #endif
7148 		}
7149 
7150 		SQRM_Q(sq, qp);
7151 		priority = MAX(qp->q_spri, priority);
7152 		qp->q_spri = 0;
7153 		head = qp->q_sqhead;
7154 		tail = qp->q_sqtail;
7155 		qp->q_sqhead = qp->q_sqtail = NULL;
7156 		qp->q_syncqmsgs = 0;
7157 
7158 		/*
7159 		 * Walk the list of messages, and free them if this is a driver,
7160 		 * otherwise reset the b_prev and b_queue value to the new putp.
7161 		 * Afterward, we will just add the head to the end of the next
7162 		 * syncq, and point the tail to the end of this one.
7163 		 */
7164 
7165 		for (bp = head; bp != NULL; bp = next) {
7166 			next = bp->b_next;
7167 			if (isdriver) {
7168 				bp->b_prev = bp->b_next = NULL;
7169 				freemsg(bp);
7170 				continue;
7171 			}
7172 			/* Change the q values for this message */
7173 			bp->b_queue = nqp;
7174 #ifdef DEBUG
7175 			bp->b_prev = (mblk_t *)func;
7176 #endif
7177 			moved++;
7178 		}
7179 		/*
7180 		 * Attach list of messages to the end of the new queue (if there
7181 		 * is a list of messages).
7182 		 */
7183 
7184 		if (!isdriver && head != NULL) {
7185 			ASSERT(tail != NULL);
7186 			if (nqp->q_sqhead == NULL) {
7187 				nqp->q_sqhead = head;
7188 			} else {
7189 				ASSERT(nqp->q_sqtail != NULL);
7190 				nqp->q_sqtail->b_next = head;
7191 			}
7192 			nqp->q_sqtail = tail;
7193 			/*
7194 			 * When messages are moved from high priority queue to
7195 			 * another queue, the destination queue priority is
7196 			 * upgraded.
7197 			 */
7198 
7199 			if (priority > nqp->q_spri)
7200 				nqp->q_spri = priority;
7201 
7202 			SQPUT_Q(nsq, nqp);
7203 
7204 			nqp->q_syncqmsgs += moved;
7205 			ASSERT(nqp->q_syncqmsgs != 0);
7206 		}
7207 	}
7208 
7209 	/*
7210 	 * Before we leave, we need to make sure there are no
7211 	 * events listed for this queue.  All events for this queue
7212 	 * will just be freed.
7213 	 */
7214 	if (sq->sq_evhead != NULL) {
7215 		ASSERT(sq->sq_flags & SQ_EVENTS);
7216 		prev = NULL;
7217 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7218 			next = bp->b_next;
7219 			if (bp->b_queue == qp) {
7220 				/* Delete this message */
7221 				if (prev != NULL) {
7222 					prev->b_next = next;
7223 					/*
7224 					 * Update sq_evtail if the last element
7225 					 * is removed.
7226 					 */
7227 					if (bp == sq->sq_evtail) {
7228 						ASSERT(next == NULL);
7229 						sq->sq_evtail = prev;
7230 					}
7231 				} else
7232 					sq->sq_evhead = next;
7233 				if (sq->sq_evhead == NULL)
7234 					sq->sq_flags &= ~SQ_EVENTS;
7235 				bp->b_prev = bp->b_next = NULL;
7236 				freemsg(bp);
7237 			} else {
7238 				prev = bp;
7239 			}
7240 		}
7241 	}
7242 
7243 	flags = sq->sq_flags;
7244 
7245 	/* Wake up any waiter before leaving. */
7246 	if (flags & SQ_WANTWAKEUP) {
7247 		flags &= ~SQ_WANTWAKEUP;
7248 		cv_broadcast(&sq->sq_wait);
7249 	}
7250 	sq->sq_flags = flags;
7251 
7252 	return (moved);
7253 }
7254 
7255 /*
7256  * Try and upgrade to exclusive access at the inner perimeter. If this can
7257  * not be done without blocking then request will be queued on the syncq
7258  * and drain_syncq will run it later.
7259  *
7260  * This routine can only be called from put or service procedures plus
7261  * asynchronous callback routines that have properly entered the queue (with
7262  * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7263  * associated with q.
7264  */
7265 void
7266 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7267 {
7268 	syncq_t	*sq = q->q_syncq;
7269 	uint16_t count;
7270 
7271 	mutex_enter(SQLOCK(sq));
7272 	count = sq->sq_count;
7273 	SQ_PUTLOCKS_ENTER(sq);
7274 	SUM_SQ_PUTCOUNTS(sq, count);
7275 	ASSERT(count >= 1);
7276 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7277 
7278 	if (count == 1) {
7279 		/*
7280 		 * Can upgrade. This case also handles nested qwriter calls
7281 		 * (when the qwriter callback function calls qwriter). In that
7282 		 * case SQ_EXCL is already set.
7283 		 */
7284 		sq->sq_flags |= SQ_EXCL;
7285 		SQ_PUTLOCKS_EXIT(sq);
7286 		mutex_exit(SQLOCK(sq));
7287 		(*func)(q, mp);
7288 		/*
7289 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7290 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7291 		 * until putnext, leavesq, or drain_syncq drops it.
7292 		 * That way we handle nested qwriter(INNER) without dropping
7293 		 * SQ_EXCL until the outermost qwriter callback routine is
7294 		 * done.
7295 		 */
7296 		return;
7297 	}
7298 	SQ_PUTLOCKS_EXIT(sq);
7299 	sqfill_events(sq, q, mp, func);
7300 }
7301 
7302 /*
7303  * Synchronous callback support functions
7304  */
7305 
7306 /*
7307  * Allocate a callback parameter structure.
7308  * Assumes that caller initializes the flags and the id.
7309  * Acquires SQLOCK(sq) if non-NULL is returned.
7310  */
7311 callbparams_t *
7312 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7313 {
7314 	callbparams_t *cbp;
7315 	size_t size = sizeof (callbparams_t);
7316 
7317 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7318 
7319 	/*
7320 	 * Only try tryhard allocation if the caller is ready to panic.
7321 	 * Otherwise just fail.
7322 	 */
7323 	if (cbp == NULL) {
7324 		if (kmflags & KM_PANIC)
7325 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7326 			    &size, kmflags);
7327 		else
7328 			return (NULL);
7329 	}
7330 
7331 	ASSERT(size >= sizeof (callbparams_t));
7332 	cbp->cbp_size = size;
7333 	cbp->cbp_sq = sq;
7334 	cbp->cbp_func = func;
7335 	cbp->cbp_arg = arg;
7336 	mutex_enter(SQLOCK(sq));
7337 	cbp->cbp_next = sq->sq_callbpend;
7338 	sq->sq_callbpend = cbp;
7339 	return (cbp);
7340 }
7341 
7342 void
7343 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7344 {
7345 	callbparams_t **pp, *p;
7346 
7347 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7348 
7349 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7350 		if (p == cbp) {
7351 			*pp = p->cbp_next;
7352 			kmem_free(p, p->cbp_size);
7353 			return;
7354 		}
7355 	}
7356 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7357 	    "callbparams_free: not found\n"));
7358 }
7359 
7360 void
7361 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7362 {
7363 	callbparams_t **pp, *p;
7364 
7365 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7366 
7367 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7368 		if (p->cbp_id == id && p->cbp_flags == flag) {
7369 			*pp = p->cbp_next;
7370 			kmem_free(p, p->cbp_size);
7371 			return;
7372 		}
7373 	}
7374 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7375 	    "callbparams_free_id: not found\n"));
7376 }
7377 
7378 /*
7379  * Callback wrapper function used by once-only callbacks that can be
7380  * cancelled (qtimeout and qbufcall)
7381  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7382  * cancelled by the qun* functions.
7383  */
7384 void
7385 qcallbwrapper(void *arg)
7386 {
7387 	callbparams_t *cbp = arg;
7388 	syncq_t	*sq;
7389 	uint16_t count = 0;
7390 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7391 	uint16_t type;
7392 
7393 	sq = cbp->cbp_sq;
7394 	mutex_enter(SQLOCK(sq));
7395 	type = sq->sq_type;
7396 	if (!(type & SQ_CICB)) {
7397 		count = sq->sq_count;
7398 		SQ_PUTLOCKS_ENTER(sq);
7399 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7400 		SUM_SQ_PUTCOUNTS(sq, count);
7401 		sq->sq_needexcl++;
7402 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7403 		waitflags |= SQ_MESSAGES;
7404 	}
7405 	/* Can not handle exclusive entry at outer perimeter */
7406 	ASSERT(type & SQ_COCB);
7407 
7408 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7409 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7410 		    (sq->sq_cancelid == cbp->cbp_id)) {
7411 			/* timeout has been cancelled */
7412 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7413 			callbparams_free(sq, cbp);
7414 			if (!(type & SQ_CICB)) {
7415 				ASSERT(sq->sq_needexcl > 0);
7416 				sq->sq_needexcl--;
7417 				if (sq->sq_needexcl == 0) {
7418 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7419 				}
7420 				SQ_PUTLOCKS_EXIT(sq);
7421 			}
7422 			mutex_exit(SQLOCK(sq));
7423 			return;
7424 		}
7425 		sq->sq_flags |= SQ_WANTWAKEUP;
7426 		if (!(type & SQ_CICB)) {
7427 			SQ_PUTLOCKS_EXIT(sq);
7428 		}
7429 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7430 		if (!(type & SQ_CICB)) {
7431 			count = sq->sq_count;
7432 			SQ_PUTLOCKS_ENTER(sq);
7433 			SUM_SQ_PUTCOUNTS(sq, count);
7434 		}
7435 	}
7436 
7437 	sq->sq_count++;
7438 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7439 	if (!(type & SQ_CICB)) {
7440 		ASSERT(count == 0);
7441 		sq->sq_flags |= SQ_EXCL;
7442 		ASSERT(sq->sq_needexcl > 0);
7443 		sq->sq_needexcl--;
7444 		if (sq->sq_needexcl == 0) {
7445 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7446 		}
7447 		SQ_PUTLOCKS_EXIT(sq);
7448 	}
7449 
7450 	mutex_exit(SQLOCK(sq));
7451 
7452 	cbp->cbp_func(cbp->cbp_arg);
7453 
7454 	/*
7455 	 * We drop the lock only for leavesq to re-acquire it.
7456 	 * Possible optimization is inline of leavesq.
7457 	 */
7458 	mutex_enter(SQLOCK(sq));
7459 	callbparams_free(sq, cbp);
7460 	mutex_exit(SQLOCK(sq));
7461 	leavesq(sq, SQ_CALLBACK);
7462 }
7463 
7464 /*
7465  * No need to grab sq_putlocks here. See comment in strsubr.h that
7466  * explains when sq_putlocks are used.
7467  *
7468  * sq_count (or one of the sq_putcounts) has already been
7469  * decremented by the caller, and if SQ_QUEUED, we need to call
7470  * drain_syncq (the global syncq drain).
7471  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7472  * one of two states, non-CIPUT perimeter, and we need to clear
7473  * it, or we went exclusive in the put procedure.  In any case,
7474  * we want to clear the bit now, and it is probably easier to do
7475  * this at the beginning of this function (remember, we hold
7476  * the SQLOCK).  Lastly, if there are other messages queued
7477  * on the syncq (and not for our destination), enable the syncq
7478  * for background work.
7479  */
7480 
7481 /* ARGSUSED */
7482 void
7483 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7484 {
7485 	uint16_t	flags = sq->sq_flags;
7486 
7487 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7488 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7489 
7490 	/* Clear SQ_EXCL if set in passflags */
7491 	if (passflags & SQ_EXCL) {
7492 		flags &= ~SQ_EXCL;
7493 	}
7494 	if (flags & SQ_WANTWAKEUP) {
7495 		flags &= ~SQ_WANTWAKEUP;
7496 		cv_broadcast(&sq->sq_wait);
7497 	}
7498 	if (flags & SQ_WANTEXWAKEUP) {
7499 		flags &= ~SQ_WANTEXWAKEUP;
7500 		cv_broadcast(&sq->sq_exitwait);
7501 	}
7502 	sq->sq_flags = flags;
7503 
7504 	/*
7505 	 * We have cleared SQ_EXCL if we were asked to, and started
7506 	 * the wakeup process for waiters.  If there are no writers
7507 	 * then we need to drain the syncq if we were told to, or
7508 	 * enable the background thread to do it.
7509 	 */
7510 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7511 		if ((passflags & SQ_QUEUED) ||
7512 		    (sq->sq_svcflags & SQ_DISABLED)) {
7513 			/* drain_syncq will take care of events in the list */
7514 			drain_syncq(sq);
7515 			return;
7516 		} else if (flags & SQ_QUEUED) {
7517 			sqenable(sq);
7518 		}
7519 	}
7520 	/* Drop the SQLOCK on exit */
7521 	mutex_exit(SQLOCK(sq));
7522 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7523 	    "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7524 }
7525 
7526 void
7527 set_qend(queue_t *q)
7528 {
7529 	mutex_enter(QLOCK(q));
7530 	if (!O_SAMESTR(q))
7531 		q->q_flag |= QEND;
7532 	else
7533 		q->q_flag &= ~QEND;
7534 	mutex_exit(QLOCK(q));
7535 	q = _OTHERQ(q);
7536 	mutex_enter(QLOCK(q));
7537 	if (!O_SAMESTR(q))
7538 		q->q_flag |= QEND;
7539 	else
7540 		q->q_flag &= ~QEND;
7541 	mutex_exit(QLOCK(q));
7542 }
7543 
7544 /*
7545  * Set QFULL in next service procedure queue (that cares) if not already
7546  * set and if there are already more messages on the syncq than
7547  * sq_max_size.  If sq_max_size is 0, no flow control will be asserted on
7548  * any syncq.
7549  *
7550  * The fq here is the next queue with a service procedure.  This is where
7551  * we would fail canputnext, so this is where we need to set QFULL.
7552  * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7553  *
7554  * We already have QLOCK at this point. To avoid cross-locks with
7555  * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7556  * SQLOCK and sd_reflock, we need to drop respective locks first.
7557  */
7558 void
7559 set_qfull(queue_t *q)
7560 {
7561 	queue_t		*fq = NULL;
7562 
7563 	ASSERT(MUTEX_HELD(QLOCK(q)));
7564 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7565 	    (q->q_syncqmsgs > sq_max_size)) {
7566 		if ((fq = q->q_nfsrv) == q) {
7567 			fq->q_flag |= QFULL;
7568 		} else {
7569 			mutex_exit(QLOCK(q));
7570 			mutex_enter(QLOCK(fq));
7571 			fq->q_flag |= QFULL;
7572 			mutex_exit(QLOCK(fq));
7573 			mutex_enter(QLOCK(q));
7574 		}
7575 	}
7576 }
7577 
7578 void
7579 clr_qfull(queue_t *q)
7580 {
7581 	queue_t	*oq = q;
7582 
7583 	q = q->q_nfsrv;
7584 	/* Fast check if there is any work to do before getting the lock. */
7585 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7586 		return;
7587 	}
7588 
7589 	/*
7590 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7591 	 * for QFULL being set.
7592 	 */
7593 	mutex_enter(QLOCK(q));
7594 	/*
7595 	 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7596 	 * Hence clear the QFULL.
7597 	 * If both q_count and q_mblkcnt are less than the hiwat mark,
7598 	 * clear the QFULL.
7599 	 */
7600 	if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7601 	    (q->q_mblkcnt < q->q_hiwat))) {
7602 		q->q_flag &= ~QFULL;
7603 		/*
7604 		 * A little more confusing, how about this way:
7605 		 * if someone wants to write,
7606 		 * AND
7607 		 *    both counts are less than the lowat mark
7608 		 *    OR
7609 		 *    the lowat mark is zero
7610 		 * THEN
7611 		 * backenable
7612 		 */
7613 		if ((q->q_flag & QWANTW) &&
7614 		    (((q->q_count < q->q_lowat) &&
7615 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7616 			q->q_flag &= ~QWANTW;
7617 			mutex_exit(QLOCK(q));
7618 			backenable(oq, 0);
7619 		} else
7620 			mutex_exit(QLOCK(q));
7621 	} else
7622 		mutex_exit(QLOCK(q));
7623 }
7624 
7625 /*
7626  * Set the forward service procedure pointer.
7627  *
7628  * Called at insert-time to cache a queue's next forward service procedure in
7629  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7630  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7631  * inserted does not have a service procedure, then q_nfsrv points to the next
7632  * queue forward that has a service procedure.  If the queue is at the logical
7633  * end of the stream (driver for write side, stream head for the read side)
7634  * and does not have a service procedure, then q_nfsrv also points to itself.
7635  */
7636 void
7637 set_nfsrv_ptr(
7638 	queue_t  *rnew,		/* read queue pointer to new module */
7639 	queue_t  *wnew,		/* write queue pointer to new module */
7640 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7641 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7642 {
7643 	queue_t *qp;
7644 
7645 	if (prev_wq->q_next == NULL) {
7646 		/*
7647 		 * Insert the driver, initialize the driver and stream head.
7648 		 * In this case, prev_rq/prev_wq should be the stream head.
7649 		 * _I_INSERT does not allow inserting a driver.  Make sure
7650 		 * that it is not an insertion.
7651 		 */
7652 		ASSERT(!(rnew->q_flag & _QINSERTING));
7653 		wnew->q_nfsrv = wnew;
7654 		if (rnew->q_qinfo->qi_srvp)
7655 			rnew->q_nfsrv = rnew;
7656 		else
7657 			rnew->q_nfsrv = prev_rq;
7658 		prev_rq->q_nfsrv = prev_rq;
7659 		prev_wq->q_nfsrv = prev_wq;
7660 	} else {
7661 		/*
7662 		 * set up read side q_nfsrv pointer.  This MUST be done
7663 		 * before setting the write side, because the setting of
7664 		 * the write side for a fifo may depend on it.
7665 		 *
7666 		 * Suppose we have a fifo that only has pipemod pushed.
7667 		 * pipemod has no read or write service procedures, so
7668 		 * nfsrv for both pipemod queues points to prev_rq (the
7669 		 * stream read head).  Now push bufmod (which has only a
7670 		 * read service procedure).  Doing the write side first,
7671 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7672 		 * is WRONG; the next queue forward from wnew with a
7673 		 * service procedure will be rnew, not the stream read head.
7674 		 * Since the downstream queue (which in the case of a fifo
7675 		 * is the read queue rnew) can affect upstream queues, it
7676 		 * needs to be done first.  Setting up the read side first
7677 		 * sets nfsrv for both pipemod queues to rnew and then
7678 		 * when the write side is set up, wnew-q_nfsrv will also
7679 		 * point to rnew.
7680 		 */
7681 		if (rnew->q_qinfo->qi_srvp) {
7682 			/*
7683 			 * use _OTHERQ() because, if this is a pipe, next
7684 			 * module may have been pushed from other end and
7685 			 * q_next could be a read queue.
7686 			 */
7687 			qp = _OTHERQ(prev_wq->q_next);
7688 			while (qp && qp->q_nfsrv != qp) {
7689 				qp->q_nfsrv = rnew;
7690 				qp = backq(qp);
7691 			}
7692 			rnew->q_nfsrv = rnew;
7693 		} else
7694 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7695 
7696 		/* set up write side q_nfsrv pointer */
7697 		if (wnew->q_qinfo->qi_srvp) {
7698 			wnew->q_nfsrv = wnew;
7699 
7700 			/*
7701 			 * For insertion, need to update nfsrv of the modules
7702 			 * above which do not have a service routine.
7703 			 */
7704 			if (rnew->q_flag & _QINSERTING) {
7705 				for (qp = prev_wq;
7706 				    qp != NULL && qp->q_nfsrv != qp;
7707 				    qp = backq(qp)) {
7708 					qp->q_nfsrv = wnew->q_nfsrv;
7709 				}
7710 			}
7711 		} else {
7712 			if (prev_wq->q_next == prev_rq)
7713 				/*
7714 				 * Since prev_wq/prev_rq are the middle of a
7715 				 * fifo, wnew/rnew will also be the middle of
7716 				 * a fifo and wnew's nfsrv is same as rnew's.
7717 				 */
7718 				wnew->q_nfsrv = rnew->q_nfsrv;
7719 			else
7720 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7721 		}
7722 	}
7723 }
7724 
7725 /*
7726  * Reset the forward service procedure pointer; called at remove-time.
7727  */
7728 void
7729 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7730 {
7731 	queue_t *tmp_qp;
7732 
7733 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7734 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7735 		for (tmp_qp = backq(wqp);
7736 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7737 		    tmp_qp = backq(tmp_qp)) {
7738 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7739 		}
7740 	}
7741 
7742 	/* reset the read side q_nfsrv pointer */
7743 	if (rqp->q_qinfo->qi_srvp) {
7744 		if (wqp->q_next) {	/* non-driver case */
7745 			tmp_qp = _OTHERQ(wqp->q_next);
7746 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7747 				/* Note that rqp->q_next cannot be NULL */
7748 				ASSERT(rqp->q_next != NULL);
7749 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7750 				tmp_qp = backq(tmp_qp);
7751 			}
7752 		}
7753 	}
7754 }
7755 
7756 /*
7757  * This routine should be called after all stream geometry changes to update
7758  * the stream head cached struio() rd/wr queue pointers. Note must be called
7759  * with the streamlock()ed.
7760  *
7761  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7762  *	 an explicit synchronous barrier module queue. That is, a queue that
7763  *	 has specified a struio() type.
7764  */
7765 static void
7766 strsetuio(stdata_t *stp)
7767 {
7768 	queue_t *wrq;
7769 
7770 	if (stp->sd_flag & STPLEX) {
7771 		/*
7772 		 * Not streamhead, but a mux, so no Synchronous STREAMS.
7773 		 */
7774 		stp->sd_struiowrq = NULL;
7775 		stp->sd_struiordq = NULL;
7776 		return;
7777 	}
7778 	/*
7779 	 * Scan the write queue(s) while synchronous
7780 	 * until we find a qinfo uio type specified.
7781 	 */
7782 	wrq = stp->sd_wrq->q_next;
7783 	while (wrq) {
7784 		if (wrq->q_struiot == STRUIOT_NONE) {
7785 			wrq = 0;
7786 			break;
7787 		}
7788 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7789 			break;
7790 		if (! _SAMESTR(wrq)) {
7791 			wrq = 0;
7792 			break;
7793 		}
7794 		wrq = wrq->q_next;
7795 	}
7796 	stp->sd_struiowrq = wrq;
7797 	/*
7798 	 * Scan the read queue(s) while synchronous
7799 	 * until we find a qinfo uio type specified.
7800 	 */
7801 	wrq = stp->sd_wrq->q_next;
7802 	while (wrq) {
7803 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7804 			wrq = 0;
7805 			break;
7806 		}
7807 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7808 			break;
7809 		if (! _SAMESTR(wrq)) {
7810 			wrq = 0;
7811 			break;
7812 		}
7813 		wrq = wrq->q_next;
7814 	}
7815 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7816 }
7817 
7818 static int
7819 pass_rput(queue_t *q, mblk_t *mp)
7820 {
7821 	putnext(q, mp);
7822 	return (0);
7823 }
7824 
7825 /*
7826  * pass_wput, unblocks the passthru queues, so that
7827  * messages can arrive at muxs lower read queue, before
7828  * I_LINK/I_UNLINK is acked/nacked.
7829  */
7830 static int
7831 pass_wput(queue_t *q, mblk_t *mp)
7832 {
7833 	syncq_t *sq;
7834 
7835 	sq = _RD(q)->q_syncq;
7836 	if (sq->sq_flags & SQ_BLOCKED)
7837 		unblocksq(sq, SQ_BLOCKED, 0);
7838 	putnext(q, mp);
7839 	return (0);
7840 }
7841 
7842 /*
7843  * Set up queues for the link/unlink.
7844  * Create a new queue and block it and then insert it
7845  * below the stream head on the lower stream.
7846  * This prevents any messages from arriving during the setq
7847  * as well as while the mux is processing the LINK/I_UNLINK.
7848  * The blocked passq is unblocked once the LINK/I_UNLINK has
7849  * been acked or nacked or if a message is generated and sent
7850  * down muxs write put procedure.
7851  * See pass_wput().
7852  *
7853  * After the new queue is inserted, all messages coming from below are
7854  * blocked. The call to strlock will ensure that all activity in the stream head
7855  * read queue syncq is stopped (sq_count drops to zero).
7856  */
7857 static queue_t *
7858 link_addpassthru(stdata_t *stpdown)
7859 {
7860 	queue_t *passq;
7861 	sqlist_t sqlist;
7862 
7863 	passq = allocq();
7864 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7865 	/* setq might sleep in allocator - avoid holding locks. */
7866 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7867 	    SQ_CI|SQ_CO, B_FALSE);
7868 	claimq(passq);
7869 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7870 	insertq(STREAM(passq), passq);
7871 
7872 	/*
7873 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7874 	 * since we are going to change q_ptr in the stream head.  Note that
7875 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7876 	 */
7877 	sqlist.sqlist_head = NULL;
7878 	sqlist.sqlist_index = 0;
7879 	sqlist.sqlist_size = sizeof (sqlist_t);
7880 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7881 	strlock(stpdown, &sqlist);
7882 	strunlock(stpdown, &sqlist);
7883 
7884 	releaseq(passq);
7885 	return (passq);
7886 }
7887 
7888 /*
7889  * Let messages flow up into the mux by removing
7890  * the passq.
7891  */
7892 static void
7893 link_rempassthru(queue_t *passq)
7894 {
7895 	claimq(passq);
7896 	removeq(passq);
7897 	releaseq(passq);
7898 	freeq(passq);
7899 }
7900 
7901 /*
7902  * Wait for the condition variable pointed to by `cvp' to be signaled,
7903  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7904  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7905  * then the wait will be non-interruptible.
7906  *
7907  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7908  */
7909 clock_t
7910 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7911 {
7912 	clock_t ret;
7913 
7914 	if (tim < 0) {
7915 		if (nosigs) {
7916 			cv_wait(cvp, mp);
7917 			ret = 1;
7918 		} else {
7919 			ret = cv_wait_sig(cvp, mp);
7920 		}
7921 	} else if (tim > 0) {
7922 		/*
7923 		 * convert milliseconds to clock ticks
7924 		 */
7925 		if (nosigs) {
7926 			ret = cv_reltimedwait(cvp, mp,
7927 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7928 		} else {
7929 			ret = cv_reltimedwait_sig(cvp, mp,
7930 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7931 		}
7932 	} else {
7933 		ret = -1;
7934 	}
7935 	return (ret);
7936 }
7937 
7938 /*
7939  * Wait until the stream head can determine if it is at the mark but
7940  * don't wait forever to prevent a race condition between the "mark" state
7941  * in the stream head and any mark state in the caller/user of this routine.
7942  *
7943  * This is used by sockets and for a socket it would be incorrect
7944  * to return a failure for SIOCATMARK when there is no data in the receive
7945  * queue and the marked urgent data is traveling up the stream.
7946  *
7947  * This routine waits until the mark is known by waiting for one of these
7948  * three events:
7949  *	The stream head read queue becoming non-empty (including an EOF).
7950  *	The STRATMARK flag being set (due to a MSGMARKNEXT message).
7951  *	The STRNOTATMARK flag being set (which indicates that the transport
7952  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7953  *	the mark).
7954  *
7955  * The routine returns 1 if the stream is at the mark; 0 if it can
7956  * be determined that the stream is not at the mark.
7957  * If the wait times out and it can't determine
7958  * whether or not the stream might be at the mark the routine will return -1.
7959  *
7960  * Note: This routine should only be used when a mark is pending i.e.,
7961  * in the socket case the SIGURG has been posted.
7962  * Note2: This can not wakeup just because synchronous streams indicate
7963  * that data is available since it is not possible to use the synchronous
7964  * streams interfaces to determine the b_flag value for the data queued below
7965  * the stream head.
7966  */
7967 int
7968 strwaitmark(vnode_t *vp)
7969 {
7970 	struct stdata *stp = vp->v_stream;
7971 	queue_t *rq = _RD(stp->sd_wrq);
7972 	int mark;
7973 
7974 	mutex_enter(&stp->sd_lock);
7975 	while (rq->q_first == NULL &&
7976 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7977 		stp->sd_flag |= RSLEEP;
7978 
7979 		/* Wait for 100 milliseconds for any state change. */
7980 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7981 			mutex_exit(&stp->sd_lock);
7982 			return (-1);
7983 		}
7984 	}
7985 	if (stp->sd_flag & STRATMARK)
7986 		mark = 1;
7987 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7988 		mark = 1;
7989 	else
7990 		mark = 0;
7991 
7992 	mutex_exit(&stp->sd_lock);
7993 	return (mark);
7994 }
7995 
7996 /*
7997  * Set a read side error. If persist is set change the socket error
7998  * to persistent. If errfunc is set install the function as the exported
7999  * error handler.
8000  */
8001 void
8002 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8003 {
8004 	struct stdata *stp = vp->v_stream;
8005 
8006 	mutex_enter(&stp->sd_lock);
8007 	stp->sd_rerror = error;
8008 	if (error == 0 && errfunc == NULL)
8009 		stp->sd_flag &= ~STRDERR;
8010 	else
8011 		stp->sd_flag |= STRDERR;
8012 	if (persist) {
8013 		stp->sd_flag &= ~STRDERRNONPERSIST;
8014 	} else {
8015 		stp->sd_flag |= STRDERRNONPERSIST;
8016 	}
8017 	stp->sd_rderrfunc = errfunc;
8018 	if (error != 0 || errfunc != NULL) {
8019 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8020 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8021 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8022 
8023 		mutex_exit(&stp->sd_lock);
8024 		pollwakeup(&stp->sd_pollist, POLLERR);
8025 		mutex_enter(&stp->sd_lock);
8026 
8027 		if (stp->sd_sigflags & S_ERROR)
8028 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8029 	}
8030 	mutex_exit(&stp->sd_lock);
8031 }
8032 
8033 /*
8034  * Set a write side error. If persist is set change the socket error
8035  * to persistent.
8036  */
8037 void
8038 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8039 {
8040 	struct stdata *stp = vp->v_stream;
8041 
8042 	mutex_enter(&stp->sd_lock);
8043 	stp->sd_werror = error;
8044 	if (error == 0 && errfunc == NULL)
8045 		stp->sd_flag &= ~STWRERR;
8046 	else
8047 		stp->sd_flag |= STWRERR;
8048 	if (persist) {
8049 		stp->sd_flag &= ~STWRERRNONPERSIST;
8050 	} else {
8051 		stp->sd_flag |= STWRERRNONPERSIST;
8052 	}
8053 	stp->sd_wrerrfunc = errfunc;
8054 	if (error != 0 || errfunc != NULL) {
8055 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8056 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8057 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8058 
8059 		mutex_exit(&stp->sd_lock);
8060 		pollwakeup(&stp->sd_pollist, POLLERR);
8061 		mutex_enter(&stp->sd_lock);
8062 
8063 		if (stp->sd_sigflags & S_ERROR)
8064 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8065 	}
8066 	mutex_exit(&stp->sd_lock);
8067 }
8068 
8069 /*
8070  * Make the stream return 0 (EOF) when all data has been read.
8071  * No effect on write side.
8072  */
8073 void
8074 strseteof(vnode_t *vp, int eof)
8075 {
8076 	struct stdata *stp = vp->v_stream;
8077 
8078 	mutex_enter(&stp->sd_lock);
8079 	if (!eof) {
8080 		stp->sd_flag &= ~STREOF;
8081 		mutex_exit(&stp->sd_lock);
8082 		return;
8083 	}
8084 	stp->sd_flag |= STREOF;
8085 	if (stp->sd_flag & RSLEEP) {
8086 		stp->sd_flag &= ~RSLEEP;
8087 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8088 	}
8089 
8090 	mutex_exit(&stp->sd_lock);
8091 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8092 	mutex_enter(&stp->sd_lock);
8093 
8094 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8095 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8096 	mutex_exit(&stp->sd_lock);
8097 }
8098 
8099 void
8100 strflushrq(vnode_t *vp, int flag)
8101 {
8102 	struct stdata *stp = vp->v_stream;
8103 
8104 	mutex_enter(&stp->sd_lock);
8105 	flushq(_RD(stp->sd_wrq), flag);
8106 	mutex_exit(&stp->sd_lock);
8107 }
8108 
8109 void
8110 strsetrputhooks(vnode_t *vp, uint_t flags,
8111     msgfunc_t protofunc, msgfunc_t miscfunc)
8112 {
8113 	struct stdata *stp = vp->v_stream;
8114 
8115 	mutex_enter(&stp->sd_lock);
8116 
8117 	if (protofunc == NULL)
8118 		stp->sd_rprotofunc = strrput_proto;
8119 	else
8120 		stp->sd_rprotofunc = protofunc;
8121 
8122 	if (miscfunc == NULL)
8123 		stp->sd_rmiscfunc = strrput_misc;
8124 	else
8125 		stp->sd_rmiscfunc = miscfunc;
8126 
8127 	if (flags & SH_CONSOL_DATA)
8128 		stp->sd_rput_opt |= SR_CONSOL_DATA;
8129 	else
8130 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8131 
8132 	if (flags & SH_SIGALLDATA)
8133 		stp->sd_rput_opt |= SR_SIGALLDATA;
8134 	else
8135 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
8136 
8137 	if (flags & SH_IGN_ZEROLEN)
8138 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8139 	else
8140 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8141 
8142 	mutex_exit(&stp->sd_lock);
8143 }
8144 
8145 void
8146 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8147 {
8148 	struct stdata *stp = vp->v_stream;
8149 
8150 	mutex_enter(&stp->sd_lock);
8151 	stp->sd_closetime = closetime;
8152 
8153 	if (flags & SH_SIGPIPE)
8154 		stp->sd_wput_opt |= SW_SIGPIPE;
8155 	else
8156 		stp->sd_wput_opt &= ~SW_SIGPIPE;
8157 	if (flags & SH_RECHECK_ERR)
8158 		stp->sd_wput_opt |= SW_RECHECK_ERR;
8159 	else
8160 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8161 
8162 	mutex_exit(&stp->sd_lock);
8163 }
8164 
8165 void
8166 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8167 {
8168 	struct stdata *stp = vp->v_stream;
8169 
8170 	mutex_enter(&stp->sd_lock);
8171 
8172 	stp->sd_rputdatafunc = rdatafunc;
8173 	stp->sd_wputdatafunc = wdatafunc;
8174 
8175 	mutex_exit(&stp->sd_lock);
8176 }
8177 
8178 /* Used within framework when the queue is already locked */
8179 void
8180 qenable_locked(queue_t *q)
8181 {
8182 	stdata_t *stp = STREAM(q);
8183 
8184 	ASSERT(MUTEX_HELD(QLOCK(q)));
8185 
8186 	if (!q->q_qinfo->qi_srvp)
8187 		return;
8188 
8189 	/*
8190 	 * Do not place on run queue if already enabled or closing.
8191 	 */
8192 	if (q->q_flag & (QWCLOSE|QENAB))
8193 		return;
8194 
8195 	/*
8196 	 * mark queue enabled and place on run list if it is not already being
8197 	 * serviced. If it is serviced, the runservice() function will detect
8198 	 * that QENAB is set and call service procedure before clearing
8199 	 * QINSERVICE flag.
8200 	 */
8201 	q->q_flag |= QENAB;
8202 	if (q->q_flag & QINSERVICE)
8203 		return;
8204 
8205 	/* Record the time of qenable */
8206 	q->q_qtstamp = ddi_get_lbolt();
8207 
8208 	/*
8209 	 * Put the queue in the stp list and schedule it for background
8210 	 * processing if it is not already scheduled or if stream head does not
8211 	 * intent to process it in the foreground later by setting
8212 	 * STRS_WILLSERVICE flag.
8213 	 */
8214 	mutex_enter(&stp->sd_qlock);
8215 	/*
8216 	 * If there are already something on the list, stp flags should show
8217 	 * intention to drain it.
8218 	 */
8219 	IMPLY(STREAM_NEEDSERVICE(stp),
8220 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8221 
8222 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8223 	stp->sd_nqueues++;
8224 
8225 	/*
8226 	 * If no one will drain this stream we are the first producer and
8227 	 * need to schedule it for background thread.
8228 	 */
8229 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8230 		/*
8231 		 * No one will service this stream later, so we have to
8232 		 * schedule it now.
8233 		 */
8234 		STRSTAT(stenables);
8235 		stp->sd_svcflags |= STRS_SCHEDULED;
8236 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8237 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8238 
8239 		if (stp->sd_servid == NULL) {
8240 			/*
8241 			 * Task queue failed so fail over to the backup
8242 			 * servicing thread.
8243 			 */
8244 			STRSTAT(taskqfails);
8245 			/*
8246 			 * It is safe to clear STRS_SCHEDULED flag because it
8247 			 * was set by this thread above.
8248 			 */
8249 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8250 
8251 			/*
8252 			 * Failover scheduling is protected by service_queue
8253 			 * lock.
8254 			 */
8255 			mutex_enter(&service_queue);
8256 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8257 			ASSERT(q->q_link == NULL);
8258 			/*
8259 			 * Append the queue to qhead/qtail list.
8260 			 */
8261 			if (qhead == NULL)
8262 				qhead = q;
8263 			else
8264 				qtail->q_link = q;
8265 			qtail = q;
8266 			/*
8267 			 * Clear stp queue list.
8268 			 */
8269 			stp->sd_qhead = stp->sd_qtail = NULL;
8270 			stp->sd_nqueues = 0;
8271 			/*
8272 			 * Wakeup background queue processing thread.
8273 			 */
8274 			cv_signal(&services_to_run);
8275 			mutex_exit(&service_queue);
8276 		}
8277 	}
8278 	mutex_exit(&stp->sd_qlock);
8279 }
8280 
8281 static void
8282 queue_service(queue_t *q)
8283 {
8284 	/*
8285 	 * The queue in the list should have
8286 	 * QENAB flag set and should not have
8287 	 * QINSERVICE flag set. QINSERVICE is
8288 	 * set when the queue is dequeued and
8289 	 * qenable_locked doesn't enqueue a
8290 	 * queue with QINSERVICE set.
8291 	 */
8292 
8293 	ASSERT(!(q->q_flag & QINSERVICE));
8294 	ASSERT((q->q_flag & QENAB));
8295 	mutex_enter(QLOCK(q));
8296 	q->q_flag &= ~QENAB;
8297 	q->q_flag |= QINSERVICE;
8298 	mutex_exit(QLOCK(q));
8299 	runservice(q);
8300 }
8301 
8302 static void
8303 syncq_service(syncq_t *sq)
8304 {
8305 	STRSTAT(syncqservice);
8306 	mutex_enter(SQLOCK(sq));
8307 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8308 	ASSERT(sq->sq_servcount != 0);
8309 	ASSERT(sq->sq_next == NULL);
8310 
8311 	/* if we came here from the background thread, clear the flag */
8312 	if (sq->sq_svcflags & SQ_BGTHREAD)
8313 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8314 
8315 	/* let drain_syncq know that it's being called in the background */
8316 	sq->sq_svcflags |= SQ_SERVICE;
8317 	drain_syncq(sq);
8318 }
8319 
8320 static void
8321 qwriter_outer_service(syncq_t *outer)
8322 {
8323 	/*
8324 	 * Note that SQ_WRITER is used on the outer perimeter
8325 	 * to signal that a qwriter(OUTER) is either investigating
8326 	 * running or that it is actually running a function.
8327 	 */
8328 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8329 
8330 	/*
8331 	 * All inner syncq are empty and have SQ_WRITER set
8332 	 * to block entering the outer perimeter.
8333 	 *
8334 	 * We do not need to explicitly call write_now since
8335 	 * outer_exit does it for us.
8336 	 */
8337 	outer_exit(outer);
8338 }
8339 
8340 static void
8341 mblk_free(mblk_t *mp)
8342 {
8343 	dblk_t *dbp = mp->b_datap;
8344 	frtn_t *frp = dbp->db_frtnp;
8345 
8346 	mp->b_next = NULL;
8347 	if (dbp->db_fthdr != NULL)
8348 		str_ftfree(dbp);
8349 
8350 	ASSERT(dbp->db_fthdr == NULL);
8351 	frp->free_func(frp->free_arg);
8352 	ASSERT(dbp->db_mblk == mp);
8353 
8354 	if (dbp->db_credp != NULL) {
8355 		crfree(dbp->db_credp);
8356 		dbp->db_credp = NULL;
8357 	}
8358 	dbp->db_cpid = -1;
8359 	dbp->db_struioflag = 0;
8360 	dbp->db_struioun.cksum.flags = 0;
8361 
8362 	kmem_cache_free(dbp->db_cache, dbp);
8363 }
8364 
8365 /*
8366  * Background processing of the stream queue list.
8367  */
8368 static void
8369 stream_service(stdata_t *stp)
8370 {
8371 	queue_t *q;
8372 
8373 	mutex_enter(&stp->sd_qlock);
8374 
8375 	STR_SERVICE(stp, q);
8376 
8377 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8378 	stp->sd_servid = NULL;
8379 	cv_signal(&stp->sd_qcv);
8380 	mutex_exit(&stp->sd_qlock);
8381 }
8382 
8383 /*
8384  * Foreground processing of the stream queue list.
8385  */
8386 void
8387 stream_runservice(stdata_t *stp)
8388 {
8389 	queue_t *q;
8390 
8391 	mutex_enter(&stp->sd_qlock);
8392 	STRSTAT(rservice);
8393 	/*
8394 	 * We are going to drain this stream queue list, so qenable_locked will
8395 	 * not schedule it until we finish.
8396 	 */
8397 	stp->sd_svcflags |= STRS_WILLSERVICE;
8398 
8399 	STR_SERVICE(stp, q);
8400 
8401 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8402 	mutex_exit(&stp->sd_qlock);
8403 	/*
8404 	 * Help backup background thread to drain the qhead/qtail list.
8405 	 */
8406 	while (qhead != NULL) {
8407 		STRSTAT(qhelps);
8408 		mutex_enter(&service_queue);
8409 		DQ(q, qhead, qtail, q_link);
8410 		mutex_exit(&service_queue);
8411 		if (q != NULL)
8412 			queue_service(q);
8413 	}
8414 }
8415 
8416 void
8417 stream_willservice(stdata_t *stp)
8418 {
8419 	mutex_enter(&stp->sd_qlock);
8420 	stp->sd_svcflags |= STRS_WILLSERVICE;
8421 	mutex_exit(&stp->sd_qlock);
8422 }
8423 
8424 /*
8425  * Replace the cred currently in the mblk with a different one.
8426  * Also update db_cpid.
8427  */
8428 void
8429 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8430 {
8431 	dblk_t *dbp = mp->b_datap;
8432 	cred_t *ocr = dbp->db_credp;
8433 
8434 	ASSERT(cr != NULL);
8435 
8436 	if (cr != ocr) {
8437 		crhold(dbp->db_credp = cr);
8438 		if (ocr != NULL)
8439 			crfree(ocr);
8440 	}
8441 	/* Don't overwrite with NOPID */
8442 	if (cpid != NOPID)
8443 		dbp->db_cpid = cpid;
8444 }
8445 
8446 /*
8447  * If the src message has a cred, then replace the cred currently in the mblk
8448  * with it.
8449  * Also update db_cpid.
8450  */
8451 void
8452 mblk_copycred(mblk_t *mp, const mblk_t *src)
8453 {
8454 	dblk_t *dbp = mp->b_datap;
8455 	cred_t *cr, *ocr;
8456 	pid_t cpid;
8457 
8458 	cr = msg_getcred(src, &cpid);
8459 	if (cr == NULL)
8460 		return;
8461 
8462 	ocr = dbp->db_credp;
8463 	if (cr != ocr) {
8464 		crhold(dbp->db_credp = cr);
8465 		if (ocr != NULL)
8466 			crfree(ocr);
8467 	}
8468 	/* Don't overwrite with NOPID */
8469 	if (cpid != NOPID)
8470 		dbp->db_cpid = cpid;
8471 }
8472 
8473 int
8474 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8475     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8476     uint32_t flags, int km_flags)
8477 {
8478 	int rc = 0;
8479 
8480 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8481 	if (mp->b_datap->db_type == M_DATA) {
8482 		/* Associate values for M_DATA type */
8483 		DB_CKSUMSTART(mp) = (intptr_t)start;
8484 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8485 		DB_CKSUMEND(mp) = (intptr_t)end;
8486 		DB_CKSUMFLAGS(mp) = flags;
8487 		DB_CKSUM16(mp) = (uint16_t)value;
8488 
8489 	} else {
8490 		pattrinfo_t pa_info;
8491 
8492 		ASSERT(mmd != NULL);
8493 
8494 		pa_info.type = PATTR_HCKSUM;
8495 		pa_info.len = sizeof (pattr_hcksum_t);
8496 
8497 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8498 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8499 
8500 			hck->hcksum_start_offset = start;
8501 			hck->hcksum_stuff_offset = stuff;
8502 			hck->hcksum_end_offset = end;
8503 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8504 			hck->hcksum_flags = flags;
8505 		} else {
8506 			rc = -1;
8507 		}
8508 	}
8509 	return (rc);
8510 }
8511 
8512 void
8513 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8514     uint32_t *start, uint32_t *stuff, uint32_t *end,
8515     uint32_t *value, uint32_t *flags)
8516 {
8517 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8518 	if (mp->b_datap->db_type == M_DATA) {
8519 		if (flags != NULL) {
8520 			*flags = DB_CKSUMFLAGS(mp) & HCK_FLAGS;
8521 			if ((*flags & (HCK_PARTIALCKSUM |
8522 			    HCK_FULLCKSUM)) != 0) {
8523 				if (value != NULL)
8524 					*value = (uint32_t)DB_CKSUM16(mp);
8525 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
8526 					if (start != NULL)
8527 						*start =
8528 						    (uint32_t)DB_CKSUMSTART(mp);
8529 					if (stuff != NULL)
8530 						*stuff =
8531 						    (uint32_t)DB_CKSUMSTUFF(mp);
8532 					if (end != NULL)
8533 						*end =
8534 						    (uint32_t)DB_CKSUMEND(mp);
8535 				}
8536 			}
8537 		}
8538 	} else {
8539 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8540 
8541 		ASSERT(mmd != NULL);
8542 
8543 		/* get hardware checksum attribute */
8544 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8545 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8546 
8547 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8548 			if (flags != NULL)
8549 				*flags = hck->hcksum_flags;
8550 			if (start != NULL)
8551 				*start = hck->hcksum_start_offset;
8552 			if (stuff != NULL)
8553 				*stuff = hck->hcksum_stuff_offset;
8554 			if (end != NULL)
8555 				*end = hck->hcksum_end_offset;
8556 			if (value != NULL)
8557 				*value = (uint32_t)
8558 				    hck->hcksum_cksum_val.inet_cksum;
8559 		}
8560 	}
8561 }
8562 
8563 void
8564 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8565 {
8566 	ASSERT(DB_TYPE(mp) == M_DATA);
8567 	ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8568 
8569 	/* Set the flags */
8570 	DB_LSOFLAGS(mp) |= flags;
8571 	DB_LSOMSS(mp) = mss;
8572 }
8573 
8574 void
8575 lso_info_cleanup(mblk_t *mp)
8576 {
8577 	ASSERT(DB_TYPE(mp) == M_DATA);
8578 
8579 	/* Clear the flags */
8580 	DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8581 	DB_LSOMSS(mp) = 0;
8582 }
8583 
8584 /*
8585  * Checksum buffer *bp for len bytes with psum partial checksum,
8586  * or 0 if none, and return the 16 bit partial checksum.
8587  */
8588 unsigned
8589 bcksum(uchar_t *bp, int len, unsigned int psum)
8590 {
8591 	int odd = len & 1;
8592 	extern unsigned int ip_ocsum();
8593 
8594 	if (((intptr_t)bp & 1) == 0 && !odd) {
8595 		/*
8596 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8597 		 */
8598 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8599 	}
8600 	if (((intptr_t)bp & 1) != 0) {
8601 		/*
8602 		 * Bp isn't 16 bit aligned.
8603 		 */
8604 		unsigned int tsum;
8605 
8606 #ifdef _LITTLE_ENDIAN
8607 		psum += *bp;
8608 #else
8609 		psum += *bp << 8;
8610 #endif
8611 		len--;
8612 		bp++;
8613 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8614 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8615 		if (len & 1) {
8616 			bp += len - 1;
8617 #ifdef _LITTLE_ENDIAN
8618 			psum += *bp << 8;
8619 #else
8620 			psum += *bp;
8621 #endif
8622 		}
8623 	} else {
8624 		/*
8625 		 * Bp is 16 bit aligned.
8626 		 */
8627 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8628 		if (odd) {
8629 			bp += len - 1;
8630 #ifdef _LITTLE_ENDIAN
8631 			psum += *bp;
8632 #else
8633 			psum += *bp << 8;
8634 #endif
8635 		}
8636 	}
8637 	/*
8638 	 * Normalize psum to 16 bits before returning the new partial
8639 	 * checksum. The max psum value before normalization is 0x3FDFE.
8640 	 */
8641 	return ((psum >> 16) + (psum & 0xFFFF));
8642 }
8643 
8644 boolean_t
8645 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8646 {
8647 	boolean_t rc;
8648 
8649 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8650 	if (DB_TYPE(mp) == M_DATA) {
8651 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8652 	} else {
8653 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8654 
8655 		ASSERT(mmd != NULL);
8656 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8657 	}
8658 	return (rc);
8659 }
8660 
8661 void
8662 freemsgchain(mblk_t *mp)
8663 {
8664 	mblk_t	*next;
8665 
8666 	while (mp != NULL) {
8667 		next = mp->b_next;
8668 		mp->b_next = NULL;
8669 
8670 		freemsg(mp);
8671 		mp = next;
8672 	}
8673 }
8674 
8675 mblk_t *
8676 copymsgchain(mblk_t *mp)
8677 {
8678 	mblk_t	*nmp = NULL;
8679 	mblk_t	**nmpp = &nmp;
8680 
8681 	for (; mp != NULL; mp = mp->b_next) {
8682 		if ((*nmpp = copymsg(mp)) == NULL) {
8683 			freemsgchain(nmp);
8684 			return (NULL);
8685 		}
8686 
8687 		nmpp = &((*nmpp)->b_next);
8688 	}
8689 
8690 	return (nmp);
8691 }
8692 
8693 /* NOTE: Do not add code after this point. */
8694 #undef QLOCK
8695 
8696 /*
8697  * Replacement for QLOCK macro for those that can't use it.
8698  */
8699 kmutex_t *
8700 QLOCK(queue_t *q)
8701 {
8702 	return (&(q)->q_lock);
8703 }
8704 
8705 /*
8706  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8707  */
8708 #undef runqueues
8709 void
8710 runqueues(void)
8711 {
8712 }
8713 
8714 #undef queuerun
8715 void
8716 queuerun(void)
8717 {
8718 }
8719 
8720 /*
8721  * Initialize the STR stack instance, which tracks autopush and persistent
8722  * links.
8723  */
8724 /* ARGSUSED */
8725 static void *
8726 str_stack_init(netstackid_t stackid, netstack_t *ns)
8727 {
8728 	str_stack_t	*ss;
8729 	int i;
8730 
8731 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8732 	ss->ss_netstack = ns;
8733 
8734 	/*
8735 	 * set up autopush
8736 	 */
8737 	sad_initspace(ss);
8738 
8739 	/*
8740 	 * set up mux_node structures.
8741 	 */
8742 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8743 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8744 	    ss->ss_devcnt), KM_SLEEP);
8745 	for (i = 0; i < ss->ss_devcnt; i++)
8746 		ss->ss_mux_nodes[i].mn_imaj = i;
8747 	return (ss);
8748 }
8749 
8750 /*
8751  * Note: run at zone shutdown and not destroy so that the PLINKs are
8752  * gone by the time other cleanup happens from the destroy callbacks.
8753  */
8754 static void
8755 str_stack_shutdown(netstackid_t stackid, void *arg)
8756 {
8757 	str_stack_t *ss = (str_stack_t *)arg;
8758 	int i;
8759 	cred_t *cr;
8760 
8761 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8762 	ASSERT(cr != NULL);
8763 
8764 	/* Undo all the I_PLINKs for this zone */
8765 	for (i = 0; i < ss->ss_devcnt; i++) {
8766 		struct mux_edge		*ep;
8767 		ldi_handle_t		lh;
8768 		ldi_ident_t		li;
8769 		int			ret;
8770 		int			rval;
8771 		dev_t			rdev;
8772 
8773 		ep = ss->ss_mux_nodes[i].mn_outp;
8774 		if (ep == NULL)
8775 			continue;
8776 		ret = ldi_ident_from_major((major_t)i, &li);
8777 		if (ret != 0) {
8778 			continue;
8779 		}
8780 		rdev = ep->me_dev;
8781 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8782 		    cr, &lh, li);
8783 		if (ret != 0) {
8784 			ldi_ident_release(li);
8785 			continue;
8786 		}
8787 
8788 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8789 		    cr, &rval);
8790 		if (ret) {
8791 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8792 			ldi_ident_release(li);
8793 			continue;
8794 		}
8795 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8796 
8797 		/* Close layered handles */
8798 		ldi_ident_release(li);
8799 	}
8800 	crfree(cr);
8801 
8802 	sad_freespace(ss);
8803 
8804 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8805 	ss->ss_mux_nodes = NULL;
8806 }
8807 
8808 /*
8809  * Free the structure; str_stack_shutdown did the other cleanup work.
8810  */
8811 /* ARGSUSED */
8812 static void
8813 str_stack_fini(netstackid_t stackid, void *arg)
8814 {
8815 	str_stack_t	*ss = (str_stack_t *)arg;
8816 
8817 	kmem_free(ss, sizeof (*ss));
8818 }
8819