xref: /illumos-gate/usr/src/uts/common/os/strsubr.c (revision 0803869a9bc0aff1c81287ac6eae15943cbba545)
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  */
30 
31 #include <sys/types.h>
32 #include <sys/sysmacros.h>
33 #include <sys/param.h>
34 #include <sys/errno.h>
35 #include <sys/signal.h>
36 #include <sys/proc.h>
37 #include <sys/conf.h>
38 #include <sys/cred.h>
39 #include <sys/user.h>
40 #include <sys/vnode.h>
41 #include <sys/file.h>
42 #include <sys/session.h>
43 #include <sys/stream.h>
44 #include <sys/strsubr.h>
45 #include <sys/stropts.h>
46 #include <sys/poll.h>
47 #include <sys/systm.h>
48 #include <sys/cpuvar.h>
49 #include <sys/uio.h>
50 #include <sys/cmn_err.h>
51 #include <sys/priocntl.h>
52 #include <sys/procset.h>
53 #include <sys/vmem.h>
54 #include <sys/bitmap.h>
55 #include <sys/kmem.h>
56 #include <sys/siginfo.h>
57 #include <sys/vtrace.h>
58 #include <sys/callb.h>
59 #include <sys/debug.h>
60 #include <sys/modctl.h>
61 #include <sys/vmsystm.h>
62 #include <vm/page.h>
63 #include <sys/atomic.h>
64 #include <sys/suntpi.h>
65 #include <sys/strlog.h>
66 #include <sys/promif.h>
67 #include <sys/project.h>
68 #include <sys/vm.h>
69 #include <sys/taskq.h>
70 #include <sys/sunddi.h>
71 #include <sys/sunldi_impl.h>
72 #include <sys/strsun.h>
73 #include <sys/isa_defs.h>
74 #include <sys/multidata.h>
75 #include <sys/pattr.h>
76 #include <sys/strft.h>
77 #include <sys/fs/snode.h>
78 #include <sys/zone.h>
79 #include <sys/open.h>
80 #include <sys/sunldi.h>
81 #include <sys/sad.h>
82 #include <sys/netstack.h>
83 
84 #define	O_SAMESTR(q)	(((q)->q_next) && \
85 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
86 
87 /*
88  * WARNING:
89  * The variables and routines in this file are private, belonging
90  * to the STREAMS subsystem. These should not be used by modules
91  * or drivers. Compatibility will not be guaranteed.
92  */
93 
94 /*
95  * Id value used to distinguish between different multiplexor links.
96  */
97 static int32_t lnk_id = 0;
98 
99 #define	STREAMS_LOPRI MINCLSYSPRI
100 static pri_t streams_lopri = STREAMS_LOPRI;
101 
102 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
103 typedef struct str_stat {
104 	kstat_named_t	sqenables;
105 	kstat_named_t	stenables;
106 	kstat_named_t	syncqservice;
107 	kstat_named_t	freebs;
108 	kstat_named_t	qwr_outer;
109 	kstat_named_t	rservice;
110 	kstat_named_t	strwaits;
111 	kstat_named_t	taskqfails;
112 	kstat_named_t	bufcalls;
113 	kstat_named_t	qhelps;
114 	kstat_named_t	qremoved;
115 	kstat_named_t	sqremoved;
116 	kstat_named_t	bcwaits;
117 	kstat_named_t	sqtoomany;
118 } str_stat_t;
119 
120 static str_stat_t str_statistics = {
121 	{ "sqenables",		KSTAT_DATA_UINT64 },
122 	{ "stenables",		KSTAT_DATA_UINT64 },
123 	{ "syncqservice",	KSTAT_DATA_UINT64 },
124 	{ "freebs",		KSTAT_DATA_UINT64 },
125 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
126 	{ "rservice",		KSTAT_DATA_UINT64 },
127 	{ "strwaits",		KSTAT_DATA_UINT64 },
128 	{ "taskqfails",		KSTAT_DATA_UINT64 },
129 	{ "bufcalls",		KSTAT_DATA_UINT64 },
130 	{ "qhelps",		KSTAT_DATA_UINT64 },
131 	{ "qremoved",		KSTAT_DATA_UINT64 },
132 	{ "sqremoved",		KSTAT_DATA_UINT64 },
133 	{ "bcwaits",		KSTAT_DATA_UINT64 },
134 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
135 };
136 
137 static kstat_t *str_kstat;
138 
139 /*
140  * qrunflag was used previously to control background scheduling of queues. It
141  * is not used anymore, but kept here in case some module still wants to access
142  * it via qready() and setqsched macros.
143  */
144 char qrunflag;			/*  Unused */
145 
146 /*
147  * Most of the streams scheduling is done via task queues. Task queues may fail
148  * for non-sleep dispatches, so there are two backup threads servicing failed
149  * requests for queues and syncqs. Both of these threads also service failed
150  * dispatches freebs requests. Queues are put in the list specified by `qhead'
151  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
152  * requests are put into `freebs_list' which has no tail pointer. All three
153  * lists are protected by a single `service_queue' lock and use
154  * `services_to_run' condition variable for signaling background threads. Use of
155  * a single lock should not be a problem because it is only used under heavy
156  * loads when task queues start to fail and at that time it may be a good idea
157  * to throttle scheduling requests.
158  *
159  * NOTE: queues and syncqs should be scheduled by two separate threads because
160  * queue servicing may be blocked waiting for a syncq which may be also
161  * scheduled for background execution. This may create a deadlock when only one
162  * thread is used for both.
163  */
164 
165 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
166 
167 static kmutex_t service_queue;		/* protects all of servicing vars */
168 static kcondvar_t services_to_run;	/* wake up background service thread */
169 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
170 
171 /*
172  * List of queues scheduled for background processing due to lack of resources
173  * in the task queues. Protected by service_queue lock;
174  */
175 static struct queue *qhead;
176 static struct queue *qtail;
177 
178 /*
179  * Same list for syncqs
180  */
181 static syncq_t *sqhead;
182 static syncq_t *sqtail;
183 
184 static mblk_t *freebs_list;	/* list of buffers to free */
185 
186 /*
187  * Backup threads for servicing queues and syncqs
188  */
189 kthread_t *streams_qbkgrnd_thread;
190 kthread_t *streams_sqbkgrnd_thread;
191 
192 /*
193  * Bufcalls related variables.
194  */
195 struct bclist	strbcalls;	/* list of waiting bufcalls */
196 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
197 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
198 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
199 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
200 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
201 
202 kmutex_t	strresources;	/* protects global resources */
203 kmutex_t	muxifier;	/* single-threads multiplexor creation */
204 
205 static void	*str_stack_init(netstackid_t stackid, netstack_t *ns);
206 static void	str_stack_shutdown(netstackid_t stackid, void *arg);
207 static void	str_stack_fini(netstackid_t stackid, void *arg);
208 
209 /*
210  * run_queues is no longer used, but is kept in case some 3rd party
211  * module/driver decides to use it.
212  */
213 int run_queues = 0;
214 
215 /*
216  * sq_max_size is the depth of the syncq (in number of messages) before
217  * qfill_syncq() starts QFULL'ing destination queues. As its primary
218  * consumer - IP is no longer D_MTPERMOD, but there may be other
219  * modules/drivers depend on this syncq flow control, we prefer to
220  * choose a large number as the default value. For potential
221  * performance gain, this value is tunable in /etc/system.
222  */
223 int sq_max_size = 10000;
224 
225 /*
226  * The number of ciputctrl structures per syncq and stream we create when
227  * needed.
228  */
229 int n_ciputctrl;
230 int max_n_ciputctrl = 16;
231 /*
232  * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
233  */
234 int min_n_ciputctrl = 2;
235 
236 /*
237  * Per-driver/module syncqs
238  * ========================
239  *
240  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
241  * perdm structures, new entries being added (and new syncqs allocated) when
242  * setq() encounters a module/driver with a streamtab that it hasn't seen
243  * before.
244  * The reason for this mechanism is that some modules and drivers share a
245  * common streamtab and it is necessary for those modules and drivers to also
246  * share a common PERMOD syncq.
247  *
248  * perdm_list --> dm_str == streamtab_1
249  *                dm_sq == syncq_1
250  *                dm_ref
251  *                dm_next --> dm_str == streamtab_2
252  *                            dm_sq == syncq_2
253  *                            dm_ref
254  *                            dm_next --> ... NULL
255  *
256  * The dm_ref field is incremented for each new driver/module that takes
257  * a reference to the perdm structure and hence shares the syncq.
258  * References are held in the fmodsw_impl_t structure for each STREAMS module
259  * or the dev_impl array (indexed by device major number) for each driver.
260  *
261  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
262  *		     ^                 ^ ^               ^
263  *                   |  ______________/  |               |
264  *                   | /                 |               |
265  * dev_impl:     ...|x|y|...          module A	      module B
266  *
267  * When a module/driver is unloaded the reference count is decremented and,
268  * when it falls to zero, the perdm structure is removed from the list and
269  * the syncq is freed (see rele_dm()).
270  */
271 perdm_t *perdm_list = NULL;
272 static krwlock_t perdm_rwlock;
273 cdevsw_impl_t *devimpl;
274 
275 extern struct qinit strdata;
276 extern struct qinit stwdata;
277 
278 static void runservice(queue_t *);
279 static void streams_bufcall_service(void);
280 static void streams_qbkgrnd_service(void);
281 static void streams_sqbkgrnd_service(void);
282 static syncq_t *new_syncq(void);
283 static void free_syncq(syncq_t *);
284 static void outer_insert(syncq_t *, syncq_t *);
285 static void outer_remove(syncq_t *, syncq_t *);
286 static void write_now(syncq_t *);
287 static void clr_qfull(queue_t *);
288 static void runbufcalls(void);
289 static void sqenable(syncq_t *);
290 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
291 static void wait_q_syncq(queue_t *);
292 static void backenable_insertedq(queue_t *);
293 
294 static void queue_service(queue_t *);
295 static void stream_service(stdata_t *);
296 static void syncq_service(syncq_t *);
297 static void qwriter_outer_service(syncq_t *);
298 static void mblk_free(mblk_t *);
299 #ifdef DEBUG
300 static int qprocsareon(queue_t *);
301 #endif
302 
303 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
304 static void reset_nfsrv_ptr(queue_t *, queue_t *);
305 void set_qfull(queue_t *);
306 
307 static void sq_run_events(syncq_t *);
308 static int propagate_syncq(queue_t *);
309 
310 static void	blocksq(syncq_t *, ushort_t, int);
311 static void	unblocksq(syncq_t *, ushort_t, int);
312 static int	dropsq(syncq_t *, uint16_t);
313 static void	emptysq(syncq_t *);
314 static sqlist_t *sqlist_alloc(struct stdata *, int);
315 static void	sqlist_free(sqlist_t *);
316 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
317 static void	sqlist_insert(sqlist_t *, syncq_t *);
318 static void	sqlist_insertall(sqlist_t *, queue_t *);
319 
320 static void	strsetuio(stdata_t *);
321 
322 struct kmem_cache *stream_head_cache;
323 struct kmem_cache *queue_cache;
324 struct kmem_cache *syncq_cache;
325 struct kmem_cache *qband_cache;
326 struct kmem_cache *linkinfo_cache;
327 struct kmem_cache *ciputctrl_cache = NULL;
328 
329 static linkinfo_t *linkinfo_list;
330 
331 /* Global esballoc throttling queue */
332 static esb_queue_t system_esbq;
333 
334 /* Array of esballoc throttling queues, of length esbq_nelem */
335 static esb_queue_t *volatile system_esbq_array;
336 static int esbq_nelem;
337 static kmutex_t esbq_lock;
338 static int esbq_log2_cpus_per_q = 0;
339 
340 /* Scale the system_esbq length by setting number of CPUs per queue. */
341 uint_t esbq_cpus_per_q = 1;
342 
343 /*
344  * esballoc tunable parameters.
345  */
346 int		esbq_max_qlen = 0x16;	/* throttled queue length */
347 clock_t		esbq_timeout = 0x8;	/* timeout to process esb queue */
348 
349 /*
350  * Routines to handle esballoc queueing.
351  */
352 static void esballoc_process_queue(esb_queue_t *);
353 static void esballoc_enqueue_mblk(mblk_t *);
354 static void esballoc_timer(void *);
355 static void esballoc_set_timer(esb_queue_t *, clock_t);
356 static void esballoc_mblk_free(mblk_t *);
357 
358 /*
359  *  Qinit structure and Module_info structures
360  *	for passthru read and write queues
361  */
362 
363 static void pass_wput(queue_t *, mblk_t *);
364 static queue_t *link_addpassthru(stdata_t *);
365 static void link_rempassthru(queue_t *);
366 
367 struct  module_info passthru_info = {
368 	0,
369 	"passthru",
370 	0,
371 	INFPSZ,
372 	STRHIGH,
373 	STRLOW
374 };
375 
376 struct  qinit passthru_rinit = {
377 	(int (*)())putnext,
378 	NULL,
379 	NULL,
380 	NULL,
381 	NULL,
382 	&passthru_info,
383 	NULL
384 };
385 
386 struct  qinit passthru_winit = {
387 	(int (*)()) pass_wput,
388 	NULL,
389 	NULL,
390 	NULL,
391 	NULL,
392 	&passthru_info,
393 	NULL
394 };
395 
396 /*
397  * Verify correctness of list head/tail pointers.
398  */
399 #define	LISTCHECK(head, tail, link) {				\
400 	EQUIV(head, tail);					\
401 	IMPLY(tail != NULL, tail->link == NULL);		\
402 }
403 
404 /*
405  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
406  * using a `link' field.
407  */
408 #define	ENQUEUE(el, head, tail, link) {				\
409 	ASSERT(el->link == NULL);				\
410 	LISTCHECK(head, tail, link);				\
411 	if (head == NULL)					\
412 		head = el;					\
413 	else							\
414 		tail->link = el;				\
415 	tail = el;						\
416 }
417 
418 /*
419  * Dequeue the first element of the list denoted by `head' and `tail' pointers
420  * using a `link' field and put result into `el'.
421  */
422 #define	DQ(el, head, tail, link) {				\
423 	LISTCHECK(head, tail, link);				\
424 	el = head;						\
425 	if (head != NULL) {					\
426 		head = head->link;				\
427 		if (head == NULL)				\
428 			tail = NULL;				\
429 		el->link = NULL;				\
430 	}							\
431 }
432 
433 /*
434  * Remove `el' from the list using `chase' and `curr' pointers and return result
435  * in `succeed'.
436  */
437 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
438 	LISTCHECK(head, tail, link);				\
439 	chase = NULL;						\
440 	succeed = 0;						\
441 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
442 		chase = curr;					\
443 	if (curr != NULL) {					\
444 		succeed = 1;					\
445 		ASSERT(curr == el);				\
446 		if (chase != NULL)				\
447 			chase->link = curr->link;		\
448 		else						\
449 			head = curr->link;			\
450 		curr->link = NULL;				\
451 		if (curr == tail)				\
452 			tail = chase;				\
453 	}							\
454 	LISTCHECK(head, tail, link);				\
455 }
456 
457 /* Handling of delayed messages on the inner syncq. */
458 
459 /*
460  * DEBUG versions should use function versions (to simplify tracing) and
461  * non-DEBUG kernels should use macro versions.
462  */
463 
464 /*
465  * Put a queue on the syncq list of queues.
466  * Assumes SQLOCK held.
467  */
468 #define	SQPUT_Q(sq, qp)							\
469 {									\
470 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
471 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
472 		/* The queue should not be linked anywhere */		\
473 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
474 		/* Head and tail may only be NULL simultaneously */	\
475 		EQUIV(sq->sq_head, sq->sq_tail);			\
476 		/* Queue may be only enqueued on its syncq */		\
477 		ASSERT(sq == qp->q_syncq);				\
478 		/* Check the correctness of SQ_MESSAGES flag */		\
479 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
480 		/* Sanity check first/last elements of the list */	\
481 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
482 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
483 		/*							\
484 		 * Sanity check of priority field: empty queue should	\
485 		 * have zero priority					\
486 		 * and nqueues equal to zero.				\
487 		 */							\
488 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
489 		/* Sanity check of sq_nqueues field */			\
490 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
491 		if (sq->sq_head == NULL) {				\
492 			sq->sq_head = sq->sq_tail = qp;			\
493 			sq->sq_flags |= SQ_MESSAGES;			\
494 		} else if (qp->q_spri == 0) {				\
495 			qp->q_sqprev = sq->sq_tail;			\
496 			sq->sq_tail->q_sqnext = qp;			\
497 			sq->sq_tail = qp;				\
498 		} else {						\
499 			/*						\
500 			 * Put this queue in priority order: higher	\
501 			 * priority gets closer to the head.		\
502 			 */						\
503 			queue_t **qpp = &sq->sq_tail;			\
504 			queue_t *qnext = NULL;				\
505 									\
506 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
507 				qnext = *qpp;				\
508 				qpp = &(*qpp)->q_sqprev;		\
509 			}						\
510 			qp->q_sqnext = qnext;				\
511 			qp->q_sqprev = *qpp;				\
512 			if (*qpp != NULL) {				\
513 				(*qpp)->q_sqnext = qp;			\
514 			} else {					\
515 				sq->sq_head = qp;			\
516 				sq->sq_pri = sq->sq_head->q_spri;	\
517 			}						\
518 			*qpp = qp;					\
519 		}							\
520 		qp->q_sqflags |= Q_SQQUEUED;				\
521 		qp->q_sqtstamp = ddi_get_lbolt();			\
522 		sq->sq_nqueues++;					\
523 	}								\
524 }
525 
526 /*
527  * Remove a queue from the syncq list
528  * Assumes SQLOCK held.
529  */
530 #define	SQRM_Q(sq, qp)							\
531 	{								\
532 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
533 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
534 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
535 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
536 		/* Check that the queue is actually in the list */	\
537 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
538 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
539 		ASSERT(sq->sq_nqueues != 0);				\
540 		if (qp->q_sqprev == NULL) {				\
541 			/* First queue on list, make head q_sqnext */	\
542 			sq->sq_head = qp->q_sqnext;			\
543 		} else {						\
544 			/* Make prev->next == next */			\
545 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
546 		}							\
547 		if (qp->q_sqnext == NULL) {				\
548 			/* Last queue on list, make tail sqprev */	\
549 			sq->sq_tail = qp->q_sqprev;			\
550 		} else {						\
551 			/* Make next->prev == prev */			\
552 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
553 		}							\
554 		/* clear out references on this queue */		\
555 		qp->q_sqprev = qp->q_sqnext = NULL;			\
556 		qp->q_sqflags &= ~Q_SQQUEUED;				\
557 		/* If there is nothing queued, clear SQ_MESSAGES */	\
558 		if (sq->sq_head != NULL) {				\
559 			sq->sq_pri = sq->sq_head->q_spri;		\
560 		} else	{						\
561 			sq->sq_flags &= ~SQ_MESSAGES;			\
562 			sq->sq_pri = 0;					\
563 		}							\
564 		sq->sq_nqueues--;					\
565 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
566 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
567 	}
568 
569 /* Hide the definition from the header file. */
570 #ifdef SQPUT_MP
571 #undef SQPUT_MP
572 #endif
573 
574 /*
575  * Put a message on the queue syncq.
576  * Assumes QLOCK held.
577  */
578 #define	SQPUT_MP(qp, mp)						\
579 	{								\
580 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
581 		ASSERT(qp->q_sqhead == NULL ||				\
582 		    (qp->q_sqtail != NULL &&				\
583 		    qp->q_sqtail->b_next == NULL));			\
584 		qp->q_syncqmsgs++;					\
585 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
586 		if (qp->q_sqhead == NULL) {				\
587 			qp->q_sqhead = qp->q_sqtail = mp;		\
588 		} else {						\
589 			qp->q_sqtail->b_next = mp;			\
590 			qp->q_sqtail = mp;				\
591 		}							\
592 		ASSERT(qp->q_syncqmsgs > 0);				\
593 		set_qfull(qp);						\
594 	}
595 
596 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
597 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
598 		if ((sq)->sq_ciputctrl != NULL) {			\
599 			int i;						\
600 			int nlocks = (sq)->sq_nciputctrl;		\
601 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
602 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
603 			for (i = 0; i <= nlocks; i++) {			\
604 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
605 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
606 			}						\
607 		}							\
608 	}
609 
610 
611 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
612 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
613 		if ((sq)->sq_ciputctrl != NULL) {			\
614 			int i;						\
615 			int nlocks = (sq)->sq_nciputctrl;		\
616 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
617 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
618 			for (i = 0; i <= nlocks; i++) {			\
619 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
620 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
621 			}						\
622 		}							\
623 	}
624 
625 /*
626  * Run service procedures for all queues in the stream head.
627  */
628 #define	STR_SERVICE(stp, q) {						\
629 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
630 	while (stp->sd_qhead != NULL) {					\
631 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
632 		ASSERT(stp->sd_nqueues > 0);				\
633 		stp->sd_nqueues--;					\
634 		ASSERT(!(q->q_flag & QINSERVICE));			\
635 		mutex_exit(&stp->sd_qlock);				\
636 		queue_service(q);					\
637 		mutex_enter(&stp->sd_qlock);				\
638 	}								\
639 	ASSERT(stp->sd_nqueues == 0);					\
640 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
641 }
642 
643 /*
644  * Constructor/destructor routines for the stream head cache
645  */
646 /* ARGSUSED */
647 static int
648 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
649 {
650 	stdata_t *stp = buf;
651 
652 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
653 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
654 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
655 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
656 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
657 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
658 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
659 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
660 	stp->sd_wrq = NULL;
661 
662 	return (0);
663 }
664 
665 /* ARGSUSED */
666 static void
667 stream_head_destructor(void *buf, void *cdrarg)
668 {
669 	stdata_t *stp = buf;
670 
671 	mutex_destroy(&stp->sd_lock);
672 	mutex_destroy(&stp->sd_reflock);
673 	mutex_destroy(&stp->sd_qlock);
674 	cv_destroy(&stp->sd_monitor);
675 	cv_destroy(&stp->sd_iocmonitor);
676 	cv_destroy(&stp->sd_refmonitor);
677 	cv_destroy(&stp->sd_qcv);
678 	cv_destroy(&stp->sd_zcopy_wait);
679 }
680 
681 /*
682  * Constructor/destructor routines for the queue cache
683  */
684 /* ARGSUSED */
685 static int
686 queue_constructor(void *buf, void *cdrarg, int kmflags)
687 {
688 	queinfo_t *qip = buf;
689 	queue_t *qp = &qip->qu_rqueue;
690 	queue_t *wqp = &qip->qu_wqueue;
691 	syncq_t	*sq = &qip->qu_syncq;
692 
693 	qp->q_first = NULL;
694 	qp->q_link = NULL;
695 	qp->q_count = 0;
696 	qp->q_mblkcnt = 0;
697 	qp->q_sqhead = NULL;
698 	qp->q_sqtail = NULL;
699 	qp->q_sqnext = NULL;
700 	qp->q_sqprev = NULL;
701 	qp->q_sqflags = 0;
702 	qp->q_rwcnt = 0;
703 	qp->q_spri = 0;
704 
705 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
706 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
707 
708 	wqp->q_first = NULL;
709 	wqp->q_link = NULL;
710 	wqp->q_count = 0;
711 	wqp->q_mblkcnt = 0;
712 	wqp->q_sqhead = NULL;
713 	wqp->q_sqtail = NULL;
714 	wqp->q_sqnext = NULL;
715 	wqp->q_sqprev = NULL;
716 	wqp->q_sqflags = 0;
717 	wqp->q_rwcnt = 0;
718 	wqp->q_spri = 0;
719 
720 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
721 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
722 
723 	sq->sq_head = NULL;
724 	sq->sq_tail = NULL;
725 	sq->sq_evhead = NULL;
726 	sq->sq_evtail = NULL;
727 	sq->sq_callbpend = NULL;
728 	sq->sq_outer = NULL;
729 	sq->sq_onext = NULL;
730 	sq->sq_oprev = NULL;
731 	sq->sq_next = NULL;
732 	sq->sq_svcflags = 0;
733 	sq->sq_servcount = 0;
734 	sq->sq_needexcl = 0;
735 	sq->sq_nqueues = 0;
736 	sq->sq_pri = 0;
737 
738 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
739 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
740 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
741 
742 	return (0);
743 }
744 
745 /* ARGSUSED */
746 static void
747 queue_destructor(void *buf, void *cdrarg)
748 {
749 	queinfo_t *qip = buf;
750 	queue_t *qp = &qip->qu_rqueue;
751 	queue_t *wqp = &qip->qu_wqueue;
752 	syncq_t	*sq = &qip->qu_syncq;
753 
754 	ASSERT(qp->q_sqhead == NULL);
755 	ASSERT(wqp->q_sqhead == NULL);
756 	ASSERT(qp->q_sqnext == NULL);
757 	ASSERT(wqp->q_sqnext == NULL);
758 	ASSERT(qp->q_rwcnt == 0);
759 	ASSERT(wqp->q_rwcnt == 0);
760 
761 	mutex_destroy(&qp->q_lock);
762 	cv_destroy(&qp->q_wait);
763 
764 	mutex_destroy(&wqp->q_lock);
765 	cv_destroy(&wqp->q_wait);
766 
767 	mutex_destroy(&sq->sq_lock);
768 	cv_destroy(&sq->sq_wait);
769 	cv_destroy(&sq->sq_exitwait);
770 }
771 
772 /*
773  * Constructor/destructor routines for the syncq cache
774  */
775 /* ARGSUSED */
776 static int
777 syncq_constructor(void *buf, void *cdrarg, int kmflags)
778 {
779 	syncq_t	*sq = buf;
780 
781 	bzero(buf, sizeof (syncq_t));
782 
783 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
784 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
785 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
786 
787 	return (0);
788 }
789 
790 /* ARGSUSED */
791 static void
792 syncq_destructor(void *buf, void *cdrarg)
793 {
794 	syncq_t	*sq = buf;
795 
796 	ASSERT(sq->sq_head == NULL);
797 	ASSERT(sq->sq_tail == NULL);
798 	ASSERT(sq->sq_evhead == NULL);
799 	ASSERT(sq->sq_evtail == NULL);
800 	ASSERT(sq->sq_callbpend == NULL);
801 	ASSERT(sq->sq_callbflags == 0);
802 	ASSERT(sq->sq_outer == NULL);
803 	ASSERT(sq->sq_onext == NULL);
804 	ASSERT(sq->sq_oprev == NULL);
805 	ASSERT(sq->sq_next == NULL);
806 	ASSERT(sq->sq_needexcl == 0);
807 	ASSERT(sq->sq_svcflags == 0);
808 	ASSERT(sq->sq_servcount == 0);
809 	ASSERT(sq->sq_nqueues == 0);
810 	ASSERT(sq->sq_pri == 0);
811 	ASSERT(sq->sq_count == 0);
812 	ASSERT(sq->sq_rmqcount == 0);
813 	ASSERT(sq->sq_cancelid == 0);
814 	ASSERT(sq->sq_ciputctrl == NULL);
815 	ASSERT(sq->sq_nciputctrl == 0);
816 	ASSERT(sq->sq_type == 0);
817 	ASSERT(sq->sq_flags == 0);
818 
819 	mutex_destroy(&sq->sq_lock);
820 	cv_destroy(&sq->sq_wait);
821 	cv_destroy(&sq->sq_exitwait);
822 }
823 
824 /* ARGSUSED */
825 static int
826 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
827 {
828 	ciputctrl_t *cip = buf;
829 	int i;
830 
831 	for (i = 0; i < n_ciputctrl; i++) {
832 		cip[i].ciputctrl_count = SQ_FASTPUT;
833 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
834 	}
835 
836 	return (0);
837 }
838 
839 /* ARGSUSED */
840 static void
841 ciputctrl_destructor(void *buf, void *cdrarg)
842 {
843 	ciputctrl_t *cip = buf;
844 	int i;
845 
846 	for (i = 0; i < n_ciputctrl; i++) {
847 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
848 		mutex_destroy(&cip[i].ciputctrl_lock);
849 	}
850 }
851 
852 /*
853  * Init routine run from main at boot time.
854  */
855 void
856 strinit(void)
857 {
858 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
859 
860 	stream_head_cache = kmem_cache_create("stream_head_cache",
861 	    sizeof (stdata_t), 0,
862 	    stream_head_constructor, stream_head_destructor, NULL,
863 	    NULL, NULL, 0);
864 
865 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
866 	    queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
867 
868 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
869 	    syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
870 
871 	qband_cache = kmem_cache_create("qband_cache",
872 	    sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
873 
874 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
875 	    sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
876 
877 	n_ciputctrl = ncpus;
878 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
879 	ASSERT(n_ciputctrl >= 1);
880 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
881 	if (n_ciputctrl >= min_n_ciputctrl) {
882 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
883 		    sizeof (ciputctrl_t) * n_ciputctrl,
884 		    sizeof (ciputctrl_t), ciputctrl_constructor,
885 		    ciputctrl_destructor, NULL, NULL, NULL, 0);
886 	}
887 
888 	streams_taskq = system_taskq;
889 
890 	if (streams_taskq == NULL)
891 		panic("strinit: no memory for streams taskq!");
892 
893 	bc_bkgrnd_thread = thread_create(NULL, 0,
894 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
895 
896 	streams_qbkgrnd_thread = thread_create(NULL, 0,
897 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
898 
899 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
900 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
901 
902 	/*
903 	 * Create STREAMS kstats.
904 	 */
905 	str_kstat = kstat_create("streams", 0, "strstat",
906 	    "net", KSTAT_TYPE_NAMED,
907 	    sizeof (str_statistics) / sizeof (kstat_named_t),
908 	    KSTAT_FLAG_VIRTUAL);
909 
910 	if (str_kstat != NULL) {
911 		str_kstat->ks_data = &str_statistics;
912 		kstat_install(str_kstat);
913 	}
914 
915 	/*
916 	 * TPI support routine initialisation.
917 	 */
918 	tpi_init();
919 
920 	/*
921 	 * Handle to have autopush and persistent link information per
922 	 * zone.
923 	 * Note: uses shutdown hook instead of destroy hook so that the
924 	 * persistent links can be torn down before the destroy hooks
925 	 * in the TCP/IP stack are called.
926 	 */
927 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
928 	    str_stack_fini);
929 }
930 
931 void
932 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
933 {
934 	struct stdata *stp;
935 
936 	ASSERT(vp->v_stream);
937 	stp = vp->v_stream;
938 	/* Have to hold sd_lock to prevent siglist from changing */
939 	mutex_enter(&stp->sd_lock);
940 	if (stp->sd_sigflags & event)
941 		strsendsig(stp->sd_siglist, event, band, error);
942 	mutex_exit(&stp->sd_lock);
943 }
944 
945 /*
946  * Send the "sevent" set of signals to a process.
947  * This might send more than one signal if the process is registered
948  * for multiple events. The caller should pass in an sevent that only
949  * includes the events for which the process has registered.
950  */
951 static void
952 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
953 	uchar_t band, int error)
954 {
955 	ASSERT(MUTEX_HELD(&proc->p_lock));
956 
957 	info->si_band = 0;
958 	info->si_errno = 0;
959 
960 	if (sevent & S_ERROR) {
961 		sevent &= ~S_ERROR;
962 		info->si_code = POLL_ERR;
963 		info->si_errno = error;
964 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
965 		    "strsendsig:proc %p info %p", proc, info);
966 		sigaddq(proc, NULL, info, KM_NOSLEEP);
967 		info->si_errno = 0;
968 	}
969 	if (sevent & S_HANGUP) {
970 		sevent &= ~S_HANGUP;
971 		info->si_code = POLL_HUP;
972 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
973 		    "strsendsig:proc %p info %p", proc, info);
974 		sigaddq(proc, NULL, info, KM_NOSLEEP);
975 	}
976 	if (sevent & S_HIPRI) {
977 		sevent &= ~S_HIPRI;
978 		info->si_code = POLL_PRI;
979 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
980 		    "strsendsig:proc %p info %p", proc, info);
981 		sigaddq(proc, NULL, info, KM_NOSLEEP);
982 	}
983 	if (sevent & S_RDBAND) {
984 		sevent &= ~S_RDBAND;
985 		if (events & S_BANDURG)
986 			sigtoproc(proc, NULL, SIGURG);
987 		else
988 			sigtoproc(proc, NULL, SIGPOLL);
989 	}
990 	if (sevent & S_WRBAND) {
991 		sevent &= ~S_WRBAND;
992 		sigtoproc(proc, NULL, SIGPOLL);
993 	}
994 	if (sevent & S_INPUT) {
995 		sevent &= ~S_INPUT;
996 		info->si_code = POLL_IN;
997 		info->si_band = band;
998 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
999 		    "strsendsig:proc %p info %p", proc, info);
1000 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1001 		info->si_band = 0;
1002 	}
1003 	if (sevent & S_OUTPUT) {
1004 		sevent &= ~S_OUTPUT;
1005 		info->si_code = POLL_OUT;
1006 		info->si_band = band;
1007 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1008 		    "strsendsig:proc %p info %p", proc, info);
1009 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1010 		info->si_band = 0;
1011 	}
1012 	if (sevent & S_MSG) {
1013 		sevent &= ~S_MSG;
1014 		info->si_code = POLL_MSG;
1015 		info->si_band = band;
1016 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1017 		    "strsendsig:proc %p info %p", proc, info);
1018 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1019 		info->si_band = 0;
1020 	}
1021 	if (sevent & S_RDNORM) {
1022 		sevent &= ~S_RDNORM;
1023 		sigtoproc(proc, NULL, SIGPOLL);
1024 	}
1025 	if (sevent != 0) {
1026 		panic("strsendsig: unknown event(s) %x", sevent);
1027 	}
1028 }
1029 
1030 /*
1031  * Send SIGPOLL/SIGURG signal to all processes and process groups
1032  * registered on the given signal list that want a signal for at
1033  * least one of the specified events.
1034  *
1035  * Must be called with exclusive access to siglist (caller holding sd_lock).
1036  *
1037  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1038  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1039  * while it is in the siglist.
1040  *
1041  * For performance reasons (MP scalability) the code drops pidlock
1042  * when sending signals to a single process.
1043  * When sending to a process group the code holds
1044  * pidlock to prevent the membership in the process group from changing
1045  * while walking the p_pglink list.
1046  */
1047 void
1048 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1049 {
1050 	strsig_t *ssp;
1051 	k_siginfo_t info;
1052 	struct pid *pidp;
1053 	proc_t  *proc;
1054 
1055 	info.si_signo = SIGPOLL;
1056 	info.si_errno = 0;
1057 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1058 		int sevent;
1059 
1060 		sevent = ssp->ss_events & event;
1061 		if (sevent == 0)
1062 			continue;
1063 
1064 		if ((pidp = ssp->ss_pidp) == NULL) {
1065 			/* pid was released but still on event list */
1066 			continue;
1067 		}
1068 
1069 
1070 		if (ssp->ss_pid > 0) {
1071 			/*
1072 			 * XXX This unfortunately still generates
1073 			 * a signal when a fd is closed but
1074 			 * the proc is active.
1075 			 */
1076 			ASSERT(ssp->ss_pid == pidp->pid_id);
1077 
1078 			mutex_enter(&pidlock);
1079 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1080 			if (proc == NULL) {
1081 				mutex_exit(&pidlock);
1082 				continue;
1083 			}
1084 			mutex_enter(&proc->p_lock);
1085 			mutex_exit(&pidlock);
1086 			dosendsig(proc, ssp->ss_events, sevent, &info,
1087 			    band, error);
1088 			mutex_exit(&proc->p_lock);
1089 		} else {
1090 			/*
1091 			 * Send to process group. Hold pidlock across
1092 			 * calls to dosendsig().
1093 			 */
1094 			pid_t pgrp = -ssp->ss_pid;
1095 
1096 			mutex_enter(&pidlock);
1097 			proc = pgfind_zone(pgrp, ALL_ZONES);
1098 			while (proc != NULL) {
1099 				mutex_enter(&proc->p_lock);
1100 				dosendsig(proc, ssp->ss_events, sevent,
1101 				    &info, band, error);
1102 				mutex_exit(&proc->p_lock);
1103 				proc = proc->p_pglink;
1104 			}
1105 			mutex_exit(&pidlock);
1106 		}
1107 	}
1108 }
1109 
1110 /*
1111  * Attach a stream device or module.
1112  * qp is a read queue; the new queue goes in so its next
1113  * read ptr is the argument, and the write queue corresponding
1114  * to the argument points to this queue. Return 0 on success,
1115  * or a non-zero errno on failure.
1116  */
1117 int
1118 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1119     boolean_t is_insert)
1120 {
1121 	major_t			major;
1122 	cdevsw_impl_t		*dp;
1123 	struct streamtab	*str;
1124 	queue_t			*rq;
1125 	queue_t			*wrq;
1126 	uint32_t		qflag;
1127 	uint32_t		sqtype;
1128 	perdm_t			*dmp;
1129 	int			error;
1130 	int			sflag;
1131 
1132 	rq = allocq();
1133 	wrq = _WR(rq);
1134 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1135 
1136 	if (fp != NULL) {
1137 		str = fp->f_str;
1138 		qflag = fp->f_qflag;
1139 		sqtype = fp->f_sqtype;
1140 		dmp = fp->f_dmp;
1141 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1142 		sflag = MODOPEN;
1143 
1144 		/*
1145 		 * stash away a pointer to the module structure so we can
1146 		 * unref it in qdetach.
1147 		 */
1148 		rq->q_fp = fp;
1149 	} else {
1150 		ASSERT(!is_insert);
1151 
1152 		major = getmajor(*devp);
1153 		dp = &devimpl[major];
1154 
1155 		str = dp->d_str;
1156 		ASSERT(str == STREAMSTAB(major));
1157 
1158 		qflag = dp->d_qflag;
1159 		ASSERT(qflag & QISDRV);
1160 		sqtype = dp->d_sqtype;
1161 
1162 		/* create perdm_t if needed */
1163 		if (NEED_DM(dp->d_dmp, qflag))
1164 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1165 
1166 		dmp = dp->d_dmp;
1167 		sflag = 0;
1168 	}
1169 
1170 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1171 	    "qattach:qflag == %X(%X)", qflag, *devp);
1172 
1173 	/* setq might sleep in allocator - avoid holding locks. */
1174 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1175 
1176 	/*
1177 	 * Before calling the module's open routine, set up the q_next
1178 	 * pointer for inserting a module in the middle of a stream.
1179 	 *
1180 	 * Note that we can always set _QINSERTING and set up q_next
1181 	 * pointer for both inserting and pushing a module.  Then there
1182 	 * is no need for the is_insert parameter.  In insertq(), called
1183 	 * by qprocson(), assume that q_next of the new module always points
1184 	 * to the correct queue and use it for insertion.  Everything should
1185 	 * work out fine.  But in the first release of _I_INSERT, we
1186 	 * distinguish between inserting and pushing to make sure that
1187 	 * pushing a module follows the same code path as before.
1188 	 */
1189 	if (is_insert) {
1190 		rq->q_flag |= _QINSERTING;
1191 		rq->q_next = qp;
1192 	}
1193 
1194 	/*
1195 	 * If there is an outer perimeter get exclusive access during
1196 	 * the open procedure.  Bump up the reference count on the queue.
1197 	 */
1198 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1199 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1200 	if (error != 0)
1201 		goto failed;
1202 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1203 	ASSERT(qprocsareon(rq));
1204 	return (0);
1205 
1206 failed:
1207 	rq->q_flag &= ~_QINSERTING;
1208 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1209 		qprocsoff(rq);
1210 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1211 	rq->q_next = wrq->q_next = NULL;
1212 	qdetach(rq, 0, 0, crp, B_FALSE);
1213 	return (error);
1214 }
1215 
1216 /*
1217  * Handle second open of stream. For modules, set the
1218  * last argument to MODOPEN and do not pass any open flags.
1219  * Ignore dummydev since this is not the first open.
1220  */
1221 int
1222 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1223 {
1224 	int	error;
1225 	dev_t dummydev;
1226 	queue_t *wqp = _WR(qp);
1227 
1228 	ASSERT(qp->q_flag & QREADR);
1229 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1230 
1231 	dummydev = *devp;
1232 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1233 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1234 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1235 		mutex_enter(&STREAM(qp)->sd_lock);
1236 		qp->q_stream->sd_flag |= STREOPENFAIL;
1237 		mutex_exit(&STREAM(qp)->sd_lock);
1238 		return (error);
1239 	}
1240 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1241 
1242 	/*
1243 	 * successful open should have done qprocson()
1244 	 */
1245 	ASSERT(qprocsareon(_RD(qp)));
1246 	return (0);
1247 }
1248 
1249 /*
1250  * Detach a stream module or device.
1251  * If clmode == 1 then the module or driver was opened and its
1252  * close routine must be called. If clmode == 0, the module
1253  * or driver was never opened or the open failed, and so its close
1254  * should not be called.
1255  */
1256 void
1257 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1258 {
1259 	queue_t *wqp = _WR(qp);
1260 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1261 
1262 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1263 		stream_runservice(STREAM(qp));
1264 
1265 	if (clmode) {
1266 		/*
1267 		 * Make sure that all the messages on the write side syncq are
1268 		 * processed and nothing is left. Since we are closing, no new
1269 		 * messages may appear there.
1270 		 */
1271 		wait_q_syncq(wqp);
1272 
1273 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1274 		if (is_remove) {
1275 			mutex_enter(QLOCK(qp));
1276 			qp->q_flag |= _QREMOVING;
1277 			mutex_exit(QLOCK(qp));
1278 		}
1279 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1280 		/*
1281 		 * Check that qprocsoff() was actually called.
1282 		 */
1283 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1284 
1285 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1286 	} else {
1287 		disable_svc(qp);
1288 	}
1289 
1290 	/*
1291 	 * Allow any threads blocked in entersq to proceed and discover
1292 	 * the QWCLOSE is set.
1293 	 * Note: This assumes that all users of entersq check QWCLOSE.
1294 	 * Currently runservice is the only entersq that can happen
1295 	 * after removeq has finished.
1296 	 * Removeq will have discarded all messages destined to the closing
1297 	 * pair of queues from the syncq.
1298 	 * NOTE: Calling a function inside an assert is unconventional.
1299 	 * However, it does not cause any problem since flush_syncq() does
1300 	 * not change any state except when it returns non-zero i.e.
1301 	 * when the assert will trigger.
1302 	 */
1303 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1304 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1305 	ASSERT((qp->q_flag & QPERMOD) ||
1306 	    ((qp->q_syncq->sq_head == NULL) &&
1307 	    (wqp->q_syncq->sq_head == NULL)));
1308 
1309 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1310 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1311 	if (qp->q_fp != NULL)
1312 		fmodsw_rele(qp->q_fp);
1313 
1314 	/* freeq removes us from the outer perimeter if any */
1315 	freeq(qp);
1316 }
1317 
1318 /* Prevent service procedures from being called */
1319 void
1320 disable_svc(queue_t *qp)
1321 {
1322 	queue_t *wqp = _WR(qp);
1323 
1324 	ASSERT(qp->q_flag & QREADR);
1325 	mutex_enter(QLOCK(qp));
1326 	qp->q_flag |= QWCLOSE;
1327 	mutex_exit(QLOCK(qp));
1328 	mutex_enter(QLOCK(wqp));
1329 	wqp->q_flag |= QWCLOSE;
1330 	mutex_exit(QLOCK(wqp));
1331 }
1332 
1333 /* Allow service procedures to be called again */
1334 void
1335 enable_svc(queue_t *qp)
1336 {
1337 	queue_t *wqp = _WR(qp);
1338 
1339 	ASSERT(qp->q_flag & QREADR);
1340 	mutex_enter(QLOCK(qp));
1341 	qp->q_flag &= ~QWCLOSE;
1342 	mutex_exit(QLOCK(qp));
1343 	mutex_enter(QLOCK(wqp));
1344 	wqp->q_flag &= ~QWCLOSE;
1345 	mutex_exit(QLOCK(wqp));
1346 }
1347 
1348 /*
1349  * Remove queue from qhead/qtail if it is enabled.
1350  * Only reset QENAB if the queue was removed from the runlist.
1351  * A queue goes through 3 stages:
1352  *	It is on the service list and QENAB is set.
1353  *	It is removed from the service list but QENAB is still set.
1354  *	QENAB gets changed to QINSERVICE.
1355  *	QINSERVICE is reset (when the service procedure is done)
1356  * Thus we can not reset QENAB unless we actually removed it from the service
1357  * queue.
1358  */
1359 void
1360 remove_runlist(queue_t *qp)
1361 {
1362 	if (qp->q_flag & QENAB && qhead != NULL) {
1363 		queue_t *q_chase;
1364 		queue_t *q_curr;
1365 		int removed;
1366 
1367 		mutex_enter(&service_queue);
1368 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1369 		mutex_exit(&service_queue);
1370 		if (removed) {
1371 			STRSTAT(qremoved);
1372 			qp->q_flag &= ~QENAB;
1373 		}
1374 	}
1375 }
1376 
1377 
1378 /*
1379  * Wait for any pending service processing to complete.
1380  * The removal of queues from the runlist is not atomic with the
1381  * clearing of the QENABLED flag and setting the INSERVICE flag.
1382  * consequently it is possible for remove_runlist in strclose
1383  * to not find the queue on the runlist but for it to be QENABLED
1384  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1385  * as well as INSERVICE.
1386  */
1387 void
1388 wait_svc(queue_t *qp)
1389 {
1390 	queue_t *wqp = _WR(qp);
1391 
1392 	ASSERT(qp->q_flag & QREADR);
1393 
1394 	/*
1395 	 * Try to remove queues from qhead/qtail list.
1396 	 */
1397 	if (qhead != NULL) {
1398 		remove_runlist(qp);
1399 		remove_runlist(wqp);
1400 	}
1401 	/*
1402 	 * Wait till the syncqs associated with the queue disappear from the
1403 	 * background processing list.
1404 	 * This only needs to be done for non-PERMOD perimeters since
1405 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1406 	 * when the last module/driver is unloaded.
1407 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1408 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1409 	 * functions remove the queue from its syncq list, so sqthread will not
1410 	 * try to access the queue.
1411 	 */
1412 	if (!(qp->q_flag & QPERMOD)) {
1413 		syncq_t *rsq = qp->q_syncq;
1414 		syncq_t *wsq = wqp->q_syncq;
1415 
1416 		/*
1417 		 * Disable rsq and wsq and wait for any background processing of
1418 		 * syncq to complete.
1419 		 */
1420 		wait_sq_svc(rsq);
1421 		if (wsq != rsq)
1422 			wait_sq_svc(wsq);
1423 	}
1424 
1425 	mutex_enter(QLOCK(qp));
1426 	while (qp->q_flag & (QINSERVICE|QENAB))
1427 		cv_wait(&qp->q_wait, QLOCK(qp));
1428 	mutex_exit(QLOCK(qp));
1429 	mutex_enter(QLOCK(wqp));
1430 	while (wqp->q_flag & (QINSERVICE|QENAB))
1431 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1432 	mutex_exit(QLOCK(wqp));
1433 }
1434 
1435 /*
1436  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1437  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1438  * also be set, and is passed through to allocb_cred_wait().
1439  *
1440  * Returns errno on failure, zero on success.
1441  */
1442 int
1443 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1444 {
1445 	mblk_t *tmp;
1446 	ssize_t  count;
1447 	int error = 0;
1448 
1449 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1450 	    (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1451 
1452 	if (bp->b_datap->db_type == M_IOCTL) {
1453 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1454 	} else {
1455 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1456 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1457 	}
1458 	/*
1459 	 * strdoioctl validates ioc_count, so if this assert fails it
1460 	 * cannot be due to user error.
1461 	 */
1462 	ASSERT(count >= 0);
1463 
1464 	if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1465 	    curproc->p_pid)) == NULL) {
1466 		return (error);
1467 	}
1468 	error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1469 	if (error != 0) {
1470 		freeb(tmp);
1471 		return (error);
1472 	}
1473 	DB_CPID(tmp) = curproc->p_pid;
1474 	tmp->b_wptr += count;
1475 	bp->b_cont = tmp;
1476 
1477 	return (0);
1478 }
1479 
1480 /*
1481  * Copy ioctl data to user-land. Return non-zero errno on failure,
1482  * 0 for success.
1483  */
1484 int
1485 getiocd(mblk_t *bp, char *arg, int copymode)
1486 {
1487 	ssize_t count;
1488 	size_t  n;
1489 	int	error;
1490 
1491 	if (bp->b_datap->db_type == M_IOCACK)
1492 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1493 	else {
1494 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1495 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1496 	}
1497 	ASSERT(count >= 0);
1498 
1499 	for (bp = bp->b_cont; bp && count;
1500 	    count -= n, bp = bp->b_cont, arg += n) {
1501 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1502 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1503 		if (error)
1504 			return (error);
1505 	}
1506 	ASSERT(count == 0);
1507 	return (0);
1508 }
1509 
1510 /*
1511  * Allocate a linkinfo entry given the write queue of the
1512  * bottom module of the top stream and the write queue of the
1513  * stream head of the bottom stream.
1514  */
1515 linkinfo_t *
1516 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1517 {
1518 	linkinfo_t *linkp;
1519 
1520 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1521 
1522 	linkp->li_lblk.l_qtop = qup;
1523 	linkp->li_lblk.l_qbot = qdown;
1524 	linkp->li_fpdown = fpdown;
1525 
1526 	mutex_enter(&strresources);
1527 	linkp->li_next = linkinfo_list;
1528 	linkp->li_prev = NULL;
1529 	if (linkp->li_next)
1530 		linkp->li_next->li_prev = linkp;
1531 	linkinfo_list = linkp;
1532 	linkp->li_lblk.l_index = ++lnk_id;
1533 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1534 	mutex_exit(&strresources);
1535 
1536 	return (linkp);
1537 }
1538 
1539 /*
1540  * Free a linkinfo entry.
1541  */
1542 void
1543 lbfree(linkinfo_t *linkp)
1544 {
1545 	mutex_enter(&strresources);
1546 	if (linkp->li_next)
1547 		linkp->li_next->li_prev = linkp->li_prev;
1548 	if (linkp->li_prev)
1549 		linkp->li_prev->li_next = linkp->li_next;
1550 	else
1551 		linkinfo_list = linkp->li_next;
1552 	mutex_exit(&strresources);
1553 
1554 	kmem_cache_free(linkinfo_cache, linkp);
1555 }
1556 
1557 /*
1558  * Check for a potential linking cycle.
1559  * Return 1 if a link will result in a cycle,
1560  * and 0 otherwise.
1561  */
1562 int
1563 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1564 {
1565 	struct mux_node *np;
1566 	struct mux_edge *ep;
1567 	int i;
1568 	major_t lomaj;
1569 	major_t upmaj;
1570 	/*
1571 	 * if the lower stream is a pipe/FIFO, return, since link
1572 	 * cycles can not happen on pipes/FIFOs
1573 	 */
1574 	if (lostp->sd_vnode->v_type == VFIFO)
1575 		return (0);
1576 
1577 	for (i = 0; i < ss->ss_devcnt; i++) {
1578 		np = &ss->ss_mux_nodes[i];
1579 		MUX_CLEAR(np);
1580 	}
1581 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1582 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1583 	np = &ss->ss_mux_nodes[lomaj];
1584 	for (;;) {
1585 		if (!MUX_DIDVISIT(np)) {
1586 			if (np->mn_imaj == upmaj)
1587 				return (1);
1588 			if (np->mn_outp == NULL) {
1589 				MUX_VISIT(np);
1590 				if (np->mn_originp == NULL)
1591 					return (0);
1592 				np = np->mn_originp;
1593 				continue;
1594 			}
1595 			MUX_VISIT(np);
1596 			np->mn_startp = np->mn_outp;
1597 		} else {
1598 			if (np->mn_startp == NULL) {
1599 				if (np->mn_originp == NULL)
1600 					return (0);
1601 				else {
1602 					np = np->mn_originp;
1603 					continue;
1604 				}
1605 			}
1606 			/*
1607 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1608 			 * ignore the edge and move on. ep->me_nodep gets
1609 			 * set to NULL in mux_addedge() if it is a FIFO.
1610 			 *
1611 			 */
1612 			ep = np->mn_startp;
1613 			np->mn_startp = ep->me_nextp;
1614 			if (ep->me_nodep == NULL)
1615 				continue;
1616 			ep->me_nodep->mn_originp = np;
1617 			np = ep->me_nodep;
1618 		}
1619 	}
1620 }
1621 
1622 /*
1623  * Find linkinfo entry corresponding to the parameters.
1624  */
1625 linkinfo_t *
1626 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1627 {
1628 	linkinfo_t *linkp;
1629 	struct mux_edge *mep;
1630 	struct mux_node *mnp;
1631 	queue_t *qup;
1632 
1633 	mutex_enter(&strresources);
1634 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1635 		qup = getendq(stp->sd_wrq);
1636 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1637 			if ((qup == linkp->li_lblk.l_qtop) &&
1638 			    (!index || (index == linkp->li_lblk.l_index))) {
1639 				mutex_exit(&strresources);
1640 				return (linkp);
1641 			}
1642 		}
1643 	} else {
1644 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1645 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1646 		mep = mnp->mn_outp;
1647 		while (mep) {
1648 			if ((index == 0) || (index == mep->me_muxid))
1649 				break;
1650 			mep = mep->me_nextp;
1651 		}
1652 		if (!mep) {
1653 			mutex_exit(&strresources);
1654 			return (NULL);
1655 		}
1656 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1657 			if ((!linkp->li_lblk.l_qtop) &&
1658 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1659 				mutex_exit(&strresources);
1660 				return (linkp);
1661 			}
1662 		}
1663 	}
1664 	mutex_exit(&strresources);
1665 	return (NULL);
1666 }
1667 
1668 /*
1669  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1670  * last queue on the chain and return it.
1671  */
1672 queue_t *
1673 getendq(queue_t *q)
1674 {
1675 	ASSERT(q != NULL);
1676 	while (_SAMESTR(q))
1677 		q = q->q_next;
1678 	return (q);
1679 }
1680 
1681 /*
1682  * Wait for the syncq count to drop to zero.
1683  * sq could be either outer or inner.
1684  */
1685 
1686 static void
1687 wait_syncq(syncq_t *sq)
1688 {
1689 	uint16_t count;
1690 
1691 	mutex_enter(SQLOCK(sq));
1692 	count = sq->sq_count;
1693 	SQ_PUTLOCKS_ENTER(sq);
1694 	SUM_SQ_PUTCOUNTS(sq, count);
1695 	while (count != 0) {
1696 		sq->sq_flags |= SQ_WANTWAKEUP;
1697 		SQ_PUTLOCKS_EXIT(sq);
1698 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1699 		count = sq->sq_count;
1700 		SQ_PUTLOCKS_ENTER(sq);
1701 		SUM_SQ_PUTCOUNTS(sq, count);
1702 	}
1703 	SQ_PUTLOCKS_EXIT(sq);
1704 	mutex_exit(SQLOCK(sq));
1705 }
1706 
1707 /*
1708  * Wait while there are any messages for the queue in its syncq.
1709  */
1710 static void
1711 wait_q_syncq(queue_t *q)
1712 {
1713 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1714 		syncq_t *sq = q->q_syncq;
1715 
1716 		mutex_enter(SQLOCK(sq));
1717 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1718 			sq->sq_flags |= SQ_WANTWAKEUP;
1719 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1720 		}
1721 		mutex_exit(SQLOCK(sq));
1722 	}
1723 }
1724 
1725 
1726 int
1727 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1728     int lhlink)
1729 {
1730 	struct stdata *stp;
1731 	struct strioctl strioc;
1732 	struct linkinfo *linkp;
1733 	struct stdata *stpdown;
1734 	struct streamtab *str;
1735 	queue_t *passq;
1736 	syncq_t *passyncq;
1737 	queue_t *rq;
1738 	cdevsw_impl_t *dp;
1739 	uint32_t qflag;
1740 	uint32_t sqtype;
1741 	perdm_t *dmp;
1742 	int error = 0;
1743 	netstack_t *ns;
1744 	str_stack_t *ss;
1745 
1746 	stp = vp->v_stream;
1747 	TRACE_1(TR_FAC_STREAMS_FR,
1748 	    TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1749 	/*
1750 	 * Test for invalid upper stream
1751 	 */
1752 	if (stp->sd_flag & STRHUP) {
1753 		return (ENXIO);
1754 	}
1755 	if (vp->v_type == VFIFO) {
1756 		return (EINVAL);
1757 	}
1758 	if (stp->sd_strtab == NULL) {
1759 		return (EINVAL);
1760 	}
1761 	if (!stp->sd_strtab->st_muxwinit) {
1762 		return (EINVAL);
1763 	}
1764 	if (fpdown == NULL) {
1765 		return (EBADF);
1766 	}
1767 	ns = netstack_find_by_cred(crp);
1768 	ASSERT(ns != NULL);
1769 	ss = ns->netstack_str;
1770 	ASSERT(ss != NULL);
1771 
1772 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1773 		netstack_rele(ss->ss_netstack);
1774 		return (EINVAL);
1775 	}
1776 	mutex_enter(&muxifier);
1777 	if (stp->sd_flag & STPLEX) {
1778 		mutex_exit(&muxifier);
1779 		netstack_rele(ss->ss_netstack);
1780 		return (ENXIO);
1781 	}
1782 
1783 	/*
1784 	 * Test for invalid lower stream.
1785 	 * The check for the v_type != VFIFO and having a major
1786 	 * number not >= devcnt is done to avoid problems with
1787 	 * adding mux_node entry past the end of mux_nodes[].
1788 	 * For FIFO's we don't add an entry so this isn't a
1789 	 * problem.
1790 	 */
1791 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1792 	    (stpdown == stp) || (stpdown->sd_flag &
1793 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1794 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1795 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1796 	    linkcycle(stp, stpdown, ss)) {
1797 		mutex_exit(&muxifier);
1798 		netstack_rele(ss->ss_netstack);
1799 		return (EINVAL);
1800 	}
1801 	TRACE_1(TR_FAC_STREAMS_FR,
1802 	    TR_STPDOWN, "stpdown:%p", stpdown);
1803 	rq = getendq(stp->sd_wrq);
1804 	if (cmd == I_PLINK)
1805 		rq = NULL;
1806 
1807 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1808 
1809 	strioc.ic_cmd = cmd;
1810 	strioc.ic_timout = INFTIM;
1811 	strioc.ic_len = sizeof (struct linkblk);
1812 	strioc.ic_dp = (char *)&linkp->li_lblk;
1813 
1814 	/*
1815 	 * STRPLUMB protects plumbing changes and should be set before
1816 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1817 	 * and cleared in the end of mlink when passthru queue is removed.
1818 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1819 	 * queue is in-place (it is not a proper module and doesn't have open
1820 	 * entry point).
1821 	 *
1822 	 * STPLEX prevents any threads from entering the stream from above. It
1823 	 * can't be set before the call to link_addpassthru() because putnext
1824 	 * from below may cause stream head I/O routines to be called and these
1825 	 * routines assert that STPLEX is not set. After link_addpassthru()
1826 	 * nothing may come from below since the pass queue syncq is blocked.
1827 	 * Note also that STPLEX should be cleared before the call to
1828 	 * link_rempassthru() since when messages start flowing to the stream
1829 	 * head (e.g. because of message propagation from the pass queue) stream
1830 	 * head I/O routines may be called with STPLEX flag set.
1831 	 *
1832 	 * When STPLEX is set, nothing may come into the stream from above and
1833 	 * it is safe to do a setq which will change stream head. So, the
1834 	 * correct sequence of actions is:
1835 	 *
1836 	 * 1) Set STRPLUMB
1837 	 * 2) Call link_addpassthru()
1838 	 * 3) Set STPLEX
1839 	 * 4) Call setq and update the stream state
1840 	 * 5) Clear STPLEX
1841 	 * 6) Call link_rempassthru()
1842 	 * 7) Clear STRPLUMB
1843 	 *
1844 	 * The same sequence applies to munlink() code.
1845 	 */
1846 	mutex_enter(&stpdown->sd_lock);
1847 	stpdown->sd_flag |= STRPLUMB;
1848 	mutex_exit(&stpdown->sd_lock);
1849 	/*
1850 	 * Add passthru queue below lower mux. This will block
1851 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1852 	 */
1853 	passq = link_addpassthru(stpdown);
1854 
1855 	mutex_enter(&stpdown->sd_lock);
1856 	stpdown->sd_flag |= STPLEX;
1857 	mutex_exit(&stpdown->sd_lock);
1858 
1859 	rq = _RD(stpdown->sd_wrq);
1860 	/*
1861 	 * There may be messages in the streamhead's syncq due to messages
1862 	 * that arrived before link_addpassthru() was done. To avoid
1863 	 * background processing of the syncq happening simultaneous with
1864 	 * setq processing, we disable the streamhead syncq and wait until
1865 	 * existing background thread finishes working on it.
1866 	 */
1867 	wait_sq_svc(rq->q_syncq);
1868 	passyncq = passq->q_syncq;
1869 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1870 		blocksq(passyncq, SQ_BLOCKED, 0);
1871 
1872 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1873 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1874 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1875 
1876 	/* setq might sleep in allocator - avoid holding locks. */
1877 	/* Note: we are holding muxifier here. */
1878 
1879 	str = stp->sd_strtab;
1880 	dp = &devimpl[getmajor(vp->v_rdev)];
1881 	ASSERT(dp->d_str == str);
1882 
1883 	qflag = dp->d_qflag;
1884 	sqtype = dp->d_sqtype;
1885 
1886 	/* create perdm_t if needed */
1887 	if (NEED_DM(dp->d_dmp, qflag))
1888 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1889 
1890 	dmp = dp->d_dmp;
1891 
1892 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1893 	    B_TRUE);
1894 
1895 	/*
1896 	 * XXX Remove any "odd" messages from the queue.
1897 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1898 	 */
1899 	error = strdoioctl(stp, &strioc, FNATIVE,
1900 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1901 	if (error != 0) {
1902 		lbfree(linkp);
1903 
1904 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1905 			blocksq(passyncq, SQ_BLOCKED, 0);
1906 		/*
1907 		 * Restore the stream head queue and then remove
1908 		 * the passq. Turn off STPLEX before we turn on
1909 		 * the stream by removing the passq.
1910 		 */
1911 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1912 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1913 		    B_TRUE);
1914 
1915 		mutex_enter(&stpdown->sd_lock);
1916 		stpdown->sd_flag &= ~STPLEX;
1917 		mutex_exit(&stpdown->sd_lock);
1918 
1919 		link_rempassthru(passq);
1920 
1921 		mutex_enter(&stpdown->sd_lock);
1922 		stpdown->sd_flag &= ~STRPLUMB;
1923 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1924 		cv_broadcast(&stpdown->sd_monitor);
1925 		mutex_exit(&stpdown->sd_lock);
1926 
1927 		mutex_exit(&muxifier);
1928 		netstack_rele(ss->ss_netstack);
1929 		return (error);
1930 	}
1931 	mutex_enter(&fpdown->f_tlock);
1932 	fpdown->f_count++;
1933 	mutex_exit(&fpdown->f_tlock);
1934 
1935 	/*
1936 	 * if we've made it here the linkage is all set up so we should also
1937 	 * set up the layered driver linkages
1938 	 */
1939 
1940 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1941 	if (cmd == I_LINK) {
1942 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1943 	} else {
1944 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1945 	}
1946 
1947 	link_rempassthru(passq);
1948 
1949 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1950 
1951 	/*
1952 	 * Mark the upper stream as having dependent links
1953 	 * so that strclose can clean it up.
1954 	 */
1955 	if (cmd == I_LINK) {
1956 		mutex_enter(&stp->sd_lock);
1957 		stp->sd_flag |= STRHASLINKS;
1958 		mutex_exit(&stp->sd_lock);
1959 	}
1960 	/*
1961 	 * Wake up any other processes that may have been
1962 	 * waiting on the lower stream. These will all
1963 	 * error out.
1964 	 */
1965 	mutex_enter(&stpdown->sd_lock);
1966 	/* The passthru module is removed so we may release STRPLUMB */
1967 	stpdown->sd_flag &= ~STRPLUMB;
1968 	cv_broadcast(&rq->q_wait);
1969 	cv_broadcast(&_WR(rq)->q_wait);
1970 	cv_broadcast(&stpdown->sd_monitor);
1971 	mutex_exit(&stpdown->sd_lock);
1972 	mutex_exit(&muxifier);
1973 	*rvalp = linkp->li_lblk.l_index;
1974 	netstack_rele(ss->ss_netstack);
1975 	return (0);
1976 }
1977 
1978 int
1979 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1980 {
1981 	int		ret;
1982 	struct file	*fpdown;
1983 
1984 	fpdown = getf(arg);
1985 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1986 	if (fpdown != NULL)
1987 		releasef(arg);
1988 	return (ret);
1989 }
1990 
1991 /*
1992  * Unlink a multiplexor link. Stp is the controlling stream for the
1993  * link, and linkp points to the link's entry in the linkinfo list.
1994  * The muxifier lock must be held on entry and is dropped on exit.
1995  *
1996  * NOTE : Currently it is assumed that mux would process all the messages
1997  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
1998  * of the mux to handle all the messages that arrive before UNLINK.
1999  * If the mux has to send down messages on its lower stream before
2000  * ACKing I_UNLINK, then it *should* know to handle messages even
2001  * after the UNLINK is acked (actually it should be able to handle till we
2002  * re-block the read side of the pass queue here). If the mux does not
2003  * open up the lower stream, any messages that arrive during UNLINK
2004  * will be put in the stream head. In the case of lower stream opening
2005  * up, some messages might land in the stream head depending on when
2006  * the message arrived and when the read side of the pass queue was
2007  * re-blocked.
2008  */
2009 int
2010 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2011     str_stack_t *ss)
2012 {
2013 	struct strioctl strioc;
2014 	struct stdata *stpdown;
2015 	queue_t *rq, *wrq;
2016 	queue_t	*passq;
2017 	syncq_t *passyncq;
2018 	int error = 0;
2019 	file_t *fpdown;
2020 
2021 	ASSERT(MUTEX_HELD(&muxifier));
2022 
2023 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2024 
2025 	/*
2026 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2027 	 */
2028 	mutex_enter(&stpdown->sd_lock);
2029 	stpdown->sd_flag |= STRPLUMB;
2030 	mutex_exit(&stpdown->sd_lock);
2031 
2032 	/*
2033 	 * Add passthru queue below lower mux. This will block
2034 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2035 	 */
2036 	passq = link_addpassthru(stpdown);
2037 
2038 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2039 		strioc.ic_cmd = I_UNLINK;
2040 	else
2041 		strioc.ic_cmd = I_PUNLINK;
2042 	strioc.ic_timout = INFTIM;
2043 	strioc.ic_len = sizeof (struct linkblk);
2044 	strioc.ic_dp = (char *)&linkp->li_lblk;
2045 
2046 	error = strdoioctl(stp, &strioc, FNATIVE,
2047 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2048 
2049 	/*
2050 	 * If there was an error and this is not called via strclose,
2051 	 * return to the user. Otherwise, pretend there was no error
2052 	 * and close the link.
2053 	 */
2054 	if (error) {
2055 		if (flag & LINKCLOSE) {
2056 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2057 			    "unlink ioctl, closing anyway (%d)\n", error);
2058 		} else {
2059 			link_rempassthru(passq);
2060 			mutex_enter(&stpdown->sd_lock);
2061 			stpdown->sd_flag &= ~STRPLUMB;
2062 			cv_broadcast(&stpdown->sd_monitor);
2063 			mutex_exit(&stpdown->sd_lock);
2064 			mutex_exit(&muxifier);
2065 			return (error);
2066 		}
2067 	}
2068 
2069 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2070 	fpdown = linkp->li_fpdown;
2071 	lbfree(linkp);
2072 
2073 	/*
2074 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2075 	 * can slow others down. It's okay to since attempts to mlink() this
2076 	 * stream will be stopped because STPLEX is still set in the stdata
2077 	 * structure, and munlink() is stopped because mux_rmvedge() and
2078 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2079 	 * respectively.  Note that we defer the closef() of fpdown until
2080 	 * after we drop muxifier since strclose() can call munlinkall().
2081 	 */
2082 	mutex_exit(&muxifier);
2083 
2084 	wrq = stpdown->sd_wrq;
2085 	rq = _RD(wrq);
2086 
2087 	/*
2088 	 * Get rid of outstanding service procedure runs, before we make
2089 	 * it a stream head, since a stream head doesn't have any service
2090 	 * procedure.
2091 	 */
2092 	disable_svc(rq);
2093 	wait_svc(rq);
2094 
2095 	/*
2096 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2097 	 * is queued up to be finished. mux should take care that nothing is
2098 	 * send down to this queue. We should do it now as we're going to block
2099 	 * passyncq if it was unblocked.
2100 	 */
2101 	if (wrq->q_flag & QPERMOD) {
2102 		syncq_t	*sq = wrq->q_syncq;
2103 
2104 		mutex_enter(SQLOCK(sq));
2105 		while (wrq->q_sqflags & Q_SQQUEUED) {
2106 			sq->sq_flags |= SQ_WANTWAKEUP;
2107 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2108 		}
2109 		mutex_exit(SQLOCK(sq));
2110 	}
2111 	passyncq = passq->q_syncq;
2112 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2113 
2114 		syncq_t *sq, *outer;
2115 
2116 		/*
2117 		 * Messages could be flowing from underneath. We will
2118 		 * block the read side of the passq. This would be
2119 		 * sufficient for QPAIR and QPERQ muxes to ensure
2120 		 * that no data is flowing up into this queue
2121 		 * and hence no thread active in this instance of
2122 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2123 		 * could be messages on the inner and outer/inner
2124 		 * syncqs respectively. We will wait for them to drain.
2125 		 * Because passq is blocked messages end up in the syncq
2126 		 * And qfill_syncq could possibly end up setting QFULL
2127 		 * which will access the rq->q_flag. Hence, we have to
2128 		 * acquire the QLOCK in setq.
2129 		 *
2130 		 * XXX Messages can also flow from top into this
2131 		 * queue though the unlink is over (Ex. some instance
2132 		 * in putnext() called from top that has still not
2133 		 * accessed this queue. And also putq(lowerq) ?).
2134 		 * Solution : How about blocking the l_qtop queue ?
2135 		 * Do we really care about such pure D_MP muxes ?
2136 		 */
2137 
2138 		blocksq(passyncq, SQ_BLOCKED, 0);
2139 
2140 		sq = rq->q_syncq;
2141 		if ((outer = sq->sq_outer) != NULL) {
2142 
2143 			/*
2144 			 * We have to just wait for the outer sq_count
2145 			 * drop to zero. As this does not prevent new
2146 			 * messages to enter the outer perimeter, this
2147 			 * is subject to starvation.
2148 			 *
2149 			 * NOTE :Because of blocksq above, messages could
2150 			 * be in the inner syncq only because of some
2151 			 * thread holding the outer perimeter exclusively.
2152 			 * Hence it would be sufficient to wait for the
2153 			 * exclusive holder of the outer perimeter to drain
2154 			 * the inner and outer syncqs. But we will not depend
2155 			 * on this feature and hence check the inner syncqs
2156 			 * separately.
2157 			 */
2158 			wait_syncq(outer);
2159 		}
2160 
2161 
2162 		/*
2163 		 * There could be messages destined for
2164 		 * this queue. Let the exclusive holder
2165 		 * drain it.
2166 		 */
2167 
2168 		wait_syncq(sq);
2169 		ASSERT((rq->q_flag & QPERMOD) ||
2170 		    ((rq->q_syncq->sq_head == NULL) &&
2171 		    (_WR(rq)->q_syncq->sq_head == NULL)));
2172 	}
2173 
2174 	/*
2175 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2176 	 * case as we don't disable its syncq or remove it off the syncq
2177 	 * service list.
2178 	 */
2179 	if (rq->q_flag & QPERMOD) {
2180 		syncq_t	*sq = rq->q_syncq;
2181 
2182 		mutex_enter(SQLOCK(sq));
2183 		while (rq->q_sqflags & Q_SQQUEUED) {
2184 			sq->sq_flags |= SQ_WANTWAKEUP;
2185 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2186 		}
2187 		mutex_exit(SQLOCK(sq));
2188 	}
2189 
2190 	/*
2191 	 * flush_syncq changes states only when there are some messages to
2192 	 * free, i.e. when it returns non-zero value to return.
2193 	 */
2194 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2195 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2196 
2197 	/*
2198 	 * Nobody else should know about this queue now.
2199 	 * If the mux did not process the messages before
2200 	 * acking the I_UNLINK, free them now.
2201 	 */
2202 
2203 	flushq(rq, FLUSHALL);
2204 	flushq(_WR(rq), FLUSHALL);
2205 
2206 	/*
2207 	 * Convert the mux lower queue into a stream head queue.
2208 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2209 	 */
2210 	rq->q_ptr = wrq->q_ptr = stpdown;
2211 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2212 
2213 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2214 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2215 
2216 	enable_svc(rq);
2217 
2218 	/*
2219 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2220 	 * needs to be set to prevent reopen() of the stream - such reopen may
2221 	 * try to call non-existent pass queue open routine and panic.
2222 	 */
2223 	mutex_enter(&stpdown->sd_lock);
2224 	stpdown->sd_flag &= ~STPLEX;
2225 	mutex_exit(&stpdown->sd_lock);
2226 
2227 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2228 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2229 
2230 	/* clean up the layered driver linkages */
2231 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2232 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2233 	} else {
2234 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2235 	}
2236 
2237 	link_rempassthru(passq);
2238 
2239 	/*
2240 	 * Now all plumbing changes are finished and STRPLUMB is no
2241 	 * longer needed.
2242 	 */
2243 	mutex_enter(&stpdown->sd_lock);
2244 	stpdown->sd_flag &= ~STRPLUMB;
2245 	cv_broadcast(&stpdown->sd_monitor);
2246 	mutex_exit(&stpdown->sd_lock);
2247 
2248 	(void) closef(fpdown);
2249 	return (0);
2250 }
2251 
2252 /*
2253  * Unlink all multiplexor links for which stp is the controlling stream.
2254  * Return 0, or a non-zero errno on failure.
2255  */
2256 int
2257 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2258 {
2259 	linkinfo_t *linkp;
2260 	int error = 0;
2261 
2262 	mutex_enter(&muxifier);
2263 	while (linkp = findlinks(stp, 0, flag, ss)) {
2264 		/*
2265 		 * munlink() releases the muxifier lock.
2266 		 */
2267 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2268 			return (error);
2269 		mutex_enter(&muxifier);
2270 	}
2271 	mutex_exit(&muxifier);
2272 	return (0);
2273 }
2274 
2275 /*
2276  * A multiplexor link has been made. Add an
2277  * edge to the directed graph.
2278  */
2279 void
2280 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2281 {
2282 	struct mux_node *np;
2283 	struct mux_edge *ep;
2284 	major_t upmaj;
2285 	major_t lomaj;
2286 
2287 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2288 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2289 	np = &ss->ss_mux_nodes[upmaj];
2290 	if (np->mn_outp) {
2291 		ep = np->mn_outp;
2292 		while (ep->me_nextp)
2293 			ep = ep->me_nextp;
2294 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2295 		ep = ep->me_nextp;
2296 	} else {
2297 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2298 		ep = np->mn_outp;
2299 	}
2300 	ep->me_nextp = NULL;
2301 	ep->me_muxid = muxid;
2302 	/*
2303 	 * Save the dev_t for the purposes of str_stack_shutdown.
2304 	 * str_stack_shutdown assumes that the device allows reopen, since
2305 	 * this dev_t is the one after any cloning by xx_open().
2306 	 * Would prefer finding the dev_t from before any cloning,
2307 	 * but specfs doesn't retain that.
2308 	 */
2309 	ep->me_dev = upstp->sd_vnode->v_rdev;
2310 	if (lostp->sd_vnode->v_type == VFIFO)
2311 		ep->me_nodep = NULL;
2312 	else
2313 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2314 }
2315 
2316 /*
2317  * A multiplexor link has been removed. Remove the
2318  * edge in the directed graph.
2319  */
2320 void
2321 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2322 {
2323 	struct mux_node *np;
2324 	struct mux_edge *ep;
2325 	struct mux_edge *pep = NULL;
2326 	major_t upmaj;
2327 
2328 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2329 	np = &ss->ss_mux_nodes[upmaj];
2330 	ASSERT(np->mn_outp != NULL);
2331 	ep = np->mn_outp;
2332 	while (ep) {
2333 		if (ep->me_muxid == muxid) {
2334 			if (pep)
2335 				pep->me_nextp = ep->me_nextp;
2336 			else
2337 				np->mn_outp = ep->me_nextp;
2338 			kmem_free(ep, sizeof (struct mux_edge));
2339 			return;
2340 		}
2341 		pep = ep;
2342 		ep = ep->me_nextp;
2343 	}
2344 	ASSERT(0);	/* should not reach here */
2345 }
2346 
2347 /*
2348  * Translate the device flags (from conf.h) to the corresponding
2349  * qflag and sq_flag (type) values.
2350  */
2351 int
2352 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2353 	uint32_t *sqtypep)
2354 {
2355 	uint32_t qflag = 0;
2356 	uint32_t sqtype = 0;
2357 
2358 	if (devflag & _D_OLD)
2359 		goto bad;
2360 
2361 	/* Inner perimeter presence and scope */
2362 	switch (devflag & D_MTINNER_MASK) {
2363 	case D_MP:
2364 		qflag |= QMTSAFE;
2365 		sqtype |= SQ_CI;
2366 		break;
2367 	case D_MTPERQ|D_MP:
2368 		qflag |= QPERQ;
2369 		break;
2370 	case D_MTQPAIR|D_MP:
2371 		qflag |= QPAIR;
2372 		break;
2373 	case D_MTPERMOD|D_MP:
2374 		qflag |= QPERMOD;
2375 		break;
2376 	default:
2377 		goto bad;
2378 	}
2379 
2380 	/* Outer perimeter */
2381 	if (devflag & D_MTOUTPERIM) {
2382 		switch (devflag & D_MTINNER_MASK) {
2383 		case D_MP:
2384 		case D_MTPERQ|D_MP:
2385 		case D_MTQPAIR|D_MP:
2386 			break;
2387 		default:
2388 			goto bad;
2389 		}
2390 		qflag |= QMTOUTPERIM;
2391 	}
2392 
2393 	/* Inner perimeter modifiers */
2394 	if (devflag & D_MTINNER_MOD) {
2395 		switch (devflag & D_MTINNER_MASK) {
2396 		case D_MP:
2397 			goto bad;
2398 		default:
2399 			break;
2400 		}
2401 		if (devflag & D_MTPUTSHARED)
2402 			sqtype |= SQ_CIPUT;
2403 		if (devflag & _D_MTOCSHARED) {
2404 			/*
2405 			 * The code in putnext assumes that it has the
2406 			 * highest concurrency by not checking sq_count.
2407 			 * Thus _D_MTOCSHARED can only be supported when
2408 			 * D_MTPUTSHARED is set.
2409 			 */
2410 			if (!(devflag & D_MTPUTSHARED))
2411 				goto bad;
2412 			sqtype |= SQ_CIOC;
2413 		}
2414 		if (devflag & _D_MTCBSHARED) {
2415 			/*
2416 			 * The code in putnext assumes that it has the
2417 			 * highest concurrency by not checking sq_count.
2418 			 * Thus _D_MTCBSHARED can only be supported when
2419 			 * D_MTPUTSHARED is set.
2420 			 */
2421 			if (!(devflag & D_MTPUTSHARED))
2422 				goto bad;
2423 			sqtype |= SQ_CICB;
2424 		}
2425 		if (devflag & _D_MTSVCSHARED) {
2426 			/*
2427 			 * The code in putnext assumes that it has the
2428 			 * highest concurrency by not checking sq_count.
2429 			 * Thus _D_MTSVCSHARED can only be supported when
2430 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2431 			 * supported only for QPERMOD.
2432 			 */
2433 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2434 				goto bad;
2435 			sqtype |= SQ_CISVC;
2436 		}
2437 	}
2438 
2439 	/* Default outer perimeter concurrency */
2440 	sqtype |= SQ_CO;
2441 
2442 	/* Outer perimeter modifiers */
2443 	if (devflag & D_MTOCEXCL) {
2444 		if (!(devflag & D_MTOUTPERIM)) {
2445 			/* No outer perimeter */
2446 			goto bad;
2447 		}
2448 		sqtype &= ~SQ_COOC;
2449 	}
2450 
2451 	/* Synchronous Streams extended qinit structure */
2452 	if (devflag & D_SYNCSTR)
2453 		qflag |= QSYNCSTR;
2454 
2455 	/*
2456 	 * Private flag used by a transport module to indicate
2457 	 * to sockfs that it supports direct-access mode without
2458 	 * having to go through STREAMS.
2459 	 */
2460 	if (devflag & _D_DIRECT) {
2461 		/* Reject unless the module is fully-MT (no perimeter) */
2462 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2463 			goto bad;
2464 		qflag |= _QDIRECT;
2465 	}
2466 
2467 	*qflagp = qflag;
2468 	*sqtypep = sqtype;
2469 	return (0);
2470 
2471 bad:
2472 	cmn_err(CE_WARN,
2473 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2474 	    (int)(qflag & D_MTSAFETY_MASK),
2475 	    stp->st_rdinit->qi_minfo->mi_idname);
2476 
2477 	return (EINVAL);
2478 }
2479 
2480 /*
2481  * Set the interface values for a pair of queues (qinit structure,
2482  * packet sizes, water marks).
2483  * setq assumes that the caller does not have a claim (entersq or claimq)
2484  * on the queue.
2485  */
2486 void
2487 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2488     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2489 {
2490 	queue_t *wq;
2491 	syncq_t	*sq, *outer;
2492 
2493 	ASSERT(rq->q_flag & QREADR);
2494 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2495 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2496 
2497 	wq = _WR(rq);
2498 	rq->q_qinfo = rinit;
2499 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2500 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2501 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2502 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2503 	wq->q_qinfo = winit;
2504 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2505 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2506 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2507 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2508 
2509 	/* Remove old syncqs */
2510 	sq = rq->q_syncq;
2511 	outer = sq->sq_outer;
2512 	if (outer != NULL) {
2513 		ASSERT(wq->q_syncq->sq_outer == outer);
2514 		outer_remove(outer, rq->q_syncq);
2515 		if (wq->q_syncq != rq->q_syncq)
2516 			outer_remove(outer, wq->q_syncq);
2517 	}
2518 	ASSERT(sq->sq_outer == NULL);
2519 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2520 
2521 	if (sq != SQ(rq)) {
2522 		if (!(rq->q_flag & QPERMOD))
2523 			free_syncq(sq);
2524 		if (wq->q_syncq == rq->q_syncq)
2525 			wq->q_syncq = NULL;
2526 		rq->q_syncq = NULL;
2527 	}
2528 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2529 	    wq->q_syncq != SQ(rq)) {
2530 		free_syncq(wq->q_syncq);
2531 		wq->q_syncq = NULL;
2532 	}
2533 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2534 	    rq->q_syncq->sq_tail == NULL));
2535 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2536 	    wq->q_syncq->sq_tail == NULL));
2537 
2538 	if (!(rq->q_flag & QPERMOD) &&
2539 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2540 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2541 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2542 		    rq->q_syncq->sq_nciputctrl, 0);
2543 		ASSERT(ciputctrl_cache != NULL);
2544 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2545 		rq->q_syncq->sq_ciputctrl = NULL;
2546 		rq->q_syncq->sq_nciputctrl = 0;
2547 	}
2548 
2549 	if (!(wq->q_flag & QPERMOD) &&
2550 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2551 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2552 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2553 		    wq->q_syncq->sq_nciputctrl, 0);
2554 		ASSERT(ciputctrl_cache != NULL);
2555 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2556 		wq->q_syncq->sq_ciputctrl = NULL;
2557 		wq->q_syncq->sq_nciputctrl = 0;
2558 	}
2559 
2560 	sq = SQ(rq);
2561 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2562 	ASSERT(sq->sq_outer == NULL);
2563 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2564 
2565 	/*
2566 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2567 	 * bits in sq_flag based on the sqtype.
2568 	 */
2569 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2570 
2571 	rq->q_syncq = wq->q_syncq = sq;
2572 	sq->sq_type = sqtype;
2573 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2574 
2575 	/*
2576 	 *  We are making sq_svcflags zero,
2577 	 *  resetting SQ_DISABLED in case it was set by
2578 	 *  wait_svc() in the munlink path.
2579 	 *
2580 	 */
2581 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2582 	sq->sq_svcflags = 0;
2583 
2584 	/*
2585 	 * We need to acquire the lock here for the mlink and munlink case,
2586 	 * where canputnext, backenable, etc can access the q_flag.
2587 	 */
2588 	if (lock_needed) {
2589 		mutex_enter(QLOCK(rq));
2590 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2591 		mutex_exit(QLOCK(rq));
2592 		mutex_enter(QLOCK(wq));
2593 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2594 		mutex_exit(QLOCK(wq));
2595 	} else {
2596 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2597 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2598 	}
2599 
2600 	if (qflag & QPERQ) {
2601 		/* Allocate a separate syncq for the write side */
2602 		sq = new_syncq();
2603 		sq->sq_type = rq->q_syncq->sq_type;
2604 		sq->sq_flags = rq->q_syncq->sq_flags;
2605 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2606 		    sq->sq_oprev == NULL);
2607 		wq->q_syncq = sq;
2608 	}
2609 	if (qflag & QPERMOD) {
2610 		sq = dmp->dm_sq;
2611 
2612 		/*
2613 		 * Assert that we do have an inner perimeter syncq and that it
2614 		 * does not have an outer perimeter associated with it.
2615 		 */
2616 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2617 		    sq->sq_oprev == NULL);
2618 		rq->q_syncq = wq->q_syncq = sq;
2619 	}
2620 	if (qflag & QMTOUTPERIM) {
2621 		outer = dmp->dm_sq;
2622 
2623 		ASSERT(outer->sq_outer == NULL);
2624 		outer_insert(outer, rq->q_syncq);
2625 		if (wq->q_syncq != rq->q_syncq)
2626 			outer_insert(outer, wq->q_syncq);
2627 	}
2628 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2629 	    (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2630 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2631 	    (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2632 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2633 
2634 	/*
2635 	 * Initialize struio() types.
2636 	 */
2637 	rq->q_struiot =
2638 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2639 	wq->q_struiot =
2640 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2641 }
2642 
2643 perdm_t *
2644 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2645 {
2646 	syncq_t	*sq;
2647 	perdm_t	**pp;
2648 	perdm_t	*p;
2649 	perdm_t	*dmp;
2650 
2651 	ASSERT(str != NULL);
2652 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2653 
2654 	rw_enter(&perdm_rwlock, RW_READER);
2655 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2656 		if (p->dm_str == str) {	/* found one */
2657 			atomic_inc_32(&(p->dm_ref));
2658 			rw_exit(&perdm_rwlock);
2659 			return (p);
2660 		}
2661 	}
2662 	rw_exit(&perdm_rwlock);
2663 
2664 	sq = new_syncq();
2665 	if (qflag & QPERMOD) {
2666 		sq->sq_type = sqtype | SQ_PERMOD;
2667 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2668 	} else {
2669 		ASSERT(qflag & QMTOUTPERIM);
2670 		sq->sq_onext = sq->sq_oprev = sq;
2671 	}
2672 
2673 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2674 	dmp->dm_sq = sq;
2675 	dmp->dm_str = str;
2676 	dmp->dm_ref = 1;
2677 	dmp->dm_next = NULL;
2678 
2679 	rw_enter(&perdm_rwlock, RW_WRITER);
2680 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2681 		if (p->dm_str == str) {	/* already present */
2682 			p->dm_ref++;
2683 			rw_exit(&perdm_rwlock);
2684 			free_syncq(sq);
2685 			kmem_free(dmp, sizeof (perdm_t));
2686 			return (p);
2687 		}
2688 	}
2689 
2690 	*pp = dmp;
2691 	rw_exit(&perdm_rwlock);
2692 	return (dmp);
2693 }
2694 
2695 void
2696 rele_dm(perdm_t *dmp)
2697 {
2698 	perdm_t **pp;
2699 	perdm_t *p;
2700 
2701 	rw_enter(&perdm_rwlock, RW_WRITER);
2702 	ASSERT(dmp->dm_ref > 0);
2703 
2704 	if (--dmp->dm_ref > 0) {
2705 		rw_exit(&perdm_rwlock);
2706 		return;
2707 	}
2708 
2709 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2710 		if (p == dmp)
2711 			break;
2712 	ASSERT(p == dmp);
2713 	*pp = p->dm_next;
2714 	rw_exit(&perdm_rwlock);
2715 
2716 	/*
2717 	 * Wait for any background processing that relies on the
2718 	 * syncq to complete before it is freed.
2719 	 */
2720 	wait_sq_svc(p->dm_sq);
2721 	free_syncq(p->dm_sq);
2722 	kmem_free(p, sizeof (perdm_t));
2723 }
2724 
2725 /*
2726  * Make a protocol message given control and data buffers.
2727  * n.b., this can block; be careful of what locks you hold when calling it.
2728  *
2729  * If sd_maxblk is less than *iosize this routine can fail part way through
2730  * (due to an allocation failure). In this case on return *iosize will contain
2731  * the amount that was consumed. Otherwise *iosize will not be modified
2732  * i.e. it will contain the amount that was consumed.
2733  */
2734 int
2735 strmakemsg(
2736 	struct strbuf *mctl,
2737 	ssize_t *iosize,
2738 	struct uio *uiop,
2739 	stdata_t *stp,
2740 	int32_t flag,
2741 	mblk_t **mpp)
2742 {
2743 	mblk_t *mpctl = NULL;
2744 	mblk_t *mpdata = NULL;
2745 	int error;
2746 
2747 	ASSERT(uiop != NULL);
2748 
2749 	*mpp = NULL;
2750 	/* Create control part, if any */
2751 	if ((mctl != NULL) && (mctl->len >= 0)) {
2752 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2753 		if (error)
2754 			return (error);
2755 	}
2756 	/* Create data part, if any */
2757 	if (*iosize >= 0) {
2758 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2759 		if (error) {
2760 			freemsg(mpctl);
2761 			return (error);
2762 		}
2763 	}
2764 	if (mpctl != NULL) {
2765 		if (mpdata != NULL)
2766 			linkb(mpctl, mpdata);
2767 		*mpp = mpctl;
2768 	} else {
2769 		*mpp = mpdata;
2770 	}
2771 	return (0);
2772 }
2773 
2774 /*
2775  * Make the control part of a protocol message given a control buffer.
2776  * n.b., this can block; be careful of what locks you hold when calling it.
2777  */
2778 int
2779 strmakectl(
2780 	struct strbuf *mctl,
2781 	int32_t flag,
2782 	int32_t fflag,
2783 	mblk_t **mpp)
2784 {
2785 	mblk_t *bp = NULL;
2786 	unsigned char msgtype;
2787 	int error = 0;
2788 	cred_t *cr = CRED();
2789 
2790 	/* We do not support interrupt threads using the stream head to send */
2791 	ASSERT(cr != NULL);
2792 
2793 	*mpp = NULL;
2794 	/*
2795 	 * Create control part of message, if any.
2796 	 */
2797 	if ((mctl != NULL) && (mctl->len >= 0)) {
2798 		caddr_t base;
2799 		int ctlcount;
2800 		int allocsz;
2801 
2802 		if (flag & RS_HIPRI)
2803 			msgtype = M_PCPROTO;
2804 		else
2805 			msgtype = M_PROTO;
2806 
2807 		ctlcount = mctl->len;
2808 		base = mctl->buf;
2809 
2810 		/*
2811 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2812 		 * blocks by increasing the size to something more usable.
2813 		 */
2814 		allocsz = MAX(ctlcount, 64);
2815 
2816 		/*
2817 		 * Range checking has already been done; simply try
2818 		 * to allocate a message block for the ctl part.
2819 		 */
2820 		while ((bp = allocb_cred(allocsz, cr,
2821 		    curproc->p_pid)) == NULL) {
2822 			if (fflag & (FNDELAY|FNONBLOCK))
2823 				return (EAGAIN);
2824 			if (error = strwaitbuf(allocsz, BPRI_MED))
2825 				return (error);
2826 		}
2827 
2828 		bp->b_datap->db_type = msgtype;
2829 		if (copyin(base, bp->b_wptr, ctlcount)) {
2830 			freeb(bp);
2831 			return (EFAULT);
2832 		}
2833 		bp->b_wptr += ctlcount;
2834 	}
2835 	*mpp = bp;
2836 	return (0);
2837 }
2838 
2839 /*
2840  * Make a protocol message given data buffers.
2841  * n.b., this can block; be careful of what locks you hold when calling it.
2842  *
2843  * If sd_maxblk is less than *iosize this routine can fail part way through
2844  * (due to an allocation failure). In this case on return *iosize will contain
2845  * the amount that was consumed. Otherwise *iosize will not be modified
2846  * i.e. it will contain the amount that was consumed.
2847  */
2848 int
2849 strmakedata(
2850 	ssize_t   *iosize,
2851 	struct uio *uiop,
2852 	stdata_t *stp,
2853 	int32_t flag,
2854 	mblk_t **mpp)
2855 {
2856 	mblk_t *mp = NULL;
2857 	mblk_t *bp;
2858 	int wroff = (int)stp->sd_wroff;
2859 	int tail_len = (int)stp->sd_tail;
2860 	int extra = wroff + tail_len;
2861 	int error = 0;
2862 	ssize_t maxblk;
2863 	ssize_t count = *iosize;
2864 	cred_t *cr;
2865 
2866 	*mpp = NULL;
2867 	if (count < 0)
2868 		return (0);
2869 
2870 	/* We do not support interrupt threads using the stream head to send */
2871 	cr = CRED();
2872 	ASSERT(cr != NULL);
2873 
2874 	maxblk = stp->sd_maxblk;
2875 	if (maxblk == INFPSZ)
2876 		maxblk = count;
2877 
2878 	/*
2879 	 * Create data part of message, if any.
2880 	 */
2881 	do {
2882 		ssize_t size;
2883 		dblk_t  *dp;
2884 
2885 		ASSERT(uiop);
2886 
2887 		size = MIN(count, maxblk);
2888 
2889 		while ((bp = allocb_cred(size + extra, cr,
2890 		    curproc->p_pid)) == NULL) {
2891 			error = EAGAIN;
2892 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2893 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2894 				if (count == *iosize) {
2895 					freemsg(mp);
2896 					return (error);
2897 				} else {
2898 					*iosize -= count;
2899 					*mpp = mp;
2900 					return (0);
2901 				}
2902 			}
2903 		}
2904 		dp = bp->b_datap;
2905 		dp->db_cpid = curproc->p_pid;
2906 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2907 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2908 
2909 		if (flag & STRUIO_POSTPONE) {
2910 			/*
2911 			 * Setup the stream uio portion of the
2912 			 * dblk for subsequent use by struioget().
2913 			 */
2914 			dp->db_struioflag = STRUIO_SPEC;
2915 			dp->db_cksumstart = 0;
2916 			dp->db_cksumstuff = 0;
2917 			dp->db_cksumend = size;
2918 			*(long long *)dp->db_struioun.data = 0ll;
2919 			bp->b_wptr += size;
2920 		} else {
2921 			if (stp->sd_copyflag & STRCOPYCACHED)
2922 				uiop->uio_extflg |= UIO_COPY_CACHED;
2923 
2924 			if (size != 0) {
2925 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2926 				    uiop);
2927 				if (error != 0) {
2928 					freeb(bp);
2929 					freemsg(mp);
2930 					return (error);
2931 				}
2932 			}
2933 			bp->b_wptr += size;
2934 
2935 			if (stp->sd_wputdatafunc != NULL) {
2936 				mblk_t *newbp;
2937 
2938 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2939 				    bp, NULL, NULL, NULL, NULL);
2940 				if (newbp == NULL) {
2941 					freeb(bp);
2942 					freemsg(mp);
2943 					return (ECOMM);
2944 				}
2945 				bp = newbp;
2946 			}
2947 		}
2948 
2949 		count -= size;
2950 
2951 		if (mp == NULL)
2952 			mp = bp;
2953 		else
2954 			linkb(mp, bp);
2955 	} while (count > 0);
2956 
2957 	*mpp = mp;
2958 	return (0);
2959 }
2960 
2961 /*
2962  * Wait for a buffer to become available. Return non-zero errno
2963  * if not able to wait, 0 if buffer is probably there.
2964  */
2965 int
2966 strwaitbuf(size_t size, int pri)
2967 {
2968 	bufcall_id_t id;
2969 
2970 	mutex_enter(&bcall_monitor);
2971 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2972 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2973 		mutex_exit(&bcall_monitor);
2974 		return (ENOSR);
2975 	}
2976 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2977 		unbufcall(id);
2978 		mutex_exit(&bcall_monitor);
2979 		return (EINTR);
2980 	}
2981 	unbufcall(id);
2982 	mutex_exit(&bcall_monitor);
2983 	return (0);
2984 }
2985 
2986 /*
2987  * This function waits for a read or write event to happen on a stream.
2988  * fmode can specify FNDELAY and/or FNONBLOCK.
2989  * The timeout is in ms with -1 meaning infinite.
2990  * The flag values work as follows:
2991  *	READWAIT	Check for read side errors, send M_READ
2992  *	GETWAIT		Check for read side errors, no M_READ
2993  *	WRITEWAIT	Check for write side errors.
2994  *	NOINTR		Do not return error if nonblocking or timeout.
2995  * 	STR_NOERROR	Ignore all errors except STPLEX.
2996  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
2997  *	STR_PEEK	Pass through the strgeterr().
2998  */
2999 int
3000 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
3001     int *done)
3002 {
3003 	int slpflg, errs;
3004 	int error;
3005 	kcondvar_t *sleepon;
3006 	mblk_t *mp;
3007 	ssize_t *rd_count;
3008 	clock_t rval;
3009 
3010 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3011 	if ((flag & READWAIT) || (flag & GETWAIT)) {
3012 		slpflg = RSLEEP;
3013 		sleepon = &_RD(stp->sd_wrq)->q_wait;
3014 		errs = STRDERR|STPLEX;
3015 	} else {
3016 		slpflg = WSLEEP;
3017 		sleepon = &stp->sd_wrq->q_wait;
3018 		errs = STWRERR|STRHUP|STPLEX;
3019 	}
3020 	if (flag & STR_NOERROR)
3021 		errs = STPLEX;
3022 
3023 	if (stp->sd_wakeq & slpflg) {
3024 		/*
3025 		 * A strwakeq() is pending, no need to sleep.
3026 		 */
3027 		stp->sd_wakeq &= ~slpflg;
3028 		*done = 0;
3029 		return (0);
3030 	}
3031 
3032 	if (stp->sd_flag & errs) {
3033 		/*
3034 		 * Check for errors before going to sleep since the
3035 		 * caller might not have checked this while holding
3036 		 * sd_lock.
3037 		 */
3038 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3039 		if (error != 0) {
3040 			*done = 1;
3041 			return (error);
3042 		}
3043 	}
3044 
3045 	/*
3046 	 * If any module downstream has requested read notification
3047 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3048 	 * down stream.
3049 	 */
3050 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3051 		mutex_exit(&stp->sd_lock);
3052 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3053 		    (flag & STR_NOSIG), &error))) {
3054 			mutex_enter(&stp->sd_lock);
3055 			*done = 1;
3056 			return (error);
3057 		}
3058 		mp->b_datap->db_type = M_READ;
3059 		rd_count = (ssize_t *)mp->b_wptr;
3060 		*rd_count = count;
3061 		mp->b_wptr += sizeof (ssize_t);
3062 		/*
3063 		 * Send the number of bytes requested by the
3064 		 * read as the argument to M_READ.
3065 		 */
3066 		stream_willservice(stp);
3067 		putnext(stp->sd_wrq, mp);
3068 		stream_runservice(stp);
3069 		mutex_enter(&stp->sd_lock);
3070 
3071 		/*
3072 		 * If any data arrived due to inline processing
3073 		 * of putnext(), don't sleep.
3074 		 */
3075 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3076 			*done = 0;
3077 			return (0);
3078 		}
3079 	}
3080 
3081 	if (fmode & (FNDELAY|FNONBLOCK)) {
3082 		if (!(flag & NOINTR))
3083 			error = EAGAIN;
3084 		else
3085 			error = 0;
3086 		*done = 1;
3087 		return (error);
3088 	}
3089 
3090 	stp->sd_flag |= slpflg;
3091 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3092 	    "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3093 	    stp, flag, count, fmode, done);
3094 
3095 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3096 	if (rval > 0) {
3097 		/* EMPTY */
3098 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3099 		    "strwaitq awakes(2):%X, %X, %X, %X, %X",
3100 		    stp, flag, count, fmode, done);
3101 	} else if (rval == 0) {
3102 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3103 		    "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3104 		    stp, flag, count, fmode, done);
3105 		stp->sd_flag &= ~slpflg;
3106 		cv_broadcast(sleepon);
3107 		if (!(flag & NOINTR))
3108 			error = EINTR;
3109 		else
3110 			error = 0;
3111 		*done = 1;
3112 		return (error);
3113 	} else {
3114 		/* timeout */
3115 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3116 		    "strwaitq timeout:%p, %X, %lX, %X, %p",
3117 		    stp, flag, count, fmode, done);
3118 		*done = 1;
3119 		if (!(flag & NOINTR))
3120 			return (ETIME);
3121 		else
3122 			return (0);
3123 	}
3124 	/*
3125 	 * If the caller implements delayed errors (i.e. queued after data)
3126 	 * we can not check for errors here since data as well as an
3127 	 * error might have arrived at the stream head. We return to
3128 	 * have the caller check the read queue before checking for errors.
3129 	 */
3130 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3131 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3132 		if (error != 0) {
3133 			*done = 1;
3134 			return (error);
3135 		}
3136 	}
3137 	*done = 0;
3138 	return (0);
3139 }
3140 
3141 /*
3142  * Perform job control discipline access checks.
3143  * Return 0 for success and the errno for failure.
3144  */
3145 
3146 #define	cantsend(p, t, sig) \
3147 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3148 
3149 int
3150 straccess(struct stdata *stp, enum jcaccess mode)
3151 {
3152 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3153 	kthread_t *t = curthread;
3154 	proc_t *p = ttoproc(t);
3155 	sess_t *sp;
3156 
3157 	ASSERT(mutex_owned(&stp->sd_lock));
3158 
3159 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3160 		return (0);
3161 
3162 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3163 
3164 	for (;;) {
3165 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3166 		sp = p->p_sessp;
3167 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3168 
3169 		/*
3170 		 * If this is not the calling process's controlling terminal
3171 		 * or if the calling process is already in the foreground
3172 		 * then allow access.
3173 		 */
3174 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3175 		    p->p_pgidp == stp->sd_pgidp) {
3176 			mutex_exit(&sp->s_lock);
3177 			mutex_exit(&p->p_splock);
3178 			mutex_exit(&p->p_lock);
3179 			return (0);
3180 		}
3181 
3182 		/*
3183 		 * Check to see if controlling terminal has been deallocated.
3184 		 */
3185 		if (sp->s_vp == NULL) {
3186 			if (!cantsend(p, t, SIGHUP))
3187 				sigtoproc(p, t, SIGHUP);
3188 			mutex_exit(&sp->s_lock);
3189 			mutex_exit(&p->p_splock);
3190 			mutex_exit(&p->p_lock);
3191 			return (EIO);
3192 		}
3193 
3194 		mutex_exit(&sp->s_lock);
3195 		mutex_exit(&p->p_splock);
3196 
3197 		if (mode == JCGETP) {
3198 			mutex_exit(&p->p_lock);
3199 			return (0);
3200 		}
3201 
3202 		if (mode == JCREAD) {
3203 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3204 				mutex_exit(&p->p_lock);
3205 				return (EIO);
3206 			}
3207 			mutex_exit(&p->p_lock);
3208 			mutex_exit(&stp->sd_lock);
3209 			pgsignal(p->p_pgidp, SIGTTIN);
3210 			mutex_enter(&stp->sd_lock);
3211 			mutex_enter(&p->p_lock);
3212 		} else {  /* mode == JCWRITE or JCSETP */
3213 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3214 			    cantsend(p, t, SIGTTOU)) {
3215 				mutex_exit(&p->p_lock);
3216 				return (0);
3217 			}
3218 			if (p->p_detached) {
3219 				mutex_exit(&p->p_lock);
3220 				return (EIO);
3221 			}
3222 			mutex_exit(&p->p_lock);
3223 			mutex_exit(&stp->sd_lock);
3224 			pgsignal(p->p_pgidp, SIGTTOU);
3225 			mutex_enter(&stp->sd_lock);
3226 			mutex_enter(&p->p_lock);
3227 		}
3228 
3229 		/*
3230 		 * We call cv_wait_sig_swap() to cause the appropriate
3231 		 * action for the jobcontrol signal to take place.
3232 		 * If the signal is being caught, we will take the
3233 		 * EINTR error return.  Otherwise, the default action
3234 		 * of causing the process to stop will take place.
3235 		 * In this case, we rely on the periodic cv_broadcast() on
3236 		 * &lbolt_cv to wake us up to loop around and test again.
3237 		 * We can't get here if the signal is ignored or
3238 		 * if the current thread is blocking the signal.
3239 		 */
3240 		mutex_exit(&stp->sd_lock);
3241 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3242 			mutex_exit(&p->p_lock);
3243 			mutex_enter(&stp->sd_lock);
3244 			return (EINTR);
3245 		}
3246 		mutex_exit(&p->p_lock);
3247 		mutex_enter(&stp->sd_lock);
3248 		mutex_enter(&p->p_lock);
3249 	}
3250 }
3251 
3252 /*
3253  * Return size of message of block type (bp->b_datap->db_type)
3254  */
3255 size_t
3256 xmsgsize(mblk_t *bp)
3257 {
3258 	unsigned char type;
3259 	size_t count = 0;
3260 
3261 	type = bp->b_datap->db_type;
3262 
3263 	for (; bp; bp = bp->b_cont) {
3264 		if (type != bp->b_datap->db_type)
3265 			break;
3266 		ASSERT(bp->b_wptr >= bp->b_rptr);
3267 		count += bp->b_wptr - bp->b_rptr;
3268 	}
3269 	return (count);
3270 }
3271 
3272 /*
3273  * Allocate a stream head.
3274  */
3275 struct stdata *
3276 shalloc(queue_t *qp)
3277 {
3278 	stdata_t *stp;
3279 
3280 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3281 
3282 	stp->sd_wrq = _WR(qp);
3283 	stp->sd_strtab = NULL;
3284 	stp->sd_iocid = 0;
3285 	stp->sd_mate = NULL;
3286 	stp->sd_freezer = NULL;
3287 	stp->sd_refcnt = 0;
3288 	stp->sd_wakeq = 0;
3289 	stp->sd_anchor = 0;
3290 	stp->sd_struiowrq = NULL;
3291 	stp->sd_struiordq = NULL;
3292 	stp->sd_struiodnak = 0;
3293 	stp->sd_struionak = NULL;
3294 	stp->sd_t_audit_data = NULL;
3295 	stp->sd_rput_opt = 0;
3296 	stp->sd_wput_opt = 0;
3297 	stp->sd_read_opt = 0;
3298 	stp->sd_rprotofunc = strrput_proto;
3299 	stp->sd_rmiscfunc = strrput_misc;
3300 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3301 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3302 	stp->sd_ciputctrl = NULL;
3303 	stp->sd_nciputctrl = 0;
3304 	stp->sd_qhead = NULL;
3305 	stp->sd_qtail = NULL;
3306 	stp->sd_servid = NULL;
3307 	stp->sd_nqueues = 0;
3308 	stp->sd_svcflags = 0;
3309 	stp->sd_copyflag = 0;
3310 
3311 	return (stp);
3312 }
3313 
3314 /*
3315  * Free a stream head.
3316  */
3317 void
3318 shfree(stdata_t *stp)
3319 {
3320 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3321 
3322 	stp->sd_wrq = NULL;
3323 
3324 	mutex_enter(&stp->sd_qlock);
3325 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3326 		STRSTAT(strwaits);
3327 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3328 	}
3329 	mutex_exit(&stp->sd_qlock);
3330 
3331 	if (stp->sd_ciputctrl != NULL) {
3332 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3333 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3334 		    stp->sd_nciputctrl, 0);
3335 		ASSERT(ciputctrl_cache != NULL);
3336 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3337 		stp->sd_ciputctrl = NULL;
3338 		stp->sd_nciputctrl = 0;
3339 	}
3340 	ASSERT(stp->sd_qhead == NULL);
3341 	ASSERT(stp->sd_qtail == NULL);
3342 	ASSERT(stp->sd_nqueues == 0);
3343 	kmem_cache_free(stream_head_cache, stp);
3344 }
3345 
3346 /*
3347  * Allocate a pair of queues and a syncq for the pair
3348  */
3349 queue_t *
3350 allocq(void)
3351 {
3352 	queinfo_t *qip;
3353 	queue_t *qp, *wqp;
3354 	syncq_t	*sq;
3355 
3356 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3357 
3358 	qp = &qip->qu_rqueue;
3359 	wqp = &qip->qu_wqueue;
3360 	sq = &qip->qu_syncq;
3361 
3362 	qp->q_last	= NULL;
3363 	qp->q_next	= NULL;
3364 	qp->q_ptr	= NULL;
3365 	qp->q_flag	= QUSE | QREADR;
3366 	qp->q_bandp	= NULL;
3367 	qp->q_stream	= NULL;
3368 	qp->q_syncq	= sq;
3369 	qp->q_nband	= 0;
3370 	qp->q_nfsrv	= NULL;
3371 	qp->q_draining	= 0;
3372 	qp->q_syncqmsgs	= 0;
3373 	qp->q_spri	= 0;
3374 	qp->q_qtstamp	= 0;
3375 	qp->q_sqtstamp	= 0;
3376 	qp->q_fp	= NULL;
3377 
3378 	wqp->q_last	= NULL;
3379 	wqp->q_next	= NULL;
3380 	wqp->q_ptr	= NULL;
3381 	wqp->q_flag	= QUSE;
3382 	wqp->q_bandp	= NULL;
3383 	wqp->q_stream	= NULL;
3384 	wqp->q_syncq	= sq;
3385 	wqp->q_nband	= 0;
3386 	wqp->q_nfsrv	= NULL;
3387 	wqp->q_draining	= 0;
3388 	wqp->q_syncqmsgs = 0;
3389 	wqp->q_qtstamp	= 0;
3390 	wqp->q_sqtstamp	= 0;
3391 	wqp->q_spri	= 0;
3392 
3393 	sq->sq_count	= 0;
3394 	sq->sq_rmqcount	= 0;
3395 	sq->sq_flags	= 0;
3396 	sq->sq_type	= 0;
3397 	sq->sq_callbflags = 0;
3398 	sq->sq_cancelid	= 0;
3399 	sq->sq_ciputctrl = NULL;
3400 	sq->sq_nciputctrl = 0;
3401 	sq->sq_needexcl = 0;
3402 	sq->sq_svcflags = 0;
3403 
3404 	return (qp);
3405 }
3406 
3407 /*
3408  * Free a pair of queues and the "attached" syncq.
3409  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3410  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3411  */
3412 void
3413 freeq(queue_t *qp)
3414 {
3415 	qband_t *qbp, *nqbp;
3416 	syncq_t *sq, *outer;
3417 	queue_t *wqp = _WR(qp);
3418 
3419 	ASSERT(qp->q_flag & QREADR);
3420 
3421 	/*
3422 	 * If a previously dispatched taskq job is scheduled to run
3423 	 * sync_service() or a service routine is scheduled for the
3424 	 * queues about to be freed, wait here until all service is
3425 	 * done on the queue and all associated queues and syncqs.
3426 	 */
3427 	wait_svc(qp);
3428 
3429 	(void) flush_syncq(qp->q_syncq, qp);
3430 	(void) flush_syncq(wqp->q_syncq, wqp);
3431 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3432 
3433 	/*
3434 	 * Flush the queues before q_next is set to NULL This is needed
3435 	 * in order to backenable any downstream queue before we go away.
3436 	 * Note: we are already removed from the stream so that the
3437 	 * backenabling will not cause any messages to be delivered to our
3438 	 * put procedures.
3439 	 */
3440 	flushq(qp, FLUSHALL);
3441 	flushq(wqp, FLUSHALL);
3442 
3443 	/* Tidy up - removeq only does a half-remove from stream */
3444 	qp->q_next = wqp->q_next = NULL;
3445 	ASSERT(!(qp->q_flag & QENAB));
3446 	ASSERT(!(wqp->q_flag & QENAB));
3447 
3448 	outer = qp->q_syncq->sq_outer;
3449 	if (outer != NULL) {
3450 		outer_remove(outer, qp->q_syncq);
3451 		if (wqp->q_syncq != qp->q_syncq)
3452 			outer_remove(outer, wqp->q_syncq);
3453 	}
3454 	/*
3455 	 * Free any syncqs that are outside what allocq returned.
3456 	 */
3457 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3458 		free_syncq(qp->q_syncq);
3459 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3460 		free_syncq(wqp->q_syncq);
3461 
3462 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3463 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3464 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3465 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3466 	sq = SQ(qp);
3467 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3468 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3469 	ASSERT(sq->sq_outer == NULL);
3470 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3471 	ASSERT(sq->sq_callbpend == NULL);
3472 	ASSERT(sq->sq_needexcl == 0);
3473 
3474 	if (sq->sq_ciputctrl != NULL) {
3475 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3476 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3477 		    sq->sq_nciputctrl, 0);
3478 		ASSERT(ciputctrl_cache != NULL);
3479 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3480 		sq->sq_ciputctrl = NULL;
3481 		sq->sq_nciputctrl = 0;
3482 	}
3483 
3484 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3485 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3486 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3487 
3488 	qp->q_flag &= ~QUSE;
3489 	wqp->q_flag &= ~QUSE;
3490 
3491 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3492 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3493 
3494 	qbp = qp->q_bandp;
3495 	while (qbp) {
3496 		nqbp = qbp->qb_next;
3497 		freeband(qbp);
3498 		qbp = nqbp;
3499 	}
3500 	qbp = wqp->q_bandp;
3501 	while (qbp) {
3502 		nqbp = qbp->qb_next;
3503 		freeband(qbp);
3504 		qbp = nqbp;
3505 	}
3506 	kmem_cache_free(queue_cache, qp);
3507 }
3508 
3509 /*
3510  * Allocate a qband structure.
3511  */
3512 qband_t *
3513 allocband(void)
3514 {
3515 	qband_t *qbp;
3516 
3517 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3518 	if (qbp == NULL)
3519 		return (NULL);
3520 
3521 	qbp->qb_next	= NULL;
3522 	qbp->qb_count	= 0;
3523 	qbp->qb_mblkcnt	= 0;
3524 	qbp->qb_first	= NULL;
3525 	qbp->qb_last	= NULL;
3526 	qbp->qb_flag	= 0;
3527 
3528 	return (qbp);
3529 }
3530 
3531 /*
3532  * Free a qband structure.
3533  */
3534 void
3535 freeband(qband_t *qbp)
3536 {
3537 	kmem_cache_free(qband_cache, qbp);
3538 }
3539 
3540 /*
3541  * Just like putnextctl(9F), except that allocb_wait() is used.
3542  *
3543  * Consolidation Private, and of course only callable from the stream head or
3544  * routines that may block.
3545  */
3546 int
3547 putnextctl_wait(queue_t *q, int type)
3548 {
3549 	mblk_t *bp;
3550 	int error;
3551 
3552 	if ((datamsg(type) && (type != M_DELAY)) ||
3553 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3554 		return (0);
3555 
3556 	bp->b_datap->db_type = (unsigned char)type;
3557 	putnext(q, bp);
3558 	return (1);
3559 }
3560 
3561 /*
3562  * Run any possible bufcalls.
3563  */
3564 void
3565 runbufcalls(void)
3566 {
3567 	strbufcall_t *bcp;
3568 
3569 	mutex_enter(&bcall_monitor);
3570 	mutex_enter(&strbcall_lock);
3571 
3572 	if (strbcalls.bc_head) {
3573 		size_t count;
3574 		int nevent;
3575 
3576 		/*
3577 		 * count how many events are on the list
3578 		 * now so we can check to avoid looping
3579 		 * in low memory situations
3580 		 */
3581 		nevent = 0;
3582 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3583 			nevent++;
3584 
3585 		/*
3586 		 * get estimate of available memory from kmem_avail().
3587 		 * awake all bufcall functions waiting for
3588 		 * memory whose request could be satisfied
3589 		 * by 'count' memory and let 'em fight for it.
3590 		 */
3591 		count = kmem_avail();
3592 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3593 			STRSTAT(bufcalls);
3594 			--nevent;
3595 			if (bcp->bc_size <= count) {
3596 				bcp->bc_executor = curthread;
3597 				mutex_exit(&strbcall_lock);
3598 				(*bcp->bc_func)(bcp->bc_arg);
3599 				mutex_enter(&strbcall_lock);
3600 				bcp->bc_executor = NULL;
3601 				cv_broadcast(&bcall_cv);
3602 				strbcalls.bc_head = bcp->bc_next;
3603 				kmem_free(bcp, sizeof (strbufcall_t));
3604 			} else {
3605 				/*
3606 				 * too big, try again later - note
3607 				 * that nevent was decremented above
3608 				 * so we won't retry this one on this
3609 				 * iteration of the loop
3610 				 */
3611 				if (bcp->bc_next != NULL) {
3612 					strbcalls.bc_head = bcp->bc_next;
3613 					bcp->bc_next = NULL;
3614 					strbcalls.bc_tail->bc_next = bcp;
3615 					strbcalls.bc_tail = bcp;
3616 				}
3617 			}
3618 		}
3619 		if (strbcalls.bc_head == NULL)
3620 			strbcalls.bc_tail = NULL;
3621 	}
3622 
3623 	mutex_exit(&strbcall_lock);
3624 	mutex_exit(&bcall_monitor);
3625 }
3626 
3627 
3628 /*
3629  * Actually run queue's service routine.
3630  */
3631 static void
3632 runservice(queue_t *q)
3633 {
3634 	qband_t *qbp;
3635 
3636 	ASSERT(q->q_qinfo->qi_srvp);
3637 again:
3638 	entersq(q->q_syncq, SQ_SVC);
3639 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3640 	    "runservice starts:%p", q);
3641 
3642 	if (!(q->q_flag & QWCLOSE))
3643 		(*q->q_qinfo->qi_srvp)(q);
3644 
3645 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3646 	    "runservice ends:(%p)", q);
3647 
3648 	leavesq(q->q_syncq, SQ_SVC);
3649 
3650 	mutex_enter(QLOCK(q));
3651 	if (q->q_flag & QENAB) {
3652 		q->q_flag &= ~QENAB;
3653 		mutex_exit(QLOCK(q));
3654 		goto again;
3655 	}
3656 	q->q_flag &= ~QINSERVICE;
3657 	q->q_flag &= ~QBACK;
3658 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3659 		qbp->qb_flag &= ~QB_BACK;
3660 	/*
3661 	 * Wakeup thread waiting for the service procedure
3662 	 * to be run (strclose and qdetach).
3663 	 */
3664 	cv_broadcast(&q->q_wait);
3665 
3666 	mutex_exit(QLOCK(q));
3667 }
3668 
3669 /*
3670  * Background processing of bufcalls.
3671  */
3672 void
3673 streams_bufcall_service(void)
3674 {
3675 	callb_cpr_t	cprinfo;
3676 
3677 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3678 	    "streams_bufcall_service");
3679 
3680 	mutex_enter(&strbcall_lock);
3681 
3682 	for (;;) {
3683 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3684 			mutex_exit(&strbcall_lock);
3685 			runbufcalls();
3686 			mutex_enter(&strbcall_lock);
3687 		}
3688 		if (strbcalls.bc_head != NULL) {
3689 			STRSTAT(bcwaits);
3690 			/* Wait for memory to become available */
3691 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3692 			(void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3693 			    SEC_TO_TICK(60), TR_CLOCK_TICK);
3694 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3695 		}
3696 
3697 		/* Wait for new work to arrive */
3698 		if (strbcalls.bc_head == NULL) {
3699 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3700 			cv_wait(&strbcall_cv, &strbcall_lock);
3701 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3702 		}
3703 	}
3704 }
3705 
3706 /*
3707  * Background processing of streams background tasks which failed
3708  * taskq_dispatch.
3709  */
3710 static void
3711 streams_qbkgrnd_service(void)
3712 {
3713 	callb_cpr_t cprinfo;
3714 	queue_t *q;
3715 
3716 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3717 	    "streams_bkgrnd_service");
3718 
3719 	mutex_enter(&service_queue);
3720 
3721 	for (;;) {
3722 		/*
3723 		 * Wait for work to arrive.
3724 		 */
3725 		while ((freebs_list == NULL) && (qhead == NULL)) {
3726 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3727 			cv_wait(&services_to_run, &service_queue);
3728 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3729 		}
3730 		/*
3731 		 * Handle all pending freebs requests to free memory.
3732 		 */
3733 		while (freebs_list != NULL) {
3734 			mblk_t *mp = freebs_list;
3735 			freebs_list = mp->b_next;
3736 			mutex_exit(&service_queue);
3737 			mblk_free(mp);
3738 			mutex_enter(&service_queue);
3739 		}
3740 		/*
3741 		 * Run pending queues.
3742 		 */
3743 		while (qhead != NULL) {
3744 			DQ(q, qhead, qtail, q_link);
3745 			ASSERT(q != NULL);
3746 			mutex_exit(&service_queue);
3747 			queue_service(q);
3748 			mutex_enter(&service_queue);
3749 		}
3750 		ASSERT(qhead == NULL && qtail == NULL);
3751 	}
3752 }
3753 
3754 /*
3755  * Background processing of streams background tasks which failed
3756  * taskq_dispatch.
3757  */
3758 static void
3759 streams_sqbkgrnd_service(void)
3760 {
3761 	callb_cpr_t cprinfo;
3762 	syncq_t *sq;
3763 
3764 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3765 	    "streams_sqbkgrnd_service");
3766 
3767 	mutex_enter(&service_queue);
3768 
3769 	for (;;) {
3770 		/*
3771 		 * Wait for work to arrive.
3772 		 */
3773 		while (sqhead == NULL) {
3774 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3775 			cv_wait(&syncqs_to_run, &service_queue);
3776 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3777 		}
3778 
3779 		/*
3780 		 * Run pending syncqs.
3781 		 */
3782 		while (sqhead != NULL) {
3783 			DQ(sq, sqhead, sqtail, sq_next);
3784 			ASSERT(sq != NULL);
3785 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3786 			mutex_exit(&service_queue);
3787 			syncq_service(sq);
3788 			mutex_enter(&service_queue);
3789 		}
3790 	}
3791 }
3792 
3793 /*
3794  * Disable the syncq and wait for background syncq processing to complete.
3795  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3796  * list.
3797  */
3798 void
3799 wait_sq_svc(syncq_t *sq)
3800 {
3801 	mutex_enter(SQLOCK(sq));
3802 	sq->sq_svcflags |= SQ_DISABLED;
3803 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3804 		syncq_t *sq_chase;
3805 		syncq_t *sq_curr;
3806 		int removed;
3807 
3808 		ASSERT(sq->sq_servcount == 1);
3809 		mutex_enter(&service_queue);
3810 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3811 		mutex_exit(&service_queue);
3812 		if (removed) {
3813 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3814 			sq->sq_servcount = 0;
3815 			STRSTAT(sqremoved);
3816 			goto done;
3817 		}
3818 	}
3819 	while (sq->sq_servcount != 0) {
3820 		sq->sq_flags |= SQ_WANTWAKEUP;
3821 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3822 	}
3823 done:
3824 	mutex_exit(SQLOCK(sq));
3825 }
3826 
3827 /*
3828  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3829  * Add the argument to the end of the sqhead list and set the flag
3830  * indicating this syncq has been enabled.  If it has already been
3831  * enabled, don't do anything.
3832  * This routine assumes that SQLOCK is held.
3833  * NOTE that the lock order is to have the SQLOCK first,
3834  * so if the service_syncq lock is held, we need to release it
3835  * before acquiring the SQLOCK (mostly relevant for the background
3836  * thread, and this seems to be common among the STREAMS global locks).
3837  * Note that the sq_svcflags are protected by the SQLOCK.
3838  */
3839 void
3840 sqenable(syncq_t *sq)
3841 {
3842 	/*
3843 	 * This is probably not important except for where I believe it
3844 	 * is being called.  At that point, it should be held (and it
3845 	 * is a pain to release it just for this routine, so don't do
3846 	 * it).
3847 	 */
3848 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3849 
3850 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3851 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3852 
3853 	/*
3854 	 * Do not put on list if background thread is scheduled or
3855 	 * syncq is disabled.
3856 	 */
3857 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3858 		return;
3859 
3860 	/*
3861 	 * Check whether we should enable sq at all.
3862 	 * Non PERMOD syncqs may be drained by at most one thread.
3863 	 * PERMOD syncqs may be drained by several threads but we limit the
3864 	 * total amount to the lesser of
3865 	 *	Number of queues on the squeue and
3866 	 *	Number of CPUs.
3867 	 */
3868 	if (sq->sq_servcount != 0) {
3869 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3870 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3871 			STRSTAT(sqtoomany);
3872 			return;
3873 		}
3874 	}
3875 
3876 	sq->sq_tstamp = ddi_get_lbolt();
3877 	STRSTAT(sqenables);
3878 
3879 	/* Attempt a taskq dispatch */
3880 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3881 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3882 	if (sq->sq_servid != NULL) {
3883 		sq->sq_servcount++;
3884 		return;
3885 	}
3886 
3887 	/*
3888 	 * This taskq dispatch failed, but a previous one may have succeeded.
3889 	 * Don't try to schedule on the background thread whilst there is
3890 	 * outstanding taskq processing.
3891 	 */
3892 	if (sq->sq_servcount != 0)
3893 		return;
3894 
3895 	/*
3896 	 * System is low on resources and can't perform a non-sleeping
3897 	 * dispatch. Schedule the syncq for a background thread and mark the
3898 	 * syncq to avoid any further taskq dispatch attempts.
3899 	 */
3900 	mutex_enter(&service_queue);
3901 	STRSTAT(taskqfails);
3902 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3903 	sq->sq_svcflags |= SQ_BGTHREAD;
3904 	sq->sq_servcount = 1;
3905 	cv_signal(&syncqs_to_run);
3906 	mutex_exit(&service_queue);
3907 }
3908 
3909 /*
3910  * Note: fifo_close() depends on the mblk_t on the queue being freed
3911  * asynchronously. The asynchronous freeing of messages breaks the
3912  * recursive call chain of fifo_close() while there are I_SENDFD type of
3913  * messages referring to other file pointers on the queue. Then when
3914  * closing pipes it can avoid stack overflow in case of daisy-chained
3915  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3916  * share the same fifolock_t).
3917  *
3918  * No need to kpreempt_disable to access cpu_seqid.  If we migrate and
3919  * the esb queue does not match the new CPU, that is OK.
3920  */
3921 void
3922 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3923 {
3924 	int qindex = CPU->cpu_seqid >> esbq_log2_cpus_per_q;
3925 	esb_queue_t *eqp;
3926 
3927 	ASSERT(dbp->db_mblk == mp);
3928 	ASSERT(qindex < esbq_nelem);
3929 
3930 	eqp = system_esbq_array;
3931 	if (eqp != NULL) {
3932 		eqp += qindex;
3933 	} else {
3934 		mutex_enter(&esbq_lock);
3935 		if (kmem_ready && system_esbq_array == NULL)
3936 			system_esbq_array = (esb_queue_t *)kmem_zalloc(
3937 			    esbq_nelem * sizeof (esb_queue_t), KM_NOSLEEP);
3938 		mutex_exit(&esbq_lock);
3939 		eqp = system_esbq_array;
3940 		if (eqp != NULL)
3941 			eqp += qindex;
3942 		else
3943 			eqp = &system_esbq;
3944 	}
3945 
3946 	/*
3947 	 * Check data sanity. The dblock should have non-empty free function.
3948 	 * It is better to panic here then later when the dblock is freed
3949 	 * asynchronously when the context is lost.
3950 	 */
3951 	if (dbp->db_frtnp->free_func == NULL) {
3952 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3953 		    (void *)dbp);
3954 	}
3955 
3956 	mutex_enter(&eqp->eq_lock);
3957 	/* queue the new mblk on the esballoc queue */
3958 	if (eqp->eq_head == NULL) {
3959 		eqp->eq_head = eqp->eq_tail = mp;
3960 	} else {
3961 		eqp->eq_tail->b_next = mp;
3962 		eqp->eq_tail = mp;
3963 	}
3964 	eqp->eq_len++;
3965 
3966 	/* If we're the first thread to reach the threshold, process */
3967 	if (eqp->eq_len >= esbq_max_qlen &&
3968 	    !(eqp->eq_flags & ESBQ_PROCESSING))
3969 		esballoc_process_queue(eqp);
3970 
3971 	esballoc_set_timer(eqp, esbq_timeout);
3972 	mutex_exit(&eqp->eq_lock);
3973 }
3974 
3975 static void
3976 esballoc_process_queue(esb_queue_t *eqp)
3977 {
3978 	mblk_t	*mp;
3979 
3980 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
3981 
3982 	eqp->eq_flags |= ESBQ_PROCESSING;
3983 
3984 	do {
3985 		/*
3986 		 * Detach the message chain for processing.
3987 		 */
3988 		mp = eqp->eq_head;
3989 		eqp->eq_tail->b_next = NULL;
3990 		eqp->eq_head = eqp->eq_tail = NULL;
3991 		eqp->eq_len = 0;
3992 		mutex_exit(&eqp->eq_lock);
3993 
3994 		/*
3995 		 * Process the message chain.
3996 		 */
3997 		esballoc_enqueue_mblk(mp);
3998 		mutex_enter(&eqp->eq_lock);
3999 	} while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
4000 
4001 	eqp->eq_flags &= ~ESBQ_PROCESSING;
4002 }
4003 
4004 /*
4005  * taskq callback routine to free esballoced mblk's
4006  */
4007 static void
4008 esballoc_mblk_free(mblk_t *mp)
4009 {
4010 	mblk_t	*nextmp;
4011 
4012 	for (; mp != NULL; mp = nextmp) {
4013 		nextmp = mp->b_next;
4014 		mp->b_next = NULL;
4015 		mblk_free(mp);
4016 	}
4017 }
4018 
4019 static void
4020 esballoc_enqueue_mblk(mblk_t *mp)
4021 {
4022 
4023 	if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
4024 	    TQ_NOSLEEP) == NULL) {
4025 		mblk_t *first_mp = mp;
4026 		/*
4027 		 * System is low on resources and can't perform a non-sleeping
4028 		 * dispatch. Schedule for a background thread.
4029 		 */
4030 		mutex_enter(&service_queue);
4031 		STRSTAT(taskqfails);
4032 
4033 		while (mp->b_next != NULL)
4034 			mp = mp->b_next;
4035 
4036 		mp->b_next = freebs_list;
4037 		freebs_list = first_mp;
4038 		cv_signal(&services_to_run);
4039 		mutex_exit(&service_queue);
4040 	}
4041 }
4042 
4043 static void
4044 esballoc_timer(void *arg)
4045 {
4046 	esb_queue_t *eqp = arg;
4047 
4048 	mutex_enter(&eqp->eq_lock);
4049 	eqp->eq_flags &= ~ESBQ_TIMER;
4050 
4051 	if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4052 	    eqp->eq_len > 0)
4053 		esballoc_process_queue(eqp);
4054 
4055 	esballoc_set_timer(eqp, esbq_timeout);
4056 	mutex_exit(&eqp->eq_lock);
4057 }
4058 
4059 static void
4060 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4061 {
4062 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4063 
4064 	if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4065 		(void) timeout(esballoc_timer, eqp, eq_timeout);
4066 		eqp->eq_flags |= ESBQ_TIMER;
4067 	}
4068 }
4069 
4070 /*
4071  * Setup esbq array length based upon NCPU scaled by CPUs per
4072  * queue. Use static system_esbq until kmem_ready and we can
4073  * create an array in freebs_enqueue().
4074  */
4075 void
4076 esballoc_queue_init(void)
4077 {
4078 	esbq_log2_cpus_per_q = highbit(esbq_cpus_per_q - 1);
4079 	esbq_cpus_per_q = 1 << esbq_log2_cpus_per_q;
4080 	esbq_nelem = howmany(NCPU, esbq_cpus_per_q);
4081 	system_esbq.eq_len = 0;
4082 	system_esbq.eq_head = system_esbq.eq_tail = NULL;
4083 	system_esbq.eq_flags = 0;
4084 }
4085 
4086 /*
4087  * Set the QBACK or QB_BACK flag in the given queue for
4088  * the given priority band.
4089  */
4090 void
4091 setqback(queue_t *q, unsigned char pri)
4092 {
4093 	int i;
4094 	qband_t *qbp;
4095 	qband_t **qbpp;
4096 
4097 	ASSERT(MUTEX_HELD(QLOCK(q)));
4098 	if (pri != 0) {
4099 		if (pri > q->q_nband) {
4100 			qbpp = &q->q_bandp;
4101 			while (*qbpp)
4102 				qbpp = &(*qbpp)->qb_next;
4103 			while (pri > q->q_nband) {
4104 				if ((*qbpp = allocband()) == NULL) {
4105 					cmn_err(CE_WARN,
4106 					    "setqback: can't allocate qband\n");
4107 					return;
4108 				}
4109 				(*qbpp)->qb_hiwat = q->q_hiwat;
4110 				(*qbpp)->qb_lowat = q->q_lowat;
4111 				q->q_nband++;
4112 				qbpp = &(*qbpp)->qb_next;
4113 			}
4114 		}
4115 		qbp = q->q_bandp;
4116 		i = pri;
4117 		while (--i)
4118 			qbp = qbp->qb_next;
4119 		qbp->qb_flag |= QB_BACK;
4120 	} else {
4121 		q->q_flag |= QBACK;
4122 	}
4123 }
4124 
4125 int
4126 strcopyin(void *from, void *to, size_t len, int copyflag)
4127 {
4128 	if (copyflag & U_TO_K) {
4129 		ASSERT((copyflag & K_TO_K) == 0);
4130 		if (copyin(from, to, len))
4131 			return (EFAULT);
4132 	} else {
4133 		ASSERT(copyflag & K_TO_K);
4134 		bcopy(from, to, len);
4135 	}
4136 	return (0);
4137 }
4138 
4139 int
4140 strcopyout(void *from, void *to, size_t len, int copyflag)
4141 {
4142 	if (copyflag & U_TO_K) {
4143 		if (copyout(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 /*
4153  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4154  * It assumes that the stream head lock is already held, whereas strsignal()
4155  * acquires the lock first.  This routine was created because a few callers
4156  * release the stream head lock before calling only to re-acquire it after
4157  * it returns.
4158  */
4159 void
4160 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4161 {
4162 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4163 	switch (sig) {
4164 	case SIGPOLL:
4165 		if (stp->sd_sigflags & S_MSG)
4166 			strsendsig(stp->sd_siglist, S_MSG, band, 0);
4167 		break;
4168 	default:
4169 		if (stp->sd_pgidp)
4170 			pgsignal(stp->sd_pgidp, sig);
4171 		break;
4172 	}
4173 }
4174 
4175 void
4176 strsignal(stdata_t *stp, int sig, int32_t band)
4177 {
4178 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4179 	    "strsignal:%p, %X, %X", stp, sig, band);
4180 
4181 	mutex_enter(&stp->sd_lock);
4182 	switch (sig) {
4183 	case SIGPOLL:
4184 		if (stp->sd_sigflags & S_MSG)
4185 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4186 		break;
4187 
4188 	default:
4189 		if (stp->sd_pgidp) {
4190 			pgsignal(stp->sd_pgidp, sig);
4191 		}
4192 		break;
4193 	}
4194 	mutex_exit(&stp->sd_lock);
4195 }
4196 
4197 void
4198 strhup(stdata_t *stp)
4199 {
4200 	ASSERT(mutex_owned(&stp->sd_lock));
4201 	pollwakeup(&stp->sd_pollist, POLLHUP);
4202 	if (stp->sd_sigflags & S_HANGUP)
4203 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4204 }
4205 
4206 /*
4207  * Backenable the first queue upstream from `q' with a service procedure.
4208  */
4209 void
4210 backenable(queue_t *q, uchar_t pri)
4211 {
4212 	queue_t	*nq;
4213 
4214 	/*
4215 	 * Our presence might not prevent other modules in our own
4216 	 * stream from popping/pushing since the caller of getq might not
4217 	 * have a claim on the queue (some drivers do a getq on somebody
4218 	 * else's queue - they know that the queue itself is not going away
4219 	 * but the framework has to guarantee q_next in that stream).
4220 	 */
4221 	claimstr(q);
4222 
4223 	/* Find nearest back queue with service proc */
4224 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4225 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4226 	}
4227 
4228 	if (nq) {
4229 		kthread_t *freezer;
4230 		/*
4231 		 * backenable can be called either with no locks held
4232 		 * or with the stream frozen (the latter occurs when a module
4233 		 * calls rmvq with the stream frozen). If the stream is frozen
4234 		 * by the caller the caller will hold all qlocks in the stream.
4235 		 * Note that a frozen stream doesn't freeze a mated stream,
4236 		 * so we explicitly check for that.
4237 		 */
4238 		freezer = STREAM(q)->sd_freezer;
4239 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4240 			mutex_enter(QLOCK(nq));
4241 		}
4242 #ifdef DEBUG
4243 		else {
4244 			ASSERT(frozenstr(q));
4245 			ASSERT(MUTEX_HELD(QLOCK(q)));
4246 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4247 		}
4248 #endif
4249 		setqback(nq, pri);
4250 		qenable_locked(nq);
4251 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4252 			mutex_exit(QLOCK(nq));
4253 	}
4254 	releasestr(q);
4255 }
4256 
4257 /*
4258  * Return the appropriate errno when one of flags_to_check is set
4259  * in sd_flags. Uses the exported error routines if they are set.
4260  * Will return 0 if non error is set (or if the exported error routines
4261  * do not return an error).
4262  *
4263  * If there is both a read and write error to check, we prefer the read error.
4264  * Also, give preference to recorded errno's over the error functions.
4265  * The flags that are handled are:
4266  *	STPLEX		return EINVAL
4267  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4268  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4269  *	STRHUP		return sd_werror
4270  *
4271  * If the caller indicates that the operation is a peek, a nonpersistent error
4272  * is not cleared.
4273  */
4274 int
4275 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4276 {
4277 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4278 	int error = 0;
4279 
4280 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4281 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4282 	if (sd_flag & STPLEX)
4283 		error = EINVAL;
4284 	else if (sd_flag & STRDERR) {
4285 		error = stp->sd_rerror;
4286 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4287 			/*
4288 			 * Read errors are non-persistent i.e. discarded once
4289 			 * returned to a non-peeking caller,
4290 			 */
4291 			stp->sd_rerror = 0;
4292 			stp->sd_flag &= ~STRDERR;
4293 		}
4294 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4295 			int clearerr = 0;
4296 
4297 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4298 			    &clearerr);
4299 			if (clearerr) {
4300 				stp->sd_flag &= ~STRDERR;
4301 				stp->sd_rderrfunc = NULL;
4302 			}
4303 		}
4304 	} else if (sd_flag & STWRERR) {
4305 		error = stp->sd_werror;
4306 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4307 			/*
4308 			 * Write errors are non-persistent i.e. discarded once
4309 			 * returned to a non-peeking caller,
4310 			 */
4311 			stp->sd_werror = 0;
4312 			stp->sd_flag &= ~STWRERR;
4313 		}
4314 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4315 			int clearerr = 0;
4316 
4317 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4318 			    &clearerr);
4319 			if (clearerr) {
4320 				stp->sd_flag &= ~STWRERR;
4321 				stp->sd_wrerrfunc = NULL;
4322 			}
4323 		}
4324 	} else if (sd_flag & STRHUP) {
4325 		/* sd_werror set when STRHUP */
4326 		error = stp->sd_werror;
4327 	}
4328 	return (error);
4329 }
4330 
4331 
4332 /*
4333  * Single-thread open/close/push/pop
4334  * for twisted streams also
4335  */
4336 int
4337 strstartplumb(stdata_t *stp, int flag, int cmd)
4338 {
4339 	int waited = 1;
4340 	int error = 0;
4341 
4342 	if (STRMATED(stp)) {
4343 		struct stdata *stmatep = stp->sd_mate;
4344 
4345 		STRLOCKMATES(stp);
4346 		while (waited) {
4347 			waited = 0;
4348 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4349 				if ((cmd == I_POP) &&
4350 				    (flag & (FNDELAY|FNONBLOCK))) {
4351 					STRUNLOCKMATES(stp);
4352 					return (EAGAIN);
4353 				}
4354 				waited = 1;
4355 				mutex_exit(&stp->sd_lock);
4356 				if (!cv_wait_sig(&stmatep->sd_monitor,
4357 				    &stmatep->sd_lock)) {
4358 					mutex_exit(&stmatep->sd_lock);
4359 					return (EINTR);
4360 				}
4361 				mutex_exit(&stmatep->sd_lock);
4362 				STRLOCKMATES(stp);
4363 			}
4364 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4365 				if ((cmd == I_POP) &&
4366 				    (flag & (FNDELAY|FNONBLOCK))) {
4367 					STRUNLOCKMATES(stp);
4368 					return (EAGAIN);
4369 				}
4370 				waited = 1;
4371 				mutex_exit(&stmatep->sd_lock);
4372 				if (!cv_wait_sig(&stp->sd_monitor,
4373 				    &stp->sd_lock)) {
4374 					mutex_exit(&stp->sd_lock);
4375 					return (EINTR);
4376 				}
4377 				mutex_exit(&stp->sd_lock);
4378 				STRLOCKMATES(stp);
4379 			}
4380 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4381 				error = strgeterr(stp,
4382 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4383 				if (error != 0) {
4384 					STRUNLOCKMATES(stp);
4385 					return (error);
4386 				}
4387 			}
4388 		}
4389 		stp->sd_flag |= STRPLUMB;
4390 		STRUNLOCKMATES(stp);
4391 	} else {
4392 		mutex_enter(&stp->sd_lock);
4393 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4394 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4395 			    (flag & (FNDELAY|FNONBLOCK))) {
4396 				mutex_exit(&stp->sd_lock);
4397 				return (EAGAIN);
4398 			}
4399 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4400 				mutex_exit(&stp->sd_lock);
4401 				return (EINTR);
4402 			}
4403 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4404 				error = strgeterr(stp,
4405 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4406 				if (error != 0) {
4407 					mutex_exit(&stp->sd_lock);
4408 					return (error);
4409 				}
4410 			}
4411 		}
4412 		stp->sd_flag |= STRPLUMB;
4413 		mutex_exit(&stp->sd_lock);
4414 	}
4415 	return (0);
4416 }
4417 
4418 /*
4419  * Complete the plumbing operation associated with stream `stp'.
4420  */
4421 void
4422 strendplumb(stdata_t *stp)
4423 {
4424 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4425 	ASSERT(stp->sd_flag & STRPLUMB);
4426 	stp->sd_flag &= ~STRPLUMB;
4427 	cv_broadcast(&stp->sd_monitor);
4428 }
4429 
4430 /*
4431  * This describes how the STREAMS framework handles synchronization
4432  * during open/push and close/pop.
4433  * The key interfaces for open and close are qprocson and qprocsoff,
4434  * respectively. While the close case in general is harder both open
4435  * have close have significant similarities.
4436  *
4437  * During close the STREAMS framework has to both ensure that there
4438  * are no stale references to the queue pair (and syncq) that
4439  * are being closed and also provide the guarantees that are documented
4440  * in qprocsoff(9F).
4441  * If there are stale references to the queue that is closing it can
4442  * result in kernel memory corruption or kernel panics.
4443  *
4444  * Note that is it up to the module/driver to ensure that it itself
4445  * does not have any stale references to the closing queues once its close
4446  * routine returns. This includes:
4447  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4448  *    associated with the queues. For timeout and bufcall callbacks the
4449  *    module/driver also has to ensure (or wait for) any callbacks that
4450  *    are in progress.
4451  *  - If the module/driver is using esballoc it has to ensure that any
4452  *    esballoc free functions do not refer to a queue that has closed.
4453  *    (Note that in general the close routine can not wait for the esballoc'ed
4454  *    messages to be freed since that can cause a deadlock.)
4455  *  - Cancelling any interrupts that refer to the closing queues and
4456  *    also ensuring that there are no interrupts in progress that will
4457  *    refer to the closing queues once the close routine returns.
4458  *  - For multiplexors removing any driver global state that refers to
4459  *    the closing queue and also ensuring that there are no threads in
4460  *    the multiplexor that has picked up a queue pointer but not yet
4461  *    finished using it.
4462  *
4463  * In addition, a driver/module can only reference the q_next pointer
4464  * in its open, close, put, or service procedures or in a
4465  * qtimeout/qbufcall callback procedure executing "on" the correct
4466  * stream. Thus it can not reference the q_next pointer in an interrupt
4467  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4468  * it can not reference q_next of a different queue e.g. in a mux that
4469  * passes messages from one queues put/service procedure to another queue.
4470  * In all the cases when the driver/module can not access the q_next
4471  * field it must use the *next* versions e.g. canputnext instead of
4472  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4473  *
4474  *
4475  * Assuming that the driver/module conforms to the above constraints
4476  * the STREAMS framework has to avoid stale references to q_next for all
4477  * the framework internal cases which include (but are not limited to):
4478  *  - Threads in canput/canputnext/backenable and elsewhere that are
4479  *    walking q_next.
4480  *  - Messages on a syncq that have a reference to the queue through b_queue.
4481  *  - Messages on an outer perimeter (syncq) that have a reference to the
4482  *    queue through b_queue.
4483  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4484  *    Note that only canput and bcanput use q_nfsrv without any locking.
4485  *
4486  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4487  * after qprocsoff returns, the framework has to ensure that no threads can
4488  * enter the put or service routines for the closing read or write-side queue.
4489  * In addition to preventing "direct" entry into the put procedures
4490  * the framework also has to prevent messages being drained from
4491  * the syncq or the outer perimeter.
4492  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4493  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4494  * qprocsoff has returned.
4495  * Note that if a module/driver uses put(9F) on one of its own queues
4496  * it is up to the module/driver to ensure that the put() doesn't
4497  * get called when the queue is closing.
4498  *
4499  *
4500  * The framework aspects of the above "contract" is implemented by
4501  * qprocsoff, removeq, and strlock:
4502  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4503  *    entering the service procedures.
4504  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4505  *    canputnext, backenable etc from dereferencing the q_next that will
4506  *    soon change.
4507  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4508  *    or other q_next walker that uses claimstr/releasestr to finish.
4509  *  - optionally for every syncq in the stream strlock acquires all the
4510  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4511  *    that no thread executes in the put or service procedures and that no
4512  *    thread is draining into the module/driver. This ensures that no
4513  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4514  *    currently executing hence no such thread can end up with the old stale
4515  *    q_next value and no canput/backenable can have the old stale
4516  *    q_nfsrv/q_next.
4517  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4518  *    have either finished or observed the QWCLOSE flag and gone away.
4519  */
4520 
4521 
4522 /*
4523  * Get all the locks necessary to change q_next.
4524  *
4525  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4526  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4527  * the only threads inside the syncq are threads currently calling removeq().
4528  * Since threads calling removeq() are in the process of removing their queues
4529  * from the stream, we do not need to worry about them accessing a stale q_next
4530  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4531  * for them can cause deadlock).
4532  *
4533  * This routine is subject to starvation since it does not set any flag to
4534  * prevent threads from entering a module in the stream (i.e. sq_count can
4535  * increase on some syncq while it is waiting on some other syncq).
4536  *
4537  * Assumes that only one thread attempts to call strlock for a given
4538  * stream. If this is not the case the two threads would deadlock.
4539  * This assumption is guaranteed since strlock is only called by insertq
4540  * and removeq and streams plumbing changes are single-threaded for
4541  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4542  *
4543  * For pipes, it is not difficult to atomically designate a pair of streams
4544  * to be mated. Once mated atomically by the framework the twisted pair remain
4545  * configured that way until dismantled atomically by the framework.
4546  * When plumbing takes place on a twisted stream it is necessary to ensure that
4547  * this operation is done exclusively on the twisted stream since two such
4548  * operations, each initiated on different ends of the pipe will deadlock
4549  * waiting for each other to complete.
4550  *
4551  * On entry, no locks should be held.
4552  * The locks acquired and held by strlock depends on a few factors.
4553  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4554  *   and held on exit and all sq_count are at an acceptable level.
4555  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4556  *   sd_refcnt being zero.
4557  */
4558 
4559 static void
4560 strlock(struct stdata *stp, sqlist_t *sqlist)
4561 {
4562 	syncql_t *sql, *sql2;
4563 retry:
4564 	/*
4565 	 * Wait for any claimstr to go away.
4566 	 */
4567 	if (STRMATED(stp)) {
4568 		struct stdata *stp1, *stp2;
4569 
4570 		STRLOCKMATES(stp);
4571 		/*
4572 		 * Note that the selection of locking order is not
4573 		 * important, just that they are always acquired in
4574 		 * the same order.  To assure this, we choose this
4575 		 * order based on the value of the pointer, and since
4576 		 * the pointer will not change for the life of this
4577 		 * pair, we will always grab the locks in the same
4578 		 * order (and hence, prevent deadlocks).
4579 		 */
4580 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4581 			stp1 = stp;
4582 			stp2 = stp->sd_mate;
4583 		} else {
4584 			stp2 = stp;
4585 			stp1 = stp->sd_mate;
4586 		}
4587 		mutex_enter(&stp1->sd_reflock);
4588 		if (stp1->sd_refcnt > 0) {
4589 			STRUNLOCKMATES(stp);
4590 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4591 			mutex_exit(&stp1->sd_reflock);
4592 			goto retry;
4593 		}
4594 		mutex_enter(&stp2->sd_reflock);
4595 		if (stp2->sd_refcnt > 0) {
4596 			STRUNLOCKMATES(stp);
4597 			mutex_exit(&stp1->sd_reflock);
4598 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4599 			mutex_exit(&stp2->sd_reflock);
4600 			goto retry;
4601 		}
4602 		STREAM_PUTLOCKS_ENTER(stp1);
4603 		STREAM_PUTLOCKS_ENTER(stp2);
4604 	} else {
4605 		mutex_enter(&stp->sd_lock);
4606 		mutex_enter(&stp->sd_reflock);
4607 		while (stp->sd_refcnt > 0) {
4608 			mutex_exit(&stp->sd_lock);
4609 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4610 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4611 				mutex_exit(&stp->sd_reflock);
4612 				mutex_enter(&stp->sd_lock);
4613 				mutex_enter(&stp->sd_reflock);
4614 			}
4615 		}
4616 		STREAM_PUTLOCKS_ENTER(stp);
4617 	}
4618 
4619 	if (sqlist == NULL)
4620 		return;
4621 
4622 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4623 		syncq_t *sq = sql->sql_sq;
4624 		uint16_t count;
4625 
4626 		mutex_enter(SQLOCK(sq));
4627 		count = sq->sq_count;
4628 		ASSERT(sq->sq_rmqcount <= count);
4629 		SQ_PUTLOCKS_ENTER(sq);
4630 		SUM_SQ_PUTCOUNTS(sq, count);
4631 		if (count == sq->sq_rmqcount)
4632 			continue;
4633 
4634 		/* Failed - drop all locks that we have acquired so far */
4635 		if (STRMATED(stp)) {
4636 			STREAM_PUTLOCKS_EXIT(stp);
4637 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4638 			STRUNLOCKMATES(stp);
4639 			mutex_exit(&stp->sd_reflock);
4640 			mutex_exit(&stp->sd_mate->sd_reflock);
4641 		} else {
4642 			STREAM_PUTLOCKS_EXIT(stp);
4643 			mutex_exit(&stp->sd_lock);
4644 			mutex_exit(&stp->sd_reflock);
4645 		}
4646 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4647 		    sql2 = sql2->sql_next) {
4648 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4649 			mutex_exit(SQLOCK(sql2->sql_sq));
4650 		}
4651 
4652 		/*
4653 		 * The wait loop below may starve when there are many threads
4654 		 * claiming the syncq. This is especially a problem with permod
4655 		 * syncqs (IP). To lessen the impact of the problem we increment
4656 		 * sq_needexcl and clear fastbits so that putnexts will slow
4657 		 * down and call sqenable instead of draining right away.
4658 		 */
4659 		sq->sq_needexcl++;
4660 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4661 		while (count > sq->sq_rmqcount) {
4662 			sq->sq_flags |= SQ_WANTWAKEUP;
4663 			SQ_PUTLOCKS_EXIT(sq);
4664 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4665 			count = sq->sq_count;
4666 			SQ_PUTLOCKS_ENTER(sq);
4667 			SUM_SQ_PUTCOUNTS(sq, count);
4668 		}
4669 		sq->sq_needexcl--;
4670 		if (sq->sq_needexcl == 0)
4671 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4672 		SQ_PUTLOCKS_EXIT(sq);
4673 		ASSERT(count == sq->sq_rmqcount);
4674 		mutex_exit(SQLOCK(sq));
4675 		goto retry;
4676 	}
4677 }
4678 
4679 /*
4680  * Drop all the locks that strlock acquired.
4681  */
4682 static void
4683 strunlock(struct stdata *stp, sqlist_t *sqlist)
4684 {
4685 	syncql_t *sql;
4686 
4687 	if (STRMATED(stp)) {
4688 		STREAM_PUTLOCKS_EXIT(stp);
4689 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4690 		STRUNLOCKMATES(stp);
4691 		mutex_exit(&stp->sd_reflock);
4692 		mutex_exit(&stp->sd_mate->sd_reflock);
4693 	} else {
4694 		STREAM_PUTLOCKS_EXIT(stp);
4695 		mutex_exit(&stp->sd_lock);
4696 		mutex_exit(&stp->sd_reflock);
4697 	}
4698 
4699 	if (sqlist == NULL)
4700 		return;
4701 
4702 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4703 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4704 		mutex_exit(SQLOCK(sql->sql_sq));
4705 	}
4706 }
4707 
4708 /*
4709  * When the module has service procedure, we need check if the next
4710  * module which has service procedure is in flow control to trigger
4711  * the backenable.
4712  */
4713 static void
4714 backenable_insertedq(queue_t *q)
4715 {
4716 	qband_t	*qbp;
4717 
4718 	claimstr(q);
4719 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4720 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4721 			backenable(q, 0);
4722 
4723 		qbp = q->q_next->q_nfsrv->q_bandp;
4724 		for (; qbp != NULL; qbp = qbp->qb_next)
4725 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4726 				backenable(q, qbp->qb_first->b_band);
4727 	}
4728 	releasestr(q);
4729 }
4730 
4731 /*
4732  * Given two read queues, insert a new single one after another.
4733  *
4734  * This routine acquires all the necessary locks in order to change
4735  * q_next and related pointer using strlock().
4736  * It depends on the stream head ensuring that there are no concurrent
4737  * insertq or removeq on the same stream. The stream head ensures this
4738  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4739  *
4740  * Note that no syncq locks are held during the q_next change. This is
4741  * applied to all streams since, unlike removeq, there is no problem of stale
4742  * pointers when adding a module to the stream. Thus drivers/modules that do a
4743  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4744  * applied this optimization to all streams.
4745  */
4746 void
4747 insertq(struct stdata *stp, queue_t *new)
4748 {
4749 	queue_t	*after;
4750 	queue_t *wafter;
4751 	queue_t *wnew = _WR(new);
4752 	boolean_t have_fifo = B_FALSE;
4753 
4754 	if (new->q_flag & _QINSERTING) {
4755 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4756 		after = new->q_next;
4757 		wafter = _WR(new->q_next);
4758 	} else {
4759 		after = _RD(stp->sd_wrq);
4760 		wafter = stp->sd_wrq;
4761 	}
4762 
4763 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4764 	    "insertq:%p, %p", after, new);
4765 	ASSERT(after->q_flag & QREADR);
4766 	ASSERT(new->q_flag & QREADR);
4767 
4768 	strlock(stp, NULL);
4769 
4770 	/* Do we have a FIFO? */
4771 	if (wafter->q_next == after) {
4772 		have_fifo = B_TRUE;
4773 		wnew->q_next = new;
4774 	} else {
4775 		wnew->q_next = wafter->q_next;
4776 	}
4777 	new->q_next = after;
4778 
4779 	set_nfsrv_ptr(new, wnew, after, wafter);
4780 	/*
4781 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4782 	 * so only reset this flag after calling it.
4783 	 */
4784 	new->q_flag &= ~_QINSERTING;
4785 
4786 	if (have_fifo) {
4787 		wafter->q_next = wnew;
4788 	} else {
4789 		if (wafter->q_next)
4790 			_OTHERQ(wafter->q_next)->q_next = new;
4791 		wafter->q_next = wnew;
4792 	}
4793 
4794 	set_qend(new);
4795 	/* The QEND flag might have to be updated for the upstream guy */
4796 	set_qend(after);
4797 
4798 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4799 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4800 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4801 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4802 	strsetuio(stp);
4803 
4804 	/*
4805 	 * If this was a module insertion, bump the push count.
4806 	 */
4807 	if (!(new->q_flag & QISDRV))
4808 		stp->sd_pushcnt++;
4809 
4810 	strunlock(stp, NULL);
4811 
4812 	/* check if the write Q needs backenable */
4813 	backenable_insertedq(wnew);
4814 
4815 	/* check if the read Q needs backenable */
4816 	backenable_insertedq(new);
4817 }
4818 
4819 /*
4820  * Given a read queue, unlink it from any neighbors.
4821  *
4822  * This routine acquires all the necessary locks in order to
4823  * change q_next and related pointers and also guard against
4824  * stale references (e.g. through q_next) to the queue that
4825  * is being removed. It also plays part of the role in ensuring
4826  * that the module's/driver's put procedure doesn't get called
4827  * after qprocsoff returns.
4828  *
4829  * Removeq depends on the stream head ensuring that there are
4830  * no concurrent insertq or removeq on the same stream. The
4831  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4832  * STRPLUMB.
4833  *
4834  * The set of locks needed to remove the queue is different in
4835  * different cases:
4836  *
4837  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4838  * waiting for the syncq reference count to drop to 0 indicating that no
4839  * non-close threads are present anywhere in the stream. This ensures that any
4840  * module/driver can reference q_next in its open, close, put, or service
4841  * procedures.
4842  *
4843  * The sq_rmqcount counter tracks the number of threads inside removeq().
4844  * strlock() ensures that there is either no threads executing inside perimeter
4845  * or there is only a thread calling qprocsoff().
4846  *
4847  * strlock() compares the value of sq_count with the number of threads inside
4848  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4849  * any threads waiting in strlock() when the sq_rmqcount increases.
4850  */
4851 
4852 void
4853 removeq(queue_t *qp)
4854 {
4855 	queue_t *wqp = _WR(qp);
4856 	struct stdata *stp = STREAM(qp);
4857 	sqlist_t *sqlist = NULL;
4858 	boolean_t isdriver;
4859 	int moved;
4860 	syncq_t *sq = qp->q_syncq;
4861 	syncq_t *wsq = wqp->q_syncq;
4862 
4863 	ASSERT(stp);
4864 
4865 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4866 	    "removeq:%p %p", qp, wqp);
4867 	ASSERT(qp->q_flag&QREADR);
4868 
4869 	/*
4870 	 * For queues using Synchronous streams, we must wait for all threads in
4871 	 * rwnext() to drain out before proceeding.
4872 	 */
4873 	if (qp->q_flag & QSYNCSTR) {
4874 		/* First, we need wakeup any threads blocked in rwnext() */
4875 		mutex_enter(SQLOCK(sq));
4876 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4877 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4878 			cv_broadcast(&sq->sq_wait);
4879 		}
4880 		mutex_exit(SQLOCK(sq));
4881 
4882 		if (wsq != sq) {
4883 			mutex_enter(SQLOCK(wsq));
4884 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4885 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4886 				cv_broadcast(&wsq->sq_wait);
4887 			}
4888 			mutex_exit(SQLOCK(wsq));
4889 		}
4890 
4891 		mutex_enter(QLOCK(qp));
4892 		while (qp->q_rwcnt > 0) {
4893 			qp->q_flag |= QWANTRMQSYNC;
4894 			cv_wait(&qp->q_wait, QLOCK(qp));
4895 		}
4896 		mutex_exit(QLOCK(qp));
4897 
4898 		mutex_enter(QLOCK(wqp));
4899 		while (wqp->q_rwcnt > 0) {
4900 			wqp->q_flag |= QWANTRMQSYNC;
4901 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4902 		}
4903 		mutex_exit(QLOCK(wqp));
4904 	}
4905 
4906 	mutex_enter(SQLOCK(sq));
4907 	sq->sq_rmqcount++;
4908 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4909 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4910 		cv_broadcast(&sq->sq_wait);
4911 	}
4912 	mutex_exit(SQLOCK(sq));
4913 
4914 	isdriver = (qp->q_flag & QISDRV);
4915 
4916 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4917 	strlock(stp, sqlist);
4918 
4919 	reset_nfsrv_ptr(qp, wqp);
4920 
4921 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4922 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4923 	/* Do we have a FIFO? */
4924 	if (wqp->q_next == qp) {
4925 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4926 	} else {
4927 		if (wqp->q_next)
4928 			backq(qp)->q_next = qp->q_next;
4929 		if (qp->q_next)
4930 			backq(wqp)->q_next = wqp->q_next;
4931 	}
4932 
4933 	/* The QEND flag might have to be updated for the upstream guy */
4934 	if (qp->q_next)
4935 		set_qend(qp->q_next);
4936 
4937 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4938 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4939 
4940 	/*
4941 	 * Move any messages destined for the put procedures to the next
4942 	 * syncq in line. Otherwise free them.
4943 	 */
4944 	moved = 0;
4945 	/*
4946 	 * Quick check to see whether there are any messages or events.
4947 	 */
4948 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4949 		moved += propagate_syncq(qp);
4950 	if (wqp->q_syncqmsgs != 0 ||
4951 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4952 		moved += propagate_syncq(wqp);
4953 
4954 	strsetuio(stp);
4955 
4956 	/*
4957 	 * If this was a module removal, decrement the push count.
4958 	 */
4959 	if (!isdriver)
4960 		stp->sd_pushcnt--;
4961 
4962 	strunlock(stp, sqlist);
4963 	sqlist_free(sqlist);
4964 
4965 	/*
4966 	 * Make sure any messages that were propagated are drained.
4967 	 * Also clear any QFULL bit caused by messages that were propagated.
4968 	 */
4969 
4970 	if (qp->q_next != NULL) {
4971 		clr_qfull(qp);
4972 		/*
4973 		 * For the driver calling qprocsoff, propagate_syncq
4974 		 * frees all the messages instead of putting it in
4975 		 * the stream head
4976 		 */
4977 		if (!isdriver && (moved > 0))
4978 			emptysq(qp->q_next->q_syncq);
4979 	}
4980 	if (wqp->q_next != NULL) {
4981 		clr_qfull(wqp);
4982 		/*
4983 		 * We come here for any pop of a module except for the
4984 		 * case of driver being removed. We don't call emptysq
4985 		 * if we did not move any messages. This will avoid holding
4986 		 * PERMOD syncq locks in emptysq
4987 		 */
4988 		if (moved > 0)
4989 			emptysq(wqp->q_next->q_syncq);
4990 	}
4991 
4992 	mutex_enter(SQLOCK(sq));
4993 	sq->sq_rmqcount--;
4994 	mutex_exit(SQLOCK(sq));
4995 }
4996 
4997 /*
4998  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
4999  * SQ_WRITER) on a syncq.
5000  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
5001  * sync queue and waits until sq_count reaches maxcnt.
5002  *
5003  * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
5004  * does not care about putnext threads that are in the middle of calling put
5005  * entry points.
5006  *
5007  * This routine is used for both inner and outer syncqs.
5008  */
5009 static void
5010 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
5011 {
5012 	uint16_t count = 0;
5013 
5014 	mutex_enter(SQLOCK(sq));
5015 	/*
5016 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
5017 	 * SQ_FROZEN will be set if there is a frozen stream that has a
5018 	 * queue which also refers to this "shared" syncq.
5019 	 * SQ_BLOCKED will be set if there is "off" queue which also
5020 	 * refers to this "shared" syncq.
5021 	 */
5022 	if (maxcnt != -1) {
5023 		count = sq->sq_count;
5024 		SQ_PUTLOCKS_ENTER(sq);
5025 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5026 		SUM_SQ_PUTCOUNTS(sq, count);
5027 	}
5028 	sq->sq_needexcl++;
5029 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5030 
5031 	while ((sq->sq_flags & flag) ||
5032 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
5033 		sq->sq_flags |= SQ_WANTWAKEUP;
5034 		if (maxcnt != -1) {
5035 			SQ_PUTLOCKS_EXIT(sq);
5036 		}
5037 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5038 		if (maxcnt != -1) {
5039 			count = sq->sq_count;
5040 			SQ_PUTLOCKS_ENTER(sq);
5041 			SUM_SQ_PUTCOUNTS(sq, count);
5042 		}
5043 	}
5044 	sq->sq_needexcl--;
5045 	sq->sq_flags |= flag;
5046 	ASSERT(maxcnt == -1 || count == maxcnt);
5047 	if (maxcnt != -1) {
5048 		if (sq->sq_needexcl == 0) {
5049 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5050 		}
5051 		SQ_PUTLOCKS_EXIT(sq);
5052 	} else if (sq->sq_needexcl == 0) {
5053 		SQ_PUTCOUNT_SETFAST(sq);
5054 	}
5055 
5056 	mutex_exit(SQLOCK(sq));
5057 }
5058 
5059 /*
5060  * Reset a flag that was set with blocksq.
5061  *
5062  * Can not use this routine to reset SQ_WRITER.
5063  *
5064  * If "isouter" is set then the syncq is assumed to be an outer perimeter
5065  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5066  * to handle the queued qwriter operations.
5067  *
5068  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5069  * sq_putlocks are used.
5070  */
5071 static void
5072 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5073 {
5074 	uint16_t flags;
5075 
5076 	mutex_enter(SQLOCK(sq));
5077 	ASSERT(resetflag != SQ_WRITER);
5078 	ASSERT(sq->sq_flags & resetflag);
5079 	flags = sq->sq_flags & ~resetflag;
5080 	sq->sq_flags = flags;
5081 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5082 		if (flags & SQ_WANTWAKEUP) {
5083 			flags &= ~SQ_WANTWAKEUP;
5084 			cv_broadcast(&sq->sq_wait);
5085 		}
5086 		sq->sq_flags = flags;
5087 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5088 			if (!isouter) {
5089 				/* drain_syncq drops SQLOCK */
5090 				drain_syncq(sq);
5091 				return;
5092 			}
5093 		}
5094 	}
5095 	mutex_exit(SQLOCK(sq));
5096 }
5097 
5098 /*
5099  * Reset a flag that was set with blocksq.
5100  * Does not drain the syncq. Use emptysq() for that.
5101  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5102  *
5103  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5104  * sq_putlocks are used.
5105  */
5106 static int
5107 dropsq(syncq_t *sq, uint16_t resetflag)
5108 {
5109 	uint16_t flags;
5110 
5111 	mutex_enter(SQLOCK(sq));
5112 	ASSERT(sq->sq_flags & resetflag);
5113 	flags = sq->sq_flags & ~resetflag;
5114 	if (flags & SQ_WANTWAKEUP) {
5115 		flags &= ~SQ_WANTWAKEUP;
5116 		cv_broadcast(&sq->sq_wait);
5117 	}
5118 	sq->sq_flags = flags;
5119 	mutex_exit(SQLOCK(sq));
5120 	if (flags & SQ_QUEUED)
5121 		return (1);
5122 	return (0);
5123 }
5124 
5125 /*
5126  * Empty all the messages on a syncq.
5127  *
5128  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5129  * sq_putlocks are used.
5130  */
5131 static void
5132 emptysq(syncq_t *sq)
5133 {
5134 	uint16_t flags;
5135 
5136 	mutex_enter(SQLOCK(sq));
5137 	flags = sq->sq_flags;
5138 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5139 		/*
5140 		 * To prevent potential recursive invocation of drain_syncq we
5141 		 * do not call drain_syncq if count is non-zero.
5142 		 */
5143 		if (sq->sq_count == 0) {
5144 			/* drain_syncq() drops SQLOCK */
5145 			drain_syncq(sq);
5146 			return;
5147 		} else
5148 			sqenable(sq);
5149 	}
5150 	mutex_exit(SQLOCK(sq));
5151 }
5152 
5153 /*
5154  * Ordered insert while removing duplicates.
5155  */
5156 static void
5157 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5158 {
5159 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5160 
5161 	prev_sqlpp = &sqlist->sqlist_head;
5162 	while ((sqlp = *prev_sqlpp) != NULL) {
5163 		if (sqlp->sql_sq >= sqp) {
5164 			if (sqlp->sql_sq == sqp)	/* duplicate */
5165 				return;
5166 			break;
5167 		}
5168 		prev_sqlpp = &sqlp->sql_next;
5169 	}
5170 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5171 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5172 	new_sqlp->sql_next = sqlp;
5173 	new_sqlp->sql_sq = sqp;
5174 	*prev_sqlpp = new_sqlp;
5175 }
5176 
5177 /*
5178  * Walk the write side queues until we hit either the driver
5179  * or a twist in the stream (_SAMESTR will return false in both
5180  * these cases) then turn around and walk the read side queues
5181  * back up to the stream head.
5182  */
5183 static void
5184 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5185 {
5186 	while (q != NULL) {
5187 		sqlist_insert(sqlist, q->q_syncq);
5188 
5189 		if (_SAMESTR(q))
5190 			q = q->q_next;
5191 		else if (!(q->q_flag & QREADR))
5192 			q = _RD(q);
5193 		else
5194 			q = NULL;
5195 	}
5196 }
5197 
5198 /*
5199  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5200  * associated with the "q" parameter. The resulting list is sorted in a
5201  * canonical order and is free of duplicates.
5202  * Assumes the passed queue is a _RD(q).
5203  */
5204 static sqlist_t *
5205 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5206 {
5207 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5208 
5209 	/*
5210 	 * start with the current queue/qpair
5211 	 */
5212 	ASSERT(q->q_flag & QREADR);
5213 
5214 	sqlist_insert(sqlist, q->q_syncq);
5215 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5216 
5217 	sqlist_insertall(sqlist, stp->sd_wrq);
5218 	if (do_twist)
5219 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5220 
5221 	return (sqlist);
5222 }
5223 
5224 static sqlist_t *
5225 sqlist_alloc(struct stdata *stp, int kmflag)
5226 {
5227 	size_t sqlist_size;
5228 	sqlist_t *sqlist;
5229 
5230 	/*
5231 	 * Allocate 2 syncql_t's for each pushed module. Note that
5232 	 * the sqlist_t structure already has 4 syncql_t's built in:
5233 	 * 2 for the stream head, and 2 for the driver/other stream head.
5234 	 */
5235 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5236 	    sizeof (sqlist_t);
5237 	if (STRMATED(stp))
5238 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5239 	sqlist = kmem_alloc(sqlist_size, kmflag);
5240 
5241 	sqlist->sqlist_head = NULL;
5242 	sqlist->sqlist_size = sqlist_size;
5243 	sqlist->sqlist_index = 0;
5244 
5245 	return (sqlist);
5246 }
5247 
5248 /*
5249  * Free the list created by sqlist_alloc()
5250  */
5251 static void
5252 sqlist_free(sqlist_t *sqlist)
5253 {
5254 	kmem_free(sqlist, sqlist->sqlist_size);
5255 }
5256 
5257 /*
5258  * Prevent any new entries into any syncq in this stream.
5259  * Used by freezestr.
5260  */
5261 void
5262 strblock(queue_t *q)
5263 {
5264 	struct stdata	*stp;
5265 	syncql_t	*sql;
5266 	sqlist_t	*sqlist;
5267 
5268 	q = _RD(q);
5269 
5270 	stp = STREAM(q);
5271 	ASSERT(stp != NULL);
5272 
5273 	/*
5274 	 * Get a sorted list with all the duplicates removed containing
5275 	 * all the syncqs referenced by this stream.
5276 	 */
5277 	sqlist = sqlist_build(q, stp, B_FALSE);
5278 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5279 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5280 	sqlist_free(sqlist);
5281 }
5282 
5283 /*
5284  * Release the block on new entries into this stream
5285  */
5286 void
5287 strunblock(queue_t *q)
5288 {
5289 	struct stdata	*stp;
5290 	syncql_t	*sql;
5291 	sqlist_t	*sqlist;
5292 	int		drain_needed;
5293 
5294 	q = _RD(q);
5295 
5296 	/*
5297 	 * Get a sorted list with all the duplicates removed containing
5298 	 * all the syncqs referenced by this stream.
5299 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5300 	 * starting to drain them; otherwise the draining might
5301 	 * cause a freezestr in some module on the stream (which
5302 	 * would deadlock).
5303 	 */
5304 	stp = STREAM(q);
5305 	ASSERT(stp != NULL);
5306 	sqlist = sqlist_build(q, stp, B_FALSE);
5307 	drain_needed = 0;
5308 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5309 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5310 	if (drain_needed) {
5311 		for (sql = sqlist->sqlist_head; sql != NULL;
5312 		    sql = sql->sql_next)
5313 			emptysq(sql->sql_sq);
5314 	}
5315 	sqlist_free(sqlist);
5316 }
5317 
5318 #ifdef DEBUG
5319 static int
5320 qprocsareon(queue_t *rq)
5321 {
5322 	if (rq->q_next == NULL)
5323 		return (0);
5324 	return (_WR(rq->q_next)->q_next == _WR(rq));
5325 }
5326 
5327 int
5328 qclaimed(queue_t *q)
5329 {
5330 	uint_t count;
5331 
5332 	count = q->q_syncq->sq_count;
5333 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5334 	return (count != 0);
5335 }
5336 
5337 /*
5338  * Check if anyone has frozen this stream with freezestr
5339  */
5340 int
5341 frozenstr(queue_t *q)
5342 {
5343 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5344 }
5345 #endif /* DEBUG */
5346 
5347 /*
5348  * Enter a queue.
5349  * Obsoleted interface. Should not be used.
5350  */
5351 void
5352 enterq(queue_t *q)
5353 {
5354 	entersq(q->q_syncq, SQ_CALLBACK);
5355 }
5356 
5357 void
5358 leaveq(queue_t *q)
5359 {
5360 	leavesq(q->q_syncq, SQ_CALLBACK);
5361 }
5362 
5363 /*
5364  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5365  * to check.
5366  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5367  * calls and the running of open, close and service procedures.
5368  *
5369  * If c_inner bit is set no need to grab sq_putlocks since we don't care
5370  * if other threads have entered or are entering put entry point.
5371  *
5372  * If c_inner bit is set it might have been possible to use
5373  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5374  * open/close path for IP) but since the count may need to be decremented in
5375  * qwait() we wouldn't know which counter to decrement. Currently counter is
5376  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5377  * in the future we might use curthread id bits to select the counter and this
5378  * would stay constant across routine calls.
5379  */
5380 void
5381 entersq(syncq_t *sq, int entrypoint)
5382 {
5383 	uint16_t	count = 0;
5384 	uint16_t	flags;
5385 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5386 	uint16_t	type;
5387 	uint_t		c_inner = entrypoint & SQ_CI;
5388 	uint_t		c_outer = entrypoint & SQ_CO;
5389 
5390 	/*
5391 	 * Increment ref count to keep closes out of this queue.
5392 	 */
5393 	ASSERT(sq);
5394 	ASSERT(c_inner && c_outer);
5395 	mutex_enter(SQLOCK(sq));
5396 	flags = sq->sq_flags;
5397 	type = sq->sq_type;
5398 	if (!(type & c_inner)) {
5399 		/* Make sure all putcounts now use slowlock. */
5400 		count = sq->sq_count;
5401 		SQ_PUTLOCKS_ENTER(sq);
5402 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5403 		SUM_SQ_PUTCOUNTS(sq, count);
5404 		sq->sq_needexcl++;
5405 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5406 		waitflags |= SQ_MESSAGES;
5407 	}
5408 	/*
5409 	 * Wait until we can enter the inner perimeter.
5410 	 * If we want exclusive access we wait until sq_count is 0.
5411 	 * We have to do this before entering the outer perimeter in order
5412 	 * to preserve put/close message ordering.
5413 	 */
5414 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5415 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5416 		if (!(type & c_inner)) {
5417 			SQ_PUTLOCKS_EXIT(sq);
5418 		}
5419 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5420 		if (!(type & c_inner)) {
5421 			count = sq->sq_count;
5422 			SQ_PUTLOCKS_ENTER(sq);
5423 			SUM_SQ_PUTCOUNTS(sq, count);
5424 		}
5425 		flags = sq->sq_flags;
5426 	}
5427 
5428 	if (!(type & c_inner)) {
5429 		ASSERT(sq->sq_needexcl > 0);
5430 		sq->sq_needexcl--;
5431 		if (sq->sq_needexcl == 0) {
5432 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5433 		}
5434 	}
5435 
5436 	/* Check if we need to enter the outer perimeter */
5437 	if (!(type & c_outer)) {
5438 		/*
5439 		 * We have to enter the outer perimeter exclusively before
5440 		 * we can increment sq_count to avoid deadlock. This implies
5441 		 * that we have to re-check sq_flags and sq_count.
5442 		 *
5443 		 * is it possible to have c_inner set when c_outer is not set?
5444 		 */
5445 		if (!(type & c_inner)) {
5446 			SQ_PUTLOCKS_EXIT(sq);
5447 		}
5448 		mutex_exit(SQLOCK(sq));
5449 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5450 		mutex_enter(SQLOCK(sq));
5451 		flags = sq->sq_flags;
5452 		/*
5453 		 * there should be no need to recheck sq_putcounts
5454 		 * because outer_enter() has already waited for them to clear
5455 		 * after setting SQ_WRITER.
5456 		 */
5457 		count = sq->sq_count;
5458 #ifdef DEBUG
5459 		/*
5460 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5461 		 * of doing an ASSERT internally. Others should do
5462 		 * something like
5463 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5464 		 * without the need to #ifdef DEBUG it.
5465 		 */
5466 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5467 #endif
5468 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5469 		    (!(type & c_inner) && count != 0)) {
5470 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5471 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5472 			count = sq->sq_count;
5473 			flags = sq->sq_flags;
5474 		}
5475 	}
5476 
5477 	sq->sq_count++;
5478 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5479 	if (!(type & c_inner)) {
5480 		/* Exclusive entry */
5481 		ASSERT(sq->sq_count == 1);
5482 		sq->sq_flags |= SQ_EXCL;
5483 		if (type & c_outer) {
5484 			SQ_PUTLOCKS_EXIT(sq);
5485 		}
5486 	}
5487 	mutex_exit(SQLOCK(sq));
5488 }
5489 
5490 /*
5491  * Leave a syncq. Announce to framework that closes may proceed.
5492  * c_inner and c_outer specify which concurrency bits to check.
5493  *
5494  * Must never be called from driver or module put entry point.
5495  *
5496  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5497  * sq_putlocks are used.
5498  */
5499 void
5500 leavesq(syncq_t *sq, int entrypoint)
5501 {
5502 	uint16_t	flags;
5503 	uint16_t	type;
5504 	uint_t		c_outer = entrypoint & SQ_CO;
5505 #ifdef DEBUG
5506 	uint_t		c_inner = entrypoint & SQ_CI;
5507 #endif
5508 
5509 	/*
5510 	 * Decrement ref count, drain the syncq if possible, and wake up
5511 	 * any waiting close.
5512 	 */
5513 	ASSERT(sq);
5514 	ASSERT(c_inner && c_outer);
5515 	mutex_enter(SQLOCK(sq));
5516 	flags = sq->sq_flags;
5517 	type = sq->sq_type;
5518 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5519 
5520 		if (flags & SQ_WANTWAKEUP) {
5521 			flags &= ~SQ_WANTWAKEUP;
5522 			cv_broadcast(&sq->sq_wait);
5523 		}
5524 		if (flags & SQ_WANTEXWAKEUP) {
5525 			flags &= ~SQ_WANTEXWAKEUP;
5526 			cv_broadcast(&sq->sq_exitwait);
5527 		}
5528 
5529 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5530 			/*
5531 			 * The syncq needs to be drained. "Exit" the syncq
5532 			 * before calling drain_syncq.
5533 			 */
5534 			ASSERT(sq->sq_count != 0);
5535 			sq->sq_count--;
5536 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5537 			sq->sq_flags = flags & ~SQ_EXCL;
5538 			drain_syncq(sq);
5539 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5540 			/* Check if we need to exit the outer perimeter */
5541 			/* XXX will this ever be true? */
5542 			if (!(type & c_outer))
5543 				outer_exit(sq->sq_outer);
5544 			return;
5545 		}
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 	mutex_exit(SQLOCK(sq));
5552 
5553 	/* Check if we need to exit the outer perimeter */
5554 	if (!(sq->sq_type & c_outer))
5555 		outer_exit(sq->sq_outer);
5556 }
5557 
5558 /*
5559  * Prevent q_next from changing in this stream by incrementing sq_count.
5560  *
5561  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5562  * sq_putlocks are used.
5563  */
5564 void
5565 claimq(queue_t *qp)
5566 {
5567 	syncq_t	*sq = qp->q_syncq;
5568 
5569 	mutex_enter(SQLOCK(sq));
5570 	sq->sq_count++;
5571 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5572 	mutex_exit(SQLOCK(sq));
5573 }
5574 
5575 /*
5576  * Undo claimq.
5577  *
5578  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5579  * sq_putlocks are used.
5580  */
5581 void
5582 releaseq(queue_t *qp)
5583 {
5584 	syncq_t	*sq = qp->q_syncq;
5585 	uint16_t flags;
5586 
5587 	mutex_enter(SQLOCK(sq));
5588 	ASSERT(sq->sq_count > 0);
5589 	sq->sq_count--;
5590 
5591 	flags = sq->sq_flags;
5592 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5593 		if (flags & SQ_WANTWAKEUP) {
5594 			flags &= ~SQ_WANTWAKEUP;
5595 			cv_broadcast(&sq->sq_wait);
5596 		}
5597 		sq->sq_flags = flags;
5598 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5599 			/*
5600 			 * To prevent potential recursive invocation of
5601 			 * drain_syncq we do not call drain_syncq if count is
5602 			 * non-zero.
5603 			 */
5604 			if (sq->sq_count == 0) {
5605 				drain_syncq(sq);
5606 				return;
5607 			} else
5608 				sqenable(sq);
5609 		}
5610 	}
5611 	mutex_exit(SQLOCK(sq));
5612 }
5613 
5614 /*
5615  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5616  */
5617 void
5618 claimstr(queue_t *qp)
5619 {
5620 	struct stdata *stp = STREAM(qp);
5621 
5622 	mutex_enter(&stp->sd_reflock);
5623 	stp->sd_refcnt++;
5624 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5625 	mutex_exit(&stp->sd_reflock);
5626 }
5627 
5628 /*
5629  * Undo claimstr.
5630  */
5631 void
5632 releasestr(queue_t *qp)
5633 {
5634 	struct stdata *stp = STREAM(qp);
5635 
5636 	mutex_enter(&stp->sd_reflock);
5637 	ASSERT(stp->sd_refcnt != 0);
5638 	if (--stp->sd_refcnt == 0)
5639 		cv_broadcast(&stp->sd_refmonitor);
5640 	mutex_exit(&stp->sd_reflock);
5641 }
5642 
5643 static syncq_t *
5644 new_syncq(void)
5645 {
5646 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5647 }
5648 
5649 static void
5650 free_syncq(syncq_t *sq)
5651 {
5652 	ASSERT(sq->sq_head == NULL);
5653 	ASSERT(sq->sq_outer == NULL);
5654 	ASSERT(sq->sq_callbpend == NULL);
5655 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5656 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5657 
5658 	if (sq->sq_ciputctrl != NULL) {
5659 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5660 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5661 		    sq->sq_nciputctrl, 0);
5662 		ASSERT(ciputctrl_cache != NULL);
5663 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5664 	}
5665 
5666 	sq->sq_tail = NULL;
5667 	sq->sq_evhead = NULL;
5668 	sq->sq_evtail = NULL;
5669 	sq->sq_ciputctrl = NULL;
5670 	sq->sq_nciputctrl = 0;
5671 	sq->sq_count = 0;
5672 	sq->sq_rmqcount = 0;
5673 	sq->sq_callbflags = 0;
5674 	sq->sq_cancelid = 0;
5675 	sq->sq_next = NULL;
5676 	sq->sq_needexcl = 0;
5677 	sq->sq_svcflags = 0;
5678 	sq->sq_nqueues = 0;
5679 	sq->sq_pri = 0;
5680 	sq->sq_onext = NULL;
5681 	sq->sq_oprev = NULL;
5682 	sq->sq_flags = 0;
5683 	sq->sq_type = 0;
5684 	sq->sq_servcount = 0;
5685 
5686 	kmem_cache_free(syncq_cache, sq);
5687 }
5688 
5689 /* Outer perimeter code */
5690 
5691 /*
5692  * The outer syncq uses the fields and flags in the syncq slightly
5693  * differently from the inner syncqs.
5694  *	sq_count	Incremented when there are pending or running
5695  *			writers at the outer perimeter to prevent the set of
5696  *			inner syncqs that belong to the outer perimeter from
5697  *			changing.
5698  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5699  *
5700  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5701  *			inner syncqs are added to or removed from the
5702  *			outer perimeter.
5703  *	SQ_QUEUED	sq_head/tail has messages or events queued.
5704  *
5705  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5706  *			setting the SQ_WRITER flag.
5707  */
5708 
5709 /*
5710  * Get write access at the outer perimeter.
5711  * Note that read access is done by entersq, putnext, and put by simply
5712  * incrementing sq_count in the inner syncq.
5713  *
5714  * Waits until "flags" is no longer set in the outer to prevent multiple
5715  * threads from having write access at the same time. SQ_WRITER has to be part
5716  * of "flags".
5717  *
5718  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5719  * until the outer_exit is finished.
5720  *
5721  * outer_enter is vulnerable to starvation since it does not prevent new
5722  * threads from entering the inner syncqs while it is waiting for sq_count to
5723  * go to zero.
5724  */
5725 void
5726 outer_enter(syncq_t *outer, uint16_t flags)
5727 {
5728 	syncq_t	*sq;
5729 	int	wait_needed;
5730 	uint16_t	count;
5731 
5732 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5733 	    outer->sq_oprev != NULL);
5734 	ASSERT(flags & SQ_WRITER);
5735 
5736 retry:
5737 	mutex_enter(SQLOCK(outer));
5738 	while (outer->sq_flags & flags) {
5739 		outer->sq_flags |= SQ_WANTWAKEUP;
5740 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5741 	}
5742 
5743 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5744 	outer->sq_flags |= SQ_WRITER;
5745 	outer->sq_count++;
5746 	ASSERT(outer->sq_count != 0);	/* wraparound */
5747 	wait_needed = 0;
5748 	/*
5749 	 * Set SQ_WRITER on all the inner syncqs while holding
5750 	 * the SQLOCK on the outer syncq. This ensures that the changing
5751 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5752 	 */
5753 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5754 		mutex_enter(SQLOCK(sq));
5755 		count = sq->sq_count;
5756 		SQ_PUTLOCKS_ENTER(sq);
5757 		sq->sq_flags |= SQ_WRITER;
5758 		SUM_SQ_PUTCOUNTS(sq, count);
5759 		if (count != 0)
5760 			wait_needed = 1;
5761 		SQ_PUTLOCKS_EXIT(sq);
5762 		mutex_exit(SQLOCK(sq));
5763 	}
5764 	mutex_exit(SQLOCK(outer));
5765 
5766 	/*
5767 	 * Get everybody out of the syncqs sequentially.
5768 	 * Note that we don't actually need to acquire the PUTLOCKS, since
5769 	 * we have already cleared the fastbit, and set QWRITER.  By
5770 	 * definition, the count can not increase since putnext will
5771 	 * take the slowlock path (and the purpose of acquiring the
5772 	 * putlocks was to make sure it didn't increase while we were
5773 	 * waiting).
5774 	 *
5775 	 * Note that we still acquire the PUTLOCKS to be safe.
5776 	 */
5777 	if (wait_needed) {
5778 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5779 			mutex_enter(SQLOCK(sq));
5780 			count = sq->sq_count;
5781 			SQ_PUTLOCKS_ENTER(sq);
5782 			SUM_SQ_PUTCOUNTS(sq, count);
5783 			while (count != 0) {
5784 				sq->sq_flags |= SQ_WANTWAKEUP;
5785 				SQ_PUTLOCKS_EXIT(sq);
5786 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5787 				count = sq->sq_count;
5788 				SQ_PUTLOCKS_ENTER(sq);
5789 				SUM_SQ_PUTCOUNTS(sq, count);
5790 			}
5791 			SQ_PUTLOCKS_EXIT(sq);
5792 			mutex_exit(SQLOCK(sq));
5793 		}
5794 		/*
5795 		 * Verify that none of the flags got set while we
5796 		 * were waiting for the sq_counts to drop.
5797 		 * If this happens we exit and retry entering the
5798 		 * outer perimeter.
5799 		 */
5800 		mutex_enter(SQLOCK(outer));
5801 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5802 			mutex_exit(SQLOCK(outer));
5803 			outer_exit(outer);
5804 			goto retry;
5805 		}
5806 		mutex_exit(SQLOCK(outer));
5807 	}
5808 }
5809 
5810 /*
5811  * Drop the write access at the outer perimeter.
5812  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5813  * decrementing sq_count.
5814  */
5815 void
5816 outer_exit(syncq_t *outer)
5817 {
5818 	syncq_t	*sq;
5819 	int	 drain_needed;
5820 	uint16_t flags;
5821 
5822 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5823 	    outer->sq_oprev != NULL);
5824 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5825 
5826 	/*
5827 	 * Atomically (from the perspective of threads calling become_writer)
5828 	 * drop the write access at the outer perimeter by holding
5829 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5830 	 * SQ_WRITER.
5831 	 * This defines a locking order between the outer perimeter
5832 	 * SQLOCK and the inner perimeter SQLOCKs.
5833 	 */
5834 	mutex_enter(SQLOCK(outer));
5835 	flags = outer->sq_flags;
5836 	ASSERT(outer->sq_flags & SQ_WRITER);
5837 	if (flags & SQ_QUEUED) {
5838 		write_now(outer);
5839 		flags = outer->sq_flags;
5840 	}
5841 
5842 	/*
5843 	 * sq_onext is stable since sq_count has not yet been decreased.
5844 	 * Reset the SQ_WRITER flags in all syncqs.
5845 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5846 	 * inner syncqs.
5847 	 */
5848 	drain_needed = 0;
5849 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5850 		drain_needed += dropsq(sq, SQ_WRITER);
5851 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5852 	flags &= ~SQ_WRITER;
5853 	if (drain_needed) {
5854 		outer->sq_flags = flags;
5855 		mutex_exit(SQLOCK(outer));
5856 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5857 			emptysq(sq);
5858 		mutex_enter(SQLOCK(outer));
5859 		flags = outer->sq_flags;
5860 	}
5861 	if (flags & SQ_WANTWAKEUP) {
5862 		flags &= ~SQ_WANTWAKEUP;
5863 		cv_broadcast(&outer->sq_wait);
5864 	}
5865 	outer->sq_flags = flags;
5866 	ASSERT(outer->sq_count > 0);
5867 	outer->sq_count--;
5868 	mutex_exit(SQLOCK(outer));
5869 }
5870 
5871 /*
5872  * Add another syncq to an outer perimeter.
5873  * Block out all other access to the outer perimeter while it is being
5874  * changed using blocksq.
5875  * Assumes that the caller has *not* done an outer_enter.
5876  *
5877  * Vulnerable to starvation in blocksq.
5878  */
5879 static void
5880 outer_insert(syncq_t *outer, syncq_t *sq)
5881 {
5882 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5883 	    outer->sq_oprev != NULL);
5884 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5885 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5886 
5887 	/* Get exclusive access to the outer perimeter list */
5888 	blocksq(outer, SQ_BLOCKED, 0);
5889 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5890 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5891 
5892 	mutex_enter(SQLOCK(sq));
5893 	sq->sq_outer = outer;
5894 	outer->sq_onext->sq_oprev = sq;
5895 	sq->sq_onext = outer->sq_onext;
5896 	outer->sq_onext = sq;
5897 	sq->sq_oprev = outer;
5898 	mutex_exit(SQLOCK(sq));
5899 	unblocksq(outer, SQ_BLOCKED, 1);
5900 }
5901 
5902 /*
5903  * Remove a syncq from an outer perimeter.
5904  * Block out all other access to the outer perimeter while it is being
5905  * changed using blocksq.
5906  * Assumes that the caller has *not* done an outer_enter.
5907  *
5908  * Vulnerable to starvation in blocksq.
5909  */
5910 static void
5911 outer_remove(syncq_t *outer, syncq_t *sq)
5912 {
5913 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5914 	    outer->sq_oprev != NULL);
5915 	ASSERT(sq->sq_outer == outer);
5916 
5917 	/* Get exclusive access to the outer perimeter list */
5918 	blocksq(outer, SQ_BLOCKED, 0);
5919 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5920 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5921 
5922 	mutex_enter(SQLOCK(sq));
5923 	sq->sq_outer = NULL;
5924 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5925 	sq->sq_oprev->sq_onext = sq->sq_onext;
5926 	sq->sq_oprev = sq->sq_onext = NULL;
5927 	mutex_exit(SQLOCK(sq));
5928 	unblocksq(outer, SQ_BLOCKED, 1);
5929 }
5930 
5931 /*
5932  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5933  * If this is the first callback for this outer perimeter then add
5934  * this outer perimeter to the list of outer perimeters that
5935  * the qwriter_outer_thread will process.
5936  *
5937  * Increments sq_count in the outer syncq to prevent the membership
5938  * of the outer perimeter (in terms of inner syncqs) to change while
5939  * the callback is pending.
5940  */
5941 static void
5942 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5943 {
5944 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5945 
5946 	mp->b_prev = (mblk_t *)func;
5947 	mp->b_queue = q;
5948 	mp->b_next = NULL;
5949 	outer->sq_count++;	/* Decremented when dequeued */
5950 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5951 	if (outer->sq_evhead == NULL) {
5952 		/* First message. */
5953 		outer->sq_evhead = outer->sq_evtail = mp;
5954 		outer->sq_flags |= SQ_EVENTS;
5955 		mutex_exit(SQLOCK(outer));
5956 		STRSTAT(qwr_outer);
5957 		(void) taskq_dispatch(streams_taskq,
5958 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5959 	} else {
5960 		ASSERT(outer->sq_flags & SQ_EVENTS);
5961 		outer->sq_evtail->b_next = mp;
5962 		outer->sq_evtail = mp;
5963 		mutex_exit(SQLOCK(outer));
5964 	}
5965 }
5966 
5967 /*
5968  * Try and upgrade to write access at the outer perimeter. If this can
5969  * not be done without blocking then queue the callback to be done
5970  * by the qwriter_outer_thread.
5971  *
5972  * This routine can only be called from put or service procedures plus
5973  * asynchronous callback routines that have properly entered the queue (with
5974  * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
5975  * associated with q.
5976  */
5977 void
5978 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5979 {
5980 	syncq_t	*osq, *sq, *outer;
5981 	int	failed;
5982 	uint16_t flags;
5983 
5984 	osq = q->q_syncq;
5985 	outer = osq->sq_outer;
5986 	if (outer == NULL)
5987 		panic("qwriter(PERIM_OUTER): no outer perimeter");
5988 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5989 	    outer->sq_oprev != NULL);
5990 
5991 	mutex_enter(SQLOCK(outer));
5992 	flags = outer->sq_flags;
5993 	/*
5994 	 * If some thread is traversing sq_next, or if we are blocked by
5995 	 * outer_insert or outer_remove, or if the we already have queued
5996 	 * callbacks, then queue this callback for later processing.
5997 	 *
5998 	 * Also queue the qwriter for an interrupt thread in order
5999 	 * to reduce the time spent running at high IPL.
6000 	 * to identify there are events.
6001 	 */
6002 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
6003 		/*
6004 		 * Queue the become_writer request.
6005 		 * The queueing is atomic under SQLOCK(outer) in order
6006 		 * to synchronize with outer_exit.
6007 		 * queue_writer will drop the outer SQLOCK
6008 		 */
6009 		if (flags & SQ_BLOCKED) {
6010 			/* Must set SQ_WRITER on inner perimeter */
6011 			mutex_enter(SQLOCK(osq));
6012 			osq->sq_flags |= SQ_WRITER;
6013 			mutex_exit(SQLOCK(osq));
6014 		} else {
6015 			if (!(flags & SQ_WRITER)) {
6016 				/*
6017 				 * The outer could have been SQ_BLOCKED thus
6018 				 * SQ_WRITER might not be set on the inner.
6019 				 */
6020 				mutex_enter(SQLOCK(osq));
6021 				osq->sq_flags |= SQ_WRITER;
6022 				mutex_exit(SQLOCK(osq));
6023 			}
6024 			ASSERT(osq->sq_flags & SQ_WRITER);
6025 		}
6026 		queue_writer(outer, func, q, mp);
6027 		return;
6028 	}
6029 	/*
6030 	 * We are half-way to exclusive access to the outer perimeter.
6031 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
6032 	 * while the inner syncqs are traversed.
6033 	 */
6034 	outer->sq_count++;
6035 	ASSERT(outer->sq_count != 0);	/* wraparound */
6036 	flags |= SQ_WRITER;
6037 	/*
6038 	 * Check if we can run the function immediately. Mark all
6039 	 * syncqs with the writer flag to prevent new entries into
6040 	 * put and service procedures.
6041 	 *
6042 	 * Set SQ_WRITER on all the inner syncqs while holding
6043 	 * the SQLOCK on the outer syncq. This ensures that the changing
6044 	 * of SQ_WRITER is atomic under the outer SQLOCK.
6045 	 */
6046 	failed = 0;
6047 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6048 		uint16_t count;
6049 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
6050 
6051 		mutex_enter(SQLOCK(sq));
6052 		count = sq->sq_count;
6053 		SQ_PUTLOCKS_ENTER(sq);
6054 		SUM_SQ_PUTCOUNTS(sq, count);
6055 		if (sq->sq_count > maxcnt)
6056 			failed = 1;
6057 		sq->sq_flags |= SQ_WRITER;
6058 		SQ_PUTLOCKS_EXIT(sq);
6059 		mutex_exit(SQLOCK(sq));
6060 	}
6061 	if (failed) {
6062 		/*
6063 		 * Some other thread has a read claim on the outer perimeter.
6064 		 * Queue the callback for deferred processing.
6065 		 *
6066 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6067 		 * so that other qwriter(OUTER) calls will queue their
6068 		 * callbacks as well. queue_writer increments sq_count so we
6069 		 * decrement to compensate for the our increment.
6070 		 *
6071 		 * Dropping SQ_WRITER enables the writer thread to work
6072 		 * on this outer perimeter.
6073 		 */
6074 		outer->sq_flags = flags;
6075 		queue_writer(outer, func, q, mp);
6076 		/* queue_writer dropper the lock */
6077 		mutex_enter(SQLOCK(outer));
6078 		ASSERT(outer->sq_count > 0);
6079 		outer->sq_count--;
6080 		ASSERT(outer->sq_flags & SQ_WRITER);
6081 		flags = outer->sq_flags;
6082 		flags &= ~SQ_WRITER;
6083 		if (flags & SQ_WANTWAKEUP) {
6084 			flags &= ~SQ_WANTWAKEUP;
6085 			cv_broadcast(&outer->sq_wait);
6086 		}
6087 		outer->sq_flags = flags;
6088 		mutex_exit(SQLOCK(outer));
6089 		return;
6090 	} else {
6091 		outer->sq_flags = flags;
6092 		mutex_exit(SQLOCK(outer));
6093 	}
6094 
6095 	/* Can run it immediately */
6096 	(*func)(q, mp);
6097 
6098 	outer_exit(outer);
6099 }
6100 
6101 /*
6102  * Dequeue all writer callbacks from the outer perimeter and run them.
6103  */
6104 static void
6105 write_now(syncq_t *outer)
6106 {
6107 	mblk_t		*mp;
6108 	queue_t		*q;
6109 	void	(*func)();
6110 
6111 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6112 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6113 	    outer->sq_oprev != NULL);
6114 	while ((mp = outer->sq_evhead) != NULL) {
6115 		/*
6116 		 * queues cannot be placed on the queuelist on the outer
6117 		 * perimeter.
6118 		 */
6119 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6120 		ASSERT((outer->sq_flags & SQ_EVENTS));
6121 
6122 		outer->sq_evhead = mp->b_next;
6123 		if (outer->sq_evhead == NULL) {
6124 			outer->sq_evtail = NULL;
6125 			outer->sq_flags &= ~SQ_EVENTS;
6126 		}
6127 		ASSERT(outer->sq_count != 0);
6128 		outer->sq_count--;	/* Incremented when enqueued. */
6129 		mutex_exit(SQLOCK(outer));
6130 		/*
6131 		 * Drop the message if the queue is closing.
6132 		 * Make sure that the queue is "claimed" when the callback
6133 		 * is run in order to satisfy various ASSERTs.
6134 		 */
6135 		q = mp->b_queue;
6136 		func = (void (*)())mp->b_prev;
6137 		ASSERT(func != NULL);
6138 		mp->b_next = mp->b_prev = NULL;
6139 		if (q->q_flag & QWCLOSE) {
6140 			freemsg(mp);
6141 		} else {
6142 			claimq(q);
6143 			(*func)(q, mp);
6144 			releaseq(q);
6145 		}
6146 		mutex_enter(SQLOCK(outer));
6147 	}
6148 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6149 }
6150 
6151 /*
6152  * The list of messages on the inner syncq is effectively hashed
6153  * by destination queue.  These destination queues are doubly
6154  * linked lists (hopefully) in priority order.  Messages are then
6155  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6156  * Additional messages are linked together by the b_next/b_prev
6157  * elements in the mblk, with (similar to putq()) the first message
6158  * having a NULL b_prev and the last message having a NULL b_next.
6159  *
6160  * Events, such as qwriter callbacks, are put onto a list in FIFO
6161  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6162  * linked list, and messages here MUST be processed in the order queued.
6163  */
6164 
6165 /*
6166  * Run the events on the syncq event list (sq_evhead).
6167  * Assumes there is only one claim on the syncq, it is
6168  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6169  * Messages here are processed in order, with the SQ_EXCL bit
6170  * held all the way through till the last message is processed.
6171  */
6172 void
6173 sq_run_events(syncq_t *sq)
6174 {
6175 	mblk_t		*bp;
6176 	queue_t		*qp;
6177 	uint16_t	flags = sq->sq_flags;
6178 	void		(*func)();
6179 
6180 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6181 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6182 	    sq->sq_oprev == NULL) ||
6183 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6184 	    sq->sq_oprev != NULL));
6185 
6186 	ASSERT(flags & SQ_EXCL);
6187 	ASSERT(sq->sq_count == 1);
6188 
6189 	/*
6190 	 * We need to process all of the events on this list.  It
6191 	 * is possible that new events will be added while we are
6192 	 * away processing a callback, so on every loop, we start
6193 	 * back at the beginning of the list.
6194 	 */
6195 	/*
6196 	 * We have to reaccess sq_evhead since there is a
6197 	 * possibility of a new entry while we were running
6198 	 * the callback.
6199 	 */
6200 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6201 		ASSERT(bp->b_queue->q_syncq == sq);
6202 		ASSERT(sq->sq_flags & SQ_EVENTS);
6203 
6204 		qp = bp->b_queue;
6205 		func = (void (*)())bp->b_prev;
6206 		ASSERT(func != NULL);
6207 
6208 		/*
6209 		 * Messages from the event queue must be taken off in
6210 		 * FIFO order.
6211 		 */
6212 		ASSERT(sq->sq_evhead == bp);
6213 		sq->sq_evhead = bp->b_next;
6214 
6215 		if (bp->b_next == NULL) {
6216 			/* Deleting last */
6217 			ASSERT(sq->sq_evtail == bp);
6218 			sq->sq_evtail = NULL;
6219 			sq->sq_flags &= ~SQ_EVENTS;
6220 		}
6221 		bp->b_prev = bp->b_next = NULL;
6222 		ASSERT(bp->b_datap->db_ref != 0);
6223 
6224 		mutex_exit(SQLOCK(sq));
6225 
6226 		(*func)(qp, bp);
6227 
6228 		mutex_enter(SQLOCK(sq));
6229 		/*
6230 		 * re-read the flags, since they could have changed.
6231 		 */
6232 		flags = sq->sq_flags;
6233 		ASSERT(flags & SQ_EXCL);
6234 	}
6235 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6236 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6237 
6238 	if (flags & SQ_WANTWAKEUP) {
6239 		flags &= ~SQ_WANTWAKEUP;
6240 		cv_broadcast(&sq->sq_wait);
6241 	}
6242 	if (flags & SQ_WANTEXWAKEUP) {
6243 		flags &= ~SQ_WANTEXWAKEUP;
6244 		cv_broadcast(&sq->sq_exitwait);
6245 	}
6246 	sq->sq_flags = flags;
6247 }
6248 
6249 /*
6250  * Put messages on the event list.
6251  * If we can go exclusive now, do so and process the event list, otherwise
6252  * let the last claim service this list (or wake the sqthread).
6253  * This procedure assumes SQLOCK is held.  To run the event list, it
6254  * must be called with no claims.
6255  */
6256 static void
6257 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6258 {
6259 	uint16_t count;
6260 
6261 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6262 	ASSERT(func != NULL);
6263 
6264 	/*
6265 	 * This is a callback.  Add it to the list of callbacks
6266 	 * and see about upgrading.
6267 	 */
6268 	mp->b_prev = (mblk_t *)func;
6269 	mp->b_queue = q;
6270 	mp->b_next = NULL;
6271 	if (sq->sq_evhead == NULL) {
6272 		sq->sq_evhead = sq->sq_evtail = mp;
6273 		sq->sq_flags |= SQ_EVENTS;
6274 	} else {
6275 		ASSERT(sq->sq_evtail != NULL);
6276 		ASSERT(sq->sq_evtail->b_next == NULL);
6277 		ASSERT(sq->sq_flags & SQ_EVENTS);
6278 		sq->sq_evtail->b_next = mp;
6279 		sq->sq_evtail = mp;
6280 	}
6281 	/*
6282 	 * We have set SQ_EVENTS, so threads will have to
6283 	 * unwind out of the perimeter, and new entries will
6284 	 * not grab a putlock.  But we still need to know
6285 	 * how many threads have already made a claim to the
6286 	 * syncq, so grab the putlocks, and sum the counts.
6287 	 * If there are no claims on the syncq, we can upgrade
6288 	 * to exclusive, and run the event list.
6289 	 * NOTE: We hold the SQLOCK, so we can just grab the
6290 	 * putlocks.
6291 	 */
6292 	count = sq->sq_count;
6293 	SQ_PUTLOCKS_ENTER(sq);
6294 	SUM_SQ_PUTCOUNTS(sq, count);
6295 	/*
6296 	 * We have no claim, so we need to check if there
6297 	 * are no others, then we can upgrade.
6298 	 */
6299 	/*
6300 	 * There are currently no claims on
6301 	 * the syncq by this thread (at least on this entry). The thread who has
6302 	 * the claim should drain syncq.
6303 	 */
6304 	if (count > 0) {
6305 		/*
6306 		 * Can't upgrade - other threads inside.
6307 		 */
6308 		SQ_PUTLOCKS_EXIT(sq);
6309 		mutex_exit(SQLOCK(sq));
6310 		return;
6311 	}
6312 	/*
6313 	 * Need to set SQ_EXCL and make a claim on the syncq.
6314 	 */
6315 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6316 	sq->sq_flags |= SQ_EXCL;
6317 	ASSERT(sq->sq_count == 0);
6318 	sq->sq_count++;
6319 	SQ_PUTLOCKS_EXIT(sq);
6320 
6321 	/* Process the events list */
6322 	sq_run_events(sq);
6323 
6324 	/*
6325 	 * Release our claim...
6326 	 */
6327 	sq->sq_count--;
6328 
6329 	/*
6330 	 * And release SQ_EXCL.
6331 	 * We don't need to acquire the putlocks to release
6332 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6333 	 */
6334 	sq->sq_flags &= ~SQ_EXCL;
6335 
6336 	/*
6337 	 * sq_run_events should have released SQ_EXCL
6338 	 */
6339 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6340 
6341 	/*
6342 	 * If anything happened while we were running the
6343 	 * events (or was there before), we need to process
6344 	 * them now.  We shouldn't be exclusive sine we
6345 	 * released the perimeter above (plus, we asserted
6346 	 * for it).
6347 	 */
6348 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6349 		drain_syncq(sq);
6350 	else
6351 		mutex_exit(SQLOCK(sq));
6352 }
6353 
6354 /*
6355  * Perform delayed processing. The caller has to make sure that it is safe
6356  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6357  * set).
6358  *
6359  * Assume that the caller has NO claims on the syncq.  However, a claim
6360  * on the syncq does not indicate that a thread is draining the syncq.
6361  * There may be more claims on the syncq than there are threads draining
6362  * (i.e.  #_threads_draining <= sq_count)
6363  *
6364  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6365  * in order to preserve qwriter(OUTER) ordering constraints.
6366  *
6367  * sq_putcount only needs to be checked when dispatching the queued
6368  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6369  */
6370 void
6371 drain_syncq(syncq_t *sq)
6372 {
6373 	queue_t		*qp;
6374 	uint16_t	count;
6375 	uint16_t	type = sq->sq_type;
6376 	uint16_t	flags = sq->sq_flags;
6377 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6378 
6379 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6380 	    "drain_syncq start:%p", sq);
6381 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6382 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6383 	    sq->sq_oprev == NULL) ||
6384 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6385 	    sq->sq_oprev != NULL));
6386 
6387 	/*
6388 	 * Drop SQ_SERVICE flag.
6389 	 */
6390 	if (bg_service)
6391 		sq->sq_svcflags &= ~SQ_SERVICE;
6392 
6393 	/*
6394 	 * If SQ_EXCL is set, someone else is processing this syncq - let them
6395 	 * finish the job.
6396 	 */
6397 	if (flags & SQ_EXCL) {
6398 		if (bg_service) {
6399 			ASSERT(sq->sq_servcount != 0);
6400 			sq->sq_servcount--;
6401 		}
6402 		mutex_exit(SQLOCK(sq));
6403 		return;
6404 	}
6405 
6406 	/*
6407 	 * This routine can be called by a background thread if
6408 	 * it was scheduled by a hi-priority thread.  SO, if there are
6409 	 * NOT messages queued, return (remember, we have the SQLOCK,
6410 	 * and it cannot change until we release it). Wakeup any waiters also.
6411 	 */
6412 	if (!(flags & SQ_QUEUED)) {
6413 		if (flags & SQ_WANTWAKEUP) {
6414 			flags &= ~SQ_WANTWAKEUP;
6415 			cv_broadcast(&sq->sq_wait);
6416 		}
6417 		if (flags & SQ_WANTEXWAKEUP) {
6418 			flags &= ~SQ_WANTEXWAKEUP;
6419 			cv_broadcast(&sq->sq_exitwait);
6420 		}
6421 		sq->sq_flags = flags;
6422 		if (bg_service) {
6423 			ASSERT(sq->sq_servcount != 0);
6424 			sq->sq_servcount--;
6425 		}
6426 		mutex_exit(SQLOCK(sq));
6427 		return;
6428 	}
6429 
6430 	/*
6431 	 * If this is not a concurrent put perimeter, we need to
6432 	 * become exclusive to drain.  Also, if not CIPUT, we would
6433 	 * not have acquired a putlock, so we don't need to check
6434 	 * the putcounts.  If not entering with a claim, we test
6435 	 * for sq_count == 0.
6436 	 */
6437 	type = sq->sq_type;
6438 	if (!(type & SQ_CIPUT)) {
6439 		if (sq->sq_count > 1) {
6440 			if (bg_service) {
6441 				ASSERT(sq->sq_servcount != 0);
6442 				sq->sq_servcount--;
6443 			}
6444 			mutex_exit(SQLOCK(sq));
6445 			return;
6446 		}
6447 		sq->sq_flags |= SQ_EXCL;
6448 	}
6449 
6450 	/*
6451 	 * This is where we make a claim to the syncq.
6452 	 * This can either be done by incrementing a putlock, or
6453 	 * the sq_count.  But since we already have the SQLOCK
6454 	 * here, we just bump the sq_count.
6455 	 *
6456 	 * Note that after we make a claim, we need to let the code
6457 	 * fall through to the end of this routine to clean itself
6458 	 * up.  A return in the while loop will put the syncq in a
6459 	 * very bad state.
6460 	 */
6461 	sq->sq_count++;
6462 	ASSERT(sq->sq_count != 0);	/* wraparound */
6463 
6464 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6465 		/*
6466 		 * If we are told to stayaway or went exclusive,
6467 		 * we are done.
6468 		 */
6469 		if (flags & (SQ_STAYAWAY)) {
6470 			break;
6471 		}
6472 
6473 		/*
6474 		 * If there are events to run, do so.
6475 		 * We have one claim to the syncq, so if there are
6476 		 * more than one, other threads are running.
6477 		 */
6478 		if (sq->sq_evhead != NULL) {
6479 			ASSERT(sq->sq_flags & SQ_EVENTS);
6480 
6481 			count = sq->sq_count;
6482 			SQ_PUTLOCKS_ENTER(sq);
6483 			SUM_SQ_PUTCOUNTS(sq, count);
6484 			if (count > 1) {
6485 				SQ_PUTLOCKS_EXIT(sq);
6486 				/* Can't upgrade - other threads inside */
6487 				break;
6488 			}
6489 			ASSERT((flags & SQ_EXCL) == 0);
6490 			sq->sq_flags = flags | SQ_EXCL;
6491 			SQ_PUTLOCKS_EXIT(sq);
6492 			/*
6493 			 * we have the only claim, run the events,
6494 			 * sq_run_events will clear the SQ_EXCL flag.
6495 			 */
6496 			sq_run_events(sq);
6497 
6498 			/*
6499 			 * If this is a CIPUT perimeter, we need
6500 			 * to drop the SQ_EXCL flag so we can properly
6501 			 * continue draining the syncq.
6502 			 */
6503 			if (type & SQ_CIPUT) {
6504 				ASSERT(sq->sq_flags & SQ_EXCL);
6505 				sq->sq_flags &= ~SQ_EXCL;
6506 			}
6507 
6508 			/*
6509 			 * And go back to the beginning just in case
6510 			 * anything changed while we were away.
6511 			 */
6512 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6513 			continue;
6514 		}
6515 
6516 		ASSERT(sq->sq_evhead == NULL);
6517 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6518 
6519 		/*
6520 		 * Find the queue that is not draining.
6521 		 *
6522 		 * q_draining is protected by QLOCK which we do not hold.
6523 		 * But if it was set, then a thread was draining, and if it gets
6524 		 * cleared, then it was because the thread has successfully
6525 		 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6526 		 * state to happen, a thread needs the SQLOCK which we hold, and
6527 		 * if there was such a flag, we would have already seen it.
6528 		 */
6529 
6530 		for (qp = sq->sq_head;
6531 		    qp != NULL && (qp->q_draining ||
6532 		    (qp->q_sqflags & Q_SQDRAINING));
6533 		    qp = qp->q_sqnext)
6534 			;
6535 
6536 		if (qp == NULL)
6537 			break;
6538 
6539 		/*
6540 		 * We have a queue to work on, and we hold the
6541 		 * SQLOCK and one claim, call qdrain_syncq.
6542 		 * This means we need to release the SQLOCK and
6543 		 * acquire the QLOCK (OK since we have a claim).
6544 		 * Note that qdrain_syncq will actually dequeue
6545 		 * this queue from the sq_head list when it is
6546 		 * convinced all the work is done and release
6547 		 * the QLOCK before returning.
6548 		 */
6549 		qp->q_sqflags |= Q_SQDRAINING;
6550 		mutex_exit(SQLOCK(sq));
6551 		mutex_enter(QLOCK(qp));
6552 		qdrain_syncq(sq, qp);
6553 		mutex_enter(SQLOCK(sq));
6554 
6555 		/* The queue is drained */
6556 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6557 		qp->q_sqflags &= ~Q_SQDRAINING;
6558 		/*
6559 		 * NOTE: After this point qp should not be used since it may be
6560 		 * closed.
6561 		 */
6562 	}
6563 
6564 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6565 	flags = sq->sq_flags;
6566 
6567 	/*
6568 	 * sq->sq_head cannot change because we hold the
6569 	 * sqlock. However, a thread CAN decide that it is no longer
6570 	 * going to drain that queue.  However, this should be due to
6571 	 * a GOAWAY state, and we should see that here.
6572 	 *
6573 	 * This loop is not very efficient. One solution may be adding a second
6574 	 * pointer to the "draining" queue, but it is difficult to do when
6575 	 * queues are inserted in the middle due to priority ordering. Another
6576 	 * possibility is to yank the queue out of the sq list and put it onto
6577 	 * the "draining list" and then put it back if it can't be drained.
6578 	 */
6579 
6580 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6581 	    (type & SQ_CI) || sq->sq_head->q_draining);
6582 
6583 	/* Drop SQ_EXCL for non-CIPUT perimeters */
6584 	if (!(type & SQ_CIPUT))
6585 		flags &= ~SQ_EXCL;
6586 	ASSERT((flags & SQ_EXCL) == 0);
6587 
6588 	/* Wake up any waiters. */
6589 	if (flags & SQ_WANTWAKEUP) {
6590 		flags &= ~SQ_WANTWAKEUP;
6591 		cv_broadcast(&sq->sq_wait);
6592 	}
6593 	if (flags & SQ_WANTEXWAKEUP) {
6594 		flags &= ~SQ_WANTEXWAKEUP;
6595 		cv_broadcast(&sq->sq_exitwait);
6596 	}
6597 	sq->sq_flags = flags;
6598 
6599 	ASSERT(sq->sq_count != 0);
6600 	/* Release our claim. */
6601 	sq->sq_count--;
6602 
6603 	if (bg_service) {
6604 		ASSERT(sq->sq_servcount != 0);
6605 		sq->sq_servcount--;
6606 	}
6607 
6608 	mutex_exit(SQLOCK(sq));
6609 
6610 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6611 	    "drain_syncq end:%p", sq);
6612 }
6613 
6614 
6615 /*
6616  *
6617  * qdrain_syncq can be called (currently) from only one of two places:
6618  *	drain_syncq
6619  * 	putnext  (or some variation of it).
6620  * and eventually
6621  * 	qwait(_sig)
6622  *
6623  * If called from drain_syncq, we found it in the list of queues needing
6624  * service, so there is work to be done (or it wouldn't be in the list).
6625  *
6626  * If called from some putnext variation, it was because the
6627  * perimeter is open, but messages are blocking a putnext and
6628  * there is not a thread working on it.  Now a thread could start
6629  * working on it while we are getting ready to do so ourself, but
6630  * the thread would set the q_draining flag, and we can spin out.
6631  *
6632  * As for qwait(_sig), I think I shall let it continue to call
6633  * drain_syncq directly (after all, it will get here eventually).
6634  *
6635  * qdrain_syncq has to terminate when:
6636  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6637  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6638  *
6639  * ASSUMES:
6640  *	One claim
6641  * 	QLOCK held
6642  * 	SQLOCK not held
6643  *	Will release QLOCK before returning
6644  */
6645 void
6646 qdrain_syncq(syncq_t *sq, queue_t *q)
6647 {
6648 	mblk_t		*bp;
6649 #ifdef DEBUG
6650 	uint16_t	count;
6651 #endif
6652 
6653 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6654 	    "drain_syncq start:%p", sq);
6655 	ASSERT(q->q_syncq == sq);
6656 	ASSERT(MUTEX_HELD(QLOCK(q)));
6657 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6658 	/*
6659 	 * For non-CIPUT perimeters, we should be called with the exclusive bit
6660 	 * set already. For CIPUT perimeters, we will be doing a concurrent
6661 	 * drain, so it better not be set.
6662 	 */
6663 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6664 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6665 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6666 	/*
6667 	 * All outer pointers are set, or none of them are
6668 	 */
6669 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6670 	    sq->sq_oprev == NULL) ||
6671 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6672 	    sq->sq_oprev != NULL));
6673 #ifdef DEBUG
6674 	count = sq->sq_count;
6675 	/*
6676 	 * This is OK without the putlocks, because we have one
6677 	 * claim either from the sq_count, or a putcount.  We could
6678 	 * get an erroneous value from other counts, but ours won't
6679 	 * change, so one way or another, we will have at least a
6680 	 * value of one.
6681 	 */
6682 	SUM_SQ_PUTCOUNTS(sq, count);
6683 	ASSERT(count >= 1);
6684 #endif /* DEBUG */
6685 
6686 	/*
6687 	 * The first thing to do is find out if a thread is already draining
6688 	 * this queue. If so, we are done, just return.
6689 	 */
6690 	if (q->q_draining) {
6691 		mutex_exit(QLOCK(q));
6692 		return;
6693 	}
6694 
6695 	/*
6696 	 * If the perimeter is exclusive, there is nothing we can do right now,
6697 	 * go away. Note that there is nothing to prevent this case from
6698 	 * changing right after this check, but the spin-out will catch it.
6699 	 */
6700 
6701 	/* Tell other threads that we are draining this queue */
6702 	q->q_draining = 1;	/* Protected by QLOCK */
6703 
6704 	/*
6705 	 * If there is nothing to do, clear QFULL as necessary. This caters for
6706 	 * the case where an empty queue was enqueued onto the syncq.
6707 	 */
6708 	if (q->q_sqhead == NULL) {
6709 		ASSERT(q->q_syncqmsgs == 0);
6710 		mutex_exit(QLOCK(q));
6711 		clr_qfull(q);
6712 		mutex_enter(QLOCK(q));
6713 	}
6714 
6715 	/*
6716 	 * Note that q_sqhead must be re-checked here in case another message
6717 	 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6718 	 */
6719 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6720 		/*
6721 		 * Because we can enter this routine just because a putnext is
6722 		 * blocked, we need to spin out if the perimeter wants to go
6723 		 * exclusive as well as just blocked. We need to spin out also
6724 		 * if events are queued on the syncq.
6725 		 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6726 		 * set it, and it can't become exclusive while we hold a claim.
6727 		 */
6728 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6729 			break;
6730 		}
6731 
6732 #ifdef DEBUG
6733 		/*
6734 		 * Since we are in qdrain_syncq, we already know the queue,
6735 		 * but for sanity, we want to check this against the qp that
6736 		 * was passed in by bp->b_queue.
6737 		 */
6738 
6739 		ASSERT(bp->b_queue == q);
6740 		ASSERT(bp->b_queue->q_syncq == sq);
6741 		bp->b_queue = NULL;
6742 
6743 		/*
6744 		 * We would have the following check in the DEBUG code:
6745 		 *
6746 		 * if (bp->b_prev != NULL)  {
6747 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6748 		 * }
6749 		 *
6750 		 * This can't be done, however, since IP modifies qinfo
6751 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6752 		 * qinfo), invalidating the check.
6753 		 * So the assignment to func is left here, but the ASSERT itself
6754 		 * is removed until the whole issue is resolved.
6755 		 */
6756 #endif
6757 		ASSERT(q->q_sqhead == bp);
6758 		q->q_sqhead = bp->b_next;
6759 		bp->b_prev = bp->b_next = NULL;
6760 		ASSERT(q->q_syncqmsgs > 0);
6761 		mutex_exit(QLOCK(q));
6762 
6763 		ASSERT(bp->b_datap->db_ref != 0);
6764 
6765 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6766 
6767 		mutex_enter(QLOCK(q));
6768 
6769 		/*
6770 		 * q_syncqmsgs should only be decremented after executing the
6771 		 * put procedure to avoid message re-ordering. This is due to an
6772 		 * optimisation in putnext() which can call the put procedure
6773 		 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6774 		 * being set).
6775 		 *
6776 		 * We also need to clear QFULL in the next service procedure
6777 		 * queue if this is the last message destined for that queue.
6778 		 *
6779 		 * It would make better sense to have some sort of tunable for
6780 		 * the low water mark, but these semantics are not yet defined.
6781 		 * So, alas, we use a constant.
6782 		 */
6783 		if (--q->q_syncqmsgs == 0) {
6784 			mutex_exit(QLOCK(q));
6785 			clr_qfull(q);
6786 			mutex_enter(QLOCK(q));
6787 		}
6788 
6789 		/*
6790 		 * Always clear SQ_EXCL when CIPUT in order to handle
6791 		 * qwriter(INNER). The putp() can call qwriter and get exclusive
6792 		 * access IFF this is the only claim. So, we need to test for
6793 		 * this possibility, acquire the mutex and clear the bit.
6794 		 */
6795 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6796 			mutex_enter(SQLOCK(sq));
6797 			sq->sq_flags &= ~SQ_EXCL;
6798 			mutex_exit(SQLOCK(sq));
6799 		}
6800 	}
6801 
6802 	/*
6803 	 * We should either have no messages on this queue, or we were told to
6804 	 * goaway by a waiter (which we will wake up at the end of this
6805 	 * function).
6806 	 */
6807 	ASSERT((q->q_sqhead == NULL) ||
6808 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6809 
6810 	ASSERT(MUTEX_HELD(QLOCK(q)));
6811 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6812 
6813 	/* Remove the q from the syncq list if all the messages are drained. */
6814 	if (q->q_sqhead == NULL) {
6815 		ASSERT(q->q_syncqmsgs == 0);
6816 		mutex_enter(SQLOCK(sq));
6817 		if (q->q_sqflags & Q_SQQUEUED)
6818 			SQRM_Q(sq, q);
6819 		mutex_exit(SQLOCK(sq));
6820 		/*
6821 		 * Since the queue is removed from the list, reset its priority.
6822 		 */
6823 		q->q_spri = 0;
6824 	}
6825 
6826 	/*
6827 	 * Remember, the q_draining flag is used to let another thread know
6828 	 * that there is a thread currently draining the messages for a queue.
6829 	 * Since we are now done with this queue (even if there may be messages
6830 	 * still there), we need to clear this flag so some thread will work on
6831 	 * it if needed.
6832 	 */
6833 	ASSERT(q->q_draining);
6834 	q->q_draining = 0;
6835 
6836 	/* Called with a claim, so OK to drop all locks. */
6837 	mutex_exit(QLOCK(q));
6838 
6839 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6840 	    "drain_syncq end:%p", sq);
6841 }
6842 /* END OF QDRAIN_SYNCQ  */
6843 
6844 
6845 /*
6846  * This is the mate to qdrain_syncq, except that it is putting the message onto
6847  * the queue instead of draining. Since the message is destined for the queue
6848  * that is selected, there is no need to identify the function because the
6849  * message is intended for the put routine for the queue. For debug kernels,
6850  * this routine will do it anyway just in case.
6851  *
6852  * After the message is enqueued on the syncq, it calls putnext_tail()
6853  * which will schedule a background thread to actually process the message.
6854  *
6855  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6856  * SQLOCK(sq) and QLOCK(q) are not held.
6857  */
6858 void
6859 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6860 {
6861 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6862 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6863 	ASSERT(sq->sq_count > 0);
6864 	ASSERT(q->q_syncq == sq);
6865 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6866 	    sq->sq_oprev == NULL) ||
6867 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6868 	    sq->sq_oprev != NULL));
6869 
6870 	mutex_enter(QLOCK(q));
6871 
6872 #ifdef DEBUG
6873 	/*
6874 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6875 	 * to trace the queue that the message is intended for.  Note
6876 	 * that the original use was to identify the queue and function
6877 	 * to call on the drain.  In the new syncq, we have the context
6878 	 * of the queue that we are draining, so call it's putproc and
6879 	 * don't rely on the saved values.  But for debug this is still
6880 	 * useful information.
6881 	 */
6882 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6883 	mp->b_queue = q;
6884 	mp->b_next = NULL;
6885 #endif
6886 	ASSERT(q->q_syncq == sq);
6887 	/*
6888 	 * Enqueue the message on the list.
6889 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6890 	 * protect it.  So it's ok to acquire SQLOCK after SQPUT_MP().
6891 	 */
6892 	SQPUT_MP(q, mp);
6893 	mutex_enter(SQLOCK(sq));
6894 
6895 	/*
6896 	 * And queue on syncq for scheduling, if not already queued.
6897 	 * Note that we need the SQLOCK for this, and for testing flags
6898 	 * at the end to see if we will drain.  So grab it now, and
6899 	 * release it before we call qdrain_syncq or return.
6900 	 */
6901 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6902 		q->q_spri = curthread->t_pri;
6903 		SQPUT_Q(sq, q);
6904 	}
6905 #ifdef DEBUG
6906 	else {
6907 		/*
6908 		 * All of these conditions MUST be true!
6909 		 */
6910 		ASSERT(sq->sq_tail != NULL);
6911 		if (sq->sq_tail == sq->sq_head) {
6912 			ASSERT((q->q_sqprev == NULL) &&
6913 			    (q->q_sqnext == NULL));
6914 		} else {
6915 			ASSERT((q->q_sqprev != NULL) ||
6916 			    (q->q_sqnext != NULL));
6917 		}
6918 		ASSERT(sq->sq_flags & SQ_QUEUED);
6919 		ASSERT(q->q_syncqmsgs != 0);
6920 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6921 	}
6922 #endif
6923 	mutex_exit(QLOCK(q));
6924 	/*
6925 	 * SQLOCK is still held, so sq_count can be safely decremented.
6926 	 */
6927 	sq->sq_count--;
6928 
6929 	putnext_tail(sq, q, 0);
6930 	/* Should not reference sq or q after this point. */
6931 }
6932 
6933 /*  End of qfill_syncq  */
6934 
6935 /*
6936  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6937  * that would be put into qp by drain_syncq.
6938  * Used when deleting the syncq (qp == NULL) or when detaching
6939  * a queue (qp != NULL).
6940  * Return non-zero if one or more messages were freed.
6941  *
6942  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
6943  * sq_putlocks are used.
6944  *
6945  * NOTE: This function assumes that it is called from the close() context and
6946  * that all the queues in the syncq are going away. For this reason it doesn't
6947  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6948  * currently valid, but it is useful to rethink this function to behave properly
6949  * in other cases.
6950  */
6951 int
6952 flush_syncq(syncq_t *sq, queue_t *qp)
6953 {
6954 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6955 	queue_t		*q;
6956 	int		ret = 0;
6957 
6958 	mutex_enter(SQLOCK(sq));
6959 
6960 	/*
6961 	 * Before we leave, we need to make sure there are no
6962 	 * events listed for this queue.  All events for this queue
6963 	 * will just be freed.
6964 	 */
6965 	if (qp != NULL && sq->sq_evhead != NULL) {
6966 		ASSERT(sq->sq_flags & SQ_EVENTS);
6967 
6968 		mp_prev = NULL;
6969 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6970 			mp_next = bp->b_next;
6971 			if (bp->b_queue == qp) {
6972 				/* Delete this message */
6973 				if (mp_prev != NULL) {
6974 					mp_prev->b_next = mp_next;
6975 					/*
6976 					 * Update sq_evtail if the last element
6977 					 * is removed.
6978 					 */
6979 					if (bp == sq->sq_evtail) {
6980 						ASSERT(mp_next == NULL);
6981 						sq->sq_evtail = mp_prev;
6982 					}
6983 				} else
6984 					sq->sq_evhead = mp_next;
6985 				if (sq->sq_evhead == NULL)
6986 					sq->sq_flags &= ~SQ_EVENTS;
6987 				bp->b_prev = bp->b_next = NULL;
6988 				freemsg(bp);
6989 				ret++;
6990 			} else {
6991 				mp_prev = bp;
6992 			}
6993 		}
6994 	}
6995 
6996 	/*
6997 	 * Walk sq_head and:
6998 	 *	- match qp if qp is set, remove it's messages
6999 	 *	- all if qp is not set
7000 	 */
7001 	q = sq->sq_head;
7002 	while (q != NULL) {
7003 		ASSERT(q->q_syncq == sq);
7004 		if ((qp == NULL) || (qp == q)) {
7005 			/*
7006 			 * Yank the messages as a list off the queue
7007 			 */
7008 			mp_head = q->q_sqhead;
7009 			/*
7010 			 * We do not have QLOCK(q) here (which is safe due to
7011 			 * assumptions mentioned above). To obtain the lock we
7012 			 * need to release SQLOCK which may allow lots of things
7013 			 * to change upon us. This place requires more analysis.
7014 			 */
7015 			q->q_sqhead = q->q_sqtail = NULL;
7016 			ASSERT(mp_head->b_queue &&
7017 			    mp_head->b_queue->q_syncq == sq);
7018 
7019 			/*
7020 			 * Free each of the messages.
7021 			 */
7022 			for (bp = mp_head; bp != NULL; bp = mp_next) {
7023 				mp_next = bp->b_next;
7024 				bp->b_prev = bp->b_next = NULL;
7025 				freemsg(bp);
7026 				ret++;
7027 			}
7028 			/*
7029 			 * Now remove the queue from the syncq.
7030 			 */
7031 			ASSERT(q->q_sqflags & Q_SQQUEUED);
7032 			SQRM_Q(sq, q);
7033 			q->q_spri = 0;
7034 			q->q_syncqmsgs = 0;
7035 
7036 			/*
7037 			 * If qp was specified, we are done with it and are
7038 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
7039 			 * waiters while we still have the SQLOCK.
7040 			 */
7041 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7042 				sq->sq_flags &= ~SQ_WANTWAKEUP;
7043 				cv_broadcast(&sq->sq_wait);
7044 			}
7045 			/* Drop SQLOCK across clr_qfull */
7046 			mutex_exit(SQLOCK(sq));
7047 
7048 			/*
7049 			 * We avoid doing the test that drain_syncq does and
7050 			 * unconditionally clear qfull for every flushed
7051 			 * message. Since flush_syncq is only called during
7052 			 * close this should not be a problem.
7053 			 */
7054 			clr_qfull(q);
7055 			if (qp != NULL) {
7056 				return (ret);
7057 			} else {
7058 				mutex_enter(SQLOCK(sq));
7059 				/*
7060 				 * The head was removed by SQRM_Q above.
7061 				 * reread the new head and flush it.
7062 				 */
7063 				q = sq->sq_head;
7064 			}
7065 		} else {
7066 			q = q->q_sqnext;
7067 		}
7068 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
7069 	}
7070 
7071 	if (sq->sq_flags & SQ_WANTWAKEUP) {
7072 		sq->sq_flags &= ~SQ_WANTWAKEUP;
7073 		cv_broadcast(&sq->sq_wait);
7074 	}
7075 
7076 	mutex_exit(SQLOCK(sq));
7077 	return (ret);
7078 }
7079 
7080 /*
7081  * Propagate all messages from a syncq to the next syncq that are associated
7082  * with the specified queue. If the queue is attached to a driver or if the
7083  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7084  *
7085  * Assumes that the stream is strlock()'ed. We don't come here if there
7086  * are no messages to propagate.
7087  *
7088  * NOTE : If the queue is attached to a driver, all the messages are freed
7089  * as there is no point in propagating the messages from the driver syncq
7090  * to the closing stream head which will in turn get freed later.
7091  */
7092 static int
7093 propagate_syncq(queue_t *qp)
7094 {
7095 	mblk_t		*bp, *head, *tail, *prev, *next;
7096 	syncq_t 	*sq;
7097 	queue_t		*nqp;
7098 	syncq_t		*nsq;
7099 	boolean_t	isdriver;
7100 	int 		moved = 0;
7101 	uint16_t	flags;
7102 	pri_t		priority = curthread->t_pri;
7103 #ifdef DEBUG
7104 	void		(*func)();
7105 #endif
7106 
7107 	sq = qp->q_syncq;
7108 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7109 	/* debug macro */
7110 	SQ_PUTLOCKS_HELD(sq);
7111 	/*
7112 	 * As entersq() does not increment the sq_count for
7113 	 * the write side, check sq_count for non-QPERQ
7114 	 * perimeters alone.
7115 	 */
7116 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7117 
7118 	/*
7119 	 * propagate_syncq() can be called because of either messages on the
7120 	 * queue syncq or because on events on the queue syncq. Do actual
7121 	 * message propagations if there are any messages.
7122 	 */
7123 	if (qp->q_syncqmsgs) {
7124 		isdriver = (qp->q_flag & QISDRV);
7125 
7126 		if (!isdriver) {
7127 			nqp = qp->q_next;
7128 			nsq = nqp->q_syncq;
7129 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7130 			/* debug macro */
7131 			SQ_PUTLOCKS_HELD(nsq);
7132 #ifdef DEBUG
7133 			func = (void (*)())nqp->q_qinfo->qi_putp;
7134 #endif
7135 		}
7136 
7137 		SQRM_Q(sq, qp);
7138 		priority = MAX(qp->q_spri, priority);
7139 		qp->q_spri = 0;
7140 		head = qp->q_sqhead;
7141 		tail = qp->q_sqtail;
7142 		qp->q_sqhead = qp->q_sqtail = NULL;
7143 		qp->q_syncqmsgs = 0;
7144 
7145 		/*
7146 		 * Walk the list of messages, and free them if this is a driver,
7147 		 * otherwise reset the b_prev and b_queue value to the new putp.
7148 		 * Afterward, we will just add the head to the end of the next
7149 		 * syncq, and point the tail to the end of this one.
7150 		 */
7151 
7152 		for (bp = head; bp != NULL; bp = next) {
7153 			next = bp->b_next;
7154 			if (isdriver) {
7155 				bp->b_prev = bp->b_next = NULL;
7156 				freemsg(bp);
7157 				continue;
7158 			}
7159 			/* Change the q values for this message */
7160 			bp->b_queue = nqp;
7161 #ifdef DEBUG
7162 			bp->b_prev = (mblk_t *)func;
7163 #endif
7164 			moved++;
7165 		}
7166 		/*
7167 		 * Attach list of messages to the end of the new queue (if there
7168 		 * is a list of messages).
7169 		 */
7170 
7171 		if (!isdriver && head != NULL) {
7172 			ASSERT(tail != NULL);
7173 			if (nqp->q_sqhead == NULL) {
7174 				nqp->q_sqhead = head;
7175 			} else {
7176 				ASSERT(nqp->q_sqtail != NULL);
7177 				nqp->q_sqtail->b_next = head;
7178 			}
7179 			nqp->q_sqtail = tail;
7180 			/*
7181 			 * When messages are moved from high priority queue to
7182 			 * another queue, the destination queue priority is
7183 			 * upgraded.
7184 			 */
7185 
7186 			if (priority > nqp->q_spri)
7187 				nqp->q_spri = priority;
7188 
7189 			SQPUT_Q(nsq, nqp);
7190 
7191 			nqp->q_syncqmsgs += moved;
7192 			ASSERT(nqp->q_syncqmsgs != 0);
7193 		}
7194 	}
7195 
7196 	/*
7197 	 * Before we leave, we need to make sure there are no
7198 	 * events listed for this queue.  All events for this queue
7199 	 * will just be freed.
7200 	 */
7201 	if (sq->sq_evhead != NULL) {
7202 		ASSERT(sq->sq_flags & SQ_EVENTS);
7203 		prev = NULL;
7204 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7205 			next = bp->b_next;
7206 			if (bp->b_queue == qp) {
7207 				/* Delete this message */
7208 				if (prev != NULL) {
7209 					prev->b_next = next;
7210 					/*
7211 					 * Update sq_evtail if the last element
7212 					 * is removed.
7213 					 */
7214 					if (bp == sq->sq_evtail) {
7215 						ASSERT(next == NULL);
7216 						sq->sq_evtail = prev;
7217 					}
7218 				} else
7219 					sq->sq_evhead = next;
7220 				if (sq->sq_evhead == NULL)
7221 					sq->sq_flags &= ~SQ_EVENTS;
7222 				bp->b_prev = bp->b_next = NULL;
7223 				freemsg(bp);
7224 			} else {
7225 				prev = bp;
7226 			}
7227 		}
7228 	}
7229 
7230 	flags = sq->sq_flags;
7231 
7232 	/* Wake up any waiter before leaving. */
7233 	if (flags & SQ_WANTWAKEUP) {
7234 		flags &= ~SQ_WANTWAKEUP;
7235 		cv_broadcast(&sq->sq_wait);
7236 	}
7237 	sq->sq_flags = flags;
7238 
7239 	return (moved);
7240 }
7241 
7242 /*
7243  * Try and upgrade to exclusive access at the inner perimeter. If this can
7244  * not be done without blocking then request will be queued on the syncq
7245  * and drain_syncq will run it later.
7246  *
7247  * This routine can only be called from put or service procedures plus
7248  * asynchronous callback routines that have properly entered the queue (with
7249  * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7250  * associated with q.
7251  */
7252 void
7253 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7254 {
7255 	syncq_t	*sq = q->q_syncq;
7256 	uint16_t count;
7257 
7258 	mutex_enter(SQLOCK(sq));
7259 	count = sq->sq_count;
7260 	SQ_PUTLOCKS_ENTER(sq);
7261 	SUM_SQ_PUTCOUNTS(sq, count);
7262 	ASSERT(count >= 1);
7263 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7264 
7265 	if (count == 1) {
7266 		/*
7267 		 * Can upgrade. This case also handles nested qwriter calls
7268 		 * (when the qwriter callback function calls qwriter). In that
7269 		 * case SQ_EXCL is already set.
7270 		 */
7271 		sq->sq_flags |= SQ_EXCL;
7272 		SQ_PUTLOCKS_EXIT(sq);
7273 		mutex_exit(SQLOCK(sq));
7274 		(*func)(q, mp);
7275 		/*
7276 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7277 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7278 		 * until putnext, leavesq, or drain_syncq drops it.
7279 		 * That way we handle nested qwriter(INNER) without dropping
7280 		 * SQ_EXCL until the outermost qwriter callback routine is
7281 		 * done.
7282 		 */
7283 		return;
7284 	}
7285 	SQ_PUTLOCKS_EXIT(sq);
7286 	sqfill_events(sq, q, mp, func);
7287 }
7288 
7289 /*
7290  * Synchronous callback support functions
7291  */
7292 
7293 /*
7294  * Allocate a callback parameter structure.
7295  * Assumes that caller initializes the flags and the id.
7296  * Acquires SQLOCK(sq) if non-NULL is returned.
7297  */
7298 callbparams_t *
7299 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7300 {
7301 	callbparams_t *cbp;
7302 	size_t size = sizeof (callbparams_t);
7303 
7304 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7305 
7306 	/*
7307 	 * Only try tryhard allocation if the caller is ready to panic.
7308 	 * Otherwise just fail.
7309 	 */
7310 	if (cbp == NULL) {
7311 		if (kmflags & KM_PANIC)
7312 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7313 			    &size, kmflags);
7314 		else
7315 			return (NULL);
7316 	}
7317 
7318 	ASSERT(size >= sizeof (callbparams_t));
7319 	cbp->cbp_size = size;
7320 	cbp->cbp_sq = sq;
7321 	cbp->cbp_func = func;
7322 	cbp->cbp_arg = arg;
7323 	mutex_enter(SQLOCK(sq));
7324 	cbp->cbp_next = sq->sq_callbpend;
7325 	sq->sq_callbpend = cbp;
7326 	return (cbp);
7327 }
7328 
7329 void
7330 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7331 {
7332 	callbparams_t **pp, *p;
7333 
7334 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7335 
7336 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7337 		if (p == cbp) {
7338 			*pp = p->cbp_next;
7339 			kmem_free(p, p->cbp_size);
7340 			return;
7341 		}
7342 	}
7343 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7344 	    "callbparams_free: not found\n"));
7345 }
7346 
7347 void
7348 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7349 {
7350 	callbparams_t **pp, *p;
7351 
7352 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7353 
7354 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7355 		if (p->cbp_id == id && p->cbp_flags == flag) {
7356 			*pp = p->cbp_next;
7357 			kmem_free(p, p->cbp_size);
7358 			return;
7359 		}
7360 	}
7361 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7362 	    "callbparams_free_id: not found\n"));
7363 }
7364 
7365 /*
7366  * Callback wrapper function used by once-only callbacks that can be
7367  * cancelled (qtimeout and qbufcall)
7368  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7369  * cancelled by the qun* functions.
7370  */
7371 void
7372 qcallbwrapper(void *arg)
7373 {
7374 	callbparams_t *cbp = arg;
7375 	syncq_t	*sq;
7376 	uint16_t count = 0;
7377 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7378 	uint16_t type;
7379 
7380 	sq = cbp->cbp_sq;
7381 	mutex_enter(SQLOCK(sq));
7382 	type = sq->sq_type;
7383 	if (!(type & SQ_CICB)) {
7384 		count = sq->sq_count;
7385 		SQ_PUTLOCKS_ENTER(sq);
7386 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7387 		SUM_SQ_PUTCOUNTS(sq, count);
7388 		sq->sq_needexcl++;
7389 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7390 		waitflags |= SQ_MESSAGES;
7391 	}
7392 	/* Can not handle exclusive entry at outer perimeter */
7393 	ASSERT(type & SQ_COCB);
7394 
7395 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7396 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7397 		    (sq->sq_cancelid == cbp->cbp_id)) {
7398 			/* timeout has been cancelled */
7399 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7400 			callbparams_free(sq, cbp);
7401 			if (!(type & SQ_CICB)) {
7402 				ASSERT(sq->sq_needexcl > 0);
7403 				sq->sq_needexcl--;
7404 				if (sq->sq_needexcl == 0) {
7405 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7406 				}
7407 				SQ_PUTLOCKS_EXIT(sq);
7408 			}
7409 			mutex_exit(SQLOCK(sq));
7410 			return;
7411 		}
7412 		sq->sq_flags |= SQ_WANTWAKEUP;
7413 		if (!(type & SQ_CICB)) {
7414 			SQ_PUTLOCKS_EXIT(sq);
7415 		}
7416 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7417 		if (!(type & SQ_CICB)) {
7418 			count = sq->sq_count;
7419 			SQ_PUTLOCKS_ENTER(sq);
7420 			SUM_SQ_PUTCOUNTS(sq, count);
7421 		}
7422 	}
7423 
7424 	sq->sq_count++;
7425 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7426 	if (!(type & SQ_CICB)) {
7427 		ASSERT(count == 0);
7428 		sq->sq_flags |= SQ_EXCL;
7429 		ASSERT(sq->sq_needexcl > 0);
7430 		sq->sq_needexcl--;
7431 		if (sq->sq_needexcl == 0) {
7432 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7433 		}
7434 		SQ_PUTLOCKS_EXIT(sq);
7435 	}
7436 
7437 	mutex_exit(SQLOCK(sq));
7438 
7439 	cbp->cbp_func(cbp->cbp_arg);
7440 
7441 	/*
7442 	 * We drop the lock only for leavesq to re-acquire it.
7443 	 * Possible optimization is inline of leavesq.
7444 	 */
7445 	mutex_enter(SQLOCK(sq));
7446 	callbparams_free(sq, cbp);
7447 	mutex_exit(SQLOCK(sq));
7448 	leavesq(sq, SQ_CALLBACK);
7449 }
7450 
7451 /*
7452  * No need to grab sq_putlocks here. See comment in strsubr.h that
7453  * explains when sq_putlocks are used.
7454  *
7455  * sq_count (or one of the sq_putcounts) has already been
7456  * decremented by the caller, and if SQ_QUEUED, we need to call
7457  * drain_syncq (the global syncq drain).
7458  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7459  * one of two states, non-CIPUT perimeter, and we need to clear
7460  * it, or we went exclusive in the put procedure.  In any case,
7461  * we want to clear the bit now, and it is probably easier to do
7462  * this at the beginning of this function (remember, we hold
7463  * the SQLOCK).  Lastly, if there are other messages queued
7464  * on the syncq (and not for our destination), enable the syncq
7465  * for background work.
7466  */
7467 
7468 /* ARGSUSED */
7469 void
7470 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7471 {
7472 	uint16_t	flags = sq->sq_flags;
7473 
7474 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7475 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7476 
7477 	/* Clear SQ_EXCL if set in passflags */
7478 	if (passflags & SQ_EXCL) {
7479 		flags &= ~SQ_EXCL;
7480 	}
7481 	if (flags & SQ_WANTWAKEUP) {
7482 		flags &= ~SQ_WANTWAKEUP;
7483 		cv_broadcast(&sq->sq_wait);
7484 	}
7485 	if (flags & SQ_WANTEXWAKEUP) {
7486 		flags &= ~SQ_WANTEXWAKEUP;
7487 		cv_broadcast(&sq->sq_exitwait);
7488 	}
7489 	sq->sq_flags = flags;
7490 
7491 	/*
7492 	 * We have cleared SQ_EXCL if we were asked to, and started
7493 	 * the wakeup process for waiters.  If there are no writers
7494 	 * then we need to drain the syncq if we were told to, or
7495 	 * enable the background thread to do it.
7496 	 */
7497 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7498 		if ((passflags & SQ_QUEUED) ||
7499 		    (sq->sq_svcflags & SQ_DISABLED)) {
7500 			/* drain_syncq will take care of events in the list */
7501 			drain_syncq(sq);
7502 			return;
7503 		} else if (flags & SQ_QUEUED) {
7504 			sqenable(sq);
7505 		}
7506 	}
7507 	/* Drop the SQLOCK on exit */
7508 	mutex_exit(SQLOCK(sq));
7509 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7510 	    "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7511 }
7512 
7513 void
7514 set_qend(queue_t *q)
7515 {
7516 	mutex_enter(QLOCK(q));
7517 	if (!O_SAMESTR(q))
7518 		q->q_flag |= QEND;
7519 	else
7520 		q->q_flag &= ~QEND;
7521 	mutex_exit(QLOCK(q));
7522 	q = _OTHERQ(q);
7523 	mutex_enter(QLOCK(q));
7524 	if (!O_SAMESTR(q))
7525 		q->q_flag |= QEND;
7526 	else
7527 		q->q_flag &= ~QEND;
7528 	mutex_exit(QLOCK(q));
7529 }
7530 
7531 /*
7532  * Set QFULL in next service procedure queue (that cares) if not already
7533  * set and if there are already more messages on the syncq than
7534  * sq_max_size.  If sq_max_size is 0, no flow control will be asserted on
7535  * any syncq.
7536  *
7537  * The fq here is the next queue with a service procedure.  This is where
7538  * we would fail canputnext, so this is where we need to set QFULL.
7539  * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7540  *
7541  * We already have QLOCK at this point. To avoid cross-locks with
7542  * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7543  * SQLOCK and sd_reflock, we need to drop respective locks first.
7544  */
7545 void
7546 set_qfull(queue_t *q)
7547 {
7548 	queue_t		*fq = NULL;
7549 
7550 	ASSERT(MUTEX_HELD(QLOCK(q)));
7551 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7552 	    (q->q_syncqmsgs > sq_max_size)) {
7553 		if ((fq = q->q_nfsrv) == q) {
7554 			fq->q_flag |= QFULL;
7555 		} else {
7556 			mutex_exit(QLOCK(q));
7557 			mutex_enter(QLOCK(fq));
7558 			fq->q_flag |= QFULL;
7559 			mutex_exit(QLOCK(fq));
7560 			mutex_enter(QLOCK(q));
7561 		}
7562 	}
7563 }
7564 
7565 void
7566 clr_qfull(queue_t *q)
7567 {
7568 	queue_t	*oq = q;
7569 
7570 	q = q->q_nfsrv;
7571 	/* Fast check if there is any work to do before getting the lock. */
7572 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7573 		return;
7574 	}
7575 
7576 	/*
7577 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7578 	 * for QFULL being set.
7579 	 */
7580 	mutex_enter(QLOCK(q));
7581 	/*
7582 	 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7583 	 * Hence clear the QFULL.
7584 	 * If both q_count and q_mblkcnt are less than the hiwat mark,
7585 	 * clear the QFULL.
7586 	 */
7587 	if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7588 	    (q->q_mblkcnt < q->q_hiwat))) {
7589 		q->q_flag &= ~QFULL;
7590 		/*
7591 		 * A little more confusing, how about this way:
7592 		 * if someone wants to write,
7593 		 * AND
7594 		 *    both counts are less than the lowat mark
7595 		 *    OR
7596 		 *    the lowat mark is zero
7597 		 * THEN
7598 		 * backenable
7599 		 */
7600 		if ((q->q_flag & QWANTW) &&
7601 		    (((q->q_count < q->q_lowat) &&
7602 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7603 			q->q_flag &= ~QWANTW;
7604 			mutex_exit(QLOCK(q));
7605 			backenable(oq, 0);
7606 		} else
7607 			mutex_exit(QLOCK(q));
7608 	} else
7609 		mutex_exit(QLOCK(q));
7610 }
7611 
7612 /*
7613  * Set the forward service procedure pointer.
7614  *
7615  * Called at insert-time to cache a queue's next forward service procedure in
7616  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7617  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7618  * inserted does not have a service procedure, then q_nfsrv points to the next
7619  * queue forward that has a service procedure.  If the queue is at the logical
7620  * end of the stream (driver for write side, stream head for the read side)
7621  * and does not have a service procedure, then q_nfsrv also points to itself.
7622  */
7623 void
7624 set_nfsrv_ptr(
7625 	queue_t  *rnew,		/* read queue pointer to new module */
7626 	queue_t  *wnew,		/* write queue pointer to new module */
7627 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7628 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7629 {
7630 	queue_t *qp;
7631 
7632 	if (prev_wq->q_next == NULL) {
7633 		/*
7634 		 * Insert the driver, initialize the driver and stream head.
7635 		 * In this case, prev_rq/prev_wq should be the stream head.
7636 		 * _I_INSERT does not allow inserting a driver.  Make sure
7637 		 * that it is not an insertion.
7638 		 */
7639 		ASSERT(!(rnew->q_flag & _QINSERTING));
7640 		wnew->q_nfsrv = wnew;
7641 		if (rnew->q_qinfo->qi_srvp)
7642 			rnew->q_nfsrv = rnew;
7643 		else
7644 			rnew->q_nfsrv = prev_rq;
7645 		prev_rq->q_nfsrv = prev_rq;
7646 		prev_wq->q_nfsrv = prev_wq;
7647 	} else {
7648 		/*
7649 		 * set up read side q_nfsrv pointer.  This MUST be done
7650 		 * before setting the write side, because the setting of
7651 		 * the write side for a fifo may depend on it.
7652 		 *
7653 		 * Suppose we have a fifo that only has pipemod pushed.
7654 		 * pipemod has no read or write service procedures, so
7655 		 * nfsrv for both pipemod queues points to prev_rq (the
7656 		 * stream read head).  Now push bufmod (which has only a
7657 		 * read service procedure).  Doing the write side first,
7658 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7659 		 * is WRONG; the next queue forward from wnew with a
7660 		 * service procedure will be rnew, not the stream read head.
7661 		 * Since the downstream queue (which in the case of a fifo
7662 		 * is the read queue rnew) can affect upstream queues, it
7663 		 * needs to be done first.  Setting up the read side first
7664 		 * sets nfsrv for both pipemod queues to rnew and then
7665 		 * when the write side is set up, wnew-q_nfsrv will also
7666 		 * point to rnew.
7667 		 */
7668 		if (rnew->q_qinfo->qi_srvp) {
7669 			/*
7670 			 * use _OTHERQ() because, if this is a pipe, next
7671 			 * module may have been pushed from other end and
7672 			 * q_next could be a read queue.
7673 			 */
7674 			qp = _OTHERQ(prev_wq->q_next);
7675 			while (qp && qp->q_nfsrv != qp) {
7676 				qp->q_nfsrv = rnew;
7677 				qp = backq(qp);
7678 			}
7679 			rnew->q_nfsrv = rnew;
7680 		} else
7681 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7682 
7683 		/* set up write side q_nfsrv pointer */
7684 		if (wnew->q_qinfo->qi_srvp) {
7685 			wnew->q_nfsrv = wnew;
7686 
7687 			/*
7688 			 * For insertion, need to update nfsrv of the modules
7689 			 * above which do not have a service routine.
7690 			 */
7691 			if (rnew->q_flag & _QINSERTING) {
7692 				for (qp = prev_wq;
7693 				    qp != NULL && qp->q_nfsrv != qp;
7694 				    qp = backq(qp)) {
7695 					qp->q_nfsrv = wnew->q_nfsrv;
7696 				}
7697 			}
7698 		} else {
7699 			if (prev_wq->q_next == prev_rq)
7700 				/*
7701 				 * Since prev_wq/prev_rq are the middle of a
7702 				 * fifo, wnew/rnew will also be the middle of
7703 				 * a fifo and wnew's nfsrv is same as rnew's.
7704 				 */
7705 				wnew->q_nfsrv = rnew->q_nfsrv;
7706 			else
7707 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7708 		}
7709 	}
7710 }
7711 
7712 /*
7713  * Reset the forward service procedure pointer; called at remove-time.
7714  */
7715 void
7716 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7717 {
7718 	queue_t *tmp_qp;
7719 
7720 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7721 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7722 		for (tmp_qp = backq(wqp);
7723 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7724 		    tmp_qp = backq(tmp_qp)) {
7725 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7726 		}
7727 	}
7728 
7729 	/* reset the read side q_nfsrv pointer */
7730 	if (rqp->q_qinfo->qi_srvp) {
7731 		if (wqp->q_next) {	/* non-driver case */
7732 			tmp_qp = _OTHERQ(wqp->q_next);
7733 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7734 				/* Note that rqp->q_next cannot be NULL */
7735 				ASSERT(rqp->q_next != NULL);
7736 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7737 				tmp_qp = backq(tmp_qp);
7738 			}
7739 		}
7740 	}
7741 }
7742 
7743 /*
7744  * This routine should be called after all stream geometry changes to update
7745  * the stream head cached struio() rd/wr queue pointers. Note must be called
7746  * with the streamlock()ed.
7747  *
7748  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7749  *	 an explicit synchronous barrier module queue. That is, a queue that
7750  *	 has specified a struio() type.
7751  */
7752 static void
7753 strsetuio(stdata_t *stp)
7754 {
7755 	queue_t *wrq;
7756 
7757 	if (stp->sd_flag & STPLEX) {
7758 		/*
7759 		 * Not streamhead, but a mux, so no Synchronous STREAMS.
7760 		 */
7761 		stp->sd_struiowrq = NULL;
7762 		stp->sd_struiordq = NULL;
7763 		return;
7764 	}
7765 	/*
7766 	 * Scan the write queue(s) while synchronous
7767 	 * until we find a qinfo uio type specified.
7768 	 */
7769 	wrq = stp->sd_wrq->q_next;
7770 	while (wrq) {
7771 		if (wrq->q_struiot == STRUIOT_NONE) {
7772 			wrq = 0;
7773 			break;
7774 		}
7775 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7776 			break;
7777 		if (! _SAMESTR(wrq)) {
7778 			wrq = 0;
7779 			break;
7780 		}
7781 		wrq = wrq->q_next;
7782 	}
7783 	stp->sd_struiowrq = wrq;
7784 	/*
7785 	 * Scan the read queue(s) while synchronous
7786 	 * until we find a qinfo uio type specified.
7787 	 */
7788 	wrq = stp->sd_wrq->q_next;
7789 	while (wrq) {
7790 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7791 			wrq = 0;
7792 			break;
7793 		}
7794 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7795 			break;
7796 		if (! _SAMESTR(wrq)) {
7797 			wrq = 0;
7798 			break;
7799 		}
7800 		wrq = wrq->q_next;
7801 	}
7802 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7803 }
7804 
7805 /*
7806  * pass_wput, unblocks the passthru queues, so that
7807  * messages can arrive at muxs lower read queue, before
7808  * I_LINK/I_UNLINK is acked/nacked.
7809  */
7810 static void
7811 pass_wput(queue_t *q, mblk_t *mp)
7812 {
7813 	syncq_t *sq;
7814 
7815 	sq = _RD(q)->q_syncq;
7816 	if (sq->sq_flags & SQ_BLOCKED)
7817 		unblocksq(sq, SQ_BLOCKED, 0);
7818 	putnext(q, mp);
7819 }
7820 
7821 /*
7822  * Set up queues for the link/unlink.
7823  * Create a new queue and block it and then insert it
7824  * below the stream head on the lower stream.
7825  * This prevents any messages from arriving during the setq
7826  * as well as while the mux is processing the LINK/I_UNLINK.
7827  * The blocked passq is unblocked once the LINK/I_UNLINK has
7828  * been acked or nacked or if a message is generated and sent
7829  * down muxs write put procedure.
7830  * See pass_wput().
7831  *
7832  * After the new queue is inserted, all messages coming from below are
7833  * blocked. The call to strlock will ensure that all activity in the stream head
7834  * read queue syncq is stopped (sq_count drops to zero).
7835  */
7836 static queue_t *
7837 link_addpassthru(stdata_t *stpdown)
7838 {
7839 	queue_t *passq;
7840 	sqlist_t sqlist;
7841 
7842 	passq = allocq();
7843 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7844 	/* setq might sleep in allocator - avoid holding locks. */
7845 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7846 	    SQ_CI|SQ_CO, B_FALSE);
7847 	claimq(passq);
7848 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7849 	insertq(STREAM(passq), passq);
7850 
7851 	/*
7852 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7853 	 * since we are going to change q_ptr in the stream head.  Note that
7854 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7855 	 */
7856 	sqlist.sqlist_head = NULL;
7857 	sqlist.sqlist_index = 0;
7858 	sqlist.sqlist_size = sizeof (sqlist_t);
7859 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7860 	strlock(stpdown, &sqlist);
7861 	strunlock(stpdown, &sqlist);
7862 
7863 	releaseq(passq);
7864 	return (passq);
7865 }
7866 
7867 /*
7868  * Let messages flow up into the mux by removing
7869  * the passq.
7870  */
7871 static void
7872 link_rempassthru(queue_t *passq)
7873 {
7874 	claimq(passq);
7875 	removeq(passq);
7876 	releaseq(passq);
7877 	freeq(passq);
7878 }
7879 
7880 /*
7881  * Wait for the condition variable pointed to by `cvp' to be signaled,
7882  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7883  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7884  * then the wait will be non-interruptible.
7885  *
7886  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7887  */
7888 clock_t
7889 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7890 {
7891 	clock_t ret;
7892 
7893 	if (tim < 0) {
7894 		if (nosigs) {
7895 			cv_wait(cvp, mp);
7896 			ret = 1;
7897 		} else {
7898 			ret = cv_wait_sig(cvp, mp);
7899 		}
7900 	} else if (tim > 0) {
7901 		/*
7902 		 * convert milliseconds to clock ticks
7903 		 */
7904 		if (nosigs) {
7905 			ret = cv_reltimedwait(cvp, mp,
7906 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7907 		} else {
7908 			ret = cv_reltimedwait_sig(cvp, mp,
7909 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7910 		}
7911 	} else {
7912 		ret = -1;
7913 	}
7914 	return (ret);
7915 }
7916 
7917 /*
7918  * Wait until the stream head can determine if it is at the mark but
7919  * don't wait forever to prevent a race condition between the "mark" state
7920  * in the stream head and any mark state in the caller/user of this routine.
7921  *
7922  * This is used by sockets and for a socket it would be incorrect
7923  * to return a failure for SIOCATMARK when there is no data in the receive
7924  * queue and the marked urgent data is traveling up the stream.
7925  *
7926  * This routine waits until the mark is known by waiting for one of these
7927  * three events:
7928  *	The stream head read queue becoming non-empty (including an EOF).
7929  *	The STRATMARK flag being set (due to a MSGMARKNEXT message).
7930  *	The STRNOTATMARK flag being set (which indicates that the transport
7931  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7932  *	the mark).
7933  *
7934  * The routine returns 1 if the stream is at the mark; 0 if it can
7935  * be determined that the stream is not at the mark.
7936  * If the wait times out and it can't determine
7937  * whether or not the stream might be at the mark the routine will return -1.
7938  *
7939  * Note: This routine should only be used when a mark is pending i.e.,
7940  * in the socket case the SIGURG has been posted.
7941  * Note2: This can not wakeup just because synchronous streams indicate
7942  * that data is available since it is not possible to use the synchronous
7943  * streams interfaces to determine the b_flag value for the data queued below
7944  * the stream head.
7945  */
7946 int
7947 strwaitmark(vnode_t *vp)
7948 {
7949 	struct stdata *stp = vp->v_stream;
7950 	queue_t *rq = _RD(stp->sd_wrq);
7951 	int mark;
7952 
7953 	mutex_enter(&stp->sd_lock);
7954 	while (rq->q_first == NULL &&
7955 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7956 		stp->sd_flag |= RSLEEP;
7957 
7958 		/* Wait for 100 milliseconds for any state change. */
7959 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7960 			mutex_exit(&stp->sd_lock);
7961 			return (-1);
7962 		}
7963 	}
7964 	if (stp->sd_flag & STRATMARK)
7965 		mark = 1;
7966 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7967 		mark = 1;
7968 	else
7969 		mark = 0;
7970 
7971 	mutex_exit(&stp->sd_lock);
7972 	return (mark);
7973 }
7974 
7975 /*
7976  * Set a read side error. If persist is set change the socket error
7977  * to persistent. If errfunc is set install the function as the exported
7978  * error handler.
7979  */
7980 void
7981 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7982 {
7983 	struct stdata *stp = vp->v_stream;
7984 
7985 	mutex_enter(&stp->sd_lock);
7986 	stp->sd_rerror = error;
7987 	if (error == 0 && errfunc == NULL)
7988 		stp->sd_flag &= ~STRDERR;
7989 	else
7990 		stp->sd_flag |= STRDERR;
7991 	if (persist) {
7992 		stp->sd_flag &= ~STRDERRNONPERSIST;
7993 	} else {
7994 		stp->sd_flag |= STRDERRNONPERSIST;
7995 	}
7996 	stp->sd_rderrfunc = errfunc;
7997 	if (error != 0 || errfunc != NULL) {
7998 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7999 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8000 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8001 
8002 		mutex_exit(&stp->sd_lock);
8003 		pollwakeup(&stp->sd_pollist, POLLERR);
8004 		mutex_enter(&stp->sd_lock);
8005 
8006 		if (stp->sd_sigflags & S_ERROR)
8007 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8008 	}
8009 	mutex_exit(&stp->sd_lock);
8010 }
8011 
8012 /*
8013  * Set a write side error. If persist is set change the socket error
8014  * to persistent.
8015  */
8016 void
8017 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
8018 {
8019 	struct stdata *stp = vp->v_stream;
8020 
8021 	mutex_enter(&stp->sd_lock);
8022 	stp->sd_werror = error;
8023 	if (error == 0 && errfunc == NULL)
8024 		stp->sd_flag &= ~STWRERR;
8025 	else
8026 		stp->sd_flag |= STWRERR;
8027 	if (persist) {
8028 		stp->sd_flag &= ~STWRERRNONPERSIST;
8029 	} else {
8030 		stp->sd_flag |= STWRERRNONPERSIST;
8031 	}
8032 	stp->sd_wrerrfunc = errfunc;
8033 	if (error != 0 || errfunc != NULL) {
8034 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
8035 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
8036 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
8037 
8038 		mutex_exit(&stp->sd_lock);
8039 		pollwakeup(&stp->sd_pollist, POLLERR);
8040 		mutex_enter(&stp->sd_lock);
8041 
8042 		if (stp->sd_sigflags & S_ERROR)
8043 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8044 	}
8045 	mutex_exit(&stp->sd_lock);
8046 }
8047 
8048 /*
8049  * Make the stream return 0 (EOF) when all data has been read.
8050  * No effect on write side.
8051  */
8052 void
8053 strseteof(vnode_t *vp, int eof)
8054 {
8055 	struct stdata *stp = vp->v_stream;
8056 
8057 	mutex_enter(&stp->sd_lock);
8058 	if (!eof) {
8059 		stp->sd_flag &= ~STREOF;
8060 		mutex_exit(&stp->sd_lock);
8061 		return;
8062 	}
8063 	stp->sd_flag |= STREOF;
8064 	if (stp->sd_flag & RSLEEP) {
8065 		stp->sd_flag &= ~RSLEEP;
8066 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8067 	}
8068 
8069 	mutex_exit(&stp->sd_lock);
8070 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8071 	mutex_enter(&stp->sd_lock);
8072 
8073 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8074 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8075 	mutex_exit(&stp->sd_lock);
8076 }
8077 
8078 void
8079 strflushrq(vnode_t *vp, int flag)
8080 {
8081 	struct stdata *stp = vp->v_stream;
8082 
8083 	mutex_enter(&stp->sd_lock);
8084 	flushq(_RD(stp->sd_wrq), flag);
8085 	mutex_exit(&stp->sd_lock);
8086 }
8087 
8088 void
8089 strsetrputhooks(vnode_t *vp, uint_t flags,
8090 		msgfunc_t protofunc, msgfunc_t miscfunc)
8091 {
8092 	struct stdata *stp = vp->v_stream;
8093 
8094 	mutex_enter(&stp->sd_lock);
8095 
8096 	if (protofunc == NULL)
8097 		stp->sd_rprotofunc = strrput_proto;
8098 	else
8099 		stp->sd_rprotofunc = protofunc;
8100 
8101 	if (miscfunc == NULL)
8102 		stp->sd_rmiscfunc = strrput_misc;
8103 	else
8104 		stp->sd_rmiscfunc = miscfunc;
8105 
8106 	if (flags & SH_CONSOL_DATA)
8107 		stp->sd_rput_opt |= SR_CONSOL_DATA;
8108 	else
8109 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8110 
8111 	if (flags & SH_SIGALLDATA)
8112 		stp->sd_rput_opt |= SR_SIGALLDATA;
8113 	else
8114 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
8115 
8116 	if (flags & SH_IGN_ZEROLEN)
8117 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8118 	else
8119 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8120 
8121 	mutex_exit(&stp->sd_lock);
8122 }
8123 
8124 void
8125 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8126 {
8127 	struct stdata *stp = vp->v_stream;
8128 
8129 	mutex_enter(&stp->sd_lock);
8130 	stp->sd_closetime = closetime;
8131 
8132 	if (flags & SH_SIGPIPE)
8133 		stp->sd_wput_opt |= SW_SIGPIPE;
8134 	else
8135 		stp->sd_wput_opt &= ~SW_SIGPIPE;
8136 	if (flags & SH_RECHECK_ERR)
8137 		stp->sd_wput_opt |= SW_RECHECK_ERR;
8138 	else
8139 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8140 
8141 	mutex_exit(&stp->sd_lock);
8142 }
8143 
8144 void
8145 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8146 {
8147 	struct stdata *stp = vp->v_stream;
8148 
8149 	mutex_enter(&stp->sd_lock);
8150 
8151 	stp->sd_rputdatafunc = rdatafunc;
8152 	stp->sd_wputdatafunc = wdatafunc;
8153 
8154 	mutex_exit(&stp->sd_lock);
8155 }
8156 
8157 /* Used within framework when the queue is already locked */
8158 void
8159 qenable_locked(queue_t *q)
8160 {
8161 	stdata_t *stp = STREAM(q);
8162 
8163 	ASSERT(MUTEX_HELD(QLOCK(q)));
8164 
8165 	if (!q->q_qinfo->qi_srvp)
8166 		return;
8167 
8168 	/*
8169 	 * Do not place on run queue if already enabled or closing.
8170 	 */
8171 	if (q->q_flag & (QWCLOSE|QENAB))
8172 		return;
8173 
8174 	/*
8175 	 * mark queue enabled and place on run list if it is not already being
8176 	 * serviced. If it is serviced, the runservice() function will detect
8177 	 * that QENAB is set and call service procedure before clearing
8178 	 * QINSERVICE flag.
8179 	 */
8180 	q->q_flag |= QENAB;
8181 	if (q->q_flag & QINSERVICE)
8182 		return;
8183 
8184 	/* Record the time of qenable */
8185 	q->q_qtstamp = ddi_get_lbolt();
8186 
8187 	/*
8188 	 * Put the queue in the stp list and schedule it for background
8189 	 * processing if it is not already scheduled or if stream head does not
8190 	 * intent to process it in the foreground later by setting
8191 	 * STRS_WILLSERVICE flag.
8192 	 */
8193 	mutex_enter(&stp->sd_qlock);
8194 	/*
8195 	 * If there are already something on the list, stp flags should show
8196 	 * intention to drain it.
8197 	 */
8198 	IMPLY(STREAM_NEEDSERVICE(stp),
8199 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8200 
8201 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8202 	stp->sd_nqueues++;
8203 
8204 	/*
8205 	 * If no one will drain this stream we are the first producer and
8206 	 * need to schedule it for background thread.
8207 	 */
8208 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8209 		/*
8210 		 * No one will service this stream later, so we have to
8211 		 * schedule it now.
8212 		 */
8213 		STRSTAT(stenables);
8214 		stp->sd_svcflags |= STRS_SCHEDULED;
8215 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8216 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8217 
8218 		if (stp->sd_servid == NULL) {
8219 			/*
8220 			 * Task queue failed so fail over to the backup
8221 			 * servicing thread.
8222 			 */
8223 			STRSTAT(taskqfails);
8224 			/*
8225 			 * It is safe to clear STRS_SCHEDULED flag because it
8226 			 * was set by this thread above.
8227 			 */
8228 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8229 
8230 			/*
8231 			 * Failover scheduling is protected by service_queue
8232 			 * lock.
8233 			 */
8234 			mutex_enter(&service_queue);
8235 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8236 			ASSERT(q->q_link == NULL);
8237 			/*
8238 			 * Append the queue to qhead/qtail list.
8239 			 */
8240 			if (qhead == NULL)
8241 				qhead = q;
8242 			else
8243 				qtail->q_link = q;
8244 			qtail = q;
8245 			/*
8246 			 * Clear stp queue list.
8247 			 */
8248 			stp->sd_qhead = stp->sd_qtail = NULL;
8249 			stp->sd_nqueues = 0;
8250 			/*
8251 			 * Wakeup background queue processing thread.
8252 			 */
8253 			cv_signal(&services_to_run);
8254 			mutex_exit(&service_queue);
8255 		}
8256 	}
8257 	mutex_exit(&stp->sd_qlock);
8258 }
8259 
8260 static void
8261 queue_service(queue_t *q)
8262 {
8263 	/*
8264 	 * The queue in the list should have
8265 	 * QENAB flag set and should not have
8266 	 * QINSERVICE flag set. QINSERVICE is
8267 	 * set when the queue is dequeued and
8268 	 * qenable_locked doesn't enqueue a
8269 	 * queue with QINSERVICE set.
8270 	 */
8271 
8272 	ASSERT(!(q->q_flag & QINSERVICE));
8273 	ASSERT((q->q_flag & QENAB));
8274 	mutex_enter(QLOCK(q));
8275 	q->q_flag &= ~QENAB;
8276 	q->q_flag |= QINSERVICE;
8277 	mutex_exit(QLOCK(q));
8278 	runservice(q);
8279 }
8280 
8281 static void
8282 syncq_service(syncq_t *sq)
8283 {
8284 	STRSTAT(syncqservice);
8285 	mutex_enter(SQLOCK(sq));
8286 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8287 	ASSERT(sq->sq_servcount != 0);
8288 	ASSERT(sq->sq_next == NULL);
8289 
8290 	/* if we came here from the background thread, clear the flag */
8291 	if (sq->sq_svcflags & SQ_BGTHREAD)
8292 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8293 
8294 	/* let drain_syncq know that it's being called in the background */
8295 	sq->sq_svcflags |= SQ_SERVICE;
8296 	drain_syncq(sq);
8297 }
8298 
8299 static void
8300 qwriter_outer_service(syncq_t *outer)
8301 {
8302 	/*
8303 	 * Note that SQ_WRITER is used on the outer perimeter
8304 	 * to signal that a qwriter(OUTER) is either investigating
8305 	 * running or that it is actually running a function.
8306 	 */
8307 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8308 
8309 	/*
8310 	 * All inner syncq are empty and have SQ_WRITER set
8311 	 * to block entering the outer perimeter.
8312 	 *
8313 	 * We do not need to explicitly call write_now since
8314 	 * outer_exit does it for us.
8315 	 */
8316 	outer_exit(outer);
8317 }
8318 
8319 static void
8320 mblk_free(mblk_t *mp)
8321 {
8322 	dblk_t *dbp = mp->b_datap;
8323 	frtn_t *frp = dbp->db_frtnp;
8324 
8325 	mp->b_next = NULL;
8326 	if (dbp->db_fthdr != NULL)
8327 		str_ftfree(dbp);
8328 
8329 	ASSERT(dbp->db_fthdr == NULL);
8330 	frp->free_func(frp->free_arg);
8331 	ASSERT(dbp->db_mblk == mp);
8332 
8333 	if (dbp->db_credp != NULL) {
8334 		crfree(dbp->db_credp);
8335 		dbp->db_credp = NULL;
8336 	}
8337 	dbp->db_cpid = -1;
8338 	dbp->db_struioflag = 0;
8339 	dbp->db_struioun.cksum.flags = 0;
8340 
8341 	kmem_cache_free(dbp->db_cache, dbp);
8342 }
8343 
8344 /*
8345  * Background processing of the stream queue list.
8346  */
8347 static void
8348 stream_service(stdata_t *stp)
8349 {
8350 	queue_t *q;
8351 
8352 	mutex_enter(&stp->sd_qlock);
8353 
8354 	STR_SERVICE(stp, q);
8355 
8356 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8357 	stp->sd_servid = NULL;
8358 	cv_signal(&stp->sd_qcv);
8359 	mutex_exit(&stp->sd_qlock);
8360 }
8361 
8362 /*
8363  * Foreground processing of the stream queue list.
8364  */
8365 void
8366 stream_runservice(stdata_t *stp)
8367 {
8368 	queue_t *q;
8369 
8370 	mutex_enter(&stp->sd_qlock);
8371 	STRSTAT(rservice);
8372 	/*
8373 	 * We are going to drain this stream queue list, so qenable_locked will
8374 	 * not schedule it until we finish.
8375 	 */
8376 	stp->sd_svcflags |= STRS_WILLSERVICE;
8377 
8378 	STR_SERVICE(stp, q);
8379 
8380 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8381 	mutex_exit(&stp->sd_qlock);
8382 	/*
8383 	 * Help backup background thread to drain the qhead/qtail list.
8384 	 */
8385 	while (qhead != NULL) {
8386 		STRSTAT(qhelps);
8387 		mutex_enter(&service_queue);
8388 		DQ(q, qhead, qtail, q_link);
8389 		mutex_exit(&service_queue);
8390 		if (q != NULL)
8391 			queue_service(q);
8392 	}
8393 }
8394 
8395 void
8396 stream_willservice(stdata_t *stp)
8397 {
8398 	mutex_enter(&stp->sd_qlock);
8399 	stp->sd_svcflags |= STRS_WILLSERVICE;
8400 	mutex_exit(&stp->sd_qlock);
8401 }
8402 
8403 /*
8404  * Replace the cred currently in the mblk with a different one.
8405  * Also update db_cpid.
8406  */
8407 void
8408 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8409 {
8410 	dblk_t *dbp = mp->b_datap;
8411 	cred_t *ocr = dbp->db_credp;
8412 
8413 	ASSERT(cr != NULL);
8414 
8415 	if (cr != ocr) {
8416 		crhold(dbp->db_credp = cr);
8417 		if (ocr != NULL)
8418 			crfree(ocr);
8419 	}
8420 	/* Don't overwrite with NOPID */
8421 	if (cpid != NOPID)
8422 		dbp->db_cpid = cpid;
8423 }
8424 
8425 /*
8426  * If the src message has a cred, then replace the cred currently in the mblk
8427  * with it.
8428  * Also update db_cpid.
8429  */
8430 void
8431 mblk_copycred(mblk_t *mp, const mblk_t *src)
8432 {
8433 	dblk_t *dbp = mp->b_datap;
8434 	cred_t *cr, *ocr;
8435 	pid_t cpid;
8436 
8437 	cr = msg_getcred(src, &cpid);
8438 	if (cr == NULL)
8439 		return;
8440 
8441 	ocr = dbp->db_credp;
8442 	if (cr != ocr) {
8443 		crhold(dbp->db_credp = cr);
8444 		if (ocr != NULL)
8445 			crfree(ocr);
8446 	}
8447 	/* Don't overwrite with NOPID */
8448 	if (cpid != NOPID)
8449 		dbp->db_cpid = cpid;
8450 }
8451 
8452 int
8453 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8454     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8455     uint32_t flags, int km_flags)
8456 {
8457 	int rc = 0;
8458 
8459 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8460 	if (mp->b_datap->db_type == M_DATA) {
8461 		/* Associate values for M_DATA type */
8462 		DB_CKSUMSTART(mp) = (intptr_t)start;
8463 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8464 		DB_CKSUMEND(mp) = (intptr_t)end;
8465 		DB_CKSUMFLAGS(mp) = flags;
8466 		DB_CKSUM16(mp) = (uint16_t)value;
8467 
8468 	} else {
8469 		pattrinfo_t pa_info;
8470 
8471 		ASSERT(mmd != NULL);
8472 
8473 		pa_info.type = PATTR_HCKSUM;
8474 		pa_info.len = sizeof (pattr_hcksum_t);
8475 
8476 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8477 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8478 
8479 			hck->hcksum_start_offset = start;
8480 			hck->hcksum_stuff_offset = stuff;
8481 			hck->hcksum_end_offset = end;
8482 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8483 			hck->hcksum_flags = flags;
8484 		} else {
8485 			rc = -1;
8486 		}
8487 	}
8488 	return (rc);
8489 }
8490 
8491 void
8492 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8493     uint32_t *start, uint32_t *stuff, uint32_t *end,
8494     uint32_t *value, uint32_t *flags)
8495 {
8496 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8497 	if (mp->b_datap->db_type == M_DATA) {
8498 		if (flags != NULL) {
8499 			*flags = DB_CKSUMFLAGS(mp) & HCK_FLAGS;
8500 			if ((*flags & (HCK_PARTIALCKSUM |
8501 			    HCK_FULLCKSUM)) != 0) {
8502 				if (value != NULL)
8503 					*value = (uint32_t)DB_CKSUM16(mp);
8504 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
8505 					if (start != NULL)
8506 						*start =
8507 						    (uint32_t)DB_CKSUMSTART(mp);
8508 					if (stuff != NULL)
8509 						*stuff =
8510 						    (uint32_t)DB_CKSUMSTUFF(mp);
8511 					if (end != NULL)
8512 						*end =
8513 						    (uint32_t)DB_CKSUMEND(mp);
8514 				}
8515 			}
8516 		}
8517 	} else {
8518 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8519 
8520 		ASSERT(mmd != NULL);
8521 
8522 		/* get hardware checksum attribute */
8523 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8524 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8525 
8526 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8527 			if (flags != NULL)
8528 				*flags = hck->hcksum_flags;
8529 			if (start != NULL)
8530 				*start = hck->hcksum_start_offset;
8531 			if (stuff != NULL)
8532 				*stuff = hck->hcksum_stuff_offset;
8533 			if (end != NULL)
8534 				*end = hck->hcksum_end_offset;
8535 			if (value != NULL)
8536 				*value = (uint32_t)
8537 				    hck->hcksum_cksum_val.inet_cksum;
8538 		}
8539 	}
8540 }
8541 
8542 void
8543 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8544 {
8545 	ASSERT(DB_TYPE(mp) == M_DATA);
8546 	ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8547 
8548 	/* Set the flags */
8549 	DB_LSOFLAGS(mp) |= flags;
8550 	DB_LSOMSS(mp) = mss;
8551 }
8552 
8553 void
8554 lso_info_cleanup(mblk_t *mp)
8555 {
8556 	ASSERT(DB_TYPE(mp) == M_DATA);
8557 
8558 	/* Clear the flags */
8559 	DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8560 	DB_LSOMSS(mp) = 0;
8561 }
8562 
8563 /*
8564  * Checksum buffer *bp for len bytes with psum partial checksum,
8565  * or 0 if none, and return the 16 bit partial checksum.
8566  */
8567 unsigned
8568 bcksum(uchar_t *bp, int len, unsigned int psum)
8569 {
8570 	int odd = len & 1;
8571 	extern unsigned int ip_ocsum();
8572 
8573 	if (((intptr_t)bp & 1) == 0 && !odd) {
8574 		/*
8575 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8576 		 */
8577 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8578 	}
8579 	if (((intptr_t)bp & 1) != 0) {
8580 		/*
8581 		 * Bp isn't 16 bit aligned.
8582 		 */
8583 		unsigned int tsum;
8584 
8585 #ifdef _LITTLE_ENDIAN
8586 		psum += *bp;
8587 #else
8588 		psum += *bp << 8;
8589 #endif
8590 		len--;
8591 		bp++;
8592 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8593 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8594 		if (len & 1) {
8595 			bp += len - 1;
8596 #ifdef _LITTLE_ENDIAN
8597 			psum += *bp << 8;
8598 #else
8599 			psum += *bp;
8600 #endif
8601 		}
8602 	} else {
8603 		/*
8604 		 * Bp is 16 bit aligned.
8605 		 */
8606 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8607 		if (odd) {
8608 			bp += len - 1;
8609 #ifdef _LITTLE_ENDIAN
8610 			psum += *bp;
8611 #else
8612 			psum += *bp << 8;
8613 #endif
8614 		}
8615 	}
8616 	/*
8617 	 * Normalize psum to 16 bits before returning the new partial
8618 	 * checksum. The max psum value before normalization is 0x3FDFE.
8619 	 */
8620 	return ((psum >> 16) + (psum & 0xFFFF));
8621 }
8622 
8623 boolean_t
8624 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8625 {
8626 	boolean_t rc;
8627 
8628 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8629 	if (DB_TYPE(mp) == M_DATA) {
8630 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8631 	} else {
8632 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8633 
8634 		ASSERT(mmd != NULL);
8635 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8636 	}
8637 	return (rc);
8638 }
8639 
8640 void
8641 freemsgchain(mblk_t *mp)
8642 {
8643 	mblk_t	*next;
8644 
8645 	while (mp != NULL) {
8646 		next = mp->b_next;
8647 		mp->b_next = NULL;
8648 
8649 		freemsg(mp);
8650 		mp = next;
8651 	}
8652 }
8653 
8654 mblk_t *
8655 copymsgchain(mblk_t *mp)
8656 {
8657 	mblk_t	*nmp = NULL;
8658 	mblk_t	**nmpp = &nmp;
8659 
8660 	for (; mp != NULL; mp = mp->b_next) {
8661 		if ((*nmpp = copymsg(mp)) == NULL) {
8662 			freemsgchain(nmp);
8663 			return (NULL);
8664 		}
8665 
8666 		nmpp = &((*nmpp)->b_next);
8667 	}
8668 
8669 	return (nmp);
8670 }
8671 
8672 /* NOTE: Do not add code after this point. */
8673 #undef QLOCK
8674 
8675 /*
8676  * Replacement for QLOCK macro for those that can't use it.
8677  */
8678 kmutex_t *
8679 QLOCK(queue_t *q)
8680 {
8681 	return (&(q)->q_lock);
8682 }
8683 
8684 /*
8685  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8686  */
8687 #undef runqueues
8688 void
8689 runqueues(void)
8690 {
8691 }
8692 
8693 #undef queuerun
8694 void
8695 queuerun(void)
8696 {
8697 }
8698 
8699 /*
8700  * Initialize the STR stack instance, which tracks autopush and persistent
8701  * links.
8702  */
8703 /* ARGSUSED */
8704 static void *
8705 str_stack_init(netstackid_t stackid, netstack_t *ns)
8706 {
8707 	str_stack_t	*ss;
8708 	int i;
8709 
8710 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8711 	ss->ss_netstack = ns;
8712 
8713 	/*
8714 	 * set up autopush
8715 	 */
8716 	sad_initspace(ss);
8717 
8718 	/*
8719 	 * set up mux_node structures.
8720 	 */
8721 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8722 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8723 	    ss->ss_devcnt), KM_SLEEP);
8724 	for (i = 0; i < ss->ss_devcnt; i++)
8725 		ss->ss_mux_nodes[i].mn_imaj = i;
8726 	return (ss);
8727 }
8728 
8729 /*
8730  * Note: run at zone shutdown and not destroy so that the PLINKs are
8731  * gone by the time other cleanup happens from the destroy callbacks.
8732  */
8733 static void
8734 str_stack_shutdown(netstackid_t stackid, void *arg)
8735 {
8736 	str_stack_t *ss = (str_stack_t *)arg;
8737 	int i;
8738 	cred_t *cr;
8739 
8740 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8741 	ASSERT(cr != NULL);
8742 
8743 	/* Undo all the I_PLINKs for this zone */
8744 	for (i = 0; i < ss->ss_devcnt; i++) {
8745 		struct mux_edge		*ep;
8746 		ldi_handle_t		lh;
8747 		ldi_ident_t		li;
8748 		int			ret;
8749 		int			rval;
8750 		dev_t			rdev;
8751 
8752 		ep = ss->ss_mux_nodes[i].mn_outp;
8753 		if (ep == NULL)
8754 			continue;
8755 		ret = ldi_ident_from_major((major_t)i, &li);
8756 		if (ret != 0) {
8757 			continue;
8758 		}
8759 		rdev = ep->me_dev;
8760 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8761 		    cr, &lh, li);
8762 		if (ret != 0) {
8763 			ldi_ident_release(li);
8764 			continue;
8765 		}
8766 
8767 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8768 		    cr, &rval);
8769 		if (ret) {
8770 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8771 			ldi_ident_release(li);
8772 			continue;
8773 		}
8774 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8775 
8776 		/* Close layered handles */
8777 		ldi_ident_release(li);
8778 	}
8779 	crfree(cr);
8780 
8781 	sad_freespace(ss);
8782 
8783 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8784 	ss->ss_mux_nodes = NULL;
8785 }
8786 
8787 /*
8788  * Free the structure; str_stack_shutdown did the other cleanup work.
8789  */
8790 /* ARGSUSED */
8791 static void
8792 str_stack_fini(netstackid_t stackid, void *arg)
8793 {
8794 	str_stack_t	*ss = (str_stack_t *)arg;
8795 
8796 	kmem_free(ss, sizeof (*ss));
8797 }
8798