xref: /titanic_51/usr/src/uts/common/os/strsubr.c (revision 81d9f076db88c1f40c85831ce1ebb444a209c5a8)
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  */
29 
30 #include <sys/types.h>
31 #include <sys/sysmacros.h>
32 #include <sys/param.h>
33 #include <sys/errno.h>
34 #include <sys/signal.h>
35 #include <sys/proc.h>
36 #include <sys/conf.h>
37 #include <sys/cred.h>
38 #include <sys/user.h>
39 #include <sys/vnode.h>
40 #include <sys/file.h>
41 #include <sys/session.h>
42 #include <sys/stream.h>
43 #include <sys/strsubr.h>
44 #include <sys/stropts.h>
45 #include <sys/poll.h>
46 #include <sys/systm.h>
47 #include <sys/cpuvar.h>
48 #include <sys/uio.h>
49 #include <sys/cmn_err.h>
50 #include <sys/priocntl.h>
51 #include <sys/procset.h>
52 #include <sys/vmem.h>
53 #include <sys/bitmap.h>
54 #include <sys/kmem.h>
55 #include <sys/siginfo.h>
56 #include <sys/vtrace.h>
57 #include <sys/callb.h>
58 #include <sys/debug.h>
59 #include <sys/modctl.h>
60 #include <sys/vmsystm.h>
61 #include <vm/page.h>
62 #include <sys/atomic.h>
63 #include <sys/suntpi.h>
64 #include <sys/strlog.h>
65 #include <sys/promif.h>
66 #include <sys/project.h>
67 #include <sys/vm.h>
68 #include <sys/taskq.h>
69 #include <sys/sunddi.h>
70 #include <sys/sunldi_impl.h>
71 #include <sys/strsun.h>
72 #include <sys/isa_defs.h>
73 #include <sys/multidata.h>
74 #include <sys/pattr.h>
75 #include <sys/strft.h>
76 #include <sys/fs/snode.h>
77 #include <sys/zone.h>
78 #include <sys/open.h>
79 #include <sys/sunldi.h>
80 #include <sys/sad.h>
81 #include <sys/netstack.h>
82 
83 #define	O_SAMESTR(q)	(((q)->q_next) && \
84 	(((q)->q_flag & QREADR) == ((q)->q_next->q_flag & QREADR)))
85 
86 /*
87  * WARNING:
88  * The variables and routines in this file are private, belonging
89  * to the STREAMS subsystem. These should not be used by modules
90  * or drivers. Compatibility will not be guaranteed.
91  */
92 
93 /*
94  * Id value used to distinguish between different multiplexor links.
95  */
96 static int32_t lnk_id = 0;
97 
98 #define	STREAMS_LOPRI MINCLSYSPRI
99 static pri_t streams_lopri = STREAMS_LOPRI;
100 
101 #define	STRSTAT(x)	(str_statistics.x.value.ui64++)
102 typedef struct str_stat {
103 	kstat_named_t	sqenables;
104 	kstat_named_t	stenables;
105 	kstat_named_t	syncqservice;
106 	kstat_named_t	freebs;
107 	kstat_named_t	qwr_outer;
108 	kstat_named_t	rservice;
109 	kstat_named_t	strwaits;
110 	kstat_named_t	taskqfails;
111 	kstat_named_t	bufcalls;
112 	kstat_named_t	qhelps;
113 	kstat_named_t	qremoved;
114 	kstat_named_t	sqremoved;
115 	kstat_named_t	bcwaits;
116 	kstat_named_t	sqtoomany;
117 } str_stat_t;
118 
119 static str_stat_t str_statistics = {
120 	{ "sqenables",		KSTAT_DATA_UINT64 },
121 	{ "stenables",		KSTAT_DATA_UINT64 },
122 	{ "syncqservice",	KSTAT_DATA_UINT64 },
123 	{ "freebs",		KSTAT_DATA_UINT64 },
124 	{ "qwr_outer",		KSTAT_DATA_UINT64 },
125 	{ "rservice",		KSTAT_DATA_UINT64 },
126 	{ "strwaits",		KSTAT_DATA_UINT64 },
127 	{ "taskqfails",		KSTAT_DATA_UINT64 },
128 	{ "bufcalls",		KSTAT_DATA_UINT64 },
129 	{ "qhelps",		KSTAT_DATA_UINT64 },
130 	{ "qremoved",		KSTAT_DATA_UINT64 },
131 	{ "sqremoved",		KSTAT_DATA_UINT64 },
132 	{ "bcwaits",		KSTAT_DATA_UINT64 },
133 	{ "sqtoomany",		KSTAT_DATA_UINT64 },
134 };
135 
136 static kstat_t *str_kstat;
137 
138 /*
139  * qrunflag was used previously to control background scheduling of queues. It
140  * is not used anymore, but kept here in case some module still wants to access
141  * it via qready() and setqsched macros.
142  */
143 char qrunflag;			/*  Unused */
144 
145 /*
146  * Most of the streams scheduling is done via task queues. Task queues may fail
147  * for non-sleep dispatches, so there are two backup threads servicing failed
148  * requests for queues and syncqs. Both of these threads also service failed
149  * dispatches freebs requests. Queues are put in the list specified by `qhead'
150  * and `qtail' pointers, syncqs use `sqhead' and `sqtail' pointers and freebs
151  * requests are put into `freebs_list' which has no tail pointer. All three
152  * lists are protected by a single `service_queue' lock and use
153  * `services_to_run' condition variable for signaling background threads. Use of
154  * a single lock should not be a problem because it is only used under heavy
155  * loads when task queues start to fail and at that time it may be a good idea
156  * to throttle scheduling requests.
157  *
158  * NOTE: queues and syncqs should be scheduled by two separate threads because
159  * queue servicing may be blocked waiting for a syncq which may be also
160  * scheduled for background execution. This may create a deadlock when only one
161  * thread is used for both.
162  */
163 
164 static taskq_t *streams_taskq;		/* Used for most STREAMS scheduling */
165 
166 static kmutex_t service_queue;		/* protects all of servicing vars */
167 static kcondvar_t services_to_run;	/* wake up background service thread */
168 static kcondvar_t syncqs_to_run;	/* wake up background service thread */
169 
170 /*
171  * List of queues scheduled for background processing due to lack of resources
172  * in the task queues. Protected by service_queue lock;
173  */
174 static struct queue *qhead;
175 static struct queue *qtail;
176 
177 /*
178  * Same list for syncqs
179  */
180 static syncq_t *sqhead;
181 static syncq_t *sqtail;
182 
183 static mblk_t *freebs_list;	/* list of buffers to free */
184 
185 /*
186  * Backup threads for servicing queues and syncqs
187  */
188 kthread_t *streams_qbkgrnd_thread;
189 kthread_t *streams_sqbkgrnd_thread;
190 
191 /*
192  * Bufcalls related variables.
193  */
194 struct bclist	strbcalls;	/* list of waiting bufcalls */
195 kmutex_t	strbcall_lock;	/* protects bufcall list (strbcalls) */
196 kcondvar_t	strbcall_cv;	/* Signaling when a bufcall is added */
197 kmutex_t	bcall_monitor;	/* sleep/wakeup style monitor */
198 kcondvar_t	bcall_cv;	/* wait 'till executing bufcall completes */
199 kthread_t	*bc_bkgrnd_thread; /* Thread to service bufcall requests */
200 
201 kmutex_t	strresources;	/* protects global resources */
202 kmutex_t	muxifier;	/* single-threads multiplexor creation */
203 
204 static void	*str_stack_init(netstackid_t stackid, netstack_t *ns);
205 static void	str_stack_shutdown(netstackid_t stackid, void *arg);
206 static void	str_stack_fini(netstackid_t stackid, void *arg);
207 
208 /*
209  * run_queues is no longer used, but is kept in case some 3rd party
210  * module/driver decides to use it.
211  */
212 int run_queues = 0;
213 
214 /*
215  * sq_max_size is the depth of the syncq (in number of messages) before
216  * qfill_syncq() starts QFULL'ing destination queues. As its primary
217  * consumer - IP is no longer D_MTPERMOD, but there may be other
218  * modules/drivers depend on this syncq flow control, we prefer to
219  * choose a large number as the default value. For potential
220  * performance gain, this value is tunable in /etc/system.
221  */
222 int sq_max_size = 10000;
223 
224 /*
225  * The number of ciputctrl structures per syncq and stream we create when
226  * needed.
227  */
228 int n_ciputctrl;
229 int max_n_ciputctrl = 16;
230 /*
231  * If n_ciputctrl is < min_n_ciputctrl don't even create ciputctrl_cache.
232  */
233 int min_n_ciputctrl = 2;
234 
235 /*
236  * Per-driver/module syncqs
237  * ========================
238  *
239  * For drivers/modules that use PERMOD or outer syncqs we keep a list of
240  * perdm structures, new entries being added (and new syncqs allocated) when
241  * setq() encounters a module/driver with a streamtab that it hasn't seen
242  * before.
243  * The reason for this mechanism is that some modules and drivers share a
244  * common streamtab and it is necessary for those modules and drivers to also
245  * share a common PERMOD syncq.
246  *
247  * perdm_list --> dm_str == streamtab_1
248  *                dm_sq == syncq_1
249  *                dm_ref
250  *                dm_next --> dm_str == streamtab_2
251  *                            dm_sq == syncq_2
252  *                            dm_ref
253  *                            dm_next --> ... NULL
254  *
255  * The dm_ref field is incremented for each new driver/module that takes
256  * a reference to the perdm structure and hence shares the syncq.
257  * References are held in the fmodsw_impl_t structure for each STREAMS module
258  * or the dev_impl array (indexed by device major number) for each driver.
259  *
260  * perdm_list -> [dm_ref == 1] -> [dm_ref == 2] -> [dm_ref == 1] -> NULL
261  *		     ^                 ^ ^               ^
262  *                   |  ______________/  |               |
263  *                   | /                 |               |
264  * dev_impl:     ...|x|y|...          module A	      module B
265  *
266  * When a module/driver is unloaded the reference count is decremented and,
267  * when it falls to zero, the perdm structure is removed from the list and
268  * the syncq is freed (see rele_dm()).
269  */
270 perdm_t *perdm_list = NULL;
271 static krwlock_t perdm_rwlock;
272 cdevsw_impl_t *devimpl;
273 
274 extern struct qinit strdata;
275 extern struct qinit stwdata;
276 
277 static void runservice(queue_t *);
278 static void streams_bufcall_service(void);
279 static void streams_qbkgrnd_service(void);
280 static void streams_sqbkgrnd_service(void);
281 static syncq_t *new_syncq(void);
282 static void free_syncq(syncq_t *);
283 static void outer_insert(syncq_t *, syncq_t *);
284 static void outer_remove(syncq_t *, syncq_t *);
285 static void write_now(syncq_t *);
286 static void clr_qfull(queue_t *);
287 static void runbufcalls(void);
288 static void sqenable(syncq_t *);
289 static void sqfill_events(syncq_t *, queue_t *, mblk_t *, void (*)());
290 static void wait_q_syncq(queue_t *);
291 static void backenable_insertedq(queue_t *);
292 
293 static void queue_service(queue_t *);
294 static void stream_service(stdata_t *);
295 static void syncq_service(syncq_t *);
296 static void qwriter_outer_service(syncq_t *);
297 static void mblk_free(mblk_t *);
298 #ifdef DEBUG
299 static int qprocsareon(queue_t *);
300 #endif
301 
302 static void set_nfsrv_ptr(queue_t *, queue_t *, queue_t *, queue_t *);
303 static void reset_nfsrv_ptr(queue_t *, queue_t *);
304 void set_qfull(queue_t *);
305 
306 static void sq_run_events(syncq_t *);
307 static int propagate_syncq(queue_t *);
308 
309 static void	blocksq(syncq_t *, ushort_t, int);
310 static void	unblocksq(syncq_t *, ushort_t, int);
311 static int	dropsq(syncq_t *, uint16_t);
312 static void	emptysq(syncq_t *);
313 static sqlist_t *sqlist_alloc(struct stdata *, int);
314 static void	sqlist_free(sqlist_t *);
315 static sqlist_t	*sqlist_build(queue_t *, struct stdata *, boolean_t);
316 static void	sqlist_insert(sqlist_t *, syncq_t *);
317 static void	sqlist_insertall(sqlist_t *, queue_t *);
318 
319 static void	strsetuio(stdata_t *);
320 
321 struct kmem_cache *stream_head_cache;
322 struct kmem_cache *queue_cache;
323 struct kmem_cache *syncq_cache;
324 struct kmem_cache *qband_cache;
325 struct kmem_cache *linkinfo_cache;
326 struct kmem_cache *ciputctrl_cache = NULL;
327 
328 static linkinfo_t *linkinfo_list;
329 
330 /* Global esballoc throttling queue */
331 static esb_queue_t	system_esbq;
332 
333 /*
334  * esballoc tunable parameters.
335  */
336 int		esbq_max_qlen = 0x16;	/* throttled queue length */
337 clock_t		esbq_timeout = 0x8;	/* timeout to process esb queue */
338 
339 /*
340  * Routines to handle esballoc queueing.
341  */
342 static void esballoc_process_queue(esb_queue_t *);
343 static void esballoc_enqueue_mblk(mblk_t *);
344 static void esballoc_timer(void *);
345 static void esballoc_set_timer(esb_queue_t *, clock_t);
346 static void esballoc_mblk_free(mblk_t *);
347 
348 /*
349  *  Qinit structure and Module_info structures
350  *	for passthru read and write queues
351  */
352 
353 static void pass_wput(queue_t *, mblk_t *);
354 static queue_t *link_addpassthru(stdata_t *);
355 static void link_rempassthru(queue_t *);
356 
357 struct  module_info passthru_info = {
358 	0,
359 	"passthru",
360 	0,
361 	INFPSZ,
362 	STRHIGH,
363 	STRLOW
364 };
365 
366 struct  qinit passthru_rinit = {
367 	(int (*)())putnext,
368 	NULL,
369 	NULL,
370 	NULL,
371 	NULL,
372 	&passthru_info,
373 	NULL
374 };
375 
376 struct  qinit passthru_winit = {
377 	(int (*)()) pass_wput,
378 	NULL,
379 	NULL,
380 	NULL,
381 	NULL,
382 	&passthru_info,
383 	NULL
384 };
385 
386 /*
387  * Verify correctness of list head/tail pointers.
388  */
389 #define	LISTCHECK(head, tail, link) {				\
390 	EQUIV(head, tail);					\
391 	IMPLY(tail != NULL, tail->link == NULL);		\
392 }
393 
394 /*
395  * Enqueue a list element `el' in the end of a list denoted by `head' and `tail'
396  * using a `link' field.
397  */
398 #define	ENQUEUE(el, head, tail, link) {				\
399 	ASSERT(el->link == NULL);				\
400 	LISTCHECK(head, tail, link);				\
401 	if (head == NULL)					\
402 		head = el;					\
403 	else							\
404 		tail->link = el;				\
405 	tail = el;						\
406 }
407 
408 /*
409  * Dequeue the first element of the list denoted by `head' and `tail' pointers
410  * using a `link' field and put result into `el'.
411  */
412 #define	DQ(el, head, tail, link) {				\
413 	LISTCHECK(head, tail, link);				\
414 	el = head;						\
415 	if (head != NULL) {					\
416 		head = head->link;				\
417 		if (head == NULL)				\
418 			tail = NULL;				\
419 		el->link = NULL;				\
420 	}							\
421 }
422 
423 /*
424  * Remove `el' from the list using `chase' and `curr' pointers and return result
425  * in `succeed'.
426  */
427 #define	RMQ(el, head, tail, link, chase, curr, succeed) {	\
428 	LISTCHECK(head, tail, link);				\
429 	chase = NULL;						\
430 	succeed = 0;						\
431 	for (curr = head; (curr != el) && (curr != NULL); curr = curr->link) \
432 		chase = curr;					\
433 	if (curr != NULL) {					\
434 		succeed = 1;					\
435 		ASSERT(curr == el);				\
436 		if (chase != NULL)				\
437 			chase->link = curr->link;		\
438 		else						\
439 			head = curr->link;			\
440 		curr->link = NULL;				\
441 		if (curr == tail)				\
442 			tail = chase;				\
443 	}							\
444 	LISTCHECK(head, tail, link);				\
445 }
446 
447 /* Handling of delayed messages on the inner syncq. */
448 
449 /*
450  * DEBUG versions should use function versions (to simplify tracing) and
451  * non-DEBUG kernels should use macro versions.
452  */
453 
454 /*
455  * Put a queue on the syncq list of queues.
456  * Assumes SQLOCK held.
457  */
458 #define	SQPUT_Q(sq, qp)							\
459 {									\
460 	ASSERT(MUTEX_HELD(SQLOCK(sq)));					\
461 	if (!(qp->q_sqflags & Q_SQQUEUED)) {				\
462 		/* The queue should not be linked anywhere */		\
463 		ASSERT((qp->q_sqprev == NULL) && (qp->q_sqnext == NULL)); \
464 		/* Head and tail may only be NULL simultaneously */	\
465 		EQUIV(sq->sq_head, sq->sq_tail);			\
466 		/* Queue may be only enqueued on its syncq */		\
467 		ASSERT(sq == qp->q_syncq);				\
468 		/* Check the correctness of SQ_MESSAGES flag */		\
469 		EQUIV(sq->sq_head, (sq->sq_flags & SQ_MESSAGES));	\
470 		/* Sanity check first/last elements of the list */	\
471 		IMPLY(sq->sq_head != NULL, sq->sq_head->q_sqprev == NULL);\
472 		IMPLY(sq->sq_tail != NULL, sq->sq_tail->q_sqnext == NULL);\
473 		/*							\
474 		 * Sanity check of priority field: empty queue should	\
475 		 * have zero priority					\
476 		 * and nqueues equal to zero.				\
477 		 */							\
478 		IMPLY(sq->sq_head == NULL, sq->sq_pri == 0);		\
479 		/* Sanity check of sq_nqueues field */			\
480 		EQUIV(sq->sq_head, sq->sq_nqueues);			\
481 		if (sq->sq_head == NULL) {				\
482 			sq->sq_head = sq->sq_tail = qp;			\
483 			sq->sq_flags |= SQ_MESSAGES;			\
484 		} else if (qp->q_spri == 0) {				\
485 			qp->q_sqprev = sq->sq_tail;			\
486 			sq->sq_tail->q_sqnext = qp;			\
487 			sq->sq_tail = qp;				\
488 		} else {						\
489 			/*						\
490 			 * Put this queue in priority order: higher	\
491 			 * priority gets closer to the head.		\
492 			 */						\
493 			queue_t **qpp = &sq->sq_tail;			\
494 			queue_t *qnext = NULL;				\
495 									\
496 			while (*qpp != NULL && qp->q_spri > (*qpp)->q_spri) { \
497 				qnext = *qpp;				\
498 				qpp = &(*qpp)->q_sqprev;		\
499 			}						\
500 			qp->q_sqnext = qnext;				\
501 			qp->q_sqprev = *qpp;				\
502 			if (*qpp != NULL) {				\
503 				(*qpp)->q_sqnext = qp;			\
504 			} else {					\
505 				sq->sq_head = qp;			\
506 				sq->sq_pri = sq->sq_head->q_spri;	\
507 			}						\
508 			*qpp = qp;					\
509 		}							\
510 		qp->q_sqflags |= Q_SQQUEUED;				\
511 		qp->q_sqtstamp = ddi_get_lbolt();			\
512 		sq->sq_nqueues++;					\
513 	}								\
514 }
515 
516 /*
517  * Remove a queue from the syncq list
518  * Assumes SQLOCK held.
519  */
520 #define	SQRM_Q(sq, qp)							\
521 	{								\
522 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
523 		ASSERT(qp->q_sqflags & Q_SQQUEUED);			\
524 		ASSERT(sq->sq_head != NULL && sq->sq_tail != NULL);	\
525 		ASSERT((sq->sq_flags & SQ_MESSAGES) != 0);		\
526 		/* Check that the queue is actually in the list */	\
527 		ASSERT(qp->q_sqnext != NULL || sq->sq_tail == qp);	\
528 		ASSERT(qp->q_sqprev != NULL || sq->sq_head == qp);	\
529 		ASSERT(sq->sq_nqueues != 0);				\
530 		if (qp->q_sqprev == NULL) {				\
531 			/* First queue on list, make head q_sqnext */	\
532 			sq->sq_head = qp->q_sqnext;			\
533 		} else {						\
534 			/* Make prev->next == next */			\
535 			qp->q_sqprev->q_sqnext = qp->q_sqnext;		\
536 		}							\
537 		if (qp->q_sqnext == NULL) {				\
538 			/* Last queue on list, make tail sqprev */	\
539 			sq->sq_tail = qp->q_sqprev;			\
540 		} else {						\
541 			/* Make next->prev == prev */			\
542 			qp->q_sqnext->q_sqprev = qp->q_sqprev;		\
543 		}							\
544 		/* clear out references on this queue */		\
545 		qp->q_sqprev = qp->q_sqnext = NULL;			\
546 		qp->q_sqflags &= ~Q_SQQUEUED;				\
547 		/* If there is nothing queued, clear SQ_MESSAGES */	\
548 		if (sq->sq_head != NULL) {				\
549 			sq->sq_pri = sq->sq_head->q_spri;		\
550 		} else	{						\
551 			sq->sq_flags &= ~SQ_MESSAGES;			\
552 			sq->sq_pri = 0;					\
553 		}							\
554 		sq->sq_nqueues--;					\
555 		ASSERT(sq->sq_head != NULL || sq->sq_evhead != NULL ||	\
556 		    (sq->sq_flags & SQ_QUEUED) == 0);			\
557 	}
558 
559 /* Hide the definition from the header file. */
560 #ifdef SQPUT_MP
561 #undef SQPUT_MP
562 #endif
563 
564 /*
565  * Put a message on the queue syncq.
566  * Assumes QLOCK held.
567  */
568 #define	SQPUT_MP(qp, mp)						\
569 	{								\
570 		ASSERT(MUTEX_HELD(QLOCK(qp)));				\
571 		ASSERT(qp->q_sqhead == NULL ||				\
572 		    (qp->q_sqtail != NULL &&				\
573 		    qp->q_sqtail->b_next == NULL));			\
574 		qp->q_syncqmsgs++;					\
575 		ASSERT(qp->q_syncqmsgs != 0);	/* Wraparound */	\
576 		if (qp->q_sqhead == NULL) {				\
577 			qp->q_sqhead = qp->q_sqtail = mp;		\
578 		} else {						\
579 			qp->q_sqtail->b_next = mp;			\
580 			qp->q_sqtail = mp;				\
581 		}							\
582 		ASSERT(qp->q_syncqmsgs > 0);				\
583 		set_qfull(qp);						\
584 	}
585 
586 #define	SQ_PUTCOUNT_SETFAST_LOCKED(sq) {				\
587 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
588 		if ((sq)->sq_ciputctrl != NULL) {			\
589 			int i;						\
590 			int nlocks = (sq)->sq_nciputctrl;		\
591 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
592 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
593 			for (i = 0; i <= nlocks; i++) {			\
594 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
595 				cip[i].ciputctrl_count |= SQ_FASTPUT;	\
596 			}						\
597 		}							\
598 	}
599 
600 
601 #define	SQ_PUTCOUNT_CLRFAST_LOCKED(sq) {				\
602 		ASSERT(MUTEX_HELD(SQLOCK(sq)));				\
603 		if ((sq)->sq_ciputctrl != NULL) {			\
604 			int i;						\
605 			int nlocks = (sq)->sq_nciputctrl;		\
606 			ciputctrl_t *cip = (sq)->sq_ciputctrl;		\
607 			ASSERT((sq)->sq_type & SQ_CIPUT);		\
608 			for (i = 0; i <= nlocks; i++) {			\
609 				ASSERT(MUTEX_HELD(&cip[i].ciputctrl_lock)); \
610 				cip[i].ciputctrl_count &= ~SQ_FASTPUT;	\
611 			}						\
612 		}							\
613 	}
614 
615 /*
616  * Run service procedures for all queues in the stream head.
617  */
618 #define	STR_SERVICE(stp, q) {						\
619 	ASSERT(MUTEX_HELD(&stp->sd_qlock));				\
620 	while (stp->sd_qhead != NULL) {					\
621 		DQ(q, stp->sd_qhead, stp->sd_qtail, q_link);		\
622 		ASSERT(stp->sd_nqueues > 0);				\
623 		stp->sd_nqueues--;					\
624 		ASSERT(!(q->q_flag & QINSERVICE));			\
625 		mutex_exit(&stp->sd_qlock);				\
626 		queue_service(q);					\
627 		mutex_enter(&stp->sd_qlock);				\
628 	}								\
629 	ASSERT(stp->sd_nqueues == 0);					\
630 	ASSERT((stp->sd_qhead == NULL) && (stp->sd_qtail == NULL));	\
631 }
632 
633 /*
634  * Constructor/destructor routines for the stream head cache
635  */
636 /* ARGSUSED */
637 static int
638 stream_head_constructor(void *buf, void *cdrarg, int kmflags)
639 {
640 	stdata_t *stp = buf;
641 
642 	mutex_init(&stp->sd_lock, NULL, MUTEX_DEFAULT, NULL);
643 	mutex_init(&stp->sd_reflock, NULL, MUTEX_DEFAULT, NULL);
644 	mutex_init(&stp->sd_qlock, NULL, MUTEX_DEFAULT, NULL);
645 	cv_init(&stp->sd_monitor, NULL, CV_DEFAULT, NULL);
646 	cv_init(&stp->sd_iocmonitor, NULL, CV_DEFAULT, NULL);
647 	cv_init(&stp->sd_refmonitor, NULL, CV_DEFAULT, NULL);
648 	cv_init(&stp->sd_qcv, NULL, CV_DEFAULT, NULL);
649 	cv_init(&stp->sd_zcopy_wait, NULL, CV_DEFAULT, NULL);
650 	stp->sd_wrq = NULL;
651 
652 	return (0);
653 }
654 
655 /* ARGSUSED */
656 static void
657 stream_head_destructor(void *buf, void *cdrarg)
658 {
659 	stdata_t *stp = buf;
660 
661 	mutex_destroy(&stp->sd_lock);
662 	mutex_destroy(&stp->sd_reflock);
663 	mutex_destroy(&stp->sd_qlock);
664 	cv_destroy(&stp->sd_monitor);
665 	cv_destroy(&stp->sd_iocmonitor);
666 	cv_destroy(&stp->sd_refmonitor);
667 	cv_destroy(&stp->sd_qcv);
668 	cv_destroy(&stp->sd_zcopy_wait);
669 }
670 
671 /*
672  * Constructor/destructor routines for the queue cache
673  */
674 /* ARGSUSED */
675 static int
676 queue_constructor(void *buf, void *cdrarg, int kmflags)
677 {
678 	queinfo_t *qip = buf;
679 	queue_t *qp = &qip->qu_rqueue;
680 	queue_t *wqp = &qip->qu_wqueue;
681 	syncq_t	*sq = &qip->qu_syncq;
682 
683 	qp->q_first = NULL;
684 	qp->q_link = NULL;
685 	qp->q_count = 0;
686 	qp->q_mblkcnt = 0;
687 	qp->q_sqhead = NULL;
688 	qp->q_sqtail = NULL;
689 	qp->q_sqnext = NULL;
690 	qp->q_sqprev = NULL;
691 	qp->q_sqflags = 0;
692 	qp->q_rwcnt = 0;
693 	qp->q_spri = 0;
694 
695 	mutex_init(QLOCK(qp), NULL, MUTEX_DEFAULT, NULL);
696 	cv_init(&qp->q_wait, NULL, CV_DEFAULT, NULL);
697 
698 	wqp->q_first = NULL;
699 	wqp->q_link = NULL;
700 	wqp->q_count = 0;
701 	wqp->q_mblkcnt = 0;
702 	wqp->q_sqhead = NULL;
703 	wqp->q_sqtail = NULL;
704 	wqp->q_sqnext = NULL;
705 	wqp->q_sqprev = NULL;
706 	wqp->q_sqflags = 0;
707 	wqp->q_rwcnt = 0;
708 	wqp->q_spri = 0;
709 
710 	mutex_init(QLOCK(wqp), NULL, MUTEX_DEFAULT, NULL);
711 	cv_init(&wqp->q_wait, NULL, CV_DEFAULT, NULL);
712 
713 	sq->sq_head = NULL;
714 	sq->sq_tail = NULL;
715 	sq->sq_evhead = NULL;
716 	sq->sq_evtail = NULL;
717 	sq->sq_callbpend = NULL;
718 	sq->sq_outer = NULL;
719 	sq->sq_onext = NULL;
720 	sq->sq_oprev = NULL;
721 	sq->sq_next = NULL;
722 	sq->sq_svcflags = 0;
723 	sq->sq_servcount = 0;
724 	sq->sq_needexcl = 0;
725 	sq->sq_nqueues = 0;
726 	sq->sq_pri = 0;
727 
728 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
729 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
730 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
731 
732 	return (0);
733 }
734 
735 /* ARGSUSED */
736 static void
737 queue_destructor(void *buf, void *cdrarg)
738 {
739 	queinfo_t *qip = buf;
740 	queue_t *qp = &qip->qu_rqueue;
741 	queue_t *wqp = &qip->qu_wqueue;
742 	syncq_t	*sq = &qip->qu_syncq;
743 
744 	ASSERT(qp->q_sqhead == NULL);
745 	ASSERT(wqp->q_sqhead == NULL);
746 	ASSERT(qp->q_sqnext == NULL);
747 	ASSERT(wqp->q_sqnext == NULL);
748 	ASSERT(qp->q_rwcnt == 0);
749 	ASSERT(wqp->q_rwcnt == 0);
750 
751 	mutex_destroy(&qp->q_lock);
752 	cv_destroy(&qp->q_wait);
753 
754 	mutex_destroy(&wqp->q_lock);
755 	cv_destroy(&wqp->q_wait);
756 
757 	mutex_destroy(&sq->sq_lock);
758 	cv_destroy(&sq->sq_wait);
759 	cv_destroy(&sq->sq_exitwait);
760 }
761 
762 /*
763  * Constructor/destructor routines for the syncq cache
764  */
765 /* ARGSUSED */
766 static int
767 syncq_constructor(void *buf, void *cdrarg, int kmflags)
768 {
769 	syncq_t	*sq = buf;
770 
771 	bzero(buf, sizeof (syncq_t));
772 
773 	mutex_init(&sq->sq_lock, NULL, MUTEX_DEFAULT, NULL);
774 	cv_init(&sq->sq_wait, NULL, CV_DEFAULT, NULL);
775 	cv_init(&sq->sq_exitwait, NULL, CV_DEFAULT, NULL);
776 
777 	return (0);
778 }
779 
780 /* ARGSUSED */
781 static void
782 syncq_destructor(void *buf, void *cdrarg)
783 {
784 	syncq_t	*sq = buf;
785 
786 	ASSERT(sq->sq_head == NULL);
787 	ASSERT(sq->sq_tail == NULL);
788 	ASSERT(sq->sq_evhead == NULL);
789 	ASSERT(sq->sq_evtail == NULL);
790 	ASSERT(sq->sq_callbpend == NULL);
791 	ASSERT(sq->sq_callbflags == 0);
792 	ASSERT(sq->sq_outer == NULL);
793 	ASSERT(sq->sq_onext == NULL);
794 	ASSERT(sq->sq_oprev == NULL);
795 	ASSERT(sq->sq_next == NULL);
796 	ASSERT(sq->sq_needexcl == 0);
797 	ASSERT(sq->sq_svcflags == 0);
798 	ASSERT(sq->sq_servcount == 0);
799 	ASSERT(sq->sq_nqueues == 0);
800 	ASSERT(sq->sq_pri == 0);
801 	ASSERT(sq->sq_count == 0);
802 	ASSERT(sq->sq_rmqcount == 0);
803 	ASSERT(sq->sq_cancelid == 0);
804 	ASSERT(sq->sq_ciputctrl == NULL);
805 	ASSERT(sq->sq_nciputctrl == 0);
806 	ASSERT(sq->sq_type == 0);
807 	ASSERT(sq->sq_flags == 0);
808 
809 	mutex_destroy(&sq->sq_lock);
810 	cv_destroy(&sq->sq_wait);
811 	cv_destroy(&sq->sq_exitwait);
812 }
813 
814 /* ARGSUSED */
815 static int
816 ciputctrl_constructor(void *buf, void *cdrarg, int kmflags)
817 {
818 	ciputctrl_t *cip = buf;
819 	int i;
820 
821 	for (i = 0; i < n_ciputctrl; i++) {
822 		cip[i].ciputctrl_count = SQ_FASTPUT;
823 		mutex_init(&cip[i].ciputctrl_lock, NULL, MUTEX_DEFAULT, NULL);
824 	}
825 
826 	return (0);
827 }
828 
829 /* ARGSUSED */
830 static void
831 ciputctrl_destructor(void *buf, void *cdrarg)
832 {
833 	ciputctrl_t *cip = buf;
834 	int i;
835 
836 	for (i = 0; i < n_ciputctrl; i++) {
837 		ASSERT(cip[i].ciputctrl_count & SQ_FASTPUT);
838 		mutex_destroy(&cip[i].ciputctrl_lock);
839 	}
840 }
841 
842 /*
843  * Init routine run from main at boot time.
844  */
845 void
846 strinit(void)
847 {
848 	int ncpus = ((boot_max_ncpus == -1) ? max_ncpus : boot_max_ncpus);
849 
850 	stream_head_cache = kmem_cache_create("stream_head_cache",
851 	    sizeof (stdata_t), 0,
852 	    stream_head_constructor, stream_head_destructor, NULL,
853 	    NULL, NULL, 0);
854 
855 	queue_cache = kmem_cache_create("queue_cache", sizeof (queinfo_t), 0,
856 	    queue_constructor, queue_destructor, NULL, NULL, NULL, 0);
857 
858 	syncq_cache = kmem_cache_create("syncq_cache", sizeof (syncq_t), 0,
859 	    syncq_constructor, syncq_destructor, NULL, NULL, NULL, 0);
860 
861 	qband_cache = kmem_cache_create("qband_cache",
862 	    sizeof (qband_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
863 
864 	linkinfo_cache = kmem_cache_create("linkinfo_cache",
865 	    sizeof (linkinfo_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
866 
867 	n_ciputctrl = ncpus;
868 	n_ciputctrl = 1 << highbit(n_ciputctrl - 1);
869 	ASSERT(n_ciputctrl >= 1);
870 	n_ciputctrl = MIN(n_ciputctrl, max_n_ciputctrl);
871 	if (n_ciputctrl >= min_n_ciputctrl) {
872 		ciputctrl_cache = kmem_cache_create("ciputctrl_cache",
873 		    sizeof (ciputctrl_t) * n_ciputctrl,
874 		    sizeof (ciputctrl_t), ciputctrl_constructor,
875 		    ciputctrl_destructor, NULL, NULL, NULL, 0);
876 	}
877 
878 	streams_taskq = system_taskq;
879 
880 	if (streams_taskq == NULL)
881 		panic("strinit: no memory for streams taskq!");
882 
883 	bc_bkgrnd_thread = thread_create(NULL, 0,
884 	    streams_bufcall_service, NULL, 0, &p0, TS_RUN, streams_lopri);
885 
886 	streams_qbkgrnd_thread = thread_create(NULL, 0,
887 	    streams_qbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
888 
889 	streams_sqbkgrnd_thread = thread_create(NULL, 0,
890 	    streams_sqbkgrnd_service, NULL, 0, &p0, TS_RUN, streams_lopri);
891 
892 	/*
893 	 * Create STREAMS kstats.
894 	 */
895 	str_kstat = kstat_create("streams", 0, "strstat",
896 	    "net", KSTAT_TYPE_NAMED,
897 	    sizeof (str_statistics) / sizeof (kstat_named_t),
898 	    KSTAT_FLAG_VIRTUAL);
899 
900 	if (str_kstat != NULL) {
901 		str_kstat->ks_data = &str_statistics;
902 		kstat_install(str_kstat);
903 	}
904 
905 	/*
906 	 * TPI support routine initialisation.
907 	 */
908 	tpi_init();
909 
910 	/*
911 	 * Handle to have autopush and persistent link information per
912 	 * zone.
913 	 * Note: uses shutdown hook instead of destroy hook so that the
914 	 * persistent links can be torn down before the destroy hooks
915 	 * in the TCP/IP stack are called.
916 	 */
917 	netstack_register(NS_STR, str_stack_init, str_stack_shutdown,
918 	    str_stack_fini);
919 }
920 
921 void
922 str_sendsig(vnode_t *vp, int event, uchar_t band, int error)
923 {
924 	struct stdata *stp;
925 
926 	ASSERT(vp->v_stream);
927 	stp = vp->v_stream;
928 	/* Have to hold sd_lock to prevent siglist from changing */
929 	mutex_enter(&stp->sd_lock);
930 	if (stp->sd_sigflags & event)
931 		strsendsig(stp->sd_siglist, event, band, error);
932 	mutex_exit(&stp->sd_lock);
933 }
934 
935 /*
936  * Send the "sevent" set of signals to a process.
937  * This might send more than one signal if the process is registered
938  * for multiple events. The caller should pass in an sevent that only
939  * includes the events for which the process has registered.
940  */
941 static void
942 dosendsig(proc_t *proc, int events, int sevent, k_siginfo_t *info,
943 	uchar_t band, int error)
944 {
945 	ASSERT(MUTEX_HELD(&proc->p_lock));
946 
947 	info->si_band = 0;
948 	info->si_errno = 0;
949 
950 	if (sevent & S_ERROR) {
951 		sevent &= ~S_ERROR;
952 		info->si_code = POLL_ERR;
953 		info->si_errno = error;
954 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
955 		    "strsendsig:proc %p info %p", proc, info);
956 		sigaddq(proc, NULL, info, KM_NOSLEEP);
957 		info->si_errno = 0;
958 	}
959 	if (sevent & S_HANGUP) {
960 		sevent &= ~S_HANGUP;
961 		info->si_code = POLL_HUP;
962 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
963 		    "strsendsig:proc %p info %p", proc, info);
964 		sigaddq(proc, NULL, info, KM_NOSLEEP);
965 	}
966 	if (sevent & S_HIPRI) {
967 		sevent &= ~S_HIPRI;
968 		info->si_code = POLL_PRI;
969 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
970 		    "strsendsig:proc %p info %p", proc, info);
971 		sigaddq(proc, NULL, info, KM_NOSLEEP);
972 	}
973 	if (sevent & S_RDBAND) {
974 		sevent &= ~S_RDBAND;
975 		if (events & S_BANDURG)
976 			sigtoproc(proc, NULL, SIGURG);
977 		else
978 			sigtoproc(proc, NULL, SIGPOLL);
979 	}
980 	if (sevent & S_WRBAND) {
981 		sevent &= ~S_WRBAND;
982 		sigtoproc(proc, NULL, SIGPOLL);
983 	}
984 	if (sevent & S_INPUT) {
985 		sevent &= ~S_INPUT;
986 		info->si_code = POLL_IN;
987 		info->si_band = band;
988 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
989 		    "strsendsig:proc %p info %p", proc, info);
990 		sigaddq(proc, NULL, info, KM_NOSLEEP);
991 		info->si_band = 0;
992 	}
993 	if (sevent & S_OUTPUT) {
994 		sevent &= ~S_OUTPUT;
995 		info->si_code = POLL_OUT;
996 		info->si_band = band;
997 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
998 		    "strsendsig:proc %p info %p", proc, info);
999 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1000 		info->si_band = 0;
1001 	}
1002 	if (sevent & S_MSG) {
1003 		sevent &= ~S_MSG;
1004 		info->si_code = POLL_MSG;
1005 		info->si_band = band;
1006 		TRACE_2(TR_FAC_STREAMS_FR, TR_STRSENDSIG,
1007 		    "strsendsig:proc %p info %p", proc, info);
1008 		sigaddq(proc, NULL, info, KM_NOSLEEP);
1009 		info->si_band = 0;
1010 	}
1011 	if (sevent & S_RDNORM) {
1012 		sevent &= ~S_RDNORM;
1013 		sigtoproc(proc, NULL, SIGPOLL);
1014 	}
1015 	if (sevent != 0) {
1016 		panic("strsendsig: unknown event(s) %x", sevent);
1017 	}
1018 }
1019 
1020 /*
1021  * Send SIGPOLL/SIGURG signal to all processes and process groups
1022  * registered on the given signal list that want a signal for at
1023  * least one of the specified events.
1024  *
1025  * Must be called with exclusive access to siglist (caller holding sd_lock).
1026  *
1027  * strioctl(I_SETSIG/I_ESETSIG) will only change siglist when holding
1028  * sd_lock and the ioctl code maintains a PID_HOLD on the pid structure
1029  * while it is in the siglist.
1030  *
1031  * For performance reasons (MP scalability) the code drops pidlock
1032  * when sending signals to a single process.
1033  * When sending to a process group the code holds
1034  * pidlock to prevent the membership in the process group from changing
1035  * while walking the p_pglink list.
1036  */
1037 void
1038 strsendsig(strsig_t *siglist, int event, uchar_t band, int error)
1039 {
1040 	strsig_t *ssp;
1041 	k_siginfo_t info;
1042 	struct pid *pidp;
1043 	proc_t  *proc;
1044 
1045 	info.si_signo = SIGPOLL;
1046 	info.si_errno = 0;
1047 	for (ssp = siglist; ssp; ssp = ssp->ss_next) {
1048 		int sevent;
1049 
1050 		sevent = ssp->ss_events & event;
1051 		if (sevent == 0)
1052 			continue;
1053 
1054 		if ((pidp = ssp->ss_pidp) == NULL) {
1055 			/* pid was released but still on event list */
1056 			continue;
1057 		}
1058 
1059 
1060 		if (ssp->ss_pid > 0) {
1061 			/*
1062 			 * XXX This unfortunately still generates
1063 			 * a signal when a fd is closed but
1064 			 * the proc is active.
1065 			 */
1066 			ASSERT(ssp->ss_pid == pidp->pid_id);
1067 
1068 			mutex_enter(&pidlock);
1069 			proc = prfind_zone(pidp->pid_id, ALL_ZONES);
1070 			if (proc == NULL) {
1071 				mutex_exit(&pidlock);
1072 				continue;
1073 			}
1074 			mutex_enter(&proc->p_lock);
1075 			mutex_exit(&pidlock);
1076 			dosendsig(proc, ssp->ss_events, sevent, &info,
1077 			    band, error);
1078 			mutex_exit(&proc->p_lock);
1079 		} else {
1080 			/*
1081 			 * Send to process group. Hold pidlock across
1082 			 * calls to dosendsig().
1083 			 */
1084 			pid_t pgrp = -ssp->ss_pid;
1085 
1086 			mutex_enter(&pidlock);
1087 			proc = pgfind_zone(pgrp, ALL_ZONES);
1088 			while (proc != NULL) {
1089 				mutex_enter(&proc->p_lock);
1090 				dosendsig(proc, ssp->ss_events, sevent,
1091 				    &info, band, error);
1092 				mutex_exit(&proc->p_lock);
1093 				proc = proc->p_pglink;
1094 			}
1095 			mutex_exit(&pidlock);
1096 		}
1097 	}
1098 }
1099 
1100 /*
1101  * Attach a stream device or module.
1102  * qp is a read queue; the new queue goes in so its next
1103  * read ptr is the argument, and the write queue corresponding
1104  * to the argument points to this queue. Return 0 on success,
1105  * or a non-zero errno on failure.
1106  */
1107 int
1108 qattach(queue_t *qp, dev_t *devp, int oflag, cred_t *crp, fmodsw_impl_t *fp,
1109     boolean_t is_insert)
1110 {
1111 	major_t			major;
1112 	cdevsw_impl_t		*dp;
1113 	struct streamtab	*str;
1114 	queue_t			*rq;
1115 	queue_t			*wrq;
1116 	uint32_t		qflag;
1117 	uint32_t		sqtype;
1118 	perdm_t			*dmp;
1119 	int			error;
1120 	int			sflag;
1121 
1122 	rq = allocq();
1123 	wrq = _WR(rq);
1124 	STREAM(rq) = STREAM(wrq) = STREAM(qp);
1125 
1126 	if (fp != NULL) {
1127 		str = fp->f_str;
1128 		qflag = fp->f_qflag;
1129 		sqtype = fp->f_sqtype;
1130 		dmp = fp->f_dmp;
1131 		IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
1132 		sflag = MODOPEN;
1133 
1134 		/*
1135 		 * stash away a pointer to the module structure so we can
1136 		 * unref it in qdetach.
1137 		 */
1138 		rq->q_fp = fp;
1139 	} else {
1140 		ASSERT(!is_insert);
1141 
1142 		major = getmajor(*devp);
1143 		dp = &devimpl[major];
1144 
1145 		str = dp->d_str;
1146 		ASSERT(str == STREAMSTAB(major));
1147 
1148 		qflag = dp->d_qflag;
1149 		ASSERT(qflag & QISDRV);
1150 		sqtype = dp->d_sqtype;
1151 
1152 		/* create perdm_t if needed */
1153 		if (NEED_DM(dp->d_dmp, qflag))
1154 			dp->d_dmp = hold_dm(str, qflag, sqtype);
1155 
1156 		dmp = dp->d_dmp;
1157 		sflag = 0;
1158 	}
1159 
1160 	TRACE_2(TR_FAC_STREAMS_FR, TR_QATTACH_FLAGS,
1161 	    "qattach:qflag == %X(%X)", qflag, *devp);
1162 
1163 	/* setq might sleep in allocator - avoid holding locks. */
1164 	setq(rq, str->st_rdinit, str->st_wrinit, dmp, qflag, sqtype, B_FALSE);
1165 
1166 	/*
1167 	 * Before calling the module's open routine, set up the q_next
1168 	 * pointer for inserting a module in the middle of a stream.
1169 	 *
1170 	 * Note that we can always set _QINSERTING and set up q_next
1171 	 * pointer for both inserting and pushing a module.  Then there
1172 	 * is no need for the is_insert parameter.  In insertq(), called
1173 	 * by qprocson(), assume that q_next of the new module always points
1174 	 * to the correct queue and use it for insertion.  Everything should
1175 	 * work out fine.  But in the first release of _I_INSERT, we
1176 	 * distinguish between inserting and pushing to make sure that
1177 	 * pushing a module follows the same code path as before.
1178 	 */
1179 	if (is_insert) {
1180 		rq->q_flag |= _QINSERTING;
1181 		rq->q_next = qp;
1182 	}
1183 
1184 	/*
1185 	 * If there is an outer perimeter get exclusive access during
1186 	 * the open procedure.  Bump up the reference count on the queue.
1187 	 */
1188 	entersq(rq->q_syncq, SQ_OPENCLOSE);
1189 	error = (*rq->q_qinfo->qi_qopen)(rq, devp, oflag, sflag, crp);
1190 	if (error != 0)
1191 		goto failed;
1192 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1193 	ASSERT(qprocsareon(rq));
1194 	return (0);
1195 
1196 failed:
1197 	rq->q_flag &= ~_QINSERTING;
1198 	if (backq(wrq) != NULL && backq(wrq)->q_next == wrq)
1199 		qprocsoff(rq);
1200 	leavesq(rq->q_syncq, SQ_OPENCLOSE);
1201 	rq->q_next = wrq->q_next = NULL;
1202 	qdetach(rq, 0, 0, crp, B_FALSE);
1203 	return (error);
1204 }
1205 
1206 /*
1207  * Handle second open of stream. For modules, set the
1208  * last argument to MODOPEN and do not pass any open flags.
1209  * Ignore dummydev since this is not the first open.
1210  */
1211 int
1212 qreopen(queue_t *qp, dev_t *devp, int flag, cred_t *crp)
1213 {
1214 	int	error;
1215 	dev_t dummydev;
1216 	queue_t *wqp = _WR(qp);
1217 
1218 	ASSERT(qp->q_flag & QREADR);
1219 	entersq(qp->q_syncq, SQ_OPENCLOSE);
1220 
1221 	dummydev = *devp;
1222 	if (error = ((*qp->q_qinfo->qi_qopen)(qp, &dummydev,
1223 	    (wqp->q_next ? 0 : flag), (wqp->q_next ? MODOPEN : 0), crp))) {
1224 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1225 		mutex_enter(&STREAM(qp)->sd_lock);
1226 		qp->q_stream->sd_flag |= STREOPENFAIL;
1227 		mutex_exit(&STREAM(qp)->sd_lock);
1228 		return (error);
1229 	}
1230 	leavesq(qp->q_syncq, SQ_OPENCLOSE);
1231 
1232 	/*
1233 	 * successful open should have done qprocson()
1234 	 */
1235 	ASSERT(qprocsareon(_RD(qp)));
1236 	return (0);
1237 }
1238 
1239 /*
1240  * Detach a stream module or device.
1241  * If clmode == 1 then the module or driver was opened and its
1242  * close routine must be called. If clmode == 0, the module
1243  * or driver was never opened or the open failed, and so its close
1244  * should not be called.
1245  */
1246 void
1247 qdetach(queue_t *qp, int clmode, int flag, cred_t *crp, boolean_t is_remove)
1248 {
1249 	queue_t *wqp = _WR(qp);
1250 	ASSERT(STREAM(qp)->sd_flag & (STRCLOSE|STWOPEN|STRPLUMB));
1251 
1252 	if (STREAM_NEEDSERVICE(STREAM(qp)))
1253 		stream_runservice(STREAM(qp));
1254 
1255 	if (clmode) {
1256 		/*
1257 		 * Make sure that all the messages on the write side syncq are
1258 		 * processed and nothing is left. Since we are closing, no new
1259 		 * messages may appear there.
1260 		 */
1261 		wait_q_syncq(wqp);
1262 
1263 		entersq(qp->q_syncq, SQ_OPENCLOSE);
1264 		if (is_remove) {
1265 			mutex_enter(QLOCK(qp));
1266 			qp->q_flag |= _QREMOVING;
1267 			mutex_exit(QLOCK(qp));
1268 		}
1269 		(*qp->q_qinfo->qi_qclose)(qp, flag, crp);
1270 		/*
1271 		 * Check that qprocsoff() was actually called.
1272 		 */
1273 		ASSERT((qp->q_flag & QWCLOSE) && (wqp->q_flag & QWCLOSE));
1274 
1275 		leavesq(qp->q_syncq, SQ_OPENCLOSE);
1276 	} else {
1277 		disable_svc(qp);
1278 	}
1279 
1280 	/*
1281 	 * Allow any threads blocked in entersq to proceed and discover
1282 	 * the QWCLOSE is set.
1283 	 * Note: This assumes that all users of entersq check QWCLOSE.
1284 	 * Currently runservice is the only entersq that can happen
1285 	 * after removeq has finished.
1286 	 * Removeq will have discarded all messages destined to the closing
1287 	 * pair of queues from the syncq.
1288 	 * NOTE: Calling a function inside an assert is unconventional.
1289 	 * However, it does not cause any problem since flush_syncq() does
1290 	 * not change any state except when it returns non-zero i.e.
1291 	 * when the assert will trigger.
1292 	 */
1293 	ASSERT(flush_syncq(qp->q_syncq, qp) == 0);
1294 	ASSERT(flush_syncq(wqp->q_syncq, wqp) == 0);
1295 	ASSERT((qp->q_flag & QPERMOD) ||
1296 	    ((qp->q_syncq->sq_head == NULL) &&
1297 	    (wqp->q_syncq->sq_head == NULL)));
1298 
1299 	/* release any fmodsw_impl_t structure held on behalf of the queue */
1300 	ASSERT(qp->q_fp != NULL || qp->q_flag & QISDRV);
1301 	if (qp->q_fp != NULL)
1302 		fmodsw_rele(qp->q_fp);
1303 
1304 	/* freeq removes us from the outer perimeter if any */
1305 	freeq(qp);
1306 }
1307 
1308 /* Prevent service procedures from being called */
1309 void
1310 disable_svc(queue_t *qp)
1311 {
1312 	queue_t *wqp = _WR(qp);
1313 
1314 	ASSERT(qp->q_flag & QREADR);
1315 	mutex_enter(QLOCK(qp));
1316 	qp->q_flag |= QWCLOSE;
1317 	mutex_exit(QLOCK(qp));
1318 	mutex_enter(QLOCK(wqp));
1319 	wqp->q_flag |= QWCLOSE;
1320 	mutex_exit(QLOCK(wqp));
1321 }
1322 
1323 /* Allow service procedures to be called again */
1324 void
1325 enable_svc(queue_t *qp)
1326 {
1327 	queue_t *wqp = _WR(qp);
1328 
1329 	ASSERT(qp->q_flag & QREADR);
1330 	mutex_enter(QLOCK(qp));
1331 	qp->q_flag &= ~QWCLOSE;
1332 	mutex_exit(QLOCK(qp));
1333 	mutex_enter(QLOCK(wqp));
1334 	wqp->q_flag &= ~QWCLOSE;
1335 	mutex_exit(QLOCK(wqp));
1336 }
1337 
1338 /*
1339  * Remove queue from qhead/qtail if it is enabled.
1340  * Only reset QENAB if the queue was removed from the runlist.
1341  * A queue goes through 3 stages:
1342  *	It is on the service list and QENAB is set.
1343  *	It is removed from the service list but QENAB is still set.
1344  *	QENAB gets changed to QINSERVICE.
1345  *	QINSERVICE is reset (when the service procedure is done)
1346  * Thus we can not reset QENAB unless we actually removed it from the service
1347  * queue.
1348  */
1349 void
1350 remove_runlist(queue_t *qp)
1351 {
1352 	if (qp->q_flag & QENAB && qhead != NULL) {
1353 		queue_t *q_chase;
1354 		queue_t *q_curr;
1355 		int removed;
1356 
1357 		mutex_enter(&service_queue);
1358 		RMQ(qp, qhead, qtail, q_link, q_chase, q_curr, removed);
1359 		mutex_exit(&service_queue);
1360 		if (removed) {
1361 			STRSTAT(qremoved);
1362 			qp->q_flag &= ~QENAB;
1363 		}
1364 	}
1365 }
1366 
1367 
1368 /*
1369  * Wait for any pending service processing to complete.
1370  * The removal of queues from the runlist is not atomic with the
1371  * clearing of the QENABLED flag and setting the INSERVICE flag.
1372  * consequently it is possible for remove_runlist in strclose
1373  * to not find the queue on the runlist but for it to be QENABLED
1374  * and not yet INSERVICE -> hence wait_svc needs to check QENABLED
1375  * as well as INSERVICE.
1376  */
1377 void
1378 wait_svc(queue_t *qp)
1379 {
1380 	queue_t *wqp = _WR(qp);
1381 
1382 	ASSERT(qp->q_flag & QREADR);
1383 
1384 	/*
1385 	 * Try to remove queues from qhead/qtail list.
1386 	 */
1387 	if (qhead != NULL) {
1388 		remove_runlist(qp);
1389 		remove_runlist(wqp);
1390 	}
1391 	/*
1392 	 * Wait till the syncqs associated with the queue disappear from the
1393 	 * background processing list.
1394 	 * This only needs to be done for non-PERMOD perimeters since
1395 	 * for PERMOD perimeters the syncq may be shared and will only be freed
1396 	 * when the last module/driver is unloaded.
1397 	 * If for PERMOD perimeters queue was on the syncq list, removeq()
1398 	 * should call propagate_syncq() or drain_syncq() for it. Both of these
1399 	 * functions remove the queue from its syncq list, so sqthread will not
1400 	 * try to access the queue.
1401 	 */
1402 	if (!(qp->q_flag & QPERMOD)) {
1403 		syncq_t *rsq = qp->q_syncq;
1404 		syncq_t *wsq = wqp->q_syncq;
1405 
1406 		/*
1407 		 * Disable rsq and wsq and wait for any background processing of
1408 		 * syncq to complete.
1409 		 */
1410 		wait_sq_svc(rsq);
1411 		if (wsq != rsq)
1412 			wait_sq_svc(wsq);
1413 	}
1414 
1415 	mutex_enter(QLOCK(qp));
1416 	while (qp->q_flag & (QINSERVICE|QENAB))
1417 		cv_wait(&qp->q_wait, QLOCK(qp));
1418 	mutex_exit(QLOCK(qp));
1419 	mutex_enter(QLOCK(wqp));
1420 	while (wqp->q_flag & (QINSERVICE|QENAB))
1421 		cv_wait(&wqp->q_wait, QLOCK(wqp));
1422 	mutex_exit(QLOCK(wqp));
1423 }
1424 
1425 /*
1426  * Put ioctl data from userland buffer `arg' into the mblk chain `bp'.
1427  * `flag' must always contain either K_TO_K or U_TO_K; STR_NOSIG may
1428  * also be set, and is passed through to allocb_cred_wait().
1429  *
1430  * Returns errno on failure, zero on success.
1431  */
1432 int
1433 putiocd(mblk_t *bp, char *arg, int flag, cred_t *cr)
1434 {
1435 	mblk_t *tmp;
1436 	ssize_t  count;
1437 	int error = 0;
1438 
1439 	ASSERT((flag & (U_TO_K | K_TO_K)) == U_TO_K ||
1440 	    (flag & (U_TO_K | K_TO_K)) == K_TO_K);
1441 
1442 	if (bp->b_datap->db_type == M_IOCTL) {
1443 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1444 	} else {
1445 		ASSERT(bp->b_datap->db_type == M_COPYIN);
1446 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1447 	}
1448 	/*
1449 	 * strdoioctl validates ioc_count, so if this assert fails it
1450 	 * cannot be due to user error.
1451 	 */
1452 	ASSERT(count >= 0);
1453 
1454 	if ((tmp = allocb_cred_wait(count, (flag & STR_NOSIG), &error, cr,
1455 	    curproc->p_pid)) == NULL) {
1456 		return (error);
1457 	}
1458 	error = strcopyin(arg, tmp->b_wptr, count, flag & (U_TO_K|K_TO_K));
1459 	if (error != 0) {
1460 		freeb(tmp);
1461 		return (error);
1462 	}
1463 	DB_CPID(tmp) = curproc->p_pid;
1464 	tmp->b_wptr += count;
1465 	bp->b_cont = tmp;
1466 
1467 	return (0);
1468 }
1469 
1470 /*
1471  * Copy ioctl data to user-land. Return non-zero errno on failure,
1472  * 0 for success.
1473  */
1474 int
1475 getiocd(mblk_t *bp, char *arg, int copymode)
1476 {
1477 	ssize_t count;
1478 	size_t  n;
1479 	int	error;
1480 
1481 	if (bp->b_datap->db_type == M_IOCACK)
1482 		count = ((struct iocblk *)bp->b_rptr)->ioc_count;
1483 	else {
1484 		ASSERT(bp->b_datap->db_type == M_COPYOUT);
1485 		count = ((struct copyreq *)bp->b_rptr)->cq_size;
1486 	}
1487 	ASSERT(count >= 0);
1488 
1489 	for (bp = bp->b_cont; bp && count;
1490 	    count -= n, bp = bp->b_cont, arg += n) {
1491 		n = MIN(count, bp->b_wptr - bp->b_rptr);
1492 		error = strcopyout(bp->b_rptr, arg, n, copymode);
1493 		if (error)
1494 			return (error);
1495 	}
1496 	ASSERT(count == 0);
1497 	return (0);
1498 }
1499 
1500 /*
1501  * Allocate a linkinfo entry given the write queue of the
1502  * bottom module of the top stream and the write queue of the
1503  * stream head of the bottom stream.
1504  */
1505 linkinfo_t *
1506 alloclink(queue_t *qup, queue_t *qdown, file_t *fpdown)
1507 {
1508 	linkinfo_t *linkp;
1509 
1510 	linkp = kmem_cache_alloc(linkinfo_cache, KM_SLEEP);
1511 
1512 	linkp->li_lblk.l_qtop = qup;
1513 	linkp->li_lblk.l_qbot = qdown;
1514 	linkp->li_fpdown = fpdown;
1515 
1516 	mutex_enter(&strresources);
1517 	linkp->li_next = linkinfo_list;
1518 	linkp->li_prev = NULL;
1519 	if (linkp->li_next)
1520 		linkp->li_next->li_prev = linkp;
1521 	linkinfo_list = linkp;
1522 	linkp->li_lblk.l_index = ++lnk_id;
1523 	ASSERT(lnk_id != 0);	/* this should never wrap in practice */
1524 	mutex_exit(&strresources);
1525 
1526 	return (linkp);
1527 }
1528 
1529 /*
1530  * Free a linkinfo entry.
1531  */
1532 void
1533 lbfree(linkinfo_t *linkp)
1534 {
1535 	mutex_enter(&strresources);
1536 	if (linkp->li_next)
1537 		linkp->li_next->li_prev = linkp->li_prev;
1538 	if (linkp->li_prev)
1539 		linkp->li_prev->li_next = linkp->li_next;
1540 	else
1541 		linkinfo_list = linkp->li_next;
1542 	mutex_exit(&strresources);
1543 
1544 	kmem_cache_free(linkinfo_cache, linkp);
1545 }
1546 
1547 /*
1548  * Check for a potential linking cycle.
1549  * Return 1 if a link will result in a cycle,
1550  * and 0 otherwise.
1551  */
1552 int
1553 linkcycle(stdata_t *upstp, stdata_t *lostp, str_stack_t *ss)
1554 {
1555 	struct mux_node *np;
1556 	struct mux_edge *ep;
1557 	int i;
1558 	major_t lomaj;
1559 	major_t upmaj;
1560 	/*
1561 	 * if the lower stream is a pipe/FIFO, return, since link
1562 	 * cycles can not happen on pipes/FIFOs
1563 	 */
1564 	if (lostp->sd_vnode->v_type == VFIFO)
1565 		return (0);
1566 
1567 	for (i = 0; i < ss->ss_devcnt; i++) {
1568 		np = &ss->ss_mux_nodes[i];
1569 		MUX_CLEAR(np);
1570 	}
1571 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
1572 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
1573 	np = &ss->ss_mux_nodes[lomaj];
1574 	for (;;) {
1575 		if (!MUX_DIDVISIT(np)) {
1576 			if (np->mn_imaj == upmaj)
1577 				return (1);
1578 			if (np->mn_outp == NULL) {
1579 				MUX_VISIT(np);
1580 				if (np->mn_originp == NULL)
1581 					return (0);
1582 				np = np->mn_originp;
1583 				continue;
1584 			}
1585 			MUX_VISIT(np);
1586 			np->mn_startp = np->mn_outp;
1587 		} else {
1588 			if (np->mn_startp == NULL) {
1589 				if (np->mn_originp == NULL)
1590 					return (0);
1591 				else {
1592 					np = np->mn_originp;
1593 					continue;
1594 				}
1595 			}
1596 			/*
1597 			 * If ep->me_nodep is a FIFO (me_nodep == NULL),
1598 			 * ignore the edge and move on. ep->me_nodep gets
1599 			 * set to NULL in mux_addedge() if it is a FIFO.
1600 			 *
1601 			 */
1602 			ep = np->mn_startp;
1603 			np->mn_startp = ep->me_nextp;
1604 			if (ep->me_nodep == NULL)
1605 				continue;
1606 			ep->me_nodep->mn_originp = np;
1607 			np = ep->me_nodep;
1608 		}
1609 	}
1610 }
1611 
1612 /*
1613  * Find linkinfo entry corresponding to the parameters.
1614  */
1615 linkinfo_t *
1616 findlinks(stdata_t *stp, int index, int type, str_stack_t *ss)
1617 {
1618 	linkinfo_t *linkp;
1619 	struct mux_edge *mep;
1620 	struct mux_node *mnp;
1621 	queue_t *qup;
1622 
1623 	mutex_enter(&strresources);
1624 	if ((type & LINKTYPEMASK) == LINKNORMAL) {
1625 		qup = getendq(stp->sd_wrq);
1626 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1627 			if ((qup == linkp->li_lblk.l_qtop) &&
1628 			    (!index || (index == linkp->li_lblk.l_index))) {
1629 				mutex_exit(&strresources);
1630 				return (linkp);
1631 			}
1632 		}
1633 	} else {
1634 		ASSERT((type & LINKTYPEMASK) == LINKPERSIST);
1635 		mnp = &ss->ss_mux_nodes[getmajor(stp->sd_vnode->v_rdev)];
1636 		mep = mnp->mn_outp;
1637 		while (mep) {
1638 			if ((index == 0) || (index == mep->me_muxid))
1639 				break;
1640 			mep = mep->me_nextp;
1641 		}
1642 		if (!mep) {
1643 			mutex_exit(&strresources);
1644 			return (NULL);
1645 		}
1646 		for (linkp = linkinfo_list; linkp; linkp = linkp->li_next) {
1647 			if ((!linkp->li_lblk.l_qtop) &&
1648 			    (mep->me_muxid == linkp->li_lblk.l_index)) {
1649 				mutex_exit(&strresources);
1650 				return (linkp);
1651 			}
1652 		}
1653 	}
1654 	mutex_exit(&strresources);
1655 	return (NULL);
1656 }
1657 
1658 /*
1659  * Given a queue ptr, follow the chain of q_next pointers until you reach the
1660  * last queue on the chain and return it.
1661  */
1662 queue_t *
1663 getendq(queue_t *q)
1664 {
1665 	ASSERT(q != NULL);
1666 	while (_SAMESTR(q))
1667 		q = q->q_next;
1668 	return (q);
1669 }
1670 
1671 /*
1672  * Wait for the syncq count to drop to zero.
1673  * sq could be either outer or inner.
1674  */
1675 
1676 static void
1677 wait_syncq(syncq_t *sq)
1678 {
1679 	uint16_t count;
1680 
1681 	mutex_enter(SQLOCK(sq));
1682 	count = sq->sq_count;
1683 	SQ_PUTLOCKS_ENTER(sq);
1684 	SUM_SQ_PUTCOUNTS(sq, count);
1685 	while (count != 0) {
1686 		sq->sq_flags |= SQ_WANTWAKEUP;
1687 		SQ_PUTLOCKS_EXIT(sq);
1688 		cv_wait(&sq->sq_wait, SQLOCK(sq));
1689 		count = sq->sq_count;
1690 		SQ_PUTLOCKS_ENTER(sq);
1691 		SUM_SQ_PUTCOUNTS(sq, count);
1692 	}
1693 	SQ_PUTLOCKS_EXIT(sq);
1694 	mutex_exit(SQLOCK(sq));
1695 }
1696 
1697 /*
1698  * Wait while there are any messages for the queue in its syncq.
1699  */
1700 static void
1701 wait_q_syncq(queue_t *q)
1702 {
1703 	if ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1704 		syncq_t *sq = q->q_syncq;
1705 
1706 		mutex_enter(SQLOCK(sq));
1707 		while ((q->q_sqflags & Q_SQQUEUED) || (q->q_syncqmsgs > 0)) {
1708 			sq->sq_flags |= SQ_WANTWAKEUP;
1709 			cv_wait(&sq->sq_wait, SQLOCK(sq));
1710 		}
1711 		mutex_exit(SQLOCK(sq));
1712 	}
1713 }
1714 
1715 
1716 int
1717 mlink_file(vnode_t *vp, int cmd, struct file *fpdown, cred_t *crp, int *rvalp,
1718     int lhlink)
1719 {
1720 	struct stdata *stp;
1721 	struct strioctl strioc;
1722 	struct linkinfo *linkp;
1723 	struct stdata *stpdown;
1724 	struct streamtab *str;
1725 	queue_t *passq;
1726 	syncq_t *passyncq;
1727 	queue_t *rq;
1728 	cdevsw_impl_t *dp;
1729 	uint32_t qflag;
1730 	uint32_t sqtype;
1731 	perdm_t *dmp;
1732 	int error = 0;
1733 	netstack_t *ns;
1734 	str_stack_t *ss;
1735 
1736 	stp = vp->v_stream;
1737 	TRACE_1(TR_FAC_STREAMS_FR,
1738 	    TR_I_LINK, "I_LINK/I_PLINK:stp %p", stp);
1739 	/*
1740 	 * Test for invalid upper stream
1741 	 */
1742 	if (stp->sd_flag & STRHUP) {
1743 		return (ENXIO);
1744 	}
1745 	if (vp->v_type == VFIFO) {
1746 		return (EINVAL);
1747 	}
1748 	if (stp->sd_strtab == NULL) {
1749 		return (EINVAL);
1750 	}
1751 	if (!stp->sd_strtab->st_muxwinit) {
1752 		return (EINVAL);
1753 	}
1754 	if (fpdown == NULL) {
1755 		return (EBADF);
1756 	}
1757 	ns = netstack_find_by_cred(crp);
1758 	ASSERT(ns != NULL);
1759 	ss = ns->netstack_str;
1760 	ASSERT(ss != NULL);
1761 
1762 	if (getmajor(stp->sd_vnode->v_rdev) >= ss->ss_devcnt) {
1763 		netstack_rele(ss->ss_netstack);
1764 		return (EINVAL);
1765 	}
1766 	mutex_enter(&muxifier);
1767 	if (stp->sd_flag & STPLEX) {
1768 		mutex_exit(&muxifier);
1769 		netstack_rele(ss->ss_netstack);
1770 		return (ENXIO);
1771 	}
1772 
1773 	/*
1774 	 * Test for invalid lower stream.
1775 	 * The check for the v_type != VFIFO and having a major
1776 	 * number not >= devcnt is done to avoid problems with
1777 	 * adding mux_node entry past the end of mux_nodes[].
1778 	 * For FIFO's we don't add an entry so this isn't a
1779 	 * problem.
1780 	 */
1781 	if (((stpdown = fpdown->f_vnode->v_stream) == NULL) ||
1782 	    (stpdown == stp) || (stpdown->sd_flag &
1783 	    (STPLEX|STRHUP|STRDERR|STWRERR|IOCWAIT|STRPLUMB)) ||
1784 	    ((stpdown->sd_vnode->v_type != VFIFO) &&
1785 	    (getmajor(stpdown->sd_vnode->v_rdev) >= ss->ss_devcnt)) ||
1786 	    linkcycle(stp, stpdown, ss)) {
1787 		mutex_exit(&muxifier);
1788 		netstack_rele(ss->ss_netstack);
1789 		return (EINVAL);
1790 	}
1791 	TRACE_1(TR_FAC_STREAMS_FR,
1792 	    TR_STPDOWN, "stpdown:%p", stpdown);
1793 	rq = getendq(stp->sd_wrq);
1794 	if (cmd == I_PLINK)
1795 		rq = NULL;
1796 
1797 	linkp = alloclink(rq, stpdown->sd_wrq, fpdown);
1798 
1799 	strioc.ic_cmd = cmd;
1800 	strioc.ic_timout = INFTIM;
1801 	strioc.ic_len = sizeof (struct linkblk);
1802 	strioc.ic_dp = (char *)&linkp->li_lblk;
1803 
1804 	/*
1805 	 * STRPLUMB protects plumbing changes and should be set before
1806 	 * link_addpassthru()/link_rempassthru() are called, so it is set here
1807 	 * and cleared in the end of mlink when passthru queue is removed.
1808 	 * Setting of STRPLUMB prevents reopens of the stream while passthru
1809 	 * queue is in-place (it is not a proper module and doesn't have open
1810 	 * entry point).
1811 	 *
1812 	 * STPLEX prevents any threads from entering the stream from above. It
1813 	 * can't be set before the call to link_addpassthru() because putnext
1814 	 * from below may cause stream head I/O routines to be called and these
1815 	 * routines assert that STPLEX is not set. After link_addpassthru()
1816 	 * nothing may come from below since the pass queue syncq is blocked.
1817 	 * Note also that STPLEX should be cleared before the call to
1818 	 * link_rempassthru() since when messages start flowing to the stream
1819 	 * head (e.g. because of message propagation from the pass queue) stream
1820 	 * head I/O routines may be called with STPLEX flag set.
1821 	 *
1822 	 * When STPLEX is set, nothing may come into the stream from above and
1823 	 * it is safe to do a setq which will change stream head. So, the
1824 	 * correct sequence of actions is:
1825 	 *
1826 	 * 1) Set STRPLUMB
1827 	 * 2) Call link_addpassthru()
1828 	 * 3) Set STPLEX
1829 	 * 4) Call setq and update the stream state
1830 	 * 5) Clear STPLEX
1831 	 * 6) Call link_rempassthru()
1832 	 * 7) Clear STRPLUMB
1833 	 *
1834 	 * The same sequence applies to munlink() code.
1835 	 */
1836 	mutex_enter(&stpdown->sd_lock);
1837 	stpdown->sd_flag |= STRPLUMB;
1838 	mutex_exit(&stpdown->sd_lock);
1839 	/*
1840 	 * Add passthru queue below lower mux. This will block
1841 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
1842 	 */
1843 	passq = link_addpassthru(stpdown);
1844 
1845 	mutex_enter(&stpdown->sd_lock);
1846 	stpdown->sd_flag |= STPLEX;
1847 	mutex_exit(&stpdown->sd_lock);
1848 
1849 	rq = _RD(stpdown->sd_wrq);
1850 	/*
1851 	 * There may be messages in the streamhead's syncq due to messages
1852 	 * that arrived before link_addpassthru() was done. To avoid
1853 	 * background processing of the syncq happening simultaneous with
1854 	 * setq processing, we disable the streamhead syncq and wait until
1855 	 * existing background thread finishes working on it.
1856 	 */
1857 	wait_sq_svc(rq->q_syncq);
1858 	passyncq = passq->q_syncq;
1859 	if (!(passyncq->sq_flags & SQ_BLOCKED))
1860 		blocksq(passyncq, SQ_BLOCKED, 0);
1861 
1862 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
1863 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
1864 	rq->q_ptr = _WR(rq)->q_ptr = NULL;
1865 
1866 	/* setq might sleep in allocator - avoid holding locks. */
1867 	/* Note: we are holding muxifier here. */
1868 
1869 	str = stp->sd_strtab;
1870 	dp = &devimpl[getmajor(vp->v_rdev)];
1871 	ASSERT(dp->d_str == str);
1872 
1873 	qflag = dp->d_qflag;
1874 	sqtype = dp->d_sqtype;
1875 
1876 	/* create perdm_t if needed */
1877 	if (NEED_DM(dp->d_dmp, qflag))
1878 		dp->d_dmp = hold_dm(str, qflag, sqtype);
1879 
1880 	dmp = dp->d_dmp;
1881 
1882 	setq(rq, str->st_muxrinit, str->st_muxwinit, dmp, qflag, sqtype,
1883 	    B_TRUE);
1884 
1885 	/*
1886 	 * XXX Remove any "odd" messages from the queue.
1887 	 * Keep only M_DATA, M_PROTO, M_PCPROTO.
1888 	 */
1889 	error = strdoioctl(stp, &strioc, FNATIVE,
1890 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
1891 	if (error != 0) {
1892 		lbfree(linkp);
1893 
1894 		if (!(passyncq->sq_flags & SQ_BLOCKED))
1895 			blocksq(passyncq, SQ_BLOCKED, 0);
1896 		/*
1897 		 * Restore the stream head queue and then remove
1898 		 * the passq. Turn off STPLEX before we turn on
1899 		 * the stream by removing the passq.
1900 		 */
1901 		rq->q_ptr = _WR(rq)->q_ptr = stpdown;
1902 		setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO,
1903 		    B_TRUE);
1904 
1905 		mutex_enter(&stpdown->sd_lock);
1906 		stpdown->sd_flag &= ~STPLEX;
1907 		mutex_exit(&stpdown->sd_lock);
1908 
1909 		link_rempassthru(passq);
1910 
1911 		mutex_enter(&stpdown->sd_lock);
1912 		stpdown->sd_flag &= ~STRPLUMB;
1913 		/* Wakeup anyone waiting for STRPLUMB to clear. */
1914 		cv_broadcast(&stpdown->sd_monitor);
1915 		mutex_exit(&stpdown->sd_lock);
1916 
1917 		mutex_exit(&muxifier);
1918 		netstack_rele(ss->ss_netstack);
1919 		return (error);
1920 	}
1921 	mutex_enter(&fpdown->f_tlock);
1922 	fpdown->f_count++;
1923 	mutex_exit(&fpdown->f_tlock);
1924 
1925 	/*
1926 	 * if we've made it here the linkage is all set up so we should also
1927 	 * set up the layered driver linkages
1928 	 */
1929 
1930 	ASSERT((cmd == I_LINK) || (cmd == I_PLINK));
1931 	if (cmd == I_LINK) {
1932 		ldi_mlink_fp(stp, fpdown, lhlink, LINKNORMAL);
1933 	} else {
1934 		ldi_mlink_fp(stp, fpdown, lhlink, LINKPERSIST);
1935 	}
1936 
1937 	link_rempassthru(passq);
1938 
1939 	mux_addedge(stp, stpdown, linkp->li_lblk.l_index, ss);
1940 
1941 	/*
1942 	 * Mark the upper stream as having dependent links
1943 	 * so that strclose can clean it up.
1944 	 */
1945 	if (cmd == I_LINK) {
1946 		mutex_enter(&stp->sd_lock);
1947 		stp->sd_flag |= STRHASLINKS;
1948 		mutex_exit(&stp->sd_lock);
1949 	}
1950 	/*
1951 	 * Wake up any other processes that may have been
1952 	 * waiting on the lower stream. These will all
1953 	 * error out.
1954 	 */
1955 	mutex_enter(&stpdown->sd_lock);
1956 	/* The passthru module is removed so we may release STRPLUMB */
1957 	stpdown->sd_flag &= ~STRPLUMB;
1958 	cv_broadcast(&rq->q_wait);
1959 	cv_broadcast(&_WR(rq)->q_wait);
1960 	cv_broadcast(&stpdown->sd_monitor);
1961 	mutex_exit(&stpdown->sd_lock);
1962 	mutex_exit(&muxifier);
1963 	*rvalp = linkp->li_lblk.l_index;
1964 	netstack_rele(ss->ss_netstack);
1965 	return (0);
1966 }
1967 
1968 int
1969 mlink(vnode_t *vp, int cmd, int arg, cred_t *crp, int *rvalp, int lhlink)
1970 {
1971 	int		ret;
1972 	struct file	*fpdown;
1973 
1974 	fpdown = getf(arg);
1975 	ret = mlink_file(vp, cmd, fpdown, crp, rvalp, lhlink);
1976 	if (fpdown != NULL)
1977 		releasef(arg);
1978 	return (ret);
1979 }
1980 
1981 /*
1982  * Unlink a multiplexor link. Stp is the controlling stream for the
1983  * link, and linkp points to the link's entry in the linkinfo list.
1984  * The muxifier lock must be held on entry and is dropped on exit.
1985  *
1986  * NOTE : Currently it is assumed that mux would process all the messages
1987  * sitting on it's queue before ACKing the UNLINK. It is the responsibility
1988  * of the mux to handle all the messages that arrive before UNLINK.
1989  * If the mux has to send down messages on its lower stream before
1990  * ACKing I_UNLINK, then it *should* know to handle messages even
1991  * after the UNLINK is acked (actually it should be able to handle till we
1992  * re-block the read side of the pass queue here). If the mux does not
1993  * open up the lower stream, any messages that arrive during UNLINK
1994  * will be put in the stream head. In the case of lower stream opening
1995  * up, some messages might land in the stream head depending on when
1996  * the message arrived and when the read side of the pass queue was
1997  * re-blocked.
1998  */
1999 int
2000 munlink(stdata_t *stp, linkinfo_t *linkp, int flag, cred_t *crp, int *rvalp,
2001     str_stack_t *ss)
2002 {
2003 	struct strioctl strioc;
2004 	struct stdata *stpdown;
2005 	queue_t *rq, *wrq;
2006 	queue_t	*passq;
2007 	syncq_t *passyncq;
2008 	int error = 0;
2009 	file_t *fpdown;
2010 
2011 	ASSERT(MUTEX_HELD(&muxifier));
2012 
2013 	stpdown = linkp->li_fpdown->f_vnode->v_stream;
2014 
2015 	/*
2016 	 * See the comment in mlink() concerning STRPLUMB/STPLEX flags.
2017 	 */
2018 	mutex_enter(&stpdown->sd_lock);
2019 	stpdown->sd_flag |= STRPLUMB;
2020 	mutex_exit(&stpdown->sd_lock);
2021 
2022 	/*
2023 	 * Add passthru queue below lower mux. This will block
2024 	 * syncqs of lower muxs read queue during I_LINK/I_UNLINK.
2025 	 */
2026 	passq = link_addpassthru(stpdown);
2027 
2028 	if ((flag & LINKTYPEMASK) == LINKNORMAL)
2029 		strioc.ic_cmd = I_UNLINK;
2030 	else
2031 		strioc.ic_cmd = I_PUNLINK;
2032 	strioc.ic_timout = INFTIM;
2033 	strioc.ic_len = sizeof (struct linkblk);
2034 	strioc.ic_dp = (char *)&linkp->li_lblk;
2035 
2036 	error = strdoioctl(stp, &strioc, FNATIVE,
2037 	    K_TO_K | STR_NOERROR | STR_NOSIG, crp, rvalp);
2038 
2039 	/*
2040 	 * If there was an error and this is not called via strclose,
2041 	 * return to the user. Otherwise, pretend there was no error
2042 	 * and close the link.
2043 	 */
2044 	if (error) {
2045 		if (flag & LINKCLOSE) {
2046 			cmn_err(CE_WARN, "KERNEL: munlink: could not perform "
2047 			    "unlink ioctl, closing anyway (%d)\n", error);
2048 		} else {
2049 			link_rempassthru(passq);
2050 			mutex_enter(&stpdown->sd_lock);
2051 			stpdown->sd_flag &= ~STRPLUMB;
2052 			cv_broadcast(&stpdown->sd_monitor);
2053 			mutex_exit(&stpdown->sd_lock);
2054 			mutex_exit(&muxifier);
2055 			return (error);
2056 		}
2057 	}
2058 
2059 	mux_rmvedge(stp, linkp->li_lblk.l_index, ss);
2060 	fpdown = linkp->li_fpdown;
2061 	lbfree(linkp);
2062 
2063 	/*
2064 	 * We go ahead and drop muxifier here--it's a nasty global lock that
2065 	 * can slow others down. It's okay to since attempts to mlink() this
2066 	 * stream will be stopped because STPLEX is still set in the stdata
2067 	 * structure, and munlink() is stopped because mux_rmvedge() and
2068 	 * lbfree() have removed it from mux_nodes[] and linkinfo_list,
2069 	 * respectively.  Note that we defer the closef() of fpdown until
2070 	 * after we drop muxifier since strclose() can call munlinkall().
2071 	 */
2072 	mutex_exit(&muxifier);
2073 
2074 	wrq = stpdown->sd_wrq;
2075 	rq = _RD(wrq);
2076 
2077 	/*
2078 	 * Get rid of outstanding service procedure runs, before we make
2079 	 * it a stream head, since a stream head doesn't have any service
2080 	 * procedure.
2081 	 */
2082 	disable_svc(rq);
2083 	wait_svc(rq);
2084 
2085 	/*
2086 	 * Since we don't disable the syncq for QPERMOD, we wait for whatever
2087 	 * is queued up to be finished. mux should take care that nothing is
2088 	 * send down to this queue. We should do it now as we're going to block
2089 	 * passyncq if it was unblocked.
2090 	 */
2091 	if (wrq->q_flag & QPERMOD) {
2092 		syncq_t	*sq = wrq->q_syncq;
2093 
2094 		mutex_enter(SQLOCK(sq));
2095 		while (wrq->q_sqflags & Q_SQQUEUED) {
2096 			sq->sq_flags |= SQ_WANTWAKEUP;
2097 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2098 		}
2099 		mutex_exit(SQLOCK(sq));
2100 	}
2101 	passyncq = passq->q_syncq;
2102 	if (!(passyncq->sq_flags & SQ_BLOCKED)) {
2103 
2104 		syncq_t *sq, *outer;
2105 
2106 		/*
2107 		 * Messages could be flowing from underneath. We will
2108 		 * block the read side of the passq. This would be
2109 		 * sufficient for QPAIR and QPERQ muxes to ensure
2110 		 * that no data is flowing up into this queue
2111 		 * and hence no thread active in this instance of
2112 		 * lower mux. But for QPERMOD and QMTOUTPERIM there
2113 		 * could be messages on the inner and outer/inner
2114 		 * syncqs respectively. We will wait for them to drain.
2115 		 * Because passq is blocked messages end up in the syncq
2116 		 * And qfill_syncq could possibly end up setting QFULL
2117 		 * which will access the rq->q_flag. Hence, we have to
2118 		 * acquire the QLOCK in setq.
2119 		 *
2120 		 * XXX Messages can also flow from top into this
2121 		 * queue though the unlink is over (Ex. some instance
2122 		 * in putnext() called from top that has still not
2123 		 * accessed this queue. And also putq(lowerq) ?).
2124 		 * Solution : How about blocking the l_qtop queue ?
2125 		 * Do we really care about such pure D_MP muxes ?
2126 		 */
2127 
2128 		blocksq(passyncq, SQ_BLOCKED, 0);
2129 
2130 		sq = rq->q_syncq;
2131 		if ((outer = sq->sq_outer) != NULL) {
2132 
2133 			/*
2134 			 * We have to just wait for the outer sq_count
2135 			 * drop to zero. As this does not prevent new
2136 			 * messages to enter the outer perimeter, this
2137 			 * is subject to starvation.
2138 			 *
2139 			 * NOTE :Because of blocksq above, messages could
2140 			 * be in the inner syncq only because of some
2141 			 * thread holding the outer perimeter exclusively.
2142 			 * Hence it would be sufficient to wait for the
2143 			 * exclusive holder of the outer perimeter to drain
2144 			 * the inner and outer syncqs. But we will not depend
2145 			 * on this feature and hence check the inner syncqs
2146 			 * separately.
2147 			 */
2148 			wait_syncq(outer);
2149 		}
2150 
2151 
2152 		/*
2153 		 * There could be messages destined for
2154 		 * this queue. Let the exclusive holder
2155 		 * drain it.
2156 		 */
2157 
2158 		wait_syncq(sq);
2159 		ASSERT((rq->q_flag & QPERMOD) ||
2160 		    ((rq->q_syncq->sq_head == NULL) &&
2161 		    (_WR(rq)->q_syncq->sq_head == NULL)));
2162 	}
2163 
2164 	/*
2165 	 * We haven't taken care of QPERMOD case yet. QPERMOD is a special
2166 	 * case as we don't disable its syncq or remove it off the syncq
2167 	 * service list.
2168 	 */
2169 	if (rq->q_flag & QPERMOD) {
2170 		syncq_t	*sq = rq->q_syncq;
2171 
2172 		mutex_enter(SQLOCK(sq));
2173 		while (rq->q_sqflags & Q_SQQUEUED) {
2174 			sq->sq_flags |= SQ_WANTWAKEUP;
2175 			cv_wait(&sq->sq_wait, SQLOCK(sq));
2176 		}
2177 		mutex_exit(SQLOCK(sq));
2178 	}
2179 
2180 	/*
2181 	 * flush_syncq changes states only when there are some messages to
2182 	 * free, i.e. when it returns non-zero value to return.
2183 	 */
2184 	ASSERT(flush_syncq(rq->q_syncq, rq) == 0);
2185 	ASSERT(flush_syncq(wrq->q_syncq, wrq) == 0);
2186 
2187 	/*
2188 	 * Nobody else should know about this queue now.
2189 	 * If the mux did not process the messages before
2190 	 * acking the I_UNLINK, free them now.
2191 	 */
2192 
2193 	flushq(rq, FLUSHALL);
2194 	flushq(_WR(rq), FLUSHALL);
2195 
2196 	/*
2197 	 * Convert the mux lower queue into a stream head queue.
2198 	 * Turn off STPLEX before we turn on the stream by removing the passq.
2199 	 */
2200 	rq->q_ptr = wrq->q_ptr = stpdown;
2201 	setq(rq, &strdata, &stwdata, NULL, QMTSAFE, SQ_CI|SQ_CO, B_TRUE);
2202 
2203 	ASSERT((rq->q_flag & QMT_TYPEMASK) == QMTSAFE);
2204 	ASSERT(rq->q_syncq == SQ(rq) && _WR(rq)->q_syncq == SQ(rq));
2205 
2206 	enable_svc(rq);
2207 
2208 	/*
2209 	 * Now it is a proper stream, so STPLEX is cleared. But STRPLUMB still
2210 	 * needs to be set to prevent reopen() of the stream - such reopen may
2211 	 * try to call non-existent pass queue open routine and panic.
2212 	 */
2213 	mutex_enter(&stpdown->sd_lock);
2214 	stpdown->sd_flag &= ~STPLEX;
2215 	mutex_exit(&stpdown->sd_lock);
2216 
2217 	ASSERT(((flag & LINKTYPEMASK) == LINKNORMAL) ||
2218 	    ((flag & LINKTYPEMASK) == LINKPERSIST));
2219 
2220 	/* clean up the layered driver linkages */
2221 	if ((flag & LINKTYPEMASK) == LINKNORMAL) {
2222 		ldi_munlink_fp(stp, fpdown, LINKNORMAL);
2223 	} else {
2224 		ldi_munlink_fp(stp, fpdown, LINKPERSIST);
2225 	}
2226 
2227 	link_rempassthru(passq);
2228 
2229 	/*
2230 	 * Now all plumbing changes are finished and STRPLUMB is no
2231 	 * longer needed.
2232 	 */
2233 	mutex_enter(&stpdown->sd_lock);
2234 	stpdown->sd_flag &= ~STRPLUMB;
2235 	cv_broadcast(&stpdown->sd_monitor);
2236 	mutex_exit(&stpdown->sd_lock);
2237 
2238 	(void) closef(fpdown);
2239 	return (0);
2240 }
2241 
2242 /*
2243  * Unlink all multiplexor links for which stp is the controlling stream.
2244  * Return 0, or a non-zero errno on failure.
2245  */
2246 int
2247 munlinkall(stdata_t *stp, int flag, cred_t *crp, int *rvalp, str_stack_t *ss)
2248 {
2249 	linkinfo_t *linkp;
2250 	int error = 0;
2251 
2252 	mutex_enter(&muxifier);
2253 	while (linkp = findlinks(stp, 0, flag, ss)) {
2254 		/*
2255 		 * munlink() releases the muxifier lock.
2256 		 */
2257 		if (error = munlink(stp, linkp, flag, crp, rvalp, ss))
2258 			return (error);
2259 		mutex_enter(&muxifier);
2260 	}
2261 	mutex_exit(&muxifier);
2262 	return (0);
2263 }
2264 
2265 /*
2266  * A multiplexor link has been made. Add an
2267  * edge to the directed graph.
2268  */
2269 void
2270 mux_addedge(stdata_t *upstp, stdata_t *lostp, int muxid, str_stack_t *ss)
2271 {
2272 	struct mux_node *np;
2273 	struct mux_edge *ep;
2274 	major_t upmaj;
2275 	major_t lomaj;
2276 
2277 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2278 	lomaj = getmajor(lostp->sd_vnode->v_rdev);
2279 	np = &ss->ss_mux_nodes[upmaj];
2280 	if (np->mn_outp) {
2281 		ep = np->mn_outp;
2282 		while (ep->me_nextp)
2283 			ep = ep->me_nextp;
2284 		ep->me_nextp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2285 		ep = ep->me_nextp;
2286 	} else {
2287 		np->mn_outp = kmem_alloc(sizeof (struct mux_edge), KM_SLEEP);
2288 		ep = np->mn_outp;
2289 	}
2290 	ep->me_nextp = NULL;
2291 	ep->me_muxid = muxid;
2292 	/*
2293 	 * Save the dev_t for the purposes of str_stack_shutdown.
2294 	 * str_stack_shutdown assumes that the device allows reopen, since
2295 	 * this dev_t is the one after any cloning by xx_open().
2296 	 * Would prefer finding the dev_t from before any cloning,
2297 	 * but specfs doesn't retain that.
2298 	 */
2299 	ep->me_dev = upstp->sd_vnode->v_rdev;
2300 	if (lostp->sd_vnode->v_type == VFIFO)
2301 		ep->me_nodep = NULL;
2302 	else
2303 		ep->me_nodep = &ss->ss_mux_nodes[lomaj];
2304 }
2305 
2306 /*
2307  * A multiplexor link has been removed. Remove the
2308  * edge in the directed graph.
2309  */
2310 void
2311 mux_rmvedge(stdata_t *upstp, int muxid, str_stack_t *ss)
2312 {
2313 	struct mux_node *np;
2314 	struct mux_edge *ep;
2315 	struct mux_edge *pep = NULL;
2316 	major_t upmaj;
2317 
2318 	upmaj = getmajor(upstp->sd_vnode->v_rdev);
2319 	np = &ss->ss_mux_nodes[upmaj];
2320 	ASSERT(np->mn_outp != NULL);
2321 	ep = np->mn_outp;
2322 	while (ep) {
2323 		if (ep->me_muxid == muxid) {
2324 			if (pep)
2325 				pep->me_nextp = ep->me_nextp;
2326 			else
2327 				np->mn_outp = ep->me_nextp;
2328 			kmem_free(ep, sizeof (struct mux_edge));
2329 			return;
2330 		}
2331 		pep = ep;
2332 		ep = ep->me_nextp;
2333 	}
2334 	ASSERT(0);	/* should not reach here */
2335 }
2336 
2337 /*
2338  * Translate the device flags (from conf.h) to the corresponding
2339  * qflag and sq_flag (type) values.
2340  */
2341 int
2342 devflg_to_qflag(struct streamtab *stp, uint32_t devflag, uint32_t *qflagp,
2343 	uint32_t *sqtypep)
2344 {
2345 	uint32_t qflag = 0;
2346 	uint32_t sqtype = 0;
2347 
2348 	if (devflag & _D_OLD)
2349 		goto bad;
2350 
2351 	/* Inner perimeter presence and scope */
2352 	switch (devflag & D_MTINNER_MASK) {
2353 	case D_MP:
2354 		qflag |= QMTSAFE;
2355 		sqtype |= SQ_CI;
2356 		break;
2357 	case D_MTPERQ|D_MP:
2358 		qflag |= QPERQ;
2359 		break;
2360 	case D_MTQPAIR|D_MP:
2361 		qflag |= QPAIR;
2362 		break;
2363 	case D_MTPERMOD|D_MP:
2364 		qflag |= QPERMOD;
2365 		break;
2366 	default:
2367 		goto bad;
2368 	}
2369 
2370 	/* Outer perimeter */
2371 	if (devflag & D_MTOUTPERIM) {
2372 		switch (devflag & D_MTINNER_MASK) {
2373 		case D_MP:
2374 		case D_MTPERQ|D_MP:
2375 		case D_MTQPAIR|D_MP:
2376 			break;
2377 		default:
2378 			goto bad;
2379 		}
2380 		qflag |= QMTOUTPERIM;
2381 	}
2382 
2383 	/* Inner perimeter modifiers */
2384 	if (devflag & D_MTINNER_MOD) {
2385 		switch (devflag & D_MTINNER_MASK) {
2386 		case D_MP:
2387 			goto bad;
2388 		default:
2389 			break;
2390 		}
2391 		if (devflag & D_MTPUTSHARED)
2392 			sqtype |= SQ_CIPUT;
2393 		if (devflag & _D_MTOCSHARED) {
2394 			/*
2395 			 * The code in putnext assumes that it has the
2396 			 * highest concurrency by not checking sq_count.
2397 			 * Thus _D_MTOCSHARED can only be supported when
2398 			 * D_MTPUTSHARED is set.
2399 			 */
2400 			if (!(devflag & D_MTPUTSHARED))
2401 				goto bad;
2402 			sqtype |= SQ_CIOC;
2403 		}
2404 		if (devflag & _D_MTCBSHARED) {
2405 			/*
2406 			 * The code in putnext assumes that it has the
2407 			 * highest concurrency by not checking sq_count.
2408 			 * Thus _D_MTCBSHARED can only be supported when
2409 			 * D_MTPUTSHARED is set.
2410 			 */
2411 			if (!(devflag & D_MTPUTSHARED))
2412 				goto bad;
2413 			sqtype |= SQ_CICB;
2414 		}
2415 		if (devflag & _D_MTSVCSHARED) {
2416 			/*
2417 			 * The code in putnext assumes that it has the
2418 			 * highest concurrency by not checking sq_count.
2419 			 * Thus _D_MTSVCSHARED can only be supported when
2420 			 * D_MTPUTSHARED is set. Also _D_MTSVCSHARED is
2421 			 * supported only for QPERMOD.
2422 			 */
2423 			if (!(devflag & D_MTPUTSHARED) || !(qflag & QPERMOD))
2424 				goto bad;
2425 			sqtype |= SQ_CISVC;
2426 		}
2427 	}
2428 
2429 	/* Default outer perimeter concurrency */
2430 	sqtype |= SQ_CO;
2431 
2432 	/* Outer perimeter modifiers */
2433 	if (devflag & D_MTOCEXCL) {
2434 		if (!(devflag & D_MTOUTPERIM)) {
2435 			/* No outer perimeter */
2436 			goto bad;
2437 		}
2438 		sqtype &= ~SQ_COOC;
2439 	}
2440 
2441 	/* Synchronous Streams extended qinit structure */
2442 	if (devflag & D_SYNCSTR)
2443 		qflag |= QSYNCSTR;
2444 
2445 	/*
2446 	 * Private flag used by a transport module to indicate
2447 	 * to sockfs that it supports direct-access mode without
2448 	 * having to go through STREAMS.
2449 	 */
2450 	if (devflag & _D_DIRECT) {
2451 		/* Reject unless the module is fully-MT (no perimeter) */
2452 		if ((qflag & QMT_TYPEMASK) != QMTSAFE)
2453 			goto bad;
2454 		qflag |= _QDIRECT;
2455 	}
2456 
2457 	*qflagp = qflag;
2458 	*sqtypep = sqtype;
2459 	return (0);
2460 
2461 bad:
2462 	cmn_err(CE_WARN,
2463 	    "stropen: bad MT flags (0x%x) in driver '%s'",
2464 	    (int)(qflag & D_MTSAFETY_MASK),
2465 	    stp->st_rdinit->qi_minfo->mi_idname);
2466 
2467 	return (EINVAL);
2468 }
2469 
2470 /*
2471  * Set the interface values for a pair of queues (qinit structure,
2472  * packet sizes, water marks).
2473  * setq assumes that the caller does not have a claim (entersq or claimq)
2474  * on the queue.
2475  */
2476 void
2477 setq(queue_t *rq, struct qinit *rinit, struct qinit *winit,
2478     perdm_t *dmp, uint32_t qflag, uint32_t sqtype, boolean_t lock_needed)
2479 {
2480 	queue_t *wq;
2481 	syncq_t	*sq, *outer;
2482 
2483 	ASSERT(rq->q_flag & QREADR);
2484 	ASSERT((qflag & QMT_TYPEMASK) != 0);
2485 	IMPLY((qflag & (QPERMOD | QMTOUTPERIM)), dmp != NULL);
2486 
2487 	wq = _WR(rq);
2488 	rq->q_qinfo = rinit;
2489 	rq->q_hiwat = rinit->qi_minfo->mi_hiwat;
2490 	rq->q_lowat = rinit->qi_minfo->mi_lowat;
2491 	rq->q_minpsz = rinit->qi_minfo->mi_minpsz;
2492 	rq->q_maxpsz = rinit->qi_minfo->mi_maxpsz;
2493 	wq->q_qinfo = winit;
2494 	wq->q_hiwat = winit->qi_minfo->mi_hiwat;
2495 	wq->q_lowat = winit->qi_minfo->mi_lowat;
2496 	wq->q_minpsz = winit->qi_minfo->mi_minpsz;
2497 	wq->q_maxpsz = winit->qi_minfo->mi_maxpsz;
2498 
2499 	/* Remove old syncqs */
2500 	sq = rq->q_syncq;
2501 	outer = sq->sq_outer;
2502 	if (outer != NULL) {
2503 		ASSERT(wq->q_syncq->sq_outer == outer);
2504 		outer_remove(outer, rq->q_syncq);
2505 		if (wq->q_syncq != rq->q_syncq)
2506 			outer_remove(outer, wq->q_syncq);
2507 	}
2508 	ASSERT(sq->sq_outer == NULL);
2509 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2510 
2511 	if (sq != SQ(rq)) {
2512 		if (!(rq->q_flag & QPERMOD))
2513 			free_syncq(sq);
2514 		if (wq->q_syncq == rq->q_syncq)
2515 			wq->q_syncq = NULL;
2516 		rq->q_syncq = NULL;
2517 	}
2518 	if (wq->q_syncq != NULL && wq->q_syncq != sq &&
2519 	    wq->q_syncq != SQ(rq)) {
2520 		free_syncq(wq->q_syncq);
2521 		wq->q_syncq = NULL;
2522 	}
2523 	ASSERT(rq->q_syncq == NULL || (rq->q_syncq->sq_head == NULL &&
2524 	    rq->q_syncq->sq_tail == NULL));
2525 	ASSERT(wq->q_syncq == NULL || (wq->q_syncq->sq_head == NULL &&
2526 	    wq->q_syncq->sq_tail == NULL));
2527 
2528 	if (!(rq->q_flag & QPERMOD) &&
2529 	    rq->q_syncq != NULL && rq->q_syncq->sq_ciputctrl != NULL) {
2530 		ASSERT(rq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2531 		SUMCHECK_CIPUTCTRL_COUNTS(rq->q_syncq->sq_ciputctrl,
2532 		    rq->q_syncq->sq_nciputctrl, 0);
2533 		ASSERT(ciputctrl_cache != NULL);
2534 		kmem_cache_free(ciputctrl_cache, rq->q_syncq->sq_ciputctrl);
2535 		rq->q_syncq->sq_ciputctrl = NULL;
2536 		rq->q_syncq->sq_nciputctrl = 0;
2537 	}
2538 
2539 	if (!(wq->q_flag & QPERMOD) &&
2540 	    wq->q_syncq != NULL && wq->q_syncq->sq_ciputctrl != NULL) {
2541 		ASSERT(wq->q_syncq->sq_nciputctrl == n_ciputctrl - 1);
2542 		SUMCHECK_CIPUTCTRL_COUNTS(wq->q_syncq->sq_ciputctrl,
2543 		    wq->q_syncq->sq_nciputctrl, 0);
2544 		ASSERT(ciputctrl_cache != NULL);
2545 		kmem_cache_free(ciputctrl_cache, wq->q_syncq->sq_ciputctrl);
2546 		wq->q_syncq->sq_ciputctrl = NULL;
2547 		wq->q_syncq->sq_nciputctrl = 0;
2548 	}
2549 
2550 	sq = SQ(rq);
2551 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
2552 	ASSERT(sq->sq_outer == NULL);
2553 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
2554 
2555 	/*
2556 	 * Create syncqs based on qflag and sqtype. Set the SQ_TYPES_IN_FLAGS
2557 	 * bits in sq_flag based on the sqtype.
2558 	 */
2559 	ASSERT((sq->sq_flags & ~SQ_TYPES_IN_FLAGS) == 0);
2560 
2561 	rq->q_syncq = wq->q_syncq = sq;
2562 	sq->sq_type = sqtype;
2563 	sq->sq_flags = (sqtype & SQ_TYPES_IN_FLAGS);
2564 
2565 	/*
2566 	 *  We are making sq_svcflags zero,
2567 	 *  resetting SQ_DISABLED in case it was set by
2568 	 *  wait_svc() in the munlink path.
2569 	 *
2570 	 */
2571 	ASSERT((sq->sq_svcflags & SQ_SERVICE) == 0);
2572 	sq->sq_svcflags = 0;
2573 
2574 	/*
2575 	 * We need to acquire the lock here for the mlink and munlink case,
2576 	 * where canputnext, backenable, etc can access the q_flag.
2577 	 */
2578 	if (lock_needed) {
2579 		mutex_enter(QLOCK(rq));
2580 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2581 		mutex_exit(QLOCK(rq));
2582 		mutex_enter(QLOCK(wq));
2583 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2584 		mutex_exit(QLOCK(wq));
2585 	} else {
2586 		rq->q_flag = (rq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2587 		wq->q_flag = (wq->q_flag & ~QMT_TYPEMASK) | QWANTR | qflag;
2588 	}
2589 
2590 	if (qflag & QPERQ) {
2591 		/* Allocate a separate syncq for the write side */
2592 		sq = new_syncq();
2593 		sq->sq_type = rq->q_syncq->sq_type;
2594 		sq->sq_flags = rq->q_syncq->sq_flags;
2595 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2596 		    sq->sq_oprev == NULL);
2597 		wq->q_syncq = sq;
2598 	}
2599 	if (qflag & QPERMOD) {
2600 		sq = dmp->dm_sq;
2601 
2602 		/*
2603 		 * Assert that we do have an inner perimeter syncq and that it
2604 		 * does not have an outer perimeter associated with it.
2605 		 */
2606 		ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
2607 		    sq->sq_oprev == NULL);
2608 		rq->q_syncq = wq->q_syncq = sq;
2609 	}
2610 	if (qflag & QMTOUTPERIM) {
2611 		outer = dmp->dm_sq;
2612 
2613 		ASSERT(outer->sq_outer == NULL);
2614 		outer_insert(outer, rq->q_syncq);
2615 		if (wq->q_syncq != rq->q_syncq)
2616 			outer_insert(outer, wq->q_syncq);
2617 	}
2618 	ASSERT((rq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2619 	    (rq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2620 	ASSERT((wq->q_syncq->sq_flags & SQ_TYPES_IN_FLAGS) ==
2621 	    (wq->q_syncq->sq_type & SQ_TYPES_IN_FLAGS));
2622 	ASSERT((rq->q_flag & QMT_TYPEMASK) == (qflag & QMT_TYPEMASK));
2623 
2624 	/*
2625 	 * Initialize struio() types.
2626 	 */
2627 	rq->q_struiot =
2628 	    (rq->q_flag & QSYNCSTR) ? rinit->qi_struiot : STRUIOT_NONE;
2629 	wq->q_struiot =
2630 	    (wq->q_flag & QSYNCSTR) ? winit->qi_struiot : STRUIOT_NONE;
2631 }
2632 
2633 perdm_t *
2634 hold_dm(struct streamtab *str, uint32_t qflag, uint32_t sqtype)
2635 {
2636 	syncq_t	*sq;
2637 	perdm_t	**pp;
2638 	perdm_t	*p;
2639 	perdm_t	*dmp;
2640 
2641 	ASSERT(str != NULL);
2642 	ASSERT(qflag & (QPERMOD | QMTOUTPERIM));
2643 
2644 	rw_enter(&perdm_rwlock, RW_READER);
2645 	for (p = perdm_list; p != NULL; p = p->dm_next) {
2646 		if (p->dm_str == str) {	/* found one */
2647 			atomic_add_32(&(p->dm_ref), 1);
2648 			rw_exit(&perdm_rwlock);
2649 			return (p);
2650 		}
2651 	}
2652 	rw_exit(&perdm_rwlock);
2653 
2654 	sq = new_syncq();
2655 	if (qflag & QPERMOD) {
2656 		sq->sq_type = sqtype | SQ_PERMOD;
2657 		sq->sq_flags = sqtype & SQ_TYPES_IN_FLAGS;
2658 	} else {
2659 		ASSERT(qflag & QMTOUTPERIM);
2660 		sq->sq_onext = sq->sq_oprev = sq;
2661 	}
2662 
2663 	dmp = kmem_alloc(sizeof (perdm_t), KM_SLEEP);
2664 	dmp->dm_sq = sq;
2665 	dmp->dm_str = str;
2666 	dmp->dm_ref = 1;
2667 	dmp->dm_next = NULL;
2668 
2669 	rw_enter(&perdm_rwlock, RW_WRITER);
2670 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next)) {
2671 		if (p->dm_str == str) {	/* already present */
2672 			p->dm_ref++;
2673 			rw_exit(&perdm_rwlock);
2674 			free_syncq(sq);
2675 			kmem_free(dmp, sizeof (perdm_t));
2676 			return (p);
2677 		}
2678 	}
2679 
2680 	*pp = dmp;
2681 	rw_exit(&perdm_rwlock);
2682 	return (dmp);
2683 }
2684 
2685 void
2686 rele_dm(perdm_t *dmp)
2687 {
2688 	perdm_t **pp;
2689 	perdm_t *p;
2690 
2691 	rw_enter(&perdm_rwlock, RW_WRITER);
2692 	ASSERT(dmp->dm_ref > 0);
2693 
2694 	if (--dmp->dm_ref > 0) {
2695 		rw_exit(&perdm_rwlock);
2696 		return;
2697 	}
2698 
2699 	for (pp = &perdm_list; (p = *pp) != NULL; pp = &(p->dm_next))
2700 		if (p == dmp)
2701 			break;
2702 	ASSERT(p == dmp);
2703 	*pp = p->dm_next;
2704 	rw_exit(&perdm_rwlock);
2705 
2706 	/*
2707 	 * Wait for any background processing that relies on the
2708 	 * syncq to complete before it is freed.
2709 	 */
2710 	wait_sq_svc(p->dm_sq);
2711 	free_syncq(p->dm_sq);
2712 	kmem_free(p, sizeof (perdm_t));
2713 }
2714 
2715 /*
2716  * Make a protocol message given control and data buffers.
2717  * n.b., this can block; be careful of what locks you hold when calling it.
2718  *
2719  * If sd_maxblk is less than *iosize this routine can fail part way through
2720  * (due to an allocation failure). In this case on return *iosize will contain
2721  * the amount that was consumed. Otherwise *iosize will not be modified
2722  * i.e. it will contain the amount that was consumed.
2723  */
2724 int
2725 strmakemsg(
2726 	struct strbuf *mctl,
2727 	ssize_t *iosize,
2728 	struct uio *uiop,
2729 	stdata_t *stp,
2730 	int32_t flag,
2731 	mblk_t **mpp)
2732 {
2733 	mblk_t *mpctl = NULL;
2734 	mblk_t *mpdata = NULL;
2735 	int error;
2736 
2737 	ASSERT(uiop != NULL);
2738 
2739 	*mpp = NULL;
2740 	/* Create control part, if any */
2741 	if ((mctl != NULL) && (mctl->len >= 0)) {
2742 		error = strmakectl(mctl, flag, uiop->uio_fmode, &mpctl);
2743 		if (error)
2744 			return (error);
2745 	}
2746 	/* Create data part, if any */
2747 	if (*iosize >= 0) {
2748 		error = strmakedata(iosize, uiop, stp, flag, &mpdata);
2749 		if (error) {
2750 			freemsg(mpctl);
2751 			return (error);
2752 		}
2753 	}
2754 	if (mpctl != NULL) {
2755 		if (mpdata != NULL)
2756 			linkb(mpctl, mpdata);
2757 		*mpp = mpctl;
2758 	} else {
2759 		*mpp = mpdata;
2760 	}
2761 	return (0);
2762 }
2763 
2764 /*
2765  * Make the control part of a protocol message given a control buffer.
2766  * n.b., this can block; be careful of what locks you hold when calling it.
2767  */
2768 int
2769 strmakectl(
2770 	struct strbuf *mctl,
2771 	int32_t flag,
2772 	int32_t fflag,
2773 	mblk_t **mpp)
2774 {
2775 	mblk_t *bp = NULL;
2776 	unsigned char msgtype;
2777 	int error = 0;
2778 	cred_t *cr = CRED();
2779 
2780 	/* We do not support interrupt threads using the stream head to send */
2781 	ASSERT(cr != NULL);
2782 
2783 	*mpp = NULL;
2784 	/*
2785 	 * Create control part of message, if any.
2786 	 */
2787 	if ((mctl != NULL) && (mctl->len >= 0)) {
2788 		caddr_t base;
2789 		int ctlcount;
2790 		int allocsz;
2791 
2792 		if (flag & RS_HIPRI)
2793 			msgtype = M_PCPROTO;
2794 		else
2795 			msgtype = M_PROTO;
2796 
2797 		ctlcount = mctl->len;
2798 		base = mctl->buf;
2799 
2800 		/*
2801 		 * Give modules a better chance to reuse M_PROTO/M_PCPROTO
2802 		 * blocks by increasing the size to something more usable.
2803 		 */
2804 		allocsz = MAX(ctlcount, 64);
2805 
2806 		/*
2807 		 * Range checking has already been done; simply try
2808 		 * to allocate a message block for the ctl part.
2809 		 */
2810 		while ((bp = allocb_cred(allocsz, cr,
2811 		    curproc->p_pid)) == NULL) {
2812 			if (fflag & (FNDELAY|FNONBLOCK))
2813 				return (EAGAIN);
2814 			if (error = strwaitbuf(allocsz, BPRI_MED))
2815 				return (error);
2816 		}
2817 
2818 		bp->b_datap->db_type = msgtype;
2819 		if (copyin(base, bp->b_wptr, ctlcount)) {
2820 			freeb(bp);
2821 			return (EFAULT);
2822 		}
2823 		bp->b_wptr += ctlcount;
2824 	}
2825 	*mpp = bp;
2826 	return (0);
2827 }
2828 
2829 /*
2830  * Make a protocol message given data buffers.
2831  * n.b., this can block; be careful of what locks you hold when calling it.
2832  *
2833  * If sd_maxblk is less than *iosize this routine can fail part way through
2834  * (due to an allocation failure). In this case on return *iosize will contain
2835  * the amount that was consumed. Otherwise *iosize will not be modified
2836  * i.e. it will contain the amount that was consumed.
2837  */
2838 int
2839 strmakedata(
2840 	ssize_t   *iosize,
2841 	struct uio *uiop,
2842 	stdata_t *stp,
2843 	int32_t flag,
2844 	mblk_t **mpp)
2845 {
2846 	mblk_t *mp = NULL;
2847 	mblk_t *bp;
2848 	int wroff = (int)stp->sd_wroff;
2849 	int tail_len = (int)stp->sd_tail;
2850 	int extra = wroff + tail_len;
2851 	int error = 0;
2852 	ssize_t maxblk;
2853 	ssize_t count = *iosize;
2854 	cred_t *cr;
2855 
2856 	*mpp = NULL;
2857 	if (count < 0)
2858 		return (0);
2859 
2860 	/* We do not support interrupt threads using the stream head to send */
2861 	cr = CRED();
2862 	ASSERT(cr != NULL);
2863 
2864 	maxblk = stp->sd_maxblk;
2865 	if (maxblk == INFPSZ)
2866 		maxblk = count;
2867 
2868 	/*
2869 	 * Create data part of message, if any.
2870 	 */
2871 	do {
2872 		ssize_t size;
2873 		dblk_t  *dp;
2874 
2875 		ASSERT(uiop);
2876 
2877 		size = MIN(count, maxblk);
2878 
2879 		while ((bp = allocb_cred(size + extra, cr,
2880 		    curproc->p_pid)) == NULL) {
2881 			error = EAGAIN;
2882 			if ((uiop->uio_fmode & (FNDELAY|FNONBLOCK)) ||
2883 			    (error = strwaitbuf(size + extra, BPRI_MED)) != 0) {
2884 				if (count == *iosize) {
2885 					freemsg(mp);
2886 					return (error);
2887 				} else {
2888 					*iosize -= count;
2889 					*mpp = mp;
2890 					return (0);
2891 				}
2892 			}
2893 		}
2894 		dp = bp->b_datap;
2895 		dp->db_cpid = curproc->p_pid;
2896 		ASSERT(wroff <= dp->db_lim - bp->b_wptr);
2897 		bp->b_wptr = bp->b_rptr = bp->b_rptr + wroff;
2898 
2899 		if (flag & STRUIO_POSTPONE) {
2900 			/*
2901 			 * Setup the stream uio portion of the
2902 			 * dblk for subsequent use by struioget().
2903 			 */
2904 			dp->db_struioflag = STRUIO_SPEC;
2905 			dp->db_cksumstart = 0;
2906 			dp->db_cksumstuff = 0;
2907 			dp->db_cksumend = size;
2908 			*(long long *)dp->db_struioun.data = 0ll;
2909 			bp->b_wptr += size;
2910 		} else {
2911 			if (stp->sd_copyflag & STRCOPYCACHED)
2912 				uiop->uio_extflg |= UIO_COPY_CACHED;
2913 
2914 			if (size != 0) {
2915 				error = uiomove(bp->b_wptr, size, UIO_WRITE,
2916 				    uiop);
2917 				if (error != 0) {
2918 					freeb(bp);
2919 					freemsg(mp);
2920 					return (error);
2921 				}
2922 			}
2923 			bp->b_wptr += size;
2924 
2925 			if (stp->sd_wputdatafunc != NULL) {
2926 				mblk_t *newbp;
2927 
2928 				newbp = (stp->sd_wputdatafunc)(stp->sd_vnode,
2929 				    bp, NULL, NULL, NULL, NULL);
2930 				if (newbp == NULL) {
2931 					freeb(bp);
2932 					freemsg(mp);
2933 					return (ECOMM);
2934 				}
2935 				bp = newbp;
2936 			}
2937 		}
2938 
2939 		count -= size;
2940 
2941 		if (mp == NULL)
2942 			mp = bp;
2943 		else
2944 			linkb(mp, bp);
2945 	} while (count > 0);
2946 
2947 	*mpp = mp;
2948 	return (0);
2949 }
2950 
2951 /*
2952  * Wait for a buffer to become available. Return non-zero errno
2953  * if not able to wait, 0 if buffer is probably there.
2954  */
2955 int
2956 strwaitbuf(size_t size, int pri)
2957 {
2958 	bufcall_id_t id;
2959 
2960 	mutex_enter(&bcall_monitor);
2961 	if ((id = bufcall(size, pri, (void (*)(void *))cv_broadcast,
2962 	    &ttoproc(curthread)->p_flag_cv)) == 0) {
2963 		mutex_exit(&bcall_monitor);
2964 		return (ENOSR);
2965 	}
2966 	if (!cv_wait_sig(&(ttoproc(curthread)->p_flag_cv), &bcall_monitor)) {
2967 		unbufcall(id);
2968 		mutex_exit(&bcall_monitor);
2969 		return (EINTR);
2970 	}
2971 	unbufcall(id);
2972 	mutex_exit(&bcall_monitor);
2973 	return (0);
2974 }
2975 
2976 /*
2977  * This function waits for a read or write event to happen on a stream.
2978  * fmode can specify FNDELAY and/or FNONBLOCK.
2979  * The timeout is in ms with -1 meaning infinite.
2980  * The flag values work as follows:
2981  *	READWAIT	Check for read side errors, send M_READ
2982  *	GETWAIT		Check for read side errors, no M_READ
2983  *	WRITEWAIT	Check for write side errors.
2984  *	NOINTR		Do not return error if nonblocking or timeout.
2985  * 	STR_NOERROR	Ignore all errors except STPLEX.
2986  *	STR_NOSIG	Ignore/hold signals during the duration of the call.
2987  *	STR_PEEK	Pass through the strgeterr().
2988  */
2989 int
2990 strwaitq(stdata_t *stp, int flag, ssize_t count, int fmode, clock_t timout,
2991     int *done)
2992 {
2993 	int slpflg, errs;
2994 	int error;
2995 	kcondvar_t *sleepon;
2996 	mblk_t *mp;
2997 	ssize_t *rd_count;
2998 	clock_t rval;
2999 
3000 	ASSERT(MUTEX_HELD(&stp->sd_lock));
3001 	if ((flag & READWAIT) || (flag & GETWAIT)) {
3002 		slpflg = RSLEEP;
3003 		sleepon = &_RD(stp->sd_wrq)->q_wait;
3004 		errs = STRDERR|STPLEX;
3005 	} else {
3006 		slpflg = WSLEEP;
3007 		sleepon = &stp->sd_wrq->q_wait;
3008 		errs = STWRERR|STRHUP|STPLEX;
3009 	}
3010 	if (flag & STR_NOERROR)
3011 		errs = STPLEX;
3012 
3013 	if (stp->sd_wakeq & slpflg) {
3014 		/*
3015 		 * A strwakeq() is pending, no need to sleep.
3016 		 */
3017 		stp->sd_wakeq &= ~slpflg;
3018 		*done = 0;
3019 		return (0);
3020 	}
3021 
3022 	if (stp->sd_flag & errs) {
3023 		/*
3024 		 * Check for errors before going to sleep since the
3025 		 * caller might not have checked this while holding
3026 		 * sd_lock.
3027 		 */
3028 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3029 		if (error != 0) {
3030 			*done = 1;
3031 			return (error);
3032 		}
3033 	}
3034 
3035 	/*
3036 	 * If any module downstream has requested read notification
3037 	 * by setting SNDMREAD flag using M_SETOPTS, send a message
3038 	 * down stream.
3039 	 */
3040 	if ((flag & READWAIT) && (stp->sd_flag & SNDMREAD)) {
3041 		mutex_exit(&stp->sd_lock);
3042 		if (!(mp = allocb_wait(sizeof (ssize_t), BPRI_MED,
3043 		    (flag & STR_NOSIG), &error))) {
3044 			mutex_enter(&stp->sd_lock);
3045 			*done = 1;
3046 			return (error);
3047 		}
3048 		mp->b_datap->db_type = M_READ;
3049 		rd_count = (ssize_t *)mp->b_wptr;
3050 		*rd_count = count;
3051 		mp->b_wptr += sizeof (ssize_t);
3052 		/*
3053 		 * Send the number of bytes requested by the
3054 		 * read as the argument to M_READ.
3055 		 */
3056 		stream_willservice(stp);
3057 		putnext(stp->sd_wrq, mp);
3058 		stream_runservice(stp);
3059 		mutex_enter(&stp->sd_lock);
3060 
3061 		/*
3062 		 * If any data arrived due to inline processing
3063 		 * of putnext(), don't sleep.
3064 		 */
3065 		if (_RD(stp->sd_wrq)->q_first != NULL) {
3066 			*done = 0;
3067 			return (0);
3068 		}
3069 	}
3070 
3071 	if (fmode & (FNDELAY|FNONBLOCK)) {
3072 		if (!(flag & NOINTR))
3073 			error = EAGAIN;
3074 		else
3075 			error = 0;
3076 		*done = 1;
3077 		return (error);
3078 	}
3079 
3080 	stp->sd_flag |= slpflg;
3081 	TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAIT2,
3082 	    "strwaitq sleeps (2):%p, %X, %lX, %X, %p",
3083 	    stp, flag, count, fmode, done);
3084 
3085 	rval = str_cv_wait(sleepon, &stp->sd_lock, timout, flag & STR_NOSIG);
3086 	if (rval > 0) {
3087 		/* EMPTY */
3088 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_WAKE2,
3089 		    "strwaitq awakes(2):%X, %X, %X, %X, %X",
3090 		    stp, flag, count, fmode, done);
3091 	} else if (rval == 0) {
3092 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_INTR2,
3093 		    "strwaitq interrupt #2:%p, %X, %lX, %X, %p",
3094 		    stp, flag, count, fmode, done);
3095 		stp->sd_flag &= ~slpflg;
3096 		cv_broadcast(sleepon);
3097 		if (!(flag & NOINTR))
3098 			error = EINTR;
3099 		else
3100 			error = 0;
3101 		*done = 1;
3102 		return (error);
3103 	} else {
3104 		/* timeout */
3105 		TRACE_5(TR_FAC_STREAMS_FR, TR_STRWAITQ_TIME,
3106 		    "strwaitq timeout:%p, %X, %lX, %X, %p",
3107 		    stp, flag, count, fmode, done);
3108 		*done = 1;
3109 		if (!(flag & NOINTR))
3110 			return (ETIME);
3111 		else
3112 			return (0);
3113 	}
3114 	/*
3115 	 * If the caller implements delayed errors (i.e. queued after data)
3116 	 * we can not check for errors here since data as well as an
3117 	 * error might have arrived at the stream head. We return to
3118 	 * have the caller check the read queue before checking for errors.
3119 	 */
3120 	if ((stp->sd_flag & errs) && !(flag & STR_DELAYERR)) {
3121 		error = strgeterr(stp, errs, (flag & STR_PEEK));
3122 		if (error != 0) {
3123 			*done = 1;
3124 			return (error);
3125 		}
3126 	}
3127 	*done = 0;
3128 	return (0);
3129 }
3130 
3131 /*
3132  * Perform job control discipline access checks.
3133  * Return 0 for success and the errno for failure.
3134  */
3135 
3136 #define	cantsend(p, t, sig) \
3137 	(sigismember(&(p)->p_ignore, sig) || signal_is_blocked((t), sig))
3138 
3139 int
3140 straccess(struct stdata *stp, enum jcaccess mode)
3141 {
3142 	extern kcondvar_t lbolt_cv;	/* XXX: should be in a header file */
3143 	kthread_t *t = curthread;
3144 	proc_t *p = ttoproc(t);
3145 	sess_t *sp;
3146 
3147 	ASSERT(mutex_owned(&stp->sd_lock));
3148 
3149 	if (stp->sd_sidp == NULL || stp->sd_vnode->v_type == VFIFO)
3150 		return (0);
3151 
3152 	mutex_enter(&p->p_lock);		/* protects p_pgidp */
3153 
3154 	for (;;) {
3155 		mutex_enter(&p->p_splock);	/* protects p->p_sessp */
3156 		sp = p->p_sessp;
3157 		mutex_enter(&sp->s_lock);	/* protects sp->* */
3158 
3159 		/*
3160 		 * If this is not the calling process's controlling terminal
3161 		 * or if the calling process is already in the foreground
3162 		 * then allow access.
3163 		 */
3164 		if (sp->s_dev != stp->sd_vnode->v_rdev ||
3165 		    p->p_pgidp == stp->sd_pgidp) {
3166 			mutex_exit(&sp->s_lock);
3167 			mutex_exit(&p->p_splock);
3168 			mutex_exit(&p->p_lock);
3169 			return (0);
3170 		}
3171 
3172 		/*
3173 		 * Check to see if controlling terminal has been deallocated.
3174 		 */
3175 		if (sp->s_vp == NULL) {
3176 			if (!cantsend(p, t, SIGHUP))
3177 				sigtoproc(p, t, SIGHUP);
3178 			mutex_exit(&sp->s_lock);
3179 			mutex_exit(&p->p_splock);
3180 			mutex_exit(&p->p_lock);
3181 			return (EIO);
3182 		}
3183 
3184 		mutex_exit(&sp->s_lock);
3185 		mutex_exit(&p->p_splock);
3186 
3187 		if (mode == JCGETP) {
3188 			mutex_exit(&p->p_lock);
3189 			return (0);
3190 		}
3191 
3192 		if (mode == JCREAD) {
3193 			if (p->p_detached || cantsend(p, t, SIGTTIN)) {
3194 				mutex_exit(&p->p_lock);
3195 				return (EIO);
3196 			}
3197 			mutex_exit(&p->p_lock);
3198 			mutex_exit(&stp->sd_lock);
3199 			pgsignal(p->p_pgidp, SIGTTIN);
3200 			mutex_enter(&stp->sd_lock);
3201 			mutex_enter(&p->p_lock);
3202 		} else {  /* mode == JCWRITE or JCSETP */
3203 			if ((mode == JCWRITE && !(stp->sd_flag & STRTOSTOP)) ||
3204 			    cantsend(p, t, SIGTTOU)) {
3205 				mutex_exit(&p->p_lock);
3206 				return (0);
3207 			}
3208 			if (p->p_detached) {
3209 				mutex_exit(&p->p_lock);
3210 				return (EIO);
3211 			}
3212 			mutex_exit(&p->p_lock);
3213 			mutex_exit(&stp->sd_lock);
3214 			pgsignal(p->p_pgidp, SIGTTOU);
3215 			mutex_enter(&stp->sd_lock);
3216 			mutex_enter(&p->p_lock);
3217 		}
3218 
3219 		/*
3220 		 * We call cv_wait_sig_swap() to cause the appropriate
3221 		 * action for the jobcontrol signal to take place.
3222 		 * If the signal is being caught, we will take the
3223 		 * EINTR error return.  Otherwise, the default action
3224 		 * of causing the process to stop will take place.
3225 		 * In this case, we rely on the periodic cv_broadcast() on
3226 		 * &lbolt_cv to wake us up to loop around and test again.
3227 		 * We can't get here if the signal is ignored or
3228 		 * if the current thread is blocking the signal.
3229 		 */
3230 		mutex_exit(&stp->sd_lock);
3231 		if (!cv_wait_sig_swap(&lbolt_cv, &p->p_lock)) {
3232 			mutex_exit(&p->p_lock);
3233 			mutex_enter(&stp->sd_lock);
3234 			return (EINTR);
3235 		}
3236 		mutex_exit(&p->p_lock);
3237 		mutex_enter(&stp->sd_lock);
3238 		mutex_enter(&p->p_lock);
3239 	}
3240 }
3241 
3242 /*
3243  * Return size of message of block type (bp->b_datap->db_type)
3244  */
3245 size_t
3246 xmsgsize(mblk_t *bp)
3247 {
3248 	unsigned char type;
3249 	size_t count = 0;
3250 
3251 	type = bp->b_datap->db_type;
3252 
3253 	for (; bp; bp = bp->b_cont) {
3254 		if (type != bp->b_datap->db_type)
3255 			break;
3256 		ASSERT(bp->b_wptr >= bp->b_rptr);
3257 		count += bp->b_wptr - bp->b_rptr;
3258 	}
3259 	return (count);
3260 }
3261 
3262 /*
3263  * Allocate a stream head.
3264  */
3265 struct stdata *
3266 shalloc(queue_t *qp)
3267 {
3268 	stdata_t *stp;
3269 
3270 	stp = kmem_cache_alloc(stream_head_cache, KM_SLEEP);
3271 
3272 	stp->sd_wrq = _WR(qp);
3273 	stp->sd_strtab = NULL;
3274 	stp->sd_iocid = 0;
3275 	stp->sd_mate = NULL;
3276 	stp->sd_freezer = NULL;
3277 	stp->sd_refcnt = 0;
3278 	stp->sd_wakeq = 0;
3279 	stp->sd_anchor = 0;
3280 	stp->sd_struiowrq = NULL;
3281 	stp->sd_struiordq = NULL;
3282 	stp->sd_struiodnak = 0;
3283 	stp->sd_struionak = NULL;
3284 	stp->sd_t_audit_data = NULL;
3285 	stp->sd_rput_opt = 0;
3286 	stp->sd_wput_opt = 0;
3287 	stp->sd_read_opt = 0;
3288 	stp->sd_rprotofunc = strrput_proto;
3289 	stp->sd_rmiscfunc = strrput_misc;
3290 	stp->sd_rderrfunc = stp->sd_wrerrfunc = NULL;
3291 	stp->sd_rputdatafunc = stp->sd_wputdatafunc = NULL;
3292 	stp->sd_ciputctrl = NULL;
3293 	stp->sd_nciputctrl = 0;
3294 	stp->sd_qhead = NULL;
3295 	stp->sd_qtail = NULL;
3296 	stp->sd_servid = NULL;
3297 	stp->sd_nqueues = 0;
3298 	stp->sd_svcflags = 0;
3299 	stp->sd_copyflag = 0;
3300 
3301 	return (stp);
3302 }
3303 
3304 /*
3305  * Free a stream head.
3306  */
3307 void
3308 shfree(stdata_t *stp)
3309 {
3310 	ASSERT(MUTEX_NOT_HELD(&stp->sd_lock));
3311 
3312 	stp->sd_wrq = NULL;
3313 
3314 	mutex_enter(&stp->sd_qlock);
3315 	while (stp->sd_svcflags & STRS_SCHEDULED) {
3316 		STRSTAT(strwaits);
3317 		cv_wait(&stp->sd_qcv, &stp->sd_qlock);
3318 	}
3319 	mutex_exit(&stp->sd_qlock);
3320 
3321 	if (stp->sd_ciputctrl != NULL) {
3322 		ASSERT(stp->sd_nciputctrl == n_ciputctrl - 1);
3323 		SUMCHECK_CIPUTCTRL_COUNTS(stp->sd_ciputctrl,
3324 		    stp->sd_nciputctrl, 0);
3325 		ASSERT(ciputctrl_cache != NULL);
3326 		kmem_cache_free(ciputctrl_cache, stp->sd_ciputctrl);
3327 		stp->sd_ciputctrl = NULL;
3328 		stp->sd_nciputctrl = 0;
3329 	}
3330 	ASSERT(stp->sd_qhead == NULL);
3331 	ASSERT(stp->sd_qtail == NULL);
3332 	ASSERT(stp->sd_nqueues == 0);
3333 	kmem_cache_free(stream_head_cache, stp);
3334 }
3335 
3336 /*
3337  * Allocate a pair of queues and a syncq for the pair
3338  */
3339 queue_t *
3340 allocq(void)
3341 {
3342 	queinfo_t *qip;
3343 	queue_t *qp, *wqp;
3344 	syncq_t	*sq;
3345 
3346 	qip = kmem_cache_alloc(queue_cache, KM_SLEEP);
3347 
3348 	qp = &qip->qu_rqueue;
3349 	wqp = &qip->qu_wqueue;
3350 	sq = &qip->qu_syncq;
3351 
3352 	qp->q_last	= NULL;
3353 	qp->q_next	= NULL;
3354 	qp->q_ptr	= NULL;
3355 	qp->q_flag	= QUSE | QREADR;
3356 	qp->q_bandp	= NULL;
3357 	qp->q_stream	= NULL;
3358 	qp->q_syncq	= sq;
3359 	qp->q_nband	= 0;
3360 	qp->q_nfsrv	= NULL;
3361 	qp->q_draining	= 0;
3362 	qp->q_syncqmsgs	= 0;
3363 	qp->q_spri	= 0;
3364 	qp->q_qtstamp	= 0;
3365 	qp->q_sqtstamp	= 0;
3366 	qp->q_fp	= NULL;
3367 
3368 	wqp->q_last	= NULL;
3369 	wqp->q_next	= NULL;
3370 	wqp->q_ptr	= NULL;
3371 	wqp->q_flag	= QUSE;
3372 	wqp->q_bandp	= NULL;
3373 	wqp->q_stream	= NULL;
3374 	wqp->q_syncq	= sq;
3375 	wqp->q_nband	= 0;
3376 	wqp->q_nfsrv	= NULL;
3377 	wqp->q_draining	= 0;
3378 	wqp->q_syncqmsgs = 0;
3379 	wqp->q_qtstamp	= 0;
3380 	wqp->q_sqtstamp	= 0;
3381 	wqp->q_spri	= 0;
3382 
3383 	sq->sq_count	= 0;
3384 	sq->sq_rmqcount	= 0;
3385 	sq->sq_flags	= 0;
3386 	sq->sq_type	= 0;
3387 	sq->sq_callbflags = 0;
3388 	sq->sq_cancelid	= 0;
3389 	sq->sq_ciputctrl = NULL;
3390 	sq->sq_nciputctrl = 0;
3391 	sq->sq_needexcl = 0;
3392 	sq->sq_svcflags = 0;
3393 
3394 	return (qp);
3395 }
3396 
3397 /*
3398  * Free a pair of queues and the "attached" syncq.
3399  * Discard any messages left on the syncq(s), remove the syncq(s) from the
3400  * outer perimeter, and free the syncq(s) if they are not the "attached" syncq.
3401  */
3402 void
3403 freeq(queue_t *qp)
3404 {
3405 	qband_t *qbp, *nqbp;
3406 	syncq_t *sq, *outer;
3407 	queue_t *wqp = _WR(qp);
3408 
3409 	ASSERT(qp->q_flag & QREADR);
3410 
3411 	/*
3412 	 * If a previously dispatched taskq job is scheduled to run
3413 	 * sync_service() or a service routine is scheduled for the
3414 	 * queues about to be freed, wait here until all service is
3415 	 * done on the queue and all associated queues and syncqs.
3416 	 */
3417 	wait_svc(qp);
3418 
3419 	(void) flush_syncq(qp->q_syncq, qp);
3420 	(void) flush_syncq(wqp->q_syncq, wqp);
3421 	ASSERT(qp->q_syncqmsgs == 0 && wqp->q_syncqmsgs == 0);
3422 
3423 	/*
3424 	 * Flush the queues before q_next is set to NULL This is needed
3425 	 * in order to backenable any downstream queue before we go away.
3426 	 * Note: we are already removed from the stream so that the
3427 	 * backenabling will not cause any messages to be delivered to our
3428 	 * put procedures.
3429 	 */
3430 	flushq(qp, FLUSHALL);
3431 	flushq(wqp, FLUSHALL);
3432 
3433 	/* Tidy up - removeq only does a half-remove from stream */
3434 	qp->q_next = wqp->q_next = NULL;
3435 	ASSERT(!(qp->q_flag & QENAB));
3436 	ASSERT(!(wqp->q_flag & QENAB));
3437 
3438 	outer = qp->q_syncq->sq_outer;
3439 	if (outer != NULL) {
3440 		outer_remove(outer, qp->q_syncq);
3441 		if (wqp->q_syncq != qp->q_syncq)
3442 			outer_remove(outer, wqp->q_syncq);
3443 	}
3444 	/*
3445 	 * Free any syncqs that are outside what allocq returned.
3446 	 */
3447 	if (qp->q_syncq != SQ(qp) && !(qp->q_flag & QPERMOD))
3448 		free_syncq(qp->q_syncq);
3449 	if (qp->q_syncq != wqp->q_syncq && wqp->q_syncq != SQ(qp))
3450 		free_syncq(wqp->q_syncq);
3451 
3452 	ASSERT((qp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3453 	ASSERT((wqp->q_sqflags & (Q_SQQUEUED | Q_SQDRAINING)) == 0);
3454 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
3455 	ASSERT(MUTEX_NOT_HELD(QLOCK(wqp)));
3456 	sq = SQ(qp);
3457 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
3458 	ASSERT(sq->sq_head == NULL && sq->sq_tail == NULL);
3459 	ASSERT(sq->sq_outer == NULL);
3460 	ASSERT(sq->sq_onext == NULL && sq->sq_oprev == NULL);
3461 	ASSERT(sq->sq_callbpend == NULL);
3462 	ASSERT(sq->sq_needexcl == 0);
3463 
3464 	if (sq->sq_ciputctrl != NULL) {
3465 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
3466 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
3467 		    sq->sq_nciputctrl, 0);
3468 		ASSERT(ciputctrl_cache != NULL);
3469 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
3470 		sq->sq_ciputctrl = NULL;
3471 		sq->sq_nciputctrl = 0;
3472 	}
3473 
3474 	ASSERT(qp->q_first == NULL && wqp->q_first == NULL);
3475 	ASSERT(qp->q_count == 0 && wqp->q_count == 0);
3476 	ASSERT(qp->q_mblkcnt == 0 && wqp->q_mblkcnt == 0);
3477 
3478 	qp->q_flag &= ~QUSE;
3479 	wqp->q_flag &= ~QUSE;
3480 
3481 	/* NOTE: Uncomment the assert below once bugid 1159635 is fixed. */
3482 	/* ASSERT((qp->q_flag & QWANTW) == 0 && (wqp->q_flag & QWANTW) == 0); */
3483 
3484 	qbp = qp->q_bandp;
3485 	while (qbp) {
3486 		nqbp = qbp->qb_next;
3487 		freeband(qbp);
3488 		qbp = nqbp;
3489 	}
3490 	qbp = wqp->q_bandp;
3491 	while (qbp) {
3492 		nqbp = qbp->qb_next;
3493 		freeband(qbp);
3494 		qbp = nqbp;
3495 	}
3496 	kmem_cache_free(queue_cache, qp);
3497 }
3498 
3499 /*
3500  * Allocate a qband structure.
3501  */
3502 qband_t *
3503 allocband(void)
3504 {
3505 	qband_t *qbp;
3506 
3507 	qbp = kmem_cache_alloc(qband_cache, KM_NOSLEEP);
3508 	if (qbp == NULL)
3509 		return (NULL);
3510 
3511 	qbp->qb_next	= NULL;
3512 	qbp->qb_count	= 0;
3513 	qbp->qb_mblkcnt	= 0;
3514 	qbp->qb_first	= NULL;
3515 	qbp->qb_last	= NULL;
3516 	qbp->qb_flag	= 0;
3517 
3518 	return (qbp);
3519 }
3520 
3521 /*
3522  * Free a qband structure.
3523  */
3524 void
3525 freeband(qband_t *qbp)
3526 {
3527 	kmem_cache_free(qband_cache, qbp);
3528 }
3529 
3530 /*
3531  * Just like putnextctl(9F), except that allocb_wait() is used.
3532  *
3533  * Consolidation Private, and of course only callable from the stream head or
3534  * routines that may block.
3535  */
3536 int
3537 putnextctl_wait(queue_t *q, int type)
3538 {
3539 	mblk_t *bp;
3540 	int error;
3541 
3542 	if ((datamsg(type) && (type != M_DELAY)) ||
3543 	    (bp = allocb_wait(0, BPRI_HI, 0, &error)) == NULL)
3544 		return (0);
3545 
3546 	bp->b_datap->db_type = (unsigned char)type;
3547 	putnext(q, bp);
3548 	return (1);
3549 }
3550 
3551 /*
3552  * Run any possible bufcalls.
3553  */
3554 void
3555 runbufcalls(void)
3556 {
3557 	strbufcall_t *bcp;
3558 
3559 	mutex_enter(&bcall_monitor);
3560 	mutex_enter(&strbcall_lock);
3561 
3562 	if (strbcalls.bc_head) {
3563 		size_t count;
3564 		int nevent;
3565 
3566 		/*
3567 		 * count how many events are on the list
3568 		 * now so we can check to avoid looping
3569 		 * in low memory situations
3570 		 */
3571 		nevent = 0;
3572 		for (bcp = strbcalls.bc_head; bcp; bcp = bcp->bc_next)
3573 			nevent++;
3574 
3575 		/*
3576 		 * get estimate of available memory from kmem_avail().
3577 		 * awake all bufcall functions waiting for
3578 		 * memory whose request could be satisfied
3579 		 * by 'count' memory and let 'em fight for it.
3580 		 */
3581 		count = kmem_avail();
3582 		while ((bcp = strbcalls.bc_head) != NULL && nevent) {
3583 			STRSTAT(bufcalls);
3584 			--nevent;
3585 			if (bcp->bc_size <= count) {
3586 				bcp->bc_executor = curthread;
3587 				mutex_exit(&strbcall_lock);
3588 				(*bcp->bc_func)(bcp->bc_arg);
3589 				mutex_enter(&strbcall_lock);
3590 				bcp->bc_executor = NULL;
3591 				cv_broadcast(&bcall_cv);
3592 				strbcalls.bc_head = bcp->bc_next;
3593 				kmem_free(bcp, sizeof (strbufcall_t));
3594 			} else {
3595 				/*
3596 				 * too big, try again later - note
3597 				 * that nevent was decremented above
3598 				 * so we won't retry this one on this
3599 				 * iteration of the loop
3600 				 */
3601 				if (bcp->bc_next != NULL) {
3602 					strbcalls.bc_head = bcp->bc_next;
3603 					bcp->bc_next = NULL;
3604 					strbcalls.bc_tail->bc_next = bcp;
3605 					strbcalls.bc_tail = bcp;
3606 				}
3607 			}
3608 		}
3609 		if (strbcalls.bc_head == NULL)
3610 			strbcalls.bc_tail = NULL;
3611 	}
3612 
3613 	mutex_exit(&strbcall_lock);
3614 	mutex_exit(&bcall_monitor);
3615 }
3616 
3617 
3618 /*
3619  * Actually run queue's service routine.
3620  */
3621 static void
3622 runservice(queue_t *q)
3623 {
3624 	qband_t *qbp;
3625 
3626 	ASSERT(q->q_qinfo->qi_srvp);
3627 again:
3628 	entersq(q->q_syncq, SQ_SVC);
3629 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_START,
3630 	    "runservice starts:%p", q);
3631 
3632 	if (!(q->q_flag & QWCLOSE))
3633 		(*q->q_qinfo->qi_srvp)(q);
3634 
3635 	TRACE_1(TR_FAC_STREAMS_FR, TR_QRUNSERVICE_END,
3636 	    "runservice ends:(%p)", q);
3637 
3638 	leavesq(q->q_syncq, SQ_SVC);
3639 
3640 	mutex_enter(QLOCK(q));
3641 	if (q->q_flag & QENAB) {
3642 		q->q_flag &= ~QENAB;
3643 		mutex_exit(QLOCK(q));
3644 		goto again;
3645 	}
3646 	q->q_flag &= ~QINSERVICE;
3647 	q->q_flag &= ~QBACK;
3648 	for (qbp = q->q_bandp; qbp; qbp = qbp->qb_next)
3649 		qbp->qb_flag &= ~QB_BACK;
3650 	/*
3651 	 * Wakeup thread waiting for the service procedure
3652 	 * to be run (strclose and qdetach).
3653 	 */
3654 	cv_broadcast(&q->q_wait);
3655 
3656 	mutex_exit(QLOCK(q));
3657 }
3658 
3659 /*
3660  * Background processing of bufcalls.
3661  */
3662 void
3663 streams_bufcall_service(void)
3664 {
3665 	callb_cpr_t	cprinfo;
3666 
3667 	CALLB_CPR_INIT(&cprinfo, &strbcall_lock, callb_generic_cpr,
3668 	    "streams_bufcall_service");
3669 
3670 	mutex_enter(&strbcall_lock);
3671 
3672 	for (;;) {
3673 		if (strbcalls.bc_head != NULL && kmem_avail() > 0) {
3674 			mutex_exit(&strbcall_lock);
3675 			runbufcalls();
3676 			mutex_enter(&strbcall_lock);
3677 		}
3678 		if (strbcalls.bc_head != NULL) {
3679 			STRSTAT(bcwaits);
3680 			/* Wait for memory to become available */
3681 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3682 			(void) cv_reltimedwait(&memavail_cv, &strbcall_lock,
3683 			    SEC_TO_TICK(60), TR_CLOCK_TICK);
3684 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3685 		}
3686 
3687 		/* Wait for new work to arrive */
3688 		if (strbcalls.bc_head == NULL) {
3689 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3690 			cv_wait(&strbcall_cv, &strbcall_lock);
3691 			CALLB_CPR_SAFE_END(&cprinfo, &strbcall_lock);
3692 		}
3693 	}
3694 }
3695 
3696 /*
3697  * Background processing of streams background tasks which failed
3698  * taskq_dispatch.
3699  */
3700 static void
3701 streams_qbkgrnd_service(void)
3702 {
3703 	callb_cpr_t cprinfo;
3704 	queue_t *q;
3705 
3706 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3707 	    "streams_bkgrnd_service");
3708 
3709 	mutex_enter(&service_queue);
3710 
3711 	for (;;) {
3712 		/*
3713 		 * Wait for work to arrive.
3714 		 */
3715 		while ((freebs_list == NULL) && (qhead == NULL)) {
3716 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3717 			cv_wait(&services_to_run, &service_queue);
3718 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3719 		}
3720 		/*
3721 		 * Handle all pending freebs requests to free memory.
3722 		 */
3723 		while (freebs_list != NULL) {
3724 			mblk_t *mp = freebs_list;
3725 			freebs_list = mp->b_next;
3726 			mutex_exit(&service_queue);
3727 			mblk_free(mp);
3728 			mutex_enter(&service_queue);
3729 		}
3730 		/*
3731 		 * Run pending queues.
3732 		 */
3733 		while (qhead != NULL) {
3734 			DQ(q, qhead, qtail, q_link);
3735 			ASSERT(q != NULL);
3736 			mutex_exit(&service_queue);
3737 			queue_service(q);
3738 			mutex_enter(&service_queue);
3739 		}
3740 		ASSERT(qhead == NULL && qtail == NULL);
3741 	}
3742 }
3743 
3744 /*
3745  * Background processing of streams background tasks which failed
3746  * taskq_dispatch.
3747  */
3748 static void
3749 streams_sqbkgrnd_service(void)
3750 {
3751 	callb_cpr_t cprinfo;
3752 	syncq_t *sq;
3753 
3754 	CALLB_CPR_INIT(&cprinfo, &service_queue, callb_generic_cpr,
3755 	    "streams_sqbkgrnd_service");
3756 
3757 	mutex_enter(&service_queue);
3758 
3759 	for (;;) {
3760 		/*
3761 		 * Wait for work to arrive.
3762 		 */
3763 		while (sqhead == NULL) {
3764 			CALLB_CPR_SAFE_BEGIN(&cprinfo);
3765 			cv_wait(&syncqs_to_run, &service_queue);
3766 			CALLB_CPR_SAFE_END(&cprinfo, &service_queue);
3767 		}
3768 
3769 		/*
3770 		 * Run pending syncqs.
3771 		 */
3772 		while (sqhead != NULL) {
3773 			DQ(sq, sqhead, sqtail, sq_next);
3774 			ASSERT(sq != NULL);
3775 			ASSERT(sq->sq_svcflags & SQ_BGTHREAD);
3776 			mutex_exit(&service_queue);
3777 			syncq_service(sq);
3778 			mutex_enter(&service_queue);
3779 		}
3780 	}
3781 }
3782 
3783 /*
3784  * Disable the syncq and wait for background syncq processing to complete.
3785  * If the syncq is placed on the sqhead/sqtail queue, try to remove it from the
3786  * list.
3787  */
3788 void
3789 wait_sq_svc(syncq_t *sq)
3790 {
3791 	mutex_enter(SQLOCK(sq));
3792 	sq->sq_svcflags |= SQ_DISABLED;
3793 	if (sq->sq_svcflags & SQ_BGTHREAD) {
3794 		syncq_t *sq_chase;
3795 		syncq_t *sq_curr;
3796 		int removed;
3797 
3798 		ASSERT(sq->sq_servcount == 1);
3799 		mutex_enter(&service_queue);
3800 		RMQ(sq, sqhead, sqtail, sq_next, sq_chase, sq_curr, removed);
3801 		mutex_exit(&service_queue);
3802 		if (removed) {
3803 			sq->sq_svcflags &= ~SQ_BGTHREAD;
3804 			sq->sq_servcount = 0;
3805 			STRSTAT(sqremoved);
3806 			goto done;
3807 		}
3808 	}
3809 	while (sq->sq_servcount != 0) {
3810 		sq->sq_flags |= SQ_WANTWAKEUP;
3811 		cv_wait(&sq->sq_wait, SQLOCK(sq));
3812 	}
3813 done:
3814 	mutex_exit(SQLOCK(sq));
3815 }
3816 
3817 /*
3818  * Put a syncq on the list of syncq's to be serviced by the sqthread.
3819  * Add the argument to the end of the sqhead list and set the flag
3820  * indicating this syncq has been enabled.  If it has already been
3821  * enabled, don't do anything.
3822  * This routine assumes that SQLOCK is held.
3823  * NOTE that the lock order is to have the SQLOCK first,
3824  * so if the service_syncq lock is held, we need to release it
3825  * before acquiring the SQLOCK (mostly relevant for the background
3826  * thread, and this seems to be common among the STREAMS global locks).
3827  * Note that the sq_svcflags are protected by the SQLOCK.
3828  */
3829 void
3830 sqenable(syncq_t *sq)
3831 {
3832 	/*
3833 	 * This is probably not important except for where I believe it
3834 	 * is being called.  At that point, it should be held (and it
3835 	 * is a pain to release it just for this routine, so don't do
3836 	 * it).
3837 	 */
3838 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
3839 
3840 	IMPLY(sq->sq_servcount == 0, sq->sq_next == NULL);
3841 	IMPLY(sq->sq_next != NULL, sq->sq_svcflags & SQ_BGTHREAD);
3842 
3843 	/*
3844 	 * Do not put on list if background thread is scheduled or
3845 	 * syncq is disabled.
3846 	 */
3847 	if (sq->sq_svcflags & (SQ_DISABLED | SQ_BGTHREAD))
3848 		return;
3849 
3850 	/*
3851 	 * Check whether we should enable sq at all.
3852 	 * Non PERMOD syncqs may be drained by at most one thread.
3853 	 * PERMOD syncqs may be drained by several threads but we limit the
3854 	 * total amount to the lesser of
3855 	 *	Number of queues on the squeue and
3856 	 *	Number of CPUs.
3857 	 */
3858 	if (sq->sq_servcount != 0) {
3859 		if (((sq->sq_type & SQ_PERMOD) == 0) ||
3860 		    (sq->sq_servcount >= MIN(sq->sq_nqueues, ncpus_online))) {
3861 			STRSTAT(sqtoomany);
3862 			return;
3863 		}
3864 	}
3865 
3866 	sq->sq_tstamp = ddi_get_lbolt();
3867 	STRSTAT(sqenables);
3868 
3869 	/* Attempt a taskq dispatch */
3870 	sq->sq_servid = (void *)taskq_dispatch(streams_taskq,
3871 	    (task_func_t *)syncq_service, sq, TQ_NOSLEEP | TQ_NOQUEUE);
3872 	if (sq->sq_servid != NULL) {
3873 		sq->sq_servcount++;
3874 		return;
3875 	}
3876 
3877 	/*
3878 	 * This taskq dispatch failed, but a previous one may have succeeded.
3879 	 * Don't try to schedule on the background thread whilst there is
3880 	 * outstanding taskq processing.
3881 	 */
3882 	if (sq->sq_servcount != 0)
3883 		return;
3884 
3885 	/*
3886 	 * System is low on resources and can't perform a non-sleeping
3887 	 * dispatch. Schedule the syncq for a background thread and mark the
3888 	 * syncq to avoid any further taskq dispatch attempts.
3889 	 */
3890 	mutex_enter(&service_queue);
3891 	STRSTAT(taskqfails);
3892 	ENQUEUE(sq, sqhead, sqtail, sq_next);
3893 	sq->sq_svcflags |= SQ_BGTHREAD;
3894 	sq->sq_servcount = 1;
3895 	cv_signal(&syncqs_to_run);
3896 	mutex_exit(&service_queue);
3897 }
3898 
3899 /*
3900  * Note: fifo_close() depends on the mblk_t on the queue being freed
3901  * asynchronously. The asynchronous freeing of messages breaks the
3902  * recursive call chain of fifo_close() while there are I_SENDFD type of
3903  * messages referring to other file pointers on the queue. Then when
3904  * closing pipes it can avoid stack overflow in case of daisy-chained
3905  * pipes, and also avoid deadlock in case of fifonode_t pairs (which
3906  * share the same fifolock_t).
3907  */
3908 
3909 void
3910 freebs_enqueue(mblk_t *mp, dblk_t *dbp)
3911 {
3912 	esb_queue_t *eqp = &system_esbq;
3913 
3914 	ASSERT(dbp->db_mblk == mp);
3915 
3916 	/*
3917 	 * Check data sanity. The dblock should have non-empty free function.
3918 	 * It is better to panic here then later when the dblock is freed
3919 	 * asynchronously when the context is lost.
3920 	 */
3921 	if (dbp->db_frtnp->free_func == NULL) {
3922 		panic("freebs_enqueue: dblock %p has a NULL free callback",
3923 		    (void *)dbp);
3924 	}
3925 
3926 	mutex_enter(&eqp->eq_lock);
3927 	/* queue the new mblk on the esballoc queue */
3928 	if (eqp->eq_head == NULL) {
3929 		eqp->eq_head = eqp->eq_tail = mp;
3930 	} else {
3931 		eqp->eq_tail->b_next = mp;
3932 		eqp->eq_tail = mp;
3933 	}
3934 	eqp->eq_len++;
3935 
3936 	/* If we're the first thread to reach the threshold, process */
3937 	if (eqp->eq_len >= esbq_max_qlen &&
3938 	    !(eqp->eq_flags & ESBQ_PROCESSING))
3939 		esballoc_process_queue(eqp);
3940 
3941 	esballoc_set_timer(eqp, esbq_timeout);
3942 	mutex_exit(&eqp->eq_lock);
3943 }
3944 
3945 static void
3946 esballoc_process_queue(esb_queue_t *eqp)
3947 {
3948 	mblk_t	*mp;
3949 
3950 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
3951 
3952 	eqp->eq_flags |= ESBQ_PROCESSING;
3953 
3954 	do {
3955 		/*
3956 		 * Detach the message chain for processing.
3957 		 */
3958 		mp = eqp->eq_head;
3959 		eqp->eq_tail->b_next = NULL;
3960 		eqp->eq_head = eqp->eq_tail = NULL;
3961 		eqp->eq_len = 0;
3962 		mutex_exit(&eqp->eq_lock);
3963 
3964 		/*
3965 		 * Process the message chain.
3966 		 */
3967 		esballoc_enqueue_mblk(mp);
3968 		mutex_enter(&eqp->eq_lock);
3969 	} while ((eqp->eq_len >= esbq_max_qlen) && (eqp->eq_len > 0));
3970 
3971 	eqp->eq_flags &= ~ESBQ_PROCESSING;
3972 }
3973 
3974 /*
3975  * taskq callback routine to free esballoced mblk's
3976  */
3977 static void
3978 esballoc_mblk_free(mblk_t *mp)
3979 {
3980 	mblk_t	*nextmp;
3981 
3982 	for (; mp != NULL; mp = nextmp) {
3983 		nextmp = mp->b_next;
3984 		mp->b_next = NULL;
3985 		mblk_free(mp);
3986 	}
3987 }
3988 
3989 static void
3990 esballoc_enqueue_mblk(mblk_t *mp)
3991 {
3992 
3993 	if (taskq_dispatch(system_taskq, (task_func_t *)esballoc_mblk_free, mp,
3994 	    TQ_NOSLEEP) == NULL) {
3995 		mblk_t *first_mp = mp;
3996 		/*
3997 		 * System is low on resources and can't perform a non-sleeping
3998 		 * dispatch. Schedule for a background thread.
3999 		 */
4000 		mutex_enter(&service_queue);
4001 		STRSTAT(taskqfails);
4002 
4003 		while (mp->b_next != NULL)
4004 			mp = mp->b_next;
4005 
4006 		mp->b_next = freebs_list;
4007 		freebs_list = first_mp;
4008 		cv_signal(&services_to_run);
4009 		mutex_exit(&service_queue);
4010 	}
4011 }
4012 
4013 static void
4014 esballoc_timer(void *arg)
4015 {
4016 	esb_queue_t *eqp = arg;
4017 
4018 	mutex_enter(&eqp->eq_lock);
4019 	eqp->eq_flags &= ~ESBQ_TIMER;
4020 
4021 	if (!(eqp->eq_flags & ESBQ_PROCESSING) &&
4022 	    eqp->eq_len > 0)
4023 		esballoc_process_queue(eqp);
4024 
4025 	esballoc_set_timer(eqp, esbq_timeout);
4026 	mutex_exit(&eqp->eq_lock);
4027 }
4028 
4029 static void
4030 esballoc_set_timer(esb_queue_t *eqp, clock_t eq_timeout)
4031 {
4032 	ASSERT(MUTEX_HELD(&eqp->eq_lock));
4033 
4034 	if (eqp->eq_len > 0 && !(eqp->eq_flags & ESBQ_TIMER)) {
4035 		(void) timeout(esballoc_timer, eqp, eq_timeout);
4036 		eqp->eq_flags |= ESBQ_TIMER;
4037 	}
4038 }
4039 
4040 void
4041 esballoc_queue_init(void)
4042 {
4043 	system_esbq.eq_len = 0;
4044 	system_esbq.eq_head = system_esbq.eq_tail = NULL;
4045 	system_esbq.eq_flags = 0;
4046 }
4047 
4048 /*
4049  * Set the QBACK or QB_BACK flag in the given queue for
4050  * the given priority band.
4051  */
4052 void
4053 setqback(queue_t *q, unsigned char pri)
4054 {
4055 	int i;
4056 	qband_t *qbp;
4057 	qband_t **qbpp;
4058 
4059 	ASSERT(MUTEX_HELD(QLOCK(q)));
4060 	if (pri != 0) {
4061 		if (pri > q->q_nband) {
4062 			qbpp = &q->q_bandp;
4063 			while (*qbpp)
4064 				qbpp = &(*qbpp)->qb_next;
4065 			while (pri > q->q_nband) {
4066 				if ((*qbpp = allocband()) == NULL) {
4067 					cmn_err(CE_WARN,
4068 					    "setqback: can't allocate qband\n");
4069 					return;
4070 				}
4071 				(*qbpp)->qb_hiwat = q->q_hiwat;
4072 				(*qbpp)->qb_lowat = q->q_lowat;
4073 				q->q_nband++;
4074 				qbpp = &(*qbpp)->qb_next;
4075 			}
4076 		}
4077 		qbp = q->q_bandp;
4078 		i = pri;
4079 		while (--i)
4080 			qbp = qbp->qb_next;
4081 		qbp->qb_flag |= QB_BACK;
4082 	} else {
4083 		q->q_flag |= QBACK;
4084 	}
4085 }
4086 
4087 int
4088 strcopyin(void *from, void *to, size_t len, int copyflag)
4089 {
4090 	if (copyflag & U_TO_K) {
4091 		ASSERT((copyflag & K_TO_K) == 0);
4092 		if (copyin(from, to, len))
4093 			return (EFAULT);
4094 	} else {
4095 		ASSERT(copyflag & K_TO_K);
4096 		bcopy(from, to, len);
4097 	}
4098 	return (0);
4099 }
4100 
4101 int
4102 strcopyout(void *from, void *to, size_t len, int copyflag)
4103 {
4104 	if (copyflag & U_TO_K) {
4105 		if (copyout(from, to, len))
4106 			return (EFAULT);
4107 	} else {
4108 		ASSERT(copyflag & K_TO_K);
4109 		bcopy(from, to, len);
4110 	}
4111 	return (0);
4112 }
4113 
4114 /*
4115  * strsignal_nolock() posts a signal to the process(es) at the stream head.
4116  * It assumes that the stream head lock is already held, whereas strsignal()
4117  * acquires the lock first.  This routine was created because a few callers
4118  * release the stream head lock before calling only to re-acquire it after
4119  * it returns.
4120  */
4121 void
4122 strsignal_nolock(stdata_t *stp, int sig, uchar_t band)
4123 {
4124 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4125 	switch (sig) {
4126 	case SIGPOLL:
4127 		if (stp->sd_sigflags & S_MSG)
4128 			strsendsig(stp->sd_siglist, S_MSG, band, 0);
4129 		break;
4130 	default:
4131 		if (stp->sd_pgidp)
4132 			pgsignal(stp->sd_pgidp, sig);
4133 		break;
4134 	}
4135 }
4136 
4137 void
4138 strsignal(stdata_t *stp, int sig, int32_t band)
4139 {
4140 	TRACE_3(TR_FAC_STREAMS_FR, TR_SENDSIG,
4141 	    "strsignal:%p, %X, %X", stp, sig, band);
4142 
4143 	mutex_enter(&stp->sd_lock);
4144 	switch (sig) {
4145 	case SIGPOLL:
4146 		if (stp->sd_sigflags & S_MSG)
4147 			strsendsig(stp->sd_siglist, S_MSG, (uchar_t)band, 0);
4148 		break;
4149 
4150 	default:
4151 		if (stp->sd_pgidp) {
4152 			pgsignal(stp->sd_pgidp, sig);
4153 		}
4154 		break;
4155 	}
4156 	mutex_exit(&stp->sd_lock);
4157 }
4158 
4159 void
4160 strhup(stdata_t *stp)
4161 {
4162 	ASSERT(mutex_owned(&stp->sd_lock));
4163 	pollwakeup(&stp->sd_pollist, POLLHUP);
4164 	if (stp->sd_sigflags & S_HANGUP)
4165 		strsendsig(stp->sd_siglist, S_HANGUP, 0, 0);
4166 }
4167 
4168 /*
4169  * Backenable the first queue upstream from `q' with a service procedure.
4170  */
4171 void
4172 backenable(queue_t *q, uchar_t pri)
4173 {
4174 	queue_t	*nq;
4175 
4176 	/*
4177 	 * Our presence might not prevent other modules in our own
4178 	 * stream from popping/pushing since the caller of getq might not
4179 	 * have a claim on the queue (some drivers do a getq on somebody
4180 	 * else's queue - they know that the queue itself is not going away
4181 	 * but the framework has to guarantee q_next in that stream).
4182 	 */
4183 	claimstr(q);
4184 
4185 	/* Find nearest back queue with service proc */
4186 	for (nq = backq(q); nq && !nq->q_qinfo->qi_srvp; nq = backq(nq)) {
4187 		ASSERT(STRMATED(q->q_stream) || STREAM(q) == STREAM(nq));
4188 	}
4189 
4190 	if (nq) {
4191 		kthread_t *freezer;
4192 		/*
4193 		 * backenable can be called either with no locks held
4194 		 * or with the stream frozen (the latter occurs when a module
4195 		 * calls rmvq with the stream frozen). If the stream is frozen
4196 		 * by the caller the caller will hold all qlocks in the stream.
4197 		 * Note that a frozen stream doesn't freeze a mated stream,
4198 		 * so we explicitly check for that.
4199 		 */
4200 		freezer = STREAM(q)->sd_freezer;
4201 		if (freezer != curthread || STREAM(q) != STREAM(nq)) {
4202 			mutex_enter(QLOCK(nq));
4203 		}
4204 #ifdef DEBUG
4205 		else {
4206 			ASSERT(frozenstr(q));
4207 			ASSERT(MUTEX_HELD(QLOCK(q)));
4208 			ASSERT(MUTEX_HELD(QLOCK(nq)));
4209 		}
4210 #endif
4211 		setqback(nq, pri);
4212 		qenable_locked(nq);
4213 		if (freezer != curthread || STREAM(q) != STREAM(nq))
4214 			mutex_exit(QLOCK(nq));
4215 	}
4216 	releasestr(q);
4217 }
4218 
4219 /*
4220  * Return the appropriate errno when one of flags_to_check is set
4221  * in sd_flags. Uses the exported error routines if they are set.
4222  * Will return 0 if non error is set (or if the exported error routines
4223  * do not return an error).
4224  *
4225  * If there is both a read and write error to check, we prefer the read error.
4226  * Also, give preference to recorded errno's over the error functions.
4227  * The flags that are handled are:
4228  *	STPLEX		return EINVAL
4229  *	STRDERR		return sd_rerror (and clear if STRDERRNONPERSIST)
4230  *	STWRERR		return sd_werror (and clear if STWRERRNONPERSIST)
4231  *	STRHUP		return sd_werror
4232  *
4233  * If the caller indicates that the operation is a peek, a nonpersistent error
4234  * is not cleared.
4235  */
4236 int
4237 strgeterr(stdata_t *stp, int32_t flags_to_check, int ispeek)
4238 {
4239 	int32_t sd_flag = stp->sd_flag & flags_to_check;
4240 	int error = 0;
4241 
4242 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4243 	ASSERT((flags_to_check & ~(STRDERR|STWRERR|STRHUP|STPLEX)) == 0);
4244 	if (sd_flag & STPLEX)
4245 		error = EINVAL;
4246 	else if (sd_flag & STRDERR) {
4247 		error = stp->sd_rerror;
4248 		if ((stp->sd_flag & STRDERRNONPERSIST) && !ispeek) {
4249 			/*
4250 			 * Read errors are non-persistent i.e. discarded once
4251 			 * returned to a non-peeking caller,
4252 			 */
4253 			stp->sd_rerror = 0;
4254 			stp->sd_flag &= ~STRDERR;
4255 		}
4256 		if (error == 0 && stp->sd_rderrfunc != NULL) {
4257 			int clearerr = 0;
4258 
4259 			error = (*stp->sd_rderrfunc)(stp->sd_vnode, ispeek,
4260 			    &clearerr);
4261 			if (clearerr) {
4262 				stp->sd_flag &= ~STRDERR;
4263 				stp->sd_rderrfunc = NULL;
4264 			}
4265 		}
4266 	} else if (sd_flag & STWRERR) {
4267 		error = stp->sd_werror;
4268 		if ((stp->sd_flag & STWRERRNONPERSIST) && !ispeek) {
4269 			/*
4270 			 * Write errors are non-persistent i.e. discarded once
4271 			 * returned to a non-peeking caller,
4272 			 */
4273 			stp->sd_werror = 0;
4274 			stp->sd_flag &= ~STWRERR;
4275 		}
4276 		if (error == 0 && stp->sd_wrerrfunc != NULL) {
4277 			int clearerr = 0;
4278 
4279 			error = (*stp->sd_wrerrfunc)(stp->sd_vnode, ispeek,
4280 			    &clearerr);
4281 			if (clearerr) {
4282 				stp->sd_flag &= ~STWRERR;
4283 				stp->sd_wrerrfunc = NULL;
4284 			}
4285 		}
4286 	} else if (sd_flag & STRHUP) {
4287 		/* sd_werror set when STRHUP */
4288 		error = stp->sd_werror;
4289 	}
4290 	return (error);
4291 }
4292 
4293 
4294 /*
4295  * Single-thread open/close/push/pop
4296  * for twisted streams also
4297  */
4298 int
4299 strstartplumb(stdata_t *stp, int flag, int cmd)
4300 {
4301 	int waited = 1;
4302 	int error = 0;
4303 
4304 	if (STRMATED(stp)) {
4305 		struct stdata *stmatep = stp->sd_mate;
4306 
4307 		STRLOCKMATES(stp);
4308 		while (waited) {
4309 			waited = 0;
4310 			while (stmatep->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4311 				if ((cmd == I_POP) &&
4312 				    (flag & (FNDELAY|FNONBLOCK))) {
4313 					STRUNLOCKMATES(stp);
4314 					return (EAGAIN);
4315 				}
4316 				waited = 1;
4317 				mutex_exit(&stp->sd_lock);
4318 				if (!cv_wait_sig(&stmatep->sd_monitor,
4319 				    &stmatep->sd_lock)) {
4320 					mutex_exit(&stmatep->sd_lock);
4321 					return (EINTR);
4322 				}
4323 				mutex_exit(&stmatep->sd_lock);
4324 				STRLOCKMATES(stp);
4325 			}
4326 			while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4327 				if ((cmd == I_POP) &&
4328 				    (flag & (FNDELAY|FNONBLOCK))) {
4329 					STRUNLOCKMATES(stp);
4330 					return (EAGAIN);
4331 				}
4332 				waited = 1;
4333 				mutex_exit(&stmatep->sd_lock);
4334 				if (!cv_wait_sig(&stp->sd_monitor,
4335 				    &stp->sd_lock)) {
4336 					mutex_exit(&stp->sd_lock);
4337 					return (EINTR);
4338 				}
4339 				mutex_exit(&stp->sd_lock);
4340 				STRLOCKMATES(stp);
4341 			}
4342 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4343 				error = strgeterr(stp,
4344 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4345 				if (error != 0) {
4346 					STRUNLOCKMATES(stp);
4347 					return (error);
4348 				}
4349 			}
4350 		}
4351 		stp->sd_flag |= STRPLUMB;
4352 		STRUNLOCKMATES(stp);
4353 	} else {
4354 		mutex_enter(&stp->sd_lock);
4355 		while (stp->sd_flag & (STWOPEN|STRCLOSE|STRPLUMB)) {
4356 			if (((cmd == I_POP) || (cmd == _I_REMOVE)) &&
4357 			    (flag & (FNDELAY|FNONBLOCK))) {
4358 				mutex_exit(&stp->sd_lock);
4359 				return (EAGAIN);
4360 			}
4361 			if (!cv_wait_sig(&stp->sd_monitor, &stp->sd_lock)) {
4362 				mutex_exit(&stp->sd_lock);
4363 				return (EINTR);
4364 			}
4365 			if (stp->sd_flag & (STRDERR|STWRERR|STRHUP|STPLEX)) {
4366 				error = strgeterr(stp,
4367 				    STRDERR|STWRERR|STRHUP|STPLEX, 0);
4368 				if (error != 0) {
4369 					mutex_exit(&stp->sd_lock);
4370 					return (error);
4371 				}
4372 			}
4373 		}
4374 		stp->sd_flag |= STRPLUMB;
4375 		mutex_exit(&stp->sd_lock);
4376 	}
4377 	return (0);
4378 }
4379 
4380 /*
4381  * Complete the plumbing operation associated with stream `stp'.
4382  */
4383 void
4384 strendplumb(stdata_t *stp)
4385 {
4386 	ASSERT(MUTEX_HELD(&stp->sd_lock));
4387 	ASSERT(stp->sd_flag & STRPLUMB);
4388 	stp->sd_flag &= ~STRPLUMB;
4389 	cv_broadcast(&stp->sd_monitor);
4390 }
4391 
4392 /*
4393  * This describes how the STREAMS framework handles synchronization
4394  * during open/push and close/pop.
4395  * The key interfaces for open and close are qprocson and qprocsoff,
4396  * respectively. While the close case in general is harder both open
4397  * have close have significant similarities.
4398  *
4399  * During close the STREAMS framework has to both ensure that there
4400  * are no stale references to the queue pair (and syncq) that
4401  * are being closed and also provide the guarantees that are documented
4402  * in qprocsoff(9F).
4403  * If there are stale references to the queue that is closing it can
4404  * result in kernel memory corruption or kernel panics.
4405  *
4406  * Note that is it up to the module/driver to ensure that it itself
4407  * does not have any stale references to the closing queues once its close
4408  * routine returns. This includes:
4409  *  - Cancelling any timeout/bufcall/qtimeout/qbufcall callback routines
4410  *    associated with the queues. For timeout and bufcall callbacks the
4411  *    module/driver also has to ensure (or wait for) any callbacks that
4412  *    are in progress.
4413  *  - If the module/driver is using esballoc it has to ensure that any
4414  *    esballoc free functions do not refer to a queue that has closed.
4415  *    (Note that in general the close routine can not wait for the esballoc'ed
4416  *    messages to be freed since that can cause a deadlock.)
4417  *  - Cancelling any interrupts that refer to the closing queues and
4418  *    also ensuring that there are no interrupts in progress that will
4419  *    refer to the closing queues once the close routine returns.
4420  *  - For multiplexors removing any driver global state that refers to
4421  *    the closing queue and also ensuring that there are no threads in
4422  *    the multiplexor that has picked up a queue pointer but not yet
4423  *    finished using it.
4424  *
4425  * In addition, a driver/module can only reference the q_next pointer
4426  * in its open, close, put, or service procedures or in a
4427  * qtimeout/qbufcall callback procedure executing "on" the correct
4428  * stream. Thus it can not reference the q_next pointer in an interrupt
4429  * routine or a timeout, bufcall or esballoc callback routine. Likewise
4430  * it can not reference q_next of a different queue e.g. in a mux that
4431  * passes messages from one queues put/service procedure to another queue.
4432  * In all the cases when the driver/module can not access the q_next
4433  * field it must use the *next* versions e.g. canputnext instead of
4434  * canput(q->q_next) and putnextctl instead of putctl(q->q_next, ...).
4435  *
4436  *
4437  * Assuming that the driver/module conforms to the above constraints
4438  * the STREAMS framework has to avoid stale references to q_next for all
4439  * the framework internal cases which include (but are not limited to):
4440  *  - Threads in canput/canputnext/backenable and elsewhere that are
4441  *    walking q_next.
4442  *  - Messages on a syncq that have a reference to the queue through b_queue.
4443  *  - Messages on an outer perimeter (syncq) that have a reference to the
4444  *    queue through b_queue.
4445  *  - Threads that use q_nfsrv (e.g. canput) to find a queue.
4446  *    Note that only canput and bcanput use q_nfsrv without any locking.
4447  *
4448  * The STREAMS framework providing the qprocsoff(9F) guarantees means that
4449  * after qprocsoff returns, the framework has to ensure that no threads can
4450  * enter the put or service routines for the closing read or write-side queue.
4451  * In addition to preventing "direct" entry into the put procedures
4452  * the framework also has to prevent messages being drained from
4453  * the syncq or the outer perimeter.
4454  * XXX Note that currently qdetach does relies on D_MTOCEXCL as the only
4455  * mechanism to prevent qwriter(PERIM_OUTER) from running after
4456  * qprocsoff has returned.
4457  * Note that if a module/driver uses put(9F) on one of its own queues
4458  * it is up to the module/driver to ensure that the put() doesn't
4459  * get called when the queue is closing.
4460  *
4461  *
4462  * The framework aspects of the above "contract" is implemented by
4463  * qprocsoff, removeq, and strlock:
4464  *  - qprocsoff (disable_svc) sets QWCLOSE to prevent runservice from
4465  *    entering the service procedures.
4466  *  - strlock acquires the sd_lock and sd_reflock to prevent putnext,
4467  *    canputnext, backenable etc from dereferencing the q_next that will
4468  *    soon change.
4469  *  - strlock waits for sd_refcnt to be zero to wait for e.g. any canputnext
4470  *    or other q_next walker that uses claimstr/releasestr to finish.
4471  *  - optionally for every syncq in the stream strlock acquires all the
4472  *    sq_lock's and waits for all sq_counts to drop to a value that indicates
4473  *    that no thread executes in the put or service procedures and that no
4474  *    thread is draining into the module/driver. This ensures that no
4475  *    open, close, put, service, or qtimeout/qbufcall callback procedure is
4476  *    currently executing hence no such thread can end up with the old stale
4477  *    q_next value and no canput/backenable can have the old stale
4478  *    q_nfsrv/q_next.
4479  *  - qdetach (wait_svc) makes sure that any scheduled or running threads
4480  *    have either finished or observed the QWCLOSE flag and gone away.
4481  */
4482 
4483 
4484 /*
4485  * Get all the locks necessary to change q_next.
4486  *
4487  * Wait for sd_refcnt to reach 0 and, if sqlist is present, wait for the
4488  * sq_count of each syncq in the list to drop to sq_rmqcount, indicating that
4489  * the only threads inside the syncq are threads currently calling removeq().
4490  * Since threads calling removeq() are in the process of removing their queues
4491  * from the stream, we do not need to worry about them accessing a stale q_next
4492  * pointer and thus we do not need to wait for them to exit (in fact, waiting
4493  * for them can cause deadlock).
4494  *
4495  * This routine is subject to starvation since it does not set any flag to
4496  * prevent threads from entering a module in the stream (i.e. sq_count can
4497  * increase on some syncq while it is waiting on some other syncq).
4498  *
4499  * Assumes that only one thread attempts to call strlock for a given
4500  * stream. If this is not the case the two threads would deadlock.
4501  * This assumption is guaranteed since strlock is only called by insertq
4502  * and removeq and streams plumbing changes are single-threaded for
4503  * a given stream using the STWOPEN, STRCLOSE, and STRPLUMB flags.
4504  *
4505  * For pipes, it is not difficult to atomically designate a pair of streams
4506  * to be mated. Once mated atomically by the framework the twisted pair remain
4507  * configured that way until dismantled atomically by the framework.
4508  * When plumbing takes place on a twisted stream it is necessary to ensure that
4509  * this operation is done exclusively on the twisted stream since two such
4510  * operations, each initiated on different ends of the pipe will deadlock
4511  * waiting for each other to complete.
4512  *
4513  * On entry, no locks should be held.
4514  * The locks acquired and held by strlock depends on a few factors.
4515  * - If sqlist is non-NULL all the syncq locks in the sqlist will be acquired
4516  *   and held on exit and all sq_count are at an acceptable level.
4517  * - In all cases, sd_lock and sd_reflock are acquired and held on exit with
4518  *   sd_refcnt being zero.
4519  */
4520 
4521 static void
4522 strlock(struct stdata *stp, sqlist_t *sqlist)
4523 {
4524 	syncql_t *sql, *sql2;
4525 retry:
4526 	/*
4527 	 * Wait for any claimstr to go away.
4528 	 */
4529 	if (STRMATED(stp)) {
4530 		struct stdata *stp1, *stp2;
4531 
4532 		STRLOCKMATES(stp);
4533 		/*
4534 		 * Note that the selection of locking order is not
4535 		 * important, just that they are always acquired in
4536 		 * the same order.  To assure this, we choose this
4537 		 * order based on the value of the pointer, and since
4538 		 * the pointer will not change for the life of this
4539 		 * pair, we will always grab the locks in the same
4540 		 * order (and hence, prevent deadlocks).
4541 		 */
4542 		if (&(stp->sd_lock) > &((stp->sd_mate)->sd_lock)) {
4543 			stp1 = stp;
4544 			stp2 = stp->sd_mate;
4545 		} else {
4546 			stp2 = stp;
4547 			stp1 = stp->sd_mate;
4548 		}
4549 		mutex_enter(&stp1->sd_reflock);
4550 		if (stp1->sd_refcnt > 0) {
4551 			STRUNLOCKMATES(stp);
4552 			cv_wait(&stp1->sd_refmonitor, &stp1->sd_reflock);
4553 			mutex_exit(&stp1->sd_reflock);
4554 			goto retry;
4555 		}
4556 		mutex_enter(&stp2->sd_reflock);
4557 		if (stp2->sd_refcnt > 0) {
4558 			STRUNLOCKMATES(stp);
4559 			mutex_exit(&stp1->sd_reflock);
4560 			cv_wait(&stp2->sd_refmonitor, &stp2->sd_reflock);
4561 			mutex_exit(&stp2->sd_reflock);
4562 			goto retry;
4563 		}
4564 		STREAM_PUTLOCKS_ENTER(stp1);
4565 		STREAM_PUTLOCKS_ENTER(stp2);
4566 	} else {
4567 		mutex_enter(&stp->sd_lock);
4568 		mutex_enter(&stp->sd_reflock);
4569 		while (stp->sd_refcnt > 0) {
4570 			mutex_exit(&stp->sd_lock);
4571 			cv_wait(&stp->sd_refmonitor, &stp->sd_reflock);
4572 			if (mutex_tryenter(&stp->sd_lock) == 0) {
4573 				mutex_exit(&stp->sd_reflock);
4574 				mutex_enter(&stp->sd_lock);
4575 				mutex_enter(&stp->sd_reflock);
4576 			}
4577 		}
4578 		STREAM_PUTLOCKS_ENTER(stp);
4579 	}
4580 
4581 	if (sqlist == NULL)
4582 		return;
4583 
4584 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4585 		syncq_t *sq = sql->sql_sq;
4586 		uint16_t count;
4587 
4588 		mutex_enter(SQLOCK(sq));
4589 		count = sq->sq_count;
4590 		ASSERT(sq->sq_rmqcount <= count);
4591 		SQ_PUTLOCKS_ENTER(sq);
4592 		SUM_SQ_PUTCOUNTS(sq, count);
4593 		if (count == sq->sq_rmqcount)
4594 			continue;
4595 
4596 		/* Failed - drop all locks that we have acquired so far */
4597 		if (STRMATED(stp)) {
4598 			STREAM_PUTLOCKS_EXIT(stp);
4599 			STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4600 			STRUNLOCKMATES(stp);
4601 			mutex_exit(&stp->sd_reflock);
4602 			mutex_exit(&stp->sd_mate->sd_reflock);
4603 		} else {
4604 			STREAM_PUTLOCKS_EXIT(stp);
4605 			mutex_exit(&stp->sd_lock);
4606 			mutex_exit(&stp->sd_reflock);
4607 		}
4608 		for (sql2 = sqlist->sqlist_head; sql2 != sql;
4609 		    sql2 = sql2->sql_next) {
4610 			SQ_PUTLOCKS_EXIT(sql2->sql_sq);
4611 			mutex_exit(SQLOCK(sql2->sql_sq));
4612 		}
4613 
4614 		/*
4615 		 * The wait loop below may starve when there are many threads
4616 		 * claiming the syncq. This is especially a problem with permod
4617 		 * syncqs (IP). To lessen the impact of the problem we increment
4618 		 * sq_needexcl and clear fastbits so that putnexts will slow
4619 		 * down and call sqenable instead of draining right away.
4620 		 */
4621 		sq->sq_needexcl++;
4622 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4623 		while (count > sq->sq_rmqcount) {
4624 			sq->sq_flags |= SQ_WANTWAKEUP;
4625 			SQ_PUTLOCKS_EXIT(sq);
4626 			cv_wait(&sq->sq_wait, SQLOCK(sq));
4627 			count = sq->sq_count;
4628 			SQ_PUTLOCKS_ENTER(sq);
4629 			SUM_SQ_PUTCOUNTS(sq, count);
4630 		}
4631 		sq->sq_needexcl--;
4632 		if (sq->sq_needexcl == 0)
4633 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
4634 		SQ_PUTLOCKS_EXIT(sq);
4635 		ASSERT(count == sq->sq_rmqcount);
4636 		mutex_exit(SQLOCK(sq));
4637 		goto retry;
4638 	}
4639 }
4640 
4641 /*
4642  * Drop all the locks that strlock acquired.
4643  */
4644 static void
4645 strunlock(struct stdata *stp, sqlist_t *sqlist)
4646 {
4647 	syncql_t *sql;
4648 
4649 	if (STRMATED(stp)) {
4650 		STREAM_PUTLOCKS_EXIT(stp);
4651 		STREAM_PUTLOCKS_EXIT(stp->sd_mate);
4652 		STRUNLOCKMATES(stp);
4653 		mutex_exit(&stp->sd_reflock);
4654 		mutex_exit(&stp->sd_mate->sd_reflock);
4655 	} else {
4656 		STREAM_PUTLOCKS_EXIT(stp);
4657 		mutex_exit(&stp->sd_lock);
4658 		mutex_exit(&stp->sd_reflock);
4659 	}
4660 
4661 	if (sqlist == NULL)
4662 		return;
4663 
4664 	for (sql = sqlist->sqlist_head; sql; sql = sql->sql_next) {
4665 		SQ_PUTLOCKS_EXIT(sql->sql_sq);
4666 		mutex_exit(SQLOCK(sql->sql_sq));
4667 	}
4668 }
4669 
4670 /*
4671  * When the module has service procedure, we need check if the next
4672  * module which has service procedure is in flow control to trigger
4673  * the backenable.
4674  */
4675 static void
4676 backenable_insertedq(queue_t *q)
4677 {
4678 	qband_t	*qbp;
4679 
4680 	claimstr(q);
4681 	if (q->q_qinfo->qi_srvp != NULL && q->q_next != NULL) {
4682 		if (q->q_next->q_nfsrv->q_flag & QWANTW)
4683 			backenable(q, 0);
4684 
4685 		qbp = q->q_next->q_nfsrv->q_bandp;
4686 		for (; qbp != NULL; qbp = qbp->qb_next)
4687 			if ((qbp->qb_flag & QB_WANTW) && qbp->qb_first != NULL)
4688 				backenable(q, qbp->qb_first->b_band);
4689 	}
4690 	releasestr(q);
4691 }
4692 
4693 /*
4694  * Given two read queues, insert a new single one after another.
4695  *
4696  * This routine acquires all the necessary locks in order to change
4697  * q_next and related pointer using strlock().
4698  * It depends on the stream head ensuring that there are no concurrent
4699  * insertq or removeq on the same stream. The stream head ensures this
4700  * using the flags STWOPEN, STRCLOSE, and STRPLUMB.
4701  *
4702  * Note that no syncq locks are held during the q_next change. This is
4703  * applied to all streams since, unlike removeq, there is no problem of stale
4704  * pointers when adding a module to the stream. Thus drivers/modules that do a
4705  * canput(rq->q_next) would never get a closed/freed queue pointer even if we
4706  * applied this optimization to all streams.
4707  */
4708 void
4709 insertq(struct stdata *stp, queue_t *new)
4710 {
4711 	queue_t	*after;
4712 	queue_t *wafter;
4713 	queue_t *wnew = _WR(new);
4714 	boolean_t have_fifo = B_FALSE;
4715 
4716 	if (new->q_flag & _QINSERTING) {
4717 		ASSERT(stp->sd_vnode->v_type != VFIFO);
4718 		after = new->q_next;
4719 		wafter = _WR(new->q_next);
4720 	} else {
4721 		after = _RD(stp->sd_wrq);
4722 		wafter = stp->sd_wrq;
4723 	}
4724 
4725 	TRACE_2(TR_FAC_STREAMS_FR, TR_INSERTQ,
4726 	    "insertq:%p, %p", after, new);
4727 	ASSERT(after->q_flag & QREADR);
4728 	ASSERT(new->q_flag & QREADR);
4729 
4730 	strlock(stp, NULL);
4731 
4732 	/* Do we have a FIFO? */
4733 	if (wafter->q_next == after) {
4734 		have_fifo = B_TRUE;
4735 		wnew->q_next = new;
4736 	} else {
4737 		wnew->q_next = wafter->q_next;
4738 	}
4739 	new->q_next = after;
4740 
4741 	set_nfsrv_ptr(new, wnew, after, wafter);
4742 	/*
4743 	 * set_nfsrv_ptr() needs to know if this is an insertion or not,
4744 	 * so only reset this flag after calling it.
4745 	 */
4746 	new->q_flag &= ~_QINSERTING;
4747 
4748 	if (have_fifo) {
4749 		wafter->q_next = wnew;
4750 	} else {
4751 		if (wafter->q_next)
4752 			_OTHERQ(wafter->q_next)->q_next = new;
4753 		wafter->q_next = wnew;
4754 	}
4755 
4756 	set_qend(new);
4757 	/* The QEND flag might have to be updated for the upstream guy */
4758 	set_qend(after);
4759 
4760 	ASSERT(_SAMESTR(new) == O_SAMESTR(new));
4761 	ASSERT(_SAMESTR(wnew) == O_SAMESTR(wnew));
4762 	ASSERT(_SAMESTR(after) == O_SAMESTR(after));
4763 	ASSERT(_SAMESTR(wafter) == O_SAMESTR(wafter));
4764 	strsetuio(stp);
4765 
4766 	/*
4767 	 * If this was a module insertion, bump the push count.
4768 	 */
4769 	if (!(new->q_flag & QISDRV))
4770 		stp->sd_pushcnt++;
4771 
4772 	strunlock(stp, NULL);
4773 
4774 	/* check if the write Q needs backenable */
4775 	backenable_insertedq(wnew);
4776 
4777 	/* check if the read Q needs backenable */
4778 	backenable_insertedq(new);
4779 }
4780 
4781 /*
4782  * Given a read queue, unlink it from any neighbors.
4783  *
4784  * This routine acquires all the necessary locks in order to
4785  * change q_next and related pointers and also guard against
4786  * stale references (e.g. through q_next) to the queue that
4787  * is being removed. It also plays part of the role in ensuring
4788  * that the module's/driver's put procedure doesn't get called
4789  * after qprocsoff returns.
4790  *
4791  * Removeq depends on the stream head ensuring that there are
4792  * no concurrent insertq or removeq on the same stream. The
4793  * stream head ensures this using the flags STWOPEN, STRCLOSE and
4794  * STRPLUMB.
4795  *
4796  * The set of locks needed to remove the queue is different in
4797  * different cases:
4798  *
4799  * Acquire sd_lock, sd_reflock, and all the syncq locks in the stream after
4800  * waiting for the syncq reference count to drop to 0 indicating that no
4801  * non-close threads are present anywhere in the stream. This ensures that any
4802  * module/driver can reference q_next in its open, close, put, or service
4803  * procedures.
4804  *
4805  * The sq_rmqcount counter tracks the number of threads inside removeq().
4806  * strlock() ensures that there is either no threads executing inside perimeter
4807  * or there is only a thread calling qprocsoff().
4808  *
4809  * strlock() compares the value of sq_count with the number of threads inside
4810  * removeq() and waits until sq_count is equal to sq_rmqcount. We need to wakeup
4811  * any threads waiting in strlock() when the sq_rmqcount increases.
4812  */
4813 
4814 void
4815 removeq(queue_t *qp)
4816 {
4817 	queue_t *wqp = _WR(qp);
4818 	struct stdata *stp = STREAM(qp);
4819 	sqlist_t *sqlist = NULL;
4820 	boolean_t isdriver;
4821 	int moved;
4822 	syncq_t *sq = qp->q_syncq;
4823 	syncq_t *wsq = wqp->q_syncq;
4824 
4825 	ASSERT(stp);
4826 
4827 	TRACE_2(TR_FAC_STREAMS_FR, TR_REMOVEQ,
4828 	    "removeq:%p %p", qp, wqp);
4829 	ASSERT(qp->q_flag&QREADR);
4830 
4831 	/*
4832 	 * For queues using Synchronous streams, we must wait for all threads in
4833 	 * rwnext() to drain out before proceeding.
4834 	 */
4835 	if (qp->q_flag & QSYNCSTR) {
4836 		/* First, we need wakeup any threads blocked in rwnext() */
4837 		mutex_enter(SQLOCK(sq));
4838 		if (sq->sq_flags & SQ_WANTWAKEUP) {
4839 			sq->sq_flags &= ~SQ_WANTWAKEUP;
4840 			cv_broadcast(&sq->sq_wait);
4841 		}
4842 		mutex_exit(SQLOCK(sq));
4843 
4844 		if (wsq != sq) {
4845 			mutex_enter(SQLOCK(wsq));
4846 			if (wsq->sq_flags & SQ_WANTWAKEUP) {
4847 				wsq->sq_flags &= ~SQ_WANTWAKEUP;
4848 				cv_broadcast(&wsq->sq_wait);
4849 			}
4850 			mutex_exit(SQLOCK(wsq));
4851 		}
4852 
4853 		mutex_enter(QLOCK(qp));
4854 		while (qp->q_rwcnt > 0) {
4855 			qp->q_flag |= QWANTRMQSYNC;
4856 			cv_wait(&qp->q_wait, QLOCK(qp));
4857 		}
4858 		mutex_exit(QLOCK(qp));
4859 
4860 		mutex_enter(QLOCK(wqp));
4861 		while (wqp->q_rwcnt > 0) {
4862 			wqp->q_flag |= QWANTRMQSYNC;
4863 			cv_wait(&wqp->q_wait, QLOCK(wqp));
4864 		}
4865 		mutex_exit(QLOCK(wqp));
4866 	}
4867 
4868 	mutex_enter(SQLOCK(sq));
4869 	sq->sq_rmqcount++;
4870 	if (sq->sq_flags & SQ_WANTWAKEUP) {
4871 		sq->sq_flags &= ~SQ_WANTWAKEUP;
4872 		cv_broadcast(&sq->sq_wait);
4873 	}
4874 	mutex_exit(SQLOCK(sq));
4875 
4876 	isdriver = (qp->q_flag & QISDRV);
4877 
4878 	sqlist = sqlist_build(qp, stp, STRMATED(stp));
4879 	strlock(stp, sqlist);
4880 
4881 	reset_nfsrv_ptr(qp, wqp);
4882 
4883 	ASSERT(wqp->q_next == NULL || backq(qp)->q_next == qp);
4884 	ASSERT(qp->q_next == NULL || backq(wqp)->q_next == wqp);
4885 	/* Do we have a FIFO? */
4886 	if (wqp->q_next == qp) {
4887 		stp->sd_wrq->q_next = _RD(stp->sd_wrq);
4888 	} else {
4889 		if (wqp->q_next)
4890 			backq(qp)->q_next = qp->q_next;
4891 		if (qp->q_next)
4892 			backq(wqp)->q_next = wqp->q_next;
4893 	}
4894 
4895 	/* The QEND flag might have to be updated for the upstream guy */
4896 	if (qp->q_next)
4897 		set_qend(qp->q_next);
4898 
4899 	ASSERT(_SAMESTR(stp->sd_wrq) == O_SAMESTR(stp->sd_wrq));
4900 	ASSERT(_SAMESTR(_RD(stp->sd_wrq)) == O_SAMESTR(_RD(stp->sd_wrq)));
4901 
4902 	/*
4903 	 * Move any messages destined for the put procedures to the next
4904 	 * syncq in line. Otherwise free them.
4905 	 */
4906 	moved = 0;
4907 	/*
4908 	 * Quick check to see whether there are any messages or events.
4909 	 */
4910 	if (qp->q_syncqmsgs != 0 || (qp->q_syncq->sq_flags & SQ_EVENTS))
4911 		moved += propagate_syncq(qp);
4912 	if (wqp->q_syncqmsgs != 0 ||
4913 	    (wqp->q_syncq->sq_flags & SQ_EVENTS))
4914 		moved += propagate_syncq(wqp);
4915 
4916 	strsetuio(stp);
4917 
4918 	/*
4919 	 * If this was a module removal, decrement the push count.
4920 	 */
4921 	if (!isdriver)
4922 		stp->sd_pushcnt--;
4923 
4924 	strunlock(stp, sqlist);
4925 	sqlist_free(sqlist);
4926 
4927 	/*
4928 	 * Make sure any messages that were propagated are drained.
4929 	 * Also clear any QFULL bit caused by messages that were propagated.
4930 	 */
4931 
4932 	if (qp->q_next != NULL) {
4933 		clr_qfull(qp);
4934 		/*
4935 		 * For the driver calling qprocsoff, propagate_syncq
4936 		 * frees all the messages instead of putting it in
4937 		 * the stream head
4938 		 */
4939 		if (!isdriver && (moved > 0))
4940 			emptysq(qp->q_next->q_syncq);
4941 	}
4942 	if (wqp->q_next != NULL) {
4943 		clr_qfull(wqp);
4944 		/*
4945 		 * We come here for any pop of a module except for the
4946 		 * case of driver being removed. We don't call emptysq
4947 		 * if we did not move any messages. This will avoid holding
4948 		 * PERMOD syncq locks in emptysq
4949 		 */
4950 		if (moved > 0)
4951 			emptysq(wqp->q_next->q_syncq);
4952 	}
4953 
4954 	mutex_enter(SQLOCK(sq));
4955 	sq->sq_rmqcount--;
4956 	mutex_exit(SQLOCK(sq));
4957 }
4958 
4959 /*
4960  * Prevent further entry by setting a flag (like SQ_FROZEN, SQ_BLOCKED or
4961  * SQ_WRITER) on a syncq.
4962  * If maxcnt is not -1 it assumes that caller has "maxcnt" claim(s) on the
4963  * sync queue and waits until sq_count reaches maxcnt.
4964  *
4965  * If maxcnt is -1 there's no need to grab sq_putlocks since the caller
4966  * does not care about putnext threads that are in the middle of calling put
4967  * entry points.
4968  *
4969  * This routine is used for both inner and outer syncqs.
4970  */
4971 static void
4972 blocksq(syncq_t *sq, ushort_t flag, int maxcnt)
4973 {
4974 	uint16_t count = 0;
4975 
4976 	mutex_enter(SQLOCK(sq));
4977 	/*
4978 	 * Wait for SQ_FROZEN/SQ_BLOCKED to be reset.
4979 	 * SQ_FROZEN will be set if there is a frozen stream that has a
4980 	 * queue which also refers to this "shared" syncq.
4981 	 * SQ_BLOCKED will be set if there is "off" queue which also
4982 	 * refers to this "shared" syncq.
4983 	 */
4984 	if (maxcnt != -1) {
4985 		count = sq->sq_count;
4986 		SQ_PUTLOCKS_ENTER(sq);
4987 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
4988 		SUM_SQ_PUTCOUNTS(sq, count);
4989 	}
4990 	sq->sq_needexcl++;
4991 	ASSERT(sq->sq_needexcl != 0);	/* wraparound */
4992 
4993 	while ((sq->sq_flags & flag) ||
4994 	    (maxcnt != -1 && count > (unsigned)maxcnt)) {
4995 		sq->sq_flags |= SQ_WANTWAKEUP;
4996 		if (maxcnt != -1) {
4997 			SQ_PUTLOCKS_EXIT(sq);
4998 		}
4999 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5000 		if (maxcnt != -1) {
5001 			count = sq->sq_count;
5002 			SQ_PUTLOCKS_ENTER(sq);
5003 			SUM_SQ_PUTCOUNTS(sq, count);
5004 		}
5005 	}
5006 	sq->sq_needexcl--;
5007 	sq->sq_flags |= flag;
5008 	ASSERT(maxcnt == -1 || count == maxcnt);
5009 	if (maxcnt != -1) {
5010 		if (sq->sq_needexcl == 0) {
5011 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5012 		}
5013 		SQ_PUTLOCKS_EXIT(sq);
5014 	} else if (sq->sq_needexcl == 0) {
5015 		SQ_PUTCOUNT_SETFAST(sq);
5016 	}
5017 
5018 	mutex_exit(SQLOCK(sq));
5019 }
5020 
5021 /*
5022  * Reset a flag that was set with blocksq.
5023  *
5024  * Can not use this routine to reset SQ_WRITER.
5025  *
5026  * If "isouter" is set then the syncq is assumed to be an outer perimeter
5027  * and drain_syncq is not called. Instead we rely on the qwriter_outer thread
5028  * to handle the queued qwriter operations.
5029  *
5030  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5031  * sq_putlocks are used.
5032  */
5033 static void
5034 unblocksq(syncq_t *sq, uint16_t resetflag, int isouter)
5035 {
5036 	uint16_t flags;
5037 
5038 	mutex_enter(SQLOCK(sq));
5039 	ASSERT(resetflag != SQ_WRITER);
5040 	ASSERT(sq->sq_flags & resetflag);
5041 	flags = sq->sq_flags & ~resetflag;
5042 	sq->sq_flags = flags;
5043 	if (flags & (SQ_QUEUED | SQ_WANTWAKEUP)) {
5044 		if (flags & SQ_WANTWAKEUP) {
5045 			flags &= ~SQ_WANTWAKEUP;
5046 			cv_broadcast(&sq->sq_wait);
5047 		}
5048 		sq->sq_flags = flags;
5049 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5050 			if (!isouter) {
5051 				/* drain_syncq drops SQLOCK */
5052 				drain_syncq(sq);
5053 				return;
5054 			}
5055 		}
5056 	}
5057 	mutex_exit(SQLOCK(sq));
5058 }
5059 
5060 /*
5061  * Reset a flag that was set with blocksq.
5062  * Does not drain the syncq. Use emptysq() for that.
5063  * Returns 1 if SQ_QUEUED is set. Otherwise 0.
5064  *
5065  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5066  * sq_putlocks are used.
5067  */
5068 static int
5069 dropsq(syncq_t *sq, uint16_t resetflag)
5070 {
5071 	uint16_t flags;
5072 
5073 	mutex_enter(SQLOCK(sq));
5074 	ASSERT(sq->sq_flags & resetflag);
5075 	flags = sq->sq_flags & ~resetflag;
5076 	if (flags & SQ_WANTWAKEUP) {
5077 		flags &= ~SQ_WANTWAKEUP;
5078 		cv_broadcast(&sq->sq_wait);
5079 	}
5080 	sq->sq_flags = flags;
5081 	mutex_exit(SQLOCK(sq));
5082 	if (flags & SQ_QUEUED)
5083 		return (1);
5084 	return (0);
5085 }
5086 
5087 /*
5088  * Empty all the messages on a syncq.
5089  *
5090  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5091  * sq_putlocks are used.
5092  */
5093 static void
5094 emptysq(syncq_t *sq)
5095 {
5096 	uint16_t flags;
5097 
5098 	mutex_enter(SQLOCK(sq));
5099 	flags = sq->sq_flags;
5100 	if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5101 		/*
5102 		 * To prevent potential recursive invocation of drain_syncq we
5103 		 * do not call drain_syncq if count is non-zero.
5104 		 */
5105 		if (sq->sq_count == 0) {
5106 			/* drain_syncq() drops SQLOCK */
5107 			drain_syncq(sq);
5108 			return;
5109 		} else
5110 			sqenable(sq);
5111 	}
5112 	mutex_exit(SQLOCK(sq));
5113 }
5114 
5115 /*
5116  * Ordered insert while removing duplicates.
5117  */
5118 static void
5119 sqlist_insert(sqlist_t *sqlist, syncq_t *sqp)
5120 {
5121 	syncql_t *sqlp, **prev_sqlpp, *new_sqlp;
5122 
5123 	prev_sqlpp = &sqlist->sqlist_head;
5124 	while ((sqlp = *prev_sqlpp) != NULL) {
5125 		if (sqlp->sql_sq >= sqp) {
5126 			if (sqlp->sql_sq == sqp)	/* duplicate */
5127 				return;
5128 			break;
5129 		}
5130 		prev_sqlpp = &sqlp->sql_next;
5131 	}
5132 	new_sqlp = &sqlist->sqlist_array[sqlist->sqlist_index++];
5133 	ASSERT((char *)new_sqlp < (char *)sqlist + sqlist->sqlist_size);
5134 	new_sqlp->sql_next = sqlp;
5135 	new_sqlp->sql_sq = sqp;
5136 	*prev_sqlpp = new_sqlp;
5137 }
5138 
5139 /*
5140  * Walk the write side queues until we hit either the driver
5141  * or a twist in the stream (_SAMESTR will return false in both
5142  * these cases) then turn around and walk the read side queues
5143  * back up to the stream head.
5144  */
5145 static void
5146 sqlist_insertall(sqlist_t *sqlist, queue_t *q)
5147 {
5148 	while (q != NULL) {
5149 		sqlist_insert(sqlist, q->q_syncq);
5150 
5151 		if (_SAMESTR(q))
5152 			q = q->q_next;
5153 		else if (!(q->q_flag & QREADR))
5154 			q = _RD(q);
5155 		else
5156 			q = NULL;
5157 	}
5158 }
5159 
5160 /*
5161  * Allocate and build a list of all syncqs in a stream and the syncq(s)
5162  * associated with the "q" parameter. The resulting list is sorted in a
5163  * canonical order and is free of duplicates.
5164  * Assumes the passed queue is a _RD(q).
5165  */
5166 static sqlist_t *
5167 sqlist_build(queue_t *q, struct stdata *stp, boolean_t do_twist)
5168 {
5169 	sqlist_t *sqlist = sqlist_alloc(stp, KM_SLEEP);
5170 
5171 	/*
5172 	 * start with the current queue/qpair
5173 	 */
5174 	ASSERT(q->q_flag & QREADR);
5175 
5176 	sqlist_insert(sqlist, q->q_syncq);
5177 	sqlist_insert(sqlist, _WR(q)->q_syncq);
5178 
5179 	sqlist_insertall(sqlist, stp->sd_wrq);
5180 	if (do_twist)
5181 		sqlist_insertall(sqlist, stp->sd_mate->sd_wrq);
5182 
5183 	return (sqlist);
5184 }
5185 
5186 static sqlist_t *
5187 sqlist_alloc(struct stdata *stp, int kmflag)
5188 {
5189 	size_t sqlist_size;
5190 	sqlist_t *sqlist;
5191 
5192 	/*
5193 	 * Allocate 2 syncql_t's for each pushed module. Note that
5194 	 * the sqlist_t structure already has 4 syncql_t's built in:
5195 	 * 2 for the stream head, and 2 for the driver/other stream head.
5196 	 */
5197 	sqlist_size = 2 * sizeof (syncql_t) * stp->sd_pushcnt +
5198 	    sizeof (sqlist_t);
5199 	if (STRMATED(stp))
5200 		sqlist_size += 2 * sizeof (syncql_t) * stp->sd_mate->sd_pushcnt;
5201 	sqlist = kmem_alloc(sqlist_size, kmflag);
5202 
5203 	sqlist->sqlist_head = NULL;
5204 	sqlist->sqlist_size = sqlist_size;
5205 	sqlist->sqlist_index = 0;
5206 
5207 	return (sqlist);
5208 }
5209 
5210 /*
5211  * Free the list created by sqlist_alloc()
5212  */
5213 static void
5214 sqlist_free(sqlist_t *sqlist)
5215 {
5216 	kmem_free(sqlist, sqlist->sqlist_size);
5217 }
5218 
5219 /*
5220  * Prevent any new entries into any syncq in this stream.
5221  * Used by freezestr.
5222  */
5223 void
5224 strblock(queue_t *q)
5225 {
5226 	struct stdata	*stp;
5227 	syncql_t	*sql;
5228 	sqlist_t	*sqlist;
5229 
5230 	q = _RD(q);
5231 
5232 	stp = STREAM(q);
5233 	ASSERT(stp != NULL);
5234 
5235 	/*
5236 	 * Get a sorted list with all the duplicates removed containing
5237 	 * all the syncqs referenced by this stream.
5238 	 */
5239 	sqlist = sqlist_build(q, stp, B_FALSE);
5240 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5241 		blocksq(sql->sql_sq, SQ_FROZEN, -1);
5242 	sqlist_free(sqlist);
5243 }
5244 
5245 /*
5246  * Release the block on new entries into this stream
5247  */
5248 void
5249 strunblock(queue_t *q)
5250 {
5251 	struct stdata	*stp;
5252 	syncql_t	*sql;
5253 	sqlist_t	*sqlist;
5254 	int		drain_needed;
5255 
5256 	q = _RD(q);
5257 
5258 	/*
5259 	 * Get a sorted list with all the duplicates removed containing
5260 	 * all the syncqs referenced by this stream.
5261 	 * Have to drop the SQ_FROZEN flag on all the syncqs before
5262 	 * starting to drain them; otherwise the draining might
5263 	 * cause a freezestr in some module on the stream (which
5264 	 * would deadlock).
5265 	 */
5266 	stp = STREAM(q);
5267 	ASSERT(stp != NULL);
5268 	sqlist = sqlist_build(q, stp, B_FALSE);
5269 	drain_needed = 0;
5270 	for (sql = sqlist->sqlist_head; sql != NULL; sql = sql->sql_next)
5271 		drain_needed += dropsq(sql->sql_sq, SQ_FROZEN);
5272 	if (drain_needed) {
5273 		for (sql = sqlist->sqlist_head; sql != NULL;
5274 		    sql = sql->sql_next)
5275 			emptysq(sql->sql_sq);
5276 	}
5277 	sqlist_free(sqlist);
5278 }
5279 
5280 #ifdef DEBUG
5281 static int
5282 qprocsareon(queue_t *rq)
5283 {
5284 	if (rq->q_next == NULL)
5285 		return (0);
5286 	return (_WR(rq->q_next)->q_next == _WR(rq));
5287 }
5288 
5289 int
5290 qclaimed(queue_t *q)
5291 {
5292 	uint_t count;
5293 
5294 	count = q->q_syncq->sq_count;
5295 	SUM_SQ_PUTCOUNTS(q->q_syncq, count);
5296 	return (count != 0);
5297 }
5298 
5299 /*
5300  * Check if anyone has frozen this stream with freezestr
5301  */
5302 int
5303 frozenstr(queue_t *q)
5304 {
5305 	return ((q->q_syncq->sq_flags & SQ_FROZEN) != 0);
5306 }
5307 #endif /* DEBUG */
5308 
5309 /*
5310  * Enter a queue.
5311  * Obsoleted interface. Should not be used.
5312  */
5313 void
5314 enterq(queue_t *q)
5315 {
5316 	entersq(q->q_syncq, SQ_CALLBACK);
5317 }
5318 
5319 void
5320 leaveq(queue_t *q)
5321 {
5322 	leavesq(q->q_syncq, SQ_CALLBACK);
5323 }
5324 
5325 /*
5326  * Enter a perimeter. c_inner and c_outer specifies which concurrency bits
5327  * to check.
5328  * Wait if SQ_QUEUED is set to preserve ordering between messages and qwriter
5329  * calls and the running of open, close and service procedures.
5330  *
5331  * If c_inner bit is set no need to grab sq_putlocks since we don't care
5332  * if other threads have entered or are entering put entry point.
5333  *
5334  * If c_inner bit is set it might have been possible to use
5335  * sq_putlocks/sq_putcounts instead of SQLOCK/sq_count (e.g. to optimize
5336  * open/close path for IP) but since the count may need to be decremented in
5337  * qwait() we wouldn't know which counter to decrement. Currently counter is
5338  * selected by current cpu_seqid and current CPU can change at any moment. XXX
5339  * in the future we might use curthread id bits to select the counter and this
5340  * would stay constant across routine calls.
5341  */
5342 void
5343 entersq(syncq_t *sq, int entrypoint)
5344 {
5345 	uint16_t	count = 0;
5346 	uint16_t	flags;
5347 	uint16_t	waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
5348 	uint16_t	type;
5349 	uint_t		c_inner = entrypoint & SQ_CI;
5350 	uint_t		c_outer = entrypoint & SQ_CO;
5351 
5352 	/*
5353 	 * Increment ref count to keep closes out of this queue.
5354 	 */
5355 	ASSERT(sq);
5356 	ASSERT(c_inner && c_outer);
5357 	mutex_enter(SQLOCK(sq));
5358 	flags = sq->sq_flags;
5359 	type = sq->sq_type;
5360 	if (!(type & c_inner)) {
5361 		/* Make sure all putcounts now use slowlock. */
5362 		count = sq->sq_count;
5363 		SQ_PUTLOCKS_ENTER(sq);
5364 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
5365 		SUM_SQ_PUTCOUNTS(sq, count);
5366 		sq->sq_needexcl++;
5367 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
5368 		waitflags |= SQ_MESSAGES;
5369 	}
5370 	/*
5371 	 * Wait until we can enter the inner perimeter.
5372 	 * If we want exclusive access we wait until sq_count is 0.
5373 	 * We have to do this before entering the outer perimeter in order
5374 	 * to preserve put/close message ordering.
5375 	 */
5376 	while ((flags & waitflags) || (!(type & c_inner) && count != 0)) {
5377 		sq->sq_flags = flags | SQ_WANTWAKEUP;
5378 		if (!(type & c_inner)) {
5379 			SQ_PUTLOCKS_EXIT(sq);
5380 		}
5381 		cv_wait(&sq->sq_wait, SQLOCK(sq));
5382 		if (!(type & c_inner)) {
5383 			count = sq->sq_count;
5384 			SQ_PUTLOCKS_ENTER(sq);
5385 			SUM_SQ_PUTCOUNTS(sq, count);
5386 		}
5387 		flags = sq->sq_flags;
5388 	}
5389 
5390 	if (!(type & c_inner)) {
5391 		ASSERT(sq->sq_needexcl > 0);
5392 		sq->sq_needexcl--;
5393 		if (sq->sq_needexcl == 0) {
5394 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
5395 		}
5396 	}
5397 
5398 	/* Check if we need to enter the outer perimeter */
5399 	if (!(type & c_outer)) {
5400 		/*
5401 		 * We have to enter the outer perimeter exclusively before
5402 		 * we can increment sq_count to avoid deadlock. This implies
5403 		 * that we have to re-check sq_flags and sq_count.
5404 		 *
5405 		 * is it possible to have c_inner set when c_outer is not set?
5406 		 */
5407 		if (!(type & c_inner)) {
5408 			SQ_PUTLOCKS_EXIT(sq);
5409 		}
5410 		mutex_exit(SQLOCK(sq));
5411 		outer_enter(sq->sq_outer, SQ_GOAWAY);
5412 		mutex_enter(SQLOCK(sq));
5413 		flags = sq->sq_flags;
5414 		/*
5415 		 * there should be no need to recheck sq_putcounts
5416 		 * because outer_enter() has already waited for them to clear
5417 		 * after setting SQ_WRITER.
5418 		 */
5419 		count = sq->sq_count;
5420 #ifdef DEBUG
5421 		/*
5422 		 * SUMCHECK_SQ_PUTCOUNTS should return the sum instead
5423 		 * of doing an ASSERT internally. Others should do
5424 		 * something like
5425 		 *	 ASSERT(SUMCHECK_SQ_PUTCOUNTS(sq) == 0);
5426 		 * without the need to #ifdef DEBUG it.
5427 		 */
5428 		SUMCHECK_SQ_PUTCOUNTS(sq, 0);
5429 #endif
5430 		while ((flags & (SQ_EXCL|SQ_BLOCKED|SQ_FROZEN)) ||
5431 		    (!(type & c_inner) && count != 0)) {
5432 			sq->sq_flags = flags | SQ_WANTWAKEUP;
5433 			cv_wait(&sq->sq_wait, SQLOCK(sq));
5434 			count = sq->sq_count;
5435 			flags = sq->sq_flags;
5436 		}
5437 	}
5438 
5439 	sq->sq_count++;
5440 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5441 	if (!(type & c_inner)) {
5442 		/* Exclusive entry */
5443 		ASSERT(sq->sq_count == 1);
5444 		sq->sq_flags |= SQ_EXCL;
5445 		if (type & c_outer) {
5446 			SQ_PUTLOCKS_EXIT(sq);
5447 		}
5448 	}
5449 	mutex_exit(SQLOCK(sq));
5450 }
5451 
5452 /*
5453  * Leave a syncq. Announce to framework that closes may proceed.
5454  * c_inner and c_outer specify which concurrency bits to check.
5455  *
5456  * Must never be called from driver or module put entry point.
5457  *
5458  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5459  * sq_putlocks are used.
5460  */
5461 void
5462 leavesq(syncq_t *sq, int entrypoint)
5463 {
5464 	uint16_t	flags;
5465 	uint16_t	type;
5466 	uint_t		c_outer = entrypoint & SQ_CO;
5467 #ifdef DEBUG
5468 	uint_t		c_inner = entrypoint & SQ_CI;
5469 #endif
5470 
5471 	/*
5472 	 * Decrement ref count, drain the syncq if possible, and wake up
5473 	 * any waiting close.
5474 	 */
5475 	ASSERT(sq);
5476 	ASSERT(c_inner && c_outer);
5477 	mutex_enter(SQLOCK(sq));
5478 	flags = sq->sq_flags;
5479 	type = sq->sq_type;
5480 	if (flags & (SQ_QUEUED|SQ_WANTWAKEUP|SQ_WANTEXWAKEUP)) {
5481 
5482 		if (flags & SQ_WANTWAKEUP) {
5483 			flags &= ~SQ_WANTWAKEUP;
5484 			cv_broadcast(&sq->sq_wait);
5485 		}
5486 		if (flags & SQ_WANTEXWAKEUP) {
5487 			flags &= ~SQ_WANTEXWAKEUP;
5488 			cv_broadcast(&sq->sq_exitwait);
5489 		}
5490 
5491 		if ((flags & SQ_QUEUED) && !(flags & SQ_STAYAWAY)) {
5492 			/*
5493 			 * The syncq needs to be drained. "Exit" the syncq
5494 			 * before calling drain_syncq.
5495 			 */
5496 			ASSERT(sq->sq_count != 0);
5497 			sq->sq_count--;
5498 			ASSERT((flags & SQ_EXCL) || (type & c_inner));
5499 			sq->sq_flags = flags & ~SQ_EXCL;
5500 			drain_syncq(sq);
5501 			ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
5502 			/* Check if we need to exit the outer perimeter */
5503 			/* XXX will this ever be true? */
5504 			if (!(type & c_outer))
5505 				outer_exit(sq->sq_outer);
5506 			return;
5507 		}
5508 	}
5509 	ASSERT(sq->sq_count != 0);
5510 	sq->sq_count--;
5511 	ASSERT((flags & SQ_EXCL) || (type & c_inner));
5512 	sq->sq_flags = flags & ~SQ_EXCL;
5513 	mutex_exit(SQLOCK(sq));
5514 
5515 	/* Check if we need to exit the outer perimeter */
5516 	if (!(sq->sq_type & c_outer))
5517 		outer_exit(sq->sq_outer);
5518 }
5519 
5520 /*
5521  * Prevent q_next from changing in this stream by incrementing sq_count.
5522  *
5523  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5524  * sq_putlocks are used.
5525  */
5526 void
5527 claimq(queue_t *qp)
5528 {
5529 	syncq_t	*sq = qp->q_syncq;
5530 
5531 	mutex_enter(SQLOCK(sq));
5532 	sq->sq_count++;
5533 	ASSERT(sq->sq_count != 0);	/* Wraparound */
5534 	mutex_exit(SQLOCK(sq));
5535 }
5536 
5537 /*
5538  * Undo claimq.
5539  *
5540  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
5541  * sq_putlocks are used.
5542  */
5543 void
5544 releaseq(queue_t *qp)
5545 {
5546 	syncq_t	*sq = qp->q_syncq;
5547 	uint16_t flags;
5548 
5549 	mutex_enter(SQLOCK(sq));
5550 	ASSERT(sq->sq_count > 0);
5551 	sq->sq_count--;
5552 
5553 	flags = sq->sq_flags;
5554 	if (flags & (SQ_WANTWAKEUP|SQ_QUEUED)) {
5555 		if (flags & SQ_WANTWAKEUP) {
5556 			flags &= ~SQ_WANTWAKEUP;
5557 			cv_broadcast(&sq->sq_wait);
5558 		}
5559 		sq->sq_flags = flags;
5560 		if ((flags & SQ_QUEUED) && !(flags & (SQ_STAYAWAY|SQ_EXCL))) {
5561 			/*
5562 			 * To prevent potential recursive invocation of
5563 			 * drain_syncq we do not call drain_syncq if count is
5564 			 * non-zero.
5565 			 */
5566 			if (sq->sq_count == 0) {
5567 				drain_syncq(sq);
5568 				return;
5569 			} else
5570 				sqenable(sq);
5571 		}
5572 	}
5573 	mutex_exit(SQLOCK(sq));
5574 }
5575 
5576 /*
5577  * Prevent q_next from changing in this stream by incrementing sd_refcnt.
5578  */
5579 void
5580 claimstr(queue_t *qp)
5581 {
5582 	struct stdata *stp = STREAM(qp);
5583 
5584 	mutex_enter(&stp->sd_reflock);
5585 	stp->sd_refcnt++;
5586 	ASSERT(stp->sd_refcnt != 0);	/* Wraparound */
5587 	mutex_exit(&stp->sd_reflock);
5588 }
5589 
5590 /*
5591  * Undo claimstr.
5592  */
5593 void
5594 releasestr(queue_t *qp)
5595 {
5596 	struct stdata *stp = STREAM(qp);
5597 
5598 	mutex_enter(&stp->sd_reflock);
5599 	ASSERT(stp->sd_refcnt != 0);
5600 	if (--stp->sd_refcnt == 0)
5601 		cv_broadcast(&stp->sd_refmonitor);
5602 	mutex_exit(&stp->sd_reflock);
5603 }
5604 
5605 static syncq_t *
5606 new_syncq(void)
5607 {
5608 	return (kmem_cache_alloc(syncq_cache, KM_SLEEP));
5609 }
5610 
5611 static void
5612 free_syncq(syncq_t *sq)
5613 {
5614 	ASSERT(sq->sq_head == NULL);
5615 	ASSERT(sq->sq_outer == NULL);
5616 	ASSERT(sq->sq_callbpend == NULL);
5617 	ASSERT((sq->sq_onext == NULL && sq->sq_oprev == NULL) ||
5618 	    (sq->sq_onext == sq && sq->sq_oprev == sq));
5619 
5620 	if (sq->sq_ciputctrl != NULL) {
5621 		ASSERT(sq->sq_nciputctrl == n_ciputctrl - 1);
5622 		SUMCHECK_CIPUTCTRL_COUNTS(sq->sq_ciputctrl,
5623 		    sq->sq_nciputctrl, 0);
5624 		ASSERT(ciputctrl_cache != NULL);
5625 		kmem_cache_free(ciputctrl_cache, sq->sq_ciputctrl);
5626 	}
5627 
5628 	sq->sq_tail = NULL;
5629 	sq->sq_evhead = NULL;
5630 	sq->sq_evtail = NULL;
5631 	sq->sq_ciputctrl = NULL;
5632 	sq->sq_nciputctrl = 0;
5633 	sq->sq_count = 0;
5634 	sq->sq_rmqcount = 0;
5635 	sq->sq_callbflags = 0;
5636 	sq->sq_cancelid = 0;
5637 	sq->sq_next = NULL;
5638 	sq->sq_needexcl = 0;
5639 	sq->sq_svcflags = 0;
5640 	sq->sq_nqueues = 0;
5641 	sq->sq_pri = 0;
5642 	sq->sq_onext = NULL;
5643 	sq->sq_oprev = NULL;
5644 	sq->sq_flags = 0;
5645 	sq->sq_type = 0;
5646 	sq->sq_servcount = 0;
5647 
5648 	kmem_cache_free(syncq_cache, sq);
5649 }
5650 
5651 /* Outer perimeter code */
5652 
5653 /*
5654  * The outer syncq uses the fields and flags in the syncq slightly
5655  * differently from the inner syncqs.
5656  *	sq_count	Incremented when there are pending or running
5657  *			writers at the outer perimeter to prevent the set of
5658  *			inner syncqs that belong to the outer perimeter from
5659  *			changing.
5660  *	sq_head/tail	List of deferred qwriter(OUTER) operations.
5661  *
5662  *	SQ_BLOCKED	Set to prevent traversing of sq_next,sq_prev while
5663  *			inner syncqs are added to or removed from the
5664  *			outer perimeter.
5665  *	SQ_QUEUED	sq_head/tail has messages or events queued.
5666  *
5667  *	SQ_WRITER	A thread is currently traversing all the inner syncqs
5668  *			setting the SQ_WRITER flag.
5669  */
5670 
5671 /*
5672  * Get write access at the outer perimeter.
5673  * Note that read access is done by entersq, putnext, and put by simply
5674  * incrementing sq_count in the inner syncq.
5675  *
5676  * Waits until "flags" is no longer set in the outer to prevent multiple
5677  * threads from having write access at the same time. SQ_WRITER has to be part
5678  * of "flags".
5679  *
5680  * Increases sq_count on the outer syncq to keep away outer_insert/remove
5681  * until the outer_exit is finished.
5682  *
5683  * outer_enter is vulnerable to starvation since it does not prevent new
5684  * threads from entering the inner syncqs while it is waiting for sq_count to
5685  * go to zero.
5686  */
5687 void
5688 outer_enter(syncq_t *outer, uint16_t flags)
5689 {
5690 	syncq_t	*sq;
5691 	int	wait_needed;
5692 	uint16_t	count;
5693 
5694 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5695 	    outer->sq_oprev != NULL);
5696 	ASSERT(flags & SQ_WRITER);
5697 
5698 retry:
5699 	mutex_enter(SQLOCK(outer));
5700 	while (outer->sq_flags & flags) {
5701 		outer->sq_flags |= SQ_WANTWAKEUP;
5702 		cv_wait(&outer->sq_wait, SQLOCK(outer));
5703 	}
5704 
5705 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5706 	outer->sq_flags |= SQ_WRITER;
5707 	outer->sq_count++;
5708 	ASSERT(outer->sq_count != 0);	/* wraparound */
5709 	wait_needed = 0;
5710 	/*
5711 	 * Set SQ_WRITER on all the inner syncqs while holding
5712 	 * the SQLOCK on the outer syncq. This ensures that the changing
5713 	 * of SQ_WRITER is atomic under the outer SQLOCK.
5714 	 */
5715 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5716 		mutex_enter(SQLOCK(sq));
5717 		count = sq->sq_count;
5718 		SQ_PUTLOCKS_ENTER(sq);
5719 		sq->sq_flags |= SQ_WRITER;
5720 		SUM_SQ_PUTCOUNTS(sq, count);
5721 		if (count != 0)
5722 			wait_needed = 1;
5723 		SQ_PUTLOCKS_EXIT(sq);
5724 		mutex_exit(SQLOCK(sq));
5725 	}
5726 	mutex_exit(SQLOCK(outer));
5727 
5728 	/*
5729 	 * Get everybody out of the syncqs sequentially.
5730 	 * Note that we don't actually need to acquire the PUTLOCKS, since
5731 	 * we have already cleared the fastbit, and set QWRITER.  By
5732 	 * definition, the count can not increase since putnext will
5733 	 * take the slowlock path (and the purpose of acquiring the
5734 	 * putlocks was to make sure it didn't increase while we were
5735 	 * waiting).
5736 	 *
5737 	 * Note that we still acquire the PUTLOCKS to be safe.
5738 	 */
5739 	if (wait_needed) {
5740 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
5741 			mutex_enter(SQLOCK(sq));
5742 			count = sq->sq_count;
5743 			SQ_PUTLOCKS_ENTER(sq);
5744 			SUM_SQ_PUTCOUNTS(sq, count);
5745 			while (count != 0) {
5746 				sq->sq_flags |= SQ_WANTWAKEUP;
5747 				SQ_PUTLOCKS_EXIT(sq);
5748 				cv_wait(&sq->sq_wait, SQLOCK(sq));
5749 				count = sq->sq_count;
5750 				SQ_PUTLOCKS_ENTER(sq);
5751 				SUM_SQ_PUTCOUNTS(sq, count);
5752 			}
5753 			SQ_PUTLOCKS_EXIT(sq);
5754 			mutex_exit(SQLOCK(sq));
5755 		}
5756 		/*
5757 		 * Verify that none of the flags got set while we
5758 		 * were waiting for the sq_counts to drop.
5759 		 * If this happens we exit and retry entering the
5760 		 * outer perimeter.
5761 		 */
5762 		mutex_enter(SQLOCK(outer));
5763 		if (outer->sq_flags & (flags & ~SQ_WRITER)) {
5764 			mutex_exit(SQLOCK(outer));
5765 			outer_exit(outer);
5766 			goto retry;
5767 		}
5768 		mutex_exit(SQLOCK(outer));
5769 	}
5770 }
5771 
5772 /*
5773  * Drop the write access at the outer perimeter.
5774  * Read access is dropped implicitly (by putnext, put, and leavesq) by
5775  * decrementing sq_count.
5776  */
5777 void
5778 outer_exit(syncq_t *outer)
5779 {
5780 	syncq_t	*sq;
5781 	int	 drain_needed;
5782 	uint16_t flags;
5783 
5784 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5785 	    outer->sq_oprev != NULL);
5786 	ASSERT(MUTEX_NOT_HELD(SQLOCK(outer)));
5787 
5788 	/*
5789 	 * Atomically (from the perspective of threads calling become_writer)
5790 	 * drop the write access at the outer perimeter by holding
5791 	 * SQLOCK(outer) across all the dropsq calls and the resetting of
5792 	 * SQ_WRITER.
5793 	 * This defines a locking order between the outer perimeter
5794 	 * SQLOCK and the inner perimeter SQLOCKs.
5795 	 */
5796 	mutex_enter(SQLOCK(outer));
5797 	flags = outer->sq_flags;
5798 	ASSERT(outer->sq_flags & SQ_WRITER);
5799 	if (flags & SQ_QUEUED) {
5800 		write_now(outer);
5801 		flags = outer->sq_flags;
5802 	}
5803 
5804 	/*
5805 	 * sq_onext is stable since sq_count has not yet been decreased.
5806 	 * Reset the SQ_WRITER flags in all syncqs.
5807 	 * After dropping SQ_WRITER on the outer syncq we empty all the
5808 	 * inner syncqs.
5809 	 */
5810 	drain_needed = 0;
5811 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5812 		drain_needed += dropsq(sq, SQ_WRITER);
5813 	ASSERT(!(outer->sq_flags & SQ_QUEUED));
5814 	flags &= ~SQ_WRITER;
5815 	if (drain_needed) {
5816 		outer->sq_flags = flags;
5817 		mutex_exit(SQLOCK(outer));
5818 		for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext)
5819 			emptysq(sq);
5820 		mutex_enter(SQLOCK(outer));
5821 		flags = outer->sq_flags;
5822 	}
5823 	if (flags & SQ_WANTWAKEUP) {
5824 		flags &= ~SQ_WANTWAKEUP;
5825 		cv_broadcast(&outer->sq_wait);
5826 	}
5827 	outer->sq_flags = flags;
5828 	ASSERT(outer->sq_count > 0);
5829 	outer->sq_count--;
5830 	mutex_exit(SQLOCK(outer));
5831 }
5832 
5833 /*
5834  * Add another syncq to an outer perimeter.
5835  * Block out all other access to the outer perimeter while it is being
5836  * changed using blocksq.
5837  * Assumes that the caller has *not* done an outer_enter.
5838  *
5839  * Vulnerable to starvation in blocksq.
5840  */
5841 static void
5842 outer_insert(syncq_t *outer, syncq_t *sq)
5843 {
5844 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5845 	    outer->sq_oprev != NULL);
5846 	ASSERT(sq->sq_outer == NULL && sq->sq_onext == NULL &&
5847 	    sq->sq_oprev == NULL);	/* Can't be in an outer perimeter */
5848 
5849 	/* Get exclusive access to the outer perimeter list */
5850 	blocksq(outer, SQ_BLOCKED, 0);
5851 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5852 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5853 
5854 	mutex_enter(SQLOCK(sq));
5855 	sq->sq_outer = outer;
5856 	outer->sq_onext->sq_oprev = sq;
5857 	sq->sq_onext = outer->sq_onext;
5858 	outer->sq_onext = sq;
5859 	sq->sq_oprev = outer;
5860 	mutex_exit(SQLOCK(sq));
5861 	unblocksq(outer, SQ_BLOCKED, 1);
5862 }
5863 
5864 /*
5865  * Remove a syncq from an outer perimeter.
5866  * Block out all other access to the outer perimeter while it is being
5867  * changed using blocksq.
5868  * Assumes that the caller has *not* done an outer_enter.
5869  *
5870  * Vulnerable to starvation in blocksq.
5871  */
5872 static void
5873 outer_remove(syncq_t *outer, syncq_t *sq)
5874 {
5875 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5876 	    outer->sq_oprev != NULL);
5877 	ASSERT(sq->sq_outer == outer);
5878 
5879 	/* Get exclusive access to the outer perimeter list */
5880 	blocksq(outer, SQ_BLOCKED, 0);
5881 	ASSERT(outer->sq_flags & SQ_BLOCKED);
5882 	ASSERT(!(outer->sq_flags & SQ_WRITER));
5883 
5884 	mutex_enter(SQLOCK(sq));
5885 	sq->sq_outer = NULL;
5886 	sq->sq_onext->sq_oprev = sq->sq_oprev;
5887 	sq->sq_oprev->sq_onext = sq->sq_onext;
5888 	sq->sq_oprev = sq->sq_onext = NULL;
5889 	mutex_exit(SQLOCK(sq));
5890 	unblocksq(outer, SQ_BLOCKED, 1);
5891 }
5892 
5893 /*
5894  * Queue a deferred qwriter(OUTER) callback for this outer perimeter.
5895  * If this is the first callback for this outer perimeter then add
5896  * this outer perimeter to the list of outer perimeters that
5897  * the qwriter_outer_thread will process.
5898  *
5899  * Increments sq_count in the outer syncq to prevent the membership
5900  * of the outer perimeter (in terms of inner syncqs) to change while
5901  * the callback is pending.
5902  */
5903 static void
5904 queue_writer(syncq_t *outer, void (*func)(), queue_t *q, mblk_t *mp)
5905 {
5906 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
5907 
5908 	mp->b_prev = (mblk_t *)func;
5909 	mp->b_queue = q;
5910 	mp->b_next = NULL;
5911 	outer->sq_count++;	/* Decremented when dequeued */
5912 	ASSERT(outer->sq_count != 0);	/* Wraparound */
5913 	if (outer->sq_evhead == NULL) {
5914 		/* First message. */
5915 		outer->sq_evhead = outer->sq_evtail = mp;
5916 		outer->sq_flags |= SQ_EVENTS;
5917 		mutex_exit(SQLOCK(outer));
5918 		STRSTAT(qwr_outer);
5919 		(void) taskq_dispatch(streams_taskq,
5920 		    (task_func_t *)qwriter_outer_service, outer, TQ_SLEEP);
5921 	} else {
5922 		ASSERT(outer->sq_flags & SQ_EVENTS);
5923 		outer->sq_evtail->b_next = mp;
5924 		outer->sq_evtail = mp;
5925 		mutex_exit(SQLOCK(outer));
5926 	}
5927 }
5928 
5929 /*
5930  * Try and upgrade to write access at the outer perimeter. If this can
5931  * not be done without blocking then queue the callback to be done
5932  * by the qwriter_outer_thread.
5933  *
5934  * This routine can only be called from put or service procedures plus
5935  * asynchronous callback routines that have properly entered the queue (with
5936  * entersq). Thus qwriter(OUTER) assumes the caller has one claim on the syncq
5937  * associated with q.
5938  */
5939 void
5940 qwriter_outer(queue_t *q, mblk_t *mp, void (*func)())
5941 {
5942 	syncq_t	*osq, *sq, *outer;
5943 	int	failed;
5944 	uint16_t flags;
5945 
5946 	osq = q->q_syncq;
5947 	outer = osq->sq_outer;
5948 	if (outer == NULL)
5949 		panic("qwriter(PERIM_OUTER): no outer perimeter");
5950 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
5951 	    outer->sq_oprev != NULL);
5952 
5953 	mutex_enter(SQLOCK(outer));
5954 	flags = outer->sq_flags;
5955 	/*
5956 	 * If some thread is traversing sq_next, or if we are blocked by
5957 	 * outer_insert or outer_remove, or if the we already have queued
5958 	 * callbacks, then queue this callback for later processing.
5959 	 *
5960 	 * Also queue the qwriter for an interrupt thread in order
5961 	 * to reduce the time spent running at high IPL.
5962 	 * to identify there are events.
5963 	 */
5964 	if ((flags & SQ_GOAWAY) || (curthread->t_pri >= kpreemptpri)) {
5965 		/*
5966 		 * Queue the become_writer request.
5967 		 * The queueing is atomic under SQLOCK(outer) in order
5968 		 * to synchronize with outer_exit.
5969 		 * queue_writer will drop the outer SQLOCK
5970 		 */
5971 		if (flags & SQ_BLOCKED) {
5972 			/* Must set SQ_WRITER on inner perimeter */
5973 			mutex_enter(SQLOCK(osq));
5974 			osq->sq_flags |= SQ_WRITER;
5975 			mutex_exit(SQLOCK(osq));
5976 		} else {
5977 			if (!(flags & SQ_WRITER)) {
5978 				/*
5979 				 * The outer could have been SQ_BLOCKED thus
5980 				 * SQ_WRITER might not be set on the inner.
5981 				 */
5982 				mutex_enter(SQLOCK(osq));
5983 				osq->sq_flags |= SQ_WRITER;
5984 				mutex_exit(SQLOCK(osq));
5985 			}
5986 			ASSERT(osq->sq_flags & SQ_WRITER);
5987 		}
5988 		queue_writer(outer, func, q, mp);
5989 		return;
5990 	}
5991 	/*
5992 	 * We are half-way to exclusive access to the outer perimeter.
5993 	 * Prevent any outer_enter, qwriter(OUTER), or outer_insert/remove
5994 	 * while the inner syncqs are traversed.
5995 	 */
5996 	outer->sq_count++;
5997 	ASSERT(outer->sq_count != 0);	/* wraparound */
5998 	flags |= SQ_WRITER;
5999 	/*
6000 	 * Check if we can run the function immediately. Mark all
6001 	 * syncqs with the writer flag to prevent new entries into
6002 	 * put and service procedures.
6003 	 *
6004 	 * Set SQ_WRITER on all the inner syncqs while holding
6005 	 * the SQLOCK on the outer syncq. This ensures that the changing
6006 	 * of SQ_WRITER is atomic under the outer SQLOCK.
6007 	 */
6008 	failed = 0;
6009 	for (sq = outer->sq_onext; sq != outer; sq = sq->sq_onext) {
6010 		uint16_t count;
6011 		uint_t	maxcnt = (sq == osq) ? 1 : 0;
6012 
6013 		mutex_enter(SQLOCK(sq));
6014 		count = sq->sq_count;
6015 		SQ_PUTLOCKS_ENTER(sq);
6016 		SUM_SQ_PUTCOUNTS(sq, count);
6017 		if (sq->sq_count > maxcnt)
6018 			failed = 1;
6019 		sq->sq_flags |= SQ_WRITER;
6020 		SQ_PUTLOCKS_EXIT(sq);
6021 		mutex_exit(SQLOCK(sq));
6022 	}
6023 	if (failed) {
6024 		/*
6025 		 * Some other thread has a read claim on the outer perimeter.
6026 		 * Queue the callback for deferred processing.
6027 		 *
6028 		 * queue_writer will set SQ_QUEUED before we drop SQ_WRITER
6029 		 * so that other qwriter(OUTER) calls will queue their
6030 		 * callbacks as well. queue_writer increments sq_count so we
6031 		 * decrement to compensate for the our increment.
6032 		 *
6033 		 * Dropping SQ_WRITER enables the writer thread to work
6034 		 * on this outer perimeter.
6035 		 */
6036 		outer->sq_flags = flags;
6037 		queue_writer(outer, func, q, mp);
6038 		/* queue_writer dropper the lock */
6039 		mutex_enter(SQLOCK(outer));
6040 		ASSERT(outer->sq_count > 0);
6041 		outer->sq_count--;
6042 		ASSERT(outer->sq_flags & SQ_WRITER);
6043 		flags = outer->sq_flags;
6044 		flags &= ~SQ_WRITER;
6045 		if (flags & SQ_WANTWAKEUP) {
6046 			flags &= ~SQ_WANTWAKEUP;
6047 			cv_broadcast(&outer->sq_wait);
6048 		}
6049 		outer->sq_flags = flags;
6050 		mutex_exit(SQLOCK(outer));
6051 		return;
6052 	} else {
6053 		outer->sq_flags = flags;
6054 		mutex_exit(SQLOCK(outer));
6055 	}
6056 
6057 	/* Can run it immediately */
6058 	(*func)(q, mp);
6059 
6060 	outer_exit(outer);
6061 }
6062 
6063 /*
6064  * Dequeue all writer callbacks from the outer perimeter and run them.
6065  */
6066 static void
6067 write_now(syncq_t *outer)
6068 {
6069 	mblk_t		*mp;
6070 	queue_t		*q;
6071 	void	(*func)();
6072 
6073 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6074 	ASSERT(outer->sq_outer == NULL && outer->sq_onext != NULL &&
6075 	    outer->sq_oprev != NULL);
6076 	while ((mp = outer->sq_evhead) != NULL) {
6077 		/*
6078 		 * queues cannot be placed on the queuelist on the outer
6079 		 * perimeter.
6080 		 */
6081 		ASSERT(!(outer->sq_flags & SQ_MESSAGES));
6082 		ASSERT((outer->sq_flags & SQ_EVENTS));
6083 
6084 		outer->sq_evhead = mp->b_next;
6085 		if (outer->sq_evhead == NULL) {
6086 			outer->sq_evtail = NULL;
6087 			outer->sq_flags &= ~SQ_EVENTS;
6088 		}
6089 		ASSERT(outer->sq_count != 0);
6090 		outer->sq_count--;	/* Incremented when enqueued. */
6091 		mutex_exit(SQLOCK(outer));
6092 		/*
6093 		 * Drop the message if the queue is closing.
6094 		 * Make sure that the queue is "claimed" when the callback
6095 		 * is run in order to satisfy various ASSERTs.
6096 		 */
6097 		q = mp->b_queue;
6098 		func = (void (*)())mp->b_prev;
6099 		ASSERT(func != NULL);
6100 		mp->b_next = mp->b_prev = NULL;
6101 		if (q->q_flag & QWCLOSE) {
6102 			freemsg(mp);
6103 		} else {
6104 			claimq(q);
6105 			(*func)(q, mp);
6106 			releaseq(q);
6107 		}
6108 		mutex_enter(SQLOCK(outer));
6109 	}
6110 	ASSERT(MUTEX_HELD(SQLOCK(outer)));
6111 }
6112 
6113 /*
6114  * The list of messages on the inner syncq is effectively hashed
6115  * by destination queue.  These destination queues are doubly
6116  * linked lists (hopefully) in priority order.  Messages are then
6117  * put on the queue referenced by the q_sqhead/q_sqtail elements.
6118  * Additional messages are linked together by the b_next/b_prev
6119  * elements in the mblk, with (similar to putq()) the first message
6120  * having a NULL b_prev and the last message having a NULL b_next.
6121  *
6122  * Events, such as qwriter callbacks, are put onto a list in FIFO
6123  * order referenced by sq_evhead, and sq_evtail.  This is a singly
6124  * linked list, and messages here MUST be processed in the order queued.
6125  */
6126 
6127 /*
6128  * Run the events on the syncq event list (sq_evhead).
6129  * Assumes there is only one claim on the syncq, it is
6130  * already exclusive (SQ_EXCL set), and the SQLOCK held.
6131  * Messages here are processed in order, with the SQ_EXCL bit
6132  * held all the way through till the last message is processed.
6133  */
6134 void
6135 sq_run_events(syncq_t *sq)
6136 {
6137 	mblk_t		*bp;
6138 	queue_t		*qp;
6139 	uint16_t	flags = sq->sq_flags;
6140 	void		(*func)();
6141 
6142 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6143 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6144 	    sq->sq_oprev == NULL) ||
6145 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6146 	    sq->sq_oprev != NULL));
6147 
6148 	ASSERT(flags & SQ_EXCL);
6149 	ASSERT(sq->sq_count == 1);
6150 
6151 	/*
6152 	 * We need to process all of the events on this list.  It
6153 	 * is possible that new events will be added while we are
6154 	 * away processing a callback, so on every loop, we start
6155 	 * back at the beginning of the list.
6156 	 */
6157 	/*
6158 	 * We have to reaccess sq_evhead since there is a
6159 	 * possibility of a new entry while we were running
6160 	 * the callback.
6161 	 */
6162 	for (bp = sq->sq_evhead; bp != NULL; bp = sq->sq_evhead) {
6163 		ASSERT(bp->b_queue->q_syncq == sq);
6164 		ASSERT(sq->sq_flags & SQ_EVENTS);
6165 
6166 		qp = bp->b_queue;
6167 		func = (void (*)())bp->b_prev;
6168 		ASSERT(func != NULL);
6169 
6170 		/*
6171 		 * Messages from the event queue must be taken off in
6172 		 * FIFO order.
6173 		 */
6174 		ASSERT(sq->sq_evhead == bp);
6175 		sq->sq_evhead = bp->b_next;
6176 
6177 		if (bp->b_next == NULL) {
6178 			/* Deleting last */
6179 			ASSERT(sq->sq_evtail == bp);
6180 			sq->sq_evtail = NULL;
6181 			sq->sq_flags &= ~SQ_EVENTS;
6182 		}
6183 		bp->b_prev = bp->b_next = NULL;
6184 		ASSERT(bp->b_datap->db_ref != 0);
6185 
6186 		mutex_exit(SQLOCK(sq));
6187 
6188 		(*func)(qp, bp);
6189 
6190 		mutex_enter(SQLOCK(sq));
6191 		/*
6192 		 * re-read the flags, since they could have changed.
6193 		 */
6194 		flags = sq->sq_flags;
6195 		ASSERT(flags & SQ_EXCL);
6196 	}
6197 	ASSERT(sq->sq_evhead == NULL && sq->sq_evtail == NULL);
6198 	ASSERT(!(sq->sq_flags & SQ_EVENTS));
6199 
6200 	if (flags & SQ_WANTWAKEUP) {
6201 		flags &= ~SQ_WANTWAKEUP;
6202 		cv_broadcast(&sq->sq_wait);
6203 	}
6204 	if (flags & SQ_WANTEXWAKEUP) {
6205 		flags &= ~SQ_WANTEXWAKEUP;
6206 		cv_broadcast(&sq->sq_exitwait);
6207 	}
6208 	sq->sq_flags = flags;
6209 }
6210 
6211 /*
6212  * Put messages on the event list.
6213  * If we can go exclusive now, do so and process the event list, otherwise
6214  * let the last claim service this list (or wake the sqthread).
6215  * This procedure assumes SQLOCK is held.  To run the event list, it
6216  * must be called with no claims.
6217  */
6218 static void
6219 sqfill_events(syncq_t *sq, queue_t *q, mblk_t *mp, void (*func)())
6220 {
6221 	uint16_t count;
6222 
6223 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6224 	ASSERT(func != NULL);
6225 
6226 	/*
6227 	 * This is a callback.  Add it to the list of callbacks
6228 	 * and see about upgrading.
6229 	 */
6230 	mp->b_prev = (mblk_t *)func;
6231 	mp->b_queue = q;
6232 	mp->b_next = NULL;
6233 	if (sq->sq_evhead == NULL) {
6234 		sq->sq_evhead = sq->sq_evtail = mp;
6235 		sq->sq_flags |= SQ_EVENTS;
6236 	} else {
6237 		ASSERT(sq->sq_evtail != NULL);
6238 		ASSERT(sq->sq_evtail->b_next == NULL);
6239 		ASSERT(sq->sq_flags & SQ_EVENTS);
6240 		sq->sq_evtail->b_next = mp;
6241 		sq->sq_evtail = mp;
6242 	}
6243 	/*
6244 	 * We have set SQ_EVENTS, so threads will have to
6245 	 * unwind out of the perimeter, and new entries will
6246 	 * not grab a putlock.  But we still need to know
6247 	 * how many threads have already made a claim to the
6248 	 * syncq, so grab the putlocks, and sum the counts.
6249 	 * If there are no claims on the syncq, we can upgrade
6250 	 * to exclusive, and run the event list.
6251 	 * NOTE: We hold the SQLOCK, so we can just grab the
6252 	 * putlocks.
6253 	 */
6254 	count = sq->sq_count;
6255 	SQ_PUTLOCKS_ENTER(sq);
6256 	SUM_SQ_PUTCOUNTS(sq, count);
6257 	/*
6258 	 * We have no claim, so we need to check if there
6259 	 * are no others, then we can upgrade.
6260 	 */
6261 	/*
6262 	 * There are currently no claims on
6263 	 * the syncq by this thread (at least on this entry). The thread who has
6264 	 * the claim should drain syncq.
6265 	 */
6266 	if (count > 0) {
6267 		/*
6268 		 * Can't upgrade - other threads inside.
6269 		 */
6270 		SQ_PUTLOCKS_EXIT(sq);
6271 		mutex_exit(SQLOCK(sq));
6272 		return;
6273 	}
6274 	/*
6275 	 * Need to set SQ_EXCL and make a claim on the syncq.
6276 	 */
6277 	ASSERT((sq->sq_flags & SQ_EXCL) == 0);
6278 	sq->sq_flags |= SQ_EXCL;
6279 	ASSERT(sq->sq_count == 0);
6280 	sq->sq_count++;
6281 	SQ_PUTLOCKS_EXIT(sq);
6282 
6283 	/* Process the events list */
6284 	sq_run_events(sq);
6285 
6286 	/*
6287 	 * Release our claim...
6288 	 */
6289 	sq->sq_count--;
6290 
6291 	/*
6292 	 * And release SQ_EXCL.
6293 	 * We don't need to acquire the putlocks to release
6294 	 * SQ_EXCL, since we are exclusive, and hold the SQLOCK.
6295 	 */
6296 	sq->sq_flags &= ~SQ_EXCL;
6297 
6298 	/*
6299 	 * sq_run_events should have released SQ_EXCL
6300 	 */
6301 	ASSERT(!(sq->sq_flags & SQ_EXCL));
6302 
6303 	/*
6304 	 * If anything happened while we were running the
6305 	 * events (or was there before), we need to process
6306 	 * them now.  We shouldn't be exclusive sine we
6307 	 * released the perimeter above (plus, we asserted
6308 	 * for it).
6309 	 */
6310 	if (!(sq->sq_flags & SQ_STAYAWAY) && (sq->sq_flags & SQ_QUEUED))
6311 		drain_syncq(sq);
6312 	else
6313 		mutex_exit(SQLOCK(sq));
6314 }
6315 
6316 /*
6317  * Perform delayed processing. The caller has to make sure that it is safe
6318  * to enter the syncq (e.g. by checking that none of the SQ_STAYAWAY bits are
6319  * set).
6320  *
6321  * Assume that the caller has NO claims on the syncq.  However, a claim
6322  * on the syncq does not indicate that a thread is draining the syncq.
6323  * There may be more claims on the syncq than there are threads draining
6324  * (i.e.  #_threads_draining <= sq_count)
6325  *
6326  * drain_syncq has to terminate when one of the SQ_STAYAWAY bits gets set
6327  * in order to preserve qwriter(OUTER) ordering constraints.
6328  *
6329  * sq_putcount only needs to be checked when dispatching the queued
6330  * writer call for CIPUT sync queue, but this is handled in sq_run_events.
6331  */
6332 void
6333 drain_syncq(syncq_t *sq)
6334 {
6335 	queue_t		*qp;
6336 	uint16_t	count;
6337 	uint16_t	type = sq->sq_type;
6338 	uint16_t	flags = sq->sq_flags;
6339 	boolean_t	bg_service = sq->sq_svcflags & SQ_SERVICE;
6340 
6341 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6342 	    "drain_syncq start:%p", sq);
6343 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6344 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6345 	    sq->sq_oprev == NULL) ||
6346 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6347 	    sq->sq_oprev != NULL));
6348 
6349 	/*
6350 	 * Drop SQ_SERVICE flag.
6351 	 */
6352 	if (bg_service)
6353 		sq->sq_svcflags &= ~SQ_SERVICE;
6354 
6355 	/*
6356 	 * If SQ_EXCL is set, someone else is processing this syncq - let him
6357 	 * finish the job.
6358 	 */
6359 	if (flags & SQ_EXCL) {
6360 		if (bg_service) {
6361 			ASSERT(sq->sq_servcount != 0);
6362 			sq->sq_servcount--;
6363 		}
6364 		mutex_exit(SQLOCK(sq));
6365 		return;
6366 	}
6367 
6368 	/*
6369 	 * This routine can be called by a background thread if
6370 	 * it was scheduled by a hi-priority thread.  SO, if there are
6371 	 * NOT messages queued, return (remember, we have the SQLOCK,
6372 	 * and it cannot change until we release it). Wakeup any waiters also.
6373 	 */
6374 	if (!(flags & SQ_QUEUED)) {
6375 		if (flags & SQ_WANTWAKEUP) {
6376 			flags &= ~SQ_WANTWAKEUP;
6377 			cv_broadcast(&sq->sq_wait);
6378 		}
6379 		if (flags & SQ_WANTEXWAKEUP) {
6380 			flags &= ~SQ_WANTEXWAKEUP;
6381 			cv_broadcast(&sq->sq_exitwait);
6382 		}
6383 		sq->sq_flags = flags;
6384 		if (bg_service) {
6385 			ASSERT(sq->sq_servcount != 0);
6386 			sq->sq_servcount--;
6387 		}
6388 		mutex_exit(SQLOCK(sq));
6389 		return;
6390 	}
6391 
6392 	/*
6393 	 * If this is not a concurrent put perimeter, we need to
6394 	 * become exclusive to drain.  Also, if not CIPUT, we would
6395 	 * not have acquired a putlock, so we don't need to check
6396 	 * the putcounts.  If not entering with a claim, we test
6397 	 * for sq_count == 0.
6398 	 */
6399 	type = sq->sq_type;
6400 	if (!(type & SQ_CIPUT)) {
6401 		if (sq->sq_count > 1) {
6402 			if (bg_service) {
6403 				ASSERT(sq->sq_servcount != 0);
6404 				sq->sq_servcount--;
6405 			}
6406 			mutex_exit(SQLOCK(sq));
6407 			return;
6408 		}
6409 		sq->sq_flags |= SQ_EXCL;
6410 	}
6411 
6412 	/*
6413 	 * This is where we make a claim to the syncq.
6414 	 * This can either be done by incrementing a putlock, or
6415 	 * the sq_count.  But since we already have the SQLOCK
6416 	 * here, we just bump the sq_count.
6417 	 *
6418 	 * Note that after we make a claim, we need to let the code
6419 	 * fall through to the end of this routine to clean itself
6420 	 * up.  A return in the while loop will put the syncq in a
6421 	 * very bad state.
6422 	 */
6423 	sq->sq_count++;
6424 	ASSERT(sq->sq_count != 0);	/* wraparound */
6425 
6426 	while ((flags = sq->sq_flags) & SQ_QUEUED) {
6427 		/*
6428 		 * If we are told to stayaway or went exclusive,
6429 		 * we are done.
6430 		 */
6431 		if (flags & (SQ_STAYAWAY)) {
6432 			break;
6433 		}
6434 
6435 		/*
6436 		 * If there are events to run, do so.
6437 		 * We have one claim to the syncq, so if there are
6438 		 * more than one, other threads are running.
6439 		 */
6440 		if (sq->sq_evhead != NULL) {
6441 			ASSERT(sq->sq_flags & SQ_EVENTS);
6442 
6443 			count = sq->sq_count;
6444 			SQ_PUTLOCKS_ENTER(sq);
6445 			SUM_SQ_PUTCOUNTS(sq, count);
6446 			if (count > 1) {
6447 				SQ_PUTLOCKS_EXIT(sq);
6448 				/* Can't upgrade - other threads inside */
6449 				break;
6450 			}
6451 			ASSERT((flags & SQ_EXCL) == 0);
6452 			sq->sq_flags = flags | SQ_EXCL;
6453 			SQ_PUTLOCKS_EXIT(sq);
6454 			/*
6455 			 * we have the only claim, run the events,
6456 			 * sq_run_events will clear the SQ_EXCL flag.
6457 			 */
6458 			sq_run_events(sq);
6459 
6460 			/*
6461 			 * If this is a CIPUT perimeter, we need
6462 			 * to drop the SQ_EXCL flag so we can properly
6463 			 * continue draining the syncq.
6464 			 */
6465 			if (type & SQ_CIPUT) {
6466 				ASSERT(sq->sq_flags & SQ_EXCL);
6467 				sq->sq_flags &= ~SQ_EXCL;
6468 			}
6469 
6470 			/*
6471 			 * And go back to the beginning just in case
6472 			 * anything changed while we were away.
6473 			 */
6474 			ASSERT((sq->sq_flags & SQ_EXCL) || (type & SQ_CIPUT));
6475 			continue;
6476 		}
6477 
6478 		ASSERT(sq->sq_evhead == NULL);
6479 		ASSERT(!(sq->sq_flags & SQ_EVENTS));
6480 
6481 		/*
6482 		 * Find the queue that is not draining.
6483 		 *
6484 		 * q_draining is protected by QLOCK which we do not hold.
6485 		 * But if it was set, then a thread was draining, and if it gets
6486 		 * cleared, then it was because the thread has successfully
6487 		 * drained the syncq, or a GOAWAY state occurred. For the GOAWAY
6488 		 * state to happen, a thread needs the SQLOCK which we hold, and
6489 		 * if there was such a flag, we would have already seen it.
6490 		 */
6491 
6492 		for (qp = sq->sq_head;
6493 		    qp != NULL && (qp->q_draining ||
6494 		    (qp->q_sqflags & Q_SQDRAINING));
6495 		    qp = qp->q_sqnext)
6496 			;
6497 
6498 		if (qp == NULL)
6499 			break;
6500 
6501 		/*
6502 		 * We have a queue to work on, and we hold the
6503 		 * SQLOCK and one claim, call qdrain_syncq.
6504 		 * This means we need to release the SQLOCK and
6505 		 * acquire the QLOCK (OK since we have a claim).
6506 		 * Note that qdrain_syncq will actually dequeue
6507 		 * this queue from the sq_head list when it is
6508 		 * convinced all the work is done and release
6509 		 * the QLOCK before returning.
6510 		 */
6511 		qp->q_sqflags |= Q_SQDRAINING;
6512 		mutex_exit(SQLOCK(sq));
6513 		mutex_enter(QLOCK(qp));
6514 		qdrain_syncq(sq, qp);
6515 		mutex_enter(SQLOCK(sq));
6516 
6517 		/* The queue is drained */
6518 		ASSERT(qp->q_sqflags & Q_SQDRAINING);
6519 		qp->q_sqflags &= ~Q_SQDRAINING;
6520 		/*
6521 		 * NOTE: After this point qp should not be used since it may be
6522 		 * closed.
6523 		 */
6524 	}
6525 
6526 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
6527 	flags = sq->sq_flags;
6528 
6529 	/*
6530 	 * sq->sq_head cannot change because we hold the
6531 	 * sqlock. However, a thread CAN decide that it is no longer
6532 	 * going to drain that queue.  However, this should be due to
6533 	 * a GOAWAY state, and we should see that here.
6534 	 *
6535 	 * This loop is not very efficient. One solution may be adding a second
6536 	 * pointer to the "draining" queue, but it is difficult to do when
6537 	 * queues are inserted in the middle due to priority ordering. Another
6538 	 * possibility is to yank the queue out of the sq list and put it onto
6539 	 * the "draining list" and then put it back if it can't be drained.
6540 	 */
6541 
6542 	ASSERT((sq->sq_head == NULL) || (flags & SQ_GOAWAY) ||
6543 	    (type & SQ_CI) || sq->sq_head->q_draining);
6544 
6545 	/* Drop SQ_EXCL for non-CIPUT perimeters */
6546 	if (!(type & SQ_CIPUT))
6547 		flags &= ~SQ_EXCL;
6548 	ASSERT((flags & SQ_EXCL) == 0);
6549 
6550 	/* Wake up any waiters. */
6551 	if (flags & SQ_WANTWAKEUP) {
6552 		flags &= ~SQ_WANTWAKEUP;
6553 		cv_broadcast(&sq->sq_wait);
6554 	}
6555 	if (flags & SQ_WANTEXWAKEUP) {
6556 		flags &= ~SQ_WANTEXWAKEUP;
6557 		cv_broadcast(&sq->sq_exitwait);
6558 	}
6559 	sq->sq_flags = flags;
6560 
6561 	ASSERT(sq->sq_count != 0);
6562 	/* Release our claim. */
6563 	sq->sq_count--;
6564 
6565 	if (bg_service) {
6566 		ASSERT(sq->sq_servcount != 0);
6567 		sq->sq_servcount--;
6568 	}
6569 
6570 	mutex_exit(SQLOCK(sq));
6571 
6572 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6573 	    "drain_syncq end:%p", sq);
6574 }
6575 
6576 
6577 /*
6578  *
6579  * qdrain_syncq can be called (currently) from only one of two places:
6580  *	drain_syncq
6581  * 	putnext  (or some variation of it).
6582  * and eventually
6583  * 	qwait(_sig)
6584  *
6585  * If called from drain_syncq, we found it in the list of queues needing
6586  * service, so there is work to be done (or it wouldn't be in the list).
6587  *
6588  * If called from some putnext variation, it was because the
6589  * perimeter is open, but messages are blocking a putnext and
6590  * there is not a thread working on it.  Now a thread could start
6591  * working on it while we are getting ready to do so ourself, but
6592  * the thread would set the q_draining flag, and we can spin out.
6593  *
6594  * As for qwait(_sig), I think I shall let it continue to call
6595  * drain_syncq directly (after all, it will get here eventually).
6596  *
6597  * qdrain_syncq has to terminate when:
6598  * - one of the SQ_STAYAWAY bits gets set to preserve qwriter(OUTER) ordering
6599  * - SQ_EVENTS gets set to preserve qwriter(INNER) ordering
6600  *
6601  * ASSUMES:
6602  *	One claim
6603  * 	QLOCK held
6604  * 	SQLOCK not held
6605  *	Will release QLOCK before returning
6606  */
6607 void
6608 qdrain_syncq(syncq_t *sq, queue_t *q)
6609 {
6610 	mblk_t		*bp;
6611 #ifdef DEBUG
6612 	uint16_t	count;
6613 #endif
6614 
6615 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_START,
6616 	    "drain_syncq start:%p", sq);
6617 	ASSERT(q->q_syncq == sq);
6618 	ASSERT(MUTEX_HELD(QLOCK(q)));
6619 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6620 	/*
6621 	 * For non-CIPUT perimeters, we should be called with the exclusive bit
6622 	 * set already. For CIPUT perimeters, we will be doing a concurrent
6623 	 * drain, so it better not be set.
6624 	 */
6625 	ASSERT((sq->sq_flags & (SQ_EXCL|SQ_CIPUT)));
6626 	ASSERT(!((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)));
6627 	ASSERT((sq->sq_type & SQ_CIPUT) || (sq->sq_flags & SQ_EXCL));
6628 	/*
6629 	 * All outer pointers are set, or none of them are
6630 	 */
6631 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6632 	    sq->sq_oprev == NULL) ||
6633 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6634 	    sq->sq_oprev != NULL));
6635 #ifdef DEBUG
6636 	count = sq->sq_count;
6637 	/*
6638 	 * This is OK without the putlocks, because we have one
6639 	 * claim either from the sq_count, or a putcount.  We could
6640 	 * get an erroneous value from other counts, but ours won't
6641 	 * change, so one way or another, we will have at least a
6642 	 * value of one.
6643 	 */
6644 	SUM_SQ_PUTCOUNTS(sq, count);
6645 	ASSERT(count >= 1);
6646 #endif /* DEBUG */
6647 
6648 	/*
6649 	 * The first thing to do is find out if a thread is already draining
6650 	 * this queue. If so, we are done, just return.
6651 	 */
6652 	if (q->q_draining) {
6653 		mutex_exit(QLOCK(q));
6654 		return;
6655 	}
6656 
6657 	/*
6658 	 * If the perimeter is exclusive, there is nothing we can do right now,
6659 	 * go away. Note that there is nothing to prevent this case from
6660 	 * changing right after this check, but the spin-out will catch it.
6661 	 */
6662 
6663 	/* Tell other threads that we are draining this queue */
6664 	q->q_draining = 1;	/* Protected by QLOCK */
6665 
6666 	/*
6667 	 * If there is nothing to do, clear QFULL as necessary. This caters for
6668 	 * the case where an empty queue was enqueued onto the syncq.
6669 	 */
6670 	if (q->q_sqhead == NULL) {
6671 		ASSERT(q->q_syncqmsgs == 0);
6672 		mutex_exit(QLOCK(q));
6673 		clr_qfull(q);
6674 		mutex_enter(QLOCK(q));
6675 	}
6676 
6677 	/*
6678 	 * Note that q_sqhead must be re-checked here in case another message
6679 	 * was enqueued whilst QLOCK was dropped during the call to clr_qfull.
6680 	 */
6681 	for (bp = q->q_sqhead; bp != NULL; bp = q->q_sqhead) {
6682 		/*
6683 		 * Because we can enter this routine just because a putnext is
6684 		 * blocked, we need to spin out if the perimeter wants to go
6685 		 * exclusive as well as just blocked. We need to spin out also
6686 		 * if events are queued on the syncq.
6687 		 * Don't check for SQ_EXCL, because non-CIPUT perimeters would
6688 		 * set it, and it can't become exclusive while we hold a claim.
6689 		 */
6690 		if (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)) {
6691 			break;
6692 		}
6693 
6694 #ifdef DEBUG
6695 		/*
6696 		 * Since we are in qdrain_syncq, we already know the queue,
6697 		 * but for sanity, we want to check this against the qp that
6698 		 * was passed in by bp->b_queue.
6699 		 */
6700 
6701 		ASSERT(bp->b_queue == q);
6702 		ASSERT(bp->b_queue->q_syncq == sq);
6703 		bp->b_queue = NULL;
6704 
6705 		/*
6706 		 * We would have the following check in the DEBUG code:
6707 		 *
6708 		 * if (bp->b_prev != NULL)  {
6709 		 *	ASSERT(bp->b_prev == (void (*)())q->q_qinfo->qi_putp);
6710 		 * }
6711 		 *
6712 		 * This can't be done, however, since IP modifies qinfo
6713 		 * structure at run-time (switching between IPv4 qinfo and IPv6
6714 		 * qinfo), invalidating the check.
6715 		 * So the assignment to func is left here, but the ASSERT itself
6716 		 * is removed until the whole issue is resolved.
6717 		 */
6718 #endif
6719 		ASSERT(q->q_sqhead == bp);
6720 		q->q_sqhead = bp->b_next;
6721 		bp->b_prev = bp->b_next = NULL;
6722 		ASSERT(q->q_syncqmsgs > 0);
6723 		mutex_exit(QLOCK(q));
6724 
6725 		ASSERT(bp->b_datap->db_ref != 0);
6726 
6727 		(void) (*q->q_qinfo->qi_putp)(q, bp);
6728 
6729 		mutex_enter(QLOCK(q));
6730 
6731 		/*
6732 		 * q_syncqmsgs should only be decremented after executing the
6733 		 * put procedure to avoid message re-ordering. This is due to an
6734 		 * optimisation in putnext() which can call the put procedure
6735 		 * directly if it sees q_syncqmsgs == 0 (despite Q_SQQUEUED
6736 		 * being set).
6737 		 *
6738 		 * We also need to clear QFULL in the next service procedure
6739 		 * queue if this is the last message destined for that queue.
6740 		 *
6741 		 * It would make better sense to have some sort of tunable for
6742 		 * the low water mark, but these semantics are not yet defined.
6743 		 * So, alas, we use a constant.
6744 		 */
6745 		if (--q->q_syncqmsgs == 0) {
6746 			mutex_exit(QLOCK(q));
6747 			clr_qfull(q);
6748 			mutex_enter(QLOCK(q));
6749 		}
6750 
6751 		/*
6752 		 * Always clear SQ_EXCL when CIPUT in order to handle
6753 		 * qwriter(INNER). The putp() can call qwriter and get exclusive
6754 		 * access IFF this is the only claim. So, we need to test for
6755 		 * this possibility, acquire the mutex and clear the bit.
6756 		 */
6757 		if ((sq->sq_type & SQ_CIPUT) && (sq->sq_flags & SQ_EXCL)) {
6758 			mutex_enter(SQLOCK(sq));
6759 			sq->sq_flags &= ~SQ_EXCL;
6760 			mutex_exit(SQLOCK(sq));
6761 		}
6762 	}
6763 
6764 	/*
6765 	 * We should either have no messages on this queue, or we were told to
6766 	 * goaway by a waiter (which we will wake up at the end of this
6767 	 * function).
6768 	 */
6769 	ASSERT((q->q_sqhead == NULL) ||
6770 	    (sq->sq_flags & (SQ_STAYAWAY | SQ_EVENTS)));
6771 
6772 	ASSERT(MUTEX_HELD(QLOCK(q)));
6773 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6774 
6775 	/* Remove the q from the syncq list if all the messages are drained. */
6776 	if (q->q_sqhead == NULL) {
6777 		ASSERT(q->q_syncqmsgs == 0);
6778 		mutex_enter(SQLOCK(sq));
6779 		if (q->q_sqflags & Q_SQQUEUED)
6780 			SQRM_Q(sq, q);
6781 		mutex_exit(SQLOCK(sq));
6782 		/*
6783 		 * Since the queue is removed from the list, reset its priority.
6784 		 */
6785 		q->q_spri = 0;
6786 	}
6787 
6788 	/*
6789 	 * Remember, the q_draining flag is used to let another thread know
6790 	 * that there is a thread currently draining the messages for a queue.
6791 	 * Since we are now done with this queue (even if there may be messages
6792 	 * still there), we need to clear this flag so some thread will work on
6793 	 * it if needed.
6794 	 */
6795 	ASSERT(q->q_draining);
6796 	q->q_draining = 0;
6797 
6798 	/* Called with a claim, so OK to drop all locks. */
6799 	mutex_exit(QLOCK(q));
6800 
6801 	TRACE_1(TR_FAC_STREAMS_FR, TR_DRAIN_SYNCQ_END,
6802 	    "drain_syncq end:%p", sq);
6803 }
6804 /* END OF QDRAIN_SYNCQ  */
6805 
6806 
6807 /*
6808  * This is the mate to qdrain_syncq, except that it is putting the message onto
6809  * the queue instead of draining. Since the message is destined for the queue
6810  * that is selected, there is no need to identify the function because the
6811  * message is intended for the put routine for the queue. For debug kernels,
6812  * this routine will do it anyway just in case.
6813  *
6814  * After the message is enqueued on the syncq, it calls putnext_tail()
6815  * which will schedule a background thread to actually process the message.
6816  *
6817  * Assumes that there is a claim on the syncq (sq->sq_count > 0) and
6818  * SQLOCK(sq) and QLOCK(q) are not held.
6819  */
6820 void
6821 qfill_syncq(syncq_t *sq, queue_t *q, mblk_t *mp)
6822 {
6823 	ASSERT(MUTEX_NOT_HELD(SQLOCK(sq)));
6824 	ASSERT(MUTEX_NOT_HELD(QLOCK(q)));
6825 	ASSERT(sq->sq_count > 0);
6826 	ASSERT(q->q_syncq == sq);
6827 	ASSERT((sq->sq_outer == NULL && sq->sq_onext == NULL &&
6828 	    sq->sq_oprev == NULL) ||
6829 	    (sq->sq_outer != NULL && sq->sq_onext != NULL &&
6830 	    sq->sq_oprev != NULL));
6831 
6832 	mutex_enter(QLOCK(q));
6833 
6834 #ifdef DEBUG
6835 	/*
6836 	 * This is used for debug in the qfill_syncq/qdrain_syncq case
6837 	 * to trace the queue that the message is intended for.  Note
6838 	 * that the original use was to identify the queue and function
6839 	 * to call on the drain.  In the new syncq, we have the context
6840 	 * of the queue that we are draining, so call it's putproc and
6841 	 * don't rely on the saved values.  But for debug this is still
6842 	 * useful information.
6843 	 */
6844 	mp->b_prev = (mblk_t *)q->q_qinfo->qi_putp;
6845 	mp->b_queue = q;
6846 	mp->b_next = NULL;
6847 #endif
6848 	ASSERT(q->q_syncq == sq);
6849 	/*
6850 	 * Enqueue the message on the list.
6851 	 * SQPUT_MP() accesses q_syncqmsgs.  We are already holding QLOCK to
6852 	 * protect it.  So it's ok to acquire SQLOCK after SQPUT_MP().
6853 	 */
6854 	SQPUT_MP(q, mp);
6855 	mutex_enter(SQLOCK(sq));
6856 
6857 	/*
6858 	 * And queue on syncq for scheduling, if not already queued.
6859 	 * Note that we need the SQLOCK for this, and for testing flags
6860 	 * at the end to see if we will drain.  So grab it now, and
6861 	 * release it before we call qdrain_syncq or return.
6862 	 */
6863 	if (!(q->q_sqflags & Q_SQQUEUED)) {
6864 		q->q_spri = curthread->t_pri;
6865 		SQPUT_Q(sq, q);
6866 	}
6867 #ifdef DEBUG
6868 	else {
6869 		/*
6870 		 * All of these conditions MUST be true!
6871 		 */
6872 		ASSERT(sq->sq_tail != NULL);
6873 		if (sq->sq_tail == sq->sq_head) {
6874 			ASSERT((q->q_sqprev == NULL) &&
6875 			    (q->q_sqnext == NULL));
6876 		} else {
6877 			ASSERT((q->q_sqprev != NULL) ||
6878 			    (q->q_sqnext != NULL));
6879 		}
6880 		ASSERT(sq->sq_flags & SQ_QUEUED);
6881 		ASSERT(q->q_syncqmsgs != 0);
6882 		ASSERT(q->q_sqflags & Q_SQQUEUED);
6883 	}
6884 #endif
6885 	mutex_exit(QLOCK(q));
6886 	/*
6887 	 * SQLOCK is still held, so sq_count can be safely decremented.
6888 	 */
6889 	sq->sq_count--;
6890 
6891 	putnext_tail(sq, q, 0);
6892 	/* Should not reference sq or q after this point. */
6893 }
6894 
6895 /*  End of qfill_syncq  */
6896 
6897 /*
6898  * Remove all messages from a syncq (if qp is NULL) or remove all messages
6899  * that would be put into qp by drain_syncq.
6900  * Used when deleting the syncq (qp == NULL) or when detaching
6901  * a queue (qp != NULL).
6902  * Return non-zero if one or more messages were freed.
6903  *
6904  * No need to grab sq_putlocks here. See comment in strsubr.h that explains when
6905  * sq_putlocks are used.
6906  *
6907  * NOTE: This function assumes that it is called from the close() context and
6908  * that all the queues in the syncq are going away. For this reason it doesn't
6909  * acquire QLOCK for modifying q_sqhead/q_sqtail fields. This assumption is
6910  * currently valid, but it is useful to rethink this function to behave properly
6911  * in other cases.
6912  */
6913 int
6914 flush_syncq(syncq_t *sq, queue_t *qp)
6915 {
6916 	mblk_t		*bp, *mp_head, *mp_next, *mp_prev;
6917 	queue_t		*q;
6918 	int		ret = 0;
6919 
6920 	mutex_enter(SQLOCK(sq));
6921 
6922 	/*
6923 	 * Before we leave, we need to make sure there are no
6924 	 * events listed for this queue.  All events for this queue
6925 	 * will just be freed.
6926 	 */
6927 	if (qp != NULL && sq->sq_evhead != NULL) {
6928 		ASSERT(sq->sq_flags & SQ_EVENTS);
6929 
6930 		mp_prev = NULL;
6931 		for (bp = sq->sq_evhead; bp != NULL; bp = mp_next) {
6932 			mp_next = bp->b_next;
6933 			if (bp->b_queue == qp) {
6934 				/* Delete this message */
6935 				if (mp_prev != NULL) {
6936 					mp_prev->b_next = mp_next;
6937 					/*
6938 					 * Update sq_evtail if the last element
6939 					 * is removed.
6940 					 */
6941 					if (bp == sq->sq_evtail) {
6942 						ASSERT(mp_next == NULL);
6943 						sq->sq_evtail = mp_prev;
6944 					}
6945 				} else
6946 					sq->sq_evhead = mp_next;
6947 				if (sq->sq_evhead == NULL)
6948 					sq->sq_flags &= ~SQ_EVENTS;
6949 				bp->b_prev = bp->b_next = NULL;
6950 				freemsg(bp);
6951 				ret++;
6952 			} else {
6953 				mp_prev = bp;
6954 			}
6955 		}
6956 	}
6957 
6958 	/*
6959 	 * Walk sq_head and:
6960 	 *	- match qp if qp is set, remove it's messages
6961 	 *	- all if qp is not set
6962 	 */
6963 	q = sq->sq_head;
6964 	while (q != NULL) {
6965 		ASSERT(q->q_syncq == sq);
6966 		if ((qp == NULL) || (qp == q)) {
6967 			/*
6968 			 * Yank the messages as a list off the queue
6969 			 */
6970 			mp_head = q->q_sqhead;
6971 			/*
6972 			 * We do not have QLOCK(q) here (which is safe due to
6973 			 * assumptions mentioned above). To obtain the lock we
6974 			 * need to release SQLOCK which may allow lots of things
6975 			 * to change upon us. This place requires more analysis.
6976 			 */
6977 			q->q_sqhead = q->q_sqtail = NULL;
6978 			ASSERT(mp_head->b_queue &&
6979 			    mp_head->b_queue->q_syncq == sq);
6980 
6981 			/*
6982 			 * Free each of the messages.
6983 			 */
6984 			for (bp = mp_head; bp != NULL; bp = mp_next) {
6985 				mp_next = bp->b_next;
6986 				bp->b_prev = bp->b_next = NULL;
6987 				freemsg(bp);
6988 				ret++;
6989 			}
6990 			/*
6991 			 * Now remove the queue from the syncq.
6992 			 */
6993 			ASSERT(q->q_sqflags & Q_SQQUEUED);
6994 			SQRM_Q(sq, q);
6995 			q->q_spri = 0;
6996 			q->q_syncqmsgs = 0;
6997 
6998 			/*
6999 			 * If qp was specified, we are done with it and are
7000 			 * going to drop SQLOCK(sq) and return. We wakeup syncq
7001 			 * waiters while we still have the SQLOCK.
7002 			 */
7003 			if ((qp != NULL) && (sq->sq_flags & SQ_WANTWAKEUP)) {
7004 				sq->sq_flags &= ~SQ_WANTWAKEUP;
7005 				cv_broadcast(&sq->sq_wait);
7006 			}
7007 			/* Drop SQLOCK across clr_qfull */
7008 			mutex_exit(SQLOCK(sq));
7009 
7010 			/*
7011 			 * We avoid doing the test that drain_syncq does and
7012 			 * unconditionally clear qfull for every flushed
7013 			 * message. Since flush_syncq is only called during
7014 			 * close this should not be a problem.
7015 			 */
7016 			clr_qfull(q);
7017 			if (qp != NULL) {
7018 				return (ret);
7019 			} else {
7020 				mutex_enter(SQLOCK(sq));
7021 				/*
7022 				 * The head was removed by SQRM_Q above.
7023 				 * reread the new head and flush it.
7024 				 */
7025 				q = sq->sq_head;
7026 			}
7027 		} else {
7028 			q = q->q_sqnext;
7029 		}
7030 		ASSERT(MUTEX_HELD(SQLOCK(sq)));
7031 	}
7032 
7033 	if (sq->sq_flags & SQ_WANTWAKEUP) {
7034 		sq->sq_flags &= ~SQ_WANTWAKEUP;
7035 		cv_broadcast(&sq->sq_wait);
7036 	}
7037 
7038 	mutex_exit(SQLOCK(sq));
7039 	return (ret);
7040 }
7041 
7042 /*
7043  * Propagate all messages from a syncq to the next syncq that are associated
7044  * with the specified queue. If the queue is attached to a driver or if the
7045  * messages have been added due to a qwriter(PERIM_INNER), free the messages.
7046  *
7047  * Assumes that the stream is strlock()'ed. We don't come here if there
7048  * are no messages to propagate.
7049  *
7050  * NOTE : If the queue is attached to a driver, all the messages are freed
7051  * as there is no point in propagating the messages from the driver syncq
7052  * to the closing stream head which will in turn get freed later.
7053  */
7054 static int
7055 propagate_syncq(queue_t *qp)
7056 {
7057 	mblk_t		*bp, *head, *tail, *prev, *next;
7058 	syncq_t 	*sq;
7059 	queue_t		*nqp;
7060 	syncq_t		*nsq;
7061 	boolean_t	isdriver;
7062 	int 		moved = 0;
7063 	uint16_t	flags;
7064 	pri_t		priority = curthread->t_pri;
7065 #ifdef DEBUG
7066 	void		(*func)();
7067 #endif
7068 
7069 	sq = qp->q_syncq;
7070 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7071 	/* debug macro */
7072 	SQ_PUTLOCKS_HELD(sq);
7073 	/*
7074 	 * As entersq() does not increment the sq_count for
7075 	 * the write side, check sq_count for non-QPERQ
7076 	 * perimeters alone.
7077 	 */
7078 	ASSERT((qp->q_flag & QPERQ) || (sq->sq_count >= 1));
7079 
7080 	/*
7081 	 * propagate_syncq() can be called because of either messages on the
7082 	 * queue syncq or because on events on the queue syncq. Do actual
7083 	 * message propagations if there are any messages.
7084 	 */
7085 	if (qp->q_syncqmsgs) {
7086 		isdriver = (qp->q_flag & QISDRV);
7087 
7088 		if (!isdriver) {
7089 			nqp = qp->q_next;
7090 			nsq = nqp->q_syncq;
7091 			ASSERT(MUTEX_HELD(SQLOCK(nsq)));
7092 			/* debug macro */
7093 			SQ_PUTLOCKS_HELD(nsq);
7094 #ifdef DEBUG
7095 			func = (void (*)())nqp->q_qinfo->qi_putp;
7096 #endif
7097 		}
7098 
7099 		SQRM_Q(sq, qp);
7100 		priority = MAX(qp->q_spri, priority);
7101 		qp->q_spri = 0;
7102 		head = qp->q_sqhead;
7103 		tail = qp->q_sqtail;
7104 		qp->q_sqhead = qp->q_sqtail = NULL;
7105 		qp->q_syncqmsgs = 0;
7106 
7107 		/*
7108 		 * Walk the list of messages, and free them if this is a driver,
7109 		 * otherwise reset the b_prev and b_queue value to the new putp.
7110 		 * Afterward, we will just add the head to the end of the next
7111 		 * syncq, and point the tail to the end of this one.
7112 		 */
7113 
7114 		for (bp = head; bp != NULL; bp = next) {
7115 			next = bp->b_next;
7116 			if (isdriver) {
7117 				bp->b_prev = bp->b_next = NULL;
7118 				freemsg(bp);
7119 				continue;
7120 			}
7121 			/* Change the q values for this message */
7122 			bp->b_queue = nqp;
7123 #ifdef DEBUG
7124 			bp->b_prev = (mblk_t *)func;
7125 #endif
7126 			moved++;
7127 		}
7128 		/*
7129 		 * Attach list of messages to the end of the new queue (if there
7130 		 * is a list of messages).
7131 		 */
7132 
7133 		if (!isdriver && head != NULL) {
7134 			ASSERT(tail != NULL);
7135 			if (nqp->q_sqhead == NULL) {
7136 				nqp->q_sqhead = head;
7137 			} else {
7138 				ASSERT(nqp->q_sqtail != NULL);
7139 				nqp->q_sqtail->b_next = head;
7140 			}
7141 			nqp->q_sqtail = tail;
7142 			/*
7143 			 * When messages are moved from high priority queue to
7144 			 * another queue, the destination queue priority is
7145 			 * upgraded.
7146 			 */
7147 
7148 			if (priority > nqp->q_spri)
7149 				nqp->q_spri = priority;
7150 
7151 			SQPUT_Q(nsq, nqp);
7152 
7153 			nqp->q_syncqmsgs += moved;
7154 			ASSERT(nqp->q_syncqmsgs != 0);
7155 		}
7156 	}
7157 
7158 	/*
7159 	 * Before we leave, we need to make sure there are no
7160 	 * events listed for this queue.  All events for this queue
7161 	 * will just be freed.
7162 	 */
7163 	if (sq->sq_evhead != NULL) {
7164 		ASSERT(sq->sq_flags & SQ_EVENTS);
7165 		prev = NULL;
7166 		for (bp = sq->sq_evhead; bp != NULL; bp = next) {
7167 			next = bp->b_next;
7168 			if (bp->b_queue == qp) {
7169 				/* Delete this message */
7170 				if (prev != NULL) {
7171 					prev->b_next = next;
7172 					/*
7173 					 * Update sq_evtail if the last element
7174 					 * is removed.
7175 					 */
7176 					if (bp == sq->sq_evtail) {
7177 						ASSERT(next == NULL);
7178 						sq->sq_evtail = prev;
7179 					}
7180 				} else
7181 					sq->sq_evhead = next;
7182 				if (sq->sq_evhead == NULL)
7183 					sq->sq_flags &= ~SQ_EVENTS;
7184 				bp->b_prev = bp->b_next = NULL;
7185 				freemsg(bp);
7186 			} else {
7187 				prev = bp;
7188 			}
7189 		}
7190 	}
7191 
7192 	flags = sq->sq_flags;
7193 
7194 	/* Wake up any waiter before leaving. */
7195 	if (flags & SQ_WANTWAKEUP) {
7196 		flags &= ~SQ_WANTWAKEUP;
7197 		cv_broadcast(&sq->sq_wait);
7198 	}
7199 	sq->sq_flags = flags;
7200 
7201 	return (moved);
7202 }
7203 
7204 /*
7205  * Try and upgrade to exclusive access at the inner perimeter. If this can
7206  * not be done without blocking then request will be queued on the syncq
7207  * and drain_syncq will run it later.
7208  *
7209  * This routine can only be called from put or service procedures plus
7210  * asynchronous callback routines that have properly entered the queue (with
7211  * entersq). Thus qwriter_inner assumes the caller has one claim on the syncq
7212  * associated with q.
7213  */
7214 void
7215 qwriter_inner(queue_t *q, mblk_t *mp, void (*func)())
7216 {
7217 	syncq_t	*sq = q->q_syncq;
7218 	uint16_t count;
7219 
7220 	mutex_enter(SQLOCK(sq));
7221 	count = sq->sq_count;
7222 	SQ_PUTLOCKS_ENTER(sq);
7223 	SUM_SQ_PUTCOUNTS(sq, count);
7224 	ASSERT(count >= 1);
7225 	ASSERT(sq->sq_type & (SQ_CIPUT|SQ_CISVC));
7226 
7227 	if (count == 1) {
7228 		/*
7229 		 * Can upgrade. This case also handles nested qwriter calls
7230 		 * (when the qwriter callback function calls qwriter). In that
7231 		 * case SQ_EXCL is already set.
7232 		 */
7233 		sq->sq_flags |= SQ_EXCL;
7234 		SQ_PUTLOCKS_EXIT(sq);
7235 		mutex_exit(SQLOCK(sq));
7236 		(*func)(q, mp);
7237 		/*
7238 		 * Assumes that leavesq, putnext, and drain_syncq will reset
7239 		 * SQ_EXCL for SQ_CIPUT/SQ_CISVC queues. We leave SQ_EXCL on
7240 		 * until putnext, leavesq, or drain_syncq drops it.
7241 		 * That way we handle nested qwriter(INNER) without dropping
7242 		 * SQ_EXCL until the outermost qwriter callback routine is
7243 		 * done.
7244 		 */
7245 		return;
7246 	}
7247 	SQ_PUTLOCKS_EXIT(sq);
7248 	sqfill_events(sq, q, mp, func);
7249 }
7250 
7251 /*
7252  * Synchronous callback support functions
7253  */
7254 
7255 /*
7256  * Allocate a callback parameter structure.
7257  * Assumes that caller initializes the flags and the id.
7258  * Acquires SQLOCK(sq) if non-NULL is returned.
7259  */
7260 callbparams_t *
7261 callbparams_alloc(syncq_t *sq, void (*func)(void *), void *arg, int kmflags)
7262 {
7263 	callbparams_t *cbp;
7264 	size_t size = sizeof (callbparams_t);
7265 
7266 	cbp = kmem_alloc(size, kmflags & ~KM_PANIC);
7267 
7268 	/*
7269 	 * Only try tryhard allocation if the caller is ready to panic.
7270 	 * Otherwise just fail.
7271 	 */
7272 	if (cbp == NULL) {
7273 		if (kmflags & KM_PANIC)
7274 			cbp = kmem_alloc_tryhard(sizeof (callbparams_t),
7275 			    &size, kmflags);
7276 		else
7277 			return (NULL);
7278 	}
7279 
7280 	ASSERT(size >= sizeof (callbparams_t));
7281 	cbp->cbp_size = size;
7282 	cbp->cbp_sq = sq;
7283 	cbp->cbp_func = func;
7284 	cbp->cbp_arg = arg;
7285 	mutex_enter(SQLOCK(sq));
7286 	cbp->cbp_next = sq->sq_callbpend;
7287 	sq->sq_callbpend = cbp;
7288 	return (cbp);
7289 }
7290 
7291 void
7292 callbparams_free(syncq_t *sq, callbparams_t *cbp)
7293 {
7294 	callbparams_t **pp, *p;
7295 
7296 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7297 
7298 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7299 		if (p == cbp) {
7300 			*pp = p->cbp_next;
7301 			kmem_free(p, p->cbp_size);
7302 			return;
7303 		}
7304 	}
7305 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7306 	    "callbparams_free: not found\n"));
7307 }
7308 
7309 void
7310 callbparams_free_id(syncq_t *sq, callbparams_id_t id, int32_t flag)
7311 {
7312 	callbparams_t **pp, *p;
7313 
7314 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7315 
7316 	for (pp = &sq->sq_callbpend; (p = *pp) != NULL; pp = &p->cbp_next) {
7317 		if (p->cbp_id == id && p->cbp_flags == flag) {
7318 			*pp = p->cbp_next;
7319 			kmem_free(p, p->cbp_size);
7320 			return;
7321 		}
7322 	}
7323 	(void) (STRLOG(0, 0, 0, SL_CONSOLE,
7324 	    "callbparams_free_id: not found\n"));
7325 }
7326 
7327 /*
7328  * Callback wrapper function used by once-only callbacks that can be
7329  * cancelled (qtimeout and qbufcall)
7330  * Contains inline version of entersq(sq, SQ_CALLBACK) that can be
7331  * cancelled by the qun* functions.
7332  */
7333 void
7334 qcallbwrapper(void *arg)
7335 {
7336 	callbparams_t *cbp = arg;
7337 	syncq_t	*sq;
7338 	uint16_t count = 0;
7339 	uint16_t waitflags = SQ_STAYAWAY | SQ_EVENTS | SQ_EXCL;
7340 	uint16_t type;
7341 
7342 	sq = cbp->cbp_sq;
7343 	mutex_enter(SQLOCK(sq));
7344 	type = sq->sq_type;
7345 	if (!(type & SQ_CICB)) {
7346 		count = sq->sq_count;
7347 		SQ_PUTLOCKS_ENTER(sq);
7348 		SQ_PUTCOUNT_CLRFAST_LOCKED(sq);
7349 		SUM_SQ_PUTCOUNTS(sq, count);
7350 		sq->sq_needexcl++;
7351 		ASSERT(sq->sq_needexcl != 0);	/* wraparound */
7352 		waitflags |= SQ_MESSAGES;
7353 	}
7354 	/* Can not handle exclusive entry at outer perimeter */
7355 	ASSERT(type & SQ_COCB);
7356 
7357 	while ((sq->sq_flags & waitflags) || (!(type & SQ_CICB) &&count != 0)) {
7358 		if ((sq->sq_callbflags & cbp->cbp_flags) &&
7359 		    (sq->sq_cancelid == cbp->cbp_id)) {
7360 			/* timeout has been cancelled */
7361 			sq->sq_callbflags |= SQ_CALLB_BYPASSED;
7362 			callbparams_free(sq, cbp);
7363 			if (!(type & SQ_CICB)) {
7364 				ASSERT(sq->sq_needexcl > 0);
7365 				sq->sq_needexcl--;
7366 				if (sq->sq_needexcl == 0) {
7367 					SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7368 				}
7369 				SQ_PUTLOCKS_EXIT(sq);
7370 			}
7371 			mutex_exit(SQLOCK(sq));
7372 			return;
7373 		}
7374 		sq->sq_flags |= SQ_WANTWAKEUP;
7375 		if (!(type & SQ_CICB)) {
7376 			SQ_PUTLOCKS_EXIT(sq);
7377 		}
7378 		cv_wait(&sq->sq_wait, SQLOCK(sq));
7379 		if (!(type & SQ_CICB)) {
7380 			count = sq->sq_count;
7381 			SQ_PUTLOCKS_ENTER(sq);
7382 			SUM_SQ_PUTCOUNTS(sq, count);
7383 		}
7384 	}
7385 
7386 	sq->sq_count++;
7387 	ASSERT(sq->sq_count != 0);	/* Wraparound */
7388 	if (!(type & SQ_CICB)) {
7389 		ASSERT(count == 0);
7390 		sq->sq_flags |= SQ_EXCL;
7391 		ASSERT(sq->sq_needexcl > 0);
7392 		sq->sq_needexcl--;
7393 		if (sq->sq_needexcl == 0) {
7394 			SQ_PUTCOUNT_SETFAST_LOCKED(sq);
7395 		}
7396 		SQ_PUTLOCKS_EXIT(sq);
7397 	}
7398 
7399 	mutex_exit(SQLOCK(sq));
7400 
7401 	cbp->cbp_func(cbp->cbp_arg);
7402 
7403 	/*
7404 	 * We drop the lock only for leavesq to re-acquire it.
7405 	 * Possible optimization is inline of leavesq.
7406 	 */
7407 	mutex_enter(SQLOCK(sq));
7408 	callbparams_free(sq, cbp);
7409 	mutex_exit(SQLOCK(sq));
7410 	leavesq(sq, SQ_CALLBACK);
7411 }
7412 
7413 /*
7414  * No need to grab sq_putlocks here. See comment in strsubr.h that
7415  * explains when sq_putlocks are used.
7416  *
7417  * sq_count (or one of the sq_putcounts) has already been
7418  * decremented by the caller, and if SQ_QUEUED, we need to call
7419  * drain_syncq (the global syncq drain).
7420  * If putnext_tail is called with the SQ_EXCL bit set, we are in
7421  * one of two states, non-CIPUT perimeter, and we need to clear
7422  * it, or we went exclusive in the put procedure.  In any case,
7423  * we want to clear the bit now, and it is probably easier to do
7424  * this at the beginning of this function (remember, we hold
7425  * the SQLOCK).  Lastly, if there are other messages queued
7426  * on the syncq (and not for our destination), enable the syncq
7427  * for background work.
7428  */
7429 
7430 /* ARGSUSED */
7431 void
7432 putnext_tail(syncq_t *sq, queue_t *qp, uint32_t passflags)
7433 {
7434 	uint16_t	flags = sq->sq_flags;
7435 
7436 	ASSERT(MUTEX_HELD(SQLOCK(sq)));
7437 	ASSERT(MUTEX_NOT_HELD(QLOCK(qp)));
7438 
7439 	/* Clear SQ_EXCL if set in passflags */
7440 	if (passflags & SQ_EXCL) {
7441 		flags &= ~SQ_EXCL;
7442 	}
7443 	if (flags & SQ_WANTWAKEUP) {
7444 		flags &= ~SQ_WANTWAKEUP;
7445 		cv_broadcast(&sq->sq_wait);
7446 	}
7447 	if (flags & SQ_WANTEXWAKEUP) {
7448 		flags &= ~SQ_WANTEXWAKEUP;
7449 		cv_broadcast(&sq->sq_exitwait);
7450 	}
7451 	sq->sq_flags = flags;
7452 
7453 	/*
7454 	 * We have cleared SQ_EXCL if we were asked to, and started
7455 	 * the wakeup process for waiters.  If there are no writers
7456 	 * then we need to drain the syncq if we were told to, or
7457 	 * enable the background thread to do it.
7458 	 */
7459 	if (!(flags & (SQ_STAYAWAY|SQ_EXCL))) {
7460 		if ((passflags & SQ_QUEUED) ||
7461 		    (sq->sq_svcflags & SQ_DISABLED)) {
7462 			/* drain_syncq will take care of events in the list */
7463 			drain_syncq(sq);
7464 			return;
7465 		} else if (flags & SQ_QUEUED) {
7466 			sqenable(sq);
7467 		}
7468 	}
7469 	/* Drop the SQLOCK on exit */
7470 	mutex_exit(SQLOCK(sq));
7471 	TRACE_3(TR_FAC_STREAMS_FR, TR_PUTNEXT_END,
7472 	    "putnext_end:(%p, %p, %p) done", NULL, qp, sq);
7473 }
7474 
7475 void
7476 set_qend(queue_t *q)
7477 {
7478 	mutex_enter(QLOCK(q));
7479 	if (!O_SAMESTR(q))
7480 		q->q_flag |= QEND;
7481 	else
7482 		q->q_flag &= ~QEND;
7483 	mutex_exit(QLOCK(q));
7484 	q = _OTHERQ(q);
7485 	mutex_enter(QLOCK(q));
7486 	if (!O_SAMESTR(q))
7487 		q->q_flag |= QEND;
7488 	else
7489 		q->q_flag &= ~QEND;
7490 	mutex_exit(QLOCK(q));
7491 }
7492 
7493 /*
7494  * Set QFULL in next service procedure queue (that cares) if not already
7495  * set and if there are already more messages on the syncq than
7496  * sq_max_size.  If sq_max_size is 0, no flow control will be asserted on
7497  * any syncq.
7498  *
7499  * The fq here is the next queue with a service procedure.  This is where
7500  * we would fail canputnext, so this is where we need to set QFULL.
7501  * In the case when fq != q we need to take QLOCK(fq) to set QFULL flag.
7502  *
7503  * We already have QLOCK at this point. To avoid cross-locks with
7504  * freezestr() which grabs all QLOCKs and with strlock() which grabs both
7505  * SQLOCK and sd_reflock, we need to drop respective locks first.
7506  */
7507 void
7508 set_qfull(queue_t *q)
7509 {
7510 	queue_t		*fq = NULL;
7511 
7512 	ASSERT(MUTEX_HELD(QLOCK(q)));
7513 	if ((sq_max_size != 0) && (!(q->q_nfsrv->q_flag & QFULL)) &&
7514 	    (q->q_syncqmsgs > sq_max_size)) {
7515 		if ((fq = q->q_nfsrv) == q) {
7516 			fq->q_flag |= QFULL;
7517 		} else {
7518 			mutex_exit(QLOCK(q));
7519 			mutex_enter(QLOCK(fq));
7520 			fq->q_flag |= QFULL;
7521 			mutex_exit(QLOCK(fq));
7522 			mutex_enter(QLOCK(q));
7523 		}
7524 	}
7525 }
7526 
7527 void
7528 clr_qfull(queue_t *q)
7529 {
7530 	queue_t	*oq = q;
7531 
7532 	q = q->q_nfsrv;
7533 	/* Fast check if there is any work to do before getting the lock. */
7534 	if ((q->q_flag & (QFULL|QWANTW)) == 0) {
7535 		return;
7536 	}
7537 
7538 	/*
7539 	 * Do not reset QFULL (and backenable) if the q_count is the reason
7540 	 * for QFULL being set.
7541 	 */
7542 	mutex_enter(QLOCK(q));
7543 	/*
7544 	 * If queue is empty i.e q_mblkcnt is zero, queue can not be full.
7545 	 * Hence clear the QFULL.
7546 	 * If both q_count and q_mblkcnt are less than the hiwat mark,
7547 	 * clear the QFULL.
7548 	 */
7549 	if (q->q_mblkcnt == 0 || ((q->q_count < q->q_hiwat) &&
7550 	    (q->q_mblkcnt < q->q_hiwat))) {
7551 		q->q_flag &= ~QFULL;
7552 		/*
7553 		 * A little more confusing, how about this way:
7554 		 * if someone wants to write,
7555 		 * AND
7556 		 *    both counts are less than the lowat mark
7557 		 *    OR
7558 		 *    the lowat mark is zero
7559 		 * THEN
7560 		 * backenable
7561 		 */
7562 		if ((q->q_flag & QWANTW) &&
7563 		    (((q->q_count < q->q_lowat) &&
7564 		    (q->q_mblkcnt < q->q_lowat)) || q->q_lowat == 0)) {
7565 			q->q_flag &= ~QWANTW;
7566 			mutex_exit(QLOCK(q));
7567 			backenable(oq, 0);
7568 		} else
7569 			mutex_exit(QLOCK(q));
7570 	} else
7571 		mutex_exit(QLOCK(q));
7572 }
7573 
7574 /*
7575  * Set the forward service procedure pointer.
7576  *
7577  * Called at insert-time to cache a queue's next forward service procedure in
7578  * q_nfsrv; used by canput() and canputnext().  If the queue to be inserted
7579  * has a service procedure then q_nfsrv points to itself.  If the queue to be
7580  * inserted does not have a service procedure, then q_nfsrv points to the next
7581  * queue forward that has a service procedure.  If the queue is at the logical
7582  * end of the stream (driver for write side, stream head for the read side)
7583  * and does not have a service procedure, then q_nfsrv also points to itself.
7584  */
7585 void
7586 set_nfsrv_ptr(
7587 	queue_t  *rnew,		/* read queue pointer to new module */
7588 	queue_t  *wnew,		/* write queue pointer to new module */
7589 	queue_t  *prev_rq,	/* read queue pointer to the module above */
7590 	queue_t  *prev_wq)	/* write queue pointer to the module above */
7591 {
7592 	queue_t *qp;
7593 
7594 	if (prev_wq->q_next == NULL) {
7595 		/*
7596 		 * Insert the driver, initialize the driver and stream head.
7597 		 * In this case, prev_rq/prev_wq should be the stream head.
7598 		 * _I_INSERT does not allow inserting a driver.  Make sure
7599 		 * that it is not an insertion.
7600 		 */
7601 		ASSERT(!(rnew->q_flag & _QINSERTING));
7602 		wnew->q_nfsrv = wnew;
7603 		if (rnew->q_qinfo->qi_srvp)
7604 			rnew->q_nfsrv = rnew;
7605 		else
7606 			rnew->q_nfsrv = prev_rq;
7607 		prev_rq->q_nfsrv = prev_rq;
7608 		prev_wq->q_nfsrv = prev_wq;
7609 	} else {
7610 		/*
7611 		 * set up read side q_nfsrv pointer.  This MUST be done
7612 		 * before setting the write side, because the setting of
7613 		 * the write side for a fifo may depend on it.
7614 		 *
7615 		 * Suppose we have a fifo that only has pipemod pushed.
7616 		 * pipemod has no read or write service procedures, so
7617 		 * nfsrv for both pipemod queues points to prev_rq (the
7618 		 * stream read head).  Now push bufmod (which has only a
7619 		 * read service procedure).  Doing the write side first,
7620 		 * wnew->q_nfsrv is set to pipemod's writeq nfsrv, which
7621 		 * is WRONG; the next queue forward from wnew with a
7622 		 * service procedure will be rnew, not the stream read head.
7623 		 * Since the downstream queue (which in the case of a fifo
7624 		 * is the read queue rnew) can affect upstream queues, it
7625 		 * needs to be done first.  Setting up the read side first
7626 		 * sets nfsrv for both pipemod queues to rnew and then
7627 		 * when the write side is set up, wnew-q_nfsrv will also
7628 		 * point to rnew.
7629 		 */
7630 		if (rnew->q_qinfo->qi_srvp) {
7631 			/*
7632 			 * use _OTHERQ() because, if this is a pipe, next
7633 			 * module may have been pushed from other end and
7634 			 * q_next could be a read queue.
7635 			 */
7636 			qp = _OTHERQ(prev_wq->q_next);
7637 			while (qp && qp->q_nfsrv != qp) {
7638 				qp->q_nfsrv = rnew;
7639 				qp = backq(qp);
7640 			}
7641 			rnew->q_nfsrv = rnew;
7642 		} else
7643 			rnew->q_nfsrv = prev_rq->q_nfsrv;
7644 
7645 		/* set up write side q_nfsrv pointer */
7646 		if (wnew->q_qinfo->qi_srvp) {
7647 			wnew->q_nfsrv = wnew;
7648 
7649 			/*
7650 			 * For insertion, need to update nfsrv of the modules
7651 			 * above which do not have a service routine.
7652 			 */
7653 			if (rnew->q_flag & _QINSERTING) {
7654 				for (qp = prev_wq;
7655 				    qp != NULL && qp->q_nfsrv != qp;
7656 				    qp = backq(qp)) {
7657 					qp->q_nfsrv = wnew->q_nfsrv;
7658 				}
7659 			}
7660 		} else {
7661 			if (prev_wq->q_next == prev_rq)
7662 				/*
7663 				 * Since prev_wq/prev_rq are the middle of a
7664 				 * fifo, wnew/rnew will also be the middle of
7665 				 * a fifo and wnew's nfsrv is same as rnew's.
7666 				 */
7667 				wnew->q_nfsrv = rnew->q_nfsrv;
7668 			else
7669 				wnew->q_nfsrv = prev_wq->q_next->q_nfsrv;
7670 		}
7671 	}
7672 }
7673 
7674 /*
7675  * Reset the forward service procedure pointer; called at remove-time.
7676  */
7677 void
7678 reset_nfsrv_ptr(queue_t *rqp, queue_t *wqp)
7679 {
7680 	queue_t *tmp_qp;
7681 
7682 	/* Reset the write side q_nfsrv pointer for _I_REMOVE */
7683 	if ((rqp->q_flag & _QREMOVING) && (wqp->q_qinfo->qi_srvp != NULL)) {
7684 		for (tmp_qp = backq(wqp);
7685 		    tmp_qp != NULL && tmp_qp->q_nfsrv == wqp;
7686 		    tmp_qp = backq(tmp_qp)) {
7687 			tmp_qp->q_nfsrv = wqp->q_nfsrv;
7688 		}
7689 	}
7690 
7691 	/* reset the read side q_nfsrv pointer */
7692 	if (rqp->q_qinfo->qi_srvp) {
7693 		if (wqp->q_next) {	/* non-driver case */
7694 			tmp_qp = _OTHERQ(wqp->q_next);
7695 			while (tmp_qp && tmp_qp->q_nfsrv == rqp) {
7696 				/* Note that rqp->q_next cannot be NULL */
7697 				ASSERT(rqp->q_next != NULL);
7698 				tmp_qp->q_nfsrv = rqp->q_next->q_nfsrv;
7699 				tmp_qp = backq(tmp_qp);
7700 			}
7701 		}
7702 	}
7703 }
7704 
7705 /*
7706  * This routine should be called after all stream geometry changes to update
7707  * the stream head cached struio() rd/wr queue pointers. Note must be called
7708  * with the streamlock()ed.
7709  *
7710  * Note: only enables Synchronous STREAMS for a side of a Stream which has
7711  *	 an explicit synchronous barrier module queue. That is, a queue that
7712  *	 has specified a struio() type.
7713  */
7714 static void
7715 strsetuio(stdata_t *stp)
7716 {
7717 	queue_t *wrq;
7718 
7719 	if (stp->sd_flag & STPLEX) {
7720 		/*
7721 		 * Not streamhead, but a mux, so no Synchronous STREAMS.
7722 		 */
7723 		stp->sd_struiowrq = NULL;
7724 		stp->sd_struiordq = NULL;
7725 		return;
7726 	}
7727 	/*
7728 	 * Scan the write queue(s) while synchronous
7729 	 * until we find a qinfo uio type specified.
7730 	 */
7731 	wrq = stp->sd_wrq->q_next;
7732 	while (wrq) {
7733 		if (wrq->q_struiot == STRUIOT_NONE) {
7734 			wrq = 0;
7735 			break;
7736 		}
7737 		if (wrq->q_struiot != STRUIOT_DONTCARE)
7738 			break;
7739 		if (! _SAMESTR(wrq)) {
7740 			wrq = 0;
7741 			break;
7742 		}
7743 		wrq = wrq->q_next;
7744 	}
7745 	stp->sd_struiowrq = wrq;
7746 	/*
7747 	 * Scan the read queue(s) while synchronous
7748 	 * until we find a qinfo uio type specified.
7749 	 */
7750 	wrq = stp->sd_wrq->q_next;
7751 	while (wrq) {
7752 		if (_RD(wrq)->q_struiot == STRUIOT_NONE) {
7753 			wrq = 0;
7754 			break;
7755 		}
7756 		if (_RD(wrq)->q_struiot != STRUIOT_DONTCARE)
7757 			break;
7758 		if (! _SAMESTR(wrq)) {
7759 			wrq = 0;
7760 			break;
7761 		}
7762 		wrq = wrq->q_next;
7763 	}
7764 	stp->sd_struiordq = wrq ? _RD(wrq) : 0;
7765 }
7766 
7767 /*
7768  * pass_wput, unblocks the passthru queues, so that
7769  * messages can arrive at muxs lower read queue, before
7770  * I_LINK/I_UNLINK is acked/nacked.
7771  */
7772 static void
7773 pass_wput(queue_t *q, mblk_t *mp)
7774 {
7775 	syncq_t *sq;
7776 
7777 	sq = _RD(q)->q_syncq;
7778 	if (sq->sq_flags & SQ_BLOCKED)
7779 		unblocksq(sq, SQ_BLOCKED, 0);
7780 	putnext(q, mp);
7781 }
7782 
7783 /*
7784  * Set up queues for the link/unlink.
7785  * Create a new queue and block it and then insert it
7786  * below the stream head on the lower stream.
7787  * This prevents any messages from arriving during the setq
7788  * as well as while the mux is processing the LINK/I_UNLINK.
7789  * The blocked passq is unblocked once the LINK/I_UNLINK has
7790  * been acked or nacked or if a message is generated and sent
7791  * down muxs write put procedure.
7792  * See pass_wput().
7793  *
7794  * After the new queue is inserted, all messages coming from below are
7795  * blocked. The call to strlock will ensure that all activity in the stream head
7796  * read queue syncq is stopped (sq_count drops to zero).
7797  */
7798 static queue_t *
7799 link_addpassthru(stdata_t *stpdown)
7800 {
7801 	queue_t *passq;
7802 	sqlist_t sqlist;
7803 
7804 	passq = allocq();
7805 	STREAM(passq) = STREAM(_WR(passq)) = stpdown;
7806 	/* setq might sleep in allocator - avoid holding locks. */
7807 	setq(passq, &passthru_rinit, &passthru_winit, NULL, QPERQ,
7808 	    SQ_CI|SQ_CO, B_FALSE);
7809 	claimq(passq);
7810 	blocksq(passq->q_syncq, SQ_BLOCKED, 1);
7811 	insertq(STREAM(passq), passq);
7812 
7813 	/*
7814 	 * Use strlock() to wait for the stream head sq_count to drop to zero
7815 	 * since we are going to change q_ptr in the stream head.  Note that
7816 	 * insertq() doesn't wait for any syncq counts to drop to zero.
7817 	 */
7818 	sqlist.sqlist_head = NULL;
7819 	sqlist.sqlist_index = 0;
7820 	sqlist.sqlist_size = sizeof (sqlist_t);
7821 	sqlist_insert(&sqlist, _RD(stpdown->sd_wrq)->q_syncq);
7822 	strlock(stpdown, &sqlist);
7823 	strunlock(stpdown, &sqlist);
7824 
7825 	releaseq(passq);
7826 	return (passq);
7827 }
7828 
7829 /*
7830  * Let messages flow up into the mux by removing
7831  * the passq.
7832  */
7833 static void
7834 link_rempassthru(queue_t *passq)
7835 {
7836 	claimq(passq);
7837 	removeq(passq);
7838 	releaseq(passq);
7839 	freeq(passq);
7840 }
7841 
7842 /*
7843  * Wait for the condition variable pointed to by `cvp' to be signaled,
7844  * or for `tim' milliseconds to elapse, whichever comes first.  If `tim'
7845  * is negative, then there is no time limit.  If `nosigs' is non-zero,
7846  * then the wait will be non-interruptible.
7847  *
7848  * Returns >0 if signaled, 0 if interrupted, or -1 upon timeout.
7849  */
7850 clock_t
7851 str_cv_wait(kcondvar_t *cvp, kmutex_t *mp, clock_t tim, int nosigs)
7852 {
7853 	clock_t ret;
7854 
7855 	if (tim < 0) {
7856 		if (nosigs) {
7857 			cv_wait(cvp, mp);
7858 			ret = 1;
7859 		} else {
7860 			ret = cv_wait_sig(cvp, mp);
7861 		}
7862 	} else if (tim > 0) {
7863 		/*
7864 		 * convert milliseconds to clock ticks
7865 		 */
7866 		if (nosigs) {
7867 			ret = cv_reltimedwait(cvp, mp,
7868 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7869 		} else {
7870 			ret = cv_reltimedwait_sig(cvp, mp,
7871 			    MSEC_TO_TICK_ROUNDUP(tim), TR_CLOCK_TICK);
7872 		}
7873 	} else {
7874 		ret = -1;
7875 	}
7876 	return (ret);
7877 }
7878 
7879 /*
7880  * Wait until the stream head can determine if it is at the mark but
7881  * don't wait forever to prevent a race condition between the "mark" state
7882  * in the stream head and any mark state in the caller/user of this routine.
7883  *
7884  * This is used by sockets and for a socket it would be incorrect
7885  * to return a failure for SIOCATMARK when there is no data in the receive
7886  * queue and the marked urgent data is traveling up the stream.
7887  *
7888  * This routine waits until the mark is known by waiting for one of these
7889  * three events:
7890  *	The stream head read queue becoming non-empty (including an EOF).
7891  *	The STRATMARK flag being set (due to a MSGMARKNEXT message).
7892  *	The STRNOTATMARK flag being set (which indicates that the transport
7893  *	has sent a MSGNOTMARKNEXT message to indicate that it is not at
7894  *	the mark).
7895  *
7896  * The routine returns 1 if the stream is at the mark; 0 if it can
7897  * be determined that the stream is not at the mark.
7898  * If the wait times out and it can't determine
7899  * whether or not the stream might be at the mark the routine will return -1.
7900  *
7901  * Note: This routine should only be used when a mark is pending i.e.,
7902  * in the socket case the SIGURG has been posted.
7903  * Note2: This can not wakeup just because synchronous streams indicate
7904  * that data is available since it is not possible to use the synchronous
7905  * streams interfaces to determine the b_flag value for the data queued below
7906  * the stream head.
7907  */
7908 int
7909 strwaitmark(vnode_t *vp)
7910 {
7911 	struct stdata *stp = vp->v_stream;
7912 	queue_t *rq = _RD(stp->sd_wrq);
7913 	int mark;
7914 
7915 	mutex_enter(&stp->sd_lock);
7916 	while (rq->q_first == NULL &&
7917 	    !(stp->sd_flag & (STRATMARK|STRNOTATMARK|STREOF))) {
7918 		stp->sd_flag |= RSLEEP;
7919 
7920 		/* Wait for 100 milliseconds for any state change. */
7921 		if (str_cv_wait(&rq->q_wait, &stp->sd_lock, 100, 1) == -1) {
7922 			mutex_exit(&stp->sd_lock);
7923 			return (-1);
7924 		}
7925 	}
7926 	if (stp->sd_flag & STRATMARK)
7927 		mark = 1;
7928 	else if (rq->q_first != NULL && (rq->q_first->b_flag & MSGMARK))
7929 		mark = 1;
7930 	else
7931 		mark = 0;
7932 
7933 	mutex_exit(&stp->sd_lock);
7934 	return (mark);
7935 }
7936 
7937 /*
7938  * Set a read side error. If persist is set change the socket error
7939  * to persistent. If errfunc is set install the function as the exported
7940  * error handler.
7941  */
7942 void
7943 strsetrerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7944 {
7945 	struct stdata *stp = vp->v_stream;
7946 
7947 	mutex_enter(&stp->sd_lock);
7948 	stp->sd_rerror = error;
7949 	if (error == 0 && errfunc == NULL)
7950 		stp->sd_flag &= ~STRDERR;
7951 	else
7952 		stp->sd_flag |= STRDERR;
7953 	if (persist) {
7954 		stp->sd_flag &= ~STRDERRNONPERSIST;
7955 	} else {
7956 		stp->sd_flag |= STRDERRNONPERSIST;
7957 	}
7958 	stp->sd_rderrfunc = errfunc;
7959 	if (error != 0 || errfunc != NULL) {
7960 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7961 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7962 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7963 
7964 		mutex_exit(&stp->sd_lock);
7965 		pollwakeup(&stp->sd_pollist, POLLERR);
7966 		mutex_enter(&stp->sd_lock);
7967 
7968 		if (stp->sd_sigflags & S_ERROR)
7969 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
7970 	}
7971 	mutex_exit(&stp->sd_lock);
7972 }
7973 
7974 /*
7975  * Set a write side error. If persist is set change the socket error
7976  * to persistent.
7977  */
7978 void
7979 strsetwerror(vnode_t *vp, int error, int persist, errfunc_t errfunc)
7980 {
7981 	struct stdata *stp = vp->v_stream;
7982 
7983 	mutex_enter(&stp->sd_lock);
7984 	stp->sd_werror = error;
7985 	if (error == 0 && errfunc == NULL)
7986 		stp->sd_flag &= ~STWRERR;
7987 	else
7988 		stp->sd_flag |= STWRERR;
7989 	if (persist) {
7990 		stp->sd_flag &= ~STWRERRNONPERSIST;
7991 	} else {
7992 		stp->sd_flag |= STWRERRNONPERSIST;
7993 	}
7994 	stp->sd_wrerrfunc = errfunc;
7995 	if (error != 0 || errfunc != NULL) {
7996 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);	/* readers */
7997 		cv_broadcast(&stp->sd_wrq->q_wait);		/* writers */
7998 		cv_broadcast(&stp->sd_monitor);			/* ioctllers */
7999 
8000 		mutex_exit(&stp->sd_lock);
8001 		pollwakeup(&stp->sd_pollist, POLLERR);
8002 		mutex_enter(&stp->sd_lock);
8003 
8004 		if (stp->sd_sigflags & S_ERROR)
8005 			strsendsig(stp->sd_siglist, S_ERROR, 0, error);
8006 	}
8007 	mutex_exit(&stp->sd_lock);
8008 }
8009 
8010 /*
8011  * Make the stream return 0 (EOF) when all data has been read.
8012  * No effect on write side.
8013  */
8014 void
8015 strseteof(vnode_t *vp, int eof)
8016 {
8017 	struct stdata *stp = vp->v_stream;
8018 
8019 	mutex_enter(&stp->sd_lock);
8020 	if (!eof) {
8021 		stp->sd_flag &= ~STREOF;
8022 		mutex_exit(&stp->sd_lock);
8023 		return;
8024 	}
8025 	stp->sd_flag |= STREOF;
8026 	if (stp->sd_flag & RSLEEP) {
8027 		stp->sd_flag &= ~RSLEEP;
8028 		cv_broadcast(&_RD(stp->sd_wrq)->q_wait);
8029 	}
8030 
8031 	mutex_exit(&stp->sd_lock);
8032 	pollwakeup(&stp->sd_pollist, POLLIN|POLLRDNORM);
8033 	mutex_enter(&stp->sd_lock);
8034 
8035 	if (stp->sd_sigflags & (S_INPUT|S_RDNORM))
8036 		strsendsig(stp->sd_siglist, S_INPUT|S_RDNORM, 0, 0);
8037 	mutex_exit(&stp->sd_lock);
8038 }
8039 
8040 void
8041 strflushrq(vnode_t *vp, int flag)
8042 {
8043 	struct stdata *stp = vp->v_stream;
8044 
8045 	mutex_enter(&stp->sd_lock);
8046 	flushq(_RD(stp->sd_wrq), flag);
8047 	mutex_exit(&stp->sd_lock);
8048 }
8049 
8050 void
8051 strsetrputhooks(vnode_t *vp, uint_t flags,
8052 		msgfunc_t protofunc, msgfunc_t miscfunc)
8053 {
8054 	struct stdata *stp = vp->v_stream;
8055 
8056 	mutex_enter(&stp->sd_lock);
8057 
8058 	if (protofunc == NULL)
8059 		stp->sd_rprotofunc = strrput_proto;
8060 	else
8061 		stp->sd_rprotofunc = protofunc;
8062 
8063 	if (miscfunc == NULL)
8064 		stp->sd_rmiscfunc = strrput_misc;
8065 	else
8066 		stp->sd_rmiscfunc = miscfunc;
8067 
8068 	if (flags & SH_CONSOL_DATA)
8069 		stp->sd_rput_opt |= SR_CONSOL_DATA;
8070 	else
8071 		stp->sd_rput_opt &= ~SR_CONSOL_DATA;
8072 
8073 	if (flags & SH_SIGALLDATA)
8074 		stp->sd_rput_opt |= SR_SIGALLDATA;
8075 	else
8076 		stp->sd_rput_opt &= ~SR_SIGALLDATA;
8077 
8078 	if (flags & SH_IGN_ZEROLEN)
8079 		stp->sd_rput_opt |= SR_IGN_ZEROLEN;
8080 	else
8081 		stp->sd_rput_opt &= ~SR_IGN_ZEROLEN;
8082 
8083 	mutex_exit(&stp->sd_lock);
8084 }
8085 
8086 void
8087 strsetwputhooks(vnode_t *vp, uint_t flags, clock_t closetime)
8088 {
8089 	struct stdata *stp = vp->v_stream;
8090 
8091 	mutex_enter(&stp->sd_lock);
8092 	stp->sd_closetime = closetime;
8093 
8094 	if (flags & SH_SIGPIPE)
8095 		stp->sd_wput_opt |= SW_SIGPIPE;
8096 	else
8097 		stp->sd_wput_opt &= ~SW_SIGPIPE;
8098 	if (flags & SH_RECHECK_ERR)
8099 		stp->sd_wput_opt |= SW_RECHECK_ERR;
8100 	else
8101 		stp->sd_wput_opt &= ~SW_RECHECK_ERR;
8102 
8103 	mutex_exit(&stp->sd_lock);
8104 }
8105 
8106 void
8107 strsetrwputdatahooks(vnode_t *vp, msgfunc_t rdatafunc, msgfunc_t wdatafunc)
8108 {
8109 	struct stdata *stp = vp->v_stream;
8110 
8111 	mutex_enter(&stp->sd_lock);
8112 
8113 	stp->sd_rputdatafunc = rdatafunc;
8114 	stp->sd_wputdatafunc = wdatafunc;
8115 
8116 	mutex_exit(&stp->sd_lock);
8117 }
8118 
8119 /* Used within framework when the queue is already locked */
8120 void
8121 qenable_locked(queue_t *q)
8122 {
8123 	stdata_t *stp = STREAM(q);
8124 
8125 	ASSERT(MUTEX_HELD(QLOCK(q)));
8126 
8127 	if (!q->q_qinfo->qi_srvp)
8128 		return;
8129 
8130 	/*
8131 	 * Do not place on run queue if already enabled or closing.
8132 	 */
8133 	if (q->q_flag & (QWCLOSE|QENAB))
8134 		return;
8135 
8136 	/*
8137 	 * mark queue enabled and place on run list if it is not already being
8138 	 * serviced. If it is serviced, the runservice() function will detect
8139 	 * that QENAB is set and call service procedure before clearing
8140 	 * QINSERVICE flag.
8141 	 */
8142 	q->q_flag |= QENAB;
8143 	if (q->q_flag & QINSERVICE)
8144 		return;
8145 
8146 	/* Record the time of qenable */
8147 	q->q_qtstamp = ddi_get_lbolt();
8148 
8149 	/*
8150 	 * Put the queue in the stp list and schedule it for background
8151 	 * processing if it is not already scheduled or if stream head does not
8152 	 * intent to process it in the foreground later by setting
8153 	 * STRS_WILLSERVICE flag.
8154 	 */
8155 	mutex_enter(&stp->sd_qlock);
8156 	/*
8157 	 * If there are already something on the list, stp flags should show
8158 	 * intention to drain it.
8159 	 */
8160 	IMPLY(STREAM_NEEDSERVICE(stp),
8161 	    (stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED)));
8162 
8163 	ENQUEUE(q, stp->sd_qhead, stp->sd_qtail, q_link);
8164 	stp->sd_nqueues++;
8165 
8166 	/*
8167 	 * If no one will drain this stream we are the first producer and
8168 	 * need to schedule it for background thread.
8169 	 */
8170 	if (!(stp->sd_svcflags & (STRS_WILLSERVICE | STRS_SCHEDULED))) {
8171 		/*
8172 		 * No one will service this stream later, so we have to
8173 		 * schedule it now.
8174 		 */
8175 		STRSTAT(stenables);
8176 		stp->sd_svcflags |= STRS_SCHEDULED;
8177 		stp->sd_servid = (void *)taskq_dispatch(streams_taskq,
8178 		    (task_func_t *)stream_service, stp, TQ_NOSLEEP|TQ_NOQUEUE);
8179 
8180 		if (stp->sd_servid == NULL) {
8181 			/*
8182 			 * Task queue failed so fail over to the backup
8183 			 * servicing thread.
8184 			 */
8185 			STRSTAT(taskqfails);
8186 			/*
8187 			 * It is safe to clear STRS_SCHEDULED flag because it
8188 			 * was set by this thread above.
8189 			 */
8190 			stp->sd_svcflags &= ~STRS_SCHEDULED;
8191 
8192 			/*
8193 			 * Failover scheduling is protected by service_queue
8194 			 * lock.
8195 			 */
8196 			mutex_enter(&service_queue);
8197 			ASSERT((stp->sd_qhead == q) && (stp->sd_qtail == q));
8198 			ASSERT(q->q_link == NULL);
8199 			/*
8200 			 * Append the queue to qhead/qtail list.
8201 			 */
8202 			if (qhead == NULL)
8203 				qhead = q;
8204 			else
8205 				qtail->q_link = q;
8206 			qtail = q;
8207 			/*
8208 			 * Clear stp queue list.
8209 			 */
8210 			stp->sd_qhead = stp->sd_qtail = NULL;
8211 			stp->sd_nqueues = 0;
8212 			/*
8213 			 * Wakeup background queue processing thread.
8214 			 */
8215 			cv_signal(&services_to_run);
8216 			mutex_exit(&service_queue);
8217 		}
8218 	}
8219 	mutex_exit(&stp->sd_qlock);
8220 }
8221 
8222 static void
8223 queue_service(queue_t *q)
8224 {
8225 	/*
8226 	 * The queue in the list should have
8227 	 * QENAB flag set and should not have
8228 	 * QINSERVICE flag set. QINSERVICE is
8229 	 * set when the queue is dequeued and
8230 	 * qenable_locked doesn't enqueue a
8231 	 * queue with QINSERVICE set.
8232 	 */
8233 
8234 	ASSERT(!(q->q_flag & QINSERVICE));
8235 	ASSERT((q->q_flag & QENAB));
8236 	mutex_enter(QLOCK(q));
8237 	q->q_flag &= ~QENAB;
8238 	q->q_flag |= QINSERVICE;
8239 	mutex_exit(QLOCK(q));
8240 	runservice(q);
8241 }
8242 
8243 static void
8244 syncq_service(syncq_t *sq)
8245 {
8246 	STRSTAT(syncqservice);
8247 	mutex_enter(SQLOCK(sq));
8248 	ASSERT(!(sq->sq_svcflags & SQ_SERVICE));
8249 	ASSERT(sq->sq_servcount != 0);
8250 	ASSERT(sq->sq_next == NULL);
8251 
8252 	/* if we came here from the background thread, clear the flag */
8253 	if (sq->sq_svcflags & SQ_BGTHREAD)
8254 		sq->sq_svcflags &= ~SQ_BGTHREAD;
8255 
8256 	/* let drain_syncq know that it's being called in the background */
8257 	sq->sq_svcflags |= SQ_SERVICE;
8258 	drain_syncq(sq);
8259 }
8260 
8261 static void
8262 qwriter_outer_service(syncq_t *outer)
8263 {
8264 	/*
8265 	 * Note that SQ_WRITER is used on the outer perimeter
8266 	 * to signal that a qwriter(OUTER) is either investigating
8267 	 * running or that it is actually running a function.
8268 	 */
8269 	outer_enter(outer, SQ_BLOCKED|SQ_WRITER);
8270 
8271 	/*
8272 	 * All inner syncq are empty and have SQ_WRITER set
8273 	 * to block entering the outer perimeter.
8274 	 *
8275 	 * We do not need to explicitly call write_now since
8276 	 * outer_exit does it for us.
8277 	 */
8278 	outer_exit(outer);
8279 }
8280 
8281 static void
8282 mblk_free(mblk_t *mp)
8283 {
8284 	dblk_t *dbp = mp->b_datap;
8285 	frtn_t *frp = dbp->db_frtnp;
8286 
8287 	mp->b_next = NULL;
8288 	if (dbp->db_fthdr != NULL)
8289 		str_ftfree(dbp);
8290 
8291 	ASSERT(dbp->db_fthdr == NULL);
8292 	frp->free_func(frp->free_arg);
8293 	ASSERT(dbp->db_mblk == mp);
8294 
8295 	if (dbp->db_credp != NULL) {
8296 		crfree(dbp->db_credp);
8297 		dbp->db_credp = NULL;
8298 	}
8299 	dbp->db_cpid = -1;
8300 	dbp->db_struioflag = 0;
8301 	dbp->db_struioun.cksum.flags = 0;
8302 
8303 	kmem_cache_free(dbp->db_cache, dbp);
8304 }
8305 
8306 /*
8307  * Background processing of the stream queue list.
8308  */
8309 static void
8310 stream_service(stdata_t *stp)
8311 {
8312 	queue_t *q;
8313 
8314 	mutex_enter(&stp->sd_qlock);
8315 
8316 	STR_SERVICE(stp, q);
8317 
8318 	stp->sd_svcflags &= ~STRS_SCHEDULED;
8319 	stp->sd_servid = NULL;
8320 	cv_signal(&stp->sd_qcv);
8321 	mutex_exit(&stp->sd_qlock);
8322 }
8323 
8324 /*
8325  * Foreground processing of the stream queue list.
8326  */
8327 void
8328 stream_runservice(stdata_t *stp)
8329 {
8330 	queue_t *q;
8331 
8332 	mutex_enter(&stp->sd_qlock);
8333 	STRSTAT(rservice);
8334 	/*
8335 	 * We are going to drain this stream queue list, so qenable_locked will
8336 	 * not schedule it until we finish.
8337 	 */
8338 	stp->sd_svcflags |= STRS_WILLSERVICE;
8339 
8340 	STR_SERVICE(stp, q);
8341 
8342 	stp->sd_svcflags &= ~STRS_WILLSERVICE;
8343 	mutex_exit(&stp->sd_qlock);
8344 	/*
8345 	 * Help backup background thread to drain the qhead/qtail list.
8346 	 */
8347 	while (qhead != NULL) {
8348 		STRSTAT(qhelps);
8349 		mutex_enter(&service_queue);
8350 		DQ(q, qhead, qtail, q_link);
8351 		mutex_exit(&service_queue);
8352 		if (q != NULL)
8353 			queue_service(q);
8354 	}
8355 }
8356 
8357 void
8358 stream_willservice(stdata_t *stp)
8359 {
8360 	mutex_enter(&stp->sd_qlock);
8361 	stp->sd_svcflags |= STRS_WILLSERVICE;
8362 	mutex_exit(&stp->sd_qlock);
8363 }
8364 
8365 /*
8366  * Replace the cred currently in the mblk with a different one.
8367  * Also update db_cpid.
8368  */
8369 void
8370 mblk_setcred(mblk_t *mp, cred_t *cr, pid_t cpid)
8371 {
8372 	dblk_t *dbp = mp->b_datap;
8373 	cred_t *ocr = dbp->db_credp;
8374 
8375 	ASSERT(cr != NULL);
8376 
8377 	if (cr != ocr) {
8378 		crhold(dbp->db_credp = cr);
8379 		if (ocr != NULL)
8380 			crfree(ocr);
8381 	}
8382 	/* Don't overwrite with NOPID */
8383 	if (cpid != NOPID)
8384 		dbp->db_cpid = cpid;
8385 }
8386 
8387 /*
8388  * If the src message has a cred, then replace the cred currently in the mblk
8389  * with it.
8390  * Also update db_cpid.
8391  */
8392 void
8393 mblk_copycred(mblk_t *mp, const mblk_t *src)
8394 {
8395 	dblk_t *dbp = mp->b_datap;
8396 	cred_t *cr, *ocr;
8397 	pid_t cpid;
8398 
8399 	cr = msg_getcred(src, &cpid);
8400 	if (cr == NULL)
8401 		return;
8402 
8403 	ocr = dbp->db_credp;
8404 	if (cr != ocr) {
8405 		crhold(dbp->db_credp = cr);
8406 		if (ocr != NULL)
8407 			crfree(ocr);
8408 	}
8409 	/* Don't overwrite with NOPID */
8410 	if (cpid != NOPID)
8411 		dbp->db_cpid = cpid;
8412 }
8413 
8414 int
8415 hcksum_assoc(mblk_t *mp,  multidata_t *mmd, pdesc_t *pd,
8416     uint32_t start, uint32_t stuff, uint32_t end, uint32_t value,
8417     uint32_t flags, int km_flags)
8418 {
8419 	int rc = 0;
8420 
8421 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8422 	if (mp->b_datap->db_type == M_DATA) {
8423 		/* Associate values for M_DATA type */
8424 		DB_CKSUMSTART(mp) = (intptr_t)start;
8425 		DB_CKSUMSTUFF(mp) = (intptr_t)stuff;
8426 		DB_CKSUMEND(mp) = (intptr_t)end;
8427 		DB_CKSUMFLAGS(mp) = flags;
8428 		DB_CKSUM16(mp) = (uint16_t)value;
8429 
8430 	} else {
8431 		pattrinfo_t pa_info;
8432 
8433 		ASSERT(mmd != NULL);
8434 
8435 		pa_info.type = PATTR_HCKSUM;
8436 		pa_info.len = sizeof (pattr_hcksum_t);
8437 
8438 		if (mmd_addpattr(mmd, pd, &pa_info, B_TRUE, km_flags) != NULL) {
8439 			pattr_hcksum_t *hck = (pattr_hcksum_t *)pa_info.buf;
8440 
8441 			hck->hcksum_start_offset = start;
8442 			hck->hcksum_stuff_offset = stuff;
8443 			hck->hcksum_end_offset = end;
8444 			hck->hcksum_cksum_val.inet_cksum = (uint16_t)value;
8445 			hck->hcksum_flags = flags;
8446 		} else {
8447 			rc = -1;
8448 		}
8449 	}
8450 	return (rc);
8451 }
8452 
8453 void
8454 hcksum_retrieve(mblk_t *mp, multidata_t *mmd, pdesc_t *pd,
8455     uint32_t *start, uint32_t *stuff, uint32_t *end,
8456     uint32_t *value, uint32_t *flags)
8457 {
8458 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8459 	if (mp->b_datap->db_type == M_DATA) {
8460 		if (flags != NULL) {
8461 			*flags = DB_CKSUMFLAGS(mp) & HCK_FLAGS;
8462 			if ((*flags & (HCK_PARTIALCKSUM |
8463 			    HCK_FULLCKSUM)) != 0) {
8464 				if (value != NULL)
8465 					*value = (uint32_t)DB_CKSUM16(mp);
8466 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
8467 					if (start != NULL)
8468 						*start =
8469 						    (uint32_t)DB_CKSUMSTART(mp);
8470 					if (stuff != NULL)
8471 						*stuff =
8472 						    (uint32_t)DB_CKSUMSTUFF(mp);
8473 					if (end != NULL)
8474 						*end =
8475 						    (uint32_t)DB_CKSUMEND(mp);
8476 				}
8477 			}
8478 		}
8479 	} else {
8480 		pattrinfo_t hck_attr = {PATTR_HCKSUM};
8481 
8482 		ASSERT(mmd != NULL);
8483 
8484 		/* get hardware checksum attribute */
8485 		if (mmd_getpattr(mmd, pd, &hck_attr) != NULL) {
8486 			pattr_hcksum_t *hck = (pattr_hcksum_t *)hck_attr.buf;
8487 
8488 			ASSERT(hck_attr.len >= sizeof (pattr_hcksum_t));
8489 			if (flags != NULL)
8490 				*flags = hck->hcksum_flags;
8491 			if (start != NULL)
8492 				*start = hck->hcksum_start_offset;
8493 			if (stuff != NULL)
8494 				*stuff = hck->hcksum_stuff_offset;
8495 			if (end != NULL)
8496 				*end = hck->hcksum_end_offset;
8497 			if (value != NULL)
8498 				*value = (uint32_t)
8499 				    hck->hcksum_cksum_val.inet_cksum;
8500 		}
8501 	}
8502 }
8503 
8504 void
8505 lso_info_set(mblk_t *mp, uint32_t mss, uint32_t flags)
8506 {
8507 	ASSERT(DB_TYPE(mp) == M_DATA);
8508 	ASSERT((flags & ~HW_LSO_FLAGS) == 0);
8509 
8510 	/* Set the flags */
8511 	DB_LSOFLAGS(mp) |= flags;
8512 	DB_LSOMSS(mp) = mss;
8513 }
8514 
8515 void
8516 lso_info_cleanup(mblk_t *mp)
8517 {
8518 	ASSERT(DB_TYPE(mp) == M_DATA);
8519 
8520 	/* Clear the flags */
8521 	DB_LSOFLAGS(mp) &= ~HW_LSO_FLAGS;
8522 	DB_LSOMSS(mp) = 0;
8523 }
8524 
8525 void
8526 lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
8527 {
8528 	ASSERT(DB_TYPE(mp) == M_DATA);
8529 
8530 	if (flags != NULL) {
8531 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO_FLAGS;
8532 		if ((*flags != 0) && (mss != NULL))
8533 			*mss = (uint32_t)DB_LSOMSS(mp);
8534 	}
8535 }
8536 
8537 /*
8538  * Checksum buffer *bp for len bytes with psum partial checksum,
8539  * or 0 if none, and return the 16 bit partial checksum.
8540  */
8541 unsigned
8542 bcksum(uchar_t *bp, int len, unsigned int psum)
8543 {
8544 	int odd = len & 1;
8545 	extern unsigned int ip_ocsum();
8546 
8547 	if (((intptr_t)bp & 1) == 0 && !odd) {
8548 		/*
8549 		 * Bp is 16 bit aligned and len is multiple of 16 bit word.
8550 		 */
8551 		return (ip_ocsum((ushort_t *)bp, len >> 1, psum));
8552 	}
8553 	if (((intptr_t)bp & 1) != 0) {
8554 		/*
8555 		 * Bp isn't 16 bit aligned.
8556 		 */
8557 		unsigned int tsum;
8558 
8559 #ifdef _LITTLE_ENDIAN
8560 		psum += *bp;
8561 #else
8562 		psum += *bp << 8;
8563 #endif
8564 		len--;
8565 		bp++;
8566 		tsum = ip_ocsum((ushort_t *)bp, len >> 1, 0);
8567 		psum += (tsum << 8) & 0xffff | (tsum >> 8);
8568 		if (len & 1) {
8569 			bp += len - 1;
8570 #ifdef _LITTLE_ENDIAN
8571 			psum += *bp << 8;
8572 #else
8573 			psum += *bp;
8574 #endif
8575 		}
8576 	} else {
8577 		/*
8578 		 * Bp is 16 bit aligned.
8579 		 */
8580 		psum = ip_ocsum((ushort_t *)bp, len >> 1, psum);
8581 		if (odd) {
8582 			bp += len - 1;
8583 #ifdef _LITTLE_ENDIAN
8584 			psum += *bp;
8585 #else
8586 			psum += *bp << 8;
8587 #endif
8588 		}
8589 	}
8590 	/*
8591 	 * Normalize psum to 16 bits before returning the new partial
8592 	 * checksum. The max psum value before normalization is 0x3FDFE.
8593 	 */
8594 	return ((psum >> 16) + (psum & 0xFFFF));
8595 }
8596 
8597 boolean_t
8598 is_vmloaned_mblk(mblk_t *mp, multidata_t *mmd, pdesc_t *pd)
8599 {
8600 	boolean_t rc;
8601 
8602 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_MULTIDATA);
8603 	if (DB_TYPE(mp) == M_DATA) {
8604 		rc = (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0);
8605 	} else {
8606 		pattrinfo_t zcopy_attr = {PATTR_ZCOPY};
8607 
8608 		ASSERT(mmd != NULL);
8609 		rc = (mmd_getpattr(mmd, pd, &zcopy_attr) != NULL);
8610 	}
8611 	return (rc);
8612 }
8613 
8614 void
8615 freemsgchain(mblk_t *mp)
8616 {
8617 	mblk_t	*next;
8618 
8619 	while (mp != NULL) {
8620 		next = mp->b_next;
8621 		mp->b_next = NULL;
8622 
8623 		freemsg(mp);
8624 		mp = next;
8625 	}
8626 }
8627 
8628 mblk_t *
8629 copymsgchain(mblk_t *mp)
8630 {
8631 	mblk_t	*nmp = NULL;
8632 	mblk_t	**nmpp = &nmp;
8633 
8634 	for (; mp != NULL; mp = mp->b_next) {
8635 		if ((*nmpp = copymsg(mp)) == NULL) {
8636 			freemsgchain(nmp);
8637 			return (NULL);
8638 		}
8639 
8640 		nmpp = &((*nmpp)->b_next);
8641 	}
8642 
8643 	return (nmp);
8644 }
8645 
8646 /* NOTE: Do not add code after this point. */
8647 #undef QLOCK
8648 
8649 /*
8650  * Replacement for QLOCK macro for those that can't use it.
8651  */
8652 kmutex_t *
8653 QLOCK(queue_t *q)
8654 {
8655 	return (&(q)->q_lock);
8656 }
8657 
8658 /*
8659  * Dummy runqueues/queuerun functions functions for backwards compatibility.
8660  */
8661 #undef runqueues
8662 void
8663 runqueues(void)
8664 {
8665 }
8666 
8667 #undef queuerun
8668 void
8669 queuerun(void)
8670 {
8671 }
8672 
8673 /*
8674  * Initialize the STR stack instance, which tracks autopush and persistent
8675  * links.
8676  */
8677 /* ARGSUSED */
8678 static void *
8679 str_stack_init(netstackid_t stackid, netstack_t *ns)
8680 {
8681 	str_stack_t	*ss;
8682 	int i;
8683 
8684 	ss = (str_stack_t *)kmem_zalloc(sizeof (*ss), KM_SLEEP);
8685 	ss->ss_netstack = ns;
8686 
8687 	/*
8688 	 * set up autopush
8689 	 */
8690 	sad_initspace(ss);
8691 
8692 	/*
8693 	 * set up mux_node structures.
8694 	 */
8695 	ss->ss_devcnt = devcnt;	/* In case it should change before free */
8696 	ss->ss_mux_nodes = kmem_zalloc((sizeof (struct mux_node) *
8697 	    ss->ss_devcnt), KM_SLEEP);
8698 	for (i = 0; i < ss->ss_devcnt; i++)
8699 		ss->ss_mux_nodes[i].mn_imaj = i;
8700 	return (ss);
8701 }
8702 
8703 /*
8704  * Note: run at zone shutdown and not destroy so that the PLINKs are
8705  * gone by the time other cleanup happens from the destroy callbacks.
8706  */
8707 static void
8708 str_stack_shutdown(netstackid_t stackid, void *arg)
8709 {
8710 	str_stack_t *ss = (str_stack_t *)arg;
8711 	int i;
8712 	cred_t *cr;
8713 
8714 	cr = zone_get_kcred(netstackid_to_zoneid(stackid));
8715 	ASSERT(cr != NULL);
8716 
8717 	/* Undo all the I_PLINKs for this zone */
8718 	for (i = 0; i < ss->ss_devcnt; i++) {
8719 		struct mux_edge		*ep;
8720 		ldi_handle_t		lh;
8721 		ldi_ident_t		li;
8722 		int			ret;
8723 		int			rval;
8724 		dev_t			rdev;
8725 
8726 		ep = ss->ss_mux_nodes[i].mn_outp;
8727 		if (ep == NULL)
8728 			continue;
8729 		ret = ldi_ident_from_major((major_t)i, &li);
8730 		if (ret != 0) {
8731 			continue;
8732 		}
8733 		rdev = ep->me_dev;
8734 		ret = ldi_open_by_dev(&rdev, OTYP_CHR, FREAD|FWRITE,
8735 		    cr, &lh, li);
8736 		if (ret != 0) {
8737 			ldi_ident_release(li);
8738 			continue;
8739 		}
8740 
8741 		ret = ldi_ioctl(lh, I_PUNLINK, (intptr_t)MUXID_ALL, FKIOCTL,
8742 		    cr, &rval);
8743 		if (ret) {
8744 			(void) ldi_close(lh, FREAD|FWRITE, cr);
8745 			ldi_ident_release(li);
8746 			continue;
8747 		}
8748 		(void) ldi_close(lh, FREAD|FWRITE, cr);
8749 
8750 		/* Close layered handles */
8751 		ldi_ident_release(li);
8752 	}
8753 	crfree(cr);
8754 
8755 	sad_freespace(ss);
8756 
8757 	kmem_free(ss->ss_mux_nodes, sizeof (struct mux_node) * ss->ss_devcnt);
8758 	ss->ss_mux_nodes = NULL;
8759 }
8760 
8761 /*
8762  * Free the structure; str_stack_shutdown did the other cleanup work.
8763  */
8764 /* ARGSUSED */
8765 static void
8766 str_stack_fini(netstackid_t stackid, void *arg)
8767 {
8768 	str_stack_t	*ss = (str_stack_t *)arg;
8769 
8770 	kmem_free(ss, sizeof (*ss));
8771 }
8772